Complex analysis and CR geometry

Giuseppe Zampieri
To my children Caterina, Sebastiano and Emanuele Zampieri

“brevissimus quisque dilucidissimus est”

 Contents

Preface vii
Chapter 1. Several Complex Variables 1
1.1. Analytic functions in complex spaces 1
1.2. Cauchy formula in polydiscs 5
1.3. Analytic functions and power series 11
1.4. Subharmonic functions 14
1.5. Separate analyticity 22
1.6. Levi forms—continuity principle (I)—Hartogs extension
theorem 24
1.7. Logarithmic supermean of Taylor radius of holomorphic
functions—continuity principle (II)—propagation of
holomorphic extendibility 30
1.8. Domains of holomorphy and pseudoconvex domains 39
1.9. L2 -estimates for ∂¯ on q-pseudoconvex domains of Cn 51
1.10. Subelliptic estimates for ∂¯ 68
Chapter 2. Real Structures 77
2.1. Euclidean spaces and their diffeomorphisms 77
2.2. Integration of vector fields and vector bundles—Frobenius
theorem 83
2.3. Real symplectic spaces—Frobenius-Darboux theorem 90
2.4. Subelliptic estimates and hypoellipticity of systems of
vector fields 96
2.5. Miscellanea: foliations—orbits 102
Chapter 3. Real/Complex Structures 109
3.1. Complex structures—real underlying structures—
complexifications 109
3.2. CR manifolds 115
3.3. CR functions and CR mappings 125
3.4. The Levi form of a submanifold M ⊂ Cn and an abstract
CR structure 134
3.5. Real/complex symplectic spaces 138
3.6. Approximation of CR functions by polynomials 146
v
vi CONTENTS

3.7. Analytic discs and the extension of CR functions: the

“edge of the wedge” theorem, the deformation of discs
for manifolds of type 2 and the Levi extension 153
3.8. Iterated commutators, finite type, Bloom-Graham normal
form: deformation of discs for manifolds of higher type 161
3.9. Partial lifts of analytic discs and CR curves 171
3.10. Defect of analytic discs—deformation of non-defective
discs—wedge extension from minimal manifolds 177
3.11. Propagation of CR extendibility 180
3.12. Separate real analyticity 186
Bibliography 193
 Preface

The geometry of the domains of holomorphic and CR functions is

one of the primary topics in contemporary mathematics. It is a field
where various techniques converge: PDE’s, real/complex vector fields,
induced systems, boundary values, analytic discs in symplectic spaces,
just to mention a few. The excellent interplay of these tools serves as a
model both for research and education. This book collects my lectures
addressed to graduate students at the University of Padova during
the last three years as well as an introduction to part of my present
research. The purpose is twofold: to provide a basic foundation in the
theory of several complex variables and CR functions and to give some
new insight and tools in current research.
Chapter 1 introduces the theory of complex analytic functions. It
starts with classical facts about power series and plurisubharmonic
functions along the guidelines of Oka and Hörmander. Next, it begins
the discussion of the equivalence between pseudoconvex domains and
domains of holomorphy and proves by geometric evidence the coinci-
dence of the pseudoconvexity of a domain with that of its boundary.
¯
Finally, it solves the ∂-Neumann problem by the method of Kohn.
This implies the solution of the Levi problem and provides the com-
plete proof of the equivalence mentioned above. It also includes some
new results concerning the regularity of ∂¯ at a weakly q-pseudoconvex
boundary.
Chapter 2 deals with real structures. Along with the classical
Frobenius theorem, it presents its symplectic counterpart, that is, the
Darboux-Frobenius theorem. Finally, it discusses subellipticity and
hypoellipticity of systems of real vector fields.
The content of Chapter 3 is CR structures. Particular emphasis is
devoted to the analysis of the conormal bundle to a real submanifold
of Cn under a canonical transformation. This serves to describe Hamil-
tonian/Levi foliations and also to interchange a higher-codimensional
submanifold with a hypersurface. Next, it recalls the theory of an-
alytic discs attached to real submanifolds and of their infinitesimal
deformations. Furthermore, it introduces a new family of discs with

vii
viii PREFACE

singular boundary. They yield a new proof and a refined statement of

the classical results on the extension of CR functions from manifolds
of finite type. The problem of the construction of lifts and partial lifts
of analytic discs is also addressed: the discussion fully discloses the
symplectic-geometric character of these objects. By means of them,
there is obtained in a natural way a “connection” on the conormal
bundle to a real submanifold of Cn . The coupling between lifts and
infinitesimal deformations is then exploited. It first shows that CR ex-
tendibility evolves along a CR curve according to the dual-inverse con-
nection of the above one. Also, when M is “minimal” in the sense of
Tumanov, it yields the wedge extension of CR functions. The discussion
ends with the theory of separate real analyticity, the real counterpart
of the Hartogs theorem, which takes a geometric explanation here.
I am greatly indebted to Luca Baracco and Alexander Tumanov for
having taught me much of the material developed here.

Padova, December 21, 2006

 CHAPTER 1

Several Complex Variables

Summary and Notes to Sections 1.1–1.6. These sections are

classical, so we comment about them all together. We recall in Sections
1.1 and 1.2 the definition and basic properties of analytic functions of
several complex variables which come from the extension of Cauchy’s
integral formula to product domains of Cn . We discuss in Section 1.3
the relation between the domains of holomorphic functions and of con-
vergent power series. This exhibits the first example of domains from
which any analytic function can be extended to a larger set. We then
discuss in Section 1.4 some basic properties of subharmonic functions
and prove in Section 1.5 Hartogs’s theorem: analyticity is separable
in each single variable. We end this discussion in Section 1.6 where
we introduce the Levi form, prove the normal form of a real hyper-
surface, and present the Hartogs extension theorem. It states that all
holomorphic functions extend from the concave side of a boundary.
We refer to the basic book by Gunning and Rossi [46] and to
Hörmander [50] from which a large part of the presentation is inspired.
The reader can find there a greater account of the material. We also
cite the books by Gunning [45], Krantz [63], Narasimhan [71], Range
[74] and Wells [98] where the reader can also test his knowledge of the
subject by clever exercises.

1.1. Analytic functions in complex spaces

We study here complex functions f : Cn → C. Being identified with
functions f : R2n → R2 , they are also particular real vector functions in
even dimension. But we are mainly interested in those √ which preserve
the complex structure. For a choice of a root i = −1 we write the
coordinates in Cn as

z = x + iy for (x, y) ∈ R2n ;

here z = (z1 ..., zn ), x = (x1 , ..., xn ), y = (y1 , ..., yn ). We have thus

∼
obtained an identification R2n → Cn given by

(1.1.1) (x, y) 7→ z = x + iy.

1
2 1. SEVERAL COMPLEX VARIABLES

In this correspondence, we write x = Re z and y = Im z; thus, the

inverse of (1.1.1) is written as z 7→ (Re z, Im z). If we want it
√ to look
“algebraic”, we need to call into play the other root −i = − −1 and
define by z̄ = x − iy the “conjugate”
√ of z. Then, in the pair (z, z̄), the
arbitrariness of the choice of ± −1 disappears, and we can describe
the real structure underlying Cn through the correspondence
(1.1.2) (x, y) 7→ (z, z̄) = (x + iy, x − iy)
whose inverse has now become linear
 
z + z̄ z − z̄
(1.1.3) (z, z̄) 7→ (x, y) = , .
2 2i
With the change (1.1.2) and its inverse (1.1.3) in hand, let us see how
derivatives and differentials behave. In general, associated to a change
of real coordinates t̃ = t̃(t) in Rm , the chain rule yields the trans-
P ∂ t̃
formation for the coordinate-derivatives ∂ti 7→ j=1,...,m ∂tji ∂t̃j . Hence,
having in mind the transformations (1.1.2) and (1.1.3), we define ∂z
and ∂z̄ by
(
∂z ∂ z̄
∂x = ∂x ∂z + ∂x ∂z̄ = ∂z + ∂z̄ ,
(1.1.4) ∂z ∂ z̄
∂y = ∂y ∂z + ∂y ∂z̄ = i(∂z − ∂z̄ ),
which reads explicitly, by inversion,
(
∂z = ∂x
∂z x
∂ + ∂y
∂z y
∂ = 12 (∂x − i∂y ),
(1.1.5)
∂z̄ = ∂x ∂ + ∂y
∂ z̄ x
∂ = 21 (∂x + i∂y ).
∂ z̄ y
 
∂z ∂zi
Here we use the notation ∂x = (∂xj )j , ∂x = ∂x j
and similarly for
ij
the other variables y, z and z̄. Thus, defined by (1.1.4) and (1.1.5),
the derivatives (∂z , ∂z̄ ) correspond to (∂x , ∂y ) via the chain rule under
the transformations (1.1.2) and (1.1.3). We have for the dual basis of
differentials the evident correspondences
(
dx = dz+dz̄
2
,
(1.1.6) dz−dz̄
dy = 2i ,
with inverse
(
dz = dx + idy,
(1.1.7)
dz̄ = dx − idy.
Again, we write dx for (dxj )j and similarly for the other variables.
In particular, for a function f : Cn → C identified with a function
 1.1. ANALYTIC FUNCTIONS IN COMPLEX SPACES 3

f : R2n → C, we have
(1.1.8) ∂x f dx + ∂y f dy = ∂z f dz + ∂z̄ f dz̄
that we also write as
(1.1.9) ¯
df = ∂f + ∂f.
Equation (1.1.9) serves as a definition ¯ . Note that in
of ∂f and ∂f
P

(1.1.8) ∂x f dx + ∂y f dy stands for j ∂xj f dxj + ∂yj f dyj and ∂z f dz +
P

∂z̄ f dz̄ stands for j ∂zj f dzj + ∂z̄j f dz̄j . Equation (1.1.8) states that
the system of derivatives ∂z and ∂z̄ is the dual system to the system of
differentials dz and dz̄. It is often used, in alternative to (1.1.4), (1.1.5),
to define ∂z and ∂z̄ from dz and dz̄.
√
Remark 1.1.1. The length 2 of dzj and dz̄j , which is imposed by
the notation, forces the dual system ∂zj and ∂z̄j to have the unnatural
length √12 ; this is why the chain rule is preferable to duality in their
definition.
We have to notice that in case f is real valued then (1.1.8) yields
(1.1.10) df = 2Re ∂f.
There is no doubt that (1.1.10) is great: it tells us that the “real” dif-
ferentials are the real parts of the “complex” differentials. It serves the
intrinsic version of the identification R2n ' Cn between the euclidean
structure and the real part of the hermitian structure: xx̃ + y ỹ = Re z z̃.
As for (1.1.9), it says that the complexified differentials, that is, the set
of the combinations of the dxj ’s and dyj ’s with complex coefficients, are
identified with the combinations of the dzj ’s and dz̄j ’s. Also, (1.1.10)
identifies real differentials with “diagonal” combinations, in which the
coefficients of the dz̄j ’s are the conjugates of those of the dzj ’s. We will
come back to this point in Chapter 3.
Differential forms which are combinations of the dzj ’s are said to
be of type (1, 0) and those which are combinations of the dz̄j ’s of type
(0, 1). Thus, ∂f and ∂f ¯ are the components of df of type (1, 0) and
(0, 1), respectively. Let Ω be a domain in Cn .
Definition 1.1.2. A function f ∈ C 1 (Ω) is said to be holomorphic
or complex analytic when df = ∂f , that is, when f satisfies the differ-
ential system ∂z̄j f = 0, j = 1, ..., n. The set of holomorphic functions
in Ω is denoted by hol(Ω).
The above system is called the Cauchy-Riemann system (CR system
for short). Thus, df , which is in general R-linear, is in this case C-linear,
4 1. SEVERAL COMPLEX VARIABLES

that is, we have the commutative diagram

df
R2n → R2
(1.1.11) || ||
∂f
Cn → C.
If we decompose f = Re f + iIm f , the CR system is a system of 2n
real equations
(1.1.12) ∂xj Re f − ∂yj Im f = 0, ∂yj Re f + ∂xj Im f = 0, j = 1, ..., n.
The elliptic estimates for the system of the ∂z̄j ’s will tell us in
Section 1.10 that the generalized solutions—in any possible sense of
the distributions or even of the hyperfunctions—are indeed “classical”
solutions. In other words, to be C 1 is a consequence of being a solution
of the system ∂z̄j , j = 1, ..., n. But we can say more: any function which
is “separately” holomorphic, that is, holomorphic with respect to each
zj when the other arguments are kept fixed, is in fact holomorphic.
Also, by the elliptic regularity of each single ∂z̄j , it suffices that each
restriction is just a generalized solution. Note that a separate solution
is not even a “weak” solution, in general. Thus, complex analyticity is
“separable” in contrast to any kind of regularity of real functions.
We now treat the inversion of holomorphic functions. We start from
an elementary remark. A holomorphic function of a single variable
f (z), z ∈ C, such that ∂z f (zo ) 6= 0 at some zo ∈ C is locally invertible
because df = ∂f 6= 0 at zo . (To show off our ability
 in the
2 R/C switch,
we can invoke that det ∂(x,y) (Re f, Im f )|zo = ∂z f (zo ) 6= 0.) On the
other hand, differentiation in z̄ and z of the identity f −1 ◦ f = id yields
∂z̄ f −1 ∂z f = 0 and ∂z f −1 ∂z f = 1.
Thus, f −1 is holomorphic and ∂z f −1 = (∂z f )−1 . This has an immediate
generalization to Cn .
Theorem 1.1.3. Let f (z) = (fj (z))j=1,...,n be analytic functions in
a neighborhood of zo for which we assume that
det (∂zi fj )ji 6= 0 at zo .
Then, there exists g(z) = (gj (z))j=1,...,n analytic in a neighborhood of
wo = f (zo ) such that g ◦ f = id.
Proof. We “double” our variables from z ∈ Cn to (z, w) ∈ Cn ×Cm
and prove a more general statement. For an analytic function h : Cn ×
Cm → Cn , h = (hj (z, w))j , such that h = 0 and det (∂z h) 6= 0 at
(zo , wo ), the equations (hj (z, w))j=1,...,n = 0 have uniquely determined
analytic solutions z = (zj )(w) in a neighborhood of wo with z(wo ) = zo .
 1.2. CAUCHY FORMULA IN POLYDISCS 5

This implies the theorem when applied for m = n and hj = wj − fj (z).

To prove the claim, we have to show that dhj = 0 and dwk = 0 for any
j = 1, ..., n and k = 1, ..., m, imply dzj = 0, j = 1, ..., n: but this follows
from det ∂z h 6= 0. Thus, by the implicit function theorem, the equations
hj = 0 define uniquely the solutions z = (zj )(w) with z(wo ) = zo . For
these solutions z(w), the identity dh = ∂z hdz +∂w hdw = 0 implies that
the dzj ’s are combinations of the dwk ’s; thus the zj ’s are holomorphic.


1.2. Cauchy formula in polydiscs

Let z o = (z1o , ..., zno ) be a point of Cn and r = (r1 , ..., rn ) a “multi-
radius” in (R+ )n .
Definition 1.2.1. Let ∆zjo rj be the disc {zj ∈ C : |zj − zjo | < rj }
of center zjo and radius rj . We define
Pz o r = Π ∆zjo rj
j=1,...,n

and call it the “polydisc” of center z o and multiradius r.

The boundary ∂P is stratified by strata of codimension 1, 2, ... which
correspond to single equations |zj1 − zjo1 | = rj1 , pairs of equations
|zj1 − zjo1 | = rj1 , |zj2 − zjo2 | = rj2 and so on. The highest codimensional
(n-codimensional) stratum is called the “distinguished” or “Shilov”
boundary of P and is denoted by
∂0 P = {|zj − zjo | = rj , j = 1, ..., n}.
Holomorphic functions on polydiscs admit the Cauchy integral repre-
sentation on ∂0 P . This also enables us to start our process of relaxing
the initial assumption of C 1 -regularity for f .
Theorem 1.2.2. Let f be a continuous function on a polydisc P̄
which is, for any j, a holomorphic function of zj (that is, a C 1 solution
of ∂z̄j f = 0) when the other variables zk for k 6= j are fixed. Then for
any z ∈ P we have
Z
−n f (ζ)
(1.2.1) f (z) = (2πi) dζ1 ∧ ... ∧ dζn .
∂0 P (ζ1 − z1 ) · ... · (ζn − zn )

Proof. We proceed by induction on the dimension n of Cn and

first treat the case n = 1. To simplify notation, we assume z o = 0 and
r = 1; thus P coincides with the standard disc ∆. Under this choice,
we want to prove the general Cauchy integral formula, which applies
6 1. SEVERAL COMPLEX VARIABLES

to a function g ∈ C 1 not necessarily holomorphic and that we state as

a separate lemma.
¯ for any z ∈ ∆ we have
Lemma 1.2.3. Let g ∈ C 1 (∆);
Z 
∂ζ̄ g(ζ)
ZZ
1 g(ζ)
(1.2.2) g(z) = dζ + dζ ∧ dζ̄ .
2πi ∂∆ ζ − z ∆ ζ −z

Proof. For z ∈ ∆, let  < dist(z, ∂∆); thus ∆ := ∆z  satisfies

¯  ⊂⊂ ∆. We note that ∂∆ and ∂∆ , as parts of the boundary of
∆
∆ \ ∆ , carry opposite orientation: +∂(∆ \ ∆ ) = +∂∆ − ∂∆ . We use
the notation ω for the 1-form g(ζ)
ζ−z
dζ; Stokes’s formula yields
Z Z ZZ 
1 1
(1.2.3) ω= ω− dω .
2πi +∂∆ 2πi +∂∆ ∆\∆

∂ζ̄ g(ζ)
We note that the last integrand in (1.2.3) is −dω = ζ−z
dζ ∧ dζ̄. Since
dω is integrable at its singular point z, then
ZZ ZZ
dω → dω.
∆\∆ ∆

We also note that Z

1
ω → g(z),
2πi +∂∆

because g is continuous at z. Thus, (1.2.3) at the limit for  → 0 turns

into (1.2.2).


In particular, if instead of a general C 1 function g we take a solution

f of ∂z̄ f = 0, (1.2.2) yields
Z
1 f (ζ)
f (z) = dζ,
2πi ∂∆ ζ − z
that is, (1.2.1) for n = 1.
End of proof of Theorem 1.2.2. We carry on our inductive
argument and suppose that (1.2.1) has been proved for holomorphic
functions of n − 1 arguments z1 , ..., zn−1 . We denote by P 0 the polydisc
Π ∆zjo rj and write (1.2.1) for z1 , ..., zn−1 with parameter zn :
j≤n−1
(1.2.4)
 n−1Z
1 f (ζ1 , ..., ζn−1 , zn )
f (z1 , ..., zn ) = dζ1 ∧....∧dζn−1 .
2πi ∂0 P 0 (ζ1 − z1 ) · ... · (ζn−1 − zn−1 )
 1.2. CAUCHY FORMULA IN POLYDISCS 7

But in the above integral we can make the substitution which we are
allowed by the 1-dimensional Cauchy formula:
Z
1 f (ζ1 , ..., ζn )
(1.2.5) f (ζ1 , ..., ζn−1 , zn ) = dζn .
2πi |ζn −zn |=rn ζn − zn
From (1.2.4), (1.2.5) we get (1.2.1) by Fubini’s theorem.

Remark 1.2.4. The separate C 1 regularity in each zj is needed
for Stokes’s formula, and the joint C 0 regularity is needed for Fubini’s
theorem.
For a function f ∈ C 0 (Ω) separately (C 1 and) holomorphic in each
zj , we wish to apply (1.2.1) to all possible P̄ ⊂⊂ Ω. Since the functions
of z produced by the integrals in ζ are C ∞ , as one can check by bringing
derivatives in the zj ’s inside the convergent integral, we get at once
Corollary 1.2.5. If f is C 0 on Ω and separately holomorphic in
each zj , then it is C ∞ and in particular jointly holomorphic.
We write the point ranging in ∂0 P (z, r) as ζ = (ζj ) for ζj = zj +
iθj
rj e . Then we have
Z
−n f (ζ)
(2πi) dζ1 ∧ ... ∧ dζn
∂0 P (z,r) (ζ1 − z1 ) · ... · (ζn − zn )
Z
= f (z + (r1 ei2πθ1 , ..., rn ei2πθn ))dθ1 ∧ ... ∧ dθn .
[0,1]n

In particular, (1.2.1) says that a holomorphic function satisfies the

mean property: the value of f at z equals its mean value on the dis-
tinguished boundary of any polydisc contained in the region where f
is holomorphic. Passing from f to |f | in (1.2.1), we see that |f (z)| is
estimated by the mean value of |f | over ∂0 P (z, r): we say that |f | en-
joys the submean property, in this case. In particular, if |f | attains a
maximum at a point z of the domain Ω where f is holomorphic, then
|f | must be constant with the same value |f (z)| in ∂0 P (z, r) for any
P (z, r) ⊂⊂ Ω. Thus, |f | is constant in a neighborhood of z. (In fact f
itself is constant since |∂z f |2 = ∂z ∂z̄ |f |2 = 0.) It follows that f satisfies
the so-called “maximum modulus principle”, a property that absolute
values of holomorphic functions share with all functions satisfying the
submean property:
Theorem 1.2.6. Let Ω be bounded and f be continuous on Ω̄ and
holomorphic on Ω. Then |f | attains its maximum on ∂Ω.
8 1. SEVERAL COMPLEX VARIABLES

Proof. The function f is locally constant at any point in the in-

terior of Ω where |f | attains its maximum. So the set of these point,
being also closed, cannot be non-empty unless it is a full connected
component of Ω. But then the same maximum value for |f | is attained
at some boundary point.

We need to introduce now the notions of multiindices, multipowers
and multiderivatives. For a multiindex α = (α1 , ..., αn ) ∈ Nn we put
α! = α1 ! · ... · αn ! and |α| = α1 + α2 + ... + αn . We define multipowers by
rα = r1α1 · ... · rnαn and multiderivatives by ∂zα f = ∂zα11 . . . ∂zαnn f that we
also denote sometimes by f (α) . We write α + 1 for (α1 + 1, ..., αn + 1)
and, if β is another multiindex such that βj < αj for any j, we define
α − β := (α1 − β1 , ..., αn − βn ). P
We say that a power series α aα (z − zo )α converges normally in
the polydisc P (zo , r) when for any r0 with rj0 < rj , for all j we have
0α α (β)
P P
α |aα |r < +∞. For any fixed β, we denote P α! α aα (z − zα−β
by ( o) )
the series of β-derivatives, that is, the series a (z − zo ) ; it
(α−β)! α
α≥β
is a routine application of Abel’s theorem to show that if a power series
centered at zo converges in the polydisc P (zo , r), then its derivative of
any index β converges in P (zo , r). By this remark we can realize any
power series as a holomorphic function wherever it converges.
Theorem 1.2.7. Let α aα (z − zo )α converge in P (zo , r); then its
P
sum f is a holomorphic function on P (zo , r).
aα (z −zo )α converge uniformly
P
Proof. The partial sums Sν :=
|α|≤ν
together with their derivatives in any compact subset of P . In particu-
lar, the relation ∂z̄j Sν = 0 for any j passes to the limit and turns into
∂z̄j f = 0.

We want to prove now the converse statement: any holomorphic
function on Ω is, in a neighborhood of any point zo ∈ Ω, the sum of
a power series convergent in a polydisc P (zo , r), necessarily its own
Taylor series at zo .
Theorem 1.2.8. Let f ∈ hol(Ω) and let zo ∈ Ω. Then
X f (α) (zo )
f (z) = (z − zo )α ,
α
α!
with normal convergence in any polydisc P (zo , r) ⊂ Ω.
 1.2. CAUCHY FORMULA IN POLYDISCS 9

Note that the choice of the polydisc is arbitrary provided that it is

contained in Ω. If one needs to stress attention to a specific variable,
say zj , it is convenient to stretch the polydisc along that direction so
that rj approaches the distance of zo to ∂Ω along the zj -plane at the
expenses of the other radii rk for k 6= j which possibly approach 0.
Proof. We choose zo = 0. We write
1 1 1
= z1
(ζ1 − z1 ) · ... · (ζn − zn ) ζ1 · ... · ζn (1 − ζ1 ) · ... · (1 − zζnn )
(1.2.6) X zα
= α+1
,
α
ζ
 
z 
with normal convergence when  ζjj  < 1 for any j. In particular, the
convergence is absolute and uniform for z ranging on compact subsets
of P and ζ on ∂0 P . Thus, we substitute (1.2.6) into Cauchy’s formula
and get
Z
−n f (ζ)
f (z) = (2πi) dζ1 ∧ ... ∧ dζn
∂0 P (ζ1 − z1 ) · ... · (ζn − zn )
Z X α
−n z
(1.2.7) = (2πi) α+1
f (ζ)dζ1 ∧ ... ∧ dζn
∂0 P α ζ
X Z
f (ζ)

−n
= (2πi) α+1
dζ1 ∧ ... ∧ dζn z α ,
α ∂0 P ζ

with normal convergence for z ∈ P . Thus, f is the sum of a power series

centered at zo = 0. But iterated derivation of the identity of f with
the sum of the series and evaluation at zo yields that the coefficients in
(α)
(1.2.7) are indeed the Taylor coefficients of f at zo = 0, that is, f α!(0) .

The proof of the theorem gives in fact the “Fourier integral” ex-
pression of the coefficients:
f (α) (zo )
Z
−n f (ζ)
(1.2.8) = (2πi) α+1
dζ1 ∧ ... ∧ dζn .
α! ∂0 P (ζ − zo )

Now that we have described locally an analytic function as the sum of

a convergent power series, we realize that the zeroes of f “propagate”
from an open part of Ω to the whole of its connected component.
Corollary 1.2.9 (Analytic continuation). Let Ωo ⊂ Ω be open
subsets of Cn with Ωo 6= ∅, Ω connected and let f ∈ hol(Ω) with f |Ωo ≡
0. Then f ≡ 0 all over Ω.
10 1. SEVERAL COMPLEX VARIABLES

In other words, two holomorphic functions on Ω which coincide over

an open non-empty set must coincide everywhere.
Since in a neighborhood of any point zo ∈ Ω, f is the sum of its
Taylor series, then in fact the corollary says that
f (α) (zo ) = 0 for any α implies f ≡ 0 on Ω.
Proof. Let Ω1 be the maximal open subset of Ω where f |Ω1 ≡ 0.
(“Maximal” makes sense: just take the union of all such subsets.) It is
non-empty since it contains Ωo . We must have Ω1 = Ω. If not, let ξ be
a point in Ω1 whose distance to ∂Ω is bigger than its distance to ∂Ω1 :
such a ξ must exist because of the connectedness of Ω. Now, the Taylor
series of f at ξ is 0 and it represents f in any polydisc contained in Ω,
in particular in those stretched along the plane in which the distance
of ξ to ∂Ω1 is attained. But these polydiscs go out of Ω1 and bring the
zeroes of f beyond their maximal set.

In the following statement there is special emphasis on the center zo
and the multiradius r of P = P (zo , r): the bigger the radii of the discs
that we can insert in the domain where f is holomorphic, the better
the estimates of its derivatives.
Theorem 1.2.10 (Cauchy’s inequalities). Let f be holomorphic in
P and continuous on P̄ . Then, for any α and for suitable c = cα r
α!
(1.2.9) |f (α) (zo )| ≤ sup |f |,
rα ∂0 P (zo ,r)

(1.2.10) |f (α) (zo )| ≤ c||f ||L1 (P (zo ,r)) .

Proof. Write ζj = zo j + rj eiθj ; then (1.2.8) becomes
f (α) (zo ) f (zo + reiθ ) −iθα
Z
−n n
= (2πi) i e dθ1 ∧ ... ∧ dθn .
α! [0,2π]n rα
Taking absolute values, we get
Z
 (α)  α! −n
(1.2.11)  f (zo ) ≤ α (2π)
 |f (zo + reiθ )|dθ1 ∧ ... ∧ dθn ,
r [0,2π] n

which yields (1.2.9). By moving the multiradius r0 with the constraint

0 ≤ r0 ≤ r, multiplying both sides of (1.2.11) by r0α+1 and integrating
with respect to r0 , we get (1.2.10).

 1.3. ANALYTIC FUNCTIONS AND POWER SERIES 11

In particular, in the space hol(Ω) the C ∞ convergence coincides with

the L1loc convergence, indeed with any “weak” (except from pointwise)
convergence. But the main consequence is about the so-called “normal
families” of holomorphic functions.
Theorem 1.2.11 (Stjelties-Vitali). Let {fν } be a bounded sequence
in hol(Ω). Then, there is a subsequence {fνk } uniformly convergent on
compact subsets of Ω to a limit f ∈ hol(Ω).
Proof. By Theorem 1.2.10, “{fν } bounded” implies “{∂zj fν }
bounded” on compact subsets. Hence {fν } is equicontinuous and by
Arzela’s theorem it converges to a limit. The limit is also holomorphic,
just by interchanging ∂z̄j with “lim”.


1.3. Analytic functions and power series

The domain of convergence of a power series of one complex variable
is a disc. For a power series in several complex variables it shows a more
interesting structure. It also serves as the first non-trivial example of
a “forced” extension for holomorphic functions which is not possessed
by functions of one variable. Let
+∞
X
(1.3.1) aα z α
|α|=0

be a power series centered at 0.

Definition 1.3.1. We denote by D the domain of convergence of
the power series (1.3.1) which consists of the set of points in whose
neighborhood the power series converges normally, that is, absolutely
uniformly.
Definition 1.3.2. We denote by B the set of “boundedness” of
(1.3.1), that is, the set of points z such that |aα z α | < c for any α.
The two above sets are related by the following theorem, which is
an elementary application of Abel’s theorem.
◦
Theorem 1.3.3. We have D = B.
◦
Proof. D is obviously open and contained in B. Conversely, if
◦
zo ∈ B, then there are a point w ∈ B, positive numbers kj < 1 and
a neighborhood V of zo such that |zj | ≤ kj |wj | for any j and z ∈ V .
12 1. SEVERAL COMPLEX VARIABLES

Let k = (k1 , ..., kn ); it follows that

X X
|aα z α | ≤ c kα
(1.3.2) α α
= c Π (1 − kj )−1 for any z ∈ V.
j=1,...,n


Therefore, the domain of convergence D of a power series shows its
first two interesting features:
• It is invariant under rotations: if z ∈ D, then eiθ z ∈ D for any
θ ∈ [0, 2π]n .
• If w ∈ D, then, for any z which satisfies |zj | ≤ |wj | for all j,
we also have z ∈ D.
An open set which enjoys the first property is called a “Reinhardt do-
main”: it can be described through its representative in the space of
moduli |D| ⊂ (R+ )n defined by |D| := {(|z1 |, ..., |zn |) : z ∈ D}. A Rein-
hardt domain which enjoys the second property is called “complete”: it
is therefore characterized as a union of polydiscs centered at 0. There-
fore, for a domain of Cn to be the domain of convergence of a power
series, we have to require it to be Reinhardt and complete. But not all
complete Reinhardt domains are precisely the convergence domain of a
power series. Some additional geometric requirements must be fulfilled.
Theorem 1.3.4. D is “logarithmically convex” in the sense that the
set D∗ := {ξ ∈ Rn : ξj = log|zj |, j = 1, ..., n, for z ∈ D} is convex.
◦
Proof. Remember that D = B; hence, what we have to prove is
that
(
|aα z α | < c,
(1.3.3)
|aα wα | < c
implies for λ ∈ [0, 1]
(1.3.4) |aα z λα w(1−λ)α | < c.
But this is a straightforward calculation.

We compare now analytic functions and convergent power series.
We have already seen that they coincide over polydiscs, hence over the
union of those which are centered at the same point: this follows easily
from analytic continuation. In particular, over a Reinhardt complete
(non-empty) domain, a holomorphic function is always represented as
a sum of a power series centered at 0. (Of course for a general domain
 1.3. ANALYTIC FUNCTIONS AND POWER SERIES 13

Ω ⊂ Cn , a holomorphic function is a family of convergent power series

but not a single one, in general.) We see now that a Reinhardt domain
need not be complete for this conclusion to be true.
Theorem 1.3.5. Let Ω be a connected Reinhardt domain which
contains 0 and let f ∈ hol(Ω); then we have
X f (α) (0)
f (z) = zα,
α
α!
with normal convergence in Ω.
Proof. We denote by Ω the connected component of 0 in the set
{z ∈ Ω : d(z, Cn \ Ω) > |z|}. The sets Ω are increasing for  & 0 and
their union over  covers the whole of Ω since this is connected. We
denote by P1+ the polydisc P1+ = {t ∈ Cn : |tj | < 1 +  for any j =
1, ..., n} and define
Z
−n f (t1 z1 , ..., tn zn )
(1.3.5) g(z) = (2πi) dt1 ∧ ... ∧ dtn .
∂0 P1+ (t1 − 1)...(tn − 1)

Since z ∈ Ω , then (1 + )z ∈ Ω; since t ∈ ∂0 P1+ and Ω is Reinhardt,

then tz ∈ Ω. Thus, the integral is defined and produces a smooth
function of z. By interchanging ∂z̄j with integration, we see that g
is holomorphic. When t leaves ∂0 P1+ and enters into P1+ , then tz
is no longer contained in Ω, in general. But this is true for small z
since Ω contains a neighborhood of 0. For these values of z, the sub-
z
stitutions ζ := tz and P1+ = {tz : t ∈ P1+ } inside (1.3.5) yield
−n
R f (ζ)
g(z) = (2πi) z
∂0 P1+ (ζ1 −z1 )...(ζn −zn )
dζ1 ∧ ... ∧ dζn . But the term in the
z
right side is the Cauchy integral of f at z for the polydisc P1+ which
is completely contained in the region where f is holomorphic. Hence
f (z) = g(z) for those values of z, but then in fact f ≡ g on Ω by
analytic continuation. We turn our attention now to the power series
1 X
= t−α−1 ,
(t1 − 1)...(tn − 1) α
P
with normal convergence for t ∈ ∂0 P1− . If we take P α out of the
integral on the right side of (1.3.5), we get f = α fα where the fα
are the analytic functions defined by
Z
−n f (tz)
fα (z) = (2πi) α+1
dt1 ∧ ... ∧ dtn .
∂0 P1+ t

In the same way as for the function g previously discussed, we can see
(α)
that fα (z) = f α!(0) z α on Ω . But then f is a sum of a power series.

14 1. SEVERAL COMPLEX VARIABLES

Since a holomorphic function on a Reinhardt domain containing 0

is represented by a power series, then the function extends analytically
wherever the power series converges. Now, the first points to be added
to the initial domain are those of its “completion”, the points z such
that for some w ∈ Ω we have |zj | ≤ |wj | for any j and then also the
points of its “logarithmically convex hull”, the points z for which there
are w1 and w2 in Ω such that log |zj | = λ log |wj1 | + (1 − λ) log |wj2 |
for any j ≤ n and for λ ∈ [0, 1]. All these points were unnaturally
“missing” from the initial domain of f . These remarks are collected in
the following statement.
Theorem 1.3.6. Let Ω be a connected Reinhardt domain containing
0 and let Ω̃ be the smallest complete logarithmically convex Reinhardt
domain containing Ω. Then we have an extension mapping
ext
hol(Ω) 99K hol(Ω̃).
Remark 1.3.7. We have purposely included the pleonastic word
“complete” in the above statement: but a Reinhardt domain logarith-
mically convex and containing 0 is always complete. In fact, let w ∈ Ω:
we know that λ |w|1−λ := (λ1 |w1 |1−λ , ..., λn |wn |1−λ ) must belong to |Ω|
for any real small  and for any λ ∈ [0, 1]. But the set of these points
whose image in the space of moduli is λ |w|1−λ covers the whole poly-
disc {z : |zj | ≤ |wj | for any j}.
We conclude this discussion by remarking that, conversely, Ω̃ is the
maximal set of analytic extension or, in other words, there is a holomor-
phic function which does not extend beyond Ω̃. This is a consequence
of Theorem 1.8.8 whose proof can be found in detail for instance in [50,
Corollary 2.5.8]. On the other hand this function is indeed represented
by a unique power series. Thus, a Reinhardt, complete, logarithmically
convex domain is the precise domain of convergence of a power series.
1.4. Subharmonic functions
It is clear from the end of the last section that our most attractive
task is now to describe the natural “domains” of holomorphic functions.
For this purpose we need to develop some preliminary technology; it
will also be the main tool in solving the problem of separate analyt-
icity. Let us declare from the very beginning that this is strictly a
one-variable theory. We recall that a C 2 function h on a domain Ω of
C with coordinate z = x + iy is said to be harmonic when ∂z ∂z̄ h = 0.
Definition 1.4.1. A real function ϕ on Ω ⊂ C with values in
[−∞, +∞) is subharmonic when
 1.4. SUBHARMONIC FUNCTIONS 15

(i) ϕ is upper semicontinuous, i.e., for any zo , ϕ(zo ) ≥ lim sup ϕ(z);
z→zo
(ii) for any K ⊂⊂ Ω and for any h continuous on K and har-
◦
monic on K,
(1.4.1) ϕ|∂K ≤ h|∂K implies ϕ|K ≤ h|K .
In particular, by taking as h the constant function on K with value
max ϕ, which exists and is finite by the upper semicontinuity of ϕ,
∂K
we have that subharmonic functions satisfy the maximum principle.
We denote by SH(Ω) the set of subharmonic functions on Ω. A func-
tion ϕ is said to be superharmonic when −ϕ is subharmonic. The
harmonic functions are precisely those which are at the same time
sub/superharmonic. This is always true once we know that any contin-
uous function on ∂∆ is endowed with a unique harmonic extension to
∆ (given by the Poisson integral as explained in Section 3.7). An easier
argument for C 2 functions is based on Theorem 1.4.9.
Subharmonicity does not appear to be a local property at first sight
but it turns out to be. Let ∆ be the standard disc in C and ∆zo r the
disc of center zo and radius r. Here is the main characterization.
Theorem 1.4.2. Let ϕ : Ω → [−∞, +∞) be an upper semicontin-
uous function. The following are equivalent,
(i) ϕ ∈ SH(Ω).
(ii) For any disc ∆zo r ⊂⊂ Ω and for any polynomial P = P (z),
ϕ|∂∆zo r ≤ Re P |∂∆zo r implies ϕ|∆¯ zo r ≤ Re P |∆¯ zo r .
(iii) (spherical submean) For any ∆ ¯ zo r ⊂⊂ Ω, we have
Z
ϕ(zo ) ≤ (2πr)−1 ϕds,
∂∆zo r

where ds is the element of the arc.

(iv) (solid submean) For any ∆ ¯ zo r ⊂⊂ Ω, we have
ZZ
2 −1
ϕ(zo ) ≤ (−2iπr ) ϕdτ ∧ dτ̄ .
∆zo r

(iv)loc (Local solid submean) For any zo there is ro < dist(zo , ∂Ω)
such that for any r ≤ ro , (iv) holds.
Proof. (i) ⇒ (ii) and (iv) ⇒ (iv)loc are obvious, whereas (iii) ⇒
(iv) is immediate by integration on dr.
(ii) ⇒ (iii): We take a continuous function f (θ) ≥ ϕ(zo + reiθ ) and
+N
ak eikθ
P
approximate f by a real trigonometric polynomial g(θ) =
k=−N
16 1. SEVERAL COMPLEX VARIABLES

with a−k = āk :

(1.4.2) ϕ(zo + reiθ ) ≤ f (θ) ≤ g(θ) ≤ f (θ) + .
N
ak ( z−z o k
P
We write P (z) = a0 + 2 r
) ; thus g = Re P |∂∆zo r and therefore
k=1
(1.4.2) can be rewritten as
ϕ|∂∆zo r ≤ Re P |∂∆zo r .
¯ zo r , in particular at
By (ii) the inequality must hold in the whole of ∆
the center zo which yields
ϕ(zo ) ≤ Re P (zo )
(1.4.3)
Z 2π
1
= a0 = g(θ)dθ,
2π 0

where the last equality follows from the fact that all non-constant har-
monics vanish when integrated over complete cycles. Finally, by the
third inequality of (1.4.2), we have
Z 2π Z 2π
1 1
(1.4.4) g(θ)dθ ≤ f (zo + reiθ )dθ + .
2π 0 2π 0
Combining (1.4.3) and (1.4.4) and then taking infimum in f and  we
get (iii).
◦
(iv)loc ⇒ (i): Let K ⊂⊂ Ω and let h be harmonic on K, continuous
on K and such that
(1.4.5) ϕ|∂K ≤ h|∂K .
Let M := max (ϕ − h); we reason by contradiction and assume M > 0.
K
We define the function v := ϕ − h and the set F := v −1 (M ): v is upper
◦
semicontinuous and F is a compact subset of K because of (1.4.5).
Let zo be a point of ∂F . Then, there is some disc ∆(zo , r) such that
∆(zo , r) ∩ (K \ F ) 6= ∅. Also, we can suppose that r is taken so small
that (iv)loc holds for such a zo and r. It follows that
M = v(zo )
Z 2π Z r
2 −1
≤ (πr ) tvdt dθ.
0 0

But the mean value on the right hand side is taken over a disc which
has an open portion outside F , the set of maximum value M for v.
Hence this mean value must be < M , a contradiction.

 1.4. SUBHARMONIC FUNCTIONS 17

Example 1.4.3. If ϕ is not upper semicontinuous, submean is not

equivalent to local submean. For example, the function ϕ on the disc
∆ which is 0 for Re τ ≤ 0 and 1 for Re τ > 0 satisfies (iii)loc and (iv)loc ,
that is (iii) and (iv) in local version, but neither (iii) nor (iv). On the
other hand, if we force ϕ to be upper semicontinuous by giving it value
1 at Re τ = 0, it looses the local submean property. In general, apart
from subharmonic functions, upper semicontinuity and the submean
property are in contrast to one another.
When ϕ is upper semicontinuous, Theorem 1.4.2 shows that any
characterization of subharmonicity can be expressed in local terms:
subharmonicity reveals therefore its local character. Also, on account
of the submean property, subharmonicity is stable under summation,
multiplication by positive constants and “sup” of arbitrary families
provided that the “sup” is already known to be < +∞ and upper
semicontinuous. We also have a similar statement for the “lim sup”
though its proof is a little less evident.
Proposition 1.4.4. Let {ϕν } be subharmonic and uniformly
bounded. Then lim sup ϕν is subharmonic if it is upper semicontinu-
ν
ous.
Proof. For any z we have
Z 2π
1
(1.4.6) lim sup ϕν (z) ≤ lim sup ϕν (z + reiθ )dθ,
ν 2π 0 ν
because of the subharmonicity of the ϕν ’s and Fatou’s lemma.

Remark 1.4.5. If lim sup ϕν is not known to be upper semicon-
ν
tinuous, nevertheless it is integrable and still satisfies (1.4.6); thus, it
always enjoys the submean property.
Hence, apart from a minor adjustment, subharmonicity is stable
under “sup” and “lim sup”. It is also stable under monotonic “inf”:
Lemma 1.4.6. If {ϕν }ν is a decreasing sequence of subharmonic
functions, then ϕ = inf ϕν is also subharmonic.
ν
S
Proof. We note that for any c we have {z : ϕ(z) < c} = ν {z :
ϕν (z) < c} which is an open set: hence ϕ is upper semicontinuous. If
now h|∂K ≥ ϕ|∂K and  > 0, then Kν := {z ∈ ∂K : ϕν (z) ≥ h(z) + }
is compact and decreasing. Since the limit for ν → +∞ is empty, then
Kν must be empty for large ν. Hence ϕν |K ≤ h|K +  for these values
of ν.
18 1. SEVERAL COMPLEX VARIABLES


Subharmonicity is preserved under composition with convex in-
creasing functions.
Theorem 1.4.7. Let χ be a function from R to R such that χ̇ ≥
··
0 and χ ≥ 0 (extended at −∞ by χ(−∞) := lim χ(t)). Let ϕ be
t→−∞
subharmonic; then χ(ϕ) is also subharmonic.
Proof. Since χ is convex, then for any to and for suitable c, we
have
(1.4.7) χ(to ) ≤ χ(t) − c(t − to ).
Replacing t with ϕ(z + reiθ ) in (1.4.7) and integrating over θ, we get
(1.4.8) Z 2π Z 2π 
−1 iθ iθ
χ(to ) ≤ (2π) χ(ϕ)(z + re )dθ − c( ϕ(z + re )dθ − 2πto ) .
0 0
−1 2π
R
We choose now to = (2π) ϕ(z + reiθ )dθ and get
0
 Z 2π 
−1 iθ
χ(ϕ(z)) ≤ χ (2π) ϕ(z + re )dθ
0
Z 2π
−1
≤ (2π) χ(ϕ)(z + reiθ )dθ,
0
where the first inequality follows from the subharmonicity of ϕ and
the fact that χ is increasing and the second from (1.4.8) under the
already-mentioned choice of to .

We have said that subharmonicity is preserved under composition
with a convex increasing χ but it is indeed strengthened: χ(ϕ) is “more”
subharmonic than ϕ is and how much more depends on how convex χ
is. According to [43], ϕ is said to be “strongly subharmonic” when there
exists χ positive, increasing, strictly convex, such that χ−1 (ϕ) is still
subharmonic. Thus, for instance, if ψ is subharmonic and positive, then
any power ψ a for a > 1 is strongly subharmonic. It is another useful
exercise to prove that if ϕ is C 2 , strongly subharmonic and satisfies
¯ > 0.
∂ϕ 6= 0, then ∂∂ϕ
We recall now that it is proved in [43] that if ϕ is strongly sub-
harmonic and has a finite harmonic majorant in the unit ball, then its
boundary measure is absolutely continuous: just plain subharmonicity
would not suffice for this conclusion. A very intense case, corresponding
to χ = exp, occurs when ϕ is logarithmically subharmonic, that is, log ϕ
is subharmonic. Let us further discuss logarithmic subharmonicity. We
 1.4. SUBHARMONIC FUNCTIONS 19

begin by the obvious remark that for f ∈ hol(Ω), the mean property
of f , and hence the submean of |f |, implies that |f | ∈ SH(Ω). We can
say more.
Lemma 1.4.8. If f ∈ hol(Ω), then log|f | ∈ SH(Ω).
Proof. Let K be a compact of Ω and P a polynomial such that
 
(1.4.9) log |f |∂K ≤ Re P ∂K .
We apply “exp” and observe that eRe P = |eP |; hence (1.4.9) yields
 
 f 
(1.4.10)   ≤ 1.
 eP  
∂K
By the maximum principle, (1.4.10) remains true when passing from
∂K to K and next, the reverse of the implication from (1.4.9) to (1.4.10)
yields  
log |f |K ≤ Re P K .
Thus, log |f | is subharmonic by (ii) of Theorem 1.4.2.

1 1
In particular, since |f | = exp ν1 log(|f |), we have that |f | is sub-
ν ν

harmonic for any ν. This will be the main tool in the problem of sepa-
rate analyticity that we will solve in Section 1.5. However, the subhar-
1
monicity of all the |f | ν ’s, that is, the strong subharmonicity of f in any
degree, does not suffice in proving such a delicate property as propaga-
tion of CR extendibility. We will see that logarithmic subharmonicity
of |f | is needed.
We want to give a differential description of subharmonicity.
Though the characterization applies to a more general situation,
we want to confine ourselves to the case of ϕ ∈ C 2 .
Theorem 1.4.9. Let ϕ ∈ C 2 (Ω); then ϕ is subharmonic if and only
if
(1.4.11) ∂z ∂z̄ ϕ ≥ 0.
Note that ∂z ∂z̄ coincides, up to a constant factor, with the cele-
brated Laplace operator
∂x2 + ∂y2 = 4∂z ∂z̄ .
Proof. We first prove that the submean property implies (1.4.11).
We rewrite the submean contained in (iii) of Theorem 1.4.2 as
Z 2π
ϕ(zo + reiθ ) − ϕ(zo ) dθ ≥ 0.
 
(1.4.12)
0
20 1. SEVERAL COMPLEX VARIABLES

If we express the difference between brackets in (1.4.12) by Taylor ex-

pansion up to order 2, (1.4.12) becomes
Z 2π 
 1
∂z ϕ(zo )reiθ + ∂z̄ ϕ(zo )re−iθ + ∂z2 ϕ(zo )r2 e2iθ

(1.4.13)
0 2

2 2 −2iθ 2
dθ + o(r2 ) ≥ 0.

+∂z̄ ϕ(zo )r e + 2∂z ∂z̄ ϕ(zo )r
Now, the only term which survives after integration is the non-harmonic
one; thus (1.4.13) implies for r sufficiently small, ∂z ∂z̄ ϕ(zo ) ≥ 0.
≥ 0. Write x = r cos θ, y = r sin θ, with
Conversely, assume ∂z ∂z̄ ϕ p
inverse correspondence r = x2 + y 2 , θ = arctg xy . The chain rule
yields
∂r ∂θ sin θ
∂x = ∂r + ∂θ = cos θ∂r − ∂θ ,
∂x ∂x r
∂r ∂θ cos θ
∂y = ∂r + ∂θ = sin θ∂r + ∂θ .
∂y ∂y r
Substitution into ∂z ∂z̄ = 41 (∂x2 + ∂y2 ) yields the expression of ∂z ∂z̄ in
(r, θ) coordinates
1 1 1
∂z ∂z̄ = (∂r2 + ∂r + 2 ∂θ2 ).
4 r r
Our assumption implies that
Z 2π
∂r2 + r−1 ∂r + r−2 ∂θ2 ϕ(zo + reiθ )dθ ≥ 0.


(1.4.14)
0
Note that under integration the term which contains the derivation in
θ vanishes. We denote by M (r) the mean value of ϕ over the r-circle
with center zo :
Z 2π
−1
M (r) := (2π) ϕ(zo + reiθ )dθ.
0

Thus, (1.4.14) reads

··
rM (r) + Ṁ (r) ≥ 0.
But the left hand side term above is the derivative of rṀ (r). Thus,
rṀ (r) is increasing and since it is 0 at r = 0, it follows that Ṁ (r) ≥ 0
and hence M (r) itself is increasing. But M (0) = ϕ(zo ) and therefore
the inequality M (0) ≤ M (r) is precisely the submean property.

Note that through its differential characterization provided by the
theorem above, subharmonicity reveals in full its local character.
 1.4. SUBHARMONIC FUNCTIONS 21

Remark 1.4.10. The statement of Theorem 1.4.7 is immediate if

··
ϕ is C 2 in addition to subharmonic. In fact, for χ̇ ≥ 0 and χ ≥ 0, we
have
··
∂z ∂z̄ χ(ϕ) = χ∂z ϕ∂z̄ ϕ + χ̇∂z ∂z̄ ϕ
≥ 0.
However, in Section 1.7, we will have a crucial application of composi-
tion in which ϕ is not C 2 .
We end this section with an application of Fatou’s lemma to the
integrals which enter into the submean property. It reflects a general
principle. For sequences of functions which admit integral represen-
tation, or estimates by integrals like submeans, in the presence of a
uniform bound, the pointwise “lim sup” enters into the integral and
becomes “uniform”.
Theorem 1.4.11. Let {ϕν }ν be a sequence of subharmonic func-
tions which are uniformly bounded on any compact subset of Ω. Let
lim sup ϕν (z) ≤ l for any z ∈ Ω; then for any  and for any K ⊂⊂
ν
Ω there is ν K such that
(1.4.15) sup ϕν (z) ≤ l +  for any ν ≥ ν K .
z∈K

Proof. Fix zo ∈ K; for |z − zo | < δ and r < d(zo , ∂Ω) − δ with

δ << r, we have by the submean property
Z 2π Z r
2 −1
ϕν (z) ≤ (πr ) tϕν (z + teiθ )dt dθ
0 0
(1.4.16) Z 2π Z r+δ
≤ (π(r + δ)2 )−1 tϕν (zo + teiθ )dt dθ + O(δ),
0 0
where the control O(δ) for the error in passing from integration over
B(z, r) to B(zo , r + δ) follows from the uniform boundedness of the
ϕν ’s. By Fatou’s lemma we can see that (1.4.16) implies
lim sup sup ϕν (z) ≤ (π(r + δ)2 )−1
ν z∈B(zo ,δ)
Z 2πZ r+δ
× lim sup tϕν (zo + teiθ )dt dθ + O(δ)
0 0 ν
≤ l + O(δ).
By a finite covering argument of K and by the arbitrariness of δ, we
conclude that
lim sup sup ϕν (z) ≤ l.
ν z∈K
22 1. SEVERAL COMPLEX VARIABLES


Theorem 1.4.11 is a variant of the famous Hartogs lemma.
Remark 1.4.12. The use of Fatou’s lemma inside the spherical
submean estimate would not suffice for the conclusion. In that case we
should apply Fatou’s lemma to the function θ 7→ sup ϕν (z + reiθ ) but
z
we do not know if its lim sup is still ≤ l. However, a proof which uses
1-dimensional integrals is also available: notice that it uses the Poisson
kernel and not just the submean (cf. [71]) as is hereby done.

1.5. Separate analyticity

If f is continuous on Ω ⊂ Cn and separately holomorphic in each
variable when the others are kept fixed, then we have already seen in
Section 1.2, as a consequence of Cauchy’s integral formula, that it is
holomorphic. We wish now to get rid of the hypothesis of continuity. We
start by proving that in this context continuity is in fact a consequence
of boundedness:
Lemma 1.5.1. Let f be separately holomorphic and bounded on com-
pact subsets of Ω. Then f is continuous and hence holomorphic on Ω.
Proof. By using iteration, we may assume that Ω is a subset of
C2 . Let z o ∈ Ω, let z1ν → z1o , let z2 move near z2o so that (z1ν , z2 ) stays
at distance > r from ∂Ω, and let c be a uniform bound for |f | in the
r-neighborhood of these points. Define Fν (z2 ) := f (z1ν , z2 ); Cauchy’s
inequalities yield
c
|∂z2 Fν (z2 )| ≤ .
r
Hence {Fν }ν is equicontinuous and in particular, if we also take z2ν →
z o , we conclude that
f (z1ν , z2ν ) → f (z1o , z2o ).

We are ready for the main result of the section.
Theorem 1.5.2 (Hartogs theorem). Let f be a function on a do-
main Ω ⊂ Cn which is separately holomorphic; then it is holomorphic.
Proof. The statement is local and can be obtained by recurrence:
so we assume from the beginning that Ω ⊃⊃ ∆ ¯ ×∆ ¯ in C2 and prove
that f is holomorphic on ∆ × ∆. For large positive l, let us define
El := {z2 ∈ ∆ : sup |f (z1 , z2 )| ≤ l}. The continuity of f (z1 , ·) for fixed
z1 ∈∆
z1 implies that if z2ν → z2o with z2ν ∈ El , then z2o ∈ El . Thus El is closed.
 1.5. SEPARATE ANALYTICITY 23

¯ of f (·, z2 ) for fixed z2 implies that SEl = ∆. Thus,

The continuity on ∆
l
Baire’s theorem applies and implies that El has non-empty interior
for large l. At this point we apply the conclusion of Lemma 1.5.1: f
◦
is holomorphic in ∆ × E l . We can repeat the same construction as
above by defining different sets El on any open subset of ∆; hence f
is holomorphic on ∆ × B for an open dense subset B ⊂ ∆. This set
contains a horizontal strip arbitrarily close to z2 = 0. At this stage
we can even forget that f is separately holomorphic in z1 (outside the
strip) while keeping the separate analyticity with respect to z2 ∈ ∆ and
prove that f is holomorphic in ∆ × ∆. Also, for this statement, it is not
restrictive to assume that the strip is centered at 0, that is, of the form
∆×∆ . (Otherwise, we have just to center expansion (1.5.1) at z2 = z2o ,
for |z2o | arbitrarily small, instead of z2 = 0.) Thus Theorem 1.5.2 is a
consequence of the following.
Theorem 1.5.3. Let f be holomorphic in ∆ × ∆ and separately
holomorphic in z2 ∈ ∆ when z1 is fixed in ∆. Then f ∈ hol(∆ × ∆).
Proof. Since we have to prove that f is holomorphic on every rel-
atively compact polydisc, we can assume that f is in fact holomorphic
over a strip slightly bigger than ∆ × ∆ . We then consider the Taylor
series of f with respect to z2 at z2 = 0:
X ∂zν f (z1 , 0)
(1.5.1) 2
z2ν .
ν
ν!
It converges normally for (z1 , z2 ) ∈ ∆×∆ , that is, absolutely uniformly
on compact subsets of ∆ × ∆ . Write
 ν 1
|∂z2 f (z1 , 0)| ν
ϕν (z1 ) := .
ν!
The family of functions {ϕν }ν∈N are subharmonic and satisfy, by
Cauchy’s inequalities,
lim sup sup ϕν ≤ −1 .
ν z1 ∈∆

By our assumption of separate analyticity in z2 , they also satisfy, for

any fixed z1 ,
lim sup ϕν (z1 ) ≤ 1.
ν
But then, by Theorem 1.4.11 they satisfy
lim sup sup ϕν ≤ 1.
ν ∆
24 1. SEVERAL COMPLEX VARIABLES

Thus by Hadamard’s theorem, the power series (1.5.1) converges nor-

mally for (z1 , z2 ) ∈ ∆ × ∆ and thus its sum produces a holomorphic
function on ∆ × ∆. This completes the proof of Theorem 1.5.3 and
therefore of Theorem 1.5.2 also.

Remark 1.5.4. We have used here the subharmonicity of all
 |∂ ν f (z ,0)|  ν1
z2 1
ν!
. This was obtained as a consequence of the subharmonic-
1 |∂zν2 f (z1 ,0)|
ity of ν
log ν!
to which we have applied the convex increasing
1 |∂zν2 f (z1 ,0)|
function exp. However the subharmonicity of ν
log ν!
itself is
not needed here.

1.6. Levi forms—continuity principle (I)—Hartogs extension

theorem
We want to describe an impressive phenomenon of the forced ex-
tension of holomorphic functions of which we have taken a small taste
in Section 1.3. In contrast to the case of real functions, some domains
of Cn are not “domains” of holomorphic functions. We start from the
local case and develop some preliminary discussion.
Definition 1.6.1. For a real C 2 function r in a domain of Cn , we
define the Levi form of r at zo as the hermitian form
X
(1.6.1) Lr (zo ) = ∂zi ∂z̄j r(zo )dzi ⊗ dz̄j .
ij

Theorem 1.6.2. The Levi form (1.6.1) is invariant under holo-

morphic changes of coordinates: if z 7→ z̃ = F (z) is a holomorphic
diffeomorphism and if r̃ is r in the new system (that is, F ∗ r̃ = r), then
(1.6.2) F ∗ Lr̃ = Lr .
(Here F ∗ denotes the pull-back by F .)
Proof. We have to prove that
(1.6.3) Lr̃ (F 0 u, F 0 v) = Lr (u, v̄) for any (u, v̄) ∈ Cn × C̄n .
Now, the coefficients of Lr = Lr̃◦F are calculated by the chain rule.
First, writing z̃(z) for F (z), we have
" #
X X
∂zh ∂z̄k (r̃ ◦ F ) = ∂zh ∂z̃i r̃∂z̄k z̃i + ∂z̄i r̃∂zk z̃i .
i i
 1.6. LEVI FORMS—HARTOGS EXTENSION 25

But the first sum vanishes because ∂z̄k z̃i = 0. We have likewise
X X
∂ zh (∂z̃i r̃)∂zk z̃i = ∂z̃j ∂z̃i r̃∂zh z̃j ∂zk z̃i
i ji
X X
+ ∂z̃j ∂z̃i r̃∂zh z̃¯j ∂zk z̃i + ∂z̄i r̃∂z̄h ∂zk z̃i .
ji ji

Again, the sums in the second line cancel. Writing F (z) for z̃(z), we
get the identity for matrices
Lr = t J(F ) · Lr̃ · J(F ),
which is another way of stating (1.6.3).

The Levi form is called the “complex hessian” of r. The full hessian,
the quadratic term of the expansion of r, written in coordinates (z, z̄) ∈
n
Cn × C is
1X 1X X
hess r = ∂zi ∂zj rdzi ⊗dzj + ∂z̄i ∂z̄j rdz̄i ⊗dz̄j + ∂zi ∂z̄j rdzi ⊗dz̄j .
2 ij 2 ij ij

The sum of the first two terms on the right side is a real harmonic
polynomial
P called the “Levi polynomial” that we denote by LPr :=
Re ij ∂zi ∂zj rdzi ⊗dzj ; thus hess r = LPr +Lr . By a coordinate change
we can see that the Levi polynomial can be canceled from a graphing
function: the Levi form is therefore the holomorphic invariant of the
hessian. A first rough way of getting rid of the Levi polynomial consists
in taking the mean of the value of the hessian over a vector u and iu.
Proposition 1.6.3. We have
1
Lr (u, ū) = [hess r(u, u) + hess r(iu, iu)] .
2
Proof. We have
(1.6.4) LPr (iu, iu) = −LPr (u, u),
(1.6.5) Lr (iu, iu) = Lr (u, ū).
It follows that
hess r(u, u) + hess r(iu, iu)
= (LPr (iu, iu) + LPr (u, u)) + (Lr (iu, iu) + Lr (u, ū))
= 2Lr (u, ū).

