ebook img

Positive-part moments via the Fourier-Laplace transform PDF

0.23 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Positive-part moments via the Fourier-Laplace transform

Positive-part moments via the Fourier-Laplace transform ∗† 0 Iosif Pinelis 1 0 Abstract: Integral expressions for positive-part moments EX+p (p > 0) 2 of random variables X are presented, in terms of the Fourier-Laplace or FouriertransformsofthedistributionofX.Anecessaryandsufficientcon- n ditionforthevalidityofsuchanexpressionisgiven.Thisstudywasmoti- a vatedbyextremalproblemsinprobabilityandstatistics,whereoneneeds J toevaluatesuchpositive-partmoments. 5 AMS 2000 subject classifications: Primary 60E10, 42A38; secondary 60E07,60E15,42A55,42A61. ] R Keywordsandphrases:positive-partmoments,characteristicfunctions, Fouriertransforms,Fourier-Laplacetransforms,integralrepresentations. P . h t Address: a DepartmentofMathematicalSciences m MichiganTechnologicalUniversity [ Houghton,Michigan49931,USA E-mail:[email protected] 2 Telephone:906-487-2108 v 4 Runninghead: 1 Positive-partmomentsviatheFourier-Laplacetransform 2 4 . 2 0 9 0 : v i X r a ∗Department of Mathematical Sciences, Michigan Technological University, Houghton, Michigan,USA. †SupportedbyNSFgrantDMS-0805946 1 imsart ver. 2005/05/19 file: rev2a.tex date: January 5, 2010 Iosif Pinelis/Positive-part moments via the Fourier-Laplace transform 2 1. Introduction In a number of extremal problems in probability and statistics (see e.g. [11, 12, 19,20,21,22,1,2,3,23,24,25,4,5,26])oneneedstoevaluatethepositive-part moments EXp of random variables (r.v.’s) X, where x := 0∨x = max(0,x), + + the positive part of x, and xp :=(x )p, for any x∈R and p>0. + + In particular, an effective procedure was needed in [26] to compute EXp + for p = 3 and r.v.’s of the form X = Γ +yΠ , where a,b,y are posi- µ,a2 b2/y2 tive real numbers, µ ∈ R, Γ and Π are independent r.v.’s, Γ has µ,a2 b2/y2 µ,a2 the normal distribution with parameters µ and a2, and Π has the Poisson b2/y2 distributionwithparameterb2/y2.Forpurelynormalr.v.’sX (withoutthePois- son component) such a computation is easy. However, the naive approach, by the formula E(Γ +yΠ )p = ∞ E(Γ +yj)p P(Π = j) did not µ,a2 b2/y2 + j=0 µ,a2 + b2/y2 work well, especially when y is smalPl. On the other hand, the series of the form ∞ ejzP(Π = j) (which is the Laplace-Fourier transform of the Poisson j=0 b2/y2 Pdistribution) is easily computable. Therefore, a natural idea was to perform a harmonic analysis of the function R ∋ x 7→ f (x) := xp. This can be easily p + done by finding its Laplace-Fourier transform. Then, once the function f is p decomposed into harmonics (i.e., exponential functions x7→ezx for z ∈C), one almost immediately obtains a general expression for the positive-part moments of a r.v. X in terms of the Laplace-Fourier transform of the distribution of X, as provided indeed by Theorem 1 in this paper. Such expressions turn out to be computationally effective, as well as the ones in terms of the characteristic function(c.f.)R∋t7→EeitX thatareobtainedascorollaries.Theorem2ofthis paper provides a necessary and sufficient condition for such a representation. A subsequent literature search has revealed only one paper, Brown [8], that contains formulas