NONLINEAR WAVE AND SCHRÖDINGER

EQUATIONS ON COMPACT LIE GROUPS

AND HOMOGENEOUS SPACES

MASSIMILIANO BERTI and MICHELA PROCESI

Abstract
We develop linear and nonlinear harmonic analysis on compact Lie groups and ho-
mogeneous spaces relevant for the theory of evolutionary Hamiltonian PDEs. A basic
tool is the theory of the highest weight for irreducible representations of compact
Lie groups. This theory provides an accurate description of the eigenvalues of the
Laplace-Beltrami operator as well as the multiplication rules of its eigenfunctions.
As an application, we prove the existence of Cantor families of small amplitude
time-periodic solutions for wave and Schrödinger equations with differentiable non-
linearities. We apply an abstract Nash-Moser implicit function theorem to overcome
the small divisors problem produced by the degenerate eigenvalues of the Laplace
operator. We provide a new algebraic framework to prove the key tame estimates for
the inverse linearized operators on Banach scales of Sobolev functions.

Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 480
2. Linear and nonlinear analysis on Lie groups . . . . . . . . . . . . . . . 490
3. Abstract Nash-Moser theorem with parameters . . . . . . . . . . . . . . 509
4. The bifurcation equation . . . . . . . . . . . . . . . . . . . . . . . . . 511
5. The range equation and the proof of Theorem 1.2 . . . . . . . . . . . . 513
6. Inversion of the linearized equation: Proof of Lemma 5.2 . . . . . . . . 516
7. The NLS equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536

DUKE MATHEMATICAL JOURNAL

Vol. 159, No. 3, 
c 2011 DOI 10.1215/00127094-1433403
Received 18 June 2009. Revision received 16 January 2011.
2010 Mathematics Subject Classification. Primary 37K55; Secondary 58C15, 58J45, 35Q55, 37G15.
Authors’ work supported by the European Research Council under FP7, “New Connections between Dynamical
Systems and Hamiltonian PDEs with Small Divisors Phenomena.”

479
480 BERTI and PROCESI

1. Introduction
The dynamics of nonlinear evolutionary Hamiltonian PDEs on a compact Riemannian
manifold strongly depends on its geometry. In this paper we consider the autonomous
nonlinear wave (NLW) equation

utt − u + μu = f (x, u) , x ∈ M, (1.1)

and nonlinear Schrödinger (NLS) equation

iut − u + μu = f (x, |u|2 )u , x ∈ M, (1.2)

where M is a compact Lie group or, more generally, a compact homogeneous space
(see Definition 2.2). The operator  is the Laplace-Beltrami operator defined with
respect to a Riemannian metric compatible with the group structure (see Section 2.2).
The mass μ > 0 is positive; the nonlinearity is finitely differentiable and vanishes at
u = 0 at least quadratically.
The existence problem of periodic and quasi-periodic solutions for Hamiltonian
PDEs has received considerable attention in the last twenty years. The main difficulty
is the presence of arbitrarily “small divisors” in the series expansion of the solutions.
For nonlinear wave and Schrödinger equations in one space dimension, a reasonably
complete answer is available. In higher spatial dimensions there is a good understand-
ing of periodic solutions when M= Tn and some results on quasi-periodic solutions
(see Bourgain [9], [11] for NLS and NLW; see Eliasson and Kuksin [19] for NLS), all
restricted to tori. For more general domains or in the case of Riemannian manifolds,
there are no results, as far as we know, even in the periodic case.
In this paper we prove the existence of Cantor families of small amplitude periodic
solutions for the wave and Schrödinger equations (1.1) and (1.2) on any homogeneous
space M , with differentiable nonlinearities (see Theorems 1.2, 1.3). These results hold
for fixed (not parameter dependent) nonlinearities. Theorem 1.2 is a PDE extension,
for the NLW, of the classical Lyapunov resonant center theorem. Theorem 1.3 is an
extension, for the completely resonant NLS, of the Weinstein-Moser theorem (see
[2]). A key tool for proving these results is the nonlinear harmonic analysis on Lie
groups developed in Section 2.
Examples of compact connected Lie groups are the standard torus Tn , the special
orthogonal group SO(n), the special unitary group SU(n), and so on. Examples of
compact homogeneous spaces are the spheres S n , the real and complex Grassmanians,
and the moving frames, namely, the manifold of the k -ples of orthonormal vectors
in Rn with the natural action of the orthogonal group O(n). We refer to [25] for the
particularly interesting case of symmetric spaces and to [12] and [37] for many other
examples.
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 481

The linear Schrödinger equation on compact Lie groups has been widely studied,
starting with the works of Weyl [41], in order to understand the rotational modes of
atoms and molecules (see, e.g., [35]). The linear Schrödinger equation on compact
homogeneous spaces appears also in high-energy physics, for example, in connection
with the quantum Hall effect and the Landau equation (see [39], [31], [27]).
Nonlinear Schrödinger equations on Lie groups can be seen as Hartree-Fock mean
field approximations (see [20]) of quantum many body lattice problems like the XY
or Heisenberg model (generalizations of the classical Ising model).
Linear and nonlinear wave equations on compact Lie groups and symmetric spaces
have been investigated (see, e.g., Helgason [24] – [25] and Taylor [38]).
For any compact Riemannian manifold, the spectrum of the Laplace-Beltrami
operator is discrete, its eigenvalues tend to infinity (with asymptotics given by the
Weyl theorem), and its eigenfunctions form an L2 -orthonormal basis. However, these
properties do not seem sufficient for proving the existence of periodic/quasi-periodic
solutions for the nonlinear equations (1.1) and (1.2). For instance, the product of two
eigenfunctions is not in general an eigenfunction (as for exponentials); neither are
finite combinations of them (as for spherical harmonics). From a dynamical point
of view, this could produce strong exchanges of energy between the modes. These
instability phenomena are also related to the possible existence of eigenfunctions of
the Laplacian normalized in the L2 -norm but with L∞ -norms not uniformly bounded.
Interestingly this really happens for certain Lie groups; nevertheless, it is compatible
with the existence of the periodic solutions of Theorems 1.2 and 1.3. This should be
seen as a consequence of the multiplication properties of the eigenfunctions inherited
by the symmetries given by the Lie group.
In this paper we restrict our attention to the search for periodic solutions, where
the small divisor problem simplifies. The proofs of Theorems 1.2 and 1.3 are based
on the following key elements:
(1) The abstract Nash-Moser implicit function theorem with parameters proved
in [5] (see Section 3): we find it convenient to use a Nash-Moser scheme,
because the eigenvalues of the Laplacian on M are highly degenerate. The main
difficulty is to prove the “tame” estimates of the inverse linearized equations
in Sobolev norm (see Section 6).
(2) The spectral and harmonic analysis of the Laplace-Beltrami operator on com-
pact Lie groups and homogeneous spaces: particularly relevant are the mul-
tiplication rules of eigenfunctions (see Section 2.4.1 and Propositions 2.14,
2.19) and the properties of the eigenvalues stated in Theorem 2.4.
The harmonic analysis in step (2) should provide many key elements also for
proving the existence of quasi-periodic solutions of (1.1) and (1.2) via KAM the-
ory (see [19]) or the Bourgain’s approach in [11]. Interestingly, many tools used in
Theorems 1.2 and 1.3 resemble the Birkhoff normal form techniques developed by
482 BERTI and PROCESI

Bambusi, Delort, Grebért, and Szeftel in [1] and [18] for PDEs on spheres and Zoll
manifolds. Moreover, we feel that there is a deep connection with the methods used
by Burq, Gérard, and Tzvetkov [14], [13] in order to prove well-posedness results of
the initial value problem on compact manifolds (see Bourgain [10] for the torus).
Before describing our techniques more in detail, we present some literature on
periodic and quasi-periodic solutions for Hamiltonian PDEs.

Short history
The first pioneering existence results of quasi-periodic solutions have been proved by
Kuksin [28] and Wayne [40] for 1-dimensional, analytic PDEs on an interval with
Dirichlet boundary conditions, via KAM theory. A key point is that, in this case, the
eigenvalues of the Laplacian are simple (see also Kuksin [29] and Pöschel [33], [30]).
For the existence of time periodic solutions of 1-dimensional PDEs with periodic
boundary conditions (i.e., M = T1 ), a more direct bifurcation approach, which does
not exclude multiplicity of the normal frequencies, was developed by Craig and Wayne
in [17] (see also [15], [16]). They applied a Lyapunov-Schimdt decomposition and
solved the small divisor problem with a Newton iterative scheme. The key step is to
prove the invertibility of the linearized equations in a neighborhood of the equilibrium,
together with estimates of the inverses in analytic norms. The main difficulty comes
from the fact that these linear PDEs have nonconstant coefficients. The Craig-Wayne
approach in [17] is based on a Fröhlich-Spencer [23] coupling-type technique. The
key ingredients are the following.
(i) The “singular sites”—namely, the Fourier indexes of the small divisors—are
more and more “separated at infinity” in clusters of uniform size.
(ii) The eigenfunctions are “localized with respect to the exponentials.”
Property (i) holds for 1-dimensional wave and Schrödinger equations. Property (ii)
holds for the eigenfunctions of Sturm-Liouville operators −∂xx + V (x) on T1 . It
implies, in particular, that the multiplication operator by an analytic function is rep-
resented, in the eigenfunction basis, as a matrix with exponentially fast off-diagonal
decay.
For higher-dimensional PDEs, a major further difficulty is that the dimensions
of the eigenspaces of the Laplacian increase to infinity. Generalizing the Craig-
Wayne approach, the first existence results of periodic solutions for analytic wave
and Schrödinger equations on Tn , n ≥ 2, have been proved by Bourgain in [8] and
[9]. (Actually, [9] proves also the existence of quasi-periodic solutions when n = 2.)
The key point is to relax the separation property (i), proving the following.
(i) A partition of the singular sites in huge clusters, with unbounded dimensions,
separated at infinity.
This is achieved by exploiting the fact that the eigenvalues of the Laplacian on Tn are
sums of integer squares.
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 483

Recently this result has been generalized by Berti and Bolle [4] for finitely
differentiable nonlinearities. The proof is based on a Nash-Moser implicit function
scheme, generalized into an abstract setting in [5]. A key role in the proof is played
by a Fröhlich-Spencer coupling technique in Sobolev spaces.
Existence of periodic solutions for analytic PDEs on Tn (or rectangles) has been
also proved by Gentile and Procesi [21], solving the small divisor problems with
Lindstedt series techniques.
Before concluding this brief survey of results, we mention that the existence of
quasi-periodic solutions for wave and Schrödinger equations with Fourier multipliers
on Tn has been proved by Bourgain in [11]. Moreover, the KAM approach has been
recently extended by Eliasson and Kuksin [19], who prove the existence of elliptic
tori for Schrödinger equations on Tn with a convolution potential. The simplification
introduced by these models is that the normal modes eigenfunctions are exponentials.

Informal presentation of ideas and techniques

We are interested in the matrix representation of the multiplication operator by a
smooth function. For any compact Riemannian manifold without boundary, the ma-
trix coefficients decay as inverse powers of the differences between the eigenvalues of
√
−. However, these eigenvalues are in general not sufficiently separated (as it hap-
pens in dimension 1). Some finer properties of the spectrum and of the eigenfunctions
seem to be required for controlling the small divisors effects.
When M = S n or, more generally, a Zoll manifold (i.e., the geodesic flow is peri-
odic), the spectrum of the Laplace-Beltrami operator is contained in disjoint intervals,
growing linearly to infinity as for 1-dimensional PDEs. Moreover, weak properties
of localization of the eigenfunctions hold. Note that spheres and Zoll manifolds are
particular symmetric spaces of rank one.
It is a general fact that the presence of continuous symmetries expressed via a Lie
group action provides information on the spectral analysis of the Laplace-Beltrami
operator. In this paper we strongly exploit this basic idea. If the action is transitive,
then, up to isomorphism,

M = (G × Tn )/N,

where G is a simply connected compact Lie group, Tn is a torus, and N is a closed

subgroup of G × Tn (see (2.13)). The functions on M can be seen as functions defined
on G × Tn and invariant under the action of N , namely,
 
L2 (M) = L2 (G × Tn /N) = f ∈ L2 (G × Tn )  f (xg) = f (x),

∀x ∈ G × Tn , g ∈ N . (1.3)
484 BERTI and PROCESI

Moreover, the Laplace-Beltrami operator on M can be identified with the Laplace-

Beltrami operator on G × Tn , acting on the functions invariant under N (see Theorem
2.7). Then we “lift” equations (1.1) and (1.2) on G× Tn , and we use harmonic analysis
on Lie groups.
The Fourier analysis on the noncommutative group G is provided by the Peter-
Weyl theorem, Theorem 2.2. It states that L2 (G) is the Hilbert sum of the subspaces
generated by all the irreducible unitary representations of the group. Moreover, each
subspace is an eigenspace of the Laplace-Beltrami operator (see Theorem 2.4). Then
we use the highest weight theory of Lie groups (see [34]), which fully describes the
eigenvalues and eigenspaces of the Laplace operator (see Section 2.2). We have focused
Section 2.2 on those properties which are relevant for the theory of Hamiltonian PDEs
on compact homogeneous spaces.
For instance, in order to get the separation property (i) for the singular sites,
we exploit the fact that the eigenvalues of  on M (see (1.4)) are essentially the
eigenvalues of the Laplacian restricted to a maximal torus (whose dimension is the
rank of the group). Moreover, they are all in Z/D for some D ∈ N. This yields the
required separation properties (i) reasoning, for NLW, as in [8] and [4] and, for NLS,
as in [9].
The other property we need for the inversion of the linearized operators is the
following.
(ii ) The multiplication operator u → bu by a Sobolev function b ∈ H s (M) is
represented in the eigenfunction basis as a block matrix with the off-diagonal
decay given by Formula (2.38).
The block structure of this matrix takes into account the (large) multiplicity of the
degenerate eigenvalues of  on M . We remark that several blocks could correspond
to the same eigenvalue. (This block decomposition follows by the highest weight
theory.) On each block we consider the operator L2 -matrix norm. Then property (ii ),
proved in Proposition 2.19, follows by the multiplication rules of the eigenfunctions
(see Theorem 2.10), which, in turn, follow by the highest weight theory.
Once property (ii ) has been guessed, the main new technique is to embed the
matrix which represents the multiplication operator into a suitable algebra of “polyno-
mially localized” block matrices satisfying interpolation inequalities (see Section 2.6).
This abstract procedure simplifies considerably also the construction in [4], providing
an intrinsic formulation of the Craig-Wayne-Bourgain approach.
Finally, since Theorems 1.2 and 1.3 hold for fixed nonlinearities, we need to solve
the bifurcation equation (infinite-dimensional for the completely resonant NLS equa-
tion). In particular, we assume that this equation possesses nondegenerate solutions.
As examples, we verify this condition for the NLW on SU(n) in Section 4.1 and for
the NLS on SO(3) in Section 7.1.
We now state precisely our main results.
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 485

1.1. Main results

We first recall some basic results of harmonic analysis on the homogeneous space
M = (G × Tn )/N . The eigenvalues and the eigenfunctions of the Laplacian on a
simply connected compact group G are

−|j1 + ρ|2 + |ρ|2 , ej1 ,σ (x1 ) , x1 ∈ G , j1 ∈ + (G) , σ = 1, . . . , dj1 ,

where + (G) is the cone generated by the natural combinations of the fundamental
r
weights wi ∈ Rr , i = 1, . . . , r (r is the rank of the group), and ρ := i=1 wi .
The degeneracy of the eigenvalues is dj1 ≤ |j1 + ρ| dim(G)−r
(see Theorems 2.3, 2.4).
Moreover, there exists D := D(G) ∈ N such that −|j1 + ρ|2 + |ρ|2 ∈ ZD −1 for all
j1 ∈ + (G) (see Lemma 2.6).
By Fubini theorem, L2 (G × Tn ) = L2 (G) ⊗ L2 (Tn ), and a complete set of
eigenfunctions of  on G × Tn are the products of the eigenfunctions of the Laplacian
on G times the exponentials. We select the N -invariant eigenfunctions, and we find
a basis of eigenfunctions of  on M = (G × Tn )/N (see (1.3)). In conclusion, the
eigenvalues and the eigenfunctions of − + μ on M are

ωj2 := |j1 +ρ|2 −|ρ|2 +|j2 |2 +μ, ej,σ (x) := ej1 ,σ (x1 )eij2 ·x2 , x = (x1 , x2 ) ∈ G×Tn ,
(1.4)
where the index j = (j1 , j2 ) is restricted to a subset M ⊂ + (G) × Zn (see (2.15))
and σ ⊂ [1, dj ] with dj ≤ dj1 (see Theorem 2.9).

