PROBABILITY

AND
MATHEMATLCAL STATISTICS
Val. U), P a s 1 (2WOhpp 169-188
STABLE PROBABILITY DISTRIBUTIONS AND
THEIR DOMAINS OF ATTRACTION: A DIRECT APPROACH
J. L. GEL WK AND L. DE HA AN (ROTTERDAM)
Abstract. The theory of stable probability distributions and their
domains of attraction is derived in a direct way (avoiding the usual
route via infinitely divisible distributions) using Fourier transforms.
Regularly varying functions play an important role in the exposition.
Key words: Domain of attraction, [generalised) regular variation,
characteristic function.
1. INTRODUCTION AND MAIN RESULTS
Let X, XI, X2 , . . . be independent random variables all of them from the
same probability distribution with distribution function F. Consider the se-
quence S,, : = XI +X2 + . . . +X, , n = 1, 2, . . . , and suppose that for some
sequences of norming constants a, > 0 and b, (n = 1, 2, . . .) the sequence
S,/a,-b, has a non-degenerate limit distribution.
In this note we shall find the general form of all the possible limit dis-
tributions and for each of these limit distributions we shall give necessary and
sufficient conditions for the distribution function F in order that S,/%-b,
converges to that particular distribution function.
The limit distributions are called stable distributions and- $he set of
distribution functions F such that S,/a,-b, converges to a particular stable
distribution is called its domain of attraction. Thus we shall identify all stable
distributions and their domains of attraction.
The indicated results have been developed more than sixty years ago. One
of the earliest systematic treatments is in Paul LCvy's famous book Thiorie de
E'addition des variables aliatoires [13]. A well-known complete description of
the theory is the book by Gnedenko and Kolmogorov [8]. Various standard
texts in probability theory offer an exposition of the subject, for example Brei-
man [2], Feller [6], Dudley [4]. In these texts the theory of stable distributions
is treated as part of the (more general and more involved) theory of infinitely
170 J. L. Geluk and L. de Haan
divisible distributions. Although infinitely divisible distributions form an
interesting and useful subject of probability theory, the stable distributions
have attracted far more attention, both in theoretical research (see e.g, the books
by Zolotarev [19] and Samorodnitsky and Taqqu [16]) as well as in applied
research (see e.g. Fama [S], Kunst [12], Mandelbrot [14], Samuelson [17]).
In contrast to the mentioned references, in this note the theory is devel-
oped ab initio, independent of results from the theory of infinitely divisible
distributions, which is too complicated to be included in a standard course of
probability theory. We have tried to present the theory of stable distributions
in a sufficiently streamlined form for presentation in such a course.
We"now set out to develop some preliminary results that allow us to
formuIate the two main theorems. We start from the limit relation:
or, equivalently,
lirn P (Sn/a,- b, < x) = G (x)
n+co
for all continuity points x of G, the distribution function of the non-degenerate
random variable I: The first question is if it is possible to have different limit
distributions for different choices of a,, and b,. Khinchine's convergence to
types theorem (Feller [6], Chapter VIII.2, Lemma 1) states that a different
choice of the norming constants can only result in a limit distribution function
of the form G (Ax + B) with A > 0 and B real. The set of all such transforms of
G will be called the type of G. From now on when we talk about a limit
distribution we shall mean the entire type so that no confusion is possible.
DEFINITION 1. Any probability distribution G that can be obtained as
a limit in (1) is called a stable distribution.
First of all we are going to reformulate the limit relation (1) in terms of the
characteristic functions (or Fourier transforms). Define for s E R the charac-
teristic functions
. -
m m
# (s) : = EeisX = j eiSX dF (x) and I) (s) : = EeisY = 1 px dG (x)
or, what is more convenient in the present setup,
2 (t) : = 4 (l/t)
and
g ( t ) : = $ (l/t)
for t E [- OO, 001\{0). By Lbvy's continuity theorem for characteristic functions
(Feller [ 6] , Chapter XV,3) relation (1) is equivalent to
(2) lim exp(-ib,/t)An(u,,t) = g(t), t ~ l - m , c~l\{O),
n+ ao
uniformly on neighborhoods of + co. Note that for t = f_ both sides equal 1.
We start with a definition and a preliminary result.
Stable probability distributions 171
DEFKNIT~ON 2. A positive measurable fuxlction f is regularly uarying if there
exists a constant y ER, the index (or order) such that
f (tx)
lim-=xY for all x > 0 .
t-.m I ( t )
In this case we will use the notation f E Ry, A function in RG is called
slowly varying. For positive measurable f the limit in (3) is either identically
0 or of the form given above.
