OXFORD UNIVERSITY

PHYSICS DEPARTMENT
3RD YEAR UNDERGRADUATE COURSE

SYMMETRY AND RELATIVITY

TUTORIAL I

Tensors. Problem Set 1.

Andrei Starinets
[email protected]

NOT FOR DISTRIBUTION

I. A BRIEF COMMENT ABOUT THE LITERATURE

The titles recommended by the lecturer are [1] and [2].

When reading about tensors, one should remember that many books dis-
cuss them twice - once in the context of Special Relativity and then in full
generality. For example, in Weinberg’s book [3] tensors appear in Chap-
ter 2 and then in Chapter 4 (and similarly in Landau-Lifshitz [4]). In this
course, we shall not be making such a distinction and always treat tensors in
the most general way, unless explicitly stated otherwise, because this is how
they appear in various branches of physics (not only in SR & GR).
A very useful collection of problems (with solutions!) in Special and Gen-
eral Relativity (tensors appear in Chapter 3) is [5].
Useful books where tensors and other structures are introduced rigorously
but in the language accessible to physicists are [6], [7], [8], [9].

II. TENSORS AND TENSOR ALGEBRA

In this course, we will be dealing mostly with 3-dimensional Euclidean

space and 4-dimensional Minkowski space. These are examples of the so
called “metric spaces”, i.e. spaces, equipped with a machinery (metric) to
measure distances between points. To be more precise, one has to define the
notion of “space” first. This is done in topology and differential geometry
courses (one starts with sets, then introduces topology to have a sense of
continuity, then gradually adds other structures, including a metric). This
is important to know for a physicist, since at small (Planckian, i.e. l ∼
p
lP = G~/c3 ∼ 10−33 cm) distances some of these structures may not be

2
adequate (e.g. Riemannian geometry may have to be replaced by a more
general construction, reducing at l  lP to the “standard” one).
Coordinates are introduced to quantify a space and objects associated
with it. One can introduce many coordinate systems for the same space,
e.g. Cartesian, spherical or cylindrical coordinates for the 3-dimensional Eu-
clidean space R3 , or Cartesian or polar for R2 . (See Morse and Feshbach
“Methods of Theoretical Physics”, Chapter 5, for a list of some useful coor-
dinate systems.)
Suppose we have two coordinate systems in n-dimensional space M n (not
necessarily Euclidean): xi = (x1 , x2 , ... xn ) and x0 i = (x0 1 , x0 2 , ... x0 n ) . Here
i = 1, 2, ...n. Each point p ∈ M n is characterised by the set of coordinates (ei-
ther xi or x0 i ), and there is one-to-one correspondence1 x0 i = x0 i (x) between
the two descriptions at a point p provided the determinant of the Jacobi
matrix
∂x0 i
Jji = (1)
∂xj
known as the Jacobian does not vanish at this point: J = det Jji 6= 0. Points
where J = 0 are known as coordinate singularities (they are singularities
associated with a given coordinate system, not the space itself).
Excercise: Compute Jji and J for the sets of Cartesian and polar coor-
dinates in R2 and Cartesian and spherical coordinates in R3 . Identify the
coordinate singularities.
Now consider vectors on M n , e.g. the velocity vector of a point moving in

1
All transformations x0 i = x0 i (x) are assumed to be smooth, e.g. of C ∞ class. The important class of discrete
transformations (such as xi → −xi ), including parity inversion (x → −x, y → −y, z → −z) and time reversal
(t → −t), are considered separately.

3
M n . Vectors are specified by their components ai (x) = (a1 (x), a2 (x), ... an (x))
at each point x ∈ M n . Consider the gradient of a function f in the direction
of ai :
∂f
ai ∇i f = ai (x) , (2)
∂xi
where summation over repeated indices is assumed (this is known as “Einstein
summation convention”). What happens to this expression if we write it in
the new coordinates x0 i = x0 i (x)? We have

i ∂f i 0 ∂f ∂x0 j 0 j 0 ∂f
a (x) i = a (x(x )) 0 j = a (x ) 0 j , (3)
∂x ∂x ∂xi ∂x
where

0j 0∂x0 j i
a (x ) = i
a (x(x0 )) (4)
∂x
is the law of transformations of vectors (old name: contravariant vectors),
or, more precisely, vector’s components, under the coordinate transformation
x0 i = x0 i (x). Eq. (4) appears naturally: indeed, we could have started in x0
coordinates, writing the gradient of the function as on the RHS of Eq. (3)
(its functional form should not depend on the choice of coordinates). In fact,
we can define a vector in a way independent of the choice of coordinates by
∂
v = ai (x) , (5)
∂xi
where ∂/∂x1 , ∂/∂x2 ...∂/∂xn can be thought as the basis in the linear vector
space, similar to the unit vectors i, j, k in V = a1 i + a2 j + a3 k. Sometimes,
the notation ei ≡ ∂/∂xi is used. Then
∂
v = ai (x) i
= ai (x) ei (6)
∂x

