Professional Documents
Culture Documents
Introduction To Involutive Structures
Introduction To Involutive Structures
Editorial Board
Béla Bollobás
William Fulton
Frances Kirwan
Peter Sarnak
Barry Simon
Burt Totaro
SHIFERAW BERHANU
Temple University
PAULO D. CORDARO
University of São Paulo
JORGE HOUNIE
Federal University of São Carlos
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo
Cambridge University Press has no responsibility for the persistence or accuracy of urls
for external or third-party internet websites referred to in this publication, and does not
guarantee that any content on such websites is, or will remain, accurate or appropriate.
Contents
Preface page ix
I Locally integrable structures 1
I.1 Complex vector fields 1
I.2 The algebraic structure of X 4
I.3 Formally integrable structures 5
I.4 Differential forms 7
I.5 The Frobenius theorem 11
I.6 Analytic structures 14
I.7 The characteristic set 15
I.8 Some special structures 15
I.9 Locally integrable structures 19
I.10 Local generators 21
I.11 Local generators in analytic structures 25
I.12 Integrability of complex and elliptic structures 26
I.13 Elliptic structures in the real plane 28
I.14 Compatible submanifolds 32
I.15 Locally integrable CR structures 36
I.16 A CR structure that is not locally integrable 38
I.17 The Levi form on a formally integrable structure 42
Appendix: Proof of the Newlander–Nirenberg theorem 47
Notes 51
II The Baouendi–Treves approximation formula 52
II.1 The approximation theorem 52
II.2 Distribution solutions 63
II.3 Convergence in standard functional spaces 69
II.4 Applications 83
Notes 99
v
vi Contents
Epilogue 361
1 The similarity principle and applications 361
2 Mizohata structures 364
3 Hypoanalytic structures 370
4 The local model for a hypoanalytic manifold 371
5 The sheaf of hyperfunction solutions on a hypoanalytic
manifold 372
Appendix A Hardy space lemmas 374
A.1 Multipliers in h1 374
A.2 Commutators 376
A.3 Change of variables 378
Bibliography 381
Index 390
Preface
ix
x Preface
planar vector fields, the solvability condition of Nirenberg and Treves is
discussed and a priori estimates are proved in Lp and in a mixed norm that
involves the Hardy space h1 R. A duality argument is then used to derive
local solvability results in Lp 1 < p < and in L R bmoR . The chapter
also includes sections on the sufficiency and necessity of condition for
local solvability in higher dimensions. The first part of Chapter V introduces
certain submanifolds in an involutive structure which are important
in the study of solutions. These submanifolds are generalizations of the totally
real and generic CR submanifolds encountered in CR manifolds. The second
part of the chapter introduces the FBI transform first in Rn and then in a locally
integrable structure. The FBI transform is then applied to derive edge-of-the-
wedge type results. It is also applied to study the microlocal singularities
of the solutions of a first-order nonlinear PDE and a generalization of the
F. and M. Riesz theorem. Chapter VI studies some boundary properties of the
solutions of locally integrable vector fields. These properties include the exis-
tence of a trace at the boundary, pointwise convergence of solutions to their
boundary values, and the validity of Hardy space-like properties. Chapter VII
describes the differential complex attached to a general involutive structure.
An invariant definition of this complex is followed by a useful representa-
tion in appropriate coordinates. An approximate Poincaré Lemma for locally
integrable structures is also proved in the chapter. Chapter VIII deals with
the local solvability theory of the undetermined systems of partial differen-
tial equations naturally associated with a locally integrable structure, that is,
the cohomology theory of its differential complex. Necessary and sufficient
conditions are studied in some detail when the structure is analytic, or elliptic,
or has corank one. Concerning the latter class, a thorough exposition of the
geometric characterization of local solvability in degree one for real analytic
structures is presented.
Finally we conclude with an epilogue which summarizes some of the
results obtained in recent years on diverse areas such as the similarity prin-
ciple, Mizohata structures, and hyperfunction solutions in hypoanalytic mani-
folds. Two applications of the similarity principle are described. The first
application concerns uniqueness in the Cauchy problem for a class of semi-
linear equations. The second application involves the theory of bending of
surfaces.
There are numerous interesting results on complex vector fields and invo-
lutive structures that have been obtained since the publication of [BT1] and
which are not covered in this book. The authors have selected the material
with which they have had first-hand experience. In the notes at the end of each
chapter, we indicate some related works and provide additional references.
xii Preface
In this chapter we introduce the main concepts which will be studied through-
out the book. In order to do so we recall some standard notions such as differ-
entiable manifolds, vector fields, differential forms, etc., with the purpose
mainly of laying down the basis for the presentation and to establish the
notations.
Nevertheless, we assume from the reader some familiarity with these
concepts. In particular, we freely use some standard results on complex vector
fields and complex differential forms on RN .
1
2 Locally integrable structures
It is easy to see that given any family ∗ = U x as above satisfying (1)
and (2) there is a unique differentiable structure over , of dimension N ,
such that ∗ ⊂ .
1
More generally, we say that a function f → C is C k (k ≥ 0) if for every U x ∈ the
composition f x−1 is C k on xU.
I.1 Complex vector fields 3
We shall denote by X the set of all complex vector fields over .
Proof. For the first statement it suffices to show that L1 = 0 and this follows
from (I.1) together with the fact that 12 = 1. We shall now prove (I.2); we
must show that if f vanishes on an open set V ⊂ then the same is true
for Lf .
Let p ∈ V be arbitrary. We select a local chart U x with p ∈ U ⊂ V
and take ∈ Cc xU such that xp = 1. Then the function g → R
defined by the rule
xq if q ∈ U
gq =
0 if q ∈ U
belongs to C R and vanishes on \V . In particular,
f = 1 − gf
and then
Lfp = 1 − gpLfp + fpL1 − gp = 0
since gp = 1.
A consequence of the preceding result is the possibility of defining the
restriction of an element L ∈ X to an open subset W of . More precisely,
there is a C-linear map
X L −→ LW ∈ XW
which turns the diagram
L
C −→ C
↓ ↓
LW
C W −→ C W
commutative (the vertical arrows denote the restriction map). Indeed, if p ∈ W
and f ∈ C W we set
LW fp = Lf̃ p
4 Locally integrable structures
2
Recall that a Lie algebra over C is a C-vector space E over which is defined a bilinear form
E × E v w → v w which satisfies
u u = 0 u v w + v w u + w u v = 0 u v w ∈ E
I.3 Formally integrable structures 5
The set of all complex tangent vectors at p, denoted by CTp , has a structure
of a C-vector space and is called the complex tangent space to at p.
If L ∈ X then Lp C p → C defined by
Lp f = Lfp f ∈ C p
belongs to CTp . Conversely, suppose that for each p ∈ an element
vp ∈ CTp is given such that
p → vp f ∈ C ∀f ∈ C
Then there is L ∈ X such that Lp = vp for all p ∈ .
Suppose now that p ∈ U and that U x is a local chart. If v ∈ CTp then,
according to (I.5),
N N
vf = gj pvxj = vxj f f ∈ C p
j=1 j=1 x j p
We shall also need the notion of a complex vector sub-bundle of CT of rank
n and corank N − n. By this we mean a disjoint union
= p ⊂ CT
p∈
The rank (resp. corank) of will be referred to as the rank (resp. corank) of
the formally integrable structure . Let be a formally integrable structure
over and fix p ∈ . There is a local chart U x with p ∈ U and vector
fields L1 Ln ∈ XU such that L1q Lnq is a basis of q for every
q ∈ U . If we write x = x1 xN and
N
Lj = ajk x
k=1 xk
then the matrix ajk has rank equal to n at every point; moreover, there are
cjk ∈ C U, j k = 1 n, such that
n
L j Lk =
cjk L j k = 1 n
=1
More generally, we can consider the concept of (weak) solutions for the
formally integrable structure over : it suffices to consider u, in the
preceding definition, belonging to the space of distributions on (we refer
to [H2] for the theory of distributions on manifolds).
X → C
8 Locally integrable structures
L = 1 − gL
If we then define
CTp∗ = dual of CTp
p v = Lp
Proof. Let U x be a local chart with p ∈ U . Formula (I.6) allows one to
define dxj ∈ NU, j = 1 N , by the rule
I.4 Differential forms 9
dxj = jk j k = 1 N
xk
Hence, if ∈ NU we have
N
= dxj (I.9)
j=1 xj
= p
p∈
N
Lj = ajk j = 1 n
k=1 xk
n
L#j = bj L j = 1 n
=1
m
L#j = + cjk j = 1 n
xj k=1 xn+k
n
= dxn+ − c dx = 1 m
=1
Then 1p mp are linearly independent for all p ∈ U0 and furthermore
Tp∗ = ∈ CTp∗ is real
I.5 The Frobenius theorem 11
T = Tp
p∈
T ∗ = Tp∗
p∈
Notice that when j ≥ 2 the vector field L#j does not involve differentiation in
the x1 -variable. Thus, in a neighborhood of the origin, we have
n
L#j L#k = #
Cjk L j k = 2 n
=2
/y2 /yn
This argument has the following consequence: returning to the original coor-
dinates x1 xN , the induction hypothesis allows us to assume from the
beginning that
Now, the coefficient of /x in the commutator L#1 L#j is equal to aj /x1 .
On the other hand,
n
n N
L#1 L#j = #
C1j L = C1j
1
+ C1j
ak
=1 x1 =2 k=2 x k
and thus
aj n
=
C1j a = 2 N j = 2 n
x1 =2
Hence for each fixed the vector a2 an satisfies a linear system of
ordinary differential equations with trivial initial condition. By the uniqueness
theorem for such systems we conclude that aj = 0 if j = 2 n and > n.
Thus we have
n
L#j = ajk j = 2 n
k=2 x k
Remark I.6.1. Suppose that is now an open subset of RN and let L ∈ X
be real-analytic. Write
N
L = aj x
j=1 x j
Let also u ∈ and take an open set C ⊂ CN , where the holomorphic
coordinates are written as z1 zN , such that
• C ∩ RN = ;
• u, aj extend as holomorphic functions ũ, ãj on C .
Then
Lu = L̃ũ (I.12)
where L̃ is the holomorphic vector field
N
L̃ = ãj z
j=1 zj
1.8 Some special structures 15
and consequently n of the tangent vectors Lj p Lj p are linearly
0 0
independent. Since this remains true in a neighborhood of p0 the result is
proved.
1 i
z= z + z − iz − iz ∈ V 0 + iV 0
2 2
Tp0 ⊗R C Tp ∩ T p ∀p ∈ (I.15)
we obtain:
I.8 Some special structures 17
Proof. Notice that (I.16) is trivial since +1 m is also a basis for
V0 ⊕ iV0 .
Next we notice that V ∩ V 1 = 0. Indeed, let z ∈ V ∩ V 1 . Then z ∈ V1 ⊂ V
and consequently z z ∈ V0 , which gives z ∈ V0 ⊕ iV0 ∩ V1 = 0. Hence
1 1 +1 m
V = Tp V0 = Tp0
18 Locally integrable structures
• The formally integrable structure is locally integrable if and only if, given
p0 ∈ and vector fields L1 Ln which span in an open neighborhood
U0 of p0 , there are an open neighborhood V0 ⊂ U0 of p0 and smooth
functions Z1 Zm ∈ C V0 such that:
dZ1 ∧ ∧ dZm = 0 in V0
Lj Zk = 0 j = 1 n k = 1 m
where
C ∈ B, then we can find real-analytic functions Z1 Zm
defined in a neighborhood of the origin and satisfying
Lj Z = 0 j = 1 n = 1 m
dZ1 ∧ ∧ dZm = 0
We write
N
Lj = ajk
k=1 xk
and take an open, connected set U ⊂ CN such that U ∩ RN = B and such that
there are ãjk , C̃ ∈ U satisfying
dW1 ∧ ∧ dWm = 0
It suffices then to set Zk = Wk V ∩B in order to obtain the desired solutions
(cf. (I.12)).
m
cjk dGk p = j j = + 1 m
k=1
We then set
m
Zj = cjk Gk − Gk p j = 1
k=1
m
W = c+k Gk − Gk p = 1 d
k=1
are linearly independent at p. We are now ready to state and prove the
following important result:
22 Locally integrable structures
Proof. The proof follows almost immediately from the preceding discussion:
it suffices to take smooth, real-valued functions t1 tn defined near p and
vanishing at p such that
dx1 dx dy1 dy ds1 dsd dt1 dtn
are linearly independent. Notice that dWk p = +k is real, from which we
derive that dk = 0 at the origin.
Since we have
Wk
0 0 0 = kk k k = 1 d
sk
we can introduce, in a neighborhood of the origin in R2+d+n , the vector
fields
d
Mk = kk z s t k = 1 d (I.25)
k =1
s k
d
k
L̃ = −i z s tMk = 1 n (I.28)
t k=1 t
If we write Zx t = Z1 x t Zm x t then Zx 0 0 equals the identity
m × m matrix. Hence we can introduce, in a neighborhood of the origin in
RN , the vector fields
m
Mk = k x t k = 1 m (I.35)
=1 x
Mk Z = k (I.36)
m
k
Lj = −i x tMk j = 1 n (I.37)
tj k=1 tj
Zk = xk + ik t k = 1 m
m
k
Lj = −i t j = 1 n (I.37 )
tj k=1 tj xk
Remark I.11.2. We point out that in the coordinates x t given by Corollary
I.11.1 it is elementary to find the unique analytic solution u to the Cauchy
problem:
Lj u = 0 j = 1 n
(I.41)
ux 0 = hx
where h is real-analytic. Indeed,
ux t = hZ1# x t Zm# x t
solves (I.41) and in order to see that this is the unique analytic solution it
suffices to notice that if v is analytic, if vx 0 = 0, and if Lj v = 0 for every
j then v must vanish identically since all its derivatives vanish at the origin.
Uniqueness for the distribution solutions of (I.41) holds when the structure
is only C . This, though, is a much deeper result and its discussion will be
postponed to Chapter II.
Proposition I.13.1. If u ∈ U and Lu ∈ L2s loc U then u ∈ Lloc U. In
2s+1
We shall now derive from (I.45) an estimate for the transpose of L which
will lead us to a solvability result. If we notice that t L = −L − ibx x t then
from (I.45) we obtain, for some constant C > 0,
Still under the assumption that L is elliptic we apply (I.50) in order to find
v ∈ C D such that
Lv = −ibx (I.51)
If we set
x
ux t = evx t dx
0
we get
x
Lux t = vt x tevx t dx + ibx tevxt
0
x
= −ibvx − ibx x tevx t dx + ibx tevxt
0
x
= −i x bev x tdx + ibx tevxt
0
= ib0 tev0t
Then if we set
t
Zx t = ux t − i b0 t ev0t dt (I.52)
0
we obtain
LZ = 0 Zx = ev = 0 (I.53)
that is, our original elliptic structure is locally integrable. We have thus
obtained a proof of the Newlander–Nirenberg theorem in the particular case
when N = 2. We emphasize for this situation the conclusion that we have
reached at the end of Section I.12:
32 Locally integrable structures
Remark I.13.5. Our discussion indeed leads to a general criterion that char-
acterizes when a rank one formally integrable structure ⊂ CT, ⊂ R2
open, is locally integrable. Suppose that is spanned, in a neighborhood of
the origin, by the vector field (I.42).
Proof. We have already presented the argument that ‡ ⇒ †. For the
reverse implication we notice that
0 = LZx = LZx + ibx Zx
and consequently
L logZx = Zx−1 LZx = −ibx
make up a family ∗ that satisfies properties (1) and (2) of Section I.1 Hence
is a smooth manifold of dimension r. We shall refer to the number N − r
as the codimension of (in ).
I.14 Compatible submanifolds 33
Let p ∈ and denote by C p the space of germs of smooth functions on
at p. It is clear that the restriction to defines a surjective homomorphism
o f C-algebras C p → C
p which gives us then a natural injection
Definition I.14.1. We shall say that is compatible with the formally inte-
grable structure if defines a formally integrable structure over .
Proof. We must prove that (I.60) implies that is a vector sub-bundle
of which satisfies the Frobenius condition.
First we observe that (I.60) and (I.59) give the existence of such that
f
f = dzj (I.63)
j=1 z j
Since
&k 1 k
= + ik k = 1 d
w 2 s
we conclude, taking into account (I.64), that FU is generic if U is taken
small enough so that
z s ≤ 1 z s ∈ U (I.68)
s 2
rank ! t = n ∀t ∈ U
where ! = 1 m . This implies that = !U is an embedded
submanifold of Rm of dimension n and it is clear that can be realized
as the CR structure induced by the complex structure on the generic subman-
ifold Rm + i of Rm + iRm = Cm .
Since L and L are linearly independent at every point it follows that defines
a CR structure which is furthermore locally integrable, since the differential
of the functions z and W = s + i z 2 span T on C × R. Notice also that the
Hans Lewy structure can be globally realized as the CR structure induced on
the hyperquadric
Q = z w ∈ C2 w = s + it t = z 2 (I.70)
by the complex structure on C2 .
More generally, given %j ∈ −1 1 , j = 1 n, we can consider the CR
structure on Cn × R spanned by the pairwise commuting vector fields
Lj = − i%j zj j = 1 n (I.71)
zj s
Such a structure is also locally integrable for the differential of the functions
z1 zn and W = s + iz, with
n
z = %j z j 2
j=1
span T on Cn × R.
Before we embark on the proof we shall state and prove the important conse-
quence of this result:
I.16 A CR structure that is not locally integrable 39
Corollary I.16.2. The vector fields (I.73) span a CR structure which is not
locally integrable in any neighborhood of the origin.
Indeed, first we notice that L#1 L#n L#1 L#n are linearly independent
over which together with property (a) shows that (I.73) define a CR structure
over .
Now, given any smooth solution h to the system
L#j h = 0 j = 1 n (I.74)
Proof of Proposition I.16.1. The first step in the proof is the construction
of the function g. In the complex plane we denote the variable by w = s + it
and consider a sequence of closed, disjoint disks Dj , all of them contained
in the sector w s < t and such that Dj → 0 as j → .
Let F ∈ C C R have support contained in the union of the disks Dj and
satisfy
Fw > 0 ∀w ∈ int Dj ∀j (I.75)
z = z1 2 − z2 2 − − zn 2 (I.76)
Then
FWz s ≤ C zHz
We then set
FWz s
gz s = (I.77)
z1 − FWz s
40 Locally integrable structures
Since
FWz s 1
gz s =
z1 1 − FWz s/z1
it follows from Lemma I.16.3 that g is smooth in an open neighborhood of
the origin in Cn × R and that g vanishes to infinite order at z1 = 0.
We shall now proceed to the proof of (a). We shall write
L#j = Lj − i%j zj gz s
s
(cf. (I.76), (I.73)). Since Lj Lk = 0 and Lj zk = 0 for all j and k we obtain
L#j L#k = −i %k zk Lj g − %j zj Lk g (I.78)
s
Now
z1
Lj g = L F W
z1 − F W2 j
and an easy computation making use of the chain rule gives
F
Lj FWz s = −2i%j zj Wz s
w
Hence from (I.78) we obtain
−iz1
L#j L#k = % z L F W − %j zj Lk F W
z1 − F W2 k k j s
−2z1
= % k zk % j zj − % j zj % k z k = 0
z1 − F W 2 s
We now start to prove (b). For this we set
(z s = hz 0 0 s
and will show that (/s0 0 = 0. We assume that ( is C 1 in a set of the
form
V = z s ∈ C × R z < r s <
and observe that
(
L( − izfz s = 0 (I.79)
s
where L is the Hans Lewy operator given in (I.79) and
Fs + i z 2
fz s =
z − Fs + i z 2
is smooth in V (contracting V if necessary).
I.16 A CR structure that is not locally integrable 41
Let U = w = s + it ∈ C s < 0 < t < r 2 and assume that Dj ⊂ U for
all j. Define
Iw = √ (z sdz w ∈ U (I.80)
z= t
Consequently,
I i 1
w = √ L(z sdz (I.81)
w 2 z= t z
(cf. (I.79)). From (I.79), (I.81) and from the fact that F is supported in the
union of the disks Dj we conclude that I is a holomorphic function of w in the
connected open set U \ ∪j Dj . Since, moreover, Iw → 0 when t → 0+ the
Schwarz reflection principle implies that I vanishes identically in U \ ∪j Dj .
In particular,
I ≡0 on Dj ∀j. (I.82)
du = L(dz ∧ dz ∧ dW
(
= izfz s dz ∧ dz ∧ ds
s
(
= −2zfz s dx ∧ dy ∧ ds
s
which in conjunction with (I.85) gives
(
zfz s dxdyds = 0 ∀j (I.87)
Sj s
Now we observe that zfz s = Fs + i z 2 *z s, where * is smooth and
satisfies *0 0 = 1. From (I.87) we conclude the existence of points Pj Qj ∈
Sj such that
( (
*Pj Pj = *Qj Qj = 0
s s
n
M= gj Lj
j=1
I.17 The Levi form on a formally integrable structure 43
x1 xn y1 yn s1 sd
44 Locally integrable structures
with ajj 0 0 = bjk 0 0 = 0 for all j j k, which span in a neighborhood
of the origin in R2n+d . Notice, moreover, that Tp0 is equal to the span of
ds1 0 dsd 0 .
Write = 1 ds1 0 + + d dsd 0 and denote by Ajj the matrix of the Levi
form Lp with respectto the basis /z1 p /zn p of p . Thus, by
definition, Ajj = Lp /zj p /zj p and then
1
Ajj = ds + + d dsd 0 Lj Lj p
2i 1 1 0
1 d
= L b − Lj bjk 0 0
2i k=1 k j j k
that is
1 d
bj k bjk
Ajj = 0 0 − 0 0 (I.89)
2i k=1 k zj zj
As an example, let us consider the CR structure defined by the vector fields
L#j given by (I.73). In this case d = 1 and we take = ds 0 . We also have
bj = −i%j zj 1 + gz s, where g vanishes to infinite order at z1 = 0. Then
bj
0 0 = −i%j jj
zj
and (I.89) gives
Ajj = diag %1 %n
Thus, Corollary I.16.2 has provided an example of a nondegenerate CR
structure, defined in a neighborhood of the origin in Cn × R, for which the
signature of the Levi form at ds 0 ∈ T00 , = 0 is equal to n − 1.
In connection with this example we mention the following deep result
which gives a positive answer to the problem of local integrability (or local
realizability, as we have seen in Proposition I.15.1) for certain classes of
CR structures. It shows that the value of the signature of the Levi form plays
a crucial role.
Recall that by Proposition I.8.4 the characteristic set of a CR structure is a
sub-bundle of the cotangent bundle.
Finally, we shall compute the expression of the matrix Ajj of the Levi form
when is locally integrable and CR. Invoking the local coordinates described
at the beginning of Section I.15, and in particular the functions (I.66) satisfying
(I.64), we see that we can take the vector fields Lj in the form (cf. (I.27))
d
k
Lj = −i z sMk j = 1 n
zj k=1 zj
where
d
Mk = kk z s k = 1 d
k =1
sk
characterized by the relations Mk sk + ik = kk . In particular, (I.64) gives
kk 0 0 = kk (I.90)
According to our previous notation, we have ajj ≡ 0 for all j j and
d
k
bjk = −i
k =1
zj k k
Again by (I.64) and by (I.90) we have
bjk 2 k
0 0 = −i 0 0
zj zj zj
and then by (I.89) we obtain
d
2 k
Ajj = k 0 0 (I.91)
k=1 zj zj
Since then T0 = n+1 = 0 , the Levi condition at the origin can be written
as:
n
2
0 0j k ≥ 0 ∀ ∈ Cn (I.93)
jk=1 z j z k
Zj = zj j = 1 n Wz s = s + iz s
span T U near the origin [we are writing s = w and considering z s
as local coordinates in U ]. From (I.91) we obtain the following equiva-
lent statement to (I.93): the Levi form of the CR structure 01 U at the
characteristic point ds 0 is positive.
10
To obtain an invariant
01 ⊥ statement let us first denote by T the orthog-
onal sub-bundle . Given an open set U with a smooth boundary U
as above, and given p ∈ U , the map '∗p CTp∗ Cn+1 → CTp∗ U induces an
isomorphism
∼
p Tp10 −→ Tp U
u" = u1 uq ∈ C M Rq
The two main results that are essential for Malgrange’s argument are:
Then there are %0 > 0, C > 0 and 0 < < 1 such that for every 0 < % ≤ %0
there is a smooth solution u" % to (I.94) on x − x0 < % satisfying the bounds
u" % x − u" 0 x ≤ C%M− + x − x0 < % ≤ M
m
Lj = + ajk z j = 1 m (I.97)
zj k=1 zk
where ajk = 0 at the origin. For technical reasons, which are going to be
clear in the argument, it is convenient to assume that ajk z = O z 2 , and
this property can be achieved after performing a local diffeomorphism of the
form z = z + Qz z, where Q is a homogeneous polynomial of degree two in
z1 zm z1 zm chosen suitably. We leave the details of this (simple)
computation to the reader.
Malgrange’s key idea is to show the existence of a local diffeomorphism
w = Hz, defined near the origin in Cm , such that, in the new variables
w1 wm , the structure has a set of generators which have real-analytic
coefficients. This implies the sought-for conclusion thanks to Theorem I.9.2.
In order to shorten the notation and make the computations more apparent,
we shall describe all the systems involved in vector and matrix notation. Thus
we set
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
L1
⎢ ⎥ ⎢ ⎥
z1
⎢ z1 ⎥
"
L = ⎣ ⎦ = ⎢ ⎥ =⎢ ⎥
z ⎣ ⎦ z ⎣ ⎦
Lm z z
m m
Appendix: Proof of the Newlander–Nirenberg theorem 49
which is clearly elliptic (in the usual sense). We conclude that there are %0 > 0,
C > 0 and < 1 such that for every 0 < % ≤ %0 there is a smooth solution H%
to (I.104) satisfying
H% − H0 C 2 z z ≤% ≤ C%2+ % ≤ %0 (I.105)
In particular, if % > 0 is small enough we can ensure that:
Notes 51
Notes
The first treatment of formally and locally integrable structures as presented
here appeared in [T4], the main point for this being the discovery of the
Approximation Formula by M. S. Baouendi and F. Treves in 1981 ([BT1]);
such structures were then studied extensively in [T5]. The pioneering work
though seems to be the article by Andreotti-Hill ([AH1]), where the concept
of what we now call a real-analytic locally integrable structure was introduced
in its full generality.
This introductory chapter contains mainly results that have already been
presented in standard textbooks. We mention, for instance, the Frobenius
theorem, whose proof was taken from L. Hörmander’s book [H4] and the
integrability of elliptic vector fields in the plane, of which we give an almost
self-contained proof, depending only on very simple facts concerning commu-
tators of certain pseudo-differential operators that can be found, for instance,
in [Fo].
As mentioned in the text, Theorem I.12.1 is due to L. Nirenberg ([N2])
and the proof we present was taken from [T5].
Proposition I.16.1 is a particular case of a more general result due to H.
Jacobowitz and F. Treves ([JT1]). We also refer to [JT2] where the same
authors study, via a category argument, the set of all formally integrable
CR structures of rank n on an open subset of R2n+1 whose Levi form has, at
each nonzero characteristic point, signature n − 1.
Theorem I.17.3 was originally due to M. Kuranishi ([Ku1], [Ku2]) in the
case n ≥ 4. Later, T. Akahori ([Ak]) presented an improvement to Kuranishi’s
argument which allowed him to prove Theorem I.17.3 also for the case n = 3.
The case n = 2 is still an open problem, whereas when n = 1 the conclusion
is false, according to [N3] (see also Theorem I.12.1). A proof of Theorem
I.17.3 can also be found in [W3].
Finally, Malgrange’s proof of the Newlander–Nirenberg theorem that we
presented in the appendix was taken from [N1], where the use of a solvability
result on elliptic determined systems of nonlinear partial differential equations
makes the argument a bit simpler.
II
The Baouendi–Treves approximation formula
In this chapter we prove what is probably the most important single result in
the theory of locally integrable structures. It states that in a small neighborhood
of a given point of the domain of a locally integrable structure , any solution
of the equation u = 0 may be approximated by polynomials in a set of a
finite number of homogeneous solutions as soon as the solutions in that set
are chosen with linearly independent differentials and the number of them is
equal to the corank of . Such a set is called a complete set of first integrals
of the locally integrable structure.
The proof is relatively simple for classical solutions and depends on the
construction of a suitable approximation of the identity modeled on the kernel
of the heat equation as shown in Section II.1. The extension to distribution
solutions is carried out in Section II.2. Section II.3 studies the convergence of
the formula in some of the standard spaces used in analysis: Lebesgue spaces
Lp , 1 ≤ p < ; Sobolev spaces; Hölder spaces; and (localizable) Hardy spaces
hp , 0 < p < . The last section is devoted to applications.
52
II.1 The approximation theorem 53
x1 xm t1 tn
k 0 0 = 0 dx k 0 0 = 0 k = 1 m
where the double bar indicates the norm of the matrix x x t = j x t/xk
as a linear operator in Rm . Modifying the functions k ’s off a neighborhood
of V may assume without loss of generality that the functions k x t,
k = 1 m, are defined throughout RN , have compact support and satisfy
(II.3) everywhere, that is
j x t 1
< x t ∈ RN (II.3 )
xk 2
degree k, Pk , of its Taylor series on a fixed polydisk that contains the set
√
Zx t − Zx 0 x x < R t < R , so replacing the exponential
in the definition of E by Pk Zx t − Zx 0 we will find polynomials
in Zx t that approximate E ux t in the C topology for x < R/4 and
t < T when k is large. Hence, from now on we fix our attention on the
II.1 The approximation theorem 57
Proof. We may write Ax 2 = t AAx · x (the dot indicates the standard inner
product in Rm and also its extension as a C-bilinear form to Cm ) so e− Ax =
2
where the branch of the square root is chosen so det C1/2 > 0 when C is
real. Since det C = det A2 the proof is complete.
Set hxux t det Zx x t = vx t. For x t fixed, the matrix Zx x t =
I + ix x t satisfies the hypotheses of the lemma in view of (II.3 ). Thus,
we may write
2
hxux t = ) −m/2 e− Zx xtx vx t dx
Rm
Rm
Then
G ux t − hxux t = I + J
58 The Baouendi–Treves approximation formula
where
2
I x t = ) −m/2 e− Zx xtx vx + −1/2 x t − vx t dx
Rm
and
J x t = ) −m/2
−1/2 x t 2 2
e− Zxt−Zx+ − e− Zx xtx vx + −1/2 x t dx
Rm
2 2 2 2
To estimate I we observe that e− Zx xtx = e− x + x xtx ≤ e−3 x /4 in
view of (II.3 ). We also observe that ,x vx t is bounded in Rm × t ≤ R
because v vanishes for large x, so the mean value theorem gives
2
I x t ≤ C −1/2 e−3 x /4 x dx ≤ C −1/2
Rm
−1/2 2
e− Zxt−Zx+ x t − e− Zx xtx dx
2
J x t ≤ C
x <K
+ C exp−K 2 /2
= vx t dZx t
II.1 The approximation theorem 59
We have
x t − x t ≤ x t − x t + x t − x t
1
≤ x − x + C t − t
2
1
≤ x − x + CT
2
because t ∈ 0 t and t ≤ T . Hence,
1
x t − x t 2 ≤ x − x 2 + 2T 2
2
and
& − Zxt−Zx t 2 &
&e & = e2T 2 − x−x 2 /2
where is a bound that depends only on and does not depend on u. Since
Lj h vanishes for x ≤ R/2 we have that x ≥ R/2 in all integrands in the
expression of R , so x − x ≥ R/4 and
2 −R2 /32
R ux t ≤ Ce2T
60 The Baouendi–Treves approximation formula
we have
jk = Mk x t Dx Zj x t − Zj x t
= −Mk x t Dx Zj x t − Zj x t
Applying this to F = e− we get, after differentiation under the integral
2
sign that
Mk G ux t = −/)m/2
Mk x t Dx e− Zxt−Zx t ux thx dZx t
2
Rm
where we have used the fact that the pullback to any slice t = const. of the
m-form dZ1 ∧ · · · ∧ dZn is given by det Zx x t dx . Next, using the ‘integra-
tion by parts’ formula
Mk v w dZ = − v Mk w dZ (II.7)
Rm Rm
II.1 The approximation theorem 61
which is valid if v and w are of class C 1 and one of them has compact support,
we get
Mk G ux t = /)m/2
2
e− Zxt−Zx t Mk ux thx + ux tMk hx dZx t
Rm
which proves (II.6). To complete the proof we show that (II.7) holds. Consider
the exact m-form defined by
' k ∧ · · · ∧ dZm
k = duv dZ1 ∧ · · · ∧ dZ
' k ∧ · · · ∧ dZm
= duv ∧ dZ1 ∧ · · · ∧ dZ
where the hat indicates that the factor dZk has been omitted. The pullback of
k to the slice t × Rm is exact, so
k = 0 (II.8)
t ×Rm
Using (II.4) to compute duv and observing that the pullback to the slice of
terms that contain a factor dtj vanish, we get
m
= + jk j = 1 n
tj k=1 xk
Lemma II.1.5.
det Zx m
jk det Zx
+ ≡ 0 j = 1 n (II.9)
tj k=1 xk
Proof. Note that (II.9) says that the vector field det Zx Lj is divergence free,
i.e., div det Zx Lj = 0, or that t Lj det Zx = 0 where t Lj is the transpose
62 The Baouendi–Treves approximation formula
observe that
e uh det Zx e uh det Zx
= det Zx + e uh
tj tj tj
m
= det Zx Lj e uh − det Zx jk e uh
k=1 xk
det Zx
+ e uh
tj
Note that the integral over Rm of the second term of the right-hand side may
be written, after integration by parts, as
m
e uh jk det Zx dx
k=1 x k
II.2 Distribution solutions 63
in view of (II.9). Since Lj e = 0, we also have that det Zx Lj e uh =
det Zx e Lj uh + det Zx e uLj h. This shows that
G̃ u t = G̃ Lj u t + e x tuLj h det Zx x t dx
tj
When = Zx t we obtain
/)m/2 e− Zxt−Zx t hx det Zx x t). This takes care of (a) and (b). To
2
where
rj x t t = /)m/2 e− Zxt−Zx t ux t Lj hx det Zx x t dx
2
Rm
(II.12)
and 0 t denotes the straight segment joining 0 to t. In other words, by
integrating first in x we may express the integral of an m + 1-form over the
cell Rm × 0 t as the integral of a 1-form over the segment 0 t . In this
form, Stokes’ theorem is just a restatement of the fundamental theorem of
calculus for a 1-form. To prove this claim, write for fixed and
2
gt = G̃ u t = e− −Zx t ux t hx det Zx x t dx
Rm
II.2 Distribution solutions 65
Then,
n
g
gt − g0 =
t dtj (∗)
0t j=1 t j
To compute the derivatives of g we write e x t = e− −Zx t , differen-
2
tiate with respect to tj under the integral sign, and recall that
e uh det Zx e uh det Zx
=
det Zx + e uh
tj tj tj
m
= det Zx Lj e uh − det Zx jk e uh
k=1 xk
det Zx
+ e uh
tj
a fact we already used in the proof of (II.10). Once again, the integral over
Rm of the second term of the right-hand side may be written, after integrating
by parts, as
m
e uh jk det Zx dx
k=1 xk
in view of (II.9). Since Lj e u = 0, we also have that det Zx Lj e uh =
det Zx e uLj h. This shows that
g
t = r̃j t (∗∗)
tj
where
r̃j t = e− −Zx t ux t Lj hx det Zx x t dx
2
Rm
Hence, (∗) for = Zx t gives an alternative proof of the fact that R u =
G u − E u as given by (II.11) and (II.12). Notice that (II.12) makes sense if
u ∈ C t ≤ R
x < R as soon as we change the integral symbol
by the duality pairing between the distribution u· t and the appropriate test
function; furthermore, R u = G u − E u is still given by (II.11) and (II.12)
in the case of distribution solutions since (∗∗) is easily seen to remain valid
in this case. Note also that R ux t is a smooth function of x t. We will
prove a stronger form of (d).
