a

r
X
i
v
:
1
4
0
5
.
3
1
1
6
v
1


[
m
a
t
h
.
D
G
]


1
3

M
a
y

2
0
1
4
NOTES ON EXTERIOR DIFFERENTIAL SYSTEMS
ROBERT L. BRYANT
Abstract. These are notes for a very rapid introduction to the basics of
exterior dierential systems and their connection with what is now known
as Lie theory, together with some typical and not-so-typical applications to
illustrate their use.
Contents
1. Introduction 2
1.1. Dierential ideals 2
1.2. Integral manifolds and elements 2
1.3. Polar spaces 3
1.4. Cartans Bound and characters 3
2. Cartan-Kahler Theory 4
2.1. A form of the Cartan-Kahler Theorem 4
2.2. Involutive tableau 5
3. First Examples and Applications 7
3.1. A version of Cartans Third Theorem 9
3.2. Variants of Cartans Third Theorem 12
4. Ordinary prolongation 19
4.1. The tableau of an ordinary element 19
4.2. The ordinary prolongation of I 20
4.3. The higher prolongations 20
4.4. Prolonging Cartan structure equations 21
4.5. Non-ordinary prolongation and the Cartan-Kuranishi Theorem 22
5. Some applications 22
5.1. Surface metrics with |K|
2
= 1 22
5.2. Surface metrics of Hessian type 23
5.3. Prescribed curvature equations for Finsler surfaces 26
5.4. Ricci-gradient metrics in dimension 3 27
5.5. Riemannian 3-manifolds with constant Ricci eigenvalues 30
5.6. Torsion-free H-structures 33
References 38
Date: December 12, 2013.
1991 Mathematics Subject Classication. 58A15 .
Key words and phrases. Exterior dierential systems, integral manifolds, Cartan-Kahler.
These notes were mostly written during a December 2013 workshop Exterior Dierential Sys-
tems and Lie Theory at the Fields Institute in Toronto, my participation in which was supported
by the Clay Mathematics Institute. My thanks go to the CMI as well as to the National Science
Foundation for its support via the grant DMS-1359583.
This is Draft Version 0.3, last modied on May 12, 2014.
1
2 R. BRYANT
1. Introduction
Around the beginning of the 20th century,

Elie Cartan developed a theory of
partial dierential equations that was well-suited for the study of local problems in
dierential geometry. His fundamental insight was that many geometric problems
(roughly speaking, those that were independent of choice of coordinates) could
be recast as problems in which one has a set of functions and dierential forms
satisfying some given set of structure equations, i.e., conditions on the exterior
derivatives of the given functions and forms. The invariance properties of the
exterior derivative then made possible an approach to dierential equations that
Cartan then developed and applied in a very large number of situations, beginning
with his theory of innite groups (what we now call pseudo-groups) and continuing
throughout his later work in classical dierential geometry.
In Cartans formulation, he was mainly concerned with systems that consisted
of functions and 1-forms, and their exterior derivatives. Erich Kahler [10] realized
that Cartans theory could be usefully extended to systems generated by forms of
arbitrary degree. The resulting extension is now known as Cartan-Kahler The-
ory and Cartans general approach is now generally called the theory of exterior
dierential systems, or simply EDS
Cartan and his students used EDS with great success in the cases in which a given
problem could be recast as a system satisfying the hypothesis of involutivity, which
many important systems did. Cartan introduced the process of prolongation as an
algorithm whose purpose is to replace a given exterior dierential system with one
that has essentially the same solutions but is also involutive, the idea being that all
of the solutions of any system should be describable as the solutions of an involutive
system. Cartan was never able to prove that his prolongation process succeeded in
all cases, but, later, Masatake Kuranishi [11] did succeed in proving a version of
the desired prolongation theorem, one that applies in nearly all cases of interest,
thus essentially completing the theory on an important point.
In these notes, I introduce the basics of exterior dierential systems, covering
the essential denitions and theorems, but I do not attempt to discuss the proofs
of fundamental results such as Cartans Bound (aka Cartans Test), the Cartan-
Kahler Theorem, or the Cartan-Kuranishi Theorem. For proofs of these, the reader
can consult any of the standard sources in the subject, such as [1].
My choice of subject matter and examples was heavily inuenced by the audience
of the workshop at which this material was written, and I have made no attempt
to give a more comprehensive view of the standard topics in exterior dierential
systems. Rather, I have focused on applications to Cartan structure equations
associated to dierential geometric problems and tried to show how Cartans theory
is connected with modern-day Lie theory.
1.1. Dierential ideals. Let M
n
be a smooth n-manifold. An exterior dierential
system on M is a graded, dierentially closed ideal I A

(M).
While it is not strictly necessary, it simplies some statements if one assumes
that I is generated in positive degrees, i.e., I
0
= I A
0
(M) = (0), so I will assume
this throughout the notes.
1.2. Integral manifolds and elements. An integral manifold of I is a subman-
ifold f : N M such that f

() = 0 for all I.
NOTES ON EDS 3
Remark 1. In most applications of exterior dierential systems, the integral man-
ifolds of a certain dimension (often the maximal dimension) of a given dierential
ideal I represent the local solutions of some geometric problem that can be ex-
pressed in terms of partial dierential equations. Thus, one is interested in tech-
niques for describing the integral manifolds of a given I.
An integral element of I is a p-plane E Gr
p
(TM) such that

E
() = 0 for
all I. The set of p-dimensional integral elements of I is a closed subset
V
p
(I) Gr
p
(TM). (It is not always a smooth submanifold of this bundle.)
Remark 2. Every tangent plane to an integral manifold of I is an integral element
of I. The fundamental problem of exterior dierential systems is to decide whether,
for a given E V
p
(I), there is an integral manifold of I that has E as one of its
tangent spaces.
1.3. Polar spaces. Fix E V
p
(I), with E T
x
M, and let e
1
, . . . , e
p
be a basis
of E. The polar space (sometimes called the enlargement space) of E is the subspace
H(E) = {v T
x
M | (v, e
1
, . . . , e
p
) = 0 I
p+1
} T
x
M.
From its denition, any E
+
V
p+1
(I) that contains E must be contatined in H(E)
and, conversely, any E
+
Gr
p+1
(TM) that satises E E
+
H(E) satis-
es E
+
V
p+1
(I). Set c(E) = dim
_
T
x
M/H(E)
_
.
While determining the structure of V
p
(I) can be dicult, one sees that the prob-
lem of understanding the (p+1)-dimensional extensions that are integral elements
of a given p-dimensional integral element is essentially a linear one.
1.4. Cartans Bound and characters. Let E V
n
(I) be xed, and let F =
(E
0
, E
1
, . . . , E
n1
) be a ag of subspaces of E, with dimE
i
= i. Thus,
(0)
x
= E
0
E
1
E
n1
E T
x
M.
Note that E
i
belongs to V
i
(I). The following result is due to Cartan and Kahler.
Proposition 1 (Cartans Bound). Given E V
n
(I) and a ag F = (E
i
) in E
as above, there is an open E-neighborhood U Gr
n
(TM) such that V
n
(I) U is
contained in a smooth submanifold of U of codimension
c(F) = c(E
0
) +c(E
1
) + +c(E
n1
).
If V
n
(I) near E actually is a smooth submanifold of Gr
n
(TM) of codimen-
sion c(F), then E is said to be Cartan-ordinary and the ag F is said to be a
regular ag of E.
Let V
o
n
(I) V
n
(I) denote the subset consisting of Cartan-ordinary integral el-
ements of I. It is an open (but possibly empty) subset of V
n
(I) that is a smooth
submanifold of Gr
n
(TM), and the basepoint projection : V
o
n
(I) M is a sub-
mersion. (This is because of the standing assumption that I
0
= (0).)
A dual version of Cartans Bound (also known as Cartans Test) is often useful:
For any E V
n
(I) and any ag F = (E
0
, E
1
, . . . , E
n1
), the character sequence
of F is the sequence of nonnegative integers
_
s
0
(F), s
1
(F), . . . , s
n
(F)
_
4 R. BRYANT
such that
s
i
(F) =
_

_
c(E
0
) i = 0,
c(E
i
) c(E
i1
) 1 i < n,
dimH(E
n1
) n i = n.
Then Cartans bound can also be expressed as saying that, near E, the subset V
n
(I)
is contained in a submanifold of Gr
n
(TM) of dimension
dimM +s
1
(F) + 2s
2
(F) + +ns
n
(F),
and that, if, near E, the subset V
n
(I) is a submanifold of Gr
n
(TM) of this dimen-
sion, then E is Cartan-ordinary and the ag F is regular.
When E is Cartan-ordinary, the character sequence
_
s
k
(F)
_
is the same for all
regular ags F = (E
0
, E
1
, . . . , E
n1
) in E. This common sequence is known as the
sequence of Cartan characters of E and simply written as the sequence
_
s
k
(E)
_
.
Moreover, the characters s
k
are constant on the connected components of V
o
n
(I).
2. Cartan-K ahler Theory
2.1. A form of the Cartan-Kahler Theorem. The main result needed in these
notes is the following version of the Cartan-Kahler Theorem.
1
Theorem 1 (Cartan-Kahler). Suppose that I is a real-analytic exterior dierential
system on M that is generated in positive degree and that E V
n
(I) is Cartan-
ordinary. Then there exists a real-analytic integral manifold of I that has E as one
of its tangent spaces.
Remark 3 (Generality). The Cartan-Kahler theorem constructs the desired integral
manifold by solving a sequence of initial value problems via the Cauchy-Kowalevski
Theorem. At each step in the sequence, one gets to choose appropriate initial data
that determine the resulting integral manifold. In fact, looking at the proof of the
Cartan-Kahler theorem, one sees that there is an open E-neighborhood U V
n
(I)
such that the initial data that determines a connected integral manifold of I whose
tangent spaces belong to U consists of s
0
(E) constants, s
1
(E) functions of 1 variable,
s
2
(E) functions of 2 variables, . . ., and s
n
(E) functions of n variables that are freely
speciable (i.e. arbitrary), subject only to some open conditions.
Thus, one usually says that the Cartan-ordinary integral manifolds of I (i.e., the
ones whose tangent spaces are Cartan-ordinary) depend on s
0
constants, s
1
func-
tions of 1 variable, s
2
functions of 2 variables, . . ., and s
n
functions of n variables.
Remark 4 (Signicance of the last nonzero character). One sometimes encounters
statements such as only the last nonzero character really matters, which the
writer usually phrases as something like the solution depends on s
q
functions of q
variables (where s
q
> 0 and s
k
= 0 for all k > q), thus ignoring all of the s
i
for
i < q.
The reason for this is that there is nearly always more than one way to describe
the local solutions of a given geometric problem as the Cartan-ordinary integral
manifolds of some exterior dierential system I. Two such descriptions might well
have dierent character sequences (some examples will be given below), but they
always have the same last nonzero character (at the same level q).
1
The standard, slightly stronger version of the Cartan-Kahler Theorem has more technical
hypotheses and so takes a bit longer to state.
NOTES ON EDS 5
Nevertheless, for any given exterior dierential system I, the full character se-
quence does have intrinsic meaning.
2.2. Involutive tableau. Let V and W be vector spaces over R of dimensions n
and m, respectively, and let A W V

be an r-dimensional linear subspace of

the linear maps from V to W.
2
One wants to understand the space of maps f :
V W with the property that f

(x) lies in A for all x V . Thus, f is being

required to satisfy a set of homogeneous, constant coecient, linear, rst-order
partial dierential equations, a very basic system of PDE.
Set up an exterior dierential system as follows: Let M = W V A and let
u : M W, x : M V , and p : M A denote the projections. Let I be the
ideal generated by the components of the W-valued 1-form = du p dx. Thus, I
is generated in degree 1 by m = dimW 1-forms and in degree 2 by the (at most)
m independent 2-forms that are the components of d = dpdx.
An n-plane E Gr
n
(TM) at (u
0
, x
0
, p
0
) M on which the components of dx
are independent will be described by equations of the form
du q(E) dx = dp s(E) dx = 0
where q(E) belongs to W V

and s(E) belongs to A V

(W V

) V

.
It will be an integral element of I if and only if, rst q(E) = p
0
, and, second
_
s(E) dx
_
dx = 0. This last condition is equivalent to requiring that s(E) lie in
the intersection
A
(1)
= (A V

)
_
W S
2
(V

)
_
Let r
(1)
denote the dimension of this space. Thus, the space V
n
(I) near E is a
submanifold of Gr
n
(TM) of codimension mn + (rn r
(1)
).
Now let (0) = V
0
V
1
V
n1
V
n
= V be a ag in V , and, for each k, let
A
k
WV

k
denote the image of A under the projection WV

WV

k
. This
denes a ag F = (E
0
, E
1
, . . . , E
n1
) in any E V
n
(I) on which dx : E V is an
isomorphismby letting dx(E
i
) = V
i
. Inspection now shows that c(E
i
) = m+dimA
i
,
so Cartans bound becomes
mn + (rn r
(1)
) c(E
0
) + +c(E
n1
) = mn +
n1

i=1
dimA
i
,
which, after rearrangement, becomes
dimA
(1)
= r
(1)
ndimA
n1

i=1
dimA
i
.
In particular, whether an integral element on which dx is independent has a regular
ag (and hence is Cartan-ordinary) depends only on the subspace A W V

.
The numbers s
i
(F, A) = dimA
i
dimA
i1
for 1 i n are called the characters
of the ag F with respect to A. In terms of the characters s
i
(F, A), the above
inequality becomes
dimA
(1)
s
1
(F, A) + 2 s
2
(F, A) + +ns
n
(F, A)
and equality holds if and only if F is a regular ag and the integral elements E
V
n
(I) on which dx : E V is a isomorphism are Cartan-ordinary. When such a
2
In the literature, A is often called a tableau, which is simply a borrowing from the French of
the word used to describe a subspace of linear maps written out as a matrix whose entries satisfy
some given linear relations.
6 R. BRYANT
ag F exists, the tableau A is said to be involutive, and the Cartan characters of A
are s
i
(A) = s
i
(F, A) (computed with respect to any regular ag F).
When A is involutive, the Cartan-Kahler theorem implies that the real-analytic
integral manifolds of A exist and depend on s
0
= m constants, s
1
(A) functions of
1 variable, s
2
(A) functions of 2 variables, etc.
In particular, if one takes the Taylor series of the general solution f : V W
of the equations forcing f

