Lie Derivatives on Manifolds

William C. Schulz
Department of Mathematics and Statistics,
Northern Arizona University, Flagstaff, AZ 86011

1. INTRODUCTION
This module gives a brief introduction to Lie derivatives and how they act
on various geometric objects. The principal difficulty in taking derivatives of
sections of vector bundles is that the there is no cannonical way of comparing
values of sections over different points. In general this problem is handled by
installing a connection on the manifold, but if one has a tangent vector field
then its flow provides another method of identifying the points in nearby fibres,
and thus provides a method (which of course depends on the vector field) of
taking derivatives of sections in any vector bundle. In this module we develop
this theory. 1

2. TANGENT VECTOR FIELDS

A tangent vector field is simply a section of the tangent bundle. For our purposes
here we regard the section as defined on the entire manifold M . We can always
arrange this by extending a section defined on an open set U of M by 0, after
some appropriate smoothing. However, since we are going to be concerned with
the flow generated by the tangent vector field we do need it to be defined on all
of M .
The objects in the T (M ) are defined to be first order linear operators acting
on the sheaf of C ∞ functions on the manifold. Xp (f ) is a real number (or a
complex number for complex manifolds). At each point p ∈ M we have
Xp (f + g) = Xp (f ) + Xp (g)
Xp (f g) = Xp (f )g(p) + f (p)Xp (g)

Xp (f ) should be thought of as the directional derivative of F in the direction X

at P . Thus X inputs a function (at p) and outputs a real (or complex) number.
We regard f as being defined on some neighborhood of p (which depends on f ).
The rules above describe how Xp acts on sums and products.
It is possible to show, although we will not do so here, that Tp (M ) is an
n-dimensional vector space (n = dim M , the dimension of the Manifold M )
∂ ∂
with a basis in local coordinates { ∂u 1 , . . . , ∂un }. Hence

∂ 
Xp = X i (p) i 
∂u p
17 Oct 2011

1
and
∂f 
Xp (f ) = X i (p)
∂ui p


If we change coordinates from ui to ũj the the local expression of X changes

like a contravariant tensor:
∂f ∂f ∂uj ∂f
X̃ i = X̃ i = Xj j
∂ ũi ∂uj ∂ ũi ∂u
so
∂uj ∂ ũi
X j = X̃ i and X̃ i = X j
∂ ũi ∂uj

3. THE LIE BRACKET

If we attempt to compose X and Y the results are not encouraging; locally
∂f
Y (f ) = Yi
∂ui  
∂ ∂f
X(Y (f )) = Xj j Y i i
∂u ∂u
i
∂Y ∂f ∂2f
= Xj j i
+ XjY i j i
∂u ∂u ∂u ∂u
which shows that XY is not a tangent vector since it contains the second deriva-
tive of f and is thus not a first order operator. However, we take heart from
the observation that the objectionable term is symmetric in i and j. Thus if we
form
∂X i ∂f ∂2f
Y (X(f )) = Y j j i
+ Y jXi j i
∂u ∂u ∂u ∂u
and subtract, the objectionable terms will drop out and we have

∂Y i ∂X i ∂f
 
X(Y (f )) − Y (X(f )) = X j j − Y j j
∂u ∂u ∂ui
which is a first order operator and hence a tangent vector. We now introduce
new notation
∂Y i ∂X i
 
∂
[X, Y ] = X(Y (·)) − Y (X(·)) = X j j − Y j j
∂u ∂u ∂ui
∂Y i ∂X i
 
[X, Y ]i = Xj j − Y j j
∂u ∂u

For practise the user may wish to verify that [X, Y ]i transforms properly under
coordinate change.
∂ ∂
X and Y are said to commute if [X, Y ] = 0. For example ∂u i and ∂uj

commute.
It is obvious that [X, Y ] is linear in each variable.

2
 Next we form

[X, [Y, Z]](f ) = X((Y Z − ZY )(f )) − (Y Z − ZY )(X(f ))

= X(Y (Z(f ))) − X(Z(Y (f ))) − Y (Z(X(f ))) + Z(Y (X(f )))

Let us abbreviate this by surpressing the f and the parentheses and then per-
mute cyclically.

[X, [Y, Z]] = XY Z − XZY − Y ZX + ZY X

[Y, [Z, X]] = Y ZX − Y XZ − ZXY + XZY
[Z, [X, Y ]] = ZXY − ZY X − XY Z + Y XZ

Now if we add the three equations the terms cancel in pairs and we have the
important Jacobi Identity

[X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0

Thus the vector fields in T (M ) form an (infinite dimensional) Lie Algebra,

which is a vector space with a skew symmetric multiplication [X, Y ] linear in
each variable satisfying the previous identity. The Jacobi identity substitutes
for associativity.

4. BACK AND FORTH WITH DIFFEOMOR-

PHISMS
The user may wonder why this area of mathematics tends to emphasize diffeo-
morphisms instead of differentiable maps. In fact, the level of generality implied
by φ : M → N is largely spurious; most of the time N = M and so diffeomor-
phisms are the natural objects to study. Nevertheless we follow convention in
this regard and formulate the results for φ : M → N
The user will recall that a mapping φ : M → N induces a mapping φ∗ = dφ
from Tp (M ) to Tq (N ) where q = φ(p). Since a vector in Tq (N ) may be identified
by its action on a function f : V → R, V a neighborhood of q in N , we can
define φ∗ by
(φ∗ X)q (f ) = Xp (f ◦ φ)
since f ◦ φ : M → R. For notational amusement we define

φ∗ f = f ◦ φ

and then we have

(φ∗ X)(f ) = X(φ∗ f )
and thus
φ∗ (X) = X ◦ φ∗

3
for X ∈ Tp (M ) and φ∗ : Tp (M ) → Tφ(p) (N ).
In local coordinates we describe φ as follows: p has coodinates u1 , . . . , un
and φ(p) has coodinates v 1 , . . . , v n and thus p → φ(p) is given by
v i (u1 , . . . , un ) i = 1, . . . , n
Then, if
∂ ∂
X = Xi Y = Yi j
∂ui ∂v
and Yφ(p) = φ∗ (Xp ) we have, (surpressing some p subscripts)

