Lesson 3 : Flatness and curvature

Notes from Prof. Susskind video lectures publicly available

on YouTube

1
 Introduction

General relativity has a reputation for being very difficult.

I think the reason is that it is very difficult. It is calculation-
intensive : symbols, indices, awe-inspiring equations. There
are ways people have invented to express things in more
condensed very neat notations. But just learning them in
itself is a task that that would take some time. Things like
deerbines ( ?), forms, spinors and twistors and all sorts of
other mathematical objects. You could call many of them
just notational devices if you like. And they do simplify the
equations.
So I sometimes feel that in presenting these things the
way I do, it is sort of like Maxwell who wrote down every
single equation of his set of Maxwell equations. At first he
wrote twenty altogether. Although now we write only four.
We don’t usually write all components of the equations.
We put them together into vector notation and so forth.
And if we are smart, we can even avoid the indices by in-
venting symbols like del, curl, laplacian, etc. The same thing
could be done to some extent for general relativity. But as
I said the computational techniques are harder.

I am by no means an expert in general relativity. I know

the subject, how to use it. But when I have to do something
involving general relativity, I go back to the indices. I can
follow if somebody else uses the fancy methods, but I can’t
get around to doing them myself. It is not that I don’t know
how to use them. But they are not natural to me. I wind up
going back to the same things we are doing here : indices,
derivatives, the whole works.
My feeling is that it is the right way to see it for the
first time. But it is a nuisance. It has indices all over the
place, notations all over the place, too much of it. We can’t

2
help it. So some aspect of it you have to get over. You
have to get over that hump of tensor notation, bunches of
indices upstairs and downstairs. And then it goes easier,
more smoothly... until you come to do a calculation. Then
you are right back into the mess again.

Now if you are really interested in doing the computations

in general relativity, there are packages. You just put in the
metric as a function of position, the metric tensor. And out
it will spit various tensors, Riemann tensors, this kind of
tensor, that kind of tensor, Einstein tensors. And then you
can say, without even looking at the results : okay, please
Mr. Computer, set the Einstein tensor equal to the energy
momentum tensor and tell me what comes out. So, yes,
computers can do a lot better than us.

As I said, general relativity is hard. And part of the diffi-

culty is just the complexity of the notations that are neces-
sary. By the way, the subject that we have been studying
up to now is Riemannian geometry. It is the geometry of
smooth curved metric spaces. They can have highly non-flat
shapes, but locally they are like planes or higher dimension
flat spaces, and their local geometry is like Euclidean geo-
metry.
I suppose we have Riemann to blame for the complexity
of the notation. Although, actually he simplified it. Einstein
simplified it further. But nevertheless it is notationally and
computationally intensive.
The basic concepts are really not so hard. But we have
to go through the mess before we get to the concepts or
before we can express those concepts.

3
 General relativity in modern physics

One of the most active areas of research among theoretical

physicists, at the beginning of the XXIst century, is the
quantization of general relativity, the quantization of black
hole physics. The other most active area is quantum field
theory.
There are also astrophysical areas where black holes are
very important. Astrophysical areas are entirely classical
general relativity.
There is numerical relativity. It means actual compu-
tations with numbers. This involves special mathematical
techniques employed with computers. It is usually used for
the purpose of solving astrophysical problems : putting in
numerically various initial situations, for instance a pair of
black holes orbiting around each other, and simply running
the computer to find out what happens, by solving Ein-
stein’s equations.
It is a very subtle business by the way. You don’t just
put in the equations and leave the computer to grind. It is
much more subtle than that. Numerical relativity is also a
very hard subject. The purpose for instance can be to find
out how much gravitational radiations are emitted when
two black holes fall into each other. These are entirely clas-
sical radiations. Bang ! A lot of gravitational radiation is
produced, gravitational waves.
Why are we interested in them ? Among other reasons,
because with the latest detectors we have a chance of de-
tecting them. Anything you can detect is a window onto a
new natural phenomenon. So gravitational wave detectors
may be able to detect gravitational waves from the colli-
sion of two black holes. It is estimated that there are every
year a few collisions of black holes that can be seen by their
gravitational radiation.

4
 So there is a lot of different active areas nowadays rela-
ted to general relativity – classical general relativity, that is
involving no quantum uncertainty. The equations of general
relativity are still full of things to teach us.
Furthermore the equations of black holes have been
found to be extremely useful in condensed matter physics
and fluid dynamics, in other words in domains having no-
thing to do with black holes and gravity. Different domains,
same equations. In short, there is a heavy amount of acti-
vity devoted to applying the equations of general relativity
to other areas of physics. It is probably one of the most
active area of research.

This preeminence of general relativity has not always been

the case. Over the career of Professor Susskind in physics
it went from being real backwaters in physics to the fore-
front. Nobody studied general relativity very much in the
1960s when he was a student, except people who were really
interested in the subject per se. They were then focused spe-
cifically on questions of general relativity. But the bulk of
theoretical physicists were not interested.

Then it started to change at the end of the XXth century.

It became interesting in astrophysics and eventually took
over questions of fundamental theoretical physics. Why ?
Because of a clash of gravity in general relativity 1 . Anyway,
for whatever reason, it is probably now more important
than the day-to-day workings of theoretical physicists in
almost any other subject – that and quantum field theory,
and the connection between them.

1. The equations lead to the hypothesis of black matter (see vo-

lume 5 of the collection The Theoretical Minimum, on Cosmology) just
like problems related to light lead to special and general relativity.

5
 Riemannian geometry

This is the last chapter in which we will be studying Rie-

mannian geometry as such, without really discussing gra-
vity. In the next chapter we will really get into gravity.
What do all these manipulations of tensors have to do with
gravity ? We saw it a little bit in the first lesson.

The problem of finding out whether there is a real gravita-

tional field, as opposed to just some funny coordinates with
which we are locating events in space-time, leading to cur-
ved lines but not intrinsically curved trajectories (see figure
4 of chapter 1), is mathematically identical to the problem
of finding out if a certain geometry – characterized by its
metric tensor – is flat or not.

Let’s think of a two dimensional space or variety to begin

with. That means a surface S of some sort, where each
point is located with two real coordinates X 1 and X 2 . See
figure 1, where it is shown embedded in the usual 3D space
for convenience :

Figure 1 : Two dimensional variety

with a system of coordinates on it.

6
For the system of coordinates, mathematicians technically
talk of a mapping from R2 into S, with nice properties of
smoothness and differentiability.

And S has a metric, defined at any point P with coordinates

X on it, by the now familiar equation

dS 2 = gmn (X) dX m dX n (1)

where m and n run over 1 and 2.

Notice that this metric defines the distance between
two infinitesimally close points, not between two points far
away.

A flat geometry is one for which Euclid’s postulates are cor-

rect. There are parallel lines, all the stuff we learned in high
school Euclidean geometry. And the surface S can be laid
out on a plane without applying any distorsion, stretching
or compression, on it. In the case of two dimensions, such
a surface is also called a developable surface.

