Chapter 3

Curvature

Problem Set #3: 3.1, 3.3, 3.5, 3.13, 3.15 (Due Monday Oct. 28th)
Midterm-exam: October 30th

3.1 Covariant derivative

In the previous chapter we have shown that the partial derivative of a non-
scalar tensor is not a tensor (see (2.34)). It does not transform as a tensor
but one might wonder if there is a way to define another derivative operator
which would transform as a tensor and would reduce to the partial derivative
in Minkowski space (note that exterior derivative does transform as a tensor,
but does not reduce to partial derivative in the limit of flat space). The
desired derivative operator (called covariant derivative and denote by ∇)
can be constructed by by enforcing certain properties such as linearity, i.e.
∇(T + S) = ∇T + ∇S, (3.1)
and product rule, i.e.
∇(T ⊗ S) = ∇T ⊗ S + T ⊗ ∇S. (3.2)
One can show (see Wlad’s book) that if the operator obeys product rule
than it can be written as a partial derivative plus a linear correction whose
coefficients are called the connection coefficients,
∇µ V ν = ∂µ V ν + Γνµλ V λ . (3.3)
Then the transformation properties of Γνµλ (which does not have to and will
not transform as a tensor) can be determined by demanding that ∇µ V ν
transforms as a (1, 1) tensor, i.e.
′
ν′ ∂xµ ∂xν
∇µ′ V = µ′ ∇µ V ν . (3.4)
∂x ∂xν

29
CHAPTER 3. CURVATURE 30

By substituting (3.3) in (3.4) we obtain

′
′ ′ ′ ∂xµ ∂xν !
∂µ′ V ν + Γνµ′ λ′ V λ = ν ν λ
"
∂µ V + Γ µλ V
# ν′ $ # λ′ $ ∂xµ ∂xν
′

µ ′ ′
∂x ∂ ∂x ν ν′ ∂x λ ∂xµ ∂xν ∂ ν ∂xµ ∂xν ν λ
V + Γ µ′ λ′ V = V + Γ V
∂xµ′ ∂xµ ∂xν ∂xλ ∂xµ′ ∂xν ∂xµ ∂xµ′ ∂xν µλ
′ λ′ ′
∂xµ ∂ ∂xν ν ν ′ ∂x λ ∂xµ ∂xν ν λ
V + Γ µλ V = Γ V
∂xµ′ ∂xν µλ
′ ′
∂xµ′ ∂xµ ∂xν ∂xλ
′ ′
′ ′ ∂xλ ∂xµ ∂xν ν λ′ ∂xλ ∂xµ ∂ ∂xν λ′
Γνµ′ λ′ V λ = Γ V − (3.5)
V .
∂xλ′ ∂xµ′ ∂xν µλ ∂xλ′ ∂xµ′ ∂xµ ∂xλ
′
Since this must be true for all V λ we get a transformation law for the con-
nection coefficients
′ ′
′ ∂xλ ∂xµ ∂xν ν ∂xλ ∂xµ ∂ 2 xν
Γνµ′ λ′ = λ′ µ′ Γ µλ − . (3.6)
∂x ∂x ∂xν ∂xλ′ ∂xµ′ ∂xµ ∂xλ

Similarly we can show that a different set of coefficients Γ̃λµν should be

used to define a covariant derivative of a one form,

∇µ ων = ∂µ ων + Γ̃λµν ωλ (3.7)

where Γ̃λµν although so far unrelated does transforms exactly as Γλµν in (3.6).
To establish a relation between Γλµν and Γ̃λµν we demand that the covariant
derivative reduces to partial derivative for scalars

∇µ φ = ∂µ φ. (3.8)

Thus

∂µ ωλ V λ = ∇µ ωλ V λ
! " ! "

(∂µ ωλ ) V λ + ∂µ V λ ωλ = (∇µ ωλ ) V λ + ∇µ V λ ωλ
! " ! "

(∂µ ωλ ) V λ + ∂µ V λ ωλ = (∂µ ωλ ) V λ + Γ̃σµλ ωσ V λ + ∂µ V λ ωλ + Γσµλ ωσ V λ

! " ! "

Γσµλ ωσ V λ = −Γ̃σµλ ωσ V λ . (3.9)

Since ωσ and V λ are arbitrary

Γσµλ = −Γ̃σµλ . (3.10)

and therefore
∇µ ων = ∂µ ων − Γλµν ωλ . (3.11)
CHAPTER 3. CURVATURE 31

The covariant derivatives of an arbitrary rank tensor are given by

∇σ T µ1 ...µkµk ...νl = ∂σ T µ1 ...µkµk ...νl + Γµσλ1 T λµ2 ...µkν1 ...νl + ...

−Γλσν1 T µ1 ...µkν1 ...νl − .... (3.12)

In general relativity the 43 = 64 independent connection coefficients of Γλµν

are uniquely specified by the metric gµν . This is accomplished by demanding
that the connection coefficient is torsion-free, i.e.

Γλµν = Γλ(µν) , (3.13)

and metric-compatible, i.e.

∇ρ gµν = 0. (3.14)

The torsion-free implies, for example, that the antisymmetrized covariant

derivative is also the exterior derivative, i.e.

∇[µ ων] = ∂[µ ων] − Γλ[µν] ωλ = ∂[µ ων] .

The metric-compatibility implies a number of nice properties. First of all the

covariant derivative of inverse metric also vanishes, i.e.

