Frobenius Theorem Two Ways

Robert Lyons

March 4, 2016

Every smooth vector field X on a manifold has integral curves which are embedded 1−dimensional
sub-spaces of M that have tangent spaces spanned by the vector field X. Let M be a manifold
of dimension m and assume that for every x ∈ M there is a p−dimensional subspace Ex of the
tangent space of M at x, Tx (M ). Is it possible to find a submanifold N ⊂ M so that at every
y ∈ N the tangent space at y ∈ N of N is just Ey = Ty (M ). Is it enough that the planes Ex
vary smoothly with x? It turns out one can construct p−dimensional sub-spaces of the tangent
spaces of M that vary in a smooth way that are not the tangent spaces of any sub-manifold.
Frobenius theorem gives necessary and sufficient conditions required of these p−planes so that
they are the tangent planes of a p−dimensional sub-manifold.

In this paper we first present Frobenius theorem using the language of vector fields. The
condition that the p−planes are the tangent spaces of a p−submanifold is that the distribution
is closed under the Lie bracket. Following this we shall give an equivalent dual presentation
using the language of differential forms. The p−planes are now the kernels of the differential
forms. This second approach is in the spirit of the Cartan-Kahler theorem where the extension
problem is turned into a problem of finding the zeros of an ideal of functions generated by the
differential ideal.

Paper source:
This paper downloaded from: www.pdx.edu/∼rlyons.
Send comments to: rlyons@pdx.edu.

Contents
1 Differentiable Manifolds 3
1.1 Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Tangent Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Integral Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Lie Derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1
2 CONTENTS

1.4.1 Exterior Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.5 Differential Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6 Coordinate Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.6.1 Linear Coordinate Transforms . . . . . . . . . . . . . . . . . . . . . . . 12
1.7 Implicit Function Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2 Frobenius Theorem And Vector Fields 13

2.1 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2 Proof of Frobenius Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3 Frobenius Theorem And Differential Forms 22

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 Frobenius Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 Differential Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4 References 37
 3

1 Differentiable Manifolds
We start with a short introduction to the definitions of differential geometry. There are many
good books on the subject. An excellent modern treatment is [5]. Fora more elementary book
that traces differential geometry back to it’s roots in non-Euclidean geometry see [3].

1.1 Manifolds
To make the definition of a manifold precise the local Euclidean structure is modeled as a
homeomorphism (continuous map with a continuous inverse) from a neighborhood to a subset
of Rn . This is called a coordinate map and is fundamental to the subject.

Definition 1.1.1. A smooth differentiable manifold M of dimension m is a second countable

Hausdorff space (for example some subset of RK for some K) together with a collection of C ∞
coordinate maps. In particular this means that there is a collection of open sets Uα ⊂ M and
a collection of homeomorphisms ϕα : Uα → RN that satisfy,
S
• Uα = M ,
α

• When Uα ∩ Uβ 6= ∅ then the mapping ϕα ◦ ϕ−1 ∞

β : ϕβ (Uα ∩ Uβ ) → ϕα (Uα ∩ Uβ ) is C .

We could replace C ∞ in this definition with ”real analytic” and we would get a real analytic
differentiable manifold.

The open sets Uα are called coordinate neighborhoods. The second countable axiom means
that M , as a topological space, has a countable base. This axiom is essential so that M has
a partition of unity which allows us to define an integral in a natural way and show every
manifold has a Riemannian metric. This axiom does not enter our discussion directly.

We denote the k-th coordinate projection by rk : RN → R which is defined as

rk (a1 , · · · , aN ) = ak . (1)

Using this we define coordinates on M by,

xk = rk ◦ ϕα : Uα → R. (2)

If x ∈ Uα then x is uniquely determined by its coordinates x1 (x), · · · xm (x) .

We use these coordinate neighborhoods to define many of the features found in Euclidean
vector calculus. For example, a smooth map f : M → R is defined to be smooth if the maps

f ◦ ϕ−1
α : ϕα (Uα ) → R
4 1 DIFFERENTIABLE MANIFOLDS

are smooth for every coordinate neighborhood (Uα , ϕα ). Notice that ϕα is a homeomorphism
so we can take the inverse map and it is continuous. We can take the derivative of a smooth
function in the direction of coordinates xk by taking,
∂(f ◦ ϕ−1
α ) ∂f
= k.
∂xk ∂r
We extend this idea and define tangent vectors and vector fields using the coordinate neigh-
borhoods we defined above.

1.2 Tangent Vectors

In Rm a smooth curve has a derivative which is the tangent to the curve. In an abstract
manifold we will define a tangent vector X to be a directional derivative that takes a smooth
function f and associates a derivative X(f ) which we think of as the directional derivative of
f in the direction X.
For example, given a smooth differentiable manifold M and given a curve γ : (−a, a) → M
the vector X tangent to γ maps a smooth function f : M → R to its directional derivative,
dγ d(f ◦ γ)
X(f ) = (f ) = . (3)
dt dt
We shall define a tangent vector, at a point x ∈ M , to be a operator that takes smooth
functions f : M → R to a real number X(f )x ∈ R. This operator is a derivation, which means
that if f, g are smooth functions at the point x ∈ M then

X(f g)x = X(f )x g(x) + f (x)X(g)x .

The set of vectors at a point x ∈ M is denoted by Tx (M ) and is called the tangent space at x.
We can add derivations and scale by real scalars a, b ∈ R,

(aX + bY ) (f ) = aX(f ) + bY (f ).

This makes the set of derivations at x into a real vector space. One can prove that this vector
space has dimension m (see [6] ). This makes intuitive sense since we think of m−dimensional
surfaces in RK have m−dimensional tangent spaces, just as 2−dimensional surfaces in R3 has
2−dimensional tangent planes.

A vector field is an assignment of a vector X ∈ Tx (M ) for every x ∈ M . A vector field

is smooth if for every smooth f : M → R the function X(f ) : M → R is a smooth mapping.
Given a coordinate system ϕ : U → RN with coordinates xj = rj ◦ ϕ we define vector fields ∂x∂ k
by,
∂(f ◦ ϕ−1 ) 

∂
(f )x = . (4)
∂xk ∂rk 
ϕ(x)
1.2 Tangent Vectors 5

∂
These m tangent vectors ∂x1
,··· , ∂x∂m form a basis of the tangent space Tx (M ). For example,
if X ∈ Tx (M ) we write,
m
X ∂
X= Xk .
∂xk
k=1

To find the coefficients Xk

we apply X to the function xj : M → R. We note that,

∂(xj ◦ ϕ−1 )  ∂(rj ◦ ϕ ◦ ϕ−1 )  ∂(rj ) 

  
∂ j
= δkj .


x = = =
∂xk ∂rk 
ϕ(x) ∂r k 
ϕ(x) ∂r k 
ϕ(x)

Using this we see that we have,

m m
∂
X k δkj = X j .
X X
X(xj ) = Xk (x j
) =
∂xk
k=1 k=1

This means that,

m
X ∂
X= X(xj ) .
∂xj
j=1

Let M and N be differentiable manifolds of dimension m and n respectively. Let φ : M → N

be a smooth mapping. If γ : I → M is a smooth curve then (φ ◦ γ) : I → N is a smooth curve
on N The differential maps Tx (M ) in a way that corresponds to this mapping of curves. Using
the association in equation 3 we write,

dγ d(φ ◦ γ)
dφ( )=
dt dt
If we apply this to a function f we get,
 
dγ d(φ ◦ γ) d(f ◦ φ ◦ γ) dγ
dφ (f ) = (f ) = = (f ◦ φ) .
dt dt dt dt

So the differential dφ maps vectors in Tx (M ) to vectors in Tφ(x) (N ). We formalize this with

the following definition.

Definition 1.2.1. The differential dφ : Tx (M ) → Tφ(x) (N ) is a mapping that maps X to the

derivation dφ(X) which takes a function f : N → R to,

dφ(X)(f ) = X(f ◦ φ).

Example 1.2.2. Let M = S 2 = x ∈ R3 |kxk2 = R2 . We define a parameterization for a



fixed R
ϕ−1 : (0, π) × (0, 2π) → S 2
6 1 DIFFERENTIABLE MANIFOLDS

given by,
 
Rsin(θ)cos(ϕ)
ϕ−1 (θ, φ) = Rsin(θ)sin(ϕ)
Rcos(θ)

The coordinates (x1 , x2 ) have,

x1 : S 2 → R defined by x1 ϕ−1 (θ, φ) = θ,

x2 : S 2 → R defined by x2 ϕ−1 (θ, φ) = φ,

1.3 Integral Curves

Given a smooth vector field X and an x ∈ M is there a curve γ : (−a, a) → M that has X as
its tangent vector? We need to solve the equation,

d(xk ◦ γ)
X(xk ) = X k =
dt

This ODE system has a unique solution which is called the integral curve of X.
Let M be a C ∞ manifold and let X be a smooth vector field on M . If x0 ∈ M has X(x0 ) 6= 0
then there is a smooth curve γ : I → M with I = (a, b) with a < 0 < b that satisfies,

γ(0) = x0

dγ 
= X(x)
dt  x

For a more complete discussion of these issues in this section see [5].
This can be done in a smooth way so that for every x0 ∈ M with X(x0 ) 6= 0 there is a
neighborhood U ⊂ M and an  > 0 and a smooth mapping ϕ : (−, ) × M → M that satisfies
the following properties.

