GRIFFIN BOOKS ON STATISTICS, &. A statistical primer FN. DAVID An introduction 10 the theory of statistics G. v, yute and M. G. KENDALL The advanced theory of statistics (three volumes) M. G. KENDALL and ‘A STUART Rank correlation methods M. G. KENDALL Exercises in theoretical statistics MG. KENDALL The design and analysis of experiment MH. QueNouRLE Sampling methods for censuses and surveys yates Biomathematies cA. a. suri Statistieal method in biological assay D., FINNEY The mathematical theory of epidemics NTs. BAILEY Probability and the weighing of evidence |. Goon Grurrin’s Stanisticat, MonocRaruis AnD Coursts: No. 1: The analysis of multiple time-series M. W. QUENOUILLE No. 2: A course in multivariate analysis M. 6, KENDALL No. 3: The fundamentals of statistical reasoning M, W. QUENOULLE "No. dA frst course in sampling A. STUART No.5: Characteristic fanctions LUKACS No. 6: Air introduction to infinitely many variates. 8, A. ROBINSONCHARACTERISTIC FUNCTIONS EUGENE LUKACS Professor of Mathematies ‘The Catholic University of America Washington, D.C. BEING NUMBER FIVE oF GRIFFIN’S STATISTICAL MONOGRAPHS & COURSES EDITED BY M. G. KENDALL, Sc.D. (DOUBLE VOLUME) CHARLES GRIFFIN & COMPANY LIMITED. LONDONCopyright © 1960 CHARLES GRIFFIN & COMPANY LIMITED 42 DRURY LANE, LONDON, W.C.2, First published in 1960 J. W, ARLOWSAUTH LEED,, Winterstoke Road, Bristol, 3PREFACE Characteristic functions and generating functions were originally developed as a tool for the solution of problems in Probability Theory and admit many important applications in this branch of mathematics as well as in Mathematical Statistics. In the present monograph we study characteristic functions for their intrinsic mathematical interest and are not concerned with their possible applications. ‘This statement could be taken as an admission that this book does not properly fit into a series of statistical monographs and courses. However, a mathe- matical statistician needs a good background in pure mathematics, and the methods discussed in a considerable part of this monograph should be contained in his tool chest. ‘The literature of the subject is widely scattered over many periodicals, and a substantial part of the important results can be found in Russian and French journals. ‘This part of the literature is often not easily accessible and its use is made even more difficult by the. language barrier. A reasonably complete account of the subject is only possible if the foreign literature is adequately covered. ‘The author hopes that such a coverage is provided in this monograph and that this feature will increase its usefulness for students of the subject, Certain parts of the theory, primarily those discussed in sections 8.2 and 8.5, are the areas in which major advances are being made at present. Important new results had been announced without proof shortly before the manuscript was completed. It was only possible to mention some of these advances (for example theorem 8.2.7) in order to indicate the direction in which research is proceeding at present. Most of the theorems discussed in this book are given with detailed proofs; occasionally it was possible only to state some results and to refer the reader to the literature for the proofs. ‘The monograph does not contain a complete bibliography; we give selected references for readers who wish cither to find the proofs not given in the book or who desire to study some background material. ‘The book can therefore not be used to decide questions of priority. ‘The author to whom a particular result is due, is often not even mentioned. If itis customary to associate the name of a particular mathematician with a theorem, then this is done, but the paper in which this result was published is not necessarily listed in the References. An article is included in the list of references only if it contains material which is mentioned in the mono- staph but not discussed in sufficient detail. T should like to express my thanks to Dr. R. G. Laha, Dr. I. R Savage and Dr. S. Vajda for reading the manuscript. ‘The author is 56 particularly indebted to Dr. R. G. Laha for valuable comments and advice. ‘Thanks are also due to the Cath versity for providing clerical help; to Miss I. Christensen, Dr. R. G, Taha, Mrs. E. C. Lukaes, and Mr. B. Ramachandran who read the proofs; and to Mrs. K. Winner who typed the manuscript. ‘Ihe book was written with the partial support of the National Science Foundation under grant NSF-G-4220 which is gratefully acknowledged here. Eucexe Lukacs Washington, D.C., December 1958CONTENTS Page PREFACE 5 1, INTRODUCTION LA Distribution functions 9 112. Examples of distribution functions 3 13) ‘The method of integral tecafoms 7 14 Moments 19 2. PRELIMINARY STUDY OF CHARACTERISTIC FUNCTIONS, 21 Elementary propertics of characterntic functions 2 22 Lebesgue decomposition of characteristic functions 26 23 Charscteritie functions and moments Fa BA ‘The second characteristic B 3. FUNDAMENTAL PROPERTIES OF CHARACTERISTIC FUNCTIONS 3A The uniqueness theorem 35 32. aversion formulse u 33. The convolution theorem 8 34 Limite of distribution fanetions a 35. The theorem of Helly % 3 ‘The continuity theorem 3 4. CRITERIA FOR CHARACTERISTIC FUNCTIONS 41 Necessary conditions 9 42. Necessary and sufficient conditions ot 43° Sateient conditions 0 $4 An estential property of characteristic functions B 5. FACTORIZATION PROBLEMS CHARACTERISTIC FUNCTIONS. SFINITELY DIVISIBLE SL Preliminary remarks B 5.2 Definition of inirtely divisible characteristic functions a 53. Elementary properties of inhaltely divisible charactenstic functions 80 34 Construction of tafintely divisible sharsetenstie functions 3 55. Canonical representations 8 $5 Alimit theorem 98 33 Stable distributions 3 6, PACTORIZATION PROBLEMS—GENERAL THEOREMS FROM ‘THE ARITHMETIC OF DISTRIBUTION FUNCTIONS 6.41 Some notations ad lemmas 108 5.2 General decorsponition theorems 1D 63 Indecomparuble charseseriatie Functions 13 7. ANALYTIC CHARACTERISTIC FUNCTIONS. LL The stip of regularity and the integral representation | 130 72. Ansiytte charsetenatic functions snd their distribution functions 136 73 Cteria for analytic characteristic functions us 74. Periodic analytic characteristic functions 158 TS. Analytic charscterstie functions sy solutions of certain differential equations 160 78 8. FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 8.1 Propertce of the factors of an analytic characteriatie function 8:2 Ractoriztion of certain entire characteratye function 8.3 Determination of certain entire characteris functint by properties of thee factors 844 Infinitely divisible analytic characteristic function 5. Extension of some factorization theorem fr analytic characteristic 9.1 Mixtures of distribution functions 912 ‘Transformations of characteristic Functions APPENDICES Appendix ‘The notations O and o Appendix B: Schwarz's inequality ‘Appendix C: Welerseass' approzimation theorem REFERENCES INDEX 179 17 12 187 190 199 202 207 308 210 a31, INTRODUCTION It is now universally recognized that probability theory is a branch of measure theory. Thus probability is defined as a bounded and normed measure. An important part of probability theory is devoted to the study of finite and measurable functions, the random variables. ‘The probability that the value of a random variable belongs to a given set is of interest; this probability is a set function and is called the proba- bility function of the random variable. It is known that this set function determines a point function, the distribution function of the random variable. Many of the most important problems concerning random variables can be expressed in terms of distribution functions. ‘This part of probability theory can be studied independently of its measure theor- etic foundation and of theorems of purely measure theoretic character. ‘The methods of classical analysis provide an efficient approach to prob- lems of this sort. i "The present monograph deals with one of the most powerful tools used for the investigation of distribution functions, namely the charac- teristic functions. Before defining characteristic functions it is neces- sary to summarize briefly some important concepts. 1.1 Distribution functions In this section, in which we discuss some of the basic properties of distribution functions, we will confine ourselves mostly to the listing of definitions and theorems. Proofs will be only given if they are short, otherwise the reader is referred to standard texts such as the books of Cramér (1946) and Loéve (1955). ‘The lack of proofs in this chapter should not impede the reader in using this monograph. In fact, it will be possible to follow the discussion of the subsequent chapters if one takes for granted the statements made in this introduction. Familiarity with proofs which are omitted is not required. For the sake of brevity we introduce the following notation, Let F(x) be a function of the real variable x, we write then Flt oo) = lim F(a) F-o) lim Fs) ey | F(x+0) = lim F(x+h) | Fee—0) = Jing Fen °10 Here 40 means that / tends to zero from above, assuming only positive values. ‘A function F(x) of a real variable x is called a distribution function if it satisfies the following three conditions:— (i) F(a) is non-decreasing, that is F(w+i) 2 F(x) if h > 0 F(x) is right-continuous, i.e. F(x+0) = F(x), Gil) F400) = 1, 00) ‘Thus distribution functions are by definition bounded and monotone and many of their basic properties follow from this fact. For instance a distribution function can have only discontinuities of the first kind, A point x is called a disco point of the distribution function F(x) if F(x+0) = F(x) # F(e—0); if F(x) = F(x —0) then x is called a continuity point of P(x). The quantity (11.2) pe = Fle +0)~ F(x—0) = Fx) —F(x—0) is called the saltus (jump) of F(x) at the point x. ‘The saltus of a distri- bution function is positive at its discontinuity points and zero at its continuity points. An interval is called a continuity interval for the distribution function F(x) if both its endpoints are continuity points of F(s). A point « is called a point of increase of the distribution func- tion F(x) if F(x+e) —F(e—e) > 0 for any « > 0. Let F(x) be a distribution function and & be positive integer. Denote by De the sct of discontinuity points of F(x) with saltus con- tained in the half open interval lives): this set contains at most +1 points. ‘The set of all discontinuity points of F(x) is the union of all sets Di(k = 1, 2, ...) and is therefore at most cenumerable, We have therefore proved ‘2 Theorem 1.1.1. The set of discontinuity points of a distribution function is at most enumerable. Let {x,} be the set of discontinuity points of the distribution func- tion F(x) and write pz, = F(x,+0)~—F(x,—0) for the saltus of F(x) at the point x,. We define the Function (13) O)= Paw the summation is here extended over all discontinuity points not exceeding x. The function (x) increases only by jumps, at points of the scquence {x,}, and is constant in every closed ‘interval not containing an x,. Such a function is called a step-funetion. ‘The (+) See E,W. Hobson (1927), pp. 301, 318.in saltus of (x) at x, is py. We form also a second function (14) ya) = Fx) Of). ‘The funetion (x) is continuous while (sx) is only continuous from the right, both functions are non-decreasing and satisfy the relations O(-0)=K-0)=0, O+o)=a 51, W+o)=b 51. therefore possible to define two distribution functions by normali (3) and (x). "The functions Fa(x) = (L/ay)®(x) and Fe(x) = (1/5)Y(2) are both distribution functions; F(x) is a step function while F(x) is continuous for all x, According to (1.1.4) we have (11.5) F(a) = (8) +9) and therefore also (11.6) Fis) = Feat OFfx) (a 20, 620, mtb = 1). ‘The decomposition (1.1.5) and therefore also the decomposition (1.1.6) are unique. Suppose that there are two decompositions F(x) = (x) +4(x) = O1(x) + 4a(x). ‘Then (x) (3) = 4als) YO). "The right-hand side is a continuous function while the left-hand side isthe difference of two step functions. ‘Therefore both sides must vanish identically and we obtain Theorem 1.1.2. Bvery distribution function F(x) can be decomposed according to (1.1.6) F(x) = arFa(x) + 5FA2). Here Fax) and F(x) are both distribution functions. Fe(x) is continucus points. {for all x sohile all the points of increase of Fa(x) are discontinuit The coefficients a, and & satisfy the relations 0 < a = 1, 0 = atb= 1. F(x) is called the discontinuous (discrete) part, F(x) is called the continuous part of F(x). We note an immediate consequence of ‘Theorem 1.1.1; There exists an enumerable set D such that {, dFa(x) = 1. ‘The integral is here and in the following a Lebesgue-Stieltjes integral. ‘The function F(x) is continuous but has not necessarily a derivative at all points; however every distribution function has a derivative almost everywhere (in the sense of Lebesgue measure). A further decomposi- tion can be obtained by means of more powerful tools of analysis.2 Applying Lebesgue’s decomposition theorem [see Lotve (1955)] one can show that it is possible to determine two distribution functions Fae(x) and F,(x) such that (LL7) Bax) = biFads)+ bok) where by 2 0, by 2 0, brtby = 1. "The function Fae(x) can be represented as the integral of its derivative, Fads) = fis Fad())4y. Moreover fxv dFa(x) = 0 if N is a set of Lebesgue measure zero, Such a distribution function is said to be absolutely continuous. ‘The distribution function F,(x) is a continuous function whose derivative is almost everywhere equal to zero; moreover there exists a set V of Lebesgue measure zero such that xv dFi(x) = 1. A distribution function with this property is called a singular dist bution, Combining (1.1.7) and (1.1.6) we obtain the following Theorem 11.3. Every distribution function F(x) can be decomposed sniquely according to (11.8) F(x) = arFalsx) + a2Facls) + an (x). Here Fas), Fadx), Fas) are three distribution functions. The points of increase of Pq(x) are all discontinuity points. The functions Fad(x) and F(x) are both continuous; however Fads) is absolutely continuous while PAs) is singular. The coefficients a, ag, ag satisfy the relations ay = 0, ay = 0, a3 © 0, a+ a2+ay = 1 ‘The distribution functions Fe(x), Fad(), Fa(x) arc called the dis- crete, the absolutely continuous, the ‘singular parts respectively of F(a). ‘A distribution function is said to be pure if one of the coefficients in the representation (1.1.8) equals one, For pure distributions we will use the expressions discrete distribution if a = 1, absolutely con- tinuous distribution if a2 = 1, and singular distribution if as = 1. In applications one encounters almost inevitably cither discrete or ab- solutely continuous distributions. Singular distributions are interesting from a theoretical viewpoint but hardly ever occur in practical work ‘This is the reason why statistical texts frequently refer to absolutely continuous distributions as continuous distributions and seemingly ignore the existence of singular distributions. Let F(x) be an absolutely continuous distribution function. ‘Then (119) Rw) = f Pay where the integral is supposed to bea Lebesgue integral. The derivative p(s) = P°(s) is called the frequency function” ofthe distribution Fx). (9 ‘The expression “probability density function" ot “density function” is often used for “frequency fuseion’ ° =1B From the definition of a distribution function it follows that every function p(x) which satisfies the conditions pia) 20 for all x, (1.1.10) f ” pfa)de = 1 is the frequency function of an absolutely continuous distribution fune- tion which is given by (1.1.9). 1.2 Examples of distribution functions In this section we list a number of important distributions which occur frequently in practical work. For the sake of brevity we re- frain from describing mathematical models which lead to these dis- tributions, (D), Discrete Distributions. ‘The simplest discrete distribution function is, a function which has a single saltus of magnitude 1 at the origin. We denote this function by 0 fore <0 (20 = fore zo. Let £ be an arbitrary fixed real number; the funetion «(x —¢) is then a distribution function. It has a single discontinuity point at x and the magnitude of the saltus at this point is 1. The distributions (x =£) are called degenerate (or improper) distributions. "The most general purely diserete distribution function is determined by the location of its discontinuity points and by the magnitude of the corresponding jumps. Let {€)} be a sequence which contains all dis- continuity points of F(x) and let p; be the saltus at the point €. ‘The corresponding distribution function is then given by (122) Ms) = Breve Here the é) are real numbers and the py satisfy the relations (123) p20 Eppa In Table 1 we present some discrete distribution functions. ‘The Pascal distribution as well as the geometric distribution are par~ ticular cases of the negative binomial distribution. The Pascal dis- tribution is obtained if r is an integer, the geometric distribution if rel. In the examples given in ‘Table 1 the discontinuity points are conse~ cutive non-negative integers.Conditions on ‘parameters rn positive integer O 1N, M, m positive integers NoM>n Oepo Tay or move severally Laplace Distribution Normal Divibution | (vith degrees of Treedomn) Cauchy Distribution | iif Fx] ormore generally | | — | Braden? Dibation | Gamma Distibution | F eta Distribution yer ° otherwise ‘Two particular cases of the gamma distribution are of interest: For A= 1 one obtains the exponential distribution while the choice 2 = n/2 (n integer) yields the chi-square distribution with s of freedom which is of great importance in_ mathematical statistica. We note also that the Cauchy Distribution is a particular case (= 1) of Student's Distribution, (IID), Singular Distributions. We give only one cxample of a singular distribution; this example is closely related to Cantor's ternary set defined over the closed unit interval V = (0, 1]. ‘The Cantor set 7 is constructed by means of a step by step procedure. In the first step we16 remove from V the open interval (4, 4). In the second step two open intervals (4, 3) and (f, 3) are removed from V. From each of the remain- ing 4 closed intervals the middle (open) interval of length (4) is deleted in the third step and this process is continued indefinitely. Let each of the (=I) numbers cy, ey, ., ce-1 assume either the value 0 or the value 15 we denote by Ase,2e,..2e the open interval with initial point 21 Bote and terminal point S22 Dal tae Our procedure consists then in removing in the kth step the 21 open intervals Aze.2,.. 9. of length (})*. The Cantor set Tis obtained from the closed interval ” by removing a denumerable number of open intervals, It is easily seen that the points of 7’ can be characterized in the following manner. Write each number x, 0 5 x S 1, in the triadic system x = ay/3 +-a3/32+...+ dn)3"-+...; the set consists of all num- bers x which can be written in at least one way in the form of a triadic fraction whose digits are only zeros or twos. ‘The points of T’ have therefore triadic expansions 2ci/3 + 2ca/32-+...