∇ρ g µν = 0 (3.15)

and thus the raising an lowering operators commute with covariant derivative

∇ρ V µ = ∇ρ (g µν Vν ) = g µν ∇ρ Vν . (3.16)

Moreover the torsion-free and metric-compatible properties of the connection

single out a unique connection known as the Christoffel (or Levi-Civita)
connection of Christoffel symbol. The formula for the Christoffel symbol
can be derived by from

0 = ∇ρ gµν = ∂ρ gµν − Γλρµ gλν − Γλρν gµλ (3.17)

0 = ∇µ gνρ = ∂µ gνρ − Γλµν gλρ − Γλµρ gνλ (3.18)
0 = ∇ν gρµ = ∂ν gρµ − Γλνρ gλµ − Γλνµ gρλ (3.19)

by subtracting (3.18) and (3.19) from (3.17) and using (3.13),

∂ρ gµν − ∂µ gνρ − ∂ν gµρ = Γλρµ gλν − Γλµρ gνλ + Γλρν gµλ − Γλνρ gλµ − Γλµν gλρ + Γλνµ
! " ! " ! "
(3.20)
gρλ
∂ρ gµν − ∂µ gνρ − ∂ν gµρ = −2Γλµν gλρ (3.21)
CHAPTER 3. CURVATURE 32

or
1
Γλµν = g λρ (∂µ gνρ + ∂ν gµρ − ∂ρ gµν ) . (3.22)
2
In Minkowski space described by Cartesian all of the Christoffel symbols
vanish, but this does not have to be the case in curvilinear coordinates. As
an example consider a two dimensional Euclidean space described by polar
coordinates with metric
ds2 = (dr)2 + r 2 (dθ)2 . (3.23)
The non-vanishing components of the inverse metric are
g rr = 1 (3.24)
g θθ = r −2 (3.25)
and for example the connection coefficient
1 rρ
Γrrr = g (∂r grρ + ∂r gρr − ∂ρ grr ) =
2
1 rr 1
= g (2∂r grρ − ∂ρ grr ) + g rθ (2∂r grρ − ∂ρ grr ) =
2 2
1 1
= 1 (2∂r grr − ∂r grr ) + 0 (2∂r grσ − ∂σ grr ) = 0 (3.26)
2 2
but
1 rρ
Γrθθ = g (∂θ gθρ + ∂θ gρθ − ∂ρ gθθ ) =
2
1 rr 1
= g (2∂θ gθr − ∂r grr ) + g rθ (2∂θ gθθ − ∂θ gθθ ) =
2 2
1 1
= 1 (2∂θ gθr − ∂r grr ) + 0 (2∂θ gθρ − ∂ρ gθθ ) =
2 2
1
= − ∂r grr = −r. (3.27)
2
It is a straightforward exercise to find all other coefficients,
Γrθr = Γrrθ = 0 (3.28)
Γθrr = 0 (3.29)
1
Γθrθ = Γθθr = (3.30)
r
θ
Γθθ = 0. (3.31)
Just like one can make the connection coefficients to be non-zero in flat
space it is possible to make the connection coefficients to vanish at some
point curved space but not everywhere.
CHAPTER 3. CURVATURE 33

Since the covariant derivatives of a vector is

∇µ V µ = ∂µ V µ + Γµµλ V λ (3.32)

and (one can also show)

1 %
Γµµν = % ∂ν |g| (3.33)
|g|

then we obtain a useful expression

1 &% '
∇µ V µ = % ∂µ |g|V µ . (3.34)
|g|

3.2 Parallel transport

Up until now we were not able to compare different tangent vectors at dif-
ferent points since they where not elements of the same vector space. With
the help of the connection coefficients we can continuously move the tangent
vectors (or higher rank tensors) from on point to another (or parallel trans-
port), but the resulting vector will usually depend on the path along which
it was moved. (Think about a parallel transport of a vector on the surface
of a sphere to convince yourself that the parallel transported vector would
depend on the path.) This is a generic property of curved spaces which is
why it makes no sense to ask what is a relative velocity of two particles in
two distinct points. In fact interpreting the cosmological expansion of space
by galaxies receding away at a speed defined by the redshift is incorrect and
can lead to paradoxes involving superluminal velocities. Of course in cos-
mology nothing is receding away, but the metric between galaxies changes
which causes the light between the object to change the wavelength (i.e. to
redshift).
In flat space the parallel transport of a tensor along a parametrized curve
xµ (λ) is given by the requirement

dxµ
# $
D
T µ1 ...µk ν1 ...νl = ∂µ T µ1 ...µkν1 ...νl = 0 (3.35)
dλ dλ

which is generalized to curved spaces as

dxµ
# $
D
T µ1 ...µk ν1 ...νl = ∇µ T µ1 ...µkν1 ...νl = 0. (3.36)
dλ dλ
CHAPTER 3. CURVATURE 34

This is the parallel transport equation which, for example, take the fol-
lowing form for vectors
D µ d µ dxσ ρ
V = V + Γµσρ V = 0. (3.37)
dλ dλ dλ
It follows, for example, that the inner product of two parallel transported
vectors is preserved, i.e.
D D D D
(gµν V µ W ν ) = (gµν ) V µ W ν + gµν (V µ ) W ν + gµν V µ (V W ν ) = 0.
dλ dλ dλ dλ
(3.38)
Next we will obtain a formal solution of the parallel transport equa-
tion (3.37). Our task is to find the so-called parallel propagator matrix
P µρ (λ0 , λ) along trajectory γ(λ) such that
V µ (λ) = P µρ (λ, λ0 )V ρ (λ0 ). (3.39)
If we define a transition matrix
dxσ
Aµρ (λ) ≡ −Γµσρ (3.40)
dλ
then (3.37) can be written as a Schrodinger equation
d µ
V = Aµρ V ρ . (3.41)
dλ
By substituting (3.39) into (3.41) we get
d ( µ
P ρ (λ, λ0 )V ρ (λ0 ) = Aµσ P σρ (λ, λ0 )V ρ (λ0 )
) ( )
(3.42)
dλ
d µ
P (λ, λ0 ) = Aµσ P σρ (λ, λ0). (3.43)
dλ ρ
By integrating both side we get
* λ
P µρ (λ, λ0 ) = δρµ + Aµσ (η)P σρ (η, λ0)dη (3.44)
λ0

which can be solved by iteration

* λ * λ * η1
µ µ µ
P ρ (λ, λ0 ) = δρ + A ρ (η1 )dη1 + Aµσ (η1 )Aσρ (η2 )dη1 dη2 +... (3.45)
λ0 λ0 λ0

or in matrix notation
* λ * λ * η2
P (λ, λ0 ) = 1 + A(η1 )dη1 + A(η2 )A(η1 )dη1 dη2 + ... (3.46)
λ0 λ0 λ0
CHAPTER 3. CURVATURE 35

which can be simplified using a path-ordered product of matrices, P[A(ηn )A(ηn−1 )...A(η1 )],
as ∞
1 λ
+ *
P (λ, λ0) = 1 + P[A(ηn )...A(η1 )]dη1 dη2 ... (3.47)
n=1
n! λ0
which is the series expansion of an exponential,
#* λ $
P (λ, λ0) = P exp A(η)dη . (3.48)
λ0

