Beijing Lectures in Harmonic Analysis-Stein-0691084181

Annals of Mathematics Studies
Number 112
BEIJING LECTURES IN
HARMONIC ANALYSIS
EDITED BY
E. M. STEIN
PRINCETON UNIVERSITY PRESS
PRINCETON, NEW JERSEY

1986
Copyright C 1986 by Princeton University Press
ALL RIGHTS RESERVED
The Annals of Mathematics Studies are edited by

William Browder, Robert P. Langlands, John Milnor, and Elias M. Stein
Corresponding editors:
Stefan Hildebrandt, H. Blaine Lawson, Louis Nirenberg, and David Vogan
Clothbound editions of Princeton University Press

books are printed on acid-free paper, and binding
materials are chosen for strength and durability. Pa.
perbacks, while satisfactory for personal collections,
are not usually suitable for library rebinding
ISBN 0-691-08418-1 (cloth)
ISBN 0-691-08419-X (paper)
Printed in the United States of America

by Princeton University Press, 41 William Street
Princeton, New Jersey
Library of Congress Cataloging in Publication data will

be found on the last printed page of this book
TABLE OF CONTENTS
PREFACE vii
NON-LINEAR HARMONIC ANALYSIS, OPERATOR THEORY
AND P.D.D.
by R.R. Coif man and Yves Meyer 3
MULTIPARAMETER FOURIER ANALYSIS

by Robert Fefferman 47
ELLIPTIC BOUNDARY VALUE PROBLEMS ON LIPSCHITZ

DOMA INS
by Carlos E. Kenig 131
INTEGRAL FORMULAS IN COMPLEX ANALYSIS

by Steven G. Krantz 185
VECTOR FIELDS AND NONISOTROPIC METRICS

by Alexander Nagel 241
OSCILLATORY INTEGRALS IN FOURIER ANALYSIS

by E. M. Stein 307
AVERAGES AND SINGULAR INTEGRALS OVER LOWER

DIMENSIONAL SETS
by Stephen Wainger 357
INDEX 423
v
PREFACE
In September 1984 a summer school in analysis was held at Peking

University. The subjects dealt with were topics of current interest in the
closely interrelated areas of Fourier analysis, pseudo-differential and
singular integral operators, partial differential equations, real-variable
theory, and several complex variables. Entitled the "Summer Symposium
of Analysis in China," the conference was organized around seven series
of expository lectures whose purpose was to give both an introduction of
the basic material as well as a description of the most recent results in
these areas. Our objective was to facilitate further scientific exchanges
between the mathematicians of our two countries and to bring the students
of the summer school to the level of current research in those important
fields.
On behalf of all the visiting lecturers I would like to acknowledge our
great appreciation to the organizing committee of the conference: Pro-
fessors M. T. Cheng and D. G. Deng of Peking University, S. Kung of the
University of Science and Technology of China, S. L. Wang of Hangzhou
University, and R. Long of the Institute of Mathematics of the Academia
Sinica. Their efforts helped to make this a most fruitful and enjoyable
meeting.
E. M. STEIN
vii
Beijing Lectures in
Harmonic Analysis
NON-LINEAR HARMONIC ANALYSIS,
OPERATOR THEORY AND P.D.E.
R. R. Coifman and Yves Meyer
Our purpose is to describe a certain number of results involving the

study of non-linear analytic dependence of some functionals arising
naturally in P.D.E. or operator theory.
To be more specific we will consider functionals i.e., functions
defined on a Banach space of functions (usually on Rn ) with values in
another Banach space of functions or operators.
Such a functional F:B1 -1 B2 is said to be real analytic around 0 in
B1 if we can expand it in a power series around 0 i.e.
00
F(f) = I Ak(f )
k=0
where Ak(f) is a "homogeneous polynomial" of degree k in f. This

means that there is a k multilinear function
Ak(fI...fk):B1 xB1...xB1 B2
(linear in each argument) such that Ak(f) = Ak(f, f, f) and
(1) IIAk(fI...fk)IIB2 < Ck

11 IIfjIIB1
j=1
for some constant C . (This last estimate guarantees the convergence of

the series in the ball IIf1I < C .)
3
4 R. R. COIFMAN AND YVES MEYER
Certain facts can be easily verified. In particular if F is analytic

it can be extended to a ball in Bi (the complexification of B1 ) and the
extension is holomorphic from B c to B2 i.e., F(f +zg) is a holomor-
phic (vector valued) function of z e C, JzJ < 1 , Yf, g sufficiently small.
The converse is also true. Any such holomorphic function can be ex-
panded in a power series, (where Ak is x the kth Frechet differen-
k!
tial at 0 ).
We will concentrate our attention on very concrete functionals arising
in connection with differential equations or complex analysis, and would
like to prove that they depend analytically on certain functional parameters.
As you know there are two ways to proceed.
1. Expand in a power series and show that one has estimates (1).
2. Extend the functional to the complexification as "formally holo-
morphic" and prove some boundedness estimates.
Let L denote a differential operator like
a(x) dx xfR,
a(z)az zEC
aij(x) = div A(x) grad, A _ (aij) x f Rn
ai j(x) x e Rn
the coefficients a(x) (or aij(x) ) will be assumed to belong to some

Banach space B1 of functions (for example L ). It is natural to ask
when such objects as:
L-1 , fL sgnL, e-tL , e-t\/L

,
or more generally, c(L) (where (k: C C ), can be defined as a bounded

operator (say on L2 or some Soboleff space), and a functional calculus
developed i.e., (k1(L)02(L) _ 0102(L).
NON-LINEAR HARMONIC ANALYSIS 5
Many questions arise:

a) Does F(a) = c(L) viewed as an operator valued function of a
depend analytically on a ?
This is equivalent to asking whether we can consider complex valued
coefficients in L and still have estimates on (A(L).
b) What is the largest domain of coefficients a for which we have
estimates for c(L) ? This question is the same as asking what is the
largest BI for which (1) holds, and what is the domain of holomorphy of
F(a) in this space.
The answer to question a) will require first that we understand methods
for expanding functionals in a power series, and second, that the nature of
the multilinear operators Ak be sufficiently well understood to provide
estimates (1). As for question b) we will see that the largest spaces
possible for the coefficients involve rough coefficients and leads us to
work with coefficients in L, B.M.O. and other "exotic spaces."
We now start with a fundamental example related to the Cauchy
integral. We let
La l+a -dx
with hall < 1 a(x), real valued.
If we define h(x) = x+A(x), A'(x) = a. We then have
Laf=\_d fh-110h=i
UhdxUhlf
where
Uhf =fh.
Of course, in this case, if we use the Fourier transform we can define
(I_ d)f r eix&O(e)f(e)de

idx
R. R. COIF MAN AND YVES MEYER
This gives, for example

00
sgn d
dx) f = J
elx6sgn 6 (e)de = n P. V.
f
-00
f t dt = H(f
x-t
Thus we can define

00
n sgn(La )f = nUhsgn li dx
a Ulf
h =pv J
r f(t)(l+a(t)
x-t+A(x)-A(t)
dt
-00
(where we used the observation that 95(ULU-1) = U(L)U-1 ).

We view
F(a) = sgn L. as an operator on L2(R)
and wish to know whether it is analytic on L or if we can replace a

by complex a and still have a bounded operator.
If we do this, writing a = a + ip jfaIl0 < 1 , we find
F(a) f =
f f(t) (1+ia+i$)
x-t+A(x) + iB(x) - A(t) - iB(t)
dt
f t)[(l+a)/(1+a)](1+a) dt
x+A(x)-t-A(t)+i(B x)-B(t))
= UhCUhif1
where
(' f(t) (1+Bi(t))

h = x+A(x) Cf =
J x-t+iB1(x)-iB1(t)dt
-00
f1(t) = f(t) i+a Bi(t) B1 = Bh-1

Since Uh is bounded on L2 it would suffice to prove that C is

bounded on L2 for all B such that B' is small.
We could also try to prove this by expanding
-irr sgn(La)f = J (xf t)+A(x)-A(t) dt = C' (-1)k r

/
(A(x)_A(t)\kf(1+a)dtJ
x-t x-t
-
Observe that the operators are of the form
T(f) = ft (Ax_ A t) f(t) dt = fk(xt) f(t)dt
We will prove Theorem I: Let T a Cc(C) and A(x) such that
IA xx_A t I
< M and T(f) = p.v. J (À(xx-y (Y) f(y) dy
Then the operator T is bounded on L2(R) (and LP 1 < p < oe ). This

result will then be extended to Rn and other settings.
We now return to the interpretation of C as the Cauchy integral for
the curve z(t) = t + iA(t) where A is Lipschitz
as we can see its boundedness in L2 is equivalent to the analytic

dependence of C(a)f on the curve a. This now is related to the lectures
by C. Kenig (to which we shall return later).
8 R. R. COWMAN AND YVES MEYER
Let us consider a more general version of the Cauchy integral.

Let I' be a rectifiable curve through 0, s be the arc length
parameter
s
z'(s) = ea(s) i.e., z(s) = r eia(t)dt
0
The Cauchy integral on r is given as:
00
cr(f) = p.v. f f t z' t dt

z(s)- z(t)
-00
00
r 1 f t z' t dt
J z s- Z (t s-t
00 s-t
rJ 0rz` ss-t-z t/I fs-tt) dt

-00
if we assume
* 0<a<lzs-ztl<1
s-t
we can take c e C o (C)t/i(z) = on 8 < lz l < 1 and obtain the bounded-

z
ness on L2 of C11 (from Theorem I).
Condition * is the so-called chord arc condition and * for S small is
equivalent to a e BMO with lIaIIBMO small (see [3]). If we think of C
as an operator valued functional of a, we will see that B.M.O. is the
space of analyticity or holomorphy of Ca.
2.
All the operators which we encountered previously had the form

T(f) = p.v. f K(x,y)f(y)dy
where IK(x,y)I < C IaxKI + Ia KI < Ix-y12

C moreover they were also
Ix-YI Y
antisymmetric i.e.,
K(x,y) = -K(y,x) .
i1
(For example, K(x,y) (A(x) -Y (Y)) --y
-0 c C4, A' c L .) Recently
G. David and J.-L. Journe found a necessary and sufficient condition for
such operators to be bounded on L2 (or LP ). This condition is simply
that T(1) must be of bounded mean oscillation.
We now would like to state certain facts concerning B.M.O. and
prove their theorem.
Recall that b c BMO(R) if
1 /2
IIbII* = uIII fI where mi(b) =III I b(x)dx
and I is an interval (or a cube in Rn ), and that this norm is equivalent

to the following "Carleson" norm
111 1 /2
rr *b12 dxtdt
JJ
I 0
where '/it = 0 (t) 0 C o C. fcidx =0 (ci a'0) (see [51).

A basic treason for the frequent occurrence of functions in B.M.O. is
the following simple fact.
PROPOSITION. If T is as above and T is bounded on L2 then T

maps L into B.M.O.
Proof. Let b c L and let I be given. Consider I = 21 and write

Tb = T(bXI) + T(b(1 - XI)) = -51 + b 2
Clearly
1 /2 f 1 /2
lill Ib-m1(b)I 2
1/2
+
r
III f ID2-ml(b2)I 2
The first term is dominated by
1 /2 1 /2
2(-L
II
JI
r1I2dx <c III fI
<cIIbII
For the second we observe that
T(b2)(X) - T(b2)(u) = I f[K(x,y) - K(u,Y))b2(y)dy
<
f
Ix-y1>111
III dylIbll,. < CIIbII .
IX-YI2
Integrating in y we get
IT(b2)(X)-mI(T'(b2))I < CIIbII,,.
which shows that second term is bounded by CIIbII. We have thus shown
the necessity of the condition T(1) c BMO. Before stating the theorem
precisely we would like to reformulate it somewhat.
Let k eCU(RI) with f6dx = 1 and =f(x). Let ft(u) _

(U-xand similarly for Vit (u) . We claim that under the preceding
assumptions on T we have
l<T'&r,Ot>I <CPt(x-y) =Ct 1
1+
(xY)2
In fact, assume for simplicity, that S6 is supported in (-1,1). Since

f c&du = 0, if we assume Ix-yj > 3t
I<Tqx, (kt >I = fr(z){j'K(z ,u) dz
< J'Ic!'(z)I t 21ot(u)ldudz <C t

IY-xl ly-xl2
(where we used the fact that ly-z I < t , Ix-ut < t, Ix-y l > 3t and the
hypothesis layK(x,y)I < ). Ix-yi-2
If Ix-yj < 3t we use the antisymmetry of k(x,y) to write
<T I2 ffK(z,u)Otu)Ot(z)-.At(z)ot(u))dzdul
but lp (u) t (z) -fir (u) (z) I < Iu and the fact that Iu-x I < t ,
t
1
t
Iu-zl < t, Ix-yj < 3t and IK(z,u)l <3 1 imply
lz-u I
,.Ot >I
Combining these estimates proves our claim.

We can now state
THEOREM (G. David, J. L. Journe) [7]. Let T:`D - D' such that for some
E>0
<T,`t'St >1 <- t 1 I+E
= pt(u-v)
1+IuyI
t
and
* I<T*,At, q5t >!

S t vll+E = pt(u-v)
1 + Iu
t
Then the necessary and sufficient condition for T to extend to a bounded

operator on L2 into L2 is that
T(1) and T*(1) be in B.M.O.
We would like to make a few comments concerning the conditions *.

We have just seen that if
T(f) = p.v. J k(x,y)f(y)dy
where
1 Ik(x,y)I < IxyI

1
2 Ik(x,yl)-k(x,y)I < Iy-yI IE for Ix-yI > 2Iy-y' I

Ix-yI'+E
and
1 E
k(x1,y)-k(x,y)I < Ix x
I I
for Ix-yI > 2Ix-xI I
E
x-yI
for some E>0.
3 I<T(c1r ), or >I + I<T*(,A ), Ot >I < t

(This last condition followed from k(x,y) = -k(x,y) and 1, 2) then the
conditions * are verified. It can be shown that if * is valid then T
can be represented as a limit of integral operators whose kernels satisfy
the conditions 1, 2, 3 We will refer to 1, 2 as standard estimates,
and to 3 as the weak cancellation (or boundedness) property. This last
condition is independent of 1, 2, and can be proved in a variety of ways,
as we shall see.
To see how the theorem can be applied to reduce the degree of non-
linearity a of a "polynomial" and to obtain estimates consider
C(aIf)
n = ft!J,A(xX_A(Y)ln
! Y J
f(y)
Y
dY
We check by a simple integration by parts that
Cn(a,l)(x) = Cn-I(a,a)(x)
If we make the induction hypothesis that Cn_I(a,f) is bounded on L2

it would follow by the preceding remarks that Cn_I(a,f) maps L to
B.M.O. from which we deduce that Cn(a,l) a B.M.O. and that 3C > 0
such that
IIC5(a,f )IIL2 <_ CnllaIILjIfIIL2
We will see later that this reduction of Cn to Cn_I is not a trick

but can be accomplished in general to reduce the study of n+l multilinear
operators to an n-multilinear.
Proof of the theorem. For simplicity of notation we'll denote
Ptf=95t*f Qtf=qit*f.
We start with the fundamental theorem of calculus writing for f e Co(R)

1 /E
Tf = li PtTPtf Eir I t (PtTPtf) dtt

o
E
We'll study only the first term (since the second is its adjoint).
We start by observing that
pt 2g
t _2 ( ) = t 2 2(te)g(e)
= 20(t6)01(te)g(e) = (Q1 tQtg)
where ip1(x) = xq5(x) and Q 1,tg = iA1,t *g. This permits us to rewrite
the first term as
00
dt
-2 f Q1 tLt'tf t
0
where
Ltg = QtTPtg = J t(x,Y)g(Y)dY
and
et(x,Y) = < T*Ot , Ot >
Also observe that Lt(l) = f et(x,y)dy = <T*cb ,1 > = At *T(1) = Pt*b

(here all integrals converge absolutely since let(x,y)I < pt(x-y) ).
Before proceeding we remark that since Iet(x,y)l < pt(x-y) we can
think of Lt as an averaging operator on the scale t, and of Ptf as
being essentially constant on that scale i.e., Lt(Ptf) would look like
Lt(1) Pt(f) . (This would be exact if Pt is a conditional expectation and
Lt had the correct measurability.) We are thus led to write
r QI,tLtPtf dt = r QL,t1Lt(1) Pt(f )t Lt

0 0
00
+ QL,tlLtPtf-Lt(1)Pt(f )Idt
J0
r
0
We consider E(f) as an error term and prove that
J <e J Ifl2dx
(this is valid using only *). In fact, let g e L2
j<g,E(f)>1 = fJQit()(LtPt)f-Lt(1)Pt(f)}dx tdtl
1 (f
4xtdt) ftxYxPt(f )(y)
< rTQ1,t(g)I2
- Pt(f )(x))dy
A simple application of Plancherel's theorem permits us to estimate

the first term. For the second we use * and Minkowsky's inequality to
dominate it by
1 /2
dxdydt
(j//'Pt(x_Y)IPt(f)(Y)_Ptf )(x)12
1 /2
dudydt
fff Pt(u)IPt(f )(Y)-Pt(f )(Y+u)12 t
- dudedt1 /2
')I2leit1 1I2f(e)2
t
1/2
(fffpt(u) 1u S I$(te)12 Itel 8 2 du ddt
` t I
1^
(6)1 t
1/2 00 1/2
= P(u) us f I0(t)12ts dt (j(2de)1/2
J
o
As for the first term, we proceed as above and are led to estimate
1 /2
(f*bX)I2t*f2 d tdt
This is dominated by llf 112 if and only if (fit * b(x))2 d tdt is a Carleson
measure i.e., b is in B.M.O.
In fact, recall that dv(x,t) on R* is a Carleson measure if for each
interva 1 I
v(I) < C111
I =1(x,t) cR2:(x-t,x+t)<If.
I
Carleson's lemma states that
fFP(xt)d(xt) <c J F*(x)t dx ,
where F*(x) = sup IF(x-y,t)j .
This is proved by observing that if OA = {x:F*(x)>A1 then outside

OA F(x,t) <A, thus
v(1F>A1) < v(OA) <cIOAI
and the LP result follows by integration.

We now would like to make a few comments concerning the principal
term:
00
f Qt((,At * b) (0t * f )) dt = 77 (b; f)
This is-essentially equivalent to a simpler looking version
00
J (art*b)(0t*f)dt
0
which is the basic bilinear operation in (b, f) commuting with transla-

tions and dilations i.e.:
If bT(x) = b(x-r) and bx(x) = b(Ax) then
Since all the problems we will be considering are invariant under such
transformations, it is not surprising that the bilinear operations which
arise look like n.
One cannot conclude this section without mentioning that IT is also

the principal term in the Bony linearization formula and is called para-
product by him i.e., given F f C, f Aa, 0 < a < 1/2 one has the
remarkable formula (Bony)
F(f)=rr(f;F'(f))+e
where
e e .12a (f EAa iff If(x)-f( y)1 < cIx-yIa) .
The proof of this linearization formula is the "same" as for the T(1)
theorem and is achieved as follows.
F(f) = lim F(f * ct) = -

do fr
0
t -2 F(f * qt) at
00
J F'(f * fit) f * Vt !Lt (here

ct = t 2- ct)
0
00
f F'(f)*(btf * btdt+e =rr(f;F'(f))+e

0
and it is quite easy to verify the e e A2a if f e Aa. (This method

should be explored further in connection with non-linear estimates [15].)
3. Multilinear Fourier Analysis

We would like to understand various explicit representations for
"homogeneous polynomials" i.e. multilinear functionals. For simplicity
and to avoid technicalities we consider Ak a multilinear operator
defined on trigonometric polynomials on T' with values in periodic con-
tinuous functions i.e.
is multilinear in the ti
We will assume that Ak commutes with simultaneous translation of ti i.e.
* Ak(th,tz,...,tk)(e) =Ak(tI...tk)(e-h)
where
th(IA ) = ti(V, - h) .
ikie,thi e-ihki
Clearly if we take ti(9) = e ( = ti(O) we find using the
linearity in each argument that
k
A h(l ki)
ik19 ikke ik10 ikkO
e Ak(e , ...,e )(e) = Ak(e ...e )(e-h)
taking h = 0 we get
Ak(e e (e) _ A(kl ... kk) e
More generally if ti(0) = F ti(j)e' 0 we find

k
ie(! j i)
A(tlt2...tk)(e) =I...LA(jl,...,jk)tl(j1)t2(j2)tk(jk)e 1
jk j1
where
A(jl ... jk) = A(ellle, ...,ellke)(O)
Before continuing let us observe that for k = I , eike form a basis

diagonalizing all linear operators commuting with translation, in fact, all
such operators are norma I. For k > 2 it isn't as simple.
Consider for a moment a finite dimensional vector space V and a
bilinear transformation AN xV - V. We can ask when is it true that
there exists a basis in V el eN a linear operator A defined on

VV diagonalized by eiej
A (eiej) = Ai,j eiej
such that
A(ei,ej) = Ai j e,(i,j) for some map n(i,j) of
(1,..,N)2 -(1,...,N)
i.e. we want the diagram to commute

ti
A VV
where
v(ei, ej) = eiej n(eiej) = en(i,j)
This is clearly the situation for periodic functions. It would be interesting

to understand the bilinear operators admitting such a diagonizable lift to
the tensor product.
This observation indicates that the hypothesis concerning commuta-
tion with translations, imposes severe restrictions on the nature of a multi-
linear operation.
We now return to the line on which we realize multilinear operations as
ffeix(e1+...ek)VeI...ek)(eI)f(e2)...f((k)deidek
Ak(fI...fk)(x) =
Rk
and
nk(eieix, ..., eix(e1+...ek)Ak(eI

elekx) (x) = ... Xk) .
(Note that this realization is only valid for multilinear operations verify-
ing some mild continuity conditions.) This realization permits us for
example to show that the study of fo lAt * b (kt * fat can easily be con-
verted to the study of
J to * (iAt*b 95t* fdtt

0
which we considered previously in the proof of Theorem [II].

In fact
0
(' t/it*b(kt*f dt _ ('eix(eI 2) (f(tei) 3(te2)dt 2
0
J (ow
if we take with support in 1 < 2 and supported in lei < lU

we have
5l'1(t(e1
where 0 I(() is any function equal to 1 on 1.9 < lel < 2.1 . In this
case we see that the two expressions are equal. This sort of analysis
comparing frequencies and interaction of various functions can be carried
out permitting one to understand the structure of some multilinear opera-
tions. We are thus led to consider the general question of studying multi-
linear multipliers. We state two known results.
PROPOSITION 1. Let I'k(61 6k)1 < 1 then

(max
k2
k
ei
A(fI...fk) = re I 4e1... ek)f(e1) ... f(ek)del ... d fk
maps L2 x x L2 L2 continuously.
This is an elementary result which we leave as an exercise. It
would be desirable to obtain more subtle results for L2.
THEOREM [5]. Let A(e,...4k) be such that
al, Ca jai _S [.n] + a=(a1"'ak)IaIai

Id ,
dial 2J
then
A(fI...fk): LPIxLp2...xLpk,L4
k
>pi>1 >Pi
More generally one can take operators of the form
o(f I" f) =f I" dd d4k

k
J k 1 1
with o(x, d) such that
1a dX o(x,d)I < Ca1R for all a,/3 sufficiently large

+IeF
and the same conclusion is valid.
We now prove the result by induction on k; for simplicity we consider
only k = 2.
We take f e LP , f 2 = a e L and write
eix(e1+e2)x(41,
A(f, a) (x) = IJ e2)a((1)f(e2)del de2
[fk(x_tx_)a(t)] f(y)dy
where k(x 1,x2) = A(x1x2) satisfies Ik(x 1, x2)I < IVkI <
x12 +Cx22 ,
3
Ixi
thus k(x, y) = f k(x-t, x-y)a(t)dt verifies conditions 1, 2, and 3 of the

T(1) theorem: thus to verify the boundedness in L2 (for a fixed) it
suffices to calculate T(1) and T*(1). These can easily be calculated,
in fact
T(1) = J eixel
X(e1, 0)9(S 1)dS l
which is a linear Calderon Zygmund operator applied to the bounded func-

tion a. Thus it is in BMO.
Since T*, (in f, for a fixed) is given by the symbol 'k*((1, e2) _
X(el, -el -'), verifying the same hypothesis the result follows for T*(1).
Property 3 can be verified once we observe that the same induction shows
that
IIT(e'xe)IIBMO < C .
and that this condition implies
f IT(-Ot)ldx < c on any interval of length t .
III
In fact
Ot(x) = feiX6 (et)de4 lIT(ot)(x)IIBMO t ,
now assume cbt(y) is supported in I.
lox-y(y)l
fIT(9t)(x)I <_ dy
-t
<
if lx I > 2 It l thus
JTt(x)Idx < c if dist (I,I) > t .
I
Since T(4t) is in B.M.O. it follows that
t > III fIT(t)(x)-miT(t)) 1

ftTdx < c ,
thus we obtain
IT(Ot) ,At 15 t .
J
provided trtOt are supported in I.

Again we conclude with a few general observations. As you all know
the Fourier transform in a basic tool in linear P.D.E. permitting the reduc-
tion for example, of initial value problems to simple O.D.E. We claim that
this can be achieved also for non-linear partial differential equations.
Before illustrating on an example recall also that a common method for
solving differential equations (linear and non-linear) is to plug into the

equation a formal power series whose coefficients are determined recur-
sively using the differential equation. We now sketch how these two
ideas can be combined to "formally" solve the Korteweg-de Vries
equation (KdV)
0 "U
+a36u"u
&A3
xER,t>0.
The set of solutions is clearly invariant under x translations and can be
parametrized by functions of x (for example by specifying u(x,O) = f(x) ).
Let us write u as a power series
u(x,t) = 1: Ak(v) = I uk(x,t)
where
k
f ix(16i)
uk(x,t) = J e 1 ak(e,t)v(e1)... v(ek) del ... d6k
(here v is some unspecified functional parameter) plugging into the

equation we obtain
n-1
r3uk ,auk
3
at ax 3 r)x Lr uiuk-l
j=1
k
it e3
we find first that ak(e, t) = e 1 ak(e), and that
/k k
( k
13 k-1
ak(6)1 6i3 l ; J )=_3(e1) Y aj(61 ... ej) 6k)
1 j=1
(Strictly speaking ak is determined only if we assume that it is invariant

under permutations of ek, so that the equality should be true only after
symmetrization.)
It can be shown (not so simply) that
e1+...+ek
ak (e1+Q (e2+N ". (ek-1+ek)
solves the recurrence above.

Thus
k
uk(x,t) =13
i J 1 1 ;k-1 k i=l vx,t(ei)de1 ...d,k
vx,t = eixe+te 3 v(S)
If we let p. v. f --L- v,
uk(x,t) = iax J
and
u(x,t) = _ J ,U-Cv)-1(1)dri _ fV 0 d71
where satisfies the integral equation
I I- ' f T4,
d
This set of remarkable formulas constitute the so-called method of

"inverse scattering" for solving K deV see [9], and was shown to us by
B. Dahlberg.
4. Functional calculus, resolvent expansions

We start by considering a small perturbation of -0
ItbIL,. <E0.
ij 1
The bij(x) are complex valued functions. We would like to prove L2

estimates for such functions as
e-tL, eit L, eitV'L, e-tVL
These provide us with estimates for solutions of the initial value or

Dirichlet problem for
2
u= -Lu, -1u=Lu, au=
2 -Lu
at
a2 -Llu=0
&2 /
or more generally for F(L) where F is a bounded holomorphic function
in a sector jarg zj < S containing the spectrum of L (this can be shown
to be true if e0 is sufficiently small).
We define formally
F(L)
2ni J
d
r
where r is the boundary of the sector.
We let
R1= 92
ax
R =
I
i 1
2
-L "
+A-Lr bil
2
(I-BR) (C +A)
= fix) =
(C-L)-I = (C+A)-1(I-BR)-t
=
, (C+A)-t(BR)k
We thus find
00 00
F(L) = I J (C+A)-t(BRekF(C)dC = IAk(B) .
0 r 0
To better understand such expressions we consider the first two terms
(A0(B)f) = J Q F(C)dI 1
2ni
1 F(C)dCf(C)
C_ICI2
r r
=F(IKI2)f(C) = m(C)f(C) .
Since F is assumed to be bounded holomorphic in a wedge containing the

c
real axis we obtain (using Cauchy's theorem) that I(m(k)(C)I < and
gIk
thus A0 is a Calderon-Zygmund kernel operator mapping L2 - L2,

L - BMO. We now consider A1(B) and to simplify the exposition we
will assume that r = iR (i.e., F is bounded holomorphic in the half
plane Re z > 0 ). This is not going to affect the argument.
00
AI(B) = f
t B a2 1 F(it) dt
it+A axiaxl itt+ t
J
00
and consider separately t > 0, t < 0. Change variables t = , to

S2
obtain 2 terms of the form 2
J 1
B s2D;D 1 (s) as (s) = F(1/s2) .
i+s2t i+s2A
0
Now write
D.D. D.D. D.D.

s2DiD) _ s2A J I
A
i+s 2O i+S2O i+s2O
00
D.D.
This gives f 1 B 1 j (s) ds
s which we recombine with the other
0 1+s2O
term corresponding to t < 0 to get
D_J
.D
A0(B) B f
The second term is of the form
If'o 1+s2O
1
B
1+s 2O
1
G) s
ds DiDI
0'
The operator in curly brackets has a kernel satisfying *. Thus, to check

boundedness in L2 we need to calculate T(1) , T*(1). This again re-
duces to the linear case when the terms are recombined.
The higher order case is much more complicated since terms appearing
lack regularity and need to be replaced by more regular terms complicating
the induction. The main idea is the same, see [7), leading to L2
estimates for F(L) for r0 sufficiently small.
In general a functional calculus is obtained by considering functions

F(L) where F is holomorphic in a neighborhood of the spectrum of L,
and using the Cauchy formula. This becomes impossible for an operator
like N, z cc, z =x1+ix2 since
e 161x12x2 = (61 -'62) el x
and the spectrum is the whole complex plane. We are thus led to consider
for 0 e Co(C1) expressions of the form
=-2ai fx)1(idxAd.
C C
These formal expressions need to be given sense and one should prove that
(oq,)(L). In the case L0 = 2 such an expression is
easily justified using the Fourier transform. Since
d
1 dX A d f (e) = 21 4(X) 1
dX
21 1(f aX- X-L
C
ax X-6
= O(Of(e)
(here we used the fact that S(z), where S is the Dirac mass
tai az =
at 0 ). z
We see that
O(L0) = fe'x6OC6)i(6)d6 = f cv(z-Y)f(y)dy .
If we now consider L = -L where a E L, Ila11 < S0 we would

ITO
like to understand the nature of
(L) = 2ni f c(A) aX L-A as A as
for functions / homogeneous of d0. For example, we would like L

(this will give us L = L L which will turn out to be very important).
We can, in this case, calculate explicitly
(L-A)-1 , in fact, L-1f = J f(l+a)fdv(w) .

"(Y-W )
We now observe that if h(z) is such
21 h=l+a and h(z)-zeL
then
= L(ei(h(z)C+i )) =
LXC CXC
from which we find
(L-C) f = XL if and XL-1 X f

X
or more precisely
n
f e'{(h(z)-h(w))C+(z-w #
f(w)dw
(L-C)-1f = 1 e"(h(z)-h(w)X+(z-w)f# dh f(w)dw
and
)ei1(h(z)-h(w))C+(z-w)Z1dr
c(L)f ( f(w)dw
f1fo
fk(h(z)-h(w),z-w) , f(w)dw
where
k(u,v) = f)eidC
C
is the Laplace transform of c. This gives us an explicit kernel realiza-

tion of the calculus.
It can be shown independently that
Lf=-1 fw dw-!!f.
L (z-w)2 0z
Since Lh = 1+a, for this to be bounded we must have !Lh c L i.e.,

a * 2 c L. In other words, in order to have a functional calculus in
z
it T (consisting of bounded operators in LZ(CI)) it is necessary to
assume that a c L and the 2-dimensional Hilbert transform (or Ahlfors
Beurling transform) of a is also bounded. This version of H is quite
interesting and should be better understood.
5.
Until now we have only shown that small perturbations of certain

operators are bounded, we would like to describe an extension method due
to G. David.
We proved earlier that the commutators
-yy
Ck(aIf) _ J(A(x)_A(y))kf(y)d A(x) = a(x)
sati sfy
IICk(a,f )II1 < Ckllallk llf 1 2
(This is A. P. Calderon's theorem [1].)

If we recombine these terms in a series we obtain, for example, that
f
f(y)
dy
x-y+iA(x) - iA(y)
is a bounded operator on L2 if IIA'II, < c and that
I A(x)-A(y)
i x-y
f(y)dy < ec lei IIA'Ilo
x-y IIf II2
L2
It was proved in [4] by a careful analysis of the functional calculus

interpretation of the Cauchy integral that in fact,
< c(1+k4) IIaIIkIICk(a,f ) I I

llf 112
L2
This more precise result implied a much stronger statement.
THEOREM 1 [4]. Let 0 E C11(R). A real valued IIA'II. < - then
x-y X-y y <clif112

(A(x)_A(y)'f(y)
L2
It is our purpose to describe a direct real variable method to obtain

this result. The main idea, due to G. David, is that on each interval a
Lipschitz function with a given norm is, in fact, equal (modulo an
additive constant) on a substantial subset to a Lipschitz function with a
fraction of the norm. By iteration this permits estimates for large norms
in terms of smaller ones.
To get results like Theorem 1 it is enough to show that for A real
A(x)-A(y)
x-y
IITe(f)112 = fe (y) dy
x-y < C(IIeajl-+1)NIIf II2
2
for some N>0, C>0.

In fact, this will imply that for (x) such that +IeIN)d5 <
we have
A( x ) - A( y )
P (
A(x)-A(y) f(y) dy
x-y ) x=y fc){fe xydy d
2 2
< fiei IlTfII2de < cf 112
To show that the norm of

. A (x)-A (y)
i x-y
TA(f)= re dy
grows like 11A'1111 we start with a number of easy observations. The

kernel of T satisfies the estimates (*) and is antisymmetric, thus it
suffices to estimate the B.M.O. norm of T(1).
The operator norm of T is unchanged if we add a constant to a , or

replace a by -a (this in fact will imply that the result is true for
a e BMO ).
We also need a lemma indicating that the space B.M.O. can be
characterized by a weak local estimate.
LEMMA (Stromberg). Let b be measurable and assume that there exist

a > 0, N > 0 such that for each interval I there is a constant C(I)
(depending continuously on I) for which
Ix eI: Ib(x)-C(I)I < aI > N III
Then b e BMO and IIbIIBMO < cNa.
The main idea to estimate the BMO norm of T(1) is to replace inside
each interval I , TA by an operator TA where AI = A on a large
I
fraction of I and Ai has a smaller Lipschitz norm, and then compare
TA (1) to TA(1). This is achieved via the following lemma, the first of
I
which is the rising sun lemma (or the one-dimensional version of the
Calderon-Zygmund decomposition).
LEMMA 1. Let A be such that
C-M < A'(x) < C+M
then for each I there exists a function AI and a constant CI such that
AI =A on a set E IEI > III

4
and
CI-3M<AI(x)< CI+3M.
Proof. We can assume C = M i.e., 0 < A' < M2. There are two cases:
a) mI(A') > M.
In that case consider the smallest function AI > A with A'(y) > 23 . We
Ik
then have AI(y) = A(y) except for disjoint intervals Ik on which

AI(y) = M . But
fiA(Y)dY = + < 2M II-UlkI +
"_.UIk 'fUlk
=2MIII-3MIUlkI
3MIUIkI < (2M - m1(Ai)) II I < (2M - M) III =III M
i.e.,
IuIkI < 4 III .
Since I< Ai(y) < M2, we have
3M-3M<Ai(y)<43 +3M.
b) m1(A') < M.
We consider the function
2M-A'(x) = Ai IA; I < M2 .
Then we have ml(Ai) > M and we construct A 11 as above.
A11 = 2M(x-a)-A(x) except on a set of meas < III
4
A(x) = 2M(x-a) - A11 = Al
A'(x)=2M-AI=A1
3M-3M<Ail< 3 +3M
2M-3M-3M<2M-A' < 3IM+3M+2M
3M-3M<AI<3M+3M.
The main result is the following.
THEOREM (G. David). Assuming that there exist S > 0, c > 0 such that
for each I there exists KI(x,y) satisfying standard estimates uniformly
in I with
T1(f) = J K1(x,y) f(y) dy
satisfying
IITI11L2(L),L2(I) < Co
and there is a subset E C I with IEI > 6111 such that
Hx rE , Vy eE , K1(x,y) = K(x,y) .
Then T maps L to BMO with
IITIIL,BMO C77 C0 .
<-
iMA(x)-A(y)
x-y
If we let a(M) = sup II f e x-y dylisMo a direct application
IIA-liar 1
of this theorem choosing for each interval I
A1(x)-A1(y)
im x-y
e
K1(x,y) = x-y
shows that
a(M) < Ca( M)
i.e., a(M) < c1(1+M)N for some N or
. A(x)-A(y) II
t x-y
e
IITf11 2= x-y f(y)dy L2 < C(1+IIA'II00IIflIL2
Proof. We consider f e L, f = fXl + f(1-X1) = f1+f2, llf1l00 < 1 . One

checks that
f IT(f2)(x)-T(f2) (x01 < clll

I
from which it follows Ix el:ITf2(x)-c1I > cI < 770111 if c is large enough.

We now claim T(f 1) is well approximated by T1(f 1) , which we con-
trol, so that we want now to estimate
f IT(f1)-TI(f1)Idx
E
where E = I - UIi , E' = I - UIi where Ti = (1+s) Ii
fIKxy)_Ki(xY)ldY
J
xEE' yd
IK(x,y)-K(x,Yi)+KI(x,Yi)-KI(x,Y)IdY
= J
xEE' I.i
where we have used the fact that K(x,y) = KI(x,y) outside Ii and yi
are endpoints of Ii, (K(x,yi) = KI(x,yi)). This integrand is dominated
by the Marcenkiewitz function of UIi . Consequently the integral is
bounded by c III enabling us to apply Lemma 1.
All of these results can be extended to Rn by various methods. The
easiest is the so-called method of rotation based on a general transfer-
ence principle, valid for multilinear operators commuting with translations,
and some nonlinear operators. We state the result in general although for
the case of Rn this is an easy application of Fubini's theorem.
TRANSFERENCE THEOREM. Let Ut be 1 -parameter group of measure

preserving transformations on a measure space (X,dx) and A(a I . . .a k, f )
a k+1 multilinear operation on R given by
A(a; f) = ff k(x-t, x-t2 ... x-tk, x-y)JJai(ti) f(y)dtdy

Rn
=J
satisfying
k
I IA(a; f)II, p<_ [I IIaiII,oIlfIIL
1=1
Then the multilinear operator A on L'(X)k x LP(X) -> LP(X) defined by
A(A,F) = J
frk(t1,t2. tk,s)[JAi(Ut .x)F(Usx)dt ds 1
satisfies
I}A(A,F)IILP(X)` 11 IIA'IIL(X)IIFIILP(X)
(note that the constants are the same).
The examples we have in mind are the following. X = Rn, dx Lebesque

measure, e a unit vector
Utx = x-te .
If we take Hf f x t dt, A f = f x -te dt . Or consider

fR
f
X = T1 Ut(ei9) = ei(e-t)
4f - c rf(0-t) 1t dt
J ctg t
If we take
k
Ak(a,f) = x-y f
f(A(x)_A(y)dy
SE--y
k
f(A(x)_A(x_t)) A? t dt
then
Nk,e(A' f) = ftf A(x-te),j

t /
k
f(x-te) dt t'
Multiplying by i2(e) and integrating on IeI = 1 in Rn we get
k
S2(e)
A(x)-A(x-te) f(x-te) dt
t t
f (A(x)-A(y)lk
x-y I K(x-y)f(y)dy where
is odd for K even
k(Y) _
WYIYI)
IYIn
even for K odd.
6.
We now are in a position to recall the various ingredients which we

discussed previously and recast them in a general setting. We considered
operators L which are "small" perturbations of an operator L0 for
e0 and
example, we took L0 = -A and L = -A+Y- blJ
ax-
-E
then proceeded to expand functions of L, F(L) as a power series in

terms of the coefficients of L-L0 i.e., we wrote
00
F(L) = F(L0) + I Ak(b)

k=1
where the Ak(b) was an operator valued homogeneous polynomial of dAk

in b. Of course such perturbations could be shown to converge only in a
little ball around 0 (in the space of coefficients). It is then appropriate
to ask how far can one extend the results and what is the natural domain
of analyticity or holomorphy of the function F(L). This question is
meaningless if the Banach space in which we prove analyticity (in a

neighborhood of 0) is not the largest possible space for which such
estimates can be proved. A first task is to identify this largest space,
which we will call the space of holomorphy. Once this space has been
found it is natural to ask for the domain of holomorphy of the correspond-
ing functional.
Let us return flow to our first example where F(a) = sgn(1i dx)
where a is a function on R, and F(a) is an operator on L2(R). The
series expansion was
f(y) (7 A (X
(1-a(y))dy, A'=a
F(a)f=1: J y
0
Let us consider for simplicity the commutator series
Fo(a) f > ` I (A(xX y (Y))k X_ y fly (f )

J =
0
We wish to find the smallest norm III III (i.e., the largest Banach space)
for which
* IIAk(a)IIL2 L2 <_ Ck IIIaIIIk
in particular for k = 1 we must have
IIA1(a)IIL2 L2 <_ Ca .
If we decide to define IIIaIII = IIAI(a)II(L2 L2 it certainly defines a

seminorm, if it actually is a norm for which the estimate * can be proved,
we would have identified the space of holomorphy. But
NON-LINEAR HARMONIC ANALYSIS
A(x)-A(y ) f(y )
AI(a)f =
f x-y xx-y
dy .
We have already shown that 1IA I(a)11 L2L2 < c Ila II, and it can easily be
shown that IIalI, is dominated by cIIAI(a)II 2 (In fact it suffices

L L 2.
to consider the L2 norm of
(x-x 0)2 A I (a) ()(I) - 2(x -x 0) A I(a) ((y-y 0) XI) + A I (a) ((y-Y o)2 X I)
on the interval I whose center is x0. By letting I shrink to x0 and

using the L2 estimate on AI(a) one obtains an estimate for a , for
simplicity one can assume that a is smooth since the estimate does not
depend on this condition.) Thus our norm is equivalent to the L
III III
norm. The estimates * have already been proved in this norm.

We may also find the domain of holomorphy of F0(a), in fact
f(y)dy
1 f(y)dy = fT-_ A (x1X -A x-y
F0(a)f -f x-y (A(x)-A(Y)) _y
We have seen by the method of boosting for the Lipschitz constant

that this will be bounded on L2 as long as
1 ,/A(x)-A(y))
x-Y for some q5 c C
1 - A(x) -A(y) 1`
x-y
clearly the case whenever infI

x,y A X_- - 1I > 0 and we conjecture
that this condition gives the domain of holomorphy. [Before discussing
other examples we urge the reader to identify the space of holomorphy for
F(a) = sgn(l1 dx) . [Caution: It is not enough to consider all Ak
separately.`]
A more interesting example arises if we reconsider the Cauchy

integral on rectifiable curves, which we chose to parametrize by arc
length (and not as graphs).
We write z(s) = fo er(a(t))dt and consider the L2 operator valued
functional
f
00
aa)f . 217i
p.v.
z(s)-z(t) dt =
Ak(a) f
_ ,p
here again one can define the norm
IIIaIII = IIAI(a)IIL2,L2
and one finds that this norm is equivalent to the BMO norm of a. On the
other hand it is easy to show that if IIaIIBMO < So then 1 -So <
Iz s -z t I<1 and this is precisely the condition permitting us to write
S -t - ,
C(a) = f 4(z ss_z t) fLt) dt giving rise to a bounded operator on

\
L2.
S-t
Thus BMOis the natural space of holomorphy. As you recall from
S. Krantz's lecture one can express the Szego projection S(a), projecting
L2(1', ds) L2(R, ds) onto H+(I', ds) (H+ is the space of functions in
L2 admitting a holomorphic H2 extension to the "left" of T ) in terms
of the Cauchy operator. This representation yields the result that the
S(a) has BMO as its natural space of holomorphy and is an entire func-
tion on the manifold of chord-arc curves. (A similar result is true for the
Riemann mapping function.)
Another remarkable example leading to an "exotic" space of
holomorphy and to interesting geometry involves the functional calculus
in L = 1L - . As already seen it was necessary to assume that a E L
z
12
and a * E L with small norm. (This was obtained by considering the
z
example T(a) = L .) It turns out that T(a) is analytic in a , relative to
the norm IIIaIII = IIail0 + Ila * II.., and the condition really means that
a
Z2
there exists a bilipschitz map h(z):C -C such that ai h = a(z) + 1 .
It is clear from these and other examples that the identification of

the space of holomorphy, or a detailed study of the first bilinear operation
AI(a)f is basic to the understanding of these functionals.
R. R. COIFMAN YVES MEYER

DEPARTMENT OF MATHEMATICS CENTRE de MATHEMATIQUES
YALE UNIVERSITY ECOLE POLYTECHNIQUE
NEW HAVEN, CONN. PALAISEAU, FRANCE
REFERENCES
[1] A. P. Calderon, Cauchy Integrals on Lipschitz curves and related
operators. Proc. Nat. Acad. Sci., U.S.A. 75 (1977, 1324-1327.
[2] R.R. Coif man, D. G. Deng and Y. Meyer. Domaine de la racine
caree de certains operateurs differentiels accretifs. Ann. Inst.
Fourier 33, 2 (1983), 123-134.
[3] R.R. Coifman, Y. Meyer, Lavrentiev's curves and conformal map-
pings, Rep 5. 1983, Mittag Leffler Inst., Sweden.
[4] R.R. Coifman, A. McIntosh and Y. Meyer. L'integrale de Cauchy
definit un operateur borne sur L2 pour les courbes lipschitziennes.
Annals of Math. 116 (1982), 361-387.
[5] R. R. Coifman and Y. Meyer, Au deli des operateurs pseudodiffer-
entiels. Asterisque 57. Societe Mathematique de France (1978).
[6] G. David, Operateurs integraux singuliers sur certaines courbes du
plan complexe. Ann. Scient. Ec. Norm. Sup. 4 serie, 17 (1984),
157-189.
[71 G. David, J. L. Journe, "A boundedness Criterion for Calderon
Zygmund operators," Annals of Math. 120(1984), 371-397.
[8] E. Fabes, D. Jerison and C. Kenig, Multilinear Littlewood-Paley
estimates with applications to partial differential equations. Proc.
Nat. Acad. Sci. U.S.A. 79(1982), 5746-5750.
[9] R. Rosales, Exact Solutions of Some Nonlinear Evolution Equations,
Studies in Applied Math. 59, 117-151.
[10] E. M. Stein, Singular Integrals and Differentiability Properties of
Function, Princeton University Press (1970).
MULTIPARAMETER FOURIER ANALYSIS
Robert Fefferman
Introduction
The article which follows is an attempt to give an exposition of some
of the recent progress in that part of Fourier Analysis which deals with
classes of operators commuting with multiparameter families of dilations.
In some sense, this field is not that new, since already in the early 1930's
the properties of the strong maximal function were being investigated by
Saks, Zygmund, and others. However, for many of the problems in this
area which seem quite classical, answers have either not been found at
all, or only quite a short time ago, so that our knowledge of the area is
still fragmentary at this time.
The article is divided into six sections. The first treats some basic
issues in the classical one-parameter theory whose multiparameter theory
is then discussed in the remaining sections. Since the reader is no doubt
quite familiar with the main elements of the classical theory, we have
omitted references to the materials in section one. The book "Singular
Integrals and Differentiability Properties of Functions" by E. M. Stein is
an excellent reference for virtually all of the material there.
Finally, it is a pleasure to thank Professors M. T. Cheng and
E. M. Stein for all of their hard work in organizing the Summer Symposium
in Analysis in China, as well as many others whose generous hospitality
made the visit to China such a very enjoyable one.
47
48 ROBERT FEFFERMAN
1. The maximal function, Calder6n-Zygmund decomposition, and

Litt lew ood-Pa ley Stein theory
We hope here to review briefly some aspects of the classical
1-parameter theory of these topics. The three are inseparable and we
hope to stress this.
We begin with the fundamental
CALDER ON-ZYGMUND LEMMA. Let f(x) > 0, f t L1(Rn), and a > 0.

Then there exist disjoint cubes Qk such that
(1) a < IQ I ' f(x)dx < 2na

k
Qk
(2) f(x)<a a.e. for x/UQk

and
(3) IU QkI << f If(x)Idx -
a
Rn
Proof. Let Rn be subdivided into a grid of congruent cubes so large

that IQ fQ f < a for all of them. Subdivide each cube in this collection
II
into 2n congruent subcubes. Select from these the cubes Q' such that
1 f, Qf > a . For these Q' we stop the bisection process. For the
10
rest, we continue until we first arrive at a cube Q' such that 1 f ,f > a
I Q
at which point we stop.
The cubes Q' at which we stop are then our Qk. By construction
1
f > a. Let Qk be the cube containing Qk which was bisected
IQkI fQk
Qk. Then and since we did not stop at k,
1IkIf1k f < a. It follows that
MULTIPARAMETER FOURIER ANALYSIS 49
1Q k 1
f f < 1Q k
k1 1 f f< 2na ,
Qk
proving (1). Notice that (3) follows from (1) because 1Qk' < a fQk f so
summing on k, we have
('
IUQk1<a Y f f
Qk Rn
Finally, (2) follows, since for each x I UQk, x belongs to a sequence

of cubes Ck whose diameters converge to zero and such that
It follows from Lebesques theorem on differentiation of
Ck fc k f < a.
integrals that 1(x) < a a.e. for such x .
We all know how important the maximal operator of Hardy-Littlewood
is in the subject of Fourier analysis. This is the operator given by
Mf(x) = sup 1 If(t)(dt

r>o IB(x;r)1
B(x,r)
Going along with this we also define
MSf(x) = sup 1 JI If(t)I dt

xEQ dyadic cube IQI Q
(Recall that a dyadic interval of RI is one of the form [j2k,(j+1)2k]

j,k E Z and a dyadic cube is a product of dyadic intervals of equal length
recall also the basic property of dyadic cubes-if Q1,Q2 are dyadic
either Q1 f1Q2 = 0, Q1 C Q2 or Q2 C QI .) Then the following simple
lemma sheds some light on the relationship of the Calderon-Zygmund
lemma to the Maximal Operator.
50 ROBERT FEFFERMAN
LEMMA. Let f > 0 c L1(Rn) and a > 0. Let Qk be Calderon-Zygmund

cubes as above, and let Qk denote the double of Qk. Then there exist
positive numbers c and C such that
(1) UQkC 1Mf > ca(
(2) UQk ) (Mf > Cal
(3) Furthermpre, if Qk are dyadic, then UQk )1MSf > Cal.
Proof. (1) Let x c Qk. Then there exists a ball B(x;r) such that
x c Qk C B(x;r) and B(xx;rl < Cn. Then
k
IQkl 1
Mf(x) >_ 1 Ifl > IfI > 1 a
B(x;r)
,/
B(x;r)
(IX)l) IV Cn
Qk
(2) Let x / UQk. Let r > 0. Then we estimate
ff I f+ I f
B(x;r) B(x,r)fld[UQk] Qi
< ajB(x;r)I }
Qi
< alB(x;r)I 2na Y IQjI
Key point : if Qj fl B(x ;r) 0 then Q) C B(x;l Or) so that
I IQiI <CIB(x;r)I and f < Cna IB(x;r)l .
QiflBj$
That is, Mf(x) < Cna.
The use of the Ca1deron-Zygmund lemma is apparent in the Calder6n-

Zygmund Theorem on Singular Integrals: Let X and N be Banach
spaces and let B(X,N) denote the bounded operators from X to N . Let
K: Rn x Rn/lx=yl -, B(X,N) satisfy
(1) B(x <

Ihl1
Ix- for IhI < Ix-yI
n++8 2
and for some S > 0.

and
(2) if Tf(x) = fRn K(x,y)f(y)dy for f c LP(X), and suppose for some
p0> 1 IITfII << CIIf1I

LPO(N) LPO(X)
Then, for T we have
I#XI ITf(x)IN > all < IIf IILI

Q (X)
and
IITf1ILP(N) < CPIIfIILP(X) for 1 0 and f c L1(X). Set

1
f dt if x cQk,
IQ kI
g(x) = Qk
f(x) if x / Qk .
and b(x) = f (x) - g(x) . Then
Il ITg(x)IN > a1I < IlgllL < a: IIgiILI SI: IIfIILI(X)

PO
L (X) - (X) <
- a
As for Tb(x), suppose x I UQk.
52 ROBERT FEFFERMAN
Let bk(x) = XQk(x) b(x) ; f k

bk(x) dx = 0. Then
Tbk(x) = J K(x,y)bk(y)dy .
Qk
Let Vk be the center of Qk; then
J K(x,yk)bk(y)dy = K(x,Vk) bk(y) dy = 0

J
Qk Qk
so
Tbk(x) = IK(x,y)-K(x,yk)Ibk(y)dy
J
Qk
and
diam(Q )s
ITbk(x)IN <_ +3 Ilbkll Lt (x)
Ix-yk1
summing over k we have
diam(Qk)s
Ilb k II t dx < CII f II t
f IT(b)(x)IN <k Y. f dist(x,Qk)n , L L (X)
X/Uijk x'!k
Thus
I{Tb(x)IN > all S Okl +1 llflll <_ SL IIflll .
From this weak (1,1) estimate, interpolate to get the LP result.

We now quote some important examples:
1. Classical Calderon-Zygmund Convolution Operators. Here If = f * K

where K(x) is a complex-valued function satisfying
MULTIPARAMETER FOURIER ANA LYSIS 53
(a) IK(x)I < C/ Ixln
(0) I K(x)dx = 0 for all 0 < P1 < p2

P1<_lxIp2
and S
(y) IK(x+h)-K(x)I < C IIh IhI < IxI
2
The Riesz transforms Ref = f * xJ/Ixln+l are especially important since

they are related to HP spaces and analytic functions.
For a Calder6n-Zygmund singular integral T , it is easily seen to be
bounded on L2(Rn), since K(e) e L. Also, since T* is also a
Calderon-Zygmund singular integral, T is bounded on the full range of
LP(Rn), 1 0. Then
Rn
g2(f )(x) =J If *c1it(x)12 dt

t
0
if *
S2(f)(x) t(y)I2 anal
=ff
r(x)
where r(x) _ {(y,01 Iy-xI < ti. Then it is a basic fact that 'IIS(f)IILp <
CPIIfIILP and Ilg(f)IILp <C when 1 -cPIIfIILp

54 ROBERT FEFFERMAN
Now take S(f) . We want to point out here that S is a singular

integral. In fact define K:Rn F. L2(F(O); dydt) by K(x)(y,t) = &t(x-y).
Then
S(f )(x) = If * K(x)IL2(r,dtdy/tn+1)
and K satisfies
IK(x+h)-K(x)I < C IhI <2 IxI .

IxIh' 1
Also by a Fourier transform argument IISfII < cilfiI so the

I- 2(R n) LZ(R n)
Calder6n-Zygmund theorem applies. In fact, the adjoint operator also maps
Lp(L2(P)) - Lp(Rn), 1 < p < 2 , because it is also C-Z, so again this
explains why we get boundedness of S on the full range 1 2,
O(x) = 0. Then define K:Rn+, L((0, 00);dt) by K(x)(t) = -ot(x) _
t-no(x/t). Then
IoxK(x)(t)I = It (n+t)0k x < CV 00 fl

Ixln1
+i
(since if t > IxI/2 , V (x/t) = 0 ).

Again
L <C InI1
IK(x+h)-K(x)I, if IhI < IxI
IxI 2
and we also have IIf * K11L(L<_ CIIf II since If * ot(x)I < II-kII1 IIf II
L
Then Mf(x) -If* K(x)I , so M is bounded on Lp(Rn) , p'> 1 and
weak 1-1.
4. The Estimates for Pointwise Convergence of Singular Integrals

on L1(Rn) . Suppose that K(x) is a classical Calder 6n-Zygmund
kernel and let KE(x) = K(x) )(IXI>E(x), for e > 0. We are interested in the
existence a.e. of lim f * KE(x) for f e Ll(Rn). In order to know this, it
E-40
is enough to show that T*f(x) = sup If * KE(x)I satisfies the weak type
E>0
estimate IIT*f(x) > all < a f If I . It turns out that by using the Hardy
R
Littlewood maximal operator it is not difficult to prove T*f(x) < CIM(Tf)(x)
+ Mf(x)l which immediately gives the boundedness of T* on Lp(Rn) for
p > 1 . However, it fails to give the weak type inequality for functions on
LI(Rn). This inequality follows easily from the observation that T* is
a singular integral.
Let
1 if IxI < 1
O(x) c C (Rn), O(x) =
0 if lxI >2
and set K,(x) = K(x)[1 -0 QL)] Then IK,(x) -K(x)I

I In XIxI<2e (x)
so that T*f(x) 0
sup If *kI is weak type (1,1). In order to do this let H:Rn , L((0,-);de)
E>0
C InI
be given by H(x)(E) = II (x). Then IH(x)-H(x+h)IL <
1
and H is
Ix
bounded from L2 - L2(L), so H is weak 1-1 .
5. The Maximal Function as a Lit tlewood-Pa ley-S tein Function. Let

f c L2(Rn), f(x) > 0 for all x. Use the Calderon-Zygmund decomposition
with a = Ci , j c Z for some C > 0 sufficiently large, to get (dyadic)
cubes Qk where 1 f f - C). Define f .l as in the Calder6n-Zygmund
IQ)kI Q)k
decomposition,
1
IQ,
f if x c Q
k
Qjk
f(x) if x U Qk
and A f =fj+1-fj, then observe that:

i
56 ROBERT FEFFERMAN
(1) jf lives on k Qk and has mean value 0 on each Q.

(2) Aif is constant on every Qk for i < j.
+Do
(3) fJ - 0 as j,--o and f.J f as j -.+00 so f = F Af.
j=-
From (1) and (2) it is clear that the A f are orthogonal so that
i
1/2 1/2
If IIL2(Rn) = (Y IlojfIIL2)= II (Y lojf(x)12) I1L2
Finally, observe that the square function (I IA f(x)12)1 /2 is

j
essentially just the dyadic maximal function. In fact, if CJ << MS(f)(x)
then x e Qk for some k and Aj(f Xx) 'ti CJ. It follows that
1 /2
Iojf(x)12) > cMSf(x)
C i 11
Before finishing this section, we shall need estimates near L1 for

the maximal function.
If Q denotes the unit cube in Rn then for k a positive integer
J Mf(log+ Mf )k-1dx < 00 if and only if r If l(log+ If I)kdx < 00 .

Q Q
The proof runs as follows': If f e L(log+ L)k then
IlxIMf(x) > all < C J f(x)dx

lf(x)l>a/2
and so
IIMf II L(log L)k-1 <
f(log a) k-1 a J lf(x)Idxda < If(x)I f Q (log a)k-ldadx

J
1 If(x)l>a/2 Qp 1
llf II
L(Iog L) k
Conversely, (Stein) Calder6n-Zygmund decompose Rn at height a > 0.

We have
J f(x)dx < J f(x)dx< Y J f < Ca j IQkl < CaI{Mf > call .

k Qk k
M&f(x)>Ca UQk
This yields
00
5 If(x)I(log+Msf(x))kdx < ra J If(x)ldx(log

a)k-lda
Qk
I M6f(x)>Cna
00
<
J
1
r a
a I
MSf(x)>Cna
If(x)Idx (log a)k-1da
< f Mf(x) [log+Mf(x)]k-1dx
Qk
2. Multi-parameter differentiation theory

During the first lecture we discussed some fundamentally important
operators of classical (and sometimes, not so classical) harmonic
analysis: the maximal operator, singular integrals, and Littlewood-Paley-
58 ROBERT FEFFERMAN
Stein operator. These operators all had one thing in common. They all
commute in some sense with the one-parameter family of dilations on Rn,
x - fix, 8 > 0. The nature of the real variable theory involved does not
seem to depend at all on the dimension n. In marked contrast, it turns
out that a study of the analogous operators commuting with a multi-
parameter family of dilations reveals that the number of parameters is
enormously important, and changes in the number of parameters drastically
change the results.
Let us begin by giving the most basic example, which dates back to
Jessen, Marcinkiewicz, and Zygmund. We are referring to a maximal opera-
tor on Rn which commutes with the full n-parameter group of dilations
(xl,x2, ,xn) - (Six1,82x2, ,Snxn), where Si > 0 is arbitrary. This
is the "strong maximal operator," M(n), defined by
M(n)f(x) = sup 1 If(t)I dt

xER IRI JR
where R is a rectangle in Rn whose sides are parallel to the axes. Un-
like the case of the Hardy-Littlewood operator, M(n) does not satisfy
IIxIM(n)(f )(x) > al I < a Ilf II I(Rn

L )
For instance when n = 2 and when f8 = 8-2X(Ixl I<8/2)x (Ix2I<8/2)

then for 1x 1 1 , Ix2I > 28,
M(2)(f8)(xl,x2) =M(1)(XIxII<8/2)(xl)M(1)(XIx2I<8/2)(x2) ti
I lI I I
and
I#x EQ0IM(2)(f8)>aI I ? I#xI Ix1I Ix2I < a , and 8< 1x11<1 II ti a log !
if a=1/8.
If we have a weak type inequality, we must have
i1M(2)(fg) > all Ilfsll

a
so that llfsll > c log . and the smallest Orlicz norm for which this holds
is the L(log L) norm. A similar computation in Rn reveals that for
M(n) to map LD boundedly to Weak LI we must have Lip C L(log L)n-I
The next theorem shows that indeed M(n) does indeed map L(log L)n-t
boundedly into Weak LI .
THEOREM OF JESSEN-MARCINKIEWICZ-ZYGMUND (1935) [1]. For

functions f(x) in the unit cube of Rn we have
lixeQo,M(n)(f)(x)>ail a llfll L)n_t

L(Iog (Q0)
The proof is strikingly simple. Define Mi to be the 1-dimensional

maximal function in the ith coordinate direction. Consider the case
n = 2, which is already entirely typical. Let R be a rectangle contain-
ing the point (x t , 3E2)1 say R = I x J . Then
R
ffif(xt.x2)idxidx2 = III f(th ff(xi.x2)dx2)dxi
(2.1)
R I J
111 lf(xl,x2)ldx2 < Mx2f(xt,-2)

J
so (2.1) is
< lIl M2f(x1,x2)dxl <Mx1(Mx2f)(x1 ,x2)

60 ROBERT FEFFERMAN
Thus, for all (x ,x2) 'r QO'
M(2)f(x1,x2)
<- Mx Mx2(f )(xl,x2)
1
We have seen that Mx2 maps L(log L)(Qo) boundedly into L1(Q0) so
that
IIMx2fIIL1(Q0) - CIIfIIL(log L)(Q0)
and finally
I1xEQoIMx1Mx2f(x)>aII < a'IIMX2fIIL1 I1fIIL(log L)(QD)

(Q <a
Now, this method of iteration in the proof above gives sharp estimates
Now,
for M(n), and it may be suspected that the whole story of the harmonic
analysis of several parameters can be told by applying this iteration
technique. That this is not the case should become clear as this lecture
proceeds. We want to describe some of the multi-parameter theory and to
do this, let us begin with maximal functions. For many years after the
Jessen-Marcinkiewicz-Zygmund theorem, there was no machinery around to
treat problems here, and then, only fairly recently, two such machines were
created. The one we describe here proceeds by means of covering lemmas
while the other, due to Nagel, Stein, and Wainger, which Wainger has
described in detail, uses the Fourier transform [2]. Though the two
methods seem totally different on the surface, they are really quite closely
related and have in common the main theme of reducing higher parameter,
complicated operators to lower parameter simpler ones, which are already
well understood.
The model for our method is the following, where the operator. M(n) is
M(n-1 )
controlled by
COVERING LEMMA FOR RECTANGLES OF THE STONG MAXIMAL

OPERATOR [3]. Let 1Rk1 be a given sequence of rectangles in
Rn C B(0,1) whose sides are parallel to the axes. Then there is a subse-
quence 1Rkl of 1RkI so that
(1) IUkkI > cnIURkI

1
n-1
/
(2) IleXp(I X'lk) II'LI(B) <
C
Before we prove this theorem, let us show that it implies the Jessen-
Marcinkiewicz-Zygmund result. Let a'> 0, and for each point
x e IM(n)f(x) > at there is a rectangle Rx containing x with
(2.2) LIfI>a.
I
Without loss of generality we assume URx = URk where Rk are certain

if the Rx's . Apply the covering lemma to get Rk with properties (1)
and (2) above. Then by virtue of (1) we need only show that
IUkkI <_ a Ilfll L(log L)n-I(Qo)
By (2.2), IRkI < a fW IfI and summing we have
IURkI r IfI 1" X,k < IIf IIL(Iog L)n-1 xRR

exp(LI /(n-1))
a k
Qo
Now, let us prove the covering theorem. We shall proceed by induction

on n. Assume the case n-1 . Let R1,R2, ,Rk, ... be ordered such
that the xn side length decreases. For a rectangle R, let Rd denote
the rectangle whose center and xi side lengths, i < n , are the same as
those of R , but whose xn side length is multiplied by 5. Then we
describe the procedure for selecting the Rk from the Rk: Let K1 = R1
Suppose R12, ')Rk have already been chosen. We continue along
the list, and each time we consider the rectangle R we ask whether or not
62 ROBERT FEFFERMAN
IR n [U(Rj)d]I < z [RI
where the above union is taken over all the Rj , j <k for which Rj n R 4 0.
If the answer is no, we move on to consider the next rectangle on the list.
If the answer is yes, we make the rectangle R = Rk+l , and start the
process over again.
Now, we prove (1) as follows: If R is an unselected rectangle, then
Rn U (Rj)d >21R1.
kinR=c
L before R
kj
Let us slice all rectangles with a hyperplane perpendicular to the xn

axis. Then if slices are indicated by using S's instead of R's ,
Isn[U(Sj)d]I > 2 ISI
so that M(n-1)()(U 1) > on UR j , where M(n-1) is acting in the

(W
jd 2
x1,x2, , xn_1 coordinates. By the boundedness of M(n-1 ) on, say, L2
(by induction) we have
IURjI < CIU(Rk)dl < C' I IRkI < C'IURkl
To obtain (2), notice that the Rj's satisfy
Rk n U (Rj)d
Wjnkk
j<k
If we again slice with a hyperplane perpendicular to the xn axis,
kn Ujl <2 (kl

j<k
so that, if Ek =-9k - lUk^`Sj we have IEkI > 2 1. and if

4) t L(log L)n-1(So) we shall show that
(2.3) <
L)n-1
X k (k C II(bII L(log
so
which will give
\1 /(n-1)
f exp X
k/
1
S
0
Integrating this estimate in xn finishes things.

