Lecture Notes in MathematicsLecture Notes in Mathematics Editors:
A. Dold, Heidelberg F. Takens, Groningen
1684
Advisor: Roberto Conti
Springer Berlin Heidelberg New York Barcelona Budapest Hong Kong
London Milan Paris Santa Clara Singapore Tokyo
C. A. Berenstein R E Ebenfelt S.G. Gindikin S. Helgason A.E.
Tumanov
Integral Geometry, Radon Transforms and Complex Analysis Lectures
given at the 1 st Session of the Centro Internazionale Matematico
Estivo (C.I.M.E.) held in Venice, Italy, June 3-12, 1996
Editors: E. Casadio Tarabusi, M. A. Picardello, G. Zampieri
Fondazione
C.I.M.E.
Springer
Authors
Carlos A. Berenstein Institute for Systems Research 221 A. V.
Williams Building University of Maryland College Park, MD
20742-0001, USA
Peter F. Ebenfelt Department of Mathematics Royal Institute of
Technology 100 44 Stockholm, Sweden
Simon Gindikin Department of Mathematics Hill Center Rutgers
University New Brunswick, NJ 08903-2101, USA
Sigurdur Helgason Department of Mathematics Massachusetts Institute
of Technology Cambridge, MA 02139-4307, USA
Alexander Tumanov Department of Mathematics University of Illinois
1409 West Green Street Urbana-Champaign, IL 61801-2943, USA
Editors
Massimo A. Picardello Dipartimento di Matematica Universith di Roma
"Tor Vergata" Via della Ricerca Scientifica 00133 Roma, Italy
Giuseppe Zampieri Dipartimento di Matematica Pura ed Applicata
Universit'~ di Padova Via Belzoni, 7 1-35131 Padova, Italy
Cataloging-in-Publication Data applied for
Die Deutsche Bibtiothek - CIP-Einheitsaufnahme
Integral geometry, radon transforms and complex analysis : held in
Venezia, Italy, June 3-12. 1996 / C. A. Berenstein ... Ed.: E.
Casadio Tarabusi ... - Berlin; Heidelberg; New York; Barcelona:
Budapest; Hong Kong; London; Milan; Paris; Santa Clara; Singapore;
Tokyo: Springer, 1998 (Lectures given at the ...session of the
Centro lnternazionale Matematico Estivo (CIME) ... ; 1996,1)
(Lecture notes in mathematics; vol. 1684: Subseries; Fondazione
CIME) ISBN 3-540-64207-2 Centro Internationale Matematico Estivo
<Firenze>: Lectures given at the ... session of the Centro
lnternationale Matematico Estivo (CIME) ... - Berlin; Heidelberg;
New York; London; Paris; Tokyo; Hong Kong: Springer Friiher
Schriftenreihe. - FriJher angezeigt u. d. T.: Centro lnternationale
Matematico Estivo: Proceedings of the ... session of the Centro l
nternationale Matematico Estivo (C1ME) 1996,1. Integral geometry,
radon transforms and complex analysis. - 1998
Mathematics Subject Classification (1991): 43-06, 44-06,
32-06
ISSN 0075- 8434 ISBN 3-540-64207-2 Springer-Verlag Berlin
Heidelberg New York
This work is subject to copyright. All rights are reserved, whether
the whole or part of the material is concerned, specifically the
rights of translation, reprinting, re-use of illustrations,
recitation, broad- casting, reproduction on microfilms or in any
other way, and storage in data banks. Duplication of this
publication or parts thereof is permitted only under the provisions
of the German Copyright Law of September 9, 1965, in its current
version, and permission for use must always be obtained from
Springer-Verlag. Violations are liable for prosecution under the
German Copyright Law.
9 Springer-Verlag Berlin Heidelberg 1998 Printed in Germany
The use of general descriptive names, registered names, trademarks,
etc. in this publication does not imply, even in the absence of a
specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general
use.
Typesetting: Camera-ready TEX output by the authors SPIN: 10649783
46/3143-543210 - Printed on acid-free paper
PREFACE
This book contains the notes of five short courses delivered at the
Italian Can- fro Internazionale Matematico Estivo (CIME) session
Integral Geometry, Radon Transforms and Complex Analysis held at
Ca' Dolfin in Venice (Italy) in June 1996.
Three of the courses (namely those by: Carlos A. Berenstein of the
University of Maryland at College Park; Sigurdur Helgason of the
Massachusetts Institute of Technology; and Simon G. Gindikin of
Rutgers University) dealt with various aspects of integral
geometry, with a common emphasis on several kinds of Radon
transforms, their properties and applications.
The lectures by C. A. Berenste in , Radon transforms, wavelets, and
applica- tions, explain the definition and properties of the
classical Radon transform on the two-dimensional Euclidean space,
with particular stress on localization and inver- sion, which can
be achieved by the recent tool of wavelets. Interesting
applications to Electrical Impedance Tomography (EIT) are also
illustrated.
The lectures by S. Helgason, Radon transforms and wave equations~
give an account of Radon transforms on Euclidean and symmetric
spaces, focusing atten- tion onto the Huygens principle and the
solution of the wave equation in these environments.
The lectures by S. G. Gindikin, Real integral geometry and complex
analysis, give an account of the deep connection between the two
main themes of this CIME session, covering several variations of
the Radon transform (RT): the projective RT; RT's taken over
hyperplanes of codimension higher than 1; and RT's over spheres. An
important and unifying tool is the ~" operator of
Gel'fand-Graev-Shapiro, used to explain analogies between inversion
formulas for the various RT's. This approach goes hand-in-hand with
~-cohomotogy and hyperfunctions, typical subjects in the field of
complex a~alysis.
In related areas, the other two courses (namely those by: Alexander
E. Tumanov of the University of Illinois at Urbana-Champaign; Peter
F. Ebenfelt of the Royal Institute of Technology at Stockholm)
share stress on CR manifolds and related problems.
The lectures by A. E. Tumanov , Analytic discs and the
extendibility of CR functions, provide an introduction to CR
structures and deal in particular with the problem of
characterizing those submanifolds of C N whose CR functions are
wedge-extendible. This property turns out to be equivalent to the
absence of proper submanifolds which carry the stone CR structure.
(The technique of the proof con- sists in an infinitesimal
deformation of analytic discs attached to CR submanifolds.)
The lectures by P. F. Ebenfel t , Holomorphic mappings between real
analytic aubmanifolds in complex space, deal with algebralcity of
locally invertible holomor- phic mappings. Along with classical
results, new criteria are introduced in terms of the behavior of
these mappings on a real-analytic CR submanifold which is generic,
minimal, and holomorphically non-degenerate in a suitable sense. To
this end a fundamental tool is afforded by the so-called Segre
sets.
VI
We wish to express our appreciation to the authors of these notes,
and to thank all the numerous participants of this CIME session for
creating a lively and stim- ulating atmosphere. We are particularly
grateful to those who contributed to the success of the session by
delivering very inspiring talks.
Enrico CASADIO TARABUSI Massimo A. P I C A R D E L L O Giuseppe
ZAMPIERI
TABLE OF CONTENTS
BERENSTEIN, C. A.
EBENFELT, P. F.
GINDIKIN, S. G.
Holomorphic Mappings Between Real Analytic Submanifolds in Complex
Space
Real Integral Geometry and Complex Analysis
Radon Transforms and Wave Equations
Analytic Discs and The Extendibility of CR Functions
35
70
99
123
Radon transforms, wavelets, and applications Carlos
Berenstein
We present here the informal notes of four lectures 1 given at Cs
Dolfin, Venice, under the auspices of CIME. They reflect the
research of the author, his collaborators, and many other people in
different applications of integral geometry. This is a vast and
very active area of mathematics, and we try to show it has many
diverse and sometimes unexpected applications, for that reason it
would impossible to be complete in the references. Nevertheless, we
hope that every work relevant to these lectures, however
indirectly, will either be explicitly found in the bibliography at
the end or at least in the reference lists of the referenced items.
I apologize in advance for any shortcomings in this respect.
The audience of the lectures was composed predominantly of graduate
students of universities across Italy and elsewhere in Europe, for
that reason, the emphasis is not so much in rigor but in creating
an understanding of the subject, good enough to be aware of its
manifold applications. There are several very good general
references, the most accesible to students is, in my view, [Hell.
For deeper analysis of the Radon transform the reader is suggested
to look in [He2] and [He3]. For a very clear explanation of the
numerical algorithms of the (codimension one) Radon transform in R
2 and R 3, see [Na] and [KS]. There have also been many recent
conferences on the subject of these lectures, for a glimpse into
them we suggest [GG] and [GM].
Finally, I would like to thank the organizers, Enrico Casadio
Tarabusi, Massimo Picardello, and Giuseppe Zampieri, for their
kindness in inviting me and for the effort they exerted on the
organization of this CIME session. I am also grateful to David
Walnut for suggestions that improved noticeably these notes.
1. T o m o g r a p h i c imag ing of space p l a s m a
Space plasma is composed of electrically charged particles that are
not uniformly distributed in space and are influenced by celestial
bodies. The problem consists in determining the distribution
function of the energy of these particles (or of their velocities)
in a region of space. A typical measuring device will take discrete
mea- surements (for instance, sample temperatures at different
points in space) and then the astrophysicist will try to fit a
"physically meaningful" function passing through these points. The
procedure proposed in [ZCMB] is based on the idea that the mea-
surements should directly determine the distribution function. We
do it by exploiting the charged nature of the particles and using
the Radon transform. (The recently launched Wind satellite carries
a measuring device based on similar interaction prin- ciples and
requires tomographic ideas for the processing of the data.)
The advantage of the tomographic principles that we shall describe
presently is that each measurement carries global information and
seems to have certain noise reduction advantages over the pointwise
measurements of temperatures, which is the
|
](
Figure 1: Schematic detector.
usual technology. We will describe everything in a two-dimensional
setting, but the more realistic three-dimensional case can be
handled similarly.
The instrument we proposed in [ZCMB] is schematically the
following. An electron enters into the instrument (a rectangular
box in the figure below) through an opening located at the origin
and is deflected by a constant magnetic field/~ perpendicular to
the plane of the paper (see Figure 1). Under the Lorentz force, the
electrons follow circular orbits and strike detectors lo- cated on
the front-inside surface of the box (along the y axis). Those that
strike a detector located at the point y have the property
that
where m is the mass of the electron, e its charge, and B the
magnitude of the magnetic field /~. In other words, all the
electrons with the same first component v~ of their velocities
strike the same detector located at the height y. The range of
velocities over a segment of width a (width of the detector)
is
Avx = (eB/2m)d
(In terms of the length of the detector plate D in Figure 1 and the
maximum velocity vm~ we have Avx = (d/D) - Vm~). If f(v~,vy)
represents the electron velocity dis- tribution, then the number dN
of electrons counted by a detector in time dt is given by
dN = Anev~Avx i f(vx, vy)dvy, dr, - - 0 0
ne is electron density and A is the area of the entrance aperture.