26 1. SEVERAL COMPLEX VARIABLES

We go on now to a geometric notion, that of the Levi form of a real

hypersurface M ⊂ Cn and refer to Section 3.4 for the case of higher
codimension. We assume that M is locally defined by an equation r = 0
with ∂r 6= 0. We denote by T C M the complex tangent bundle to M
defined by T C M = T M ∩ iT M . For any z ∈ M its fiber TzC M is
the space of vectors orthogonal to ∂r(z) under the hermitian product.
Sometimes, we also denote it by ∂r(z)⊥ .
Definition 1.6.4. The Levi form of the hypersurface M at a point
z ∈ M is defined by
LM (z) := Lr (z)|TzC M .
We have to note at the very beginning that the conformal class of
LM does not depend of the choice of the equation r = 0 of M ; thus the
rank of LM and, under the choice of the orientation, the signature are
well-defined. In fact, let r · g = 0 be another equation with g 6= 0, and
consider
Lr·g = gLr + rLg + ∂r ⊗ ∂g ¯ + ∂r
¯ ⊗ ∂g.
First, we have rLg = 0 on r = 0, and second, both ∂r ⊗ ∂g ¯ and ∂r¯ ⊗ ∂g
⊥
vanish when restricted to ∂r . Thus
Lr·g |T C M = g · Lr |T C M .
We denote by s+
r , s−
s0r the numbers of the eigenvalues of Lr which
r ,
−
are > 0, < 0, = 0, respectively. We denote by s+ 0
M , sM , sM the corre-
sponding numbers for LM . Here there is an ambiguity on the sign since
M is defined both by ±r = 0. We always agree to think of M as the
boundary of Ω : {z : r(z) < 0} with the orientation induced by the
outward conormal ∂r.
Theorem 1.6.5 (Normal coordinates for a hypersurface). Let M be
a hypersurface of Cn and zo a point of M . Then, by a holomorphic, in-
deed quadratic, change of coordinates and with the notation z = (z1 , z 0 ),
we can put locally at zo = 0 the equation of M in the form
 
+ +1
sX s+ +s − +1
X
(1.6.6) y1 = − |zj |2 + |zj |2  + o2 (x1 , z 0 ).
j=2 j=s+ +2

Proof. By a linear change of coordinates we can assume that zo =

0 and that the equation r = 0 of M satisfies 2Re ∂r(0) = dy1 . Thus, M
is locally a graph
y1 = f (z 0 , z̄ 0 , x1 ) with ∂f (0) = 0.
2
Hence f = O and we can therefore decompose
f = Q(z 0 , z̄ 0 ) + L(z 0 , z̄ 0 )x1 + Cx21 + o2 (x1 , z 0 ),


 1.6. LEVI FORMS—HARTOGS EXTENSION 27

where Q is quadratic, L linear, C constant with all of them real. We

write explicitly
X X X
Q= aij zi0 zj0 + āij z̄i0 z̄j0 + 2bij zi0 z̄j0
ij ij ij

(with bji = b̄ij ) and X

L= ci zi0 + c̄i z̄i0 .
i
We make now a quadratic coordinate change
(
z1 7→ z1 − i ( i 2ci zi z1 + Cz12 ) ,
P

z 0 7→ z 0 .
By the implicit function theorem the equation of M becomes
y1 = Q(z 0 , z̄ 0 ) + o2 (x1 , z 0 ).
We now want to eliminate the Levi polynomial inside Q. We perform
the quadratic transform
n
z1 7→ z1 − i ij 2aij zi0 zj0 ,
P

and our equation becomes

X
y1 = 2 bij zi0 z̄j0 + o2 (x1 , z 0 ).
ij

We are ready to conclude. For this we perform a last, orthonormal,

change of coordinates inside the plane Cn−1 ' T0C M of the z 0 variables
which diagonalizes the Levi matrix (bij )ij = LM and next a dilation
which makes the eigenvalues unitary. This provides the normal form
(1.6.6) in the statement of the theorem.

Example 1.6.6. Let Ω be defined by y1 < −2x22 + y22 + ... where the
dots denote terms of degree ≥ 3 in z2 , z̄2 ; we also write the inequality
which defines Ω as r < 0. Hence, for the real structure of the x2 , y2
plane, the set M = ∂Ω is neither convex nor concave. But it is “more”
convex than concave and thus it is convex in the complex sense. In fact,
if e2 = (0, 1, 0...), then
1
LM (e2 , ē2 ) = [hess r(e2 , e2 ) + hess r(ie2 , ie2 )]
2
1
= > 0.
2
Hence e2 is a positive eigenvector for the Levi form. According to
the normal form proved in Theorem 1.6.5 there must be a coordinate
28 1. SEVERAL COMPLEX VARIABLES

change z̃ = z̃(z) under which the inequality which defines the domain
changes into
(1.6.7) ỹ1 < −|z̃2 |2 + ....
For this purpose, it is enough to consider the transformation
(
z̃1 = z1 + i 23 z22 ,
z̃2 = √z22 ,

and the initial inequality changes into (1.6.7).

Thus, the convexity of the Levi form represents the “complex” con-
vexity. It is invariant under complex holomorphic change of coordinates
since the Levi form is, whereas the real convexity can be affected by
these transformations as we have just seen.
Remark 1.6.7. The C 2 regularity of M = ∂Ω at some point zo
implies the property of Ω to contain a ball which is tangent to M at zo
from the interior of Ω. Its radius is the minimum curvature radius in
different directions. On the other hand, strict positivity of the Levi form
LM (zo ) > 0 is equivalent, under a holomorphic change of coordinates,
to the fact that Ω is contained in a neighborhood of zo = 0 in the set
defined by
y1 < −|z 0 |2 − x21 + o2

< − (|z 0 |2 + x21 ).
2
Thus, instead, Ω is now contained in a ball (of center −i 2 and radius
2

).
We describe now the (deformed) Hartogs figures. This technique is
far weaker than the “continuity principle” presented in the next section,
but we think it is worth talking about it because of its simplicity and
its historical relevance. A holomorphic disc is the holomorphic bounded
image into Cn of the standard disc ∆ of C. We consider a holomorphic
foliation of a domain Ω ⊂ Cn by means of holomorphic discs. This
can be described by taking a domain U ⊂ Cn−1 and a holomorphic
one-to-one mapping Φ:
Φ
∆ × U → Ω, w 7→ z = Φ(w),
with each disc being defined by Aw0 := Φw0 (∆) for Φw0 (·) := Φ(·, w0 ).
We can also assume that Φ extends as a holomorphic diffeomorphism
 1.6. LEVI FORMS—HARTOGS EXTENSION 29

to a neighborhood of ∂∆ and set ∂Aw0 := Φw0 (∂∆). For an open subset

Ωo ⊂ Ω we assume
(
∂Aw0 ⊂⊂ Ω̄o for any w0 ∈ U,
(1.6.8)
Āw0 ⊂⊂ Ω̄o for w0 in an open subset U 0 ⊂ U .

Under these circumstances, we encounter one of the most elementary

phenomenon of holomorphic extension which is essentially a conse-
quence of Cauchy’s formula:
ext
hol(Ωo ) 99K hol(Ω).

In fact, we begin by remarking that we have the freedom of shrinking

our discs a little. This is the same as replacing Ω̄o by Ωo in (1.6.8).
R Φ∗ f (ζ1 ,w0 )
Next, we define Φ g∗ f (w) := 1 dζ1 . The integral makes
2πi ∂∆ ζ1 −w1
sense because Φw0 (∂∆) ⊂⊂ Ωo where f is defined. The function Φ ∗f
g
∗ 0
is holomorphic for w ∈ ∆ × U , coincides with Φ f on ∆ × U , and
hence it coincides on the whole of ∆ × U by analytic continuation.
Thus f˜ := Φg∗ f ◦ Φ−1 is the holomorphic extension of f from Ω to Ω.
o
This elementary principle suffices for the following celebrated theorem.
Theorem 1.6.8 (Hartogs extension theorem). Let Ω be a domain
with C 2 boundary, let zo ∈ M = ∂Ω and assume that LM has at least
one negative eigenvalue. Then for any neighborhood B of zo there is
another neighborhood B 0 ⊂ B such that
ext
hol(Ω ∩ B) 99K hol((Ω ∩ B) ∪ B 0 ).
Proof. By our assumption about the Levi form, we can as-
sume that Ω contains in a neighborhood of zo (that we suppose, by
rescaling, to be the unit ball) the set yn < |z1 |2 − Q(z2 , ..., zn−1 ) −
o2 (xn , z2 , ..., zn−1 ) where Q denotes a quadratic term in z2 , ..., zn−1 . We
consider this quadratic term and o2 as an error term E = E(xn , z2 , ...)
and define U := {z 0 ∈ Cn−1 : yn < 1 − E} and U 0 := {z 0 : yn < −E}.
Now, the family of straight discs in which z1 ranges through ∆ and z 0
is fixed in U satisfies

∂Az0 ⊂⊂ Ω for z 0 ∈ U, A ⊂⊂ Ω for z 0 ∈ U 0 .

On the other hand it is clear that this family {Az0 } covers a full neigh-
borhood B 0 of zo = 0. The conclusion is then a consequence of the
extension criterion herewith proved.

30 1. SEVERAL COMPLEX VARIABLES

1.7. Logarithmic supermean of Taylor radius of holomorphic

functions—continuity principle (II)—propagation of
holomorphic extendibility

Summary. Cauchy’s inequalities and the Cauchy-Hadamard

theorem relate the holomorphic extension of a function with the esti-
mates of the moduli of its derivatives. Since these are logarithmically
subharmonic, their inverses, which represent the radii of convergence
of the Taylor expansion, are logarithmically superharmonic. By this,
we get the proof of the theorem of propagation of holomorphic ex-
tendibility: a complex curve in a hypersurface M is a propagator of
holomorphic extendibility from either side of M .

Logarithmic supermean of Taylor radius of holomorphic

functions.

Definition 1.7.1. We say that a real upper semicontinuous func-

tion ϕ in Ω ⊂ Cn is plurisubharmonic when its restriction to any disc
A = Φ(∆) ⊂ Ω is subharmonic.

We say that ϕ is plurisuperharmonic when −ϕ is plurisub-

harmonic. When ϕ is C 2 , plurisubharmonicity is characterized by
∂zi ∂z̄j ϕ ≥ 0 for any i, j. Thus, subharmonicity along straight discs
τ 7→ wo + τ w, indeed just coordinate rays in which w = (...0, 1, 0...),
suffices. But this is in fact always true, not just for C 2 functions.

Proposition 1.7.2. Let ϕ 6= −∞ be upper semicontinuous and

subharmonic along each cartesian ray. Then, there is a sequence {ϕν }ν
of C ∞ plurisubharmonic functions on Ων := {z ∈ Ω : d(z, ∂Ω) > ν1 }
such that ϕν & ϕ. In particular, ϕ is plurisubharmonic.

Proof. We take a function χ ∈ Cc∞ of |z1 |, ..., |zn | with support

R dζ∧dζ̄
contained in |z| < 1, satisfying χ ≥ 0 and χ (−2i)n = 1, and set

dζ ∧ dζ̄
Z
ζ
ϕν (z) = ϕ(z − )χ(ζ) .
∆×...×∆ ν (−2i)n

It is obvious that every ϕν is C ∞ in Ων . We prove now that it is plurisub-

harmonic. Since it is smooth, it suffices to check the property along each
cartesian complex plane. By iterated integration, it suffices to prove it
in dimension 1 for the integration in a single variable ζ ∈ ∆. But in
 1.7. SUPERMEAN—CONTINUITY PRINCIPLE—PROPAGATION 31

fact,
Z 2π
1
ϕν (wo + reiθ )dθ
2π 0
Z Z  Z 2π 
1 iθ ζ dζ ∧ dζ̄
= ϕ(wo + re − )dθ χ(ζ)
∆ 2π 0 ν −2i
dζ ∧ dζ̄
ZZ
ζ
≥ ϕ(wo − )χ(ζ) = ϕν (wo ).
∆ ν −2i
We come back to general dimension n. Since the ϕν ’s are plurisub-
harmonic, then the (ϕν ) 1 ’s are decreasing with  & 0 (by iterated 2-

dimensional submean over -discs). By letting ν → ∞, the ϕ 1 ’s them-

selves are decreasing: writing ν in place of 1 , we conclude ϕν & ϕ.
Last, we show that ϕ is plurisubharmonic. It is upper semicontinuous
for the same reason as in one variable (cf. the beginning of the proof of
Lemma 1.4.6). Moreover, by taking composition with the parametriza-
tion of a disc A = Φ(∆), we have
ϕν ◦ Φ & ϕ ◦ Φ.
Since each ϕν ◦ Φ is subharmonic, then also the monotonic limit ϕ ◦ Φ
is subharmonic on account of Lemma 1.4.6.

Let Ω be an open domain, v a unit vector in Cn , f a holomorphic
function on Ω. We consider the function
 ν  1 !−1
∂ f (z) ν
rfv (z) = lim sup  v

(1.7.1)  , z ∈ Ω.
ν ν! 

The function rfv (z) represents the radius of the maximal disc in the v
direction in which the Taylor series of f at z is separately convergent
and hence f is separately holomorphic. When we uniformize over z 0
ranging in a neighborhood B of z in the (n − 1)-plane orthogonal to v
passing through z, we get
Proposition 1.7.3. We have
(1.7.2) inf
0
rfv (z 0 ) = sup{r : f is holomorphic in v∆r × B}.
z ∈B

Proof. The inequality “≥” follows from Cauchy’s inequalities. As

for “≤”, we suppose that the direction of v is that of the z1 axis and
denote by z 0 the complementary variables. In any polydisc centered at
ξo and contained in Ω, f is a sum of a double power series in z1 − ξ1
32 1. SEVERAL COMPLEX VARIABLES

and z 0 − ξ 0 that we may rearrange as

X ∂ ν f (z 0 )
v
(1.7.3) (z1 − ξ1 )ν .
ν
ν!
  ν 0  1 −1
ν
Let r < inf v 0
rf (z ), that is, r < lim sup  ∂v fν!(z )  for any

0
z ∈B ν
z 0 ∈ B. It follows that
1
 ∂v f (z 0 )  ν
 ν
(1.7.4) sup lim sup   ≤ r−1 .
z 0 ∈B ν ν! 
Hence, if we apply Theorem 1.4.11 we may interchange “sup” and “lim
sup” in (1.7.4); in particular the power series (1.7.3) converges normally
for z1 − ξ1 ∈ ∆r , uniformly for z 0 ∈ B and thus defines a holomorphic
function on v∆r × B which matches f ; hence r stays in the set on the
right side of (1.7.2) which completes the proof of Proposition 1.7.3.

The property expressed by (1.7.2) is essential in the present section.
As for the purpose of the subsequent Section 1.8, we need not to uni-
formize on z 0 ∈ B but, instead, on v ∈ S 2n−1 , the unit (2n − 1)-sphere
in Cn .
Proposition 1.7.4. Let f be holomorphic in a neighborhood of z.
Then
(1.7.5) inf rv (z) = sup{r : f ∈ hol(Bn (z, r))},
2n−1 f
v∈S
n
where B (z, r) is the n-ball with center z and radius r.
Proof. “≥” is clear by Cauchy’s inequalities. For the opposite, let
r < rfv (z) for any v. Then f is separately holomorphic along each ray
z + v∆r and, in addition, holomorphic in a neighborhood of z. By an
obvious variant of Theorem 1.5.3 it is holomorphic in Bn (z, r). In fact,
let us prove that f is holomorphic in a conical neighborhood in B(z, r)
of z + v∆r for each v. We can suppose, without loss of generality,
0
that v = (1, ..., 0). Under the transformation, (z1 , z 0 ) 7→ (z1 , zz1 ), the
-conical neighborhood of z + v∆r is transformed into the polydisc
∆r × (∆ × ∆ × ...). By applying Theorem 1.5.3 we get the proof of
our claim. By a finite covering we conclude that f is holomorphic in
B(z, r).

We will make essential use of the following elementary consequence
of Proposition 1.7.3. Let z ∈ Ω and zo ∈ ∂Ω be a pair of points such that
 1.7. SUPERMEAN—CONTINUITY PRINCIPLE—PROPAGATION 33

−z
zo − z is normal to ∂Ω at zo . We write v := |zzoo −z| and d(z) := |zo − z|;
thus, d(z) is the distance of z to ∂Ω. In this situation, for a holomorphic
function on Ω we have
if inf
0
rfv (z 0 ) > d(z) for some B, then f extends holomorphically across ∂Ω at zo .
z ∈B

In fact, take r with inf

0
rfv (z 0 ) > r > d(z); then, the Taylor series of
z ∈B
f at z converges in v∆r × B and provides the extension of f across
the boundary ∂Ω at zo . This will be especially useful in combination
with the following statement which is also a consquence of the results
of Section 1.4
Theorem 1.7.5. The function log rfv enjoys the supermean property
along any disc.
 ν 
Proof. Set ϕν := ν log  ∂ν!
1  vf 
. Since
− log( lim sup ·) = − lim sup (log ·),
ν ν
then
 ν  1 !−1
∂ f ν
log rfv = log lim sup  v 
ν ν!
= − lim sup ϕν .
ν
Here
(
ϕν are plurisubharmonic (by Lemma 1.4.8),
ϕν are uniformly bounded (by Cauchy’s inequalities).
Restrict the ϕν ’s to a disc A, that is, take the composition ϕν ◦ Φ with
its parametrization Φ. By Remark 1.4.5, log rfv ◦ Φ = − lim sup ϕν ◦ Φ
ν
enjoys the supermean property.

Note that in the above proof we cannot apply Proposition 1.4.4 as
it stands because we do not know if log rfv ◦ Φ is lower semicontinuous.
On the other hand, we can force it to be semicontinuous by replacing
rfv by
∨v
rf := sup inf
0
rfv (z 0 ).
B z ∈B
This is the maximal lower semicontinuous minorant of rfv . In this way,
we gain semicontinuity but, as a price, we may loose supermean prop-
∨
erty (cf. Example 1.4.3). We call rvf (z) the Taylor radius of f in the v
direction in a neighborhood of z.
34 1. SEVERAL COMPLEX VARIABLES

Continuity principle. We now wish to take advantage of the powerful

tool that we have just developed and generalize the procedure of exten-
sion by means of families of analytic discs. This has already been ex-
ploited, in a particular case, in the course of the proof of the Hartogs ex-
tension theorem. We consider a family of discs {At }t parametrized over
a connected space of parameters T . We suppose that the parametriza-
tion is a foliation of a domain Ω ⊂ Cn that is a C 1 -diffeomorphism
∼
Φ : T × ∆ → Ω which is holomorphic with respect to τ ∈ ∆ and which
extends as a diffeomorphism to a neighborhood of ∂∆. Differently from
the case of the Hartogs figure, Φ is not necessarily a holomorphic dif-
feomorphism. We set Φt = Φ(t, ·), At = Φt (∆) and ∂At = Φt (∂∆).
Theorem 1.7.6. In the above situation, let Ωo be a subdomain of
Ω which satisfies
(i) ∂At ⊂⊂ Ω̄o ,
(ii) Āt ⊂⊂ Ω̄o for t ranging in an open set of parameters To ⊂ T .
Then, there is an extension mapping
ext
(1.7.6) hol(Ωo ) 99K hol(Ω).
Proof. It is not restrictive to reduce Ωo to
¯ 1− )) ∪ (To × ∆)


Φ (T × (∆ \ ∆
for a suitable annulus ∆ \ ∆ ¯ 1− = {τ : 1 −  < |τ | < 1}. Also, by
exhausting T × ∆ by relatively compact subsets T 0 × ∆0 , it suffices
to prove extension to the set Φ(T 0 × ∆0 ). In this way we may assume
that our discs are shrunk so that ∂At ⊂⊂ Ωo . Finally, though it is not
essential, we suppose T connected by arcs.
Fix t1 and let γ be a connected curve connecting t1 to a point
to ∈ To . Fix a holomorphic function f ∈ hol(Ωo ) and let γ 0 = {t ∈ γ :
f is holomorphic in a neighborhood of At }. We have
(a) γ 0 6= ∅ (by (ii)),
(b) γ 0 is open.
But γ 0 is also closed. In fact, let tν → ξ for tν ∈ γ 0 and let zo be a
generic point in Aξ . By reparametrizing ∆, we may assume that zo is
the “center”, that is, corresponds to the value τ = 0. We also denote
by zν the centers of the discs Atν ’s. Set v := Tzo (Φ(γ 0 × {0})); by (i) we
have for a neighborhood B of 0 and a constant c > 0 independent of ν
rfv |∂Atν +B ≥ c.
Since (rfv )−1 satisfies the maximum principle, this implies
rfv |zν +B ≥ c.
 1.7. SUPERMEAN—CONTINUITY PRINCIPLE—PROPAGATION 35

It follows that the Taylor expansion of f at zν converges in polydiscs of

a fixed radius in the v direction; in particular these polydiscs contain
S
zo for zν → zo . Since the intersection of these polydiscs with Ωo ∪ At
t<ξ
is connected, the sum of this series coincides with the extension of f
for t < ξ. It follows that f extends holomorphically to a neighborhood
of zo . In conclusion, γ 0 is closed and hence it must coincide with the
whole of γ by connectedness.

It is quite obvious that the assumption that Φ is a C 1 -diffeomor-
phism is not essential. We can give a similar proof for any family of
discs which is topologically equivalent to T × ∆.
Remark 1.7.7. The statement can be adapted in many ways. The
key point in the proof is that the Taylor series of f in the polydiscs
which contain zo coincides with f previously extended. For this reason
there do not arise “monodromy” problems in the above construction.
This shows that what is crucial are the geometric properties of the re-
gion Ω swept by the discs rather than those of their parametrization.
Thus, for instance, if Ω is pseudoconvex in the sense of Section 1.8, a
very general extension principle should hold. In fact what is true for a
general family of discs, regardless of their parametrization, is an exten-
sion to a domain sheeted over Cn , the envelop of holomorphy, whose
projection contains Ω. Moreover, the assumption of pseudoconvexity is
exactly that the envelop is single-sheeted. The book [83] should be of
help in understanding this matter.
As an application, we give a “global” version of Theorem 1.6.8.
Corollary 1.7.8. Let Ω be a convex bounded domain of Cn . Then
holomorphic functions extend from a neighborhood of ∂Ω to the whole
of Ω.
Proof. We write the variable as z = (z1 , z 0 ) and denote by Ω0 the
projection of Ω in the z 0 -plane and by Ωo a neighborhood of ∂Ω inside
Ω. We “slice” Ω by the sets
{Az0 }z0 ∈Ω0 = {(C × {z 0 }) ∩ Ω}z0 ∈Ω0 .
These are convex bounded subsets of C × {z 0 }, hence straight discs
by the Riemann mapping theorem. They satisfy ∂Az0 ⊂⊂ Ω̄o and also
Āz0 ⊂⊂ Ω̄o when (z1 , z 0 ) approaches a point of strong convexity. Fi-
nally, they are parametrized by a C 1 -diffeomorphism over Ω0 × ∆. The
conclusion follows from Theorem 1.7.6.

36 1. SEVERAL COMPLEX VARIABLES

Remark 1.7.9. Corollary 1.7.8 holds in wider generality and ap-

plies to any bounded domain Ω ⊂⊂ Cn with connected boundary ∂Ω
(Bochner-Hartogs theorem). There are several proofs. One uses the
Bochner-Martinelli kernel, another the solution of ∂¯ with compact sup-
port. Our proof has an immediate generalization to domains which are
pseudoconvex in the sense of Section 1.8, but, according to [38], it does
not cover all cases.
Remark 1.7.10. If we are solely interested to the case of a convex
set as in Corollary 1.7.8, we can follow an alternative proof. It is not
restrictive to assume ∂Ω ∈ C ω ; in fact, we can approximate ∂Ω by a C ω -
hypersurface inside Rthe “shell” where f is holomorphic. We then define
(ζ,z 0 )
f˜(z1 , z 0 ) := (2πi)−1 ∂A 0 fζ−z 1
dζ; this is a C ω -function, holomorphic in
z
z1 for any z 0 and coinciding with f when z belongs to a neighborhood
of a point of strong convexity. In particular, ∂z̄j f˜ = 0 for these values
of z. But then, the principle of analytic continuation applied to the
∂zj f˜’s for j = 1, ..., n implies that f˜ is holomorphic in Ω and coincides
with f wherever defined.
Propagation of holomorphic extendibility. We start from a gen-
eral principle of propagation of analyticity. We consider in the begin-
ning a domain Ω = ∆ × Ω0 and a function f of z = (z1 , z 0 ) ∈ Cn defined
over ∆ × Ω0 and such that
• f is holomorphic with respect to z1 ,
• f is holomorphic in a neighborhood of {0} × Ω0 .
We want to show that f is in fact holomorphic on ∆ × Ω0 . If f ∈ C 1 ,
then the conclusion is elementary. In fact,
(
df ∧ dz1 ∧ ... ∧ dzn has holomorphic coefficients,
df ∧ dz1 ∧ ... ∧ dzn ≡ 0 for |z1 | < .
By analytic continuation the above form is ≡ 0 for any z1 ; thus, f
is holomorphic in ∆ × Ω0 . This argument can be extended in two di-
rections. First, we can give the discussion in the “weak sense”. If f
is a distribution endowed with distributional restriction to each plane
z1 ≡ const, which is certainly the case if f is e.g. continuous, then, in
the above situation, f is still holomorphic. (Note here that when f is
bounded, then it is easily verified to be in fact continuous because of a
“normal family” argument.) In fact, we have to define in this case
g(z1 ) := hf (z1 , ·), dψ(·)i,
where ψ is any (n − 1, n − 2)-form with Cc∞ -coefficients in z 0 and h·, ·i
is the pairing between (0, 0)- and (n − 1, n − 1)-forms. Again, since g is
 1.7. SUPERMEAN—CONTINUITY PRINCIPLE—PROPAGATION 37

holomorphic and is 0 for |z1 | < , then it is ≡ 0; thus, f is holomorphic.

Another generalization consists in deforming the complex lines z 0 ≡
const into a general foliation by holomorphic curves, the graphs z1 7→
ϕ(z1 , z 0 ) of a function ϕ which is holomorphic in z1 , smooth in (z1 , z 0 )
and normalized, say, by ϕ(0, z 0 ) = z 0 . Assume f is holomorphic along
each curve {(z1 , ϕ(z1 , z 0 )) : z1 ∈ ∆} and holomorphic in (z1 , z 0 ) for
|z1 | < ; it then follows that f is holomorphic. In this case the proof
F
consists in using the C ∞ change of coordinates (z1 , z 0 ) → (z1 , z̃ 0 ) =
(z1 , ϕ(z1 , z 0 )) whose inverse F −1 is the straightening of the curves. We
then define
g(z1 ) = hf ◦ F, F ∗ dψi{z1 }×Ω0 ,
for any (n − 1, n − 2)-form ψ with Cc∞ -coefficients in z 0 . As before,
g(z1 ) ≡ 0. Thus f annihilates ∂z̃0j for j ≥ 2 (and F∗ ∂z1 from the begin-
ning) and is C 0 and hence it is holomorphic. In what follows, we see a
much more refined principle of propagation in which in particular the
presence of a single complex curve, not of a full foliation, suffices.
What we aim to describe is a phenomenon of propagation of holo-
morphic extendibility across a hypersurface M for functions which are
holomorphic on one side, say Ω, of M . Complex curves γ ⊂ M are
“propagators” of extendibility: if f ∈ hol(Ω) extends across M at a
point zo ∈ γ, then it extends across M at any other point z ∈ γ. We
could rephrase the propagation of extendibility in terms of propagation
of “microlocal” regularity in the inward conormal for the generalized
boundary value of f on M . It then looks like a particular case of a gen-
eral phenomenon of propagation of regularity for the induced system
of ∂¯ on M . But it is not. First, according for instance to [78] or [53],
the propagators of the regularity of the solutions of the P.D.E.’s should
∗
be the complex curves of the characteristic set, in this case TM Cn , the
conormal bundle to M in Cn (cf. for instance Definition 3.2.5 and the
remark which follows). But the correspondence induced by the pro-
jection π : T ∗ Cn → Cn between complex curves of TM ∗
Cn and M is
one-to-one only if M is pseudoconvex (cf. Sections 1.8 and 3.5). For
instance, the hypersurface defined by y1 = z2 z̄3 + z2 z̄3 contains the
complex curve z2 = 0 but its conormal bundle is “totally real” lacking
therefore any complex tangency; this follows from the fact that the Levi
form is non-degenerate (cf. Section 3.5). Also, the general theory of the
propagation, such as in [78] and [53], concerns the case of foliations
by complex leaves, bicharacteristic leaves of involutive submanifolds of
the characteristic set, whereas, as we have already said, we are here in
the presence of a single leaf.
38 1. SEVERAL COMPLEX VARIABLES

Theorem 1.7.11. Let M be a hypersurface of Cn , Ω one side of

M , γ a complex curve of M , zo a point of γ, f a holomorphic function
on Ω. Then, if f extends across M at zo , it also extends at any other
point z 1 ∈ γ.
(Cf. Hanges and Treves [47] for the case in which f has slow growth
and hence a distributional boundary value at M .)
Proof. We first assume that γ is a disc A and suppose that the
point z 1 to which we wish to propagate is its “center”. By “reducing”
A, that is, taking the restriction of Φ to a piece of the parametric disc
∆, we can suppose that the point of initial extension zo is a boundary
point. We denote by v the unit normal to M at z 1 pointing outside Ω
and set for small η > 0
Aη = −ηv + A, zo η = −ηv + zo , zη1 = −ηv + z 1 .
We assume that the diameter of A is small, say , and that the extension
at zo is to the δ-ball with η << δ << . Let η 0 be the maximal radius in
the v-plane for a disc centered at points of ∂Aη and contained in Ω; this
is a small perturbation of η. In fact, if ϕ ∼  is the angle between v and
0
the conormal to M along ∂A, we have ηη = cos ϕ for cos ϕ = 1 − O(2 ).
We have therefore for a neighborhood B of 0
rfv ∂Aη +B ≥ η 0

(1.7.7)
= η(1 − O(2 )).
But we also have in an α-neighborhood Γη of zo η in ∂Aη

(1.7.8) rfv Γη +B ≥ δ.
Taking the “log” of (1.7.7), (1.7.8) and using the supermean property
of log rfv , we get
α log δ + (2π − α) log η 0
(1.7.9) log rfv |zη1 +B ≥ .
2π
Applying “exp” back, we get
α
  2π
0 δ
rfv |zη1 +B ≥η
(1.7.10) η0
> η,
for small η and η 0 . This shows that f is holomorphic in a neighborhood
of z 1 . If we pass from a disc A to a general curve γ, we take a real curve
Γ ⊂ γ connecting zo to z 1 and denote by ξ the “extremal” point along
Γ of the holomorphic extension for f . We have that ξ must coincide
 1.8. DOMAINS OF HOLOMORPHY AND PSEUDOCONVEX DOMAINS 39

with z 1 for otherwise by means of discs centered at ξ, the argument of

propagation by discs would violate the “extremality” of ξ.

Remark 1.7.12. We are using here the full power of the submean
property of log((rfv )−1 ). Note that the submean property of (rfv )−1
would not suffice. In fact, this would imply, on account of (1.7.7), (1.7.8)
2π
rfv |zη1 +B ≥ 0 −1
(η ) (2π − α) + δ −1 α
(1.7.11) 1
= η0 α η0 α
= η 0 (1 + O(α)).
1 − 2π + δ2π
Thus, the loss from η to η 0 = η(1 − O(2 )) is not balanced by the gain
from η 0 to η 0 (1 + O(α)) and the last term of (1.7.11) is possibly smaller
than η regardless of how small η is chosen.

Suggested research. Develop the discussion contained in Re-

mark 1.7.7 about families of discs which sweep out pseudoconvex do-
mains (after having a look at Section 1.8).

Notes. The continuity principle belongs to mathematical folk-

lore and can be found in many basic books such as the one of Vladimirov
[96]. It underlies an enormous variety of technical passages in SCV
(several complex variables). However, for a clever example of how “not
everything is doable with it” see the counterexample by Fornaess [38].
Notice also that it becomes dangerous when “monodromy” is missing.
The proof that we propose here is a byproduct of the preparation of
the propagation theorem. The latter is due to Hanges and Treves [47]
where it is obtained, instead, by means of microlocal techniques. Our
proof, fully geometric, is new.

1.8. Domains of holomorphy and pseudoconvex domains

Summary. We have already encountered in Section 1.3 domains

where all holomorphic functions extend to a larger set. We call a do-
main for which this does not happen a domain of holomorphy. It turns
out that this is, in fact, the domain of a single holomorphic function, a
function which does not have analytic continuation anywhere across the
boundary. Next, we define holomorphically convex domains and prove
40 1. SEVERAL COMPLEX VARIABLES

their equivalence with the domains of holomorphy. A more explicit

convexity property, pseudoconvexity, is introduced and its equivalence
with Levi convexity of the boundary is proved. Also, holomorphic con-
vexity implies pseudoconvexity, essentially by the use of the analytic
tools prepared in Section 1.7. But the inverse implication, the solution
of the so-called Levi problem, is only stated here whereas its proof is
left to Section 1.9.

Our main task is to characterize the “domains of holomorphic func-

tions”. To introduce this notion properly, we need new terminology. We
say that a function f ∈ hol(Ω) does not extend beyond ∂Ω at zo ∈ ∂Ω
(or is critical at zo ) when for any neighborhood Bzo of zo there is no
function f˜ ∈ hol(Bzo ) which induces f in some open component of
Bzo ∩ Ω. By abuse of terminology, we say in this situation that f has
no “extension” f˜ at zo .

Definition 1.8.1. A domain Ω ⊂ Cn is said to be a domain of

holomorphy if for any boundary point zo ∈ ∂Ω there exists fzo ∈ hol(Ω)
which does not extend beyond ∂Ω at zo .
1
Any domain Ω of C is a domain of holomorphy (having fzo = z−z o
as critical functions at the various zo ∈ ∂Ω). In Cn , n > 1, according
to the conclusions of Sections 1.3 and 1.6, not all domains enjoy the
property of Definition 1.8.1. A striking phenomenon is that in a domain
of holomorphy there is indeed a single function f ∈ hol(Ω) which is
critical at any boundary point; thus, Ω is “the domain” of f . Along with
the discussion on various types of non-extendibility, there naturally
arises the question if Ω always admits a “holomorphy hull” that is
ext
a domain of holomorphy Ω̃ ⊃ Ω such that hol(Ω) 99K hol(Ω̃). Thus,
holomorphic functions extend from Ω to Ω̃ but some of them not beyond
∂ Ω̃. This is true if we allow Ω̃ to “go out” of Cn on manifolds spread
over Cn ; we refer to [50] and [82] for this point.
The property of being a domain of holomorphy concerns the local
non-extendibility at the boundary points zo ∈ ∂Ω but it involves global
sections fzo ∈ hol(Ω). In particular, it is not clear whether it is local:
it holds equivalently for Ω or for the various Ω ∩ B where {B} is a
covering of ∂Ω. The point is that non-extendibility for sections on Ω∩B
is easier than for Ω. But they eventually turn out to be equivalent (cf.
Proposition 1.8.14 and the comments which follow).
We now want to get better insight into the property expressed by
Definition 1.8.1. This requires some preparation.
 1.8. DOMAINS OF HOLOMORPHY AND PSEUDOCONVEX DOMAINS 41

Definition 1.8.2. If K ⊂⊂ Ω, we define the “hol(Ω)-hull” of K

by
K̂Ω = {z ∈ Ω : |f (z)| ≤ sup |f | for any f ∈ hol(Ω)}.
K

If Ω1 ⊂ Ω2 , then K̂Ω1 ⊂ K̂Ω2 ; also, it is easy to check that regardless

of how Ω is chosen, the set K̂Ω is in any case contained in the convex
hull that we denote K̂. In fact, let K be contained in Hzo ,ζ , the half-
space through zo with conormal ζ:
K ⊂ {z : Re hz − zo , ζi ≤ 0}.
Here h·, ·i is the hermitian product in Cn defined by hz, ζi = j zj ζ̄j .
P
Since we can consider among all possible holomorphic functions f the
entire functions f (z) = ehz−zo ,ζi , we then have just by our definition of
hull
sup |ehz−zo ,ζi | = sup eRe hz−zo ,ζi
K̂Ω K

≤ 1.

Thus also K̂Ω ⊂ Hzo ,ζ . Since K̂ is an intersection of half-spaces Hzo ,ζ ,

the inclusion K̂Ω ⊂ K̂ follows. Fix a function f ∈ hol(Ω) and a direc-
tion v ∈ Cn ; recall the function rfv defined in (1.7.1). Despite the fact
that K̂Ω may be bigger than K, we have the following anyhow.
Lemma 1.8.3.
(1.8.1) inf rfv = inf rfv .
K K̂Ω

Proof. “≥” is obvious. As for “≤”, let r < inf rfv . We then have
K

(1.8.2) |∂vν f (w)| < ν!r−ν for any w ∈ K and ν ≥ νo .

But by definition of K̂Ω and on account of the fact that any ∂vν f ∈
hol(Ω), we have that (1.8.2) holds indeed for any w ∈ K̂Ω . This implies
r < inf rfv .
K̂Ω


Lemma 1.8.4. Let Ω be a domain of holomorphy. Then
d(z, ∂Ω) = inf rfv (z).
v∈S 2n−1 , f ∈hol(Ω)
42 1. SEVERAL COMPLEX VARIABLES

Proof. “≤” follows from Cauchy’s inequalities. Conversely, Ω be-

ing a domain of holomorphy, if r > d(z, ∂Ω), there must exist f ∈
hol(Ω) such that f ∈ / hol(Bn (z, r)). By Proposition 1.7.4 there must
then exist v such that rfv (z) < r.

The set K̂Ω is always closed in Ω but need not be compact in Ω.
However, this is the case when Ω is a domain of holomorphy. More
precisely, if d denotes the euclidean distance, then
Theorem 1.8.5. Let Ω be a domain of holomorphy and K ⊂⊂ Ω
be a compact subset. Then
d(K, ∂Ω) = d(K̂Ω , ∂Ω).
Proof. We apply Lemmas 1.8.3 and 1.8.4 and get
d(K, ∂Ω) = inf d(z, ∂Ω)
z∈K
= inf inf rfv (z)
z∈K v, f

= inf inf rfv (z)

z∈K̂Ω v, f

= d(K̂Ω , ∂Ω).

Definition 1.8.6. A domain Ω is said to be holomorphically convex
when for any K compact in Ω, the set K̂Ω is still compact in Ω.
In particular, Theorem 1.8.5 says that a domain of holomorphy is
holomorphically convex. Also, since K̂Ω is controlled by the convex
hull K̂, it follows that a convex domain is holomorphically convex.
Holomorphic convexity is stable under intersection.
Proposition 1.8.7. T If Ωj is a domain of holomorphy for any j,
then the interior Ω of j Ωj is holomorphically convex.
Proof. We have
d(K̂Ω , ∂Ωj ) ≥ d(K̂Ωj , ∂Ωj )
(1.8.3)
= d(K, ∂Ωj ),
where the first inequality follows from the inclusion K̂Ωj ⊃ K̂Ω and the
second equality is a consequence of Theorem 1.8.5. If we take “inf” in
j
(1.8.3) we conclude
(1.8.4) d(K̂Ω , ∂Ω) = d(K, ∂Ω).

 1.8. DOMAINS OF HOLOMORPHY AND PSEUDOCONVEX DOMAINS 43

Theorem 1.8.8. The following are equivalent:

(i) Ω is holomorphically convex.
(ii) Ω is a domain of holomorphy.
(iii) There is f ∈ hol(Ω) which does not extend holomorphically to
any larger set Ω̃ ⊃ Ω.
Proof. (iii) ⇒ (ii) is obvious and (ii) ⇒ (i) has already been
proved in Theorem 1.8.5. We prove now that (i) ⇒ (iii). In fact, let
P be a polydisc with center at 0 and denote by Pz for z ∈ Ω the
largest “P -shaped” polydisc centered at z and contained in Ω. Thus,
Pz = z + rP for a suitable r > 0 and P̄z hits ∂Ω at some point. Let
T ⊂ Ω be countable dense: we want to construct f ∈ hol(Ω) which
cannot be continued to a neighborhood of P̄z for any z ∈ T . Let z1 , z2 ...
be a sequence in T repeating each term in T infinitely many times and
K 1 ⊂⊂ K 2 ... a sequence of compact sets of Ω such that any compact
subset of Ω belongs to some of the K j ’s. Since K cj Ω ⊂⊂ Ω, then there
is ζj ∈ Pzj such that ζj ∈ /Kcj Ω . In particular, there is fj ∈ hol(Ω) such
that sup |fj | < |fj (ζj )|. By replacing fj by its powers, we may assume
Kj

fj (ζj ) = 1, sup |fj | < 2−j .

Kj
We can also suppose that the fj ’s are not ≡ 1 in any component of
+∞
Ω. Define f := Π (1 − fj )j ; the product converges uniformly on each
l
P −jj=1
K since j2 < +∞ and defines a holomorphic function f on Ω.
This function has a zero of order j at ζj . It cannot therefore have
continuation to a neighborhood of any P̄z ’s for z ∈ T because otherwise
it would have there a zero of infinite order and hence be identically 0
in a component of Ω. The zero of infinite order would be taken at any
subsequential limit of ζjk if zjk is the sequence which repeats z infinitely
many times. These limits are dense in ∂Ω since T is dense.

Remark 1.8.9. Proposition 1.8.7 in combination with Theo-
rem 1.8.8 ensures that the interior of the intersection of domains
of holomorphy is a domain of holomorphy. In particular, if Ω is a do-
main of holomorphy, then for any boundary point zo there is a system
of neighborhoods {B} such that B ∩ Ω is a domain of holomorphy.
(For this we just have to take a system of balls B.) This answers the
easy half of the question about localizability of the property.
Example 1.8.10. Let Ω be a “holomorphic polyhedron” in the
form Ω = {z : |fj | < 1} for a family of holomorphic functions fj ∈
44 1. SEVERAL COMPLEX VARIABLES

hol(Cn ), j = 1, ..., n; then Ω is a domain of holomorphy. In fact, let

K ⊂⊂ Ω; we must have |fj |K ≤ 1 −  for any j and for suitable . This
inequality remains true in K̂Cn which is therefore compact in Ω. Hence
K̂Ω is also compact because K̂Ω ⊂ K̂Cn .
The example has the following generalization.
Proposition 1.8.11. Let Ω and Ω0 be domains of holomorphy and
F : Ω → Cn a holomorphic mapping. Then F −1 (Ω0 ) is a domain of
holomorphy.
Proof. For K ⊂⊂ F −1 (Ω0 ) we have
(1.8.5) K̂F −1 (Ω0 ) ⊂ F −1 (F
d K Ω0 ),
because F ∗ (hol(Ω0 )) ⊂ hol(F −1 (Ω0 )) (where F ∗ denotes the pull-back
by F ). Now, F d K Ω0 is compact in Ω0 , hence closed in Cn and thus
F −1 (F
d K Ω0 ) is closed in Ω. In combination with (1.8.5) this implies
that K̂F −1 (Ω0 ) , which is closed in F −1 (Ω0 ), is in fact closed in Ω. But
then, since
K̂F −1 (Ω0 ) ⊂ K̂Ω ⊂⊂ Ω,
we conclude that K̂F −1 (Ω0 ) ⊂⊂ F −1 (Ω0 ).

Notice that Example 1.8.10 is obtained from Proposition 1.8.11 un-
der the choice Ω0 = ∆ × ... × ∆, F = (f1 , ..., fn ). Notice also that when
F is proper, Ω is automatically a domain of holomorphy.
We remember that an upper semicontinuous function is plurisub-
harmonic when the restriction to any disc is subharmonic; we denote
by d(z) the distance to ∂Ω. We go on to define the most crucial notion
in several complex variables.
Definition 1.8.12. A domain Ω is said to be pseudoconvex when
− log d is plurisubharmonic.
Every domain of C is pseudoconvex. Also, since plurisubharmonic-
ity is stable under “sup”, then pseudoconvexity is stable under inter-
◦
T_
section: for a family {Ωj }j of pseudoconvex domains, we have that Ωj
j
is pseudoconvex. Since balls, and more generally convex sets, are pseu-
doconvex (cf. Theorem 1.8.13), any pseudoconvex domain is “locally
pseudoconvex”. The opposite is also true and will be proved in Propo-
sition 1.8.14. Before the statement we need to define the notion of an
“exhaustion” function, that is, a function ψ : Ω → R such that the
 1.8. DOMAINS OF HOLOMORPHY AND PSEUDOCONVEX DOMAINS 45

sets {z : ψ(z) ≤ c} for c % +∞ are a fundamental system of compact

subsets of Ω. We also need to define the cohomology of the ∂¯ system
on Ω. Let
∂¯k−1 ∂¯k
(1.8.6) ...C ∞ (Ω)k−1 → C ∞ (Ω)k → C ∞ (Ω)k+1 ...
¯ P0
be the ∂-complex acting on antiholomorphic forms f = fJ dz̄J
|J|=k
0
where the suffix denotes ordered multiindices J = j1 < ... < jk and
dz̄J stands for dz̄j1 ∧ ... ∧ dz̄jn . Here the coefficients fJ are in C ∞ (Ω)
0
and the action of ∂¯ is defined via ∂f ¯ = P P
∂z̄j fJ dz̄j ∧ dz̄J ; in
j=1,...,n |J|=k
particular, ∂¯k ◦ ∂¯k−1 = 0. We denote the cohomology of (1.8.6) by
ker ∂¯k
Kk (Ω) := .
range ∂¯k−1
The complete characterization of pseudoconvexity is contained in the
following theorem, which is referred to as the “solution of the Levi
problem”. Note that in this section we prove the full characterization
apart from the fact that pseudoconvexity implies Kk (Ω) = 0, k ≥
1 ((ii) ⇒ (iii) in the next statement) which will be the content of
Section 1.9.
Theorem 1.8.13. The following are equivalent:
(i) Ω is pseudoconvex.
(ii) There is an exhaustion function ψ of Ω which is continuous
and plurisubharmonic.
(iii) Kk (Ω) = 0, k ≥ 1.
(iv) Ω is a domain of holomorphy.
Proof. (i) ⇒ (ii): If Ω is bounded, we just have to set
ψ := − log d and otherwise, ψ := − log d + λ|z|2 , which is contin-
uous plurisubharmonic and exhaustive.
(ii) ⇒ (iii): This is the most difficult part that will be proved in
the next section.
(iii) ⇒ (iv): We proceed by induction over the dimension n and use
the hypothesis in three occurrences. We fix a boundary point zo ∈ ∂Ω,
say zo = 0, and suppose that Ω has non-empty intersection ω := Ω∩{z :
zn = 0}. Let j : ω ,→ Ω and π : Cn  {z : zn = 0} be the natural
injection and projection, respectively. We first prove that any k-form
¯
f in ω which is ∂-closed, ¯ = 0, can be “lifted” to
that is, satisfies ∂f
a ∂-closed form f on Ω: this means ∂ f = 0 and j ∗ f˜ = f . To prove
¯ ˜ ¯ ˜
it, we note that ω and Ω \ π −1 (ω) are relatively closed in Ω and take
χ ∈ C ∞ (Ω) which is 1 at ω and 0 at Ω \ π −1 (ω); thus, χπ ∗ f is well
46 1. SEVERAL COMPLEX VARIABLES

¯
defined on Ω and induces f on ω. It is not necessarily ∂-closed; to
obtain it, we have to use a “correction” term of type zn v for a solution
¯ ∗f
¯ = ∂χ∧π
v of ∂v and set
zn

f˜ := χπ ∗ f − zn v.
The existence of the solution v to the above equation is assured by the
vanishing hypothesis of the cohomology groups ( for degree k + 1 if f
has degree k). Thus, f can be “lifted” to f˜ ∂-closed.
¯ Also, ω inherits
from Ω the vanishing of the cohomology:
(1.8.7) Kk (ω) = 0, k ≥ 1.
¯ = 0 on ω, we first lift it to Ω, then solve ∂¯ũ = f˜
In fact, if f satisfies ∂f
on Ω and finally define u := j ∗ ũ. This is the solution of ∂u¯ = f on ω.
By the inductive assumption, (1.8.7) implies that ω is a domain of
holomorphy (which is trivially true at the first inductive step ω ⊂ C1 )
and therefore there is a function f ∈ hol(ω) which does not extend at
zo . Next, using again that K1 (Ω) = 0, we can find f˜ ∈ hol(Ω) such that
f˜|ω = f ; hence f˜ cannot extend holomorphically and Ω is a domain of
holomorphy.
(iv) ⇒ (i): In a domain of holomorphy Ω we have (Lemma 1.8.4)
d(z) = inf rfv (z), z ∈ Ω.
v, f

Thus − log d = sup(− log rfv ); now, the functions − log rfv ’s enjoy the
v, f
submean property along discs according to Theorem 1.7.5 and the same
is true for their “sup”. Since, in addition, − log d is continuous, then
it is plurisubharmonic. The proof of the theorem is complete.

Proposition 1.8.14. Let Ω be a domain of Cn and suppose that
for every point in ∂Ω there is a basis of neighborhoods {B} such that
B ∩ Ω satisfies (ii) of Theorem 1.8.13; then Ω also satisfies (ii).
Proof. We denote by ψB∩Ω the plurisubharmonic exhaustion func-
tion on each B ∩ Ω. By taking a locally finite covering of ∂Ω by the B’s
and gluing the various ψB∩Ω by taking their sup over overlappings, we
find ψ 1 plurisubharmonic on a neighborhood Ω1 of ∂Ω in Ω, such that
ψ 1 → +∞ as z → ∂Ω, and ψ 1 is locally bounded at ∂(Ω\Ω1 ). Let ψ 2 be
a plurisubharmonic function on Cn with ψ 2 → +∞ if |z| → +∞ such
that ψ 2 > ψ 1 over a neighborhood V of ∂(Ω \ Ω1 ). Define ψ to be ψ 2 on
Ω2 := (Ω \ Ω1 ) ∪ V and sup(ψ 1 , ψ 2 ) on Ω1 ; ψ is an exhaustion function
for Ω and it is plurisubharmonic because this property is stable under
“sup” over finite indices.
 1.8. DOMAINS OF HOLOMORPHY AND PSEUDOCONVEX DOMAINS 47


Proposition 1.8.14 reveals that pseudoconvexity is a local property.
In combination with Theorem 1.8.13, it shows that the property of
being a domain of holomorphy is also local. Moreover, it implies a
version of this property in terms of germs.
Remark 1.8.15. If Ω is a domain of holomorphy, then for any
zo ∈ ∂Ω there does not exist a forced extension map for germs of
holomorphic functions
ext
(1.8.8) hol(Ω)zo 99K hol(Cn )zo .
This means that there is a neighborhood B of zo and a holomorphic
function f ∈ hol(Ω ∩ B) (indeed f ∈ hol(Ω)), such that f ∈ / hol(B 0 ) for
any B 0 ⊂ B. It is an open question whether the opposite implication
holds, that is, whether non-extendibility passes from germs to global
sections. Of course the question makes sense only if it is formulated for
any boundary point zo ∈ ∂Ω. In fact, for a single point zo we can well
have extension for global holomorphic functions but not for germs. For
instance, if Ω is defined in C2 by y2 > ϕ(y1 ) where ϕ ≡ 0 for |y1 | < 1
ext
and ϕ = (|y1 | − 1)2 for |y1 | ≥ 1, then hol(Ω) 99K hol(C2 ) holds but
ext
hol(Ω)0 99K hol(C2 )0 does not.
We come back to Theorem 1.8.13. We have proved, in conjunction
with the conclusions of Theorem 1.8.13, that if there exists an exhaus-
tion plurisubharmonic function ψ for Ω, then − log d is also plurisub-
harmonic. This is a deep point in which we use the full strength of the
characterization. A direct but not simple proof could be given along
the lines of [50, Theorem 2.3.7, p. 46]. Note that in the next geometric
characterization and for ∂Ω of class C 2 we deal in fact with − log(−r)
where r is a defining function for ∂Ω with r < 0 on Ω; but it is clear that
this function could replace − log d in Theorem 1.8.13. We also need an
extra amount of Levi positivity which is provided by an additional term
λ|z|2 for large λ.
Let Ω be a domain with C 2 boundary defined by r < 0 with ∂r 6= 0
at points where r = 0, set M = ∂Ω, let LM = Lr |T C M be the Levi
−
form of M and let s+ 0
M , sM , sM be the number of its eigenvalues which
are > 0, < 0, = 0, respectively. We describe M , in a neighborhood of
a boundary point zo , as a graph y1 = f (x1 , z 0 , z̄ 0 ), set r = −y1 + f ,
and denote by z 7→ z ∗ the projection on M along the y1 -axis. The
following relates, in a neighborhood of zo , the signature of the Levi
form of the exhaustion function ψ = − log(−r) + λ|z|2 for suitable λ
to the signature of the Levi form of the boundary LM .
48 1. SEVERAL COMPLEX VARIABLES

Theorem 1.8.16. We have

(
s− − ∗
ψ (z) = sM (z ),
(1.8.9)
s+ + ∗ 0 ∗
ψ (z) = sM (z ) + sM (z ) + 1.

Hence, in particular, s0ψ ≡ 0.

All degeneracy of LM has turned into positivity of Lψ with an extra
gain of 1 because of the jump of dimension from M to Cn .
Proof. To relate Lψ (z) to LM (z ∗ ), we first note that for r = −y1 +
f we have
(
(∂zi ∂z̄j r(z))ij = (∂zi ∂z̄j r(z ∗ ))ij ,
(1.8.10)
∂r⊥ (z) = ∂r⊥ (z ∗ ).
We then remark that
X X
(1.8.11) ∂zi ∂z̄j ψdzi ⊗ dz̄j = |r|−1 ∂zi ∂z̄j rdzi ⊗ dz̄j
ij ij
X
¯ + λ dzj ⊗ dz̄j .
+ |r|−2 ∂r ⊗ ∂r
j
∗
For a vector w ∈ C and for a point z ∈ M we decompose w = (wτ , wυ )
n

where wτ and wυ are the tangential and normal components to M at

z ∗ , respectively. We have
(1.8.12)
Lψ (w, w̄) ≥ |r|−1 Lr (wτ , w̄τ )
+ (|r|−2 − c|r|−1 )|wυ |2 − c|r|−1 |wτ ||wυ | + λ|w|2 ,
 

where c is a bound for the second derivatives of r. Then for suitable

λ = λc we have
λ
(1.8.13) Lψ (w, w̄) ≥ |r|−1 Lr (wτ , w̄τ ) + |w|2 .
2
− − ∗
This implies sψ (z) ≤ sM (z ) whereas the opposite is obvious: thus,
the first of (1.8.9) is proved. The second also follows immediately from
(1.8.13).

In particular, we can relate the pseudoconvexity of a domain Ω,
that is, the positivity of the Levi form of an exhaustion function of Ω,
to the pseudoconvexity of its boundary, that is, the positivity of LM .
Since the exhaustion function is now meant to be global, we have to
modify a little the argument of Theorem 1.8.16 which has only a local
character: for this, we replace −r by d, the Eucledian distance to ∂Ω.
 1.8. DOMAINS OF HOLOMORPHY AND PSEUDOCONVEX DOMAINS 49

If z 7→ z ∗ , Ω → ∂Ω is now the orthogonal projection (the projection

¯ at z and z ∗ is converted
along the vector field ∂d), the equality of ∂ ∂r
into
(1.8.14) ¯
∂ ∂d(z)| ¯ ∗
∂d⊥ (z) ≤ ∂ ∂d(z )|∂d⊥ (z ∗ ) .

In fact, the inclusion {z ∈ Ω : d(z) > } ⊂ {z ∈ Ω : r < −}, forces

inequality for real hessians in real tangential directions hess d(z) ≤
hess(−r)(z ∗ ). Use of the inequality for u and iu with u ∈ ∂d⊥ and
application of Proposition 1.6.3 yields (1.8.14). Thus, LM ≥ 0 implies
that − log d+λ|z|2 is plurisubharmonic on a neighborhood of ∂Ω inside
Ω; notice that, in contrast to − log(−r) + λ|z|2 , this is defined globally
and not only on a local patch. We can say better
Corollary 1.8.17. We have LM ≥ 0 if and only if L− log d ≥ 0.
In other words, pseudoconvexity of M and Ω are equivalent.
Proof. The argument above shows that − log d + λ|z|2 is plurisubhar-
monic over an annulus in the interior side of the boundary. This can be
continued to a plurisubharmonic exhaustion function to the whole Ω
(cf. Proposition 1.8.14). Thus Ω is pseudoconvex (cf. Theorem 1.8.13).
Conversely, if Ω is pseudoconvex and hence a domain of holomorphy,
then LM cannot take any negative eigenvalue (cf. 1.6.8).

In the terms of the Corollary, pseudoconvexity reveals, at least for
2
C boundaries, its fully local character.
− −
Remark 1.8.18. The relation between s+ 0 + 0
ψ , sψ , sψ and sM , sM , sM
in Theorem 1.8.16 does not depend on the choice of r but could depend
on that of λ (and of the specific structure of the exhaustion function
as ψ = − log(−r) + λ|z|2 ). However, if Lψ ≥ 0 for ψ = − log(d) + λ|z|2 ,
then it is also ≥ 0 for ψ = − log d. Therefore, the term λ|z|2 may
eventually be discarded; but it did a great job in the course of the
proof of Theorem 1.8.16 and Corollary 1.8.17.
Corollary 1.8.19. Let Ω be a domain of Cn with C 2 boundary;
the following two conditions are equivalent:
(i) Ω is pseudoconvex.
(ii) There is no analytic disc A such that
(1.8.15) ∂A ⊂⊂ Ω, A ∩ ∂Ω 6= ∅.
Proof. (ii) ⇒ (i): If Ω has a boundary point, say zo = 0, where
s− > 0, then by putting the equation of ∂Ω in canonical form (1.6.6),
that is, y1 = −|z2 |2 + O2 (z3 , ...) + o2 (x1 , z2 , z3 , ...), we have that the disc
{0} × ∆ × {0}... contradicts (1.8.15).
50 1. SEVERAL COMPLEX VARIABLES

(i) ⇒ (ii): Let A be a disc which satisfies (1.8.15). Assume with-

out loss of generality that the point where it touches ∂Ω is its center
A(0) and that it coincides with the origin 0. Then the family of trans-
lated discs {Az } for Az = z + A, |z| < , satisfies (1.6.8) and provides
an extension of holomorphic functions from Ω to a neighborhood of
0. Thus Ω is not a domain of holomorphy, or equivalently, it is not
pseudoconvex.


In Rn , a domain Ω is convex if and only if there is no line seg-

ment whose boundary is in Ω but not all points in between. Hence, the
characterization of Corollary 1.8.19 is the perfect complex-invariant
analogous to this condition with the line segments replaced by discs.

Exercise 1.8.20. Prove that a domain Ω ⊂ Cn with C 2 bound-

ary is pseudoconvex if and only if the slice with every 2-dimensional
complex plane is pseudoconvex. Hint: use the characterization of Corol-
lary 1.8.17.
Try a different argument which sustains the case of general bound-
ary (not known).

Suggested research. (1) For a pseudoconvex domain Ω ⊂ Cn ,

is the function − log(−r) plurisubharmonic even without the additional
term λ|z|2 ? We know that this is true for − log d.
(2) (Generalization of (1)) For a general domain Ω ⊂ Cn , is the
conclusion of Theorem 1.8.16 true for ψ = − log(−r) + λ|z|2 replaced
by − log(−r) or − log d?
(3) When ∂Ω ∈ C 2 , the absence of an extension map (1.8.8) for
germs of holomorphic functions at any zo ∈ ∂Ω assures that Ω is pseu-
doconvex (cf. the proof of Corollary 1.8.19). Is the same conclusion true
if ∂Ω is not C 2 ?