Any geometry is characterized by its metric tensor. A flat

one is not necessarily one whose metric, given by equation
(1), has the special form of the Kronecker delta tensor. That
is not the right statement. A flat geometry is one where we
can find another system of coordinates Y such that at any
point P , now located with those Y coordinates, the metric
has a simple form

dS 2 = (dY 1 )2 + (dY 2 )2 (2)

Notice that to find such a transformation is always possible

locally at any given point P , because locally any surface
is like a plane. But it is not always possible to find such a

7
transformation globally over the whole surface.

In other words, more generally, given an arbitrary metric

tensor which varies from place to place – assuming it is
smooth, that it has all the good differential properties and
all – the problem of determining whether the corresponding
space is flat is a somewhat difficult problem.

The bad way to approach the problem is to search through

all possible coordinate systems, that is for each possible co-
ordinates Y to transform the initial metric tensor and see
whether its new form is the Kronecker tensor. Of course,
this will take an infinite amount of time. We will never get
there.

So we need a better technique. The better technique is to

search for a diagnostic quantity, that is a quantity which is
built out of the metric and its derivatives, that we can cal-
culate 2 . If it is zero everywhere throughout the space that
will say the space is flat. If it is non-zero at some place, it
will say that the space is curved there.

The diagnostic quantity that does the job is called the cur-
vature tensor.

The goal of this chapter is to get to the curvature tensor of

a Riemannian geometry. It is a complicated goal – I warn
you about this. You will have to pay attention or else you
will fall asleep. It can be a very soporific subject, so we will
try from time to time to tell a joke to keep you awake. But
on the other hand, if you can follow it, you will learn a lot.

2. it will be a quantity computed at every point, so in reality it

will be a scalar field.

8
And it is very interesting. I think you will find it very in-
teresting. We will try not to make it too boring.

What do we start with ? We start with a space. And a space

means first of all a number of dimensions. In Riemannian
geometry the number of dimensions can be any integer. In
principle you can even have a zero-dimensional space. But
that is just a point ! There isn’t much to say about the
geometry of a point. So let’s go to the next number of di-
mensions.

A one dimensional space is either an infinite line or a closed

curve – a loop. If it is a closed curve what is it characterized
by ? It is characterized by only one thing : the total length
of the curve. Every loop is equivalent to every other loop of
the same length. In other words – just think about it for a
moment – take a piece of rope that closes on itself to form
a loop and has a certain length. Wiggle it or curve it in
any kind of ways, it can always be mapped or put on top of
another piece of rope of exactly the same length. It can be
done in a precise way without stretching either of them. A
bug crawling successively around the two loops of the same
length would not notice any difference.

So one dimensional spaces are all equivalent to each other,

up to their length. There are only two general kinds : either
they are infinite, or they form a loop of a certain length L.
And every loop of length L is identical to any other loop of
the same length.
And it is an intrinsic property. The bug walking along
the closed rope cannot distinguish the shape we give it in
3D. All it can do is count the number of steps it takes to
walk around the rope. For instance, the bug makes an ini-

9
tial mark someplace, goes around the rope till it comes back
to the mark, and records the number of steps it took it to
do it. That is the only thing the bug can say, or measure,
about the closed rope.

Two dimensional spaces are where things start to be more

complicated – and more interesting. There are flat ones.
There are curved ones. A flat one is a plane. A curved one
could be a sphere. It could a space with lumps, the surface
of the Earth including the mountains and the valleys.
It can even have a weird topology, for instance the sur-
face of a donut – also called a torus. You can poke another
hole in the torus and make a torus with two holes, and so
forth. So by the time you get the two dimensional spaces
they are pretty complicated. Nevertheless they are charac-
terized by the metric tensor.

We are concerned with whether they are flat. The first ques-
tion then is : find something which distinguishes whether
they are flat. If gmn is not equal to δmn , that doesn’t mean
the space is not flat. It just means you may be in the wrong
coordinates. You can’t diagnose the space by asking whe-
ther gmn is δmn .

We want to find a quantity called R – not necessarily a

scalar – such that if it is zero everywhere the space is flat,
and if it is not zero someplace the space is not flat there. It
would be best if it is a tensor, because if it is a tensor and
it is non-zero, it is non-zero in every frame of reference. R
stands for Riemann. He was the first person who asked this
question and then answered it.

So we want to find a tensor R such that if R is equal to

10
0, that implies that the space is flat and that there exist
coordinates such that

gmn = δmn (3)

That is what we are going to search for : the curvature

tensor.

Recall a few properties of the metric tensor and what you

do with it. First of all gmn is by definition a function of
position in general : gmn (X). If it were not a function of
position, if it was a constant matrix, that is a constant set
of coefficients, then it would be very easy to find coordi-
nates in which gmn is δmn . You simply have to straighten
out the angles of the coordinate axes and adjust the units so
that they correspond to length. So if gmn is not dependent
on position, then the space is definitely flat.

So we start with gmn (X), where X is the set of coordinates

of a point P . What do we do with it ? We compute

gmn (X) dX m dX n

That gives you the square of the length of the little diffe-
rential displacement dX.

Figure 2 : Little displacement dX.

11
We call the displacement dX for simplicity, but we could
call it dP . It is indeed a small vector which will have dif-
ferent components in different coordinate systems. In a Y
system, we will call its components dY . So the square of
the length of the little displacement is

dS 2 = gmn (X) dX m dX n (4)

The metric tensor is a tensor. We proved that in the last

lesson. We used the fact that dS 2 is an invariant quantity.
All observers, that is all coordinate systems, should agree
about the length of a given little vector as shown in figure 2.
You express the vector in Y coordinates instead of X co-
ordinates. The length has to be the same and that leads
immediately to the proof that gmn really is a tensor, that
is, transforms like a tensor.

Next property of gmn is that it can be viewed as a matrix

and it has an inverse. For instance in four dimensions, the
covariant metric tensor gmn is also the symmetric matrix
 
g11 g12 g13 g14
 g12 g22 g23 g24 
gmn =   g13 g23 g33 g34 


g14 g24 g34 g44

The inverse of gmn is denoted g mn . It is a matrix, and it is

also a contravariant tensor. It satisfies the following equa-
tion
g mn gnr = δ m
r (5)

A remark on notations : equation (5) is a matrix equation

as well as a tensor equation. As matrix equation it says that

12
g mn times gmn yields the unit matrix. As a tensor equation
it says that the tensor product of g mn with gnr , when it
is then contracted along the index n, gives the Kronecker
delta tensor.
Up to now, we haven’t paid much attention to the pla-
cing of the indices of the Kronecker delta symbol, see for
instance equation (3) where we wrote both of them downs-
tairs because we equated it to gmn . The point to note is that
in matrix algebra, the Kronecker delta symbol is generally
denoted δmn . While in tensor algebra of general relativity
or Riemannian geometry it is most of the time denoted
δmn with one upstairs index and one downstairs index (the
downstairs index slightly shifted to the right if we like neat
typography) – at least when the Kronecker delta symbol is
treated as a tensor.
This is a thing to prove : if you take the Kronecker
delta symbol and you transform it pretending it is a tensor
of whichever type, you just get back the Kronecker delta
symbol of the same type.

There is another important point to mention : in Rieman-

nian geometry, as opposed to Einsteinian geometry, all lengths
and their squares are positive. This says something about
the properties of the metric tensor : there are no directions
along which if you evaluate dS 2 you get anything but a po-
sitive number. That is equivalent to a statement about the
eigenvalues of gmn . It says that all of the eigenvalues of the
matrix are positive.