PROPOSITION. If(1) holds, then lg (t)I2 = exp (-c ltl-")for some ol ~( 0, 21 and
c > 0. Moreover,
lim
-log 12(tx)l = X-=
for x > 0,
t j m - i ~g~a( t ) l
i.e. -1oglAl is regularly varying with index -a.
Proof. From (2) we have
locally uniformly near fa. It follows that
lirn - n log ]A (an t)J = -log lg (t)]
n+m
for each ~ E R , t # 0, for which g(t) # 0. For such t it follows that
log t)l -, 0; hence a, -, + cx, (note that a, > 0 by assumption). Moreover,
replacing n with n + 1 gives
which in combination with (5) implies an+,/a, -t 1 as n + co since convergence
in (5) is uniform on neighborhoods of infinity. Application of Lemma 9 below
then shows that the function -log 1J.l is regularly varying and its order, say
- E, has to be non-positive since lirn,,, -log [I (t)] = 0 by (5). DiGiding *, , , (5) by
its counterpart for t = 1 we find
whence log ]g (t)l/log Jg (I)] = t-" for t > 0. Since lg (t)I2 = g (t) g (- t) is an even
function, we have log lg (t)l/log lg(l)l = Itl-" for t # 0. Note that Ig (t)t2 is a
characteristic function as a product of two characteristic functions.
The restriction a > 0 stems from the fact that Y is non-degenerate. Next
we show that necessarily u < 2: the assumption a > 2 would lead to
a non-constant characteristic function with a vanishing second order derivative
at 0, which is a contradiction.
172 J. L. Geluk and L, de Haan
The proposition provides a partition of the class of stable distributions
into subclasses indexed by the parameter u.
D ~ T J O N 3. Fix a ~ ( 0 , 21. Any probability distribution function G ob-
tained as a limit in (1) and with characteristic function g satisfying
is called a stable distribution with index a or an a-stable distribution.
DEFINITION 4. The class of distribution functions F for which (I) holds with
a limit distribution G satisfying (6) is called the a-stable domain of attraction.
Notration: FED,.
The classes of distributions introduced above are useful for the rest of this
note. However, the a-stabIe distributions do not form one type. We shall see
that we need another (skewness) parameter to describe the full class of all stable
distributions. Note that the characteristic functions lg(t)I2 from (6) represent
probability distributions that are symmetric about zero.
We are now in a position to formulate the main results. Define
U(t ): =Rel (t ) and V(t):=ImA(t)
and for 0 < or < 2
m
sa = J X-'I sin xdx
0
and
The constants s, and c, can also be written in terms of the gamma func-
tion. We have for 0 < u < 2, a # 1,
an an 1
sa = r(1-a)cos- and ca = r(1-a)sin---.
2 2 I-a
In the case a = 1 one should replace the formulas with the corresponding limit
as ol+ 1 : s, = n/2 and cl = r' (1). (- r' (1) is Euler's constan&.)
Further we adopt the convention that the function (ta - l)/a is defined for
aJl t > 0, aER and reads as log t for a = 0 (by continuity). Also the function
(t ) in formula (7) below is defined to be 1 at t = 0 and (1 -a) tan @/2) is
defined to be 2/n at a = 1 (by continuity).
THEOREM 1. Suppose 0 < a < 2. Every a-stable distribution (or rather a dis-
tribution type) has a characteristic function of the following form:
with 0 < p < 1.
Stable probability distributions 173
The following statements are equivalent:
(i) FED, .
(ii) 1 - F ( t ) + F (- t) E RV-, and there exists a constant p E [O, 11such that
the following tail balance condition holds:
1 - F (t)
lim = p.
**m 1-F(t )+F(-t j
(iii) 1 - U (t ) E RV-, and there exists a constant p E [0, 11such that
(8) lim , =R\ {o}.
Further, if any of (i), (ii) or (iii) holds, then
1 - U(t>
lim
-
t - , 1 -F(t)+F(-t) -
and
v(t)-t-I f 0 ( i - ~ ( ~ ) - ~ ( - s ) ) d ~
(10) lim = (2p- l)c,.
t - m 1-F(t)+F(-t)
Remar k 1. The parameter u is the same in the three equivalent state-
ments of Theorem 1. The theorem is also true if one keeps and p fixed in the
three statements. Statement (i) then reads: (1) holds with G such that its charac-
teristic function $ is as in (7).
Remar k 2. Unlike in other texts here and in the'proof we do not treat the
case a! = 1 separately. However, for ct # 1 the statements of the theorem can be
simplified: line (7) reads (remember we need only one member of the type):
$, , , (s) = exp (- (Isla - is ( 2p - 1) tan (oln/2) [XI"- '1).
