Lecture Notes On General Relativity Columbia University: January 16, 2013

Lecture Notes on General Relativity
Columbia University
January 16, 2013

Contents
1 Special Relativity 4
1.1 Newtonian Physics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 The Birth of Special Relativity . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 The Minkowski Spacetime R3+1 . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.1 Causality Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.2 Inertial Observers, Frames of Reference and Isometies . . . . . . . . . 11
1.3.3 General and Special Covariance . . . . . . . . . . . . . . . . . . . . . . 14
1.3.4 Relativistic Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4 Conformal Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.1 The Double Null Foliation . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.2 The Penrose Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5 Electromagnetism and Maxwell Equations . . . . . . . . . . . . . . . . . . . . 23
2 Lorentzian Geometry 25
2.1 Causality I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Null Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.3 Global Hyperbolicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.4 Causality II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3 Introduction to General Relativity 42

3.1 Equivalence Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.2 The Einstein Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3 The Cauchy Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4 Gravitational Redshift and Time Dilation . . . . . . . . . . . . . . . . . . . . 46
3.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4 Null Structure Equations 49

4.1 The Double Null Foliation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 Connection Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.3 Curvature Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.4 The Algebra Calculus of S-Tensor Fields . . . . . . . . . . . . . . . . . . . . . 59
4.5 Null Structure Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.6 The Characteristic Initial Value Problem . . . . . . . . . . . . . . . . . . . . . 69
1
5 Applications to Null Hypersurfaces 73
5.1 Jacobi Fields and Tidal Forces . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.2 Focal Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.3 Causality III . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.4 Trapped Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.5 Penrose Incompleteness Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.6 Killing Horizons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6 Christodoulou’s Memory Effect 90

6.1 The Null Infinity I + . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.2 Tracing gravitational waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.3 Peeling and Asymptotic Quantities . . . . . . . . . . . . . . . . . . . . . . . . 95
6.4 The Memory Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7 Black Holes 97
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.2 Black Holes and Trapped Surfaces . . . . . . . . . . . . . . . . . . . . . . . . 98
7.3 Black Hole Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.4 Spherical Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.4.1 General Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.4.2 Schwarzschild Black Holes . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.5 Kerr Black Holes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
8 Lagrangian Theories and the Variational Principle 104

8.1 Matter Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
8.2 The Action Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
8.3 Derivation of the Energy Momentum Tensor . . . . . . . . . . . . . . . . . . . 106
8.4 Application to Linear Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
8.5 Noether’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
9 Hyperbolic Equations 111

9.1 The Energy Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
9.2 A Priori Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
9.3 Well-posedness of the Wave Equation . . . . . . . . . . . . . . . . . . . . . . . 115
9.4 The Wave Equation on Minkowski spacetime . . . . . . . . . . . . . . . . . . 118
10 Wave Propagation on Black Holes 127

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
10.2 Pointwise and Energy Boundedness . . . . . . . . . . . . . . . . . . . . . . . . 127
10.3 Pointwise and Energy Decay . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
2
Introduction
General Relativity is the classical theory that describes the evolution of systems under
the effect of gravity. Its history goes back to 1915 when Einstein postulated that the laws of
gravity can be expressed as a system of equations, the so-called Einstein equations. In order
to formulate his theory, Einstein had to reinterpret fundamental concepts of our experience
(such as time, space, future, simultaneity, etc.) in a purely geometrical framework. The goal
of this course is to highlight the geometric character of General Relativity and unveil the
fascinating properties of black holes, one of the most celebrated predictions of mathematical
physics.
The course will start with a self-contained introduction to special relativity and then
proceed to the more general setting of Lorentzian manifolds. Next the Lagrangian formula-
tion of the Einstein equations will be presented. We will formally define the notion of black
holes and prove the incompleteness theorem of Penrose (also known as singularity theorem).
The topology of general black holes will also be investigated. Finally, we will present explicit
spacetime solutions of the Einstein equations which contain black hole regions, such as the
Schwarzschild, and more generally, the Kerr solution.
3
Chapter 1
Special Relativity
In both past and modern viewpoints, the universe is considered to be a continuum composed
of events, where each event can be thought of as a point in space at an instant of time. We
will refer to this continuum as the spacetime. The geometric properties, and in particular the
causal structure of spacetimes in Newtonian physics and in the theory of relativity greatly
differ from each other and lead to radically different perspectives for the physical world and
its laws.
We begin by listing the key assumptions about spacetime in Newtonian physics and then
proceed by replacing these assumptions with the postulates of special relativity.
1.1 Newtonian Physics

Main assumptions
The primary assumptions in Newtonian physics are the following
1. There is an absolute notion of time. This implies the notion of simultaneity is also
absolute.
2. The speed of light is finite and observer dependent.
3. Observers can travel arbitrarily fast (in particular faster than c).
From the above one can immediately infer the existence of a time coordinate t ∈ R such that
all the events of constant time t compose a 3-dimensional Euclidean space. The spacetime
is topologically equivalent to R4 and admits a universal coordinate system (t, x1 , x2 , x3 ).
Causal structure
Given an event p occurring at time tp , the spacetime can be decomposed into the fol-
lowing sets:
• Future of p: Set of all events for which t > tp .
• Present of p: Set of all events for which t = tp .
• Past of p: Set of all events for which t < tp .
4
More generally, we can define the future (past) of a set S to be the union of the futures
(pasts) of all points of S.
FUTURE
PRESENT p p
PAST
Figure 1.1: The Newtonian universe and its causal structure
From now on we consider geometric units with respect to which the light travels at speed
c = 1 relative to observers at rest. If an observer at p emits a light beam in all directions
of space, then the trajectory of this beam in spacetime will be a null cone with vertex at p.
We can complete this cone by considering the trajectory of light beams that arrive at the
event p.
FUTURE
PRESENT p
PAST
Figure 1.2: The Newtonian universe and light trajectories
It is important to emphasize that in Newtonian theory, in view of the existence of the

absolute time t, one only works by projecting on the Euclidean space R3 and considering all
quantities as functions of (the space and) time t.
Newton’s theory gives a very accurate theory for objects moving at slow speeds in absence
of strong gravitational fields. However, in several circumstances difficulties arise:
1. A more philosophical issue is that in Newtonian theory an observer is either at rest

or in motion. But how could one determine if an observer O is (universally) at rest?
Why can’t a uniformly moving (relative to O) observer P be considered at rest since
P is also not affected by any external influence?
5
2. For hundreds of years it has been known that in vacuum light propagates at a very high
but constant speed, and no material has been observed to travel faster. If an observer
P is moving at speed c/5 (relative to an observer O at rest) towards a light beam
(which is moving at speed c relative to O), then the light would reach the observer
P at speed c + c/5. However, astronomical observations of double stars should reveal
such fast and slow light, but in fact the speeds turn out to be the same.
3. Light is the propagation of an electromagnetic disturbance and electromagnetic fields
are governed by Maxwell’s equations. However, these equations are not well-behaved
in Newtonian theory; in particular, in this context, these laws are observer dependent
and hence do not take the desired form of universal physical laws.
One of Einstein’s contributions was his persistence that every physical law can be ex-
pressed independently of the choice of coordinates (we will return to this point later). It
was this persistence along with his belief that Maxwell’s equations are flawless that led to
what is now known as special relativity.
1.2 The Birth of Special Relativity

In 1905 Einstein published a paper titled “On the electrodynamics of moving bodies”, where
he described algebraic relations governing the motion of uniform observers so that Maxwell
equations have the same form regardless of the observer’s frame. In order to achieve his
goal, Einstein had to assume the following
1. There is no absolute notion of time.
2. No observer or particle can travel faster than the speed of light c. The constant c
should be considered as a physical law and hence does not depend on the observer who
measures it.
The above immediately change the Newtonian perception for the spacetime, since under
Einstein’s assumptions the future (past) of an event p is confined to be the interior of the
future (past) light cone with vertex at p.
In 1908, Hermann Minkowski showed that Einstein’s algebraic laws (and, in particular,
the above picture) can be interpreted in a purely geometric way, by introducing a new kind
of metric on R4 , the so-called Minkowski metric.
6
1.3 The Minkowski Spacetime R3+1
Definition
A Minkowski metric g on the linear space R4 is a symmetric non-degenerate bilinear

form with signature (−, +, +, +). In other words, there is a basis {e0 , e1 , e2 , e3 } such that
g(eα , eβ ) = gαβ , α, β ∈ {0, 1, 2, 3} ,
where the matrix gαβ is given by

 
−1 0 0 0
0 1 0 0
g=
0
.
0 1 0
0 0 0 1
Given such a frame (which for obvious reason will be called orthonormal), one can readily
construct a coordinate system (t, x1 , x2 , x3 ) of R4 such that at each point we have
e0 = ∂t , ei = ∂xi , i = 1, 2, 3.
Note that from now on in order to emphasize the signature of the metric we will denote the
Minkowski spacetime by R3+1 . With respect to the above coordinate system, the metric g
can be expressed as a (0,2) tensor as follows:
g = −dt2 + (dx1 )2 + (dx2 )2 + (dx3 )2 . (1.1)
Note that (for an arbitrary pseudo-Riemannian metric) one can still introduce a Levi–Civita
connection and therefore define the notion the associated Christoffel symbols and geodesic
curves and that of the Riemann, Ricci and scalar curvature. One can also define the volume
form such that if Xα , α = 0, 1, 2, 3, is an orthonormal frame then (X0 , X1 , X2 , X3 ) = ±1.
In the Minkowski case, the curvatures are all zero and the geodesics are lines with respect
to the coordinate system (t, x1 , x2 , x3 ).
1.3.1 Causality Theory

The fundamental new aspect of this metric is that it is not positive-definite. A vector
X ∈ R3+1 is defined to be:
1. spacelike, if g(X, X) > 0,
2. null, if g(X, X) = 0,
3. timelike, if g(X, X) < 0.

If X is either timelike or null, then it is called causal. If X = (t, x1 , x2 , x3 ) is null vector
at p, then
t2 = (x1 )2 + (x2 )2 + (x3 )2
and hence X lies on cone with vertex at p. In other words, all null vectors at p span a
double cone, known as the double null cone. We denote by Sp ,
Sp = X ∈ Tp R3+1 : g(X, X) > 0 ,

7
the set of spacelike vectors at p, by Ip ,
Ip = X ∈ Tp R3+1 : g(X, X) < 0 ,

the set of timelike vectors at p, and by Np ,

Np = X ∈ Tp R3+1 : g(X, X) = 0 ,

the set of null vectors at p. The set Sp is open (and connected if n > 1). Note that the set Ip
is the interior of the solid double cone enclosed by Np . Hence, Ip is an open set consisting
of two components which we may denote by Ip+ and Ip− . Then we can also decompose
Np = Np+ ∪ Np− , where
Np+ = ∂Ip+ , Np− = ∂Ip− .
The following questions arise: How is Np+ (or Ip+ ) related to the future and past of p?
How can one even discriminate between the future and past? To answer this we need to
introduce the notion of time-orientability. A time-orientation of (R3+1 , g) is a continuous
choice of a positive component Ip+ at each p ∈ R3+1 . Then, we call Ip+ (resp. Np+ ) the set
of future-directed timelike (resp. null) vectors at p. Similarly we define the past-directed
causal vectors. We also define:
• The causal future J + (p) of p by J + (p) = Ip+ ∪ Np+ .
• The chronological future of p to simply be Ip+ .
We also define the causal future J + (S) of a set S by
[
J + (S) = Jp+ .
p∈S
Causality, observers and particles, proper time
We can readily extend the previous causal characterizations for curves. In particular,
.
a curve α : I → R3+1 is called future-directed timelike if α(t) is a future-directed timelike
vector at α(t) for all t ∈ I. Note that the worldline of an observer is represented by a
timelike curve. If the observer is inertial, then he/she moves on a timelike geodesic. On the
other hand, photons move on null geodesics (and hence information propagates along null
geodesics).
8
We, therefore, see that the Minkowski metric provides a very elegant way to put the
assumptions of Section 1.2 in a geometric context.
The proper time τ of an observer O is defined to be the parametrization of its worldline α
. .
such that g(α(τ ), α(τ )) = −1. Note that no proper time can be defined for photons since for
. .
all parametrizations we have g(α(τ ), α(τ )) = 0. However, in this case, it is usually helpful
.
to consider affine parametrizations τ with respect to which ∇α. α = 0.
Hypersurfaces
Finally, we have the following categories for hypersurfaces:
1. A hypesurface H is called spacelike, if the normal Nx at each point x ∈ H is timelike.

In this case, g is positive-definite (i.e. H is a Riemannian manifold).

Tx H
2. A hypesurface
H is called null, if the normal Nx at each point x ∈ H is null. In this
case, g is degenerate.

Tx H
3. A hypesurface H is called timelike, if the normal Nx at each point x ∈ H is spacelike.

In this case, g has signature (−, +, +).

Tx H
Examples of spacelike hypersurfaces
1. The hypersurfaces
Hτ = {t = τ }

are spacelike, since their unit normal is the timelike vector field ∂0 . Then, Hτ , g

Tx Hτ
is isometric to the 3-dimensional Euclidean space E3 .
2. The hypersurface
H 3 = {X : g(X, X) = −1 and X future-directed }
can easily verify that X is the normal to H 3 at

is a spacelike hypersurface.Indeed, one

the endpoint of X. Then, H 3 , g is isometric to the 3-dimensional hyperbolic

3
Tx H
space. The following figure depicts the radial projection which is an isometry from H 3
to the disk model.
9
Note that the hyperboloid consists of all points in spacetime where an observer located
at (0, 0, 0, 0) can be after proper time 1.
We remark that the Cauchy–Schwarz inequality is reversed for timelike vectors. If
X, Y are two future-directed timelike vector fields then g(X, Y ) < 0 and
p
−g(X, Y ) ≥ g(X, X) · g(Y, Y ).
Hence, there exists a real number φ such that
−g(X, Y )
cosh(φ) = p .
g(X, X) · g(Y, Y )
Then φ is called the hyperbolic angle of X and Y .
Examples of timelike hypersurfaces
1. The hypersurfaces
Tτ = {x1 = τ }

are timelike, since their normal is the spacelike vector field ∂x1 . Then, Tτ , g is

Tx Tτ
isometric to the 3-dimensional Minkowski space R2+1 .
2. The hypersurface
3
H+ = {X : g(X, X) = 1 and X future-directed }
3
is a timelike hypersurface. Indeed, one can easily verify that X is the normal to H+
at the endpoint of X.
Examples of null hypersurfaces
10
1. Let n = (n0 , n1 , n2 , n3 ) be a null vector. The planes given by the equation
Pn = (t, x1 , x2 , x3 ) : n0 t = n1 x1 + n2 x2 + n3 x3

are null hypersurfaces, since their normal is the null vector n.
2. The (null) cone

n p o
C = (t, x1 , x2 , x3 ) : t = (x1 )2 + (x2 )2 + (x3 )2
is also a null hypersurface. Its tangent plane at the endpoint of n is the plane Pn and
hence its normal is the null vector n. Note that C = N + (O), where O is the origin.
Note also that N − (O) is given by
n p o
C = (t, x1 , x2 , x3 ) : t = − (x1 )2 + (x2 )2 + (x3 )2 .
The above can be summarized in the following figure:

null
timelike geodesic = timelike hypersurface
inertial obesrver hypersurface
sic
de
timelike curve =
eo
accelerating observer
ll g
nu
spacelike
hypersurface
1.3.2 Inertial Observers, Frames of Reference and Isometies

Inertial Frames
Let O be an inertial observer moving on a timelike geodesic α. Let t be the proper time
of O and (x1 , x2 , x3 ) be a Euclidean coordinate system of the (spacelike) plane orthogonal
to α at α(0). We will refer to the frame (t, x1 , x2 , x3 ) as the frame associated to the inertial
observer O.
Relativity of Time
11
Let O0 be another (inertial or not) observer moving on the timelike curve α0 which
can be expressed as α0 (τ ) = (t(τ ), x1 (τ ), x2 (τ ), x3 (τ )) with respect to the frame associ-
ated to O. Note that since α0 is future-directed curve, the function t 7→ t(τ ) is invert-
ible and hence we can write τ = τ (t) giving us the following parametrization α0 (t) =
(t, x1 (τ (t)), x2 (τ (t)), x3 (τ (t))). Note that the observer O at its proper time t sees that O0 is
located 0 1 2 3 0
d at0 the point αO (t) = (x (τ (t)), x (τ (t)), x (τ (t))). Hence, the speed of O relative to
O is dt αO (t) . As we shall see, we do not always have that t = τ , in other words the time

is relative to each observer.
Let O0 be an inertial observer passing through the origin (0, 0, 0, 0) with respect to the
frame of O and also moving in the x1 -direction at speed v with respect to O. Then, O0 moves
. .
on the curve α0 (t) = (t, vt, 0, 0); however, t is not the proper time of O0 since g(α(t), α(t)) =
−1 + v 2 . However, if we consider the following parametrization α0 (τ ) = (γτ, γτ v, 0, 0), where
1
γ=√ ,
1 − v2
.0 .0
then τ is the proper time for O0 since g(α (τ ), α (τ )) = −1. Note that γ = cosh(φ), where φ
is the hyperbolic angle of α(t = 1) = (1, 0, 0, 0) and α0 (τ = 1).
Time dilation: If τ is the proper time of O0 , then from the above representation we obtain
that t = γτ along α0 . Note that γ > 1 and so t > τ , which implies that the proper time
for the moving (with respect to O) observer O0 runs slower compared to the time t that O
measures for O0 .
Isometries: Lorentz transformations

One of the most striking properties of Minkowski spacetime is that there exists an isom-
etry F which maps O to O0 . Relative to the frame associated to O this map takes the
form:
F (t, x1 , x2 , x3 ) = (γ · (t + vx1 ), γ · (x1 + vt), x2 , x3 ).
Then
F [O(λ)] = F (λ, 0, 0, 0) = (γλ, γλv, 0, 0) = α0 (λ) = O0 (λ),
12
where O(λ), O0 (λ) denotes the position of the observers O, O0 , respectively, at proper time
λ. Clearly, the map F , being an isometry, leaves the proper time of observers invariant.
Note that although there is no absolute notion of time (and space consisting of simultane-
ous events), null cones are absolute geometric constructions that do not depend on observers
(in particular, the isometry F leaves the null cones invariant).
The isometry F is known as Lorentz transformation, or Lorentz boost in the x1 −direction
or hyperbolic rotation. The latter name arises from the fact that such an isometry can be
represented by a matrix whose form is similar to a Euclidean rotation but with trigonometric
functions and angles replaced by hyperbolic trigonometric functions and angles. Note that
the isometries F = Fv (with v 2 ≤ 1) correspond to the flow of the Killing field
Hx1 = t∂x1 + x1 ∂t .
Clearly, one can define boosts in any direction.
Relativity of simultaneity
Having understood the relativity of time, we next proceed by investigating the relativity
of space. The points simultaneous with observer O at the origin are all points such that
t = 0. Then,
F [{t = 0}] = F (0, x1 , x2 , x3 ) = (γx1 v, γx1 , x2 , x3 ).

We extend t0 such that {t0 = 0} = F [{t = 0}] and define the coordinates (x1 )0 , (x2 )0 , (x3 )0
on the hypersurface t0 = 0 such that if p ∈ {t0 = 0}, then p = (x1 )0 , (x2 )0 , (x3 )0 if F −1 (p) =

(0, x1 , x2 , x3 ). Note that the plane t0 = 0 consists of all points simultaneous with 0
O at the
0 1 0 2 0 3
origin and that one can define a global coordinate system t , (x ) , (x ) , (x ) . This is the0
reference frame associated to O0 . In view of the fact that F is an isometry, the metric takes
the form (1.1) with respect to the system t0 , (x1 )0 , (x2 )0 , (x3 )0 .
´
´
´
13
Length Contraction: First note that distance between events is meaningful only for ob-
servers who consider the events to be simultaneous. At proper time τ = 1 of O0 , observer O0
measures that his distance from observer O is v. However, when τ = 1, the measurements of
O are such that O0 is located at the point (γ, γv, 0, 0) with respect to the frame of O. Hence,
the distance (after τ = 1) that O measures between O and O0 is γv > v. In other words, O0
measures shorter distances than O does. This phenomenon is called length contraction and
is intimately connected to time dilation.
´
´
´
O
O´
Figure 1.3: The difference of the hyperbolic lengths of the blue (resp. red) segments repre-
sents length contraction (resp. time dilation).
1.3.3 General and Special Covariance

The general covariance principle allows us to put physics in a geometric framework:
• General covariance principle: All physical laws are independent of the choice of
a particular coordinate system. In other words, the equations expressing physical laws
must be written in terms of tensors.
Since the only tensor that we have in special relativity is the metric g, then we can in
fact assume the following:
• Strong general covariance principle: All physical laws can be expressed in terms
of the metric g and its tensorial expressions.
Hence, physical laws change covariantly (i.e. tensorially) under change of coordinates
(i.e. under diffeomorphisms). If we apply this principle for Minkowski spacetime and also
restrict to its isometries, then we obtain the following:
• Special covariance: The physical laws take exactly the same form when expressed in
terms of the reference frames associated to inertial observers.
14
1.3.4 Relativistic Mechanics
1. Time dilation
We will here generalize the result of the previous section on time dilation. Let O be
an inertial observer and O0 another (inertial or accelerating) observer. Let τ be the proper
time of O0 such that his/her trajectory is α0 (τ ) = (t(τ )), x1 (τ ), x2 , (τ ), x3 (τ )). As before, we
can write τ = τ (t) and hence α0 (τ (t)) = (t, x1 (τ (t), x2 (τ (t), x3 (τ (t)). If v is the speed of O0
.0 .0
relative to O, the fact that g(α (τ ), α (τ )) = −1 implies that
dt 1
=√ ≥ 1 ⇒ t(τ ) ≥ τ.
dτ 1 − v2
2. Absoluteness of speed of light
If a particle moves of a null curve α0 (τ ) = (t(τ )), x1 (τ ), x2 , (τ ), x3 (τ )) then the speed of

the particle with respect to an inertial observer O is
dαO0
dτ
v= = 1,
dτ dt
.0 .0
where the last equation follows from g(α (τ ), α (τ )) = 0.
3. Energy-momentum
Let α(τ ) represent the trajectory of a particle p with mass m. Then we have the following
definitions:
. dα
• The 4-velocity of p is the vector U = α(τ ) = dτ .
• The energy-momentum of p is the vector P = mU .

Let’s now see how an inertial observer O measures the above quantities. If the particle p
moves at speed v relative to O, then by considering the frame associated to O, we obtain:
dα dt m dα
P =m =√ .
dt dτ 1 − v 2 dt
1. The spatial component of the energy-momentum vector P is
m dαO
PO = √ ,
1 − v dt
2
which is called the momentum of p as measured by O. (Compare this definition with

the Newtonian definition in case v ∼ 0.)
2. The temporal component of the energy-momentum vector P is
m 1
EO = √ = m + mv 2 + O(v 4 ),
1−v 2 2
which is called the total energy of p as measured by O. (Note, in particular, that
EO contains the kinetic energy 21 mv 2 ). Hence, the mass m is seen to be energy. If
v = 0, then energy of p as measured by a co-observer is E = m, and by converting in
conditional units, we obtain Einstein’s famous equation
E = mc2 .
15
1.4 Conformal Structure
One is often interested in investigating the properties of isolated systems. In such situation
one should only consider the local system and hence ignore the influence of matter at far
distances. In other words, the asymptotic structure of the spacetime describing the geom-
etry of an isolated system should like the asymptotical structure of Minkowski (recall that
Minkowski spacetime represents the geometry a vacuum static highly symmetric topolog-
ically trivial universe). The goal of this section is to describe the global and asymptotic
causal structure of Minkowski spacetime
One could start by considering the following foliation of Minkowski:
[
R3+1 = Hτ ,
τ ∈R
where the spacelike hypersurfaces Hτ = {t = τ } are as defined in Section 1.3. Note, however,
this foliation does not capture the properties of null geodesics whose importance is manifest
from the fact that signals travel along such curves. Indeed, an observer (like ourselves on
earth) located far away from an isolated system under investigation must understand the
asymptotic behavior of null geodesics in order to be able to measure radiation and other
information sent from this system. For these reason, we will consider a foliation of Minkowski
spacetime which captures the geometry of null geodesics emanating from points of a timelike
geodesic. This is the so-called double null foliation.
As a final remark, note that since we want to understand the asymptotic structure, it will
be convenient to apply conformal transformations on Minkowski spacetime in order to bring
points at ‘infinity’ to finite distance. This procedure will allows us to reveal the structure
of infinity. Note that conformal transformations preserve the causal structure, since they
send timelike curves to timelike curves, spacelike curves to spacelike curves and null curves
to null curves. In fact, conformal transformations send null geodesics to null geodesics (see
Section 2.2).
1.4.1 The Double Null Foliation

Let us consider the timelike geodesic α(t) = (t, 0, 0, 0) where the coordinates are taken with
respect to an inertial coordinate system (t, x1 , x2 , x3 ). Recall that the future-directed null
cone CO with vertex at O = α(0) is given by
n p o
C0 = (t, x1 , x2 , x3 ) : t = (x1 )2 + (x2 )2 + (x3 )2 ≥ 0 ,
whereas the past-directed null cone C O with vertex at O = α(0) is given by

n p o
C 0 = (t, x1 , x2 , x3 ) : t = − (x1 )2 + (x2 )2 + (x3 )2 ≤ 0 .
In order to simplify the above expressions and capture the spherical symmetry of the null
cones, it is convenient to introduce spherical coordinates (r, θ, φ) for the Euclidean hyper-
surfaces Hτ such that r = 0 corresponds to the curve α. Then, in (t, r, θ, φ) coordinates, the
Minkowski metric takes the form
g = −dt2 + dr2 + r2 · gS2 (θ,φ) ,
16
where gS2 (θ, φ) = dθ2 + (sin θ)2 dφ2 is the standard metric on the unit sphere. Then the
future-directed null cone Cτ with vertex at α(τ ) is given by
Cτ = {(t, r, θ, φ) : t − r = τ, τ ∈ R} ,
whereas the past-directed null cone C τ with vertex at α(τ ) is given by
C τ = {(t, r, θ, φ) : t + r = τ, τ ∈ R} .
The above suggest that it is very convenient to convert to null coordinates (u, v, θ, φ) defined
such that
u = t − r,
v = t + r.
Note also that

v≥u (1.2)
and v = u if and only if r = 0. The metric with respect to null coordinates (u, v, θ, φ) takes
the form
1
g = −dudv + (u − v)2 · gS2 (θ,φ)
4
and the double null folation is given by the equations
Cτ = {(u, v, θ, φ) : u = τ, τ ∈ R} ,
C τ = {(u, v, θ, φ) : v = τ, τ ∈ R} .
Note that ∂v (resp. ∂u ) is tangential to the null geodesics of the null cones Cτ (resp.
C τ ). For a generalization of the double null foliation see Section 4.1.
17
1.4.2 The Penrose Diagram
The aim of the section is to describe the asymptotic structure of Minkowski space. In
particular, we want to draw a “bounded” diagram whose boundary represents infinity and
somehow respects the causal structure of Minkowski.
Clearly, v → +∞ along the null cones Cτ and similarly u → +∞ along the null cones C τ .
In order to bring the endpoint of null geodesics in finite distance, we consider the following
change of coordinates:
tan p = v,
(1.3)
tan q = u,
π π

with p, q ∈ 2, 2 and p ≥ q. Then in (p, q, θ, φ) coordinates the metric takes the form

1 1 2
g= −dp dq + sin (p − q) · gS2 (θ,φ) .
(cos p)2 · (cos q)2 4
As expected, a consequence of the boundedness of the range of p, q is that the factor

1
(cos p)2 ·(cos q)2
blows up as p → ± π2 or q → ± π2 . In order to overcome this degeneracy,
we consider the metric g̃ which in (p, q, θ, φ) takes the form
1
g̃ = −dp dq + sin2 (p − q) · gS2 (θ,φ) . (1.4)
4
Clearly the metric g̃ is conformal to g. Note that ∇p = −∂q , ∇q = −∂p , where the ∇ is
considered with respect to g̃, and therefore, the hypersurfaces
eτ = {(p, q, θ, φ) : q = τ, τ ∈ R} ,
C
e τ = {(p, q, θ, φ) : p = τ, τ ∈ R} .
C
are null (with respect to g̃). Hence, if we suppress one angular direction, we can globally
depict the manifold (M,
f g̃) covered by the coordinates (p, q, θ, φ) as follows:
18
,
f g̃), which is conformal to Minkowski R3+1 .