1. ϕ(0, x) = x

dγ 
2. γ(t) = ϕt (x) satisfies dt t = X(ϕt (x)).

3. If |t|, |s|, |t + s| <  and if ϕt (x) ∈ U then ϕs+t (x) = ϕ(s) ◦ ϕt (x).

We denote ϕ(t, x) = ϕt (x). This is called a local 1−parameter sub-group of diffeomorphisms.

1.4 Lie Derivative 7

1.4 Lie Derivative

Let X and Y be smooth vector fields. We define the Lie Bracket as,

[X, Y ](f ) = X (Y (f )) − Y (X(f ))

One can show that the Lie Bracket is a derivation and so [X, Y ] is a vector field on M .

The bracket is easily computed in coordinates. Let (U, ϕ) be a coordinate neighborhood

with coordinates x1 , · · · , xm . We write X and Y in coordinates,
m m
X ∂ X ∂
X= Xj j
and Y = Yj j.
∂x ∂x
j=1 j=1

Now compute the coordinates of [X, Y ] by applying to the coordinate function xk ,

[X, Y ](xk ) = X(Y (xk )) − Y (X(xk )) = X(Y k ) − Y (X k )

m  k k

j ∂Y j ∂X
X
= X −Y
∂xj ∂xj
j=1

Using the 1−parameter group of diffeomorphisms we can define the Lie Derivative.

Definition 1.4.1. Let X be a vector field and let Z be a tensor field.

1



LX (Z) = lim t dϕ−t (Zϕt (x) )p − Zp
t→0

The bracket and Lie derivative are, in fact, related.

Proposition 1.4.2. Let X and Y be smooth vector fields then,

LX (Y ) = [X, Y ].

1.4.1 Exterior Algebras

Let V be a vector space of dimension n. A k−tensor T ∈ T (V ) is a multi-linear mapping,

T : V × V × · · · × V = V k → R.

A k−tensor is alternating if,

T (v1 , · · · vi , · · · vj , · · · vk ) = −T (v1 , · · · vj , · · · vi , · · · vk ).
8 1 DIFFERENTIABLE MANIFOLDS

This means permuting any two argument elements results in a negative value. We denote the
alternating k−tensors by Λk (V ). From a k−tensor we can construct an alternating tensor with
the definition,
1 X
Alt(T )(v1 , · · · vk ) = T (vσ1 , vσ2 , · · · , vσk ).
k!
σ∈Sk

If dim(V ) = n and T ∈ Λn (V ). If e1 , · · · en is a basis for V and vi = aji ej then

P
j

n
aj11 aj22 · · · ajnn T (ej1 , ej2 , · · · , ejn ))
X
T (v1 , · · · , vn ) =
j1 ,···jn =1
X
= aσ1 σ2 σn
1 a2 · · · an T (eσ1 , eσ2 , · · · , eσn ))
σ∈Sn
X
= (−1)sgn(σ) aσ1 σ2 σn
1 a2 · · · an T (e1 , e2 , · · · , en ))
σ∈Sn

= det(aji )T (e1 , · · · en )).

Definition 1.4.3. Let ω ∈ Λk (V ) and η ∈ Λm (V ) the the wedge product is defined by

(deg(ω) + deg(η))!
ω∧η = Alt(ω ⊗ η).
deg(ω)!deg(η)!

Although this abstract formula might seem rather oblique, in simple cases is comes out to
simple expressions. For example, if ω1 , ω2 , ω3 ∈ Λ1 (V ) then,
 
1 2 3 1 2! 2 3
ω ∧ω ∧ω =ω ∧ Alt(ω ⊗ ω )
1!1!
2! 3!
Alt ω 1 ⊗ Alt(ω 2 ⊗ ω 3 )


=
1!1! 2!1!
= 3!Alt ω 1 ⊗ ω 2 ⊗ ω 3

This means that for v1 , v2 , v3 ∈ V we have,

ω 1 ∧ ω 2 ∧ ω 3 (v1 , v2 , v3 ) = 3!Alt ω 1 ⊗ ω 2 ⊗ ω 3 (v1 , v2 , v3 )

1 X 1
ω ⊗ ω 2 ⊗ ω 3 (vσ1 , vσ2 , vσ3 )


= 3!
3!
σ∈S3
X
ω 1 (vσ1 )ω 2 (vσ2 )ω 3 (vσ2 ) = det ω i (vj )


=
σ∈S3

More generally we have the following,

1.5 Differential Forms 9

Proposition 1.4.4. Let θ1 , · · · θk ∈ Λ1 (V ) and v1 , · · · , vk ∈ V then

  X
θ1 ∧ θ2 · · · ∧ θk (v1 , · · · , vk ) = θ1 (vσ1 )θ2 (vσ2 ) · · · θk (vσk ) = det θi (vj )

σ∈Sk

Given a basis e1 , · · · en of V with dual basis ω 1 , · · · ω n then every k−form η can be written
as,
n
X
η= ωi1 i2 ···ik ω i1 ∧ ω i2 ∧ · · · ∧ ω ik .
i1 ,i2 ,···ik =1

There is redundancy in the coordinates ωi1 i2 ···ik . They are alternating, in the sense that

ωi1 ···ia ···ib ···ik = −ωi1 ···ib ···ia ···ik .

1.5 Differential Forms

At x ∈ M the dual space to the tangent space is denoted by T ∗ (M ) and it consists of 1−forms
ω : Tx (M ) → R. A differential form is an element of the dual space at every point x ∈ M . A
differential form is smooth if ω(X) : M → R for all smooth vector fields X. For example, for
f ∈ C ∞ (M ) we can construct the 1−form we call the gradient and for every vector field X we
define
df (X) = X(f ). (5)
We can define the set of multi-linear functionals on the tangent space

ω : Tx (M ) × · · · × Tx (M ) → R

We define an operator ∧ defined for ω ∈ Λk (Tx (M )) and η ∈ Λm (Tx (M )) and results in a

form ω ∧ η ∈ Λk+m (Tx (M )). For example, if k = m = 1 and for v, w ∈ Tx (M ), we have,

(ω ∧ η) (v, w) = ω(v)η(w) + (−1)k ω(w)η(v) = ω(v)η(w) − ω(w)η(v).

A k−form ω on M is a k−form at every point x ∈ M . The form ω is smooth if for every

smooth vector fields X1 , · · · Xk the function,

x → ωx (X1 , · · · Xk ),

is a smooth function on M . We denote the smooth k−forms on M by Λk (M ).

Given a coordinate neighborhood U , the vector fields ∂x∂ k form a basis of Tx (M ) for every
x ∈ U . The dual forms to these vector fields are denoted dxj and satisfy
∂(rj ◦ ϕ ◦ ϕ−1 ) 
  
∂ ∂
j j
= δkj .


dx = x =
∂xk ∂xk ∂rk 
ϕ(x)
10 1 DIFFERENTIABLE MANIFOLDS

If f ∈ C ∞ (M ) then df is a differential defined by equation 5. We can write this 1−form in

terms of the basis dxj ,
Xm
df = λj dxj .
j=1
∂
To find the coefficients we apply this 1−form to the vector ∂xk
,
 
∂ ∂ ∂f
df = (f ) = = λk .
∂xk ∂x k ∂xk

This means we can write df as,

m
X ∂f j
df = dx .
∂xj
j=1

We define an operator d : Λk (M ) → Λk+1 (M ) called the differential. Given a form with

coordinates
X
ω= ωi1 ,··· ,ip dxi1 ∧ dxi1 ∧ · · · ∧ dxip
i1 ,···kp
X ∂ωi1 ,···ip
dω = dxi1 ∧ dxi1 ∧ · · · ∧ dxip
∂xj
i1 ,···kp

Example 1.5.1. The sphere is the constant surfaces of the function,

f (x, y, z) = x2 + y 2 + z 2 − R2 .

The differential of this is given by,

ω = 2xdx + 2ydy + 2zdz.

The tangent to the sphere is the kernel of this 1-form.

