Generalized Hardy Operators and Normalizing Measures
TIELING CHENandGORD SINNAMON
Departmentof Mathematics, University ofWesternOntario,London, Ontario, N6A5B7, Canada
(Received15February, 2001;in finalform May, 2001)
Necessaryandsufficient conditions on theweightvand the measureafor theoperator
b(s)
k(s, y)f(y)dy Kf(s)
.
tobebounded fromLo[0 o) toLq(s) aregiven. Here a(s)andb(s) are similarly ordered functions and k(s,y) satisfies a modified GHO condition. Nearly block diagonal decompositions of positive operators are introduced as is the concept ofa normalizing measure.Anapplicationismade to estimates fortheremainder in aTaylor approximation.
Keywords: Hardy inequality; Weight; Normalizingmeasure
Classification: 1999 MathematicsSubjectClassification.Primary 26D15;Secondary 46E30, 42B25.
1 INTRODUCTION:
MONOTONICITY
GeneralizedHardyOperatorsare instances of integral operators having non-negative kernels:
Tf (s)
k(s, y)f (
y)dy.*Correspondingauthor. E-mail: sinnamon@uwo.ca
ISSN 1025-5834 print; ISSN 1029-242X. (C)2002Taylor&FrancisLtd DOI:
10.1080/1025583021000022522
Sincetheearly 1970sthere has beencontinualprogessonthe following question:
BetweenwhichweightedLebesguespacesisTabounded operator
(1.1)
Underlying thesuccessesof thelast 25y has been the exploitation ofthe monotonicity of the kernel k. The kernelinHardy’s integral operatoris k(s, y),Vo,s)(y)
which isnon-decreasinginsandnon-increasinginy.The generalized Stieljtes kernel k(s, y)=
(s + y)-
and the Riemann-Liouville kernel k(s,y)=
A’(0,s)(s-y)’,
2 > 0, are also monotone in eachvariable. See [9] for references. The above question has been an- swered forGeneralizedHardy Operators, those whose kernelk(s,
y) is supported in{(s,
y):0 <y<s} and satisfiesthe GHO condition:D
-
k(s, y)<k(s,t) +
k(t, y) <Dk(s, y) for y< < s.HereD is some fixedpositiveconstant. This condition, imposed in and
[7]
and later in [2, 3, 5, 8, 12] was sometimes accompanied by (superfluous) monotonicityconditions.However, it islargelyamonoto- nicity condition itselfas we willsee in Lemma2.2below.Recently, Question(1.1)hasbeenansweredfor someoperators whose kernelsare notmonotone. This is aimportant step, especially sincethe necessary andsufficient conditionsgiven haveretained the simple char- acterof those givenforpreviously studiedoperators.The newoperators includevariable limits onthe defining integral, essentially restricting the support of the kerneltothe regionbetween two curves.In [4], Question (1.1) wasresolved for theoperator
with a andb smooth functionson [0,cxz) which increase from 0 to o withs. The paper [3] looks atthemoregeneral operator
k(s,y)f(y)dy Ki(s)
withaandb non-decreasing butnotnecessarilysmoothand k satisfying amodified GHO condition. The boundedness ofK is established be- tween certainBanachfunctionspaces includingtheweighted Lebesgue spacesK"
Lo[0, cx)
-+Lqu[0, c)
for p <q butnot for q <p. The case q <pwas thedifficult case in[4]and necessitatedtheintroduction there of the concept ofanormalizing function.Inthispaperweanswer(1.1)fortheoperatorKin the case q <p.We also drop the monotonicity assumptions on a and b andas a resultwe are ableto take thevariable s offthe half line and allow itto be in a general measure space. We explore the normalizing function concept further, placingit in themoregeneral andmorenatural contextofnor- malizing measures. We examine the GHO condition in some depth, showingits connection withmonotonicity assumptionsandformulating itforuse whens is inageneral measure space.
An orderly presentation of this investigation requires that we begin with our look at the GHO condition and prove some needed results overgeneral measurespaces. This is donein Section2. Section 3 con- tains technicalresultsonnearly block diagonal decompositionofopera- tors with positive kernels. These results are quite generally applicable andmaybeof independentinterest. InSection4wedefinenormalizing measuresanduse ablock diagonal decompositiontoproveourmain re- sultgiving necessary and sufficient conditions forK to be bounded from
[0, cxz)
toLq(s)
foran arbitrarymeasurespace(S,a). The exis- tence of normalizingmeasuresforalarge class ofpairs (a,b)
is estab- lished in Section 5 where wealsoseetheinterestingformtaken bywhat remains ofour monotonicity assumptions. The final section is abrief presentation of the applicationoftheseresults toapproximationbyTay-
lor polynomials. The integral formof theTaylorremainder isreadilyre- cognizedasone ofthe operators we have been studying.Thenotationofthe paperisstandard. The harmonic conjugateof the Lebesgue index p is denoted
p’
so that 1/p+
1/p’-- 1. Weight func-tions are non-negative and allowed to take the value oe. As usual, 0.o 0.The supremum oftheemptyset istaken to be zero. Integrals with limits are assumedto includethe endpoints whenpossible sothat
[a,b]
but
[a, cx)
The expression "Aiscomparableto
B,"
writtenA B,meansthat there are positiveconstantsC
andC2
such thatCA
<_B <_C2A.
IfX0
CXthen countingmeasure on
X0
isthe measure definedonthea-algebraof all subsetsofXwhose value onE isjust(ENX0), the number ofele- ments inEf3X0.
2
THE CASE a(s)=
0The operators we consider in this section take the function f(y),y
[0,
oe) to thefunctionKf(s),
s S, withthe formulat(s)
Xf(s)
k(s,ylf(y)dy.dO
Here (S, a) is anarbitrarymeasure space, b S--> [0,o) is a-measur- able, and k Sx [0, cx) -->
[0,
cxz) satisfiestheGHOconditiongivenin Definition2.1below. Themainresult ofthissection, Theorem2.6, gives simple integralconditions onk, b,oandawhicharenecessaryandsuf- ficient for the operator K to be bounded as a map fromL,,[0, cxz)
toLq(s).