In other words,
1 dN f(v~, v~)dvy = AneAv~--~
- - o o
so tha t the count of hits provides the integral of f along a line
vx = constant in the velocity plane. By rota t ing the detector or
changing the orientation of the magnet ic field we obtain the Radon
transform of f .
As a realistic example, consider a p lasma of nominal electron
density ne = 10 c m -3 ,
velocity in the range Vmi n to Vmax of 1.2 X 10 s to 3.0 • 109 cm s
-1, average velocity = 6.5 • 108 cm s -1, and we assume a Gaussian
distr ibution function
so tha t
dt - const, e x p , , 2~2, ]
with individual detector area and aper ture of 0.04 c m 2 for a
small instrument one gets tha t the dis tr ibut ion function f
varies from 1 to 10 -5 while d N / d t varies from 102 to 105s -1.
The s tandard measurement methods make the a priori assumption tha
t f is the sum of a Gaussian centered at V and per turbed by adding
a finite collection of Gaussians, often located in the region where
f varies from 10 -4 to 10 -5, but the previously described
instrument does not require any such assumption, on the other hand,
experimental ly one sees that such large variations, like from 1 to
10 -5 as in the example, are realistic. We shall see in Section 2
tha t this is an embodiment of the Radon transform in R 2. The more
realistic case of 3-d is handled by an instrument where there is a
plane which contains the entrance aper ture and a 2-d array of
detectors in the plane (x, y). One shows tha t at each detector
location (x, y) one obtains an integral over a planar curve and
that the addi t ion of overall elements with the same x component
leads to a 2-d plane integral of the density dis t r ibut ion so
tha t we have the Radon transform in R 3. (This is an observation
we made joint ly with M. Shahshahani.)
Before concluding this section, let us remark tha t the large
variations expected from the velocity density function f make the
inversion of the Radon transform very ill-conditioned, even if f is
assumed to be a smooth function. This is due to the continuity
propert ies of the Radon transform and its inverse as seen in the
next section. The remarkable point is that in medical applications,
like CAT scans, the unknown density is natural ly discontinuous
along some curves but otherwise it has small local variations, and
it is this reason the inversion problem is u l t imate ly easier
for medical applications.
Source
Figure 2: Schematic CAT scanner.
2. T h e R a d o n T r a n s f o r m in R ~
Let w E S 1, w = (cos 0, sin 0), and take p E R. The equation x-w =
p represents the line l which has (signed) distance p fi'om origin
and is perpendicular to the direction 02.
For any reasonable function f (e.g., continuous of compact
support), we can com- pute the line integral, with respect to
Euclidean arc length ds,
oo
:= / f(x)ds = / f(x0 Rf(w,p) + tw• )dt (1) X , o 2 ~ p - -00
where x0 is a fixed point in l, i.e., satisfying the equation x0 -
w = p, and w • = ( - sin 0, cos 0) is the rotate of w by ~/2.
The map f ~-~ R f is called the Radon transform and R f is called
the Radon transform of f . Clearly R f is a function defined on S 1
• R (that is, the family of all lines in R 2) with the obvious
compatibility condition:
(R f )(-w, -p) = Rf(w, p). (2)
There are several reasonable domains of definition for R such as
LI(R~), $(R2), etc., but in many applications it is enough to
consider functions which are of compact sup- port, with
singularities which are only jumps along reasonable curves, and
otherwise smooth. This is obviously the transformation appearing in
Section 1. The full 3-d instrument there corresponds to integration
over planes in R 3, perpendicular to a unit vector w. A big source
of interest of this transform lies in CAT (Computerized Axial
Tomography) as a radiological tool where each planar section of a
patient is scanned by X-rays as in Figure 2.
In this particular case it can be seen that
I0 f log ~ ~ / p d s (3)
J l
where I0 is the radiation intensity at the source and I is the
intensity measured at the detector. The attenuation is a
consequence of traversing a tissue of density #. So the data
collected from this X-ray scanning appears in the form of the Radon
transform R# of the density #, computed for a finite collection of
directions wt, w2, . . . , w~v (usually equally spaced) and a
finite collection of lines, i.e., values Pl,P2,...,PM for each
direction. This is called a "parallel beam" CT scanner. The
configuration that it is now most used but we shall not discuss
here is the "fan beam" CT scanner, we refer to [Na], [KS] for a
discussion of the differences of these two cases, they really only
appear at the implementation level of the inversion algorithms
because only a limited amount of data can be obtained in the real
world.
Some easy properties of the Radon transform are obtained by
observing that Rf can be written using distributions. In fact, if
we introduce the unit density 5 ( p - x . w ) which is supported by
the line x 9 w = p, then
Rf(w, p) = f f(x)~(p - x . w)dx (1') R ~
with the usual abuse of language. It is also convenient to
write
R~f(p) = Rf(w,p). (4)
Formula (1') can be used to extend Rf to (R2\{0}) x R, using the
fact that 5(p-x.w) is homogeneous of degree -1 ; indeed, one
defines
nf(~, s) = ~ n s , (5)
One can therefore take derivatives of (1') with respect to the
variables ~j(~ = (~1, ~2)) and obtain
~---~jRf(~,s)= f f(x)~-~j~i(s-x.~)dx (j = 1,2), (6)
but
and
O f f (x)x jh(s- :c'~)dx
0 Os (R(xjf)(~, s)).
On the other hand the Radon transform of the derivative of f
is:
In particular, for
= ~jff---~(RJ)(s).
(7)
Rr = (~ + ~ ) (RJ ) ( s ) . (7')
When ~ is restricted to be an element co E S 2, we get,
02 ( R A f )(co, s) = -~s2s2 R f (co, s); (S)
0 2 In other words, R intertwines A and ~ when the arguments are
restricted to S 1 • R. Another useful property is the
following:
P ~ ( f 9 g) = P ~ f | P~g, (9)
where the symbol 9 on the left side of (9) denotes the convolution
in R 2 and | denotes the convolution product in R. The easiest way
to verify (9) is via the Fourier Slice Theorem, which we recall
here:
Let ~1 denote the Fourier transform of a function in R and f or F2
the Fourier transform of a function f in R 2. Then
.T'l(P~f)('r) = f(Tco). (10)
oo
= f e-2'm"f( tco + scoX)dsdt, R 2
Letting now x = tco + sw • one has t = x- w and dtds = dx, the
Lebesgue measure in R 2, in the previous equation we obtain
.T'I(/~/)(T) = f e-2"(:"')'V(z)dz It.2
9 F'2(f * g)(~) = ](~)~(~),
we can easily prove (9). Indeed, from (10) we have
~I (P~( f 9 g))(.) = ~2( : 9 g)(.~) = / ( . ~ ) ~ ( . ~ )
= y l (P~/ ) ( . )~: l (p~g)(~) = ~:l(p~f | p~g)( . )
and therefore, by the injectivity of 5vl, we get P ~ ( f 9 g) = P ~
f | Let us also note that if Ta denotes the translation by a, i.e.,
T , f ( x ) = f ( x - a),
then
R ( 7 - , ~ f ) ( w , p ) = R i f ( x - a ) ] ( w , p ) = P ~ f ( p
- w . a) = % . a R f ( p ) .
We now proceed to state some inversion formulas, which give
different ways to recover f from R f .
F o u r i e r I n v e r s i o n F o r m u l a :
oo
(nj)O-)d~ (11) 0 $1
The proof is clear, we begin with the Inversion Theorem for the
Fourier transform. We have
f (x) = [ R2
oo
We now apply the Fourier Slice Theorem to get
oo
0 wES1
This inversion formula can be implemented numerically using the
Fast Fourier Transform (EFT) (see [Na]). Quite often the points T w
where the data J : l ( R ~ , f ) ( ~ - )
is known do not have a lat t ice structure. This causes problems
for the F F T but we can use rebinning algori thms like [ST] to
obviate this problem.
To obtain another kind of inversion formula we observe the
following:
f i P~f(s)g(s)ds= fi ; S(sw + tw• - o o - o o - o o
Let x -- sw + tw • so that s -- x 9 w, dx = dtds, and
therefore
- r 1%2
i.e., the adjoint of P~ is the operator R~ defined by
Rig(x) = 9(x. ~). (12)
We now consider for an arbi t rary function g(w, s), having the
symmetry g(-w, -s) =
f Rf(w,s)g(w,s)dwds= fsl dw fi R~f(s)g(w,s)ds S l x R - ~
oo
(with the usual subst i tut ion, x = sw + tw • etc., we get)
= f d ~ f f ( x ) g ( ~ , ~ . x ) d 2 x $1 R 2
= /f(x)R#g(x)d2x. l:t2
(13)
The opera tor R # defined by (13) is known by the name of
"backprojection operator" . Note, in this regard, tha t g(w, s) is
a function of "lines" and that R#g(x) is its integral over all
lines passing through x. It is easy to prove the following useful
proper ty of the backproject ion operator
(R#g) 9 f = R#(g | Rf), (14)
where the convolution | in the second member clearly takes place in
the second variable. This identi ty plays an impor tant role in the
numerical inversion of the Radon transform.
Finally, we get to the following impor tant result:
2 9 f . (15) R # R f =
Indeed,
R#Rf (x ) = S R f ( w ' w ' x ) d w S~
= Sd,~z S f ( ( m - x ) m + s w • S1 --00
S~ 0
By set t ing y = s~o • s = lyl, dy = s&ods we get
R#RS(x) = ~ S ~S(x § y)<~y R 2
1 d = 2ST;--m~s(~,)y,
R2
1 1 (see e.g. [He2, page 134]), one deduces which is exactly (15).
Since one has 2 ~ = tha t
m2(n#n/)(~) = ~](~)
One can therefore conclude tha t the inversion operator A is such
tha t
X(~) =
'S f(x) = ~ e2~'~l~l(R#Rf)~(~)d~ = AR#Rf(x) , (16) R2
which is sometimes called the backprojection inversion formula.
Reorganizing the terms in the last formula one can rewrite it to
obtain a more
s tandard form, where the filtering is one dimensional.
10
where
Hg(s)= l-Tr ? f(~t-)tdt' - - 0 0
where the last integral is understood in the sense of principal
value. In other words,
1 (Rf)'(O,t_)dtdO (18) f (~)- (2~)~ L, f x .O- t 9
R
No introduction to the Radon inversion formula can be complete
without at least mentioning the inversion formula due to Radon,
which among other things, is akin to the inversion formula for the
hyperbolic Radon transform due to ttelgason, which will be
mentioned below. Consider for a fixed x E R 2, the average of Rf
over all lines at a distance q > 0 from x, namely, let
1 fRf(~,w.x+q)dw. Fz(q) := S 1
Radon found that
0
We refer to [GM] for the original 1917 paper and
commentaries.
An approximate implementation of (17) can be given by using the
Fourier inversion formula
71-
where, as above, w = (cos 0, sin 0) and
o(t) b
o o
1]
This last approximation constitutes a band limiting process, and it
can also be ob- tained from (14) as follows: Let wb be a
"band-limiter ' , i.e, supp(~b) C [-b, b] and Wb = R#wb. Then (by
letting g = wb in (14)), we obtain
Wb * f = R#(wb | R f)
that is, we want Wb to be an approximate di-function (cf. [Na, Ch.