To obtain (2.3) we write
fI
S0
Xgk0dxl..dxn-1 < I
k J 0 <C I
'k
'
IEkI-L
k
fO
<
<C
f
S0
M(n-1)(4))dxldx2...dxn-1 <C1II0II
L(log L) n-1
Notice that in our argument above the slicing was the most important
idea. If you try the proof without it, you will not wind up with the estimate
you want, on the exp( )1 /(n-1 )norm, but rather on the exp( ) 1 /n norm
instead. Also the slicing is the mechanism by which we control M(n) by
the lower parameter operator M(n-1) and here this enables us to proceed
by induction. Of course, in the end the theory of the boundedness of M(n)
had been known for some 40 years before the covering lemma. But the
lowering of the number of parameters, and the induction procedure will be
used in what follows as the key ingredient to prove new theorems.
64 ROBERT FEFFERMAN
To illustrate the method of the machine, we consider a maximal

operator whose relation to multiplier theory will be studied below.
Suppose in R2 we consider the class 91 of all rectangles of arbitrary
side lengths which are oriented in one of the directions ek = 2-k,
k = 1,2, , measured from some fixed direction, say the positive
x-direction. Define the maximal operator m by
mf(x) = sup 1 f If(t)I dt

xEREQ IRI
R
This operator was considered following the ideas of the covering approach
outlined above by Stromberg [4] and Cordoba-Fefferman [5]. Somewhat
later Nagel, Stein, and Wainger [6] used Fourier transform methods to
extend the result we shall discuss below.
What we prove is that
IImf(x)>afl <
a IIfIIL2fR21
The proof consists of showing that, given a sequence of rectangles 3Rkl
belonging to S?j , there exists a subfamily {RkI such that
(1) IURkI > cIURkI

and
(2) k11L2(R2) < IURkI1/2
ti
To prove this we give a rule for selecting Rk, given that we have
k
already selected Rj for j < k. Assume that the Rk all have their
longest side in a direction in the 1st quadrant, and are ordered so thai
their longer side lengths are decreasing. Then consider the rectangle R
ti
following Rk_1- Consider in particular
IRI Y IRURjI = IRI J I XW1dx .

j<k R j<k
If this is less than 1/2 , select R as Rk. If not go to the next

rectangle on the list, and apply the same test to it. In this way we obtain
the desired Rk. Notice that
II y XWkI2 = f X
j,kXk,Xkk =
2j
1
j
j<k
xAk +
k
IRkI
(2.4)
2 L.r fXWi + I IRkI < 2 Y IRkI < CIURkI

k 1r j<k k
k
This is (2). To show (1), we let R be some rectangle which was un-
selected. This implies that
IRI
I
before R
IRfRjI > 1
.
Now draw the following picture. Let S be the envelope rectangle to

R whose sides are parallel to the coordinate axes.
66 ROBERT FEFFERMAN
Then by dilating S if necessary, we can assume that Rj is centered at

the same center as S. Now the point is that
(2.5) iSI - c JRI
In fact,
IRjnSI hH/B' h
IS 1
hL - LO'
and
IRjnRI ti h h O'-9 _ h
IRI h e(e'-0)
and our inequality (2.5), taking into account that f ti L, is
I > I or 9'- 6 > c O'

0' - (0'-0) -
and this in our case of Bk = 2-k is valid with c = 1/2 . (If ek = (1 _ )k,
c = e.) Then summing over j in (2.5) we have
I$11 fL
S
& be re
R
. I I
JR before R
2
c
in other words
M(2)(I X, } > 2 c on URj
and by the boundedness of M(2) we see that
IUR;I _CIIYX&112<c'IURji
by (2.5). We have controlled m by M(2) here in just exactly the way

M(n) was previously controlled by M(n-I). And while m is a 3-parameter
maximal operator, M(2) is a 2-parameter one.

Finally, we should note that, as before, this covering lemma implies
the maximal theorem for m as claimed. In fact, suppose we have shown
that given !Rkl there exists lkk} a subsequence so that
(1) IURkj > cIURkl
(2) IILXR IIp.<cIURkiI"P', (here P+1 =1).

k p
Then
IIfIIp1P
;mf> all < CC .
In fact we have, by definition IRkl so that tmf> of C URk and

1 fR If I > a for all k. By the covering lemma, select the class
I I kl.
12k
k
Then Ikki < a f& IfI and so
Iv kl < a ff 1 X" k `< IIf IIpII Y )<kkilp,
a
IIfIIp}uRk)I/P.
a IIflIpiURkII"p <_ C
and the estimate on Ilmf > all follows by a division of both sides by
I U'k I I /P'
Our last topic for this lecture will be the so-called Zygmund Conjec-
ture. I believe it was Zygmund who was the first to realize the difference
between the one-parameter and several parameter harmonic analysis.
Particularly, he remarked that in differentiation theory a "big picture"
was evolving. He considered n functions (A I1 '021 ,'On of the positive
real variable t , with each Oi(t) increasing and the family of rectangles
68 ROBERT PEFFERMAN
1Rt{t>o given by Rt = iIIt t2(t) , t2 t)J . Form a maximal operator M

defined by
f lf(x+y)I dy
M(f )(x) = u JR._
Rt
Then M is of weak type 1-1 , just as in the special case of the Hardy-
Littlewood operator where fi(t) = t. Zygmund noticed that the proof of
this was virtually the same as the Hardy-Littlewood theorem. All one had
to do was to prove a Vitali-type covering lemma for Rt's using the fact
that if is the class of all translates of the Rt and if R,S r 3 and
R R S 10 and if R corresponds to a bigger value of t than does S ,
then S C k, the 5-fold dilation of R. Next, he considered the collection
of rectangles Rs,t , s,t > 0 where
r t t >' L c(s,t) i(s,t)

E flx l -2'2J
=[-2'2]
Rs,t -2 L2
where 0 is a function increasing in each variable separately, fixing the
other variable. In other words, Zygmund next conjectured that since 1
is a 2-parameter family of rectangles in R3, the corresponding maximal

operator, which we shall call MZ should behave like the model 2-parameter
operator M(2) in R2:
11Mzf(x) > a, JxJ < 1 #{ < a 1{"L(1og L)(Ixl<1)
Not long ago, using the methods we have just discussed, Cordoba was
able to prove this [7]. Let us give the proof. Suppose IRkl is a
sequence of rectangles with side lengths s,t, and 0(s,t) in R3. We
must show there exists a subcollection 1Rk4 such that
(1) IURkI > c}URkf
(2) 111 kllexp(L) < C

To prove this, order the Rk so that the z side lengths are decreasing.
With no loss of generality, we may assume that IRkf1 ['Uk Rj11 < 2 (Rkl ,
that there are finitely many Rk and that the Rk are dyadic. (In fact, we
may assume this because if IRI fR IfI > a for some R e 91 containing x ,
then there exists a dyadic R1 whose R1 (double) contains x such that
IR' 1R1 +fj> C .) Now let R1 =R and, given R1, ,Rk, select Rk+l
as follows: Let Rk+1 be the first R on the list of Rk so that
1 fexP(x)dx<c.
(R j<k j
Rk be R1,
ti
We claim that the Rk satisfy
, RN and let
`x
j = Rj
i exp(I X
- U Rk. Then
) < C'. To see this let the
k
k>j
N N (' N-1
I exp X = r exp Xk )dx +J ex
X+...+ Jexxk=1
J 1
Ukj il N 'kN-1 t1
and
j
fexp X <C fex(kk=1
k
k<j }
so we have
Now let us show that 1URjI > clURjl . Let R be an unselected

rectangle. Then
70 ROBERT FEFFERMAN
1 fex(x)dx>C
1R I
ti
where the sum extends only over those chosen R k which precede R.
Let us slice R with a hyperplane in the xl,x2 direction. Call S,Si the
slices of R and R Then
1 fexP())dxidx2>C.
I S1
S
ti
(Again we sum only over those Sj which appear before S.) Now, each
Rj appearing before R has the property that its x3 (or z ) side length
ti
exceeds that of R. It follows that each corresponding Si . has either its
xl or x2 side length longer than that of S. Call those j havyng
longer x t side length than x 1 length of S of type 1, and the other of
type II. Put S = I x j . Then
Jbexp(,.+jX
X.9i)dxidx2 = I1111J1 f ,,dxldx2
C < II1 f f exp( I J II I
IxJ
so it follows that
III
I
exp
C X dxt -Lfexp
IJI
J
X-,
;}
dx2
MxI[exp(YXWAMx2Cexp(lXk A >C
on URA; hence
URAC MxI [ex()]> U Mx[exp(X)]>/

LL \\
so
JURjI < C' Ilexp(y' Xj )IIL' < C_IUkjI
So far what has been done suggests the following general conjecture
of Zygmund which says: Let 0, i = be func-
tions which are increasing in each of the variables ti > 0 separately.
Define a k-parameter family of rectangles
Rt1 ,t2,...,tk
and a maximal operator on Rn by

i=1
2,
0(tl,...,tk) +S6(t1,...,tk)
MJ)(f)(x) =
t1 t 2 supsk >o Rt1,...,tk
1 fR
t1, ,tk
lf(x+y)ldy
Then
lUxJlxI < 1,M,(D(f)(x) > all <Q lIfIl

L(log L) k-1
Quite recently a beautifully simple counterexample to the general con-

jecture was given by Fernando Soria of the University of Chicago [8].
Soria's counterexample was: in R3, consider all rectangles 6 of the
form s x tc1(s) x tc2(s) where Ebi are increasing on [0,1]. In fact we
claim that given any small rectangle R with side length x,y and z
such that x < y < z < 1 in R3 , there exists a rectangle S E 6 such that
R C S and ISI < CIRI . Clearly, then % cannot be any better than the
3-parameter operator M(3). To do this let (k1 and 02 be increasing,
0, and continuous on [0,11 so that on each interval 2-(k+1) < s < 2-k ,
2(s)
() takes every value between 1 and C.2(k+1) . Then given three
1s
side lengths p1, p21 p3 < 1 of R , simply choose s according to the
72 ROBERT FEFFERMAN
following: say 2-k-1 < p1 < 2-k. Choose s e (2-k-1, 2-k) satisfying
'Ys P3 (since 1 < P3 < < 2k+1 we can do this). Then choose t
1(s) P2 - P2 - P1
so that t(Ai(s) = p2. This guarantees t-02(s) = p3, and we are finished.
01(2(ss))
We can make assume every value between 1 and C2k on
[2-k-1 2-k] bt letting 01(s) = e-1 /s and 42(s) = 1(s) on
2-k-1, 2 -k] and
L2 .
952(s) - 'k1(2 , 2-k-1 for all s e L2-k-1, 2 2 kl
3. Multiparameter weight-norm inequalities and applications to multipliers

In this lecture we want to describe further applications of the ideas
centering around the covering lemma for rectangles previously described.
We shall begin with more about maximal operators, and then move on to
multiparameter multiplier operators, and the connection they have with our
maximal functions.
The first topic we take up is that of classical weight norm inequalities,
which have proven of enormous importance throughout Fourier analysis.
Here, we want to know which locally integrable non-negative weight func-
tions w(x) on Rn have the property that some operator T is bounded
on Lpw(x)dx . The most basic examples are the Hardy-Littlewood maximal
operator, and Calderon-Zygmund singular integrals Tf = f * K. The theory
was developed in R1 by Muckenhoupt [9] and Hunt, Muckenhoupt and
Wheeden [10], and in Rn by Coifmart and C. Fefferman [11]. We will
present only a small segment of that theory now and list some relevant
facts for which the interested reader should see the Studia article of
Coifman-C. Fefferman [ii].
It is no coincidence that the class of weights w for which the Hardy-
Littlewood maximal operator is bounded on LP(w) is exactly the same as
the class of w for which all Calderon-Zygmund operators are bounded on
LP(w). This is the so-called AP class of Muckenhoupt.
A nonnegative locally integrable function w(x) on Rn is said to

belong to AP if and only if for each cube Q C Rn
p-1
w-1/(p-1)dx <C.
IQI IQI ('
fwdx)
Q Q
The smallest such C is called the AP norm. We say that w e A if and

only if, whenever Q is a cube and E C Q, if IEI/IQI > 1/2 then
w(E)/w(Q) > -q for some rl > 0.
Let us list some properties of AP classes:
(a) If p > 0 and w e AP then pw a AP with the same norm as w.
(/3) If w e AP and S > 0 then w(3x) a AP with the same norm as w.
AP,
(y) If w e AP then w-1 /(P-1) a where 1 1, = 1.
P P
(8) If w e AP then w e A. In fact, if w e AP by (a) and (13) it is
enough to show that if IQI = 1 , and fQ w = 1 then IEI > 1/2 implies
w(E) > r).
(For, in general if Q is arbitrary of side 8 and I E I > 1/2IQI , consider
w(8x) on Q/8 and multiply w(8x) by the right constant p to have
fl pw(Sx) = 1 . Then pw(Sx) on E/8 would have
/8
[pw(Sx)](E/8) > w (E) >
rl
[pw(bx)](Q/8) w(Q)
But by the AP condition,
r)_1 p-1
< (f W-1/(P-1) < rw-1/(P-1) <C
JE E JQ
and so
w>C.
JE
74 ROBERT FEFFERMAN
So far all the properties of AP weights listed are obvious and follow
straight from the definitions. There are some deeper properties which
though not difficult to prove are not immediate.:
(E) If w c AP then w satisfies a Reverse Holder Inequality:
1/(1+8)
IQI f
Q
w1+s
<
C8
IQI
fw
Q
for all cubes Q with 8 > 0. The constant CS may be taken arbitrarily
close to 1 as 8>0.
(C) From (E) it is immediate (see also (y)) that w c AP implies w c Aq
for some q < p.
(r)) If f is a locally integrable function in some LP space and
0 < a < 1 then (Mf )a c A 1 , i.e., M((Mf )a)(x) < C(Mf )a(x) (for w c A
implies w c AP for all p > 1 ).
To prove this let f e Lp(Rn) be given and a c (0,1). Let Q be a
cube centered at z, and Q its double. Then write f = XQf + XcQf =
f1 + f0. We must show that
IQI I M(f 1)a dx < CM(f )a(z )

Q
and
1 M(fo )adx <- CM(f)a(x) .

IQI .
Q
As for the first inequality,
If
)1/a
II ff1.
M(f1)adx iIdx
J J If
Q Z1
(This is an immediate consequence of the weak type estimate for M on

L1 .) This shows that
f M(f1)adx < <M(f)a(x)

iQl ICI J if I
Q
As for the second inequality, choose a cube C centered at z such that
ICI f If of dx > 2 M(f d(i)

C
Now, if x f'Q then any cube C' centered at x which intersects c is

contained inside a cube of comparable volume centered at i so that
Ilfol < A Id flfoldx.

IC II
C C
We have proven that for all x EQ, M(foxx) <A ICI fc Ifoldx so that
I1 G
lf <AaM(f)a(1) .
fM(fo)adx<Aa(_i_
Q C
Let us begin to discuss the weight theory by showing that the Hardy-
Littlewood maximal operator is bounded on LP(w) if and only if w e AP,
1 < p < - [121. In the first place if f = w-1 /(P-1) XQ and if M is
bounded on LP(w) one sees right away that
p
w(Q)
IQI f Q
w-1/(P-1)dx
<C f
Q
w-P/(P-1)wdx
or
76 ROBERT FEFFERMAN
w(() 1 fw-l /(P-1) <C

1QI IQI
Conversely, assume w c AP. Calder6n-Zygmund decompose f c LP(w) at

heights Ck where k c Z , and C is large (to be described later) and
get Calderon-Zygrpund cubes {Q } 1 so that
IQ
f1
f- Ck and
k
Qj
(y is a large constant dependent only on n ). Then
P
1
f (Mf)Pwdx<C'Y w(QkCk<C'I w(Qk) (TQkl J
Rn k,j k,j j
Now w c AP w c A therefore w(Qk) < C"w(Qk) and so the above

J J
expression is
P
<C- W(Q) k (Ql) 1
fo odx < by the AP condition on w
1
k, j IQk Qk
1
p
(*) < C "" o(Qk) 1 f (f o 1) odx where o = w-i /(p-1)
kj o(Qk) k
Qi
So far we have used only arithmetic. Now we come to the main point.
If Ek = Qk - U Q then choosing C large enough insures that
Q>k
IEkI > IQkI , and since a c AP u .E A we have o(Ek) > rlo(Qk)
J 2 j J J
and so
(*) < C I o(Ek) (0,(Qjk)

I (fo I)do < C Mp(for) do
JQig R
n a
where Mo(f )(x) =scupp g(Q) fQ IfIda. Now the same proof that works to
show that the Hardy-Litt lewood maximal operator is bounded on LP(Rn)

proves that if or satisfies a doubling condition, then for p > 1 , Mo is
bounded on LP(do).
It follows that the second inequality is
r
<C fPol-Pdx = j fPw dx .
Note that the operator M11 f(x) = sup 1 f IfId and its boundedness
x(Q (Q) Q
on LP(d) enter in a crucial way the proof of the weight norm inequalities
for the Hardy-Littlewood operator. This operator M is interesting in its
own right, since it is natural to ask what happens if we replace the God-
given Lebesgue measure by another measure d.
In fact, if is any measure finite on compact sets and if, in the
definition of M we insist that the balls be centered at x then M is
bounded on LP(), p > 1 . The proof of this remarkable theorem relies
on a refinement of the usual Vitali covering lemma due to Besocovitch.
We should also remark that the original proof of the weight norm inequali-
ties for M on Rn made use of M as well. In fact, if w c AP it is not
hard to see that
M(f )(x) < CMw(fP)I /p(x)
78 ROBERT FEFFERMAN
(Indeed,
1 /p /p11
(ffPwdx) (JrW_P'IPdXV'
I f f dx = I1 II
ffw'/Pw4/Pdx<_L 1Q1
Q Q Q Q
p (p-1)/p
w 1/p I fw_h/(P_1)dx) < CMw(fP)1/p(x) . )
1Q1 ( JfPw)
Q
I
Q
The proof would be complete if Mw(fP)1/P were bounded on LP(w). Un-

fortunately, Mw(fP)1/P is bounded on Lq(w) only for q > p. But if we
notice that w e App w e Ap_f then the proof is complete.
Anyway, it is clear that the operator M arises naturally. Next we

shall study its n-parameter variant M(n) just the strong maximal operator,
only taken relative to the measure
M()(f)(x) =sup R) r IfI d

R
R a rectangle with sides parallel to the axes. Unlike the case where
n = 1 , restrictions must be placed on whether or not R is centered
at x .
The following result gives conditions on which are rather unre-
strictive, and which guarantee the boundedness on LP(IA) on Mn) [13]:
THEOREM. Suppose is absolutely continuous and non-negative on Rn,

and that the Radon-Nikodym derivative of , w(x) satisfies an A
condition in each variable separately, uniformly in the other variables.
Then M) is bounded on LP(d) for all p > 1 .
We require the following lemma.

LEMMA. If is uniformly in A in each of the variables

separately, then w satisfies the following: If R is any rectangle with
its sides parallel to the axes and E C R is such that ELI > L, then
IRI
fE w
>q, for some q>0.
IR w
Proof. The proof is by induction on n. Assume this for n -1 . Consider
a rectangle R as above, R= I x J where I is n-1 dimensional and J
one-dimensional, and a subset E of R such that
For each x e I let Jx = i(x,y)IyEJI be a vertical segment. Since

Ei
> it is easy to see that for x in a set I' of measure > 100 III we
R
2
have IE n JX I > 100 IJX1 Now.since w is uniformly A in the xn
variable,
(3.1) wdxn>77 wdxn , x e I' .

J
JX,nE
But also if we fix any xnEJ then
(3.2) fwdx'? il' J wdx'

I I
by induction. It follows by integrating (3.1) in x' c I' that
J wdx>r) f w dx
E It'JxJ
80 ROBERT FEFFERMAN
and integrating (3.2) in xn e J gives,
fw dx
f wdx>71'
I'xJ IxJ
which shows that- fE w>rfr?'fRw and proves the lemma.
Proof of the Theorem. We prove a covering lemma, namely, if 1Rk1 are

rectangles with sides parallel to the axes, there exists 1Rk1 so that
w(URkL < Cw(URk) and
III XkIILp(W) <

Cw(URk)1/p'
To prove this order Rk by decreasing xn side length, and then select a

rectangle R when
IR n [U(Rk)d]I < 2 IR I
where the union is taken over all those k such that Rk precedes R and
RknR=0.
Now if we slice R , an unselected rectangle, with a hyperplane per-
pendicular to the xn direction we have
Is n [U(k)d]I > 2 I,SI -= w(S n [U('k)d]) > (S)
so that Mwn-1 71 on URk. By induction, w(URk) < Cw(U(Rk)d).

k
Now the Rk have disjoint parts property with respect to dx and so they
have this property with respect to w dx also. It follows that
w(U(Rk)d) < 5' wOk)d < C

L w(Rk) < C'w(URk) .
Therefore
w(URk) < Cw(URk) .

Now to estimate III let us slice the Rk with a hyperplane

perpendicular to xn , calling the slices 9k* Then 2
so since w fA` in we have w('Sk - U Sj) > r7w(Sk) (if
j<k
w(E) = JE and we test
fRn-1I Xkow Ir Owdx < C y w(tk) 1

w (gk)
f
k
ow dx
< C J M(n-1)(O)wdx.
w
U SjBy induction
k - -9k - j<k is bounded on LP(w) so this
last integral is
IIX1j IILP(w)<
LP(w)
pth
This shows that III X3rk1I (w) < Cw(USk)l /P'. Raising this to the
LP
power and integrating in xn we have
11 1 XWk1ILP(w) < Cw(URk) .
REMARK. Given this covering lemma, cover the set IMwn)(f) > al by Rk
such that fw >a. Then we need only estimate w(URk) of
w(k) fRk
the covering lemma. But
w(URk) < w(Rk) < a ff Y Xkrkw 1 and we are finished by
interpolation.
One application of this theorem is that with it, one can obtain weighted
norm inequalities for multi-parameter maximal operators which cannot be
handled directly through iteration. We give the following example.
Suppose % denotes the family of rectangles with side lengths of the
form s, t , and s and t > 0 are arbitrary. (Suppose
the sides are also parallel to the axes.) Define the corresponding maximal
operator M by
Mf(x) = sup 1 f IfIdt.

xERER IRI
R
Then it is natural to ask for which weights w do we have
fMfPw<CJfpw.
The answer is AP(N) where this class is defined in the obvious way [14]:
P-1
w c AP(R) if and only if (_'_. rw <C
lR II
IwJJ
R
for all R E %. To prove this result we need the following.
LEMMA. If w c AP(R) then w satisfies a reverse Holder inequality.
Proof. Since w is uniformly AP in the x1 variable, w will satisfy a

reverse Holder inequality uniformly in that variable:
1/(1+)
fw(xlfrP)i4dxI) <C III fw(xiP)dxi).
I I
(C independent of p ). Fix I, an interval of the x1-axis, and define a

measure in the x2, x3 plane whose Radon-Nikodym derivative W(p) is
defined by
1/(1+8)
W(P) = III fw1+&(xi:P)dxi)
I
We claim that W satisfies an A condition uniformly (in I) relative to

the class of rectangles whose side lengths are t , II It in the x2,x3 plane.
Then let S be such a rectangle in the x2,x3 plane and E C S such
that IEI/ISI < 1/2. Then
JW(P)dP C. fIII fw(xiP)dxi)dP =C III f f w(xi,P)dxldp .
E I. E I ExI
Since w c A(t), ffExl w < (1- 71) AR w and so fE W(p)dp < C(1-rl) fsW(P)dp ,
and by taking S small enough, C will be so close to 1 that C(1-71) < 1
and W is uniformly A on the collection of all such S. Therefore since
S is just 1-parameter (just a linear change of variable in one of the x2
or x3 variable away from squares) we have that W satisfies a reverse
Holder inequality: (For 8' some value < S )
1/(1+8')
ISI f W14dp <C ISI rW.
w S
This shows that

1
(1+8) 1(1+8') (1+S')
I111 fwixi) < C-

- ISIf(Ti wdx
f1
1 1 1
dP dp
S I S I
and so
84 ROBERT FEFFERMAN
I1 rW1+ <C'IRI f w, Re9l.

R R
Now on to the theorem. Because w e AA(JR) and w-1 /(P-1) a AP'(91)

satisfies a reverse Holder inequality, w EAp-E(at) for some E> 0. It
follows that
Mf(x) <
and it just remains to show that Mw is bounded on Lp(w).

A quick review of the proof that Mwn), n = 3 , is bounded on LP(w)
reveals that all we really used was that w satisfy an A condition in
the x1 and x2 variables as well as a doubling condition in the x3
variable: w((R)d) < Cw(R). All of these are satisfied by our w here, and
this concludes the proof since Mw(f) < M(3)(f ).
Now we wish to relate some of our results on multi-parameter maximal

functions to the theory of multiplier operators.
We shall work in R2, and consider the following basic question: For
which sets S C R2 is XS(e) a multiplier on LP(R2) for some p 2 ?
For XS to be a multiplier of course means that, if for f e C R2) we set
Tf (6) = XS(e)f(6) then we have the a priori estimate
IITf1ILp(R2) <- CIIfliLp(R2)
In his celebrated theorem, Charles Fefferman showed that if S is a nice

open set in R2 whose boundary has some curvature then XS(e) will
only be an L2 multiplier [15]. The other nice regions left are those
whose boundaries are comprised of polygonal segments. If S is a convex
polygon then XS will obviously be an LP multiplier for all p, 1 02 > 03 > On > 0n-1 ~ 0 be a given sequence of angles 0 and let
N
N
I I I I
0 1 2 4
Figure 2
R0 be the polygonal region pictured above. Then we shall define TO by
T0f (f) = XRe(04)
Consider as well the maximal operator M0 defined by
Mef(x) = sup 1
XEREB0 IRI ,J
r if I dt
R
where B0 is the family of all rectangles in R2 which are oriented in

one of the directions 0k, but whose side lengths are arbitrary. We claim
86 ROBERT PEFFERMAN
that the behavior of T9 on LP(R2) for p > 2 is linked with the be-
havior of Me on L(P/2)((p/2)' is the exponent dual to p/2 ). More pre-
cisely, suppose that To is bounded on LP(R2) and we assume the
weakest possible estimate on M0, namely J{MOXE > 2 I1 < CIEI. Then
Me is of weak type on L(P12) . Conversely, if Me is bounded on
L(P12) (R2) then TB is bounded on Lq(R2), for p'< q < p [16].
To prove this assume first that To is bounded on LP(R2). Then
the first step is to notice that this implies that
I(YITkfkl2) /2 (fkl2)1/2IILp(R2)
1ILP(R2)<CII
This is proven by observing that if we dilate the region Re by a huge

RP"
p to get RB and translate R0 properly (by Tk) then R8" kk

looks like a half plane with boundary line making an angle of ek with
the positive x axis. Then if rk(t) are the Rademacher functions, and
T
T8 ' k f XRp,T k f then
e
T0,Tkf
_
and
(lII
li ( Lp ` `Te(eTk xf k)I2/ II
' LP
iT x p \IIP
J
r f i Irk(t)T0(e k fk)(x) dtdx = TT (Irk(t) e
irk *x
fk)I dxdt
J
R2 0 0 R2
1
(' (' P /2:1
<C
fJ R2
J
0
k(x) dtdx<C' (IfkxI2)'
lam ll
LP
Taking the limit as p -> oc, we have

1/2 /211LP(R2)
11LP(R2) < CII (lfkl2)'
12
II (YW JTkfk )
where Tk is the Hilbert transform in the direction 9k.
The next step is to use the above inequality to prove a covering
lemma for rectangles in 30. Let 1Rk1 be a sequence of such rectangles.
Select R given that R1, have been chosen provided
IR n [ U R ] I < 2 IR I . Then if R is unselected M0(XU ) > on R
j<k
2
so that
IURkl
6KXUi'1k > 2}I << CIURkl .
Let Ek = Rk - U and let fk = Xt . Then looking at the picture

j<k k
below, since IEkl > 2 IRkI on at least a set of measure > I-rkI the
100
Figure 3
88 ROBERT FEFFERMAN
segments pictured contain at least 1 /100 of their measure in EEk. If we

duplicate the rectangle Rk as shown, on these segments Tkfk > 1/100.
Applying Tk = Hilbert transform in the direction perpendicular to 0k to
Tkfk we see that Tk(Tkfk) > 1/100 on all of Rk. Repeating twice more
we get
/2I1LP 1/2
<C
<c Ia(tk12)
IfI
1 I+
Il (1 LP
IIxUiz k11 LP
This shows that M0 is of weak type (p/2)'.

Conversely, assume that M0 is bounded on L(P12) . Define
Sk < 61 < 2k+1L Then if Sk also stands for the
= 16 _ (61 e2)I2k
multiplier operator corresponding to Sk , we see that
I1 /2IIq )1/2IIq
IITOfIIq
Y
I{ (skTofI2) (IskTkf 12
To estimate II(11 ISkTkfl2)1/2IIq , let IIII(q/2)' =1 and let us estimate
fITkskfI20 f IT k(Skf
)12.b
But in R1 we have the classical weight norm inequality for the Hilbert
transform:
fjHfj20 <C J IfI2M(01+e)1/(1+e)

forall e> O.
It follows that (3.3) is
(3.4) < if ISk(f)I2Me(kl+e)t /(l+s)
M0 is bounded on L(q/2)/(1+e) if a is sufficiently small. It follows

that (3.4) is
< CN (sf2\fl110/2 C'IIfIIL9
proving that TO is bounded on L9 .
4. HP spaces - one and several parameters

In this lecture we wish to discuss another chapter of harmonic
analysis relating to differentiation theory and singular integrals, namely
Hardy Space theory. In this lecture, we shall discuss the one-parameter
theory, and, in the next, the theory in several parameters. In the begin-
ning when Hp spaces were first considered, they were spaces of complex
analytic functions in R+ _ 1z=x+iyIxcR1,ycR1 ,y>01 which satisfies
the size restriction
(1+00 I /p
IF(x +iy)Ipdx < C for all y c R+ .
00
One of the main reasons for introducing these spaces was the connection
with the Hilbert transform. If F(z) = u(z) + iv(z) is analytic with u and
v real, and if F is sufficiently nice then F will have boundary value
u(x) + iv(x) where v(x) is the Hilbert transform of u(x). It turns out,
since
+00
f
_
IF(x+it)Idx
increases as t 0, we have
+00
IIFIIHI defsu IF(x+it)Idx ti IIuHI1 + IIvIII(RI)

N J
-00
1
So we may view the space HI through its boundary values as the space
of all real valued functions f of LI(RI) whose Hilbert transforms are LI
as well.
90 ROBERT FEFFERMAN
If we want a theory of HP(Rn) then, following Stein and Weiss we

may consider the functions F(x,t) in R++1 = 1(x,t)IxeR11,t>01 whose
values lie in Rn+1 : F(x,t) = where the ui(x,t)
satisfy the "Generalized Cauchy-Riemann equations,"
n Cal
i (x, t) 0 (t = x 0)
i=0 1
and
dui dtrj
for all i,j .
J 1
These Stein-Weiss analytic functions are then said to be Hp(R++1) if

and only if
1 /n
sup J iF(x,t)ipdx = JIF11 [171.
t>O HP(R)
Again, these functions have an interpretation in terms of singular

integrals, since if a Stein-Weiss analytic function F(x,t) is sufficiently
"nice" on Rn+1 , then the boundary values u1(x) satisfy ui(x)
Ri[u0](x) where Ri is the ith Riesz transform given by R1(f )(x) _
In particular we may consider an H1(Rn+1) function (by

f *Ixln+i
identifying functions in Rn+1 with their boundary values) as a function
f with real values in L1(Rn) each of whose Riesz transforms Rif also
belong to L1(Rn). An interesting feature of Hp spaces is that they are
intimately connected to differentiation theory as well as singular integrals.
To discuss this, let us make some well-known observations. For a
harmonic function u(x,t) which is continuous on Rn+1 and bounded
there, u is given as an average of its boundary values according to the
Poisson integral:
cnt
u(x,t) = f * Pt(x); f(x) = u(x,0) and Pt(x) =
(Ix I 2 +t2)(n+1)12
Let F(x) _ ((y,t)l lx-yl < t(. Then since convolving with Pt at a point
x can be dominated by an appropriate linear combination of averages of
f over balls centered at x of different radii, it follows that
if u*(x) = sup lu(y,t)l , then u*(x) < cMf(x)

(y,t)cI'(X)
Unfortunately, if u(x,t) is harmonic, for p < 1 , u = PLO, and

1Rn lu(x,t)lpdx < C then the domination u* < CMf is not useful, since
M is not bounded on LP , and it is not true in general that u*(x) < 00
for a.e. x c Rn. On the other hand, suppose F is Stein-Weiss analytic,
say F c H1(R++1)
Then a beautiful computation shows that if 1 > a > 0 is close enough
to 1 (a > nn1) then A(lF la) > 0 so that lF la is subharmonic. If
s(x,t) is subharmonic and has boundary values h(x) then s is dominated
by the averages of h, i.e.,
s(x,t) < P[h](x,t) .
Applying this to G = [F la (which has JG1 'a(x,t)dx < C for all t > 0 )
we see that G* < M(h) for some h c L1 'a. Now M is bounded on
L1 Ia so that M(h) c L1 La and so G* c L1 Ia. It follows that F* a L1 .
Just as for a random f c L1(Rn) we do not necessarily have Rif c L1(Rn)
(singular integrals do not preserve L1 ) it is also not true that for an
arbitrary L1 function f that for u = P[f], u* c L1 . But if f c H1(Rn+1)
then u* c L1(Rn). Thus the nontangential maximal function
F*(x) = sup lF(y,t)l c L1(Rn)

(y, t) (F(X)
if and only if the analytic function f c H1(Ri+1)

92 ROBERT FEFFERMAN
We know so far that we can characterize Hp functions in terms of

singular integrals and maximal functions. There is another characteriza-
tion which is of great importance. To discuss it, let us return to Hp
functions in R+ as complex analytic functions, F = u+iv. It is an
interesting question as to whether the maximal function characterization
of Hp can be reformulated entirely in terms of u. That is, is it true
that F* c LP if and only if u* c LP ? In fact, this is true, and the best
way to see this is by introducing a special singular integral, the Lusin-
Littlewood-Paley-Stein area integral,
S2(u)(x)
= ff JVu12(y,t)dtdy
r(x)
which we already considered in the first lecture. As we shall see later,

for a harmonic function u(x,t), IIS(u)IILp ti IIu*IILP for all p > 0 [18].
The importance of S here is that the area integral is invariant under the
Hilbert transform, i.e.,
S(u) = S(v), since Ipvl = IVul .
When we combine the last two results, we immediately see that
IIu*IILP<00<1IF*IILP<00 FcHP(R+)
It is interesting to note that the first proof of IIS(u)IILP 11u* 11

LP '
1 > p > 0 was obtained by Burkholder, Gundy, and Silverstein [19] by
using probabilistic arguments involving Brownian motion. Nowadays
direct real variable proofs of this exist as we shall see later on.
To summarize, we can view functions f in Hp spaces by looking at
their harmonic extensions u to R++1 and requiring that u* or S(u)
belong to Lp(Rn).
It turns out that there is another important idea which is very useful
concerning Hp spaces and their real variable theory. So far, we have
spoken of Hp functions only in connection with certain differential
equations. Thus, if we wanted to know whether or not f e Hp we could
take u = P[f] which of course satisfies Au = 0.
This is not necessary. If f is a function and 0 E C- (Rn) with
IR 4 n=1 , then we may form f* (x) = sup If *Ot(y)I , (kt(x)
(t,y)EF(x)
C n O(x/t) and if lr a C0 (Rn) is suitably non-trivial (say radial, non-zero)
and fci = 0 we may form
S2 (f Xx) = ff If * At(y)I2 dndt

t
Then C. Fefferman and E. M. Stein have shown that
IIf ll IIf *II IISq'(f )II for 0 < p < -. .

Hp(Rn) Lp(Rn) Lp(R )
Thus, it is possible to think of Hp spaces without any reference to

particular approximate identities like Pt(x) which relate to differential
equations.
In addition to understanding the various characterizations of Hp
spaces, another important aspect is that of duality of HI with BMO,
which we shall now discuss.
A function O(x) , locally integrable on Rn is said to belong to the
class BMO of functions of bounded mean oscillation provided
I fI0(x)-cQldx < M for all cubes Q in Rn

IQI
Q
where 0Q = IQI fQ 0. The BMO functions are really functions defined
modulo constants and 11 IIBMO is defined to be sup IQI fQ 10 -IQI

94 ROBERT FEFFERMAN
According to a celebrated theorem of C. Fefferman and Stein, BMO

is the dual of H1 (181. This result's original proof involves knowing
that singular integrals map L to BMO and also a characterization of
BMO functions in terms of their Poisson integrals which we now describe.
Suppose p > 0 is a measure in Rn+1 and Q C Rn is a cube. Let
S(Q) = 1y,t)Iy Q, 0 < t < side length (Q)1. Then we say that A is a
Carleson measure-on R++1 iff g(S(Q)) < CIQI . The basic property that
characterizes Carleson measures is
ff Iu(x,t)Ipdi (x,t) < Cp J u*(x)pdx p> 0

Rn+1 Rn
for all functions u on R++1 In connection with this type of measure

there is the characterization of functions in BMO(Rn) in terms of their
Poisson integrals. A function O(x) on Rn with Poisson integral u(x,t)
is in BMO if and only if the associated measure
d(x,t) = IVuI2(x,t)t dxdt
is a Carleson measure. C. Fefferman proved this and used it to prove that

every function in BMO acts continuously on H1 :
J f(x)O(x)dx < CIIfIIH1(Rn)10IIBMO(Rn) .
Rn
These are the basic facts of Hp spaces that will concern us here and
which we shall later generalize to product spaces.
Let us now prove that for a harmonic function u(x,t) in R++1
IIS(u)IILp IIu*IILP for p> 0 .

We begin with the estimate flu*llp < CpiIS(u)llp , and to do this we

shall show that
12
ltu*>Call < c S2(u)(x)dx +
a J
S(uYa
From this our claim follows. This is because for )Lg(a) = If Igi > all we
have
00 00 a
Ifu*IILp J ap IXu*(a)da < Jr ap-1 r P'g(u)W)dO +Xg(u)(a da

a
0 0 0
<
f
0
#xS(u)(,6) f aP3dadfl +
00 f aP_1As(aa
00
0 00 0
Assuming p < 2 as we clearly may, this is
f0
OP-1's(u)(R)df3
NS(u)IIP .
To prove the estimate on Itu* > Calf , we set the notation that
M(XE) > 2 , and then claim
1. 5 S2(u)(x)dx > c fJ lVu(y,t)l2t dtdy where R = Ur(x)

R x4iS(u)>at
s(u)<a
In fact, if (y,t) c R then
IB(y,t) flf S(u)>ajI < 2 lB(y,t)I

96 ROBERT FEFFERMAN
Then
f
S(u)<a
S2(u)(x)dx = J
xJS(u)aI
J ivul2(y,t)tl-ndydt
T(x)
dx
(4.1)
Ioul2(y,t)tl-nIIxl(y,t)tr(i),x/IS(u)>a#ildydt
= ff
R n+l
But for (y,t) e R ,
IIxk(y,t)cI(x),x/IS(u)>aiij = IB(y,t)ncIs(u)>alI >2 IB(y,t)I

and (4.1) is
(4.1) > c JJ IVu12(y,t)t dydt
as claimed.
II. The next step is to write IVu12 = A(u2), and apply Green's theorem
to R :
ffA(U2)(y,t)t dydt = f !) t - u2 Alt da
R aR
Now > c > 0 for some c so the above gives
f u2da < C J S2(u)(x)dx + u(pu) tda .

J
aR S(u)<a aR
Since, for purposes of all estimates we may assume that u is rather nice,
we may assume u(Vu)t vanishes at t = 0, so
u(Vu)t do, = J u(Vu) t da

faR aR
where OR is the part of aR above IS(u) > al. It is not hard to see that
lpult < a on aR so that
1 /2
(+
J Jul lpult da< all' /2 (f u2da
aR aR
Putting all of our estimates together, we see that
f u2da < C J S2(u)(x)dx + a2 IS(u) > all .
aR S(u)<a
ti ti
III. Next we wish to define a function f by f( x) = u(x, r(x)) where
(x,r(x)) c aR defines the function r. We claim that in the region R
(4.2) lul<P[fl+Ca.
This is done by harmonic majorization. It is enough to show this on OR
and this in turn is just saying that for any point p c aR , JU(p)l is
dominated by the average over a relative ball on aR of u + Ca. This
follows from the estimate JVult < Ca on aR . Anyway, from (4.2) we
have, for x I IS(u) > al, u*(x) < CP[f ]*(x) + Ca , so that finally
l1u*>Call <4M('(x)>all <a2 IIf112
<C r S2(u)(x)dx +CI1S(u) > all

a2 J
S (u)<a
This completes the proof.

98 ROBERT FEFFERMAN
The proof that IIu*IIp < CPIIS(u)IIp which we just gave has been lifted
from Charles Fefferman and E. M. Stein's Acta paper [18]. To prove the
reverse inequality we want to go via a different route, and we shall follow
Merryfield here [20]. We prove the following lemma. In the next lecture we
show how this lemma proves IIu*IIp ? CpIIS(u)IIp
LEMMA. Let f(x) and g(x) c L2(Rn), and suppose 0 c C (Rn) radial
and u = P[f ] . Then
ff IouI2(x,t)Ig*ot(x)12t dtdx
R n+1
+
< fn'
R Rn+1
for some 1i c C'(Rn) with f;& = 0 real-valued).
Proof.
ff A(u2)(x,t)Ig*ot(x)I2t dtdx = f
Rn+1 Rn+1
rfV(u2)(x,t) V[(g * ot(x,t)2)t]dt dx = -2 fJu(xt)Vu(xt). V[(g * ot)2](x,t) t dt dx

Rn+1 Rn++1
+ +
f
-2 ,
R n+1
u(x,t)
_ -2 ffuVu 2(g*q5t)(x)p[g*q5t(x)]tdtdx -2 ffu i (g*q5t)2dtdx = 1+ 11 ,
n+t n+1
R+ R2
where
I
ff
Rn+l
+
u(tV(g*.kt))t-1/2. Vu(g*0t)t1/2dtdx
1/2 1/2
< ff U2 1 g*otl 2 dt dx I rIV,I21g*ot(x)I2 t dtdx
Rn++l RJn++I
and
II =fu (g*ot)2dtdx = - ffu + J u2(x,0)g2(x)dx

Rn+l R n+l Rn
u !AL (g*ot)2dtdx
ff
Rn+l
-JJ u22(g*(kt)
A fn
R
f2g2dx .
We see that
2 ff u (g*q5t)2dtdx < 55u2(g*) (g*)dtdx + fn f2g2dx

Rn++l
+ +
but
(g*-Ot)dtdx = ff u2(g*ot)t I (g*ot) dt dx

Rn+1 Rn+1
100 ROBERT FEFFERMAN
But
so
ff_2u -5211 (g*Ot)t(xi(k)t * gldt dx + r u2t (g*(kt)(g dt dx

S= 1 * (xio))
(ffIvu2(g*ct)2 t dtdx)\
1)/2 1 /2
I fu2(g*(xjc)t)2 d tdt
+(fg/ 2
Ydx
1/2,
*(xio))2 dtdxi)t
t
1/2
Putt ing this together gives
fflvuI2(g*t)2tdxdt
/
f
2
C [Y 1 (
1u2 g* 14) dt dx )rru2(g * ( X d tdx
+f f2g2
(ai4) 2
dt dx
5. More on Hp spaces
At this point we wish to discuss the theory of multi-parameter Hp
spaces and BMO. We saw, in the last lecture, that HP(Rn) could be
defined either by maximal functions or by Littlewood-Paley-Stein theory.
All of these spaces, Hp and BMO were invariant under the usual dila-
tions on Rn, x -Sx, and this is hardly a surprise, since they can be
defined by the maximal functions and singular integrals which are
invariant under these dilations. Here we shall define Hp and BMO

spaces which are invariant under the dilations (in R2 ) (x1,x2)
(S1x1,82x2), For convenience we shall work in R2 but all
61,82>0.
of this could just as well be carried out in R n x Rm, n, m > 1 . We shall

call our Hp and BMO spaces "product Hp and BMO" and denote
them by Hp(R+ x R+) and BMO(R+ x R+) so that we reserve HP(R2)
for the one-parameter space of functions on R2.
Let (x 1,x2) c R2 and denote by 1'(x) the set
i(yi,tl,y2,t2)I Iyi -xiI < ti, ti > 01 c R+ X R+
Let u(x,t) be a function in R+ x R+, x c R2 , t c R x R+, which is

biharmonic, i.e., harmonic in each half plane separately. Then the non-
tangential maximal function and area integral of u are defined by
u*(xi,x2) = sup Iu(yl,tl,y2,t2)I

(y,t)EF(x1,x2)
and
S2(u)(x) = ff IV1V2u(y,t)I2dy1dy2dtidt2 .
F(x)
More generally, if S6 c Cc (R2) and if
x x'
I 0 = 1, otI,t2(xi,x2) = tl1t21 tz
then for a function f on R2
f*(x) = sup if *0t t (x)

1 2
(y,t)cr'(x)
and if tA c Cc '(R2) and

fb(x1x)dx1 = 0 for all x2 ER1
f(J(xix2)dx2 =0 for all x1 a R1 ,
then
S ,(f )(x) =
ffIf*&tt(yi,y2)I2 dyl d Y2 dt dt 2
I 't t 2
t 2
F(x)
Given f(x1, x2), we define its bi-Poisson integral by u(xl,tl,x2,t2) =

P[f](x,t) = f * Ptl't2(x1,x2), where the bi-Poisson (or just Poisson for
short) kernel is defined by Pt1) t (xl,x2) = P x1 P 2
and
2 tl (T2
where P is the 1-dimensional Poisson kernel. Then, of course, P[f]
is bi-harmonic in R+ x R+.
In analogy with the 1-parameter case, we define f c Hp(R+xR+) if
and only if u* E LP(R2) where u = P[f 1. It is not hard to see, just as in
the single parameter case that f or any O .E C- (R2), f.0 = 1
IIf*IILP - IIu*IILP for p > 0 ,
so that we may use any approximate identity which is sufficiently nice to

define product Hp spaces. In terms of area integrals, we also have, for
O(x1,x2) suitably nontrivial, say i/i even in x1,x2 and not - 0,
I1SV'fIILP ti I!S(u)IILP p > 0, u = P[f] .
To complete the chain of equivalences , we would like to know that

IIS(u)IILP IIu*IILp
In fact, this is true, but is not obvious, and so we intend to present

the proof here. The proofs are by iteration, but they are not of the same
totally straightforward nature as the iteration in the Jessen-Marcinkiewicz-
Zygmund theorem. Often, this is the case in the analysis of product
domains, namely, the proof is by iteration, but this requires a different way
of looking at the one-parameter case than one is used to.
Proof that of S(u)e LP , then, u* a LP (Gundy-Stein) [211. We can

assume that u = PR]. Then since the 2-parameter area integral is in-
variant under taking Hilbert transforms in each variable separately, we
see that we may write
f =f+++f+_+f_++f__
where f++ is supported in eI, e2 > 0 and f+_ is supported in e > 0,

`'2 < 0, etc., and we have S(f++) a LP. By reflection we may assume f
is analytic (i.e., P[f] is bi-analytic) and then show that f* a LP. But
for u = [f] a bi-analytic function, we know that for a > 0, Iu(xl,tl,x2,t2)Ia
is subharmonic in each half plane (xi,ti) a R+ separately; this implies
that
lu(x,t)la < P[If(x)la] .
If a<p,
u*(x)a < M(Ifla) e LP/a(R2), if f f LP
So what we must show is that
IIf1ILp <_ Cplls(f )IILp, P>0.
To show this let us define some notation. In R I , if f( x) has

Poisson integral u, we let Qt be the operator (or kernel) which takes f
to tpu(x,t) , so that
S2(f)(x)= fflf*Qt(Y)I 2 dy2t

t
F(x)
Then going back to our present situation where f is given on R2, we

define
00
-Qtf(x1,x2) _ f(xl-y,x2)Qt(y)dy
and Q t similarly. Then let us define a Hilbert space valued function
dydt
F(x1,x2) a L2 (0);
t
by
F(x1,x2)(y2,t2) =Qt f(x1,x2+y2)

2
Now we know that in the one-parameter case iiS1fIIp > cpIlflip and a
glance at the proof of this fact reveals that it remains valid for Hilbert
space valued functions. Fix x2. Then
00 00
S1(F)(xl,x2)pdx1 > cp f IF(xl,x2)Ip 2 dxl

L (r)
-00 -00
and integrating this in x2 ,
(5.1) J1S1(F)(xl,x2)Pdxldx2 > cpJjIF(xl,x2)Ip 2 dxldx2
R2 R2
But fixing x1, since IF(x1,x 2)1 L2(iis the value of the one-parameter
)
integral of f(x 1, - ) at x2 , we have
J
R
IF(x1,x2)Ip 2
L J,)
dx2 > cp f
R 1
If(xl,x2)IPdx2
and so (5.1) is greater than or equal to
cp ff Iflpdxldx2
R2
On the other hand when the SI operator acts on the first variable we have
SI(F)(xl,x2) = S(f )(xl,x2)
the two-parameter area integral of f.

Next, let us prove that IIS(u)IILP < cpllu*IILp [21]. The proof is a
simple iteration of the one-parameter case given previously. To begin
with, we recall Merryfield's lemma: Let q c Cc (R I) be supported in
[-1, +1] and have f q5 = 1 . Then there exists Vi c Cc (RI) whose support
is also contained in [-1, +1] with f i/' = 0 and such that if u = P[f],
fJIvul2(x,t)(g*t(x))2t dtdx < C J f2(x)g2(x)dx +

R2 RI
jJu2(x,t)(g*cfrt(x))2 dt dx
+
R2
Introduce the notation tVu(x,t) = Qtf(x) , u(x,t) = Ptf(x), g*qt(x) _

Pt(g)(x) and -Qt(g)(x), Qt, i = 1,2, will denote the operator
acting in the ith variable. Then we estimate
(5.2)
jf
R2xR2
[Qt1Qt2f(xl,x2)]2[tt Pt2g(xt,x2)]2
dt1dt2
t it2
dxldx2
Fix x2, t2. Then (5.2) is
< ir ff (PtlQt2f(x))2(Q11P2 g(x))2

dtdx
x2,t2 xl,t2
f{Q?f(x)]2{g(x)]2 dx1 dt2 xl =1+11.

+ f t2
X
2 2
t x1
Now
ffPt2Pt1f(x))2(?g(x))2 dtdx
I= d
X1t1 x2 t2
t
fjfI f(Pf(x))2(tg(x))2dx2!1 dxI

+
x1,t1 x2
Then
II = fff(Pf(x))2(Qtg(x))2dx2dt2/t2dxi + fff2(x) g2(x)dx1dx2

X1
So the inequality we seek in the product case is

(5.3) J (Qtf )2(x)(Ptg)2(x)dxdt/t < c J

(R+)2 (R+)2
+ (Pt2f(x))2(Qt2)2(x)dx2dt2/t2
J J
x1ER1 (x2,t2)ER+
f J J f2(x)g2(x)dx
Rz
x2ER1\(x1,t1)ER+
It is now easy to see that IIS(u)IIL

P < CpIIu*IILP, P > 0.
In order to simplify things a little, we shall take a modified definition
of u* in what follows, namely
u*(x) = sup Iu(y,t)I

(y,t)E 1 0(x)
10
where
r1010(x) = {(y.t)I ly;-xiI < 1010ti, i =1,21 .
This is an irrelevant change, since a trivial computation shows that

IIu*Iip
iIu*IIp is, for a larger aperture, no more than a constant times for
a smaller aperture, the constant depending only on the apertures involved.
In (5.3), take c(x) = 1 for all Ixl < 1/3, and g(x) = Xu*(x)<a(x)'
Let us estimate
I
M()( u*>a )<1 /200
S2(u)(x)dx when u - P[f]
<J J IV1V2uI2(Y,t)IR(Y,t)flIM(Xu*> )<1/20011dydt
< ff 1V1V2ul2(y,t)t1t2dydt
R*
where R* = t(y,t)I IR(y,t)fllu*>aII < IR(y,t)il and where R(y;t) is

200
the rectangle in R2 with sides parallel to the axes and with side lengths
2t 1,2t2 centered at y. Notice that if IR(y; t) fl lu*>all < 100 IR(y; t)I
then g * (kt(y) = Ptg(y) > c for some c > 0. It follows that
f
M(Xl *>a #)<1 /2 00
u
S2(uXx)dx < ff IV,V2u12(Y,t)Pt(g)2(Y)dyt1t2dt
(R+)2
<C 5J u2(Y,t)ot(g)2(Y)dy tat2 + ff P2 f (Y)2 Qt 2(g)2(Y)dy a22

J
(R+)2 Y1 (Y2,t2)eR+
f If
f Pt1)f(Y)2.
t
t
(g)2(Y)dydt
l+ ff2g2dy
+ J 1 1 1 J
Y2 (Yi,tl)fR+2 R2
= i + ii + iii + iv.
C onsider If Qt(gXy)' 0 then u*(x) < a for some x f R(y; t).

i:
But then lu(y,t)l < a so i is less than or equal to

fft(g)2(y)dy tdt = a2
fft(1_g)2(y1y dt < a2II1-gIIL2 < a2lju*>a}I
a2 12
(R+)2 (R+)2
Consider ii: If Qt2)(g)(y1,y2) 1 0, then there exists x2 such that
Ix2 -Y2I < t2 and u * (yl,x2) < a. This implies that IPt2)f(y1,y2)1 <a so
2
ii is less than or equal to
ff(2)()2(Y)th2dY)dYl
a2 J 2 t2
Yl Y2't2)
Ag ain
ff Qt2)(g)2(Y)
dt 2
dY2 dyl g)2(Y)dY2 t2 dYl
2 J Jv 2
Yl Y2.t2 Yl 2,t2
< f(1-g)2(Y1,Y2)dY2 dYl < Ilu*>atI .

J
Y1 Y2
(iii) is similar to (ii).

Finally, (iv) is less than or equal to
J f 2(x) dx < J (u*)2(x) dx .
u*(x)<a u(x)<a
So we have
S2(u)(x)dx <C a2IIu*>atl+ u*2(x)dx

J J
{u*<al u*(x)<a
and we have seen before that this implies that
1IS(u)II < CpIIu*Il 2>p>0.

LP LP'
The next topic that we shall consider is that of duality of H1 and

BMO in the product setting. In the classical case there were four results
which expressed this duality.
1) The characterization of Carleson measures p for which the Poisson
transform f -- P[f] is bounded from LP(dx) to LP(dp), p > 1 .
2) The characterization of functions in BMO(R1) by a condition on
their Poisson integrals in terms of Carleson measures.
3) The characterization of functions in the dual of H1 by the BMO
condition.
4) The atomic decomposition of H1 .
Let us try to guess what the analogous theory should look like in
product spaces. For simplicity we consider the dual of H1(R+xR2).
Then what should an element of BMO(R2 xR+) look like? We might look
at tensor products of functions in BMO(R1) to get a feel for the answer.
So, for example if 951 and 952 are in BMO(R1) then S1(x1)S2(x2)
might be our model. Of course, this function S(x 1,x2) satisfies
(5.4) 195(x1,x2)-c1(xl)-c2(x2)j2dx2dx2 <C

J1 d
R
for the appropriate choice of functions c1 and c2 of the x1,x2 variable.

A Carleson measure in R+ x R2 would be a non-negative measure p
for which
(5.5) ffP[fIPdii < Cp ffIf(x)I1'dx, p>1

(R+) 2 R2
where P is the bi-Poisson integral. The obvious guess is that p

satisfies (5.5) if and only if p(S(R)) < CIRI for all rectangles R C R2
with sides parallel to the axes, where the Carleson region S(R) is
defined by S(I x J) = S(I) x S(J) for R = I x j . In terms of these Carleson
measures, it is not hard to show that 95 satisfies (5.4) if and only if its
bi-Poisson integral u satisfies
dp = IO172uI2(y,t)tlt2dt is a Carleson measure.
And finally, all of this in some sense is equivalent to asserting that every
f e HI(R2 xR2) can be written as I Akak where Y' IAkj < CIIf1I 1 and
H
ak(x1,x2) are "atoms," i.e., ak is supported in a rectangle Rk = IkxJk
such that
r ak(x1,x2)dx1 = 0 for all x2

Ik
ak(x1,x2)dx2 = 0 for all x1

J
Jk
and
1
IIak112<_
1 /2
IRkI
In 1974 [22], L. Carleson showed that p(S(R)) < CIRI was not suffi-
cient to guarantee the inequality
ffIP[f]Pdl < Cp J Iflpdx

(R2)2 R2
From here it is not difficult to produce examples of functions O(x1,x2)

which satisfy
1J
I
Ic(xi,X2)-C1(x1)-C2(x2)12dxldx2 < C
R
where C 1, C2 depend on R, yet I LP(R2) for any p > 2. Therefore,

this condition is not strong enough to force 0 to belong to the dual of H1
In other words the simple picture of the structure of H1(R2 xR+) and
BMO(R2xR2) suggested above as the obvious guess is completely wrong.
Rather one considers the role of rectangles to be played instead by
arbitrary open sets. Although this may seem a bit frightening at first
glance, it turns out, and this is of course the final test of the theory, that
nearly all the classical theory of Hp and BMO can easily be carried out
using the approach suggested here.
By way of introduction, we shall prove that for any function
e H1(R2 xR2)*, if u = P[qS] we have a Carleson condition with respect
to open sets satisfied by the appropriate measure. To describe this result,
we make the following definition ([23], [24], and [251).
Let 9 C R2 be an arbitrary open set, and let R(y; t) be the rectangle
in R2 centered at (y1,y2) = y and with side lengths 2t1 and 2t2. Then
S(i2) the Carleson region above SZ is defined as
S(Q) = U S(R) = f(y,t) E (R+)21 R(y; t) C SZ# .

RCSZ
Then we say that > 0 in (R4)2 is a Carleson measure if and only if

,u(S(SI)) < C101 for every open set SZ C R2 f c H1(R2 xR+)* if and only if
for u = P[f ] ,
d = 1V1V2uI2tlt2dt1dt2dy1dy2 is a Carleson measure.
In fact, this follows immediately from the inequality (5.3). To see this,
notice that IV1V2uI is invariant under the Hilbert transform HXi(i = 1,2)
so that if we prove this when f e L(R2), we will have proven it also
when f is of the form
gl + Hx 92 + Hx2 93 + Hx Hx 2g4 gi c L .
1 1
A function a(x) on R2 will be in H1(R2xR+) if and only if a c L1(R2),

H a,Hx2 a, and H xl H a cL1.
xl x2
In fact, if a c H1(R2xR+) then S(a) e L1(R2) hence so are S(Hx a),
1
S(HX a) and S(HX Hx2 a); therefore Hxia, Hx1 Hx a e L1 . Conversely,
2 1 2
if a, Hxia , and Hx Hx2 a e L1(R2) then we can form F , F+_, F_+
1
and F__ c L1(R2) such that a = c F++ and reflections of the F++ are
boundary values of bi-analytic functions. A bianalytic function F with
(distinguished) boundary values in L1(R2) has F* c L1 by a subhar-
monicity argument applied to IF I', a < 1 . So a* a L1 and a e H1. Let
$ e H1(R+xR+)*. Define a map from H1(R+xR+) L1(R2)i by
i=1
0(f) = (f, Hx l f, H x2 L H xl Hx 2f) .
Then Ij6f jI
l lif 11
HI .
0 is obviously one to one, so 0-1 = s exists
and is bounded on Im(6). The map $ extends, by Hahn-Banach to
an element of the dual L1 = L. Then
$(f)=P s(f,Hx l f,Hx2 f,Hxl Hx2 f)
= rfgl + Hx f g2 + Hx f g3 + Hx Hx2f g4
rJ 1 2 1
=J f (g1Hx 1g2+Hx 2g3+Hx Hx2g4) fcH1.

1
Thus every element of (Hl)* is of the form

So it suffices to show that if f c L(R2) with u = P[f] then
dp = Ip1V2I2(y,t)tlt2dydt is a Carleson measure.
But in (5.3), take g = XU(x1,x2), and notice that if f 0 = 1 , supp fi(x) C

[-1,1] then Ptg(x) = 1 if (x,t) c S(Sl). This is because for such (x,t),
R(x; t) C SZ and g * 95t(x) = fR2 ot(x -u)du = 1 . It follows from (5.3) that
ffviv2u2tit2dxdt < CIIfIIoo f IQtgi2I at dx
X1 \(x2t2) / X2 \(X1,t1)
11f 112
< CIIf11211gII2 < C ICI
6. Duality of H1 and BMO and the atomic decomposition

In this lecture we shall consider in greater detail the spaces
HP(R+xR+) and BMO(R+xR+), which we discussed briefly in section 5.
There we saw that in product spaces, the most obvious guesses at char-
acterizations of Hp atoms of BMO failed. In order to circumvent these
difficulties we must take a slightly different approach than we are used to
in the classical 1-parameter case.
In what follows we shall be working with functions in HP(R+xR+) or
BMO(R2 xR2) only. The theory for R2 xR2 x x R2 or for Rn f1 x Rm+1
is quite similar and only requires minor changes. Now let I1 be the
family of all rectangles with sides parallel to the axes and sR4d be the
subfamily of 91 whose sides are dyadic intervals.
If f( x1,x2) is a sufficiently nice function on R2, and q' c C(R1),
iii is even, ib' 4 0 real valued and supp(qi) C [-1,1], and i/i has a large
number of moments vanishing, then for
00
ft 1
1
2
(x1,x2) _
tl
X
1)0(x2)
t2
t11t21,
f
0
=1
we have
dtldt2
f(x1,x2) = f *tPtl,t2(Yl,y2)otl,t2(xl -y1,x2-y2)dyldy2 tit t2
R+x R+
In fact, taking Fourier transforms of both sides, for the right-hand

side we have
00 00
ff f(St dtia22 = f(e) r r I (t1C1,t2C2)12 dtldt2 =j(e)

J J
RZxR2 0 0
We can use this representation to decompose the function f as

follows: R c 8Rd . Set ?1(R) = 1(y,t) c R+xR+I y c R , P1 < t1 < 2 Pi where
Pi , i = 1,2 is the side length of R in the xi direction 1. Since
R2xR2
+ +
= U W(R), if we define
Rc91d
fR(x1,x2) _ ff
f(y,t) f= 1 fR, and each fR is supported in
Rc9? d
R the double of R and has the property that

ffRx2x2 dx1 = 0 for all x2

I
ffR(XlIX2 )dX2 = 0 for all x1
where 14 =TxT.
It will be convenient to define a norm I I R on functions supported on
a rectangle R , as follows.
N
_ oaf
IfIR IIIa111ia2
IaI=O axa1ax22
where N is a large integer. With these preliminaries we can pass to a

theorem characterizing (H1)* in a number of useful ways.
THEOREM [2$]. For a function on R2 the following are equivalent:

(1) e H 1(R+ x R+)*.
(2) 95 = g, +Hx1(g2) + Hx2(g3) + Hx1Hx2(g4) for some 91,92,93 and
g4 in L(R2).
(3) If u = P[] in R+ x R+, then
ffIviv2uI2(y.t)tit2dt < CIt I, for all open sets Q C R2 .
S(Q)
(4) If (y,t) _ * lt(y), then
ffIy,o2dy.<cti, fore!! open sets [i C R2
S(fl)
(5) 0 can be written in the form 7- cRbR where bR(xl,x2) are

ReRd
supported in k IbRIR < 1 and I cR < CIcI for all 9 open
RCSZ
in R2.
Proof. To begin with, we proved in section 5 that (1) --> (2). It is also
trivial that (2) - (1), since if f e H(R2xR+
f
J f(x) Hx Hx (g)(x)dx =
1 2
fHx1 Hx (f Xx) g(x) dx
2
and since f eH1, Hx Hx 2 (f) rLl.

