Pseudotensors

Math 1550 lecture notes, Prof. Anna Vainchtein

1 Proper and improper orthogonal transfor-
mations of bases
Let {e
1
, e
2
, e
3
} be an orthonormal basis in a Cartesian coordinate system
and suppose we switch to another rectangular coordinate system with the
orthonormal basis vectors {e
1
, e
2
, e
3
}. Recall that the two bases are related
via an orthogonal matrix Q with components Q
ij
= e
i
e
j
:
e
i
= Q
ij
e
j
, e
i
= Q
ji
e
i
. (1)
Let
= detQ (2)
and recall that = 1 because Q is orthogonal. If = 1, we say that the
transformation is proper orthogonal ; if = 1, it is an improper orthogonal
transformation. Note that the handedness of the basis remains the same
under a proper orthogonal transformation and changes under an improper
one. Indeed,

V = (e
1
e
2
)e
3
= (Q
1m
e
m
Q
2n
e
n
) Q
3l
e
l
= Q
1m
Q
2n
Q
3l
(e
m
e
n
) e
l
, (3)
where the cross product is taken with respect to some underlying right-
handed system, e.g. standard basis. Note that this is a triple sum over m,
n and l. Now, observe that the terms in the above some where (m, n, l) is a
cyclic (even) permutation of (1, 2, 3) all multiply V = (e
1
e
2
) e
3
because
the scalar triple product is invariant under such permutations:
(e
1
e
2
) e
3
= (e
2
e
3
) e
1
= (e
3
e
1
) e
2
.
Meanwhile, terms where (m, n, l) is a non-cyclic (odd) permutations of (1, 2, 3)
multiply V , e.g. for (m, n, l) = (2, 1, 3) we have (e
2
e
1
) e
3
= (e
1

e
2
) e
3
= V . All other terms in the sum in (3) (where two or more of the
three indices (m, n, l) are the same) are zero (why?). So (3) yields

V =(Q
11
Q
22
Q
33
+ Q
13
Q
21
Q
32
+ Q
12
Q
23
Q
31
Q
12
Q
21
Q
33
Q
11
Q
23
Q
32
Q
13
Q
22
Q
31
)V,
1
or, recalling (2) and the denition of the determinant,

V = V. (4)
Hence

V and V have the same sign (bases have the same handedness) if and
only if = 1 (a proper orthogonal transformation). Otherwise, if = 1,
a right-handed basis transforms into a left-handed one and vice versa.
Example 1. A counterclockwise rotation by the angle about x
3
-axis is a
proper orthogonal transformation. Indeed, in this case
[Q
ij
] =

cos sin 0
sin cos 0
0 0 1

.
which yields = cos
2
+ sin
2
= 1. The handedness is preserved under
such transformation (or any other pure rotation or combination thereof).
Example 2. Consider now a reection in the (x
2
, x
3
) plane: x
1
= x
1
,
x
2
= x
2
, x
3
= x
3
. We have
[Q
ij
] =

1 0 0
0 1 0
0 0 1

, = 1,
so the transformation is improper orthogonal. It is easy to see that such
transformation reverses the handedness of the basis, as does any orthogonal
transformation that involves an odd number of reections.
2 Pseudotensors
Denition. A pseudotensor A of order n in R
3
is a quantity with 3
n
com-
ponents which under orthogonal transformations of coordinate system trans-
form as follows:

A
i
1
i
2
...in
= Q
i
1
j
1
Q
i
2
j
2
. . . Q
injn
A
j
1
j
2
...jn
.
In other words, pseudotensors transform the same way as ordinary tensors
under proper orthogonal transformations but dierently under the improper
ones.
2
Example. Consider two vectors, a and b, and let
c = a b
be dened as their cross product computed in the current basis. Then we
have
c
i
= a
j
b
k
a
k
b
j
, c
i
= a
j

b
k
a
k

b
j
, (5)
where for any i, the indices j and k are selected so that (i, j, k) is the cyclic
permutation of (1, 2, 3).
Question: does c so dened transform as a Cartesian tensor, i.e. do we
have c
i
= Q
ik
c
k
under the change of basis (1)? Lets check:
c
i
= a
j