66 The Baouendi–Treves approximation formula
Proposition II.2.1. Let u ∈ W satisfy the system (II.5). Then,
where (x is a cut-off function that vanishes for x ≤ R/4 such that
(x Lj hx = Lj hx . Let us write
maps continuously L2s Rm onto L2s+2k Rm and the latter is contained in
L Rm ∩ C 0 Rm if s + 2k > m/2 by Sobolev’s embedding theorem. Hence,
rj x t t is continuous with respect to t and converges to 0 uniformly
for x ≤ R/2 t ≤ t ≤ T , as → , since the derivatives in 1 − $x k
produce powers of that are dominated by the exponential e−c . Hence,
R ux t → 0 uniformly as → and it is easy to see, by differentiating
(II.11), that the same holds for the derivatives of any order with respect to x
and t of R ux t, as we wished to prove.
Finally, it is enough to prove that (c) holds assuming that u ∈ C 0 t ≤
R L2k
loc BR for some integer k. Let us start with the case k = 0. We assume
that u ∈ C t ≤ R Lloc BR (with R slightly larger that R) and we wish
0 2
to prove that
G ux t − ux t 2 dx → 0 uniformly in t ≤ T
x ≤R/4
II.2 Distribution solutions 67
where the convolution is performed in the x variable and t plays the role of
a parameter. Since F L1 = F1 L1 = C, Young’s inequality for convolution
implies
sup G u· tL2 Rm ≤ C sup u· tL2 Rm (II.14)
t ≤T t ≤T
On the other hand, we proved in Section II.1 that if u ∈ Cc V then G u → u
uniformly in U = BR/4 × t < T , which implies convergence in the mixed
norm space C 0 t ≤ T L2 BR/4 . So the operator G U convergesto the
restriction operator u → u U , as → , on a dense subset of C 0 t ≤
T L2 BR and the family of operators G U is equicontinuous
because
of
(II.14). Thus, G u U → u U in the whole space C 0 t ≤ T L2 BR .
Assume now that u ∈ C 0 t ≤ T L21 BR , R > R. Introducing
a cut-off
function we may assume that u ∈ C 0 t ≤ T L21 Rm without modifying
u for x < R. Thus, for t ≤ T fixed, we see that u, u/xk and u/tj
are in L2 Rm for 1 ≤ k ≤ m, 1 ≤ j ≤ n. Since we are assuming that x t
is compactly supported, the coefficients of Lj and Mk are bounded, with
bounded derivatives. In particular, Lj u and Mk u are in L2 Rm for 1 ≤ k ≤ m,
1 ≤ j ≤ n, uniformly in t ≤ T . To obtain the convergence result for k = 1 we
will be able to reason as with the case k = 0 as soon as we prove an estimate
analogous to (II.14) for the L21 norm, i.e.,
sup G u· tL21 Rm ≤ C sup u· tL21 Rm (II.15)
t ≤T t ≤T
sup G u· tL2k Rm ≤ Ck sup u· tL2k Rm k = 1 2 (II.17)
t ≤T t ≤T
≤ Cu· tL2k
where we have used (II.17) for the positive integer k in the last inequality.
This extends (II.17) to all integers k ∈ Z, proving the equicontinuity of G
in all spaces C 0 t ≤ T L2k BR , k ∈ Z, which together with the conver-
gence of G u U to u U for the space of test functions Cc BR × t ≤ T
which is dense in any C 0 t ≤ T L2k BR proves that G u → u in
C 0 t ≤ T L2k BR/4 for any u ∈ C 0 t ≤ T L2k BR . This proves
(c) and concludes the proof of part (i) of Theorem II.1.1.
II.3 Convergence in standard functional spaces 69
To prove part (ii) of the theorem—this is the fifth and final step of the
proof—using the same method of proof, it will be enough to prove the
equicontinuity of G on the spaces
C j t ≤ T Cbk Rm j k = 0 1 2
For j k ≤ 1 one expresses the derivatives in terms of the vector fields Lj and
Mk and reduces the equicontinuity for the norms of C j t ≤ T Cbk Rm to
the case j = k = 0 by introduction of the commutators G Lj and G Mk ,
as was done before for Sobolev norms; iteration of this process gives the
result for k = 2 3 This concludes the proof of Theorem II.1.1.
II.3.1 Convergence in L p
The main result of this subsection is:
In the proof of Theorem II.3.1 we may assume from the start by shrinking
W that u ∈ Lp W and we will do so. We are also tacitly assuming that we
are using special coordinates x t adapted to a given set of local generators
dZ1 dZm of ⊥ with linearly independent real parts so that Z = x +
ix t, where x t is smooth, real, has compact support and satisfies
(II.3 ). Once the special coordinates x t are fixed, the operator E referred
to in (II.19) and (II.20) is defined precisely as in the proof of Theorem II.1.1.
We will also prove below theorems similar to Theorem II.3.1 for different
norms and in all of them the first step will be to choose special local coordi-
nates where Z has this special form where the operators E and G are defined
and have good convergence properties. To avoid repetitions we will always
assume that this step has already been carried out, even if not mentioned
explicitly.
According to the considerations made at the beginning of the section, we
need only prove that
G u −→ hu in Lp W −→ u ∈ Lp W (II.21)
For 1 ≤ p < , the space Cc0 W is dense in Lp W and (II.20) will be a
consequence of
G u −→ hu uniformly −→ u ∈ Cc0 W
(which we already know by Theorem II.1.1) and the uniform bound that we
will prove later:
G up ≤ Cup u ∈ Lp RN > 0 (II.22)
where p denotes the Lp -norm.
II.3 Convergence in standard functional spaces 71
Proof. We take functions ∈ Cc Bx and * ∈ Cc Bt . We know that t →
#Tt u $ is a C -function defined in Bt , t → #ut $ belongs to Lp Bt and
#Tt u $*t dt = ux tx dx *t dt
(II.23)
= #ut $*t dt
If we take *t = (j t −t0 , (j t = j n (jt, 0 ≤ ( ∈ Cc t ≤ 1 , (dt = 1,
and let j → , the left-hand side of (II.23) converges for every t ∈ Bt to
#Tt u $ while the right-hand side converges a.e. to #ut $. Hence, there is
a null set N ⊂ Bt such that
If we apply the last identity to a dense sequence k ⊂ Cc Bx and set
N = Nn we obtain that Tt u = ut as elements of Bx when t is not in
the null set N .
Remark II.3.3. One cannot expect in general that, under the conditions of
Lemma II.3.2, Tt u ∈ Lp for all t. For instance, if = −1 1 × −1 1 ⊂ R2 ,
Z = x + it2 /2, L = t − itx is the Mizohata operator and ux t = 1/Zx t,
it is simple to verify that u ∈ Lp for 1 ≤ p < 3/2, Lu = 0 in the sense of
distributions and Tt u ∈ C −1 1 ⊂ L −1 1 ⊂ Lp −1 1 for t = 0 but
for t = 0 we have T0 u = pv1/x − i)x % Lp −1 1.
We claim that, for u ∈ L1 , there exists a constant C > 0 such that
where
F x = C m/2 e−3 x
2 /4
The last inequality follows from the fact that F1 x = Ce−3 x /4 is radial
2
decreasing and belongs to L1 Rm (see, for instance, [S1, page 62]). Thus,
(II.24) is proved.
If u ∈ Cc0 W, we know that G ux t → hxux t → uniformly.
The standard properties of the maximal operator allow us to conclude that
for any t ∈ Bt such that Tt ux ∈ L1 Bx there exists a subset Nt ⊂ Bx with
Nt = 0 such that
G ux t → hxux t x % Nt
Hence, if we choose x t ∈ U such that Tt u ∈ L1 Bx and x % Nt , we get
(recalling that R u → 0 uniformly in U )
E ux t → hxux t = ux t ae in U
and therefore E ux t → ux t a.e. in U as we wished to prove.
We now prove (II.22). We observe that
G ux t ≤ F ∗ h Tt u det Zx
and then Young’s inequality for convolution implies
G u· tLp dx ≤ F 1 h Tt u det Zx Lp dx ≤ CTt uLp dx
II.3 Convergence in standard functional spaces 73
since the L1 norm of F does not depend on and h det Zx is bounded. Raising
this inequality to the pth power and integrating with respect to t we obtain
(II.22). Since G u → hu uniformly in W as → when u is continuous,
the usual density argument shows that (II.21) holds for 1 ≤ p < . Thus,
(II.19) and (II.20) have been proved. Finally, since E u can be approximated
in C U by polynomials in Z for fixed , the proof is complete.
It is obvious that (II.20) is, in general, false for p = because the uniform
limit of a sequence of continuous functions, such as E ux t, is continuous.
A simple consequence of Theorem II.3.1 is:
where #s fx = −1 1 + 2 s/2 f x and denotes the Fourier trans-
form in RN (#s is the Bessel potential and denotes the space of tempered
distributions). For k ∈ Z+ and p in the range 1 < p < the space Lpk RN
is exactly the subspace of the functions in Lp RN whose derivatives of
74 The Baouendi–Treves approximation formula
The space Lps
loc is the subspace of of the distributions u such that
*u ∈ Ls R for all test functions * ∈ Cc , equipped with the locally
p N
Indeed, (II.27) and (II.28) imply as usual, by density and triangular approx-
imation, that G w − hwps → 0 as → for any w ∈ Lps RN ∩ W—
where W denotes the space of distributions compactly supported in
W —which implies that G w → w in the topology of Lps loc U. We know
that for u ∈ Cc U, G u → u in C U, thus (II.27) is clearly true and we
need only worry about proving (II.28), which we prove first for a positive
integer s = k ∈ Z+ . The vector fields Lj and Mk form a basis of CT Rn and
we may express the derivatives D in (II.26) in terms of the vector fields Lj ,
j = 1 n, Mk , k = 1 m. This gives
G wLpk ≤ C M 1 L2 G wp (II.29)
1 + 2 ≤k
We write
Lj G w = G Lj w + Lj G w
Mk G w = G Mk w + Mk G w
Thus, for 1 ≤ j ≤ n, 1 ≤ k ≤ m,
and
G wpk ≤ Cwpk w ∈ Cc RN
≤ Cs wLps RN
76 The Baouendi–Treves approximation formula
where in the last inequality we used (II.28) with q in the place of p and
−s > 0 in the place of s. Thus, (II.28) is completely proved and the proof of
Theorem II.3.5 is complete.
defined as
C = u ∈ C k u <
where
u = u + u 0
u 0 = sup ux
x∈
ux − uy
u = sup 0 < ≤ 1
xy∈ x−y
x=y
u= D u −k k < ≤ k + 1 k ∈ Z+ u ∈ C k
≤k
The spaces C RN are defined similarly. The approximation theorem is:
Proof. As always, since Cc W is dense in Cc W for the C norm, we
need only prove
G u −→ u in C W u ∈ Cc W
and the inequality
G u ≤ Cu u ∈ Cc W
It is obvious that G u − u → 0 when → , u ∈ Cc W, because by
Theorem II.1.1 G u → u, → in C k W for every positive integer k.
We may assume without loss of generality, as we always do, that Zx t =
x + ix t is defined and satisfies (II.3 ) throughout RN and reduces to
II.3 Convergence in standard functional spaces 77
We assume first that 0 < < 1. It will be useful to use the following
well-known characterization of C RN ([S2, page 256]):
It also follows that the best constant K in (i) and (ii) above is proportional to
u . Such a sequence is usually called a sequence of best approximation for
u. We start by writing u = uk with uk a sequence of best approximation
for u. Then, G u = G uk and we need to estimate the essential supremum
of G uk and ,G uk . Taking account of (II.22) with p = and (i) of Lemma
II.3.7 we derive
Thus, (II.34), (II.35) and Lemma II.3.7 imply that (II.33) holds for 0 < < 1.
Let us assume next that there is a positive integer k such that = k + ,
0 < < 1 and we wish to estimate
G u ∼ D G u ≤ C M 1 L2 G u
≤k 1 + 2 ≤k
78 The Baouendi–Treves approximation formula
Then ([S2])
H p RN = f ∈ RN M! f ∈ Lp RN
Notice that if the support of a is contained in a cube Q such that (i) holds and
the side of Q has length ≥ 1, then a is an atom, as condition (ii) is vacuous
and only (i) is required in this case.
As always, (II.37) follows from the convergence G v → v in Cc ,
v ∈ Cc . So, to prove Theorem II.3.9, we need only show (II.38) and
the density of Cc RN in hp RN . To prove the density, it is enough to
approximate hp atoms by smooth hp atoms in the hp norm. This is simply
approximating a rough atom a by the convolution a- = a ∗ *- , where *- z =
-−N *z/-, and * ∈ Cc RN has integral equal to 1. Then, a- satisfies the
vanishing moments condition (ii) because a does and satisfies (i) for a cube
Q slightly larger than the one that worked for a, if - > 0 is sufficiently small.
Moreover, a- → a in the hp ‘norm’ as - → 0. To check the last fact use
Hölder’s inequality to write
m! a − a- zp dz ≤ Q 1−p/2 m! a − a- L2
Indeed,
" #p " #p
m ! G k ak dz ≤ k m ! G ak dz
k k
≤ k p
m! G ak p dz
k
Here and in the sequel, (A will denote the characteristic function of a measur-
able set A. Let Q∗1 (resp. Q∗∗
1 ) be the cube in R concentric with Q1 having
m
twice (resp. four times) the side length. Then (II.40) shows that
m! G ax t p dx dt ≤ C (II.41)
Q∗∗
1 ×R
n
Assuming that !x t = !1 x!2 t, !1 and !2 supported in the unit ball of
x x
Rm and Rn respectively, we are led to consider the convolution !-1 ∗ a ∗ F .
In order to simplify the notation we simply write !-1 ∗ a ∗ F , letting t play
the role of a parameter. Let us assume first that the side r of the cube Q
is ≥ 1. Since !1 is supported in the unit ball, !-1 ∗ a = a- , 0 < - ≤ 1, is
supported in Q∗1 . Therefore, if x % Q∗∗ 1 , letting x0 be the center of Q1 and
CL = supx∈Rn x L Fx, we have
& &
& &
!-1 ∗ a ∗ F x t ≤ (Q2 t & a- y tF x − y dy&
( )
x − x0 −L
≤ CCL (Q2 t Q −1/p Q∗1 −m
where we have used that x − x0 ∼ x − y for y ∈ Q∗1 and x % Q∗∗1 . Since
∗
Q1 = 2r ≤ 2 x − x0 and
m m L−m
≤ 1 if we take L > m, we obtain for a
large integer d = L − m
x t
!- ∗ F ∗ ax t ≤ C Q −1/p x − x0 −d
!-2 ∗(Q2 t
≤ C Q −1/p x − x0 −d
(Q∗2 t
assumption r ≥ 1.
Let us assume now that r < 1, so az satisfies the moment conditions (ii).
It is clear that these properties are inherited by a- z, i.e., z a- z dz = 0,
≤ N1/p − 1. We start by writing Fx as a convergent series in Rm ,
Fx = k F k x with F 0 supported in the unit ball B = B0 1 and each
F k supported in some ball of radius 1. We aim at proving (II.42) with F k
in the place of F . Using the vanishing of the moments of a
x
a- ∗ Fk x t = (Q∗2 t ay t Gk
- x − y dy
= ay t Gk
- x − y − qx- y dy (II.43)
where Gk - = !- ∗ F and qx- y is the Taylor polynomial of degree d of
1 k
the function y → Gk - x − y expanded about x0 and d is the integral part of
N1/p − 1. The usual estimates for the remainder of the Taylor expansion
imply that the integrand in (II.43) is ≤ C Q −1/p −d+1+m r d+1 . We assume
first that k = 0 so F 0 is supported in the unit ball. Since x − x0 ≤ C x − y
when y ∈ Q∗1 and x % Q∗∗ 1 , x − y ≤ on the support of F x − y, and a is
0
supported in the cube Q∗1 of measure 2rm it follows that for any 0 < % ≤ 1
and 0 < ≤ 1
d+m+1p
r
a- ∗ F0 x t p ≤ C0 (Q2 t x % Q∗∗
1
x − x0
which after integration gives
sup !- ∗ F0 ∗ ax t p dx dt ≤ C0 (II.44)
Rm \Q∗∗
1 ×R 0<-≤1
n
II.4 Applications
In this section we discuss two typical applications of the Baouendi–Treves
approximation formula. The first one deals with extensions of CR functions
and the second with uniqueness of solutions of the equation u = 0 where
is a locally integrable structure. The principle that governs the first appli-
cation is conceptually very simple: suppose that we know that a sequence of
polynomials P , ∈ Cm , converges uniformly in a compact set K ⊂ Cm ,
then it converges uniformly in the holomorphic convex hull K $ of K in Cm .
We recall that
*
$=
K ∈ Cm P ≤ sup P
P∈ K
is not continuous but, say, belongs to Lp , things are technically more involved
but essentially the same principle works.
This type of approach may also be seen at work in the following simple
example. Consider the operator in R2
L= − 3it2 (a)
t x
with first integral Zx t = x +it3 . Indeed, it is easily verified that LZ = 0 and
clearly dZ never vanishes. The operator L has real analytic coefficients and
is elliptic off the x-axis but is not elliptic at t = 0, nevertheless it shares with
elliptic vector fields with real analytic coefficients the following regularity
property: if u is a C 1 solution of Lu = 0, then u is real-analytic ([M]). This
is also true for distribution solutions (thus, (a) is analytic hypoelliptic) but to
keep matters simple let us restrict ourselves to classical solutions. To prove the
claim, it will be enough to prove that u is real-analytic at any point x 0 of
its domain, since for points x t with t = 0 this follows from ellipticity. Let
us prove, for instance, that u is real-analytic at the origin in case it is defined
in a neighborhood of the origin. By Theorem II.1.1 we may find > 0 such
that for x ≤ and t ≤ the uniform limit ux t = lim→ P x + it3
holds for a certain sequence of polynomials P , ∈ Z+ . This implies that
the sequence P z = P x + iy is a Cauchy sequence in the space C 0 K
where K = − × −3 3 . Hence, lim→ P z = $ uz is a continuous
function on K which is a holomorphic function on − × −3 3 and
we have that ux t = $ ux + it3 for x t ≤ . Since $ u is real-analytic in
a neighborhood of the origin and so is Zx t = x + it3 , it follows that u is
real-analytic in a neighborhood of the origin as we wished to prove.
Then Hn can be topologically identified with the boundary of the Siegel upper
half-space
n
Dn+1 = z1 zn+1 ∈ Cn+1 zn+1 > zj 2
j=1
II.4 Applications 85
Observe that the vector fields L̃j are left-invariant under the action of
Hn . The components of the map (II.47), that is, the functions Z1 z s =
z1 Zn z s = zn , Wz s = s + i z 2 satisfy (II.48) and it is of interest to
determine which solutions of (II.48) may be expressed as the composition of
the map (II.47) with a holomorphic function defined in Dn+1 and having a
suitable trace in Dn+1 . It is known ([FS]) that a function f ∈ C 1 Hn is a CR
n+1
function if and only if there exists a function F ∈ C 1 D which is holo-
morphic in Dn+1 and whose composition with the map (II.47) is equal to f .
There is also a similar local result due to Hans Lewy ([L1]) which holds in
the general set-up of CR structures of hypersurface type with nondegenerate
Levi form which we now describe. Consider a hypersurface in Cn+1 with
the CR structure induced by the standard anti-holomorphic differentiations
of Cn+1 . We may assume that, in a suitable neighborhood of the origin in
Cn+1 , is given by
where
n
2 !
!z s = 0 0zj z̄k + O z 3 + s z + s2
ij=1 z j z̄ k
Lj u = 0 (II.50)
u=F Z
where each %j may assume the values +1, −1, or 0. We will assume initially
that the remainder terms vanish identically because the proof is very simple
in this case. Hence, we assume that
n
!z s = !z = z1 2 + %j z j 2 (model case)
j=2
as
a a 2) Rz t
I= ds dt dx dy d urei z s t p rdr
−a −a $ 0a 0 0
where, for a given z , z indicates the value of t below which the disk
Dz t becomes empty (if this ever happens) or −a, whichever is larger.
Now the substitution = Rz t in the integral with respect to t (so that
t = ! z ) yields, assuming a is sufficiently small,
a 2√a 2)
I≤ ds dx dy √ 3 d uei z s ! z p d
−a $ 0a −2 a 0
a √
2 a 2)
≤ ds dx dy √ d uei z s ! z p d
−a $ 0a −2 a 0
≤2 √ u Z p dxdyds
$02 a×−aa
Lemma II.4.2. There exist t0 r0 > 0 such that for all 0 < t ≤ t0 and ∈
% , D̃m + t is a relatively compact simply connected open subset of the
disk D0 r0 . Furthermore, there exists C > 0 such that for every harmonic
function u defined in a neighborhood of the closure of D̃m + t and any
1 ≤ p < , the following a priori inequality holds:
u p dA ≤ C diam D̃m + t u p d (II.54)
D̃m+t D̃m+t
Proof. After a translation, we may assume that z0 = 0. For small t > 0, the
level curve x y = m + t, which is implicitly given in polar coordinates
by r 2 A + rBr = t where A = cos2 + 2 sin cos + sin2
and all derivatives of B with respect to x and y are bounded, may also be
explicitly expressed by r = r t. Observe that if % is small, and are
close to 1 and is close to zero. Implicit differentiation shows that
r rA + r 2 B √
r = =− = O t t → 0
2A + 3rB + r Br
2
√
dilation of D̃m + t, t = 1/ tD̃m + t, whose boundary is given by
√ √
Rt = r t/ t = A−1/2 + O t. Observe that we also have
√
dn Rt /dn = dn A−1/2 /dn + O t for n ≥ 1.
Since (II.54) is invariant under dilations of the domain, it will be enough
to prove it for the dilate t that converges in C to the domain 0 with
equation R < A−1/2 as t → 0. To do so it is enough to show that, for small
t, the derivative Ft of the Riemann map Ft from t to the unit disk satisfies
1/C ≤ Ft ≤ C. Indeed, if u is harmonic in t and continuous up to the
boundary, so will be v = u Ft−1 on the unit disk, and starting from (II.52)
applied to v, the change of variables w = Ft z will give
u p dA ≤ C u p d (II.55)
t t
Remark II.4.4. Stronger results than Theorem II.4.1 are known. In fact, it is
possible to sweep V + with suitable translates of so that the Lp norm of the
restriction of F to those translates is uniformly bounded ([Ro]). Theorem II.4.1
then follows from an application of Fubini’s theorem.
(i) u = 0 in ;
(ii) u 0 = 0;
of generators of , L = L1 L2 Ln such that in appropriate coordinates
x t, x < 1, t < 2, that rectify the flow of X1 = X we have
(i)
X1 = L1 =
t1
m
Xj = Lj = + jk j = 2 n
tj k=1 xk
and p = 0 0;
(ii) s = s − 1 0 0, q = 0 = −1 0 0;
(iii) for some a > 0 the embedded closed m-ball given by x ≤ a, t = 0,
t1 = −1 does not meet K (here t = t2 tn ). Since it is an embedded
submanifold with boundary we may denote this m-ball as 00 ∪00 , where
00 is the corresponding open m-ball.
We claim that
00 ∩ K = ∅ (B)
Indeed, since dist00 K = 0, this is certainly true if we replace 00 by its
closure 00 which amounts to adding to 00 its boundary points 00 . But,
for any ∈ 0 1 , 0 is given by x = a, t1 = −1, t2 = · · · = tn = 0, so (iii)
shows that 0 ∩ K = ∅. Hence, 00 ∩ K = 00 ∩ K = ∅. However, applying
Theorem II.4.6 to 00 , (A) implies that u vanishes in a neighborhood of 00
in x < a, t < 2. This contradicts (B).
Let be a manifold and consider a collection D = X of locally defined,
smooth, real vector fields X. In Chapter III, the notion of orbit of D is
II.4 Applications 95
Example II.4.10. P. Cohen ([Co]) (see also [Zu] and the references therein)
constructed smooth functions ux y and ax y defined on R2 such that
u u
(1) Lux y = + ax y = 0;
y x
(2) ux y = ax y = 0 for all y ≤ 0;
(3) supp u = supp a = x y y≥0 .
Thus, the formally integrable structure spanned by the vector field L fails to
have the Uniqueness in the Cauchy Problem property for the noncharacteristic
curve 0 = t = 0 and, by Corollary II.4.9, cannot be locally integrable in
any open set that intersects the x-axis. The construction of ax y shows
that ax y is real-analytic for y = 0, so for any point p = x y with y = 0
we may find a function Z defined in a neighborhood of p such that LZ = 0
and dZp = 0. On the other hand, if Z is a smooth function defined in a
neighborhood of p = x 0 such that LZ = 0 we must have that dZp = 0,
otherwise would be locally integrable in some open set that intersects the
x-axis, a contradiction. A nonlocally integrable vector field was first exhibited
by Nirenberg ([N1]) who used a completely different method to construct a
vector field whose only homogeneous solutions are contant.
II.4.3 An extension
In the applications to uniqueness we have seen so far, the ‘initial’ maximally
real manifold t = 0 is in the interior of the domain where the solution u of
u = 0 is defined. This is quite convenient because in this case the trace
u· t exists and t → u· t is a continuous function of t valued in the space
of distributions. However, in the study of one-sided Cauchy problems or
boundary values of solutions, it is desirable to consider the case where the
solution is not defined in a neighborhood of the ‘initial’ manifold. We will
say that a set + ⊂ Rn \0 is a cone (or a cone with vertex at the origin to be
explicit) if t ∈ + ⇐⇒ &t ∈ + ∀ 0 < & < . A set +T ⊂ Rn \0 , 0 < T , will be
called a truncated cone if there exists a cone + such that +T = + ∩ t < T .
An open truncated cone is a truncated cone which is an open set. Notice that
the origin is in the closure of + and +T but it does not belong to them. A cone
+ is said to be a proper subcone of + if + ∩ x = 1 is a compact subset of
+. This is, for instance, the case if + and + are circular cones with the same
II.4 Applications 97
and
e− Zxt−Zx t ux thx det Zx x t dx
2
G ux t = /)m/2
Rm
where
rj x t t = /)m/2 e− Zxt−Zx t ux t Lj hx det Zx x t dx
2
Rm
98 The Baouendi–Treves approximation formula
Letting % → 0 we obtain
R u = G u − E u
with R given by
m
R ux t = rj x t t dtj
0t j=1
rj x t t = /)m/2 e− Zxt−Zx t ux t Lj hx det Zx x t dx
2
Rm
The proof of Theorem II.1.1 now shows that there is a ball Bx = Bx 0 and
proper subcone +& ⊂ +T such that R u → 0 uniformly in Bx × +& as → .
Indeed, we can find a fixed k such that vj x t = 1 − $x −k Lj hx ux t
detZx x t is continuous in Bx ×+& , since the distributions x→Lj hx ux t
detZx x t lie in a bounded set of some Sobolev space when t ranges
over a compact subset of +T ∪ 0 because u ∈ C 0 +T ∪ 0 Bx . Since
the continuity of +T ∪ 0 t → Tt ux ∈ Bx implies the continuity of
+T ∪ 0 t → Dx Tt ux = Tt Dx ux ∈ Bx and equation (II.5) allows
us to express the derivatives of u with respect to t as a linear combination
with smooth coefficients of derivatives of u with respect to x for t = 0, we
conclude that actually u ∈ C +T ∪ 0 Bx . The derivatives of R u can
be estimated in the same fashion and we obtain
Notes
The approximation formula of Section II.1 for classical solutions was first
proved by Baouendi and Treves in [BT1], building upon their previous work
([BT2]) that dealt with a corank one system of real-analytic vector fields.
For distribution solutions, the proof in [BT1] relied on a local representa-
tion formula proved under a supplementary hypothesis on the locally inte-
grable structure. This representation formula, which is of independent interest
and states that any distribution solution u of u = 0 may be written as
u = Px Dv, where v is a classical solution of v = 0 and Px D is a
differential operator that commutes with the local generators Lj , 1 ≤ j ≤ n, of
. This representation formula was proved in general by Treves in [T4], who
also stated and proved the approximation formula for distribution solutions
in all generality. Metivier studied the case of a nonlinear first-order analytic
single equation and proved an approximation formula for solutions of class
C 2 , obtaining as a consequence uniqueness in the Cauchy problem ([Met]).
100 The Baouendi–Treves approximation formula
The equivalence classes of this relation will be called the orbits of D. In [Su],
Sussmann showed that these orbits can be equipped with a natural topology
and differentiable structure which makes them immersed submanifolds of .
We will next briefly describe the orbit topology and C structure (the reader
is referred to [Su] and [BER] for more details). If X ∈ D is defined near p in
, let !tX p denote the integral curve of X which at t = 0 equals p and is
defined on a maximal interval. If Y = X1 Xm ∈ Dm (i.e., each Xi ∈ D),
101
102 Sussmann’s orbits and unique continuation
We will use +D to denote the smallest set of locally defined, smooth
vector fields which contains +D and is invariant under the group of local
diffeomorphisms generated by +D. It is not hard to see that the dimension
is constant as x varies in the orbit . Suppose now q ∈ .
of the fibers +D x
By lemmas 5.1 and 5.2 in [Su], there exist Y ∈ Dn for some n, q ∈ and
s ∈ Y q such that
!Y q s = q
and the rank k of the differential of
!Y q Y q −→
for
at the point s is maximal, and that in fact, this rank equals dim +Dx
any x ∈ . By the rank theorem, we can find neighborhoods U of s in Rn , V
of q in , diffeomorphisms F from U onto C n , G from V onto C N (N =
dimension of ) such that
G !Y q F −1 x1 xn = x1 xk 0 0
Here C l denotes the cube
x1 xl ∈ Rl xi < 1 ∀i
III.1 Sussmann’s orbits 103
G# = x1 xk 0 0
rational numbers, each orbit is diffeomorphic to the real line. In this case the
orbits are dense in T2 , and hence are not embedded submanifolds.
k
fj Xj
j=1
and let
= z w ∈ C2 \0 &z w = 0
P = v ∈ V *0 v ∈
106 Sussmann’s orbits and unique continuation
basis for the topology of . Hence we only need to show that each Ai consists
of a countable number of slices in Uim . We will do this by induction on m.
