Caltech Ph106 Fall 2001

Math for physicists: dierential forms

Disclaimer: this is a rst draft, so a few signs might be o.
1 Basic properties
Dierential forms come up in various parts of theoretical physics, including advanced
classical mechanics, electromagnetism, thermodynamics, general relativity, and quan-
tum eld theory. So theyre well worth knowing about. This is supposed to be a
self-contained exposition for someone who has some knowledge of multivariable calcu-
lus.
Forms are like innitesimal objects, but this is (or can be made) a completely rigorous
subject. The most basic object is exterior derivative, d. For a function F(x
1
, . . . , x
n
),
dF =
F
x
i
dx
i
(1)
by denition. Summation over i from 1 to n is implicit. In some subjects (particularly
relativity), there is a preference to write x
i
and/or dx
i
, but I will eschew this here. Eq.
(1) looks like an innitesimal variation. The special thing about forms is that they
involve an anti-commuting wedge product, , dened so that
dx
i
dx
j
= dx
j
dx
i
. (2)
A general p-form (for p n) is
=
1
p!

i
1
...i
p
dx
i
1
dx
i
2
. . . dx
i
p
. (3)
Again there is an implicit sum on i
1
through i
p
, each from 1 to n. The coecient
functions
i
1
i
p
may be assumed to be antisymmetric in i
1
through i
p
. So for instance,
if p = 2,
ij
=
ji
. The 1/p! is there in (3) because the sum contains p! copies of
each term. Again for the case of 2-forms, say now in 2 dimensions, we would have
=
1
2

ij
dx
i
dx
j
=
12
dx
1
dx
2
. (4)
The middle expression, explicitly, is
1
2
(
12
dx
1
dx
2
+
21
dx
2
dx
1
), but the last ex-
pression is obviously equal to this.
1
The exterior derivative of a p-form can be dened as
d =

i
1
...i
p
x
j
dx
j
dx
i
1
. . . dx
i
p
. (5)
Its crucial to put the dx
j
consistently where I did. You can verify the following product
rule for d:
d(
(p)

(q)
) = d
(p)

(q)
+ (1)
p

(p)
d
(q)
, (6)
where by
(p)
I mean a p-form. And you can show that d
2
= 0, in the sense that if
= d, then d = 0. An important and non-trivial theorem says that if d = 0, then
= d for some .
1
Dierential forms were invented for integration. Heres how it works. Suppose you
have some integration region M, which could be any smooth p-surface in R
n
, and
suppose you have a set of smooth 1-1 functions, (
1
, . . . ,
p
) (x
1
, . . . , x
n
), which map
a ducial p-dimensional integration region I into M. For instance, I could be [0, 1]
p
in
R
p
. Then, for any p-form , we dene
_
M
=
_
I
1
p!

i
1
...i
p
(x
i
1
, . . . , x
i
p
)
(
1
, . . . ,
p
)
d
p
. (7)
Here (x
i
1
, . . . , x
i
p
)/(
1
, . . . ,
p
) denotes a Jacobian: that is, the determinant of the
matrix of partial derivatives x
i
j
/
k
, where j and k run from 1 to p. The denition
(7) would follow, formally, if we wrote dx
i
=
x
i

j
d
j
. (Try it!)
The most important theorem in this subject is Stokes Theorem, which says that if
is a (p 1)-form and M is a p-suface, then
_
M
d =
_
M
, (8)
where M is the boundary of M. Actually, theres something a little tricky about
(8): both M and M have orientations which must be specied, and picking them in
some ways gives a minus sign in front of the right hand side of (8). To understand this
orientation business, note that if you replaced
1
by
1
, or swapped
1
and
2
, in (7),
it would change the sign of the Jacobian, and hence the sign of the integral. Its not
a priori obvious how to relate the orientations of M and M in general, but this is
getting beyond the scope of what we really need to know.
2 Div, grad, and curl
Lets now consider some examples. Start with a function f on R
3
. Its exterior deriva-
tive, v = df, can be written as v = v
i
dx
i
, where v = (v
1
, v
2
, v
3
) = f is the gradient.
1
This is true up to cohomology. So its precisely true if were working on R
n
, or on any other space
thats contractable to a point.
2
So vectors can be turned into one-forms. Now take v to be any one-form, and you get
dv =
v
i
x
j
dx
j
dx
i
=
_
v
3
x
2

v
2
x
3
_
dx
2
dx
3
+
_
v
1
x
3

v
3
x
1
_
dx
3
dx
1
+
_
v
2
x
1

v
1
x
2
_
dx
1
dx
2
= w
1
dx
2
dx
3
+ w
2
dx
3
dx
1
+ w
3
dx
1
dx
2
w
(9)
where w = (w
1
, w
2
, w
3
) = v. Observe that if v = f, then w = v = 0. And
recall that if a vector eld has no curl, then its the gradient of some function: thats
why we can write