One can define this without coordinates. In two dimensions this formula is,

dω(X, Y ) = X(ω(Y )) − Y (ω(X)) − ω([X, Y ]). (6)

1.6 Coordinate Transforms

In this section we describe how vectors and forms transform under coordinate transforms. We
use the formal description of vectors and forms to deduce the transform properties but try to
establish a link with older notation systems.
1.6 Coordinate Transforms 11

Let M be a differentiable manifold of dimension m and (U, ϕ) be a coordinate neighborhood

with coordinates (x1 , x2 , · · · xm ) where xk = rk ◦ ϕ : U → R. Recall that rk : Rm → R is the
projection onto the kth coordinate so that,
 1 
a
 a2 
rk  .  = ak .
 
 .. 
am

If we transform the open set ϕ(U ) with a diffeomorphism T : Rm → Rm then we get a new
coordinate mapping,
ϕ T
U− → ϕ(U ) −→ Rm , (7)
and we define ψ = T ◦ ϕ. The mapping T is a diffeomorphism so ψ is a homeomorphism and
it has coordinates,
y k = rk ◦ ψ = rk ◦ T ◦ ϕ.
∂
The vector fields ∂xk
form a basis of Tx (M ) for all x ∈ U are defined by,

∂ ∂(f ◦ ϕ−1 )
(f ) =
∂xk ∂rk
∂(f ◦ ϕ−1 ◦ T −1 ◦ T ) ∂(f ◦ ψ −1 ◦ T )
= =
∂rk ∂rk
−1 j
∂(f ◦ ψ ) ∂(r ◦ T )
=
∂rj ∂rk
j
∂(r ◦ T ) ∂
= (f ) .
∂rk ∂y j
Since this is true for all f we write this as,
∂ ∂(rj ◦ T ) ∂
= (8)
∂xk ∂rk ∂y j
In a similar way, we can see that,
∂ ∂(f ◦ ϕ−1 ◦ T −1 )
(f ) =
∂y k ∂rk
∂(f ◦ ϕ−1 ) ∂(ri ◦ T −1 )
=
∂ri ∂rk
This leads to the inverse transform,
∂ ∂(rj ◦ T −1 ) ∂
= (9)
∂y k ∂rk ∂xj
12 1 DIFFERENTIABLE MANIFOLDS

We can write this in a different way by performing the following computation,

∂xj ∂  k  ∂(y k ◦ ϕ−1 ) ∂(rk ◦ T ◦ ϕ ◦ ϕ−1 ) ∂(rk ◦ T )
= y = = =
∂y k ∂xj ∂rj ∂rj ∂rj
This expression is the same as is found in equation 8. Similarly, we see that,
∂y j ∂  k  ∂(xk ◦ ϕ−1 ◦ T −1 ) ∂(rk ◦ T −1 )
= x = =
∂xk ∂y j ∂rj ∂rj
Using these we re-write equations 8 and 9 as following,
∂ ∂(rj ◦ T ) ∂ ∂y j ∂
= = (10)
∂xk ∂rk ∂y j ∂xk ∂y j
∂ ∂(rj ◦ T −1 ) ∂ ∂xj ∂
= = . (11)
∂y k ∂rk ∂xj ∂y k ∂xj
We say that vectors transform contravariantly.
Using this formula we can compute the change of basis for differential forms,
j
   
k ∂ k ∂(r ◦ T ) ∂
dy = dy
∂xi ∂ri ∂xj
∂(rj ◦ T ) k ∂(rj ◦ T ) k
 
∂
= dy = δj
∂ri ∂xj ∂ri
∂(rk ◦ T )
=
∂ri
This means that,
∂y k j
dy k = dx . (12)
∂xj
The converse is computed in a similar fashion,
∂xk j
dxk = dy . (13)
∂y j
We say that vectors transform covariantly.

1.6.1 Linear Coordinate Transforms

If we assume that T : Rm → Rm is linear then the coordinates transform described in equation
7. Let (U, ϕ) be a coordinate neighborhood with coordinates (x1 , · · · , xm ) and let e1 , · · · em be
the standard basis for Rm . If T : Rm → Rm is a linear transform then

y j = rj ◦ T ◦ ϕ = rj ◦ T (ei )(ri ◦ ϕ) = Tij xi . (14)

1.7 Implicit Function Theorem 13

Now we compute the coordinate transform matrix,

∂y k ∂(rk ◦ T −1 ) −1 k


= = T j
, (15)
∂xj ∂rj
∂xk ∂(rk ◦ T )
= = Tjk . (16)
∂y j ∂rj

1.7 Implicit Function Theorem

The implicit function theorem is needed in several proofs in the following sections. Wew state
this theorem here without proof.
m m
 i  Theorem. Let F : R → R be a smooth mapping
Theorem 1.7.1. The Inverse Function
on an open set V ⊂ Rm . If det ∂F
∂rj
6= 0 on V . Then for all x0 ∈ V there is a neighborhood
W , x0 ∈ W ⊂ V such that F : W → F (W ) is a diffeomorphism (e.g. F −1 exists and is a
diffeomorphism).

For a proof see [4].

Theorem 1.7.2. The Implicit Function Theorem. Let F : Rn+k → Rk be a smooth function
and assume that F (a) = 0. If the Jacobian Df (a) has rank k then there is an open neighborhood
U ⊂ Rn+k and a smooth diffeomorphism h : Rm+n → Rm+n so that F ◦ h : U → Rk so that,

(F ◦ h) x1 , x2 , · · · xm+n = xn+1 , · · · xn+m .

This follows readily from Theorem 1.7.1. For a proof see [4].
An immediate consequence of this is that if F has maximal rank at the point a and F (a) = 0
then n o
(F ◦ h)−1 (0) ∩ U = (x1 , · · · , xn ) (x1 , · · · , xn , xn+1 , · · · xn+k ) ∈ U


2 Frobenius Theorem And Vector Fields

In this section we shall discuss Frobenius thoerem using the language of vector fields. All the
details of this section are explained nicely in [6].
Every smooth vector field generates a family of smooth curves that have tangent vector X.
We would like to examine whether this has an analogue in higher dimensions. For example,
given two smooth vector fields X, Y on a manifold M , is there a 2−dimensional submanifold N
that has span{X, Y } as its tangent space. Of course we must have X and Y linear independent
at every point so that the span is actually 2−dimensional. Are there any other conditions? In
Frobenius theorem we shall see that there is one condition, along with linearly independence,
that guarantees the existence of this submanifold. In this section we discuss the theorem using
14 2 FROBENIUS THEOREM AND VECTOR FIELDS

the language of vector fields and in section 3 we present Frobenius’ theorem using the language
of differential forms and exterior differential analysis.
But we state Frobenius theorem we start with a few simple examples. Let M = R3 and
define two vector fields on the manifold M , which we write write in terms of the coordinate
basis and as standard R3 vectors.
 
1
∂ ∂
X(x, y, z) = +x = 0 (17)
∂x ∂z
x
 
0
∂ ∂
Y (x, y, z) = +y = 1 (18)
∂y ∂z
y

Fix (x, y, z) ∈ R3 . Can we find a 2−dimensional surface N that has tangent spaces formed
by the basis X, Y . If γ is a curve with tangent vector X and it has ϕ(γ(0)) = (x, y, z) then it
must satisfy,

d(x ◦ γ)
= 1,
dt
d(y ◦ γ)
= 0,
dt
d(z ◦ γ)
= x.
dt
Let (x, y, z) ∈ R3 be any point in M and let γ(0) = (x, y, z) then we can solve for γ uniquely.
We get the following solution as a 1−parameter group of diffeomorphisms,
 
x+t
ϕt (x, y, z) =  y  (19)
1 2
z + xt + 2 t

Similarly we can solve Y globally and get the 1−parameter group given by η,
 
x
ηs (x, y, z) =  s+y  (20)
1 2
z + ys + 2 s

So, around the point (x, y, z) we can construct a 2−dimensional manifold given by
 
t+x
Σ(t, s) =  s+y 
1 2 1 2
z + xt + ys + 2 t + 2 s
 15

This parameterization is the inverse of a coordinate neighborhood. In this case we can weave
the two 1-dimensional integral curves for X and Y together to form a smooth 2-dimensional
manifold.
To see what can go wrong we look at another simple 2−dimensional example, Let M = R3
and construct two smooth vector fields,
∂ ∂
X(x, y, z) = +y (21)
∂x ∂z
∂ ∂
Y (x, y, z) = −x . (22)
∂y ∂z
We have coordinates xj with j = 1, 2, 3 where x1 = x, x2 = y, x3 = z.

