Contents lists available atScienceDirect
Journal of Mathematical Analysis and Applications
www.elsevier.com/locate/jmaa
On discretizing integral norms of exponential sums
András Kroó1
AlfrédRényiInstituteofMathematicsandBudapestUniversityofTechnologyandEconomics,Budapest, Hungary
a r t i cl e i n f o a b s t r a c t
Articlehistory:
Received5January2021 Availableonline21October2021 Submittedby S.Tikhonov
Keywords:
Marcinkiewicz-Zygmund BernsteinandMarkovtype inequalitiesforgeneralexponential sums
DiscretizationofLpnorm Multivariateexponentialsums Exponentialsumswithnonnegative coefficients
InthispaperwestudyLp Marcinkiewicz-Zygmundtypeinequalities
c1
1≤j≤N
wj|g(xj)|p≤ gpLp(K)≤c2
1≤j≤N
wj|g(xj)|p
forgeneralexponentialsumsoftheformg(x)=
1≤j≤najeλj,x, x,λj∈Rd,aj∈ R.Oneof themainresultsofthepaperassertsthatwhen1≤p<∞,K= [a,b]
andtheexponentssatisfyrelationsλj+1−λj≥,1≤j≤n−1, max1≤j≤n|λj|≤ Λ then Marcinkiewicz-Zygmund type inequalities hold for certain point sets of cardinality
N≤cnln1p+1Λ .
Since the dependence of cardinality on the parameters,Λ appears only in the logarithmic term, this bound is “almost” degree and separation independent.
Moreover,thediscretemeshesxjandweightswj=xj+1−xjaregivenexplicitlyand theyareuniversalinthesensethattheyworkforanyexponentialsumasabove.This resultwillrelyonsomenewdegreeindependentBernstein-Markovtypeinequality for exponentialsums.Moreoverwewillextendour considerationstomultivariate exponentialsums.Inaddition,itwillbeshownthatmuchstrongerresultsholdfor exponentialsumswithnonnegativecoefficients.
©2021TheAuthor(s).PublishedbyElsevierInc.Thisisanopenaccessarticle undertheCCBYlicense(http://creativecommons.org/licenses/by/4.0/).
1. Introduction
ConsiderthespaceLp(K),1≤p≤ ∞endowedwith someprobabilitymeasure onthecompactset K⊂ Rd. Then givenasubspaceU ⊂Lp(K) theMarcinkiewicz-Zygmundtype problemfor 1≤p<∞consists
E-mailaddress:kroo.andras@renyi.mta.hu.
1 SupportedbytheNKFIH- OTKAGrantK128922.
https://doi.org/10.1016/j.jmaa.2021.125770
0022-247X/©2021TheAuthor(s). PublishedbyElsevierInc.ThisisanopenaccessarticleundertheCCBYlicense (http://creativecommons.org/licenses/by/4.0/).
infindingdiscretepointsetsYN={x1,...,xN}⊂K andcorrespondingpositiveweights wj >0,1≤j≤N suchthatforanyg∈U wehave
c1
1≤j≤N
wj|g(xj)|p≤ gpLp(K)≤c2
1≤j≤N
wj|g(xj)|p (1)
with someconstants c1,c2 >0 depending onlyon p,dand K. Incasewhen K = [a,b] ⊂R isan interval on thereal line anatural choiceof theweights associated with the pointsa < x1 < x2 < ... < xN < b is givenbywj=xj+1−xj,1≤j≤N−1.InthisrespecttheclassicalMarcinkiewicz-Zygmundinequalityfor trigonometricpolynomialsstatesthat(1) holdsforthespaceU=Tnoftrigonometricpolynomialsofdegree
≤nand uniformlydistributed pointsontheperiodwith wj= n1 andN = 2n+ 1.Clearly thecardinality ofthediscretepointsetN = 2n+ 1 isoptimalhere.Thisequivalencerelationturnedouttobeaneffective tool usedforthediscretizationoftheLpnormsoftrigonometricpolynomialswhichiswidelyappliedinthe study oftheconvergenceofFourierseries,LagrangeandHermiteinterpolation,positivequadratureformu- las,scattereddatainterpolation,seeforinstance[9] forasurveyontheunivariateMarcinkiewicz-Zygmund typeinequalities.AnimportantgeneralizationoftheclassicalMarcinkiewicz-Zygmundinequalityfortheso called doubling weights wasgivenbyMastroianniand Totik[10].Various extensionsof theMarcinkiewicz- Zygmund inequality for multivariate algebraic and trigonometric polynomials can be found in [1], [4]
and [2].