∂f ∂f ◦ φ ∂ui
Yφ(p) (f ) = Y j j
= Yj
∂v ∂ui ∂v j
and
∂f ◦ φ
Yφ(p) (f ) = Xp (f ◦ φ) = X i
∂ui
from which we see that
∂ui
Xi = Y j
∂v j
and symmetrically
∂v j
Y j = Xi
∂ui
which resembles the coordinate change rules. This is for the best of reasons;
because φ is a diffeomorphism a coordinate patch on N becomes, via φ, a
coordinate patch on M . Thus, locally, a diffeomorphism looks like a coordinate
change. This is not particularly helpful in keeping things straight in ones mind,
although occasionally technically useful.
Now if φ were just a differentiable mapping then it could not be used to
map vector fields on M to vector fields on N . The obvious way to do this is as
follows: given q ∈ N , select p ∈ M so that φ(p) = q and then let Yq = φ∗ Xp .
This won’t work for two reasons. First, there might be q ∈ N not in the range
of φ, and even if q is in the range of φ we might have φ(p1 ) = φ(p2 ) = q for
p1 6= p2 so that Yq would not be uniquely defined. However, neither of these is
a problem for diffeomorphisms so that
Def If X is a vector field on M and φ : M → N is a diffeomormorphism then
we can define a vector field Y = φ∗ (X) on N by
(φ∗ X)q = Xφ−1 q ◦ φ = φ∗ (Xφ−1 q )
If f : N → R then (recall φ∗ (f ) = f ◦ φ)
= Xφ−1 q ◦ φ (f ) = Xφ−1 q (φ∗ f )


(φ∗ X)q (f )
= Xφ−1 q (f ◦ φ)
Expressed slightly differently
if Y = φ∗ X
then Yq (f ) = Xφ−1 q (f ◦ φ)

4
 Notice that if φ1 : M1 → M2 and φ2 : M2 → M3 are diffeomorphisms then

φ2∗ ◦ φ1∗ = (φ2 ◦ φ1 )∗

This is just an abstract expression of the chain rule, but we can say it really
fancy: ∗ is a covariant functor from the category of Manifolds and Diffeomor-
phisms to the category of vector bundles and isomorphisms.
Now we want to show that the Lie Bracket is preserved under diffeomor-
phisms. First we note that if φ : M → N is a diffeomorphism then it can be
regarded locally as a coordinate change. Those persons who verified that the
Lie Bracket was invariant under coordinate change when I requested them to
do so need not read the following.
Let φ : M → N be a diffeomorphism and let u1 , . . . , un be coordinates
around p ∈ M and v 1 , . . . , v n be coordinates around q = φ(p) ∈ N . Then we
have, for f : N → R,

(φ∗ X)q f = Xφ−1 q (f ◦ φ)

∂
= Xui i (vj ) i (f ◦ φ)
∂u
∂f ∂v j ∂f
= Xui i (vj ) i i = X̃vjk j
∂v ∂u ∂v
where we set
∂v j
X̃vjk = Xui ℓ (vk )
∂ui
Now we can calculate the Lie Bracket. We surpress the subscripts on X, X̃, Y, Ỹ
because they are always the same. Then

∂ Ỹ i j ∂ X̃
i
[X̃, Ỹ ]i = X̃ j − Ỹ
∂v j ∂v j
j
∂v ∂  ∂v i  ∂uℓ k ∂v
j
∂  m ∂v i  ∂uℓ
= Xk k ℓ Y m m − Y X
∂u ∂u ∂u ∂v j ∂uk ∂uℓ ∂um ∂v j
m ℓ j i j 2 i ℓ
∂Y ∂u ∂v ∂v ∂v ∂ v ∂v
= Xk + X k m
Y
∂uℓ ∂v j ∂uk ∂um ∂uk ∂ui ∂um ∂uj
m
∂X ∂u ∂v ∂v ℓ j i
k m ∂v
j
∂ 2 v i ∂v ℓ
− Yk − Y X
∂uℓ ∂v j ∂uk ∂um ∂uk ∂ui ∂um ∂uj
m i m i
∂Y ∂v ∂X ∂v
= Xk δ ℓ
− Y k
δ ℓ
∂uℓ k ∂um ∂uℓ k ∂um
2 i
∂ v ∂ 2 vi
+ X k Y m δkℓ ℓ m − Y k X m δkℓ ℓ m
∂u ∂u ∂u ∂u
 ∂Y m m i 2 i 2 i
∂X ∂v k m ∂ v k m ∂ v
= Xℓ − Y ℓ
+ X Y − Y X
∂uℓ ∂uℓ ∂um ∂uk ∂um ∂uk ∂um
i
∂v
= [X, Y ]m m + 0
∂u
i
^
= [X, Y]

5
5. FLOWS
A vector field (section of T (M )) gives rise to a flow φt (p) = φ(p, t) : M × I → M
where the φt are diffeomorphisms and the interval I is (−ǫ, ǫ), where ǫ may, in
general, depend on the value of p. Let {u1 , . . . , un } be local coordinates and the
vector field be given locally by
∂
X = Xi
∂ui
The φt also have a local expression where φt (p) is given by {u1 (t), . . . , un (t)}
(where p corresponds to ui (0) = ui0 ). Then φ(t) is determined by differential
equations with initial conditions. We denote the ui (t) with initial conditions
ui (0) = ui0 ) by ui (uk0 , t). The differential equations are:

duk (ui0 , t)
= X k (uj (ui0 , t)) uk (ui0 , 0) = uk0
dt
The ui0 function as parameters in these equations. The flowline begins with the
point p that has these coordinates and flows out of p to points with the coor-
dinates uj (ui0 , t). The Picard-Lindelöf theorem guarantees that the solutions
will be as differentiable as the input data and depend as differentiably on the
parameters as the X i do. Solutions will exist for some interval (−ǫ, ǫ) where
ǫ > 0. However, in general the ǫ will depend on the intial point p. This won’t
do us any harm. Uniqueness of the solutions guarantees that

φt1 (φt2 (p)) = φt1 +t2 (p)

as long as t1 , t2 , t1 + t2 remain within (−ǫ, ǫ). If M is compact we can say much

more; φt is defined for all t ∈ R. However, for Lie purposes the local existence
is sufficient, so we will not pursue the matter here.

