General coordinates

February 19, 2013

In order to describe curved spaces, it is necessary to assign a set of labels coordinates to each point.
There are infinitely many ways to do this, and we need to write our physical laws in a way that doesnt
depend on which choice we make. While we often use certain coordinate systems, e.g., Cartesian, polar,
spherical, in flat space, we need a more general formulation.

1 General coordinate transformations

1.1 From Lorentz transformations to general coordinate transformations
We have defined (Lorentz) vectors to be objects which transform in the same way as the coordinates,
v = v
where may be any Lorentz transformation. Suppose instead, we have a fully general change of coordinates,

y = y x

Then the differential of the new coordinates changes by

y
dy = dx
x
Notice the similarity between this and the Lorentz transformation law. The differentials transform linearly
and homogeneously, with the transformation matrix being the Jacobian matrix,
y
J =
x
This transformation is invertible provided the determinant of J , called the Jacobian, is nonzero.
There is another object with a similar transformation law. Consider a parameterized curve, x (). The
derivative,
dx
t =
d
is tangent to this curve.

If we change coordinates as before, we can write the curve in the new coordinates,
y () = y x () , and we may use the chain rule to write

x dy
t =
y d
x
= t
y

where t = dy
d is the tangent expressed in terms of the y coordinates.
Below, we make these notions precise by defining the vector space of differentials and the vector space of
tangents, but we will wait until we can make the definitions in a way that works for curved spaces as well as
flat ones.

1
1.2 Derivatives
There is an important difference between the Lorentz transformations and general coordinate transforma-
tions. For Lorentz transformations, we were able to differentiate vectors in the usual way to get other vectors.
 
1
Thus, if we wanted the derivatives of a vector v , we could simply compute them all, and have a type-
1
tensor,
v
T =
x

Then T is a tensor, because in any other Lorentz frame,

T v


=
x

= v

x
= T

This happens only because is constant. However, if a vector v transforms with a change of coordinates
as
y
v = v
x
its derivative is not a tensor. Instead,
 
y
v = v
y y x
  
x y
= v
y x x
x y x 2 y
= v + v
y x x y x x
x y x y x

  2
y
= v + v
y x x y x y x x
x y x 2 y
 
= v + v
y x x y x x
so the change is inhomogeneous.
We define the covariant derivative in such a way as to correct this problem and produce a tensor. The
idea is to add another term to the partial derivative, and let the extra term change in just the right way to
cancel the extra, inhomogeneous part. Define

D v = v + v

and require to transform so that D v transforms as a tensor when we change coordinates. That is,
we require the covariance condition,
x y
D v = D v
y x
where

D v = v + v
y
y
v = v
x
The symbol is called the connection.

2
 Substituting into covariance condition,
x y
D v = D v
y x
x
   
x y v
 
y y
v + v =
+ v
y x x x y x x
x y x 2 y
  
x y v
  
y

v + v + v =
+ v
y x x y x x x y x x
x y x 2 y y x y
 

v + v = v
y x y x x x y x
This must hold for every vector, v , so
x y x 2 y y x y
+ =
y x y x x x y x
y x y x y x 2 y
=
x y x
y x y x x
x y
x x y
x x y x 2 y
=
y x y y x
y y x y x x

x 2 y
 
x x y
=

y y x y x x
Fortunately, we do not need to use this formula very often.
The covariant derivative satisfies the basic properties of any derivation: it is linear and Leibnitz. Linearity
is easy to see, but to understand the Leibnitz (product rule) property, we need to consider how the covariant
derivative acts on general tensors. We determine this by requiring the product rule and applying it to a
product of vectors. Thus, we demand that
D u v = (D u ) v + u D v

= u + u v + u v + v

= ( u ) v + u v + u v + (u v )

= u v + u v + (u v )

 
2
If we define a type- tensor T = u v then we see from the calculation above that its covariant
0
derivative must be
D T = T + T + T
Using linearity, we can sum outer products of pairs of vectors to produce a general rank-2 tensor. It is easy
to see the pattern here: each index of T needs to be contracted with a copy of the connection. A rank-3
tensor will have three terms containing and so on for higher rank tensors.
Exercise: Prove that D (au + bv ) = aD u + bD v for arbitrary constants a, b and arbitrary vectors

u ,v .