4
are contravariant vectors (or just vectors) with components ai (x). More
precisely, they are vector fields, since ai are not constant but depend on x.
Similarly, consider the differential of a function, df = bi (x)dxi . Under
x0 i = x0 i (x), we have

∂xi 0 j
df = bi (x)dxi = bi (x(x0 )) 0 j
dx = b0j (x0 )dx0 j , (7)
∂x
where
∂xi
b0j (x0 ) = 0 j bi (x(x0 )) (8)
∂x
is the law of transformation of covectors or differential forms (old name: co-
variant vectors) under the coordinate transformation x0 i = x0 i (x). In fact,
we can define a covariant vector in a way independent of the choice of coor-
dinates by

v ∗ = bi (x)dxi , (9)

where dx1 , dx2 ...dxn can be thought as the basis in the linear vector space,
i.e. notation ei ≡ dxi is used. Then

v ∗ = bi (x)dxi = bi (x) ei (10)

are the covariant vectors with componets bi (x). More precisely, they are
covector fields, since bi are not constant but depend on x.
Denoting the space of all vectors by V and all covectors by V ∗ , we see
that there us a natural map V ⊗ V ∗ → R (in principle, other fields such as
C can be used as well, but some care should be exercised then, especially in
the case of curved spaces) given by
 
∗ i ∂
j
v (v) = bj dx a i = ai bi ∈ R . (11)
∂x

5
 We can think of generalising these constructions to objects with more than
one index. For example,

w∗ = cij (x)dxi dxj (12)

is an obvious generalisation of (9). A more highbrow notation is

w∗ = cij (x)dxi ⊗ dxj (13)

but it is really the same thing. An operation

 
∂
w∗ (v) = cij (x)dxi ⊗ dxj ak k = cij aj dxi ∈ V ∗ (14)
∂x
is a map V ∗ ⊗ V ∗ ⊗ V → V ∗ which can be seen as linear operators (matrices)
acting on vectors.
Obviously, we can add more components

w∗ = pijk (x)dxi ⊗ dxj ⊗ dxk , (15)

and so on. For vectors we have,

∂ ∂ ∂
w = hijk (x) ⊗ ⊗ , (16)
∂xi ∂xj ∂xk
and we can have mixed objects as well, such as
∂ ∂
t = sij
k (x) i
⊗ j ⊗ dxk , (17)
∂x ∂x
A generic tensor (more precisely - tensor field, since components depend
on x) then is an object

i i ...i ∂ ∂
T = Tj11j22...jpq (x) i
⊗ · · · ⊗ i
⊗ dxj1 · · · ⊗ dxjq , (18)
∂x 1 ∂x p

i i ...i
whose components Tj11j22...jpq (x) transform under a continuous x0 i = x0 i (x) such
that each upper index transforms as in (4) and each lower index - as in (8),

6
i.e.

0 i i ...i ∂x0 i1 ∂x0 i2 ∂x0 ip ∂xl1 ∂xl2 ∂xlq k1 k2 ...kp

Tj1 j12 ...j
2 p
(x0 ) = k k
· · · k 0 j 0 j
· · · 0 j
Tl1 l2 ...lq (x(x0 ))
q
∂x ∂x1 2 ∂x ∂x ∂x
p 1 2 ∂x q

(19)
Tensors are often called rank (p, q)-tensors, specifying the number of upper
(contravariant) and lower (covariant) components. In N -dimensional space,
a generic rank (p, q)-tensor has N p+q components.
The simplest example of the transformation law (19) is a transformation
of a scalar ϕ(x) (under continuous x0 i = x0 i (x)):

ϕ0 (x0 ) = ϕ(x) . (20)

Note: if, in addition to the property (20) under a continuous transforma-

tion, ϕ(x) changes sign under a parity transformation xi → x0i = −xi , it is
called a pseudoscalar. If the sign remains the same it is sometimes called a
true scalar. The same terminology applies to higher tensors, e.g. we have
pseudovectors etc.
Another important example of a (0, 2) tensor is the metric gij (x):

0 ∂xρ ∂xσ
gµν (x0 ) = 0 µ 0 ν gρσ (x(x0 )) (21)
∂x ∂x
Note: since rank 2 tensors are represented by matrices, the transformation
(21) can be written as

G0 (x0 ) = S T G(x)S , (22)

where
∂xρ
S ρµ (x) = . (23)
∂x0 µ

7
In special relativity, the matrix Λ representing Lorentz transformations x0 =
Λx is independent of x and is given by
 
γ −γβ 0 0
 
−γβ γ 0 0
 
Λ= 
,
 (24)
 0 0 1 0
 
0 0 0 1

whereas
 
γ γβ 0 0
 
γβ γ 0 0
 
Λ−1 = 

,
 (25)
0 0 1 0
 
0 0 0 1
p
where β = v/c and γ = 1/ 1 − β 2 . Thus, S = Λ−1 . Now, G is the
Minkowski metric tensor (normally denoted by η)
 