Inthispaper wewillconsiderproblem (1) forgeneralrealexponentialpolynomials oftheform
g(x) =
1≤j≤n
ajeλj,x, aj ∈R, x∈Rd (2)
with arbitrarygiven λj ∈Rd. Ingeneral,resultsrelated tothese exponential sumsdependonthe number n of terms in the sums and the size of λj-s. The “degree” of these exponential polynomials is given by max1≤j≤n|λj|. Here and throughout the paper |· | denotes the usual Euclidian norm of vectors in Rd. Naturally oneshould tryto aim for discrete meshes with possibly smallestcardinality. A particularly in- teresting problem consists inobtaining discrete meshesof cardinality depending onlyon the dimensionn of theexponential sums(2), thatis independent of their degree. For the trigonometric exponential sums with λj-s being of the form λj =inj,nj ∈Z this question is discussed in detailin the survey paper [2].
In particular in[12] Theorem1.1 itis shownthatfor generaltrigonometric exponential sumsonecanget meshes ofcardinality∼n.Intrigonometric casethebasicproperty whichisthemaintool ofinvestigation isthepairwiseorthogonalityofexponentials.Inthegeneralalgebraiccasetheexponentsin(2) donothave this crucial feature. So instead of orthogonalityour considerations will rely heavily on Bernstein-Markov type inequalities for the size of derivatives of exponential sums. We will verify in this paper some new degree independentBernstein-Markov typeinequalitiesforexponential sums.This will allowus toextend Marcinkiewicz-Zygmund typeinequalitiesto generalrealexponential sums.Inparticular, itwill beshown (see Theorem1below) thatthere existmeshesYN={x1,...,xN}⊂[a,b]⊂Rofcardinality
N ∼nln1p+1 Λ
(3)
so that (1) holds with weights wj := xj+1−xj for every 1≤ p <∞ and every exponential polynomial (2) withany λj ∈ R satisfying λj+1−λj ≥ >0,1≤j ≤n−1 andmax1≤j≤n|λj| ≤Λ. Animportant featureof thisresultis thefactthatthe cardinalityofthe discretemesh isofoptimal ordern, whiletheir degreeΛ andtheseparationparameteroftheexponentsessentiallyappearonlyinthelogarithmicterm.
So inthis sense ourbound is“almost”degree andexponent independent.The explicitlygivendiscrete sets usedinTheorem1areuniversalinthesensethattheyworkforallexponentialsumsasabove.Wewillalso includesimilar resultsfor multivariateexponentialsums.Finally,we willpresentsomenewdimensionand
degreeindependentBernstein-Markovtypeinequalitiesformultivariateexponential sumswithnonnegative coefficients.ThiswillleadtoaconsiderableimprovementofthecorrespondingMarcinkiewicz-Zygmundtype inequalitiesformultivariateexponentialsumswithnonnegativecoefficients.Itshouldbealsonotedthatall discretemeshesareconstructedinthispaperexplicitly.