1.1.1. The nonlinear wave equation

The solutions of the linearized equation utt − u + μu = 0 on M are the normal
modes

  
dj

v= vl,j,σ eilωj t ej,σ , vl,j,σ ∈ C.

l=±1 j ∈M σ =1

Fixed j0 ∈ M , we want to prove the existence of small amplitude periodic solutions

of the nonlinear equation (1.1) with frequencies close to ωj0 . Assuming that μ is an
irrational number, the normal mode frequencies ωj commensurable with ωj0 satisfy
ωj = ωj0 . Actually, if μ is diophantine, that is, there are constants γ0 > 0, τ0 > 1,
such that
γ0
|μ m + n| ≥ , m := max(1, |m|) , ∀(m, n) ∈ Z2 \ {(0, 0)}, (1.5)
mτ0
then first-order Melnikov nonresonance condition (1.6) holds. (It is proved at the end
of Section 2.2.)
486 BERTI and PROCESI

LEMMA 1.1
Let μ be diophantine. There exist γ1 , τ1 > 0 such that
γ1
|ωj20 l 2 − ωj2 | ≥ , ∀ (l, j ) ∈ Z × M , (|l|, ωj ) = (1, ωj0 ). (1.6)
lτ1

Concerning the nonlinearity, we assume that f ∈ C k (M × R, R), k > 2, and, for

some m ∈ [2, k],

f (x, 0) = · · · = (∂um−1 f )(x, 0) = 0, (∂um f )(x, 0) = 0. (1.7)

Rescaling in (1.1) amplitude and time u(t, x) → δu(ωt, x), δ > 0, we look for
2π -periodic solutions of

ω2 utt − u + μu = ε f(δ, u), ε := δ m−1 , (1.8)

where
f (x, δu) (∂um f )(x, 0)
f(δ, u) := = a(x)um + O(δ), a(x) := ,
δm m!
s
in the real Sobolev space Heven := {u ∈ H s (T × M, R) | u(−t, x) = u(t, x)} for
some s ≤ k .
We define the resonant indices Q ⊂ Z × M as
 
Q := (l, j ) ∈ Z × M such that (|l|, ωj ) = (1, ωj0 ) ,

P := (Z × M ) \ Q . (1.9)

Then we perform a Lyapunov-Schmidt reduction according to the decomposition

s
Heven = (Q ∩ Heven
s
) ⊕ (P ∩ Heven
s
),

where
  
dj 
Q := q = eilt ql,j,σ ej,σ ∈ Heven
0
and P := Q⊥ . (1.10)
(l,j )∈Q σ =1

We set the frequency-amplitude relation

ω2 − ωj20 = εχ with χ ∈ {−1, +1}

NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 487

chosen later in (1.14). We project equation (1.8) on Q and P . Writing u = q + p ,

q ∈ Q, p ∈ P and using qtt = −q , equation (1.8) is equivalent to

q = −χ Q f(δ, q + p) (bifurcation equation),
(1.11)
ω2 ptt − p + μp = εP f(δ, q + p) (range equation),

where Q , P denote the orthogonal projectors onto Q, P .

For δ = 0 and p = 0, the bifurcation equation reduces to


q = −χ Q a(x)q m , (1.12)

and we suppose that

Q (a(x)q m ) ≡ 0. (1.13)

Taking

−1 if ∃ q ∈ Q such that T×M a(x)q m+1 > 0,
χ :=  (1.14)
1 if ∃ q ∈ Q such that T×M a(x)q m+1 < 0,

equation (1.12) possesses at least a “mountain pass” critical point q̄ ∈ Q \ {0} for the
functional

q2 q m+1
(q) := + χa(x) dt dx.
T×M 2 m+1
We assume the following nondegeneracy condition.
(ND) There is an invariant subspace H s ⊂ Heven s
of the NLW equation (1.8) and a
solution q̄ ∈ Q ∩ H \ {0} of (1.12) which is nondegenerate; namely, h = 0
s

is the unique solution of the linearized equation

h + χQ (m a(x)q̄ m−1 h) = 0, h ∈ Q ∩ H s. (1.15)

In Lemma 4.2 we verify (ND) for SU(n) and f = u3 + r(x, u), where r(x, u) satisfies
(1.7) for some m > 3 and r(x, u) = r(x −1 , u).

THEOREM 1.2 (Wave equation)

Suppose that μ > 0 is diophantine, and fix j0 ∈ M . There is s0 := s0 (M),
k := k(M) ∈ N such that ∀f ∈ C k satisfying (1.7) for some m ∈ [2, k] and (ND),
there exist δ0 > 0, a curve

u ∈ C 1 ([0, δ0 ]; H s0 ) with u(δ) − δ q̄s0 = O(δ 2 ) ,

488 BERTI and PROCESI

and a positive measure Cantor set C ⊂ [0, δ0 ] such that, ∀δ ∈ C , u(δ)(ωt, x) is a

solution of (1.1) with ω2 = ωj20 + χ δ m−1 , and χ is chosen in (1.14).

1.1.2. The nonlinear Schrödinger equation

Under the transformation u → eiμt u, equation (1.2) becomes

iut − u = f (x, |u|2 )u. (1.16)

This nonlinear Schrödinger equation is called completely resonant because the normal
mode frequencies |j1 + ρ|2 − |ρ|2 + |j2 |2 ∈ Z/D , D ∈ N, are all commensurable.
For fixed j0 ∈ M , we want to prove the existence of small amplitude periodic
solutions of (1.16), with frequencies close to ωj20 , in the form of wave packets. Note that
(1.16) possesses standing wave solutions eiωt U (x), where U (x) satisfies the elliptic
PDE

−U − ωU = f (x, |U |2 )U.

On the contrary, the more complex wave packet periodic solutions of (1.16) proved in
Theorem 1.3 are not obtained by separation of variables and correspond, in general,
to quasi-periodic solutions of (1.2). We assume that f (x, y) ∈ C k (M × R, R) and,
for some m ∈ [1, k],

f (x, 0) = · · · = (∂ym−1 f )(x, 0) = 0 , (∂ym f )(x, 0) = 0. (1.17)

Rescaling amplitude and time u(t, x) → δu(ωt, x), we look for solutions in H s :=
H s (T × M, C) of

iωut − u = ε f(δ, |u|2 )u , ε := δ 2m , (1.18)

where f(δ, |u|2 ) := f (x, δ 2 |u|2 )/ε = a(x)|u|2m + O(δ 2 ) and a(x) := (∂ym f )(x, 0)/
m!
We define the infinite-dimensional resonant subspace

  
dj 
Q := q = e ilt
ql,j,σ ej,σ ∈ H 0 , (1.19)
(l,j )∈Q σ =1

where
 
Q := (l, j ) ∈ Z × M | ωj20 l − ωj2 = 0 .

Then we perform a Lyapunov-Schmidt reduction along Q and P := Q⊥ as for the

wave equation. We set the frequency-amplitude relation ω = ωj20 + χε for some
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 489

χ ∈ {−1, +1}. The bifurcation equation at δ = 0, p = 0, reduces, using iqt =

ωj−2
0
q , to


iqt = χQ a(x)|q|2m q . (1.20)

We suppose that
(ND) there exists an invariant subspace H s ⊂ H s and a solution q0 ∈ Q ∩ H s \ {0}
of (1.20) for some χ ∈ {−1, 1} satisfying q0 (t, x) = eilt U (x) for all l ∈ N.
The solution q0 is nondegenerate in Q ∩ H s ; namely, the linear operator


J0 [h] := −iht +χQ a(x) (m+1)|q0 |2m h+m|q0 |2(m−1) q02 h̄ , h ∈ Q∩H s ,

is invertible, and |J0−1 |s ≤ C(s) for all s ≥ 0 (see Definition 2.4).

If q0 = eilt U (x) for some l ∈ N, the periodic solution found in Theorem 1.3 would
reduce to a standing wave solution. The last assumption on J0−1 is required since Q is
infinite-dimensional.
In Section 7.1 we verify (ND) for SO(3), f = |u|2 + r(|u|2 ), where r(y) satisfies
(1.17) for some m > 1.

THEOREM 1.3 (Schrödinger equation)

Fix j0 ∈ M . There is s0 := s0 (M), k := k(M) ∈ N, such that ∀f ∈ C k satisfying
(1.17) for some m ∈ [1, k] and (ND), there exist δ0 > 0, a curve

u ∈ C 1 ([0, δ0 ]; H s0 ) with u(δ) − δq0 s0 = O(δ 2 ),

and a positive measure Cantor set C ⊂ [0, δ0 ] such that, ∀δ ∈ C , u(δ)(ωt, x) is a

solution of (1.16) with ω = ωj20 + χ δ 2m . Then eiμt u(δ) is a (quasi-periodic) solution
of (1.2).

The paper is organized as follows. We first collect in Section 2 various elements of

nonlinear analysis on Lie groups of general interest for Hamiltonian PDEs. In Section
2.1 we introduce the basic notions of Fourier analysis on groups, and in Section 2.2 we
describe the spectral decomposition of the Laplace-Beltrami operator on Lie groups
and homogeneous spaces. We exemplify the theory for the classical Lie groups SU(n)
and SO(3) in Section 2.3. In Section 2.4 we discuss the properties of the multiplication
of eigenfunctions. Then we introduce in Section 2.5 the Sobolev scale H s , and we
define in Section 2.6 the algebra of block matrices which are polynomially localized
in the vicinity of the diagonal. A key point of the paper is the block off-diagonal
decay property (2.38) for the matrix which represents the multiplication operator for
a Sobolev function. Finally, we discuss composition operators on H s .
Section 3 presents the abstract Nash-Moser theorem [5, Theorem 1]. As an ap-
plication, using the previous nonlinear functional analysis on Lie groups, we devote
490 BERTI and PROCESI

Sections 4 – 7 to the proof of Theorems 1.2 and 1.3. In Section 4 we solve the bifurca-
tion equation for NLW on SU(n). In Sections 5 and 6 we prove the tame estimates for
the inverse linearized operator. Finally, in Section 7 we prove the existence of periodic
solutions for NLS.

2. Linear and nonlinear analysis on Lie groups

A basic tool of harmonic analysis, essential to this paper, is representation theory of
compact Lie groups. For a general introduction we refer to [12], [22], [36], and [34].

2.1. Harmonic analysis on compact groups

Let G be a compact topological group, and let L2 (G) := L2 (G, C) be the Lebesgue
space defined with respect to the normalized Haar measure μ of G.
Let (V , ρV ) be a finite-dimensional unitary representation of G, namely, a con-
tinuous homomorphism x → ρV (x) which maps G into the group of unitary transfor-
mations U (V ) ⊂ End(V ), where V is a finite-dimensional complex vector space V .
Given an orthonormal basis {v1 , . . . , vn } of V , the representation is described by the
unitary matrices

U (x) := U V (x) := {Ul,k

V
(x)} = {(ρV (x)vl , vk )} ,
l, k = 1, . . . , n, n := dim(V ).
(2.1)
The dual representation is given by the complex conjugated matrix Ū (x).
A subspace W ⊂ V is invariant for a representation ρV of G if ρV (x)(W ) ⊂ W ,
∀x ∈ G. If a representation space V has no nontrivial invariant subspaces, then V and
ρV are irreducible.
The representations (V , ρV ) and (W, ρW ) of G are equivalent if there is an
isomorphism ψ : V → W such that ρV (x) = ψ −1 ◦ ρW (x) ◦ ψ for each x ∈ G.
Let MV be the vector space generated by the matrix coefficients Ul,k (x), namely,
   
MV := x → tr(BU (x)) , B ∈ Mat(Cn ) = x → tr(AρV (x)) , A ∈ End(V ) .
(2.2)
Subspaces MV , MW corresponding to nonequivalent irreducible representations ρV ,
ρW are orthogonal in L2 (G) (Schur lemma). We also have the following.

LEMMA 2.1
If ρV is irreducible, the map A → tr (AρV (x)) is an isomorphism.

Given two representations ρV , ρW and A ∈ End(V ⊕ W ), we have




tr A ρV ⊕W (x) = tr AV ρV (x) + tr AW ρW (x) , ∀x ∈ G, (2.3)
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 491

where

AV := V A|V ∈ End(V ), AW := W A|W ∈ End(W ),

and V , W denote the orthogonal projections on V , W .

Since G is compact, a finite-dimensional representation decomposes as a direct
sum of irreducibles. Hence the formula (2.3) fully describes MV in terms of matrix
coefficients of irreducible representations.
The product of two functions of MV and MW is in MV ⊗W . Indeed, ∀A ∈ End(V ),
B ∈ End(W ),




tr AρV (x) tr BρW (x) = tr (A ⊗ B)(ρV ⊗ ρW )(x) , ∀x ∈ G. (2.4)

The Fourier analysis on the group is provided by the Peter-Weyl theorem. The irre-
ducible representations of a group play the role of the exponential basis in the case of
the standard torus.

THEOREM 2.2 (Peter-Weyl)

Let Ĝ be the set of equivalence classes of irreducible unitary representations of the
compact group G, and, for each j ∈ Ĝ, let Mj := MVj . We have the Hilbert
decomposition


L2 (G) = Mj .
j ∈Ĝ

For f ∈ L2 (G), we have the L2 -convergent “Fourier series”


f (x) = tr (fj ej (x)), fj := f (x)ēj (x) dμ, (2.5)
G
j ∈Ĝ

where ej (x) := (dimVj )1/2 Uj (x) and the matrices Uj (x) := U Vj (x) are defined in
(2.1). In (2.5) the matrix ēj (x) is the complex conjugate of ej (x), and fj are the Fourier
coefficients of f (x).

With a slight abuse of notation, we denote the matrix coefficients of ej (x) by ej,σ (x),
σ = 1, . . . , dimVj2 .
The Schur orthogonality relations



 tr(AB † )
tr AρV (x) tr BρV (x) dμ(x) = , ∀A, B ∈ End(V ),
G dim(V )

imply that the ej,σ (x), σ = 1, . . . , dimVj2 , form an L2 -orthonormal basis for Mj .
492 BERTI and PROCESI

Note that the trivial representation of G on V0 := C, defined by ρC (x) = I ,

∀x ∈ G, is irreducible. We denote by M0 the corresponding space (of constant
functions) in the Peter-Weyl theorem.
For a product group G = G1 ×G2 , the Fubini theorem implies that L2 (G1 ×G2 ) =
L2 (G1 ) ⊗ L2 (G2 ) and all the irreducible representation spaces are Vj1 ⊗ Vj2 with
j1 ∈ Ĝ1 , j2 ∈ Ĝ2 .
We denote by χV (x) := tr(ρV (x)) the character of a representation ρV of G. The
characters χVj , j ∈ Ĝ form a Hilbert basis for the subspace of L2 (G) formed by the
functions satisfying the symmetry condition f (gxg −1 ) = f (x), ∀x, g ∈ G, called
central functions.
All the previous results are valid for any compact topological group. We now
specialize to Lie groups.

2.2. Laplace operator on compact Lie groups and homogeneous spaces

By a standard classification theorem (see, e.g., [12]), any connected compact Lie group
G = (G× Tn )/N , where G is simply connected and N is finite and central. Moreover,
the simply connected Lie group G is product of groups of simple type, which are
r the special unitary group SU(n), n ≥ 2,
r Spin(n), the double cover of the special orthogonal group SO(n),
r SP (n) := Sp(n; C) ∩ U (2n), where Sp(n; C) is the group of the 2n × 2n
complex symplectic matrices, and
r the five exceptional groups called G2 , F4 , E6 , E7 , E8 .
Then, to understand harmonic analysis for any G , we study the Laplace-Beltrami
operator, first on simply connected groups of simple type and then on the product
space G × Tn . At the end we discuss how the Laplacian passes to the quotient space.
Let G be a simply connected compact Lie group of simple type. Recall that the
Killing form

−(X, Y ) := tr(Ad(X) ◦ Ad(Y )), where Ad(X)(·) := [X, · ], (2.6)

is negative definite and hence defines a Riemannian metric on G. We define the

Laplace Beltrami operator  on G with respect to this metric ( is called a Casimir
in most books on representation theory).
The next theorem (see [34, Proposition 1, p. 349]) describes all the irreducible
representations.
We recall that the rank of a Lie group G is defined as the dimension of any
maximal connected commutative subgroup of G (maximal torus).

THEOREM 2.3
For a simply connected compact Lie group G of rank r , there is a one-to-one correspon-
dence between the set of equivalence classes Ĝ of irreducible unitary representations
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 493

and a discrete cone

 
r 
+ := + (G) := j = ni wi , ni ∈ N ⊂ Rr
i=1

generated by r independent vectors wi ∈ Rr . The {w1 , . . . , wr } are called the fun-

damental weights of the group, and + is called the cone of dominant weights. The
irreducible representation of G corresponding to the dominant weight j = 0 is the
trivial representation on V0 = C.

The matrix coefficients of an irreducible representation are eigenfunctions of the

Laplacian.

THEOREM 2.4 ([34, Lemma 1])

Each Mj is an eigenspace of the Laplace Beltrami operator  with eigenvalue


r
−|j + ρ|2 + |ρ|2 , ρ := wi . (2.7)
i=1

We have dj := dim(Mj ) ≤ |j + ρ|dim(G)−r .

