General Relativity Fall 2019

Lecture 7: basic operations on tensors; tensor fields

Yacine Ali-Haı̈moud
September 24th 2019

MORE ON TENSORS

• Component transformations. Remember that a tensor of rank (k, l) can be written as

T = T µ1 ...µk ν1 ...νl e(µ1 ) ⊗ ... ⊗ e(µk ) ⊗ e∗(ν1 ) ⊗ ... ⊗ e∗(νl ) ,

where T µ1 ...µk ν1 ...νl are the components of the tensor T . Let us consider a change of basis, e(µ0 ) = M µ µ0 e(µ) . We
0
define the inverse of M µ µ0 by M µ µ , i.e.
0 0 0
M µ µ0 M µ ν = δνµ , M µ µ M µ ν 0 = δνµ0 .
0 0
We then have e∗(µ ) = M µ µ e∗(µ) . Combining these results, and the fact that the tensor T itself is unchanged under
a coordinate transformation, we find that its components change according to
µ01 ...µ0k µ01 µ0k ν1 νl µ1 ...µk
T ν10 ...νl0 =M µ1 ...M µk M ν 0 ...M ν 0 T ν1 ...νl [general change of basis]
1 l

This0 holds for

0
arbitrary basis transformations. If we now specialize to transformations between coordinate bases,
M µ µ = ∂xµ /∂xµ , we get
0 0
µ01 ...µ0k ∂xµ1 ∂xµk ∂xν1 ∂xνl µ1 ...µk
T ν10 ...νl0 = ... 0 ... 0 T ν1 ...νl . [change of coordinate basis]
∂xµ1 ∂xµk ∂xν1 ∂xνl
Note that this is sometimes used a a definition of a tensor: some object which trnasforms according to this rule
under coordinate transformations. While this is an equivalent definition, the one that we started with – a linear map
T : (Vp∗ )k × Vpl → R – is somewhat more fundamental, in the sense that it does not require any basis.

• Tensor product in abstract notation

We have already defined the tensor product T ≡ X ⊗ Y of two vectors X, Y . In any basis, the components of
X ⊗ Y are T µν = X µ Y ν . Using abstract index notation, we can directly define the tensor T αβ ≡ X α Y β as the tensor
product of X α and Y β .
Similarly, we can define the tensor produt of two tensors of arbitrary rank T ≡ R ⊗ S. This notation does not
convey the rank of the tensors, though. So we can instead write something like T αγβδλ ≡ Rαγ S β δλ to indicate that
the rank-(2, 3) tensor T is the tensor product of the rank-(1, 1) tensor Rαγ and the rank-(1, 2) tensor S β δλ . You can
think of the abstract index notation as meaning “in any basis, the components of such tensor are equal to...”.

• Contraction of indices. Given a rank-(1,1) tensor T αβ , with components T µν in a specific basis, one can
compute the contraction T µµ . This is quantity is independent of the basis:
0 0
T µµ0 = M µµ M νµ0 T µν = δµν T µν = T µ µ .
As a consequence, we may use the abstract index notation T αα to denote the contraction of T αβ . More generally,
one can contract up and down indices for a tensor of any rank (k ≥ 1, l ≥ 1), to get a tensor of rank
(k − 1, l − 1). For instance, given a rank-(1, 2) tensor T αγδ , we can compute the following rank-(0,1) tensors (i.e. dual
vectors), which are both contractions of a pair of indices:
T ααβ , T αβα .
Note that these two tensors are in general different. For instance, consider the tensor product T αβγ = X α Yβ Zγ .
We then have
T ααβ = (X · Y )Zβ 6= (X · Z)Yβ = T αβα .
 2

• Symmetry, antisymmetry. A tensor T αβ is said to be symmetric if T βα = T αβ . This statement holds in

any basis if it holds in one, as you can check explicitly, thus one can write it in abstract index notation. Similarly,
T αβ is antisymmetric if T βα = −T αβ . The same definition holds if both indices are down. However, it does not
make any sense to talk about symmetry or antisymmetry of two indices which are not both up or both
down.
To understand this, think of tensors as linear maps. Consider for instance a rank-(0,2) tensor Tαβ . This is a linear
map T : Vp2 → R. The fact that the tensor is symmetric/antisymmetric implies (as you should check for yourselves!)
that for any pair of vectors X, Y ,

T (X, Y ) = ±T (Y , X), [+ for symmetric, −for antisymmetric].