DEFINITION 2.1 Suppose that
(S,a)
is a measure space and b’ S--+[0, oo)
isa-measurable.Akernel ksatisfies
theGHOconditionon
{ (s,y)
0<_
y< .b(s) }
provided thereexists aD>_
such thatD-k(s,y)
<k(s,b(t))+
k(t,y) <Dk(s,y)fory<b(t)
<b(s) (2.1)
andD-k(s,
y)<k(s,w)
<Dk(s, y)fory< w <b(s),
wb(S). (2.2)
IfS [0, o) andb(s) sthen the case (2.2) doesnot ariseandwe seethat this definition agrees withthe usual GHO condition.
LEMMA2.2 Suppose
(S, a)
is ameasure space, b" S--+[0, c,)
is a-measurable and k
satisfies
theGHOconditionon{ (s, y),
0<_
y<_ b(s) }.
Then there exists a kernel l satisfying the GHO condition on
{ (x, z)"
0<
z< x}
such thatl(x, z)
is non-decreasing in x,l(x, z)
isnon-increasingin z, and
k(s,y) , l(b(s),y) for
0<
y< b(s).
Proof
Define l"{ (x,z)
0<_
z<_ x} -- [0, eel
byl(x,
z)
sup{k(t, y) z <y<b(t)
<x}.(2.3)
Itisclear thatl(x,z)
isnon-decreasinginx andnon-increasinginz.Itis also clear thatk(s, y)< l(b(s), y) whenever0 <y <b(s).
Let Dbe the constant in the GHO condition satisfied by k. If we show thatl(b(s), z)
<D2k(s,
z) whenever 0 < z <b(s)
we will have shown that k(s, y)l(b(s),
y). To this end, fixz > 0 ands 6S such thatz <b(s)
and suppose that y> 0 and 6S satisfyz <y <b(t)
<b(s).
First ob- serve that k(t,y)< Dk(s,y) by the second inequality in (2.1). If yq b(S)
we havek(s,
y) <Dk(s,z)
by the second inequality in (2.2) but ify b(S), say yb(tl),
then k(s, y) k(s,b(t))
<Dk(s,z) by the second inequality in (2.1). In either case we have k(t,y)<Dk(s,
y) <D2k(s, z)
and, taking the supremumover allyand we getl(b(s),
z)<D2k(s, z)
as required.To
completetheproofit remains toshowthat satisfies theGHOcon- dition on {(x, z)’0<z <x}. To do this it is enough to show thatD-1 l(x, z)
<l(x,w) +
l(w,z)
<2l(x,z)
for0< z <w< x. (2.4) Themonotonicityof1,already established, provesthe secondinequality in(2.4).
To prove the first we suppose that y and satisfy z <y <b(t)
<xand show thatk(t, y)<D(l(x, w)
+ l(w, z)) (2.5)
wheneverz < w<xby looking atfour cases.
Case 1 z <y <
b(t)
<w <x. The definition of l yieldsk(t,
y) <_l(w,
z)
so(2.5)
holds. (Recall thatD> 1.)Case 2 z<w<y<b(t)<x. The definition of shows that k(t, y) < l(x,
w)
so again (2.5)holds.Case 3 z <y <_ w <b(t) <x ando3
’ b(S). By
the firstinequality in(2.2), k(t, y) <Dk(t,
w)
and by the definition ofl,k(t, w) < l(x,w)
sowe havek(t, y) < Dl(x, w)and (2.5) follows.
Case 4 z <y< w <
b(t)
<x and w b(s) for somes 6S. The first inequality in(2.2),
with s and t interchanged, shows that k(t, y) <D(k(t, b(s)) +
k(s, y)). The definition of l, used twice, shows thatk(t, b(s))<
l(x, w) and k(s, y) <.l(w, z) so in this case too we have (2.5).Taking the supremum over all and y satisfying z <y <
b(t)
<x, (2.5) becomes l(x, z) <D(l(x,w)+
l(w,z))
which completes the proof of(2.4) and the lemma.Lemma 2.2permits usto move from the kernel k depending on the variable s 6S to a kemel defined in the familiar triangle
{(x,
y):0 <y <x}. We must also be able to move from the measure aonStoameasureon [0,o) and,inordertoapply Stepanov’sresults on Generalized Hardy Operators, from there to weight functions on [0,cx). Somewhatsurprisingly, the lattermoveproves tobemorepro- blematic than the former.LEMMA2.3 Suppose
(S, a)
isameasurespace andb"S-- [0, cxz)
isa-measurable. Then there exists a measure l
defined
on the Borelsubsets
of [0, )
and satisfyingF(x)dp(x)
Is F(b(s))da(s)
(2.6)for
every Borelmeasurablefunction F’[0, c)
--+[0,
cx).Proof
Since b is a-measurable,b-(E)
is a-measurable for every Borel setE C[0, cxz).
Define #byp(E) a(b-1
(E)).
(2.7)It is routine tocheck that p is a measure and that(2.6) holds.
THEOREM 2.4 Suppose
(S, a)
isameasurespace and b"Sis a-measurable. Let k be a kernel satisfying the GHO condition on
{ (s,y)"
0<_
y<_ b(s)}
anddefine
lby (2.3).Define
lu by(2.7). Ifq>
0then
j (Ii(S)
k(s,Y) f (y)dY) qdr(s) I(o,o) (Ji
l(x,Y)f (y)dY) qdl(x)
for
allf
>_ O.Proof
The work has beendone.By
Lemma2.2,k(s,y) l(b(s),y)
so wehaveIs (Ji(S)
k(s, y)f(y)dy) qdr(s) Is (Ii()
l(b(s),y)f(y)dy) qda(s)
withconstantsindependent
off.
NowletF(x) (
l(x, y)f(y)dy)qand notethatFisnon-decreasing and hence Borelmeasurable. Lemma2.3 providesJs (Ii(S) l(b(s), Y)f (y)dY) qdr(s) I[o,) (Ii
l(x,Y)f (y)dY) qdlt(x).