−1 0 0 0
 
0 1 0 0
 
G=η= 
.
 (26)
0 0 1 0
 
0 0 0 1

One can easily check that η 0 = S T ηS = η, i.e. the Minkowski metric is

invariant under the Lorentz transformations.
Important note: NOT EVERY OBJECT WITH INDICES IS A
TENSOR. A canonical example here is the connection coefficient Γijk (x) of
GR which is not a tensor (see e.g. [3]). To check whether an object with
indices is a tensor, one has to check the transformation law (19) explicitly or
use some simple facts of tensor algebra:

8
 • A linear combination of (p, q)-tensors tensors is a (p, q)-tensor

• A contraction of tensors is a tensor

If S ijk and Tijlm are tensors, so is

S ijk Tijlm = Ulm

k
,

where summation over repeated (“dummy”) indices is assumed. An

important contraction is

T = Tii

known as “trace” (or “Spur” in German), with the notation tr (or Sp).

Tensor product: For two tensors, A and B, one can define a tensor
product S = A ⊗ B. For example, if A = Ai and B = B j are vectors,
S ij = Ai B j . E.g. in d = 2 we have
 
1 1 1 2
AB AB
S= .
2 1 2 2
AB AB
This can be extended to tensors of any rank. Note that the tensor product
operation is generically not commutative (A ⊗ B 6= B ⊗ A) but associative.

A. Special tensors

• The Kronecker tensor δji (the identity matrix) is a (1, 1)-rank tensor. It
is the same in all coordinate systems,

δj0i = δji ,

as can be seen from the transformation law of tensors. Note that lowering or
raising indices of δji we get the metric tensor or its inverse,

gij δkj = gik , g ij δjk = g ik , g ij gjk = δki .

9
In this sense, the notations δij and δ ij only make sense in Euclidean space,
where the metric itself is a unit matrix, gij = δij .
• The Levi-Civita absolutely antisymmetric (pseudo) tensor. In R3 , we
had a useful object εijk , where ε123 = +1 and any interchange of indices
changes the sign. In 4d Minkowski space, we define a similar object with
ε0123 = −ε1023 = · · · = +1. Note that ε0123 = −1. (Also note that some
authors define ε0123 = −1.)
In general curvilinear coordinates, one can introduce a generalisation of
this object (written here in 4 dimensions)
p
ijkl (x) = |g(x)|εijkl ,

where g = det gij is the determinant of a metric tensor. Such an object is a

covariant tensor. The corresponding contravariant tensor is
1
ijkl (x) = p εijkl ,
|g(x)|
whereas εijkl is known as tensor density (more details can be found e.g. in
[3] or in the exercises in [5]).

B. Vector components in curvilinear coordinates

It is helpful to consider a number of standard examples familiar from earlier

studies, such as the orthogonal curvilinear coordinates (polar, cylindrical,
spherical) in R3 . It is important to emphasize that these are coordinates in
flat space (the criterium for this is simple - all components of the Riemann
curvature tensor for a given metric are zero and thus there exists a coordinate
transformation bringing the metric into the form ds2 = dx2 + dy 2 + dz 2 ).

10
Nevertheless, such coordinates exhibit non-trivial features. For example, the
connection coefficients or Christoffel symbols for them are non-zero (they
are known as flat connections since the curvature tensor remains zero), and
therefore the covariant derivative is non-trivial, and so on.
For a cylindrical coordinate system, we have (here xµ = (x, y, z) are Carte-
sian coordinates, and x0µ = (r, φ, z)):

x = r cos φ (27)

y = r sin φ (28)

z=z (29)

The metric tensor is

 
1 0 0
 
 2 
gij = 0 r 0 .
 
0 0 1

The inverse metric is given by

 
1 0 0
 
ij  1 
g = 0 r 2 0  .
 
0 0 1

To represent a vector in the new system, usually a set of basis unit vectors r̂,
φ̂, ẑ, similar to the Cartesian unit vectors î, ĵ and k̂, is introduced, so that

~ = Âx î + Ây ĵ + Âz k̂ = Âr r̂ + Âφ φ̂ + Âz ẑ .

A (30)

Vector components with hats such as Âr were introduced to distinguish them
from the contravariant and covariant componets Ai and Ai used earlier. We
shall explain the difference shortly.