Figure 1.4: The manifold (M,
We define:
• Future null infinity I + to be the endpoints of all future-directed null geodesics along
which r → +∞.
• Future timelike infinity i+ to be the endpoints of all future-directed timelike geodesics.
• Spacelike infinity i0 to be the endpoint of all spacelike geodesics. This is in fact a point,
and not a sphere, which can be thought of as the point at infinity of the one-point
compactification of, say, the spacelike hypersurface t = 0.
Similarly, we define the past null infinity I − and past timelike infinity i− . Note that if we
want to study an isolated system (located say at r = 0), then we should think of ourselves as
moving along the future null infinity I + and receiving the radiation/information sent from
the system.
The boundary of the Figure 1.4 nicely depicts the above asymptotic structures of Minkowski
space R3+1 .
19
timelike
timelike infinity
geodesic
null
null infinity
geodesic
spacelike spacelike
infinity infinity
spacelike
geodesic
Figure 1.5: Identification of the boundary of the conformal diagram of Minkowski
Note that if a curve α(τ ) = (t(τ ), r(τ ), θ(τ ), φ(τ )) is such that as τ → +∞
• t → +∞ and r < ∞, then α approaches i+ .
• t ∼ r → +∞, then α approaches I + .
• |t| < ∞ and r → +∞, then α approaches i0 .
• t ∼ −r → −∞, then α approaches I − .
• t → −∞ and r < ∞, then α approaches i− .
Remark 1.4.1. Although the above figures for (M, f g̃) are very intuitive as been thought
3+1
embedded in the Minkowski space R , they are not rigorous since the metric g̃ is not
isometric to the Minkowski metric. Indeed, the factor sin2 (p−q) highly distorts the geometry
eτ ∩ C
of the spheres Sτ,τ 0 = C e τ 0 compared to the Minkowskian case. Not only this factor
vanishes for the points where p = q (which correspond to the points where u = v and hence
r = 0) but also degenerates as p → π2 and q → − π2 , and hence the point denoted by i0 above
collapses to a point. In fact, the metric g̃ is locally isometric to the Einstein static metric
which is the natural Lorentzian metric on R × S3 . The latter can be thought of as the metric
induced on the cylinder R × S3 embedded in the Minkowski space R4+1 .
20
Figure 1.6: The red segments correspond to the boundary of the conformally compactified
Minkowski spacetime in the Einstein static universe E = R × S3 . The points i+ , i0 , i− are
also points in E. Every other point in M,
f apart from those that lie on the line connecting
+ −
i and i , represent a two-sphere in E.
We can proceed further by suppressing all angular directions. Formally speaking, we

consider the quotient M/SO(3).
f Then, by (1.4), the metric on the quotient is simply
gquot = −dp dq, which coincides with the Minkowski spacetime R1+1 . Hence, if we consider
a planar section of Figure 1.5, then the resulting bounded 2-dimensional domain is embedded
in the Minkowski spacetime R1+1 :
timelike
infinity
null
infinity
spacelike
infinity
Figure 1.7: The Penrose diagram of Minkowski spacetime R3+1 .
This diagram is called the Penrose diagram of Minkowski spacetime R3 . All cones are
21
collapsed to lines. Using the above diagram, one can read off the causal structure of the
spacetime as is shown below:
Figure 1.8: The Penrose diagram and causal structure.
More generally, one defines the Penrose diagram of a spherically symmetric spacetime to
be the image of a bounded conformal transformation of the quotient spacetime in Minkowski
spacetime R1+1 . As already mentioned, the importance of such diagrams stems from fact
that they allow one to read off the causal structure and recognize the asymptotic structure
of a spacetime.
For example, we see that in Minkowski, the past of future null infinity I + (where r = ∞)
is the whole spacetime. That is to say, any point can send signals and thus communication
with I + . However, we can construct spacetimes for which this is not the case. In other words,
there are spacetimes which contain points which cannot communicate I + . The conformal
diagram of such spacetime would be as follows:
22
black hole event
horizon
ll
nu nity
fi
in
Figure 1.9: The shaded region is known as the black hole region.
The shaded region cannot send signal to I + and for this reason is called black hole.
Black holes are one of the most celebrated predictions of (general) relativity and, in fact,
of mathematical physics. These notes will investigate geometric and analytic properties of
such regions.
Remark 1.4.2. The conformal symmetries of Minkowski spacetime are closely related to
the global structure of the Penrose diagrams. For more details see Part II of these notes.
1.5 Electromagnetism and Maxwell Equations

We conclude this section by an account of the relativistic version of Maxwell’s equation gov-
erning the propagation of electromagnetic fields. After all, special relativity was discovered
in order to incorporate Maxwell’s theory.
The electric E and magnetic B fields are now incorporated in a new 2-form F on
Minkowski space R2 and the Maxwell equations take the form
dF = 0, DivF = −4πJ,
where J is a 1-form representing the current density of electric charges. Note that since F
is antisymmetric DivDivF = 0 and hence DivJ = 0. If denotes the volume form in R3+1 ,
then E, B are recovered from F by the equations
 
0 E1 E2 E3
−E1 0 −B3 B2 
Fµν = −E2 B3
.
0 −B1 
−E3 −B2 B1 0
Since dF = 0, by Poincaré lemma we have that there exists a 1-form A, called the Maxwell
potential, such that F = dA. If A is such a 1-form, then for any function f , the 1-form A+df
23
is also a Maxwell potential. Note that one can choose f such that the Maxwell potential
satisfies DivA = 0. This choice is called the Lorentz gauge. Under this gauge, the Maxwell
equations take the form:
DivA = 0,
2g A = DivdA = −4πJ.
Clearly, these equations are covariant under change of coordinates and invariant under isome-
tries.
24
Chapter 2
Lorentzian Geometry
Introduction
As we have already seen, one needs to deal with curved spaces even in the framework of
special relativity (recall that the set {X : g(X, X) = 1} is a “curved” timelike hypersurface).
In this chapter we provide the general framework for curved spaces and introduce the notions
of Lorentzian geometry which are necessary for understanding the mathematical aspects of
general relativity.
• A Lorentzian manifold (M, g) is a differentiable manifold of dimension n + 1, endowed

with a Lorentzian metric g, namely a differentiable assignment of a symmetric, non-
degenerate bilinear form gx with signature (−, +, · · · , +) in Tx M at each x ∈ M.
We will mostly consider the case of four spacetime dimensions (and hence n = 3); though
our results apply for the general case too.
• A spacetime manifold (M, g) is an orientable four-dimensional Lorentzian manifold.
In contrast to the Riemannian case, a given differentiable manifold might not admit a
Lorentzian metric. The reason being that in a Lorentzian manifold there are “distinguished”
directions which basically correspond to the (−) of the signature. One can thus easily show
the following
Proposition 2.0.1. A smooth manifold M admits a Lorentzian metric if and only if it
admits a non-vanishing vector field; in other words, M admits a Lorentzian metric if and
only if it is either non-compact or compact with vanishing Euler characteristic.
The formulas in Riemannian geometry for geodesics, parallel transport, curvatures etc. carry
over to Lorentzian manifolds.
2.1 Causality I
Basic notions
The fundamental new aspect of Lorentzian metrics g is that gx is not positive-definite

in Tx M. In fact, for each x ∈ M, the linear space1 (Tx M, g) is isometric to the Minkowski
1
For simplicity, we will drop the index and thus by g we will also mean gx .
25
spacetime R3+1 and therefore there exists a basis (E0 , E1 , E2 , E3 )) of Tx M such that
g(Eα , Eβ ) = mαβ ,
where mαβ is the Minkowski diagonal matrix (−1, +1, +1, +1). Then, for any vector X ∈
Tx M we have X = α X α Eα and thus
P
g(X, X) = −(X 0 )2 + (X 1 )2 + (X 2 )2 + (X 3 )2 .
Note that unless (M, g) is locally isometric to Minkowski space the frame (Eα , α = 0, 1, 2, 3)
does not correspond to a coordinate frame. As in the flat Minkowski case, we say that a
vector X ∈ Tx M is:
1. spacelike, if g(X, X) > 0,
2. null, if g(X, X) = 0,
3. timelike, if g(X, X) < 0.
If X is either timelike or null, then it is called causal. If X is null vector at p, then
(X 0 )2 = (X 1 )2 + (X 2 )2 + (X 3 )2
and hence X lies on cone (which depends on the basis Eα ) with vertex at x. In other words,
all null vectors at x span a double cone, known as the double null cone. We denote by Nx ,
Nx = X ∈ Tx R3+1 : g(X, X) = 0 ,

the set of null vectors in Tx M, by Ix ,
Ix = X ∈ Tx R3+1 : g(X, X) < 0 ,

the set of timelike vectors in Tx M, and by Sx ,
Sx = X ∈ Tx R3+1 : g(X, X) > 0 ,

the set of spacelike vectors in Tx M. The set Sx is open (and connected if n > 1).
Note that the set Ix is the interior of the solid double cone enclosed by Nx . Hence, Ix
is an open set consisting of two components which we may denote by Ix+ and Ix− . Then we
can also decompose Nx = Nx+ ∪ Nx− , where
Nx+ = ∂Ix+ , Nx− = ∂Ix− .
26
A time-orientation of (M, g) is a continuous choice of a positive component Ix+ at each
x ∈ M. Then, we call Ix+ (resp. Nx+ ) the set of future-directed timelike (resp. null) vectors
at x. Similarly we define the past-directed causal vectors.
Causality, observers and particles, proper time
We can readily extend the previous causal characterizations for curves. In particular,
.
a curve α : I → M is called future-directed timelike if α(t) ∈ Tα(t) M is a future-directed
timelike vector at α(t) for all t ∈ I. Note that the worldline of an observer is represented
by a timelike curve. As we shall see in the sequel, if an observer is inertial, then he/she
moves on a timelike geodesic. On the other hand, photons move on null geodesics (and
hence information propagates along null geodesics).
The proper time τ of an observer O is defined to be the parametrization of its worldline α
. .
such that g(α(τ ), α(τ )) = −1. Note that no proper time can be defined for photons since for
. .
all parametrizations we have g(α(τ ), α(τ )) = 0. However, in this case, it is usually helpful
.
to consider affine parametrizations τ with respect to which ∇α. α = 0.
Submanifolds
Let N be a submanifold of M. Then N is called:

1. spacelike (or Riemannian), if the induced metric g is positive-definite (and thus

Tx N
Riemannian).

2. timelike (or Lorentzian), if the induced metric g has signature (−, +, · · · , +) and

Tx N
thus is Lorentzian.

3. null, if the induced metric g is degenerate. Recall that a symmetric bilinear form

Tx N
d on a linear space V is called degenerate if there exists a vector X ∈ V such that
d(X, Y ) = 0 for all Y ∈ V .
By Sylvester’s law of inertia, these are all the possible case for the induced metric in a
Lorentzian manifold. The particularly interesting case is when N is a hypersurface (of
codimension one). In such a case, the hypersurface N can be characterized in terms of its
normal (in M) vector field N . Indeed:
1. A hypesurface N is called spacelike, if the normal Nx at each point x ∈ N is timelike.
2. A hypesurface N is called timelike, if the normal Nx at each point x ∈ N is spacelike.
3. A hypesurface N is called null, if the normal

Nx at each point x ∈ N is null. In this
case, Nx is tangential to Tx N and hence g is degenerate.

Tx N
The above are a consequence of the fact that the orthogonal complement of a timelike (resp.
spacelike) vector in a Lorentzian space only consists of spacelike (resp. timelike) vectors.
For more details about degenerate hyperplanes and hypersurfaces see Section 2.2.
27
Causality theory
Clearly, a Lorentzian manifold M is generally not a linear space and hence does not
coincide with the tangent planes at its points. For this reason, the definition of global causal
structures require some additional care.
Definitions
We have the following definitions:
• The causal future J + (S) of a set S is defined to be the set of all points which can be
connected with a point of S through a future-directed causal curve.
• The chronological future I + (S) of a set S is defined to be the set of all points which
can be connected with a point of S through a future-directed timelike curve2 .
• A set S is called achronal if there is no timelike curve which intersects S twice. Hence,
/ I + (q) and q ∈
for all p, q ∈ S we have p ∈ / I + (p).
• A set F is called future if for all x ∈ F we have I + (x) ⊂ F .
• Let ∂S = S ∩ M/S be the (topological) boundary of S in M. An achronal piece B1

of the boundary ∂S of S is called future if for all x ∈ B1 we have I + (x) ∩ S = ∅.
Similarly, an achronal piece B2 of the boundary ∂S of S is called past if for all x ∈ B2
we have I − (x) ∩ S = ∅.
Important properties
The definition of the exponential map expx at each point x ∈ M carries over in the case
of Lorentzian manifolds. Let Ux ⊂ M be a normal neighborhood around x where expx is
2
Note that I + (x) ⊂ M denotes the chronological future of x whereas Ix+ ⊂ Tx M denotes the set of all
timelike vectors in Tx M.
28
a diffeomorphism: expx : Vx ⊂ Tx M → Ux ⊂ M, where Vx = exp−1 x (Ux ). The following
proposition determines locally the chronological future I + (x) of a point x.
Proposition 2.1.1. Let (M, g) be a Lorentzian manifold and x ∈ M. Then,
I + (x) ∩ Ux = expx Ix+ ∩ Vx .

(2.1)
In particular, if y ∈ I + (x) ∩ Ux , then y can be connected with x through a timelike geodesic.
Furthermore, for any set S the chronological future (resp. past) I + (S) (resp. I − (S)) is open
in M.
Proof. First note that
expx Ix+ ∩ Vx ⊆ I + (x) ∩ Ux .

Indeed, if Y ∈ Ix+ ∩ Vx then expx (Y ) = αY (1), where αY is the unique affinely-parametrized

.
geodesic such that αY (0) = x and αY (0) = Y . The geodesic αY is timelike since Y is timelike
and the norm of the tangent to a geodesic is preserved along the geodesic:
. .
. .
∇α. Y g(αY , αY ) = 2g(∇α. Y αY , αY ) = 0.
We now prove the inverse inclusion. Suppose that it does not hold, that is to say suppose
that there is a point y ∈ I + (x) ∩ Ux and a timelike curve γ such that γ(0) = x and γ(1) = y
and such that exp−1x (y) ∈/ Ix+ ∩ Vx . Then the curve exp−1
x (α) initially enters the interior of
the null cone at x and subsequently exits it.
Recall the hyperboloids Hτ = {X : g(X, X) = −τ, τ > 0} ⊂ Tx M. These are spacelike
hypersurfaces in Tx M and foliate Ix+ . By a continuity argument we can easily deduce
that there exists a point where the curve exp−1 x (α) is tangent to a hyperboloid Hτ for
some τ > 0. Indeed, if exp−1 x (α) were always transversal to Hτ then the function f (s) =
−1 −1
g(expx (α(s)), expx (α(s))) would be decreasing; however, this contradicts the fact that for
some s0 the point α(s0 ) lies on the null cone and thus f (s0 ) = 0 (note that f (0) < 0).
The proposition now follows from the following lemma which is the analogue of Gauss
lemma in Riemannian geometry.
Lemma 2.1.1. (Gauss Lemma) Let x be a point in a Lorentzian manifold (M, g) and
Y ∈ Tx M ∩ Vx be a timelike vector. Let also αY be the unique affinely-parametrized geodesic
.
such that αY (0) = x and αY (0) = Y . Then, the timelike geodesic αY (t) is perpendicular to
the hypersurfaces expx (Hτ ) in M. In particular, this implies that the hypersurfaces expx (Hτ )
are spacelike.
The proof of the lemma is identical to the Riemanian case and thus is omitted.
Returning to the proof of the proposition, we have that the curve exp−1 x (α) cannot
be tangent to Hτ for some τ , because if it were then the timelike curve α would have to
be tangent to the spacelike hypersurface expx (Hτ ), which, in view of Gauss’ lemma, is a
contradiction.
29
The above allows us the locally “compute” the chronological futures. WeSwill use this
result to show that the chronological future of sets is open. Since I + (S) = x∈S I + (x) it
suffices to show that I + (x) is open for all points x ∈ M. Let y ∈ I + (x) and γ a curve
such that γ(0) = x and γ(1) = y. There exists a point p on γ close to y such that y ∈ Up ,
where Up is a normal neighborhood of p. Applying exp−1 p we can pass to Tp M and obtain an
−1 +
neighborhood Oq of expp (q) lying in Ip ⊂ Tp M. Then, by virtue of the previous results,
the open set expp (Oq ) lies in I + (p) and thus in I + (x).
The following proposition describes the relation between the seemingly opposite notions
of achronal and future sets.
Proposition 2.1.2. Let F be a future set in a Lorentzian manifold M. Then the topological
boundary ∂F is a closed achronal three-dimensional locally Lipschitz submanifold of M such
that ∂mani ∂F = ∅, where ∂mani denotes the boundary in the sense of manifolds.
Proof. The topological boundary of a set is by definition closed.
Let now x, y ∈ ∂F such that y ∈ I + (x). Then there exists an open neighborhood Ox of
x in M such that Ox ⊂ I − (y). Since x ∈ ∂F there exists a point p ∈ Ox such that p ∈ F
and y ∈ I + (p). There also exists an open set Oy of y such that Oy ⊂ I + (p). However, since
y ∈ ∂F , we have that Oy ∩ M/F 6= ∅, which contradicts the futureness of F .
The second part of the proposition deals with local notions and so it suffices to restrict
our attention to a normal neighborhood Ux of a point x ∈ ∂F . Let Vx ⊂ Tx M be such that
expx (Vx ) = Ux and consider inertial coordinates (t, x1 , x2 , x3 ) on Vx .
Consider the lines γ(t) = (t, c1 , c2 , c3 ), i.e. the integral curves of ∂t in Vx . These are all
timelike. Consider also their image expx (γ(t)). In general, since expx is not an isometry
in general, expx (γ)(t) might fail to be timelike; however, since dexpx is the identity when
resticted at the tangent plane of the origin of Tx M, by the stability of timelike vectors, we
can assume that Ux is sufficiently small so the curves expx (γ(t)) are timelike in Ux for all
curves γ. Hence these curves can intersect ∂F at most once. We will show that the curves
expx (γ(t)) intersect ∂F exactly once.
First note that I + (x) ⊂ F . Indeed, if q ∈ I + (x) then there exists a x0 ∈ F close to x
such that q ∈ I + (x0 ) and hence q ∈ F . Similarly, we can show that I − (x) ⊂ M/F . Consider
now the picture in Vx ⊂ Tx M. The curves γ intersect both Ix+ ∩ Vx and Ix− ∩ Vx . Passing
to Ux in M by applying expx and in view of the previous results and Proposition 2.1.1 we
have that the curves expx (γ) intersect F and M/F and hence intersect ∂F at least (in fact
exactly) once.
30
We will show that ∂F is a locally Lipschitz submanifold with no bounary (in the sense
of manifolds). It suffices to prove that exp−1 x (∂F ∩ Ux ) satisfies these properties. Since the
integral curves γ of ∂t intersect exp−1
x (∂F ∩ Ux ) exactly once, we have that exp−1
x (∂F ∩ Ux )
1 2 3
is the graph of the function t = t(x , x , x ) defined by

−1
expx (∂F ∩ Ux ) = 1 2 3 1 2 3 1 2 3
t(x , x , x ), x , x , x : (x , x , x ) ∈ {t = 0} ∩ Vx ∼
= R3 .
We will prove that the function t = t(x1 , x2 , x3 ) is Lipschitz (and thus continuous). Indeed,
a point q ∈ Tx M cannot belong (under expx ) to ∂F ∩ Ux if it belongs in a “modified” null
cone with vertex at a point p ∈ exp−1x (∂F ∩ Ux ). By continuity around x we obtain that for
a sufficiently small Ux (the size of which does not depend on F ) we have
t(x1 , x2 , x3 ) − t(x̃1 , x̃2 , x̃3 )
≤ 2.
k(x1 , x2 , x3 ) − (x̃1 , x̃2 , x̃3 )k
and hence the function t = t(x1 , x2 , x3 ) is indeed Lipschitz.
The manifold ∂F has no boundary since for any point x ∈ ∂F the neighborhood ∂F ∩ Ux
of x in ∂F is homeomorphic to R3 .
Remarks
1. We will be particularly interested in the case where the future set F is such that
F = J + (S) for a set S ⊂ M.
2. Proposition 2.1.2 holds in fact for more general achronal sets A and not just for the
boundary of future sets. The only additional assumption we need to impose is that
the achronal set A has no edge. This assumption implies that the curves expx (γ(t))
do intersect the set A. The precise definition of the edge is the following:
• The edge of an achronal set A consists of all points x ∈ A such that every
neighborhood U of x contains a timelike curve which intersects I + (x) ∩ U and
I − (x) ∩ U but does not intersect A.
31
3. Note that the weak regularity of Proposition 2.1.2 cannot be improved. Note, for
example, that the null cone (being the boundary of the future of a point in Minkowski)
has a differentiable singularity at the vertex.
2.2 Null Geometry

Null geometry is a new kind of geometry that naturally lives on hypersurfaces of Lorentzian
manifolds.
Definition and Applications

Let (M, g) be a (3 + 1)−dimensional Lorentzian manifold.
• A hypersurface H of M is called null, if at each point x ∈ H the normal L ∈ Tx M to

Tx H is a null vector, that is
g(L, L) = 0 and g(L, X) = 0 : ∀X ∈ Tx H.
Then, since dim(Tx H) = 3 we have Tx H = hLi⊥ and since L is null we have that
L ∈ Tx H and hence for each x ∈ M the hyperplane Tx H is degenerate.
The structure of these hypersurfaces is of paramount importance in relativity. The

following lemma has important applications for the geometry of a null hypersurface.
Lemma 2.2.1. Let x ∈ M and H be a null hupersurface on M with Tp H = hLi⊥ . Then, if

X ∈ Tx H then either X is null and X ∈ hLi or X is spacelike.
Proof. If X = (x0 , x1 , ..., xn ) and L = (l0 , l1 , ...ln ) with respect to an orthonormal basis, then
g(L, X) = 0 and g(L, L) = 0 can be written as x0 l0 = x1 l1 + ... + xn ln and l02 = l12 + ... + ln2 ,
respectively. By Cauchy–Schwarz inequality we obtain
(l12 + ... + ln2 )(x21 + ... + x2n ) ≥ (l1 x1 + ... + ln xn )2 = (x0 l0 )2 ,
which implies that g(X, X) ≥ 0 and we have equality if and only if X ∈ hLi.
32
Null Generators
Therefore, we have a distinguished line bundle on H, which is the null lines spanned by the
normal L to the hupersurface (recall that in this case the normal is also tangential to the
hypersufrace). We next investigate the integral curves of this line bundle. We consider the
covariant derivative ∇L L and take: If X ∈ T H, then
1
g(∇L L, X) = −g(L, ∇L X) = −g(L, ∇X L) − g(L, [L, X]) = − X g(L, L) = 0, (2.2)
2
since [L, X] ∈ T H. Hence the integral curves are geodesics. We arrive at the following
Proposition 2.2.1. A null hypersurface H is generated (ruled) by null geodesics whose

tangent is the normal to H.
We will refer to these null geodesics as the null generators of H. Note also that a natural
choice for the normal L to H is such that L satisfies the geodesic equation ∇L L = 0.
Corollary 2.2.1. Conformal transformations between Lorentzian manifolds map null hy-
persurfaces to null hypersurfaces and hence null geodesics to null geodesics.
Optical Functions
Equation (2.2) holds only if X is tangential to H. We will next present a method to derive
a similar formula for any vector X in M.
A null hypersurface H can be considered to be a non-critical level set of a smooth function
u, i.e. H = {u = c}, where c is a constant. In this case, the vector ∇u is normal to H and
hence null, and hence u must satisfy the so-called eikonal equation
g(∇u, ∇u) = g µν (∂µ u)(∂ν u) = 0. (2.3)
For simplicity, let us denote L = ∇u. If x ∈ H then for any X ∈ Tx M we have:

g(∇L L, X) = ∇L L [ (X) = ∇L L[ (X) = ∇L (du)(X)
1
= ∇2 u (L, X) = ∇2 u (X, L) = ∇X (du)(L) = g(∇X L, L) = ∇X g(L, L) ,

2
which indeed generalizes (2.2). The musical isomorphism [ corresponds to the standard
index-lowering. Hence, if L = ∇u then we obtain
1
∇L L = ∇ g(L, L) . (2.4)
2
A particularly important application of formula (2.4) is when the function u is an optical
function, namely:
• A differentiable function u : M → R is called optical if all level sets {u = c} are null

hypersurfaces. In this case, the eikonal equation is satisfied along all level sets of u.
If u is an optical function, and L = ∇u, then g(L, L) = 0 everywhere and hence the gradient
of this function also vanishes. Hence, the vector field L is a geodesic null normal to H since
∇L L = 0.
33
Remark: An example of a function u such that u = 0 is a null hypersurface but u is
not an optical function is the following:
u(t, x1 , x2 , x3 ) = −t2 + (x1 )2 + (x2 )2 + (x3 )2
on Minkowski spacetime (with the origin deleted).

For more about optical functions see Section 4.1.
Sections of null hypersurfaces
• A section S of a null hypersurface H is a two-dimensional submanifold of H which

intersects each null geonerator of H transversally.
Clearly, by virtue of Lemma 2.2.1 and the above transversality assumption, any section
S is a two-dimensional Riemannian manifold. Note that, since H is null, all null generators
intersect S orthogonally.
We are mainly interested in the case where S is topologically homeomorphic to the

two-dimensional sphere S2 .
Spacelike surfaces and their associated null normal geodesic congruences
Let us now start with a surface S, namely a two-dimensional Riemannian manifold

(homeomorphic to S2 ). For every point x ∈ S we have dim(Tx S) = 2 and g Tx S is positive
definite. Hence, by Sylvester’s law of inertia we have that the orthogonal complement Px =
(Tx S)⊥ of Tx S in Tx M is a two-dimensional Lorentzian plane. In other words, (Px , g Px ) is
isometric to the two-dimensional Minkowski spacetime.
34
Since all vectors orthogonal to S at x belong in Px the above discussion shows that there
are exactly two null lines orthogonal3 to S at x. Let Lx and Lx be two future-directed null
vectors in Px . Suppose that Lx projects to the exterior of S and Lx projects to the interior
of S. We will call Lx the outer null normal to S at x and Lx the inner null normal to S
at x. Although there is no natural normalization4 for Lx (and Lx ), we require the choice of
Lx (and Lx ) to depend differentiably on x so the resulting vectorfield L (and L) along S is
differentiable.
The vector bundle [
P= Px
x∈S
is called the normal bundle of S in M. Note that Px = hLx , Lx i.

Let now (L, L) be given future-directed null normal vector fields to S. For each x ∈ S
there is a unique affinely parametrized geodesic Gx with initial conditions (x, Lx ), that is to
.
say Gx (0) = x and Gx (0) = Lx . Similarly, there is a unique affinely parametrized geodesic
Gx with initial conditions (x, Lx ). We consider the sets in M formed by these geodesics:
[ [
C= Gx , C= Gx . (2.5)
x∈S x∈S
The null geodesics Gx (resp. Gx ) are called the null generators of C (resp. C). We
assume that C, C are smooth hypersurfaces (or that at least the part of C, C are smooth
hypersurfaces). Note that in general C, C will not be globally smooth hypersurfaces. For
example, the incoming null geodesic congruence of a standard sphere in Minkowski spacetime
forms a cone which is singular at the vertex. For more about the regularity of C, C see Part
II of these notes.
We have the following
Proposition 2.2.2. The sets C, C defined above are null hypersurfaces.
Before we prove Proposition 2.2.2 it is convenient to define affine foliations of C, C. From

now on we focus on the (smooth) hypersurface C.
The vector field L on S extends to a vector field L on C as the tangent field to the
affinely parametrized null generators of C. Then L satisfies
∇L L = 0
3
Clearly, the null lines span a null cone at x, however there are only two null lines normal to S.
4
For more about the normalization see the end of this section.
35
on C. Consider the flow Fτ of L which is a one-parameter family of diffeomorphisms of C.
We define the surfaces
Sτ = Fτ (S)
which foliate C: [
C= Sτ . (2.6)
τ ≥0
Clearly S0 = S and Sτ are sections of C. Note that τ is the affine parameter of L, namely
Lτ = 1 and τ = 0 on S. For this reason we call the foliation defined by (2.6) an affine
foliation of C. Understanding the geometric properties of the surfaces Sτ is of fundamental
importance in Lorentzian geometry.
Let y ∈ C. Suppose that y ∈ Gx , and in fact y = Gx (τ ). Then y ∈ Sτ . Let (E1 , E2 ) be

a basis of Tx S. Then (L, E1 , E2 ) is a basis for Tx C. We can propagate this basis along Gx
according to
[L, Ei ] = 0, i = 1, 2.
Then the vectors E1 , E2 are tangential to the sections Sτ and in fact we have:

(Ei )Gx (τ ) = dFτ (Ei )x , i = 1, 2.
The vector fields Ei , i = 1, 2 are called normal5 Jacobi fields along the null geodesic Gx . They
can be thought of as the infinitesimal displacement field of Gx through nearby null generators
whose feet on S span a curve tangential to Ei , i = 1, 2. Indeed, if γ(s) : (−, ) → S is a
.
curve on S such that γ(0) = x and γ(0) = Ei , then we have a mapping
Φ : (−, ) × [0, ∞) → C
given by
Φ(s, τ ) = Gγ(s) (τ ).
Then, if we fix τ , the curve α(s) = Gγ(s) (τ ) represents a displacement of Gx (τ ) along nearby
d

null generators. The infinitesimal displacement is given by Ei = ds Gγ(s) (τ )s=0 .
5
Note that g(L, Ei ) = 0, i = 1, 2.
36
The triple (L, E1 , E2 ) is a basis for Ty C, the linear map dFτ : Tx C → Ty C being
an isomorphism (see Section 5.2 for a discussion of the case where this map is not an
isomorphism).
Proof of Proposition 2.2.2. Let y = Gx (τ ) ∈ C and (L, E1 , E2 ) be a Lie propagated frame

along the null generator Gx . We will show that Ty C is a null hyperplane. If X ∈ {L, E1 , E2 },
then
1
∇L g(L, X) = g(∇L L, X) + g(L, ∇L X) = g(L, ∇X L) = ∇X g(L, L) = 0 (2.7)
2
along Gx , where we used the fact that [L, X] = 0 and that L is null on C. However, by
assumption L ⊥ hL, E1 , E2 i at x and so g(L, X) = 0 for τ = 0. In view of (2.7), we have
that g(L, X) = 0 along Gx and hence the hyperplane Ty C is indeed null.
Note that had we chosen an arbitrary null vector Lx (not necessarily normal to S) then
we would not have obtained a null hypersurface. We will see several examples of timelike
hypersurfaces spanned by null geodesics in the sequel.
We have the following definition:
• Given a surface S in M, the outgoing null hypersurface C is called theouter null

geodesic congruence normal to S. Similarly, the incoming null hypersurface C is called
theinner null geodesic congruence normal to S.
The importance of the null geodesic congruences will become apparent in Section 2.4.
Remarks:
1. As we mentioned above, there is no natural normalization for the null normal to S

vector fields L, L. However, we can impose on the pair (L, L) the normalization condition:
g(L, L) = −1.
If L transforms according to
L 7→ aL,
where a is an arbitrary positive differentiable function on S, then L transforms according to
L 7→ a−1 L.
37
Given a normalized null pair (L, L), we define the pair:
1 1
T = √ (L + L), N = √ (L − L).
2 2
One can easily see that for each x ∈ S, the pair (Tx , Nx ) forms a positively-oriented orthonor-
mal basis of Px . If (L0 , L0 ) is another normalized pair related to (L, L) via the positive dif-
ferentiable function α, then the pair (T 0 , N 0 ), obtained as above, gives another orthonormal
basis of Px for each x ∈ S. Note that the group SO(1, 1) of Lorentz boosts in two spacetime
dimensions (described in Section 1.3.2) connects such positively-oriented orthonormal bases.
If S is contained in a spacelike hypersurface Σ such that S is the boundary of a compact
region K in Σ, then a natural choice is to take T to be the unit future-directed timelike
normal to Σ at S and N the unit outward normal to S in Σ.
2. Similar results apply for the (future) null geodesic cone with vertex a point O ∈ M.
This is defined as the set of all (future) null geodesic which emanate at the point O.
2.3 Global Hyperbolicity

The causal structure of Lorentzian manifolds might in some cases exhibit unphysical be-
havior. For example, there are Lorentzian manifolds with closed timelike curves. This kind
of behavior is very pathological and so we want to impose conditions on the spacetimes in
order to exclude it.
• An achronal hypersurface H is a Cauchy hypersurface if every intextendible causal

curve intersects Σ exactly once.
• A spacetime (M, g) which possesses a Cauchy hypersurface is called globally hyperbolic.
Figure 2.1: The spacetime R1+1 minus a point is not globally hyperbolic since any achronal
hypersurface does not intersect one of the causal inextendible curves γ1 or γ2 .
The existence of a Cauchy hypersurface is a global causal property of the spacetime.