When m = 1, Ai1 contains at most one slice. Suppose the result holds for all
tuples j = i1 ik−1 of length k − 1. Then Aj is the union of countably
many slices in Uik−1 . Fix a slice 0 in Aj . Since slices are open sets in , the
intersection of 0 with each slice in Uik is an open set. Moreover, since the
slices in Uik are pairwise disjoint, and 0 is homeomorphic to an open set in
Rd , it follows that 0 can intersect only a countable number of slices in Uik .
Thus each Ai is the union of a countable number of slices and therefore is
second countable.
The preceding lemma can be used to show that orbits possess properties
not shared by a general immersed submanifold. To see one such property, call
an immersed submanifold N of a manifold weakly embedded if whenever
A is a manifold and f A −→ is smooth with fA ⊆ N then f A −→ N
is smooth. This notion was introduced by Pradines in [Pr]. For an example
of an immersed submanifold that is not weakly embedded, see remark 6.8 in
[Boo].
∩ *T × V = *T × v
v∈P
Proof. Let q ∈ and assume q = !tX p for some X ∈ + and t ∈ R. Set
W = !tX W. Here we may assume W has been contracted enough to lie in
the domain of the flow of !tX . Since !tX W −→ W is a diffeomorphism, the
submanifold !tX W ∩ is an embedded submanifold of W . It is also easy
to see that
!tX W ∩ = W ∩
Hence is an embedded submanifold of .
Before we provide the proof, we will recall some definitions and results from
a paper of Bony ([Bo]).
Since L is in the tangential direction on 0, each aj ∈ C 0. Let *jn r =
fj r ein , where fj r is C and fj 1 = jl for 0 ≤ j ≤ k. Note that the
l
transpose of L is given by
w w
t
Lw = − + irYr + i2Yr + rY rw
r
and so
t
L*kn = iein rYrfk r + 2Yr + rY r − nfk
Moreover,
m
0 m<k
rYrfk r r=1 =
r kY 1 m=k
and
m
0 m<k
2Yr + rY r − nfk r r=1 =
r Y 1 − n m = k
Thus, we get:
0 = #Lu *kn $
= #ut L*kn $
2)
= ak e d k + 1Y 1 − n
in
(III.1)
0
Since we may assume that ak does not vanish identically, there is an
integer M for which
112 Sussmann’s orbits and unique continuation
2)
ak eiM d = 0 (III.2)
0
ak = ce−iM
k
2) m
#u *r $ = bm *1 d
m=0 r
0
Set bk = e−iM . Each bj can be determined from the equation #Lu *jM $ =
0. To see this, note that #Lu *k−1M $ = 0 is equivalent to
⎛ ⎞
2)
d k
0 = 2)
rYfk−1 + rY + 2Y − Mfk−1 1 + ⎝ bk−1 eiM d⎠
dr
0
k−1
d
× rYfk−1 + rY + 2Y − Mfk−1 1
dr
2)
The coefficient of bk−1 eiM d in the latter equation is −2kh1 − M =
0
2h1 = 0, and hence we can get a constant ck−1 such that if we set
In general, we can determine bl from #Lu *lM $ = 0. This leads to bl =
cl e−iM for some constant cl since
l
d
rYfl + 2Y + rY − Mfl 1 = 2k − lh1 = 0
dr
III.2 Propagation of support and global unique continuation 113
(1) The orbits of 2L through p1 = 1 1 1 and p2 = 1 1 −1 are compact
but all other orbits are noncompact.
(2) Depending on the value of
2)
xdx
0
there may not be any solution supported on either of the compact orbits.
(3) Global unique continuation is valid for continuous solutions.
Let F R3 −→ T3 be given by
Consider the orbit 1 through the point p1 = 1 1 1. F0 0 0 = p1 and
the orbit in R3 of X Y through 0 0 0 is Rx × Ry × 0 . Therefore, 1 =
S 1 × S 1 × 1 . Likewise, for the point p2 = 1 1 −1, the orbit 2 = S 1 ×
S 1 × −1 . Consider now a typical point p = 1 1 eis0 for some 0 < s0 < ).
If t = xt st is the integral curve of X with 0 = 0 s0 , we will
see that the orbit through p is given by
Indeed, xt = t and s t = t sinst, s0 = s0 . If for some t0 , st0 = ),
then the curves t and 1 t = t ) will both be integral curves of X
passing through t0 ) at t = t0 . This implies that st ≡ ), contradicting the
assumption that s0 = s0 < ). Likewise, st can never equal zero. Thus,
0 ≤ st ≤ ) and s t ≥ 0. Suppose
Then s0 ≤ st ≤ a for all t ≥ 0. Therefore, s t ≥ ct for some c > 0,
which in turn leads to
lim st =
t→
Hence,
lim st = )
t→
We have
#Lk gx y s$ = #k t Lgx y s$
2) 2) k ( )
g
=− x sin s + cos sg x y 0 dxdy
0 0 s s
2) 2) k+1
=− x sin sg x y 0 dxdy
0 0 s
k−l l
2) 2) k + 1
k+1
=− cos s gx y 0 dxdy
l=0 0 0 l s s
Thus,
k−2
k + 1 k−l
Lk = −k + 1xk − x s cos s0l
l=0 l
Let Mk = +i −k+1x, for k = 0 1 2 If vx y ∈ C S 1 ×S 1 ,
x y
it follows that
k + 1 k−l
k−2
Lvk = Mk vk − v s cos s 0l
l=0 l
III.2 Propagation of support and global unique continuation 115
Suppose now
N
u= uk x yk
k=0
is a solution. Then
N
N −2
N
k+1
Lu = Ml ul l − uk sk−l cos s 0l
l=0 l=0 k=l+2 l
Let
1 2)
0 = x dx
2) 0
and define 1 = − 0 . Since Lu = 0 and uN may be assumed nontrivial,
we must have N + 10 ∈ Z. Thus if 0 is not a rational number, there
are no solutions supported on the orbit 1 . If 0 is rational, with 0 = p/q
where p and q are relatively prime, then N = q − 1 is the lowest possible
transversal order of a nontrivial solution supported on 1 . This follows from
the injectivity of Ml for l < N and the fact that MN has a nontrivial kernel.
Since Ml is also surjective for l < N (as is easily seen using Fourier series),
one can correct the ‘errors’ to obtain a solution u iteratively.
Finally, we remark that there are solutions supported on the closure of any
noncompact orbit. This will follow from Theorem III.2.12 as stated below,
or can be constructed explicitly as in [BM]. Thus, global unique continuation
is not valid for distribution solutions. However, it is valid for continuous
solutions.
We will now place these two examples in a more general context following
[BM]. Given a locally integrable structure , let 0 be an orbit such
that dim 0 < dim = m + n, where n is the rank of . Assume that 0 is an
embedded submanifold of . Fix p ∈ 0 and let Z1 Zm be a complete
set of first integrals defined in a neighborhood U in of p. Let L1 Ln
be smooth, local generators of in U such that the brackets Li Lj = 0 for
all i, j. Complete this to a basis
L1 Ln M1 Mm
of CT in U such that
(1) Li Mk = 0, and
(2) Mk Zi = ik .
− hLj ∧ dZ
U
= − hLj ∧ dZ
U
= − #h Lj ∧ dZ$ ∀j = 1 n
It follows that for the distribution u supported on U ∩ 0 as before, if ∈
Cc U, we have:
=− #u M Lj ∧ dZ $ ∀j = 1 n
≤N
Assume now that u is also a solution. We will next show that each u is a
solution of the induced structure 0 . Fix a point q ∈ U ∩ 0. The restrictions
of Zl (l = k + 1 m) to 0 are solutions of 0 . By the Baouendi–Treves
approximation theorem, for each such Zl , there is a sequence Pil i=1 of
holomorphic polynomials such that
in C V ∩ 0 for some neighborhood V of q in . For each l = k + 1 m,
define the sequence fil
i=1 by
lim fil = 0
i→
= −#Lj u ∧ dZ $ + Ei
0 = #Lj u * ∧ dZ$
= #Lj u M * ∧ dZ $
≤N −1
for any * ∈ Cc U. Plugging * = fi with = N − 1 and ∈ Cc V in
these latter equations will likewise lead to
Lj u = 0 whenever = N − 1
118 Sussmann’s orbits and unique continuation
= M 0
where the i are exact and Z1 Zm is a complete set of first integrals.
We will assume that the restrictions of Z1 Zk to U ∩0 form a complete
set of first integrals for 0 . If u is a solution supported on 0 of transversal
order zero, then we know that it is given by distributions u in U ∩ 0 in the
sense that for any ∈ Cc U ,
where in the right-hand side we mean the pullback of the form on 0. Let
V = U ∩ 0 and whenever V ∩ V = ∅, let g ∈ C V ∩ V satisfy
where for a form in , i∗ denotes the pullback to 0. Note that the g are
nonvanishing and on V ∩ V , g u = u . Therefore, 0 = Lj u = Lj g u .
If Lj g is not zero on an open set, then u will be zero there. But then u
will vanish on this open set and hence on 0, contradicting the nontriviality
120 Sussmann’s orbits and unique continuation
g V ∩ V −→ C
Definition III.3.1. The structure satisfies the strong uniqueness prop-
erty if every solution u ∈ L1loc that is zero on a set of positive measure
vanishes identically.
It turns out that orbits play a role in the validity of the strong uniqueness
property. Before stating the main results, we need to introduce refinements
of the concept of an orbit.
In the proof of Theorem III.3.15, we will use the following lemma where
0 is a smooth hypersurface in Cn , f is a CR function on 0, and f ∈ Lp 0
for some 1 ≤ p ≤ . Suppose also that f extends to a holomorphic function
F on a side 0+ , that is, f is the boundary value of F in the distribution sense.
Then we have:
Lemma III.3.16. For any 0 ⊂⊂ 0, and a sufficiently small ball B in Cn
containing zero, the restrictions of F to the hypersurfaces z ∈ B dist z 0 =
t have uniformly bounded Lp norms. In particular, F ∈ Lp B ∩ 0+ .
D% = y ∈ D Gx y > %
2% z y is the Green’s function for
have smooth boundaries. Observe that if G
D% , then
2% x y = Gx y − %
G
P% x 1−1
% yJ% y −→ Px y
for some open convex cone + ⊂ Rd and > 0. We may suppose that
dx y s < 41 for x y s < 2. Without loss of generality, assume
that
+ = v = v vd v < 2vd
Let
2
+ = y t ∈ Rn+d y t < td t = t td
t = s s ∈ U
Notice that we have included (iv) here since it follows from (iii) and the
Cauchy–Riemann equations satisfied by → Asv at = 1. The meaning
of (i) and (ii) is that Asv is partially attached to 0 at p = s s and (iii)
implies that we can choose a neighborhood Ũ ⊂ U of the origin and a small
% > 0 such that for every p = s0 + is0 , s0 ∈ Ũ , the map
and writing v in polar coordinates we see that for some 0 < % < 2% our claim
is true for 2
E % . We fix such an % > 0 and, dropping any reference to the
dependence on % , simply write 2
E% = 2
E.
Consider now the map
U0 × S m−1 0 % ∩ V0 × 1 − % 1 s v r −→ Asv r ∈ Cm (III.3)
Taking account of (iv) we note that this map has rank 2m for small % > 0 and
maps s × S m−1 0 % ∩ V0 × 1 − % 1 onto Bp \p , where Bp is a C 1+ -
differentiable m-ball that intersects 0 orthogonally at p = s + is. Indeed,
the respective tangent spaces at p are
Tp 0 = s + is + v + i s · v v ∈ Rm and
Tp Bp = s + is + iv − s · v v∈R m
the wedge . Then the composition FAsv rei is defined for −% < < %,
1 − % < r < 1, is holomorphic and has a weak boundary value which, for
a.e. s v ∈ 2 E, is given by—and the proof of this fact is our second step—
fAsv ei . The third step is to prove that for a.e. s v, the restriction of f to
the curve %/2 % → Asv ei is in Lp . Hence, by Corollary III.3.17 and
the classical theorem of Priwaloff, the holomorphic function of one complex
variable FAsv rei vanishes identically for −% < < %, 1 − % < r < 1, in
particular for = 0. But we know that letting s v r vary on 2 E × 1 − % 1
and keeping = 0, the union of Asv rei covers $ E. Thus, F vanishes a.e.
on $E and so must vanish identically. The proof of the second step involves a
discussion about the trace which will be developed next.
We begin our considerations by looking at the simplest case of a holomor-
phic function of one complex variable Fx + iy defined for x < 1, 0 < y < 2
which satisfies the inequality
Fx + iy ≤ C log y x < 1 0 < y < 2 (III.4)
We assume (III.4) for simplicity but the argument below can be iterated to
handle the case Fx + iy ≤ C y −N . The standard manner of defining the
weak trace f of F as an element of is through the formula
#f *$ = lim Fx + i- *x dx * ∈ Cc −1 1 (III.5)
-(0
In formula (III.5) we see that for each fixed x the argument of F describes
a straight vertical segment - → x + i- that flows toward x as - → 0. We
wish to see what happens if we change each vertical segment to a curve
- → x + x - + i-. We will assume that −1 1 × 0 1 x - → is of
class C 2 (we would need class C N +2 if we were assuming Fx +iy ≤ C y −N
instead of (III.4)) and that x 0 = 0, x < 1. The latter assumption simply
means that the curve - → x + x - + i- flows toward x as - → 0. Thus,
j
x 0 = 0 j = 0 1 2 (III.6)
x
We now define
#2
f *$ = lim Fx + x - + i- *x dx * ∈ Cc −1 1
-(0
Fx+x - + i- *x dx = Fx + x - + i*x dx
1
*x
+i Fx + x - + it dt dx
- x 1 + x x -
The first fact follows from proposition 7.2.22 in [BER]. The second fact is
true because theorems 6.5.4 and 7.4.12 in [BER] are valid with the same
proofs if the analytic disks are taken to be in C k for a fixed positive integer
k. In the proof of theorem 7.4.12, the function h has to be modified so that
one gets a C k extension.
Set s = s1 sm−1 . We will assume without loss of generality that
We may now reason as in Lemma II.3.2 to conclude that for a.e. s , s < %,
f̃s ∈ Lp −% % and f̃s = fs . If s is in the set (III.7) and f̃s = fs
holds, then → FAs 0v0 has an Lp boundary value that vanishes on
a set of positive measure which implies that → FAs 0v0 vanishes
identically. We conclude that for a.e. s on the set (III.7), FAs 0v0 = 0,
or equivalently, that the set
E0 v0 = s s < % such that F As 0v0 ≡ 0
has positive measure. A similar conclusion could have been reached for the set
Esm v = s s < % such that F As sm v ≡ 0
where sm is a small number and v − v0 is small. Thus, the set s v
such that F Asv ≡ 0 has positive measure and so does the union of the
corresponding partially attached disks.
132 Sussmann’s orbits and unique continuation
Proof. We first flatten 0 near the origin so that in coordinates x y, 0 =
x 0 . By Proposition III.5.1, L is not tangent to 0 and so if Zx y is
a first integral near the origin, then Zx 0 0 = 0. Assume Z0 0 = 0. Let
s = Zx y and t = y. Then in s t coordinates,
Z = s + is t
and we may take
it
L= −
t 1 + is s
134 Sussmann’s orbits and unique continuation
We may assume > 0 near the origin. For % small, let = ZD% 0 where
D% 0 denotes the disk centered at 0 of radius %. Let be a smooth subdomain
of such that 0 ∈ and the boundary part of near 0 is s s . By
the Riemann mapping theorem, there exists a holomorphic function H which
is a diffeomorphism up to such that
where ˜ > 0 near the origin. Let x = H Zs t and y = ts˜ t k1 . It can
easily be checked that these are coordinates near 0 and in these coordinates,
Wx y = x + iyk
is a first integral.
Proof. Since k is odd, if we take the first integral Zs t = s + is t with
s t = s + tk s t, 0 = 0 as in Proposition III.5.2, we see that L
is hypocomplex. Therefore, to prove the necessity, we only need to do it for
this first integral. Suppose then x y are coordinates in which x + iyk is a
solution. Let F = U + iV denote this diffeomorphism and we may assume
F0 = 0. Then F maps the characteristic set of L to that of
− ikyk−1
y x
The latter, together with the fact that H 0 ∈ R (we assume 0 = 0),
implies that F is a diffeomorphism near the origin. Clearly, using F F
as new coordinates, we get x + iyk as a first integral for L.
Thus, we have:
The preceding corollary was also proved in [Me1]. Proposition III.5.5 can
be generalized as follows:
Proposition III.5.8. Suppose L1 and L2 are two vector fields of the same
uniform odd type on their respective characteristic sets 01 and 02 . Then there
exists a local diffeomorphism mapping the structure generated by L1 to the
one generated by L2 if and only if for any first integrals Z1 and Z2 of L1 and
L2 respectively, there exists a local biholomorphism mapping Z1 01 onto
Z2 02 .
Z1 = x + ix2 y + y3
Z2 = x + ix4 y + y3
It follows that the type at the origin is odd. Since b does not change sign, L
is locally solvable (Theorem IV.1.6) and hence locally integrable.
We are now ready to state and prove the key lemma from [Cor2] concerning
approximate solutions:
Lemma III.5.13. Let L be locally integrable near the origin with a first
integral Zx t = x + i!x t. Suppose that
!t x t = ax ttk
where k is a positive integer and a is never zero for x , t ≤ , > 0. Let
u be a nontrivial function satisfying on x < , 0 < t < ,
Lux t ≤ M ux t
and continuous up to t = 0. Then the set
x x < ux 0 = 0
has zero Lebesgue measure.
Proof. We may assume that ax t > 0 for every x t. The map x t →
x !x t is a diffeomorphism from the region x < , 0 < t < onto the
open set in the plane:
= z = x + iy x < !x 0 < y < !x
Denote by z → x .x y the inverse of this diffeomorphism and set
vx y = ux .x y x + iy ∈ (III.8)
By the chain rule, we have
& &
& v v &
& & −1
& x + i y x y& ≤ K!t x .x y vx y (III.9)
138 Sussmann’s orbits and unique continuation
Since
!t x t ≥ %tk % > 0
we get
k/k+1
y − !x 0
!t x .x y ≥ %.x y = %
k
Ax .x y
Consequently, (III.9) implies for x + iy ∈ :
& &
& v v &
& & −k/k+1
(III.9 )
& x + i y x y& ≤ K y − !x 0 vx y
Observe that since x t → x !x t is also a homeomorphism from x <
, 0 ≤ t < onto
Lemma III.5.14. Let D be the unit disk in the complex z-plane and let v ∈ CD
be not identically zero and satisfy
& &
& v &
& z& ≤ K1 − z − vz z ∈ D (III.10)
& z̄ &
Proof. The main step in the proof is to show the following property:
3
There is a solution S ∈ <1 C D of the equation
S 1 v
= in D satisfying (∗)
z̄ v z̄
2)
sup Srei d <
r<1 0
Let us show right away that (∗) implies the conclusion of Lemma III.5.14.
Write v = expS h with h ∈ D. There is p ∈ Z+ such that v/zp is continuous
in D and does not vanish at the origin. Moreover, (III.10) is satisfied when
v/zp is substituted for v. Summing up, this argument shows that there is
no loss of generality in assuming from the outset that v0 = 0. Applying
Jensen’s inequality to the holomorphic function h gives, if r < 1,
1 2) 1 2)
log v0 ≤ S0 − Srei d + log vrei d
2) 0 2) 0
and consequently (∗) implies
1 2)
log v0 ≤ C + log vrei d (III.11)
2) 0
where C > 0 is independent of r. A standard application of Fatou’s lemma in
(III.11) shows that log− vei ∈ L1 T, whence the sought conclusion.
We now proceed to the proof of (∗). To simplify the notation, we set
F = vz̄ /v. We observe that there is p > 1 such that F ∈ Lp D (indeed it
suffices to take 1 < p < 1/). We set
1 Fz
Sz = dx dy
D z−z
)
Then
S
=F
z̄
140 Sussmann’s orbits and unique continuation
moreover, since p > 1, it also follows (cf. [V], theorem 1.35) that
S
= 1F
z
where 1 denotes the singular integral operator
1 gz
1gz = − dx dy
D z − z
) 2
Since 1 is a bounded linear operator in Lp D if 1 < p < (cf. [V], page
64) we obtain S ∈ Lp1 D.
Since F ∈ L z < R for R < 1, any solution of the equation u/z̄ =
3
F belongs to <1 C D. Hence (∗) will follow if we can establish the
following property:
2)
sup Srei d < (III.12)
1/2≤r<1 0
We observe that /r = ei /z + e−i /z̄, / = irei /z − e−i /z̄
from which we derive that r → Srei belongs to the Sobolev space
L11 1/4 1 × 0 2) . Thus r → Srei is absolutely continuous for almost all
. By first integrating on 1/2 r and afterwards on 0 2) we conclude that
2) 2) && 1 && 2) 1 && S &
&
Sre d ≤
i & S i &
e & d + & i &
r e & dr d
0 0
& 2 0 1/2
& r
for every r ∈ 1/2 1 , from which (III.12) follows. This completes the proof
of Lemma III.5.14.
In what follows, for any set S and a number k, k S will denote the
k-dimensional Hausdorff measure of S.
We can now prove:
(1) If u does not vanish identically on 0 < x < % t < % , then S ∩ x > 0
has a trivial one-dimensional Hausdorff measure (likewise for x < 0).
(2) If F0 = 0 × −% %, then S ∩ F0 = ∅ ⇒ F0 ⊂ S.
(3) If F0 = 0 0 and if u does not vanish identically then S has a trivial
one-dimensional Hausdorff measure.
Proof. Assume first that F0 = 0 × −% %. Then L = u/t over F0 (since
Zt 0 t = i!t 0 t = 0), which gives
& &
& u &
& 0 t& ≤ M u0 t
& t &
and thus
L0 u0 x s ≤ M u0 x s
showing that u0 is a C 1 -approximate solution for the structure defined by L0
in x < , s < 1. Moreover, u0 vanishes for s = 0. Therefore, by Corollary
III.6.2 and a standard propagation argument, u0 vanishes identically for x <
s < 1. Hence ux t0 = 0 for all x < .
Let be a real-analytic locally integrable structure over a connected, real-
analytic manifold of dimension N . When m = 1 ( has then dimension
n + 1) the orbits of the structure have either dimension n + 1 (open subsets
of ) or dimension n.
Introduce the projection over of the characteristic set of :
0 = p ∈ Tp ∩ Tp = 0
It is easy to see that 0 is an analytic subset of . Since is connected we
either have dim 0 ≤ n or 0 = .
Assume first that 0 = : in this case defines a real structure on in
the sense that = C ⊗ 0 , where 0 is an involutive vector sub-bundle of
T of rank n. The leaves of the foliation defined by 0 are precisely the
n-dimensional (Nagano) leaves.
Next suppose that the dimension of the analytic set 0 is ≤ n. On \0
the bundle defines an elliptic structure and every n-dimensional leaf is
contained in 0; in particular, it follows that the union of all n-dimensional
leaves is a set of n + 1-dimensional measure zero. We now prove:
0∗ = p ∈ 0 dim p = 1
U0 ∩ 0 = tn = 0
Notes
As indicated in the introduction, the concept of orbits and its basic prop-
erties were presented in Sussmann’s paper [Su]. Lemma III.1.8 and some
148 Sussmann’s orbits and unique continuation
149
150 Local solvability of vector fields
smooth. The sufficiency of condition for local solvability in any number
of variables is discussed in Section IV.3, while Section IV.4 is devoted to its
necessity.
Since our point of view is local, most of the time the behavior of L outside
a neighborhood of the point under study is irrelevant. This means that we
can modify the coefficients of L off that neighborhood in order to assume
that they are defined throughout R2 and we shall often do so. The sort of
properties of L we shall be interested in will not change by multiplication of
L by a nonvanishing factor. Since (IV.2) implies that either A or B does not
vanish in a neighborhood of a given point (assume as well that it is A), we
may multiply L by A−1 and obtain the new vector field L̃ = A−1 L which has
the form
u u
L̃u = + B̃x t (IV.3)
t x
Write B̃x t = ãx t + ib̃x t with ã and b̃ real, and assume that they are
defined for x < &, t < &.
Remark IV.1.3. Observe that Definition IV.1.2 means that given p there
exists a fixed open subset U p such that for every f ∈ C there exists
u ∈ such that the equation Lu = f holds on U . A moment’s reflection
shows that we would get an equivalent definition by requiring instead that
for every f ∈ Cc U there exists u ∈ U such that Lu = f in U . It is
less evident that we also get an equivalent definition if we require that for
every f ∈ C there exists u ∈ such that Lu − f vanishes on a
neighborhood Up f of p that may depend on both f and p. However,
a category argument shows that if this happens we may always take U
independent of f for fixed p and the apparently weaker requirement is in fact
equivalent to that given in Definition IV.1.2 (cf. Theorem VII.6.1).
L= + ibt
t x
and we wish to study the local solvability of L in a neighborhood of the
origin. In other words, we wish to find a distribution u such that Lu = f
where f ∈ Cc R2 is given. We shall perform a partial Fourier transform in
the variable x and denote by $u and $
f the transforms of u and f respectively,
so the transformed equation becomes
d$
u
− bt$u =$f where $ f t = e−ix fx t dx
dt R
Using a standard formula for the linear ODE with parameter , we find a
solution $
u
t t
$
u t = eBt−Bs$
f s ds where Bt = b d
T 0
sign of b0. Let us assume now that b0 = 0. If b does not change
sign in −T T for some T > 0 we already know how to proceed. What if
b changes sign in −T T? Well, suppose there is a point t0 ∈ −T T
such that b ≥ 0 for ∈ −T t0 and b ≤ 0 for ∈ t0 T . In this case,
t
we take T + = t0 and notice that s b d ≤ 0 both for t0 < s < t and for
t < s < t0 . It is easy to convince oneself that those are all the cases for which
a good choice of T + is possible. Indeed, if b0 < 0 and b1 > 0 for some
−T < 0 < 1 < T no choice of T + will work. We would be forced to take
t
T + > 1 to guarantee that s b d ≤ 0 for t < s, s t close to 1 , but this
t
would imply that s b d > 0 for t < s, s t close to 0 . In other words, we
must prevent that bt changes sign from minus to plus as t increases. The
analysis of the case < 0 and the choice of T − will tell us that we must as
well prevent that bt changes sign from plus to minus as t increases and both
conclusions imply together that bt cannot change sign at all.
We will not prove Theorem IV.1.6 here. The ‘if’ part of the theorem will
follow from Corollary IV.1.10 presented later in this section while the ‘only
if’ part will be discussed in Section IV.4 under the assumption that L is
locally integrable.
Using Parseval’s identity it is easy to derive that for fixed t < s < T the L2 R
norm of
x → 2)−1 eix+Bt−Bs$
f s d
0
Before embarking on the rather long proof of Theorem IV.1.9, let us state a
standard consequence that implies the local solvability of L.
uLq
R Lp R ≤ CT f Lq
R Lp R
Proof. We shall use the notation T = R×−T T. Let p, q be the conjugate
exponents, p = p /p − 1, q = q /q − 1. Take C and T0 as granted by
Theorem IV.1.9 and for some 0 < T < T0 consider the linear functional
t
LCc T t Lx t → t L = fx t x t dx dt
R2
The inequalities
of t L it will easily imply the estimate for t L since the contribution of the
bounded zero-order term ibx can be absorbed by taking T0 small enough. In
other words, it is enough to prove (IV.11) with L in the place of t L.
When dealing with the case bt we already saw the advantage of consid-
ering separately the cases > 0 and < 0 (microlocalization) and this corre-
sponds to writing 1 = H + H−, where H is the Heaviside function,
defined as H = 1 for > 0 and H = 0 for < 0. It will be convenient—
although not strictly necessary—to substitute this rough partition of unity by a
smooth one, so we consider a test function ( ∈ Cc −2 2 such that ( = 1
for ≤ 1 and set
+ 1 − ( if ≥ 0
* =
0 if ≤ 0
and
− 0 if ≥ 0
* =
1 − ( if ≤ 0
+
where we have used that P̃ is bounded in L R for 1 < p < because it
p
so integrating between −T and T with respect to t and taking the 1/qth power
we obtain
K + f Lq −TT Lp R ≤ CT f Lq −TT Lp R (IV.16)
IV.1 Planar vector fields 159
Next we have to see the effect of K + on L, ∈ Cc T = Cc R×−T T.
Observe that since x ±T ≡ 0 it follows that + ±T = P + ±T ≡ 0,
in particular + Tx = 0 for any ∈ R. Let us compute
t '
d + s d
K + t+ x t = eix+Bxt−Bxs ds
Tx R ds 2)
where
t
R+ + x t = b e−Bxs−Bxt Dx x+ x s ds (IV.19)
Tx
so (IV.20) will imply (IV.11) for + with L in the place of t L if the error
term R+ + Lq R Lp R can be absorbed. At this point we need
Proof. We have
−- D
b e x
x x = bx − byP- x − y y dy
+ Lq −TT Lp R ≤ CT L+ Lq −TT Lp R + CT + Lq −TT Lp R
+ Lq −TT Lp R ≤ CT LLq −TT Lp R + CT Lq −TT Lp R
IV.1 Planar vector fields 161
Lu = f (IV.22)
in the sense of distributions and set
fx s
vx s = ux s gx s =
s log2 s
Lemma IV.1.13. Let L, ux t, fx t, vx s, gx s be as above. Then,
vx s ∈ L , gx s ∈ L1loc and
1
v= g for − 1 < s < 1 (IV.23)
z 2
in the sense of distributions. In particular, if w is any solution of
w = g
z
in a neighborhood of the origin, w must be essentially bounded in a neigh-
borhood of the origin.
By the lemma, we will have our example if we show that for an appropriate
choice of f ∈ L , equation (IV.23) has a solution which is not locally bounded
in any neighborhood of the origin. We choose f so that F = f . is the
characteristic function ( of the sector K described in polar coordinates by
0 ≤ r ≤ 1/2, 0 ≤ ≤ )/4. Hence, g = ( ∈ L1c R2 and a solution wx s
of (IV.23) is obtained by convolution of g/2 with the standard fundamental
solution of the Cauchy–Riemann operator. Thus,
1 x − x 1
wx s = dx ds
2) K x − x 2 + s − s 2 s ln2 s
We see that for x s = 0 0, the integral above is given by
1 )/4 1/2 − cos 1 )/4 1
2
drd = d
2) 0 0 r sin log r sin 2) 0 sin log sin /2
= −
integral
operators. Let us choose a function ! ≥ 0 ∈ Cc −1/2 1/2 , with
!dz = 1. Write !- z = -−1 !z/-, z ∈ R, and set
M! fz = sup |!- ∗ fz|
0<-<
Then [S2]
H 1 R = f ∈ L1 R M! f ∈ L1 R
A space of distributions is called semilocal if it is invariant under multipli-
cation by test functions. The space H 1 R, is not: *u may not belong to
H 1 R for * ∈ Cc R and u ∈ H 1 R. A way around this is the definition
of the semilocal (or localizable) Hardy space—better suited for the study of
PDEs—h1 R ([G], [S2]) by means of the truncated maximal function
m! fz = sup |!- ∗ fz|
0<-≤1
f h1 ∼ inf j j , where the infimum is taken over all atomic representa-
tions. Another useful fact is that the atoms may be assumed to be smooth
functions. A simple consequence of the atomic decomposition is that h1 R is
stable under multiplication by Lipschitz functions bx: if a satisfies (1) with
I ≥ 1 it follows that axbx/bL also does. If I < 1 and the center of I is
x0 we may write axbx = bx0 ax + bx − bx0 ax = 1 x + 2 x.
Then 1 x/bL satisfies (1) and (2) (with the same I) while 2 x/K satis-
fies (1) for the interval I of center x0 and length 1, where K is the Lipschitz
constant of ax. It follows that f → bf is bounded with constant ≤ bL +K
in h1 R. A refinement of this argument shows that h1 R is stable under
multiplication by more general continuous functions including Hölder func-
tions, as we now describe. Let be a modulus of continuity, meaning that
0 −→ R+ is continuous, increasing, 0 = 0 and 2t ≤ Ct,
0 < t < 1. Consider the Banach space C R of bounded continuous functions
f R −→ C such that
fy − fx
f = sup <
x=y x − y
C
equipped with the norm f C = f L + f C . Note that C is only deter-
mined by the behavior of t for values of t close to 0. We will show in
Lemma A.1.1 in the Appendix that if the modulus of continuity t satisfies
1 h 1 −1
t dt ≤ C 1 + log 0 < h < 1 (IV.25)
h 0 h
then h1 R is stable under multiplication by functions ∈ C R. Note that the
modulus of continuity t = tr , 0 < r < 1, that defines the Hölder space C r ,
satisfies (IV.25).