2. Marcinkiewicz-Zygmundtype inequalitiesforgeneralexponentialsums
ThenexttheoremwhichisoneofthemainresultsofthispaperpresentsaMarcinkiewicz-Zygmundtype inequalityforgeneralreal univariateexponentialsumsbasedonpointsetsofcardinalityN ≤cnlnp1+1 Λ. Theorem 1.Let 1 ≤ p < ∞,[a,b] ⊂ R,0 < ≤ 1,n ∈ N,Λ > 1. Then there exist discrete points sets YN={x1< ...< xN}⊂(a,b)of cardinality
N ≤cpnln1p+1Λ
wherec>0isanabsolute constant,sothat forevery exponentialsum (2) witharbitrary λj ∈R satisfying λj+1−λj ≥
b−a,1≤j≤n−1, max
1≤j≤n|λj| ≤Λ wehave
1 2
1≤j≤N−1
(xj+1−xj)|g(xj)|p≤ gpLp([a,b])≤2
1≤j≤N−1
(xj+1−xj)|g(xj)|p. (4)
Remark1.Oneshouldnote the factthatthe discrete nodesprovided byTheorem 1are rather universal, they work for all exponential sums satisfying the separation condition λj+1−λj ≥ b−a,1 ≤ j ≤ n−1 and the upper bound max0≤j≤n|λj| ≤ Λ. In addition the dependence of cardinality on the parameters ,Λ appearsonly inthe logarithmic term, so the bound on cardinalityis “almost” degreeand separation independent. It shouldbe alsomentioned thattheconstants 12 and2in(4) canbe replaced by1−ξ and 1+ξ, respectively,with anarbitrarily small 0< ξ <1.Of course, onewouldhave topay aprice forthis inthe sense that this will make theconstant inthe upper bound forthe cardinality of YN dependent on ξ. The proof ofTheorem 1 given belowindicatesthatconstants of order −lnξξ canbe used for theupper boundof cardinality.Inarecentpaper[8] theprecise dependenceofthis constantonparameterξ >0 was givenintheclassicalunivariate trigonometriccase.
TheproofoftheaboveMarcinkiewicz-ZygmundinequalitywillbebasedonanewdegreeindependentLp Bernsteintypeinequalityforderivativesofexponentialsums(2).WewillalsousethefollowingLp,1≤p<
∞Bernsteintypeinequalityforderivativesofunivariateexponential sumsfn(x)=
1≤j≤ncjeλjxgivenin [5],Theorem3.4
fnLp[−1+δ,1−δ]≤ 2n−1
δ fnLp[−1,1], 0< δ <1. (5) ThiselegantresultprovidesexponentindependentupperboundsforLpnormsofderivativesofexponential sums inside the interval. The drawback of the above estimate is the appearance of the term 1δ in the upperestimate whichleadsto largerthan requireddiscretizationsets inLp Marcinkiewicz-Zygmund type inequalities. Our next lemma shows that introducing a weight 1−x2 into the Lp norms of derivatives of exponential sums allows to replace 1δ by a substantially smaller term ln2δ. This improvement will be subsequentlyusedinordertoverifynearoptimaldiscretizationmeshes.
Lemma 1.Let1 ≤p< ∞,0 < δ < 1,n ∈ N. Then forany distinct real numbers λ1,...,λn ∈ R and any exponential sum fn(x)=
1≤j≤ncjeλjx,∀cj ∈R wehave
(1−x2)fn(x)Lp[−1+δ,1−δ]≤9nln1p 2
δfnLp[−1,1]. (6)
Proof. Using(5) withany 0< t<1 andmultiplyingthep-thpowerofthisestimatebytp−1 yields
tp−1 1−t
−1+t
|fn(x)|pdx≤ (2n−1)p t
1
−1
|fn(x)|pdx, 0< t <1.
Integratingaboveinequalitywithrespect tot∈[δ2,1] wehave 1
δ 2
1−t
−1+t
tp−1|fn(x)|pdxdt≤(2n−1)p 1
δ 2
dt t
1
−1
|fn(x)|pdx= (2n−1)pln2 δ 1
−1
|fn(x)|pdx.
Furthermore applyingFubinitheoremfortheintegralontheleft handsideofaboveestimateimplies 1
δ 2
1−t
−1+t
tp−1|fn(x)|pdxdt=
1− δ2
−1+δ2 1−|x|
δ 2
tp−1|fn(x)|pdtdx= 1 p
1− δ2
−1+δ2
|fn(x)|p((1− |x|)p−(δ 2)p)dx
≥1 p
1−δ
−1+δ
|fn(x)|p((1− |x|)p−(δ
2)p)dx≥ 1 p
1−δ
−1+δ
|fn(x)|p((1− |x|)p−((1− |x|)/2)p)dx
≥ 1 p2p+1
1−δ
−1+δ
|fn(x)|p(1−x2)pdx.
Thus combiningthelast twoestimatesabovewearriveat
1−δ
−1+δ
|fn(x)|p(1−x2)pdx≤p22p+1npln2 δ 1
−1
|fn(x)|pdx.