According to the following proposition, if a Cauchy hypersurface exists, then its topology is
“unique”.
Proposition 2.3.1. If a time-orientable spacetime (M, g) admits two Cauchy hypersurfaces

Σ1 , Σ2 then Σ1 is homeomorphic to Σ2 .
38
Proof. Since (M, g) is time-orientable, there exists a non-vanishing timelike vector field T .
The integral curves of T are timelike and hence intersect Σ1 and Σ2 exactly once.
The projection of Σ1 onto Σ2 is a continuous bijective map whose inverse is the projection
of Σ2 onto Σ1 .
Furthermore, the topology of a globally hyperbolic spacetime is completely determined

by that of a Cauchy hypersurface:
Proposition 2.3.2. If Σ is a Cauchy hypersurface for (M, g), then M is homeomorphic to

Σ × R. In particular, this implies that there exists a global ‘time’ function t : M → R such
that each level set Στ = {t = τ } is a (spacelike) Cauchy hypersurface and thus the vector
field ∇t is everywhere timelike. Furthermore, the hypersurfaces Στ foliate M.
Global hyperbolicity plays a role similar to that of completeness of Riemannian manifolds.
Proposition 2.3.3. Let (M, g) be a globally hyperbolic spacetime. Let also x, y ∈ M with
y ∈ I + (x). Then there exists a timelike geodesic γ which connects x, y and maximizes the
length function defined by the following formula
Z s 1
. . 2
L(γ) = − g(γ(t), γ(t)) dt.
0
A spacetime may not be globally hyperbolic; however there may exist subsets of such
spacetimes which are indeed globally hyperbolic.
• Let A be an achronal three-dimensional hypersurface of a spacetime (M, g). Then

the Cauchy development D(A) of A is the biggest globally hyperbolic subset of M
which admits A as a Cauchy hypersurface. In other words, D(A) consists of all points
x ∈ M such that every inextendible causal curve through x intersects A. The future
development is defined to be D+ (A) = J + (A) ∩ D(A), and similary we can define the
past development.
• The future Cauchy horizon of the development D+ (A) in M is defined to be the future
boundary of D+ (A) in M.
39
By virtue of Proposition 2.1.2, the Cauchy horizon is a three-dimensional Lipschitz subman-
ifold.
2.4 Causality II
The following two propositions are extremely important for our analysis. They provide
additional information regarding causal structures of globally hyperbolic spacetimes.
Recall that the topological boundary ∂S of set S in M is defined to be ∂S = S ∩ M/S.
Proposition 2.4.1. Let K be a compact subset of a globally hyperbolic spacetime (M, g).
Then J + (K) is closed in M and hence ∂J + (K) ⊂ J + (K).
Figure 2.2: If we remove the point p from R1+1 then the set J + (x) is not closed anymore.
Proposition 2.4.2. Let S be a compact orientable two-dimensional surface in a globally

hyperbolic spacetime (M, g) (for example let S be a surface homeomorphic to the sphere S2 ).
Let C and C denote the (future) outgoing and incoming null geodesic congruence normal to
S, respectively. Then, [
∂J + (S) ⊂ C C.
40
Sketch of proof of Proposition 2.4.2: In view of Proposition 2.4.1, if x ∈ ∂J + (S) then
there is a caucal curve γ which connects x and a point of S. Basically, this is the only point
in the proof where the global hyperbolicity is needed; contrast this with Figure 2.2. Clearly
γ cannot be timelike, since otherwise x would not lie in the boundary of the future of S but
in the interior instead. If γ is causal and not everywhere differentiable, then by a variation
of γ we can find a strictly timelike curve which connects x with some point in S. Hence γ
must be a null curve, and in fact by a variation argument again, it must be a null geodesic.
One more variation of γ shows that γ must be orthogonal to S.
We will further characterize the boundary ∂J + (S) in Section 5.3.
41
Chapter 3
Introduction to General Relativity
General relativity is a successful geometric theory of gravitation. In chapter 1 we saw

why special relativity provides the natural framework for electromagnetism and Maxwell’s
equations. In this chapter, we describe the main geometric features of spacetimes modeling
gravitational systems.
3.1 Equivalence Principle

It is not a priori obvious why the setting of Lorentzian geometry allows one to extend
special relativity and thus incorporate gravity as well. However, in 1907 Einstein suggested
the following principle (also known as Einstein’s happiest thought):
Equivalence principle
• One cannot discriminate (at least locally) between accelerating systems and systems
fixed in a gravitational field.
In other words, an observer who is fixed somewhere about the surface of the earth is accel-
erating. However, the observer cannot decide whether his acceleration is due to gravity or
another force that forces him to accelerate. This also implies that the gravitational mass is
equal to the inertial mass.
Since there is no ‘magic’ with gravity we have no way to distinguish gravity from other
forces. For this reason, in general relativity, gravity is not described as a force field; rather
we are forced to view gravity as an aspect of spacetime structure.
The equivalence principle may also be thought of as the following statement, known as
the geodesic hypothesis:
• Inertial freely falling observers in a gravitational field move on timelike geodesics and
photons move on null geodesics.
Note that freely falling observers are not accelerating. If one freely falls on earth then he
will feel strange because his state is rapidly changing from the previous accelerating state
of motion. Indeed, if a freely falling observer tries to lift his hand during the free fall then
he will feel no weight and thus no external forces on him. This means that his motion is
inertial and hence his worldline is a geodesic.
42
3.2 The Einstein Equations
The natural framework to develop general relativity is that of Lorentzian geometry. Indeed,
in Minkowski spacetime the geodesics are just lines (with respect to inertial systems) and
thus freely falling observers do not experience non-trivial dynamics. The only way to have
a spacetime with non-trivial geodesics is by considering non-flat metrics. Hence, we seek for
equations/laws that will determine the metric of the spacetime in the presence of matter
fields.
According to Newtonian physics, laws of physics should depend on the second order
derivative. Applied to our setting, this means that we need to impose conditions on the
curvature of the spacetime (M, g). These conditions will depend on the matter model
present M.
Einstein equations:
1
Ric(g) − Rsc (g) g = 8πT, (3.1)
2
where Ric, Rsc (g) denote the Ricci and scalar curvature, respective, and T denotes the
energy-momentum tensor of the present matter. Note that T is a (0, 2) symmetric divergence
free tensor field.
We will be mostly interesting in the case where no matter is present. Then T = 0 and
3.1 imply that the vacuum equations are:
Einstein-vacuum equations:
Ric(g) = 0. (3.2)
Minkowski spacetime is the trivial solution of the vacuum equations. As we shall see in the
sequel, most of the qualitative properties of (3.1) are already present in (3.2).
3.3 The Cauchy Problem

As with any physical theory, it is very important to confirm that general relativity is a
locally well posed theory. In particular, we are interested in showing that given appropriate
smooth initial data then there exists (at least locally in time) a unique smooth solution to
the Einstein equations.
First we need to determine the “correct” notion for the initial data. We expect to have
well-posedness only as long as (M, g) is globally hyperbolic. In this case, there exists a time
functions t, such that the level sets Hτ are spacelike Cauchy hypersurfaces and M = ∪t Ht .
Then, the vector field ∇t is normal to Hτ . We define the lapse function Φ by
1
Φ= p .
−g(∇t, ∇t)
Then,
T = −Φ2 · ∇t ⇒ T ⊥ Hτ and T t = 1
43
and
1
N= · T ⇒ N ⊥ Hτ and g(N, N ) = −1.
Φ
We can now consider a coordinate system (x1 , x2 , x3 ) on H0 . This can be propagated to
any Hτ by the flow of T , constructing a coordinate system (t, x1 , x2 , x3 ) of M. Then, the
metric with respect to this system takes the form
g = −Φ2 dt2 + ḡ,
where ḡ = ḡ(t) is the induced metric on Ht .

We look for initial data to prescribe on H0 . Since the equations Ric(g) = 0 are of second
order in g, we must prescribe the metric ḡ and also somehow the transversal derivatives ∂t ḡij
of ḡ on H0 . In the context, however, of the Cauchy problem, we want to a priori describe the
initial data set only in terms of the geometry of H0 , and hence we want to avoid prescribing
the t-derivatives. This can be done by introducing the second fundamental form k of H0 in
M. More generally, the second fundamental form of Ht is defined to be the following (0, 2)
tensor field on Ht :
k(X, Y ) = g(∇X N, Y ),
where X, Y ∈ Tp Ht . If X, Y ∈ {∂1 , ∂2 , ∂3 }, where ∂i = ∂xi , then we obtain1 :
1 ∂ḡij
kij = ,
2Φ ∂t
which is known as the first variational formula. This formula shows that the initial data set
∂ √
consists of the triplet (H0 , g, k). Note that the first variational formula implies ∂t ( ḡ) =
√
Φ · trk · ḡ.
The trace of the Codazzi equation and the double trace of the Gauss equation of the
embedding of Ht in (M, g) give us respectively:
divk − dk =Ric(N, ·),

Rsc + (trk)2 − |k|2 =Rsc + 2Ric(N, N ),
where the operators and quantities with the bar above refer to the induced geometry of
(Ht , ḡ).
In view of Ric(g) = 0, the right hand side of the Codazzi and Gauss equations vanishes.
Hence, (g, k) must be prescribed so they satisfy these constraint equations. This completes
the formulation of the initial value problem for the Einstein equations.
Let us now show that given such an initial data set, there exists a unique smooth local
solution (M, g). Since the differential structure of M is known, it suffices to show the
existence of g with respect to the coordinate system (t, x1 , x2 , x3 ) defined above. Without
loss of generality, we can take Φ = 1 on H0 . Then,
   
−1 0 0 0 ∂t gtt ∂t gtx1 ∂t gtx2 ∂t gtx3
0
 , ∂t g| = ∂t gtx1
  
g|H0 = 
0 H
. (3.3)
g ij  0  ∂t gtx2 2kij 
0 ∂t gtx3
1
We will not prove equations here. However, we prove all the analogous null structure equations in detail
in Chapter 4
44
Note that ∂t gtxα can be freely chosen initially for α = 0, 1, 2, 3.
We next briefly describe Choque–Bruhat’s argument for the existence of solutions to
Ric(g) = 0 under the above initial conditions. Let us suppose that t = x0 . We define
Γa = g µν Γaµν ,
where Γaµν are the Christoffel symbols of g with respect to the coordinate system xa , a =
0, 1, 2, 3. We also define
Γa = gab Γb , Hµν = Rµν − Sµν .
The crucial observation is that
P.P.(Hµν ) = g αβ ∂α ∂β gµν = P.P.(2g gµν ),
where P.P. =Principal Part. Given initial data as in (3.3), by standard theory of quasilinear
equations, there exists a unique smooth local solution g to
Hµν (g) = 0.
See Chapter 9 for an introduction to the energy method which can be used to address the
local well posedness of such equations.
However, we want g to satisfy Rµν (g) = 0 rather than Hµν = 0. In other words, we want
to make sure that the solution g to Hµν (g) = 0 somehow also satisfies Sµν (g) = 0 or, in fact,
Γa = 0.
If Hµν = 0 then
1 1 1
Rµν − Rgµν = Sµν − Sgµν ⇒ Div(Sµν − Sgµν ) = 0,
2 2 2
since for any metric g we have Div(Gµν ) = 0, where Gµν = Rµν − 12 Rgµν . The second crucial
observation is that
1
P.P. Div(Sµν − Sgµν ) = g αβ ∂α ∂β Γµ = P.P.(2g Γµ ).
2
Hence, for a metric g solving Hµν (g) = 0, Γµ can be shown to be identically zero if Γµ and
∂t Γµ are zero on H0 . It can be easily shown that
∂t gtxi = Γi , ∂t gtt = 2trk ⇒ Γµ |H0 = 0.
However, we have now completely exhausted the freely chosen initial data and hence we
would need a miracle so that ∂t Γµ |H0 = 0. The third crucial observation is that if Hµν = 0
then
∂t Γi |H0 = 2Gtxi , ∂t Γt |H0 = 2Gtt .
However, the constraint equations (satisfied by the initial data g, k) are precisely the equa-
tions Gtt = Gtxi = 0. Hence, Γµ do in fact vanish to first order on H0 and hence, the
previously constructed g is a solution to Ric(g) = 0.
Remarks
1. Note that the condition Γα = 0 is equivalent to 2g xα = 0, i.e. the coordinate functions
satisfy the wave equation. For this reason, these coordinates are called wave coordinates and
45
this choice of coordinates is called wave gauge. In the Riemannian case, this is also known
as harmonic gauge.
2. It is important to emphasize the role of global hyperbolicity to the uniqueness of the
Einstein equations, and in fact, of general hyperbolic equations. This is also known as finite
propagation speed. See Chapter 9 for a detailed discussion about this phenomenon.
3. The local well posednedss being understood (at least in the smooth class), one is inter-
ested in understanding global properties of the solutions to the Einstein equation. In general,
this is extremely difficult. A priori, however, Choquet–Bruhat and Geroch have shown that,
given an initial data set, there exists a uniqne maximal global hyperbolic spacetime (known
as the maximal development) which solves the Einstein equations. By maximality here we
mean the fact that any other globally hyperbolic solution can be isometrically embedded
in the maximal development (in fact, for this reason, one should think of the maximal
development as the “maximum” development).
3.4 Gravitational Redshift and Time Dilation

Based on qualitative properties of the postulates of relativity, (i.e. without precise knowledge
of the spacetimes (M, g), which satisfy the Einstein equations) we can easily deduce that
the proper time of observers depends on the gravitational potential of the region they live
in. In particular, the closer an observer is at the gravitational source the slower his proper
time passes.
Consider first the following thought experiment. Imagine two observers A, B in an accel-
→
erating spaceship such that their acceleration is congruent to the vector AB. Suppose that
A emits a photon towards B. Since photons travel with finite speed it will take some time
for this photon to reach B. Since the spaceship is accelerating, at the time of reception of
the photon by B, the speed of B is faster than the speed of A at the time of emission of
the photon (note, however, that the distance of A, B is constant at all times). Invoking the
Doppler effect we infer that B records the photon’s frequency on arrival as lower than it was
on departure from A. Hence, the frequency recorded by B is redshifted compared to the
frequency of the photon at the time of emission from A. Since frequency is a basic measure
of time, this implies that the proper time of B runs faster than the time of A as measured
by B. In other words, 1sec for A is measured as 0.9secs for B (and hence B finds the time
of A to be slower than his time), and conversely, 1sec for B is measured (by A) as 1.1secs
for A. The latter is due to the blueshift of photons emitted from B as received by A.
Consider now two observers A, B who are constantly located at distance, say, 1km and
100km, respectively, away from the surface of the earth. In view of the equivalence principle,
these two (non-freely falling) observers can be thought of as accelerating observers2 , both
→
being accelerated in the outward direction AB. Hence, the same principles apply as above
(there also exists a relativistic version of the Doppler effect). When light travels from A to B
it gets redshifted (as measured by B at the time of reception compared to the measurements
of A at the time of emission) and blueshifted when it travels from B to A. This phenomenon
is known as gravitational redshift. According to the above discussion, the time of B as
measured by A is faster than the proper time of A and the time of A as measured by
B is slower than the proper time of A. This phenomenon is known as gravitational time
2
The acceleration is needed in order to counteract gravity.
46
dilation. Confirmation of this effect has been achieved by comparison of atomic clocks flown
in airplanes with clocks on the ground.
3.5 Applications
We next describe some predictions of general relativity without involving equations.
1. Global Positioning System (GPS)

The Global Positioning System uses the effect of relativity in order to make correct
measurements. The GPS is a network of over 30 satellites orbiting at distance 20,000km
above the surface of the earth and moving at speed 14,000km/h, arranged such that at least
four satellites are visible from any point on earth. Each satellite constantly transmits signals
containing information on where the satellite is and what time the signal was transmitted.
A GPS device on earth (constantly) receives this signal and calculates the distance of the
device from the point where the signal was sent using the formula d = ct, where d is the
distance, c the speed of light and t the time that it took for the signal to reach the device.
Performing the same calculation using the signals sent from at least four visible satellites,
the device is able to locate its distance from four known points in the space, and hence it
can calculate its location. Note that although three satellites suffice, four or more satellites
are needed for more accurate results.
It is of paramount importance that the GPS devices calculate the time that signals had
to travel from the satellites to reach the device. However, in view of the gravitational time
dilation, the time on the satellites runs 45 microseconds/day faster compared to the time
on the surface of the earth. Furthermore, in view of the motion of the satellites compared
to the device, the effect special relativistic time dilation gives us that the time on satellites
runs 7 microseconds/day slower. When special and general relativistic effects are combined,
we deduce that satellite clocks run 38 microseconds/day faster. For this reason, the atomic
clocks are adjusted when constructed on earth so they run 38 microseconds/day slower
(and hence when on space they agree with the clocks on earth). If uncorrected, the total
measurements would be off by 11km/day!
2. Light bending
Null geodesics in a general Lorentzian manifold are curves with non trivial curvature. In
particular, according to relativity, a light ray that glazes the surface of the sun deflects from
47
its original path. This bending of starlight passing near our sun has been observed during
solar eclipses beginning with the 1919 expedition led by Eddington. This made relativity a
very popular theory of gravitation overnight.
Furthermore, the position of the astronomical objects may be such that a galaxy appears
to be in many positions at the same time:
In fact, in the case of black holes, relativity predicts the existence of lights which orbit
the black hole. The region where orbiting photons live is called photon sphere.
3. Mercury’s Perihelion Precession

Recall that the nearest point of the (elliptic) trajectory of a planet to the sun is called
perihelion. A number of effects in our solar system cause the perihelia of planets to precess
(rotate) around the sun. Mercury deviates from the precession predicted from these New-
tonian effects by 43 arcseconds (=43/3600 of a degree) per century. On the other hand,
relativity agrees closely with the observed amount of perihelion shift. Note that this effect
is small; it requires a little over twelve million orbits for a full excess turn.
48
Chapter 4
Null Structure Equations
4.1 The Double Null Foliation

Null Foliations and Optical Functions
Consider an open neighborhood of the spacetime M. This region can be foliated by outgoing
and incoming null hypersurfaces. This section is concerned with the formal definition of the
double null foliation.
Let S0 be an embedded 2-surface and C0 , C 0 be the null hypersurfaces spanned by
outgoing and incoming null geodesics normal to S0 .
Let Ω : S0 → R be a smooth function on S0 and L0 be a null vector field normal to S0

(and tangential to C0 ). Let now L0 be the null vector field normal to S0 and tangent to the
null geodesics of C 0 such that
g(L0 , L0 ) = −Ω−2 .
We extend L0 , L0 on C0 , C 0 , respectively, such that
∇L0 L0 = 0, ∇L0 L0 = 0.
We extend Ω to a function on the hypersurfaces C0 and C 0 and consider the vector fields
L = Ω2 L0 , L = Ω2 L0 (4.1)
Define the function u on C0 by
Lu = 1, with u = 0 on S0
49
and similarly define the function u on C 0 by
Lu = 1, with u = 0 on S0 .
Let Sτ be the embedded 2-surface on C0 such that u = τ , and similarly, let S τ be the
embedded 2-surface on C 0 such that u = τ .
We also define L0 on C0 such that L0 is null and normal to Sτ and g(L0 , L0 ) = −Ω−2 .
Consider the affinely parametrized null geodesics which emanate from the points on Sτ
with initial tangent vector L0 . These geodesics, whose tangent we denote by L0 , span (in view
of Proposition 2.2.2) null hypersurfaces which we will denote by C τ . Hence, C0 ∩ C τ = Sτ .
Similarly, we define L0 globally (and the hypersurfaces Cτ such that their normal is L0 ).
Extend the vector field L, L to global vector fields such that
L = Ω2 L0 , L = Ω2 L0 .
We will refer to Ω as the null lapse function. We also extend the functions u, u to global
functions such that
Lu = 0, L u = 0.
Therefore,
Cτ = {u = τ } , C τ = {u = τ } ,
and hence u, u are optical functions. The importance of the renormalized vector fields L, L
is manifest from the following
Proposition 4.1.1. The optical functions u, u satisfy the following relations
∇u = −L0 , ∇u = −L0
and
Lu = 1, Lu = 1.
50
Proof. Since u, u are optical functions, by the results of Section 2.2 we have that ∇u, ∇u
satisfy the geodesic equation. Since L0 , L0 satisfy the geodesic equation as well, it suffices
to show, for instance, that ∇u = −L0 on C0 . Expressing ∇u in terms of the null frame
(L0 , L0 , e1 , e2 ), where e1 , e2 is a local frame on the spheres Sτ , we obtain
0 0 0 0 0
(∇u)L = g L L · (L0 u) + g L L · (L0 u) = −Ω2 · Ω−2 = −1,
on C0 , and similarly, we obtain that the remaining components with respect to the above
frame are zero. This proves the first equation. For the second, it suffices to notice that
Lu = g(L, ∇u) = g(L, −L0 ) = −Ω2 · g(L0 , L0 ) = 1.
In this way, we foliate the spacetime with null hypersurfaces (being the level sets of u, u)
as depicted below
Gauge Freedom
The above analysis implies that the vector field L0 on S0 and the function Ω on C0 ∪ C 0 can
be freely chosen. This freedom reflects the freedom for the functions u0 = u0 (u), u0 = u0 (u).
Fixing Ω and L0 determines up to additive constants the (optical) functions u, u.
The Canonical Coordinate System

We saw above that we can construct at each point a null frame (L, L, e1 , e2 ) adapted to the
double null foliation (recall that e1 , e2 is a local frame for the spheres Su,u ). However, these
vector fields do not arise from a coordinate system.
Using the optical functions u, u we will introduce a coordinate system suitably adapted
to the corresponding double null foliation of the spacetime.
If p ∈ M then p ∈ Cu0 ∩ C u0 and hence u(p) = u0 , u(p) = u0 . We next prescribe angular
coordinates for the point p on the 2-surface Cu0 ∩ C u0 .
Let (θ1 , θ2 ) denote a coordinate system on a domain of S0 . Suppose that the null gener-
ator of C u0 through p intersects C0 at the point q and that the null generator of C0 through
the point q intersects the sphere S0 at the point with coordinates (θ1 , θ2 ). Then we as-
sign to p the angular coordinates (θ1 , θ2 ). Hence, the point p corresponds to the spacetime
coordinates (u0 , u0 , θ1 , θ2 ).
51
By construction we have everywhere:
∂
=L
∂u
and
∂ ∂
, ∈ T Su,u ,
∂θ1 ∂θ2
whereas
∂
= L : on C0 .
∂u
Note that the latter equation will not in general hold everywhere, as it is easily seen from
the picture below:
∂
From now on, for simplicity we denote ∂u = ∂u , and so on. In general we have:
∂u = L + bi ∂θi .
52
By virtue of the equations L = ∂u and [∂u , ∂u ] = [∂u , ∂θi ] = 0 we obtain:
∂bi
[L, L] = − ∂ i ∈ T Su,u ,
∂u θ
and therefore,
∂bi
= −dθi ([L, L]), and bi = 0 on C0 = {u = 0} . (4.2)
∂u
Hence, the S-tangent vector field b = bi ∂θi is the obstruction to the integrability of hL, Li =
⊥
T Su,u . In order to compute b it suffices to compute [L, L]. Since [L, L] = ∇L L − ∇L L,
it suffices to compute the connection coefficients; we will do this in the next section.
The metric g with respect to the canonical coordinates is given by
g = −2Ω2 dudu + (bi bj g/ ij )dvdu − 2(bi g/ ij )dθj du + g/ ij dθi dθj , (4.3)
where g/ denotes the induced metric on the 2-surfaces Su,u = Cu ∩ C u . We immediately

obtain
det(g) = −Ω2 · det(g/ ). (4.4)
Null Frames

From now on we denote Su,u = Cu ∩ C u . If {e1 , e2 } = ea a=1,2 is an arbitrary frame on the
spheres S(u,u) then we, in fact, have the following null frames:
• Geodesic frame: (e1 , e2 , L0 , L0 ),
• Equivariant frame: (e1 , e2 , L, L),
• Normalized frame: (e1 , e2 , e3 , e4 ),
where
e3 = ΩL0 , e4 = ΩL0 .
We symmetrically rescaled L0 , L0 above so that the equations take a symmetric (dual) form
in the 3 and 4 directions.
Note that e3 , e4 satisfy the normalization property
g(e3 , e4 ) = −1. (4.5)
53
4.2 Connection Coefficients
We consider the normalized frame (e1 , e2 , e3 , e4 ) defined above. We define the connection
coefficients with respect to this frame to be the smooth functions Γλµν such that
∇eµ eν = Γλµν eλ , λ, µ, ν ∈ {1, 2, 3, 4}
Here ∇ denotes the connection of the spacetime metric g. We are mainly interested in
the case where at least one of the indices λ, µ, ν is either 3 or 4 (otherwise, we obtain the
Christoffel symbols with respect to the induced metric g/ ). These coefficients are completely
determined by the following components:
The components χ, χ, η, η, ω, ω, ζ:
χAB = g(∇A e4 , eB ), χAB = g(∇A e3 , eB ),

ηA = g(∇3 e4 , eA ), η A = g(∇4 e3 , eA ),
(4.6)
ω = −g(∇4 e4 , e3 ), ω = −g(∇3 e3 , e4 ),
ζA = g(∇A e4 , e3 )
where A, B ∈ {1, 2} and ∇µ = ∇eµ . Note that ζ = −ζ. The covariant tensor fields χ, χ, η, η, ζ
are only defined on Tx Su,v . We can naturally extend these to tensor fields to be
defined on Tx M by simply letting their value to be zero if they act on e3 or e4 .
Such tensor fields will in general be called S-tensor fields. Note that a vector
field is an S-vector field if it is tangent to the spheres Su,u . The importance the
S-tensors originates from the fact that one is interested in understanding the embedding of
Su,u in Cu and C u .
The connection coefficients Γ can be recovered by the following relations:
∇A e B = ∇
/ A eB +χAB e3 + χAB e4 ,
∇3 eA = ∇
/ 3 eA + ηA e3 , ∇4 eA = ∇
/ 4 eA + η A e4 ,
∇ A e 3 = χA ]B
eB + ζA e3 , ∇A e4 = χA ]B eB − ζA e4 , (4.7)
∇3 e4 = η ]A eA − ωe4 , ∇4 e3 = η ]A eA − ωe3 ,
∇3 e3 = ωe3 , ∇4 e4 = ωe4 ,
Remarks:
1. Recall, that the second fundamental form of a manifold S embedded in a manifold
M is defined to be the symmetric (0, 2) tensor field II such that for each x ∈ S we have
IIx : Tx S × Tx S → (Tx S)⊥ ,
where the ⊥ is defined via the decomposition Tx M = Tx S ⊕ (Tx S)⊥ . Specifically, if X, Y ∈

Tx S then ⊥
IIx (X, Y ) = ∇X Y
and hence
∇X Y = ∇
/ X Y + II(X, Y ).
54
Here ∇
/ denotes the induced connection on S (which is taken by projecting the spacetime
connection ∇ on Tx S).
The S-tensor fields χ, χ give us the projections of IIAB on e3 and e4 , respectively. Indeed
II(X, Y ) = χ(X, Y )e3 + χ(X, Y )e4 .
For this reason we will refer to χ, χ as the null second fundamental forms of Su,u with
respect to the null hypersurfaces Cu , C u , respectively. One can easily verify that χ and χ
are symmetric (0,2) S-tensor fields. Indeed, a simple calculation shows that if X, Y are
S-tangent vector fields then
χ(X, Y ) − χ(Y, X) = g(e4 , [X, Y ]),

χ(X, Y ) − χ(Y, X) = g(e3 , [X, Y ]).
Hence, χ, χ are symmetric if and only if [X, Y ] ⊥ e3 and [X, Y ] ⊥ e4 and thus if and only if
he3 , e4 i⊥ 3 [X, Y ] ∈ T Su,u . The symmetry of χ, χ is thus equivalent to the integrability of
the orthogonal complement he3 , e4 i⊥ .
Furthermore, we can decompose χ and χ into their trace and traceless parts by
1 1
χ = χ̂ + (trχ)g/ , χ = χ̂ + (trχ)g/ . (4.8)
2 2
The trace of the S-tensor fields χ, χ (and more general S-tensor fields) is taken with respect
to the induced metric g/ . The trace trχ is known as the expansion and the component χ̂ is
called the shear of Su,u with respect to Cu .
2. Note also that ω = ∇4 (log Ω) = and ω = ∇3 (log Ω).
3. Let X be a vector tangential to a given sphere Su,u at a point x. Then, if we extend
X along the null generator γ of Cu passing through x according to the Jacobi equation
[L, X] = [Ωe4 , X] = 0, then we obtain an S-tangent vector field along γ. Note that in this
case we obtain
∇4 X = ∇X e4 + ∇X log Ω e4 .
On the other hand, if we simply extend X such that [e4 , X] = 0 then, although X will
be tangential to Cu , X will not be tangential to the sections Su,u of Cu . This is because
the sections the optical functions u, u are the affine parameters of the vector fields L, L,
respectively.
4. The S 1-form ζ is known as the torsion. If d
/ denotes the exterior derivative on Su,u
then the S 1-forms η, η are related to ζ via
η =ζ +d
/ (log Ω), η = −ζ + d
/ log Ω.
Proof. Let X be S-tangential and extend along the null generator of Cu according to the
Jacobi equation, then
η(x) = g(∇4 e3 , X) = −g(e3 , ∇4 X) = −g(e3 , ∇X e4 )+∇X (log Ω) = −ζ(X)+∇X log Ω. (4.9)
We similarly show the analogous relation for η.
55
The above imply also that
1 1
ζ = (η − η), d
/ log Ω = (η + η).
2 2
The 1-forms η, η can be thought of as the torsion of the null hypersurfaces with respect to
the geodesic vector fields. Indeed, the previous relations imply
ηA = Ω2 g(∇A L0 , L0 ).
5. We have

[L, L] = ∇L L − ∇L L = Ω · ∇4 (Ωe3 ) − ∇3 (Ωe4 )

= Ω2 · ∇4 e3 − ∇3 e4 + (∇4 log Ω)e3 − (∇3 log Ω)e4

= Ω2 · (η ]A − η ]A )eA − ωe3 + ωe4 + (∇4 log Ω)e3 − (∇3 log Ω)e4

= Ω2 · η ] − η ]
= −2Ω2 ζ ] .
The above completely determines the vector field b defined in (4.2). In particular, it shows
i
that ∂b 2 ] i
∂u = 2Ω (ζ ) and hence the torsion ζ is the obstruction to the integrability of the
timelike planes he3 , e4 i orthogonal to the spheres Su,u .
6. It is easy to show that for the standard spheres of radius r in the Minkowski spacetime
we have
2 2
χ̂ = χ̂ = 0, trχ = , trχ = − . (4.10)
r r
Indeed, using the double null coordinate system we easily obtain, for example,
1
χ(∂θ , ∂θ ) = χ(∂φ , ∂φ ) = , χ(∂θ , ∂φ ) = 0.
r
Note that trχ → −∞ as r → 0 (that is to say, the incoming expansion diverges we approach
the vertex). For a generalization of this fact, see Section 5.2.
4.3 Curvature Components

We next decompose the Riemann curvature R in terms of the normalized null frame. First,
we define the following components, which contain at most two S-tangential components
(and hence at least 2 null components):
αAB = RA4B4 , αAB = RA3B3 ,

βA = RA434 , β A = RA334 ,
1
ρ = R3434 , σ = / AB RAB34 .
2
Note that R(·, ·, e3 , e4 ), when restricted on Tx Su,u , is an antisymmetric form and hence
collinear to the volume form / on Su,u . Furthermore, if
(∗R)3434 = ∗ρ,
56
then ∗ρ = 2σ. Here, the dual ∗R of the Riemann curvature is defined to be the (0,4) tensor:
(∗R)αβγδ = µναβ Rµν γδ .
Clearly, the (0,2) S-tensor fields α, α are symmetric.