Note that this is a very strong statement: for a generic rank-(0,2) tensor, T (X, Y ) and T (Y , X) in general have
nothing to do with one another – in other words, tensors need not be either symmetric or antisymmetric:
they can also have no symmetry property at all.
Now consider instead a rank-(1, 1) tensor S αβ , this is a map S : Vp∗ × Vp → R. It taks a dual vector as its first
argument, and a vector as its second argument. It makes no sense to switch the arguments: the first slot can
only take a dual vector, not a vector.
The symmetry or antisymmetry of a tensor can be generalized to any pair of indices for a higher-rank ten-
sor. For instance, a tensor Tαβγδ is symmetric in its first two indices and antisymmetric in its last two if
Tαβγδ = +Tβαγδ = −Tαβδγ . A tensor S αβγδ is symmetric in first and third indices if S αβγδ = S γβαδ .

We can go further, and say that a tensor is fully symmetric of fully antisymmetric is a group of indices (all
of them up or all of them down, always!), if it is symmetric or antisymmetric under exchange of any pair within this
group. For instance, a tensor Tαβγδ is fully symmetric in its first 3 indices if

Tαβγδ = Tβαγδ = Tγβαδ = Tαγβδ ,

And a tensor Sαβγ is fully antisymmetric if

Sαβγ = −Sβαγ = −Tγβα = −Sαγβ .

Any permutation σ of {1, ...N } can always be written as a succession of pairwise exchanges. The signature s(σ) of a
permutation is +1 if it can be factorized as en even number of pairwise exchanges, and -1 if it is an odd number – it
turns out that this is a unique property, i.e. the signature is not changed by factorizing a permutation in a different
way. So more generally, a tensor is said to be fully symmetric in N indices if, for any permutation σ of N indices,

Tασ(1) ...ασ(N ) = Tα1 ...αN ,

and it is said to be fully antisymmetric in N indices if

Tασ(1) ...ασ(N ) = s(σ)Tα1 ...αN .

These properties are defined in general for a tesnor having more than N indices, but I didn’t try to be comprehensive
to not clutter the notation. Note that the dimension n of the manifold/tangent space is not to be confused
with the number N of indices that can be symmetric/antisymmetric.

• Symmetrization and antisymmetrization. Given a tensor Tα1 ...αN with (at least) N indices (all up or all
down), we can define the following tensors:

1 X 1 X
T(α1 ...αN ) ≡ Tασ(1) ...ασ(N ) , T[α1 ...αN ] ≡ s(σ)Tασ(1) ...ασ(N ) ,
N! N!
σ∈SN σ∈Sn

where the sum is over the set SN of the N ! distinct permutations of {1, ..., N }. You can check for yourselves
that these two tensors are, respectively, fully symmetric and fully antisymmetric – what you’ll need is the property that
the signature of the composition of two permutations is the product of the signatures. These two tensors are called
the symmetric and antisymmetric parts of T . In general T is made of a symmetric part, an antisymmetric
part, and a non-symmetric part.
 3

Let’s look at examples. For N = 2 (again, this has nothing to do with n, the dimension of the manifold!), there are
two perumtations of {1, 2}: the identity (signature +1) and the exchange of the two elements (signature -1). So we
get
1 1
T(αβ) = (Tαβ + Tβα ) , T[αβ] = (Tαβ − Tβα ) .
2 2
It turns out, for N = 2 only, that any tensor with (at least) 2 indices can be decomposed into a symmetric and an
antisymmetric part:
Tαβ = T(αβ) + T[αβ] .
Now let us look at the case N = 3, i.e. 6 permutations. The symmetric and antisymmetric parts of a tensor are,
respectively
1
T(αβγ) = (Tαβγ + Tβγα + Tγαβ + Tβαγ + Tγβα + Tαγβ )
6
1
T[αβγ] = (Tαβγ + Tβγα + Tγαβ − Tβαγ − Tγβα − Tαγβ ) .
6
Note that cyclic permutations of 3 indices {1, 2, 3} 7→ {2, 3, 1} and {1, 2, 3} 7→ {3, 1, 2} can be obtained by two pairwise
exchanges, hence have signature +1. It should hopefully be clear that, in general,
T(αβγ) + T[αβγ] 6= Tαβγ ,

which means that there is more to a rank-3 tensor – whether it is (0, 3) or (3, 0) – than its symmetric part and its
antisymmetric part. The same thing holds for N ≥ 3.

