SR → MATHEMATICAL TOOLS I: TENSOR CALCULUS

↓ Lecture 8 [05.12.23]

3.5. The metric tensor

A differentiable manifold M does not automatically allow us to measure the length of curves, the angles
of intersecting lines, or the area/volume of subsets of the manifold; to do so, we need a metric on M
(which is an additional piece of information). While the continuity structure (an atlas) that comes with
M determines its topology, the metric determines its geometry (= shape). The same manifold M can be
equipped with different metrics; this corresponds to different geometries of the same topology (a potato
and an egg both have the topology of a sphere, nonetheless they are geometrically distinct).
A differentiable manifold together with a (pseudo-)metric is called ↑ (pseudo-)Riemannian manifold. In spe-
cial relativity and general relativity, spacetime is modeled by such (pseudo-)Riemannian
manifolds where the metric is used to represent spatial and temporal distances between events.

25 | Motivation:
On linear spaces V , it is convenient to define an ↓ inner product (like in quantum mechanics where
you consider Hilbert spaces and use their inner product to compute probabilities and transition
amplitudes).
Recall the definition of a (real) inner product:

hji W V  V ! R with … (3.44a)

Symmetry: hxjyi D hyjxi (3.44b)
(Bi)linearity: hax C byj´i D ahxj´i C bhyj´i (3.44c)
Positive-definiteness: x ¤ 0 ) hxjxi > 0 (3.44d)

Once you have an inner product, you get a norm, and subsequently a metric for free:
p
hxjyi ) kxk WD hxjxi ) d.x; y/ WD kx yk (3.45)
„ƒ‚… „ ƒ‚ … „ ƒ‚ …
Inner product Norm Metric

Thus an inner product is a rather versatile structure and nice to have!

Problem: We cannot define a inner product on the manifold directly because M is not a linear space.
However: We can introduce an inner product on each of its tangent spaces Tp M ! !
26 | ⁂ Riemannian (Pseudo-)Metric ds 2 := Symmetric, non-degenerate .0; 2/-tensor field:

ds 2 W M 3 p 7! dsp2 W Tp M  Tp M ! R


(3.46a)
„ ƒ‚ …
Bilinear & symmetric & non-degenerate

dsp2 bilinear ) dsp2 2 Tp M ˝ Tp M

D
X
) dsp2 D gij .x/ dx i ˝ dx j  gij .x/ dx i dx j (3.46b)
i;j D1

with gij D gj i (symmetry) and g D det.gij / ¤ 0 (non-degeneracy).

NICOLAI LANG • INSTITUTE FOR THEORETICAL PHYSICS III • UNIVERSITY OF STUTTGART PAGE

92
SR → MATHEMATICAL TOOLS I: TENSOR CALCULUS

• The tensor product is non-commutative: dx i ˝ dx j ¤ dx j ˝ dx i . However, you can always

decompose a tensor product as
1 1
dx i ˝ dx j D .dx i ˝ dx j C dx j ˝ dx i / C .dx i ˝ dx j dx j ˝ dx i / (3.47)
2
„ ƒ‚ … „ 2 ƒ‚ …
DWdx i _dx j DWdx i ^dx j

with the symmetrized tensor product dx i _ dx j and the anti-symmetrized tensor product
dx i ^ dx j (↑ wedge product).
Since gij is assumed to be symmetric, only the symmetric component survives:

gij .x/dx i ˝ dx j D gij .x/dx i _ dx j  gij .x/dx i dx j (3.48)

This means that when writing dx i dx j in the above formula, you can be sloppy and either
mean dx i ˝ dx j or, equivalently, dx i _ dx j . You will find both conventions in the literature.
I will use dx i dx j  dx i _ dx j so that dx i dx j D dx j dx i .
• It would be more appropriate to write g D gij dx i dx j for the metric .0; 2/-tensor; it is
conventional, however, to reserve g for the determinant det.gij / so that we are stuck with
ds 2 for the metric. Note that the d in ds 2 does not refer to an ↑ exterior derivative, it is purely
symbolical.
• To define a proper ↓ inner product on Tp M , we should demand ↓ positive-definiteness instead
of non-degeneracy. This, however, is often (for example in relativity) too restrictive; as
it turns out, non-degeneracy is all we need for an isomorphism between Tp M and Tp M
(“pulling indices up and down”, → below). This is why negative eigenvalues of gij are fine for
many purposes, and motivates the concept of a → signature:
27 | Signature:
Since gij .x/ D gj i .x/ and det.gij .x// ¤ 0
! gij .x/ has r positive and s negative real eigenvalues for all p 2 M
Since det.gij .x// ¤ 0, these numbers must be the same for all p 2 M .
! .r; s/: ⁂ Signature of the metric ds 2
This classification does not depend on the coordinate basis (↑ Sylvester’s law of inertia).