for EXp in terms of the c.f. of X. However, the constant- + sign argument used in [8] does not actually seem to work for p ∈ (0,2). The resultsin[8]arebasedonthemethoddevelopedbyvonBahr[29]toexpressthe absolutemomentsE|X|p intermsofthec.f.ofX.Otherpaperscontainingsuch expressions for absolute moments include [14, 8, 9]; see also [18, §§1.8.6–1.8.8] and [16, §11.4]. Aswasexplained,ourapproachdiffersfromthepreviousones.Webeginwith expressions of positive-part moments in terms in the Fourier-Laplace transform z 7→EezX with Rez 6=0, and then use the Cauchy integral theorem to express EXp intermsofthec.f.ofX.Itshouldbeclearthatexpressionsfortheabsolute + moments will follow trivially from expressions for the positive-part moments. 2. Results Let X be any r.v. Let s and s be any real numbers such that s 606s and 1 2 1 2 EesX <∞ for all s∈[s ,s ]. Let p be any positive real number, and then let 1 2 k :=k(p):=⌊p⌋ and ℓ:=ℓ(p):=⌈p−1⌉, respectively the integer part of p and the smallest integer that is no less than p−1, so that k 6p<k+1, ℓ<p6ℓ+1, and ℓ6k. For all complex z and all imsart ver. 2005/05/19 file: rev2a.tex date: January 5, 2010 Iosif Pinelis/Positive-part moments via the Fourier-Laplace transform 3 m=−1,0,1,..., let m zj m z2j e (z):=ez − , c (z):=(−1)m+1 cosz− (−1)j , m 2m X j! (cid:16) X (2j)!(cid:17) j=0 j=0 m z2j+1 s (z):=(−1)m+1 sinz− (−1)j , 2m+1 (cid:16) X (2j+1)!(cid:17) j=0 with the convention −1 a := 0 for any a ’s, so that e (z) ≡ ez, c (z) ≡ j=0 j j −1 −2 cosz, and s (z)≡siPnz. For any complex number z =s+it, where s and t are −1 real numbers and i stands for the imaginary unit, let Rez := s and Imz := t, the real and imaginary parts of z. In this paper, we shall present a number of identities involving certain in- tegrals. For the sake of brevity, let us assume the following convention (unless specifiedotherwise):whensayingthatsuchanidentitytakesplaceundercertain conditions, we shall actually mean to say that under those conditions the cor- responding integral exists in the Lebesgue sense (but may perhaps be infinite) and the identity takes place. Theorem 1. For any s∈(0,s2] and any j =−1,0,...,ℓ such that E|X|j+ < ∞, Γ(p+1) ∞ Ee (s+it)X EXp = j dt (1) + 2π Z (s(cid:0)+it)p+1 (cid:1) −∞ Γ(p+1) ∞ Ee (s+it)X = Re j dt. (2) π Z (s(cid:0)+it)p+1 (cid:1) 0 Remark 1. Of course, the statement of Theorem 1 is devoid of content in the casewhens2 =0.AsfortheconditionE|X|j+ <∞,hereweusetheconvention 00 :=1(any other realnumber inplace of 1would dohereaswell), to interpret |X|j+ whenX =0andj ∈{−1,0}. So,thecondition E|X|j+ <∞willtrivially hold if j ∈ {−1,0}. On the other hand, if j > 1 (and hence p > 1), then the expressionEe (s+it)X undertheintegralsontheright-handsidesof (1)and j (2) would lose(cid:0)meaning w(cid:1)ithout the condition E|X|j+ <∞. Corollary 1. If p ∈ N and E|X|p < ∞, then for any s ∈ [s ,0) and any 1 j =−1,0,...,ℓ p! ∞ Ee (s+it)X EXp =EXp+ j dt (3) + 2π Z (s(cid:0)+it)p+1 (cid:1) −∞ p! ∞ Ee (s+it)X =EXp+ Re j dt. π Z (s(cid:0)+it)p+1 (cid:1) 0 Corollary 2. If E|X|p <∞ then EXk Γ(p+1) ∞ Ee (itX) EXp = I{p∈N}+ Re ℓ dt. (4) + 2 π Z (it)p+1 0 imsart ver. 2005/05/19 file: rev2a.tex date: January 5, 2010 Iosif Pinelis/Positive-part moments via the Fourier-Laplace transform 4 In particular, one has the