E = in electrostatics. These statements are special cases of
d
2
= 0 and the deeper theorem that if d = 0 then = d for some . An important
point is that a vector w is naturally associated to a 2-form, in the way we wrote in
(9). This is a special property of three dimensions: in n dimensions, a vector would be
naturally associated to either a 1-form or a (n 1)-form.
Continuing on, lets consider any w and the associated 2-form w. Dierentiating
once more gives
dw =
_
w
1
x
1
+
w
2
x
2
+
w
3
x
3
_
dx
1
dx
2
dx
3
= ( w)dx
1
dx
2
dx
3
. (10)
Obviously, if w = dv, then dw = 0. This is precisely the statement that the curl of
a vector eld has no divergence. And recall also that if a vector eld is divergence-
free, then its the curl of something: thats why we can always write

B =

A in
electromagnetism. Again, these statements are special cases of d
2
= 0 and d = 0
= d.
We can pursue our vector calculus examples further to get a couple of special cases
of Stokes Theorem. For instance, if we have a vector eld w on R
3
, then we know how
to integrate its divergence over a simply connected region M (simply connected just
means, no holes):
_
M
wd
3
x =
_
M
n w

g d
2
, (11)
where n is the outward-pointing unit vector normal to the boundary M. This is
sometimes called Gausss Theorem. Saying that n points outward amounts to specify-
ing the orientation of M. In (11),

g d
2
is by denition the area element on M.
So if M were a ball, and we used angular coordinates, then

g d
2
= sin dd. Ive
slipped in the

g because something has to tell you that theres a non-trivial measure

on the surface. A more concise way to phrase (11), which avoids needing to discuss
complicated things like measures, is just
_
M
dw =
_
M
w. (12)
3
The left side hardly needs explaining: we just plug in (10) for dw. On the right hand
side, to integrate over M, we require a map (
1
,
2
) (x
1
, x
2
, x
3
), and then we use
the denition (7). For the case where M is a unit ball, we could take (
1
,
2
) = (, ),
and x = (r sin cos , r sin sin , r cos ). With some work, you can show that the
right hand side of (12) is the same as our previous expression in (11).
2
Another example, in two dimensions, is Greens theorem: if we have a vector eld
v = (v
x
, v
y
), and a simply connected region M, then
_
M
_
v
y
x

v
x
y
_
=
_
M
v d

=
_
M
(v
x
dx + v
y
dy) , (13)
where we traverse M counterclockwise (this is the specication of orientation). You
should be able to convince yourself in short order that this can be concisely rephrased
as
_
M
dv =
_
M
v, where the right hand side is dened via a map that circles around
Mcounterclockwise as ranges from 0 to 1. We could give a similar treatment of the
classic statement of Stokes Theorem (really another small special case of (8)) relating
with a line integral over the boundary of a curved surface in three dimensions... but
hopefully the basic ideas are by now obvious.
3 Partial derivatives and forms
An important aspect of dierential forms is their relation to partial derivatives. Sup-
pose we have dierent ways of coordinatizing the same space: say (x
1
, x
2
) and (y
1
, y
2
).
If you like, the x coordinates could be Cartesian coordinates on R
2
while the y coordi-
nates are polar coordinates. But this is only an example. Another example is for the x
coordinates to be q and p for a classical system with one degree of freedom, while the
y coordinates are some other Q and P related by a canonical transformation. What
we want in general is for there to be a smooth 1-1 map from the x coordinates to the
y coordinates. In generic situations (that is, barring some accident like x
1
= y
2
),
we can choose any two of the four variables (x
1
, x
2
, y
1
, y
2
) to parametrize R
2
. So if
we start with a function f(x
1
, x
2
), we could write it instead as a function of y
1
and
y
2
, or x
1
and y
2
, or whatever we please. With this in mind, we should write partial
derivatives of f like this:
df =
_
f
x
1
_
x
2
dx
1
+
_
f
x
2
_
x
1
dx
2
, (14)
2
The best way to go about this is to say (
0
,
1
,
2
) = (r, , ), and then show
(x
1
, x
2
, x
3
)/(
0
,
1
,
2
) = r
2
sin . Now the result should be obvious at least for w = f(r) sin dd,
which corresponds to w pointing in the radial direction.
4
where the ()
u
notation means that u is held xed. But it seems we could as well have
written
df =
_
f
y
1
_
y
2
dy
1
+
_
f
y
2
_
y
1
dy
2
. (15)
Equating these two reproduces the well-known multi-variable chain rule:
_
f/y
1
f/y
2
_
=
_
x
1
/y
1
x
2
/y
1
x
1
/y
2
x
2
/y
2
__
f/x
1
f/x
2
_
, (16)
where now to avoid clutter I just leave it implicit that the partials on the left hand
side hold either y
1
or y
2
xed, whereas the ones on the right hand side hold either x
1
or x
2
xed. A less obvious application of forms is to prove the relation
_
f
x
1
_
y
1
=
_
f
x
1
_
x
2

_
f
x
2
_
x
1
(y
1
/x
1
)
x
2
(y
1
/x
2
)
x
1
(17)
To prove this, we can set
dy
1
=
_
y
1
x
1
_
x
2
dx
1
+
_
y
1
x
2
_
x
1
dx
2
= 0 , (18)
solve for dx
2
, and plug back into (14) to get
df =
__
f
x
1
_
x
2