To be complete, the following statements are equivalent :

1. dS 2 , of equation (4), is always positive
2. the quadratic form gmn (X) dX m dX n is positive de-
finite

13
 3. the matrix gmn has only positive eigenvalues
4. the matrix gmn is positive definite

That will change when we go to the theory of real gravity

and have to use the Einsteinian geometry.

The metric tensor allows you to do something : it allows you

to raise and lower indices. Consider a vector with contra-
variant components
Vm
You can put it in correspondence with another vector with
covariant components. It is really the same vector. We tal-
ked about this a little bit in the last chapter. We saw the dif-
ference between contravariant and covariant components 3 .
To construct the covariant components of a vector, given
its contravariant components, we just multiply it by the
covariant metric tensor

Vn = V m gmn (6)

Likewise, and this is something that needs to be proved,

the following holds

V m = Vn g mn (7)

So if you know the upper indices, that is the contravariant

components of a vector V , equation (6) is the way you build
its covariant components. And if you know the covariant
3. V m and Vn denote the same geometric vector V . The V m ’s are
the components of V in the basis of ei ’s, see figure 7 of chapter 2. The
components Vn ’s are defined by equations (11) of chapter 2. The Vn ’s
can also be seen as the components of a dual vector. See vector spaces
and their dual in any book of linear algebra.

14
components of V , equation (7) is the way you build its
contravariant components. This requires a bit of proof that
is left to the reader.

Gaussian normal coordinates

To what extent can you force the metric to look like a flat
metric ? Now we mean not only that the space is flat, but
that we are in a coordinate system such that gmn is equal
to δmn . If that is the case, the space is definitely flat – that
is the definition of flatness. But to what extent given any
space can you find coordinates such gmn might be the Kro-
necker delta over some limited region.

Here is a theorem : at any given point P on the surface,

we can find a system of coordinates which are good in the
immediate vicinity of the point, see figure 3. They will not
be globally flat. Equation (4) will not reduce over the whole
variety to Pythagoras theorem.

But in the immediate vicinity of point P , with a suitable

choice of coordinates the metric tensor gmn can be made
to be approximately like the Kronecker delta. Such coordi-
nates are called Gaussian normal coordinates.

Here is the way we proceed. We position ourselves at point

P and we move along any first direction as straight as we
possibly can. Later we will learn what is meant by "as
straight as we possibly can". It will mean along a geodesic.
So you make as straight a curve as you can.

15
 Figure 3 : Displacement of length ∆S along the surface
and along the tangent plane in the same direction. The coordi-
nates are represented on the tangent plane. We could also have
represented them – slightly curved – on the surface itself.

Say, you are a little bug. You move on the surface following
your nose going straight ahead. That defines one coordi-
nate axis. Then you come back to point P . You have some
surveying tools to figure out which other directions make a
right angle with the fist line. On a two dimensional surface
there is only one other direction (in one sense or the other).
In three dimensions, there is a whole plane. And you go
off in an orthogonal direction as straight as you can. You
build that way a complete set of coordinates based on those
directions.

The theorem says that at every point P of the surface you

can choose Gaussian normal coordinates such that at that
point, whose coordinates are, say, X0

gmn (X0 ) = δmn (8)

16
You can do that in more than one way. If you found coor-
dinates with which equation (8) is true, you can rotate the
coordinates. This will produce a different set of axes such
that equation (8) in the new set is still true. In figure 3,
think of pivoting the coordinate system around P .

The theorem says furthermore that, at point P , once you

have chosen the directions, you can also choose the X’s
such that the derivative of any element of the metric tensor
gmn (X) at that point with respect to any direction in space,
X r , be set equal to zero
∂gmn
=0 (9)
∂X r

The proof is actually very simple. It is just a counting pro-

blem. You count how many independent variables you have,
and how many constraints they must satisfy.

Equation (9) will be true, at a given point, only for the

first derivatives. Unless the space is flat, the derivatives of
higher order at that point won’t be zero

∂ 2 gmn
6= 0 (10)
∂X r ∂X s

So, at a point, there is no content really in saying that the

metric can be chosen to be, let’s say, flat-like. Up to the
first derivatives included that can always be done.

It is in the second derivatives of the metric tensor that

the flatness or non-flatness of the space somehow starts to
show up.

17
How do we prove it ? As said, this is actually not hard. Let’s
take up the point of interest that we called P , of coordinates
X0 , to be the origin.
X0 = 0
Now suppose we have some general metric and some coor-
dinates Y in which the metric has some form, which does
not satisfies equations (9).

Let’s look for some X’s which will be functions of the Y ’s,
and choose them in the following way : at the place where
X = 0, in other words at the origin, let’s also assume that
Y = 0. So the two sets of coordinates have the same origin.
That means that X will start out just equal to Y plus
something quadratic in Y

X m = Y m + Cnr
m n r
Y Y (11)

plus some more complicated terms. We are simply expan-

ding each X m in powers of Y 1 , Y 2 ,... Y N , where N is the
number of dimensions of the variety.

How many such Cnr m are there ? Suppose we work in four

dimensions. Then there 10 distinct combinations Y n Y r , be-

cause Y n Y r = Y r Y n . And for each n and r we have four
m , when m runs from 1 to 4. That means there are 40
Cnr
independent coefficients.

Now how many independent components of g are there ?

Answer : ten. So there are 40 equations (9).

Finally we reached 40 equations to solve for 40 unknowns.

That allows us to be sure, at point P , that not only gmn (X) =
δmn , but also that the derivatives of gmn and δmn will match

18
up to quadratic order 4 . That means that we will be able
to solve the forty equations (9), but that we will fail to set
the left hand sides of equations (10) to be equal to zero.

To summarize, at any point P , a smooth variety is locally

flat. We can approximate it by its tangent space, see figure
3. And we can construct coordinates X’s such that P is
located at the origin, and the metric tensor has the form

gmn (X) = δmn + o(X) (12)

where o(X) represent terms of second order and higher. We

also interpret equation (12) as saying that the metric is lo-
cally Euclidean up to second order.

The reason we introduced this theorem is because we want

to give some meaning to derivatives of tensors. We have
talked about various operations we can do on tensors to
make new tensors : we can add them, subtract them, mul-
tiply them, contract them... We already can do all sorts of
things.

Can we differentiate them ? In particular, if we have a ten-

sor which is a function of position – that means all of its
components are functions of position –, can we construct
another tensor which is in some appropriate sense the de-
rivative with respect to position of the tensor that we have ?

The answer is yes. We shall see how to do it and why it is

a little bit tricky.

4. It is easy to check that we are in a case where the 40 equations

with 40 unknowns do lead to an existing and unique solution. The
reader is invited to verify it in two dimensions.

19
 Covariant derivatives

To differentiate a tensor with respect to position, we could

think : "Okay. Let’s take the components of the tensor – for
instance contravariant components – and just differentiate
them." That would yield a new set of components, with one
more index, which would be simply the derivatives of the
components of the tensor. But there is the following pro-
blem.

Think for instance of the derivatives of a vector. We could

differentiate each component with respect to each direction.
We sure can do that. This would produce a two dimensional
collection of values. But it would not be a tensor 5 . Let us
see why.