Finally takingthep-throotaboveyieldstherequiredestimatewithaconstantc≤4e2e ≤9.
ByastandardlineartransformationLemma1canbeextendedtoanyinterval [a,b]⊂R.
Corollary1.Let1≤p<∞,[a,b]⊂R,0< δ <b−2a,n∈N.Thenforanydistinctrealnumbersλ1,...,λn∈R and anyexponential sum fn(x)=
1≤j≤ncjeλjx,∀cj∈Rwe have (b−x)(x−a)fn(x)Lp[a+δ,b−δ] ≤9(b−a)
2
lnb−a δ
1p
nfnLp[a,b]. (7) AnimportantfeatureofLemma1consistsinthefactthatitprovidesanestimateforthederivativesof exponential sums
1≤j≤najeλjx, x,aj ∈R whichis independentof theexponentsλj. Infactthesize of thederivativesessentiallydependsjustonn,thenumberoftermsintheexponentialsum.Whatmakesthis
happenis thepresence oftheweight(b−x)(x−a) inthenorm ofthe derivatives.This weightenforces a lineardeclineattheendpointsoftheinterval.Anaturalquestionarisesinthisrespect:cananupperbound similar toLemma1holdwithaweight(b−x)(x−a)1−ε forsomeε>0?Thenextsimpleexampleshows thatingeneral,theanswertothisquestionisnegative,i.e.,lineardeclineoftheweightiscrucialhere.
Example.Let n = 1 and consider the exponent g(x) = eλx,λ > 4. Clearly with any 0 < ε < 1 and 0< δ <1/3 wehave
x(1−x)1−εg(x)2L2[δ,1−δ]≥
1−δ
1 2
x2(1−x)2−2ελ2e2λxdx≥λδ2−2ε
8 (e2λ(1−δ)−eλ).
Thussettingδ=λ1 implieswhenλ>4
x(1−x)1−εg(x)2L2[δ,1−δ]≥ λ2ε−1e2λ
8 (e−2−e−λ)≥ce2λλ2ε−1. Sinceinaddition,
g(x)2L2[0,1] ≤e2λ λ wearriveat
x(1−x)1−εg(x)2L2[1
λ,1−1λ]≥ce2λλ2ε−1≥cλ2εg(x)2L2[0,1].
Note thatwhen n = 1 and δ = 1λ we get an upper bound of magnitude O(logλ) in the Bernstein type estimate(6) whichmeansthatnomatterhowsmallisε>0 anupperboundsimilarto(6) isimpossiblefor theweight(b−x)(x−a)1−ε.Obviouslythesameconclusionholdsfortheweight(b−x)1−ε(x−a),aswell.
ThenextlemmaprovidesanexplicitconstructionofnodesusedforthediscretizationoftheLpnormsof theexponentialsums.Essentially thesenodesarechosento beequidistributedwithrespect tothemeasure
μ1(E) :=
E
dx
x(1−x), E⊂(0,1) appearingintheBernsteintypeinequality(6).
Lemma2.Forany 0< h<1/2andm≥1set
xj,m:= 1
1 + 1−hh e−(j−1)/m, 1≤j≤N =Nm:= [2mln1−h h ] + 2.
Then1−h≤xNm,m≤1−h/e and
xj+1,m−xj,m≤4x(1−x)
m , x∈(xj,m, xj+1,m), 1≤j≤Nm−1. (8) Proof. Clearly
2 ln1−h
h ≤Nm−1
m ≤2 ln1−h h + 1
and thusweobtain
1−h= 1
1 +1−hh e−2 ln1−hh ≤xNm,m ≤ 1
1 +1−hh e−1 ≤1−h e.
Sethj:= 1−hh e−(j−1)/m.Thenevidently xj+1,m−xj,m
1−xj+1,m
= hj−hj+1
hj+1(1 +hj)≤ hj
hj+1 −1 =em1 −1≤ 2 m.
Hence
xj+1,m−xj,m≤ 2
m(1−x), x∈(xj,m, xj+1,m).
Similarly,
xj+1,m−xj,m xj,m
= 1 +hj
1 +hj+1 −1≤ hj
hj+1 −1 =em1 −1≤ 2 m,
yielding that
xj+1,m−xj,m≤ 2
mxj,m≤ 2
mx, x∈(xj,m, xj+1,m).