I
Next, we recall that (2) _- (3) was also proven in the preceding
section.
Now we claim that (3) or (4) implies (1). We show that (4) implies (1),
the other proof being similar. We.do this via the atomic decomposition of
H1 which we shall describe here only enough to derive our implication.
We shall present the decomposition of H1 in greater detail later. Let
f e H1(R+xR+). Then SO(f) e L1(R2) (this follows by vector iteration,
just as in the argument that S(f) e L1 implies f e L1 ).
Consider the sets SZk = ISO(f) > 2k1, k e Z. Set
ak(x l ,x2) = I
Re91d
f R(x)
IRflUk1 >1/21RI
IRf1Slk+1I < 1 /2I RI
ak(x I x )
Then, as we shall elaborate later on ak(xlPx2) = 2 is an
H1(R+xR+) atom where
l(xl,x2)IM(2)(Xci )(xl,x2) > 1/101

k
Then f = 1 Akak where Ak = and by the strong maximal theorem

IS'EkI < CISZkI so that
I Ak < C IIS f,(f )IIL1 < C,IIf 11

Hi
Now consider c(x1,x2) satisfying (4). Then it will be enough to show

that
J ak(xl,x2)'O(xlx2)dx <C
R2
and then simply sum over k. But
(6.1) 5ak(xl.x2)(xlx2)dxldx2 =2ki I

fIf(x)c(x)dx
2
k
R
where the sum is taken over
k ={Rdtdl IRrH kI>2 IRI but IRfQ2k+1I <2 IRI}.
Then (6.1) becomes
56(x) fffYttx_Y) d y dt dx 2ki kl

RY k
91(R)
1
t i
tit
R9k Sff(y,ty,t'y__ 2
(R)
2k1 RU
ff If
(y,t)I2dY
ate
1 /2
ff I56(Y.t)12dy d t2c2
/2
kI c%k RU
-1 91k
MR) 91(R)
(6.2)
J
IS,,(f)2(x)dx >_ U
RE9
ff If(Y,t))2dy td
1t2
fl k/Qk+1 k 91(R)
To show this
S, (f Xx)dx =
J
ff If
(Y,t)12dy
t2
I)k/uk+l xE?k/SZk+1 F(x)
> J J jf(y't)12Ifx (? k/Ik+1 I(y,t) Er(x)1I dy 2t2

12
Suppose that (y,t) c %(R), R c R k . Then if the aperture of I' is large

enough, then (y,t) E I'(x) for all x c R. Since
IN x E(?k/Ik+1) 'R' I> IRI , R 9k

2
we see that
{x t1t2 (observe that t1t2 ' IRI )

2
for (y,t) c U %(R), and this proves (6.1).
RcRk
But then
SV(f)2(x)dx < (2k+1)21 kI
il kAlk+1
and combining this with (6.1) yields
U ff If(y,t)I2dy t2t2 < C2k IS kI

RED 12
k W(R)
As for
R(R.
91(R)
if we observe that for any R E 91k I 91(R) is contained in S(?lk), we get
U
ff I0(y,t)I2dy
tat2 < if
195(y,t)I2dy tat- < C ?1
RE9lk
91(R) S(? k)
by (4). This shows that I f a k Odx I < C and completes the proof that (4)
implies (1).
We shall show next that (2) implies (4). Let g e L(R2). We claim
that if 1 C R2, then
dt CIIgII2lgl
f Ig(y,t)12dy tit-00
00
where g(y,t) = g *Vit t (y). To show this, observe that since

1 2
supp(i4') C [-1,1], C R(y,t). Hence, if (y,t) f S((1),
g *0t(y) _ (gX 1) * qt(y) and so
ff
S((1)
Ig(y,t)12dy tit 2 ff (g)(y,t)dy dt
(R+)2
12
An easy application of Plancherel's formula says that this is, in turn,
CIO
IIgX 1112 f
0
Since, f +1 1 q, = 01
i(0) = 0 and fr E Co
() =
0(161 -N) as I CI - 00, for each N > 0.
so
00
f
0
de
<
It follows that
f
S(U)
Ig(Y,t)I2dy
dt < C IIgII,2oIjjl
tlt2
as claimed.
If we wish to prove that the same Carleson condition holds for g
replaced by Hx1 Hx2 (g), then we proceed as follows. Observe that
(.rf EHxHx()l *t(Y)I2dY t

S(f)
1/2
=
JJJ
S(SZ)
t12
1/2
where T = Hx1Hx2(0). The function T splits into a product of

gJ(1)(xl).T(2)(x2) where 'y(1) is odd, Co and decreasing at oo like
Ixil-N (depending on how many moments of rr vanish). Now suppose we
choose r?(x) on R1 so that supp(q) C (1/4,4), 71 E C(R1), 77 even and
E -q1/ x = 1. Let i 0(x) = E ,Ix 1l and for k > 0, let 77(x)

k _
k=-o
00 \2k/ k<0 \2k/
rl/ 2kl Set Then
(a) supp(Wk,j) C 4R(0; 2k,2i)
(b) 'Yk,j is odd in each variable separately

(c) Tkj is Cc (R2)
and
(d)
(.9)aq,j = 0(2-kN-jN) as k,j o if Ial <2
If
By Minkowski's inequality, we have
1 /2 1 /2
(6.3) (ffig * 'Ftl2dy dt <

Jf19*
("k,j)tl2dy tdt
12
tlt2 k,j
S(fl) s(1)
Now, to estimate
I9 *(Tk,j)t12dy tat
12
kj
we use the same argument as that given above, except that now
supp(q1k j) < 4R(0; 2k,2)) and not the unit square. If (y,t) a S(cl), then
R(y;t) C B and the support of (Wk,j )t(. -y) will be contained in
R(y,2kt1,2jt2) C IM(2)(X[2) > 2-(k+j)I =

jj
k,j
Thus
I 5 Ig*('Fk,j)t(y)I2dY
dt2 = ff 1(9,,,,
kj) *'Ykjl2dy tat2
S(B) S(Q)
I'kj(6,,C2)12
2 <1i Ill 1 2
11gx
112 ff
kj 2
0
f 0
1
1 2
kit
JJ
0
0
00
2
By the strong maximal theorem ISkj I < C(k+j)2k+1-, and it is easy to

de
see that fo 00 fo kj2 decreases like a large power of 2-(k+l) as
k,j oo. So our desired estimate follows from (6.2).
So far we have proven the equivalence of (1), (2), (3) and (4). We shall
not go into the details of the equivalence of (5) except to say the proof is
given in the Annals paper of Chang-Fefferman [25]. Rather, let us point
out a beautiful application of the equivalence of (5) with the other defini-
tions of BMO which occurs already in the one-parameter setting. This
is the theorem of A. Uchiyama [26], which tells us which families of
multipliers homogeneous on Rn of degree 0 determine H1(Rn). He
showed that for multiplier operators I, KI, Km with multipliers
1, ei(e) that f, Kif E L1(Rn) implies f E H1(Rn) if and only if the ei
,
separate antipodal points of Sn-I , i.e., if and only if for every e t Sn-I
there exists i such that ei(e) I ei(-e).
The way Uchiyama proves this is to show that the dual statement is
true, namely, every (k E BMO(Rn) can be written as
(6.4) go + I Kigi for some 90,91,-,gm c
This depends on a simple lemma.
LEMMA. If ei are as above, then given f c L2, and a vector v c Cm+1 ,

there exist functions go,---, gm E L2 so that
go + I Kigi = f and g(x) = (go(x), g1(x),..., gm(x)) v for all x E Rn
To prove the formula (6.4), Uchiyama decomposes 0 = I CIO1 as in

our (5), and applies the lemma to get functions g1(x) such that Kg1(x) =
C1-O1(x) for which g(x) is perpendicular to the correct v, and the result,
when modified only slightly to gI, has the property that gI E L. For
the details see Uchiyama's recent paper in Acta [26].
Now, finally we wish to discuss the atomic decomposition of
H1(R+xR .) in greater detail. There are interesting applications of this
decomposition besides duality with BMO(R+xR+) which was presented

above. We shall be content with one more application here which sheds a
good deal of light on the nature of these atoms. Namely, we intend to
give a second proof, directly by real variables, that on R+ x R+ if
S(f) c L1(R2) then f* or L'(R2) [27). Suppose S,&(f) E L'(R2). Then
in our discussion of duality we defined atoms
ak(x) = fR(xl,x2)
2kls kl RERk
To simplify this notation we define w = itk and A(x) = 2klSkrak(x) . Let

q51 and q52 c Cc (R1) with 0'(x) > 0, f q5' = 1 , and supp(O') C [-1, +11.
Set
1t2101(t1).02(x)
-011't2(x1,x2) = ti
Define w = Mt21(X(a ) > to . We need to estimate A * q5t 1t (x) for

10 2
x / CO. To do this let us make the following definitions. If R is a
rectangle then R1, R2 will be its sides, so that R = R1 xR2 . Let
AI(x) = I fR(x), A?(x) = L f R(x) .

91k.
JR11=23 IR2l=2J
Then to estimate A * 6t1,t2(x) , since supp(q5t( -x)) C R(x; t) = S, in

the definition of A we need only consider those fR for which R f1S 0.
For any such rectangle R c Rk , since R C (0,
minimumr (sill ,
R il <
IS
Slsi q11 /2 =P
1` 2
where w denotes again Mt21(Xw) >

1010
Then
A * ctlt2x) = IA I A * q5t1t2(x)
2i/Is2I<p
2)/Is1I<p
Y fR * Otlt2(x)
RE8
iR2II
where '23 C 91k consists of rectangles R so that iRi < p, I
< p and
SII S2
kf1S 0. Thus R C''`9 for all R E 18, and the reason this subtracted
term occurs is that we have double counted these fR whose R sides are
both very small.
In order to estimate A' * q5t1t2x) we use the following trivial lemma.
LEMMA. On R1 suppose that fi(x) E C and is supported in an interval

0. Suppose a(x) is supported on disjoint subintervals of $, Ik whose
lengths are all < yI$I. Assume that a(x) has N vanishing moments
over each Ik. Then
f a(x)-O(x)dx
aq
We estimate Al * (btt(x) using the fact that for each fixed x2,
A1(- ,x2) has N vanishing moments over disjoint x1 intervals over
l
length 2 .2) . (Actually, we sould have to break up A into 3 pieces to
insure this, but we spare the reader this trivial complication.) It follows
from the lemma that
2i )N+I(IA' I
IA *1 0t 1 (x)I < * 1 X[-t 1 t 1 ] (x)) '
t1 1 I
Convolving in the x2 variable, we have
N+1 (' IA'

-0t1,t2(x)I < C 11
I dx'
IAA *
IS -gI
For this we get
2'/IS1 I<p
Aj <CpN/2 1
IiI J
f (IA2)
-9
II
1/2
23/IS1 I< I g )N1 /2
< CpN /2 1 Y 1 /2
r(IAhl2)
By symmetry
1 /2
2 .0tlt2(x) < CpN/2 1 121
L. Aj * IAi
.
2'/IS2I<p
IR R 2I
Now let RCS , with 1I < IS2I . Then
S11
and also
<C\1SI/N/2 IS I 1
IfR * .0t1,t2(x)I IfRI
Thus
r
L IfR *1t t (X)I
2
<- C --L J
fI fR(x)
1 /2
(
(Is
R
IlN /21 /2 RN/4
?
ReAap 1 IS I RERk / RE%k
I
1 /2
CpN/4 1 Y IfRI2
I I R E991k
To sum up our findings, we have seen that if x / Ei then

I)N/4
(6.5) JA * otlt2(x), C fJ(Y)dY
Snl SI
where
1/2 1/2 `1/2
+((A2)2)
-++ (I fR)
RE%k
To finish the proof, we need another lemma:
LEMMA. Let g(x) _ I fR(x) where B is a collection of dyadic

REI
rectangles. Then
1IgU12 < U J If(Y,t)12dy dt

RE tlt2
91(R)
Proof. Let = 1. Then

11h11
22
f(x)h(x)dx =f I ff f(Y ,tt(x-y)dy 2 h(x)dx
R8
_ R2 I fffthYt)dYTTdt
W(R)
rf f &'t)12dy dt2)
/2 /2
< (Pz JJ
9I(R)
1 f
R+)2
Ih(y,t)12dy dt
tIt2)
1/2 1/2
c IlS .(h)II2 < (Y't)Izdy dt
(51If(Y,t)12dY tat
tit 2
(ffif tIt2
Now, notice that, by the lemma,
IIJI12 < ff If(y,t)I2dy

at
12
<_
f S ,(f)(x)dx < C 22k. Ia 1
R E(R )
k kk+1
/0
The same estimate holds for IIA I12 . Then
1 /2
A* < 111 /2 (fA*2 < c1,011/2 IIA112 < c2k 1.1 .
ti J
Also away from ('0
A*(x) < M(2)(.1)(x) >
so
1 /2
fA*dx < (fM(2)(x2o(xdx)'(fM2a)2(x)dx
R /
< C I,,11121(,11 /22k = C2kIw1 .
It follows that IIA*111 < C2kkwi and also IIakIII < C.
ROBERT FEFFERMAN
DEPARTMENT OF MATHEMATICS
UNIVERSITY OF CHICAGO
CHICAGO, ILLINOIS 60637
BIBLIOGR APHY
Ill B. Jessen, J. Marcinkiewicz and A. Zygmund, Notes on the Differ-
entiability of Multiple Integrals, Fund. Math. 24, 1935.
[2] E. M. Stein and S. Wainger, Problems in Harmonic Analysis Related
to Curvature, Bull. AMS. 84, 1978.
[3] A. Cordoba and R. Fefferman, A Geometric Proof of the Strong

Maximal Theorem, Annals of Math., 102, 1975.
[4] J. 0. Stromberg, Weak Estimates on Maximal Functions with
Rectangles in Certain Directions, Arkiv fur Math., 15, 1977.
[5] A. Cordoba and R. Fefferman, On Differentiation of Integrals, Proc.
Nat. Acad. of Sci., 74, 1977.
[6] A. Nagel, E. M. Stein, and S. Wainger, Differentiation in Lacunary
Directions, Proc. Nat. Acad. Sci., 75, 1978.
[7] A. Cordoba, Maximal Functions, Covering Lemmas, and Fourier
Multipliers, Proc. Symp. in Pure Math., 35, Part I, 1979.
[8] F. Soria, Examples and Counterexamples to a Conjecture in the
Theory of Differentiation of Integrals, to appear in Annals of Math.
[9] B. Muckenhoupt, Weighted Norm Inequalities for the Hardy Maximal
Function, Trans. of the AMS, 165, 1972.
[10] R. Hunt, B. Muckenhoupt, and R. Wheeden, Weighted Norm Inequali-
ties for the Conjugate Function and Hilbert Transform, Trans. AMS,
176, 1973.
[11] R. R. Coifman and C. Fefferman, Weighted Norm Inequalities for
Maximal Functions and Singular Integrals, Studia Math., 51, 1974.
[12] M. Christ and R. Fefferman, A Note on Weighted Norm Inequalities
for the Hardy-Littlewood Maximal Operator, Proceedings of the AMS,
84, 1983.
[13] R. Fefferman, Strong Differentiation with Respect to Measures, Amer.
Jour. of Math., 103, 1981.
[14] , Some Weighted Norm Inequalities for Cordoba's Maximal
Function, to appear in Amer. Jour. of Math.
[15] C. Fefferman, The Multiplier Problem for the Ball, Annals of Math.,
94, 1971.
[16] A. Cordoba and R. Fefferman, On the Equivalence between the
Boundedness of Certain Classes of Maximal and Multiplier Operators
in Fourier Analysis, Proc. Nat. Acad. Sci., 74, No. 2, 1977.
[17] E. M. Stein and G. Weiss, On the Theory of Hp Spaces, Acta. Math.,
103, 1960.
[18] C. Fefferman and E. M. Stein, Hp Spaces of Several Variables,
Acta Math., 129, 1972.
[19] D. Burkholder, R. Gundy, and M. Silverstein, A Maximal Function
Characterization of the Class Hp, Trans. AMS, 157, 1971.
[20] K. Merryfield, Ph.D. Thesis: Hp Spaces in Poly-Half Spaces,

University of Chicago, 1980.
[21] R. Gundy and E. M. Stein, Hp Theory for the Polydisk, Proc. Nat.
Acad. Sci., 76, 1979.
[22] L. Carleson, A Counterexample for Measures Bounded on Hp for the
Bi-Disc, Mittag-Leffler Report No. 7, 1974.
[23] S.Y. Chang, Carleson Measure on the Bi-Disc, Annals of Math., 109,
1979.
[24] R. Fefferman, Functions on Bounded Mean Oscillation on the Bi-Disc,
Annals of Math., 10, 1979.
[25] S.Y. Chang and R. Fefferman, A Continuous Version of the Duality
of H1 and BMO on the Bi-Disc, Annals of Math., 1980.
[26] A. Uchiyama, A Constructive Proof of the Fefferman-Stein Decom-
position of BMO(Rn), Acta. Math., 148, 1982.
ELLIPTIC BOUNDARY VALUE PROBLEMS
ON LIPSCHITZ DOMAINS
Carlos E. Kenig*
Dedicated to the memory of lack P. Burke
PREFACE
This paper is an outgrowth of a series of lectures I presented at the

Summer Symposium of Analysis in China (SSAC), held at Peking University
in September, 1984. The material in the introduction and parts (a) and (b)
of Section 1 comes from the expository article `Boundary value problems
on Lipschitz domains' ([191), which I wrote jointly with D. S. Jerison in
1980. The rest of the paper can be considered as a sequel to that article.
Some of the material in part (b) of Section 2, and all of Section 3 comes
from the recent expository article "Recent progress on boundary value
problems on Lipschitz domains" ([231). The results explained in Section
2, (b) and Section 3 are unpublished. Full details will appear elsewhere
in several joint papers.
Acknowledgements. I would like to thank Peking University, and the

organizing committee of the SSAC, Professors M. T. Cheng, S. L. Wang,
S. Kung, D. G. Deng and R. Long for their invitation to participate in the
SSAC, and for their warm hospitality during my visit to China. I would
also like to thank Professor E. M. Stein for his many efforts to make the
SSAC a success. Thanks are also due to Mr. You Zhong and Mr. Wang
Wengshen for taking careful notes of my lectures.
Finally, I would like to thank B. Dahlberg, E. Fabes, D, Jerison and
G. Verchota for the many discussions and fruitful collaborations that we
Supported in part by the NSF. 131

132 CARLOS E. KENIG
have had throughout the years, which resulted in the work explained in
this paper.
Introduction
A harmonic function u is a twice continuously differentiable function
on an open subset of Rn , n > 2, satisfying the Laplace equation
132u
Au = = 0. Harmonic function arise in many problems in mathe-
i=I 9X?
7
matical physics. For example, the function measuring gravitational or

electrical potential in free space is harmonic. A steady state temperature
distribution in a homogeneous medium also satisfies the Laplace equation.
Moreover, the Laplace equation is the simplest, and thus the prototype, of
the elliptic equations, or systems of equations. These in turn also have
many applications to mathematical physics and geometry. A first step in
the understanding of this more general situation is the study of the
Laplacian. This will be illustrated very clearly later on.
Initially we will be concerned with the two basic boundary value
problems for the Laplace equation, the Dirichlet and Neumann problems.
Let D be a bounded, smooth domain in Rn and let f be a smooth
(i.e. C ) function on oD , the boundary of D . The classical Dirichlet
problem is to find and describe a function u that is harmonic in D,
continuous in 5, and equals f on 09D . This corresponds to the problem
of finding the temperature inside a body D when one knows the tempera-
ture f on 3D. The classical Neumann problem is to find and describe a
function u that is harmonic in D, belongs to C1(D), and satisfies
N = f on r3D , where represents the normal derivative of u on XD .
09
a
This corresponds to the problem of finding the temperature inside D when
one knows the heat flow f through the boundary surface X.
Our main purpose here is to describe results on the boundary behavior
of u in the case of smooth domains, and to study in detail the extension
ELLIPTIC BOUNDARY VALUE PROBLEMS 133
of these results to the case of minimally smooth domains, where we allow

corners and edges, i.e. Lipschitz domains. This class of domains is
important from the point of view of applications, and also from the mathe-
matical point of view. Their importance resides in the fact that this is a
dilation invariant class of domains with some smoothness. They have the
borderline amount of regularity necessary for the validity of the results
we are going to expound on.
In a smooth domain, the method of layer potentials, (which we are
going to develop soon) yields the existence of a solution u to the
Dirichlet problem with boundary data f e Ck,a(aD), and the bound
llullck,a(D) < Ck,allf llck,a(3 k = 0,1,2,...

0<a<1
What happens if the size of f is measured in some other norm, like
the L2 norm? This is of interest as a measure of the variation in data
even if we are only concerned with continuous functions: if fl-f2 has
small L2 norm we want to know that the corresponding solutions ul
and u2 are near each other. The wisdom of hindsight tells us that as
long as we are going to examine all continuous functions in L2 norm, it
is no harder to consider arbitrary functions in L2. Another reason to
consider the L2 norm is that it is better suited for the Neumann problem,
even on smooth domains. We will also consider how our results change if
we consider LP norms, p ? 2 , as the smoothness of the domain
decreases.
We will then show the flexibility of our methods by considering exten-
sions of our results to systems of elliptic equations in Lipschitz domains.
The specific systems of equations that we will study are the Navier
systems which arise in the linear and infinitesimal theory of elasticity,
when the displacements or the surface forces are given on the boundary
of a homogeneous and isotropic elastic body D. These systems are the
prototype of the second order elliptic systems of equations. We will also
134 CARLOS E. KENIG
study the so-called Stokes problem; this is the linearized stationary

problem of the mathematical theory of viscous incompressible flow.
Before going on to study the general situation, we will formulate
appropriate theorems, by examining a model case, namely the Laplacian
in the unit ball B. In this case we have a lot of symmetry at our dis-
posal and everything can be done explicitly. Let do denote surface
measure of aB.
THEOREM. Suppose that 1 < p < oc and f c LP(3B,do). Then, there

exists a unique harmonic function u in B such that lim u(rQ) = f(Q)
r-,1
for almost every Q e aB , and
(*) fu*(Q)Pdo(Q) <_ Cp ff(Q);Pdc7(Q)

aB aB
where u*(Q) = sup Iu(rQ)I.

O<r<1
The theorem asserts that fr(Q) = u(rQ) converges to f(Q) not only in
LP norm, but also in the sense of Lebesgue's dominated convergence.
In the analogous estimates to (*) in the Neumann problem, u is re-
placed by the gradient of u. In that case the estimate fails for p = 00,
even if aN is continuous.
In both the case of the Dirichlet problem and the Neumann problem,
the radial limit can be replaced by a non-tangential limit: if X tends to
Q with IX-QI < (l+a) dist(X, dB), for some fixed a > 0, then u(X)
has the limit f(Q) for almost every Q .
The theorem is most easily proved by writing down a formula for the
2
solution, u(X) = fas P(X,Q) f(Q) do(Q) , where P(X,Q) = n Iin
IX-Q
The estimate now follows as an easy consequence of the Hardy-Littlewood
maximal theorem. An analogous formula holds for the Neumann problem.
This time, it is more difficult to obtain the estimates. One needs to use
the Calderon-Zygmund theory of singular integrals, and the Hardy-
Littlewood maximal theorem.
The case of the Laplacian in the ball is relatively easy because of
the existence of explicit formulas for the solution. What should we do in
the case of a general domain, where explicit formulas are not available?
What should we do to study systems of equations? What happens to our
solutions as the domains become less smooth? We hope to give a system-
atic answer to these questions in the rest of this paper.
1. Historical comments and preliminaries

(a) The method of layer' potentials for Laplace's equation on smooth
domains.
DEFINITION. A bounded domain D is called a Lipschitz domain with

Lipschitz constant less than or equal to M if for any Q e aD there is a
ball B with center at Q, a coordinate system (isometric to the usual
coordinate system) x'= (xi, , xn_I), xn, with origin at Q and a
function (h: Rn-I R such that
4(O)=O, IOW)--0(0I <Mlx'-y'I and Df1B = (X =(x,xn):xn>o(x'){(1B .
If for each Q the function 0 can be chosen in CI(Rn-I), then D

is called a C1 domain. If in addition, 74 satisfies a Holder condition
of order a,
IV(*')-V (y')I <-Clx'-y'la ,
we call D a Cl,' domain.
Notice that the cone re = l(x,xn):xn<-Mlx'1{ satisfies ren B CCD.
Similarly, ri = 4(x ,xn) : xn>M1x'I ! satisfies ri n B CD. Thus, Lipschitz
domains satisfy the interior and exterior cone condition.
The function 0 satisfying the Lipschitz condition 10(x')--*(y')I <
Mlx'- y'l is differentiable almost everywhere and 170 e L00(Rn-I), 111741I,, M.
136 CARLOS E. KENIG
Surface measure a is defined for each Borel subset E C d D fl B by
a(E) = f (1 + IO95(x')12)I
/2 dx
where E* = ix': (x', O(x')) eE 1.

The unit outer normal to dD given in the coordinate system by
/2
exists for almost every x'. The unit
(NO(x'), -1)/(1 + lvq,(X,)I 2)1
normal at Q will be denoted by NQ. It exists for almost every Q e dD ,
with respect to da.
In order to motivate the use of the method of layer potentials, we need
to recall some formulas from advanced calculus, and some definitions. We
will start with the derivation corresponding to the Dirichlet problem.
We first recall the fundamental solution F(X) to Laplace's equation
in Rn : AF = 8, where
1
(n-2)wnjX1n-2
n>2
F(X) =
log jXJ n 2
TIT
1-
where wn is the surface area of the unit sphere in Rn. F(X) is the
electrical potential in free space induced by a unit charge at the origin.
It provides a formula for a solution w to the equation Aa with
( eCp(Rn),
w(X) = F *#(X) = fF(xYY)dv.
Rn
It will be convenient to put F(X,Y) = F(X-Y). Notice that tYF(X,Y) _

5(X-Y). The fundamental solution in a bounded domain is known as the
Green function G(X,Y). It is the function on D xD continuous for X j Y
and satisfying A G(X,Y) = 5(X-Y), X e D ; G(X,Y) = 0, X e D, Y e dD.
G(X,Y) as a function of Y is the potential induced by a unit charge at

X that is grounded to zero potential on aD . The Green function can be
obtained if one knows how to solve the Dirichlet problem. In fact, let
uX(Y) be the harmonic function with boundary values uX(Y)laD = F(X,Y)IcD.
Then,
(1) G(X,Y) = F(X,Y)-uX(Y) .
On the other hand, if we know G(X,Y), we can formally write down the
solution to the Dirichlet problem.
In fact,
u(X) = 5u(Y)(XY)dY = r u(Y)AYG(X,Y)dY =

D JD
= 5Eu(Y)AyG(xv) - Au(Y) G(X,Y)ldY =

D
f[u(Q) (X,Q) - (Q)G(X,Q)J da(Q) =

Q Q
aD
= Ju(Q) aN G(X,Q)da(Q) ,
4
aD
where the fourth equality follows from Green's formula. Thus, we have
derived the formula
(2) u(X) = 5f(Q) aN (X,Q)da(Q)

Q
aD
for the harmonic function u with boundary values f. The problem with
138 CARLOS E. KENIG
formula (2) is of course that we don't know G(X,Q). Because of formulas

(1) and (2), C. Neumann proposed the formula
w(X) = 5f(Q) aN (X,Q)da(Q) _

Q
0
1
n IX--Q InI<X_Q,NQ> da(Q)
aD
as a first approximation to the solution of the Dirichlet problem, Au = 0

in D, uI D=f.
w(X) is known as the double layer potential of f. First of all w is
harmonic in D. Also, one can show that as X - Q e W, w(X) 2 f(Q) +
Kf(Q), where K is the operator on aD given by
Kf(Q) = wn < f(P)do(P) .

J p Q Inp
an
If Kf were zero, we would be done, and in some sense it is true that Kf

is small compared to 2 f, when the domain D is smooth. In fact, aD
has dimension n-1 , while it is easy to see that if aD is C,
<Q -P,Np C
IP-Qin IP-QIn-2
Thus, the operator K: C(3D) C(aD) is compact. Therefore, by the

general theory of Fredholm, the operator T = 2 I+K is invertible modulo
a finite dimensional subspace of C(3D). If D and cD are connected,
T is actually invertible on C(3D). Therefore, the solution to the
Dirichlet problem may be written
<X-Q,N >
u(X) = wn g(Q) IX-Ql do(Q),
dD
where g = T-1(f). The operator K is compact on C(O) even in C1'a

<Q-P,Np > C
domains because in this case I
_Qln I< l+a) and so this
P n
Q-
procedure solves the classical Dirichlet problem in that case too. If D
is a C domain, K is compact from Ck'a(aD) to Ck,a(3D),
k = 1, 2, , 0 < a < 1. Hence, if f e C(aD), u c C'(5). This approach
can also be used to obtain results for f c Lp(aD) in C domains, and
even on C1'a domains. We will now sketch the extension of the theorem
for the ball stated in the introduction to C1'a domains. We first define
non-tangential approach regions as follows:
r (Q) _ IX eD : IX-Q I < (1+(3)dist (X, 3D)} .
The non-tangential maximal function, with opening 16, of a function w

defined in D is
Na(w) (Q) = sup I lw(x)j : X er0(Q)} .
IP--P,QNp>
Cl,a
Because of the estimate I <Q I on
n
domains, it is easy to see that K is compact as a mapping on LP(dD).
Also, standard arguments show that
N16 (w) (Q) <Ca}M(f)(Q)+M(Kf)(Q)1 ,
where M is the Hardy-Littlewood maximal operator on X, and w is

the double layer potential of f. Finally, from LP bounds for M, K and
T-1 = \2 I+K)-1
, one obtains:
140 CARLOS E. KENIG
THEOREM. Let D be a Cl,a domain, 1 
and u(x) = f3D
T-If(Q)da(Q), then JIN0uIILp(do)
(On IX-Qln <
C 1 1 f 11
w and the harmonic function u tends to f non-tangentially.
p LP(da)
The difficulty in the case of CI and Lipschitz domains, is that, in

this case the size-estimate on the kernel of K is
<Q-P,Np C
lP-Qln < IP-QIn-I
and so, even the LP boundedness, much less the compactness of K, is
far from obvious.
Let us now turn to the Neumann problem. Let D be a smooth domain.
We seek to solve Au = 0 in D, !AL I = f. By Green's formula, we
aD
must have f3D fda = 0. When D and cD are connected this is the only
compatibility condition needed. We will only consider that case for
simplicity. A good first guess at the solution u is the so-called single
layer potential of f given by v(x) = faD f(Q)F(X,Q)da(Q) =
IX-QlaNda(Q). Once again v is harmonic in D, and

1 n-2
f(Q)
Cn faD
(Q) = 2 f(Q)-K*f(Q), where K* is the adjoint of K above, i.e.

*Q 1 <P-Q,Np>
K f(P) = Wn 13D IP-Qln f(Q)da(Q). Thus, K is compact from
Ck,a(aD) to Ck,a(aD), and Fredholm's theory says that =2 I-K* is

invertible on the subspace of Ck'a (aD) of functions of mean value 0.
Therefore the solution to the Neumann problem can be written
u(X) = f(T1f)(Q)F(X,Q)da(Q),
aD
and if f a C(3D), u e C(D). If D is CI'a, T is also invertible on

the subspace of Lp((3D) of functions with mean value, 1 < p < w. Hence:
THEOREM. Let D be a C I'a domain, (D and cD connected). Let

1 , and
JIN,(Vu)JJLp(da) < CplifilLp(do).
What do we do when c9D is merely CI , or even merely Lipschitz?

As I mentioned before, the LP boundedness of K is even in doubt. In
1977, A. P. Calderon ([1]) showed that for any CI -domain, K: LP(OD) -,
Lp(09D), 1 < p < oo is a bounded operator. Shortly afterwards, Fabes,
Jodeit and Riviere ([11]) showed that K is in fact compact in this case.
They were thus able to extend the theorems above to the case of CI
domains.
Before going on to the main subject matter of this paper, i.e. the
method of layer potentials on Lipschitz domains, I want to discuss
another important method for the Dirichlet problem for Laplace's equation.
(b) The method of harmonic measure

Another way of studying the Dirichlet problem for Laplace's equation
is in terms of the notion of harmonic measure. Let D be a bounded
Lipschitz domain in Rn. As we mentioned before, then D satisfies the
exterior cone condition, and so, by a classical result of Zaremba and
Lebesgue, we can solve the classical Dirichlet problem for A in D ,
for any f e C(dD). Given X e D , the maximum principle shows that the
mapping f " u(X) defines a positive continuous linear functional on
C(09D). Therefore, by the Riesz representation theorem, there is a
unique positive Borel probability measure wX on 3D such that
u(X) = ff(Q)da(Q)
dD
142 CARLOS E. KENIG
wX is called the harmonic measure for D , evaluated at X . For example,

harmonic measure for the unit ball B , evaluated at the origin is a constant
multiple of surface measure: wo = a/a(cB). (This follows from the mean
value property of harmonic functions.) For different X, the measures wX
are mutually absolutely continuous (a simple consequence of Harnack's
X
principle). We fix X0 e D, and denote w = w 0. The importance of
harmonic measure to the boundary behavior of harmonic functions on
Lipschitz domains can be illustrated by the following theorem of Hunt and
Wheeden (1967): If u is a positive harmonic function in a Lipschitz
domain D , then u has finite non-tangential limits almost everywhere
with respect to w. Conversely, given any set E C OD, with w(E) = 0,
there is a positive harmonic function u in D with lim u(X) = + oo as
X Q, for every Q e E. Despite its advantages, harmonic measure has
some inherent difficulties. First, it is hard to calculate it explicitly.
Second, it is tied up to the maximum principle, positivity, and the Harnack
principle, and so it is not useful for the Neumann problem, or for the
Dirichlet problem for systems of equations.
In general, harmonic measure may be very different from surface
measure. If D is a CI,a domain, then harmonic measure and surface
measure are essentially identical in that each is a bounded multiple of
the other. This can be proved by the classical method of layer potentials.
Along the same lines, as we saw before, one can use layer potentials to
solve the Dirichlet and Neumann problems with boundary data in L.
On CI domains, it is no longer true that harmonic measure is a
bounded multiple of surface measure, or vice versa. Moreover, as was
explained before, the applicability of the method of layer potentials is not
obvious. The situation for general Lipschitz domains is even less obvious.
In 1977, B. E. J. Dahlberg ([4]) proved that on a CI or even a
Lipschitz domain, harmonic measure and surface measure are mutually
absolutely continuous. Using a quantitative version of mutual absolute
continuity, and the theory of weighted norm inequalities, he proved ([5])
that in a Lipschitz domain D one can solve the Dirichlet problem as in

the theorem above with f ( L2(aD, d o) . In fact, he showed that given a
Lipschitz domain D , there exists s = (D) such that this can be done
for f ( LP(dD, do), 2-e < p < o. Also, simple examples to be presented
later show that given p < 2, we can find a Lipschitz domain D for
which this cannot be done in L. By establishing further properties for
harmonic measure on C 1 domains, he was able to show the results above
in the range 1 < p < oo for C1 domains. (The best possible regularity
result for harmonic measure on C1 domains is due to D. Jerison and
C. Kenig (1981): if k =4, then log k ( VMO(aD).)
A shortcoming of Dahlberg's method of proof, as was explained before,
is that, by studying harmonic measure, it relied on positivity and the
Harnack principle. This made the method inapplicable to the Neumann
problem, or to systems of equations. Also the method does not provide
useful representation formulas for the solution.
(c) The method of layer potentials revisited

In 1979, D. Jerison and C. Kenig [16], [17] were able to give a simpli-
fied proof of Dahlberg's results, using an integral identity that goes back
to Rellich ([30]). However, the method still relied on positivity. Shortly
afterwards, D. Jerison and C. Kenig ([18]) were also able to treat the
Neumann problem on Lipschitz domains, with L2(aD, do) data and optimal
estimates. To do so, they combined the Rellich type formulas with
Dahlberg's results on the Dirichlet problem. Thus, it still relied on
positivity, and dealt only with the L2 case, leaving the corresponding
LP theory open.
In 1981, R. Coifman, A. McIntosh and Y. Meyer [2] established the
boundedness of the Cauchy integral on any Lipschitz curve, opening the
door to the applicability of the method of layer potentials to Lipschitz
domains. This method is very flexible, does not relie on positivity, and
does not in principle differentiate between a single equation or a system
of equations. The difficulty then becomes the solvability of the integral
144 CARLOS E. KENIG
equations, since unlike in the C I case, the Fredholm theory is not

applicable, because on a Lipschitz domain operators like the operator K
in part (a) are not compact, as simple examples show.
For the case of the Laplace equation, with L2(3D, da) data, this
difficulty was overcome by G.C. Verchota ([33]) in 1982, in his doctoral
dissertation. He made the key observation that the Rellich identities
mentioned before are the appropriate substitutes to compactness, in the
case of Lipschitz domains. Thus, Verchota was able to recover the L2
results of Dahlberg [5] and of Jerison and Kenig [18], for Laplace's equa-
tion on a Lipschitz domain, but using the method of layer potentials.
These results of Verchota's will be explained in the first part of Section 2.
In 1984, B. Dahlberg and C. Kenig ([16]) were able to show that
given a Lipschitz domain D C Rn , there exists E = E(D) > 0 such that
one can solve the Neumann problem for Laplace's equation with data in
Lp(r7D, do) , 1 < p < 2 +E E. Easy examples (to be presented later) show
that this range of p's is optimal. Moreover, they showed that the solu-
tion can be obtained by the method of layer potentials, and that Dahlberg's
solution of the LP Dirichlet problem can also be obtained by the method
of layer potentials. They also obtained end point estimates for the Hardy
space HI(dD,da), which generalize the results for n = 2 in [20] and
[211, and for C I domains in [12 ]. The key idea in this work is that one
can estimate the regularity of the so-called Neumann function for D , by
using the De Giorgi-Nash regularity theory for elliptic equations with
bounded measurable coefficients. This, combined with the use of the so-
called àtoms' yields the desired results. These results will be explained
in the second part of Section 2.
Also in 1984, B. Dahlberg, C. Kenig and G. Verchota ([8]) and
E. Fabes, C. Kenig and G. Verchota ([13]) were able to extend the ideas
of Verchota to be able to obtain results for L2 boundary value problems
for some systems of equations on Lipschitz domains. The systems
treated are those that arise in linear elastostatics and in linear hydro-
statics. The results obtained had not been previously available for
general Lipschitz domains, although a lost of work had been devoted to
the case of piecewise linear domains. (See [24], [25] and their bibli-
ographies.) For the case of CI domains, these results for the systems
of elastostatics had been previously obtained by A. Gutierrez ([15]),
using compactness and the Fredholm theory. This is of course, not
available for the case of Lipschitz domains. The authors use once more
the method of layer potentials. Invertibility is shown again by means of
Rellich type formulas. This works very well in the Dirichlet problem for
the Stokes system (see part (b) of Section 3), but serious difficulties
occur for the systems of elastostatics (see part (a) of Section 3). These
difficulties are overcome by proving a Korn type inequality at the
boundary. The proof of this inequality proceeds in three steps. One first
establishes it for the case of small Lipschitz constant. One then proves
an analogous inequality for non-tangential maximal functions on any
Lipschitz domain, by using the ideas of G. David ([10]), on increasing the
Lipschitz constant. Finally, one can remove the non-tangential maximal
function, using the results on the Dirichlet problem for the Stokes system,
which are established in part (b) of Section 3.
As a final comment, I would like to point out that even though through-
out this paper we have emphasized non-tangential maximal function esti-
mates, also optimal Sobolev space estimates hold. All the Sobolev esti-
mates can be proved in a unified fashion, using square functions and a
variant of some of the real variable arguments used in part (b) of Section 3.
The details will appear in a forthcoming paper of B. Dahlberg and
C. Kenig, [7].
2. Laplade's equation on Lipschitz domains

(a) The L2 theory
A bounded Lipschitz domain D C Rn is one which is locally given by
the domain above the graph of a Lipschitz function. Such domains satisfy
146 CARLOS E. KENIG
both the interior and exterior cone condition. For such a domain D, the
non-tangential region of opening J8 at a point Q c aD is I'18(Q) _
IX cD : IX-Qj < (1+,8)dist (X, 3))J. All the results in this paper are valid,
when suitably interpreted for all bounded Lipschitz domains in Rn,
n > 2, with the non-tangential approach regions defined above. For
simplicity, in this exposition we will restrict ourselves to the case n > 3
(and sometimes even to the case n = 3 ), and to domains D C Rn ,
D = I(x,y): y>4i(x)), where : Rn-I R is a Lipschitz function with
Lipschitz constant M, i.e. j0(x)-4(x')I < Mix-x'I . D- _ I(x,y):y<cb(x)1
For fixed M'< M, I'e(x) _ I(z,y): (y-(k(x)) < -M'Iz-xI4 C D and I'i(x) _
I(z,y): (y-0(x)) > M'Iz-xII C D. Points in D will usually be denoted by
X, while points on aD by Q = (x, (k(x)) or simply by x, Nx or NQ
will denote the unit normal to aD = A at Q = (x, 4(x)). If u is a func-
tion defined on Rn'A and Q c aD , u t(Q) will denote lim u(X) or
X-3Q
XEIi(Q)
lim u(X), respectively. If u is a function defined on D, N(u)(Q) _

X-Q
Xc e(Q)
sup Iu(X)I.
XcF1(Q)
We wish to solve the problems
Au = 0 in D Au = 0 in D
(D) , (N)
u' aD = f c L2(aD,do) f c L2(3D,do)
a-N
aD
The results here are
THEOREM 2.1.1. There exists a unique u such that N(u) c L2((9D, do),
solving (D), where the boundary values are taken non-tangentially a.e. .
Moreover, the solution u has the form
- 1 <X-Q'NQ>
u(X) (On g(Q) do(Q)
J
IQ_XIn
aD
for some g c L2(3D, da).
THEOREM 2.1.2. There exists a unique u tending to 0 at -, such

that N(Vu) E L2(dD, do), solving (N) in the sense that f(Q)
as X -* Q non-tangentially a.e.. Moreover, the solution u has the form
u(X) = w n (n_2) 1 n-2 g(Q)do(Q) ,

1X-Q1
aD
for some g ( L20D, d u) .
In order to prove the above theorems, we introduce the double and

single layer potentials
<X-QNQ>
g(x) = (an J IX-QIn g(Q)do(Q)
aD
and
Sg(X) _ - wn( f
aD
I n 2 g(Q)do(Q)
IX-QI
If Q = (x, 4(x)) , X = (z ,y) , then
g(x) dx
g(z.Y)=w f [Ix-z I2 + [fi(x)-0(z)121n/2
Rn-I
Sg(z ,Y) _ -
ca
1
n-2)
'J r 1+IV _(X) I2
-n 2- g(x) dx
2
Rn-I [Ix-zI2+I4(x)-y]21
THEOREM 2.1.3. a) If g c LP(aD, do), 1 < p < -, then N(VSg), N(Xg)

also belong to Lp(aD, do) and their norms are bounded by CilgjI p
L ((9D,du)
148 CARLOS E. KENIG
(b) lim 1
r z)-q(x)-(z-x) VS6(x) g(x) dx = Kg(z) exists
Wn [IX-zI2 +[S6(x)-0(z)]2]n-2
Ix zI>E
a.e. and IIKgII <_CIIgII 1 <p<..

LP(aD,do) Lp(aD,do)
(z-x,(h(z)-S(x)) 1 +IW,(x)I2 g(x)dx

lim 1 exists a.e. and in
E- 06)n [Iz-x 12 + [S6(z) -
S6(x)j2]./2
1z-X I>E
Lp(aD, do), and its LP norm is bounded by CIIgII , 1 E
COROLLARY 2.1.4. (NzVSg)(z) g(z)-K*g(z), where K* is the

L2(aD,do) adjoint of K. 2
The proof of Theorem 2.1.3 a) follows by well-known techniques from

the deep theorem of Coifman, McIntosh and Meyer ([2]).
THEOREM ([2]). Let 0: R - R be even, and C'. Let A,B: Rn-1 - R

be Lipschitz. Let K(z,x) = A(z) -A(x 916 z-x6 x 1 . Then, the maximal
I
operator
M*g(z) = sup
E-0
I
f
It XI>E
K(z,x)g(x)dx(
is bounded on LP(Rn-1), 1 < p < o, with

IIM*gllp << CIlgllp
where C = C(M,0,n,p), and IIVAII00 < M, IIVBII. < M.
The proof of (b), (c) follow from the theorem above, together with the
following simple lemmas.
LEMMA. If f cCo(Rn-1), then
lim 1 c(z)-c(x)-(z-x) VOW f(x)dx -

E +0 Wn iz-xI>E [Ix-zI2+[O(x)-95(z)]2]n/2
n--1
_-
1 zk-xk
X
[(z)_(x)1 (x)dx
J n-I x-z axk
1
where a(0) = 0, X'(t) = (1 +t2)-n/2,

and where the equality holds at
every z at which 95 is differentiable, i.e. for a.e.z.
LEMMA. If a c Rn-1 , a (a), f c Co(Rn-1) and A is as in the

previous lemma, then
1
c.n
I a-O(a)- (a-x) V (x) f(x)dx =
[Ia-xI2+[0(x)-a]2]n/2
n-1
fix) a
=
2
sign (a --O(a)) f(a) - cI
on n fI xk ak
Ix-amn-I x--a)
c3f (x)dx
c7xk
Moreover, the integral on the right-hand side of the equality is a continu-

ous function of (a,a) c R.
It is easy to see that (at least the existence part) of Theorems 2.1.1
and 2.1.2 will follow immediately if we can show that (2 I+K) and
150 CARLOS E. KENIG
I+K* are invertible on L2(aD,do). This is the result of

G. Verchota ([33]).
THEOREM 2.1.5. I+K) , 2 I+K*) are invertible on L2(3D,do).

( 2 (+
In order to prove this theorem, it suffices to show that ( I+K*)J

2
L2(c3D,`do),
are invertible. In order to do so, we show that if f e

II 1 I+K* fIi
2
1 I-K fll
II , where the constants
L2(dD,do) (2 L2(aD,do)
of equivalence depend only on the Lipschitz constant M. Let us take
this for granted, and show, for example, that 2 I+K* is invertible. To
do this, note first that if T = 2 I+K*, IITf11L2 ? CIIfIIL2, where C
depends only on the Lipschitz constant M. For 0 < t < 1 , consider the
operator Tt = I + Kt , where Kt is the operator corresponding to the
2 by to. Then, To = 2 1, TI = T, and 03Tt :
domain defined y LP(Rn-1)
cit
Lp(Rn-1) , 1 C Ilf II L2 , C
L2
independent of t. The invertibility of T now follows from the continuity
method:
LEMMA 2.1.6. Suppose that Tt : L2(Rn-1) -. L2(Rn-1) satisfy

(a) IITtfIIL2 > C111flIL2
(b) IITtf-Tsf1IL2 < C2lt-sIIf!IL2. 0 < t,S < 1 .
(c) To: L2(Rn-1) , L2(Rn-1) is invertible.
Then, T1 is invertible. The proof of 2.1.6 is very simple. We are

I-K*fII
thus reduced to proving (2.1.7) II(2 I+K*)fII 2
Od a` 11
JI 2
In order to prove (2.1.7), we will use the following formula, which goes
back to Rellich [30] (also see [28], [29], [27]).
LEMMA 2.1.8. Assume that u e Lip (6), Du = 0 in D, and u and its

derivatives are suitably small at -. Then, if en is the unit vector in

the direction of the y-axis,
f<Nen>iVui2da = 2 ay ado.
aD aD
Proof. Observe that div(enIVu12) _ 1 Vu 12 = 2 Vu - Vu, while
div a Vu = - Vu Vu + . div Vu = - Vu Vu . Stokes' theorem now

gives the lemma.
We will now deduce a few consequences of the Rellich identity. Recall

that N xx (-V (x),1)/X/1 + IVO(x)12, so that 1 < <Nx ,en> < 1 .
(1+M2)I/2
COROLLARY 2.1.9. Let u be as in 2.1.8, and let T1(x), T2(x),

Tn_I(x) be an orthogonal basis for the tangent plane to dD at (X, O(X)).
n-1
Let Iptu(x)12 = I I<Vu(x),Tj(x)> 12. Then,
j=1
J( )2da<C fVtuI2da.
aD` t3D
Proof. Let a = en-<Nx,en>Nx, so that a is a linear combination of

T1 (x), T2(x), ...,Tn-1(x). Then, = <Nx,en> aN`+ <a,Vu>. Also,
IDu12 jVtu12, and so faD ( )2do+
= ( /2 + <Nx,en> faD<Nx,en>1VtuI2do=
2 faD
<Nx,en>( )2do+2
M faD <a,Vu> TN
('u) do. Hence faD<Nx,en>(o)2do=
faD <Nx,en> IVttu12da - 2 f <a,Vu> do. So, f jVtuI2do+

C faD
(faD 2do)1/2, and the corollary follows.
C( faD IVtuj2da)1/2
N
152 CARLOS E. KENIG
COROLLARY 2.1.10. Let u be as in 2.1.8. Then,
flVtul2da< c f( )2 do.
aD aD
Proof. faD IVuj2do < 2(f,3D IVu12d(,)1/2 GD I I2do1/2, by 2.18,

and the corollary follows.
olu
COROLLARY 2.1.11. Let u be as in 2.1.8. Then, faD IVtuI2da ti faD I I2 do
In order to prove 2.1.7, let u = Sg. Because òf 2.1.3c, ptu` is con
C0N`
tinuous across the boundary, while by 2.1.4, = (T. z I-K*) g . We
This
now apply 2.1.11 in D and 5, to obtain 1.1.7. finishes the proof
of 2.1.1 and 2.1.2.
We now turn our attention to L2 regularity in the Dirichlet problem.
DEFINITION 2.1.12. f E LP(A), 1 < p < ., if f(x, (x)) has a distribu-

tional gradient in Lp(Rn-1). It is easy to check that if F is any exten-
sion to Rn of f, then VxF(x,95(x)) is well defined, and belongs to
LP(A). We call this Vtf.The norm in LP(A) will be IIVtf1ILP(A)
THEOREM 2.1.13. The single layer potential S maps L2(A) into L2(A)
boundedly, and has a bounded inverse.
Proof. The boundedness follows from 2.1.3a). Because of the L2-Neumann

theory, and 2.1.11, IIDtS(f)I1 2(A) > CIIN(f)i 2 > CllfIi 2 . The
L L (A) L (A )
argument used in the proof of 2.1.5 now proves 2.1.13.
THEOREM 2.1.14. Given f c L2(A), there exists a harmonic function u ,

with IIN(Vu)IlL2(A) < CllVtfIl 2(A), and such that ptu = Vtf (a.e.) non-
L
tangentially on A . u is unique (modulo constants), and we can chose
u = S((), where g f L2(A).
The existence part of 2.1.14 follows directly from 2.1.13.
(b) The LP theory

We will start out our treatment of the LP theory by discussing some
counterexamples.
Let z = x +iy E C , and for 0<,8<2n, let
D18= 1z EC:jargzI <J3/21.
We will consider the holomorphic function f(z) = zn/R, which maps DR

conformally onto Dn, the right-hand plane. We will also consider a
bounded domain SZR CDR, with the property that an \0 is smooth, and
such that an fl 1zI < 1 # = aDR n ilzI < 1 !. Let u(x,y) = Ref(z), and
v(x,y) = u(x,y)jn v is harmonic in SZR, and v is identically zero near
R
the corner of aSZR, and is smooth everywhere else in 11 . Let s be
the arc length parametrization of an , starting at 0. Then, it is clear
that as E L(aS2R). Let w(x,y) = Im f(z)l9 . By the Cauchy-Riemann
equations,
EL(3113).
I I - ICI I
However, N(Vv) (s) = N(Vw) (s) ti s- 1+771 16This function belongs to

Lp(ds) if and only if p v/R - p > -1. Fix now a p > 2, and choose R
so close to 277 that p n/R-p < -1 . Then, N(Vw) / Lp(ai2R). If
I - K*J were invertible/ in LP(a10) , then, since a-N E L(a11R), we
\2
would have that S((2 I-K*) I (has a non-tangential
maximal function in LP(a11 R). By the L2-uniqueness in the Neumann
problem, w-w is constant in 9R, but this is a contradiction. This
shows that given p > 2 , we can find a Lipschitz domain so that
I -- K*J is not invertible in Lp. The example can also be used to
\2
154 CARLOS E. KENIG
show that I+K is not always invertible in Lq , when q < 2. In

fact, fix q < 2 , and let p satisfy + 9 = 1 . Choose f3 so that
p'-''-p<-1. Let B = and fix XeQ\IJzl<11.
X
Let o = w * be harmonic measure evaluated at X*, and k = d(''. We
rs
first claim that k / Lp(ds). In fact, let G(X) be the Green's function of
0 with pole at X*. Then, for s near 0, k(s) = (s) _
lim G(s+EN)-G(s) = lim G s+EN > C lim v s+eN = C av (s) s-1+v/I3,
E-e 0 a
e-.0 E E O E (3N
where the first inequality follows from the fact that both G and v are
positive, and harmonic on B, and 0 on ()iZP fl B (this is Lemma 5.10
in [19]). Assume now that 2 I+K were invertible on Lq(ds). Let
g ? 0 e C(dl3 ), and h(X) be the solution of the Dirichlet problem with
data g. Then,
h(X*) = gda, = J gkds
J
aiiP 0-n
also, by the L2-theory, h(X*) = x[(2 I+K)-1(g]

(X*), where x is
the double layer potential. Let U be a ball centered at X*, contained
in 9P. By the mean value property of harmonic functions and Harnack's
principle, we have
1 /q 1 /q
h(X*) < JJ fhc(fNhds) < C J ggds
because of the second formula for h(X*), and the assumed Lq bounded-
1
ness of (2 1+K . But this implies that k ( LP(ds), a contradiction.
We now turn to the positive results. They are-

THEOREM 2.2.1. There exists e = e(M) > 0 such that, given

f c L'(aD, da) , 2- e 
u(X) -
X Ql g(Q)da(Q) ,
n I
aD
for some g c LP(iD,da) .
THEOREM 2.2.2. There exists E = e(M) > 0, such that, given

f c Lp(3D, da), 2-E < p < ac, there exists a unique u harmonic in D,
tending to 0 at oc, with N(Vu) c LP(3D, da) , such that NQ Vu(X) con-
verges non-tangentially a. e. to f(Q). Moreover, u has the form
u(X) = g(Q)da(Q) ,
wn(n-2) J IX-Q In-2
aD
for some g e Lp(3D, d a).
THEOREM 2.2.3. There exists E = E(M) > 0 such that given f c LP(A),
1 < p < 2+E, there exists a harmonic function u, with IIN(Vu)lt <
L'(A) -
CIIVtfIILp(A), and such that Vtu = Vtf (a.e.) non-tangentially on
is unique (modulo constants). Moreover, u has the form
1 1
u(X) _ - g(Q)da(Q) ,
wn(n-2) fIX-QIn-2
aD
for some g c Lp(3D, d a).

156 CARLOS E. KENIG
The case p - 2 of the above theorems was discussed in part (a). The
first part of 2.2.1 (i.e. without the representation formula), is due to
B. Dahlberg (1977) ([51). Theorem 2.2.3 was first proved by G. Verchota
(1982) ([331). The representation formula in 2.2.1, Theorem 2.2.2, and
the proof that we are going to present of 2.2.3 are due to B. Dahlberg and
C. Kenig (1984) ([61). Just like in part (a), 2.2.1, 2.2.2, and 2.2.3 follow
from.
THEOREM 2.2.4. There exists E = E(M)/> 0 such that r+ - I - K*) is

i\`nvertible
invertible in Lp(dD, do), 1 < p < 2 + E, 2 1+9 is in
\
LP(aD, da), 2-E < p < co , and S : LP(dD, do) Lp(aD, do) is invertible
1 <p<2+E.
In order to prove Theorem 2.2.4, just as in part (a), it is enough to
show that if u = Sf , f nice, then, for 1 < p < 2+E, IIVtuIILp(aD,da) ti
This will be done by proving the following two theorems:

IIONI LP(dD,da)
THEOREM 2.2.5. Let Au = 0 in D. Then JJN(Vu)JJ <

LP(aD,da)
1 < p < 2+E .
C IIaNIILp(aD,da)'
THEOREM 2.2.6. Let Au = 0 in D. Then 1lN(Vu)JJ <

Lp(dD,da)
Clivtull , 1 <p<2+E.
LP( dD,do)
We first turn our attention to the case I < p < 2 of Theorem 2.2.5. In
order to do so, we introduce some definitions. A surface ball B in A
is a set of the form (x,4(x)), where x belongs to a ball in Rn-I .
DEFINITION 2.2.7. An atom a on A is a function supported in a

surface ball B, with hail L < 1/a(B), and with fA ada = 0. Notice
that atoms are in particular L2 functions. The following interpolation
theorem will be of importance to us.
THEOREM 2.2.8. Let T be a linear operator such that llTffI <

L2(A)
C IIf11L2(A), and such that for all atoms a, IITaII I,(A) < C. Then, for
1 < p <2, IITfIILp(A) IIf1ILp(A).
For a proof of this theorem, see [3].

Thus, in order to establish the case 1 < p < 2 of 2.2.5, it suffices to
show that if a = au is a atom, then IIN(Vu)II < C . By dilation and
L1(A)
translation invariance we can assume that 0(0) = 0 , supp a C BI =
Let B* be a large ball centered at (0,0) in Rn,
[(x, O(x)) : xl < 11.
which contains (x, qS(x)), jxl < 2. The diameter of e depends only
on M. Since IIa II = C, by the L2-Neumann theory,
o(Bjl
L2(A) /2
I
faDnB* N(Vu) < C(faDnB* N(Vu)2da)1 /2 < C. Thus, we only have to

estimate JCB*naD N(Vu)do). We will do so by appealing to the regu-
larity theory for divergence form elliptic equations. Consider the bi-
Lipschitzian mapping (D: D - D- given by 4D(x,y) = (x, O(x) - [y-(x)1).
Define u* on D- by the formula u* = u at-1 , u* verifies (in the weak
sense) the equation d iv (A(x,y) Vu*) = 0, where A(x,y) _ J X) (X) ,
where X = 1-1(x,y) . It is easy to see that A c L(D-), and
< A(x,y)e, t; > > CI; I2 . Notice also that supp dN C B 1 < B* fl aD. Define now
I for (x,y) e D u(x,y) for (x,y) c D

B(x,y) = , and u (x, y)
A(x,y) for (x, y) c D u*(x,y) for (x,y) c D-
Because 0 = 0 in 3D\B* , it is very easy to see that u is a (weak)

solution in Rn\B* of the divergence form elliptic equation with bounded
measurable coefficients, Id' =div B(x,y)ju = 0. In order to estimate u,
(and hence Vu ) at -, we use the following theorem of J. Benin and
H. Weinberger ([31]).
158 CARLOS E. KENIG
THEOREM 2.2.9. Let u solve Lu = 0 in Rn\B*, and suppose that

Ilall < -. Let g(X) solve Lg = 0 in IXI > 1 , with
g(X) IXI2-n. Then, u(X) = u. + ag(X) + v(X), where Lv = 0 in Rn\B*,

and Iv(X)I < C Ilull . IXI2-n -v, where v > 0, C > 0 depend
L(Rn\B * )
only on the ellipticity constants of L. Moreover, a = C JB(X) Vu(X) . V &(X) ,
where i/i c C(Rn),- = 0 for IXI in 2B*, and 0 = 1 for large X.

Let us assume for the time being that u is bounded and let us show
that if a is as in 2.2.9, then a = 0. Pick a tA as in 2.2.9. In D,
B(x) = I , and so fD = f Vu V & = lim f Vu Vi , where
D D E-0 DP
DP = {(x,y): I(x,y)I < p, y > 4(x)+e 1, and p is large. The right-hand
side equals lim faD1 aN
= lim J since, by the
e-0 ayo aDp
harmonicity of u, Jape = 0. Let ct! 1 = {(x,y) c aDp : y > fi(x)+e
a-N
P
and cDe = aDe\ (VP, I. Then, lim f [t/i-1 ]
au = lim f [0-11 au +
P12 P 3Dp -3N e-,0 aDP 1
TR
1] = f [0-11 a = JaD t/ia - JOD a a = 0, since

E -0 J3DP,2 [ J0
i/i = 0 on supp a. Moreover, JD_ BVuo = fD Vu where 1*

by our construction of B. The last term is also 0 by the same argument,
and so a = 0. We now show that u (and hence u ) is bounded. We will
assume that n > 4 for simplicity. Since Ila 11 < C, we know that
L2(A)
u(X) f(Q)2 da(Q), with IIfII <C. Now, for X cD1 =
=cn JaD
IX-QI L2(A)
(x,y): y > 4(x)+1 {, 1
IX-QIn-2 < C c L2(A), and so u c L(D1).
IQIn-2
1+
Let now B be any ball in Rn so that 2B C Rn\B*, B is of unit size,

and such that a fixed fraction of B is contained in D 1. Since
N(Vu) c L2(A), with norm less than C, J2Bf1D IDuI2 < C, and moreover
on B f1D 1 , lu(X)I < C. Therefore, by the Poincare inequality
J I2 < C. But, since ' solves Lu = 0, max [u1 < C(f rul21 /2 < C
2B B 2B
([261). Therefore, u E L(Rn\B*), IrII LOO(Rn\B*) < C. Hence, since
a = 0, Vu=Vv, and Iv(x,y)I VC/(IxI +IyI)n-2+v'
v > 0. For R > R0 =

diam B*, set b(R) = fA N(Vu)2, where, A ( x, I 2R
R
For each fixed R , let N1(Vu)(x) = sup I Ipu(x,y)I : (z,y) E I'i(x),
dist((z,y), 9D) < SRl, N2(Vu)(x) = sup flVu(z,y)I : (z,y) E Fi(x),
dist((z;y), (3D) > SR}. In the set where the sup in N2 is taken, u is
harmonic, and the distance of any point X to the boundary is comparable
to IXI. Thus, using our bound on v, we see that N2(Vu)(x) <
C/IXIn-1+v ti C/Rn-l+v, and so JA N2(Vu)2 < CR1-n-2v
Let now
R /
R(x,y) : 4(x) < y < O(x) + CR, rR < Ix I < r--1R 1, r c \4 , 2) . By the
L2-Neumann theory in fZr, JA, N1(Vu)2da < C J Ioul2da. Integrating
R r
in r from 1 /4 to 1 /2 gives
f NI(Du)2du < I Ipul2dX < 3 r u2

R ,J R J
AR Q1/4\Q1 /2 C1R<IXI<C2R
since Lu = 0 (see [26] for example). The right-hand side is bounded by

C 1 Rn = CR1-n-2v, and hence b(R) < CR1-n-2v. Then,
R3 R2(n-2)-2v
n-1
/2
JA R N(Vu) < C(fAR N(Vu)2)1 R 2 < CR-v. Choosing now R = 2j,
and adding in j , we obtain the desired estimate.

We now turn to the case I < p < 2 of 2.2.6. We need a further
definition.
DEFINITION 2.2.10. A function a is an H atom if A = Vta satisfies

-1
(a) supp A C B, a surface ball, (b) IIAII < 1/a(B), (c) f Ada= 0.
We will use the following interpolation result:
160 CARLOS E. KENIG
THEOREM 2.2.11. Let T be a linear operator such that IITfII <

I.2(A)
CIIfIIL2(A) and IITa1ILI(A)<C for all Hi atoms a. Then, for
1
1 < p <2, IITfIILP(A)<CIIf1ILP(A).
1
Hence, all we need to show is that if Au = 0, Qtu = Vta, and a is

a unit size Hi atom, N(Vu) E L1(A). But note that if we let
u(x,y) (x,y) c D
u(x,y) _
-u*(x,y) (x,y) E D_
then u is a weak solution of Lu = 0 in Rn\B*, since uLaD\B 0.