Each of these smooth vector fields generates a local 1-parameter group of diffeomorphisms
ϕ : I × M → M where I ⊂ R1 . The curves through the point x ∈ M is t → ϕt (x). For the
vector field X the 1−parameter group of diffeomorphisms satisfy,
d(x ◦ ϕ))
= X(x) = X 1 (x, y, z) = 1,
dt
d(y ◦ ϕ)
= X(y) = X 2 (x, y, z) = 0,
dt
d(z ◦ ϕ)
= X(z) = X 3 (x, y, z) = y.
dt
From these conditions we can write down the solution,
 
x+t
ϕt (x, y, z) =  y  . (23)
z + yt
For the vector field Y let η : I × M → M be the 1 =parameter group of diffeomorphisms
then η satisfies,
d(x ◦ η))
= Y (x) = Y 1 (x, y, z) = 0,
dt
d(y ◦ η)
= Y (y) = Y 2 (x, y, z) = 1,
dt
d(z ◦ η)
= Y (z) = Y 3 (x, y, z) = −x.
dt
From these conditions we can write down the solution,
 
x
ηs (x, y, z) =  y + s  . (24)
z − xs
16 2 FROBENIUS THEOREM AND VECTOR FIELDS

Do these two 1−dimensional solutions combine to form a 2−dimensional surface that has a
tangent space that is spanned by the vector fields X and Y . We start at a point (x, y, z) ∈ M
and progress along a rectangle of coordinates. First we proceed along X integral curves by t
and then along Y by s,
       
x x+t x+t x
ϕt ηs ϕ −t
y  −→  y  −→  y+s  −− → y+s  (25)
z z + yt (z + yt) − (x + t)s z − xs − 2ts
We compare this when we proceed along Y by s and then X by t.
   
x x
ηs
y  −
→  y + s . (26)
z z − xs
A comparison of equation 25 and equation 26 we see that the x− and y−coordinates are
the same however the z−coordinate is different. In this case there is no integral submanifold.

2.1 Distributions
Let M be a differentiable manifold of dimension m and let N ⊂ M be a p−dimensional
submanifold. The tangent space Tx (N ) is a p−dimensional subspace of Tx (M ). The two
examples discussed earlier both describe two vector fields X and Y that span p−dimensional
subspace (where p = 2) of the original manifold M = R3 . These p−planes are an important
part of the our discussion so we define the following.

Definition 2.1.1. Let M be a manifold of dimension m and let p be a positive integer p ≤ m.

A p−dimensional distribution D is a choice of a p−dimensional subspace of Tx (M ) for every
x ∈ M . We denote the plane at x by D(x). The distribution D is smooth if for every x ∈ M
there is a neighborhood x ∈ U ⊂ M and smooth vector fields X1 , · · · Xp defined on M so that
D(x) is spanned by X1 (x), · · · , Xp (x) for every x ∈ U .

Both of the examples discussed above generate smooth 2−dimensional distributions but
only one of these distributions was the tangent bundle of a sub-manifold. The 2−dimensional
manifold is called an integral manifold of the 2−dimensional distribution.

Definition 2.1.2. A sub-manifold N ⊂ M with i : N → M is an integral manifold of the

distribution D if,
di(Ty (N )) = D(i(y)),
for each y ∈ N .

This definition just states that the tangent space of N at y corresponds to the distribution
D(y). The key property that insures the existence of integral sub-manifolds is the following.
2.1 Distributions 17

Definition 2.1.3. A smooth distribution D is called involutive if [X, Y ] ∈ D whenever X and

Y are smooth vector fields that lie in D. This means that if X1 , · · · Xp are smooth vector fields
that span D in a neighborhood U then,
k
[Xi , Xj ] = Cij Xk .

Before we get to the main result we need one preliminary technical proposition. The
coordinates described in the following proposition are called flowbox coordinates.

Proposition 2.1.4. Let M be a manifold of dimension m and let X be a smooth vector field
on M . If x ∈ M has X(x) 6= 0 then there is a coordinate neighborhood U ⊂ M of x with
coordinates x1 , · · · xm such that,
∂
X|U = ,
∂x1
Proof. Let x0 ∈ M and find a coordinate system (U, φ) with coordinates y 1 , · · · y m such that
x0 ∈ U and 
∂ 
Xx0 =
∂y 1 x0

We simplify the coordinate system by translating so that y k (x0 ) = 0. In, perhaps, a smaller,
neighborhood U of x0 we have a 1−parameter family of diffeomorphisms ϕt : U → ϕt (U ) ⊂ M .
We look at the surface y 1 = 0 which includes x0 and move along the integral curves of X to
form a coordinate system. To do this we define a mapping ψ : V ⊂ Rm → U by

ψ(t, a2 , a3 , · · · am ) = ϕt (φ−1 (0, a2 , · · · , am )).

First we note that ψ(0, · · · , 0) = x0 . At a2 = a3 = · · · = am = 0 we have

ψ(t, 0, · · · , 0) = ϕt (φ−1 (0, 0, · · · , 0)) = ϕt (x0 ).

This is a curve along the curve X and the derivative at t = 0 is just Xx0 . At t = 0 we also have

ψ(0, a2 , a3 , · · · am ) = φ−1 (0, a2 , · · · , am ).

This means that at t = 0 the curves along the coordinates ak have tangent vector ∂x∂ k . So, at the
point (0, · · · , 0) the Jacobian of ψ takes the m−linearly independent vectors e1 , · · · em to linearly
independent vectors Xx0 , ∂x∂ 2 x , · · · ∂x∂m x0 . This mapping has an invertible Jacobian at x0 and
0
by the inverse mapping theorem the mapping ψ is a diffeomorphism in some neighborhood of
x0 . The inverse of this diffeomorphism is a coordinate mapping with the desired properties.

In Theorem 2.2.1 we shall prove that there is an integral sub-manifold of a distribution if

and only if the distribution is involutive.
18 2 FROBENIUS THEOREM AND VECTOR FIELDS

2.2 Proof of Frobenius Theorem

The main result is the following.

Theorem 2.2.1. Let M be a manifold of dimension m. Let D be a p−dimensional C ∞ dis-

tribution. If D is involutive then for every x ∈ M there exists an integral sub-manifold N that
contains x. In fact, there is a coordinate neighborhood U centered at x such that the slices,

xi = Ki , for all i ∈ p + 1, · · · N,

are integral sub-manifolds of D. If N is a connected integral sub-manifold of U then i(N ) ∩ U

corresponds to one of these slices.

Proof. Let N is a connected integral sub-manifold with di(N )|x = D(i(x)). For every x ∈ N
there is a neighborhood U of M with coordinates x1 , · · · xm so that N is given by xp+1 = · · · =
xm = 0. If the Xi are tangent to N then
p
∂
Xkj
X
Xk = ,
∂xj
j=1

where Xkj is a p × p invertible matrix. Let Y = X −1 . Now compute the bracket,

p X
p
∂ i ∂
[Xkj
X
[Xk , Xm ] = j
, Xm ]
∂x ∂xi
j=1 i=1
p X p  i i
∂Xm j ∂Xk ∂
Xkj
X
= − Xm j
∂xj ∂x ∂xi
j=1 i=1
p X p  i i
∂Xm j ∂Xk
Xkj
X
= − Xm j Yil Xl .
∂xj ∂x
j=1 i=1

This means that the distribution is involutive.

We will prove the converse by induction. Let p = 1 then we have a single smooth vector
field X1 that spans the 1−dimensional distribution D(x) at every x ∈ M . The integral curve
of X is the 1−dimensional manifold through x that has tangent space equal to the span of X.
This proves the theorem for p = 1. In this case [X, X] = 0 so the involutive assumption is
superfluous.

Assume the theorem is true for p − 1 and we shall show it is true for p. Let X1 , · · · Xp
be smooth vector fields that span the distribution D in a coordinate neighborhood of x ∈ M .
2.2 Proof of Frobenius Theorem 19

By Proposition 2.1.4 we can take a neighborhood U1 of x contained in U with coordinates

y 1 , · · · , y m such that,

∂ 
X1 (x) = .
∂x1 x

We shall form a p−1 distribution on the m−1 dimensional manifold formed by setting y 1 = K 1 .
Let W ⊂ U1 be the m − 1 sub-manifold determined by the condition y 1 = K 1 where K 1 is a
constant. For every x ∈ W the distribution is determined by X1 and X2 , · · · Xp . We know that
X1 does not lie in W so we use this to find projections of Xk along W as follows. Define Yk for
k = 2, · · · p by,
Yk = Xk − Xk (y 1 )X1 . (27)

We also define Y1 = X1 so that

D(x) = span{X1 (x), · · · , Xp (x)} = span{Y1 (x), · · · , Yp (x)}.

Notice, however, that Y2 , · · · Yk are tangent to the sub-manifold determined by y 1 = K 1 for

reasonable choices of K 1 . To see this we can show that Yk we compute the derivative of y 1 in
the direction Yk ,
Yk (y 1 ) = Xk (y 1 ) − Xk (y 1 )X1 (y 1 ) = 0. (28)

The sub-manifold W is one leaf of a foliation and we can define a mapping π : U1 → W by,

π(y 1 (x), · · · , y m (x)) = (y 2 (x), · · · , y m (x)).