Consider now a first-order linear differential operator in two variables
L= + ibx t + cx t x t ∈ R (IV.26)
t x
We assume that for some 0 < r < 1
Of course, (iii) means that the operator L given by (IV.26) satisfies condition
. We now introduce the space L1 Rt h1 Rx of measurable functions
ux t such that, for almost every t ∈ R, x → ux t ∈ h1 R and
u· th1 dt ≤ C <
R
The dual of the space L R h1 R is (canonically isomorphic to) the space
1
This is a technical proposition that does not have an immediate duality conse-
quence due to the fact that the norm on the left-hand side of the estimates is
a weaker norm than that on the right-hand side and it should be regarded as
an intermediate step towards a better estimate to be obtained later. The proof
of Proposition IV.1.14 is similar to that of Theorem IV.1.9; in particular, the
operators K and R referred to in (IV.27) were implicitly used in its proof, for
instance, K = K + + K − + K0 , R = R+ + R− + R0 with K + given by (IV.14),
R+ given by (IV.19) and so on. So the first step will be to prove the analogue
of (IV.27) for K + . This will follow from a slight modification of (IV.15). Let
us consider a restricted maximal function
g ⊥ x = sup Py ∗ gx
0<y<1
Notice that the sup is now taken for values of y between 0 and 1 instead
of 0 < y < as we did in (IV.15), but we keep the same notation g → g ⊥ .
Assuming without loss of generality that bx t has compact support and
IV.1 Planar vector fields 167
taking T small we may assume that Bx t − Bx s < 1 in formula (IV.14),
so we get
T
K + ux t ≤ ũ+ ⊥ x s ds (IV.30)
−T
Before we continue with the proof of the estimates, we state and prove
some lemmas. The first one deals with the nonlocal space H 1 .
Lemma IV.1.16. Let 0 < < , let P be the Poisson kernel in R2+ and let Q
be an integrable function satisfying Q x ≤ C/1 + x 2 as in the previous
lemma. There exists C > 0 such that
sup Py ∗ fx dx ≤ C f h1 R f ∈ h1 R
− 0<y<1
sup Qy ∗ fx dx ≤ C f h1 R f ∈ h1 R
− 0<y<1
Proof. The first inequality follows from the second one, as P satisfies the
hypothesis required for Q. To prove the second inequality we need only show
that there exists C > 0 such that
sup Qy ∗ a ≤ C
− 0<y<1
with
1 dP −2x2
Qy x = Qx/y Qx = x x =
y dx 1 + x2 2
IV.1 Planar vector fields 169
and
⎧
⎨ bx t − by t if y = x
y t =
x
x−y
⎩
bx x s if y = x
Using once more Lemma IV.1.16 we see that the norm in L1 − ×
−T T of the term Py ∗ bx u+ x t is dominated by
≤ 2L QL1
Let us tackle the case r ≤ 1/2 assuming initially that x0 = 0. The estimate
2r
sup Q- ∗ x ax dx ≤ CraL ≤ C
−2r 0<-<1
170 Local solvability of vector fields
is, as usual, easily obtained. Keeping in mind that aydy = 0 and writing
Since x > 2r and y < r imply that x − y ≥ x /2, using the decay of Q
and Q we see that
1 C C
sup Q x − y/- ≤ 2 ≤ 2
2
- y ≤r - +x 2 x
1 C- C
sup Qx − y/- ≤ 2 ≤
- y ≤r - + x2 x2
for x > 2r so
r r
Q- ∗ x ax ≤ C Q ay dy
x2 −r
r
≤ C Q 2
x
Thus,
1
sup Q- ∗ x ax dx ≤ C Q r dt ≤ C Q
x >2r 0<-<1 2r t2
Q- ∗ x ax = Q- ∗ 2
x−x02
ax − x0
so
sup Q- ∗ x ax dx = sup Q- ∗ 2
x2
ax dx ≤ C Q
0<-<1 0<-<1
IV.1 Planar vector fields 171
Returning to the estimate of the second term Qy ∗ x u+ x in the expres-
sion of b e−y Dx u+ we point out that Lemma IV.1.17 can indeed be applied
for any fixed t to x y = x y t, so using Lemma IV.1.17 and (IV.32) we
get
R+ uL1 −×−TT ≤ CT u+ L1 R h1 R (IV.33)
Notice that we cannot absorb the term L1 R h1 R by taking T small because
it involves a stronger norm than that of the left-hand side. Thus, we wish to
obtain a similar but sharper estimate in which the norm L1 R h1 R also
appears as well on the left-hand side. To achieve this we make use of the
4
2 defined by Hf
2 = 1 − (Hf ' , where H denotes
mollified Hilbert transform H
the usual Hilbert transform, ( ∈ Cc −2 2, = 1, for ≤ 1. Here the
2 which is a pseudo-differential operator of order zero, derives
usefulness of H,
mainly from the fact that it can be used to define an equivalent norm on
hp R without appealing to maximal functions, as granted by the following
estimates (cf. [G]):
2 h1 ≤ f h1 ≤ C2 f L1 + Hf
C1 Hf 2 L1 f ∈ h1 R
where ( ∈ Cc −2 2 with (y = 1 for y ≤ 1. Then there exists -0 such
that for all 0 < - ≤ -0 there exist constants C1 = C1 -, C2 = C2 - > 0 such
that
f h1 ≤ C1 f L1 + r1- Df L1 ≤ C2 f h1 (IV.35)
Proof. For each - > 0, r1- D is a pseudo-differential operator of order zero,
thus bounded in h1 , so
f L1 + r1- Df L1 ≤ f h1 + r1- Df h1 ≤ C2 -f h1
On the other hand, r2- Df L1 ≤ K2- L1 f L1 and K2- L1 → 0 as - → 0.
Therefore, there exists -0 > 0 such that K2- L1 ≤ 1/2C for 0 < - ≤ -0 . Thus
Remark IV.1.20. Notice that r1- D is given by convolution with a distribu-
tion supported in the interval −2/- 2/-, in particular if u ∈ −r r —
i.e., if u is distribution supported in the interval −r r —r1- Du is supported
in the interval −r − 2-−1 r + 2-−1 .
We are now able to prove a stronger estimate. We will show that there exist
constants C and T0 > 0 such that for any 0 < T ≤ T0 and ∈ Cc −a a ×
−T T,
L1 −TTh1 Rx ≤ CT LL1 −TTh1 Rx (IV.36)
IV.1 Planar vector fields 173
2t = 1 − (H·
H· t
where H is the Hilbert transform and ( ∈ Cc −2 2, ( = 1 for ≤ 1.
The symbol of H 2 is equal to h = * + − * − , where * + and * − are the
symbols of the operators P + and P − already used. We see that H 2 is a pseudo-
differential operator satisfying the hypotheses of Lemma IV.1.19 and we
may write it as a sum H 2=H 22- where H
21- + H 21- −a a → −a a
satisfies (IV.35), i.e.,
21- · tL1 −a a
· th1 Rx ≤ C· tL1 −aa + H (IV.37)
Theorem IV.1.21. Let the operator L given by (IV.26) satisfy (i), (ii) and
(iii) and let a > 0. Then there exist constants C > 0 and T0 > 0 such that
terms of the original variables from estimates obtained in the new variables by
the behavior of local Hardy norms under composition with diffeomorphisms.
For this reason we now deal with operators having C 2+r coefficients in
the principal part. Since we are dealing with mixed norms, the roles of t
and x cannot be interchanged and we must consider changes of variables
that preserve the privileged role of t. Consider a general first-order operator
defined in an open subset ⊂ R2 that contains the origin
u u
Lu = Ax t + Bx t + Cx t u
t x
with complex coefficients A B ∈ C 2+r , 0 < r < 1, C ∈ C . Assume
that the lines t = const. are noncharacteristic, which amounts to saying that
Ax t > 0, x t ∈ . Since the properties we are studying do not change
if L is multiplied by a nonvanishing function of class C 2+r , we may assume
without loss of generality that A ≡ 1, i.e.,
u u
Lu = + Bx t + Cx t u
t x
Write Bx t = ãx t + ib̃x t with ã and b̃ real. In convenient new local
coordinates = x t, s = t, the expression of L is
for all u ∈ Cc −a a × −T T , 0 < T ≤ T0 . Hence, for every function
f ∈ X = L Rt bmoRx there exists a function u ∈ X which solves Lu = f
in a neighborhood U of the origin, with norm
uX ≤ CT f X
IV.2 Solvability in C
In the last section we introduced the local solvability condition in Defi-
nition IV.1.5 assuming that the vector field L was in the special form
L= + ibx t (IV.41)
t x
with bx t real, smooth, and defined for all x t ∈ R2 . However, to require
that t → bx t does not change sign is not per se a coordinate-free definition
because we are demanding that a particular coefficient (namely, bx t does
not take opposite signs on sets of a special kind (namely, x × R). It order to
find more invariant ways to formulate condition it is convenient to find
larger sets on which bx t keeps its sign unchanged. Assume that L given
by (IV.41) satisfies . Then the sets
A+ = x ∈ R sup bx t > 0 and A− = x ∈ R inf bx t < 0
t t
are open and disjoint, and the complement of its union F = R\A+ ∪ A− is a
closed set with the property that bx t = 0 on F × R. Write A+ and A− in
terms of their connected components
A+ = a+ +
j bj A− = a− −
j bj
j j
If x ∈ a+ +
j bj there exists t ∈ R such that bx t > 0 so we see that bx t ≥ 0
+ +
on aj bj × R and similarly bx t ≤ 0 on a− −
j bj × R. There is an easy
IV.2 Solvability in C 177
a− −
j bj × R: they are the orbits of dimension two of the pair of vector fields
X = L Y = L . Indeed, ± j is a union of vertical lines, so invariant
under the flow of X, and it is also invariant under the flow of Y because Y
vanishes on its boundary, so if p ∈ ± j the orbit p of X Y through
±
p is contained in j . Now, p is an orbit of maximal dimension, thus
open and connected, and being invariant under the flow of X it is of the
form a b × R with a± ±
j ≤ a < b ≤ bj . Since a × R is contained in the
boundary of p, bx t must vanish identically on a × R so a % a± ±
j bj
± ± ±
and similarly b % aj bj , which proves that j = p. On the other hand,
the sets x × R, x ∈ F , are precisely the orbits of dimension one of X Y .
Since a+ + − −
j bj aj bj ∈ F we see that a two-dimensional orbit is bounded
by two one-dimensional orbits in case its orthogonal projection onto the x-
axis is a finite interval, by one one-dimensional orbit if its projection has
exactly one finite endpoint and, of course, the boundary is empty if the
projection is the whole real line. To give a coordinate-free formulation of the
fact that bx t does not change sign on two-dimensional orbits we look at
5 5
X ∧ Y ∈ C R2 2 TR2 . Since 2 TR2 has a global nonvanishing
section e1 ∧ e2 , X ∧ Y is a real multiple of e1 ∧ e2 and this gives a meaning to
the requirement that L ∧ L does not change sign on any two-dimensional
orbits of L L . Note that when L has the form (IV.41) we have seen that
this happens if and only if L satisfies .
Consider now a vector field defined in an open subset ⊂ R2
u u
Lu = Ax t + Bx t (IV.42)
t x
with complex coefficients A B ∈ C such that
Lemma IV.2.2. Let Zx t and L be given by (IV.43) and (IV.44) respectively.
Then, L satisfies in R2 if and only if R t → x t is monotone for
every x ∈ R.
Proof. We have
t
X= + t x Y =−
t 1 + x2 x 1 + x2 x
so
t x y
X ∧Y = ∧
1 + x x t
2
Note that X and Y are linearly dependent at a point if and only if t vanishes at
that point. Thus, the one-dimensional orbits of L are vertical lines x = constant
on which t vanishes identically. Since the two-dimensional orbits of L are
bounded by 0, 1 or 2 one-dimensional orbits we see that each two-dimensional
orbit j , j = 1 2 , is of the form aj bj × R. If L satisfies then t
does not assume opposite signs on j , say, t ≥ 0 on j so t → x t is
monotone increasing for all aj < x < bj . If x % aj bj for any j it follows
that the point of coordinates x 0 belongs to a one-dimensional orbit, so
t x t = 0, − < t < , and t → x t is constant. This shows that
t → x t is monotone for every x ∈ R. Conversely, assume that t → x t
is monotone for every x ∈ R and let aj bj × R be a two-dimensional orbit.
Given x0 ∈ aj bj we have that t → t x0 t has a consistent sign, say
t x0 t ≥ 0. We must show that t x t ≥ 0 for all aj < x < bj . Indeed,
if t x1 t < 0 for some x1 ∈ aj bj and t ∈ R, it is easy to see that there
exist an intermediate point x2 between x0 and x1 such that t x2 t = 0 for
all t ∈ R. Then x2 × R is a one-dimensional orbit and must be disjoint
of the two-dimensional orbit aj bj × R, a contradiction to the fact that
x2 ∈ aj bj .
From now on, we assume that L given by (IV.44) satisfies condition and
we wish to find a local solution Lu = f with u ∈ C when f ∈ C . We start
from estimate (IV.11) in Theorem IV.1.9, with L in the place of t L, q = p = 2.
There exists a T C > 0 such that, for every u ∈ Cc −a a × −T T,
M = Zx−1 x D = −L2 − M 2
where d denotes the Poisson bracket performed in all variables. Note that
Zt Z2 +
dx t = 2 − 2 + t 2 2
Zx Zx
so for large d ≤ C d and also dx t = 0 implies = = 0,
i.e., D is a uniformly elliptic second-order operator with smooth bounded
coefficients.
Consider a pseudo-differential operator Px t Dx Dt of order s and type
& = 1 0 with symbol px t , that is,
1 ix+t
Pux t = e px t $
u dd
2)2 −
The first term in the expansion of the symbol of the commutator L P is given
by −i p x t by a well-known formula from the calculus of pseudo-
differential operators. Thus, L P is a pseudo-differential operator with the
same order s. However, if px t = Fdx t with F holomorphic on
the range of d, it follows that
• Lu ∈ L2 R2 .
Let us now prove that for every f ∈ C R2 there is u ∈ C R2 such
that Lu = f in K. Denote by C K the quotient of C R2 by the subspace
of those functions which vanish on K to infinite order. This is a Fréchet
space and its dual may be identified with K, the distributions in R2
supported in K.
182 Local solvability of vector fields
for which L and −L are formal transposes of each other, i.e., #Lu v$ =
−#u Lv$, u v ∈ Cc R2 . This pairing can be extended to u ∈ C R2 and
v ∈ R2 and if v ∈ K the value of #Lu v$ only depends on the residue
class u of u ∈ C R2 in C K and u → #u v$ is clearly continuous.
Conversely, given a continuous linear functional on C K, the contin-
uous linear functional C R2 u → u is represented by a compactly
supported distribution v ∈ R2 such that u = #u v$, u ∈ C R2 . Since
#u v$ must vanish when u vanishes to infinite order on K we see that v is
supported in K. Furthermore, it is clear that v = 0 if = 0.
Consider the continuous linear map T C K −→ C K defined by
T u = Lu , where u denotes the residue class of u ∈ C R2 in C K.
Then the range of T is dense; in fact, if is a continuous linear functional
on C K such that # T u $ = 0, u ∈ C K, regarded as an element of
K, satisfies the equation L = 0 which implies that = 0. Thus, to
show that T is onto we need only show that the range of T is closed and by
the Banach closed range theorem for Fréchet spaces this will follow if we
prove that the range of the dual operator T is closed for the weak∗ topology.
However, C K is reflexive, a consequence of the reflexivity of C R2 , and
in this case it is enough to prove that the range of T is closed for the strong
topology (see, e.g., [T1], chapter 37). Let the sequence j = T j = −Lj ,
j ∈ K, converge to ∈ K. There exist s such that j ⊂ H s R2
and j H s ≤ C, j = 1 2 This implies that j ∈ H s R2 and by (IV.48)
j H s ≤ Csj H s ≤ C
Most of the results we have proved so far in this section are summed up
in the following:
Proof. Only the converse part has not been proved already, and we prove it
now. Assume that
u u
Lu = Ax t + Bx t
t x
with complex coefficients A B ∈ C such that
Ax t + Bx t > 0 x t ∈
is locally solvable in C . Given a point p ∈ , that we may as well assume to
be the origin, we wish to prove the existence of a smooth function Z, defined
in a neighborhood of the origin, such that LZ = 0 and dZ = 0. Set
Ax t Bx t
dx t = +
t x
and find u ∈ C such that Lu = d in a rectangle U centered at the origin.
Then the 1-form
= Bx te−uxt dt − Ax te−uxt dx
is closed, since
Be−u Ae−u
+ = e−u d − e−u Lu = 0 in U
x t
Furthermore, does not vanish. Since U is simply connected, there exists
Z ∈ C U such that dZ = . So dZ = 0 in U and also LZ = #L $ =
e−u #At + Bx Bdt − Adx$ = 0.
It is clear that the above definition is coordinate-free. We will now see that
it is invariant under multiplication by a nonvanishing factor.
Theorem IV.3.5. Let L given by (IV.49) satisfy (IV.50) and condition
in a neighborhood of the origin and fix 1 < p < . Then, there exist a
IV.3 Vector fields in several variables 187
neighborhood U of the origin and a constant C > 0 such that the following
a priori estimate holds for every ∈ Cc U:
Proof. The proof of this theorem requires six steps. Since Theorem IV.3.5
follows from Theorem IV.1.9 when n = 1, we will assume in the proof that
n ≥ 2.
The first step. Renaming coordinates if necessary we may assume that
A0 0 = 0. Then, dividing by A in a neighborhood of the origin and applying
Lemma IV.3.1 we put L in the form (IV.51). The new vector field thus
obtained still satisfies condition by its invariance under multiplication
by nonvanishing factors and change of coordinates. If is a test function
supported in a small neighborhood of the origin and ! is the diffeomorphism
induced by the change of variables, the Lp norm of and the Lp norm
of ! are comparable because the Jacobian determinant det! satisfies
c1 ≤ det! ≤ c2 in a neighborhood of the origin for some positive constants
c1 c2 . Note that the derivatives of order k of the coefficients bj , j = 1 n,
may be estimated in terms of bounds for the derivatives of order up to k + 1
of the original coefficients A B1 Bn , as one extra derivative is consumed
by the change of coordinates. Furthermore, by multiplying the coefficients bj ,
j = 1 n, by a non-negative cut-off function equal to 1 on a neighborhood
of the origin, we may assume that b1 bn ∈ Cc Rn+1 . Hence, it is enough
to prove the theorem when L is given by (IV.51) and its coefficients are
compactly supported, provided that we prove that the constant C in (IV.53)
depends only on p and the L norms of the derivatives of order at most one
of the coefficients of L.
The second step. We assume that L is given by (IV.51) and its coefficients
are compactly supported, then denote by bx " t the vector field in Rn given
n
by j=1 bj x t/xj . In view of Proposition IV.3.4 and its proof, the fact
that L verifies implies that there exists a unit vector field v"x defined
on Rn such that
" t = bx
bx " t v"x x ∈ Rn t ∈ R
" 0 t = 0 for all t. Set
Note that v"x0 may be defined arbitrarily if bx
, -
N = x ∈ Rn " t = 0 t < 1
bx (IV.54)
188 Local solvability of vector fields
and
" t
&x = sup bx x ∈ Rn
t <1
so that N is precisely the set where &x vanishes. From now on we use the
notations = Rn × −1 1 and T = Rn × −T T, 0 < T < 1.
The third step. We prove inequality (IV.55). The proof of (IV.55) is easy
because L( = (L = (t , so
t
(xx t = L(x s ds
−T
Hence,
t
(·· tLp Rn ≤ L(· sLp Rn ds
−T
≤ 2T1/p L1 − (Lp Rn+1
with p −1 + p−1 = 1. Raising both sides to the power p and integrating with
respect to t between −T and T we obtain (IV.55) with C = 2.
The fourth step. We introduce a partition of unity that reduces the proof of
inequality (IV.56) to the proof of local estimates for test functions. Note that
the function 1 − ( is not even continuous which, of course, is a source
IV.3 Vector fields in several variables 189
This shows that &x − &y ≤ ,x b " L x − y and interchanging x and y
"
we are led to &x − &y ≤ ,x bL x − y for all x y ∈ Rn . This implies
(IV.57).
Next, given x0 % N select t ≤ 1 such that &x0 = bx " 0 t > 0. Then
" t is positive and differentiable in a neighborhood of x0 , so
bx
& & & &
& b" & & , b" " && 2, b
" L
& & & x , b
, v"x0 = &,x x0 t& ≤ & + b" ⊗ x
&≤
& b" & & b b" 2 & &x
(1) Qj = = Rn \F ;
j
(2) the Qj ∈ are mutually disjoint;
(3) diam Qj ≤ inf fx ≤ sup fx ≤ 5 diam Qj .
Qj Qj
190 Local solvability of vector fields
Proof. Let 0 denote the family of cubes with side length one and vertices
with integral coordinates. For every integer k we define
k = 2−k Q Q ∈ 0
so the cubes in k form a mesh of cubes of side length 2−k and diameter
√ −k
n2 . Each cube ∈ k gives rise to 2n cubes ∈ k+1 by bisecting the sides.
Set for any integer k
√ √
k = x ∈ Rn 2 n2−k < fx ≤ 4 n2−k
Note that k ⊂ and = k k .
We now define
0 = Q ∈ k Q ∩ k = ∅
k
√ √
Let Q ∈ 0 ∩ k . There exists x ∈ Q such that 2 n2−k < fx ≤ 4 n2−k .
Given y ∈ Q we have
fx − y − x ≤ fy ≤ fx + y − x
√
so using that ≤ 1 and y − x ≤ n2−k = diam Q we get
diam Q ≤ inf fx ≤ sup fx ≤ 5 diam Q
Q Q
where we have used that Q1 satisfies (3) of Lemma IV.3.8 to estimate fx.
Now, (3) applied to Q2 gives diam Q2 ≤ supz∈Q2 fz ≤ 6 diam Q1 . Since
the quotient diam Q2 /diam Q1 is a power of 2, the latter estimate implies
that diam Q2 /diam Q1 ≤ 4.
The family that disjointly fills up with closed cubes gives rise to a cover
by open cubes that has the bounded intersection property. We fix 0 < - < 1/4
and for any Q ∈ denote by Q∗ the cube with the same center as Q but with
side dilated by the factor 1 + -. Let Q1 and Q2 ∈ do not touch. We claim
that Q∗1 and Q2 cannot intersect. Indeed, the union of Q1 with all the cubes
of that touch Q1 (among which Q2 is not) contains, by Proposition IV.3.9,
the cube 5/4Q1 whose interior contains Q∗1 . This shows that Q∗1 ∩ Q2 = ∅.
Consider now a point x ∈ and select Q ∈ such that x ∈ Q. If x ∈ Q∗j
for some Qj ∈ then Q ∩ Q∗j = ∅, which implies that Q and Qj touch.
Then Proposition IV.3.10 shows that x belongs to at most 12n cubes Q∗j . If
z ∈ Q∗ then fz ≥ inf Q f − - diam Q ≥ 3/4diam Q ≥ 3/5diam Q∗ .
Similarly, fz ≤ 5 diam Q + - diam Q ≤ 5 diam Q∗ . Thus, for every
Q ∈ we have
1
diam Q∗ ≤ inf∗ fx ≤ sup fx ≤ 5 diam Q∗ (IV.59)
2 Q Q∗
192 Local solvability of vector fields
This estimate implies that Q∗ ⊂ and since the interior Int Q∗ ⊃ Q we see
that Int Q∗ is an open cover of with the bounded intersection property.
Lemma IV.3.11. Let N ⊂ Rn be the closed set defined in (IV.54) and let
0 < ≤ 1, 1 < p < . There exists a covering of Rn \N by open cubes
with sides parallel to the coordinate axes Int Q∗j , j = 1 2 , such that
the intersection of 12n cubes of the family is always empty and for any
j = 1 2 we have the estimate:
1
diam Q∗j ≤ inf∗ &x ≤ sup &x ≤ 5 diam Q∗j (IV.60)
2 Qj Q∗j
Furthermore, there are functions j ∈ Cc Rn \N such that pj is a partition
of unity in Rn \N subordinated to the covering Int Q∗j and for a certain
constant C > 0,
C
,j L ≤ j = 1 2 (IV.61)
diam Q∗j
Proof. From now on we assume without loss of generality that ,x b " L ≤ 1.
We apply Lemma IV.3.8 with fx = &x so F = N. The hypotheses are
satisfied because the Lipschitz constant of &x is 1 by (IV.57) and the
complement of N is bounded so N = ∅. Thus we obtain the collection of
disjoint cubes Qj which, dilated by the factor 1 + -, yields the associated
collection Q∗j of cubes whose interiors cover Rn \N, have the bounded
intersection property, and satisfy (IV.59). This proves (IV.60). Fix a function
0 ≤ * ∈ Cc Rn supported in x < 1 + -/2 such that * p x is smooth
and *x = 1 if x ≤ 1/2 (such a function is easily constructed). If Qj ∈ ,
denote by xj its center and by j its side length. Then *j x = *x −xj /j ∈
Cc Int Q∗j and *j x = 1 on Qj . We have
,*L C
,*j L ≤ ≤ (IV.62)
j diam Q∗j
Note that . = j *jp is smooth and ≥ 1 in Rn \N. Let us estimate ,.x on
the support of *j . If x ∈ Q∗j and *k x = 0 for some k ∈ Z+ it follows that
Q∗j ∩ Q∗k = ∅. We know that Q∗j is contained in the union of Qj with those
cubes of which touch it and the same can be said about Qk . This implies
that there are cubes Qj and Qk in such that
(1) Qj touches Qj ;
(2) Qk touches Qk ;
(3) Qj ∩ Qk = ∅ so they either coincide or touch.
IV.3 Vector fields in several variables 193
Applying Proposition IV.3.9 three times we obtain that diam Qk ≥ 4−3
diam Qj and Proposition IV.3.10 tells us that there are less than N = 123n
integers k such that Q∗j ∩ Q∗k = ∅. This shows that at most N terms *kp x of
the infinite sum that defines .x are not zero if x ∈ supp *j . Thus, using
the analogue for *kp of (IV.62) we obtain
C 43 NC
sup ,.x ≤ sup ,*kp x ≤ ≤ (IV.63)
Q∗j k Q∗j k diam Qk diam Q∗j
∗
Since
1
,. −1/p x ≤ . −1−1/p L ,.x ≤ ,.x
p
because . ≥ 1, (IV.63) implies
C
sup ,. −1/p x ≤ (IV.64)
Q∗j diam Q∗j
Set
*j x
j x =
. 1/p x
Then, pj is a partition of unity in Rn \N with the required properties. Indeed,
to prove (IV.61) we use the Leibniz rule and invoke (IV.62) and (IV.64).
The fifth step. We prove estimate (IV.56) when x t is supported in Q∗j ×
−T T, Qj ∈ . Assume that is supported in Q∗k × −T T for a certain
cube ∈ ; the value of T < 1 will be chosen momentarily. Since we are
" L ≤ 1, (IV.58) yields
assuming that ,x b
2
, v"x ≤ for x % N
&x
This shows, in view of (IV.60), that , v"x ≤ 4/diam Q∗j on Q∗j . Further-
more, Rn \N is bounded so diam Q∗j ≤ C, j ∈ Z. Hence, v"x is approximately
constant on Q∗j if is small; this allows us to rectify its flow as follows. Since
v" is a unit vector, we may assume without loss of generality that at the center
√ √
xj of Q∗j we have v1 xj ≥ 1/ n. Then, v1 xj − v1 x ≤ 4 < 1/2 n
for fixed once for all, small but independent of j, and we may assume that
√
v1 x ≥ 1/2 n on Q∗k . Solving the differential equations
dxj vj x
= xj 0 = yj j = 2 n (IV.65)
dy1 v1 x
we obtain a change of variables on a neighborhood of Q∗k given by x1 = y1 ,
xj = xj y1 y2 yn , 1 < j ≤ n, where the right-hand side denotes the
194 Local solvability of vector fields
where the Lp norms are taken in the variables y1 t and the map y → xy
takes Q†j onto Q∗j . Integrating this estimate with respect to y we get
which may be regarded as estimate (IV.56) for ∈ Cc Q∗j × −T T.
The sixth step. We prove (IV.56) in general. Let ∈ T and set j = j
where j is the collection of functions described by Lemma IV.3.11. We
have
1 − (x x t p = j 1 − (xx t p
j
≤ CT 1 − (LpLp + CT Lj 1 − (pLp
j
where we have used the Leibniz rule and the fact that j pj = 1. The second
term on the right-hand side is dominated by CT 1 − (pLp . Indeed,
in view of the definition of &, (IV.60) and (IV.61). Hence, Lj x p ≤ C
and since Lj x p = 0 except for at most 12n values of j we also have
j Lj x ≤ C. Thus,
p
and the last term can be absorbed as soon as CT < 1/2. This proves (IV.56).
We have already seen in steps 1 and 2 that (IV.53) follows in general once
(IV.55) and (IV.56) are proved for L of the form (IV.51), so the proof
of Theorem IV.3.5 is now complete for L and we may also replace L by
−L + cx t in (IV.53) if cx t is any bounded function provided we shrink
the neighborhood U of the origin, in particular, we may replace L by the
transpose operator t L = −L − idivx b."
Let us assume now that we are dealing with a locally integrable vector field
L in an open set of Rn+1 that contains the origin. After an appropriate local
change of coordinates x t we may assume that there are functions Zj x t,
j = 1 n defined on a neighborhood of the origin of the form
holds for every u ∈ Cc Rn × −T T for some C > 0, as guaranteed by
the proof of Theorem IV.3.5. The estimate can be extended to any u ∈ L2c Rn
−T T × −T T such that Lu ∈ L2c Rn −T T × −T T by Friedrich’s
lemma. It follows that b% → b = (1 + ds/2 in the symbol space S10
s
and that
us ∼ BuL2 if B is the pseudo-differential with symbol b and u ∈ Hcs Rn ×
−T/2 T/2. Furthermore, L B has order s − 1 on R × −T/2 t/2. If
u ∈ Hcs−1 Rn × −T/2 T/2 is such that Lu ∈ Hcs Rn × −T/2 T/2 we
may apply (IV.70) to B% u. Letting % → 0 we obtain
if u ∈ Hcs−1 Rn × −T/2 T/2 is such that Lu ∈ Hcs R × −T/2 T/2 and to
the existence of local smooth solutions, as described in the proof of Theorem
IV.2.3. We leave details to the reader.
While the method to obtain smooth solutions starting from the existence of
L2 solutions is essentially the same independently of the number of variables,
the proof that smooth local solvability implies local integrability is rather
different if n = 1 or n > 2. In the proof of Theorem IV.2.3 it was shown that,
for n = 1, solving Lu = f for a specific f obtained from the coefficients of
L was enough to produce locally a smooth Z such that LZ = 0 and dZ = 0.
Nothing like this is available if n > 1 and we must proceed indirectly. Assume
that L given by (IV.51) is locally solvable in C and we wish to find n first
integrals with linearly independent differentials defined in a neighborhood of
a given point p that we may as well assume to be the origin. The first step
is to find a complete set of approximate first integrals, namely, n smooth
functions Zj# , j = 1 n, such that LZj# = fj vanishes to infinite order at
the origin—i.e., fj x = O x k , k = 1 2 —and dZ1# 0 dZn# 0 are
linearly independent. To find Zj# we solve first the noncharacteristic Cauchy
problem
LUj = 0
Uj x 0 = xj
in the sense of formal power series. The coefficients of the formal series
Uj corresponding to monomials that do not contain t are determined by the
initial condition Uj x 0, i.e., they are all zero with the exception of the
coefficient of xj which is 1. The coefficients of monomials of the form t x
are determined from LUj = 0 inductively on . Once the formal series Uj
has been found we take as Zj# any smooth function that has Uj as its Taylor
series at the origin (the existence of such a function is usually called Borel’s
lemma). By their very definition Z1# Zn# are approximate first integrals.
To obtain exact first integrals by correction of Z1# Zn# we must solve the
equations Luj = fj , j = 1 n, in a neighborhood of the origin and then
define Zj = Zj# − uj . Clearly, LZj = 0, so the problem is now to verify that
dZ1 0 dZn 0 are linearly independent. This will be guaranteed if we
can make sure that duj 0 is small. Let K be a ball centered at the origin
such that LC K = C K and let denote the subspace of C K of
the (equivalence classes of) functions h such that Lh = 0. Then L defines
a continuous linear map from C K/ onto C K which, by the open
IV.4 Necessary conditions for local solvability 199
mapping theorem for Fréchet spaces, has a continuous inverse. This means, in
particular, that given % > 0 there exists > 0 and m ∈ Z+ such that for every
f ∈ C K such that D f L K < for all ≤ k there exist u ∈ C K
such that Lu = f and duL K < %. Let (x t ∈ Cc Rn+1 be equal to 1 for
x 2 + t2 < 1 and set fj& x t = fj x t(&x &t. Since fj vanishes to infinite
order at the origin we see that, choosing & big enough, D fj& L < for
all ≤ k. Choose now uj such that Luj = fj& and duj L K < %. Since
fj& = fj for x 2 +t2 < 1/& we see that the functions Zj = Zj# −uj , j = 1 n
form a complete set of first integrals in a neighborhood of the origin if % is
taken small enough.
means that L and −L are each other’s formal transpose with respect to this
pairing. The formula is also valid by continuity if v ∈ provided that
we replace the integration by the standard duality between distributions and
test function, i.e.,
One of the basic tools in the study of necessary conditions for local solv-
ability is Hörmander’s lemma ([H6]), of which we give the following version.