The point0 may be omitted from the rangeofintegrationbecause the integrand iszero there. This completes the proof.
Theorem2.4takesus fromthemeasure space
(S, a)
backtothehalf linebut themeasure#maynotbe aweightedLebesguemeasure. How- ever, the monotonicity of enables us to overcome this difficulty and approximateintegralswithrespecttod#
byintegralswithrespecttoab- solutelycontinuousmeasures.LEMMA2.5
flit
isa measure on[0, cx)
then thereexists asequenceUnof
non-negativefunctions
such thatF(x)un(x)dxincreases withn to F(x)dit(x)and
(2.8)
,0
lim
F(x) u(z)dz
Un(X)dxF(x)
n 0,o) x,o)
(2.9)
for
every > 0and everynon-negative, non-decreasing,left
continuousfunction
F.Proof
SetU(y) J’(y,) dit(x)
fory>
0 andnotethatUx(o,)
is non-increasing and fightcontinuous for each integern
>_
1. Set Un(X)n[U(x)X{o,n)(X) U(x + 1/n)X’{o,o)(x + 1/n)].
Ify<n-lthen
un(x)dx n
U(x)dx-
nU(x + 1/n)dx
n,yI
y+linU(x)dx.
Since Uisnon-increasing,thissequences ofaveragesisnon-decreasing and
U(y
+ In)
<_u.(x)dx
<_ U(y).The fight continuity ofU shows that
’u.(x)dx increaseswithnto
(2.10)
Suppose thatFisnon-negative, non-decreasing and left continuous.
Standard arguments 10, p. 262ff]show thatthere exists a measureq6on theBorel subsets of
[0, cxz)
such thatF(x) f[0,x)
ddp(y)forx> 0.Now(2.10)
and the Monotone Convergence Theorem show thatJ Un(X)dx ddp(y)
increases withntoI I(
i0,o) y,o) [0,cxz) y,x)
dU(x)d (y).
Interchange the order ofintegration andthis becomes
l(o,o ) I[O,x)
dq (y)u"(x)dxincreases withntiO,x)
ddp(y)d#(x) which establishes
(2.8).
Nowwerepeatthelast partof theabove argumentwith
Un(X)
replaced by(.Ix Un(Z)dz)lUn(X)
andd#(x)
replaced by(x,)d#(z))d#(x)
Theconclusion
(2.9)
will followonce we show thatI(I?bln(Z)dz)flUn(X)
dxincreases with n to something equivalent to
I(y,o) (J[x,) dl2(z)) /d#(x)"
Performing the integration, wehave
cx
bln(X)dx
(i ign(X)dx)
fl+which increases to
(f(y,cx)dla(X))
l+1 by (2.10). Itremains to show thatReplacing theinterval
Ix,
o) by (y,o)
in thefight hand integralshows that the left hand integraldominates it. Toprove the other direction, sup- pose that#(y,o) < c and choosey0 >y such thatdp(x)<2
]
IY,Yold#(x)
and.I
(y,oo)d#(x)
<2.I
Iy0,oo)It iseasy to seethat sucha Y0 must exist. Now
AlthoughsuchaY0maynot exist inthecasep(y,
o)
o,theconclu- sion remains valid. Weomit the details.Generally speaking, theresult of thelastlemma cannotbe extendedto include functionsFwhicharenotleftcontinuous.Thisleadsus to make the followingtechnical restriction onthefunction b and thekernel k. If 0 < z <xthen
sup{k(t, y) z <y<b(t) <x} sup{k(t, y) z <y <_b(t)<x}. (2.11) Thiswill ensurethatthekernell(x,z), definedby (2.3),is leftcontinu- ous in x.
THEOREM 2.6 Let p,q
(1,cx)
and o be a non-negative weightfunction
on(0, cxz).
Suppose that(S, tr)
is a measure space, b S--,[0, cxz)
is a-measurable, ksatisfies
the GHO condition on{(s,y)
0_<
y_< b(s)}
and(2.11) holds whenever0<
z<
x.Let Cbe the leastconstant,.finite
or infinite,for
which the inequalityk(s, y)f(y)dy
dtr(s)
<C’v(y)dy)
lip
holds
for
all non-negativefunctions f. If
1 <p < q< c thenC
max(A0, A)
andif
<q <p < c then C, max(B0, B1)
whereAo sup(j
y>O [s:b(s)>y}k(s,
y)qda(s) v(z)l
-t dz[t:b(t)>b(s)}
dr(t))
1/q(If
(s)k(s,y)P’v(y)l-p’dy)
[s:b(s)>y}
r/q r/q’
)
1/rk(s, Y)qdty(s)) (I v(z)l-p’dz) v(y)l-p’dy
[t:b(t)>_b(s)}
dr(t))r/p
(s)k(s, y)P’v(y)l-P’dy dr(s)
l/r
Herer is
defined
by1/r
1/q- 1/p.Proof
Define1 and pby(2.3)
and(2.7)respectively.LetC’
betheleast constant, finite orinfinite, such that(l(o,oo)(Jil(x’Y)f(y)dy)qd#(X))
1/q<_Ct(ioo
0f(yy’v(y
holds for all non-negativef.
By
Theorem2.4, CC’.
NowletUnbe the sequence from Lemma2.5 and defineC(n)
tobethe least constant,fi- nite orinfinite, suchthat(I(Iil(x,Y)f(y)dY)qun(x)dx) 1/q< C(n)(If(Y)Pv(y)dy )
1/pholds for all non-negative
f.
Theassumption(2.11)showsthatl(x, y)is leftcontinuous inthevariable xanditfollows that([
l(x, y)f(y)dy)q isnon-negative, non-decreasing, and leftcontinuousfor each non-negative
f. By
Lemma 2.5increasesto
I(o,o) (Ji
l(x’Y)f (y)dy) qdp(x)
as n-- cx so
C(n)
is an increasing sequence and supnC(n)--
limn C(n) C’.