This mapping is smooth. On the sub-manifold W we define vector fields Zk on W for k = 2, · · · p

by,
Zk (w) = Yk (w).

We define a distribution on W by,

E(w) = span < Z2 , · · · Zp > .

The Zk span a distribution that is the projection of D onto Tx (W ) for every w ∈ W . To use
the induction hypothesis we would like to show that this distribution is involutive. To show
this we must compute the commutator of Zk . Let f : W → R be a smooth function and extend
f to a smooth function f˜ : U1 → R by,

f˜(y 1 , · · · y m ) = f (y 2 , · · · y m ).
20 2 FROBENIUS THEOREM AND VECTOR FIELDS

We compute the commutator for i, j = 2, · · · m,

[Zi , Zj ](f ) = [Yi , Yj ](f˜)

m
X
1
= Cij Y1 (f˜) + k
Cij Yk (f˜)
k=2
m
X
= k
Cij Yk (f˜)
k=2
Xm
k
= Cij Zk (f )
k=2

This means that [Zi , Zj ]w ∈ E(w) for all w ∈ W and the distribution E is involutive. By
induction there is a p − 1 dimensional sub-manifold S ⊂ W with,

Tw (S) = E(w),

for all w ∈ W . By the induction hypothesis for any w ∈ W there is a coordinate system
w2 , · · · wm on W so that S is the sub-manifold determined by the condition,

S = w ∈ W | wp+1 (w) = K p+1 , · · · wm (w) = K m .



By varying K j we form a foliation in W . We meed to extend our p − 1 sub-manifold S ⊂ W

to a p dimensional sub-manifold N ⊂ U1 . To do this we define convenient coordinates xk . For
x ∈ U1 define
x1 (x) = y 1 (x), x2 (x) = w2 (π(x)), · · · xm (x) = wm (π(x)).
We can easily extend S along x1 = y 1 by defining,

N = x ∈ U1 | (x2 (x), · · · xp (x)) ∈ W and xp+1 = · · · xm = 0



This is clearly a smooth sub-manifold of U1 but it is not obvious that the vectors Y1 , · · · Yp are
tangent to N . Clear Y1 is tangent to N as any curve that starts in N and is tangent to Y1
remains in N . To show that Y1 is always tangent to N note that,
∂
Y1 (xp+r ) = xp+r = 0, for r = 1, · · · m − p.


1
(29)
∂x
The vectors Y2 , · · · Yp are more problematic. We must show that that for all k = 2, · · · , m we
have,
Yk (xp+r ) = 0, for r = 1, · · · m − p.
We know that this holds true on the sub-manifold W as,

Yk (xp+r )x1 (x)=0 = Zk (wp+r )x1 (x)=0 = 0.

 
2.2 Proof of Frobenius Theorem 21

To show that this is true for general x compute the derivative in the x1 direction,
∂
Yk (xp+r ) = Y1 Yk (xp+r )



∂x 1

= Y1 Yk (xp+r ) − Yk Y1 (xp+r )

= [Y1 , Yk ](xp+r )
m
j
X
1
= C1k Y1 (xp+r ) + C1k Yj (xp+r )
j=2
m
j
X
= C1k Yj (xp+r )
j=2

where we have used that Y1 , · · · Yp are involutive and we have used equation 29 again. We get
p − 1 differential equations for the p − 1 functions Yk (xp+r ),
m
∂ p+r

X j
Yk (x ) = C1k Yj (xp+r ).
∂x1
j=2

The system is a homogeneous linear system that must have a unique solution. However, we
know that the p − 1 functions Yk (xp+r ) vanish at x1 = 0 so they must vanish for all x1 for
r = 1, · · · m − p since that is a valid solution and it is unique. We conclude conclude that for
every k we have,
Yk (xp+r ) = 0, for r = 1, · · · , m − p.
This means we can write each Yk in terms of the basis,
m
∂
Ykj
X
Yk = .
∂xj
j=1

Apply this derivation to xp+r and we get,

Yk (xp+r ) = 0 = Ykp+r .
This means that we can write,
p
∂
Ykj
X
Yk = .
∂xj
j=1
By equation 27 this means we can write,
p
∂
Xkj
X
Xk = .
∂xj
j=1

The manifold N is formed by the equations xp+1 = · · · = xm = 0. This means that D(x) =
Tx (N ) for all x ∈ U1 . This proves the case for p and the theorem follows by induction.
22 3 FROBENIUS THEOREM AND DIFFERENTIAL FORMS

Frobenius theorem insures that for every x ∈ M there is a submanifold N ⊂ M so that

x ∈ N and for every y ∈ N the tangent space Ty (N ) coincides with the distribution through
y ∈ M . This slicing is called a foliation.

3 Frobenius Theorem And Differential Forms

In this section we discuss Frobenius theorem using the language of differential forms. For a
more complete discussion see [2].

3.1 Introduction
We start with a simple example. Let M = R3 and define,

f (x, y, z) = x2 + x2 + z 2 − r2 .

The implicit function theorem says that

f −1 (0) = (x, y, z)| x2 + y 2 + z 2 = r2 = 0 = S 2 (r)



is a smooth manifold. If we take the differential of f we get,

df = 2xdx + 2ydy + 2zdz.

Let θ = df . The mapping θ : Tx (M ) → R has a two dimensional kernel,

Ex = ker(θ) ≤ Tx (M ).

Let v ∈ Ex and let

∂ ∂ ∂
v = α1 + α2 + α3 .
∂x ∂y ∂z
The condition that v ∈ Ex is
     
1 ∂ 2 ∂ 3 ∂
θ(v) = 2xdx α + 2ydy α + 2zdz α
∂x ∂y ∂z
1 2 3


= 2 xα + yα + zα = 0.

We can write down solutions for this,

       
α1 y 0 z
α2  = −x , −z  ,  0  .
α3 0 y −x
3.2 Frobenius Theorem 23

These three vectors are not linearly independent but, as long as x, y, z are not all 0 then we
can span the plane Ex by two of these three vectors. These vectors span the tangent spaces of
Tx (M ). It turns out, because of the topology of S 2 there are two non-zero vector fields on the
entire surface S 2 .

These three vectors are written as, We write these three vectors in abstract notation as,
∂ ∂ ∂ ∂ ∂ ∂
y − x ,z − y ,z −x .
∂x ∂y ∂y ∂z ∂x ∂z
The kernel spaces of the smooth differential form df form a integrable distribution. We can
find a sub-manifold N (e.g. N = S 2 ) that Tx (N ) = kern(df ). This is the form of Frobenius
theorem that we shall discuss.

This simple example shows the basic structure of Frobenius Theorem. A collection of linear
independent 1−forms θ1 , · · · θk have joint kernels that form a m − k dimensional sub-space of
Tx (M ). Can we find a manifold N that has tangent spaces Tx (N ) that is the joint kernels of
θ1 , · · · , θk . We call this manifold N the integral manifold of the 1−forms θ1 , · · · θk . In our first
example we found a manifold, N = S 2 , which has tangent spaces which are equal to ker(θ).
This is not true for every collection of forms. The 1−form above is closed, but this is not
required, as our next example demonstrates.
Example 3.1.1. Define θ = f (x, y, z)dx where f (x, y, z) > 0 for all (x, y, z). In this case we
have
∂f ∂f
dθ = dy ∧ dx + dz ∧ dx.
∂y ∂z
We can write down the kernel,
  
∂ ∂ 
ker(θ) = a + b  a, b ∈ R .
∂y ∂z
There are integral manifolds which consist of planes perpendicular to the x−axis.

3.2 Frobenius Theorem

Frobenius Theorem described the conditions required to create a smooth submanifold using
the kernels of a collection of 1−forms. We start with a statement of the theorem.
Theorem 3.2.1. Frobenius Theorem. Let M be a smooth manifold of dimension m. Let
θ1 , θ2 , · · · θm−p ∈ Λ1 (M ) be smooth pointwise linearly indepdendent forms. If there exist 1−forms
αji ∈ Λ1 (M ) such that,
m−p
X
a
dθ = αba ∧ θb , (30)
b=1
24 3 FROBENIUS THEOREM AND DIFFERENTIAL FORMS

for all a = 1, · · · m − p, then for any x ∈ M there exists a unique k−dimensional manifold

i : N ,→ M,

such that x ∈ N and such that i∗ (θa ) = 0 for all a = 1, · · · m − k. Further, for this x ∈ M there
is a coordinate neighborhood U ⊂ M with coordinates (x1 , · · · xm ) so that N has coordinates
(x1 , · · · xp ) and
Xm
a
θ = Aab dxb ,
b=p+1

so that θa are generated by dxa for a = p + 1, · · · m and the joint kernel of the θa corresponds
to the joint kernel of dxa .