Thus forany x∈(xj,m,xj+1,m) wehavethat xj+1,m−xj,m≤ 2
mmin{1−x, x} ≤ 4x(1−x)
m .
WewillalsoneedbelowanL∞Markovtypeinequalityverifiedin[7] formultivariateexponentialsumson convexbodies.Letusdenoteby∇gthegradientofadifferentiablefunctiong,|∇g|standsforitsEuclidian norm which isused indefining∇gL∞(K):= supx∈K|∇g(x)|. Inaddition, c will denote possiblydistinct positive absoluteconstants.
Lemma 3.Let K⊂Rd,d≥1be aconvex body with rK being theradius of its largestinscribed ball. Then for every exponential sum g(w) =
1≤j≤ncjeλj,w,w ∈ Rd with λj ∈ Rd satisfying max1≤j≤n|λj| ≤ Λ, |λj+1−λj|≥ rK,j=k,0< ≤1
∇gL∞(K)≤ cd3n3Λ
gL∞(K).
In addition,when d= 1thestronger inequality
gL∞[a,b] ≤cnΛ
gL∞[a,b]
holds.
AstandardconsequenceofanL∞MarkovtypeinequalityisacorrespondingNikolskiitypeupperbound forsupnormsviatheLp norms.NamelywecaneasilydeducefromLemma3thenext
Corollary2. Forany convexbody K⊂Rd,d≥1and every exponentialsum g(w)=
1≤j≤ncjeλj,w,w∈ Rd with λj ∈ Rd satisfying |λj+1−λj| ≥ rK,j = k,0 < ≤ 1, max1≤j≤n|λj| ≤ Λ we have for any 1≤p<∞and someconstant c(K,d)>0depending onK,d
gL∞(K)≤c(K, d) n3Λ
dp
gLp(K). (9) Whend= 1 wehave thestrongerestimate
gL∞[a,b] ≤c(a, b) nΛ
1p
gLp[a,b]. (10)
Proof. FirstweneedtoobservethatwhenK⊂RdisaconvexbodythenforanyddimensionalballBr⊂Rd ofradius rcenteredinKits intersectionwith K willhaveLebesguemeasure atleast c(K,d)rd with some fixed constantdepending only onK and d. Furthermore, assuming thatgL∞(K) = 1=g(y),y∈ K we have by Lemma3 thatg(x) ≥ 12 whenever x ∈ K∩Br with Br being the ball centered at y and radius r:= 2cd3n3Λ.Hence
gpLp(K)≥
K∩Br
|g|p≥2−pc(K, d)rd≥2−pc(K, d) n3Λ
d
withaproperc(K,d)>0.Thisclearlyimplies(9).Thesecond claimfollowsanalogously.
Nowwehaveallthetoolsneededinorder toproveTheorem 1.
Proof of Theorem1. Clearly,itsufficestoverifythetheoremwhen[a,b]= [0,1],thegeneralcasewillthen followbyastandardlineartransformation.Wewill usetheelementaryestimate
b a
g(x)dx−(b−a)g(a)
≤(b−a) b a
|g(x)|dx
applied to the function g(x) = |f(x)|p on the interval [xj,m,xj+1,m] where xj,m,1 ≤ j ≤ Nm−1 is the discretepointset specifiedinLemma2.Thenitfollowsthat
xj+1,m
xj,m
|f(x)|pdx−(xj+1,m−xj,m)|f(xj,m)|p
≤(xj+1,m−xj,m)
xj+1,m
xj,m
p|f(x)|p−1|f(x)|dx.
Applyingestimate(8) ofLemma2wehave
xj+1,m
xj,m
|f(x)|pdx−(xj+1,m−xj,m)|f(xj,m)|p ≤4p
m
xj+1,m
xj,m
x(1−x)|f(x)|p−1|f(x)|dx, 1≤j≤Nm−1.
Summingupaboveupperboundsfor1≤j ≤Nm−1 wearriveat
xNm
x1,m
|f(x)|pdx−
1≤j≤Nm−1
(xj+1,m−xj,m)|f(xj,m)|p ≤4p
m
xNm
x1,m
x(1−x)|f(x)|p−1|f(x)|dx.