-+ 2en/3* + Whete C1, €2, son Gy = #8 8 sequence of “O” and “1” We introduce the function dif f 0 an absolutely continuous distribution function. An elemen- tary computation shows that ax = Bk = m|(m —R) if k < m while the moments of order equal or greater than m do not exist. ‘As another example we mention Student's Distribution. ‘This distribution has moments up to order (n—1)- For n = I this shows that the Cauchy distribution does not possess any moments. Itis of interest to know whether a given sequence of real numbers can be the sequence of the moments of a distribution. ‘The discussion of this difficult problem is beyond the scope of this monograph, we refer the20 reader to the book of J. A. Shohat-J. D. ‘Tamarkin (1943), However, re remark that a motnent sequence does not necessarily determine @ distribution function uniquely. Tt may happen that two different distributions have the same set of moments. As an illustration we give the following example. ‘The function Cexpl—x" cosun] if x >0 (1.4.2) pala) 0 if x<0 where o = meosun Fie) and 0 < w < fis a frequency function, ‘The moments of the cor- responding distribution are NI coun) (m= 01, (143) ay = Let _ {CLL +sin(e* sin pn] expl—x* cosa] if.e > 0 (49) RO) = Ig ifx <0. It is known [see Pélya-Szegé (1925), vol. 1, pp. 114 and 286] that for O oo. ‘Then the moments of any order s < k exist. For the meaning of the symbol O we refer the reader to Appendix A. A proof of theorem 1.4.3 may be found in Cramér (1946). ‘We give several inequalities for absolute moments which supplement2 the relations (1.4.1). Suppose that the kth moment of a distribution exists, then (14.6) Bia S PreBe “This inequality follows almost immediately from Schwar2’s inequality [see appendix B or Cramér (1946)]. By elementary operations we obtain finally from (1.4.6) the inequalities (47) Bury s Bele (1.4.8) Bi S Ba S Bat S .. S Ba!*. For a rather wide class of discrete distributions one can obtain recurrence relations for the moments. This class was discussed by Noack (1950) and contains many important distribution functions such as the binomial, Poisson and negative binomial distributions. Recurrence formulae for the moments of singular distributions (especially for the distribution function described in section 1.2) were given by G. C. Evans (1957). Finally we list in ‘Table 3 formulae for the moments of a few common absolutely continuous distributions. ne of Distribution ‘Rectangular Distribution ‘witha =O ian a Exponential Disirotion | an srbaton Distribution | Normnal Distribution with w=0,2=1 ieee afin Senior Gael Rea) Sadent’s Distribution (degrees of freedom) cay = 0, am even) dee aR(o— I) Te 1)/21 eT Pre2. PRELIMINARY STUDY OF CHARACTERISTIC FUNCTIONS 2.1 Elementary properties of characteristic functions In the preceding chapter we denoted distribution funetions by capital letters, as F(x), and the characteristic function of F(x) by the correspond- ing small letter, as/(¢). We adhere to this notation throughout this mono- graph; if subscripts are used on the symbol for a distribution function then the same subscript is attached to its characteristic function. We defined the characteristic function /(¢) of a distribution function F(a) by (13.6) 9 = { ate d(x). ‘Lhe properties of characteristic functions, stated in theorem 2.1.1, follow immediately from this formula. Theorem 2A.1. Let F(x) be a distribution function with characteristic function f(Q). Then @ fO=1 Gi) [fl st Gi) 1-9 =F ‘We use here the horizontal bar atop of f(t) to denote the complex conjugate of f(t). Theorem 2.12. Beery characteristic function is uniformly continuous on ici ee Ler —f00) sf Jeet] anes = 2 f (singohy arts) so that | Fle+h) 09) We note that the right side of this inequality is independent of t; it is possible to make the first and the third integral on the right side arbi- trarily small by choosing A > 0, B > 0 sufficiently large. Moreover, the second integral on the right side can be made arbitrarily small by selecting ft > 0 sufficiently small, so that the statement is proved. 223 Let F(a) be a distribution function and let @ and b be two real num- bers and suppose that a > 0. ‘Then 241) Ge) = (=) is also distribution function, We say that two distribution functions F and G belong to the same ype if they are connected by relation (2.1.1) (with a > 0). More generally, we consider a distribution function F(s) and two real numbers a and b and suppose only that a ¥ 0. We define * a ) if a@>0 @ Qa) GQ) = : LA(-0) if aco. “Then G(s) isalsoa distribution function, under the integral defining g(t) shows Q12) ai) = f(a’. For a= --1 and b= 0 we sce that g(t) = (1) is a characteristic function whenever f(t) is a characteristic function. Tet ay oy dy be m real numbers such that imple change of the variable and let F(x), «., Fa(x) be m distribution functions. ‘Then Gta) = 3 ayh(s) is also a distribution function; the corresponding characteristic function is = 3 aiff. Theorem 2.1.3. Suppose that the real numbers ar, .., dy satisfy the conditions and that fit), Jot) are characteristic functions. Then = 3 aff is also a characteristic function.Ey ‘Asa particular case we obtain the following corollary. Corollary to Theorem 2.1.3. Let ft) be a characteristic function; then FU aswell as Re ft) are characteristic functions. Here Refi) = aO+TO = | costars) is the real part of f(t). ae We showed in the preceding chapter that any purely discrete distri- bution can be written in the form (122) Rs) = Bayee6) where the £j are real while the py satisfy the relations (1.23) py 20, My "The Lebesgue-Stieltjes integral with respect to the distribution func- tion (1.2.2) reduces to a sum, so that the characteristic function f(t) of F(x) becomes 2.13) f= Eye If, in particular, F(x) is a lattice distribution then we can write (214) fy = atid where a and dare real numbers. ‘The characteristic function of a lattice distribution has therefore the form B15) f= te S py, where the pj satisfy (1.2.3), It is immediately seen that iW) ss [dierarai oe 1, Every lattice distribution F(x) has therefore the following property: there exists a real number fo # 0 such that the modulus of its charac~ teristic function f(t) assumes the value I for ¢ = to, We show next that this property characterizes lattice distributions, Suppose that the characteristic function f{¢) of a distribution function F(x) has this pro- . We assume therefore that there exists a fo 7 0 such that [f(to)| = 1. This means that for some real & (to) = e'%E or Ph erane = a25 tis then easily seen that F(x) satisfies the relation 2.16) f° [-costofe-]dF (x) = 0. Since the funetion 1 —cos f9(x—€) is continuous and non-negative, 2.1.6) can hold only if F(x) is a purely discontinuous distribution Whose discontinuity points are contained in the set of zeros of the function 1 ~cos fol =€). ‘The discontinuity points of F(x) have then necessarily the form §-+ 2n]ta) (+ integer) so that F(x) i lattice distri= bution, We have therefore the following result. Theorem 2.1.4, A characteristic function f(t) is the characteristic function of a lattice distribution if, and’ only if, there exists a real to % 0 such that | f(ta)| = 1. ‘Theorem 2.14 implies that | f()| < 1 almost everywhere. Suppose now that /() is the characteristic function of a non-degenerate distri- bution and that a < 0, then [/(#)]* cannot be a characteristic function. ‘This remark leads to the following corollary. : Corollary 1 to Theorem 2.1.4. The only characteristic functions whose reciprocals are also characteristic functions belong to degenerate distri- hnutions.. We obtain easily from ‘Theorem 2.1.4 the following corollary: Corollary 2 to Theorem 2.14. If a characteristic function f(t) has the property that for two incommensurable real values t, and ty the relations S(3)| = |8)| = 1 hold then | fl) = 1. We conclude this section by listing in tabular form the characteristic functions of some of the distributions given in Tables 1 and 2. (See “Table ¢ on page 26.) Tn section 1.2 we constructed [see (II1)] a singular distribution func tion. It ean be shown that the characteristic function of this distribution is given by @17 sy = tim TT cos bi For details and further examples of characteristic functions of singular distributions we refer the reader to papers by B. Jessen-A. Wintner (1935), A. Wintner (1936), R. Kershner (1936), and the thesis of M. Girault (1954).26 Table 4 "Characteristic Functions — Name of Distribution Function F(a) | Charscteriatie Function 10 ~ fT ear oli for more generally Gamma Distribution : Teta Bets Distribution — Te) z Rorennrar 22 Lebesgue decomposition of characteristic functions ‘The decomposition theorem 1.1.3 induces immediately a decomposi- tion of the characteristic function, Every characteristic function f(t) ‘can be written in the form (2.21) f= arfalt)+arfadt)+aape() ), dy = O and ay +da-+ay = 1, Here fa(t), fae(t) and f(t) istic functions of (purely) discrete, absolutely nuous and singular distributions. For a pure distribution one of the coefficients a, dz, ay is equal to one while the other two are zero, We add a few remarks concerning characteristic functions of a pure type. (A) If a= 1 then f(0) is the characteristic function of a (purely) discrete distribution. It follows from (2.1.3) that f(t) is almost periodic! so that . Jim sup [/(0) (‘The conditions on the parameters afe stated in Table 2 for discrete dstibutions and in ‘Table 3 for absolutely continuous distributions, We denote here e4 hy expt}, © See H. Bohs (1932), (1947)2 Pan = 1 then belongs to an abeoltely continuous distribution, t follows from the Riemann-Lebesgue lemma [sec ‘Titchmarsh (1939) p. 403] that Him sO =0. (C), If ag = 1 then f(t) is the characteristic function of a singular distri- bation. In this ease f(#) does not necessarily go to zero as |t| tends to co ‘Thus L Jim sup [/(0) ‘may be any number between zero and one. In fact, examples are known where L assumes the value zero [Girault (1954)]. An example of a characteristic function of a singular distribution for which L = 1 was communicated to the author by A. Wintner. "The behaviour of the characteristic function at infinity permits there- fore some inference concerning its type. For instance, if Tim sup [f(] = 05 then /(¢) belongs to a continuous distribution, 2.3 Characteristic functions and moments ‘There is a close connection between characteristic functions and moments. In order to discuss this relation we introduce the following notation. Let hi(3) be an arbitrary function, we define its first (central) difference with respect to an increment ¢ by a AMG) = BG) = AOED—M—1) and the higher differences by SewhQ)= AHO), for k = 1, 2,.... It can be shown by induction that (23.1) shag = 3 ¢ oe (7) aot 20y In particular, for the function f(y) = e'*¥ we have (23.2) Ag ee we eau(etst—e-lan ws oi2¥ Disinat}. Theorem 2.3.1. Let ft) be the characteristic function of a distribution function F(x), and lei : Aixf) (ure (Lat 0) = fhcos nti flows fom a result 0F By Jessen. Winner (1988) {theorem 11} tht jt) belongs to a purely singular dstsbution. Movcoer, for integer ‘it is easily seen that 1—f(2ak!)= Of E Rin!) = ofl) as k-> 00 50 that Tmsup [fo = 128 bbe the 2kth (central) difference quotient of f(t) at the origin. Assume that anf (0) mi | epee Then the 2ith moment aoe of Fs) exists, as do all the moments of order 5 < 2k, Moreover the derieatves f¥(0) exist forall t and for + = 1, 2, von 2h and FX) = ef. stetedF(x) (6 = 1,2,.04.2h) so that a= #90), From the assumptions of the theorem it is seen that there exists a finite constant M such that Ade f 0) 233) timing A2ZO Bat opp It follows from (2.3.2) that Abef(Q) = je 02(2i sin xf?* AF). "The difference quotient at the origin is then (yaw 7 We see therefore from (2.3.3) that ann (2 (sinstyt M= limgot f° (A) ‘dF(x) and hence that MM & lim int cE (22 )"are = [ stkdF(x) for any finite @ and 6 It follows then that the 2kth moment ee = J sh dF(x) exists and that M > 29x. Let s be a positive integer such that s < 2k; then it follows from theorem 1.4.2 that the moments fs and a exist for s = 1, 2, su 2k. From the existence of the moments as (= 1, 2, w+» 2k) we see immediately that {2 x! e di(x) exists and converges absolutely and29 uniformly for all real t and s = 2k, Tt follows then from a well-known theorem [sce for instance Cramér (1946) pp. 67-68] that all derivatives, up to order 2k exist and are obtained by differentiating under the integral sign. ‘This proves theorem 2.3.1 If a characteristic function has a derivative of even order, then the conditions of theorem 2.3.1. are satisfied and we obtain Corollary 1 to Theorem 2.3.1. If the characteristic function of a distri- bution F(x) has a derivative of order k at t= 0, then all the moments of F\x) up 10 order k exist if k is even, respectively up to order k—1 if k is odd. ‘The following example, due to A. Zygmund, shows that the result cannot be improved. then It can be shown that /'(t) exists and is continuous for all values of t, in particular for t = 0, However, the first moment of F(x) is infinite. Corollary 2 to Theorem 2.3.1. If the moment as of order s of a distribution. function’ F(x) exists then the characteristic function f(t) of F(x) can be differentiated s times and FX) = vf sell dF(x), Corollary 2 follows immediately from the argument used in the last part of the proof of theorem 2.3.1. "Theorem 2.3.1 yields also another result: Theorem 2.3.2. Let ft) be the characteristic function of a distribution F(x) and assume that Jor an infinite sequence of even integers {2ne} My = lim inf | is finite (but not necessarily bounded) for le = 1,2, ... Then all moments a30 af the distribution function F(x) exist and f(t) can be differentiated for real ‘airy number of Himes, with fo) = i fe tells dR(x), Corollary 10 Theorem 2.3.2, Let f(t) be the characteristic function of F(x); if all the derivatives of f(t) exist at the origin then all ihe moments of Fx) exist. Itis worthwhile to note that a characteristic function may be nowhere het & "This is the characteristic function of the purely discrete distribution Sol no~ Sod Let us assume that the first moments of a distribution function F(x) exist and denote by ft) its characteristic function. Then f(?) has the Macl.aurin expansion *” where Ral?) LOH OI << 1). Te follows then from theorem 23.1 that $00 = Sayre aw te por (or the meaning of the symbol “‘o” see Appendix A). ‘This connection is precisely described in the following manner: Theorem 2.3.3. Let Ps) be a distribution function and assume. that the nth moment of M(x) exist. Then the characteristic function f(0) of P(s) admits the expansion ("The remainder term used bere is a modification of Lagrange’s form for the semainderan (234) y= 143 fio soe) as 120. Conversely, suppose that the characteristic function f() of a distribution F(x) has én expansion (2.34). Then the distribution function F(x) has ‘moments up to the order nif nis even, but only up to the order (n—1) if nis odd. Moreover cj = aylj! for j = 1, my nif n is even but only for j = 1, 2, wns (n=) if is odd. We have already established the first part of the theorem. To prove its second part we compute An! f(0) according to formula (2.3.1) using the assumption (2.3.4). ‘This yields an expansion in powers of £5 it is easily seen that the constant term of this expansion vanishes and one obtains (23.8) 880) = 3 verAyth rote) as t>0 where 0 if jn is odd Ay " ; (-98(" \(n—2ey fj +n is even. Zon) We can prove by induction that at oe iflss0. f(t) = 14 Sean bot) as t+ 0. The argument just used proves therefore that the moment of order (u=1) of F(x) exists, The last part of the theorem follows immediately from (2.3.7).32, ‘To illustrate the situation described by theorem 2.3.3 we give an example, due to A. Wintner (1947). Tt is easily seen that the funetion 0 if [x] <2 (238) pey=| log is frequency function, provided that C is determined so as to make (2, p(s) de = 1, From if |x| > 2 4 de f log log A—log log? 2 wloge it follows that the distribution determined by (2.3.8) does not have any moments. ‘The characteristic function f{¢) of p(x) is given by the integral f09~ 20 | Si oe costs, > 1—cost 2 loge bai we loge so that 1 ~f(0) is a real, non-negative and even function of t. For any real = one has 0 = 1—cos = = Min (2, 2°) so that [1 —/(4)] has as a majorant a constant multiple of Me de p> dx = O(-tHlogt) = oft) as t +0. , i toes 2A ati Cee ae 1+o(0) admits an expansion of the form (2.34) with 0, even though the first moment does not exist. If the moments am of F(x) exist for all orders n and if lim sup (|x is finite, then the characteristic function /(0) of F(x) is regular at the ori- gin and has the power series expansion f= me «ni mi and p = 1/L is the radius of convergence of this series. Tis possible to define symmetrie moments 239) (ee = tm f * ck any).33 Symmetric moments may exist for distributions which do not possess ‘moments; the connection between the existence of symmetric moments and symmetric Ath derivatives hip LO ae of the characteristic function was investigated by A. Zygmund (1947). We mention here only the simplest case: Theorem 2.3.4. Suppose that the characteristic function f(t) of a distri- ‘ution function F(x) satisfies the ‘smoothness condition” Bat f(0) = ot) ast -r0, then a necessary and sufficient condition for the existence of f'(0) is the existence of the symmetric moment of order one, and (se) = lim fre = -if'(0). Zygmund’s “smoothness condition’ is expressed in terms of the characteristic functions. E. J. G. Pitman (1956) replaced it by a condi- tion on the distribution function. 24 The second characteristic We have shown that every characteristic function /(t) is continuous and that /(0) = 1. "Therefore there exists a neighbourhood of the origin in which f(t) is different from zero, let |¢| < A be this neighbourhood. "The function $(¢) = log f(¢) can be defined uniquely for |t| (so large that the half open interval (—m, m) contains J. We define then fin(x) as a continuous periodic 3836 function of period 2m such that fim(x) = g(x) for —m< x =m. Then G.1.3) holds for the function fin(x). Since f(x) is a function of bounded Variation it is possible to choose m so large that the variation of A(x) for |x| % m becomes arbitrarily small; the integral J, f(x) B(x) approaches therefore f%, g(x) d8(x) as m tends to infinity. Hience Fame = f atare =o for every continuous function which vanishes outside a fixed interval J. Tt follows then easily from the uniform boundedness of g(x) that frasaaey =o provided that a and b are continuity points of f(x) and that g(x) is Leela id et rn arin a Sertant on fet ofits continuity points so that F(x) and F(x) must agree in all continuity points and are therefore identical. ‘We would emphasize here that two characteristic functions f(t) and fo(t) must agree for all values of t in order to assure that the cor~ responding distribution functions Fi(x) and F(x) should be identical. "This requirement can only be weakened in a trivial way: one could suppose that the functions fi(?) and fa(t) agree for t-values which form a set which is dense on the positive real axis. It follows then from theorem 2.1.1 (condition ii) that they must agree for a set dense on the whole real axis, and one can conclude then from theorem 2.1.2 that fi(t) = fe(0). Agreement over a finite interval is, in general, not sufficient for the identity of the corresponding distribution functions. In fact, it is not difficult to construct a pair of characteristic funetions which belong to different distributions and which agree over a finite interval. It is also possible to show that a pair of different characteristic functions can agree everywhere with the exception of two symmetrically located intervals. We will give these examples in section 4.3. Let F(x) be an arbitrary distribution function. It is then easily seen that the function 1—F(—x—0) is also a distribution function. ‘This function is called the conjugate distribution of F(x) and is denoted by G18) Ae) = 1-R—s-0), Let f(t) be the characteristic function of the distribution F(x); an elementary computation shows that the characteristic function of the conjugate distribution F(x) is J ee ds) = f(—) = FO. A distribution function is said to be symmetric if it is equal to its conjugate. 