Recall that Λk (M ) denotes the vector space of smooth k−forms on a manifold M . Notice
that equation 30 is more general than closed, although a collection of closed forms will suffice.
Later, we shall see that dθa is in the algebraic ideal generated by the forms θa . Before we prove
the theorem we prove some preliminary facts. Then we will prove theorem 3.2.1 for the case of
p = 1.
It is important to note that properties satisfied by the forms θ1 , · · · θm−p in Theorem 3.2.1
are also satisfied by linear combinations of these forms, as long as there are m − p linearly
independent 1−forms. To make this concrete we add the following elementary proposition,

Proposition 3.2.2. Let θ1 , · · · θk ∈ Λ1 (M ) be 1−forms that satisfy the criteria in Theorem

3.2.1. If Aab is a k × k invertible matrix with smooth entries then the forms η a satisfy the
criteria in Theorem 3.2.1 where η a are defined by,
k
X
a
η = Aab θb for a = 1, · · · k.
b=1

Proof. The forms η a are linearly independent because Aab are linearly independent. We take
the differential,

dη a = d(Aab ) ∧ θb + Aab dθb

= d(Aab ) ∧ (A−1 )bc θc + Aab αdb ∧ θd
= d(Aab ) ∧ (A−1 )bc η c + Aab αdb ∧ (A−1 )dc η c
 
= d(Aab )(A−1 )bc + Aab αdb (A−1 )dc ∧ η c
= βca ∧ η c

where βca ∈ Λ1 (M ).
3.2 Frobenius Theorem 25

Another important fact is that the kernels of the collection of differential forms form smooth
distributions. In the language of bundles, we will show that the kernels form a sub-bundle of
the tangent bundle.
Theorem 3.2.3. Let M be a smooth manifold of dimension m. Let θ1 , θ2 , · · · θm−p ∈ Λ1 (M ) be
point-wise linearly independent forms. The set of vectors X ∈ Tx (M ) satisfying θb (X) = 0 for
b = 1, · · · m−p form a m+p dimensional differentiable manifold. In fact, it is a p−dimensional
vector bundle over M .
Proof. We set k = m − p so that the forms θ1 , θ2 , · · · θk are linearly independent. Given any
x0 ∈ M we shall construct a coordinate neighborhood U so that the vectors in the perpendicular
space is diffeomorphic to U ×Rm−k . Let (U1 , ϕ) be a coordinate neighborhood with coordinates
(y 1 , · · · y m ). We can write each θa as,

θa = θ1a dy 1 + · · · + θm
a
dy m .

We write down the matrix, which depends on the value x ∈ U ,

 1
θ21 · · · θm1

θ1
θ 2 θ22 · · · θm2 
 1
T (x) =  .

.. .. .. 
 .. . . . 
θ1k θ2k · · · k
θm
For our fixed point x0 ∈ M we use Gaussian elimination to change T (x0 ). There is a k × k
invertible matrix R so that left-multiplication change T (x0 ) into an upper diagonal matrix of
the form,  
θ̃11 θ̃21 θ̃31 θ̃41 · · · 1
θ̃k−1 θ̃k1 ··· θ̃m1
 0 θ̃2 θ̃2 θ̃42 · · · 2
θ̃k−1 θ̃k2 ··· θ̃m2 
 2 3 
 0 0 θ̃33 θ̃43 · · · 3
θ̃k−1 θ̃k3 ··· 3
θ̃m 
 
T̃ (x0 ) = RT (x0 ) = 
 .. .. .. .. . . .. .. .. .. 
. . . . . . . . . 

k−1
0 ··· θ̃kk−1 · · · k 

0 0 0 θ̃k−1 θ̃m 
0 0 0 0 ··· 0 θ̃kk · · · θ̃mk

where θ̃jj = 1 for j = 1, · · · k. We use this linear transform R to change coordinate systems as
described in Section 1.6.1. We get a new coordinates (y 1 , · · · y m ) so that θ̃ are coordinates of
θa in the y j coordinates. We get the form above at the point x0 . There is a neighborhood V
of x0 ∈ V such that θjj (y) > 0.5 and all the 00 s in the sub-diagonal are < 0.5. This insures
that all the vectors are linearly independent. The new coordinates (y 1 , · · · y m ) are defined on,
a possible open subset, of V . In these coordinates, given a vector
X ∂
Y = Yj j.
∂y
j
26 3 FROBENIUS THEOREM AND DIFFERENTIAL FORMS

This means that Y ∈ KT if,

X
θ̃ja Y j = 0,
j

for all a. If a = 1 then we have,

1  1 2 1 m

Y1 = − θ̃ 2 Y + · · · + θ̃ m Y .
θ̃11

We proceed in this way to determine values for a = 2, 3, · · · k,

1  
Y2 = − θ̃12 Y 1 + θ̃32 Y 3 + · · · + θ̃m
1
Ym
θ̃22
1  
Y3 = − 3 θ̃12 Y 1 + θ̃22 Y 2 + θ̃43 Y 4 + · · · + θ̃m
1
Ym
θ̃3
..
.
1  
Yk = − k θ̃12 Y 1 + · · · + θ̃k−1
k
Y k−1 + θ̃k+1
k
Y k+1 + · · · + θ̃m
1
Ym
θ̃k

Since k = m − p we have coordinates y j that form a bundle coordinate system with coordi-
nates, (y 1 , · · · y m , Y m−p+1 , Y m−p+2 , · · · Y m ). All the remaining coordinates Y 1 , · · · , Y m−p are
determined. This neighborhood has the topology of Ũ × Rp .

Proposition 3.2.4. Let M be a smooth manifold of dimension m. Let θ1 , θ2 , · · · θm−1 ∈ Λ1 (M )

be pointwise linearly indepdendent forms. Then for all x ∈ M there is a coordinate neighborhood
x ∈ U ⊂ M with coordinates (x1 , · · · xm ) that satisfy
m
X
θa = Aab dxb .
b=2

Proof. Given forms θ1 , · · · θm−1 then the perpendicular space is just a line. Theorem 3.2.3 says
that this is spanned by a smooth vector field. This smooth vector field has integral curves
which are then the integral submanifolds of the distribution.

To get the coordinates xj we use the Flow box coordinates that we discussed in Proposition
2.1.4. Let X be a smooth vector field that is in the kernel of the θa . We pick flow box
coordinates with ∂x∂ 1 in the direction of X so it is in the kernel of θa . This means that,

∂
θa ( ) = 0, for a = 1, · · · m − 1.
∂x1
3.2 Frobenius Theorem 27

Since dx1 , · · · dxm span T ∗ (M ) we have,

m
X m
X
θa = Aaj dxj = Aab dxb ,
j=1 b=2

where Aab is an (m − 1) × (m − 1) dimensional invertible matrix and Aab is a smooth function

of M .

Proposition 3.2.5. Frobenius theorem 3.2.1 is true for any p.

Proof. We shall prove the theorem by induction on p. The case for p = 1 was proved in
Proposition 3.2.4. Now assume the theorem is true for p − 1 we will show it is true for p. Let
θ1 , · · · θm−p be linearly independent 1−forms that satisfy,
m−p
X
a
dθ = αba ∧ θb , for a = 1, · · · m − p,
b=1

where αba ∈ Λ1 (M ). We will add another form to this collection and show that the new ex-
panded collection still satisfies the condition of equation 30. Since the theorem is true for p − 1
this addition collection has m − (p − 1) = m − p + 1 forms and the induction hypothesis is
satisfied. We start by adding a 1−form to our collection of one forms.

Claim 3.2.6. If p > 0 then there is a function f : M → R with df, θ1 , · · · θm−p are linearly
independent.

Let U ⊂ M be a coordinate neighborhood with coordinates (y 1 , · · · y m ). Since the forms

θ1 , · · · θm−p cannot span the entire span of forms dy 1 , · · · dy m−p+1 there must be some for dy k
that is linearly independent of the θa . This means that for every x ∈ U there is a k so that one
can set f = y k .

With f defined as in the claim we have a collection of m − p + 1 = m − (p − 1) forms given

by,
θ1 , · · · θm−p , df. (31)

We are assuming the theorem is true for p − 1 so we must show the forms 31 satisfy the criteria
of the theorem. These forms clearly satisfy condition 30 since d(df ) = 0 and, by assumption,

m−p
X
a
dθ = αba ∧ θb ,
b=1
28 3 FROBENIUS THEOREM AND DIFFERENTIAL FORMS

for a = 1, · · · m − 1. Therefore, by Frobenius theorem for p − 1 we know that for any x0 ∈ M

there is a neighborhood V and coordinates (y 1 , · · · y m ) that satisfy x0 ∈ V and for all for all
y ∈V,
m
X
a
θ (y) = Bba (y)dy b for a = 1, · · · m − p,
b=p
Xm
df (y) = Cb (y)dy b
b=p

∂
Notice that the vector fields ∂y 1
· · · ∂y∂p−1 are in

ker(θ1 ) ∩ · · · ker(θm−p ) ≤ ker(θ1 ) ∩ · · · ∩ ker(θm−p ) ∩ ker(df ).