Since the Laplacian is a real operator, for any eigenvector ej,σ (x) ∈ Mj also ēj,σ (x)
is an eigenvector of  with the same eigenvalue. Therefore ēj,σ (x) ∈ Mj̃ for some
j̃ ∈ + . Moreover, Vj̃ = Vj∗ because the matrix ēj (x) is the dual representation on
Vj∗ of the matrix ej (x) .
The eigenvalues and the eigenfunctions of the Laplace operator are described by
introducing the positive simple roots α1 , . . . , αr ∈ Rr of the group G (see [34]). They
satisfy the relations
1
(wi , αj ) = δij |αj |2 , ∀i, j = 1, . . . , r, (2.8)
2
where δij is the Kronecker symbol. We define the cone
 
r 
R+ := α = ni αi , ni ∈ N ⊂ Rr
i=1

generated by the natural combinations of the positive simple roots. We define the
lattice
 
r 
 := j = ni wi , ni ∈ Z ⊂ Rr
i=1
494 BERTI and PROCESI

generated by the fundamental weights w1 , . . . , wr (see Theorem 2.3) and ++ :=

ρ + + .

Definition 2.1 (Dominance order)

For i, j ∈ , we say that i ≥ j if i = j + α for some α ∈ R+ .

Remark 2.1
r r
By the theory of roots, ρ := i=1 wi = i=1 Ni αi , where the Ni are given positive
integers. Hence, by (2.8), the cones + , R+ ⊂ Rr intersect the hyperplane ρ⊥ ⊂ Rr
only in zero, and the other elements are contained in the half-space (ρ, y) > 0, y ∈ Rr .

LEMMA 2.5
There is c ∈ (0, 1) such that ∀i ≥ j , i ∈ + , j ∈  with (ρ, j ) > 0, one has
|i + ρ| ≥ c|j + ρ|.

Proof
We claim that there is a ∈ (0, 1) such that, ∀j ∈  with (ρ, j ) > 0 and ∀α ∈ R+ ,
one has

(j + ρ, α) ≥ −a|j + ρ||α| . (2.9)

Let x := (j + ρ)|j + ρ|−1 , y := α|α|−1 , and −a := minX×Y (x, y), where X is

the intersection of the unitary sphere with the half-space (x, ρ) ≥ 0 and Y is the
intersection of the unitary sphere with the prisma spanned by the wi . Note that X × Y
is compact. By Remark 2.1, the cone and the plane (x, ρ) = 0 do not have any
common line, and so a = 1. If i ≥ j , namely, by Definition 2.1, i = j + α for some
α ∈ R+ , then |i + ρ|2 ≥ (1 − a 2 )|j + ρ|2 by (2.9) and minimizing on |α|. 䊐

Many properties of the eigenvalues are known.

LEMMA 2.6
For any simply connected Lie group G, there is D ∈ N such that (wi , wj ) ∈ D −1 Z.
Hence

|j |2 , |j + ρ|2 , (ρ, j ) ∈ D −1 Z , ∀j ∈ + .

Proof
Consider the nonsymmetric Cartan matrix
(αi , αj )
C := C (G) := (Ci,j )i,j =1,...,r , where Ci,j := 2 .
|αj |2
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 495

By definition of roots, we get (wi , wj ) = |αi |2 (C −T )i,j /2. The lemma follows because
the matrix elements of the Cartan matrix satisfy Ci,j ∈ Z (actually |Ci,j | ≤ 3) and
|αj |2 ∈ N/2 (see [34]). 䊐

The eigenspaces of the Laplace operator on G × Tn are

Mj1 eij2 ·x2 with x1 ∈ G, x2 ∈ Tn , (j1 , j2 ) ∈ + × Zn ,

the eigenfunctions ej1 ,σ (x1 )eij2 ·x2 , 1 ≤ σ ≤ dj , and the eigenvalues −|j1 + ρ|2 +
|ρ|2 − |j2 |2 .
We finally need to understand how to pass to the quotient space (see [25], [26],
[34, pp. 81–85], [37]).

THEOREM 2.7
Let H be a closed subgroup of a Lie group G . Then there is a unique manifold structure
on the quotient space G /H , such that the projection map π : G → G /H is a smooth
submersion.
Moreover, given a biinvariant metric on G , the projection π induces on G /H a
Riemannian structure such that the Laplace-Beltrami operator on C ∞ (G /H, C) is
identified with the Laplace-Beltrami operator on
∞
 
Cinv (G , C) := f ∈ C ∞ (G , C) such that f (x) = f (xg) , ∀x ∈ G , g ∈ H
(2.10)
and the diagram commutes:
π∗
C ∞ (G /H, C) −−−−→ Cinv
∞
(G , C)
⏐ ⏐
⏐
G /H 
⏐
 G (2.11)
∗
π
C ∞ (G /H, C) −−−−→ Cinv
∞
(G , C).

Given an irreducible representation ρV of G , we consider the subspace

  
W := w ∈ V  ρV (g)w = w , ∀g ∈ H ⊂ V

and the subspace of functions NV ⊂ MV defined by

 
NV := Span (ρV (x)wk , vl ), k = 1, . . . , dim(W ), l = 1, . . . , dim(V ) , (2.12)

where (vl )l=1,...,dim(V ) is a basis of V and (wk )k=1,...,dim(W ) is a basis of W .

LEMMA 2.8
The NV are the functions of MV satisfying (2.10). Moreover, dim NV = dimV dimW .
496 BERTI and PROCESI

Proof
For any w, v ∈ V , v = 0, the matrix coefficient (ρV (x)w, v) satisfies (2.10) if and
only if




ρV (xg)w, v = ρV (x)ρV (g)w, v = ρV (x)w, v , ∀x ∈ G , g ∈ H,

and therefore (ρV (x)(ρV (g)w − w), v) = 0, ∀x ∈ G , g ∈ H . This implies—by the

irreducibility of ρV —that ρV (g)w − w = 0, ∀g ∈ H , that is, w ∈ W . 䊐

Let (g, x) → gx denote the action G × X → X of a group G on a set X . We recall

that the orbit of x is O(x) := {gx ∈ X , g ∈ G }. The stabilizer of x is the subgroup
Gx := {g ∈ G , gx = x}. The action of a group G on X is called transitive if, ∀x ∈ X,
the orbit O(x) = X .

Definition 2.2
A compact manifold M is said to be homogeneous if there is a compact Lie group
G which acts on M transitively and differentiably; that is, for each g ∈ G , the map
x → gx is differentiable in M.

Consider the group G = (G × Tn )/N1 , where N1 is a finite central subgroup. The

action of G × Tn on any p ∈ M induces a diffeomorphism
φ
M ↔ (G × Tn )/N, where N := N1 Gp (2.13)

and Gp is the stabilizer of p . (Note that Gp is closed.) By Theorem 2.7, a biinvariant

metric on G × Tn induces a metric on (G × Tn )/N and, then, on M (see [6]).

Remark 2.2
If M is also a Riemannian manifold with a G -invariant metric, it is natural to ask if it
can be derived by submersion from a biinvariant metric on G × Tn . In general, this
is not the case. However, this is true for the sphere S n with the canonical metric (see
[26]) and in several applications arising from physics (see, e.g., [37] or [27] and [32]
or [27] for an application to the Landau equation).

For each index j = (j1 , j2 ) ∈ + × Zn of the irreducible representations of G × Tn ,

we consider the subspace, defined in (2.12), of N -invariant functions

Nj ⊂ Mj := Mj1 eij2 ·x2 . (2.14)

Each Nj is an eigenspace of the Laplacian with dimension dim Nj ≤ dim Mj .

NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 497

It could happen that for some j ∈ + × Zn , the subspace Nj = {0}. The set
 
M := j ∈ + × Zn such that Nj = {0} (2.15)

is closed under sum. By the Peter-Weyl theorem (Theorem 2.2), we deduce the spectral
theory of the Laplace Beltrami operator on a compact homogeneous space.

THEOREM 2.9

We have L2 (M) = j ∈M Nj . A basis for Nj ⊂ Mj is, up to a reordering of the
index σ ,

ej,σ (x) = ej1 ,σ (x1 ) exp(ij2 · x2 ), σ = 1, . . . , dj ,

for some 1 ≤ dj ≤ dj .

As a first application of the properties of the eigenvalues stated in Lemma 2.6, we

prove Lemma 1.1.

Proof of Lemma 1.1

Let j0 = (j01 , j02 ) ∈ M ; ∀(l, j ) ∈ Z × M such that (|l|, ωj2 ) = (1, ωj20 ), by (1.4),
(1.5), and Lemma 2.6,


|ωj20 l 2 − ωj2 | =  |j01 + ρ|2 l 2 − |j1 + ρ|2 + |j02 |2 l 2 − |j2 |2 + |ρ|2 (1 − l 2 )
 c γ1
+ μ(1 − l 2 ) ≥ ≥ 2α
|1 − l |
2 α l
for some constant γ1 > 0. 䊐

2.3. Examples: SU(n) and SO(3)

We exemplify the above theory on the classical Lie groups SU(n), n ≥ 2, and SO(3).
The special unitary group SU(n) is a matrix Lie group with real dimension
n2 − 1. Its rank is n − 1 since a maximal torus is the group of diagonal matrices
diag(eiθ1 , . . . , eiθn ) with determinant equal to 1. The Lie algebra of SU(n) is the space
of anti-Hermitian matrices with trace zero.
The fundamental weights and the positive simple roots form two bases over R
for the Lie algebra of the maximal torus of diagonal matrices, that is, the (n − 1)-
n
dimensional subspace of Rn with coordinates θ ∈ Rn and i=1 θi = 0. For SU(n)
498 BERTI and PROCESI

they are

√ k √
n
2n wk = e1 + e2 + · · · +ek − ei , 2n αk = ek −ek+1 , k = 1, . . . , n−1,
n i=1
(2.16)
n ∗
where the ei , i = 1, . . . , n, form the standard basis of R . Formula (2.16) allows us
to compute the eigenvalues of the Laplace-Beltrami operator  on SU(n) according
to (2.7).
One can also describe explicitly the irreducible representations (Vj , ρVj ) and also
the corresponding eigenspaces Mj defined through formula (2.2).
Given 1 ≤ i1 ≤ · · · ≤ ik ≤ n, denote by (i1 , . . . , ik ) the function that associates
to y ∈ SU(n) the determinant of the minor extracted from the first k columns and
from the rows i1 , . . . , ik . We define a partial order on this finite set of functions saying
that (i1 , . . . , ik ) ≤ (j1 , . . . , jh ) if and only if k ≥ h and il ≤ jl for l = 1, . . . , h.
A product of these functions f1 · f2 · · · fs is standard if fi ≤ fi+1 for all i . Given

j = i ni wi , consider all standard monomials formed by nk k × k minors for each
k . These standard monomials form a basis of a space of functions representative of
Vj , where ρVj (x)f (y) = f (x −1 y). This fully describes the eigenfunctions of the
Laplacian on SU(n) (see [34, Chapter 13]).
As an example, if j = w1 , then (Vj , ρVj ) is equivalent to the defining
k represen- 
tation (V , x) with

V =
C. If j = wk , then (Vj , ρVj ) is equivalent to
n
V , ∧k x ,
where ∧k x is the nk × nk -matrix with entries the determinants of the (k ×k)-minors

k 
n−k 
of x . Finally, the dual representation to V , ∧k x is V , ∧n−k x (by Binet
formulas for the determinant).
For SU(2) the fundamental weight is w1 = 1/4(1, −1), and the dominant weights
are j = m/4(1, −1), m ∈ N. The integer D in Lemma 2.6 is equal to 8, and all the
eigenvalues of − on SU(2) are
1
 N
(m + 1)2 − 1 ∈ . (2.17)
8 8
All the unitary representations ρVj of SU(2) are self-dual, namely, ρ̄Vj = ρVj , and
hence the characters χj := χVj are real.
The orthogonal group SO(3) = SU(2)\{±I } is also a homogeneous space (see
Definition 2.2). The set SO(3) ⊂ + defined as in (2.15) results in
 m 
SO(3) = j ∈ + : j = (1, −1), m ∈ 2N . (2.18)
4
∗
In some books the normalization in (2.16) is different due to a multiplicative factor in front of the Killing form
(2.6), that is, the scalar product. This changes the Laplace-Beltrami operator and its eigenvalues by a factor.
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 499

2.4. Nonlinear analysis on Lie groups

2.4.1. Multiplication of eigenfunctions

The product of two eigenfunctions of the Laplace-Beltrami operator,

a = tr (A ρVi ) ∈ Mi , b = tr (BρVj ) ∈ Mj

is, by (2.4), given by ab = tr (A ⊗ B ρVi ⊗Vj ). Then we need to express Vi ⊗ Vj as the

direct sum of irreducible representations
 cij
Vi ⊗ Vj = V , cij ∈ Z ∪ {0}. (2.19)
∈+

The cij are called the Clebsch-Gordan coefficients of the group.

A key property of compact Lie groups is that the product of two eigenfunctions
is a finite sum of eigenfunctions according to the following theorem.

THEOREM 2.10 (Multiplication of eigenfunctions)

Let a ∈ Mi , b ∈ Mj . Then a b ∈ ⊕≤i+j M .

Proof
By a theorem of Cartan (see [34, p. 345]), Vi ⊗ Vj = Vi+j ⊕ V  , where V  is the sum
of irreducibles with dominant weights  ∈ + ,  < i + j . Then we have cij = 0 if
 ≤ i + j . 䊐

The L2 -orthogonal projection of ab on the eigenspace M is



M ab = tr (A ⊗ B)|,s ρV ,
s≤cij

where (A ⊗ B)|,s denotes the restriction of A ⊗ B to the s th copy of V in (2.19).

If A = B = Id, we obtain the formula for the characters χi := χVi , namely,

χi χj = cij χ . (2.20)
≤i+j

2.4.2. An example on SU(n)

By (2.7) and (2.16), the eigenvalues corresponding to w1 , wn−1 are equal. Actually
Mw1 ⊕Mwn−1 is the eigenspace of smallest nonzero eigenvalue of the Laplace-Beltrami
operator on SU(n).
As an example that will be relevant for the bifurcation equation in Section 4, we
compute explicitly the nonlinear functional

Mw1 ⊕ Mwn−1  q → Mw1 ⊕Mwn−1 q 3 .

500 BERTI and PROCESI

Since the defining representation ρVw1 (x) = x and its dual is ρVwn−1 (x) = x̄ , any
function in Mw1 ⊕ Mwn−1 can be represented uniquely (by Lemma 2.1) as
(2.3)

qA (x) = tr(A1 x) + tr(A2 x̄) = tr A(x ⊕ x̄) , A := A1 ⊕ A2 , (2.21)

where A1 , A2 ∈ Mat(Cn ). The function qA (x) is real if and only if A1 = Ā2 . By (2.4)
we have

3


qA3 = tr(A(x ⊕ x̄)) = tr A⊗3 (x ⊕ x̄)⊗3 = tr A ρU (x) , (2.22)

where A := A⊗3 := A ⊗ A ⊗ A ∈ End(U ), ρU (x) := (x ⊕ x̄)⊗3 , and

U := (V ⊕ V ∗ )⊗3 = V ⊗3 ⊕ (V ⊗ V ⊗ V ∗ ) ⊕ (V ⊗ V ∗ ⊗ V )
⊕(V ∗ ⊗ V ⊗ V ) ⊕ duals. (2.23)

We decompose in irreducible subspaces each representation space in the direct sum

(2.23).

LEMMA 2.11
Let n = 2, 4. In the direct sum decomposition of:
(i) V ⊗3 , (V ∗ )⊗3 there are no copies either of V or of V ∗ ;
(ii) V ⊗ V ⊗ V ∗ , V ⊗ V ∗ ⊗ V , V ∗ ⊗ V ⊗ V there are exactly 2 copies of V and
none of V ∗ . Hence their duals have two copies of V ∗ and none of V .

Proof
This lemma follows by the Pieri formulas for SU(n) (see [34, p. 289]) but also by the
first fundamental theorem of invariant theory (see [34, p. 387]). For SU(n) it states that
the only invariant polynomial functions F (v1 , . . . , vk , φ1 , . . . , φh ), vi ∈ V , φj ∈ V ∗ ,
are polynomials in the generators

F (φi , vj ) = φi (vj ), F (vi1 , . . . , vin ) = det(vi1 , . . . , vin ),

F (φi1 , . . . , φin ) = det(φi1 , . . . , φin ).

Now a representation space W contains k copies of V if and only if W ⊗ V ∗ contains

k independent invariants. (This is an immediate consequence of the Schur lemma.)
Since there is a one-to-one correspondence between the invariant multilinear functions

and invariants, the only invariants in V ⊗ V ⊗ V ∗ ⊗ V ∗ are i,k ei ⊗ ek ⊗ ei ⊗ ek

and i,k ei ⊗ ek ⊗ ek ⊗ ei . 䊐
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 501

Remark 2.3
For SU(4) the only difference is that there is one copy of V in V ⊗3 . For SU(2) there
is only one fundamental weight, and V ⊗3 contains two copies of V . We deal with this
case for NLS in Section 7.1.