From Lemma 1 below it follows that in the case 0 < u < 1 (iii) is equivalent to
1-U(t)RV-, and V(t)-(2p-l)tan(an/2)(1-U(t)) as t + m.
If 1 < a < 2, then (iii) is equivalent to
-,-
1 - U (t) E RV-,, tV(t) + p for some constant p
and
p - t V (tj - - (2p - 1) tan (a,rc/2) t (1 - U (t))
as t -+ oo .
In view of (10) we must have p = EX, which is finite in this case.
Remar k 3. Suppose any of (i), (ii) or (iii) holds. We now indicate how to
choose the normalizing constants a, > 0 and b, in terms of either the
distribution function F or the characteristic function 4 (ie. in terms of the
functions U and v).
174 J. L. Gel uk and L. de Ha a n
The relation (1) holds with G such that the function $ is exactly as in (7)
(i,e. this distribution and not another one of the same type) i f we choose a, and
b, such that
and
See (9) and- part (iii) -(i) of the proof. Note that the above choice of the
sequence;a, is always possible since 1 - F (x) + F (- x) is regularly.varying. By
Lemma 1 again, it follows that the above choice of a, and b, implies
b, + ( 2p - 1) tan (an/2) for 0 < a < 1
and
b, - np/a, -+ (2p - 1) tan ( ~ 4 2 ) for 1 < a < 2.
It follows from relations (9) and (10) that the same limit distribution is
obtained with the alternative choices of a, and b,:
l i m l - ( a ) ) = and bn=nV(a,).
n+ m
Remar k 4. The behavior of U and V at - c~ follows from (9) and (10)
since U is an even function and V is an odd one.
The case a = 2 is covered by the following result:
THEOREM 2. Every 2-stable distribution (or rather a distribution type) has
a characteristic function of the following form:
corresponding to the normal distribution.
The following statements are equivalent:
(i) FED, .
(ii) The function H, (t ) : = So u (1 - F (u) + F (- u)) du is slowly uarying.
(iii) 1 - U (t ) E RV- and ,,?
If (i) hoids, then as t -, ao
(13)
1 - U (t ) -- HI (t)/t2
and
where p = EX.
Stable probability distributions 175
Remark. The behavior of U and V at -co follows from (13) and (14)
since U is an even function and V is an odd one.
Using the results of Theorems 1 and 2 one verifies easily that the stable
distribution functions are precisely those distribution functions G such that if
Y, Yl, Y2, .. . are i.i.d. G, then there exist constants A, > 0 and B,, such that, for
n 2 1, (Yl + Y2 + . . . + ',)/A, -B, has the same distribution as Y
. -
2. AUXILJARY RESULTS
7i z.
Before we prove the theorems we collect some basic facts about regularly
varying functions in a sequence of lemmas. Lemmas 1-7 are standard results
that are useful in other contexts as well. Lemma 8 (preparing for the use of
Lebesgue's theorem on dominated convergence) and Lemma 9 (on replacing
a sequence by a continuous variable in the limit relation) are specific for the
present setup.
LEMMA 1 (see [7], Theorems 1.9 and 1.10). Suppose f is a measurable
function and there is a positive function a such that for all x > 0
f ( t x) - f ( t ) xY-1
lim . =-
t +m a( t ) Y
where y is a real parameter. (The right-hand side is interpreted as log x for y = 0.)
I f (15) holds with y > 0, then a(t ) -- yf (t ) as t -, co, both functions tend to
inJinity, and hence f E RTj.
If (15) holds with y < 0, then lim,, , f (t ) = : f (a) exists and
Hence f (m)- f ( t ) is regularly varying of order y.
If (15) holds with y = 0, then a (t ) = o ( f (t)) (t + CQ) and a is regularly
varying of order 0, i.e. slowly varying. Also lim,, , f (t ) =: f (m) exists (finite or
+ a). I f f ( m) = a, then f E R V ~ . I f f (a) < a, then f ( m) - f ( t ) is-slowly * ern vary-
ins and a( t ) = o ( f ( m) - f (t)) as t+ ao.
Remar k 1. For f measurable the limit in (15), if not identically zero, is
necessarily of the given form.
Re ma r k 2. If the limit in (15) exists and is identically 0 for x > 0 with
QERV, , then
for y >O, f (t ) =o( a( t ) ) as t -r CQ,
for y < 0, f (co) exists and f ( m) - f (t)= o(a(t )) as t -+ a.
LEMMA 2 (see 171, Theorem 1.20). Suppose that the function f is integrable
over finite intervals and that (15) holds with y = 0.