Then, u = -i ag + v , but a = 0 since 'u -'U. must change sign at .
The argument is then identical to the one given before.
Before we pass to the case 2 < p < 2 +e, we would like to point out
that using the techniques described above, one can develop the Stein-
Weiss [32] Hardy space theory on an arbitrary Lipschitz domain in Rn.
This generalizes the results for n = 2 obtained in [20] and [21], and the
results for CI domains in [12].
Some of the results one can obtain are the following: Let Hit(3D) _
I Xiai : IIXiI < oo, ai is an atoml, Hi at(aD) = I Aiai: XIAiI < +,
ai is an Hi atoms.
THEOREM 2.2.12. a) Given f c Hat(cD) , there exists a unique harmonic

function u, which tends to 0 at -, such that N(Vu) ( LI(3D), and
such that NQVu(X) -, f( Q) non-tangentially a.e. Moreover, u(X) =
S(g)(X), g ( Hat . Also, ulaD E Hl,at(aD) b) Given f c H1,at, there
exists a unique (modulo constants) harmonic function u , such that
N(Vu) F L1(cD), and such that VtulaD = Vtf a.e. Moreover, u = S(g),
g c Hat , and a E Hat(aD). c) If u is harmonic, and N(Vu) E L1(oD),
then d) f c Hat(aD) if and only if

a E Hat(3D), ulaD E
-N
Hl,at(c3D).
N(VSf) E L1(3D), if and only if \2 I-K*) f E Ha't(dD).
We turn now to the LP theory, 2 < p < 2 +E. In this case, the results
are obtained as automatic real variable consequences of the fact that the
L2 results hold for all Lipschitz domains. We will now show that
IIN(?u)IILp(n) <CÌIILp(n) for 2 <p <2+E.
The geometry will be clearer if we do it in Rn , and then we transfer
it to D by the bi-Lipschitzian mapping (D: R+ -, D , 4D(x,y) = (x,y+q5(x)).
We will systematically ignore the distinction between sets in R+ and
their images under It.
Let y=t(x,y)ER+:Ix(<y1, y*=((x,y)ER+:alxl<yt, where a
is a small constant to be chosen. Let m(x) = sup IVu(z,y)I , and
(Z' Y) Ex+y
*(x)
m = sup IVu(z,y)I . Our aim is to show that there is a small
(z, y) E x+y*
Eo > 0 such that fm2+Edx < c f Ifl2+Edx

, for all 0 < E < EQ, where
f= clu . Let h = M(f2)1/2, where M denotes the Hardy-Littlewood
maximal operator. Let EX = Ix E Ra-1 m* (x) > At. We claim that
f m 2 < CA2IEAI + Ca f m2. Let us assume the claim, and
Im*>A,h<AI Im*>X
prove the desired estimate. First, note that
fm2< f<X1 m2 + m2 < CX2IEXI + Ca J m2 + J m2 ,

J
EA t m *>I h>A t m *>A 1 I h>Jl l
by the claim. Choose now and fix a so that 1/2 . Then,

fEX f m2. For E>0, fm2+E=E fAE-1 f m2dA<
EA Ih>At 0 Im>AI
E f0 AE-1 fEA m2dA < C E fO Al+E IIm* >AIIdA+CE fQ c'o AE 1(fh>A m2)dA.
162 CARLOS E. KENIG
By a well-known inequality (see [141 for example), IEAI <

Thus, fm2+E < CE fo AI+E Iim>a)lda+CE fo aE-1( fh>A m2)dA < Cc f M2+,+
C fm2 hE. If we now choose EO so that CEO < 1/2, for E < EO, fm2+E <
C f m2 hE. If we now use Holder's inequality with exponents 2 2 E and
2 2+E E
2+E , we see that fm2+E < C(f m2+E)2+E(fM(f2) 2 )2+E, and the
E
desired inequality follows from the Hardy-Littlewood maximal theorem.

It remains to establish the claim. Let {Qkl be a Whitney decomposi-
tion of the set EA = Im*>AI, such that 3Qk C EA, and 13Qk1 has
bounded overlap. Fix k , we can assume that there exists x c Qk such
that h(x) < X, and hence, f2Q f2 < Ca2lQk l . For 1 <r <2, let Qk,r =
k
rQk, and Qk,r = (x.y) : x c rQk , 0 < y < r length (Qk)}. Qk,r (and
(D(Qk,r) ) is a Lipschitz domain, uniformly in k,r. Also, by construction
of Qk, there exists xk with dist(xk,Qk) ti length (Qk), and such that
m*(xk)<A. Let
Ak,r = aQk,r fl xk + y
Bkr = aQkrflR+\Akr
so that aQk,r = Qk,r U Ak,r U Bk,r. Note that the height of Bk,r is
dominated by Ca length (Qk), and that IVul <,k on Ak,r. Let ml
be the maximal function of Vu, correspond ing to the domain Qk,r (i.e.
where the cones are truncated at height ti P(Qk) ) Then, for x c Qk,
m(x) < mI(x) + X. Also,
f m i < (using the L2-theory on
C J IVuI2da+ c flvuI2da+c f f2 < C J IVul2da+c)2IQkl .
Bk,r Ak,r 2Qk Bk,r

ELLIPTIC BOUNDARY VALUE PROBL EMS 163
Integrating in r between 1 and 2, we see that

a length(Qk)
5 m21 <- length(Qk)
C
f0
2Qk 2Qk
Thus, fQ m2 < Ca f2Q m2 + CA2IQkl. Adding in k, we see that

k k
f < Ca2IEAl + Ca f m2 , which is the claim. Note also

im*>A,h<,1} im*>AI
that the same argument gives the estimate IIN(Vu)IIp < CIIVtullp
2 < p < 2+e, and the LP theory is thus completed.
3. Systems of equations on Lipschitz domains

(a) The systems of elastostatics.
In this part we will sketch the extension of the L2 results for the
Laplace equation to the systems of linear elastostatics on Lipschitz
domains. These results are joint work of B. Dahlberg, C. Kenig and
G. Verchota, and will be discussed in detail in a forthcoming paper ([81).
Here we will describe some of the main ideas in that work. For simplicity
here we restrict our attention to domains D above the graph of a
Lipschitz function 0: R2 y R.
Let a > 0, > 0 be constants (Lame moduli). We will seek to
solve the following boundary value problems, where u = (u1,u2,u3)
0u+(a+)Vdivu=0 in D
(3.1.1)
ulaD =f eLZ(aD,da)
0u + (k +) p div u = 0 in D
(3.1.2)
A(div u)N+l pu+(pu)tlNlaD =f EL2(aD,da).
164 CARLOS E. KENIG
(3-1 -1) corresponds to knowing the displacement vector u on the

boundary of D, while (3.1.2) corresponds to knowing the surface stresses
on the boundary of D. We seek to solve (3.1.1) and (3.1.2) by the method
of layer potentials. In order to do so, we introduce the Kelvin matrix of
fundamental solutions (see [24] for example), r(X) = (rij(X)), where
..(X) = A S'1 'XI r1+ 1

and A J2Lu C= 1 1_
17 4n IXI 41r- IXI3 2 2[
We will also introduce the stress operator T, where Tu = A (div u) N +
ulVu + Vu tIN.
The double layer potential of a density g(Q) is then given by u(X)
Xg(X) = fdD (T(Q)F(X-Q)}tg(Q)da(Q), where the operator T is applied
to each column of the matrix F.
The single layer potential of a density g(Q) is
u(X) = Sg(X) _ fr(xQ). g(Q)da(Q)

dD
Our main results here parallel those of Section 2, part a). They are
THEOREM 3.1.3. (a) There exists a unique solution of problem 3.1.1 in

D, with N(u) c L2(aD,da). Moreover, the solution u has the form
u(X) = ((X), g c L2(dD, do).
(b) There exists a unique solution of (3.1.2) in D, which is 0 at infinity,
with N(17q) c L2(r7D,do). Moreover the solution u has the form u(X) _
Sg(X), g c L2(dD,da).
(c) if the data f in 3.1.1 belongs to Li(dD1 da), then we can solve
(3.1.1), with N(V) c L2(iD,da). Moreover, we can take
u(X) = Sg(x), g c L2(aD, da) .
The proof of Theorem 3.1.3 starts out following the pattern we used
to prove 2.1.1, 2.1.2 and 2.1.14. We first show, as in Theorem 2.1.3. that
the following lemma holds:
LEMMA 3.1.4. Let Kg', Sj be defined as above, so that they both solve
0u + (A+)p div u = 0 in R3\oD. Then:
(a) IIN(xg)IILP(aD,do,) < C11111

LP((3D,da)'
IIN(osg)IILP(aD,da) <- CII g IILp((3D,da) for 1 <p<oo.
(b) (xi)(P) = g(P) + Kg(P)

2
: (Sg )j) = jA 2 ni(P)g1(P) - ni(P) nj(P) < N(P), g(P)>} +

C 1 I ll )))
+ (.v. f
aD
aPi
F(P-Q)g(Q)da(Q)
j
,
where K(P) = p.v. faD 1T(Q)F(P-Q)#t g(Q)da(Q), and A, C are the

constants in the definition of the fundamental solution.
Thus, just as in Section 2, part (a) is reduced to proving the inverti-

bility on L2(0,D, da) of 2 1+K, 2 I+K*, and the invertibility from
L2(aD, da) onto L2(aD, da) of S. Just as before, using the jump rela-
tions, it suffices to show that if u(X) = Sg(X), then IITulI ti
LZ(aD,da)
Il7tullL20D,da). Before explaining the difficulties in doing so, it is very
useful to explain the stress operator T (and thus the boundary value
problem 3.1.2), from the point of view of the theory of constant coefficient
second order elliptic systems. We go back to working on Rn, and use
the summation convention.
Let ars, 1 < r, s < m , 1 < i, j < n be constants satisfying the
ellipticity condition
alb eiejifrls > clel21ii12

166 CARLOS E. KENIG
and the symmetry condition ars = asr. Consider vector valued functions
u = (u1, ,um) on Rn satisfying the divergence from system
ars us = 0 in D. From variational considerations, the most
TT ij TX- T
natural boundary conditions are to Dirichlet condition (43D =f ) or the
s
Neumann type conditions, = n ars = f r . The interpretation of
av 1 ij
problem (2) in this context is that we can find constants ail , 1 < i, j < 3 ,
1 < r, s < 3, which satisfy the ellipticity condition and the symmetry con-
dition, and such that p0u + (A+p) V div u = 0 in D if and only if
aXi a.. dXsj = 0 in D, and with Tu = T u . In order to obtain the

equivalence between the tangential derivatives and the stress operator we
need an identity of the Rellich type. Such identities are available for
general constant coefficient systems (see [29], [27]).
LEMMA 3.1.5 (The Rellich, Payne-Weinberger, NeCas identities). Suppose

that ars a us = 0 in D, ars = a, h is a constant vector in Rn ,
axi tj axj >
and u and its derivatives are suitably small at oo. Then,
ars our aus dv = 2

f h ene ij axi aj rh our naarsej aus da
1 axi TX-j
aD aD
Proof. Apply the divergence theorem to the formula
3a [(h,ars - h rs -hj ars) ()ur

ij ej1 if axi am = 0
auaus
REMARK 1. Note that if we are dealing with the case m = 1 , aij = I ,

and we choose h = en , we recover the identity we used before for
Laplace's equation.
REMARK 2. Note that if we had the stronger ellipticity assumption that

ars
tj
i Sj >_ C 12
we would have, if X = (x, Rn-1 - R,
e,t
JJv Il < M1, that IiV uII In fact, if we take
O0 t L2(aD,da) ti Ildvll L2(aD,do)
h = en, then we would have
J Ipurl2do < C udo

fhnEajjrs TXi 3T
aD aD
I/2 I/2
('
2C fh, n
E
ars
Lj
s do, < 2C
i
J pur12 da J da
aD 030 D
(9D
Thus,
fa D IVur12du < C I3 D j
2 do.
For the opposite inequality, observe that, for each r,s,j fixed, the
vector hinLaes - hLnfars is perpendicular to N. Because of Lemma 3.1.5,
J j
I hen Pars X do = 2 f(hneahinLar) W Nj do.

a`D aD
Hence, faD DuI2do< C(j jptu12da)1/2('3D IVul2da)I/2 and so
Jii I2da<c Jf Ipui2do<

aD ft
REMARK 3. In the case in which we are interested, i.e. the case of the
systems of elastostatics,
168 CARLOS E. KENIG
s
ars au au r j i
au-
2
(div u)2 + ,
i , JX-i ' Fxj z ax; dx.
i.i
which clearly does not satisfy aid e, es > C t i

i2, since the
f
quadratic form involves only the symmetric part of the matrix ( ;r). In
this case, of course = Tu = A (div UN + u1Vu+pa IN.
REMARK 4. The inequality 11vu 11 2 < C11ptu11 L2 holds

L (aD,da) (3D,da)
in the general case, directly from Lemma 3.1.5, by a more complicated
algebraic argument. In fact, as in Remark 2, fft henears our aus

,, ax; axj da =
2 faD (h n a ij - h. n ars) aue , aus da and for fixed r, s, j ,
e e a e, ax; axj
(heeij
n ars - h' n ee,
ars) is a tangential vector. Thus, f
aD
h n ars our aus da <
N
2 I 12 -2 12
C (j IVtuI do) (f3D IVu1 do)' Consider now the matrix drs =
(ark n;nj)-1 . This is a strictly positive matrix, since ark e; ej77 t77 s >
C 1e121'i12 ta"
Moreover, d rs UV (a'A - ars
`àv i, Ri Ni
r
ùs = d rs n. art aut , n sm
i, Ni ek 3Xk 1
0,U m
ars our aus d rs nk art aut nmast our - atr at our = d rs nk art òut
ij JX; aXj kgaXQ my aXv vec)Xv aXe kv3Xv
nmameN, ave7v a Q
= 1drsnkarkv n ma ml - avf v X Now, note
that for t,r, f fixed, Idrsnkak1mast - atT is perpendicular to N, by

me ve
our definition of drs, and the symmetry of ail = drsnkak8mamenv - avenv =
n= atr
art
kvaknvdrs ast
me nm atr
mem vlv nkdrs asr
mem n -- atr
memn--Sis ast
mem n -atrmenm
amenm-amenm = 0. Therefore, JaD henedrsCdV (av da <

C( faD IVtuI2da)1/2(faD IVul2da)1/2. Now, N-S

il r - akjnkn] aN
rs (3us
niai] rs
-- akinknjni rs
aus = niai]rs c3us akinkni ni aus
ox rs -
()X] - =;niaij
A
arsn I aus __ (mars - arsn au. But, for i, r,s fixed, ars -
ki knin J aX i] ik k I J
I
3Xi i]
arsn n is perpendicular to N, and so f h n d artn n !

ikkl aD e e rs kjk] aN
a Qnineaur Iptul2da)1/2(faD loul2da)1/2+faD
da<CI(faD IVtul2dol.
We now choose h = en , so that hene > C , and recall that (drs) and
(artnkn) are strictly positive definite matrices. We then see that
I
aD (ID
Iotul2da J
OD
+ J jVtul2da
OD
.
-. 2 ,12
Now, as IVul = Iptul + aN, 2 , the remark follows.
2
REMARK 5. In order to show that IaD Ivtul do < C fan TuJ2da, it
suffices to show that faD Ipul2da< C faD IA(div u)I + pipu+Qut}I2da.
In fact, if this inequality holds, we would clearly have that f3D Ipul2da<
C faD Ipu+putl2 da (Korn type inequality at the boundary). The Rellich-
Payne-Weinberger-Necas identity is, in this case (with h = en ),
faD nn I 2 Ipu+vutl2 +A(div u)2da=2 IX(div

faD
But then, Ipu12da< C(faD 1Vu1
da)1/2(faD Ik(div u)N+p{pu+putlNl2do)1/2
431)
The rest of part (a) is devoted to sketching the proof of the above inequality.
THEOREM 3.1.6. Let u solve pAu + (A+)V div u = 0 in D, u = S(g),

where g is nice. Then, there exists a constant C, which depends only
on the Lipschitz constant of 95 so that
170 CARLOS E. KENIG
J IVu-12 da<C J jA(div)I+lpu+putl{2da.

aD aD
The proof of the above theorem proceeds in two steps. They are:
LEMMA 3.1.7. Let u be as' in Theorem 3.1.6. Then,
fN(v;)2 d a<C fN(A(div u)I+pIVu +putI)2do .

aD aD
LEMMA 3.1.8. Let u be as in Theorem 3.1.6. Then,
fN(x(div u)I+)pu+Vut})2da<C fijVu)I+Ipu+putIj da.

CID aD
Lemma 3.1.7 is proved by first doing so in the case when the Lipschitz
constant is small , and then passing to the general case by using the
ideas of G. David ([91 ). Lemma 3.1.8 is proved by observing that if v
is any row of the matrix A(div u)I+l4pu+putj, then v is a solution of
the Stokes system
Av = pp in D
(S) div v = 0 in D
vl3D = f f L2(dD,da)
This is checked directly by using the system of equations

pAu + (A+)p div u = 0. One then invokes the following Theorem of
E. Fabes, C. Kenig and G. Verchota, whose proof will be presented in
the next section.
THEOREM 3.1.9. Given f f L2(0,D, do), there exists a unique solution

(v, p) to system (S) with p tending to 0 at oc, and N(v) e L2(0,D,da).
Moreover {{N(v){{L2((?D,da) I1
< CIIf L2(aD,do)
We now turn to a sketch of the proof of Lemma 3.1.7., We will need

the following unpublished real variable lemma of G. David ([10]).
LEMMA 3.1.10. Let F : RxRn -> R be a function of two variables t c R,

x = (x1,. ,xn) E Rn. Assume that for each x , the function t H F(t,x)
is Lipschitz, with Lipschitz constant less than or equal to M, and for
each i , 1 < i < n , the function xi H F(t,x) is Lipschitz, with Lipschitz
constant less than or equal to Mi, for any choice of the other variables.
Given an interval I x J = I x J1 x x Jn , where the Ji's and I are
1 dimensional compact intervals, there exists a function G(t,x) : Rx Rn.R
with the following properties:

(a) G(t,x) > F(t,x) on I x j .
(b) If E = I(t,x) E I x J : F(t,x) = G(t,x)I, then IEI > 8 III IJ I
(c) For each i, the function is
Lipschitz, with Lipschitz constant less than or equal to Mi, and one
of the following statements is true:
Either for each x , -M < 29 (t,x) < 45 or

49t - ,
for each x (t,x) < M

- 45 < .
The proof of this lemma is the same as in the 1 dimensional case,

treating x as a parameter (see [9]).
Before we proceed with the proof of Lemma 3.1.7, we would like to
point out that in the analogue of Lemma 3.1.7 for bounded domains, a
normalization is necessary since if u(X) solves the systems of
elastostatics, so does u(X) + a + BX , where a is a constant vector,
while B is any antisymmetric 3x3 matrix. The right-hand side of the
inequality in the lemma of course remains unchanged, while the left-hand
side increases if B ìncreases.' The most convenient normalization is
that for some fixed point X* in the domain pu(X*)-pu(X*)t = 0. This
gives uniqueness modulo constants to problem 3.1.2 in bounded domains.
172 CARLOS E. KENIG
We now need to introduce some definitions. Let Do C Rn be a fixed,

C domain with {(x, 0) : IIIxIII = max Ixil < 1} C aDo, I(x,y):0 < y < 1 ,
IIIxIII<1ICDoCI(x,y):0<y<2, IIIxIII<2I. If 95:Rn-I->R is
Lipschitz, with IIIVoIII < M, we construct the mapping TO: Rn Rn by
TT(x,y) _ (x,cy+rly*O(x)) where rJ E Co(Rn-I) is radial, f rl = 1 , and
C = C(M) is chosen so that TOR+) C ((x,y) : y > o(x)I, and so that To
is a bi-Lipschitzian mapping. Also, it is clear that To is smooth for
(x,y) with y > 0, and T0(x,0) _ (x, O(x)). We will denote by Ao the
point T00,1). Lemma 3.1.7 is an easy consequence of
LEMMA 3.1.11. Given M > 0 and 0 with IIIVOIII < M, there exists a
constant C = C(M) such that for all functions u in Do, which are
Lipschitz in Do, which satisfy pAu +(A+iz)V div u = 0 in Do and
pu(Ao) = pu(Ao)t , we have IINo(V)IIL2(aD,do) < CIINo(A(div U _)I +
y1vu+DutIIIL20D,da). Here No is the non-tangential maximal operator
corresponding to the domain Do.
This lemma will be proved by a series of propositions. Before we

proceed we need to introduce one more definition. We say that Proposition
(M,E) holds if whenever 0 is such that IIIVOIJI < M, and there exists a
constant vector a with III a III < M so that III7O -a III < e, then, for all
Lipschitz functions u on Do with pDu + (a+)V div u = 0 in Do,
with Vu(Ao) = pu `(Ao) we have
IINO(ou)IIL2(aD,da) CIINo(a(div u)I+(Du+putIIIL2(dD,da)

< ,
where C = C (M, e) .
Note that if Proposition holds, then the corresponding estimates
automatically hold for all translates, rotates or dilates of the domains
Do, when 95 satisfies the conditions in Proposition (M,E). In the rest
of this section, a coordinate chart will be a translate, rotate or dilate of a
domain Do. The bottom Bo of aDo will be T95(dDOU (x,0): x ERn-I).
PROPOSITION 3.1.12. Given M > 0, there exists a = e(M) so that

Proposition (M,e) holds.
We will not give the proof of Proposition 3.1.12 here. We will just
make a few remarks about its proof. First, in this case the stronger
estimate IINA(Vu)IIL2(aD,do)<CIIA(div u)N+1A Vu+VutINIIL2(dD,do)
holds. This is because in this case, the domain DS6 is a small perturba-
tion of the smooth domain D. . For the smooth domain D_ , we can
ax aX
solve problem 3.1.2 by the method of layer potentials (see [24], for
example). If a is small, a perturbation analysis based on the theorem of
Coifman-Mclntosh-Meyer ([2]) shows that this is still the case. This
easily gives the estimate claimed above.
PROPOSITION 3.1.13. For all M > 0, e > 0, a e (0,0.1), if Proposition

(M,e) holds, then Proposition ((1-a)M,1.1e) holds.
We postpone the proof of Proposition 3.1.13, and show first how

Proposition 3.1.12 and Proposition 3.1.13 yield Lemma 3.1.11.
Proof of Lemma 3.1.11. We will show that Proposition (M,a) holds for any
M,e. Fix M,a, and choose R so large that if e(10M) is as in Proposition
R
3.1.12, then (1.1)Re(10M) > E. Pick now aj > 0 so that II (1-aj)=1/10
j=1
Then, since Proposition (10M, e(10M)) holds, by Proposition 3.1.12, apply-

ing Proposition 3.1.13 R times we see that Proposition (M,a) holds. We
will not sketch the proof of Proposition 3.1.13. We first note that it
suffices to show that
IINO(Du)IIL2(3D,da) <CIIRo(,k(div u)I+l pu+vutIIIL2(dD,da)
where N,6 is the nontangential maximal operator with a wider opening of

the non-tangential region. This follows because of classical arguments
relating non-tangential maximal functions with different openings (see [14]
174 CARLOS E. KENIG
for example). Pick now 0 with IIIVO-a III < 1-1e, 111-a 1115 (1-a)M. We
will choose No as follows: Since a1) 0\Bo is smooth, it is easy to
see that we can find a finite number of coordinate charts (i.e. rotates,
translates and dilates of Do ), which are entirely contained in Do ,
such that their bottoms By are contained in c3Do, such that
Tq((x,0): I IIx III < 1/2) cover dD(k, and such that the pi's involved
satisfy IIIVIII < (1 - 2) M, and there exist u such that III ao III
1.11E . The non-tangential region defining N0l, on TV (x,0) : 11141 < 2
is defined as follows: let F C be a closed set. Con

x,0): IIIxIII
2)
sider the cone on R+, y = I(x,y) E R+: b;xI < y}, where b is a small
constant. Consider now the domain DF on R+, given by
DF = U ((x,0)+y). Then DF is the domain above the graph of a
x(F
Lipschitz function 0, for which IIIVOIII < Cb, for some absolute
constant C (independent of F ). It is also easy to see that we can take
now b so small (depending only on M and E) that T,(DF) is the
domain above the graph of a Lipschitz function 0, with and
which satisfies IIIVq JII < (1 - 10) M, IIIVO 1.111E. The non
tangential region defining No, for Q e To ((x,0): IIIxIII < 2 is then
the image under TV of (x,0) + y, with b chosen as above, suitably
truncated, and where Q = TV,((x,0)). Let now, to lighten notation,
m=N,k(pu), m =NO(A(div)I+givu+V VI).
For t > 0, consider the open-set Et = Im > ti. We now produce a
Whitney type decomposition of Et into a family of disjoint sets IUj I
with the property that each Uj is contained in TO ((x,0): IIIxIII < 2)
for a coordinate chart Do , each Uj contains T0(I.), where Ij is a
cube in IIIxIII < 2 and is contained in TV,(Ij), where Ij is a fixed
multiple of Ij . Finally, we can also assume that there exists a constant
770 such that if diam (Uj) < no, there exists a point Qj in 3D0, with
dist (Qj,Uj) ti diam Up such that m(Qj) < t. Let now /3 > 1 be given.
We claim that there exists 5 > 0 so small that if Ej = Uj fl Im > Pt, Fn < 80
then a(E3) < (1-rJM)a(Uj) where 77M > 0. Assume the claim for the time
being. Then
f
00
f m2dc=2 J ta(Et)dt = 2162 ta(E/3tlt = 2/32 f ta(Ui n E/3t)dt <

dD0 0 0 0
00 00 00
<!r2/32 r +2/32 r ta(m>St)dt < 2/32(1-,1M) r ta(U )dt +

0 0 0
00
2
tof m>tIdt = (32 (1-77M) f m2da+ 5 m2da. Thus, if
S
0 aDo 3D0
we choose /3 > 1 , but so that /32 . (1- riM) < 1 , the desired result follows.
It remains to establish the claim. We argue by contradiction. Suppose not,
then a(EJ) > (1- r)M) a(Uj). Let E) = T,I(E)). If 71M is chosen suffi-
ciently small, we can guarantee that IEi -9911-1. Let now Fi _
Ei n Ij , and construct now the Lipschitz function Vi corresponding to it,
as in the definition of NS6 . Thus, > , III < (i_)vi
IIo - aIII < 1.111E . We now apply Lemma 3.1.10 to Vi, one variable
at a time, to find a Lipschitz function f, with f > 0 on Ij , such that if
F = Ix (Ii : f =01, then IFS nFjI > C a(Uwith IIIVfIII < (1 - 10) M,
and such that there exists af, with III afIII < (1 - 10} M so that
IIIVf-afIII < 5 1.111E < e . We can also arrange the truncation of our non-
tangential regions in such a way that on the appropriate rotate, translate
and dilate of Df (which of course is contained in the corresponding
coordinate chart associated to Dv,, which is contained in Do ),
IA(div u)I + jA Vu+ Vut I I <5 t . To lighten the exposition, we will still
denote by Df the translate, rotate and dilate of Df. Note that Proposi-
tion (M,e) applies to it. We divide the sets Ul into two types. Type I
176 CARLOS E. KENIG
are those with diam Ui > 710, and type 11 those for which diam Ui < ri0.
We first deal with the Uj of type I. In this case, Df has diameter of
the order of Because of the solvability of problem 3.1.2 for balls,
1 .
and our normalization, we see that on a ball B C Do, diam B 1 ,

A eB, we have fB Ipul2 <C fB Joining Af
to A,0 by a finite number of balls, and using interior regularity results
for the system EcAu + (A+) V div u = 0, we see that Ipu(A f)l < Cat, for
some absolute constant C. Then
C o(Uj)162t2 <
I
(Fif1Fi)
m2do<C f N2(pu)do,<
aDf
<C C J Nf v LVu(Af) - Vu t(Af)

2
do <-
a Df - 1)
< C a(U)52t2 + C J Nf(A(div u )I+(pu+putF)2da

aDf
by (M,e). The last quantity is also bounded by C a(Ui )a2t2, which is

a contradiction for small S.
Now, assume that Ul is of type II. Note that in this case there
exists Q) c 3Do, with dist(Qj,Uj) diam U), and such that IV (X)I < t
for all X in the nontangential region associated to Qj. Because of this,
it is easy to see, using the arguments we used to bound IV (Af)I in
case 1, that for all X in a neighborhood of Af and also on the top part
of Df, we have that IVu(X)I < t + Cat. Since for Q cTp(F,fIFj),
m(Q) > 16t, and (3 > 1 , if a is small enough, we see that we must have
Dut(Af)
Nf( Q) > m(Q). Hence, Nf Vu -r2u(Af) 2 (Q) > (R-1-Ca)t >
L
1
2
)t if S is small and Q E TIA (F f1Fj). Thus, applying (M,e) to Df,
[Vu(Af)
we see that CJ((,3-1)2t2o(U.) < J Nf pu
T (FJ.(1F.) C
2 Vut(Af
1)
2da<
J
LVu(Af) 20u t(Af1 )

faD Nf tpu - 2da < Ca(U)82t2 , a contradiction if S
f
is small. This finishes the proof of Proposition 3.1.13, and hence of
Lemma 3.1.11,
(b) The Stokes system of linear hydrostatics

In this part I will sketch the proof of the L2 results for the Stokes
system of hydrostatics. These results are joint work of E. Fabes,
C. Kenig and G. Verchota ([13]). We will keep using the notation intro-
duced in part (a).
We seek a vector valued function u = (u1,u2,u3) and a scalar valued
function p satisfying
Au = Vp in D
(3.2.1) div u = 0 in D
ulaD = f E L2(c3D, do) in the non-tangential sense.
THEOREM 3.2.2 (also Theorem 3.1.9). Given f r L2(0-D, do), there

exists a unique solution (u,p) to (3.2.1), with p tending to 0 at oo,
and N(u) e L2(c7D, da) . Moreover, u(X) = Kg(X) , with g e L2(r7D, d a) .
In order to sketch the proof of 3.2.2 we introduce the matrix r(X) of

fundamental solutions (see the book of Ladyzhenskaya [251), r(X)
1 aij
(T'ij(X)), where rij(X) = 8n IXI + 1 XiXj , and its corresponding pressure
8n X 3
X1
vector q(X) = q'(X)), where q'(X) = Our solution of (3.2.2) will
4nIXI3
be given in the form of a double layer potential, u(X) = Kg(X) _
faD (H'(Q)F(X-Q){g(Q)do(Q), where (H'(Q) r(X-Q))if = Sijge(X-Q) nj(Q) +
178 CARLOS E. KENIG
arig
(X-Q)nj(Q). We will also use the single layer potential u(X) _
Sg(X) = faD r(X-Q)g(Q)do(Q).
In the same way as one establishes 3.1.4, one has:
LEMMA 3.2.3. Let Kg, Sg be defined as above, with g c L2(cD, do,).

Then, they both solve Au = Vp in D, and D-, div u = 0 in D and
D-. Also
(a) IIN(Kg)IIL2((?D,da) III g IIL2(aD,aa)
'
(b) (Kg)(P) =
2 g(P) - p.v. fdD IH (Q)F(P-Q)Ij(Q)da(Q)
(c) IIN(VS0IIL2((3D,da) CII9IIL2(OD,da)

5
+ nl(P)g3(P) nl(P)n)(P)
(d) aXl (Sg )j (P) == +-.) <N(P), g(P)>
2 2
+ P.V. faD F(P - Q) g (Q) d a (Q)

aPi
(e) (HSg)(P) 2 g(P) + P.V. faD {H(P)r(P-Q))g(Q)da(Q) ,
where (H(X)r(X-Q))ig = ni(x) aXIP (X-Q) - Sil qf(X-Q)nj(X).

J
For the proof of this lemma in the case of smooth domains, see [25].
The proof of Theorem 3.2.2 (at least the existence part of it), reduces
to the invertibility in L20D, da) of the operator 2 I+K, where Kg(P)
- P.V. faD {H'(Q)T'(P-Q)jg(Q)da(Q). As in previous cases, it is enough
to show
(3.2.4)
I2
I-K*)2L (aD,da) +I(2
I+K*) 9
L2OD,da)
1
This is shown by using the following two integral identities.
LEMMA 3.2.5. Let h be a constant vector in Rn, and suppose that

Du = Vp, div u = 0 in D, and that u,p and their derivatives are
suitably small at co. Then,
f
aD
hFnPaS 1
a 1
Sdo=2 J
aD
hFaX
F
da-2 f
3D
PnS
h
fa F
da.
LEMMA 3.2.6. Let h,p and u be as in 3.2.5. Then,
J hFnFp2da=2 fhr aNpda-2 fhrA i aNda+
aD 3D 3D
+2 fhns_.da.
ass our
3D
The proofs of 3.2.5 and 3.2.6 are simple applications of the properties
of u,p, and the divergence theorem.
Choosing h = e3, we see that, from 3.2.6 we obtain
COROLLARY 3.2.7. Let u,p be as in 3.2.5. Then,
fP2do<C fIvn2da
aD aD
where C depends only on M.
A consequence of Corollary 3.2.7 and Lemma 3.2.5, is that, if
av aN
- then we have
180 CARLOS E. KENIG
COROLLARY 3.2.8. Let u,p be as in 3.2.5. Then,
fVt2da I fnS2da,
IN, 2
dam
l
aD aD i aD
where the constants of equivalence depend only on M.
Proof. 3.2.5 clearly implies, by Schwartz's inequality, that f D 1Vu 12dor <
2
C f,9D ICI do. Moreover, arguing as in the second part of the Remark 2
after 3.1.5, we see that 3.2.5 shows that
f pul2da<C Jf Vtu12da+I fPnsh-cda

e
aD aD I aD
By Corollary 3.2.7, the right-hand side is bounded by
1 /2 1 /2
(' j2da)
CJ Vu-1 2do
fins
aD ' (9D
3.2.8 follows now, using 3.2.7 once more.
To prove 3.2.4, let u = S(g). By d) in 3.2.3, Vtu and ns !2!!! are

1
continuous across 3D. Using this fact, 3.2.3 e) and Corollary 3.2.8,
3.2.4 follows.
In closing we would like to point out another boundary value problem

for the Stokes system, which is of physical significance, the so-called
slip boundary condition
(Au = Vp in D
(3.2.9) divu=0 in D
((pu+Vut)N-p-N)1j aD = f e L2(3D,do) .
This problem is very similar to (3.1.2). Using the techniques introduced

in part (a), together with the observation that if Au = Vp, div u = 0 in
D , the same is true for each row v of the matrix [Vu V t - p-11, we
have obtained
THEOREM 3.2.10. Given f c L2(3D,da) there exists a unique solution

(u ,p) to (3.2.9), which tends to 0 at -o, and with N(V ) e L2(3D,da).
Moreover, u(X) = S(g)(X) , with g e L2(3D, da) .
UNIVERSITY OF CHICAGO
CHICAGO, ILLINOIS 60637
REFERENCES
[1] A. P. Calderon, Cauchy integrals on Lipschitz curves and related
operators, Proc. Nat. Acad. Sc. U.S.A. 74(1977), 1324-1327.
[2] R. R. Coifman, A. McIntosh and Y. Meyer, L'integrale de Cauchy
definit un operateur borne sur L2 pour les courbes lipschitziennes,
Annals of Math. 116(1982), 361-387.
[3] R. R. Coifman and G. Weiss, Extensions of Hardy spaces and their
use in analysis, Bull. AMS 83 (1977), 569-645.
[4] B.E.J. Dahlberg, On estimates of harmonic measure, Arch. Rational
Mech. and Anal. 65 (1977), 272-288.
[5] , On the Poisson integral for Lipschitz and CI domains,
Studia Math. 66(1979), 13-24.
[6] B.E.J. Dahlberg and C. E. Kenig, Hardy spaces and the LP Neumann
problem for Laplace's equation in a Lipschitz domain, to appear,
Annals of Math.
[7] , Area integral estimates for higher order boundary value
problems on Lipschitz domains, to appear.
[8] D.E.J. Dahlberg, C.E. Kenig and G.C. Verchota, Boundary value
problems for the systems of elastostatics on a Lipschitz domain, in
preparation.
182 CARLOS E. KENIG
[9] G. David, Operateurs integraux singuliers sur certaines courbes du

plan complex, Ann. Sci. del'Ecole Norm. Sup. 17 (1984), 157-189.
[10] , personal communication, 1983.
[11] E. Fabes, M. Jodeit, Jr., and N. Riviere, Potential techniques for
boundary value problems on C1 domains, Acta Math. 141, (1978),
165-186.
[12] E. Fabes and C. E. Kenig, On the Hardy space H1 of a C1 domain,
Ark. Mat. 19(1981), 1-22.
[13] E. Fabes, C. E. Kenig and G.C. Verchota, The Stokes system on a
Lipschitz domain, in preparation.
[141 C. Fefferman and E. Stein, Hp spaces of several variables, Acta
Math. 129 (1972), 137-193.
[15] A. Gutierrez, Boundary value problems for linear elastostatics on
C1 domains, University of Minnesota preprint, 1980.
[161 D. S. Jerison and C. E. Kenig, An identity with applications to
harmonic measure, Bull. AMS Vol. 2 (1980), 447-451.
[17] , The Dirichlet problem in non-smooth domains, Annals of
Math. 113 (1981), 367-382.
[181 , The Neumann problem on Lipschitz domains, Bull. AMS
Vol. 4 (1981), 203-207
[19] , Boundary value problems on Lipschitz domains, MAA
Studies in Mathematics, Vol. 23, Studies in Partial Differential
Equations, W. Littmann, editor (1982), 1-68.
[20] C.E. Kenig, Weighted Hp spaces on Lipschitz domains, Amer. J.
of Math. 102 (1980), 129-163.
[21] , Weighted Hardy spaces on Lipschitz domains, Proceedings
of Symposia in Pure Mathematics, Vol. 35, Part 1, (1979), 263-274.
[22] , Boundary value problems of linear elastostatics and
hydrostatics on Lipschitz domains, Seminaire Goulaouic-Meyer-
Schwartz, 1983-84, Expose no. XXI, Ecole Polytechnique, Palaiseau,
France.
[231 , Recent progress on boundary value problems on Lipschitz
domains, to appear, Proceedings of Symposia in Pure Mathematics,
Proceedings of the Notre Dame Conference on Pseudodifferential
Operators, Volume 43 (1985), 175-205.
[24] V. D. Kupradze, Three dimensional problems of the mathematical
theory of elastocity and thermoelasticity, North Holland, New York,
1979.
[251 O. A. Ladyzhenskaya, The mathematical theory of viscous incom-
pressible flow, Gordon and Breach, New York, 1963.
[26] J. Moser, On Harnack's theorem for elliptic differential equations,

Comm. Pure Appl. Math. Vol. XIV (1961), 577-591.
[27] J. Necas, Les methodes directes en theorie des equations
elliptiques, Academia, Prague, 1967.
[28] L. Payne and H. Weinberger, New bounds in harmonic and biharmonic
problems, J. Math. Phys. 33 (1954), 291-307.
[29] , New bounds for solutions of second order elliptic partial
differential equations, Pacific J. of Math. 8 (1958), 551-573.
[30] F. Rellich, Darstellung der Eigenwerte von Au + Au durch ein
Randintegral, Math Z. 46 (1940), 635-646.
[31] J. Serrin and H. Weinberger, Isolated singularities of solutions of
linear elliptic equations, Amer. J. of Math. 88 (1966), 258-272.
[32] E. Stein and G. Weiss, On the theory of harmonic functions of several
variables, 1, Acta Math. 103 (1960), 25-62.
[33] G. C. Verchota, Layer potentials and boundary value problems for
Laplace's equation in Lipschitz domains, Thesis, University of
Minnesota, (1982), also, J. of Functional Analysis, 59(1984),
572-611.
Steven G. Krantz*
1. Three basic methods for obtaining integral formulas

We begin by discussing three ways to think about integral formulas on
domains in Cl, with a view to finding techniques which might generalize
to Cn. These are
(I) Exploit symmetry of the domain;
(II) Use differential forms and Stokes's theorem;
(III) Use functional analysis.
Discussion of f. Let A = (z EC : Jz j < 11. For n c Z, define On(Ceie) =

rlnlein9 Then direct calculation shows that
n(0) fnd0, all n

0
.
Now if f is harmonic on a neighborhood of A, then f has an L2

convergent Fourier expansion
00
f(reie) =n=`w an0n(reie) (1.2)
By linearity, (1.1) and (1.2) yield
Work supported in part by the National Science Foundation. The splendid

lecture notes prepared by Li Hui Ping, Li Xin Min and Ye Ke Ying greatly simpli-
fied the task of writing this paper.
185
186 STEVEN G. KRANTZ
f (0) = f(ei0)d0 . (1.3)

f21T
0
Of course formula (1.3) holds in particular for holomorphic f ; then we

rewrite (1.3) as
2n
fC(C0
f (O)
Tir-i f eie- 0 (iei0d= d (1.4)
o Y
where y(6) = ei0, 0 < 0 < 2rr.

Now (1.4) is the Cauchy integral formula on the disc for the point
z = 0. In order to obtain the general Cauchy formula, we exploit more
symmetry. Recall that if z r A is fixed then the function
0(0_oz(0=i +zC ,
called a Mbbius transformation, has the following properties:

(a) 0: A -. A is biholomorphic (i.e. holomorphic, one-to-one, and
onto, with a holomorphic inverse).
(b) 0(0) = z
(c) 0, 0-1 are smooth on a neighborhood of A.
If f is holomorphic on a neighborhood of A, define g(e) = f oO(e).
Then (1.4) applied to g yields
f(z) = g(0) = 21 g(e)de= 1

Y
f Y
00(e) de
Change variables by e _ 0-1(C) = Then
f( z) = 'ri f f(O
1 1d C
Y
171 _(o
INTEGRAL FORMULAS IN COMPLEX ANALYSIS 187
and logarithmic differentiation of 0-I gives
27i f 00 +
z dC (1.5)
1
f(z) = -z 1=z
y
= 1 fCQ
-Z
dC 27ri
+
fr
Now the numerator of the second integrand in (1.6) is a holomorphic
function of C which vanishes at 0. By (1.4), the second integral
vanishes. Formula (1.6) now becomes
f( z) _
2ni
fy
Cf(C)
-z dC
which is the Cauchy Integral Formula for the disc.
REMARK 1. If f is only assumed to be harmonic, then we cannot argue

that the second integral in (1.6) vanishes. Instead, a little algebra
applied to (1.5) gives
2n
y,
f(z) = Zn f(ei9) 110_ z2 2 d
le -z10
This is the Poisson Integral Formula.
REMARK 2. Among bounded domains in C, only the disc (and domains

biholomorphic to it) has a transitive group of biholomorphic self maps.
(This follows from the Uniformization Theorem; see also [321.) Thus
approach I has serious limitations in CI . In Cn the limitations are even
more severe. Indeed in Section 7, after we develop a lot of machinery, we
shall return to the concept of symmetry in Cn and gain some new insights.
Discussion of H. We need some notation. Recall that in real differential

analysis on R2 we use the basis ax , for the tangent space (i.e. all
ay
linear first order differential operators are linear combinations of these)
and dx , dy for the cotangent space. We have the pairings
<dx>=<$,dy>=1,
<dy>=<1,dx> =0.
In complex analysis, it is convenient to define differential operators
a -1 (ax
a a a _1 (a a .
, aj_2c+iay)
az 2 lay
The motivation for this notation is twofold. First,
a
3j7z z=1, z= -z=0
Secondly, if f( z) = u(z) + iv(z) is a C 1 function, with u and v real
valued, then
-fi and ay=-1,

//
(1.7)
which is the Cauchy-Riemann equations. Thus = 0 means that f is

ho lom orph ic.
We also define
dz = dx + idy, dz = dx - idy .
It is immediate that
<,dz> = 1 ,
<az, dz> _ < a,dz> = 0.

az
An arbitrary 1-form is written
u(z) = a(z)dz + b(z)dz (1.8)
and we define the exterior differentials
au = ab dz A dz , X = . dz A dz . (1.9)
Clearly du = au + au. Recall
STOKES' THEOREM. If ci CC Rn is a bounded domain with smooth

boundary and u is a smooth form on ci then
fu = fdu.
an sZ
In our new notation, if ci C CI ti R2 and u is a 1-form as in (1.8), (1.9),

then Stokes' Theorem becomes
fu= fau+u
ale 11
(1.10)
(t as j)dzA.
a
Now we can prove
THEOREM. If SZ C C is smoothly bounded and f is holomorphic on a

neighborhood of ci then
f(z)
2ni
f.!2dC, all zci .
= -z
aci
Proof. Fix z e it. Let E < distance (z, aft). Define D(z,E) = 1C E C : IC-zI < El
and it = it\D(z,E). We apply Stokes' Theorem to the 1-form u(C) _
f(t
dC on the domain it (note that u has smooth coefficients on a
-z ti
neighborhood of the closure of it, but not on all of it ). Thus, by (1.10),
fu(i= fdu=_j(2)dAd.
ti ti ti
ti
This last is 0 by (1.7). Thus, since aft = dl2 U 3D(z,e) (with suitable
orientations), we have
fu() = J u(O
aft 3D(z,E)
Parametrizing 3D(z,E) by C = z + Eeie, 0 < 0 < 21r, we obtain
21r
f-!d= J ('
f(z+Eei0)id9 2aif(z)
o
as E -. 0+. This completes the proof. 0
REMARK 1. In the proof of the theorem, if we assume that f is smooth

but not necessarily holomorphic, then the right side of (1.11) does not
vanish; instead it equals dC A dC. Thus the proof of the

theorem yields
f(/a A dJ. (1.12)

f(z)
2ai ff
C -z 2ni J -z
This formula, valid for all f c C'(KI), will be valuable later on.
REMARK 2. As it stands, the method of Stokes' theorem will not general-

ize to Cn. Namely, we used the fact that is holomorphic with an
isolated singular point at z . In Cn , n > 2 , holomorphic functions
never have isolated singularities. A more sophisticated approach will
therefore be needed.
Discussion of III. If Q C C is a smoothly bounded domain and if E > 0,

define S1E _ {z a iZ : dist(z, 00) > E 1. If E is sufficiently small, say
0 < E < E0, then 0E will also be smoothly bounded. Define
H2(fl) = f holomorphic on 0: sup

O<E<E
f If(C)I2ds(C) < 4 .
0
do E
(Here ds is the element of arc length.) That this definition is

unambiguous is a technical matter (see (31, Ch. 81). We shall need
PROPOSITION ([311). Each f e H2(fl) has associated to it a unique

f e L2((3Q). The Poisson integral of f is f. Also
sup
0
f If(()I2da( )I/2 I II f
L2 (f)
E
If ITE: 09 E i3f is normal projection then
(f I
au
E) (17 E)-I' f in L2 (do)
<f,g> = ffda for f,g a H2(SZ) . (1.13)
an
BASIC LEMMA. if K C Sl is compact then there is a constant C = C(K)

such that
sup If(z)I < C1101 H 2s all f E H2(0) . (1.14)

zEK
Proof. Fix z E K. Let Et < U (distance (K, 311) ). Then
1f(z)1 1
I
2m z dC
3D(z,el) I
(Stokes)
1
dC
2771
C -z
a1Z
el
< --L- ffIds

- 2nE1
aloe
1
(Schwartz)
1/+
< C(E1) S)l2ds(OI/2

If(y
J
< C(K) IIf1IH2 .
Now (1.13) and (1.14) imply that H2(fl) is a Hilbert space. Fix
z c fl and define the functional
Oz :H2(11), C
fF-f(z).
Then the lemma, with K = {z#, shows that z is continuous. The

Riesz Representation Theorem yields a unique kz f H2(il) such that
95z(f)=<f,kz>, all f eH2.

In other words,
f(z) ff())dsQ) , all f fH2(Sl), z eSl. (1.15)
ail
Formula (1.12) is called the Szego formula and kz(c) = S(z, C) the Szego
kernel (see [31, Ch. 11 for further details).
REMARK 1. Formula (1.15) has the advantage of working on any smoothly

bounded domain (even in Cn ), and the disadvantage of being non-
explicit. As an exercise, check that when Sl = A C C and z = 0 then
S(z,0 _= 21r. Then use Mobius transformations to calculate for
any z f A. You will rediscover the Cauchy formula!
2. The Cauchy -Fanta ppie Formula

Now we begin to consider integral formulas in Cn. For purposes of
differential analysis we introduce the notation
a 1 !a
=2 '
aa 1
2
(1)"ij
dzj = dxj +idyj , dzj = dxj - idyj, j
It is easily checked that
a_ zj = T =1, <-l, dzj> = <-, > = 1 , j = 1,...,n ,

j
and all other pairings are 0.

If a = are tuples of non-negative

integers (mufti-indices) then we write
dza = dza, A ... A dzak , di- 13 = dz16A ... A d

1 Q
A differential form is written
u=Iaa,6 dzaACZ /3 (2.1)

a,f3
with smooth coefficients a a/. (If 0 < p,q e Z and the sum in (2.1)
ranges over jai = p, 1131 = q only, then u is called a form of type
(p,q) .) We then define
au = t dzj n dza F. dz ru = a# dzj A dza Adz

c3i
a calculation (or functoriality), du = au + au. A C1 function f is

called holomorphic if Ju = 0. (Note: this means that df = 0,
(9zj
j so f is holomorphic in the one variable sense in each

variable separately.)
Finally, we introduce two special forms: if w = is an
n-tuple of smooth functions then we define the Leray form to be
77(w) = I (-1)l+lwj A dw1 A A dwj-1 A dwj+1 A - A dwn .

j=1
Likewise
m(w) =dw1 A... Adwn.

We define a constant
W(n)= J W(C)AW(C)
B(0,1)
Here B(z,r) = IC a Cn : IC -z I < r 1.

Now we may formulate a generalization of approach II in Section 1.
THEOREM (The Cauchy-Fantappi(; formula). Let Sl C Cn be a smoothly

bounded domain. Assume that w = cC(SZ x SZ\A),
wj and
n
I wj(z,4)'(Cj-zj)-1 on iZxSZ\A. (2.2)
jj=1
If f e Cl(1) is holomorphic on 0, then for any z e 11 we have
f(z)
_W (n)
f 1(i) i w) A W (C) . (2 . 3)
do
Before proving this result, we make some detailed remarks.
REMARK 1. In case n = 1, then w = wl = 1 (of necessity). So

4-z
rl(w)_---, 1 _2ni
1
C-z nW(n)
The Cauchy-Fantappie formula becomes
f(z)
m
f f(C) dC,
z
do
which is just Cauchy's formula.

REMARK 2. As soon as n > 2 , the condition (2.2) no longer uniquely

determines w. However an interesting example is given by
(wi(z,0,...,wn(z,0)
1-z1 .,
w(z,0 _ _ Cn-n .
I.
Let us calculate what the theorem says for this w in case n = 2. Now
r!(w) A c(C) _ (w 1dw2-w2dw 1) A dC1 Ad C2
w1
aa1
2dc1 + w I - dC2
aC2
1
- w2 w2 1 1A dC, A dC2
a 1 (X2
which by direct calculation
- 2z q dc-1 + IC _z q A dC1 A dC2

IC -z
Thus, by the theorem, we get a form of the Bochner-Martinelli formula:
4=z2)
f(z) = z(2) 5 f(O IC-zI
A dC1 Ad 2
for f c C 1(Sl), holomorphic on Q.

Now we turn to the proof of the Cauchy -Fa ntappie formula. For sim-
plicity, we restrict attention to n = 2 . Let
2
a=(a1,a2) cC00(SlxSZ\O): 1 aj(z,c) (cj-zj) =1 .
j=1
If al,a2 E `.f then we define
B(a',a2) =det(a',dal) _I aES2

E(a)ao(I) A
We claim that B has three key properties:

(1) B(a', al) does not depend on aI ;
(2) c3CB(a1, a2) = O ;
(3) If a I, al, RI, R2 E f then B(a1, al) - B((31, R2) is J exact.
Assuming (1), (2) and (3) for the moment, let us complete the proof (note
that (1) is used only to prove (3)). Letting
wI c
aI =a2 = EJ
(C2-z2)/K zI2
we have (observe that B(at,a2) = q(w) )
J f(())7(w)AW(C)= ff()B(ata2) A w(C) (2.5)
ag do
Note that, by (2),
dC(f(() B(at , a2) A W(C))
= a.f(C) A B(a1, a2) A Ev(C) + f(C) dCB(a1, a2) A W(C)
=0+0=0.
Hence, letting 0 < e < dist (z, SO), we have by Stokes' Theorem that (2.5)
= f(C)B(a1,a2) A 0)(C)
J
aB(z,E)
f f(C)IB(al,a2)-B(j61, 62)1 Atd(C) (2.6)

a6(z,E)
+ B(j6 1,62)AW(C).
JaB(z,E)
But the first integral
= J f(()dA A W(C)
aB(z,e)
(for some A, by (3))
= f d(f(() A A A W(C))
3B(z,e) (2.7)
=0,
by Stokes' Theorem, since a9B(z,E) = 0. Thus (2.5)-(2.7) give
f f f(C)B(j61,,62) A W(C)
Al 3B(z,E)
which, by (2.4),
= A W(C)
J I C-z l4
3B(z,E)
= 1 f (C) rl(- z) A W(C) .

E4 J
3B(z,E)
J f (z) rK -) A 0)(t) + ((E)

=13B(z,E)
4
(Stokes)
= 4f(z) 2((-z)Aw(C)+O(E)
64 J
B(z,E)
= 4f(z)E4 2(a(J)Au(C)+0(E)
J
B(0,1)
= f(z) 2W(2) + O(E)
Letting E - 0 yields the desired result. 0
We conclude this section by proving (1)-(3). For (1), we have
B(a1, a2) = det
1 det
(C1-z1)(C2-z2)
which, by adding row 2 to row 1,
1 det
(C1-zl)(C2-z2)
((C22)a2 aC ((C2-z2)a2)
_
(Cl z1) ? a2
This calculation is correct for C1 z 1 , C2 z2. The full result follows

by continuity.
For (2), imitate the proof of (1).
For (3), use (1) to write
B(a1, a2) - B((31, p2) = B(a1, a2-P2)

ai (ai-bi)
= det
a21 'j (a 2-b2
Now an easy calculation, as in (1), shows that this last equals aA where
A = det
a2 a2 2
This completes our discussion of the Cauchy-Fantappie formula.
3. Introduction to the a problem

One of the principal problems in complex function theory is the con-
struction of holomorphic functions with specified properties. In one
dimension, there are a number of highly developed techniques: Runge and
Mergelyan theorems, power series, infinite products, integral formulas,
and so on. In several variables, these techniques are either unavailable,
much less useful, or much less accessible.
The two most prevalent techniques for constructing functions in

several complex variables are sheaf theory and the inhomogeneous
Cauchy-Riemann equations. The latter interact strongly with the subject
of integral formulas, and in any case are a more flexible technique than
the former. To these we now turn our attention.
n
The setup for our study is that, for a given form f = ,SI f j(z) dz j , we
seek a function u such that au = f . Notice that since 0 = d2 = a2 +

+
;72, linear independence considerations yield that '2 = 0.
Hence au = f necessitates of = 0.
A simple calculation shows that, for n > 2 , this compatibility condi-
aft af k
all j, k. Notice, however, that when
tion is equivalent to
n=1
vkj = ,
the condition df = 0 is always vacuously satisfied. This differ-

ence can be explained in part by the fact that the equation au = f is
really n equations f one unknown (namely u ). For n > 1
1
the system is then "over-determined" and a compatibility condition is
necessary. For n = 1 the system is not over-determined.
The three basic considerations about a PDE are existence, uniqueness
and regularity. It is easy to check that a is elliptic on functions in the
interior of a given domain; hence, if u exists, it will be smooth whenever
f is (we will see this in a more elementary fashion later). So interior
regularity is not a problem. Also, since the kernel of 3 consists of all
holomorphic functions, uniqueness is out of the question. So, for us,
existence is the main issue.
The following example shows that the compatibility condition of = 0
does not by itself guarantee existence of u.
EXAMPLE. Let S1 C C2 be given by
11 = (B(0,4)\B(0,2)) U B ((2,0), 2/
Let U = B ((1,0), 2) and V = B ((1,0), 4) as shown. Let 0 e C (U)

satisfy --1 on V. Finally, let
Then f is smooth and j 'closed on Q since zl

1 is well-defined and
holomorphic on supp (drl) fl 11. If there existed a u satisfying Al = f

then the function h =_ u would be holomorphic (Jh=0) on
zt-1
cl\IB (1,0),41 n 1zt=11 . But u would necessarily be smooth near

(1,0) (since f is) hence h has a singularity at, for instance, (1,0).
Thus we have created a function holomorphic on B(0,4)\B(0,2) which
does not continue analytically to (1,0). This contradicts the Haitogs
extension phenomenon (an independent proof of this phenomenon will be
given momentarily). C1
Now that we know that du = f is not always solvable, let us turn to

an example where it is useful to be able to solve the a equation.
EXAMPLE. Consider the following question for an open domain it CC Cn :
If w = fl fl 1zn = OI' 0 and if g is holomorphic

(3.1) on co (in an obvious sense), can we find G
holomorphic on ft such that GIC0 = g ?
If 1 is the unit ball, then the trivial extension

g(zi,,zn_1,0) will suffice. However if 0 = B(0,2)\B(0,1) C C2 then
g(z1,0) = 1/z, is holomorphic on a) but could not have an extension G
(else the Hartogs extension phenomenon would be contradicted). o
THEOREM. Suppose that w C Cn is a connected open set such that

whenever f is a smooth a-closed (0.1) form on fZ then the equation
au = f has a smooth solution. Then the answer to (3.1) is "yes."
Proof. Let u: Cn - Cn be given by (z1,..,zn-1,0). Let

B = ;z f l : nz / col. Then B, w are disjoint relatively closed subsets
of fZ, so there is a C function 4 on fZ such that 0 = 1 on a rela-
tive neighborhood of co and = 0 on a relative neighborhood of B.
ti
Define F on fZ by
c(z) f(n(z)) if z e supp q
ti
F(z) _
0 else.
ti
Then F gives a C (but certainly not holomorphic) extension of f to Q.
We now seek a v such that F + v is holomorphic and F + vIw = f. With

this in mind, we take v of the form zn u and we want
d(F+v) = 0
or
=0.
Now f o n is holomorphic on supp (b and zn is holomorphic so all that

remains is
(fn) - zn' 3u=0
or
n)
zn (3.2)
The critical fact is that, by construction, 4 = 0 in a neighborhood of

n f1 {zn=0# so the right-hand side of (3.2) is smooth on Q. Also it is
easily checked to be 3 closed. Thus our hypothesis is satisfied and a
ti
u satisfying (3.2) exists. Therefore F - F +v has all the desired
properties.
Our two examples show that solving the 3 equation is (i) subtle and
(ii) useful. Thus we have ample motivation to prove our next result.
LEMMA. Let (he CS(C) , k > 1 , and define f = (z)dz. Then
u(z) =
2i dC A dC
J J ; -z
C
satisfies u e Ck(C) and 3u = f .
Proof. We have
0 "U
az
a (-L
)g-2ni
ff
C
f(a/)(+z) do A dC
2ni fC
w d6 (o d A dC
II
1
2ni
D (0, R)
where D(0,R) is a disc which contains supp 0. We apply Remark 1 of

II in Section 1 to 0 on D(0,R) to obtain that the last line
(C)dC
2ni J z
aD(0,R)
The integral vanishes since 0 = 0 on (91)(0,R), hence au = f . Observe

finally that u e Ck by differentiation under the integral sign. o
REMARK. In general, the u given by the lemma will not be compactly

supported. Indeed if f f 0 0 and if u were compactly supported (say
u C D(0,R) ) then a contradiction arises as follows:
0= J udC= J
aD(0,R) D(O,R)
The supports of solutions to the a problem explain many phenomena in

one and several complex variables. We explore this theme later. Mean-
while, contrast the Lemma and Remark with the following result.
THEOREM. Let n > 1 and let D(z) _ q5 1dz 1 +-.. + (kndz 1 be Xclosed
on Cn and suppose each (kj CC(Cn). Then for any 1 < j < n the
of
function
(kl(z dC A dC
uj(z) _ - 217i
1
ff
C
C_zj
satisfies uj a Ck(Cn) and aut = 40. Moreover uj = of for all j,
Proof. Fix 1 < m < n. We need to check that auj = (kj 1 < m < n. If
fim
m = j then the result follows from the lemma. If m ' j then use the
compatibility condition
d0j
=
-m to write
Sam azj
j (zl'...,zj-iX'zj+l,...,zn)
uj(z)
'3z_rn dC A dC
M- C-zj
C
4m
dCAdC.
2ni
ff C
By the lemma, this last equals <bm. Thus 9uj = 0. Notice that, if
f V j , then uj = 0 for zf large (since then (kj = 0 ). Also uj is
holomorphic for zQ large (since du = 40 is then 0 ). So, by analytic
continuation, u = 0 off a compact set. Next, uj - up - 0 since it is
compactly supported and holomorphic. Finally, uj e CCn) by differen-
tiation under the integral sign. 0
REMARK. The proof actually shows that u is zero on the unbounded

n
component of c(U supp (kj). It also shows that there is at most one
j=1
compactly supported solution to au = f . In the present case, one exists

and is given by an integral formula.
4. The Hartogs Extension phenomenon and more on the a problem

We have cited the Hartogs extension phenomenon in the examples of
Section 3. The reader will want to check that the proof of it that we now
give is independent of those examples.
THEOREM (The Hartogs Extension Phenomenon). Let fZ C Cn, n > 1 ,

be a bounded, connected open set. Let K C 9 be compact. Assume that
Q\K is connected. If f is holomorphic on Q\K then there is a holo-
morphic F on 9 such that F Ia\K = f .
Proof. Choose (k F C(S1) such that (k = 1 in a neighborhood of an

and (k = 0 in a neighborhood of K. Define
O(z) f(z) if z FQ\K

ti
F(z) =
0 if zcK.
ti
Then F is a C extension of f to 9, but it is not in general holomor-
ti
phic. We now seek v such that F + v is holomorphic on Q (and
ti
F +vjQ\K = f ). Thus we seek v satisfying
a(F +v) = 0
or
49 ((k f)+av=0
or
0'v = (-) f (4.1)
since f is holomorphic on supp 0. Now (k - I near M so (- f

is smooth and compactly supported in Q. The theorem of Section 3 now
guarantees that there exists a v satisfying (4.1). Moreover, the remark
following the theorem guarantees that v = 0 near aa. Thus F + v is
holomorphic and, near ail, F?+v = V _ f = f. By analytic continua-

tion, F + v = f on U and the result follows. o
Notice how the hypothesis n > 1 was used in the proof to control
supp v. The Hartogs extension phenomenon has several interesting
consequences:
(i) A holomorphie function f in Cn , n > 2, cannot have an isolated
singularity. If it did, say at P, then f would be holomorphic on
B(P,2E)\B(P,E) for a small hence, by the Hartogs phenomenon, on
B(P,2E), and hence at P. That is a contradiction
(ii) A holomorphic function f in Cn, n > 2 , cannot have an isolated
zero. If it did, say at P, then apply (i) to 1/f to obtain a
contradiction.
(iii) If U C Cn is open, E C U, f is holomorphic on U\E, and E is
a complex manifold of complex codimension at least 2, then f
continues analytically to all of U. To see this, notice that for
n = 2 the set E is discrete and the result follows from (i). For
n > 2 , the result follows from the case n = 2 by considering
f j(il\E)ne ranging over all two dimensional complex affine spaces
QC Cn.
Now we return to discussion of the d operator. There are essentially
four aspects to this matter:
(1) Existence of solutions;
(2) Support of d data and d solutions;
(3) Choosing a good solution, where "good" means smooth or
bounded;
(4) Estimates and regularity.
Regarding (4), we have noted that ellipticity considerations imply
that when av = g then v is smooth wherever g is. As an exercise, use
the theorem of Section 3 to give another proof of this assertion. (Hint: if
g is smooth on B(P,r), then let 0 E C (B,(P,r)) satisfy = 1 on
B (P, 2) . Let E Cc 1B 2satisfy c = 1 on B (P, 4/ . Define

`Àpply
u =95-v and f = d( v) = v/+c g. When Ju = f. the theorem
of Section 3, decomposing f as f = (V J0 v) + (1 i)( v)+ g.)