• .r > 0; s D 0/
! ds 2 : Riemannian metric ! .M; ds 2 /: ⁂ Riemannian manifold
I.e., gij has only positive eigenvalues for all p 2 M and is therefore ↓ positive-definite. This
produces a true, positive-definite inner product on Tp M .
• .r > 0; s > 0/
! ds 2 : pseudo-Riemannian metric ! .M; ds 2 /: ⁂ pseudo-Riemannian manifold
I.e., gij has both positive and negative eigenvalues and is therefore ↓ indefinite.
– .r > 0; s D 1/ or .r D 1; s > 0/:
! ds 2 : Lorentzian metric ! .M; ds 2 /: ⁂ Lorentzian manifold
In relativity we are only interested in metric tensors with one positive and three
negative eigenvalues (equivalently: three positive and one negative eigenvalue). Math-
ematically speaking, spacetime is then a four-dimensional Lorentzian manifold and a
special case of a pseudo-Riemannian manifold.

NICOLAI LANG • INSTITUTE FOR THEORETICAL PHYSICS III • UNIVERSITY OF STUTTGART PAGE

93
SR → MATHEMATICAL TOOLS I: TENSOR CALCULUS

28 | Example: (Details: → Problemset 4)

i| ^ D D 2 Euclidean space E2  .R2 ; dsE

2
/
The Euclidean metric in Cartesian coordinates x 1 D x and x 2 D y reads:
 
2 1 0
dsE WD dx 2 C dy 2 D gij .x/ dx i dx j with .gij / D : (3.49)
0 1
„ ƒ‚ …
Signature
.2;0/

This is consistent with the notion of dx and dy as infinitesimal shifts in coordinates and ds 2
as the infinitesimal distance (squared) that corresponds to this shift:

ii | We can now transition to a new chart, namely polar coordinates xN 1 D r and xN 2 D  . The
induced basis change on the cotangent space is given by the total differential of the coordinate
functions Eq. (3.14):
(
1 x D r cos. / Eq. (3.14) dx D cos. / dr r sin. / d
' W ) (3.50)
y D r sin. / dy D sin. / dr C r cos. / d

iii | We find the components of the metric tensor field in the new basis fdxN 1 D dr; dxN 2 D dg:

2 2 2 2 i j 1 0
ds $ dr C r d D gN ij .x/
N dxN dxN with .gN ij / D : (3.51)
0 r2
„ ƒ‚ …
Signature
.2;0/

This expression is again compatible with infinitesimal shifts in the (new) coordinates r and  :

• The Euclidean plane E2 is therefore an example for a Riemannian manifold with metric
signature .2; 0/; its distinctive feature is that it is flat.

NICOLAI LANG • INSTITUTE FOR THEORETICAL PHYSICS III • UNIVERSITY OF STUTTGART PAGE

94
SR → MATHEMATICAL TOOLS I: TENSOR CALCULUS

• Note that here we compute the same infinitesimal length in different coordinates (with
the same result)! We did not change the metric, only the coordinates and thereby the
coordinate basis in which we express the metric tensor. This is flat Euclidean space in
↑ curvilinear coordinates. By contrast, later in general relativity we will study
curved (non-flat, non-Euclidean) metric tensors, i.e., we will modify the geometry of
space(time) itself.

29 | Since the metric ds 2 is a .0; 2/-tensor field:

N xN i dxN j D ds 2 D gij .x/dx i dx j

gN ij .x/d (3.52)

ı
Eq. (3.14) !