5]). To begin with, choose Wb radial, e.g.,
27r " b "
where 0 < •(a) < 1, + = 0 for a > 1; this implies that wb
= const. I~l+(~b ), The previous example is given by the ideal
lowpass filter defined by
and so
1 i f 0 < a < l ~ ' = 0 i f a > _ l
Wb(x) lbzJ~(blxl) = 2 ~ (bill) '
where J1 is the Bessel function of the first kind and order one. We
shall see below that one of the wavelet-based inversion formulas is
inspired by
(20). The formulas (15), (16), (17) allow for rather precise
estimates of the degree in
which the Radon transform and its inverse preserve the smoothness
of the function f and data Rf. One way to measure this is to do it
using Sobolev norms defined in an obvious way in the space of
functions in the space of all lines. For instance, if f E C ~ ( B )
, where B is the unit disk in R 2, then one can find in [Na,
Theorem 3.1], that for any real a, there are constants c, C > 0
such that
cllf[lH~'(m) ~ Ilnf[[~.§ ~ CIIflIH~'(B) (21)
In the particular case of a = 0, we see that for f E L0~(R 2) and
supp(f) CC B one cannot expect better than control of one-half
derivatives of Rf.
Useful variations of the estimates (21) for other kinds of Radon
transform can be found using that R#R is a Fourier integral
operator of elliptic type [GS].
12
3. L o c a l i z a t i o n o f t h e R a d o n t r a n s f o r
m
Returning to the problem of p lasma physics that s tar ted these
lectures, besides the fact t ha t the functions we are t rying to
detect seem to have a very large variation, that is, a large H 1
norm, we have the added difficulty that the amount of da t a one
can process or send down to Earth is fairly limited. One knows
experimental ly that , on a first approximat ion, all the
variations from being Gaussian occur in the region where the values
of f have gone down by 4-5 orders of magnitude. That is, if f has a
value 1 for the bulk velocity, then we are interested in the region
where the values of f lie between 10 -4 to 10 -5 . The way this
problem was t radi t ional ly solved (for conventional measuring
devices) was to assume f had the form of a linear combination of a
small number of Gaussians, and one just tries to est imate the
variances and coefficients of these per turba t ion terms. If one
does not want to impose these a priori restrict ions on f , and we
have only l imited amount of da ta to use, a natural idea is to jus
t use those lines tha t cross only the annular region where the
main Gaussian varies between 10 -4 and 10 -s . (In the case of
dimension 3 we would be dealing with a shell instead of an
annulus.) This requires a localization of the Radon Transform.
There are two ways to proceed. One, the most obvious (or naive) way
is to try to localize the Radon transform as follows:
Reconstruct a function f in a disk B(a, r) from the data Rf(g),
using only lines g passing through B(a, r).
This cannot be true in dimension 2, as observed already by F. John.
The reason is the well-known fact tha t waves cannot be localized
in 2-dimensions, namely, if we drop a pebble in the water, the
ripples propagate along ever-expanding disks with time. In other
words, an arbi t rary per turbat ion confined to a disk at t ime t
= 0 does not necessarily remain confined to the same disk (or any
concentric disk, for tha t mat te r ) at all future times. On the
other hand, as F. John pointed out [J], if u is a solution of the
wave equation Au - utt = 0, then its Radon transform v = Ru is a
solution of the one dimensional wave equation vss - vu = 0, as it
is seen immedia te ly from the relat ion (8). For the
one-dimensional wave equation with initial conditions at t ime t =
0, v(s, O) = Vo(S) and ~ 0) = vl(s) , we have
1 I r~+t v(8, t) = (vo(s - t) + vo(s + t)) + I
S--t
If we th ink the pebble as being given by u(x ,0) = no(s) = 0 for
{x I _> c, 0 < ~ < < 1, and ut(x,O) = ul(x) =--- O,
then for a fixed w E S 1 and any later t ime t > 0 we would have
with v(s,t) = R(u(.,t))(w, s) that vo(s) = v(s,O) = 0 for {s I >
e while vl(s) = vt(s, 0) - 0. Thus, at any later t ime t > c we
have that v(s, t) is only different from 0 for t - e < Is{ <
t + e. Thus, the strict localization of the Radon transform would
impose tha t the support of u(x, t) be in the annulus t - ~ < {x
I < t + e, which contradicts our observations. (Nevertheless, we
shall see that some sort of localization takes place.)
The other al ternative, which fits the p lasma problem, is to try
to see whether we could reconstruct the values of f outside of a
disk from the values of Rf(g), with never crossing tha t disk. This
turns out to be possible! It is the exterior problem for the Radon
transform. We follow here the work of Quinto [Q]], [Q2] (and
references therein.)
13
The starting point, as recognized in the pioneering work of A. M.
Cormack, is to expand both the function f and its Radon transform g
= Rf in a Fourier series. (For R n, n > 2, one uses spherical
harmonics [Na, p. 25 ft.]) That is,
f(~) = ~ f~(r)~ ~~ x= (rcosO, rsinO),
l~--oo
Then, the Fourier coefficients ft and g~ are related by the two
formulas
oo 2
O0
:~(r) -- ~ ~i ~ - ~)-~%i(~)d,, (23) T
where T~ is the Chebyshev polynomial of the first kind. One of the
consequences of the Fourier Slice Theorem is that g cannot be an
arbitrary function in the space of lines, it must satisfy certain
compatibility conditions, usually called the moment
conditions,
j s'~-lg(w, s)ds e span{e 'k~ Ikl < m}, w = e ~.
This allows for a modification of (23) that makes it far more
practical for numerical purposes [Na, p. 29-30]. This pair of
equations show that the values of Rf(i) over all lines exterior to
the disk B(0, r) are thought to determine f in the exterior of B(0,
r). In particular, if one has
supp f C B(0, P0),
then the values of f in the annulus Pl < Ixl < P0 are
entirely determined by the mea- surements of the Rf(~), only for
lines ~ that intersect this annulus. (The uniqueness of the
exterior problem and its variants is usually called the support
theorem. It was first proved by Helgason in 1965, we refer to
[Hell, [He2] for details and generalizations.)
Quinto [Q1], [Q2] has successfully used this kind of ideas to
obtain a very effective tomographic algorithm to determine cracks
in the exterior shell of (usually large) circular objects, for
instance, rocket nozzles. The method of Quinto is based on two
things. First, the known characterization of the kernel of the
exterior Radon transform in L 2 spaces with convenient radial
weights (this is due to Perry for n -- 2 and Quinto for n > 3).
For the case of interest at hand, n = 2, we consider the kernel of
the exterior Radon transform in L2(B~, rdx), then the Fourier
coefficients f~(r) must be given by the rule
fe(r) = linear combination of r 2-k, 0 ~ k ~_ I~1, [~[- k even.
(24)
For instance,
14
fo = O, f l = O , / 2 = c r - 2 , f3 = c r - 3 , f4 = c l r - 2 + c
2 r - 4 , . . .
The second observation is the fact that the Radon transform
maps/-/1 := L2(B[, r(1 - r)U2dx) into H2 = L2(S ~ x (1, cr g -~)
and one has an explicit diagonalization procedure for R, so that
there are orthonormal bases ~j and Cj, respectively of / /1 @ ker R
and of ImR c H2, so that
R~j = gj~bj with aj > 0
and aj explicitly computable. Thus, for a given f of L2c(B~, dx)
(i.e., of compact support), we have
oo
f = ~ , aj~j + ] with ] E ker R N n2c(S~, dx) j - - 1
so that
1 aj = - - < R f , Cj >H2
~j
This determines exactly ~ = f - ] . One expands now f in a Fourier
series ~ f~(r)e ~e, with f~ of the form (22) as mentioned earlier.
Now for r > > 1 we know that f -- 0,
so that f = -~ , thus, for r > > 1 we have f~(r) = - ~
e-~~176 but the --Tr
coefficients f~(r) are polynomials in 1/r, so they are completely
determined everywhere (up to r = 1) by their values for r > >
1. It is here that one uses a sort of analytic continuation, so it
is fairly unstable, but Quinto has modified further this algorithm
if one assumes f to be known in the small annulus 1 < Ix] < 1
+ e, to give it further stability [Q2].
In the context of the plasma problem, we compared numerically the
use of the same number of data measurements Rf(e), either spread
throughout the whole disk versus the measurements taken only (and
thus more densely) in the annulus of interest. We found the
surprising result (to us) that the standard algorithm, with more
thinly spread measurements did better. It was this numerical
observation that led to the search of a different way to localize
the Radon transform using wavelets.
Let us first review briefly two other localization methods that had
appeared earlier in the literature.
The first one is the following. Let us assume that the unknown
function f has support in the disk B1 of center 0 and radius 1, but
that we are only interested on the values of f in Bb, 0 < b <
1, while we collect data on Ba, 0 < b < a < 1. (Note that
all the disks are centered at 0). This is the situation considered
for the interior Radon transform [Na, VI.4]. The basic idea is to
make Rf(g) = 0 when e doesn't intersect B , and apply the standard
reconstruction algorithm. In other words, we
]5
want to obtain (even approximately) the values f(x) , Ixl S b, from
R(fxa), where Xa stands for the characteristic function of the disk
B~. The first problem is that there are many non-zero functions f
that have R(fx~) -- 0. Luckily, these functions do not vary much on
Bb [Na], so one could just try to find f up to an additive constant
(and try to find that constant by other means). One can see from
the table or the formula (4.4) in [Na, p. 170], that one needs a =
4b to obtain a maximum L ~ error of 1.6% of the L 2 norm of f in
B1. In particular, this procedure could not be applied if we are
interested in f(x), for x E B(xo,a) C B1 with x0 close to cOB1. A
typical such example is that of spinal chord studies. Usually, one
study involves 40-60 CAT scans, that is, 40-60 scans along body
sections perpendicular to the spine at different heights. The
spinal chord area is about 15% of any such cross section of the
body, and there would be a substantial reduction of radiation
received by the patient if one localizes the CAT scan to only those
lines passing through or near the spinal chord area.
Another alternative that has been proposed is that of A-tomography
[FRK], where one only attempts to reconstruct to discontinuities of
the function f , i.e., perform edge detection in the image. The
principle is based on the formula (15) namely, consider the
"approximate" inversion
] = AR#Rf ,
so that / = 4~rAf. This formula preserves the "edges" (=
discontinuities of f ) but not the actual values of f . A variation
of this formula has been implemented in the Mayo Clinic to study
angiograms [FRK].
Another interesting consequence of this kind of approximate formula
is that it can also be applied to the attenuated Radon
transform,
o o
R,f (w, s) = / f(sw + tw• § tw • ?1))dr -oo
where #(x,w) is assumed to be real analytic in R 2 • S 1 and
nonnegative. This appears in SPECT tomography and, usually, both f
and # are unknonwn. As observed by Kuchment and collaborators [KLM]
the function A R # R j will have the same singularities as f . The
point is that R#R, is still an elliptic Fourier integral operator.