Notes. The characterization of the domains of holomorphy as

holomorphically convex, stated in Theorem 1.8.8, is due to Cartan and
Thullen [31]. That domains of holomorphy are pseudoconvex was es-
sentially found by Levi and Hartogs. We also refer to the fundamental
work by Oka contained in the papers [73]. The presentation that we
follow here is the one of the book by Hörmander [50] revisited by
the method of Section 1.7. The link between the pseudoconvexity of
a domain and its boundary, Theorem 1.8.16, follows the approach of
 1.9. L2 -ESTIMATES FOR ∂¯ 51

Zampieri [101] (cf. also Hörmander [50, Theorem 2.6.12] for another
proof).

1.9. L2 -estimates for ∂¯ on q-pseudoconvex domains of Cn

Summary. We study the existence on a bounded domain Ω ⊂

n
C of solutions to the Cauchy-Riemann equations

(1.9.1) ¯ =f
∂u ¯ = 0,
for f satisfying ∂f

where the datum f and the solution u are (0, k)- and (0, k − 1)-forms,
respectively. We prove solvability of (1.9.1) in C ∞ (Ω̄) for all k ≥ q + 1
when Ω is bounded, has smooth boundary, and is “q-pseudoconvex”. If
Ω is unbounded or does not have C ∞ boundary, but can be exhausted
by bounded, C ∞ , q-pseudoconvex domains Ων % Ω, then (1.9.1) can
be solved in C ∞ (Ω) in corresponding degrees k. In particular, all pseu-
doconvex domains have this exhaustion and this implies the vanishing
of all cohomology groups of ∂¯ in degree k ≥ 1 (the implication (ii) ⇒
(iii) of Theorem 1.8.13, that is, the solution of the Levi problem).
The key point in solving (1.9.1) is to also call into play the adjoint
operator ∂¯∗ : in particular, we take full advantage of the ellipticity in
the interior of the system (∂,¯ ∂¯∗ ).
To study equation (1.9.1), the most effective tool is the Hilbert space
¯
techniques in the context of the ∂-Neumann problem. To exploit them,
we first establish the a priori estimates of Morrey, Kohn and Hörmander
in the weighted L2 spaces. By a suitable choice of the weight, these
¯
ensure the existence of the ∂-Neumann operator and its continuity in
the Sobolev spaces H . The C ∞ regularity up to the boundary of ∂¯
s

follows by an approximation argument.

Weighted H 0 estimates for the (∂, ¯ ∂¯∗ )-system. Let T ∗ Cn be the

n
cotangent bundle to C ; we decompose forms of the fiberwise com-
plexified bundle C ⊗R T ∗ Cn into (1, 0)- and (0, 1)-forms and denote by
C ⊗R T ∗ Cn = (T 1,0 Cn )∗ ⊕ (T 0,1 Cn )∗ this decomposition. We take an or-
thonormal basis ω1 , ..., ωn of (1, 0)-forms and the dual basis ∂ω1 , ..., ∂ωn
of (1, 0)
P0vector fields. We write (0, k)-forms (or k-forms for short) as
u = uJ ω̄J where ω̄J := ω̄j1 ∧ ... ∧ ω̄jk ; here J = j1 < ... < jk
|J|=k
is a multiindex with ordered entries and 0 denotes summation over
P
ordered indices. When the multiindices are no more ordered, then
they are assumed to be alternant: if J decomposes into jK, then
52 1. SEVERAL COMPLEX VARIABLES

J

P0
ujK := sign jK
uJ . We define |u|2 := |uJ |2 ; this norm is inde-
|J|=k
pendent of the choice of the orthonormal basis. We consider a bounded
domain Ω ⊂⊂ Cn with boundary ∂Ω of class C ∞ defined by r(z, z̄) < 0
with ∂r 6= 0 over the points where r = 0. We choose our orthonormal
frame of (1, 0)-forms such that ∂ωn r = 1 (and for this we need possibly
to renormalize r) and ∂ωj r = 0 for any j ≤ n − 1; hence ∂ωj , j ≤ n − 1,
are a basis for T 1,0 ∂Ω. The coefficients of our forms are taken in vari-
ous spaces Λ such as C ∞ (Ω̄), C ∞ (Ω), Cc∞ (Ω), L2 (Ω), H s (Ω) and the
corresponding spaces of k-forms are denoted by Λk . Though our a pri-
ori estimates are proved over smooth forms, they are stated in Hilbert
norms. Thus, let ||u||H 0 be the H 0 = L2 -norm and, for a real function
ϕ, let the weighted L2 -norm be defined by
X0 Z
(1.9.2) 2
||u||Hϕ0 := e−ϕ |uJ |2 dV,
|J|=k Ω

where dV is the element of volume in Cn . We begin by noticing that ∂¯

is a closed densely defined operator on (Hϕ0 )k . Moreover, smooth forms
certainly belong to the domain D∂¯ of ∂¯ and
X0 X
(1.9.3) ¯ =
∂u (∂ω̄i ujK − ∂ω̄j uiK )ω̄i ∧ ω̄j ∧ ω̄K + ...,
|K|=k−1 ij=1,...,n
i<j

where the dots denote terms in which no differentiation of u occurs.

These error terms just come from differentiation of the coefficients of
the vector fields ∂ωj .
We denote by δωj the formal Hϕ0 -adjoint of −∂ω̄j . (“Formal adjoint”
means adjoint in the Hϕ0 scalar product restricted to compactly sup-
ported test functions.) We have

δωj = ∂ωj − ϕj + ...,

where ϕj stands for ∂ωj ϕ and the dots denote a 0-order operator which
does not involve ϕ. We denote by ∂¯ϕ∗ the Hϕ0 -adjoint of ∂.
¯ The operator
∂¯ϕ is still closed, densely defined but it is no longer true that smooth
∗

forms belong to D∂¯ϕ∗ . For this, they must satisfy certain boundary con-
ditions. Namely, there must exist g ∈ (Hϕ0 )k−1 such that

(1.9.4) ¯ = hg, ψi for any ψ ∈ (H 0 )k−1 .

hu, ∂ψi ϕ
 1.9. L2 -ESTIMATES FOR ∂¯ 53

If this is the case, one sets ∂¯ϕ∗ u := g. But integration by parts shows
that (1.9.4) cannot hold unless
X Z
e−ϕ ∂ωj (r)ujK ψ̄K dS = 0 for any K and ψK ,
j=1,...,n ∂Ω
P
that is, ∂ωj (r)ujK |∂Ω ≡ 0 for any K. (Here dS is the element of
j=1,...,n
hypersurface in ∂Ω.) Since we have chosen our basis with the property
∂ωj r|∂Ω = κjn , we then conclude
(1.9.5) u belongs to D∂¯ϕ∗ iff uJ |∂Ω = 0 whenever n ∈ J.

As a set, Hϕ0 coincides with H 0 ; from (1.9.5) it also follows that D∂¯ϕ∗ =
D∂¯∗ . We call a form which belongs to D∂¯∗ “tangential”. Over such a
form the action of the Hilbert adjoint of ∂¯ coincides with that of its
“formal adjoint” and is therefore expressed by a “divergence operator”:
X0 X
(1.9.6) ∂¯ϕ∗ u = − δωj (ujK )ω̄K + ... for any u ∈ D∂¯∗ ,
|K|=k−1 j

where the dots denote an error term in which u is not differentiated

and ϕ does not occur. When taking the restriction of (1.9.6) to ∂Ω,
the index j does not take value n because, in that case, the coefficient
unK is missing. We leave unspecified in what follows the indices such
as i, j, meaning that they run from 1 to n when we are in the interior
of Ω and from 1 to n − 1 over the boundary ∂Ω. We get from (1.9.6)
and (1.9.3)
 X0 X Z
 ¯∗ 2
||∂ u|| = e−ϕ δωi uiK δωj ujK dV + R,
 ϕ Hϕ0


|K|=k−1 ij ZΩ



 X0 X
¯ 20 =− e−ϕ ∂ω̄j uiK ∂ω̄i ujK dV

(1.9.7) ||∂u||Hϕ
Ω
Z ij
|K|=k−1



 X 0 X
e−ϕ |∂ω̄j uJ |2 dV + R,



 +
Ω

|J|=k j

where we denote by R different error terms which involve integration of

products of type ∂ω̄j (uiK )ū or δωj (ujK )ū with coefficients independent
of ϕ and only depending on the derivatives of the coefficients of the
ωj ’s. (We omit indices in the u’s when unnecessary.) The key technical
result is contained in the following proposition.
Proposition 1.9.1 (Hörmander, Kohn and Morrey). Let u ∈
∞
C (Ω̄)k ∩ D∂¯∗ and fix arbitrarily an index qo with 0 ≤ qo ≤ n − 1; then
54 1. SEVERAL COMPLEX VARIABLES

for a suitable C > 0, independent of u and ϕ,

(1.9.8)
¯ 2 0 + ||∂¯∗ u||2 0 + C||u||2 0
||∂u||Hϕ ϕ Hϕ Hϕ
X0 X Z X0 X Z
−ϕ
≥+ e ϕij uiK ūjK dV − e−ϕ ϕjj |uJ |2 dV
|K|=k−1 ij Ω |J|=k j≤qo Ω

X0 X Z X0 X Z
+ e−ϕ rij uiK ūjK dS − e−ϕ rjj |uJ |2 dS
|K|=k−1 ij ∂Ω |J|=k j≤qo ∂Ω
 
X0 X X0 X
+ (1 − )  ||δωj uJ ||2Hϕ0 + ||∂ω̄j uJ ||2Hϕ0  .
|J|=k j≤qo |J|=k j≥qo +1

(Note that there is no relation between k and qo in this statement.)

Proof. We replace the suffix Hϕ0 by ϕ in all our norms. We recall

(1.9.7); we wish to rewrite the integrals δωi uiK δωj ujK and ∂ω̄j uiK ∂ω̄i ujK
in the right sides of the two lines of (1.9.7). By integration by parts we
get
(1.9.9)
R
−ϕ
e−ϕ r̄j δωi (uiK )ūjK dS
R

 Ω
e δω i
uiK δ ωj
ujK dV = ∂Ω
− Ω e−ϕ ∂ω̄j δωi (uiK )ūjK dV + R,

 R

− Ω e ∂ω̄j uiK ∂ω̄i ujK dV = − ∂Ω e−ϕ ri ∂ω̄j (uiK )ūjK dS

R −ϕ R



+ Ω e−ϕ δωi ∂ω̄j (uiK )ūjK dV + R.
 R

Here rj stands for ∂ωj r. Recall that ∂ωj r = h∂ωj , ωn i and this is 0
or 1 according to j < n or j = n. On the other hand, for j = n,
it is the coefficient unK which vanishes when restricted to ∂Ω. We
thus conclude that the boundary integrals vanish in both equalities
of (1.9.9). Note also that by taking the sum of the two terms in the
right side of (1.9.9), after discarding the boundary integrals, we put
in evidence the commutators [δωi , ∂ω̄j ]. In the last sum of the second
equality of (1.9.7), we now want to interchange ∂ω̄j with δωj for indices
j ≤ qo . A double integration by parts yields
Z
||∂ω̄j uJ ||ϕ = ||δωj uJ ||ϕ − e−ϕ δωj , ∂ω̄j uJ ūJ dV + R,
2 2
 
(1.9.10)
Ω

since, again, the boundary integrals vanish because rj |∂Ω ≡ 0 for

j ≤ qo . Thus, the sum of the two right hand sides of (1.9.9) is
 1.9. L2 -ESTIMATES FOR ∂¯ 55

e−ϕ [δωi , ∂ω̄j ](uiK )ūjK dV + R. What we have obtained so far is

R
Ω
(1.9.11)
X0 X Z
¯ 2
||∂u|| + ||∂¯ϕ∗ u||2ϕ = e−ϕ [δωi , ∂ω̄j ](uiK )ūjK dV
ϕ
|K|=k−1 ij Ω

X0 X Z
− e−ϕ [δωj , ∂ω̄j ](uJ )ūJ dV
|J|=k j≤qo Ω
!
X0 X X
+ ||δωj uJ ||2ϕ + ||∂ω̄j uJ ||2ϕ + R.
|J|=k j≤qo j≥qo +1

We have to describe now the commutators which appear in (1.9.11).

¯ Observe that
Recall that ωn = ∂r and hence ∂ ω̄n = ∂ ∂r.
(1.9.12) ¯ ∼ Lr
∂ ∂r
in the sense that
(1.9.13) ¯ u ∧ v̄i = Lr (u, v̄) for any u and v in T Cn .
h∂ ∂r,
(Incidentally, we remark that the identification contained in (1.9.13)
looks strange because Lr is hermitian whereas ∂ ∂r ¯ is alternant. They
coincide just because their action is restricted to (T 1,0 Cn ⊕ {0}) ×
({0} ⊕ T 0,1 Cn ).) Denote by (rij )ij the matrix of ∂ ∂r ¯ or Lr in the basis
ω1 , ..., ωn ; we wish to relate rij for j ≤ n − 1 to the component of
[∂ωi , ∂ω̄j ] along ∂ωn . For this, we recall Cartan’s formula
¯ n , ∂ω ∧ ∂ω̄ i
hωn , [∂ω , ∂ω̄ ]i = −h∂ω
i j i j

+ ∂ωi (hωn , ∂ω̄j i) − ∂ω̄j (hωn , ∂ωi i).

By the orthogonality hωn , ∂ω̄j i = 0 and hωn , ∂ωi i = 0, and by the iden-
tity ∂ ∂¯ = −∂∂,
¯ we have
¯ n , ∂ω ∧ ∂ω̄ i
hωn , [∂ω , ∂ω̄ ]i = −h∂ω i j i j

= rij .
Similarly
hω̄n , [∂ωi , ∂ω̄j ]i = −h∂ ω̄n , ∂ωi ∧ ∂ω̄j i
= −rij .
It follows that
(1.9.14) [∂ωi , ∂ω̄j ] = rij ∂ωn − rij ∂ω̄n + ...,
where the dots denote combinations of ∂ωh ’s and ∂ω̄h ’s for h ≤ n − 1.
We denote by chij = hωh , [∂ωi , ∂ω̄j ]i(= −h∂ω ¯ h , ∂ω ∧ ∂ω̄ i), resp. dh =
i j ij
hω̄h , [∂ωi , ∂ω̄j ](= −h∂ ω̄h , ∂ωi ∧∂ω̄j i), the coefficients with which the ∂ωh ’s
56 1. SEVERAL COMPLEX VARIABLES

and ∂ω̄h ’s, respectively, occur in the dotted terms of (1.9.14). (Inciden-
tally, they are related by the relation
dhij = hω̄h , [∂ωi , ∂ω̄j ]i
= hωh , [∂ω̄i , ∂ωj ]i
= −hωh , [∂ωj , ∂ω̄i ]i = −chji .)
Let (ϕij )ij be the matrix of Lϕ in the basis ω1 , ..., ωn . We have
n−1
X
ϕij = ∂ωi ∂ω̄j (ϕ) + (∂ ω̄n )ij ϕω̄n + (∂ ω̄h )ij ϕω̄h
h=1
(1.9.15) n−1
X
= ∂ωi ∂ω̄j (ϕ) + rij ϕω̄n − dhij ϕω̄h .
h=1

It follows that
[−ϕi , ∂ω̄j ] = ∂ω̄j ∂ωi (ϕ)
= −[∂ωi , ∂ω̄j ](ϕ) + ∂ωi ∂ω̄j (ϕ)
(1.9.16) n−1
X
= ϕij − rij ϕωn − chij ϕωh .
h=1

Using (1.9.14) together with (1.9.16), we get

[δωi , ∂ω̄j ] = [∂ωi − ϕi , ∂ω̄j ]
n−1 n−1
(1.9.17) X X
= ϕij + rij δωn − rij ∂ω̄n + chij δωh + dhij ∂ω̄h .
h=1 h=1

With (1.9.17) in hand, we wish to rewrite the third line of (1.9.11).

What we want is to interchange δωn with ∂ω̄n in the expressions for
[δωi , ∂ω̄j ] since the other derivatives, of either type δ or ∂, ¯ are all error
terms, as we will see soon. Now,
Z Z
−ϕ
e rij δωn (uiK )ūjK dV = e−ϕ rij rn uiK ūjK dS
Ω
(1.9.18) Z ∂Ω
− e−ϕ rij uiK ∂ω̄n (ujK )dV,
Ω

where rn = h∂ωn , ∂ri = 1 under our normalization. All terms involving

squares of u or products of u by δωj u for j ≤ n − 1 or ∂ω̄j u for j ≤ n
are denoted by R. Also, to keep track of former errors described after
(1.9.7) coming from ||∂¯∗ u||2ϕ , we include into R integrals of products
 1.9. L2 -ESTIMATES FOR ∂¯ 57

δωn (un K )ū. We also denote by S the sum in the last line in (1.9.11).
Thus, we can rewrite the right side of (1.9.11) as
!
X0 XZ XZ
e−ϕ ϕij · − e−ϕ ϕjj ·
|K|=k−1 ij Ω j≤qo Ω
!
X0 XZ XZ
+ e−ϕ rij · − e−ϕ rjj · + S + R,
|J|=k ij ∂Ω j≤qo ∂Ω

where · denotes products of the u’s. To conclude our proof, we only

need to prove that
(1.9.19) R ≤ S + C ||u||2ϕ .
In fact, if we point our attention at terms which involve ∂ω̄j (u)ū for
qo + 1 ≤ j ≤ n or δωj (u)ū for j ≤ qo , then (1.9.19) is clear since S
carries the corresponding square ||∂ω̄j u||2ϕ or ||δωj u||2ϕ . Otherwise, we
note that for j ≤ n − 1 we may interchange ∂ω̄j with δωj by means of
integration by parts: boundary integrals do not occur because rj |∂Ω = 0
for j ≤ n−1. As for δωn , notice that it only hits coefficients whose index
contains n and hence unK |∂Ω = 0. So δωn (unK )ū is also interchangeable
with unK ∂ω̄n (ū) by integration by parts. This concludes the proof of
Proposition 1.9.1. 
Remark 1.9.2. The term (1 − )S in the last line of (1.9.8) is not
relevant in our line of discussion but it is by itself very significant. In
fact, let us point our attention at forms u ∈ (Cc∞ (Ω))k . Then, each
||∂ω̄j u||2ϕ can be interchanged with ||δωj u||2ϕ even for j = n by the
vanishing of the boundary integral which occurs because u has compact
support. Thus,
!
1 X0 X X
S + C||u||2ϕ ≥ ||∂ω̄j uJ ||2ϕ + ||δωj uJ ||2ϕ .
2 j≤n j≤n
|J|=k

For the same reason, in the right side of (1.9.8), the boundary integrals
vanish. If we then apply the variant of (1.9.8) with, say, ϕ = 0, we get
an estimate which fully expresses the elliptic regularity in the interior
¯ ∂¯∗ ):
of the system (∂,
(1.9.20) ||∂¯∗ u||2 + ||∂u||
¯ 2 + ||u||2 > ||u||2 1
H for any u ∈ (Cc∞ (Ω))k ,
∼

where “>” means “≥” up to multiplicative constant and where H 1 is

∼
the Sobolev space of index 1.
58 1. SEVERAL COMPLEX VARIABLES

Remark 1.9.3. In the proof of (1.9.8) we have interchanged all

||∂ωj uJ ||2ϕ with ||δωj uJ ||2ϕ for any j ≤ qo . If we do it only when j belongs
to J, that is, for P J =P jK, we get an obviousP variant
Pof (1.9.8), in which,
0 0
in particular, − · is replaced by − · in the second and
|J|=k j≤qo |K|=k−1 j≤qo
third lines.

As we have already said, (1.9.8) is true for any qo and any k. We

clarify now how they must be specified in order to exploit the full
strength of (1.9.8).

Weighted H s estimates on q-pseudoconvex domains. Let M =

∂Ω and let (rij )ij=1,...,n−1 be the matrix of LM under a choice of a C 2 ,
orthonormal basis of (1, 0)-forms ω1 , ..., ωn with ωn = ∂r. We order the
eigenvalues of LM (z) as λ1 (z) ≤ λ2 (z) ≤ ... ≤ λn−1 (z) and denote by
−
s+ 0
M (z), sM (z), sM (z) the numbers of them which are positive, negative
and null, respectively. We go on to describe our geometric hypotheses
on M . Let q ≤ n − 2; following [101] and [1], we introduce

Definition 1.9.4. We say that M is q-pseudoconvex (for the ori-

entation induced from Ω) if there exists a covering of the boundary
and, on each patch, a C 2 smooth bundle V ⊂ T 1,0 M of rank qo ≤ q,
say V = Span {∂ω1 , ..., ∂ωqo }, such that

q+1 qo
X X
(1.9.21) λj (z) − rjj (z) ≥ 0.
j=1 j=1

The index qo may vary on different patches. It is evident that

(1.9.21) implies λq+1 ≥ 0; hence (1.9.21) is still true if we replace the
first sum q+1
P Pk
j=1 · by j=1 · for any k such that q + 1 ≤ k ≤ n − 1.

Example 1.9.5. A rough estimate of the index q = qo for which

(1.9.21) holds in a neighborhood B of a boundary point zo is

(1.9.22) q = s− (zo ) + s0 (zo ).

In some cases we can do better. For instance, assume that s− (z) is con-
stant; then λs− < 0 ≤ λs− +1 and hence the negative eigenvectors span
− −
a bundle V s that,
P identified withP the span of the first s coordinate
vectors, yields j≤s− +1 λj (z) − j≤s− rjj (z) ≥ 0. Note that a pseudo-
convex domain is characterized by s− (z) ≡ 0; thus, it is 0-pseudoconvex
in our sense.
 1.9. L2 -ESTIMATES FOR ∂¯ 59

Now, (1.9.21) is equivalent to

X0 X X0 X
rij uiK ūjK − rjj |uJ |2 ≥ 0
(1.9.23) |K|=q ij |J|=q+1 j≤qo

for any u tangential of degree q + 1.

By what was remarked after (1.9.21), it follows that (1.9.23) is in fact
true for any u tangential of degree k ≥ q + 1. Sometimes, we also use
a variant of (1.9.23) which is still sufficient for our existence problem:
X0 X X0 X
(1.9.24) rij uiK ūjK − rjj |ujK |2 ≥ 0
|K|=q ij |K|=q j≤qo

for any u tangential of degree q + 1.

Again, (1.9.24) is in fact true for any tangential form of degree k with
q + 1 ≤ k ≤ n − 1.
Theorem 1.9.6. Let Ω be q-pseudoconvex or let it satisfy (1.9.24);
then, for ϕt := (t + C)|z|2 and for any u ∈ C ∞ (Ω̄)k ∩ D∂¯∗ , we have
(1.9.25) ¯ 2 0 + ||∂¯∗ u||2 0 if k ≥ q + 1.
t||u||2 0 ≤ ||∂u||
Hϕ Hϕ ϕ Hϕ

Proof. We choose qo as in (1.9.21); we have

(1.9.26)
 P0 P P0 P

 rij uiK ūjK − rjj |uJ |2 ≥ 0,
|K|=k−1 ij |J|=k j≤qo
P0 P P0 P

 ϕij uiK ūjK − ϕjj |uJ |2 ≥ (k − qo )(t + C)|u|2 .
|K|=k−1 ij |J|=k j≤qo

Using (1.9.8) and noticing that k − qo ≥ 1, we get (1.9.25).

If we are assuming (1.9.24) instead of (1.9.21), then we replace uJ
by ujK in the second sums of the two lines of (1.9.26). Taking into
account Remark 1.9.3, we conclude that (1.9.25) is still true.

For s integer, let Hϕs t be the weighted s-Sobolev space: recall that
u ∈ Hϕs t when Dα u ∈ Hϕ0 t for any |α| ≤ s. As a set, Hϕs t coincides with
H s and the two norms are equivalent. If we wish to bring the estimate
(1.9.25) into H s -norm, we have to control the commutators [∂, ¯ Dα ] and
[∂¯t , D ] where we are writing ∂¯t for ∂¯ϕt . To this end, we use the full
∗ α ∗ ∗

power of the weight ϕt .

Theorem 1.9.7. Let Ω be q-pseudoconvex or let it satisfy (1.9.24).
Then for any s, for suitable t = ts and for any u ∈ C ∞ (Ω̄)k ∩ D∂¯∗ , we
have
(1.9.27) ¯ 2 s + ||∂¯∗ u||2 s if k ≥ q + 1.
||u||2 s < ||∂u||
H H t H
∼
60 1. SEVERAL COMPLEX VARIABLES

Proof. We show how to derive (1.9.27) from (1.9.25). We denote

by T a general tangential vector field, that is, a combination of ∂ωj , ∂ω̄j ,
j ≤ n − 1, and ∂ωn − ∂ω̄n and by T s a tangential operator of order s.
We denote by N the normal vector field ∂ωn − ∂ω̄n . We first remark
that for any u with H s cefficients and for any i ≥ 1, we have for any
choice of t
(1.9.28)
¯ 2 s−1 + ||∂¯∗ u||2 s−1 + ||T s u||2 0 ) + cst ||u||2 s−1 .
||N i T s−i u||2Hϕ0 ≤ cs (||∂u||
t Hϕ t Hϕ Hϕ t Hϕ
t t t

(It is meant here that the inequality is true for any T s−i and for suitable
T s .) Classically, this formula is referred to as a consequence of the
“noncharacteristicity” of (∂,¯ ∂¯∗ ). If i = 1 and s = 1 it immediately
follows from the splitting

(1.9.29) N = ∂ω̄n + T ,
¯ 2 +||∂¯∗ u||2 +||u||2 contains ||∂ω̄n u||2 .
and the fact that the “energy” ||∂u||
For general i and s, one replaces ∂ω̄n inside (1.9.29) (or ∂ωn inside the
twin decomposition N = ∂ωn + T for a new T ) according to the (non-
characteristicity) relations


 N uj K = ∂ω̄n uj K + T uj K



 ¯ j n K + T uj K + P 0 u,
= ∂ω̄j un K − (∂u)
N un K = ∂ωn un K + T un K

(1.9.30)

 n−1
X

 = (∂¯∗
u)K − ∂ωj uj K + T un K + Pt0 u,
t



j=1

where P 0 and Pt0 are 0-order operators. If we then apply N i−1 T s−i
to the two terms of (1.9.30), remark that ∂ω̄j and ∂ωj j ≤ n − 1 are
derivatives of type T , and put into an s − 1 order term the error from
commutators for bringing N in head position, we get (1.9.28). Now,
(1.9.28), together with induction, shows that in order to estimate the
various ||N i T s−i u||’s, that is, the full norm||u||s , it is sufficient to es-
timate the ||T s u||’s. This is the central point of the proof. To see it,
we begin by remarking that T s u still belongs to D∂¯∗ . Next, we observe
that we can describe the commutator [∂¯t∗ , T s ] as As + Ats−1 where As
and Ats−1 are operators of order s and s − 1, respectively, the first being
independent of t and involving at most one non-tangential derivative.
¯ T s ] has order ≤ s and is independent of t; hence it can be
Also, [∂,
 1.9. L2 -ESTIMATES FOR ∂¯ 61

“absorbed” by As . It follows that

(1.9.31)
¯ s u||2 0 + ||∂¯∗ T s u||2 0
t||T s u||2Hϕ0 ≤ ||∂T
t
Hϕ t
t Hϕ t
s¯ s ¯∗
≤ ||T ∂u||2Hϕ0 + ||T ∂t u||2Hϕ0 + ||As u||2Hϕ0 + ||Ats−1 u||2Hϕ0 .
t t t t

Now, if As is fully tangential, it can be absorbed in the left side of

(1.9.31) by a suitable choice of t = ts . Otherwise, if As carries some
normal derivative N i , we use (1.9.28) and get the estimate
(1.9.32)
¯ 2 s−1 + ||∂¯∗ u||2 s−1 + ||T s u||2 0 ) + cst ||u||2 s−1 .
||As u||2Hϕ0 ≤ cs (||∂u||
t
H t H H H

Again, the term involving T s u can be absorbed in the left side of

(1.9.31). This proves
(1.9.33) ¯ 2 s + ||∂¯∗ u||2 s + cs t ||u||2 s−1 .
t||T s u||2 0 < ||∂u||
H H t H H
∼

As it has already been said, (1.9.33), inserted inside (1.9.28) implies

(1.9.27) apart from an error of order s − 1 which can be removed by
induction.
This completes the proof of (1.9.27).

In contrast to the statement of Theorem 1.9.6, when passing from
H - to H s -norm, we need a multiplicative constant in estimate (1.9.27).
0

Remark 1.9.8. The constant t has disappeared from (1.9.27) ex-

cept from the adjoint operator ∂¯t∗ . However, the weight ϕt did a great
job: by means of its part C|z|2 , it served in Proposition 1.9.1, in com-
bination with the sum of derivatives that we denoted by S, to control
the error terms R. Also, through its part t|z|2 , it took care of the errors
given by the commutators [∂,¯ T s ] and [∂¯∗ , T s ].
t

¯
The ∂-Neumann operator. We give the outline of the construction
by Kohn. We begin by recalling from [50, Lemma 4.1.3] and, more
closely, [36, Lemma 4.3.2] that in a domain Ω with C ∞ boundary,
C ∞ (Ω̄)k ∩ D∂¯∗ is dense in H 0 (Ω)k ∩ D∂¯∗ ∩ D∂¯ for the triplet norm
¯ 2 0 + ||∂¯∗ u||2 0 .
||u||2 0 + ||∂u||
H H t H

For this reason, (1.9.25) is in fact true for any u ∈ H 0 (Ω)k ∩ D∂¯t∗ ∩ D∂¯
instead of C ∞ (Ω̄)k ∩ D∂¯t∗ . We define now the “complex Laplacian” by
t := ∂¯∂¯t∗ + ∂¯t∗ ∂.
¯
This is a linear closed densely defined operator on forms with coeffi-
cients in H 0 which leaves unchanged their degree. A form u belongs
62 1. SEVERAL COMPLEX VARIABLES

to Dt iff u ∈ D∂¯t∗ and ∂u ¯ ∈ D∂¯∗ , that is, u satisfies the celebrated
t
¯
“∂-Neumann conditions”. In our normalization ωn = ∂r, they are de-
scribed over a smooth form u by

un K |∂Ω ≡ 0,

(1.9.34) ∂ω̄n uj K |∂Ω = (∂ω̄n uj K − ∂ω̄j un K )|∂Ω

 ¯ n K |∂Ω ≡ 0.
= (∂u)
It is readily checked that
¯ 2 0 + ||∂¯∗ u||2 0 .
ht u, uiH 0 = ||∂u||H t H

Thus, (1.9.27) for the choice s = 0 (extended for all u ∈ H 0 ∩ Dt ) can
be rewritten as
||u||2H 0 < ht u, uiH 0
∼
≤ ||t u||H 0 ||u||H 0 ,
which yields
||u||H 0 < ||t u||H 0 for any u ∈ H 0 (Ω)k ∩ Dt .
∼

The range of t is therefore closed and t is injective. There is then a

¯
well-defined “∂-Neumann” inverse operator Nt := −1 t which acts as
Nt : H 0 (Ω)k → Dt
and which satisfies
t||Nt u||H 0 ≤ ||u||H 0 for any u ∈ H 0 (Ω)k .
It is easy to check that it satisfies, among other properties
¯ t = Nt ∂¯
(i) ∂N
and
(ii) ∂¯t∗ Nt = Nt ∂¯t∗ .
We have used (1.9.27) for s = 0 (or rather (1.9.25)) so far. Coming to
a general s, we have a statement similar as to Theorem 1.9.7 with the
system (∂, ¯ ∂¯∗ ) replaced by the operator t , that is,
t

(1.9.35) ||u||2H s < ||t u||2H s when u ∈ D ∩ C ∞ (Ω̄)k for k ≥ q + 1.

∼

The proof is the same as above and consists in getting control of the
commutators [Λs , t ] by a suitable choice of t = ts . By the density of
C ∞ (Ω̄) into H s , (1.9.35) can be rewritten as
(1.9.36)
||Nt u||H s < ||u||H s for any u ∈ H s (Ω) such that Nt u ∈ H s (Ω)k .
∼
 1.9. L2 -ESTIMATES FOR ∂¯ 63

Note that we only know that Nt u ∈ H 0 ; so, the restraint Nt u ∈ H s is

substantial. To eliminate it from (1.9.36), we have to use the method
of the “elliptic regularization” (cf. [57]) and define a perturbed elliptic
operator Nt : H s → Dt ∩ H s+2 which satisfies
(1.9.37) ||Nt u||2H s ≤ cs ||u||2H s for any u ∈ H s (Ω)k ,
where cs is independent of . Let u ∈ H s (Ω)k ; because of (1.9.37) we
have a subsequential H s -weak limit Nt u → v as  → 0. It is clear that
Nt u → Nt u in H 0 . In particular Nt u = v, which proves that
(1.9.38) Nt u ∈ H s (Ω)k if u ∈ H s (Ω)k .
We are ready to prove
Theorem 1.9.9. Let Ω be q-pseudoconvex or let it satisfy (1.9.24).
Then, if k ≥ q +1, for any f ∈ H s (Ω)k with ∂f ¯ = 0, we can find u = us
in H s (Ω)k−1 such that
¯ = f,
(
∂u
(1.9.39)
||u||H s < ||f ||H s .
∼

Proof. We set u := ∂¯t∗ Nt f for t = ts ; we have

¯ ∂¯∗ Nt f ) = t Nt f − ∂¯∗ ∂N
∂( ¯ tf
t t

= f − ∂¯t∗ Nt ∂f
¯
= f.

Remark 1.9.10. If under the choice t = ts , we have (1.9.27) for
H , we also have it for any H j , j ≤ s. In particular, the solution u
s

of Theorem 1.9.9 satisfies not only the estimate in the second line of
(1.9.39) but also ||u||2H j < ||f ||2H j for any j ≤ s.
∼

We did not just find a solution in Hϕs t , but the one which is orthog-
¯ in fact ∂¯∗ u = 0 because ∂¯∗ ◦ ∂¯∗ = 0. This serves
onal to the kernel of ∂; t t t
to define the “Bergman projection”
Bs : H s → ¯
ker ∂,
f 7→ f − ∂¯t∗ Nt ∂f. ¯

In particular, the estimate ||∂¯∗ Nt ∂f

¯ ||H s < ||f ||H s shows that Bs
t ϕt ϕt
∼
s
is continuous in H -norm. However, we do not know if the canonical
solution in H 0 (unweighted), u = ∂ ∗ N0 f , belongs to H s nor if the
projection B0 accordingly defined, is H s -continuous. Let us denote by
Kk (Ω̄) (resp. Kk (Ω)) the cohomology in degree k of the complex ∂¯
64 1. SEVERAL COMPLEX VARIABLES

over forms with coefficients in C ∞ (Ω̄) (resp. C ∞ (Ω)) (cf. the lines after
(1.8.6)).
Theorem 1.9.11. Let Ω be bounded, smooth and q-pseudoconvex or
let it satisfy (1.9.24). Then
Kk (Ω̄) = 0 if k ≥ q + 1.
Proof. We follow [58]. We have to solve the equation ∂u ¯ = f for
∞ k ¯
f ∈ C (Ω̄) such that ∂f = 0. We claim that we can find a sequence
{uν }ν of H ν solutions such that ∂¯t∗ u = 0 for t = tν and
||uν+1 − uν ||H ν ≤ 2−ν .
Suppose we have carried out our construction up to uν and now look
for uν+1 . Take ũν+1 ∈ H ν+1 with (∂¯ũν+1 = f, ∂¯t∗ ũν+1 = 0) according
to Theorem 1.9.9. Put h := −ũν+1 + uν ; the problem that we have to
overcome is that h is only H ν and not yet H ν+1 . Thus, in each patch
of a finite covering {Bj }j≥1 of ∂Ω we “push points of Ω̄ inward to Ω”
by means of the mapping
Ω̄ ∩ Bj → Ω, z 7→ z + vj ,
where vS
j is an “inner normal” to Ω in Bj . Let B0 be a neighborhood
of Ω \ Bj , take a partition of the unity {ψj }j≥0 subordinate to the
j≥1
covering {Bj }j≥0 and put
X
h (z) = ψ0 (z)h(z) + ψj (z)h(z + vj ).
j≥1
¯ ∂¯∗ ) in the interior of Ω, the function
Note that by the ellipticity of (∂, tν
∞
ψj hj . We have
P
h is C (Ω̄). We also rewrite the above sum as
j≥0
X
¯ =
∂h ¯ j ∧ hj
∂ψ 
j≥0
X
= ¯ j ∧ (hj − h),
∂ψ 
j≥0
P¯ ¯ P ψj h) = 0. Thus,
where the last equality follows from ∂(ψj )h = ∂(
j≥0 j≥0
¯  ||H ν → 0 as  → 0. At this point we recall Theorem 1.9.7 and find
||∂h
a solution v of
¯  = ∂h ¯ ,
(
∂v
¯  ||H ν .
||v ||H ν < ||∂h
∼
The solution v can be found in any Sobolev space, in particular in
¯  ∈ C ∞ (Ω̄); its norm is controlled by the norm of ∂h
H ν+1 , because ∂h ¯ 
 1.9. L2 -ESTIMATES FOR ∂¯ 65

in any H j for j ≤ ν +1, in particular in H ν . (Cf. Theorem 1.9.9 and the

subsequent remark.) Define uν+1 := ũν+1 + h − v ; by choosing  small,
−ν
P
we can achieve ||uν+1 − uν ||H ν ≤ 2 . Define u = uνo + (uν+1 − uν );
ν≥νo
then the above power series converges and defines u ∈ H νo for any νo ;
thus, u is in C ∞ (Ω̄).

¯
We have thus achieved our goal of solving the ∂-problem with
smooth solutions up to the boundary over a q-pseudoconvex domain
provided that it is bounded and has smooth boundary. Note that this
does not mean that the canonocal solution in H 0 is smooth; there are
counterexamples in case of the “worm” domains (cf. e.g. [36]). We go
on to consider the situation in which Ω no longer has a C ∞ bound-
ary but can be exhausted by a sequence Ων % Ω of C ∞ , bounded,
q-pseudoconvex domains. In this context we are not able to prove the
regularity of ∂¯ at the boundary, that is, the existence of C ∞ (Ω̄) solu-
tions. However, our techniques suffice to produce solutions in C ∞ (Ω).
We start with a definition. For a domain Ω, possibly with non-smooth
boundary and unbounded, and for a smooth exhaustion function ψ of
Ω, we denote by λψ1 ≤ λψ2 ≤ ... ≤ λψn the ordered eigenvalues of the
Levi form Lψ |∂ψ⊥ , where ∂ψ ⊥ is the complex (n − 1)-bundle orthogo-
nal to ∂ψ. We introduce a notion related to q-pseudoconvexity which
applies to a boundary which is not necessarily smooth. Let ψ be a C ∞
exhaustion function in B ∩ Ω for a neighborhood B of ∂Ω.
Definition 1.9.12. The function ψ is said to be q-plurisubharmonic
if there is a smooth bundle V qo on Ω̄ ∩ B orthogonal to ∂ψ, say
V qo = Span {∂ω1 , ..., ∂ωqo }, such that
q+1 qo
X X
(1.9.40) λψj (z) − ψjj (z) ≥ 0 for any z ∈ B ∩ Ω.
j=1 j=1

We have also a variant of this definition on the line of (1.9.24). In

contrast to Definition 1.9.4, the latter Definition 1.9.12 is a condition
on Ω instead of ∂Ω; but it is not substantially different because it is
uniform up to ∂Ω.
The case q = 0 deserves particular attention. First, if we assume in
addition that ∂Ω is C ∞ , Definition 1.9.4 mathches Definition 1.9.12 for
the choice ψ = − log d (Corollary 1.8.17). In general, we have already
introduced in Section 1.8 another notion of pseudoconvexity which also
applies to non-smooth domains. We wish to show that it is consistent
with the new one.
66 1. SEVERAL COMPLEX VARIABLES

Proposition 1.9.13. Let Ω have a continuous plurisubharmonic

exhaustion function ψ. Then we can regularize it to a C ∞ , strictly
plurisubharmonic exhaustion function ψ̃.
In particular, pseudoconvexity is equivalent in the sense of Defini-
tion 1.8.12 and Definition 1.9.12.
Proof. (Cf. [50, Theorem 2.6.11].) Set Ωj := {z ∈ Ω : ψ(z) < j};
let χ = −2n χ( z ) be a set of smooth functions which approximate the
Dirac measure δ as defined in Section 1.7. We wish to regularize ψ by
convolution with χ ; however, some care is needed when approaching
the boundary ∂Ω. Let
Z
ψj (z) = ψ(ζ)χ (z − ζ)dV (ζ) + |z|2 .
Ωj+1

For sufficiently small  = j , we have that ψj ∈ C ∞ (Cn ) and ψ < ψj <

ψ +1 on a neighborhood of Ω̄j ; also, from ∆ψj = ∆ψ ∗χ +, we deduce
that ∆ψj |Ω̄j > 0. Next, we take the composition χ(ψj + 1 − j) where
χ is a C ∞ function such that χ(t) ≡ 0 for t ≤ 0 and χ̇ > 0, χ̈ > 0 for
t > 0; this is a strictly plurisubharmonic function on Ω̄j \ Ωj−1 . Hence
we can find a sequence of positive numbers {aj } such that
µ
X
ψµ := ψjo + aj χ(ψj + 1 − j),
j=jo +1

is > ψ, and is strictly plurisubharmonic on Ω̄µ . Since different ψµi and

ψµk agree over Ω̄j when µi , µk > j, we have that ψ̃ := limψµ is well
µ
defined, > ψ, and strictly plurisubharmonic C ∞ on Ω.

We come back to general q and are ready for the conclusive state-
ment.
Theorem 1.9.14. Let Ω have a smooth q-plurisubharmonic exhaus-
tion function; then
Kk (Ω) = 0, k ≥ q + 1.
Notice that Ω is not necessarily smooth and bounded.
Proof. Let ψ be the C ∞ exhaustion function as in Defini-
tion 1.9.12, and define Ων := {z ∈ Ω : ψ(z) < ν}; the Ων ’s are
bounded, C ∞ , q-pseudoconvex. We have to solve the equation ∂u ¯ =f
∞
in the spaces of forms with C (Ω)-coefficients, when deg(f ) = k ≥ q+1
¯ = 0. We first assume that f ∈ H 0 (Ω)k where H 0 = H 0 for
and ∂f ϕt
2
ϕt = e−(C+t)|z| and show how to find a solution u ∈ H 0 (Ω)k−1 . In
 1.9. L2 -ESTIMATES FOR ∂¯ 67

the present discussion, t can be chosen as t = 0. According to Theo-

rem 1.9.6, for any ν, there is a solution uν := ∂¯t∗ Nt f ∈ H 0 (Ων )k−1 ; in
particular, uν satisfies ∂¯t∗ uν = 0. Also, the uν ’s have the estimate
t||uν ||H 0 (Ων ) ≤ ||f ||H 0 (Ων )
≤ ||f ||H 0 (Ω) .

Hence, there is a subsequence uνj which converges in H 0 (Ω) to a weak

limit u which is a solution of (∂u ¯ = f, ∂¯∗ u = 0). If we assume that
t
the coefficients of f are C ∞ (Ω) (in addition to Hϕ0 t (Ω)), then (1.9.20)
(elliptic regularity in the interior) implies u ∈ C ∞ (Ω)k−1 .
We pass to the general case in which f is not necessarily in Hϕ0 t . We
then change the weight into ϕt + η where η is plurisubharmonic and so
rapidly increasing at ∂Ω and at ∞ that f ∈ Hϕ0 t +η . As before, we first
solve in Hϕ0 t +η and next regularize in C ∞ .


For the case q = 0, when Ω has a plurisubharmonic exhaustion

funtion which is C 0 , then it has also another which is C ∞ (Proposi-
tion 1.9.13) and hence Kk (Ω) = 0 for any k ≥ 1, which is the impli-
cation (ii) ⇒ (iii) of Theorem 1.8.13. This completes the last detail in
the solution of the “Levi problem”.

¯
Notes. The ∂-Neumann problem was suggested by Garabedian
and Spencer [44]. The basic a priori estimates were first proved for
1-forms by Morrey [69]. Next, Kohn [55] derived the general estimates
and proved the regularity up to the boundary on strongly pseudocon-
vex domains. The L2 estimates with weights which are 0 at the bound-
ary were used by Hörmander [49] to bypass the boundary regularity
problem. The regularity at the boundary by use of the weighted ∂- ¯
Neumann problem, that is, the conclusion of Theorem 1.9.11 for q = 0,
was proved by Kohn in [57]. A very clear and exhaustive explanation
¯
of the ∂-Neumann method by Kohn can be found in the book by Chen
and Shaw [36]. As for the condition of q-pseudoconvexity and its use
¯
for the ∂-Neumann weighted estimates, we refer to Zampieri [101] and
Ahn [1]. Another approach, through the integral method, is mainly due
to Henkin (cf. [4]) and is very well explained in the last six chapters of
the book by Boggess [27].
We now want to make some complementary comments. The solution
¯ = f that we have constructed in Theorem 1.9.11 is obtained by
u of ∂u
approximation and is not necessarily the “canonical” solution. This is
the one in the form ∂¯∗ N f and hence, it satisfies ∂¯∗ u = 0; in particular,
68 1. SEVERAL COMPLEX VARIABLES

it is orthogonal to the kernel of ∂. ¯ As pointed out by Kohn, the regu-

larity of the canonical solution would have significant consequences for
mapping problems. In fact, for f ∈ C ∞ (Ω̄), it would imply that the
Bergman projection B(f ) := f − u where u is the solution of ∂u ¯ = ∂f
¯
¯ ∞
which satisfies u ⊥ Ker ∂ would continue to belong to C (Ω̄). But
then, according to a result by Bell and Ligocka [23], any holomorphic
one-to-one self-mapping of Ω would extend smoothly up to Ω̄. However,
Barret discovered smooth pseudoconvex domains, the so called “worm”
domains, for which the canonical H 0 solution is not smooth. Even more
delicate is the problem of the regularity of the solution u in the part of
∂Ω where f is regular. Consider for example (cf. Kohn [58]) a domain
Ω ⊂ C2 with smooth boundary defined by y2 < 0 for |z1 | < 1. Let χ
be a real C ∞ function which satisfies χ ≡ 1 for |z1 | < 41 and χ ≡ 0 for
¯ χ ) is a ∂-closed
|z1 | > 12 . Then, f := ∂( ¯ 1-form which is C ∞ in the part
z2
1 1
of ∂Ω where |z1 | < 4 and |z1 | > 2 . However, no solution u of ∂u ¯ =f
can be regular in that part of the boundary, because, otherwise, the
function u − zχ2 would be holomorphic in Ω but would be uniformly
bounded in the boundary of the disc ∂∆ × {−iη} and unbounded at
their center in violation of the maximum principle.

1.10. Subelliptic estimates for ∂¯

Summary. We say a few words about the subellipticity and hy-

poellipticity of the system (∂, ¯ ∂¯∗ ) on bounded domains Ω ⊂ Cn with
smooth boundary. This is “1-subelliptic”, that is, elliptic in the inte-
rior of Ω as we have noticed in Remark 1.9.2: so our interest is con-
fined to the boundary ∂Ω. If Ω is strongly q-pseudoconvex, we prove 21 -
< ¯ 2 +||∂¯∗ u||2 +||u||2
subelliptic estimates. They are of type |||u|||21 ∼ ||∂u|| 0 0 0
2
¯∗
for any form in the domain of ∂ of degree ≥ q + 1. These are classical
in a situation which can be referred to as the celebrated “Z-condition”
of Kohn [42]. We end the section by proposing a new and simple proof
of subellipticity of “decoupled finite type” domains. For these domains,
the finite type conditions in the sense of commutators or contact order
of complex curves do coincide. In whatever case, subelliptic estimates
entail local hypoellipticity at the boundary for (∂, ¯ ∂¯∗ ), that is, any H 0
solution is C ∞ (Ω̄) precisely in the part of Ω̄ where the datum is C ∞ (Ω̄).
 1.10. SUBELLIPTIC ESTIMATES FOR ∂¯ 69

We keep the notation of Section 1.9, suppose that M is defined

by r = 0 with ∂r 6= 0 and choose an orthonormal system of (1, 0)-
vector fields ∂ωj , j = 1, ..., n, with ∂ωj r = κjn ; thus, ∂ω1 , ..., ∂ωn−1 and
∂ω̄1 , ..., ∂ω̄n−1 when restricted to M are a basis of the space of complex
vector fields tangential to M of type (1, 0) and (0, 1), denoted by T 1,0 M
and T 0,1 M respectively. Let ∂ω be a vector field in T 1,0 M ; the commu-
tator [∂ω , ∂ω̄ ] is still a tangent vector of C ⊗ T M but not necessarily
of C ⊗ T C M = T 1,0 M ⊕ T 0,1 M . If [∂ω , ∂ω̄ ] ∈ T 1,0 M ⊕ T 0,1 M , we iter-
ate the commuation and consider expressions of type [∂ω1 , [∂ω2 [..., ∂ωm ]]]

where the suffix j in ∂ωj denotes either choice of ∂ω or ∂ω̄ . We call
these expressions commuators of order m.
Definition 1.10.1. ∂ω is said to be of finite type m if iterated com-
mutators of ∂ω and ∂ω̄ of order m produce the “totally real tangential”
vector field TR = √i2 (∂ωn − ∂ω̄n ).
Recall the Levi form LM whose matrix in the ωj ’s basis we have
denoted by (rij )ij . Remember that (cf. (1.9.14)) by contraction with
the (1, 0)-form ωn = ∂r, Cartan formula yields
[∂ω , ∂ω̄ ] = (rij )(u, ū)∂ωn − (r̄ji )(u, ū)∂ω̄n + ...
√
(1.10.1) 2
= (rij )(u, ū)TR + ...
i
where the dots denote combinations of the ∂ωi ’s and ∂ω̄i ’s for i ≤ n − 1.
Thus, [∂ω , ∂ω̄ ] produces TR if and only if (rij )(u, ū) 6= 0. There is a
“real” construction of iterated commutators underlying the complex
one. We write ∂ω = X − iY and identify
T 1,0 Cn (a + ib)∂ω
∼ ∼
& &
(1.10.2) T R2n , aX + bY.
% %
∼ ∼
T 0,1 Cn (a − ib)∂ω̄
This identification induces an identification between tangent vector
fields
T 1,0 M
∼
&
T C M.
%
∼
T 0,1 M
We call these identifications “Re”. We give the “real” counterpart of the
properties of the iterated commutators of ∂ω and ∂ω̄ . First, the iterated
70 1. SEVERAL COMPLEX VARIABLES

commutators of the X and Y can go out of T C M but stay inside T M .

Also, if we denote by Lj either choice of X or Y , then [∂ω1 , [∂ω2 [..., ∂ωm ]]]
produces TR if and only if the same is true of [L1 , [L2 [..., Lm ]]].
Since we need to work with fractional or negative index s, we can not
hesitate any more to introduce the tangential Sobolev space with the
corresponding tangential norm ||| · |||Hϕs (Ω) (or ||| · |||s for short); we also
denote by (·, ·)Tan the associated scalar product. For this, we use the
defining function r of M as the last coordinate, supplement it to a full
system of coordinates (a, r) ∈ R2n−1 × R and denote by ξ = (ξa , ξr ) the
dual coordinates. We then consider the standard s-order elliptic symbol
s
Λs (ξa ) = (1+|ξa |2 ) 2 , denote by FTan the partial Fourier transform with
respect to a, define the tangential microdifferential operator Λs = ΛsTan
−1
with symbol Λs (ξa ) by Λs (u) = FTan (Λs (ξa )FTan (u)), and set
Z 0Z
2
(1.10.3) |||u|||s := Λs (ξa )2 |FTan u(ξa , r)|2 dξa dr.
∞ R2n−1

When ∂ω is of finite type m there is a remarkable consequence.

For any function u ∈ Cc∞ (Ω̄ ∩ B) we have the celebrated “subelliptic
estimate”
(1.10.4) |||TR u|||2−1+ 1 ' ||∂ω u||20 + ||∂ω̄ u||20 + ||u||20 .
2m−1

1
Thus we get control of “(TR ) 2m−1 ” by means of ∂ω and ∂ω̄ . Estimate
(1.10.4) follows from Theorem 2.4.6 below which is in turn a conse-
quence of Lemma 2.4.7 applied to the pair Re ∂ω and Im ∂ω . We call the
attention to this point. Finite type of systems of complex vector fields
implies subellipticity but under the strict assumption that the system
is closed under conjugation (cf. the counterexample by Kohn [60]). We
postpone to Remark 2.4.9 below the analysis of this point.
We state now 12 -subelliptic estimates. In these estimates, the
Sobolev spaces are unweighted and the forms u have compact support
in Ω̄. They apply to a domain Ω which is strongly q-pseudoconvex in
the sense that (1.9.21) holds with the inequality “≥ 0” replaced by
“≥ c” for c > 0. In particular, LM is not the 0-form and M has type
2: we have 12 -subelliptic estimates in this case:
Theorem 1.10.2. Let Ω be smooth and strongly q-pseudoconvex in
a neighborhood B of zo . Then, for any u ∈ Cc∞ (Ω̄ ∩ B)k ∩ D∂¯∗ , we have
¯ 2 + ||∂¯∗ u||2 + ||u||2
(1.10.5) |||u|||21 + |||∂r u|||− 1 < ||∂u|| if ≥ q + 1.
0 0 0
2 2 ∼

Proof. We go back to the proof of Proposition 1.9.1 with the

choice ϕ = 0, and thus in particular δωj = ∂ωj , and modify a little the
 1.10. SUBELLIPTIC ESTIMATES FOR ∂¯ 71

argument therein. We interchange the third sums of the right side of

(1.9.8) according to the following scheme:
||∂ωj u||20 →(1 − )||∂ωj u||20 + ||∂ω̄j u||20
Z
+  [∂ωj , ∂ω̄j ](u)ūdV, +... for j ≤ qo ,
Ω

||∂ω̄j u||20 →(1 − )||∂ω̄j u||20 + ||∂ωj u||20

Z
−  [∂ωj , ∂ω̄j ](u)ūdV + ... for qo + 1 ≤ j ≤ n − 1.
Ω
Hence, the estimate becomes
(1.10.6)
!
X0 X X
¯ 2 + ||∂¯∗ u||2 ≥ 
||∂u|| ||∂ωj uJ ||20 + ||∂ω̄j uJ ||20 + Ru
0 0
|J|=k j≤n−1 j≤n
X0 X Z
+ e−ϕ rij uiK ūjK dS
|K|=k−1 ij ∂Ω

X0 X Z
− (1 − ) rjj |uJ |2 dS
|J|=k j≤qo ∂Ω

X0 X Z
− rjj |uJ |2 dS,
|J|=k qo +1≤j≤n−1 ∂Ω

where Ru is an error term. We denote by Su the term before Ru

in the right side of (1.10.6). For suitable , the sum of the boundary
integrals is positive because Ω is strongly q-pseudoconvex. Also, |Ru| ≤

2
Su + C ||u||20 . Thus, we get in conclusion
 ¯ 2 + ||∂¯∗ u||2 .
(1.10.7) Su − C ||u||20 ≤ ||∂u||0 0
2
By the strong q-pseudoconvexity of M , its Levi form has at least one
positive eigenvalue: in particular, commutators of the ∂ωj ’s and ∂ω̄j ’s
for j ≤ n−1 generate the vector field TR = √i2 (∂ωn −∂ω̄n ) which belongs
to C ⊗ T M but is transversal to C ⊗ T C M when restricted to M . Along
the same lines as in the proof of (1.10.4), we get
(1.10.8) |||TR u|||2− 1 < ||∂ωj u||20 + ||∂ω̄j u||20 + ||u||20 .
2 ∼

On the other hand, ∂ω̄n is not in C ⊗ T M since ∂ω̄n r = 1. Thus, the

system (∂ωj )j≤n−1 , (∂ω̄j )j≤n supplemented by TR spans the whole of
C ⊗ T Cn which implies
(1.10.9) |||u|||21 + |||∂r u|||2− 1 < Su + ||u||20 .
2 2 ∼
72 1. SEVERAL COMPLEX VARIABLES

Combining (1.10.9) with (1.10.7), we get the conclusion.


Remark 1.10.3. Estimate (1.10.5) is said to be the “ 21 -subelliptic
estimate” for (∂, ¯ ∂¯∗ ). It ensures the “hypoellipticity” of the system
¯ ∂¯∗ ), that is, the property
(∂,
¯ ∈ (C ∞ (Ω̄))k+1 , ∂¯∗ u ∈ (C ∞ (Ω̄))k−1 ,
∂u zo zo
0 ∞
u ∈ (H (Ω))kzo ⇒ u∈C (Ω̄)kzo .
This can be checked essentially by the same argument as in Theo-
rem 2.4.5 below but requires a little extra work (cf. [61]). As a result
of hypoellipticity, not only do we have a solution u ∈ (C ∞ (Ω̄))k to
¯ = f when f ∈ (C ∞ (Ω̄))k satisfies ∂f
∂u ¯ = 0, but any H 0 solution u
with ∂¯ u = 0, such as the canonical solution, is smooth in Ω̄. We regain,
∗

in the case of strongly pseudoconvex domains, that is, for q = 0, the

celebrated conclusion that the Bergman projection acts continuously
from (C ∞ (Ω̄))k into itself for any k ≥ 1.
The conclusion of Theorem 1.10.2 is classical since, when Ω is
strongly q-pseudoconvex, then λq+1 > 0 and hence s+ ≥ n − 1 − q.
In this situation, and for any k ≥ q + 1, Ω enjoys the celebrated “Z(k)
condition” (see [42]) which yields subelliptic estimates. We present
now a new result (cf. [2]) which deals with the intermediate situation
between q-pseudoconvexity and strong q-pseudoconvexity. Note that,
in the absence of additional assumptions, the weak q-pseudoconvexity
does not suffice for (1.10.5) as the example by Kohn shows (cf. the
Notes of Section 1.9). Let Ω be q-pseudoconvex and let ∂ωjo be a vector
field in V ⊥ (the orthogonal bundle to V in T 1,0 M for the metric). We
will deal with the property
(1.10.10)
qo
n−1
!
X X X0
rij (z)uiK ūjK − rjj (z)|u|2 − rjo jo (z)|ujo K |2 ≥ 0
ij=1 j=1 |K|=k−1

for any z ∈ M and u ∈ D∂¯∗ of degree k ≥ q + 1.

In Theorem 1.10.2 it has been discussed the finite type 2. As for the
higher type, we have
Theorem 1.10.4. Let Ω be a q-pseudoconvex domain and assume
that for any vector field ∂ωjo of a system {∂ωj }j=q+1,...,n−1 ⊂ V ⊥ of rank
n − 1 − q, we have
(i) (1.10.10) is fulfilled,
(ii) ∂ωjo is of finite type mjo .
 1.10. SUBELLIPTIC ESTIMATES FOR ∂¯ 73

Then, for k ≥ q + 1 and for any u ∈ Cc∞ (Ω̄ ∩ B)k ∩ D∂¯∗ , we have
(1.10.11)
¯ 2 +||∂¯∗ u||2 if  < 1 where m = max mjo .
|||u|||2 +|||∂r u|||−1+ < ||∂u||0 0
∼ 2m−1 jo

Remark 1.10.5. According to Remark 2.4.8 below, we have that

(1.10.11) holds in fact for any  < m1 ; note that, according to [32],
 = m1 is optimal.

Remark 1.10.6. The theorem applies in particular to a pseudo-

1
convex domain, that is, for q = 0: -subelliptic estimates for  = 2m−1
(m = maxj mj ) hold over forms of degree k ≥ 1; in fact, they also
hold for  = m1 according to the remark above. But in this situation we
have that complex curves have order of contact ≤ m with M and the
conclusion follows from [33]; also in order to apply [33], (i) is needless.
However our description of  is more accurate and the method more
elementary.

Remark 1.10.7. Similarly to the conclusions of Remark 1.10.3, the

estimate (1.10.11) implies the hypoellipticity at the boundary for the
¯ ∂¯∗ ). In particular, the canonical solution that we already
system (∂,
know from (1.9.25) to exist in H 0 , is in fact smooth.

