COMPUTING LAPLACE TRANSFORMS FOR NUMERICAL INVERSION VIA CONTINUED FRACTIONS

by Joseph Abate1 AT&T retired

Ward Whitt2 AT&T Labs

May 14, 1998

1 900 Hammond Road, Ridgewood, 2 Room A117, AT&T Labs,

NJ 07450-2908

180 Park Avenue, Building 103, Florham Park, NJ 07932-0971; email: wow@research.att.com

Abstract

It is often possible to e ectively calculate cumulative distribution functions and other quantities of interest by numerically inverting Laplace transforms. However, to do so it is necessary to compute the Laplace transform values. Unfortunately, convenient explicit expressions for required transforms are often unavailable. In that event, we show that it is sometimes possible to nd continued-fraction representations for required Laplace transforms that can serve as a basis for computing the transform values needed in the inversion algorithm. This property is very likely to prevail for completely monotone probability density functions, because their Laplace transforms have special continued fractions called S fractions, which have desirable convergence properties. We illustrate the approach by considering applications to compute rst-passage-time distributions in birth-and-death processes and various cumulative distribution functions with non-exponential tails, which can be used to model service-time distributions in queueing models.

Subject classi cations: Probability distributions: calculation by transform inversion. Queues, algorithms: Laplace transform inversion. mathematics, functions: Laplace transforms and continued fractions.

Other keywords: computational probability, numerical transform inversion, continued fractions, Laplace transforms, S fractions, complete monotonicity, Pade approximants, cumulative distribution function, birth-and-death process.

1. Introduction
Many descriptive quantities of interest in queueing models and other probability models arising in operations research can be e ectively computed by numerically inverting Laplace transforms; see Abate, Choudhury and Whitt 1], Abate and Whitt 4], 5] and references therein. The biggest challenge in this approach, when there is a challenge, is usually computing the required Laplace transform values, because convenient closed-form expressions for Laplace transforms often are not available. In this paper we point out that continued fractions can sometimes serve as a basis for e ectively computing the required Laplace transform values needed in the inversion algorithms. A simple motivating example is the steady-state waiting-time distribution in the M/G/1 queue. The classical Pollaczek-Khintchine (transform) formula gives the Laplace transform of the steady-state waiting-time density function in terms of the Laplace transform of the servicetime density function. Thus we can compute the waiting-time transform values in order to compute the waiting-time density function or cumulative distribution function by numerical inversion whenever we can compute the service-time transform values. A possible di culty, however, is that we might want to consider service-time density functions for which convenient explicit expressions for the Laplace transform are unavailable. Indeed, this di culty often arises when we consider distributions which have non-exponential tails, e.g., which cannot be represented as phase-type distributions. The present paper provides a way to address this problem: Under favorable circumstances, we may be able to construct a continued-fraction representation of the service-time Laplace transform that enables us to compute the servicetime Laplace transform values, which in turn enables us to compute the waiting-time Laplace transform values needed to perform the desired numerical inversion. For background on continued fractions and their use for numerical computation, see Baker and Graves-Morris 12], Bender and Orszag 13], Chapter 12 of Henrici 26], Jones and Thron 28], Section 5.2 of Press, Flannery, Teukolsky and Vetterling 32] and Wall 35]. Applications of continued fractions in statistics and applied probability are described in Bowman and Shenton 15] and Bordes and Roehner 14]. More recently, Guillemin and Pinchon 20], 21], 22], 23] have used continued fractions to analytically derive important properties of queueing models. A summary of that work is contained in Dupuis and Guillemin 16]. However, continued fractions evidently have not been suggested previously as a way to numerically compute transform values in order to perform numerical transform inversion. 1