Consider a surface and a point P on it, figure 4. We have

two sets of coordinates on the surface, coordinates X and
coordinates Y .
If the space is flat, for X we just use flat ordinary Car-
tesian coordinates. Or if the space is curved, we use a set of
Gaussian normal coordinates at P , that is coordinates X
that are locally, at P , as straight and orthogonal as possible
so to speak. And there is another set of coordinates Y , for
instance the initial ones. For convenience, at P we chose X
such that the X 2 -axis is tangent to the Y 2 -axis.

5. A tensor is an abstract object which has, in each coordinate sys-

tem, a multi-indexed collection of components. And these collections
of components must transform from one system to another according
to equations (14) to (16) of chapter 2 or their obvious generalizations.
If we differentiated a vector field, at the same point, using two different
coordinate systems, we would obtain two collections of double-indexed
values. But they would not be linked by a tensor equation like those
mentioned, which are necessary to define a tensor.

20
 Figure 4 : Surface viewed at P with Gaussian normal
coordinates X, and any coordinates Y .

Think of a vector field defined over the surface. The vector

field is made of a different vector at every point. In order
not to clutter the picture, these vectors are not shown on fi-
gure 4 yet. There is one at P and there are plenty around P .

Before we ask how to differentiate a vector field, let’s ask

what it would mean for the vector field to be constant in
space. We run into the following difficulty : because the
space is curved it becomes hard to compare the vector at
one point with the vector at another point.

The coordinates X cannot be chosen to be everywhere flat.

Then what exactly do we mean by saying that a vector at
one point is equal to a vector at another point ? It does
not really mean anything because to compare a vector at
P with a vector at Q, unless we have nice flat coordinates
over the whole surface, there is no unique way to do that.
So let’s look at it for a moment.

21
If the space is really flat, then we know what it means for
a vector at one point P of the surface to be the same or
to be equal to a vector at another point Q of the surface,
see figure 5. It means they point in the same direction and
have the same length. Therefore in the X coordinates they
have the same components.

Figure 5 : Two equal vectors, at P and at Q.

But what about their components in the Y coordinate sys-

tem ? In fact let’s take a special case. Suppose both vectors
point vertically in the X axes, figure 6. In that case, VP
only has an X 2 component in the flat coordinates. And so
does VQ . They have the same components in the X axes.

Are the components of VP and VQ in the Y coordinates

the same ? Answer : No. Along the Y axes, VQ has a Y 1
component and a Y 2 component, while VP only has a Y 2
component.

22
 Figure 6 : Two equal vertical vectors, at P and at Q.

So it’s clear that even though the vectors VP and VQ are

the same, their components are not the same in the Y co-
ordinate system.
This will be true, be it for covariant vectors or for contra-
variant vectors : when we have curvilinear coordinates, we
cannot judge the equality of two vectors. And in general we
will be dealing with curvilinear coordinates.

Another way to say the same thing is that the derivative of

the m-th component of V , with respect to the r-th direction
of the coordinate system, might 0 in one coordinate system
and not 0 in the other coordinate system.

∂Vm
=0
∂X r
∂Vm0
6= 0
∂Y r

23
It might even be the case – as in figure 5 or 6 – that all of
the derivatives of Vm are 0 in one coordinate system and not
0 in the other coordinate system. That will be because the
coordinates are shifting, not because the vector is changing.

You see what we are driving at : the equation saying that

the derivatives of a vector are all equal to 0 may true in one
reference frame, but untrue in another. Therefore it cannot
be a tensor equation. In other words

the ordinary derivatives of the components of a vector with

respect to the coordinates do not form themselves a tensor.

If they were components of a tensor we would think of this

quantity
∂Vm
Tmr =
∂X r
as a rank 2 tensor with an m index and an r index.

But it were a tensor, the following fact would have to be

true : if Tmr is zero in one frame or one coordinate system,
it is zero in every coordinate system. And it is simply not
true with that Tmn – not because the vector may change
from point to point, but because the orientation of the co-
ordinates changes.

So we need a better definition of the derivative of a vector

than just differentiating its components. We need some-
thing which if it is zero in one frame is zero in every frame.

Here is how we will define the derivative of a vector. As a

preliminary, notice that to define the derivative at a point
P we only need to look at points in the vicinity of P .

24
The first thing we do is to construct a set of Gaussian nor-
mal coordinates at point P . Remember : Gaussian normal
coordinates are as straight as possible near P . They are
well defined over the whole variety, and they make up an
approximately Euclidean system of coordinates in the vi-
cinity of P . So we re-express all the vectors of the vector
field in the new coordinates X that are locally flat and Eu-
clidean.

To follow the procedure geometrically, let’s look again at

two vectors, one at P and the other at Q nearby, and for
clarity make the second also slightly different from the first.

Figure 7 : Two vectors nearby.

Then pretend that the Gaussian normal coordinates are

really nice flat coordinates over the whole vicinity of P . Vi-
sually translate the vector VQ so that its origin be the same
as VP ’s, and look at the difference between VQ and VP .

25
The components, in Gaussian normal coordinates, of the
difference between VQ and VP are the kind of elements that
we will use to define the derivatives of the vector field at P .
For instance if we were interested in the derivatives along
−−−−−→
the P Q direction, it would approximately be VQ − VP divi-
ded by the small distance between P and Q. But we shall
be interested in the derivatives of V along the X axes.

Now the derivatives ∂Vm /∂X r , in the Gaussian normal co-

ordinates, define the derivatives of V at P .

Finally, if we want to work with our initial Y coordinates,

we take the double-indexed collection of partial derivatives
produced by the differentiation in Gaussian normal coor-
dinates at P . We treat it as a tensor, that is, we consider
this collection as the components of a tensor in the X coor-
dinates. And we transform it back into the Y coordinates
using the tensor equations linking X and Y .

This necessarily produces a tensor, since when transformed

from any Y coordinates into the X coordinates, it gives the
same thing.

Now when we look at the derivatives of a vector in the gene-

ral coordinates Y , we will have the addition of two terms :
one term because the vector may be changing and the other
term because the coordinates may be changing. As we saw,
the coordinates may shift, may be rotating out from under
you, even if the vector doesn’t change (see figures 5 and 6).

But before we look into these two terms, let’s repeat the
prescription as a methodical procedure.

26
 You have a vector field V on a variety, equipped with
any coordinate system Y . You want to calculate the
derivative of V at P . Then follow these steps :
1. Change coordinates to use Gaussian normal coor-
dinates at P , let’s call them X (notice that they
are valid over the whole surface, and approxima-
tely flat at P ).
2. Differentiate V at P in the usual way, using the
X coordinates.
3. Consider the collection of partial derivatives you
got as the components of a tensor of rank 2 in the
X coordinates.
4. Switch back to your original coordinate system
Y , and re-express in that original system the ten-
sor you got, using the tensor equations linking
X and Y .

Let’s look at what we get and then comment on each term.

We find that, with this new definition, the derivatives at P
are the old ones corrected by the addition of another term :

Dr Vm = ∂r Vm − Γtrm Vt (13)

Here is how to read equation (13) :

1. The notation Dr Vm is by definition the partial deriva-
tive of Vm with respect to the r-th direction in X ob-
tained from the above procedure, that is, in the Gaus-
sian normal coordinates X, and then re-expressed in
the Y coordinates.