Moreover, applyingtheHölderinequalitytotheintegralontherighthandside itfollowsthat
xNm
x1,m
|f(x)|pdx−
1≤j≤Nm−1
(xj+1,m−xj,m)|f(xj,m)|p ≤ 4p
mfp−1Lp[0,1]x(1−x)f(x)Lp[x1,m,xNm]. Thus
fpLp[0,1]−
1≤j≤Nm−1
(xj+1,m−xj,m)|f(xj,m)|p ≤ 4p
mfpL−p[0,1]1 x(1−x)f(x)Lp[x1,m,xNm]
+
x1,m
0
|f(x)|pdx+ 1 xNm
|f(x)|pdx. (11)
Now weneedtoestimatethethreetermsontherighthandsideof(11).
InordertoestimatethefirsttermrecallthatbyLemma2wehave1−h≤xNm,m≤1−h/eandx1,m=h.
Therefore
x(1−x)f(x)Lp[x1,m,xNm]≤ x(1−x)f(x)Lp[h/e,1−h/e]. Furthermore, applyingestimate(7) with[a,b]= [0,1] andδ:= he yields
x(1−x)f(x)Lp[h/e,1−h/e] ≤9
2(1−lnh)1pnfLp[0,1]. Using thelastupperboundforthefirsttermontherighthandsideof (11) weobtain
4p
mfp−1Lp[0,1]x(1−x)f(x)Lp[x1,m,xNm]≤ 18p
m (1−lnh)1pnfpLp[0,1].
Forthesecond andthirdtermsontherighthandsideof(11) wecanproceed using(10) asfollows
x1,m
0
|f(x)|pdx+ 1 xNm,m
|f(x)|pdx≤2hfpL∞[0,1]≤c1h nΛ
1p
fpLp[0,1].
Substitutingthelast twoboundsinto (11) wearriveat
fpLp[0,1]−
1≤j≤Nm−1
(xj+1,m−xj,m)|f(xj,m)|p ≤18p
m (1−lnh)1pnfpLp[0,1]+c1h nΛ
1p
fpLp[0,1]. (12) Now settingh:=ξ
nΛ
1p
andm:= pn(1−lnh)
1 p
ξ evidentlyyields
fpLp[0,1]−
1≤j≤Nm−1
(xj+1,m−xj,m)|f(xj,m)|p
≤(18 +c1)ξfpLp[0,1].
Since 0< ξ <1 canbe chosen arbitrarily thelast estimate obviously impliestherequired Marcinkiewicz- Zygmundtypeinequalityafterproperchoiceoftheconstantξ.
Itremainsnowto checkthecardinalityNm= [2mln1−hh ]+ 2 ofthediscretepointset.Clearly Nm≤cmln 1
h ≤cpnln1p+1 1
h ≤cpnln1p+1nΛ . Recallingthat|λj+1−λj|≥,j =k andmax1≤j≤n|λj|≤Λ it followsthat
(n−1)≤2Λ (13)
i.e.,lnnΛ ≤clnΛ.ThisclearlyyieldsthatNm≤cpnln1p+1 Λ.
Remark 2.Theorem 1 provides explicit discrete meshes xj of cardinality ∼ nln1p+1 Λ for the weighted Marcinkiewicz-Zygmund type inequalitieswith weights xj+1−xj. The nodes xj are equidistributedwith respect to themeasure μ1(E)=
E
x(1−x)dx , E ⊂(0,1).It shouldbe noted thatTheorem 1remainsvalid for any sequence of nodes satisfying (8) together with theproper restrictions onthe first and last nodes.
In addition, therestrictions of theexponents imposedin Theorem 1wereused only for thefirst and last intervals ofthe partition.This meansthat“near”discretization ofthe normfLp[δ,1−δ] ispossible using orderofnlog1δ nodeswithouttheseparationandthedegreerequirementsontheexponents.