6.1.6)37 ‘The following characterization of symmetric distributions is easily established. Theorem 3.1.2. A distribution function is symmetric if, and only if, its characteristic function is real and even. ‘The necessity of the condition follows from (3.1.6) while the suffi- ciency is a consequence of the uniqueness theorem. Moreover, we see from (3.1.6) that 617 f= [" costars, ‘This formula can be used to establish the following property of sym- metric distributions. Theorem 3.1.3. Let F(x) be a symmetric distribution with characteristic function f(t) and suppose that the moments a (j= 1, 2, +» 2) of F(x) eT : cof faim F292) = (=1)1 fe seostedFx) (jf sinted(x) (Gj ) asses Be ‘The theorem is a consequence of formula (3.1.7) and the Corollary to ‘Theorem 2 Corollary to Theorem 3.1.3. Let F(x) be a symmetric distribution with characteristic function f(t) and suppose that the moments 2 (j = 1, 5 2) of Fx) exist. Then @ ar 50 G = 12k) oo GH) gO = ty) é tg. ea : Gi) fig POON) aay tin F907 Part = ae (= Levk-D. ‘The corollary follows immediately from ‘Theorem 3. 3.2 Inversion formulae ‘Theorem 3.1.1 of the preceding section establishes a one-to-one correspondence between characteristic functions and_ distribution functions. However, theorem 3.1.1 does not give a method for the deter- mination of the distribution function belonging to a given characteristic38 function. ‘The theorems discussed in the present section deal with this problem, Theorem 3,2. (The Inversion Theorem). Let f(t) be the characteristic ‘function of the distribution function F(x). Then 1 ph toeitn 21) Fla+h)—Fa) = tim = [wa g(ayar Be Jog ae provided that a and a+h (with h > 0) are continuity points of Fs). For the proof of the inversion theorem we need the following well- known lemma, Lemna 3.2.1. The integral §: (sin y)fy dy is bounded for x > 0 and ‘approaches x2 as x tends to infinity. "The first part of the lemma ise integration into segments of length 7 by contour integration. *) Tet ily proven by dividing the range of 3 the second statement is obtained eae alo y aJ Se nan lites From this definition and from lemma 3.2.1 it is seen that A(h, 7) is canst cba oe M=hy T) = =D) while } if hoo (22) limMi,7)=) 0 if A=0. } if h<0 We introduce now the integral 1p? etta—e-tuarn Tee We substitute here for f(t) = {% et AF{s) and note thatthe absolute value of the integrand does not exceed fh. Hence the order of integra- tions may be reversed and one obtains seo BE [ff elon We replace in the inner integral the exponentials by trigonometric + Sce Titchmarsh (1937), setion 3.122,functions and obtain G23) Jp~ ale | [Stead sine eM aidan dalle 7 We write ae, T) = Mea, T)-Ne—a-h, 7) and conclude from the boundedness of A(x, 7’) and from (3.2.2) that ag(é, T)is also bounded for all x and all and that, 0 if xath Moreover, (3.23) can be written as 625) Je= J” ats MAR. Let ¢ > 0 be an arbitrary positive number. We assumed that @ and +h are continuity points of F(x) and we have shown that g(x, 7) is bounded. ‘Therefore itis possible to select a 8 > 0 which is so small that the three inequalities ty [J se.2arI9] <« me cy fees rare 5 « (is) fi, aR Pr ar) <« hold simultaneously. Moreover, we conclude from (3.24) that 7 can be chosen so large that the relations (uy) [eas T)ARx) oe ars) < and (Is) iE * ess TAF) + fe Ae TAM) Se40 are both satisfied. We decompose the range of integration of the integral (3.2.5) into the five intervals ( —c0, a—8), (a3, a+8) (a+3, a+h—8), (a+h—8, a+h+6) and (a+h+8, 00) and obtain, using (11), (Ia) and (ls), a ay [ye [7 ae r)aMt 5 56 ‘We see then from (1g), (14) and (Ig) that | yn— f°" ane s5e if only 7 is sufficiently large. ‘This is equivalent to (3.2.1) so that the proof of the inversion theorem is completed. We remark that the uniqueness theorem could easily have been obtained from the inversion theorem. Such an approach would have shortened the presentation, however it is of some methodological interest to distinguish between these two theorems. ‘The inversion formula can be written in a more symmetrical form by putting a = x—8and h = 26. In this way we obtain, (3.2.6) F(e+8)—F(e—8) = lim tft Sint nesta) dt Progr t rovided that x—8 and x+6 are continuity points of F(x). The last ermula can also be written as F(x+8)—Fe—8) 1p? sinsd - a os lim ete f(t)adt. 8 toate me Let us now assume that j(¢)is absolutely integrable over (—c0, +0), the integrand is then dominated by the absolutely integrable function ‘A(d) and converges to e'"* f(t) as 8 tends to zero. Hence one may go to the limit under the integral sign and one sees that F(x) exists for all x and one obtains the following result. Theorem 3.2.2. If a characteristic function f(0) is absolutely integrable over (—00, +90) then the corresponding distribution function F(x) is absolutely continuous and the formula 1p pey= Fay= 5] ewesiae expresses its density function p(x) in terms of the characteristic function. The density function p(x) is continuous. We have already proved the absolute continuity of the distribution4 corresponding to f{¢) and must still show that p(x) is continuous. We see easily that Lace) —p00) 3 Ff, snc2| oar El +E i; aga iMOHPALFOa We choose A so large that the second integral becomes arbitrarily small and can then make the first integral as small as we wish by selecting h sufficiently small. This completes the proof of the theorem. ‘We note that this inversion formula is here not derived for all ab- solutely continuous distributions but only for absolutely continuous distribution functions which have absolutely integrable characteristic functions. We will give later [p. 72] examples of characteristic functions which belong to absolutely continuous distributions but which are not absolutely integrable. However, we will see that under certain conditions the inversion formula is still valid. Let again f(¢) be the characteristic function of an arbitrary distri- bution F(x). We consider the integral i |» We write in (3.2.7) /(0) = J2, et” dF(2) and sce easily that the order of the two integrations may be exchanged so that ie fame aii B27) Irs ete f(a)at. f Ds won Ty Here h is a positive number to be chosen later. Let « > 0 be an arbitrary positive number and denote the saltus of F(@) at the point x by ps, that is pz = F(x) ~F(x —0). It is then possible to choose a value fio > 0 which is so small that the inequality g28) |f ary+s) | Se Jan holds. "The integral converges to zero as T tends to infinity, since the integrand is dominated.2 by +1 and converges to zero as T’-> co. We can therefore choose T’s0 large that 6.29) if 107) a my+a) woke Ty Let now fy be a number such that (3.2.10) iy > n> 0. It follows then from (3.2.8) that Os pemes f ayy +x) wish Son, POT) 5 Poe so that J sea FOF -Pal 5 « (3.211) while os Srcocn aFO+ts) 56 We are still free to choose /, subject only to the restriction (3.2.10). We select Ja so small that lees sin Ty Ty for |y| < fy. According to the law of the mean there exists real (10) < 1) such that = (1-0) { ayF(x+y) where 0 $ 0, 5 1, From this we sce immediately that sin) Gog, eee Pe sf dull type tie wan Dy vem with 0S Op S 1. Hence we conclude from (3.2.11) that, jaz) | f sin TY 4 retype S2e wien Ty fee |si sin. since|°%22| 1, we se from the second formula (3.2.11) that My8 sin Ty = eros) s ayy +3) oe ee ‘We combine the last inequality with (3.2.9) and (3.2.12) and obtain (3.213) [r—pel 5 4e if T is sufficiently large. Thus we have proved the following statement Theorem 3.2.3. Let f{d) be an arbitrary characteristic function. For ery real x te limit tim + [" etepinar Pex tine [mney exists and is equal to the salins of the distribution function of ft) at the point s. Corollary 1 10 Theorem 3.2.3. The characteristic function of a singular distribution cannot bean almost periodic function. Let f(t) be the characteristic function of a singular distribution. Since a singular distribution is everywhere continuous we see immediately from Theorem 3.23 that 1 opt 21) tim [ ereae for all real x. We give an indirect proof for the corollary and suppose that f() is almoct periodic. Formula (3.2.14) means then that all Fourier coefficients of f() vanish; we conclude then from the uniqueness theorem for almost periodic functions [I1. Bohr (1932), (1947) English edition, p. 60] that /(s) = 0 which contradicts the assumption that fle) is characteristic function. Corollary 2 to Theorem 3.2.3. A distribution function is purely diserete if, and only if, its characteristic function is almost periodic. ‘The necessity of the condition follows from our remarks in Section 2.2. To prove the sufficiency we note that an almost periodic charac- teristic function (2) satisfies necessarily the relation Tim sup [/(1 = in view of Corollary 1 to Theorem 3.2.3 (d) must then belong to a purely discrete distribution. 3.3 The Convolution Theorem fe consider next two distribution functions F(x) and x(x) and their characteristic funetions /;(¢) and fx(f) respectively. We form the function“4 G31) Fa) f F(a—)dFa(e) ‘The function Fi(s—2) is bounded so that the integral (3.3.1) exists. Moreover it is easily seen that F(2) is a distribution function; this follows from the fact that the necessary passages to the limit can be carried out under the integral sign. We wish to determine the characteristic function f(t) of F(2). Clearly, So = [~ ettare. We consider first the integral [ e!*" dIf(2) (where a and & are finite) and write its a limit of generalized Darboux sums. We usea sequence of subdivisions {2)'"7} of decreasing modulus om ym aaa cay exp(itej=) (Rapa) — Flay). In view of (3.3.1) this can be written as ° iE exp((st) dF(2) =i f° Desvltese—a)] x x [F(zjl 2) — Fy ay —x)] explits) dP). ees Ponany= [| [era] ten, We take the limit as a + —co and b > oo and obtain 33.2) f() = AC) AO. ‘Suppose conversely that a characteristic function /(t) is the product, (3.3.2) of two characteristic functions. According to the uniqueness theorem relation (3.3.1) holds between the corresponding distribution functions, Formula (3.3.1) defines an operation between distribution functions, it indicates how a new distribution F can be obtained from two given 19) See Cramée (1946), pp 62, 72.8 distributions F, and F. This operation is called convolution (sometimes composition or Faltung) and is called the convolution of Fy and Fy and is written as a symbolic product F=Fth. Itis seen from (3.3.2) that the convolution is a commutative and asso- Giative operation. ‘We summarize our results as Theorem 3.3.1 (Convolution Theorem). A distribution function F is the convolution of two distributions F; and Fe, that is Fa) = fe Fy(z—s) dPe{a) = i Ps—s) dP (3) = Fy" Fe ‘and only if, the corresponding char acteristic functions satisfy the relation fH) = AORO- Hence the genuine multiplication of the characteristic functions and the symbolic multiplication of the distribution functions correspond to each other uniquely. "he following corollaries follow almost immediately from the con- volution theorem. Corollary 1 to Theorem 3.3.1. The product of two characteristic functions is a characteristic function. Corollary 2 to Theorem 3.3.1. If f\Q) is a characteristic function, f(t)? is also @ characteristic function. ‘This follows from theorem 2.1.1 and formula (3.1.6). We mention some properties of the convolution operation which follow easily from its definition. Theorem 3.3.2. Let F = Fi*Fs be the convolution of t100 distributions F, and Fy, If one of the components of F is a continuous distribution, then the symbolic product is also a continuous distribution. If one of the’ com- ponents of F is absolutely continuous then F is also absolutely continuous. Remark 1. Let F = Fi*F, be the convolution of two distributions F, and F; and suppose that F is a discrete distribution. ‘Then both components f; and F; are also discrete distributions. Remark 2. However, it is not known at present whether it is possible to conclude from the assumptions (i) F = Fi*Fs, (ii) F is absolutely continuous, that at least one of the distributions F and F is absolutely continuous, Corollary 1 to Theorem 3.3.2. If ft) is the characteristic function of a continuous (respectively absolutely continuous) distribution then | f(t)?46 also belongs to a continuous (respectively absolutely continuous) distribution. Corollary 2 to Theorem 3.3.2. Let F(x) and Fax) be two absolutely continuous distribution functions and denote by px(s) and pax) their frequency functions. Let F = F\*F2 and jx) = F'(x) be the density of F(x); then 2) = | pile —x)pa(a) de. ‘We consider next the convolutions of two purely discrete distributions. Theorem 3.3.3. Let Fy(x) and Fels) be two purely discontinuous distri= butions with discontinuity points {E,} and {7,} respectively. Then F = F,*F, is also purely discontinuous and the discontinuity points of F are the points af the sequence {£,+7,) Moreover, let a, be the salts of Fy at &y ard by be the saltus of Fs at n,. and suppose that € is a discontinuity point of F. The saltus of F at € is then eee ‘The characteristic functions fi(t) and fe(#) of F(x) and Fa(s) respec tively are Ail) = Z ayexp(te) Fat) = Lhpexplitn,. ‘The characteristic funetion f(t) of F(x) is, according to the convolution theorem, B33) fO = fAOhO = LE abpexplitE,+a))- ‘Theorem 3.3.3 follows immediately from this formula. Suppose that Fi(x) and F(x) are two purely discrete distribution fanctions and that at least one of these has infinitely many discontinuity points. Tt follows then from (3.3.3) that F(x) = Fi(x)*Fa(s) has also infinitely many discontinuity points. ‘Assume next that each of the purely discrete distributions F(x) and F(a) has only finitely many discontinuity points, then (3.3.3) becomes B34 f= HOMO = FE abpesplig, +7] where m and m are the numbers of discontinuity points of F(x) and F(x) respectively. ‘The last formula indicates that ” = Fy"; has also a finite number, say N, of discontinuity points. Let us consider the set {E.+ny} (0 = Ty oy3u = 1, mm). ‘Then itis no restriction to assumere that the £1, 2 =» Gn and 73,2)» Nm are arranged in increasing order. ‘The set of discontinuity points of F(x) contains at most nm elements, while it is seen immediately that the n-+m—1 numbers +m, €2-m, Ean, Ext on En-+7m are distinct, We formulate this result in the following manner. Corollary to Theorem 3.3.3. Let F = Fy*F be the convolution of two purely discrete distributions Fy and Fe, The distribution function F(x) has a finite number of discontinuity points if, and only if, each of the functions Fy and P, has finitely many discontinuity points. Denote by ‘N, nm and m the number of discontinuity points of Fix), Fi(x) and Fal); then n-+-m—1 5 N's nm, Next let F(x) be an arbitrary distribution function. If F(x) has a discontinuity at the point «x = gj with saltus py then the conjugate distribution function F(x) of Fx) has a discontinuity at x = —& with the same saltus pj. According to theorem 3.3.3 the convolution F*F has the saltus Eppat the point x = 0. On the other hand one can determine the saltus of F*F at x = 0 from theorem 3.2.3 and'we obtain the following result. Theorem 3.3.4. Let F(x) be a distribution function and f(t) its charac- teristic function. Then 1? lim — (Olde = Ter dim sz [oa = Er where the py are the saltus of F(x) and there the summation on the right is to be taken over all discontinuity points of F(x). We conclude this section by stating the connection between the existence of the moments of a convolution and the moments of its components. Theorem 3.3.5, Let Fi and Pz be two distributions and suppose that the moments of order ke exist for F, as well as for Fy. The same is then true for F = Fi*F,, "This follows easily from the elementary inequality legit 5 26 alt iit 3.4 Limits of distribution functions In this section we study sequences of distribution functions and their limits and we introduce a specific definition for the convergence of such sequences. In order to motivate this definition we consider first two examples,48 Example 1, Let 0 if xeon a Seen Fis)=) 5 it -as dinates a (n = 1, 2, .«) be a sequence of rectangular distribution functions. ‘This Sequence converges for all x and lim F(x) = 4. ‘We note that the limiting function of this sequence of distribution fune- tions is not a distribution function. Example 2. Let roy eae i2)dy (w= 1,2, Ws) = if exp(—n?y2/2)dy (= 1,2...) be a sequence of normal distributions. It is easly seen that ao 0 if x<0 Jim Fy(s) = lim — f evtde ={h if x ae eae 1 if x>0. Xf ye look at the graphs of the functions Fs) we might expect in- tuitively that the sequence Fy(x) should converge to the degenerate distribution «(x). This agrees also with the fact that lim Fy(x) = 0 if x <0 and limFa(x) = 1 if x > 0. However, we have lim Fg(0) = $ while <(0) = 1. We observe therefore that it seems to be too restrictive to require that a sequence of distribution functions should converge at all points to a limiting distribution function. Example 2 suggests that exceptions should be permitted for the discontinuity points of the limiting distri- bution, Moreover, we see from example 1 that a sequence of distribution functions may converge at all points but that the limiting function is, not necessarily a distribution function. In view of the situation revealed by these two examples the following definitions seem to be appropriate. ‘A sequence of functions {/ta(2)} is said to converge weakly to a limiting fanction f(s) if ition h(s) ‘amda(s) = Ms)for all continuity points x of h(x). We write then Lim ii(s) = Me), that is we use the symbol “Lim” for weak convergence to distinguish it from “lim” used for ordinary convergence. Using this terminology we introduce the following definition. A sequence {Fq(2)} of distribution functions is called a convergent sequence if there exists a function F(x) such that Lim Fy(s) = Fs). We note (sce example 1) that the weak limit of a sequence of distri- bution functions is not necessarily a distribution function. However the (weak) limit is always a bounded and non-decreasing function. We are primarily interested in obtaining a necessary and sufficient condition for the weak convergence of a sequence of distribution fune- tions to a limiting distribution. In order to obtain this condition we need some results which are also of independent interest. ‘These will be given in the next section. 