following: for any m∈N such that EX2m <∞, EX2m (2m)! ∞ Es (tX) EX2m = + 2m−1 dt; (5) + 2 π Z t2m+1 0 for any m∈N such that E|X|2m−1 <∞, EX2m−1 (2m−1)! ∞ Ec (tX) EX2m−1 = + 2m−2 dt. (6) + 2 π Z t2m 0 Corollary 3. If Y is another r.v. then Γ(p+1) ∞ EeitX −EeitY EXp =EYp+ Re dt (7) + + π Z (it)p+1 0 provided that E|X|p+E|Y|p <∞ and EXj =EYj for all j =1,...,k. Moreover, one has the following: for any m∈N (2m)! ∞ EsintX −EsintY EX2m =EY2m+(−1)m dt (8) + + π Z t2m+1 0 if EX2m+EY2m <∞ and EX2j+1 =EY2j+1 for all j =0,...,m−1,m− 1; 2 for any m∈N (2m−1)! ∞ EcostX −EcostY EX2m−1 =EY2m−1+(−1)m dt (9) + + π Z t2m 0 ifE|X|2m−1+E|Y|2m−1 <∞andEX2j =EY2j forallj =0,...,m−1,m−1. 2 Foranyr.v.X asinCorollary2,itisalwayspossible(andeasy)toconstruct a r.v. Y as in Corollary 3 such that the characteristic function EeitY, and hence its real and imaginary parts EcostY and EsintY, are easily computable. In particular, one can always take Y to be a r.v. with finitely many values; more specifically, k+1 values would suffice for (7) and m+2 values for (8) or (9). An obvious advantage of the formulas given in Corollary 3 (over those in Corollary 2) is that the corresponding integrals will converge (for p>1) faster near ∞, and just as fast near 0. Also, Corollary 3 will be just one step closer thanCorollary2topossibleapplicationstotheconvergenceofthepositive-part moments in the central limit theorem; see Example 4 in Section 3 for further details. Remark 2. If EXp < ∞ for an even natural p = 2m, then it is clear that + identity(4)(or,equivalently,(5))cannotholdwithouttheconditionE|X|p <∞, sinceoneneedsthemomentEXk =EXp ontheright-handsideof (4)toexist. Similarly, if EXp < ∞ for an odd natural p = 2m−1, then identity (4) (or, + equivalently, (6)) cannot hold without the condition E|X|p <∞. However, if p > 0 is not an integer, the term EXk I{p ∈ N} on the right- 2 hand side of (4) disappears. One may then wonder as to what, if any, moment condition on the left tail of the distribution of X is needed in order for identity imsart ver. 2005/05/19 file: rev2a.tex date: January 5, 2010 Iosif Pinelis/Positive-part moments via the Fourier-Laplace transform 5 (4)tohold–havinginmindthatthefinitenessofthepositivepartmomentEXp + depends only on the right tail of the distribution of X. Perhaps surprisingly, it turnsoutthataleft-tailconditionisnecessaryfor(4),eveniftheintegralin(4) is allowed to be understood as an improper one (at 0); moreover, the moment conditionE|X|p <∞inCorollary2canbeonlyslightlyrelaxed,evenwhenpis not an integer. Actually, one can obtain a necessary and sufficient condition for such a representation of the positive-part moment in terms of the characteristic function, as described in the following theorem. Theorem 2. Take any p ∈ (0,∞)\N and any r.v. X with EXp < ∞. Then + the following two conditions are equivalent to each other: Γ(p+1) ∞ Ee (itX) (I) E|X|ℓ <∞ and EXp = Re ℓ dt; + π Z (it)p+1 0+ (II) P(X >x)=o(1/xp) as x→∞. − Here and in what follows, x :=(−x) for all x∈R; also let xp :=(x )p. − + − − Note that the part E|X|ℓ < ∞ of condition (I) of Theorem 2 will trivially hold if 0 < p < 1 (and hence ℓ = 0). On the other hand, if p > 1 (and hence ℓ > 1), then the expression Ee (itX) in condition (I) of Theorem 2 would lose ℓ meaning if E|X|ℓ =∞; cf. Remark 1. Notealsothat,ifcondition(II)or,equivalently,(I)ofTheorem2holds,then EXr <∞ for all r ∈(0,p). Indeed, condition (II) of Theorem 2 implies − EX−r =R0∞rxr−1 P(X− >x)dx=O R0∞rxr−1(1∧x−p)dx <∞. (10) (cid:0) (cid:1) On the other hand, it is easy to give examples where condition (II) of Theo- rem2holds,whileEXp =∞,andso,Corollary2isnotapplicable;forinstance, − take any r.v. X such that P(X >x)=1/(xplnx) for all large enough x>0. − InviewofCorollary2andTheorem2,fortheidentity (4)toholdforagiven p ∈ (0,∞)\N (with the integral understood in the Lebesgue sense) and a r.v. X with EXp < ∞, it is sufficient that EXp < ∞ and it is necessary that + − EXr <∞ for all r ∈(0,p). − Yet, no “simple” necessary and sufficient condition for (4) to hold – in the Lebesguesense–foragivenp∈(0,∞)\Nappearstobeknown;itisapparently an open question whether such a condition exists at all. However, one can see that condition (II) of Theorem 2 is not sufficient (although, by Theorem 2, it is necessary) for (4). Indeed, for any p ∈ (0,∞)\N, there exists a r.v. X with EXp <∞suchthatthetwoequivalentconditions–(I)and(II)–ofTheorem2 + hold, while the integral ∞ReEeℓ(itX)dt does not exist in the Lebesgue sense. 0 (it)p+1 For instance, one can coRnsider Example 1. The idea is to take a r.v. X with a lacunary distribution and then construct a matching lacunary set A ⊂ (0,∞) such that the integrand ReEeℓ(itX) is of constant sign and large enough in absolute value for t ∈ A. (it)p+1 Take indeed any p∈(0,∞)\N and let X be a discrete r.v. such that, for some d∈(1,∞), one has P(X =−dn)= c for all n∈N, where c is the constant ndnp chosen so that P(X =−dn)=1. Then it is easy to check that condition n∈N P imsart ver. 2005/05/19 file: rev2a.tex date: January 5, 2010 Iosif Pinelis/Positive-part moments via the Fourier-Laplace transform 6 (II) (and hence condition (I)) of Theorem 2 holds. Moreover, X =0 a.s., and + so, EXp < ∞. Further, one can show that there exist some d = d(p) > 1, + ε=ε(p)∈(0,1−1/d), and m=m(p)∈N such that (−1)mReEeℓ(itX) > 1 (it)p+1 tlnε for all t ∈ A := ∞ [ε(1−ε), ε ], whence (−1)m ReEeℓ(itX)dt = ∞. Fotr n=1 dn dn A (it)p+1 details, see [27, PrSoposition 2.6 and its proof]. R Theorem 2 and Example 1 are similar in spirit to a number of known if- and-only-if conditions and counterexamples, respectively; those results in the literaturemainlyconcernsuchsimplerrelationsastheonesofthetails/moments of the distribution with the derivatives of the characteristic function at 0 (but alsowithotherlinearfunctionalsofthec.f.)–seee.g.[13,30,28,6,7].Ibragimov [15] provides necessary and sufficient conditions – both in terms of (truncated) moments/tails and the c.f. – for rates O(n−p) in the central limit theorem to hold; cf. Example 4 in the next section. 3. Applications Here let us give examples of known or potential applications of identities pre- sented in Section 2. Example 2. Let X ,...,X be independent r.v.’s such that X 6 y for some 1 n i real y > 0 and EX 6 0, for all i. Let S := X +···+X and assume that i 1 n EX2 6 σ2 for some real σ > 0. Also, let ε := 1 E(X )3, so that i i σ2y i i + P0 < ε < 1. In the paper [26], mentioned in the IntroductiPon, an optimal in a certain sense upper bound on the tail P(S >x) was obtained, of the form E3(η−t )2 