The argumentsgiven in theproof of Theorem 1canbe used to derive aMarcinkiewicz-Zygmund type resultwith equalweights 1foruniformly distributeddiscrete meshesx0:=h,xj :=h+1−2hm j,0≤j ≤m leadingtothenext
Proposition 1.Let 1 ≤ p< ∞,0 < ≤ 1,n ∈ N,Λ > 1. Then given equidistributed discrete points sets x0:=h,xj :=h+1−m2hj,0≤j≤m, h=c
nΛ
1p
of cardinality
m≤c1n Λ
2p
wherec,c1>0areproper absoluteconstants, wehavethat 1
2m
1≤j≤m−1
|g(xj)|p≤ gpLp[0,1] ≤ 2 m
1≤j≤m−1
|g(xj)|p (14)
forevery exponentialsum (2) satisfyingλj+1−λj ≥,1≤j≤n−1andmax1≤j≤n|λj|≤Λ.
Proof. SincetheproofofthepropositionusesthesametechniqueasTheorem1wegiveabriefoutline.The BernsteintypeinequalityofLemma1canbereplacedbyestimate(5).Thensimilarlyto(12) wewillarrive at
fpLp[0,1]− 1 m
0≤j≤m−1
|f(xj,m)|p ≤ cpn
mhfpLp[0,1]+c1h nΛ
1p
fpLp[0,1]. Now setting h :=ξ
nΛ
1p
and m := pnξh with a proper 0< ξ <1 will lead to aMarcinkiewicz-Zygmund typeresultwithequalweights 1anduniformly distributedmeshofcardinality
m≤cn nΛ
p1
≤cn Λ
p2
. Notethatthelastestimateabovefollowsfromtheinequality(13).
Remark 3.As noted at the end of the proof of Theorem 1under given assumptions on exponents λj we have that(n−1) ≤ 2Λ. Thus the quantity Λ appearing both in Theorem 1 and Proposition 1 ideally is of order n for proper distributions of exponents λj. This results in discrete meshes of cardinality ∼ nlnp1+1n, 1 ≤p<∞ inTheorem 1. Similarly inProposition 1when all weights equal 1the cardinality will be of order ∼ n1+2p for every 1 ≤ p < ∞. In a recent paper [3], Theorem 2.3 the authors verified a weighted Marcinkiewicz-Zygmund type inequality for Lp,1 ≤ p ≤ 2 norm with meshes of cardinality
∼ nln3n and arbitrary n-dimensionalsubspaces of L∞. This is a rather general beautiful result, but in contrast to Theorem 1 theweights andmeshes are notgiven explicitly. In additionin [3], Theorem 2.2 a Marcinkiewicz-Zygmund typeresultwith weights1isverifiedfor Lp,1≤p≤2 normswiththecardinality ofthediscretemeshbeingeffectedbythequantityappearinginarequiredNikolskiitypeinequality.Onthe other handit isknownthatforgeneral exponentialsumsthe magnitudeof thefactorsinaNikolskii type inequality depends onthe exponentsλj and it canbe verylarge, see [6], Theorem 1. Henceeven though Theorem 1and Corollary 3are lessgeneral incomparison to [3] in termsof subspacesconsideredfor the exponential sums our results provide a compensation in terms of explicit nature of meshes and weights, theiruniversality,smallercardinalityofdiscretemeshesand theirvalidityforevery1≤p<∞.
We canextend Marcinkiewicz-Zygmundtype inequalitiesfor generalmultivariate exponential sums, as well. Similarly to the univariate case this can be done by applying the Bernstein-Markov type estimates verifiedearlier.Inaddition,wewillhavetoovercomeatechnicaldifficultyrelatedtotheseparationcondition
|λj+1−λj|≥δ whichneeds tobe imposedon theexponents.Forthis end thenextauxiliaryproposition, see [7], Lemma5willberequired.
Lemma 4.Letλj ∈Rd,1≤j ≤ n,d ≥2satisfy the separationcondition |λj −λk| ≥δ > 0,j =k. Then for any w∈Rd,w=0any every >0there existu∈Rd,|w−u|≤, |w|=|u| sothat with some cd >0 depending only ondwe have
|λj−λk,u| ≥ cdδd−1
|w|d−2n2, ∀j=k. (15) The above lemma shows that the separation condition |λj −λk| ≥ δ > 0,j = k is preserved in a certain form when the exponentsare restricted to small perturbations of arbitrarily chosen lines. Clearly the orthogonal projections of λj-s into lines {tu : t ∈ R},u ∈ Sd−1 are given by λj,uu, where these projectionsareseparatedbyquantities|λj−λk,u|.