27
 2. ∂r Vm is the ordinary partial derivative of Vm with
respect to the r-th direction in Y calculated directly
in the Y coordinates. ∂r is a short hand for ∂Y∂ r .
3. −Γtrm Vt is the additional term due to the fact that
the coordinates Y themselves evolve in the vicinity of
P . The minus sign is a pure convention. This whole
second term on the right hand side of equation (13)
must clearly be proportional to Vt . If you double the
size of Vt it must be twice as big. And the coefficient
Γtrm in front of Vt is a new mathematical object ap-
pearing in the differentiation procedure. We shall talk
about it.

The term Γtrm Vt does not have a derivative related to the

vector because if doesn’t come from the fact that the vector
is changing. It comes from the fact that the coordinates are
changing in the vicinity of P .

The right hand side of equation (13) is what you will get
if you take a vector, differentiate it in Gaussian normal co-
ordinates and then transform the double-indexed collection
of derivatives to other coordinates as a tensor. In any other
coordinate system what you will get is the usual derivative
in that coordinate system minus an object times the com-
ponents of V themselves. As usual Γtrm Vt means of course
the sum over t.

Equation (13) holds in any arbitrary coordinate system. Of

course the Vm ’s or Vt ’s and the Γtrm ’s depend on the coor-
dinate system. Notice that there is not only one Γtrm : they
form a three-indexed collection.

28
Dr Vm is called the covariant derivative of Vm 6 . This is
not because the index m is downstairs. The terminology
has another origin. If we had differentiated V m , we would
have obtained a formula analogous to equation (13). And
it would also be called the covariant derivative of V m .

And, by its very construction, Dr Vm is a tensor.

Christoffel symbols

The coefficients Γtrm have two names : connection coeffi-

cients and Christoffel symbols.

The name connection coefficients comes from the fact that

they connect neighboring points and tell us how to calcu-
late the rate of change of a vector field from one point to
another, even though the coordinate system may be chan-
ging.

They are also named Christoffel symbols after Elwin Bruno

Christoffel 7 . They have been known on occasion as the
"Christ awful" symbols because they seem complicated.
With some practice, however, the reader will discover that
they are not that complicated. They are just an extra li-
near term. But I grant you that they are complicated and

6. In other contexts it goes under the names of material derivative

or convective derivative or Lagrangian derivative, etc.
7. Elwin Bruno Christoffel (1829-1900), German mathematician
at the University of Strasbourg – when it was in the German empire
between 1871 and 1918 – who did fundamental work in differential
geometry.

29
unlikeable enough.

Let’s investigate what follows from the definition of the co-

variant derivative and the Christoffel symbols. We are not
going to prove every single fact we state, because there are
just too many little pieces. But they are easy to check.

It follows from the definition of the covariant differentiation

– namely, to differentiate a vector V at a point P , go to a
set of Gaussian normal coordinates at P , differentiate the
vector in the ordinary manner, treat the object you obtain
as a tensor with two indices, change coordinates, etc. – that
the Christoffel symbols have a symmetry :

Γtrm = Γtmr (14)

There are generalized Riemannian geometries, also called

geometries with torsion, in which this symmetry is not true.
But those geometries are not widely in use in ordinary gra-
vitational theory. The geometry of general relativity is the
Minkowski-Einstein geometry which is an extension of Rie-
mannian geometry with a non-positive definite metric. But
it doesn’t involve torsion. So the Christoffel symbols we will
use will be symmetric in the sense of equation (14).

A remark to build our physical intuition : doing the deriva-

tive in Gaussian normal coordinates which are almost flat,
or as flat as can be, and then treating what we obtain as
an object in its own right, is very similar to what we do in
gravitational theory when we evaluate something in a freely
falling frame.
For example in a freely falling frame we calculated how
light moved across an elevator and then we transformed it

30
to the frame of reference in which the elevator was accele-
rating.
That is closely related to the operations we have been
doing in this chapter : we calculate something because we
know how to do it in coordinates which are as flat as pos-
sible. That would be a freely falling frame in general rela-
tivity. And then we transform it in any coordinate we like,
accelerated coordinates or anything we like, and we trans-
late the statement from one coordinate system to another.
In the construction of the covariant derivative, the cal-
culation of the variation of a vector from point to point is
done first in Gaussian normal coordinates, and then it is
transformed in any coordinate system. Equation (13), re-
produced below,

Dr Vm = ∂r Vm − Γtrm Vt (15)

is the form that you get for the corresponding collection of

components. It is a tensor. However, ∂r Vm is not a tensor.
Therefore Γtrm Vt cannot be a tensor. And Γtrm cannot be
a tensor either.

We will see that the Γtrm ’s are build-up out of the deriva-
tives of the metric ∂r gmn . In fact in a coordinate system
in which the derivatives of the metric are 0, the Christoffel
symbols are 0. But a tensor, if it is zero in one coordinate
system, it is 0 in every coordinate system. So that is ano-
ther way to see that they can’t be tensors.

Let’s look now at the covariant derivative of higher rank

tensors, because we will need this for curvature. Suppose
we have a tensor with more than one index, say

Tmn

31
and we want to differentiate it covariantly along the r-th
axis. We denote the resulting tensor

Dr Tmn

Its expression is the analog of equation (15), except that

for every index in the tensor Tmn there will be a term like
Γtrm Vt . Let’s see you how it works :

We start by working only on the m index, letting n be

passive, and writing the equivalent of (15)

Dr Tmn = ∂r Tmn − Γtrm Ttn − ...

But we are not finished. We have to do exactly the same

with the n index, letting this time m be passive.

Dr Tmn = ∂r Tmn − Γtrm Ttn − Γtrn Tmt (16)

That is the form of the covariant derivative at point P of

the tensor Tmn . The rule is the same : we switch to Gaussian
normal coordinates at P , we do the ordinary differentiation
of the tensor with respect to each direction X r . This adds
one more index to the collection of components that formed
Tmn 8 . And we re-express the new tensor in the original co-
ordinate system with the usual tensor equations (equations
(15) and (16) of chapter 2 and their generalizations).

This allows us to to differentiate any tensor. At the moment

we are only dealing with tensors with covariant indices. We
will come in a moment to tensors with contravariant indices.

8. Notice that in order not to clutter equations we no longer use

prime signs for one system and un-prime for the other. The last time
we did it in this chapter was in our comments following figure 6.

32
The reader may wonder : what is all this intricate business
of covariant differentiation of tensors for ? Answer : it is for
comparing things at different points. We want to be able
to talk about rates of variation of things along coordinate
lines, with objects which have an existence irrespective of
the system of coordinates we work with.
Remember that a vector in ordinary 3D has an existence
irrespective of the basis we are using. For certain work and
calculations with it – not all of them – we need a represen-
tation of the vector in a basis. The collection of components
to represent it and work with it is different from one ba-
sis to another, but the vector we are talking about is the
same 9 .
Where are we going to use covariant derivatives ? Ans-
wer : in field equations. Field equations are going to be
differential equations which represent how a field changes
from one place to another. But we want them to be the same
equations in every reference frame. We don’t want to write
down equations which are special to some peculiar frame.
We want them to be valid in general. That is, if they are
true one frame they will be true in all frames. That means
they have to be tensors equations. So we have to know how
to differentiate tensors to get other tensors.