Turning back to the components notation gives us

# * λ
dxσ
$
P µν (λ, λ0) = P exp − Γµσν dη . (3.49)
λ0 dη

In quantum field theory the same formula is known as Dyson’s formula which
is due to the fact that (3.41) is mathematically equivalent to the Schrodinger
equation. The parallel transport transformation around a closed loop is called
the holonomy of the connection around the loop. For the metric-compatible
connections the group of holonomy transformations at a point is a Lorentz
group. Note that the knowledge of the holonomy group at each point is
sufficient to determine the metric and thus might question what is more
fundamental: holonomies of all loops or metric at all point. In the canonical
quantum gravity the metric is treated as fundamental, when in the loop
quantum gravity the holonomies are more fundamental.

3.3 Geodesics
Geodesic connecting two points can be defined as a path which parallel trans-
ports its own tangent vector, i.e.

D dxµ d2 xµ ρ
µ dx dx
σ
= + Γ ρσ = 0. (3.50)
dλ dλ dλ2 dλ dλ
This is the geodesic equation which is obtained by substituting the tangent
µ
vector dx
dλ
into (3.37). For the metric-compatible connection coefficients the
geodesic is also the path of the largest proper time (there is no smallest
proper time since one can always construct a path of zero length made out
of null segments) defined as
,
dxµ dxν
*
τ [x] = dλ −gµν (3.51)
dλ dλ
CHAPTER 3. CURVATURE 36

that can be obtained using variational derivative, i.e.

δτ [x]
= 0. (3.52)
δxµ
By Taylor expanding the metric

gµν (x + δx) ≈ gµν (x) + δxσ ∂σ gµν (x) (3.53)

we can vary the proper time we get

- -
µ +δxµ ) d(xν +δxν )
δτ δτ [x + δx] − δτ [x]
* −gµν dλ dλ − − (gµν + δxρ ∂ρ gµν ) d(x dλ
dxµ dxν
dλ
= = dλ
δxσ δxσ δxσ
- -
µ dxν µ dxν µ dxν µ dxν µ dδxν "
−gµν dx − −gµν dx − δxρ ∂ρ gµν dx − gµν dδx − gµν dx
* !
dλ dλ dλ dλ dλ dλ dλ dλ dλ dλ
= dλ
δxσ
- . - /
dxµ dxν dxµ dxν −1 dxµ dxν dδxµ dxν
! " ! "
* −gµν dλ dλ 1 − 1 + −gµν dλ dλ ρ
−δx ∂ρ gµν dλ dλ − 2gµν dλ dλ
= dλ
δxσ
- 0 ! "1
µ dxν dxµ dxν −1 dxµ dxν dδxµ dxν
−gµν dx 1
" ! ρ
*
dλ dλ 2
−g µν dλ dλ
−δx ∂ρ g µν dλ dλ
− 2g µν dλ dλ
= dλ σ
δx
dxµ dxν −1/2 dxµ dxν µ dxν "
! 1 ρ
− 2 δx ∂ρ gµν dλ dλ − gµν dδx
! "
−gµν dλ dλ
*
dλ dλ
= dλ (3.54)
δxσ
By changing the integration variable from λ to proper time
$−1/2
dxµ dxν
#
dλ = dτ −gµν (3.55)
dλ dλ
we get
! 1 ρ µ dxν µ dxν "
δτ − 2 δx ∂ρ gµν dx − gµν dδx
*
dτ dτ dτ dτ
= dτ
δxσ δxσ
dxµ dxν ν"
! 1
− 2 ∂ρ gµν dτ dτ δxρ + dτd gµν dx
! "
δxµ
*
dτ
= dτ
δxσ
dx dxν
µ
dxν
# # $$
1 d
*
= dτ − ∂σ gµν + gσν (3.56)
2 dτ dτ dτ dτ
CHAPTER 3. CURVATURE 37

or by demanding that δτ /δxσ vanishes for all variations

dxµ dxν dxν

# $
1 d
− ∂σ gµν + gσν = 0
2 dτ dτ dτ dτ
µ ν µ ν
1 dx dx dx dx d2 xν
− ∂σ gµν + ∂µ gσν + gσν 2 = 0
2 dτ dτ dτ dτ dτ
d2 xν 1 dxµ dxν
gσν 2 + (−∂σ gµν + ∂µ gσν + ∂ν gµσ ) = 0
dτ 2 dτ dτ
d2 xρ 1 ρσ dxµ dxν
+ g (−∂σ gµν + ∂µ gσν + ∂ν gµσ ) = 0. (3.57)
dτ 2 2 dτ dτ
The last equation is equivalent to the geodesic equation (3.50) for the Christof-
fel connection, but may different for more general connections. The equation
describes the motion of unaccelarated particle independent on its mass, but
it is also straightforward to include forces. For example, if the particle has a
charge q and mass m, then the geodesic equation would be given by
d2 xµ ρ
µ dx dx
σ
q µ dxν
+ Γ ρσ = F . (3.58)
dτ 2 dτ dτ m ν dτ
Note that the geodesics equation (3.50) is valid not only for proper time,
but for any affine parameter defined as linear function of proper time, i.e.