1.3 Example: polar coordinates

There is an easier way to do this, but this example shows that the construction so far actually works.
Consider what happens when we change from Cartesian to polar coordinates. We know that in Cartesian
coordinates in flat space, the covariant derivative is the same as the usual partial derivative,
Di v j = i v j

3
If we change to polar coordinates, the components of the vector v i change to
r x r y
vr = v + v
x y
x y
v = v + v
x y
Since
p
r = x2 + y 2
y
tan =
x
we have:
r x
=
x r
r y
=
y r
for the r derivatives. For we take the differential,
y
d (tan ) = d
x
1 1 y
d = dy 2 dx
cos2 x x
1 y2
Because cos2 = 1 + tan2 = 1 + x2 , this becomes

dx + dy = d
x y
 
2 1 y
= cos dy 2 dx
x x
 
1 xdy ydx
= 2
1 + xy 2 x2
xdy ydx
=
x2 + y 2
and therefore
y
=
x x2 + y 2
y
= 2
r
x
=
y r2
The components of the vector are therefore
r x r y
vr = v + v
x y
xv x yv y
= +
r r
= v x cos + v y sin

4
 x y
v = v + v
x y
y x
= 2 vx + 2 vy
r r
vx vy
= sin + cos
r r
Now find the connection. We will learn a much easier way to compute this later. Moreover, once we have
them, we can use them to compute many properties of the geometry and all covariant derivatives in that
geometry. For now, however, we use the general transformation,

x x y x 2 y
 

=
y y x y x x

but since the original coordinate system is Cartesian, the original connection vanishes, = 0. Therefore,

x x y x 2 y
=
y y x y x x
x x 2 y
=
y y x x
x x 2 y

=
y y x x
2
Notice that = because the transformation depends on x yx . This is always the case: if the
connection is symmetric in one coordinate system, = , it remains symmetric in all coordinate
systems. In two dimensions, there are therefore 6 different terms to compute: rrr , rr , rr , r , r , .
Substituting,

x x 2 r
rrr =
r r x x
x x 2 r x y 2 r y y 2 r
= 2
r r x2 r r xy r r y 2
2
y xy x2
= cos2 3 + 2 cos sin 3 sin2 3
r  r r
1 2 1
= cos2 sin2
r r r
= 0

where we have used

2r x
=
x2 x r
1 x2
= 3
r r
y2
=
r3
2r x2
=
y 2 r3
2r x
=
yx y r
xy
= 3
r

5
The other components proceed similarly, with the only nonzero cases being r , r . There is an easy way
to see this, once we find the relationship between the connection and the metric. For now, we compute these
two directly:

x x 2 r
r =
x x
x x 2 r x y 2 r y y 2 r
= 2
x2 xy y 2
2
y xy x2
= sin2 3 2 cos sin 3 cos2 3
r r
r
= r sin4 + 2 cos2 sin2 + cos4

2
= r sin2 + cos2
= r

and
x x 2
r =
r x x
x x 2 x y 2 y x 2 y y 2
= 2
+
r x r xy r yx r y 2
x x 2xy x y x y 2
 2
y x x2 y 2
  
y y 2xy
= + + +
r r4 r r4 r r4 r r4
 2
x y2 x2 y 2
  
2xy 2xy
= r cos sin 4 + r cos2 4
r sin2 + r sin cos 4
r r r4 r
1
2 cos2 sin2 + cos2 cos2 sin2 sin2 cos2 sin2 + 2 sin2 cos2




=
r
1
cos4 + sin4 + 2 sin2 cos2


=
r
1
=
r
where we have used
2 y
=
x2 x r2
2xy
=
r4
2 2xy
= 4
y 2 r
2 2y 2
 
1
= 4
yx r2 r
2 2
x y
=
r4
We may now take the covariant derivative of the vector v i ,

Dk v i = k v i + v j ijk

There are four components,

Dr v r = r v r + v r rrr + v rr

6
 = r v r
= r (v x cos + v y sin )
v x v y
= cos + sin
r r
and
Dr v = r v + v r rr + v r
1
= r v + v
 rx
vy
 x
vy
 
v 1 v
= sin + cos + sin + cos
r r r r r r
 x x
 y y
vx vy
 
v 1 v v 1 v
= 2 + sin + 2 + cos 2 sin + 2 cos
r r r r r r r r
sin v x cos v y
= +
r r r r
and
D v r = v r + v r rr + v r
= v r + rv
 x
vy

v
= (v x cos + v y sin ) r sin + cos
r r
x y
v v
= v x sin + cos + v y cos + sin + v x sin v y cos

v x v y
= cos + sin

and finally,
D v = v + v r r + v
= v + v r r
1
= v + v r
 xr
vy

v 1
= sin + cos + (v x cos + v y sin )
r r r
x y
1 v 1 v vx vy 1 1
= sin + cos cos sin + v x cos + v y sin
r r r r r r
1 v x 1 v y
= sin + cos
r r

2 Covariant derivatives
2.1 The Laplacian
Even in flat space, the covariant derivative is useful. For example, suppose we want to compute the Laplacian
in spherical coordinates. We know that the Laplacian may be written as the divergence of the gradient,
2 f = f
In general coordinates, the divergence of a vector must be a proper contraction,
D v