(x) to lie in A for all x, one gets

f(x) = f
0
+f
1
(x) +f
2
(x) + +f
k
(x) +
where f
k
is a W-valued homogeneous polynomial of degree k on V and hence lies
in the subspace
A
(k1)
=
_
W S
k
(V

)
_

_
A S
k1
(V

)
_
.
which has dimension
dimA
(k1)
=
n

j=1
_
j +k 2
k 1
_
s
j
(A) ,
which is exactly what one would expect if f were to be thought of as being comprised
of s
1
(A) functions of 1 variable, s
2
(A) functions of 2 variables, etc.
The concept of involutivity turns out to be fundamental, so it is worthwhile
to examine how this is connected with the notion of a Cartan-ordinary integral
element in general.
Thus, x E V
n
(I), with E T
x
M. Let E

x
M be the space of forms that
vanish on E, and note that the ideal (E

) (T

x
M) generated by E

consists of
the forms that vanish on E T
x
M. Let I
x
(T

x
M) denote the set of values of
forms in I at the point x, Then I
x
is contained in (E

) because E is an integral
element of I. Consider the quotient
I
E
= (I
x
+ (E

)
2
)/(E

)
2
(E

)/(E

)
2
E

(E

).
This I
E
E

(E

) should be thought of as the linearization of the ideal I

at E. It generates an ideal (I
E
) in the space of forms on T
x
M/E E (whose dual
space is E

) that has 0 E T
x
M/E E as an integral element.
If E is, in addition, Cartan-ordinary, then it is not dicult to show that 0 E is
a Cartan-ordinary integral element of (I
E
), that a ag of the form 0 E
i
is regular
for 0 E if and only if F = (E
i
) is a regular ag for E, and that one has equality
of characters s
i
(0 E) = s
i
(E).
This motivates the following: Given two vector spaces W and V (of dimensions
m, and n, respectively), a graded
3
subspace I W

(V

) is involutive if 0V
W V is a Cartan-ordinary integral element of the ideal (I)
_
(W V )

_
generated by I. In this case, set s
i
(I) = s
i
(0 V ) and let
A
I
=
_
f Hom(V, W) |
f
V
n
_
(I)
_ _
W V

,
where
f
= {(f(x), x) | x V } WV ; the subspace A
I
is said to be the tableau
of I. When I is involutive, A
I
is also involutive and its characters are
s
i
(A
I
) = s
i
(I) + s
i+1
(I) + + s
n
(I).
3
This means that I is the direct sum of its subspaces I
q
= I

q
(V

)

.
NOTES ON EDS 7
3. First Examples and Applications
I will now give a basic set of examples illustrating the concepts and applica-
tions of the Cartan-Kahler Theorem. Some are intended just to help the reader
gain familiarity with the concepts, while others will turn out to have signicant
applications.
Example 1 (The Frobenius theorem). Suppose that I on M
n+s
can be locally
generated algebraically by s linearly independent 1-forms
1
, . . . ,
s
. In particular,
since I is dierentially closed, it follows that there are (local) 1-forms
a
b
such that
d
a
=
a
b

b
.
In particular, there is a unique n-dimensional integral element at each point x
M, namely the n-dimensional subspace E
x
T
x
M on which each of the
a
vanish.
Thus, V
n
(I) Gr
n
(TM) is simply a copy of M, in fact, the image of a smooth
section of the bundle Gr
n
(TM), so it is a smooth manifold of dimension n+s.
Meanwhile, for any ag F = (E
0
, . . . , E
n1
) in E
x
, one has H(E
p
) = E
x
, so
c(E
i
) = s for 0 i < n. In particular, s
0
(F) = s and s
i
(F) = 0 for 0 < i < n.
Since dimV
n
(I) = n+s = dimM + s
1
+ 2s
2
+ + ns
n
, it follows that Cartans
bound is saturated, and all of the elements of V
n
(I) are Cartan-ordinary and all
their ags are regular.
By the Cartan-Kahler Theorem, every E
x
is tangent to an integral manifold of I
and the local integral manifolds near E depend on s
0
= s constants.
Now, in this particular case, there is another way to get the same result, which is
to use the Frobenius Theorem (which is even better since it applies in the smooth
setting). This Theorem says that, locally, it is possible to choose closed genera-
tors
a
= dy
a
for some functions y
1
, . . . , y
s
that form part of a coordinate sys-
tem x
1
, . . . , x
n
, y
1
, . . . , y
s
. Then the local n-dimensional integral manifolds of I
are the leaves dened by holding the y
a
constant, so that the general local n-
dimensional integral manifold depends on s constants, in agreement with the pre-
diction of the Cartan-Kahler Theorem.
Example 2 (A non-ordinary integral element). Let M = R
3
, with coordinates x, y, z,
and let I be generated by the 2-forms dxdz and dydz. Then the 2-plane eld
dened by dz = 0 consists of 2-dimensional integral elements, and these are the only
2-dimensional integral elements, so that V
2
(I) is a smooth 3-manifold in Gr
2
(TM).
Since I
1
= 0, one has c(E
0
) = 0 for all E
0
. Letting E
1
be spanned by a
x
+ b
y
,
where (a, b) = 0, one nds that H(E
1
) has dimension 2 and is dened by dz = 0,
so c(E
1
) = 1. Thus, c(E
0
) +c(E
1
) = 1 while the codimension of V
2
(I) in Gr
2
(TM)
is 2. Thus, E
2
= H(E
1
) has no regular ag and hence is not Cartan-ordinary.
It may seem disappointing that the Cartan-Kahler Theorem does not apply
to prove the existence of 2-dimensional integral manifolds, especially, since there
evidently does exist an integral manifold tangent to every 2-dimensional integral
element, namely, a horizontal plane z = z
0
.
However, to see why one should not expect Cartan-Kahler to apply in this case,
consider a modication of this example got by instead considering the ideal I

generated by dxdz and dy(dz y dx). The ideals I and I

are algebraically
equivalent at each point, and so one sees that, there is also a unique 2-dimensional
integral element of I

through each point, namely the one that satises dzy dx = 0.

Since the algebra of the polar equations is essentially the same for I

as it is for I,
8 R. BRYANT
these integral elements of I

also are not Cartan-ordinary, and this is good because

there evidently are not any integral surfaces of the equation dz y dx = 0.
Example 3 (Lagrangian submanifolds). Let M = R
2n
and let I be generated by
the symplectic form
= dp
1
dx
1
+ + dp
n
dx
n
.
Then the n-plane E spanned by the
x
i is an integral element of I, and, if one takes
the ag F = (E
0
, . . . , E
n1
) so that E
i
is spanned by the
x
j with 1 j i, then
one computes that, for 0 < i < n, the polar space H(E
i
) is the subspace dened
by dp
1
= dp
2
= = dp
i
= 0. Thus, c(E
i
) = i.
By Cartans bound, V
n
(I) has codimension at least
C = 1 + 2 + + (n1) =
1
2
n(n1)
in Gr
n
(TM) near E. Meanwhile, any

E Gr
n
(TM) on which the dx
i
are linearly
independent will be dened by unique equations of the form
dp
i
s
ij
(

E) dx
i
= 0
for some numbers s
ij
(

E), and these functions s
ij
, together with the x
i
and the p
i
dene a local coordinate system on an open subset of Gr
n
(TM) that contains E
(which is dened by s
ij
(E) = 0).
By Cartans Lemma, such an

E will be an integral element of I if and only
if s
ij
(E) s
ji
(

E) = 0. This is
1
2
n(n1) independent equations on

E, so that
V
n
(I) has codimension
1
2
n(n1) = C in Gr
n
(TM) near E. Consequently, E is
Cartan-ordinary, and the ag F is regular.
Of course, one already knows that Lagrangian manifolds exist, so this is not a
surprise. Note, however, that what the Cartan-Kahler theorem would say is that
one can specify an integral manifold on which the x
i
are independent uniquely by
choosing p
n
to be an arbitrary function of the x
i
, then choosing p
n1
subject to the
condition that its partial in the x
n
-direction equals the partial of p
n
in the x
n1
direction (which determines p
n1
up to the addition of a function of x
1
, . . . x
n1
),
then choosing p
n2
subject to the conditions that its partials in the x
n
- and x
n1
-
directions are determined by those of p
n
and p
n1
(which determines p
n2
up to
the addition of a function of x
1
, . . . x
n2
), etc. Thus, the integral manifolds are
described by a choice of 1 = s
n
(E) function of n variables, 1 = s
n1
(E) function
of n1 variables, etc., in agreement with the general theory.
Of course, one can also specify a Lagrangian using only one function of n variables
simply by taking p
i
=
u
x
i
for some function u of x
1
, . . . , x
n
. However, for general I,
one cannot nd such a formula that combines the arbitrary functions in the general
Cartan-ordinary integral manifolds of I in this way.
Another way to interpret this discrepancy is to note that the Lagrangian
manifolds on which dx
1
dx
n
= 0 are, by the above formula put in cor-
respondence with the arbitrary local function u of n variables, which, via its
graph
_
x
1
, . . . , x
n
, u(x
1
, . . . , x
n
)
_
is seen to be an integral manifold of the trivial
ideal I = (0) on R
n+1
, which has Cartan characters
(s
0
, s
1
, . . . , s
n1
, s
n
) = (0, 0, . . . , 0, 1).
Thus, this provides an example of the phenomenon that I mentioned earlier of
two dierent exterior dierential systems describing (local) solutions to the same
NOTES ON EDS 9
problem. Note that they have the same last nonzero character, namely, s
n
= 1,
while their lower characters are dierent.
3.1. A version of Cartans Third Theorem. Suppose that C
i
jk
= C
i
kj
and
F

i
(with 1 i, j, k n and 1 s) are given functions on R
s
, and one wants
to know whether or not there exist linearly independent 1-forms
i
on R
n
and a
function a = (a

) : R
n
R
s
that satisfy the Cartan structure equations
(3.1) d
i
=
1
2
C
i
jk
(a)
j

k
and da

= F

i
(a)
i
.
Such a pair (a, ) will be said to be an augmented coframing satisfying the structure
equations (3.1).
Applying the fundamental identity d
2
= 0 yields necessary conditions in order
for such a pair (a, ) to exist: One must have d(C
i
jk
(a)
j

k
) = d(d
i
) = 0
and d
_
F

i
(a)
i
_
= d(da

) = 0. Expanding these identities using (3.1) and the

assumed independence of the
i
then yields that, if, for each u
0
R
s
, an augmented
coframing (a, ) on some n-manifold M exists satisfying (3.1) with a(x) = u
0
for
some x M, then one must have
(3.2) F

j
C
i
kl
u

+F

k
C
i
lj
u

+F

l
C
i
jk
u

=
_
C
i
mj
C
m
kl
+C
i
mk
C
m
lj
+C
i
ml
C
m
jk
_
and
(3.3) F

i
F

j
u

j
F

i
u

= C
l
ij
F

l
.
Cartan proved the converse statement [5]:
Theorem 2 (Cartans Third Fundamental Theorem). Suppose that C
i
jk
= C
i
kj
and F

i
are real-analytic functions on R
s
that satisfy (3.2) and (3.3). Then, for
any u
0
R
s
, there exists an augmented coframing (a, ) on R
n
that satises (3.1)
and has a(0) = u
0
. (Moreover, any two such augmented coframings agree on a
neighborhood of 0 R
n
up to a dieomorphism of R
n
that xes 0 R
n
.)
Proof. Let M = GL(n, R) R
n
R
s
, and let p : M GL(n, R), x : M R
n
,
and u : M R
s
be the projections. Consider the ideal I generated on M by the
n 2-forms

i
= d(p
i
j
dx
j
) +
1
2
C
i
jk
(u)(p
j
l
dx
l
) (p
k
m
dx
m
)
and the s 1-forms

= du

i
(u) (p
i
j
dx
j
).
Note that one can write

i
=
i
j
dx
j
for some 1-forms
i
j
= dp
i
j
+ P
i
jk
dx
k
for some functions P
i
jk
on M and that the
forms
i
j
, dx
k
, and

dene a coframing on M, i.e., they are linearly independent

everywhere and span the cotangent space everywhere.
Now, the hypothesis that d
2
= 0 be a formal consequence of the structure equa-
tions (i.e., the equations (3.2) and (3.3)) is easily seen to be equivalent to the
equations
d
i
=
1
2
C
i
jk
u

(p
j
l
dx
l
) (p
k
m
dx
m
) +C
i
jk

j
(p
k
m
dx
m
)
10 R. BRYANT
and
d

=
F

i
u

(p
i
j
dx
j
) +F

i

i
.
Thus, these hypotheses imply that I is generated algebraically by the
i
and the

. This makes it easy to choose an integral element and compute the Cartan
characters:
Fix a point z M and let E T
z
M be the n-dimensional integral element
dened by
i
j
=

= 0. Let F be the ag in E dened so that E

i
is also annihilated
by the dx
j
for j > i. Then one nds that H(E
i
) is dened by

=
j
k
= 0
where k i, and hence that c(E
i
) = s + ni for 0 i n1. In particular, it
follows that V
n
(I) must be contained in a submanifold Gr
n
(TM) of codimension
at least C = ns +
1
2
n
2
(n1).
Meanwhile, any n-plane

E on which the dx
i
are linearly independent is specied
by knowing the ns +n
3
numbers s

i
(

E) and s
i
jk
(

E) such that

E satises

i
j
s
i
jk
(

E) dx
k
=

k
(

E), dx
k
= 0.
The condition that such an

E be an integral element of I is then that s

k
(

E) =
s
i
jk
(

E) s
i
kj
(

E) = 0, which is ns +
1
2
n
2
(n1) = C equations on

E. Thus, E is
Cartan-ordinary, and F is a regular ag.
Now, since the functions C
i
jk
and F

i
are assumed to be real-analytic, the Cartan-
Kahler Theorem applies and one concludes that there is an integral manifold of I
tangent to E. This integral manifold is described by having the p
i
j
and the u

be
certain functions of the x
1
, . . . , x
n
, say, p
i
j
= f
i
j
(x) and u

= a

(x). These then

give the desired (a

,
i
) =
_
a

(x), f
i
j
(x) dx
j
_
.
Remark 5 (Cartans original theorem). The result I have just proved
4
is only a very
special case of the theorem that Cartan proves in the rst of his innite groups
papers [5]. However, this version suces for applications that I have in mind in
these notes, and it is much easier to state than Cartans full theorem.
Remark 6 (Uniqueness). The general Cartan-ordinary integral of I depends on n
arbitrary functions of n variables (since the last nonzero character is s
n
= n), but
this is to be expected because, given any integral as a (local) coframing on R
n
,
one can get others by simply pulling back by an arbitrary dieomorphism of R
n
.
5
To get uniqueness up to local dieomorphism for data (a, ) in which a takes on a
specic value a
0
R
s
, one shows that two such solutions are locally equivalent by
an application of Cartans technique of the graph.
Note, by the way, that when s = 0 (i.e., there are no functions a