200 Local solvability of vector fields
Lemma IV.4.1. Let L be as described above and suppose that for every
f ∈ Cc there exists u ∈ such that Lu = f . Then, for any compact
set K ⊂ there exist constants C > 0, M ∈ Z+ such that
& &
& &
& fv detZx dxdt& ≤ C Dxt
f L Dxt
LvL (IV.74)
≤M ≤M
Proof. Let K ⊂⊂ with nonempty interior be given and consider the bilinear
form (IV.73) restricted to pairs f v ∈ Cc K × Cc K. Endow the first
factor with the topology defined by the seminorms Dxt
f L —so it becomes
a Fréchet space—and the second factor with the countable family of semi-
norms Dxt
LvL . Our solvability hypothesis implies that the latter topology
is Hausdorff, indeed, if v ∈ Cc K is such that Lv = 0 we may choose for
any f ∈ Cc K a distribution u ∈ such that Lu = f , so we have
for any f ∈ Cc K, which implies that v = 0. For fixed v, the bilinear form
clearly depends continuously on f . The solvability hypothesis implies that
the dependence on v is also continuous for f fixed. Indeed, we may assume
that f = Lu for some u ∈ . Hence
fv detZx dxdt = #Lu f detZx $ = −#detZx u Lf $
in view of (IV.73), which shows the continuity with respect to f for fixed v.
A bilinear form defined on the product of a Fréchet space and a metrizable
space which is separately continuous is continuous in both variables. This
proves (IV.74).
The last lemma shows that in order to prove that L is not solvable it is
enough to violate the a priori inequality (IV.74). We now describe a method
to violate (IV.74) provided we find a solution h of the homogenous equation
Lh = 0 with certain geometric property. Let g ∈ C 0 be a real function
and K ⊂⊂ be compact. We say that g assumes a local minimum over K if
there exists a ∈ R and V open, K ⊂ V ⊂ such that
(1) g ≡ a on K;
(2) g > a on V \K.
Note that we may always replace the open set V with one of its open subsets
with compact closure that contains K. In this case, still denoting the new set
IV.4 Necessary conditions for local solvability 201
by V we have
inf g = a1 > a
V
Then, taking a < b < a1 we see that the set W = g < b ∩ V has compact
closure contained in V and g ≥ b > a on V \W .
The proof of the next lemma shows how (IV.74) may be violated.
(i) Lh = 0;
(ii) h assumes a local minimum over some K1 ⊂⊂ .
Next, choose * ∈ Cc V,0 ≤ * ≤ 1, such that * = 1 on K1 and hx t < -/2
on the support of *. Define
*x t
f& x t = e&hxt
detZx x t
Then
Dxt
f& L ≤ C&M e-&/2 (IV.76)
≤M
which together with (IV.75) and (IV.76) shows that (IV.74) cannot hold for
the pair f& v& ∈ Cc K × Cc K if & is large enough.
202 Local solvability of vector fields
and the difficulty is to show under assumption (IV.77) that for an appropriate
choice of ≤ 1 and x0 ≤ 1 our function h assumes a local minimum over
a compact set. Writing h in terms of its real and imaginary parts,
hx t x0 = ux0 x t + ivx0 x t
we obtain
ux0 x t = x − x0 2 − x t − x0 0 2
Thus, if we can find points x1 t1 , x0 t0 , x2 t2 in x0 such that t1 < t0 < t2
and
ux0 x1 t1 > ux0 x0 t0
ux0 x2 t2 > ux0 x0 t0
it follows that there is a compact set K ⊂ x0 such that ux0 x t assumes a
local minimum over K. To study the variation of ux0 along x0 we consider
the parameterization x0 s = xx0 s s and differentiate
ux0 xx0 s s
204 Local solvability of vector fields
Lemma IV.4.4. Assume that there exists x0 t0 ∈ and ∈ Rn such that
" 0 t0 · = 0;
(i) bx
"
(ii) bt x0 t0 · = 0.
Proof. By Lemma IV.4.2 we need only show that there exists a solution h
of Lh = 0 such that h assumes a local minimum at p = x0 t0 . Set Z =
Z1 Zn = Zx−1 x0 t0 Zx t − Zx0 t0 . Then LZj = 0, j = 1 n,
Z p = 0, Zx p = I. Then, the change of coordinates x = x − x0 , t = t − t0 ,
206 Local solvability of vector fields
shows that there is no loss of generality in assuming from the start that
x0 t0 = 0 0. Write !j x t = Zj x t − xj , so !j 0 0 = x !j 0 0 = 0,
j = 1 n. Set
n
Wx t = Zx t · = j Zj x t
j=1
!tt 0 0 · = 0
Set
where the dots indicate terms of order > 2. Thus, the Hessian of u at the origin
with respect to x t is positive definite and u has a strict local minimum at
the origin for small.
Case 2. b"j 0 0 = 0 for some 1 ≤ j ≤ n. After a linear change in the
x-variables we may assume that
⎧
⎨ b1 0 0 = 1
b 0 0 = 0 j = 2 n
⎩ j
= 0 2 n
Since (ii) implies that = 0 this case can only occur if n ≥ 2. Set
n
Wx t = iZx t · = i j Zj x t
j=2
IV.4 Necessary conditions for local solvability 207
In order to find a solution h of Lh = 0 with the property that its real part
assumes a local minimum over a compact set we need only worry about those
cases not covered by Lemma IV.4.4, i.e., we may always assume that
Let us assume that L does not satisfy condition in any cube centered on
the origin and let us try to produce the required homogeneous solution h. As
in the case of two variables we will look for solutions h = u + iv such that the
Hessian matrix uxx is everywhere positive definite and the critical points of
x → ux t are located on a certain curve so that when looking for a local
minimum of u we only need to direct our attention to the restriction u . Then,
assuming by contradiction that u is monotone on and that this happens for
all the functions u of this type, we must conclude that L is forced to satisfy
in some neighborhood of the origin. The first step is then to show the
abundance of solutions of this type, which is taken care of by the next lemma
that describes a family of solutions depending on two parameters, x0 ∈ B
and ∈ Rn . The general form of these solutions is based on the function h
introduced in case (i) of Lemma IV.4.4.
Lemma IV.4.6. If T and are small enough there exists a smooth function
h ∈ C × B × Rn ,
Proof. Set
k
hx t x0 = 1 + 2 1/2 Zx t − Zx0 02
j=1
B = vtx − vxx A
Note that the jth component of Vux is Vux j = utxj −uxx Aj = utxj −utxj = 0,
j = 1 n so ux is constant along the integral curves of V . A similar
computation shows that V −vx j = 0, j = 1 n so −vx is also constant
along the integral curves of V . It follows that V is tangent to the submanifold
of × B × Rn × Rn of dimension 2n + 1
0 = x t x0 ux x t x0 = 0 = vx x t x0
Since x0 0 x0 ∈ 0 by (ii) of Lemma IV.4.3 the partial derivative of
x t x0 → ux x t x0 − vx x t x0
with respect to x0 at 0 0 0 0 0 0 is the identity. Thus, 0 may be
parameterized by x t for x < 1 t < T1 , < 1 as the graph of a
smooth map
x t → x0 x t x t
with values in x0 < 2 × < 2 . We may assume, if and T are further
shrunken, that the image of x < 1 t < T1 , < 1 by the map
x t → x0 x t t x t
covers × B × < . Thus, the vector field
V∗ = − x t · + x t · (IV.81)
t x
where
x t = Ax t x0 x t x t
x t = Bx t x0 x t x t
210 Local solvability of vector fields
Note that the identity ux = b" ·vx is just the real part of the equation Lh = 0. This
reduced the problem of finding a homogeneous solution h = u + iv whose
real part assumes a local minimum over a compact set for an appropriate
choice of x0 to the problem of finding a curve ˜ x0 such that the function
" t· changes from negative to positive along ˜ x . Thus, from
qx t = bx 0
the fact that is not satisfied in any neighborhood of the origin—which
amounts to saying that any cube centered at the origin contains an integral
curve of X = t along which qx t changes sign—we must derive that
there exists an integral curve of V∗ along which qx t changes sign. The
tool to compare the changes of sign of a function along the integral curves of
two different vector fields is provided by
IV.4 Necessary conditions for local solvability 211
which may vary in the range − . The mean value inequality states that
if f ∈ C 0 a b there exists c ∈ a b such that fb − fa ≤ D+ fca − b.
If fa = fb it is enough to choose c ∈ a b so that fc = inf fx and
the general case is reduced to this one by subtracting the affine function
fa + x − afb − fa/b − a. It follows that if D+ fx ≤ 0, x ∈ a b,
then fx is monotone nonincreasing.
Let V be a Lipschitz vector field in U ⊂ RN , that is, Vx − Vy ≤
K x − y , x y ∈ U . We denote by !t x, the forward flow of V stemming
from x, i.e., the solution !t x defined in a maximal interval 0 ≤ t < Tx of
the ODE
d
! x = V!t x
dt t
!t 0 = x
Let F ⊂ U be a closed set. We say that F is positively V -invariant, or just
V -invariant for brevity, if
x ∈ F &⇒ !t x ∈ F for all t ∈ 0 Tx
The characterization of V -invariant sets given below is due to Brézis ([Br]).
The following properties are equivalent:
Proof. Let U − be the V∗ -flow out of the set x ∈ U qx < 0 , i.e., a point
x ∈ U − if x = !t y for some y ∈ U with qy < 0 and 0 ≤ t < Ty, where !t
is the flow of V∗ . Hence, U − is an open set and qx < 0 ⊂ U − ⊂ qx ≤ 0
because of •. By its very definition, U − is positively V∗ -invariant and so is
its closure F = U − . Indeed, if x ∈ F there exist a sequence xj ⊂ U − such
that xj → x. If 0 < t < Tx, then 0 < t < Txj for large j because s → Tx
is lower semicontinuous. Then !t xj ∈ U − by the V∗ -invariance of U − and
!t x = limj !t xj ∈ U − .
To prove the lemma we will show that F is X-invariant, which clearly
implies that X has property • because F ⊂ qx ≤ 0 . We must show that
dist x + -Xx F
lim = 0 x ∈ F (IV.82)
-(0 -
If qx < 0 this is trivially true, since x + -Xx ∈ F for small - > 0. If
qx = dqx = 0, (2) implies that Xx = V∗ x so
dist x + -Xx F dist x + -V∗ x F
=
- -
and the right-hand side → 0 as - ( 0 because F is V∗ -invariant.
If qx = 0 and dqx = 0, the set qy = 0 ∩ W is a C 1 manifold where
W is a convenient ball centered at x. It is easy to find a smooth unit vector
field Ny that meets qy = 0 transversally and points toward qy < 0 , so
Nq < 0 on W ∩qy = 0 . Let .t y denote the flow of the vector X = X +
N , ≥ 0. Then, (1) implies that X qy < 0 on W ∩qy = 0 for any > 0.
214 Local solvability of vector fields
Note that no integral curve of X , > 0, that stems from a point in W ∩qy <
0 can cross W ∩ qy = 0 (this would amount to traveling against the flow
at W ∩ qy = 0 ) and this implies that q.t x < 0 for > 0, t > 0
small, in particular .t x ∈ U − . Hence, .t x 0 = lim(0 .t x ∈ F
where the limit holds by the continuous dependence on the parameter .
Thus, the flow .t x 0 of X does not exit F for small values of t > 0, which
easily implies (IV.82), as in the proof of ‘(i) &⇒ (ii)’ of the characterization
of flow-invariant sets.
Notes
A few years after the publication of Hans Lewy’s example [L1], Hörmander
([H6], [H7]) shed new light on the nonsolvability phenomenon explaining it
in a novel way. Although his results are set in the framework of general order
operators of principal type we will describe its consequences for vector fields.
He proved that if a (nonvanishing) vector field L is locally solvable in then
the principal symbol of the commutator L L between L and its conjugate
must vanish at every zero of the principal symbol x of L. A vector field
with this property is said to satisfy condition . For the Lewy operator
condition is violated at every point. If the coefficients of L are real
or constant L L vanishes identically. This was a most remarkable advance
because it explained a phenomenon that had appeared as an isolated example
in terms of very general geometric properties of the symbol, an invariantly
defined object. However, it turns out that condition does not tell apart
the solvable vector fields from the nonsolvable ones among some examples
considered by Mizohata ([M]), which we now describe. Let k be a positive
integer and consider the vector field in R2 defined by
Mk = − iyk
y x
If k = 1 condition is violated at all points of the x-axis so, in particular,
M1 is not locally solvable at the origin. For k ≥ 2 condition is satisfied
everywhere. On the other hand, it follows from relatively simple arguments
that Mk is locally solvable at the origin if and only if k is even ([Gr], [Ga]).
The principal symbol of Mk is m1 = −i − iyk . The crucial difference
between k odd and k even is that in the first case the function yk changes
sign and in the second case it doesn’t. Nirenberg and Treves ([NT]) elabo-
rated these examples and identified a property that turned out to be the right
condition for local solvability of vector fields, i.e., condition . When L
Notes 215
with bj real-valued and Lipschitz and satisfies then it is locally solvable
in L2 . Since there is loss of one derivative in the process of obtaining coordi-
nates in which L has this form one must require, in general, that derivatives
up to order one of the coefficients of L be Lipschitz. However, in two vari-
ables (i.e., when n = 1) it is possible to prove L2 solvability directly without
assuming that L is in the special form (a) ([HM1]). Hence, planar vector
fields with Lipschitz coefficients that satisfy are locally solvable in L2 .
This result is essentially sharp in the sense that there are counterexamples to
L2 solvability and to the existence of L2 a priori estimates if the coefficients
are only restricted to belong to the Hölder class C for any 0 < < 1 ([J1],
[HM1], [HM2]). Whether any vector field with Lipschitz coefficients that
satisfies in three or more variables is locally solvable in L2 is an open
problem at the time of this writing.
It is a characteristic feature of locally solvable operators of order one that
the L2 a priori estimates that they satisfy can be extended to Lp estimates
for 1 < p < , a fact that turns out to be false for second-order operators
in three or more variables (for results in that direction see [Li], [K], [KT1],
[KT2], [Gu], [Ch1]). Solvability in Lp for vector fields was first considered in
[HP], where the method involved pseudo-differential operators and demanded
smooth coefficients. On the other hand, using the method of H. Smith ([Sm]),
Lp a priori estimates in the range 1 < p < can be proved in one stroke
under the same regularity hypothesis on the coefficients initially known to
guarantee just L2 estimates ([HM2]). This is the point of view used in the
presentation of a priori estimates in this book, although for simplicity we
have not included the proof that in two variables Lp estimates for vector
fields with Lipschitz coefficients are valid without assuming they are in
the canonical form (a) ([HM2]). The proof of a priori estimates in several
variables is reduced, thanks to the geometry of that prevents the existence
of orbits of dimension higher than 2, to two-dimensional a priori estimates
that are glued by a partition of unity associated with a convenient Whitney
decomposition in cubes. The presentation in this chapter owes much to the
discussion in [S1] about decomposition of open sets in cubes.
216 Local solvability of vector fields
While it is true that for any locally solvable vector field L and 1 < p <
the equation Lg = f can locally be solved in Lp if f is in Lp , this is false,
in general, for p = as we saw in the example after Remark IV.1.12 that
was taken from [HT2]. This difficulty can be dealt with by introducing the
space X = L Rt bmoRx of measurable functions ux t such that, for
almost every t ∈ R, x → ux t ∈ bmoR and ut ·bmo ≤ C < for a.e.
t ∈ R, where bmoR is a space of bounded mean oscillation functions, dual
to the semilocal Hardy space h1 R of Goldberg. This was first observed
in [BHS], where it is proved that for a substantial subclass of the class of
locally solvable vector fields L, the equation Lu = f can be locally solved
with u ∈ X if f ∈ L . This result was later improved by showing that for
any locally solvable vector field L the equation Lu = f can be locally solved
with u ∈ X for any f ∈ X ([daS], [HdaS]) which can be regarded as an ersatz
for p = of the Lp local solvability valid for 1 < p < . The presentation
in Section IV.1.2 follows closely [HdaS] but replaces lemma 4.5 of that
paper—which is true but incorrectly proved—by Lemma IV.1.17 which is
sharper.
A priori estimates in L2 easily give a priori estimates in L2s for any s ∈ R
but the absorption of lower-order terms requires shrinking of the neighborhood
in which the estimate holds in a way that makes its diameter tend to zero
when s → . Therefore, the technique of a priori estimates gives solutions
of arbitrary high but finite regularity for smooth right-hand sides. Using
a different approach, Hörmander ([H9]) proved solvability for differential
operators of arbitrary order that satisfy by studying the propagation
of singularities of the equation Pu = 0 mod C , showing the existence
of semiglobal solutions, i.e., solutions defined on a full compact set under
the geometric assumption that bicharacteristics do not get trapped in the
given compact set. Furthermore, the solutions can be taken smooth if f is
smooth. In Sections IV.2 and IV.3 of this chapter, the construction of smooth
solutions is simplified by the assumption that the vector fields are locally
integrable. Since vector fields that satisfy are indeed locally integrable,
the local integrability hypothesis is superfluous, however this fact depends on
the difficult and long theorems on smooth solvability by Hörmander ([H9],
[H5]). Thus, it would be interesting to have a shorter ad hoc proof of the
local existence of smooth solutions for vector fields that satisfy without
invoking local integrability.
Concerning the necessity of , Nirenberg and Treves had shown in their
seminal paper [NT] that local solvability implies for vector fields with
real-analytic coefficients and conjectured the same implication should hold
for smooth coefficients. This state of affairs remained unchanged for 15 years
Notes 217
until Moyer ([Mo]) removed in 1978 the analyticity hypothesis for operators
in two variables in a never published manuscript. His ideas, however, were
applied by Hörmander [H4] to extend the result for operators in any number
of variables with smooth coefficients. The discussion of the necessity of
in Section IV.4 of this chapter is again simplified by the assumption of local
integrability and follows the presentation in [T3] (see also [T5] and [CorH2]).
V
The FBI transform and some applications
218
V.1 Certain submanifolds of hypoanalytic manifolds 219
and
" #
Ul Vl
J dF0 =J +
xj l xj ul l xj vl
Ul Vl
= −
l xj vl l xj ul
equality also holds in the same fashion for y . Thus, J is independent of the
j
holomorphic coordinates. This also means that J can be defined on the real
tangent space of any complex manifold. Note that J extends to a C-linear
map from CTp CN into itself and the extension still satisfies J 2 = −Id. We
will also denote this extension by J . The fact that J 2 = −Id implies that
220 The FBI transform and some applications
J CTp CN −→ CTp CN has only two eigenvalues: i and −i. Define Tp10 CN
to be equal to the eigenspace associated with i, and Tp01 CN will be the
eigenspace associated with −i. We get corresponding vector bundles T 10 and
T 01 . Observe that T 10 is generated by z z . Hence T 10 is the bundle
1 N
of holomorphic vector fields introduced in Chapter I (see the discussion
preceding Theorem I.5.1). Likewise, T 01 is generated by z z .
1 N
The spaces Tpc and p are related. To see this, we recall the following
result from [BER] where p denotes the real parts of elements of
p :
Proof. Observe first that for any x ∈ Tp CN x +iJx ∈ Tp01 CN . Let x ∈ Tpc .
Then x and Jx ∈ Tp and so x + iJx ∈ p . Thus x ∈ p .
Conversely, if x ∈ p , then there is y ∈ Tp CN such that x+iy ∈ p ⊆
CTp implying that x ∈ Tpc since y = Jx and y ∈ Tp . We have thus
proved (a) and (b) follows from (a) trivially. The proof of (c) is also contained
in that of (a).
From Proposition V.1.6 we see that
dim Tpc = 2 dimC p
Example V.1.18. Let z w denote the coordinates of Cn+d where z = x+iy ∈
Cn and w = s + it ∈ Cd . Let 1 z s d z s be smooth, real-valued
functions. Then
= z w tj − j z s = 0 1 ≤ j ≤ d
= T0 + JT0
T0 = T0c ⊕ Y
that the map A commutes with J and hence after a complex linear map (see
Remark V.1.20 below) we are in coordinates z w p ∈ Cn+r+l where
= span of 1 ≤ j ≤ n 1 ≤ k ≤ r
xj yj sk tk
and T0 = span of x y s . It follows that near 0, can be expressed
j j k
as a graph of the form:
Proof. Let p ∈ and &1 &d be defining functions of near p. Since
p = 0 , we must have:
at the point p. Hence d ≥ N . If is also generic, then &1 &d are
linearly independent and so d = N .
The map J can be used to characterize CR, generic CR, and totally real
submanifolds.
Proof. (i) follows from the definition of Tpc and Proposition V.1.6. (ii) also
follows from Proposition V.1.6. To prove (iii), if is generic and &1 &d
are local defining functions, then the linear independence of &1 &d is
equivalent to:
dimC p = N − d
Hence, by Proposition V.1.6, dim Tp ∩ JTp = 2N − d. But then
dim Tp + JTp = 2N , implying that Tp + JTp = Tp CN for all
p ∈ . Conversely, if Tp + JTp = Tp CN , then dimTp ∩ JTp =
2N − d and so by Proposition V.1.6, dimC p = N − d showing that
is generic.
We will next describe certain submanifolds in hypoanalytic structures that
play important roles in the analysis of the solutions of the sections of the
associated vector bundle. Let be a hypoanalytic structure. is
a smooth manifold of dimension N and is an involutive sub-bundle of
CT of fiber dimension n whose orthogonal bundle T in CT ∗ is locally
generated by the differentials of m = N − n smooth functions. Recall from
0
Chapter I that T 0 = Tp denotes the characteristic set of the structure
p∈
.
→ ux
1
eix−y·− t−y −% 2 uy dtdyd
2 2 m
= 3 m lim
4) 2 %→0
1 2 m
= 3 m lim Fu t eix−t·−% 2 dtd
4) 2 %→0
The following characterization of analyticity by means of an exponential
decay of the FBI transform may be viewed as an analogue of the Paley–Wiener
Theorem.
(i) u is real-analytic at x0 ∈ Rm .
(ii) There exist a neighborhood V of x0 in Rm and constants c1 , c2 > 0 such
that
Fu x ≤ c1 e−c2 for x ∈ V × Rm
228 The FBI transform and some applications
Proof. We will assume that u is continuous and leave the general case for
the reader.
i ⇒ ii Suppose u is real-analytic at x0 . Let 0 ≤ ≤ 1 ∈ C0 Rm ≡
1 near x0 , and supp ⊆ x u is analytic at x . The integrand in Fu x has
a holomorphic extension in a neighborhood of y = x0 in Cm . We will denote
by u the holomorphic extension of u near x0 . In the integration defining
Fu x , we deform the contour from Rm to the image of Rm under the map
y → y = y − isy where s is chosen small enough so that u is defined
on the image Rm . We then have
Fu x = eQxy uy det y dy (V.2)
Rm
≤ ce− 2
2
Hence, for x − x0 ≤
2
and any ∈ Rm ,
Fu x ≤ c1 e−c2 for some c1 c2 > 0
(ii)⇒(i) Assume without loss of generality that x0 = 0. Suppose then that for
some c1 c2 > 0,
Fu x ≤ c1 e−c2
for all ∈ Rm , and for all x near 0. We will use the inversion given by
Theorem V.2.2. Write
2 m
Fu t eix−t·−% 2 dtd = I1% x + I2% x + I3% x + I4% x
V.2 Microlocal analyticity and the FBI transform 229
t2 A2
− + 162
≤ ce
4
Using the latter we see that after integrating in t, the integrand in I3% is
uniformly bounded by a constant multiple of
A22
e− 16
This allows us to complexify as in I1% to conclude that I3% z converges
uniformly to a holomorphic function in a neighborhood of 0. Write
eix−y·− t−y −% 2 uy dydtd
2 2 m
I4% x =
R
where
R = y t ≥ B A1 ≤ t ≤ A2 y ∈ supp u
230 The FBI transform and some applications
Note that the function → has a holomorphic extension #$ in the region
< , where
" # 21
m
#$ = j2
j=1
and an appropriate branch of the square root is taken. We change the contour
in the integration from Rm to its image under the map = +is x −y
for s small, s > 0. The number s is chosen to be small enough to ensure that
for = 0 < . We then have, modulo entire functions that
converge uniformly to an entire function,
m
I4% x = ePxyt% #$ 2 uy dydtd
where
y0
(a) ym0 ≥ C1
≥ 2 y ≥ 2ym and
y0 y0
(b) ym0 − ym ≥ ym0 − ym ≥ C1
− y > 2C1
.
Using the growth condition on f and the fact that f is holomorphic, we can
integrate by parts and arrive at
CC
D hy ≤ for all where C = sup D
yk
Let = k. We will estimate D hy on +0 . Assume first that k ≥ 2. For
y ∈ +0 ,
& 1 &
& &
D hy − D hy = & DD hty + 1 − ty · y − y dt&&
0 & 0 0
0
" #
1 1
≤ C C dt
=k+1 0 ty + 1 − ty0 k
" #
1 1
≤ C C dt
0 tym + 1 − tym
0 k
=k+1
" #
1 1 1
≤ C C 0 −
=k+1
ym − ym ymk−1 ym0 k−1
232 The FBI transform and some applications
" #
1
≤ C C k−1
=k+1
y
We have used (a) and (b) and the fact that since + is convex, ty +1−ty0 ∈ +.
Thus there is C > 0 such that for all , = k,
" #
1
D hy ≤ C
C whenever y ∈ +0
=k+1
y k−1
Note that this inequality also holds when k = 1. Fix yk−2 ∈ +k−2 . Let k−1 =
yk−2
2C1
. Using the preceding inequality, for y ∈ +k−1 , we can easily get:
" #
Dhy ≤ C C log y for some C > 0
=k+1
Let now y y ∈ +k−1 . We have:
& 1 &
& &
hy − hy ≤ && Dhty + 1 − ty · y − y dt&&
0
" #& &
& 1 &
≤ C C & & log ty + 1 − ty dt&& y − y
=k+1 0
" #" #
1 1
≤ C C 1 dt y − y
=k+1 0 tym + 1 − tym 2
" #" #
ym + ym
≤ C C √ +
=k+1 ym + ym
" #
≤ C C y + y
=k+1
with C independent of .
Remark V.2.7. We note here that when m = 1, the theorem above says that if
a holomorphic function f defined on a rectangle Q = −a a×0 b satisfies
V.2 Microlocal analyticity and the FBI transform 233
the growth condition fx + iy ≤ yck , then the traces f + iy converge in
−a a to a distribution of order k + 1.
Proof. Let u ∈ Rm . There exist an integer N and a constant c > 0 such
that the Fourier transform $
u satisfies the estimate $
u ≤ c1 + N . Let
j 1 ≤ j ≤ k be open, acute cones such that
k
Rm = j
j=1
+j = v ∈ Rm v · > 0 ∀ ∈ j
1
= $
u$
−d
2)m j
Hence
k
#u $ = #bfj $
j=1
Example V.2.10. Let f1 x y = 1
x+iy
for y > 0 and f2 x y = − x+iy
1
for y < 0.
Then it is not hard to show that
−2)i0 = bf1 + bf2
Granted this, since u ∗ 0 = u for any u ∈ R, we get an explicit decompo-
sition of u as a sum of two distributions each of which is the boundary value
of a tempered holomorphic function on a half-plane.
Observe that from Definition V.2.13 it can easily be shown that the analytic
wave front set is invariant under an analytic diffeomorphism, and hence
microlocal analyticity can be defined on any real-analytic manifold. The
following theorem provides a very useful criterion for microlocal analyticity
in terms of the FBI transform:
V.2 Microlocal analyticity and the FBI transform 235
The proof uses the inversion formula of Theorem V.2.2 and ideas similar to
those in the proof of Theorem V.2.4 (see also Theorem V.3.7). The reader is
referred to [Sj1] for the proof of this theorem.
It was shown in [BT5] that this system is analytic hypoelliptic at 0, i.e., every
solution u of Lk u = 0 k = 1 n is analytic at 0 if and only if, for every
∈ Rm , the function
t → t ·
does not have a local minimum at 0. This result was proved using the FBI
transform. The authors also proved a microlocal version of this result.
236 The FBI transform and some applications
' Ck
u ≤ on +
1 + k
Definition V.3.3. The C wave front set of a distribution u denoted WFu
is defined by
WFu = x u is not microlocally smooth at x
=k
Remark V.3.6. Lemma V.5.1 in Section V.5 shows that each smooth function
of one real variable has an almost analytic extension. Such extensions also
exist in higher dimensions (see [GG]).
The following theorem characterizes microlocal smoothness in terms of
almost analytic extendability in certain wedges.
Theorem V.3.7. Let u ∈ Rm . Then x0 0 % WFu if and only if there
exist a neighborhood V of x0 , open acute cones + 1 + N in Rm \0 , and
almost analytic functions fj on V + i+j (for some > 0) of tempered growth
such that u = Nj bfj near x0 and 0 · + j < 0 for all j.
Proof. Suppose x0 0 % WFu. Let ∈ C0 Rm , ≡ 1 near x0 such that
'
u decays rapidly in a conic neighborhood of 0 . By the Fourier inversion
formula,
1 ix· '
u = e u d
2)m
where the formula is understood in the duality sense, that is, for * ∈ C0 Rm ,
1 ix·
'
#u *$ = e *x dx u d
2)m
Let j 1 ≤ j ≤ N be open, acute cones such that
N
Rm = j
j=1
Definition V.4.1. For + an open convex cone with vertex at 0 ∈ Rd , the set
U h + = z ∈ U hz ∈ +
is called a wedge with edge . The wedge is said to be centered at p and
to point in the direction of +.
The reader is referred to [BER] for the proof of this proposition. We mention
that if az is a d × d smooth invertible matrix satisfying g = ah near p, then
the matrix B = ap −1 .
for any smooth function * of sufficiently small compact support near the
origin in R2n+d .
Proof. The proof will use arguments similar to those used in the proof of
Theorem V.2.6. For *x y s smooth, supported near the origin, set
hv = fz s + iz s + iv*x y sdxdyds
R2n+d
for v ∈ + , v < %. We will estimate the derivatives of h. We have
h f
v = i z s + iz s + iv*x y sdxdyds
vj R 2n+d wj
for each j = 1 d. Observe that since
d d
f k
fz s + iz s + iv = z s + iz s + iv km + i
dsm k=1 wk sm
and the matrix I + is is invertible near the origin, there are smooth functions
ajm z s such that for each k = 1 d,
f d
d
z s + iz s + iv = akm z s fz s + iz s + iv
wk m=1 ds m
It follows that
d
h d
v = fz s + iz s + ivajm z s*x y sdxdyds
vj m=1 R
2n+d ds
m
We can thus integrate by parts and iterate the procedure to conclude that for
some constant C > 0 and every multi-index ,
CC
D hv ≤
vk
where C = sup D * . It then follows, as in the proof of Theorem V.2.6,
that hv has a limit as + v → 0. Set #u *$ = limv→0 hv. Note that u
242 The FBI transform and some applications
The hypoanalytic wave front set for solutions in structures of positive codi-
mension is defined by restriction to a maximally real submanifold as follows
([BCT]). Let be a hypoanalytic structure and u a distribution solution
near p ∈ . Select a maximally real submanifold through p. We recall
that the restriction u is well-defined and by Proposition V.1.29 inherits
a hypoanalytic structure of codimension 0. Hence the hypoanalytic wave
front set WFha u is defined and lives in T ∗ \0 . Since is maximally
real, by Propositions V.1.27 and V.1.29, the inclusion i → induces
an injection i∗ T 0 → T ∗ . We will say a covector ∈ Tp0 \0 is in the
hypoanalytic wave front set of u if i∗ ∈ WFha u . This set will be denoted
by WFhap u This definition is independent of the choice of the maximally
real submanifold through p (see [BCT] for the proof) and thus for any
such , we have a bijection i∗ WFhap u → WFhap u , where WFhap
denotes the hypoanalytic wave front set at p.
We will next recall the FBI transform of [BCT] which gives a very useful
Fourier transform criterion for microlocal hypoanalyticity. X is a hypoanalytic
structure of codimension 0 as above. If p ∈ X, by the results in Chapter I (see
V.5 Application to the C wave front set 243
Zj = xj + ij x 1 ≤ j ≤ m
where z ∈ C w
m 2
= w12 +· · ·+wm2 ,
and for any ∈ Cm with < #$ =
2 21
1 + · · · + m (the principal branch of the square root).
2
F 2 u z ≤ c1 e−c2 ∀z ∈ V ∀ ∈ +
Here U is a neighborhood of 0 in Rm .
ut = fx t u ux
Chemin’s ([Che]) result that the C wave front set of any C 2 solution is
contained in the characteristic set of the linearized operator
m
f
Lu = − x t u ux
t j=1 j xj
Proof. The conditions that u has to satisfy determine the Taylor coefficients
of the formal series
ux t = uj x tj
j=0
j
where uj x = t ux0
j!