The use of continued fractions is an alternative to computation of Laplace transforms via series representations, which we recently discussed in Abate and Whitt 9]. We make an explicit numerical comparison to show that continued fractions can be far superior in some circumstances, even when the series converges geometrically. (See Section 7.) Here is how the rest of this paper is organized. In Section 2 we brie y de ne continued fractions and specify the basic recursive algorithm for numerical computation. In Section 3 we discuss the relation between continued fractions and power series. There we show how to compute the continued fraction elements from the moments of a probability distribution (which are related to the coe cients of a power series | the moment generating function). In Section 4 we point out that completely monotone probability density functions can be identi ed with special continued fractions called S fractions, which have nice convergence properties. In Section 5 we show how continued fractions can be used to compute the Laplace transforms of rst passage times in birth-and-death processes. We can exploit S fractions for this purpose because rst passage times to neighboring states have completely monotone probability density functions. The rest of the paper is devoted to numerical examples. In Section 6 we consider the M/M/1 busy period, which is a special case of a rst passage time in a birth-and-death process. In Section 7 we consider the beta mixture of exponential (BME) distributions from Abate and Whitt 8] 9] and show that continued fractions can be much more e ective for computing Laplace transform values than the previously considered series representations. In Section 8 we show how the continued fractions associated with the BME distributions can be used to compute the Laplace transforms of other probability distributions related to the BME distributions, including a Pareto distribution. Finally, in Section 9 we discuss continued fraction representations of other probability distributions.

2. Continued Fractions
An (in nite) continued fraction (CF) associated with a sequence fan : n 1g of partial numerators and a sequence fbn : n 1g of partial denominators, which are complex numbers with an 6= 0 for all n, often called elements, is the sequence fwn : n 1g, where

wn = t1 t2 : : : tn (0); n 1 ;
and
k tk (u) = b a+ u ; k 1 ; k

(2.1) (2.2)

i.e., wn is the n-fold composition of the mappings tk (u) in (2.2) applied to 0, called the nth approximant. If limn!1 wn = w, then the CF is said to be (properly) convergent and the limit w is called the value of the CF. We write
2 w = 1 an or w = ba1 ba+ ba3 n=1 b n 1+ 2 3+

(2.3)

When we consider Laplace transforms of probability density functions, the CF elements an and bn will be functions of the complex variable s. In particular, we will consider special CFs called S fractions (S for Stieltjes), which can be expressed as 1 w w(s) = 1+ a2 s a3 s a4 s 1+ 1+ 1+ (2.4)

has the same approximants as the CF in (2.3); see p. 478 of Henrici 26]. We call such CFs equivalent and use the notation '. It is signi cant that there is a relatively simple recursion for calculating the successive approximants of a CF, due to Euler in 1737. In particular, given the CF in (2.3),

where ak is real and positive for all k. However, S fractions may appear in other forms, because CFs have many equivalent representations. Indeed, for any sequence of complex numbers fcn : n 0g with c0 = 1, the CF 1 cn?1 cn an (2.5) n=1 c b
n n

P wn = Qn ;
n

(2.6)

where P0 = 0, P1 = a1 , Q0 = 1, Q2 = b1 and

Pn = bn Pn?1 + an Pn?2 Qn = bn Qn?1 + anPn?2

(2.7)

for n 2. In performing numerical calculations, it is prudent to renormalize after, say, every 10 iterations by dividing the current values of Pk , Qk , Pk?1 and Qk?1 all by Qk , and then proceed. We will be interested in the special case of S fractions, as in (2.4). Based on our computational experience, we conclude that the S fraction converges rapidly and is easy to compute if an = O(1) as n ! 1. If an = O(n) as n ! 1, then the S fraction converges, but more slowly and requires more work to calculate. If the elements grow much faster, then computation is likely to be infeasible. (We will give examples later.) The S fraction with an = nk can be shown to be convergent if and only if k 2; see p. 486 of Henrici 26]. 3

3. Power Series and Continued Fractions

Continued fractions are intimately related to power series. This relationship is useful in probability applications because the moments of probability distributions can be regarded as coe cients of a power series, namely, the moment generating function (mgf). Let f be a probability density function (pdf) on the nonnegative real line with associated R cumulative distribution function (cdf) F (t) 0t f (u)du, t 0, and associated complementary cdf (ccdf) F c (t) 1 ? F (t), t 0. Assume that the pdf f has nite moments of all orders, i.e.,

mn (f )

