AVERAGES IN THE PLANE OVER CONVEX CURVES

AND MAXIMAL OPERATORS

By

J. B O U R G A I N

1. I n t r o d u c t i o n

The main goal of this paper is to give a somewhat simpler exposition of the result
on circular means proved in [1] and extensions to more general curves. The
simplifications with respect to [1] consists in avoiding the more combinatorial part
and most of the interpolation. The basic approach however is the same and also
here information coming from both harmonic analysis and geometry will be
combined.
Let [' be the boundary of a compact convex centrally symmetric body in R 2. Thus
also U = { x ER2,11xH = 1}, where II II is the norm on R 2 induced by this convex
symmetric body. We assume [" sufficiently smooth and of non-vanishing curvature
(extensions will be discussed later). Let ~r denote the arc length measure of U.

T h e o r e m 1. Let f be a bounded measurable function on the plane and define

for 0 < t < ~ the average

(1) A~(x) = f]'(x + ty)~r(dy).

Denote Mf the corresponding maximal operator

(2) M[ = sup I AJI.

Then for 2 < p < =, there is an inequality

(3) ttMfll, --< c(r,p)ll/ll,.

Here II lip denotes the L "(R2)-norm.

This result was proved [1] in the special case U in the unit circle {x E R'; I x I -- l).
Theorem 1 has the following consequence on differentiation of functions.

Corollary 2. Under the hypothesis of Theorem 1

(4) f = !iomAJ a.e.

69
JOURNAL D'ANALYSEMATHI~MATI(~HUE,VoL 47 (1986)
70 J. BOURGAIN

If F is the unit circle, Theorem 1 answers a problem posed by E. Stein and S.

Wainger in [4]. The restriction p > 2 is essential in (3), as is clear from the simple
examp[e
I
f(x) = Ix [log(l/Ix [) Xl,~,~,l
considering the operator

(5) Tf(x ) = d f(x + Ix ) y ) , ~ ( d y )

(o" = arc length measure of unit circle). Clearly I Tfl <=IM[I and Tf = ~ everywhere
while Ilfl12 < ~ .
Related operators, more precisely

T[(x)=f[(x+(l§ o- = O'r,

will play a special role in proving Theorem 1.

In ter/ns of Fourier-transforms, the study of (5) reduces to the pseudo-differential
operator

(6) f f(r162
where ~p @ 5e(R 2) satisfies if, = I on a neighborhood of 0. In this paper, we will in
particular show the LP-boundedness of (5) for p > 2 by direct geometrical
considerations rather than working with (6).

2. T h e L2-estimation

Although (3) fails for p = 2, the study of the L2-behaviour is important in our
approach. Denote P, (t > 0) the Poisson-semigroup on R 2, ~b,(sr = e -'~t, and define
for e > 0 small

MJ=s, u p l f f(x + ty)(o'* P,)(dy) I

thus replacing o- by an "e-mollification". The failure of the L2-bound of M is

"logarithmic" in the following sense.

Lemma 1.

(7) IIM ,ll2 c F,(,og )ff,ll

Relation (7) follows from a general LZ-estimate on convolution maximal
operators obtained in [2].
 CONVEX CURVES 71

temma 2. ConsiderakernelKEL'(RZ) andletK,(x)=t-2K(t-'x)[ort>O.

Define for j E Z the following quantities :

(8) a;= sup

2J~1~1~2)+l
I/~(~)1 and /3,= sup
2i~1~1~2~+l
I<vg(~),~>l.
Then

suplt,,(,lll cr(K)llrll,,
I >l) 2

denoting

r ( r ) = Z a; ''~(a, +/3,) ';2.

j~z

To derive (7), we let K = (tr * P,) - P,. Thus (tr * P,), = K, + P, and the maximal
operator associated to P, is taken care of by the Hardy-Littlewood theorem, while
the expression F(K) reduces to
,2. b,),,2
C+~aj taj+

where

a, = supldr(st)[ and b; = supl(Vdr(st),~)l.

If F is smooth (up to order 5 say) and has everywhere positive gaussian curvature,
we have a, < C2 -i'2, b, < C2 ';2 as in the circle case. This can be shown by a simple
direct computation. In fact, C. Herz has shown in [3] that the essential part of dr is
given by the expression (J st[ > 1)

P(St) + P(/5) where p(st) = [st I-"2K-"~(~')e 2"luli'-''"t

where K(~') is the curvature of F at the point with unit outer normal vector ~" and
II II. is the norm on R 2 dual to II II,

Ilxll. = sup I(x,y>[.