Now we apply the results of [12] to get
C(n),,max(Ao(n), Al(n))
when <p < q < c andC(n) max(B0(n), B (n))
when<q<p<cxwhere
Ao(n) sup l(x,
y)qu,,(x)dx v(z)l-p’dz
y>0
A
l(n)
supun(z)dz
l(x,yy
v(y) dyx>0
r/q
’
r/q’ -P’Bo(n)
l(x,y)qun(x)dx
/)(Z)1-t dz v(y) dyB (n)
u,,(z)dz l(x,yy" v(y)-P’dy
u,,(x)dx We show sup,Ao(n) Ao,
sup,A1
(n) A,suPn Bo(n) Bo,
and supn B(n) B
to completetheproof.For each fixedy,
,(y,)(x)l(x,y)
q is non-negative, non-decreasing, and left continuous so, byLemma2.5,l(x,
y)qttn(x)dx X(y,)(x)l(x, y)qun(x)dx
increases with n to
X(y, oo)(x)l(x, y)qd(x) j X(y, oo)(x)l(x, y)qdlt(x).
Lemma2.3 shows thatthe lastexpression isequal to
2((y,oo)(b(s))l(b(s), y)qdtr(s) Ils:b(s)>y}
l(b(s),y)qdr(s)
which is equivalent, by Lemma2.2,to
{s:b(s)>y}
k(s, y)q
dr(s).
Thus,
SuPnAo(n ) Ao
and, by the MonotoneConvergence
Theorem, SUPnBo(n) Bo.
The proofthat SUPn
A(n) A
also relies on the leftcontinuity inxof
l(x,
y). As abovewe find thatfx Un(Z)dz
increases toX(x,)(z)dp(z) Jlt:b(t)>xl
dtr(t).Observe that since {t"
b(t)
>x] C {t"b(t)
>inf(b(S)N[x, oo))}
we haveda(t)<_ sup
j
t:b(t)>x} {s:b(s)>x} [t:b(t)>b(s)}
Now
(Jb )l/q(Ii dY)
I/psup
da(t)
l(x,y)"
v(y)-p’X>0 (t)>x
<sup sup dr(t)
l(b(s),y)P’v(y)
dyx>0b(s)>x X, b(t)>b(s)
)""(l: _., )"’
_< sup
d(O (b(s),yy’
v(y) dy(s)>O (t)>_b(s)
< sup lim
d(t) l(x,y)n v(y)-p’
b(s)>O xb(s)- (t)>x
_< sup dr(t)
l(x,y)P
v(y) -Pdyx>O (t)>x
(2.2) Becausethe firstandlastexpressionscoincide alltheinequalitiesabove areequalities and sinceLemma 2.2 showsthat the expression (2.12)is equivalentto
A
we have sup, A (n) A as required.Fortheproof ofsupn
B (n) Bl
weapply Lemma2.5 withfl
ripto seethatsup
B (n)
is equivalentto(o,oo)
[x,o)dp(z))
l(x,y)t" v(y)-P’dy
dp(x) which Lemma2.3, applied twice, shows to be just(L(J
[t:b(t)>_b(s)}l(b(s), yy"
v(y)-P’dy da(s)
l/rBy
Lemma 2.2 the lastexpression isequivalent toBl.When the kernel k
-=
the weightconditions simplify and theresult extendsto include thecase 0 < q < 1.COROLLARY 2.7.
Suppose
0<
q<
cx,<
p<
c,o is a non-nega-tive weight
function
on(0, c),(S, tr)
is a measure space, andb" S
-- [0, cx)
is a-measurable. Let C be theleast constant,finite
or infinite,for
which the inequalityIS
fb(s)"
q 1/qf (y)Pv(y)dy)
1/pholds
for
allf
>O.If
<p<qO <q <p < cx then C Bwhere
then CA and
if
A sup
da(s) v(z)l-p’dz
y>0 {s:b(s)>y}
B
dtr(t) v(y)l-p’dy dtr(s)
{t:b(t)>_b(s)}
1/r
Here
1/r
1/q 1/p.Also,if
q > 1orO <q < 1 ando-P’
is locally integrable then{s:b(s)>y}
r/q r/q’
dy)
1/rda(s)) (Ji v(z) l-p’az)
v(y)I-p’Proof
The case<
p<
q<
cxfollows fromTheorem 2.6by taking k since in this caseAA0
andit is not difficult to seethatA <
A.Inthecase0
<
q<
p<
wedefineC(n)
asinTheorem 2.6.Westill have limn--,oC(n)
C. Using [11, Theorem 2.4] we havel/r
Un(X)dx
Inthesameway thatweshowedsup
B (n) B
in Theorem 2.6we see that the fight handsideconvergestoB.Thefinal assertionfollowsfrom the remarkonpage 93 of[11 ]. This completes the proof.COROLLARY2.8 weight
function
Suppose0
<
q<
c,<
p<
oe,o is anon-negative on(0, cx),(S,a)
is a measure space, anda" S---.
[0, x)
is a-measurable. Let C be the leastconstant,finite
or infinite,for
which theinequalityl/p
holds
for
allf
>O. If <p<q<oo O < q <p< oo then CB’
wherethen C
, A’
andif
y>O [s:a(s)<y}
[t:a(t)<a(s)}
rip oo riP’ 1/r
dtr(t)) (l
(s)v(y)I-p’dY) dr(s)
Here
1/r
1/q lip. Also,if
q > 1orO <q < and0I-p’ is locallyintegrable then
{s:a(s)<v}
r/q oo r/q’
)
I/rda(s)) (Ig, v(z)l-p’dz) v(y)I-p’dY
Proof
Make the change ofvariabley1/y
and apply Corollary2.7with
b(s) 1/a(s).
We omitthe details.3
DECOMPOSITION OF NEARLY BLOCK
DIAGONAL OPERATORSBlock diagonalmatrices are well understood. Thereare direct sum de- compositions of both thedomainandcodomainspacessothat theaction ofthe whole matrix isbroken downintothe actionof the blocksontheir individual summands. A similar process can be carried out for more general linear operators whose domain and codomain can be decom- posedinsuch afashion. Werestrict our attention to positive linearop-
erators, those that take non-negative functions to non-negative func- tions. This restriction allows us to consider operators which do not have a strictly block diagonal decomposition but which decompose into blocks whose natural domains
(and
codomains) may overlap to some extent. Ourdecompositiontheoremforthese nearlyblockdiago- nal operators is Theorem3.3.DEFINITION 3.1
If
Kisa linear operator taldngnon-negative v-mea- surablefunctions
tonon-negativea-measurablefunctions
wedefine
thenorm
of
Kto beIlKIlg
sup[. jiKf()e>(),S.(s):f
>_O.g>_O. Ilfll,
_<1. Iigl14-< 1.1"
Weidentifya
function
q9 on themeasurespace(X, )
with the multipli- cation operatorf---qgfso thatif
q9 X--+ [0,cx) then114ollt-+t
suplJxqg(x)f(x)g(x)d(x ) "f>0
g>0_Ilfllc
<1_Ilgllc<_ 1}.
DEFINITION 3.2 A non-negative, linear operator K is nearly block diagonal provided there exists a measure space
(X, ),
r-measurable subsetsSx
of(S, a),
v-measurablesubsetsYx of(Y, v),
anda positiveconstantMsuch that
(1/M)Kf(s)
<.It XSx(S)K(fXrx)(S)d(x)
<MKf(s),
s S,f> O;(3.1)
M-1
<I
Jd(x)
<_M,s S; and{x:sSx}
M-l <
1" d(x)
<M,y
Y.{x:yeYx
(3.2)
In this casewesay that
(, {(Sx, Yx)
Xex})
is a nearlyblock diagonal decomposition ofK.
Theassertion of(3.1) isthat the action of the operatorK canbeex-
pressedin terms oftheactionof the blocks and(3.2)controls theextent of the overlap ofthe decompositions ofthe spaces Y andS.
THEOREM 3.3
Suppose
that(X,)
is a measure space and(, { (Sx, Yx)
xEX})
isanearly block diagonal decompositionof
K.If Kx f XsxK J’X
Yr then(3.3)
If
iscountingmeasure on a subsetof
XthenIIKll(Yx)Z(sx)ll(x)(x) M+/P’+I/qIIKIIL(Y)Lq(s).
(3.4) HereMis the constantfrom Definition
3.2.Proof
Fix non-negative functionsf
and g withand
IlgllLz’(s)<-
1. SetF(x)- M-1/PlIfXr.IIL(rx)
M-l/q’
IlgXs.llf(s.).
NotethatIl fllL(Y <-
and
G(x)=
by (3.2). In a similar waywe see that
[[G(x)llL,(x
<_ 1.To establish
(3.3)
we useDefinition 3.1.Taking the supremumoverallchoices
off
andg we haveIIKIIL(y)-->Lq(s) MI-I/p+I/q’I]
which is
(3.3).
Suppose now that is counting measure on some subset ofX. In- equality (3.4) is trivial if
Ilgll(r)--,q(s)
is infinite so we assume that it is finite. It is clear from the definition ofKx
thatKxf(S)< Kf(s)
for all x6X, all s 6S and all non-negativef.
It follows thatIIKxIIL(Yx)Lq(sx)
<o forall x6X.Fix2 6
(0, 1).
Foreachx6X choose non-negativefunctionsfx
andgx such that
I[fx[[L(Y)
_< 1,[[gx[lL’(s,) -<
andAIIKxlIL(Yx)L(x) I
Kxfx(S)gx(s)da(s).(3.5)
Replacingfx
byfxX n
andgxbygxXSx
does not affect(3.5) and cannot increase the norms offx
and gx so we may assume henceforth thatfx fxXrx
andgxgxXSx.
Let F(x) and
G(x)
be non-negative functions on (X,) with[[F[[c(x)
_< and[[Gl[cq,(x) -<
and setjZ(y) M-/p’
f
Jx F(x)fx(y)d(x)
and
(S)
M-l/qJ’x
G(x)gx(S)d(x).
Since is counting measure, it is clear that
F(x)fx(y)<
M/P’.(y)
and G(x)gx(S)<M/q(s)
forWeall yuse6dualityY,s 6S andto estimate the norm ofxin the supportof9
.
v inLv(Y
P).
Suppose H is non-negativeandIIHII ,, ’
(r) < ThenIIHX.,.II ,,
,’(rx)I1’
(x)(L j’r H(y)P’,r.,.(y)dv(y)d(x))
/’’g(yy"
Xr.,.(y)d{(x)dv(y) <M/p’Y
sowe have
Ir
"T’(Y)H(y)dv(y)M-’
/P’IrIx
F(x)’(y)d(x)H(y)dv(y) M- /P’Ix F(x) Irfx(y)H(y)dv(y)d(x)
F(X)Jrfx(y)H(y)Xr.,.(y)dv(y)d(x)
<
L F(x)lllL(rx)llHrll ’ r,)d(x)
M-/P’ x F(x)IIHXr.
5
M-/P’IIF[[L(X)[I
[IHXyx<M-1/p’M/P’=1.
Taking the supremumover thenctions Hwe have
IIll(r
1.A similarargument shows that
IIll
L,r (S)qt <NOW
2
Ix IIKllLvyr)__>L(s)F(x)G(x)d(x)
<
J’x Is
Kxfx(s)gx(s)da(s)F(x)G(x)d(x)1; Jx
Kx(F(x)fx)(s)G(x)gx(s)d(x)da(s)<M’/P’+l/q
ls lxKxf’(s)G(s)d(x)da(s)
<_MI+I/p’+I/q
IsK’(s)G(s)da(s)
Taking the supremum over all non-negative
F(x)
andG(x)
withIIFIlx) IIKxlI(Yxq<s
< andIlGIIc,(x IIL<X)Lqc(x)
_< and letting2--
1- wehaveThis completes the proof.