Proposition 4.3.1. Suppose that Ric = 0. Then,
tra = 0, tra = 0,
where the trace (of the S-tensor fields) is taken with respect to g/ , and the remaining com-
ponents (with at least three spherical components or R(·, e3 , ·, e4 )) are given by the following
expressions:
RA3BC = g/ AB β C − g/ AC β B ,
RA4BC = −g/ AB βC + g/ AC βB ,
1 1
RA3B4 = σ/ AB − ρg/ AB ,
2 2
RABCD = 2ρ(g/ AD g/ BC − g/ AC g/ BD ).
Remark 4.3.1. It is important to recall the basic properties of the volume form. If is the
(spacetime) volume form with respect to g, and / is the volume form on the spheres Su,u
with respect to the induced metric g/ , then
1234 = (e1 , e2 , e3 , e4 ) = 1 and AB34 = / AB .
We note that, when working with / it is very convenient to consider a (local) orthonormal
S-frame e1 , e2 on the spheres Su,u .
Hodge dual relative to Su,u
• The (left) dual ∗ξ of an S 1-form ξ is defined to be the following S 1-form
∗ξ = / (·, ξ ] ),
or, using the frame:

(∗ξ)A = / AB ξ ]B .
Note also that
∗ ∗ ξ = −ξ.
• The dual ∗θ of an (0,2) S-tensor field θ is defined to be the following (0,2) S−tensor
field
∗θ = / (·, θ] (·)),
or, using the frame:
(∗θ)AB = / AC θ]CB .
Note that ∗θ is symmetric if θ is traceless and ∗θ is traceless if θ is symmetric.
57
Proof of Proposition 4.3.1. We have Ricµν = (g −1 )κλ Rκµλν . The inverse metric g −1 can be
written in terms of the normalized frame as follows:
g −1 = −e3 ⊗ e4 − e4 ⊗ e3 + g/ −1 .
Hence,
Ricµν = −R3µ4ν − R4µ3ν + (g/ −1 )CD RCµDν .
1. The equations Ric33 = Ric44 = 0

Note that R44 = (g −1 )CD RC4D4 = trα and similarly, R33 = trα.
2. The equations Ric3A = Ric4A = 0
We have R4B = −R344B + (g/ −1 )CA RC4AB and therefore,
(g/ −1 )CA RC4AB = −βB .
Observe now that R(eC , e4 , ·, ·) is an (0,2) antisymmetric S-form on Tx Su,u and hence it
is proportional to the induced volume form / , hence:
RC4AB = ξC / AB
where ξ is an S 1-form. Then,
(g/ −1 )CA RC4AB = ξC / ]CB = −(∗ξ)B .
Therefore,
∗ξ = β =⇒ ξ = −(∗β)
and
RC4AB = −(∗β)C / AB = −g/ CA βB + g/ CB βA .
3. The equations RicAB = Ric34 = 0

A reasoning similar as above implies that there exists a function f such that
RABCD = f / AB / CD .
Since RAB = 0 we obtain
R3A4B + R4A3B = (g/ −1 )CD RCADB = f g/ AB (4.11)
and therefore, taking the trace (and using that trg/ = g/ AB g/ AB = 2) we obtain
1
f = (g/ −1 )AB (R3A4B + R4A3B ) = (g/ −1 )AB RA3B4 .
2
However, equation Ric34 = 0 gives us
(g/ −1 )AB RA3B4 = R4334 = −ρ (4.12)
and hence f = −ρ, which in turn determines RABCD . Finally, we compute the (0,2) S-tensor
field RA3B4 . For we consider its symmetric and antisymmetric parts:
RA3B4 + RB3A4 = −ρg/ AB ,
RA3B4 − RB3A4 = RAB34 = σ/ AB ,
where, in the last line above, we used the first Bianchi identity.
58
4.4 The Algebra Calculus of S-Tensor Fields
Consider a spacetime (M, g) and assume that it admits a double null foliation of spheres
Su,u . Let g/ be the induced metric on Su,u . Let / denote the induced volume form on Su,u .
Definition of S-tensor fields:
A 1-form ξ of M is an S 1-form if ξ(e3 ) = ξ(e4 ) = 0. In other words, xi is specified by a
smooth assignment of 1-form on Su,u for each Su,u . Similarly, a p-covariant tensor θ on M
is an S-tensor if θ gives the value zero whenever it acts on either e3 or e4 . A vector field X
is an S-tangent vector field (or S-vector field) if it is tangent to Su,u .
Musical isomorphisms
The (positive definite) metric g/ induces canonical, known as musical, isomorphisms
between the tangent bundle T Su,u and the cotangent bundle T ∗ Su,u . Specifically, given a
S-vector field X we define the S 1-form X[ such that X[ (Y ) = g/ (X, Y ), for all S-vector fields
Y . Similarly, given an S 1-form ω we define the S-vector field ω ] such that g/ (ω ] , Y ) = ω(Y )
for all S-vector fields Y .
The isomorphisms [, ] can be extended to more general S-tensor fields. Let, for example,
T : T Su,u → T Su,u be a (1, 1) tensor field. Then T[ is a (0, 2) tensor field, given by
T[ (X, Y ) = (T (X))[ (Y ) = g(T (X), Y ).
More generally, if T ∈ T Su,u ⊗n T ∗ Su,u is a tensor field of type (1, n) then T[ ∈ ⊗n+1 T ∗ Su,u
is a tensor field of type (0, n + 1) given by
T[ (x1 , x2 , ..., Xn , Y ) = (T (X1 , X2 , ..., Xn ))[ (Y ) = g(T (X1 , X2 , ..., Xn ), Y ).
On the other hand, if T ∈ ⊗2 T ∗ Su,u is of type (0, 2), then T ] ∈ T Su,u ⊗T ∗ Su,u = End(T Su,u )
is of type (1, 1) such that if X ∈ T Su,u then
T ] (X) = (T (X, ·))] ∈ T Su,u .
Note that one could also define
T ] (X) = (T (·, X))] ∈ T Su,u ,
and these two definitions coincide if and only if T is symmetric. Otherwise one needs to
prescribe explicitly how ] acts, as above.
Trace
The trace of a (1,1) S-tensor field T is defined to be the contraction of the matrix of T
in an arbitrary basis. If Ei is an S orthonormal frame then
n
X
trT = g/ (T (Ei ), Ei ).
i=1
Given now a (0, 2) tensor field T , we define
trT = trT ] .
59
In view of the fact that T ] depends on the metric g/ , the above trace is also called metric
trace. For (0, 2) S-tensors one does not need to prescribe how ] acts since
] ]
tr X 7→ T (X, ·) = tr X 7→ T (·, X) .
This can also be seen as the full contraction of the tensor product of the inverse metric
−1
g/ and T , i.e. trT = (g/ −1 )AB TAB .
For higher order S-tensor fields, the trace is not uniquely defined and so one needs to
prescribe how ] acts (and hence which indices to raise and contract). We usually consider
tensor fields T (X, Y, Z, W ) of type at most (0, 4). In that case, the trace in, say, X, Z
variables is
n
X

trX,Z T (Y, W ) = T (Ei , Y, Ei , W ).
i=1
Operations
Let θ, φ be (0,2) symmetric S-tensor fields. We define the (0,2) S-tensor field θ × φ such
as
θ × φ = g/ (θ] (·), φ] (·)), i.e. (θ × φ)AB = g/ CD θA]C φB]D = θAC φB]C .
Note that (θ × φ)(X, Y ) = (φ × θ)(Y, X) := (φ ^ × θ)(X, Y ), where ˜ denotes transposition.

For such S-tensors we also define the functions (θ, φ) and θ ∧ φ to be
(θ, φ) = tr(θ × φ) = θAB φ]]AB

1
θ ∧ φ = / AB (θ × φ − φ × θ)AB .
2
For example, g/ × θ = θ and hence (g/ , g/ ) = tr(g/ × g/ ) = trg/ = 2.
Differential operators
The exterior derivative on Su,u is denoted by d/ . If ξ is an S 1-form then we define the

curlξ to be the following function
curl ξ = ∗d
/ ξ,
where ∗ denotes the Hodge dual.
If T is an arbitrary S-tensor field, then we denote the induced gradient of T on Su,u by
∇
/ T and we extend ∇ / T to be an S-tensor field. Here ∇/ should be viewed as the connection
on Su,u associated to g/ .
The divergence of an S-vector field T is given by

div
/ T = tr(∇ / T ) = tr X 7→ ∇
/ XT ,
whereas the divergence of an S 1-form ξ is given by

]
div / ξ) = tr X 7→ ∇
/ ξ = tr(∇ / Xξ ,
60
More generally, if T is of type (0, n) then div/ T is of type (0, n − 1):
]
/ T (Y1 , ..., Yn−1 ) = tr X 7→ (∇
div / X T )(·, Y1 , ..., Yn−1 )
n
X
= ∇
/ Ei T (Ei , Y1 , ...Yn ),
i=1
where Ei , i = 1, ..., n form an orthonormal frame.

Let T be p-covariant S-tensor field. We define the projected covariant derivative ∇
/ 3 T to
be the S-tensor field with the property
(∇
/ 3 T )(X1 , X2 , . . . , Xp ) = (∇3 T )(X1 , X2 , . . . , Xp ),
for all S-vector fields X. Similarly we define ∇

/ 4 T . Note that ∇
/ 3 g/ = ∇
/ 4 g/ = 0 and hence
∇
/ 3 (trT ) = tr(∇
/ 3 T ), ∇
/ 4 (trT ) = tr(∇
/ 4 T ).
If ξ is an S 1-form and X is an S-vector field then

]
/ 3 (ξ ] ) = ∇3 ξ ,
∇ ∇
/ 3 (X[ ) = ∇/ 3X .
[
Let T be p-covariant S-tensor field. We define the projected lie derivative L

/3 T to be the
S-tensor field with the property
(L
/3 T )(X1 , X2 , . . . , Xp ) = (L3 T )(X1 , X2 , . . . , Xp ),
for all S-vector fields X. Similarly we define L /4 T . The importance of L

/4 lies on the fact
that, unlike ∇
/ 4 , it depends only on the differential structure of Cu .
In view of (4.7), and since LX Y = ∇X Y − ∇Y X, we obtain:
L / 4 ξ + χ] · ξ, if ξ is an S 1-form,
/4 ξ =∇
L / 4 θ + χ × θ + θ × χ, if θ is an S (0,2) tensor field,
/4 θ =∇
Finally, if ξ is an S 1-form we define the symmetrized traceless covariant derivative ∇ ˆ

/ ⊗ξ
by the following
∇ ˆ =∇
/ ⊗ξ /ξ + ∇ /fξ − (div [
/ ξ)g/ = L
/ξ] g/ ,
where L [
/ξ] g/ denotes the traceless part of L
/ξ] g/ := Lξ] g/ on Su,u . It is easy to verify that
1 ˆ
−2∇ / ⊗ is the adjoint of the restriction of div
/ on symmetric and traceless (0,2) S-tensor
fields.
4.5 Null Structure Equations

The ultimate goal is to understand the geometry g/ of the spheres Su,u and also their em-
bedding in Cu , C v , and more generally, in M. For this reason, in this section, we will derive
equations for the connection coefficients Γ defined in Section 4.2. Specifically, we will derive
propagation equations along the null hypersurfaces Cu , C u and also elliptic equations on
Su,u .
61
The propagation equations are of the following general form
∇
/ 4Γ = R + Γ · Γ + D
/ Γ,
∇
/ 3Γ = R + Γ · Γ + D
/ Γ,
/ ∈ {d
where D /,∇ / }.
/ , div
We will work with the normalized null frame (e1 , e2 , e3 , e4 ).
The null structure equations
The first variational formulas
L
/4 g/ =2χ,
(4.13)
L
/3 g/ =2χ.
The second variational formulas

∇
/ 4 χ = − χ × χ − α + ωχ,
(4.14)
∇
/ 3 χ = − χ × χ − α + ωχ.
The transversal propagation equations for χ, χ
1 1f 1
∇
/ 4χ = ∇/η + ∇
/η + η ⊗ η − (χ × χ + χ × χ) − ωχ + ρg/ ,
2 2 2 (4.15)
1 1f 1
∇
/ 3χ = ∇/η + ∇
/η + η ⊗ η − (χ × χ + χ × χ) − ωχ + ρg/ .
2 2 2
The equations for the torsions η, η
/ 4 η = −χ] · (η − η) − β,
∇ / 3 η = −χ] · (η − η) + β,
∇ (4.16)
/ 3 η = −2χ · η ] − β + 2d
∇ / ω + (η + η)ω. (4.17)
curl η = − curl ζ = − curl η = χ ∧ χ − σ. (4.18)

The propagation equations for ω, ω
1 h i
/ 4 ω = − ωω + Ω2 · 2(η, η) − |η|2 − ρ ,
∇
2 (4.19)
1 h i
/ 3 ω = − ωω + Ω2 · 2(η, η) − |η|2 − ρ .
∇
2
The Gauss equation
K = (χ, χ) − trχtrχ − ρ (4.20)
The Codazzi equations
/ trχ + χ] · ζ − (trχ) · ζ = −β,

/ χ−d
div
(4.21)
/ trχ − χ] · ζ + (trχ) · ζ = β.
/ χ−d
div
Remarks:
62
In view of the decomposition χ = 12 (trχ)g/ + χ̂, the second variational formulas imply:
/ 4 trχ = − |χ|2 + ωtrχ,

∇
(4.22)
/ 3 trχ = − |χ|2 + ωtrχ.
∇
and
∇
/ 4 χ̂ =ω χ̂ − (trχ)χ̂ − α,
(4.23)
∇
/ 3 χ̂ =ωχ̂ − (trχ)χ̂ − α,
where for the former equations we used that [∇

/ 3 , tr] = [∇
/ 4 , tr] = 0 and for latter ones we
used that
1 1
|χ|2 = (trχ)2 g/ + |χ̂|2 and χ̂ × χ̂ = |χ̂|2 g/ , (4.24)
2 2
since χ̂ is symmetric and traceless (0,2) S-tensor field. The equations (4.22) are also known
as Raychaudhuri equations.
Note also that
1
(χ, χ) = trχtrχ + (χ̂, χ̂).
2
Therefore, if
ρ̌ = −ρ + (χ̂, χ̂). (4.25)
then the Gauss equation can be written as
1
K = ρ̌ − trχtrχ.
2
Note that the product trχtrχ does not depend on the normalization of e3 , e4 , i.e. it is
invariant under the transformations e3 7→ ae3 , e4 7→ a1 e4 , with a > 0. The mass aspect
functions µ, µ are defined by
1
µ =K + trχtrχ − div / ζ,
2 (4.26)
1
µ =K + trχtrχ + div / ζ.
2
q
R Au,u
If Au,u = Su,u dg/ is the area of Su,u and ru,u = 4π is the ‘radius’ of Su,u , then we define
the Hawking mass m of Su,u by
Z
r 1
m= 1+ trχtrχ . (4.27)
2 8π S
The Gauss–Bonnet theorem on Su,u implies
Z Z
2m 1 1
= µ= µ.
r 4π S 4π S
Geometric structure equations

We first derive differential equations which rely only on the geometry of the double null
foliation, without assuming the Einstein equations.
63
We will show these relations only for the 4-direction. The proof is identical for the
3-direction.
First variation formula:
Let X, Y ∈ Tp Su,u and extend them along the null generator of Cu passing through p
so they satisfy the Jacobi equation [L, X] = [L, Y ] = 0. Then, [e4 , X] = (∇X log Ω)e4 and
[e4 , Y ] = (∇Y log Ω)e4 . Then,

(L/4 g/ )(X, Y ) = (L4 g/ )(X, Y ) = ∇4 g/ (X, Y ) − g/ ([e4 , X], Y ) − g/ (X, [e4 , Y ])
= g/ (∇4 X, Y ) + g/ (X, ∇4 X) = g/ (∇X e4 , Y ) + g/ (X, ∇Y e4 ) = 2χ(X, Y ).
Since this is true for all X, Y ∈ Tp Su,u we obtain the result.
Second variation formula:
so they satisfy the Jacobi equation [L, X] = [L, Y ] = 0. Then, the commutators [e4 , X] =
(∇X log Ω)e4 and [e4 , Y ] = (∇Y log Ω)e4 are proportional to e4 . Recall also that ∇X e4 =
χ] (X) − ζ(X)e4 . Then,

/ 4 χ)(X, Y ) = ∇4 χ(X, Y ) − χ(∇4 X, Y ) − χ(X, ∇4 Y )
(∇

= ∇4 g(∇X e4 , Y ) − χ(∇X e4 , Y ) − χ(X, ∇Y e4 )
= g(∇4 ∇X e4 , Y ) + g(∇X e4 , ∇4 Y ) − χ(χ] (X), Y ) − χ(X, χ] (Y ))

= g R(e4 , X)e4 + ∇X ∇4 e4 + ∇[e4 ,X] e4 , Y − (χ × χ)(X, Y )
= R(Y, e4 , e4 , X) + ωg(∇X e4 , Y ) − (χ × χ)(X, Y )

= − α − χ × χ + ωχ (X, Y ).
The torsion equation:

Let X ∈ Tp Su,u . Then

/ 4 η)(X) = (∇4 η)(X) = ∇4 (η(X)) − η(∇4 X) = ∇4 g(∇3 e4 , X) − g(∇3 e4 , ∇4 X)
(∇
= g(∇4 ∇3 e4 , X) + g(∇3 e4 , ∇4 X) − g(∇3 e4 , ∇4 X)

= g R(e4 , e3 )e4 + ∇3 ∇4 e4 + ∇[e4 ,e3 ] e4 , X

= R(X, e4 , e4 , e3 ) + ωg(∇3 e4 , X) + g ∇(η] −η] −ωe3 +ωe4 ) e4 , X

= −β(X) + ωη(X) + g ∇(η] −η] ) e4 , X − ωη(X)

= −β + χ · η ] − η ] (X).
Furthermore, for all scalar functions f we have the following commutation formula:
1
[∇ / ]f = (η + η)∇3 f − χ · (∇
/ 3, d / f)
2
Proof. For any scalar function f and X ∈ Tp Su,u we have,
∇ / f )(X) = ∇3 (d
/ 3 (d / f )(X) = ∇3 (Xf ) − (∇
/ 3 X)f
= X∇3 f + [∇3 , X]f − (∇
/ 3 X)f = X∇3 f + (∇3 X − ∇X e3 )f − (∇
/ 3 X)f
The result follows from the formulas (4.7).
64
Since η = −η + 2d
/ log Ω we obtain
∇
/ 3 η = −∇
/ 3 η + 2∇
/ 3 (d
/ log Ω)
= χ · (η ] − η ] ) − β + 2∇
/ 3 (d
/ log Ω)
= χ · (η ] − η ] − 2∇
/ log Ω) − β + 2d
/ ω + (η + η)ω
= −2χ · η ] − β + 2d
/ ω + (η + η)ω.
The propagation equation for ω:

Recall the equivariant pair (L, L) = (Ωe4 , Ωe3 ). The following relations can be easily
proved:
1
Ωω = − Ω−2 g(∇L L, L),
2
g ∇L L, ∇L L = −4Ω4 ωω,

g ∇[L,L] L, L − 2Ω2 g ∇ζ ] L, L = Ω4 · η − η, η ,

g(∇L L, L) = 0, since ∇L L is S-tangent,

g ∇L ∇L L, L = −Ω4 · (η, η),

g(R(L, L)L, L) = R(L, L, L, L) = Ω4 R4343 = Ω4 ρ.
Then,
1 2 LΩ 1
g ∇L , L, L − Ω−2 · g(∇L L, ∇L L) + g ∇L ∇L L, L

∇
/ 4 (Ωω) = 2
2Ω Ω 2
2 2 1 −2
= −2Ω ωω + 2Ω ωω − Ω g ∇L ∇L L + R(L, L)L + ∇[L,L] L, L
2
1 2
= Ω · (η, η) − ρ − (η − η, η)
2
1
= Ω2 · 2(η, η) − |η|2 − ρ .

2
The transversal propagation equation for χ and the curl equation for η can be obtained
as follows:
so they satisfy the Jacobi equation [L, X] = [L, Y ] = 0. Then, [e4 , X] = (∇X log Ω)e4 and
[e4 , Y ] = (∇Y log Ω)e4 . Then, using (4.9) we obtain
∇X e4 = χ] · X − ζ(X) · e4 , ∇4 X = χ] · X + η(X)e4
and similarly for Y . Furthermore, using also (4.7) one can easily verify the following:
h 1 1 i
g(∇X e3 , ∇4 Y ) = (χ × χ) + (η ⊗ η) − (η ⊗ η) (X, Y ),
2 2
g(∇X ∇4 e3 , Y ) = (∇X η)(Y ) − ωχ(X, Y ),
1 1
g ∇[∇4 ,X] e3 , Y = (η ⊗ η)(X, Y ) + (η ⊗ η)(X, Y ),
2 2
g(∇4 ∇X e3 , Y ) = g(∇X ∇4 e3 + ∇[∇4 ,X] e3 , Y ) + R(Y, e3 , e4 , X).
65
We have:

∇
/ 4 χ (X, Y ) = ∇4 χ (X, Y ) = ∇4 χ(X, Y ) − χ(∇4 X, Y ) − χ(X, ∇4 Y )
= ∇4 g(∇X e3 , Y ) − χ(χ] · X, Y ) − χ(X, χ] · Y )

= g(∇4 ∇X e3 , Y ) + g(∇X e3 , ∇4 Y ) − (χ × χ)(Y, X) − (χ × χ)(X, Y )

= (∇X η)(Y ) + (η ⊗ η)(X, Y ) − (χ × χ)(X, Y ) − R(X, e4 , Y, e3 ) − ωχ(X, Y ).
(4.28)
If we consider the antisymmetric part of the above equation, then since the left hand
/ η = ∇η − ∇η,
side is symmetric and since d f we obtain

/ η = χ × χ − χ × χ (X, Y ) + R(X, e4 , Y, e3 ) − R(Y, e4 , X, e3 )
d

= χ × χ − χ × χ (X, Y ) − R(X, Y, e3 , e4 ),
where in the last equation we used the first Bianchi identity. We have thus obtained the
equality of two 2-forms on Su,u . Hence, their Hodge duals are also equal:
1
/ η = χ ∧ χ − / AB RAB34 = χ ∧ χ − σ.
curl η = ∗d
2
On the other hand, the symmetric part of (4.28) reads:
1 1f 1
∇
/ 4χ = ∇/η + ∇/ η + η ⊗ η − (χ × χ + χ × χ) − ωχ − symm R(·, e4 , ·, e3 ) . (4.29)
2 2 2
The remaining two equations relate the geometry of (Su,u , g/ ) and the second fundamental
forms χ, χ with the geometry of the spacetime manifold (M, g).
The Gauss equation:
Recall that if ∇
/ denotes the induced connection on Su,u and X, Y are S vector fields
then
∇X Y = ∇/ X Y + II(X, Y ),
where II(X, Y ) ∈ (Tx S)⊥ . If X, Y, Z, W are S vector fields and R
/ denotes the induced
Riemann curvature of Su,u then the Gauss equation reads

R/(W, Z, X, Y ) = R(W, Z, X, Y ) + g II(Z, X), II(W, Y ) − g II(W, X), II(Z, Y ) .
Proof. We have

R(W, Z, X, Y ) =g ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y ] Z, W

=g ∇X ∇/ Y Z + ∇X II(Y, Z) − ∇Y ∇ / X Z − ∇Y II(X, Z) − ∇
/ [X,Y ] Z, W

=g ∇/ X∇
/Y Z − ∇ /Y ∇
/ XZ − ∇
/ [X,Y ] Z, W − g II(Y, Z), ∇X W + g II(X, Z), ∇Y W

/(W, Z, X, Y ) − g II(W, X), II(Z, Y ) + g II(Z, X), II(W, Y ) .
=R
66
Let {eA : A = 1, 2} be a local frame field for Su,u . Setting W = eA , Z = eB , X = eC , Y =
eD and using that (see (4.7))
II(eA , eB ) = χAB e3 + χAB e4 . (4.30)
we can rewrite the Gauss equation as
/ABCD + χAC χBD + χBD χAC − χAD χBC − χBC χAD .

RABCD = R (4.31)
Since Su,u is a two-dimensional (sub-)manifold, we have
R
/ABCD = K/ AB / CD ,
where K is the Gaussian curvature of Su,u . Therefore, we should think of (4.31) as an

equation between functions. In fact, we can explicitly do this by taking the double trace of
(4.31) and using / AB / AB = 2:
1
RABCD g/ AC g/ BD = K + trχtrχ − (χ, χ). (4.32)
2
The Codazzi equation:

The vector fields X, Y, Z are S-tangential. The Gauss equation expresses the S-tangential
components of R(X, Y )Z in terms of the intrinsic geometry of S and the second fundamental
form. The Codazzi equation expresses the normal to S components of R(X, Y )Z in terms of
the derivative (i.e. normal connection) of the second fundamental form II. Let us recall the
definition of the normal connection. Suppose that X is an S vector field and V is a normal
to S vector field, i.e. V ∈ (T S)⊥ . Then, the normal connection n ∇ is defined such that
n

∇X V = nor ∇X V ,
where nor ∇X V denotes the normal projection of ∇X V on (T S)⊥ . We can also extend the

normal connection to more general (normal) tensor fields on S. In particular, we have

n
∇X II (Y, Z) = n ∇X II(Y, Z) − II(∇

/ X Y, Z) − II(Y, ∇
/ X Z). (4.33)
The Codazzi equation then reads

n
∇X II (Y, Z) − ∇nY II (X, Z).

norR(X, Y )Z = (4.34)
Proof.
norR(X, Y )Z =nor∇X ∇Y Z − nor∇Y ∇X Z − nor∇[X,Y ] Z

/ Y Z + ∇nX II(Y, Z) − II Y, ∇X Z − ∇nY II(X, Z) − II([X, Y ], Z)

=II X, ∇
/ Y X, Z − ∇nY II(X, Z)

=II(X, ∇
/ Y Z) + II ∇
/ X Y, Z) + ∇nX II(Y, Z)

− II Y, ∇/ X Z − II(∇
= ∇nX II (Y, Z) − ∇nY II (X, Z).

67
Observe now that (4.7), (4.30) and (4.33) imply
h i h i
n
∇X II (Y, Z) = χ(Y, Z)ζ(X) + (∇ / X χ)(Y, Z) · e3 + χ(Y, Z)ζ(X) + (∇
/ X χ)(Y, Z) · e4 ,
(4.35)
where ζ = −ζ. Therefore, in our setting, we can rewrite the Codazzi equation (4.34) as
follows
h i
/ X χ)(Y, Z) − χ(X, Z)ζ(Y ) − (∇
norR(X, Y )Z = χ(Y, Z)ζ(X) + (∇ / Y χ)(X, Z) · e3
h i
+ χ(Y, Z)ζ(X) + (∇ / X χ)(Y, Z) − χ(X, Z)ζ(Y ) − (∇
/ Y χ)(X, Z) · e4 .
(4.36)
The above finally gives us
R(e3 , Z, X, Y ) = −χ(Y, Z)ζ(X) − (∇

/ X χ)(Y, Z) + χ(X, Z)ζ(Y ) + (∇
/ Y χ)(X, Z),
(4.37)
R(e4 , Z, X, Y ) = −χ(Y, Z)ζ(X) − (∇
/ X χ)(Y, Z) + χ(X, Z)ζ(Y ) + (∇
/ Y χ)(X, Z).
In view of the fact that
R(ei , Z, X, Y ) = ξi (Z) · / (X, Y ), i = 3, 4,
each of the above equations should be considered as an equation of one-forms acting on Z.

By setting X = eA , Y = eB , Z = eC and taking the trace in the Z, X-entries we obtain:
(g/ −1 )AC R4CAB = −χB]A ζA − (div

/ χ)B + (trχ)ζB + tr∇
/ B χ. (4.38)
/ 4 , tr] = 0 (since ∇
Since [∇ / 4 g/ = 0) we finally obtain:

(g/ −1 )AC R4CAB = − χ] · ζ − div / χ + (trχ) · ζ + d
/ trχ (eB ). (4.39)
Proof of null structure equations

Let us now impose the Einstein equations (see Proposition 4.3.1).
1. The equations R44 = R33 = 0:
The above equations are equivalent to trα = trα = 0 which in turn are equivalent to
(4.22).
2. The equations R3A = R4A = 0:
The equations R4A = 0 are equivalent to (g/ −1 )AC R4CAB = βB which in turn, in view of
(4.39), are equivalent to 4.21.
3. The equations R34 = RAB = 0:
First note that (4.11), (4.12) and (4.29), (4.32) imply the null structure equations (4.15)
and (4.20). Conversely, given (4.15) we can infer R34 = 0 from (4.29). Finally, given also
(4.20) we have RAB = 0.
68
4.6 The Characteristic Initial Value Problem
In Section 3.3 we discussed about the Cauchy problem for the Einstein equations. In par-
ticular, we saw that the initial data set consists of the triplet (H0 , g, k), where H0 is a
three-dimensional Riemannian manifold, g is the metric on H0 and k is a symmetric (0,2)
tensor field on H0 and such that g, k satisfy the constraint equations. Recall that g, k are to
be the first and second fundamental forms of H0 in M, respectively.
In this section, we will discuss in detail the formulation of the characteristic initial value
problem, i.e. the case where the initial Riemannian (spacelike) Cauchy hypesurface H0 is
replaced by two degenerate (null) hypersurfaces C ∪ C intersecting at a two-dimensional
surface S.
Motivation
Let us first motivate the formulation of the characteristic initial value problem. Let us
assume that g/ is a given degenerate metric on C ∪ C and let M be the arising spacetime
manifold and g the Lorentzian metric which satisfies the Einstein equations extending g/ on
C ∪C. Let us consider the double null foliation of (M, g) such that Ω = 1 on C ∪C. Let L be
the geodesic vector field on C, which coincides with the normalized and equivariant vector
field, and let u be its affine parameter such that u = 0 on S. Then, we obtain a foliation of
C which consists of the (spacelike) surfaces Sτ = {u = τ }. The crucial observation is that
the null second fundamental form χ on C, which recall that is defined to be the following
(0,2) tensor field on C
χ(X, Y ) = g(∇X e4 , Y ),
where X, Y ∈ Tp C, is in fact, an tensor field which depends only on the intrinsic geometry of
C (although ∇X L depends on the spacetime metric g). Indeed, the first variational formula
gives us
1
χ= L / g/ ,
2 4
and since the Lie derivative L /L is intrinsic to the hypersurface C, we deduce that g/ com-
pletely determines χ on C. On the other hand, by the Raychaudhuri equation we have
e4 (trχ) = −|χ|2 − trα,
and since χ and trχ (and ω = L(log Ω) = 0) are determined from g/ , we deduce that trα is
also determined. However, in view of the Einstein equations (see Section 4.3) we have
trα = Ric(e4 , e4 ) = 0.
This shows that one cannot arbitrarily prescribe a degenerate metric g/ on C ∪ C, since
otherwise trα would in general be non-zero.
69
A naive approach would of course be to mimic the formulation of the Cauchy problem
in which case the characteristic initial data would consist of a degenerate metric g/ , and
possibly other tensor fields on C ∪ C, subject to constraint equations. However, it turns
out that there is another very elegant approach according to which the initial data do not
contain the full knowledge of the degenerate metric g/ on C ∪ C. On the other hand, if we
want to consider the characteristic initial value problem, we need to make sure that C ∪ C
are to be null hypersurfaces in the to be constructed spacetime (M, g). The crucial step
here is to recall Corollary 2.2.1 according to which conformal transformation respect null
hypersurfaces and their null generators. In other words, instead of prescribing the full metric
g/ on C ∪ C, we could just prescribe the conformal class of the metric, namely
Conf(g/ ) = {Ag/ , A ∈ C ∞ (C ∪ C), A > 0} .
In order to see if this approach is fruitful, we need to understand the conformal properties
of the double null foliation.
Conformal properties of the double null foliation
Before, we proceed with an exposition of the conformal properties of the double null
foliation, it is important to fix a foliation of C ∪ C relative to which our quantities are
constructed.
Fixing a foliation of C ∪ C
Let L, L be null vector fields tangential to S. We consider the (unique, given the vector
fields L, L on S) foliation of C ∪ C such that Ω = 1 with respect to the actual (but unknown
for the time being) metric g/ . This choice uniquely determines the vector fields L, L on
C, C, respectively. Note that this choice is analogous to the choice Φ = 1 in Section 3.3.
The affine parameters u, u of these vector fields define the sections of the foliation of the
null hypersurfaces. In fact, the above choice determines the optical functions u, u globally.
Hence, the equivariant vector fields for any metric g̃ ∈ Conf(g/ ) coincide with L, L.
For all metrics in Conf(g/ ), we assume that all the associated tensor fields (i.e. the connec-
tion coefficients and curvature components) are defined with respect to the above foliation.
For example, if g̃ ∈ Conf(g/ ) then the second fundamental form χ̃ is defined by
˜ X L, Y ),
χ̃(X, Y ) = g̃(∇
where ∇˜ is the connection of g̃. Note that L coincides with the geodesic, the normalized and
the equivariant vector field with respect to the metric g/ . However, this√ is not the case for
the metric g̃ = Ag/ , since in this case the corresponding null lapse Ω̃ = A 6= 1 (this relation
follows immediately from the fact that (L, L) is an equivariant pair for both g/ and g̃). This
is why it is important to fix the vector field L and L with respect to which we calculate the
connection coefficients and curvature components. Note that in view of the first variational
formula we have
1 1
χ= L/ g/ , χ̃ = L/ g̃.
2 L 2 L
Since g̃ = Ag/ , we have
1 1 1 1
χ̃ = L/ g̃ = L/ (Ag/ ) = (LA)g/ + Aχ = (LA)g/ + Aχ.
2 L 2 L 2 2
70
ˆ χ̂ we have
Therefore, for the traceless parts χ̃,
ˆ = χ̃ − 1 (trχ̃)g̃ = 1 (LA)g/ + Aχ − 1 A−1 (LA + Atrχ)Ag/ = Aχ̂,

χ̃
2 2 2
where we used that g̃ AB = A−1 g/ AB . Therefore,
ˆ 2g̃ = g̃ AB g̃ CD χ̃
|χ̃| ˆBD = A−2 g/ AB g/ CD A2 χ̂AB χ̂CD = |χ̂|2 .
ˆAC χ̃ (4.40)
g/
ˆ 2 of the shear is conformally invariant! We denote e = |χ̃|
Hence, the size |χ̃| ˆ 2.
g̃ g̃
Hence, so far we have fixed the optical functions u, u, the conformal geometry of C ∪ C is
known as well as e. It is convenient from now on to work with a distinguished representative
of Conf(g/ ). At this point, we need to assume that the metric g/ is completely known on S.
Then, using canonical coordinates (u, θ) along C (and similarly for C) we obtain consider
the unique metric g̃ ∈Conf(g/ ) such that
p p
detg̃ = detg/ .