176 9. L. Geluk and L, de Haan
(i) Lei k: R+ + R be a function which is bounded on [0, A] for some A > 0.
Then as t += m
(ii) Let k: R+ + R be a function such that jy .r%(s)ds < 03 for some
A, E > 0. Then
LE& 3 (cf. Bingham et al. [l], Chapter 4). Suppose that the function g is
integrable over Jinite intervals and that (3) holds with f positive. Assume
g ( t ) / f ( t ) + c 3 o as t + oo.
(i) Suppose y > -1 in (3). Let k: Rf + R be a function which is bounded
on [ O, A] for some A > 0. Then as t -, co
(ii) Let k: R+ + R be a function such that xY+' Ik(X)I dx < m for some
A, E > 0. Then
Re mark. If the limit in (15) is identically zero, then the limit in Lem-
ma 2 is also identically zero.
LEMMA 4 (see e.g. Ibragimov and Linnik [lo], the proof of Lemma 2.6.1).
Suppose g is a non-increasing function and g(t)/ f (t ) + C E [0, oo) as t + oo for
some function f E RV-, (0 < a < 2). For any E > 0 there exist constants A. and
to such that for all t 2 to and A > A,
im
Proof. By the second mean value theorem for all B > A
g (tx) g ttA)
j-sinxdx = -jsinxdx for some CECA, 31,
A f (t ) f (0 A
and hence
The proof of the first statement is complete since the right-hand side tends to
zero as A + co. The proof of the second statement is similar.
Stable probability distributions 177
Next we give a version of the monotone density theorem (see e.g. Bingham
et al. [I], Chapter 1.7.3).
LEMMA 5. Iff (t) : = fo $ ( s) ds is regularly varying with index a > O and $ is
monotone, then $ E RK',- I.
In the sequel we need a modification of the above lemma.
LEMMA 6. Suppose f is nun-decreasing. If there exists fl 2 0 and a positiue
function o s wh that the function 7 d&ned by f ( t ) := t-' Sof ( s) ds snti.$es
- - T( t x) - T( t ) xP- l
T - 3-
B
for x > O , t - +m,
a ( t )
then
f (tx) - f It)
xfl - 1
+-
P
for x > O , t + c o .
a ( t )
Proof. Define the function $ by
It is easy to see that this definition implies
Hence we have for x > 0 and t + co
" 9 (t s) as f ( t x) - f ( t ) xfl - 1
-
itif if -
+-,
P a (0
Since $ is non-decreasing, for x > 1 the left-hand side is at least
and hence
$( t ) xfl -1
lim sup - 6
t+, t a( t ) P( 1 - x - I )
This shows that lim sup,, , $ (t )/ t a ( t ) < 1 by letting x J 1. Starting with
0 < x < 1 and applying a similar inequality we get lim id,, , 9 (t )/ t a ( t ) 2 I. It
follows that $ ( t ) - t a( t ) ( t + a), which combined with (17) gives
12 - PAMS 20.1
178 J. L. Gel uk and L, de Haan
Hence as t + og
f (tx) -f (t) - - J(t x) -f(t)
xfi- 1
+o(l) -+-.
B a ( t ) a (t)
Remark. If the limit in (16) is identically zero, the corresponding limit for
f is also identically zero.
The next lemma is a special case of Feller [dl, Chapter VIII.9, Theorem 2.
LEMMA 7. Suppose Fa is a distribution function on [0, a). Thefunction U2
is defined. by U2 (t) := so s2 dF,(s]. Then U, ERV, i f and only if
Remark. An integration by parts shows that the above statements are
also equivalent to
t" (1 - F, (0)
-0 ( t - m) .
S (1 - Fo(s)) ds
The following result is a modification of a result in Pitman [15].
LEMMA 8. Assume the conditions of Theorem 1 (iii) (or Theorem 2 (iii)) are
satisfied. For every y > 0 there is a constant c such that for euery T > 0 and
o < x < y
and
Proof. Since the other statements can be proved similarly, we only prove
the first statement. Note that if (8) holds with 0 < a d 1, there exists to such
that IV(l/t)l 6 tat' for 0 < t < to < 1. Define
x sin txdt
li!y costxdt 1 =: L, +L2,
< c.
where L1 and Lz are the integrals over (0, t o) and (to, T) , respectively. It follows
that L, is bounded if 0 < a < 1. For 1 < a < 2 it follows from (8) that
lim,,, tV(t) =: p exists, in the case ol = 2 this follows from (13). Hence L, is
bounded. Next we estimate L,. Integration by parts gives
m 0 m
V (l/t) = J sin t xdF (x) = j sin txdF (x) + j sin txd (F (x) - 1)
-03 - m 0
m
= t 1 K ( y) cos tydy ,
0
Stable probability distributions 179
where K(y) := 1-Fly)-F(-y). Hence
T m
(191
Lz = 1 1 K (y) cos ty cos txdydt.
t o 0
Using the second mean value theorem for each M > 0 we infer that there exists
{ E 10, MI such that
M
11(1 - F (y)) cos ty cos txdyl
n
v
5
' - = (1 - F (0)) lcos tx j cos tydy) d 2/ t d 2/to for to d t G T.