Another point worth stressing : the Christoffel coefficients

will be present in equation (16) even in a flat space – like
a plane, or this page, or the ordinary 3D Euclidean space –
if you chose funny coordinates (see figure 3 of chapter 2).
That is an important point : terms like Γtrm Ttn are there

9. It is even like the difference between things and their notations.

When we talk about 12 in the decimal system or 1100 in the binary
system, we are actually talking about the same thing : the number
twelve – which some other people call douze...

33
even in flat space if you are using funny coordinates. In fact
if you choose any coordinates in which the derivatives of the
gmn ’s are not zero, that is, in which the coordinates vary
from point to point sinuously (viewed from an embedding
space for instance), Γtrm Ttn will be there.

The presence of terms like Γtrm Ttn in the covariant deri-

vative of a tensor is not a characteristic feature of curved
spaces, it is a feature of curved coordinates.

To begin to use our new tool, let’s apply equation (16) to the
metric tensor itself. There is something special, however,
about the metric tensor : in Gaussian normal coordinates,
its derivatives are all zero. That means that the covariant
derivative of the metric tensor is zero. That simple observa-
tion is what is going to allow us to compute the Christoffel
symbols. Let’s write equation (16) with the metric tensor.
We are in a coordinate system that is general, curvilinear
or not, flat of not, whatever :

Dr gmn = ∂r gmn − Γtrm gtn − Γtrn gmt

But we know that this is zero. Why ? Because the ordinary

derivative of the metric tensor in Gaussian coordinates is
zero. So, in any coordinate system, we have

∂r gmn − Γtrm gtn − Γtrn gmt = 0 (17a)

Now let’s write the same equation, except permuting the

indices. It is a little trick to get as much information about
the Christoffel symbols as we can, and eventually, via some
nice eliminations, to be able to isolate one Christoffel sym-
bol and express it in terms of the ordinary partial deriva-
tives of g with respect to the axes in any coordinate system.

34
So equation (17a) becomes

∂m grn − Γtmr gtn − Γtmn grt = 0

But the middle term, by symmetry, can be rewritten inter-

changing m and r

∂m grn − Γtrm gtn − Γtmn grt = 0 (17b)

And for exactly the same reasons we can write

∂n grm − Γtrn gtm − Γtmn grt = 0 (17c)

Let’s write the three equations of interest next to each

other, to look at them more conveniently :

∂r gmn − Γtrm gtn − Γtrn gmt = 0 (a)

∂m grn − Γtrm gtn − Γtmn grt =0 (b)
∂n grm − Γtrn gtm − Γtmn grt =0 (c)

How can we add them, or subtract them, or do something

clever, that will isolate only one of the terms with a gamma ?

Let’s add equation (b) to (c) and subtract (a). Of course

we will get ∂n grm + ∂m grn − ∂r gmn plus some other terms.
But the middle term of (a), Γtrm gtn , will disappear, and so
will the last term of (a), Γtrn gmt . And we will be left with
twice the same last term with a gamma, Γtmn grt . So we are
in luck : (b) + (c) − (a) yields

∂n grm + ∂m grn − ∂r gmn = 2Γtmn grt (18)

We are still not done. We would like to get Γtmn by itself.

Our goal, indeed, is to find out what the Christoffel sym-
bols are in terms of derivatives of the metric. We are almost

35
there.

Fist of all, equation (18) shows that if all the derivatives of

the metric are zero, the Christoffel symbols must be 0.

But how are we going to get rid of the grt on the right hand
side of equation (18) ? The answer comes from recalling that
grt has an inverse. We saw that in the form of matrix equa-
tions, as well as in the form of tensor equations, see equation
(33) of chapter 2. We multiply both sides of equation (18)
by the inverse tensor, and move also the factor 2, and we
get
1
Γtmn = g rt [ ∂n grm + ∂m grn − ∂r gmn ] (19)
2

This is the expression of the Christoffel symbols in terms

of the ordinary derivatives of the metric tensor.

It looks pretty simple. The indices m and n are symmetric.

You can interchange them, the Christoffel symbol won’t
change. There is one negative term and two positive terms.
It is not very complicated.
The problem is that there is a boatload of them. When
you think about a four dimensional space and let all the
coefficients range from 1 to 4, there is just a lot of Christoffel
symbols.
That is what makes doing calculations in general rela-
tivity a very tedious business. Intrinsically there is nothing
hard about it. But in fact doing a calculation in a general re-
lativity context usually fills page after page of nothing more
complicated than just computing these derivatives and as-
sembling them together.

Equation (19) holds for any coordinate system and any

36
metric tensor. Notice that all our calculations are at one
point P . Whatever coordinate system our variety is equip-
ped with, we position ourselves at a point on it, consider
the metric tensor gmn there, calculate the gammas there
with equation (19).
The use of Gaussian normal coordinates at P was just
for intermediate reasoning, calculation and proof purposes.
We are now back in the initial coordinate system of our
space.
The gmn , g mn and Γtmn depend on P . But equation (19)
is general. At every point, it expresses the connection co-
efficients – the other name of the Christoffel symbols – in
terms of the derivatives of g. These connection coefficients
enable us to figure out how any vector or tensor varies, in
our space, when we move a little bit along the coordinate
lines.

The problem with the Christoffel symbols is that they are

not tensors. They can be zero in one frame of reference,
and not zero in another. For example, in a set of Gaussian
normal coordinates at point P , all of the Γtmn are equal
to zero. This can be seen in many ways. Since the metric
tensor in that case is constant (even equal to the Kronecker
delta tensor, but that is not necessary), equation (19) tells
us that Γtmn = 0. But in some other coordinate systems the
Christoffel symbols are not.
Remember that even in an intrinsically flat space, we
can have coordinates such that the metric tensor is not
constant. Then the Chritoffel symbols won’t be zero. As
said, they are related to the coordinate system not to the
intrinsic geometry of the space.
A sphere is intrinsically non flat. In the polar coordi-
nates θ and φ (see chapter 1, figure 14), the components

37
of g are not constant, therefore the Christoffel symbols are
not zero in that system of coordinates. Even on a sphere,
however, at any given point we can build a set of Gaussian
normal coordinates (like maps do – you just tinker with the
longitude), and then the Christoffel symbols at that point
will be zero.

Exercise 1 : Explain why the space can be flat

and nevertheless the Christoffel symbols not
zero.

Exercise 2 : Explain why the covariant derivative

of the metric tensor is always zero.

Exercise 2 : On the Earth, with the polar coor-

dinates θ for latitude, and φ for longitude, find
1. the metric tensor gmn
2. its inverse g mn
3. the Christoffel symbols at point (θ, φ).

All this is conceptually tricky when we meet it for the first

time. But at the end of the day the rule is simple : calcu-
late the Christoffel symbols and, in many contexts, replace
ordinary derivatives with covariant derivatives.
You could write your equations in Gaussian normal co-
ordinates. Then they would just involve ordinary deriva-
tives and we would not have to wade through a river of

38
Christoffel symbols. But if you want the same equations in
general coordinates then replace the ordinary derivatives by
covariant derivatives.