), this result
becomes Lies Third Theorem giving the existence of a local Lie group for any given
Lie algebra.
Remark 7 (Smoothness and Globalization). While this treatment assumes real-
analyticity, so that the Cartan-Kahler Theorem can be applied, it is now known
that the theorem is true in the smooth category as well. The proof in the smooth
4
The proof in the text is Cartans; I have merely simplied his proof as possible in this special
case.
5
Alternatively, one should think of the Cartan-ordinary integral manifolds of I as giving a
(local) augmented coframing (a, ) satisfying the structure equations (3.1) plus a local coordinate
system x = (x
i
) on the domain of (a, ).
NOTES ON EDS 11
case is not dicult, but requires a little more insight than this simple application
of Cartan-Kahler.
The reader will probably also have noticed that nothing is really used about
the domain of the functions C
i
jk
and F

i
other than that it is a smooth manifold
of some dimension s. This observation spurred the development of a globalized
version of Cartans Theorem, which becomes the subject of Lie algebroids, in which
R
s
is replaced by a smooth manifold A. For these details on these developments,
as well as the smooth theory, the reader should consult treatises devoted to these
subjects, but I will sketch the translation here to aid in comparison with a somewhat
generalized construction associated to a variant of Cartans Theorem that I will
describe in the next subsection.
Recall that a Lie algebroid is a manifold A endowed with a vector bundle Y A
of rank n whose space of sections (Y ) carries a Lie algebra structure
_
,
_
: (Y ) (Y ) (Y )
together with a bundle map : Y TA that induces a homomorphism of Lie
algebras on the spaces of sections
6
and that satises the Leibnitz compatibility
condition
(3.4)
_
U, fV
_
= (U)f V +f
_
U, V
_
for all U, V (Y ) and f C

(A).
A realization of
_
A, Y, {, },
_
is a triple (M, a, ), where M is an n-manifold,
a : M A is a (smooth) mapping, and : TM a

Y is a vector bundle
isomorphism, such that = da : TM TA and such that induces an
isomorphism of Lie algebras on the space of sections of TM and a

Y .
To see the translation from Cartans language to that of Lie algebroids, start
with the data of functions C
i
jk
= C
i
kj
and F

i
on A = R
s
. Set Y = A R
n
with
a basis for sections U
i
and set
_
U
i
, U
j
_
= C
k
ij
U
k
and dene : Y TA = TR
s
by
(U
i
) = F

.
Then (3.2) and (3.3) are precisely the equations necessary and sucient in order
that (3.4) hold, that {, } dene a Lie bracket on the space of sections of Y , and
that : Y TA induce a homomorphism of Lie algebras.
Moreover, an augmented coframing (a

,
i
) on a manifold M
n
satises Cartans
structure equations if and only if, when one sets
= U
i

i
,
and denes a : M R
s
to be a = (a

), the data (M, a, ) is a realization in the

above sense.
This approach to globalizing Cartans theorem has been very fruitful, and the
reader is encouraged to consult the literature on Lie algebroids for more on this
development.
However, it should be borne in mind that Cartans original formulation in terms
of what I am calling augmented coframings turns out already to be very well
6
Here, (TA), the set of vector elds on A, is given its standard Lie algebra structure via the
Lie bracket.
12 R. BRYANT
suited for applications to dierential geometry, as I hope to show in the discussion
of examples below.
3.2. Variants of Cartans Third Theorem. Cartans Third Theorem is one of
a number of existence results that are all proved more or less the same way, at
least in the real-analytic category. In this subsection, I will give two such variants,
and, in the following sections in the notes, I will illustrate their use in a range of
dierential geometry problems.
Throughout this rst variant, the index ranges 1 i, j, k n, 1 s, and
1 , r will be assumed.
Suppose that C
i
jk
(u) = C
i
kj
(u) are given functions on R
s
while F

i
(u, v) are
given functions on R
s+r
, and suppose that one wants to know whether or not there
exist linearly independent 1-forms
i
on an n-manifold M, a function a = (a

) :
M R
s
, and a function b = (b

) : M R
r
that satisfy these Cartan structure
equations
(3.5) d
i
=
1
2
C
i
jk
(a)
j

k
and da

= F

i
(a, b)
i
.
Such a triple (a, b, ) on M will be said to be an augmented coframing satisfy-
ing (3.5).
Note that this is a dieomorphism invariant notion, since, if f : N M is a dif-
feomorphism, then
_
f

a, f

b, f

_
will be an augmented coframing on N that satis-
es (3.5). In many geometric problems (see some examples in the next section), one
is interested in understanding the general augmented coframing satisfying (3.5)
and one regards two such augmented coframings that dier by a dieomorphism as
equivalent.
One should think of the b

as unconstrained derivatives of the functions a

.
Thus, this version of Cartans structure equations covers situations in the more
typical case in which one does not have formulae for all of the derivatives of the
geometric quantities that appear in the problem. Informally, one speaks of the
functions b

as free derivatives.
To understand necessary and sucient conditions for such augmented coframings
to exist, one again wants to consider the consequences of the identity d
2
= 0, but
now, because of the free derivatives appearing in the structure equations, one cannot
simply expand this fundamental identity formally and arrive at complete necessary
and sucient conditions on the functions C and F.
Now, the equations d(d
i
) = d(C
i
jk
(a)
j

k
) = 0 do make good sense, so one
should require that C and F at least satisfy
(3.6) F

j
C
i
kl
u

+F

k
C
i
lj
u

+F

l
C
i
jk
u

=
_
C
i
mj
C
m
kl
+C
i
mk
C
m
lj
+C
i
ml
C
m
jk
_
.
(Because the F

i
contain the variables v

while the right hand side of (3.6) does

not, this equation places constraints on how the v

can appear in the F

j
.)
Meanwhile, expanding d(da

) = d
_
F

i
(a, b)
i
_
= 0 yields
0 =
F

i
v

(a, b)db

i
+
1
2
_
F

i
(a, b)
F

j
u

(a, b) F

j
(a, b)
F

i
u

(a, b) C
l
ij
(a) F

l
(a, b)
_

j
,
and the simplest way for these equations to be satisable by some expression of
the db

in terms of the
i
would be for there to exist functions G

j
on R
s+r
such
NOTES ON EDS 13
that
(3.7) F

i
F

j
u

j
F

i
u

C
l
ij
F

l
=
F

i
v

j

F

j
v

i
,
for then the above equations can be written in the form
0 =
F

i
v

(a, b)
_
db

j
(a, b)
j
_

i
.
The conditions (3.6) and (3.7) will at least ensure that there are no obvious
incompatibilities derivable by taking the exterior derivatives of the structure equa-
tions. However, they arent enough to guarantee that there wont be higher order
incompatibilities. To rule this out, it will be necessary to impose conditions on how
the free derivatives v

appear in the functions F

i
. Let u
1
, . . . , u
s
be a basis of R
s
,
and let x
1
, . . . x
n
be a basis of the dual of R
n
. Let A(u, v) Hom(R
n
, R
s
) denote
the subspace (i.e., tableau) spanned by the r elements
(3.8)
F

i
v

(u, v) u

v
i
, 1 r.
This A(u, v) is known as the tableau of free derivatives of the structure equations
at the point (u, v) R
s+r
.
Here is a useful variant
7
of Theorem 2.
Theorem 3. Suppose that real analytic functions C
i
jk
= C
i
kj
on R
s
and F

i
on R
s+r
are given satisfying (3.6) and that there exist real analytic functions G

i
on R
s+r
that satisfy (3.7). Finally, suppose that the tableaux A(u, v) dened by (3.8)
have dimension r and are involutive, with Cartan characters s
i
(1 i n) for
all (u, v) R
s+r
. Then, for any (u
0
, v
0
) R
s+r
there exists an augmented cofram-
ing (a, b, ) on an open neighborhood V of 0 in R
n
that satises (3.5) and has
_
a(0), b(0)
_
= (u
0
, v
0
).
Proof. Let M = GL(n, R) R
n
R
s
R
r
, and let p : M GL(n, R), x : M R
n
,
u : M R
s
, and v : M R
r
be the projections. Consider the ideal I generated
on M by the n 2-forms

i
= d(p
i
j
dx
j
) +
1
2
C
i
jk
(u)(p
j
l
dx
l
) (p
k
m
dx
m
)
and the s 1-forms

= du

i
(u, v) (p
i
j
dx
j
).
Note that, as in the proof of Theorem 2, one can write

i
=
i
j
dx
j
for some 1-forms
i
j
= dp
i
j
+P
i
jk
dx
k
for some functions P
i
jk
on M and that the forms

i
j
, dx
k
,

, together with

= db

i
(p
i
j
dx
j
) dene a coframing on M, i.e., they
are linearly independent everywhere and span the cotangent space everywhere.
Now, the hypotheses of the theorem imply that
d
i
=
1
2
C
i
jk
(u)
a

(p
j
l
dx
l
) (p
k
m
dx
m
) +C
i
jk
(u)
j
(p
k
m
dx
m
)
while
d

=
F

i
b

(p
i
j
dx
j
) +
F

i
(u)
a

(p
i
j
dx
j
) +F

i

i
.
7
Note that the proof is very closely patterned on Cartans proof of Theorem 2.
14 R. BRYANT
Thus, I is generated algebraically by the
i
, the

, and the 2-forms

=
F

i
b

(p
i
j
dx
j
).
This makes it easy to choose an integral element and compute the Cartan characters:
Fix a point z = (I
n
, 0, u
0
, v
0
) M, and let E T
z
M be the n-dimensional
integral element dened by
i
j
=

= 0.
Choose a regular ag for the tableau A(u
0
, v
0
) (which, by hypothesis, exists).
By rotating the x
i
if necessary, one can assume that the ag F in E dened so that
E
i
is also annihilated by the dx
j
for j > i is such a regular ag. Then one nds
that H(E
i
) is dened by

=
j
k
=
F

k
b

= 0
where k i, so c(E
i
) = s+ni+dimA(u
0
, v
0
)
i
= s+ni+s
1
+ +s
i
for 0 i n1.
Meanwhile, any n-plane E

Gr
n
(TM) on which the dx
i
are independent will
be dened by equations of the form

i
(E

) dx
i
=
l
k
q
l
ki
(E

) dx
i
=

i
(E

) (p
i
j
dx
j
) = 0
for some numbers q

i
(E

), q
l
ki
(E

), q

i
(E

). The conditions that E

be an integral
element of I then imply that
q

i
(E

) = q
l
ki
(E

) q
l
ik
(E

) =
F

i
b

(u, v)q

j
(E

)
F

j
b

(u, v)q

i
(E

) = 0,
and, by the hypothesis that F = (E
i
) is a regular ag for the tableaux A(u
0
, v
0
)
(and hence is also regular for A(u, v) for (u, v) near (u
0
, v
0
)), it follows that this is
c(E
0
) +c(E
1
) + +c(E
n1
) = ns +
1
2
n
2
(n+1) +(n1)s
1
+(n2)s
2
+ +s
n1
equations on the quantities q

i
(E

), q
l
ki
(E

), q

i
(E

). Thus, Cartans bound is satu-

rated, and the ag F is regular for E.
8
The ideal I is real analytic, so the Cartan-Kahler Theorem applies, and one con-
cludes that there is an integral manifold of I tangent to E. This integral manifold
is described by having the p
i
j
, the u

, and the v

be functions of the x
1
, . . . , x
n
, say,
p
i
j
= f
i
j
(x) and u

= a

(x) and v

= b

(x). These then give the desired augmented

coframing (a

, b

,
i
) =
_
a

(x), b

(x), f
i
j
(x) dx
j
_
.
Remark 8 (Generality). Theorem 3 as stated only gives existence for specied
(u
0
, v
0
), but, as will be seen, the (local) augmented coframings that satisfy the
structure equations depend (modulo dieomorphism) on s constants, s
1
functions
of 1 variable, s
2
functions of 2 variables, etc., but to make precise sense of this, I
will need to discuss prolongation, which comes in the next section.
Remark 9 (Globalization). Just as in the case of Theorem 2, which has a modern
formulation in terms of Lie algebroids, there is a global version of Theorem 3.
9
The appropriate global data structure,
_
A, B, , Y, {, },
_
, starts with two mani-
folds, A of dimension s and B of dimension r +s, and a submersion : B A. For
8
Essentially, the involutive tableau A(u
0
, v
0
) is being combined with a tableau already shown
to be involutive in the proof of Theorem 2, one for which every ag is regular. Perhaps, I should
also remind the reader that the s
i
are the characters of the tableaux A(u, v) and not of the ideal I
constructed above. In fact, one has s
0
(I) = s and s
i
(I) = s
i
+n for 1 i n.
9
I will not actually need this formulation in these notes, but since there were questions about
this during the lectures, I will briey describe it here.
NOTES ON EDS 15
notational convenience, let K = ker

TB, and let Q = TB/K be the quotient

bundle over B. For a vector eld X on B, let X
K
(i.e., X modulo K) denote the
corresponding section of Q.
Next, the data structure includes a vector bundle Y A of rank n, and a Lie
algebra structure {, } on the space C

(Y ) of sections of Y over A. For U C

(Y ),
let U

Y ) denote the pullback section of the pullback bundle over B, i.e.,

U

(b) = U
_
(b)
_
.
Finally, the data includes a bundle map :

Y TB, that satises

(3.9)
_
{U, V }

_
K
=
_
(U

), (V

)

K
and the requirement that there exist an anti-symmetric, bilinear product
10
{{, }} on
C

Y ) that satises the compatibility condition

(3.10) {{U

, f V

}} =
_
(U

)f
_
V

+f {U, V }

for U, V C

(Y ) and f C

(B).
A realization of the data structure
_
A, B, , Y, {, },
_
is a triple (M, b, ), where M
is an n-manifold, b : M B is a smooth mapping, and : TM (b)

Y is an
isomorphism of bundles that induces an isomorphism of Lie algebras on the appro-
priate spaces of sections and that satises d(b) =

.
(Note that if : B A is a dieomorphim (e.g., r = 0), then the data
_
A, Y, {, },
_
denes a Lie algebroid, and the notion of a realization is the standard
one.)
Now, there is a map : K Q(

Y )

of B-bundles, uniquely determined by

the condition that it satisfy
(X)(U

) = [X, (U

)]
K
for any X (K) and U C

(Y ). One says that the data

_
A, B, , Y, {, },
_
is nondegenerate if injective, and, further, that it is (uniformly) involutive if
(K)
b
Q
b
(

Y )

b
is an involutive tableau for all b B (and the Cartan
characters s
i
_
(K)
b
) are constant, independent of b B).
Then Theorem 3 asserts the local existence of realizations (M, b, ) of uniformly
involutive, nondegenerate real analytic data structures
_
A, B, , Y, {, },
_
with b :
M B taking any specied value b
0
B.
To see the translation from the notation of Theorem 3 to this global formulation,
let A = R
s
(with coordinates u

), let B = R
s+r
(with coordinates u

and v

),
let : R
s+r
R
s
be the projection on the rst s coordinates, let Y = R
s
R
n
(with the standard basis of sections U
i
), let
{U
i
, U
j
} = C
k
ij
(u) U
k
,
and let
(U

i
) = F

i
(u, v)

+G

i
(u, v)

.
The reader can now verify that (3.6) and (3.7) are the necessary and sucient
conditions that {, } dene a Lie bracket on C

(Y ), that (3.9) hold, and that there

exists an extension {{, }} of {, } to sections of C

Y ) that satises (3.10).