This fact had already been used effectively by many people, most
notably Boman and Quinto [BQ], and it is the key observation in the
work of Quinto [Q3], Ramm and Zaslavsky [RZ], and others.
The method of localization we want to discuss here with a bit more
detail is that of using wavelets to invert and localize the Radon
transform in dimension 2. This general principle, which is joint
work with David Walnut, was presented first in a 1990 NATO
conference [W], and independently in [Ho]. Since then, similar
ideas have appeared elsewhere in the literature (see, for instance,
the recent volume [AU], the papers [BWl], [BW2], [DB], [DO], [O],
and references therein.) True localization using discrete wavelets
and filter banks is clearly developed in [FLBW]. (See also [FLB]
for the fan beam case.)
There are many excellent books on the subject of wavelets, at all
levels of so- phistication and different points of view, the
following is a very partial list [M], [D],
16
[Ka]. There are actually two different, albeit related concepts,
the continuous wavelet transform (CWT) (easier to understand) and
the discrete wavelet transform (DWT) (easier to work with).
The idea of CWT originates from the standard properties of the
Fourier transform representation of nice functions. For f 6 L2(R)
or f E S(R), we have both
f(~) = f f ( x ) e-2"ix~dx - - 0 0
O0
II/II = IIfII : ( If(x)12dx) ~/2 - o o
If we translate f by b 6 R, Tbf(x) := f ( x -- b), then (vbf)'(~) =
e2"ib~/(~), and for dilations we have D~f(x) := -~af(x/a)(a >
0), so that IIfJI2 = IID~f[12 and
(D~f)'(~) = D(1/~)/(~).
In other words, the group x --+ ax + b (a > O, b q R) operates
via unitary operators in L2(R), and has a corresponding
representation on the space of Fourier transforms (which happens to
coincide with L2(R)). The "problem" of the Fourier transform rep-
resentation is that the behavior of f at a point ~ depends on the
values of f everywhere, for that reason, the idea of a "windowed"
Fourier transform has been introduced long ago, namely, introduce a
cut-off function g (say, a "smooth" approximation of X[-1,1]) and
consider
9 T'l((Tbg)f)(~) = f g(x -- b)f(x)c-2"i~dx. - o o
Note that 5rl((7-bg)f)(~) is r 9 f(~), where r is the wavelet r =
g(x)e2~iz~,(b(x) = r If we want to consider also the behavior at f
at different scales we are led naturally to the CWT: Given a
wavelet r 6 L2(R), and f 6 L2(R) we define
4- A ../+
-oo
for 0 < a < cx), b 6 R, r denotes the complex conjugate of r
and <, > denotes the L2-scalar product. We assume the wavelet
is "oscillatory", that is, it is an arbitrary function in L2(R)
which satisfies the condition
j I (012 _ cr :=-oo - - - ~ a r < oo.
17
OO
This condition implies that f r = O. (For instance, when r is
continuous at - - C O
= 0, which occurs if r E LI(R) N L2(R).) In fact, later on we will
be interested in wavelets with many vanishing moments
f xkr = O, - - 0 0
A typical wavelet is the Haar wavelet
0 < k < N .
so that
D1/2r = v~(X[0:/4] - X[1 /4 ,1 /2 ] )
which shows that for k -+ co, D2-~r "analyzes" smaller and smaller
details of the "signal" f .
Moreover, Wcf determines f as seen from the following relation
valid for any pair f ,g C L2(R)
j f dadb Wr = cr < f,g > [1r - - 0 0 - - 0 0
usually called Calderon's identity. If 11r = 1 one also has the
L2-approximation property
1 b) DaTbr I I (26) Ilf-~ f Wcf(a, ---~0 A l ~ l a l < _ A
2
Ibl_<B
as A2 ---+ 0+,A2 ~ +co, B --+ +co.
The generalization to R ~ is easy. A function r E L2(R ~) is a
wavelet if
For a radial wavelet r E L2(R ~) and f C L2(R ~) we define the CWT
by
Wr f o r a e R \ ( 0 } , b C R ~,
where this time, D~r = lal-~/2r The interest of the CWT for
tomograpy lies in the following two propositions from
[BW2].
P ropos i t i on 1. Let p r L2(R) be real valued, even, and
satisfying
OO ^ 2
18
Define a radial function r in R 2 by 3v2r = 2~([~1)/1~1, then r is
a wavelet a n d
w,d(a, b) = a-'/2 f (w,P~/)(a, b. o) ) do3 (28) S a
P r o o f . Using the Fourier Slice Theorem we have for "y 9
R
It is then easy to verify that (27) implies that r is a wavelet in
R 2. Recall that the Riesz transform of order a, I"~ , of a
function ~ 9 S ( R ) is defined
by ( I~) ' (9 ' ) -- h ' l -~ (9 ' ) , thus the identity (29) can
be rewritten as
p(t) = l : - ~ ( P ~ r
Extend p to a function in the space of lines by making it
independent of the slope of the line, p(w, t) = p(t) for every w 9
S ~, then we have
= 8 9 1 6 2 = r R#p(x)
since the last formula is a rewritting of formula (16) in terms of
the Riesz transform. More generally, for any a > 0 and every w E
S 1, we have
(R#Dapo,)(x) = aV2Df,(b(x),
so that, using identity (14) and the fact that p is real valued, we
obtain
Wcf(a,x) = (f.D.(p)(x) = a-I/2(f. R#D~,:.,)(x)
= a-~/~R#(P~f | D~
= a - 1 / 2 / ( W p P ~ f ) ( a , x . w ) d w S 1
This concludes the proof of the proposition. 9
A similar relation between the Radon transform and the CWT can be
found using "separable" wavelets in R 2.
P r o p o s i t i o n 2. Given a separable 2-dimensional wavelet of
the form
r = r162
where each r satisfies Ir < (71(1 + 171) -1 for all 7 9 R,
define the family of one-dimensional functions {P~}~es' by
1 ^1 ^2
where w = (wl,w2) 9 S t. Then, for every f 9 L I ( R 2) VI
L2(R2),
19
0.,
0.6
0.4
0.2
C
-0.2
-0.4' -10
G a u s s i a n a n d I ts H i l b e r t t r a n s f o r m
'\] i,j/ I I I I I I J I I
~8 --6 --4 --2 0 2 4 6 8 10
Figure 3: Gaussian and its Hilbert transform.
( W r = a -1/2 J f)(a,x . ~ ) d ~
s 1
The point of Proposition 2 is the observation that the wavelet
transform of a function f ( x ) with any mother wavelet and at any
scale and location can be obtained by backprojecting the wavelet
transform of the Radon transform of f using wavelets that vary with
each angle, but which are admissible for each angle.
So far we have not yet shown that the inversion formulas of the
Radon transform based on wavelets do a good localization job. Using
Proposition 1 the problem is clear, find a function p such that p
has small support and simultaneously r has small support. From the
relation (29) we see that we have overcome the Reisz operator of
order - 1 , its symbol is 171 = (sgnv)v, so it is the composition
of the differentiation and the Hilbert transform. (This is exactly
the content of the inversion formula (18).) The problem, of course,
is the Hilbert transform, but if we choose p with many vanishing
moments, then we can overcome the difficulty. For the sake of
comparison we show in Figure 3 the Hilbert transform of a Gaussian,
its effective support is about four times the effective support of
the Gaussian (defined by making zero those points below 1% of
maximum value), which tails exactly with the result about the
interior Radon transform mentioned earlier in this section.
The key to explain the success of the wavelet method of
localization is the following proposition [BW1], which in spirit is
similar to the general principles about Calderon- Zygmund operators
stated in [BCR].
P r o p o s i t i o n 3. Suppose that n is an even integer and the
compa~iy supported function h E L2(R) is such that for some integer
m >_ 0 we have that h is n + ra - 1 times differentiable and
satisfies
20
9O
80
~, 70
10 i L 20 40 60 80 10~ 120 140
Radius of the region of interest in pixels
Figure 4: Exposure versus the radius of the ROI.
(a) 7jh(k)(7 ) E L~(R) N L2(R) for 0 __ j _~ m, 0 < k < m + n
- 1
(b) ~ tJh(t)dt = 0 for 0 < j _< m - - o o
Then
and
t~+m-1II-~h C L2(R).
The proof is rather elementary, it depends on the fact that if h is
a function of compact support with m + 1 vanishing moments then
171~-1h(7) has n + m - 1 continuous derivatives.
For ease of application it is better to work with the discrete
wavelet transform (DWT). This is basically obtained by diseretizing
the CWT or appealing to the mul- tiresolution analysis of Mallat
and Meyer [D], [M]. We have done this in detail in [FLBW] using
coiflets [D] in order to be able to implement the inversion process
using filter banks. One can show that to obtain a relative error of
0.5~o one only needs a margin of security of 12 pixels around the
region of interest (ROI). For instance, to recover within this
error bound an image occupying a disk of radius 20 pixels in a 256
• 256 image, one only needs about 25% of exposure, as shown in
Figure 4.
Figure 5 below is the Shepp-Logan phantom and its reconstruction
from global fan beam data using the standard algorithm, in Figure 6
we use local data and our wavelets algorithm.
The following figures are the reconstruction of a heart from real
CAT scanner data using our wavelet method, and the reconstruction
of the central part from local data and our wavelet method is found
below.
21
(a) (b)
Figure 5: (a) The Shepp-Logan head phantom; (b) the standard
filtered backprojec- tion in fan beam geometry (4).
Figure 6: Reconstruction from wavelet coefficients.
22
Figure 7: Reconstruction of heart from wavelet coefficients.
Figure 8: The local reconstruction of of central portion of
heart.
23
We leave to the discussion and references in [FLBW] and [BW2] the
comparison with other methods of inversion of the Radon transform
using wavelets. One should add to the references in those two
papers, the very recent work of Rubin [R], which is based on a
systematic use of the Calderon reproducing formula and it is thus a
development of the original ideas in [Ho].
24
4. T h e hype rbo l i c R a d o n t r a n s f o r m and Elec t r
ica l I m p e d a n c e Tomography
In this section we discuss the role tomography plays in a classical
problem of Ap- plied Mathematics, the inverse conductivity problem.
Several of the earlier attempts to solve this problem involve
generalizing the Radon transform to other geometries, that is,
integrating functions over other families of curves beyond straight
lines in the Euclidean plane. There are many examples of such
transforms, in fact, the integration over great circles in S 2 was
a transform considered by Minkowski and which inspired Radon in his
work. The two we shall introduce presently are the generalized
Radon transform of Beylkin [By] and the Radon transform on the
hyperbolic plane [He1].