0
Note that a similar argument holds for the integral containing F(-y). Hence
we may reverse the order of integration in (19) to find
m T
L, = j j K (y) cos ty cos txdtdy
0 to
1 cos T {x + y) cos T fx - y) cos to (x + y) cos to (x - y)
+
- -
X + Y x-Y X + Y x-Y
The latter integral is bounded since
exists as a finite (semiconvergent) integral for all real x.
LEMMA 9 (the extension of Kendall [ll], cf. Bingham et al. [I], Chap-
ter 1.9). Suppose
lim sup x, = cc, lim sup x,, l/xn = 1
n+ w n+ m
and f is a continuous function.
1. Suppose 0 < b < c < m and for some sequence a,
IF*
a,f(Ax,)+$(A.)~(O,m) f o r a l l L~ ( b , c ) a s n h m ;
then f varies regularly.
2. Suppose 0 < b < c < m, the function a is regularly varying and
lim f(Axn)-f (xn) + i~ (A)
for aii A E ( b, c) ;
n4s, a(x,)
then there exist constants c, y ER such that
180 J. L. Geluk and L. de Haan
Proof. The continuity of f is the key assumption.
1. With V = (b, c) there exists a non-empty interval K such that
Vnu- ' V#O for all UEK. If t , u t ~ K we have
Hence if we write f * (t) = f (ud/ f (4 for u > 0 fixed and x: = log x,, then
f * (t +x$) converges as n -t oo for all t in a non-empty open interval J. Choose
E > 0 and .define for k E 2, m E N the closed sets
By Baire's category theorem (see Hewitt and Stromberg [9]), since J is
non-empty and open, one of the sets C,,, contains an open interval I. This
means that
kg- E 6 f*(t+x,*) < ~ E + E for n 2 m, ~ E I .
Since by assumption xz -, co, xi + - x : + 0, it follows that U, xz + I con-
tains an interval of the form [ t o, m]; hence
KE- E G f *(t) < ~ E + E for all t 2 to.
Hence lim,,, f * (t) exists and is finite and positive for all UEK, i.e.
exists and is finite for all UEK. It follows that the function f is regularly
varying.
2. In a similar way as above, using the fact that a is regularly varying we
obtain for u > 0 fixed and all t in a non-empty open interval
lim
f Ixn t ~ ) -f ( ~ n U)
= $* (t).
n-+m ~ ( x , u )
Define for u > 0 fixed the function
*em
(with xX = logx, as before). Then in a similar way as above we can show that
lim,, , f * (t) exists and is finite.
3. PROOF OF THE MAIN THEOREMS
Proof of Theorern 1. We first prove the equivalence of (i), (ii) and (iii). In the
part (iii) -(i) of the proof we obtain the characterization (7) (see (45)).
Stable probability distributions 181
(i) +- (iii). It follows from (2) that for all real t + 0
(20)
lim n log X (LZ, t)- ib,/t = log g (t),
n+ m
and hence
lim aR (a,, t ) = - Re (log g (t)),
n+ m
and
(22)
lim nI (a, t ) - bn/t = Irn (log (t)),
. - .
m- r m
. ~
where R ( t ) =TL Re log R (t) and I (t) = Irn log A ( t ) (Re and Im denote the real
and the imaginary part, respectively). Note that there exists a unique version of
log R (log g) satisfying log ll (t) 4 1 (log g (t) -, 1) as t + cn (see e.g. Feller [6],
Chapter XV).
Application of Lemma 9 (note that a,, + a, a,+ Jan -+ 1 (TI + co) as in the
proof of the Proposition) shows that the function R is regularly varying and
- Re(1ogg (t)) = It1-" for t # 0.
Next we focus on (22). By setting t = 1 we get
lim n1 (a,) - b, = Irn log g (I),
I n'm
i and hence
i
Combining this with (21) for t = 1, we get for all real t # 0
I
(23)
t l (a, t ) - a, I (a,)
t Im log g (t) - Im log g (1)
litq - - =: z(t).