(I should point out that this global formulation is not perfect, because, ideally,
one should only have to specify the functions G

i
up to a section of the prolongation
10
N.B.: It is easy to see that there is at most one such product {{, }} satisfying (3.10). In
general, this extended product is not a Lie algebra structure on C

Y ).
16 R. BRYANT
of the tableau bundle, i.e., one should regard two such structures
_
A, B, , Y, {, },
_
and
_
A, B, , Y, {, },

_
as the same if the dierence =

, which is a section
of TB (

Y )

, is actually a section of the kernel K

(1)
of the composition
K (

Y )

id
Q(

Y )

Y )

Q
2
_
(

Y )

_
.
Thus, one should probably formulate the data structure with the notion of non-
degeneracy built into the axioms and with taking values in the quotient bun-
dle (TB (

Y )

)/K
(1)
instead of in TB (

Y )

. However, this is turns out to

be awkward, as checking that the axioms even make sense becomes cumbersome.)
While this global formulation may be more satisfying than the coordinate
formulation in Theorem 3, one should bear in mind that there is little (and, more
often than not, no) hope of proving the global realization theorems that one has
in the more familiar case of Lie algebroids. For the general such structure, there
is no obvious notion of completeness of a realization and there is also no obvious
way to classify even the germs of realizations up to dieomorphism. (However,
there is a way to test two such germs for dieomorphism equivalence, at least in
the real-analytic category. I will say more about this in Remark 10.)
Remark 10 (Local equivalence of realizations). The reader may be wondering how
one distinguishes two realizations of the data in Theorem 3 up to dieomorphism.
After all, as Cartan proved, given two augmented coframings (M, a, ) and (

M, a, )
satisfying (3.1) and points x M and x

M such that a(x) = a( x), there will exist
an x-neighborhood U M, an x-neighborhood

U

M, and a dieomorphism f :

U U such that ( a, ) = (f

a, f

) and f( x) = x.
In contrast, for augmented coframings (M, a, b, ) and (

M, a,

b, ) satisfying (3.5),
having points x M and x

M with
_
a(x), b(x)
_
=
_
a( x),

b( x)
_
is not sucient
to imply that there is a dieomorphism f :

U U for some x-neighborhood U and
x-neighborhood

U such that ( a,

b, ) = (f

a, f

b, f

).
A sucient condition (due, of course, to Cartan [7]) for local dieomorphism
equivalence does exist in this more general case but it is more subtle.
An augmented coframing (a, b, ) on M
n
satisfying (3.5), is regular of rank p
at x M if there is an x-neighborhood U M, a smooth submersion h : U R
p
,
and a smooth map (A, B) : h(U) R
s+r
such that A : h(U) R
s
is a smooth
embedding and such that (a, b) = (Ah, Bh) holds on U. Note, in particular, that
this implies that the image (a, b)(U) R
s+r
is a smoothly embedded p-dimensional
submanifold that is a graph over its projection a(U) R
s
(also a smoothly embed-
ded p-dimensional submanifold). Equivalently, (a, b, ) is regular of rank p at x M
if some p of the functions a

have independent dierentials at x and, moreover, on

some x-neighborhood U M, all of the other a

and all of the b

can be ex-
pressed as smooth functions of those p independent functions. For an augmented
coframing (a, b, ) satisfying (3.5), being regular of rank p at a point x M is a
dieomorphism-invariant condition.
Cartan showed that, if (M, a, b, ) and (

M, a,

b, ) satisfy (3.5), are regular of

rank p at points x M and x

M with
_
a(x), b(x)
_
=
_
a( x),

b( x)
_
, and there
are an x-neighborhood U M and x-neighborhood

U

M such that (a, b)(U)
and ( a,

b)(

U) are the same p-dimensional submanifold of R

r+s
, then, after possibly
shrinking U and

U, there exists a dieomorphism f :

U U such that ( a,

b, ) =
(f

a, f

b, f

) and f( x) = x.
NOTES ON EDS 17
The reader should have no trouble rephrasing Cartans sucient condition in a
form suitable for the global data structure version described in Remark 9. (The
reader may feel that the hypotheses of Cartans equivalence theorem are absurdly
strong, but, without knowing more about a specic set of structure equations (3.5),
it is not possible to weaken these hypotheses in any signicant way and still get the
conclusion of local equivalence, as examples show.)
I conclude this subsection with another useful variant of Cartans Third Theorem.
Let V be a vector space of dimension n. For each V -valued coframing : TM V
on an n-manifold M, there will be a unique function C : M V
2
(V

), the
structure function of , such that
(3.11) d =
1
2
C( ).
Given a basis v
i
of V with dual basis v
i
, one has = v
i

i
and C =
1
2
C
i
jk
v
i
v
j
v
k
,
and (3.11) takes the familiar form d
i
=
1
2
C
i
jk

j

k
.
Now, let A V
2
(V

) be a submanifold.
11
A V -valued coframing : TM
V will be said to be of type A if its structure function C : M V
2
(V

) takes
values in A. The goal is to determine the generality of the space of (local) V -valued
coframings of type A when two such that dier by a dieomorphism of M are
regarded as equivalent.
For example, if A consists of a single point a
0
=
1
2
c
i
jk
v
i
v
j
v
k
, then Lies The-
orem asserts that a necessary and sucient condition that such a coframing exist
is that J(a
0
) = 0, where J : V
2
(V

) V
3
(V

) is the quadratic mapping

(sometimes called the Jacobi mapping) that one gets by squaring, contracting, and
skewsymmetrizing:
V
2
(V

)
_
V
2
(V

)
_

_
V
2
(V

)
_
V V

2
(V

) V
3
(V

).
Given a basis v
i
of V with dual basis v
i
, the formula for J is
J
_
1
2
c
i
jk
v
i
v
j
v
k
_
=
1
6
(c
i
jm
c
m
kl
+c
i
km
c
m
lj
+c
i
lm
c
m
jk
) v
i
v
j
v
k
v
l
.
Of course, in this case, all V -valued coframings of type A = {a
0
} are locally
equivalent up to dieomorphism.
This motivates the following denitions: A submanifold A V
2
(V

) is said
to be a Jacobi manifold if
(3.12) J(a) (T
a
A V

)
for all a A, where : V
2
(V

) V

V
3
(V

) is the skewsymmetrization
mapping dened by exterior multiplication. The condition (3.12) is an obvious
necessary condition in order for there to exist a V -valued coframing : TM V
whose structure function takes values in A and assumes the value a A. It is not,
in general, sucient.
A Jacobi manifold A is involutive if each of its tangent spaces T
a
V
2
(V

)
is involutive, with characters s
i
(T
a
) = s
i
.
I can now state a useful existence result that I will apply in some examples.
11
In most applications, A will be an ane subspace of V
2
(V

), but the extra generality
of allowing A to be a submanifold is frequently useful.
18 R. BRYANT
Theorem 4. Let V be a vector space, and let A V
2
(V

) be a real-analytic,
involutive Jacobi manifold. Then, for any a
0
A, there exists a V -valued cofram-
ing of type A on a neighborhood U of 0 V such that its structure function C
satises C(0) = a
0
.
Proof. The proof follows the by-now familiar pattern laid down by Cartan.
The result is local, so one can suppose that A has dimension s and is parametrized
by a real-analytic embedding T : R
s
A V
2
(V

) with T(0) = a
0
. Write
T =
1
2
T
i
jk
(a) v
i
v
j
v
k
where the T
i
jk
are some real analytic functions on R
s
. By hypothesis, for each a =
(a

) R
s
, the tableau A
a
V
2
(V

) spanned by the s independent elements

A

(a) =
T
i
jk
a

(a) v
i
v
j
v
k
, 1 s,
is involutive, with characters s
i
for 1 i n. By changing the chosen basis of V
if necessary, it can even be supposed that the ag such that V
i
V is spanned
by v
1
, . . . , v
i
is a regular ag for A
0
(and hence it will be regular for A
a
for all a in
a neighborhood O of 0 R
s
). For the rest of the proof, I use this basis to identify V
with R
n
.
Also, by the hypothesis that A is a Jacobi manifold, the linear equations for
quantities R

i
given as
T
i
kl
a

(a) R

j
+
T
i
lj
a

(a) R

k
+
T
i
jk
a

(a) R

l
=
_
T
i
mj
(a)T
m
kl
(a)+T
i
mk
(a)T
m
lj
(a)+T
i
ml
(a)T
m
jk
(a)
_
.
are solvable, and the associated homogeneous linear system for the R

i
has, for each
value of a, a solution space of dimension s
1
+2 s
2
+ +ns
n
. Thus, the equations
are compatible and have constant rank, so there exist real-analytic functions R

i
(a)
on a neighborhood of 0 R
s
(which can be supposed to be O) that furnish solutions
to the above inhomogeneous system.
Let M = GL(n, R) R
n
O, where O R
s
is the neighborhood of 0 R
s
selected above. Let p : M GL(n, R), x : M R
n
, and a : M O be the
respective projections. Set
i
= p
i
j
dx
j
and

= da

i
(a)
i
.
Now let I be the ideal on M generated by the 2-forms

i
= d
i
+
1
2
T
i
jk
(a)
j

k
.
Note that there exist 1-forms
i
j
such that
i
=
i
j

j
and such that the 1-forms

i
j
,
i
, and

are linearly independent and hence dene a coframing on M.

Now, by the way the functions R

i
on O were chosen, one has
d
i
=
1
2
T
i
jk
a

(a)

j

k
+T
i
jk
(a)
j

k
,
so that I is generated algebraically by the 2-forms
i
=
i
j

j
and the 3-forms

i
=
T
i
jk
a

(a)

j

k
.
Since A is involutive, the integral elements in V
n
(I) dened at each point of M
by
i
j
=

= 0 are all Cartan-ordinary. By the Cartan-Kahler Theorem, there

is an n-dimensional integral manifold I tangent to this integral element at the
point (I
n
, 0, 0) M.
NOTES ON EDS 19
This integral manifold is written as a graph of the form
_
p
i
j
(x), x, a

(x)
_
for x in
a neighborhood of 0 R
n
. Now, setting
i
= p
i
j
(x) dx
j
, one sees that the structure
function of the coframing = v
i

i
is
C =
1
2
T
i
jk
_
a

(x)
_
v
i
v
j
v
k
,
which takes values in A and, in particular, takes the value a
0
A at x = 0.
Remark 11 (Checking the hypotheses). Note that, in practical terms, checking the
condition that A V
2
(V

) be an involutive Jacobi manifold can be reduced

to a relatively simple calculation:
A coframing satisfying the structure equations (3.11) will necessarily satisfy
0 = d(d) =
1
2
dC ( ) +
1
2
C
_
C( ) ),
or, relative to a basis v
i
of V with dual basis v
i
,
0 =
1
2
dC
i
jk

j

k
+
1
6
(C
i
mj
C
m
kl
+C
i
mk
C
m
lj
+C
i
ml
C
m
jk
)
j

k

l
.
Regarding the v
i
as linear coordinates on V and regarding the C
i
jk
= C
i
kj
as the
components of the embedding of A into V
2
(V

), one can consider the algebraic

ideal I
A
generated on M = A V by the 3-forms

i
=
1
2
dC
i
jk
dv
j
dv
k

1
6
(C
i
mj
C
m
kl
+C
i
mk
C
m
lj
+C
i
ml
C
m
jk
) dv
j
dv
k
dv
l
.
(N.B.: Just this once, I do not want to consider the dierential closure of I
A
.)
Then I
A
has an integral element E of dimension n based at (a, 0) A V
on which the dv
i
are independent if and only if (3.12) is satised. Moreover, this
integral element is Cartan-ordinary if and only if T
a
A is an involutive subspace
of V
2
(V

).
Remark 12. It will turn out that the s
i
for a involutive Jacobi manifold A have a
signicance for describing the dierential invariants of V -valued coframings taking
values in A. As will be shown below, in an appropriate sense, the V -valued cofram-
ings whose structure functions take values in A depend (modulo dieomorphism)
on s
1
functions of 1 variable, s
2
functions of 2 variables, etc.
4. Ordinary prolongation
It is time to take a closer look at the geometry of V
o
n
(I).
4.1. The tableau of an ordinary element. Recall that the basepoint pro-
jection : V
o
n
(I) M is a smooth submersion, so the ber over x, which is
V
o
n
(I) Gr
n
(T
x
M), is a smooth submanifold of Gr
n
(T
x
M). For a given E V
o
n
(I),
the tangent space to this ber is an involutive tableau
A
E
T
E
Gr
n
(T
x
M)
_
T
x
M/E
_
E

of dimension s
1
(E) +2s
2
(E) + +ns
n
(E), and its Cartan characters are given by
s
i
(A
E
) = s
i
(E) +s
i+1
(E) + +s
n
(E).
20 R. BRYANT
4.2. The ordinary prolongation of I. Set M
(1)
= V
o
n
(I). Dene a subbun-
dle C T