λ = aτ + b (3.59)

and for non-linear parametrizations (3.50) would be modified

d2 xµ ρ
µ dx dx
σ
dxµ
+ Γ ρσ = f (α) (3.60)
dα2 dα dα dα
where f (α) is some function which depends on the parametrization. Con-
versely if (3.60) is satisfied along some curve, then one can always find an
affine parameter λ for which (3.50) is satisfied.
In addition to parallel transport the geodesics xµ (λ) passing through some
point p can be used to map point from the tangent space Tp to a neighborhood
of p. Such mapping is called the exponential map

expp : Tp → M (3.61)

which is defined as
expp (k µ ) = xν (λ = 1) (3.62)
where
dxµ (λ = 0)
kµ = (3.63)
dλ
CHAPTER 3. CURVATURE 38

is the tangent vector at point p where we have set λ = 0. Of course the map
is well defined and invertible on a subset of Tp sufficiently close to k µ = 0,
and so (3.61) should not be taken literally. The range of the map can fail
to be all of the manifold given that there can be points not connected by a
geodesic, and the range of the map can fail to be all of the tangent space if
the manifold is geodesically incomplete (has singularities or boundaries).

3.4 Riemann tensor

The role of the Riemann (or curvature) tensor is to be able to represent
the features of the connection coefficients which would manifest the presence
of curvature. Such manifestations include parallel lines remain parallel or
parallel transports around closed loops change vectors. The change of par-
allel transported vectors depends on the loop, and for a local description of
the curvature it would be more useful to consider parallel transport around
infinitesimal loops.
We can guess what kind of object the Riemann tensor R should be by
considering a closed loop defined by two vectors Aµ and B µ . Then a trans-
formation around the loop should produce a vectorial change δV µ to a vector
V µ such that
δV ρ ∝ Rρσµν V σ Aµ B ν . (3.64)
Thus we anticipate the Riemann tensor to be of rank (1,3). It should also be
antisymmetric, i.e.
Rρσµν = −Rρσνµ , (3.65)
so that the parallel transport in opposite direction comes with a negative
sign.
To obtain the exact expression for the Riemann tensor in terms of con-
nection coefficients we consider instead the action of a commutator of two
covariant derivatives on the vector, i.e.

[∇µ , ∇ν ]V ρ . (3.66)

Indeed, the covariant derivative of a vector in the direction of a parallel

transport is zero by definition, and thus the covariant derivative measure
by how much the vector changes compared to what it would have been if
parallel transported. Then the commutation of two covariant derivatives
(when contracted with Aµ and B ν ) measures the changes to a vector V µ
if it was first parallel transformed one way (first along Aµ and then along
B µ ) compared to the other way (first along B µ and then along Aµ ). The
CHAPTER 3. CURVATURE 39

commutator of a vector is easy to express in terms of connection coefficients

& '
[∇µ , ∇ν ]V ρ = 2∇[µ ∇ν] V ρ = 2 ∂[µ ∇ν] V ρ − Γσ[µν] ∇σ V ρ + Γρ[µ|σ| ∇ν] V σ =
& & ' '
= 2 ∂[µ ∂ν] V ρ + ∂[µ Γρν]σ V σ + Γρ[ν|σ| ∂µ] V σ + −
& '
− 2 Γσ[µν] ∂σ V ρ + Γσ[µν] Γρλσ V λ + 2 Γρ[µ|σ| ∂ν] V σ + Γρ[µ|σ Γσλ|ν] V λ =
! "
& '
= 2 ∂[µ Γρν]σ + Γρ[µ|λ Γλσ|ν] V σ − 2Γσ[µν] ∇σ V ρ . (3.67)

The second term is a torsion tensor, i.e.

Tµν σ = −2Γσ[µν] (3.68)

which vanishes for Christoffel connections, and the first term defines the
Riemann tensor,
& '
Rρσµν = 2 ∂[µ Γρν]σ + Γρ[µ|λ Γλσ|ν] = ∂µ Γρνσ − ∂ν Γρµσ + Γρµλ Γλσν − Γρνλ Γλσµ . (3.69)

One can check that it is a legitimate (1,3) tensor, although it is not immedi-
ately clear that it is the same tensor as in (3.64) (see Wald for details). It is
also straightforward to determine the action of the commutator [∇ρ , ∇σ ] on
a tensor of arbitrary rank,

[∇ρ , ∇σ ]X µ1 ...µkν1 ...νl = −2Γλ[ρσ] ∇λ X µ1 ...µkν1 ...νl +Rµ1 λρσ X λµ2 ...µkν1 ...νl ...−Rλν1 ρσ X µ1 ...µkλν2 ...νl ....
(3.70)
Sometimes it useful to express the relevant quantities in a coordinate
independent way. Then one can think of the (1,2) torsion tensor as a map
from two vectors to a third vector,

T (X, Y ) = ∇X Y − ∇Y X − [X, Y ] (3.71)

and the (1,3) Riemann tensor as a map from three vectors to a fourth vector,

R(X, Y )Z = ∇X ∇Y Z − ∇X ∇Y Z − ∇[X,Y ] Z (3.72)

where the commutator of two vectors is

[X, Y ]µ ≡ X λ ∂λ Y µ − Y λ ∂λ X µ (3.73)

and the directional covariant derivative is

∇X ≡ X µ ∇ µ . (3.74)
CHAPTER 3. CURVATURE 40

Note that the third term in (3.72) vanishes when X and Y are coordinate
basis since
[∂µ , ∂ν ] = 0. (3.75)
In the case of metric compatible connection the Riemann tensor vanishes
if the metric is constant everywhere on the manifold,

gµν = const ⇒
∂σ gµν = 0⇒
Γρµν = 0⇒
∂σ Γρµν = 0⇒
Rρσµν = 0. (3.76)

One can also show that the opposite is true, i.e.