3. R e d u c t i o n of the p r o b l e m

We restrict ourselves in the definition (2) of M to values t _-<1 (by rescaling, this is
no loss of generality).
For k = 1,2 ..... denote ~ the it-algebras generated by the 2-k-size squares in
the plane obtained by successive diadic refinement of the unit square partition. Let
Ek be the corresponding expectation operators, thus

E~ [f] = Elf [ ~kl.

72 J. BOURGAIN

If f is a bounded measurable function on R 2, we have

f -- ~ Aft defining Aft = E,f - E,_,f.

*=1

For v E Z+ and 2-" _-<t < 2 -~+', write

f = E.tfl + ,Y_-,A4,

,a,f= A,(E.[/])+ Y- A,(A4).

A moment's reflection shows that

IA,(E. If')'-<-Cf*

where f* stands for the Hardy-Littlewood function of f.

Therefore, we are reduced to estimate

which is dominated by (fixing some p > 2)

The L*(It')-norm of latter expression is dominated by

(9) x,l:P.I IPlz/P

/ .

Our main goal will be to prove the following estimate:

(I0) IsuplA,(g)l]~ C2-*""llgll,,

P
provided E,+.[g] =0.

Once (10) is obtained, (9) may be majorized by (since p > 2)

2 -a'" I"-" ' ' 2--< 2-~',"II/II,--<Cll/ll,.

Z II~,+,/II;)
Notice that again by rescaling, it is enough to prove (10) with , = 0. Also, in this
situation, the problem trivially localizes to the case g is supported by the unit
square [0,1] 2.
Consider the following decomposition for the arc length measure o. of F:

o. = o.o+ ~ 2*-'o',
 CONVEX CURVES 73

where
or o = ,~ln~llyll<21

and

O'k = Xit~f[yll~u+2-tl--~ l l + 2 - t < l [ y f l ~ l + 2 - t + l I

(analogous to the expansion of the Dirac measure &) on R in the Haar system). Thus

(11) ,A,w0t< [f i(~ + ty)~o(y)dy[+ ~ 2'-' ]f f(~ + ty)~,(y)dyl.

k=l

It will thus suffice to have an estimate

(12, 0su<P21f "~ +',)~.(,)",l I =<r

for some constant a = a ( p ) > O. Summation over k yields then indeed the bound

I <
p )
2-~" )11111 < C(p)NJ[N

In fact, as is easily seen, an estimate

p
P

is obtained provided EJ -- O, since the initial terms in (11) can be estimated with the
inequality

If f(x +ty)o~.(y)dyl< C,,f,,p2-'"'.

Let thus k be fixed and denote n = 2 k. Consider a radius function t(x) ranging in
[1,2] and satisfying at the point x

(13) sup [f f(x + ty)o.t(y)dYl= i f

|<#<2
[(x + t(x)y)o'~(y)dyl-

Denoting
y-x and Vx -- I tTx [ (absolute value),
~'. = 0~ (t(x) )
expression (13) becomes

I f f(Y)V.(y)dy I

and we estimate its Le-norm by dualization. Consider thus a function g on

{Ixl-< c}, Ilgll0 - 1, q =p/(p-1, and write
74 J. BOURGAIN

I(f f(y)Q~(y)dy,g)[ = IIf g(x)V,ax,rll

L r (I o. t I z)

Thus the question reduces to proving an inequality :.

I[I ~,x,~.~xII ~ c~-,.~.,~, ,

9~ q
~ g q"

Let 1 < qo< q < q~ < 2. By the Marcinkicwicz interpolation theorem and the fact
that

[L%",L~,"]o,q = L ~ = [Lq,,,Lr ~ if 1 = 1 - 0 + 0 ,
q qo ql

it suffices to consider a region a C[I x IN c] and estimate (n = 2 ~)

(14) Ill Q~dxl[<=Cn-'-~'lll]'"~"

q
fl

Writing
1 -1--+0 0
q 1 2'

it follows from H61der's inequality that

0 0 0

Of course, there is the trivial inequality

(16) I[~ Q~dx[[,<=llfV"dx[I,=f llv.,,dx--~l,~l.

fl II 0

The L~-norm may be estimated using Lemma 1, One clearly has

and hence from Lemma 1

,~7, Ill ~dxll ~lLf ~dxlt ~c,o,o,~,..2

2 2 n
0 0
 CONVEX CURVES 75

Substitution of (16), (17) in the right member of (15) shows that in order to get (14),
it suffices for given 1-1 to improve a bit on (16) or on (17), i.e.