This is all that we shall say about interior regularity.
Topic (2) has been discussed vis a vis the Hartogs phenomenon.
Topics (1) and (3) are more subtle. First note that if au = f then also
a(u+h) = f for any holomorphic h. Given f , one cannot expect all
u + h to be nice (i.e. bounded, or L2 , or C up to the boundary). How
does one find a nice solution?
An idea from Hodge theory is to study the solution u to au = f which
is orthogonal to the kernel of a, i.e. which is orthogonal to holomorphic
functions. This solution has been studied by Kohn [25], [27], [28],
Catlin [3], [4], Greiner-Stein [16], and others. It is often called the Kohn
solution or canonical solution to the a equation.
We now briefly review some of what is known about the Kohn solution,
and other solutions, to the a problems on domains in Cn. (In this
section we shall take "pseudoconvex" and "strongly pseudoconvex" as
undefined terms. These terms will be discussed in detail in Section 5;
for now, a pseudoconvex domain is a domain of existence for the a
operator.)
(a) If SZ is a bounded pseudoconvex domain in Cn and f = ifi dzj is a

a-closed form with all fj e L2(f), then there is a u e L2(f) with
dAU = f . Also lull 2 < C(f) i 2
See [20]. (Exercise: the
L j 1L
canonical solution also satisfies this estimate.)
(b) If Q C Cn is smoothly bounded and pseudoconvex and f = ifi daj is

a a-closed (0,1) form with all fj a C(1) then there is a u e C(Q)
satisfying au = f. See [26]. It is not known whether the canonical
solution has this property.
(c) If Q C Cn is strongly pseudoconvex with C2 boundary and if

f = ifjda] is a a-closed (0,1) form with bounded coefficients, then
there is a u satisfying au = f and lull < C11lf]1IL.. The
L
canonical solution has this property. See [11], [17], [22], [16], [35].
Sibony [39] has shown that there are smooth pseudoconvex domains
on which uniform estimates for d do not hold. It is not known on
which parameters the uniform estimates depend (however see [13]).
Range [38], Henkin [17], and others have proved uniform estimates on
certain weakly pseudoconvex domains.
(d) Complete, and sharp, estimates have been computed on strongly

pseudoconvex domains in Lipschitz, Sobolev, Besov and other norms.
See [16], [30]. These estimates hold for the canonical solution. One
feature of the theory is that the operator assigning the canonical
solution Kf to a -closed (0,1) form f is compact in these norms.
This compactness is best exp-essed as a "subelliptic estimate" (see
[27]). Catlin [4] has announced a characterization of those domains
on which d satisfies a subelliptic estimate.
Many times an estimate tells us how to choose the right solution to

du = f. We conclude this section with an example of how estimates can be
useful.
DEFINITION. Let SZ C Cn be a domain and P r 30. A holomorphic

function f : St C is called singular at P if for every e > 0, f Inns(P.F)
is unbounded.
It is useful to be able to construct singular functions. Often we'can

nearly do this in the sense that we can find a neighborhood U of P and
a holomorphic function on u fl a which is singular at P (this is'called
a local singular function). Then the problem reduces to extending local
singular functions to global ones.
LEMMA. Let Sl C Cn be a domain on which the d operator satisfies

uniform estimates. If P E 30 and there exists a local singular function
at P which is bounded off any B(P,e) then there exists a global one.
Outline of proof. Let g : U no - C be a local singular function at P.

Let V be an open neighborhood of P such that V C U. Let k eC(U)
satisfy q = 1 near P and i e 0 off V . Set f = i g + u and solve a
a problem to find a bounded u. Then f is a global singular function at
P
Fix a strongly pseudoconvex domain Q. We shall prove later that (i)
uniform estimates for the a operator hold on ct and (ii) local singular
functions satisfying the hypotheses of the lemma exist for each P e Al.
By taking a suitable root of the local singular function and applying the
lemma, we may construct for each P e act a singular function Fp at P
which is in L2(1Z). Now we will prove that there is an L2 holomorphic
function F on ct that cannot be holomorphically continued past any
boundary point. This shows that ct is a domain of holomorphy and
essentially solves the Levi problem (see [311).
For the construction of F , let IPili I be a countable dense set in
act. Let Hij be the L2 holomorphic functions on ct U B (p11
j = 1,2,... . Let A2(1l) be the L2 holomorphic functions on ct. Con-
sider the restriction map yij : Hij A2(cl). Define Xij = image yij C A2(ct).
Because FP exists for each i , Xij A2(cl) for all i, j . We claim that
i
i.i '
U X A2(9). Assume the claim for now. Take F e A2(ct)\ U X...
This is the F we seek. The claim now follows from-
i.i Il
PROPOSITION. Let X and Y be Banach spaces, and T : X - Y a

continuous linear map. Then the following are equivalent:
(1) T(X) is not of first category in Y .
(2) T is an open mapping.
(3) T is onto.
Proof. This a variant of the Open Mapping Theorem for Banach spaces.
(I am grateful to R. Huff for this proposition.) o
5. Convexity and pseudoconvexity

Let Q C RN be an open set. Then fZ is called geometrically convex
if whenever P, Q E Q and 0 < t < 1 then (1-t) P + tQ ( fZ. In calculus,
however, a C2 function y = f(x) is called convex if f"> 0. How are
these ideas related?
If 11 has smooth boundary, then we may think of f1 as given by
0 =}x(RN:p(x)<01
for a smooth function p with V/p 0 on d11 (Exercise: use the

implicit function theorem). The function p is called a defining function
for f1. If P f aQ, let
Tp(au) = (al,...,aN) E RN :fax (p) 0}

)1)
J
Then Tp((At) is the (real) tangent space to c3 at P. If SZ = }p< 0}

we say that f1 is convex at P E all if
In
j, k=1
a2
dx dx
J J
(P) a] ak > 0 Aa E TP(dul) . (5.1)
We say that f1 is strongly convex at P if strict inequality obtains in

(5.1) when 0 a c TP(dfZ). A domain is convex (strongly convex) if each
boundary point is.
EXERCISES (see [31 ]):

(i) For a smoothly bounded domain, geometric convexity is equiva-
lent to convexity.
(ii) If 1 is smoothly bounded and convex, then iZ can be written
as an increasing union of strongly convex domains.
(iii) The above definitions are independent of the choice of p.
In order to understand the role of convexity in complex analysis, we

need to discuss inner products. If z, w (Cn , we define the Hermitian
inner product
n
<z,w>H = I zjwj
j=1
and if we identify
z = (zl,...,zn) = (x1+iy1,...,xn+iyn) ti (x1,y1,...,xn,yn)
(t 1,...,t2n)
and likewise
w = (wl,...,wn) ti (S,,---,s 2n)
then we define the real inner product
2n
<z,w >Re = I tjsj .
j=1
Notice the following facts:
(1) <z,w>Re =Re(<z,w>H) ;
(2) If Sl = (z a Cn : p(z) < 01 has smooth boundary, P e an, then

Tp(a1l) = Ia ( Cn : <a, ap(P)>Re = 01 .
(3) With Tl, P as in (2), let 5' p(9) = {a a Cn : <a, ap(P)>H = 01. We
call `.11(atZ) the complex tangent space to c%) at P. If
a e J'p(9) then is e `.'p(atZ). Also .Ì p(9S2) C Tp( ) and it is
the largest subspace of Tp(atl) which is closed under multiplica-
tion by i .
Now if Il = #z a Cn : p(z) < 01 is smooth and convex and P e at1 then

I lies on one side of Tp(ai2). Thus if we define fp(z) = <z-P, ap(P)>H
then the zero set Z(fp) of fp lies in P + Tp(af2). In particular
Z(fp) n S c ac (and if is strongly convex then Z(fp) n aft = 4PI).
Tl
Thus 1/fp is singular at P. Also Re fp < 0 on t2 so we may choose

0 < N e Z such that 1/(f p)1 IN is holomorphic on 0 and in L2(1l).
Thus each P e afl has an L2 singular function. By the argument at the

end of Section 4, fl supports an L2 holomorphic function which cannot
be analytically continued to any large open set. So any convex domain is
a domain of holomorphy.
If we want to understand domains of holomorphy, convexity will not
tell the whole story. For convexity is not a biholomorphic invariant: con-
sider 0: A C given by '(z) = (z+3)3. What is needed is a new notion
called pseudoconvexity:
DEFINITION. If fl =;z a Cn : p(z) < 01 is smoothly bounded we say that

fl is (Levi) pseudoconvex at P e afl if
a2p
0 Vw e 3'p(on) . (5.2)
!mar aZ.aZk
i, k=l 7
We call fl strongly pseudoconvex at P if strict inequality holds in (5.2)

for all 0 / w e fp(afl). The domain is pseudoconvex (strongly pseudo-
convex) if each boundary point is.
EXERCISE. If P e afl is strongly pseudoconvex prove that if A > 0 is

sufficiently large and p(z) _ (eAp(z)-1)/A then
(P)w]`k > ClwI2, VP E afl,Vw E p(afl) . (5.3)

I aZ az k
i,k=l 1
See [311 for details.
The rather technical notion of pseudoconvexity is vindicated by the

following deep theorem (see [311):
THEOREM. If U C Cn is smoothly bounded then the following are

equivalent:
(i) fl is pseudoconvex
(ii) fl is a domain of holomorphy
(iii) the equation au = f, f a 3 closed (p,q) form, is always
solvable.
This theorem means that pseudoconvex domains are the natural arena for
complex function theory. Also (exercise, or see [31]) any pseudoconvex
domain is the increasing union of smooth strongly pseudoconvex domains.
So strongly pseudoconvex domains, in a certain sense, are generic.
In order to unify and illustrate the ideas introduced so far in this
section we prove
LEMMA. If iZ C Cn is smoothly bounded and P e all is a point of

convexity then P is a point of pseudoconvexity.
Proof. Let p be a defining function for Q. Let a e J-p((3il). Writing

the definition of convexity in complex notation we have
n
d2
(P)aj ak
1,k=1 j k j,k=1 l
n 2
ap (P)aa
+12 , k=1
1
dzjdzk jk >0
But a similar inequality also obtains for is e fp(dil). Adding the two
inequalities yields the result. o
6. Solutions for the d problem

We briefly describe the Hilbert space setup for Hormander's L2 theory
of the d problem. We fix a smoothly bounded ft C Cn and introduce the
notation
o o =T L20 o
L2(fl) 1)(1) L(o 1)(ft)
11
H1 H2 H3
The operators T,S are of course unbounded, but they are densely defined
(since C' is dense in L2 ). It is easy to check that T, S are closed;

also S o T= 0 so if F= ker S then Range T C F.
An existence theorem for the a equation amounts to proving that
Range T = F . Moreover, it is an exercise in functional analysis to
check that this is equivalent to proving an inequality of the form
1IYIIH2 < CIIT*YIIH , dY c5) * n F . (6.1)

T
1
See [201 for details. Rather than prove (6.1), it is more convenient to
study the symmetric inequality
{IYIIH2 <C(IIT*YIIH1+IISYIIH3) dy c T* n1DS. (6.2)
Notice that when y E F (6.2) reduces to (6.1).

,
Unfortunately this program, as stated, fails. The difficulty is that the

computation of T* gives rise to boundary terms which, in general, cannot
be controlled. (However, in the strongly pseudoconvex case and on weakly
pseudoconvex domains satisfying a non-degeneracy condition, there are
delicate techniques for handling the boundary terms. See [251.)
Hormander's idea [201 was to work not in Euclidean L2 but rather in
L2 of the measure space a-Odx. If 0 is chosen to have certain con-
vexity properties and to blow up rapidly at ail, then the boundary is
effectively suppressed ( f becomes a complete Riemannian manifold)
and the Hilbert space program outlined above works. After the existence
problem is thus tamed, the weights can be eliminated (provided SZ is
bounded) and one obtains an existence theorem in L2(i2,dx). A leisurely
exposition of all these ideas can be found in [311. We now formulate a
version of HSrmander's result, which we will use freely in what follow:
THEOREM (Hormander). If SZ C Cn is smoothly bounded and pseudo-

convex, and if f is a a closed (0,1) from on SZ with L2(Sl,dx) coeffi-
cients, then there exists u E L2(fl,dx) such that i9u = f.
REMARK. Hormander's theorem can be used to prove most of the theorems

at the end of Section 5 characterizing domains of holomorphy. See [311 for
details.
Our next main goal is to obtain an integral formula for a solution to

the 0 problem on a convex domain. We first need:
THEOREM (Bochner-Martinelli). Let fl be a domain in Cn with CI

boundary. Let f c C I (1) . Then for all z c fl we have
f(z) = 1 P077
A W(C)
nW(n)
an
- ff()A7J
Q
Proof. Apply Stokes' theorem to the form
-z
(0 = f(i)n A W(C)
on the domain fl\B(z,e). Imitate the proof of the Cauchy Integral Formula
(or see [311).
Now fix 1 a bounded, C2, convex domain. Choose e > 0 so small
that 1 == l z c Cn : dist (z, 11) < e I is convex (hence pseudoconvex). Let f
be a smooth, 0-closed (0,1) form on fl. By Hormander's theorem, there
is a smooth u on fl such that du = f. We apply the Bochner-Martinelli
formula to u (which is certainly in C(fl)
I ). Thus
Z2
u(z) = fu(017
WW(n)
au
nW(n) f An
-z{
z2
A)(t) (6.3)
fu()1? ff(4)A rl Z-a2 A.W(C).

nW(n) z2 Ate(()-n 1
The first term on the right is not useful, since it involves u , so we will
remedy matters by subtracting an appropriate holomorphic function from
the right side of (6.3) (see the discussion in Section 4 on choosing a good
solution). The Cauchy-Fantappie formalism now comes into play:
If 0 =(p<01, let
ap -ap
Xn
(D(z, 0 4)(z,
with 4(z, I d
(C) (zj -Cj) and observe that
j=1 Oxj
H(z) = -1 fu(xw(z)) A
is well defined and holomorphic in z (because w is).

Now let v(z) = u(z) - h(z). Then certainly 3v = du = f and, by (6.3),
v(z) = nW(n)
1 {fu n
/ A - J uw-z
12
asp ale
nW(n) f
Q
A ri
z_z
=z2
Aw(C)-I-II.
z-z
Let G = SZ x [0,1] and define g(z, C, A) = (1-A) + Xw(z,0 to be a
K z I2
form on G. Then
I= n
W(u) full(s) A 0)(C)
aG
By Stokes' theorem, applied on G, this
1
nW n)
f d(u(C)rl(b) A
G
But
d(u(C)rl(g) A au A ?7(g) A W(C)
+ u(C)d(n(g)) A W(C)
= f A-J(g)AW(C)
(this last equality takes advantage of the special algebraic properties of

Cauchy-Fantappie forms). So we finally obtain
HENKIN'S INTEGRAL FORMULA. 11 f = Ip < 01 is C2 and convex and

f is a smooth, a-closed (0,1) form on a neighborhood of 12 then the
function
v(z) - n . W(n)
5 f(c) A n(g) A W(C)
-J A 17
-z
a
satisfies dv = f on 11. Here
g(z, , X) _ (1-A) - z` 2+ Aw(z, 0 ,

VIP
(D(z,
and
(P(z ( )(z- ) .
The standard reference for Henkin's work is [181; see also [311 Similar
formulas were derived by Grauert-Lieb [111, Kerzman [221, and (6vrelid [35].
Now we assume that i1 is strongly convex, and show how to use
Henkin's formula to obtain uniform estimates for solutions to the
problem. For simplicity we work in C2 only. So
a
f ,(g) Aa )f ^ [g1 dg2 dA-g2 dA AdC1
After some algebra, and integrating out a , the Henkin formula is

1
a 1 (C)(r,2-z2) +
2
(C)(1-z1)
v(z) = 2W(2)J
4'(z, S)I4_zI2
all
A (f +f2( ) n
1 f1(b) (_I= 1) +
2W(2)
I K z I4
d1S n d Cz n dCl ^ d s2
1
2W(2)
f ndC1 AdC 2
as
+ 2W(2) ( f2(C)A2(z, n dC1 ^ d62
ag
+ 2W(2) J f(C)B1(z, n dC1 ^ dC2

a
+ 2W(2) f2(C)B2(z,C)dz1 n dZ2 n dC1 ^ dC2

J
Q
In order to prove an inequality of the form
10 L,o < CIif11Lw (6.4)
it suffices to check that

flAj(z)idcr()<C, j =1,2 (6.5)
190
and
J IBj(z,C)jdV(C) <C, j=1,2 (6.6)
with the estimates uniform over z e 11. (Here d or is area measure on .)

By symmetry we check only j = 1 .
For B1 , choose R > 0 such that B(z, R) D t1 for all z e Q. Then
f IB1(z,C)IdV(() < r C dV(C)
Sl B(z,R)
<C f r3dr=CR<- .
r3
0
For A 1 , we must work harder. We need to know something about the

degree to which the denominator 4D(z, C) IC-z i2 of Al vanishes. (This
is where strong convexity plays a role.) Think of z r St as fixed. Notice
that, writing the Taylor expansion for p about t in complex coordinates,
we have
dp ap
p(z)=p(()+2Re (C)(z2-C2)
+ (quadratic terms) + (error terms)
> 0 + 2Re (D(z, C) + C lz- l2 , C>0.
The last estimate is by strong convexity - see Section S. Thus

IRe (D(z,S)I > CIIz-CI2 + Ip(z)II
Let nz be the normal projection of z to Al. Let t1 be a coordi-

nate in the complex normal direction at nz (that is, i times the real
normal direction) in ail. Recall that
2
ap
(C) a j= 0 a c `.`P(R)
and
2
dp
Re 5 (C)aj = 0 a ETp(SZ).
j=1
It follows that
2 2
Im I (Nzj-cj) = Im I z)(zj-Cj) + 00Z-0)

=1 7 j=1 J
measures (essentially) the complex normal component of z - C at 1Z.

As a result
IImtI >C1t1I.
Introducing two additional coordinates t2 , t3 centered at uz , which

span the complex tangential directions at nz , we conclude that
I(D(z,C)I >2 (IRe w(z,C)I + IIm(z,C)I)
>C 1t1+t2+t3+Ip(z)I1
provided C is near nz, say K-rrzI <r0. Then
fAi(z)da(i)= J + J .
an B(nz,r0)lao as\B(rrz,r0)
The second integral is trivially bounded since when iz-Cl > r0 then Al
is bounded. The first is majorized by
C ,/t2 +t2 +t3 +p(z)2 dt ldt2dt3

J (t2I +t22 +t23 +p(z)2) (it,1 I + t 22 + t 23 + Ipz)I)
ti+t2+t3<Cr0
=C
IpI+ItlI <t2+t3 IpI+It11>t2+t3
tl+t2+t3<CrO ti+t2+t3<Cr0
Y1+Y2
Now
1
IY11<C dttdt2dt3
I (t2 +t3)
IPI+I t1 I<t2+t3
t1 +t2+t3<C r0
< C J 1 dt2dt3
t22 + t23
t2+t3<Cr0
r0
<C J
Likewise
dtldt2dt3
IY21 < C
t22 +t23 (it,1 I + IPI)
IPI+It1I>t2+t3
ti+t2+t3<Cr0
<C
I
t2+t3<C r 0
t2+t
1
2
3
(Ilog(t2+t2)I+llog(Crdl)dt2dt3
2 3
C r0
<C
f
0
(Ilog rl + log CI)rdr
<C-<00.
Thus
f JAIIda<C <-o .
an
The estimates for A2 , B2 follow by symmetry, as already noted. We

have proved
THEOREM. With fZ strongly convex and f, v as in Henkin's integral

formula for solutions of the a problem,
IV II < C Y-IIfjIl L, .
L(tt)
REMARKS. 1) To eliminate the hypothesis that f is defined on a neigh-

borhood of ft we observe that for E > 0 small enough the domain
06 = fz Ef : dist (Z, 3f) > e y

is smoothly bounded and strongly convex. Moreover, the estimates in the

theorem on 1lE depend boundedly on E (by a calculation). Thus esti-
mates can be obtained for f a smooth form on 11 with bounded coeffi-
cients by applying a limiting argument to the solution uE of d(*) = f on
QE.
2) The results of the theorem actually hold on smoothly bounded strongly
pseudoconvex domains. This is most easily seen by using the following
important result:
THE FORNAESS IMBEDDING THEOREM [9]. Let it CC Cn be a

strongly pseudoconvex domain with C2 boundary. Then there is a neigh-
borhood SZ of _Q, a k > 0, a C2 strongly convex domain U C Cn+k
and a holomorphic imbedding F : c Cn+k such that
(i) F(12) C U
(ii) F(6\5) C Cu
(iii) F(4) C au
(iv) image F is transversal to X.
The upshot of this theorem is that the Henkin singular function (D which
we know how to construct on U can be pulled back to Q. The construc-
tion of the Henkin solution to the a equation and the uniform estimates
follow just as before.
3) The singular function c can be constructed more directly on a

strongly pseudoconvex domain fl as follows: first write it = { p < Of
where
(C)wjwk>CIwl2 dw (Cn,
J k
(see (5.3)). For C E 9Sl fixed we define
2, y
L(z, ) _Cj) + aP
J J k k
i=t J J k-t
The function L, called the Levi polynomial, has the property that there
a neighborhood Uc of C such that sz n UC n #z : L(z, t) = 0} _;c } . One
can modify L, by solving a a problem (see [181 or [311) to obtain 4i
n
such that 5 n {z : d>(z, c) = 01 = I I and d>(z, C) =I Pi (z, 0 (z)-Cj)
j=1
with Pi holomorphic in z . Henkin's program may be carried out using
this and w(z, C) = (-PI(z, 011b(z, 0, ..., -Pn(z, 0/(D(z, 0)
Notice that, by the discussion in the preceding paragraph, 1/L( , C)
is a local singular function at C. This, together with the uniform esti-
mates for the a equation which we have obtained, completes the program
outlined in Section 4 to show that a strongly pseudoconvex domain is a
domain of holomorphy.
7. Connections between various integral formulas and applications

In Section 1 we constructed the Szego kernel for domains in CI . How-
ever the construction goes through for domains in Cn once one has the
basic lemma, and that follows in Cn from the Bochner-Martinelli formula.
We leave the details of the basic Szego theory in Cn as an exercise.
Recall that the Szego kernel for a domain 1Z is the reproducing kernel
for H2([Z) . Now fix z e Q. By construction, S(z, ) a H2(fl) . By the
reproducing property, it follows that
S(z,4) = fS(w)S(z.w)da(w)
aQ
= f
ag
= S(C,z )
=S(C,z)
Thus the operator
S:f i+
has the following three properties:

(a) S : L2(dQ) - H2(fl)
(b) S is self-adjoint
(c) S is idempotent.
Therefore S is the Hilbert space projection of L2(8l1) onto H2(1l).
It turns out that the Henkin operator on a strongly pseudoconvex
domain very nearly has properties (a)-(c). First, by a theory of non-
isotropic singular integrals developed especially for boundaries of
strongly pseudoconvex domains (see [81, [361), the Henkin operator
H:fi.nWn) f(4)r!(w)A
J
w produced from the Fornaess theorem as in Section 6) maps

L2(afl) onto H2(C1). Also H is idempotent. Now H is not quite self-
adjoint, but it is nearly so. To see this, one needs to write (7.1) in the
form
N(z,
1
nW(n)
f pn(z,
4)
da
where da is area measure on d fl. This is a straightforward but tedious

calculation (see [231). It turns out that N is real. It also turns out that
(D(z, ) - I(4,z) vanishes to higher order at z = 4 than does (P. (Try
this when Il is the ball to see how this works - details are in [231.)
As a result of the preceding observations, the kernel
N(z, C) N((,z)
n(z, 0
is a kernel which is less singular than the original Henkin kernel. Thus
H - H* , rather than being a non-isotropic singular integral operator (as is
H ), is a smoothing operator. This observation of Kerzman and Stein is
now exploited as follows. Denote H* -H = A .
The reproducing properties of S and H guarantee that
(1) S = HS
and
(2) H=SH.
Thus
(3) S = S* = (HS)* - S*H* = SH*
Subtracting (2) from (3) gives
S -H - S(H* - H) = SA
This is an operator equation on L2. We may resubstitute the equation

into itself as follows:
S = H + SA
= H + (H+SA)A
=H+HA+SA2
(7.2)
= H + HA + (H+SA)A2
-H+HA +HA2 + SA3
.. = H + HA + HAk + SAk+1
Now we know that each of the operators HA, HA2, -- are smoothing. If
we apply both sides of (7.2) to a sequence Oj E Cc '(4) such that ) -, S
in the weak`* topology on 91, we obtain an equation relating S(z, C)

and H(z, C). In particular, H and S are equal modulo terms which are
less singular. From this fundamental result, many basic mapping proper-
ties of S can be determined (see [36]).
The basic construction of Kerzman and Stein can be used in other
contexts. Let us turn now to one of these: the Bergman kernel. Fix a
domain SZ CC (:n and define
A242) = f holomorphic on SZ: J If(z)I2dVol(fl) <
(Notice that, for SZ smoothly bounded, H2(fl) is a proper subspace of

A2(fl) - Exercise.) Define
<f,g>= ffdv
ft
[fil = f If 12 dV' /2
Q
for f,g a A2(fl). The basic lemma in this context,
sup If(z)I 5 CKIIf II

K
for K CC [Z, is easily derived from the mean value property for holomor-
phic functions. As in Section 1, the abstract Hilbert space theory yields
a reproducing kernel for A2 which we call the Bergman kernel.
Just like the Szego kernel, the Berman kernel (denoted by the letter K)
satisfies K(z, C) = K(C,z). Thus the associated operator
B:f H ff()K(zs)dV()
12
is self-adjoint. It maps onto A2 (by construction) and is idempotent. So

B : L2(1l) - A2(1l) is Hilbert space projection.
A remarkable construction of S. Bergman (see [311) is as follows:
note that, for z E it fixed,
K(z,z) = J K(z,w)K(w,z)dV(w)
12
dV(w) > 0 .
I K(z,w)
Therefore we may set
2
gij(z) _ log K(z,z)
l 0-Til
By a calculation (see [31]), the matrix (glj(z)) gives a non-degenerate

Kfihler metric on iZ (called the Bergman metric) which is invariant under
biholomorphic mappings. In particular it holds that if ( : flI - fl2 is
biholomorphic then
distBerg(Z,w) = distBerg(c(z), NW))
As a result, metric geodesics and curvature are preserved.

The Bergman metric and kernel are potentially powerful tools in func-
tion theory, provided we can calculate them. To do so, we exploit the
idea of Kerzman and Stein [231 to compare K with the Henkin kernel.
However a complication arises: the Henkin integral (7.1) is a boundary
integral while the Bergman integral is a solid integral. How can we com-
pare functions with different domains? What we would like to do is apply
Stokes' theorem to the Henkin integral and turn it into an integral over Q.
However, for z c fI fixed, Henkin's kernel has a singularity at = Z. So
Stokes' theorem does not apply.
The remedy to this situation is to use an idea developed in [19], [30],
[331: for each fixed z c fl, let
)
N(z,i
(Dn(z,
S) Easl
Now construct a smooth extension Tz of 1Az to 9. The Cauchy-

Fantappie formula is still valid with Tz replacing qz (since the integral
takes place on the boundary where Tz = Oz ). Thus Stokes' theorem can
be applied to the new Henkin formula containing Tz. The resulting solid
integral operator on L2(il) can be compared with the Bergman integral
via the program of Kerzman and Stein (details are in [331). The result is
that
K(z, C) = 9'z(C) + (terms which are less singular) .
As a result, curvature, geodesics, etc. of the Bergman metric may be

calculated. Also the dependence of these invariants on deformations of
aQ can be determined (see [12], [13]; it should be noted that the methods
of [1] or [6] may be used for the deformation study instead of the Kerzman-
Stein technique). The following are the three principal consequences of
these calculations for a smoothly bounded strongly pseudoconvex it :
(a) As Q 3,z - a11, the Bergman metric curvature tensor at z converges

to the constant Bergman metric curvature tensor of the unit ball. The con-
vergence is uniform over M.
(,B) The kernel and the curvature vary smoothly with smooth perturbations
of M.
(y) fl, equipped with the Bergman metric, is a complete Riemannian
manifold.
Now we conclude this paper by coming full circle and discussing

once again the topic of symmetry of domains. The reader should consider
that, up to now, all of our effort has been directed at obtaining (a), (,3), (y).
Now we use those to derive concrete information about symmetries. If
Q C Cn is a domain, let Aut a denote the group of biholomorphic self-
mappings. If two domains f11 and a2 are biholomorphic we will write
a1 1 Q2'
THEOREM (Bun Wong [41]). If a CC Cn is smoothly bounded and
strongly pseudoconvex and if Aut Q acts transitively on fl, then
Q ti ball.
Proof (Klembeck). Let P0 a fl be any fixed point. Let IPi I C 0 satisfy

Pj a g. By hypothesis, choose (k] a Aut a such that (k](P0)=pi . Then
the holomorphic sectional curvature tensor K for the Bergman metric
satisfies
K(P0) = K(cbj(Po)) = K(PH) - (constant curvature tensor of the ball). (*)
Thus the Bergman metric curvature tensor is constant on Q. We now use
THEOREM (Lu Qi-Keng [34]). If M is a complete connected Kahler

manifold with the constant holomorphic sectional curvature of the ball
then M ball.
This theorem, together with (*), completes the proof. o
THEOREM (Greene-Krantz [13]). If f, C Cn is smoothly bounded and if

is C sufficiently close to the unit ball B then either
(i) 0tiB
or
(ii) Q B and Aut a is compact and has a fixed point.

Proof. Step 1. If (1 B then Aut 1Z is compact. For if Aut 0 is not

compact then a normal families argument [12] implies that for P0 E 1 3
Oj E Aut f such that iij(P0) 3(1. As in the proof of the preceding
theorem, it follows that Q ti ball.
Step 2. Recall the following result of Cartan-Hadamard (see [241):
THEOREM. If M is a complete Riemannian manifold of non-positive

curvature and if K is a compact group of isometries on M then K has
a fixed point.
Now we prove the theorem by denying (i) and proving (ii).
Step 3. By a calculation, the ball B has negative (bounded from zero)

Bergman metric curvature. But the stability result (B) implies that this
statement holds for domains t which are C sufficiently close to B.
By Step 1, Aut iZ is compact. So the result follows from (y) and Step 2.
Now we turn to a conjecture of Lu Qi-Keng (see [34]):
CONJECTURE. If 0 C Cn is a bounded domain then the Bergman kernel

never vanishes on a x 12.
On the disc and the ball this conjecture is correct; for the ball in Cn one
can calculate (see [31]) that
K(z, ) = nn! 1
nn (1-z. )n+1
However it turns out that in C1 the conjecture is true if and only if Q

is simply connected (see [40]). From this it follows that the conjecture
is not always true in Cn either. To see this, let C2 ) ci = disc x
annulus. Then the uniqueness of the Bergman kernel easily implies that
the kernel for 1Z is the product of those for the disc and annulus. Now
0 = UGj where fj C fj+1 and each 0i is smooth and strictly pseudo-
convex (see (311). By a theorem of Ramadanov [37], Koj - K11 normally.
By Hurwitz's theorem [31], KQi vanishes for j large enough. So there

exist smooth strictly pseudoconvex domains with vanishing Bergman
kernels. Thus we have the
MODIFIED Lu QI-KENG CONJECTURE. If fZ CC Cn is smoothly

bounded and diffeomorphic to the ball, then K0 never vanishes.
Some results about the modified conjecture may now be formulated. Let
us agree to topologize the collection of all smoothly bounded strictly
pseudoconvex domains by equipping their defining functions with the C
topology. Then we have (see [12], [13]):
(i) If = In: Kfl never vanishes l then a is closed.

(ii) If 'U = In: Kfl is bounded from 01 then Ù is open.
Statement (i) follows from Hurwitz's theorem. Statement (ii) follows

from (f3). Facts (i) and (ii), together with the fact that e, 11 are non-
empty (since both contain the ball), nearly provide a connectedness argu-
ment to verify the modified Lu Qi-Keng conjecture. The conjecture was
recently resolved in the negative by Boas and by Catlin.
Now we turn to a semi-continuity result for automorphism groups:
THEOREM (Greene-Krantz [14]). If noCC Cn is a smoothly bounded

strongly pseudoconvex domain and if fZ is a sufficiently small smooth
perturbation of no then
(i) Aut iZ C_ Aut 00

subgroup
(ii) 34): 0 -. no a diffeomorphisrn such that

Aut 0 3 a H 4 o a -q5_1 E Aut Q0
is a univalent group homomorphism.
Sketch of proof. We may as well suppose that no ball, else the result
is straightforward. Then normal families arguments show that, for it
sufficiently near S10, [1 ball. Thus Aut 0 is compact and, by
averaging the Euclidean metric, one can construct a new metric y, smooth
across aU, which is invariant under Aut Q. By patching this metric

with the Bergman metric, and modifying it near o5n, we can arrange that
Isom(y) = <Aut SZ, Aut Q> and that y is a product metric near ffi.
Finally, we construct the metric double M of 0 equipped with y.
Known theorems [5] about semi-continuity of isometry groups of deforma-
tions of a compact Riemannian manifold now give the result.
We introduce our final result by recalling a corollary of the Uniformiza-

tion Theorem (see [2]):
THEOREM. If SZ CC C, P e 0, and the isotropy group IP of Aut 0 is

infinite, then Q Z A,
The generalization of this result to Cn would require new ideas since,

in that context, there is no uniformization theorem. Also, on dimensional
grounds, the infinitude of IP is an insufficient hypothesis when n > 1 .
Instead we have
THEOREM (Greene-Krantz [15]). If M is any n dimensional, connected,

non-compact complex manifold, and if IP has a compact subgroup H
which acts transitively on real tangent directions at P, then M is
biholomorphic to either the ball or Cn.
Idea of proof. First create an H-invariant metric on M by averaging over

.H. By a continuity argument, we show that metric balls B(P,r) centered
at P are biholomorphic to the unit ball in Cn. This last is the heart of
the argument: It involves analysis of geodesics and equivariance proper-
ties of the Bergman metric on B(P,r) and of the canonical solution to
the 5 equation. Since (by inspection of the proof), the biholomorphisms
of B(P,r) to B vary continuously with r, and since the biholomorphisms
match up as r increases, the conclusion follows.
Let us conclude by briefly reviewing the course we have come. We

began by exploiting the many symmetries of the disc to derive an integral
reproducing formula on the disc. Since generic domains possess few

symmetries, we developed two alternate techniques to find integral
formulas - via Stokes' theorem and via Hilbert space theory. The latter
method has the advantage of being canonical while the former is explicit.
We used explicit integral formulas in several complex variables to
establish a number of basic results in the theory. Then, using an idea of
Kerzman-Stein, we were able to relate the explicit formulas to the
canonical ones. Finally, we used this connection between explicit and
canonical formulas to return to the question of symmetries of domains. We
established results which explain how the automorphism group of a domain
iZ depends on the geometry of ail.
There are still many open problems in the study of automorphism groups
of domains. One of the most compelling is to decide which domains have
non-compact automorphism groups. Another is to relate the dimension of
Aut (St) as a Lie group to the rank of the Levi form on ail. I hope that
the survey presented here will inspire some new people to consider these
questions.
THE PENNSYLVANIA STATE UNIVERSITY
UNIVERSITY PARK, PA. 16802
BIBLIOGRAPHY
[11 L. Boutet de Monvel and J. Sjostrand, Sur la Singularite des noyaux
de Bergman et Szego, Soc. Mat. de France Asterisque 34-35 (1976),
123-164.
[2] R. Burckel, An Introduction to Classical Complex Analysis,
Birkhhuser, Basel, 1979.
[3] D. Catlin, Necessary conditions for subellipticity of the a-Neumann
problem, Ann. of Math. (2)117(1983), 147-172.
[4] , Boundary invariants of pseudoconvex domains, to appear.
[5] D. Ebin, The Manifold of Riemannian metrics, Proc. Symp. in Pure

Math., Vol XV (Global Analysis), AMS (1970), 11-40.
[6] C. Fefferman, The Bergman kernel and biholomorphic mappings of
pseudoconvex domains, Invent. Math. 26(1974), 1-65.
[7] G. Folland and J. J. Kohn, The Neumann Problem for the Cauchy-
Riemann Complex, Princeton Univ. Press, Princeton, 1972.
[8] G. Folland and E. M. Stein, Estimates for the ab complex and
analysis of the Heisenberg group, Comm. Pure Appl. Math. 27(1974),
429-522.
[9] J. Fornaess, Strictly pseudoconvex domains in convex domains, Am.
Jour. Math. 98 (1976), 529-569.
[10] B.A. Fuks, Introduction to the Theory of Analytic Functions of
Several Complex Variables, Translations of Mathematical Monographs,
American Mathematical Society, Providence, 1963.
[11] H. Grauert and I. Lieb, Das Ramirezche Integral and die Gleichung
of = a im Bereich der Beschrankten Formen, Rice Univ. Studies
56 (1970), 29-50.
[12] R. E. Greene and S.G. Krantz, Stability of the Bergman kernel and
curvature properties of bounded domains, in Recent Developments in
Several Complex Variables, J. E. Fornaess, ed., Princeton Univ.
Press, Princeton, 1981.
[13] . , Deformations of complex structures, estimates for the
d equation, and stability of the Bergman kernel, Adv. Math.
43 (1982), 1-86.
[14] , The automorphism groups of strongly pseudoconvex
domains, Math. Ann., 261 (1982), 425-446.
[15] -, Characterization of complex manifolds by the isotropy
subgroups of their automorphism groups, preprint.
[16] P. Greiner and E. M. Stein, Estimates for the a-Neumann Problem,
Princeton Univ. Press, 1977.
[17] G.M. Henkin, A uniform estimate for the solution of the 3-problem
on a Weil region, Uspekhi Math. Nauk. 26 (1971), 221-212 (Russ.).
[18] , Integral representation of functions holomorphic in
strictly pseudoconvex domains and some applications, Mat. Sb.
78 (120) (1969), 611-632; Math. U.S.S.R. Sbornik 7 (1969), 597-616.
[19] G. M. Henkin and A. Romanov, Exact Holder estimates of solutions
of the d equation, Izvestija Akad. SSSR; Ser. Mat. (1971), 1171-1183,
Math. U.S.S.R. Sb. 5(1971),1180-1192.
[20] L. Hormander, L2 estimates and existence theorems for the a
operator, Acta Math. 113 (1965), 89-152.
[21] T. Iwinski and M. Skwarczynski, The convergence of Bergman func-
tions for a decreasing sequence of domains, in Approximation Theory,
Reidel, Boston, 1972.
[22] N. Kerzman, Holder and LP estimates for solutions of c3u = f on

strongly pseudoconvex domains, Comm. Pure Appl. Math. XXIV (1971),
301-380.
[23] N. Kerzman and E.M. Stein, The Szego kernel in terms of Cauchy-
Fantappie kernels, Duke Math. J. 45 (1978), 85-93.
[24] S. Kobayashi and K. Nomizu, Foundations of Differential Geometry,
Vol. 1, II, Interscience, New York, 1963, 1969.
[25] J. J. Kohn, Harmonic integrals on strongly pseudoconvex manifolds I,
Ann. Math. 78(1963), 112-148; II. ibid 79(1964), 450-472.
[26] Global regularity for c3 on weakly pseudoconvex mani-
folds, Trans. Am. Math. Soc., 181 (1973), 273-292.
[27] , Sufficient conditions for subellipticity on weakly pseudo-
convex domains, Proc. Nat. Acad. Sci. (USA) 74 (1977), 2214-2216.
[28] , Subellipticity of the 3-Neumann problem on pseudo-convex
domains: sufficient conditions, Acta Math. 142(1979), 79-122.
[29] , Methods of partial differential equations in complex
analysis, Proc. Symp. Pure Math. 30, Part 2 (1977), 215-237.
[30] S. Krantz, Optimal Lipschitz and LP estimates for the equation
Au = f on strongly pseudoconvex domains, Math. Ann. 219 (1976),
233-260.
[31] , Function Theory of Several Complex Variables, John
Wiley and Sons, New York, 1982.
[32] , Characterization of smooth domains in C by their
biholomorphic self-maps, Am. .Math. Monthly (1983), 555-557.
[33] E. Ligocka, The Holder continuity of the Bergman projection and
proper holomorphic mappings, preprint.
[34] Lu Qi-Keng, On Kahler manifolds with constant curvature, Acta Math.
Sinica 16(1966), 269-281 [Chinese]; Chinese J. Math. 9(1966),
283-298.
[35] N. Ovrelid, Integral representation formulas and LP estimates for
the 3 equation, Math. Scand. 29(1971), 137-160.
[36] D.H. Phong and E.M. Stein, Estimates for the Bergman and Szego
projections on strongly pseudoconvex domains, Duke Math. ]our.
44 (1977).
[37] I. Ramadanov, Sur une propriete de la fonction de Bergman,
C. R. Acad. Bulgare des Sci. 20(1967), 759-762.
[38] R. M. Range, The Caratheodory metric and holomorphic maps on a
class of weakly pseudoconvex domains, Pac. Jour. Math. 78 (1978),
173-189.
[39] N. Sibony, Un exemple de domain pseudoconvexe regulier ou

1'equation u = f n'admet pas de solution bornee pur f bournee,
Invent. Math. 62 (1980), 235-242.
[40] N. Suita and A. Yamada, On the Lu Qi-Keng conjecture, Proc. Am.
Math. Soc. 59 (1976), 222-224.
[41] B. Wong, Characterizations of the ball in Cn by its automorphism
group, Invent. Math. 41 (1977), 253-257.
VECTOR FIELDS AND NONISOTROPIC METRICS
Alexander Nagel*
The main object of this paper is to show how nonisotropic metrics

constructed from vector fields play an important role in certain recent
developments in partial differential equations and several complex
variables. As we shall see, these metrics are useful in describing
boundary behavior of holomorphic functions in pseudoconvex domains, in
estimating the kernel of the Szego projection in some of these domains,
and in estimating the size of approximate fundamental solutions to certain
nonelliptic, hypoelliptic partial differential operators.
The exposition is divided into three parts. In the first, we set the
stage by recalling certain classical theorems which are models for and
which motivate the more recent results. In the second part, we outline
the construction of metrics from a given family of vector fields. In the
third part, we show how to apply this construction to some examples from
several complex variables and partial differential equations, and obtain in
this way analogues of the results sketched in part one.
It is a pleasure to thank Professor M. T. Cheng, and my other hosts at
the University of Peking for their invitation to participate in the Summer
Symposium in Analysis in China. It was an honor and a privilege to attend
the Symposium, and I am grateful for the very warm hospitality I received.
It is also a pleasure to thank E. M. Stein for organizing and directing the
Symposium. All of his efforts are greatly appreciated.
Research supported in part by an NSF grant at the University of Wisconsin,

Madison.
241
242 ALEXANDER NAGEL
Much of this paper is an exposition of joint work with Eli Stein and
Steve Wainger, and I am particularly grateful to them for many years of
stimulation, encouragement, and collaboration.
Part 1. Some classical theorems and examples

In order to motivate our later discussion, we begin by considering
three examples of metrics: the standard Euclidean metric; a nonisotropic
but translation invariant metric on Rn ; and the translation invariant
metric on the Heisenberg group. In the Euclidean case, we see how the
balls and metric are involved in Fatou's theorem on nontangential limits
of Poisson integrals, and in estimates for the Newtonian potential and re-
lated singular integral operators. In the other two examples, we see how
analogous estimates can be made for kernels related to the heat operator
and to the Kohn Laplacian, and how nonisotropic balls on the Heisenberg
group are involved in Koranyi's extension of Fatou's theorem to existence
of admissible limits of holomorphic functions.
In each of these settings, there is a naturally given family of first
order linear homogeneous differential operators, or vector fields. In part II
of this paper, we shall see that the general construction of metrics applied
to these families of vector fields gives back the natural metric in these
classical settings.
The discussion of results in these examples will be very brief, but
references are given for the complete proofs of all the results.
1. The isotropic Euclidean metric and the Laplace operator

The standard metric on Rn is defined by
n 1/2
Ix-yI = Ixj-yj12
j=1
In this example, the important first order operators are just the partial
derivatives with respect to the n variables The Laplace
O-A 1 n
VECTOR FIELDS AND NONISOTROPIC METRICS 243
operator
is of course just the sum of squares of these first order operators.

We first study the role of the Euclidean metric in the solution of the
Dirichlet problem for Rn+1 = {(x1 xn,y) = (x,y)Iy>01. We identify the
boundary of R++1 with Rn = Rn x {0}, and, given a function f on Rn,
we want a function u(x,y) harmonic in Rn+1 such that u(x,y) - f(x0)
as (x,y) -> (x 0, 0) .
For continuous boundary data, the problem is completely solved by
the Poisson integral formula. Thus suppose f is continuous on Rn
and f c L1(Rn) + L(Rn) . Set
Pf(x.v) = P. * f(x) = c_v I f (t) dt

+1
2
Rn 1kX-t12+y21
where
n+1
cn=r (n21`/n 2
Then:
(a) Pf is harmonic on R++1
(b) Pf extends continuously to the boundary and takes on the
boundary values f.
(c) fRn IPf(x,y)ipdx < Iliii for 1 0 and fs(x) = u(x,s),

P(fs)(x,y) = u(x,y+s)
Proofs of these assertions, along with many other of the results dis-
cussed here, can be found in Stein [16], Chapter III.
Assertions (c) and (d) above suggest a generalization of the Dirichlet
problem to certain classes of discontinuous boundary functions. For
1 o.f
Rn
sup Iu(x,y)I = IIuII ,o<00 if p=c

R n+i h
Now we are interested in the boundary behavior (along y =0) of functions

u in hp , and it is here that the Euclidean metric begins to play an im-
portant role for us. To study the boundary behavior, a fundamental con-
cept is that of a nontangential approach region. For a > 0 and xo t Rn
define:
ra(xo) = (x,y) F R++1I Ix-xo1 < ayl
Note that if B(x0, S) = lx r RnI Ix-xoI < 8l are the balls defined by the
standard Euclidean metric, then
ra(xo) = I(x,y) a R++'Ix a B(xo,ay)l
Thus the nontangential approach regions in Rn+1 are really defined in

terms of the projection rr(x,y) = x of Rn+1 onto the boundary, the
"height function" h(x,y) = y, and the family of Euclidean balls on the
boundary. We shall later see that in other examples, natural approach
regions can be defined in essentially the same way.
We say that a function u(x,y) has a nontangential limit at x0 a Rn

if and only if for all a > 0, lim u(x,y) exists as (x,y) approaches
(x0,0) and (x,y) a ra(xp). In 1906 Fatou [4] proved:
THEOREM 1. For I < p < -, if u e hp , then u has a nontangential

limit at almost every point of Rn.
A standard modern approach to this theorem involves two main esti-

mates. The first involves the Hardy-Littlewood maximal operator. Let
f r L1oc(Rn) and set
Mf(x0) = sup IBL-i f If(Y)I dy

B
where the supremum is taken over all Euclidean balls B which contain
x0. The basic estimates for the maximal operator are given in:
THEOREM 2 (Hardy and Littlewood). For I All < AiA-i llflll if p =1.
The very definition of the maximal operator involves the family of

Euclidean balls, and the proof of the crucial estimate (ii) depends on a
covering lemma for these balls.
The second basic estimate needed to prove Fatou's theorem involves
the nontangential supremum of a function defined on Rn+i Thus for any
a > 0 and any v(x,y) defined on Rn+I set
Nav(x0) = sup Iv(x,Y)I .

(x.Y)fra(x0)
For Poisson integrals, this non-tangential supremum is point-wise domi-

nated by the Hardy-Littlewood maximal function of the boundary data:
246 ALEXANDER NAGEL
THEOREM 3 (Hardy and Littlewood). For a > 0 there exists a constant

Ca < , so that if f E L1(Rn) + L(Rn) and if u(x,y) = P,, * f(x), then
for all x c Rn
Nau(x) < CaM'(x) .
These are the two quantitative estimates which underlay the qualita-
tive statement of Fatou's theorem. Complete proofs of these results can
be found for example in Stein [16), Chapters I and III. However, since we
shall appeal to this kind of argument again, we now recall how Fatou's
theorem follows from these two theorems.
Let p < oo and let u c hp. If s > 0 and if we let fs(x) = u(x,s),
then
sup Iu(x,y+s)I = Na[P(fs)I(x0)
(x,y)Era(xp)
< CaIIM(fs)](x0) .
Therefore if A > 0
I}x0fRnI sup Iu(x,Y+s)I > A}I

(x.Y)fra(xo)
< I}x0ERnI M(fs)(xo) > Ca'A}I
< [CaA-IIIM(fs)IIp]P
< [CaAp011fs Iip]P
[CaApA-1 IIu IIh JP

P
Since s > 0 was arbitrary, we obtain for any u c hp
I;xocRnINau(xo)>,'}I < [CaApA-1IIuHIhp]P . (1)

Now let u of hp be real valued, and set
Slau(xo) = lim sup u(x,y) - lim inf u(x,y)
where the limits are taken as (x,y) approaches (x0,0) and (x,y) E Fa(xO).
Then the following facts are easy to verify:
(a) fau(x) < 2Nau(x)
(b) Sla(u+v)(x) < Slau(x) + Slav(x)
(c) Slau(xo) = 0 if and only if u has a limit within Fa(xo)
(d) Slau(x) - 0 if u = Pf and f is continuous.
Now let un(x,y) = u (,y + n) . Then
Slav = SZa(u-un + un) < Sla(u-un) + Sh(un)
= fla(u--un) < 2Na(u-un)
Hence, by inequality (1), we have:
Ilx ERnlfau(x)>A}I < Ilx ERnI Na(u-un)>A/21l
< [2CaApk-Illu-un11hp]p
Since p < o, llu_unllhp 0 as n - oc, and since A > 0 is arbitrary, it

follows that
Ilx ERnjjlau(x)>011 = 0 .
By taking a countable sequence of a's which increase to infinity, we

obtain a proof of Fatou's theorem.
It is clear that the family of Euclidean balls plays an important role
in this theorem, not only in the definition of nontangential approach
regions, but also crucially in the definition of the Hardy-Littlewood maxi-
mal operator and the proof of its boundedness. We now recall how these
248 ALEXANDER NAGEL
balls are also involved in studying the fundamental solution for the
Laplace operator.
An important fundamental solution for A is given by the Newtonian
potential:
N(x) =
/r( . Then AN = S as distributions. In particular,

where con = 2nn/2/P(2)
if C U (Rn)
c(x) = (2)
Rn
and if
'/r(x) = fN(x_Y)(Y)dY (3)
Rn
then Ar/r = 95.

Proofs of these facts can be found in Folland [5], Chapter 2.
A great deal is known about the operator
f N * f(x) = r N(x-y)f(y)dy .
Rn
Basically, the fundamental idea is that, when measured with appropriate

norms, N * f has two more orders of smoothness than f itself. For
example, if f satisfies a Holder continuity condition of order a,
0 < a < 1 , then f * N is of class C2 and all second derivatives again
satisfy a Holder continuity condition of order a. (See Bers, John, and

Schechter [1], page 232.) Proofs of these continuity properties of the
operator f -, N * f depend on size estimates of the kernel N(x,y) _
N(x-y) which can be written:
,32IB(x,8)I-I
IN(x,y) < C
(4)
IV N(x,y)I + IVyN(x,y)I < C sIB(x, s)I-1
where S = Ix-yI. Written in this way, these inequalities again make clear
the important role played by the Euclidean metric and the Euclidean balls.
This importance can also be seen when we consider certain singular
integral operators. We claimed earlier that N * f is two orders smoother
2
than f . Using formula (2), this means that for all i, j , should be
9256
as smooth as A S6. Thus we are led to the study of the operator which
32
carries A-0 to . There are two ways in which we can think of
i j
this.
Formally differentiating equation (2) we see that
(x) _ 5 kij(x-y)A S(y)dy (5)
Rn
where kij(y) = cn 2
ylyl
IyI
IyI-n
for an appropriate constant cn 0. We note
that the kernel kij is not locally integrable at 0 , so we must study the
integral in equation (5) in the principal value sense. Now the kernel
kij(x,y) = kij(x-y) satisfies the following estimates in terms of the
Euclidean metric:
250 ALEXANDER NAGEL
Ikij(x.y)I <CIB(x,6)1-1
S)I-t
IVxkij(x,y)I + Ioykij(x,y)I <
CS-1IB(x,
where S = Ix-yl. It is well known how to combine estimates of this type

with the Calderon-Zygmund decomposition of L1 functions to prove
2
LP(Rn) boundedness of the operator 0 - provided that one
J
knows that this operator is bounded on L2(Rn) (see for example, Stein
[161, Chapter 2). Boundedness on L2(Rn) follows from the cancellation
property
f kij(x)dx = 0 ,
a<Ixl<b
2
and can also be established by studying the operator AO in
terms of Fourier transforms. Define:
(e) = r
Rn
so that
-O(x) = f e
Rn
if, for example, 0 E C o (Rn) . Then:
(A) (e) = _4V2 lei 2 (e)
(e) _ -4n2 eiej Vi(e)

u&i/
so
a295 ei el
lei I
Our operator is thus given on the Fourier transform side by multiplication

by a bounded function. It follows from Plancherel's theorem that the
operator is bounded on L2(Rn); Thus we have
THEOREM 4. For 1 < p < 0 there are constants Ap < 00 so that if

95 E C p (Rn)
(9
2o
II IILp < ApIIok IILp
2. Spaces of homogeneous type

In our discussion of the Laplace operator, we emphasized the role of
the Hardy-Littlewood maximal operator and the Calderon-Zygmund decom-
position of Ll functions. These basic tools of Euclidean harmonic
analysis can be used in much more general settings, and can be applied to
a variety of interesting non-Euclidean examples. What is really necessary
for the theory to work is a measure space together with an appropriate
family of balls, and many people have been involved in these generaliza-
tions; see for example Koranyi and Vagi [10], Riviere [14], and Coifman
and Weiss [3]. Here I want to briefly sketch the approach of Coif man and
Weiss to what they call "spaces of homogeneous type."
A locally compact space X is a space of homogeneous type if there
is a continuous map p : X xX [0, 00) and a non-negative Borel measure
d on X so that:
(1) p is a pseudometric; i.e., for all x,y,z E X ,
(a) p(x,y) = 0 if and only if x = y
(b) P(x,Y) = P(Y,x)
(c) p(x,z) < K;p(x,y)+p(y,z)].
252 ALEXANDER NAGEL
(2) The measure has the "doubling property" relative to the family of
falls B(x,5) = IyeXJp(x,y)<5I: There is a constant A so that for all
x eX, 5>0
g(B(x,25)) < AW(B(x, 5)) .
In order to emphasize the crucial importance of properties (lc) and (2),

we now recall the proof of part (ii) of Theorem 2 in this general setting.
Thus if f e Lloc(X,d) , define
Mf(x0) = sup IBI-1 f If(y)Idsx(y)
where the supremum is taken over all "balls" B which contain x0. We
prove:
THEOREM 5. There is a constant Al so that if f e L1(X,dtt),
p(xeXIMf(x)>hI < A1A-1II01l .
Proof. For X > 0, let E = IMf >X1, and let F C E be any compact sub-
set. If x e E, since Mf(x) > A there is a ball Bx containing x so that
(Bx 1-1 I
If(y)Id(y) > A
Bx
The balls IBxlxel cover F, and since F is compact we can find a

subcover, which we call B1, , BN. Suppose Bj = B(xj,5j) so Bj has
"radius" 3 J.. Choose BI so that 5i1 > S i for all j. We can then
l -
inductively choose Bil, , Bik so that
(1) B. is disjoint from B. B.
`k 11 `k-l
(2) S ?a for all j such that B is disjoint from B B'k-1

`k > > 't
In this way we get a finite subsequence Bt , , Bt m which are disjoint.
1
Suppose B is a ball from our original finite sequence which does not
appear in the subsequence. Then there is a first k so that B; fl B 0.
k
By property (2), > &.. Let xik be the center of and xj the
Sik k
center of B1. Then if z < Bik fl Bl ,
P(xikxj) < K[p(xik,z) + P(z,xj)l
< 2KSi
k
If y c Bl , then p(xi,y) < Si , and so
P(xik.Y) < K[P(xik,xj) + p(xj,y)l
< K(2K+1)S. .
k
Let B k = B(xik,K(2k+l)Sik). Then Bi C B* and so
N 1m`
L.t C U Bj C U Bi k
j=1 k=1
m
Hence (F) < Y p(Bi ). But by property (2) of spaces of homogeneous
k=1 k
type
* 1+1092K(2K+1)
P(Bih) < A (Bik)
=A1p.(B'k).
Thus
Al I u(Bik) < Ala 1 If(Y)I dY < A 1X 1 IIfII1

k=1 k=1 J
B.
'k
254 ALEXANDER NAGEL
since the balls (BikI are disjoint. Since I was an arbitrary compact
subset of E , we obtain the same estimate for the measure of E.
3. The heat operator and a nonisotropic metric

We turn now to an example which, though elementary, involves truly
nonisotropic phenomena. We consider the heat operator
3t
In
j=J
a2
'= a
x
on Rn+1, where we use coordinates (x,t). Unlike the

Laplace operator A on Rn , L is not elliptic. Nevertheless there is a
remarkable fundamental solution for L. Define
if t>0
E(x,t) =
0 if t<0.
Then E is C on Rn+1\1(0,0)L, and LE = S in the sense of distribu-
tions. (See Folland [5], Chapter 4.) Thus if 0 c C(Rn+1)
0 +
r00 n
(k(x,t) = J (4rrs) 2 e-IYI2 /4s I(x-Y,t-s) dy ds
J
0 Rn
= f S (4(t-s)
2e--2/4(t-(Y,s)dyds
J
where L = - - A is the formal adjoint of L.

In order to study the operator f -+ E * f , we would like to obtain size
estimates on the kernel E and its derivatives analogous to those for the
Newtonian potential N in equation (4). However, if B((x,t), S) denotes

the ball in the standard Euclidean metric, the estimate
JE(x,t)l < C 321B((x,t), 8)l-1 8-n+1
with a= (Ix12+t2)1 /2
is false. To see this, let t =x12 with lxi
small, so E(x,t) =CIxt-n, while S =(Ix12+Ixl4)1/2 ti IxJ, so
S-n+l ti xj-n+1 Thus we cannot obtain the appropriate kind of estimate
with the Euclidean metric.

To remedy this situation, we make the crucial observation that, like
the Newtonian potential N, the fundamental solution E possesses a
certain homogeneity, though now this homogeneity is nonisotropic. For
A > 0 define
SX(x,t) = (Ax,A2t) .
It is easy to check that
E(SA(x,t)) = E(Ax,A2t) = A-nE(x,t) .
We associate to the family of dilations a pseudometric
(Y's)) (lx-y14 + (t-S)2)1 /4

P((x,t), _
so that
P(Sa(x,t), Sx(Y,s)) = kp((x,t), (Y's))
The corresponding family of balls
BP((x,t), S) = {(Y,s) a Rn+1 jp((x,t), (Y,s)) < SI
are now ellipsoids of size S in the directions of x1, ,xn, and of size
S2 in the direction of t. Thus
8n+2
IBP((x,t), S)1 ti
256 ALEXANDER NAGEL
Moreover, it follows from the homogeneity of E that we now have:
(a) IE(x,t)I < C 821Bp((x,t), 6)1-1
(b) IV E(x,t)I < C6IBp((x,t),6)1-1 (7)
(c) IdE (x,t) < C 1Bp((x,t), 6)1-1
(d)- (x,t)I < C IBp((x.t), 6)x-1
where 6 = p((0,0), (x,t)).

Thus we obtain estimates for the fundamental solution E(x,t) which
are exactly analogous to those we have for N(x), provided we view the
operator as acting like a second order operator, so in equation (7c)
we loose two powers of 6 rather than one.
We can now use the general theory of spaces of homogeneous type to
show that L satisfies certain subelliptic estimates analogous to the
elliptic estimates for N given in Theorem 4. For example, one can prove:
THEOREM 6. For 1 < p < oc there are constants Ap < oo so that if

0 e Co(Rn+1) then
< Ap IIL0IIp
Ilp
a20
A p IIL0li p
p <-
The operator L is the sum of squares of the first order operators

(9 a minus the operator , which we shall count as an operator
1 a
of order two. We shall later see how the appropriately weighted vector
fields i-, , _ , and as give back the nonisotropic metric p.
1 n
4. The Siegel upper half space and the Heisenberg group

Nonisotropic balls and metrics play an important role in the theory of
boundary behavior of holomorphic functions in strongly pseudoconvex
domains. This is discussed for example in the monograph by Stein [171.
Here we recall what happens in the case of a model strictly pseudoconvex
domain, the generalized upper half space, and its boundary, the Heisenberg
group.
We let
H= {(z1,...,zn'zn+1)
n
= (z,zn+l) c Cn+1 I1mzn+l
> I izjI2 = IzI2
j=1
n+1
Recall that U is the image of the unit ball B = (w1,-',wn+1)I jwj12<1
under the biholomorphic map
wk
zk = 1<k <n
l+wn+1
l-wn+1
zn+1 = I l+wn+l
The boundary of H is the set
an = I(z,zn+1) c Cn+1IImzn+1 = Iz121
= I(z,t+ilz12)Iz c Cn,t c RI
Thus we can identify afl with Cn x R by associating to the point

(z,t+ilz12) c dfl the point (z,t) c Cn x R. We can make Cn x R into a
group Hn, the Heisenberg group, by defining
(z,t) (w,s) = (z+w,t+s+2Im<z,w>)

n
where <z,w> = I zjwj. This definition can be motivated and the group
j=1
258 ALEXANDER NAGEL
properties verified by the following considerations: to each point

h = (w,s) f Cn x R we associate a holomorphic map Th : Cn+1 Cn+1 by
the formula
Th(z,zn+1) _ (z+w'zn+1 + +2i<z;w>) .
Th is clearly holomorphic. Moreover, if hj = (wj,sj), j = 1,2, then a

simple calculation shows that
Th2 ' Th1(z,zn+l) = Th3(z,zn+l)
where
h3 = (w1+w2, s1+s2+2Im<w2w1 >)
= h2 h1 .
From this it follows that Hn is a group, and
Th2 Thl = Th2,hl .
Let p(z,zn+1) = lmzn+l - Iz12 be the "height function" for f1 so

that (z,zn+1) f ,f1 if and only if p(z,zn+1) > 0, and (z,zn+1) f c3S1 if
and only if p(z,zn+1) = 0. If h = (w,s) f Hn we compute
p(Th(z,zn+l)) = Im(zn+l+s+ilwl2+2i<z,w>) - (z+wl2
= Imzn+l - Iz I2
= p(z,zn+l) '
Thus Th carries i1 to itself and to itself. If h = (w,s) or Hn, we

f3I1
have identified (w,s) with the point (w,s+iiw)2) f d2, and this is the
same as Th(0,O). Thus under the identification, Hn acts on (90 as
follows: if (z,t) fan and if h = (w,s) f Hn , then
Th(z,t) = ThT(z.t)(0,0)
= Th.(z,t)(0,0)
_ (w,s) (z,t)
(z +w, t+s -2 Im< z,w >)
We now make Hn (or ail) into a space of homogeneous type. Define

d : Hn x Hn -+ [0, c) by
d((z,t); (w,s)) _ [Iz-wI4+It-s+2ImCz,w>I211/4 ,
or, if we identify Hn with rail,
d((z,zn+1); (w,wn+l)) _ [Iz-wI4+Re(zn+l-wn+l-2i<z,w>)I211 /4
The function d is invariant under the action of H Thus if

n.
h = (u,r) r Hn , easy algebraic manipulations show that
d(Th(z,t), Th(w,s)) = d((z,t), (w,s))
If we define the balls
B((w,s),8) _ I(z,t)Id((z,t), (w,s)) < 81
then this invariance property means that
B((w,s),8) = T(w,s)(B((0,0), 8))
We claim that d is a pseudometric. For if z = (z,t), w = (w,s) and

u = (u,r) are in Hn with
d(z,w) < 8, d(w u ) < 8
then
1z-w( < 8 It-s+2lm<z,w>I < 82

1w-ul < 8 Is-r+2lm<w,u>I < 82
260 ALEXANDER NAGEL
Hence
Iz-u I < Iz-w I + Iw-u I < 28 ,
and
It-r+2Im<z,u>I < It-s+21m<z,w>l + is-r+2Im<w,u>I
+ 12Im(<z,u>-<z,w>-<w,u>)I
< 232 + 21Im<z-w, u-w >1
< 432 = (2S)2 .
Theref ore
d(z u) < 2 m a x (d(z,w ), d(w,u ))
< 2 [d(-Z-, w ) + d(w, u ) I .
What do the corresponding balls B(i,S) look like? We see that

w c B(Z,S) is essentially equivalent to the pair of inequalities:
Iz-wI < S
)Re (zn+i-wn+1-2i<z ,w >)I < 82
Fix z = (z,zn+1) a asl. The complex tangent space to asl at z is

given by the equation
(''n+1 - Zn+1 - 2i<w,z > = 0 .
If w = (w,wn+1) a afl, the distance from w to this complex tangent plane

is essentially
Iwn+i-zn+1-2i<w,z >1
= IRe(wn+1-zn+l)+i(lw12+Iz12-2Cw,z>)I
= IRe (wn+l-zn+1-2i<z,w >)+i lw-z 121
ti lw-zI2+(Re(wn+1-zn+1-2i<z,w>)l
ti< S2
Thus the balls B(z,S) are essentially "ellipsoids" of size S in the

directions of the complex part of the tangent space to df at z , and of
size S2 in the orthogonal real direction, and hence in particular
B(- Z,
ti 82n+2
Thus the doubling property of the balls is verified, and di1 (or Hn )
equipped with the pseudometric d is indeed a space of homogeneous type.
We now want to discuss the analogue of Fatou's theorem for boundary
behavior of holomorphic functions in Q. This problem was first studied
by Koranyi [9] for domains like 1, and was later generalized by Stein [17]
to general smoothly bounded domains in Cn. Here we want to emphasize
the role of the nonisotropic balls on the boundary, in analogy with the
role of Euclidean balls on Rn in Fatou's theorem.
We begin by defining appropriate nonisotropic approach regions in it.
Let n : SZ K1 be the projection
n(z,zn+1) = (z,zn+l-ip(z,zn+1)) -
For a> 0 and w = (w,s+iIwl2) c dSl let
Aa(w) = l(z,zn+l)cg!n(z,zn+l)EB(w,ap(z,zn+l)
where of course B(w,S) is the nonisotropic ball defined by the pseudo-

metric d . It is clear that this definition is analogous to our earlier
definition of nontangential approach regions Fa(xo) in Rn+1
Now (z,t+ilzl2+iy) f Aa(w) is essentially equivalent to the following
pair of inequalities:
Iz-wI <ay1/2
It-s+21m<z,w>I < ay .
As above, It-s+2Im<z,w>I is essentially the distance from (z,t+ilz12)

to the complex part of the tangent space at W. Thus Aa(w) allows
tangential approach to w of order two in the complex directions at w,
262 ALEXANDER NAGEL
but requires nontangential approach in the complementary real direction,

and so the sets (Aa(w )1 are essentially the admissible approach regions
introduced by Koranyi [9].
The invariance of the balls B(w, S) under the action of Hn , together
with the fact that the height function p is invariant under the mappings
Th, imply that the approach regions Aa(w) are carried into each other
under the action of Hn H. Thus it is easy to check that if h = (w,s) c Hn,
then (z,zn+1) t Aa((0,0)) if and only if Th(z,zn+t) E Aa(w,s+iIwl2).
Let F be a function defined on 11, and for C c 3O , set
NaF(Co) = sup IF(z,zn+1)I

(z,zn+l)tAa(y SO)
If f c Lloc(3t1) , define
Mf(Co) = sup IBI-1 lf(C)I da( )

J
B
where the supremum is taken over all nonisotropic balls B containing

Co, and do is surface area measure on Al.
We now have the following analogues of Theorems 2 and 3, which were
used to prove Fatou's theorem.
THEOREM 7. For 1 0 there is a constant Ca (independent of u) so that for all
Ct do
Nau(C) < CaMu(C) .
Theorem 7 of course follows from Theorem 5 and the Marcinkiewicz

interpolation theorem (see Stein [16], Chapter I), since we already know
dfl is a space of homogeneous type. Before proving Theorem 8, we point
out some of the consequences of these results.
For 1<p< 0, we let Hp(l) denote the space of holomorphic func-
tions on 9 which satisfy
sup IF(z,t+ilzl2+iy)IPda(z,t) = IIFIIH < o if p < 00

y>o p
an
sup IF(z,zn+l)I = IIFIIH 0. There are constants Ap a < 00
so that if F f Hp(11)
(i) INaFNLP < Ap,aIIFIIH for '<P< oo
P
(ii) IlCE3f INaF(C)>AII < A1,aa-1IIFIIHI if p=1.
Proof. Put FE(z,zn+l) Then FE is continuous on fl,

=F(z,zn+l+ie).
and pleurisubharmonic on 0, so by Theorem 8
NaFe(C) < CaMFE(C) .
Since IIFEI{LP < 1F for 1 ,'MI < AICaa-1IIFIIHI
and
IINaFEIILp _< ApCaIIFIIH if P>1 .