1
0
Z

tn f (t)dt; n 1 ;
1

(3.1)

and let f^ be the Laplace transform (LT) of f , i.e.,

f^(s)

e?st f (t)dt ;

(3.2)

where s is a complex variable. We often are most interested in computing the cdf F (t) or ccdf ^ ^ F c (t), which we can do via their LTs F (s) = f^(s)=s and F c (s) = (1 ? f^(s))=s. The associated mgf is f^(?s). The Laplace transform f^ is analytic for all s with Re(s) > 0. The nth moment of f can be recovered from the nth derivative of the transform, i.e.,

mn(f ) = ? dds f^(s) s=0 ;

(n)

(3.3)

where limits are taken through real positive s, and the LT can be represented as a formal power (Maclaurin or Taylor) series

f^(s) =
where c0 = 1 and

1 X
n=0

cn sn ;

(3.4) (3.5)

by which we mean that, for each N

cn = (?1)n mn(!f ) ; n 1 ; n
1,
N

X f^(s) ? cn sn = O(sN +1 ) as s ! 0 :

n=0

(3.6)

However, in general we cannot conclude that the power series (3.4) has a positive radius of convergence. It will if and only if the LT is analytic at 0, which is not always the case.

Given a formal power series such as (3.4), we can construct rational approximants, called Pade approximants, that match the coe cients of (3.4) as far as possible; see Chapter 1 of ^ Baker and Graves-Morris 12]. The L=M ] Pade approximant to f is the rational function
L + 1 L=M ] a0+ bas s++ ++b aLss ; 1 1 M M

(3.7)

where the rst L + M +1 coe cients of the Maclaurin series of (3.7) match the rst L + M +1 coe cients in (3.4). Given a formal power series such as (3.4), it is also possible to construct an associated CF of the form n (3.8) f^(s) = a0 + 1 a1 s ; n=1 whose approximants also have Maclaurin series that match the initial terms of the power series (3.4). (A CF with the variable s in each coe cient as in (3.8) is sometimes called a RITZ fraction; e.g., see p. 515 of Henrici 26].) To obtain the CF representation, we use the notion of the reciprocal of a power series to create an appropriate recursion. In particular, we consider the reciprocal of the series 1 + cc2s + c3s + c
1 1 2

= (1 + c(1)s + c(1)s2 + 1 2

)?1

(3.9) (3.10)

to obtain

1 X
n=0

cnsn = c0 +
(1) 2

from (3.4). Next consider the reciprocal of the series

c1 (1)s + c(1)s2 + 1 + c1 2
= (1 + c(2)s + c(2)s2 + )?1 1 2

s 1 + c2(1)s + c3 (1) + c1 c1
to obtain
1 X
n=0

(1)

(3.11)

: (3.12) 1 + 1+c sc+c s s + Proceeding by induction, we obtain (3.8) with a0 = c0, a1 = c1 and an = c(n?1) , n 2. 1 Since the approximants of the CF in (3.8) are rational functions, it should come as no surprise that there is a link between the CF in (3.8) and Pade approximants. It turns out that the CF approximants w2M and w2M +1 are \diagonal" Pade approximants; in particular, they are precisely the M=M ] and M + 1=M ] Pade approximants, respectively; see Theorem 4.2.1 of Baker and Graves-Morris 12]. (If the CF is terminating, corresponding to a rational function, then the approximants may be of lower order.)
(2) 1 (1) 1 (2) 2 2

cnsn = c0 +

c1s

Thus, methods for computing diagonal Pade approximants are equivalent to methods for computing approximants for CFs associated with power series. A powerful computational procedure is the quotient-di erence (QD) algorithm; see Chapter 3 and Section 4.3 of Baker 0 and Graves-Morris 12]. The QD algorithm has a1 = c0 , a2k = ?qk and a2k+1 = ?e0 , k 1, k j and ej are de ned recursively by where qk k
j j qk+1 = ej+1 qk+1 =ej k k