3.5 The theorems of Helly We first prove the following lemma. Lemma 3.5.1. Let {Fy(x)} be a sequence of non-decreasing functions of the real variable x and let D be a set which is dense on the real line. Suppose that the sequence {Fy(x)} converges to some function F(x) in all points of the set D, then Lim Fa(s) = Fe). Let x be an arbitrary point and choose two points x'eD, x”«D so that a” sx Sx” Then Fy(s’) S Fy(s) S Fa(2") hence Tim F(x’) < lim inf Fy(s) lim sup Fy(s) Slim Fy(s"). From the assumption of the lemma we conclude that Flw) S lim inf Fy(s) = lim sup Fats) = Fe"), Since D is dense on the real line we have F(x—0) ¢ lim inf Fy(x) = lim sup Fy(x) S F(e+0). From this relation one obtains immediately the lemma. Theorem 3.5.1. (Helly’s First Theorem). Every sequence {Fu(x)} of uniformly bounded non-decreasing functions contains a subsequence {Fy,(s)} which converges weakly to some non-decreasing hounded function F(x).50 The proof is carried by the standard diagonal method and uses the fact that the sct of rational numbers is enumerable and can be arranged in a sequence {r3}. We form first the sequence {Fy(r)}. This is a bounded sequence of real numbers and has therefore at least one accumulation point. Thus it is possible to select a convergent subsequence {Fi,9(71)}. Let lim Fiala) = O71). In the second step we consider the sequence of functions {Fi,9(%)}. We select again from the bounded sequence of real numbers {Fiya(r2)} a convergent subsequence {F2,q(/2)} and write Him Fan(rs) = (02). ‘The sequence of functions (F(x) isa subsequence of the original sequence {Fy(x)} which converges for x = ry and x = re. We continue this procedure and obtain a sequence of subsequences of {Fa(x)} Finals), Piaf), Fi,9(%) - Final), Fa,s(%)s F222), Fe,a(%), ony Fan(s) Fn.x(2), Fm 22), Fn, €2), "These sequences are selected in such a manner that each sequence is a subsequence of the preceding sequence; morcover the mth sequence Fm) Fm,2%), Fm,a(%)s oy Frn(%)s converges at the first m rational points 7, rz, ..., rm and we denote Him Falta) = (2), B= 1,2, 09m We form the diagonal sequence {F,n(x)} and conclude that Him Fan) = ©(02) for all rational arguments rg. ‘The functions {Fy,9(x)} are non-decreasing and uniformly bounded so that the function (72), defined for all rational values of the argument, is also bounded and non-decreasing. We introduce now Fx) = gl.b. (re) rox51 ‘The function F(a) is defined for all real x and agrees with ®(x) for rational values of the argument. ‘The function F(x) is bounded and non- decreasing and we see from lemma 3.5.1 that GSA) Lim Pr.a(s) = Fs). qed. "The limiting function F(x) is not necessarily right continuous; however it is always possible to change the values of F(x) at its dis- continuity points in such a manner that it becomes right continuous. Such a change obviously does not affect the validity of the relation G5.1). Theorem 3.5.2. (Helly's Second Theorem). Let fx) be a continuous function and assume that {Fi{x)} is a sequence of uniformly bounded, non-decreasing functions which converge weakly to some function F(x) in all points of a continuity interval (a, 6] of F(x), then : i: fim [i seearee) = '700)4R69. Since f(x) is continuous it is possible to construct a subdivision ee (35.2) [f()-se) S «for 4) Sa Sap Here « is an arbitrary positive constant. Moreover, the subdivision points can be selected so that they are all continuity points of F(x). Hence |Fr(x;)—F(x))| can be made arbitrarily small if k is sufficiently large. Let M = max,’ f(s) and choose K so large that (53) (Fae) Fpl S a We define a step function /.(x) in the interval [a, 6] by B54) FO) =f) for xy Sx < apn and see from (3.5.2) that | f(x) —f.(x)| S$ « Clearly we have for k2K. 33.5) If f(s) aF(x)— fis anc) | =| fireraren [separ] + +| [rear 10) aree|+ . . ' +| eo) args) [fy are))We oe ht em on te ght of 5) a ae if 1/0) AM = € f APs) = Ce. In the same manner we get an estimate for the last term in (3.5.5) | Pranee)— f'eyance) = Ge: the existence of the constant C; is assured by the assumption of uniform boundedness of the Fi{s) Finally i follows from (25.3) and (3.5.4) at | 0) aR0- . Js) dF] Xe a = [Zfeeateten)— Peal &, Kon Faless)— Fie) “The last three inequalities and (3.5.5) yield the estimate aso | ie f(a) dF) — ie Ss) dP (| = 5 Al4Qie = Ce if k2K. But this means that Jim [sear = [sep ar, which is the statement of the theorem. Corollary to Theorem 3.5.2. (Extension of Helly’s Second Theorem). Let (8) be continuous and bounded in the infinite interoal <0 < x < 0 ‘and let {Fx{x)} be a sequence of non-decreasing, uniformly bounded functions which converges weakly to some function F(X). Suppose that Him Fy{—o0) = P(—oo) and Jim Fi-+ 00) = F{-+00), then oe Pe Bn J sonar) = [" soyare53 To prove the corollary we consider the three expressions b= {fi porary f° sorares) A=|f- sean |" sear) [FP renariey- f° separ ’ , J where a <0 0 such that f(3)| = -M. Let ¢ > 0; a8 a consequence of the uniform boundedness of sequence {Fy(x)} it is possible to determine ja| and |p| so large that Ji = cand Js 5 c. From theorem 3.5.2 we see that there exists a K so large that Jo = Ce for k 2 K. Therefore : [ speaneey f° seoare = (C42) for ke K so that the corollary is proved. statement, analogous to the corollary, holds if the range of the inte- gration is a semi-infinite interval. 3.6 The Continuity Theorem In this scction we derive a necessary and sufficient condition for the weak convergence of a sequence of distribution functions to a limiting distribution. ‘The theorems of the preceding section will serve as tools; we need however one more lemma which we deduce next from the inversion formula and the convolution theorem. Lemma 3.6.1. Let F(x) be a distribution function with characteristic ‘function f(0), then i. ° 1p tcoskt freer J Bovey == f : ‘ ‘f(t e for any real positive h. We denote by R(x) the uniform distribution over the interval (a, +a); the characteristic function (see ‘Table 4) of R(x) is then inta54 Let F(x) be an arbitrary distribution funetion and consider the distri- bution H(x) = F(a)*R(x); clearly 1 ye poe (6.1) He) ~ 5 [me dy = 5, fo Rea while sinta ey = Fa, We apply (3.2.6) and obtain 1p? sinta Hehe) Hea) = i= [ me eefioeae Using (3.6.1) we see that 1p 1-cos2a i Treo 2a te Faye, 1 @ Tong * Ure +0)—F(e—e)]do = lim ° ‘The funetion (1~cos af)/¢®is absolutely integrable over (—00, + 00) so that the integral on the right converges absolutely. We write h = 2a and go to the limit so that 1-coshe s 1 ie [Fs +2)-Re—o)]du = = ertinf(jde. ao @ ‘We finally put x = 0 and transform the integral on the left and obtain the formula of the lemma. ‘We now proceed to the main theorem of this section. Theorem 3.6.1 (Continuity Theorem). Let {Fy(x)} be a sequence of distribution functions and denote by {u(0)} the sequence of the corresponding characteristic functions. The sequence {Fq(x)} converges weakly to a distribution function F(x) if, and only if, the sequence { fu(t)} converges for every t to a function f(t) which is continuous at t = 0. The limiting ‘Junction is then the characteristic function of F(x). ‘This theorem indicates that the one-to-one correspondence between distribution functions and characteristic functions is continuous. ‘The necessity of the condition follows immediately from the exten- sion of Helly’s second theorem (Corollary to Theorem 3.5.2). To prove the sufficiency we assume that the sequence /a(t) converges for all ¢ to a function f(t) which is continuous at t = 0. Let {Fy(x)} be the se- quence of distribution functions corresponding to the sequence {fa(t)} of characteristic functions. According to Helly’s first theorem we can select a subsequence {F,(x)} such that Lim F(x) = Fis)35 where F(x) is a non-decreasing and bounded function which is continu- ous to the right. Since the F;,(x) are distribution functions we conclude that the limiting function F(x) satisfies the inequality 0 F(x) <1. In order to show that F(x) is a distribution funetion we must only show that F(c0) —F( —o0) ‘We apply lemma 3.6.1 to the functions Fy,(x) and find that ‘* 0 1 pe 1—cosht i} Fr) dy— i Fro)ey = =| wi fut) at. ‘ - oa Its easily seen that the passage to the limit, k > oo, ean be carried out under the integral signs so that © 1—cosy 1 Paya 4” Aoyay =* f Ih) dy. 7 J Oe-7 Lr = 2] ome. ‘The expression on the left side of this equation tends to F(0) —F( —co) as ht > op, Since we assumed that j(¢) is continuous at ¢ = 0 we see that fim f(Uh) = f(0) = lim flO) = 1. Moreover, it i agnin permissible to carry out the passage to the limit under the integral sign s0 that 1-cosy f ToS ay, =~ # Foo) -F{-0) = Tt is well known? that “= 1—cosy 7 ay =" iF aaa? so that F(co) —F{ oo) = 1. ‘The limiting function F(x) of the sub- pe eae ae te neil Dalles just used applies to every convergent subsequence of {Fy(:x)}. It follows then from the uniqueness theorem that every convergent subsequence of {Fa(z)} converges weakly to the same limiting distribution F(x). "This means however that Lim Fu(s) = Fe). Corollary 1 to Theorem 3.6.1. If a sequence { fa(t)} of characteristic Junctions converges to a characteristic function f(t) then the convergence is uniform in every finite t-interoal [-T, T]. (©) We integrate by parts and see that fo aaa For the last integral see Titchmarsh (1937), setion 3.122. Boy, y56 Denote by Fy(x) and F(x) the distribution function of fa(t) and f(t) respectively. ‘Then (362) (4at)-F00| s | ean) f° etearen) + Fu) Pada) +1 FO) + F. Let « be an arbitrary positive number and select for a and b two con- tinuity points of F(3) and take a and 8 so large that 1-F(b)+Fla) < «. Lim Fa(s) = F(x) Since wwe have, for sufficiently large values of n, 1-Fy(b)+ Fala) S 1-F(b)+ Fla) +e < 2c ‘The inequality (3.6.2) can then be written as o re , .63) [fal)—FO) = [fe ettedFa(x)— [ett da) +3. ‘We wish to estimate the difference of the two integrals on the right side of (3.6.3) for values of ¢ from the interval [ ‘odo this we sub- divide the interval (a, 5] by means of the subdivision points a= mem 0 is the function fd) = 1. Let f(t) be a characteristic function and assume that f(¢) as 0. It follows then from Theorem 2.3.3 that 2; ay = [%, 22 dF(x) [where F(x) isthe distribution function corresponding to /(0] itis seen that F(x) must be constant over every interval which does not contain the point + = 0; that is F(x) = <(x) and f(t) = 1. ‘The function f(t) = e-* satisfies the conditions of the’ ‘Theorems 211 and 2.1.2 but et’ = 14+0(f%. Therefore we see from ‘Theorem 4.1.1 that e~t' is not a characteristic function. ‘This example shows that the necessary conditions stated in ‘Theorems 2.1.1 and 2.1.2 are not sufficient. Corollary to Theorem 4.1.1. Let w(t) = at) as t +0 and suppose that 10 (—t) = nt). Then the only characteristic function of the form fit) = 14d) +0(2), as t > 0, isthe function f(t) = 1. We know [Corollary 2 to ‘Theorem 3.3.1] that f(d) f(—2) is a charac- teristic function. Under the assumptions of the Corollary f(t) f(—t) = [1+ (2) +o(t%)][1 —u(t)+0(t2)] = 1 +0(¢%), hence according to ‘'heo- rem 41.1 | fO)2 = A) f(—1) = 1 or f(t) = e'@t (a real). ‘Therefore Ait) = Veiat late? +o(t2) and this has the form 1+ 10(t) + o(22) (with, ‘2(t) = o(8)) only if a = 0 so that f(2) = 1. 590 We discuss next an inequality which every characteristic function must satisfy and which can therefore also be regarded. as a necessary condition for characteristic functions. ‘This inequality is also of some independent interest. Let F(x) be a distribution function and f(t its characteristic function. ‘Then Gy Rett=s(9)= [(1ome}aR It follows from the elementary relation & tie L-costx = 2sint= > 2sin?— cost costy = 2sint> 2 os and (4.1.1) that (41.2) Re{i—f(9] 2 #Re[1-F(29]. By induction we obtain casily the following condition. sintts = {1 cos 2¢x) Theorem 4.1.2, Let n be a non-negative integer, the inequality 1 RefL=f(O] 2 G_Rell -f2%)] is satisfied for every characteristic function. We consider now a characteristic function which has the property that 413) [ols al if Qe. From Theorem 4.1.2, applied to the function | f(¢)|® we see that 1 1-I/OF 2 GU fF Let ¢ be a fixed value such that t] < B, and choose n so that B B Ss\ti< 2 SS Qt ‘Then—according to (4.1.3)— dee Quy < AP while > — sano Pe Hence we obtain from ‘Theorem 4.1.2 the inequality @ —\nor > 2a WP > geft-49 A\ slight modification yields the following corollary.a Corollary to Theorem 4.1.2. Let f(t) be a characteristic function rohich satisfies (4.1.3) then ea?) Wn) <1-S ‘for all t such that 0 < |t| < B. We discuss next another important property of characteristic functions. Theorem 4.1.3, Let f(t) be a characteristic function and let N be an arbitrary positive integer. Denote by sak & ry tebe te where ty, tay v5 ty are arbitrary real and &1, £2, « €xy are arbitrary complex numbers. Then S is real and non-negative for any choice of N, 1, fay vary tty Ey 85 vary Eve Let F(x) be the distribution function which corresponds to-f(t). ‘Then xX 8-22 oe |” expties—tos) arte) - [7 (See) (Shem) aco it oe | DI & owl F(x). "The last expression is real and non-negative. 42 Necessary and sufficient conditions ‘The property of characteristic functions which is described by the last theorem suggests the introduction of a concept which is useful in formulating necessary and sufficient conditions for characteristic functions. A complex valued function f(t) of the real variable t is said to be non-negative definite for —c"< t< +00 if the following two con- ditions are satisfied: @ f{d) is continuous; (i) for any positive integer NV and any real ft, .., ty and any complex £15 + €2y the sumco 2B su-wosie is real and non-negative. We establish next a few properties of non-negative definite functions Theorem 4.2.1. Let f(t) be non-negative definite. Then (@) KO) isreat and f0) = 0 O f-)=) © LKOl <0) Proof of (a). Put N = 1, t: = 0, £1 = 1 then (ii) implies (a). Proof of (8). Put N = 2, ty = 0, t = t and choose arbitrary € and ‘Then S = (O[|él?+ léel*]+/( Df 8+/O EK. It follows therefore from (a) and (ii) that f( ~2)£:£2+ f(e)éebi is real for any & and &. We write f(—#) = 01 +181, ft) = a + iBa, f:f2 = y +18. Then (G1 +181) y +18) + (ca + HBaNy—18) is real, so that (B; +f2)y-+(ai —a2)8 = 0 for any y and 8. ‘This is only possible if fi: +2 = 0 and x a9 = 0, so that (b) is satisfied. ‘The property (8) is sometimes expressed by stating that the function At) is “Hermitian”. Proof of (c). We put again N = 2, t) = 0, t2 = t but choose & = f(t), £2 = =| (0)|. Using (ii) and (6) we see that S = 2f(O)/F(O?-21f(Of 2 0. In case | f(f)| > 0 this inequality yields immediately (c). S (@)] (0) & 0, relation (c) holds in a trivial way if |f( = ‘We can now formulate a criterion for characteristic functions. ce [by Theorem 42.2. (Bochner's Theorem). A complex valued function of a real variable t isa characteristic function if, and only if, fi) (0) is non-negative definite GAO) = 1 ‘The necessity of the conditions is established by ‘Theorems 2.1.1 and 4.1.3 so that we must prove only their sufficiency. We assume there- fore that f(t) is a non-negative definite function and choose positive integers x and N and a real number x and put t) = j/m, j= e~Ut, It follows then from the definition of non-negative definiteness that “ES SAE N 3 oy Josrl-1j-¥5] z063 for all x. The difference jk = r occurs in N —[r| terms of this sum; here r is an integer between —N+1 and N—1. We collect these terms Leper i () way (BE We introduce now the funetion (42.1) BI (0 Ey ramerws 2 0 ‘Therefore ° for x < —2 1p (423) Fea) zl Body for a sx 0 we have F'(—n—e) = 0 while Fy(r+e) = 1. We see from (4.2.4) that Flo) = fim (1 = ne) fen = lim J evsares ) Applying Helly's second theorem we obtain (425) fly = [7 evearmsy“o for all integers s. We introduce the distribution functions Fala) = FO). “The points of increase of the distribution function a(x) are all located in the interval (—nn, +nn); the characteristic function of Fg(x) is, fil) = [eran = f “ etn aP1y) It follows then from (4.2.5) that for any integer (4.2.6) fallen) = (Rin). Let t be an arbitrary real number. It is then possible to determine a sequence k = R(t, n) such that (42.7) 0 5 t-(fn) < (In). We write 0 = ¢—(kjn), then 0 5 0 < (1Jn) and (pte fcnny = | tering) aF()| s eC lee 1]dR,0). We apply Schwvar's inequality (see Appendix B) tothe las integral nlf] = (2 [" (1 ondary =p f ed cos ns) Fm(=)p 2 Since n <1 we have 1—cos nz < 1—cos # for |z| S 7; hence \falt)—FalRjn)] = [2 f © (leon same ~ QU —Refa(1/n)e = (2[1— Ref (t)n)}} 2. By assumption f(0) is a continuous funetion such that f{0) = 1, and the right-hand member of this inequality can be made arbitrarily small by selecting m sufficiently large; this means that (4.2.8) Tim |falt)—fal/n)l = 0. Moreover, we see from (4.2.6) and (4.2.7) that (829) JW) = Tem 40h) = lim fn. Since His fu(O) = Ti Lfult)—ful@elm)] + fal in)wwe see from (4.2.8) and (4.2.9) that forall ¢ Jim fat) = 0 ‘The continuous function f{@) is therefore the limit of the sequence {falt)} of characteristic functions and is therefore [Corollary 2 to ‘Theorem 3.6.1] also a characteristic function. ‘This completes the proof of Bochner’s Theorem, We derive next another criterion which is due to H. Cramér. Theorem 4.2.3 (Cramér’s Criterion). A bounded and continuous function {ois a characteristic function if, and only if, @ =1 ae Hea) = f° [fe—mesotise—19] dean is real and non-negative forall rel and for all A> 0. We first prove the necessity of the condition by substituting H(t) = [" explit—uy] dF) into (ii), Since the inversion of the order of integration is permissible wwe obtain ely f° [1—cos A(w+y)] (x, A) = 2 — ath % eas "Vhis shows that (x, A) is always real and non-negative. ‘To prove the sufficiency of the condition we assume (i) and (ii) so that dF). 