Rρσµν = 0 ⇒ gµν = const. (3.77)

Consider normal coordinates, i.e.

gµν = ηµν (3.78)

at some point p and with basis vectors ê(µ) whose dot product is

gσρ (p)êσ(µ) (p)êρ(ν) (p) = ηµν . (3.79)

This set of basis vectors can be parallel transported to some other point q.
Since the Riemann tensor is zero the result of the transport does not depend
on the trajectory and the dot product remains unchanged,

gσρ (q)êσ(µ) (q)êρ(ν) (q) = ηµν . (3.80)

Moreover the commutator of the parallel transported basis vectors vanishes

for torsion-free connections,

[ê(µ) , ê(ν) ] = ∇ê(µ) ê(ν) − ∇ê(ν) ê(µ) − T (ê(µ) , ê(ν) ) = 0. (3.81)

Then according to the Forbenius’s Theorem one can always find coordinates
xµ such that
∂
ê(µ) = µ (3.82)
∂x
in which the metric would have the form of ηµν .
CHAPTER 3. CURVATURE 41

Not all of the components of the Riemann tensor are independent and
one can show that the components must obey,

Rρσµν + Rσρµν = 0 (3.83)

Rρσµν − Rµνρσ = 0 (3.84)
Rρσµν + Rρµνσ + Rρνσµ = 0. (3.85)

which reduces the number of independent components to only 20 out of

44 = 256. (The number is 0, 1 and 6 in 1, 2 and 3 dimensions respectively.)
In addition algebraic identities there is a differential Bianchi identity,

∇λ Rρσµν + ∇ρ Rσλµν + ∇σ Rλρµν = 0 (3.86)

which is closely related to the Jacobi identity since it is nothing but

[[∇λ , ∇ρ ], ∇σ ] + [[∇ρ , ∇σ ], ∇λ ] + [[∇σ , ∇λ ], ∇ρ ] = 0. (3.87)

Some other useful tensor that can be formed from the Riemann tensor
are:
• Ricci tensor:
Rµν = Rλµλν (3.88)
which is symmetric for Christoffel connections.
• Ricci scalar:
R = Rµµ (3.89)
which can be shown (using Bianchi identity) to obey
1
∇µ Rρµ = ∇ρ R. (3.90)
2
• Einstein tensor:
1
Gµν = Rµν − R gµν (3.91)
2
which is also symmetric for Christoffel connections and can be shown
to obey
∇µ Gµν = 0. (3.92)

• Weyl tensor:
1! " 1
Cρσµν = Rρσµν − gρ[µ Rν]σ − gσ[µ Rν]ρ + R gρ[µ gν]σ (3.93)
2 3
which is essentially the Riemann tensor with all of the contractions
removed (i.e. with zero traces) while maintaining its symmetries.
CHAPTER 3. CURVATURE 42

Consider an example of two dimensional sphere with metric

ds2 = a2 dθ2 + sin2 θdφ2

! "
(3.94)

whose non-zero connection coefficients are

Γθφφ = − sin θ cos θ

Γφθφ = Γφφθ = cot θ. (3.95)

Then

Rθφθφ = ∂θ Γθφφ − ∂φ Γθθφ + Γθθλ Γλφφ − Γθφλ Γλθφ =

= (sin2 θ − cos2 θ) − (− sin θ cos θ cot θ) = sin2 θ (3.96)

and
Rθφθφ = a2 sin2 θ. (3.97)
It follows that

Rθθ = g φφ Rφθφθ = 1
Rθφ = Rφθ = 0
Rφφ = g θθ Rθφθφ = sin2 θ (3.98)

and
R = g θθ Rθθ + g φφ Rφφ = 2a−2 . (3.99)
Note that the Ricci scalar is positive as it should be for a positively curved
space such as the two-sphere, but for negatively curved spaces such as a
saddle it would be negative. The considered two-sphere is an example of a
maximally symmetric space generically defined by

Rρσµν = a−2 (gρµ gσν − gρν gσµ ) . (3.100)

Consider a one parameter family of geodesics γs (t) on a manifold M such

that for each s ∈ R, γs is parametrized by the affine parameter t. For a
non-crossing family of geodesics (which may or may not be the case) we can
choose the affine parameters in such a way that t and s form a system of
coordinates on a two-dimensional surfaces embedded in M, i.e.

xµ (s, t) ∈ M. (3.101)

There are two vector fields “tangent vector”,

∂xµ
Tµ = (3.102)
∂t
CHAPTER 3. CURVATURE 43

and “deviation vector”

∂xµ
Sµ = (3.103)
∂s
whose commutator vanishes,

[S, T ] = 0. (3.104)

This means that they can be used as basis vectors for a coordinate system
of our two-dimensional surface. Then the quantities defined as

V µ = (∇T S)µ = T ρ ∇ρ S µ (3.105)

and
aµ = (∇T V )µ = T ρ ∇ρ V µ (3.106)
we can call the “relative velocity” and “relative acceleration” of the geodesics.
For a connection with vanishing torsion equation (3.104) implies

S ρ ∇ ρ T µ = T ρ ∇ρ S µ

and the “relative acceleration” is

aµ = T ρ ∇ρ (T σ ∇σ S µ )
= T ρ ∇ρ (S σ ∇σ T µ )
= (T ρ ∇ρ S σ ) ∇σ T µ + T ρ S σ (∇ρ ∇σ T µ )
(S ρ ∇ρ T σ ) ∇σ T µ + T ρ S σ ∇σ ∇ρ T µ + Rµνρσ T ν
! "
=
= (S ρ ∇ρ T σ ) ∇σ T µ + S σ ∇σ (T ρ ∇ρ T µ ) − (S σ ∇σ T ρ ) (∇ρ T µ ) + Rµνρσ T ν T ρ S σ
= S σ ∇σ (T ρ ∇ρ T µ ) + Rµνρσ T ν T ρ S σ
= Rµνρσ T ν T ρ S σ (3.107)

where in the last line we used that T ρ is a tangent vector to geodesic and
thus
T ρ ∇ρ T µ = 0. (3.108)
The resulting equation is known as the geodesic deviation equation,

D2 µ
aµ = S = Rµνρσ T ν T ρ S σ (3.109)
dt2
which can be interpreted as the gravitational tidal force due to curvature
Rµνρσ .
CHAPTER 3. CURVATURE 44

3.5 Pullback and pushforward

Consider a map
φ:M →N (3.110)
where the manifolds M and N with respective coordinates xµ and y α might
have different dimensions. Then one can think of φ as a function which maps
function on N,
f :N →R (3.111)
to functions on M,
φ∗ f : M → R, (3.112)
where φ∗ f defined as a composite map

φ∗ f ≡ f ◦ φ. (3.113)

Such map is called the pullback since it is defined by pulling back the action
of function f from manifold N to manifold M (i.e. in the direction opposite
to the way φ was defined).
Clearly, it is not be possible to define a pushforward of a function, but
it is possible to define a pushforward of a vector V which, as we know, is
nothing but a map from functions on manifolds to real numbers. Then the
pushforward of the vector field defined on manifold M, i.e.