To use the above theorem we must understand the norm
IIL.x)-x).
This is not difficult.Aproofof thefollowing simple pro- positionmay be foundin [6].PROPOSITION 3.4
If (X, )
is a measure space,<_
q<
p<_
xz and1/r- a/q- lip
thenfor
anynon-negatived?. If
iscountingmeasure on asubsetof
Xand<p<q<_cxzthen
4
CONDITIONS
FOR BOUNDEDNESS OFKTo give necessary andsufficient conditions fortheboundedness ofthe operator
b(s)
k(s,y)f(y)dy
Kf (s)
Ja(s)
(4.1) from
L/v[0, oo)
toLq(s)
weapplythe decomposition theorem of the pre- vious section. The action of the operator on the resulting blocks is handled using the results of Section 2. The necessary and sufficient conditions for boundedness on the blocks combine to give integral conditions similar in form to those ofTheorem 2.6.Thevalues
off
offYtOss[a(s), b(s)]
haveno effecton the values ofKf
so it is natural to consider the functionsf
to be defined on Y.It is easy to see that
K’L[O, oo)--+ Lq(s)
if and only ifK" L’,(Y)
--+Lq(s).
Webeginby introducingtheconceptof anormalizingmeasurewhich providesus with anearlyblockdiagonal decomposition of the operator K.
DEFINITION4.1 Let
(S, a)
bea measurespaceand suppose thataand b arenon-negative a-measurablefunctions
on Ssuch thata(s) < b(s) for
alls. A measure on[0, c]
is calleda normalizing measurefor
(a, b)
provided thereexistpositiveconstants cl andc2 such thatb(s)
Cl <
d(x)
<c2(4.2)
aa(s)
for
alls S.If
inaddition, iscountingmeasure on asubsetof[O,
then iscalleda discretenormalizingmeasure.Nextweshowthatanormalizingmeasureisall thatisrequiredforthe operatorK of(4.1) tobenearlyblock diagonal.
LEMMA
4.2 Let(S, a)
be a measure space and suppose that a and b arenon-negative a-measurablefunctions
on Ssuch thata(s) <_ b(s)
Jbr
all s.If
is a normalizing measurefor (a,b)
then(, {(Sx, Yx)
"X EX})
is a nearly block diagonal decompositionof
Kwhere X Y
Uss[a(s), b(s)], Sx {s
Sa(s) <
x< b(s) },
andrx {y [0, Sy Sx #
Proof
Letcl and 2 bepositive constants for which satisfies (4.2)andsetM max
(1/c, 2c2).
Sincej
{x:s6Sx}a (x) fb(s)
da(s)a (x)
for eachs 6S, the firstinequalityin
(3.2)
follows from(4.2).Notethat
Y (_Js[a(s), b(s)]
which isaunionofintervalscontain-ingxso
Yx
is an interval. The symmetry inthe definition ofYx
showsthat {x’y Y}
Yy
and sinceYy
is an interval there exist sequences Sn ands’
n of pointsinSy
suchthat(Yy)
lim[a(Sn), b(stn)].
Sinceyis inboth
[a(Sn), b(s,)]
and[a(S’n), b(s’n)
thelastexpressionisnogreater than
lim
[a(sn), b(sn)] + [a(s’,), b(s’,)]
_<2c
< M.n---- o
For y6
X,
there exists some s witha(s)<
y<b(s)
so we have[a(s), b(s)] c Yy
andhence1/M
<Cl <[a(s), b(s)]
<(Yy).
Wehave shownthat
1/M
<(Yy)
<Mwhich establishes the second in- equality in(3.2).
It remains to show that (3.1) holds. An interchange of the order of integration yields
Ix Xsx(s)K(fXrx)(s)a(x) Ix XSx(S) f
b,s)k(s, y)f(y)2(r,(y)dy d(x)
[()
k(s, y)f(y)f 2(s (s)Xv
(y)d(x)dy.aa(s) dx
The innerintegralin the last expressionisjust
[a(s), b(s)]
so(4.2) im- plies(3.1).
This completes theproof.The main results of the paper are presented in Theorems 4.3 and 4.4. It is convenient to split up the cases <q<p< and
<p<q<c.
THEOREM4.3 Let
<
q<
p<
x,v bea non-negativeweight,(S, a)
be a measurespace, a and b be a-measurable
functions
on S, and kbe a non-negative kernel satis.ing the GHO condition on
{(s,y)
0<
y< b(s)}
andalso(2.11).Suppose
that isanormalizing measurefor (a, b).
Let C be the leastconstant,finite
or infinite, suchthat
(is([l(s)
k(s,y)f(y)dy)q dtr(s) )l/q
X.aa(s)
holds
for
allf
>O. Then C is bounded above by a multipleof max(/3,
/32,/33,/34)
wherex(i
ax_b(s))r/q(! I )r/q’
]3 IX Ix
a(s,<xA(S, y)qdo(s) v(z)l-p’dz 1)(y)l-p’dyd(x)
y<b(s)
b(s) (s)
-P’
r/F
] l(s)
b(s)<_b(t)(t)<x da(t) k(s,y)P’v(y)
dyd(x)da(s)
b(s) x r/p’
dtr(t)
-(x, y)P’v(y)-P’dy d(x)dtr(s).
Here k(x, y) sup{k(t, y)
b(t)
x}.If
isa discretenormalizingmeasure then C isalsobounded below bya multipleof
max(Bl, /32, /33,B4).
Proof
LetX,Y, Sx
andYx
be as in Lemma 4.2. Then(, {(Sx, Yx)
"xEX})
isanearlyblock diagonal decompositionofK. Itfollows from Theorem 3.3 and Proposition3.4 thatMl+l/p+1/
IIKxllx)--,tq(sx)
whereM depends onlyontheconstantsc andC2inthedefinitionofthe normalizingIf isa discretemeasurenormalizing measure,
.
the inequality maybe essen- tially reversedto giveIIKx ll(rx)--,tZ(sx) llt(x) M
+ /P’+ /qllKllz(r)--->t(s).