That is the procedure. It will require the reader to think

about it. He or she will have to sit down, to follow carefully
the reasonings, to do exercises. Then what we are doing will
become clear.

Curvature tensor

What is curvature ? It is easiest to start with two-dimensional

curvature. Intuitively it is easy to understand : it is a cha-
racteristic of something that is round and cannot be flatte-
ned out. But we are going to give it some more mathema-
tical definition. How do we probe for curvature ?

Let’s begin by drawing a space which is curved. The sphere

is obviously curved. I don’t want, however, to deal with a
sphere. I want to look at a cone.

Figure 8 : Cone with a rounded summit.

39
But it is going to be a cone with a round summit, figure 8.
Think of the top of a mountain the sides of which are nice
and flat like those of a volcano, and the top is round.

So if you are away from the top of the mountain, below

the dotted line, around you the surface is flat. It may not
look flat because, like the furled page in figure 1 of chapter
2, we represented the mountain embedded in 3D Euclidean
space. But the surface is what mathematicians call develo-
pable : any section with no hole in it, cut from the side of the
mountain, can be flattened onto a plane without distorsion.

The rounded cone only differs from a flat space in this vi-
cinity of the summit. To see that just take the same space
below the dotted line but continue it so that it really does
form a genuine cone.

Figure 9 : Genuine cone.

Then slice it along a generatrix, that is, a straight line on

the cone going to the top. And open it up. You can lay out
flat on a plane the shape that you get. It is a disk with a
missing piece, see figure 10.

40
 Figure 10 : Cone opened up and laid flat
(smaller scale than figure 9).

The missing piece is called the deficit angle, or the conical

deficit. We can see that the bigger the conical deficit is, the
pointier the cone will be.

Now, on the flat surface of figure 10, let’s consider a col-

lection of identical vectors arranged around the shape as
shown in figure 11.

Figure 11 : Identical vectors.

41
On the flat surface, all the vectors point in the same direc-
tion. But when we fold again the shape to form the cone, we
see that the vectors no longer point in the same direction.
Think of them as very small so that they don’t have to be
bent. The first one on the left is along a generatrix, but the
last one on the right points in another direction.

So the tip of the cone is such that if you carry a vector

around (in the same direction on the flattened version)
when you get around to the other side it has been rota-
ted. The rotation is due to the conical deficit.

Exactly the same is true on the rounded cone of figure 8 :

if we take a vector on the flat side, below the dotted line,
and we carry it around the mountain in such a way that,
when it is opened up and laid on a plane, it is always poin-
ting in the same direction, by the time we get back to the
other side it will be pointing in some other direction. That
is because the summit of the shape is curved. That is the
effect of curvature.

A region of curvature has the property that if you take a

vector pointing in the same direction as much as you can
do, that is locally parallel to itself, and take it around the
region of curvature, then the vector will rotate.

The vector will rotate when it goes around any region of

curvature even though you took pains to make sure that at
every point you were moving it parallel to itself. In other
words you took pains to make sure that the covariant deri-
vative of the vector was always zero. But nevertheless, when
you came around to the other side, it is shifted.

42
There is another way to say this, which is equivalent and
actually more useful.

Consider a curved space with some curvature at point P ,

figure 12.

Figure 12 : Displacements to differentiate a vector field

along two axes.

Take a vector field and differentiate it along one axis (first

displacement in figure 12). Then differentiate it along the
second axis (second displacement in figure 12). That is, you
consider the vector field at P . Then you move a bit along
one axis, and consider the new value of the vector field at
I. Then you move another bit along the second axis, and
consider the value of the vector field at Q.

Over each displacement, the vector will change. How will it

change ? It will change typically by differentiating the vec-
tor along the two axes in sequence. We first differentiate
the vector along one axis and then differentiate it along the

43
second axis. And this will produce a small change in the
vector due to the two derivatives.

The total change in the vector consists of two changes. And

that total change is proportional to a second derivative 10 .
That is true in any coordinates : if, to compare the vectors
at Q and at P , you compared the vector at I with the vector
at P , and then compared the vector at Q with the vector
at I, what you would be calculating is the second partial
derivative of the vector with respect to the two directions.

In figure 12, if the first displacement is along the direction

X s and the second displacement along the direction X r ,
then the variation of the m-th component of the vector V
– let’s say it has covariant indices – would be

Dr Ds V m (20)

This expression is calculated covariantly. And in Gaussian

normal coordinates, expression (20) would just contain or-
dinary derivatives.

We could have also gone in the other direction, figure 13.

That is to say we could have gone first in the r direction and
then in the s direction and calculated the way the vector
changes from P to J and then from J to Q. The variation
of V would then be
Ds Dr Vm (21)

10. To get a feel for this, think of a function f of two independent

variables a and b. View it in 3D. Suppose f is nice and smooth every-
where, and in particular, at the point (0, 0), is tangent to the plane
formed by the a and b axes. Suppose we are interested in f (a, b) −
f (0, 0). If a and b are small, this variation is approximately equal to
∂2f
ab ∂a∂b (0, 0).

44
 Figure 13 : Displacements in the other order.

Ordinarily, and in flat space in general, expression (20) and

expression (21) are equal to each other. That is
Dr Ds Vm = Ds Dr Vm (22)
This is just a version of the fact, in calculus, that the partial
derivatives of a function of several variables can be taken
in the order you like (see Interlude 3 of Volume 1 in the
collection The Theoretical Minimum).

But equation (22) is not true in curved space. In curved

space the difference
Dr Ds Vm − Ds Dr Vm
can be thought of as taking the vector around the closed
loop
P →I→Q→J →P

Let’s go back to our cone, either the genuine cone (figure 9),
or the cone with a rounded top but looking at the part be-
low the dotted line (figure 8). Consider a vector field which

45
when the cone is opened and laid flat is constant. Fold the
flat shape to form the cone. We discovered that if we follow
the vector field on a closed loop around the top, we don’t
get back to the same vector we started with. This is due to
the following fact, that is important enough to note with
italics.

In curved space covariant derivatives are not interchan-

geable. In flat space they are interchangeable.

That is the way we will test whether the space is flat or not.

We will test whether differentiating tensors, and in parti-

cular vectors, in opposite order gives the same result.
- If the answer is yes everywhere in the space for any
vector, then the space is flat.
- If we discover that there are places in the space where
the order of differentiation gives different answers,
then we know the space has some kind of defect in
it (like the point of the genuine cone) or has curva-
ture (like the summit of the rounded cone).

So all we have to do is compute the second covariant deri-

vatives of a vector in opposite order and compare them. In
principle it is not complicated. In practice it will be a little
complicated, but will remain manageable. We have all the
tools to do it. Now it is a mechanical operation, consisting
of pure plug-ins. We sketch the steps below, and then give
the answer.

We start with a vector expressed with covariant components

Vn

46
We compute its covariant derivative in the r direction

Dr Vn

And then we differentiate this, still covariantly, in the s

direction
Ds Dr Vn (23)
Later on, we will interchange s and r and subtract.