By (2.23) and Lemma 2.11, we get the decomposition


U = (V ⊕ V ∗ )⊕6 ⊕j =w1 ,wn−1 Vj
cj
. (2.24)

An orthonormal basis for the six copies of V (resp., V ∗ ) in U is explicitly given below.

LEMMA 2.12
Consider the two orthogonal subspaces Vj := Spank=1,...,n Ek ⊂ V ∗ ⊗ V ⊗ V ,
(j )

j = 1, 2, where

e(1) ne(2) − ek(1)

Ek(1) := √k , Ek(2) := √k ,
n n3 − n
 
ek(1) := ei ⊗ ei ⊗ ek , ek(2) := ei ⊗ ek ⊗ ei ,
i i

and ei ∈ V , ei ∈ V ∗ , i = 1, . . . , n, are the standard basis of V and V ∗ .

(j )
Similarly, consider the two orthogonal subspaces Vj := Spank=1,...,n Ek ⊂
∗
V ⊗ V ⊗ V , j = 3, 4, where

e(3) ne(4) − ek(3)

Ek(3) := √k , Ek(4) := √k ,
n n3 − n
 
ek(3) := ei ⊗ ei ⊗ ek , ek(4) := ek ⊗ ei ⊗ ei .
i i

Finally, consider the two orthogonal subspaces Vj := Spank=1,...,n Ek ⊂ V ⊗V ⊗V ∗ ,

(j )

j = 5, 6,

e(5) ne(6) − ek(5)

Ek(5) = √k , Ek(6) = √k ,
n n3 − n
 
ek(5) := ek ⊗ ei ⊗ ei , ek(6) := ei ⊗ ek ⊗ ei .
i i

The representation spaces Vj are equivalent to V ; then the dual spaces Vj∗ are
equivalent to V ∗ . Finally, the {Ek }k=1,...,n are an orthonormal basis in Vj .
(j )
502 BERTI and PROCESI

For A ∈ End(U ), we represent its restriction Vj A|Vj to Vj by the (n × n)-matrix

(j ) (j )
Aj with entries (Aj )h,k = (Eh , AEk ). By (2.3) we have

 6

 6
Mw1 tr AρU (x) = tr(Aj x), Mwn−1 tr AρU (x) = tr(Āj x̄), (2.25)
j =1 j =1

where Āj is the matrix that represents Vj∗ A|Vj∗ .

Let us briefly discuss other nonlinearities and groups. The main point is to have
the analogue of Lemma 2.11. For E6 one can proceed essentially as above. For the
other groups, we have a unique defining representation W with minimum eigenvalue
of the Laplace operator, and W and W ∗ are equivalent. This allows us to construct
an invariant tensor J ∈ W ⊗ W which, under the isomorphism between W and W ∗
mapping W ⊗ W to W ⊗ W ∗ = End(W ), corresponds to the identity map. By the
theory of symmetry on tensor spaces, we see that we have copies of the representation
on each odd tensor power W ⊗2k+1 . One copy of W is J ⊗k ⊗ W , and other copies are
obtained by permuting the tensors.
So one treats the case when G is any simply connected group and the func-
tion is q → q p with p odd essentially as the case SU(n). Clearly the problem is
computationally harder.

2.5. The Sobolev scale H s

We now introduce the Banach scale of the Sobolev spaces on a group. This is a
fundamental tool in the theory of Hamiltonian PDEs, especially for the Nash-Moser
implicit function theorem.
Consider a homogeneous space M = (G× Tn )/N , where G is a simply connected
Lie group of dimension d and rank r .
The Peter-Weyl theorem, Theorem 2.2, combined with the standard Fourier theory
in the time variable t ∈ T, implies the orthogonal decomposition

L2 := L2 (T × M, C) = Nk , where k := (l, j ), E + := Z × M ,
k∈E +

Nk := eilt  ⊗ Nj . (2.26)

The Fourier series of u ∈ L2 is defined by


u= uk , where uk := Nk u and Nk : L2 → Nk
k∈E +

are the spectral projectors. In components,

 
dj
 
dj

u(t, x) = eilt ul,j,σ ej,σ and u2L2 = 2π |ul,j,σ |2 .

(l,j )∈E + σ =1 (l,j )∈E + σ =1
(2.27)
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 503

For s ≥ 0, we define the Sobolev scale of Hilbert spaces

   
H s := H s (T × M, C) := u = uk | u2s := |k + ρ |2s uk 2L2 < +∞ ,
k∈E + k∈E +

where ρ = (0, ρ, 0) ∈ Z × + × Zn . We have H 0 = L2 . Since T × M is a compact

C ∞ -Riemannian manifold without boundary, for any s ∈ N, H s is equivalent to the

usual Sobolev space H s = {u ∈ L2 | D α u ∈ L2 , ∀|α| ≤ s , u2s = |α|≤s D α u2L2 }.

LEMMA 2.13
For s ≥ s0 > (dim(M) + 1)/2, ∀u, v ∈ H s ,
(1) (Sobolev embedding) uL∞ ≤ C(s)us ,
(2) (algebra) uvs ≤ C(s)us vs , and
(3) (interpolation) uvs ≤ C(s, s0 )(us vs0 + us0 vs ).

Proof
These inequalities are well known for C ∞ -functions with compact support in
Rdim(M)+1 (see, e.g., [38]), and, therefore, by a partition of unity argument, they
still hold true for any compact C ∞ -Riemannian manifold without boundary. 䊐

We now introduce a preorder relation on E := Z ×  × Zn .

Definition 2.3 (Dominance order)

Given k = (l, j1 , j2 ), k  = (l  , j1 , j2 ) ∈ E , we say that k ≥ k  if |l| ≥ |l  |, j1 ≥ j1
(see Definition 2.1), and |j2 | ≥ |j2 |.

The product of two eigenfunctions splits as a finite sum of eigenfunctions.

PROPOSITION 2.14 (Multiplication of eigenfunctions)

Given a ∈ Nk and b ∈ Nk , then

ab ∈ Nk 1 ,
k1 ∈D(k,k  )

where
D(k, k  ) := {k  ∈ E + | l  = l + l  , j1 ≤ j1 + j1 , j2 = j2 + j2 },
where k = (l, j1 , j2 ), k  = (l  , j1 , j2 ), k  = (l  , j1 , j2 ).

Proof
The conditions l  = l+l  , j2 = j2 +j2 are the multiplication rules for the exponentials
   
eilt eil t = ei(l+l )t , eij2 ·x2 eij2 ·x2 = ei(j2 +j2 )·x2 . The condition j1 ≤ j1 + j1 comes from
Theorem 2.10. 䊐
504 BERTI and PROCESI

The next lemma is the key for proving the decay of the matrix representing the
multiplication operator.

LEMMA 2.15

Let s > (d +n+1)/2. If u ∈ k≥k0 ,k∈E + Nk with k0 := (l0 , j0 ) ∈ E , j0 = (j01 , j02 ) ∈
+ (G) × Zn , and (ρ, j01 ) ≥ 0, then
C(s)us
uL∞ ≤ , where k := max(1, |k|). (2.28)
k0 s−(d+n+1)/2
Proof
By (2.27), Definition 2.3, and Cauchy inequality, we have
   
    1/2
|u(t, x)| ≤  ul,j,σ ej,σ  ≤ Nl,j u0 |ej,σ |2 .
(l,j )∈E + σ =1,...,dj (l,j )∈E + σ =1,...,dj
(l,j )≥(l0 ,j0 ) (l,j )≥(l0 ,j0 )

(2.29)
Now ej,σ (x) = ej1 ,σ (x1 )eij2 ·x2 , and since ej1 (x1 ) is a unitary matrix, then

|ej,σ |2 ≤ tr (ej1 e†j1 ) = n2j1 = dj1 (2.30)
σ =1,...,dj

†
(where ej1 denotes the adjoint matrix). Applying again Cauchy inequality in (2.29),

u2L∞ ≤ Nl,j u20 (|l|2 + |j1 + ρ|2 + |j2 |2 )s
(l,j )∈E +
(l,j )≥(l0 ,j0 )

 dj1
×
(l,j )∈E +
(l 2 + |j1 + ρ|2 + |j2 |2 )s
(l,j )≥(l0 ,j0 )

 dj1
= u2s ≤ u2s
(l,j )∈E +
(l 2 + |j1 + ρ|2 + |j2 |2 )s
(l,j )≥(l0 ,j0 )

 |j1 + ρ|dim(G)−r
× (2.31)
(l,j )∈E +
(l 2 + |j1 + ρ|2 + |j2 |2 )s
(l,j )≥(l0 ,j0 )

by Theorem 2.4. By Lemma 2.5, j ≥ j0 implies |j1 + ρ| > c|j01 + ρ| with c ∈ (0, 1),
and we estimate the last term in (2.31) by

1 d r+n+1 x dy
≤ ≤ C
k∈E +
|k + ρ |2s−d+r
ρ | |x|
|x|>c|k0 +
2s−d+r
ρ| y
y>c|k0 +
2s−d−n
k≥k0

C2 (s)
≤
1 + |k0 |2s−(d+n+1)
deducing (2.28). 䊐
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 505

Note that the previous theorem provides a proof of the embedding H s → C(T × M)
for s > dim(T × G × Tn )/2 > dim(T × M)/2. We did not attempt to recover the
optimal bound on s given in Lemma 2.13.

2.6. Polynomially localized block matrices

According to the orthogonal splitting

Hs = Nk ,
k∈E +

we identify a linear operator A acting on H s with its matrix representation A =

 
(Akk )k,k ∈E + with blocks Akk ∈ L(Nk , Nk ).

Definition 2.4
We define the polynomially localized block matrices
 
 

As := A = (Akk )k,k ∈E + : |A|2s := sup k − k  2s Akk 20 < ∞ ,
k∈E +
k  ∈E +

 
where Akk 0 := supu∈Nk ,u0 =1 Akk u0 is the L2 -operator norm in L(Nk , Nk ).

If s  > s , then As  ⊂ As because | · |s  > | · |s . Let s0 > (r + n + 1)/2.

LEMMA 2.16 (Algebra)

For s > s0 , As is an algebra:

|AB|s ≤ C(s)|A|s |B|s , ∀A, B ∈ As . (2.32)

A similar matrix algebra has been used in [7]. Applying (2.32) iteratively, we get,
∀m ∈ N,

|Am |s ≤ C(s)m−1 |A|m

s . (2.33)

The algebra property of As is implied by the more general interpolation inequality

(2.34).

LEMMA 2.17 (Interpolation)

For all s ≥ s0 and for all A, B ∈ As ,

|AB|s ≤ C(s)(|A|s |B|s0 + |A|s0 |B|s ) (2.34)

and, ∀u ∈ H s ,

Aus ≤ C(s)(|A|s us0 + |A|s0 us ). (2.35)

506 BERTI and PROCESI

Proof
By definition,
 

|AB|2s = sup k − k  2s (AB)kk 20 ≤ sup k − k  2s
k∈E + k∈E +
k  ∈E + k  ∈E +
  
2
× Akk1 0 Bkk1 0 = S1 + S2 ,
k1 ∈E +

where
   k − k  s 2
S1 := sup Akk1 0 k1 − ks0 Bkk1 0 k  − k1 s
k∈E +
k  ∈E + k1 ∈E +
k1 − ks0 k  − k1 s

k  − k
if ≤2
k  − k1 
and
   k − k  s 2
S2 := sup Akk1 0 k1 − ks Bkk1 0 k  − k1 s0
k∈E +
k  ∈E + k1 ∈E +
k1 − ks k  − k1 s0

k − k
if > 2.
k  − k1 
Note that in this second case, k − k  / k − k1  ≤ 2. By Hölder inequality, and
exchanging the order of the sum,
  
S1 ≤ sup Akk1 20 k1 − k2s0 Bkk1 20 k  − k1 2s
k∈E +
k  ∈E + k1 ∈E +


k  − k2s 
×
k1 ∈E +
k1 − k2s0 k  − k1 2s
  

≤ sup Akk1 20 k1 − k2s0 Bkk1 20 k  − k1 2s
k∈E +
k1 ∈E + k  ∈E +
 C(s)
× ≤ |B|2s |A|2s0 K(s)
k ∈E +
k1
2s0
1

for s0 > (r + n + 1)/2. Similarly S2 ≤ |B|2s0 |A|2s K(s), proving (2.34). Then (2.35)
follows similarly. 䊐

Applying (2.34) iteratively, we obtain, ∀m ∈ N,

m−1
|Am |s ≤ m C(s)|A|s0 |A|s . (2.36)
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 507

LEMMA 2.18
Let A ∈ As , 1 , 2 ⊂ E + , and 1 ∩ 2 = ∅. Then A
2 0 ≤ C(s)|A|s /
1

d(1 , 2 )s−(r+n+1)/2 .

Proof

For any u = k ∈1 uk we have
 
  k  2  
  1 2
A
2 u0 =
1 2
 Ak uk  ≤ Akk  k − k  s uk 0
k  ∈1
0
k  ∈1
k − k  s
k∈2
  
k∈2
  1 
≤ uk 20 Akk 2 k − k  2s
k  ∈1 k∈2 k  ∈1
k − k  2s
≤ u20 C(s)2 |A|2s /d(1 , 2 )2s−(r+n+1) ,

where r + n + 1 is the dimension of E + . 䊐

Since H s is an algebra, any function b ∈ H s defines the multiplication operator

u(t, x) → b(t, x)u(t, x), ∀u ∈ H s , (2.37)

 
which is represented by (Bkk )k,k ∈E + with Bkk := Nk b(t, x)| Nk ∈ L(Nk , Nk ). The
main property of multiplication operators on Lie groups is the following.

PROPOSITION 2.19 (Multiplication operators)

 
If b ∈ H s is real, then the matrix (Bkk )k,k ∈E + is self-adjoint, that is (Bkk )† = Bkk , and,
∀s > (d + n + 1)/2,
 C(s)bs
Bkk 0 ≤ . (2.38)
k − k  s−(d+n+1)/2

Proof
The self-adjointness is a consequence of b being real. In order to prove (2.38), we first
suppose that k = (l, j ) and k  = (l  , j  ) satisfy (ρ, j1 − j1 ) ≥ 0. We split
 
b = b≥k−k + b≥k−k ,


 

where b≥k−k := Nk1 b, b≥k−k := Nk1 b.
k1 ≥k−k  ,k1 ∈E + k1 ≥k−k  ,k1 ∈E +

(Note that the relation ≥ is only a partial order; see Definition 2.3.) By Proposition
2.14 and Definition 2.3,

Nk (b≥k−k uk ) = 0, ∀uk ∈ Nk ,
508 BERTI and PROCESI

and therefore
 
Bkk 0 := sup Nk (buk )0 = sup b≥k−k uk 0
uk  0 =1 uk  0 =1
uk  ∈Nk  uk  ∈Nk 

 (2.28) C(s)bs
≤ b≥k−k L∞ ≤
k − k  s−(d+n+1)/2
applying Lemma 2.15 because (ρ, j −j  ) ≥ 0. If (ρ, j −j  ) < 0, then (ρ, j  −j ) > 0
and
 C(s)bs
Bkk 0 = (Bkk )† 0 = Bkk 0 ≤
k  − ks−(d+n+1)/2
by the previous estimate. Then (2.38) holds in any case. 䊐

By Proposition 2.19, we have the following corollaries.

COROLLARY 2.20

For any real b ∈ H s+a with a ≥ (d + r + 2n + 3)/2, the matrix B = (Bkk )k,k ∈E + ,
which represents the multiplication operator (2.37), is self-adjoint, it belongs to the
algebra of polynomially localized matrices As , and

|B|s ≤ K(s)bs+a . (2.39)

We need to consider restricted matrices. Given a set of indexes I ⊂ E + , let

  
 

As (I) := A = (Akk )k,k ∈I : (Akk )† = Akk and|A|2s := sup k −k  2s Akk 20 < ∞ .
k∈I
k  ∈I

COROLLARY 2.21
 
For A := (Akk )k,k ∈E + ∈ As , its restriction AI := (Akk )k,k ∈I ∈ As (I) satisfies

|AI |s ≤ |A|s . On the other hand, any A := (Akk )k,k ∈I ∈ As (I) can be extended to a

matrix in As setting Akk = 0 for k, k  ∈
/ I without changing the norm |A|s .

Therefore all the properties (algebra, interpolation, etc.) hold for As (I) with constants
which do not depend on I . Moreover, the matrices II which represent the projectors

I : H s → HI := Nk satisfy |II |s = 1, ∀s ≥ 0. (2.40)
k∈I∩E +
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 509

2.7. Composition operators on Sobolev spaces

We now present a general result for composition operators for smooth functions in
Sobolev spaces, discussing the regularity of the operator


u(t, x) → f(δ, u)(t, x) := δ −m f x, δu(t, x)

acting between the Sobolev spaces H s .