Since C
IIKIIL[O,o)LgS)= IIKIIL(Y)Lg(S),
this reduces the pro-blem to looking at thenorms of
K
for each x inX. Towork withK
we decompose it into three operators and apply the results ofSection 2. Fix x 6Xand take jr>_0 to be supported in
Y.
ThenXxf (S) XSx(S)
a(s)
k(s, y)f(y)dy
Xs, (s) k(s,
y)f(y)dy+ Xs, (s)
k(s, y)f(y)dy.(s)
(4.3)
Note that, according to the definition ofSx,
a(s)
<x <b(s)
wheneverXsx (s)
O.Wenow usetheGHOconditiononktofurther decompose the first summand. Ifxq[b(S)
thenk(x,
y) 0 and if x b(S), sayx=b(t),
then it follows from the condition(2.1)
on k that k(t,y) < k(x,y)< Dk(t,y). In either case we have (using (2.1) or(2.2)
asappropriate)D-1
k(s, y)<k(s,x) + -(x,
y) <Dek(s,
y)whenevery<_ x< b(s).Applyingthis estimate tothe kernel kinthefirst summand of(4.3) shows that
Kxf(s)
is bounded above and below by multiples of’s,. (s)k(s, x) f
(y)dy+ ,Vs, (s)
k(x,y)f(y)dy(s) (s)
+ R’s.,.(s)
k(s,y)f(y)dy=_Kl!f(s) + KZ)f(s) + Kx3f(s).
Since theoperators
Kx0), Kx2),
andKxt3),
areall non-negativeand hence
IIKll:(v.)_(s.,.) llv.e(x
(2)
To complete the proofwe showthat
IIK(x2)llt(v.)-t(s.)llt,ex)
B4, and[[K(xa) IIL(r.,.)__.Lg(S.,.) llL(X)
max(B2,/33).The norm
IlK(x)llL(y.,.)L(S.,.
is the least constant for which the in-equality
(L. ( li(s)f (y)dy) qk(s,x)qdtT(s)) l/q<_ C( Pv(y)dy)
holds forall
f
> 0. It is straightforwardto see that it is also the least constantforwhich(j(j(s)f(y)dY)qdo(xl)(s))
< C y)(y
holdsfor all
f
_>0 whereda(xl)(s) Xs.(s)k(s,
x)qda(s),V(xl)(y)
v(y) for y [0,x] Yx
andV(xl(y)
cxotherwise.By
Corollary2.8wehaveIlg(xl)ll(yx)--,Lg(Sx) ’Yx(Y) ’Sx(S)k(s,x)qd6(s)
(s)<y
(I; v(z) l-P’dz) r/q’v(y) l-P’dy.
Fromthis itreadily follows that
The norm
IlK(xZ)llL(rr)L(Sx)
isthe leastconstant for whichthe inequal- ity(ISx (Ji(s)-(x’ Y)f (y)dY) qda(s)) /q<- c ( Ir. f (y)p v(y)dy)
1/pholdsforall
f
>O.Making the substitutiong(y) k(x, y)f(y), we see thatit isalso theleast constant for which
(j(Ia()g(y)dY)qda(x:)(s)) 1/q< C(J g(yv(xz)(y)dy)
1/pholds for all g>0 where
da(x2)(s)=2(s(s)da(s), V(x2)(y)=
-(x,y)-Pv(y)
for y[0, x]
f3Yx
andv(xl)(y)=
cxz otherwise. Again we appealto Corollary 2.8. WegetII/q
t,,(.,.)tZcs.)Xs,.(t)da(t)
(t)<a(s)
x k(x,
y)P
v(y)-p’Xs.,. (s)da(s)
(s)
and so, with an interchange inthe order of integration,
The norm
Ilg(x3)llL(.OLq(s.,.)
is the least constant for which the in- equality(Is.,. (Ji’(S)
k(s,Y)f (y)dY) qda(s)) /q<
C( .I’y.,.f (Y)P
v(y)dy)
1/pholds for all
f
> 0. It isalso the leastconstant forwhich(JS(II
(S)k(s, y)f(y)dy)
qda(x3)(s) )l/q-c(Iof(Y)PV(x3)(y)dy)
< l/pholds for all
f
>0 whereda(x3)(s) Xs,.(s)da(s ), V(x3)(y)
v(y) for y[x,
co)f3Y
andV(x3(y)
cx otherwise. This time we applyTheo- rem2.6 to seethatIlgx(3)ll:(n)_,&q(s.,.
iscomparabletothemaximumofJ ’’Y. (Y) (Iy<b(s)
k(s,Y)q "s, (s)da(s))
"/q(Ji’ v(z)
l-t"dz)
"/q’v(y)l-p’dy
and
r/q
XSx (s)d (s).
Fromtheseweconclude that
Ilgx(3) IIc(r)q(s) II(x)
max(/2,B3)
to complete the proof.
THEOREM4.4 Let 1
<
p<
q<
cxz,v beanon-negativeweight,(S, a)
be a measure space, a and b be a-measurable
functions
on S witha(s) < b(s),
and k be a non-negative kernel satisfying the GHO con- dition on{(s,y):
0<
y< b(s)}
and also (2.11). Suppose that(a,b)
admits a discrete normalizing measure. Let C be the least constant,
finite
or infinite, such thatk(s,
y)f(y)dy) da(s))
(L(fdii:
q 1/q 1/pholds
for
allf
> O. Then Cmax(,A, ,A2)
where(_[a ) l/qtl
.A1
sup (s)<_xk(s,y)qda(s) v(z) -P’dz
[(x,y):x<y}
\"y<_b(s)
.42
sup{(x,s):x<b(s)}
1/q
(l
(s)k(s,z)P’v(z)-P’dz)
Proof
Supposethat countingmeasureonX0
isa discretenormalizing measure for(a, b).
Thenfor anychoice ofx
andx2, countingmeasureon
Xo
U{Xl,X2}
is also a discrete normalizing measure for(a,b).