Pin(x):= x +, E2(η−t )3 x + where x ∈ R; η := Γ +yΠ˜ ; Γ and Π˜ are any indepen- (1−ε)σ2 εσ2/y2 (1−ε)σ2 εσ2/y2 dentr.v.’swhosedistributionsare,respectively,thecenteredGaussianonewith variance (1−ε)σ2 and the centered Poisson one with variance εσ2/y2; t ∈R is x theonlyrealrootoftheequationm(t )=x;andm(t):=t+E(η−t)3/E(η−t)2. x + + Thus,tocomputetheboundPin(x),oneneedsafastcalculationofthepositive- part moments of the r.v. X :=η−t. While the direct approaches mentioned in the Introduction do not work here, the formulas of Theorem 1 and Corollary 2 proveveryeffective,sincetheFourier-Laplacetransformofther.v.η−tisgiven by a simple expression: Eez(η−t) =exp −zt+ λ2 (1−ε)σ2+ ezy−1−zy εσ2 ∀z ∈C. (11) 2 y2 (cid:8) (cid:9) It takes under 0.5 sec in Mathematica on a standard Core 2 Duo laptop to produce an entire graph of the bound Pin(x) for a range of values of x using either (2) or (4). Example 3. Aspointedoutbyareferee,positive-partmomentsnaturallyarise in mathematical finance. For example, (S −K) is the value of a call option + imsart ver. 2005/05/19 file: rev2a.tex date: January 5, 2010 Iosif Pinelis/Positive-part moments via the Fourier-Laplace transform 7 with the strike price K when the stock price is S. Exact bounds on E(S−K) + under various conditions were recently given in [10]. Example 4. Just as it was done in [8] for absolute moments, the identities given in Corollary 3 could be used to prove the convergence of the positive-part moments in the central limit theorem, or even to obtain bounds on the rate of suchconvergence.However,weshallnotbepursuingthismatterhere,sincethe mainmotivationforthepresentpaperwastheneedforaneffectivecomputation of the bound Pin(x), as described in Example 2. 4. Proofs For any two expressions a and b, let us write: a < b or, equivalently, b > a if ⌢ ⌢ |a| 6 Cb for some positive constant C depending only on p; a ≍ b if ab > 0, a<b, and b<a. ⌢ ⌢ Letusalsowritea<<b(toberead“aismuchlessthanb”)or,equivalently, b >> a (“b is much greater than a”) if b > 0 and |a| = o(b). Accordingly, if A is an assertion and a > 0, then “A holds if a << b” will mean “for all p ∈ (0,∞)\N there exists some constant K = K(p) > 0 such that A holds whenever |a| < b/K”, that is, A will hold “eventually”. Similarly, “if a << b and A holds, then a << b ” will mean “for all p ∈ (0,∞)\N and all K > 0 1 1 1 there exists K = K(p,K ) > 0 such that one has the implication ‘if |a| < b/K 1 and A holds, then |a |<b /K ’”. 1 1 1 Also, let us write a∼b if a/b→1. Proof of Theorem 1. Note that (2) immediately follows from (1), since the in- tegrandin(2)isevenint∈R.Notealsothat zj dz =0forallj =0,...,ℓ and s 6= 0, where C := {z ∈ C: Rez = s}. SRoC,siztp+is1enough to prove (1) with s j = −1. The key observation here is the following: for any x ∈ R and any complex number z =s+it with s:=Rez >0, ∞ x ∞ Γ(p+1) ezu(x−u)p du= ezu(x−u)pdu=ezx vpe−zvdv = ezx; Z + Z Z zp+1 −∞ −∞ 0 thisisobviousforrealz >0,andthenonecanuseanalyticcontinuation. Thus, for each s ∈ (0,∞) the Fourier transform R ∋ t 7−→ ∞ eituesu(x−u)p du −∞ + of the integrable function R ∋ u 7−→ esu(x−u)p is Rthe function R ∋ t 7−→ + Γ(p+1) e(s+it)x , which is integrable as well. So, by the inverse Fourier trans- (s+it)p+1 form, oneobtains (1)(with j =−1)with therealconstant xinplace ofthe r.v. X.