To prove the theorem we need to add an additional coordinate to this kernel space.

The forms θa and df are non-zero so Cb (x0 ) 6= 0 for at least one b and we can find, a possible
sub, neighborhood of V such that Cb (x) 6= 0 for all x ∈ V . We see that we can assume, without
a loss of generality, that Cp (x) 6= 0 for all x ∈ V . We can then write,
 
  X m
1 1 
dy p = df − Cb dy b 
Cp Cp
b=p+1

We use this formula to eliminate the form dy p from our collection and replace it with the form
df . We get the following set of equations,
m
X
a
θ = Dba dy b + da df for a = 1, · · · m − p,
b=p+1

df = df

where da are smooth functions of Cp and Bpa and x. Notice this is the first step in Gaussian
elimination. The matrix elements Dba are functions of the remaining Cb and Bba . Notice that
Dba is now a (m − p) × (m − p) matrix.

The m−p+1 forms df, dy p+1 , · · · , dy m span the same sub-space of T ∗ (M ) as θ1 , · · · θm−p , df .
We write this in matrix form as,
 a  a a  b
θ Db d dy
= . (32)
df 0 1 df
3.2 Frobenius Theorem 29

This relationship is invertible so det(Dba (x)) 6= 0 for all x ∈ V . We can write down the inverse
and using this inverse we multiply both sides of 32,
 a   −1 a
(D )b −(D−1 )ab db θb
    −1 a a
(D )b θ − (D−1 )ab db df

  a
η + ea df

dy
= = =
df 0 1 df df df

We have defined η a and ea by,

η a (y) = (D−1 )ab (y)θb and ea (y) = (D−1 )ab db .

Using these definitions we have,

dy a
  a
η + ea df
 
=
df df
There is an m − p + 1 dimensional subspace of T ∗ (M ) that has a basis given by the forms
{df, dy p+1 , · · · , dy m }. There are also basis given by {df, θ1 , · · · θm−p } and {η p+1 , · · · η m , df }.
We can span the sub-space of T ∗ (M ) using the basis {df, dy a } or the basis {df, η a }. From
the matrix equation we get,

η a = dy a − ea df, where a = p + 1, · · · m. (33)

Claim 3.2.7. The 1−forms η a satisfy equation 30 in that,

m
X
a
dη = Eba η b for a = p + 1, · · · m.
b=p+1

This is simply Theorem 3.2.2.

Claim 3.2.8. The functions ea (y) are function of only (y p , · · · , y m ).
From the computation in Claim 3.2.7 we have the following,
m
X
dη a = Eba ∧ η b ,
b=p+1

But we also have equation 33,

η a = dy a − ea df, where a = p + 1, · · · m.

We take the differential d of this equation and write everything in terms of the basis {η a , df }
and equate this with above,
dη a = −dea ∧ df = γba ∧ η b
30 3 FROBENIUS THEOREM AND DIFFERENTIAL FORMS

The means that all the terms in dea ∧ df must contain a term of the form β ∧ η a . From this we
deduce that (note that df ∧ df = 0),

m
X
dea = Cba η b + g a df, where a = p + 1, · · · , m.
b=p+1

We can now write this in terms of the dy a and we have,

m m
X X ∂ea X ∂ea b
dea = Fba dy b = b
dy b = dy , for a = p + 1, · · · m.
∂y ∂y b
b b=1 b=p+1

This means that,

∂ea
= 0, for a = p + 1, · · · m, j = 1, · · · p − 1.
∂y j
This concludes claim 3.2.8.

The forms η a satisfy Proposition 3.2.5 on the m − p + 1 dimensional manifold generated

by the coordinates (y p , · · · y m ). Using this proposition we can find coordinates (xp , · · · xm ) so
that,
Xm
ηa = ηba dxb for a = p + 1, · · · , m.
b=p+1

We also have ∂x∂ p ∈ ker(η p+1 ) ∩ · · · ∩ ker(η m ). Notice that these expressions are true for
all (y 1 , · · · , y p−1 ). We define a coordinate system by xj = y j for j = 1, · · · p and xp+1 =
y p+1 , · · · , xm = y m . The forms η p+1 , · · · , η m are linear combinations of the forms θ1 , · · · , θm−p .
This means that our original forms are linear combinations of the dxp+1 , · · · dxm ,
m
X
θa = θba dxb for a = p + 1, · · · m
b=p+1

The transverse manifold has coordinates (x1 , · · · xp ) and is determined by the conditions xp+1 =
K p+1 , · · · , xm = K m . Thus we have proved the theorem for p assuming p−1 and this concludes
the theorem.

Remark 3.2.9. We have two separate proofs of Frobenius theorem. The first uses vector fields
and the second uses differential forms. One can also prove the two forms of Frobenius theorem
are equivalent using a generalization of the formula in equation 6.
3.2 Frobenius Theorem 31

Example 3.2.10. Frobenius theorem has a close relationship to the implicit function theorem
1.7.2. Let G : M → Rk so that,  1 
g (p)
 g 2 (p) 
G(p) = 
 
.. 
 . 
g m−k (p)
We form the closed forms dg j for j = 1, · · · m − k. The intersection of the kernel planes is,

Ep = ker(dg 1 ) ∩ ker(dg 2 ) ∩ · · · ∩ ker(dg m−k )

The conditions Frobenius are that dg j are linearly indepdendent which is the same as the rank
condition of the implicit function theorem. We look at the Jacobian,
∂g 1 ∂g 1 ∂g 1
 
1 2 · · · ∂x m
 ∂x
∂g 2
∂x2
∂g ∂g 2 
 ∂x 1 ∂x 2 · · · ∂xm 

Dg(p) =  .
 . .
 .. .. ··· .. 
∂g m−k ∂g m−k ∂g m−k
∂x1 ∂x2
··· ∂xm

If the dg j are linearly independent then the Jacobian

Dg has rank m − k. The other condition
j j
is trivial since the forms dg are closed as d dg = 0. Frobenius then says that there is a
sub-manifold i : N → M such that i∗ (dg j ) = 0 for all j so that,

dg j (di(X)) = dg j (X) = X(g j ) = 0,

for all j = 1, · · · m − k so that X ∈ ker(dg j ) for all j. This means that for all z ∈ N we have

Tz (N ) ≤ Ep .

We see that the implicit function theorem is really a special case of Frobenius Theorem.

Example 3.2.11. There are systems that are not Frobenius. Let M = R3 and define a 1−form
θ by
θ = dz − ydx + xdy.
We compute the differential,

dθ = −dy ∧ dx + dx ∧ dy = 2dx ∧ dy.

We see that dθ 6= η ∧ θ so the distribution is not integrable. In fact, there are no sub-manifolds
of R3 with these planes as tangent planes.
32 3 FROBENIUS THEOREM AND DIFFERENTIAL FORMS

Example 3.2.12. Let M = R3 and define a 1−form by,

θ = θ1 dx + θ2 dy + θ3 dz. (34)

We want to find general conditions where θ satisfies Frobenius Theorem 3.2.1. The kernels of
this 1−form are given by vectors

∂ ∂ ∂
X = X1 + X2 + X3 ,
∂x ∂y ∂z

that have the property that θi X i = 0. Is this system of planes integrable? We compute dθ.
     
∂θ2 ∂θ1 ∂θ1 ∂θ3 ∂θ3 ∂θ2
dθ = − dx ∧ dy + − dz ∧ dx + − dy ∧ dz.
∂x ∂y ∂z ∂x ∂y ∂z

When is dθ = α ∧ θ for some α ∈ Λ1 (M ). If we extend θ to a full basis of T ∗ (M ) using forms

ω2 , ω3 then we can write,

dθ = p1 (ω1 ∧ ω2 ) + p2 (ω3 ∧ θ) + p3 (θ ∧ ω3 ).

Now we see that the condition that dθ = α ∧ θ is equivalent to the condition that dθ ∧ θ = 0.
So the distribution of equation 34 is integrable if and only if
     
∂θ2 ∂θ1 ∂θ1 ∂θ3 ∂θ3 ∂θ2
dθ ∧ θ = − θ3 + − θ2 + − θ1 = 0.
∂x ∂y ∂z ∂x ∂y ∂z

If the form θ satisfied Frobenius Theorem (3.2.1 ) then we could, locally, find coordinates
(x1 , x2 , x3 ) so that,
θ = A(x)dx3 .

In the classic literature this condition is written, Vθ · curl(Vθ ) = 0.