(18) f f'xdx , =< O, -'-~ [C~l

D

or

2
fl

In fact, we will split [l = flo + l'l~ where

$

s-I

each lq,., satisfying

(20)
I
[I,as

and

(21)
fl I

Then

tl th~ fll

where the first term satisfies (18) (by adding up (20)) and the second term satisfies
(19).
It should be pointed out that both (18) and (19) considered separately may fail for
suitably chosen f~ and radius function t(x).

4. G e o m e t r i c a l estimates

In this section, we first prove some estimates on (V~, Vy) and (Vx, (",) which will
be used later on. Again c, C will stand for constants (possibly depending on F), By
the curvature hypothesis on F, one has the following property:

(22) Ilpll = l=llqll~ 2P-~] = 1 - c l l p - q l l 2.

76 J. BOURGAIN

We will use the following corollary. Suppose lip [[- 1 ~ I[q [[ and lip - q II< c[Ip II,
then
(23) dist(q, Rp)- [lip - q II- [lip II-IIq II []"2lip - q I[''2.
To see this, assume lip II > Ilq II and let [[0 II, = 1 satisfy (p, 0) =[IP [[. Denote

8 = Ilq II+ lip - q II-lip II.

Writing p = q + (p - q), we get
(q,O>~llqll-8 and (p-q,o>>llp-qll-8.
Therefore, by (22)

, s i n A n g l e ( p - q , q ) , - I[[q-
~- P-q,,l- ( ~ ) '/2
IIp-q
dist(q, Re) ~ IIp - q [I [sin Angle(p - q, q)[ - 8 ,/2lip _ q I1%
proving (23).
The following lemma generalizes the estimates in the circle case (see [1]).
L e m m a 3. Denote r~ the radius t(x) introduced in previous sections when
denning Vx. Then
(1) The diameter of each of the components of Vx N Vy (vesp. of the unique
component in case of coincidence) is

(24) ---n (ix -Y n [[Ix-yH-[r,-r,[[+ /n )

-"~
(2)

(25) I< ~'.. ',:', )1-<- r + n lllx (v"

- y l v,>
l - I , . - r , ll"
(We assume I x - y l < c <89 + ry)).
FhrooL Let x =(0,1), y =(a,b), r, =[[xl[, ry =IlYlI. Assume
Ilx - y II- Ix -y [-[a F+[1-b[<'.
Let h(p) stand for [[PI[. Notice that by (23)

(26) [a [ - dist(y, Rx) -IIx - y 11"2[11x - y II- f r, - r, II ''2.

Consider the component of V, N 1/, (Vx, I/, referring to the support of the
respective functions) containing O.
For I t ] < 1/n, let p(t)= (A(t),/~(t)) be the solution of the equations

(27)
f 4,(-~(t), l-~(t))-4,(o,D--t,
1
I, ~(a - A(t), b-p(t))-cb(a,b)=O,
 CONVEX CURVES 77

belonging to the 0-component. T h e n

I/n

(28) diamcomponent(% 0 V~)~ f (I;~(t)l+ll2(t)l)dt,

-IIn

IIn
(29) 1
(V,, Vy)-- n f (lJ((t)l+ll2(t)l)dt,
-IIn

IIn

(30) ](f/., f',)l<---~ f (l~'(t)l+lt2(t)l)dt.

-I/n

Taking t-derivative in equations (27) yields

f - O=,b(- A, 1 - tL),( - ,9,4~( - A, 1 - tz)~ -- 1,

(31)
t- O~cb(a - A,b - Ix),( - O,ck(a - X,b - ~)/2 = O;

- &,~.( - A, 1 - .u,)A" - ,9,6( - A, 1 - p,)/2 =

2
- a ~ 4 ~ ( - a,1 - / z ) ( ; / )

(32) - 2a 2~,~ ( - A. 1 - .u. ),(.ti, - a 2.4~ ( - A, 1 -/.~)(/2) ~,

- a . 6 ( a - ~,b - t~)s - a, cb(a - ,Lb - l~)12 =

-- 02x~(a - - i~, b - / , L ) ( ~ ) 2

-20~,~b(a - a , b - ~t),(12 - 0 2 , 6 ( a - a , b - / ~ ) 0 2 ) 2.