Let ~ be an open subset of R 2 and r E C ~ ( ~ • (R 2 \ (0))) be
such that
(a) r A~) ---- Ar ~) for )~ > 0
(b) V=r r 0 for all (x,~) E ~ x (R 2 \ {0})
Then, for any s E R and w E S ~ we can define the smooth
curve
= {x e a : r = s} ,
that is, the level curves of ~b. We let da denote the Euclidean arc
length in such a curve. For u E C~(~) define the "Radon
transform"
R+u(~,s) = f u(x)l V: r 5)ld,~(x) / ' / s ,~
Let h(x,~) be the Hessian determinant of r with respect to the
second variables, 0 2 x h(x, ~) = d e t f ~ l then the
"backprojection" operator Rr # is defined by
h(x,~) R v(x) = f I
w E S 1
Introducing K as the operator of convolution by 1/Ixl, Beylkin
proved the following approximate inversion formula for the Radon
transform as an operator
R+: Lc2(a) --+ L~or
R#cKRr = I + T (30)
T : L~(~) -+ L~or )
is a compact operator. In fact, Beylkin gives a recipe for a family
of backprojection operators and generalized convolution operators K
so that a decomposition of the type (30) holds. This gives his
transform great flexibility and applicability to many problems,
especially inverse acoustic problems, of course, the reader can
easily verify
25
that for convenient choices of r the transform Re yields the
Euclidean Radon trans- form studied earlier and the hyperbolic one,
which we now introduce. (The reader should consult [Hell, [He3] for
more details on this subject.)
Let D, the unit disk of the complex plane C, be endowed with the
hyperbolic metric of arc-length element ds given by
4'dzl~ (31) ds2 - (1 -Iz[2) 2'
where Idzl denotes the Euclidean arc-length element. This metric is
clearly conformal to the Euclidean metric but has constant
curvature
-1 . The geodesics of this metric are the diameters of D and the
segments lying in D of the Euclidean circles intersecting the unit
circle COD perpendicularly. One can introduce geodesic polar
coordinates z ++ (w, r), where w = z / H , r = d(z, 0). Note that
Iz[ = tanh(r/2). In these coordinates the metric (31) can be
rewritten as
ds 2 = dr 2 + sinh 2 r dw 2
where dw 2 indicates the usual metric on cOD. The hyperbolic
distance between two points is given by
Iz- ! d(z ,w) arcsinh ((1 -I~I~),~(I -l l )V
The Laplace-Beltrami operator AH on D can be written in terms of
the Euclidean Laplacian A as
a . - 0-1zP)2a 4
02 O 02 - Or 2 + c o t h r + s i n h - ~ r o w 2. (32)
The classical Moebius group of complex analysis is the group of
orientation preserving isometrics of the hyperbolic plane D.
One can define the hyperbolic Radon transform RH by
Rf( '~) = RHf (7 ) = ~ f ( z )ds ( z ) , 7 geodesic in D (33)
which is well defined for, say, continuous functions of compact
support, or functions decaying sufficiently fast. Observe that to
be integrable on the hyperbolic ray [0, oc[ (which is just the
straight line segment from 0 to 1 in the complex plane C), f has to
decay a bit faster than e -r. We denote by F the space of all
geodesics in D, then the dual transform R # (or backprojection
operator) is given by
R#r -- fr~ r (34)
where Fz is the collection of geodesics through the point z and
dttz is the normalized measure of Fz. Since a geodesic through z is
determined by its starting direction w C S 1, then Fz ~ S 1 and d#z
is naturally associated to ~ d w when we use this particular
parameterization of Fz.
26
In order to invert R H one can proceed in the spirit of Radon's
inversion formula (19). This was done by Helgason [He2, p. 155]. Or
one can try to find a filtered backprojection type formula like
(16). For that purpose we need to define convolution operators with
respect to a radial kernel k. For k E L~oc([0, cxD)) and f E Co(D)
we define
f
k * f (z) = k *H f(z) := ]D f(w)k(d(z, w))dm(w) (35)
where dm(w) stands for the hyperbolic area measure, which in polar
coordinates is given by
d m = sinh r drdw.
Corresponding to the Euclidean formula (15) we have
R#HRHf : k * f, where k(t) - 1 Tr sinh t
One can prove [BC1] that if
(36)
which is the exact analogue of (16).
It is convenient to recall here that in the hyperbolic disk D we
have a Fourier
transform [He2]. It is easier to work it out for "radial" functions
as we interpret our kernel k, then the Fourier transform is defined
with the help of the Legendre functions P~(r) by means of the
following formula
o o
For radial functions k, m, we have
(k * m)'(1) : k(~)~(~)
So that, if/~(A) # 0 for all I E R, in principle, that is, for a
convenient class of functions f , the convolution operator f , >
k *H f is invertible.
We refer to [He2], [BC1], [BC2], [Ku] for corresponding inversion
formulas in the higher dimensional hyperbolic spaces, and the
characterization of the range of the Radon transform. In
particular, [Ku], [BC2] exploit the "intertwining" between RH and
the Euclidean Radon transform as well as the Minkowski-Radon
transform on spheres.
Let us cxplain now what the above hyperbolic Radon transform has to
do with Electrical Impedance Tomography (EIT) and what EIT
is.
Let us consider the following tomographic problem: using a
collection of electrodes of the type used in electrocardiograms
(EKG) uniformly distributed around the breast
27
of a patient and all lying in the same plane, introduce
successively (weak) currents at each one of the electrodes (as done
in EKG) and measure the induced potential at the remaining ones.
The objective is to obtain an image of a cross section of the lungs
to determine whether there is a collapsed lung or not. This was
what Barber and Brown set up to do in 1984 [BB1], [BB2]. The point
being that this equipment is cheap, transportable and provides a
non-intrusive test (that is, no punctures have to be done to the
chest cavity). Similarly, one can try to determine the rate of
pumping of the heart using this kind of equipment. Notc that the
pulse only determines the rate of contracting and expanding of the
heart but not how much blood is being pumped by it. Another
completely different problem arises in the determination of the
existence and lengths of internal cracks in a plate, by using
electrostatic measurements on the boundary [FV], [BCW], [W]. These
three are examples of the following inverse problem. (The best
reference for the general facts about this problem is the supply
[SU]. See also the nice explanation for the general public [C],
[S]):
Assume/5 is a strictly positive (nice) function in the closed unit
disk D. If we were to introduce a current at the boundary OD,
represented by a function ~ satisfying fOB Cds = 0, then the
Neumann problem
div (/3 g radu) = 0 i n D (39)
/3~ = r on OD
has a solution u which is unique up to an additive constant. If r
is a nice function then ~ (that is, the tangential derivative of u)
is well defined on OD, so we have the input-output map
0u
which is a linear continuous map from the Sobolev space H~(OD) into
itself. (This statement holds for any domain D with nice boundary,
not just the disk.)
Consider now the (very non-linear) map
/3, >A/~ (40)
is it injective? Can one find the inverse to this map? This problem
was originally posed by A. Calder6n, who proved that (40) was
locally invertible near/3 = constant, more recently Nachman IN1],
IN2] proved global invertibility. Since/3 is usually called the
conductivity and 1//3 the impedance, this is the reason for the
name EIT of this inverse problem. In the biological applications we
know the value/3 for the different constituents like blood, lung
tissue, etc., so one only looks for a profile of the areas occupied
by them. In the determination of cracks, one can assume/3 "known",
except for curves where/3 = 0, and one wants to determine this
curve, or whether any exists. One can find in [SU] many important
inverse problems that are equivalent to EIT: in acoustics,
radiation scattering, etc. Note that in the problem of the rate of
pumping of the heart, we can think that all we want to determine is
just a single number, this rate. Isaacson, Newell and collaborators
have in fact patented [C], [I] a device that measures this rate
with the help of EIT. We also know that this problem, being an
inverse elliptic problem is very ill-conditioned, so in any case
one is willing to restrict oneself to find the deviation of/3 from
an assumedly known conductivity/30. In the simplest case we
assume/30 --= 1, so that/3 = 1 +5/3, 15/31 < < 1, and we
further assume
28
5/3 = 0 on OD (One can always reduce matters to this case). Thus u
= U + 5U, where U is the solution of (39) for the same boundary
value, and/3 = 1. In other words
AU = 0 in D (41) OU = r on OD
Here A is the Euclidean Laplacian. The perturbation 5U then
satisifes
A(SU) = - < grad (&3), grad U > in D o(~u) (42)
o~ -- -(8/~)~b on OV
We have at our disposal the choice of inputs r Their only
restriction is that rOD ~bd8 : O. For that reason, they can be well
approximated by linear combinations of dipoles. A dipole at a point
oo c OD is given by o
It turns out that the solution Uw of
{ AU~ = 0 i n D (43) -r~~ on cOD
has level curves which are arcs of circles passing through w and
perpendicular to 0D. That is, the level curves of U~, are exactly
the geodesics of the hyperbolic metric. This fact passd unnoticed
to Barber and Brown but they definitely realized that the
value
O(SU) (44) # - Os
at a point a E OD must be some sort of integral of 53 over the
level curve of U~ that ends at a, precisely the geodesic starting
at w and ending at a. In other words, # is a function in the space
of geodesics in D considered as the hyperbolic plane, all the
geodesics are obtained this way by changing w and a. Without
expressly stat- ing this, Barber and Brown introduced a
"backprojection" operator that turned out to be exactly R#H and
gave the approximation to ~ as R#H#. Santosa and Vogelius
recognized explicitly that some sort of Radon transform was
involved and used the generalized Byelkin transform and a
convenient choice of K in (30) to stabilize nu- merically the
inversion of EIT. Casadio and I, prodded by a question of Santosa
and Vogelius, saw that RH was involved and developed the inversion
formula (38) for this purpose. As it turns out, all of these
approaches are just approximations to the lin- earized problem.
Only in [BC3], [BC4], we realized the fact that the exact
formulation of the linearized problem in terms of hyperbolic
geometry requires also a convolution operator! Namely, let
~r = c~ - 3 cosh-4(t) (45) 8 r
and # the boundary data (44) considered as a function on the space
of geodesics in D, then one has that the exact relation between 5/?
and # is given by
RH(a *H ~/~) = # (46)
R# # = R# RH(g *H ~1~) (47)
29
1 G a H ( S *H (n~,,)) = ~ *H 6fl (48)
which requires to invert the convolution operator of symbol ~. One
can compute its hyperbolic Fourier transform k exactly and find out
that k(.~) :fi 0 for every A E R, so that the operator ~. is, in
principle, invertible, but the numerical implementation of this
inversion has proven difficult so far. (Although Kuchment and his
students have made in [FMLKMLPP] some progress towards implementing
a numerical Fourier transform in D, which we hope will prove useful
to compute 6ft.) One can recognize in (47) and (48) the same
principle that lead to the numerical approach in [BB1], [SV] and
others. Due to the importance of this problem there have been many
other interesting approximate inversion formulas, under special
assumptions on the conductivity fl, for instance, fl is "blocky",
that is the linear combination with positive coefficients of a
finite number of disjoint squares [DS]. Their approach is
variational, and one may wonder whether one could not use some
version of the Mumford-Shah edge detection algorithms [MS] to
obtain a rather sharp solution of the inverse conductivity problem
(40).