,-tm anR(a,J - Re log g (1)
In a similar way, by Lemma 9 this implies
where C E R is a constant. Since RE RV-,, it follows from Lemma 1 that c = 0
or, if c + 0, then y = 1 - E. Using the fact that I is an odd function we now have
We have now (iii) with 1 -U replaced with R and V replaced with I.
Since for complex z, lzl < 1/2,
(ez- 1 -z[ < 1zI2,
182 J. L. Geluk and L. de Haan
we have
In (a,) - 1 -log A (a31 < llog a (a,)l
for n sufficiently large. From (21) and (22) we obtain
It follows that
(27) lim nl-logA(a,)-l+A(a,)l = 0;
n-t m
. - .
hence we-may replace -1ogA in (20) with 1 -A. Consequently, we can repeat
the above argument with -Re log 1 replaced with 1 - U and Im log A replaced
with V to obtain (iii).
(ii) -(iii). Define the functions W and K by
H( t ) : =l - F( t ) +F( - t j and K(t ): =l -F(t j -F(-t ).
First we prove that
~(t )-1-i t p1 So ~( s ) d s
lim = -sa+i(2p-l)c,.
t-+m H (t )
Now for any A > 0
~ ( t ) - 1 -i t -' f, ~(s)ds
H (0
A H(tx) A K (tx) A K (tx)
= -5sinx-dx-iJ(1-cosx)-dx+iJ-dx
o H(t) 0 H (t) 1 H(t)
* H(tx) " K(tx)hrm Y
- J sinx-dx+i j cosx-
A H (t) A H (0
Take E > 0. By Lemma 4 the last two integrals are less than E for t > to and
A > A,. For fixed A > 0 the first three integrals converge by Lemma 3 to
Now (28) follows if we take A + oo. By separating the real and imaginary parts
in (28) we get the limiting behavior as t -+ + co in (9) and (10). The limiting
behavior as t + - m foI10ws since U is an even function and V an odd one.
Obviously, (9) implies that 1 - U E RV-/_, (since HE RV-,). Note that s, # 0 for
0 <a < 2. Now (10) implies that for x > 0
tx V (tx) - I ~ K (s) ds
Iim = (2p-l)caxl-a
z* m t H (t)
Stable probability distributions 183
(use H E RV-,). Combination with (10) gives
(29) lim
txV(tx)-tV(t) - - ' K( l s I ds = (Zp-l)ca(xl-'-l).
t tH(t) I H (t)
Note that the integral on the left-hand side converges to
by Lemma 3. Now (8) follows since 1 - U satisfies (9).
- -
(iii) -(ii)..rn this part of the proof c denotes a constant which may take
daerent values at each occurrence. In order to prove the results in this part we
make use of Lkvy's inversion relation
valid for all x, x + h for which F is continuous. See e.g. Chow and Teicher [3].
Note that the above integral is to be understood as the limit as A + m of the
integral over (-A, +A). A similar remark holds for the other inversion
integrals below. Using the relation (301, the equality +( t ) = A(l/t) = U(l/t)+
+iV(l/t) and the fact that
" sin x n
J -dx = -,
O X 2
we obtain the following inversion formula for H:
2 " 1-U(l/t)
H(x) = - j sintxdt, x > 0 .
no t
-First we prove that H is regularly varying with order - u. For t $ 0 define
where H, (t) = st xH (x) dx = $ fi H(&) du as in Theorem 2 (ii). By (31) it
follows that
2 " 1 - U(l/s)
=;I J x sin sxdsdx.
0 0 s
From Lemma 8 it follows that we may reverse the order of integration, and so
2 " sin ts - ts cos ts
1 ) = - 1 - 1 9 ds.
no
184 J. L. Gel uk and L. de Haan
Since this integral is absolutely convergent, by Fubini's theorem we have
(32)
Z m * sin xs - xs cos xs
= ; J - W/ s ) ]
s3
dxds
0 0
2"O 2 (1 - cos ts) - ts sin ts
= - J (1 - U(l/s))
s4
ds.
n o
Hence
H,(t) = 27 1-U(t/s)2(l-coss)-ssins - .
s4
ds.
- t3(1 - U(t)) X 0 1 - U(t1
Since 1 - U is regularly varying with index -a, in view of Lemma 3 (substitute
s = x-l) the right-hand side converges to
As a consequence, H2 E RV3 -, By the monotone density theorem (Lemma 5) it
follows that Hl ~ R V 2 - , ; then HERV-, . In order to prove the tail balance
condition we need an inversion relation for K. As for the inversion relation for
H we obtain
(33)
2 " V(l/t)
K(x)-K(y) =;I - (cos tx - cos ty) dt,
x, y > 0;
o t
hence by Lemma 8 the function
2 v (lit)
K(x)-; I - cos txdt
o t
is constant for x > 0. The constant is necessarily 0. This follows by taking the
limit as x + co and applying the Riemann-Lebesgue lemma in (18). See e.g.