1 (4 pa (4.2.10) ple) Jaks. A) = f fj fe-epemeve dude 0 In the last integral we introduce new variables teue ‘This change of variables transforms the original region of integration into a parallelogram. One diagonal of this parallelogram is located on the z-axis and decomposes it into two triangles. ‘The function p(x, A) is then computed by adding the two integrals taken over these triangular regions, Thus eA) . f, novel f | aes f 7 ‘foe | a66 We introduce the function {a (1 a)ro if |e] 0 (is) Jim = 0. Then f(t) is the characteristic function of an absolutely continuous distri- bution F(s). Since f(t) is a convex function it has everywhere a right-hand deriva- tive which we denote by (2). ‘The function f(t) is non-decreasing for t > 0. It follows from (jv) that f/(f) = 0 for t > 0 and that lim f’@ = 0 60) A funtion (is maid 0 be convex fort > 0 ficken) OER) as forall > 0, t4 > 0, Fora survey of the properties of convex functions we refer the reader fo'G. HE Hardy J.B Litcwood-Ge alps (988), pp. 10-72, 91-96n Its casily scen that the integral %, e-t# f(t) dt exists for all x # 0. We write Lf 431) pe) = f etef(t) at, We see from (ii) and (4.3.1) that fie 43.2) px) =f, F(O costed ‘The conditions of Fourier’s inversion theorem? are satisfied and we obtain fo= f taps) de, It follows from (i) that J&, p(x) dv = 1 and the proof of ‘Theorem 4.3.1 is completed as soon as we show that p(x) is non-negative, Integrating by parts and writing g(t) = —f'(2) we get 1 pe 433) p)=— f g(t) sinxtde where g(t) is a non-increasing, non-negative function for t > 0 while lim (0) = 0. ‘Then wo 2 [Scemle one Let x > 0; the series Zeist f) is an alternating series whose terms decrease in absolute value; since the first term of the series is non-negative one secs that the integrand is non-negative. ‘Thus p(x) 2 0 for x > 0. Formula (4.3.2) indicates that p(x) is an even function of x so that p(x) = 0 if x # 0. Therefore (©) We use here the following theorem due to Pringsheims (Titchmarsh, (1937), p16] the function) te non-inerenting over (Oy 00) and ifs ie integrable over Every finite neerval(0,@) (where a> O)andat hans) = Othen the nversion formula HACHOSE-O] = iy J cm Cf) com yu doe holds for any postive f. A short proof of Pringshel’s theorem can be found in MM, Rices-A. B! Livingstone (1955).2 (8) is a frequency function and ft) is the characteristic function of the absolutely continuous distribution. F(x) = {*,, p(y) dy. ‘We will occasionally call functions which’ satisfy’ the conditions of "‘Vheorem 43.1 Pélya-type characteristic functions. From the preceding proof it is clear that the frequency funetion p(x) of a Pélya-type characteristic function f(t) can always be obtained bby means of the Fourier inversion formula (4.3.1), even if the condition of Theorem 3.2.2 that f(t) should be absolutely integrable is not satisfied, We list next a few Pélya-type characteristic functions. (4340) f(Qe el 1 q (4340) f() = -ld for 0S || s¥ G84) f= for |i) 24 1-\¢ for |t) £1 0 for |e Using the inversion formula (4.3.1) we see easily that (4.3.4a) is the characteristic function of the Cauchy distribution. ‘The characteristic functions (4.344) and (434d) are absolutely integrable; how- ever (4.3.4b) and (4.3.4c) are examples of characteristic functions of absolutely continuous distributions which are not absolutely integrable (see page 41). The corresponding frequency functions can nevertheless be computed by means of formula (4.3.1) but lead to higher transcen- dental functions. The frequeney function of the characteristic function (4.3.44) is the function zl“ | Pélya’s condition permits us to construct examples which help us to get a better insight into the assumptions of the uniqueness theorem. ") (34a) 1 = { Example 1. Let f(t) be any Pélya-type characteristic function whose right-hand derivative f(t) is strictly increasing for t > 0. Replace an arbitrarily small ar¢ of the right side of f(t) by a chord and change the left side symmetrically. In this manner one obtains a new function Jit) which also satisfies the conditions of Theorem 4.3.1. ‘Thus fi(f) is a Polya-type characteristic function which agrees with f(t) everywhere, (2) See pase 36,B except on two symmetrically located arbitrarily small intervals. As a consequence of the uniqueness theorem fi2) and f() belong to two different distributions. Example 2. Let fi(t) be the characteristic function (4.3.4¢). while FAt) is the function (4.3.44). ‘These are examples of two characteristic functions which agree over a finite interval but belong to different distributions. Pélya’s condition can be used to derive another sufficient condition which is applicable to certain periodic functions. Theorem 43.2. Let f(t) be a real valued function which satisfies the ‘following conditions @ 0) = (i) A 1) (iii) _f(0)is convex and continuous in the interval (0, 7) (iv) f(t) is periodic with period 2r. () fo) =0. Then fit) is the characteristic function of a lattice distribution, We consider the funetion fi(t) defined by if We 039 60-(0° Se toe Clearly, fi@) satisfies the conditions of Pélya’s theorem and is therefore a characteristic function. Hence it follows from (43.1) that iE ete f(gdt > 0 for all x. Combining this with (4.3.5) we obtain (43.6) f f(O)costedt 2 0 while condition (i) implies that, 43.7) f ‘f()sintede = 0. We substitute x = an[r (n integer) in (4.3.6) and (4.3.7) and see that Ay = : ff so costar = 0 (4.3.8) te a Boot f sitar =4 ‘The quantities y and By are the Fourier coefficients of the function f(t). It follows from Dirichlet’s conditions [Titchmarsh (1939)] that ‘Ct is equal? to its Fourier series in the interval (—r, +7). On account ‘of the periodicity of f{¢) one has then 439) = M+ SAncas™ wt for any real value of t, Formula (4.3.9) indicates that f(¢) is the charac- teristic function of a lattice distribution whose lattice points are the points (nnjr) (m= 0, 1, £2, .)o "The functions f(t) and fi(t) discussed in this proof are also examples of characteristic functions which agree over a finite interval. The first example of this kind is due to Khinchine. Extensions of Pélya’s condition can be found in Girault (1954) and in Dugué (1957b). These authors also obtained some interesting results concerning Pélya-type characteristic functions [D. Dugué-M. Girault (1955). ‘The decision whether a given funetion f(t) is a characteristic func- tion can sometimes be made by means of the results derived in earlier chapters. The continuity theorem is frequently useful in this connec- tion; we consider next a simple example, Let 1 2 cosht ~ ettet f= ‘The function cosh ¢ is an entire function which has zeros at the points in(2j—1)]2. Applying Weierstrass’ Theorem on the factorization of entire (integral) functions we get ° 42 ont TI [+g al so that = 40 a Ho= | Jen whee 240 = [+ oan Let Kt) = 1/(1+#) be the characteristic function of the Laplace distribution, then ( D. Dugué (1955), (1957) investigated the Fourier series of a characteristic function and shosted that a characteristic function is, in a certain interval, equal tothe sum of is Fourier series.1s is also a characteristic function, ‘Then [Corollary 1 to Theorem 3.3.1] inl) = T [oe) st is also a characteristic function. Finally we conclude from the continuity theorem that S(t) = tim f(t) is also a characteristic function. In a similar way one can show that the reciprocal of an entire function of order 1 which has only purely imagi- nary zeros and which equals 1 at the origin is always a characteristic function. 44 An essential property of characteristic functions ‘We mentioned already that every distribution function has a charac- teristic function and discussed in Chapter 3 some other very important theorems concerning characteristic functions such as the uniqueness theorem, the convolution theorem and the continuity theorem. ‘The great usefulness and importance of characteristic functions in’ probability theory is largely explained by the fact that these properties make them a very convenient tool for the solution of many problems. The present section deals with the question whether there are any other integral transforms of distribution functions which have these properties. Let G(s) be a distribution function and consider its integral trans- form by means of the kernel Kis, x), that is (441) 9) = £ (sx) dG(s). In the following we denote by Gi(x), G(x) distribution functions and by gi(2), go(s) their respective transforms (4.4.1). In this section we show that the uniqueness and convolution pro- perties determine essentially the kernel. The following theorem gives a precise formulation of this statement. Theorem 44.1. Suppote that a hernel Kit) ster the folowing conditions (1) K(x) is a complex valued function defined for all values of the real variables s and x and is bounded and measurable in ». () Criucres Property): (8) = 92(6) if, and only if, G(x) (II) (Convolution Property): If Gla) = Gi? Ge = i Gilx—NdGxt) then (s) = 9x(s)90(8).76 Then K(s, #) has the form K(s,x) = clea where A(s) is a real valued function of s such that the values assumed by |A(s)| form a set which is dense on (0, +0). ‘The converse of this statement is also true, We sce from assumption (I) that every distribution function G(x) has a transform given by (4.4.1). We write assumption (ITI) in terms of the kernel and obtain le Kis 2)dy f ” Gi(e—w)dGa(u) = fe Kis, Gt) if * Kis wydGe(u) = iE iE (6, )K(s,n)dG(\dGalu). On the other hand “ i ” K(s,3)de if * Gx(x—uydGu(u) = le ic K (t+ u)dGy(t)dGo(u) so that (442) f : f * K(5,t+u)dGu()dGelu) - iE le (5, DK (6, w)dGy(QdGx(u) holds for every pair of distributions Gy(x) and Gy(a). Let ¢ and » be arbitrary real numbers and put [ Gus) = (=n) (Gafe) = (=) +<—-9)]. Substituting (4.4.3) into (44.2) we get (444) Kon) +K(60048) = KK 0)+K(6,8)] for any real £ and 7, We obtain in particular for ¢ = 0 2K(6,2) = 2K(6,0)K(647). “Therefore (4.44) reduces to (445) KG n+ 8) = Kis MBs. It is known (see for instance Hahn-Rosenthal (1948), pages 116-118] that every measurable solution of the functional equation, Ha+8) = HOAs) (443)7 has the form $(E) = °€ where C is a constant. Since K(s, x) is by assumption (I) measurable in x, every solution of (4.4.5) is of the form (44.6) K(6,x) = eto, Let p(s) = B(s) +éA(s); then [K(s, 2)| = ¢*80, Since K(s, x) is bounded wwe have B(s) = 0. The kernel has therefore the form (44.7) K(x) = cea, ‘The transform (4.4.1) of a distribution G(a) is therefore f etedto dO (x) while the characteristic function g(t) of G(x) is 449) a) = [ etraow), (448) g() It follows from (4.4.8) and (4.4.9) that (44.10) gf4(9)] = 96). ‘We show next by an indirect proof that |4(s)| must assume all values of a set dense in (0, +00). Suppose tentatively that |4(s)| omits an arbitrary interval I ~ (a, a+A) on (0, +00) and denote by 1’ = (—a, —a—h) the interval which is symmetric to I with respect to the origin. It is then possible to construct two Pélya-type characteristic functions g(t) and go(t) which agree everywhere except on I and I’, The two corresponding transforms (4.4.1) are 94(s) = g1[4()] (j = 1, 2). Since |A(s)| does not assume values of I we see that $4(s) and go(s) agree for all values s but belong to different distribution functions in contradiction to the uniqueness assumption (II). We still have to prove the converse statement. Suppose that the kernel is given by (4.4.7), then it is immediately seen that (I) holds. The proof of (II) can be carried out in the customary manner with the aid of Weierstrass’ approximation theorem. Finally it is easy to show that (TIT) is also satisfied. ‘We see therefore that an integral transform (4.4.1) which is defined for every distribution function and for which the uniqueness and the convolution theorems hold, is obtained from the characteristic function by a simple change of the variable. We note that we arrived at this conclusion without using the continuity theorem.5. FACTORIZATION PROBLEMS—INFINITELY. DIVISIBLE CHARACTERISTIC FUNCTIONS 5.1 Preliminary remarks We showed in Chapter 3 that the product of two characteristic functions is always a characteristic function. It is therefore obvious that some characteristic functions can be written as products of two or more characteristic functions Every characteristic function f(t) can be written as the product of the two characteristic functions f(t) = e™ (m real) and f(t) = f{t)e™, We say that the representation of a characteristic function as the pro- duct of two characteristic functions is trivial if one of the factors has the form et, In order to avoid trivial product representations, we intro- duce the following definition, ‘A characteristic funetion f(@ written in the form G11 sO = AOA where fi(t) and f.(t) are both characteristic functions of non-degenerate distributions. ‘We say then that fi(#) and fe(t) are factors of f(t). A characteristic function which admits only trivial product represen- tations is called indecomposable. ‘We show next that there exist indecomposable characteristic functions. Theorem 5.1.1. Let F(x) be a purely discrete distribution function twhich hhas too discontinuity points. Then its characteristic function is indecom- posable, We see from the Corollary to Theorem 3.3.3 that, the components of F(x) are necessarily purely discrete distributions with a finite num- ber of discontinuity points. ‘The inequality, given in this corollary, in- dicates that at least one of the components must be a degenerate distri- bution. ‘This proves our assertion. ‘The factorization of a characteristic function into indecomposable factors is somewhat similar to the factorization of integers into prime factors. This is the reason why the theory of the decomposition of characteristic functions is often called the arithmetic of distribution functions. However, this analogy does not go very far; as an illustration wwe give an example which shows that the factorization of a characteristic function into indecomposable factors is not always unique. said to be decomposable if it can be Example. Let fit) 4 Sel and write fe) = 4(1 +s et, * fat) = A+, gilt) = Met +e%), golt) = 3(1 +e), 87” It follows from theorem 2.1.3 that the functions fi(t), fo(t), gx(t), g(t) and f(®) are characteristic functions. Moreover it is easily seen that Ald) = fad) fot) = uldgelt). We conciude from Theorem 5.1.1 that f(t) tnd go(t) are indecomposable. It follows from ‘Theorem 3.3.3 that a factorization of gx(t) must necessarily have the form (51.2) gilt) = [pet (1 —phetfgettr + (1 —g)ettn] where G13) O p. Then Ald = C1-Clm]+Ceingcoye = [1+ 2EO— BP isalso a characteristic function, We see then from the continuity theorem that J) = lim fat) = expple— 1} is also a characteristic function. ‘The function [ f(é (« > 0) satisfies also the conditions of the lemma and is also a characteristic function. We conclude then from the Corollary to Theorem 5.3.3 that f(¢) is infinitely divisible, ‘We use this lemma to prove the following theorem. Theorem (De Finetti’s Theorem). A characteristic function is in- finitely divisible if, and only if, it has the form S(0) = lien exp(Paln(®)— 1} where the Pm are positive real numbers while the gy(t) are characteristic functions. ‘The sufficiency of the condition of the theorem follows immediately from Lemma 5.4.1 and from the continuity theorem. We show next that the condition is necessary and assume that f(t) is infinitely divisible. It follows then from the Corollary to Theorem 5.3.3 and from Lemma SAL that £0) = expe{{(AO)"—H]} is, for any real postive «a characteristic function. Since 4(0) = lim ft) wwe sce that (t) can be represented in the above form with Pm =m and gn(t) = [fh Theorem The limit of a sequence of finite products of Poisson type characteristic functions is infinitely divisible. The converse is also true.Co Every infinitely divisible characteristic function can be voritten as the limit of a sequence of finite products of Poisson type characteristic functions. ‘The first part of the theorem is a consequence of the closure theorems. ‘To prove the second part we assume that f(t) is infinitely divisible; according to DeFinetti’s theorem it can be represented as Gott) f= lim exptpalen()—1) where the ga(t) are the characteristic functions of some distributions Gy{x) so that ie ant) = [ada ‘Then we see that (642) palga()—1) = a tim on [ete dGa0). We wish to approximate the integral by Darboux sums and introduce therefore the subdivision -A=m. It is easily seen that £0) = > 0-Cincpre, 2 and we note that this is the characteristic function of the geometric85 distribution listed in Table 1. We expand feel Vee wai ws ne a) 5) f= T [nl ie »}. We can apply Theorem 5.4.2 and see that f(t) is infinitely divisible. 5.5 Canonical representations ‘The results of the preceding section can be used to deduce explicit formulae for infinitely divisible characteristic functions. For their derivation we need several auxiliary theorems. Lemma 5.5.1. Let a be a real constant and let Ox) be a real valued, bounded and non-decreasing function of the real variable x such that ®—co) = 0. Suppose that a function f(t) of the real variable t admits the representation 51) og) = tas [” (a1) My 5.1) toay(t) = ita [" (ae—1 55) Sane. The integrand is defined for x = 0 by continuity, and is therefore equal to — 1/2 if x = 0. Then fit) is an infinitely divisible characteristic function. Moreover, the constant a and the function 0x) are uniquely determined by ft). Let t belong to an arbitrary fixed interval, then it is seen that the integrand of (5.5.1) is bounded and continuous in x so that the integral exists for all values of t. We first prove by repeated applications of the continuity theorem and of the closure theorems that f(t) is an infinitely divisible characteristic function, Let 0 < « < 1 and define Gm pf = ee cinco \ ise ‘The function 1,(¢) is continuous at t = 0; we can write it as a limit of Darboux sums Sn(t) where aoe), Swalt) = > Patel” —1) iat] with Ay AEE so) oe ah & pe = Alten) S(t) is the second characteristic of a product of Poisson type charac- teristic functions. 1,(t) is the limit of these functions and therefore the86 logarithm of an infinitely divisible characteristic function, Let now 653) 14) = lim 1(0) =f (et) enn Clearly {o(t) is continuous at t = 0; we concude again from the con- tinuity theorem that exp[Zo(0)] is a characteristic function and then from ‘Theorem 5.3.3 that it is infinitely divisible. Finally it follows from (5.5.1) that ® logs) = 140) +ita — F140) ~H(-0)] ‘The last equation shows that f(t) is the product of the infinitely divisible characteristic function exp[/o(#)] and the characteristic function of a normal distribution so that f(t) is also infinitely divisible. ‘We show next that the constant a and the function (x) are uniquely determined by (5.5.1). We write (¢) = log f(t) for the second charac- teristic and see easily that 654) Ht)—¥4Ce+W)+ge—A)] - L ott cons) 40 We introduce now the function 1 I+ $e—A) ne) = f, [ro - se _ Dan, Since the integrand in (5.5.4) is bounded, we can integrate with respect to h under the integral sign and obtain vp = [7 oo(t 222) 2a, We introduce the function Ag) = [7 (1- ‘Then xe as) = i ee while . Ae = | etedA(ey.7 Ic is easily sen that there exist two positive constants cx and ca such at siny) 14 2) se o 0, nl = 1} = nel —1} = n(n) se) +4 1/m)} = 4(0)+0(1) ‘so that (553) $0) = lim afffQ}”*~ 1). Since, by assumption, f(t) is infinitely divisible [/(¢)]}/* is a charac- teristic function, Denote the corresponding distribution function by Fal); then“389 (5.5.6) afer} =m [ (e##—1)dF a(x) = vit f° i arayen (ote wo 1 +32 + ) aut If we write (6.5.7) then we see from (5.5.5) and (5.5.6) that $q(t) has the form (5.5.1) and that . dim da(t) = #0. "The three lemmas permit the derivation of the desired canonical representations for infinitely divisible characteristic functions. ‘Let now f(¢) be an infinitely divisible characteristic function. Since ‘f(t) cannot Vanish, the function 4(t) = log f(¢) is defined for all values of t. According to Lemma 5.5.3 there exists a sequence of functions a(t) which has the following ‘property: each y(t) has the form (5.5.1) and the sequence qy(t) converges to g(t) as n tends to infinity. It follows then from Lemma 5.5.2 that 4(t) has also the form (5.5.1). If we combine this with the result of Lemma 5.5.1 we obtain the following theorem, ‘Theorem 5.5.1 (The Lévy-Khinchine Canonical Representation). The function f(t) is an infinitely divisible characteristic function if, and only if, it can be written in the canonical form (5.5.1) logf(t) = ita+ L. (ot-- i ) Z Ham) tohere a is real and where 6x) is a non-decreasing and bounded function such that 6(—<0) = 0. The integrand és defined for x = 0 by continuity to be equal to —(t8/2). The representation (8.5.1) is unique. "The canonical representation given by the last theorem can be some- what modified. We define two functions, M(u) and N(u) and a constant o? by writingMw) = f * 12 ame) foru<0 (558) Nw) = - faa for u>0 of = H(4-0)-H(—0). ‘The functions M(u) and N(u) are non-decreasing in the intervals (=, 0) and (0, +c0) respectively and M(—c0) = N(+.00) = 0. For every finite ¢ > 0 the integrals f',12dM(u) and {5 tdN(u) are finite. Conversely, any two functions M(u) and N(w) and any constant 2 satisfying these conditions determine, by (3.5.8) and (5.5.1), an infinitely divisible characteristic function.’ We have therefore obtained a second canonical form, ‘Theorem §.5.2 (The Lévy Canonical Representation). The function f(t) is an infinitely divisible characteristic fection if, and only if, tt ean Be triten in the form 559) toast) = ita Sere (etn — 7, Jan 59) es ~ ta Fee f( ina) ° itu ees ae) where M(u), N(u) and o® satisfy the following conditions: (@) M(x) and N(u) are non-decreasing in the intervals (—co, 0) and (0, +00) respectively. (i) M(-«) = M+) = 0. (ii) The integrals %,u®dM(u) and fgutdN(u) are finite for every e > 0. (iv) The constant o? is real and non-negative. The representation (8.5.9) is unique. ‘The canonical representations (5.5.1) and (5.5.9) are generalizations of a representation, due to Kolmogorov, which is valid only for the characteristic functions of infinitely divisible distributions with finite ‘Theorem 5.5.3 (The Kolmogorov Canonical Representation). The func tion f(t) is the characteristic function of an infinitely divisible distribution toith finite second moment if, and only if, it can be written in the form (5.5.10) logy (t) = iet+ £ (et-1- ing KO1 where c is a real constant while K(u) is a non-decreasing and bounded function such that K(—co) = 0, The representation is unique. ‘The integrand (el##—1—its)/x® is defined for x =0 to be equal to -(#)2). Let f(t) be an infinitely divisible characteristic function and suppose that the second moment of its distribution function exists and is finite, ‘Then f(t), and therefore also 4(t) = log f(t), can be differentiated twice. We form the second central difference quotient AMO (eae and conclude that | B8600)_ (S510) tip pr| We note that f{¢) admits a representation (5. show that the integral [°, (1+22)d0(x) is finite. Then also finite. We write K(a) = ft, (1-+38)40()) and ¢ = a-+J%, yd0(y) and obtain (5.5.10). Conversely, sentation (5. suppose that the function 4(¢) = log f(t) admits a repre= 10). "Then f dK) wo 143? satisfies the conditions of Theorem 5.5.1 so that f(t) is an infinitely divisible characteristic function. Moreover, it is easily seen that (5.5.10) may be differentiated twice under the integral sign so that the second moment of the distribution function corresponding to f(t) exists. The uniqueness of the representation is an immediate consequence of the uniqueness of the representation (5.5.1). ‘As an illustration we determine the canonical representation for a given infinitely divisible characteristic function, The procedure repeats the steps of the proofs of Lemmas 5.5.3 and 5.5.2, We consider as an example the Gamma distribution ro-(T9 [irre for x >0, 0 for x <0. ‘The corresponding characteristic function is w= (-4)"92, where 8 and A are two positive parameters. It follows from the form of f(t) that it is an infinitely divisible characteristic function so that [/(d)]}!/* is also a characteristic function. We denote the corresponding distribution function by Fa(x); clearly F(x) is also a Gamma distribu- tion and gam pe 7 eo =[ Tom mietdy for x >0 0 for x <0. Substituting this into (5.5.7) we get apesiety ZL oytnetvdy for > 0 wo (Fan haerrns TAMA) Jo Tey o for <0 and 1 _yvnetvdy, a oye en We note that fig ae me DQln) — ™™(Am)Aln) > LYn) +1] and obtain from Lemma 5.5.2 fs) = Li te ( af iw for #>0 0 for «<0 while Sen a a o 1+? Itis then easy to compute the other canonical representations. ‘Table 5 lists the canonical representations of some of the more common infinitely divisible characteristic functions. ‘We conclude this section with a few remarks concerning the factori- zation of infinitely divisible characteristic functions. Theorem 5.5.4. Let f(t) be an infinitely divisible characteristic function ‘and suppose that it can be decomposed into two infinitely divisible factors, Hl) = falt) felt). Then fit) and fut) determine f(t) uniquely. ‘The theorem follows immediately from the uniqueness of the canonical representation; it shows that the cancellation law holds if we restrict the decompositions to infinitely divisible factors.93 o< #13 | oc eer | hel 3044 | lara + Hee | | elie we you o> zayo| BOSS] 0 o> #250, paneton 1 |_a-m ele} amr o (1=sh6cI0), wossieg eae ojo | svsuate ane iar ares 3 wonenuansdend ON pole sme] 0=* # | saree | o> #240 |alv usa ere [xls0¢—vadxe | peunony ox we esas j opmuaresdey onmuosard wonoung | wounguisig {sos0200031 | Suu [_eammeny | jo sum suopouny opspiajgexeyo opaysiasp Aya1fuyEy Jo woReIUasexdas [eOyUOUED s oraer4 However, these are not the only possible decompositions; infinitely divisible characteristic functions can have factors which are not them= selves infinitely divisible. We give the following example, Let a and & be two positive real numbers and write o = a+ib. It can be shown that the function NIU + ef) Coe oN — Ge) is a characteristic function if (65.13) 6 2 2ay2, Then f(t) is also a characteristic function, as is 89 = SOND = VO? = Temas We will show later [Theorem 8.4.1] that f{) and therefore also f( -t) are not infinitely divisible. ‘The product g(t) = f(@)(—t) is the charac- teristic function of the Laplace distribution which is known to be in- finitely divisible (gee page 81), ‘The function f(, determined, by (5.5.12) and (5.5.13), has the following interesting property: f(t) is a characteristic function but is not infinitely divisible, however | f(@)/? and therefore also | f(t)| are infinitely divisible characteristic functions. ‘Thus, the infinitely divisible characteristic function | f{¢)? admits two decompositions, fl)? = [FO - FO) = FOF(-9- ‘The first decomposition has two infinitely divisible fact factors of the second decomposition are not infinitely divisible. This example shows that two different characteristic functions, namely f(t) and | f(2)|, can have the same absolute value. ‘We give next another example which shows that an infinitely divisible characteristic function may have an indecomposable factor. Let p and q be two positive real numbers such that p>q>0 and ptg=1. ‘The funetion a(t) = prac is then—according to Theorem 5.1.1—an indecomposable characteristic (©) To show this, one expands (0) into partial fractions and computes 1 Ef ew soer by integrating the expansion term by term It is then not difficult to show that the resulting expression is non-negative W (5.3.13) ie satshedfunction. We write ra(é) = logen(t) = logp + log{l + (alp)e] Tis then easily seen that no- > (yer. & Let Se #0 => a-(5) [etn 1] B12) sy m= Dan(5) [ene], then nl) = $O)-r2(0). The functions fit) =e and ga(t) = er are infinitely divisible characteristic functions; moreover, S(O) = ar(thga(t). ‘The infinitely divisible characteristic function f(¢) has therefore an indecomposable factor 10). 5.6 A limit theorem We have shown (Theorem 5.3.3) that a characteristic function which is the limit of a sequence of infinitely divisible characteristic functions is also infinitely divisible. In the present section we show that under certain conditions the limit of a sequence of characteristic functions is, infinitely divisible, even if the elements of the sequence are not infinitely divisible characteristic functions. We consider in the following an infinite sequence of finite sets of characteristic functions. Such a system {fus(t)} (j= 1, 2, os ni n= 1, 2, vy ad. inf.) can be arranged in a two-dimensional arra Fale), firlt) oy f(s Fart), frat) os fret) (5.6.1) ae Salt) faalt),96 We form the (finite) products Salt) = T] fas) ft of the functions in each row of the scheme (5.6.1) and wish to investigate their limits. As usual, we denote by Fpj(x) the distribution function which corresponds to fyj(t). The following theorem contains a very important result. Theorem 5.6.1. Let {fus(t)) (j= 1, 2 son Rani = 1, 2; ny ade inf) bea system of characteristic functions and suppose that 6.6.2) lim [ sup. [fas(t)— 11] = 0. Denote by gn(t) the characteristic function determined by (6.63) tog en(t) = Sten f (ce—1) dF 209} A where i) ong and where + > 0 is a constant. The necessary and sufficient condition for the convergence of the sequence of characteristic functions (5.64) — falt) = [| fast) ja to a characteristic function fit) is that the sequence gq(t) converge to a limit. Then the limits of the sequences fy(t) and gq(t) coincide. ‘The characteristic functions gq(t) are infinitely divisible. be seen by writing them in the canonical form (3.5.1) with e 40) = > [i thal rou) fil ond . oS fo [ apstoeesi or by noting that the g4(f) are finite products of limits of Poisson type characteristic functions. Theorem 5.6.1 indicates that it is possible7 to replace the investigation of the limit of a system of arbitrary charac- teristic functions [subject to the restriction (5.6.2)] by the investigation of the limit of a sequence of infinitely divisible characteristic functions. ‘This circumstance explains the great importance of Theorem 5.6.1 in connection with the study of limit distributions for sums of inde- pendent random variables. We do not intend to discuss in this monograph the limit theorems and will therefore not be in a position to appreciate the full significance of this theorem. For its proof we refer the reader to B. V. Gnedenko~ A. N. Kolmogorov (1954) where this result can be found in its proper context. In connection with our investigation of factorization problems we will use a Corollary to Theorem 5.6.1. Corottary to Theorem 5.6.1. Let fit) be a characteristic function and suppose that f(t) admits a sequence of decompositions where the fast) (j = 1, 2, 2, .) form a system of charac- teristic functions which satisfy (36.2). Then Qh 8 infintely divsble. ‘The corollary follows immediately from Theorem 5.6.1 if we observe that (= (0. 5.7 Stable distributions In this section we discuss a class of infinitely divisible distribution functions, the so-called stable distributions. Stable distributions and their characteristic functions are important in connection with certain limit theorems and were originally introduced in this context. Our study of these distributions is motivated by the fact that the class of stable characteristic functions is of independent interest and occurs also in some problems not related to the limit theorems. A distribution function F(x) is said to be stable if to every bi > 0, by > 0, and real cy, 2 there corresponds a positive number and @ real number c such that the relation (6.7.1) *( holds. "The characteristic function of a stable distribution is called a stable characteristic function. ‘The equation (5.7.1) is not so much a property of an individual distribution function’ F(x) but is rather a characteristic of the type to which F(x) belongs. It would therefore be more appropriate to say that98 a distribution belongs to a stable type if its type is closed with respect to convolutions. The defining relation (6.7.1) can be expressed in terms of charac- teristic functions as (5.7.2) f(brt) (bat) = (bres where y = e~a cn Let Br B's, ny B'n be n positive real numbers; it follows then from (5.7.2) that SO fO2t)..f Cat) = fer where y' is some real number while 8 is a positive number. If we put 6; = 1 (= 1, ...m)and write by for the corresponding value of 8’, then We get FOI" = F(bntetrt 10-((5,) ell ‘The last formula implies the following result. Theorem 5.7.1. A stable characteristic function is infinitely divisible, We sce therefore that a stable characteristic function has no real zeros. We can take logarithms in (5.7.2) and express this equation in terms of the second characteristic 4(f). We obtain (5.73) $i +dbat) = Hot) + it. Since 4(¢) isthe logarithm of an infinitely di ible characteristicfunction, ‘we can write it in the canonical form (5.5.1) as Ho tas f” (ena) a, It follows then that, $(0t) Since the function 2/(1+6%2) is bounded, we see that the integral =f 2. ’ fe er ie) = ie a) exists, We write ay = ba+ (1-6) f Tato)99 and obtain, by means of an elementary computation, ity) 1+d-%y2 iss 674) Het) = oo fo (w-1-; 2 We introduce again the functions Mu) = i Fw where v <0 vo fA and write o? = (+0) ~0). With this notation we obtain from (8.7.3) and (5.7.4) the relation ieee ps itay, Pra [° (ety—1 - annoys an, pee (ee ine) ory beta? +f. (er-1 as 5) aNor tm -a 406-49). -40(9) where uw >0 fifo amen (rao ity se = itay— Fes fe (e +f (# From the uniqueness of the canonical representation we see that (5.7.52) 04(62—by?—ba?) = 0 (8.7.5d) My) = MGry)+MGry) if y <0 (5.7.5e) Ny) =Nb'y)+NGry) if y > 0. letermine first the function M(u) (w < 0). Let 1, Bay ws Pu be m ositive real numbers; it follows from (5.7.5b) that there exists a positive number f = A(f1, Ba, ~-» Bn) such that M(exy)+M@Bay)+ + MIB) = MiG). We substitute here fj = 1 (j = 1, ..., m) and write (1, 1, ... 1) = An and see that MQ) = Min) (1m)MQ) = Milas). Here y < 0 and An > 0. Using this reasoning we see that to every positive rational number r= m/n (m, positive integers) there 2 Jan A an(e-ty) it Satna) or100 corresponds a positive real number A = A(?) = Am/4n such that (3.7.6) MQ) = M(4y) (y < 0). ‘The function A = A(r) is defined for all rational r > 0; we show next that A() is non-increasing for rational values of the argument, Let ry and rz be two rational numbers and suppose that ri < re. Since Mu) 2 0 we sec that nM(u) < reMlu) or, according to (5.7.6), M[A(riia] S M[A(rayd]. Since M(u) is non-decreasing and u <0 we conclude that A(r1) = (rz). By the same reasoning we can show that A(r) is strictly decreasing, provided that M(u) 0. Let us suppose from now on that A1(u) % 0. ‘We define now a function for all positive real values of x by means of At) if x is a positive rational number Clee Lub.d(r) ify > O is irational. It follows from this definition that B(x) is non-increasing and it is easy to show that B(x) is strictly decreasing. Let now x be an arbitrary positive real number, then there exist two sequences {re} and {r’s} of rational numbers such that the rv approach x from below while the r'e tend to x from above. Since ro < x Bix) > B(r's) and hence yB(r») < yB(x) < yB(rs) for any y <0. Since Miu) is non-decreasing We se¢ 1 MpyBe.)) $ MiyB(s)] $ MfyBe'.)) It follows then from (5.7.7) and (5.7.6) that reM(y) $ M[yB(s)] $ r'eM(). We let © tend to infinity and see that for every real positive x there exists a B(x) > 0 such that (678) 2MQ)= MBG] (y <0). Since the function Miu) is non-decreasing and has the property that M(=00) = 0, we see that BO)= 0, B= 1, Bo) = 0. The strictly decreasing function x = B(x) has an inverse function x = f(z). This function is defined for z 2 0 and is single-valued and non-negative. We rewrite (5.7.8) in terms of f(z) and see that to every real z > 0 there corresponds a A(z) > 0 such that (679) Ale)M() = Myvz) is satisfied. Let m(y) and ma(y) be two solutions of (5.7.9) and suppose that m(y) #0. We put os ma m(y)101 and see that mdey) _ leony) _ mat : . =O ng (©) ~ sm) ~ Bl=im(y) my) — "0 ‘This indicates that the quotient of two solutions of (5.7.9) is a constant, Moreover mi(y) = |y|-+ is a solution and (2) = |z|-% ‘Therefore the general solution of (5.7.9) has the form MQ) = Gi Since M(—co) = 0 we must have «1 > 0 and since M(y) is non- decreasing we see that C, 2 0. We know (Theorem 5.5.2) that the integral f°, x2dM(u) is finite; this permits the conclusion 23 < 2. We have therefore found that 67.10) Mw) = Glu (C2 0,0 0) and notes that (5.7.11) Ca[B—by%s— bo] = 0. We show next that o #0 implies Ci = Cz = 0 so that Mu) = 0 and Mu) = 0. We conclude from (5.7.5a) that o® #0 implies bY —by? —bx? = 0. Since a < 2, a < 2, we infer from (5.7.11a) and (5.7.11b) that Cy, = Cy =. If, on the other hand M(u) [or N(u)] is not identically zero, then Cy > 0 for C2 > 0}. We put bi = be = Land conclude from (5.7-1la) [or (5.7.11b)] that 6* (or 6) = 2, ‘Then necessarily b? % 2 so that it follows from (5.7.5a) that o® = 0. ‘We finally show that a1 = az. Suppose that Ci > 0, C2 > 0, and put again 61 = b= 1; it follows then from (5.7.11a) and (5.7.11b) that bv = 2 = 6% 50 that a = a2. We have therefore determined the canonical representation (in Léxy's form) of stable distributions and summarize our result, Theorem 5.7.2. The characteristic function of a stable distribution has the canonical representation itu 7 Jam (3.7.13) log f(t) +H +f (ot one +a102 ‘here either 0 a0 and Mi Mw) Cilul-*(@ <0), Nu) = —Cau-* (w > 0) The parameters are here subject to the restrictions O0. Conversely, any characteristic function of the form (5.7.13) is stable. ‘The last statement of the theorem is exsly verified by elementary computations. ‘The parameter a is called the exponent of the stable distribution, tis possible to obtain an explicit formula forthe second characteristic of stable distributions by evaluating the integrals. Flora itaita mt [olemt- i) Se which occur in their canonical representation. ‘The computations are carried out separately for the three cases 0 < a 0. ‘The relation 1B| 5 I can then be written as siny cos) cosy sin” ‘The constant ya ae [eet | cK isa scale factor; by a suitable choice of the variable it can be made equal to 1. In this case we denote the characteristic function of a stable law in the restricted sense by fa,,(¢) and have (5.7.16) logf..(t) = —|t|* (cosy Let f(t) be the characteristic function of a stable distribution. It follows from (5.7.14) that | f()| = exp ~citle]. Tt is then easily seen that f(¢) is absolutely integrable over (—co, +00) and we obtain then from ‘Theorem 3.2.2 the following result, sin). Theorem 5.7.4. All stable distributions are absolutely continuous. We denote by u(x; a 8, ¢) the frequency function of the stable distri- bution with parameters’ a, 2, 8, ¢ and write in case a = 0 p(x; a B, ¢) Poles a B, c)- Clearly’ pals a B, ¢) = plea; a, B,6), and it is104 therefore sufficient to study the function p(x; 2, 8, ¢). We can find this function by means of Theorem 3.2.2 and obtain easily ‘ [° exp{—ite—cer[1 + Bot, Dae 7.17) peiaBe) => | : = i explite—ct>[1 iB], a))} de where o(|t|, a) is given in Theorem 5.7.3. It follows immediately from (6.7.17) that (8.7.18) p(¥}% 8,¢) = p(—xia —B,e). Itis easily seen that the two integrals in formula (5.7.17) can be differen tiated under the integral sign any number of times; in this manner we get ee ese ae (6.7.19) we] s 2 [meat = r( _ where (5% Bye) Using Stirling's formula we see easily that (5.7.20) fim [ety c if a>1 me Laima a lot if aad From (5.7.19) and (5.7.20) we obtain immediately the following result. Theorem 5.7.5. A stable distribution function with characteristic exponent a & L has a frequency function p(x; 2, B, ¢) which is regular for all real x. For a > 1 itis an entire function while for «= 1 the radius of convergence of the Taylor series of pix; 1, B, ¢) in the neighbourhood of a point x of the real axis is at least equal to ¢ We study next p(x; a, B, ¢) in case « < 1. Then alt], a) = tan(na/2) is constant and we simplify our notation by writing 4 sigan Gedne>d} Formula (5.7.17) becomes then (6.7.21) peviaBid= x [ etestar sD i eA at, Ido ane We consider first the case x > 0 and substitute tx in the integrals105 on the right side of (5.7.21). If we introduce the functions 1 pe fee ve)= = if wemedn ale) = 5 if cane dy, wwe obtain Pin Bre) = (spar) rode] if x > 0. Using (5.7.18) we get also for x < 0 P(x; Bye) = pix, 2% Bye) = (I fsl)[e(Alst)+ af 4|a|)] A. V. Skorohod (1954a) showed that the functions 7(z) and q(z) are entire functions and obtained therefore the following result. Theorem ‘The frequency function of a stable distribution with characteristic exponent a <1 has the form * _ (Ai2aie-) for x>0 Head = [vena *) fer ect there (=) and (2) are entire functions. A number of authors recently investigated the analytic properties of the frequency functions of stable distributions. ‘The asymptotic be- haviour of these density functions was studied by H. Bergstrdm (1952), ‘Yu. V. Linnik (1954) and A. V. Skorohod (1954b) who gave a compre- hensive survey of these formulae, A convergent series for the symmetric stable frequency functions with exponent « <1 was given by H. Pollard (1946) and later for the general case by H. Bergstrim (1952). We list here, without proof, the convergent series for the frequency functions of ‘stable distributions in the restricted sense which were derived by H. Bergstrim (1952). Let p.,(x) be the frequency function which corresponds to the characteristic function f.,(t) given by (5.7.16). If0 0 for x <0.106 An interesting property of stable distributions follows from (5.7.22). ‘We choose the parameter B'so that |8| = 1, it