M
(1)
by letting
C(E) =

(E

),
where E

(E)
M is the annihilator of E T
(E)
M. This subbundle of rank
dimM n is known as the contact bundle on M
(1)
.
Let I
(1)
A

_
M
(1)
_
denote the dierential ideal generated by the sections
of C. The ideal I
(1)
on M
(1)
is known as the ordinary prolongation of I on M.
(Technically, the denition of the prolongation depends on the choice of n, but, in
nearly all applications, the choice of n is determined by the problem that I was
designed to study, so I will not make this part of the notation.)
Every Cartan-ordinary integral manifold f : N M has a canonical lift f
(1)
:
N M
(1)
, dened by f
(1)
(x) = f

(T
x
N) V
o
n
(I) = M
(1)
. It follows directly from
the denition that f
(1)
: N M
(1)
is an integral manifold of I
(1)
and, moreover,
any integral manifold h : N M
(1)
that is an integral of I
(1)
and has the property
that h : N M is an immersion is of the form h = f
(1)
, in fact, with f = h.
At the integral element level, every

E V
n
_
I
(1)
_
with

E T
E
M
(1)
such that

:

E T
(E)
M is injective actually satises

(

E) = E. Moreover, each such

E
is Cartan-ordinary, with Cartan characters
s
i
(

E) = s
i
(E) +s
i+1
(E) + +s
n
(E),
and with a ag

F = (

E
0
, . . . ,

E
n1
) of

E being regular if and only if the ag F =
(E
0
, . . . , E
n1
) with E
i
=

(

E
i
) is a regular ag of E.
4.3. The higher prolongations. In particular, one can repeat the prolongation
process, but, now considering M
(2)
V
o
n
_
I
(1)
_
to be the open subset consisting
of those

E that satisfy the transversality condition

(

E) = E (and retaining
the corresponding condition for all the higher prolongations, etc). This denes a
sequence of manifolds M
(k)
with ideals I
(k)
, such that
_
M
(0)
, I
(0)
_
= (M, I) while,
for k 1, the manifold M
(k)
is embedded as an open subset of V
o
n
_
I
(k1)
_
. By
induction, one sees that the ideal I
(k)
has Cartan characters
s
(k)
j
= s
j
+
_
k
1
_
s
j+1
+
_
k + 1
2
_
s
j+2
+ +
_
k +n j 1
n j
_
s
n
.
One should think of M
(k)
as the space of k-jets of n-dimensional Cartan-ordinary
integral manifolds of I in the sense that two Cartan-ordinary integral manifolds f :
N M and g : N M represent the same k-jet of an integral manifold at x N
if and only if f
(k)
(x) = (g h)
(k)
(x) for some dieomorphism h : N N such that
h(x) = x.
Note that
dimM
(k)
= n +
_
k
0
_
s
0
+
_
k + 1
1
_
s
1
+
_
k + 2
2
_
s
2
+ +
_
k +n
n
_
s
n
,
which is what one would expect for a solution space that depends on s
0
constants,
s
1
functions of 1 variable, s
2
functions of 2 variables, . . ., and s
n
functions of n
variables.
NOTES ON EDS 21
4.4. Prolonging Cartan structure equations. This idea can also be applied
to understanding the dierential invariants of the solutions to a system of Cartan
structure equations such as (3.5). Starting with these equations, one can augment
them with a system for the b

, namely
(4.1) db

=
_
G

i
(a, b) +H

i
(a, b)c

i
where the the functions H

i
for 1 dimA(a, b)
(1)
are a basis for the rst
prolongation space of the tableau A(a, b), i.e., they give a basis for the solutions of
the homogeneous equations
F

i
b

(a, b)h

j

F

j
b

(a, b)h

i
= 0.
Using Cartans ideas, it is not dicult to show that, if the system (3.5) satises
the hypotheses of Theorem 3, then the prolonged system of structure equations
consisting of (3.5) and (4.1) will also satisfy the hypotheses of Theorem 3 and that
the Cartan characters of the tableau of the prolonged system will be
s
(1)
i
= s
i
+s
i+1
+ +s
n
.
In particular, in this case, for any given (a
0
, b
0
, c
0
) there will exist an augmented
coframing (a, b, c, ) satisfying the prolonged structure equations for which (a, b, c)
assumes the value (a
0
, b
0
, c
0
).
This leads naturally to the notion of dierential invariants for distinguish-
ing augmented coframings (a, b, ) satisfying (3.5) up to dieomorphism. Re-
call that two such coframings (a, b, ) on M
n
and ( a,

b, ) on

M
n
are equiva-
lent up to dieomorphism if there exists a dieomorphism h :

M M satis-
fying ( a,

b, ) = h

(a, b, ). Obviously, this will imply that, if db

= b

i

i
and
d

=

b

i

i
, then

b

i
= h

(b

i
) and similarly for all of the derivatives of the b

j
expanded in terms of the
i
.
Following Cartans terminology, one often speaks of the a

as the primary (or

fundamental ) invariants of the augmented coframing and the b

and b

i
, etc. as
derived invariants. (Here invariant means invariant under dieomorphism equiv-
alence.)
Thus, the import of Theorem 3 is that one sees that, in addition to being able to
freely specify the values of the s primary invariants (i.e., the a

) of an augmented
coframing (a, b, ) satisfying (3.5) at a point, one can also freely specify their rst
derived invariants (i.e., the b

), which are r = s
1
+s
2
+ +s
n
in number, at the
point, and freely specify a certain number of second derived invariants (i.e., the c

)
which are r
(1)
= s
1
+ 2s
2
+ +ns
n
in number, at the point, and so on.
Applying prolongations successively, one sees that the number of freely speci-
able dierential invariants of augmented coframings satisfying (3.5) of derived order
less than or equal to k is equal to
s +
_
k
1
_
s
1
+
_
k + 1
2
_
s
2
+ +
_
k +n 1
n
_
s
n
.
In a sense that can be made precise, this is the dimension of the space of k-jets of
dieomorphism equivalence classes of augmented coframings satisfying (3.5).
It is in this sense that one can assert that, up to dieomorphism, the gen-
eral augmented coframing satisfying a given involutive system of Cartan structure
22 R. BRYANT
equations depends on s
1
functions of 1 variable, s
2
functions of 2 variables, and so
on.
Similar remarks apply to the structure equations of Theorem 4. In fact, the
rst prolongation of these structure equations yield structure equations to which
Theorem 3 applies, so that one could have simply quoted Theorem 3 to prove
Theorem 4. This may make the reader wonder why this latter theorem is useful.
The reason is this: It is often simpler to check the hypotheses of Theorem 4 for a
given set of structure equations than it is to check the hypotheses of Theorem 3 for
the prolonged set of structure equations (as the reader will see in the examples).
4.5. Non-ordinary prolongation and the Cartan-Kuranishi Theorem. In
most cases, V
n
(I) does not consist entirely of Cartan-ordinary integral elements,
and even when the open subset V
o
n
(I) V
n
(I) is not empty, one is often interested
in at least some components of the complement and would like to know when there
exist integral manifolds tangent to these non-ordinary integral elements.
Cartans prescription for treating this situation was to prolong the non-ordinary
integral elements as well: Let M
(1)
V
n
(I) be any submanifold of V
n
(I) (in most
applications, it will be a component of a smooth stratum of V
n
(I) that does not lie
in V
o
n
(I)). Then, again, one can construct the ideal I
(1)
generated by the sections
of the contact subbundle C T

M
(1)
and one can consider V
n
_
I
(1)
_
, looking for
Cartan-ordinary integral elements of this ideal whose projections to M are injective.
If one nds them, then one has existence for integral manifolds tangent to these
non-regular integral elements. If one does not nd them, one can continue the
prolongation process as long as it results in ideals that have integral elements.
Cartan believed that continuing this process would always eventually result in
either an ideal with no integral elements of dimension n or else one that had Cartan-
ordinary integral elements. He was never actually able to prove this result, though.
Finally, a version of this prolongation theorem was proved by Kuranishi (in the
real analytic category, of course).
The hypotheses of the Cartan-Kuranishi Prolongation Theorem are somewhat
technical, so I refer you to Kuranishis original paper [11] for those. In practice,
though, one uses the Prolongation Theorem as a justication for computing suc-
cessively higher prolongations until one reaches involutivity (i.e., the existence of
Cartan-ordinary integral elements), which, nearly always, is what one must do
anyway in order to prove existence of solutions via Cartan-Kahler.
5. Some applications
There are many applications of these structure theorems in dierential geometry.
Here is a sample of such applications meant to give the reader a sense of how they
are used in practice. For further applications to dierential geometry, the reader
can hardly do better than to consult Cartans own beautiful collection of instructive
examples [9].
5.1. Surface metrics with |K|
2
= 1. Consider the metrics whose Gauss curva-
ture satises |K|
2
= 1. The structure equations are
d
1
=
12

2
d
2
=
12

1
d
12
= K
1

2

1

2

12
= 0,
NOTES ON EDS 23
where
dK = cos b
1
+ sinb
2
.
for some function b. (Here, b is the free derivative.)
Now d
2
= 0 is an identity for the forms in the coframing = (
1
,
2
,
12
), while
0 = d(dK) = (db
12
) (sinb
1
+ cos b
2
).
It follows that the hypotheses of Theorem 3 are satised, with the characters of
the tableau of free derivatives being s
1
= 1, s
2
= s
3
= 0. Thus, the general (local)
solution depends on one function of one variable.
The prolonged system will have
db =
12
+c (sinb
1
+ cos b
2
)
where c is now the new free derivative, etc.
(Of course, it is not dicult to integrate the structure equations in this sim-
ple case and nd an explicit normal form involving one arbitrary function of one
variable, but I will leave this to the reader.)
5.2. Surface metrics of Hessian type. Now, an application of Cartans original
theorem. The goal is to study those Riemannian surfaces (M
2
, g) whose Gauss
curvature K satises the second order system
Hess
g
(K) = a(K)g +b(K)dK
2
for some functions a and b of one variable.
Writing g =
1
2
+
2
2
on the orthonormal frame bundle F
3
of M, the structure
equations become
d
1
=
12

2
d
2
=
12

1
d
12
= K
1

2
dK = K
1

1
+K
2

2
and the condition to be studied is encoded as
_
dK
1
dK
2
_
=
_
K
2
K
1
_

12
+
_
a(K) +b(K) K
1
2
b(K) K
1
K
2
b(K) K
1
K
2
a(K) +b(K) K
2
2
__

2
_
.
Applying d
2
= 0 to these two equations yields
_
a

(K) a(K)b(K) +K
_
K
i
= 0 for i = 1, 2.
Thus, unless a

(K) = a(K)b(K)K, such metrics have K constant.

Conversely, suppose that a

(K) = a(K)b(K)K. The question becomes Does

there exist a solution (F
3
, ) to the following system?
d
1
=
12

2
d
2
=
12

1
d
12
= K
1

2

1

2

12
= 0,
where
_
_
dK
dK
1
dK
2
_
_
=
_
_
K
1
K
2
0
a(K) +b(K) K
1
2
b(K) K
1
K
2
K
2
b(K) K
1
K
2
a(K) +b(K) K
2
2
K
1
_
_
_
_

12
_
_
.
Since d
2
= 0 is formally satised for these structure equations, Theorem 2 applies
and guarantees that, for any constants (k, k
1
, k
2
), there is a local solution with the
invariants (K, K
1
, K
2
) taking the value (k, k
1
, k
2
).
24 R. BRYANT
In fact, the above equations show that, on a solution, the R
3
-valued func-
tion (K, K
1
, K
2
) either has rank 0 (if K
1
= K
2
= a(K) = 0) or rank 2. Moreover,
one sees that

_
a(K) +b(K)(K
1
2
+K
2
2
)
_
dK +K
1
dK
1
+K
2
dK
2
= 0,
so that the image of a connected solution lies in an integral leaf of this 1-form,
which only vanishes when K
1
= K
2
= a(K) = 0. Setting L = K
1
2
+K
2
2
, this
expression becomes
2
_
a(K) +b(K)L
_
dK + dL = 0,
which has an integrating factor: If (K) is a nonzero solution to

(K) = b(K)(K),
then
2(K)
2
a(K) dK + d
_
(K)
2
L) = 0,
so that the curvature map has image in a level set of the function F(K, K
1
, K
2
) =
(K)
2
(K
1
2
+K
2
2
) (K), where

(K) = 2(K)
2
a(K). (This function has critical
points only where K
1
= K
2
= a(K) = 0.)
On any solution (F
3
, ), the vector eld Y dened by the equations

1
(Y ) = (K)K
2
,
2
(Y ) = (K)K
1
,
12
(Y ) = (K)a(K),
is a symmetry vector eld of the coframing (since the Lie derivative of each of
1
,

2
,
12
with respect to Y is zero). It is nonvanishing on a solution of rank 2, and,
up to constant multiples, it is the unique symmetry vector eld of the coframing
on any connected solution.
For simplicity, I will only consider the case b(K) 0 in the remainder of this
discussion. In this case, a

(K) = K, so a(K) =
1
2
(C K
2
) for some constant C
and

(K) = 0, so one can take (K) 1.