6. LIE DERIVATIVES
Here is the idea of the Lie Derivative. Given a tangent vector field X on M , the
derivative in the X direction of an object is the rate of change along the flowline
φt (p) if this makes sense. Unfortunately, it only makes sense for functions. To
find derivatives of objects that live in bundles we must change the game slightly,
since objects in neighboring fibres cannot be directly compared. Let E be a
bundle over M . To compare an object Wq in the fibre Eq over q to an object
Wp in the fibre Ep over p we need an isomorphism from Eq to Ep . This is not
generally available, but if q is on the flowline out of p determined by the vector
field X then we may be able to find such an isomorphism and this suffices to
define the Lie Derivative. Indeed if q = φt (p) then p = φ−1 t (q) = φ−t (q) and if
we can find (φ−t )∗ : π −1 [q] → π −1 [p] then we can compare (φ−t )∗ Wq with Wp .
In fact
(φ−t )∗ Wφt (p) ∈ π −1 [p]

6
is, for small t, a curve in π −1 [p] and hence can be differentiated at t = 0, yielding
an object in Tp (π −1 [p]) = Tp Wp . This is the Lie Derivative for the bundle E.
The importance of understanding this construction is easy to underesti-
mate. If we understand it we understand a) where the Lie Derivative lives (in
the tangent space to the fibre of the bundle). If the fibre is a vector space it is
possible to identify the tangent space with the fibre itself, and this is often done.
b) Knowing that we are dealing with a curve in the fibre Wp focuses our atten-
tion properly when we are doing the technical calculations. c) it prepares us for
the idea of connections in vector bundles and principal fibre bundles where the
ideas are somewhat similar.
Naturally for different bundles (φ−t )∗ will have a different form, but this is
just technical stuff; the basic idea is given above.
First we will deal with the Lie Derivative of a function. In this case (φ−t )∗ =
Identity and (φ−t )∗ f = f . Hence
d 
£X (f ) = (f ◦ φt )t=0
dt
d
(f (ui (t)))t=0

=
dt
∂f dui 
=
∂ui dt t=0

∂f i
= X
∂ui
= X(f )

Our next project is the Lie Derivative of a Tangent Vector Field Y ∈ T (M ).

d
X is still the vector field with flow φt (p) with dt φt |t=0 = X. Recall that
−1
φ−t = φt . The isomorphism is

(φ−t )∗ : Tφt (p) → Tp (M )

where p = φ0 (p) and q = φt (p). The rest is mere intricate calculation. Let p
have coordinates u1 (0), . . . , un (0) and q have coordinates u1 (t), . . . , un (t). For
ease of presentation we will use v i = ui (t). Then (φt )∗ : Tp (M ) → Tq (M ) is
given in coordinates, with Yq = (φt )∗ Zp , by

∂v i j
Yqi = Z
∂uj p
and then (φ−t )∗ = (φ−1
t )∗ is given by
 1   1 
 i −1 Yq Zp
∂v  ..   .. 
 .  =  .  ∈ Tp (M )
∂uj
Yqn Zpn
Notice that
∂v i ∂ui (t)
   
= = Id
 
∂uj ∂uj
 
t=0 t=0

7
We are going to need
−1  −2
∂v i ∂v i ∂v i
   
d d
= (−1)
 
dt ∂uj ∂uj dt ∂uj
 
t=0 t=0
i −2
   
∂v ∂ d i
= (−1) u (t) 

∂uj ∂uj dt t=0
 
∂
= (−1)(Id) Xi
∂uj
∂X i
 
= −
∂uj

Now we have, for Yq ∈ Tq (M ), or more precisely

Yφt (p) ∈ Tφt (p) (M )

Zp1 Yφ1t (p)
   
 i
−1
 ..  ∂v ..
 .  =
 
∂uj
 . 
Zpn Yφnt (p)

a curve in the fibre Tp (M ) whose derivative at t = 0 is the Lie Derivative

£X (Y ) ∈ Tp,0 (Tp (M )). We now compute
 1 
Zp 
d  . 
£X (Y ) =  .. 
dt 
Zpn t=0
 1  1
Yφt (p) Yφt (p)
  
 i −1 !  i −1
d ∂v .. ∂v d  ..
=  +
  
dt ∂uj
 . ∂u j dt
 . 
Yφnt (p) Yφnt (p)
t=0
 1    1
Y
 
Yp −1 φt (p)
∂X i  .   ∂v i  dv j 
  
∂  ..
= − j
.
. + j j .
∂u ∂u ∂v dt
     
n n
Yp Yφt (p)
t=0
 1   1 
Yp Yp
∂X i  . 
 
∂  .  j
= − .
. + Id .
. X
∂uj ∂uj
  
n
Yp Ypn

If the coordinates of £X (Y ) in the local basis are denoted by £X (Y )i this gives

∂ i ∂
£X (Y )i = X j Y − Y j j X i = [X, Y ]i
∂uj ∂u
and we have proved
£X (Y ) = [X, Y ]

8
We note that £X (Y ) has two terms because it is the derivative of (φ−t )∗ Yφt (p)
both factors of which depend on t, and the origin of the negative sign lies in the
inverse: φ−t = φ−1
t . We also note that the above derivation is highly dependent
on the continuity of the second derivatives.
Next we want to derive a formula for the Lie Derivative of the Lie Bracket
of two Vector Fields. This is immediate from the Jacobi Identity. Indeed recall

[X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0

from which we get

[X, [Y, Z]] = [Y, [X, Z]] − [Z, [X, Y ]]

£X ([Y, Z]) = £Y ([X, Z]) − £Z ([X, Y ])
£X ([Y, Z]) = £Y £X (Z) − £Z £X (Y )

In contrast to the preceding the calculation of the Lie Derivative of a 1-form

is somewhat easier as the formulation of (φ−t )∗ is a bit easier. It is simply the
familiar pullback of differential forms. If φt (p) = q then

φ∗t : Tq∗ (M ) → Tp∗ (M )

is the pullback and we set, for this bundle, (φ−t )∗ = φ∗t . (Remember, to get a
Lie Derivative on a bundle we must have an isomorphism (φ−t )∗ from π −1 [q] to
π −1 [p]. We hope to find such an isomorphism which makes sense for the bundle.
In the sequel we will also have to worry about whether the Lie Derivatives for
the various bundles fit together well. For T ∗ (M ) the above turns out to be
a good choice, as well as the only obvious choice.) Let’s now recall how φ∗
works. If φ : M → N and u1 , . . . , un are coordinates on M and v 1 , . . . , v n are
coordinates on N and φ is given locally by v i (u1 , . . . , un ) and a section ω is
given locally on N by ωi dv i then ω̃ = φ∗ (ω) is given locally on M by