(3.13) (3.14)

and

j j ej = ej+1 + qk+1 ? qk k k?1

j for k 1 and j 0, with ej = 0, j 1, and q1 = cj +1 =cj , j 0. The QD algorithm is 0 related to the -algorithm; e.g., see Chapter 3 of Baker and Graves-Morris 12] and Chapter 8 of Wimp 37]. An essentially equivalent computational procedure is the product-di erence (PD) algorithm due to Gordon 19]. The PD algorithm involves only a single array pj , instead of the two arrays i j and q j in the QD algorithm. In particular, the recursion is ei i

?1 ?2 pj = pj?1 pj+1 ? pj?2 pj+1 ; 1 i 1 i i

(3.15)

where, for i 1, p1 and p2 are initialized to p1 = 1, p1 = 0 for i 2, and p2 = ci?1 for i 1. 1 i i i i Then n+1 1 an = p?1 n : (3.16) n Hence, given the moments of a pdf as in (3.1), we can obtain a CF representation of its LT in (3.2), with the property that the CF approximants coincide with diagonal Pade approximants. We now discuss additional structure that ensures that the CF is actually convergent.

p1 p1

4. Complete Monotonicity and S Fractions

Probability applications of CFs are especially appealing when we have a completely monotone (CM) pdf because then the associated LTs can be represented by special CFs called S fractions, which are known to converge under minor regularity conditions. A CF of the form (3.8) is an S fraction if all of the coe cients an (not considering the complex variable s) are positive. With such a simple characterization, it is also often possible to verify that a CF is an S fraction directly.

A function f on 0; 1) is completely monotone (CM) if it possesses derivatives of all orders that alternate in sign, i.e., (?1)n f (n) (t) 0 for all t 0 ; (4.1)

see p. 439 of Feller 18] and p. 66 of Keilson 29]. All CM pdf's are log convex, i.e., log f (t) is convex, and thus have decreasing failure rate (are DFR); see p. 74 of Keilson 29]. The family of CM pdf's is closed under mixtures. A pdf f on 0; 1) is CM if and only if f is a mixture of exponential pdf's, i.e., Z 1 y ?1 e?t=y dH (y) (4.2) f (t) = for some cdf H on 0; 1). Examples of CM pdf's appear in Abate and Whitt 2], 6], 8] and references therein. By making the change of variables x = y ?1 , we see that the CM pdf f in (4.2) can also be represented as Z 1 ~ xe?xt dH (x) ; (4.3) f (t) = ~ ~ where dH (x) = x?2 dH (x?1) and dH (x) = x?2 dH (x?1). We call H in (4.2) the mixing cdf ~ and H in (4.3) the spectral cdf. (See 2] for the case of the M/M/1 busy period.) Let X ~ be a mean-1 exponential random variable; let Y and Y be random variables with cdf's H ~ ~ and H , respectively. Then the representations (4.2) and (4.3) are the pdf's of XY and X=Y , respectively. The CM pdf f in (4.2) has Laplace transform
0 0

f^(s)

1
0

e?st f (t)dt =

1
0

(1 + sy )?1 dH (y ) :

(4.4)

Associated with any cumulative distribution function (cdf) F (t) is a complementary cdf (ccdf) F c (t) 1 ? F (t). Associated with the CM pdf f is a dual ccdf

Gc(t) =
and associated dual pdf

1
0

e?xt dH (x)

(4.5) (4.6)

g(t) =

1
0

xe?xt dH (x) ;

~ which are obtained by switching the roles of H and H . Clearly, Gc (t) and g (t) are CM too. It is evident from (4.4) and (4.5) that the dual Laplace transforms are related by ^ Gc (s)
Z

e?st Gc (t)dt =

1
0

^ (s + x)?1 dH (x) = s?1 f (s?1 ) :