Notice (0x~b( - A, 1 - / ~ ) , 0,d~ ( - A, 1 - / x ) ) and (0x6 (a - A, b - / t ) , 0,~b (a - A, b - / z ) )

give the normal directions at the respective points ( - A, 1 - / x ) and (a - A, b - / ~ ) .
H e n c e , by the curvature hypothesis

(33) = I(1 - b)A + a ~ - a I-

Clearly we may suppose Ix - y I - l a I + I 1 - b I ,> n -~ and in proving (25) that

IIx - y II-Irx - r,l'> n-' or equivalently, by (26), that
(34) ha2 >> I x - y I.

If (34) does not hold, i.e. if

(35) ha2< Clx- y l,

we only have to prove

(36) diam comp(Vx A Vy) < C n - m ( I x - y I + l / n ) -m-

78 J. BOURGAIN

If (34) holds, then, by (26), (28), (29), (30), we need to prove

IIn

(37) f (I)~ [+ I/2 I)dt l_na

-IIn

and
IIn

(38) f ([ A"[+ 112[)dt <=Cl Xna-'Y l

-IIn

Assume first (35). We may suppose [a [<]1 - b[ since otherwise [x - y I< cn-' and
there is nothing to prove. It follows from (31) and (33) that

(39) [i [+[/2 I< C I ( 1 - b)A +atz-al-'

hence

[1- b l lA~( l <=2la l [~( [+ C.

Integrating, for It I_-
< 1/n, we get
!1 - b l a ( O 2 <-_ C(1/n + / a J l~(t)])
thus

IA (t)l <=Cn-"2l l - b [-''z

or

I a ( t ) l _-< c l , , I l l - b I-'---- On-"21 x - y I-''2

by (35). It is geometrically clear that I/~ (t)[ ~ I a (t)[. Hence (36) is proved, under the
assumption

n [11x - y II- [rx - r, [l < C.

Assume next (34).
Then the determinant value
(40) la - (1 - b)A - a/~l-la I

as long as
(41) [A [ < ~ [ a [ [ 1 - b [ - ' .
By (39), it follows that IA[<Clal -j, thus [A[<C/na for Itl<l/n, on the
component of V, r Vy containing 0.
It follows from (34) that n - ' [ a [-' "~ Ja [ [ 1 - b [-'. By previous reasoning (40)
holds on the 0 containing component of V, tq Vr
 CONVEX CURVES 79

It follows from (31) and (40) that

I;( I + 1/2 I-]a ]-t ~ (37).
It remains to prove (38).
From (32) and (40), we clearly get the following bound on 1,~"] + I/2 1,
C[-O,dp(a-A,b-l.t) a,4,(-A,I-.)]( ~
lal a~6(a-A,b-lz) - 0 ~ 6 ( - X , 1-/-~) n J
<=Clal-'ll~-nl+lx-yl(l~l+lnl)},
where ~,7/ are the right members in equations (32). Hence
1A'1+1/21-
-< Cla ]-'[x - y I(I,(I 2+ I/2 I:) <-- Eta [-3Ix- y1=),(38).
This completes the proof.
We will also use the following geometrical fact.
L e m m a 4. Let p < ~ and x, y, z E R z satisfy

(42) ~<llxtl, llyll< p,

(43) Ilzll> 1+ ~p, ~l,

(44) fllz-xll-l-Ilxlll~ and [llz-Yll-l-IlYlll~',

where r ,~ #. Then

(45) IIx - y II - III x II- II y III > c~ Ix - y I - c~-,

/
80 J. BOURGAIN

P r o o f of L e m m a 2. Assume Ilx II----tly II. Let w = - xlllx II and ~o,qJ the angles
indicated above. It follows from the discussion at the beginning of this section that

(46) IIx - y II + Ily - z II---- IIx - z II + c IIx - y Ilsin2 ~.

From (44)

(47) IIx - y II+ IIy II + 2~ -> IIx II+ c IIx - y Ilsin 2 ~.