∂v j i
ω̃i (uj ) dui = ωj (v k (uj )) du
∂ui
In our case φt is expressed by v i = v i (uj ), and we know that

dv k 
= Xk
dt t=0


since
dφt 
=X
dt t=0


Thus
d
£X (ω) = [(φ−t )∗ ω(φt (p))]t=0
dt
d
= [(φt )∗ ω(φt (p))]t=0
dt

9
 ∂v i j
 
d
= ωi j du
dt ∂u t=0
k i
d ∂v i j
 
∂ωi du ∂v j
= du + ω i du
∂uk dt ∂uj dt ∂uj
 t=0
∂ dv i j

∂ωi k i j
= X δj du + ωi j du
∂uk ∂u dt t=0
∂ωi k i ∂ i j
= k
X du + ωi j X du
∂u ∂u
∂ωi j ∂X j
= X + ωj dui
∂uj ∂ui
= X(ωi ) dui + ωi dX i

This last is an interesting formula and we will find use for it from time to time.
The following calculations are not strictly necessary; they can be derived
abstractly from the fact that (φt )∗ and d commute. However, the formulas may
come in handy and and there is (for me) a certain amusement in the formulas.
We first derive
£X (df ) = d£X (f )
by computing both sides.
∂f
df = j
duj
∂u
∂f ∂X j
 
∂ ∂f j
£X (df ) = X + dui
∂ui ∂uj ∂uj ∂ui
 2
∂f ∂X j

∂ f j
= X + dui
∂uj ∂ui ∂uj ∂ui
 
∂f j
d£X (f ) = d(Xf ) = d X
∂uj
 2
∂f ∂X j

∂ f j
= X + dui
∂ui ∂uj ∂uj ∂ui
Next I want
k
 
k k j ∂u
d£X (u ) = d(X(u )) = d X
∂uj
= d(X j δjk ) = dX k

and then

£X (duk ) = d£X (uk )

= dX k

Now we can verify that Leibniz’ rule works here; for ω ∈ Λ1

£X (ω) = £X (ωi dui )

10
 = £X (ωi ) dui + ωi £X (dui )
= X(ωi ) dui + ωi dX i

which coincides with our previous formula. We can read this in two ways; if we
have Leibniz’ formula for products then this says the general formula for £X (ω)
is derivable from the formulas for £X (f ) and £X (duk ), or we can view it as a
justification for Leibniz’s formula given our previous work.
We can sweat a little more out of the formula

£X (ω) = X(ωi ) dui + ωi dX i

by applying it to a vector field Y . We get

£X (ω)(Y ) = X(ωi ) dui (Y ) + ωi dX i (Y )

= X(ωi )Y i + ωi Y (X i )
= X(ωi Y i ) − ωi X(Y i ) + ωi Y (X i )
= X(ω(Y )) − ωi [X, Y ]i
= X(ω(Y )) − ω([X, Y ])
= £X (ω(Y )) − ω([X, Y ])

This is often written as

£X (ω(Y )) − £X (ω)(Y ) = ω([X, Y ])

This is a handy formula and we will meet it again in another context.

7. THE dω(X0, . . . , Xr ) FORMULA

In this section we wish to derive a formula which shows how dω acts on T (M ).
Specifically, for ω in Λr (M ), we want the formula
r
X
dω(X0 , X1 , · · · Xr ) = (−1)i X i (ω(X0 , · · · , Xi−1 , Xi+1 , · · · , Xr )
i=0
X
+ (−1)i+j ω([Xi , Xj ], · · · , Xi−1 , Xi+1 , · · · , Xj−1 , Xj+1 , · · · , Xr )
0≤i<j≤r

This way of expressing the formula is a little cumbersome so we will make of

use of the convention that a hat on a letter means the letter is omitted . The
formula then reads
r
X
dω(X0 , X1 , · · · Xr ) = (−1)i X i (ω(X0 , · · · , X̂i , · · · , Xr )
i=0
X
+ (−1)i+j ω([Xi , Xj ], · · · , X̂i , · · · , X̂j , · · · , Xr )
0≤i<j≤r

11
 This interesting formula explicitly gives the value of dω evaluated on vec-
tors. It is even possible to use it to define dω, as Godbillion[2], does instead of
determining dω by the three rules
∂f
df = dui
∂ui
d(ω 1 ∧ ω 2 ) = dω 1 ∧ ω 2 + (−1)deg(ω1 ) ω 1 ∧ dω 2
ddω = 0

However, we will follow the classical route which requires us to prove this for-
mula.
It is worth doing the special case of a one-form first to get the feel for the
situation. We have ω ∈ Λ1 (M ) and X0 and X1 are two vector fields on M .
Then locally

ω =ωi dui
∂ωi j
dω(X0 , X1 ) = du ∧ dui (X0 , X1 )
∂uj
∂ωi  j i i j

= du (X 0 )du (X 1 ) − du (X 0 )du (X 1 )
∂uj
∂ωi  j i i j

= X X
0 1 − X 0 1X
∂uj
∂ωi ∂ωi
= X0j j X1i − X1j j X0i
∂u ∂u
∂   ∂X i
= X0j j ωi X1i − X0j ωi j1
∂u ∂u
i
∂   ∂X
− X1j j ωi X0i + X1j ωi j0
∂u ∂u
 ∂X i i
j ∂X0
= X0 (ω(X1 ) − X1 (ω(X0 ) − ωi X0j 1
− X 1
 ∂uj  ∂uj
= X0 (ω(X1 ) − X1 (ω(X0 ) − ωi [X0 , X1 ]i
 
= X0 (ω(X1 ) − X1 (ω(X0 ) − ω [X0 , X1 ]

This verifies the formula for the case r = 1. Unfortunately for r > 1 the situation
is not so simple. As Helgason[3] remarks, this formula is essentially trivial but
it is a bit tricky to push through in detail. In fact, it’s not too easy to find a
source where this is done. The steps are, in the big picture, the same as those
for r = 1 but there are so many things to keep straight that the proof is a bit
unwieldy.
Before starting we remind the reader of the formula for diffentiating a
determinant. It is, for r = 3, simply
 f g h   ∂f g h   f ∂g h   f g ∂h 
       