@x l @x m
gN ij .x/
N D glm .x/ (3.53)
@xN i @xN j
The metric (components) transforms as any other .0; 2/ tensor. Nothing special!
Side note:
Eq. (3.53)
Let g WD det.gij / and gN WD det.gN ij / !
ˇ  ˇ
p @x ˇˇ p
N D ˇˇdet
ˇ
jgj jgj (3.54)
@xN ˇ

! jgj is a pseudo scalar tensor density of weight w D C1. The “pseudo” indicates that the
p

absolute value of the Jacobian determinant shows up, cf. Eq. (3.37).
Eq. (3.39) p
^g <0 ! dDx g is a scalar (→ later)!
30 | Length of curves on M :
One immediate benefit of having a Riemannian manifold is that we can now compute the length of
curves .t/ on M (parametrized by t 2 Œa; b and given in some chart):

s
b
d i .t / d j .t /
Z Z
LŒ   ds WD gij . .t // dt (3.55)
a dt dt
Z b
 k P .t /k .t/ dt (3.56)
a

¡! If ds 2 is a true pseudo metric (i.e., gij has at least one negative eigenvalue), one must make sure
that the chosen curve does not produce negative values under the square root. In relativity
these will be ↑ time-like curves.
Example:
Let be the circle with radius R in the Euclidean plane E2 . A possible parametrization in Cartesian
coordinates (with origin in the center of the circle) is Exy .t/ D .x t ; y t / D .R cos.t/; R sin.t//
with 0  t < 2 so that one finds for the circumference:
Z p Z 2 q
LD dx 2 C dy 2 D xP t2 C yP t2 dt $ 2R (3.57)
0

NICOLAI LANG • INSTITUTE FOR THEORETICAL PHYSICS III • UNIVERSITY OF STUTTGART PAGE

95
SR → MATHEMATICAL TOOLS I: TENSOR CALCULUS

The same length can of course be calculated with the parametrization Er .t/ D .r t ;  t / D .R; t/
and 0  t < 2 in polar coordinates:
Z p Z 2 q
LD dr 2 C r 2 d 2 D rPt2 C r t2 Pt2 dt $ 2R (3.58)
0

Details: → Problemset 4
31 | Besides computing lengths of curves (and other geometric quantities, → later), there is another
benefit of having a metric tensor:
Pulling indices down:

i : : :  : : : ip  : : :  i1 : : : k : : : ip  : : : 
TQ 1 : : : i : : :  j : : : j WD gi k T  : : :  : : :  j : : : j (3.59)
1 q 1 q

! TQ is a tensor of type .p 1; q C 1/
• In Eq. (3.59) we indicate “empty” slots for indices by  to emphasize that in each index
“column” an index can either be up (contravariant) or down (covariant). It is conventional to
omit the -markers. Note that this explains why you never should write two indices directly
above each other (except for special cases, → below).
Furthermore, since g is fixed, it makes sense to label TQ again by T (note that the difference
between the original tensor and the new one is manifest in the different index patterns!):

i : : :  : : : ip  : : :  i1 : : : : : : ip
TQ 1 : : : i : : :  j :::j 7! T i j 1 : : : jq (3.60)
1 q

Example:

Ai j k l WD gj m Aimkl (3.61)

• This convention matches perfectly with the computation of an inner product (which is
determined by the metric tensor g) of two contravariant vectors:

def def
hA; Bi D gij Ai B j D Ai Bi (3.62)
„ƒ‚…
Scalar

32 | Pulling indices up:

We would like to have a .2; 0/-tensor g ij with the property
Š def
ıjk T j D T k D g ki Ti D g ki gij T j : (3.63)

g ij allows us to revert the pulling-down of indices defined by the metric gij . Note that g ij is a
different tensor than gij , we could call it gQ ij ; however, it is conventional to denote it with the same
label due to the following close relationship with g:

Š
g ki gij D ıjk (3.64)

This is an implicit equation for g ki !

NICOLAI LANG • INSTITUTE FOR THEORETICAL PHYSICS III • UNIVERSITY OF STUTTGART PAGE

96
SR → MATHEMATICAL TOOLS I: TENSOR CALCULUS

! g ij is the inverse matrix of gij

(Which always exists because ds 2 is non-degenerate: det.gij / ¤ 0.)
! In general:

i : : : ip  : : : j : : :  j k i1 : : : ip  : : :  : : : 
TQ 1 : : :  j : : :  : : : j WD g T  : : :  j : : : k : : : j (3.65)
1 q 1 q

! TQ is a tensor of type .p C 1; q 1/
• Again we relabel TQ to T and omit the -markers:
i : : : ip  : : : j : : :  i1 : : : ip j
TQ 1 : : :  j ::::::j 7! T j1 : : : : : : jq (3.66)
1 q

• Example:

Aij kl WD g lm Aij k m (3.67)

• With these new definitions, we can now raise and lower contractions:
j
Ai Bi D Ai ıi Bj D Ai gik g kj Bj D Ai gik B k D Ak B k D Ai B i (3.68)

• What happens if you pull the indices of the Kronecker symbol up or down?