(u,θ) (0,θ)
Note that
g/ = φ2 g̃
for some φ which is uniquely determined by the above condition. Clearly, we have φ|S = 1.
Therefore, specifying of the conformal class Conf(g/ ) is equivalent to specifying the metric g̃.
The free data vs the full data
By the Raychaudhuri equation (for g/ and hence Ω = 1, ω = 0) we obtain

1 1
∂u trχ = ∇L trχ = −|χ|2g/ = − (trχ)2 − |χ̂|2g/ = − (trχ)2 − e.
2 2
Recall now that
p 1
φ2 (u, θ) = b(θ) · detg/ , b(θ) = √ . (4.41)

(u,θ) detg (0,θ)
Then, if χ is the null second fundamental form associated with g/ , the first variational formula
gives us
1 p p
χAB = ∂u g/ AB ⇒ ∂u ( detg/ ) = (trχ) · detg/ . (4.42)
2
Equations (4.41) and (4.42) imply:
∂u φ2 2
trχ = 2
= · ∂u φ. (4.43)
φ φ
Using now (4.41) and (4.43) we obtain that φ satisfies the following linear(!) equation:
1
∂u ∂u φ = − eφ.
2

Note that φ|S = 1 and, by (4.43), ∂u φS = trχ. Therefore, if we also know trχ and trχ on
S then we can determine φ on C ∪ C. This in turn determines g/ , and by first variational
formula, the second fundamental forms χ, χ.
71
Let us try to determine the rest of the connection coefficients. By the choice of our gauge
Ω = 1 we have ω = ω = 0 on C ∪ C. Hence, it remains to determine the torsion η (note that
since Ω = 1 we have η = ζ = −η. Note that we can eliminate the curvature component β
from the Codazzi equations (4.21) and the torsion equations (4.16) and hence we obtain a
linear propagation equation for the torsion (recall that χ, χ were just determined) and hence
can be explicitly solved provided we know the value of the torsion η at S.
The curvature components can now be very easily computed using the null structure
equations.
Formulation of characteristic initial value problem
Our previous analysis suggests the following:

The characteristic initial data set for the Einstein equations consists of a pair of three-
dimensional hypersurfaces intersecting at a two-dimensional surface along with the (free)
specification of the conformal class Conf(g/ ) of the degenerate metric g/ on C ∪ C as well as
the full metric g/ , the expansions trχ, trχ and the torsion η on S.
Local well-posedness
Rendall has shown that for smooth characteristic initial data there exists a unique solu-
tion to the Einstein equations in a neighborhood of the surface S.
Luk has recently extended the above result to appropriate neighborhoods of the initial
hypersurfaces C, C.
72
Chapter 5
Applications to Null Hypersurfaces
We restrict to one hypersurface C in a Lorentzian manifold (M, g), and hence for conve-
nience, we fix the gauge such that Ω = 1 on C. Then, the geodesic, the normalized and the
equivariant null frames (e1 , e2 , L, L) all coincide on C (and L is tangential to C). Let Sτ be
the level sets of the affine parameter of L on C. We will refer to ∪τ Sτ as the affine foliation
of C. Since Ω = 1 we have, in particular,
∇L L = 0, η = ζ = −η.
5.1 Jacobi Fields and Tidal Forces

A vector field X on C is called normal Jacobi vector field if LL X = 0. It is clear that X is
tangent to the sections Sτ of the affine foliation of C. The reason X is called normal Jacobi
is because X represents the infinitesimal displacement of the null generators around a fixed
null generator.
Fermi frame
73
It is possible that the null generators approach another one as is the case for the incoming
null geodesic congruence of a standard sphere in Minkowski. In this case, all null geodesics
intersect at the vertex of the cone. In order to understand quantitatively the behavior of
the null generators we need to express the normal Jacobi fields along a given null generator
Gx in terms of an orthonormal basis Ei (τ ), i = 1, 2 which is tangent to the sections Sτ .
Let us first see how one can consider such basis. Starting with an orthonormal frame
Ei (0), i = 1, 2 of Tx S0 , one ideally wants to propagate this along the null generator Gx such
that at each point Gx (τ ) one obtains an orthonormal frame of TGx (τ ) Sτ . One way to do
this, is by considering the so-called Fermi frame constructed by the following propagation
equation:
∇L Ei = −ζi L
where ζi = ζ(Ei ). The above relation immediately implies that

L g(Ei , K) = 0,
where K ∈ {Ej , L, L} , j = 1, 2, confirming the fact that if Ei (0), i = 1, 2, is an orthonormal

frame of Tx S0 then Ei (τ ), i = 1, 2, is an orthonormal frame of TGx (τ ) Sτ .
Propagation equation of Jacobi fields
If now X is a normal Jacobi field along Gx , where x ∈ S0 , then it is very convenient
to express X in terms of the Fermi frame Ei . Recall that X represents the infinitesimal
displacement of null generators nearby Gx and Ei , on the other hand, remains orthonormal.
We have:
X = X i (τ ) · Ei (τ ), (5.1)
where as usual we sum over the repeated index i. Then,
dX i (τ )
∇L X = · Ei − X i · ζi L. (5.2)
dτ
On the other hand,
g(∇L X, Ej ) = g(∇X L, EJ ) = χ(X, Ej ), (5.3)
giving us
dX i (τ )
= χij X j . (5.4)
dτ
This shows that the components X i satisfy a linear system of first order equations. However,
the connection coefficients χij depend on the regularity of null hypersurface and not on the
regularity of the ambient metric g. In order to obtain propagation equations for X i which
only depend on the Riemann curvature of g, we need to look at the second order equations.
74
The vector field X satisfies the more general Jacobi equation:
JL X = ∇L ∇L X − R(L, X)L = 0. (5.5)
Indeed, we have
R(L, X)L = ∇L ∇X L − ∇X ∇L L − ∇[L,X] L = ∇L ∇L X,
since [L, X] = 0 and ∇L L = 0. The operator JL is also known as the tidal operator and
appears naturally in the second variation of the length of Gx . Now, equation (5.5) implies
g(∇L ∇L X, Ej ) = g R(L, X)L, Ej = αij X i ,

where α is the curvature component introduced in Section 4.3. On the other hand, (5.2)
implies
i d2 X i (τ )
∇L ∇L X = .
dτ 2
Hence,
d2 X i (τ )
= αij X i . (5.6)
dτ 2
Therefore, as long as the spacetime metric g remains smooth (and hence the curvature
component α is smooth), the components X i of the normal Jacobi vector field X with
respect to the Fermi frame satisfy the above linear second order system of equations. In
view of (5.4), the initial data X i (0), dX i (0)/dτ depend nearly on X i (0) and so does the
solution X i (τ ). Therefore, there exists a smooth curve of matrices M (τ ) such that
X i (τ ) = Mji (τ )X j (0). (5.7)
Clearly
M (0) = I, and hence detM (0) = 1. (5.8)
The matrix M measures the change of the normal Jacobi field X along Gx with respect to
its initial value X(0) = Xx . For this reason, M is called the deformation matrix.
Using equation (5.4) we can obtain a propagation equation for M . Indeed, we have
dX i (τ ) dMki (τ ) k
= X (0) = χij X j (τ ) = χij Mkj X k (0).
dτ dτ
Therefore,
dMki (τ )
= χij Mkj = (χ × M )ik ,
dτ
or simply,
dM (τ )
= χ × M. (5.9)
dτ
75
5.2 Focal Points
Recall that as long as the Riemann curvature is smooth then so is the solution of (5.6). On
the other hand, the deformation matrix M is invertible as long as the flow Fτ of L induces
isomorphisms between Tx S0 and TGx (τ ) Sτ . However, the null generators nearby Gx may
converge to Gx in such a way such that X = 0 at some point Gx (τf ) along Gx . In this case,
the point Gx (τf ) is called a focal point of C. For convenience, let τf denote the first time for
which we have a focal point along Gx . The regularity of the null hypersurface breaks down
at the focal points. We will see the exact meaning of this by expressing the vanishing of the
normal Jacobi field X in terms of deformation matrix M .
Note that the matrix M (τ ) is invertible for all τ ∈ [0, τf ) but not for τ = τf and so
dM
χ= × M −1 . (5.10)
dτ
Recalling the formula for the derivative of the determinant

ddetM dM −1
= (detM ) · tr ×M
dτ dτ
we obtain
1 ddetM
trχ = . (5.11)
detM dτ
Note that the trace of χ is taken with respect to the metric g/ of Sτ which, with respect to
the Fermi frame, is simply the identity matrix. Since
ddetM
≤ 0, and detM → 0+
dτ
as τ → τf , and therefore,
trχ(τf ) = −∞. (5.12)
Hence the geometry of the null hypersurface C breaks down at the focal points. The quantity
min τf (Gx ) (5.13)

Gx ,x∈S0
is called the radius of conjugacy of the null hypersurface C.

It is not always the case that nearby null generators intersect Gx at a focal point. If this
happens then the point is called a caustic point.
Example
76
Recall that trχ = − 2r and hence note that indeed trχ → −∞ as r → 0 in Minkowski
spacetime. Hence, the radius of conjugacy of the incoming null geodesic congruence normal
to a standard sphere is equal to the radius of the sphere.
This behavior of the incoming null congruence of S2 (r) in Minkowski spacetime will be
generalized in Section 5.4.
5.3 Causality III

Recall from Section 2.4 that if S is a closed two-dimensional surface in a globally hyper-
bolic time-orientable spacetime (M, g) and C, C are its (future) outgoing and incoming null
geodesic congruence normal to S then
[
∂J + (S) ⊂ C C.
Note that we always have [

C C ⊂ J + (S),
however, it is not always the case that
[
C C ⊂ ∂J + (S).
Although the boundary J + (S) of the future of S is always part of C ∪ C, it may be the case
that part of C ∪ C lies in the interior of the future J + (S). In this case, this part can be
connected to S with timelike curves (and hence so does a neighborhood which thus belongs
in the interior of the future of S).
For example, let us consider again the case a standard sphere S in Minkowski spacetime.
In this case the null congruences C ∪ C are depicted below
whereas the boundary of the future of S is the following
77
Hence, we can see that the part of C after the focal point pf is not any more in the
boundary of the future of S. This is in fact true for any globally hyperbolic Lorentzian
manifold. Indeed, we have the following
Proposition 5.3.1. Let S be a closed two-dimensional surface in a globally hyperbolic time-
orientable spacetime (M, g). Let C and C denote the (future) outgoing and incoming null
geodesic congruence normal to S. Let C ∗ and C ∗ denote the part of C and C, respectively,
which do not contain any focal points. Then,
[
∂J + (S) ⊂ C ∗ C ∗ .
The proof of this proposition relies on a second variation of the part of the null generator
Gx between the foot x in S and the focal point. The variation is appropriately taken to
be in the direction of the normal Jacobi field which vanishes at the focal point. Then, this
variation gives rise to infinitesimally close to Gx timelike curves connecting point of S and
the focal point. Hence, the focal point is in the interior of J + (S), from which it easily
follows that the rest of the null generator Gx also lies in the interior of J + (S).
Remark
A similar property of focal points is known in Riemannian, in fact even in Euclidean,

geometry. For example, consider the following piece H of a circle in two-dimensional flat
plane:
Note that if P is between A and the center O then the length of P A is the least distance
of P from H. On the other hand, the focal point O (where the normals to H intersect) is
equidistant from all points of H. Furthermore, if Q does line between O and A, then the
length QA is not the least distance of Q from H.
The above can be generalized to any complete Riemannian manifold. Note, finally, the
importance of completeness for Riemannian geometry and global hyperbolicity for Lorentzian
geometry.
78
5.4 Trapped Surfaces
Let S be a closed surface and C ∪ C the null geodesic congruences normal to S and assume
again Ω = 1 on C ∪ C. We consider the geodesic vector fields L, L of C, C, respectively, and
let τ, τ be their affine parameters such that S = {τ = 0} = {τ = 0} ≡ S0 .
The area and the second fundamental forms χ, χ

Consider canonical coordinates (τ, θ1 , θ2 ) on C (and similarly for C), where (θ1 , θ2 ) ∈ U ⊂ R2
(see also Section 4.1). Let g/ denote the induced metric on the sections Sτ of C. In view
of the first variationa formula, and the formula for the derivative of the determinant of a
matrix, we have p p
∇L detg/ = (trχ) detg/ ,
where detg/ (τ ) denotes the determinant of the induced metric g/ (τ ) on Sτ with respect to
the coordinates (θ1 , θ2 ) ∈ U ⊂ R2 . Note that
Z p
Area(Sτ ) = detg/ (τ ) dθ1 dθ2
U
and hence Z
∇L Area(Sτ ) = trχ dµg/ , (5.14)
U
and more generally,
Z
∇f L Area(Sτ ) = f · trχ dµg/ , (5.15)
U
for any smooth non-negative function f on Sτ . The equation (5.15) expresses the relation of
the second fundamental form χ and the rate of change of the area of Sτ under infinitesimal
displacements along the null generators of C. This is the reason trχ is called the expansion
of Sτ .
Of course, the analogous equation is true for the hypersurface C.
Definition of trapped surfaces

A trapped surface is, by definition, a closed two-dimensional surface S in (M, g) for which
the area decreases under arbitrary (infinitesimal) displacements along the null generators of
both null geodesic congruences C ∪ C normal to S.
If (M, g) is globally hyperbolic, then since C ∪ C bounds the future of S, we obtain that
a trapped surface cannot expand in its future (hence the term trapped).
79
In view of formula (5.15), we have that for a trapped surface S, the integrals
Z Z
f · trχ dµg/ < 0, f · trχ dµg/ < 0
U U
for all smooth non-negative functions f on S. Hence, an equivalent definition of the trapped
surface is the following:
A trapped surface is a closed two-dimensional surface S in a Lorentzian manifold (M, g)
such that the expansions
trχ < 0, trχ < 0 (5.16)
everywhere on S.
Trapped surfaces and focal points

The null generators of the incoming null hypersurface C contain focal points. Note also that
the null expansion of a section of C is negative. The crucial observation is the following that
these two properties are not unrelated:
Proposition 5.4.1. Suppose that S is a closed two-dimensional surface (not necessarily
trapped) in a Lorentzian manifold (M, g) which satisfies the Einstein equations Ric(g) = 0.
If trχ < 0 at a point x ∈ S then there exists a focal point on the null generator Gx of C
emanating from the point x. A similar result holds for C.
Proof. In view of the Raychaudhuri equation (4.22), and (4.24), noting that ω = 0 since
Ω = 1 on C ∪ C we have
1
∇L trχ = − (trχ)2 − |χ̂|2 ≤ 0.

(5.17)
2
Hence, if T = trχx = trχ(0) < 0 then trχ(τ ) < 0 for all τ ≥ 0. Therefore, by dividing with
(trχ)2 and forgetting about the shear term we obtain
1 1
∇L − ≤− (5.18)
trχ 2
and therefore
1 1 τ
− ≤− − . (5.19)
trχ T 2
Therefore, for
2 2
τ∗ = = , (5.20)
−T −trχx
80
we obtain
trχ(τ∗ ) = −∞ (5.21)
and hence the point Gx (τ∗ ) is the first focal point on Gx .
Remark
The equality part of (5.17) is true in view of the fact that Ric(L, L) = trα = 0. On the other
hand for the inequality part of (5.17) we only need trα = Ric(L, L) ≥ 0. The latter condition
is known as the positive null energy condition and is weaker than the Einstein equations.
Hence, Proposition 5.4.1 holds if (M, g) satisfies the positive null energy condition.
Trapped surfaces in Minkowski spacetime?

We saw earlier that the incoming null geodesic congruence C normal to a standard sphere
S in Minkowski spacetime has everywhere negative expansion (trχ = −2/r) and every null
generator of C contains a focal point (which is the vertex of the cone). On the other hand,
the outgoing expansion is always positive (trχ = 2/r). The following question arises:
• Can we deform the standard sphere in Minkowski spacetime so we obtain a trapped

surface? More generally, is there a trapped surface in Minkowski spacetime?
We will answer this question in the negative in the next section. For the time being,
we remark that there are surfaces in Minkowski for which there are arbitrarily large regions
where the incoming and outgoing null expansions trχ, trχ are both negative. Indeed, the
idea is to deform a sphere so that it remains a section of the incoming null congruence C
for which the expansion trχ (of any section of C) is everywhere negative. For example, we
can consider a surface S a large part of which lies in the intersection of the past null cones
of two points p, q:
Note, however, that these regions do not cover the whole surface S. As mentioned above,
the obstruction to obtaining surfaces with globally negative (both) null expansions is the
subject of the next section.
5.5 Penrose Incompleteness Theorem

Without further delay, let us state the theorem
81
Theorem 5.5.1 (Penrose Incompleness Theorem). Let (M, g) be a globally hyperbolic time-
orientable (Hausdorff ) spacetime with a non-compact Cauchy hypersurface H such that M
contains a trapped surface S. If, in addition, (M, g) satisfies Ric(L, L) ≥ 0 for all null
vector fields L, then M is future null geodesically incomplete. In fact, at least one of the null
generators of C ∪ C, the future null geodesic congruences normal to S, cannot be extended
for all τ ≥ 0 in M.
Corollary 5.5.1. In view of the geodesic completeness of Minkowski spacetime, there are
no trapped surfaces.
Remark 5.5.1. The theorem, as shall be obvious by the proof, does not explain the reason
why the null generator cannot be extended for infinite time in M. It may be either that there
is a metric (curvature) singularity somewhere along this null generator or that this null
generator enters an extended manifold M f which is not globally hyperbolic. Both scenaria
are depicted below:
Quite surprisingly, the Penrose incompleteness theorem relies on a global topological

argument. For this reason, we first summarize the basic results needed from differential
topology:
1. A bijective continuous map from compact spaces to a Hausdorff space is a homeomor-
phism.
2. Let N ⊂ M be an (injective) immersed topological submanifold with dim(N ) =dim(M ) =
n such that N is compact and M is a Hausdorff connected non-compact topological
manifold. Then, ∂mani N 6= ∅, where ∂mani N denotes the boundary of N in the sense
of (topological) manifolds.
The second property can be easily proved by contradiction. Assume that ∂mani N = ∅.
Then every point in N has a neighborhood homeomophic to Rn and hence, in view of
the equal dimensions, it is also open in M. By compactness, it follows that N is open
in M. However, N is compact and thus closed in M, since M is Hausdorff. Hence, by
connectedness, M = N , which contradicts the fact that M is non-compact.
Proof of Theorem 5.5.1. Recall that, from the results of Section 5.2, if trχx = −kx < 0,
x ∈ S = {τ = 0}, then for the first focal point on the null generator Gx ⊂ C appears at time
τ = 2/kx . In view of the compactness of S we have

sup trχ = kC < 0, sup trχ = −kC < 0, sup kC , kC = k < 0.
S S
82
We assume that (M, g) is future null geodesically complete. We consider the union V of
all parts of the null generators of C ∪ C for which 0 ≤ τ ≤ 2/k, that is
[
V= Gx (τ ) ∪ Gx (τ ) .
τ ∈[0,2/k]
x∈S
In view of our last assumption we have V ⊂ M. Furthermore, V is compact, since it is

the union of two compact spaces, each being the image of the compact space S × [0, 2/k]
under the continuous mapping which maps (x, τ ) ∈ S × [o, 2/k] to Gx (τ ) ∈ V.
Clearly every null generator in V contains at least one focal point. Therefore, by the
results of Section 5.3, we have
∂J + (S) ⊆ V.
Since V is compact and ∂J + (S) (the topological boundary is closed by definition), we obtain
that ∂J + (S) is compact. We will next show that a global topological argument leads us to
contradiction, and hence (M, g) cannot be future null geodesically complete.
In view of the time-orientability of M, there is a global timelike vector field T whose

integral curves are timelike foliate M and intersect the Cauchy hypersurface H exactly once.
Furthermore, these integral curves intersect exactly once ∂J + (S), since J + (S) is a future
set and the (topological) boundary of a future set is achronal three-dimensional Lipschitz
submanifold without boundary in the sense of (topological) manifolds (see Proposition 2.1.2).
The projection of ∂J + (S) on H via the integral curves of T is thus an injective continuous
mapping from ∂J + (S) onto a subset T of H. Therefore, by the property 1 above, we have
that ∂J + (S) is homeomorphic to T and thus T is a Lipschitz three-dimensional compact
submanifold without boundary in the sense of (topological) manifolds (∂mani T = ∅) in the
non-compact three-dimensional manifold H. Property 2 above shows that T must have
non-empty boundary, which is contradiction.
83
Remark 5.5.2. Note that the general condition Ric(L, L) ≥ 0 for all null vectors L guar-
antees that the volume of conical regions around null geodesics is smaller (or equal) than the
Minkowskian regions and hence it can be considered as a focusing condition. Therefore, the
Penrose theorem guarantees that an almost collapsed surface (i.e. a trapped surface) must
lead to entirely collapsed geodesics (i.e. to focal points).
5.6 Killing Horizons

We finish this chapter with a discussion about a special but very important class of null
hypersurfaces. This special class of null hypersurface is characterized by the following prop-
erty: Suppose that a null hypersurface C admits a normal ξ which can be extended in a
spacetime neighborhood of H+ such that the extended ξ is Killing with respect to the metric
g. Note that ξ does not have to be null outside the null hypersurface C.
Recall that a vector field ξ of M is called Killing if
Lξ g = 0.
This then implies that the map Tξ = ∇ξ : X 7→ ∇X ξ is antisymmetric since
g(∇X ξ, Y ) + g(∇Y ξ, X) = 0, (5.22)
for all vector fields X, Y of M.

If C admits a Killing normal then it is called a Killing horizon. The Killing normal ξ
does not have to coincide with the geodesic normal; if it does, i.e. if ∇ξ ξ = 0, then C is
called extremal horizon. More generally, we will merely have
∇ξ ξ = κξ, (5.23)
for some function κ on C. The function κ is known as the surface gravity of C.
Remark
We could have defined a Killing horizon to be one such that Lξ g/ = 0, where g/ is the induced
metric on C and ξ null (and hence ξ is null Killing with respect to g/ ). However, by the
first variational formula, this is simply equivalent to χ = 0 and does not depend on the
normalization of ξ.
Properties of Killing horizons
1. Vanishing of the second fundamental form: χ = 0
Proof. As before we assume that the null lapse function Ω = 1 on C and L denote the
geodesic vector field normal to C. We define the following (0, 2) tensor field on C:
χξ (X, Y ) = g(∇X ξ, Y ),
where X, Y ∈ Tp C, p ∈ C. Note that since X, Y are tangential to C, the values of χξ depend

only on the restriction of ξ on C. Since by assumption ξ is normal to C we have
ξ = f · L,
84
where the function f on C is given by f = −g(ξ, L). Then, since L is normal to C we obtain
χξ (X, Y ) = g(∇X f L, Y ) = f · χ(X, Y ). (5.24)
Therefore, in view of the symmetry of χ, the tensor field χξ is a symmetric (0,2) tensor field
on C. On the other hand, in view of the Killing equation (5.22), the tensor field χξ is also
antisymmetric. Hence, χξ = 0 which, in view of (5.24), implies χ = 0.
2. The relation of the surface gravity and the spacetime curvature: ∇

/ κ = g(ξ, L) · β.
Proof. First note that κ is constant along the null generators of C; this can be easily shown
by taking the pushforward of (5.23) under the flow of ξ (which consists of isometries). Hence,
Lκ = 0.
Since ξ is Killing it satisfies
∇2X,Y ξ := ∇X ∇Y ξ − ∇∇X Y ξ = R(X, ξ)Y. (5.25)
A proof of this relation is given at the end of this section. If X, Y are tangential to C, then
we obtain
g(∇X Y, L) = −g(Y, ∇X L) = 0
since χ = 0. Hence ∇X Y is also tangential to C and thus all the terms in (5.25) depend
only on the restriction of ξ on C. Let us now assume that X is tangential to the sections Sτ
of the affine foliation of C and Y = ξ. Then (5.25) becomes
∇X ∇ξ ξ − ∇∇X ξ ξ = R(X, ξ)ξ.
Note that g(∇X ξ, ξ) = 21 X g(ξ, ξ) = 0 and that, if Z is any S vector field then g(∇X ξ, Z) =

χξ (X, Z) = 0. Therefore,
∇X ξ = h · ξ
for some function h on C (note that this function is related to the function f and the torsion
ζ). Then, (5.25) becomes

R(X, ξ)ξ = ∇X (κ · ξ) − h∇ξ ξ = ∇X κ · ξ + κ∇X ξ − h∇ξ ξ = ∇X κ · ξ.
Taking the inner product with L we obtain
f 2 · g R(X, L)L, L = −f ∇X κ

and hence,
∇X κ = −f · R(L, L, X, L) = g(ξ, L) · R(X, L, L, L) = g(ξ, L) · β(X).
85
3. The bifurcate sphere S
Recall that ∇ξ ξ = κξ, ∇L L = 0 and ξ = f · L. Then,
∇ξ ξ = f ∇L (f L) = f (Lf ) · L = (LF ) · ξ.
Therefore, Lf = κ, or in other words,
f (t) = κ · t + c,
where t is the affine parameter of L and c ∈ R. Then,
ξ = (κ · t + c) · L
and hence, if κ 6= 0, then for the section S of C which corresponds to t = − κc is called the
bifurcate sphere of C. Note that the bifurcate sphere S exists if the horizon is non-extremal.
Furthermore, S coincides with a section of the affine foliation if and only if the surface
gravity κ is globally constant on C. Clearly, an extremal horizon does not admit a bifurcate
sphere.
4. Conservation of the transversal second fundamental form: LL χ = 0 for extremal horizons.
Proof. Let C be an extremal horizon, i.e. κ = 0. In this case, we can take the geodesic vector
field L = ξ. Let X, Y be S tangential normal Jacobi vector fields, i.e. [L, X] = [L, Y ] = 0.
In view of Corollary 5.6.2 we obtain

LL ∇X Y = ∇LL X Y + ∇X LL Y = 0. (5.26)
On the other hand, in view of the first variational formula and χ = 0, we have LL g/ = 0.
Hence,
LL ∇ / X Y = 0. (5.27)
Moreover,
∇X Y = ∇
/ X Y + χ(X, Y ) · L + χ(X, Y ) · L = ∇
/ X Y + χ(X, Y ) · L.
Therefore, in view of (5.26), (5.27) we obtain

LL χ(X, Y ) = 0, (5.28)
which completes the proof (since LL X = LL Y = 0).
Properties of Killing fields

We finish this section by recalling several very important properties of Killing fields:
Lemma 5.6.1. Let ξ be a Killing field of a pseudo-Riemannian manifold (M, g). Then for
all vector fields X, Y the following holds
∇2(X,Y ) ξ = R(X, ξ)Y. (5.29)
86
Proof. By taking the covariant derivative of the Killing equation (5.22) with respect to a
vector field Z we obtain
g(∇Z ∇X ξ, Y ) + g(∇X ξ, ∇Z Y ) + g(∇Z ∇Y ξ, X) + g ∇Y ξ, ∇Z X) = 0,
or, using (5.22)
g(∇Z ∇X ξ, Y ) − g(∇∇Z Y ξ, ∇X ) + g(∇Z ∇Y ξ, Y ) − g ∇∇Z X ξ, Y ) = 0,
and thus
g ∇2(Z,X) ξ, Y + g ∇2(Z,Y ) ξ, X = 0.
Similarly we obtain

g − ∇2(X,Y ) ξ, Z + g − ∇2(X,Z) ξ, Y = 0,

g ∇2(Y,Z) ξ, X + g ∇2(Y,X) ξ, Z = 0.
By summing the previous three equations and the definition of the Riemann curvature we
obtain

g ∇2(Z,Y ) ξ, X + g ∇2(Y,Z) ξ, X + g R(Z, X)ξ, Y + g R(Y, X)ξ, Z = 0.
Furthermore,

2g ∇2(Z,Y ) ξ, X + g R(Y, Z)ξ, X + g R(Z, X)ξ, Y + g R(Y, X)ξ, Z = 0,
or
2g ∇2(Z,Y ) ξ, X + R(X, ξ, Y, Z) + R(Y, ξ, Z, X) + R(Z, ξ, Y, X) = 0.
Using R(Y, ξ, Z, X) = R(X, Z, ξ, Y ), R(Z, ξ, Y, X) = −R(X, Y, Z, ξ) and the first Bianchi

identity which reads
R(X, ξ, Y, Z) + R(X, Z, ξ, Y ) = −R(X, Y, Z, ξ),
we obtain
2g ∇2(Z,Y ) ξ, X = 2R(X, Y, Z, ξ)
from which the required result follows by relabeling.
Corollary 5.6.1. A Killing field ξ is completely determined everywhere by the values of ξ

and ∇ξ at a point p ∈ M. Hence, the dimension of the linear space of Killing vector fields
is at most n(n+1)
2 , where n = dimM.
Proof. Consider another point q and a curve γ which connects p and q. Then (5.29) cor-
responds to a system of linear second order ODEs for the components of ξ, ∇ξ. Moreover,
ξp lies in an n-dimensional vector space, whereas (∇ξ)p lies in an n(n−1)
2 -dimensional vector
space, since it is axisymmetric.
87
Corollary 5.6.2. Let ξ be a Killing vector field. Then

Lξ ∇X Y = ∇Lξ X Y + ∇X Lξ Y .
Proof. The proof follows immediately from Lemma 5.6.1 and the identity LX Y = ∇X Y −
∇Y X.
Another corollary of Proposition 5.6.1 is the following:
Corollary 5.6.3. Let ξ be a Killing field. Then we have
Lξ R = Lξ Ric = Lξ Rsc = 0,
where R, Ric, Rsc is the Riemann, Ricci and scalar curvature, respectively.
Alternatively, one could use the definition of Lξ using the flow of ξ, which consists of
isometries.
Proof. From (5.6.1), for all vector fields X, Y we have
Lξ (∇X Y ) = ∇Lξ X Y + ∇X (Lξ Y ),
where we have used that Lξ X = ∇ξ X − ∇X ξ. This applied several times implies

Lξ R(X, Y )Z = R(Lξ X, Y )Z + R(X, Lξ )Z + R(X, Y )Lξ Z,
from which it follows that Lξ R = 0. Since now Ric is the contraction of the tensor product
of g −1 and R and the Lie derivative Lξ commutes with contractions, satisfies the Leibniz
rule and also Lξ g −1 = 0 we obtain Lξ Ric = 0. Similar argument applies for the scalar
curvature.
We can now prove the following
Proposition 5.6.1. Let ξ be a Killing field. Then the following identities hold
Div(ξ) = 0,

Div Rsc ξ = 0,
2g ξ = Div(∇ξ) = Ric(ξ, ·),

Div Ric(ξ, ·) = 0,
and hence the current

J = Div(∇ξ) (5.30)
is conserved (i.e. divergence-free).
Proof. The first identity is trivial since for Killing ξ the (1, 1) tensor ∇ξ is antisymmetric
and hence its trace vanishes. For the second identity observe that
Dic(Rsc ξ) = Rsc Div(ξ) + ∇ξ Rsc = 0.
88
Regarding the third identiy we have

2g ξ = Div(∇ξ) = traceX,Y ∇2(X,Y ) ξ = traceX,Y R(X, ξ)Y = Ric(ξ, ·),
where we have used (5.29). For the last one, recall that the Einstein tensor
1
G = Ric − Rsc g
2
satisfies Div(G) = 0. Moreover,
Div G(ξ, ·) = 0,
since
n
X
Div G(ξ, ·) = Div(G) (ξ) + G(∇Ei ξ, Ei ),
i=1
P the fact that G is symmetric and that ∇ξ is antisymmetric and noting that
and using
∇Ei ξ = nj= g(∇Ei ξ, Ej )Ej . Furthermore,
1
Div G(ξ, ·) = Div Ric(ξ, ·) − Div Rsc g(ξ, ·)
2
and since

Div Rsc g(ξ, ·) = Div Rsc (ξ[ ) = Div (Rsc ξ)[ = Div Rsc ξ = 0,
we obtain the required result.

