Connections and covariant derivatives

Problem: How do we dierentiate a vector or tensor eld and preserve the tensor property?
There are a number of levels at which this can be discussed. The simplest, as mentioned in
the previous section, is to note that the components of the obvious thing, namely V
a
/x
b
,
do not behave properly under change of basis - they are not the components of a tensor of
rank (1, 1).
In at space, theres no problem; we just dierentiate. What were really doing is moving
all the vectors parallel to themselves to a single vector space, and doing the dierentiation
there. In a curved space, this is not possible: its easily seen (as shown in class) that you
can only move a vector (or tensor) parallel to itself along a curve. You cannot propagate it
into an open set in such a way that its parallel unless the space is at.
With this in mind, we will try to see what we can do, instead of what we cant. Our
considerations are purely local, and dont involve the metric tensor initially.
Propagation of a vector along a curve
Suppose t (t) is a curve in some open set U with local coordinates (x
a
), so the curve is
given parametrically by t x
a
(t) : 1 a n. Suppose p, q, and r are any three points on
the curve. We want to dene a map, called parallel propagation and denoted
pq
such
that
1.
qp
: T
p
(U) T
q
(U) is a vector space isomorphism depending only on the points
comprising the curve, and not on the parametrization.
2.
rp
=
rq

qp
for all points p, q, r on .
3. Parallel propagation preserves the scalar product: if V, W T
p
(U), then g(V, W) =
g(
qp
(V ),
qp
(W)).
Since the curve is given parametrically, lets use the symbol
t
to denote parallel propagation
from T
p
= T
x(0)
to T
x(t)
. Let V
0
T
p
, and let V (t) =
t
V
0
. (We are assuming that parallel
propagation exists, and were going to deduce some equations that must be satised.) then
dV
dt
=
d
t
dt
V
0
=
d
t
dt

1
t
(
t
V
0
) =

d
t
dt

1
t

V
t
, or
dV
dt
+ V = 0, where =
d
t
dt

1
t
In local coordinates, this will read
dV
a
dt
+
a
b
V
b
= 0. (1)
1
Now will depend upon the path: it will be a function of x, x, x, . . ., all of which are functions
of the parameter t. The simplest assumption, and the one that will work for us, is that the
matrix depends only on x(t) and x(t). Moreover, the dependence on x is assumed to be
linear. If the curve is reparametrized with u, we nd that (

= d/du):
dV
a
du
+
a
b
(x, x

)V
b
=
dV
a
dt
dt
du
+
a
b
(x, x
dt
du
)V
b
=

dV
a
dt
+
a
b
(x, x)V
b

dt
du
(linearity of )
That is, under our assumptions, parallel propagation (indicated by the vanishing of these
quantities) is independent of the parametrization, and is said to dene a linear or ane
connection in U. Specically, this means that

a
b
(x, x) =
a
b
(x, x
c

c
) =
a
b
(x,
c
) x
c
,
and we dene the Christoel symbols

a
bc
(x) =
a
b
(x,
c
). (2)
Equation (1) now reads
dV
a
dt
+
a
bc
V
b
x
c
= 0; 1 a n . (3)
This coupled system of n rst order ODEs denes the operation of parallel transport along
the curve . Note that the Christoel symbols are functions of position only and can be
dened in the whole of U in principle. If the curve is given, then x(t) and x(t) are known,
the s are known, and Eqn (3) is a standard initial value problem with unique solutions
- the IC is V (0) = V
0
. Moreover, the equations are linear and homogeneous in V , as they
should be. A vector eld on satisfying eqn (3) is said to be parallel along . So, given
initial conditions V
0
at a point p ,
Theres a unique solution V (t) to the system (3).
If V
0
T
p
U, and q = x(t), then
qp
(V
0
) = V (t), where is the parallel propagator.
If V
0
, W
0
T
p
U, and V (t), W(t) are the corresponding solutions to (3), then
qp
(c
1
V
0
+
c
2
W
0
) = c
1
(V
0
) + c
2
(W
0
), so
qp
is linear, as advertised.
The matrix is a matrix of 1-forms, since each entry
a
b
is a linear function on the tangent
space. However, itself is not a tensor; if we change coordinates to ( x
a
), then of course well
require the existence of such that, for the same parallel vector eld V
1
d

V
dt
+

V = 0.
1
To take a break from the indices, which are a pain to type, we are using matrix notation here. So V is
the vector of components and not, as weve previously used this symbol, V
a

a
.
2
If P
a
b
= x
a
/ x
b
is the change of basis matrix, then in matrix notation we have
d