ı ij WD g j k ı i k D g ij and ıij WD gik ı kj D gij (3.69)

¡! ı ij  g ij and ıij  gij denote the metric and its inverse!

! We never use the notation ı ij and ıij to prevent confusion!
• Note that in general
g j k T ik D T ij ¤ T j i D g j k Tk i : (3.70)

This means that the “column” in which the index is located is important, and notations like
Tki are ill defined (if you pull k up by g j k , do you get T ij or T j i ?). However, if the tensor is
symmetric, T ij D T j i , this does not matter and you can get away with the sloppy notation Tki .
This explains why writing ıki for the Kronecker symbol is fine: g j i D g j k ıki is symmetric.
33 | Mathematical side note:
“Pulling indices up and down” is mathematically the application of an ↓ isomorphism between Tp M
and Tp M :

g.; / W Tp M 3 A 7! g.A; / 2 Tp M (3.71)

This has nothing to do with differential geometry or manifolds in particular; it is a general feature
of non-degenerate bilinear forms on vector spaces. In differential geometry, this canonical iso-
morphism between the tangent bundle TM and the cotangent bundle T  M is know as ↑ musical
isomorphism.
For example, you are using the same kind of isomorphism all the time in quantum mechanics,
namely whenever you “dagger” a ket j‰i to obtain a bra h‰j:
Š
./Ž W H 3 j‰i 7! h‰j  j‰iŽ 2 H  with h‰jjˆi D h‰jˆiH for all jˆi 2 H . (3.72)

NICOLAI LANG • INSTITUTE FOR THEORETICAL PHYSICS III • UNIVERSITY OF STUTTGART PAGE

97
SR → MATHEMATICAL TOOLS I: TENSOR CALCULUS

Note how the bra bra h‰j associated to the ket j‰i is defined via the inner product hjiH (and
therefore metric) of the Hilbert space (↑ Riesz representation theorem)!
This leads to a nice dictionary between concepts in tensor calculus (and therefore relativity)
and the bra-ket formalism of quantum mechanics:
Relativity (fixed p 2 M ) Quantum mechanics
Inner product space Tp M H
Basis f@i g fji ig
Vector A D Ai @i j‰i D ‰i jii
Dual space Tp M H
Dual basis fdx i g fhi jg
… dx i .@j / D ıji hi jj i D ıij
Covector B D Bi dx i h‰j D ‰i hi j
Inner product g.A1 ; A2 / D gij Ai1 Aj2 h‰jˆi
Tensor A D Aij @i ˝ @j j‰i ˝ jˆi  j‰ijˆi
… B D Bij dx i ˝ dx j h‰j ˝ hˆj  h‰jhˆj
Operator T D T ij @i ˝ dx j jˆi ˝ h‰j  jˆih‰j
Trace T ii TrŒjˆih‰j
Scalar BA D Bi Ai D gij B i Aj h‰jjˆi D h‰jˆi
Pulling indices down Ai D gij Aj h‰j D j‰iŽ
Pulling indices up Ai D g ij Aj j‰i D h‰jŽ

3.6. Differentiation of tensor fields

34 | Remember: @i ˆ is covariant vector if ˆ is scalar. However:

^ Contravariant vector Ai :
" #
@ANi @x m @ @xN i l @2 xN i @x m l @x m @xN i @Al
ANi ;k  kD A D A C (3.73)
@xN @xN k @x m @x l @x m @xƒ‚
„
l @x
N k … „@xN k @x l m
ƒ‚ @x …
¤ 0 (in general) / .1; 1/-tensor ,

Here we used the transformation of ANi [Eq. (3.8)] and @N k [Eq. (3.5)] and the product rule.
@ANi
! In general: @xN k
is not a tensor!
35 | How to define a derivative of tensor fields that again transforms as a tensor?
To solve this problem, we first need a new field:
! ⁂ Christoffel symbols (of the second kind):

1
€ ikl WD g im gmk;l C gml;k


gkl;m (3.74)
2

• The Christoffel symbols are symmetric in the lower two indices: € ikl D € ilk
• ¡! Despite the index notation, the Christoffel symbols are not tensors:
@xN i @x n @x p m @x n @x p @2 xN i
€N ikl $ m k € np l @x n @x p
(3.75)
@x @xN @xN l @xN k @xNƒ‚
„ …
No tensor!