_
f
x
2
_
x
1
(y
1
/x
1
)
x
2
(y
1
/x
2
)
x
1
_
dx
1
with dy
1
= 0. (19)
Now dividing by dx
1
gives (17). That wasnt quite rigorous, since I switched from
regarding df and dx
1
as one-forms to regarding them as innitesimals. A more rigorous
method would be to equate (14) to
df =
_
f
x
1
_
y
1
dx
1
+
_
f
y
1
_
x
1
dy
1
(20)
and then set dy
1
= 0, using (18). A special case of (17) is to set f = x
2
; then we obtain
_
x
2
x
1
_
y
1
=
(y
1
/x
1
)
x
2
(y
1
/x
2
)
x
1
=
_
x
2
y
1
_
x
1
_
y
1
x
1
_
x
2
. (21)
Plugging back into (17), we obtain the simpler identity
_
f
x
1
_
y
1
=
_
f
x
1
_
x
2
+
_
f
x
2
_
x
1
_
x
2
x
1
_
y
1
(22)
5
4 Applications to physics
The relations (21) and (22) are quite useful in thermodynamics. For instance, we could
them to relate specic heats at xed volume versus xed pressure:
C
P
= T
_
S
T
_
P
= T
_
S
T
_
V
+ T
_
S
V
_
T
_
V
T
_
P
= C
V
+ T
_
P
T
_
V
_
V
T
_
P
= C
V
T
_
P
V
_
T
_
V
T
_
2
P
.
(23)
In the third equality we used a Maxwell relation, (S/V )
T
= (P/T)
V
, which follows
from the fact that both sides can be expressed as
2
F/V T, where F = F(V, T) is
the Helmholtz free energy. Dening the isothermal compressibility, K
T
=
1
V
_
V
P
_
T
,
and the expansion coecient
0
=
1
V
_
V
T
_
P
, we nd (c.f. Mandl, p. 123)
C
P
C
V
= TV
2
0
/K
T
, (24)
which has been called the most beautiful relation in thermodynamics (I think because
its hard to remember how to derive it). The crucial step in the derivation is the
second equality in (23), which came from parametrizing the possible equilibrium states
of the system either by (V, T) (the canonical ensemble) or by (P, T) (which we might
call the Gibbs ensemble since its associated with the Gibbs free energybut dont
get confused between this and the grand canonical ensemble, where particle number
uctuates). Specializing to an ideal gas, we have PV = NkT, and (24) becomes
C
P
= C
V
+ Nk , (25)
which you probably learned in high school. Incidentally, the equipartition theorem
applies most directly to C
V
.
After all this, hopefully the exposition of canonical transformations given in class
makes better sense. Its worth making a few additional comments. The Poisson bracket
structure is equivalent in information content to specifying a so-called symplectic
form on phase space:
=
N

k=1
dq
k
dp
k
. (26)
Canonical transformations by denition preserve the symplectic form: that is, if the
canonically transformed coordinates are Q
k
and P
k
, then
=
N

k=1
dQ
k
dP
k
(27)
6
is the same form as we had in (26). Once we realize that Hamiltonian ow generates
canonical transformations, we can give an intuitive proof of Liouvilles Theorem in just
a few lines. Observe that
vol =
1
N!
N times
..
. . . = dq
1
. . . dq
N
dp
1
. . . dp
N
. (28)
Ignoring the sign (which only depends on N), we see that vol is precisely the volume
element on phase space. It is preserved by Hamiltonian ow because is. This is the
essential content of Liouvilles Theorem: phase space volume is invariant.
Dierential forms give a very natural formulation of electromagnetism. In R
3,1
(Minkowski space), suppose we construct
F = dt (E
x
dx + E
y
dy + E
z
dz) B
x
dy dz B
y
dz dx B
z
dx dy

F = dt (B
x
dx + B
y
dy + B
z
dz) + E
x
dy dz + E
y
dz dx + E
z
dx dy
J = dx dy dz dt (J
x
dy dz + J
y
dz dx + J
z
dx dy)
A = dt A
x
dx A
y
dy A
z
dz .
(29)
Here

E and

B are the electric and magnetic elds; and

J are charge and current
density; and and

A are the electrostatic potential and vector potential. F is called
the eld strength tensor;

F is its Hodge dual; and A is called the gauge potential. Its
straightforward to verify that the basic equations of electromagnetism (in conventions
where c = 1c.f. Jackson p. 548),

E = 4

E +


B
t
= 0

B = 0

B


E
t
= 4

E =


A
t

B =

A

t
+

J = 0 ,
(30)
can be rephrased as
dF = 0 d

F = 4J
F = dA dJ = 0 .
(31)
Note that the second line is entirely a consequence of the rst, given d
2
= 0 and
d = 0 = d. Incidentally, its also obvious now that sending A A + d,
for any function (t, x), doesnt change F at all. This is a gauge transformation.
7