The most interesting case is when C > 0, and, by scaling the metric g by a
constant, one can reduce to the case C = 1. Thus, the equations simplify to
_
_
dK
dK
1
dK
2
_
_
=
_
_
K
1
K
2
0
1
2
(1K
2
) 0 K
2
0
1
2
(1K
2
) K
1
_
_
_
_

12
_
_
.
and these functions satisfy
F(K, K
1
, K
2
) = K
1
2
+K
2
2
+
1
3
K
3
K = C
where C is a constant (dierent from the previous C, which is now normalized to 1).
There are two critical points of F, namely (K, K
1
, K
2
) = (1, 0, 0), and these
correspond to the surfaces whose Gauss curvature is identically +1 or identically
1. These clearly exist globally so it remains to consider the other level sets.
The level sets with C <
2
3
are connected and contractible, in fact, they can be
written as graphs of K as a function of K
1
2
+K
2
2
. C =
2
3
contains the critical
point (K, K
1
, K
2
) = (1, 0, 0), but away from this point, it is also a smooth graph.
When
2
3
< C <
2
3
, the level set has two smooth components, a compact 2-sphere
that encloses the critical point (1, 0, 0) and a graph of K as a smooth function
of K
1
2
+K
2
2
. The level set C =
2
3
is singular at the point (1, 0, 0), but, minus
this point, it has two smooth pieces, one bounded and simply connected, and one
unbounded and dieomorphic to RS
1
. For C >
2
3
, the level set is connected and
contractible.
NOTES ON EDS 25
According to the general theory, for each contractible component L of a (smooth
part of a) level set F = C, there will exist a simply-connected solution mani-
fold (F
3
, ) whose curvature image is L and whose symmetry vector eld Y is
complete. Moreover, the time-2-ow of the vector eld X
12
(i.e., the vector eld
that satises
1
(X
12
) =
2
(X
12
) = 0 while
12
(X
12
) = 1) is a symmetry of the
coframing and hence is the time-T-ow of Y for some T > 0. Dividing F by the
Z-action that this generates produces a solution manifold (

F, ) that is no longer
simply-connected but on which the ow of X
12
is 2-periodic, and this is the nec-
essary and sucient condition that

F be the oriented orthonormal frame bundle of
a Riemannian surface (M
2
, g) satisfying the desired equation.
However, for the components of the level sets that are dieomorphic to the 2-
sphere, this global existence result does not generally hold, i.e., the corresponding
solution manifold (F
3
, ) need not be the orthonormal frame bundles of complete
Riemannian surfaces (M
2
, g). I will explain why for the 2-sphere components of
the level sets F =
2
2/3 where > 0 is small.
Suppose that a connected solution manifold (F
3
, ) whose curvature map has,
as image, such a 2-sphere component is found and that the symmetry vector eld Y
as dened above is complete on it. Then the metric h =
1
2
+
2
2
+
12
2
must be
complete on F. Now, for small positive , one has that K is close to 1 while K
1
and
K
2
are close to zero, so it follows from a computation that the sectional curvatures
of h are all positive. In particular, the completeness of the metric on F
3
implies,
by Bonnet-Meyers, that it is compact, with nite fundamental group.
By passing to a nite cover, one can assume that F is simply connected. I
claim that the symmetry vector eld Y has closed orbits and that its ow generates
an S
1
-action on F. To see this, note that the map (K, K
1
, K
2
) : F R
3
submerses
onto the 2-sphere leaf. Hence the bers over the two points where K
1
= K
2
= 0
must be a nite collection of circles that are necessarily integral curves of the vector
eld Y , which has no singular points. In particular, the ow of Y on one of these
circles must be periodic, but, because the ow of Y preserves the coframing , if
some time T > 0 ow of Y has a xed point, then the time T ow of Y must be the
identity. Thus, the ow of Y is periodic with some minimal positive period T > 0,
so it generates a free S
1
-action on F. The quotient by this free S
1
-action is a
connected quotient surface that is a covering of the 2-sphere. Since this covering
must be trivial, the orbits of Y are the bers of the map (K, K
1
, K
2
) to the 2-sphere.
In particular, F, being connected and simply-connected, must be dieomorphic to
the 3-sphere.
Now, consider the vector eld X
12
on F as dened above. This vector eld is
(K, K
1
, K
2
)-related to the vector eld
K
2

K
1
+K
1

K
2
on R
3
whose ow is rotation about the K-axis with period 2.
It also follows that the ow of X
12
preserves the two circles that are dened
by K
1
= K
2
= 0. If (F, ) is to be a covering of the orthonormal frame bundle of
a Riemannian surface (M
2
, g), then X
12
must be periodic of period 2k for some
integer k > 0. As already remarked, by the structure equations, the 2-ow of X
12
,
say , is a symmetry of the coframing and hence must be the time R > 0 ow of Y
for some unique R (0, T].
26 R. BRYANT
Now, along each of the two circles in F dened by K
1
= K
2
= 0, one has
Y = a(K)X
12
= 0. The two points where K
1
= K
2
= 0 satisfy K = K

() where
K

() < 1 < K
+
() and
1
3
K

()
3
K

() =
2

2
3
. In fact, one nds expansions
K

() = 1
1
6

5
72

3

and this implies that
a
_
K

()
_
=
1
2
(1 K

()
2
) =
1
3

2
+ .
Thus the ratios of X
12
to Y on these two circles are not equal or opposite, and
hence Y cannot have the same period on these two circles, which is impossible.
Thus, there cannot be a global solution surface for such a leaf.
5.3. Prescribed curvature equations for Finsler surfaces. For an oriented
Finsler surface (M
2
, F), Cartan showed that the tangent indicatrix (i.e., the analog
of the unit sphere bundle) TM carries a canonical coframing (
1
,
2
,
3
)
generalizing the case of the unit sphere bundle of a Riemannian metric. It satises
structure equations
(5.1)
d
1
=
2

3
d
2
=
3

1
I
2

3
d
3
= K
1

2
J
2

3

3
= 0,
where I have written
3
for what would be
12
in the Riemannian case. The
functions I, J, and K are the Finsler structure functions.
One can check that Theorem 4 applies directly to these equations, with V of
dimension 3 and A V
2
(V

) an ane subspace of dimension 3 (and on which

I, J, and K are coordinates). The Cartan characters are s
1
= 0, s
2
= 2, and s
3
= 1.
Thus, the general Finsler surface depends on one function of 3 variables, which is
to be expected, since a Finsler structure on M is locally determined by choosing a
hypersurface in TM (satisfying certain local convexity conditions) to be the tangent
indicatrix TM. In fact, if = (
i
) is any coframing on a 3-manifold
3
that
satises (5.1) such that the space M of leaves of the system
1
=
2
= 0 can
be given the structure of a smooth surface for which the natural projection :
M is a submersion, then has a natural immersion : TM dened by
letting (u) =

_
X
1
(u)
_
for u , where X
1
is the vector eld on dual to
1
,
and, locally, this denes a Finsler structure on M.
Taking the exterior derivatives of (5.1), one nds that they satisfy identities (the
Bianchi identities of Finsler geometry)
(5.2)
dI = J
1
+I
2

2
+I
3

3
,
dJ = (K
3
+KI)
1
+J
2

2
+J
3

3
,
dK = K
1

1
+K
2

2
+K
3

3
.
for seven new functions I
2
, I
3
, . . ., K
3
. These are the free derivatives of the structure
theory. As expected from the general theory, the tableau of free derivatives of
the prolonged system, i.e., (5.1) together with (5.2), is involutive with characters
s
1
= s
2
= 3 and s
3
= 1.
Now, by the structure equations (5.2), if I = 0, then J = 0 and K
3
= 0, so that
the Bianchi identities reduce to
dK = K
1

1
+K
2

2
,
NOTES ON EDS 27
which is simply the Riemannian case. Note that in this case, the tableau of free
derivatives has s
1
= s
2
= 1 while s
3
= 0, corresponding to the fact that Riemannian
surfaces depend locally on one function of 2 variables (up to dieomorphism).
One can, of course, study other curvature conditions. For example, the Landsberg
surfaces are those for which J = 0. They satisfy structure equations
(5.3)
dI = 0
1
+ I
2

2
+I
3

3
,
dK = K
1

1
+ K
2

2
KI
3
.
The tableau of free derivatives now has s
1
= s
2
= 2 and s
3
= 0, so that the general
Landsberg metric depends on 2 functions of 2 variables. (By the way, this is only a
microlocal description of the solutions; constructing global solutions is much more
dicult. However, it does suce to show how exible the microlocal solutions
are.)
Another common curvature condition is the K-basic condition, i.e., when, K,
the Finsler-Gauss curvature, is constant on the bers of the projection M.
This is the condition K
3
= 0, so that the structure equations become
(5.4)
dI = J
1
+I
2

2
+I
3

3
,
dJ = KI
1
+J
2

2
+J
3

3
,
dK = K
1

1
+K
2

2
+ 0
3
.
The tableau of free derivatives now has s
1
= s
2
= 3 and s
3
= 0, showing that these
Finsler structures depend on 3 functions of 2 variables.
Even more restrictive are the Finsler metrics with constant K. These satisfy
(5.5)
dI = J
1
+ I
2

2
+I
3

3
,
dJ = KI
1
+J
2

2
+ J
3

3
,
dK = 0
1
+ 0
2
+ 0
3
.
The tableau of free derivatives now has s
1
= s
2
= 2 and s
3
= 0, showing that
these Finsler structures depend on 2 functions of 2 variables. (For those who know
about characteristics, note that, in this case, a covector =
1

1
+
2

2
+
3

3
is
characteristic for this tableau if and only if
1
= 0. Thus, the arbitrary functions
are actually functions on the leaf space of the geodesic ow
2
=
3
= 0. This
suggests (and, of course, it turns out to be true) that these structures are actually
geometric structures on the space of geodesics in disguise.)
5.4. Ricci-gradient metrics in dimension 3. Here are some sample problems
from Riemannian geometry. In the following, for simplicity of notation, I will
consider only the 3-dimensional case, but the higher dimensional cases are not
much dierent.
Consider the problem of studying those Riemannian manifolds (M, g) for which
there exists a function f such that Ric(g) = (df)
2
+H(f) g, where H is a specied
function of one variable. Most metrics g will not have such a Ricci potential, and
it is not clear how many such metrics there are.
The problem can be set up in structure equations as follows: On the orthonormal
frame bundle F
6
M
3
of g, one has the usual rst structure equations
(5.6) d
i
=
ij

j
28 R. BRYANT
and the second structure equations (in dimension 3) can be written in the form
(5.7)
_
_
d
23
d
31
d
12
_
_
=
_
_

12

31

23

12

31

23
_
_

_
R
1
2
tr(R) I
3
_
_
_

2
_
_
where R = (R
ij
) is the symmetric matrix of the Ricci tensor. By hypothesis, there
exists a function f such that
R
ij
= f
i
f
j
+H(f)
ij
where
(5.8) df = f
1

1
+f
2

2
+f
3

3
.
The four functions (f, f
1
, f
2
, f
3
) will play the role of the a

in the structure equa-

tions. Since d(df) = 0, there exist functions f
ij
= f
ji
such that
(5.9) df
i
=
ij
f
j
+f
ij

j
.
The symmetry of R implies that the equations d(d
i
) = 0 are identities, but, when
one computes d(d
ij
) = 0, one nds that these relations can be written as
_
2(f
11
+ f
22
+f
33
) H

(f)
_
df = 0.
Thus, either df = 0, in which case f is constant (so that the metric is Einstein), or
else the relation
f
11
+f
22
+f
33

1
2
H

(f) = 0
must hold. So impose this condition, and rewrite the above equation in the form
(5.10) df
i
=
ij
f
j
+
_
b
ij
+
1
6
H

(f)
ij
_

j
.
where the (new) b
ij
= b
ji
are subject to the trace condition b
11
+ b
22
+ b
33
= 0.
These b
ij
will play the role of the b

in the structure equations.

Thus, the problem can be thought of as seeking coframings = (
i
,
ij
) and
functions (f, f
i
) on a 6-manifold F
6
that satisfy the equations (5.6), (5.7), (5.8),
and (5.10), where the b
ij
= b
ji
are subject to b
11
+b
22
+b
33
= 0.
The tableau of the free derivatives is involutive, with characters s
1
= 3, s
2
= 2,
and s
k
= 0 for 3 k 6. Moreover, the equations d(d
i
) = d(d
ij
) = d(df) = 0
are identities while the equations d(df
i
) = 0 are satisable in the form
db
ij
= b
ik

kj
b
kj

ki
+F (3f
i

j
+ 3f
j

i
2
ij
f
k

k
) +b
ijk

k
where F =
1
10
_
f
1
2
+f
2
2
+f
3
2
+H(f) +
1
3
H

(f)
_
and where b
ijk
= b
jik
= b
ikj
and
b
iik
= 0. Hence, there are 7 = s
1
+ 2 s
2
+ + 6 s
6
independent free derivatives of
the b
ij
, the maximum allowed by the characters of their tableau.
Thus, the hypotheses of Theorem 3 are satised. Consequently, when H is an
analytic function, the pairs (g, f) that satisfy Ric(g) = (df)
2
+ H(f) g depend on
2 functions of 2 variables (up to dieomorphism).
(For those who know about the characteristic variety: A nonzero covector =

i
+
ij

ij
is characteristic if and only if
ij
= 0 and
1
2
+
2
2
+
3
2
= 0. Thus,
the real characteristic variety is empty, so the solutions are all real analytic when
H is real analytic.)
More generally, one can consider the problem of studying those Riemannian
manifolds (M, g) for which there exists a function f such that
(5.11) Ric(g) = a(f) Hess
g
(f) +b(f) (df)
2
+c(f) g
NOTES ON EDS 29
where a, b, and c are specied functions of one variable and Hess
g
(f) = f is the
Hessian of f with respect to g, i.e., the quadratic form that is the second covariant
derivative of f with respect to the Levi-Civita connection of g. For example, when
a(f) = 1, b(f) = 0, and c(f) = (a constant), (5.11) is the equation for a
gradient Ricci soliton. For simplicity, in what follows, I will assume that a, b, and
c are real-analytic functions.
If a(f) b(f) 0, then (5.11) implies that g is an Einstein metric, and so the
only solutions (g, f) are ones for which c(f) is a constant. In particular, if c

(f)
is not identically vanishing, then the only solutions (g, f) are when g is Einstein
and f is a constant.
If a(f) 0 and b(f) > 0, one can reduce (5.11) to the case b(f) 1 (which was
treated above) by replacing (g, f) by
_
g, (f)
_
, where

(f)
2
= b(f). (Meanwhile,
when b(f) < 0, one can reduce to b(f) 1 by replacing (g, f) by
_
g, (f)
_
where

(f)
2
= b(f). The reader can easily check that the local analysis of this case is
essentially the same as the case a(f) 0 and b(f) 1, with a few sign changes.)
In the generic case, in which a is nonvanishing, one can reduce to the case
b(f) 0 by replacing (g, f) by
_
g, (f)
_
where is a function that satises