Alternatively, note that Div Div(∇ξ) = 0, because ∇ξ is antisymmetric and the di-
vergence of the divergence of an antisymmetric tensor vanishes.
Corollary 5.6.4. If (M, g) is Ricci flat (i.e. Ric = 0) and ξ is a Killing field, then ξ
satisfies the Maxwell equations.
89
Chapter 6
Christodoulou’s Memory Effect
In this chapter we describe the celebrated nonlinear memory effect of the gravitational field
due to Christodoulou. The only way we can study isolated systems in the universe is by
investigating the radiation/signals that reaches us from these systems. We therefore, need a
well-defined notion for the region where radiation scatters. This is precisely the future null
infinity I + .
6.1 The Null Infinity I +

Heuristically, the future null infinity I + consists of all ideal limit points of null geodesics
which reach arbitrarily large spatial distances.
Definition
One can give a precise meaning to this construction by considering conformal transforma-
tions. Indeed, conformal transformations do not change the causal structure and, in fact,
send null geodesics to null geodesics. Recall (see Section 1.4.2) that Minkowski spacetime
can be conformally embedded in Einstein’s static universe E such that the closure of the
image is compact and hence has a boundary in E. Recall also that this boundary consists of
several components each of which is of fundamental importance in the study of the causal
structure of the spacetime.
In this section we shall generalize these notions to more general Lorentzian manifolds.
Let, therefore, (M, g) be a globally hyperbolic time-orientable Lorentzian manifold. Sup-
pose that (M, g) can be conformally embedded in another Lorentzian manifold (M̃, g) such
that the closure of the image (also known as the conformal compactification) has a bound-
90
ary, which consists of two components; the future boundary and the past boundary, both
intersecting at a point i0 . We shall refer to this point as the spacelike infinity.
Let us consider the future boundary B of the conformal compactification of (M, g). Let
2
Ω be the conformal factor which we assume that it satisfies the following properties
1. Ω > 0 on M,
2. Ω extends appropriately to B such that Ω = 0 on B,
3. Ω satisfies the eikonal equation g(∇Ω, ∇Ω) = 0 on B.
Then the future boundary B is a null hypersurface since it is a level set of Ω and hence
its normal is the null vector field ∇Ω. We shall call this null hypersurface the future null
infinity I + . Note that I + is an incoming null hypersurface. For, I + is the future boundary
of a set.
Asymptotical flatness and null infinity

Of course there are spacetime which do not admit such a null infinity. However, any asym-
potically flat spacetime does so. Let us explain this further:
We are mostly interested in studying isolated systems in the universe. Hence, we can
assume that arbitrarily far away from these systems the spacetime approaches Minkowski
spacetime. Since general relativity is a dynamical theory we can only impose restrictions on
a Cauchy hypersurface (say Σ ∼ R3 ), since then the domain of dependence of Σ is uniquely
determined. Hence, let us assume that the data on Σ are asymptotically flat, i.e. approach
Minkowskean data at infinity. Then, there exists a sphere S0 in Σ such that the data on its
complement B in Σ is a small perturbation of Minkowski data.
Then by the stability of Minkowski spacetime proved by Christodoulou–Klainerman and
Klainerman–Nicolo we can conclude that one can attach a piece of future null infinity at the
domain of dependence D+ (B) of B:
91
Note that the aforementioned works on the stability of Minkowski proved that the con-
1
formal factor Ω2 extends to I + as a function in C 1, 2 . In fact, Christodoulou has shown that
for physically relevant spacetime the conformal factor Ω2 cannot belong in C 2 (for more see
also Part II of these notes).
The works on the stability of Minkowski showed also that the null generators of this
piece of null infinity are past complete in the following sense: Let C0 be the outgoing null
geodesic congruence normal to S0 . Let L be the outgoing null vector field normal to S0
given by L = √12 (T + N ), where T is the unit future directed vector field normal to Σ and
N is the outward normal to S0 in Σ. This choice then yields an affine foliation Sτ of C0
(for which we can take τ → +∞). For each section of this foliation we consider the affine
foliation of the incoming null hypersurfaces as depicted below:
Then the affine time for which null generators of the incoming null hypersurfaces intersect
Σ tends to infinity as τ → +∞.
Completeness of null infinity

As we just explained, for asymptotically flat spacetimes the future null infinity is past com-
plete. However, it is of fundamental significance to know if the future null infinity is also
future complete, in a similar sense as above:
In this case, we measure the affine time uτ for which the depicted null generators remain
in the domain of dependence of the sections Sτ of C0 and we require uτ → +∞ as τ → +∞.
92
Weak cosmic censorship
Generic initial data for the Einstein equations lead to developments with complete future
null infinity.
6.2 Tracing gravitational waves

In this section we describe the basic setup of a gravitational wave experiment. One of
the fundamental assumptions is that we, the observers, are located very far away from the
source of the gravitational waves and hence we only rely on the gravitational radiation which
reaches us. Hence we assume the experiment takes place along future null infinity.
The experiment is based on the following principle: “A gravitational wave forces masses
to change their position”. However, how can one measure this displacement? Considering
just one mass would be hopeless, since masses move on timelike geodesics. Therefore, one
needs to consider two or more masses and study their relative motion. If we find a difference
in the relative distance of the masses then we could deduce that a gravitational wave has
passed.
The most basic setup would be to consider two masses. However, in order to study com-
pletely the problem, we consider three (free) masses m0 , m1 , m2 which lie initially (i.e. when
the masses are at rest) on a plane Π = Π(0). We assume that m0 , m1 , m2 form initially a
right isosceles triangle on Π(0) and that the size of the equal sides is d.
We are then interested in investigating the relative motions of the pairs (m0 , m1 ) and
(m0 , m2 ) and comparing these motions too.
93
Let T be tangent to the timelike geodesic γ of m0 such that ∇T T = 0. Let also assume
that initially
m0 m1 = X1 (0) = d · E1 (0), m0 m2 = X2 (0) = d · E2 (0),
where (E1 (0), E2 (0)) form an orthonormal frame on Π(0). We then completely the or-
thonormal frame of the three-dimensional space: (E1 (0), E2 (0), E3 (0)). We parallely propa-
gate this frame along γ according to ∇T Ei = 0, obtaining therefore an orthonormal frame
(E1 (t), E2 (t), E3 (t)) known as the Fermi frame. Here t denotes the affine parameter of T .
A priori the motion of the masses may not be constraint in the planes Π(t) = hE1 (t), E2 (t)i.
However, under very reasonable physical assumptions, the timelike geodesics of m1 , m2 can
be approximated by the associated Jacobi fields X1 , X2 along γ given by following relations:
X1 = d · E1 , ∇T X1 = 0,
X2 = d · E2 , ∇T X2 = 0.
Note that the relations ∇T Xi = 0, i = 1, 2 represent the fact that the masses are assumed
to be at rest with respect to m0 (hence no gravitational waves are present at the beginning
of the experiment).
Recall that the whole configuration is supposed to propagate along a null generator of
future null infinity. We can assume therefore, that the planes hE1 (t), E2 (t)i are tangential
to the sections of null infinity along a fixed null generator. In this case, E3 (t) points towards
the direction of the source:
94
6.3 Peeling and Asymptotic Quantities
Consider the sections Sτ of the affine foliation of C0 and the incoming null geodesic congru-
ence C τ normal to Sτ . Renormalize L, L on C0 ∩ C τ such that trχ + trχ = 0. Consider then
the associated affine foliation S u of C τ . Then, from the work of the stability of Minkowski
spacetime (see Part II), we have precise limits for all appropriately rescaled geometric quan-
tities (metric, connection coefficients, curvature components) of S u . The rescalings involve
the radius coordinate r given by the area of the sections. Note that r → +∞ at I + .
By renormalizing the null structure equations we obtain the asymptotic null structure
equations for the asymptotic gravitational field along the null infinity I + .
6.4 The Memory Effect

Assume the source is perpendicular to the horizontal plane spanned by E1 , E2 .
We will next describe the qualitative conclusions of the memory effect. For more precise
quantitative results see Part II of these notes.
1. The motion of the masses is constrained in the horizontal plane hE1 , E2 i. Hence no
movement in the vertical E3 direction is to be observed.
95
2. Acceleration of m1 relative to m0 is at each instance the rotation by 90o degrees of
the acceleration of m2 relative to m0 .
3. The masses m(i) suffer a permanent displacement ∆x(i) given by asymptotic quantities
of the gravitational field (i.e. the asymptotics of the Riemann tensor).
This permanent displacement is known as the Christodoulou memory effect.
96
Chapter 7
Black Holes
In this chapter we introduce the notion of a black hole region. We will discuss the intimate
relation of black holes and trapped surfaces and then prove, using the machinery that we
have developed so far general properties of such regions (known as black hole mechanics). We
will next describe the main properties of the most basic black hole, namely the Schwarzschild
black hole, and then proceed by presenting the Kerr metric.
7.1 Introduction
Recall that the future null infinity consists of the ideal endpoints of future-directed null
geodesics which reach arbitrarily large distances. A region which cannot send signals to
arbitrarily large distances is called black hole. Formally speaking, the black hole region
in a spacetime M is the non-empty complement of the past of the future null infinity,
i.e. M/J − (I + ). The complementary region, i.e. the region which can communicate with I +
is called domain of outer communications. Their common boundary is called future event
horizon.
Remark 7.1.1. A usual additional condition in the definition of the black hole region is the
completeness of future null infinity I + . This condition guarantees that ideal observes along
I + (such as ourselves) live forever yet never receive signals from the black hole region.
black hole event

horizon
ll
nu nity
fi
in
97
7.2 Black Holes and Trapped Surfaces
The presence of a trapped surface in a vacuum spacetime implies the existence of a black
hole region.
Proposition 7.2.1. Let (M, g) be a vacuum spacetime. If S is a closed trapped surface in

M then S cannot lie in the domain of outer communications J − (I + ). Hence M must have
a black hole region and S must lie in it.
Proof. Assume in the contrary that S ⊂ J + (I + ). Clearly there exists p1 ∈ I + such that
/ J + (S) and p2 ∈ I + such that p2 ∈ J + (S). Hence, there must exist p ∈ I + such that
p1 ∈
p ∈ ∂J + (S). However, in view of Proposition 2.4.2 we have p ∈ C, where C is the outgoing
null geodesic congruence normal to S. Hence p lies on a null generator γ of C and since
p ∈ I + it follows that γ must be complete and always in ∂J + (S). However, as is shown
in Section 5.4, since S is trapped, all null generators of J + (S) have finite affine length,
contradiction.
Although the above proposition does not show that I + must be complete, nonetheless
it makes clear the need to study evolutionary criteria for the formation of trapped surfaces.
For more on this see Part II of these notes.
7.3 Black Hole Mechanics

Let (M, g) contain a black hole region. The event horizon H+ must be a null hypersurface
since it is the boundary of the past of a set. Of particular importance is the case where H+
is a Killing horizon (see Section 5.6). We then have the following
Proposition 7.3.1 (Zeroth law of black hole mechanics). If a Lorentzian manifold (M, g)
satisfies the Einstein equations Ric(g) = 0 and contain a Killing horizon H+ . Then the
surface gravity κ is constant on H+ .
Proof. Recall that in Section 5.6 it was shown that
∇
/ κ ∼ β.
Hence in order to show that κ is constant on H+ is suffices to show that β = 0. However,

this follows immediately from the null Codazzi equation 4.21 and the fact that χ = 0 on the
Killing horizon H+ .
The remaining laws can be found in any standard textbook. We here mention that
the first law is related to the conservation of energy and the second law is related to the
non-decrease of the entropy. Finally, according to the third law, spacetimes which contain
extremal black holes, i.e. such that κ = 0, should not form dynamically (see Chapter 10 for
more a discussion of the wave propagation on such spacetimes).
98
7.4 Spherical Symmetry
7.4.1 General Setting
• A spacetime (M, g) is called spherically symmetric if SO(3) acts by isometry and
the orbits of this group are two-dimensional spheres. The area-radius coordinate r =
(A/4π)1/2 is very important (A is the area of the orbits). Then, in spherical coordinates
(θ, φ) the induced metric on the symmetry spheres is r2 (dθ2 + sin2 θdφ2 ).
Birkhoff ’s theorem: The one-parameter Schwarzschild metrics gM is the unique spher-

ically symmetric one-parameter of metrics g which satisfy Ric(g) = 0.
It turns out that these metrics admit additional symmetries. We have the following
definitions:
• A spacetime (M, g) is called stationary, if it admits a timelike Killing field, that is

to say, if it admits a one-parameter family of isometries (the flow of the Killing field)
whose orbits are timelike curves.
• A spacetime (M, g) is called static, if it is stationary and if, in addition, there exists
a spacelike hypersurface Σ such which is orthogonal to the stationary Killing vector
field.
Derivation of the metric
First, it is somewhat easier to assume that the spacetime is static and spherically sym-
metric (recall that Birkhoff’s theorem does not require staticity; see below). Let us assume,
in fact, that the spacetime admits a unique static Killing field T . Let Σ be a spacelike hy-
persurface orthogonal to T . One can easily see that Σ is foliated by spheres of symmetry. To
show this it suffices to show that T is orthogonal to the spheres of symmetry. Indeed, since
T is unique in the above sense, it must be invariant under the action of SO(3); therefore, its
projection on the spheres of symmetry must be an invariant vector field and hence it must
vanish.
Coordinates on Σ: Let (θ, φ) be coordinates on a sphere of symmetry S embedded in
Σ. If p ∈ S then there exists a unique geodesic starting at p with initial velocity v such
that v ∈ Tp Σ and v ⊥ Tp S. Then we define (θ, φ) to be constant along such geodesics. We
parametrize the spheres by the area coordinate r, obtaining the system (r, θ, φ) for Σ.
Coordinates on M: We next use the staticity. Let Σ = Σ0 . Let t be the parameter of
integral curves of T such that t = 0 on Σ0 . Then, M = ∪Σt , where Σt is the image of Σ0
under the flow of T . In this way, we can pushforward the system of Σ0 onto M obtaining
the system (t, r, θ, φ). It is easy to see that the metric with respect to this system must take
the form
g = −f (r)dt2 + h(r)dr2 + r2 (dθ2 + sin2 dφ2 ),
where f, h > 0. Note that this system breaks down at the points where T and ∇r become
collinear.
By the equations Ric(g) = 0, one then obtains that there exists a constant M ∈ R such
that
1 2M
h(r) = − =1− .
f (r) r
99
We next describe how to arrive to (7.4.1) without assuming staticity (completing the
proof of Birkhoff’s theorem). Consider a point p ∈ S where S is a sphere of symmetry.
There exists a 1-parameter subgroup Ip of SO(3) which fixes p. Let O(p) be the set of
fixed points of Ip . Then O(p) consists of two connected-components and is a orthogonal
to S at q (and similarly it is orthogonal to any other sphere of symmetry at their common
point). Hence O(p) is a timelike two-dimensional plane. Let (θ, φ) be spherical coordinates
on S. Extend (θ, φ) such that they are constant on each connected component of O(p),
for each p ∈ S. We can now introduce coordinates (t, r) coordinates on O(p) and extend
them globally such that t and r are constant on the spheres of symmetry. Then the metric
with respect to (t, r, θ, φ) looks like (7.4.1), with the only difference being that f = f (t, r)
and h = h(t, r). However, by solving the equations Ric(g) = 0 we again obtain the same
solutions.
7.4.2 Schwarzschild Black Holes

The Schwarzschild metric in (t, r, θ, φ) is thus given by
1 2
g = −Ddt2 + dr + r2 gS2 (θ,φ) ,
D
where
2M
D = 1− ,
r
and M ∈ R is a constant. We will be mainly interested in the case where M > 0. The metric
is manifestly spherically symmetric and the vector field T = ∂t is Killing and timelike for
r > 2M and null at r = 2M , where however the metric component grr = D−1 is singular
(see below for more about this “singularity”). We note the topology of M is such that it is
covered by the coordinate system (t, r, θ, φ) with respect to which the metric g makes sense.
Killing fields
Since SO(3) acts by isometry, we have that there exists three linearly independent Killing
fields Ω1 , Ω2 , Ω3 which arise from this action. As mentioned above, the vector field T = ∂t
is also Killing. It turns out that all Killing fields are only those spanned by T, Ω1 , Ω2 , Ω3 .
The geodesic flow
Let γ(τ ) = (t(τ ), r(τ ), θ(τ ), φ(τ )) be a geodesic of (M, g). In view of the spherical
symmetry we may assume that γ lies on the equatorial plane θ = π/2. Suppose that γ is
timelike or null. Without loss of generality we may assume that Ω1 = ∂φ = Φ. Then we
have the following integrals of motion:
· ·
1. g(γ, γ) = −κ, where κ = 0 or κ = −1.
·
2. E = −g(T, γ), which we will refer to as the energy per unit mass of the particle.
·
3. L = g(Φ, γ), which we will refer to as the angular momentum per unit mass of the
particle.
100
We can see that the Killing fields provide enough integrals of motion so the Hamilton–
Jacobi equations for the geodesic flow are completely integrable. In particular, r = r(τ )
satisfies the following (separated) ODE:
2
dr
+ V (r) = E 2 ,
dτ
where the potential V is given by
1 M L2 L2
V (r) = κ − κ + 2 −M 3.
2 r 2r r
One can easily derive properties of causal geodesics depending on the values of E, L
relative to M . In particular one can see that there exists stable circular (i.e of constant r)
timelike geodesics for r > 6M , whereas the unstable circular timelike geodesics are restricted
to 3M < r < 6M . On the other hand, in the case of null geodesics, we only have unstable
circular null geodesics at r = 3M .
The maximal extension
It turns out that the “singularity” at r = 2M (for M > 0) is merely a coordinate

singularity which can be overcome by a coordinate transformation. Consider the tortoise
coordinate r∗ = r∗ (r) such that
dr∗ 1
= .
dr D
∗ ∗
Then r → −∞ as r → 2M and r ∼ r for large r. If
v = t + r∗
then the metric in the coordinates1 (v, r, θ, φ) takes the form
g = −Ddv 2 + 2dvdr + r2 gS2 (θ,φ) , (7.1)
where again D = 1 − 2M r . Clearly, this metric satisfies Ric(g) = 0 and can be extended
beyond r = 2M (although r = 0 is indeed a curvature singularity). In fact, the Kruskal
maximal extension is depicted below:
1
These coordinates are known as ingoing Eddington–Finkelstein.
101
We restrict to the domain covered by the (v, r, θ, φ) coordinate in the Schwarzschild
spacetime. In this domain T = ∂v whereas Y = ∂r is tangential to incoming null geodesics
(and is normalized so it differentiates with respect to r). In particular, Y is transversal to
H+ .
7.5 Kerr Black Holes

The Kerr metric with respect to the Boyer-Lindquist coordinates (t, r, θ, φ) is given by
g = gtt dt2 + grr dr2 + gφφ dφ2 + gθθ dθ2 + 2gtφ dtdφ,
where
∆ − a2 sin2 θ ρ2 2M ar sin2 θ
gtt = − , grr = , gtφ = − ,
ρ2 ∆ ρ2
(r2 + a2 )2 − a2 ∆ sin2 θ
gφφ = sin2 θ, gθθ = ρ2
ρ2
with
∆ = r2 − 2M r + a2 , ρ2 = r2 + a2 cos2 θ. (7.2)
Schwarzschild corresponds to the case a = 0, subextreme Kerr to |a| < M and extreme Kerr
to |a| = M .
Note that the metric component grr is singular precisely at the points where ∆ = 0. To
overcome this coordinate singularity we introduce the following functions r∗ (r), φ∗ (φ, r) and
v(t, r∗ ) such that Z 2
r + a2
Z
∗ ∗ a
r = , φ =φ+ , v = t + r∗
∆ ∆
In the ingoing Eddington-Finkelstein coordinates (v, r, θ, φ∗ ) the metric takes the form
g = gvv dv 2 + grr dr2 + gφ∗ φ∗ (dφ∗ )2 + gθθ dθ2 + 2gvr dvdr + 2gvφ∗ dvdφ∗ + 2grφ∗ drdφ∗ ,
where

2M r
gvv =− 1− 2 , grr = 0, gφ∗ φ∗ = gφφ , gθθ = ρ2
ρ
(7.3)
2
2M ar sin θ
gvr = 1, gvφ∗ = − , grφ∗ = −a sin2 θ.
ρ2
102
For completeness, we include the computation for the inverse of the metric in (v, r, θ, φ∗ )
coordinates:
a2 sin2 θ ∆ ∗ φ∗ 1 1
g vv = , g rr = , gφ = , g θθ =
ρ2 ρ2 ρ2 sin2 θ ρ2
r2 + a2 ∗ a ∗ a
g vr = , g vφ = , g rφ = .
ρ2 ρ2 ρ2
103
Chapter 8
Lagrangian Theories and the

Variational Principle
8.1 Matter Fields

Lagrangian mechanics, introduced by French mathematician Lagrange (1788), is a re-formulation
of classical mechanics that combines conservation of momentum with conservation of energy.
It was first introduced as an alternative way to calculate the trajectory of a particle in a
dynamical system, but its applications are now far reaching.
The Lagrangian is a function associated to the system and contains all necessary infor-
mation about its dynamics. Christodoulou defined the Lagrangian to be a n-form given
by the composition of sections of appropriate vector bundles. In our case, the system is
described by the maps φ : M → R and its Lagrangian is a function
L : R, T ∗ M, g −1 → R

(8.1)
where g −1 is the inverse of the pseudo-Riemannian metric on M. In this way, the Lagrangian
L = L φ, dφ, g −1 .

8.2 The Action Principle

According to the action principle the evolution of a dynamical system are is done in accor-
dance with a quantity that takes an extremum value. For example, in Newtonian Mechanics,
the evolution is such that K − V takes the least value, where K is the total kinetic energy
and V the total potential. Therefore, we only have to consider that L = K − V at each
point and then take the integral of L over our manifold (in this case the real line-time) in
order to obtain the total energies.
More generally, we need to integrate the LagrangianR over M. For this reason, we use
the natural volume form dg. Then, the number S = M Ldg is called the action of the
Lagragian.
Action Principle: “The actual evolution of the system (i.e. the actual φ) is such that
S attains an extremum value.”
Some remarks on this important principle:
104
• In case of Newtonian Mechanics, action principle is equivalent to the third law of
Newton. However, Lagranges equations hold in any coordinate system, while Newtons
are restricted to an inertial frame of reference.
• All the fundamental laws of physics can be written in terms of an action principle. This
includes electromagnetism, general relativity, the standard model of particle physics,
etc.
• This is an example of a variational principle and so the use of variational calculus is

fundamental.
If a system admits a Lagrangian, then by definition the action principle holds. This
principle will give us the laws of evolution and this is what we study next.
Euler Lagrange Equations

The Euler Lagrange equations are the equations that the action principle yields. Consider
the coordinate system x = (x1 , ..., xn ). We want to extremize
Z Z
−1
1 n
∂φ ∂φ −1
S (φ) = L φ, dφ, g dg = L φ x ,...,x , 1,..., n,g dg.
M M ∂x ∂x
The first variational principle gives us

d
0= S (φ + h) ,
d =0
for any h : M → R that vanishes on the boundary ∂M so that the “endpoints” of the system
are the same in the variation. Chain rule yields
n
Z !
∂L X ∂L ∂h
0= h + dg (8.2)
M ∂φ ∂φk ∂xk
k=1
h ik
∂φ ∂L ∂L
where φk = ∂xk
. If is the vector field with components
∂dφ ∂φk (w.r.t. the system x)
h i
∂L
then applying divergence theorem for X = h ∂dφ we take
!
∂L k ∂h
Z Z
∂L ∂L
Div h dg = + h Div dg
M ∂dφ M ∂dφ ∂xk ∂dφ
Z (8.3)
∂L
= h · n dgn−1 .
∂M ∂dφ
However, the boundary condition for h makes the boundary integral vanish. Therefore
!
∂L k ∂h
Z Z
∂L
dg = −h Div dg
M ∂dφ ∂xk M ∂dφ
and so equation (8.2) can be written as

Z
∂L ∂L
0= h − h Div dg. (8.4)
M ∂φ ∂dφ
105
Since (8.4) is true for any such h, the fundamental lemma of calculus of variation gives

∂L ∂L
− Div = 0. (8.5)
∂φ ∂dφ
This is exactly the Euler Lagrange equation. In general, if φ : M → N then we have such
an equation for every component of φ.
8.3 Derivation of the Energy Momentum Tensor

We now have all the necessary machinery to compute the energy momentum tensor of
Lagrangian matter fields. We remind that we are seeking for a (0, 2) tensor field T that is
symmetric and divergence-free. The first condition is rather physically reasonable condition
and the second expresses the
infinitesimal conservation of energy. Let Rus suppose that the
Lagrangian is L φ, dφ, g −1 and the action of the Lagrangian is S (φ) = M L φ, dφ, g −1 dg.
We know that the Lagrangian contains all physical information concerning the system and
the forces acting on it and the action of the Lagrangian gives us the laws of evolution of the
system. Therefore, if we want to have a version of the conservation of energy then we should
study more closely (and geometrically!) the action S.
Suppose we want to compute the integral S (φ). For such a computation we definitely
need a coordinate system x on M. On the other hand, every diffeomorphism f : M → M
induces a pullback coordinate system y = f ∗ x. Therefore, our problem now is to construct
diffeomorphisms of M. But what are the most natural diffeomorphisms we can consider on
M? Let us look at vector fields of M. Therefore, let X be an arbitrary vector field. Since
we seek for an object on M with local properties we may assumme that X is compactly
supported, i.e. X = 0 outside a compact region U ⊂ M. Now let Ft be the associated flow
of X. Each such diffeomorphism Ft defines a pullback coordinate system yt = Ft∗ x whose
change of coordinates is Ft . Then
Z Z
−1
Ft∗ L φ, dφ, g −1 dg .

L φ, dφ, g dg = (8.6)
M Ft−1 (M)
Clearly Ft−1 (M) = M and so

Z Z
−1
Ft∗ L φ, dφ, g −1 dg .

L φ, dφ, g dg = (8.7)
M M
Equation (8.7) is the key identity! Indeed, (8.7) gives us

Z
L φ, dφ, g −1 dg − Ft∗ L φ, dφ, g −1 dg

0=
M
or
Ft∗ L φ, dφ, g −1 dg − L φ, dφ, g −1 dg
Z
0=
M t
or !
Ft∗ L φ, dφ, g −1 dg − L φ, dφ, g −1 dg
Z
0= lim
M t→0 t
106
and by the definition of the Lie derivative we get
Z
LX L φ, dφ, g −1 dg .