• The identity tensor. Let us define the rank-(1, 1) tensor δ such that
δ(X, Y ) = X · Y .
∗(µ)
Now, to find its components δνµ on the basis {e ⊗ e(ν) }, all we have to do is compute
(
µ ∗(µ) ∗(µ) 1 if µ = ν,
δν = δ(e , e(ν) ) = e · e(ν) =
0 otherwise
Thus we see that the components of δ on any basis are always the Kronecker delta symbols. We can then use δβα to
both mean the Kronecker delta, as well as the geometric object which is the identity tensor. You can ckeck explicitly
that the components of δ are independent of the basis. Note that this would not apply, for instance, for the Minkowski
symbol: this is invariant only under Lorentz transformations, but not under general changes of basis. Thus, while δβα
is a bona fide tensor, ηµν is only a Lorentz-tensor, but not a true tensor.

VECTOR AND TENSOR FIELDS

So far all we have said about vectors, dual vectors and tensors applies to one point p ∈ M, at which we can
define Vp , Vp∗ , and subsequently, tensors. Now we can define the tangent space at every point in the manifold –
the union of all tangent spaces is called the tangent bundle. As a consequence, we can define vectors, dual vectors,
and tensors at every point in M (or, more generally, on a subset of M, for example, a curve, as is the case for the
tangent vector d/dλ along a curve). These are vector and tensor fields.

It is best to work with smooth vector fields. How to define smoothness? Recall that a vector V |p at p is defined
as a linear map from smooth functions to R. For a given function f ∈ F, we can define the function
(
M→R
fV :
p 7→ V |p (f )

The vector field V is said to be smooth if the function fV is smooth for any smooth function f . So we’ve built the
notion of smooth vector fields out of the notion of smooth scalar fields. Similarly, we can build the notion of smooth
dual vector fields (they must give smooth functions when applied to smooth vector fields), and smooth tensor fields
(must give smooth functions when applied to smooth vector fields and dual vector fields).

• Commutator of two vector fields: You will study this in the homework.
 4

THE METRIC TENSOR FIELD

We can finally formally introduce our old acquaintance, the metric tensor. This is a rank-(0, 2) smooth tensor
field gαβ , which is symmetric, and moreover non-degenerate, i.e. such that gαβ X β = 0 if and only if X α = 0.
This tensor field has special physical meaning, as it serves to compute “square norms” (which may be of either signs)
of tangent vectors, ||X||2 ≡ gαβ X α X γ . In a coordinate basis, we have

g = gµν dx(µ) ⊗ dx(ν) ,

which, for short, we have been writing as

ds2 = gµν dxµ dxν .

We will keep using the latter notation because it is shorter, even though the first expression is the most rigorous one.
The metric components transform exactly as we had seen under coordinate transformations, and as can be simply
derived from the chain rule.
From now on, we will specialize to 4-dimensional spaces, and suppose that the metric has signature (−1, 1, 1, 1),
i.e. that at every point p, one can find a basis {e(µ) } (which need not be a coordinate basis) such that g = ηµν e∗(µ) e∗(ν) .
Sylvester’s law of inertia tells us that if the signature is unique, i.e. one cannot find bases in which it would be,
say, (-1, -1, 1, 1) if it is (−1, 1, 1, 1) in one basis. The fact that the metric is a smooth tensor field and that it is
non-degenerate further enforces that it has the same signature throughout the manifold.
As we already discussed, at any given point p ∈ M, an appropriate choice of coordinates {xµ } can always be made
such that, for points q nearby p,

gµν (q) = ηµν + Oµν (xµp − xµq )(xνp − xνq ),

i.e. such that not only the metric components are equal to Minkowski at p, but also that they deviate from it at
most at quadratic order in the coordinate separations. This is what we called a locally inertial coordinate
system.