V : F (M) → R. (3.114)

is a vector field on manifold N, i.e.

φ∗ V : F (N) → R (3.115)

defined by the action of vector V on the pullback of functions, i.e.

(φ∗ V ) (f ) ≡ V (φ∗ f ). (3.116)

In coordinate basis, the pushforward of a vector can be written as

(φ∗ V )α ∂α f = V µ ∂µ (φ∗ f )
= V µ ∂µ (f ◦ φ)
∂y α
= V µ µ ∂α f (3.117)
∂x
or
∂y α
(φ∗ V )α = V µ . (3.118)
∂xµ
CHAPTER 3. CURVATURE 45

Then we can defined the pushforward matrix operator as

∂y α
(φ∗ )αµ = (3.119)
∂xµ
so that
(φ∗ V )α = (φ∗ )αµ V µ . (3.120)
Note that the pushforward matrix reduces to the general coordinate trans-
formations matrix for M = N.
Similarly to how it is not possible to define a pushforward of a function
it is not possible to define a pullback of a vector, but it is possible to define
a pullback of a one-form which is nothing but a map from vectors to real
numbers. Then the the pullback of a one-form ω on manifold N, is a one-
form φ∗ ω on manifold M, defined by how it acts on pushforward vectors,
i.e.
(φ∗ ω) (V ) = ω (φ∗ V ) . (3.121)
Then we can also define the pullback matrix operator as
∂y α
(φ∗ )µα = (3.122)
∂xµ
so that
(φ∗ ω)µ = (φ∗ )µα ωα . (3.123)
The pushforward and pullback can respectively be generalized to arbitrary
(k, 0) and (0, l) tensors,

(φ∗ S) ω (1) , ..., ω (k) = S φ∗ ω (1) , ..., φ∗ω (k) ,

! " ! "
(3.124)

(φ∗ T ) V (1) , ..., V (l) = T φ∗ V (1) , ..., φ∗V (l)

! " ! "
(3.125)
or in indices notation
∂y α1 ∂y αk µ1 ...µk
(φ∗ S)α1 ...αk = ... S (3.126)
∂xµ1 ∂xµk
∂y α1 ∂y αl
(φ∗ T )µ1 ...µl = ... Tα ...α . (3.127)
∂xµ1 ∂xµl 1 l
The pushforward and pullback of other tensors is defined only if the map φ
is invertible and then
"∗ "∗
(φ∗ S) ω (1) , ..., ω (k), V (1) , ..., V (l) = S φ∗ ω (1) , ..., φ∗ ω (k), φ−1 V (1) , ..., φ−1 V (l)
! " ! ! ! "

(3.128)
CHAPTER 3. CURVATURE 46

or

(φ∗ T ) ω (1) , ..., ω (k), V (1) , ..., V (l) = T φ−1 ∗ ω (1) , ..., φ−1 ∗ ω (k) , φ∗ V (1) , ..., φ∗V (l) .
! " !! " ! " "

(3.129)
−1
Clearly the pullback of φ is nothing but a pushforward of φ .
Consider an example of a two-sphere S 2 with coordinates (θ, ϕ) embedded
into a three dimensional Euclidean space R3 with coordinates (x, y, z) such
that there is a natural map
φ : S 2 → R3 (3.130)
defined by
φ(θ, ϕ) = (sin θ cos ϕ, sin θ sin ϕ, cos θ). (3.131)
Then the metric on R3 as a (0, 2) tensor can be pullback to S 2 . Then the
pullback matrix operator
# $
cos θ cos ϕ cos θ sin φ − sin θ
φ∗ = (3.132)
− sin θ sin ϕ sin θ cos φ 0

can be used to determine the pullback of the metric on R3 , i.e.

⎛ ⎞
1 0 0
g= ⎝ 0 1 0 ⎠ (3.133)
0 0 1

to the metric on S 2 , i.e.

⎞⎛ ⎛ ⎞
# 1 0 0 cos θ cos ϕ − sin θ sin ϕ
$
cos θ cos ϕ cos θ sin ϕ − sin θ ⎝
φ∗ g = 0 1 0 ⎠ ⎝ cos θ sin ϕ sin θ cos ϕ ⎠
− sin θ sin ϕ sin θ cos ϕ 0
0 0 1 − sin θ 0
⎛ ⎞
# $ cos θ cos ϕ − sin θ sin ϕ
cos θ cos ϕ cos θ sin ϕ − sin θ ⎝
= cos θ sin ϕ sin θ cos ϕ ⎠
− sin θ sin ϕ sin θ cos ϕ 0
− sin θ 0
# $
1 0
= (3.134)
0 sin2 θ

As a result we will obtain a (0, 2) tensor on S 2 which is also called the

induced metric, (φ∗ g)µν .

3.6 Killing vectors

Given a one parameter family of diffeomorphism
φt : M → M (3.135)
CHAPTER 3. CURVATURE 47

such that
φs+t = φs ◦ φt (3.136)
we can consider trajectory of a given point p under φt . These trajectories
xµ (t) can used to define vector fields by evaluating tangent vector at t = 0:
. µ/
dx
V µ (x) = . (3.137)
dt t=0

Similarly, we can define a one parameter family of diffeomorphism (or inte-

gral curves) from a given vector field (or generator of the diffeomorphism)
by solving the differential equation
dxµ
= V µ. (3.138)
dt
Then for any tensor field we can ask how it changes as we travel along the
integral curves, i.e.