Feller [6], Chapter XV.4. For y > 0 we have
Y
2 a, V(l/t)
v n
K,(y) := j ~ ( x ) d x = -J J- cos txdtdx
0 no0 t
2 V(l/t) 2 V (l/t) sin ty
=- J J-costxd~dt = -J-- dt.
n o 0 t n o t t
Interchanging the order of integration is justified by Lemma 8. Now we inte-
grate once more, use (34) and Lemma 8 to find for t > 0
1 *
2 " v (l/s) (1 - cos st)
Kl(t) : = ; [ ~ ~ ( y ) d y =- 1 ds.
0 nt 0 s
Stable nrobabilitv distributions 185
It follows that for b, t > 0
(35)
K, (bt) - K, {t) 2 btsV(bts)- tsV(ts) a(ts)
= - j
- (1 - cos( s- l ) ) as,
a ( t ) no a (ts) a (0
where a (t ) : = t ( 1 - U (t)).
Taking the limit as t + co, using (8) and Lemmas 1, 2 and 3 we find for
x > 0
lim
R1 ( t x) - Kl ( t ) xl i a- 1
= C
t+-m a (t ) 1-a '
Application gf Lemma 6 then shows that
It follows from (9), since HERV-,, that as t + co for x > 0
Adding both sides of (36) and (38) we obtain for x > 0, as t -+ a,
t 71- F( t s )
-=- 1
ds -, c
I a @) 1-a '
In view of (9) this implies
For x > 1 the left-hand side is at most
Hence
1 - F( t ) x l - a - l
lim inf 2 c -,,,
t +m 1- F( t ) +P( - t ) ( 1 - ~ ) ( x - l ) .
Letting x J 1 then gives
lim inf
1 - F (t)
2 c.
t+, 1- F( t ) +F( - t )
Starting with 0 < x < 1 in (39) and applying similar inequalities we obtain
where c equals the constant in (40).
186 J. L. Geluk and L. de Haan
(iii) - (i). Define the sequence a,, n = 1, 2, . . . , such that
Note that this is possible since 1 - F (t)+ F (- t) is regularly varying. Moreover,
a, + co as n + m. By (9) we have, since 1-UERV-,,
(42) limn(1-U(ra,t))=Itl-" for a11 t gR, t #O.
n+ m
Define the sequence b,, n = l , 2, . . . , by
Then as n + for all t E W, t # 0, by (9) and (42)
l a. t v( ant ) - onV( ~ I ~(an)-a; ' J?~(s)ds 29-1
N - +- --
El (a,)
C,.
t a (1 - u a ) ) t S a ts,
Substituting relations (8) and (10) on the right-hand side we find
(43)
l t y a -
R+ OO l -a
Combining (42) and (43) we get
i ( 2p - 1) ltll-u-l
(44) lim n(1-l(a,t))+ib,/t = Itl-"-
n+ m tsu l -a
We want to prove that
(45)
lim A" (an t ) exp (- ib,/t)
n+ m
Now for lzl < 112 we have 18- 1 -zl < 1zI2. In particular, for fixed t ER, t # 0,
there exists no such that for n > no
Iexp(-l+i(a, t))-~(a, t)( < 11-ll(a,t)12,
and hence
So it is sufficient to prove that n 11-R. (an t)I2 + 0 as n + CO. This follows from
(42) and (43).
Stable probability distributions 187
Roof of Theorem 2
(i) * (iii). Following the reasoning of the proof of Theorem 1, part (i) =
(iii), we find that 1 - UERV-,. Since (24) now holds with y = - 1, application
of Lemma 1 (or its extended form from Remark 2 following the lemma) shows
that lim,,, tl(t) =: co exists. Hence (25) holds with c possibly 0 and the
right-hand side equals
z (t) = t Imlogg (t)- Irn log g (1) = - c (It1 - - 1).
Since - Re log g (t) = t-2, t # 0, we have
. -
I g(t) = e ~ p ( - t - ~ + i t - l ( c ~ +c41tlf1)},
where c3 and 12, are constants. Since any bounded continuous function w with
w(0) = 1 is a characteristic function only if for all x and E > 0
(see Feller [6], Chapter XIX.21, we must have c4 = 0 (see Steutel [18]). Hence
Il.(t) = g(t-') = exp (- t2) and (24) holds with c = 0. Remark 2 following Lem-
ma 1 now shows that lim,,, tV(t) =: p exists and (12) holds.