(4.7)

Note that the moments of the CM pdf f in (4.2) are related to the moments of the mixing cdf H by mn (f ) = mn (H )n! for all n 1 : (4.8) Thus if f has moments of all orders, then its LT has the (formal) power series representation 1 n 1 1 n n ^(s) = X f^ (0)s = X mn (F )(?s) = X mn (H )(?s)n ; f (4.9) n! n=0 n! n=0 n=0 ^ where f^(0) (0) f (0) m0(F ) m0 (H ) = 1. If the LT f^(s) is analytic at 0, then the power series (4.9) has a positive radius of convergence and is not formal, but it su ces to have only a formal power series. We now relate these probabilistic quantities to continued fractions. The main connection is that, under the complete monotonicity assumption, the CF associated with the power series of f^(s) in (4.9) is an S fraction; see Chapter 5 of Baker and Graves-Morris 12]. Except for the normalization f^(0) = 1, which holds for pdf's, Laplace transforms of CM pdf's with all moments nite, having the integral representation (4.4), coincide with Stieltjes functions in the theory of continued fractions; see p. 193 of Baker and Graves-Morris 12]. As a regularity condition to avoid the case of terminating CFs, it is usually assumed that the underlying cdf H has in nitely many points of increase (does not have nite support). The associated series P1 n n=0 mn (H )(?s) in (4.9) then is called the associated Stieltjes series. The series is called formal because it may not converge for any s (except s = 0). In Section 3 we saw that there is a CF with denominator elements 1 associated with any formal power series. The fact that we have a Stieltjes series implies that the numerator CF elements (not counting the complex variable s) are all positive, i.e., that we have an S fraction. The advantage of S fractions is that there is more supporting convergence theory. We now note duality properties of the CM pdf f (t) in (4.2) and the dual ccdf Gc (t) in ^ (4.5). Since the LTs are related by Gc (s) = s?1 f^(s?1 ), it is immediate that the CM for f^(s) ^ converges at s if and only if the CM for Gc (s) converges at s?1 . We can also relate the two CFs.

Proposition 4.1. The LT of a CM pdf f (t) has an S fraction representation

with a1 = 1 and ak > 0 for all k if and only if the LT of the dual ccdf Gc (t) has the S fraction representation
?1 ?1 a1 ^ Gc (s) = s?1 1+ a2s a3 s 1+ 1+

a1 f^(s) = 1+ a2 s a3s a4 s 1+ 1+ 1+

(4.10)

a1 a2 a3 a4 ' s+ 1+ s+ 1+

(4.11)

resentation in (4.11). Then we obtain the second relation in (4.11) by applying the equivalence transformation in (2.5). The convergence of S fractions is intimately linked to the classical moment problem; see Akhiezer 11], Chapter 5 of Baker and Graves-Morris 12] and Sections 12.7{12.14 of Henrici 26]. For the following result, we apply Theorems 12.14b and 12.8e in Henrici 26]. We also ^ use the fact that the CF associated with Gc (s) converges to s if and only if the CF associated ^ with f (s) converges at s?1 .

^ ^ Proof. Since Gc(s) = s?1f^(s?1), from (4.10) we immediately see that Gc(s) has the rst rep-

Theorem 4.1. Let f (t) be a CM pdf as in (4.2) with all moments nite. Then the CFs

^ associated with the LTs f^(s) in (4.4) and Gc (s) in (4.7) are convergent (uniformly in compact subsets of the complex plane minus the negative real axis) if and only if the moments mn (H ) uniquely determine the cdf H . Note that an assumption in Theorem 4.1 is that the pdf f (t) in (4.2) has all moments nite. By (4.8), H too necessarily has all moments nite. Note that Theorem 4.1 does not ^ preclude f (s) having a convergent S fraction when the pdf f (t) does not have all moments nite. Indeed the dual ccdf Gc (t) may fail to have all moments nite when the conditions of ^ Theorem 4.1 are satis ed, so that Gc (s) has a convergent S fraction. It is also possible that the cdf H has all moments nite and is uniquely determined by those moments, but that the cdf F is not uniquely determined by its moments. An example of such an F is the Weibull cdf with exponent 1/2. Thus this Weibull cdf is not uniquely determined by its moments, but its LT nevertheless has a convergent S fraction. (We discuss this example in Section 9.) When the complex variable s in an S fraction is real and positive, we also have a convenient a posteriori bound on the numerical errors in the approximants, because then the even approximants w2m are increasing, while the odd approximants are decreasing, so that

