. 33
( 48 .)


’ n’2 δi § ωi .
dρ =

We therefore have a situation similar to that in §3.3 (cf. (3.40, 3.41)), with
equations formally like those in the conformal geometry equivalence problem of
§3.1.2. We can mimic the derivation of conformal structure equations in the
present case by ¬rst setting βi = n’2 δi , and then we know that this results in
an equation

d±i + ±i § ±k + βi § ωj ’ βj § ωi = 2 Ai ωk § ωl ,
j k j jkl

where the quantity (Ai ) has the symmetries of the Riemann curvature tensor.
Furthermore, we know that there are unique functions tij = tji such that replac-
ing βi ; βi + tij ωj will yield the preceding equation with Al = 0. However,
it will simplify matters later if we go back and replace instead
n’2 o j
δi ; δ i + 2 tij ω ,

where to = tij ’ n δij tkk is the traceless part (note that only a traceless addition
to δi will preserve the structure equations on B ’ M ). In terms of the new δi ,
we de¬ne
’ n Kωi ),
2 1
βi = (δ (4.7)
n’2 i

K = ’ n’2 tjj .
This K was chosen so that de¬ning βi by (4.7) gives the correct conformal
structure equation (4.6) with Al = 0; it re¬‚ects the di¬erence between the
pseudo-connection forms for the Poincar´-Cartan form and the Cartan con-
nection forms for the induced conformal geometry on the submanifold. One
interpretation is the following.
The function K on Bf is a fundamental invariant of a stationary
submanifold (N, [ds2]f ) ’ (M, Π) of an Euler-Lagrange system, and
may be thought of as an extrinsic curvature depending on up to third
derivatives of the immersion f.
In the classical setting, N is already the 1-jet graph of a solution z(x) of an Euler-
Lagrange equation, so an expression for K(x) depends on fourth derivatives of
Now suppose that our integral manifold f : N ’ M , with δi , βi , and K as
above, is the initial manifold F0 = f in a Legendre variation F : N —[0, 1] ’ M .
Then we can rewrite our formula (4.5) for the second variation as

g(dgi + nρgi ’ gj ±j +
δ 2 (FΛ )N0 (g) = ’ § ω(i) + g2 Kω .
2 gβi )

Part of this integrand closely resembles the expression (3.19) for the second co-
variant derivative of a section of a density line bundle, discussed in constructing
the conformal Laplacian in §3.1.3. This suggests the following computations.
First, consider the structure equation dθ = ’(n ’ 2)ρ § θ ’ πi § ωi . Along
W0 (but not yet restricted to W0 ), where θ = g dt and πi = gidt, this reads
dg § dt = ’(n ’ 2)ρ § g dt ’ gidt § ωi ,
so that restricted to W0 , we have
dg + (n ’ 2)ρg = giωi .
This equation should be interpreted on Bf , which is identi¬ed with the principal
bundle associated to the conformal structure [ds2]f . It says that g is a section
of the density bundle D 2n , and that gi are the components of its covariant
derivative (see (3.18) ¬.). We can now write
dgi + nρgi ’ gj ±j + = gij ωj ,
i 2

and by de¬nition
∆f g = gii ∈ “(D ),

where ∆f is the conformal Laplacian on N induced by f = F0 : N ’ M . We
now have a more promising version of the second variation formula:

δ 2 (FΛ )N0 (g) = ’ (g∆f g + g2 K)ω. (4.8)

It is worth noting that the sign of this integrand does not depend on the sign
of the variation™s generating function g, and if we ¬x an orientation of N , then
the integrand Kω on Bf has a well-de¬ned sign at each point of N .