LEMMA 2.22 (Composition operator)

Fix s0 > 1+(d +n+1)/2. Assume that f ∈ C k satisfies (1.7) for some 0 ≤ m ≤ k−4.
Then, ∀ s ∈ [s0 , k − 2], the map

(δ, u) → f(δ, u) is in C 2 ([0, δ0 ] × H s ; H s )

and satisfies the tame estimates

f(δ, u)s , (∂u f)(δ, u)s , (∂δ f)(δ, u)s , (∂δ,u

2
f)(δ, u)s , (∂uu f)(δ, u)s
≤ C(s, us0 )(1 + us ). (2.41)

The proof is similar to that in [3, Section 2]. The above properties of the composition
operator hold true in any compact Riemannian manifold (i.e., they do not exploit Lie
group properties).

3. Abstract Nash-Moser theorem with parameters

We recall now the abstract Nash-Moser theorem [5] (with a few notations adapted to
the present setting).
Let (Xs ,  s )s≥0 be a scale of Banach spaces such that

∀s ≤ s, Xs  ⊆ Xs , us ≤ us  ,
∀u ∈ Xs  .

Let (H (N) )N≥0 be an increasing family of closed subspaces of s≥0 Xs with projectors
(N) : X0 → H (N) satisfying the “smoothing” properties: ∀s ≥ 0, ∀d ≥ 0,
(S1) we have

(N) us+d ≤ N d us , ∀u ∈ Xs ; (3.1)

(S2) we have

(I − (N) )us ≤ N −d us , ∀u ∈ Xs+d . (3.2)

We consider a C 2 -map F : [0, δ0 ) × Xs0 +ν → Xs0 , where s0 ≥ 0 , ν > 0, δ0 > 0,

satisfying
510 BERTI and PROCESI

• (F1) F (0, 0) = 0 and the tame properties ∃ S ∈ (s0 , ∞] such that ∀s ∈ [s0 , S),
∀u ∈ Xs+ν with us0 ≤ 2, ∀δ ∈ [0, δ0 ),
• (F2) ∂δ F (δ, u)s ≤ C(s)(1 + us+ν ), Du F (δ, 0)[h]s ≤ C(s)hs+ν ,
• (F3) Du2 F (δ, u)[h, v]s ≤ C(s)(us+ν hs0 vs0 + vs+ν hs0 +
hs+ν vs0 ), and
• (F4) ∂δ Du F (δ, u)[h]s ≤ C(s)(hs+ν + us+ν hs0 ).
The main assumption concerns the invertibility of the linear operators
L(N) (δ, u) := (N) Du F (δ, u)|H (N ) . We consider two parameters κ ≥ 0, σ ≥ 0,
such that

σ > 4(κ + ν), s̄ := s0 + 4(κ + ν + 1) + 2σ < S. (3.3)

For all γ > 0, we define appropriate subsets


Jγ(N)
,κ ⊆ (δ, u) ∈ [0, δ0 ) × H (N) |L(N) (δ, u) is invertible and ∀s ∈ {s0 , s̄},

Nκ 
L(N) (δ, u)−1 [h]s ≤ (hs + us hs0 ) , ∀h ∈ H (N) . (3.4)
γ

Given k > 0, we define U(N)

k := {u ∈ C 1 ([0, δ0 ), H (N) ) | us0 ≤ 1, ∂δ us0 ≤ k}
and, for u ∈ Uk ,
(N)

  

,κ (u) := δ ∈ [0, δ0 ) (δ, u(δ)) ∈ Jγ ,κ .
Gγ(N) (N)
(3.5)

ASSUMPTION (L)
There exist σ ≥ 0, κ ≥ 0 satisfying (3.3), γ̄ > 0, M ∈ N, C > 0, such that

c 
∀γ ∈ (0, γ̄ ], ∀δ ∈ (0, δ0 ],  Gγ(M) 
,κ (0) ∩ [0, δ) ≤ Cγ δ. (3.6)

For all γ ∈ (0, γ̄ ], k̄ > 0, ∃δ̃ := δ̃(γ , k̄) ∈ (0, δ0 ] such that, ∀δ ∈ (0, δ̃], N  ≥ N ≥
(N  )
M , u1 ∈ U(N)
k̄ , u2 ∈ Uk̄ with u2 − u1 s0 ≤ N −σ ,

c 
(N) c  
 (N  )  γδ
 Gγ ,κ (u2 ) Gγ ,κ (u1 ) [0, δ) ≤ C . (3.7)
N
The following Nash-Moser-type theorem holds.

THEOREM 3.1 ([5])

There exists C > 0, and, ∀γ ∈ (0, γ̄ ), there are δ0 > 0 and a C 1 -map u : [0, δ0 ) →
Xs0 +ν such that u(0) = 0 and F (δ, u(δ)) = 0 except in a set Cγ of Lebesgue measure
|Cγ | ≤ Cγ δ0 . Moreover, for all δ ∈ (0, δ0 ), |Cγ ∩ [0, δ)| ≤ Cγ δ .
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 511

4. The bifurcation equation

By the hypothesis (ND), the subspace H s ⊂ Heven s
is invariant for the NLW equation
(1.8), and there is a nondegenerate solution of the equation (1.15) in Q ∩ H s . Then,
by the standard implicit function theorem, we solve in H s the bifurcation equation
(1.11).

LEMMA 4.1
There exist δ0 , R0 > 0, and a C 2 -map (δ, p) → q(δ, p) defined on

[0, δ0 ] × {p ∈ P ∩ H s : ps < R0 } → Q ∩ H s , ∀ s ∈ [s0 , k − 2],

such that q(0, 0) = q̄ and q(δ, p) solves the bifurcation equation in (1.11). Moreover,


∂p q(δ, p)[h] = −χJ −1 Q (∂u f)(δ, q(δ, p) + p) h , ∀h ∈ Q ∩ H s , (4.1)

where J : Q → Q is the linear operator


J [H ] := J (δ, p)[H ] := H + χQ (∂u f)(δ, q(δ, p) + p)H , ∀H ∈ Q, (4.2)

and

q(δ, p)s , ∂δ q(δ, p)s , ∂p q(δ, p)s , ∂δ,p

2
q(δ, p)s , ∂pp q(δ, p)s
≤ C(s)(1 + ps ). (4.3)

Proof
Note that the operator J (δ, p) in (4.2) reduces, for δ = 0, p = 0, to h → h +
χQ (ma(x)q̄ m−1 h), which is invertible in Q ∩ H s by (ND). Finally, (4.3) holds
simply because Q is finite-dimensional. 䊐

4.1. The nondegeneracy condition

We now consider M = SU(n). We choose j0 = w1 corresponding to the smallest
nonzero eigenvalue of −. By Section 2.4.2, the space Q defined in (1.10) is Q =
cos(t) × (Mw1 ⊕ Mwn−1 ).

LEMMA 4.2
Let n = 2, 4. If f (x, u) = u3 + r(x, u) where r(x, u) satisfies (1.7) for some
m > 3 and r(x, u) = r(x −1 , u), then the non-degeneracy condition (ND) holds in the
invariant subspace
 
H s := u(t, x) ∈ Hevens
| u(t, x) = u(t, x −1 ) . (4.4)
512 BERTI and PROCESI

Remark 4.1
A real function u(x) satisfies u(x) = u(x −1 ) if and only if its Fourier coefficients
uj := SU(n) u(x)ēj (x) = u†j (defined in (2.5)) are self-adjoint. We use that x → x −1
is measure preserving. This is the analogue of the even functions.

Proof of Lemma 4.2

The bifurcation equation (1.12) reduces to

q = Q q 3 , q ∈ Q ∩ H s, (4.5)

because a(x) = 1, χ = −1 by (1.14), and m = 3. By (2.21) and Remark 4.1,

 
Q∩H s = qA := (cos t)tr(A(x⊕ x̄)) where A := A1 ⊕ Ā1 , A1 ∈ Mat(Cn ), A1 = A†1 .

In order to solve (4.5), we use the formulas in (2.25).

LEMMA 4.3 √
The function ( 2/3) (cos t) χV ⊕V ∗ (x) is a real nondegenerate solution of (4.5).

Proof
= tr (x ⊕ x̄) = tr (x) + tr (x̄). The function αqI =
The character χV ⊕V ∗ (x) √
α(cos t)χV ⊕V ∗ with α := 2/3 is a solution of (4.5) in Q ∩ H s . Indeed, by (2.22)
and (2.25) with Aj = I , we get

3
Q (αqI )3 = α 3  cos t (cos t)3 Mw1 ⊕Mwn−1 tr (x) + tr (x̄)
3

= α 3 (cos t)6 tr (x) + tr (x̄) = αqI .
4
We have to prove that the linearized equation hA − 3α 2 Q qI2 hA = 0, hA ∈ Q ∩ H s
has the only solution hA = 0. This amounts to showing that



2 tr (A1 x) + tr (Ā1 x̄) − Mw1 ⊕Mwn−1 (tr (x) + tr (x̄))2 tr ((A1 ⊕ Ā1 )(x ⊕ x̄)) = 0
(4.6)
has the only solution A = A1 ⊕ Ā1 = 0. By (2.3) and (2.4), we have

2

tr ((I ⊕I )(x ⊕ x̄)) tr A(x ⊕ x̄) = tr (BρU ), where B := (I ⊕I )⊗2 ⊗(A1 ⊕ Ā1 ).

As in (2.25), we find


6 
6
Mw1 tr (BρU ) = tr (Bj x), Mwn−1 tr (BρU ) = tr (B̄j x̄),
j =1 j =1
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 513

where Bj represents Vj B |Vj and B̄j represents Vj∗ B |Vj∗ . Since V , V ∗ are irre-
ducible, (4.6) becomes


6 
6
2A1 = Bj , 2Ā1 = B̄j . (4.7)
j =1 j =1

LEMMA 4.4
We have B1 = B3 = A1 , B4 = B2 ,
n A1 tr (A1 )
B2 = tr (A1 )I − 2 , B5 = I,
n2 − 1 n −1 n
n2 + 1 2A1
B6 = tr (A1 )I − 2 .
n −n
3 n −1

Proof
(j ) (j )
The coefficients (Eh , BEk ) are computed by Lemma 2.12 and recalling that A1 =
A†1 . 䊐

By Lemma 4.4, the equations (4.7) reduce to A1 = ntr (A1 )I . Then we get tr (A1 )(n2 −
1) = 0, and, for n = 1, tr (A1 ) = 0 and A1 = 0. 䊐

5. The range equation and the proof of Theorem 1.2

We solve the range equation

F (δ, p) := D(ω)p − εF (δ, p) = 0, (5.1)

where ω := ω(δ) := (ωj20 + χδ m−1 )1/2 , ε := δ m−1 ,


D(ω) := ω2 ∂tt −  + μ and F (δ, p) := P f δ, q(δ, p) + p , (5.2)

applying the abstract Nash-Moser implicit function Theorem 3.1 in the scale Xs :=
s
Heven ∩ P . To simplify notation, we solve the range equation in the larger space
s
Heven ∩ P (not in the subspace H s ∩ P ).
For all N ∈ N, we consider the finite-dimensional subspaces
 
H (N) := Nk ∩ Heven
0
⊂ Xs , where EN+ := {k ∈ E + | |k + ρ | ≤ N}
k∈EN+ ∩P s≥0

(P is defined in (1.9) and Nk in (2.26)) and the corresponding projectors (N) : Xs →


H (N) , (N) u := k∈EN+ ∩P Nk u. The smoothing properties (S1) and (S2) in (3.1)
and (3.2) hold.
514 BERTI and PROCESI

LEMMA 5.1
Let s0 > 1 + (d + n + 1)/2 and f ∈ C k with k > s0 + 2. For all s ∈ [s0 , k − 2], the
map F ∈ C 2 ([0, δ0 ) × Xs+2 , Xs ), F (0, 0) = 0, and tame properties (F2)–(F4) hold
(with ν = 2 and S = k − 2).

Proof
The composition operator (δ, u) → f(δ, u) is C 2 by Lemma 2.22, and, by Lemma
4.1, also the map (δ, q) → q(δ, p) is C 2 . Properties (F2)–(F4) follow by (2.41) and
(4.3). 䊐

The main issue is to prove Assumption (L) concerning the invertibility of the linearized
operators

L(N) (δ, p) := (N) (D(ω) − εDp F (δ, p))|H (N ) . (5.3)

It will be a consequence of the following lemma proved in the next section.

LEMMA 5.2 (Invertibility)

For all τ > 1, there is s0 := s0 (M, τ ) such that if
γ
|ω2 m − n| ≥ , γ ∈ (0, 1), ∀(m, n) ∈ Z2 \ {(0, 0)}, (5.4)
n3/2
ε(ps0 + 1) ≤ c(γ , k) is small enough, and
  4K τ
 (K) 
∀1 ≤ K ≤ N, (L (δ, p))−1  ≤ , (5.5)
0 γ

then L(N) (δ, p) is invertible and, ∀s ∈ [s0 , k − 1],

  Nκ
 (N) 
(L (δ, p))−1 h ≤ (hs + ps hs0 ), ∀h ∈ H (N) , (5.6)
s γ
with κ := (4τ + 3r + 2n + d + 7)/2.

We now complete the proof of property (L). We fix, in (5.5),

 m−2 
τ ≥ max τ1 , 2 + d + n + (τ1 + 2) , γ ∈ (0, γ̄ ], γ̄ ≤ γ1 , (5.7)
m−1
(τ1 and γ1 are given in (1.6)) and, correspondingly, the parameter
 
σ := max 4(κ + 2), d + n + 3 + 1.
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 515

(The first condition is the first inequality in (3.3) with ν = 2.) We then take f ∈ C k
with

k > s0 + 4(κ + 3) + 2σ + 2, (5.8)

which is the second condition in (3.3) with S := k − 2.

For δ0 > 0 small enough, by Lemma 5.2, defining
 
(N) 

,κ := (δ, p) ∈ [0, δ0 ) × H
Jγ(N) ps0 ≤ 1, ω(δ) satisfies (5.4) and (5.5) holds ,

the inclusion (3.4) is satisfied. Given a function δ → p(δ) in U(N) (N)

k , the set Gγ ,κ defined
in (3.5) is
 
,κ (p) =
Gγ(N) GK G, (5.9)
1≤K≤N

where
   Kτ 
 
GK := δ ∈ [0, δ0 )  (L(K) (δ, p(δ)))−1  ≤ 4
0 γ
and
  
G := δ ∈ [0, δ0 )  ω(δ) satisfies (5.4) .

LEMMA 5.3
Let μ be diophantine. If ε0 γ1−1 M τ1 +2 is small enough, then Gγ(M)
,κ (0) = G . Moreover,
G satisfies (3.6).

Proof
The eigenvalues of L(K) have the form −ω2 l 2 +ωj2 +O(ε) (l ∈ Z, (j1 , j2 ) ∈ + × Zn )
with |(l, j1 + ρ, j2 )| ≤ K . Hence, by (1.6), if ε0 γ1−1 K τ1 +2 is small enough, all the
eigenvalues of L(K) have modulus ≥ γ1 /(4K τ1 ) ≥ γ /(4K τ ). Then GK = [0, δ0 ). The
measure estimate for G is standard. 䊐

LEMMA 5.4
The measure estimate (3.7) holds.

The proof is in the appendix.

All the assumptions of the Nash-Moser theorem (Theorem 3.1) are satisfied, and we
deduce Theorem 1.2.
516 BERTI and PROCESI

6. Inversion of the linearized equation: Proof of Lemma 5.2

It is more convenient to prove Lemma 5.2 in the larger scale of complex Sobolev

spaces Xs = H s ∩ k∈P Nk . Then the same estimates hold in Heven
s
∩ P.

According to the orthogonal decomposition H (N)
= k∈EN+ ∩P Nk , we represent


h → L(N) [h] := (N) D(ω) + εb(t, x)(h + (∂p q)[h]) , ∀h ∈ H (N) ,

where b(t, x) := (∂u f)(δ, q(δ, p) + p) ∈ Heven

s
, ∂p q := ∂p q(δ, p) by the block matrix

L(N) = D + εT ,

where D := diagk∈EN+ ∩P (Dk Ik ) with small divisors

Dk := −ω2 l 2 +|j1 +ρ|2 −|ρ|2 +|j2 |2 +μ, k := (l, j1 , j2 ) ∈ Z ×+ × Zn , (6.1)

Ik is the identity map in Nk , and

 
T := (Tkk )k,k ∈EN+ ∩P , Tkk := Nk L(N) |Nk ∈ L(Nk , Nk ). (6.2)

The main properties to establish are

(i) separation of the singular sites (see Section 6.1);
(ii) decay estimates for the block matrix T (see Section 6.2).

6.1. Regular and Singular sites

Definition 6.1
For fixed η ∈ (0, 1), we define the regular sites R and the singular sites S as
     
R := k ∈ E + ∩ P  |Dk | ≥ η and S = R c = k ∈ E + ∩ P  |Dk | < η .
(6.3)
We shall assume that there is a partition of the singular sites

S= α , (6.4)
α∈N

where the “clusters” α ⊂ E + are pairwise disjoint and satisfy

(H1) (dyadic) Mα ≤ 2mα , ∀α , where Mα := maxk∈α |k + ρ |, mα := mink∈α |k +
ρ |,
(H2) (separation) ∃λ, C̄ > 0 such that d(α , β ) ≥ C̄(Mα + Mβ )λ , ∀α = β,
where d(α , β ) := mink∈α ,k ∈β |k − k  | and λ depends only on M .
We denote the restrictions of S , R , α to EN+ (for each N ) with the same symbols.
Actually, the Diophantine condition (5.4), (6.1), and Lemma 2.6 imply the
following.
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 517

LEMMA 6.1
Assume (5.4). There exists η0 (γ ) such that, if η ∈ (0, η0 (γ )], there exists a partition
of the singular sites S like (6.4) satisfying (H1) and (H2).