follows then from (5.7.16a) that K = Land thaty = ~(ra/2) for < x < 1,8 = 1 while-y = (na/2) if 0 0, B Pa) =0 if #<0, BEL O 0 we have Fla~e) = 0 while Fla+e) > 0. Similarly we say that F(x) is bounded to the right, and that 6 is its right extremity; in symbols 6 = rext [F, if for any positive « F(b—<)< 1 while F(6) = 1. Distributions which are bounded either to the right or to the left are called one-sided distributions, distributions which are bounded both to the right and to the left are called finite distri- butions. Our preceding result can now be formulated in the following ‘Theorem 5.7.7. ‘The stable distribution functions with exponent 0 < a < 1 and parameter || = 1 are one-sided distributions. They are bounded to the right if B = +1 and bounded to the left if B = ~1. V. M. Zolotarev (1954) found an expression for the density of a stable distribution with exponent greater than I in terms of a density with exponent 1/x, We give this relation without proof; in view of equation (5.7.18) it is sufficient to study the case where x > 0 while -1S BS +1; moreover it is no restriction to assume that ¢ = 1. Zolotarev obtained the following result. Theorem 5.78. Let v= ~Btan(ma2) and ay = (1-+02))!2, the equa- tion 1, O 0, —1$ 8S +1 and any «> 1. Hered = —B Ge and (tm zy" inf (2 are tan o+@-n]}, oe 2 V. M. Zolotarev (1954) as well as P. Medgyessy (1956) obtained differential equations for stable frequency functions with rational expo nents. V. M, Zolotarev (1956) derived also a number of relations between stable distribution functions (density functions). A simple relation of this type is equation (5.7.18). Explicit expressions for stable frequency functions are known only in a few isolated cases. We obtain from (5.7.14) for « = 2 the charac- teristic function of the normal distribution, for a = 1, 8 = 0 the characteristic function of the Cauchy distribution. P. Lévy (1939) has107 shown that the stable distribution function with a = 0,¢ = 1,8 = —1, a = thas the frequency function® 0 if x<0 (6.7.23) pele} 1 5/2 10m ages so: Om Apart from these three cases—the Normal distribution, the Cauchy distribution and the distribution given by (5.7.23)—no stable distri- butions are known whose frequency functions are elementary functions. However V. M. Zolotarev (1954) expressed the standardized") fre- quency function of stable laws for certain combinations of the para- meters « and 8 in terms of higher transcendental functions. ‘These combinations of the parameters are: (« = $, P = 1), («= 3, B = 1), = 4,2 =0),@= 4,8 = 1), (@ = 4, Barbitrary) | We defined the stable distributions by means of the functional equa- tion (5.7.2). It is possible to introduce other classes of characteristic functions in a somewhat similar manner. As an example we mention the characteristic functions f(t) which obey the relation F(t) = FFL, for every ¢(0 <¢ <1) where fe(t) is some characteristic function. ‘These characteristic functions were introduced by P. Lévy and A. Ya. Khinchine [see Lévy (1937a) p. 192]; one sometimes calls this class the L-class [B. V. Gnedenko-A. N. Kolmogorov (1954) § 29] or the class of self-decomposable characteristic functions [Loéve (195S)]. Tt can be shown that self-decomposable characteristic functions at infinitely divisible; they are important in connection with certain limit theorems. We finally mention the semi-stable distributions introduced by P. Lévy (1937a). They are defined by means of the funetional equation Hat) = FH) @#0, g#1) for the second characteristic. *) B.V. Gnedenko-A. N. Kolmogorov (1954) mention that this frequency function was also found by N. V. Smirnov. 0 ie. those obtained by putting a= 0, €= 1 in (5.7.14)6. FACTORIZATION PROBLEMS— GENERAL THEOREMS FROM THE ARITHMETIC OF DISTRIBUTION FUNCTIONS In the preceding chapter we discussed a number of examples which indicated that the analogy between the factorization of integers and the decomposition of characteristic functions is rather limited. While a great number of remarkable decompositions of characteristic functions is known, we have only few general results, and one has the impression that the arithmetic of distribution funetions has not yet reached a final stage in its development. In this chapter we present the most important general theorems concerning the factorization of characteristic functions. ‘This chapter will be supplemented by Chapter 8 which treats factor tion problems of analytic characteristic functions. This separation is justified by the different tools used: in the present chapter we deal with problems which can be handled without using the theory of functions of a complex variable, while complex variable methods are essential in deriving the results discussed in Chapter 8. 6.1 Some notations and lemmas For the investigation of the general factorization theorems we need certain lemmas which we discuss in this section. Lemma 6.1.1. Let f(t) be the characteristic function of a symmetric distribution, then 1-f(2e) § 41-f] for any real t. Since the characteristic function of a symmetric distribution is real (Theorem 3.1.2), the assertion of lemma 6.1.1 follows immediately from Theorem 4.1.2. Corollary to Lemma 6.1.1. Let fit) be a characteristic function and suppose that | f(t)| = 1 in some neighbourhood \t) S 5 of the origin ‘Then fit) is the characteristic function of a degenerate distribution. ‘To prove the corollary we apply repeatedly the lemma to the function AO)? and see that | f(t)| = 1 in every finite interval. We introduce next an operation which is applicable to any characteris- tic function. Let f(t) be an arbitrary characteristic function; then there exists a real number a such that IA@| > 0 for 0S 108 Sa.109 For a fixed a satisfying this relation we define (1A) Naf] = Nats) = ~ [Tog] seo ‘The following properties of this operator are easily established: @ Naif) 20 (i) Neem) = 0 (iii) TE fle) = fi fat) then Na(f) = Na(fa)+ Nal fo) (iv) No(f) 2 LOI lee (v) Ne(f) = 0 if, and only if, (tis the characteristic function of degenerate distribution. Properties (i), (ii) and (iil) follow immediately from (6.1.1); (iv) is a consequence of the inequality log | f(2)| = —logf1-(1-|f@))] 2 1-1 Ff while (v) is easily obtained from (ii) and from (iv). ‘The quantity Na(f) is a measure of the departure of the distribution belonging to fit) from the degenerate distribution. We will refer to Na(f) a8 the Ng-value of f(2). ‘The main object of this section is the proof of the following lemma. Lemma 6.1.2. Let (Fy(x)} be a sequence of distribution functions and denote by {fu(t)} the corresponding sequence of characteristic functions. Suppose that = = 0 is a median of Fy(x) (n = 1, 2, ..) and that there exists a real a > 0 such that (61.2) tim Na(fx) = 0, then Lim F(x) ~ «(). ‘We say that the point x = mis a median of the distribution function Fg) if the inequalities F(m—e) S 4, F(m-+e) % 4 hold for any « > 0. ‘Phe assumptions of the lemma imply that for m sufficiently large St) #0 for 0S tS a. Using (iv) we see that fu- f(t de s 2 fio- [fult)|]dt s 2No( fn). Moreover it follows from Lemma 6.1.1 that ‘ “ fy incerta = 2 [" o-La@omee s 8" o-lpcnreras : : : We combine the last two inequalities and conclude from assumption110 (6.1.2) ofthe lemma that Jim [CLAP] ae Itis then easily seen that for n sufficiently large f(t) # Ofor 0 St S 2a so the argument which we used can be repeated. In this way we see that, for every T > 0, 0. 0, (613) lim fro- rte) ]ae *), We denote by Fy(x) = 1—F,(~x—0) the conjugate distribution of Py(s) and write (6.14) Pala) = Fala) * Fa(x) for the symmetric distribution whose characteristic function is |fa(t)|?. We denote by tere 90) = [" evigy the standardized normal distribution and consider the distribution defined by (6.15) Galx) = By(x)* O(x) whose characteristic function is Balt) = e-P!2| fult)/2 From the inversion formula we see that Gala) Gal-3) = =f" Ecoa pena, Since Gy(x) is a symmetric distribution this can be written as Gye)-B = 2 [BE eraipopan a ae 22 Beeman PSE omg pteyenae ie Sere ftolP—tat as and see thatmn Since for any T > 0 1p? sintx =f, Feet acor—nal s caer we conclude from (6.1.3) that : f, @-lnoriee ° lim In(s) = 0 ‘so that (6.1.6) Lim Gale) = (x). It follows then from the continuity theorem that dim ga) = 1 tim | f(t)? = 2 so that Jim Jato? = 1 and therefore (64.7) Lim Fy = a). We write (6.1.4) in the form Fa) = [Fala and get Bis) = i Ploy) dFy(y) 2 Pale o\Fule) 2 = Fale (1 Fal 09] where «ia an arbitrary postive number. Since by assumption = = 0 is a median of Fa(s) we see that Fy(x) 2 UFale—e). We conclude from the last relation and (6.1.7) that for any x < 0 (6.1.8) Jim Fe) = 0. On the other hand, we see by a similar reasoning that 1=Fate) 2 ie (1 Fi(e—y)]4Fu(9) = Fae +9]2 so that for any = > 0 (61.9) lim Fale) = ‘The formulae (6.1.9) and (6.1.8) imply the assertion of the lemma, 6.2 General decomposition theorems In this section we discuss three general theorems concerning the factorization of distribution functions and characteristic functions, ‘The first two of these theorems are due to A. Ya. Khinchine, the last is due to H. Cramér. ‘Theorem 6.2.1. Every characteristic function can be represented as the product of at most twoo characteristic junctions which have the following property: one does not have any indecomposable factors tokile the other is the convergent product of a finite or denumerable sequence of indecom- posable factors. Let f(t) be an arbitrary characteristic function and denote the cor- responding distribution function by F(2). Since /(t) is continuous and f(0) = 1, there exists a real a such that /(¢) # 0'if |e! $ a; in the fol- Towing we fix such a value a and write Na(f) = a If f(t) does not have any indecomposable factors then the theorem holds. We suppose therefore that f(t) has indecomposable factors. Then it is possible that f(e) has a prime factor) px(t) such that Ne(x) > 9/2; it follows then from (ji) [see page 109] that one can write FO) = AOAO where Na(pr) > a/2 while Nu(fa) < a/2. In this case we repeat the procedure with fi(f) but use a/4 instead of a/2 as the lower bound for the Na-value of its prime factor. If fi(t) has an indecomposable factor whose Na-value exceeds «4, then one obtains a decomposition FO = prlthpal) fl) where Na(p)) > a/4 (j = 1, 2)? while every indecomposable factor at) of fa?) has the property that Na(g) < a/4. In case no indecom- posable factor with Ng-value greater than «/2 exists, we search for prime factors p(t) which satisfy the relation No(p) > a4. In case such factors exist one obtains a decomposition into at most four factors L(O) = Prlt) -- Past) felt) where 1 < ng S 3. Here the pj(t) are indecomposable factors and satisfy the inequality Na(pj) > a/4, while every prime factor g(t) of fa(t) has the property that Na(e) < 2/4. 1) We use the tetms “prime factor” and indecomposable factor synonymous. Note that sceording to our construction Nui?) > a2113 ‘We repeat this procedure and see that f(t) can be decomposed in the following manne (6.21) f(t) = Pa(Ohpalt) «Pa ()ful0) ashere the (fare indecomposable factors such that api) > o/2* =, 2 om ‘Emp &2*—1) and where every prime factor B00 of fa) has the property that Nelg) < a)24. a hhappen that for some k 2 1 the characteristic function fu(0) has no indecomposable factors. Then our process terminates and we see that the theorem holds, We must therefore prove the theorem only in the case where the factorization process does not terminate; the factors py(t) form then an infinite sequence. Since PR Nelos) < Naf) ‘we see that the series EM is convergent so that the sum EN joes Nels) converges to zero as & tends to infinity; this convergence is uniform in m(m > 0). We apply now lemma 6.1.2 and see that there exist real numbers Aj, such that lim eft,, TI put) = 1 uniformly in every finite t-interval |i] $ T and o' > ©. We write p(t) = putt) explian(t)] and see that (62.2) tw + Son(s) = 2B o()+-0(1) a8 0» 00 where By,»(t) assumes only integer values. The left side of (6.2.2) is continuous, moreover By,y(0) = 0; hence By,.(t) = 0 for sufficiently large v and we have Aye as 0 +0. It is no restriction to assume that ox(1) = 0 s0 that Av,» = o(1) (9) "his can be seen if one multiplies each p,(?) with expt —itou(1)]ut and Sonny = ol) a orm. ‘We see therefore that (623) tim] [pst = 1 vod 1 uniformly in |t) s Tand e” > v. "The infinite product Tle ya is then convergent, let v(t) be its limit. It follows from (6.2.3) that 2 (0) = im T [pat) i where the convergence is uniform in every finite t-interval. We sec then from the continuity theorem and its coroilaries that o(¢) is a characteris tic function. Let ¢ > 0 be an arbitrary positive number; according to (6.2.3) we have kim \TI pill] 0, |e] s 7) pea if k is sufficiently large. Since eam Sit) = fevnlt) TT pt) ra one can conclude easily that |fi)—firm(d)| < € (m>0, || s 7). ‘This means that the sequence {fi(0)} converges also to a characteristic function, Denote u(t) = Him ft. Tt follows then from (6.2.1) that LO = (OMe.15 ‘The function u(t) has no indecomposable factor, this follows easily fram the fet dnt each indecompostble factor of ‘sf) must be a prime factor of fx(t) for all k. But such a factor cannot exist since the No~ values of the prime factors of fi(¢) tend to zero as k goes to infinity. ‘This completes the proof of ‘Theorem 6.2.1 ‘The second decomposition thearem supplements the preceding re ch have no indecomposable sult by characterizing the distributions wi factor. Theorem factor is Let fit) be a characteristic function which has no indecomposable factors and denote by D @) LO = AO LO ~ fall) an arbitrary decomposition of f(¢) where all f(t) are characteristic func- tions. Suppose that a is a positive number such that f(0) + 0 if |e] 3 a, wwe write then 4D) = Max Nf) and denote by » the greatest lower bound of the »(D) for all possible decompositions D of f(¢). ‘We derive first a lemma: Lemma 6.2.1, Let f(t) be a characteristic function which has no indecom- posable factor, then v = 0 It follows from the definition of » that there exists a sequence of decompositions {Dx}, say (Dn) $= far) fa.2(t)—-fr,e(f) (m= for which v(Da) converges to v so that YS Dy) 0 and sufficiently large values of a > 0 and x one has Fis(—a) <_ while Fys(a) > 1- We carry the proof indirectly and assume that one of these inequali- ties, for instance the second, is not satisfied for a > 0 and arbitrarily large n. ‘Then for any 5 > 0 tno) = [a niso-ppanety 2 [7° 1-nme-yy] arin) 2 1 -FM G+] —F(— 0] 2 br ‘This however contradicts the assumption that F(s) is a distribution function. ‘The proof of the inequality '*(~a) < for sufficiently large a> 0 and m is carried in a similar manner and we so have shown that Lim F(x) = Fe) is a distribution function, We write fi(t) for the corresponding charac- teristic function, We consider next the sequence Fa'®x), it contains also a convergent subsequence and we use again the notation F(x) for this convergent subsequence. Let PAs) = Lim Fy), we show that F(x) is also a distribution function. We have for any a>Oand5>0 1-F@)> [° 1-B%e-w1dFIG) 2 1-Fi"(-4))01 -Fo™(a+s)]. For sufficiently large 8 we have Lim Fyim(—6) = F(—8) < ee 1-Fy(a+b) < 2[1- Flay} "This indicates that the left member of this inequality tends to zero a8 d increases; in a similar manner one can show that F(z) tends to7 zero as x goes to — co provided that n is sufficiently large. Thus F(x) is a distribution function and Ale) = im f%) is its characteristic function. It follows then from (6.2.4) that FO = filha) No(fi) = » We show next by means of an indirect proof that » < $Na(f). Let us therefore suppose that » 2 4Na(f), it follows then from (iii) (page 109) that Na( fa) S v. If we decompose fi and fs once more we obtain a decomposition (DY) f = aigesaes which has the property that (D*) < v. But this contradicts the defini- tion of » so that we can conclude that » < 4Na(f). ‘Therefore there exists a decomposition D of f(t) such that oD) < Nal). Each factor of D is a characteristic function without indecomposable factors and we can apply the result to a factor of D and see that » < HD) < NAS) We iterate this procedure to obtain the statement of the lemma. ‘We proceed to prove ‘Theorem 6.2.2 Let {en} be a sequence of decrea- sing positive numbers such that lim en = 0. while It follows from Lemma 6.2.1 that there exists, for each n, a decomposi- tion (Ds) F(8) = fn fo.0(8)--Fn,e, (2) such that Na(fn,1) < €n for j = 1, 2,5 x while a : lim fy = «0. Using Lemma 6.1.2 we conclude then that it is possible to find con- stants ans (j= 1, 0.» Anim = 1, 2, ...) such that li (f, (0) exit s]} = 118 where the convergence is uniform for 1 = j kx and in every finite interval |t| 3 7. We write { Sa 2) = pa, (0) explo, K2)) FO) = a) explio(e)] where on,,(0) = (0) = 0. Then (6.2.5) Lim pn.s(t) = 1 and lim [on (2) +tan,1] = 0 (uniformly in |e] $7 and for 1 1) of the geometric distribution which we considered at the end of Section 5.4 permits an even more remarkable factorization. It is easily seen that pol _ ey tte poe pag Leet We see therefore that the characteristic function of the geometric distri- bution is infinitely divisible but admits nevertheless a representation as a product of an enumerable sequence of indecomposable charac- teristic functions. Our next theorem indicates that the existence of infinitely divisible distributions with indecomposable factors is not a rare occurrence. ‘Theorem 6.2.3. An infinitely divisible characteristic function g(t) whose Lévy canonical representation is determined by the constants a= a = 0 ‘and by the functions ae mend if0 O and c > 0 are arbi- trary real constants. ‘The characteristic function g(t) is then given by fi (es-1- Let ebea real number such that 0 < « < } and introduce the functions < G-ee cu Gtee logs(t) (628) alu) = | 1 eenhere in Gd) 0 outside (0, c) wom ape[ ier TES st (6.2.8) loggy(t) = fi (1 twa G12). Clearly {2a(u)-+aa(u)] = N"(u) so that (62.9) — logglt) = loggr(t)+ logge(?). ‘The function go(?) is, according to the representation theorem 5.5.2, the characteristic function of an infinitely divisible distribution. We show next that gi(t) is also a characteristic function provided that « is sufficiently small. ‘We define a sequence of functions f(x) by means of the recurrence ae r(x) = Raa(x) als) = J Prateneutde = ff se— A We conclude from these relations and from (6.2.8a) th: (x) = if either x S 0 or x ae and me also that « ae Pa) Ber We remark that By(x) is the n-fold convolution of i(x) with itself so that [Pew aatey as = aot f° eas) xp ‘The series °° : Ein pata) is absolutely and uniformly convergent; therefore toa1y fe ot [San] expt [taxa ds].11 We write & [i aveyae and =k + antes and obtain from (6.2.8) and (6.2.10) al [os free S609 exp(-A~ity). Let G4) = facet f’ [Sa0-+n] Ps then alt= [ aede(. In order to show that gi(t) is a characteristic function we must prove that Gi(x) is a distribution function. Clearly Gy (— co) = 0, while we see from (6.2.8c) that Gi(+ 00) = gx(0) = 1, We must still show that G(x) is non-decreasing; we do this by proving that Sts) is non-negative for all x, provided that « is chosen sufficiently small. We first remark that (x) is non-negative except in the interval (G-9c < x < (4+e)c, moreover it is easily seen that Ba(x) tends to #X(c— |e) uniformly in the interval 0 < x < 2c as e tends to zero. ‘One can also show that a(x) and Ba(x) as well as 1 1 Bulx) + pp Pal) + s9Pale) are non-negative for all real x if « is sufficiently small. Let ¢ = ep be such a value. For 2 4 we can rewrite the relations defining the