0= (8.8)
M
We were led to equation (8.8) by only using the geometric structure of M. Applying the
chain and Leibnitz rule we obtain
Z
∂L ∂L ∂L
0= LX φ + LX dφ + −1 LX g −1 dg + L · LX dg . (8.9)
M ∂φ ∂dφ ∂g
We now have the following:
1. Cartan’s identity gives us
LX dφ = d(Xφ) = d (LX φ) . (8.10)
h ik h i
∂L ∂L ∂φ
2. Consider the vector field ∂dφ = ∂φk , where φk = ∂xk
. Then

∂L ∂L ∂L
Div LX φ = d (LX φ) + (LX φ) · Div . (8.11)
∂dφ ∂dφ ∂dφ
h i
∂L
3. Applying divergence theorem for the vector field ∂dφ LX φ we get
Z Z
∂L ∂L
Div LX φ dg = LX φ · n dgn−1 = 0, (8.12)
M ∂dφ ∂M ∂dφ
since, the boundary condition X = 0 implies that LX φ = X (φ) = 0 on ∂M.
4. Equations (8.10), (8.11) and (8.12) imply that

Z Z
∂L ∂L
LX dφ dg = d (LX φ) dg
M ∂dφ M ∂dφ
Z (8.13)
∂L
= − (LX φ) · Div dg.
M ∂dφ
Therefore, using equation (8.13) we can write (8.9) as

Z
∂L ∂L ∂L −1
0= LX φ − (LX φ) · Div + −1 LX g dg + L · LX dg
M ∂φ ∂dφ ∂g
or Z
∂L ∂L ∂L −1
0= − Div LX φ + −1 LX g dg + L · LX dg .
M ∂φ ∂dφ ∂g
However, according to Euler Lagrange equation (8.5) we obtain
Z
∂L −1
0= LX g dg + L · LX dg ,
M ∂g −1
107
which can be written as
Z
∂L 1
0= LX g −1 + L · gij (LX g)ij dg
M ∂g −1 2
or Z
∂L ij 1
0= ij
LX g −1 + L · gij (LX g)ij dg
M ∂g 2
or, Z
∂L 1 ij
0= − ij + gij L (LX g) dg. (8.14)
M ∂g 2
The left term in the integrad of (8.14) is independent of the vector field X and so let’s define
the tensor Q by
∂L 1
Qij = − gij L . (8.15)
∂g ij 2
Clearly Q is symmetric. Also, since the integral of the product of Q and a derivative is zero
then by integrating by parts we expect the integral of the product of the divergence of Q
and a vector field to be zero. Since the latter is going to be true for any vector field then
we expect that the divergence of T vanishes. This is exactly what we were looking for and
so it is the most natural candidate for the energy momentrum tensor. First, let us prove in
detail the following
Theorem 8.3.1. The tensor Q defined by (9.1) is divergence- free.
Proof. We have
(LX g)ij = Xi;j + Xj;i
therefore equation (8.14) can be written as
Z
Qij Xi;j + Xj;i dg

0=
M
or Z
Qij Xi;j + Qij Xj;i dg

0=
M
and since Q is symmetric we equivalently have
Z
Qij Xi;j + Qji Xj;i dg

0=
M
or Z
Qij Xj;i dg.

0= (8.16)
M
We have
Div (QX) = (Div (Q)) (X) + Q (∇X)
and so Z Z Z
Qji Xj;i dg.

Div (QX) dg = (Div (Q)) (X) dg + (8.17)
M M M
Applying divergence theorem for the vector field Y = QX = Q (X) we get
Z Z
Div (QX) dg = ((QX) · ndgn−1 ) = 0 (8.18)
M ∂M
108
since we have assumed that X is zero on the boundary ∂M. Therefore, equations (8.16),(8.17)
and (8.18) give Z
0= (Div (Q)) (X) dg, (8.19)
M
which is true for any compactly supported X. This implies that Div (Q) = 0. Indeed, if
there is a point p ∈ M where Div (Q) 6= 0 then we can find a vector Xp at p such that
Div (Q) (Xp ) > 0. By continuity we can extend Xp to a compactly supported vector field
whose support contains p and is sufficiently “small” and such that Div (Q) (Xp ) > 0. Then
integrating this we get a strictly positive number with contradicts (8.19).
We now have the following definition:
Definition 8.3.1. The tensor Q given by (9.1) is called the energy momentum tensor of the
Lagrangian matter field φ.
8.4 Application to Linear Waves

One of the most important examples of matter fields are those given by the following La-
grangian
L φ, dφ, g −1 = g −1 (dφ, dφ) = g µν ∂µ φ∂ν φ(= g(∇φ, ∇φ)).

(8.20)
Then the Euler Lagrange equations (8.5) imply that φ satisfies
4g φ = 0, (8.21)
where 4g is the Laplacian on (M, g). If g is positive definite, then φ is a harmonic function
on the Riemannian manifold (M, g). If g is a Lorentzian metric then φ is a linear wave. In
this case, there is a special symbol for the Laplacian on Lorentzian manifolds:
2g φ = 0. (8.22)
Equation (9.1) gives us the energy momentum tensor of scalar linear waves on Lorentzian
manifolds
1
Qµν = ∂µ φ∂ν φ − g −1 (dφ, dφ) gµν . (8.23)
2
This tensor plays a fundamental role in the analysis of the evolution of waves.
8.5 Noether’s Theorem

We saw how exploiting the geometry of the space and the Lagrangian structure of the system
to derive an infinitesimal conservation of energy. If one wants to derive stronger versions
of these laws, then stronger conditions have to be imposed on M too. Noether’s theorem
expresses in the most general way these conditions.
Let us see this more closely. The geometry of the space gives us equation (8.7) which is
the key for the derivation of energy momentum tensor. It says that the “average” of those
Lagrangians over the manifold M is the same. Obviously, a stronger condition that may be
imposed is the following
L φ, dφ, g −1 = Ft∗ L φ, dφ, g −1 .

(8.24)
109
This expresses a one-parameter symmetries of the Lagrangian. Then Noether’s theorem
gives a conservation law for the Euler Lagrange equation. For example, if a physical system
behaves the same regardless of how it is oriented in space, its Lagrangian is rotationally
symmetric; from this symmetry, Noether’s theorem shows the angular momentum of the
system must be conserved.
In the case of linear waves, Noether’s theorem states that every Killing fields gives rise
to a conservation law.
110
Chapter 9
Hyperbolic Equations
In this chapter we will study the properties of hyperbolic equations and emphasize the
importance of global hyperbolicity in connection to hyperbolic equations.
We shall first consider the case of the scalar wave equation.
9.1 The Energy Method

Let (M, g) be a Lorentzian manifold and ψ : M → R a scalar function. The energy momen-
tum tensor associated to ψ is defined to be the following symmetric (0, 2) tensor field
1
Q[ψ] = dψ ⊗ dψ − g −1 (dψ, dψ)g. (9.1)
2
The following proposition reveals the connection of Q and the wave equation.
Proposition 9.1.1. For any function ψ we have the following identity
Div Q[ψ] = (2g ψ) · dψ (9.2)
Proof. Let us first compute its divergence for arbitrary ψ. Recall that the covariant derivative
∇ satisfies
∇(A ⊗ B) = (∇A) ⊗ B + A ⊗ (∇B)
and that Div A = (∇µ A)µ . Therefore,
1
(∇a T )µν = ((∇a dψ) ⊗ dψ)µν + (dψ ⊗ (∇a dψ))µν − g −1 (dψ, dψ)(∇a g)µν
2
1 a −1

− ∇ (g (dψ, dψ)) gµν
2
= (∇a dψ)µ · (dψ)ν + (dψ)µ · (∇a dψ)ν − g −1 (∇a dψ, dψ) · gµν ,
where we used that ∇a g −1 = 0. Let now a = µ. Observe also that
g −1 (∇µ dψ, dψ) · gµν = (∇µ dψ)b · (dψ)b · gµν = (∇b dψ)µ · (dψ)b · gµν = (∇b dψ)ν · (dψ)b ,
where we used that the Hessian is symmetric (recall the Einstein’s notation for summation).
Therefore,
Div Q = (Div dψ) · dψ = (2g ψ) · dψ.
111
On top of equation (9.2), we also have the following
Proposition 9.1.2. Let V1 , V2 be two future directed timelike vector fields. Then the quantity
Q(V1 , V2 ) is positive definite in dψ, that is to say
3
X
Q(V1 , V2 ) ≥ c (∂α ψ)2 .
α=0
By continuity, if one of V1 , V2 is null then Q(V1 , V2 ) is non-negative definite in dψ.

Proof. Let x ∈ M and V1 , V2 ∈ Tx M be future-directed timelike vectors. Consider the
plane M = hV1 , V2 i, which is manifestly timelike. The orthogonal complement Π = hM i⊥ is
spacelike, and let e1 , e2 be an orthonormal frame of it. Let also L, L be future-directed null
vectors in M such that g(L, L) = −1. Then L, L, e1 , e2 form a frame for Tx M.
With respect to this frame, we have
|∇ψ|2 = −2(Lψ)(Lψ) + (e1 ψ)2 + (e2 ψ)2 .
Therefore,
Q(L, L) = (Lψ)2 , Q(L, L) = (Lψ)2 , Q(L, L) = (e1 ψ)2 + (e2 ψ)2 . (9.3)
The result now follows from the fact that Vi = ai L + bi L for some ai , bi > 0, i = 1, 2.
See Appendix 8 for more about the tensor Q.
The Vector Field Method
For investigating the evolution of linear waves we will use the so-called vector field
method. This is a geometric and robust method and involves mainly L2 estimates. The main
idea is to construct appropriate (0,1) currents and use the divergence identity in appropriate
regions. We usually consider currents Pµ that depend on the geometry of (M, g) and are
such that both Pµ and ∇µ Pµ depend only on the first order derivatives of ψ. Such currents
are in general called compatible currents. This can be achieved by using the wave equation to
make all second order derivatives disappear and end up with something that highly depends
on the geometry of the spacetime. There is a general method for producing such currents
using the energy momentum tensor Q.
Energy Currents
Since Q is a (0, 2) tensor we need to contract it with vector fields of M. It is here where
the geometry of M makes its appearance. We have the following definition
112
Definition 9.1.1. Given a vector field V we define the J V current by
JµV [ψ] = Qµν [ψ]V ν = Q[ψ](V, ·). (9.4)
Note that we will usually drop ψ and thus simply write Q instead of Q[ψ] etc. We say
that we use the vector field V as a multiplier 1 if we apply the divergence identity for the
current JµV . The divergence of this current is
Div(J) = Div (QV ) = Div (Q) V + Q (∇V ) , (9.5)

j j
where (∇V )ij = g ki ∇k V = ∇i V . If ψ is a solution of the wave equation then Div (Q) =
0 and, therefore, ∇µ JµV is an expression of the 1-jet of ψ.
Definition 9.1.2. Given a vector field V the scalar current K V is defined by

j 1
K V [ψ] = Q[ψ] (∇V ) = Qij [ψ] ∇i V = Qµν [ψ]πVµν , (9.6)
2
where πVµν = (LV g)µν = (∇µ V )ν + (∇ν V )µ is the deformation tensor of V .
The first term of the right hand side of (9.5) does not vanish when we commute the wave
equation with a vector field that is not Killing (or, more generally, when ψ does satisfy the
wave equation). Therefore, we also have the following definition.
Definition 9.1.3. Given a vector field V we define the scalar current E V by
E V [ψ] = (Div Q) V = (2g ψ) dψ (V ) = (2g ψ) (V ψ) . (9.7)
Note that Stokes’ theorem gives us:

Z Z Z
1 µν
(2g ψ) (V ψ) + Qµν [ψ]πV = Q(V, n∂R ),
R R 2 ∂R
where the direction of the normal to the boundary vector n∂R is determined by the following
figure (thought of as embedded in R1+1 )
9.2 A Priori Estimate

Consider two spacelike hypersurfaces Σ0 , Στ which enclose the spacetime region R foliated
by hypersurfaces Σt , where t is a time function.
1
the name comes from the fact that the tensor Q is multiplied by V.
113
We suppose that ψ satisfy the inhomogeneous equation
2g ψ = F.
Let also V be a timelike vector field and
Z Z X
f (t) = Q(V, nΣt ) ∼ (∂α ψ)2 .
Σt Σt α
Then, Z Z
f (τ ) + F · (V ψ) + K V [ψ] = f (0),
R R
where K V [ψ] is given by (9.6). Since V and nΣt are timelike we have
Z
V
K [ψ] ≤ Cf (t).
Σt
Recall now the coarea formula
Z τ Z Z Z
f dt = f |∇t| ∼ f.
0 Σt R R
By the energy identity and coarea formula, there exists a uniform constant C > 0 such that
for all solutions ψ to the inhomogeneous wave equation we have
Z τ Z τ
f (τ ) ≤ f (0) + F 2 dt + Cf (t)dt. (9.8)
0 0
The fundamental theorem of calculus gives us
Z τ
2 2 2 2
ψ (τ, ·) ≤ ψ (0, ·) + C ψ (t, ·) + (∂ψ) (t, ·) dt. (9.9)
0
Recall also the Grönwall inequality: If f, A, B are non-negative functions and A is non-
decreasing then
Z τ Rτ
f (τ ) ≤ A(τ ) + B(t)f (t)dt =⇒ f (τ ) ≤ e 0 B(t)dt A(τ ).
0
By adding (9.8) and (9.9) and using Grönwall inequality we then obtain the following esti-
mate:
kψk2H 1 (Στ ) ≤ eCτ kψk2H 1 (Σ0 ) + kF k2L2 (R) . (9.10)
Symmetries and Conservation Laws
If the spacetime admits a timelike Killing field V and apply it as a multiplier then, since
its deformation tensor identically vanishes, if ψ satisfies the inhomogeneous equation, then
the divergence of the current JµV [ψ] is simply equal to F · (V ψ). Then,
Z τZ Z τ
sup f (t) ≤ f (0) + |F · V ψ| ≤ f (0) + kF kL2 (Σt ) · kV ψkL2 (Σt )
t∈[0,τ ] 0 Σt 0
! Z
τ
≤ f (0) + sup kV ψkL2 (Σt ) · kF kL2 (Σt )
t∈[0,τ ] 0
! 2
Z τ
1
≤ f (0) + sup f (t) +C kF kL2 (Σt )
2 t∈[0,τ ] 0
114
Hence,
Z τ 2
sup f (t) ≤ f (0) + C kF kL2 (Σt ) . (9.11)
t∈[0,τ ] 0
Furthermore, we can easily obtain an improved estimate for ψ itself
1/2 2
Z Z Z τ
2 2
sup ψ ≤ ψ +C f (t) . (9.12)
t∈[0,τ ] Σt Σ0 0
Using (9.11) in (9.12) and added those two together we take

Z τ 2 !
sup kψk2H 1 (Σt ) ≤ Cτ kψk2H 1 (Σ0 ) + kF kL2 (Σt ) . (9.13)
t∈[0,τ ] 0
Note that (9.13) is stronger than (9.10).

Remark 9.2.1. Estimate (9.13) holds for perturbations of Minkowski (even though they do
not, in general, admit (timelike) Killing fields).
If now V is Killing and ψ satisfies the homogeneous wave equation (i.e. F = 0) then
Z Z
Q(V, nΣτ ) = Q(V, nΣ0 ). (9.14)
Στ Σ0
This statement corresponds to a conservation law. This is partly the content of a deep
theorem of Noether2 . This conservation law is particularly important if V is timelike since
in this case the associated integrand quantities are positive definite in dψ.
9.3 Well-posedness of the Wave Equation

Let Σ0 be a spacelike hypersurface in a Lorentzian manifold (M, g). We next study the
Cauchy problem to the wave equation
2g ψ = F,
ψ|Σ0 = f0 , nΣ0 ψ|Σ0 = f1 .
The regularity of F, f0 , f1 is always such that the right hand sides of our estimates is finite.
Uniqueness: finite speed propagation
By virtue of the global hyperbolicity of D+ (Σ0 ) there exists a time function t and hence
we can consider a spacelike foliation Σt of D+ (Σ0 ).
2
According to this theorem, any continuous family of isometries gives rise to a conservation law.
115
Then if we suppose that ψ1 , ψ2 satisfy the same inhomogeneous wave equation with same
initial data then ψ = ψ1 − ψ2 satisfies the linear wave equation with trivial data. Then,
(9.10) implies that kψkH 1 (Σt ) = 0 for all t and hence ψ = 0 in D+ (Σ0 ) which shows the
required uniqueness result.
In fact, we could have also considered the following foliation
and note that the null boundaries contribute non-negative definite terms in the energy
identity. Estimate (9.10) again shows that if the data are trivial on Σ0 , then the solution is
trivial in D+ (Σ0 ). In other words, non-trivial data on the dotted line outside the hypersurface
Σ0 take some time to affect the solutions to the wave equation (in particular, such data cannot
affect ψ in D+ (Σ0 )). This phenomenon is a characteristic feature of hyperbolic equations
and is called finite speed propagation (and also known as weak Huygen’s principle and domain
of dependence property).
Existence of solutions to the linear wave equation
We first prove the existence of solutions to the inhomogeneous wave equation with trivial
initial data. Indeed, a solution to the general problem can be obtained as follows: Let ψ̃
be any function with the same initial data as the given one and let ψ be the solution
2ψ = F − 2g ψ̃ with trivial data. Then, ψ + ψ̃ is the solution of the general problem.
Hence, we look for solutions to 2g ψ = F , with f0 = 0, f1 = 0.
Fix τ > 0 and consider the spacetime region R = ∪t∈[0,τ ] Σt . We will show that a solution
exists in R, and since this will be true for any τ > 0 the global existence in the interior of
D+ (Σ0 ) follows.
Since our estimates hold for specific functions spaces, namely Sobolev spaces, it follows
that the solutions will also have to exist in Sobolev spaces. For this reason, we first introduce
116
the notion of a weak solution. Let
Co (R) = φ ∈ C ∞ (R), φ|Στ = ∂φ|Στ = φ|Σ0 = ∂φ|Σ0 = 0 .

Then a weak solution to 2ψ = F , with trivial initial data, is a distribution ψ such that
Z
hψ, 2g φi = F ·φ (9.15)
R
for all φ ∈ Co (R). The motivation of this definition is of course the

R case where ψ is smooth
and thus (9.15) holds in the classical sense, where hψ, 2g φi = R ψ · 2g φ. We will next
show that if F is sufficiently regular, then ψ can in fact be represented by a function in an
appropriate functional space such that (9.15) makes sense in the following sense:
Z Z
ψ · 2g φ = F · φ,
R R
where both integrals make sense provided φ ∈ Co (R).
Remark 9.3.1. We need to introduce the important notation for functional spaces we will
be considering. If u ∈ X([0, τ ]; Y (Σt )), with t ∈ [0, τ ], then u|Σt belongs in the functional
space Y (Σt ) and the function ũ : [0, τ ] → Y lies in the spaceRX([0, τ ]), where ũ(t) = uΣt .
τ
For example, if u ∈ L1 ([0, τ ]; H 1 ), then kukL1 ([0,τ ];H 1 ) = 0 kukH 1 (Σt ) dt is finite. There-
fore, estimate (9.13) can also be written as

kψkL∞ ([0,τ ];H 1 ) ≤ Cτ kψkH 1 (Σ0 ) + kF kL1 ([0,τ ];H 1 ) .
Assume that F ∈ L2 (R). Then

Z Z Z Z τZ
2
|hψ, 2g φi| = F · φ ≤ 2 2
φ2

F φ ≤ C kF kL2 (R) ·
R R R 0 Σt
(9.10)
Z Z
2 2
≤ Cτ kF kL2 (R) (2g φ) = CF,τ (2g φ)2 ,
R R
where we used the a priori estimate (9.10) for the smooth function φ (note also that by
assumption the H 1 norm of φ on Σ0 vanishes). It is important to emphasize that the a
priori estimate allowed us to obtain the above estimate where the constant CF,τ , of course,
does not depend on φ. Hence, if we consider the linear subspace
S(R) = {2g φ : φ ∈ C0 (R)}
of the Banach space L2 (R) then we have that the distribution ψ is in fact a bounded linear
operator on S(R) (equipped with the L2 (R)-norm). Therefore, by Hahn-Banach we can
extend ψ to bounded linear functional on L2 (R). By Riesz representation theorem (or in
other words, by duality) the L2 (R) function ψ is in fact represented by a function ψ ∈ L2 (R)
so that Z Z
hψ, 2g φi = ψ · 2g φ = F · φ,
R R
for all φ ∈ Co (R).
117
Note that for the general problem with initial data, by the trace theorem, it suffices to
take f0 ∈ H 3/2 (Σ0 ), f1 ∈ H 1/2 (Σ0 ). In fact, one can show that if k ≥ 0 and
f0 ∈ H k+1 (Σ0 ), f1 ∈ H k (Σ0 ), F ∈ H k (R)
then there exists a unique solution
ψ ∈ H k+1 (R) and ψ ∈ H k+1 (Στ ), ψ ∈ H k (Στ ).
This can be shown by generalizing (9.10) for all k ≥ 0 and using Cauchy-Kowalewsky
theorem. Indeed, one can approximate (in the appropriate Sobolev norms) the differential
operators and the initial data by analytic functions and solve in the analytic class the new
equations. By the a priori estimates we have that the H k+1 (R) norm of these analytic
functions are uniformly bounded and hence the sequence converges weakly in H k+1 (R).
This limit is then the solution to the original problem.
Remark 9.3.2. We can obtain slightly stronger results if (M, g) admits a timelike Killing
field V and F ∈ L1 ([0, τ ]; L2 (Σt )). Indeed, letting φ ∈ Co (R) we obtain
Z Z τZ Z τ
|hψ, 2g φi| = F · φ ≤

|F · φ| ≤ kF kL2 (Σt ) · kφkL2 (Σt )
R 0 Σt 0
Zτ Z T
≤ sup kφkL2 (Σt ) · kF kL2 (Σt ) = CF,τ sup kφkL2 (Σt ) ≤ CF,τ k2g φkL2 (Σt )
t∈[0,τ ] 0 t∈[0,τ ] 0
= CF,τ k2g φkL1 ([0,τ ];L2 ) .
Clearly, ψ is a bounded functional on Co (R) with respect to the L1 ([0, τ ]; L2 )-norm and hence
by Hahn-Banach extends to a bounded functional of L1 ([0, τ ]; L2 ). Hence, by duality ψ is in
fact represented by a function ψ ∈ L∞ ([0, τ ]; L2 ) such that
Z Z
hψ, 2g φi = ψ · 2g φ = F · φ.
R R
In fact, we can improve the last inequality used above such that k2g ψkL2 (Σt ) is replaced
by k2g ψkH −1 (Σt ) . This then implies that ψ ∈ L∞ ([0, τ ]; H 1 ). Assuming that F is smooth,
one can show that
ψ ∈ C([0, τ ]; H k ) ∩ C 1 ([0, τ ]; H k−1 )
for all k ∈ N. To remove the smoothness assumption on F , we choose a sequence of smooth
1 l
functions Fm such that Fm → F in L [0, τ ]; H for a fixed l. For each of these Fm we
obtain a solution ψm with trivial initial data, as above, which can be easily seen to form a
Cauchy sequence in the Banach space C([0, τ ]; H l ) ∩ C 1 ([0, τ ]; H l−1 ). Hence the limit of this
sequence solves the equation in C([0, τ ]; H l ) ∩ C 1 ([0, τ ]; H l−1 ).
9.4 The Wave Equation on Minkowski spacetime

Pointwise and energy boundedness
Consider the Minkowski spacetime R3+1 . Let ψ solve the homogeneous wave equation
2g ψ = 0 with initial data (f0 , f1 ) prescribed on a Cauchy hypersurface Σ0 (for example, we
can take Σ0 = {t = 0}). Clearly, the vector field
T = ∂t
118
is globally timelike and Killing and hence its application as a multiplier gives us the following
By applying the divergence identity in the region R = ∪0≤t≤τ Σt and using the fact that
Στ is spacelike for all τ (and hence the normal nΣτ is timelike), we obtain a uniform constant
C > 0 such that for all solutions ψ to the wave equation we have
Z X Z X
(∂α ψ)2 ≤ C (∂α ψ)2 . (9.16)
Στ a Σ0 a
Vector fields can be applied as multipliers but also as commutators. In other words, since
∂t is Killing, we have that ∂t ψ also satisfies the wave equation. In fact, we can do the same
for all translation vector fields ∂a , α = 0, 1, 2, 3. Repeating this, we can in fact bound the
following higher order quantity:
Z X Z X
k 2
(∂α ψ) ≤ C (∂αk ψ)2 , (9.17)
Στ a,k Σ0 a,k
with k ≥ 1.
It remains to bound kψkL2 (Στ ) . Estimate (9.10) (or even (9.13)) is not at all satisfactory
since it does not provide a uniform in τ bound. Consider spherical coordinates (t, r, θ, φ).
Then for Στ = {t = τ } we obtain the 1-dimensional identity
Z ∞ Z ∞
2 2 r=∞
ψ dr = [rψ ]r=0 − 2 rψ(∂r ψ)dr
0 0
Z ∞ Z ∞
1
≤ ψ 2 dr + (∂r ψ)2 r2 dr,
0 0
where we assumed that rψ 2 → 0 at infinity (recalling the finite speed propagation, this is
true if, for example, the initial data are compactly supported). Hence, there exists a uniform
constant C > 0 such that
Z ∞ Z ∞
1 2 2
2
ψ r dr ≤ C (∂r ψ)2 r2 dr.
0 r 0
Integrating over S2 we obtain

Z Z Z X
1 2 2
2
ψ ≤C (∂r ψ) ≤ C (∂α ψ)2 . (9.18)
Στ r Στ Σ0 a
119
Although we can only bound a weight L2 norm of ψ, this bound suffices to apply Sobolev
inequality and thus obtain uniform pointwise bounds.
Pointwise and energy decay
Let Z Z X
f (t) = Q(V, nΣt ) ∼ (∂α ψ)2 .
Σt Σt α
In order to prove that f (τ ) decays as τ → +∞ one would ideally show that there exists a
uniform constant C > 0 such that
Z τ2
f (t)dt ≤ Cf (τ1 ).
τ1
In view of the boundedness of energy, this would then imply that f (τ ) decays faster than any
polynomial rate. It turns out, however, that such an estimate does not hold in Minkowski.
Indeed, as we shall see, the spacetime integral of the energy can only be bounded by quan-
tities with stronger weights (in other words, the left hand side above cannot just be f (τ1 )).
For this reason, we divide the problem into two separate ones. The first one has to do
with the region r ≤ R, for some R > 0, and the second one with r ≤ R.
Spacetime estimates in the spatially compact region are traditionally called Integrated
Local Energy Decay (ILED). Such an estimate in Minkowski goes back to Morawetz 1968.
Since we want to obtain bounds for a spacetime integral, we want to apply vector field
multipliers for which the current K[ψ] is non-negative definite. So, clearly the Killing field
T is not an option. If we use the vector field
X = f (r)∂r
(with respect to (t, r, θ, φ) coordinates) then

0 0
f0

f f f f
K X [ψ] = + (∂t ψ)2 + − (∂r ψ)2 − |∇/ ψ|2 ,
2 r 2 r 2
where K X is given by (9.6). Unfortunately, there is no f which makes all coefficients above
positive. For this reason we need to modify the above current. For we consider the current
JµX,h1 ,h2 ,w [ψ] = JµX + h1 (r) ψ∇µ ψ + h2 (r) ψ 2 (∇µ w) ,
where h1 , h2 , w are functions on M. Then we have
K̃ X = ∇µ JµX,h1 ,h2 ,w = K X + ∇µ (h1 ψ∇µ ψ) + ∇µ h2 ψ 2 (∇µ w)

= K X + h1 (∇a ψ∇a ψ)+ (∇µ h1 + 2h2 ∇µ w) ψ∇µ ψ + (∇µ h2 ∇µ w + h2 (2g w)) ψ 2 .
By taking h2 = 1, h1 = 2G, w = −G we make the coefficient of ψ∇µ ψ vanish. Therefore, let

us define
.
JµX,1 [ψ] = Q(X, ·) + 2Gψ (∇µ ψ) − (∇µ G) ψ 2 = JµX + 2Gψ (∇µ ψ) − (∇µ G) ψ 2 . (9.19)
Then

.
K X,1 = ∇µ JµX,1 = K X + 2G − (∂t ψ)2 + (∂r ψ)2 + |∇
/ ψ|2 − (2g G) ψ 2 .
120
Therefore, if we take G such that3
f0 f f0

f
−2G = − + ⇒G= + (9.20)
2 r 4 2r
then
f
K X,1 [ψ] = f 0 (∂r ψ)2 +/ ψ|2 − (2g G) ψ 2
|∇
r (9.21)
0 2 f 2 1 000 1 00
= f (∂r ψ) + |∇/ ψ| − f + f ψ2
r 4 r
If we take
1
JµX,1 ψ, f = 1, G=
2r
then K X,1 [ψ] = 1r |∇
/ ψ|2 . We will apply the energy identity in the region r ≥ , and we will
let → 0.
Since r = is timelike and its unit normal is n = −∂r we obtain

Z Z Z
X,1 X,1 µ
K [ψ] = Jµ [ψ]n + JµX,1 [ψ]nµ .
R∩{r≥} Σ0 ∪Στ r=
The boundary integral over Σ0 ∪ Στ can be estimated using (9.17) and (9.18). The integral
over r = is equal to
Z Z τZ 1 1
JµX,1 [ψ]nµ = − JrX [ψ] − ψ(∇r ψ) − 2 ψ 2 2 dgS2 dt
r= 0 S2 (θ,φ)
Therefore, as → 0, the only term that remains is negative, and hence has the correct sign.
Therefore, we obtain,
Z Z τ Z
1
|∇ 2
/ ψ| + 2
ψ (r = 0)dt ≤ C JµT [ψ]nµΣ0 . (9.22)
R r 0 Σ0
Morawetz’s original argument was that, although the above estimate bounds the angular
derivatives only, one can obtain exactly the same estimate by simply translating the origin
of the spheres. In order to bound all spatial derivatives, it suffices to add three estimates
for which the origins of the spheres are not colinear. This way we can retrieve the ∂r ψ
derivative in (9.22). This argument however heavily uses the symmetries of Minkowski in
3 f
We could have also chosen G = 2r
, however this choice is not very suitable for more general spacetimes.
121
a very non-robust way. It turns out that we can derive similar estimates for ∂r ψ using the
energy method. Indeed, if we consider the current
1
JµX,1 [ψ], f =−
r+1
then,
1 1
K X,1 [ψ] = 2
(∂r ψ)2 + γ(r)ψ 2 − / ψ|2 ,
|∇
(r + 1) r(r + 1)
where γ(r) > 0 and γ(r) ∼ r14 for large r. The angular derivatives have already been
bounded. Since G ∼ 1r for small r, the boundary integrals can be bounded as previously
noting again the factor r2 in the volume form of the integral over Στ (which is needed as
r → 0+ ). Hence, Z Z
1 2
(∂ ψ) ≤ C
2 r
JµT [ψ]nµΣ0 . (9.23)
R (r + 1) Σ0
Finally, in order to bound the derivative ∂t ψ we use the current
1
JµX [ψ], f=
r+1
Then,
(r + 2) 1
K X [ψ] = 2
(∂t ψ)2 + α1 (r)(∂r ψ)2 + / ψ|2
|∇
2r(r + 1) 2(r + 1)2
where α1 , ∼ r12 for large r and α1 ∼ 1r as r → 0+ . Hence, all the error terms and the the
boundary integrals can be bounded using the previous estimates and noting the r2 term in
the volume form.
Therefore, given R > 0 there exists a constant CR such that for all r ≤ R we have the
following integrated local energy decay estimate
Z τ2 Z Z Z
2 2 2
Q(T, nΣt ) ∼ / ψ| ≤ CR
(∂t ψ) + (∂r ψ) + |∇ Q(T, nΣτ1 ).
τ1 Σt ∩{r≤R} R∩{r≤R} Στ1
(9.24)
We now consider the region r ≥ R. Dafermos and Rodnianski have introduced a very
elegant and robust method to derive non-degenerate estimates in this region. In particular,
their method uses only r-weights and no t-weights. We assume that Στ coincides with t = τ
for r ≤ R and u = τR for r ≥ R, as depicted below:
122
Suppressing all angular variables we obtain the following diagram
Consider now Σ̃τ = ϕτ (Σ̃0 ), where ϕτ is the flow of T . For arbitrary τ1 < τ2 we define
R̃ττ21 = ∪τ ∈[τ1 ,τ2 ] Σ̃τ ,