5. F inal r e m a r k
The objective of these short notes (and the corresponding CIME
course) was only to indicate how, beyond the well-known
applications of tomography to Medicine, there are many other
possible ones. Moreover, even to solve them approximately, they
require deep mathematical tools, showing once more that the
applicability of "pure" and "abstract" mathematics is not a
fairy-tale but a concrete reality. It also indicates that it pays
to "invest" one's time trying to communicate with those, be they
physicists, or physicians, etc., that have the ready made
applications. A lesson often lost by graduate students in
Mathematics.
30
6. References
[AA] S. Andrieux and A. Ben Alda, Identification de fissures planes
par une donn~e de bord unique, C.R. Acad. Sci. Paris 315 I(1992),
1323-1328.
[AU] A. Aldroubi and M. Unser, editors, "Wavelets in Medicine and
Biology," CRC Press, 1966, 616 pages.
[BB1] D. C. Barber and B. H. Brown, Recent developments in applied
potential, in "Information processing in Medical Imaging," S.
Bacharach (ed.), Martinus Nijhoff, 1986, 106-121.
[BB2] D. C. Barber and B. H. Brown, Progress in Electrical
Impendance Tomog- raphy, in "Inverse problems in partial
differential equations," D. Colton et al. (eds.), SIAM, 1990,
151-164.
[BC1] C. A. Berenstein and E. Casadio Tarabusi, Inversion formulas
for the k- dimensional Radon transform in real hyperbolic spaces,
Duke Math. J. 62 (1991), 613-632.
[BC2] C. A. Berenstein and E. Casadio Tarabusi, Range of the
k-dimensional Radon transform in real hyperbolic spaces, Forum
Math. 5 (1993), 603-616.
[BC3] C. A. Berenstein and E. Casadio Tarabusi, The inverse
conductivity prob- lem and the hyperbolic x-ray transform, in "75
years of Radon transform," S. Gindikin and P. Michor, editors,
International Press, 1994, 39-44.
[BC4] C. A. Berenstein and E. Casadio Tarabusi, Integral geometry
in hyperbolic spaces and electrical impedance tomography, SIAM J.
Appl. Math. 56 (1996), 755-764.
[BCW] C. A. Berenstein, D. C. Chang and E. Wang, A nondestructive
inspection method to detect a through crack by electrostatic
boundary measurements, ISR- TR 96-1.
[BW1] C. A. Berenstein and D. Walnut, Local inversion of the Radon
transform in even dimensions using wavelets, in "75 years of Radon
transform," S. Gindikin and P. Michor, editors, International
Press, 1994, 45-69.
[BW2] C. A. Berenstein and D. Walnut, Wavelets and local
tomography, in "Wavelets in Medicine and Biology," A. Aldroubi and
M. Unser, editors, CRC Press, 1966.
[BQ] J. Boman and E. Quinto, Support theorems for real analytic
Radon transforms, Duke Math. J. 55 (1987), 943-948.
[BV] K. M. Bryant and M. Vogelius, A computational algorithm to
detect crack locations from electrostatic boundary measurements,
Int. J. Eng. Sci. 32 (1994), 579-603.
[By] G. Beylkin, The inversion problem and applications of the
generalized Radon transform, Comm. Pure Appl. Math. 37 (1984),
579-599.
31
[BCR] G. Beylkin, R. Coifman, and V. Rokhlin, Fast wavelet
transforms and nu- merical algorithms I, Comm. Pure Appl. Math. 44
(1991), 141-183.
[C] B. Cipra, Shocking images from RPI, SIAM News, July 1994,
14-15.
[D] I. Daubechies, "Ten lectures on wavelets," SIAM, 1992.
[DB] A.H. Delaney and Y. Bresler, Multiresolution tomographic
reconstruction using wavelets, ICIP-94, 830-834.
[DO] J. DeStefano and T. Olson, Wavelet localization of the Radon
transform, IEEE Trans. Signal Proc. 42 (1994), 2055-2057.
[DS] D. C. Dobson and F. Santosa, An image enhancement technique
for electrical impedance tomography, Inverse Problems 10 (1994),
317-334.
[FRK] A. Faridani, E. Ritman and K. T. Smith, Local tomography,
SIAM J. Applied Math. 52 (1992), 1193-1198.
[FLBW] F. Rashid-Farrokhi, K. J. R. Liu, C. A. Berenstein and D.
Walnut, Wavelet- based multiresolution local tomography, ISR-TR
95-73, see also ICIP-95, Wash- ington, DC.
[FLB] F. Rashid-Farrokhi, K. J. R. Liu and C. A. Berenstein, Local
tomography in fan-beam geometry using wavelets, ICIP-96,
Laussane.
[FMP] B. Fridman, D. Ma, and V. G. Papanicolau, Solution of the
linearized inverse conductivity problem in the half space, preprint
Wichita St. U., 1995.
[FMLKMLPP] B. Fridman, D. Ma, S. Lissianoi, P. Kuchment, M.
Mogitevsky, K. Lancaster, V. Papanicolaou, and I. Ponomaryov,
Numeric implementation of harmonic analysis on the hyperbolic disk,
in preparation.
[FV] A. Friedman and M. Vogelius, Determining cracks by boundary
measurements, Indiana U. Math. J. 38 (1989), 527- 556.
[GG] I. M. Gelfand and S. Gindikin, editors, "Mathematical problems
of tomogra- phy," AMS, 1990.
[GM] S. Gindikin and P. Michor, editors, "75 years of Radon
transform," Interna- tional Press, 1994.
[GIN] D. Gisser, D. Isaacson, and J. Newell, Current topics in
impedance imaging, Clin. Phys. Physiol. 8 (1987), 216-241.
[GS] V. Guillemin and S. Sternberg, "Geometric asymptotics," AMS,
1977.
[He1] S. Helgason, "The Radon transform," Birkh~user, 1980.
[He2] S. Helgason, "Groups and geometric analysis," Academic Press,
1984.
[He3] S. Helgason, "Geometric analysis on symmetric spaces," AMS,
1994.
32
[Ho] M. Holschneider, Inverse Radon transform through inverse
wavelet transforms, Inverse Problems 7 (1991), 853-861.
[J] F. John, "Plane waves and spherical means," Springer-Verlag,
reprinted from originial edition Interscience, 1955.
[Ka] G. Kaiser, A friendly guide to wavelets, Birkhguser,
1994.
[KS] A. C. Kak and M. Slaney, "Principles of computerized
tomographic imaging," IEEE Press, 1988.
[KaS] P. G. Karp and F. Santosa, Non-destructive evaluation of
corrosion damage using electrostatic measurements, preprint
1995.
[KR] A. I. Katsevich and A. G. Ramm, New methods for finding values
of jump of a function from its local tomography data, Inverse
Probl. 11 (1995), 1005-1023.
[Ke] F. Keinert, Inversion of k-plane transforms and applications
in computer to- mography, SIAM Riview 31 (1989), 273-289.
[KLM] P. Kuchment, K. Lancaster and L. Mogilevskaya, On local
tomography, In- verse Problems 11 (1995), 571-589.
[KSh] P. Kuchment and I. Shneiberg, Some inversion formulas for
SPECT, Applica- ble Analysis 53 (1994), 221-231.
[Kul] A. Kurusa, The Radon transform on hyperbolic space,
Geometriae Dedicata 40 (1991), 325-336.
[Ku2] A. Kurusa, Support theorems for the totally geodesic Radon
transform on constant curvature spaces, Proc. Amer. Math. Soc. 122
(1994), 429-435.
[M] Y. Meyer, "Ondelettes et op&ateurs," 3 vols., Herman,
1990.
[MS] J. M. Morel and S. Solimini, "Variational methods in image
segmentation," Birkhguser, 1995.
[N1] A. I. Nachman, Reconstruction from boundary measurements,
Annals Math. 128 (1988), 531-576.
IN2] A. I. Nachman, Global uniqueness for a two-dimensional inverse
boundary value problem, Annals Math. 143 (1996), 71-96.
[Na] F. Natterer, "The mathematics of computerized tomography,"
Wiley, 1986.
[O] T. Olson, Optimal time-frequency projections for localized
tomography, Annals of Biomedical Engineering 23 (1995),
622-636.
[Q1] E. T. Quinto, Tomographic reconstruction from incomplete
data-numerical in- version of the exterior Radon transform, Inverse
Problems 4 (1988), 867-876.
[Q2] E. T. Quinto, Singularities of the X-ray transform and limited
data tomography in l:t 2 and R 3, SIAM J. Math. Anal. 24 (1993),
1215-1225.
33
[Q3] E. T. Quinto, Computed tomography and rockets, Springer
Lecture Notes in Math. 1497 (1991), 261-268.
[QCK] E. T. Quinto, M. Cheney, and P. Kuchment, eds., "Tomography,
impedance imaging, and integral geometry," Lect. Appl. Math. 30,
Amer. Math. Soc., 1994.
[R]B. Rubin, Inversion and characterization of Radon transforms via
continuous wavelet transforms, Hebrew Univ. TR 13, 1995/96.
[RS] A. Ramm and A. I. Zaslavsky, Singularities of the Rdaon
transform, Bull. Amer. Math. Soc. 25 (1993), 109-115.
IS] F. Santosa, Inverse problem holds key to safe, continuous
imaging, SIAM News, July 1994, 1 and 16-18.
[ST] H. Schonberg and J. Timmer, The gridding method for image
reconstruction by Fourier transformation, IEEE Trans. Medical
Imaging 14 (1995), 596-607.
[SCII] E. Sommersalo, M. Cheney, D. Isaacson, and I. Isaacson,
Layer stripping: a direct numerical method for impedance imaging,
Inverse Probl. 7 (1991), 899-926.
[SU] J. Sylvester and G. Uhlmann, The Dirichlet to Neumann map and
applications, in "Inverse problems in partial differential
equations," D. Colton et al., eds., SIAM, 1990, 101-139.
[SV] F. Santosa and M. Vogelius, A backprojection algorithm for
electrical impedance imaging, SIAM J. Appl. Math. 50 (1990),
216-243.
[W] D. Walnut, Applications of Gabor and wavelet expansions to the
Radon trans- form, in "Probabilistic and stochastic methods in
analysis," J. Byrnes et al., ed., Kluwer, 1992, 187-205.
[Wa] E. Wang, Ph.D. thesis, University of Maryland, College Park,
1996.
[ZCMB] Y. Zhang, M. A. Coplan, J. H. Moore and C. A. Berenstein,
Computerized tomographic imaging for space plasma physics, J. Appl.
Phys. 68 (1990), 5883- 5889.