Consider the unit cube in Rd given by Id := [0,1]d. Using Bernstein-Markov type estimates verified aboveandseparationLemma4we canprovethenextMarcinkiewicz-Zygmundtypeinequalityforgeneral multivariateexponentialsumsonthecube.WeshallusebelowthenotationA∼Binordertoindicatethat c1(d,p)A≤B≤c2(d,p)Awithsomepositiveconstants c1(d,p),c2(d,p) dependingonlyonp,d.
Theorem 2.Let1≤p<∞,d,n∈N,Λ>1,0< δ <1,andconsiderany λj ∈Rd,1≤j≤nsatisfying
|λj−λk| ≥δ >0, j=k, max
1≤j≤n|λj| ≤Λ.
Then thereexistpositiveweights b1,...,bN anddiscretepointsetsYN={w1,...,wN}⊂Id of cardinality N ≤c(d, p)ndlndp+dΛ
δ, so thatforevery exponential sum g(w)=
1≤j≤ncjeλj,w,w∈Rd wehave
gpLp(Id)∼
1≤i≤N
bi|g(wi)|p. (16)
Proof. The proof of the theorem would be a straightforward product type argument if the separation condition|λj−λk|≥δ >0,j =k wastruefortheorthogonal projectionsofλj-sto everycoordinateaxis in Rd. Indeed, if this was the case then successive integration plus standard inductionwould accomplish theproof. Thus if|λj−λk,es|≥δ∗ >0,j =kwith someδ∗ >0 and es := (δs,i)1≤i≤d ∈Rd,1≤s≤d being thestandard basis in Rd then for certain positive weights b1,...,bN and a discrete point set YN = {x1,...,xN}⊂Id ofcardinalityN ≤c(d,p)ndlndp+dδΛ∗
gpLp(Id)∼
1≤i≤N
bi|g(xi)|p, g(x) =
1≤j≤n
cjeλj,x. (17)
Nowwewill applyLemma4tocertainproperlychosenvectors. Forany< 4d1 set e∗s:= (1−)es+
k=s
ek, 1
2 <|e∗s|<2, 1≤s≤d.
Note that each coordinate of e∗s is not smaller than . By Lemma 4 for every 1 ≤ s ≤ d there exist us∈Rd,|e∗s−us|≤,|e∗s|=|us| sothat
|λj−λk,us| ≥ cdδd−1
n2 , ∀j =k. (18)
Sinceallcoordinatesofe∗s are≥thecondition|e∗s−us|≤impliesthat us∈Rd+:={x= (x1, ..., xd)∈Rd:xj≥0,|x|= 1}. ConsidernowtheparallelepipedinRd givenby
Jd:={t1u1+...+tdud, 0≤ts≤ 1
1 + 2d, 1≤s≤d}. Sinceus∈Rd+ and|e∗s−us|≤,1≤s≤ditiseasyto checkthatJd ⊂Id.
Denote by T : Rd → Rd theregular matrixtransformation definedby T(es)= 1+2dus ,1≤s≤d. Then evidently,T(Id)=Jd.Furthermore,notethat
us
1 + 2d−es ≤
e∗s
1 + 2d−es
+≤ |e∗s−es|
1 + 2d + (2d+ 1)≤(3d+ 1), 1≤s≤d.
Evidently,thismeansthatforeveryx=
1≤s≤dtses∈Id,0≤ts≤1 wehave
|x−T(x)|=
1≤s≤d
tses−
1≤s≤d
ts us
1 + 2d =
1≤s≤d
ts
es− us
1 + 2d
≤d(3d+ 1).
Henceforanyx∈Id\JdwehavethatT(x)∈Jdand|x−T(x)|≤d(3d+ 1).ObviouslythismeansthatId iscontainedina1+d(3d+ 1)dilationofJd abouttheoriginwhichinturnimpliesthatμ(Id\Jd)≤cd whereμstandsfortheLebesguemeasure.
Considernowanyexponentialsumg(w)=
1≤j≤ncjeλj,w,w∈Rd with|λj−λk|≥δ >0,j=k.Since Jd⊂Id andμ(Id\Jd)≤cdwehaveby(9)