V
dt
=
d
dt
(P
1
V )
=

P
1
V + P
1

V
=

P
1
V P
1
V
=

V
= P
1
V
Equating the third and fth lines above, we get

P
1
P
1
= P
1
,
or
= P
1
P

P
1
P
= P
1
P + P
1

P (4)
where in the last line, weve used the fact that
d
dt
(P
1
P) = 0 =

P
1
P + P
1

P.
Its the last term of eqn (4) which shows that is not a tensor; if you write it out, youll nd
that it contains the second derivatives of the coordinates. This is the same correction term
we had to add to
a
V
b
to get a tensor in the previous section; its the term which vanishes
if we only use linear coordinate systems.
Denition: For any vector eld V dened along the curve x(t), the covariant derivative
of V along the curve is the vector eld
DV
dt
=
dV
dt
+ V.
In components, this reads
DV
a
dt
=
dV
a
dt
+
a
b
V
b
=
dV
a
dt
+
a
bc
V
b
x
c
.
Proposition: Under a change of coordinates, where P is the change of basis matrix, we
have
D

V
dt
= P
1
DV
dt
.
In other words, the covariant derivative of V along is another vector eld along .
3
Proof: We have
D

V
dt
=
d

V
dt
+

V
=
d
dt
(P
1
V ) +

P
1
P + P
1
dP/dt

P
1
V
=
dP
1
dt
V + P
1
dV
dt
+

P
1
P
dP
1
dt
P

P
1
V
= P
1

dV
dt
+ V

= P
1
DV
dt
.
Of course, if V is a parallel vector eld along , then DV/dt = 0, as above. In any case, we
can just compute DV/dt to get another vector eld.
Exercise: What is the covariant derivative of
c
along the curve? Write it out as a vector,
not just its components. Reverting to the usual notation, if V = V
a

a
, show that
DV
dt
=
D
dt

V
a

=
dV
a
dt

a
+ V
a
D
a
dt
Suppose now that V is a vector eld dened and smooth in some neighborhood of instead
of just on . Then we can expand dV
a
/dt and write
DV
a
dt
=
dV
a
dt
+
a
bc
V
b
x
c
=
V
a
x
c
x
c
+
a
bc
V
b
x
c
=

V
a
x
c
+
a
bc
V
b

x
c
Since the left hand side of this is the component of a tensor, and x
c
is the component of a
tensor, we conclude that the quantity in parentheses is likewise a tensor. It has nothing to
do with the specic path , since it only depends on x.
Denition: If V is a vector eld, dened in some open set U with local coordinates (x
a
),
then the quantities

a
V
b
=
V
b
x
a
+
b
ca
V
c
(5)
are the components of a tensor eld of rank (1, 1) called the covariant derivative of V .
Putting in the basis elements, we have
V =
a
V
b
dx
a

b
(6)
4
We can extend the covariant derivative to operate on tensors of any rank by requiring that it
be appropriately linear, that it satisfy Leibnizs rule, and that it commute with contraction:
Denition: A covariant derivative is an operator on tensor elds which satises the
following conditions:
1. If T is of rank (r, s), then T is of rank (r, s + 1); the covariant rank increases by 1.
2. For any function f, (f) = df.

a
(f) =
a
(f)

3. For any function f and tensor T, (fT) = df T + fT.

a
(fT

) =
a
(f)T

+
f
a
T

.
4. More generally, for any tensors S and T, (S T) = S T + S T.
Assuming we know how to dierentiate vector elds, we can use the above to determine the
covariant derivative of a one form . Consider the tensor eld V with components
b
V
c
.
From (2) and (3) above, we have

a
(
b
V
c
) = (
a

b
)V
c
+
b
(
a
V
c
)
= (
a

b
)V
c
+
b

a
V
c
+
c
da
V
d

Contracting on b and c, this reads

a
(
b
V
b
) = (
a

b
)V
b
+
b

a
V
b
+
b
da
V
d

.
But now the left hand side is just a function and by (2) we have
a
=
a
on functions. The
whole thing now becomes

a
(
b
V
b
) = (
a

b
)V
b
+
b
(
a
V
b
)
= (
a

b
)V
b
+
b
(
a
V
b
) +
b

b
da
V
d
= (
a

b
)V
b
+
b
(
a
V
b
) +
d

d
ba
V
b
,
where we have switched two indices in the last term. Equating the rst and third lines of
this, cancelling the two identical terms and rearranging gives

d
ba

V
b
= 0.
Since this holds for any vector eld V , the quantity in parentheses must vanish, and we
obtain

b
=
a

d
ba

d
(7)
So now we know how to handle both upper and lower indices, and this means we can take
the derivatives of tensor elds of any rank. See the exercises below.
5
Note: The formulas are complicated by the Christoel symbols because the standard bases
for TU and T