∆t T µ1 ...µkν1 ...νl (p) = φt∗ T µ1 ...µkν1 ...νl (p) − T µ1 ...µkν1 ...νl (p)
! "
(3.139)

as well as define the so-called Lie derivative, i.e.

∆t T µ1 ...µkν1 ...νl
# $
! µ1 ...µ "
LV T k
ν1 ...νl = lim (3.140)
t→0 t

which is a map from (k, l) to (k, l). The Lie derivative is linear

LV (aT + bS) = aLV T + bLV S (3.141)

and obeys product rule

LV (T ⊗ S) = (LV T ) ⊗ S + T ⊗ (LV S) . (3.142)

Although the Lie derivative does not depend on the coordinate system it
does not require specification of connections. It is easy to show that the Lie
derivative reduces to an ordinary derivatives for scalars

LV f = V (f ) = V µ ∂µ f (3.143)

and to a commutator (also called the Lie bracket) for vectors

LV U µ = [V, U]µ . (3.144)

CHAPTER 3. CURVATURE 48

Then the Lie derivative of a one-form can be derived by considering the Lie
derivative of a scalar

LV (ωµ U µ ) = V ν ∂ν (ωµ U µ )
(LV ωµ ) U µ + ωµ (LV U)µ = V ν U µ ∂ν ωµ + V ν ωµ ∂ν U µ
(LV ωµ ) U µ + ωµ (V ν ∂ν U µ − U ν ∂ν V µ ) = V ν U µ ∂ν ωµ + V ν ωµ ∂ν U µ
(LV ωµ ) U µ = V ν U µ ∂ν ωµ + ωµ U ν ∂ν V µ
L V ωµ = V ν ∂ν ωµ + ων ∂µ V ν . (3.145)

For an arbitrary tensor the Lie derivative is given by

LV T µ1 ...µkν1 ...νl = V σ ∂σ T µ1 ...µkν1 ...νl −(∂λ V µ1 ) T λµ2 ...µkν1 ...νl ...+ ∂ν1 V λ T µ1 ...µkλν2 ...νl ....
! "

(3.146)
Although the above expression is coordinate independent one can rewrite it
in a more covariant form

LV T µ1 ...µkν1 ...νl = V σ ∇σ T µ1 ...µkν1 ...νl −(∇λ V µ1 ) T λµ2 ...µkν1 ...νl ...+ ∇ν1 V λ T µ1 ...µkλν2 ...νl ....
! "

(3.147)
A diffeomorphism φ is called a symmetry of some tensor T if it leaves
the tensor invariant, i.e.
φ∗ T = φ∗ T = T. (3.148)
If the symmetry is generated by a vector field V µ (x) then the symmetry
implies that the Lie derivative vanishes

LV T = 0. (3.149)

In that case one find a coordinate system in which the components of T

are independent of one of the coordinates (pointing in the direction of the
integral curves of V ).
One of the most important symmetries are of the metric

φ∗ gµν = gµν (3.150)

in which case the diffeomorphism φ is called an isometry. The vector field

V µ (x) is called a Killing vector field if it generates the isometries, i.e.

LV gµν = 0 (3.151)

or

0 = V σ ∇σ gµν + (∇µ V σ ) gσν + (∇ν V σ ) gµσ

0 = ∇ µ V ν + ∇ν V µ
0 = ∇(µ Vν) (3.152)
CHAPTER 3. CURVATURE 49

which is the Killing equations. The maximal number of Killing vector

field in n-dimentional manifold is the sum of n translations and n(n − 1)/2
rotations (from any axis n to any other axis n − 1 but without double-
counting). The sum of rotations and translations is n(n + 1)/2 or ten for the
four-dimensional space-time. The space-time of general relativity is called
maximally symmetric if its metric tensor has ten independent Killing
vectors which is the same as the number of independent parameters in a
symmetric (0, 2) tensor. For example,

ds2 = dx2 + dy 2 (3.153)

has three Killing vectors (i.e. satisfy the Killing equations (3.152)) corre-
sponding to two translation

X = (1, 0) (3.154)
Y = (0, 1) (3.155)

and one rotation

Rµ = (−y, x) (3.156)
µ
Consider a free particle moving along a geodesic x (λ) whose tangent
vector is
dxµ
Uµ = . (3.157)
dλ
Then for a Killing vector V µ we have

U ν ∇ν (Vµ U µ ) = U ν U µ ∇ν Vµ + Vµ (U ν ∇ν U µ )
1 ν µ
= (U U ∇ν Vµ + U µ U ν ∇µ Vν )
2
= 0. (3.158)

which implies that the quantity Vµ U µ is conserved along the trajectory of

motion.

3.7 Tetrads
So far we were using the coordinate basis for the tangent space

ê(µ) = ∂µ (3.159)

defined as partial derivatives with respect to coordinate functions and for

dual space
θ̂(µ) = dxµ (3.160)
CHAPTER 3. CURVATURE 50

defined as gradients of the coordinate functions. Of course, we are free to

use any basis we like and there is one particularly useful set of bias vectors,
known as tetrads or vierbein, defined at each point by
! "
g ê(a) , ê(b) = ηab . (3.161)

The coordinate basis vectors can be expressed in tetrad basis as

ê(µ) = eaµ ê(a) (3.162)

where the eaµ matrix an invertible matrix whose inverse eµa expresses tetrad
basis vectors in coordinate basis

ê(a) = eµa ê(µ) . (3.163)

As usual (with the abuse of notations) the component eaµ and eµa are often
called the tetrads and inverse tetrads respectively. Clearly, they must satisfy

eµa eaν = δνµ (3.164)

or
eaµ eµb = δba , (3.165)
and (3.161) can be rewritten in the indices notation as

gµν eµa eνb = ηab (3.166)

or
gµν = eaµ ebν ηab (3.167)
which is why tetrads are sometimes called the “square root” of the metric.
After defining the orthonormal basis for vectors we can define an orthonormal
basis for one-forms by setting

θ̂(a) ê(b) = δba .