P
But NaFE(() ,' NaF(C) as a 0, and

264 ALEXANDER NAGEL
INaF>Al = U INaFE>AI
E<0
so the corollary follows by the monotone convergence theorem and the

regularity of the measure do.
We can now apply exactly the same argument we used for Fatou's
theorem to prove
THEOREM 9 (Koranyi). If F c Hp(Q) , then F has admissible limits at

almost every point of caf2.
It remains to prove Theorem 8. For simplicity, we deal with the case

n =1 . Since the approach regions Aa(C) are carried into each other by
the action of Hn , it suffices to prove the estimate at the origin (0,0).
We first obtain an estimate on the radial maximal function:
2
Iu(O,iy)1 < -I
Ay If 2
IS12+It-yI2< Y)
f211
< 2
Ir2y2 If 2 0
u(N/tei0,s+it)Id9dsdt
Is I2+I t_Y 12<(2)
y 3y
2 rr 2 2
r
0
f Iu(V/teie,s+it)Idtdsd9.
Letting r = V/t , and interchanging the order of integration, we get

Iu(O,iy)I < 4 u(re10, s+ir 2) Irdrd gds

- n2y2
< 4
Try2 f
B((0,0),2\)
Z IB((0,O),2,Vy-)I-1
J Iu(C)Ida(C) .
B((0,0),2 J )
Thus, by translation invariance, we see there is a constant A so that for

(z,zn+l) E lZ
Iu(z'zn+1)I <_ AIBI-1 J lu(C)Ida(C)

B
where B=B(rr(z,zn+l),2p(z,zn+1)1/2),
Now let (z,zn+l) E Aa((0,0)) . Then ir(z,zn+1) E B((O,O), ap(z,zn+1)1 /2)

so
d(n(z,zn+1), (0,0)) < ap(z,zn+1)1/2
If Cc B(rr(z,z n+1),2p(z ,zn+1)1 /2) it follows that
d (C, (0,0)) < (2a+4) p(z,zn+1)1 /2
But now using the doubling property of the balls, it follows that there is
a constant A so that if (z,zn+1) E Aa(0,0)
I(z,zn+1)I < A'IBI-1 Ju()1da() (8)

B
266 ALEXANDER NAGEL
where B = B((0,0), (2a+4)p(z,zn+1)112), and this last average is certainly

dominated by A'Mu(0,0). We remark that estimates of type (7) are
actually false for Poisson integrals of harmonic functions.
The nonisotropic balls on dV are also useful in studying singular
integrals and fundamental solutions. For example, the Cauchy-Szego
projection of L2(ail) onto H2(SZ) is given by the operator
S((z,zn+1) =
asp
where
=cn[i(wn+t-zn+1)-2<z,w>]-(n+1)
S((z,zn+1),(w,wn+i))
with cn = 2n-1n!/nn+1 (see Nagel and Stein [11], page 23). In particular,
for z = (z,t) and w = (w,s) on dQ we have,
.
S(z,w) = cn(i)-(n+1)[(s-t-2Im<z,w>)-ilz-wl2]-(n+1)
Thus IS(z,w )$ IB(i,8)I-1 where 8 = d(i,w ). It is also easy to check

that derivatives of S yield estimates with corresponding negative powers
of S. Thus S behaves like a singular integral kernel relative to this
family of nonisotropic balls, and from this follows LP and Lipschitz
estimates for the operator f -> Sf (see Kornyi and Vagi [101).
Finally, we consider fundamental solutions. On Hn let
Xj =
+2YjYj = -2xj. , T
where we write zj = xj + iyj. These vector fields form a basis for the
left invariant vector fields on Hn. Put
Zj = (Xj-iYj), Zj = (Xj+iYj)
2 2
and consider for a c C.
n
'a=-2 1 (ZjZj+ZjZj)+iaT
j=1
This second order operator arises in the following way: if we identify Hn

with asp, the vector fields Zj annihilate the boundary values of holo-
morphic functions. Thus, in analogy with the operator a, we consider
n
abf = Zjdzj on functions,
j=1
and we extend this in the usual way to (0,q) forms on Al. In L2(Hn)
we can define a formal adjoint (ab)*, and the Kohn Laplacian is then
b=ab+ab .
On q forms, obq) acts diagonally, and is given by the operator '2a

where a = n - 2q. The operator ob is not elliptic but Kohn's fundamental
work [8a] showed that one can obtain subelliptic estimates for a.
Folland and Stein [6] discovered a fundamental solution for a. Define:
_jn+a a
(1212+jt) -(n 2
`` 2
Y'a(Z,t) = (IZ I2-it)
THEOREM 10 (Folland and Stein). '2acba = cab in the sense of distribu-

tions, where ca is a constant, and ca 0 if a n, (n+2), (n+4),
etc.
Thus except for the exceptional values of a, c Oa is a fundamental

a
solution for 2a , and it is easy to verify that
C52IB((0,0),5)I-1
I'ka(z,t)I <
where S = d((0,0),(z,t)). One also obtains corresponding estimates for

derivatives of Oa, so that again there is a complete analogy with the
estimates (4) for the Newtonian potential.
268 ALEXANDER NAGEL
In the case of the Heisenberg group, the basic vector fields are
X1,"',Xn'Y1,"',Yn, and the vector field T which is given "weight"
two. We shall later see how the general construction applied to these
vector fields gives the nonisotropic pseudometric d .
Part II. Metrics defined by vector fields

Our object in this part of the paper is to outline the construction of
metrics from certain families of vector fields. Many details of the argu-
ments will be omitted, and complete proofs can be found in [13]. In [71,
Hormander studied differentiability along noncommuting vector fields, and
used the techniques of exponential mappings and the Campbell-Hausdorff
formula. The case of vector fields of type 2 was studied in [11]. Balls
reflecting commutation properties of vector fields have also been studied
by Fefferman and Phong [4a], by Folland and Hung [5a], and by Sanchez-
Calle [15a 1.
Let Q C RN be a connected open set, and let be C
real vector fields defined on a neighborhood of Q. We associate to each
vector field Yj an integer dj = d(Yj) > 1 which we call the formal
degree, and we make two fundamental assumptions about this collection
of vector fields.
(1) For each x c Q, the vectors span RN.
(2) For all j, k, we can write [Yj,Yk' = Y-
ck
. (x)YE where
d1<di+dk 1
cek c C(Q). Here [X,Y] --- XY -YX is the commutator of the two
J
vector fields.
There are several basic examples to keep in mind.
(A) Let N = q = n , let Yi = and let dj = 1 for 1 < j < n . In this
case we shall recover the standard Euclidean metric on Rn.

(B) Let N=q=n+1, let Y and let d1 for 1<j<n and
let do+1 = 2 . In this case we shall obtain the nonisotropic metric
on Rn+1 appropriate for the study of the heat operator
n+l I
I c3xj
012
(C) Let N = q =2n4-1 and let
Yj +2yj 1<j<n
with dj = 1
Yn+j = - 2xj 1<j<n
Yj
Yen+1 = [Yj,Yn+j] _ -4 at with den+1 = 2
In this case of course we are dealing with the Heisenberg group, and
we shall recover the invariant metric on Hn defined earlier.
(D) (An example of Grushin type). Let N = 2, q = 3, and let
X Y3 = [Y 1,Y 2] = a
2
0x2
with dl=d2=1 and d3 = 2 . This example leads to a metric

2 2
appropriate for studying the hypoelliptic operator + x2
(2
2
See Grushin [6a]. 1 2
(E) (This example generalizes (A), (C), and (D).) Let be C

real vector fields on Sl. Let
X(1) =
iX1,...,xp1
x(2) = ; . . . , [x1 x I, ...1

X(3) = #..., [Xi, [XJ,Xk]], 1, etc.
so that X(k) is a vector whose components are all the commutators

of of length k. We say that the vector fields
are of Hormander type m or finite type m if at each x e 1, the
components of X(1) ,X(m) span RN.
270 ALEXANDER NAGEL
Let be some enumeration of the components of

and if Yi is an element of XU) we set d(Yi) = j.
Then property (1) follows from the assumption of finite type, while
property (2) follows from the Jacobi identity for commutators.
This example leads to a metric appropriate for studying the hypo-
elliptic operator Xi + + Xp and its variants.
(F) (This gerreralizes (B).) Let Xo,Xl,.",X, be C real vector fields
as in (E) of finite type, but this time let d(X3) = 1 for 1 < j < n
while d(X0) = 2, with appropriate weighting of higher commutators.
This example leads to a metric appropriate for studying the generali-
zation of the heat operator Xi+Xi Xp-X0.
Our object now is to construct natural metrics out of the family of

vector fields Y1, .,Yq . I want to distinguish between two general
approaches to this problem.
On the one hand, one can define a metric in terms of a "global defini-
tion." Here, we let the metric be given as the infimum of some functional
over a large class of curves. Then the defining properties of a metric,
such as the triangle inequality, are relatively easy to verify, but the local
geometry of the corresponding family of balls is hard to understand.
On the other hand, we can define a metric in terms of a "local defini-
tion"; here we want the metric to be given in terms of an exponential
mapping. Now the local geometry is clearer, but it is much more difficult
to verify that one really has a metric.
We now discuss each of these approaches, and then try to sketch why
the two definitions are in fact equivalent.
5. Global definitions
DEFINITION 1. Let xp,xl c f2, and say p(xp,xl) < S if and only if
there is an absolutely continuous map 0: [0,1] -+ fl with c(j) = xj ,
j = 0, 1 , so that for almost all t c (0,1)
q
(t) _ I aj(t)Yj(cS(t))
j=1
d(Y.)
where laj(t)I < S
PROPOSITION. p is a metric on Q. For every compact set F CC i2

there are constants C 1, C 2 so that if x0,x 1 < F
C11x0-x11 <p(xo,x1) <C21xo-x11m
where 1x0-xII is the standard Euclidean metric.
We sketch the proof:
(1) That p(x,y) = 0 if and only if x = y is clear.

(2) p(x,y) = p(y,x) since we can replace ca(t) by 0(1-t).
(3) Given x and y , there is a smooth curve joining them, so p(x,y) <
(4) The triangle inequality: Suppose x,y,z e Q. Given e > 0, there are
curves 0, 0: [0,1] -.1Z with 0(0) = x, b(1) = y , 0(0) = Y, l'(1) = z ,
q q
4'(t) = a j(t) Y -(O(t)) , &'(t) = F b j(t) Y j('(t)) , with
1
d(y,z).+E)d(Yi)
Iaj(t)I < (d(x,y)+e)d(Y1) Ibj(t)I <
Define 0: [0,1] - 9 by
0(at) 0<t<1/a
0(t) =
at-1 1<t<1
as -1 a
d(y,z) + E
where a = 1 + Then 0(0) = x, 0(1) = z and 0'(t)
d(x,y) } E
q
E cj(t)Yj(0(t)) where
j=1
Icj(t)I < (d(x,y)+d(y,z)+2E)d(Yi) .

This shows, since a is arbitrary, that d(x,z) < d(x,y)+d(y,z).
Now let F C it be compact. Then there is a constant C = C37 so
that if x,y c F , there i s a smooth curve cb joining x to y with
l,0'(t)I < Clx-yl. Since span RN at every point, we can
q
write (t) = I cj(t)Yj with Icj(t)I < CI0'(t)I . But then:
j=1
C[Ix-yId(Yj)]d(Yj)
Icj(t)I < CIx-Y1 =
< C[Ix_yIm]d(Yj)
and so p(x,y) < CIx_yIm

Conversely, if x,y c I and p(x,y) = 8, there is a curve joining
q d(Y )
x to y with 0'(t) = F aj(t)Yj(c(t)) and Iaj(t)I < (28) i almost
j=1
everywhere. But then
1
O'(t)dt
J0
1
q
< I laj(t)I IYj(O(t))Idt
0
j=1
< C8
since the lengths of the vectors (Yjl are uniformly bounded on 0.

We now give two other possible global definitions of metrics.
DEFINITION 2. Let x0,x1 E Q and say p2(x0,x1) < 8 if and only if

there is a C curve : [0,1 ] U with (j) = xj , j = 0,1 , and
q d(Y)
q'(t) = I ajYj(c(t)) where aj c R and Iajl < 8 I . Note that in this
j=1
definition, we have somewhat restricted the class of curves over which we
take an infimum. However, in general it is not true that p2(x,y) is
finite for any two points of Q. For example, if 9 is not convex,
N=q=n, Yj = and dj =1, 1 <j<n, then p2(x,y)<oo if and
1
only if x and y can be joined by a straight line contained in SZ so p2
is not even a pseudometric in this case. On the other hand, it is clear
that at least in this example, the metric p2 is locally equivalent to the
standard Euclidean metric.
Now suppose we are in the case of vector fields of

finite type m.
DEFINITIQN 3. Let x0,x1 E Q and say p3(x0,x1) < 3 if and only if

there is an absolutely continuous map 0: [0,1] - 9 with q(j) = xj,
a
j = 0,1, with 0'(t) = E aj(t)X.(c(t)) and Iaj(t)I < S almost everywhere.
j=
Note that in this definition, we only allow the derivative 0'(t) to

belong to the span of which may not be all of RN.
It is again easy to verify that p3 satisfies the triangle inequality, but it
is not a priori clear that p(x,y) is finite for any two points of Q. This
is in fact a consequence of the finite type hypothesis, and was first proved
by Caratheodory in 1909 [2].
To get some feeling for why p3 is finite, and to begin to understand
the role of commutators, let us consider the following example of Grushin
type in R2. Let the coordinates be x and y and let X1 = &,
X2 = xk Note that these vectors fail to span R2 along x = 0, but
[XI,X2] = kxk-1 , [X1[x1,X2]] = k(K-1)xk-2 19 I...

13Y O
[X1, [x1,x2] ]] = k! vy
where the commutator is of length k + 1 , so (X1,X21 are of finite type

k+1.
What points belong to the ball centered at (x0,0) of radius 3 ? We
shall consider two special cases: x0 = 1 and x0 = 0.
274 ALEXANDER NAGEL
Case 1. If we let
1(t) _ (1 +St,0) so ci(t) = S
and
952(1) = (1,St) so c?(t) = S (1)k
then this shows that the points (1 +5,0) and (1, S) belong to the p3
ball centered at (1,0) of radius S. In fact, this ball is essentially the
Euclidean ball of radius S centered at (1,0).
Case 2. Again, starting at (0,0) we can go to (5,0) by using the vector

field (3 , but it is not immediately clear how to get from (0,0) to other
points on the y axis. To do this, we need to use a curve which is only
piecewise smooth. Let
.0 1(t)_(St,O) 0<t<1 so 95i =SX1
QSZ(t) = (S, Sk+lt) 0<t<1 so 952 = SXZ
c63(t)=(S-St,Sk+1) 0<t<1 so 95g=-6X1
This shows that the p3 distance from (0,0) to (0,8k+1) is at most 38.
In fact the p3 ball with center (0,0) is an "ellipsoid" of size S in
the x direction and size Sk+1 in the y direction.
Thus we have given three possible definitions of a metric or pseudo-
metric in terms of the family of vector fields It is also clear
from the definitions that we have the following inequalities:
P(x,y) < P2(x,y)
p(x,y) < p3(x,y) (when p3 is defined).
In fact, these three quantities are locally equivalent. One can prove that
for every x0 E (I there is a neighborhood U of x0 and constants C1,C2
so that for all x,y c Q, x 4 y , and j =2,3.
P3(x,y)
0<cI<P(x,y)<c2<oo.
6. Local definitions
In order to begin to see why these pseudometrics are locally equivalent,
we need to consider a local definition of metric, and this in turn relies on
the notion of an exponential map.
Given a point p c (1, we let TpiZ denote the tangent space to SZ
at p. Suppose we are given C vector fields defined near
p which form a basis for RN at p. Then we can construct a map from
a neighborhood of 0 c T 1 to a neighborhood of p in [ as follows:
N
every tangent vector v at p can be uniquely written as I
j_1 J j
(p) = v
N
with E RN, and soj=1
I J J
a smooth vector field defined
near p. We can flow along the integral curve of this vector field for unit
time if IlajI is sufficiently small, and the result is by definition
N
exp jS (p) , the exponential map of V.
J
Given vector fields we can identify T f with RN via

N
1Y'
ajSj(p), and then the exponential mapping
(al,...,aN) .exp (ajSj) (p)
introduces a coordinate system centered at p , the so-called canonical

coordinates relative to the vector fields The Jacobian of this
exponential map at 0 c RN is just the volume of the
"parallelopiped" spanned by S1, ,SN. It is important to remember how-
ever that this exponential map from TP( to [ depends on the choice of
vector fields S1, SN.
Now we return to the general situation of vector fields in
1 C R.N On each tangent space TxI there is a natural notion of length.
276 ALEXANDER NAGEL
If V_ E Tf) we say
Nx(v)<a
if and only if
q
v = ajYj(x)
j=1
d (Y ) -.
where J <5 i . Of course this representation of v need not be
Ia j
unique. Nevertheless questions about the set
IV ETxSZINx(v)<6
are presumably just problems in elementary linear algebra. In giving a

local definition of metric, we want to transfer these "balls" in the
tangent space at x to genuine balls in 11, and this suggests the use of
an exponential map. The question is: how does one choose an appropri-
ate N-tuple of vector fields in order to construct such a map?
To motivate the answer, we first ask a simple question: what is the
volume in TxQ of the set { vlNx(v)<SJ ? This amounts to the following
problem. Let be vectors in RN which span, and consider
the map
q
(a 1 ,...,aq) = ajyj
j=1
What is the volume of the image under of the box
d (Y .)
QS = 1(a1,...,aq)(RgI IajI<S I?
For any N-tuple I = let d(I) = d(Yi )+ ... + d(Y. ), and let
1 N
AT = det (Yi1,.. ,YiN).
LEMMA. There are universal constants C1,C2 so that 0 < C1 <

Ip(QS)I/ I IXII Sd(I) < C2 < 00. (Here the sum is taken over all N-tuples I.)
I
Proof. For each N-tuple 1, the image 0(Q3) contains all vectors of
N
d(Yi)
the form ajY1 where Iaj I < and this is just the image under
i=1
a linear map`from RN to RN with determinant AI. Thus
1A11 sd(I) < 1(ns)1
5d(I)
IXII < (;l)
We must now prove the reverse inequality. Pick an N-tuple 10 so that
IAI01 ad(I0) > 1A J I ad (J)
for all N-tuples J. By renumbering, we may assume 10 =

Since AI0 0, we can write
N
Yj=4 bjkYk 1<j<q,
k=1
and using Cramer's rule, we see that
A,
Ijk
bjk X1
0
where Ijk is obtained from 10 by replacing Yk by Yj. From our

choice of I0 , it now follows that
d(Yk)-d(Y j) .
Ib jkI _ 5
278 ALEXANDER NAGEL
q
Now let v c O(Q8), so v -_ ajYj with Iajl < 8 d(Y I) . Then
I
j=1
N 9
= " ajbjk Y k
V k- j=1
and
q
1: ad(Y
< J)Sd(Yk)_d(Yj) = q Sd(Yk)
j=1
(gS)d(Yk)
Hence
d(I) I d(1 )
!)L1
IB(Q8)I < q
0
< qmN IA118d(I)
This lemma suggests the following further definitions. For each

x c 12 and each N-tuple I = let
N d(Yi
BI(x,$) = YcHIY=exp(I ajYi) (x) with jajI <8
1 )
Clearly for every I
BI(x, 8) C (y c 12Ip2(x,Y)<8j B2(x, 6) .
Hence U BI(x,6) C B2(x,S). We also have

I
B2(x,8)C B(x,6) = iycI p(x,y)<81
since p(x,y) < p2(x,y).

It is now reasonable to conjecture that for each x c 11 and each 6 > 0,

if we choose an N-tuple 10 so that
IA1
(x)I ad(10) > IAJ(x)I
8d(J) for all J
o
then for this particular 10
B(x, S) C Bl (x, ca)

0
where C is a constant independent of x and 8. This would of course

show that p and p2 are locally equivalent, and also show that the local
and global definitions of metric are equivalent.
We give a very brief sketch of a proof. Suppose y f B(x,6). Then
there is 0 : [0,1] - t with 0(0) = x , 6(1) = y and
q
0,(t) _ bj(t)Yj(o(t))
j=1
Sd(Y'
with lbj(t)I < almost everywhere. Now assume without loss that
I0 = and let denote canonical coordinates near x
relative to the exponential map using Then the curve c(t)
is given in canonical coordinates by (uI(t),...,uN(t)) and our object is
d(Y.)
to show that there is a uniform constant C with Iuj(1)I < (CS) . But
uj(1) = uj(1) - uj(0) = r [uj(t)]dt

0
1 q
f Y, bk(t)Yk(c(t))(uj)(t)dt
k=1
0
Now since IY11...IYNI span near x we can write

2$O ALEXANDER NAGEL
N
Yk =1 akYP
=1
where ak (C". Thus

q
u (1) _ Lr I k(t)ak((b(t))(Ypu))(t)dt
b
P=1 k=1
0
We now need to make two kinds of estimates:
(1) Iak(c (t))I < Cad(Yprd(Yk)
(2) IYPuj(t)I <
Sd(Y k), d(Y

Combined with the estimate Ibk(t)I < this gives Iuj(1)I < CO
which is what we want.
Now estimates (1) are obtained by using Cramer's rule to write
ak(y) _ A J(y)/A1 (y) as in our earlier lemma. At x , our choice of Io
0
Sd(Yp)-d(Yk)
shows that IAJ(x)I/IAI (x) I < and in order to obtain a
0
similar estimate for nearby points y, we expand at in a Taylor series
in canonical coordinates about x .
The proof of (2) is more complicated, and involves the Campbell-
Hausdorff formula. However, note that at x , Ypui = S)p, which is the
right estimate there. Complete proofs can be found in (13].
As a corollary of our analysis of the metric p , we can estimate the
volume of the balls B(x,8). For every compact I CC it there are con-
stants C1 and C2 so that
C1 (IIXi(x)I6')) < IB(x,6)I < C2Ì

I IAI(x)Isd(I)/
I
In particular, the balls B(x, S) satisfy the doubling property, and so Sl

with metric p and Lebesgue measure is a space of homogeneous type.
Part III. Applications

In this part of the paper, we show how the constructions outlined in
part II can be applied in partial differential equations and several complex
variables.
7. Estimates for approximate fundamental solutions

We can ue the metrics constructed from vector fields to obtain esti-
mates for the integral kernels of parametricies of certain hypoelliptic
differential operators. We briefly describe the setting. In 1967 Hormander
[7] obtained a far reaching generalization of Kohn's result on the hypo-
ellipticity of b. Suppose are C real vector fields on
a C RN of finite type m. Hormander showed that the second order opera-
tors or are hypoelliptic. As in the work
of Folland and Stein [6] on b , Rothschild and Stein [15] want to con-
struct parametricies for these more general operators by inverting model
operators on appropriate nilpotent groups. There is in general no nilpotent
group of dimension N which works, but Rothschild and Stein overcome
this difficulty by proving their "lifting theorem." Given on
iZ C RN (with coordinates of type m, they show that one
can find additional variables E Rs and form new vector fields
S
X =X+ ajf(x,t) .
f=1
These new vector fields will again be of type m on a neighborhood of

iZ x 101 C iZ x RS , and in addition they are free up to step m , so that one
can model by vector fields on a free nilpotent group of step m
with p generators.
Rothschild and Stein are able to construct a parametrix for
'X-2 +... +
Xp on 9 x Rs, given by a kernel k((x,t), (y,s)) which comes
from a homogeneous kernel on the nilpotent group. In particular, this
kernel satisfies
282 ALEXANDER NAGEL
Ik((x,t),(y,s))I < CS21]j((x,t),6)I-t
where S = p((x,t), (y,s)), and p is the metric constructed from the vector
fields They then define a restriction operator
Rk(x,y) = fk((xiO)(Yis))(s)ds
RS
where 0 E C(Rs) is supported near s = 0, and if k((x,t), (y,s)) is the

parametrix for Xi R2p, then D(x,y) = Rk(x,y) is a parametrix for
X +... + X2 X. Using the properties of the metric p constructed from
1
one can now prove, for example:
THEOREM 11. Let D(x,y) be the Rothschild-Stein parametrix for

Then, if N>3
ID(x,y)I <- C821B(x,3)I-t
and it N>2,
<CS2-]IB(x,5)I-1
J
where S = p(x,y).
Thus we get estimates analogous to those for the Newtonian potential

and Kohn Laplacian discussed in part I. These estimates have also been
obtained by Sanchez-Calle [15a]. One also gets estimates of this type for
the Rothschild-Stein parametrix of the operator X1+X2+ +X,, in
analogy with the estimates in part I for the heat operator. For details of
the argument. see [13].
8. Nonisotropic metrics on domains of finite type

We now apply the theory of metrics to boundaries of domains in Cn+t
(Some of these results were announced in [12J.) Thus let p c C(Cn+l),
with dp 4 0 when p = 0 and let
it = Iz cCn+llp(z)<01
so that fl is a domain with smooth boundary. If we fix a point Co c ait,

then near Co we can find n linearly independent tangential holomorphic
vector fields Ln. For example, if ' (Co) 0, we can take
n+1
Lj (9P a (9P a
Nn+I dzj - Nj
We write Lj = 2 (Xj-iXn+j), 1 < j < n, so that X1, ,X2n are C

real vector fields defined near Co on aft . Since the real dimension of
sit is 2n + 1 , near Co we can find an additional real tangential vector
field T, so that is a basis for the real tangent space to
ap
aft near Co. Again if az
(CO) 0, we can let
n+1
TaP
[fin+i
a
Wzn+l
_ dp a
()n+l Wn+1
The notion of type of a point was introduced by Kohn [8]. If C c aft

is near Co, C is of type m if every commutator of of
length at most m - 1 at i lies in the span of while
some commutator of length m does not lie in this span, and hence has
nonzero T component. One easily checks that the type of a point does
not depend on the particular choice of the vector fields and T,
but is really a biholomorphic invariant.
We say that the domain it is of finite type m if every point c sit
is of type < m. In this case of course, the vector fields
viewed as being defined on an open set in R2n+1 , are of finite type as
defined earlier. In particular, we then have a metric defined on sit, and
various equivalent descriptions of the associated family of balls.
284 ALEXANDER NAGEL
Our first object is to show that we can define an equivalent family of

balls on ail in terms of a single exponential map using the vector fields
For each finite sequence of integers with
1 < ij < 2n, we can write the commutator
Xil'...'ik = [Xik, [Xik_I,..., [Xi2'XiII...11
Ai1'ik T (mod X1''**,x 2n)
where Ai,,---,ik c C(il) (near CO), since is a basis for

the tangent spaces near Co. Let 5k be the ideal in C(l) generated
by the functions with f < k. It is easy to check that this
ideal does not depend on the choice of the vector fields or T.
Also
(A11'...,ik) - Aii...... c 9k
Xrk+i k+1
so in particular, Xjk+1 (Ai1' ik ) E 9k+i One can verify this claim by
noting that X.li,...,rk+i - A. T, and also Xli'...,1k+1 _

r1' 'in+1
[Xrk+l,XiI,,,,,ik1 and then expanding this commutator.
DEFINITION. For Cc ail near CO, set
Ak(0 =I IAil,...'if(S )I
and
m
A(C.8) = Aj(t)8j
j=2 2
where the first sum is over all the generators of 9k. Note that Ak(O is
a function whose size measures how much T component the commutators
of XI,---'X2, of length < k can have. In particular, since c3il is of
type m, then Am(0 > 0 > 0.
We are now in a position to define balls in terms of a single exponen-

tial map. For E Al set
2n
77co-QIrl=exp ajXj+yT (c), where
1
IajJ<S, 1<j<2n, and byl<A(C,&) .
Thus B(C,S) is the image of a box of size S in the "complex direc-

tions" given by and of size A(C,S) in the complementary
real direction. Our first result is then:
THEOREM 12. The "balls" B((, S) are equivalent to the balls B(C, S)
defined in part fl in terms of the vector fields i.e., there
is a constant C so that B((, S) C B(C, CS) and B((, S) C B(C,C&).
We sketch only part of the proof. Let q c B((,3), so that
n=exp tajXj+yT (0
M
It 1
with lajJ <S, Iyj <A(C,S)= A.(C)SI. Since A.(C)=F1ai1...

2
it follows that
M
y=IbjAI
2
where IbjI < SJ and Ij is a j-tuple of integers. Hence
m 2n
yT = Ibi X1. + I .6I.,EXE
2 Q=1
where 0,j,f c R and X1 is the jth order commutator defined earlier.

Hence we can join C to q with a curve 4(t) (the integral curve of
2n q
ajXj+yT) with 4'(t) _ ajYj and JajI < CSdIYi. Thus
III
r) E B(c, CS).
286 ALEXANDER NAGEL
The opposite inclusion is more technical, involving ideas used in

the proof of the equivalence of the families of balls introduced in part H
and we omit the details. A major ingredient is the use of Taylor series
expansions of functions in canonical coordinates. If f is a smooth func-
tion defined near t; , and if
g(a1,..., a2n, Y) = f (exp ( ajxj+YT) (c)
then g has a formal Taylor series at the origin, and it is given by:
00
k
g(at,...,a2n,Y)
^' I
k!IajXj+YT\\)
flo
k=0
where (lajXj+yT)k is a kth order differential operator. (See

Rothschild-Stein [15] for further details.)
To give an idea of how this formula is used, we can easily see that
the function n A(n,8) is essentially constant on B((,8); i.e. there is
a constant C so that for n e 4(C, 8),
IA(n,6)I < CIA((,S)I
A typical term in IA(71,8)I is SklAill,,,lik(n)I . We can expand
Ail ,,,,,ik(n) about C in canonical coordinates. Any term involving

applying T to this Ail, . k
is certainly of the right size since
Iyl < A(C, S). On the other hand, if we differentiate ik with one
of the vector fields this gives us an element of 9k+1 ,
so
again we get the correct estimate.
9. Balls in terms of a polarization

We now want to find a description of the balls on the boundary of a
domain Il C Cn+1 directly in terms of inequalities involving the defining
function for U. As before, let
SZ =(z ECn+llp(z)<01
where p : Cn+1 - R is C" and dp 0 when p = 0. A polarization of

p is a function R : Cn+1 x Cn+1 - C of class C" satisfying:
(a) R(z,z) = p(z)
(b) JzR(z,w) vanishes to infinite order on the diagonal z = w.
(c) R(w,z) vanishes to infinite order on the diagonal z = w.
Polarizations always exist, and are unique up to functions vanishing
to infinite order on z = w. When p is a polynomial or is real analytic,
it is easy to construct a polarization. If
P(z) = I as zap
then
R(z,w) = S` as za W13
is a polarization which is holomorphic in z and antiholomorphic in w ,

and in all that follows, we shall be dealing with polynomial p.
We now define a new family of balls on the boundary. For w e d 'Q
and S > 0 set
B'(w,8) = 1z CaQ 11z-wI <8 jR(z,w)j <A(w,8)1 .
Note that for w e dg, Vw = 1z ECn+1IR(z,w)=01 is a holomorphic hyper-

surface tangent to c3i at w . Thus B#(w,8) is essentially the set of
points z e 31 within Euclidean distance 8 of w, and within Euclidean
distance A(w, 8) of Vw.
Let us quickly calculate what all this means for the Siegel upper half
space discussed in part I, where p(z) = n Imzn+1 .

Then
n
R(z,w)=I zjwj-Zi(zn+l-wn+l).
j=1
288 ALEXANDER NAGEL
Also, on the Heisenberg group [Xj, Xn+jI _ -4T so A(w, S) 32 for all
w E M. Thus in this case
1 n
B#(w,S) = z -F ul Iz-wI <6,
L
14
j=1
zjwj 2i (zn+1 -wn+1) <32
and it is easy to check that this defines essentially the same balls as we
did earlier in part I.
We now want to show that the balls B'a(w,S) are equivalent to the
balls B(w,S) defined in terms of the exponential mapping. For simplicity
we will do this only in the special case
a = SZ0 = 1(z1,z2)EC2IIm z2>'b(z1)'
where 0: C R is a polynomial of degree m. Here the defining

function is
P(zl'z2) = b(zl) - 2i (z2-z2) .
It is easy to check that if
L _ a - 2i a", (zl) a
(9f1 azl az2
then L globally spans the space of tangential antiholomorphic vector

fields. We also set
T= a2+ a
02
We want to investigate the geometry of the balls near a fixed boundary
point, say (0,ij(0)). Since all our families of balls are invariant under
biholomorphic mappings, we can choose a special coordinate system to
study the geometry.
Let
O(z)_Cz)-MO)-2Re( (0)zl}
If we make the change of variables:
zl=zl
m j
z2 = z2 - i c (0)+2 1 i (0)zi
j=1 dzi
1
our original domain now has the form
1(zi,z2) f C211m z2 > 0(z, )I
and moreover 0 now satisfies
j
a-o
00 1<j<m
chi 00
We shall work in this coordinate system near (0,0) f a g.

We identify 311 with C xl{ so that (z, t+iq (x)) corresponds to (z,t).
Then our basic vector fields are:
L= +i'(z)N, 5j (97
(Z) Jt
T=?.
We can calculate the various functions XiI", ik using the vector fields L
and L (instead of Re L, Im L ). Thus:
[L,L]=-2i'!
2
so the ideal 92 is generated by O or Ac . Thus
aza
A2(z,t) = IoO(z)I
290 ALEXANDER NAGEL
alo
Ak(z,t) = I
a+,B<k aza&-#
a,13>1
To investigate the exponential map exp (aL+uL+yT) (0) with a e C,

y e R, we must find a curve
[0,1] CxR
with
q (t)=a
G
12 (t) = y + is iff (01 (t))
with (1(0),2(0)) = (0,0). Thus 1(t) = at and
q2(t) = yt + i f (a (as) a (as)) ds.

J0
CV 0
And so
exp (aL+aL+yT) (0) = a, y+i (as) - a (as)\ dsl
We can estimate the integral by expanding a and about 0. Since

aj
(0) = 0 1 < j < m, all the terms will be dominated by the correspond-
Jz-j
ing Aj(0), so
f (aas)
0
_a (as) ds I2A2(0) +...+ JalmAm(0)
= A(O, lab
In particular, we see that the image under the exponential map of the box
I(a,y) e C xRI IQI < s, Iy) < A(O,6)1
is essentially
{(z,t) a CxRI Izl < s, Itl < A(0,8)1.
On the other hand, the polarization of our defining function is
R((zIz2)(WIw2)) _ !1Sia_zQ+a5_ (0) ziw - 2i (z2 W2)

a'
and
R((zlz2), (0,0)) = -2i1 z2 = -1

2i (t+iV'(z1))
if (zt,z2) = (zI, t+ici(z1)) c aft. Thus the inequalities
IR((zIz2). (0,0))1 < A(0,8), Iz1I <a
are equivalent to the inequalities
ItI<A(o,a), IzII<8
and so the families of balls are equivalent.
10. Estimating pleurisubbarmonic functions

We continue our study of domains of the form
"= 1(z1,z2) EC2IIm z2 > q(zi)1 ,
but we now make the additional hypothesis that L4 > 0; i.e. di is a

subharmonic polynomial of degree m, which is not harmonic. Our object
is to obtain an analogue of Theorem 8 in part I. We will prove:
292 ALEXANDER NAGEL
THEOREM. Suppose u is continuous on G and pleurisubharmonic on

fl . Let n : iZ Al be the projection onto the boundary. Then if
(z 1,z2) a fl
lu(zl,z2)I <CIBI-'
f lu(C)Ida(()
B
where B = B(n(z1 z2),AS) and A(n(zi,z2),S) = Im z2-c(zl). Here A,C

are constants independent of u and (zi,z2).
As before, by making a holomorphic change of variables, we can

assume Co) _ (0) _ (0) = 0, 1 < j < m, and it is enough to
i
estimate u at the point (0, iy). We begin as we did in part I:
4 u(0,s+it)I ds dt .
lu(0,iy)I < (8)
7ry2
I s 12+l t-y 12<(v /2 )2
Now for each s + it we want to imbed an analytic disc into fl, whose
boundary is mapped to A) and whose center is mapped to (0,s+it). To
do this, we try to find S > 0 and a function G(4), continuous for
141 < 1 , holomorphic on 1 , so that the holomorphic map
(84,G(4))
has the required properties. This means we want
Im G(e'0) _,(Sei0) 0 < 0 < 2n
and
G(0)=s+it.
Let 0&(ei0) _ O(Sei0). Let P[0S] be the Poisson integral of (kS , and
let Q[958] be the conjugate Poisson integral, with Q[0,51 (0) = 0. Then
G(() = Gs t(o = s + i[P[gS](o + Q1081 (01
has the required properties provided that S is chosen so that
t
2n f .0(8 ei0)d9 . (9)
But by Green's theorem
2n
dS 2 f 0(S ei0)de =
2 775
r AO(x,y) dx dy > 0
0 X2+Y2<5 2
fo21r
so the function t(S) = Zn j(S ei0)d0 is a monotone increasing func-
tion of S . Thus given t, there is a unique S = S(t) so that equation (9)
holds. We thus obtain
3Y /2
Iu(0),iy)1 < na 22
J Iu(s(t)e
Y
0
-Y /2 Y12
s+ dt dOds
We want to make the change of variables r = S(t). Let
A(S) = 1
nS2
ff Ad!(x,y)dxdy
x2+y2<52
so that t'(S) = SA(S). We get:

2
294 ALEXANDER NAGEL
Y/2 2s 32
u(O,iy)< Lff
77Y
2
J
-Y/2 0
S(I)
2
We now need the following result, which is where the hypothesis 0

is used:
LEMMA. There are constants C1, C2 which depend only on m so that

for S>0,
(i) 0 < C1 < A(0,S)/t(S) < C2 < 00
(ii) 0 < C1 < A(0, 8)/82A (8) < C2 < 00
(iii) IWS(eie)I < C2t(S)
We defer the proof for a moment, and return to the estimate for u(0,iy).
In our last integral, when r is between S 2 and S , t(r) is
between and 2 . 2
It follows from the lemma that in this range,
2
A(r) ti y/6(y)2 , and it follows from part (iii) of the lemma that the range
of s integration in the integral is contained in { Is I < Cy I for some
constant C. Thus:
VT
CS(y)
Iu(0,iy)I < C
yS(y)2 ff
IsI<Cy 0
f Iu(rei0,s+ic(reie)Irdrdeds
0
But finally, rdrdOds is essentially surface area measure, and on the

surface r7lZ we are integrating over a region centered at (0,0) of size y
in the "real" s direction, and of size S(y) in the "complex" directions.
But this is exactly our nonisotropic ball B(0,5) where A(0, S) = y and
since IB(0,8)I .ti yS(y)2, we have shown: there are constants A and C
so that
Iu(0,iy)I < C`B(0,A8)j-I J ju(J)jda(()

B(0,AS)
where A(0,5) = y . Thus we have managed to estimate u along the

"radius" from (0,0). Just as in part 1, we now easily extend this result
to obtain th'e theorem.
We finally turn to the proof of the lemma. The inequalities t(S) <
CA(0, S) and S2A(S) < CA(0, S) follow easily by expanding c or Aq
in a Taylor series about 0. The main content of the lemma is contained
in the opposite inequalities, and follows from a homogeneity and compact-
ness argument.
We claim for every integer m > 2 there is a constant Am so that if
J
q(z) is a real polynomial of degree at most m with a (0) = 0,
az
0 < j < m , and if AO(z)>O for JzI <8 D then for S < S0
t(S)=2n
('
J0
2n
L
0(Sei0)d9>AmL r as+(30
azar}i
(0) I Sa+R = AmA(0, a)
2n S
f
I
0
f 0
Ab(rei0)rdrde> Am
llazaaza
(0) Sa+1i = AmA(0, S) .
In fact, if we can prove this for S = So = 1 , then given c , we apply the

result to t#(z) = q(Sz), and the result for general S follows, so it
suffices to study S = So = 1 . But now we let
b(z) real polynomials of degree < m such that
J
j (0)=0, 0<j<m; 66(b(z)>0 if ,zj<1 ;
and
I as
+(3
o (0)
1(gzaaZo
1
296 ALEXANDER NAGEL
It is easy to see that I is a compact set of polynomials and that the

maps
0 fe10o
0
1 71 L7T r (reie) rdr d9

J0
are continuous on 1. Moreover, these functions are strictly positive

since AO > 0 and AO' 0, on JzJ < 1 , and so the functions are bounded
below by a constant Am > 0. The general result now follows by dividing
a general polynomial 4 by I laza(90 ( 0)1 .
Finally, in order to show
sup IQ.0g(ei9)J < Ct(S)
when 0 is a polynomial of degree < m, it again suffices to check this

for S = 1 . But
sup IQ.08(e1O)I = 111.0111
0
is a norm on this space of polynomials. Hence
III(bPHH < CA(0,1) < Ct(1).
As an easy corollary of the theorem we obtain an estimate for the

Szego kernel S(z,C) for the domain 1 on the diagonal z = C. Recall
that the orthogonal projection S of L2(af) onto H2(1Z) is called the
Szego projection, and formally, this projection is given by integrating
against a kernel:
Sf(z) = ff()S(zs)da() , z e iZ .
0 il
In fact S(z, J) = E 0j(z)0j(C) where (0j! is a complete orthonormal

basis for H2(H), and this series converges uniformly on compact sub-
sets of SZ x Q. (See Krantz [l0a], Chapter 1, for further details.) Now it
is easy to check that for z f SZ
S(z,z)1 12 = sup IF(z)I
where the supremum is taken over all F e H2(1) with IIFII 2 < 1 . But
H
by our theorem, if F E H2(SZ),
IF(z)I < CIBI-1 r IF( )I da(C)

B
1 /2
< CIBI-1 /2 r IF( )I2 da(C)
aJiZ
= CIBI-1 /2IIFIIH2
where B = B(n(z),S) is the ball centered at the projection n(z) of z ,

and A(n(z), S) = - p(z) . Thus we obtain:
COROLLARY. If z E SZ
S(z,z) < CIB(n(z),8)I-1
where A(n(z),S) = Im z2 - O(z1).
11. Estimates for the Szego kernel on aQ

As a final application, we show how one can make estimates of the
Szego kernel S(z,C) on the boundary, at least for certain very special
domains Q. Thus, let
f' ='(zt,z2)EC2IIm z2>q(zi){
298 ALEXANDER NAGEL
where 95 is a subharmonic, non-harmonic polynomial of degree m. We

also make the very restrictive assumption that
e95(z) = AO(x+iy)
is actually independent of y .
Our approach is the following: if
L= -2ic3z
1 1
(z )a 2
then L is a global tangential antiholomorphic vector field, and
H2((qf2) =If fL2(a[l)jLf=0 as distribution) ,
so we identify H2 with the kernel of the differential operator. When we

identify dfl with C x R in the usual way, and write z = x + iy , this
operator becomes:
L
07 Tt
2[aao 49
+2 (I-
We now make a change of variables on C x R
D(x,y,t) = (x,y,t-A(x,y))
where
A(x,y) _ - f (t,y)dt .
0
Then if we put
x
b'(x) = f A95(t)dt
0
and
L = +i r + b'(x) a
it is easy to check that L(f ocb) = 2(Lf) o (D.

Thus our object is to obtain estimates for the orthogonal projection
onto the kerpel of
L=
X +i +b'(x)
10i
11 it]
11
where b"(x) > 0. Our finite type hypothesis is now E IbW)(x)l > go > 0,
i=2
and we want to obtain estimates in terms of balls defined by the vector
fields and 22- + b'(x) in R3.
dly -t d
If u = u(x,y,t) is a reasonable function we define partial Fourier

transforms by
`.fu = u(x,rl,r) = f e
R2
so that
u(x,y,t) = ff e21ri(Yri+tr)u(x,i7,r)di7dr
R2
is an isometry of L2(R3) and
Luf-1ifu
where
Lu . e2"(nx+b(x)r) dx (e 2rr(rix+b(xp)u)
Let
q,(x,rr,r) = e 2rr(nx+b(x)r)
300 ALEXANDER NAGEL
and let
M.g(x,rl,r) = O(x,r1,r)g(x,71,r) .
Then
Lu=f 0- 1-iA,fu.
dx
Now
Mq,: L2(R3, dx di7dr) - L2(R3, e4n(r/x+rb(x))dx d77 d,)
and
MO-1 : L2(R3, e477(71x+rb(x)) dx drldr) -+ L2(R3, dx dr/dr)
are isometries. Thus L is similar to the operator N acting on func-

tions which satisfy
f g(x,r1,r)12e4r7(t7x+rb(x))dxdr7dr
< +- .
The kernel of this operator thus consists of functions g(77, r) so that
fJ'g(iir)I2[j'e4T'dx] drldr < 00 .
Let
Y = (r),r)ER2l fe477(?7x+rb(x))dx < o
Then since b"(x)> 0, 1 = {(71,r)lr<0). For (>),r) a 2, the space

L2(e4rr(rlx+rb(x))dx) contains the constants. Let P. be the projection

of L2(e4n(r)x+rb(x))dx) onto the constants, if (,l,r) a E, and let P'7,7 =0
otherwise. Let
P: L2(e4n(rlx+rb(x))dxdrldr) -, L2(e4n(rlx+rb(x))dxdrldr)
be defined by
Pg(x,rl,r) =
where g,7 r(x) = g(x,rl,r). It is then easy to check that P is an orthogonal

projection whose range is precisely the null space of dx on
L2(e4n(rlx+rb (x)) dx drl dr) . Thus if we set
P .` -IM PMtA
then P is the orthogonal projection of L2(dxdydt) onto the null space

of L.
Now if (rl, r) e
<g,1>1
Pr1,r g = < 1,1 >
00
r g(r)e4n(rlr+rb(r))dr
00
00
r e41r(Tlr+rb(r))dr
Thus
00
P77,rg(x) = J g(y) It 7,r(Y)dy

302 ALEXANDER NAGEL
where
Kll,r(x,Y) =
Thus if f f L2(R3, dx dy dt)
Pf(x,y,t) If f(r,s,u)S((x,y,t); (r,s,u))drds dy
where
S((x,Y,t); (r,s,u)) = ff
K,?,r(x,r) dry dr
00 00
e21rq((x+r) + i(y-s))
fe 27rrt(b(x)+b(r))+i(t-u)]
foe 00
di dr.
0
f
-00
e41r(nr-rb (r)) dr
This is the kernel we have to estimate.

We begin by estimating the inner integral. For r > 0, set
00
e21r-q(,k+it)
F A+it r d
-00 r e4n[r7r-rb(r)]dr
-00
Then replacing r by r + 2 and 77 by r) + rb'(2) it follows that
F(A+it, r) _
I_
2rrr 2b +itb' (`]
/1
00
f+ e2ni77td77
e (2)
-00
J
00
2(r+)+rbe
47rl77r+rrb ' (-rb
LLL
2(2)1J
-00
Let G(r) = rrb'(2) - rb (r+2) + rb(2) . Then
G'(r) = rb'C2) - rb'(r+2)
G"(r)= -rb"(r+2)
Since G(0) = G'(0) = 0, we have
m
G(r) r 1 b()) 1
j=2
( -) r)
Hence
00
[2b(3) + itb'(2)1 e2rrjr7td,7
F(a+it,r) = e
277r
L JJ f
J
-00
00
4n nr-r
m f
1 b (j) ( ) r7
Now choose p = (A, r) so that
I 1 b(j)(a>rj12
In
2 =1
j=2 1!
and in the last integral, make the change of variables r - gr, 77 77.
304 ALEXANDER NAGEL
Then
e2ar [2b() + itb "%] e2ai7,(

F(a+it, r) = rl
A-2 M
477 r)r- .j
e dr
00
-00
where ai = 1 j 2
and hence Ym
2
Ia.2 = 1 .
7
We now make two observations. First, in terms of size,
m
(X, r) 1 ti Y Ib(j) ( l /i
z
\2/ I rl
This is clear from the definition of . Second, the collection of functions
f4a r - a i
6a (n) e m
2 dr
m
where a = am), Y-IajI2 = 1 , and r 'I' air) is convex, is a com-
2
pact set of functions in the Schwartz class S(R). Thus
e2rrr L2b rj'l + itb (2/] t1o, r)-1 .

F(A+it, r) = \/
ea 1
From this, one can make estimates on the size of F(A+it,r) and its
derivatives.
Finally we have
00
e-2ar[b(x)+b(r)+i(t-u)]F(x+r+i(y-s),r)dr
S((x,y,t); (r,s,u)) 5
0
and we can use the estimates on F to estimate S. A consequence is,

for example:
IS((x,y,t); (r,s,u))l <
CIB((x,y,t),6))-I
where S is the nonisotropic distance between (x,y,t) and (r,s,u).
ALEXANDER NAGEL
DEPARTI4IENT OF MATHEMATICS
UNIVERSITY OF WISCONSIN
MADISON, WISCONSIN
REFERENCES
[1] Bets, L., John, F., and Schechter, M., Partial Differential Equations,
Interscience Publishers, John Wiley and Sons, Inc., New York 1964.
[2] Caratheodory, C., "Untersuchungen fiber die Grundlagen der
Thermodynamik," Math. Ann. 67 (1909), 355-386.
[3] Coifman, R. R., and Weiss, G., Analyse harmonique non-commutative
sur certain espaces homojenes, Lecture Notes in Math. #242,
Springer-Verlag, 1971.
[4] Fatou, P., "Series trigonometriques et series de Taylor," Acta Math.
30 (1906), 335-400.
[4a] Fefferman, C., and Phong, D. H., "Subelliptic eigenvalue problems"
in Proceedings of the Conference on Harmonic Analysis in Honor of
Antoni Zygmund, 590-606, Wadsworth Math. Series, 1981.
[5] Folland, G. B., Introduction to Partial Differential Equations,
Mathematical Notes Series, #17, Princeton University Press,
Princeton, N. J. 1976.
[5a] Folland, G., and Hung, H. T., "Non-isotropic Lipschitz spaces" in
Harmonic Analysis in Euclidean Spaces, Part 2, 391-394; Amer. Math.
Soc., Providence, 1979.
[6] Folland, G. B., and Stein, E. M., "Estimates for the complex and
analysis on the Heisenberg group," Comm. Pure Appl.Math. 27 (1974),
429-522.
[6a] Grushin, V. V., "On a class of hypoelliptic pseudo-differential opera-
tors degenerate on a sub-manifold," Math. USSR Sbornik 13(1971),
155-185.
[7] H6rmander, L., "Hypoelliptic second order differential equations,"
Acta Math. 119 (1967), 147-171.
[8] Kohn, J. J., "Boundary behavior of d on weakly pseudoconvex
manifolds of dimension two," J. Diff. Geom. 6 (1972), 523-542.
306 ALEXANDER NAGEL
[8a] Kohn, J. J., "Boundaries of Complex Manifolds," in Proceedings of

the Conference on Complex Analysis, Minneapolis, 1964; 81-94;
Springer-Verlag, New York, 1965.
[9] Koranyi, A., "Harmonic functions on Hermetian hyperbolic space,"
Trans. Am. Math. Soc. 135 (1969), 507-516.
[10] Koranyi, A., and Vagi, S., "Singular integrals in homogeneous
spaces and some problems of classical analysis," Ann. Scuola Norm.
Sup. Pisa 25 (1971), 575-648.
[10a] Krantz, S., Function theory of several complex variables, John Wiley
and Sons, New York, 1982.
[11] Nagel, A., and Stein, E.M., Lectures on Pseudo-differential Opera-
tors, Mathematical Notes Series, #24, Princeton University Press,
Princeton, N.J. 1979.
[12] Nagel, A., Stein, E. M., and Wainger, S., "Boundary behavior of
functions holomorphic in domains of finite type," Proc. Natl. Acad.
Sci. USA, 78 (1981), 6596-6599.
[13] "Balls and metrics defined by vector fields I: Basic
properties" to appear in Acta Math.
[14] Riviere, N., "Singular integrals and multiplier operators," Ark. for
Mat. 9(1971), 243-278.
[15] Rothschild, L. P., and Stein, E. M., "Hypoelliptic differential opera-
tors and nilpotent groups," Acta Math. 137 (1976), 247-320.
[15a]Sanchez-Calle, A., "Fundamental solutions and geometry of the sum
of squares of vector fields," Inventiones Math. 78, 143-160 (1984).
[16] Stein, E. M., Singular Integrals and Differentiability Properties of
Functions, Princeton University Press, Princeton, N.J. 1970.
[17] , Boundary behavior of holomorphic functions of several
complex variables, Mathematical Notes Series, #11, Princeton Univ.
Press, Princeton, N. J. 1972.
OSCILLATORY INTEGRALS IN FOURIER ANALYSIS
E. M. Stein
Introduction
Oscillatory integrals in one form or another have been an essential
part of harmonic analysis from the very beginnings of that subject.
Besides the obvious fact that the Fourier transform is itself an oscillatory
integral par excellence, one needs only bear in mind the occurrence of
Bessel functions in the original work of Fourier (1822), the study of
asymptotics related to such functions in the early works of Airy (1838),
Stokes (1850), and Lipschitz (1859), Riemann's use in 1854* of the
method of "stationary phase" in finding the asymptotics of certain
Fourier transforms, and the application of all these ideas to number theory,
initiated in the first quarter of our century by Voronoi (1904), Hardy (1915),
van der Corput (1922) and others.
Given this long history it is an interesting fact that only relatively
recently (1967) did one realize the possibility of restriction theorems for
the Fourier transform, and that the relation of the above asymptotics to
differentiation theory had to wait another ten years to come to light!
The purpose of these lectures is to survey part of this theory and at
the same time to describe some new results. We have found it convenient
to divide our discussion into oscillatory integrals of the "first kind," and
those of the "second kind." The main difference between the two is that
for the first kind we are studying the behavior of only one function as the
parameter increases to infinity, while for the second kind we are dealing
*In Section XIII of his paper on trigonometric series.
307
308 E. M. STEIN
with the boundedness properties of an operator which carries an oscilla-

tory factor in its kernel. However this distinction need not be taken
literally since sometimes these different types merge.
We begin by considering the more-or-less standard facts about
oscillatory integrals of the first kind, first in one dimension and then in
n dimensions. Next as a first application we deal with some estimates
of the Fourier transform of smooth surface-carried measures in Rn . This
leads us naturally to restriction theorems. (Differentiation theorems,
which are another application, are not dealt with here; but these are the
subject of Wainger's lectures [3].)
Next we discuss oscillatory integrals (of the first kind) arising in the
theory of Hilbert transform along curves and their generalizations. We
then turn to oscillatory integrals of the second kind suggested by twisted
convolution on the Heisenberg group and the theory of Radon singular
integrals. Finally we return to restriction theorems and the oscillatory
integrals of the second kind they give rise to, which operators are closely
related to Bochner-Riesz summability.*
1. Oscillatory integrals of the first kind, n = I

We are interested in the behavior for large positive X of the integral
b
1(X) = r e"(x),A(x)dx
a
where q5 is a real-valued smooth function (the "phase"), and i' is

complex-valued and smooth; often, but not always, one assumes that ,Ji
has compact support in (a,b).
*The reader will note that there are several related topics not touched on in
this survey. Chief among them is the subject of oscillatory integrals arising in
the solution of hyperbolic equations and their generalizations - the class of
"Fourier integral operators." For an elegant introduction to that subject see [1],
Chapter 4.
OSCILLATORY INTEGRALS IN FOURIER ANALYSIS 309
The basic facts about I(A) can be presented in terms of three

principles.
(a) Localization : The asymptotic behavior of I(A) is determined by

those points where q'(x) = 0, (assuming that t has compact support in
(a,b) ). More precisely,
PROPOSITION 1. Suppose 0 e C 0 (a,b), and q5'(x) 0 for x in [a,b].

Then I(A) = 0(A-N), as A ao for every N > 0.
The proof is very simple. Let D denote the differential operator

Df = dx , and let !D denote its transpose, -Df = d
Then clearly DN(er40) = erA for every N , and integration by parts

shows that
fe "'0 0 dx = fDN(e),dx = (-1)N fe"(Wtdx.
Thus clearly jI(A)I < ANA-N, and the proposition is proved.

I
(b) Scaling : Suppose we only know that dk'(x) > 1 for some fixed k,
dxk
and we wish to obtain an estimate for f b which is independent
a
of a and b. Then a simple scaling argument shows that the only possi-
ble estimate for the integral is 0(A-Ihk). That this is indeed the case
goes back to van der Corput.
PROPOSITION 2. Suppose 0 is real-valued and smooth in [a,b]. If

I0tkl(x)I > 1 , then
b
fa
ea(bi"ldx <ck0/k
holds when
310 E. M. STEIN
(i) k>2
(ii) or k = 1 , if in addition it is assumed that 0'(x) is monotonic.
Proof. Let us show (ii) first. We have
b b b b
e'4dx =f D(eiAO)dx = - r e1X4t-D(1)dx + i1 ix 4v
e
a a a a
The boundary terms are majorized by 2/A, while
b b
r ei t_D(1)dx
r eat i-j1 x (-1) dx '0,
aJ
a a
b
<1 I d 1
_ W
dx
a
by the montonicity of 0'. The last expression equals 1 Iq

(b) 0'(a)1
which is dominated by 2/A. This gives the desired conclusion with
c1 = 4.
We now prove (i) by induction on k. Let us suppose that'the case k
is known, and assume (taking complex conjugates if necessary) that
0(k+1)(x) > 1. Let x = c be the point in [a ,b] where 14,(k)(x)l takes
its minimum value. If O(k )(c) = 0, then outside the interval (c-S,c+S),
we have that I.0(k)(x)l > 5, (and of course 0'(x) is monotonic when
k = 1 ). Write
a
b
=
f
a
c-S
}f
J
c+S
r
c& +d
b
f
c+S
By the previous case
< ck/(XS)l /k
0
Similarly
e'4 dx /k
L
Clearly however
c+s
fC-S
e'kO dx <2S.
Thus
f a
e'40 dx < 2ck
(xs)l /k
+25.
If j(k)(c) 1 0, and so c is one of the end-points of [a,b], a similar

argument shows that ck/(tb)1 /k + S is an upper bound to the integral. In
either situation the case k + 1 follows by taking a = 'k-1 /k+1 , which
proves (1.1) with ck+l - 2ck+2 .
COROLLARY. Under the assumptions on -0 in Proposition 2, we can

conclude that
(1.2)
f
0
e"W(x)0(x)dx < Ck,\-1 /k
J
Jv(x)Idx .
312 E. M. STEIN
This follows from (1.1) by integrating by parts an estimate of the form
f a
eaO(x)dx <ck1-1/k, for a<x<b.
r b e'AOV
(c) Asymptotics : We already know that the behavior of Ja dx is
determined by those points x0, where 0'(xo) = 0, (the "critical" points

of (k ). Assuming that the support of i/r is so small that it contains only
one critical point of j , the character of the asymptotic expansion then
depends on the smallest k so that 0(k)(x0) 10, and is given in terms
of powers of A in a way which is consistent with Proposition 2.
PROPOSITION 3. Suppose * is real and smooth, k > 2, and

d(k-1)(x0) = 0, while 0(k)(xo) j 0. If Co and
j(x0) _,k'(x0)... _
the support of 0 is a sufficiently small neighborhood of x0, then
00
(1.3) 1(X) = f eix(b(x)q,(x)dx A-1/k I ajX-j/k ,
j=o
in the sense that for any non-negative integers N and r
N
(1.3') (1E (A) - ajA-jO(a-(N+1)/k-)
as A, 00
j=0
REMARK. When k is even, then aj = 0 for j odd.
We shall give a proof of the proposition when k = 2. There are three

steps.
Step 1. This is the observation that

00
00
(1.4) x xPe x z
eiXX2 dx - 1 /2-Q/2 ,
cj(Q) -j
J
-,o
j=0
where F. is a non-negative integer; in the case a is odd the integral

vanishes. In fact the left-side of (1.4) is j e-(I-1A)C2xedx which by
a change of variables equals (1-v1)-`/=-P12 L a "2xedx. However, when
A > 0, (1-ik)-y'4/2 where we have fixed the
principal branch of z-'/r4/2 in the plane slit along the negative half-
axis. The power series expansion of (w-i)-`/'è/2 (which holds for
lwl < 1 ), then gives the desired asymptotic expansion (1.4).
Step 2. Observe next that if -q e Co and a is a non-negative integer,

then
00
(1.5)
f CIO
eax2xe77(x)dx < AX-1/2 /2
To prove this let a be a C function with the property that a(x) = 1

for lxl < 1 , and a(x) = 0 when lxl > 2 , and write
J eikx2xe>)(x)dx = feiXx2xfrl(x)a(x/E)dx + J eikx2xeri(x)(1-a(x/E))dx .
The first integral is dominated by CEe+1 . The second integral can be

written as
5e2(tr)N [xf-q(x) [1-a(x/E)]]dx
1
with tDf = A simple computation then shows that this term
dx f .
is majorized by
CN
,\N
f lxl>E
lxle-2N dx = C' \-NEe-2N-1
N
if e - 2N < - 1 . Altogether then the integral in (1.5) is bounded by

314 E. M. STEIN
CN#E/+1 +A-NCQ-2N+1 I and we need only take e = .1-112
(with N > e 21) , to get the conclusion (1.5).

A similar (/but simpler) argument of integration by parts also shows
that
(1.6) fe2ex)dx = 0(A-N), every N > 0
whenever e c 8, and 6 vanishes near the origin.
Step 3. We prove the proposition first in the case q(x) = x2 . To do this

write
fe2r(x) dx = feiAXe x2(ex2
where is a Co function which is 1 on the support of Vi. Now for

each N, write the taylor expansion
N
eX2c (x) _ Y` bjxj + xN+IRN(x) = P(x) + XN+1RN(x)
jj=0+
Substituting in the above gives three terms
00
(2)
f 00
eiAX2 e_ C2
xi dx
00
(b)
J00
(C)
fe2P(x)e2(1_(x))dx.
For (a) we use (1.4); for (b) we use (1.5); and for (c) we use (1.6). It is
then easy to see that their combination gives the desired asymptotic ex-
eiAx2
pansion for f c/i(x)dx .