(x) > 0
and

(x) =
_
b(x)/a(x)
_

(x). Hence, I will consider only the case b(f) 0 in the

remainder of this discussion.
Thus, the equation to be studied is encoded with the same structure equa-
tions (5.6), (5.7), and (5.9) but now with the relations
R
ij
= a(f) f
ij
+c(f)
ij
,
where a is a nonvanishing function. The equations d(d
ij
) = d(df
i
) = 0 then turn
out to imply the relation
d
_
L(f)
a(f)
_
+
_
1 +
a

(f)
a(f)
2
_
dH(f) +
_
2a(f)c(f)c

(f)
_
a(f)
2
df = 0
where L(f) = f
11
+f
22
+f
33
and H(f) = f
1
2
+f
2
2
+f
3
2
. Taking the exterior deriva-
tive of this relation yields
_
a(f)a

(f) 2a

(f)
2
a(f)
3
_
df dH(f) = 0.
At this point, the study of these equations divides into cases, depending on
whether aa

2(a

)
2
vanishes identically or not.
If the function aa

2(a

)
2
does not vanish identically, then any pair (g, f) that
satises the original equation must also satisfy equations of the form
f
1
2
+f
2
2
+f
3
2
= h(f)
and
f
11
+f
22
+f
33
= a(f)l(f)
for functions l and h of a single variable that satisfy
l

(x) +
_
1 +
a

(x)
a(x)
2
_
h

(x) +
_
2a(x)c(x)c

(x)
_
a(x)
2
= 0.
The rst of these equations implies, upon dierentiation,
2f
ij
f
j
= h

(f)f
j
which, as long as h(f) > 0, gives three equations on the free derivatives f
ij
= f
ji
.
Moreover, the equation f
11
+f
22
+f
33
= a(f)l(f) is independent from these three.
30 R. BRYANT
This means that there is only a 2-parameter family of possible variation in the f
ij
.
In fact, the tableau of free derivatives in this case is involutive with s
1
= 2 and
all s
i
= 0 for i > 1, so that solutions of this system depend on at most
12
two
functions of one variable (three if you count the function h). Thus, the pairs (g, f)
that satisfy the above equation are rather rigid.
On the other hand, if aa

2(a

)
2
vanishes identically, then a(f) = 1/(c
0
+c
1
f)
for some constants c
0
and c
1
, not both zero.
If c
1
= 0, then, by scaling f, one can reduce to the case a(f) = 1 and the original
equation becomes
R
ij
= f
ij
+c(f)
ij
,
while the relation above becomes
f
11
+ f
22
+f
33
+f
1
2
+f
2
2
+f
3
2
c(f) + 2C(f) = ,
where C

(f) = c(f), and where is a constant. Adding this relation on the free
derivatives f
ij
yields a tableau of free derivatives that has s
1
= 3, s
2
= 2 and
s
j
= 0 for j > 2. Moreover, a short calculation reveals that this relation satises
the conditions of Theorem 3, so, up to dieomorphism, the local general pairs (g, f)
that satisfy a relation of the form Ric(g) = Hess
g
(f)+c(f)g (for a xed real-analytic
function c(f)) depend on two functions of two variables.
Meanwhile, if c
1
= 0, then by translating and scaling f, one can reduce to
the case a(f) = 1/f, and one gets a similar result, that, up to dieomorphism,
the local general pairs (g, f) (with, say f > 0) that satisfy a relation of the form
Ric(g) = (Hess
g
(f))/f +c(f)g (for a xed real-analytic function c(f)) also depend
on two functions of two variables.
5.5. Riemannian 3-manifolds with constant Ricci eigenvalues. In dimen-
sion 3, a dierent way of writing the structure equations on the orthonormal frame
bundle F
6
of (M
3
, g) is to write them in vector form as
(5.12) d =
and
(5.13) d = +
_
R
1
4
tr(R)I
3
_

t
+
t

_
R
1
4
tr(R)I
3
_
where = (
i
) takes values in R
3
(thought of as columns of real numbers of height 3)
and

g =
t
, while =
t
= (
ij
) takes values in so(3), the space of skewsym-
metric 3-by-3 matrices, and R =
t
R is the 3-by-3 symmetric matrix that represents
the Ricci curvature, i.e., R = (R
ij
) and

_
Ric(g)
_
= R
ij

i

j
.
5.5.1. The general metric. Setting V = R
3
so(3) (so that, again, n = 6), then =
(, ) is a V -valued coframing, and the above structure equations take the form d =

1
2
C(), where C takes values in a 6-dimensional ane subspace A V

2
(V

).
The exterior derivatives of these structure equations then give the compatibility
conditions: One has d(d) = 0, and, setting = dR +R R, one nds
(5.14) d(d) =
_

1
4
tr()I
3
_

t
+
t

_

1
4
tr()I
3
_
,
so A V
2
(V

), an ane subspace, is a Jacobi manifold. Inspection shows that

its tableau has characters s
0
= s
1
= 0, s
2
= s
3
= 3, and s
k
= 0 for k = 4, 5, 6. Now,
12
The reason for the at most is that I have not veried that the torsion is absorbable, so I
cannot claim that this prolonged system is involutive.
NOTES ON EDS 31
the three 3-forms d(d) place 21 restrictions on the 36 coecients of S Hom(V, R
6
)
in order that the equation S(, ) = 0 should dene an integral element. Since
21 = c
0
+c
1
+c
2
+c
3
+c
4
+c
5
, it follows that the tableau is involutive, so that A
is an involutive Jacobi manifold.
Thus, Theorem 4 yields the expected result that the general metric in dimension
3 modulo dieomorphism depends on 3 functions of 2 variables and 3 functions of
3 variables. Applying prolongation to the structure equations would yield that the
number of dierential invariants of the coframing of order at most k + 1 is
6

j=0
_
k +j 1
j
_
s
j
=
k(k + 1)(k + 5)
2
,
which is the classically known number of independent derivatives of the curvature
functions R
ij
of order at most k1 (as expected, since the R
ij
themselves are the
rst derivatives of the coframing ).
5.5.2. Constant Ricci eigenvalues. More interesting are the proper submanifolds
of A that are involutive Jacobi manifolds. For example, suppose that one wanted
to determine the generality (modulo dieomorphisms) of the space of metrics whose
Ricci tensor has constant eigenvalues. Thus, one takes the above structure equations
and imposes that
(5.15) R =
t
PCP = P
1
CP,
where C is a constant diagonal matrix with diagonal entries c = (c
1
, c
2
, c
3
) where c
1

c
2
c
3
and P lies in SO(3). Restricting R to take this form in the structure equa-
tions denes a (non-ane) submanifold B
c
A V
2
(V

) that has dimension 3

(and is dieomorphic to the quotient of SO(3) by its diagonal subgroup) when the
c
i
are distinct, dimension 2 (and is dieomorphic to RP
2
) when two of the c
i
are
equal, and has dimension 0 (and is a single point) when the c
i
are all equal.
One can write the structure equations in a relatively uniform way by setting
= P and = dPP
1
PP
1
=
t
, for then the above equations can be
written
0 = Pd(d)P
1
= (C C)
t
+
t
(C C)
and the three 3-forms in the skew-symmetric matrix on the righthand side of this
equation are seen to be

1
=
_
(c
3
c
1
)
2

2
(c
1
c
2
)
3

3
_

1

2
=
_
(c
1
c
2
)
3

3
(c
2
c
3
)
1

1
_

2

3
=
_
(c
2
c
3
)
1

1
(c
3
c
1
)
2

2
_

3
where = (
ij
) = (
ijk

k
). Note that the 1-forms
1
,
2
,
3
complete the compo-
nents of and to a basis on the frame bundle cross SO(3).
In particular, this formula yields that B
c
is a Jacobi manifold for any choice
of c = (c
1
, c
2
, c
3
) and that its tableau has rank 3 when the c
i
are distinct, rank 2
when exactly two of the c
i
are equal, and rank 0 when all of the c
i
are equal.
When all of the c
i
are equal, the tableau is trivial, and so there is a regular ag
(with characters s
i
= 0) by denition.
32 R. BRYANT
5.5.3. Three distinct, constant eigenvalues. When the c
i
are distinct, one sees that
there is a regular ag for the integral elements described by
i
= 0 with charac-
ters s
2
= 3 and s
i
= 0 otherwise. In fact, a hyperplane in this integral element
fails to be the end of a regular ag if and only if it is described by an equation of
the form =
1

1
+
2

2
+
3

3
= 0 with
1

3
= 0. Consequently, Theorem 4
applies, and it follows that, up to dieomorphism, Riemannian 3-manifolds with
distinct constant eigenvalues of the Ricci tensor depend on 3 arbitrary functions of
2 variables.
5.5.4. Two distinct, constant eigenvalues. However, when exactly two of the c
i
are
equal, there is no regular ag: One easily checks that the codimensions of the
polar spaces of a generic ag for this tableau are c
0
= c
1
= 0, while c
k
= 2
for k 2. However, the codimension of the space of integral elements is 9 >
c
0
+ c
1
+ c
2
+ c
3
+ c
4
+ c
5
, as the reader can check. Thus, when two of the c
i
are
equal, the 2-dimensional Jacobi manifold B
c
is not involutive.
This does not mean that there are not Riemannian metrics for which the Ricci
tensor has two distinct, constant eigenvalues. To check this, though, one must
prolong the structure equations and use Theorem 3 instead of Theorem 4 as follows:
Suppose that Ric(g) =
t
R has two distinct constant eigenvalues, say c
1
= c
2
and c
2
(of multiplicity 2). This means that there is a circle bundle F
4
over M
3
consisting of the g-orthonormal coframes such that Ric(g) = c
1

1
2
+c
2
_

2
2
+
3
2
_
.
As the reader can check, this implies that the structure equations on F can be
written in the form
(5.16)
d
1
= 2a
1

2

3
d
2
=
23

3

_
a
2

2
+ (a
1
+a
3
)
3
_

1
d
3
=
23

2
+
_
(a
1
a
3
)
2
+a
2

3
_

1
d
23
= c
2

2

3
where a
1
, a
2
, a
3
are functions satisfying a
1
2
a
2
2
a
3
2
=
1
2
c
1
and the relations
(5.17)
da
1
= 2b
3

2
+ 2b
4

3
da
2
= 2a
3

23
+ (b
4
+b
1
)
2
+ (b
3
+b
2
)
3
da
3
= 2a
2

23
(b
3
b
2
)
2
+ (b
4
b
1
)
3
for some functions b
1
, b
2
, b
3
, and b
4
.
Conversely, given an augmented coframing (a, ) satisfying the structure equa-
tions (5.16) and (5.17) and a
1
2
a
2
2
a
3
2
=
1
2
c
1
, the form g =
1
2
+
2
2
+
3
2
denes
a metric on the space of leaves of
i
= 0 that satises Ric(g) = c
1

1
2
+c
2
_

2
2
+
3
2
_
.
Now, because d(a
1
2
a
2
2
a
3
2
) =
1
2
d(c
1
) = 0, the b
i
must satisfy the relations
a
2
b
1
+ a
3
b
2
(a
3
+2a
1
) b
3
+ a
2
b
4
= 0,
a
3
b
1
+ a
2
b
2
+ a
2
b
3
+ (a
3
2a
1
) b
4
= 0,
so that there are really only two free derivatives among the b
i
, as these two relations
are independent except when (a
1
, a
2
, a
3
) = (0, 0, 0) (and this can only happen if
c
1
= 0; but when c
1
= 0, I will remove the locus where the a
i
all vanish from further
consideration).
NOTES ON EDS 33
The reader can check that there exist 1-forms
i
db
i
mod {
1
,
2
,
3
,
23
} such
that
a
2

1
+ a
3

2
(a
3
+2a
1
)
3
+ a
2

4
= 0,
a
3

1
+ a
2

2
+ a
2

3
+ (a
3
2a
1
)
4
= 0,
and such that the relations
d(da
1
) 2
3

2
+ 2
4

3
d(da
2
) + (
4
+
1
)
2
+ (
3
+
2
)
3
d(da
3
) (
3

2
)
2
+ (
4

1
)
3
are identities modulo the above structure equations.
Meanwhile, the tableau of free derivatives is involutive, with s
1
= 2 and s
i
= 0
for i > 2. Thus, Theorem 3 applies, and one sees that the general such metric
depends on 2 functions of 1 variable.
(For those who know about the characteristic variety, one can compute that a
covector is characteristic i it is of the form =
2

2
+
3

3
where (
2
,
3
) satisfy
(a
1
+a
3
)
2
2
2a
2

3
+ (a
1
a
3
)
3
2
= 0
In particular, the characteristic variety consists of two complex conjugate points
when c
1
> 0, a double point when c
1
= 0, and two real distinct points when
c
1
< 0. Consequently, the metrics with c
1
> 0 will be real-analytic in harmonic
coordinates.)
5.6. Torsion-free H-structures. This last set of examples are applications to the
geometry of H-structures on n-manifolds.
Let m be a vector space over R of dimension m, and let H GL(m) be a
connected Lie subgroup of dimension r with Lie algebra h gl(m) = mm

.
One is interested in determining the generality, modulo dieomorphism, of the
(local) H-structures that are torsion-free, and, more generally, of torsion-free con-
nections on m-manifolds with holonomy contained in (a conjugate of) H.
Remark 13. When the rst prolongation space of h vanishes, i.e., when
h
(1)
= (h m

)
_
mS
2
(m

)
_
= (0),
these two questions are essentially the same, since, in this case, an H-structure that
is torsion-free has a unique compatible torsion-free connection, while a torsion-free
connection on M whose holonomy is conjugate to a subgroup K H denes an
P/N-parameter family of torsion-free H-structures, where P GL(m) is the group
of elements p GL(m) such that p
1
Kp H, while N H is the group of elements
such that p
1
Kp = K.
Now, the geometric objects being studied are the H-structures : B M
m
endowed with a torsion-free compatible connection. Letting : TB m be the
canonical m-valued 1-form on B, then the torsion-free compatible connection denes
an h-valued 1-form : TB h satisfying the rst structure equation
(5.18) d = ,
and having the equivariance R

h
() = Ad(a
1
)
_

_
for all h H.
One then has the second structure equation
(5.19) d = +
1
2
R( )
for a unique curvature function R : B h
2
(m

).
34 R. BRYANT
Conversely, any manifold B endowed with a coframing
= (, ) : TB mh = V
satisfying the equations (5.18) and (5.19) for some function R : B h
2
(m