7
while the gradient of a function is a form,
D f = f
No connection is required for this because the function is a scalar its value at a point does not change if
we change coordinates. To combine these, we need the metric to change the gradient to a vector,

D f = g D f

and now it becomes possible to write the Laplacian,

D D f = D g D f

In our example of polar coordinates, the metric is

 
1 0
gij =
0 r2

and its inverse is  

1 0
g ij = 1
0 r2
We have also found the components of the connection,

r = r

and
1
r =
r
The Laplacian is therefore,

D D f = D g D f

= g D f + g D f

= + g D f

The first term expands as

g D f = r g r D f + g D f

= r (g rr Dr f + g r D f ) + (g r Dr f + g D f )
 
1
= r (r f + 0 D f ) + 0 r f + 2 f
r
2 1 2
= f+ 2 f
r2 r 2
which we know to be correct.

2.2 The connection at a point

We have shown that the connection transforms as
x x y x 2 y
 
=
y y x y x x

Now, we know that a manifold is a space which is Rn to lowest order in a sufficiently small neighborhood
of any point P. This means that, again in a sufficiently small neighborhood, we can choose Cartesian

8
coordinates where the connection vanishes. There are deviations as we move away from P, but the fact
remains that there exists a coordinate system in which the connection vanishes at P.
We can show this explicitly by asking for a coordinate transformation from a general to a set of new
coordinates where is simpler. Imposing this condition gives

x x y x 2 y
 
=

y y x y x x
and since the matrices
y
J =
x
x
J =
y
are invertible, we may eliminate the leading factors to get
x 2 y
=
y x x
2
y y
= 0
x x x
Now expand y , and around the point P (with coordinates x
0)

1 1
y = x
0 + a (x x0 ) + b (x x0 ) (x x0 ) + c (x x0 ) (x x0 ) (x x0 ) . . .
2 3!

= (P) + (x x0 ) (P) + . . .
x

= (P) + (x x0 ) (P) + . . .
x


We may choose teh constant coefficients in the expansion of y x in any way we choose.
Writing out the lowest order terms,

b a (P) = (P)

c (x x0 ) b (x x0 ) (P) a (x x0 ) (P) = (x x0 ) (P)
x x
Choose b = a (P). This makes the new connection vanish at x = x
0,

(P) = 0

At next order we have

c (x x0 ) a (P) (P) (x x0 ) a (x x0 ) (P) = (x x0 ) (P)
x x
Since a must be invertible, we multiply by its inverse, a , leaving

a c (P) (P) (P) = a (P)
x x
Our only free choice here is a c , but this must be totally symmetric in the three lower indices. We
therefore have only four independent choices for the lower three indices, and four for , giving a total of 16
constants. The remaining terms on the left include constants x (P). The symmetry on means that
(P) contains 40 constants, and the derivatives increase this to 160. Therefore, we do not have enough
choice left to eliminate the first derivatives of the new connection, so a x (P) is in general nonzero.

9
2.3 The covariant derivative of the metric
The covariant derivative is required to satisfy the product rule for differentiation. Therefore, if we differentiate
the inner product of two vectors, we have

D (g u v ) = u v D g + g v D u + g u D v

and the question immediately arises: what is the covariant derivative of the metric, D g ?
1
In flat space and Cartesian coordinates, the metric is the identiy matrix, gij = ij =
.. ,

.
1
and its derivatives are zero,
i gjk = 0
This relationship is identical to
Di gjk = 0
in Cartesian coordinates, but the second form is true independent of coordinate system. Therefore, in flat
space we expect the covariant derivative of the metric to vanish.
We have a similar result in curved space because curved manifolds are Euclidean in a neighborhood of a
point. In general, the covariant derivative of the metric is

D g = g g g

As we showed above, at any point P, there exists a change of coordinates

1 1
y = x
0 + a (x x0 ) + b (x x0 ) (x x0 ) + c (x x0 ) (x x0 ) (x x0 ) . . .
2 3!
where we may choose the coefficients b so that the connection vanishes at P. This choice leaves the linear
coefficients, a , free. In these new coordinates, the new metric at P is given by

x x
g (y) = g
y y
g (P) = a a g (P)

and we may use the matrix a to transform the original g to orthonormal form, . Then

D g (y) = g (P) g (P)

= 0

Since this is a tensor equation, it must hold in every coordinate system, and we conclude

D g = 0

regardless of our choice of chart.