(Equivalently,onecanuseheretheLaplaceinversion.)Now(1)followsbythe Fubini theorem, since E e(s+it)X = EesX for all s>0 and t∈R. (s+it)p+1 (s2+t2)(p+1)/2 (cid:12) (cid:12) Proof of Corollary 1. H(cid:12)ere, just(cid:12)as in the proof of Theorem 1, without loss of generality (w.l.o.g.) j = −1. Consider first the case when s > 0. W.l.o.g., 2 s < 0. By the convexity of EesX in s, one has Ee(s+it)X 6 Ees1X∨Ees2X 1 (s+it)p+1 |t|p+1 (cid:12) (cid:12) (cid:12) (cid:12) imsart ver. 2005/05/19 file: rev2a.tex date: January 5, 2010 Iosif Pinelis/Positive-part moments via the Fourier-Laplace transform 8 for all s ∈ [s ,s ] and all t 6= 0, so that s2 Eej (s+it)X ds → 0 as |t| → ∞. Hence, by1th2e Cauchy integral theoremRsa1n(cid:12)d t(hs(cid:0)e+ict)opn+t1in(cid:1)u(cid:12)ity of the integral (cid:12) (cid:12) I(s) := ∞ Ee(s+it)X dt in s ∈ [s ,s ], the difference between the value of I(s) −∞ (s+it)p+1 1 2 for s ∈ (R0,s ] and that of I(s) for s ∈ [s ,0) is 1 EezX dz, where C is the circle in C o2f radius ε centered at 0 and 1traced oiuRtCcεouzpn+t1erclockwise, pεrovided that ε∈(0,s ∧s ). So, (3) (in the case when s >0 and with j =−1) follows 1 2 2 from (1) by taking ε ↓ 0, since ezX = e(zX) +O(|zX|p+1eε|X|) over z ∈ Cε; p recall that here it is assumed that p∈N. Assume now that s =0 (and, again, j =−1). Then, by what has just been 2 proved, one has (3) for any s ∈ [s ,0) with X ∧N in place of X, where N is 1 any real number. By the dominated convergence theorem and in view of the condition E|X|p < ∞, one has E(X ∧N)p → EXp and E(X ∧N)p → EXp + + as N → ∞. To complete the proof of Corollary 1, it remains to use again the dominated convergence theorem, in view of the inequalities Ee(s+it)(X∧N) 6 (s+it)p+1 (s2E+ets2()X(p∧+N1))/2 and es(X∧N) 61+es1X for all N >0, s∈[s1,0),(cid:12)(cid:12)and t∈R. (cid:12)(cid:12) Proof of Corollary 2. Let us first prove identity (4) in the case when the r.v. X takes on only one value, say x ∈ R. For any ε > 0, consider the integration contour C := {it: t 6 −ε}∪C+ ∪{it: t > ε}, where C+ := {z ∈ C: |z| = ε ε ε, Rez >0}. Then, by (1) (with j =−1) and the Cauchy integral theorem, 2πi ezx e (zx) xp = dz = ℓ dz, (12) Γ(p+1) + Z zp+1 Z zp+1 C C because zj dz =0forallj =0,...,ℓ.Next,e (zx)= zℓ+1 xℓ+1+O(|z|ℓ+2) C zp+1 ℓ (ℓ+1)! as z →0R, and so, e (zx) xℓ+1 dz ℓ dz = +O(εℓ−p+2) ZCε+ zp+1 (ℓ+1)!ZCε+ zp−ℓ xℓ+1 =iπ I{p∈N}+O εℓ−p+1(I{p∈/ N}+ε) (ℓ+1)! (cid:0) (cid:1) xℓ+1 xk =iπ I{p∈N}+o(1)=iπ I{p∈N}+o(1) (ℓ+1)! Γ(p+1) as ε↓0. This and (12) imply xk Γ(p+1) ∞ e (itx) xp = I{p∈N}+ Re ℓ dt, (13) + 2 π Z (it)p+1 0+ since the latter integrand is even in t∈R. Next, observe that |e (iu)|<|u|ℓ∧|u|ℓ+1 (14) ℓ ⌢ over all u∈R, and so, for all x∈R ∞ e (itx) ℓ dt=c |x|p, (15) Z (cid:12)(it)p+1(cid:12) ℓ 0 (cid:12) (cid:12) (cid:12) (cid:12) imsart ver. 2005/05/19 file: rev2a.tex date: January 5, 2010 Iosif Pinelis/Positive-part moments via the Fourier-Laplace transform 9 where c := ∞| eℓ(iu) |du < ∞ for p ∈ (0,∞)\N. Therefore, (4) follows from ℓ 0 (iu)p+1 (13) by the FRubini theorem – provided that p∈(0,∞)\N. The case p∈N is treated quite similarly. Namely, if p=2m for some m∈N, then Re eℓ(itx) = s2m−1(tx) for all x ∈ R and t > 0, and so, (5) follows by the (it)p+1 t2m+1 Fubini theorem from (13), the estimate |s (u)| < |u|2m−1∧|u|2m+1 for all 2m−1 ⌢ u∈R, and the condition EX2m <∞. If p=2m−1 for some m∈N, then (6) similarly follows using the estimate |c (u)| < |u|2m−2∧|u|2m for all u ∈ R 2m−2 ⌢ and the condition E|X|2m−1 <∞. Proof of Corollary 3. This immediately follows from Corollary 2, applied the r.v. Y, as well as to X. (Note that k = p and ℓ = p−1 if p ∈ N, and k = ℓ if p∈/ N.) Proof of Theorem 2. Step 0. Take indeed any p ∈ (0,∞)\N and any r.v. X with EXp < ∞. Recalling (10), note that either one of the conditions (I) and + (II) implies E|X|ℓ <∞. (16) Step 1. For v >0, let ∞ e (−iu) I(v):=I (v):= Re ℓ du. (17) p Z (iu)p+1 v Thisdefinitioniscorrect,sincetheintegralexistsintheLebesguesense,inview of estimate (14). Note that I (0)=0; (18) p this is a special case of (4), with X = −1 almost surely (a.s.). In turn, (18) implies v e (−iu) I (v)=− Re ℓ du. (19) p Z (iu)p+1 0 Step 2. At this step, let us prove that if p∈(0,1) and v ∈(0,∞), then I (v)>0. (20) p Take indeed any p∈(0,1). Then ℓ=0, and (17) and (19) can be rewritten as ∞ sinπp −sin(πp +u) I(v)= 2 2 du (21) Z up+1 v v sin(πp +u)−sinπp = 2 2 du (22) Z up+1 0 for all v > 0. Observe next that vp+1I′(v) = sin(πp +v)−sinπp for all v > 2 2 0, whence the only local minima of I in (0,∞) are at the points 2π,4π,.... Therefore, and because of (18) and the obvious equality I(∞−) = 0, to prove (20) it suffices to show that I(2jπ)>0 for all j ∈N. imsart ver. 2005/05/19 file: rev2a.tex date: January 5, 2010 Iosif Pinelis/Positive-part moments via the Fourier-Laplace transform 10 Using (21) and (twice) applying the Fubini theorem, for all j ∈N and q ∈N such that j <q one has 2qπ sin(πp +2jπ)−sin(πp +u) I(2jπ)−I(2qπ)= 2 2 du Z up+1 2jπ 2qπ du u πp =− dt cos +t Z2jπ up+1 Z2jπ (cid:16) 2 (cid:17) 2qπ πp 2qπ du =− dt cos +t Z2jπ (cid:16) 2 (cid:17)Zt up+1 1 2qπ πp 1 1 =− dt cos +t − p Z2jπ (cid:16) 2 (cid:17)(cid:16)tp (2qπ)p(cid:17) 1 2qπ πp 1 =− dt cos +t p Z2jπ (cid:16) 2 (cid:17)tp 1 2qπ πp ∞ =− dt cos +t up−1e−tudu pΓ(p) Z2jπ (cid:16) 2 (cid:17) Z0 1 ∞ 2qπ πp =− up−1du dt cos +t e−tu Γ(p+1) Z0 Z2jπ (cid:16) 2 (cid:17) 1 ∞ sinπp −ucosπp = up−1 2 2 (e−2jπu−e−2qπu)du; Γ(p+1) Z 1+u2 0 letting now q →∞ and using again the equality I(∞−)=0, by the dominated convergence and a comparison of the integrands one obtains 1 ∞ sinπp −ucosπp I(2jπ)= up−1 2 2 e−2jπudu Γ(p+1) Z 1+u2 0 1 ∞ sinπp −ucosπp > up−1 2 2 e−2jπudu Γ(p+1) Z 1+tan2 πp 0 2 (2jπtanπp −p) cos3 πp (jπ2−1) cos3 πp = 2 2 > 2 >0, (23) (2jπ)p+1p (2jπ)p+1 which completes the proof of (20). Step 3. Still assuming that p ∈ (0,1), it follows from (23) that I(2jπ) > (jπ2(−2j1π))cpo+s13 π2p ⌢>j−p overallj ∈N.Next,foreveryj ∈N,theonlylocalminima ofI ontheinterval[2jπ,(2j+2)π]aretheendpoints;so,I(v)>I(2jπ)∧I (2j+ 2)π >j−p >v−p overallv ∈[2jπ,(2j+2)π]andallj ∈N.Thatis,I(v)(cid:0)>v−p ⌢ ⌢ ⌢ over(cid:1)all v ∈ [2π,∞). On the other hand, by (21), |I(v)| 6 ∞ 2 du < v−p v up+1 ⌢ over all v ∈(0,∞). Thus, I(v)≍v−p over all v ∈[2π,∞). R Yetontheotherhand,by(22),I(v)≍ v u du≍v1−p overallv inaright 0 up+1 neighborhood of 0. R Also, I is obviously continuous on (0,∞). So, in view of (20), for each p ∈ p (0,1), I (v)≍v−p(v∧1) over all v >0. (24) p imsart ver. 2005/05/19 file: rev2a.tex date: January 5, 2010

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.