. Set u0 x = fx . For each j, since we want
Lu = Ot , we must have tj−1 Lux 0 = 0. This then leads to
j+1
( N
1 1 up q a
uj x = − x qk x 0
j p+q=j−1 q! k=1 xk t
)
up
M
q b
+ x qk x 0
k=1 k t
for j ≥ 1. Note that the functions uj x are C and holomorphic in . Let
( ∈ C0 R be such that ( ≥ 0 ( ≡ 1 in − 21 21 and supp ( ⊂ −1 1 Then
there exists a sequence Rj > 1 Rj / + such that the series
ux t = (Rj tuj x tj
j=1
V.5 Application to the C wave front set 245
m
L= + aj x t
t j=1 xj
LZx t = Otk k = 1 2
and that
Remark V.5.3. If L is C , then Lemma V.5.1 insures that the Zj exist and
the proof below will show that in this case, we only need to assume that
f ∈ C 0 T Rm .
Mj Zk = kj Mj Mk = 0
where each cjs = Otk k = 1 2 Indeed, the latter can be seen by expressing
Mj L in terms of the basis L M1 Mm and applying both sides to the
m + 1 functions t Z1 Zm . For any C 1 function g, observe that the
differential
" #
m m
dg = Mk gdZk + Lg − Mk gLZk dt (V.6)
k=1 k=1
For ∈ Rm s ∈ Rm , let
Es x t = i · s − Zx t − s − Zx t 2
where for w ∈ Cm , we write w 2 = m 2
j=1 wj . Let B denote a small ball centered
at 0 in R and ∈ C0 B, ≡ 1 near the origin. We will apply (V.7) to the
m
function
gs x t = xfx teEsxt
where s are parameters. We get:
m
dgdZ = Lf + fLE − Mk f + fMk ELZk eE dt ∧ dZ
k=1
(V.8)
where dZ = dZ1 ∧ · · · ∧ dZm . Next by Stokes’ theorem we have, for t1 > 0
small:
t1
gs x 0dx = gs x t1 dx Zx t1 + dgdZ (V.9)
B B 0 B
for some constants c C > 0. In the latter, we have used the constancy of
near 0. It follows that the integral
t1
LfeE dt ∧ dZ
B 0
decays rapidly in . The term fLEeE is estimated using the fact that for
any k, LE ≤ ck tk for some constant ck and that eE ≤ e−c1 t . This shows
l
that
t1
fLEeE dt ∧ dZ
B 0
decays rapidly in . The integrals of the terms fMk ELZk eE and Mk f
LZk eE are estimated in the same fashion. Thus
t1
dgdZ
B 0
248 The FBI transform and some applications
and for each j ≥ 1 there exists a vector vj satisfying vj · +j > 0 and vj · 0 < 0.
We now write
n
I3 % = Kj %
j=0
V.5 Application to the C wave front set 249
where Kj equals the integral over +j . The decay in the FBI established in
(V.12) shows us that K0 is a smooth function even after setting % = 0. Each
of the remaining functions Kj , after setting % = 0, is a boundary value of a
tempered holomorphic function in a wedge whose inner product with 0 is
negative. Hence
0 0 % WFa Kj 0+
where WFa denotes the analytic wave front set. By Corollary V.3.8, the latter
implies that
0 0 % WFKj 0+
We have thus proved that
0 0 % WFfx 0
So the condition
· . 0 0 0 = 0
for some ∈ 0 2) is equivalent to saying that 0 0 is not in the
characteristic set of L. Suppose now hx t is a C 1 function with the following
property: there exist C 1 functions h x t s such that h x t 0 = hx t and
L h is s-flat at s = 0. If 0 0 0 0 is not in the characteristic set of L, we
know that there is ∈ 0 2) such that
0
· . 0 0 0 = 0
0
By replacing by + ) or − ) if necessary, we may assume that
0
· . 0 0 0 < 0
0
and we can apply what we saw in the proof of Lemma V.5.2 to an FBI in
x t-space to conclude the following: there exist a conic neighborhood + of
0 0 in Rm+1 \0 and a neighborhood of the origin in Rm+1 such that
Fh 0 x t
#x −x$2 +t −t2
= ei ·x −x+t −t − h x t 0dxdt
B×J
= Fhx t
is rapidly decreasing for ∈ + and x t ∈ . We have thus proved:
Lemma V.5.4. For each ∈ 0 2) let L = s − e−i L and suppose there
exist .1 .m+1
∈ C 1 × J such that Z = x t + s. x t s is an
approximate solution of L Z = 0 in the sense that L Z is s-flat at s = 0.
Suppose moreover that there exist h ∈ C 1 × J such that h x t 0 =
hx t and L h is s-flat at s = 0. Then
WFh 0 ⊂ charL 0
Consider
m
f
= − x t 0 (V.15)
t j=1 j xj
and
m
f
Lu = − x t u ux
t j=1 j xj
Lu v = gx t v (V.16)
where
m
f
g0 x t 0 = fx t 0 − j x t 0
j=1 j
and
f
gi x t 0 = fxi x t 0 − i x t 0 i = 1 m
0
Consider now the principal part of the holomorphic Hamiltonian of (V.16)
m
H= + g0 + gj
0 j=1 j
and
& &
& &
& 3̃z t 0 & ≤ Ck z k
& z &
j
Z̃ 3̃
z t 0
z 0
then Az t 0 ≤ Ck z k for all k. It follows that for each k
& &
& Q & & &k
& 0 & & &
& Z̃ t 3̃& ≤ Ck &PZ̃ t 3̃& ∀j = 1 m
& Z̃ &
j
Hence Q0
Z̃x t 0 0 3̃x t 0 is t-flat at t = 0, which in turn
Z̃j
implies that for all k ∈ N, there exists Ck > 0 such that
H.x t 0 ≤ Ck t k
Hence Lu . v = v . v = H.v is t-flat at t = 0, and so we have found
hx t = . v x t such that Lu h is t-flat at t = 0 and hx 0 = ux 0. Now
ux t is also a solution of the equation
us = e−i ut − fx t u ux
which is of the same kind as (V.14), and the associated vector field as in
(V.15) is given by
= − e−i
s
with as before. Note that
v
= − e−i v = − e−i Lu = Lu
s s
It follows that there exists a C 1 function h x t s such that
−i u
L h =
u
− e L h
s
254 The FBI transform and some applications
is s-flat at s = 0 and h x t 0 = ux t. We apply Lemma V.5.4 and conclude
that WFu 0 ⊂ char Lu 0 . By translation we may apply the same argument
to all points of and state
x1 xn y1 yn s1 sd
vanishing at p and smooth, real-valued 1 d defined near the origin such
that the differentials of
zj = xj + iyj j = 1 n
⊥
is an open cone in p ∩ Tp , it follows that ∈ p ∩ Tp . Therefore
the corollary follows from the fact that
⊥
i∗ Tp0 = p ∩ Tp
We will next present a result on the hypoanalytic wave front set of the trace
of a solution when the vector field in question is locally integrable.
We consider a smooth vector field L = X + iY where X and Y are real
vector fields defined in a neighborhood U of the origin. Let 0 be an embedded
hypersurface through the origin in U dividing the set U into two regions, U +
and U − , where U + denotes the region toward which X is pointing. We assume
that L is noncharacteristic on 0, which means (after multiplying L by i if
necessary) that X is noncharacteristic. Our considerations will be local and
so after an appropriate choice of local coordinates x t and multiplication of
L by a nonvanishing factor, the vector field is given by
m
L= + aj x t (V.17)
t j=1 xj
Theorem V.6.9. Let L = t + m
j=1 aj x t xj be locally integrable. Suppose
f ∈ U+ has a boundary value at t = 0 and
Lfx t = 0 x t ∈ U +
Assume that there is a sequence pk ∈ + , pk → 0 such that for each k =
1 2 , Xpk and Ypk are linearly independent. Then there exists a unit
vector v such that
0 ∈ Rn v · 0 > 0 &⇒ 0 0 % WFha bf
In particular, the hypoanalytic wave front set of bf at the origin is contained
in a closed half-space.
Mj Zk = kj Mj Mk = 0
It is readily checked that for each j = 1 m,
Mj L = 0 (V.18)
For any C 1 function g, the differential may be expressed as
m
dg = Lg dt + Mk g dZk (V.19)
k=1
(1) !0 tk = 0;
(2) !0 t ≤ !0 tk for 0 ≤ t ≤ tk ;
(3) lim !0 tk / !0 tk = −v.
tk →0
We will postpone the proof of Lemma V.6.10 and continue our reasoning
with v given by (3) in Lemma V.6.10. The assumptions on ! allow us to
write
!x t = !0 t + ex t ex t ≤ A xt + B x 2 (V.23)
for some positive constants A and B. Suppose first !t 0 0 = 0, which is
the case that is needed for Theorem V.6.4. Then there is < 0 such that
!t 0 0 = v. Since !0 0 = 0 and !x 0 0 = 0, we can write
!x t · 0 = !t 0 0 · 0 + O x 2 + t2
= v · 0 + O x 2 + t2
262 The FBI transform and some applications
Hence given 2 > 0, we can find t1 r and > 0 such that (i) and (ii) above hold.
We may therefore assume that !t 0 0 = 0 and so the quotient !0 t /t2 ≤
C for 0 t ∈ U + . We have !0 tk + !0 tk v = o !0 tk . We recall
that by hypothesis 0 · v > 0. Hence,
for tk small and 0 < c < 1. We now take r = !0 tk /tk , with and tk
small to be chosen later. Hence, for x ≤ r and 0 ≤ t ≤ tk , we can choose
small enough (depending on A, B and C but not on tk ) so that
t !0 tk
ex t ≤ A !0 tk + B2 !0 tk
tk tk2
(V.24)
!0 tk
≤c
2
This implies that on the support of x we have
c
−1 + c !0 tk ≤ !x tk · 0 ≤ − !0 tk
2
Let 2 = %/ !0 tk . A consequence of (V.23), (V.24) and the fact that
!0 t ≤ !0 tk for 0 ≤ t ≤ tk is
!x t ≤ 1 + c !0 tk
!x t 2 ≤ 1 + c2 !0 tk 2 (V.25)
2 !x t ≤ %1 + c !0 tk
2 2
L = ax y + bx y
x y
is said to be locally integrable in an open set if each p ∈ is contained
in a neighborhood which admits a smooth function Z with the properties that
LZ = 0 and the differential dZ = 0.
with jk x x̃ t real and smooth. The proof of theorem 1.1 in [BH1] shows
that for any ∈ C −a a we have
T
# $ = fx T!k x T ds + fx t Lt !k x t dxdt (V.27)
0 B0a
k
tj
where !k x t = j x t 0 x t = x (V.28)
j=0 j!
V.7 Application to the F. and M. Riesz theorem 267
% n
and j x t = − j−1 x t − x − x̃ j x x̃ tj−1 x t
t s=1 xs
Notice that A x x̃ t ≤ C, for x ∈ B0 a, x̃ ∈ F , t ∈ 0 T, ≤ k, and
= 1 n because the coefficients of L have uniformly bounded derivatives
on B0 a. Hence, we obtain from (V.29) and (V.30) the estimate
& &
& &
& fx T!k x T dx& ≤ C dx F Dx x dx (V.31)
≤k+1 B0a
where dx F = inf x̃∈F x − x̃ . We next consider the second integral on the
right in (V.27). We will first show that for any j,
j+1 j
Lt !j = t (V.32)
j!
To see this, note first that (V.32) holds for j = 0 from the definition of 1 .
To proceed by induction, assume (V.32) for j ≤ m. Then
m+1 m+1
Lt !m+1 = Lt !m + Lt t
m + 1!
m+1 m m+1 m+1
= t + Lt t
m! m + 1!
Lt m+1 m+1
= t
m + 1!
m+2 m+1
= t
m + 1!
This proves (V.32). Next we observe that since the coefficients bj x t vanish
on F × 0 T , each j has the form
j x t = c x tDx x (V.33)
≤j
268 The FBI transform and some applications
c ≤ Cdx F
The form (V.33) is clearly valid for 0 = . Assume it is valid for j . Then
it will also be valid for j+1 since by definition, j+1 = Lt j . If we now
choose k = N + 1, (V.32) and (V.33) imply that
& T & T
& & k+1 x t k
& &
& 0 B0a fx t L ! x t dxdt& ≤ 0 B0a fx t
t k
t dxdt
k!
T
≤C k+1 x t dxdt (V.34)
0 B0a
≤C dx F Dx x dx
≤k+1 B0a
Thus the second integral on the right-hand side of (V.27) also satisfies an
estimate of the kind in (V.31). Consider now a compact subset K ⊂ F with
Lebesgue measure K = 0 and choose a sequence
such that (i) % x = 1 for all x ∈ K; (ii) % x = 0 if dx K > %; (iii)
Dx % x ≤ C %− . Note that % x converges pointwise to the characteristic
function of K as % → 0 while D % x → 0 pointwise if > 0. Let * ∈
C B0 a and use (V.31) and (V.34) with = % * keeping in mind the
trivial estimate dx F ≤ dx K. By the dominated convergence theorem,
# % *$ → * d
K
while
dx K Dx % xL1 ≤ % Dx % xL1 → 0
as % → 0 (when = 0 one uses the fact that K = 0). Thus, (V.31) and (V.34)
show that
* d = 0 * ∈ C B0 a
K
which implies that the same conclusion holds for any continuous function *
on K (first extend * to a compactly supported function on B0 a and then
approximate the extension by test functions). Thus the total variation K
of on K is zero and by the regularity of it follows that F = 0
whenever F ⊂ F is a Borel set with F = 0. This proves that F is absolutely
continuous with respect to Lebesgue measure.
V.7 Application to the F. and M. Riesz theorem 269
End of the proof of Theorem V.7.1. We may assume that the vector field
has the form
L = + ibx t
t x
where bx t is real and smooth on a neighborhood of U = B−a a ×
−T T of the origin in R2 . Since the trace bf is a measure, by the Radon–
Nikodym theorem, we may write
bf = g +
Notes
For a more detailed account of CR manifolds the reader is referred to
the books [Bog] and [BER]. The book [T5] contains a detailed discus-
sion of hypoanalytic manifolds. The characterization of microlocal analyt-
icity (Theorem V.2.14) was proved by Bony. Microlocal analyticity was
generalized to microlocal hypoanalyticity in the work [BCT]. Several mathe-
maticians have used the FBI transform to study the regularity of solutions in
involutive structures and higher-order partial differential equations. Some of
these applications can be found in the works [BCT], [BT3], [BRT], [Hi] and
[HaT], [Sj1], and [EG1]. Theorem V.5.5 was proved by Chemin [Che] by
using para-differential calculus. The main ideas for the proof presented here
are due to Hanges and Treves ([HaT]), who proved the analytic version of
Chemin’s result. Subsequently, Asano [A] used the techniques in [HaT] to
give a new proof of Chemin’s result. Most of the material in Section V.6 is
taken from a paper of Eastwood and Graham ([EG1]). Section V.7 is taken
from [BH1]. For a generalization of the F. and M. Riesz theorem to systems
of vector fields, we refer the reader to [BH7].
VI
Some boundary properties of solutions
271
272 Some boundary properties of solutions
(ii) If limy→0+ f· + iy exists in Dk −A A, then for any 0 < A < A, and
0 < B < B, there exists C such that
Because of the local equivalence of L1 and sup norms for solutions in the
elliptic (Cauchy–Riemann) case, the preceding theorem asserts that a holo-
morphic function f on Q has a trace at y = 0 if and only if for some integer
N > 0,
fx + iy yN dxdy <
Q
The preceding result suggests that for a locally integrable vector field, in
general, one should seek a growth condition that is weaker than a tempered
growth expressed in terms of dist x .
VI.1 Existence of a boundary value 273
1 1
uN x y ≤ =
x y N Zx y − Zx 0 N
Observe that may be chosen so that uN x y is not bounded by any power
of y as y → 0+ . In general, if L is locally integrable, Z is a first integral of
L near the origin and Lu = 0 in the region y > 0, then the growth condition
We wish to show that f has a boundary value at y = 0. Let * ∈ C0 −A A.
Fix 0 < T < B. For each integer m ≥ 0, choose *m x y ∈ C −A A ×
0 B such that
then it is easy to see that (i) and (ii) hold. Note that since f is holomorphic,
for any 0 < % < T , and g ∈ C01 −A A, integration by parts gives:
A A
fx + i%gx % dx = fx + iTgx Tdx
−A −A
A T
+ 2i fx + iygx y dxdy
−A %
hand side in the formula converges. This proves that fx + iy has a boundary
value at y = 0.
We will prove now the sufficiency of (VI.2) in a more general set-up.
Let L be a smooth, locally integrable vector field defined near the origin in
Rm+1 . In appropriate coordinates x t we may assume that L possesses m
smooth first integrals of the form Zj x t = Aj x t + iBj x t j = 1 m
defined on a neighborhood of the closure of the cylinder Q = Br 0 × −T T
where Br 0 is a ball in x space Rm and Zx 0 0 is invertible. Thus, after
multiplication by a nonvanishing factor, L may be written as
m
Zk
L= − M (VI.3)
t k=1 t k
where the Mk are the vector fields in x space satisfying Mk Zj = kj 1 ≤
k j ≤ m. The next theorem gives, in particular, a sufficient condition for the
existence of a boundary value of a continuous function f when f is a solution
of Lf = 0.
(i) Lf ∈ L1 Q+ ;
(ii) there exists N ∈ N such that
T
Zx t − Zx 0 N fx t dxdt <
0 Br 0
Proof. Note first that by taking complex, linear combinations of the Zj ’s,
we may assume that Zx 0 0 = Id, the identity matrix. This will not affect
hypothesis (ii) in the theorem. Let * ∈ C0 Br 0. For each integer k ≥ 0,
we will show that there exists *k x t ∈ C Br 0 × 0 T such that
Assuming for the moment that such a uk with these properties exists, we set
*k x t = uk Ax t Bx t
where
Ax t = A1 x t Am x t Bx t = B1 x t Bm x t
Then *k x 0 = *x so that (i) above holds. To check (ii), observe that from
the equations
LZj = LAj + iBj = 0 j = 1 m
we have
m
m
uk u u
L*k = LAj + k LBj = 2 LAj k (VI.4)
j=1 xj yj j=1 zj
It follows that
L*k x t ≤ C1 uk Ax t Bx t
≤ C2 distAx t + iBx t 0k
≤ C2 Zx t − Zx 0 k
Thus if uk satisfies (a) and (b), then *k x t will satisfy (i) and (ii). We will
next write a formula for the uk . Since the map x → Ax 0 is invertible, there
is a smooth map G = G1 Gm such that
Zx 0 = Bx 0 = GAx 0
This and some of what follows may require decreasing the neighborhood
around the origin. Note that since dB0 0 = 0, and dA0 0 = 0, dG0 0 = 0.
276 Some boundary properties of solutions
Let Vj be the vector fields satisfying Vj xs + iGs x = js 1 ≤ j s ≤ m. For
each k = 1 2 define
i
uk x y = ˜
V *xy − Gx
≤k
!
˜
where by definition, *x = *Ax 0−1 . Clearly, uk Zx 0 = *x. We
claim that for each j = 1 m,
u 1
2 k = ik ˜
V *x y − Gx (VI.5)
zj =k
! x j
In particular, the claim implies property (b) for uk . Indeed, after contracting
the neighborhood of the origin, we may assume that 0 = x + iGx . Since
dG0 0 = 0, it follows that
which proves the claim for k = 1. Assume next that (VI.5) holds for k − 1,
k ≥ 1. We can write
where
1
Ek x y = ik ˜
V *x y − Gx
=k
!
VI.1 Existence of a boundary value 277
Observe that
Ek 1 ˜
x y = i k
V * ˜ y − Gx
y − Gx + V *
xj =k
! xj xj
(VI.9)
and
Ek ˜
V *
x y = ik y − Gx (VI.10)
yj =k
! y j
Using the expression for xj
from (VI.6), (VI.8) can be written as
uk−1 k m
1 Gs
2 =i xVs V *˜ y − Gx
zj =k−1 s=1
! xj
1
+ ik−1 Vj V *˜ y − Gx (VI.11)
=k−1
!
which establishes property (b) for uk . Hence for each k we have *k which
satisfies (i) and (ii) and has the form
*k x t = ˜
P x t Dx *Ax tBx t − GAx t (VI.12)
≤k
We can clearly modify *n by dropping the tilde in its definition and use
(VI.14) to conclude:
lim fx %.N x % dZx % exists (VI.15)
%→0 Br 0
Let Px t = Bx t − GAx t. For gx t ∈ C Br 0 × −T T whose
x-support is contained in a fixed compact set independent of t, and n a
non-negative integer, define
Tn gx t = P x t Dx gx tPx t T0 gx t = gx t (VI.16)
≤n
for any g = gx t. To see this, for * = *x, we change variables y = Ax t
in (VI.15) to write
fx t.N x t dZx t = fHy t tQy t Dy *y dy
it follows that
lim fHy t tQy t Dy *y t dy exists
t→0
for any smooth *y t with a fixed compact support in y. Going back to the
x coordinates, we have shown that
lim fx tSN gx t dZx t exists (VI.18)
t→0 Br 0
VI.1 Existence of a boundary value 279
where by definition
Sn gx t = P x t Dx gAx t tPx t
≤n
for any smooth g = gx t. Observe that the integral in (VI.18) can be written
in the form
ux tgAx t t dx
where this latter integral denotes the action of a distribution u t on the
smooth function x → gAx t t. Now since x t → Ax t t is a diffeo-
morphism near the origin, any function *x t is of the form gAx t t
for some g = gx t. We can therefore use (VI.18) to conclude that for any
gx t,
lim fx tTN gx t dZx t exists (VI.19)
t→0 Br 0
which proves (VI.17). For *x t ∈ C Br 0 × −T T whose x-support is
contained in a fixed compact set and a given multi-index with = N , plug
gx t = *x tPx t = *x tBx t−GAx t in (VI.19). Note that
we may write
TN *P x t = *P + * e x tP + h x tP (VI.20)
=N >N
Indeed, this follows from applying the integration by parts formula (VI.13)
to the m-form fx th x tPx t dZ1 ∧ · · · ∧ dZm , using the hypotheses on
f , and the bound Px t ≤ Zx t − Zx 0 . From (VI.19) and (VI.21) we
conclude that
" #
lim fx t *P + * e x tP dZx t exists. (VI.22)
t→0 Br 0
=N
It follows that
lim fx t * P dZx t exists (VI.24)
t→0 Br 0
=N
whenever the functions * x t ∈ C Br 0 × −T T have their x-support
contained in a fixed compact set independent of t. We now return to a general
gx t ∈ C Br 0 × −T T with x-support contained in a fixed compact
set independent of t. From (VI.19) and (VI.24) we conclude that
lim fx tTN −1 gx t dZx t exists (VI.25)
t→0 Br 0
for any gx t ∈ C Br 0 × −T T with x-support contained in a fixed
compact set independent of t. We will prove by descending induction that for
any such gx t and 0 ≤ k ≤ N ,
lim fx tTk gx t dZx t exists,
t→0 Br 0
which for k = 0 and gx t = *x ∈ Cc Br 0 gives us the desired limit. To
proceed by induction, suppose 1 ≤ k ≤ N and assume that for any multi-index
with = k, the limits
lim fx tP x tgx t dZx t and
t→0 Br 0
(VI.26)
lim fx tTk−1 gx t dZx t
t→0 Br 0
both exist for any gx t ∈ C Br 0 × −T T with x-support contained in
a fixed compact set independent of t. We have already seen that (VI.26) is
true for k = N as follows from (VI.24) and (VI.25). Fix with = k − 1.
Plug gx t = *x tPx t in the limit on the right in (VI.26) and observe
that Tk−1 g may be written as
Tk−1 gx t = *P + * e x tP + h x tP (VI.27)
=k−1 ≥k
where the e and h are smooth, the x-supports of the h x t are contained
in a compact set that is independent of t, and all order derivatives of the e
VI.2 Pointwise convergence to the boundary value 281
for all with = k − 1 and *x t ∈ C Br 0 × −B B with x-support
contained in a fixed compact set independent of t. Hence, taking account of
(VI.26) and (VI.29) we conclude that
lim fx tTk−2 gx t dZx t exists. (VI.30)
t→0 Br 0
We have thus proved that (VI.26) holds for k − 1, completing the inductive
step. Therefore,
lim fx %*x dZx % exists (VI.31)
%→0 Br 0
Theorem VI.2.1. (Theorem 744 in [Zy].) Let C0 be any simply closed curve
passing through z = 1 situated, except for that point, totally inside the circle
z = 1, and tangent to the circle at that point. Let C be the curve C0 rotated
around z = 0 by the angle . There is a Blaschke product Bz which, for
almost all 0 , doesn’t tend to any limit as z → expi0 inside C0 .
This theorem shows us that for nonelliptic vector fields, we can’t expect
nontangential convergence. Indeed, by the theorem, if
Lk = − ik + 1tk k = 1 2 3
t x
then for each k, we can get a bounded solution fk = Fk x + itk+1 of Lk with
Fk holomorphic in a semidisk in the upper half-plane, bfk x = bFk x ∈
L1 −1 1, but each fk x t doesn’t converge nontangentially on a subset of
−1 1 of positive measure. It suffices to take Fk holomorphic and bounded
on the semidisk z z < 1 z > 0 such that on a set of full measure in
−1 1, Fk has no limit in certain appropriate regions. By considering the Lk
with k even, we see that nontangential convergence may fail even for vector
fields that are C and analytic hypoelliptic. Note that for each k, and for
almost all p ∈ −1 1, there is an open region +k p with p ∈ + k p such
that fk x t converges to bfk p in +k p. On the other hand, if we take the
real vector field t , and the solution ux t ≡ bux = (, the characteristic
function of a Cantor set C of positive measure in −1 1, the only sets of
approach for which ux t → bux x ∈ C, are the vertical segments. Thus
for a general locally integrable vector field, we cannot get approach sets for
convergence larger than curves. Suppose now L = X + iY is a smooth, locally
integrable vector field near the closure of a planar domain . Assume 0 ⊆
is a smooth curve that is noncharacteristic for L, f ∈ L1loc Lf = 0 and f
has a trace bf ∈ L1 0. Multiplying by i if necessary, we may assume that
X is not tangent to 0 anywhere and that it points toward . For each p ∈ 0,
let p be the integral curve of X through p and set p+ = p ∩ . We shall
classify the points of 0 into two types:
(I) A point p ∈ 0 is a type I point if the vector fields X and Y are linearly
dependent on an arc p+ s 0 < s < % for some % > 0.
(II) A point q ∈ 0 is a type II point if there is a sequence qk ∈ p+ converging
to q such that L is elliptic at each qk .
VI.2 Pointwise convergence to the boundary value 283
Proof. Since the problem is local, we may assume that we are in coordi-
nates x t where = −1 1 × 0 1, 0 = −1 1 × 0 , and Zx t =
x + ix t is a first integral of L with real, 0 0 = 0 and x 0 0 = 0.
Modulo a nonvanishing factor,
t
L= −i
t 1 + ix x
and so
t x −t
X= − Y=
t 1 + x2 x 1 + x2 x
Observe that L is elliptic, i.e., X and Y are linearly independent precisely
at the points where t = 0. Assume now that 0 ∈ 0 is a type II point. Then
t → 0 t can’t vanish on any interval 0 % % > 0. Indeed, otherwise, we
would conclude that L = X on 0 × 0 %—contradicting the hypothesis that
0 is a type II point. For > 0 small, define
mx = inf x t Mx = sup x t
0≤t≤ 0≤t≤
Then since m0 < M0, we may choose A > 0 so that mx < Mx for
x ≤ A. After decreasing A and , by the boundary version of the Baouendi–
Treves approximation theorem in Chapter II (Theorem II.4.12), there is a
sequence of entire functions Fk satisfying:
Set
A = = + i < A m < < M
We may assume that the sequence Fk converges uniformly on compact
subsets of A to a holomorphic function F and ux t = FZx t for
x t ∈ Z−1 A . Indeed, this is clearly true if ux t is continuous for t > 0.
In general, we can use the fact that we can express u as Qh where h is a
284 Some boundary properties of solutions
We have:
u% x t ≤ G∗ x ∈ L1 A2 A3
Let
Gx + ix t ≤ G∗ x and a.e. Gx + ix t → bux as t → 0
we conclude that
is not empty (otherwise, we would be placed under Case 1), Fz ≡ Gz on
1 .
Case 3: Suppose x t ≤ x 0 on A2 A3 × 0 T T > 0 A2 < A3 . We
may assume that there exists x0 ∈ A2 A3 and sj → 0 such that x0 sj <
x0 0. Indeed, otherwise, matters will reduce to Case 1. By theorem 3.1
286 Some boundary properties of solutions
We have thus shown that if p is a type II point, then there is an interval around
it such that a.e. in the interval, pointwise convergence holds as asserted.
Consider now a type I point x0 0. Then Zx0 t ≡ Zx0 0 for t in some
interval 0 % . This implies that Fk Zx0 t ≡ Fk Zx0 0 for t ∈ 0 % , and
so because of the a.e. convergence stated in (a) and (b), we conclude that for
almost every type I point x, ux t → bux as t → 0.
at y = 0. From the local version of the classical Hardy space (H p theory for
holomorphic functions in the unit disk, we have:
The main results of Section VI.4 will extend (i) and (ii) above to solutions
of complex vector fields that satisfy a one-sided solvability condition. In the
elliptic case, property (i) follows easily from part (ii) of Theorem VI.1.1. We
will show in Section VI.4 that in general, property (i) follows from property
(ii) above and a boundary solvability condition. When a vector field exhibits
property (ii), we will say that it has the H p property. To describe the class of
vector fields with the H p property, consider a curve 0 in such that \0
has two connected components, \0 = + ∪ − . It turns out that the local
solutions of the equation Lu = 0 on + possess the H p property at q ∈ 0 if
and only if there is a neighborhood U of q such that L satisfies the solvability
condition of Nirenberg and Treves ([NT]) on U ∩ + . This leads to a
one-sided version of that we denote by + (or − if + is replaced
by − ) to indicate the side where it holds. If holds at q, then both +
and − hold at q. However, + and − may hold at q ∈ 0 and yet
may not hold in a neighborhood of q. The Mizohata vector field provides an
example illustrating this. Write L = X + iY with X and Y real. Let ⊂ U be
5
a two-dimensional orbit of L in U and consider X ∧ Y ∈ C U 2 TU.
52
Since TU has a global nonvanishing section e1 ∧ e2 , X ∧ Y is a real
multiple of e1 ∧ e2 and this gives a meaning to the requirement that X ∧ Y
does not change sign on any two-dimensional orbit of X Y in U . Recall
from Chapter IV that the vector field L satisfies condition at p ∈ 0 if
there is a disk U ⊆ centered at p such that X ∧ Y does not change sign on
any two-dimensional orbit of L in U .
Lemma VI.3.4. Let Zx t and L be given by (VI.33) and (VI.34) respectively.
Then, L satisfies + at the origin if and only there exist T a > 0 such that
0 T t → x t is monotone for every x ∈ −a a.
We now recall from [BH6] the local equivalence between + and one-sided
solvability. More precisely,
Theorem VI.3.5. Let Zx t and L be given by (VI.33) and (VI.34) respec-
tively. The following properties are equivalent:
Before proving Theorem VI.4.1, we will need to recall some concepts and
results from the classical theory of Hardy spaces for bounded, simply connected
domains in the complex plane. Let D be a such a domain with rectifiable
boundary. There are several definitions of a Hardy space for such a domain
(see [L] and [Du]). For our purpose here, we need to recall two of the
definitions:
where the inf is taken over all sequences of rectifiable curves Cj in D tending
to D.
When is the unit disk, it is a classical fact that both definitions of Hardy
spaces agree ([Du]). By the Riemann mapping theorem, this is also true for
any bounded, simply connected domain with a smooth boundary. In the work
[L], it is shown that when 1 < p < , these spaces agree if is bounded,
simply connected with a Lipschitz boundary.
where Cu∗ denotes the nontangential maximal function of the Cauchy inte-
gral of u and c is a positive constant depending exclusively on the aperture
of the cone + p.