b
k
a
k

b
j
= Q
jm
a
m
Q
kn
b
n
Q
kn
a
n
Q
jm
b
m
= Q
jm
Q
kn
(a
m
b
n
a
n
b
m
).
Interchanging the summation indices m and n in the second term (Q
jm
Q
kn
a
n
b
m
=
Q
jn
Q
km
a
m
b
n
), we get
c
i
= (Q
jm
Q
kn
Q
jn
Q
km
)a
m
b
n
. (6)
Now, observe that the terms with m = n in the above sum are zero, so
only terms with m = n remain. Now, if m = 1 and n = 2, say, we get
(Q
j1
Q
k2
Q
j2
Q
k1
)a
1
b
2
, while m = 2 and n = 1 yields (Q
j2
Q
k1
Q
j1
Q
k2
)a
2
b
1
,
so that the two terms combined give us
(Q
j1
Q
k2
Q
j2
Q
k1
)(a
1
b
2
a
2
b
1
).
Similarly, the terms with (m, n) = (2, 3) and (m, n) = (3, 2) yield
(Q
j2
Q
k3
Q
j3
Q
k2
)(a
2
b
3
a
3
b
2
)
and the terms with (m, n) = (3, 1) and (m, n) = (1, 3) yield
(Q
j3
Q
k1
Q
j1
Q
k3
)(a
3
b
1
a
1
b
3
).
Combining these, we can see that (6) becomes
c
i
= (Q
j1
Q
k2
Q
j2
Q
k1
)(a
1
b
2
a
2
b
1
)
+ (Q
j2
Q
k3
Q
j3
Q
k2
)(a
2
b
3
a
3
b
2
)
+ (Q
j3
Q
k1
Q
j1
Q
k3
)(a
3
b
1
a
1
b
3
),
(7)
3
where we recall that (i, j, k) is the cyclic permutation of (1, 2, 3).
We now claim that
Q
jm
Q
kn
Q
jn
Q
km
= Q
il
, (8)
where (i, j, k) and (m, n, l) are both cyclic permutations of (1, 2, 3). To see
this, note that by (1),
e
j
e
k
= Q
jm
Q
kn
e
m
e
n
,
where both cross products are taken with respect to the same underlying
right-handed basis, e.g. standard basis. Then
(e
j
e
k
) e
l
= Q
jm
Q
kn
(e
m
e
n
) e
l
.
As in the earlier calculation, we note that (e
m
e
n
) e
l
) = (e
1
e
2
) e
3
= V
if (m, n, l) is a cyclic permutation of (1, 2, 3), (e
m
e
n
) e
l
= V if it is
non-cyclic and zero otherwise. This yields
(e
j
e
k
) e
l
= (Q
jm
Q
kn
Q
jn
Q
km
)V, (9)
where (m, n, l) is a cyclic permutation of (1, 2, 3). But (i, j, k) is also a cyclic
permutation of (1, 2, 3), so we have
e
j
e
k
=

V e
i
,
and hence
(e
j
e
k
) e
l
=

V (e
i
e
l
) =

V Q
il
. (10)
Combining (9) and (10), we get

V Q
il
= V (Q
jm
Q
kn
Q
jn
Q
km
),
which by (4) yields (8).
Now observe that (8) with (m, n) = (1, 2) (and thus l = 3) yields
Q
j1
Q
k2
Q
j2
Q
k1
= Q
i3
. Similarly, with (m, n) = (2, 3) we get Q
j2
Q
k3

Q
j3
Q
k2
= Q
i1
and with (m, n) = (3, 1) we have Q
j3
Q
k1
Q
j1
Q
k3
= Q
i2
.
Substituting these in (7), we get
c
i
=

Q
i3
(a
1
b
2
a
2
b
1
)+Q
i1
(a
2
b
3
a
3
b
3
)+Q
i2
(a
3
b
1
a
1
b
3
)