ChooseXl andx2 suchthat
[y:x <y}
\y<b(s)
I/q
V -p’
sup
{s:x2<b(s)}
1/q
k(s,
z)P’v(z) -P’dz
\dX2
and
Let be countingmeasure on
Xo
U {x, X2}. Wedecomposetheoperator Kjust as inTheorem 4.3 and applythe resultsofSection 3 togetAswehaveseenabove, thenorm
IlKx)llL(r)_L(Sx)
istheleastconstantfor whichthe inequality
f f(yY’4’)(y)dy
l/p
holds forall
f>
0wheredtrl)(s) ,.G(s)k(s,
x)qdo(s),V(x)(y)
v(y)fory
[O,x] Yx
andv(x)(y)
otherwise. Corollary2.8 shows that...()
II.G IIL’Y.,.)-.L(S.,.)IIL(X)
iscomparable to
(Ja )’/q
sup sup k(s,
x)qdty(s)
x_X y>O (s)<y<x<_b(s) (x)>O
(4.4)
The norm
IlKx(2)ll(g)_+q(s
isthe least constant for whichthe inequal- ity(JS(Ja(s)g(y)dY)qdtT(x2)(S))
1/q l/pholds for all g>0 where
dtr(x2)(s)=XSx(S)da(s), V(x2)(y) (x, y)-Pv(y)
fory[0, x]
t-IYx
andV(x)(y)
c otherwise. Corollary 2.8 shows thatiscomparable to
(la
sup sup
dtr(s)
x" y>0 (s)<y<x<b(s) (x)>0
Sincek(x, z) 0whenx
b(S)
andk(x, z) k(t,z)
when xb(t)
the last expression iscomparable to(Ia )l/q([b(t)
sup sup
dtr(s) k(t,z)P’v(z)
dztS y>O (s)<y<b(t)<_b(s) \y
(b(t))>0
(4.5)
(3) isthe leastconstantforwhichthe inequal- Thenorm
IlK’
xity
(IS(If (S’k(s,y)f(y)dy) q) do’(x3’(s)
1/q_<C(J’
0f(Y)Pl)(x3’(y)dy)
1/pholds for all
f
> 0 wheredtr(x3)(s) ?(s,(s)dtr(s), V(x3)(y)
v(y) fory
. [x, o) Yx
andV(x3)(y)
oe otherwise.By
Theorem2.6 thenormis comparableto the maximumof sup
sup([
xX y>O\J
(x>O s:a( <_x<y<b(s)
k(s,
y)qdo(s))
l/q(4.6)
and
(I )’/q(I
(s))
sup sup
dtr(t) k(s,y)P’v(y)l-p’dy
xeX seS {t:a(t)<x<b(s)<b(t)]
(x)>O
(4.7) Themaximumof the expressions (4.4) and (4.6) is comparable to
sup
{(x,y):x<y}
(x)>0
( k(s, y)qda(s) )l/q(Jir t
lip’v
v(z) -p’dz
"y<b(s)
which is comparableto
.A
because(x)
> 0.Inasimilarwaywe seethat themaximum of(4.5)and (4.7) is com- parable to
M2.
This completes the proof.5
NORMALIZING MEASURES
The resultsof the previoussectiondependontheexistence of a discrete normalizingmeasureforthefunctionsaandb. Hereweprovethatsuch a measureexistswheneveraand baresinailarly orderedin thefollowing sense.
DEFINITION 5.1 Let2"-
{ [c, d]"
0_<
c<
d_< cx}
anddefine
apar-tialorderon by
[c, d] -< [-, d]
providedc<_ -andd <_
d. Wesay that
non-negative
functions
aand b onSare similarly ordered provided the set{[a(s), b(s)]
sS}
is a totallyorderedsubsetof Z.
Toconstruct a discretenormalizingmeasure
,
we need thesetX0
ofatoms of
.
This set is constructed inthenexttheorem.THEOREM5.2
If T
isa totallyorderedsubsetof Z
then thereexists asubset
Xo of [0, cxz]
such that< (Xo
q[c, d]) <
3for
all[c, d]
ET.
Proof
A straightforward application ofZom’s Lemma showsthat we may assume withoutloss of generalitythatT
is a maximal totally or- dered subset of 2. Thatis, wemayassumethat theonlytotally ordered subset ofZ
which contains7"isT
itself.Itfollowsfromthisassumption thatT [0, ].
If
x6[0, c]
define Lx=inf{’x6[,d]67-} and Mx=sup{d’x6[, d] 7"}. Clearly, Lx< x<Mx. Ifx<y and x6[,
d]
67- theneithery6[, d]
ory> d. Inthe formercaseMy
> d bydefinition and in the latter we haveMy
>_ y > d. Taking the supremum over all such dproves thefirsthalf of:Ifx<ythenMx<_
My
andLx<Ly. (5.1)
The other halfisproved similarly.
Wenowestablish the first halfof:
Foreachx,
[Lx,
x]T
and[x,
Mx]T. (5.2)
Onceagain the second half may be proved similarly.By
themaximality ofT,
ifwe show that{[Lx,. x]}
A7" is totally orderedthen[Lx, x]
7"will follow. To do this we fix [c,d] 7" and show that either
[c, d]
-<[Lx,x]
or[Lx,
x] -<[c, d].
If x < c then [Lx, x] -<[c, d].
If c <x_< d thenLx < cbydefinition so again [Lx,x]
-<[c,
d]. Inthere- maining case, when d <x,weseethatwhenever x6[, d]
67"wehave d <x <d so[, d]-<
[, d] becauseT
is totally ordered. Thus c <and, taking the infimum over all such
[,
d], we conclude that c <Lx soIt, d] - [Lx,
x]. We have shown that{[Lx, x]} T
istotally ordered and hence[Lx,x]T.
Foreachx6
[0, cx]
define the subsetEx
of[0,cx]
as follows.Ex= (k= [Lkx, L’-lx])
A(k= [Mk-lx’Mkx])"
Here the exponents represent repeated application ofthe operator and