Let us now consider the general case when k = 2. We can then write
q(x) = c(x-x0)2 + 0(x-x0)3 with c 4 0 and set q(x) = c(x-x0)2[1+e(x)] ,
where a is a smooth function which is 0(x-x0), and hence 1
when x is sufficiently close to x0. Moreover, q'(x) 0, when x x0

but x lies sufficiently close to x0. Let us now fix such a neighborhood
of x0, and let y = (x-x0)(1+e(x))I12. Then the mapping x y is a
diffeomorphism of that neighborhood of x0 to a neighborhood of y = 0,
and of course cy2 = 4(x). Thus
fei'(o)clr(x)dx = reilcy2 (y)dy
with te Co if the support of 0 lies in our fixed neighborhood of x0.

The expansion (1.3) (for k=2 ), is then proved as a consequence of the
special case treated before.
REMARKS:
(1) The proof for higher k is similar and is based on the fact that
f 0
eixxke xkxedx = ck,Q(1- i'\)-R+I)/k .
(2) Each constant aj that appears in the asymptotic expansion (1.3)

depends on only finitely many derivatives of and Vi at x0. Note
e.g. that when k = 2, we have a0 = \r1-T(-i0'(x0))-I"2Vi(x0). Similarly
the bounds occurring in (1.3') depend on upper bounds of finitely many
derivatives of 0 and Vi in the support of sli , the size of the support of
t,, and a lower bound for 0(k)(x0).
316 E. M. STEIN
References : The reader may consult Erde1yi [8], Chapter II, where further
citations of the classical literature may be found.
2. Oscillatory integrals of the first kind, n > 2

Only some of the above results have analogues when n > 2, but the
extension of Proposition 1 is simple. Continuing a terminology used above
we say that a phase function (k defined in a neighborhood of a point xa
in R has x0 as a critical point, if (Vg)(xo) = 0.
PROPOSITION 4. Suppose cu c Co(Rn), and 0 is a smooth real-valued

function which has no critical points in the support of Vi. Then
I(A) = r O(A-N), as A for every N > 0 .
Rn
Proof. For each xo in the support of ci, , there is a unit vector 6 and a
small ball B(x0), centered at x0, so that (e,vx)o(x)> c > 0, for
x c B(xp). Decompose the integral fek'0(x)t4(x)dx as a finite sum
I fe(')clFk(x)dx
n
where each c/ik is C and has compact support in one of these balls. It
then suffices to prove the corresponding estimate for each of these
integrals. Now choose a coordinate system xt,x2, ,xn so that xI
lies along 6. Then
=f(fe"(x
'...,xn)Vk(xt,...,xn)dxI
fe('c)ci/k(x)dx I dx2,...,dxn .
But the inner integral is 0(A-N) by Proposition 1, and so our desired

conclusion follows.
We can only state a weak analogue for the scaling principle, Proposi-
tion 2; it, however, will be useful in what follows.
PROPOSITION S. Suppose 0 t Co , q is real-valued, and for some

multi-index a, at > 0,
(d)a
(x),>1
throughout the support of t# . Then
(2.1) f eikO(x)tp(x)dx < Ck('A)' A-I /k (IIqj .L00+ II174IIL1)

Rn
with k = ial , and the constant ck() is independent of A and t/i and
remains bounded as long as the Ck+1 norm of 0 remains bounded.
Proof. Consider the real linear space of homogeneous polynomials of

degree k in Rn. Let d(k,n) denote its dimension. Of course
{xa}1a`=k is a basis for this space. However it is not difficult to see that
there are d(k,n) unit vectors 1` (1), (2), , e (d(k,n)) so that the
homogeneous polynomials ( (J) x)k , j - 1, , d(k,n) , give another basis.
This means that if
0(x0I
axa >
for some lal = k , there is a unit vector e = ;(x), so that
I(e, px)kO(xO)l > ak, with ak> 0 .
Moreover since we can assume that the Ck+1 norm of is bounded we

can also conclude that I(e, p,)(k(x)l > ak/2 whenever x c B(xo), where
B is the ball centered at x of fixed radius. (The radius of B can be
taken to be a small multiple of the Ck+1 norm of (b.) Next choose an
318 E. M. STEIN
appropriate covering of Rn by such balls of fixed radius, and a corre-

sponding partition of unity, 1 = F 173(x), with 0 < ijj < 1 , {pnj{ < bk ,
and each rlj supported in one of our balls. So k
fe'Acj'dx= feuicci7j dx feuic lij dx.
To estimate Jei-k'ipj dx , with a determined as above, choose a

coordinate system so that x1 lies along C. Then
fe'1c1rjdx xn)dx1) dx2,...,dxn

=J W
For the inner integral we invoke (1.2) giving us an estimate of the form
-1 /kA-1 / a f' f
ckak (x1,...,xn) dx1
L
J I1 (x1,...,xn) dx1
I
A final integration in the other variables then leads to (2.1).
REMARK. Let us note that in R2 if q(x) = x 1 x2 , the above proposition

gives no better than a decrease of order 'k-1 /2 , while the asymptotics of
the proposition below shows that the true order is 0 .
Let us go back for the moment to the case of one dimension. If 95
has a critical point at x0, and q' does not vanish of infinite order at
x0, then after a smooth change of variables (b can be transformed to a
simple canonical form , with '(x) = xk (for x near 0 ). There is
no analogue of this in higher dimensions, except for k = 1 and in a
special case corresponding to k = 2. To the asymptotics of the latter
situation we now turn.
Suppose 4 has a critical point at xO. If the symmetric n x n matrix

is invertible, then the critical point is said to be non-
r7x
(((
degenerate. It is an easy matter to see by the use of Taylor's expansion

that if xO is a non-degenerate critical point, then in fact it is an isolated
critical point.
PROPOSITION 6. Suppose q(xo) = 0, and (A has a non-generate

critical point at x0. If i/r a Co and the support of q1 is a sufficiently
small neighborhood of x, then
(2.2)
e"(x)0(x)dx A-n/2 ai X-j , as A
j=o
Rn
where the asymptotics hold in the same sense as (1.3), (1S).
Note. Again each of the constants aj appearing in the asymptotic ex-

pansion depends on only finite many values of derivatives of (A and 0
/
at x0. Thus e.g. ao = (an/2 R
n
(-ij)-1/21 O(x0), where
i=I
a2(A(x0)
1A1' 2' "' n are the eigenvalues of the matrix 2 axk
1 Similarly
each of the bounds occurring in the error terms depend only on upper
bounds for finitely many derivatives of 0 and 0 in the support of Vi ,
a2(b(xo)
the size of the support of and a lower bound for det
xk
The proof of the proposition follows closely the same pattern as that
of Proposition 3. First, let Q(x) denote the unit quadratic form given by
Q(x) xn, where 0 < m < n , with m fixed.
The analogue of (1.4) is
00
(2.3) r eikQ(x)e Ix12xedx ti A-n/2-JeJ/2 5" c)(m,e)),-i

s% j=0
Rn
320 E. M. STEIN
with P = a multi-index, Pl = and

El P
xP = xn ; also note that if one Pi is odd then (2.3) is identically
zero. To prove (2.3) write it as a product
n
eax2x_x2
J Pdx J
(' e 2
P
x J dx (1
(f j=1
-oo
n -lh-P./2
and expand the function 11 (1/X+i) (for large A ) in a power
j=1
series in 1/X .
The analogue of (1.5) is the statement that
(2.4)
f
Rn
ei,\Q(x)xPrl(x) dx < AX-n/2-IP!/2;
if r, a C0(Rn)
To prove it we consider the two-sided cones T i defined by

Ci = { s Ixil2 > Ix12 y, and the smaller cases hI _ {xi Ix)I2 > Ix121 .
tsince 2n n
Then
j=1i 1
J
= Rn we can find functions lZ ,
J
i = 1, , n, each
homogeneous of degree 0, and Co away from the origin, so that
n
1= fI.(x), x = 0, with ft. supported in F.. Then we can write
j=1
fe V Q(x)xPn(x)dx = fe'AQ(x)xP,r(x)SZ(x)dx
i
In the cone C'i one uses integration by parts via
pie"\Q(x) = eixQ(x) with Di(f) _- I Of

NT T 1 7
This, together with the fact jxii > 1 Ixl in F., and I (tD.)Nj .(x)1
J2n
CN Ixl-2N allows one to conclude the proof of (2.4) in analogy with that
of (1.5).
A similar argument also show that whenever e 8 and e vanishes
near the origin, then
(2.5) 1 e 4Q(x)e(x)dx = O(A-N), for every N > 0 .
We then combine (2.3), (2.4) and (2.5) as before to obtain the

asymptotic formula (2.2) in the special case when c (x) = Q(x).
To pass to the general case one is then fortunate to be able to appeal
to the change of variables guaranteed by Morse's lemma: Since cb(xo) = 0,
(V )(x) = 0, and the critical point is assumed to be non-degenerate,
there exists a diffeomorphism of a small neighborhood of x0 in the
x-space to the y-space, under which cb is transformed ym -
2 Y2.
ym+1 yn' Observe that the index m is the same as that of the
form corresponding to 1 .92`b(xO)
2 oPx j(?xk -
References. For a proof of Morse's lemma see Milnor [201, 2.
D. Fourier transforms of surface-carried measures

Let S denote a smooth m-dimensional sub-manifold of Rn (not
necessarily closed). We let do denote the measure on S induced by the
Lebesgue measure on Rn, and fix a function 0 in Co (Rn) . Consider
now the finite Bore] measure d = O(x)do on Rn, which is of course
carried on S . The problem we wish to deal with is that of finding esti-
mates at infinity of the Fourier transform of , i.e. d(e). We shall
consider two cases of this problem.
(1) Suppose first dim S = n -1 , and S has non-zero Gaussian curvature

at each point. By this we mean the following: Let x0 be any point of S,
322 E. M. STEIN
and consider a rotation and translation of the underlying Rn so that the

point xO is moved to the origin, and the tangent plane of S at xD
becomes the hyperplane xn = 0. Then near the origin (i.e. near x0 ) the
surface S can be given as a graph xn = ,Xn_1) with e C*,
and 0(0) = 0, (%)(0) = 0. Now consider the (n-1) x (n-1) matrix
a2d?(o) Its eigenvalues are called the princi-
1
xk 1<i, k n-1
_
/
pal curvatures of S at x0, and their product (= det 2 r) is the
Gaussian curvature at x . `
THEOREM 1. Suppose S is a smooth hypersurface in Rn, with non-
vanishing Gaussian curvature at each point, and let d = tfida as above.
Then
AIfI-(n-1)/2 .
(3.1) J(d)^(()I <_
Proof. It would be convenient (in applying Proposition 6 above) to change

notation momentarily by taking n to be n + 1, Now by the compactness
of the support of tfi (and since when the surface is given as a graph
xn+1 = then da = 1+IVbl2 we can reduce
matters to showing that
(3.2)
fRn
< AA-n/2
where 45(x,'1) = x181 +x2'22 +... + xnnn + (b(x1 ?".'xn)rln+1 ,

with
n+1
iF1'1 =1, and A>0; also c'(0)=(42)(0)=0,
det (a) 0, and the support of tri is a sufficiently small neighbor-
hood of the origin.

We divide consideration of the unit vector n into three cases.
10 when n is sufficiently close to the "north pole" nN =

20 when n is sufficiently close to the "south pole" nN =
30 the complementary set on the unit sphere .
Let us consider 1 first. The function ((x, r,) has the property that
(Vx(k)(0, 71N) = 0. We want to see that for each q sufficiently close to
)IN there is a (unique) x _ x(,1) , so that (Vx) c(x(n), rl) = 0. The latter
is a series of n equations and one can find the desired solution by the
implicit function theorem, which requires that we check that the Jacobian
determinant det ((VxpxqS) (0, RN)) 4 0; but this of course is our assump-
tion of non-vanishing curvature. Notice that if the -q-neighborhood of 'IN
is sufficiently small, then det (a2(Jl

10,77
k
4 0, so we can invoke
Proposition 6 (with xe = x(rl) ), as long as the support of is suffi-

ciently small. This proves (3.2) when n is in region 1. The proof when
ri is in 2 is the same. So we now come to the region 30. Since c(x, r,)
n
then vxo(x, rl) _
4k(x)r/n+1 , (rll,..., nn) + How-
i=1 l
ever ( 7 7 2+"'+ nn)1 /2 > c > 1 , since -q is in 3 and V (x) = 0(x) as
x -, 0; thus IV I(x, q)j > c'> 0, if the support of 0- is a a sufficiently
small neighborhood of the origin. Hence for the n in region 3 we may
use Proposition 4 to conclude that the left-side of (3.2) is actually 0(11-N)
for every N . The proof of Theorem 1 is therefore concluded.
REMARK. We have used only a special consequence of the asymptotic

formula (2.2), namely the "remainder estimate" analogous to (1.3') when
N = r = 0. Had we used the full formula we can get an asymptotic expan-
sion for d(e); its main term is explicitly expressible in terms of the
Gaussian curvature at those points x c S , for which the normal is in the
direction 6 or -6.
(2) We shall now consider the problem in a wider setting. Here S will
be a smooth m-dimensional sub-manifold, with 1 < m < n-1 , and our
assumptions on the non-vanishing curvature will be replaced by the more
324 E. M. STEIN
general assumption that at each point S has at most a finite order contact
with any hyperplane. We shall call such sub-manifolds of finite type.
(These have some analogy with the finite-type domains in several complex
variables, which are also discussed in Nagel's lectures [21].) The pre-
cise definitions required for our considerations are as follows. We shall
assume that we are considering S in a sufficiently small neighborhood of
'
a given point, and then write S as the image of mapping <b: Rm -> Rn,
defined in a neighborhood U of the origin in R n. (To get a smoothly
embedded S we should also suppose that-the vectors , ,
axI ax2
, NM
are linearly independent for each x, but we shall not need that assump-
tion.) Now fix any point x0 c U C Rm , and any unit vector in Rn . 71
We shall assume that the function does not vanish of

infinite order as x --)x0. Put another way, for each x0 E U and each
unit vector 17, there is a multi-index a , with 1 < lal , so that
(3)a (fi(x) 0. Notice that if (x, ii') are sufficiently close to
(x 0, rj), then also (d)a(x.).,).I , '0. The smallest k so that for

x=x
each unit vector q then 3a , jal < k, with (O(x) 77)1 0 0 will
UJL a x=x
be called the type of (A at x0. Also if UI is a compact set in U,

the type of 0 in UI will be the least upper bound of the types for x0
in Ut .
THEOREM 2. Suppose S is a smooth m-dimensional manifold in Rn of

finite type. Let d = Vida, with t/i E Co(Rm). Then
(3.3) du(e)l<Al l_ , for some e>0,
and in fact we can take e = 1/k, where k is the type of S inside the
support of t/i .
Proof. By a suitable partition of unity we can reduce the problem to

showing that
0(IeI-I /k)
ei (x)* (x)dx =
J
R`"
ti
with (A as described above, and the support of 0- sufficiently small.
Now we can write l; = AY), with Jill = , and X > 0 . Then we know that
11
there is an a , with tat < k, so that ()t2qs(x).?1 0, whenever x is

in the support of t (once the size of the support has been chosen small
enough). Thus the conclusion (3.3) follows from (2.1) of Proposition 5.
References. Theorem 1 in its more precise form alluded to in the remark

goes back to Hlawka [14]. See also Herz [13], Littman [18], Randol [25],
and Hormander [16]. When S is a real-analytic sub-manifold not contained
in any affine hyper-plane, then it is of finite type as defined above. For
such real-analytic S estimates of the type (3.3) were proved by Bjork [2].
4. Restriction theorems for the Fourier transform

The Fourier transform of a function in Lp(Rn) , 1 < p < 2 is most
naturally thought of as an LP function (via the Hausdorff-Young Theorem)
and so at first sight it is viewed as defined only almost-everywhere. This
impression is further supported by the case p = 2 , when clearly the
Fourier transform can be completely arbitrary on any given set of zero
Lebesgue measure. It is therefore a noteworthy fact that whenever n > 2
and S is a sub-manifold of Rn (with some appropriate "curvature")
then there exists a p0 = p(S), p0 > 1 , so that every function in LP,
1 < p < p0 has a Fourier transform restricting to S (i.e. with respect to
the induced measure on S ). Let us make this precise.
Suppose that S is a given smooth sub-manifold in Rn, with da its
induced. Lebesgue measure. We shall say that the LP restriction property
holds for S , if there exists a q = q(p), so that the inequality
326 E. M. STEIN
1 /q
(4.1) If(e)lgda(e) < Ap,q(So) IIf tIp
JSo
holds for each f c 8, whenever So is an open subset of S with compact

closure in S.
THEOREM 3. Suppose S is a smooth hypersurface in Rn with non-zero

Gaussian curvature. Then the restriction property (4.1) holds for
1<p< 2n+2, (with q=2).
Proof. Suppose c > 0 and e Co C. It will suffice to prove the

inequality
(4.2) f 1 /2
o
AIIfII p
Rn
for PO = 2n + 32, and f e 5; the case 1 < p < p0 will then follow by
interpolation.* By covering the support of V1 by sufficiently many small
open sets, it will be enough to prove (4.2) when (after a suitable rotation
and translation of coordinates) the surface S can be represented (in the
support o f Vi) as a graph: en Now with d = V'da we
have that
J )?(e)I2d = J f(e)f(()d, = J T(f)(x)13dx
where (Tf) (x) _ (f * K) (x), with
K(x) = fe
*In fact the interpolation argument shows that we can take q so that (4.1)
holds with q = (n+1} p'' which is the optimal relation between p and q.
Thus (4.2) follows from Holder's inequality if we can show that
(4.3) IIT(f )Ilpo < AIIfIIpo
where po is the dual exponent to p0.

To prove (4.3) we consider the function Ks (initially defined for
Re(s) > 0 ) by
e s2 7
(4.4) Ks(x) =
1'(s,'2) fe 2nix
)I
1+s
n( W(e')de
Rn
Here we have abbreviated by e'; we have set (C') =

+Iph(e')I2)I/2 , so that (C')dC'= d; also ri is a Co(R)
function which equals near the origin.
1
Now the change of variables en -+ en + 0(') in the above integral

shows that it equals
(' e 2ni(x'-e1+xn0(e'))Yf(e')de'
s(xn) = Cs(xn)K(x)
Rn-I
with
00
2
2nix nenlfnl-l+s n(fn)den
e
Cs(xn) 1'(s/2) ./
-00
Now it is well known that 4s has an analytic continuation in s

which is an entire function; also Co = 1 ; and I4 (xn)I < clxnl-Re(s),
where Ixnl > 1 , and the real part of s remains bounded. From these
facts it follows that Ks has an analytic continuation to an entire function
s (whose values are smooth functions of x 1'..., xn of at most polynomial
growth). One can conclude as well that
328 E. M. STEIN
(a) KO(x) = K(x) ,

(b) IK-n/2+it(x)I < A, all x e R, all real t
(c) IK1+it(e)I < A, all Rn, all real t
In fact (c) is immediate from our initial definition (4.4), and (b) follows
from Theorem 1.
Now consider the analytic family Ts of operators defined by T5(f) _
f * Ks' From (b) one has
(4.5) IIT_n/2+it(f )IIL00 < AIIf1ILI , all real t
and from (c) and Plancherel's theorem one gets
(4.6) IITI+it(f )IIL2 < AIIfIIL2 , all real t ,
An application of a known convexity property of operators (see [281) then

shows that IIT0(f )II LP, < AIIfJJ PO , with PO0 = 2n + 2 , and the proof of
n+3
Theorem 3 is complete.
REMARKS:
(i) For hypersurfaces with non-zero Gaussian curvature this theorem
is the best possible, only insofar as it is of the form (4.1) with
q > 2 . If q is not required to be 2 or greater, then it may be
conjectured that a restriction theorem holds for such hypersurfaces
in the wider range 1 < p < 2n/(n+1). This is known to be true
when n = 2 (see also 7 below).
(ii) For hypersurfaces for which only k principal curvatures are non-
vanishing, Greenleaf [121 has shown that then the corresponding
results hold with 1 < p < 2k 2 , giving an extension of
+
Theorem 3.
(iii) In the case of dim(S) = 1 (i.e. in the case of a curve) there are
a series of results extending our knowledge of the case n = 2
alluded to above. For further details one should consult the
references cited below.
It would of course be of interest to know what are the exponents p

and q (if any) for which the restriction holds if we are dealing with a
given sub-manifold S. This problem is highlighted by the fact quite
general sub-manifolds S (those which are of finite type in the sense
described in 3) have the restriction property:
THEOREM 4. Suppose S is a smooth m-dimensional sub-manifold of Rn

of finite type. Then there exists a po = po(S), 1 < po, so that S has
the LP restriction property (4.1) with q = 2 , and 1 < p < po. (In fact if
the type of S is k, we can take po = 2nk/(2nk-1 .)
COROLLARY. Suppose S is real analytic and does not lie in any affine
hyperplane. Then S has the LP restriction property for 1 1 .
Proof. As we saw above, it suffices to prove (4.3). However Tf = f *K,

and K(x) = d(-x), Theorem 2 tells us that IK(x)l < AIxI-11k, So
according to the theorem of fractional integration, (see [26], Chapter V),
we therefore get (4.3) with __ = P -- where a=n-1/k, and this
n,
0 0
relation among exponents is the same as PO = 2nknk1 Q.E.D.
'
Further bibliographic remarks. The initial restriction theorem dates from

1967 but was unpublished. The sharp result for n = 2 was observed by
C. Fefferman and the author and can be found essentially in [9]; see also
Zygmund [33]. Further results are in Thomas [30], [31], Strichartz [29],
Prestini [24], Christ [4], and Drury [7].
5. Oscillatory integrals of the first kind related to singular integrals

A key oscillatory integral used in the theory of Hilbert transforms
along curves is the following:
00
(5.1) P.V. Jr elpa(t) dt

t,
330 E. M. STEIN
d
where Pa(t) is a real polynomial in t of degree d, Pa(t) = F a.O. It
j=o
was proved by Wainger and the author in [27], that the integral is bounded
with a bound depending only on the degree d and independent of the
coefficients ao,al,...,ad. The relevance of such integrals can be better
understood by consulting Wainger's lectures [32]. We shall be interested
here in giving an n=dimensional generalization of this result. We formulate
it as follows. Let K(x) be a homogeneous function of degree -n;
suppose also that IK(x)I < AIxI-n (i.e. K is bounded on the unit sphere);
moreover, we assume the usual cancellation property: f x,l-1 K(x') d a (x') = 0.
We let P(x) _ I aaxa be any real polynomial of degree d.
Ial<d
THEOREM 5:
(5.2) P.V.
f
Rn
eiP(x)K(x)dx < Ad
with the bound Ad that depends only on K and d, and not on the
coefficients as .
Nagel and Wainger observed that if K were odd, one could prove (5.2)
from the one-dimensional form (5.1) by the method of rotations (passage to
polar coordinates). To deal with the general case we need two lemmas.
d
Let Pa(t) = F a tJ denote a real polynomial on R1 , and write also
j=1 j
d
Pb(t) = F
j=1
LEMMA 1:
d
Jipa(t) 1Pb(t) dt
(5.3) (e -e )
t
<Ad 1+ log`-II,
j=1
e
with Ad independent of (aj f , I bj { and e > 0.
LEMMA 2. Let P(x) = F aaxa be a homogeneous polynomial of

Iaj=d
degree d on R'. Write
mp =f IP(x')I da(x') .
Ix k=1
Then,
(5.4)
f
Ix'I=1
Ilog (I P 1.)) da(x') < B ,
with Bd independent of P.*
One can if one wishes give an elementary (but complicated) proof of

Lemma 2. It may however be more interesting to obtain it as a conse-
quence of a general property of polynomials in Rn related to the class of
functions of bounded mean oscillation. This property can be stated as
follows. It is very well known that the function log jxj is in B.M.O., and
this is usually the first example discussed in that theory. It is surprising
therefore that the following natural generalization seems to be been
overlooked.
Of course in writing (5.4) we assume that P is not identically zero, i.e.

mp > 0.
332 E. M. STEIN
THEOREM 6. Let P(x) be any polynomial of degree < d in Rn . Then

logIP(x)I is in B.M.O. and in addition
11 log IP(x)1 IIBMO < Bd
where Ba depends only on d, and not otherwise on P.
The proofs of Theorem 6 and Lemma 2 will be given in an appendix.

We now pass to the proof of Lemma 1. We prove it by induction on d.
The case d = 1 , i.e. the estimate
00
(eiat_eibt) dt <A(1+Ilog .
5 E
t )
is classical. Let us now assume (5.3) for polynomials of degree d - 1 ,

and observe that the estimates (5.3) we wish to prove are unchanged if we
replace t by bt , 5 ;E 0. Thus we may assume that bd = 1 and Iadi < 1 .
Now write
00 00
r (e1Pa(t) - e iPb(t)) dt = 5+ r
J t J
E E 1
and we treat these two integrals separately. (If e'> 1 , we have only
f00
f"0 and that integral is estimate like .) Let us consider the second
integral. It equals
00 00
f eiPa(t) dt _ (' eiPb(t) dt
J t
f t
Now since bd = 1, we see that (d/dt)d Pb(t) = d!, and hence

(' eipb(t) dt <cd ,

t
J
1
by the corollary of Proposition 2 in 1. Next, by a change of variables

t lad-1 /d t , the integral
00
t
1
becomes
00 1
e iVa(t) dt =
f
+f
ladll/d ladll/d 1
where Pa(t) is a polynomial of degree d , with td having coefficient

one. Again
e"Va(t) dt <cd,
t
while
r dt_1
<
J t d log l(Tl=a to 1
ladl1/d
since bd = 1 . Next
f(e Pa(t)
i
a
IPb(t)
)t-J
d_ 1
(e
1Qa(t) e 'Qb(t) )dtt+ 0('dt
J t,
E E E
334 E. M. STEIN
with
d-1 d-1
Qa(t) = I ajtj ,
J=1
Qb(t) = 4 bjti
j=1
since IPa(t)-Qa(t)I < Iti and IPb(t)-Qb(t)I < Iti . However, by induction
hypothesis (using (5.3) for E' = E , and E =1, and d - 1 ),
r (elQa(t) _ e iQb(t) dt Ad-1 Ilog

1+
JE
t J=1
Gathering all these terms together then proves (5.3).

Armed with Lemmas 1 and 2 we can now prove Theorem 5. We may
assume that P(x) = F axaa has no constant term, and using polar
IaI<d d
coordinates x = tx', t > 0, Ix'I = 1, we write P(x) P-(x')ti ,
j=1 j
where Pj(x') are restrictions to the unit sphere of homogeneous poly-
nomials of degree j . Let us also set mj = f x'I-I I Pj(x')I dv(x) , and
write K(x) = t-n1Z(x'), with 1 bounded and f x'I-11Z(x')dv(x') = 0.
Then to prove (5.2) it suffices to show that
el1(x)K(x)dx <Ad,
with Ad independent of E1 , E2 and P. The above integral can be

written as
(' (fE2 1FPj(x')tj
J e SZ(x")do(x')
Ix l=1 E1
Since 0 has vanishing mean-value this integral may be rewritten as

f f
2
Ix'I=1 \ E1
However by Lemma 1 the inner integral is bounded by
d
+
Aa 1 m)
and so an appeal to Lemma 2 shows that
fE2
exp (i Pj(x')ti) - exp (i mit) dt (sup IH(x')I)da(x') < Ad

f
Ix'1=1 EI
proving (5.2) and the theorem.
6. Oscillatory integrals of the second kind: an example related to the

Heisenberg group
To motivate the interest in this example we recall the definition of the
Heisenberg group Hm. The underlying space of Hm is Cm x R, i.e.
Hm = #(z,t){, with z e Cm, t e R; the multiplication here is (z,t)(w,s) _
M
(z+w, t+s+<z,w>), where <z,w> = 2 Im II zjwj, with z = (zj), w = (wj).
1
Now on the Heisenberg group one can consider two types of dilations and
their corresponding quasi-distances. The first are the usual dilations
(z,t) (pz, pt), p > 0, and the metric could be defined in terms of the
usual distance. The second are the dilations (z,t) -+ (pz,p2t), and the
appropriate quasi-distance (from the origin) is then (Izl4+t2)1 /4. The
latter dilations and metric are closely tied with the realization of the
Heisenberg group as the boundary of the generalized upper half-space
holomorphically equivalent with the unit ball in Cn+1 . This point of view,
as well as related generalizations, is elaborated in Nagel's lectures [211.
336 E. M. STEIN
In the present context the first type of dilations and corresponding

metric would be appropriate if one considered expressions related to
ordinary potential theory in Hm viewed as R2m+1 . However the two
conflicting types of dilations (and related metrics) occur in e.g. the
solutions of Ju = f . (One sees this for example in Krantz's lectures [17],
where in the formula of Henkin we have a kernel made of products of
functions each belonging to one of the two above homogeneities.)* Other
expressions of this type occur in the explicit formulae for the solutions
of the a-Neumann problem (see [1], Chapter 7).
Let us now consider the simplest operator on the Heisenberg group
displaying simultaneously these two homogeneities. The prime example
is given by
(6.1) Tf =f*K
where convolution is with respect to the Heisenberg group, and the kernel
K is a distribution of the form
(6.2) K(z,t) = L(z)S(t) .
L(z) is a standard Calder6n-Zygmund kernel in Cm = R2m, i.e. L(pz) _

p-2m L(z), L is smooth away from the origin, and L has vanishing
mean value on the unit sphere. Here 6(t) is the Dirac delta function in
the t-variable, and in an obvious sense is homogeneous 3(pt) = p-1 S(t) .
Thus K is homogeneous at degree -2m - 1 with respect to the standard
dilations, and at the same time homogeneous of degree -2m - 2 with
respect to the other dilations; in both instances the degrees are the criti-
cal ones.
We turn next to the question of proving that the operator (6.1) is
bounded on L2(Hm). The most efficient way is to proceed via the
Fourier transform in the t-variable. This leads to the problem of showing
that the family of operators TA defined by
*In particular the terms AI and A2 that appear in 6 of [17].

(6.3) (TA)(F)(z) = i L(z-w)e1A<z,w>F(w)dw

Cm
(with <z,w> the anti-symmetric form which occurs in the multiplication

law for the Heisenberg group) is bounded on L2(Cm) to itself, uniformly
in A, -DO<OO.*
We now change our notation and call 2m = n, L = K, ->t< , >_
B( , ) , and F = f. Then the operators TA have the form
(6.4) (Tf)(x) = I
K(x-y)eiB(x,Y)f(y)dy
Rn
We shall suppose B is a real bilinear form, but we shall not suppose

that B is necessarily anti-symmetric nor that K is homogeneous of
degree -n.
THEOREM 7. Suppose K is homogeneous of degree -, 0 < g < n,
smooth away from the origin, and with vanishing mean-value when p = n.
(a) If B is non-degenerate, then the operator T given by (6.4) is

bounded on L2(Rn) to itself, for 0 < p < n ; when 1 < p < oo , the opera-
tor is bounded on Lp(Rn) to itself if -11 I
< 2n
I1 .
(b) If we drop the assumption that B is non-degenerate but require that

g = n, then T is bounded on LP(Rn) to itself for 1 < p < -o. The
bound of T can then be taken to be independent of B.
We shall give only the highlights of the proof, leaving the details,
further variants, and applications to the papers cited below. Let us con-
sider first the L2 part of assertion (a) when n/2 < g < n. Suppose ri
*For further details see Mauceri, Picardello and Ricci [19] and Geller and
Stein [10].
338 E. M. STEIN
is a Co function, with 77(x)=l for IxI < 1/2, and ii(x) = 0, for
1x1 > 1 . We write T = To is defined as in (6.4), but
with K replaced by Ko = riK, and T with K replaced by K = (1-77)K.
Observe first that since Ko(x-y) is supported where Ix-yl < 1 ,
estimating T0(f)(x) in the ball IxI < 1 involves only f(y) in the ball
IYI < 2. We claim
(6.5)
5Ixl<I ITo(f)(x)12dx < A f If(Y)I2dy .
IYI<_2
In fact when IxI < 1 ,
fKo(x_y)ei'cf(y)dy -J Ko(x-Y)eis(y,Y)f(y)dy
<cJ Ix-YI-'`+'If(Y)IdY ,
Ix-YI<I
and thus the L2 theory for f Ko (which is non-trivial only when

= n ) proves (6.5).
While operators of the type (6.4) are not translation invariant they do
satisfy
(6.6) r_hTrhf = e1B(h,h)e1B(x,h)T(eiB(h*)f( ))
with rh(f)(x) = f(x-h). Applying this to To gives the following generali-

zation of (6.5)
J
ITo(f)(x)I2dx < A r If(Y)I2dY
Iy-hI<2
Ix-hi<l
and an integration in h shows that as a consequence
f IT0(f)(x)I2dx < A2n J If(y)12dy
Rn Rn
We now turn to the proof of
J IT f(x)I2dx < A f If(x)I2dx

Rn Rn
This will be done by proving the corresponding result for the operator
T,*,T. The kernel L of this operator is given by
L(x,y) = J e-iB(z,x-Y)1c(z-x)K.(z-y)dz
Now since K. is in L2(Rn) (here the assumption n/2 < is used),

Schwarz's inequality implies
IL(x,y)I < A .
We next integrate by parts in the definition of L(x,y), using the fact

(Dz)Ne-iB(z,x-y) = e-iB(z,x-Y),
that where Dz = i(a,Vz)/Ix-yI , with
a = B-I (Ix yI , and B denotes the matrix so that B(x,y) _ (Bx,y).
` x-
The result is IL(x,y)I < ANIx-yI-N, for every N > 0, and hence
(6.7) IL(x,y)I < AN(1 + Ix-yI)-N , N>0.
This shows that L is the kernel of a bounded operator on L2

proving the boundedness of T ,T and thus of T.. The proofs of the
L2 boundedness when 0 < < n/2 (in part (a) of the theorem), and the
L2 boundedness when = n but when B is not assumed to be non-
degenerate, are refinements of the above argument.
340 E. M. STEIN
Let us now describe the main idea in proving the LP inequalities

stated in (a) and (b) above. We shall need a generalization of BMO (and
of H1 ) which may be of interest in its own right. Suppose E = leQI is
a mapping from the collection of cubes Q in Rn to complex-valued
functions on Rn so that
IeQ(x)I = , Q(x), all x
where XQ denotes the characteristic function of the cube Q. Let us

define on "E-atom" to be a function a so that for some cube Q
(i) a is supported in Q
(ii) Ia(x)I < 1/IQI
(iii) ,f a(x) Q(x) dx = 0
The space HE is then given by If if = I Ajaj , with each aj an E atom,
and I IAil < ool. In a similar vein the function fE will be defined as
(6.8) fE(x) su IQI I !f -fQldx ,
where
fQ IQI f f U_Q
Q
and we take BMOE = IfIfE E LI .

Some of the basic facts about the standard H1 and BMO spaces* go
through for HE and BMOE, and sometimes these come free of charge.
One such case is the following assertion: Suppose f c Lpo , 1 < p < p0
< oo, and fE f LP. Then f c LP and
*The standard situation arises of course when eQ = yQ, all Q.

(6.9) OilLp Lp
To prove this we need only observe that (if I)# < 2f0 , and use the result
(see [10]) for the standard # function.
The point of all of this is that for operators of the form (6.4), there is
a naturally associated HE and BMOE theory, and it is given by
choosing
-iB(x,cQ)
(6.10) eQ(x) = e ,
where cQ is the center of the cube Q. The basic step in the LP

theory (besides an appropriate interpolation which goes via (6.9)), is the
proof that when u = n our operator T maps L to BMOE . Let us
give the proof in the case (a). We may assume that ilf 11 < 1 , and
suppose first that Q is a cube centered at the origin. Then we have to
show that there exists a constant yQ , so that
(6.11) 1
Q,
f JTf -yQ dx < A
Q
The corresponding inequality for a cube centered at another point, say

CQ , then follows from the translation formula (6.6), (and this is the reason
for defining eQ as we do). Turning to (6.11), the argument is not exactly
the same as in the standard case (see e.g. Coifman's lectures [5] or [91),
since we must split f into three parts to take into account the oscilla-
tions of e1B(x,y) . Suppose Q = QS , has side-lenghts S , then write
f = fl+f2+f3, where
Q28' fl = 0 elsewhere,
CQ25 n QS-1 ,
f2 = 0 elsewhere,
SQ2S n cQ 5-1 , f3 = 0 elsewhere.

342 E. M. STEIN
(Note that f2 occurs only when S < x/2/2 .) We have F = T(f) =

F1+F2+ F3, where FJ =T(fj). For F1 we make the usual estimate,
using the fact that T is bounded on L2. Next observe that
IK(x-y)eis(x,Y)- K(-Y)I < cS 1 + 1

IYIn+1 IYIn-1 '
if x cQs and y e'-Q2s. Thus if yQ = f K(-y)f2(y)dy, we get that for

xCQ3
IF2(x)-yQI <A 8
f Q2S
dy
IYIn+1
+S
Qs-1
dy
IYI-1
l<A'.
Finally
K(x-Y)eiB(x,Y)f3(Y)dY
F3(x) = I
= J (K(x-Y)-K(-Y))eiB(x,Y)f3(Y)dY + J K(-Y)f3(Y)eiB(x,Y)dY
For F3 we again make the standard estimates, and for F3 we use

Plancherel's formula (which we may since we have assumed that B(x,y)
is non-degenerate). The result is
5 IF(x)I2dx
3
< J IF3(x)I2dx = A J IK(-Y)f3(Y)I2dy < Al J den = A5"-
IYI
Q Rn Rn cQ
S-1
Combining these estimates proves (6.11), and hence the fact that T takes
L to BMOE .
We shall now state a generalization of this theorem which also in-

cludes the oscillatory integral result given in Theorem 5 (in 5). Suppose
P(x,y) is a real polynomial on Rn xRn of total degree d . Consider the
operator
(6.12) (Tf) (x) = p.v. I e'p(x'y) K(x-y) f (y) dy
Rn
where K is homogeneous of degree -n, smooth away from the origin and
with vanishing mean-value.
THEOREM 8. The operator T given by (6.12) is bounded on L2(Rn) to

itself, with a bound that can be taken to depend only on K and the degree
d of P, and is otherwise independent of P.
This is a recent result obtained jointly with F. Ricci. The proof is

based in part on a combination of ideas used in the proof Theorems 5 and
7. This result has also many variants, and we now state some of these:
(i) One may also show that the operators (6.12) are bounded on LP,
1<p<00.
(ii) Given p, with 1 < p < o, then there is an e =E(p,d), so that if K
is homogeneous of degree - g, n - E < < n , T is still bounded on
LP. However now the bounds may depend on P, and in addition one
must assume that P(x,y) is not of the term P(x,y) = P0(x) + Pl(y).
(iii) One can replace K(x-y) in (6.12) by a more general "Calder6n-
Zygmund kernel" K(x,y), a distribution for which the operator when
P e 0 is bounded in L2 , and which in addition is a function (when
x y ) which satisfies IK(x,y)l < Alx-yl-n, lpxK(x,y)l + IoyK(x,y)I
< Alx-yl-n-I
References. For the detailed proof of Theorem 7, other variants, and

applications to the a Neumann problem see the papers of Phong and the
author [22], [231.
344 E. M. STEIN
7. Further oscillatory integrals related to restriction theorems and

Bochner-Riesz summability
We have seen that if S is a hypersurface in Rn with non-vanishing
Gaussian curvature, then
fei(x)du(x) = 0(I6I-(n-I)/2)
as e 00
whenever 1 e Co , and this is a typical oscillatory integral of the first

kind. We may pass to an oscillatory integral of the second kind when we
replace the Co function l by an Lr function f , and consider the re-
sulting linear operator on f . The resulting operator, as is easy to
observe, is in fact the dual to the restriction operator considered in 4.
Hence by Theorem 3 we can state that operator is in fact bounded from
L2(da) to L p (Rn) , where p' is the dual exponent to 2n + 2 We shall
n+3
now describe the sharper result in this setting that can be obtained for
n = 2.
We fix a curve t -Y(t) = (t, y2(t)) , 0< t < 1, lying in R2 , with
y(t) a C2 , and having non-vanishing curvature, i.e. Iy 2(t)I > c > 0. Con-
sider the transformation T, which maps function on the interval [0,1]
to functions on R2 , given by
(7.1) (Tf)( ) - f e4'y(t)f(t)dt .
THEOREM 9. Under the assumption above T is bounded from Lp[0,1]

to Lq(R2), whenever 3/q + 1/p = 1 and 1 < p < 4.
(Note that when p -. 4 , then q -. 4 in the above relation between p

and q .)
r(7.2)
Proof. Write
F() = ((Tf)( ))2

f
0
J0
e e-(Y(s)+Y(t))f(s)f(t)ds dt ,
and we shall try to apply Plancherel's theorem (more precisely, the

Hausdorff-Young inequality) to F. To do this break the above integral
into two essentially equal parts according to t > s or t < s , which
divides [0,11 x [0,1] into the union of two regions R1 and R2 . We then
consider the mapping of R1 R2 given by x = y(s) + y(t), i.e. x1 =
S + t , X2 = y2(s) + y2(t). It is easy to verify on the basis of our assump-
tions that this mapping is one-one, and its Jacobian j satisfies IJI =
Iy2(s)-y2(t)I > cls-tI . Therefore
(7.3) r e'6'(Y(s)+y(t))f(s)f(t)dsdt = f e'C'xf(x1,x2)dxldx2

R1
R2
with f(x1,x2) = f(s)f(t)1J1_1

So if we denote by the quantity appearing in (7.3) then when-
ever 1 < r < 2 , and 1/r'+ 1/r = 1 , we know that
(7.4) IIF111Lr(R2) <- cOf1ILr(R2)
However
IIfIILr(R2) 5f12t12
= f If(t)Ir If(t)IÌJI I-r ds dt
<c f If(t)Ir If(t)Ir Is-tl I-rds dt

346 E. M. STEIN
To estimate the last integral we need to invoke the theorem of fractional

integration in one dimension in the form
fg(s)g(t)(st)_1+adsdt < AIIgflu , u - 1 = a, 0 < a < 1 .
So we take g(t) = If(t)Ir, then Ilgllu = IIfIIp when p = ur. Then if we fix
a so that -1 +a =1-r, then 2 rr = u , and 3r The limitation
0 < a becomes r < 2, and with q = 2r' we obtain from (7.4) that
1 X2/P
IIFiIILr (R2) < c'
0
with a similar estimate for F2(e) which is the analogue of (7.4), but
taken over R2 . Since F = FI +F2 and F = (Tf )2 we obtain
IIT(f )IILq(R2) <- Ai1fIILp(0,ii
Note that Q =_i=32r3=1-P , so q +P =1, and the limitation

1 < r < 2 is equivalent with 1 < p = 3 < 4. Theorem 9 is therefore
r
proved.
It is clear that inequalities for the Fourier transform play a key role
in the above argument. If we want to generalize Theorem 9 it is natural
to look for a corresponding extension of the L2 boundedness of the
Fourier transform and the Hausdorff-Young theorem. One result along
these lines is as follows.
Suppose we consider the family of operators T, depending on the
parameter A , A > 0, defined by
TA(f)(e) = J ei1'(x,4)&(x,e)f(x)dx
Rn
where 0 is a fixed Co (Rn x Rn) cut-off function; (D is a real-valued

C phase function which we assume satisfies the assumption that its
Hessian is non-vanishing, i.e.
axij
(7.6) det a2L(x, e) 0
PROPOSITION:
A)C-n/2IIfJIL2(R')
(7.7) [ITA(f)IIL2(Rn) < .
COROLLARY:
(7.8) IITA(f )II < AAn/P IIfff ,

Lp (Rn) LP(Rn)
where 1 < p < 2, and 1/p + 1/p' = 1.
REMARK. The boundedness of TA for any fixed A is trivial, but what

is of interest is the decrease in the norm as A - oc. This decrease is
consistent with the special case when t(x,6) is bilinear (and non-
degenerate); when we take A oc in that case we recover the usual
(LP,LP) inequalities for the Fourier transform. Notice also that the
corollary follows from the proposition by the use of the M. Riesz con-
vexity theorem.
To prove (7.7) we argue as in the proof of Theorem 7; as in the treat-
ment of the operator T. it suffices to show that the operator norm of
T*TA is bounded by AK-n.
Now this operator has as its kernel the
function KA(e,n) given by
(7.9) J
Rn
348 E. M. STEIN
Now since
L
(a,Vx)[c(x,n)-F(x,e)] =l _ _ a,17-e'\ + OIn-ej2
OXOV
we can find a = (a,,...,an), so that the aj depend smoothly on x and

IA(x,e,n) >- cIe-nI on the support of KA(e,n). Set Dx = (a,px).
0
= ei\(((x,n)-D(x,e))
Then since (Dx)NeA(t(x,n}-'P(x,6))
, we can
integrate by parts N times in (7.9) and obtain
(7.10) IKA(e,n)I < AN(1 +ale-nI)-N , N > 0 .
It follows from (7.10) with N = n+1 , that the operator TX*TA which has
kernel KA has a norm bounded by AKn and the proposition is proved.
We shall now formulate some theorems for oscillatory integrals of the
form
(7.11) (TAf)(e) = J ei (t.4)c&(t,e)f(t)dt, 6 E Rn
Rn-I
which will generalize the restriction theorems (Theorem 9 above, as well

as Theorem 3 in 4) and also give results for Bochner-Riesz summability.
Notice that (7.11) are mappings from functions on Rn-1 to functions
on Rn. The basic assumptions on the real phase function cF are as
follows: We consider for each fixed (t,e 0) the associated bilinear form
Rn-1
B(u,v) defined by B(u,v) = (v,Vt) (u, Ve) (c) (to,4 0), with u E
v e Rn . Our first assumption is that
(7.12a) B is of rank n - 1 .
Thus there exists (an essentially unique), u E R n, IuI = 1 , so that the

scalar function t - (u,V,(D(t,e)) has a critical point at (t,e). We
shall also assume that this critical point is non-degenerate, i.e. we
suppose the non-vanishing of the (n-1)x(n-1) determinant:
(7.12b) det (pt (G,Oe1)) (t, e 0) - 0
These assumptions will be supposed to hold at all (t,6 ) in the

support of qi(t,e), where i/i is a fixed function in Cc (Rn-1 xRn).
THEOREM 10. Under the assumptions above the operator (7.11) satisfies
(7.13) IITx(f )IIq < Ak n/gllf IIp
with q = n-1 p', p +p 1 1,

(n+1)
if 1<p<4
(a) when n = 2,
(b) when n>3, if 1<p<2.
REMARKS:
(1) When (D(t,6)=t161+t262+...+tn-1en-1+c(ti,...,tn-1).en , and
(V)(0) = 0, then the conditions (7.12) are near the origin equivalent with
the non-vanishing Gaussian curvature of the graph tn if
we apply the result (7.13), letting A - oo , it is not difficult to recover
Theorem 9 from part (a), and Theorem 3 from part (b).
(2) The proof of part (a) follows the same lines as the proof given for
Theorem 9, once we use (7.8) as the substitute for the Hausdorff-Young
theorem; further details as well as relations with Bochner-Riesz summa-
bility may be found in the papers of Carleson and Sjolin [3] and Hormander
[15]. Since part (b) has not appeared before, we will outline its proof.
This will also serve as a good review of many of the notions we have dis-
cussed here.
Proof of part (b). It suffices to prove the case p = 2 , since the case
p = 1 is trivial and the rest follows by interpolation. Now the case p = 2
is equivalent by duality to the statement
(7.14) LIT, (F)IIL2(Rn-1) < AX n/r'IIFIILT(Rn)

350 E. M. STEIN
with r = 2(n+ ,
where
(Ta)(F)(t) = J e-iAD(t14)0(t,e)F(e)de, t e Rn-1
Rn
We can calculate
J TX*(F)Tj*(F)dt
Rn-I
and write as
KA(6,r!)F(e)F(n)d6drl
J
RnxRn
with
(7.15) KA(e,r1) = J e'

Rn-1
It suffices therefore to see that K,\ is the kernel of a bounded operator

from Lr(Rn) to Lr(Rn), with norm not exceeding AA 2n/r'
Because of our assumptions on 4) we can construct a phase function
on Rn x Rn so that the following holds: we will write x e Rn, as
(t,xn) with t = Rn-1 .
The we can construct will
satisfy:
(i) $(x,e)_O(t,e)+(Dp(6)xn
(ii) the determinant of the nxn matrix pxVe$ is non-vanishing.
In fact Vtpe(b already has rank n-1 by assumption (7.12a), and

so we need only choose (Do(e) so that (u,V,)Io(e) 0 to increase
the rank of px e to n .
Now, as in the proof of Theorem 3 in 4, we form Ka defined by
es2
f eia( (x,n
I'(s/2) J
Rn
with dx = dtdxn , and where v is a Co function which equals 1 near

the origin. We easily verify
(7.16)
since
(V(x, when x = (t, 0) .
Next
Kx+it is the kernel of a bounded operator from

(7.17)
L2(Rn) to itself with norm < AX -n/2.
This follows by applying the estimate (7.7) of the proposition above and
using the non-degeneracy of the Hessian of $(x,e). Finally we claim
that
(7.18) I K-n /2+1 /2+it(e ,l)I < A

A
To see this write KAs(6,rq) as KA(6,r?)v(k(4)o(rl)-(Do(6))) where
00
2
vs(u) = F(ss/2) r eixnuv(xn)+xnl-1+sdxn .
-Do
352 E. M. STEIN
Then since Iv_n/2+1/2+it(u)1-< cluln/2-1/2 , as u 00 we see that to

prove (7.18) it suffices to show that
(7.19) IKX(e,q)I <A(AIq-eI)-n/2+1/2
In proving this estimate for the integral KA given by (7.15) we may

suppose that the integrand is supported in a sufficiently small neighbor-
hood of a given point t = to , (for otherwise we can write it as the sum of
finitely many such terms). When we write I(t,ri) = I(t,e) = (VeO)(t,77)
('i-e) + 0(77-e)2 we see that these are two cases to consider as in the
proof of Theorem 1 in 3: 1 when the directions n - 6 or e - n are
close to the critical direction u arising in condition (7.12b); or 2 in
the opposite case. In the first case we use stationary phase (i.e. Proposi-
tion 6 in 2) to obtain (7.19). In the second case, we actually get
0(AI,7-e I)-N , for every N > 0 as an estimate, by Proposition 4. This
completes the proof of (7.18), and shows that Kin/2+1 /2+it is the kernel
of a bounded operator from L1(Rn) to L(Rn), with bounds uniform in A.
The proof of the theorem is then concluded by applying the interpolation
theorem, as in the proof of Theorem 3.
8. Appendix
Here we shall prove Lemma 2 and Theorem 6 which were stated in 5.
First let `f'd denote the linear space of polynomials in Rn of degree
< d . We claim that there is a constant Ad , so that
1 /2
(8.1) IP(x)I2 dx < Ad f IP(x)I dx

(1
IQI J
Q
IQI
holds for all P c 'd , and all cubes Q.

The space Td is invariant under translations and dilations, and so a
moment's reflection shows that to prove (8.1) for all P t Td , it suffices
to prove it for Q = Q0 , the unit cube centered at the origin. However
(f IP(x)I2 dx)1
2 and f IP(x)ldx are two (equivalent) norms on the
Q0 Q0
finite-dimensional space 5d , so (8.1) holds for Q = Q0, and then for
general Q.
Now it is well known (see e.g. [6]) that a function which satisfies a
"reverse Holder" inequality belongs to the weight space A.. Examining
the proof of this fact one obtains an r = r(d), 0 < r < co, and a constant
Cd , so that
1Jr
(8.2)
(7Q,1 f IP(x)I dx
Q (fQ_o
I f IP(x)I"dx << Cd
for all cubes Q.

From (8.2) and Jensen's inequality, Theorem 6 follows easily.
Let us now assume that P is homogeneous of degree d. Observe
also that since (8.2) holds, if we normalize P by the condition that
mp = fIXI_1 I P(x)l da(x) = 1 , we can conclude that
(8.3)
fIXI=1
I P(x)I-r da(x) < ca .
However, when u > 0, Ilog ul = log+u + log+u < u + u -r. Therefore

(8.3) implies (5.4) whenever mp = 1 , and so that result also holds in i
general.
E. M. STEIN
PRINCETON UNIVERSITY
PRINCETON, NEW JERSEY 08544
REFERENCES
[1] M. Beals, C. Fefferman, and R. Grossman, "Strictly pseudo-convex
domains," Bull. A.M.S. 8(1983), 125-322.
[2] J. E. Bjorck, "On Fourier transforms of smooth measures carried by
real-analytic submanifolds of Rn," preprint 1973.
3S4 E. M. STEIN
[3] L. Carleson and P. Sjolin, "Oscillatory integrals and a multiplier

problem for the disc," Studia Math. 44(1972), 287-299.
[4] M. Christ, "On the restriction of the Fourier transform to curves,"
Trans. Amer. Math. Soc., 287(1985), 223-238.
[5] R. Coifman and Y. Meyer, in these proceedings.
[6] R. Coifman and C. Fefferman, "Weighted norm inequalities for
maximal functions and singular integrals," Studia Math. 51 (1979),
241-250.
[7] S. Drury, "Restrictions of Fourier transforms to curves," preprint.
[8] A. Erdelyi, "Asymptotics Expansions," 1956, Dover Publication.
[9] C. Fefferman, "Inequalities for strongly singular convolution opera-
tors," Acta Math. 124 (1970), 9-36.
[10] C. Fefferman and E.M. Stein, "HU spaces of several variables,"
Acta Math. 129 (1972), 137-193.
[11] D. Geller and E. M. Stein, "Estimates for singular convolution
operators on the Heisenberg group," Math. Ann. 267 (1984), 1-15.
[12] A. Greenleaf, "Principal curvature in harmonic analysis," Ind.
Univer. Math. J. 30(1981), 519-537.
[13] C.S. Herz, "Fourier transforms related to convex sets," Ann. of
Math. 75 (1962), 81-92.
[14] E. Hlawka, "Uber Integrale auf konvexen Korper. I," Monatsh.
Math. 54 (1950), 1-36.
[15] L. Hormander, "Oscillatory integrals and multipliers on FLU," Ark.
Mat. 11 (1973), 1-11.
[16] , "The analysis of linear partial differential operators. I,"
1983, Springer Verlag.
[17] S. Krantz, "Integral formulas in complex analysis," in these
proceedings.
[18] W. Littman, "Fourier transforms of surface-carried measures and
differentiability of surface averages," Bull. A.M.S. 69(1963),
766-770.
[19] G. Mauceri, M.A. Picardello, and F. Ricci, "Twisted convolutions,
Hardy spaces, and Hormander multipliers," Supp. Rend. Cir. Mat-
Palermo 1(1981), 191-202.
[20] J. Milnor, "Morse Theory," Annals of Math. Study #51, 1963,
Princeton University Press.
[21] A. Nagel, "Vector fields and nonisotropic metrics," in these proceedings.
[22] D. H. Phong and E. M. Stein, "Singular integrals related to the Radon

transform and boundary value problems," Proc. Nat. Acad. Sci. USA
80(1983), 7697-7701.
[23] , "Hilbert integrals, singular integrals, and Radon trans-
forms," preprint.
[24] E. Prestini, "Restriction theorems for the Fourier transform to some
manifolds in Rn in Harmonic analysis in Euclidean spaces," Proc.
Symp. in Pure Math. 35, part 1(1979), 101-109.
[25] B. Randol, "On the asymptotic behaviour of the Fourier transform of
the indicator function of a convex set," Trans. Amer. Math. Soc.
139(1%9), 279-285.
[26] E.M. Stein, "Singular integrals and differentiability properties of
functions," 1970, Princeton University Press.
[27] E.M. Stein and S. Wainger, "The estimation of an integral arising in
multiplier transformations," Studia Math. 35(1970), 101-104.
[28] E. M. Stein and G. Weiss, "Introduction to Fourier analysis on
Euclidean spaces," 1971, Princeton University Press.
[29] R. S. Strichartz, "Restrictions of Fourier transforms to quadratic
surfaces and decay of solutions of wave equations," Duke Math. J.
44 (1977), 705-713.
[30] P.A. Tomas, A restriction theorem for the Fourier transform," Bull.
A.M.S. 81 (1975), 477-478.
[31] , "Restriction theorems for the Fourier transform in
Harmonic Analysis in Euclidean spaces," Proc. Symp. in Pure Math.
35, part 1 (1979), 111-114.
[32] S. Wainger, "Averages and singular integrals over lower dimensional
sets," in these proceedings.
[33] A. Zygmund, "On Fourier coefficients and transforms of functions of
two variables," Studia Math. 50(1974), 189-201.
AVERAGES AND SINGULAR INTEGRALS
OVER LOWER DIMENSIONAL SETS
Stephen Wainger(l)
I. Introduction
These lectures deal with work primarily due to Alex Nagel, Nestor
Riviere, Eli Stein, and myself dealing with certain averages of and singu-
lar integral operators on functions, f , of n variables, n > 2 . These
averages and singular integrals differ in character from the classical
theory in that the integration is over a manifold of dimension less than n.
Let us begin with an example of the type of problem we have in mind.
The classical differentiation theorem of Lebesgue asserts for any locally
integrable function f
a.e.
f(x) = limo IQ
r
I f f(x-y)dy
(where Qr is the square, Q. =jxcRnIsupIxil<rt, and tQrj denotes the

Lebesgue measure of Qr ), and
f(x) = lim 1
r-+0 JBr) J
r f(x-y)dt a.e.,
Br
(where B. is the ball, Br = {xI lxI <rl ).
*Supported in part by a grant from the National Science Foundation.
357
358 STEPHEN WAINGER
Our first problems are the following:
Problem IA:
Does
1) lim f(x-y)dar(y) = f(x) a.e.?

r-+0
aQr
Here aQr denotes the boundary of Q. and dar is n-1 dimensional

Lebesque measure on aQr normalized so that dar(aQr) = 1 .
Problem IB:
Does
2) lim J f(x-y)dp = f(x) a.e.?

r-O
aBr
Here di is the unit rotationally invariant mass on aBr.

1) and 2) trivially hold if f is continuous, and the questions only
become interesting when we consider functions in a class like L-, L2,
or LI.
In questions IA and IB, we are considering certain averages
3) MQ f(x) =
r
r
,J
f(x-y)dar(y)
aQr
and
4) MB f(x) = J f(x-y)dkr(y)
r
aBr
We are asking if
5) MQr f(x) f(x) a.e.

AVERAGES AND SINGULAR INTEGRALS 359
and
6) MBrf(X) - f(x) a.e.
The standard approach to this type of problem involves consideration

of appropriate maximal functions.
We define the maximal functions
7) )lIQf(x) = sup MQ (If1)(x)

r>0 r
and
8) )RBf(x) = sup MB (IfI)(x)

r> O r
9fIBis called the spherical maximal function.

Since 1) and 2) hold for f which are continuous 1) would follow for
every f in LP, 1 AlI <--9 1'fI'p
for every f in LP, and 2) would follow if we could show
10) I1xI)RBf(x)>x1I <-L IIfIIP
The argument showing that 9) and 10) imply 1) and 2) is the same as the
argument showing that Lebesgue's differentiation theorem follows from
the weak type inequality for the Hardy-Littlewood maximal function given
in chapter 1 of [S]. While it is not quite as well known, there are appropri-
ate estimates on maximal functions that guarantee 1) and 2) hold for all
L functions. In our case this means the following;
Let E be a measurable set and XE its characteristic function.
Then if
(11) I1xIRQXE(x)>a1I < C(a) I E I

360 STEPHEN WAINGER
where C(a) may depend on A but not on E, then 1) holds for every f
in L. If
12) I1xI911BXE(x)>a1I < C(A)JEJ ,
then 2) holds for every f in A discussion of this can be found in

[BF].
Let us try to see if 9) or 10) could be true in some simple cases. We
consider for example the one-dimensional case. Here Br = Qr = lxl-r<x<r},
and
MQrf(x) = MBrf(x) = [f(x+r)+f(x-r)]
2
So if we take f(x) to be log L near x = 0, have compact support,
Ix
and be in C away from the origin, we would have a function f in
every LP class such that )IIQf(x) =)IIBf(x) = oo for every x . We can
also see that 11) and 12) are false in one dimension. We just take
EE = Ix 10<x<E} . Then J EEl 0 but
)IlQf(x) = 911Bf(x) > 1
for all x.
We could still ask if 1) and 2) hold in some interesting class even
though 9), 19), 11), and 12) fail. However an important idea of Stein
shows that the failure [SI] of 9), 10), 11), and 12) implies that 1) and 2)
fail even in the class of locally bounded functions. The statement of the
main theorem of [SI] requires that the underlying space be compact. But
if 1) or 2) were true for an LP class on Rn, it would also hold for the
corresponding LP class on the torus. Furthermore, the theorem of Stein
requires the hypothesis that 1 < p < 2. However due to the positive
nature of the averages under consideration, his ideas can be modified to
show that 1) fails for at least some L functions. See [SW]. Thus we
obtain negative results in one dimension. Similar reasoning gives the
same negative conclusion for question IA in any number of dimensions.
One need only consider f is as

above and h is a nice function.
So there are no interesting positive results in problem IA. However
as we shall see later there are positive results for problem IB in 3 or more
dimensions. One might ask if there is a simple geometric reason why
there should be positive answers for the sphere and only negative answers
for the boundaries of squares. It turns out that the underlying basic
reason that we have positive results for the boundary of balls and nega-
tive results for the boundary of squares is that spheres are round and
boundaries of squares are flat. In other words an important word for us
will be
CURVATURE.
We will come back to the role of curvature in our problem in a little
while, but first we shall discuss the other problems that we will consider.
Problem II: Let y(t) be a curve passing through the origin in Rn . Is it

true that lim h J f(x-y(t))dt = f(x) a.e., for f in L' or L2 or LI ?
Problem III: Let v(x) be a smooth vector field in Rn. Does
h
1
lim 1 f(x-tv(x))dt = f(x) a.e.
h-'0 h _
0
for f in L or L2 or LI ?
Corresponding to problems II and III there are interesting singular
integrals. We let y(t) be a curve and v(x) be a smooth vector field as
in problems II and III. We set
13) Hyf(x) = f (x-y(t)) Lt .

J-a
(where sometimes we wish to think of a as finite and sometimes as 00 ),
and
362 STEPHEN WAINGER
14) Hvf(x) f(x-tv(x)) dt

-1
We call Hy the Hilbert transform along the curve y and Hv the

Hilbert transform along the vector field v(x). We then have the following
two problems:
Problem II': Can we have an estimate
15) IIHyf 11 < CPIIIIILP

LP
for some p's ?
Problem III': Can we have an estimate
16) IIHyfIILp <_ CpIIfilLP
for some values of p ?
The classical development of singular integrals and maximal functions

suggests that problems II' and III' should be related to problems II and III.
In fact the progress on problems I, II, III, II', and III' is all interrelated.
We have presented our problems as variants of Lebesgue's Theorem on
the differentiation of the integral. These particular variants arose from
other considerations.
Riviere was led to problem II' from the consideration of a problem of
singular integrals, namely from trying to generalize the method of Rota-
tions of Calderon and Zygmund. Calderon and Zygmund developed the
method of rotations to reduce the study of operators
Tf = K*f
where K is a kernel having "standard homogeneity" that is
17) K(Ax) = A-nK(x) h>0

(K(x) is a function on Rn ) to the one-dimensional Hilbert transform
Hf(x) = ff(x-t) dt
(f a function on R1 ).
We will explain how the method of rotations can lead to problem II'.
Let K(x,y) be a function of two variables x and y which is odd,
K(-x,-y) = -K(x,y)
and which has a "parabolic homogeneity," that is
18) K(Ax.X2y) = 3 K(x,y)

X3
We wish to consider the LP boundedness of the transformation
00 00
19) Tf(u,v) = I f(u-x,v-y)K(x,y)dx dy .
We now introduce parabolic polar coordinates into 19)
x = rcos 0
y = r2sinO
and find
00 277
20) Tf(u,v) = J J f(u-rcos 0, v-resin 0)

0 0
K(rcos 0, resin 0)r2N(0)drd0
where r2N(0) is the Jacobian factor in the change of variables. N(0) is

smooth and N(0+n) = N(0). By 18 we see that
364 STEPHEN WAINGER
27r o0
21) Tf(u,v) = r N(O)K(cos d,sin O)dO r C f(u-rcos 0,v-rsin 0) dr

0 0
2rr 00
-J N(O)K(cos(O+n),sin(O+n))dd f
=
0 0
i f(u-rcos 0, v-rsin(O)) dr
since K is odd. Thus
217
fTf(u,v) r N(O)K(cos 6,s in 0) dd i f(u+rcos O,v+rsin 0) dr

0 0
since
N(6+n) = N(6) .
Finally
277 00
21A) Tf(u,v) = r N(O)K(cost,sinO)dO r r f(u-rcos6;v-r2sinO)dr

Now adding 21) and 21A) we find that
2n ao
f
Tf(u,v) =
2 0
N(O)K(cos O,sin 0)dd f
-00
i (u-rcos6,v-r2sin6)dr
f
If K(cos O,sin 0) is in LT of [0,2n] , we can apply Minkowski's

inequality
n
IITfIILp < C f de IIHIILp
0
where
f
00
Hef = dt
f(x-yg(r)) ,
-00
with
ye(r) = (rcos O,r2sin O) .
Now we prove
IITfIILp<cI{fIILP
by showing
IIHefIILP <cplif11Lp
This is a problem of the type II'.