11
 More generally, one can write

d~r = ei dxi = e0k dx0k , (31)

with
∂xi
e0k = 0k ei . (32)
∂x
For the cylindrical cordinates, with e1 = ex ≡ î, e2 = ey ≡ ĵ, e3 = ez ≡ k̂
and e01 = er , e02 = eφ , , e03 = ez , we have from Eq. (32)

er = cos φ î + sin φ ĵ + 0 k̂ , (33)

eφ = −r sin φ î + r cos φ ĵ + 0 k̂ , (34)

ez = 0 î + 0 ĵ + 1k̂ . (35)

This can also be written as

∂ ∂x ∂ ∂y ∂ ∂z ∂ ∂ ∂ ∂
= + + = cos φ + sin φ +0 , (36)
∂r ∂r ∂x ∂r ∂y ∂r ∂z ∂x ∂y ∂z
∂ ∂x ∂ ∂y ∂ ∂z ∂ ∂ ∂ ∂
= + + = −r sin φ + r cos φ +0 ,(37)
∂φ ∂φ ∂x ∂φ ∂y ∂φ ∂z ∂x ∂y ∂z
∂ ∂x ∂ ∂y ∂ ∂z ∂ ∂ ∂ ∂
= + + =0 +0 +1 , (38)
∂z ∂r ∂x ∂r ∂y ∂r ∂z ∂x ∂y ∂z
which is an explicit form of the identification ei = ∂/∂xi .
Note that the basis vectors er , eφ , ez are not normalised to unity, e.g.
√
|eφ | = eφ · eφ = r. One can introduce the normalised vectors
er eφ ez
r̂ = , φ̂ = , ẑ = ,
|er | |eφ | |ez |
whose Cartesian coordinates are r̂ = (cos φ, sin φ, 0), φ̂ = (− sin φ, cos φ, 0),
ẑ = (0, 0, 1). In general, êi = ei /|ei |, where |ei |2 = ei · ei = gαβ eα(i) eβ(i) .
~ will be different in these two bases,
Clearly, components of a vector A

~ = Ai ei = Ai |ei |êi = A1 |er |r̂ + A2 |eφ |φ̂ + A3 |ez |ẑ .

A (39)

12
 Raising the indices with the metric g ij , we get

er = cos φ î + sin φ ĵ + 0 k̂ , (40)

1 1
eφ = − sin φ î + cos φ ĵ + 0 k̂ , (41)
r r
ez = 0 î + 0 ĵ + 1k̂ . (42)

Note that ei · ej = gij , ei · ej = g ij , and ei · ej = δij . We can introduce êi =

ei /|ei |. For diagonal metrics, êi = g ij ej /|g ij ej | = êj sgn(g ij ). In particular,
êr = êr , êφ = êφ , êz = êz . This is convenient, since in the expansion of a
vector
~ = Ai ei = Ai |ei |êi = Âi êi = Ai ei = Ai |ei |êi = Âi êi
A (43)

we have Âi = Âi , i.e. there is no difference between components with upper
and lower indices in this basis. For details, see [? ] and [? ]. This basis is
typically used when dealing with orthogonal curvilinear coordinates in R3 .
We note that for a generic curved space-time with the metric tensor gµν (x)
the standard bases ∂/∂xi (for vectors) and dxi (for covectors) are used, and,
respectively, one has the standard contravariant and covariant components
Ai and Ai .

C. Differential operators

In curved space (and even in flat space when using curvilinear coordinates)
one has to generalise various differential operations accordingly. This is fully
considered in GR courses but we mention some operations here. When dif-
ferentiating tensors, ordinary derivatives should be replaced with covariant
derivatives. For example, acting on vectors, the covariant derivative is

∇i Aj = ∂i Aj + Γjik Ak ,

13
where Γjik are Christoffel symbols (coefficients of the metric connection). We
also have

∇i Aj = ∂i Aj − Γkij Ak ,

Some operations can be easily generalised any dimension, for example, the
divergence ∇i Ai , whereas others, such as curl, may be dimension-specific and
are replaced in other dimensions by more general constructions.
~ in 3d can be written as
The curl of a vector A
 i
~ = ijk ∇j Ak ,
curlA

where the tensor ijk is defined as

1
ijk (x) = p εijk ,
|g(x)|
whereas εijk is the permutation coefficient, with ε123 = 1. For example, in
spherical coordinates in R3 , g = r4 sin2 θ and e.g. the r-component of a curl
is given by
 r
~ = εrθφ
curlA (∇θ Aφ − ∇φ Aθ ) ,
r2 sin θ
where

∇θ Aφ = ∂θ Aφ − Γkθφ Ak ,

and

∇φ Aθ = ∂φ Aθ − Γkφθ Ak .

Since for metric connection Γkθφ = Γkφθ , and εrθφ = 1, we have


~
r 1
curlA = 2 (∂θ Aφ − ∂φ Aθ ) .
r sin θ
14
Remembering our discussion of different bases for curvilinear coordinates in
flat space, we note that Aφ = r sin θÂφ and Aθ = rÂθ . Correspondingly, we
have

~
r 1 h i
curlA = ∂θ (sin θÂφ ) − ∂φ Âθ .
r sin θ
Typically, this is the expression that appears in the standard literature such
as ref. [2]. Finally, we note that one can also write the coordinate-free ex-
pression for curl in 3d space as

~ = ? dA
curlA ~,

where ? denotes the Hodge dual operator.