∂    ∂u
∂i
  ∂u
∂j
  ∂u
∂k 

i j k  =  ∂u j k  +  i ∂u k  +  i j ∂u
∂u 
 
l m n   ∂u ∂l
m n   l ∂m ∂u n   l m ∂n ∂u


12
as is immediate from the formula
 1 
 a1 a12 a13 
π(1) π(2) π(3)
 2 X
 a1 a22 a23  =

 3  sgn(π) a1 a2 a3
3 3
 a1 a2 a3  π∈S3

By linearity the preceding formula gives, for a vector field X,

       
 f g h   X(f ) g h   f X(g) h   f g X(h) 
   
X  i j k  =  X(i) j k  +  i X(j) k  +  i j X(k) 
  
 l m n   X(l) m n   l X(m) n   l m X(n) 

We also remind the reader that the hat on an expression means that it is omitted,
so that for example
   
 f g ĥ i   f g i 
   
 j k l̂ m  =  j k m 
   
 n o p̂ q   n o q 

We are now going to prove the general formula

r
X
dω(X0 , X1 , · · · Xr ) = (−1)i X i (ω(X0 , · · · , X̂i , · · · , Xr )
i=0
X
+ (−1)i+j ω([Xi , Xj ], X0 , · · · , X̂i , · · · , X̂j , · · · , Xr )
0≤i<j≤r

By linearity, it suffices to prove the formula for

ω = f dui1 ∧ . . . ∧ duir

We recall that, with Xk = Xkj ∂

∂uj , that

dui1 ∧ . . . ∧ duir (X1 , . . . , Xr ) = det |duij (Xk )|

Thus we are trying to prove that

r
X
(−1)i Xi f dui1 ∧ . . . ∧ duir (X0 , . . . , X̂i , . . . , Xr )


dω(X0 , . . . , Xr ) =
i=0
X
+ (−1)i+j f dui1 ∧ . . . ∧ duir ([Xi , Xj ], X0 , . . . , X̂i , . . . , X̂j , . . . , Xr )
0≤i<j≤r

This is in some sense straightforward, and there is only one substantive step;
everything else is cosmetic rearrangement. I will point out the substantive step
when we get there. For ease of reading I will leave out the wedges in the
calculation. Here we go....
r
X ∂f
dω(X0 , . . . , Xr ) = j
duj dui1 . . . duir (X0 , X1 , . . . , Xr )
j=1
∂u

13
  

 duj (X0 ) duj (X1 ) . . . duj (Xr ) 

X ∂f  dui1 (X0 ) dui1 (X1 ) . . . dui1 (Xr ) 
=  
∂uj  ··· ··· ··· ··· 
j  
 duir (X0 ) duir (X1 ) . . . duir (Xr ) 
X0j X1j
 
 ... Xrj 
X0i1 X1i1

X ∂f  ... Xri1 
= 
∂uj  ··· ··· ··· · · · 
j 
 X0ir X1ir ... Xrir 

Now we do an expansion off the ith column to get

 i1
. . . X̂ii1 . . . Xri1

 X0 
r  i
. . . X̂ii2

XX ∂f  X2 . . . Xri2
(−1)i j Xij  0

= 
∂u  ··· ··· ··· ··· 
j i=0 
 X ir . . . X̂iir . . . Xrir 
0
 i1
X̂ii1 Xri1 

 X0 . . . ...
r  i
X  X 2 ... X̂ii2 ... Xri2 
= (−1)i Xi (f )  0
 ··· ··· ··· ··· 
i=0 
 X ir . . . X̂iir ... Xr ir 
0

Now comes the one substantive step; we use X(f )g = X(f g) − f X(g):
  i1
 X0 . . . X̂ii1 . . . Xri1 

r
  X i2 . . . X̂ i2 . . . X i2 
X  
i 0 i r
= (−1) Xi f 
  
· · · · · · · · · · · · · · · 
i=0 
 X ir . . . X̂ ir . . . X ir 
0 i r
 i1 i
. . . Xri1 

 X0 . . . X̂i 1
r 
X  X i2 . . . X̂ i2 . . . X i2 
− (−1)i f Xi   0 i
 · · · · · · · · · · · · · · · 
r 

i=0  
 X ir . . . X̂ ir . . . X ir 
0 i r
 dui1 (X ) . . . du\
  
i1 (X ) . . . dui1 (X ) 
 0 i r 
r
i2 \i2 i2
(−1)i Xi f  du (X0 ) . . . du (Xi ) . . . du (Xr ) 
X   
=
  
i=0
  · · · · · · · · · · · · · · · 
 duir (X ) . . . du\
 
ir (X ) . . . duir (X ) 
0 i r
 i1
 X0 . . . Xi (Xki1 ) · · · X̂ii1 . . . Xri1 

r
 X i2 . . . Xi (X i2 ) · · · X̂ i2 . . . X i2 
XX  
− (−1)i f  0 k i r 

i=0 k6=i  ··· ··· ··· · · · · · · · · · · · · 

 X ir . . . X (X ir ) · · · X̂ ir . . . X ir 
0 i k i r
Xr
(−1)i Xi f dui1 · · · duir (X0 , . . . , X̂i , . . . Xr )


=
i=0

14
 X0i1 . . . Xi (Xki1 ) · · · X̂ii1 . . . Xri1 
 

r X
X0i2 . . . Xi (Xki2 ) · · · X̂ii2

X . . . Xri2 
(−1)i f 

−
··· ··· ··· ··· ··· · · · · · · 
i=0 k<i 
 X0ir . . . Xi (Xkir ) · · · X̂iir . . . Xrir 
X0i1 . . . X̂ii1 . . . Xi (Xki1 ) . . . Xri1 
 

r X 
X
i
 X0i2 . . . X̂ii2 . . . Xi (Xki2 ) . . . Xri2 
− (−1) f 
··· ··· ··· ··· 
i=0 k>i  
 X0ir . . . X̂iir . . . Xi (Xkir ) . . . Xr ir 

r
X
= (−1)i Xi (ω(X0 , . . . , X̂i , . . . , Xr )
i=0
 Xi (Xki1 ) X0i1 · · · X̂ki1 · · · X̂ii1 · · · Xri1
 

r i−1  i2 i2

XX
i k  Xi (Xk ) X0
 · · · X̂ki2 · · · X̂ii2 · · · Xri2 
− (−1) f (−1)  
··· ··· ··· ··· ··· ··· ··· ··· 
i=0 k=0  
 X (X ir ) X ir · · · X̂kir · · · X̂iir · · · Xrir 
i k 0
 Xi (Xki1 ) X0i1 · · · X̂ii1 · · · X̂ki1 · · · Xri1 
 