Proof of Theorem 1.10.4. We assume, by reordering, that V =

Span{∂ω1 , ..., ∂ωqo }. We fix ∂ωjo for which (i) and (ii) are satisfied
and assume by reordering that ∂ωjo = ∂ωn−1 . We go back to the basic
estimate in Proposition 1.9.1 that we apply twice. Once as it stands
and another time by interchanging
(1.10.12)
X0 X0 X0 X0 Z
2 2 2
||∂ω̄jo uJ ||0 = ||∂ω̄jo uJ ||0 + ||∂ωjo ujo K ||0 − rjo jo |ujo K |2 dS.
|J|=k J63jo |K|=k−1 |K|=k−1 ∂Ω
74 1. SEVERAL COMPLEX VARIABLES

In both cases, the term which involves boundary integrals can be dis-
carded since it is positive. What we get is
(1.10.13)
 
 X 0 X X 0 X
2
||∂ω̄j uJ ||20 





 ||∂ω j
uJ ||0 +
|J|=k j≤qo |J|=k qo +1≤j≤n





 < ||∂u||¯ 2 + ||∂¯∗ u||2 + ||u||2 ,
0 0 0

∼

 

 X0 X X0 X X0 
2 2
||∂ωjo ujo K ||20 





 ||∂ u
ωj J 0|| + ||∂ u
ω̄j J 0|| + 

 |J|=k j≤qo |J|=k qo +1≤j≤n |K|=k−1

j6=jo or j ∈J

 /

¯ ¯
 2 ∗ 2 2
< ||∂u|| 0 + ||∂ u||0 + ||u||0 ,


∼

in the two cases respective. We denote by S1 and S2 the two terms

between brackets involving derivatives in the first and third line of
(1.10.13) respectively. We select a multiindex K; by combining the two
inequalities in (1.10.13) we get in particular
(1.10.14) ¯ 2 + ||∂¯∗ u||2 + ||u||2 .
||∂ωjo ujo K ||20 + ||∂ω̄jo ujo K ||20 < ||∂u||0 0 0
∼

Let TR = √i2 (∂ωn − ∂ω̄n ) be the totally real vector field tangential to M .
Since ∂ωjo has finite type mjo , we know that TR is obtained by mjo − 1
iterated commutators of ∂ωjo and ∂ω̄jo . We then have (cf. Section 2.4)

|||TR ujo K |||2−1+ 1 < ||∂ωjo ujo K ||20 + ||∂ω̄jo ujo K ||20 + ||u||20
2
mjo −1 ∼
(1.10.15)
¯ 2 + ||∂¯∗ u||2 + ||u||2 .
< ||∂u||0 0 0
∼

Recall what has already been said: subellipticity follows from finite type
for systems of real vector fields. Therefore, the same is true for systems
of complex vector fields when these are closed under conjugation such
as the system ∂ωjo and ∂ω̄jo (but it is not necessarily true for general
complex systems). We carry out our proof. We reason in the same
way for any index j of the system {∂ωj } ⊂ V ⊥ for which (i) and (ii)
of the statement of Theorem 1.10.4 hold. By changing the metric in
¯
M (which does not affect the ∂-Neumann conditions) we can assume
that this system is orthonormal. We remark now that, u having degree
k ≥ q + 1, any coefficient uJ can be written as uJ = ujK for some index
j of the system or else uJ = unK . We start from the first; they satisfy
(1.10.16)
¯ 2 + ||∂¯∗ u||2 + ||u||2
|||TR uJ |||2−1+ < ||∂u|| for  = inf 2−(mj −1) .
0 0 0
∼ j
 1.10. SUBELLIPTIC ESTIMATES FOR ∂¯ 75

In this situation, we say that the energy, that is, the quantity in the
right of (1.10.16) contains  derivative TR of these coefficients. The
energy also contains a full ∂ωj (resp. ∂ω̄j ) derivative for 1 ≤ j ≤ qo
(resp. qo + 1 ≤ j ≤ n). The missing barred and unbarred derivatives,
except from ∂ωn , can be “-interchanged” with the complemantary ones
on account of
(1.10.17)
   
|||∂ω uJ |||2−1+ − |||∂ω̄ uJ |||2−1+  ≤  [∂ω , ∂ω̄ ]uJ , uJ  + ||uJ ||20


j j j j −1+

< ||TR uJ ||2−1+ + ||uJ ||20 .

∼

This concludes the proof of (1.10.11) for any coefficient uJ such that
n∈ / J. As for the coefficients unK , we have that barred and unbarred
¯
derivatives are fully interchangeable on account of the ∂-Neumann con-
dition unK |M ≡ 0. In other terms, the energy contains the full ||unK ||1 -
norm.

We have a result of subellipticity for an important class of domains,
the so-called “finite type decoupled” domains. We choose q = qo and
write z 0 = (z1 , ...zq ) and z 00 = (zq+1 , ..., zn−1 ).
Theorem 1.10.8. Let Ω be defined by r < 0 for r = 2Re zn −
n−1
h(z 0 ) + |fj (zj00 )|2 where hij (z 0 ) ≥ 0 in Cqo and the fj (zj00 )’s are
P
j=q+1
holomorphic functions with vanishing order mj at zo = 0. Then -
1
subelliptic estimates hold in degree k ≥ q + 1 for  = 22m−1 for m =
maxj mj .
Proof. Let π : Cn → Cn−1 be the projection over the first n − 1 coor-
dinates; since h and the fj ’s do not depend on zn , that is Ω is “rigid”,
then we have an identification
π0
∼ ¯ n−1
Lr |T C M → ∂ ∂r|Cz ,...,z .
1 n−1

We take ∂ωj = ∂zj − rzj ∂zn , j = 1, ..., n − 1 and ∂ωn = ∂zn as a basis of
(1, 0) vector fields and its dual system ω1 , ..., ωn = ∂r as a basis of (1, 0)
forms; we choose a Hermitian metric in which this is orthonormal. In
this basis we have
 
−(hij ) 0
(rij )i,j≤n−1 = .
0 m2j |zj00 |2(mj −1) κij
It readily follows that Ω is q-pseudoconvex and that (1.10.10) is fulfilled.

76 1. SEVERAL COMPLEX VARIABLES

Example 1.10.9. Let z 0 ∈ Cq , z 00 ∈ Cn−1−q and define

q n−1
X X
Ω = {z : yn > − |zj0 |2 + |zj00 |2mj }.
j=1 j=q+1
1
Then (1.10.11) holds for k ≥ q + 1 and with  = 22m−1
.
According to Remark 2.4.8, we have in fact subelliptic estimates for
1
any  < 2m . Again, this result could be obtained from [33] but for a
worse index .

Suggested research. Suppose that Ω is q-pseudoconvex and

satisfies the finite contact order condition (in the sense of D’Angelo)
with every (k − qo )-dimensional complex submanifold transversal to V.
In this situation, do we have subelliptic estimates for forms of degree
k ≥ q +1? Are the finite contact conditions necessary for the subelliptic
estimates?

Notes. “Finite type” is here understood in the sense of bracket

as was originally introduced by Kohn in [56]. Subsequently, Kohn in-
troduced “finite ideal type” and proved that it is sufficient for subel-
liptic estimates. D’Angelo [37] proposed another notion of finite type
through the order of contact of complex varieties. The necessity of this
for subelliptic estimates was proved by Catlin who gave subsequently in
[33] the ultimate characterization. All this literature deals with pseudo-
convex domains. As for strongly q-pseudoconvex domains, subelliptic
estimates were established by Folland and Kohn in [42] under the so-
called “Z-condition” which turns out to hold for any k ≥ q + 1. Our
approach is inspired by the guidelines [42] assimilated into the style of
Section 1.9. For weakly q-pseudoconvex domains of finite type which
satisfy the conditions of Theorem 1.10.4 the estimates have been re-
cently established by Ahn, Baracco and Zampieri [2].
 CHAPTER 2

Real Structures

Summary of Chapter 2. In Section 2.1 we start with analysis on

Euclidean spaces and then pass to manifolds, spaces which locally look
like Rn . We then present, in Section 2.2, systems of vector fields and
deal with the problem of their integration. It turns out that the formal
integrability, the so-called involutivity, ensures the effective integrabil-
ity (Frobenius theorem). Next, in Section 2.3, we present real symplec-
tic manifolds and prove that they are all symplectically diffeomorphic
to T ∗ Rn (Darboux theorem). We also give the canonical form of the
submanifolds of T ∗ Rn over which the canonical 2-form σ has constant
rank. In particular, this describes all involutive/isotropic/lagrangian
submanifolds. We then present, in Section 2.4, the Hörmander num-
bers associated to a system of vector fields and the related condition
of finite type m. We prove that it ensures subelliptic estimates with a
1
gain of 2m−1 derivatives; these latter, regardless of the amount of the
gain, imply in turn local hypoellipticity. We end the discussion by in-
troducing, in Section 2.5, the notion of the orbit of a set of vector fields.
In case the set is C ω , the orbit is uniquely defined and coincides with
the integral manifold of the Lie span (Nagano theorem). Otherwise, for
a smooth set, we can find in any case a unique minimal manifold, the
local orbit, stable under the set (Sussmann’s theorem).

2.1. Euclidean spaces and their diffeomorphisms

Let F be a map of class C k , k ≥ 1, from Ω ⊂ Rn to Ω̃ ⊂ Rm . Let
xo be a point of Ω and F 0 (xo ) : Rn → Rm be the tangent map to F at
xo . It is defined over tangent vectors v of Rn , say column vectors, as
F (xo + tv) − F (xo )
(2.1.1) F 0 : v → lim .
t→0 t
Let (fj )j=1,...,m be the components of F and let
 
∂x1 f1 . . . ∂xn f1
 . . 
JF =  .

. 
∂x1 fm . . . ∂xn fm
77
78 2. REAL STRUCTURES

be the Jacobian matrix of F ; sometimes, we also denote it by Jx F or

∂x F . The linear map F 0 can be identified with the matrix JF in the
sense that
(2.1.2) F 0 (v) = JF · v,
where in the right side we have the usual product between matrices
and vectors. Let t F 0 (x̃o ) : Rm → Rn be the cotangent application to
F at x̃o = F (xo ) defined over cotangent vectors w at x̃o , say column
vectors w ∈ Rm , by the identity over tangent vectors v at xo
ht F 0 (x̃o )w, vixo = hw, F 0 (xo )vix̃o .
Here h·, ·i denotes the duality between tangent and cotangent vectors.
These dualities at xo and x̃o are in fact identified with the inner product
which defines the euclidean structure in Rn and Rm
X
hw, vi = w j vj .
j

It is obvious that the matrix associated to the linear map t F 0 is

t
(Jx F ), the transpose of the Jacobian.
Remark 2.1.1. Sometimes, it is more convenient to denote cotan-
gent vectors as row vectors t w; in this case the inner product is written
hw, vi = t w · v
and the map t F 0 is now represented by Jx F .
We now assume that F is a diffeomorphism of an open subset
Ω ⊂ Rn ; this is also called a change of coordinates F : x 7→ x̃ = x̃(x).
In particular, we now have n = m. Along with F 0 (xo ) : Rn → Rm we
now also have at our disposal t F 0−1 (xo )(= t (F −1 )0 (xo )) : Rn → Rm .
(Because of inversion, t F 0−1 (xo ) “goes in the same sense” as F 0 (xo ).)
By means of F 0 and t F 0−1 we can thus describe the effect of a change of
coordinates over tangent and cotangent vectors. How does it act over
derivations and differentials? First, the coordinate derivatives trans-
form according to the rule
X
(2.1.3) ∂ xj → ∂xj (x̃i )∂x̃i .
i

In fact, for a function f = f (x̃), Punder the change
x̃ = x̃(x), the

chain rule yields ∂xj (f (x̃(x))) = ∂
i xj (x̃ )∂
i x̃i f (x̃(x)).
PFor a general
derivation along the direction v = (vj )j , that is, ∂v := j vj ∂xj , if we
apply (2.1.3), we get
XX
(2.1.4) ∂v → ( ∂xj (x̃i )vj )∂x̃i .
i j
 2.1. EUCLIDEAN SPACES 79

But in the basis ∂x̃i the coefficients on the right side of (2.1.4) are
exactly those of ṽ := Jx x̃ · v. We have thus obtained
∂v → ∂ṽ .
Hence, under coordinate change, derivations and tangent vectors trans-
form accordingly. This allows
P us to identify, as we intensively do, vectors
v and derivations ∂v = j vj ∂xj . As for the basis of differential dxj ,
dual to ∂xj in the inner product h·, ·i, the chain rule applied to the
inverse transformation xj = xj (x̃) yields
X
dxj → ∂x̃i (xj )dx̃i .
i
P
For a general differential, that is, a combination w = j wj dxj , we get
!
X X
(2.1.5) w→ ∂x̃i (xj )wj dx̃i .
i j

But the coefficients on the right are those of (t Jx̃ x)w = t (Jx x̃)−1 w
which we also denote by w̃, and therefore (2.1.5) reads
w → w̃.
We are ready to enlarge the setting of our analysis and deal with the
diffeomorphic images of the euclidean space which is the local structure
of an immersed manifold. In greater generality, a manifold M is the
∼
union of the homeomorphic images Uj → Mj of open sets Uj ⊂ Rn such
Φj
that if Φjk (resp. Φkj ) is the homeomorphism induced by Φj (resp. Φk )
on overlappings Mj ∩ Mk 6= ∅, then
Φjk ◦ Φ−1
kj is a diffeomorphism.

We are mainly interested in immersed manifolds, which, in real geom-

etry, is not a significant restriction.
Definition 2.1.2. A real l-codimensional submanifold M of Rn of
class C k , k ≥ 1, is, in a neighborhood B ⊂ Rn of each point xo ∈ M ,
∼
the image U → M ∩ B of an open set U ⊂ Rn−l under a map Φ whose
Φ
differential has maximal rank.
Thus, M is locally a C k smooth deformation of an open set of the
euclidean space Rn−l . The mapping Φ is a diffeomorphism of U with its
image and is called a local chart or a local parametrization of M and
its inverse Φ−1 is called a local system of coordinates. There is another
possible definition, so-called implicit, of a manifold.
80 2. REAL STRUCTURES

Definition 2.1.3. A subset M of Rn is said to be a submanifold

of codimension l and class C k if for any xo ∈ M there is V ⊂ Rn and
a system of C k functions g1 , ..., gl with independent differentials such
that M 0 = {x ∈ V : gj (x) = 0 for any j}.
We want to show that the parametrization can be chosen to be
“cartesian”, that is, with U being an open set of a coordinate plane in
a group x0 of variables or equivalently that the equations can take the
form x00 − E(x0 ) = 0 where the x00 are the complementary variables.
(For this reason it has been better to take U instead of the whole of
Rn−l in Definition 2.1.2.)
Theorem 2.1.4 (Implicit function theorem). Definitions 2.1.2 and
2.1.3 are equivalent.
Proof. We set G = t (g1 , ..., gl ) and denote by JΦ and JG the
Jacobians of Φ and G, respectively; these are n × (n − l) and l × n
matrices, respectively. Remember that according to Definitions 2.1.2
and 2.1.3 they are assumed to have rank n − l and l, respectively. Thus,
we can choose a group of n − l variables x0 and the complementary
variables x00 such that Jx0 Φ and Jx00 G have at xo maximal rank n − l
and l, respectively. In both cases it follows that the projection π : Rn →
∼
Rn−l , x 7→ x0 induces a local diffeomorphism M 0 → U between an open
π
neighborhood M 0 of xo on M and an open set U ⊂ Rn−l . We denote
by x0 7→ (x0 , E(x0 )) the inverse of π|M 0 : thus, M 0 is now parametrized
by Φ = id × E or defined by G = 0 for G = x00 − E(x0 ).

We also deal with manifolds with boundary M + . They are manifolds
in the usual sense at interior points whereas, at a boundary point x ∈
∼
∂M + , there is a diffeomorphism U + → M + where U + is a half-space in
Φ
Rn−l and Φ extends as a diffeomorphism beyond the boundary. They
are therefore manifolds truncated by a transversal inequality gl+1 > 0.
The tangent space to M at x, denoted by Tx M , is defined as the
space of vectors of Rn proportional to the limits
xν − x
(2.1.6) v = xlim .
ν →x |x − x|
xν ∈M ν

Since M is defined by G = 0, then differentiation yields

(2.1.7) Tx M = {v ∈ Rn : JG · v = 0}.
On the other hand, differentiation of the identity G ◦ Φ ≡ 0 yields
range(JΦ) ⊂ ker(JG); but the two spaces have the same dimension by
 2.1. EUCLIDEAN SPACES 81

our assumption, which implies

(2.1.8) range(JΦ) = ker(JG).
So far, there could be some ambiguity in our description of tangent
vectors. First, if we change coordinates in Rn , how does this affect the
limit of our differences (2.1.6)? Next, if we change parametrization Φ or
equation G = 0, how does this affect the description of Tx M ? Now, we
have seen that under a change x̃ = x̃(x) the limits change as ṽ = Jx x̃·v.
As for the case of two different parametrizations
U1
Φ1
∼
!
Θ
=
M 0,
∼
 Φ2
U2
differentiation of the identity Φ2 = Φ1 ◦ Θ−1 yields
JΦ2 = JΦ1 · (JΘ)−1 .
Last, two different systems of equations G = 0 and G1 = 0 would be
related by gj1 = i αji gi for an invertible square matrix (αji )j,i=1,...,l .
P
This yields
JG1 = J(αji )ji · JG,
which shows again the independence of the tangent space from the
choice of the equation of M .
The distribution of the tangent planes Tx M, x ∈ M , is called the
tangent bundle and is denoted by T M . The distributions Tx∗ M, x ∈ M ,
of linear forms on Tx M , is called the cotangent bundle and is denoted
by T ∗ M . A C k smooth section L ∈ T M , resp. θ ∈ T ∗ M , is called a
vector field, resp. a differential form. It consists of a smooth function
which assigns to each x ∈ M a vector Lx ∈ Tx M , resp. θx ∈ Tx∗ M . In
a local coordinate system x = (x1 , ..., xn ), L and θ can be represented
by X X
Lx = aj (x)∂xj and θx = αj (x)dxj ,
j j
where each coefficient aj and αj is defined over a local chart U →
Φ
M 0 , t 7→ x(t) and depends in C k fashion on t. A vector field L is
described by its action over smooth functions f that we denote by
L(f ) or Lf , and a differential θ is described by the pairing with vector
fields hL, θi. Sometimes, different coordinate systems are in use near
the same xo ∈ M . Our previous work explains how the coordinates of
82 2. REAL STRUCTURES

a vector field or a differential are described in two different systems x

0 00 0 00
on
P M and x̃ on PM at common points of M ∩ M . Let L be written as
j aj ∂xj and j bj ∂x̃j : we have already seen that the coefficients are
related by (bj ) = (Jx x̃) · (aj ). (This could be reproved by remarking
that ∂x̃j x̃k = κjk (Kroneker symbol), and therefore

bj = L(x̃j )
X
= ak ∂xk (x̃j ).)
k
P
In the same way, if θ is written in the two systems as θ = j αj dxj and
θ = j βj dx̃j , then the coefficients are related by (βj ) = t (Jx̃ x)(αj ).
P

Let F be a C k map from a manifold M to a manifold N . Recall

that tangent vectors (cotangent vectors) are the same as derivations
(differentials). Our previous discussion explains that F induces a tan-
gent map

T M → M ×N T N, (x, v) 7→ (x, F (x); F 0 (x) · v),

whose action over a vector field L on M , denoted by F∗ L and called

the “push forward” of L by F , is defined by means of composition with
F:
F∗ L(g) = L(g ◦ F ) for any g smooth on N .
We identify F∗ = F 0 . In the same way, F induces a cotangent map

M ×N T ∗ N → T ∗ M, (x, F (x); w) 7→ (x, t F 0 (x) · w),

whose action on smooth sections θ, called the “pull back” of θ and

denoted by F ∗ θ, is given by

hF ∗ θ, Lix = hθ, F∗ LiF (x) for any L.

Again, we identify F ∗ = t F 0 . Note that under the choice F = Φ, where

Φ is a parametrization of M , we can define by the aid of F∗ and F ∗ two
“distinguished” bases {Lj } and {θj } of T M and T ∗ M , dual to each
other. If t = (tj )j=1,...,n−l are coordinates in Rn−l , these are defined by

Lj = F∗ ∂tj , j = 1, ..., n − l,

resp.
θj = (F −1 )∗ dtj , j = 1, ..., n − l.
 2.2. INTEGRATION OF VECTOR BUNDLES 83

2.2. Integration of vector fields and vector

bundles—Frobenius theorem
A vector field of class C m , m ≥ 1, over an open domain Ω ⊂ Rn ,
is a C m -section n
P of T R |Ω ; in a system of local coordinates x it takes
m
the form L = j aj (x)∂xj with the aj ’s of class C . We know from the
local existence and uniqueness of the solution of the Cauchy problem
that for any point x ∈ Ω there is a unique integral curve of L, that
is, a curve Γ parametrized, say, by t 7→ ϕ(t), such that ϕ̇(t) is parallel
to L(ϕ(t)) for any t. Also, if Σ is any hypersurface transversal to L,
that is, satisfying Tx Σ ∩ L(x) = {0} for any x, then we have a smooth
parametrization of this system of curves by means of the intersection
point ξ of Γ with Σ. In other words, we have a local diffeomorphism
(2.2.1) (ξ, t) 7→ x = ϕξ (t).
This tells us that (ξ, t) serves as a new system of coordinates in which
the integral curves Γξ are “straightened”, that is, they take the trivial
equation
ξ ≡ const.
We want to generalize this conclusion to general systems of vector fields.
Definition 2.2.1. A Frobenius distribution or bundle E of rank k
on Ω is a distribution of vector spaces Ex , x ∈ Ω, of constant dimension
k.
Let L be a vector field; if L(x), which we also denote Lx , belongs
to Ex for any x, then L is said to be a “section” of E.
Definition 2.2.2. A Frobenius distribution or bundle E of rank
k over Ω is said to have class C m when it is spanned by a system of
vector fields L1 , ..., Lk of class C m , a basis of E.
A vector field L is in a natural way an operator of derivation: its
action on a smooth function g is denoted by L(g).
Definition 2.2.3. The commutator or Lie bracket of two vector
fields L and W is defined by
[L, W ] = L ◦ W − W ◦ L.
P P
Proposition 2.2.4. Let L = j aj (x)∂xj and W = j bj (x)∂xj be
vector fields of class C m , m ≥ 2, on a domain Ω ⊂ Rn . We have
n n
!
X X
(2.2.2) [L, W ] = ai ∂xi (bj ) − bi ∂xi (aj ) ∂xj .
j=1 i=1
84 2. REAL STRUCTURES

Proof. Let g be a smooth function on Ω. We have

(2.2.3)
Xn
(ai ∂xi (bj ) − bi ∂xi (aj )) ∂xj (g)
j,i=1
n n
! n
X X X
= ai ∂xi bj ∂xj (g) − ai bj ∂x2i xj (g)
i=1 j=1 i,j=1
n n
! n
X X X
− bi ∂ x i aj ∂xj (g) + aj bi ∂x2j xi (g)
i=1 j=1 j,i=1
n n
! n n
!
X X X X
= ai ∂xi bj ∂xj (g) − bi ∂ x i aj ∂xj (g)
i=1 j=1 i=1 j=1

= (L ◦ W − W ◦ L)(g) = [L, W ](g),

where the second equality follows from the commutativity of the second
derivatives ∂x2i xj (g) = ∂x2j xi (g).

Remark 2.2.5. The Lie bracket [L, W ] between vector fields of class
m ≥ 2 enjoys the following properties whose proof is straightforward:
(i) [L, W ] = −[W, L] and in particular [L, L] = 0 (alternance).
(ii) [L + W, Z] = [L, Z] + [W, Z] and the same for the second term
(additivity).
(iii) For C 1 functions f and g, we have [f L, gW ] = f L(g)W −
gW (f )L + f g[L, W ].
(iv) [L, [W, Z]] + [W, [Z, L]] + [Z, [L, W ]] = 0 (Jacobi’s identity).
Definition 2.2.6. A C m , m ≥ 2, bundle E is said to be “involu-
tive” or “integrable” when for any pair of sections L, W of E, their Lie
bracket [L, W ] is still a section of E.
Thus, E is involutive when it is stable under Lie bracket. To check
the involutivity of E, it suffices to test the property over the elements
of a basis. In fact we have
Proposition 2.2.7. Let L1 , ..., Lk be a basis of E, that is, at any
x ∈ Ω, the vectors L1 (x), ..., Lk (x) are a basis of Ex . Then, E is invo-
lutive if (and only if ) for any j = 1, ..., k the Lie bracket [Li , Lj ] is a
section of E.
Proof. We have to prove only the “if” part of P the statement.
For this, we remark that for general vector fields L = j aj Lj , W =
 2.2. INTEGRATION OF VECTOR BUNDLES 85
P
j bj Lj ,we have
X X
[L, W ] = [ ai L i , bj L j ]
i j
X X X
= ai bj [Li , Lj ] + ai Li (bj )Lj − bj Lj (ai )Li .
i,j i,j i,j

The last two terms in the second equality belong to E. Thus, we have
the fact that [L, W ] ∈ E is reduced to [Li , Lj ] ∈ E for any i, j.

Definition 2.2.8. A k-dimensional manifold Γ is said to be an
integral leaf of the bundle E when Tx Γ = Ex for any x ∈ Γ.
Our problem is to state under which condition on E there is the
unique integral leaf of E through each point. If E has rank k = 1,
we have already seen that the problem has a positive answer. In the
general situation, the answer is contained in the following
Theorem 2.2.9 (Frobenius theorem). Let E be a bundle on Ω of
rank k and class C m , m ≥ 2. The following are equivalent:
(i) E is involutive.
(ii) For any xo ∈ Ω there is a unique local integral leaf Γ of E
passing through xo .
(iii) For any xo ∈ Ω there are local coordinates x = (x0 , x00 ), x0 =
(x01 , ..., x0k ), such that E = Span{∂x01 , ..., ∂x0k } in a neighborhood
of xo .
Remark 2.2.10. A system of manifolds {Γ} as in (ii), that is, a local
covering of Ω by disjoint manifolds, is said to be a “local foliation” of
Ω. According to (iii), it is possible to reduce a foliation, under a local
coordinate change, to the system of k-planes defined by x00 = const. In
fact, reducing the bundle to Span{∂x0j }j=1,...,k is equivalent to reducing
its integral leaves to x00 = const. Thus, under a C 1 diffeomorphism, the
leaves of a foliation are always “straightenable”. We remark now that
the changes of coordinates which preserve the leaves x00 = const, or the
bundle Span{∂x0j }j=1,...,k , are those in the form
(
x̃0 = x̃0 (x0 , x00 ),
(2.2.4)
x̃00 = x̃00 (x00 ).
We also remark that these changes do not need to preserve each single
∂x0j but, as has already been said, the set of their linear combinations.
This freedom is the key point in the proof of Theorem 2.2.9, which
consists of two steps: linear algebra to adjust the basis {Lj } so that
86 2. REAL STRUCTURES

the terms commute, that is, [Li , Lj ] = 0 and a coordinate change of type
(2.2.4) provided by the solution of Cauchy’s problem, which transforms
Lj for j ≤ k into ∂x0j .

We start from this latter point, which we present as a separate

statement.

Proposition 2.2.11. Let Lj , j = 1, ..., k, be a system of indepen-

dent vector fields which satisfy

[Li , Lj ] = 0, i, j = 1, ..., k.

Then, in new local coordinates x, we have Lj = ∂x0j , j = 1, ..., k.

Proof. We reason by induction on k. When k = 1, on one hand

we have a bundle spanned by a single vector field which is always
involutive. On the other hand, this can always be reduced to a co-
ordinate derivative ∂x1 by Cauchy’s theorem. Suppose now that the
proposition is true for k. There are coordinates x = (x0 , x00 ) with
x0 = (x01 , ...,P
x0k ), x00 = (x00k+1 , ..., x00n ), such that L1 = ∂x01 , ..., Lk = ∂x0k .
Let Lk+1 = j aj (x)∂xj ; from the hypothesis [Lk+1 , Lj ] = 0 for j ≤ k,
it follows that aj (x) = aj (x00 ). We may also assume without loss of
generality that ak+1 6= 0. We look for new coordinates x̃ such that
Lj = ∂xj → ∂x̃j for j ≤ k and Lk+1 → ∂x̃k+1 . To obtain the first, the
transformation must be of type (2.2.4). Moreover, since we are requir-
ing in fact that each single ∂xj , j ≤ k, is preserved, we must ask that
the transformation be of the type
(
x̃0j = xj + bj (x00 ),
(2.2.5)
x̃00 = x̃00 (x00 ).

As for the second point, the reduction of Lk+1 to ∂x̃k+1 , we are asking
that
X X
aj ∂ x j = ∂x̃k+1 (xj )∂xj ,
j j

that is,

∂x̃k+1 (xj ) = aj , j = 1, ..., n.

 2.2. INTEGRATION OF VECTOR BUNDLES 87

Now, to find x = x(x̃), whose inverse is the transform we are looking

for, we just have to solve the ordinary system of differential equations
in the variable x̃k+1 , with the other arguments x̃j as parameters, with
initial values at x̃k+1 = 0:


∂x̃k+1 (xj ) = (
 aj , j = 1, ..., n,
(2.2.6) 6 k + 1,
x̃j , j =
xj |x̃k+1 =0 =

0, j = k + 1.

Since aj = aj (x00 ), then the solutions of (2.2.6) satisfy x00 = x00 (x̃00 )
which is the crucial point of the proof. These solutions yield the explicit
description of the change of coordinates

(
xj = x̃j + aj (x00 (ξ˜j , x̃00k+2 , ..., x̃00n ))x̃k+1 for j 6= k + 1,
(2.2.7)
xk+1 = ak+1 (x00 (ξ˜k+1 , x̃00k+2 , ..., x̃00n ))x̃k+1 ,

where ξ˜j is the point in the interval (0, x̃00k+1 ) of mean value for
∂x̃k+1 xj (·, x̃00k+1 , ..., x̃00n ); thus (2.2.7) is certainly of type (2.2.5).


Proof of Theorem 2.2.9. (iii) ⇒ (ii) is obvious. (ii) ⇒ (i) is easy:

if rj = 0 is a system of equations for Γ with Tx Γ = Ex , then from
Lrj |Γ ≡ 0 and W r|Γ ≡ 0, we get (L ◦ W − W ◦ L)rj |Γ ≡ 0. P
(i) ⇒ (iii): let L1 , ..., Lk be a basis for E and write Li = j aij (x)∂xj .
We want to find a new basis which commutes. The matrix (aij (x))ij has
rank k at every point. By a linear change of coordinates, we can assume
that the k × k submatrix of the first k columns is non-degenerate. In
particular,

Span{L1 , ..., Lk } ∩ Span{∂xk+1 , ..., ∂xn } = {0}.

We can then solve in the unknown (zih )i=1,...,k the linear systems

k
X
aij zih = κjh , j = 1, ..., k,
i=1
88 2. REAL STRUCTURES

where κjh is the Kronecker symbol. These are a set of k systems with
k
index h = 1, ..., k. We then put L̃h = zih Li . We have
P
i=1

k
X n
X
h
L̃h = zi aij (x)∂xj
i=1 j=1
k k n
!
X X X
= zih aij (x)∂xj + aij (x)∂xj
i=1 j=1 j=k+1
(2.2.8) ! !
k
X k
X n
X k
X
= aij (x)zih ∂xj + aij (x)zih ∂xj
j=1 i=1 j=k+1 i=1
n
X
= ∂xh + chj ∂xj ,
j=k+1

k
aij (x)zih for any j ≥ k +
P
where we have used the notation chj :=
i=1
1. Hence, [L̃i , L̃j ] belongs to Span{∂xk+1 , ..., ∂xn } but it also belongs
to Span{L̃1 , ..., L̃k } which is transversal to the former. It follows that
[L̃i , L̃j ] = 0; we can thus apply Proposition 2.2.11, which yields the
conclusion of the proof of Theorem 2.2.9.

For further application to symplectic geometry we need the follow-
ing variant.
Corollary 2.2.12. Let E = Span{L1 , ..., Ln } be an involutive
bundle and let the constants cijr be defined by the relations [Li , Lj ] =
∞
P
r cijr Lr . Let h1 , ..., hk ∈ C satisfy
X
(2.2.9) Li hj − Lj hi = cijr hr .
r

Let S be a manifold with tangent plane supplementary to E and pick a

function uo ∈ C ∞ (S). Then, there is a unique solution u ∈ C ∞ of the
initial value problem
(
Lj u = hj , j = 1, ..., k,
(2.2.10)
u|S = uo .

Note that (2.2.9) in clearly necessary for (2.2.10) (regardless of the

boundary condition on S).
 2.2. INTEGRATION OF VECTOR BUNDLES 89

Proof. The hypothesis (2.2.9) as well as the conclusion (2.2.10) is

obviously invariant under change of coordinates. According to Theo-
rem 2.2.9, we can find coordinates in which E = Span{∂xn−k+1 , ...∂xn }.
It then suffices to prove the theorem for the system ∂xn−k+1 , ..., ∂xn .
Under this choice, (2.2.9) becomes
∂xi hj − ∂xj hi = 0.
P
By this, for fixed x1 , ..., xn−k , the form hj dxj is closed. Let S be
j≥n−k+1
graphed over x0 = x1 , ..., xn−k by x00 = s(x0 ) and consider the function
Z x00 X
0 0
u(x) = uo (x , s(x )) + hj dxj ,
s(x0 ) j≥n−k+1

where integration is performed along a curve of Rk connecting s(x0 ) to

x00 . This is well defined and provides the unique solution to (2.2.10).

Example 2.2.13. Let M ⊂ Rn be a manifold of class C 2 . We can see
that in the neighborhood of any point of M there is a C 1 “orthogonal
projection”:
p : Rn → M, x 7→ p(x),
which associates to x the point of minimal distance p(x) on M . In fact,
let Tx M, x ∈ M , be the tangent space at x and Tx M ⊥ the orthogonal
space in the euclidean product. We set
(2.2.11) (T M ⊥ ) = {(x, ξ) : x ∈ M, ξ ∈ (Tx M )⊥ , |ξ| = },
(2.2.12) (Rn \ M ) = {x ∈ Rn \ M : d(x, M ) = },
where d is the euclidean distance. We then have
∼
(2.2.13) (Rn \ M ) → (T M ⊥ )  M,
(2.2.14) y 7→ (x, ξ) 7→ x,
where the first correspondence identifies y to the unique (x, ξ) such
that y = x + ξ and the second is the projection on the first component.
Note that we have a natural foliation of a neighborhood of Rn \ M by
the family of hypersurfaces
M = {x ∈ Rn \ M : d(x, M ) = }.
We have
Tx M = (x − p(x))⊥
= Tp(x) M.
90 2. REAL STRUCTURES

Thus, the family {M } is the system of the integral leaves of

(2.2.15) Ex = Tp(x) M.
In particular, the bundle E defined by (2.2.15) is involutive. Note how-
ever that not every bundle E defined by (2.2.15), without some ad-
ditional assumptions about the “projection” p, that in the previous
discussion is “orthogonal”, is involutive. To prove it, we reason back-
wards. We take a non-integrable (n − 1)-dimensional bundle E in Rn
and choose M as the unit sphere. We define locally p : Rn \ M → M
by continuously sending x to either of the two points p(x) ∈ M which
satisfies (2.2.15) or equivalently
p(x) − 0 = Ex⊥ .
Though E is related to a projection p by means of (2.2.15), it is nev-
ertheless non-integrable by the way we chose it.

2.3. Real symplectic spaces—Frobenius-Darboux theorem

Let M be Rn or a real submanifold of dimension n in Rn+l . Let T M
be the tangent bundle to M and T ∗ M its dual bundle, the cotangent
bundle. The fiber Tx∗o M at xo ∈ M is the space of differentials in
xo , that is, the linear forms on Txo M . Take coordinates x in M and
the canonically associated coordinates (x, ξ) in T ∗ M : the ξ are the
coordinates of a form in the canonical basis dx1 , ..., dxn . Remember
that if the base coordinates change as x̃ = f (x), then the ξ change
through ξ˜ = (t f 0 )−1 (x) · ξ (the ξ are here column vectors). We define on
T ∗ M a “symplectic homogeneous structure” by means of the canonical
1-form n
X
ω= ξj dxj
j=1
and the canonical 2-form
n
X
σ = dω = dξj ∧ dxj ,
j=1

respectively. We often write ξdx and dξ ∧ dx instead of the above sums.

Exercise 2.3.1. Let π : T ∗ M → M be the projection, let t π 0 :
T M ×M T ∗ M ,→ T ∗ T ∗ M be the transposed injection and let δ :
∗

T ∗ M ,→ T ∗ M ×T ∗ M T ∗ M be the diagonal immersion. Prove that

ω(x, ξ) = t π 0 ◦ δ(x, ξ).
(Hint: in coordinates, t π 0 is the mapping
((x, ξ), (x, η)) 7→ (x, ξ; ηdx, 0dξ),
 2.3. REAL SYMPLECTIC SPACES—FROBENIUS-DARBOUX THEOREM 91

δ the embedding (x, ξ) 7→ ((x, ξ), (x, ξ)) and hence t π 0 ◦ δ(x, ξ) = ξdx.)
Thus we learned from Exercise 2.3.1 that the forms ω and σ can be
intrinsically defined without use of coordinates. A change of coordinates
˜ that is, a C 1 diffeomorphism of T ∗ M , is said to be
F : (x, ξ) → (x̃, ξ),
“symplectic” when it preserves σ, that is, (F −1 )∗ σ = σ̃.
Note that σ is non-degenerate, that is, if v is a vector of T T ∗ M such
that σ(v, ·) ≡ 0 on T T ∗ M , then v = 0. Let h·, ·i be the duality between
T T ∗ M and T ∗ T ∗ M . Associated to σ, there is therefore a “Hamiltonian”
isomorphism which interchanges the duality with the inner product σ:

(2.3.1) H : T ∗ T ∗ M → T T ∗ M, u 7→ Hu,
where Hu is defined by σ(v, Hu) = hv, ui, for any v ∈ T T ∗ M .
In coordinates (x, ξ) of T ∗ M and for u = u0 dx + u00 dξ, the isomorphism
H is written as
H(u0 dx + u00 dξ) = u00 ∂x − u0 ∂ξ .
For a real smooth function f = f (x, ξ) on T ∗ M , we define its Hamil-
tonian vector field by
Hf = H(df )
= ∂ξ f ∂x − ∂x f ∂ξ .
We define the “Poisson bracket” of two functions on T ∗ M by
{f, g} = Hf (g)
= ∂ξ f ∂x g − ∂x f ∂ξ g.
We have the equality
H{f,g} = [Hf , Hg ],
whose proof is straightforward. In particular, the product {·, ·} enjoys
the same properties as [·, ·]: alternancy, additivity, Jacobi identity. For
˜ we can check whether it is
a change of coordinates F : (x, ξ) 7→ (x̃, ξ),
symplectic just by inspecting if it preserves the Poisson relations
{x̃i , x̃j } = 0, {ξ˜i , ξ˜j } = 0, {x̃i , ξ˜j } = κij .
Note that these relations are trivial for the initial system (x, ξ).
We use the superscript ⊥ σ to denote the orthogonal with respect to
σ.
Definition 2.3.2. A submanifold V of T ∗ M is said to be involutive
if and only if T V ⊥ σ ⊂ T V .
92 2. REAL STRUCTURES

Remark 2.3.3. Let V be defined by f1 = 0, ...., fl = 0 with df1 ∧...∧

dfl 6= 0 (that is, rankt (f1 ...fl )0 = l). Then T V ⊥ σ = Span{Hf1 , ..., Hfl }.
In fact, Span{df1 , ..., dfl } is the orthogonal space to V in the duality
h·, ·i, and H interchanges this duality with the inner product σ.
The following relates the notion of involutivity of manifolds to the
former involutivity of bundles.
Theorem 2.3.4. The submanifold V ⊂ T ∗ M is involutive if and
only if the bundle T V ⊥ σ is involutive, that is, closed under [·, ·].
Proof. The submanifold V is involutive iff
Hfi (fj )|V ≡ 0 for any i, j
iff
{fi , fj }|V ≡ 0 for any i, j
iff
l
X
{fi , fj } = chij fh for any i, j and for suitable coefficients chij
h=1
iff
l
X
[Hfi , Hfj ] = chij Hfh for any i, j and for suitable coefficients chij
h=1
iff
[Hfi , Hfj ] ∈ Span{Hf1 , ..., Hfl }
= TV ⊥σ for any i, j.

The above proof shows that involutivity is also characterized by
{fi , fj }|V ≡ 0 for fi and fj ranging through a system of equations of
V . We can define a symplectic homogeneous structure in an abstract,
conic, real even-dimensional manifold Y just by requiring the presence
of a closed non-degenerate homogeneous 2-form σ; in this manifold
Y , we can define accordingly involutive submanifolds V . But this is
only an illusory generalization: locally, Y still turns out to be T ∗ Rn
under different coordinates. Before proving it, we want to define the
above notions properly . We say that a manifold Y is conic if it is
locally parametrized over a cone of Rm \ {0} by a mapping Φ and
endowed with a “multiplication” Mt for t ∈ Ṙ+ , the push-forward via
Φ∗ of the multiplication by t in Rm \ {0}. In this situation, Φ−1 is
a system of homogeneous coordinates (of degree 1). When we choose
Y = T ∗ M \(M ×{0}), the conic structure consists in the multiplication
 2.3. REAL SYMPLECTIC SPACES—FROBENIUS-DARBOUX THEOREM 93

only along the fibers. Thus, the homogeneous map Φ−1 is expressed, in
terms of usual coordinates, by (x, ξ) 7→ (x|ξ|, ξ). In a conic manifold
there is intrinsically defined a “radial” vector field ρ. Its action on a
C 1 function f : Y → R is expressed by
d
(2.3.2) ρ(f ) = Mt∗ f |t=1 .
dt
In particular, in the coordinates of T ∗ M , we have ρ =
P
j ξj ∂ξj . It
is readily seen that a function is homogeneous of degree µ, that is, it
satisfies
Mt∗ f = tµ f,
if and only if it satisfies the Euler relation
ρf = µf.
In particular, in T ∗ M , x and ξ are homogeneous of degree 0 and 1, re-
spectively. When Y is symplectic, in the sense that it is endowed with
a non-degenerate closed 2-form σ, we ask for σ to satisfy the homo-
geneity condition Mt∗ σ = tσ and call Y a conic symplectic manifold.
One can then define a homogeneous 1-form ω by putting
(2.3.3) ω(v) = σ(ρ, v) for any vector field v.
Note that (2.3.3) is equivalent to ω = H −1 (ρ). In other words, in a
symplectic manifold, a homogeneous structure is equivalently defined
by a radial vector field ρ or a canonical 1-form ω. There naturally arises
the question if, σ being closed, it is in fact exact and namely satisfies
σ = dω. Since this is true for T ∗ M , one expects that it is also true for
a general Y . The question is hard and its answer stated positively is
equivalent to saying that Y is symplectically diffeomorphic to T ∗ M .
Theorem 2.3.5 (Darboux-Frobenius theorem). Let Y be a symplec-
tic conic manifold and V a conic involutive submanifold of Y of codi-
mension l in a neighborhood of po = (xo , ξo ). We also assume that V is
regular, that is, ω|V 6= 0. Then we can find local homogeneous symplec-
tic coordinates (x, ξ) in a neighborhood of po , that is, a local symplectic
∼
homogeneous diffeomorphism Y → T ∗ Rn , such that xo = 0, ξo = eio
(the io -th coordinate vector) for io ≤ n − l, and
(2.3.4) V = {(x, ξ) : ξn−l+1 = 0, ..., ξn = 0}.
Cf. Theorem 2.3.7 for the proof.
Remark 2.3.6. (a) In particular, if we choose Y = V , the theorem
says that any symplectic homogeneous manifold Y is locally symplec-
tically diffeomorphic to T ∗ Rn (Darboux theorem).
94 2. REAL STRUCTURES

(b) Theorem 2.3.4 in combination with Theorem 2.2.9 asserts that

we can find real coordinates y ∈ Y ' R2n such that V = {y : y2n−l+1 =
0, ..., y2n = 0}. What Theorem 2.3.5 says additionally is that the trivi-
alization can be obtained in a symplectic system of coordinates.
(c) Associated to an involutive bundle such as T V ⊥ σ , there is
the foliation by its integral leaves {Γ}. In the present situation, this
is called the “bicharacteristic” or “Hamiltonian” foliation. The leaves
are “isotropic”, that is, satisfy the inclusion T Γ ⊂ T Γ⊥ σ . Under the
choice of coordinates (x, ξ) they are “straightened” into the planes
x1 , ..., xn−l = const, ξ1 , ..., ξn−l = const, ξn−l+1 , ..., ξn = 0.
When V is involutive of codimension l, then corank(σ|V ) ≡ l; for
a prescribed codimension l of V , this is the maximal corank. Theo-
rem 2.3.5 has a generalization to manifolds of constant σ-corank.
Theorem 2.3.7. Let V be a regular (in the sense that ω|V 6= 0),
conic submanifold of codimension l in a conic symplectic manifold Y
such that corank σ|V ≡ const in a neighborhood of po = (xo , ξo ); write
this constant as l − 2k. Then, there are local symplectic homogeneous
coordinates (x, ξ) in Y such that po = (0, e1 ) and
(2.3.5) V = {(x, ξ) : ξi = 0, i ≥ n−(l−k)+1, xj = 0, j ≥ n−k+1}.
Proof. It suffices to show that any partial system of coordinates
xj , j ≥ n − k + 1, and ξi , i ≥ n − (l − k) + 1, in a neighborhood of
(0, e1 ) which satisfy
(2.3.6) {ξi1 , ξi2 } = 0, {xj1 , xj2 } = 0, {ξi , xj } = κij
if i1 , i2 , i ≥ n − (l − k) + 1 and j1 , j2 , j ≥ n − k + 1,
can be supplemented to a full symplectic system (x, ξ) which satisfies
therefore the analogue of the Poisson relations (2.3.6) for all indices.
Suppose we have proved this assertion. Consider first the case k = 0:
V is regular involutive and is therefore defined in a neighborhood of
po = (0, e1 ) by, say, fi = 0, i ≥ n − l + 1, satisfying {fi1 , fi2 }|V ≡ 0.
By linear algebra (cf. the proof of Theorem 2.2.9) we see that we can
replace this set of equations by a new one which satisfies {fi1 , fi2 } ≡ 0
on the whole of Y instead of V . We then just have to rename fi = ξi
and supplement them to full coordinates.
Otherwise, if k ≥ 1, we know that some Poisson bracket, say
{fio , gjo }, does not vanish in a neighborhood of (0, e1 ). We rename
fio = ξn ; from ∂xn gjo (= {ξn , gjo }) 6= 0 and from the implicit function
theorem, we can replace gjo by x̃n = xn − h(x1 , ..., xn−1 , ξ1 , ..., ξn ). We
have thus achieved {ξn , x̃n } ≡ 1 and hence we complete these two co-
ordinates to a full system. In these new coordinates, V is in fact a
 2.3. REAL SYMPLECTIC SPACES—FROBENIUS-DARBOUX THEOREM 95

submanifold of T ∗ Rn−1 and hence Theorem 2.3.7 follows by the induc-

tive hypothesis.
We have to prove therefore the statement about supplementing sym-
plectic coordinates. To begin with, we state the linear version of the as-
sertion in the tangent plane Tpo Y : the vectors vj = ∂xj for j ≥ n−k +1,
and νi = ∂ξi for i ≥ n−(l −k)+1, can be supplemented to a symplectic
basis (v, ν). In fact, we first choose vi00 for n − (l − k) + 1 ≤ i ≤ n − k
such that
(
σ(νi , vj ) = κij , j ≥ n − (l − k) + 1,
σ(vi1 , vi2 ) = 0, i1 , i2 ≥ n − (l − k) + 1,
P
which is possible since σ is non-degenerate. If w = xj vj +
j≥n−(l−k)+1
ξj νj = 0, then σ(w, vj )(= −ξj ) = 0 and σ(w, νj )(= −xj ) = 0, j ≥
n − (l − k) + 1 and hence the extended system of vectors is still lin-
early independent. Let V1 = Span{vj , νj }j≥n−(l−k)+1 and V2 = V1⊥σ .
What we have to do is to find a symplectic basis for V2 . We reason by
induction on dim(V2 ) = 2m. If it is true for m − 1, we pick νio ∈ V2
and its “companion” vio related by σ(νio , vio ) = 1, which is possible on
account of the non-degeneracy of σ. We then set Vo = Span{vio , νio }
and decompose V2 = Vo ⊕ Vo⊥ . By applying the inductive hypothesis to
Vo⊥ , we conclude the discussion of the linear case.
Coming back to the general case, we can supplement the coordinates
xj , ξi to a full system by adding gj , j ≤ n − k, and fi , i ≤ n −
(l − k), such that Hgj = −∂ξj , and Hfi = ∂xi at po = (0, e1 ) where
(x, ξ) are symplectic coordinates in T ∗ Rn ' Tpo Y : this follows from the
discussion of the linear case. Let jo ≤ n − k and Y1 = {ξj = 0, j ≥
n − k + 1, and xi = 0, i ≥ n − (l − k) + 1}. We claim that there exists
gjo with dgjo = dxjo at 0 and which satisfies
(2.3.7)
(
Hgj (gjo ) = 0, j ≥ n − k + 1, Hfi (gjo ) = κijo , i ≥ n − (l − k) + 1,
gjo |Y1 = xjo .
In fact, from H{·,·} = [·, ·] and from (2.3.6), we know that all vector
fields Hgj , Hfi for i ≥ n − (l − k) + 1, j ≥ n − k + 1 commute. They
are also clearly transversal to Y1 . Thus, Corollary 2.2.12 applies and
gives the solution to (2.3.7). Also, if jo ≥ n − (l − k) + 1, then gjo |Y1 =
xjo |Y1 ≡ 0 which implies, in combination with the first line of (2.3.7),
that dgjo = dxjo . If, instead, jo ≤ n − (l − k), then
dgjo = d(gjo |Y1 )
= dxjo ,
96 2. REAL STRUCTURES

where the first and second equalities follow from the first and second
lines of (2.3.7), respectively.
We can do the same to find fio for io ≤ n − (l − k) which satisfies
the symplectic relations with the coordinates previously found and such
that dfio = dξio . The proof is complete.


Theorem 2.3.7 says that when rank σ|V ≡ const, then ker σ|V is
integrable. It also provides the explicit straightening of its integral
leaves: under a suitable choice of local symplectic coordinates, they
are reduced to the (l − 2k)-dimensional planes whose tangent bun-
dle is Span{∂xi }i=n−(l−k)+1,...,n−k , that is, the planes defined by xi =
const for i ≤ n − (l − k), xi = 0 for i ≥ n − k + 1, ξj = const for j ≤
n − (l − k), and ξj = 0 for j ≥ n − (l − k) + 1.

Suggested research. Let V1 and V2 be conic regular involutive

submanifolds of T ∗ Rn which have regular involutive intersection. Then
we can find symplectic homogeneous coordinates (x, ξ) such that V1 is
defined by x0 = 0 and V2 by ξ 00 = 0. The suggested research consists
in finding the most general conditions under which a pair of involutive
submanifolds can be put “simultaneously” into canonical form. This is
of interest, for instance, in the study of boundary value problems for a
system M of PDE’s over a domain Ω ⊂ Rn where V1 = T ∗ Rn |∂Ω and
V2 = char M, the characteristic variety of M.

2.4. Subelliptic estimates and hypoellipticity of systems of

vector fields
Let M be a smooth (C ∞ ) manifold, let L1 , ..., Lk be a system of
smooth vector fields of T M and set L1 = Span{L1 , ..., Lk }. Here L1
may happen to be closed under Lie bracket . Otherwise, we set L2 =
Span{Li , [Li1 , Li2 ] for i, i1 , i2 ∈ {1, ..., k}} and in general




(2.4.1) Lj = Span Li , [Li1 , Li2 ], ..., [Li1 , [Li2 , ...[Lij −1 , Lij ]...]]

 | {z }
j



for any i, i1 , ..., ij ∈ {1, ..., k} .


 2.4. SUBELLIPTIC ESTIMATES FOR SYSTEMS OF VECTOR FIELDS 97

Suppose that for a point xo ∈ M and for an integer m1 (xo ) ≥ 2 we

have
Ljxo = L1xo , j ≤ m1 − 1, Lm 1
xo ⊃ L xo .
1
6=
m
Lxo1
Let dim 1
L xo
=: l1 (xo ); in this situation we call m1 (xo ) the first
“Hörmander number” of the system L1 , ..., Lk at xo and l1 (xo ) its
“multiplicity”. We continue this process and find integers mh (xo ) and
lh (xo ) such that
Ljxo = Lm
xo
h−1
if mh−1 ≤ j ≤ mh − 1, Lm mh−1
xo ⊃ Lxo
h
6=

and
Lm
xo
h
lh (xo ) := dim
m .
Lxoh−1
This gives rise to a chain of spaces of iterated commutators which
becomes stationary after, say, r steps:
(2.4.2) L1xo ⊂ L2xo ⊂ ... ⊂ Lm
xo .
r

Notice that the entries of (2.4.2) are not necessarily bundles (constant
rank) but just distributions of tangent vectors.
Definition 2.4.1. The system L1 , ..., Lk is said to be of “finite
type” or to satisfy the “Hörmander condition” at xo when the chain
ends after finitely many, say r, steps with Lm
xo = Txo M .
r

In the above situation, we say that mr is the “type” of the system.

Otherwise, if Lm xo 6= Txo M , we say that the type is +∞.
r

Let F be the Fourier transform defined, over functions u in the

Schwartz R space S of rapidly decreasing C ∞ functions, by Fu(ξ) ≡
û(ξ) = Rn e−ihx,ξi u(x)dx. With the notation ∂x2 = nj=1 ∂x2j , let Λs =
P
s
(I − ∂x2 ) 2 be the standard elliptic pseudodifferential operator of (non
s
necessarily integer) order s defined by Λs u = F −1 (1 + |ξ|2 ) 2 û(ξ) . In


s
particular, if s > 0 is integer and even, then Λs u = (I − ∂x2 ) 2 u. We
introduce a scalar product by
hu, viH s = hΛs u, Λs viH 0
= h(1 + |ξ|2 )s û, v̂iH 0 ,
and define H s to be the completion of S under the associated norm.
We denote by h·, ·iH s and || · ||H s the scalar product and the norm
in H s , respectively. By the Plancherel theorem we have,
P forα integer s,
H s = {u : Dα u ∈ H 0 , |α| ≤ s} with ||u||2H s ' ||D u||2H 0 . Let
|α|≤s
 > 0.
98 2. REAL STRUCTURES

Definition 2.4.2. The system L1 , ..., Lk is said to be “-subelliptic”

at xo when, for a neighborhood B of xo ,
X
(2.4.3) ||u||2H  < ||Lj u||2H 0 + ||u||2H 0 for any u ∈ Cc∞ (B).
∼
j=1,...,k

Definition 2.4.3. The system L1 , ..., Lk is said to be “hypoelliptic”

at xo when for a neighborhood B of xo and for any x ∈ B,
(2.4.4) u ∈ H 0 , Lj u ∈ Cx∞ , j = 1, ..., k ⇒ u ∈ Cx∞ ,
where Cx∞ denotes the space of germs of C ∞ -functions at x.
The system is said to be subelliptic, resp. hypoelliptic, when (2.4.3),
resp. (2.4.4), is true for any xo . To prove that subellipticity implies
hypoellipticity we need a technical preliminary. Let χν ’s be a sequence
of smooth functions which approximate the Dirac measure.
Lemma 2.4.4. (Friedrichs’) Let g ∈ H s have compact support and
suppose Lg ∈ H s . Then,
(2.4.5) L(g ∗ χν ) → Lg in H s .
Proof. It suffices to prove that
(2.4.6) kL(g ∗ χν ) − (Lg) ∗ χν kH s < kgkH s .
∼

It is immediate to see that (2.4.6) implies (2.4.5). In fact, if we ap-

proximate gk →s g with gk ∈ Cc∞ , we have, by applying (2.4.6) to the
H
difference g − gk ,
kL(g ∗ χν ) − (Lg) ∗ χν ks ≤ kgk − gks + kL(gk ∗ χν ) − (Lgk ) ∗ χν ks ,
(I) (II)

with (I) → 0 for k → ∞ and (II) → 0 for ν → ∞. We prove now

(2.4.6) for g ∈ Cc∞ ; by the density of Cc∞ in H s , (2.4.6)
P follows in full
generality. We write L = a(x)∂x or, in detail, L = j aj (x)∂xj . We
start by proving (2.4.6) for s integer and denote by ∂xα a derivative of
order |α| ≤ s. We have
(2.4.7)
 
∂x L(g ∗ χν ) − L(g) ∗ χν = L(∂xα g ∗ χν ) − (L∂xα g) ∗ χν + [∂xα , L](g ∗ χν ) − [∂xα , L](g) ∗ χν
α

XZ 
= (aj (x) − aj (x − t))∂xα g(x − t)∂tj χν (t)
j

+ ∂tj aj (x − t)∂xα g(x − t)χν (t) dt1 ∧ ... ∧ dtn
+ [∂xα , L](g ∗ χν ) − [∂xα , L](g) ∗ χν .
 2.4. SUBELLIPTIC ESTIMATES FOR SYSTEMS OF VECTOR FIELDS 99

Denote by Ijαν each integral in the summation on the right of (2.4.7).

If M is a bound for ||a||C 1 , we have
Z   
|Ij | ≤ M |∂xα g(x − t)| χν (t) + |t| ∂tj χν (t) dt1 ∧ ... ∧ dtn .
αν

Young’s inequality for convolution applied to the above estimate, yields

 
||Ijαν ||0 ≤ M k∂xα gk0 kχν + |t| ∂tj χν kL1
(2.4.8)
≤ ck∂xα gk0 < ckgks ,
∼

where c is independent
R of ν. (This
 independence Rfollows from the fact
that kχν k = 1 and |t| ∂tj χν  dt1 ∧ ... ∧ dtn ≤ |∂tj χ|dt1 ∧ ... ∧ dtn
for any ν.) On the other hand, [∂xα , L] is of order s and hence the H 0 -
norm of the two terms on the last line of (2.4.7) is controlled by kgks
uniformly in ν.
By taking all choices of |α| ≤ s, we get (2.4.6) for s integer. As it
has already been said, this implies (2.4.5), that is, the continuity from
H s to H s of the mapping
(2.4.9) g 7→ L(g ∗ χν ) − L(g) ∗ χν .
Next, by interpolation lemma, the mapping (2.4.9) is continuous also
for s fractionary. This concludes the proof of the Lemma.

Theorem 2.4.5. If L1 , ..., Lk is subelliptic, then it is also hypoel-
liptic.
Proof. (i): Let ψ be a smooth cut-off function such that ψ ≡ 1
at x. We show that ψu ∈ H 0 and Lj (ψu) ∈ H 0 for any j implies
ψu ∈ H  . In fact, let us approximate ψu by uν := χν ∗ (ψu) ∈ Cc∞ ; by
Lemma 2.4.4 applied with s = 0 and g = χu, we have that uν → ψu
and Lj uν → Lj (ψu) in H 0 -norm. From
X
||uν − uµ ||2H  < ||Lj uν − Lj uµ ||2H 0 + ||uν − uµ ||2H 0 ,
∼
j

we deduce that {uν }ν is Cauchy in H  and therefore its limit ψu is also

in H  .
(ii): Let Lj (u) ∈ H s for any j over supp ψ and suppose that we
already know that u belongs to H σ over supp ψ for σ ≤ s: we wish
to prove that, in fact, ψu ∈ H σ+ . Let ψ 1 ∈ Cc∞ with ψ 1 ≡ 1 over
supp(ψ); be begin by proving that Lψ 1 Λσ ψu ∈ H 0 . In fact, we have
the elementary equality
Lψ 1 Λσ uν = Λσ Lψ 1 uν + [L, ψ 1 Λσ ]uν .
100 2. REAL STRUCTURES

Now, the first term in the right is H 0 convergent, for ν → ∞, to Λσ Lψu;

this follows from Lemma 2.4.4. The second is trivially H 0 convergent to
[L, ψ 1 Λσ ]ψu. Thus, the first is also H 0 convergent and, by H −1 unique-
ness, it must converge to Lψ 1 Λσ ψu; so this belongs to H 0 . Hence, we
apply (i) to ψ 1 Λσ (ψu) and conclude that ψu ∈ H σ+ . Iterated use of
this argument, with a gain of Sobolev index  and a shrinkig of supp(ψ)
at each step, makes it possible to conclude that ψu ∈ H s . (Note here
that the cut off in next step must have support where the former is
≡ 1.)