The proof, which essentially follows the scheme of [4], is in the appendix.

6.2. Block off-diagonal decay estimates

We now consider the operator B0 : H s ∩ P → H s ∩ P defined by

 (4.1)

B0 (h) := P b(t, x)(∂p q)[h] = −χP b(t, x)J −1 Q (b(t, x)h) .

LEMMA 6.2
The matrix which represents B0 is self-adjoint and belongs to As (P ), and |B0 |s ≤
K(s)|B|s .

Proof
The fact that B0 is self-adjoint is immediate by writing


B0 (h) = −χP b(t, x)Q J −1 Q (b(t, x)P h) .

The matrix which represents J is in As (Q), and Q is finite-dimensional (see (1.9)). By

the algebra property, J −1 ∈ As (Q), as well, and |J −1 |s ≤ C(s). If B represents the
multiplication by b(t, x), by Lemma 2.20 the matrix B ∈ As as well as the matrices
which represent the projections P , Q (see (2.40)). Applying Lemmas 2.16 and
2.17, we deduce |B0 |s ≤ C(s)|B|s |B|s0 . 䊐

By Lemmas 2.20 and 6.2, we finally get the following.

LEMMA 6.3

Consider a real b ∈ Hevens+a
with a ≥ (d + r + 2n + 3)/2. The matrix T = (Tkk )k,k ∈EN+
defined in (6.2) is self-adjoint and belongs to the algebra of polynomially localized
matrices As (EN+ ) with

|T |s ≤ K(s)bs+a . (6.5)

Note also that, since the index |k + ρ|, |k  + ρ| ≤ N , then ∀s > a ,

(6.5)
|T |s ≤ (2N)a |T |s−a ≤ K  (s)N a bs . (6.6)
518 BERTI and PROCESI

6.3. Reduction along the regular sites

According to the decomposition
 
H (N) := HR ⊕ HS , where HR := Nk , HS := Nk ,
k∈R∩EN+ k∈S∩EN+

we represent L(N) as the self-adjoint block matrix

 
LR LSR
L =
(N)
, where LSR = (LRS )† , L†R := LR , L†S = LS .
LRS LS
For the sequel, we fix
(2 + λ)(r + n + 1) + 2τ + 1
s1 := , (6.7)
2λ
and we always assume that
εγ −1 η−1 (1 + |T |s1 ) ≤ c(k) is small enough. (6.8)
Moreover, we consider the index s ∈ [s1 , k − 1].

LEMMA 6.4
The operator LR is invertible, |L−1 −1
R |s1 ≤ 2η , and for all s ∈ [s1 , k − 1],

|L−1 −1 −1
R |s ≤ C(s)η (1 + εη |T |s ), (6.9)


L−1
R hs ≤ C(s)η
−1
(1 + εη−1 |T |s ) hs1 + hs , ∀h ∈ H (N) . (6.10)

Proof
By (6.3), the diagonal matrix DR satisfies |DR−1 |s ≤ η−1 , ∀s ≥ 0. By (2.32), the
Neumann series

L−1
R = (−ε)m (DR−1 TR )m DR−1 (6.11)
m≥0

is totally convergent in | · |s1 with |L−1 −1 −1

R |s1 ≤ 2η , taking εη |T |s1 ≤ c(s1 ) small
enough.

Proof of (6.9)
We obtain, ∀m ∈ N,
(2.32)
εm |(DR−1 TR )m DR−1 |s ≤ εm C(s)|(DR−1 TR )m |s |DR−1 |s
(2.36)
m−1 −1
≤ C(s)εm m C(s)|DR−1 TR |s1 |DR TR |s η−1
(2.32)
m−1
≤ C  (s)εmη−2 εC(s)η−1 |T |s1 |T |s .

Then (6.9) follows by (6.11) for εη−1 |T |s1 < c(s1 ) small enough. 䊐
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 519

Proof of (6.10)
By the interpolation estimates (2.35) and (6.9),

L−1 −1 −1
R hs ≤ C(s)(|LR |s hs1 + |LR |s1 hs )


≤ C  (s) η−1 (1 + εη−1 |T |s )hs1 + 2η−1 hs

since |L−1 −1
R |s1 ≤ 2η . 䊐

The invertibility of L(N) is reduced to that of the “quasi-singular” matrix

L := LS − LRS L−1 S
R LR (6.12)

via the “resolvent-type” identity

   −1  
(N) −1 I −L−1 R LR
S
LR 0 I 0
(L ) = . (6.13)
0 I 0 L−1 −LRS L−1
R I

Note that L ∈ As (S) because L is the restriction to S of the polynomially localized

matrix
 
−1 I 0
IS (L − IS LIR L̃ IR LIS )IS ∈ As , where L̃ = ∈ As
0 LR

and IS , IR ∈ As are the projection matrices on S , R (see (2.40)).

6.4. The quasi-singular matrix L

By (6.4), the singular sites restricted to EN+ are

S= α , where IN := {α ∈ N | mα ≤ N}
α∈IN

and α ≡ α ∩ EN+ . According to the decomposition HS := α∈IN Hα , where

k∈α Nk , we represent L as the block matrix L = (Lα )α,β∈IN , where
β
Hα :=
Lβα := Hα L|Hβ . We write

L=D +T , where D := diagα∈IN (Lα ), Lα := Lαα , T := (Lβα )α=β .

LEMMA 6.5 (Decay of T )

For all ∀α = β,
C(s)ε|T |s
Lβα 0 ≤ . (6.14)
d(α , β )s−r+n+1/2
520 BERTI and PROCESI

Proof
Since T and L−1
R belong to the algebra of polynomially localized matrices As , the
matrix T = εTS − ε 2 TSR L−1 S
R TR also belongs to As . Then (6.14) follows by Lemma
2.18 once we prove that

|T |s ≤ ε|T |s + ε2 |TSR L−1

R TR |s ≤ C(s)ε|T |s .
S
(6.15)

The last bound in (6.15) follows from

(2.34)
|TSR L−1
R TR |s
S
≤ C(s)(|TSR |s1 |L−1 −1 S
R TR |s + |TS |s |LR TR |s1 )
S R

(2.34),(2.32)

≤ C  (s) |T |s1 (|L−1 −1 −1
R |s1 |T |s + |LR |s |T |s1 ) + |T |s |LR |s1 |T |s1
(6.9)
≤ C1 (s)η−1 |T |s1 |T |s

using |L−1 −1 −1
R |s1 ≤ 2η , |T |s1 ≤ |T |s , and εη |T |s1 ≤ 1. 䊐

By the assumption (5.5), we deduce the following lemma (proved in the appendix).

LEMMA 6.6
For all ∀α ∈ IN , Lα is invertible and L−1
α 0 ≤ Cγ
−1
Mατ .

Then the block diagonal matrix D is invertible. Moreover, we have the following.

LEMMA 6.7 (Inversion of D )

For all s ≥ 0, D −1 s ≤ C(s)γ −1 N τ .

Proof

Let u = α∈IN uα , uα ∈ Hα . By Lemma 6.6, and since each α is dyadic (see (H1)),
 (3.1)  C  2(s+τ )
D −1 u2s = L−1
α u α s ≤
2
Mα2s L−1
α u α 0 ≤
2
Mα uα 20
γ2
α∈IN α∈IN α∈IN

(3.2) C  2(s+τ ) uα 2s (H 1) C  2(s+τ ) s uα 2s C(s)

≤ Mα ≤ 2 Mα 2 ≤ 2 N 2τ u2s
γ2 m2s
α γ M 2s
α γ
α∈IN α∈IN

because Mα ≤ N . 䊐

We now prove that L is invertible and L−1 satisfies an interpolation inequality in high
Sobolev norm.
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 521

LEMMA 6.8 (Inversion of L)

L is invertible, L−1 s1 ≤ Cγ −1 N τ , and, ∀s ∈ [s1 , k − 1],
K(s) κ1
L−1 us ≤ N (|T |s us1 + |T |s1 us ), ∀u ∈ HS , (6.16)
γ
where κ1 := 2τ + r + n + 1.

Proof
If C(s1 )D −1 T s1 < 1/2, then the Neumann series

L−1 = (I + D −1 T )−1 D −1 = (−1)m (D −1 T )m D −1 (6.17)
m≥0

is convergent in operator norm  s1 and, using Lemma 6.7, L−1 s1 ≤ Cγ −1 N τ .

LEMMA 6.9
We have C(s1 )D −1 T s1 < 1/2.

Proof

For all u = β∈IN uβ ∈ HS , uβ ∈ Hβ ,
  
  −1 β 2 (3.1)  2s1   −1 β 2

D −1 T u2s1 =  Lα Lα uβ  ≤ Mα  Lα Lα uβ 
s1 0
α∈IN β∈IN α∈IN β∈IN

C  2s1   τ β 2
≤ M α M α Lα  0 u β  0 (6.18)
γ2
α∈IN β∈IN

(3.2)
using Lemma 6.6. Since uβ 0 ≤ uβ s1 /msβ1 , Cauchy inequality in (6.18) gives

  Lβ 2 
D −1 T u2s1 ≤ γ −2 Mα2(s1 +τ ) α 0
uβ 2s1
α∈IN β∈IN
m2s
β
1
β∈IN

(6.14),(H 2)  M 2(s1 +τ ) C(s1 )2 ε2 |T |2s1

≤ γ −2 α
u2s1
m2s
β
1
(Mα + Mβ )λ(2s1 −(r+n+1))
α,β∈IN

(6.7)  1  1
≤ γ −2 C(s1 )2 ε2 |T |2s1 u2s1
α∈N
Mαr+n+2 β∈N
mr+n+2
β

−2
≤ γ 2 2
K(s1 ) ε |T |2s1 u2s1 , (6.19)

 
−(r+n+2) 2
where K(s1 )2 = C(s1 )2 k∈E + k is independent of N . By (6.8), we prove
the lemma. 䊐
522 BERTI and PROCESI

LEMMA 6.10
We have D −1 T us ≤ εK(s)γ −1 (N κ0 |T |s us1 + |T |s1 us ), where κ0 := τ + r +
n + 1.

Proof
We estimate D −1 T u2s as in (6.18) (with s1 substituted by s ) as follows:
   2
−1 −2 s
D T us ≤ γ 2
2
Mα2s Mατ Lβα 0 uβ 0
α∈IN β∈IN :Mα ≤5mβ

   2 
+ Mα2s Mατ Lβα 0 uβ 0 =: γ −2 2s (S1 + S2 ).
α∈IN β∈IN :Mα >5mβ

Estimate of (S1 ). Proceeding as in (6.19) with s1 substituted by s , we obtain

S1 ≤ K12 ε2 |T |2s1 u2s ,

where K1 is independent of N , because

 Mα2(s+τ )  C(s1 )2
K12 =
α∈IN
Mαλ(2s1 −(r+n+1)) β∈IN :Mα ≤5mβ
m2s
β

 Mα2(s+τ )  C(s1 )2
≤ λ(2s1 −(r+n+1))
Mα k∈E + : k≥Mα /5
k2s
α∈IN

 Mα2(s+τ ) C1 (s)  Mα2τ +r+n+1 (6.7)

≤ ≤ C1 (s) ≤ K(s).
α∈IN
Mαλ(2s1 −(r+n+1)) Mα2s−(r+n+1) α∈IN
Mαλ(2s1 −(r+n+1))

Estimate of (S2 ). Applying Cauchy inequality in S2 , we get

(3.2)   Lβα 20 

S2 ≤ Mα2(s+τ ) uβ 2s1
α∈IN β∈IN :Mα >5mβ
m2s
β
1
β∈IN

(6.5)   ε2 |T |2s C(s)2

≤ Mα2(s+τ ) u2s1
α∈IN β∈IN :Mα >5mβ
m2s1
β d(α , β )
2s−(r+n+1)

  ε2 |T |2s C(s)2
≤ Mα2(s+τ ) 2s1 2s−(r+n+1)
u2s1
α∈IN β∈IN
m β Mα
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 523

because Mα > 5mβ implies mα > 5Mβ /4 and d(α , β ) > Mα /10. Then
  1
S2 ≤ C(s)2 ε2 |T |2s u2s1 Mα2τ +r+n+1 ≤ C  (s)2 ε2 |T |2s u2s1 N 2(τ +r+n+1)
α∈IN β∈N
m2s
β
1

using |IN | ≤ CN r+n+1 and 2s1 > r + n + 2. 䊐

LEMMA 6.11
Letting A := D −1 T , we have, ∀m ∈ N,

m
Am us ≤ εγ −1 K(s) (mN κ0 |T |s |T |m−1
s1 us1 + |T |s1 us ).
m
(6.20)

Proof
We proceed by induction. Lemma 6.10 implies (6.20) for m = 1. Then, again by
Lemma 6.10,

Am us ≤ εγ −1 K(s)(N κ0 |T |s Am−1 us1 + |T |s1 Am−1 us )

(6.19)

≤ εγ −1 K(s) N κ0 |T |s (K(s1 )εγ −1 |T |s1 )m−1 us1 + |T |s1 Am−1 us
(6.20)
m
≤ εγ −1 K(s) (mN κ0 |T |s |T |m−1
s1 us1 + |T |s1 us )
m

using K(s) ≥ K(s1 ). 䊐

By (6.17),

L−1 us ≤ D −1 us + (D −1 T )m (D −1 u)s ≤ D −1 us
p≥1

+D −1 us (εγ −1 K(s)|T |s1 )m
m≥1

(6.20) 
+ N κ0 K(s)εγ −1 |T |s D −1 us1 m(K(s)εγ −1 |T |s1 )m−1
m≥1

(6.8)
≤ γ −1 N κ0 +τ K1 (s)(us + ε|T |s us1 )

and using Lemma 6.7. This completes the proof of Lemma 6.8. 䊐

Proof of Lemma 5.2

By (5.4) and Lemma 6.1, taking η := η0 (γ ), we deduce the existence of a partition
like that in (6.4) satisfying (H1) and (H2).
By (6.5), the smallness condition (6.8) is implied by taking s0 := s1 + a , with
s1 defined in (6.7) and a := (d + r + 2n + 3)/2, and ε(1 + ps0 ) ≤ c(γ , k) small
524 BERTI and PROCESI

enough. Note that s0 depends only on M and τ because λ, a depend only on M (see
(H2) and Lemma 6.1).
By the resolvent identity (6.13), setting u = uS + uR with uS ∈ HS , uR ∈ HR ,

(L(N) )−1 us ≤ L−1 −1 R −1 S −1 −1 S −1

R uR +LR LS L (uS +LR LR uR )s +L (uS +LR LR uR )s .

To prove (5.6), we apply repeatedly the interpolation estimates (2.34) and (2.35). We
apply Lemma 6.4 to bound L−1 −1
R s1 and LR us ; we use Lemma 6.8 for L us
−1

and (2.35) to estimate LR us . By applying (6.6), we finally obtain

S

 (N) 

(L (δ, p))−1 u ≤ K1 (s) N κ us + εDp F (δ, p)s us1
s γ
Nκ
≤ (us + ps us0 )
γ
with κ := κ1 + a + 1 = 2τ + r + n + a + 2 and N sufficiently large. 䊐

7. The NLS equation

We look for periodic solutions of the vector NLS equations

iωu+t − u =
+
εf(δ, u+ u− )u+ ,
(7.1)
−iωut − u = εf(δ, u+ u− )u− ,
− −

where u := (u+ , u− ) ∈ Hs := H s ⊕ H s . Note that (7.1) reduces to the scalar NLS

equation (1.18) in the invariant subspace
  
V s := u := (u+ , u− ) ∈ Hs  ū+ = u− . (7.2)

Remark 7.1
The reason for doubling the equations, instead of working directly with the scalar
NLS, is that the matrix which represents u → ū is not Töplitz.