Ba(x) in the form Possls) = J ante—npaeat (n= 2,3, ..) and see that Bnja(x) 2 0 for all « and all n 2 2 if e = 6 ‘Therefore > Pala) 0,122 and G(x) is the distribution function whose characteristic function is 4(0).. We show next by means of an indirect proof that g(t) is not infinitely divisible. Let us therefore assume tentatively that gi() is an infinitely divisible characteristic function. Then y(t) admits a Lévy canonical representation with some non-decreasing function Ni(t). ‘The derivative N'x(t) of Ni(t) exists almost everywhere and is non- negative. It follows from the uniqueness of the canonical representation that the relation Nix) = k= Ni{x)+hao(s) 2 haa(s) is valid almost everywhere in the interval (0, ¢); but this contradicts (6.28b), Hence gi(t) cannot be infinitely divisible. It follows then from ‘Theorem 6.2.2 that g1(¢) must have an indecomposable factor and we see from (6.2.9) that this is alo true for g() so that Theorem 6.2.3 is prove Tnexactly the same way in which we proved Theorem 6.2.3 it ean be shown that an infinitely divisible characteristic function whose Lévy canonical representation is determined by the constants a =o = 0 and by the functions N(u) = 0 and M(u) = ku+c) for —c 0 has always an indecomposable factor. We can therefore reformulate our result: Corollary 1 to Theorem 6.2.3. Let f(t) be an infinitely divisible charac~ teristic function and suppose that the functions M(u) and N(w) which occur in its ‘canonical representation satisfy the following condition: There exist two positive constants k and e such that at least one of the relations M'(u) > he almost everyahere in (—c, 0) or N'(u) > almost everywhere in (0, 6) holds. Then f(@) has an indecomposable factor. We consider only the case where the condition is satisfied in the interval (0, c). ‘Then f(¢) has an infinitely divisible factor of the form required by Theorem 6.2.3 s0 that the corollary is established. ‘Let f(t) now be a characteristic function which satislies the conditions of this corollary. The function f(¢) ean be written as an infinite product f=] vor. Each factor [ f(#)] satisfies also the conditions of the Corollary, so that we obtain the following result. Corollary 2 to Theorem 6.2.3. Suppose that the infinitely divisible characteristic function f(t) satisfies the condition of the Corollary 1, then it is divisible by the product of an infinite sequence of indecomposable characteristic functions. We finally remark that the conditions of Corollary 1 are satisfied by the Gamma distribution and also by all non-normal stable distri- butions.123 6.3 Indecomposable characteristic functions ‘Wehhave already given several examples of indecomposable character- istic functions, and showed at the end of the last section that a rather wide class of infinitely divisible characteristic functions has indecomposable factors. This however is almost the only general theorem concerning indecomposable characteristic functions which we know at present. ‘There is no general method for finding the prime factors of a given characteristic function, our knowledge consists mostly of interesting special examples. In the present section we will make a few general remarks about prime factors and list also a number of remarkable decompositions. We consider next an arbitrary distribution function (which is not assumed to be of a pure type) and suppose that it is the convolution of two distribution funetions F(x) and x(x); Fea f” Fi nary). Suppose that £ is a point of increase of Fi(x) and 7 a point of increase of F(a); it is then easy to see that £+1 is a point of increase of F(x). Similarly, if € is a discontinuity point of Fi(x) and » a discontinuity Pint of (a) then fis 2 discontinuity point of F(x)? Tt can also 1e shown that every discontinuity point (resp. point of increase) J of F(x) can be written in the form Z = £-+7 where £ is a discontinuity point (resp. point of increase) of Fi(x) while 7 is a discontinuity point (resp. point of increase) of Fx(x). Let {€1} and {yj} be the (finite or enumerable) sets of discontinuity points of Fi(x) and Fx(x) respectively and denote the discontinuity ints of F(x) = Fi(x)* Fa(x) by {Zs}. Since the elements of the set Gi} can be written in the form {8:-+9)} we see that every difference £—Ee must occur among the differences of £j~ {2 at least as many times as there are different 7-values. From these considerations we obtain easily the following results: (1) Suppose that all the differences fj £4 between the discontinuity ints of a distribution function F(x) are different. Then F(x) is an indecomposable distribution function.” (2) Suppose that the distribution function (x) has at least n? dis- continuity points and that it is not possible to find n pairs of discon- tinuity points which differ by the same number. ‘Then F(x) is inde~ composable, (9) ‘These statement follow from the inequality PEL YE) —MEEn— hy) BURUEE aby) PE hs IEP he) — Ph) I> 0, y>0, B>0, b> 0. (©) We say thatthe distribution function Fx) is indecomposable if its characteristic funedion i indecomposable. Here124 (3) A finite distribution which has two discontinuity points, one at each extremity, is always indecomposable. ‘The last statement follows from the earlier remarks and from the relations lext [Fi * Fo] = lext (F)] +lext [F2] rext [Fy * Fa] = ext (Fi) +rext [Fa] which hold for the convolution of two finite distributions F and Fa In Section 3 (Corollary to ‘Theorem 3.3.3) we showed that the num- ber N of discontinuity points of a purely discrete distribution function has lower and upper bounds which are determined by the numbers of discontinuities of its factors. It can be shown that this corollary is also truc if the distribution is not purely discrete, the same limits (n-+m— 1 = NS nm) are valid also in the general case. We obtain then by in- duction the following result: (4) A purely discrete distribution which has exactly n-+1 discon- tinuity points has at most n indecomposable factors. ‘This maximum. can only be attained if the discontinuity points are the (7+ 1) consecu- tive terms of an arithmetic series. We consider next @ purely discrete distribution whose discontinuity points are the consecutive terms of a finite arithmetic series. It is then no restriction) to assume that this series consists of the integers 0, 1, 2, =, m In studying this class of distributions it is more con- venient to use the probability generating function than the characteristic function. Let ay be the saltus of the distribution at the point k (k = 0, 1, .«, n), the probability generating function is then the polynomial (63.1) PQ) = Sot (@>0, Zar= 1) “The substitution y = e transforms P(y) into the characteristic function St) = P(e), the corresponding distribution function is given by Fe) = Sash). Each decomposition of f() corresponds to a factorization of the generat- ing funetion P(y) into the product of polynomials with non-negative coefficients. If no such decomposition exists then f(t) is indecomposable. "The number of factors of F(x) reaches its possible maximum if, and only if, the generating function has n real, negative zeros. We assume next that all coefficients of the polynomial (6.3.1) are equal and write Pag) = iS ey (©) Since the decomposition properties of distribution functions are invariant under linear transformations,125 for the generating functions of this sub-class. Since Pam) = Pa(y)Pm(o") = Pm(y)Pa(™) we see that the distributions of this class admit multiple decompositions provided that the index n is not a prime number. Using the present notation we could rewrite the example at the end of Section 5.1 in the form Fey) = Px)PsG8) = Ps)FO"). IF n= PPh Peo (ur tagtuntas = h) is the decomposition of n into prime factors, then one can obtain in this way At aalagl nae different. decompositions of Px(y) into prime factors. M. Krasner B, Ranulac (1937) have shown that these are the only decompositions of Pa). ‘The problem of decomposition into prime factors is therefore com- pletely olved™ for the family of distributions with equal and equally Spaced jumps. “The indecomposable distributions which we have so far studied all had aufinite set of discontinuity points. We show now that an indecomposable distribution can have an enumerable set of discontinuity points and can also be absolutely continuous. Let {po} be the sequence of prime numbers and suppose that the distribution function (x) has its discontinuity points at £ ~ log pe (@ = 1, 2, .), Itis then clear that the diferences between discontinuity points are ail different so that F(x) is necessarily indecomposable. "The following lemma will be used in our construction of a charac- teristic function which belongs to an absolutely continuous, inde- composable distribution function Lemma 6.3.1. Let p(x) be a frequency function thick has a normal com- ponent, then p(0) > 0. If p(x) has a normal component then it can be written in the form ate inte sehen ee eee eet eee -_ Jt (+91 a 20) = pe [” exe" P*] are > 0 and the lemma is proved. (©) According to a remark by J. Hadamard, which is appended to the paper by Kragner-Ranulac, this problem war alo solved idependentiy and simultancously by A. Lignard and D. A. Raikow. PI126 Let now He a (632) f= (1-@eta = ~ 4 Since exp[its- =] as, we see that 1 © V8 J It follows then that (0) is the characteristic function of the density [sen fin-* 1 . (2) 2 ates", 20) = ae Since p(0) = 0 we see from Lemma 6.3.1 that f(t) has no normal com- ponent and therefore also no component of the form (1-28) exp(— 0/2) (8 <1). ‘The only possible factors” have then either the form (1~f)e~#/® or the ee ee ee of Theorem 2.1.1 and cannot therefore be characteristic functions. Hence f(t) is indecomposable. ‘We have thus demonstrated that an absolutely continuous distribu- tion function can be indecomposable. We add'a few remarks which refer to the presence of normal components, Since the characteristic function (6.3.2) is indecomposable we see that We) = (= aejeta is also an indecomposable characteristic function. We consider next the characteristic function (633) elt) = (MOP = (1A tate and show that it has a normal component. This is the case if silt) = (1-24 2¢8) expl~ 42/2) is a characteristic function for some A such that |A| < 1. Let os iG) = 5 [7 etenio as (©) ‘This follows from "Pheorern 8.1.2127 it is easily seen that (6.3.4) ~ae lz) grandee ze) ‘The function p,(x) is a frequency function if, and only if, the polynomial is non-negative for all real 2. It is easy to see that this is the case if AP 2 3, Moreover, for A? = 3 we see that pil +(3/2)/4] = 0 so that an(é) can have no normal component. It follows then that for A? = the characteristic function alt) is indecomposable. We have just shown that the product of two characteristic functions without normal components can have a normal component and have also obtained the factorization a(t) = I- P+ beet8 = [aot (2) fet] where the factors (1—4f2)e/4 and (1 — ##/2)8e-#"/8 are indecomposable characteristic functions. Our next example will show that the Poisson distribution has a similar property. The product of two characteristic functions which have no Poissonian factor may have a Poissonian component. Let A> 1; (1-#+2¢9) exp(—307/8) 1 Oe ete and matin ‘falt) = e i) Clearly f(t) is the characteristic function of an indecomposable law, it is not difficult to show that f(t) is also a characteristic function Moreover we can write fo) = emp Sn] oa128 Ho ~ V-FFeo[ SAE] ‘Then _ = sons (alee) a is a characteristic function which has the Poisson factor exp{(1/a)(e#—1)} while neither f(t) nor fo(t) have Poisson factors. This is trivial for the indecomposable characteristic function fi(t). If, on the other hand, At) would have a Poissonian factor exp[ia(et!—1)] (ue > 0), then aut sof-aee 2 Salt) exp[-—n(e*—1)] = ety/(1- would be a characteristic function; if we expand the right side of this equation according to powers of e'* then we see that the coefficient of et is negative, so that f(t) exp[—n(et"—1)] cannot be a characteristic function. This shows that fa() has no Poissonian component. ‘The characteristic function (6.3.2) is an example of a characteristic function of an absolutely continuous, unbounded distribution function which is indecomposable. P. Lévy has also constructed an indecompos- able characteristic function which belongs to an absolutely continuous, finite distribution. This example is reproduced in D. Dugué (19576). ‘We conclude this section with a few remarks concerning the factori- zation of the rectangular distribution. ‘The characteristic function of the rectangular distribution over the range (—1, +1) is Clearly We iterate this procedure and get ote Eh Since129 ‘we see that r(¢) can be written as the product of an infinite sequence of indecomposable factors: sint yy t (638) = = i cose We know that the rectangular distribution is not infinitely divisible; formula (6.3.5) indicates that it is possible to decompose a characteris tic function which is not infinitely divisible into a product of infinitely many indecomposable factors. ‘The decomposition (6.3.5) is not unique. Iris easy to verify that wae _ sit 2 ey, t th 3 We repeat the procedure which led to (6.3.5) and obtain sint p2cos(2t/3)41] 727 t (636) = ~~[ 3 Loser We have therefore obtained two different representations of the characteristic function 7(t) a8 an infinite product of indecomposable factors. Finally we note that r(t) is the characteristic function of a finite distribution; we will show later [Corollary 3 to Theorem 8.4.1] that the characteristic function of a finite distribution is always the product of a finite or of a denumerable sequence of indecomposable factors.7. ANALYTIC CHARACTERISTIC FUNCTIONS ‘We introduce now the class of analytic characteristic functions. This class includes many of the characteristic functions which are important in probability theory and in mathematical statistics. In the present chapter we consider the general theory and study in Chapter 8 factorization problems. In the following we will denote by t and y real variables and by x= t+iy a complex variable. We introduce the following definition, A characteristic function f(t) is said to be an analytic characteristic function if there exists a function A(z) of the complex variable = which is regular in the circle |2| < p(p > 0) and a constant A > 0 such that Ald) = fit) for |e) < A. ‘We can express this in an informal manner by saying that an analytic characteristic function is a characteristic function which coincides with a regular analytic function in some neighbourhood of the origin in the complex z-plane. As examples of distributions with analytic characteristic functions we mention: the Normal Distribution, the Gamma Distribution, the Poisson Distribution. ‘The stable distributions with exponent @ < 2 provide examples of characteristic functions which are not analytic characteristic functions, 7A The strip of regularity and the integral representation From now on we assume that f(t) is an analytic characteristic function. We know then (Corollary to Theorem 2.3.2) that all moments of the corresponding distribution exist and that it admits a Macl.aurin ex- pansion (7.1.1) < p (= complex) where p > 0 is the radius of convergence of the series. ‘We write for the even part of f(2) fda) = WU @)+4(-2)] and for the odd part of f(z) Al2) = 8Uf@)-f(-2)], 130131 ‘Then the two series (7.1.2) converge also in circles about the origin. We denote the radii of con= vergence of these series by po and pi. From the inequality [p23] 5 468+ x22) we see that ce Baw ook iy arc * 5a * ea Here and in the following we write again ay and Br for the algebraic and absolute moments of order &. ‘We conclude from (7.1.3) that p 2 po 2 p and also that the series St Zon converges for |2| < Tet ¢ be a real number and denote the radius of convergence ofthe Taylor series of fo(z) [respectively of fi(z)] around € by po(é) [respec- tively pi(é)]- According to Cortay to Theorem 2.3.1 we have [f2(E)] S axe and | f-P(E)| < Bara so that polé) 2 po(0) = po 2 p and pr(£) & pi(0) = mr & p- The Taylor series of fo(z) and of fi(2) around & converge therefore in circles of radii at least equal to p. The same is therefore true for the expansion of f(2) around £ so that f(z) is regular at least in the strip ‘Im(2)] < p. : ‘We have already mentioned that the series SB abe132 converges for |y| < p. Clearly S Bees SPH 4 ae yet Qah= > Wr |g Part) = f° eves art) for any A and |y| < p. Therefore the integral [oman exists for |y| < p, hence the integral [ara is convergent whenever |e2| < ell, where z = ¢+y. This means that the integral f ees ds) is convergent for any t and |y| < p. This integral is a regular function in its strip of convergence and agrees with f(z) for real 2, therefore it must agree with f(s) also for complex values z ~ t+éy provided y| < p. ‘The integral [" et* dF(x) converges ina strip ~a < Im(s) < +8 where « 2 p, B & p and is regular inside this strip. We write 0 f(s) = [jeans [ eee dF (x) = Py(2) + Lele). ‘The functions (2) and Ya(s) are Laplace integrals, convergent in the half-planes y > — a andy < p respectively. Let = it; then w= —iz = witty and Zilia) ~ fp e** dF(s) = O(a) is convergent for Rea) y > a Tt is Known that the Laplace transform g(s) = {6 e-# dG(t) of a monotonic function G() has a singularity at the real point ofits axis fof convergence, For proof we refer the reader to D. V. Widder (1946) [page 58, Theorem Sb]. Since F(s) is non-decreasing we can apply this result to (eo) and we conclude that —a is a singular point of (a), Thus ia is a singular point of ff). In the same way itis also seen that if isa singular point of f(a). We summarize these results and obtain the following theorem Theorem 7.1.1, Ifa characteristic fneton fs) i regular in neighbour. hood of the origin, then itis also regular ina horizontal strip and can be represented in this strip by a Fourier integral. This strip is either the whole plane, or it has one or to horizontal boundary lines. The purely imaginary ‘points on the boundary of the strip of regularity (if this strip is not the whole plane) are singular points of fe).133 From Theorem 7.1.1. we obtain immediately the following result. Corollary to Theorem 7.1.1. A necessary condition that a function, analytic in some neighbourhood of the origin, be a characteristic function is that in either half-plane the singularity nearest to the real axis be located on the imaginary axis. ‘The corollary can sometimes be used to decide whether a given func tion could be a characteristic function, We illustrate this by an example. © = [(-E)-atal ia with a2 5 > 0. = Tt is easy to see that both these functions are analytic characteristic functions. ‘Their quotient, £0, A) satisfies the elementary necessary conditions for characteristic functions, A—1) = f{O, | | S f(0) = 1 for real t. However the condition of the Corollary to Theorem 7.1.1 is violated since f(t) has no singularity on the imaginary axis while it has a pair of conjugate complex poles +b~ia. Therefore f(t) cannot be a characteristic Function. Let f(s) be an analytic characteristic function ; according to ‘Theorem 7.1.1 it can be represented as the Fourier integral fi f eedF(xs) [a < Im(z) < 8] "Therefore cae and falt) fe e ° Pa) = Fle = ip ateter P(x) We write » = a+éy where a and y are real and where a < y < By then \yr(a+iy)| s i 20 AR(x) Ifr = 2k(k = 0, 1,2, [amarin| s J” sthesvantey = 72H), is an even integer, then this becomes so that max, [fe (a+iy)] 5 |fe)] We have therefore derived the following result,