NICOLAI LANG • INSTITUTE FOR THEORETICAL PHYSICS III • UNIVERSITY OF STUTTGART PAGE

98
SR → MATHEMATICAL TOOLS I: TENSOR CALCULUS

This is why they are called “symbols” and not “tensors”!

• There are also Christoffel symbols of the first kind:
1
€ikl WD gij € jkl D


gik;l C gil;k gkl;i (3.76)
2

• Mathematically, the Christoffel symbols are the coefficients (in some basis) of the ↑ Levi-
Civita connection which is determined by the metric tensor g ij (→ later).
36 | ^ Contravariant vector ANi and contract it with €N ikl :

@xN i @x n
 p 
@x n @2 xN i @x p Nl
 
@x Nl
€N ikl ANl D m k € mnp A A (3.77)
@x @xN @xN l @xN k @x n @x p @xN l
„ ƒ‚ … „ ƒ‚ …
Ap Ap
„ ƒ‚ … „ ƒ‚ …
.1; 1/-tensor , Problematic term in Eq. (3.73)

Idea: Add Eq. (3.73) and Eq. (3.77) to cancel the problematic term:

@x m @xN i h l i
ANi ;k C €N ikp ANp D k l
A ;m C € l
mp A p
(3.78)
„@xN @x ƒ‚ …
.1; 1/-tensor ,,

37 | This motivates the definition of the ⁂ Covariant derivative:

Scalar: ˆIk WD ˆ;k (3.79a)

Contravariant vector: Ai Ik WD Ai ;k C € ikl Al (3.79b)

Covariant vector: Bi Ik WD Bi;k € lik Bl (3.79c)

• With this definition, Ai Ik is a .1; 1/-tensor and Bi Ik is a .0; 2/-tensor!

• With this definition, the product rule is valid for the covariant derivative:

.Ai Bi /Ik D .Ai Bi /;k $ Ai Ik Bi C Ai BiIk (3.80)

• The construction of higher-rank tensors by tensoring contra- and covariant vectors Eq. (3.32)
and the definitions of the covariant derivative above Eq. (3.79) can be used to construct
covariant derivatives of arbitrary tensor fields. For example:

T ikIl WD T ik;l C € iml T mk € mkl T im (3.81)

• With this generalization, we can apply the covariant derivative multiple times. For example:
 
Ai IkIl  Ai Ik (3.82)
Il

• The covariant derivative is not commutative in general:

Ai IkIl Ai IlIk ¤ 0 (3.83)

! Riemann curvature tensor ! general relativity (→ later)

(This is not the case for the “normal” derivative: Ai ;k;l D Ai ;l;k .)

NICOLAI LANG • INSTITUTE FOR THEORETICAL PHYSICS III • UNIVERSITY OF STUTTGART PAGE

99
SR → MATHEMATICAL TOOLS I: TENSOR CALCULUS

38 | Conclusion:
If you can formulate an equation that describes a physical theory in terms of tensors, it can always
be brought into the form

T IJ .x/ D 0 : (3.84)

(This equation is meant to hold for all values of indices I and J and all coordinate values x.)
Here is an example:
The (inhomogeneous) Maxwell equations on an arbitrary (potentially curved) spacetime read:

4 
F I C J D0 (3.85)
„ ƒ‚ c …
DWT  .x/

with current density J  and field strength tensor F  D g  g  .AI AI /.

How does Eq. (3.84) look like in any other coordinate system xN D '.x/?
Easy:

@xN I @x N M
TN IJ .x/
N D M T N .x/ D 0 , TN IJ .x/
N D 0: (3.86)
@x @xN J „ ƒ‚ …
D0

This means:

Tensor equations are automatically form-invariant under arbitrary coordinate

transformations; we say they exhibit ⁂ (manifest) general covariance.

The “manifest” means that checking general covariance is just a matter of checking whether the
equation “looks right”, i.e., whether it is built from tensors following the rules discussed in this
chapter. If a property of an equation is manifest, you don’t have to do calculations to verify it!
In the next chapter, we take a step back and specialize the allowed coordinate transformations to
the Lorentz transformations of special relativity. We can then use the form-invariance of
equations built from “Lorentz tensors” to construct Lorentz covariant equations from scratch –
which was our original goal!

NICOLAI LANG • INSTITUTE FOR THEORETICAL PHYSICS III • UNIVERSITY OF STUTTGART PAGE

100