The Fourier analysis of Section 2.5 can be applied to u := (u+ , u− ) ∈ L2 (T ×M, C2 )

expanding

u= uk , where uk := Nk u = (Nk u+ , Nk u− ).
k∈E +

We then consider the linear operators on Hs represented by polynomially localized


matrices defined as in Section 2.6, with the only difference now the blocks Akk ∈
L(Nk ⊕ Nk , Nk ⊕ Nk ). All the results of Section 2.5 remain valid. We point out that
Proposition 2.19 has to be stated as follows.
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 525

LEMMA 7.1 (Multiplication operator)

Given f (t, x), g(t, x) ∈ H s with f (t, x) ∈ R real,
 
f (t, x) g(t, x)
B :=
ḡ(t, x) f (t, x)

is self-adjoint and, ∀s > (d + n + 1)/2,

 (f s + gs )
Bkk  ≤ C(s) . (7.3)
k − k  s−(d+n+1)/2

Proof
The B is self-adjoint since f is real. By definition,
 
k Nk f (t, x)Nk Nk g(t, x)Nk
Bk := .
Nk ḡ(t, x)Nk Nk f (t, x)Nk

Recall that k = (l, j1 , j2 ) with l ∈ Z, j1 ∈ + , and j2 ∈ Zn . If (ρ, j1 − j1 ) > 0, we

proceed for each component as in Proposition 2.19, obtaining (7.3). The bound for
 
Bkk with (ρ, j1 − j1 ) < 0 is obtained by the relation Bkk = (Bkk )† . 䊐

The resonant subspace for the vector NLS equation (7.1) is Q := Q ⊕ Q∗ , where

  
dj 
∗
Q := q = e ilt
ql,j,σ ej,σ ∈ H 0 | ωj20 l + ωj2 = 0 .
(l,j )∈Z×M σ =1

We perform the Lyapunov-Schmidt reduction as in the previous sections along Q

and P := Q⊥ . In view of (ND) we restrict further the NLS equation in the invariant
subspace H s ⊕ H s . Restricting p := (p + , p − ) to V s (defined in (7.2)), the implicit
function theorem and the nondegeneracy assumption (ND) (lifted to V s ) imply the
existence of a C 2 -solution q(δ, p) ∈ V s ∩ Q of the bifurcation equation. The tame
properties as in Lemma 4.1 follow by the assumption |J0−1 |s ≤ C(s).
It remains to solve the range equation. Setting u(δ, p) := p + q(δ, p), p ∈ V s ,
we define
 
iω∂t −  0

D(ω) := , F (δ, p) := f δ, |u+ (δ, p)|2 u(δ, p),
0 −iω∂t − 

and we apply the abstract Nash-Moser theorem (Theorem 3.1) to

F (δ, p) := D(ω)p − εF (δ, p) = 0.

526 BERTI and PROCESI

We have to prove Assumption (L) concerning the linearized operators


L(N) (δ, p) := (N) D(ω) − εDp F (δ, p) |H(N ) ,

where
 
f(δ, |u+ |2 ) + f (δ, |u+ |2 )|u+ |2 f (δ, |u+ |2 )(u+ )2
Dp F (δ, p) := ε  + 2 − 2
f (δ, |u | )(u ) f(δ, |u | ) + f (δ, |u+ |2 )|u+ |2
+ 2


1 + Dp q(δ, p) .

Property (L) is a consequence of the following lemma, similarly to Section 5.

LEMMA 7.2 (Invertibility)

For all τ > 1, there are constants κ > 0, s0 := s0 (M, τ ), such that, if ε(ps0 + 1) ≤
c(γ , k) is small enough and
  CK τ
 (K) 
∀1 ≤ K ≤ N, (L (δ, p))−1  ≤ , (7.4)
0 γ

then L(N) (δ, p) is invertible and an estimate like (5.6) holds.

Proof of Lemma 7.2

It requires only small modifications in the proof of Lemma 5.2. First, note that the
diophantine hypothesis (5.4) is not necessary to get a partition of the singular sites as
in (6.4). It can be proved for any ω as in the case M = Tn by convexity arguments
as in [9] or [21]. The key observation is that the eigenvalues ωj2 in (1.4) are similar to
those of a torus, and the small divisors appear close to the “paraboloid” ωl − ωj2 = 0.
Then, following Section 6.2 and using Lemma 7.1, we obtain the bound (6.5).
Sections 6.3 and 6.4 remain valid because they are based on purely algebraic tech-
niques: they require only that the operator T belong to the algebra of polynomially
localized matrices.
We can finally prove property (L). The sets Gγ(N),κ (p) are defined as in (5.9), and
G ≡ [0, δ0 ). The analogue Lemma 5.3 states that, if ε0 M is small, then Gγ(M),κ (p) = G .
We then follow Section A.1 word by word. To prove (A.3) we proceed as follows.
We consider the parameter λ := ω−1 and


L(K) (λ, p) := L(K) δ(λ), p λ

which is a self-adjoint perturbation of

 
(K) i∂t − λ 0
 .
0 −i∂t − λ |H(K)
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 527

We consider frequencies ω close to ωj0 > 0. Note that δ → λ(δ) = ω−1 (δ) is a
diffeomorphism in [δ1 /2, δ1 ] with λ (δ) ≥ cδ12m−1 . Now we estimate the measure of
  1 
λ ∈ ωj−1 ,  ∃ an eigenvalue of L(K) (λ, u) with modulus in
0
ωj0 − δ1
2m

 −τ
+ Cδ12m N −σ 
−τ γ K
ωj−1 γ K ,
0
ωj0 − δ12m
by following [4, Lemma 3.2] (see also Section A.1). The derivative with respect to λ
of an eigenvalue E(λ) of L(K) (λ, u), with eigenspace HE , satisfies


∂λ E(λ) ≥ min ∂λ L(K) (λ, u)v, v 0 ≥ min ∇v20 + O(1). (7.5)
v∈HE ,v0 =1 v∈HE ,v0 =1

We consider eigenvalues with modulus

 −τ
+ Cδ12m N −σ
−τ γ K
|E(λ)| ∈ ωj−1 γ K , ⊂ [−γ , γ ] (7.6)
0
ωj0 − δ12m

where the last inclusion holds for ε small, and N, K large. The projection of any
v ∈ HE , v0 = 1, on the subspace P0 := v(t, ·) orthogonal to ej,σ , ∀|j | ≤ C, ∀t
is ε -small. Then, by (7.5), we get ∂λ E(λ) ≥ 1/2. 䊐

7.1. The NLS equation on SO(3)

We want to prove the existence of small amplitude periodic solutions of the NLS
equation

iut − u = f (|u|2 )u, f (|u|2 ) = |u|2 + r(|u|2 ), r(0) = r  (0) = 0, (7.7)

defined on SO(3) = SU(2)\{±I } (see Section 2.3). The cone of dominant weights is
+SU(2) = N, and henceSO(3) = 2N, by identifying SO(3)  m/4 (1, −1), m ∈ 2N,
with m ∈ 2N (see (2.18)).
The eigenvalues of − in (2.17) are denoted by ωm 2
= ((m + 1)2 − 1)/8 and the
characters by χm .
The multiplication rules (2.20) for SU(2) reduce to


min(h,m)
χh χm = χh+m−2k (7.8)
k=0

with h, m, k ∈ N (see [34, p. 47]).

Since the nonlinearity is x -independent, we look for solutions of the corresponding
rescaled equation (1.18) in the invariant subspace of central functions with real Fourier
528 BERTI and PROCESI

coefficients
  
H s = u(t, x) = ul,m eilt χm (x) ∈ H s , ul,m ∈ R .
(l,m)∈Z×2N

Since i∂t and  leave H s invariant, we only have to prove the following lemma.

LEMMA 7.3
If u ∈ H s , then f (|u|2 )u ∈ H s .

Proof
The space H s is closed under product (use the multiplication rule (7.8)) and complex
conjugation. Then we only need to show that f (|u|2 ) ∈ H s for all u ∈ H s . The main
point is that the space of real-valued functions in H s coincides with the real-valued
functions f (t, x) which are central and even; that is, f (t, x) = f (−t, x −1 ). Indeed,
for f ∈ H s , we have

 

f (t, x) = fl,m e−ilt tr ρm (x) = fl,m e−ilt tr ρm (x −1 ) = f (−t, x −1 )
(l,m)∈Z×2N (l,m)∈Z×2N

using the definition of unitary representation. Hence if f ∈ H s is real, then it is even.

Conversely, if f (t, x) is real valued, central, and even, then fl,m = fl,m . Now if
u ∈ H s , then |u|2 = uū it is in H s and it is real valued. By the above arguments, |u|2
is even. In conclusion, f (|u|2 ) is real valued, central, and even and hence is in H s . 䊐

We look for solutions of (7.7) with frequency close to ω02 = 1/8 = 1/D . Notice that
1/8 is not an eigenvalue of −. This simplifies the calculations because the space Q
in (1.19) simplifies to
  2

Q ∩ H s = q(t, x) = qm e8iωm t χm (x)
m∈2N

(i.e., Q = {(l, m) ∈ Z × 2N : (l/8) − ωm

2
= 0}). The bifurcation equation (1.20) is

iqt + Q |q|2 q = 0, q ∈ Q ∩ H s. (7.9)

The remaining part of this section is devoted to proving the following lemma.

LEMMA 7.4
The bifurcation equation (7.9) admits a nondegenerate solution of the form

q0 (t, x) = q6 ei48t χ6 (x) + q8 ei80t χ8 (x). (7.10)

NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 529

In Fourier coefficients, the equation (7.9) is


8ωh2 qh = qm1 qm2 q̄m3 c(h, m1 , m2 , m3 ) 2
such that ωm 1
+ ωm
2
2
− ωm
2
3
= ωh2 ,
m1 ,m2 ,m3

 (7.11)
where c(h, m1 , m2 , m3 ) := SO(3) χh χm1 χm2 χm3 . We find solutions of (7.11) in finite-
dimensional subspaces.

LEMMA 7.5
For all m > 4, the subspace
 2 2 
Qm := q(t, x) = qm e8iωm t χm (x) + qm+2 e8iωm+2 t χm+2 (x) ⊂ Q ∩ H s (7.12)

is invariant under the map q → Q |q|2 q . The bifurcation equation (7.11) restricted
to Qm is




(m + 1)2 − 1qm = qm (m
+ 1)|qm | + 2(m + 1)|qm+2 |2 , 
2 2
(7.13)
(m + 3) − 1 qm+2 = qm+2 (m + 3)|qm+2 |2 + 2(m + 1)|qm |2 ,

which has a unique solution q0 (t, x).

Proof
By (7.11), in order to prove the invariance of Qm it is sufficient to show that, if
m1 , m2 , m3 ∈ {m, m + 2}, then ωm 2
1
+ ωm2
2
− ωm
2
3
= ωh2 is verified only if h ∈
{m, m + 2}. This holds true because, for m = 4,
2
8(2ωm − ωm+2
2
) = (m − 1)2 − 9 = (h + 1)2 − 1,
8(2ωm+2 − ωm ) = (m + 5)2 − 9 = (h + 1)2 − 1,
2 2

using that the only natural solutions of A2 − B 2 = 8 are (A, B) = (3, 1). We claim
that

c(h, h, k, k) = min{h, k} + 1, (7.14)

which, by (7.11), implies (7.13). By (7.8), we have


h  
h min(k,2(h−i))
χh2 χk = χ2(h−i) χk = χ2(h−i−m)+k .
i=0 i=0 m=0

Now 2(h − i − m) + k = k only if m = h − i which is smaller than 2(h − i). If

h ≤ k , then there are solutions for all 0 ≤ i ≤ h. If h > k , then h − i ≤ k holds for
h − k ≤ i ≤ h. This proves (7.14). 䊐
530 BERTI and PROCESI

We now prove that the solution q0 (t, x) ∈ Qm found in Lemma 7.5 is nondegenerate
for m = 6.

LEMMA 7.6
Let m = 6. The unique solution h ∈ Q ∩ H s of

J0 [h] := iht + 2|q0 (t, x)|2 h + (q0 (t, x))2 h̄ = 0 (7.15)

is h = 0. Moreover, J0 is (2 × 2)-block diagonal, and so |J0−1 |s ≤ C(s).

Proof
By the definition of H s, q6 , q8 , and for hk ∈ R, equation (7.15) is written as
 2 2
Q e8iωk t 8ωk2 hk χk − 2e8iωk t (q62 χ62 + q82 χ82 )hk χk
k∈2N

−2e8i(ω6 −ω8 +ωk )t q8 q6 hk χ6 χ8 χk

2 2 2

−2e8i(ω8 −ω6 +ωk )t q8 q6 hk χ6 χ8 χk − e8i(2ω6 −ωk )t q62 hk χ62 χk −e8i(2ω8−ωk )t q82 hk χ82 χk
2 2 2 2 2 2 2


−2e8i(ω6 +ω8 −ωk )t q6 q8 hk χ6 χ8 χk = 0.
2 2 2
(7.16)

The result is that the unique solution with h, k ∈ 2N of ω62 − ω82 + ωk2 = ωh2 is
(h, k) = (6, 8) and that of ω82 − ω62 + ωk2 = ωh2 is (h, k) = (8, 6). Moreover, for all
h, k ∈ 2N, 2ω62 − ωk2 = ωh2 , 2ω82 − ωk2 = ωh2 . Finally,

ω82 + ω62 − ωk2 = ωh2 has the only solutions (h, k) = (6, 8), (8, 6), (2, 10), (10, 2).

This implies that the equation for hk is diagonal for all k = 2, 6, 8, 10 and has two
(2 × 2)-blocks for the harmonics 6, 8 and 2, 10.
By (7.14), we have c(6, 6, k, k) = min(6, k) + 1, c(8, 8, k, k) = min(8, k) + 1.
With similar computations using the Clebsh-Gordan multiplication rules (see (7.8)),
the result is c(8, 6, 2, 10) = 3.
Then equation (7.16) gives, for k = 2, 6, 8, 10,
! "
(k + 1)2 − 1 − 2((min(6, k) + 1)q62 + (min(8, k) + 1)q82 ) hk = 0

and, for (h2 , h10 ) and (h6 , h8 ), the (2 × 2)-equations




4 − 3(q 2
6 + q8
2
) h2 − 3q6 q8 h10 = 0,
60 − (9q82 + 7q62 ) h10 − 3q6 q8 h2 = 0,


48 − 7(3q62 + 2q82 )h6 − 28q6 q8 h8 = 0,
80 − (27q82 + 14q62 )h8 − 28q6 q8 h6 = 0.
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 531

A direct computation using that, by (7.13), q62 = (48/7) − (32/19) and q82 = 16/19,
shows that hk = 0, ∀k . 䊐

Appendix

A.1. The measure estimate: Proof of Lemma 5.4

As in the proof of Lemma 5.3, for all N, N  ≤ Nε0 := (cγ /ε0 )1/(τ1 +2) , the result is
(N  )
,κ (u1 ) = Gγ ,κ (u2 ) = G , and thus (3.7) is trivially satisfied in such cases. In order
Gγ(N)
to prove (3.7) in the other cases, it is sufficient to prove that, ∀δ1 ∈ [0, δ0 ],

c 
(N) c   δ1 
 (N  ) γ δ1
 Gγ ,κ (u2 ) Gγ ,κ (u1 ) , δ1  ≤ C . (A.1)
2 N
#
Indeed, we can decompose [0, δ0 ] = n≥1 [δ0 2−n , δ0 2−(n−1) ]. For N  ≥ N , let us
consider δ1 fixed and consider always the complementary sets in [δ1 /2, δ1 ):

(N  ) c
c
(N  ) c
Gγ ,κ (u2 ) \ Gγ(N),κ (u1 ) = Gγ ,κ (u2 ) ∩ Gγ(N) ,κ (u1 )
#
c 
⊂ K≤N GK (u2 ) ∩ GK (u1 ) ∩ G
#
∪ K>N GK (u2 ) ∩ G .
c

As we have just seen, if K ≤ Nε1 , then GcK (u2 ) ∩ G = ∅. Hence it is enough to prove
that, if u1 − u2 s0 ≤ N −σ , then
  γ δ1
B := |GcK (u2 ) ∩ GK (u1 )| + |GcK (u2 )| ≤ C̄ . (A.2)
Nε <K≤N K>max{N,Nε }
N

Since L(K) (u) is self-adjoint, (L(K) (u))−1 0 is the inverse of the eigenvalue of smallest
modulus.
Since L(K) (u2 ) − L(K) (u1 )0 = O(εu2 − u1 s0 ) = O(εN −σ ), if one of the
eigenvalues of L(K) (u2 ) is in [−4γ K −τ , 4γ K −τ ], then by the variational charac-
terization of the eigenvalues of L(K) (u), one of the eigenvalues of L(K) (u1 ) is in
[−4γ K −τ − CεN −σ , 4γ K −τ + CεN −σ ]. As a result,
 

GcK (u2 ) ∩ GK (u1 ) ⊂ δ ∈ [δ1 /2, δ1 ]  ∃ at least one eigenvalue of L(K) (δ, u1 )

with modulus in [4γ K −τ , 4γ K −τ + CεN −σ ] .
532 BERTI and PROCESI

We now use [4, Lemma 3.2 ], which is based on a simple eigenvalue variation argument:
if ε is small enough and I is a compact interval in [−γ , γ ] of length |I |, then
 $ % &
 δ1 
 δ∈ , δ1 s.t. at least one eigenvalue of L (δ, u1 ) belongs to I 
(K)
2
K d+n+1 |I |
≤C . (A.3)
δ1m−2
As a consequence,

|GcK (u2 ) ∩ GK (u1 )| ≤ CK d+n+1 δ1−(m−2) εN −σ ≤ Cδ1 N −σ K d+n+1 .