Stein was led to consider problem II by his study of Poisson integrals
on symmetric spaces. We are not going to launch into a discussion of
symmetric spaces, but instead we consider an example. Let
MEf(x,Y) = E2 ff f(x-r,Y-s)
(is)(i+
1
drd s
E2
(")
If K(x,y) were dominated by a decreasing, radial, LI function, the
classical theory would imply
lim MEf(x,y) = f(x,y) a. e.

E-0
see [SWE]. However
1
K(x,Y) =
(1+x2)(1+y2)
366 STEPHEN WAINGER
so the smallest radial majorant of K is 1 which is not integrable.

1+x2+y2
In effect K has too much of its mass along the coordinate axis. The
extreme case of this phenomena would be to have a kernel with all of its
mass on the coordinate axis.
In other examples, kernels have too much of their mass along curves,
and the extreme case of difficulties arising in problems of Poisson Inte-
grals on symmetric spaces lead to Problem II.
Appropriate positive results to problem III would have implications for
the boundary behavior of functions holomorphic in pseudoconvex domains
in Cn. The natural balls in these problems are long, thin and twisting.
The idealized situation is that of a vector field. In the case of a strictly
pseudo convex domain, the balls satisfy the standard properties that en-
sure that the usual covering arguments apply. See [SBC]. For progress
in the case of pseudoconvex domains see [NSW] and [NSWB].
Now that we have seen some of the roots of our problems, let us con-
sider why these problems don't fit into the framework of the standard
theory of Maximal functions and singular integrals as presented for
example in [S].
In the standard treatment of averages over Euclidean balls an important
geometric property of the Euclidean balls is used. If two balls B1 and
B2 of the same radius, r, intersect, then B2 , the ball having the same
center as B2 but having radius 3r contains B1 . To see how badly
this property fails for our problems let us suppose we were considering
averages
h t2+E
1 ff
0 t2
f(x-r,y-s) dr ds
over slightly thickened parabolas or balls
BE,h = i(r,s)I0<r<h, t2<s<t2 +E I .

We now consider the intersection of two of these balls of the same size
Clearly one of these "balls" is not contained in a fixed multiple of the

other (uniform in r ).
We can also see the difficulty of using the Calderon-Zygmund theory
to study Problem 11'.
Suppose y(t) is the parabola (t,t2) in R2 and
Tf(x,y) = J f(x-t,y-t2)dt
Then we may formally write
Tf(x,y) = fff(x-t,y-st2)- dsdt
f(x-t,y-v) . S (1- -1 dv dt
= JJ 0)
Or
22) Tf=K*f,
where
22A) K(x,Y)=XS(1--X ),
368 STEPHEN WAINGER
The Calderon-Zygmund theory deals with convolution operators with

kernels K(x,y), but in their theory K(x+h,y+k)- K(x,y) should be much
less than K(x,y) if h and k are much smaller than x or y. However
for our K if (x,y) is a point on the curve y = x2 and (x+h,y+k) is not
on the curve, no cancellation in the difference K(r+h,y+k)-K(x,y) can
occur, no matter how small h and k are.
The Calder6n-Zygmund Theory is based on 4) tools
a) The Fourier transform (The Fourier transform is even used in the
L1 theory)
b) Interpolation
c) Covering lemmas
d) Calderon-Zygmund decomposition.
Perhaps the natural attack on our problems would be to find appropriate
covering lemmas and suitable variants of the Calder6n-Zygmund decom-
position. Some progress in finding covering lemmas for related problems
was made by Stromberg [Str] and [STRO] and Cordoba [COR1], [COR2],
Cordoba and Fefferman [CF1], [CF2], [CF3], and Fefferman [FEf].
Our approach will however be different. We shall try to use the Fourier
transform or other orthogonality methods and interpolation to reduce our
problems on averages and singular integrals to the more standard averages
and singular integrals. In retrospect we see that some of these ideas
occurred in [SPL], [CS], and in [KS].
We have said earlier that curvature and Fourier Transform would be
important for us. Actually they go together. If one has a nice measure on
a curved surface, the Fourier transform of that measure decays at infinity
even though the measure is singular. Let us consider some examples.
Define, for a test function ,,
r O(t,O)dt .
0
u is supported on a straight line, namely the x-axis, and

1
g(e'exel)?y) = r eietdt
0
which is independent of n and hence cannot decay at infinity along the

Y7-axis. Now let us consider a measure supported on a parabola,
00
23) v(q) = r e_t2rb(t,t2)dt .
_
Then
v(e exei7y)
00
fe_t2eietet2dt.
This integral may be computed exactly by completing the square, and it

is easy to see that
e2
Ce
1+1771
Thus v(e,q) tends to zero at infinity.

Another example is afforded by rotationally invariant Lebesgue
measure on the n-1 dimensional sphere Ixl = 1 in Rn. If we denote
this measure by d, we have for e c Rn
= CnJ(lei) Ii (v).
2
370 STEPHEN WAINGER
See [SWE]. Thus
^ _ (n 1
Idu(4)I < Cn(1 + I.I) 2
Of course we want to have a tool to estimate the Fourier transform of

measures in general, not in just a few specific cases.
This tool is a lemma of Van Der Corput.
VAN DER CORPUT'S LEMMA. Let h(t) be a real function. For some j ,
assume Jh(1)(t)I > A in an interval a < t < b. If j = 1, assume also that
h'(t) is monotone, then
b
r exp (ih(t))dt
a
For the proof of Van Der Corput's lemma for j = 1 and 2 see [Z]. The
proof for higher j is similar.
Let us consider the measure
24) du(O) f1
2
(k(t,t2)dt.
Then
2
du(e,i) = el4telnt dt
1
This integral cannot be evaluated explicitly, but we wish to see that Van
Der Corput's lemma may be applied. We take h(t) = e t4 rt2 . First we
use the fact that h"(t) = rt . Thus by Van Der Corput's lemma with j = 2 ,
we see
25)
n
ldu(, , C
1
(1 + Inl)I12
Now if Ir1I < igl,
Ih'(t)I = Ie+2,itI > e/8 .
So by Van Der Corput's lemma with j = 1 ,

n
26)
I .I
if
161 > 8lnl
Putting 25) and 26) together we have
27) C
for some 8 > 0.

Stein pointed out in retrospect that we can already see from an
estimate like 27) that du has interesting properties from the point of
harmonic analysis - namely even though du is singular,
Tf=du*f
maps LP into L2 continuously for some p < 2. For
372 STEPHEN WAINGER
f(Tf)2 = ITf(6,r/)12 dtd77
= f Idu(e,17)12If(e, )I2 de d,1
=f +e 2 +1771)2,5
If( ,ii)I 2
(1+e 2+I17I
If(e,,7)12
(1+e2+1,,127
2q ) /q 1 /q
If(e,rl)I
J (Ll+e2+'2')
The second integral is bounded if q' is sufficiently large which means
for some q > 1 . But then the first integral is bounded for f e LP where
P +2q=1
II. The Hilbert transform along curves

The first progress in our series of problems was made on the Hilbert
transform along curves. The Hilbert transform along a curve can be
thought of as a multiplier transformation
28) Hyf(e) = my(e) f (e )
where
29) my(e) = f
_,p
Lt
To see that 29) is true we may either substitute the formula

f(x) je-e'x f(e)dC
into 13) or recognize the fact that
where D is a distribution
So
and b may be computed by evaluating D on an exponential.

Thus to prove that Hy is bounded on L2 one needs to show that
mY is bounded. The first result of this type was obtained by Fabes [F].
Fabes showed HY is bounded on L2 in 2-dimensions for the curve
y(t) = (t, ltlasgnt) , a > 0.
So Fabes' proof consisted in showing that the integral
m((,q) = r exp(ite+iIt1a(sgnt)rt) Lt
-00
is uniformly bounded in 6 and q. To this end Fabes employed the

method of steepest descents. The method of steepest descents is a
method of obtaining very precise asymptotic information for large A about
integrals of the form
f exp (iah(t))dt
374 STEPHEN WAINGER
by contour integration. However to employ the method one has to have

very precise information on where the real part of h(z) is positive and
negative in the complex plane. Thus already to employ the method of
steepest descents for the curve (t,t2,t3), one would have to understand
the zero set of
Real Part
uniformly in 61 , e2 and 63. So it is hard to imagine using the method

of steepest descents, and for the curve t, t2, t3, t4, is it would seem close
to impossible. Fabes' result was very important in that it gave the first
clue that problems such as II and III could have positive answers.
However a better method would have to be found - a method that
needed less precise information about h(t). The next step was to
show that if y(t) = (t,tal,ta2'...'tan-1), 1 <a1 <a2
(here t 1 can mean either Itlal or Itlalsgnt ) then Hy was bounded on
L2(Rn) [SWA]. Here we had to prove the boundedness of the integral
00
r dt
J
The proof was by way of the Van Der Corput lemma but was unnecessarily
complicated because at that time we only knew the lemma for j = 1,2 .
Let us see how Van Der Corputs'lemma works in the case y(t) =
(t,t2'..., tn). We then have to show that
30)
I
E <Itl<R
dt < C(n) .
where C(n) does not depend on We shall prove 30) by

induction on n. By changing variables, replacing t by 1/n we
41
may assume en = 1 in 30). Then by using Van Der Corput's lemma

with j = n , we find
t
exp (iels + ... + ien- sn-1 + < C(n) .
31) iSn)
An integration by parts together with 31) shows that
32)
1
1<JtJ<R
Now
f exp (ieit+...+ien_ltn-1 +itn) dt
I exp (ieit+... + ien-ltn-1) Lt

t
<
t
Jr to dt < C(n) .
E<I tl<1 E<t< I
Hence we have reduced the proof 30) for n to proving 30) for n - 1.
Also the case of n = 1 is easy. So we are done by induction.
We now turn to the LP theory. We wish to emphasize how curvature
and Fourier Transform are joining together to help us. So we shall com-
pare the case of the parabola (t,t2) to the straight line (t,t). In the
case of the parabola we are studying
34) Hpf = DP * f
where DP is a distribution. For a test function 0

376 STEPHEN WAINGER
35) Dpi dt
In the case of a straight line we are studying
36) HLf=DL*f,
where
37) DL95 = J o(t.t) dt
00
eieteirlt2 dt
38) Dp(e,rl) = Dp(e1eXei"ly) = F
-00
and
39) DL(e,rl) = DL(e exe"7y)
f00ete'it.
-00
t
We can calculate DL explicitly, and we find
40) DL(e,rl) =c sgn((+n) .
Notice that DL is discontinuous along the line We shall

show in contrast that Dp(e,rl) is continuous away from the origin. It is
very easy to see that Dp(e,rl) is C away from the line i = 0 by com-
plex integration. If for example rl > 0, we think of t as a complex
z
variable and integrate along the line lmt = Ret. Then the factor eirlt
decays as fast as e-c7ItI2 , and one can easily justify differentiation
under the integral sign as long as it > e , for some positive E.
We shall now show that Dp(6,0 is continuous near 161 = 1 . What we

must show is that l m0 Dp(4,-q) exists for 6 near 1 . We shall show
00 00
41) lim
,q-to J
r ei 'tei77t2 dt _
t
-00
ei6t dt
t
Assume for simplicity that 71 > 0.
Of course
1
1/3
77 00
42) lim r eiCt dt = r ei4t

Lt
7,-0 J t
-00
,17 1 /3
and
I
,71 /3
43)
1
eiet(ei??t2_1) dt
t
.771 /3
711r/ 3
< 271 tdt < 171/3 - 0 as 17 ->0.

J
0
In view of 42) and 43) we can show 41) by showing
44)
I
f
1/3 <I fl
e'eteigt2 dt o.
378 STEPHEN WAINGER
We shall prove 44) by using Van Der Dorput's lemma. By using

Van Der Corput's lemma with j = 2, we see
f -2/3
eiese"gs
2
ds < c
XFII
(Let h(s) = es + qs2 , then h"(s) = 277 .) So an integration by parts

shows
00
45) eieteir7t2 d t
00
< 2/3 + 1 (' dt

77
1/2 ,J t2
77
2/3
77-
< C 77 2/3171/2 < C171/6
Note that if t < 77-2 /3
ie+217tl > iei - 2,71/3
So
-2/3
77
e tese l'7s 2 ds
1/3
77
is bounded if 6 is close to minus one. Hence an integration by parts

shows
1
13 00
77
46)
f eieteil7t2 dt
t <C f 1
dt +77 1/3<C711/3
t2
1/3
77
45) and 46) together with similar estimates for negative t prove 44)
and hence the continuity of Dp(e,-9) away from the origin. If one is a
little more careful in the above argument, one can prove that Dp(C,.9)
satisfies a Lipschitz condition away from the origin. One may then use
Riviere's [R] version of Hermander's multiplier theorem [H] to obtain
some LP results for p 2 . In fact if y(t) = (t,t2) one obtains
47) IIHyfIILpCC IIfllLp, 3 <p<4.
For quite a while we tried to prove the range of p, 3 < p < 4 , in

47) was optimal - with no success. Also, there was a suspicion that the
use of the Hormander Riviere theorem lost something. In the Hormander
argument, one wishes to estimate an expression of the form
J=
f
R<Ixl<2R
IK(x+h)-K(x)Idx
where K is a kernel, in terms of K. One does this by using Schwartz's

inequality and Plancherel's theorem.
J=
f
R<IXI<2R
I la
IxIaIK(x+h)-K(x)ldx
112
<(
f
R<I x l< 2 R
IXI2a
1
d)'
x (J
IxI2aI(x+h)-K(x)I2dx)
/
< ,_n fI1ei2i K( )12

R 2
where the sum is over all a'th derivatives of K.

380 STEPHEN WAINGER
Now there was the feeling that a use of Schwartz inequality like that
above lost too much, and that more careful estimates for j for particular
kernels might lead to better results.
To get an idea of what to do we calculated D(e,n) very precisely by
the method of steepest descents. We found
2
48) Dp(6.77) = sgne + C I77I e1e2/n
`- 2/
+ Better terms ,
where i/i is a Co function on R1 which is one near the origin. Hence,

the crux of the matter was to study the transformation Tf given by
49) Tf (e,11) = m(e,rl)f((,rl) ,
where
50) m(,n) =
Ir111/2
1
r)2 1
e
This suggests introducing an analytic family of operators in the sense of

[SI] as follows:
51) T (e,rl) = mz(e,rl)f(e,i)
where
2)
(171
52) mz(e,rl) = m(e,v)
Tzf is clearly bounded on L2 if Re z = -1/2 . So if we could

prove that the kernel Kz corresponding to Tz for Re z positive satis-
fied a condition of the Calderon-Zygmund type, we could prove Tz was
bounded in each LP if Re z > 0. Hence by Stein's interpolation
theorem we would know that To = T was bounded in all LP, p < 2.
Then by a duality argument T would be bounded in all LP, 1 0.
Let us try to understand why the kernel for Tz , Re z > 0, might be
a little better than the kernel for To. T is essentially HY y = (t,t2),
and so the kernel K0 of To is essentially
K0(x,Y)=X8(1-z )
from 22) and 22A). If we introduce "parabolic polar coordinates"

y =resin 0 x =rcos 0 , we see
K0(x,y) = a S(9-90)
r3
where
sin0 0
cos200
We might expect if Re z = e > 0 Kz to be E better than KO. So we

might expect
1 1
Kz(x,Y) =
r3 19-0011-e
Now we would like to explain why a 1 singularity is better than

10-0011-E
a S(0-90) singularity. To see the situation more clearly, let us examine

the analogous situation for the standard polar coordinates with 00 = 0.
Suppose
S3) K0(x,Y) _ O
r
where x = rcos 0 and y = rsin 0 , and

382 STEPHEN WAINGER
1
54) KE(x.y) =
r 21011-E
(The factor for ordinary polar coordinates plays the same role as
r
r
for parabolic polar coordinates.) We are trying to see whether
55) jK(x,y)-K(x-h,y-k)j < C .

1
x2+y2>c(h2+k2)
For either K = KO or K = KE . Let us take h = 0 and k = 1 and con-

sider first K = Ko. Let us look at the contribution from y's which are
very close to 0. We have
)Id9rdr
J J r2
I-D- r20
r>C nearo
If y is very close to 0 161 '" r
y
r
0 y-1
r
So the left-hand side of 55) is at least fC i dr = 00 . So 55) can't

hold. Let us put the matter a little differently. If we consider the 0's
with 0 > 0 where the difference
K0(r, 0) - Ko(r , 0)
offers no cancellation, we find there is only one bad 0, 0 = 0. But still

n/4
f
J0
JK0(r, 6)- K0(r', O')j dO = 1 .
Let us consider now what happens with We should expect no

help from the difference K(x,y) - K(x,y-1) when y > 0, if y-1 < 0.
But this can only happen if y < 1 or 0 < 1/r . But over this set KE(x,y)
is integrable at infinity
00 1 /r
KE(r, 0)d0rdr
5
1 50
00 1/r
Irfderdr
2
el-F
5 0
f1
00
< dr .
,J rl+e
5
It is not difficult to complete the argument and to show 55).

After a laborious calculation one could prove that the kernel Kz
corresponding to Tz of 51) was for Re z > 0 an operator of Calderon-
Zygmund type. This proved that Hy was bounded in LP 1 < p < oo if
y = (t,t2). However it would be extremely difficult to carry over this
proof to a three dimensional curve. For example it would be hard to
derive an analogue of 48) for the curve (t,t2,t3). Essentially one
needed a way to define a suitable analytic family Tz without using the
asymptotic formula 48). Recall that Hyf = DP * f where
56) Dp(e,rl) = reieteint2 dt

384 STEPHEN WAINGER
So one might be tempted to define
57) Hy,f = DP * f
w here
58) Dz(e,rl)
p
e a irlt2 dt
(1+t2)z/2
It turns out that 58) is not a good idea for a very important reason. By
changing variables in formula 56) we see that Dp(Ac, A271) = Dp(c,rl) , for
any A > 0. Note that also the function mz(e,r/) defined in 52) also has
this type of homogeneity, namely mz(Ac,A2r7) = mz((,77), for A > 0.
Now experience has shown that homobeneity is a powerful friend not
to be tossed away lightly. However Dp does not have this homogeneity.
This situation can be remedied by defining
59) Hyf=Dzxf
where
00
+r72t4)-z /4 Xteir7t2 dt
60) Dp(c,rl) = (1
I
-00
Note that for A > 0
61) Dp(Ac, A2rl) = DP(c ,77) .
Let us see how formula 61) can help us. We would like to show
62) C(z)
if Re z > -2 . By formula 61) we may assume r/ = 1 , let us say 77 =1.

Then by Van Der Carput's lemma with } = 2 , we see that
t
2
eXsel1 ds
ft
d ese's2 dsI <C
So an integration by parts shows
00
J (1+t4)-z/4eieteirlt2 dt < C(z)

t
1
Now
j
-1
(1+t4)-z /4 Xt eirlt2 dt
t
f
-1
e'et t2 dt
t
<C(z) r t2dtt<C(z).
J-1
But we already know that
r dt
J T <C.
-1
386 STEPHEN WAINGER
One may finally deal with -00 < t < -1 in the same manner that one
treated 1 < t < 00. Hence 62) is proved.
Let us calculate the kernel, Kz , corresponding to Hy if z = E > 0.
Then
eiCx e myDp(e,rl)dCdrl
KE(x,Y) =fv
00 00
f f
eirlt2
Lt e-177Y (1
+,72t4)-E/4
t
- 00
00 00
f e-'6(x-t)dx = r dt
f
i?7(t2-Y)(1+772t4) -E/4 ei 7t2&(x -t)dt
J J
-00
t
00
00
00
-.1 ir7(x2-y)(1 + 772x4)-E/4 drl

X
_00
0o
Ir7(x2y)
x3 f e 1
(1+,72)1/4
dr7
-00
x2'-Y2
1p
x3 e/2 x2
(
where PE/2 is a modified Poisson kernel. PE/2 decays exponentially
fast at oo and PE/2(u) ti JuJICE/2 as u -, 0. See [SWE].
Thus KE(x,y) has a singularity near the curve (t,t2) of the form
1 which is just the improvement over the 8-function that we
1e_eo1i-E/2
seek.
A modification of these ideas worked for curves
a'
,ta2,... , tan)
Y(t) = (t
al < an. See [NRW].

However, there is a natural generalization of these curves. All of
these curves satisfy an equation of the form
63) Ye(t) = A Y(t)
where A is a real nxn matrix such that the real parts of the eigen-
values of A are positive. For example if
y(t) = (t,t2)
A curve satisfying 63), where all the eigenvalues of A have positive

real part is called a homogeneous curve.
A will generate a group of transformations
TX = exp (A loga) .
Then
64) Hyf - Dy*f
where
65) Dy(Tx) = Dy(e )
Moreover there is a distance pA(x) defined on Rn such that
P(TAX) _ AP(x)
In the case of the cure (t,t2) we may, as we said before take

38$ STEPHEN WAINGER
0
A =C1
0 2 )
Then
0 ),
T- ,
and
p(x,y) = (x4+y2)1 /4
It turns out that in the case of a general homogeneous cure, we can

obtain a satisfactory analytic family of operators by defining
66) Hzf=Dz*f
where
67) z(e) =pA*(() j ItI-zexp(ie'y(t))dt
p * is the distance function corresponding to A*, the adjoint of A.

A
For a detailed description of the argument see [SW]. Here we shall
just make a comment. If some of the eigenvalues of A have non-zero
imaginary part, y(t) can be an infinite spiral. For example the curve
y(t) = (tacos (j3logt), tasin(/3logt))
is an example. So one could believe it might be rather messy to prove

integrals involving exp to be bounded. It might be difficult to
show that at each t some derivative of , y(t) would be non-zero.
However, if one makes a change of variables t = eu , we would be led to
consideration of integrals involving n(u) where n(u) = y(eu). If
A
y'(t) = y(t), n(u) satisfies
68) n'(u) = Ai(u) .

We shall show that if rj(u) is a curve in Rn satisfying 68) where the

eigenvalues of A have positive real part, then either r)(u) lies in a
proper subspace of Rn or for every 0 and u there is a j
1<j<n such that
69) rl(u) 0.
du)
From 68) we see that
dI+17
(u) Ajrl
(u)
du1+1
By the Cayley-Hamilton theorem, we can find numbers aj , 0 < j < n,

such that
jAj=0.
j=0
So
n Yd j+t
j=O
aJ duI+1 '7(u) = 0
and
j=0
n
aj a j
duJ
. rr'(u) =0.
In other words rJ,(u) e satisfies an nth order constant coefficient

differential equation. So if for some u and
d, n'(u) . =0 j = 1 ,2,...,n-1
du3
rj'(u) 6 = 0 for all u. Thus r)(u) 6 is a constant. But r!(u) 6 , 0

390 STEPHEN WAINGER
as u -, -oo since the eigenvalues of A have positive real part. Hence

rt(u) is in the subspace of Rn orthogonal to e.
We shall conclude this section with the statement of some theorems
that follow from the reasoning discussed above.
THEOREM 1. Let y(t) satisfy
A
Y '(t) - Y(t)
Suppose the span in Rn of y(t) for positive t and the span in Rn

of y(t) for negative t agree. Then
IIHyf1I < IIf1ILp 1<p<00 .

LP Cp/Y
We say that a curve y(t) in Rn is well curved if
dJy(t) j = 1,2,...
dyj(t) t= O
span Rn. It turns out that well curved curves can be approximated by
homogeneous curves. We can then prove
THEOREM 2. Let y(t) be well curved then,
dt <ApyIIfIILp, 1 <p<oo.
f f(x-y(t))
-
A general theorem in L2 for curves which are approximately
homogeneous was obtained by Weinberg (We].
III. Maximal functions and g-functions

We turn now to a discussion of maximal functions. We are especially
interested in how the Fourier transforms and g-functions may be used as
a tool to relate our maximal functions to more classical ones. The story
began with the study of maximal functions along the curve (t,t2). Thus
we wish to consider averages
h
70) Mhf(x,y) -- r f(x-t,y-t2)dt .
F 0
After much frustration it was decided to take Fourier Transforms and

try to see if anything could be learned. It is easy to see that
Mhf(e,r1) = mh(e,r))f(e,rl)
where
h
eit e eit2 'Idt
mh(e,r1) = h
0
Now mh(e,rl) cannot be evaluated explicitly. If one hopes to gain some

insight by staring at a formula, one should have a formula that is as
explicit as possible. Now a similar situation arose in the path integral
approach to Quantum Mechanics. See [FHI. Feynman and Hibbs wished
to have an explicit expression for the probability amplitude that a particle
lies in a sphere of radius t. In essence, they had to consider an integral
in Rn of the form
f exp Q(x)dx
I. J<F
where Q(x) was a quadratic function of x . Instead they considered
(' a
Ixl2
+Q(x)
e J e F dx
R
392 STEPHEN WAINGER
which could be calculated explicitly. This suggests to consider instead

of 70)
71) vh * f(x,y)
f exp( t2)f(x-t,y-t2)dt
h/
.
Then
72) vh (e,q) = v(he,h2rl) f (C,rl)
where v is the measure considered in 23). In particular
00
v(he,h277) = f e t2eiheteih2r)t2dt ,
and this integral can be computed explicitly by completing the square. We

find
v(h,h2) = (nice smoothly decaying function) exp i

h477
2
73)
1 +h 4,772
One might guess that the appearance of the oscillatory factor
exp i h42 't is a reflection of the fact that is a singular measure.

1+h477 2
On the other hand we see that if h2r) is large
expii h27,expie77
which is independent of h. Thus one might try to write (from 72)
2
vh* f (e,n) _ (i(h,h27l)exp(_ l ((exp )f
rl // i
One might now hope that if one defines a measure vh by the formula
h(6,r!) = v(he,h2rl) exp - i

r7
vh could be dealt with by classical arguments, while
{exp 'I ) f( ,14 =

(
g is another L2 function having the same norm as f. So one
could hope that
74) lsup vh * fllL2 = IISUp vh * gllL2
< CIIgIIL2 = CIIfIIL2
Roughly speaking this works out. See [NRWM].

The proof of 74) was a hint on how to proceed. However, it depended
(as had happened before) on very special computations. What was needed
was a way to compare averages like vh to more classical averages by
using only the decay of vh and not so much the explicit expression of
vh as was used above.
Stein [Ssp] and [SH] succeeded in doing this by introducting appropri-
ate g-functions. Stein's first argument with g-functions dealt with the
averages MB f of equation 4). Recall
r
75) MB f(x) = J f(x-y)dgr(y)

r
Br
where Br is the ball of radius r centered at the origin, and dur is the
unit rotationally invariant measure on Br. We set, as before,
76) 911f(x) = sup IMB f(x)I

r>O r
394 STEPHEN WAINGER
Stein used g-functions to prove
THEOREM 3:
77) Cp,n 110

LP
if p>nn-1 and n>3.
Simple examples of the form
n-1
IxI O log' IXI
where 1 near the origin and has compact support show that p > nn1
is necessary in order that 77) hold. The situation for n = 2 , p > 2 is
unknown at this time.
I would like to present here Stein's original argument which proved
77) for p = 2 and n = 4 . We define
00 1 /2
78) g(f) (x) _

f tI dt Mtf(x)I2 dt
Assume that we could prove
79) IIg(f )IIL2 < C(n)IIf1IL2 '
and let us see how 77) would follow. Now
rnMrf(x) CT- snMsf(x)ds

0
=n r sn-1 Msf(x)ds + r sn ds Msf(x)ds

Thus
r r
Mrf(x) < n r sn'1 Msf(x) ds + n r sn d Msf(x)ds
0 0
= I(r) + 11(r) .
Now l(r) is dominated by the Hardy-Littlewood Maximal function and
II(r) < n
rJ f
0
sn-1 /2 s1 /2 Msf(x)ds
r 1 /2
< fr
1 s2n-1 ds g(f) (x)
-rn J 0
<Cg(f)(x).
So if we assume 29), we have
Ilsup Mrf(x)IIL2 <- CIIf1IL2
We turn now to the proof of 79).
Ig(f)(x)I2dx= J t J IdtMtf(x)I2dxdt
JRn 0 Rn
0
=
J0
r t Jr Mtf(6)12d6dt
Rn
00
_ tJ Imtr(6)I2dedt .
0 Rn
396 STEPHEN WAINGER
But Mtf(e) = f(()m(tlel), where
m(r) = n12 Jn-2(r)

r 2 2
Here Jn_2 is the usual Bessel function. We shall need to know

2
dmr I < C 1 =C 1
dr n-2 n-1
r 2 r1/2 r 2
and dorm
is bounded. See [SWE]. Thus
f lg(f)(x)I2dA < f f Jtm(t()Ide

0
00
Rn
00
< f If( )12 5 t at m(t 1j1)2 dt

Rn 0
Thus to obtain 79) we need to show
00
r tIdm(tIibI2dt<C
0
First since is bounded,
f
0
1/161
dt <1e12f
0
1 /IeI
tdt<C.
Next, since Im'(t)I < C ,

n-1
t 2
00
dt m(tIeI)I2 dt
t
/IeI
00
Ifi2 f t )n-1
dt < C ,
(t lel
1 /ICI
if n>4.
Let us be more precise about the counterexample in 2-dimensions. We
take
1
x very near 0
IxIlogI1
in Co away from 0 .
We can disprove 77) for p = 2, n = 2 by showing
80) MBIxIf(x) _ 00
for all small x.

Because of rotational symmetry it suffices to prove 80) for points
(a,0) with a small. In that case
dO
MIxlf(x) _
(a2(1-cos)2+a2sin2ej1 /2 log 1
-n a2[(l-cos)2+sin2O]
ti dO = 00 .
IeI In IeI
-E
398 STEPHEN WAINGER
A similar argument shows
81) sup IMB f(x)I = 00

1<r<2 r
a.e. for an appropriate f .

We wish now to prove a theorem indicating how bad 81) fails in
L2(R2). We set
82) )Rkf(x,Y) = sup IMB

f(x)YI
2k<j<2k+1
2k
where f is in L2(R2). We shall show
83) Il)KkfpL2 < CkIIfIIL2
To prove 83), it suffices to show that for each function j(x) taking
the values
84) IIMBI+j(x)2-k f(x)II < CkpfIIL2 ,
with C independent of the function j(x). We show 85) by induction on

k. That is given a j(x) taking the values we shall define
a function j*(x) so that j*(x) takes values in the set
and
85) IIMBI+j(x)2-kf fIIL2 < CIIfIIL2

_MB1+j*(x)2-k+l
85) provides the inductive step to prove 86). If j(x) is given define
if j(x) is even
if j(x) is odd .
Then, if we set
2k
86) gf(x) _ I IMB f(x)-MB f(x)I2
J=0 k
1+ 1+11
2 2k
we see
IMBfI < gf(x)
Thus to prove 86) and hence 85) it suffices to prove
87) 119(fAL2 < CIIfIIL2
We shall prove 87) by using the Fourier Transform.
(' 2k
J Igf(x)I2 dx = I fMf(x)-MB
1 0 l+2k l+-
2k
=I IMB W) - MB
j=0
f 2
where m(r) = J0(r).

So to prove 87) it suffices to show
88)
400 STEPHEN WAINGER
But 88) follows because for s positive and r positive
89) IJ0(r)) < C/'
and
90) IJ0(r+s)-Jo(r)I < Ct.
To prove 88) we use 89) if ICI > 2k and 90) if 161 < 2k
Finally we will show how Stein [SH] proved the maximal function
along the parabola (t,t2) is bounded by using g-functions.
We start with the measure d defined in 24)
2
qS(t,t2)dt
d(O) = J .
We set
91) dh(qS) = f -O(ht,ht2) dt .
Then
dh * f(x,y) = r f(x-ht,y-ht2)dt ,
or
92) dh * f(x,y) _ f 2h
f(x-t,y-t2)dt .
We choose a function &i(x,y) E C0m(R2) with %&(0) = 1. We set

93) 'Ph(x,Y) = h3 0 x , h)
and
94) g(f)(x,Y) _ I Idh *f(x,Y)-Oh *f(x,Y)I
Let us first assume
95) 119(f)IIL2 5 C IIf1IL2
Note that
sup
E>0
f Idh*f(x,Y)-h*f(x,Y)Idh
0
sup 1f
E>0 E1/2
E
Idh *f(x,Y)-oh *f(x,Y)I2dh

0
1 /2
< r Idh *f(x,Y)I2 dh
0
< g(f) (x,Y)
So
96) sup I E f dh *f(x,Y)dh1

t>o
0
<- Of )(x,Y) + sup 10h *f(x,Y)I

h>0
402 STEPHEN WAINGER
A classical argument (see (RI) shows
sup Oh *fp < C tIfIILp

h>0 Lp
Thus by 95), we see
sup I e fd1zh*f(xY)dhl < CIIf11L2

E>0
0 L2
If f>0,
E
('
J dh*f(x,Y)dh
0
E 2h
f
0 h
f(x-t,y-t2)dtdh
E E
> f r f(x-t,y-t2)
0
f/2
dh
('E
>
E
f
J
0
f(x-t,y-t2)dt .
So from 96) we infer
sup If f(x-y,y-t2)dt <AIIf1IL2

E>0
eJ0 L2
It remains to prove 96).

fI(gf)(x,y)I2dxdY
= r T J Idh*f-ch*fl2dxdy
0
ff
0
00
rd h(e,-0)h(e,7)I2 If( ,77)I2 d dii
f
a*
.77)12
f0,1f( dry.
0 0
So to prove 96) it suffices to prove
00
97) J I h( ,rl)I2 < C .

0
The integral on the left side of 97) is
98) JT
0
Thus by replacing h by Ah A > 0 we may write the expression in

98) as
0
(Xhe,)12h271)-0(Ahe,42h277)I2
J0
By choosing A so that \2e2+X4772 = 1 , we see that it suffices to esti-
mate 99) when e 2 +772 = 1 . In this case we see
1 1
('
99) JJ dh d(he,h27)-cb(he,h27)I2 < C r h2 dh < C
0 0
since d(0) = 0(0) = 1 .

404 STEPHEN WAINGER
Then from 27)
Id(he,h27l)I < Ch-3
for some 6>0. So

('00 ('00
100) J T Id(he,h2rl)I < J ld26 < C

J
1 1
h
CN
Also k1;'"?) < for any N, so
(1+e2+772)IJ
00
('
101) 1 dh Ii&(he,h277)I < C .
Now we obtain 98) and hence 96) by combining 100), 101) and 102).
In this section we have emphasized L2 methods. LP results for
p > 1, can be obtained by combining the L2 estimates presented here
with the techniques of section 2. Altogether one can prove the following
theorems :
THEOREM 4. If y(t) satisfies y'= A y(t) where all the eigenvalues of

A have positive real part,
IIo<np Ff If(x-y(t))IdtIILp<C IIfIILp, 1<P<

<00
THEOREM 5. Let y(t) be a curve in Rn. If the vectors

y'(0), y"(0), y(3)(0) ... span R ,
f
h
II sup If(x-y(t))IdtII < CpIIfIILP 1<p<

0<h<1 LP
0
IV. Vector field problems

We turn now to problems III and 1II'. To study problems III and III' it
is convenient to make a change of variables. It is possible to make a
change of variables so that the integral curves of the vector field v(x)
become lines parallel to the x-axis. Under this change of variables the
vectors v(x) transform into curves y which vary from point to point. So
we are led to studying problems IV and IV'. We facilitate the statement
of these problems with 3 definitions.
If y(x,t) is for each x a smooth curve in t with y(x,0) = 0, let
!'1
102) Hyf(x) = f f(x-y(t,x)) dt

-1
let
!'h
103) Myf(x) _ f(x-y(t,x))dt ,

0
and
104) )11 f(x) = sup Ahf(x)l .

Y
1>h>o y
We are now ready for the statement of problem IV and W.
Problem IV: When do we have
UHyf(x)11Lp <- Cpilf11Lp ?
Problem IV': When do we have
f(myf11Lp <- CpIIfhILp ?
The change of variables described above preserves tangency and

curvature conditions. So for example N It=o will be parallel to the x
406 STEPHEN WAINGER
axis, and if the curvature of the integral curves of v(x) never vanishes
a2y
will not be zero. It turns out that one can prove the following
at2 t=0
theorems:
THEOREM 6. Let y(t,x) = (t,r(t,x)) be a smooth curve in R2 satisfying
ar x
It=o = 0 and a2rlt=ono.
at
Then
105) IIHyf IIL2(R2) <- Of I!L2(R2)
and
106) IItyfIIL2(R2) <- CIIfllL2(R2)
A consequence of Theorem 6 for vector fields will then be
THEOREM 7. If v(x) is a smooth vector field in R2 such that the

integral curves of v have nowhere vanishing curvature,
IIHvf1IL2(R2) <_ IIfIIL2(R2)
and
h
lim h1 , f(x-tv(x))dt = f(x) a.e.
h-00 1
0
for f in L2(R2).
These theorems are announced in NSWV. Here we shall give some
discussion of the ideas. Because y(t,x) depends on x, the Fourier
Transform no longer seems like a good tool to study Hy and 59 y. We
must find a different way to employ orthogonality. Here we're motivated
by work of Kolmogorov and Silveristov [Z] on the partial sums of Fourier

Series and an approach to the Poisson Integral by Paley [P]. The idea is
to consider Hy H* and Mh(x) . (Mh(x))* where denotes composition
of operators and * signifies Hilbert space adjoint. In order to gain some
insight into this method we shall just discuss it in the case y(t) = (t,t2)
where, of course, we already know the results by a different method.
Let us consider K = Hy H*. The kernel of K will have support on
t(u,v)Iu=t-s,v=t2-s2J. Hence one might hope that K would have a much
smoother kernel than Hy H. (Of course if K were bounded on L2 so
would Hy.) One might hope then, that K would be a Calderon-Zygmund
operator. Let us see if this could be the case. One can see that
H*f(x,y) = J f(x+t,y+y(t)) dt
and hence that
HYH*f(x,y) f(x+t-s,y+t2-s2) dt
tss
=ff
Let u=s -t, v =s2-t2, and we find
HYH*f(x,y) =ff f(x-u,y-v)k(u,v)dudv
where
k(u,v) = C .
IuI u-u u+u
Now k(u,v) does have its support spread out. However the singularities
of k across the curves v = u2 are not locally integrable away from the
408 STEPHEN WAINGER
origin. Hence HYHY cannot be a Calderon-Zygmund operator. Let us

try to see what goes wrong on the level of the Fourier Transform. Recall
HYf=Qp*f
where
p(e,r/) _ f e(iet+ir7t2)dt
-00
We know from 48) that
Qp((,ri) = sgn + C IjII/2 e 77
6
+ better terms.
The multiplier for HY HY would be essentially IQp12 . Notice that

IQp12 doesn't look any nicer than Q. However IQp(e,r7)-sgneI2 is
much nicer than Qp(e,rj) or Qp(e,rj) - sgne. Now sgne corresponds
to the operator
Lf = ff(x-t,y) dt
L is known to be bounded in L2. This suggests that we try to consider
M = (L-HY)(L*-HY) .
If M is bounded in L2 so will be L - HY . Hence so will be HY . This

actually works and is the basic idea in the proof of Theorem 4. We turn
now to the idea of the proof of Theorem 5. Paley showed that
107) Ph(x) (Ph(x))*f(x) < CIPh(x)f(x)+P*(x)f(x)]
where Ph is the Poisson Kernel. It follows from 107) that

IIPh(x)f Il L2 < Clip h(x)f(X)ll

L2
and hence
IIPh(x)f1IL2 <C .
Since the function h(x) is arbitrary we have
Ilsup Ph*f1IL2 <_ CIIf1IL2

h
The fact that we had to modify Hy suggests that we should not expect
107) to hold, but we might expect a variant to hold - perhaps involving an
operator
h(x,y)
(x,Y) = f(x-t,Y)dt .
Rh(x,Y)f h(x,y) f
0
(We know
IIRh(x,y)f(x,Y)IIL2 <- Cllf(x,Y)lIL2 )
We might hope to prove
108) Mh(x)(Mh(x))*f(x) < C(Rh(x)Mh(x)f(x)+Mh(x)Rh(x)f(x)+Bh(x)f(x))
where Bh(x) is some bounded operator. Even 107) is not quite right.
We refer the reader to NSWV for the correct technical modification of 107).
This concludes our discussion of the vector field problem.
V. Recent developments
The positive results of Theorem 2 and Theorem 5 assume that the
curve y(t) has some curvature at the origin. There have been a number
of papers trying to understand what happens when this curvature condition
is dropped. See [C], [CNVWW], [NVWW] 1), 2), 2), and [NE]. Let us note
that we don't have positive results for all C curves.
410 STEPHEN WAINGER
Suppose for example y(t) is odd and y(t) _ (t,r(t)) where
0 for 0<t<1
r(t) =
t-1 for t>1
Then
nHrf(6,71) = my(6,17)f(6,q)
where
00
my(4,i7) = r ei4teirir(t) dt
-00
= r
1
sintt dt + 00
at
0 1
which is easily seen to be unbounded if In fact it is easy to see

the following: Suppose y(t) is odd and linear on a sequence of
intervals ai < t < bi < 1 and assume that the linear extension of y on
b
[ai,bi] does not pass through the origin. Then if the ratios a are
t
unbounded,
E--0
lim I f(x-y(t)) dt
ItI<E
cannot exist in the L2 sense for every f in L2 .

It is somewhat more difficult to produce counterexamples for the
operator 1R y. For example in the case of the two dimensional curve
(t,r(t)) described above, it is easy to see that Vy is bounded in every
LP, because it is dominated by
!'h h
sup I
0
If(x-t,y)1 dt + sup h
f
0
f(x-t,y-t+I)dt t .
Both of these operators are bounded by the classical theory. We refer to

[SWI for an example to show that the maximal function along a Coo curve
has no non-trivial positive results.
The above authors have been investigating the behavior of the Hilbert
Transform and maximal functions related to convex curves. Let y(t) _
(t, F(t)) a plane curve with r(t) convex for positive t. Let h(t) =
tr'(t)- F(t). -h(t) represents the y-intercept of the line tangent to the
curve y at y(t). We then have
THEOREM 8. Let y(t) = (t,P(t)) with r(t) convex and increasing for
t > 0. If r(t) is even
IIHyfIIL2 <- CIIflIL2
if and only if
Ir'(ct)I > 21r'(t)I t>0
for some C>0.

If y(t) is odd
IIHyfIIL2 <- CNfIIL2
if and only if
h(Ct) > 2h(t) , t>0
for some C>0.
There are generalizations of Theorem 8 to higher dimensions, and an

investigation of the LP theory has begun. We know that
412 STEPHEN WAINGER
II yfliL2 <- CIIfIIL2
if * holds for some C > 0, however the maximal functions can be

bounded on L2 (and in fact in LP for any p > 1) for some convex
curves even if * fails.
Recently Phong and Stein [PS] introduced a general problem of which
our problems are special cases. They consider at each point P in Rn
a submanifold Mp of dimension say a and an a dimensional Calderon-
Zygmund kernel K(P,Q). Then they consider
Tf(P) T (Q) K(P,Q)dm(Q)
where dm(Q) is a measure on Mp M. They show that if n > 3 and

k = n-1
IITfIILp <_ CIIfIILP
if Mp satisfies a kind of generalized curvature condition. See [PSI.

Various authors have also considered multiple parameter problems
which are essentially multiple Hilbert transforms on surfaces and multi-
parameter maximal functions on surfaces. See [NW2], [V], [STR1], and
[CSS].
Appendix 1. An introduction to the method of steepest descents.

Here we shall try to give an explanation of the main ideas of the
method of steepest descent. The interested reader can find a more
detailed description in [B].
Let us first consider the behavior of the integral
00
A-1) 1(I) = f e ax2 dx

z
for large A. Of course we can make a change of variables
and observe
A-2) I(A) _
V"X
where
00
e-t2
A-3) B= dt .
-00
The point we wish to make here is that if A is large most of the

contribution to the integral I(A) comes from a small neighborhood of the
origin. In fact
A__t2
('
5e e
At2 dt <
fA21/5<t<1
At2 dt + J e 2
t>1
2 dt
It1,A21/5
<e-AI/5+e,\/2
-gyp very fast as A. o .
Thus the main contribution to the integral I(A) comes from the small
interval - 1 < t < 1 . Now if we perturb the integrand in I(A), we
A2/5 - -A2/5
can expand the integrand in a power series in that little interval. For
example we might consider
00
JA) = e Xh(t)dt
J
414 STEPHEN WAINGER
where h(t) > t2 for large t, h(t) = t2 +0(t3) for small t and h(t) > 0
for t > 0.
Then one can easily see that
f dt
Itl>A2/5
is exponentially small as before. Now
e Ah(t)dt
J = f e-At2 . (1+O(at3)
ItIU 2/5 Itl
00
e-'fit2
dt + Ea I /5 e-at2 dt
ltl2/5 -00
B + O(e-AI /5)+0
(1 l
f
= .
In the method of steepest descents we try to choose a contour of

integration so that on the new contour the situation would be essentially
that of J. Thus for example, if we had
K= r eikt2+.k P(t)dt
-00
and P(t) were very negative at infinity and 0(t3) near t = 0 we would
try to write t = a + it and integrate on the line a = r for It l < 1/, 2 /5 ,
and we would expand in a power series in this small interval.
More, generally, if we were concerned with the asymptotic behavior

for large values of k of an integral of the form
fg(z)eAh(z)dz
g(z) and h(z) are holomorphic, we would try to choose a contour on

which Re h(z) had only a finite number of maxima, and argue that the
main contribution to the integral should come from a small neighborhood
of the largest maximum or perhaps an endpoint and we then expand g and
h in a Taylor series at such points. At such a maximum, e, h'(C) = 0.
Thus the main contribution to our integral should come from a point where
h'(C) = 0 or an endpoint. This is also reflected in Van Der Corput's
lemma. A variant of this principle for non-analytic functions is called the
principle of stationary phase and is discussed in Professor Stein's
lectures in these proceedings.
Appendix 2. The method of stationary phase and quantum mechanics

The method of stationary phase lends itself to a formulation of quantum
mechanics that is very appealing to at least some mathematicians. The
principle of stationary phase asserts that the integral
b
I= r eiAf(t)dt
a
with f(t) real gets most of its contributions for large A, near a, b, or a
zero of f'(t).
If we had no endpoints for example if f were periodic with period
b - a or if the interval of integration was from -oc to oo and f oscil-
lated very rapidly for large t , we would expect the main contribution to
the integral to come from small neighborhoods of a zero of V.
416 STEPHEN WAINGER
Let us now turn to quantum mechanics. In particular let us consider a

particle moving from a point xa to xb as time evolves from time ta to
time tb. According to classical physics, the particle will follow a path
for which the classical action is stationary. If x(t) is any path, and our
particle has mass m and is moving under the influence of a potential
V(x,t), the classical Lagrangian is defined by
L = 2 [x(t)]2 - V(x(t),t) .
The action along a path x is defined by
tb
S(x) = r L(x(t))dt .
ta
The path on which the classical particle moves will be a path x(t) such
that z(ta) = xa , z(tb) = xb , and such that
S(x +Sx)IS=o 0.
=
One of the most important principles of quantum mechanics asserts that

motion on a classical scale must be essentially described by the laws of
classical mechanics. In our example it means the only paths that are im-
portant are paths near the path x of *. This principle is expressed in
terms of a small number h. The principle says that the only classical
paths that should be important are paths which differ from x only on an
h scale of measuring. Now the Feynman path integral formulation is in
terms of probability amplitudes of events. In our case the probability
amplitude of passing from xa at time ta to xb at time tb is given by
an integral
s(x)
Fje"h Dx
where the integration is an integral over all paths x(t) such that
x(ta) = xa and x(tb) = xb. Leaving aside the question of how such an
integral can be precisely defined, let us try to guess what paths contribute
the most to the integral F . Since h is very small, the principle of
stationary phase would indicate that the main contribution to the integral
F should come from a small neighborhood of a path of which some kind
of a derivative of S was zero. Thus we might expect the main contribu-
tion to the integral F to come from paths which are very close (on a
scale of h ) to the path x defined by *.
There have been many papers in the mathematical and physical litera-
ture dealing with the problem of making sense out of the definition F.
See for example [CS] and references cited there. However, the original
definition in [FH] serves the purpose of making many formal calculations.
We may imagine dividing the t interval into 21 subintervals, Ij , of
equal length
[b_ta]k
Ik = + ta, [tb-ta]2 + to
l nl .
We consider only paths which are linear on Ik .

Such paths are determined by
xk = X (ta + [tb-ta])
2
k= We then take
1 S(x1 ... x2j-1)

F = lim A. eh dx 1dx2,...,dxk
J
j-,
where S(x1,...,x2j_1) is the action along the polygonal path determined

by x1'...,x2j_1 , and Aj is a normalizing factor. For more details see
[FH]. We would like to make one final remark about the book [FH]. It's
418 STEPHEN WAINGER
great. It explains quantum mechanics in terms of mechanics and does not

use notions of atomic physics as many of the standard books do. Thus it
is accessible to many more mathematicians than standard quantum
mechanics texts. The book also elucidates the differences between the
nature of physicists and mathematicians. If you don't want to know h to
3 significant figures in ergs/sec, whatever they are, you probably would
rather be a mathematician than a physicist. Finally, many mathematicians
could probably learn a great deal to improve themselves as mathematicians
by reading the book.
STEPHEN WAINGER
UNIVERSITY OF WISCONSIN
MADISON, WISCONSIN
REFERENCES
[B] N. DeBruijn, Asymptotic Methods in Analysis, North Holland
Publishing Co., Amsterdam, 1958.
[BF] H. Busemann and W. Feller, "Zur Differentiation des
Lebesguesche Integrale," Fund. Math, Vol. 22, 1934,
pp. 226-256.
[CS] R. Cameron and D. Storvick, A simple definition of the Feynman
integral with applications, Amer. Math. Soc., Providence, 1983.
[CSS] H. Carlsson, P. Sjogren, and J. Stromberg, "Multiparameter
maximal functions along dilation-invariant hypersurfaces" to
appear in Trans. of the A.M.S.
[CW] H. Carlsson and S. Wainger, "Maximal functions related to con-
vex polygonal lines," to appear.
[C] M. Christ, preprint.
[CS] J. L. Clerc and E. M. Stein, "LP multipliers for non-compact
symmetric spaces," Proc. Nat. Acad. Sci., U.S.A., Vol. 71,
1974, pp. 3911-3912.
[CORI] A. Cordoba, "The Kekeya maximal function and the spherical
summation multipliers," Amer. J. of Math., Vol. 99, 1977,
p. 1-22.
[COR2] "Maximal functions, covering lemmas and Fourier

multipliers," Proc. Symp. in Pure Math., Vol. XXXV, Part I,
1979, pp. 29-50.
[CF1] A. Cordoba and R. Fefferman, "A geometric proof of the strong

maximal theorem," Annals of Math., Vol. 102, 1975, pp. 95-100.
[CF2] , "On differentiation of integrals," Proc. Nat. Acad.
Sci., U.S.A., Vol. 74, 1977, pp. 2211-2213.
[CF3] "On the equivalence between the boundedness of
certain classes of maximal and multiplier operators in Fourier
Analysis," Proc. Nat. Acad. Sci., U.S.A., Vol. 74, 1977,
pp. 423-425.
[CNVWW] A. Cordoba, A. Nagel, J. Vance, S. Wainger, and D. Weinberg,

"LP bounds for Hilbert Transforms along convex curves,"
preprint.
[F] E. B. Fabes, "Singular integrals and partial differential equa-
tions of parabolic type," Studia Math., Vol. 28, 1966,
pp. 81-131.
[FEF] R. Fefferman, "Covering lemmas, maximal functions, and multi-
plier operators in Fourier Analysis," Proc. Symp. in Pure Math.,
Vol. XXXV, Part 1, pp. 51-60.
[FH] R. Feynman and A. Hibbs, Quantum Mechanics and Path
Integrals, McGraw Hill, New York.
[H] L. HOrmander, "Estimates for translation invariant operators in
LP spaces," Acta Math., Vol. 104, 1960, pp. 93-139.
[KS] R. Kunze and E. Stein, "Uniformly bounded representations
and harmonic analysis of the 2x2 unimodular group," Amer. J.
of Math., Vol. 82, 1960, pp. 1-62.
[NRW] A. Nagel, N. Riviere, and S. Wainger, "On Hilbert transforms
along curves, It, "Amer. J. Math., Vol. 98, 1976, pp. 395-403.
[NRWM] A. Nagel, N. Riviere and S. Wainger, "A maximal function
associated to the curve (t,t2)," Proc. Nat. Acad. of Sci., U.S.A.,
Vol. 73, 1976, pp. 1416-1417.
[NSWB] A. Nagel, E. Stein, and S. Wainger, "Balls and metrics defined
by vector fields I; Basic Properties," to appear in Acta
Mathematica.
[NSW] A. Nagel, E. Stein, and S. Wainger, "Boundary behavior of
functions holomorphic in domains of finite type," Proc. Nat.
Acad. Sci. U.S.A., Vol. 78, 1981, pp. 6595-6599.
[NSWV] A. Nagel, E. Stein, S. Wainger, "Hilbert transforms and maximal
functions related to variable curves," Proc. of Symposia in
Pure Math., Vol. XXXV, part 1, 1979, pp. 95-98.
[NVWW1] A. Nagel, J. Vance, S. Wainger, and D. Weinberg, "Hilbert
transforms for convex curves," Duke Math. J., Vol. 50, 1983,
pp. 735-744.
420 STEPHEN WAINGER
[NVWW2] A. Nagel, J. Vance, S. Wainger, and D. Weinberg, "The

Hilbert transform for convex curves in Rn," to appear in
Amer. J. of Math.
[NVWW3] - , "Maximal functions for convex curves," Preprint.
[NW] A. Nagel and S. Wainger, "Hilbert transforms associated with
plane curves," Trans. Amer. Math. Soc., Vol. 223, 1976,
pp. 235-252.
[NW2] -, "L2 boundedness of Hilbert transforms along
surfaces and convolution operators homogeneous with respect
to a multi-parameter group," Amer. J. of Math., Vol. 99, 1977,
pp. 761 785.
[NE] W. Nestlerode, "Singular integrals and maximal functions
associated with highly monotone curves," Trans. Amer. Math.
Soc., Vol. 267, 1981, pp. 435-444.
[P] R. Paley, "A proof of a theorem on averages," Proc. Lond.
Math. Soc., Vol. 31, 1930, pp. 289-300.
[R] N. Riviere, "Singular integrals and multiplier operators," Ark.
Mat., Vol. 9, 1971, pp. 243-278.
[PS] D. Phong and E. Stein, "Singular integrals related to the Radon
transform and boundary value problems," Proc. Nat. Acad. Sci.,
U.S.A., Vol. 80, 1983, pp. 7697-7701.
[S] E. Stein, Singular Integrals and Differentiability Properties of
Functions, Princeton University Press, 1970.
[SBC] , Boundary Behaviour of Holomorphic Functions of
Several Complex Variables, Princeton University Press,
Princeton, 1972.
[SHI , "Maximal functions: Homogeneous curves," Proc.
Nat. Acad. Sci., U.S.A., Vol. 73, 1976, pp. 2176-2177.
[Ssp] , "Maximal functions: Spherical means," Proc. Nat.
Acad. Sci., U.S.A., Vol. 73, 1976, pp. 2174-2175.
[SPL] , Topics in Harmonic Analysis related to the Littlewood-
Paley Theory, Princeton University Press, Princeton, 1970.
[SI] , "Interpolation of linear operators," Trans. Amer.
Math. Soc., Vol. 88, 1958, pp. 359-376.
[SWA] E. Stein and S. Wainger, "The estimation of an integral arising
in multiplier transformations," Studia Math., Vol. 35, 1970,
pp. 101-104.
[SW] , "Problems in harmonic analysis related to curvature,"
Bulletin of the A.M.S., Vol. 84, 1978, pp. 1239-1295.
[SWE] E. Stein and G. Weiss, Introduction to Fourier Analysis on

Euclidean Spaces, Princeton University Press, Princeton.
[STR1] R. Strichartz, "Singular integrals supported on submanifolds,"
Studia Math., Vol. 74, 1982, pp. 137-151.
[STR] J. Stromberg, "Weak estimates on maximal functions with
rectangles in certain directions," Ark. Mat., Vol. 15, 1977,
pp. 229-240.
[STRO] J. Stromberg, "Maximal functions associated to rectangles
with uniformly distributed directions," Ann. of Math., Vol. 107,
1978, pp. 399-402.
[V] J. Vance, "LP boundedness of the multiple Hilbert transform
along a surface," Pacific J. of Math., Vol. 108, 1983,
pp. 221-241.
[WE] D. Weinberg, "The Hilbert transform and maximal function for
approximately homogeneous curves," Trans. Amer. Math. Soc.,
Vol. 267, 1981, pp. 295-306.
[Z] A. Zygmund, Trigonometric Series, Vols. I & II, Cambridge
University Press, London, 1959.
INDEX
approach regions convex curves, 411

admissible, 245 curvature, 321, 361
non-isotropic, 261 and Fourier transform, 321, 325,
non-tangential, 244 268, 375
Ap classes, 73, 353
area integral, 92 DeGiori-Nash regularity theory,
144, 158
atomic decomposition, 114, 156,
159, 340 Dirichlet problem, 132, 243
domain of holomorphy, 211
Bergman kernel, 230 duality of H1 and BMO, 114, 340
B.M.O. (Bounded mean oscillation),
9, 94, 331, 340
electrostatics, 163
BMO(R+xR+), 101
exponential mapping, 267, 275
Bochner-Martinelli formula, 196
Bochner-Riesz summability, 344
Fatou's theorem, 245
finite type, 269, 283, 324
Calder6n-Zygmund decomposition,
48 Fornaess imbedding theorem, 226
Campbell-Hausdorff formula, 267 functional calculus, 27
canonical coordinates, 275
Carleson measure, 16, 94 g-functions, 393
Cauchy-Fantappie formula, 195
Hardy space, 144
Cauchy integral, 5, 8, 186
on a Lipschitz curve, 143 Hp spaces, 89
HP(R+ x R+) , 101
Cauchy-Riemann equation, 201
harmonic measure, 141
convergenge of averages over
spheres, 358, 361 Hartogs extension phenomenon, 207
along curves, 361
along vector fields, 361, 406 heat operator, 254
covering lemmas, 60 Henkin integral formula 219, 336
423
424 INDEX
Heisenoerg group, 257, 335 non-commuting vector fields, 267

Hilbert transforms along curves,
361, 372 oscillatory integrals (first kind),308
along vector fields, 362,406 (second kind), 335, 344, 349
hydrostatics, 177
hypoelliptic differential opera- Poisson integral, 187, 243
tors, 281 bi-Poisson integral, 102
Kohn (canonical) solution, 209 Rellich-type formulas, 150, 166

Laplacian, 266
restriction theorems, 325, 344
Korn-type inequalities, 145
Korteweg-de Vries equation, 25 Sobolev estimates, 145
space of holomorphy, 42
Laplace equation, 132, 242
space of homogeneous type, 251
Leray form, 194
stationary phase and quantum
Levi-pseudoconvex, 214, 257 mechanics, 415
polynomial, 227
Stein-Weiss spaces, 96, 160
Lipschitz domain, 133, 145
steepest descents, 412
Littlewood-Paley-Stein theory,
48, 53 Stokes theorem, 189
local singular function, 210 surfaces (non-zero curvature), 321
Lu Qi-Keng conjecture, 235 systems of elliptic equations, 133,
163
maximal functions, 48, 49, 245 Szego kernel, 193, 265, 296, 297
strong, 60
spherical, 359, 393 Transference theorem, 39
on curves, 391, 400
on vector fields, 406
method of layer potentials, 133, 143 Van der Corput lemmas, 309, 370
Mobius transformation, 186
weight norm inequalities, 72
multilinear Fourier analysis, 18
multiparameter differentiation Zygmund conjecture, 67
theory, 57
multipliers, 72
Library of Congress Cataloging-in-Publication Data
Beijing lectures in harmonic analysis.
(Annals of mathematics studies ; no. 112)
Bibliography: p.
Includes index.
1. Harmonic analysis. I. Stein, Elias M.,
1931- . H. Series.
QA403.B34 1986 515'.2433 86-91452
ISBN 0-691-08418-1
ISBN 0-691-08419-X (pbk.)
Elias M. Stein is Professor of Mathematics at

Princeton University

Beijing Lectures in Harmonic Analysis-Stein-0691084181

Enviado por

Dados do documento

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Beijing Lectures in Harmonic Analysis-Stein-0691084181

Enviado por

Direitos autorais:

Formatos disponíveis

Annals of Mathematics Studies

PRINCETON UNIVERSITY PRESS

PRINCETON, NEW JERSEY

The Annals of Mathematics Studies are edited by

Clothbound editions of Princeton University Press

ISBN 0-691-08418-1 (cloth)

ISBN 0-691-08419-X (paper)

Printed in the United States of America

Library of Congress Cataloging in Publication data will

MULTIPARAMETER FOURIER ANALYSIS

ELLIPTIC BOUNDARY VALUE PROBLEMS ON LIPSCHITZ

INTEGRAL FORMULAS IN COMPLEX ANALYSIS

VECTOR FIELDS AND NONISOTROPIC METRICS

OSCILLATORY INTEGRALS IN FOURIER ANALYSIS

AVERAGES AND SINGULAR INTEGRALS OVER LOWER

In September 1984 a summer school in analysis was held at Peking

Our purpose is to describe a certain number of results involving the

where Ak(f) is a "homogeneous polynomial" of degree k in f. This

(linear in each argument) such that Ak(f) = Ak(f, f, f) and

(1) IIAk(fI...fk)IIB2 < Ck

for some constant C . (This last estimate guarantees the convergence of

Certain facts can be easily verified. In particular if F is analytic

aij(x) = div A(x) grad, A _ (aij) x f Rn

the coefficients a(x) (or aij(x) ) will be assumed to belong to some

L-1 , fL sgnL, e-tL , e-t\/L

or more generally, c(L) (where (k: C C ), can be defined as a bounded

Many questions arise:

If we define h(x) = x+A(x), A'(x) = a. We then have

Of course, in this case, if we use the Fourier transform we can define

(I_ d)f r eix&O(e)f(e)de

This gives, for example

Thus we can define

(where we used the observation that 95(ULU-1) = U(L)U-1 ).

F(a) = sgn L. as an operator on L2(R)

and wish to know whether it is analytic on L or if we can replace a

(' f(t) (1+Bi(t))

f1(t) = f(t) i+a Bi(t) B1 = Bh-1

Since Uh is bounded on L2 it would suffice to prove that C is

-irr sgn(La)f = J (xf t)+A(x)-A(t) dt = C' (-1)k r

T(f) = ft (Ax_ A t) f(t) dt = fk(xt) f(t)dt

We will prove Theorem I: Let T a Cc(C) and A(x) such that

Then the operator T is bounded on L2(R) (and LP 1 < p < oe ). This

as we can see its boundedness in L2 is equivalent to the analytic

Let us consider a more general version of the Cauchy integral.

The Cauchy integral on r is given as:

cr(f) = p.v. f f t z' t dt

rJ 0rz` ss-t-z t/I fs-tt) dt

we can take c e C o (C)t/i(z) = on 8 < lz l < 1 and obtain the bounded-

All the operators which we encountered previously had the form

T(f) = p.v. f K(x,y)f(y)dy

where IK(x,y)I < C IaxKI + Ia KI < Ix-y12

IIbII* = uIII fI where mi(b) =III I b(x)dx

and I is an interval (or a cube in Rn ), and that this norm is equivalent

where '/it = 0 (t) 0 C o C. fcidx =0 (ci a'0) (see [51).

PROPOSITION. If T is as above and T is bounded on L2 then T

Proof. Let b c L and let I be given. Consider I = 21 and write

The first term is dominated by

For the second we observe that

T(b2)(X) - T(b2)(u) = I f[K(x,y) - K(u,Y))b2(y)dy

IT(b2)(X)-mI(T'(b2))I < CIIbII,,.

Let k eCU(RI) with f6dx = 1 and =f(x). Let ft(u) _

l<T'&r,Ot>I <CPt(x-y) =Ct 1

In fact, assume for simplicity, that S6 is supported in (-1,1). Since

I<Tqx, (kt >I = fr(z){j'K(z ,u) dz

Also observe that Lt(l) = f et(x,y)dy = <Tcb ,1 > = At T(1) = Pt*b

f F'(f)(btf btdt+e =rr(f;F'(f))+e

J to * (iAtb 95t fdtt