2.4 Connection in terms of the metric

The vanishing of the covariant derivative of the metric allows us to find an expression for the connection in
terms of the metric. Expanding the covariant derivative,

D g = g g g
= g

10
where we define g . Equating to zero and moving the connection terms to the left, we have

+ = g

This relationship holds no matter what we name the indices, as long as we maintain the correspondence of
free indices across the whole equation. We may therefore equally well write

+ = g
+ = g

For any choice of , , all three of these must hold. Combining the three equations, we add the first two
and subtract the third,

+ + + = g + g g
( + ) + ( ) + ( ) = g, + g, g,

Now, because the connection is symmetric on the last two indices, = , the left side simplifies to

2 = g, + g, g,
1
g = g (g, + g, g, )
2
and therefore,
1
= g (g, + g, g, )
2
This immensely simplifies computing the connection. For example, in polar coordinates the metric is
 
1 0
gij =
0 r2

so the only component with nonvanishing derivative is g = r2 . This means that g, + g, g, will
vanish unless two of the indices are and one is r. There are only three ways to make this happen:
1
r = (gr, + gr, g,r )
2
1
r r2


=
2
= r
1
r = r = (gr, + g,r gr, )
2
1
= (g,r )
2
= r

Using the inverse  

1 0
g ij = 1
0 r2
we find the same result we struggled to get before,

r = g rr r + g r
= r
r = r = g r
1
=
r

11
2.5 Conditions on the metric and curvature at a point
Now that we know that, if D g = 0, then the connection can be built out of first derivatives of the metric,
we can prove the results above in a more direct way. Consider a Taylor series for a coordinate transformation
of the metric. Expand the old coordinates in terms of the new, and the old and new metrics,

1 1
x (y) = x
0 + a (y y0 ) + b (y y0 ) (y y0 ) + c (y y0 ) (y y0 ) (y y0 ) . . .
2 3!
1
g (x) = g (P) + g (P) (y y0 ) + g (P) (y y0 ) (y y0 ) + . . .
2
1
g (y) = g (P) + g (P) (y y0 ) + g (P) (y y0 ) (y y0 ) + . . .

2
where the coefficients in y are symmetric,

b = b
c = c()

The coordinate transformation matrix is

x 1

= a + b (y y0 ) + c (y y0 ) (y y0 ) . . .
y 2!
Then, expanding the new metric in terms of the old, we have
x x
g (y) = g (x)
y y
g (P) + g (P) (y y0 )
 
1 1
+ g (P) (y y0 ) (y y0 ) + . . . = g (P) + g (P) (y + g (P) (y y0 ) (y y0 ) + . . .

y0 )
2 2
 
1
a + b (y y0 ) + c (y y0 ) (y y0 ) . . .
2!
 


1


a + b y y0 + c y y0 (y y0 ) . . .
2!
Collect all zeroth order terms,

g (P) = g (P) a a

and first order terms,

 
g (P) (y y0 )= g (P) a b y y0 + b (y y0 ) a + g (P) (y y0 ) a a

g (P) = g (P) a b + b a + g (P) a a

Finally, collect the second order terms,


1 1
g (P) (y y0 ) (y y0 ) a c y y0 (y y0 )


= g (P)
2 2!


1
+b (y y0 ) b y y0 + c a (y y0 ) (y y0 )
2!
 
+ g (P) (y y0 ) a b y y0 + a b (y y0 )

1
+ g (P) (y y0 ) (y y0 ) a a
2

12
so that
 
1 1 1
g (P) = g (P) a c + b b + c a
2 2! 2!
+ g (P) a b + a b

1
+ a a g (P)
2
Now we ask how simple we can make the expansion of g . First, the transformation
g (P) = g (P) a a
where a is arbitrary is a general linear transformation of the symmetric matrix g (P), and this is sufficient
to make g (P) orthonormal,
g (P) =
This determines the coefficients a up to a Lorentz transformation. For the next order, we have

g (P) = g (P) a b + b a + g (P) a a

and the b have just enough freedom to make the right side vanish. To see this, define

B g (P) a b

and notice that since b is arbitrary, so is B . Then the right side is

B + B + g (P) a a
The sum
B + B
is symmetric on , but otherwise arbitrary, while the symmetry of the metric makes g (P) a a
symmetric on as well. Therefore, we may choose B to make the right side vanish, so that the new
metric has vanishing derivatives at P,
g (P) = 0
Now look at the second order equation. Defining C = g (P) a c , the right side is
1 1
(C + C ) + g (P) a b + a b + b b + a a g (P)


2! 2
The last term contains g (P). This is symmetric on and on so the 10 components of the metric
each have 10 derivatives, giving 100 degrees of freedom. However, C is built from c which is totally
symmetric on . This means that there are 456
123 = 20 independent ways to choose these three indices, for
each of the 4 values of , giving only 80 degrees of freedom. In general, we cannot make this side vanish by
any choice of c , and the second derivatives of the metric remain.
In conclusion, our coordinate choice can make the metric orthonormal and the first derivatives of the
metric vanish at P,
g (P) =
g (P) = 0
The second condition is equivalent to the vanishing of the Christoffel connection, since
1
(P) = (g, (P) + g, (P) g, (P))
2
= 0
This reproduces the conclusions of the previous sections.

13