= Q
il
(a
m
b
n
a
n
b
m
),
4
where again (m, n, l) is a cyclic permutation of (1, 2, 3). But this implies that
c
i
= Q
il
c
l
,
so that c dened as the cross product in the current basis is indeed a pseu-
dotensor of rst order (a pseudovector, or axial vector) rather than a regular
vector. It transforms as a regular vector under proper orthogonal transfor-
mations ( = 1), c
i
= Q
il
c
l
, but not under improper ones ( = 1) where
we have c
i
= Q
il
c
l
.
Note that if we just compute c = a b in some basis and then consider
the transformation of the resulting quantity under a change to another basis
(without recomputing the cross product with respect to the new basis), then
of course the quantity will transform as a regular vector, c
i
= Q
il
c
l
. But
the point is that the new components will not be components of the cross
product ab as computed in the new basis unless Q is proper orthogonal. So
it is really the cross product operation that gives us a pseudovector because
we use the convention (right hand rule in a right-handed basis and left hand
rule in a left-handed one). The convention allows us to compute the cross
product the same way,
a b =

e
1
e
2
e
3
a
1
a
2
a
3
b
1
b
2
b
3

,
in any orthonormal basis {e
1
, e
2
, e
3
}, regardless of its handedness, but the
price we pay is that it gives us a pseudovector if we compute the cross product
in the current basis.
Remark. In general bases, the cross product is given by
a b =

e
1
e
2
e
3
a
1
a
2
a
3
b
1
b
2
b
3

G =

e
1
e
2
e
3
a
1
a
2
a
3
b
1
b
2
b
3

G
,
where G = V
2
= det[g
ik
] is the determinant of the metric tensor. In the
orthonormal case G = 1 and the covariant and contravariant components
coincide, as do regular and reciprocal basis vectors.
5
3 Properties of pseudotensors
1. The sum of two pseudotensors of order n is a pseudotensor of order n.
Indeed, if C
i
1
...in
= A
i
1
...in
+ B
i
1
...in
and

A
i
1
...in
= Q
i
1
j
1
. . . Q
injn
A
j
1
...jn
,

B
i
1
...in
= Q
i
1
j
1
. . . Q
injn
B
j
1
...jn
,
then

C
i
1
...in
= Q
i
1
j
1
. . . Q
injn
C
j
1
...jn
, so C = A+B is a pseudotensor.
2. The product of a pseudotensor of order m and a regular tensor of order
n is a pseudotensor of order m + n.
Let A be a pseudotensor of order m and B be a regular tensor of or-
der n. Then we have

A
i
1
...im
= Q
i
1
j
1
. . . Q
imjm
A
j
1
...jm
and

B
k
1
...kn
=
Q
k
1
l
1
. . . Q
knln
B
l
1
...ln
. Thus for C
i
1
...imk
1
...kn
= A
i
1
...im
B
k
1
...kn
we have

C
i
1
...imk
1
...kn
= Q
i
1
j
1
. . . Q
imjm
Q
k
1
l
1
. . . Q
knln
C
j
1
...jml
1
...ln
, so C is a pseu-
dotensor of order m + n.
3. The product of two pseudotensors of orders m and n is an ordinary
(Cartesian) tensor of order m + n.
Proved the same way as above, except now we get for both A and
B, so we obtain
2
= 1 in the transformation law for C.
4. Contraction of a pseudotensor of order n 2 gives a pseudotensor of
order n 2.
Indeed, if

A
i
1
i
2
...in
= Q
i
1
j
1
Q
i
2
j
2
. . . Q
injn
A
j
1
j
2
...jn
, then its contraction
in the kth and mth index (replace i
m
by i
k
and sum over these) satises

A
i
1
...i
k
...i
k
...in
= Q
i
1
j
1
. . . Q
i
k
j
k
. . . Q
i
k
jm
. . . Q
injn
A
j
1
...j
k
...jm...jn
.
But since Q
i
k
j
k
Q
i
k
jm
=
j
k
jm
this yields

A
i
1
...i
k
...i
k
...in
=
Q
i
1
j
1
. . . Q
i
k1
j
k1
Q
i
k+1
j
k+1
. . . Q
i
m1
j
m1
Q
i
m+1
j
m+1
. . . Q
injn
A
j
1
...j
k
...j
k
...jn
,
so A
i
1
...i
k
...i
k
...in
are components of a pseudotensor of order n 2.
6
4 Levi-Civita symbol
Consider

ijk
= (e
i
e
j
) e
k
, (11)
where the cross product is taken with respect to the orthonormal
basis {e
1
, e
2
, e
3
}. Then we have

ijk
=

+1, if (i, j, k) is a cyclic (even) permutation of (1, 2, 3)