15
III. PROBLEM SET I: SOLUTIONS

Problem 1
If we have two successive transformations from ui = ui (x1 , x2 , . . . xN ) to v i =
v i (y 1 , y 2 , . . . y N ), and from v i to wi = wi (z 1 , z 2 , . . . z N ), with i = 1, 2, . . . N ,

i∂y i j
v = ju , (44)
∂x
and

i∂z i j
w = jv , (45)
∂y
show that we can perform the transformation in one step via

i∂z i j
w = ju . (46)
∂x

Solution: What we have here is the change of coordinates

x → y = y(x) → z = z(y) .

This change induces the change of the components of a vector (rank (1, 0)
tensor)

ui = ui (x) → v i = v i (y) → wi = wi (z)

according to the standard rules of transformation of a tensor, here given

explicitly by Eqs. (44) and (45). Combining Eqs. (44) and (45), we can write
(note that we can use any letter to label dummy indices, and freely change
it as long as it doesn’t coincide with the one already used in the expression)

i∂z i ∂y k j ∂z i j
w = k ju = ju , (47)
∂y ∂x ∂x

16
where the right hand side is a consequence of the rule of differentiating a
composite function z = z(y(x)):
∂z ∂z ∂y
= .
∂x ∂y ∂x

17
Problem 2
If Aij ij
k is a mixed tensor, Bk is another tensor of the same kind, and α and

β are scalar invariants, show that αAij ij

k + βBk is yet another tensor of the

same kind.
Solution: This is a particular case of a tensor algebra property: a linear
combination of tensors is a tensor. To prove it, we only need to use the
definition of a tensor (19). Specifically, the (2, 1)-rank tensors Aij ij
k and Bk

transform under x → x0 = x0 (x) as

∂x0 i ∂x0 j ∂xr p q
A0kij (x0 ) = p q 0 k
Ar (x(x0 )) , (48)
∂x ∂x ∂x

∂x0 i ∂x0 j ∂xu s t

Bk0 ij (x0 ) = s t 0 k
Bu (x(x0 )) , (49)
∂x ∂x ∂x
How does tensor Ckij = αAij ij
k + βBk transform? Well,

∂x0 i ∂x0 j ∂xr p q ∂x0 i ∂x0 j ∂xu s t

Ck0 ij (x0 ) = α 0
A0ij
+β 0
Bk0ij
=α p 0
A (x(x )) + β s B (x(x0 ))
k
∂x ∂xq ∂x0 k r ∂x ∂xt ∂x0 k u
∂x0 i ∂x0 j ∂xr 0 0 ∂x0 i ∂x0 j ∂xr p q
= p q 0 k
pq pq
(αAr (x(x )) + βBr (x(x ))) = p q 0 k
Cr (x(x0 )) ,
∂x ∂x ∂x ∂x ∂x ∂x
since the scalars transform as α0 (x0 ) = α(x), β 0 (x0 ) = β(x), and the dummy
indices can be changed at will. Thus, Ckij transforms as a (2, 1)-rank tensor.

18
Problem 3
If Aij are the components of a mixed tensor, show that Aii transforms as a
scalar.
Solution: Under x → x0 = x0 (x), the tensor Aij itself transforms as

∂x0 i ∂xq p
A0ji (x0 ) = p 0 j
Aq (x(x0 )) . (50)
∂x ∂x
The contracted tensor Aii therefore transforms as
∂x0 i ∂xq p
A0i i (x0 ) = p 0 i
Aq (x(x0 )) . (51)
∂x ∂x
But
∂xq ∂x0 i ∂xq
0 i p
= p = δpq .
∂x ∂x ∂x
Therefore,

A0i i (x0 ) = App (x(x0 )) = Aii (x(x0 )) . (52)

This means that Aii (x) is a scalar.

Note: For any tensor, full contraction of upper indices with lower ones is
a scalar. More generally, a contraction of a pair of indices (one upper and
one lower) lowers the rank of the tensor from (p, q) to (p − 1, q − 1).

19
Problem 4
Assuming x and y transform as the components of a Euclidean vector, de-
termine which of the following matrices are tensors:
     
2 2 2
x xy xy y y xy
Aij =   , B ij =   , C ij =  .
2 2 2
xy y x −xy xy x

Solution: To make this question more clear, introduce notations for coor-
dinates x1 = x, x2 = y and components of a vector v i = (v 1 , v 2 ) = (vx , vy ).
In this particular case v i is a vector with components vx = x and vy = y. So,
for example,
 
vx2 vx vy
Aij =  ,
vx vy vy2

and so on. Under the change of coordinates

x → x0 = x0 (x, y) , y → y 0 = y 0 (x, y)

the components v i transform as

∂x0 i p
v 0 i (x0 ) = p
v (x(x0 )) . (53)
∂x
Explicitly, we have
∂x0 ∂x0
vx0 (x0 ) = 0
vx (x(x )) + vy (x(x0 )) , (54)
∂x ∂y
0
∂y ∂y 0
vy0 (x0 ) = vx (x(x0 )) + vy (x(x0 )) . (55)
∂x ∂y
If Aij is a tensor, its components should transform as
∂x0 i ∂x0 j p q
A0 ij (x0 ) = p q
A (x(x0 )) . (56)
∂x ∂x