(ii) * (iii). By Lemma 7 (take Fo(t) = 1 -H(t), t 2 0), tZH(t)/H, (t) -+ 0 as
t + a. Note that
A (t) - 1 - ip/t + H1 (t)/t2
Hl (t)/t2
( t ~ ) ~ H (tx) sin x - x (tx)' H (tx) sin x
= - j dx- J - ax
0 Hl(t) x2 1 HI (4 x2
' ( t ~ ) ~ K (tx) cos x - 1 ( t ~ ) ~ K (tx) cos x - 1
+ i j dxf i j dx.
* Hl (t) x2 1 Hl(t) x2
Application of Lemma 3 shows that the integrals on the right-hand side all
tend to.zero as t +a. Now (13) and (14) follow by taking the red-and imagina-
ry part and (iii) follows from (ii), (13) and (14).
.ern
(iii)-(ii). Compared to the corresponding part in the proof of Theo-
rem 1 we have to integrate once more in order to get an absolutely convergent
integral. For the function H, defined by H3 (t) = So H, (s) ds an expression simi-
lar to (32) can be given. A similar calculation shows that HI is slowly varying.
(iii) *(i). With the sequences a, and b,, n = 1, 2, . . ., defined by
nH1(an)+l as n + m and b.=np/an,
a:
the proof is similar to the proof of the corresponding part of Theorem 1. We
omit the details.
188 J. L. Gel uk and L. de Haan
REFERENCES
[I] N. H. Bingham, C. M. Gol di e and J. L. Teugels, Regular variation, Encyclopedia Math.
Appl. 27, Cambridge Univ. Press, 1987.
[Z] L. Breiman, Probability, Addison-Wesley, Reading MA, 1968.
[3] Y. S. Chow and H. Teicher, Probability Theory, Independence. Interchangeability, Mmtin-
gales, Springer, Berlin 1978.
[4] R. M. Dudley, Real Analysis and Probability, Wadsworth and Brooks/Cole, 1989.
[5] E. Fama, The behavior of stock prices, Journal of Business 38 (19651, pp. 34105.
161 W. Feller, An Introduction to Probability Theory nnd Its Applications 2,2nd ed., Wiby, New
York 1971.
[7] 1. L. Ti el u k and L. de Ha an, Regulnr Variation. Extensions and Tauberian Theorems, CWI
Tract 40, Amsterdam 1987.
[8 1 B. V. Gneden ko and A. N. Kol mogo r o v, Limit Distributions f i r Sum of Independent
Random Variables, Addison-Wesley, Reading MA, 1954.
[9] E. Hewi t t and K. S t romberg, Real and Abstract Analysis, Springer, Berlin 1969.
[lo] I. A. I br agi mov and Yu. V. Linnik, Independent and Stationary Sequences of Random
Variables, Wolters-Noordhoff, Groningen 1971.
[ll] D. G. Kendall, Delphic semigroups, infinitely divisible regenerative phenomena and the arith-
metic of pfunctions, 2. Wahrsch. verw. Gebiete 9 (1968), pp. 163-195.
1121 R. M. Kunst , Apparently stable increments infinance data: Could ARCH eflects be the cause?,
J. Statist. Comput. Simulation 45 (1993), pp. 121-127.
[I31 P. Li vy, Thkorie de l'addrtion des variables al.4utoires, 2nd ed., Gauthier Villars, Paris 1954.
[I41 B. Mandel br ot , The variation of certain speculative prices, Journal of Business 36 (1963),
pp. 394419.
[15] E. J. G. Pi t man, On the behaviour ofthe characteristic function of a probability distribution in
the neighbowhood of the origin, J. Austral. Math. Soc. Ser. A 29 (1968), pp. 337-347.
C16] G. Samo r odni t s k y and M. S. Taqqu, Stable non-Gaussian random processes, Chapman
and Hall, London 1994.
[17] P. Sam u e 1 son, Eficient portfolio selection for Par et eUvy investments, Journal of Financial
and Quantitative Analysis 2 (1967), pp. 107-117.
[I81 F. W. Steutel, Problem 979, Problem Section, Nieuw. Arch. Wisk. 15 (3), p. 252.
[I91 V. M. Zol otarev, One-dimensional stable distributions, Transl. Math. Monographs, Val. 65,
American Mathematical Society, 1986.
Econometric Institute
Erasmus University Rotterdam
P.O. Box 1738 z,n
NL-3000 DR Rotterdam, The Netherlands
Received on 25.10.1999;
revised version on 17.4.2000