1, if (i, j, k) is a non-cyclic (odd) permutation of (1, 2, 3)
0, otherwise, i.e. whwen any two indices coincide
(12)
Note that there are only six nonzero components:

123
=
231
=
312
= 1,
132
=
321
=
213
= 1
and observe that for any matrix A, detA =
ijk
A
1i
A
2j
A
3k
(show this as an
exercise). Observe also that

mnk
=
nkm
=
kmn
, (13)
i.e. the components are invariant under the cyclic permutation of the indices.
We claim that
ijk
is a third order pseudotensor. Indeed, in a new basis
{e
1
, e
2
, e
3
} we have

ijk
= (e
i

e
j
) e
k
,
where the bar above the cross product sign indicates that we are taking the
cross product with respect to the new basis. Recalling (1), we have

ijk
= (Q
im
e
m

Q
jn
e
n
) Q
kl
e
l
= Q
im
Q
jn
Q
kl
(e
m

e
n
) e
l
,
where again we are taking the cross product in the new coordinate system.
But we know that the cross products taken in the new and old systems coin-
cide, (e
m

e
n
) e
l
= (e
m
e
n
) e
l
if the two bases have the same handedness
(and thus Q is a proper orthogonal transformation, = 1) and have oppo-
site signs, (e
m

e
n
) e
l
= (e
m
e
n
) e
l
if they have dierent handedness
and thus the transformation is improper ( = 1). This means that
(e
m

e
n
) e
l
= (e
m
e
n
) e
l
=
mnl
,
and thus we have

ijk
= Q
im
Q
jn
Q
kl

mnl
,
7
proving that
ijk
is a third order pseudotensor.
Observe that if c = a b, we can now write
c
i
=
ijk
a
j
b
k
(14)
in any Cartesian coordinate system. Indeed,
c
1
=
1jk
a
j
b
k
=
123
a
2
b
3
+
132
a
3
b
2
= a
2
b
3
a
3
b
2
,
where we have used the fact that by (12) the only nonzero components
1jk
are
123
= 1 and
132
= 1. Comparing this to the rst equality in (5)
with (i, j, k) = (1, 2, 3), we see that this is indeed the rst component of the
cross product. Similarly, one can show that (14) yields c
2
= a
3
b
1
a
1
b
3
and
c
3
= a
1
b
2
a
2
b
1
, which are the second and third components of the cross
product according to the rst equality in (5) taken with (i, j, k) = (2, 3, 1)
and (i, j, k) = (3, 1, 2), respectively.
Recalling the properties of pseudotensors, one can see that if is a scalar,
then
ijk
are components of a third order pseudotensor, and if T
ij
are com-
ponents of a second order tensor, then
ijk
T
jk
are components of a rst order
pseudotensor, or pseudovector (obtained by the double contraction of a fth
order pseudotensor). For example, T
jk
= a
j
b
k
are components of a regular
Cartensian tensor, ans so c
i
=
ijk
T
jk
=
ijk
a
j
b
k
are components of a pseu-
dovector. If T
ijk
are components of a third order tensor, then
ijk
T
ijk
is a
pseudoscalar (meaning that it changes sign as we change from a right handed
system to a left handed or vice versa). For example, T
ijk
= a
i
b
j
c
k
, we get
the pseudoscalar
d =
ijk
T
ijk
=
ijk
a
i
b
j
c
k
= a (b c).
Meanwhile,
ijk

mnl
is a sixth order regular Cartesian tensor, which sat-
ises the following identity:

ijk

mnl
=

im

in

il

jm

jn

jl

km

kn

kl

.
In particular, expanding this determinant about the third row and setting
l = k (which means summing over k), we obtain the fourth order tensor