We notice now that if the system L1 , ..., Lr is not of finite type but
the Lie span L has locally constant rank a < n, then in new variables
L = Span{∂x0j }1≤j≤a ; this follows from the Frobenius theorem. Any
function u of x00j , j ≥ a + 1 only, is then a solution of the system
Lj u = 0 but it is not necessarily regular. Thus the sufficient condition
in next theorem is also essentially necessary.

1
Theorem 2.4.6. Let L1 , ..., Lk be of finite type m. Then it is 2m−1
subelliptic (and hence in particular hypoelliptic).

To prove the theorem, we need an auxiliary result which holds in

full generality (even for systems of non-finite type).

Lemma 2.4.7. We have

(2.4.10) ||[L1 , [L2 , [...Lm ]]]u||2H −1+

| {z }
m
X 1
< ||Lj u||2H 0 + ||u||2H 0 for  = .
∼
j
2m−1

Proof. We first consider the case of a simple commutator [L1 , L2 ].

s
Recall the pseudodifferential operator Λs with symbol (1 + |ξ|2 ) 2 . We
have

||[L1 , L2 ]u||2H −1+ = hΛ−1+ [L1 , L2 ]u, Λ−1+ [L1 , L2 ]uiH 0

= h[L1 , L2 ]u, Λ−2+2 [L1 , L2 ]uiH 0 .
 2.4. SUBELLIPTIC ESTIMATES FOR SYSTEMS OF VECTOR FIELDS 101

We write in the left side [L1 , L2 ] = L1 L2 − L2 L1 and consider only the

term L1 L2 since the other can be handled likewise. We have
(2.4.11)
|hL1 L2 u, Λ−2+2 [L1 , L2 ]uiH 0  = hL2 u, L1 Λ−2+2 [L1 , L2 ]uiH 0 
  

≤ |hL2 u, [L1 , Λ−2+2 [L1 , L2 ]]uiH 0 | + |hL2 u, Λ−2+2 [L1 , L2 ]L1 ui|
| {z } | {z }
order −1 + 2 order −1 + 2
(
||L1 u||2H −1+2 + ||L2 u||2H 0 + ||u||2H 0
< ||L2 u||2H −1+2 +
∼ ||L2 u||2H −1+2 + ||L1 u||2H 0 + ||u||2H 0 ,
where we can choose either of the two alternatives in the third line
of (2.4.11). When m = 2, we let  = 21 , choose either of the alterna-
tives and get (2.4.10). Otherwise, we reason by induction and suppose
(2.4.10) true for commutators of m − 1 vector fields. When we go on to
m, we have to use the second alternative in (2.4.11) with L2 replaced
by [L2 , [L3 , [..., Lm ]]]. Notice that, since we get a factor 2 for each com-
mutator, the choice  = 2−(m−1) is the one needed in the induction. We
then conclude
||[L1 , [L2 , [..., Lm ]]]u||2H −1+
< ||[L2 , [..., Lm ]]u||2H −1+2 + ||L1 u||2H 0 + ||u||2H 0
∼
X
< ||Lj u||2H 0 + ||L1 u||2H 0 + ||u||2H 0 .
∼
2≤j≤k


End of Proof of Theorem 2.4.6. We are assuming that L1 , ..., Lk
have finite type, say, m. Hence all derivatives of T M can be expressed
through commutators involving at most m vector fields Lj ’s. Hence,
(2.4.10) yields the subelliptic estimate for  = 2−(m−1) .

Remark 2.4.8. If L1 , ..., Lk has finite type m, then u “gains”  =
2−(m−1) derivatives with respect to (Lj u, j = 1, ..., k). A more efficient
proof shows that we have, in fact, the better gain of any  with  < m1
(cf. [75]). However, for application to hypoellipticity, this improvement
is irrelevant.
Remark 2.4.9. We spend a few words about the “complex” ver-
sion of Theorem 2.4.6 and Lemma 2.4.7. If a system of vector fields
Lj ∈ T 1,0 Cn ⊕ T 0,1 Cn is closed under conjugation then, identified to
a system in T R2n , it can enjoy the conclusions of Lemma 2.4.7 and
Theorem 2.4.6. Otherwise, Lemma 2.4.7 holds only if an  amount of
102 2. REAL STRUCTURES

kL̄j uk2H 0 norm is added on the right. On the other hand, on account
of the estimate kL̄j uk2H 0 < kLj uk2H 0 + kuk2 1 , the norm kuk2 1 is ab-
∼ H2 H2
sorbed in the left side provided that m = 2 (and taking summation
over all bracket of order ≤ 2). This explains why Theorem 2.4.6 stays
true in complex version when m = 2 and, instead, it fails for m > 2
(cf. Kohn [60]).

2.5. Miscellanea: foliations—orbits

We have seen that the involutivity of a bundle E over a manifold M
characterizes the existence of a local foliation by integral leaves Γ of E,
that is, manifolds which satisfy Tx Γ = Ex for any x ∈ Γ; in particular,
the various leaves have the same dimension, equal to the rank of E.
When E is not involutive, let L(E) denote the Lie algebra with respect
to the product [·, ·] generated by E. Two situations may occur:
• L(E) is a bundle: thus, we cannot apply the Frobenius theorem
to E but we can to L(E).
• L(E) is not a bundle: the Frobenius theorem fails. The failure
is substantial: there cannot be a system of integral leaves of
L(E) with the same dimension.
We can also enlarge our setting and, instead of a bundle E, deal with
Span{X}, the span of a set of vector fields X = {Lj }j=1,...,k whose rank
is not necessarily constant (the span meant over the ring of functions
on M of the same regularity as X).
Surprisingly, in case M and X are real analytic, there is in any case
the local integral leaf of L(X), the Lie algebra engendered by Span(X)
(having possibly different dimension leaf by leaf).
Theorem 2.5.1 (Nagano theorem). Let M and X be real analytic;
then for any xo there is a real analytic leaf Γ through xo such that
Tx Γ = Lx (X) for any x ∈ Γ in a neighborhood of xo .
Thus, the rank of L(X) may vary but it is forced by the real an-
alyticity to remain constant along each Γ. The manifold Γ is referred
to as the “Nagano leaf” of X (or L(X)). It is unique: we postpone the
proof to Remark 2.5.7(1) and (2).
Proof. Let n be the dimension of M and ko the rank of Lxo (X).
The cases ko = 0, ko = n, n = 1 are trivial. We assume by induction
that the theorem is true in dimension n − 1. Let X = {L1 , L2 , ...}; by
a change of coordinates x = (x1 , x0 ) in M , we may assume xo = 0 and
L1 = ∂x1 . Let Lj = aj1 ∂x1 + aj2 ∂x2 + ... with aji extending holomorphi-
cally to a neighborhood of xo . We can replace Lj by Lj −aj1 ∂x1 without
 2.5. MISCELLANEA: FOLIATIONS—ORBITS 103

changing L(X); thus it is not restrictive to assume Lj = Lj (x, ∂x0 ) for

∂x0 = (∂x2 , ∂x3 , ...). We write Lj = +∞ ν 0
P
ν=0 x1 Ljν (x , ∂x0 ), j = 2, 3, .... We
have
[L1 , [L1 , ...[L1 , Lj ]...]] = µ!Ljµ + x1 Rjµ (x, ∂x0 ).
| {z }
µ times

It follows that, if L is the Lie span of the Ljµ ’s, then L0 (x0 ) ⊂ L(0, x0 ).
0

Thus, let Γ0 be the integral leaf of L0 in the x0 -space Rn−1 which exists
by the inductive hypothesis. Then (−, +) × Γ0 is the integral leaf of
L in Rn .

When M or X is not C ω , we cannot expect the existence of leaves
such that T Γ = L(X)|Γ . Instead, we can look for manifolds Γ of minimal
dimension which satisfy
(2.5.1) Tx Γ ⊃ Xx for any x ∈ Γ.
As was already noticed, since T Γ is stable by linear combinations and
commutators, (2.5.1) implies
(2.5.2) Tx Γ ⊃ Lx (X) for any x ∈ Γ.
It turns out that, if M and X are C ∞ , there are no more Frobenius
or Nagano leaves but, nevertheless, there exist locally the manifolds of
minimal dimension which satisfy (2.5.2). To describe them, we need
new terminology.
Definition 2.5.2. A “piecewise smooth integral curve”, or “polyg-
onal integral path” of X in M , is the junction of finitely many integral
curves γ1 , ..., γs of smooth sections of X with each γj being issued from
the ending point of the former γj−1 .
Definition 2.5.3. We call the set of all ending points of piecewise
smooth integral curves of X issued from xo the “orbit” of X through
xo ∈ M and denote it by O(xo , X, M ). We call the set of the ending
points of the piecewise smooth integral curves contained in B the orbit
of X in a neighborhood B of xo ∈ M and denote it by O(xo , X, B).
Some examples are in order.
Example 2.5.4. Let X = Span{L1 , L2 } in R3 where L1 = ∂x1 and
L2 = ∂x2 + x1 x3 ∂x3 . Note that [L1 , L2 ] = x3 ∂x3 . It follows that
(2.5.3) rank X = 2,

3 for x3 6= 0,
(2.5.4) rank L(X) =
2 for x3 = 0.
104 2. REAL STRUCTURES

According to the Nagano theorem, we must have three orbits, the

Nagano leaves, of different dimensions. These are in fact
(2.5.5) O+ = {x ∈ R3 : x3 > 0},
(2.5.6) O0 = {x ∈ R3 : x3 = 0},
(2.5.7) O− = {x ∈ R3 : x3 < 0}.
When X is no longer C ω , the situation changes strikingly.
Example 2.5.5. For a function χ = χ(x1 ) in C ∞ such that
χ = 0 for x1 ≤ 0,
χ 6= 0 for x1 > 0,
we define in R2 the vector fields L1 = ∂x1 , L2 = χ(x1 )∂x2 and put
X = {L1 , L2 }; we have

1 for x1 ≤ 0,
rank L(X) = rank X =
2 for x1 > 0,
The Frobenius theorem fails (rank X being non-constant), the Nagano
theorem fails (X not being real analytic). In this situation, which are the
orbits? We choose as a system of neighborhoods B of xo the rectangles
centered at xo . If xo1 > 0, we readily have O(xo , X, B) = B. If xo1 < 0
and B does not intersect the half-plane x1 > 0, we easily see that
O(xo , X, B) reduces to a horizontal ray (−a, −b) × {xo2 }. If xo1 < 0 and,
instead, B intersects x1 > 0, then O(xo , X, B) = B. This requires a
few words of explanation: if we want to reach from xo a point ξ ∈ B
with ξ1 < 0 and ξ2 6= xo2 , we first move along the horizontal line till we
reach the half-plane x1 > 0, next we move on the vertical line till we
reach x2 = ξ2 , and finally we move backwards along the horizontal line
till we reach x1 = ξ1 . Of course, the global orbit through xo is always
the full plane R2 regardless of xo1 > 0 or xo1 < 0.
On one hand, the conclusion is unpleasant because Tx O =
6 Lx (X) for
x1 < 0. On the other hand, we have found that the orbits are manifolds.
These are far from being unique, and the dimension itself may vary with
B. However, if M and X are C ∞ , this ambiguity disappears in germs:
when shrinking B, the corresponding orbit stabilizes and is the germ
of a unique, well-defined, manifold.
Theorem 2.5.6 (Sussmann). For any xo ∈ M , there is a unique
smooth submanifold Γ ⊂ M in a neighborhood of xo , passing through
xo satisfying T Γ ⊃ X and such that
(i) if Γ0 is a manifold which satisfies T Γ0 ⊃ X, then for some
neighborhood B of xo we have B ∩ Γ0 ⊃ B ∩ Γ,
 2.5. MISCELLANEA: FOLIATIONS—ORBITS 105

(ii) for any neighborhood B of xo , there are an index jo and neigh-

borhoods B1 ⊂ B2 contained in B such that any point x ∈
Γ ∩ B1 can be reached by a polygonal path of jo integral curves
of X contained in Γ ∩ B2 .
Moreover, Γ is C ω if M and X are C ω .
We refer to [8] for a proof.
Remark 2.5.7. (1) The uniqueness of Γ is explicit from (i). We call
the manifold Γ of Theorem 2.5.6 the “local orbit” or “Sussmann orbit”
and sometimes denote it by ΓS or Oloc .
(2) If X is C ω , then the Nagano leaf ΓN coincides with the Sussmann
orbit ΓS . In fact, from (i) we know that B ∩ ΓS ⊂ B ∩ ΓN . On the other
hand,
L(E) ⊂ T ΓS ⊂ T ΓN = L(E).
Thus ΓS = ΓN . This identity shows the uniqueness of ΓN (and also
makes more explicit the independence of ΓS of the neighborhood where
it is defined).
We want to adapt (2.5.1) to general closed subsets Σ.
Definition 2.5.8. A closed subset is said to be stable under X if,
whenever it contains a point xo , it also contains the orbit O(xo , X, B)
for a suitable neighborhood B of xo .
This condition is the natural substitute of (2.5.1) to which it is
indeed equivalent when Σ is a manifold. In this case, it is also obviously
equivalent to Γ ⊃ O(xo , L(X), B). But, if X is C ∞ , this equivalence is
not specific to a manifold Γ.
Corollary 2.5.9. Let X be C ∞ and Σ be a closed subset of M
stable under X. Then Σ is stable under L(X).
Proof. Let ΓS be the Sussmann orbit: by hypothesis Σ contains
ΓS . Since T ΓS ⊃ X, then T ΓS ⊃ L(X) and thus Σ itself is stable under
L(X).


The fact is that, if X is C ∞ , the equivalence holds for orbits:

O(xo , X, B) = O(xo , L(X), B).
Corollary 2.5.10. Let X be C ∞ , assume Span(X) involutive of
constant rank k, and let Σ be closed and stable under X. Then for any
x ∈ Σ there is an integral leaf Γ of Span(X) through x contained in Σ.
106 2. REAL STRUCTURES

Proof. We know that Σ ⊃ ΓS , the Sussmann orbit through x.

We also know from Section 2.2 that there exists the integral leaf Γ of
Span(X) through x. We have T ΓS ⊃ X|ΓS ; by (i) of Theorem 2.5.6,
Γ = ΓS .

We now discuss some examples.
Example 2.5.11. For coordinates x = (x0 , x00 ) in Rn , we consider
the transformations
x̃0 = x̃0 (x),

(2.5.8) T1 =
x̃ = x̃00 (x00 ),
00
 0
x̃ = O(x0 ) + O(x0 )σ(x),
(2.5.9) T2 =
x̃00 = x̃00 (x00 ),
 0
x̃ = O(x0 ),
(2.5.10) T3 =
x̃00 = O(x00 ).
We denote by ∼1 , ∼2 , ∼3 the equivalence relations in the set C of co-
ordinates in Rn which identify those which are related by T1 , T2 and
T3 , respectively. We consider the foliation x00 = const, the projection
x 7→ x0 and the product Rnx01 × Rnx002 , or equivalently the pair of pro-
jections x 7→ x0 and x 7→ x00 , and denote by F, P and Π these three
structures with all possible choices of coordinates. It is clear that the
transformations which preserve the above three structures are those of
T1 , T2 and T3 , respectively. This yields
C
F ' ∼1

↑ p1 p1 ↑
C
P ' ∼2

↑ p2 p2 ↑
C
Π ' ∼3
.
We explain p1 and p−1 n
1 . A projection h : R → S to a submanifold
S induces a foliation defined by its fibers {h−1 (x)}x∈S ; a foliation and
a choice of a transversal manifold of dimension complementary to the
fibers give a projection which assigns to each x the point h(x) in which
the leaf through x meets S. The mapping p2 is clear by itself.
Example 2.5.12. We take two local foliations R = {R} and S =
{S} with leaves of dimension l + k and l, respectively. We claim that
 2.5. MISCELLANEA: FOLIATIONS—ORBITS 107

there is a local diffeomorphism F in Rn which transforms S in an l-

dimensional foliation {F (S)} which is “embedded” in R, that is, for
any S there is R such that F (S) ⊂ R. To see this, we choose an
n − (l + k)-foliation K = {K} transversal to R and S and a manifold
Σ transversal to S of dimension n − l which is a union of K-leaves. Let
ξS be the point of intersection S ∩ Σ = {ξS }, and let RξS be the R-leaf
through ξS . We consider the projection induced by K by the choice of
the transversal manifold RξS . This is defined by
pS : Rn → RξS , pS (x) := Kx ∩ RξS ,
where Kx is the K-leaf through x. Note that by restriction, pS induces
a diffeomorphism between S and pS (S), which we denote by FS . We
then have to define our diffeomorphism F by gluing the various FS .
Notes to Chapter 2. We refer to the book by Boggess [27] for
a general account about real spaces, real manifolds, vector fields and
the Frobenius theorem. For the presentation of the symplectic spaces,
we refer to the book by Hörmander [51] that we follow closely in the
proof of Theorem 2.3.7. Subelliptic estimates and hypoellipticity can
be found in many books, as for instance Folland and Kohn [42]. We
also refer to Rothschild and Stein [75] for the sharp estimate on the
gain of derivatives, that is, m1 instead of 2m−1
1
. Section 2.5 contains two
very relevant results in the integration of vector fields: Theorem 2.5.1
due to Nagano [70] and Theorem 2.5.6 due to Sussmann [84].
 CHAPTER 3

Real/Complex Structures

3.1. Complex structures—real underlying

structures—complexifications

Summary. We exhibit the identification of the real underlying

space R2n to Cn with the diagonal Cn ×Cn C̄n ; this is the same as re-
placing (x, y), the real and imaginary parts of z = x + iy, with (z, z̄).
The complexification of R2n in C2n coincides with the embedding of
Cn ×Cn C̄n in Cn × C̄n : coordinates are extended from (x, y) to (xC , y C )
or (z, z̄) to (z, w̄). We also discuss the intrinsic version, for manifolds,
of this operation. We then introduce complex structures and their gen-
eralization, the almost complex structures, and state the integrability
theorem for the latter when they are involutive.

For coordinates (x, y) in R2n , let z = x + iy be coordinates in Cn ;

here x = (x1 , ..., xn ), y = (y1 , ..., yn ), z = (z1 , ..., zn ). If z̄ = x − iy is
the conjugate to z, we identify
(3.1.1) R2n → Cn ×Cn C̄n , (x, y) 7→ (z = x + iy, z̄ = x − iy),
with inverse
z + z̄ z − z̄
(3.1.2) Cn ×Cn C̄n → R2n , (z, z̄) 7→ (
, ).
2 2i
Associated to (3.1.1) and (3.1.2), there is the transformation of differ-
entials
(dx, dy) 7→ (dz, dz̄) = (dx + idy, dx − idy).
We define (∂z , ∂z̄ ) by
∂x − i∂y ∂x + i∂y
(∂x , ∂y ) 7→ (∂z , ∂z̄ ) = ( , ).
2 2
Thus (∂z , ∂z̄ ) satisfies the chain rule related to (3.1.1); it is the dual
basis to (dz, dz̄). This yields an identification for tangent vectors, that
is, derivatives,
(3.1.3) T R2n ' T Cn ×T Cn T C̄n , a∂x + b∂y 7→ (a + ib)∂z + (a − ib)∂z̄
109
110 3. REAL/COMPLEX STRUCTURES

and the dual identification for cotangent vectors, that is, forms,
(3.1.4)
a − ib a + ib
T ∗ R2n → T ∗ Cn ×T ∗ Cn T ∗ C̄n , adx + bdy 7→ dz + dz̄.
2 2
Under (3.1.3) and (3.1.4), the duality between T R2n and T ∗ R2n in the
left hand side is interchanged with the real part of the duality between
T Cn and T ∗ Cn on the right hand side
ha∂x + b∂y , a0 dx + b0 dyi = Re h(a + ib)∂z , (a0 − ib0 )dzi,
that is,
aa0 + bb0 = Re ((a + ib) · (a0 − ib0 )) .
This identifies the euclidean structure of R2n with the real part of the
hermitian structure of Cn . After we have extracted the real structure
R2n underlying Cn , we want to complexify it, that is, to go on to C2n .
We first revisit the above construction in an intrinsic way and prefer to
reason on a complex manifold X , a union of local holomorphic patches
∼
{Xj } parametrized by means of homeomorphisms Uj → Xj over open
Φj
domains Uj ⊂ Cn which are holomorphically compatible over over-
lappings Ui ∩ Uj . We denote by X̄ the conjugate, the manifold with
local holomorphic parametrizations Φ̄j : Ūj → X̄j , and define the real
underlying manifold X R through the diagonal identification
∼
(3.1.5) X R → X ×X X̄ .
δX

Going on to tangent vectors, we get

∼
(3.1.6) TXR →
0
T X ×T X T X̄ ,
δX

and to cotangent vectors,

∼
(3.1.7) T ∗X R →
0−1
T ∗ X ×T ∗ X T ∗ X̄ .
tδ
X

Thus, on the right side of (3.1.6), (3.1.7), we continue to find diagonals,

that is, real underlying manifolds: T X ×T X T X̄ = (T X )R and T ∗ X ×T ∗ X
T ∗ X̄ = (T ∗ X )R . We thus have evidence that
∼
(3.1.8) δX0 = δT X : T X R → (T X )R ,
t 0−1 ∼
(3.1.9) δT ∗ X = δT ∗ X : T ∗ X R → (T ∗ X )R .
We now complexify X R to (X R )C by the embedding j : X ×X X̄ ,→
X × X̄ . When complexifying in fiber the bundles T X R and T ∗ X R , we
have, instead, two options: either to complexify the coefficients of vector
 3.1. REAL UNDERLYING STRUCTURE—COMPLEXIFICATION 111

fields and forms or to use the mapping j 0 and t j 0−1 . The two ways are
related by the diagonal correspondences
δ 0C
(3.1.10) C ⊗R T X R → T X ×X T X̄
and
t δ 0−1C
(3.1.11) C ⊗R T ∗ X R → T ∗ X ×X T ∗ X̄ .
If we wish to express (3.1.10) and (3.1.11) in coordinates, we have just
to use (3.1.3), (3.1.4) and there replace real coefficients a and b by
complex ones. We note that the “diagonal” relations a − ib = a + ib
correspond precisely to the choices of a and b real.
Definition 3.1.1. A complex structure on X is the morphism J
induced on T X R by the multiplication by i on T X :
i
TX → TX
|| ||
J
TX R
→ T X R.
Note that we have the identity J 2 = −id as the real counterpart
of i2 = −1. Also, in T X R , the complex structure is given in local
holomorphic coordinates z = x + iy of X by J ∂x = ∂y , J ∂y = −∂x .
With the aid of J we can write (3.1.10) and (3.1.11) as
(3.1.12)
ξ + i t J −1 ξ ξ − i t J −1 ξ
  
X − iJ X X + iJ X
X 7→ , , ξ 7→ , ,
2 2 2 2
with inverses
(3.1.13) (Z, W̄ ) 7→ Z + W̄ , (ζ, µ̄) 7→ ζ + µ̄.
When the mappings in (3.1.13) are restricted to the diagonal, they
take values in T X R and T ∗ X R , respectively. We denote by “Re 00 the
composition of (3.1.13) restricted to the diagonal with (π1 |T X ×T X T X̄ )−1
and (π1 |T ∗ X ×T X T ∗ X̄ )−1 , respectively:
(3.1.14)
Re : T X → T X R , Z 7→ Z + Z̄, Re : T ∗ X → T ∗ X R , ξ 7→ ξ + ξ. ¯
We remark that in C ⊗R T X R there are two “distinguished” bundles
T 1,0 X = { X−iJ
2
X
: X ∈ T X R } and T 0,1 X = { X+iJ
2
X
: X ∈ T X R },
the bundles of holomorphic and antiholomorphic vector fields or of type
(1, 0) and (0, 1), respectively. They satisfy
(3.1.15) T 1,0 X ∩ T 0,1 X = {0}, T 1,0 X ⊕ T 0,1 X = (T X R )C .
They can be characterized as the eigenspaces of the eigenvalues +i and
−i for J C , respectively. In coordinates, they are described as T 1,0 X =
112 3. REAL/COMPLEX STRUCTURES

Span{∂zj }j=1,...,n and T 0,1 X = Span{∂z̄j }j=1,...,n , respectively. What we

have just seen to be true for T X R continues to hold for a general real
vector space V . We start by extending Definition 3.1.1.
Definition 3.1.2. A morphism of complex structure J on V is a
morphism which satisfies J 2 = −id.
We have the analogues of the bundles T 1,0 X and T 0,1 X . In fact,
given a complex structure J , there remain well-defined eigenspaces of
the eigenvalues +i and −i of J C that we denote L and L̄. They satisfy
(3.1.16) L ∩ L̄ = {0} and L ⊕ L̄ = V C .
Conversely,
Theorem 3.1.3. Let L be a subspace of an even-dimensional vector
space V , let L̄ be the conjugate and assume that (3.1.16) is satisfied.
Then, there is a unique morphism J : V → V such that L and L̄ are
the eigenspaces of the J C -eigenvalues +i and −i, respectively.
Proof. Define
(3.1.17) J C L = iL for L ∈ L, J C L̄ = −iL̄ for L̄ ∈ L̄.
It is clear that J C is an endomorphism of V C and that L and L̄ are
the eigenspaces of +i and −i. What remains to prove is that J maps
V into V . Define
(
X := L + L̄,
(3.1.18)
Y := i(L − L̄).
We have J C X = Y and J C Y = −X; in particular, J C fixes V .

We write after (3.1.18)
X − iY X − iJ X X + iY X + iJ X
(3.1.19) L = = , L̄ = = ,
2 2 2 2
which agrees well with (3.1.12). We need to make our abstract setting
richer and call into play the involutivity properties of systems of vector
fields. So, we leave the abstract vector space V and specify our discus-
sion for V = T M where M is a real manifold, not necessarily embedded
in Cn nor in a complex manifold X . We start by observing that T 1,0 X
and T 0,1 X are involutive, that is, closed under Lie bracket. For general
bundles L and L̄ this is an additional requirement.
Definition 3.1.4. (i) An “almost complex structure” over a man-
ifold M is a bundle L ⊂ C ⊗ T M such that
L ∩ L̄ = {0}, L ⊕ L̄ = C ⊗ T M.
 3.1. REAL UNDERLYING STRUCTURE—COMPLEXIFICATION 113

(ii) An almost complex structure is said to be “involutive” when

the bundle L is involutive.
(iii) An involutive almost complex structure L is a “complex struc-
ture” if it is realized as L = T 1,0 X for a complex manifold X .
The following shows that (ii) implies (iii).
Theorem 3.1.5 (Newlander-Nirenberg). A smooth, involutive, al-
most complex structure is locally a complex structure.
Proof (Kohn [58]). We prove the existence of local holomorphic
coordinates (u1 , ..., un ) (for n := dimM
2
), that is, complex functions
which satisfy
L̄uj = 0 for any L̄ ∈ L̄ and du1 ∧ ... ∧ dun 6= 0.
Associated to the splitting of vector fields L⊕L̄ and to the dual splitting
¯
of forms, there is a well-defined ∂-operator, the “(0, 1)-component of
d”. Now, the condition of integrability of the structure is equivalent to
∂¯2 = 0.
This is a classical and instructive exercise. We take now n complex
functions w = (w1 , ..., wn ) vanishing at the reference point zo = 0 and
such that (Re w, Im w) are a system of real local coordinates. In these
coordinates, the operator ∂¯ is expressed, for suitable coefficients, by
!
X X j ∂u X j ∂u
¯ =
∂u ak + ãk dwk
k j
∂w j j
∂ w̄ j
(3.1.20) !
X X j ∂u X j ∂u
+ bk + b̃k dw̄k .
k j
∂w j j
∂ w̄j

We can assume that at zo = 0 the structure coincides with the standard

structure, that is, ajk (0) = ãjk (0) = bjk (0) = 0 and b̃jk (0) = κjk , the
Kronecker symbol. By means of the dilation  : z 7→ z, we may define
a structure on the unit ball Bn1 that we denote by ∂¯ ; its coefficients
are obtained by applying ∗ to the coefficients of (3.1.20). We claim
that for suitable  we can find uj , j = 1, ..., n, complex such that
du1 ∧ ... ∧ dun 6= 0 and ∂¯ uj = 0. We start by remarking that ∂ ∂¯ |z|2
¯ 2 and therefore it stays positive definite. We
is a perturbation of ∂ ∂|z|
¯
then go back to the L2 estimates (1.9.25) for the ∂-Neumann problem.
¯
They hold unchanged and provide a ∂-Neumann operator N . (This is
not always the case for a general structure if it is not endowed with
114 3. REAL/COMPLEX STRUCTURES

a strictly plurisubharmonic function such as |z|2 .) This allows us to

“correct” the zj ’s, which are not necessarily ∂¯ -holomorphic, to
(3.1.21) uj := zj − ∂¯∗ N ∂¯ zj .
Moreover, the elliptic regularity in the interior for  = ∂¯ ∂¯∗ + ∂¯∗ ∂¯
yields
(3.1.22) ||χN ∂¯ zj ||2H k+2 ≤ ck ||∂¯ zj ||2H k for χ ∈ Cc∞ (Bn1 ).
Because of (3.1.22) the uj ’s stay independent.

Alternative proof in a special case. For the reader who does
¯
not feel comfortable with the ∂-Neumann problem, we give the choice of
a simplified proof in case L is already known to be the bundle L = T 1,0 Y
of (1, 0) vector fields tangent to a real submanifold Y ⊂ X , that is,
T 1,0 Y = (C ⊗R T Y) ∩ T 1,0 X . Note that in this case L is involutive and
satisfies well L ∩ L̄ = {0}. The argument is still local. Let us define
Re
∼
T C Y := T Y ∩ iT Y and identify T 1,0 Y → T C Y. Hence our hypothesis is
T C Y = T Y,
and our task is to show that
if T Y is complex, then Y is complex.
Since T Y and its orthogonal T Y ⊥ are complex, that is, J -invariant, of
complex dimension, say, n − l and l, respectively, then we can choose
for them a pair of real basis in the form
(3.1.23) T Y = Span{vh0 , J vh0 }, T Y ⊥ = Span{vk00 , J vk00 }.
Keeping in mind that for coordinates (x, y) in R2n we have J ∂x = ∂y ,
we define an isomorphism of tangent vectors at zo = 0
∼
Φ : T0 Y ⊕ T0 Y ⊥ → T0 R2n , v 7→ ∂x , J v 7→ ∂y .
Under this identification, we get ΦT0 Y = Cn−l z 0 ×{0}. It follows that Y is
locally in a neighborhood of zo = 0 the graph of a function z 0 7→ h(z 0 ),
that is, Y = {(z 0 , h(z 0 )) : z 0 ∈ Cn−l }. Now, our aim is to show that
h is holomorphic. If we define Ψ := id × h and take any vector field
L ∈ (T Cn−l )R , we get
J (Ψ∗ L) = J (L, h∗ L) = (J L, J h∗ L).
We have
(L, h∗ L)|z=0 ∈ T0 Y.
 3.2. CR MANIFOLDS 115

Since T0 Y is J -invariant, then (J L, J h∗ L)|z=0 ∈ T0 ΨY and hence

there is a vector field W such that

(J L, J h∗ L) = (W, h∗ W ),

which implies W = J L and hence also

h∗ J L = J h∗ L.

Notes. For a general background on complex structures, holo-

morphic and antiholomorphic vector fields, real underlying structures,
complexifications and almost complex structures, the reader may con-
sult the books by Baouendi, Ebenfelt and Rothschild [8], Boggess [27]
and Chern [34]. Theorem 3.1.5 was proved by Newlander and Nirenberg
¯
[72]. The proof that we present here is an application of the ∂-Neumann
problem and is due to Kohn [58]. Another proof, which uses integral
kernel methods, is due to Webster [97].

3.2. CR manifolds

Summary. We present here embedded CR structures and give

the normal form of a generic CR submanifold of Cn . We also discuss
some celebrated classes of manifolds which fail to be CR and still have
a nice classification. In these cases we are in the presence of “isolated
complex tangencies”.

Linear models for CR structures are not geometrically significant.

We introduce them just to warm up with terminology and notation.
Let E be a complex vector space of dimension n. We can think of E
as a real vector space E R of dimension 2n endowed with a morphism
J : E → E which is an antiinvolution, that is, satisfies J 2 = −id.

Remark 3.2.1. We can identify C-spaces to R-spaces endowed with

such a J . Given an R-space, we can realize it as a C-space by defining
iv := J v for any v ∈ E and we can go the other way around by
defining J v := iv.

Let F ⊂ E be a real subspace. We have

F ∩ J F ⊂ F ⊂ F + J F,
116 3. REAL/COMPLEX STRUCTURES

where the term on the left (resp. the right) is the maximal C-space
contained in F (resp. the minimal C-space containing F ). We set
dimCR (F ) = dimC (F ∩ J F ).
It is clear that
(3.2.1) dimR F = dimC (F ∩ J F ) + dimC (F + J F ),
which is obtained by dividing by 2 the two sides of
(3.2.2) dimR F + dimR J F = dimR (F ∩ J F ) + dimR (F + J F ).
On account of (3.2.1) we have
dimR F
(3.2.3) dimR F − n ≤ dimCR F ≤ .
2
We set l = codRCn F, s = codCCn (F + JF ); one can check that if rkR and
rkC are the real and complex rank of a system of (linear) equations for
F , respectively, then l = rkR and s = l − rkC , which yields
dimCR F = n − l + s.
There are two extremal cases:
(
s = 2l , dimCR F = n − 2l = dimF
2
,
(3.2.4)
s = 0, dimCR F = n − l.
The first corresponds to F being complex. The second means that
F + J F = Cn and F is said to be “generic” in that case. In the sequel,
we refer to the number s as the “defect of genericity” of F . Note that
when F is generic, then dimR F ≥ n.
Lemma 3.2.2. For a pair of linear subspaces F, G ⊂ E the following
are equivalent:
(i) dimR F = dimR G and dimCR F = dimCR G.
∼
(ii) There is an isomorphism ϕ : F → G which is C-complex, that
is, ϕ ◦ J = J ◦ ϕ.
Proof. It is not restrictive to assume
(
F + J F = G + J G = E,
(3.2.5)
F ∩ J F = G ∩ J G.
We choose a basis e1 , . . . , en−l , J e1 , . . . , J en−l of F ∩ J F = G ∩ J G
and two completions f1 , . . . , fl and g1 , . . . gl for F and G, respectively.
Thus, we can explicitly define our morphism ϕ by
ϕ(ej ) = ej , ϕ(J ej ) = J ej , ϕ(fj ) = gj .
Such a ϕ fulfills all requirements of the statement.

 3.2. CR MANIFOLDS 117

Definition 3.2.3. (i) We say that F is totally real when F ∩J F =

{0}, that is, dimCR F = 0.
(ii) We say that F is totally real maximal when Cn = F ⊕ J F .

Note that if F is totally real (resp. totally real maximal), then

dimF ≤ n (resp. dimF = n). However, it is not just a question of
dimension. For instance, in C2 , we have
• R × R totally real maximal,
• R × {0} totally real,
• C × R generic,
• C × {0} complex non-generic.
We now discuss embedded CR structures. Let M be a real submanifold
of Cn . We set
TzC M = Tz M ∩ J Tz M, z ∈ M,
and call it the maximal complex tangent space at z.

Definition 3.2.4. The manifold M is said to be a CR manifold

when rk(TzC M ) ≡ const, that is, when T C M is a bundle.

We refer to the rank of T C M as the CR dimension of M and de-

note it by dimCR M . We define the CR codimension by codimCR M :=
dimM − 2dimCR M . There are two distinguished situations:

(3.2.6) TzCo M = {0}, that is, M is totally real at zo ,

(3.2.7) Tzo M + J Tzo M = Tzo Cn , that is, M is generic at zo .

We note that (3.2.6) and (3.2.7) are stable: if they hold at zo , they
hold also in a neighborhood of zo . This is clear for (3.2.6), which is a
condition of transversality T M ∩ J T M = {0}. As for (3.2.7), we have
to remark that this is the condition of C-independence of a system of
equations of M : ∂r1 ∧ · · · ∧ ∂rl 6= 0. Another distinguished situation is
given by

TzC M = Tz M for any z close to zo ,

(3.2.8)
that is, M is a complex manifold at zo .

(Cf. here the Newlander-Nirenberg theorem.) If s is the defect of gener-

icity of M , recall from the discussion of the linear case that codCR M =
2n − l − 2(n − l + s) = l − 2s. In particular,

“M generic” is equivalent to “codCR M = l and dimCR M = n − l”.

118 3. REAL/COMPLEX STRUCTURES

For a CR manifold M , we define the bundles of vector fields tangential

to M of type (1, 0) and (0, 1), respectively, by
T 1,0 M = T 1,0 Cn ∩ (C ⊗ T M ),
T 0,1 M = T 0,1 Cn ∩ (C ⊗ T M ).
The second is also called the bundle of CR vector fields. We note that
we have two identifications
X − iJ X
(3.2.9) j1 : T C M → T 1,0 M, X → ,
2
X + iJ X
(3.2.10) j2 : T C M → T 0,1 M, X → ,
2
with inverse
(3.2.11) L → Re L = L + L̄.
We wish to discuss these identifications in coordinates z = x + iy of
Cn and under a choice of C-independent equations for M , say r1 =
0, ..., rl = 0, such that ∂r1 ∧ · · · ∧ ∂rl 6= 0. We also write r for (rj )j ; we
have
¯ = 0}
T M ' {(L, L̄) ∈ T 1,0 Cn ×T 1,0 Cn T 0,1 Cn : hL, ∂ri + hL̄, ∂ri
' {a∂x + b∂y ∈ (T Cn )R : a∂x r + b∂y r = 0 for a, b real},
T C M ' {(L, L̄) ∈ T 1,0 Cn ×T 1,0 Cn T 0,1 Cn : hL, ∂ri = 0}
' {a∂x + b∂y ∈ (T Cn )R : a∂x r + b∂y r = 0 and − b∂x r + a∂y r = 0
for a, b real}.
It follows that
¯ = 0}
C ⊗ T M ' {(L1 , L̄2 ) ∈ T 1,0 Cn ×Cn T 0,1 Cn : hL1 , ∂ri + hL̄2 , ∂ri
' {α∂x + β∂y ∈ C ⊗ (T Cn )R : α∂x r + β∂y r = 0 for α, β complex},
C ⊗ T CM
¯ = 0}
' {(L1 , L̄2 ) ∈ T 1,0 Cn ×Cn T 0,1 Cn : hL1 , ∂ri = 0, hL̄2 , ∂ri
' {α∂x + β∂y ∈ C ⊗ (T Cn )R : α∂x r + β∂y r = 0, −β∂x r + α∂y r = 0
for α, β complex}.
There is a geometric set which deeply represents the CR system.
Definition 3.2.5. We define the characteristic set of the system
of CR vector fields tangential to M by
char(T 0,1 M ) := T 0,1 M ⊥ ∩ T ∗ M
 ∗
TM
= ,
T CM
 3.2. CR MANIFOLDS 119

with the orthogonal “⊥ ” taken in C ⊗ T ∗ M .

It is readily seen that the natural projection T ∗ Cn |M  T ∗ M com-
bined with J ∗ induces an injection
 ∗
∗ n J
∗ TM
TM C → .
T CM
This is an isomorphism precisely when M is generic; thus char(T 0,1 M )
∗
' TM Cn in this case. Let N ⊂ M ; when N is also generic, we have the
transversality condition
TN∗ M ∩ char(T 0,1 M ) = {0}.
This condition is crucial for taking the restriction or trace to N of a
solution of the CR system over M .
We have a preliminary description of generic submanifolds which
only uses the basics of linear algebra. Let M ⊂ Cn be a submanifold of
codimension l.
Proposition 3.2.6. The following are equivalent, in a neighbor-
hood of a point zo ∈ M :
(i) M is generic.
(ii) dimCR M is minimal, that is, equals n − l.
(iii) There is a system of equations for M , r1 = 0, ..., rl = 0, such
that ∂r1 , ..., ∂rl are C-independent.
(iv) dz1 |M , ..., dzn |M are independent.
(v) There are local coordinates z = x + iy, z = (z 0 , z 00 ), z 0 = z1 , ..., zl ,
such that zo = 0 and M is graphed over x0 , z 00 by
yj0 = hj (x0 , z 00 ) with hj (0) = 0, ∂hj (0) = 0.
(vi) There is N ⊂ M totally real maximal.
Proof. The equivalence of (i), (ii) and (iii) has already been seen.
That (i) and (iv) are equivalent follows from the fact that
dz1 |M , ..., dzn |M are C-independent
if and only if
dz1 |T M +J T M , ..., dzn |T M +J T M are independent
if and only if
T M + J T M = Cn .
That (v) implies (iii) is obvious. Conversely, modulo linear complex
transformations, we may arrange that zo = 0 and
Span{dr1 (zo ), ..., drl (zo )} = Span{dy10 , ..., dyl0 }.
120 3. REAL/COMPLEX STRUCTURES

Hence, by the implicit function theorem, we can find hj with hj (0) =

0 and ∂hj (0) = 0, such that M is defined by yj0 = hj , j = 1, ..., l.
Finally, for the equivalence of (v) and (vi), it suffices to define N by
supplementing the equations yj0 = hj by the new equations yj00 = 0. The
proof is complete.

We now want to discuss the normal form of M analogous to the
one stated in Section 1.6 about the “2-jet”.
Theorem 3.2.7. Let M ⊂ Cn be generic of codimension l and class
k
C and let zo be a point in M . By a complex holomorphic change of
coordinates we can assume that M is graphed over x0 , z 00 at zo = 0 by
yj0 = hj (x0 , z 00 ), j = 1, ..., l,
such that
|I|+|J| |I|+|J|
(3.2.12) ∂xI z00J h(zo ) = ∂xI z̄00J h(zo ) = 0 if |I| + |J| ≤ k.
Proof. By a linear change of coordinates, we first arrange that
zo = 0 and h(0) = 0, ∂h(0) = 0. For the Taylor expansion of h = (hj )
up to order k, we have
k
X
h= aI J K x0I z 00J z̄ 00K + ok ,
|I|+|J|+|K|≥2

where a = aI J K is an l-vector a = (aj ). We “complexify” from x0 ∈ Rl

to z 0 ∈ Cl and from (z 00 , z̄ 00 ) ∈ Cn−l ×Cn−l C̄n−l to (z 00 , w̄00 ) ∈ Cn−l × C̄n−l
and consider the polynomial map
k
X
0 00 00
k
h : (z , z , w̄ ) 7→ aIJK z 0I z 00J w̄00K .
|I|+|J|+|K|≥2

By the implicit function theorem, there is a unique map Φ = Φ(z 0 , z 00 )

such that
(3.2.13) z 0 = Φ(z 0 + ihk (z 0 , z 00 , 0), z 00 ).
We define a polynomial change of coordinates in a neighborhood of 0
by
(
z̃ 0 = z 0 − ihk (Φ(z 0 , z 00 ), z 00 , 0),
(3.2.14)
z̃ 00 = z 00 .
We notice that if z ∈ M , that is, y 0 = h(x0 , z 00 , z̄ 00 ), then
ỹ 0 = y 0 − Re hk (Φ(z 0 , z 00 ), z 00 , 0)
(3.2.15)
= Re h(x0 , z 00 , z̄ 00 ) − hk (Φ (x0 + ih(x0 , z 00 , z̄ 00 ), z 00 ) , z 00 , 0) .


 3.2. CR MANIFOLDS 121

To obtain an equation for the image M̃ of M under the change (3.2.14),

we must replace (x0 , z 00 ) by (x0 (z̃ 0 , z̃ 00 ), z̃ 00 ), which yields
  
00
(3.2.16) ỹ 0 = Re h x0 (z̃ 0 , z̃ 00 ), z̃ 00 , z̃
   
k 0 0 00 0 0 00 00 00 00 00
−h Φ x (z̃ , z̃ ) + ih(x (z̃ , z̃ ), z̃ , z̃ ), z̃ , z̃ , 0 .

We write (3.2.16) as ỹ 0 = Re h̃. By the implicit function theorem, we

can remove ỹ 0 from h̃; thus Re h̃ has become a graphing function for
M̃ . It remains to show that h̃ satisfies the requirement (3.2.12) in the
statement of the theorem. Now, consider the function
(3.2.17) (x̃0 , z̃ 00 ) 7→ x0 (x̃0 , z̃ 00 , z̃ 00 ).
By taking the k-jet, x0k , of (3.2.17) and complexifying the variables
00 00 00
from (x̃0 , z̃ 00 , z̃ ) to (z̃ 0 , z̃ 00 , w̃ ), we get for x0k = x0k (z̃ 0 , z̃ 00 , w̃ )
 
00 00
(3.2.18) h̃(z̃ 0 , z̃ 00 , w̃ ) = hk x0k , z̃ 00 , w̃
     
0k 0k 00 00 00 00
− h Φ x + ih x , z̃ , w̃ , z̃ , z̃ , 0 + ok .
k k

To establish (3.2.12), we have to show that

00
h̃(z̃ 0 , z̃ 00 , 0) = ok , h̃(z̃ 0 , 0, w̃ ) = ok .
We prove the first, the second following by conjugation. On account of
(3.2.13), we have, for x0 and z 00 replaced by x0k = x0k (z̃ 0 , z̃ 00 , 0) and z̃ 00 ,
respectively,
x0k = Φ x0k + ihk (x0k , z̃ 00 , 0), z̃ 00 + ok .


(3.2.19)
Substitution of (3.2.19) into (3.2.18) and evaluation of hk and Φ for
00
w̃ = 0 yields
(3.2.20)
h̃ = hk (x0k , z̃ 00 , 0) − hk Φ x0k + ihk (x0k , z̃ 00 , 0), z̃ 00 , z̃ 00 , 0 + ok

= hk (x0k , z̃ 00 , 0) − hk x0k + ok , z̃ 00 , 0 + ok

= ok .

Remark 3.2.8. There is a C ω version of the above statement. In
that case the equations yj0 = hj can be found such that (3.2.12) holds
for any I, J, K and not just for those whose sum of lengths is ≤ k.
122 3. REAL/COMPLEX STRUCTURES

CR singularities—isolated complex tangencies. We exhibit now

some examples of manifolds which are not CR. In some fortunate case,
we still have a nice classification, a normal form for the equation, and
an efficient theory of attached analytic discs similar to the one that will
be developed in Section 3.7 for CR manifolds.
Let M be a 2-codimensional real submanifold of C2 with Tzo M = C.
We first arrange, by a linear change, that zo = 0 and Tzo M = {z : z2 =
0}. Thus M is defined by z2 = h(z1 ) for h(0) = 0 and ∂h(0) = 0. The
points (z1 , h(z1 )) where M has a complex tangency correspond to the
zeroes z1 of
 
−∂z1 h 1
det = ∂z1 h̄
(3.2.21) −∂z1 h̄ 0
= 0.

The only chance for M to be CR is that ∂z1 h̄ ≡ 0, or equivalently, that

M is a complex manifold in accordance with Theorem 3.1.5. Other-
wise, M is not CR. If other assumptions are not made, M is hard to
handle. To have a nicer situation, we first ask for zo to be an isolated
singularity. But the condition which is really crucial to develop some
serious analysis is that zo be a non-degenerate singularity in the sense
that

∂z1 ∂z̄1 h(zo ) 6= 0.

Thus h = a|z1 |2 + bz12 + cz̄12 + O3 for a 6= 0. By a quadratic change

of variables we may arrange that a = 1, b = c and, by a rotation in
the z1 -plane, we may reduce the common value of b and c to be real
positive, say γ ≥ 0. Hence the equation takes the form

(3.2.22) z2 = |z1 |2 + γ(z12 + z̄12 ) + O3 .

We call q = q(z1 ) the quadric in (3.2.22). We have that the sublevel

sets {q ≤ c} are compact or not according to γ < 21 or γ ≥ 21 . The
complex tangent is said to be “elliptic”, “parabolic” and “hyperbolic”
when γ < 12 , γ = 21 and γ > 12 , respectively; if γ 6= 12 , the singularity is
isolated. We are mainly concerned with the elliptic case.
We pass now from C2 to Cn and consider a manifold which is gener-
ically totally real maximal and has a complex tangency at zo = 0 that
we suppose to be non-degenerate. By combining (3.2.22) with the nor-
mal form of Theorem 3.2.7, we get the canonical form for a system of
 3.2. CR MANIFOLDS 123

equations


 z2 = |z1 |2 + γ(z12 + z̄12 ) + O3 (x3 , ..., xn , z1 ),
y = O3 (x , ..., x , z ),

3 3 n 1
(3.2.23)

 ...,

yn = O3 (x3 , ..., xn , z1 ).


Note that the xj , j = 3, ..., enter into the error term of the first equa-
tion because by a holomorphic transformation of type z̃1 = z1 + zj , j =
3, ..., terms like zj z̄1 are absorbed by |z̃1 |2 . Also, |z1 |2 enters into the
errors of the remaining equations because, by the first,
|z1 |2 = z2 − 2Re (γz12 ) + O3
= Re (z2 − 2γz12 ) + O3 .
In (3.2.23), the first error term O3 is complex and the others are real.
We write the first line as z2 = h and assume γ < 21 . We note that
the set of the complex tangencies forms an (n − 2)-dimensional real
submanifold. In fact, they are defined as the 0-set of the determinant
of the system
(3.2.24)
   
∂z (z2 − h) −(z̄1 + 2γz1 ) 1 ∗ ∗ · ∗
 ∂z (z̄2 − h̄)  −(z̄1 + 2γ z̄1 ) 0 ∗ ∗ · ∗ 
   i

 ∂z (y3 − O3 )  =  ∗ 0 −2 ∗ · ∗ 
    + error.
 ...   ... ... ... ... ... ... 
∂z (yn − O )3
∗ 0 ∗ ∗ · − 2i
This determinant is
i
(− )n−2 (z̄1 + 2γz1 ) + error,
2
and therefore it has only one zero near z1 = 0.
We pass to the general case and suppose that M is generic of CR
dimension m apart from some CR singularities where the CR dimension
grows to m + 1. We divide the variables as (z1 , z2 , z 0 , z 00 ); by (3.2.23)
and Theorem 3.2.7, we can put the equations of M in the form
(
z2 = q(z1 , z 00 ) + O3 ,
y 0 = Q(z1 , z 00 ) + O3 .
We assume that zo = 0 is a non-degenerate isolated CR singularity in
the (z1 , z2 )-plane. By the discussion of the C2 case, we have q(z1 , 0) =
|z1 |2 + γ(z12 + z̄12 ). We point out the following.
• The points of higher-dimensional complex tangencies, that is,
the points where dimCR M = m + 1, do not necessarily form
124 3. REAL/COMPLEX STRUCTURES

a manifold Σ and, if this is a manifold, it does not necessarily

satisfy dimΣ = dimM − 2. Note that the Jacobian matrix
whose dropping rank gives the extra complex tangency is no
longer squared.
• At a higher-dimensional complex tangency, the quadric in the
(z1 , z2 )-plane stays non-degenerate, by continuity.
Example 3.2.9. In C4 , let M be defined by
(
z2 = |z1 |2 + O3 (x3 , z1 , z4 ),
y3 = Q(z1 , z4 ) + O3 (x3 , z1 , z4 ),
where Q is a quadric. Clearly 0 is an elliptic CR singularity with γ = 0.
The set of singularities is defined by two complex equations in the plane
of the parameters (x3 , z1 , z4 ) whose rank may vary or not according to
how the O3 terms are chosen. When the rank is constant, a regular set
parametrized over R or R × C is defined.
Remark 3.2.10. When M is generically totally real maximal
(which is always the case in C2 ) and when the singularities are elliptic,
the method of Bishop [24] permits us to attach an (n − 1)-parameter
family of discs. In the space of parameters these are defined by
|z1 | ≤ r, (x3 , ..., xn ) = const. Their union gives rise to an (n + 1)-
dimensional manifold M̃ with boundary M foliated by discs. It is
quite easy to prove, with the aid of the implicit function theorem, that
this manifold is smooth outside the manifold which carries the CR
singularities. But indeed, Kenig and Webster prove in [54] that M̃ is
smooth at any point without exceptions.

Suggested research. (1) Let M be generic apart from some

non-degenerate elliptic CR singularities which form a manifold Σ of
maximal rank, that is, dimΣ = dimM − 2. Is it possible to construct
a manifold M̃ , with dimM̃ = dimM + 1, foliated by discs attached to
M by a similar procedure as in [54] where M is generically totally real
maximal? Is this manifold M̃ smooth at any point without exceptions?
(2) We say that a CR submanifold N of a CR manifold M is “relatively
generic” when
(3.2.25) codimM N = dimCR M − dimCR N.
It is clear that if M is already generic, N is relatively generic in M if
and only if it is generic in Cn . As the terminology itself suggests, the
failure of (3.2.25) is quite exceptional. The most pathological case, the
occurrence of a submanifold N ⊂ M such that dimCR N = dimCR M ,
 3.3. CR FUNCTIONS AND CR MAPPINGS 125

is referred to as “non-minimality” of M and is the main argument of

Section 3.10. When M is a hypersurface and N is a submanifold such
that dimCR N = dimCR M , then N is necessarily a complex hypersur-
face. In this situation, if M is defined by r = 0, then, for some s which
satisfies s|N ≡ 0, we have

(3.2.26) ∂r|N ∧ i∂s|N = 0.

By differentiating (3.2.26) along N , we get ∂∂r|¯ T C N = i∂∂s|

¯ T C N and
¯ being hermitian, it is 0. In particular, if the Levi form LM
hence, ∂∂r
is not 0 at a point zo , a complex hypersurface cannot pass through zo .
For the same reason, if there are no isotropic vectors for LM , there are
no complex curves. In particular, this occurs when LM > 0, that is, M
is strictly pseudoconvex. Otherwise, if M is only weakly pseudoconvex,
that is, LM ≥ 0, the problem becomes much more involved and higher-
order tensors than the Levi form might enter into play. The occurrences
of complex submanifolds are all instances of failure of genericity: com-
plex is non-generic.
The suggestion for research is to get, in wide generality, a deeper un-
derstanding of the obstructions to the existence of submanifolds which
are not relatively generic. By the richness of the situations that may
occur, this leads in fact to a galaxy of problems and exercises.

Notes. We refer to the books by Berhanu, Cordaro and Hounie

[21], Baouendi, Ebenfelt and Rothschild [8] and Boggess [27] for the
presentation of CR structures. For the discussion on the normal form
of a submanifold, the main references are Baouendi, Jacobowitz and
Treves [9], Chern and Moser [35], Jacobowitz [52] and Treves [88].
Finally, we refer again to Chern and Moser [35] for the discussion on
complex tangencies.

3.3. CR functions and CR mappings

Summary. We present CR functions and CR mappings. We

prove that any CR real analytic submanifold is embedded in a “par-
tial” or “intrinsic” complexification, a complex manifold in which it is
generic. In particular, CR real analytic functions holomorphically ex-
tend, with uniqueness, to this complexification. This is a consequence
of the Cauchy-Kowalevsky theorem. Also, we prove that any CR mani-
fold, regardless of whether it is real analytic or not, is CR isomorphic to
a generic submanifold of a complex manifold obtained via projection.
126 3. REAL/COMPLEX STRUCTURES

We then present abstract CR structures and prove that, in the C ω set-

ting, they are embeddable, that is, CR diffeomorphic to submanifolds
of Cn .

Definition 3.3.1. A C 1 function f : M → C is said to be CR

when
(3.3.1) L̄f = 0 for any L̄ ∈ L̄.
In case M is a submanifold of Cn , defined by r = 0 for r =
(rj )j=1,...,l , then (3.3.1) means that L̄f = 0 for any L such that Lr = 0.
Let us discuss this case in further detail.
Lemma 3.3.2. Let M ⊂ Cn be a CR manifold defined by r = 0
for r = (rj )j=1,...,l . A C 1 function f : M → C is CR if and only
if for any extension of f to Cn , which we still denote by f , we have
¯ ∧ ∂r
∂f ¯ 1 ∧ ... ∧ ∂r
¯ l = 0.

Proof. Let ∂¯M f be the component of ∂f ¯ |M which is orthogonal

¯
t¡o the ideal generated by {∂rj }j=1,...,l . One readily sees that f is CR
precisely when ∂¯M f = 0, and the lemma follows.


A large class of CR functions on M ⊂ Cn is given by the restrictions

to M of functions f holomorphic in a neighborhood of M : since ∂z̄j f =
0 for any j, then L̄(f |M ) = 0 for any L̄ ∈ T 0,1 M . In particular, for
M = Cn , holomorphic and CR functions coincide. We will see that if
M and f are C ω , then the converse is true. But, in general, CR functions
are a larger family than just restrictions of holomorphic functions. For
example, if M = R×C, then any (C 1 ) function f (x1 , z2 ) which satisfies
∂z̄2 f = 0 is CR. But the holomorphic extension requires f to be C ω . We
want to start our discussion about extension with the basic uniqueness
criterion.
Theorem 3.3.3. Let N and M be CR generic submanifolds with
N ⊂ M ⊂ Cn and let f be CR on M . Then
f |N ≡ 0 implies f |N = 0∞ ,
that is, f vanishes at infinite order along N . In particular, for M = Cn ,
holomorphic functions which vanish on N generic are identically 0.
Proof. Let {Lj }j=1,...,m , {L̄j }j=1,...,l be a basis for T 1,0 N ⊕ T 0,1 N ,
let {Xj }j=1,...,a be a basis for TTCNN and let {Nj = J Xj }j=1,...,b with b ≤ a
be a basis for T M |TNC∩J
N
TN
. The vector fields Lj , L̄j and Xj are tangential
 3.3. CR FUNCTIONS AND CR MAPPINGS 127

to N and the Nj ’s normal to N in M , respectively. We observe that by

our hypothesis
L̄j f = 0 and Nj f = iXj f for any j.
We wish to prove that Dα f |N = 0 for any differential operator tangen-
tial to M —combination of the Lj , L̄j , Xj , Nj . We proceed by double
induction on the degree p in the Nj ’s and the total degree q of Dα and
suppose that the conclusion holds for any Dα with degrees p and q. If
we compose Dα with a tangential vector field, then we are taking the
derivative of the identity to 0 along N which is still 0. Thus, what we
need to prove is that Nj Dα f |N = 0. Now,
Nj Dα f = Dα Nj f + [Nj , Dα ]f
= Dα iXj f + [Nj , Dα ]f.
The first term still contains p normal derivatives and hence it is 0. The
second still has order q and so is 0 likewise.

Let M be a real smooth manifold of dimension d.
Definition 3.3.4. An abstract CR structure on M is a bundle
L ⊂ C ⊗ T M such that
(i) L ∩ L̄ = {0},
(ii) L is involutive.
It is clear that if M ⊂ Cn is a CR manifold, then L = T 1,0 M is a
CR structure.
Definition 3.3.5. Let (M, LM ) and (N, LN ) be a pair of CR struc-
tures and f : M → N a C 1 map. Then, f is said to be CR if it satisfies
(3.3.2) f∗ LM ⊂ LN .
In particular, if f : M → Cn , then f is CR if and only if each
component fj , for j = 1, ..., n, is a CR function. It is more natural to
reason directly with T M instead of C ⊗ T M . Thus, let us define
T C M = {L + L̄ : L ∈ L}.
If we have an embedded structure L = T 1,0 M for M ⊂ Cn , then
T C M = T C M (= T M ∩ J T M ).
Theorem 3.3.6. Let f : M → N be a C 1 map between CR struc-
tures (M, LM ) and (N, LN ). Then, f is CR if and only if
(
f∗ T C M ⊂ T C N,
(3.3.3)
f∗ ◦ JM = JN ◦ f∗ .
128 3. REAL/COMPLEX STRUCTURES

In other words, f is CR if it preserves the maximal tangent complex

space and f∗ is complex linear on it.
Proof. We begin by proving that f∗ LM ⊂ LN implies (3.3.3).
First, if L + L̄ ∈ T C M , then f∗ (L + L̄) = f∗ L + f∗ L ∈ T C N . Moreover,
since LM and L̄M are the +i and −i eigenspaces for JM , then
f∗ JM (L + L̄) = f∗ (iL − iL̄)
(3.3.4) = i(f∗ L − f∗ L)
= JN (f∗ L + f∗ L) = JN f∗ (L + L̄).
Conversely, let us prove that (3.3.3) implies f∗ LM ⊂ LN . Remember
that any vector field L ∈ LM can be written as L = X−iJ2 M X for
X := L + L̄. It follows that
f∗ X − if∗ (JM X)
f∗ L =
(3.3.5) 2
f∗ X − iJN f∗ X
= .
2
The vector in the second line is certainly a vector of LN , which proves
our claim.