r r i2
X0i2 · · · X̂ii2 · · · X̂ki2

X X
i k−1  Xi (Xk )
 · · · Xri2 
− (−1) f (−1)  ··· ··· ··· ··· ··· ··· · · · · · · 
i=0 k=i+1 
 X (X ir ) X0ir · · · X̂iir · · · X̂kir · · · Xrir 
i k
Xr
= (−1)i Xi (ω(X0 , . . . , X̂i , . . . , Xr )
i=0
 Xi (Xki1 ) − Xk (Xii1 ) X0i1 · · · X̂ii1 · · · X̂ki1 · · · Xri1 
 
i2 i2
X0i2 · · · X̂ii2 · · · X̂ki2

X
i+k  Xi (Xk ) − Xk (Xi )
 · · · Xri2 
− (−1) f
··· ··· ··· ··· ··· ··· · · · · · · 
0≤i<k≤r 
 X (X ir ) − X (X ir )
i k k i X0ir · · · X̂iir · · · X̂kir · · · Xrir 
r
X
= (−1)i Xi (ω(X0 , . . . , X̂i , . . . , Xr )
i=0

· · · du\ · · · du\
 
 dui1 ([X , X ]) dui1 (X ) i1 (X ) i1 (X ) · · · dui1 (Xr ) 
 i k 0 i k 
i
i+k  du 2 ([Xi , Xk ]) du 2 (X0 )
i
· · · du\
i2 (X ) · · · du\
i2 (X ) · · · dui2 (Xr )
X  
− (−1) f i k 
··· ··· ··· ··· ··· ··· ··· ···

 
0≤i<k≤r
· · · du\ · · · du\
 
 duir ([Xi , Xk ]) duir (X0 ) ir (X )
i
ir (X )
k
ir
· · · du (Xr ) 
Xr
= (−1)i Xi (ω(X0 , . . . , X̂i , . . . , Xr )
i=0
X
− (−1)i+k f dui1 dui2 · · · duir ([Xi , Xk ], X0 , . . . , X̂i , . . . , X̂k . . . , Xr )
0≤i<k≤r
r
X
= (−1)i Xi (ω(X0 , . . . , X̂i , . . . , Xr )
i=0

15
 X
− (−1)i+k ω([Xi , Xk ], X0 , . . . , X̂i , . . . , X̂k . . . , Xr )
0≤i<k≤r

and the proof is complete.

8. SOME ALGEBRAIC IDEAS

The calculation in the previous section might be regarded by some people as
unpleasant. In the next section we will develop what we will call the French
Method for proving identities. This will give us a less computational mode of
attack. However, to do this efficiently we want to bring on board some algebraic
ideas which will clarify the structure of the theory.
We suppose that we have an Algebra A (i.e. a vector space with an associa-
tive multiplication respecting the vector space operations) which is GRADED
in the sense that
∞
M
A = Ak
k=0
Ai · Aj ⊆ Ai+j
L∞
Elements of Ai are called homogeneous of degree i. A0 and k=0 A
2k
form
natural subalgebras of the graded algebra A. A paradigmatic example of this
concept is the Exterior Algebra Λ(M ) = ∞ k
L
k=0 Λ (M )
In recent years the term communtative has come to be applied to graded
algebras that satisify the equation

xi xj = (−1)ij xj xi xi ∈ Ai , xj ∈ Aj

However, to prevent confusion we will refer to this as a Graded Commuatitive

Algebra. There is a commuatator that goes with this situation given by

[xi , xj ] = xi xj − (−1)ij xj xi

and of course the algebra is graded communtative if and only if the commutator
vanishes.
The case of interest to us is when the graded algebra A is generated by
the elements of degrees 0 and 1. In the case of A = Λ(M ) this would be the
functions and 1−forms. Abstractly we say

A is generated by A0 ⊕ A1

Next we introduce the concept of a Graded Derivation of degree p. We will

temporarily use the letters a and b for graded derivations. A function a : A → A
is a graded derivation if for each r it satisfies

a: Ar → Ar+p
a(xy) = a(x)y + (−1)rp xa(y) for x ∈ Ar

16
We also define a bracket [a, b] of two graded derivations of degrees p and q
respectively by
[a, b] = ab − (−1)pq ba
The theorem is then
Theorem If a and b are graded derivations of degrees p and q respectively then
[a, b] is a graded derivation of degree p + q.
Proof The proof is straightforward if a trifle intricate. Recall that

a : Ar → Ar+p
b : Ar → Ar+q
a(xy) = a(x)y + (−1)rp xa(y) x ∈ Ar
b(xy) = b(x)y + (−1)rq xb(y) x ∈ Ar

By the definition of bracket for graded derivation we have

[a, b] = ab − (−1)pq ba

Hence

[a, b](xy) = (ab)(xy) − (−1)pq (ba)(xy)

a (bx)y + (−1)rq x(by) − (−1)pq b (ax)y + (−1)rp x(ay)
   
=
= (abx)y + (−1)p(r+q )(bx)(ay) + (−1)rq (ax)(by) + (−1)rp x(aby)
−(−1)pq (bax)y + (−1)q(r+p) (ax)(by) + (−1)rp (bx)(ay) + (−1)rp (−1)rq x(bay)
 

= (ab − (−1)pq ba)x y + (−1)r(p+q) x (ab − (−1)pq ba)y

+ (−1)p(r+q) − (−1)pq+rp (bx)(ay) + (−1)rq − (−1)pq+qr+qp (ax)(by)

= ([a, b](x) y + (−1)(p+q)r x[a, b](y)

because the 3rd and 4th terms in the penultimate equation are 0 due to cancel-
lation of the powers of -1 in the coefficients.
The popular terminology for derivations is as follows.
If p is even a graded derivation is called a derivation
If p is odd a graded derivation is called an antiderivation
Let a1 and a2 be antiderivations with odd degrees p1 and p2
Let b1 and b2 be derivations with even degrees q1 and q2
Then our result about brackets of graded derivations tell us that

[a1 , a2 ] = a1 a2 + a2 a1 is a derivation of degree p1 + p2

[a1 , a1 ] = 2a21 is a derivation of degree 2p1

[a1 , b1 ] = a1 b1 − b1 a1 is an antiderivation of degree p1 + q1
[b1 , b2 ] = b1 b2 − b2 b1 is a derivation of degree q1 + q2