20
For example, we should have
∂x0 ∂x0 1 1 ∂x0 ∂x0 1 2 ∂x0 ∂x0 2 2
A0 11 (x0 ) = A (x(x0 )) + 2 A (x(x0 )) + A (x(x0 ))
∂x ∂x ∂x ∂y ∂y ∂y
 0 2 2
∂x0 ∂x0 ∂x0
 
∂x
= vx2 + 2 vx vy + vy2 . (57)
∂x ∂x ∂y ∂y
Does this hold? We know that A0 11 (x0 ) = vx02 , and we know that vx0 trans-
forms as in Eq. (54). Then vx02 transforms exactly as Eq. (57), i.e. exactly
as expected from a tensor component. Similar checks for the remaining 3
components show that Aij is indeed a tensor. Similarly, one can check that
B ij and C ij are not tensors.

21
Problem 5
Show that if the components of a contravariant vector vanish in one coor-
dinate system, they will vanish in all coordinate systems. What can be said
of two contravariant vectors whose components are equal in one coordinate
system?
Solution: A contravariant vector Ai transforms under x → x0 = x0 (x) as

0i 0 ∂x0 i j
A (x ) = j
A (x(x0 )) . (58)
∂x
If all components Aj (x) vanish, then all components A0 i (x0 ) vanish as well.
Note: this fact is of fundamental importance and explains why tensors are
used widely in physics. Indeed, fundamental physical laws can written as
tensor identities: S ν = 0, T µν = 0 etc. For example, Maxwell equations are
written in the covariant form as S ν ≡ ∂µ F µν − µ0 J ν = 0, Einstein equations
in vacuum as Rµν = 0 etc. Clearly, the laws should not depend on the
coordinate system used to write them, i.e. we should have S 0 ν (x0 ) = 0 etc.
This is guaranteed by the tensorial transformation property such as the one
in Eq. (58). If all components of a tensor vanish in some coordinate system,
they vanish in all coordinate systems.

22
Problem 6
Let Aij be a skew-symmetric tensor with Aij = −Aji , and Sij a symmetric
tensor with Sij = Sji . Show that the symmetry properties are preserved in
coordinate transformations. Also show that the quantities with raised indices,
Aij and S ij , possess the same properties. From this, show that Aij Sij = 0
and Aij S ij = 0.
Solution:
• We need to show that Aij (x) = −Aji (x) implies A0ij (x0 ) = −A0ji (x0 ). The
transformation properties of Aij are
∂xm ∂xn
A0ij (x0 ) = 0 i 0 j
Amn (x(x0 )) . (59)
∂x ∂x
Since m and n are dummy indices, we can write
∂xm ∂xn ∂xn ∂xm ∂xm ∂xn
0 i 0 j
Amn (x(x )) = 0 i 0 j Anm (x(x )) = − 0 j 0 i Amn (x(x0 )) = −A0ji (x0 ) .
0 0
∂x ∂x ∂x ∂x ∂x ∂x
(60)
For the symmetric tensor, exactly the same argument works (except for the
minus sign).
• For the same tensors with upper indices, we have

0 ij 0∂x0 i ∂x0 j p q 0 ∂x0 i ∂x0 j q p 0 ∂x0 j ∂x0 i p q

S (x ) = p q
S (x(x )) = q p
S (x(x )) = p q
S (x(x0 )) = S 0 ji (x0
∂x ∂x ∂x ∂x ∂x ∂x
(61)
and similarly for Aij (x) = −Aji (x).
• Finally, using the fact that Aij Sij is a scalar (all indices are contracted)
and symmetry propeties of Aij and Sij , we find

Aij Sij = Aji Sji = −Aij Sij (62)

which implies Aij Sij = 0. Similarly, Aij S ij = 0.

23
 Note: the last property is widely used in practical calculations to simplify
expressions. Note also that any (0, 2) tensor Tij can be decomposed into
symmetric and antisymmetric parts:

Tij = Tijsym + Tijasym ,

where
1 1
Tijsym = (Tij + Tji ) , Tijasym = (Tij − Tji ) .
2 2
Note: symmetry properties reduce the number of independent components
of a tensor. For example, a generic tensor Tij in N -dimensional space has
N 2 componets. How many components do Tijsym and Tijasym have?