Similarly

(48) II x - y II + II y - w II --< 1 + II x II + C IIx - y IIsin2 ~/

and subtraction from (47) yields

(49) 2~ + II y II + 1 - II y - w II --> c II x - y IIsin2 ~o - C II x - y IIsin24'.

Also

IIx II + IIz II ~ IIx - z II+ C(sin 2 ~o + sin 24,)11x II

and again from the hypothesis on z

(50) sin 2 r + sin 2 ~ _->cK - C "r .

P
From (49), (50)

(51) CIIx- yllsin2~ --- - 2 ~ ' + l l y - w l l - l - I l y l l + c x l l x - y l l - c ~ ' .

Assume IIY-w II > l. Then again

l+llyll-Ily-wll=<c
-dist(y, Rx) 2
Ilyll ----
clX-yrsin~,/,
p

and substituting in (51),

IIx - y II-IIx II + Ily II- CIx - y Isin2 ~ _-> cKlx - y I- c~-.

If 1 => Ily - w II---Ily + x/llx II II, clearly IIx - y II----IIx II and (45)is trivial.

5. T h e basic c o n s t r u c t i o n

Let now ~ C [0,1] 2 (actually by further splitting, we may assume that ~ has an
arbitrary small diameter; in particular, [Ix - y II< 89 + ry) for x , y E ~ ) . For x E f~,
define the set

~ --{y ~ ft;J Ilx - y l l - [r~ - r, I l < n-'+'}.

An easy inductive argument allows us to obtain points xl,x2 . . . . . x~ in fl and a

c o m p l e m e n t a r y set ~ C ~ such that
 CONVEX CURVES 81

ffL, l> n -'§

In.,\n.,l > n -'§

If},,\ U ['}x, > n -l+'

i<J

and f}~ = D,\Ui~.,f}~,, with the property that

(52) If}, fl f}x I < n -'+" for x E i.}.

Notice that in this construction J_-< n l-*. Define for j = I .....J

f}', = f}x, and [l~ = [l~j\(12., O"" O 12~j_,),

also
J
t~o = U ~;.
i=!

It will suffice from previous discussion to show that the t}; satisfy (20), thus

(53)
]If Q"dxll,<
n'j
and also

(54)
lit

We start with (54), writing by Lemma 3(ii)

(v,, v,}
(55)
l+nlllx-yll-lr.-r, idxdy.
fl I fltxllt

Defining
= {(x,y) ~,0,, x fl, I y e f}~,}

and using Lemma 3(i), the right member of (55) may be estimated by

(56)

From the LLtheory, the first term in (56) is dominated by

(57) tl -"
IIf
.~
tit
X
fl2<
2
= n -~ o g n 2.
82 J. BOURGAIN

It follows from (52) that the sections ~ ( x ) of ~ CI'I~ • l'l~ satisfy

I (x)l < n
It easily follows from a r e a r r a n g e m e n t argument that given a E R 2, D CR 2,

(58) f l a - y I-'/=dy --< CI O 13''.

D
Hence, the second term in (56) is d o m i n a t e d by

Cn -3/2 I ~'~1I ~ I x -- Y I-l'2dy ~ CFl-3/2n-3/4+(3/4)'~ J~"~1I

D(x)

(59) = Cn-~"~+sII'~. I.

Taking ~ < 88 the proof of the L2-estimate (54) follows from (57) and (59).

6. E n d o [ t h e p r o o f

It remains to prove (53) for the sets tq'j. H e r e we use L e m m a 4. By hypothesis,

IO~l > n -~§ and there is a point xi with O'i C f~x, Of course, we may assume x i = 0,
rx~ = 1. Denoting ~ ' = fYj, we then have the property

(60) IIIxll-lrx-lll<n -'§ for x E O ' .

Divide ~ ' in the respective regions [rx --- 1] and [rx > 1]. T h e computations for both
are analogous and we therefore only treat the second case, i.e.

(61) I1 + l l x l [ - r ~ l < , - ' " for x E I T .

(To handle the first case, L e m m a 4 has to be suitably restated.) Since clearly

we may also assume JJx JJ-~ n'/2 for x E ~'.

For l = 0,1,2 ..... log n, define

fl, = {x E IT;2'n -'`~ = IIx II----2 ' + ' " - ' %

It will suffice to prove for fixed l an inequality

,62, If
Nt

since (53) then simply follows from addition over I.