17
 The following theorem is almost obvious.
L
Theorem Suppose that the graded algebra A = Ai is generated by A0 ⊕ A1
as an algebra and suppose a and b are graded derivations of degree p and suppose
that they agree on A0 ⊕ A1 (that
L is, on scalars and ”vectors”). Then they agree
on the whole graded algebra Ai .
Proof Since the elements of Ai are generated by products of vectors from A1
and scalars from A0 , the value of a on an element of Ai is determined by the
values on A0 and A1 . Since a and b coincide on A0 and A1 , they coincide on
the entire algebra.
Now the obvious next question is whether one can create a graded derivation
on an algebra generated by A0 ⊕ A1 by specifying it it on the elements of A0
and A1 and requiring it to satisfy the graded derivation law. The answer, in
general, is NO because an element of Ai may be a product of elements of A0
and A1 in many different ways, and the different ways may lead to inconsistent
values of the graded derivation on that element. Algebraically this means that
the graded algebra has relations.
But suppose the algebra A is a FREE graded Algebra, which means there
are no relations between the elements except those that are forced by being
a graded algebra. Then indeed the construction of the derivation is possible.
An example of such a free graded algebra is the Exterior Algebra of Forms
Λ(M ) on a manifold M . It is for precisely this reason that the operator d may
be uniquely determined by its value on functions f and one forms ω and the
law d(ωη) = d(ω)η + (−1)deg ω ωd(η) which says precisely that d is a graded
derivation of degree 1. This bit of trickery works because Λ(M ) is free.

9. INTERACTION OF THE LIE DERIVATIVE

WITH d AND INNER PRODUCT
In this section we will use the ideas in the previous section to develop what
we will call the French Method for proving identities. This will give us a less
computational mode of attack.
First however we show that £X commutes with d. Recall that if X ∈
Γ(T (M )) (i.e. a vector field) and φt is the corresponding flow then for ω ∈
Λr (T ∗ (M )) = Λr (M )
φ∗ (ω) − ω
£X ω = lim t
t→0 t
Recall that φ∗t (ω1 ∧ ω2 ) = φ∗t (ω1 ) ∧ φ∗t (ω2 ). Then

φ∗t (ω1 ∧ ω2 ) − ω1 ∧ ω2 = φ∗t (ω1 ) ∧ φ∗t (ω2 )ω1 ∧ ω2

= φ∗t (ω1 ) ∧ φ∗t (ω2 ) − ω1 ∧ φ∗t (ω2 )
+ ω1 ∧ φ∗t (ω2 ) − ω1 ∧ ω2
φ∗ (ω1 ∧ ω2 ) − ω1 ∧ ω2
£X (ω1 ∧ ω2 ) = lim t
t→0 t

18
 (φ∗t (ω1 ) − ω1 ) ∧ φ∗t (ω2 ) + ω1 ∧ (φ∗t (ω2 ) − ω2 )
= lim
t→0 t
= £X (ω1 ) ∧ ω2 + ω1 ∧ £X (ω2 )
as we would expect for a reasonable product.
In a similar way, since φ∗t (d ω) = d φ∗t (ω) we have
φ∗t (dω) − d ω d φ∗t (ω) − d ω
£X (d ω) = lim = lim = d £X (ω)
t→0 t t→0 t
Note that t is not one of the variables involved in d. We previously proved the
formula above for 1−forms but this proof is valid for r−forms. A different proof
of this is possible using methods we are about to introduce.
Next notice that £X is a graded derivation of degree 0 and d is a graded
derivation of degree 1. Hence by the theorem in the previous section,
[d, £X ] = d ◦ £X − (−1)−1·0 £X ◦ d = d ◦ £X − £X ◦ d
is a graded derivation of degree 1. We have just shown it is identically 0. How-
ever, note that this fact could be proved by showing that [d, £X ] is identically
0 on functions and 1−forms. It would then follow that it is 0 on all forms, by
the discussion at the end of the previous section. We will call this the French
Method.
Clearly there is no fun to be had here with d and £X because they commute.
For real fun we need another graded derivation which does NOT commute with
d or £X .
Our first example is simply to use two Lie Derivatives, £X and £Y . We
will show that
[£X , £Y ] = £[X,Y ]
Since these are graded derivations of degree 0, it suffices to show that they
coincide on functions and one forms. Note that we have
[£X , £Y ] = £X ◦ £Y − (−1)0·0 £Y ◦ £X = £X ◦ £Y − £Y ◦ £X
On functions this gives
[£X , £Y ](f ) = £X ◦ £Y (f ) − £Y ◦ £X (f ) = £X (Y (f )) − £Y (X(f ))
= X(Y (f )) − Y (X(f )) = (XY − Y X)(f ) = [X, Y ](f )
= £[X,Y ] (f )
For one-forms it is perfectly possible to perform a direct attack and fill a page
with computations that establish that the two derivations coincide on one forms.
However, we would like to do this in a smarter way. The miminal one-form we
can work with is dui and we can use the commutativity of £X and d effectively
here; we don’t even need to look at the formula for £X on one-forms.
(£X £Y − £Y £X )(dui ) = d(£X £Y − £Y £X )(ui )
= d £[X,Y ] (ui )
= £[X,Y ] (dui )

19
Since the two derivations £X £Y − £Y £X and £[X,Y ] agree on functions f and
the special one forms dui , they will agree on all one forms ω = fi dui , as required.
Another very important example is the inner product of a vector field and
a differential form. This is defined by
Def

X |f = 0 for f ∈ Λ0 (M )
X |ω = ω(X) for ω ∈ Λ1 (M )
X | is a graded derivation of degree − 1

This completely defines X | as we discussed previously, and no further formulae

are theoretically necessary. But it is nice to know that for ω ∈ Λr (M )

X | ω(Y2 , Y3 , . . . , Yr ) = ω(X, Y2 , Y3 , . . . , Yr )

However, in order not to break up the derivation we will prove the general
formula in appendix 1 of this section. The case r = 2 we need below.
We now want to use all our equipment to prove Cartan’s Homotopy formula
Theorem For X a vector field on M and ω ∈ Λr (M ).