Let M be a CR submanifold of Cn . We want to discuss the existence
of an intermediate complex manifold Y , with M ⊂ Y ⊂ X such that
M is generic in Y . A positive answer comes first.
Theorem 3.3.7. Let M ⊂ Cn be CR and C ω . Then, there is a
complex submanifold Y ⊂ X which contains M and satisfies T Y |M =
T M + J T M.
∼
Proof. Let Φ : R2n−l → M be a C ω parametrization of M and
∼
ΦC : C2n−l → M C its complexification. Let δ : M ,→ Cn ×C̄n C̄n be
the diagonal embedding, δ C : M C ,→ Cn × C̄n its complexification,
p1 : Cn × C̄n → Cn the first projection, and let j be the embedding
p1 ◦ δ|M and k the mapping k = p1 ◦ δ C |M C . This yields a commuting
diagram
Φ
∼ δ
R2n−l → M ,→ Cn ×C̄n C̄n
j& ↓ p1
(3.3.6) ↓ ↓ Cn
k% ↑ p1
∼
C2n−l → M C ,→ C × C̄n .
n
ΦC δC
 3.3. CR FUNCTIONS AND CR MAPPINGS 129

We define
Y := k(M C ).
We wish to prove that
T Y |M = k 0 (T M C )|M = T M + iT M ;
in particular, T Y having constant (complex) rank, Y is a (complex)
manifold and fulfills the requirements in the statement. In fact, we
have
(3.3.7) δ C0 T M C = {α∂z + β̄∂z̄ : α∂z r + β̄∂z̄ r = 0},

(3.3.8) k 0 T M C = {α∂z : for some β̄, α∂z r + β̄∂z̄ r = 0},

(3.3.9) T M + iT M = {(σ + ρ)∂z :

σ∂z r + σ̄∂z̄ r = 0 and ρ∂z r − ρ̄∂z̄ r = 0}.
We denote by (I) the equation for α, β in (3.3.8) and by (II) the system
for σ, ρ in (3.3.9). We observe that the correspondence
α = σ + ρ, β = σ − ρ,
with inverse
α+β α−β
σ= , ρ= ,
2 2
interchanges the solutions of (I) with those of (II). This implies T Y |M =
T M + iT M , which proves our claim.

Guided Exercise 3.3.8. Let M be CR and C ω , parametrized by
Φ
∼
Rd → M . Sometimes, we use the notation ∂¯b to denote the system
induced by the vector fields T 0,1 M in the space of parameters, that is,
∂¯b := (Φ−1 )∗ T 0,1 M . We want to describe ∂¯b explicitly in an example.
We choose M to be the strongly pseudoconvex hypersurface in C2 de-
fined by −y2 + |z1 |2 = 0 that we also rewrite as r = 0. In this situation,
T 0,1 M is engendred by the single vector field
L̄ = ∂z̄2 r∂z̄1 − ∂z̄1 r∂z̄2
i
= − ∂z̄1 − z1 ∂z̄2 .
2
On the other hand, M is parametrized over R3 by Φ : (x1 , y1 , x2 ) 7→
(x1 + iy1 , x2 + i|z1 |2 ). This induces a transformation of vector fields

∂ Φ̄1 ∂Φ2 ∂ Φ̄2
∂z̄1 → ∂ z̄1 ∂z̄1 + ∂ z̄1 ∂z2 + ∂ z̄1 ∂z̄2

= ∂z̄1 + iz1 (∂z2 − ∂z̄2 ),

∂ → ∂ + ∂ .
x2 z2 z̄2
130 3. REAL/COMPLEX STRUCTURES

We then have
i z1
∂¯b = − ∂z̄1 − ∂x2 ,
2 2
as we can check from
i z1 z1
Φ∗ (∂¯b ) = − ∂z̄1 + (∂z2 − ∂z̄2 ) − (∂z2 + ∂z̄2 )
2 2 2
= L̄.
We refer to the manifold Y of the preceding theorem as an “intrinsic
complexification” of M . It can be characterized by
Theorem 3.3.9. Let M be C ω and Y be a complex manifold con-
taining M and such that T Y |M = T M + iT M . Then, any f which is
CR and C ω , extends uniquely as holomorphic to Y in a neighborhood
of M .
Proof. Let f C : M C → C be the complexification of f , defined,
e.g., through a parametrization Φ of M and its complexification ΦC .
Our hypothesis that f is CR ensures that L̄f C = 0 for any L̄ ∈ T 0,1 Cn ∩
T M C . Thus, f C is constant on the fibers of k defined by (3.3.6). Hence,
there is a well-defined holomorphic function f˜ on Y = kM C given by
f˜(z) = f C (z̃) for z̃ ∈ M C with k(z̃) = z.

In general, if M is not C ω , we cannot expect that there is a complex
submanifold Y ⊂ Cn in which M is embedded as generic. However,
this is true if we allow a projection instead of an immersion. Before
discussing this, we recall that we have already treated the inversion
problem for holomorphic mappings: the inverse is holomorphic, if it ex-
ists. The same is not true for CR mappings. For instance, the mapping
R × R → C, (x, y) → z = x + iy is one-to-one and CR but its inverse
Re × Im is not CR. However, it is an instructive exercise to verify that
if dimCR M = dimCR M 0 , then any f : M → M 0 which is CR and is a
C 1 -diffeomeorphism is in fact a CR diffeomorphism, that is, f −1 is also
CR. With this preliminary we can state
Theorem 3.3.10. Let M ⊂ Cn be CR. Then, there is a complex
manifold Y and a complex projection j : Cn  Y such that
(
j is a CR diffeomorphism between M and M 0 := jM ,
T Y |M 0 = T M 0 + iT M 0 .
Proof. We choose Y as the complex plane Tzo M + iTzo M , identi-
fied in a choice of coordinates to a plane of Cn , and choose a transversal
 3.3. CR FUNCTIONS AND CR MAPPINGS 131

projection j : Cn  Y . We notice that j is one-to-one over M since

the fibers of the projection are transversal to M . Also, j is clearly a
CR mapping between M and its image since it is the restriction of a
complex projection. We have to show that its inverse is still CR: but
this is just a consequence of the remark prior to the statement.

Remark 3.3.11. The manifold Y , produced by the proof of Theo-
rem 3.3.10, is not a CR hull of M as it was under the assumption of
C ω regularity. However, such a kind of hull exists, indeed, though it
has a very strange appearance. According to the theory by Tumanov
that we will introduce in Section 3.10, if M is “minimal”, then there
is a manifold of wedge type W of dimension n − s, with edge M , and
which is contained in the polynomial hull of M . (Here s is the “defect
of genericity”.) In particular, CR functions f on M extend uniquely
as holomorphic in W and their boundary values on M coincide with
f . Thus, the unique holomorphic extension to manifolds of dimension
n − s generically “takes” place, minimality being a “generic” condition,
provided that we restrict the extension to a “wedge type” domain in
an (n − s)-dimensional complex manifold.
We have introduced abstract CR structures and almost complex
structures. In one way the latter are more restrictive because they must
satisfy L ⊕ L̄ = C ⊗ T M . On the other hand, they are not necessarily
assumed to be involutive. If this is the case, they are an example of
CR structures and, in fact, they are complex structures according to
the Newlander-Nirenberg theorem. There naturally arises the question
whether a general CR structure L is “embeddable”, that is, if it can
be realized as L = T 1,0 M for a submanifold M ⊂ Cn . We will see that
the answer, in general, is no, even if the question is put in an intrinsic
way, that is, by allowing “CR diffeomorphisms” of M as it is clarified
in what follows. We refer to the number dimC L as the dimension of
M
the CR structure and to dimC C⊗T L⊕L̄
as its CR codimension. There is a
complex structure morphism J : T M → T M such that the eigenspaces
of the extension of J to L ⊕ L̄ are L and L̄ for the eigenvalues +i and
−i, respectively (cf. Section 3.1). Abstract, C ω and CR, structures are
embeddable.
Theorem 3.3.12. Let (M, L) be a C ω , CR structure. Then, there is
an embedding M ,→ Cn which identifies L with T 1,0 M . More precisely,
∼
there is a CR diffeomorphism Φ : M → M 0 ⊂ Cn such that
Φ∗ L = T 1,0 M 0 .
132 3. REAL/COMPLEX STRUCTURES

Proof. We start from the case of a trivial CR structure L = {0}.

We denote by d the dimension of M : we want to find an embedding of
M into a complex manifold X of dimension d. In fact, if U is open in
∼
Rd and Φ : U → M is a local C ω chart, we take a complexification ΦC
which gives a holomorphic parametrization of a d-dimensional manifold
∼
X : ΦC : U C → X . Here, U C is open in Cd . We have the commuting
diagram
j
U ,→ U C
Φ↓ ↓ ΦC
M X.
The desired embedding is then obtained by the composition
ΦC ◦ j ◦ Φ−1 : M ,→ X .
Our purpose is to generalize this elementary construction. We use
the notation m = dimCR M and l = codCR M . In coordinates x =
(x1 , ..., x2m+l ) ∈ R2n+l , we write
2m+l
X
L̄j = aji (x)∂xi ,
i=1

where each aji is C ω and C-valued. Since rkC {L̄j }j=1,...,m is m, we can
assume, by a linear change of coordinates, that the m × m minor A :=
(aji )j,i≤m is non-degenerate. We decompose the coordinates as
x = (x0 , x00 ), x0 = (x1 , ..., xm ), x00 = (xm+1 , ..., x2m+l ).
A linear change of coordinates x̃0 = A−1 x0 has the effect of multiplying
the coefficients of the ∂x0 ’s in L̄j by A−1 , thus giving
2m+l
X
L̄j = ∂xj + ãji ∂x00i .
i=m+1

It follows that
[L̄j , L̄i ] ∈ Span{∂x00i }i=m+1,...,2m+l .
On the other hand, we have
[L̄j , L̄i ] ∈ Span{L̄j }j=1,...,m ,
by the assumption of involutivity. Since the two above “Span”’s are
transversal one to another, we conclude that
[L̄j , L̄i ] = 0 for all i, j,
that is, we have found a basis of vector fields for L̄ which commute.
We now complexify R2m+l , with variable x, to C2m+l , with variable z,
 3.3. CR FUNCTIONS AND CR MAPPINGS 133

and extend holomorphically the coefficients from aji (x) to aji (z). We
define
X
L̄Cj = ∂zj + aji (z)∂zi .
i

We recall that [L̄Cj , L̄Ci ]|R2m+l ≡ 0 and hence, by analytic identity, we

conclude that
[L̄Cj , L̄Ci ] ≡ 0.
Recall the Frobenius theorem and observe that there is a “complex
analytic” version of it for vector fields with holomorphic coefficients (cf.
[27, pp. 56–58]). This means that there are holomorphic coordinates
z̃ = z̃(z), z̃ = (z̃ 0 , z̃ 00 ) in which the vector fields L̄Cj are interchanged
with ∂z̃0j , j ≤ m. More precisely, we have
(
L̄Cj z̃i00 = 0, j ≤ m, i ≥ m + 1,
z̃i00 (0, z 00 ) = zi .
∼
Let us denote by Φ : U → M the parametrization of M and denote by
Ψ the mapping z 7→ z̃ 00 (z); we define

M̃ := Ψ(Φ−1 (M )).

There is no difficulty in proving that M̃ is still a real manifold (cf.

[27, pp. 171–172]). What is crucial is to observe that, indeed, the CR
structure is preserved, that is, z̃ 00 (z) is a CR diffeomorphism along M .
This follows from
L̄j (z̃ 00 |M ) = (L̄Cj z̃ 00 )|M = 0.


We point out that the embedding fails for C ∞ , instead of C ω ,

structures. Nirenberg has given an example (cf. [27] and the refer-
ences therein) of a three-dimensional CR structure which cannot be
embedded in C2 .

Notes. We refer again to the books by Baouendi, Ebenfelt and

Rothschild [8] and Boggess [27] for the presentation of CR structures
and CR mappings. As for the notion of partial or intrinsic complexifi-
cation, we refer, among others, to Rossi [76], Tomassini [85], Andreotti
and Hill [6]. As for Theorem 3.3.12 (embedding of C ω , CR structures)
we refer to the book by Boggess [27] for an alternative proof.
134 3. REAL/COMPLEX STRUCTURES

3.4. The Levi form of a submanifold M ⊂ Cn and an

abstract CR structure

Summary. We present the so-called “intrinsic” Levi form of a

real submanifold M ⊂ Cn which is defined by means of commutators
of tangential holomorphic and antiholomorphic vector fields. We show
that it is “tensorial”, that is, independent of the vector field extension,
and it measures the failure of involutivity of the bundle T C M . We also
present the “extrinsic” Levi form, defined by contraction with a system
of differentials of equations of M .

We first consider the Levi form through the so-called “intrinsic ap-
proach” which applies to any abstract CR structure (M, L) as well. By
definition, L and L̄ are involutive, that is, [L1 , L2 ] ∈ L and [L̄1 , L̄2 ] ∈ L̄
for any pair of vector fields L1 , L2 ∈ L. Also, [L1 , L̄2 ] belongs to C⊗T M
but it does not necessarily belong to L ⊕ L̄. The Levi form precisely
measures how much L ⊕ L̄ fails to be involutive. Let

C ⊗ TM
p : C ⊗ TM →
L ⊕ L̄

be the projection. Let zo be a point of M .

Definition 3.4.1. The intrinsic Levi form at zo is defined over a

tangent vector u ∈ Tz1,0
o
M as

1 
LM (zo )(u) := p[L, L̄]zo ,
i

where L is any section of T 1,0 M which extends u from zo , that is,

L(zo ) = u.

The factor 1i is introduced in order to make the form real.

Sometimes, we drop the symbol p. The Levi form is well defined,
that is, independent of the choice of the L-vector field extension L such
that L(zo ) = u ∈ Tz1,0
o
M.

Lemma 3.4.2. If N is another vector field extension of u, that is,

N (zo ) = L(zo ) = u, then

p[L, L̄]|zo = p[N, N̄ ]|zo .

 3.4. LEVI FORM OF EMBEDDED AND ABSTRACT STRUCTURES 135
P
Proof. We write in a basis {L1 , ..., Lm } of L, L = j aj Lj and
P
N = j bj Lj . We know that aj (zo ) = bj (zo ). We have
X
p[L, L̄]|zo = aj (zo )āi (zo )p[Lj , L̄i ]|zo
ji
X
= bj (zo )b̄i (zo )p[Lj , L̄i ]|zo
ji

= p[N, N̄ ]|zo .

Proposition 3.4.3. L ⊕ L̄ is involutive if and only if LM ≡ 0.
Proof. The “only if” part is obvious. “If”: we have to show that if
[L, L̄] ∈ L ⊕ L̄ for any L ∈ L, then [L1 , L̄2 ] ∈ L ⊕ L̄ for any L1 , L2 ∈ L.
Now, we have
[L1 + L2 , L1 + L2 ] = [L1 , L̄1 ] + [L2 , L̄2 ]
(3.4.1)  
+ [L1 , L̄2 ] − [L1 , L̄2 ] .
We also have
[L1 + iL2 , L1 + iL2 ] = [L1 , L̄1 ] + [L2 , L̄2 ]
(3.4.2)
− i([L1 , L̄2 ] + [L1 , L̄2 ]).
By (3.4.1) and (3.4.2), LM ≡ 0 implies [L1 , L̄2 ] − [L1 , L̄2 ] ∈ L ⊕ L̄
and [L1 , L̄2 ] + [L1 , L̄2 ] ∈ L ⊕ L̄, respectively, and hence in conclusion
[L1 , L̄2 ] ∈ L ⊕ L̄.

We wish to discuss in further detail LM when M is an embedded
CR generic manifold M ⊂ Cn . In this case L = T 1,0 M, L̄ = T 0,1 M are
M
both involutive bundles and C⊗T L⊕L̄
= C ⊗ TTCMM is the complexification
of the totally real part of the tangent bundle. The Levi form is the map
Tzo M
LM (zo ) : Tz1,0
o
M → TzCo M
,
1
u 7→ i
p[L, L̄]|zo ,
for L(zo ) = u. We want to revisit Proposition 3.4.3 in this situation.
What must be said is that T 1,0 M and T 0,1 M are involutive whereas
T C M is not, in general. Though these bundles can all be identified
∼ ∼
by the isomorphisms Re : T 1,0 M → T C M and Re : T 0,1 M → T C M ,
nevertheless Re [·, ·] 6= [Re ·, Re ·]. In detail, let M be defined by r1 =
0, ..., rl = 0 with ∂r1 ∧...∧∂rl 6= 0 and write r = t (r1 , ..., rl ). Take a pair
of vector fields L1 , L2 in T 1,0 M ; these are characterized by Lj (ri ) = 0
136 3. REAL/COMPLEX STRUCTURES

for any j = 1, 2 and i = 1, ..., l. Clearly this condition is stable under

[·, ·], that is,P[L1 , L2 ](ri ) = 0Pbut it is not preserved by “Re ”. In fact,
write L1 = j aj ∂zj , L2 = j bj ∂zj , put
X
cj := āh ∂z̄h bj − b̄h ∂z̄h aj ,
h
P
and define R := j cj ∂zj . We have
1
[Re L1 , Re L2 ] = [L1 + L̄1 , L2 + L̄2 ]
4
1 
= [L1 , L2 ] + [L1 , L2 ] + R + R̄ .
4
What we have to prove is that R(r) = 0: this characterizes the invo-
lutivity of T C M . It is evident that Re R(r) = 0, which is equivalent to
saying that T M is involutive. However,
X X
R(r) = − ∂zi ∂z̄j rāj bi + ∂zi ∂z̄j rb̄j ai
ji ji
X
= −2iIm ∂zi ∂z̄j rāj bi .
ji

Also, we have R(r) = 0 for any such R if and only if (∂zi ∂z̄j r)i j |T C M ≡ 0,
that is, if and only if M is “Levi flat”. On the other hand, the condition
R(r) = 0 for any R characterizes the involutivity of T C M . We have
thus obtained the proof of the following statement which rephrases
Proposition 3.4.3.
Theorem 3.4.4. T C M is involutive if and only if M is Levi flat.
Indeed, what we proved is more than we stated. Whether M is
Levi flat or not, the vector field R + R̄ obtained by commutation of
Re L1 ∈ ker LM and Re L2 ∈ ker LM still belongs to ker LM .
Theorem 3.4.5. ker LM is involutive.
Thus, Theorem 3.4.4 is a corollary of Theorem 3.4.5 applied for
ker LM = T C M . For an embedded submanifold M ⊂ Cn , we want to
develop further the relation between LM and the Levi form that we
already introduced in Section 1.6 through a system of equations of M .
We note that, in case M is generic, the natural projection combined
with J , i.e.,
TM
→ TM Cn , [u] mod T C M 7→ [J u] mod T M,
T CM
 3.4. LEVI FORM OF EMBEDDED AND ABSTRACT STRUCTURES 137

induces an isomorphism that we still denote by J . By the choice of

∗
a basis {∂rj }j=1,...,l of TM Cn and the dual basis of TM Cn , we get an
∗
identification K : TM Cn → TM Cn . In composition with J this yields

T M J∼ K ∗
→ TM Cn → TM Cn .
T CM
∗
(Here, as always, ∂rj are identified to sections of TM Cn by the aid of the
∼
correspondence 2Re ∂rj → drj .) We define, in analogy to Section 1.6
where the hypersurface case was treated, the “extrinsic Levi” form LM
by
X
(3.4.3) LM (u, ū) = ¯ j (u, ū)∂rj .
∂ ∂r
j

¯ j (ū, u) are real and so (3.4.3) is identified

Note that the coefficients ∂ ∂r
∗
with a vector field of M × Rl ' TM Cn .
Theorem 3.4.6. We have
(3.4.4) LM (u, ū) = K ◦ J LM (u).
Proof. What we have to prove is that by coupling the right side of
(3.4.4) with ∂rj (in the contraction between forms and vector fields), we
¯ j (u, ū). In the contraction, the projection p can be disregarded.
get ∂ ∂r
Let L be a vector field continuation of u; we begin by noticing that
1 1
h∂rj , J [L, L̄]i = hJ ∗ ∂rj , [L, L̄]i
(3.4.5) i i
= h∂rj , [L, L̄]i,
because J ∗ ◦ ∂ = i∂. Cartan’s formula applied to the 1-form ∂rj yields
¯ j , L ∧ L̄i + L(h∂rj , L̄i) − L̄(h∂rj , Li).
h∂rj , [L, L̄]i = −h∂∂r
In the right side, we have h∂rj , Li = 0 since L ∈ T 1,0 M and h∂rj , L̄i = 0
since the coupling of a (1, 0)-form with a (0, 1)-vector field gives 0. Also,
by the identity ∂∂¯ = −∂ ∂,¯ we get

¯ j , L ∧ L̄i = ∂ ∂r
h∂rj , [L, L̄]i = h∂ ∂r ¯ j (u, ū).


Remark 3.4.7. We have already encountered this circumstance in
¯ is alternant on T 1,0 ⊕T 0,1 , but, when restricted
Section 1.9: the form ∂ ∂r
to (T ⊕ {0}) × ({0} ⊕ T 0,1 ), it becomes hermitian.
1,0
138 3. REAL/COMPLEX STRUCTURES

∗
Cn )zo ∼ Rl and set rξ := j ξj rj ;
P
Definition 3.4.8. Let ξ ∈ (TM
the “microlocal” Levi form is the 1-component hermitian form defined
by

LξM (zo )(u, ū) := hξ, LM (z)(u, ū)i

(3.4.6)
¯ ξ (u, ū) for u ∈ T C M .
= ∂ ∂r zo

Note that this is a real-valued 1-component function: LξM (zo ) :

T M ×T 1,0 M T 0,1 M → R.
1,0

Notes. Intrinsic/extrinsic Levi form, involutivity of T C M and

Levi flatness of M are classical notions. We refer to the books by
Boggess [27] and Baouendi, Ebenfelt and Rothschild [8] for a more
detailed account of these topics.

3.5. Real/complex symplectic spaces

Summary. The content of this section is the link between the

Levi form LM of a submanifold M ⊂ Cn and the symplectic 2-form
∗
σ = dζ ∧ dz restricted to the conormal bundle TM Cn . We show that if
0 ∗
M has constant Levi corank s , or equivalently, TM Cn is a CR sub-
manifold of T ∗ Cn , there is a complex Hamiltonian foliation of TM ∗
Cn .
By a symplectic “blow up”, we show how to interchange a higher-
codimensional submanifold with a hypersurface preserving s0 . Next, we
assume in addition that M is C ω . In this case, the foliation is straight-
∗ ∗ n−s0 0
enable, that is, TM Cn takes the form TM 0C × Cs for M 0 a Levi
0
non-degenerate hypersurface of Cn−s . By an argument of geometric
optics, we also show how to “deform” M into a hypersurface passing
through M and keeping s0 unchanged. Application of this result to the
¯
tangential ∂-system is presented.

Let z = (z1 , ..., zn ) be complex variables in Cn and (z, ζ) the canon-

ically associated coordinates in T ∗ Cn for the choice of dz1 , ..., dzn as a
basis of 1-forms. Let
Xn
ω= ζj dzj
j=1

and
n
X
σ = dω = dζj ∧ dzj
j=1
 3.5. REAL/COMPLEX SYMPLECTIC SPACES 139

be the canonical 1- and 2-forms, respectively: we also write ω = ζdz

and σ = dζ ∧ dz for short. The form σ is “symplectic”, that is, non-
degenerate. Decompose σ = σ R + iσ I and note that both σ R and σ I
induce a real symplectic—non-degenerate—structure on (T ∗ Cn )R '
T ∗ R2n . For the form σ we call a complex submanifold V such that
T V ⊥σ ⊂ T V (resp. T V ⊂ T V ⊥σ ) involutive (resp. isotropic). We call
a submanifold which is both involutive and isotropic lagrangian: note
that its dimension must be n, half of the dimension of T ∗ Cn . We use
similar terminology for the forms σ R and σ I . Thus, a real submanifold
V is R- (resp. I-) involutive/isotropic/lagrangian when it is involu-
tive/isotropic/lagrangian for σ R (resp. σ I ). Note that the real dimen-
sion of an R- or I-lagrangian submanifold must be 2n, half the real
dimension of (T ∗ Cn )R . Let M be a real generic submanifold of Cn
∗
and let TM Cn be its conormal bundle, the set of the forms whose real
part vanishes on T M . Let r = 0 for r = t (r1 , ..., rl ) be a system of
C-independent equations for M , that is, ∂r1 ∧ ... ∧ ∂rl 6= 0. Through
∗ ∼
the basis ∂r1 , ..., ∂rl of TM Cn , we have an identification M × Rl →
∗
Cn , (z, ξ) → (z, ∂rξ (z)) where ∂rξ := j ξj ∂rj . First note that for
P
TM
∗
any (z, ξ) ∈ TM Cn we have ω R (z, ξ)|Tz M = 0 and therefore, by differ-
entiating,
(3.5.1) σ R (z, ξ)|T(z,ξ) TM∗ Cn = 0.
∗ ∗
Thus, TM Cn is R-lagrangian and in particular σ = iσ I over T TM Cn .
∗ n
Also, note that ω(z, ξ)|Tz M 6= 0 or, in other words, TM C is “regular”
since, otherwise, we would have ∂rξ |T M = 0 which is forbidden by the
C-independence of the ∂rj ’s, that is, the genericity of M . By (3.5.1)
and by C-linearity we get
∗
(3.5.2) ker σ|T TM∗ Cn = T C TM Cn .
∗
Thus, in particular, TM Cn is totally real (that is, its T C is 0) if and only
∗
if σ|T TM∗ Cn , or equivalently, σ I |T TM∗ Cn , is non-degenerate. We call TM Cn
∗
“I-symplectic” in this case. More generally, TM Cn is a CR submanifold
∗ n
of T C , that is, its T has constant rank, if and only if σ|T TM∗ Cn or
C

σ I |T TM∗ Cn have constant rank. We go further with our discussion. We

differentiate ξ∂r(z)|T TM∗ Cn and get
(3.5.3) ¯ ξo |Tz M .
σ|T(zo ,ξo ) TM∗ Cn = dξ ∧ ∂r|Rl ×Tzo M + ∂∂r o

We take a closer look of (3.5.3) which we will also describe in coor-

dinates in Exercise 3.5.1. If the alternant form ∂∂r ¯ is restricted from
1,0 0,1 1,0 0,1
T ⊕ T to (T ⊕ {0}) × ({0} ⊕ T ), it becomes hermitian, but
if, instead, it is restricted to the diagonal(T Cn )R × (T Cn )R , it becomes
140 3. REAL/COMPLEX STRUCTURES

¯ = i∂∂r
purely imaginary, that is, ∂∂r ¯ I on vectors (ż + ż, ż 1 + ż 1 ). More
precisely
¯ ż + ż, ż 1 + ż 1 ) = 2iLr (ż, ż 1 ).
∂∂r(
Also, in the first term in the right of (3.5.3), we have ∂rR |T M = 0.
Thus (3.5.3) shows once more that σ R |T TM∗ Cn = 0. Moreover, if we
restrict (3.5.3) to vectors which are not only in T M but also in T C M ,
then the first term on the right is 0 and we get
(3.5.4) σ|T(zo ,ξo ) TM∗ Cn ∩π−1 (T C M ) = 2iLrξo |TzCo M .
In particular, this restriction applies to ker σ|T TM∗ Cn which is contained
in π −1 T C M . Thus (3.5.4) shows that the projection π induces an iden-
tification
∼
(3.5.5) ker σ → ker LM .
Before going ahead with our discussion, we want to relax and have an
alternative proof of what we have seen so far.
Guided Exercise 3.5.1. Prove (3.5.2) and (3.5.5) in coordinates.
(a) Start by proving
∗
(3.5.6) T(zo ,ξo ) TM Cn = {(ż, ζ̇) : ż ∈ Tzo M, ζ̇ = ξ∂r(zo )
+ ∂ż ∂rξo (zo ) + ∂ż ∂rξo (zo ) for ξ ∈ Rl }.
(Here (3.5.6) is obtained by differentiating the parametrization M ×
∗
Rl → TM Cn , (z, ξ) 7→ (z, ξ∂r(z)).)
(b) Describe the relation between σ|T TM∗ Cn and LM . First prove that
σ(zo , ξo ) = d(ξ∂r)(zo , ξo )
¯ ξo (zo ).
= dξ ∧ ∂r(zo ) + ∂∂r
1
Next, observe that for any ξ, ξ 1 ∈ Rl , ż+ż ∈ Tzo M and ż 1 +ż ∈ Tzo M ,
we have
1 1
hd(ξ∂r)(zo , ξo ) , (ξ, ż+ż)∧(ξ 1 , ż 1 +ż )i = 2i(∂rξI ż 1 −∂rξ I ż+Lrξo (ż, ż 1 )).
1
Conclude that if ż + ż ∈ TzCo M and ż 1 + ż ∈ TzCo M , then
1
hd(ξ∂r)(zo , ξo ) , (ż + ż) ∧ (ż 1 + ż )i = 2iLrξo (ż, ż 1 ).
This proves once more (3.5.4).
(c) Carry out the proof. Show that if
∗
(ż, ζ̇) = i(ż 1 , ζ̇ 1 ) ∈ T TM Cn ,
 3.5. REAL/COMPLEX SYMPLECTIC SPACES 141

then (3.5.6) implies

(
ż = iż 1 ∈ T M, that is, ż ∈ T C M,
(3.5.7)
2∂ż ∂rξo = (−ξ + iξ 1 )∂r, that is, ż ∈ ker Lrξo .
Thus (3.5.2) and (3.5.5) have been re-proved.
We want to go ahead with our discussion and remark that, after we
achieved our goal of relating LM to σ, we have evidence that s0M , the
∗
number of the null eigenvalues of LM , is a symplectic invariant. If TM Cn
∗ n 0 0
is transformed to TM1 C , then sM1 = sM . Of course, this is not true for
s±
M ; a symplectic transformation can interchange + with − and also
affect the codimension l1 6= l. We well see through explicit examples
how this may occur. We wish to adopt a new notation. We write M 99K
M1 when M and M1 are related by a symplectic transformation, say
χ, which interchanges their conormal bundles, that is, when we have a
commutative diagram
∗ ∗ χ
TM Cn → TM 1
Cn
π↓ ↓π
M 99K M1 .
First, we want to show that, under a change of coordinates in T ∗ Cn , “a
submanifold is generically a hypersurface”. Let M be a submanifold of
Cn defined, in coordinates z = (z 0 , z 00 ) with z 0 = (z1 , ..., zl ) at zo = 0,
by the equations
(3.5.8) y 0 = hj (x0 , z 00 , z̄ 00 ), hj (0) = 0, ∂hj (0) = 0.
We consider the transformation
(3.5.9) χ : T ∗ Cn → T ∗ Cn , χ = χ̃ × id,
where
ζ0
(3.5.10) χ̃ : T ∗ Cl → T ∗ Cl , (z 0 , ζ 0 ) 7→ (z 0 +  P 02 ; ζ 0 ).
j ζj

Here  is a small real parameter. We check readily that χ is (holo-

∗
morphic and) symplectic. We then define M1 := πχ(TM Cn ); since
T M1 = T M ⊕ T S where S is the -sphere in the plane Rl of
l−1 l−1

the y 0 ’s and M1 is a hypersurface and remains a hypersurface also un-

der any homogeneous transformation close to χ. As we have already
noticed, we must have s0M1 = s0M . On the other hand, the “gain” of
l − 1 in dimension passing from M to M1 is obtained by “adding” the
sphere Sl−1 (from the outward side). Since this has l − 1 positive Levi
eigenvalues, we get for suitable 
s− −
M1 = sM , s+ +
M1 = sM + (l − 1).
142 3. REAL/COMPLEX STRUCTURES

This can be readily checked. When LM , or equivalently, σ|T TM∗ Cn , have

constant rank, then we have developed in Sections 2.3 and 3.4 two
different theories of integration of their (involutive) kernels; we want
to put them now in a unitary frame. We know on one hand that
Ker LM is involutive on account of Theorem 3.4.5. Thus, M is foli-
ated by the integral leaves of Ker LM —the Levi leaves—and these are
complex manifolds since their tangent bundle, Ker LM , is a complex
∗
bundle. In general, the complex structure of TM Cn being poorer than
the one of M , the complex leaves of M cannot be “lifted” to com-
∗
plex leaves of TM Cn . However, in the present case of Levi leaves, owing
π0
∗ n ∼
to T TM C → Ker LM , for any leaf Γ
C
on M through z and for any
∗
ξ ∈ TM Cnz there is the unique complex section Γ∗ through (z, ζ) over
∗
Γ, that is, satisfying πΓ = Γ. This is also in accordance with Theo-
rem 2.3.5 which gives in addition the trivialization of the foliation (but
under a C ∞ , not holomorphic, change of coordinates). In either way
we have obtained
Theorem 3.5.2. Let M be generic and have constant Levi rank.
∗
Then TM Cn is foliated by the integral leaves of ker σ = π 0 −1 ker LM .
We come back to the foliation {Γ} on M and notice that it can by
0
no means be trivialized, that is, reduced to a “product” M = M 0 × Cs
by a complex holomorphic change of coordinates in Cn . This is true
only for a smooth transformation of R2n . A celebrated example is the
“tube” X
M = {z ∈ Cn : y12 = yj2 }.
j≥2
We choose one of the two cones with vertex at y = 0, say the positive
cone y1 > 0, and get what is called the “future tube” (“past tube”
corresponding to the complementary choice y1 < 0). This is a relevant
model of symmetric space in quantum mechanics. Here we have a “ra-
dial” Levi foliation given by the system of radii defined for ζ = ξ +iη by
Γζ = ξ +iCη. These rays are not straightenable as is shown, e.g., in [41]
by an investigation on the 3-jet of M . However they are straightenable
at least when M is C ω if we tackle the problem from “upstairs”.
Theorem 3.5.3. Let M be generic, C ω , and have constant Levi
∗
rank. Then, the Hamiltonian foliation of TM Cn is straightenable: there
are complex symplectic homogeneous coordinates (z, ζ) in T ∗ Cn such
∗
that TM Cn takes the form
∗ ∗ 0 0
TM Cn ' TM 0C
n
× Cs ,
0
where M 0 is a Levi non-degenerate hypersurface of Cn for n0 = n − s0 .
 3.5. REAL/COMPLEX SYMPLECTIC SPACES 143

∗
Proof. Let V1 be a partial or intrinsic complexification of TM Cn
∗ n
in T C (cf. Section 3.3) and let f1 = 0, ..., fso = 0 be a system of
holomorphic equations which define V1 in T ∗ Cn . We choose ζ̃n−so +1 =
f1 , ..., ζ̃n = fs0 and supplement to a full system of coordinates according
to the obvious holomorphic version of Theorem 2.3.5. In the new coor-
0 0 0
dinates, V1 is interchanged with T ∗ Cn × Cs and TM ∗
Cn with Λ0 × Cs
where Λ0 is an R-lagrangian submanifold. But it is in fact I-symplectic
0 0 0
because (Λ0 ×Cs )⊥ = {0}×Cs . By choosing new coordinates in T ∗ Cn
such that Span{∂ζj0 }j=1,...,n−s0 ∩T(zo ,ζo0 ) Λ = ζo0 R, the real line engendered
n0
by the Euler vector field, we achieve that Λ0 = TM ∗
0C for a Levi non-
0 n0
degenerate hypersurface M ⊂ C .


Remark 3.5.4. The numbers s+ , s− and s0 are invariant under

holomorphic change of coordinates in Cn . In T ∗ Cn , as has already been
noticed, the only invariant is s0 . In fact, we can give s+ and s− the
values we wish provided that s0 is respected. For instance, let M = ∂Bn
be the boundary of the unit ball in Cn , and let us “microlocalize” our
attention at (zo , ζo ) = ((1, 0, ...), (−1, 0, ...)); then, s−
M = n − 1, that is,
M is strongly pseudoconcave. By moving the boundary points along
the rays of Bn with a “caustic” at 0, we can reverse the curvature. In
detail, let
ζ
(3.5.11) χ : (z, ζ) 7→ (z − λ P 1 ; ζ) for (z, ζ) close to (zo , ζo ),
( j ζj2 ) 2

where λ > 1. Though this transformation is not globally defined in

T ∗ Cn because of the vanishing of the denominator, however it makes
∗
sense along TM Cn . If M 99K M̃ is the base correspondence underlying
(3.5.11), it transforms s− +
M = n − 1 into sM̃ = n − 1. Thus M̃ is strongly
pseudoconvex at (z̃o , ζ̃o ) = χ(zo , ζo ) whereas M was strongly pseudo-
concave at (zo , ζo ). However, we cannot always achieve s+ M̃
= n − 1,
0 0
that is, M̃ strongly pseudoconvex, unless sM̃ = sM = 0.

Theorem 3.5.3 is the main tool in solving the problem that we de-
scribe now. Let ξo be a conormal to M at zo and assume rk LM (z, ξ) ≡
const in a neighborhood of (zo , ξo ): is there a way to find a hypersur-
face S ⊃ M with conormal ξo at zo such that s0S = s0M ? If M is Levi
non-degenerate, defined, say, by r1 = 0, ..., rl = 0, we may define our
“extension” S of M for instance by

S = {z : rξo (z) + c(r12 (z) + ... + rl2 (z)) = 0},

144 3. REAL/COMPLEX STRUCTURES

for large c. Otherwise, if M is degenerate but in the favorable case in

0
which M is a product M = M1 × Cs , we can do a similar construc-
0
tion with parameters which leads to S = S1 × Cs . In general, it is
Theorem 3.5.3 which comes to our rescue.
Theorem 3.5.5. If s0 ≡ const in a neighborhood of (zo , ξo ) and M
is C ω , then there exists a hypersurface S ⊃ M , with conormal ξo at zo
such that s0S ≡ s0M .
Proof (Cf. [102, Theorem 1]). Here is the scheme of the proof.
By a symplectic transformation we interchange M with a hypersurface
and reduce it to a product. Next, deform the non-degenerate term so
that it is brought back to a hypersurface by the inverse transforma-
tion: the deformation, as well as the symplectic transformation, does
not affect s0 since this is a symplectic invariant. In detail, we choose
coordinates such that ξo ∈ Rl (by the genericity of M ) and consider
the transformation defined by (3.5.9) and (3.5.10) that we denote by
χ1 . It interchanges M 99K M1 where M1 is a hypersurface and satis-
fies s+ +
M1 = sM + (l − 1). Next, we use the symplectic transformation
provided by Theorem 3.5.3, which we denote by χ2 and which inter-
0
changes M1 99K M2 = M20 × Cs without affecting the numbers s± . We
now want to deform M2 . We take R20 ⊂ M20 of dimension 2(n − so ) − l
∗
transversal to χ2 χ1 (TM Cnzo ). We take a real analytic hypersurface S20
which intersects M2 along R20 with order of contact 2 and stays in the
0
0
−χ2 χ1 (po ) side of M20 . We also set S2 = S20 ×Cs . We go now back along
χ−1 −1 −1 ∗ n
2 and χ1 and see that χ2 (TS2 C ) is still the conormal bundle to ¡a
hypersurface, say TS∗1 Cn , and χ−1 ∗ n
1 (TS1 C ) is still the conormal to a hy-
persurface, say, TS∗ Cn . We can describe the scheme of our construction
by
0
M 99K M1 99K M20 × Cs
↓
0
S L99 S1 L99 S20 × Cs .
It is clear that S still contains M and
s0S = s0M ,
because of the symplectic invariance of the s0 ’s. Also, by the choice
of our orientations, s+ +
S = sM + (l − 1); thus, the jump of l − 1 in the
dimension results in a full gain for s+ .

Theorem 3.5.5 is applied in [102] to the study of the tangential
¯
∂-system. Let Ω = Ω± be the two local components of Cn \ S in a
neighborhood of a point zo ∈ S. We denote by Kj (Ω̄) the cohomology
 3.5. REAL/COMPLEX SYMPLECTIC SPACES 145

of ∂¯ on C ∞ (Ω̄)-forms of degree j and by Kj (Ω̄)zo its “germ” at zo (cf.

Section 1.9). Assume that s− ∂Ω ≡ q in a neighborhood of zo , for the
choice of the outward conormal to Ω; then
(3.5.12) Kj (Ω̄)z ≡ 0 for any j 6= 0, q and z close to zo .
In fact, the vanishing of (3.5.12) for j ≥ q + 1 is proved in Section 1.9.
The vanishing for j ≤ q − 1 is easier; this is related to the q-concavity
of Ω and can be proved, e.g., by the integral method (cf. [27, Chap-
ter 24]. Conversely, if (3.5.12) holds for any j ≥ 1, then Ω must be
pseudoconvex as the implications (iii) ⇒ (iv) ⇒ (i) of Theorem 1.8.13
show. The statement therein concerns the cohomology for C ∞ (Ω)-, in-
stead of C ∞ (Ω̄)-, forms. But it is readily seen that non-extendibility
for functions in hol(Ω) or in hol(Ω) ∩ C ∞ (Ω̄) are the same. To see this
− √1
at z = 0, just use f = e z instead of f = z1 . Thus, if s− ≡ q for q 6= 0,
since the only non-trivial cohomology may occur in degree q, then this
must be 6= 0:
(3.5.13) Kq (Ω̄)z 6= 0 for any z close to zo .
We consider now a generic submanifold M ⊂ Cn of codimension l > 1
and denote by Kj (M ) the cohomology in degree j of the tangential, or
induced, system ∂¯M over C ∞ (M )-forms. If S is a hypersurface which
passes through M , the natural restriction induces a morphism in the
cohomology
(3.5.14) Kj (S) → Kj (M ).
Again, the restriction from the two half spaces Ω̄± to S induces a
decomposition
(3.5.15) Kj (Ω̄+ ) ⊕ Kj (Ω̄− )  Kj (S),
which is an isomorphism for any j ≥ 1 (cf. for example [6]). We then
have the following consequence of Theorem 3.5.5.
Theorem 3.5.6. Let M be a C ω , generic, submanifold of Cn ,
∗
(zo , ξo ) a point of TM Cn , and assume
s± ± ∗ n
M (z, ξ) ≡ q for (z, ξ) ∈ TM C close to (zo , ξo ).

Then, for any z close to zo

Kj (M )z 6= 0 if j = 0, q + , and q − .
Proof. We prove that Kj (M )z 6= 0 for j = q − . On account of
Theorem 3.5.5, we take S which contains M , has conormal ξo at zo and
satisfies s− − ∗ n
S (z, ξ) ≡ q for any (z, ξ) ∈ TS C close to (zo , ξ). According
−
to (3.5.13), we then have Kq (Ω̄+ )z 6= 0 and hence also Kj (S)z 6= 0
146 3. REAL/COMPLEX STRUCTURES

because of (3.5.15). Finally, owing to [53] or [4], we know that (3.5.14)

is injective for j = q − and this implies Kj (M )z 6= 0. The non-vanishing
+
of Kq (M )z can be proved likewise, by switching from ξo to −ξo and
from Ω̄+ to Ω̄− .


Suggested research. (1) When s0M is constant, it is not nec-

essary to assume N to be C ω for the Hamiltonian foliation to exist:
this is a consequence of Theorem 2.3.7. Is this foliation straightenable
in T ∗ Cn if M is only C ∞ ?
(2) Theorem 3.5.5 and its corollary Theorem 3.5.6 are likely to
be true if M is not necessarily C ω but just C ∞ but their proof does
not apply to this situation. Try an alternative proof which sustains the
case C ∞ .

Notes. The relation between the symplectic and the Levi forms
is classical and very fruitful in the geometric study of PDE’s. We fol-
lowed here the guidelines of Sato, Kashiwara and Kawai [78] and Kashi-
wara and Schapira [53]. The result on the straightening of a Levi foli-
ation in the symplectic space (Theorem 3.5.5) follows the presentation
by Zampieri in [102]. The non-straightening in the base space is dis-
cussed, among others, by Ebenfelt in [41]. The deformation argument
contained in Theorem 3.5.3 is developed by Zampieri in [102] as well
as the non-vanishing of the tangential cohomology contained in Theo-
rem 3.5.5. The latter was already obtained by Andreotti, Fredricks and
Nacinovich [5] in case the Levi form is non-degenerate; note that, in
their setting, M need not be C ω .

3.6. Approximation of CR functions by polynomials

Summary. We present here the result of [13]: CR functions f

on generic submanifolds M ⊂ Cn can be approximated by polynomials
uniformly in controlled neighborhoods of any point. The result consists
in two steps: an approximation of the restriction of f to a totally real
submanifold N ⊂ M by means of convolution with the the heat kernel
and a deformation of the integration chain in a family of N ’s which
sweep out M . The deformation leaves the integral unchanged, because
of Stokes’s theorem and since f is CR. This result will play a crucial
 3.6. APPROXIMATION OF CR FUNCTIONS BY POLYNOMIALS 147

role in Sections 3.7, 3.8 and 3.10 where the holomorphic extension of
CR functions is studied.

We first discuss CR functions which are not C 1 . In Rn , we have an

identification of locally integrable functions with distributions given by
Z
1 n 0 n
(3.6.1) Lloc (R ) → D (R ), f 7→ [v 7→ f vdV ],
Rn
where v ranges through the family of Cc∞
functions and dV denotes the
element of volume in Rn . Let M be an oriented submanifold of Cn of
codimension l: the identification analogous to (3.6.1) requires
V the choice
of a (2n − l)-form α (an element of the exterior algebra T ∗ M ) with
2n−l
C ∞ coefficients everywhere 6= 0. This yields
Z
0
L1loc (M ) → D (M ), f 7→ [v 7→ f vα].

Let M be generic of codimension l. We choose a local basis dt1 , ..., dtn−l

of closed forms in (T 0,1 M )∗ and complete into a full basis of C ⊗ T ∗ M :
dξ1 , ..., dξn , dt1 , ..., dtn−l .
We denote the dual basis of C ⊗ T M by
X1 , ..., Xn , L̄1 , ..., L̄n−l ;
thus, L̄1 , ..., L̄n−l are a system of CR vector fields which span T 0,1 M .
We choose α := dξ1 ∧ ... ∧ dξn ∧ dt1 ∧ ... ∧ dtn−l as the 2n − l integration
form on M . For the adjoint differential operators we have
t t
(3.6.2) Xj = −Xj ,
L̄j = −L̄j ,
R R R
in R the sense that Xj (u)vα = − uXj (v)α and L̄j (u)vα =
− uL̄j (v)α for any u, v ∈ Cc∞ . This is easily proved by means of
Stokes’s formula. For example,
Z Z Z
L̄j (u)vα = − uL̄j (v)α + L̄j (uv)α
Z Z ∨
n+j+1
= − uL̄j (v)α + (−1) d(uv... ∧ dtj ∧ ...)
Z Z ∨
= − uL̄j (v)α + uv... ∧ dtj ∧ ...
∂(supp(uv))
Z
= − uL̄j (v)α.

We have used here the symbol “∨” to mean that the corresponding item
is omitted. Since every differential operator P on M is a combination
148 3. REAL/COMPLEX STRUCTURES

of the Xj ’s and the L̄j ’s, then we can define t P starting from (3.6.2).
With this preliminary, we are in a position to define “weak solutions”
of the tangential CR equations.
Definition 3.6.1. We say that u is a CR distribution, and we write
0
u ∈ DCR , when L̄u = 0, for all L̄ ∈ T 0,1 M , in the sense that
(3.6.3) hu, t L̄(v)i = 0 for any v ∈ Cc∞ ,
where h·, ·i denotes the duality between D0 and Cc∞ .
The adjunction being made under a choice of the basis Xj , L̄j , the
1
R tdefinition is not invariant. If u ∈ Lloc , thenV(3.6.3) is equivalent
above V
to u L̄(v)α = 0 for any L̄ and v where α = dξj ∧ dtj
j=1,...,n j=1,...,n−l
which declares the dependence of the condition on the choice
R of the
0
density α. In particular, we can characterize u ∈ CCR by udβ = 0
for any β with compact support of type (n, n − l − 1), that is, of type
P ∨
β = j βj ...dtj ....

Proposition 3.6.2. If u is C 1 , then the notions of “weak solution”

and “classical solution” coincide.
P ∨
Proof. Let β = j βj ...dtj ...: we have
Z Z Z
udβ = − du ∧ β + d(uβ)
Z
= − du ∧ β (by Stokes’s theorem)
X Z
j+n
= (−1) L̄j (u)βj α.
j
R
Thus, L̄j (u) = 0, for all j, is equivalent to udβ = 0 for any β of
type (n, n − l − 1), which proves the proposition.


We state the approximation theorem due to Baouendi and Treves

[13].
Theorem 3.6.3. Let M be a C ∞ , CR submanifold of Cn and zo
a point of M . Then, there exists a relatively compact neighborhood M 0
0 k ∞
of zo on M such that any u ∈ CCR (M ) (resp. CCR (M ), CCR (M ),
0
DCR (M )) is approximated by a sequence of polynomials in the conver-
0
gence of CCR (M 0 ) (resp. CCR
k
(M 0 ), CCR
∞
(M 0 ), DCR
0
(M 0 )).
 3.6. APPROXIMATION OF CR FUNCTIONS BY POLYNOMIALS 149

Proof. The proof is given in several steps throughout the rest of

the section. We begin by remarking that M is CR diffeomorphic to a
0
generic submanifold (of some subspace Cn for n0 ≤ n): hence it is not
restrictive to assume M generic from the beginning. We declare first the
kind of the approximation. We take a germ of a totally real maximal
submanifold N ⊂ M . We assume that the equations of M and N are
normalized at zo as follows. The manifold M is defined by yj0 = hj for
any j ≤ l and N is defined by yj = hj for any j ≤ n with hj (0) =
0, ∂hj (0) = 0. In particular, |Im z| <  on suitable neighborhoods M 0
and N 0 of zo on M and N , respectively. With this information and
by taking u with compact support u ∈ E 0 (M ), we define, for a large
parameter λ, the convolution of u with the heat kernel:

  n2
λ 2
(3.6.4) Tλ u(z) := hu(ξ)|N , e−λ(z−ξ) i, z ∈ N.
π

If u ∈ L1 (N ) has compact support, then this definition

R −λ(z−ξ)coincides
2
with the one provided by the integral Tλ u(z) := N e u(ξ)dξ.

n
λ 2 −λ(z−ξ)2
We denote by Kλ (z, ξ) = π e the heat kernel. We notice
that for the CR distribution u, the restriction u|N makes sense. To
see this, according to [51], it suffices to show that the wave front set
satisfies

W F (u) ∩ TN∗ M = {0}.

In fact, on one hand we have that, u being a solution of the sys-

tem of CR vector fields, W F (u) is contained in the characteristic set
⊥ ∗ TM ∗
0,1 0,1


char(T M ) = T M ∩ T M ' T C M (defined in Definition 3.2.5).

∗
On the other hand, N being generic, we have TN∗ M ∩ TTCMM = {0}.
We now prove that Tλ u → u over N . In Rn , we use the notation
dξ = dξ1 ∧ ... ∧ dξn and ξ 2 = ξ12 + ... + ξn2 . We claim that it suffices to
prove the convergence C 0 . In fact, for the different types of convergence
we have the implications
(i) C ∞ ⇒ E 0 : this follows from

Z Z
Tλ uvdξ = hu, Kλ ivdξ
Z
= hu, Kλ vdξi = hu, Tλ vi.
150 3. REAL/COMPLEX STRUCTURES

(ii) C 0 ⇒ C k and C ∞ : in fact, recalling that supp(u) is compact,

Z
Tλ (Xj u) = Kλ Xj (u)dξ
Z Z
= Xj (Kλ u)dξ − Xj (Kλ )udξ
Z
= − Xj (Kλ )udξ (by Stokes’s formula)
Z
= Xj ( Kλ udξ) = Xj (Tλ u).

Thus, what we have to show is that for any totally real maximal
submanifold N ⊂ M and for a suitable N 0 ⊂ M , we have Tλ u ⇒
u uniformly on N 0 . We first consider the case N = Rn (and next de-
form it by a diffeomorphism). We start from the identity
  n2 Z
λ 2
(3.6.5) e−λ(x−ξ) dξ = 1.
π Rn
√
By using (3.6.5) and by means of the coordinate change s := λ(x − ξ)
for x ∈ Rn , we get
  n2 Z
λ 2
Tλ u(x) − u(x) = e−λ(x−ξ) [u(ξ) − u(x)] dξ
π Rn
(3.6.6)   n2 Z  
1 −s2 s
= e u(x − √ ) − u(x) ds.
π Rn λ
The term between bracket, which we denote by uλ , satisfies uλ → 0
2 2
and e−s |uλ | < e−s ; thus, by the Lebesgue dominated convergence
∼
theorem, the integral converges to 0 for fixed x. We have to show that
this convergence is indeed uniform. In fact,
Z Z
|Tλ u − u| < ·+ ·
∼ |s|≥R |s|≤R
−1
= (R ) + R (λ−1 ),
where the first is infinitesimal uniformly in λ. The uniformity follows
and the case N = Rn is proved.
We take now a general, totally real maximal, submanifold N with
a local, orientation-preserving parametrization over Rn :
Rn → N, ξ 7→ Φ(ξ).
We decompose the n × n complex Jacobian matrix Φ0 as Φ0 = B + iC
and suppose C(0) = 0 in our normalization. We need
 3.6. APPROXIMATION OF CR FUNCTIONS BY POLYNOMIALS 151

Lemma 3.6.4. We have the identity

  n2 Z
λ 2
e−λ((B+iC)ξ) det(B + iC)dξ = 1.
π Rn

Proof. The above integral is an entire function of the coefficients

of B +iC and, when B +iC is real, that is, C = 0, its value is 1 because
Z   n2 Z  π  n2
−λ(Bξ)2 1 2
e det Bdξ = e−ξ̃ dξ˜ = .
Rn λ Rn λ
By the identity principle for holomorphic functions, its value is identi-
cally 1 for any choice of C.

End of proof of Theorem 3.6.3 We start from the equality
  n2 Z
λ 0 2
(3.6.7) u(Φ(x)) = e−λ(Φ (x)·(ξ−x)) det Φ0 (x)u(Φ(x))dξ,
π Rn

which is a consequence of Lemma 3.6.4. For z = Φ(x), we have to

estimate
(3.6.8)
  n2 Z
λ 2
Tλ u(z) − u(z) = e−λ(Φ(ξ)−Φ(x)) u(Φ(ξ)) det Φ0 (ξ)dξ
π Rn
Z 
−λ(Φ0 (x)·(ξ−x))2 0
− e u(Φ(x)) det Φ (x)dξ
Rn
  n2 Z
λ 2
= e−λ(Φ(ξ)−Φ(x)) [u(Φ(ξ)) det Φ0 (ξ) − u(Φ(x)) det Φ0 (x)] dξ
π Rn
  n2 Z 
λ −λ(Φ(ξ)−Φ(x))2 −(λΦ0 (x)·(ξ−x))2

+ e −e u(Φ(x)) det Φ(x)dξ.
π Rn

We denote by [∗] the term between brackets in the second to last term
of (3.6.8). We can assume that
(
Re (Φ(x) − Φ(ξ))2 ≥ c|ξ − x|2 − c1 ,
(3.6.9)
Re (Φ0 (x) · (ξ − x))2 ≥ c|ξ − x|2 − c1 .
It follows that
 
e−λ(Φ(ξ)−Φ(ξ))2 [∗] < e−λ|ξ−x|2 E(|ξ − x|),
∼
(3.6.10)  
e−λ(Φ(ξ)−Φ(x))2 − e−λ(Φ0 (x)·(ξ−x))2  < e−λ|ξ−x|2 E(|ξ − x|),
∼
152 3. REAL/COMPLEX STRUCTURES

where E(|ξ − x|) is infinitesimal (in fact we could replace E(|ξ − x|) by
O(|ξ − x|) in the second line). By a similar argument to that which
follows (3.6.6), we see that
  n2 Z
λ 2
(3.6.11) lim e−λ|ξ−x| Edξ = 0.
λ→∞ π Rn
√
To prove it, we just have to set s = λ|ξ − x| and apply the domi-
nated convergence theorem. By (3.6.8), (3.6.10) and (3.6.11) we have
Tλ u(z) → u(z) as λ → ∞. Next, by the same argument as for N = Rn ,
we conclude that Tλ u ⇒ u uniformly on a suitable N 0 ⊂ N . Now,
we want to show that this convergence also takes place out of N 0 in
a controlled neighborhood M 0 of N 0 in M . For this, we remark that
any point z in M 00 ⊂⊂ M 0 can be reached by a deformation Nz0 of N 0 ,
parametrized over Rn by Φz such that
(
N 0 and Nz0 coincide near ∂M 0 ,
|Im Φ0z | < 21 .

If we then define Tλ (u) by integration along Nz0 , we get Tλ u → u

on Nz0 ∩ M 0 . But, in fact, this integration coincides with the former
one which was performed along N 0 . In fact, let W be an (n + 1)-
dimensional manifold in M with (oriented) boundary +N − Nz and
take u ∈ Cc∞ (M 0 ), CR over W̄ , and an integration form α of degree n
over W̄ . Then
Z Z Z
Kλ uα − Kλ uα = d(Kλ uα)
Nz N W
Z
= ¯ = 0.
Kλ ∂uα
W
0
This shows that the C approximation Tλ u → u obtained by integra-
tion along a prescribed N 0 also extends to M 00 . The various kinds of
convergences, C k , C ∞ , E 0 , follow in the same way as before.

Exercise 3.6.5. Suppose that M has some non-degenerate elliptic
CR singularities in the sense of Section 3.2 which form a manifold Σ.
Prove that there is no local uniformity at points of Σ for the polynomial
approximation of CR functions on M \ Σ.
Hint: in the simplest case of a 2-codimensional submanifold M ⊂ C2
with a non-degenerate elliptic singularity at 0, take a real function on
M . If f is approximated by polynomials on K ⊂⊂ M \ {0}, then it is
constant over discs whose boundary is in K.
 3.7. ANALYTIC DISCS AND THE EXTENSION OF CR FUNCTIONS 153

Notes. Approximation by polynomials of continuous real func-

tions goes back to Stone-Weierstrass. In the specific case of CR func-
tions (Theorem 3.6.3), the result is due to Baouendi and Treves [13].
Further generalizations can be found in Treves [88].

3.7. Analytic discs and the extension of CR functions: the

“edge of the wedge” theorem, the deformation of discs
for manifolds of type 2 and the Levi extension

Summary. We first describe the construction of analytic discs

by solving a functional equation, Bishop equation, in the spaces of dif-
ferentiable functions with fractional regularity. Once this theory is set-
tled, we prove two problems. First we prove the edge of the wedge theo-
rem by [3]: CR functions on M which extend to additional directions vj
extend in fact to their convex combinations. This is obtained by a con-
struction of a family of discs, parts of whose boundaries belong to the
different regions of extension, and in proving that their centers cover
a wedge W of edge M and normal cone Γ = convex hull {vj }. The
approximation of CR functions by polynomials “propagates” from the
boundary of the discs to their interior and yields the holomorphic exten-
sion to W . Second, we prove the CR extension in Levi directions of [29].
For this, we show that if L ∈ T 1,0 M and 2i1 [L, L̄](zo ) = v 6= 0 in TTCMM ,
then there is a one-parameter family of discs {Aη } with Aη |η=0 = {zo }
and CR components ηL(zo )τ , τ ∈ ∆, such that ∂η Aη points to a di-
rection whose projection modulo T M is J v. Again, by approximation
and by the maximum principle, we get CR extension to the Levi cone.