Our example is immediately generalized to 4 or even higher dimension, although the notation
gets more complicated since the space of 2−forms in R4 is 6 dimensional. We could easily derive
a condition so that 3−planes in R4 are integrable.

3.3 Differential Forms

The set of alternating forms is a vector space and the product operator ∧ turns the alternating
forms into a Graded algebra. We can formally add and multiply forms by real number. A form
θ has order p if θ ∈ Λp .
3.3 Differential Forms 33

Definition 3.3.1. A graded module A is a module along with a grading that assigns sub-modules
Ak to grade k. The module A a direct sum of the graded sub-modules,
M
A= An = A0 ⊕ A1 ⊕ · · ·
n∈N

A graded module A is a graded algebra if there is a product and the product takes Ai Aj ⊂ Ai+j .

Definition 3.3.2. A sub-space I ⊂ Ω∗ (M ) is an algebraic ideal if

1. I is a direct sum of homogeneous sub-spaces I k ⊂ Ωk (M ).

2. I is an ideal under ∧ so that for any ω ∈ I and η ∈ Ω∗ (M ) we have ω ∧ η ∈ I.

This is called a graded ideal of the graded algebra Ω∗ (M ).

Example 3.3.3. Let φ1 , · · · , φs is a collection of homogeneous elements. The algebraic ideal

generated by these forms is
( s  )
X 
(γ i ∧ φi ) γ 1 , · · · , γ s ∈ Ω∗ (M )

1
φ , · · · , φs alg =
 

k=1

All of the above works for alternating forms on a vector space. To add derivatives we add
the differential of a form to the mix. We start with the basic definition.

Definition 3.3.4. An algebraic ideal I is a differential ideal if I is closed under exterior

product, so that dω ∈ I for every ω ∈ I.

Definition 3.3.5. Let M be a manifold M and let D be a distribution. The annihilator of the
distribution is the set of forms θ ∈ Λp (M ) that have the property that for any X1 , · · · Xp vector
fields that are contained in the distribution D then

θ(X1 , X2 , · · · Xp ) = 0.

The forms in the annihilator will annihilate the vectors fields in the distribution.

Notice that if D is a distribution then the collection of annihilators form an ideal in Λ∗ (M ).

This follows since if θ ∈ Λp (M ) annihilates a distribution D and if ω ∈ Λk (M ) then for any
X1 , · · · Xp+k ∈ D we have,
(ω ∧ θ)(X1 , · · · , Xp+k ) = 0.
Frobenius theorem can be stated in our new language. In this language Frobenius theorem
is more closely linked to Cartan-Kahler theorem.
34 3 FROBENIUS THEOREM AND DIFFERENTIAL FORMS

Theorem 3.3.6. Let M be a C ∞ differentiable manifold of dimension m. Let I be an ideal

generated algebraically by linearly independent 1−forms θm−p+1 , · · · , θm . If I is a differential
ideal then through any x ∈ M there is a integral manifold of I. Further, for any x ∈ M ,
there is a neighborhood U with x ∈ U ⊂ M with coordinates (y 1 , · · · y N ) so that I is generated
by dxm−p+1 , · · · dxm . The coordinates x1 , · · · xp are coordinates to a sub-manifold N that any
Y ∈ Tx (N ) has θa (X) = 0 for all a = m − p + 1, · · · m.
Example 3.3.7. Let M = R3 . In this example we model a PDE system where u : M = R3 → R
and the equations are given by,
∂u
= F (x, y, z) (35)
∂x
∂u
= G(x, y, z) (36)
∂y
If γ : R → M is a curve that lies in the solution to the system surface then it has form
   
x0 + at a
γ(t) =  y0 + bt  with dγ (t) =  b 
dt
u(x0 + at, y0 + bt) aux + buy
   
∂ ∂ ∂ ∂
=a +F +b +G
∂x ∂z ∂y ∂z
This two dimensional space is the kernel of the following form,

θ = dz − F (x, y, z)dx − G(x, y, z)dy. (37)

We want to find the differential ideal generated by this single form. We take the differential of
θ,
∂F ∂F ∂G ∂G
dθ = − dy ∧ dx − dz ∧ dx − dx ∧ dy − dz ∧ dy
∂y ∂z ∂x ∂z
 
∂F ∂G ∂F ∂G
= − dx ∧ dy − dz ∧ dx − dz ∧ dy
∂y ∂x ∂z ∂z
Let’s use θ in equation 37 to eliminate dz = θ + F dx + Gdy. We get,
 
∂F ∂G ∂F ∂G
dθ = − dx ∧ dy − (θ + F dx + Gdy) ∧ dx − (θ + F dx + Gdy) ∧ dy
∂y ∂x ∂z ∂z
 
∂F ∂G ∂F ∂G
= − dx ∧ dy − (θ + Gdy) ∧ dx − (θ + F dx+) ∧ dy
∂y ∂x ∂z ∂z
     
∂F ∂G ∂F ∂G ∂F ∂G
= − dx ∧ dy + G −F dx ∧ dy − θ ∧ dx + dy
∂y ∂x ∂z ∂z ∂z ∂z
3.3 Differential Forms 35

We would like to use Frobenius theorem to find solutions of the system. For θ to satisfy the
theorem dθ must be in the algebraic ideal generated by θ. This means that all the dx ∧ dy terms
must vanish. We arrive at the compatibility condition,

∂F ∂G ∂F ∂G
− +G −F = 0. (38)
∂y ∂x ∂z ∂z

This equation is not unexpected. Using equations 35 and 36 we have,

∂F ∂F ∂F ∂F
uxy = + uy = +G
∂y ∂z ∂y ∂z
∂G ∂G ∂G ∂F
uyx = + ux = +F
∂y ∂z ∂y ∂z

The condition that uxy = yxy is equivalent to equation 38.

Example 3.3.8. Let M = R3 and define a PDE system by,

∂y
= F (x, y, z)
∂x
∂z
= G(x, y, z)
∂x
A solution to this system is a curve
 
at  
dγ ∂ ∂ ∂
γ(t) = y(at) so that
  =a +F +G
dt ∂x ∂y ∂z
z(at)

This tangent vector is the kernel of the two forms,

θ1 = dy − F (x, y, z)dx
θ2 = dz − G(x, y, z)dx

We compute,

∂F ∂F
dθ1 = − dy ∧ dx − dz ∧ dx
∂y ∂z
∂F 1
∂F 2

=− θ + F dx ∧ dx − θ + Gdx ∧ dx
∂y ∂z
∂F 1 ∂F 2
=− θ ∧ dx − θ ∧ dx
∂y ∂z
36 3 FROBENIUS THEOREM AND DIFFERENTIAL FORMS

∂G ∂G
dθ2 = − dy ∧ dx − dz ∧ dx
∂y ∂z
∂G 1
∂G 2

=− θ + F dx ∧ dx − θ + Gdx ∧ dx
∂y ∂z
∂G 1 ∂G 2
=− θ ∧ dx − θ ∧ dx
∂y ∂z

Both dθ1 and dθ2 bot satisfy the criteria of Frobenius theorem. This means that there are
integral submanifolds. This example is trivial since we know the smooth vector field,
∂ ∂ ∂
X(x, y, z) = +F +G ,
∂x ∂y ∂z
has integral curves which are solutions to the system.

For more details see [2] and [1].

 37

4 References

References
[1] R.L. Bryant S.S. Chern R.B. Gardner H.L. Goldschmidt and P.A. Griffiths. Exterior Dif-
ferential Systems. Springer-Verlag, 1991.

[2] Thomas Ivey and J.M. Landsberg. Cartan for Beginners. American Mathematical Society,
2003.

[3] John McCleary. Geometry From a Differentiable Viewpoint. Cambridge Univerity Press,
1994.

[4] Michael Spivak. Calculus on Manifolds. W. A. Benjamin, Inc, 1965.

[5] Shlomo Sternberg. Lectures on Differential Geometry. Prentice Hall, Inc, 1964.

[6] Frank Warner. Foundations of Differentiable Manifolds and Lie Groups. Scott Foresman
and Company, 1970.
Index
algebraic ideal, 33
Alt operator, 8
alternating functions, 7

contravariant, 12
coordinate neighborhood, 3
coordinate transform, 10
coordinate transforms
linear, 12
covariant, 12

derivation, 4
differentiable manifold, 3
differential, 5, 10
differential forms, 9
differential ideal, 33
distribution, 16

exterior algebra, 7

flowbox coordinates, 17
foliation, 22

graded algebra, 33
graded ideal, 33

Integral Curve, 6
integral manifold, 16
involutive distribution, 17

Lie Derivative, 7

perpendicular bundle, 25

tangent vector, 4
The Implicit Function Theorem, 13
The Inverse Function Theorem, 13

vector field, 4

wedge product, 8

38