Let r = x − p . We have
u
2)iCux = d
−p >2r − p
u u
+ − d
−p >2r −x −p
u
+ d
−p <2r − x
r
≤ 2 + u d
j j+1
j=1 2 r< p− <2 r 2j r2
1 1
≤ 22 + j
u d
j=1 2 2j+1 r p− <2j+1 r
≤ cMup
Finally, in order to estimate I3 we observe that x ∈ + p and ∈ imply
1 1+
≤
−x r
Using the latter estimate we obtain:
1 +
I3 ≤ u d ≤ cMup
2)r p− <2r
Our next aim is to prove that E p = H p for a particular class of
domains that includes the domains Uk that will appear in the proof of
Theorem VI.4.1. We consider smooth regions U that are bounded by two
smooth curves C1 and C2 that cross each other at two points A and B where
294 Some boundary properties of solutions
they meet at angles 0 ≤ A B < ). If A B > 0 then U has a
Lipschitz boundary and by the result in [L] we know that E p U = H p U for
p > 1. Our methods will show that this equivalence still holds when the values
A = 0, B = 0, and p = 1 are allowed. By a conformal map argument
we may assume that
(1) A = 0 and B = 1;
(2) the part C1 in the boundary of U is given by 0 1 t → t;
(3) the part C2 in the boundary of U is given by 0 1 t → xt+iyt where
xt yt are smooth real functions such that x0 = y0 = y1 = 0,
x1 = 1.
whatever the value of > 0, q ∈ + z for all q ∈ Cj and gq ≤ g ∗ z for
any function g defined on U . Given g ∈ H p U we must show that
sup gz p dz ≤ M < (VI.39)
j Cj
We have
gq p dq = gj2 z p j2 z dz
Cj2 −1 C
j2 j
≤ g ∗ z p j2
z dz
−1 C
j2 j
≤C g ∗ z p dz (VI.40)
C2
integral representation
1 bf
fz = d z ∈ U
2)i U − z
is valid ([Du], theorem 10.4). Furthermore, f E p U f + Lp U . Next we
recall Lemma VI.4.8 that gives the estimate
The same technique leads to the inclusion E p U ⊂ H p U for p > 1 because
T∗ and M are bounded as well in Lp U for 1 < p < but the method breaks
down for p = 1. This case will be handled in the proof of Theorem VI.4.1
using the fact that if f ∈ E p U, 1 ≤ p < , f has a canonical factorization
f = FB where F has no zeros, and B ≤ 1. This is classical for the unit disk
296 Some boundary properties of solutions
We will consider three essential cases, in each of which we will show that the
assertions of the theorem are valid on a half-interval 0 a . Since the same
arguments also apply to the half-intervals −a 0 , the theorem will follow.
Case 1: Assume that M0 = m0 and Ma = ma for some a > 0. In this
case we will first assume that the solution f is smooth on Q. If Mx = mx
for every x ∈ 0 a , then L would be t in 0 a and fx t = fx 0 for all
t ∈ 0 B , which trivially leads to the inequality we seek on the half-interval
0 a . Hence we may assume that there is x ∈ 0 a for which mx < Mx.
Then the set Z0 a × 0 B has nonempty interior. Every component of
the interior of this set has the form
where is a component of the open set x ∈ 0 a Mx > mx . Let
either mx = x 0 and Mx = x B or mx = x B and Mx =
x 0 on k k . Without loss of generality, we may assume that mx =
x 0 and Mx = x B for every x ∈ k k . Let Uk = the interior of
Zk k × 0 B. Thus
+p = z ∈ Uk z − p ≤ 2 distz Uk
Define the maximal functions Fk∗ and T∗ Fk on Uk (except at the two cusps)
by
Fk∗ p = sup Fk
∈+p
and
& &
& 1 &
T∗ Fk z = sup && F d && z ∈ Uk
%>0 ∈U
k −z >% −z k
Recall the Hardy–Littlewood maximal function
1 +
MFk z = sup f d z = k + iMk k + iMk
I I
where the sup is taken over all subarcs I ⊆ Uk that contain z and I denotes
the arclength of I. Next, since each Uk is Ahlfors-regular, Lemma VI.4.8
gives the estimate
and x 0 < x t < x B. Let p = x + ix B and q = x + ix 0.
We claim that z ∈ +p ∪ +q . Indeed suppose first
x B − x t ≤ x t − x 0 (VI.44)
Then for any y:
1
x + ix t − y − iy B ≥ x − y + x t − y B
2
1
≥ x − y + x t − x B − x B − y B
2
1 1
≥ x t − x B since x ≤
2 2
1
= z−p (VI.45)
2
We also have:
1
x + ix t − y − iy 0 ≥ x − y + x t − y 0
2
1
≥ x t − x 0
2
1
≥ x B − x t by (VI.44)
2
1
= z−p (VI.46)
2
From (VI.45) and (VI.46) we see that if (VI.44) holds, then z ∈ +p . By a similar
reasoning, if (VI.44) does not hold, then z ∈ +q . We have thus shown that
Uk ⊆ +p (VI.47)
p∈Uk
Next fix x t ∈ k k × 0 B. If x + ix t ∈ Uk , i.e., if x 0 <
x t < x B, then by (VI.47),
Fk x + ix t ≤ Fk∗ x + ix 0 + Fk∗ x + ix B (VI.48)
On the other hand, if x t = x 0, then since x 0 < x B, there
exists t ≤ y < B such that x y = x 0 = x t and y is the maximum
such. Let ym → y ym > y. Then by (VI.48),
Fk x + ix ym ≤ Fk∗ x + ix 0 + Fk∗ x + ix B
Letting m → , we get
Fk x + ix t = Fk x + ix y
≤ Fk∗ x + ix 0 + Fk∗ x + ix B
VI.4 The H p property for vector fields 299
for any 0 < t < B. Finally, we use a refinement of the approximation theorem
as in Theorem II.4.12 to remove the smoothness of f .
Case 2: Assume that M0 = m0 and Mx > mx for every 0 < x ≤ A.
We will need to use the boundary version of the Baouendi–Treves approx-
imation formula. Let hx ∈ C0 −A A, hx ≡ 1 in a neighborhood of 0.
For > 0, define
E fx t = /)1/2 e− Zxt−Zx 0 fx 0hx Zx x 0 dx
2
R
300 Some boundary properties of solutions
and
e− Zxt−Zx t fx thx Zx x t dx
2
G fx t = /)1/2
R
R fx t → 0 in C −a a × 0 b
Let F z be the entire function satisfying F Zx t = E fx t. Let Ua =
the interior of Z0 a × 0 b. Recall that m0 = M0 but mx < Mx
for any 0 < x ≤ A. The domain Ua is also an Ahlfors-regular domain. There-
fore, we can apply the arguments in Case 1 to the smooth functions E f to
arrive at:
a
E fx t p dx ≤ C F fz p dz (VI.56)
0 Ua
Note that this time Ua has three pieces and so (VI.56) leads to:
a a a
E fx t p dx ≤ C E fx 0 p dx + E fx b p dx
0 0 0
b
+ p
E fa s s a s ds 0 < t < b (VI.57)
0
We thus need only compute the limit of the s integral in (VI.57). We will
show that for almost all a ,
b b
E fa s p s a s ds → fa s p s a s ds (VI.58)
0 0
We know that Mx > mx for every 0 < x ≤ A. We may also assume that
x t > x 0 for every x ∈ 0 A t ∈ 0 b . Indeed, otherwise, we will be
placed in the context of Case 1. The approximation theorem then implies that
for each x > 0, f is continuous at x t for t > 0 small. Since R fx t → 0
uniformly in 0 a × 0 b , (VI.58) will follow if we show that for almost all
a ,
b b
G fa s p s a s ds → fa s p s a s ds (VI.59)
0 0
Choose two numbers a1 a2 such that 0 < a1 < a < a2 ≤ A. By the approxima-
tion theorem, after decreasing b, since f is continuous at x t for t = tx > 0
small, there exists F continuous in Za1 a2 × 0 b, holomorphic in W =
the interior of Za1 a2 × 0 b such that FZx t = fx t. Observe that
p
Since F has an L boundary value, it is well known (see, for example, [Ro])
that F ∗ ∈ Lploc a1 a2 . Let * ∈ C0 a1 a2 * ≥ 0, *x ≡ 1 near a. Write
G fx t = G1 fx t + G2 fx t, where
G1 fx t = /)1/2 e− Zxt−Zx t *x fx thx Zx x t dx
2
and = G fx t − G1 fx t. Consider first G2 fx t for x near a.
G2 fx t
Observe that the integrand is zero for x near a and hence for x near a and
t ∈ 0 b ,
G2 fx t → 0 uniformly (VI.60)
In the integrand of G1 fx t, fx t can be replaced by FZx t = Fx +
ix t and hence we have:
2
G1 fx t ≤ C/)1/2 e− 2 x−x *x F ∗ x dx
1
(VI.61)
R
x2
x = ) −1/2 e− 2 , and x = 1/2 1/2 x, then (VI.61) says that
Pick a point x0 where this supremum is finite and where F ∗ x0 < . Then at
such a point, the functions G1 fx0 t are bounded on 0 b . Since pointwise,
it follows that
b b
G fx0 s p s x0 s ds → fx0 s p s x0 s ds (VI.63)
0 0
for almost all a . We can therefore let → in (VI.57) and conclude that
for almost all a:
a a a
fx t dx ≤ C
p
fx 0 p dx + fx b p dx
0 0 0
b
+ p
fa s s a s ds 0 < t < b (VI.65)
0
Case 3: Assume M0 > m0. Let a > 0 such that Mx > mx for every
x ∈ −a a. If Wa = Z−a a × 0 B, there is a function F holomorphic
on the interior of Wa such that fx y = FZx y. This time the boundary
of Wa has four pieces. One can then reason as in the previous case to get
the required estimate on the interval −a a. Finally, observe that estimates
on the interval of the form −a 0 are also valid under Cases 1 and 2. The
theorem for 1 < p < follows from these three cases.
We consider next the case when p = 1.
Assume we are in the situation of Case 1 where M0 = m0 and Ma =
ma for some a > 0. As before we assume first that fx t is smooth on
Q+ , Fk ∈ C Uk , holomorphic in Uk and fx y = Fk Zx y on k k ×
0 B . Since Uk is simply connected, by a classical result (see the corollary of
theorem 10.1 in [Du]), Fk has a factorization Fk = Gk Bk where each factor is
VI.4 The H p property for vector fields 303
Summing up over k and adding the contributions from the set S = 0 a\ ∪k
k k , we get:
a a a
fx t dx ≤ C fx 0 dx + fx B dx (VI.68)
0 0 0
Ma = ma for some a > 0. Assume first that fx t ∈ C Q and for k
fixed as before, let
Uk = x + iy k < x < k x 0 < y < x B
and fx y = Fk Zx y on k k × 0 B , Fk holomorphic on Uk and
continuous on the closure. We apply the maximum modulus principle to Fk
and use the constancy of f on the vertical segments x = k and x = k to
conclude that
fx y ≤ f 0 L 0a + f B L 0a ∀x y ∈ k k × 0 B
If S is the set as before with
" #
0 a = k k S
k
Next we consider Case 2 where M0 = m0 and Mx > mx for every
0 < x ≤ A. As before, let a b > 0 such that
E fx t → fx t a.e. in −a a × 0 b (VI.75)
Let Ua = Z0 a × 0 b and consider the holomorphic function F such
that F Zx t = E fx t. The maximum principle applied to F on Ua
leads to
E fx y ≤ E f 0 L 0a + E f b L 0a
We therefore only need to estimate the term E fa L 0b for which it
suffices to estimate G fa L 0b . Let 0 < a1 < a < a2 < A be as before,
F holomorphic such that
fx y = Fx + ix y on a1 a2 × 0 b
Since bF = bf ∈ L a1 a2 , by the generalized maximum principle applied
to F there exists M > 0 such that
Since f
t
= −ax t f
x
+ gx t, (VI.78) implies that for some constant C ,
a && f &2
&
& x t& dx ≤ C ∀t ∈ 0 T
&
−a t
&
Notes
The results of this chapter in the holomorphic case are classical. For a discus-
sion of the conditions that guarantee the existence of a boundary value we
refer to the books [BER] and [H2]. The basic theory of Hardy spaces for
bounded, simply connected domains in the complex plane is exposed in [Du]
(see also [Po]). The paper [L] and the references in it contain more recent
developments on the subject. The planar case of Theorem VI.1.3 as well
as the necessity in the real-analytic, planar situation was proved in [BH5].
Lemma VI.4.8 is taken from [L]. Theorem VI.4.1 and its corollaries appeared
in [BH6]. The work [HH] extends Theorem VI.4.1 to the case 0 < p < 1 for
vector fields with real-analytic coefficients.
VII
The differential complex associated with a
formally integrable structure
Notice that, according to Section I.4, we have N1 = N; notice also that
Nq has, for each q, the structure of a C -module. We then generalize
the concept of one-forms introduced in Section I.4 and call the elements
of the direct sum ⊕ q=0 Nq differential forms over . If ∈ Nq we
shall say that is a differential form of degree q (or q-form for short).
The exterior product between ∈ Nq and ∈ Nr is the q + r-form
∧ ∈ Nq+r defined by the formula
∧X1 Xq+r = sg X1 Xq Xq+1 Xq+r
AB
(VII.1)
where Xj ∈ X and the summation is over all partitions A B of 1
q + r with A = q, B = r and ∈ S q+r is such that 1 q = A,
q + 1 q + r = B. It is easy to see that (VII.1) defines indeed a q + r-
form, that the map
308
VII.2 The local representation of the exterior derivative 309
→ ∧
is C -bilinear, and that the operation so defined is associative. It follows
that ⊕
q=0 Nq has a structure of a graded C -algebra. We also remark
that
∧ = −1qr ∧ ∈ Nq ∈ Nr (VII.2)
The exterior differentiation operator is a C-linear map
d ⊕
q=0 Nq → ⊕q=0 Nq
and our claim follows immediately from Lemma I.4.1 applied to the one-form
X → X X2 Xq . In particular, we can restrict a q-form over to an
310 The differential complex
open set W ⊂ , that is, given ∈ Nq there is W ∈ Nq W which makes
the diagram
X × × X −→ C
↓ ↓
W
XW × × XW −→ C W
where fJ ∈ C D × .
Fix x0 ∈ D and set, for J = j1 jq ,
q
J x x0 = 4j ∧ ∧ dxj
−1r−1 xjr − xj0r dxj1 ∧ ∧ dx r q
r=1
G N•q D × → N•q−1 D ×
which are formulae that can be proved by direct computation, using (VII.6).
In particular we derive, if q ≥ 1,
dx Gf = f if dx f = 0
VII.5 Localization
If W ⊂ is open there is a well-defined complex homomorphism
Uq d −→ Uq W d
which is induced by restriction.
Let p ∈ and consider an open neighborhood W of p over which there
are defined m differential forms 1 m ∈ NW that span T W at every
point. After contracting W around p and a linear change on 1 m ,
we can obtain a coordinate system x1 xm t1 tn defined on W and
centered at p in such a way that
n
k = dxk − bjk x t dtj k = 1 m
j=1
with bjk ∈ C W. Next we introduce the linearly independent vector fields
over W
m
Lj = + bjk x t
tj k=1 xk
Since k Lj = 0 for all j = 1 n and k = 1 m it follows that
L1 Ln span W at each point.
Next the C W-module Nq W is spanned by the q-forms
J ∧ dtK J + K =q
and since
J > 0 &⇒ J ∧ dtK ∈ Nq W
VII.6 Germ solvability 313
it follows that Uq W can be identified with the submodule of Nq W spanned
by dtK K = q .
If f ∈ C W then it is plain that
n
df = Lj fdtj mod 1 m
j=1
since dtj Lj = jj . From this we obtain the representation of the operator
d under the preceding identification: if f = J fJ dtJ ∈ Uq W then
n
d f = Lj fJ dtj ∧ dtJ (VII.16)
J =q j=1
Remark VII.5.1. Since satisfies the Frobenius condition and since further-
more the vector fields Lj Lj do not involve any differentiation in the
t-variables it follows that Lj Lj = 0 for every j j = 1 n. Now it is
easily seen that this condition is equivalent to the fact that formula (VII.16)
defines a differential complex, i.e., that d d = 0.
are continuous. Now the fact that (VII.17) is exact at x0 means that
The same argument gives a version of Theorem VII.6.1 where the solutions
are now allowed to be distributions.
Theorem VII.6.2. Assume that for every f ∈ C C satisfying Qx Df =
0 in a neighborhood of x0 there is u ∈ C solving Px Du = f in a
neighborhood of x0 . Then the following holds:
in the place of E .
VII.7 -cohomology and local solvability 315
H q p = lim H q W q≥0 (VII.20)
W →p
1
We recall that for a sheaf of C-vector spaces U → FU over a topological space X, the direct
limit
lim FW
W →p
at p ∈ X is the space of all pairs W f, with W an open subset of X that contains p and
f ∈ FW, modulo the following equivalence relation: W1 f1 ∼ W2 f2 if there is an open
neighborhood W of p, W ⊂ W1 ∩ W2 , such that f1 W = f2 W .
316 The differential complex
m
k
Lj = −i x tMk j = 1 n (VII.24)
tj k=1 tj
We have dZ1 y t ∧ ∧ dZm y t ∧ dt = det Zy y tdy ∧ dt. Hence
, -
uy t det Zy y t dy dt = tuy t
"0 t j B0 ×"0 t j
a fact that follows from Stokes’ theorem in conjunction with the identity
, -
'j ∧ ∧ dtn
d tuy tdZ1 y t ∧ ∧ dZm y t ∧ dt1 ∧ ∧ dt
= −1m+j−1 Lj u y t dZ1 y t ∧ ∧ dZm y t ∧ dt
We now let
fx t = fJ x tdtJ ∈ Uq B0 × "0 (VII.27)
J =q
dt F z t = (VII.29)
m2
e− z−Zyt d 0 .y t ∧ fy tdet Zy y t dy
2
)
318 The differential complex
In (VII.29) the integral is over B0 \B, since . is identically equal to one over
B. On the other hand, the real part of the exponent equals −z y t, where
z y t = z − y 2 − z − y t 2
Now, thanks to (VII.23) we have
1
y t ≤ 0 t + y
2
and then
1 2
y t 2 ≤ 2 0 t 2 + y
2
Denote by b > 0 the radius of B and use the fact that 0 0 = 0: there is an
open ball " ⊂⊂ "1 , centered at the origin in Rn , such that
1
2 0 t 2 ≤ b2 t ∈ "
4
If y ∈ B0 \B and t ∈ " then we obtain
1 2 1
y − 2 0 t 2 ≥ b2
0 y t ≥
2 4
and consequently, by continuity we conclude that there are r > 0 and > 0
such that
y t ∈ B0 \B × " z < r &⇒ z y t ≥
We can state:
Lemma VII.8.2. Given ∈ Zm
+ , ∈ Z+ there is a constant C > 0 such that
n
in the topology of Uq B0 × "0 , we obtain from (VII.31) and (VII.32) the
following result:
in the topology of Uq B × ". This justifies us referring to the result stated
in Theorem VII.8.4 as the Approximate Poincaré Lemma for the differential
complex d .
Tp0 ∩ N ∗ 0p = 0 ∀p ∈ 0 (VII.34)
Tp ∩ CN ∗ 0p = 0 ∀p ∈ 0 (VII.35)
H q p • = lim H q U ∩ • (VII.38)
U →p
0 ∩ U = y1 yN yN = 0
• ∩ U = y1 yN yN > 0
Notice that
Zk
0 = kr (VII.40)
yr
in particular it follows from (VII.40) that (VII.39) defines a local diffeo-
morphism in a neighborhood of the origin. In the new variables x1 xm
t1 tn we have
Zk x t = xk + ik x t
where the functions k are smooth, real-valued and vanish at the origin.
Furthermore, we have
and consequently
k N
Zk y
= yx t x t
xs =1 y x s
322 The differential complex
For the proof we shall first establish some lemmas. We return to the local
coordinates and conclusions provided by Proposition VII.10.1. We call atten-
tion, in particular, to the properties of the vector fields L1 Ln as described
at the beginning of Section VII.11. Recall that L1 Ln−1 are tangent to 0
and so they have well-defined restrictions to 0 ∩ B0 × "0 :
L0j = Lj tn =0
We shall work in an open set of the form W0 = B0 × "0 × J0 ⊂ B0 × "0 ,
where "0 (resp. J0 ) is an open ball (resp. open interval) centered at the origin
in Rn−1 (resp. R).
Given a smooth function (or even a differential form) g on W0 the notation
g ∼0 0 will indicate that g vanishes to infinite order on 0 ∩ W0
where ∈ Cc R, s = 1 for s < 1, s = 0 for s > 2, and j is a
suitably chosen sequence of real numbers satisfying j < j+1 , j → .
VII.12 A Mayer–Vietoris argument 325
For the uniqueness it suffices to observe that u and v must a fortiori have
identical formal power series expansions in tn , whence the assertion.
Proof. We write
g= gI x t dtI
I =q I⊂1n−1
and apply Lemma VII.12.2 in order to solve, for each I, the approximate
Cauchy problems
L n GI ∼0 0
(VII.48)
GI tn =0 = gI
If we set
G= GI x t dtI ∈ Uq W0
I =q
Lemma VII.12.4. Let G ∈ Uq W0 satisfy '0 ∗ G = 0 and d G ∼0 0. Then:
(a) If q = 0 then G ∼0 0.
(b) If q ≥ 1 then there exists u ∈ Uq−1 W0 such that G − d u ∼0 0.
where u1 ∈ Uq−1 W0 and 1 ∈ Uq W0 do not involve dtn . Since '0 ∗ G = 0
we have GI tn =0 = 0 if n ∈ I. Consequently, all the coefficients of 1 vanish
when tn = 0 and then we can further write
G = d tn u1 + tn h1 (VII.50)
where u+1 ∈ Uq−1 W0 and g ∈ Uq W0 do not involve dtn .
Then
B C +1
tnj
+1
tn
G− u = tn h − d u+1
j=1 j +1
tn+1
= tn h − tn dtn ∧ u+1 − d u+1
+1
t+1
= tn h + tn tn g − h − n d u+1
+1
1
= tn+1 g − d u+1
+1
Defining h+1 = g − d u+1 / + 1 completes the proof of the inductive
argument.
VII.12 A Mayer–Vietoris argument 327
Next we observe that any element v ∈ Uq−1 W0 which does not involve
dtn can be written as
v = v tn =0 + tn v1
where v1 ∈ Uq−1 W0 also does not involve dtn . Reasoning again by induction we
then obtain, from (VII.51), a sequence v ⊂ U 0 W0 such that, for every ,
B C
G=d vj tn + Otn
j
(VII.54)
j=1
Finally we select s and j as in the proof of Lemma VII.12.2 in such a
way that
j tn vj tnj
j=0
converges in Uq−1 W0 . Call u ∈ Uq−1 W0 this sum: for every , (VII.54)
gives
B C
G−d u = d vj tn − u + Otn = Otn
j
j=1
We start by proving (a). Let g ∈ Uq 0 p satisfy d 0 g = 0. By Lemma VII.12.3
there is f ∈ Uq p satisfying '0 ∗ f = g, d f ∼0 0. We then define F ∈ Uq+1 p
by the rule
d f in • p
F=
−d f in \• p
Then F ∈ Uq+1 p and d F = 0. We now apply our hypothesis: we can find
f ∈ Uq p solving d f = F and also u• ∈ Uq−1 p • , u•• ∈ Uq−1 p \•
solving d u• = f − f in • p, d u•• = f + f in \• p. We then set
328 The differential complex
1
u = '0 ∗ u• + '0 ∗ u••
2
We obtain
1
d 0 u = ' ∗ d u• + '0 ∗ d u••
2 0
1 !
= '0 ∗ f − f + '0 ∗ f + f
2
= '0 ∗ f
=g
f − d u + V ∼0 0 (VII.55)
Define
f − d u + V in • p
G=
0 in \• p
n
= dx − bj x t dtj
j=1
is d -closed.
In other words, the structure will be locally integrable near p if the class of
f0 in H 1 p vanishes.
Proof. Assume that (i) holds. If we differentiate the identity
Ztj + bj Zx = 0
with respect to x we obtain
Zx tj + bj Zx x = −bj x Zx
which gives
Lj log Zx = −bj x
Thus
d − log Zx = f0
which proves (ii).
330 The differential complex
Then
x
Lj Gx t = bj x teuxt + utj y teuyt dy
0
x
= bj x teuxt − bj y tuy y t + bj y y t euyt dy
0
x
= bj x teuxt − bj y teuyt y dy
0
= bj 0 teu0t
If we set
n
Bt = bj 0 teu0t dtj
j=1
Lemma we can find Ft smooth near the origin such that dt F = B and then if
we set Zx t = Gx t − Ft we obtain Zx = expu = 0 and d Z = 0.
Notes
The differential complex associated with a formally integrable structure, first
presented in [T4], is the natural generalization of the de Rham, Dolbeault,
and tangential Cauchy–Riemann complexes, associated respectively with real,
complex, and CR structures.
Much of the material of this chapter is preparatory for Chapter VIII, and
we should just point out that the Approximate Poincaré Lemma is due to
Treves ([T4]), whereas Theorem VII.12.1 is a consequence of the existence
of a natural Mayer–Vietoris sequence, whose existence for hypersurfaces in
the complex space was first proved in [AH1].
VIII
Local solvability in locally integrable structures
331
332 Local solvability
that is, the d -complex is nothing other than the standard de Rham complex
along the leaves of the foliation defined by . In particular, if we apply the
Poincaré Lemma (Section VII.3) we conclude that local solvability holds for
an essentially real structure near any point and at any degree.
Proposition VIII.2.1. Let f ∈ Uq p have analytic coefficients and satisfy
d f = 0. If q ≥ 1 then there is u ∈ Uq−1 p, also with analytic coefficients,
solving d u = f .
Proof. We let
f= fI x t dtI
I =q
and
f2 = −1q−1 fJ ∪r dtr ∧ dtJ
J =q−1J ⊂1r−1
Lr Ls uJ = 0 s > r
r−1
d u = Lj uJ dtj ∧ dtJ +
J =q−1J ⊂1r−1 j=1
Lr uJ dtr ∧ dtJ
J =q−1J ⊂1r−1
and hence (VIII.9) implies that the d -closed form d u − f only involves
dt1 dtr−1 . The proof can then be concluded by an elementary inductive
argument, whose details are left to the reader.
Let m ∈ Z+ and write the variables in Cm × Rn as
z t = z1 zm t1 tn
We shall also write zj = xj + iyj , j = 1 m, and n = m + n .
The elliptic complex on Cm ×Rn is defined as follows: given ⊂ Cm ×Rn
open and 0 ≤ q ≤ n we set C # as being the space of all smooth
q
where C U × " # is the space of forms of the kind
rs
f= fJK dzJ ∧ dtK
J =r K =s
VIII.3 Elliptic structures 335
such that D = z + dt .
Let f ∈ C U × " #q satisfy Dq f = 0 and decompose f = rs frs ,
where frs ∈ C U × " #rs and the sum runs over the pairs r s such that
r + s = q, r ≤ m, s ≤ n . Consider, in this decomposition, the term frs whose
value of s is maximum. From the fact that Df = 0 it follows that dt frs = 0
and consequently we can apply the Poincaré Lemma (Section VII.3) in order
to find h ∈ C U × " #rs−1 such that dt h = frs . If we set f • = f − Dq−1 h
it follows that in the analogous decomposition f • = rs frs •
the maximum
•
value of s that occurs has dropped by one and Dq f = 0.
If we iterate the argument we will, after a finite number of steps, either
solve the equation Dq−1 u = f or at least find v ∈ C U × " #q−1 such that
g = f − Dq−1 v does not involve dt1 dtn . If this is the case we can write
g= gJ z tdzJ
J =q
and the fact that Dq g = 0 gives in particular that dt gJ = 0 for all J , that is,
gJ are independent of t. Hence g defines a Dolbeault class in U and by the
standard complex analysis theory we can determine w = L =q−1 wL zdzL
solving z w = g. If we set u = v + w we obtain Dq−1 u = f , which completes
the proof.
Likewise we can introduce the spaces #q , which are the spaces of
all currents of the form (VIII.11) where now the coefficients are allowed to
be elements of . By the same expressions (VIII.12) and (VIII.13) we
obtain new differential complexes
Dq #q −→ #q+1 (VIII.14)
#q = f ∈
#q supp f ⊂⊂ (VIII.17)
defined by
f * −→ f ∧ dz ∧ *
which allows us to identify #q with the topological dual of Cc
#n−q , when the latter carries its natural structure of an inductive limit of
Fréchet spaces. We shall use the standard notation of the de Rham theory: if
T ∈ #q and * ∈ Cc #n−q we shall set
T * = T ∧ * 1 = T ∧ dz ∧ *
Proof. Using the fact that C #q ⊂ #q as well as Cc #q ⊂
#q are dense inclusions we can assume that T = f is smooth. We have
Since
d2m+n −1 f ∧ dz ∧ * = 0
0 is hypoelliptic in Cm × Rn (VIII.26)
n
Moreover, since any open subset of Cm ×R is P-convex for singular supports
([H3]), we also have
Given any open set ⊂ Cm × Rn the maps (VIII.27)
q #q −→ #q are surjective
338 Local solvability
Proposition VIII.4.1. For any open set ⊂ Cm × Rn the maps (VIII.15)
are isomorphims. More precisely:
q−1 w = u
for some w ∈ #q−1 . Then, by (VIII.24),
(ii) is proved.
Finally let f be as in (iii) and solve
HDn = 0
VIII.4 The Box operator associated with D 339
we obtain
q E U = q E U = U (VIII.30)
Dq−1 D∗q−1 E U + D∗q E Dq U = U (VIII.31)
We push the argument further. Let be a regular, bounded open subset of
Cm × Rn . If f ∈ C #q we consider f ∈ Cm × Rn #q , where
denotes the characteristic function of . We obtain, from (VIII.31),
f = Dq−1 D∗q−1 E f + D∗q E Dq f + D∗q E D0 ∧ f
(VIII.32)
If we now introduce the operators
Gq C #q −→ C #q−1 Hq C #q −→ C #q
defined by the expressions
Gq f = D∗q−1 E f Hq f = D∗q E D0 ∧ f (VIII.33)
formula (VIII.32) then gives a natural extension of the so-called Bochner–
Martinelli formula:
Theorem VIII.4.4. If is a regular, bounded open subset of Cm × Rn with
a smooth boundary and if f ∈ C #q then
Dq−1 Gq f + Gq+1 Dq f + Hq f = f (VIII.34)
340 Local solvability
defined by
Iq f " = f ∧ dz ∧ "
The intersection number for the pair in degree 0 is the C-bilinear
form defined on the product
f ∈ C D0 f = 0 ×" ∈ Cc #n F dz ∧ " = 0
∀F ∈ C Cm × Rn D0 F = 0
defined by
I0 f " = f dz ∧ "
for some sequence u in C #q−1 , and if " ∈ Cc #n−q satisfies
Dn−q " = 0, we have
Iq f " = lim Dq−1 u ∧ dz ∧ " = lim −1q+m u ∧ dz ∧ Dn−q " = 0
→ →
F −→ f in C
Proof. The proof that (VIII.36) implies the vanishing of I0 is immediate.
We then prove the converse and for this we argue by contradiction as in the
proof of Proposition VIII.5.1. Thus we assume that there is f0 ∈ C
satisfying D0 f = 0 for which no sequence F ⊂ C Cm × Rn as stated
342 Local solvability
exists and apply once more the Hahn–Banach theorem: there is T ∈ #n
such that
T f0 = 1 (VIII.37)
T F = 0 ∀F ∈ C Cm × Rn D0 F = 0 (VIII.38)
We next observe that the vanishing of HD1 Cm × Rn (Theorem VIII.3.1)
implies, in particular, that the homomorphism of Fréchet spaces
D0 C Cm × Rn −→ C Cm × Rn #1
has closed image. Consequently its transpose, which is the map
Dn Cm × Rn #n−1 −→ Cm × Rn #n
has a weakly closed image, that is, its image is precisely the orthogonal of
the kernel of D0 C Cm × Rn → C Cm × Rn #1 in Cm × Rn #n .
Returning to our argument we conclude from (VIII.38) that there exists
S ∈ Cm × Rn #n−1 such that Dn−1 S = T .
As in the proof of Proposition VIII.5.1 we introduce once more the regu-
larizations
"% = &% T ∈ Cc Cm × Rn #n
There is %0 > 0 such that "% ∈ Cc #n if 0 < % ≤ %0 . Furthermore, if
F ∈ C Cm × Rn satisfies D0 F = 0 then
F dz ∧ "% = F dz ∧ &% Dn−1 S = F dz ∧ Dn−1 &% S
= −1m−1 D0 F ∧ dz ∧ &% S = 0
(a) The exterior derivative defines a map with closed image when defined
on an arbitary smooth manifold.
(b) The d-cohomology is isomorphic to the standard singular cohomology
with complex coefficients for any smooth manifold.
These two properties in conjunction with Proposition VIII.5.1 prove the asser-
tion for q ≥ 1. Furthermore, since a scalar function is d-closed if and only if
it is locally constant, the assertion for q = 0 is an immediate consequence of
Proposition VIII.5.2.
We shall now draw an important corollary of Propositions VIII.5.1 and
VIII.5.2. Let be a regular, bounded open subset of Cm × Rn . Since we are
dealing with an elliptic structure on Cm × Rn it follows that is nonchar-
acteristic and consequently we can apply the one-sided approximate Poincaré
Lemma (Theorem VII.8.4) and obtain:
Corollary VIII.5.5. Let p ∈ . Given any neighborhood W of p in Cm ×Rn
there is another such neighborhood W ⊂⊂ W such that IqW ∩W ∩ = 0 for
all q = 0 m + n .
f •
=0 in /2 (VIII.40)
w
We must find g ∈ C such that
for every F ∈ C #q−1 satisfying Dq−1 F = 0 and every T∈ #n−q+1
satisfying Dn−q+1 T = 0 (when q = 1 this condition must be replaced by
T G = 0 for every G ∈ C Cm × Rn satisfying D0 G = 0).