Analytic discs A are the holomorphic bounded images into Cn of the

standard disc ∆ of C. These are “attached” to a generic submanifold
M ⊂ Cn , that is, ∂A ⊂ M . The basic idea is that “small” analytic
discs attached to M are perturbations of analytic discs in T C M . We first
introduce the Hilbert transform. If u : ∂∆ → R is a smooth function, it
has a unique harmonic extension to ∆: ¯ this extension u has a harmonic
¯
conjugate v on ∆, that is, a function v such that u + iv is holomorphic;
this is determined up to a constant. The Hilbert transform is the
correspondence T : u|∂∆ 7→ v|∂∆ and is denoted by T1 when it is
normalized by the condition v(1) = 0. Note that u = −T1 v + u(1).
It is classical that T1 is not a continuous functional over functions
with integer regularity C k . Instead, we let it act on the spaces C k,α =
C k,α (∂∆), for 0 < α < 1, of functions f endowed with continuous
154 3. REAL/COMPLEX STRUCTURES

derivatives up to order k which satisfy

|∂τk f (eiθ1 ) − ∂τk f (eiθ2 )|
||f ||k,α = ||f ||k + sup < +∞.
θ1 , θ2 |θ1 − θ2 |α
Here, ||f ||k is the usual C k -norm ||f ||k = sup |∂τj f (eiθ )|. Endowed with
jθ
k,α
the norm || · ||k,α , the space C turns out to be Banach. We want
to show that T : C k,α → C k,α is continuous and for this we start by
describing through Poisson integrals the harmonic extension of u from
values τ = eiθ to z = reiϕ , r ≤ 1:
Z 2π
iϕ 1 (1 − r2 )u(eiθ )
u(re ) = dθ
2π 0 (1 + r2 − 2rcos(θ − ϕ))
1 − |z|2
Z
1 dτ
(3.7.1) = 2
u(τ )
2πi {|τ |=1} |τ − z| τ
 Z   
1 τ +z dτ
= Re u(τ ) .
2πi {|τ |=1} τ − z τ
The function between brackets [ ], say F , is holomorphic in z and
satisfies Re F = u. Note also that Im F (0) = 0 since u is real. Thus, the
harmonic conjugate v = T0 u, normalized by v(0) = 0, is the boundary
value on ∂∆ of Im F :
 Z   
1 τ +z dτ
T0 u|z=eiθ = Im u(τ ) | iθ .
2πi {|τ |=1} τ − z τ z=e
Proposition 3.7.1. The functional T : C k,α (∂∆) → C k,α (∂∆) is
continuous.
Proof. Let K be the Cauchy integral defined by
Z
1 u(τ )
Ku(z) = dτ, |z| < 1.
2πi {|τ |=1} τ − z
Since τ +z
τ
is smooth for |τ | = 1, |z| < 1, it suffices to prove the conti-
nuity of K. It is not a big deal to estimate the norm ||Ku||k by ||u||k,α .
So, we have to establish the difference estimate for ũ := ∂θk u. Now, for
z1 = eiθ1 and z2 = eiθ2 in ∂∆, we set  = |z1 − z2 |, B2 (z1 ) = {τ :
|τ − z1 | < 2} and get
Z    
1 ũ(τ ) − ũ(z1 ) ũ(τ ) − ũ(z2 )
K ũ(z1 ) − K ũ(z2 ) = − dτ
2πi ∂∆ τ − z1 τ − z2
+ (ũ(z1 ) − ũ(z2 ))
Z Z
1 1
= · + · + (ũ(z1 ) − ũ(z2 )).
2πi ∂∆∩B2 (z1 ) 2πi ∂∆\B2 (z1 )
 3.7. ANALYTIC DISCS AND THE EXTENSION OF CR FUNCTIONS 155

Now,
Z
1
· + (ũ(z1 ) − ũ(z2 ))
2πi ∂∆\B2 (z1 )
Z
1
= (ũ(τ ) − ũ(z1 ))
2πi ∂∆\B2 (z1 )
 
1 1
× − dτ
τ − z1 τ − z2
Z
1 (−ũ(z1 ) + ũ(z2 ))
+ dτ + (ũ(z1 ) − ũ(z2 )).
2πi ∂∆\B2 (z1 ) τ − z2
1 1
R
Because of the identity 2πi ∂∆ τ −z2
dτ = 1, the sum of the second and
third terms in the right side is B2 (z1 )∩∂∆ ũ(z1τ)−ũ(z 2)
R
−z2
dτ , and therefore
α
its absolute value is estimated by ||ũ||α |z1 − z2 | . The absolute value of
the first term on the right side can be estimated by
Z
<
||ũ||α |z1 − z2 | 2|θ − θ1 |−2+α dθ ∼ ||ũ||α |z1 − z2 |α .
∂∆\B2 (z1 )

As for the remaining integral over ∂∆ ∩ B2 (z1 ), we have the estimate
Z Z
1 <
· ∼ ||ũ||α (|θ − θ1 |α−1 + |θ − θ2 |α−1 )dθ
2πi ∂∆∩B2 (z1 ) ∂∆∩B2 (z1 )
<
∼ ||ũ||α |z1 − z2 |α .

We are ready for the construction of discs. Let M be defined by the
system of equations
yj0 = hj (x0 , z 00 ), hj (0) = 0, ∂hj (0) = 0.
We fix a point zo = (zo0 , zo00 ) ∈ M and prescribe the “z 00 ” component
of our discs as a holomorphic function w(τ ), τ ∈ ∆, whose boundary
value is small in C k,α -norm. We look for a disc A, with A(1) = zo , in
the form A = zo + (u(τ ) + iv(τ ), w(τ )).
Proposition 3.7.2 (Solution of Bishop’s equation). Let M be de-
fined by yj0 = hj with h ∈ C k+m+2 . Then for any w ∈ C k,α (∂∆, Cn−l )
small and normalized by w(1) = 0 and for any zo close to 0, there is a
unique u ∈ C k,α (∂∆) which solves the equation
(3.7.2) u = −T1 h(u, zo00 + w(τ )) + x0o .
Moreover, if w depends on some parameter η ∈ Rd so that Rd →
C k,α , η 7→ wη is C m , then also (η, x0o , zo00 ) 7→ uη,x0o ,zo00 , Rd × Cn−l →
C k,α (∂∆) is C m .
156 3. REAL/COMPLEX STRUCTURES

Proof. We consider the mapping

(3.7.3)
F : C k,α (∂∆, Rl ) × C k,α (∂∆, Cn−l ) × Rl × Cn−l → C k,α (∂∆, Rl ),
(3.7.4) (u, w, x0o , zo00 ) 7→ u + T1 h(u, zo00 + w) − x0o .
The functional F is C 1 in its action between function spaces. For the
Jacobian ∂u F with respect to u, we have ∂u F : u̇ 7→ u̇ − T1 ∂x hu̇.
In particular, evaluation at (u, w, zo ) = (0, 0, 0) implies that ∂u F is
invertible since ∂x h(0) = 0. Thus, by the implicit function theorem,
there is a unique solution to (3.7.2). If w depends on some parameter
η, then u̇ = ∂η uη is a solution to the linearized equation
(3.7.5) ∂η uη − T1 ∂x h∂η u = 0.
Let us write (3.7.5) as G(∂η u) = 0. Since ∂x h is C k+m+1 , then G is C 1 as
an application on C k,α . In particular, ∂η uη ∈ C k,α and (η, zo ) 7→ uη,zo is
C 1 . Iterated derivations in η yield the C m dependence on the parameter
η.

For a solution u of (3.7.2), we set v = T1 u + yo0 (for yo0 = h(x0o , zo00 ))
and define A = Aη,zo ,w by
A(τ ) = ((u(τ ) + iv(τ )), zo00 + w(τ )), τ ∈ ∂∆.
Since v = h(u, zo00 +w(τ )) on ∂∆, then ∂A ⊂ M and, since u+iv|1 =
zo0 , then A(1) = zo . Thus, A fulfills all requirements.
We can also handle the case in which M is CR non-generic. In this
pr 0 pr
case we choose a projection Cn → Cn such that M → M 0 is a CR
∼
0
diffeomorphism and M 0 is generic in Cn (cf. Section 3.2). Note that
the projection pr is the identity on the “z 00 -component” of a disc. So,
by the first part of the proof, we may find a disc A0 attached to M 0 with
prescribed z 00 -component and through a fixed point of M 0 . By the ap-
proximation theorem, the CR mapping (pr|M )−1 extends holomorphi-
cally inside A0 . Its holomorphic extension provides the supplementary
component of a disc A which projects via pr over A0 .
Note also that the solutions of Bishop’s equation describe all the
small discs attached to M with prescribed z 00 -component through a
fixed point of M . By the uniqueness of the inversion in the functional
spaces, the solution, hence the disc, is unique: thus, through a point
zo ∈ M , the small disc A is uniquely determined by w(τ ), its projection
on TzCo M . This is true even in case M is CR non-generic: A0 is unique
and hence A is also unique because ∂A corresponds in a one-to-one
manner to ∂A0 .
 3.7. ANALYTIC DISCS AND THE EXTENSION OF CR FUNCTIONS 157

We need to introduce now a tricky method due to Tumanov (cf.

[94]) of also prescribing the so-called “additional normal components”.
We consider a manifold of wedge type W with edge M ; this is the
“twisted” version of a plane wedge T W = T M ⊕ Γ where Γ is a cone in
the normal plane TM Cn . We can describe W by “adding” to M defined,
say, by yj0 = hj (x0 , z 00 ), the extra directions of Γ, that is, by extending
h(x0 , z 00 ) to h(x0 , z 00 , t) for t belonging to the “quadrant” Rd+ = {t : ti >
0} and setting
W = {z : yj0 = hj (x0 , z 00 , t), t ∈ Rd+ }.
Then, we can restate the existence result of Proposition 3.7.2: given
zo ∈ M, w(τ ) and t(τ ) small, there is a unique solution u of u =
−T1 h(u, zo00 + w(τ ), t(τ )) + x0o ; setting v = T1 u + yo0 and A(τ ) = ((u(τ ) +
iv(τ )), zo00 + w(τ )), we have A(1) = zo and yj0 = hj (u(τ ), zo00 + w(τ ), t(τ ))
on ∂∆, that is, ∂A ⊂ W . We are ready to prove the following theorem
by Ajrapetyan and Henkin [3]
Theorem 3.7.3 (Edge of the wedge). Let M ⊂ Cn be generic and
let Mj , j = 1, ..., l, be a family of manifolds with boundary M and one
extra direction vj ∈ TM Cn , that is, T Mj = T M ⊕ vj R+ , such that
Span{vj } = TM Cn ; let Γ be the convex conic hull of {vj }. Then, there
is a wedge W with edge M and directional cone Γ, that is, T W |M =
0 0
T M ⊕ Γ, such that any f ∈ CCR (M ) which extends as CCR to each
(Mj ) extends holomorphically to W .
Proof. We can find local coordinates at 0 such that M is defined
by yj0 = hj (x0 , z 00 , 0) and Mj is defined by yj0 = hj (x0 , z 00 , 0, ..., tj , ..., 0),
tj ∈ R+ , with ∂tj hS = vj . We denote by A(τ ) = ((u(τ )+iv(τ )), zo00 +w(τ ))
a disc attached to j Mj and we denote by t(τ ) its “normal component”
defined as the solution of the equation
v(τ ) = h(u(τ ), zo00 + w(τ ), t(τ )).
S
Note that A is attached to j Mj precisely when tj (τ ) = 0 for all
but one of the indices jo and tjo (τ ) ≥ 0. We decompose ∂∆ into a
l
γj of arcs of length 2πl , take χj (τ ) ≥ 0 with supp χj ⊂ γj and
S
union
j=1
1
R
2π
χj dθ = 1, and define
tλ (τ ) = (λ1 χ1 (τ ), ..., λl χl (τ )) for all λ ∈ S l−1 ∩ Rl+ ,
where S l−1 is the (l − 1)-dimensional sphere. Take w(τ ) ≡ 0, zo00 ∈
Cn−l , xo ∈ Rl , set yo0 = h(x0o , zo00 ), and consider Bishop’s equation
u = −T0 h(u, zo00 , tλ ) + x0o , v = T0 u + yo0 .
158 3. REAL/COMPLEX STRUCTURES

Set Ax0o ,zo00 ,λ = (u + iv, zo00 ) where T0 is normalized by the condition

T0 ·|0 = 0. By our choice of the function tλ (τ ) (whose components are all
0 except forSone which is ≥ 0), we have that v = yo0 + T0 u = h(u, zo00 , tλ )
belongs to j Mj . We consider the evaluation mapping of the centers
of the discs
E : (x0o , λ, zo00 ) 7→ Ax0o ,λ,z000 (0) = (z 0 (0), z 00 (0))
= (x0o + iv(0), zo00 ).
For the Jacobian of E at 0 we have
 
id i∂t h ∗
(3.7.6) JE = .
0 0 id
R 2π
In fact, from v(0) = (2π)−1 0 v(eiθ )dθ for v = h, we get
Z 2π
−1
∂x0o v(0) = (2π) ∂x h∂x0o udθ.
0
R 2π
We also have ∂λ v(0) = (2π)−1 0 (∂x h∂λ u+∂t h∂λ tλ )dθ. At λ = 0, x0o =
0, wo = 0, we have ∂x h = 0, ∂λ tλ = id. This implies ∂x0o z 0 (0) =
id, ∂λ z 0 (0) = i∂t h, which completes the proof of (3.7.6). Thus, when
the parameters (x0o , λ, zo00 ) describe Rl × Rl+ × Cn−l , the union of the
centers of the discs covers a wedge W such that T W = T M ⊕ Γ where
Γ is the image of the quadrant Rl+ under the l × l matrix ∂t h.


It is a useful exercise to prove the analogous statement of Theo-

rem 3.7.3 when the convex hull of the vj ’s is a cone of dimension d < l
in TM Cn . In this case, the wedge of extension W is no more open in
Cn but in a (l − d)-codimensional submanifold.
We now treat CR extensions in the Levi normal directions. Let
L ∈ T 1,0 M with L(zo ) = wo be such that 2i1 p[L, L̄] = v 6= 0 in TTCMM .
Let zo = 0 and suppose that M is locally defined by yj0 = hj (x0 , z 00 , z̄ 00 )
where hj (0) = 0, ∂hj (0) = 0, ∂x0I ∂z00J h(0) = 0 and ∂x0I ∂z̄00J h(0) = 0
according to the normalization of Theorem 3.2.7. By the basis ∂rj for
∼ ∗ ∼
rj = −yj + hj , we get an isomorphism TTCMM → TM Cn → Rl which
K◦J
identifies v ' (∂wo ∂w̄o hj )j .
Theorem 3.7.4 (CR extension in Levi directions). Let M ∈
C k , k ≥ 5, and let v be a normal Levi direction. Then CR functions
on M extend to a manifold with boundary M1 which points to the
additional direction v, that is, T M1 = T M ⊕ R+ v.
 3.7. ANALYTIC DISCS AND THE EXTENSION OF CR FUNCTIONS 159

Proof. The argument of the proof consists in finding a family of

discs {A} attached to M with A(1) = z for z describing a neighbor-
hood of zo which are transversal to M . More precisely, they satisfy
pr ∂t A(1) ∼ −v (where pr is the projection T Cn → TM Cn ) with a uni-
form bound for the angle that they form with T M . Hence the union of
their “radii” A|[0,1] , in a neighborhood of zo , forms the desired manifold
M1 . We set wη = ηwo (1 − τ ) and solve, for z ∈ M close to zo , Bishop’s
equation
(3.7.7) u(τ ) = −T1 h(u(τ ), z 00 + wη (τ )) + x0 .
If we denote by u = uη the solution of (3.7.7) and put v = T1 u + y 0 and
Az,η (τ ) = (u(τ )+iv(τ ), z 00 +wη (τ )), then this is a family of analytic discs
attached to M , with Az,η (1) = z and such that (z, η) 7→ Az,η , M ×R →
C 1,α , is C 2 . We now fix z = zo and consider the Taylor development of
∂t A (for τ = teiθ ) up to order 2 with respect to η. Since Aη |η=0 ≡ {zo },
it follows that ∂t Aη |η=0 = 0. Also, from
∂η v = ∂η h(uη , wη )
= ∂x h∂η u + ∂z00 h∂η w + ∂z̄00 h∂η w̄,
evaluation at η = 0 yields ∂η v|η=0 = 0. By differentiating in η the
identity u = −T1 v + x0o and interchanging ∂η with T1 , we get
η2
∂t A = −wo η + ∂t ∂η2 A|η=0 + o2 .
2
We now want to prove that the projection of 2i1 ∂t ∂η2 A|η=0 in the plane
∗
Rl ' TM Cn is −2(∂wo ∂w̄o hj )j . In fact, recalling that our equations are
taken in normal forms (3.2.12), we have
∂η2 v|η=0 = 2(∂wo ∂w̄o hj )j |1 − τ |2 on ∂∆.
Now, |1 − τ |2 |∂∆ = 2 − 2 cos θ = 2Re (1 − τ )|∂∆ . Hence the harmonic
extension of ∂η2 v|η=0 from ∂∆ to ∆ is (∂wo ∂w̄o hj )j 4Re (1 − τ ) and there-
fore
1 2
∂ ∂t Aη |η=0 ∼ −2(∂wo ∂w̄o hj )j .
2i η
Thus, if we fix ηo close to 0 and move z near zo , we have succeeded
in constructing a family of discs transversal to M which point at z
to the direction −∂t Az,ηo ∼ 4i(∂wo ∂w̄o hj )j . We Spick up from the discs
their radii Iz,ηo = Az,ηo |[−1,1] and define M1 := Iz,ηo . Since the Iz,ηo ’s
z∈M
are transversal to M with a uniform estimate from below for their
angle with M at τ = 1, then M1 is a manifold with boundary M and
additional direction v. After we managed to fill such a manifold by
discs attached to M , we can extend CR functions from M to M1 by
160 3. REAL/COMPLEX STRUCTURES

the approximation technique already exploited in the edge of the wedge

theorem. The proof is complete.

TM
We define the “Levi cone” Γ ⊂ T CM
by
1
(3.7.8) Γz = convex hull { [L, L̄]|z modulo T C M }.
1,0
L∈T M 2i
This is a convex cone of TM Cn that we also identify with a cone of
∗
TM Cn ' Rl via J combined with the contraction K with the basis
∗
∂r1 , ..., ∂rl of TM Cn . Note that, by the aid of Theorem 3.7.4, for any v ∈
Γzo we can construct M1 such that T M1 = T M ⊕ R+ v1 . This is a union
of discs which enjoys the property that continuous CR functions on M
extend to M1 . By the edge of the wedge theorem, we can “combine”
all different extensions. Note here that the statement therein can be
generalized to the case of dimΓ < l. This yields the following corollary
to Theorem 3.7.4
Theorem 3.7.5. Let Γ be the Levi cone of M . Continuous CR
functions on M extend to a wedge type domain W with edge M and
directional cone Γ, that is, T W = T M ⊕ Γ.

Suggested research. In the statement of Theorem 3.7.3, take

M totally real and Mj Levi flat. To simplify the situation, even assume
Mj to be C ω so that it can be reduced indeed to R × ... × C × ... × R
with a single occurrence of C in the j-th position. Assume there is a set
of points Ξ = {ξ} of M and a set of CR curves {γξj } in Mj issued from
the points ξ which are excluded from the hypothesis of CR extension,
for instance, the rays ξ1 × ... × (ξj + iR+ ) × ... × ξn in the linear case.
How small must the set Ξ be for the conclusion of Theorem 3.7.3 to
continue to hold?

Notes. The technique of the analytic discs attached to a real

submanifold M ⊂ Cn has a long history which begins with the work
of Lewy [67] who also established the first example of non-solvable
first order PDE. Other early basic work is due to Bishop [24]. Since
then, analytic discs have obtained enormous development. In particu-
lar, in connection with the forced CR extension, we mention Ajrapetyan
and Henkin [3], Bogges and Polking [29] and all the references in Sec-
tions 3.8 and 3.10. We also recall here the related theory of station-
ary/extremal discs developed by Lempert [66] in connection with the
study of the Kobayashi metric (cf. also the recent higher-codimensional
 3.8. ITERATED COMMUTATORS AND FINITE TYPE 161

version due to Tumanov [95]). The edge of the wedge theorem is due to
Ajrapetyan and Henkin [3] and the Levi extension theorem was estab-
lished by Lewy in codimension 1 and next by Ajrapetyan and Henkin
and Boggess and Polking [29], in general codimension. For the presen-
tation of the material that we propose in this section, we followed the
book by Boggess [27], the book by Baouendi, Ebenfelt and Rothschild
[8] and, more closely, the survey by Tumanov [91].

3.8. Iterated commutators, finite type, Bloom-Graham

normal form: deformation of discs for manifolds of
higher type

Summary. We discuss here the CR extension in normal direc-

tions obtained by iterated commutators. We follow the method of [18],
based on analytic discs with Lipschitz boundary: this seems to be the
most appropriate to describe the geometry of the problem. For a sub-
manifold M ⊂ Cn , we introduce the normal form in the sense of Bloom
and Graham that we refine by specifying a CR direction. We show the
relation between the normal form and the properties of the commu-
tators of vector fields in T 1,0 M and T 0,1 M . “Finite type”, in either
sense, is the same. We then present the CR extension theorem which
is obtained by introducing a new space of discs, of “sector” type, and
by proving that they can be approximated by regular discs. But the
approximation for their components which are normal to M occurs in
fact in the C 1 -norm because of the vanishing order of the equations.
The CR extension then goes along the same lines as in Section 3.7.

For a generic manifold M ⊂ Cn , we have seen in Section 3.4 the

relation between the commutators [L, L̄], L ∈ T 1,0 M , and the complex
hessian (∂zh ∂z̄k rj )j=1,...,l for a system of equations r1 = 0, ..., rl = 0
of M . We wish to extend this relation to iterated commutators and
higher-order Taylor terms. For this purpose we need to introduce the
notion of Hörmander’s numbers of M .
The hypersurface case has already been treated in Section 1.10 to
which we refer for notations and terminology. We continue to use the
notation T C M = T M ∩ iT M for the complex tangent bundle to M and
C ⊗R T C M = T 1,0 M ⊕ T 0,1 M for its complexification. As it has already
been seen, C⊗R T M is integrable, that is, closed under Lie bracket, but
C ⊗R T C M is not, in general. We introduce an interpolation between
C ⊗R T C M and C ⊗R T M . We set L1C = C ⊗R T C M and denote by LjC
the distribution of vector spaces spanned by Lie bracket of holomorphic
162 3. REAL/COMPLEX STRUCTURES

and antiholomorphic vector fields of length ≤ j − 1. The commutators

of length 0 are the vector fields themselves; those of length 1 are simple
commutators defining the Levi form and so on. We describe them in
greater detail. We choose an orthonormal system of (1, 0)-vector fields
∂ωk , k = 1, ..., n, with ∂ωk rh = κkh ; thus, ∂ω1 , ..., ∂ωn−l and ∂ω̄1 , ..., ∂ω̄n−l ,
when restricted to M , are a basis of T 1,0 M and T 0,1 M respectively.
With this choice, we have
 

 

j k1 k2 kj
(3.8.1) LC = Span ∂ωk , ∂ω̄k , [∂ωk1 , ∂ω̄k2 ], ..., [∂ωk1 , [∂ωk2 [...∂ωkj ]]] ,

 | {z }

j

where the suffix ks in ∂ωkkss denotes either choice of ∂ωks or ∂ω̄ks . Suppose
that for an integer m1 ≥ 2 we have
(3.8.2) LjC po = Tp1,0
o
M ⊕Tp0,1
o
M, j ≤ m1 −1, Lm 0,1 1,0
C po ⊃ Tpo M ⊕Tpo M.
1

6=
m
LC p1o
Let dim 1
LC po
= l1 ; in this situation we refer to m1 as the first Hörmander
number of M at po and to l1 as its multiplicity. In case LjC = L1C for
any j, we set m1 = +∞ with multiplicity l1 = l. Next, we look for
m2 > m1 such that
(3.8.3) LjC po = LmC po , j < m2 ,
1
Lm 2 m1
C po 6= LC po ,
 Lm2 
C po
and set l2 = dim Lm 1 ; again, m2 is possibly +∞. We continue the
C po
above process. We say M is of finite type when commutators span
the whole of C ⊗R TpCo M . Thus, the above chain ends with a number
mr < +∞ or mr = +∞ according to whether the type is finite or not.
There is a “real” construction of chains of spaces of iterated commu-
tators underlying the complex ones. We recall from Section 1.10 the
identifications Re : T 1,0 M → T C M and Re : T 0,1 MP→ T C M . If ∂ωk =
Pk − iYk , k = 1, ..., n − l, these
X P are defined by Re P : k (ak + ibk )∂ωk 7→
k ak Xk + bk Yk and Re : k (ak − ibk )∂ω̄k 7→ k ak Xk + bk Yk respec-
tively. Let us define LjR by means of iterated commutators of the vector
fields Xk , Yk that we denote by a common symbol Lk , i = 1, ..., 2(n−l):
LjR = Span {Lk , [Lk1 , Lk2 ], [Lk1 , [Lk2 , [..., Lkj ]]]}.
| {z }
j

It is an immediate exercice to verify that LjC = C ⊗ LjR ; thus the

interpolation between C ⊗ T C M and C ⊗ T M by means of the LjC ’s
induces, via Re , a corresponding interpolation between T C M and T M
by means of the LjR ’s.
 3.8. ITERATED COMMUTATORS AND FINITE TYPE 163

We take a system of equations yj = hj , j = 1, . . . , l, for M at zo = 0,

with h(0) = 0 and ∂h(0) = 0, and also set rj = −yj + hj and r = (rj ).
∼
We identify TTCMM → TM Cn with the aid of the complex structure J
∼
and TM Cn → Rl with the aid of the dual basis to ∂rj . We can reason,
instead of over commutators, over the equations r = 0. We fixP numbers
m1 < ... < mr (perhaps mr = +∞) and multiplicities P i l with i li = l.
We take multiindices I1 = (1, . . . , l1 ), . . . , Ir = (( i<r li + 1), . . . , l),
divide the variables as (zI1 , ..., zIr , w), give weight 1 to the w’s and
mi to the xIi ’s, and define the weighted vanishing order for a function
f = f (..., xIi , ..., w) by putting f = O+∞ when mr = +∞ and f
contains some monomial in the xIr ’s, and, otherwise, putting f = Omi
when f (tm1 xI1 , ..., tmi−1 xIi−1 , tw) = O(tmi ). We suppose that associated
to M there are numbers m1 , ...., mr and multiplicities l1 , ..., lr so that
M is described by “normal equations”


 yI1 = PI1 (w) + Om1 +1 ,
y = P (x , w) + Om2 +1 ,

I2 I2 I1
(3.8.4)

 ...

yIr = PIr (xI1 , ..., xIr−1 , w) + Omr +1 ,


where the PIj ’s are weighted homogeneous polynomials of degree mj .

We denote by hImj the right sides of (3.8.4). We point out that this is
not necessarily the normal form in the Bloom-Graham sense. In fact, we
are not assuming the hξ, PIj i’s non-M -pluriharmonic for any j and any
ξ ∈ Rli . To compare the system (3.8.4) of the equations with the chain
Lj , j = 1, ..., r, of the commutators, we need to describe a basis {Xj }
of vector fields for T 1,0 M . We put rIi = −yIi + hIi , r = t (r1 , . . . , rl ),
define an (n − l) × l matrix A = (ajh ) by

A = −t (∂w r) t (∂z r)−1 ,

and set Xj = lh=1 ajh ∂zh + ∂wj . We have

P
(3.8.5)
X
ajh ∂zh (rIi ) + ∂wj (rIi ) = 0 for any i = 1, . . . , r and j = 1, ..., r.
h

Derivation of (3.8.5) yields

(3.8.6)
αi−1
 β
α1
P
∂
 ww̄ ∂xI1 . . . ∂xIi−1 (aj,Ii ) = 0 for |β| + j≤i−1 mj |αj | ≤ mi − 2,
P β α1 αi−1 β α1 αi−1
 h ∂ww̄ ∂xI1 . . . ∂xIi−1 (aj,h ) = −2i∂ww̄ ∂xI1 . . . ∂xIi−1 ∂wj (rIi )
P
for |β| + j≤i−1 mj |αj | ≤ mi − 1.

164 3. REAL/COMPLEX STRUCTURES

Once the equations are ordered as in (3.8.4), we can introduce for

any i ≤ r a diagram
TM ϕ
T CM
→ Rl
(3.8.7) ↓ ↓
TM ψ
Lmi −1
→ Rli +...+lr ,
where ϕ is defined by [v] 7→ ∂rJ (v) and ψ by [v] 7→ ∂(rIi , ..., ∂rIr )J (v)
(where we are performing multiplication rows by columns). The fact
that ψ is well defined (in which case the diagram (3.8.7) is commuta-
tive) is a consequence of the first part of the next proposition. Also, the
link between the normal form of the equations (3.8.4) and the chain of
the Lj ’s is expressed by
Proposition 3.8.1. Let Xo ∈ T 1,0 M satisfy Xo (po ) = ∂w1 , and
denote by Xo either choice Xo or X̄o . Then
(3.8.8)
m −2
J [Xo1 , ..., [Xo i , [Xo , X̄o ]]...](rIj )(po ) = 0 for any j > i and any .
If moreover hξo , hIi i is in the form P + Omi +1 for P = P (w) homoge-
neous of degree mi with mi < +∞, then for some choice of 1 , ..., m1 −2
with k occurrences of Xo and k̄ occurrences of X̄o , we have
(3.8.9)

J [Xo1 , ..., [Xom1 −2 , [Xo , X̄o ]]...]hξo , rIi i(po ) = −2∂wk 1 ∂¯wk̄ 1 (hξo , hIi i)(p0o ).
We refer to [18, Proposition 3.3] for an account of the proof. We
now go on to treat the problem of the CR extension in an additional
normal direction v o ∈ TM Cn .
Theorem 3.8.2. Let M be a generic manifold of Cn and, for a
choice of a complex tangent direction w1 , let (3.8.4) be a normal system
of equations for M̃ = M ∩ (Clz × C1w1 ). Assume that for some j, for
ξ o ∈ Rlj +···+lr , and with the notation P := hξ o , PIj i, we have
(3.8.10)

P = P (w1 ) is a homogeneous non-harmonic polynomial of degree

mj which does not depend on xI1 , ..., xIj−1 ,
P ≥ 0 for w in a sector S of width > π .

1 mj

Then CR functions on M extend to a manifold M1 with boundary M

which points to an additional direction v o satisfying hξ o , v o i > 0.
Remark 3.8.3. (i) The first condition of (3.8.10) is a sort of semi-
rigidity in direction w1 and codirection ξ o . There are large classes of hy-
persurfaces M for which, when (3.8.10) is violated, we can find a “bar-
rier” that is a holomorphic function F such that M \ {0} ⊂ {Im F < 0}
 3.8. ITERATED COMMUTATORS AND FINITE TYPE 165

(cf. [18, §4, Proposition 4.2 and Corollary 4.3]). In particular, for these
M , there always exist CR functions f ∈ CRM which do not extend to
any extra direction v such that hv, d(Im F )i > 0. This shows that the
statement in Theorem 3.8.2 is sharp.
(ii) When j = 1, the first condition of (3.8.10) is always fulfilled.
Also, since PI1 is not harmonic, then it is divisible by |w1 |2 and therefore
it has at most 2m1 − 2 zeroes on the unit circle |w1 | = 1. In particular,
for either of ±P , the second condition of (3.8.10) is satisfied.
(iii) There is a sort of “hierarchy” between the Hörmander numbers
mj whose geometric meaning will be fully clear from the proof of The-
orem 3.8.2. According to it, (3.8.10) gives the control not of v o itself,
but of its projection on Rlj +...+lr . The higher j becomes, the weaker the
control of the extension gets.
(iv) When M is of finite type and semirigid (in all its arguments
w), then our proof provides an alternative proof of the extension of
f to a wedge W . The first conclusion in this direction is due to [10]
where a description of W is also given. We improve this description
by specifying the vanishing order in a precise direction w1 . Also, the
semirigidity in the first condition of (3.8.10) can be released, as well as
the hypothesis that the equations are in canonical form as in (3.8.10).
What is indeed essential is the weighted vanishing order.
Proof of Theorem 3.8.2. (a) Preliminaries on F α spaces. Let
0 < α < 1 and denote by τ = reiθ the variable in the standard disc
∆. Let us recall from [93], [18] and [15] some basics about attaching
analytic discs to M in the subspace F α of the Hölder space C α . This
is the space of real continuous functions σ(θ), θ ∈ [−π, π], which are
C 1,α out of 0 and for which the following norm is finite:
||σ||F α := ||σ||C 0 + ||θσ (1) ||C α .
(Here ·(1) denotes the first derivative.) We remark that for σ ∈ F α
we must have θσ (1) |θ=0 = 0 for otherwise θσ (1) → c 6= 0 which implies
|σ| ≥ log |c|
2
−log |θ| which contradicts the boundedness of σ. This shows
that F α is continuously embedded into C α . It is easy to check that F α
is a Banach algebra. Also, if σi ∈ F αi , i = 1, 2, then σ1 · σ2 ∈ F α1 +α2
for α1 + α2 < 1, resp. σ1 · σ2 ∈ C 1,β with β := (α1 + α2 ) − 1 for
α1 + α2 > 1. In both cases the multiplication is continuous with values
in the respective spaces.
Let T1 denote the Hilbert transform normalized by the condition
T1 (·)(1) = 0; it is easy to see that T1 is a bounded operator in F α as has
already been proved to be true in C α . We come back to our manifold M .
We write coordinates in Cn ' Cl ×Cn−l as (z, w) with z = x+iy, choose
166 3. REAL/COMPLEX STRUCTURES

a distinguished direction, say w1 , and describe M̃ := M ∩ (Clz × C1w1 )

by the system of equations yIi = hIi (x, w1 ) with hIi = PIi + Omi +1 . We
consider in Cn analytic discs A(τ ) = (z(τ ), w(τ )), τ ∈ ∆ (where ∆ is
the standard disc in C), attached to M̃ , that is, satisfying A(∂∆) ⊂ M̃ .
If we prescribe an analytic function w1 (τ ), τ = teiθ ∈ ∆, the so-called
CR component of the disc, and a point p = (z, w1 ) with y = h(x, w1 )
and look for an analytic completion z(τ ) for A(τ ) = (z(τ ), w1 (τ )) with
A(1) = p, we are lead to Bishop’s equation
(3.8.11) u(τ ) = −T1 h(u(τ ) + x, w1 (τ )), τ ∈ ∂∆.
We consider equation (3.8.11) in the spaces F α , F mi α and C 1,β for
which T1 is bounded. We also use the composition properties of hIi for
i ≥ j, with functions in the above classes as stated in [15]. To take
advantage of this composition, we assume mj α > 1 (and, to be sharp,
α(mj − 1) < 1). Here is our main technical tool.
Proposition 3.8.4. Let hIi be of class C mi +3 , and let it satisfy
hIi = Omi . For any  there is δ such that if ||hIi ||C 1,α < δ, ||w1 ||F α <
δ, |x| < δ, then equation (3.8.11) has a unique solution u ∈ F α
with ||ukF α < . Moreover, uI1 ∈ F m1 α , . . . , uIj−1 ∈ F mj−1 α and
(uIj , . . . , uIr ) ∈ C 1,β for β = mj α − 1 and, if w1 depends on some
parameter λ ∈ Rd so that λ 7→ w1λ , Rd → F α is C k for k ≤ mi , then
(λ, x) 7→ (uIi )λ x , Rd+l → C 1,β is also C k . In particular, there exist
mixed derivatives in λ, x and t up to order k and 1, respectively, and
they commute, that is,
0 0
(3.8.12) ∂t ∂λk x u = ∂λk x ∂t u, for any k 0 ≤ k.
Proof. One first solves Bishop’s equation (3.8.11) in the F α -
spaces with the aid of the implicit function theorem. To this end one
considers the mapping F : (λ, x, w1 , u) 7→ u − T1 h(u + x, w1 ), Rd ×
Rl × F α × F α → F α . Then, for the partial Jacobian ∂u F with respect
to u, one has ∂u F : u̇ 7→ u̇ − T1 ∂x hu̇. In particular, if we evaluate at
(λ, x, w1 , u) = (0, 0, 0, 0), then this is invertible since ∂x h|0 = 0. The
differentiability with respect to the parameters in the space F α is also
clear in view of [100, Proposition 11].
We show that the components uIi , i ≥ j, of the solution to Bishop’s
equation, as well as their harmonic conjugates vIi , are in fact in F mi α
for i < j (resp. C 1,β for i ≥ j with β := mj α − 1) and we also prove
differentiability in the parameters with values in this space. The key
point is that, when ϕ = Omi , the composition ϕ((1 − τ )α ) and, in
greater generality, ϕ(w1 ) for w1 ∈ F α with w1 (1) = 0 belong to F mi α
for any i < j (resp. C 1,β for any i ≥ j). We put z(τ ) = u(τ ) + iv(τ ) + z
 3.8. ITERATED COMMUTATORS AND FINITE TYPE 167

with v = T1 u and z = x+ih(x, w) and also write τ = eiθ on ∂∆. We can

check that if zIi (τ ) ∈ F kα for k ≤ mi − 2, then in fact z(τ ) ∈ F (k+1)α .
In fact, v gains regularity at each step because hIi = Omi together
with the fact that if σ(θ) ∈ F kα and σ(0) = 0, then |θ|α σ(θ) ∈ F (k+1)α
because of
|(|θ|α σ(θ))(1) | = ||θ|α−1 σ(θ) + |θ|α σ (1) (θ)|
(3.8.13)
≤ c|θ|(k+1)α−1 .
But the Hilbert transform transfers the F (k+1)α regularity from v to u
and thus z(eiθ ) ∈ F (k+1)α . This completes the proof when i < j. On
the other hand, when i ≥ j, in order to pass from F (mi −1)α to C 1,β ,
we have to prove that (θα u)(1) = θα−1 u + θα−1 (θu(1) ) belongs to C β .
But, in fact, since both u and θu(1) are in C (mi −1)α and are 0 at θ = 0,
then their product with θα−1 is in C β as one can easily check by the
Hardy-Littlewood principle. It follows that (θα u)(1) ∈ C β and hence
θα u ∈ C 1,β . Thus, u(eiθ ) and hence z(eiθ ) itself are in C 1,β . As for
the differentiability on x and on the parameters, it is a variant of [15,
Proposition 15] by the same feedback argument as above. The proof of
Proposition 3.8.4 is complete.

We can think of the family of discs produced by the above statement
as a deformation of the disc A(τ ) ≡ 0 which is, of course, a solution to
Bishop’s equation. By the next statement we show how it is possible
to make infinitesimal deformations of discs which are no longer small.
Proposition 3.8.5. Let hIi ∈ C mi +3 satisfy hIi = Omi , let w̃1 (τ ) ∈
C with w̃1 (1) = 0 be small in F α (not necessarily in C 1,β ), and let
1,β

ũ(τ ) ∈ F α be a solution to Bishop’s equation ũ = −T1 h(ũ, w̃). In

particular, ũIi ∈ C 1,β for any i ≥ j according to Proposition 3.8.4.
Then, for any w1 (τ ) with ||w1 − w̃1 ||C 1,β < δ, |x| < δ, there is a unique
solution to Bishop’s equation u ∈ F α with uIi (τ ) ∈ C 1,β , i ≥ j, and
||uIi − ũIi ||C 1,β < , i ≥ j. Moreover, if λ 7→ (w1 )λ is C k , k ≤ mi , then
(λ, x) 7→ (uIi )λ is also C k .
Proof. In the present situation we define F : Rd × Rl × C 1,β simi-
larly to as in the proof of Proposition 3.8.4 and wish to prove that ∂u F
is still invertible. For this purpose it is enough to show that ∂u hIi (ũ, w̃)
is small in C 1,β -norm. But, in fact, recall that |∂x hIi (u, w)| = O(|w|2 )
and therefore
(1)
||∂x hIi (ũ, w̃)(1) ||C β ≤ c||w̃1 ||C β ||w̃1 ||C β
(1)
≤ c||w̃1 ||F α ||w̃1 ||C β ,
168 3. REAL/COMPLEX STRUCTURES

where the last inequality follows from β < α. Since we are assuming
||w̃1 ||F α small, the conclusion follows.

(b) Construction of a singular disc attached to M with con-
trolled normal component. Let us suppose that (3.8.10) is fulfilled.
It is not restrictive to assume that the sector of the plane Cw1 where
g ≥ 0 contains (1 − τ )α iel+1 , τ ∈ ∆. (Here el+1 is the unit vector of
the w1 -plane.) Let α satisfy αmj > 1, α(mj − 1) < 1. We define, for a
small real parameter η > 0,
(3.8.14) w1 (τ ) = (w1 )η (τ ) := η(1 − τ )α iel+1 .
We attach to M̃ a family of F α -discs A(τ ) = Aη (τ ) whose “w1 -
component” is w1 (τ ). We recall from (a) that for i ≥ j we have that
η 7→ (zIi )η (τ ), R → C 1,β is C mi . We also write zIi (τ ) instead of
(zIi )η (τ ), zIi (τ ) = uIi (τ ) + iT1 uIi (τ ), and finally A(τ ) = (z(τ ), w(τ )).
We note that we have
(3.8.15) ∂ηs vIi |η=0 ≡ 0, ∂ηs uIi |η=0 ≡ 0, s ≤ mi − 1.
This is clear for s = 0, 1. If it is true for any s ≤ mi − 2, then it is also
true for s = mi − 1 because of hIi = Omi by a “feedback” procedure.
If we then take the Taylor expansion of ∂r vIi at η = 0, we get
∂ηmi ∂r vIi |η=0 mi
(3.8.16) ∂r vIi = η + o(η mi ).
mi !
By a feedback argument we can also prove that
(3.8.17) |vIi | ≤ c|w1 |mi , |uIi | ≤ c|w1 |mi .
In fact, in the classes F kα , regularity and vanishing order are coincident;
thus, the equation vIi = hIi gives control from below of the vanishing
order of vIi which is transferred as regularity to uIi through Hilbert
transform and again as vanishing order to vIi . In this way, we can
prove that each vIi and uIi belongs to F mi α (and also to C [mi α],{mi α}
where [mi α], resp. {mi α}, is the integer, resp. fractional, part of mi α).
Recall that if ξo is, say, the unit vector in the l0 := (l1 + ... + lj−1 + 1)-
m
direction, we have hξo , hi ≥ 0 if w1 is in a sector S of width > πj and
if |xIi | ≤ c|w1 |mi . We first observe that this latter condition, |xIi | ≤
c|w1 |mi , is automatically fulfilled by the components xIi = uIi of our
discs A(τ ) because of (3.8.17). We show now that ∂r vl0 < 0. In fact, in
our situation we have
(3.8.18) ¯
hξo , ∂ηmj vl0 i|η=0 ≥ 0 for any τ ∈ ∆.
 3.8. ITERATED COMMUTATORS AND FINITE TYPE 169

Hence (3.8.18) yields, through Hopf’s lemma,

(3.8.19) hξo , ∂r ∂ηmj vl0 i|τ =1, η=0 = −c < 0.
By (3.8.16) we conclude hξo , ∂r vl0 i|τ =1 = −c0 η mj < 0, for any η suffi-
ciently small.
(c) Polynomial approximation of (1 − τ )α in F γ (∆) ¯ for γ < α
and construction of a smooth disc transversal to M and of its
infinitesimal deformation.
We have the Taylor expansion
α(1 − α) 2 α(1 − α)(2 − α) 3
(1 − τ )α = 1 − ατ − τ − τ + ...
2! 3!
(3.8.20) +∞  
X  α  j
=1−  j τ .
 
j=1

We call SN = SN (τ ) the partial sum of the power series (3.8.20) for

1 ≤ j ≤ N . We have the following approximation result for whose proof
we refer to [18, Theorem 2.3]:
(3.8.21) ¯ for any γ < α.
SN (τ ) → (1 − τ )α in F γ (∆)
We now put wN (τ ) = SN (τ ) − SN (1), let uN be the solution in F γ
to Bishop’s equation uN = −T1 h(uN , wN ), and let zN = uN + ivN for
vN = T1 uN . Let u be the solution to u = −T1 h(u, (1 − τ )α ), and set
z = u + iv for v = T1 u. Since
¯
wN (τ ) → (1 − τ )α in F γ (∆)
and since wN (1) ≡ 0 for any N , then (zIj ,...,Ir )N (τ ) → zIj ,...,Ir (τ ) in
0
¯ by Proposition 3.8.4. (Clearly, we are supposing γ close enough
C 1,β (∆)
to α so that β 0 := mj γ − 1 > 0.) In particular, for any  and for large
0
N , the discs AN η = (zN η , wN η ) for wN η := ηwN are in C 1,β and satisfy
(cf. (3.8.19))
(3.8.22) |∂r (vIj ,...,Ir )N η (1) − ∂r (vIj ,...,Ir )| < η mj .
0 0
P explain (3.8.22), we still denote by l any index satisfying l ≥
To
li + 1 and notice that
i≤j−1

||vl0 − vl0 N ||C 1,γ = ||hl0 (uη , ηw) − hl0 (uN η , SN η )||C 1,γ
m
= η mj ||h(mj ) (0)wmj − h(mj ) (0)SN j ||C 1,γ
m
< η mj ||wmj − SN j ||F α0 .
∼

This proves (3.8.22). Also, once N has been fixed so that (3.8.22) holds,
one chooses η such that ||ηwN ||C 1,γ << 1 so that the hypotheses of
170 3. REAL/COMPLEX STRUCTURES

Proposition 3.8.5 are fulfilled. We call à = (z̃, w̃) one of these discs for
the choices of the parameters N and η which have been pointed out,
and use the notation ṽ o := ∂r (ṽIj ,...,Ir ). With a smooth transversal disc
in our hands, we are in the same situation as for the Levi extension dis-
cussed in Section 3.7 (where the discs were smooth from the beginning
since we had α = 1 therein). By a deformation of the disc Ã, we are
ready to construct a manifold M1 with boundary M which gains one
more direction and has the property that CR functions extend from M
to M1 . For this, we consider Bishop’s equation

(3.8.23) u = −T1 h(u + x, w + w̃),

for x ∈ Rl , w ∈ Cn−l with |x| < δ, |w| < δ. According to Propo-

sition 3.8.5, for any  and for suitable δ = δ , there is a unique so-
lution u which satisfies ku − ũkC 1,β0 <  for β 0 < β := kα − 1.
We write p = (x + ih(x, w), w) with v = T1 u and define Ap (τ ) =
p + (u(τ ) + iv(τ ), w̃(τ )). We also write Ip = Ap |[−1,+1] and define
[
(3.8.24) M1 = Ip ([1 − , 1]).
p

In the same way as in Section 3.7 we can see that M1 is a manifold with
boundary M which satisfies T M1 = T M +R+ ṽ o . By the approximation
Theorem 3.6.3 and by the maximum principle, we can extend f from
the boundaries ∂Ap ⊂ M to the “radii” Ip which fill the manifold M1 .


Corollary 3.8.6. Let M be a manifold of finite type and “semi-

rigid” in the sense that it has normal equations of type (3.8.4) with
mr < ∞ and with each PIj non-pluriharmonic and independent of the
xIi ’s (cf. [10]). Then CR functions extend from M to an open wedge
W with edge M .

Proof. We start from j = 1 and consider all related directions of

extension v o given by Theorem 3.8.2 (cf. Remark 3.8.3(ii)). By taking
their convex hull according to the edge of the wedge theorem, we obtain
a first approximate extension open cone Γ1 ⊂ Rl1 . Next, we apply
Theorem 3.8.2 for j = 2 and obtain a perturbation of Γ1 in the form
Γ1 + 2 Γ2 as a second approximate cone of extension. We continue this
process and end up with a full-dimensional approximate extension cone
in the form Γ = Γ1 + 2 Γ2 + ... + r Γr .

 3.9. PARTIAL LIFTS OF ANALYTIC DISCS AND CR CURVES 171

Notes. The CR extension in directions shown by iterated

bracket up to the first Hörmander number goes back to Boggess
and Pitts [28] and to Baouendi and Rothschild [10]. The method of
[10] also applies to further Hörmander numbers but under an addi-
tional hypothesis of “semirigidity” which is required by the technique
based upon Fourier operators and microlocal analysis. Theorem 3.8.2
is stated in Baracco and Zampieri [18]. More recent work by Baracco,
Zaitsev and Zampieri [15] treats the CR extension from a wedge to
a bigger wedge. We refer to the latter paper for a further study of
the “sectors” which develop the “thin discs” originally introduced by
Tumanov in [94].

3.9. Partial lifts of analytic discs and CR curves

Summary. We prepare the geometric background for the prob-

lem that we will shortly anticipate here and that will be completely
solved in Section 3.11. We introduce “partial lifts” of CR curves and
analytic discs. We show that any CR curve γ in M ⊂ Cn is endowed
with partial lifts, CR curves γ ∗0 in (TM∗
Cn )0 (the “orthogonal” projec-
∗0
tion of the conormals) such that π(γ ) = γ. Also, any small analytic
disc A ⊂ Cn with ∂A ⊂ M is endowed with partial lifts A∗0 , holo-
morphic sections of (T ∗ Cn )0 with ∂A∗0 ⊂ (TM ∗
Cn )0 and π(A∗0 ) = A.
On the other hand, the uniqueness of the lift through a fixed conor-
mal, both for curves and discs, is obvious. For this reason, the lifts
∗
provide a “connection”, a morphism of TM Cn between the fibers over
different points of γ and ∂A, and a dual-inverse connection of TM Cn .
Moreover, the approximation of CR curves γ ∗0 by discs A∗0 , induces
an approximation for the respective connections. We end the section
with a geometric application of the lifts: “microlocally pseudoconvex”
manifolds “absorb” tangent analytic discs.

Let f be a CR function on a generic CR manifold M ⊂ Cn which

has CR extension at some point zo ∈ M to some “additional direction”
vo normal to M in Cn , that is, belonging to (TM Cn )zo . This means
that there is a germ of a manifold M 0 with boundary M of dimension
2n − l + 1, that is, dimM 0 = dimM + 1, which “points to” vo in the
sense that Tzo M 0 = Tzo M ⊕ R+ vo , to which f extends as CR. Let γ
be a CR curve in M , that is, a curve which runs in a complex tangent
direction, i.e., T γ ⊂ T C M |γ , let zo = γ(0), and let z1 = γ(1) be another
point of γ. Is there a rule to find a “connection” Φ : (TM Cn )zo →
(TM Cn )z1 , vo 7→ v1 so that if f extends to vo at zo , then it extends to
172 3. REAL/COMPLEX STRUCTURES

∼
v1 at z1 ? In the identifications (TM Cn )z → Rl through the dual basis to
∂r1 (z), ..., ∂rl (z), what is the l×l matrix B associated to Φ, that is, such
that v1 = Bvo ? In practice, there is no natural connection in TM Cn .
∗
But there is in fact in TM Cn and it turns out that its dual-inverse is the
n
connection in TM C we are looking for (cf. Section 3.11). Let us start
the construction of t Φ−1 and t B −1 . Let z and (z; ζ) be coordinates in Cn
and T ∗ Cn , respectively. Let M be described in a neighborhood of 0 as
a graph yj0 = hj (x0 , z 00 ) with hj (0) = 0 and ∂hj (0) = 0 for j = 1, ..., l.
Let prCn−l be the projection in the fibers of T ∗ Cn modulo Cn−l , given
by Cn → Cl , ζ → ζ 0 , and define
∗
(3.9.1) (TM Cn )0 := prCn−l (TM
∗
Cn ).
∗
We denote by T C M := T M ∩ iT M and T C (TM Cn )0 := T (TM∗
Cn )0 ∩
∗ n 0 ∗ n 0
iT (TM C ) the complex tangent bundles to M and (TM C ) , respec-
tively.
Theorem 3.9.1. The natural projection π : T ∗ Cn → Cn induces
an identification
∗ ∼
(3.9.2) T C (TM Cn )0 →0 T C M.
π

Proof. To begin our proof, we need to describe tangent and com-

∗
plex tangent vectors to (TM Cn )0 . We put rj := −yj + hj , r = (rj ), and
∗
observe that ∂rj is a basis for the conormal bundle TM Cn . We fix a
l
base point zo ∈ M and a conormal to ∂r for to ∈ R that we also denote
by ∂ro , write p = (zo ; ∂ro ), and denote by p0 := (zo ; ∂ 0 ro ) its projection
under prCn−l . We have (cf. (3.5.6))
(3.9.3) Tp0 (T ∗ Cn )0 = {(u; t∂ 0 r+∂u ∂ 0 ro + ∂¯u ∂ 0 ro ) : u ∈ Tzo M, t ∈ Rl }.
M

It follows that
(3.9.4)
∗
TpC0 (TM Cn )0 ={(u; t∂ 0 r + ∂u ∂ 0 ro + ∂¯u ∂ 0 ro ) : u ∈ T C M,
1
∂¯u ∂ 0 ro = (−t + is)∂ 0 r for some t and s in Rl }.
2
Now, since ∂ 0 r is non-degenerate, for any u ∈ T C M it is possible to find
a unique λ = 21 (−t + is) ∈ Cl which solves the equation ∂¯u ∂ 0 ro = λ∂ 0 r;
∗
hence, over any vector u ∈ T C M , there is a vector of T C (TM Cn )0 . Also,
as already mentioned, this vector is unique.

∗ ∗
For a CR curve γ in M , we call a CR curve in TM Cn , resp. (TM Cn )0 ,
which projects over γ a “lift”, resp. a “partial lift”, of γ. We denote
by γ ∗ and γ ∗0 a lift and a partial lift of γ, respectively.
 3.9. PARTIAL LIFTS OF ANALYTIC DISCS AND CR CURVES 173

Corollary 3.9.2. Let γ be a CR curve in M and p0 a point in

∗
(TM Cn )0 .
Then, there is an unique partial lift γ ∗0 such that γ ∗0 (1) = p0 .
Let zo and z1 be two points of M connected by a CR curve γ with
γ(0) = zo , γ(1) = z1 .
Associated to γ, there is naturally defined the connection
t
Φ−1 : (TM
∗ ∗
Cn )zo → (TM Cn )z1 ,
∗
defined as follows. For any po ∈ TM Cn with π(po ) = zo we take the
∗0 ∗0
unique “partial lift” γpo such that γpo (0) = po and define
t
Φ−1 : po 7→ p1 where p1 = pr−1 (γ ∗0 )(1).
Cn−l po
∗
(Recall here that prCn−l is one-to-one when restricted to TM Cn .) There
is no doubt that this definition contains a great deal of geometric sense:
the section γ ∗0 of (TM∗
Cn )0 that we have chosen is a distinguished one.
It is the unique CR. Note also that the partial lifts of γ are a vector
space. In fact, the equation λ∂ 0 r = to ∂ū ∂ 0 r which defines the vector ũ
such that π 0 ũ = u is linear in to . Thus, t Φ−1 is also linear. We denote
by t B −1 the matrix associated to t Φ−1 in the basis ∂r1 , ..., ∂rl (with
coordinates of differentials being column vectors). We can summarize
the aforementioned construction by
t Φ−1
∗ ∗
(TM Cn )zo → (TM Cn )z1
(3.9.5) || ||
Rl →
t B −1
Rl .

Guided Exercise 3.9.3. Let 0

us come back to our former transfor-
mation χ : (z; ζ) 7→ (z 0 + P ζ 02 1 , z 00 ; ζ) already introduced in (3.5.9)
( j≤l ζj )2
∗ ∗
and (3.5.10); it interchanges M 99K M̃ , that is, TM Cn → TM̃ Cn . Prove
that χ interchanges CR curves γ 99K γ̃.
Hint: let zo be a point of γ and γ ∗0 = γp∗00o its partial lift through
p0o ∈ (TM
∗
Cn )0 ; define γ̃ ∗0 := χ(γ ∗0 ) and γ̃ := π(γ̃ ∗0 ). Then, γ̃ ∗0 is a CR
∗
curve in (TM̃ Cn )0 and therefore γ̃ a CR curve in M̃ . The situation is
described by the diagram
∗ ∗ χ χ
TM Cn → TM̃ Cn γ ∗0 → γ̃ ∗0
(3.9.6) π↓ ↓π π↓ ↓π
M 99K M̃ , γ 99K γ̃.
Thus, by means of the partial lift, we have explained that χ inter-
changes CR curves.
174 3. REAL/COMPLEX STRUCTURES

We go on to discuss lifts of analytic discs. We denote by A both

an analytic disc and its parametrization over the standard disc ∆. The
disc is said to be “attached” to M when ∂A ⊂ M . A lift of A is an an-
alytic section of T ∗ Cn over A. If A is attached to M , we require its lift
∗
to be attached to TM Cn . A disc endowed with analytic lifts is said to
be “defective” and its “defect” is the dimension of the space of its lifts.
The presence of defective discs is related to the “non-minimality” of
M in the sense of Tumanov that will be discussed in Section 3.10. Let
us recall here that the minimality of M is equivalent to the existence
of small non-constant discs of defect 0 near any point of M . Generi-
cally, CR manifolds are minimal and therefore defective discs are quite
exceptional. However, any disc has “partial lifts” and the dimension of
their space is maximal (equal to the codimension l of M ).
Definition 3.9.4. A∗ 0 is said to be a partial lift of A when A∗ 0 is
a holomorphic section of (T ∗ Cn )0 over A and is attached to (TM
∗
Cn )0 ,
that is, ∂A∗0 ⊂ (TM
∗
Cn )0 .
The family of all partial lifts of a disc can be found by solving a
Riemann-Hilbert problem as described by the following lemma.
Lemma 3.9.5. Let M be defined by a family of “normalized” equa-
tions of class C 2,α
yj = hj (x0 , z 00 ), j = 1, ...l, with hj (0) = 0 and ∂hj (0) = 0 for any j.
We also set rj := −yj + hj . Let A be a disc attached to M which is
small in C 1,α -norm. Then, there is a real l ×l matrix G = G(τ ) of class
C 1,α on ∂∆ close to idl×l such that
G · ∂ 0 r(A) extends holomorphically from ∂∆ to ∆ and G(1) = idl×l .
Proof. We denote by T1 the Hilbert transform normalized by the
condition T (·)(1) = 0. We then have to find G such that G(i · idl×l +
∂x h(A)) extends holomorphically and G(1) = idl×l , which is equivalent
to solving the linear Bishop equation
G = T1 (G · ∂x h(A)) .
This equation can be solved, by the implicit function theorem, in the
spaces with fractional continuity such as C 1,α , because ∂x h(A) is small
(and because T1 is continuous in these spaces).

Remark 3.9.6. The system r1 , ..., rl is normalized in Cn but
∂r1 , ..., ∂rl is not in T ∗ Cn . The matrix G serves as a renormalization:
∗
the lines of G · ∂r(A) are a basis of TM Cn whose “prime” components
extend holomorphically from ∂∆ to ∆.
 3.9. PARTIAL LIFTS OF ANALYTIC DISCS AND CR CURVES 175

∗
Let z1 = A(1) and take ξ1 ∈ (TM Cn )z1 . Then, for any A attached to
M with A(1) = z1 , there is a unique partial lift A∗ 0 with A∗ 0 (1) = ξ10 ;
in fact, the disc
A∗ 0 = A; t ∂ 0 r(A)t Gξ10


(3.9.7)
clearly serves the purpose. Also, if ζ 0 (τ ) = t ∂ 0 rξ 0 (τ ) is another lift, then
if we define g(τ ) := (t ∂ 0 r(A)t G)−1 ζ(τ ), we have
(
g(τ ) ∈ R for any τ ∈ ∂∆,
g(τ ) extends holomorphically.
It follows that g(τ ) ≡ ξ10 (constant) and therefore ζ 0 = t ∂ 0 rt Gξ10 ; thus,
ζ 0 (τ ) is uniquely determined by ζ 0 (1) = ξ10 . Also, through the partial
lifts of A, we can define a connection Φ over ∂A similar to that obtained
over γ. Let zo = A(−1) and z1 = A(1), take ξo0 ∈ (TM ∗
Cn )0zo , and let
A∗0 ∗0 0
(zo ,ξo0 ) be the unique partial lift such that A(zo ,ξo ) (−1) = ξo . We define
∗
a morphism (TM ∗
Cn )0zo → (TM Cn )0z1 by
(3.9.8) t
Φ−1 : ξo0 7→ ξ10 where ξ10 = pr−1 (A∗0
(zo ,ξo ) (1)).