D̃ττ12 = R̃ττ21 ∩ {r ≥ R} ,
Ñτ = Σ̃τ ∩ {r ≥ R} ,
∆ττ21 = R̃ττ21 ∩ {r = R} .
123
Proposition 9.4.1. (Dafermos–Rodnianski hierarchy) There exists a constant C > 0
such that for all solutions ψ to the wave equation we have
Z τ2 Z Z Z
Q(T, nNτ ) dτ ≤ C Q(T, nΣτ1 ) + C r−1 (∂v Ψ)2 ,
τ1 Nτ Στ1 Nτ1
Z τ2 Z Z Z
r−1 (∂v Ψ)2 dτ ≤ C Q(T, nΣτ1 ) + C (∂v Ψ)2 ,
τ1 Nτ Στ1 Nτ1
where Ψ = rψ and ∂v is considered with respect to the null system (u, v).
This hierarchy is an application of the following more general r-weighted estimate of
Dafermos and Rodnianski:
Proposition 9.4.2. Suppose p ∈ R, ψ is a solution to the wave equation and Ψ = rψ .
Then,
2 2
rp−1
Z Z Z
p (∂v Ψ) p−1 (∂v Ψ)
r 2
+ r p 2
+ / ψ|2
(2 − p) |∇
Nτ2 r τ2
Dτ1 r Dτ1
τ2 4
Z Z 2
(9.25)
(∂ v Ψ)
≤C Q(T, nΣτ1 ) + rp
Στ1 Nτ1 r2
Proof. We first consider the cut-off function ζ : [R, +∞) → [0, 1] such that
ζ(r) = 0 for all r ∈ [R, R + 1/2] ,
ζ(r) = 1 for all r ∈ [R + 1, +∞) .
Let q = p − 2. We consider the vector field
V = r q ∂v
which we apply as multiplier acting on the function ζΨ in the region Dττ12 . Then
Z Z
V V
τ
K [ζΨ] + E [ζΨ] = τ
Q[ζΦ](V, nµ ).
Dτ12 ∂Dτ12
Note that for r ≥ R + 1 we have K V [ζΨ] = K V [Ψ] and E V [ζΨ] = E V [Ψ]. Then,
K V [Ψ] = Qµν (∇µ (rq ∂v ))ν = Qµν ((∇µ rq ) ∂v )ν + Qµν rq (∇µ ∂v )ν

= rq K V + Qµv (∇µ rq )
rq−1
= 2rq−1 (∂u Ψ)(∂v Ψ) + qrq−1 (∂v Ψ)2 − q / Ψ|2 .
|∇
4
Note that since ψ solves the wave equation, Ψ satisfies
4∂u ∂v Ψ − 4
/Ψ = 0
and so
2
2g Ψ = − (∂u Ψ − ∂v Ψ) − 4(∂u ∂v Ψ) + 4
/Ψ
r
2
= − (∂u Ψ − ∂v Ψ),
r
124
which, as expected, depends only on the 1-jet of Ψ. Therefore,
E V (Ψ) = rq (∂v Ψ)(2g Ψ)
= −2rq−1 (∂u Ψ)(∂v Ψ) + 2rq−1 (∂v Ψ)2 .
Thus
rq−1
/ Ψ|2 .
K V [Ψ] + E V [Ψ] =(q + 2)rq−1 (∂v Ψ)2 − q
|∇
4
In view of the cut-off function ζ all the integrals over ∆ vanish. Clearly, all error terms that
arise in the region4 W = supp(ζ − 1) = {R ≤ r ≤ R + 1} are quadratic forms of the 1-jet of
ψ and,
Z therefore, in view of the local integrated energy decay, these integrals are bounded
by Q(T, nΣτ1 ). The boundary integrals are:
Στ1
Z Z Z Z
2 2 1
Q[ζΨ](V, n) = q
r (∂v ζΨ) − q
r (∂v ζΨ) − / Ψ|2 .
|∇
τ
∂Dτ12 Nτ1 Nτ2 I+ 4
The last two integrals on the right hand side appear with the right sign. The error terms
produced by the cut-off ζ in the region W are controlled by the flux of T through Στ1 .
The second estimate of Proposition 9.4.1 follows trivially by taking p = 2. Regarding

the first estimate, taking p = 1 implies
Z Z Z
1 2 1 2
2
(∂v Ψ) + 2 |∇
/ Ψ| ≤ C Q(T, nΣτ1 ) + C r−1 (∂v Ψ)2 . (9.26)
Dτ2 r r
τ
1
Στ1 Nτ1
Note now that |∇ / Ψ|2 = r2 |∇ / ψ|2 , (9.26) and

Z Z Z Z Z
1 2 1 2 1 2 1 2 1
τ2 r 2
(∂v ζΨ) = τ2 r 2
(∂ v ζψ) + τ2 2r 2
∂ v (rζψ ) = τ2 r 2
(∂v ζψ) + (ζψ)2 .
Dτ Dτ Dτ Dτ I + 4r
1 1 1 1
Again all the error terms can be bounded by the local integrated energy decay.
Remark 9.4.1. The reason we introduced the function Ψ is because the weight r that it
contains makes it non-degenerate (ψ = 0 on I + but Ψ does not vanish there in general).
The reason we have divided by r2 in (9.25) is because we want to emphasize the weight that
corresponds to ψ and not to Ψ.
Therefore, if Ψ = rψ and
Z Z Z
1
f (τ ) = Q(T, nΣτ ), h1 (τ ) = (∂v Ψ)2 , h2 (τ ) = (∂v Ψ)2
Στ Nτ r Nτ
and τ1 < τ2 then

f (τ1 ) ≤ f (τ2 ),
Z τ2
f (t)dt ≤ Cf (τ1 ) + Ch1 (τ1 ),
τ1
Z τ2
h1 (t)dt ≤ Cf (τ1 ) + h2 (τ1 ).
τ1
4
The weights in r play no role in this region.
125
The first two inequalities imply that
D
f (τ ) ≤ C ,
τ
where D depends on the initial data. The last inequality implies that there exists a dyadic
sequence5 τn such that
D
h1 (τn ) ≤ C ,
τn
where D depends again on the initial data. The second estimate then gives us
Z τn+1
D
(τn+1 − τn )f (τn+1 ) ≤ f (t)dt ≤ Cf (τn ) + Ch(τn ) ≤ C ,
τn τ
which, by the properties of the dyadic sequence (and the first inequality), implies
D
f (τ ) ≤ C .
τ2
Pointwise bounds: First note that for r0 ≥ R > 0 we have
Z +∞ 2 Z +∞ Z +∞ Z +∞
2 2 2 1 1 2 2
ψ (r0 , ω) = (∂ρ ψ) dρ ≤ (∂ρ ψ) ρ dρ dρ = (∂ρ ψ) ρ dρ ,
r0 r0 r0 ρ2 r0 r0
where ∂ρ is a derivative tangential to Στ . Therefore,

Z Z Z +∞ Z
1 C
ψ 2 (r0 , ω)dω ≤ (∂ρ ψ)2 ρ2 dρdω ≤ Q(N, nΣτ ), (9.27)
S2 r0 S2 r0 r0 Στ ∩{r≥r0 }
For the interior region r ≤ R we can commute with T and apply elliptic estimates. Hence,
commuting with either T and the angular operators and using Sobolev inequality gives us
D D
|ψ| ≤ C , |ψ| ≤ C .
r1/2 ·τ r · τ 1/2
5
A dyadic sequence τn has the property that (τn+1 − τn ) ∼ τn ∼ τn+1 .
126
Chapter 10
Wave Propagation on Black Holes
10.1 Introduction
Recall that the conformal diagram of a black hole looks like the following
black hole event

horizon
ll
nu nity
fi
in
10.2 Pointwise and Energy Boundedness

We will study the properties of the linear homogeneous wave equation on black hole back-
grounds. Although we shall be particularly interested in the Schwarzschild case, our analysis
will cover more general black holes for which the following hold:
1. The spacetime is asymptotically flat, that is to say the metric g approaches the
Minkowski metric for large r, where r is an appropriate radius coordinate.
2. There exists a Killing field T which is timelike outside the event horizon H+ and null
and normal on H+ .
3. The coordinate r is such that it is constant along the integral curves of T and such that
H+ = {r = rH+ }. The domain of outer communications then corresponds to r > rH+ .
127
4. The event horizon has positive surface gravity κ > 0. The surface gravity is defined
by the following relation
∇T T = κT.
Let Σ0 be a spacelike hypersurface which crosses the event horizon. We will study the
following Cauchy problem
2g ψ = 0,
ψ|Σ0 = f0 , T ψ|Σ0 = f1 .
We assume that f0 , f1 are sufficiently regular and we want to understand the evolution
of ψ in D+ (Σ0 ) up to and including the horizon H+ .
Basic a priori estimate
We define Στ = Fτ (Σ0 ), where Fτ is the flow of T .
128
Applying the Killing field T as a multiplier in region R(0, τ ) = ∪0≤t≤τ Σt and noting that
Q(T, nH+ ) ≥ 0, we obtain the following estimate
Z Z
Q(T, nΣτ ) ≤ Q(T, nΣ0 ). (10.1)
Στ Σ0
In view, however, of (9.3), we can see that
Q(T, nΣτ ) ∼ (T ψ)2 + (e1 ψ)2 + (e2 ψ)2 ∼ (T ψ)2 + |∇

/ ψ|2 ,
on the horizon H+ where T is null and tangential to it and nΣτ timelike. Note that ∇
/
denotes the gradient on the sphere Sr0 = Στ ∩ {r = r0 } .
This implies that (10.1) does not provide an estimate for all four derivatives of ψ up to
and including the horizon, but in fact, the left hand side of (10.1) degenerates with respect
to the transversal to H+ derivative (note that T, e1 , e2 are all tangential to H+ ). For this
reason we will refer to estimate (10.1) as the boundedness of the degenerate energy.
This however loss of derivatives is very bad since this degeneracy does not allow us to
apply Sobolev inequalities and obtain pointwise estimates up to and including H+ . One of
the fundamental difficulties of the study of the wave equation on black holes is precisely to
overcome this degeneracy.
The Schwarzschild case: Consider a spacelike hypersurface Σ0 which crosses H+ and
coincides with t = 0 for sufficiently large r. We define Στ = Fτ (Σ0 ), where Fτ is the flow of
T . Then,
2 2M
Q(T, nΣτ ) ∼ (T ψ) + 1 − (Y ψ)2 + |∇
/ ψ|2 .
r
On other hand, if N is a strictly timelike vector field then
Q(N, nΣτ ) ∼ (T ψ)2 + (Y ψ)2 + |∇

/ ψ|2 .
Boundedness of the non-degenerate energy
We will next present the argument of Dafermos and Rodnianski (2005) for the bounded-
ness of the non-degenerate energy.
It is clear from the above discussion that in order to obtain non-degenerate bounds we
need to apply a timelike vector field as multiplier; such a multiplier, however, will not have
vanishing spacetime term K N [ψ]. In this context, Dafermos and Rodnianski showed that
under assumptions 1–4 above there exists rH+ < r0 < r1 and a timelike Fτ −invariant vector
field N such that
129
1. K N [ψ] ∼ Q(N, nΣτ ), for rH+ ≤ r ≤ r0 .

2. K N [ψ] ≤ CQ(T, nΣτ ), for r0 ≤ r ≤ r1 .
3. N = T (and hence K N [ψ] = 0), for r ≥ r1 ,

where K N is given by (9.6).
Proof. For simplicity, we will prove this result only for metrics which take the form (7.1) for
a general D. Consider the following ansatz
N = N v (r)T + N r (r)Y
Then
K N [ψ] =Fvv (T ψ)2 + Frr (Y ψ)2 + F∇ / ψ|2 + Fvr (T ψ) (Y ψ) ,

/ |∇
where the coefficients are given by

(∂r N r ) N r N r D0 2N r

v 1 r v
Fvv = (∂r N ) , Frr = D − − , F∇
/ = − (∂ r N ) , Fvr = D (∂ r N ) − ,
2 r 2 2 r
where D0 = dD v 2 v r v r
dr . Note that since g (N, N ) = −D (N ) + 2N N , g (N, T ) = −DN + N ,
r
and so N (r = M ) can not be zero (otherwise the vector field N would not be timelike).
The crucial observation is the following: The surface gravity is given by κ = 12 D0 and,
therefore, by assumption we must have D0 > 0.
We can thus take for r = rH+ (which is a root of D) N v , ∂r N v to be very large, N r
negative and ∂r N r negative. Then by Cauchy–Schwarz inequality we obtain the required
positivity of K N [ψ] on H+ and thus, by stability considerations, on a neighborhood of H+
where rH+ ≤ r ≤ r0 . We can then smoothly extend N such that N = T for r ≥ r1 > r0 .
We now have all tools in place in order to prove the boundedness of the non-degenerate
energy. Writing the energy identity for JµN [ψ] in region R(τ 0 , τ ) we obtain
Z Z Z Z
Q(N, nΣτ ) + Q(N, nH+ ) + K N [ψ] = Q(N, nΣτ 0 ),
Στ H+ R Στ 0
where K N [ψ] = Div Q(N, ·) . Since

Z Z Z
N N
K [ψ] = K [ψ] + K N [ψ]
R r≤r0 r0 ≤r≤r1
130
and since the flux on H+ is non-negative definite we obtain
Z Z Z Z
N N
Q(N, nΣτ ) + K [ψ] ≤ −K [ψ] + Q(N, nΣτ 0 )
Στ r≤r0 r0 ≤r≤r1 Στ 0
Z Z
≤C Q(T, nΣ ) + Q(N, nΣτ 0 ).
r0 ≤r≤r1 Στ 0
R R
We can now add a big multiple of r≥r0 Q(T, nΣ ). If f (τ ) = Στ Q(N, nΣτ ), in view of the
properties of N , the coarea formula and the boundedness of the degenerate energy we obtain
Z τ
f (τ ) + b f (t)dt ≤ C(τ − τ 0 )f (0) + f (τ 0 ).
τ0
Rearranging terms, dividing by (τ − τ 0 ) and taking the limit as τ 0 → τ tends
f 0 (τ ) + bf (τ ) ≤ Cf (0).
This implies
C 0
f (τ )ebτ − f (0)ebτ ≤ 0
b
and so
f (τ ) ≤ Bf (0)
for some uniform constant B > 0.
Remark 10.2.1. The positivity of the surface gravity is intimately related to the so-called
redshift effect that takes place along the horizon H+ . For this reason, the vector field N is
sometimes referred to as the redshift vector field.
Remark 10.2.2. As in the Minkowksi case, the vector field N can be applied as a multiplier
as well as a commutator to give us higher order bounds.
Remark 10.2.3. For a special class of black holes, the surface gravity κ vanishes. Such
black holes are called extremal. In these cases, the above construction for N does not work.
Relevant results for such black holes have been obtained by Aretakis.
Remark 10.2.4. Although for the general subextremal family of axisymmetric Kerr black
holes (see also Section 10.3) the surface gravity is positive and so the above construction
for N applies, the Killing field T fails to be timelike (in fact causal) everywhere outside the
event horizon H+ . This means that in Kerr one does not have the trivial estimate (10.1).
Of course, this implies that the above proof for the boundedness of the non-degenerate energy
fails. It turns out that for such backgrounds, in order to show any boundedness estimate one
must in fact show something stronger, i.e. a decay estimate.
10.3 Pointwise and Energy Decay

The discussion about energy decay on Minkowski in Section 9.4 suggests that one needs to
prove a local integrated energy decay and also show an appropriate hierarchy of estimates in
a neighborhood of infinity. It turns out, that the Dafermos–Rodnianski hierarchy still holds
for Schwarzschild (and in fact more general) backgrounds. Hence, the problem is reduced in
131
obtaining estimate in the spatially compact region {rH+ ≤ r ≤ R}, for (arbitrary but fixed)
R > rH+ . In fact, in view of the redshift construction presented in Section 10.2, it suffices
to show a spacetime estimate in region B depicted below
In order to obtain spacetime estimates in region B we need to know specific properties

of the metric, and in particular, of the geodesic flow. Hence, the general assumptions 1–4
of Section 10.2 which lead to boundedness do not suffice to yield decay too (in fact, one
does not expect to obtain decay using only assumptions). For this reason, we restrict our
attention to Schwarzschild and Kerr (see below) spacetimes.
Schwarzschild
Let us first consider Schwarzschild backgrounds. As in Minkowski spacetime, one uses
appropriately modified energy currents with multiplier ∂r . In fact, it turns out that the
multiplier
X = ∂r ∗ ,
with respect to (t, r∗ , θ, φ) is more suitable. The constructions of the multipliers (first pre-
sented by Dafermos and Rodnianski), although they are similar to the constructions for
Minkowski, arise several technical difficulties. For example, one needs to decompose ψ in
angular frequencies; that is to say, one needs to consider the Fourier series of ψ on the
spheres of symmetry and derive uniform estimates for every projection separately. However,
the most striking new feature of these constructions comes from the non-trivial dynamics of
the geodesic flow. The most relevant property of the geodesic flow of Schwarzschild is that
the three-dimensional timelike hypersurface r = 3M is spanned by null geodesics, which or-
bit the black hole region (and hence are not orthogonal to the spheres of symmetry). Indeed,
the (continuity) principle that applies to satellites orbiting the earth
EARTH
132
can be extended for Schwarzschild and thus obtain (trapped) null geodesics which neither
scatter to null infinity (where r = +∞) nor cross the event horizon H+ .
BLACK HOLE
The hypersurface r = 3M is known as the photon sphere.

Returning now to the integrated local energy decay estimate, the presence of the photon
sphere implies that the energy gets concentrated for long time around the photon sphere.
This implies that an estimate in region B must either degenerate on the photon sphere
or lose derivatives. Specifically, one can only show the following
Z Z
(∂r∗ ψ)2 + ψ 2 + (r − 3M )2 (T ψ)2 + |∇
/ ψ|2 ≤ Q[ψ](T, nΣ0 ).
B Σ0
Note that it is in fact the tangential to the photon sphere derivatives that degenerate above.
Hence, commuting with the Killing T (and thus replacing ψ above by T ψ) we obtain
Z X Z
2 2
(∂α ψ) + ψ ≤ C Q[ψ](T, nΣ0 ) + Q[T ψ](T, nΣ0 ).
B a Σ0
This loss of derivatives is known as the trapping effect. One can now obtain the pointwise
and energy decay results (which also lose one more derivative compared to Minkowski).
Remark 10.3.1. The redshift effect, as already mentioned, degenerates for extremal black
holes (κ = 0). For this reason, one needs to derive separate estimates in region A. Work
of Aretakis has shown that, in the extremal case, the trapping effect takes place along the
horizon H+ too.
Kerr
The Kerr metric admits less Killing fields and this makes the proof of estimates on
such backgrounds more elaborate. Dafermos and Rodnianski (2010) have proved decay for
133
Kerr for all |a| < M . We next describe the main features that arise from passing from
Schwarzschild (M > 0, |a| = 0) to Kerr (M > 0, M > |a| ≥ 0) and also sketch the main
points of the approach of Dafermos and Rodnianski.
Ergoregion and superradiance
6 0 then the Killing field T = ∂t = ∂v is spacelike in a region E outside H+ known

If |a| =
as ergoregion. In the context of 2g ψ = 0, this means that the flux Q[ψ](T, , nΣτ ) fails to be
non-negative definite. Hence, we can have
Z Z
Q(T, nI + ) > Q(T, nΣ0 ).
I+ Σ0
This phenomenon is called superradiance. Note that H+ ⊂ E and supE r = 2M . As we have

already mentioned, the lack of the a priori degenerate (at H+ ) estimate for the T -flux implies
that one needs to show decay directly. One can prove decay for Minkowski and Schwarzschild
by constructing currents with positive definite divergence and using the conservation of the
T -flux to bound the future boundary terms. Hence, in Kerr one needs to find an alternative
approach in order to bound the boundary terms. This is in fact done using a bootstrap
argument (see below).
Remark 10.3.2. Superradiance is effectively absent if ψ is assumed to be axisymmetric.

Indeed, if Φψ = 0 then
JµT [ψ]nµΣτ ∼ (T ψ)2 + D(r)(Y ψ)2 + |∇

/ ψ|2
for r ≤ R where the constants in ∼ depend on M and R and D degenerates only on H+ .
Trapping
In Schwarzschild, all trapped null geodesics either live on r = 3M or approach this

hypersurface. In Kerr, on the other hand, there are trapped null geodesics for an open range
of the radius coordinate r and therefore they can only be understood in phase space (and
not just in physical space). This implies that the trapping effect, which must be taken into
account for any dispersive estimate, is much more complicated than in the Schwarzschild
case.
Remark 10.3.3. If ψ is axisymmetric (Φψ = 0), then it only “sees” the trapped null
geodesics which are orthogonal to the vector field Φ. The behavior of the latter geodesics is
much more favorable since they all approach a unique r = ra,M hypersurface.
The bootstrap
The desired estimates close in a bootstrap setting which makes use of the smooth depen-
dence of the Kerr metric on the angular momentum a. Specifically, given any 0 < a0 < M
and R sufficiently large, one considers the set S ⊂ [0, a0 ] such that |a| ∈ S if there exists a
134
constant C = C(a) which depends on M, a such that for all solutions ψ to the wave equation
we have
Z X Z X
(∂α ψ)2 ≤ C (∂α ψ)2 ,
Στ α Σ0 α
Z X Z X
(χ∂α ψ)2 + ψ 2 ≤ C (∂α ψ)2 .
R∩{r≤R} α Σ0 α
for all τ ≥ 0, where χ = χ(r) ≥ 0 and vanishes only in the physical space projection of the
trapped set.
Since [0, a0 ] is a connected set, it suffices to show that S is non-empty, open and closed
in [0, a0 ]. The non-emptiness follows from the Schwarzschild result (0 ∈ S).
In order to show that S is open, it suffices to show that if a ∈ S then (a − , a + ) ⊂ S
for some > 0. Then, if a ∈ S, the smooth dependence on the angular momentum of the
metric allows us to use Cauchy stability considerations. We note that these considerations
have to be localized away from the trapped set where the above estimate degenerates. In
the openness part of the proof, we expect to allow C(a + e) → +∞ as e → .
In order to show that S is closed in [0, a0 ], it suffices to prove that if (a − , a + ) ⊂ S
then a − , a + ∈ S. The way to show this is the following: Since (a − , a + ) ⊂ S we have
that C(ã) < ∞ for ã ∈ (a − , a + ). It, therefore, suffices to improve the boundedness of
the constants C(ã), that is to say, it suffices to prove that
C(ã) < ∞ ∀ã ∈ (a − , a + ) ⇒ C(ã) < C(a0 ) ∀ã ∈ (a − , a + ).
Appealing again to the smooth dependence of the metric on the angular momentum, we
obtain that C(a − ) < C(a0 ) and C(a + ) < C(a0 ).
Frequency localization
As we mentioned above, one needs to understand the the properties of the geodesic
flow, and in particular, of the trapped null geodesics. In Schwarzschild, on top of the
stationary Killing field T , one has a complete set of spherical Killing fields traditionally
denoted Ω1 , Ω2 , Ω3 . These Killing fields provide enough conserved quantities to deduce that
the Hamilton–Jacobi equations separate. In Kerr, although the only Killing vector fields
are T and Φ, the Hamilton–Jacobi equations still separate in view of a third non-trivial
conserved quantity discovered by Carter. Penrose and Walker showed that the complete
integrability of the geodesic flow has its fundamental origin in the existence of an irreducible
Killing tensor.
A Killing 2-tensor is a symmetric 2-tensor K which satisfies ∇(α K βγ) = 0. For example,
the metric is always a Killing tensor. A Killing tensor will be called irreducible if it can not
be constructed from the metric and other Killing vector fields.
A Killing tensor K yields a conserved quantity for geodesics. Indeed, if γ is a geodesic
.α.β
then Kαβ γ γ is a constant of the motion. It turns out that in Ricci flat spacetimes the
separability of Hamilton–Jacobi, the separability of the wave equation and the existence
of an irreducible Killing tensor are equivalent. Moverover, a Killing tensor gives rise to a
symmetry operator K = ∇α K αβ ∇β with the property [K, 2g ] = 0. In the Kerr spacetime

the symmetry operator K associated to Carter’s irreducible Killing 2-tensor takes the form
/ S2 ψ − Φ2 ψ + a2 sin2 θ T 2 ψ.

Kψ = 4 (10.2)
135
One of the main insights of the approach of Dafermos and Rodnianski is the use of the
separability of the wave equation (in (t, r, θ, φ) coordinates) as a way to frequency (micro-
)localize the energy currents. This now allows us to construct multipliers for each frequency
separately and thus capture trapping.
We have the following frequency decomposition of solutions ψ to the wave equation in
Kerr:
Fourier Expansion
Z zX }| {
1 (aω) −iωt
ψ(t, r, θ, φ) = √ Rω,m,` (r) · Sm` (θ, φ) · e dω.
2π ω∈R | {z
m,`
}
Oblate Spheroidal Expansion
(aω)
The functions Sm` (θ, φ), m, ` ∈ Z, l ≥ |m|, are the eigenfunctions of the elliptic operator
/ S2 − (aω)2 cos2 θ.
P(aω) = −4
(aω)
If λm` are the corresponding eigenvalues, then we define the following angular frequency
(aω) (aω)
Λm` = λm` + (aω)2 ≥ 0.
(aω) p
If we also define um` (r) = (r2 + a2 ) · Rω,m,` (r) then by suppressing the indices, the
∗
function u = u(r ) satisfies the Carter’s equation
u00 + (ω 2 − V (r))u = 0,
where V is a potential function.

In view of Parseval’s identity we have
Z +∞ Z Z +∞ X
(aω) 2

ψ 2 r2 dgS2 dt ∼ u
ml (r) dω,
−∞ S2 −∞ m,l
Z Z +∞ X
(T ψ)2 ∼ ω 2 |u(r = c)|2 dω,
r=c −∞ m,l
Z Z +∞ X
(aω)
/ ψ|2 ∼
|∇ Λm` |u(r = c)|2 dω,
r=c −∞ m,l
" 2 #
Z Z +∞ X d
2 2

(∂r∗ ψ) ∼ dr∗ u(r = c) + |u(r = c)| dω.

r=c −∞ m,l
Hence, in order to derive spacetime estimates, it suffices to construct microlocal currents

(aω)
Jm` [u] such that for rH+ < r0 < R0 there exists a uniform in frequencies constant b such
that Z R0 h i
|u|2 + |u0 |2 + Λ + ω 2 |u|2 dr∗ ≤ J[u](r = R0 ) − J[u](r = r0 )

b
r0
for all frequencies (ω, Λ), where Λ is an angular frequency.

Major technical difficulty: This frequency decomposition of ψ requires taking the Fourier
transform in time which in turn requires knowing some decay in time. Since this is not known
a priori when proving the openness part of the bootstrap, we need to cut off in time as follows:
Let ξτ be a cut-off function such that ξτ (τ̃ ) = 0 for τ̃ ≤ 0 and τ̃ ≥ τ and ξτ (τ̃ ) = 1 for
1 ≤ τ̃ ≤ τ − 1. Then the support of ∇ξτ is the region Sξ = {0 ≤ τ̃ ≤ 1} ∪ {τ − 1 ≤ τ̃ ≤ τ }:
136
1
Let ψQ = ξτ ψ be the compactly supported in time function which arises from the cut-off
ξτ multiplied to the solution ψ of the wave equation. This cut-off version of ψ then satisfies
the following inhomogeneous wave equation
2g ψQ = F,
where
F = 2∇µ ξτ ∇µ ψ + (2ξτ ) ψ. (10.3)
It turns out that one can still frequency localize ψQ and thus try to apply the above frame-
work. The only difference is that the cut-off creates error terms that need to be estimated.
For closeness, since we have established estimates from the openness part of the proof, we
can remove the future cut-off and thus try to prove quantitative estimates with precise con-
trol on the constants C(ã). For details of the constructions see the work of Dafermos and
Rodnianski.
137
Index
trace, 59
metric trace, 60
138

Lecture Notes On General Relativity Columbia University: January 16, 2013

Uploaded by

Copyright:

Available Formats

Lecture Notes On General Relativity Columbia University: January 16, 2013

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lecture Notes On General Relativity Columbia University: January 16, 2013

Uploaded by

Copyright:

Available Formats

Lecture Notes on General Relativity

January 16, 2013

3 Introduction to General Relativity 42

4 Null Structure Equations 49

6 Christodoulou’s Memory Effect 90

8 Lagrangian Theories and the Variational Principle 104

9 Hyperbolic Equations 111

10 Wave Propagation on Black Holes 127

1.1 Newtonian Physics

The primary assumptions in Newtonian physics are the following

2. The speed of light is finite and observer dependent.

• Future of p: Set of all events for which t > tp .

• Present of p: Set of all events for which t = tp .

• Past of p: Set of all events for which t < tp .

Figure 1.1: The Newtonian universe and its causal structure

Figure 1.2: The Newtonian universe and light trajectories

It is important to emphasize that in Newtonian theory, in view of the existence of the

1. A more philosophical issue is that in Newtonian theory an observer is either at rest

1.2 The Birth of Special Relativity

A Minkowski metric g on the linear space R4 is a symmetric non-degenerate bilinear

g(eα , eβ ) = gαβ , α, β ∈ {0, 1, 2, 3} ,

where the matrix gαβ is given by

g = −dt2 + (dx1 )2 + (dx2 )2 + (dx3 )2 . (1.1)

1.3.1 Causality Theory

3. timelike, if g(X, X) < 0.

Sp = X ∈ Tp R3+1 : g(X, X) > 0 ,

the set of timelike vectors at p, and by Np ,

Causality, observers and particles, proper time

Finally, we have the following categories for hypersurfaces:

1. A hypesurface H is called spacelike, if the normal Nx at each point x ∈ H is timelike.

3. A hypesurface H is called timelike, if the normal Nx at each point x ∈ H is spacelike.

Examples of spacelike hypersurfaces

H 3 = {X : g(X, X) = −1 and X future-directed }

can easily verify that X is the normal to H 3 at

Hence, there exists a real number φ such that

Then φ is called the hyperbolic angle of X and Y .

Examples of timelike hypersurfaces

Examples of null hypersurfaces

are null hypersurfaces, since their normal is the null vector n.

2. The (null) cone

The above can be summarized in the following figure:

1.3.2 Inertial Observers, Frames of Reference and Isometies

Isometries: Lorentz transformations

Clearly, one can define boosts in any direction.

1.3.3 General and Special Covariance

2. Absoluteness of speed of light

If a particle moves of a null curve α0 (τ ) = (t(τ )), x1 (τ ), x2 , (τ ), x3 (τ )) then the speed of

• The energy-momentum of p is the vector P = mU .

which is called the momentum of p as measured by O. (Compare this definition with

1.4.1 The Double Null Foliation

whereas the past-directed null cone C O with vertex at O = α(0) is given by

g = −dt2 + dr2 + r2 · gS2 (θ,φ) ,

whereas the past-directed null cone C τ with vertex at α(τ ) is given by

Note also that

As expected, a consequence of the boundedness of the range of p, q is that the factor

f g̃), which is conformal to Minkowski R3+1 .

• Future timelike infinity i+ to be the endpoints of all future-directed timelike geodesics.

Figure 1.5: Identification of the boundary of the conformal diagram of Minkowski

• t → +∞ and r < ∞, then α approaches i+ .