 CONVEX CURVES 83

Let thus 1 be fixed and denote p=2t+~n -m. Let K = n - L In order to use
L2-estimates, we need to restrict the functions I~', to the set [llx II > 1 + Kp]. Let thus
~0 be a [0, 1J-valued function on R" satisfying the conditions

~(z)=0 iftlzll=<l+,cp, ~(z)=l ifllzll>l+2Kp;

(63) IVr C/tcp

and define for x ~ Ill

~,~ = , ~ . ~ .

Thus

f'x(z)r 1 + Kp.

Let x, y E l~,. Clearly ( Q., f'y) = 0 unless there is a point z E V. tq Vy satisfying

IIz II=> 1 + ,~p. From (61). we then have

I1 + IIx II- IIx - z l l I< .-'+" + 2n =<2n-'+L

Similarly,

I1 + Ily II-Ily - z II1< 2 n -'+~

in which case, by Lemma 4 (r = 2n -'+') and (61),

IIIx yll-lrx r, l l > c K : l x Yl- -'"

thus

(64) 1 / . + l l l x - yll-Irx - r, ll ~ . - " ( I x - y I+ 1 / . ) .

If (Q~, Qy) ~ O, then Vx N Vy consists of 2 (possibly 1) components of diameter (by
Lemma 3,(i)) at most

Cn-,+2.(I x _ y I+ 1 / n ) ~'

on each of which, by (63), ~p oscillates at most

-Cn-'+2~(,x- y I+I )-'"

Kp n

Hence

and by Lemma 3 and (64)

--" f§
P
84 J. BOURGAIN

Estimate the left member of (62) by

~t !tit

Since [Ix [[> p/2 for x E [~t, r~ > 1 + 0/4 by (61) and it follows from construction of tp
that for each x E ['t'

II v~(1 - ~o)11,----l l e n g t h ( F x n [1 -< II~ II ~ 1 + 2xp])

and thus (cf. proof of Lemma 4)

II vx (1 - q,)ll, --< c
I1

Hence, the first term in (66) is dominated by Cn -"2-' lilt I. Again since ]]x [[_-<p for
x E l)t, the function fa, Vxdx is supported by a set of measure at most Cp, and thus
by the Cauchy-Schwartz inequality and (65)

tlf
~1t
r llt• I
I -2 112

<__cln'l ,,~
,, ( n l w l ) ''~
and hence (62), with e ' = 6 / 2 - 5 e , since 11)'[> n -~+~.
Thus the estimates (53) are obtained and the proof of Theorem 1 is completed.

7. Extension to vanishing curvature

The differentiation result (4) may be generalized to curves I" that are a bit more
general than those covered by T h e o r e m 1. Let F be the smooth boundary of a
compact convex body in the plane containing 0 as inner point. Assume that the
curvature of V only vanishes at finitely many points. We may then approximate the
arc length measure o" of U by a sum o'~ + o'2 + " " + O's,

!1O" - (O', +"" + O's)ll~,,~2~ <

where the O's (1 -< s -< S) are the arc length measures of disjoint curves Fs CF (for
which the curvature stays away from 0). Each of the Fs can be embedded in a F'~
satisfying the conditions of Theorem 1 and hence we have a maximal inequality

This reasoning yields

 CONVEX CURVES 85

Corollary 3, Let cr be the arc length measure o f a curve F as described above

and A , [ defined by (1), for f a bounded measurable function. Then

f = !i~ A,[ a.e.

It is easily seen that Corollary 3 fails when, for instance,

F = {max(lx.l,lx2l) = 1}

is boundary of the square.

REFERENCES

1. J. Bourgain, On the spherical maximal function in the plane, preprint IHES, June 1985.
2. J. Bourgain, High dimensional maximal functions associated to convex bodies, Am. J. Math., to
appear.
3. C. Herz, Fourier transforms of convex bodies, Ann. Math. 75 (1962), 81-92.
4. E. Stein and S. Wainger, Bull. Am. Math. Soc. g4 (1978), 1239-1295.

I.H.E.S.
BURES-SUR-YVE'ITE,FRANCE
AND
UNIVERSITYOF ILLINOIS
URBANA, ILLINOIS,U S A

(Received May 31, 1985 and in revised form November 11, 1985)