Moreover, still by (A.3), |GcK (u2 )| ≤ CK d+n+1 γ K −τ δ1−m+2 . Hence B defined in (A.2)
satisfies
  γ   
B ≤ Cδ1 K d+n+1 N −σ + C m−2 K d+n+1−τ
K≤N
δ1 K>max{N,N } ε

≤ Cδ1 N d+n+2−σ
+ C  δ1−(m−2) γ (max{N, Nε })d+n+2−τ ≤ C̄γ δ1 N −1

for (5.7). This proves the measure estimate (A.1).

A.2. Separation properties of the singular sites

Consider the bilinear symmetric form ϕω : Rr+n+1 × Rr+n+1 → R defined by

ϕω (x, x  ) := J · J  − ω2 l l  , ∀ x = (l, J ) , x  = (l  , J  ) ∈ R × Rr+n ,

and the corresponding quadratic form Qω (x) = ϕω (x, x) := |J |2 − ω2 l 2 .

Note that for ρ = (0, ρ, 0) and for all k = (l, j1 , j2 ) ∈ Z × + × Zn , one has
Qω (k + ρ ) = Dk + |ρ|2 , where Dk are the small divisors. For notational convenience,
we denote x = k + ρ = (l, J ), where x ∈ Z × ++ × Zn since j1 ∈ + and
++ = ρ + + .
A vector x = (l, J ) ∈ Z × ++ × Zn is said to be “weakly singular" if |Qω (x)|
≤ C for some constant C fixed once and for all. In particular, if (l, j ) ∈ Z × + × Zn
is singular according to Definition 6.3, then x = (l, j + ρ) ∈ Z × ++ × Zn is
“weakly singular” with |Qω (x)| ≤ 1 + m + |ρ|2 .

Definition A.1
A sequence x0 , . . . , xK ∈ Z×++ ×Zn of distinct, weakly singular vectors satisfying,
for some B ≥ 2, |xk+1 − xk | ≤ B , ∀k = 0, . . . , K − 1, is called a B -chain of length
K.
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 533

THEOREM A.1
If ω2 satisfies (5.4), then any B -chain has length K ≤ B C /γ p for some C := C(G) >
0 and p := p(G) > 0.

Proof
We proceed with the same strategy as [4].
Given lattice vectors fi ∈ Z ×  × Zn , i = 1, . . . , n, 1 ≤ m ≤ r + n + 1,
linearly independent on R, we consider the subspace F := SpanR {f1 , . . . , fm } of
Rr+n+1 and the restriction ϕω|F of the bilinear form ϕω to F , which is represented
by the symmetric matrix Aω := {ϕω (fi , fi  )}m
i,i  =1 . Introducing the symmetric bilinear
   
forms R(x, x ) := J · J and S(x, x ) := l l , we write

ϕω = R − ω2 S and Aω = R − ω2 S,

where R := {R(fi , fi  )}m i,i  =1 = (R1 , . . . , Rm ), S := {S(fi , fi  )}i,i  =1 = (S1 , . . . , Sm )

m

are the matrices that represent, respectively, R|F and S|F in the basis {f1 , . . . , fm }.
Here Ri , Si , i = 1, . . . , m denote the column vectors, respectively, of R and S . The
main lemma is the following.

LEMMA A.2
If M = (G × Tn )/N , then the matrices R , S have coefficients in D −1 Z for some
D ∈ N.

Proof
Let us represent each fi ∈ Z ×  × Zn by the components {fi,j }r+n+1 with fi,j ∈ Z
r j =0
in such a way that fi,0 is the projection on Z, j =1 fi,j wj is the projection on , and
j =r+1 is the projection on Z . By definition, Si,i  = fi,0 fi  ,0 and hence is integer
(fi,j )r+n n

valued for all i, i  . Similarly, one has


r 
r+n
Ri,i  = fi,j fi  ,j  (wj , wj  ) + 2
fi,j ,
j,j  =1 j =r+1

and by Lemma 2.6, one has (wj , wj  ) ∈ ZD −1 . 䊐

LEMMA A.3
Assume that ω2 satisfies (5.4). Then Aω is invertible, and
c(m, D)
A−1
ω  ≤ ( max |fi |)5m−2 . (A.4)
γ i=1,...,m
534 BERTI and PROCESI

Proof
The matrix S has rank at most 1 because it represents the restriction to F of a bilinear
form of rank 1. Since any two columns of S are colinear, the development in ω2 of
detAω reduces to

det Aω = det(R1 − ω2 S1 , . . . , Rm − ω2 Sm ) = det(R1 , . . . , Rm ) − ω2


m
× det(R1 , . . . , Si , . . . , Rm ). (A.5)
i=1

Therefore D m det Aω = P (ω2 ) is a polynomial in ω2 of degree at most 1 with

integer coefficients since det R , det(R1 , . . . , Si , . . . , Rm ) ∈ D −m Z, by Lemma A.2.
Furthermore, P (·) is not identically zero because P (−1) = D m det(R + S) is positive
(the matrix R + S being positive definite).
By (A.5), if det R = 0, then | det Aω | ≥ ω2 D −m , and if det R = 0, since ω2
satisfies (5.4), | det Aω | ≥ γ D 5m/2 | det R|−3/2 . Since R + S = tF F with F =
(f1 , . . . , fm ), we have 0 ≤ det R ≤ det(R + S) = (det F )2 ≤ |f1 |2 . . . |fm |2 ≤ M 2m ,
where M := maxi=1,...,m |fi |. Hence

D 5m/2 γ
| det Aω | ≥ . (A.6)
M 3m
By (5.4), ω2 ≥ γ , and (A.6) holds in both cases. Applying the Cramer rule,

c(m, D) 2 (m−1) (A.6) c (n, D) 3m+2(m−1)

|(A−1
ω )i,i  | ≤ (M ) ≤ M
| det Aω | γ
whence (A.4) follows. 䊐

The remainder of the proof of Theorem A.1 follows [4] word by word, and we omit
it. 䊐

A.3. Proof of Lemma 6.6

It is sufficient to prove that Lα w0 ≥ w0 γ /(cMατ ), ∀w ∈ Hα . For all w ∈ Hα ⊂
HS , we have
 (6.12)
Lα w + Lαβ w = Lw = LS w − LRS L−1 S
R LR w
β=α

= HS L(N) h, where h := w − L−1 S

R LR w. (A.7)
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 535


Step 1. We have β=α Lαβ w0 ≤ ε|T |s1 C(s1 )Mαr+n+1−ν w0 , where ν := λ(s1 −
(r + n + 1)/2) > 0:
  Cε|T |s1 w0  1
Lαβ w0 ≤ ≤ Cε|T |s1 w0 .
β=α β=α
s
d(α , β ) 1 −(r+n+1)/2
β=α
(Mα + Mβ )ν

Then, for ν > r + n + 1, step 1 follows by

  +∞
1 1 y r+n dy
≤ ≤ C
β=α
(Mα + Mβ )ν k∈E +
(Mα + |k + ρ |)ν 1 (Mα + y)ν

C(ν)
≤ .
(1 + Mα )ν−(r+n+1)

Step 2. We have L(N) h0 ≥ 2−τ −1 γ Mα−τ w0 . Decompose h = h + h with
h := (K) h, h := h − (K) h, and K := 2Mα . We have
(E + )c
HS L(N) h0 ≥ (K) L(N) h0 ≥ L(K) h 0 − εTE +K h 0
K

γ
≥ h 0 − εC|T |s1 h 0 (A.8)
(2Mα )τ

by assumption (5.5) and Lemma 6.3. Moreover, since h = w − LR −1 LSR w and

w ∈ Hα ⊂ H (K) ,

h = −(1 − (K) )L−1 −1 R

R LR w = −[LR ]R∩(E + )c LR w.
S α
K

Now d(α , R ∩ (EK+ )c ) ≥ Mα , and we derive

Cε|T |s1
h 0 ≤ s1 −(r+n+1)/2
w0 . (A.9)
Mα

Furthermore, since w ∈ Hα and h − w = −(K) L−1 

R LR w ∈ HR , we have h 0 =
S

(h − w20 + w20 )1/2 ≥ w0 , and (A.8), (A.9) imply

γ w0 Cε2 |T |2s1 γ w0  C  ε2 |T |2s1 

L(N) h0 ≥ − s −(r+n+1)/2
w 0 ≥ 1 − s −τ −(r+n+1)/2
(2Mα )τ Mα1 (2Mα )τ γ Mα1
γ
≥ 2−τ −1 τ w0
Mα

because s1 > τ + (r + n + 1)/2 and provided that εγ −1 is small enough.

Acknowledgments. The authors warmly thank Claudio Procesi for many enlightening
explanations on Lie groups. We also thank Fulvio Ricci, Norman Wallach, and Michele
536 BERTI and PROCESI

Vergne for their disponibility. Finally, we thank Philippe Bolle for useful comments
on the paper.

References

[1] D. BAMBUSI, J.-M. DELORT, B. GRÉBERT, and J. SZEFTEL, Almost global existence for
Hamiltonian semilinear Klein-Gordon equations with small Cauchy data on Zoll
manifolds, Comm. Pure Appl. Math. 60 (2007), 1665 – 1690. MR 2349351
[2] M. BERTI, Nonlinear Oscillations of Hamiltonian PDEs, Progr. Nonlinear Differential
Equations Appl. 74, Birkhäuser, Boston, 2007. MR 2345400
[3] M. BERTI and P. BOLLE, Cantor families of periodic solutions of wave equations with
C k nonlinearities, NoDEA Nonlinear Differential Equations Appl. 15 (2008),
247 – 276. MR 2408354
[4] ———, Sobolev periodic solutions of nonlinear wave equations in higher spatial
dimensions, Arch. Ration. Mech. Anal. 195 (2010), 609 – 642. MR 2592290
[5] M. BERTI, P. BOLLE, and M. PROCESI, An abstract Nash-Moser Theorem with
parameters and Applications to PDEs, Ann. Inst. H. Poincaré Anal. Non Linéaire
27 (2010), 377 – 399. MR 2580515
[6] A. L. BESSE, Einstein Manifolds, reprint of the 1987 ed., Classics Math. Springer,
Berlin, 2008. MR 2371700
[7] L. BIASCO and L. DI GREGORIO, A Birkhoff-Lewis type theorem for the nonlinear wave
equation, Arch. Ration. Mech. Anal. 196 (2010), 303 – 362. MR 2601076
[8] J. BOURGAIN, Construction of periodic solutions of nonlinear wave equations in
higher dimension, Geom. Funct. Anal. 5 (1995), 629 – 639. MR 1345016
[9] ———, Quasi-periodic solutions of Hamiltonian perturbations of 2D linear
Schrödinger equations, Ann. of Math. (2) 148 (1998), 363 – 439. MR 1668547
[10] ———, Global Solutions of Nonlinear Schrödinger Equations, Amer. Math. Colloq.
Publ. Soc. 46, American Math. Soc., Providence, 1999. MR 1691575
[11] ———, Green’s Function Estimates for Lattice Schrödinger Operators and
Applications, Ann. of Math. Stud. 158, Princeton Univ. Press, Princeton, 2005.
MR 2100420
[12] T. BRÖCKER and T. TOM DIECK, Representations of Compact Lie Groups, Grad. Texts
in Math. 98 Springer, New York, 1995. MR 1410059
[13] N. BURQ, P. GÉRARD, and N. TZVETKOV, Bilinear eigenfunction estimates and the
nonlinear Schrödinger equation on surfaces, Invent. Math. 159 (2005), 187 – 223.
MR 2142336
[14] ———, Multilinear eigenfunction estimates and global existence for the three
dimensional nonlinear Schrödinger equations, Ann. Sci. École Norm. Sup. (4) 38
(2005), 255 – 301. MR 2144988
[15] L. CHIERCHIA and J. YOU, KAM tori for 1D nonlinear wave equations with periodic
boundary conditions, Comm. Math. Phys. 211 (2000), 497 – 525. MR 1754527
[16] W. CRAIG, Problèmes de petits diviseurs dans les équations aux dérivées partielles,
Panor. Synthèses 9, Soc. Math. Montrouge, France, 2000. MR 1804420
NONLINEAR WAVE EQUATIONS ON COMPACT LIE GROUPS 537

[17] W. CRAIG and C. E. WAYNE, Newton’s method and periodic solutions of nonlinear
wave equations, Comm. Pure Appl. Math. 46 (1993), 1409 – 1498. MR 1239318
[18] J.-M. DELORT and J. SZEFTEL Long-time existence for semi-linear Klein-Gordon
equations with small Cauchy data on Zoll manifolds, Amer. J. Math. 128 (2006),
1187 – 1218. MR 2262173
[19] L. H. ELIASSON and S. B. KUKSIN, KAM for the nonlinear Schrödinger equation, Ann.
of Math. (2) 172 (2010), 371 – 435. MR 2680422
[20] L. ERDO⁄ ⁄S and H.-T. YAU, Derivation of the nonlinear Schrödinger equation from a
many body Coulomb system, Adv. Theor. Math. Phys. 5 (2001), 1169 – 1205.
MR 1926667
[21] G. GENTILE and M. PROCESI, Periodic solutions for a class of nonlinear partial
differential equations in higher dimension, Comm. Math. Phys. 289 (2009),
863 – 906. MR 2511654
[22] J. FARAUT, Analysis on Lie Groups: An Introduction, Cambridge Stud. Adv. Math.
110, Cambridge Univ. Press, Cambridge, 2008. MR 2426516
[23] J. FRÖHLICH and T. SPENCER, Absence of diffusion in the Anderson tight binding
model for large disorder or low energy, Comm. Math. Phys. 88 (1983), 151 – 184.
MR 696803
[24] S. HELGASON, “Wave equations on Homogeneous Spaces” in Lie Group
Representations, III (College Park, Md., 1982/1983), Lecture Notes in Math.
1077, Springer, Berlin, 1984, 254 – 287. MR 765556
[25] ———, Geometric Analysis on Symmetric Spaces, Math. Surveys Monogr. 39, Amer.
Math. Soc. Providence, 1994. MR 1280714
[26] ———, Groups and Geometric Analysis: Integral Geometry, Invariant Differential
Operators, and Spherical Functions, corrected reprint of the 1984 original, Math.
Surveys Monogr. 83. Amer. Math. Soc., Providence, 2000. MR 1790156
[27] L. HU, The Laplacian on homogeneous spaces, J. Math. Phys. 49 (2008), no. 053513.
MR 2421923
[28] S. B. KUKSIN, Hamiltonian perturbations of infinite-dimensional linear systems with
imaginary spectrum (in Russian), Funktsional Anal. i Prilozhen. 21 (1987), no. 3,
22 – 37. MR 911772
[29] ———, Analysis of Hamiltonian PDEs, Oxford Lecture Ser. Math. Appl. 19, Oxford
Univ. Press, Oxford, 2000. MR 1857574
[30] S. B. KUKSIN and J. PÖSCHEL, Invariant Cantor manifolds of quasi-periodic
oscillations for a nonlinear Schrödinger equation, Ann. of Math. (2) 143 (1996),
149 – 179. MR 1370761
[31] G. MENG, Geometric construction of the quantum Hall effect in all even dimensions,
J. Phys. A 36 (2003), 9415 – 9423. MR 2006735
[32] J. NEGRO, M. A. DEL OLMO, and A. RODRÍGUEZ-MARCO, Landau quantum systems:
An approach based on symmetry, J. Phys. A 35 (2002), 2283 – 2307.
MR 1908725
[33] J. PÖSCHEL, A KAM-theorem for some nonlinear partial differential equations, Ann.
Scuola Norm. Sup. Pisa Cl. Sci. (4) 23 (1996), 119 – 148. MR 1401420
[34] C. PROCESI, Lie Groups: An Approach through Invariants and Representations,
Universitext, Springer, New York, 2007. MR 2265844
538 BERTI and PROCESI

[35] G. RACAH, Group Theory and Spectroscopy, Ergeb. Exakt. Naturwiss. 37, Springer,
Berlin, 1965. MR 0213113
[36] M. R. SEPANSKY, Compact Lie Groups, Grad. Texts in Math. 235, Springer, New York,
2007. MR 2279709
[37] Y. TANIGUCHI, Normal homogeneous manifolds and their spectra, Osaka J. Math. 18
(1981), 555 – 576. MR 635721
[38] M. E. TAYLOR, Partial Differential Equations, III: Nonlinear Equations, corrected
reprint of the 1996 original, Appl. Math. Sci. 117, Springer, New York, 1997.
MR 1477408
[39] S. C. ZHANG and J. P. HU, A Four-Dimensional Generalization of the Quantum Hall
Effect, Science, 294, 823 (2001).
[40] C. E. WAYNE, Periodic and quasi-periodic solutions of nonlinear wave equations via
KAM theory, Comm. Math. Phys. 127 (1990), 479 – 528. MR 1040892
[41] H. WEYL, The Theory of Groups and Quantum Mechanics, Dover, New York, 1950.

Berti
Dipartimento di Matematica e Applicazioni “R. Caccioppoli,” Università Federico II, I-80126,
Napoli, Italy; [email protected]

Procesi
Dipartimento di Matematica e Applicazioni “R. Caccioppoli,” Università Federico II, I-80126,
Napoli, Italy; [email protected]