£X ω = d(X | ω) + X | (dω)

Since d is a graded derivation of degree 1 and X | is a graded derivation of

degree -1 we know that

[d, X | ] = d ◦ X | + X | ◦ d

will be a graded derivation of degree 0. Since the only graded derivation of

degree 0 that we know is £X the result is not surprising. Moreover, as we
discussed above we can prove the formula by verifying it for f ∈ Λ0 (M ) and
ω ∈ Λ1 (M ). Since £X and [d, X | ] coincide on Λ0 (M ) and Λ1 (M ), they must
coincide on all Λr (M ). We proceed to verify the two cases. for f ∈ Λ0 (M )

d(X | f ) + X | (d f ) = X | (d f )
= d f (X) = X(f ) = £X (f )

and for ω ∈ Λ1 (M ), using the formula (X | ω)(Y ) = ω(X, Y ),

(d(X | ω) + X | (dω))(Y ) = d(ω(X))Y + dω(X, Y )

= Y (ω(X)) + X(ω(Y )) − Y (ω(X)) − ω([X, Y ])
= X(ω(Y )) − ω([X, Y ])
= £X (ω)(Y )

where we used the formula £X (ω(Y )) − £X (ω)(Y ) = ω([X, Y ])

20
 Since we have three graded derivations, d, X | and £X , we should be able
to find formulas using any two of them. We just did the first two, and the pair
d, £X is dull since the two commute, so the remaining possibility is £X and
X | . The commutator of these two should give us a graded derivation of degree
-1. Let us see if we can turn it up and simultaneously prove the formula. For
f ∈ Λ0 (M )
£X (Y | f ) − Y | £X (f ) = 0
No help there. For ω ∈ Λ1 (M )

£X (Y | ω) − Y | £X (ω) = £X (ω(Y )) − (£X (ω))(Y )

= ω([X, Y ])
= [X, Y ] | ω

AHA!. We now see that

[£X , Y | ] = £X ◦ Y | − Y | ◦ £X = [X, Y ] |

which is the formula we were seeking. The formula is true on functions and
1−forms, and since it is a graded derivation is true on all r−forms.
We have now exhasted our supply of graded derivations. However, I wish
to present one more formula which has a slightly different feel to it and is not
proved in the same way. For this formula we must define the Outer Product or
Exterior Product.
Def Let ω ∈ Λ1 (M ). Then the Outer Product ǫ(ω) is defined by

ǫ(ω)(η) = ω ∧ η for η ∈ Λr (M )

Note that ǫ(ω) is not a derivation. However, it does have an interesting formula
which looks a little like a derivation formula. First, recall that X | is a graded
derivation of degree -1 so we have for ω ∈ Λ1 (M )

X | (ω ∧ η) = (X | ω) ∧ η − ω ∧ (X | η)

Rewriting this with the outer product we have

X | (ǫ(ω)(η)) = (ω(X))η − ǫ(ω)(X | η)

X | (ǫ(ω)(η)) + ǫ(ω)(X | η) = (ω(X))η

and so we have
X | ◦ ǫ(ω) + ǫ(ω) ◦ X | = ω(X)
where the right hand side is to be interpreted as the operator of scalar multi-
plication by ω(X).
Appendix 1 Here we prove the formula for the inner product

(X | ω)(Y2 , . . . , Yr ) = ω(X, Y2 , . . . , Yr )

21
for ω ∈ Λr (M ). We begin in a leisurely manner by proving the formula for
r = 2. For this we have

X | (ω1 ∧ ω2 ) = (X | ω1 )ω2 − ω1 (X | ω2 )
= ω1 (X)ω2 − ω2 (X)ω1
|
(X (ω1 ∧ ω2 ))(Y ) = ω1 (X)ω2 (Y ) − ω2 (X)ω1 (Y )
= (ω1 ∧ ω2 )(X, Y )

For the general case we remind the reader of some formulae. First
 
ω1 (Y1 ) ω1 (Y2 ) · · · ω1 (Yr )
 ω2 (Y1 ) ω2 (Y2 ) · · · ω2 (Yr ) 
ω1 ∧ ω2 ∧ . . . ∧ ωr (Y1 , . . . , Yr ) = det 
 
··· ··· ··· ··· 
ωr (Y1 ) ωr (Y2 ) · · · ωr (Yr )
r
X
ω1 ∧ ω2 ∧ . . . ∧ ωr (Y1 , . . . , Yr ) = (−1)i−1 ω1 (Yi )ω2 ∧ . . . ∧ ωr (Y1 , . . . , Ŷi , . . . , Yr )
i=1
r−1
Thus for ω ∈ Λ
r
X
ω1 ∧ ω(Y1 , . . . , Yr ) = (−1)i−1 ω1 (Yi )ω(Y1 , . . . , Ŷi , . . . , Yr )
i=1

Now we begin our proof by induction. Assume the formula is true for r − 1.
Then for ω1 ∈ Λ1 (M ) and ω ∈ Λr (M )

Y1 | (ω1 ∧ ω)(Y2 , . . . , Yr )
= (Y1 | ω1 )ω(Y2 , . . . , Yr ) − (ω1 ∧ (Y1 | ω))(Y2 , . . . , Yr )
r
X
= ω1 (Y1 )ω(Y2 , . . . , Yr ) − (−1)i ω1 (Yi )(Y1 | ω)(Y2 , . . . , Ŷi , . . . , Yr )
i=2

Now using the induction assumption we have

Y1 | (ω1 ∧ ω)(Y2 , . . . , Yr )
r
X
= ω1 (Y1 )ω(Y2 , . . . , Yr ) − (−1)i ω1 (Yi )ω(Y1 , . . . , Ŷi , . . . , Yr )
i=2
r
X
= (−1)i−1 ω1 (Yi )ω(Y1 , . . . , Ŷi , . . . , Yr )
i=1
= (ω1 ∧ ω)(Y1 , . . . , Yr )

This completes the proof.

22
References
[1] Berline,N; Getzler, E. & Vergne, M. Heat Kernals and Dirac Operators
Springer Verlag, Berlin, 1992
[2] Godbillon, Claude ; Géométrie Différentielle et Méchanique Analytique
Hermann Paris 1969
[3] Helgason, Sigurdur; Differential Geometry and Symmetric Spaces Ameri-
can Mathematical Society, Tel Aviv, 2000
[4] Loomis, Lynn H.; Advanced Calculus 2nd edition, Jones and Bartlett,
Boston, 1990
[5] Morrey, Charles B.; Multiple Integrals in the Calculus of Variations,
Springer Verlag, Berlin, 1966

23