! "
(3.168)

Then one can show that using the very same inverse tetrads eµa and tetrads
eaµ we can respectively express

θ̂(µ) = eµa θ̂(a) (3.169)

and
θ̂(a) = eaµ θ̂(µ) . (3.170)
CHAPTER 3. CURVATURE 51

Of course, no only the basis vectors but any vectors expressed in terms of
the coordinate basis can be re-expressed in terms of tetrad basis. In other
words
V = V a ê(a) = V µ ê(µ) (3.171)
implies
V a = eaµ V µ . (3.172)
Similarly for tensors

V ab = eaµ V µb = eνb V aν = eaµ eνb V µν (3.173)

and one example of such transformation we had already seen before for metric
tensor (3.166).
For a non-coordinate tetrad basis we can illustrate how the change of
tetrad basis would change components of tensors even without changing co-
ordinates. If we go from one tetrad basis to another the orthonormality must
be preserved and thus the metric ηab must remain flat. Of course the group
of such transformation is known as Lorentz group and the transformation
matrices are known as Lorentz matrices. In contrast to the global Lorentz
transformations in special relativity we are now free to have different changes
of basis at different points (local Lorentz transformations), i.e.

ê(a) → ê(a′ ) = Λa′a (x)ê(a) (3.174)

where Λaa′ (x) (actually an inverse Lorentz transformation) is a function of

position such that
Λaa′ (x)Λb′b (x)ηab = ηa′ b′ (3.175)
at every point. But since the local Lorentz transformations only transform
basis, we can still transform the coordinates using the general coordinate
transformations. So even if the tensor is expressed in mixed basis, we know
exactly how to transform it
# µ′ $ # ν′ $
a′ µ′ ∂x ∂x
a′
T b′ ν ′ = Λ a µ
Λb′ b
T aµbν . (3.176)
∂x ∂xν

Things become a little bit more complicated with covariant derivatives

where the connection coefficients Γλµν are replaced with the so-called spin
connections ωµab . For example, the covariant derivative of a (1,1) tensors
components in tetrads basis is given by

∇µ X ab = ∂µ X ab + ωµ ac X cb − ωµ c b X ac . (3.177)
CHAPTER 3. CURVATURE 52

Of course in mixed basis one can get terms with connection coefficients and
spin coefficients in the same expression, e.g.

∇µ X aν = ∂µ X aν + ωµac X cν + Γνµλ X aλ . (3.178)

But since the same object (e.g. covariant derivative of vector) can be
written in coordinate basis

∇X = ∇µ X λ dxµ ⊗ ∂λ = ∂µ X λ + Γλµν X ν dxµ ⊗ ∂λ

! " ! "
(3.179)

as well as in tetrads basis

∇X = (∇µ X a ) dxµ ⊗ ê(a)

= ∂µ X a + ωµa b X b dxµ ⊗ ê(a)
! "

= ∂µ (eaν X ν ) + ωµa b ebν X ν dxµ ⊗ eλa ∂λ

! ! "" ! "

= eaν ∂µ X ν + ∂µ eaν X ν + ωµa b ebν X ν dxµ ⊗ eλa ∂λ

! " ! "

= ∂µ X λ + eλa ∂µ eaν X ν + ωµa b eλa ebν X ν dxµ ⊗ ∂λ

! "
(3.180)

we can express the connection coefficients in terms of spin coefficients

Γλµν = eλa ∂µ eaν + ωµa b eλa ebν (3.181)

and
ωµab = eaν eλb Γνµλ − eλb ∂µ eaλ (3.182)
which is equivalent to
∇µ eaν = 0. (3.183)
Just like the connection coefficients the spin connections are not legitimate
tensors, but the lower Greek index does transform as a one-form. The trans-
formation law for other indices is given by

ωµ′a b′ = Λa a Λb′ b ωµab − Λb′ c ∂µ Λa c.

′ ′ ′
(3.184)

which can be derived by demanding that covariant derivatives transform as

tensors.
It is sometime useful to view objects with mixed indices as tensor-valued
(described by Latin indices) differential from (described by Greek indices).
For example, Aµν ab which in antisymmetric in µ and ν can be thought of
as (1,1) tensor valued differential two-form, or Xµa can be thought of as a
vector valued one form. Then one can show that the exterior derivative

(dX)µν a = ∂µ Xν a − ∂ν Xµ a (3.185)
CHAPTER 3. CURVATURE 53

does not transform as a tensor, but

(dX)µν a + (ω ∧ X)µν a = ∂µ Xν a − ∂ν Xµ a + ωµ ab Xν b − ων ab Xµ b (3.186)

is a legitimate tensor. Then one can view torsion as a vector valued two-
form Tµν a and Riemann curvature tensor as a (1,1) valued two-form Rabµν .
By suppressing indices of the differential forms the torsions and curvature
tensors we can written as

T a = dea + ω ab ∧ eb (3.187)

and
Rab = dω ab + ω ac ∧ ω cb (3.188)
known as Maurer-Cartan structure equations. Similarly the Bianchi
identities (3.85) and (3.86) can be written respectively as

dT a + ω ab ∧ T b = Rab ∧ eb (3.189)

and
dRab + ω ac ∧ Rcb − Rac ∧ ω cb = 0. (3.190)
The metric compatibility (i.e. ∇µ gνλ = 0)

0 = ∇µ ηab = ∂µ ηab − ωµ ca ηcb − ωµ cb ηac = −ωµba − ωµab (3.191)

implies the antisymmetric of the spin connection

ωµab = −ωµba . (3.192)

For Christoffel connection the torsion (3.187) vanishes and thus

dea = −ω ab ∧ eb = ω ab ∧ eb (3.193)

The two equations can be used to solve for the spin connections in terms of
tetrads.