ijk

mnk
=
im

jn

jm

in
. (15)
8
Indeed, the expansion gives us

ijk

mnl
=
kl

im

in

jm

jn

kn

im

il

jm

jl

+
km

in

il

jn

jl

.
Setting l = k, we get

ijk

mnk
=
kk

im

in

jm

jn

kn

im

ik

jm

jk

+
km

in

ik

jn

jk

.
But
kk
=
11
+
22
+
33
= 1 + 1 + 1 = 3, and due to the Kronecker Deltas
in front of the determinants in the second and third terms, we can simplify
this further as

ijk

mnk
= 3

im

in

jm

jn

im

in

jm

jn

in

im

jn

jm

Observing that interchaging the columns in the last determinant changes its
sign to the opposite, we get

ijk

mnk
= 2

im

in

jm

jn

im

in

jm

jn

im

in

jm

jn

,
which yields (15).
Contracting (15) further, we obtain the second order tensor

ink

mnk
= 2
im
.
Indeed, setting j = n in (15) (and thus summing over n), we get
ink

mnk
=

im

nn

nm

in
. But
nn
= 3 (see above), so we get
ink

mnk
= 3
im

im
=
2
im
.
In particular, we can use (15) to prove
a (b c) = b(a c) c(a b).
Note that a (b c) is a regular Cartesian vector (not a pseudovector)
because we apply the cross product twice. To prove the identity, let d =
a (b c). Then by (14)
d
i
=
ijk
a
j
(b c)
k
=
ijk
a
j

kmn
b
m
c
n
9
But by (13)
kmn
=
mnk
, so that
d
i
=
ijk

mnk
a
j
b
m
c
n
= (
im

jn

jm

in
)a
j
b
m
c
n
=
im

jn
a
j
b
m
c
n

jm

in
a
j
b
m
c
n
,
where we used (15) to get the rst equality. Hence
d
i
= a
n
b
i
c
n
a
m
b
m
c
i
= b
i
(a
n
c
n
) c
i
(a
m
b
m
) = b
i
(a c) c
i
(a b).
Thus,
a (b c) = d = b(a c) c(a b),
proving the identity.
5 Antisymmetric second order tensors and
pseudovectors
Let A
ij
be components of an antisymmetric second order tensor, i.e. A
ji
=
A
ij
. The matrix representation of the tensor is
[A
ij
] =

0 A
12
A
13
A
12
0 A
23
A
13
A
23
0

,
so that there are only six nonzero components. Now observe that we can
write
A
ij
=
ijk
a
k
,
or
[A
ij
] =

0 a
3
a
2
a
3
0 a
1
a
2
a
1
0

.
Indeed, A
23
=
231
a
1
= a
1
, A
12
=
123
a
3
= a
3
, A
31
=
312
a
2
= a
2
, while
A
32
=
321
a
1
= a
1
, A
21
=
213
a
3
= a
3
and A
13
=
132
a
2
= a
2
. The
diagonal components are all zero because
iik
= 0 (no sum). It follows that
a
i
=
1
2

ijk
A
jk
. (16)
Indeed, A
1
=
1
2
(
123
A
23
+
132
A
32
) =
1
2
(A
23
A
32
), a
2
=
1
2
(
231
A
31
+
213
A
13
) =
1
2
(A
31
A
13
) and a
3
=
1
2
(
312
A
12
+
321
A
21
) =
1
2
(A
12
A
21
). We write
a =
1
2
A

= vecA.
10
Since a is obtained by a double contraction of a second order tensor and
a third order pseudotensor, the result must be a pseudovector (recall the
properties of pseudotensors). And indeed, under a change of variables,
a
i
=
1
2

ijk

A
jk
=
1
2
Q
im
Q
jn
Q
kl

mnl
Q
jp
Q
kq
A
pq
.
But Q
jn
Q
jp
=
np
and Q
kl
Q
kq
=
lq
, so we have
a
i
=
1
2
Q
im

mnl

np

lq
A
pq
=
1
2
Q
im

mnl
A
nl
= Q
im
a
m
,
since by (16)
mnl
A
nl
= 2a
m
. So a, whose components are related to the
components of the antisymmetric tensor A via (16), transforms as a pseu-
dovector, or an axial vector.
11