4.1.3 Intrinsic Integration by Parts
In order to detect local minima using the second variation formula (4.8), it is
often helpful to convert an integral like g∆g dx into one like ’ || g||2dx. In
the Euclidean setting, with either compact supports or with boundary terms,
this is done with integration by parts and is straightforward; the two integrands
di¬er by the divergence of g g, whose integral depends only on boundary data.
We would like to perform a calculation like this on N , for an arbitrary
Legendre variation g, using only intrinsic data. In other words, we would like
to associate to any N and g some ξ ∈ „¦n’1(N ) such that dξ is the di¬erence
between (g∆f g)ω and some quadratic expression Q( g, g)ω, possibly with
some additional zero-order terms. A natural expression to consider, motivated
by the ¬‚at case, is

ξ = ggi ω(i) . (4.9)

Here, in order for gi to make sense, we are assuming that we have a local
coframing on M adapted to the integral manifold N , so that at points of N ,
θ = g dt, πi = gi dt, and so that restricted to N , θ|N = πi |N = 0. We can then

= dg § gi ω(i) + g dgi § ω(i) + ggi dωj § ω(ij)
dξ (4.10)
= (dg + (n ’ 2)ρg) § giω(i) + g(dgi + nρgi ’ gj ±j ) § ω(i) (4.11)

(gi )2 + g∆f g ω ’ n’2 2
§ ω(i) .
= 2 g βi (4.12)

Now, the ¬rst term is exactly what we are looking for, and second is fairly
harmless because it is of order zero in the variation g, and in practice contributes
only terms similar to g 2 Kω.
The problem is that ξ (4.9) is de¬ned on the total space Bf ’ N , and
although semibasic for this bundle, it is not basic; that is, there is no form on
N that pulls back to Bf to equal ξ, even locally. The criterion for ξ to be basic
is that dξ be semibasic, and this fails because of the appearance of βi in (4.12).
But suppose that we can ¬nd some canonical reduction of the ambient B ’
M to a subbundle on which δi becomes semibasic over M ; in terms of the Lie
algebra (4.1), this means that we can reduce to the subgroup having Lie algebra
given by {di = 0}. In this case, each βi = n’2 (δi ’ K ωi ) is a linear combination
of θ, ωi , πi , and is in particular semibasic over N . Consequently, ξ is basic over
N , and we can perform the integration by parts in an invariant manner.
Unfortunately, there are cases in which no such canonical reduction of B is
possible. An example is the homogeneous Laplace equation on Rn ,

∆z = 0,

which is preserved under an action of the conformal group SO o (n + 1, 1). The
associated conformal geometry on the trivial solution z = 0 is ¬‚at, and our
second variation formula reads

δ 2 (FΛ )0 (g) = g∆g dx

for a variation g ∈ C0 („¦), „¦ ‚ Rn . It follows from our construction that this
integrand is invariant under a suitable action of the conformal group. However,
the tempting integration by parts

|| g||2dx
g∆g dx = ’
„¦ „¦

leaves us with an integrand which is not conformally invariant. It is this phe-
nomenon that we would like to avoid.
To get a sense of when one might be able to ¬nd a canonical subbundle of
B ’ M on which the δ i are semibasic, recall that

(n ’ 2)dρ ≡ ’δi § ωi (mod {θ, πi}).

Working modulo {θ, πi} essentially amounts to restricting to integral manifolds
of the Euler-Lagrange system. The preceding then says that a choice of sub-
bundle B ‚ B on which δi are semibasic gives a subbundle of each conformal
bundle Bf ’ N on which dρ is semibasic over N . Now, typically the role of ρ in
the Cartan connection for a conformal structure is as a psuedo-connection in the
density line bundle D; a special reduction is required for ρ to be a genuine con-
nection, and the latter requirement is equivalent to dρ being semibasic. In other
words, being able to integrate by parts in an invariant manner as described above
is equivalent to having a connection in D represented by the pseudo-connection
ρ. One way to ¬nd a connection in D is to suppose that D has somehow been
trivialized, and this is equivalent to choosing a Riemannian metric representing
the conformal class. This suggests that Euler-Lagrange systems whose integral
manifolds have canonical Riemannian metrics will have canonical reductions of
this type.
In fact, we have seen an example of a Poincar´-Cartan form whose geometry
B ’ M displays this behavior. This is the system for Riemannian hypersurfaces
having prescribed mean curvature, characterized in §2.5 in terms of di¬erential
invariants of its neo-classical, de¬nite Poincar´-Cartan form. To illustrate the
preceding discussion, we calculate the second variation formula for this system.
The reader should note especially how use of the geometry of the Poincar´- e
Cartan form gives a somewhat simpler derivation of the formula than one ¬nds
in standard sources.1
1 See for example pp. 513“539 of [Spi75], where the calculation is prefaced by a colorful
warning about its di¬culty.