U are not parallel: As an example, for a one-form , we can also write

c
dx
a
dx
c
+
c
(dx
c
)
To compute (dx
c
) =
a
dx
a
, where
a
=
c
a
, we use (7) and nd that

a
=
b

c
a

d
ab

c
d
=
c
ab
.
So
(dx
c
) =
b

a
dx
b
dx
a
=
c
ab
dx
b
dx
a
.
Plugging this into the expression above, and massaging the indices, we nd
=

c
ab

dx
b
dx
a
.
And we see that the complications come from the fact that the covariant derivatives of the
basis one-forms {dx
a
} are not zero.
Denition: A tensor eld T is parallel if T = 0.
Exercise:
1. Compute g, where g = g
ab
dx
a
dx
b
using the same techniques as in the above
example. What is
c
g
ab
?
2. Without doing any work, guess the form of
a
T
bcd
.
3.
a
T
b
c
=
a
T
b
c
+
b
da
T
d
c

d
ca
T
b
d
.
Riemannian geometry
The alert reader has surely noticed that thus far weve made no use of the metric tensor. So
there are no constraints on other than that it behave reasonably as a dierential operator.
And correspondingly, there is no unique way to choose a connection on U. But if we have a
non-degenerate metric g, then its natural to ask that parallel propagation respect the scalar
product.
Denition: (Notation) Suppose V is a vector eld along some curve . If V is smooth, it
is always possible to extend V smoothly to a neighborhood of . In this case, we can write
DV
b
dt
= T
a

a
V
b
, where T = T
a

a
6
is the tangent vector to the curve. See eqn (5) and the derivation above it. We can also
denote this, in index free notation, by
T
V , where

T
V = T
a
(
a
V
b
)
b
=
a
V
b
dx
a
(T)
b
Its clear that the result is independent of the particular smooth extension of V to a neigh-
borhood of the curve, so well often use this notation without mentioning the extension
specically. The same is true for tensor elds of any rank.
Denition: The connection is said to be a metric connection, or compatible with
the metric if the scalar product between two parallel-propagated vector elds is constant.
Suppose that U
0
, V
0
T
p
U, is a curve at p and U, V are parallel along the curve, agreeing
with U
0
, V
0
at t = 0. If the connection is compatible with the metric, the function g(U, V )
is constant along . So
d
dt

g(U, V )

= (
T
g)(U, V ) + g(
T
U, V ) + g(U,
T
V ) = 0.
Since U and V are parallel along ,
T
U =
T
V = 0. At the point p,
(
T
g)(p)(U
0
, V
0
) = 0.
But U
0
and V
0
are arbitrary in T
p
U. Therefore
T
g = 0 at p. But the curve is arbitrary,
so T is arbitrary and therefore g(p) = 0. Since p is arbitrary as well, we must have g = 0
everywhere in U. That is, the metric is covariantly constant (or parallel). Writing this out
in detail,

a
g
bc
=
a
g
bc

d
ba
g
dc

d
ca
g
bd
= 0 (8)
We now make one last assumption, that the connection is symmetric:
Denition: is symmetric if, in any coordinate system,
a

b
=

a
. Or equivalently,
the connection is symmetric if the Christoel symbols are symmetric in the two lower indices:

a
bc
=
a
cb
.
If is symmetric and compatible with the metric, the connection turns out to be unique:
From Eqn (8),
a
g
bc
+
b
g
ac

c
g
ab
= 0. Using the symmetry of the connection, this yields
2
d
ab
g
dc
=
a
g
bc
+
b
g
ac

c
g
ab
,
Now multiply both sides by (1/2)g
ce
, obtaining

e
ab
=
1
2
g
ce

a
g
bc
+
b
g
ac

c
g
ab

(9)
Denition: Given the metric, the connection uniquely determined by Eqn (9) is called the
Riemannian or Levi-Civita connection. Dierential geometry, done with this connection,
7
is called Riemannian geometry, although some authors reserve this term for the case
of positive-denite metrics, and call it pseudo-Riemannian geometry if the metric is
Lorentzian.
Exercise: On a sphere S
2
of radius r, the components of the metric tensor in the (, )
coordinate system are given by
(g
ab
) =

r
2
0
0 r
2
sin
2

.
Find the Christoel symbols, using Eqn (9).
8