Institute for Systems Research University of Maryland College Park,
MD 20742
[email protected]
H O L O M O R P H I C M A P P I N G S B E T W E E N R E A L
A N A L Y T I C S U B M A N I F O L D S IN C O M P L E X S P A C
E
PETER EBENFELT
I00 44 Stockholm Sweden
1. Introduction 2. Preliminaries on algebraic mappings and CR
geometry
2.1. Algebraic mappings 2.2 Real analytic CR submanifolds in C N
2.3 Finite type and minimality 2.4 Normal forms for generic
submanifolds 2.5 The complexification of a real analytic
submanifold
3. Holomorphic nondegeneracy, finite nondegeneracy, and re-
flection identities for holomorphic mappings 3.1. Holomorphic
nondegeneracy of real analytic CR sub-
manifolds 3.2. Finite nondegeneracy of real analytic CR
submanifolds 3.3. Reflection identities for holomorphic
mappings
4. The Segre sets 4.1. The Segre sets of a real analytic CR
submanifold 4.2. Homogeneous submanifolds of CR dimension 1 4.3.
Proof of Theorem 4.1.21 (CR dimension 1 case)
5. An application to holomorphic mappings between real al- gebraic
submanifolds 5.1. A reformulation of Theorem 1.1 5.2. Proof of
Theorem 5.1.1
6. Other applications and concluding remarks 6.1. The algebraic
equivalence problem 6.2. Uniqueness of biholomorphisms between real
analytic,
generic submanifolds
1. INTRODUCTION
In this paper we shall present, in fairly self-contained form, some
recent ideas and concepts concerning real analytic submanifolds in
C N. One of the main construc- tions described here is an invariant
sequence of sets--called the Segre sets--at tached to a real
analytic submanifold in cN; this sequence of sets was introduced in
joint work by the author together with Baouendi and Rothschild
[BER1]. The first Segre set coincides with the so-called Segre
variety, introduced by Segre [Seg] and suc- cessfully used in
mapping problems for real analytic hypersurfaces by a number of
authors: e.g. Webster [W1], Diederich-Webster [DW],
Diederich-Fornaess [DF], Huang [H2], and others. Subsequent Segre
sets turn out to be unions of Segre varieties.
One of the merits of the Segre sets is that they allow one to
analyze "reflection identities" (see section 3.3) for higher
codimensional submanifolds--the idea of using reflection identities
to analyze CR mappings goes back to e.g. Lewy ILl,
Baouendi-Jacobowitz-Treves [BJT], Baouendi-Rothschild [BR1]. The
Segre sets also allow a new characterization of the notion of
finite type, as introduced by Bloom-Graham [BG]. In fact, the Segre
sets provide a way of constructing the CR orbits of a real analytic
CR submanifold without solving any differential equations (see
Theorem 4.1.21). As a result, one finds e.g. that the CR orbits of
a real algebraic CR submanifold are algebraic (Corollary
4.1.31).
As an application and illustration of these techniques, we shall
prove the following result (which we shall also reformulate in
terms of more c]assical CR geometry, Theorem 5.1.1) from
[BER1].
T h e o r e m 1.1 ( [BER1]) . Let A C C N be an irreducible real
algebraic set, and Po a point in A such that Po C Ares. Suppose the
following two conditions hold.
(1) There is no hoIomorphic vector field (i.e. a vector field with
holomorphic coe~cients and values in T I '~ which is tangent to an
open piece of
Areg. (2) If f is a germ, at a point in A, of a holomorphic
]unction in C N such that
the restriction of f to A is real valued, then f is constant.
Then, if H is a holomorphic map from an open neighborhood in C N of
po E A into C N, with Jac H ~ O, that maps A into another real
algebraic set A' with dim A' = dim A, necessarily the map H is
algebraic.
The first result along these lines goes back to Poincar6 [P] who
proved that a biholomorphic map H: C 2 ~-* C 2 defined near a point
on a sphere S C C 2 and mapping S into another sphere S' C (;2 is
rational. This was later extended to mappings between spheres in C
g by Tanaka ITs]. Webster [W1] then proved that a biholomorphic
mapping H: C g H C N, defined in some open subset of C N, tak- ing
a real algebraic, somewhere Levi-nondegenerate (i.e. with
nondegenerate Levi form at some point) hypersurface M into another
real algebraic hypersurface M ' is algebraic. Recently, Baouendi
Rothschild [BR3] showed that if the real algebraic hypersurface M
satisfies condition (1) of Theorem 1.1 above (which is a weaker
condition than being Levi nondegenerate somewhere) then any
holomorphic map- ping H : C N ~-* C N, defined in some open subset
of C N and with Jac H ~ 0, taking M into another real algebraic
hypersurface M I is algebraic. Moreover, they show
37
that this condition is also necessary for such a conclusion to hold
in the sense that if (1) is violated then there is a non-algebraic
biholomorphism of M into itself. The sufficiency of condition (1)
in this result by Baouendi-Rothschild is contained in Theorem 1.1
above, because a real analytic hypersurface that satisfies
condition (1) automatically satisfies condition (2). We would also
like to mention that the conditions (1) and (2) in Theorem 1.1 are
essentially necessary for the conclusion of the theorem to hold. We
refer the reader to [BER1] for details on this (see also section
6).
Condition (1) was first introduced, and named holomorphic
nondegeneracy (see section 3.1 for a detailed treatment of this
notion), by Stanton [Stl] in connection with the study of
infinitesimal CR automorphisms of real hypersurfaces. It deserves
to be mentioned here that holomorphic nondegeneracy is fairly easy
to verify be- cause it turns out to be closely related to another
property (finite nondegeneracy, see section 3.2), which is very
computational and is a direct generalization of Levi
nondegeneracy.
Results of the type above for mappings between hypersurfaces in
different di- mensional spaces have been obtained by e.g. Webster
[W2], Forstneric IF], and Huang [H1]. For higher codimensional
submanifolds, work has been done by e.g. Senkin-Tumanov [TH],
Tumanov [Wu2], and Sharipov-Sukhov [SS].
Other applications to rigidity properties of holomorphic mappings
between real analytic submanifolds will be briefly discussed in the
last section of this paper.
The paper is organized as follows. In chapter 2, we give the basic
definitions and facts, mostly without proofs, concerning algebraic
mappings and CR geometry. More recent concepts such as holomorphic
nondegeneracy and finite nondegener- acy, as well as reflection
identities for holomorphic mappings, are introduced and discussed
in chapter 3. The construction of the Segre sets and proofs of the
main results concerning these are given in chapter 4. In chapters
5-6, applications of the techniques are discussed. A reformulation
of Theorem 1.1 and a proof is given in chapter 5. Applications to
uniqueness questions and some open problems are discussed in the
final chapter.
A c k n o w l e d g e m e n t . The author would like to thank
Professors M. S. Baouendi and L. P. Rothschild for agreeing to have
results and arguments from our joint papers [BER1-3] included in
these notes. As the reader will no doubt notice, the results
presented here, for which the author can claim any credit, are due
to this above mentioned joint work.
2. PRELIMINARIES ON ALGEBRAIC MAPPINGS AND CR GEOMETRY
2.1. A lgeb ra i c m a p p i n g s . We denote by ON(po) the ring
of germs of holomor- phic functions in C g at p0, and by .AN (P0)
the subring of ON (Po) consisting of those germs that are also
algebraic, i.e. those germs for which there is a nontrivial poly-
nomial P(Z,x) e C[Z,x] (with Z E C N and x C C) such that any
representative f(Z) of the germ satisfies
(2.1.1) P(Z, f(Z)) =_ O.
In particular, any function in .AN(Po) (throughout this paper we
shall, without comment, identify a germ of a function with some
representative of it) extends
38
as a possibly multi-valued holomorphic function in C N \ V, where V
is a proper algebraic variety in C N. We list here some basic
properties of algebraic holomorphic functions that will be used in
the proof of Theorem 1.1. We use the notat ion AN for AN(O).
L e m m a 2.1.2. The following holds:
(i) If f E `AN then O~" f E r for any multi-index a. (ii) I f f G
`AlV and gj E ,AN with gj(O) = O, for j = 1, ..., I(, then
f (g l (Z) , ...,9K(Z)) E `AN.
(iii) (The Algebraic Implicit Function Theorem) Let F( Z, x) be an
algebraic holo- morphic function near 0 in C y x C, i.e. F E
.,4N+1, and assume that
OF F(o, o) = o , -g-;(o, o) # o.
Then there is a unique function f ~ .AN such that x = f ( Z )
solves the equation F ( Z , x ) = O, i.e.
F(Z, f ( Z ) ) = 0
The arguments needed to prove this lemma are standard (see e.g.
[BM], and also [BR3] for further properties of algebraic
functions), and the proof is omitted.
We say that a germ of a holomorphic mapping H : C N --~ C K at p0
is algebraic if the components of H (we write H = (H1,..., Hh'))
are all algebraic. It follows from Lemma 2.1.2 (ii) that this
property is invariant under algebraic changes of coordinates in C x
and C K at P0 and p~ = H(p0), respectively.
2.2. R e a l a n a l y t i c C R s u b m a n i f o l d s in C N. In
sections 2.2-2.5, we shall set up the notation, and give the basic
definitions and results from CR geometry needed for subsequent
sections. Most facts and results in these sections will be stated
informally, and without proofs. Unless otherwise specified, proofs
can be found in e.g. [B].
We should point out that only real analytic submanifolds will be
considered. The definitions presented in these sections can be made
in the broader category of smooth (e.g. C ~176 submanifolds, but
some of the facts stated fail to be true in that general setting.
For instance, the two notions "finite type" and "minimality"
presented in section 2.3 coincide for real analytic CR
submanifolds, but do not coincide in general for merely smooth CR
submanifolds.
Let M be a real analytic submanifold in C N and P0 a point in M.
Let m be the (real) codimension of M. We may describe M near P0 as
the zero locus
(2.2.1) M = {Z c cN: p(z, 2) = 0},
where p = (Pl ..... Pro) are real valued, real analytic functions
near P0 with linearly independent differentials dpl, ..., dpm; we
use the notation h(Z, 2) for a real analytic function in C N to
indicate that we think of such objects as restrictions to ~ = Z
of
39
holomorphic functions of (Z, 4) 6 C N • C N. We say that M is real
algebraic if the pj can be taken to be real valued polynomials in Z
and 2.
The complex tangent space of M at p 6 M is defined as
(2.2.2) T;(M) = Tp(M) A Jp(TR(M))
where Jp: Tp(C N) ~ Tp(C N) denotes the complex structure in C N.
The (real) dimension of Tp(M) is even and satisfies
(2.2.3) 2N -- 2m <_ d im~T; (M) < 2N - m.
If M is a hypersurface, i.e. a real codimension 1 submanifold, then
d i m i T y ( M ) is necessarily 2N - 2. In general, the dimension
varies with p 6 M.
D e f i n i t i o n 2.2.4. M is said to be CR at p G M if dim~Tq(M)
is constant for all q in some neighborhood ofp in M.
We decompose d as d = 0 + c3. The real analytic submanifold M is CR
at p 6 M if and only if the rank of the covectors Opl, ..., Opm is
constant at all q 6 M near p. Hence, any real analytic submanifold
M C C N is CR outside a proper real analytic subset V C M. We say
that a connected submanifold is CR if it is CR at every
point.