4.1.4 Prescribed Mean Curvature, Revisited
In §2.5, we considered a de¬nite, neo-classical Poincar´-Cartan form (M, Π)
whose associated geometry (B ’ M, •) had invariants satisfying

T ijk = 0, U ij = »δ ij .

We further assumed the open condition

» < 0,

and this led to a series of reductions of B ’ M , resulting in a principal sub-
bundle B3 ’ M , having structure group with Lie algebra
±« 

00 0
 
0  : ai + a j = 0 ,
 0 ai
g3 = j j i
 
0 0 ’ai

on which the original structure equations (4.2, 4.3, 4.4) hold, with U ij = ’δ ij ,
Sj = T ijk = 0, ρ = ’ 2n θ for a function H on B3 , and dρ = ’ n’2 δi § ωj , where
H 1

δi ≡ 0 (mod {θ, ωj , πj }). In this case we computed that dH ≡ 0 (mod {θ, ω i }),
so that H de¬nes a function on the local quotient space Qn+1, which also inherits
a Riemannian metric (ωi )2 . The contact manifold M can be locally identi¬ed
with the bundle of tangent hyperplanes of Q, and the integral manifolds of the
Euler-Lagrange di¬erential system EΛ are the tangent loci of hypersurfaces in Q
whose mean curvature coincides with the background function H. In this case
B3 ’ M ’ Q is locally identi¬ed with the orthonormal frame bundle of the
Riemannian manifold Q.
An important point here is that the Riemannian geometry associated to Π
only appears after reducing to B3 ’ M . However, if our goal is to see the
formula for the second variation, then we face the following di¬culty. That
formula required the use of coframes of M adapted to a stationary submanifold
N ’ M in a certain way, but while adapted coframes can always be found
in B ’ M , there is no guarantee that they can be found in the subbundle
B3 ’ M , where the Riemannian geometry is visible.
We will overcome these di¬culties and illustrate the invariant calculation of
the second variation by starting only with the Riemannian geometry of (Q, ds 2).
This is expressed in the Levi-Civita connection in the orthonormal frame bundle,
where we can also give the Poincar´-Cartan form and Euler-Lagrange system
for prescribed mean curvature. We then introduce higher-order data on a larger
bundle, which allows us to study the second fundamental form. In fact, this
larger bundle corresponds to the partial reduction B2 ’ M on which ρ and δi
are semibasic, but σij is not. The end result of our calculation is formula (4.16).
In the following discussion, index ranges are 0 ¤ a, b, c ¤ n and 1 ¤ i, j, k ¤ n.
We begin with a generalization of the discussion in §1.4 of constant mean
curvature hypersurfaces in Euclidean space. Let (Q, ds2) be an oriented Rieman-
nian manifold of dimension n + 1. A frame for Q is a pair f = (q, e) consisting

of a point q ∈ Q and a positively-oriented orthonormal basis e = (e0 , . . . , en)
for Tq Q. The set F of all such frames is a manifold, and the right SO(n + 1, R)-
a a a
(q, (e0 , . . . , en )) · (gb ) = (q, ( ea g0 , . . . , ea gn ))
gives the basepoint map
q:F ’Q


. 33
( 48 .)