Let’s replace the first covariant derivative of Vn , with res-

pect to r, by its expression given in equation (13). We get

Ds Dr Vn = Ds ∂r Vn − Γtrn Vt
 

But notice that ∂r Vn − Γtrn Vt is a tensor. So we know how

 

to differentiate it : use equation (16). Continue to crank me-

chanically the calculations.

In the end, the difference between the two second order co-
t , multiplied
variant derivatives yields a tensor, denoted Rsrn
by Vt
t
Ds Dr Vn − Dr Ds Vn = Rsrn Vt (24)

And here is R
t
Rsrn = ∂r Γtsn − ∂s Γtrn + Γpsn Γtpr − Γprn Γtps (25)

There are two terms involving derivatives of Christoffel

symbols, and two terms which are sums over p of products
of Christoffel symbols.

t is the curvature tensor, also called Riemann curvature

Rsrn
tensor or Riemann–Christoffel tensor.

47
It has a complicated expression. And it is even more com-
plicated when you remember that the Christoffel symbols
are given by equation (19), which we reproduce below
1 rt
Γtmn = g [ ∂n grm + ∂m grn − ∂r gmn ]
2

Let’s see what are the elements in the curvature tensor gi-
ven by equation (25).

The Christoffel symbols involve derivatives of g. So differen-

tiating again produces second derivatives of g. Remember :
the second derivatives of g are the things that we cannot
generally set equal to zero.

For the first derivatives of g, we saw that we can find a

frame of reference where they are equal to zero. But for
the second derivatives of g we can’t. So by the time we are
finished calculating the curvature tensor, the second deri-
vatives of g have come into it.

The second derivatives are testing and probing out the geo-
metry of the surface a little more thoroughly than just the
first derivatives. In a similar way, in the theory of functions,
when at a point x you know f (x) and f 0 (x) and f ”(x), you
are better off than with just the value of f and of its first
derivative.

So the curvature tensor has second derivatives of the me-

tric g, and it has squares or quadratic things involving first
derivatives of g. It is a complicated thing. If we were to ac-
tually write it in terms of the metric, or we were to try to
calculate it for a given metric, it could rapidly fill up pages.

48
But conceptually what it is doing is simply calculating the
difference in a vector if you transport it around the loop
in figure 13, keeping it parallel to itself, as much as you
can locally at every point, until you have come all the way
around. It calculates the little change in a vector in parallel
transport going around a loop.

The curvature tensor has a complicated formula, but we

can calculate it. We can put the metric tensor into a com-
puter and ask the computer : "Is the curvature tensor 0 ?"
It is even better if you have a software that can do algebra.
If you have the metric in some algebraic form, you can do
all the operations of equations (19) and (25) and then test
out whether the curvature tensor is zero everywhere. If the
curvature tensor is zero everywhere, that is all its compo-
nents are zero everywhere, then the space is flat.

We shall study the curvature tensor a little more. As we

said, it is a complicated thing. Its main use is to tell us
whether the space is flat. And, if not, how unflat it is.

It is closely related to a quantity in gravitational physics.

Can you guess which one ? A local quantity which tells you
that the space is not flat. It must be something telling you
whether there is really a gravitational field present or not.
Answer : tidal forces. It is exactly related to tidal forces,
those things which in a gravitational field squeeze bodies
one way and stretch them another way. Tidal forces are re-
presented by the curvature tensor.

Here is another way to get a feel for what the curvature

tensor is. Imagine a surface which is flat away from a point
in the center where it has a bulge. It doesn’t have to be a

49
rounded cone. It can simply be a plane with a bulge, as in
figure 14.

Figure 14 : Probing the curvature with

a little structure of Tinkertoy sticks.

And you have a small structure of Tinkertoy sticks, all hin-

ged at their extremities, so that their directions can move
freely from each other, while remaining attached. At first,
the probing structure lays flat, without stress or distorsion,
in a flat part of the surface, because the probe is itself flat.

And then you start moving the probe. While you move it in
the flat region, nothing happens to it. It remains perfectly
happy. It doesn’t get stretched, it doesn’t get distorted or
deformed. This would have also been the case on the side
of the rounded cone away from the summmit, by the way.

But now what happens when you try to move the probe
into the curved region ? When you are trying to move it
into the curved region, it simply can’t follow the curvature

50
without having to stretch or compress some of its lengths.
It has to follow the metric properties of the curved space.

In particular if you go around the probe, what you are doing

somehow is sampling the double covariant derivatives of
equation (20) or (21). You are going to find out that va-
rious angles between sticks change from their value in flat
space. The lengths of the sticks shift too, they get stressed,
they get deformed.

The measure of how much the probe gets stressed locally

is given by the curvature tensor. The curvature is an im-
portant property because, if you are in region where there
is curvature, you can feel it, either with tidal forces in a
gravitational field, or with the probe in the experiment of
figure 14.

Uniform gravitational fields don’t have curvature. That is

why in free-fall, in a perfectly uniform gravitational field,
you simply feel nothing. Indeed they don’t cause tidal forces.
Of course perfectly uniform gravitational fields don’t really
exist in nature. You can simulate one with acceleration,
but you cannot see one in nature. They exist only approxi-
mately on the surface of big massive objects, if you limit
yourself to a small solid angle. This leads us to a last re-
mark.

Tidal forces, or curvature on a surface, have a bigger effect

on bigger objects. The 2000-mile man in free-fall toward
the Earth will feel tidal forces more strongly than a free-
falling bacteria. Similarly, in figure 14, if the probe is small
compared to the bulge, it won’t be much deformed when it
goes over it. Whereas if it were a bigger Tinkertoy struc-

51
ture, made for instance of many more hexagones, covering a
larger region of the plane, like floor tiles, but still hinged so
that any two connected sticks can change their directions
from each other, then the probe would feel the curvature
more strongly.

At the end of this third chapter we have reached the cur-

vature tensor. It is complicated. Its expression is given by
equation (25). I wish we could do without the sea of sym-
bols and indices.

But we can compute it. Often you will be presented with

the metric tensor in some analytic form. There will be a for-
mula for it. With the formula you can do differentiations.
Everything consists in analytic functions that you can cal-
culate. So we have finally reached our initial aim.

Remember that our aim was to find a method to deter-

mine whether a space is flat. By definition, it is flat if
there exist a set of coordinates in which the metric ten-
sor is everywhere equal to the Kronecker delta tensor. But
trying out every possible set of coordinates, and checking
them at every point, was not a practical solution. So we
found the curvature tensor. If it is zero everywhere, then
we can find a set of coordinates such the metric tensor is
everywhere equal to the Kronecker delta tensor. You just
position yourself at any fixed point and start to build Eu-
clidean coordinates, like we did when we built Gaussian
normal coordinates. Now no curvature will limit us to a
small vicinity.

In summary, the curvature tensor has a complicated form

but it is practical tool.

52
We are finished with our mathematical study of Rieman-
nian geometry, metrics, tensors, curvature, etc. The inter-
ested reader who wants to go further into the mathematical
aspects of these topics is invited to take up any good ma-
nual in differential geometry oriented toward applications.
As far as we are concerned, our new tools will now be put
to use.

In the next chapter, we will enter into gravity land. We

will see what has to change to go from Riemann geome-
try to Einstein geometry. Then we will study a particular
example : the Schwarzschild geometry. It is the geometry
of a black hole, a star, or any gravitating mass.

53