The complexified complex tangent space CTp(M) = C | T;(M)
decomposes as
the direct sum T(I '~ + T(~ where
T(I '~ = CTp(M) A Tp(l'~ N) (2.2.5)
T(~ : CTp(M) N T(~
here, T( I '~ and T(~ N) denote the spaces of (1,0)-vectors and
(0,1)- vectors in C N respectively, and CTp(M) denotes the
eomplexified tangent space of M, G | Tp(M). If M is CR then
(2.2.6) T(~ = O Tp (~ pcM
forms a complex vector bundle over M, called the CTl bundle.
Sections of the CR bundle are called CR vector fields. A function
(or distribution) defined on (a piece of) M is said to be a CR
function (or CR distribution) if it is annihilated by all the CR
vector fields on (that piece of) M. The restriction to M of a
holomorphic function defined in a neighborhood of M is CR, and any
real analytic CR function on M is the restriction of such a
holomorphic function. In general, there are smooth (C ~ ) CR
functions on M that are not restrictions of holomorphic
functions.
D e f i n i t i o n 2.2.7. M, of codimension m, is said to be
generic at p E M if the dimension of T;(M) is minimal, i.e. d i m ~
T ; ( M ) = 2 N - 2m.
The real analytic submanifold M is generic at p 6 M if and only if
the rank of the covectors Opl,...,Opm equals m at p. In particular,
if M is generic, then it is also CR.
40
If M is CR, we call the complex dimension of T(~ (= d imiTy(M) /2 )
the CR dimension of M and the real dimension of T~(M)/T;(M) the CR
codiraen- sion of M. For a generic submanifold M C C N, the CR
codimension equals the codimension rn and
(2.2.8) N = CR dim (M) + CR codim (M).
Another important fact is that a real analytic CR submanifold M is
generic if and only if it is not contained in a proper complex
submanifold of C N. If _~4 is not generic, then there is proper
complex submanifold 2( C C N (unique if we consider X as a germ of
a submanifold) such that M is generic as a real analytic
submanifold of X. We shall refer to X as the intrinsic
complexifieation of M.
The Levi form of a CR submanifold M at p ~ M is the (vector valued)
quadratic
form on T(p~ defined as follows
(2.2.9) T(p~ ~ Lp ~ ~r,([L,L]) E CTp(M)/CT;(M),
where rr, is the projection 7r,: CTp(M) ~ CTp(M)/CT;(M) and L is
some section
of T(~ that equals Lp at p. A real hypersurface M is said to be
Levi nonde- generate at p E M if the quadratic form (2.2.9) is
nondegenerate. In this paper, we shall not impose any conditions on
the Levi form of a CR submanifold. Instead, we shall introduce (in
section 3.1) a weaker condition, which also turns out to be
essentially necessary in most applications we shall consider.
We conclude this section by giving a few examples of generic, CR,
and non-CR submanifolds.
E x a m p l e 2.2.10. A complex subrnanifold of C N is CR but not
generic.
E x a m p l e 2.2.11. A real hypersurface in C N is generic.
E x a m p l e 2.2.12. Consider the real analytic 4-dimensional
submanifold M C C 3 defined by
(2.2.13) I m Z 3 - l Z l ] 2 - l Z 2 1 2 = 0 , I m Z 2 = 0 .
If we write pl(Z, Z) = 0 for the first equation and p2(Z, 2) = 0
for the second, then it is easy to verify that
1 Z 21dZ1 Z~dZ2 , Op2 ldz2. (2.2.14) Off1 = ~ d 3 - - = 2
The rank is two and, hence, M is generic.
E x a m p l e 2 .2 .15. Consider the real analytic 2-dimensional
submanifold M C C 2 defined by
(2.2.16) Im Z~ -]Z112 = 0 , Re Z2 = 0.
As above, we write pl(Z ,Z) = 0 for the first equation and p2(Z,Z)
= 0 for the second. We find that
(2.2.17) Opl = ~dZ2 - 2~dZ~ , Op2 = dZ2.
Outside {Z~ --- 0} the rank is two, and along {Z1 --= 0} the rank
is one. Thus, M is generic outside the origin but is not even CR at
the origin.
41
2.3. F i n i t e t y p e a n d m i n i m a l i t y . Let M C C N be
a real analytic CR subman- if old.
Definit ion 2.3.1. M is said to be of finite type at Po E M if the
CR vector fields, the complex conjugates of the CR vector fields,
and their commutators, evaluated at Po, span CTp0(M ).
Equivalently, M is said to be of finite type at P0 E M if the
sections of TO(M) and their commutators, evaluated at p0, span Tp0
(M).
More generally, we define the Harmander numbers of M at P0 E M as
follows. We let E0 = T~o(M) and Pl the smallest integer > 2 such
that the sections of TO(M) and their commutators of lengths < #1
evaluated at P0 span a subspace E1 of Tpo(M ) strictly bigger than
E0. The multiplicity of the first Hhrmander number #1 is then gl =
dim~E1 - dim~E0. Similarly, we define/*2 as the smallest integer
such that the sections of of TO(M) and their commutators of lengths
</*2 evaluated at P0 span a subspace E~ of Tpo(M) strictly
bigger than El , and we let ga = dim~E2 - dimt~E1 be the
multiplicity of/*2. We continue inductively to find integers 2 <
/ .1 < #2 < . . . </**, and subspaces T~o(M ) = Eo C E1 C
. . . C E, C Tp0(M), where E , is the subspace spanned by the Lie
algebra of the sections of Tr evaluated at P0. The multiplicity gj
of each /*j is defined in the obvious way as above. It is
convenient to denote by ml < m2 _< ' " _< mT the H6rmander
numbers with multiplicity by taking ml = ms . . . . . mt~ =/ .1 ,
and so on. The number r coincides with the CR codimension of M if
and only if M is of finite type
at P0.
Definit ion 2.3.2. M is said to be minimal at Po E M if M contains
no proper CR submanifold through po with the same CR dimension as
M.
We define the (local) CR orbit of po in M as the Nagano leaf of the
CR vector fields through P0. The CR orbit of p0 is a minimal CR
submanifold Wpo C M through P0 such that T;o(M ) C Tpo(Wpo). In
fact, using the notation above, the tangent space Tpo(Wpo ) equals
the space E, . For a real analytic CR submanifold M, the notions
introduced above are related as follows:
M is minimal at po ~=:=>
Wp0 contains an open neighborhood of P0 in M r
M is of finite type at P0.
We refer the reader to [Tul] and the paper by ~hamartov in these
Proceedings for further reading on minimality and its connection to
wedge extendibility of CR functions.
E x a m p l e 2.3.3. Consider the real analytic hypersurface M C C
2 defined by
(2.3.4) Im Z2 = (Re Z2)IZll 2.
Note that the complex hyperplane {Z2 = 0} is contained in M. It
follows that M is not minimal (not of finite type) at the points
(Z1,0) E M. However, M is minimal (of finite type) at all other
points.
42
If M is a real anMytic hypersurface and M is not minimal at a point
P0 E M, then M contains a complex hypersurface through tha t point.
This follows from the fact tha t the CR orbit of P0 in M, being a
proper CR submanifold of M with CR dimension N - 1 , necessarily
has dimension 2 N - 2 and, hence, is a complex manifold by the
Newlander-Nirenberg theorem. Also, if a real analyt ic hypersurface
M is not minimal at most points (outside a proper analyt ic
subvariety), then M is Levi flat (i.e. M is locally biholomorphical
ly equivalent to a real plane or, which is the same, the Levi form
of M vanishes identically). Both of these facts fail in general for
higher codimensionat CR submanifolds as the following example i l
lustrates.
E x a m p l e 2.3.5. Consider the codimension two, real analytic ,
generic submanifold M C C a defined by
(2.3.6) Im Z 3 ---- IZ112 , Im Z2 = 0.
M is fol iated by the CR submanifolds Nx, of the same CR dimension
as M , defined by
(2.3.7) Im Z 3 : I N 1 12 , Z 2 : X
for x C R. Hence, it is not minimal at any point. Also, M is not
locally biholomor- phically equivalent to a real plane, and is
therefore not Levi flat.
2.4. N o r m a l f o r m fo r g e n e r i c s u b m a n i f o l d s
. In this section, we shall describe a convenient normal form for
generic real analyt ic submanifolds. Let M C C N be a connected
such submanifold. We write rn = CR codim (M) ( = codim (M) ) and n
= C R d i m ( M ) , so that N = n + r n . We let P0 be a p o i n t
in M. Then there are holomorphic coordinates Z = (z, w) vanishing
at P0, with z E C ~ and w E C m, such tha t M is given near P0 = 0
by
(2.4.1) Im w = r z, Re w),
where r X,s) is a Cm-valued holomorphic function, Nm-valued for X =
2 and s E R m, such tha t
(2.4.2) r -- r X,s) _= 0.
Such coordinates are called normal coordinates, and (2.4.1) is
called normal form for M; note that all examples given so far have
been in normal form. We shall sketch a proof of the existence of
normal coordinates at the end of this section. We refer the reader
to [CM] or [BJT] for a detai led proof of the existence.
If M is real algebraic, then there are algebraic normal coordinates
at P0, i.e. the change of coordinates is algebraic and the function
~(z, 2, s) is algebraic. This follows readi ly from the proof of
the existence of normal coordinates; as we shall see, the proof is
based on an appl icat ion of the implicit function, which preserves
algebraici ty in view of Lemma 2.1.2.
By wri t ing Re w = (w + t~)/2 and Im w = (w - w) /2 i , we can
solve for w in (2.4.1) by the implicit function theorem. We find
that M consists of the set of points (z, w) for which
(2.4.3) w = Q(z, 2, w),
43
where Q(z, x , T) is a Cm-valued holomorphic function. It is
straightforward to check that (2.4.2) implies
(2.4.4) O(z, O, v) =_ Q(O, x, v) - r.
Note that (2.4.3) is not a R'~-valued equation for M. However,
there is an m x m matrix-valued function a(z, w, X, r) such
that
(2.4.5) w - Q(z, 2, ~) = a(z, w, 2, e ) ( I m w - r 2, Re
w)).
By complex conjugating (2.4.3), we can also describe M by the
equation
(2.4.6) ~ = O(~, z, w);
(2.4.7) h(Z) = h(Z)
for a holomorphic function h(Z). An explicit basis L1, ...,L,, for
the CR vector fields on M near p0 can be given
in normal coordinates as follows
(2.4.8) Lj = ~zj + ~)k,~(2, z ,w ) , j = 1 , . . . ,n, k = l
where we use the notation
(2.4.9) Ok,~ (7, z, w) = ~ ( ~ , z, w).
Observe that the vector fields L1,..., Ln all commute. We conclude
this section by sketching a proof of the existence of normal
coordi-
nates. We assume that M is given by (2.2.1) near P0. We may assume,
by applying an affine change of coordinates in C N, that p0 = 0 and
that
i - (2.4.10) dpj(0, 0) = ~(dZ~+ i - dZn+j).
We write z = (Z1,..., Z,,), w = (Zn+l,. . . , Z~+m), and Z = (z,
w). I