Since A∗0 t 0 t 0
(zo ,ξo ) is in the form ∂ r(A) Gξ1 , then if we want it to pass
through ξo0 at −1, we must have ξ10 = t G−1 (−1)ξo0 . In other words,
we have t B −1 = t G−1 (−1). Again, the definition contains a great deal
of geometric foundation. We have chosen a distinguished section A∗0
∗
of (TM Cn )0 over ∂A: the one uniquely endowed with holomorphic ex-
tension to Ā. Since CR curves γ ∗0 can be approximated by chains of
analytic discs A∗0 , then the matrix B associated to Φ is approximable
by a composition of matrices G(−1)’s.
Guided Exercise 3.9.7 (Twin of Exercise 3.9.3). Let χ be the
transformation of Exercise 3.9.3 which interchanges M 99K M̃ . Prove
that it interchanges attached analytic discs A 99K Ã.
Hint: take a partial lift A∗0 (τ ) = (A, ζ 0 (τ )) for ζ 0 (τ ) = t ∂z0 r(A)· t Gξo ,
and define Ã∗0 = χ(A∗0 ) and à = π(Ã∗0 ). Now, in contrast to full
lifts, we have that the partial lifts A∗0 and Ã∗0 are holomorphic. Thus
à = π(Ã∗0 ) is holomorphic. Also, ∂ à ⊂ M̃ since ∂ Ã∗0 ⊂ (TM ∗
Cn )0 . The
conclusion is illustrated by the diagram
χ
A∗0 → Ã∗0
(3.9.9) π↓ ↓π
A 99K Ã.
We end this section with a geometric application of the construction
of lifts. This part is not essential for what follows.
176 3. REAL/COMPLEX STRUCTURES

For zo = A(1), z̃o = Ã(1), assume that

hξo , Tzo Ai = 0;
it follows that
hξo , Tz̃o Ãi = 0.
In fact, for the 1-form ω, we have ω|Tzo A = 0 if and only if χ∗ ω|Tz̃o à = 0.
This yields the following higher-codimensional version of the fact that
pseudoconvex boundaries “absorb” tangent analytic discs (cf. [12]).
Theorem 3.9.8. Let s−M ≡ 0 in a neighborhood of po = (zo , ξo ),
and let A be a small analytic disc attached to M and orthogonal to ξo
at zo . Then A ⊂ M and, moreover, there is an analytic lift A∗ through
(zo , ξo ).

Proof. By means of the transformation χ, we have transformed

M into a pseudoconvex hypersurface M̃ (because of the symplectic
invariance of s0 ) and A into à such that ∂ à ⊂ M̃ and Tz̃o à ⊂ TzCo M̃ .
We now prove that à ⊂ M̃ . In fact, if we denote by Ω̃ the open half-
space with boundary M̃ and outward conormal ξo , it is clear that à ⊂
Ω̃ because otherwise pseudoconvexity is violated. Next, we reason by
contradiction and suppose that à is not contained in M̃ . If ψ̃ is a
bounded plurisubharmonic exhaustion function for Ω̃ which satisfies
ψ̃ < 0 in Ω̃ and ψ̃ = 0 in M̃ (cf. Diederich and Fornaess [40]), we have
(
ψ̃ ◦ Ã ≤ 0 all over ∂∆,
ψ̃ ◦ Ã < 0 somewhere on ∂∆.
But then, Hopf’s lemma yields

∂τ (ψ̃ ◦ A)1 > 0,

which contradicts Tz̃o à ⊂ Tz̃Co M̃ . Thus, à ⊂ M̃ . Also, by Bedford and

Fornaess [22], M̃ being pseudoconvex, Ã is endowed with a holomorphic
lift. Since its “prime” components coincide with Ã∗ 0 , this lift is Ã∗ .
Thus, A∗ = χ−1 Ã∗ is the holomorphic lift of A contained in TM ∗
Cn .

Remark 3.9.9. According to Remark 3.5.4, we do not need to as-
sume s−M ≡ 0 in order to interchange M 99K M̃ where M̃ is a pseu-
doconvex hypersurface: s− M ≡ const suffices. However, the symplectic
transformation that we need in this case is not as simple as the one
defined by (3.5.9), (3.5.10): we need to take z̃ = z̃(z, ζ) instead of
 3.10. DEFECT OF ANALYTIC DISCS AND WEDGE EXTENSION 177

z̃ = z̃(z, ζ 0 ). For this reason, the “transformed” disc à is not neces-

sarily holomorphic unless A is endowed with a full, not only partial,
lift.

Notes. Lifts of discs are discussed by, among others, Tumanov

[89], [90], Trépreau [87] and Baouendi, Rothschild and Trépreau [12]
(cf. also the book [8] for a survey). A connection related to the lifts
of CR curves is introduced by Trépreau in [87] and Tumanov [90]
where the relation to the connection by discs is explained. Partial lifts
are introduced by Baracco and Zampieri in [16] and [17], which also
present the related connections. Theorem 3.9.8 for the hypersurface
case can be found in Baouendi, Rothschild and Trépreau [12] and in
Zampieri [103] in general codimension.

3.10. Defect of analytic discs—deformation of non-defective

discs—wedge extension from minimal manifolds

Summary. We present here the general theory by [89] of wedge

extendibility of CR functions from a submanifold M ⊂ Cn . This is
founded on the notion of “minimality”, that is, the absence of proper
submanifolds S ⊂ M with the same CR structure. Note that if M has
finite type, then it is minimal: otherwise, the iterated commutators of
(1, 0) and (0, 1) tangential vector fields would be forced to belong to
C ⊗ T S. But the converse is true only in some particular cases such as
if M ∈ C ω : if the type is ∞, then the Nagano leaf is the proper sub-
manifold S, which violates minimality. When M is minimal, there exist
arbitrarily small discs of defect 0. These are endowed with infinitesi-
mal deformations which span all normal directions to M . Combining
all these directions by the edge of the wedge theorem, we get a common
wedge to which CR functions are forced to extend. In contrast to the
theory developed in Section 3.8, there is no explicit description for the
wedge W .

Let M be a CR manifold of Cn ; we have already explained the trick

for making M generic up to a CR diffeomorphism. So, we assume M
generic from the beginning.

Definition 3.10.1. A disc A attached to M is said to be “defec-

tive” when there exists a holomorphic section A∗ which projects via π
∗
over A and which is attached to TM Cn .
178 3. REAL/COMPLEX STRUCTURES

Small defective discs are quite exceptional. For example, if M is

∗
Levi non-degenerate, we saw in Section 3.5 that TM Cn is totally real;
on the other hand, the construction of small discs through Bishop’s
equation reveals that there is a one-to-one correspondence induced by
π : T ∗ Cn → T C TM∗
Cn between discs on T ∗ Cn attached to TM ∗
Cn and
C ∗ n
discs on T TM C . But the latter is {0} so no small disc is defective
unless it shrinks to a point. The section A∗ = (z(τ ), ζ(τ )) of T ∗ Cn
π
∼
with A∗ → A is said to be a “lift” of A. Let M be defined by yj0 =
hj (x0 , z 00 , z̄ 00 ) with h(0) = 0, ∂h(0) = 0, and also write rj = yj − hj .
Associated to a disc A attached to M , there exists a real l × l matrix
G over ∂∆, a solution of the linear Bishop equation
G = T1 (t ∂x h(A)G),
such that t ∂z0 h(A)G extends holomorphically to ∆ (cf. Lemma 3.9.5).
∗
In particular, given ξo ∈ Rl ' (TM Cn )zo , we can find a partial lift
A = (z(τ ), ζ (τ )) with ζ (τ ) = ∂z r(A)t Gξo0 . We have seen that ζ 0 (τ )
∗0 0 0 t 0

is unique under the condition ζ 0 (1) = ξo0 . Thus, A is defective if and

only if ζ 0 (τ ) can be supplemented to a full holomorphic lift ζ(τ ) =
(ζ 0 (τ ), ζ 00 (τ )) which, by uniqueness, is forced to be in the form ζ(τ ) =
t
∂z r(A)t Gξo0 . The lifts form a linear space that we denote by VA ; by the
uniqueness of the (partial) lift through a prescribed conormal point,
the dimension of the space of lifts is constant along ∂A.
Definition 3.10.2. We call dim(VA ) the “defect” of A.
We call any smooth family of discs {Aη } passing through a fixed
point Aη (1) = zo such that A = Aη |η=0 the “deformation” of a disc.
We denote by Ȧ the derivative with respect to η evaluated at η = 0.
The existence of a lift for A is a constraint for its deformations.
Proposition 3.10.3. If A has a holomorphic lift A∗ , then hA∗ , Ȧi ≡
0 (all over A).
Proof. Since Ȧ(τ ) ∈ T MA(τ ) and A∗ (τ ) ∈ (TM
∗
Cn )A(τ ) , then we
have
(
Re hA∗ , Ȧi|∂∆ ≡ 0,
(3.10.1)
hA∗ , Ȧi extends holomorphically from ∂∆ to ∆.
Hence, hA∗ , Ȧi is constant. But since Ȧ(1) = 0, this is in fact 0.

We have just seen that Ȧ ⊂ VA⊥ at any point τ ∈ ∆; in particular,
if we denote by ∂t Ȧ(1) the radial derivative at 1, then ∂t Ȧ(1) ⊂ VA⊥ (1).
Now, denote by FA the space of deformations of A at −1, that is,
 3.10. DEFECT OF ANALYTIC DISCS AND WEDGE EXTENSION 179

FA = {Ȧ(−1)} in T M and denote by GA the equivalence classes mod-

ulo T M of their radial derivatives at +1: GA = {[∂t Ȧ(1)]} in TM Cn .
One can prove the converse of the inclusions which have followed from
Proposition 3.10.3.
Proposition 3.10.4 ([89], [8], [12]). We have
FA = VA⊥ (−1) (in T M ), GA = VA⊥ (1) (in TM Cn ).
(We refer to [91, pp. 137–140] for a (short) proof.) With Propo-
sition 3.10.4 in hand, we can state the main theorem of wedge ex-
tendibility due to Tumanov. We say that M is “minimal” at zo if there
does not exist any proper CR submanifold S ⊂ M through zo with
dimCR S = dimCR M . Note that if M has finite type, then it is minimal.
Otherwise, all spaces LjM made up of iterated commutators, would be
contained in C ⊗ T S and therefore could not reach the full tangent
space C ⊗ T M . When M is real analytic, the converse is also true. In
fact, if M is not finite type, then its Nagano leaf through zo would be
proper and would carry the same CR structure as M , thus violating
its minimality.
Theorem 3.10.5 (Tumanov). Let M be minimal at zo . Then, there
exists an open wedge W of Cn with edge M such that continuous CR
functions on M extend holomorphically to W .
The converse is also true (cf. [11]). Theorem 3.10.5 and its converse
generalize [86]. We remark here that in contrast to the theory developed
in Section 3.8, the directional cone Γ ⊂ TM Cn of W is not explicitly
described from the minimality.
Proof. (Cf. [93].) We start by a complementary result: the exis-
tence of a family of small transversal discs Aj with ∂t Aj = vj implies
the CR extension to the directions vj and hence, by the edge of the
wedge theorem, to their convex hull. More precisely, let {Aj }j=1,...,m be
attached to M with Aj (1) = zo and let ∂t Aj = vj 6= 0 in TM Cn . Define
Γ = convex hull {vj }j=1,...,m .
We claim that continuous CR functions on M extend to a wedge W
with Tzo W = Tzo M ⊕ Γ. As already remarked, it suffices to prove the
result for a single disc A1 with radial direction v1 . Let M be defined
by yj0 = hj (x0 , z 00 , z̄ 00 ) and let w(τ ) be the projection of A1 on TzCo M .
Then, for any z = (z 0 , z 00 ) close to zo we can attach a small disc with
Az (1) = z and whose projection on T C M is w(τ ). By moving z ∈ M ,
for these discs Az we have that ∂t Az (1) is close to ∂t A1 (1) and hence
180 3. REAL/COMPLEX STRUCTURES

through them we extend f to a manifold M1 with additional direction

∂t A1 (1).
We now show that there exists A small with defect 0 through
zo ; hence Γ := convex hull GA has non-empty interior by Proposi-
tion 3.10.4, and the conclusion follows from the above argument. In
fact, let d = lim inf def(A) for A → {zo } with A 3 zo and let Ao be
a small disc with defect d. It is not restrictive to assume Ao (−1) =
zo (for otherwise we just have to reparametrize Ao (τ 2 )). Note that
the discs close to Ao have the same defect d as Ao . It follows that
rk FA ≡ dim(M ) − d for A close to Ao ; thus, the evaluation mapping
A 7→ A(−1) transforms a neighborhood of Ao into a manifold S ⊂ M
of codimension d. On other hand, for any disc A close to Ao we have
TA(−1)
C
M ⊂ Range FA0 = TA(−1) S. Thus, S has the same CR dimension
as M and, M being minimal, this can happen only if S = M . Thus,
d = 0, which completes the proof.


Suggested research. Describe the wedge of holomorphic ex-

tendibility obtained by minimality. This is a hard problem which does
not yet have a complete answer. A partial conclusion is provided by
Section 3.8.

Notes. Minimality was introduced by Tumanov in [89] who ex-

plained its relation to the defect of discs and obtained the ultimate
result on (unspecified) wedge extension (Theorem 3.10.5). The neces-
sity of minimality was proved by Baouendi and Rothschild (reported in
the book [8]). The one-side holomorphic extension from a hypersurface,
that is, the version of Theorem 3.10.5 in codimension 1, was obtained
by Trépreau in [86].

3.11. Propagation of CR extendibility

Summary. In Section 1.7, we saw that analytic discs which lie in

a hypersurface M ⊂ Cn are “propagators” of holomorphic extendibility
from either side of M . We see now that the same is true in higher codi-
mension, for CR functions instead of holomorphic functions from one
side, and for discs that are no longer contained but just attached to M
provided that they are tangent. Also, the connection that we prepared
in Section 3.9 by means of partial lifts of analytic discs and which was
 3.11. PROPAGATION OF CR EXTENDIBILITY 181

represented by a matrix G(−1) describes how the directions of CR ex-

tension evolve along the boundary of the disc. The analytic argument
for proving the result on propagation consists in coupling deformations
of discs, that is, discs in T Cn , with lifts, that is, discs in T ∗ Cn . The
deformation takes the boundary of the disc out of M in the direction of
the initial extension; in particular it can be chosen with null CR com-
ponents. In this way, instead of the full lift, it only interacts with the
partial lift which is holomorphic. At the end point, by Hopf’s lemma,
the deformation points in a new normal direction. The evolution of this
direction is ruled by the dual inverse connection to that of the lift (by
effect of the coupling). The passage from propagation along boundaries
of (tangent) discs to CR curves is obtained by approximation. We end
this section by explaining the relation with the construction of [87] and
[90] where the connection and its dual-inverse are given over a subbun-
∗
dle of TM Cn and a projection of TM Cn , respectively. In this case the
“tangency” of the discs, that is, the orthogonality to the subbundle, is
automatic.

Let γ be a CR curve in M with γ(0) = zo and γ(1) = z 1 . We ap-

proximate γ by a chain of discs {Aν } attached to M . We go backwards.
We start from A1 with A1 (1) = z 1 and whose projection on TzC1 M is
z 100 + γ̇(1)(1 − τ ); we notice that the distance of A1 (−1) to some point
z 2 = γ(t2 ) in γ is o(diam(A1 )). Next, we take A2 with A2 (1) = A1 (−1)
and whose projection in TzC2 M is z 200 + γ̇(t2 )(1 − τ ). In this way we
find a chain ending at Am (−1) with |Am (−1) − zo | < . Associated to
any partial lift of γ ⊂ M to γ ∗0 ⊂ (TM ∗
Cn )0 passing through (z 1 , ξ 10 )
at t = 1, there is a unique chain of lifts of the Aν ’s, {A∗0ν }ν=1,...,m ,
∗
attached to (TM Cn )0 . Moreover, we have
|A∗0m (−1) − γ ∗0 (0)| < c uniformly over ξ 1 ∈ (TM
∗
Cn )z1 .
∗
In fact, according to Section 3.9, for the vectors ũ ∈ T(zC1 ,ξ10 ) (TM Cn )0
such that π 0 ũ = u, we have |ũ| ≤ c|u| for c independent of ξ 1 . In
this way, we approximate the connection obtained by means of the
partial lifts γ ∗0 with the composition of those associated to the lifts
{A∗0ν }ν=1,...,m . In other words,
B is approximated by G1 (−1) · ... · Gm (−1).
Let E ∗ be a subbundle of TM ∗
Cn and denote by (E ∗ )0 its projection sim-
ilarly as to (3.9.1). Note that (E ∗ )0 is a CR bundle which has the same
∗
CR structure as (TM Cn )0 , identified through π 0 with the CR structure
of M . In particular, (E ∗ )0 and (TM ∗
Cn )0 have the same families of CR
curves and of attached analytic discs once they share a common point.
182 3. REAL/COMPLEX STRUCTURES

By this, it is clear that t Φ−1 induces a connection over E ∗ . Define

Σ := (E ∗ )⊥ ,
where ⊥ denotes the orthogonal. Here Σ is a bundle which contains
T M and satisfies
T Cn 

E= .
Σ M
Theorem 3.11.1. Let γ be a CR curve in M . We assume that the
chain of discs whose boundaries {∂Aν }ν approximate γ satisfies
(3.11.1) T Aν (1) ⊂ Σ ∩ iΣ.
Then, if f is a CR function on M which extends at zo to the direction
vo ∈ TM Cn with vo ∈
/ Σ, it follows that f extends at z1 to the direction
n
v1 ∈ TM C , such that
prΣ (v1 ) is close to prΣ (Φvo ),
n
pr n

where prΣ is the projection TTCM →Σ T ΣC .
(Different chains of discs are needed according to the accuracy in
the approximation of the direction Φvo that we desire.)
Prior to the proof, we make some comments and give an example.
The above theorem applies in particular when there is a CR subman-
ifold N properly contained in M with T N + iT N ⊂ T M |N . In fact,
if we approximate a curve γ ∗0 by a family of discs {A∗ 0ν }, attached to
∗
(TM Cn )0 whose CR components are prescribed in (π 0 )−1 (T C N ) (where
∼
π 0 : T C (TM
∗
Cn )0 → T C M ), then in particular we have T Aν |∂Aν ⊂
T N + iT N |∂Aν ⊂ T M |∂Aν , which makes it possible to apply Theo-
rem 3.11.1 for E ∗ = TM ∗
Cn . The above remark applies in particular
when N is a complex curve and γ ⊂ N ; this generalizes to higher-
codimensional submanifolds the Hanges-Treves propagation theorem
of [47] for hypersurfaces. But it holds, in fact, in greater generality:
Example 3.11.2. Let N and M be defined in C3 by
N = {z : z1 = 0, y3 = y22 }, M = {z : y1 = (y3 − y22 )2 }.
We have T N + iT N ⊂ T M |N = R × C2 and thus Theorem 3.11.1
applies.
Proof. Propagation of CR extendibility is reduced to propagation
along the boundaries {∂Aν } of the projections of the discs which pro-
vide the approximation of γ ∗0 . For these discs, which we denote by A,
we are assuming condition (3.11.1) of complex tangency at b := A(1);
we also set a := A(−1). Let M be defined by yj0 = hj (x0 , z 00 , z̄ 00 ) normal-
ized at a = 0, and let its extension M1 be defined, e.g., by introducing
 3.11. PROPAGATION OF CR EXTENDIBILITY 183

a new parameter t ∈ R+ ∪ {0} and extending the domain of h from

Rl ×Cn−l × C̄n−l to Rl ×Cn−l × C̄n−l ×(R+ ∪{0}) with ∂t h(zo ) = vo . We
write A(τ ) = (u(τ ) + iv(τ ), w(τ )) and call w(τ ) the “w components”
of A. We deform A to Aη by prescribing “t-components” tη (τ ) = ηχ(τ )
where η is a small parameter and χ is Cc∞ , χ ≥ 0 and supp(χ) is con-
tained in a neighborhood of 1. We assume w(1) = 0, take wo ∈ Cn−l
and s ∈ Rl small, and find a unique solution u = uη wo s (τ ) in Ck,α (∂∆)
of Bishop’s equation
(3.11.2) u = −T1 (h(u, w + wo , tη ) + s).
Moreover, if we define v = vη wo s by v = T1 (u) + h(s, wo ) and set
0
Aη wo s (τ ) = (u(τ ) + iv(τ ), w(τ ) + wo ), we have that Aη wo s is C k,α and
0
∂τ Aη wo s is C k−1,α with respect to (η, wo , s) for any α0 < α (cf. Section
3.8). We denote by Ȧ the derivative with respect to η at η = 0 and
begin by fixing wo = 0, s = 0 and η = 0. Let G be the real l × l real
matrix on ∂∆ such that G∂z0 r(A) extends holomorphically. We note
that G∂z r(A) · Ȧ also extends holomorphically since the w-components
of A are constant in η and hence those of Ȧ are 0. We have
(3.11.3) GRe ∂z r(A) · Ȧ = G∂t h(u, w + wo )χ.
Recall that ∂t h|a = vo and that χ has support in a neighborhood of a.
We evaluate the derivative ∂τ of the harmonic vector function (3.11.3)
at τ = 1, recall that G(1) = id and conclude by Hopf’s lemma that
Re ∂z r(z1 ) · ∂τ Ȧ|τ =1 ∼ G(−1)vo ,
where ∼ means equality up to a small error. Thus, ∂τ Ȧ ∼ G(−1)vo in
the basis of (TM Cn )b dual to ∂z rj , j = 1, ..., l. Let prM and prΣ be the
projections modulo T M and Σ, respectively. On one hand, prM ∂τ Ȧ|1 ∼
G(−1)vo ; on the other hand, prΣ ∂τ A|1 = 0 (because ∂τ A|1 ∈ Σ ∩ iΣ).
Hence, for the Taylor expansion of ∂τ Aη in η, (3.11.3) yields
prΣ ∂τ Aη |1 ∼ (G(−1)vo + )η + o(η),
where  is infinitesimal with the diameter of supp χ. So far, wo and s
were frozen to 0. We now take a 2n − l − 1 plane Σ in T Mzo = Rl × Cn−l
transversal to ∂θ Aη |1 , solve (3.11.2) for all values of the parameters
(wo , s) ∈ Σ and denote by Aη wo s the corresponding solutions. In par-
ticular ASη wo s |wo =0,s=0 coincides with A = Aη considered above. We set
M2 := Aη wo s and parametrize M2 by D : Σ × ∆ → M2 , (s, wo , τ ) 7→
wo s
Aη wo s (τ ). We can see easily that
rk ∂s wo τ D|s=0, wo =0, τ =1 = 2n − l + 1.
184 3. REAL/COMPLEX STRUCTURES

Thus, M2 is a manifold at z 1 which contains M and “points” to an ex-

tra direction v1 such that prΣ (v1 ) is close to prΣ (G(−1)vo ). From now
on, we can carry out the proof in the same way as in Sections 3.7, 3.8
and 3.10. We approximate a CR function f by polynomials in a neigh-
borhood of M ∪ M1 (localized where approximation
S takes place).
S The
approximating sequence is Cauchy over ∂Aη wo s , hence over Aη wo s ,
by the maximum principle. Thus, it converges over M2 , the union of
the discs, to a CR function, the extension of f .

We want to show now how the theory of the propagation of CR
extendibility by [87] and [90] can be reduced to Theorem 3.11.1. Let
S be the local CR orbit of M at zo , that is, the union of all piecewise
smooth small CR curves issued from zo ; note that we have T C M |S =
T C S. In what follows, we assume that M is non-minimal, that is, it
n|
properly contains S. Set Σ|S := T M |S + iT S and E|S := T MT |CS +iT S
S
;
n |S
remark that we have an identification T MT |CS +iT S
' T TMS|S . Extend Σ|S
and E|S from S to M as smooth bundles. We are then in the situation
described before Theorem 3.9.1: T C (E ∗ )0 may be identified with T C M
via π 0 . But in this quite exceptional case the identification holds in fact
for the whole of E ∗ |S and not only its projection (E ∗ )0 |S as was already
remarked in [87] and [90]:
Proposition 3.11.3. Let p = (z; ζ) ∈ (E ∗ )|S ; then (E ∗ )|S is a CR
manifold such that
∼
TpC E ∗ →0 TzC M.
π

Proof. We have to modify the proof of Theorem 3.9.1. We have a

situation analogous to (3.9.4) but with full vectors in Cn = Tz Cn and
not only their components in Cl . We are thus reduced to prove that
(3.11.4)
for any u ∈ T C M there is a unique λ ∈ Cl s.t. ∂¯u ∂ro = λ∂r.
Since p ∈ E ∗ = TM
∗
Cn ∩ iTS∗ Cn , then from the identity
∂ro = i∂so for suitable so ,
we get, by differentiation along S,
¯ o |T C S = i∂∂s
∂∂r ¯ o |T C S
(3.11.5)
= 0,
where the last equality is a consequence of the fact that ∂∂r ¯ o |T C S is
Hermitian. Thus, for any u ∈ T S, we conclude that ∂¯u ∂ro is a combi-
C

nation of differentials ∂rj , or, in other words, (3.11.4) is satisfied.

 3.11. PROPAGATION OF CR EXTENDIBILITY 185

In the present situation, we can “lift” CR curves and analytic discs

from S to the whole of E ∗ |S and not only (E ∗ )0 |S . Also, if we denote
lifts of analytic discs attached to E ∗ by (A(τ ); ζ(τ )) for τ ∈ ∆, we have
Re hiτ ∂τ A, ζ(τ )i|∂∆ ≡ 0. We denote by Re F the function on the left
side of the above equality. Since F extends holomorphically from ∂∆ to
the whole of ∆, then the equality implies that F is constant and, since
its value at τ = 0 is 0, then in fact F ≡ 0 on the whole of ∆. ¯ From this,
we deduce that any disc A attached to M is orthogonal to E ∗ along its
boundary and therefore T A ⊂ T M + iT S. Condition (3.11.1) is thus
fulfilled by any disc which approximates a CR curve. Theorem 3.11.1
can therefore be applied and it yields

Corollary 3.11.4 (Cf. [87] and [90]). Let M be non-minimal at

n|
zo , let S be the local CR orbit, and set E = T MT |CS +iT
S
S
. For any γ with
z1 = γ(1) and zo = γ(0) we get a partial connection

Φ : Ezo → Ez1 ,

in such a way that if a CR function f extends at zo to the direction vo ,

then it also extends at z1 to v1 such that

prT M +iT S (v1 ) is close to prT M +iT S (Φvo ).

Notes. Propagation of holomorphic extendibility goes back to

the celebrated Hanges-Treves theorem [47], which was obtained by
means of the microlocal theory of PDE’s. A more geometric approach
is contained in Section 1.7 of the present book. Propagation in general
codimension was described by Trépreau [87] and Tumanov [90] and
was later refined by Merker [68]. In particular, Trépreau was the first
to relate, at the points of a subbundle E ∗ ⊂ TM ∗
Cn , the CR structure
∗
of TM Cn to that of M and to show that the wave front set of a CR
function is a union of CR orbits in E ∗ . Tumanov described propagation
of CR extendibility by using explicitly the pairing between lifts and in-
finitesimal deformations of discs. Baracco and Zampieri introduced in
[16] and [17] partial lifts and pointed out that they suffice for propa-
gation. It is worth noticing that lifts to E ∗ are not partial but full. In
general, existence of full lifts imposes the constraint of the vanishing of
the Levi form of M . However, according to [87] and [90], this occurs
when M is not minimal. If S ⊂ M with T C S = T C M , then LM (p) = 0
for any p ∈ E ∗ := TM∗
Cn ∩ J TS∗ Cn .
186 3. REAL/COMPLEX STRUCTURES

3.12. Separate real analyticity

Summary. We end our review by discussing separate real ana-

lyticity in the framework of CR geometry. This is the real counterpart
of the problem encountered in Section 1.5. We first present the problem
in C2 ; the general case follows next. In C2 = R2 + iR2 with coordinates
z = (z1 , z2 ), z = x + iy, we consider a separately real analytic function
f : this means that f is real analytic in x1 (resp. x2 ) when x2 (resp.
x1 ) is kept fixed. We make, indeed, the additional assumption that f
is “CR extendible” in one of the two arguments, say x2 , which means
that for x1 fixed, f extends holomorphically in z2 for y2 <  with 
locally uniform in (x1 , x2 ). We show that f is real analytic.
 ∂ ν f The result
is classical for a function f which satisfies the estimate  xν!h  ≤ −ν o for
 
h = 1, 2 and ν ∈ N. In this case f is CR extendible in both arguments
and, moreover, all of its derivatives ∂xνh f are uniformly bounded; in
particular, (∂x2ν1 + ∂x2ν2 )f are bounded for any ν. But then, the elliptic
regularity of the operators ∂x2ν1 + ∂x2ν2 yields f ∈ C ∞ . In contrast, we
only assume that f is C ω in x1 and CR extendible in x2 ; thus our
 ν 1  ν 1
∂ f ν  ∂x 2 f  ν
assumptions are  xν!1  ≤ cx2 −ν x2 and  ν!  ≤ cx1 −ν
o .
We stress paying attention to the fact that CR extension in at least
1
one argument is needed as the example of f (x1 , x2 ) = x1 x2 exp(− x2 +x 2)
1 2
∞ ω ω
shows. This function is C , separately C , not C : the point is that it
is not CR extendible in either argument. Also notice that no assump-
tion is made on uniform continuity or uniform boundedness for the
different extensions: these come as consequences of the statement. In
any case, uniform boundedness for the different extensions is the main
issue: once this is proved, continuity follows by an argument of normal
family similar to that of Lemma 1.5.1. Holomorphic extension is then
a consequence of the edge of the wedge theorem of Section 3.7.
As for the proof, the main technical tool is a refined version of the
Hartogs lemma which is contained in Theorem 3.12.1: it is a combi-
nation of Fatou’s lemma and the Phragmén-Lindelöf principle. For the
rest, we exploit the power of CR theory. There is an open set in R2
from which the separate holomorphic extensions to y1 6= 0 and y2 6= 0
are bounded and hence continuous. By the edge of the wedge theorem
there is a holomorphic extension to C2 . Next, by the theorem of Hanges
and Treves we propagate this extension along the planes y2 = const.
At this point we apply our generalized Hartogs lemma to the sequence
 3.12. SEPARATE REAL ANALYTICITY 187

of the ν-th roots of the Taylor coefficients of the expansion of f in z2

to fill all missing values of x2 in the domain of f .

We begin by some preliminary result on functions of one complex

variable. Let C be the complex plane with coordinate τ , ∆ = {τ ∈
C : |τ | < 1} the unit disc, ∆+ = {τ ∈ ∆ : Im τ > 0} the upper
half-disc, I = {t ∈ R : |t| < 1} the unit real interval. We recall that
a subharmonic function is a function which is upper semicontinuous
and whose value at any point τ is dominated by the mean value on the
boundary of any disc centered at τ . The following generalization of the
Hartogs lemma plays a crucial role in what follows.
Theorem 3.12.1. Assume given a sequence {ϕν }ν of functions up-
per semicontinuous and uniformly bounded on ∆ ¯ + , subharmonic on ∆+
and which satisfy for some positive constant l

lim sup ϕν (τ ) ≤ l for any τ ∈ ∂∆+ ,
ν
(3.12.1)
lim sup ϕν (t) ≤ 0 for any t ∈ I.
ν

Then, for any η, there is νη such that for all ν ≥ νη

(3.12.2) ϕν (τ ) ≤ lκ Im τ for Im τ ≥ η, |τ | ≤ 1 − β,
where κ is a uniform constant and β is fixed.
Proof. We denote by χ the function on ∂∆+ which is 0 for Im τ =
0 and l for |τ | = 1. Let Pz (ζ), z ∈ ∆+ , ζ ∈ ∂∆+ , be the Poisson kernel
of ∆+ . For any τ ∈ ∆+ we have
Z
dζ
lim sup ϕν (τ ) ≤ Pτ (ζ)lim sup ϕν (ζ)
ν + ν iζ
(3.12.3) Z∂∆
dζ
≤ Pτ (ζ)χ(ζ) ,
∂∆+ iζ
where the first inequality comes from the subharmonicity of each ϕν
and the second from Fatou’s lemma. Let h(τ ) = ∂∆+ Pτ (ζ)χ(ζ) dζ
R
iζ
; h is
2l 1+τ
harmonic, h = π arg 1−τ , and in particular h ≤ κlIm τ for |Re τ | > 1−β.
Since lim sup ϕν is estimated by h over ∂∆+ , then it is also estimated
ν
over ∆+ . Thus
(3.12.4) lim sup ϕν (τ ) ≤ κlIm τ.
ν

To make the estimate (3.12.4) uniform in τ , we need to make another

use of Fatou’s lemma: in the present case inside double integrals. For
τo ∈ ∆+ , we decompose its distance to ∂∆+ as r + δ with δ << r. By
188 3. REAL/COMPLEX STRUCTURES

the submean property, we have for any τ in ∆δ (τo ), the disc of center
τo and radius δ,
Z 2π Z r
2 −1
ϕν (τ ) ≤ (πr ) tϕν (τ + teiθ )dt dθ
0 0
(3.12.5) Z 2π Z r+δ
≤ (π(r + δ)2 )−1 tϕν (τo + teiθ )dt dθ + O(δ),
0 0

where the error can be controlled by O(δ) because of the uni-

form boundedness of the ϕν ’s. On the other hand, by (3.12.4),
lim sup ϕν (τo + teiθ ) ≤ κl(Im τo + r + δ) ≤ 2κlIm τo (because
ν
Im τo > r + δ, the distance to ∂∆+ ). It follows by (3.12.5) that
(3.12.6) lim sup sup ϕν (τ ) ≤ 2κlIm τo + O(δ).
ν τ ∈B(τo ,δ)

We now cover Kη+ = {τ ∈ ∆+ : d(τ, ∂∆+ ) > η, |τ | ≤ 1 − β} by δ-discs

centered at various points τo . For these discs to intersect the set Kη+ ,
we must have Im τo > η − δ. In the δ-discs centered at these points we
δ
get therefore inf Im τ ≥ (1 − η−δ )Im τo . Thus, we can conclude that
τ ∈B(τo ,δ)
for some νη and for a new κ, we have
(3.12.7)
ϕν (τ ) ≤ κlIm τ uniformly for τ ∈ Kη+ and for any ν ≥ νη .

Remark 3.12.2. We get the same conclusion as (3.12.2) if we re-
place the half-disc ∆+ by the strip Uδ+ = I + i(0, δ); in this case Im τ
needs an extra factor δ −1 because of rescaling.
We go on to the discussion in general dimension. Let z = x + iy be
the coordinates in Cn , Ω an open domain of Rn and f a function on Ω.
Definition 3.12.3. We adopt the following terminology.
• We say that f is separately real analytic in xj if its restriction
to the section of Ω with each line parallel to the xj -axis is
real analytic. Thus, when all the other coordinates are fixed,
f extends to |yj | < x .
• We say that f is separately CR extendible to yj if it is sep-
arately real analytic in xj and, moreover, it has holomorphic
extension to |yj | <  with  having positive lower bound locally
uniform in x.
Let Cn = Cn1 × Cn2 with coordinates z = (z 0 , z 00 ) = (x0 + iy 0 , x00 +
iy 00 ). Separate analyticity in the group of variables x00 means analyticity
 3.12. SEPARATE REAL ANALYTICITY 189

in x00 when the other variables are fixed. CR extendibility to all the di-
rections of the y 00 -plane Rn2 means holomorphic extendibility to |y 00 | < 
with  locally uniform in x. CR extendibility to a cone Γ2 ⊂ Rn2 means
that f extends holomorphically for y 00 ∈ Γ2 , |y 00 | <  with  locally
uniform in x; we also assume that, for fixed x0 , f is locally uniformly
continuous up to y 00 = 0. Clearly real analyticity or CR extendibility
in a group of variables x00 is more restrictive than the combination of
real analyticity or CR extendibility in any single x00j . We now state the
main result of the section; the proof will follow next.

Theorem 3.12.4. Let f be a function on Ω which is separately real

analytic in x0 and CR extendible to y 00 . Then f is real analytic.

By iteration, Theorem 3.12.4 implies that if f is separately real

analytic in x1 and CR extendible to each single variable y2 , ...,yn , then
it is in fact real analytic.

Remark 3.12.5. Let Ω = Ω1 ×Ω2 ⊂ Rn = Rn1 ×Rn2 . We denote by

Γ an open convex cone in the y-space Rn and by Γ its truncation by
|y| < ; we denote by Γ2 or Γ2  the analogous open cones in the y 00 -space
Rn2 . The argument of our following proof can be adapted (cf. [18]) to
prove that if f is a continuous function on Ω which is real analytic in
x0 and CR extendible to Γ2 , uniformly continuous, for fixed x0 , up to
y 00 = 0, then it extends holomorphically to a wedge W = Ω + iΓ where
Γ is a conic neighborhood of an arbitrarily large proper subcone of Γ2
and is uniformly continuous up to y = 0.

Proof of Theorem 3.12.4. We can assume, without loss of general-

ity, that n1 = 1. Also, by using iteration, we may assume that n2 = 1.
We use the notation I = (−1, 1), I 2 = I × I, I = (−, ), ∆ = {τ :
|τ | < 1}, ∆ = {τ : |τ | < }, U = {τ ∈ C : |Re τ | < 1, |Im τ | < },
U+ = {τ ∈ U : Im τ > 0}. The statement being local, we can suppose
that f is defined and continuous in I 2 , extends to ∆x2 × {x2 } and
{x1 } × ∆, and is holomorphic with respect to τ ∈ ∆ and τ ∈ ∆x1 ,
respectively. We may assume that f extends to discs of radius slightly
bigger than 1 or x2 .
(a) We first prove that there are an open interval Iδ ⊂ I and an
open strip Uδ = I + iIδ such that f is continuous in Iδ × ∆ and in
Uδ × Iδ . Hence it is a continuous CR function therein. We start from
the proof of the continuity on Iδ × ∆. Let

Kl = {x1 ∈ I : sup |f (x1 , z2 )| ≤ l}.

z2 ∈∆
190 3. REAL/COMPLEX STRUCTURES
S
We note that Kl ⊂ Kl+1 and that l Kl = I since, for each x1 ,
sup |f (x1 , z2 )| < +∞. We claim that
z2 ∈∆
(
Kl is closed,
(3.12.8)
f is continuous on Kl × ∆.
In fact, let xν1 → xo1 with xν1 ∈ Kl and use the notation Fν (z2 ) :=
f (xν1 , z2 ) − f (xo1 , z2 ). The sequence {Fν }ν is equicontinuous in a neigh-
borhood of ∆. In fact, remember that f was supposed to be holo-
morphic for |z2 | slightly bigger than 1; so the conclusion follows from
the uniform bound |Fν | ≤ 2l in addition to Cauchy’s inequalities. We
claim that Fν → 0. Otherwise, by the equicontinuity and on account
of Arzela’s theorem, there is a subsequence {Fνk }k which converges to
a limit F 6= 0. But this limit is holomorphic in ∆ and 0 in I, a con-
tradiction.
S This proves the claim and thus (3.12.8) follows. The union
l K l being the whole of I, it follows from Baire’s theorem that the
sets Kl must contain an open interval for large l (cf. also [50] and
[80] for a similar use of Baire’s theorem). Also, such an interval can
be found centered at points xo1 arbitrarily close to 0. We can even as-
sume that xo1 = 0 (by the propagation in z1 -direction contained in the
point (b) which follows). Thus, we can assume that f extends to Iδ × ∆
as a continuous function, holomorphic in z2 , that is, a continuous CR
function.
We now go on to prove that f is a continuous CR function on Uδ ×Iδ .
For this purpose, we define Jl = {x2 : f (·, x2 ) extends to |y1 | < 1l and
|f (·, x2 )| ≤ l}. If xν2 → xo2 with xν2 ∈ Jl , then by boundedness, there is
a subsequence which converges to a holomorphic function on U 1 ; this
l
must be f (·, xo2 ). As before, we have |f (·, xo2 )| ≤ l and hence Jl is closed
and f |U 1 ×Jl is continuous. By Baire’s theorem we conclude again that
l
for large l, the set Jl contains an open interval that we can suppose to
be centered at xo2 = 0 (because otherwise we just need to center the
following series (3.12.9) at xo2 instead of 0). Thus, f is continuous on
Uδ × Iδ for small δ. This concludes the proof of the claim.
(b) At this point we apply the Ajrapetyan-Henkin edge of the wedge
Theorem 3.7.3 and conclude that f extends holomorphically to a do-
main of type ∆δ × ∆δ . We notice now that
• f is continuous and CR on Uδ × Iδ ,
• Uδ × Iδ is foliated by the complex leaves Σx2 := Uδ × {x2 } for
x2 ∈ Iδ ,
• f extends to ∆δ × ∆δ ,
• each leaf Σx2 intersects ∆δ × Iδ .
 3.12. SEPARATE REAL ANALYTICITY 191

But then, the propagation of the holomorphic extendibility of CR func-

tions along complex leaves yields the extension of f to an open domain
Uδ × ∆δ of C2 , for small δ. We now make a short digression which is
crucial for adapting the proof of the present theorem to the statement
of Remark 3.12.5 about the boundary values from wedges. Notice that
∆δ × ∆δ is swept by discs with boundary in the region where f is
bounded and Uδ × ∆δ is swept by discs with boundary in the union of
Uδ × Iδ and ∆δ × ∆δ where f is also bounded. Hence, by the maximum
modulus principle, f is bounded in Uδ × ∆δ . Since by such an intervals
Iδ , where f has (different) bounds, we cover an open dense set D ⊂ I,
we conclude that f is continuous up to y = 0 over I × D. By this,
in the case of wedge extendibility for y belonging to cones, and for
f ∈ C 0 (I × I), we can see that the “extension” is uniformly continuous
up to y = 0 with limit f (cf. [18]). We come back to the main subject.
(c) We consider now the Taylor series of f with respect to z2 cen-
tered at z2 = 0, i.e.,
X ∂zν f (z1 , 0)
(3.12.9) 2
z2ν .
ν
ν!
This represents a holomorphic function on Uδ × ∆δ ; when x1 is fixed
in I, this extends holomorphically for z2 ∈ ∆. We write ϕν (z1 ) :=
1 |∂zν2 f (z1 ,0)|
ν
log ν!
and note that these are subharmonic functions of z1 . We
have
(3.12.10) lim sup sup ϕν (z1 ) ≤ −log δ,
ν z1 ∈Uδ

and hence, in particular, the ϕν ’s are uniformly bounded in Uδ . We also

have
(3.12.11) lim sup ϕν (x1 ) ≤ 0 for x1 ∈ I.
ν

(In (3.12.10) and (3.12.11) the domains ∆ and I should be shrunk in

an arbitrarily small fashion.) But then Theorem 3.12.1 applies to the
sequence of the ϕν ’s over the pair of half-strips x1 + Uδ± . It implies that
for any η > 0, for suitable νη and for any ν ≥ νη , we have
ϕν (z1 ) ≤ δ −1 κlog δ|y1 | when x1 ∈ I and y1 ≥ η.
Here the factor δ −1 comes from covering Uδ by δ-discs. In other words,
we have  ν 1
∂z f (z1 , 0) ν −1
2
≤ δ −κ|y1 |δ .
ν!
−1
Let  = α δ satisfy δ −κδ < (1 − α)−1 ; then, the power series
(3.12.9) converges normally for (z1 , z2 ) ∈ U̇ × ∆1−α , where U̇ stands
192 3. REAL/COMPLEX STRUCTURES

for I + i(I \ {0}), and its sum is therefore a holomorphic function on

U̇ × ∆1−α . Since we already know that this function extends to y1 = 0
when z2 ∈ ∆δ , then it is in fact holomorphic on U × ∆1−α . Note that
if the initial strip of analyticity is not centered at z2 = 0 but, say, at
z2 = z2o , then we have to replace 0 by z2o in the expansion (3.12.9). The
proof goes through without change and yields the analyticity of f in
∆ × ∆1−|z2o |−α . The conclusion follows since |z2o | and α are arbitrarily
small. The proof is complete.


Suggested research. Let f = f (x1 , x2 ) be a continuous func-

tion in I × I which is separately C ω in x1 and extends separately CR
in y2 but only for positive values y2 > 0. According to Remark 3.12.5 a
refinement of the argument of the proof of Theorem 3.12.4 shows that
f extends to a wedge defined by y2 > c|y1 | for suitable c. Does the
wedge of maximal extension have the half-space y2 > 0 as asymptotic
at 0? This is likely to be true.
However, the general challenging problem is to get the extension
to a “quadrant”. Thus, let f extend to 0 ≤ y1 <  for fixed x2 and
to 0 ≤ y2 <  for fixed x1 . Does f extend to a wedge W such that
T W |I×I coincides with the quadrant y1 ≥ 0, y2 ≥ 0? Of course, one
can then also try the cases where one of the two lengths of extension 
is non-uniform.
Recall once more that if the extensions are continuous or bounded,
which turns out, by normal family argument, to be the same, the answer
is affirmative because of the edge of the wedge theorem

Notes. The problem of separate analyticity goes back, in the

complex frame, to Hartogs [48]. As for the real setting, holomorphic
extension to the quadrant y1 > 0, y2 > 0 for separately CR extendible
functions, with uniformly continuous extension, was first obtained by
Malgrange and Zerner. A more refined version of this result was next
established by Kashiwara under the name of “local Bochner tube the-
orem” (cf. the presentation by Komatsu in [62]). The real analyticity
of functions which are separately CR extendible in all arguments and
which have, moreover, all derivatives ∂xνh f, h = 1, 2, ν ∈ N bounded,
is due to Browder [30] and Lelong [65]. The same result, without the
assumption of bounded derivatives, is due to Siciak [80]. For the proof
of Theorem 3.12.4 we refer to Baracco and Zampieri [?] whose presen-
tation we follow closely in this section.
 Bibliography

¯
[1] H. Ahn—Global boundary regularity of the ∂-equation on q-pseudoconvex
domains, Math. Reich. 280 (2007), 343–350
[2] H. Ahn, L. Baracco and G. Zampieri—Subelliptic estimates and reg-
ularity of ∂¯ at the boundary of a Q-pseudoconvex domain of finite type,
Math. Reich. (2009)
[3] R. Ajrapetyan and G. Henkin—Analytic continuation of CR functions
through the “edge of the wedge”, Dokl. Acad. Nauk. SSSR 259 (1981),
777–781
[4] R. Ajrapetyan and G. Henkin—Integral representation of differential
forms on Cauchy-Riemann manifolds and the theory of CR functions, II,
Math. USSR-Sb 55 (1986), 91–111
[5] A. Andreotti, G. Fredricks and M. Nacinovich—On the absence
of Poincaré lemma in tangential Cauchy-Riemann complexes, Ann. S.N.S.
Pisa 27 (1981), 365–404
[6] A. Andreotti and C.D. Hill—Complex characteristic coordinates and
tangential Cauchy-Riemann equations, Ann. S.N.S. Pisa, 26 (1972), 299–
324
[7] S. Baouendi, C.H. Chang and and F. Treves—Microlocal analyticity
and extension of CR functions, J. Diff. Geom. 18(1983), 331–391
[8] M.S. Baouendi, P. Ebenfelt and L.P. Rothschild—Real Subman-
ifolds in Complex Space and Their Mappings, Princeton Math. Series,
Princeton Univ. Press (1999)
[9] M.S. Baouendi, H. Jacobowitz and F. Treves—On the analyticity of
CR mappings, Ann. Math. 122 (1985), 365–400
[10] M.S. Baouendi and L.P. Rothschild—Normal forms for generic mani-
folds and holomorphic extension of CR functions, J. Diff. Geom. 25 (1987),
431–467
[11] M.S. Baouendi and L.P. Rothschild—Cauchy Riemann functions on
manifolds of higher codimension, Invent. Math. 101 (1990), 45–56
[12] M.S. Baouendi, L.P. Rothschild and J.M. Trépreau—On the geom-
etry of analytic discs attached to real manifolds, J. Diff. Geometry (1994),
379–405
[13] M.S. Baouendi and F. Treves—A property of the functions and distri-
butions annihilated by a locally integrable system of complex vector fields.
Ann. Math. (2) 114 (1981), 387–421
[14] M.S. Baouendi and F. Treves—About holomorphic extension of CR
functions on real hypersurfaces in complex space, Duke Math. J. 51 (1984),
77–107

193
194 BIBLIOGRAPHY

[15] L. Baracco, D. Zaitsev and G. Zampieri—Rays condition and exten-

sion of CR functions from manifolds of higher type, J. Anal. Math. 101
(2007), 95–121
[16] L. Baracco and G. Zampieri—Analytic discs and extension of CR func-
tions, Compositio Math. 127 (2001), 289–295
[17] L. Baracco and G. Zampieri—Propagation of CR extendibility along
complex tangent directions, Compl. Var. Theory Appl. 50 (12) (2005), 967–
975
[18] L. Baracco and G. Zampieri—CR extension from manifolds of finite
type, Canad. J. Math. 60 (6) (2008), 1219–1239
[19] L. Baracco and G. Zampieri—Separate real analyticity and CR ex-
tendibility, Forum Math. 21 (3) (2009), 519–531
[20] C.A. Berenstein and R. Gay—Complex variables: an introduction,
Graduate Texts in Math., Springer Verlag, N.Y., 125 (1991)
[21] S. Berhanu, P.D. Cordaro and J. Hounie—Introduction to involutive
structures, Cambridge Univ. Press (2007)
[22] E. Bedford, J. Fornaess—Complex manifolds in pseudoconvex bound-
aries, Duke Math. J. 48, n. 1 (1981), 279–288
[23] S. Bell and E. Ligocka—A simplification and extension of Fefferman’s
theorem on biholomorphic mapings, Invent. Math. 57 (1980), 283–289
[24] E. Bishop—Differentiable manifolds in complex Euclidean space, Duke
Math. J. 32 (1965), 1–22
[25] T. Bloom and I. Graham—On type conditions for generic real subman-
ifolds of Cn , Invent. Math. 40 (1977), 217–243
[26] S. Bochner and W.T. Martin—Several complex variables, Princeton
(1948)
[27] A. Boggess—CR manifolds and the tangential Cauchy-Riemann complex,
CRC Press, Boca Raton, FL (1991)
[28] A. Boggess and J. Pitts—CR extension near a point of higher type,
Duke Math. J. 52 (1985), 67–102
[29] A. Boggess and J.C. Polking—Holomorphic extension of CR functions,
Duke Math. J. 49 (1982), 757–784
[30] F.E. Browder—Real analytic functions on product spaces and separate
analyticity, Canadian J. of Math. 13 (1961), 650–656
[31] H. Cartan and P. Thullen—Regularitäts und Konvergenzbereiche,
Math. Ann. 106 (1932), 617–647
¯
[32] D. Catlin—Necessary conditions for subellipticity of the ∂-Neumann prob-
lem, Ann. of Math. 117 (1983), 147–171
¯
[33] D. Catlin—Subelliptic estimates for the ∂-Neumann problem on pseudo-
convex domains, Ann. of Math. 126, n. 1 (1987), 131–191
[34] S.S. Chern—Complex manifolds without potential theory, 2nd ed.,
Springer-Verlag, New York (1979)
[35] S.S. Chern and J.K. Moser—Real hypersurfaces in complex manifolds,
Acta. Math. 133 (1974), 219–271
[36] S.C. Chen and M.C. Shaw—Partial differential equations in several
complex variables, Studies in Adv. Math., AMS and Int. Press, 19 (2001)
[37] J. D’Angelo—Real hypersurfaces, order of contact, and applications, Ann.
of Math. 115 (1982), 615–637
 BIBLIOGRAPHY 195

[38] J.E. Fornaess—The disc method, Math. Z. 227, n. 4 (1998), 705–709

[39] J.E. Fornaess and C. Rea—Local holomorphic extendability of CR func-
tions, Ann. Sci. SNS Pisa 12 (1985), 491–502
[40] K. Diederich and J. Fornaess—Pseudoconvex domains: bounded
strictly plurisubharmonic exhaustion functions, Invent. Math. 39 (1977),
129–141
[41] P. Ebenfelt—Uniformly Levi degenerate CR manifolds; the 5 dimensional
case, Duke Math. J. 110 (2001), no. 1, 37–80
[42] G.B. Folland and J.J. Kohn—The Neumann problem for the Cauchy-
Riemann complex, Ann. Math. Studies, Princeton Univ. Press, Princeton,
N.J., 75 (1972)
[43] L. Garding and L. Hörmander—Strongly subharmonic functions,
Math. Scand. 15 (1964), 93–96
[44] P.R. Garabedian and D.C. Spencer—Complex boundary value prob-
lems, Trans. AMS 73 (1952), 223–242
[45] R.C. Gunning—Introduction to holomorphic functions of several va-
iables, Vols. I, II, II, Wadsworth & Brooks/Cole Adv. Books & Software,
Pacific Grove, CA, 1990
[46] R.C. Gunning and H. Rossi—Analytic functions of several complex
variables, Prentice-Hall, Englewood Cliffs, NJ, 1965
[47] J. Hanges and F. Treves—Propagation of holomorphic extendability of
CR functions, Math. Ann. 263 (1983), no. 2, 157–177
[48] F. Hartogs—Zur Theorie der analytischen Funktionen mehererer
Veränderlichen, Math. Ann. 62 (1906), 1–88
[49] L. Hörmander—L2 estimates and existence theorems for the ∂¯ operator,
Acta Math. 113 (1965), 89-152
[50] L. Hörmander—An intoduction to complex analysis in several complex
variables, Van Nostrand, Princeton, N.J. (1973)
[51] L. Hörmander—The analysis of linear partial differential operators I, II,
III and IV, Grundlehren, Springer-Verlag, 256 (1984)
[52] H. Jacobowitz—An introduction to CR structures, Math. Surveys and
Monographs AMS, 32, Providence, RI, 1990
[53] M. Kashiwara and P. Schapira—Sheaves on manifolds, Grundlehren
mat., Springer-Verlag, 292 (1990)
[54] C. Kenig and S. Webster—On the hull of holomorphy of an n-manifold
in Cn , Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 11-2 (1984), 261–280
[55] J.J. Kohn—Harmonic integrals on strongly pseudoconvex manifolds, I,
Ann. Math. 78 (1963), 112–148; II, Ann. Math. 79 (1964), 450–472
[56] J.J. Kohn—Boundary behavior of ∂¯ on weakly pseudo-convex manifolds
of dimension two, J. Diff. Geom. 6 (1972), 523–542
[57] J.J. Kohn—Global regularity for ∂¯ on weakly pseudo-convex manifolds,
Transactions of the A.M.S. 181 (1973), 273–292
[58] J.J. Kohn—Methods of partial differential equations in complex analysis,
Proceedings of Symposia in Pure Mathematics 30 (1977), 215–237
¯
[59] J.J. Kohn—Subellipticity of the ∂-Neumann problem on pseudoconvex
domains: sufficient conditions, Acta Math. 142 (1979), 79–122
[60] J.J. Kohn—Hypoellipticity and loss of derivatives, Annals of Math. 162
(2005), 943–286
196 BIBLIOGRAPHY

[61] J.J. Kohn and L. Nirenberg—Non-coercive boundary value problems,

Comm. Pure Appl. Math. 18 (1965), 443–492
[62] H.Komatsu—A local version of the local Bochner tube theorem, J. Fac.
Sci. Univ. Tokyo (1972), 201–214
[63] S.G. Krantz—Function theory of several complex variables, J. Wiley &
Sons Inc., New York, 1982
[64] S.G. Krantz—Partial differential equations and complex analysis,
Wadsworth, Belmond, CA, 1992
[65] L. Lelong—Fonctions plurisousharmoniques et fonctions analytiques de
variables reélles, Ann. Inst. Fourier 11 (1961), 515–562
[66] L. Lempert—La métrique de Kobayashi et la représentation des domaines
sur la boule, Bull. SMF 109 (1981), 427–474
[67] H. Lewy—On the local character of the solutions of an atypical linear
differential equation in three variables and a related theorem for regular
functions of two complex variables, Ann. of Math. 64 (1956), 514–522
[68] J. Merker—Global minimality of generic manifolds and holomorphic ex-
tendibility of CR functions, Int. Math. Res. Notices 1 (1997), 21–56
[69] C.B. Morrey—The analysis embedding of abstract real analytic mani-
folds, Ann. of Math. 40 (1958), 62–70
[70] T. Nagano—Linear differential systems with singularities and an applica-
tion to transitive Lie algebras, J. Math. Soc. Japan 18 (1966), 398–404
[71] R. Narasimhan—Several complex variables, Chicago Lectures in Mathe-
matics, The University of Chicago Press, Chicago, Ill., London (1971)
[72] A. Newlander and L. Nirenberg—Complex coordinates in almost com-
plex manifolds, Ann. Math. 64 (1957), 391–404
[73] K. Oka—Sur les fonctions de plusieurs variables, J. Sci. Hiroshima Univ.,
(I) 6 (1936), 245–255, (II) 7 (1937), 115–130, (III) 9 (1939), 7–19
[74] R.M. Range—Holomorphic functions and integral representations in sev-
eral complex variables, Graduate Texts in Math., Springer-Verlag, N.Y.,
108 (1986)
[75] L.P. Rothschild and E.M. Stein—Hypoelliptic differential operators
and nilpotent groups, Acta Math. 137, no. 3-4 (1976), 247–320
[76] H. Rossi—Differentiable manifolds in complex euclidean space, Proc. Int.
Congr. Math., Moscow, Izdat. Mir (1966), 512–516
[77] W. Rudin—Real and complex analysis, McGraw-Hill (1987)
[78] M. Sato, M. Kashiwara and T. Kawai—Hyperfunctions and pseudo-
differential operators, Springer Lect. Notes in Math. 287 (1973), 265–529
[79] B. Shiffman—Separately meromorphic functions and separately holomor-
phic mappings, Proc. Symp. Pure Math. 52 (1991), 191–198
[80] J. Siciak—Analyticity and separate analyticity of functions defined on
lower dimensional subsets of Cn , Zeszyty Nauk. U.J. 13 (1969), 53–70
[81] J. Siciak—Separately analytic functions and envelopes of holomorphy of
some lower dimensional subsets of Cn , Ann. Pol. Math. 22 (1969), 145–171
[82] L. Stout—A domain whose envelope of holomorphy is not a domain, Ann.
Polon. Math. 89, n. 2 (2006), 197–201
[83] L. Stout—Polynomial convexity, Progress in Mathematics, Birkhäuser
Boston, Inc., Boston, MA, 261 (2007)
 BIBLIOGRAPHY 197

[84] H.J. Sussmann—Orbits of families of vector fields and integrability of

distributions, Trans AMS 180 (1973), 171–188
[85] G. Tomassini—Tracce delle funzioni olomorfe sulle sottovarietà analitiche
reali d’una varietà complessa, Ann. S.N.S. Pisa 20 (1966), 31–44
[86] J.M. Trépreau—Sur le prolongement holomorphe des fonctions CR
definies sur une hypersurface réelle de classe C 2 dans Cn , Invent. Math.
83 (1986), 583–592
[87] J.M. Trépreau—Sur la propagation des singularités dans les varietés CR,
Bull. Soc. Math. de France 118 (1990), 403–450
[88] F. Treves—Hypo-analytic structures, Local theory, Princeton Univ. Press,
Princeton (1992)
[89] A. Tumanov—Extending CR functions on a manifold of finite type over
a wedge, Mat. Sb. 136 (1988), 129–140
[90] A. Tumanov—Connection and propagation of analyticity of CR functions,
Duke Math. J. 71, n. 1 (1994), 1–24
[91] A. Tumanov—Analytic discs and the extendibility of CR functions,
Springer Lect. Notes in Math. 1684, Berlin (1998)
[92] A. Tumanov—Extending CR functions from manifolds with boundaries,
Math. Res. Lett. 2 (5) (1995), 629–642
[93] A. Tumanov—Analytic discs and the extendibility of CR functions, In-
tegral geometry, Radon transforms and complex analysis (Venice, 1996)
Lecture Notes in Math., Springer, Berlin, 1684 (1998), 123–141
[94] A. Tumanov—Thin discs and a Morera theorem for CR functions, Math.
Z. 226 (2) (1997), 327–334
[95] A. Tumanov—Extremal discs and the regularity of CR mappings in
higher codimension, American J. of Math. 123 (2001), 445–473
[96] V.S. Vladimirov—Metody teorii funktsii mnogikh kompleknykh pere-
mennykh, Nauca, Moscow (1964)
[97] S. Webster—A new proof of the Newlander-Nirenberg theorem, Math. Z.
201 (1989), 303–316
[98] R.O. Wells Jr.—Differential analysis on complex manifolds, Grad. Texts
in Math. 65, Springer Verlag, New York, Berlin, 1980
[99] D. Zaitsev and G. Zampieri—Extension of CR functions on wedges,
Math. Ann. 326 (2003), no. 4, 691–703
[100] D. Zaitsev and G. Zampieri— Extension of CR-functions into weighted
wedges through families of nonsmooth analytic discs, Trans. Amer. Math.
Soc. 356 (2004), no. 4, 1443–1462
[101] G. Zampieri—q-Pseudoconvexity and regularity at the boundary for so-
¯
lutions of the ∂-problem, Compositio Math. 121 (2000), 155–162
[102] G. Zampieri—q-Pseudoconvex hypersurfaces through higher codimen-
sional submanifolds of Cn , J. Reine Angew. Math. 544 (2002), 83–90
[103] G. Zampieri—Analytic discs in pseudoconvex submanifolds of CN of
higher codimension, Manuscripta Math. 116 (4) (2005), 397–400