24
Problem 7
Let C kl = Aijk Bijl be a rank-2 contravariant tensor given by contracting the
N 3 functions Aijk with the tensor Bmn
l
, which is symmetric in the mn indices
l l
but otherwise arbitrary, i.e. Bmn = Bnm . Show that Aijk + Ajik is a rank-3
contravariant tensor. Give reasons why the same is not true for Aijk and
Ajik separately.
Solution: We need to establish transformation properties of the object Aijk
under x → x0 (x). Since C kl is a tensor, we can write
0 kl 0 ∂x0 k ∂x0 l m n 0 ∂x0 k ∂x0 l qrm p
C (x ) = m n
C (x(x )) = m p
A Bqr (x(x0 )) . (63)
∂x ∂x ∂x ∂x
On the other hand,
0l
0 kl 0 0ijk 0ijk ∂x ∂xq ∂xr p
C (x ) = A Bij0l =A p 0i 0j
Bqr (x(x0 )) . (64)
∂x ∂x ∂x
Subtracting (63) from (64), we have
0k
 q r
 0l
∂x ∂x ∂x ∂x p
A0ijk 0i 0j − m Aqrm p
Bqr (x(x0 )) = 0 . (65)
∂x ∂x ∂x ∂x
In (65), k and l are free (uncontracted) indices. We can simplify this expres-
sion by multiplying it by e.g. ∂xs /∂x0l and summing over l. We get
q r
∂x0 k qrm
 
0ijk ∂x ∂x
A 0i 0j
− mA s
Bqr (x(x0 )) ≡ P kqr Bqr
s
= 0. (66)
∂x ∂x ∂x
s s s
Components of Bqr are not independent: Bqr = Brq . The sum in (66) is thus
of the form

P k11 B11
s
+ P k12 + P k21 B12

s
+ ··· = 0. (67)

s
For generic independent Bqr (with q ≤ r) this implies P krq + P kqr = 0 (this
may not be true individually for P krq ), i.e.
0ijk 0jik

∂xq ∂xr ∂x0 k
A +A 0i 0j
− m (Aqrm + Arqm ) = 0 . (68)
∂x ∂x ∂x
25
Thus, the sum Aijk + Ajik transforms as a rank (3, 0) tensor:

0ijk 0jik ∂x0i ∂x0j ∂x0 k

(Aqrm + Arqm ) .


A +A = (69)
∂xq ∂xr ∂xm

26
Problem 8
In this problem, we consider a transformation from Cartesian to polar coor-
dinate systems in two Euclidean dimensions. Let x1 = x and x2 = y for the
Cartesian system and x01 = r and x02 = θ for the polar, with the transforma-
tions

x1 = x = r cos θ = x01 cos x02 , (70)

x2 = y = r sin θ = x01 sin x02 . (71)

The metric for the Cartesian system is gij = δij . Derive the metric tensor
gij0 (x0 ) for the polar coordinate system, its reciprocal g 0ij , and the covariant
polar coordinates x01 and x02 in terms of r and θ. Why might it not be appro-
priate to calculate a length from the origin to a point specified by finite values
of r and θ using these covariant components?
Show that the components of the metrics gij and gij0 (x0 ) do not change
under rotations of the coordinate system through a fixed angle α around the
origin.
Solution: A metric is a (0, 2)-rank tensor (field) used to determine infinites-
imal distances via

ds2 = gij (x)dxi ⊗ dxj . (72)

Here we have gij = δij , and so

ds2 = dx2 + dy 2 (73)

(the symbol ⊗ is usually omitted). In matrix form:

 
1 0
gij =   .
0 1

27
Under x → x0 = x0 (x) the metric transforms as
∂xm ∂xn
gij0 (x0 ) = 0 i 0 j
gmn (x(x0 )) . (74)
∂x ∂x
Explicitly,
2  2
∂x1 ∂x1 ∂x2 ∂x2

0 ∂x ∂y
g11 (x0 ) = + = + , (75)
∂x0 1 ∂x0 1 ∂x0 1 ∂x0 1 ∂r ∂r
0 ∂x1 ∂x1 ∂x2 ∂x2 ∂x ∂x ∂y ∂y
g12 (x0 ) = + = + , (76)
∂x0 1 ∂x0 2 ∂x0 1 ∂x0 2 ∂r ∂θ ∂r ∂θ
0 ∂x1 ∂x1 ∂x2 ∂x2 ∂x ∂x ∂y ∂y
g21 (x0 ) = + = + , (77)
∂x0 2 ∂x0 1 ∂x0 2 ∂x0 1 ∂θ ∂r ∂θ
 2  2 ∂r
0 ∂x1 ∂x1 ∂x2 ∂x2 ∂x ∂y
g22 (x0 ) = + = + , (78)
∂x0 2 ∂x0 2 ∂x0 2 ∂x0 2 ∂θ ∂θ
resulting in
 
1 0
gij0 (x0 ) =  ,
0 r2
so the line element (often also called the metric) is

ds2 = dr2 + r2 dθ2 . (79)

Note that a more sraightforward way of transforming the metric is just

substituting x = r cos θ and y = r sin θ into ds2 = dx2 + dy 2 (we have
dx = cos θdr − r sin θdθ and dy = sin θdr + r cos θdθ, so the result (79)
follows immediately).
We now introduce a fundamental definition of the inverse metric:

The inverse metric g ij is a (2, 0)-rank tensor obeying g ik gkj = δji . (80)

In other words, this is just a matrix inverse to gij . In our particular case,
 
1 0
g 0 ij (x0 ) =  ,
1
0 r2

28
which is singular at r = 0. This is an example of a coordinate singularity: the
space itself is not singular (as can be seen by choosing a different cordinate
system, e.g. a Cartesian one, where the origin is an ordinary point). The
same type of singularity is exhibited by black hole horizons.
The metric and its inverse can be used for raising and lowering indices of
tensors, e.g.

x01 = g1j (x0 )x0j , x02 = g2j (x0 )x0j ,

Explicitly, we find

x01 = g11 (x0 )x01 + g12 (x0 )x02 = r , x02 = g21 (x0 )x01 + g22 (x0 )x02 = r2 θ ,

The objects x01 and x02 are not related to distances (in Cartesian coordinates,
x1 = x1 = x and x2 = x2 = y, but this is not true in general). The distance
(infinitesimal) is determined by Eq. (81) which can be written as

ds2 = gij (x)dxi ⊗ dxj = gij0 (x0 )dx0i ⊗ dx0j = dx0i dx0i = dr2 + r2 dθ2 . (81)

Computing finite distances, areas, volumes will in general require integration

p
with an invariant measure |g(x)|dx, where g(x) is the determinant of the
p
metric tensor. For polar coordinates, |g(x)| = r, so the measure of inte-
gration (in two dimensions) is the familiar rdrdθ (in Cartesian coordinates,
it is dxdy). More precisely, integrating a function f (x) over a region Ω in an
n-dimensional space with metric gij (x) is done via
Z p
I= |g(x)|f (x)dn x .
Ω

In particular, the volume of Ω corresponds to f (x) = 1:

Z p
V = |g(x)|dn x .
Ω

29
For example, in d = 2, the “volume” (area) of the circle of radius R is given
by
ZR Z2π
V2 = rdrdθ = πR2 .
0 0

How to compute the length of a curve? First of all, a curve is given by an

equation (e.g. in polar coordinates in d = 2 it is r = f (θ)). The Archimedean
spiral, for example, is described by the equation r = a + bθ. These equations
describe an embedding of our curve (or other subspace) into the ambient
space. We have dr = f 0 (θ)dθ. The induced metric on the cirve is given
by substituting this dr into the ambient metric ds2 = dr2 + r2 dθ2 giving
ds2induced = (f 02 + f 2 )dθ2 . The length is then
Z p Z θf p
L= |ginduced (x)|dx = f 02 + f 2 dθ .
θi
Ω

For the Archimedean spiral (for simplicity, we can set a = 0 and integrate
from θ = 0 to θ = θmax ) we have
Z θmax p
2
b p
2

L = b 1 + θ dθ = θmax 1 + θmax + arcsinh θmax
0 2
bh p
2
 p
2
i
= θmax 1 + θmax + ln θmax + 1 + θmax . (82)
2
A rotation by the angle α around the origin in e.g. counterclockwise di-
rection is described by z 0 = eiα z, where z = x + iy. This can be written in
the form
    
0
x cos α − sin α x
 =  .
y0 sin α cos α y

Computing dx0 and dy 0 , one can see that the line element ds2 = (dx0 )2 +
(dy 0 )2 = dx2 + dy 2 , and so the metric tensor gij is not affected by this

30
transformation. The same can be also seen in polar coordinates, where
r0 = r and θ0 = θ + α, so that dr0 = dr and dθ0 = dθ, and thus
ds2 = dr02 + r02 dθ02 = dr2 + r2 dθ2 .

[1] A. Steane, Relativity Made Relatively Easy (Oxford University Press, 2012).
[2] J. Jackson, Classical Electrodynamics (Wiley, 1998).
[3] S. Weinberg, Gravitation and Cosmology (John Wiley and Sons, 1972).
[4] L. Landau and E. Lifshitz, The Classical Theory of Fields: Volume 2 (Course of Theoretical Physics)
(Butterworth-Heinemann, 1980).
[5] A. Lightman, W. Press, R. Price, and S. Teukolsky, Problem Book in Relativity and Gravitation (Prince-
ton University Press, 2017).
[6] B. Dubrovin, A. Fomenko, and S. Novikov, Modern Geometry - Methods and Applications: Part I: The
Geometry of Surfaces, Transformation Groups, and Fields (Graduate Texts in Mathematics) (Springer,
1992).
[7] B. Dubrovin, A. Fomenko, and S. Novikov, Modern Geometry Methods and Applications: Part II: The
Geometry and Topology of Manifolds (Graduate Texts in Mathematics) (Springer, 1985).
[8] M. Nakahara, Geometry, Topology and Physics (CRC Press, 2003).
[9] S. Carroll, Spacetime and Geometry: An Introduction to General Relativity (Cambridge University Press,
2019).

31