Fix w0 ∈ C and let f ∈ C w0 #q−1 , " ∈ Cc w0 #n−q be both
D-closed (we assume " ∈ g ∈ C C ×Rn D0 g = 0 ⊥ when q = 1). Thanks
to our hypotheses we can apply Proposition VIII.6.2 in order to obtain F ∈
C #q−1 satisfying Dq−1 F = 0 and F w=w0 = f .
346 Local solvability
we have Dn−q+1 T" = 0 and also T" ∈ G ∈ C Cm × Rn D0 G = 0 ⊥ when
q = 1. Then (VIII.42) gives
Iq−1
w0 w0 f " = F w=w0 ∧ dz ∧ " = ± F ∧ dz ∧ dw ∧ T" = 0
In other words, we shall assume that is defined by & > 0, where & is real-
valued, real-analytic and such that is defined by & = 0, with &/z = 0
near p.
The next result is the key tool for the proof of the result. In all the arguments
that follow we shall denote by ) C × Rn → C the projection )z t = z.
We assume that the central point in the analysis is p = z0 t0 ∈ in
C × Rn . By applying the implicit function theorem we can assume that
Proposition VIII.8.1. Assume that both (VIII.44) and condition 0 hold.
Then given any D ∈ Vp there is D• ⊂⊂ D also belonging to Vp and a
constant M > 0 such that, for any z ∈ C, any two points in z ∩ D• can
be joined by a piecewise real-analytic curve contained in z ∩ D and with
length ≤ M.
Before we embark on the (rather long) proof of this result, we will show
how it can be used to derive our one-sided solvability result.
Corollary VIII.8.3. Assume (VIII.44) and that condition 0 holds. Then
given any D0 ∈ Vp there is D ⊂⊂ D0 also belonging to Vp such that
for every f ∈ C ∩ D0 #1 satisfying D1 f = 0 there is v ∈ C ∩ D
satisfying D0 v = f in ∩ D and
Notice that, in particular, (VIII.44) and condition 0 imply solvability for
D near p in degree 1 with respect to (cf. Remark VIII.7.3).
Proof. Write
n
f = f0 dz + fj dtj
j=1
E C ∩ D −→ C D
350 Local solvability
But then some subsequence vk converges weakly to a function v which satis-
fies the required properties. This concludes the proof of Corollary VIII.8.3.
for every a and thus the proof will be completed if we can show that, for
some suitable choice of the partition of unity *j , we can obtain a solution F
to (VIII.51) on R ∩ , with R ⊂ R• another square centered at z0 , satisfying
and take a closer look on the coverings Ua and Wa . We have
= z ∈ R• ∃t ∈ "• &z t > 0
352 Local solvability
Wa = z ∈ R• &z a > 0
We set
z = sup &z t = max &z t
t∈"• t∈"•
There are *j ∈ Cc Qj , 0 ≤ *j ≤ 1, such that *j = 1 and (VIII.55)
& &
& *j &
& & −1
& z z& ≤ Cz
j
satisfies
Fz ≤ M1 dt uL ∩D z ∈ R
Let us then assume that z0 = 0. We now take R as stated and such that
and then
∩ R = z ∈ R∗ y > .x
i!tj
Lj = − j = 1 n (VIII.61)
tj 1 + i!x x
Hence • is exactly the locally integrable structure defined on a neigh-
borhood of the point p in Rn +1 which is orthogonal to the sub-bundle of
CT ∗ Rn +1 spanned by dZ, where Zx t = x + i!x t.
The reverse argument is also true, that is, any smooth locally integrable
structure of corank one, say in a neighborhood of the origin in Rn +1 , arises
VIII.10 Solvability for corank one analytic structures 355
from the restriction of the elliptic structure • on C × Rn to a hypersurface 0
in C × Rn . Indeed if we choose local coordinates x t = x t1 tn in a
neighborhood of the origin in Rn +1 in such a way that the orthogonal of is
generated by the differential of Zx t = x +i!x t, with ! smooth and real-
valued, and if we denote by 0 the image of the imbedding x t → Zx t t,
it follows easily that = • 0.
Keeping this notation, and recalling Corollary I.10.2, we can (and will)
even assume that !0 0 = !x 0 0 = 0. We emphasize that is spanned
by the pairwise commuting vector fields (VIII.61). We further take a small
open neighborhood V of the origin in C × Rn and set
Definition VIII.10.1. We shall say that condition (P)0 holds at the origin
for the locally integrable structure if condition 0 holds at the origin in
C × Rn with respect to both V+ and V− .
Thus
M 2k Lwk − f = 0
2r +2ks0
Since wk ∈ Lloc 0 U it follows that
2r +2k−1s0 −1
Lwk − f ∈ Lloc 0 U
2s −1
and hence gjk ∈ Lloc 0 " . Consequently Gjk ∈ Lloc 0 " and then, if we set
2s
2k−1
uk = wk + Gjk t Zx tj
j=0
we have
2r +2ks0
Luk = f uk ∈ Lloc 0 U (VIII.64)
Notes
Until now, complete answers for local solvability in locally integrable struc-
tures, besides the cases n = 1 (a situation which has already been discussed in
Chapter IV), defines an essentially real structure (Section VIII.1) and when
defines an elliptic structure (Theorem VIII.3.1) are known in the following
cases: (i) defines a nondegenerate locally integrable CR structure of codi-
mension one ([AH2]); (ii) defines a nondegenerate real-analytic structure
([T9]); (iii) m = 1 ([CorH3]).
We also mention a necessary condition for nondegenerate CR structures
of arbitrary codimension proved in [AFN], which was extended to general
locally integrable structures with additional nondegeneracy conditions in [T5].
The notion of intersection number and the necessary condition given in
Theorem VIII.11.4 is due to [CorT1].
As far as sufficiency is concerned, we point out the works by Kashiwara–
Schapira ([KaS]) and Michel ([Mi]), which deal with locally integrable CR
structures of codimension one and whose Levi form satisfies weaker nonde-
generacy conditions.
Locally integrable structures with m = 1: for this class of locally integrable
structures we have seen in Sections VIII.7 and VIII.8 that condition (P)0 is
necessary and (in the real-analytic category) sufficient for local solvability
near the origin (cf. Corollary VIII.7.5 and Theorem VIII.10.2). This result
358 Local solvability
(P)q Given any open neighborhood V of the origin there is another such
neighborhood V ⊂ V such that, for every regular value z0 ∈ C of the
map Z, either Z−1 z0 ∩ V = ∅ or else the homomorphism
H̃q Z−1 z0 ∩ V −→ H̃q Z−1 z0 ∩ V
induced by the inclusion map
Z−1 z0 ∩ V ⊂ Z−1 z0 ∩ V
vanishes identically. Here H̃∗ denotes the reduced homology with complex
coefficients.
where no compatibility condition occurs. This makes this case, in some sense,
the closest to the single equation situation.
Before we introduce the solvability condition for (VIII.66) we introduce
the following definition: a real-valued function F defined on a topological
space X is said to assume a local minimum over a compact set K ⊂ X if there
exist a ∈ R and K ⊂ V ⊂ X open such that F = a on K and F > a on V \K.
Notes 359
Indeed k can be taken any integer ≥ n/2 + 1 and equal to zero when the
structure is real-analytic ([CorH1]). The completion of the argument is quite
standard, and holds whatever the value of m: by applying the Hahn–Banach
theorem it is easily seen that (VIII.67) implies the existence of weak solutions
to (VIII.66), and a general result due to [T5] proves the existence of smooth
solutions.
For the tube structures it is not difficult to prove that property Pn−1
implies (VIII.67) and consequently the preceding discussion shows that our
conjecture is also satisfied for this particular class.
When defines a CR structure of codimension one then condition Pn−1
is equivalent to the existence of an open neighborhood U of the origin such
that at every characteristic point over U the Levi form is not definite. In this
case a partial answer was given in [Mi], where the existence of hyperfunction
solutions is proved.
360 Local solvability
In this epilogue we describe briefly some results that are closely connected with the
theory and tools developed in previous chapters and have been obtained in recent
years but, in spite of their importance, could not be fully treated without increasing
too much the size of this book.
Lu = Au + Bu (1)
where L is a complex vector field in the plane and A and B are bounded, measurable
functions. We will also present two applications of equation (1). The first application
concerns uniqueness in the Cauchy problem for a class of semilinear equations. The
second application will be to the theory of bending of surfaces.
Equation (1) generalizes the classical elliptic equation
u
= Au + Bu (2)
z
which was investigated by numerous researchers (see for example [Be], [CoHi], [Re],
and [V]). In the literature, solutions of (2) are called pseudo-analytic functions or
generalized analytic functions. Such functions share many properties with holomorphic
functions of one variable. These properties follow easily from the similarity principle
according to which every continuous solution of (2) has the local form
u = expg h (3)
where h is a holomorphic function and g is Hölder continuous. Thus, for example, the
zero set of u is the same as that of h. The similarity principle holds for any elliptic
vector field L (where the holomorphy of h is replaced by the condition Lh = 0) since
any such vector field is a multiple of z in appropriate coordinates. In [Me2] Meziani
explored the validity of the similarity principle for the following three classes of vector
361
362 Epilogue
fields:
Lk = − iy2k Kn = − ixn M= − iy
y x y x y x
where k and n are non-negative integers. It was proved in [Me2] that the similarity
principle is valid for the Lk and Kn (under some vanishing assumption on Bx y on
the characteristic sets of the vector fields) in the following sense: if w is a continuous
solution of
Lw = Aw + Bw
where L ∈ Lk Kn , then w has the local form w = expg h where Lh = 0 and g
is Hölder continuous. It was also shown in [Me2] that this principle does not hold
for M. The vector fields Lk and Kn are locally solvable while M is not. With this
observation as a point of departure, it was shown in [BHS] and [HdaS] that a weaker
version of the similarity principle is valid for all locally solvable vector fields L. In
this weaker version, the functions g and h in the representation w = expg h may no
longer be continuous. However, this representation was still good enough to yield the
uniqueness result mentioned below.
fx t − fx t ≤ K −
If L and f are as above, the following result on uniqueness in the Cauchy problem
was proved in [HdaS] (see also [BHS]):
Theorem 1.1. Suppose ux t wx t ∈ Lp , p ≥ 2, satisfy Lu = fx u u, Lw =
fx w w, and ux 0 = wx 0. Then u ≡ w in a neighborhood of the origin.
If the coefficients of L are smooth, Theorem 1.1 was proved in [BHS] under the
weaker assumption that u and w belong to Lp p > 1. These results were proved by
applying the similarity principle to the difference v = u − w which in view of the
assumptions satisfies an equation of the form Lv = Av + Bv with A and B bounded.
The fact that L satisfies condition is then used to reduce matters to a planar situation.
1 The similarity principle and applications 363
Lu = Au + Bu
(L a planar vector field) and the study of infinitesimal deformations of surfaces with
non-negative curvature. Here we will summarize some of the results in [Me4] to
indicate this link.
Let S be a surface of class C l , l > 2, embedded in R3 and given by parametric
equations as
satisfying
dRs t · dUs t = 0 ∀s t ∈ D (7)
Rs · Us = 0 Rt · Ut = 0 Rs · Ut + Rt · Us = 0 (8)
E = Rs · Rs F = Rs · Rt G = Rt · Rt (9)
where
Rs × Rt
N=
Rs × Rt
is the unit normal to S. The Gaussian curvature of S is
eg − f 2
K=
EG − F 2
We will assume that the curvature K ≥ 0 everywhere on S. The (complex) asymptotic
directions of S are given by the quadratic equation
2 + 2f + eg = 0
364 Epilogue
That is,
+
= −f + i eg − f 2
w = LR · U (12)
CLw = Aw + Bw
Lu ≤ M u a.e. in U .
2 Mizohata structures
The vector field in R2 , where the coordinates are denoted x t, given by
M= − it (13)
t x
2 Mizohata structures 365
is called the (standard) Mizohata vector field (or operator) after the work of S. Mizohata
([M]) who studied the analytic hypoellipticity of a class of related operators of which
M is the simplest example. A globally defined first integral of M is the function
Zx t = x + it2 /2. Notice that t → t2 fails to be monotone in any neighborhood of
a point x0 0, i.e., condition in not satisfied at any point of the x-axis and,
as discussed in Chapter IV, fails to be locally solvable at those points. Thus, it is
the simplest example of a nonlocally solvable operator and, in fact, its lack of local
solvability at points of the x-axis can be proved by ad hoc elementary arguments, as
shown by L. Nirenberg ([N1]). Off the x-axis, M is elliptic. In his Lectures Notes,
Nirenberg constructed a perturbation of the Mizohata operator
L= − it1 + &x t (14)
t x
with &x t real-valued and vanishing to infinite order at t = 0, which is not locally
integrable in any neighborhood of the origin. As a matter of fact, any smooth function u
that satisfies the homogeneous equation Lu = 0 in a connected open set U that contains
the origin must be constant. In spite of the fact that the perturbed vector fields L and
M behave differently with respect to local integrability, they have important geometric
features in common. We have
(1) M and its conjugate M are linearly dependent precisely on the x-axis;
(2) M and M M are linearly independent whenever M and M are linearly dependent.
We also say that a Mizohata vector field L is of standard type at p ∈ 0 if there exist
local coordinates x t in a neighborhood of p in terms of which 0 is given by t = 0
and has the form (13). A Mizohata structure on is a structure which is locally
generated in the neighborhood of every point by a Mizohata vector field.
Notice that (1) means that 0 is the image of the characteristic set p ∈ T ∗
p = 0 , being the symbol of L, under the canonical projection 1 T ∗ −→ .
With this terminology, the vector field (13) is a Mizohata vector field of standard type
and (14) is also a Mizohata vector field but not of standard type. Indeed, (14) cannot
be of standard type because it is not locally integrable.
Notice that a Mizohata vector field is elliptic on \0, which is a relatively small
set, since an application of the implicit function theorem shows that 0 is an embedded
curve. The following question was considered by Treves [T7]: when is a Mizohata
vector field L of standard type at a given point? Of course, since this is a local
question, it is enough to study the case when L is defined in a neighborhood of the
origin in R2 . He showed that local coordinates can be found so that L becomes of the
form (14) with &x t real-valued and vanishing to infinite order at t = 0, in other
words, every Mizohata vector field has this form locally and it will be of standard type
366 Epilogue
if we are able to take & ≡ 0. Furthermore, L is of standard type at the origin if and only
if it is locally integrable. Then Sjöstrand ([Sj2]) took a closer look into the nonlocally
integrable case. To describe his results, let us consider the problem of finding a smooth
function Z+ x t satisfying dZ0 0 = 0 and LZ+ = 0 on U + = U ∩ t ≥ 0 , where
U is a small disk centered at the origin. By the proof of Lemma I.13.4, to find Z+ it
is enough to find a smooth function u that satisfies Lu = t&x on U + . This is, in fact,
possible because L satisfies condition for t > 0 ([BH6]). Similarly, shrinking U
if necessary, we can also find a smooth function Z− x t satisfying dZ− 0 0 = 0
and LZ− = 0 on U − = U ∩ t ≤ 0 . We can always choose Z+ and Z− satisfying
Z± 0 0 = 0, JZx± 0 0 = 0, and Zx± 0 0 > 0 and we will do so. If we are so
lucky that Z+ x 0 = Z− x 0, x 0 ∈ U , we may patch Z+ and Z− to get a single
continuous solution Z of LZ = 0 on U and it is easy to see using the equation that
Z is actually smooth. So the obstruction to the local integrability of L is related to
the difficulty of finding a pair Z+ Z− such that LZ± = 0 on U ± and Z+ = Z− on
U + ∩ U − . Given such a pair, it can be shown that the range of Z± lies on one side
of the smooth curve Z± x 0 (in fact, above the curve because Zx± 0 0 > 0), so
let H ± z be a smooth function defined on the range of Z± and holomorphic in its
interior with H ± 0 = 0, H ± 0 = H ± 0 > 0. Then, Z̃± = H ± Z± satisfies
dZ̃± 0 0 = 0 and LZ̃± = 0 on U ± . By the Riemann mapping theorem we may find
H + and H − so that the range of Z̃+ and Z̃− is the upper half-plane. In other words, we
may restrict ourselves to consider pairs Z+ Z− such that Z± U ± = Jz ≥ 0 and
Z± U + ∩ U − = R. Given such a pair and a smooth function H defined on Jz ≥ 0,
holomorphic on Jz > 0, real for z real and satisfying H0 = 0, H 0 > 0, a new
pair Z+ Z̃− = Z+ H Z− may be considered and L will be locally integrable
if Z+ x 0 = Z̃− x 0. It turns out that L is locally integrable if and only if there
exists a pair Z+ Z− such that Hz = Z+ Z− −1 z is holomorphic for Jz > 0 and
smooth up to Jz = 0. Since Hz is real for z real, H has, by the reflection principle,
an extension to a holomorphic function. By uniqueness, Hx + iy is determined by its
trace bHx = Hx + i0 so it is enough to look at the restrictions bZ± x = Z± x 0
and check whether 2 = bZ+ bZ− −1 R −→ R has a holomorphic extension to a
neighborhood. Summing up, to each Mizohata vector field L we have associated an
increasing diffeomorphism 2 R −→ R such that L is locally integrable if an only if
2 = bH for some H ∈ C, i.e., 2 has a holomorphic extension. More generally, we
may consider the following question: given two Mizohata vector fields L1 , L2 , when are
they equivalent in the sense that one can be locally transformed into a multiple of the
other by a change of variables? The answer, due to Sjöstrand, can be stated as follows.
Consider the associated diffeomorphisms 21 = bZ1+ bZ1− −1 and 22 = bZ2+ bZ2− −1 ,
then L1 and L2 are equivalent if and only if there are holomorphic functions, H1 z,
H2 z, real and increasing for z real, such that 21 H1 x = 22 H2 x, x ∈ R.
The local questions of standardness and equivalence for Mizohata vector fields
have their global counterpart. For instance, it was established in [BCH] that a locally
standard Mizohata planar vector field has a first integral globally defined in a tubular
neighborhood of the characteristic set 0. The standardness of a particular class of
Mizohata structures on the sphere S 2 was proved in [Ho4] and Jacobowitz ([J2])
studied Mizohata structures on compact surfaces , in particular, he proved that the
existence of a first integral is equivalent to the fact that the genus is even. In the case
of the sphere, he gave a classification of Mizohata structures in the spirit of Sjöstrand’s
2 Mizohata structures 367
Lj = − i-j tj 1 + &j x t j = 1 n (16)
tj x
where the functions &j x t, j = 1 n, vanish to infinite order at t = 0. In other
words, every Mizohata structure has at a given point a contact of infinite order with
a standard Mizohata structure of the same type. In particular, if we can take all the
functions &j identically zero will have a first integral in U and will be standard in
U . Conversely, if has a first integral it is possible to choose the coordinates so that
is generated by the vector fields (15).
For the case = 1, i.e., if the type is 1 n − 1 , Treves showed the existence
of functions &j x t vanishing to infinite order at t = 0 such that the structure
spanned by (16) is formally integrable (i.e., ⊂ ) and not locally integrable. On
the other hand, Meziani proved in [Me7] that Mizohata structures of all other types
n − = 1 n − 1 are always locally integrable. His proof is delicate and beyond
368 Epilogue
the scope of this book: he first constructs first integrals on the connected components
of x t t ∈ Rx × Rt t 2 = t 2 which can be 2 (if n > 2 and < n − 2), 3
(if n > 2 and = n − 2), or 4 (if n = 2 and = 0). When the components are 2
or 4, these first integrals can be patched together to yield a globally defined first
integral of class C 1 which, by the hypoellipticity of the structure, is in fact smooth.
The possibility of patching together these partially defined first integrals depends on a
careful analysis of the holonomy of a certain foliation with leaves of dimension n − 1
defined by the structure. For the case of type 1 n − 1 he gives a classification of
Mizohata structures analogous to Sjöstrand’s result for a single vector field. The local
integrability for Mizohata structures of type 0 n , n ≥ 3, was first proved in [HMa2],
by techniques akin to those used in the proof of Kuranishi’s embedding theorem for
CR structures ([Ku1], [Ku2], [Ak], [W2], [W3]), which also fall beyond the scope
of this book. The restriction n ≥ 3 comes from a technical fact: Kuranishi’s approach
depends on the existence of certain so-called homotopy formulas that do not exist
when n = 2 ([HMa3]). However, the local integrability of Mizohata structures of type
0 n in Rn+1 , n ≥ 2, can be proved by elementary methods. Consider a system of n
commuting vector fields
Lj = − itj 1 + &j x t j = 1 n (17)
tj x
Here a generic point is described by coordinates x t1 tn and the smooth func-
tions &j x t vanish to infinite order at + = t = 0 = Rx × 0 . We regard the Lj ’s
as perturbations of the Mizohata vector fields
Mj = − itj j = 1 n
tj x
A simple computation using polar coordinates, t = r, r > 0, ∈ S n−1 shows that the
standard Mizohata structure spanned by the Mj ’s is also spanned on Rn+1 \+ by
⎧
⎪
⎪ M= − ir
⎪
⎨ r x
⎪
⎪
⎪
⎩k =
k=1,…,n −1
k
with 1 n−1 angular variables in S n−1 . Then, the change of variables s = r 2 /2
(x and are kept unchanged) takes M into a multiple of the Cauchy–Riemann operator
1
z̄ = +i z = x + is s > 0
2 x s
and does not change k . If we perform the same operations on the perturbed system (17)
we may find a set of generators of in the variables x s ∈ Rx × R+
s ×
n−1
of the
form
2 Mizohata structures 369
⎧
⎪
⎪L̃1 = + 1
⎪
⎨ z̄ z
(18)
⎪
⎪
⎪
⎩L̃k =
+ k k=2,…,n
k−1 z
with smooth coefficients j x s , j = 1 n, that converge to zero when s ( 0
together with their derivatives of any order. Thus, we may smoothly extend the
coefficients j as zero for Jz = s ≤ 0 and obtain an elliptic system defined on C ×
S n−1 Rx × Rs × S n−1 which for Jz < 0 has the first integral z = x + is. The process
that produced an elliptic system starting from a nonelliptic one was obtained by a
combination of singular changes of variables (polar coordinates that are singular at
the origin of Rnt and s = r 2 /2 which is singular at r = 0) and blows up the line
Rx × t = 0 to the n-manifold Rx × S n−1 . Although we know from Theorem I.12.1
that elliptic structures are locally integrable, applying that result to (18) would only
give us a first integral defined in a neighborhood of a point s = 0, x = 0, = 0 ∈ S n−1
while only a first integral defined for all ∈ S n−1 can give us a first integral defined
in a neighborhood of the origin of the original variables x t. Let’s consider first the
case n = 2, that is the system of two vector fields
⎧
⎪
⎪L̃1 = + 1 z = x + is ∈ C
⎪
⎨ z̄ z
(∗)
⎪
⎪
⎪
⎩L̃ = + 0 ≤ ≤ 2)
2 2
z
defined in C × S 1 , where the j x s , j = 1 2, are C functions, 2)-periodic in
, and vanish for s = Jz ≤ 0. Choose a smooth function = x s such that
X = L̃2 + L̃1 is a real vector. It is easy to check that this is possible if 1 < 1 (in
particular for x s close to the origin). Thus, X is a real generator of the structure ˜2
spanned by L̃1 and L̃2 for x < 1, s < - and 0 ≤ ≤ 2). It is clear that X = /
for s ≤ 0, and that the orbits of X stemming from points x0 s0 0, s0 ≤ 0, are the
closed circles → x0 s0 , 0 ≤ ≤ 2). Notice also that the component of X along
/ is 1, i.e.,
X= + &1 + &2
x s
for some smooth functions &1 and &2 which are 2)-periodic in and vanish for s ≤ 0.
Since the commutator X L̃1 ∈ ˜2 it must be a linear combination of L̃1 and L̃2 ;
on the other hand, it does not contain derivations with respect to so it has to be
proportional to L̃1 . This shows that there exists a smooth function = x s such
that
X L̃1 = L̃1 (19)
Now pick once and for all a local solution Wx s of
W W
L̃10 W = + 1 x s 0 = 0 (20)
z̄ z
Wx 00 = 0
370 Epilogue
We may assume that in a neighborhood of the origin any other solution W x s of
L̃10 W = 0 is a holomorphic function of W , in fact, W is a local diffeomorphism that
takes L̃10 into a multiple of the Cauchy–Riemann operator. Let denote the closed
orbit of X stemming from 0 0 0, given by → 0 0 , 0 ≤ ≤ 2). We now solve
the Cauchy problem
XV = 0 (21)
Vx s 0 = Wx s
XU − U = 0
Ux s 0 = 0
3 Hypoanalytic structures
Let be a smooth manifold of dimension N . By a hypoanalytic structure on (cf.
[T5]) we mean a collection of pairs = U Z , with U an open subset of and
4 The local model for a hypoanalytic manifold 371
• = U ∩ • Z U ∩•
To each hypoanalytic structure = U Z on we can canonically associate a
locally integrable structure on in the following way: for each its orthogonal on
U is defined by
T U = span dZ1 dZm
By properties (H)1 , (H)2 , and (H)3 it follows that T is indeed a subbundle of CT ∗
of rank m.
Notice however that two different hypoanalytic structures can define the same
locally integrable structure. Indeed, to give an example it suffices to take = R and
consider the hypoanalytic structure R Id , where Idx = x, and the hypoanalytic
structure R f , where f R → R is smooth but not real-analytic and f = 0 at each
point.
By a hypoanalytic manifold we shall mean a pair , where is a smooth
manifold and is a hypoanalytic structure on . Notice that if is a hypo-
analytic manifold, endowed with the hypoanalytic structure = U Z , if is
another smooth manifold and if f → is a smooth submersion, then we can pull
back the hypoanalytic structure to a hypoanalytic structure f ∗ on by defining
f∗ = f −1 U Z f
Finally we shall say that two hypoanalytic manifolds and are equivalent
if there is a smooth diffeomorphism f → such that f ∗ = .
A.1 Multipliers in h1
We recall that is a modulus of continuity if 0 −→ R+ is continuous,
increasing, 0 = 0 and 2t ≤ Ct, 0 < t < 1. A modulus of continuity determines
the Banach space C R of bounded continuous functions f R −→ C such that
fy − fx
f = sup <
x=y x − y
C
equipped with the norm f C = f L + f C . Note that C is only determined by
the behavior of t for values of t close to 0. Consider a modulus of continuity t
that satisfies
1 h 1 −1
tt dt ≤ K 1 + log
n−1
0 < h < 1 (A.1)
hn 0 h
Lemma A.1.1. Let b ∈ C Rn and f ∈ h1 Rn . Then bf ∈ h1 Rn and there exists
C > 0 such that
Proof. Let bx ∈ C . It is enough to check that bf ≤ CbC for every h1 -atom
a with C an absolute constant. This fact is obvious for atoms supported in balls
B with radius & ≥ 1 without moment condition because b is bounded so ba/bL
is again an atom without moment condition. If B = Bx0 &, & < 1, we may write
axbx = bx0 ax + bx − bx0 ax = 1 x + 2 x. Then 1 x/bL is
again an atom while 2 x is supported in B and satisfies
C
2 L ≤ 2bL aL ≤
&n
C
2 L1 ≤ CaL x − x0 dx ≤
B 1 + log &
374
A.1 Multipliers in h1 375
We wish to conclude that m! 2 L1 < . Let B∗ = Bx0 2&. Since m! 2 x ≤
2 L , we have
J1 = m! 2 x dx ≤ C B∗ &−n ≤ C
B∗
It remains to estimate
J2 = m! 2 x dx = m! 2 x dx (A.2)
R\B∗ 2&≤ x−x0 ≤2
(observe that m! 2 is supported in Bx0 2 because supp ! ⊂ B0 1). If 0 < - < 1
and !- ∗ 2 x = 0 for some x − x0 ≥ 2& it is easy to conclude that - ≥ x − x0 /2,
which implies
& & C 1 C x − x0 −n
& &
!- ∗ 2 x ≤ & !- y2 x − y dy& ≤ 2 L
≤
-n 1 + log &
so
C
m! 2 x ≤ for x − x0 ≥ 2& (A.3)
x − x0 n 1 + log &
It follows from (A.2) and (A.3) that
C
J2 ≤ dx ≤ C
2&≤ x−x0 ≤2 x − x0 n 1 + log &
which leads to
bah1 ≤ 1 h1 + 2 h1 ≤ C1 + J1 + J2 ≤ C2
Inspection of the proof shows that C2 may be estimated by CbC .
so criterion (A.1) is satisfied. This shows that (A.5) is strictly more stringent than (A.1).
A.2 Commutators
We consider now a bounded smooth function *, ∈ R, such that
& k &
&d & 1
& &
& dxk *& ≤ Ck 1 + k ∈ R k = 0 1 2
Then * is a symbol of order zero and defines the pseudo-differential operator
1 ix
*Dux = e *$ u d u ∈ R
2) R
In particular, *D is bounded in h1 R. The Schwartz kernel of *D is the tempered
distribution kx − y defined by $k = * which is smooth outside the diagonal
x = y. Moreover, kx − y may be expressed as
1 ix−y−- 2
kx − y = lim e * d = lim k- x − y
-→0 2) -→0
where the limit holds both in the sense of and pointwise for x = y. Furthermore,
the approximating kernels k- x − y satisfy uniformly in 0 < - < 1 the pointwise
estimates
Cj
k- x − y ≤ j = 1 2 (A.6)
x−y j
which of course also hold for kx − y itself when x = y.
We consider a function bx of class C 1+ , 0 < < 1, and wish to prove that the
commutator *D bx is bounded in h1 R. A simple standard computation that
includes an integration by parts gives
*D bx ux = k x − yby − bxuy dy − *Db u
where
1
x y = b x + 1 − y d and k1 x = −xk x
0
= T1 ux + T2 ux
The first operator T1 is obviously bounded in h1 because it is the composite of
*1 D with multiplication by a C function. To analyze T2 we check—writing
k1 = lim-→0 k1- and using (A.6) for k1- —that its Schwartz kernel is a locally
integrable distribution given by the integrable function k2 x y = k1 x − y x −
y rx y. Hence, k2 x y ≤ C1 k1 x − y x − y = k3 x − y. Observe that k3 x ≤
C min x −1 x −2 so k3 ∈ L1 R. We will now show that
m! k3 x = sup !- ∗ k3 x ∈ L1 R
0<-<1
where ! ≥ 0 ∈ Cc −1/2 1/2 , !dz = 1, !- x = -−1 !x/-. Since m! k3 is
pointwise majorized by the restricted Hardy–Littlewood maximal function
1 x+-
mk3 x = sup k3 t dt
0<-<1 2- x−-
Proposition A.2.2. If *, ∈ R, is a smooth symbol of order 0 and bx ∈ C 1+ R,
0 < < 1, the commutator
*D bx
1
is bounded in h R.
Proposition A.3.1 (S. Chanillo, [Ch2]). If F satisfies A9 and A10, the map
h1 Rn g → g F is bounded in h1 Rn .
A.3 Change of variables 379
Then there exists a constant C > 0 depending on the dimension n, on K and on ! but
not on u such that
U ∗ x dx ≤ Cuh1 (A.12)
Proof. In view of the atomic decomposition it is enough to prove (A.12) when ux
is an atom, that we denote by ax. We must show that if ax is an h1 -atom and
Ax t = !t Hx − Hzaz dz 0 < t < 1
∗
then A L1 ≤ C with C independent of ax. Let ax be an atom supported in ball
B = Bz0 r such that aL ≤ B −1 . Note that in view of (A.11) and the hypothesis
on !
x − z ≥ Kt &⇒ Hx − Hz ≥ t &⇒ !t Hx − Hz = 0
for 0 < t < 1 so
1 C
Ax t ≤ aL !L dz ≤ n
z−x <Kt tn r
showing that
C
A∗ x ≤ (A.13)
rn
If we write B∗ = Bz0 2r we see right away that
A∗ x dx ≤ C
B∗
and
A∗ x dx ≤ C
B∗
it is enough to estimate
& &
& &
v∗ HL1 = sup & !t Hx − zgFz dz& dx
0<t<1
because H = F . Notice that uy = ±gy det H y ∈ h1 Rn by Lemma A.1.1
−1
and (A.10); furthermore, uh1 ≤ Cgh1 . Using Lemma A.3.2 we get I ≤ Cuh1 ≤
C gh1 , as we wished to prove.
Bibliography
381
382 Bibliography
390
Index 391
manifold section, 7
differentiable (smooth), 2 sheaf of hyperfunctions, 372
hypoanalytic, 371 similarity principle, 361
real-analytic, 2 application to bending of
maximal function surfaces, 363
Hardy–Littlewood, 292 applications to uniqueness, 362
nontangential, 285, 291 solution
microlocal analyticity, 234 classical, 7
microlocal smoothness, 237 hyperfunction, 373
minimality, 118 weak, 7
Mizohata, 214 solvability condition q , 347
structures, 365 solvability in top degree, 358
in higher dimensions, 367 submanifold
local integrability, 368 embedded, 32
of standard type, 367 codimension, 32
392 Index