)
is locally dieomorphic to the canonical coframing constructed above from the
data of an H-structure on a manifold M endowed with a compatible, torsion-free
connection.
Now, because d(d) = 0, the function R satises the rst Bianchi identity,
0 = d(d) = d + d = (d + ) =
1
2
R( ) = 0.
I.e., R takes values in the kernel K
0
(h) h
2
(m

) of the natural map

h
2
(m

) mm

2
(m

) m
3
(m

).
(This is the algebraic content of the rst Bianchi identity.)
In particular, the combined structure equations (5.18) and (5.19) dene a system
of equations for the coframing = (, ) taking values in V = mh for which the
structure function is required to take values in an ane space A
h
V
2
(V

)
that is modeled on the linear subspace K
0
(h) h
2
(m

) V
2
(V

).
Dierentiating (5.19) yields, after some algebra, the second Bianchi identity
0 = d(d) =
1
2
_
dR +

0
()R
_
( ),
where
0
: H GL
_
K
0
(h)
_
is the induced representation of H on K
0
(h), and

0
:
h gl
_
K
0
(h)
_
is the induced map on Lie algebras. This means that
dR =

0
()R +R

(),
where R

: B K
0
(h) m

takes values in the kernel K

1
(h) K
0
(h) m

of the
natural linear mapping dened by skew-symmetrization
K
0
(h) m

h
2
(m

) m

h
3
(m

).
This is the algebraic content of the second Bianchi identity.
In particular, A
h
is a Jacobi manifold, and it is natural to ask when it is in-
volutive, which is a condition on the Lie algebra h gl(m). In fact, the test for
involutivity is quite simple in this case: One computes the characters s
i
of K
0
(h)
considered as a tableau in h
2
(m

). Then Cartans bound implies that

dimK
1
(h) s
1
+ 2 s
2
+ +ms
m
with equality if and only if K
0
(h), and, consequently, A
h
are involutive. Thus, this
is a purely algebraic calculation.
Example 4 (Riemannian metrics). In the case that H = SO(m), the structure
equations take the familiar form
d
i
=
ij

j
with
ij
=
ji
satisfying
d
ij
=
ik

kj
+
1
2
R
ijkl

k

l
,
where the components of the Riemann curvature function R
ijkl
satisfy the famil-
iar relations R
ijkl
= R
jikl
= R
ijlk
and R
ijkl
+ R
iklj
+ R
iljk
= 0. For the
NOTES ON EDS 35
tableau K
0
_
so(m)
_
, the character s
p
when 1 p m is the number of indepen-
dent quantities R
ijkp
subject to the above relations that have 1 k < p, which one
nds to be
s
p
=
1
2
m(p 1)(mp + 1).
(Of course, s
p
= 0 for m < p <
1
2
m(m+1).) As expected,
s
1
+ +s
m
=
1
12
m
2
(m
2
1) = dimK
0
_
so(m)
_
and one also nds
s
1
+ 2 s
2
+ +ms
m
=
1
24
m
2
(m
2
1)(m+2) = dimK
1
_
so(m)
_
,
as this latter number is the number of independent R

ijklq
that show up in the
formulae for the derivatives of the R
ijkl
:
dR
ijkl
= R
qjkl

qi
R
iqkl

qj
R
ijql

qk
R
ijkq

ql
+R

ijklq

q
,
which are subject to the classical second Bianchi identity R

ijklq
+R

ijqkl
+R

ijlqk
= 0.
Thus, as expected, A
so(m)
is involutive, and the Riemannian metrics in dimen-
sion m (up to dieomorphism) depend on s
m
=
1
2
m(m1) functions of m variables.
The above characters then determine the number of independent covariant deriva-
tives of the curvature functions to any given order of dierentiation.
Example 5 (Ricci-at Kahler surfaces). When H = SU(2) GL(4, R), one is, in
eect, considering Riemannian 4-manifolds with holonomy contained in SU(2). In
this case, one nds that dimK
0
(h) = 5 and that the representation
0
of SU(2) is
irreducible. Indeed, one nds that the structure equations take the form
_
_
_
_
d
0
d
1
d
2
d
3
_
_
_
_
=
_
_
_
_
0
1

2

3

1
0
3

2

2

3
0
1

3

2

1
0
_
_
_
_

_
_
_
_

3
_
_
_
_
and
_
_
d
1
d
2
d
3
_
_
=
_
_
2
2

3
2
3

1
2
1

2
_
_
+
_
_
R
11
R
12
R
13
R
21
R
22
R
23
R
31
R
32
R
33
_
_
_
_

2
_
_
,
where R
ij
= R
ji
and R
11
+R
22
+R
33
= 0.
It has already been shown that this denes a Jacobi manifold in V
2
(V

)
where V = R
4
su(2) R
7
, and its involutivity follows by inspection, since the
characters are visibly s
2
= 3, s
3
= 2, and s
k
= 0 all other k, and since the dimension
of K
1
(h) is easily computed to be 12 = 2s
2
+ 3s
3
.
Thus, Theorem 4 applies and justies Cartans famous assertion that metrics in
dimension 4 with holonomy SU(2) depend on s
3
= 2 arbitrary functions of three
variables up to dieomorphism.
13
Example 6 (Segre structures of dimension 2m). One can also apply these theorems
to the study of higher order H-structures, i.e., structures for which there is no
canonical connection until after a prolongation has been performed.
13
Les espaces de Riemann precedents dependent de deux fonctions arbitraires de trois
arguments... ([6], pp. 5556). As far as I know, Cartan never gave any justication for this
assertion, which is the earliest case I know of in which an irreducible holonomy group is discussed,
other than the case of symmetric spaces. It seems highly likely to me, though, that he was already,
at that time (1926), aware of some version of Theorem 4.
36 R. BRYANT
Consider the generality of torsion-free GL(2, R) GL(m, R)-structures on R
2m
. In
this discussion, Im going to assume that m > 2, since the case m = 2 is equivalent
to conformal structures of type (2, 2) on R
4
, which (as Ill point out below) turns
out to have a dierent set of structure equations.
If F U R
2m
is a torsion-free GL(2, R) GL(m, R)-structure on U R
2m
,
then there is a prolongation of F to a second-order structure F
(1)
, with structure
group a semi-direct product of GL(2, R) GL(m, R) with R
2m
, on which there exists
a Cartan connection with values in SL(m+2, R), say
=
_

i
j

i

_
,
where the index ranges are understood to be 1 i, j, k 2 and 1 , , m,
and the forms that are entries of satisfy the single trace relation
i
i
+

= 0
but are otherwise linearly independent. These components are required to satisfy
structure equations of the form
14
d

j
=

i

i
j
d
i
j
=
i
k

k
j

i

j
d

i

i

+F

2
d
i

=
i
j

j

+G
i

2
dF

= F

+F

+F

+F

+R
i

i
dG
i

= G
j

i
j
+G
i

+G
i

+G
i

+Q
ij

j
.
The functions F, G, R, and Q must satisfy the relations
F

= F

= F

, F

= 0,
G
i

= G
i

= G
i

as well as the relations

R
i

= P
i

+
1
m+3
_

G
i

G
i

G
i

(m+2)

G
i

_
,
where P
i

is fully symmetric in its lower indices and satises P

i

= 0. Finally,
Q
ij

must be fully symmetric in its lower indices.

Note that, in the application of Theorem 4, the 1-forms play the role of the

i
, the independent coecients in F and G play the role of coordinates on the
appropriate Jacobi manifold A, while the independent coecients in P and Q play
the role of coordinates on A
(1)
.
While the number n is actually (m+2)
2
1 = m
2
+4m+3, its also clear from the
structure equations that only the

i
are eectively involved in the computation of
the characters (since it is only these terms that appear with non-constant coecients
in the structure equations). Thus (modulo what should be thought of as Cauchy
characteristics), the eective dimension is n = 2m.
14
Here is where the assumption that m > 2 is important. The correct structure equations
for m = 2 have nontrivial curvature terms in the structure equations for d
i
j
, as the reader can
easily check. In fact, for m = 2, the structure equations as I have written them are the structure
equations for the so-called half-at conformal structures of type (2, 2), i.e., the ones for which
the self-dual part of the Weyl curvature vanishes.
NOTES ON EDS 37
As the reader can check, the formal d
2
= 0 conditions needed for Theorem 4 are
satised. Using the symmetries of the coecients, the dimensions
dimA = (m+2)
_
m+2
3
_

_
m+1
2
_
=
1
6
m(m+1)(m
2
+4m+1)
and
dimA
(1)
= (2m+4)
_
m+3
4
_
2
_
m+2
3
_
=
1
12
m(m+1)(m+2)(m
2
+5m+2)
are easily computed.
It remains to compute the characters, which turn out to be
s
k
= (k1)
_
m
2
(k4)m2k + 3
_
for 1 k m+1 and s
k
= 0 for k > m+1. Thus, A is an involutive Jacobi
manifold.
In particular, up to dieomorphism, the general such torsion-free structure de-
pends on s
m+1
= m(m+1) functions of m+1 variables and there exists such a
structure taking any given desired curvature value.
Remark 14 (Torsion-free H-structures). For many other examples of this kind,
examining the generality up to dieomorphism of local torsion-free H-structures
for various groups H GL(m, R), the reader might consult [2] and [3]. Essentially
all questions about the existence and generality of local torsion-free structures of
this kind can be resolved by an application of Theorem 4.
Sometimes one wants to consider a proper submanifold of A
h
in order to investi-
gate H-structures with some extra condition on the curvature that captures some
geometric property.
Example 7 (Einstein-Weyl structures). Consider Cartans analysis of the so-called
Einstein-Weyl structures on 3-manifolds. These structures are CO(3)-structures
on 3-manifolds endowed with a compatible torsion-free connection whose curvature
function takes values in a certain 4-dimensional submanifold W A
co(3)
.
Here are their structure equations as Cartan writes them (with a very slight
change in notation):
_
_
d
1
d
2
d
3
_
_
=
_
_

0

3

2

3

0

1

2

1

0
_
_

_
_

3
_
_
and
_
_
_
_
d
0
d
1
d
2
d
3
_
_
_
_
=
_
_
_
_
0

2
_
_
_
_
+
_
_
_
_
2H
1
2H
2
2H
3
H
0
H
3
H
2
H
3
H
0
H
1
H
2
H
1
H
0
_
_
_
_
_
_

2
_
_
,
where the functions H
0
, H
1
, H
2
, and H
3
are coordinates on W. This is a set of
structure equations of the type to which Theorem 4 might apply, where the ane
subspace W V
2
(V

) has dimension 4 and where V = R

3
Rso(3) R
7
. It
is easy to verify that W is a Jacobi manifold and is involutive with s
2
= 4 and all
other s
k
= 0. Thus, Theorem 4 applies, and one recovers Cartans result that the
general Einstein-Weyl space depends on four arbitrary functions of two variables [8].
38 R. BRYANT
When A
h
is not involutive, one can ask whether its prolongation, which is got
by adjoining the equation
(5.20) dR =

0
()R +R

()
to the pair (5.18) and (5.19), is involutive, where R

takes values in the sub-

space K
1
(h) K
0
(h) m

that is the kernel of the natural mapping

K
0
(h) m

h
2
(m

) m

h
3
(m

).
The combined system of equations (5.18), (5.19), and (5.20) is of the type that
Theorem 3 was intended to treat, with R playing the role of the a

and R

playing
the role of the b

.
It may be necessary to repeat this prolongation process several times in order to
arrive at a system of structure equations to which Theorem 3 can be applied.
Example 8 (Bochner-Kahler metrics). An interesting example is when m = C
n
and H = U(n) GL(m). In this case, one nds that K
0
(h) is decomposable as a
U(n)-module into three irreducible summands,
K
0
(h) = S(h) Ric
0
(h) B(h),
where S(h) R corresponds to the space of curvature tensors of Kahler mani-
folds with constant holomorphic sectional curvature, Ric
0
(h) corresponds to the
space of traceless Ricci curvatures of Kahler metrics, and B(h), known as the space
of Bochner curvatures, corresponds to the space of curvature tensors of Ricci-at
Kahler manifolds. A Kahler metric is said to be Bochner-Kahler if the B(h)-
component of its curvature tensor vanishes, i.e., if its curvature tensor takes values
in Ric
0
(h) S(h).
This denes a Jacobi manifold A K
0
(h) that is not involutive, but, after a
succession of applications of the prolongation process (in fact, three prolongations),
one arrives at a set of structure equations that has no free derivatives but satises
the hypotheses of Theorem 2, thus showing that germs of Bochner-Kahler metrics
depend on a nite number of constants. For details, see [4].
References
[1] R. Bryant, et al, Exterior Dierential Systems, Springer-Verlag, 1991. 2
[2] , Classical, exceptional, and exotic holonomies: a status report, in Actes de la Table
Ronde de Geometrie Dierentielle (Luminy, 1992), Semin. Congr., 1 (1996), pp. 93165, Soc.
Math. France, Paris. 37
[3] , Recent advances in the theory of holonomy, Asterisque, 266 (2000), 351374 (Expose
No. 861). 37
[4] , Bochner-Kahler metrics, Journal of the AMS, 14 (2001), 623715. 38
[5]

E. Cartan, Sur la structure des groupes innis de transformations, Ann.

Ec. Norm. 21 (1904),
153206. (Especially, see paragraphs 1925.) 9, 10
[6] , Le geometrie des espaces de Riemann, Mem. Sci. Math. IX (1925), Gauthier-Villars,
Paris. 35
[7] , Les probl`emes dequivalence, uvres Compl`etes, Partie II, Volume 2, 13111334.
(Especially, see 13171321.) 16
[8] , Sur une classe despaces de Weyl, Ann.

Ec. Norm. 60 (1943), 116. 37
[9] , Les Syst`emes Dierentiels Exterieurs et leurs Applications Geometriques, Paris,
Hermann, 1945. 22
[10] E. Kahler, Einf urung in die Theorie der Systeme von Dierentialgleichungen, Hamburger
Math. Einzelschriften 16 (1934). 2
[11] M. Kuranishi, On

E. Cartans prolongation theorem of exterior dierential systems, Amer.
J. Math. 79 (1957), 147. 2, 22
NOTES ON EDS 39
Duke University Mathematics Department, PO Box 90320, Durham, NC 27708-0320
E-mail address: bryant@math.duke.edu
URL: http://www.math.duke.edu/
~
bryant