w2n(s) < f^(s) < w2n+1 (s) for all n ;

(4.12)

see Theorem 12.11d of Henrici 26]. Moreover, we can numerically verify convergence for all complex s (except negative real s) by considering the case of any single positive real s, by Theorem 12.8e of 26]. Various a priori and a posteriori bounds on the error are also given in 26]. The fact that property (4.12) holds for all positive real s motivates using inversion methods based on positive real s. One such method is the Gaver-Stehfest inversion algorithm in 4]. 9

However, in our numerical examples we use the Fourier-series method with Euler summation, which is the main method described in 1] 4], 5]. It requires computing Laplace transform values at complex s. We close this section by showing what happens for a typical diverging S fraction. We let an = (n ? 1)3 for n 2. Table 1 shows the values of the even and odd approximants as a function of n for s = 1 and s = 10. First, monotonicity as in (4.12) and non-convergence are evident from the numerical results. Second, the odd and even approximants converge remarkably slowly to their limits, so that we cannot easily make use of the limits of the odd and even approximants either.

s=1 s = 10 even odd even odd 4; 000 0:6637 0:7927 0:1777 0:7312
index

8; 000 0:6643 0:7921 0:1783 0:7301 12; 000 0:6646 0:7918 0:1785 0:7296 16; 000 0:6648 0:7917 0:1787 0:7293 20; 000 0:6649 0:7916 0:1788 0:7291
Table 1: Values of even and odd S fraction approximants for s = 1 and 10 when an = (n ? 1)3 for n 2.

When the power series (4.9) is only formal, it is typically divergent. However, we have seen that the associated CF may nevertheless be convergent. On the other hand, as illustrated by Table 2, the CF may be divergent. When the CF converges, we have a way to sum a divergent series | called Stieltjes summation; see Chapter 19 of Wall 35].

5. First Passage Times in Birth-and-Death Processes

In this section we show how CFs can be used to compute the LT of a rst-passage-time pdf in a birth-and-death (BD) process. Let Ti;j be a random variable representing the rst passage time from state i to state j . It is elementary that such rst passage times can be expressed in terms of rst passage times to neighboring states; e.g., if i < j , then

Ti;j = Ti;i+1 + Ti+1;i+2 +

+ Tj ?1;j ;

(5.1)

10

where the random variables on the right are mutually independent, and similarly if i > j . Let fi;j (t) be the pdf of Ti;j and let f^i;j (s) be its LT, i.e.,

f^i;j (s)
From (5.1), we have

e?st fi;j (t)dt Ee?sT :

i;j

(5.2) (5.3)

f^i;j (s) =

k=j ?1 Y k=i

f^k;k+1 (s)

if i < j . Hence, in order to compute the LT f^i;j (s), it su ces to be able to compute the LT of the rst passage time to a neighboring state. First passage times to neighboring states are especially tractable because their pdf's are always CM. For nite-state BD processes, this CM property is an elementary consequence of the spectral theory associated with these reversible Markov processes; see Section 3.4B of Keilson 29]. For rst passage times up in in nite-state BD processes, the states above the destination state play no role, so that the state space may be considered nite. For rst passage times down in in nite-state BD processes, the CM property can be deduced by considering the limit of the nite-state approximations for which in the BD process n the original birth rate (in the in nite-state model) in state n, n , is set equal to 0. (The holding time in state n is then exponential with mean ?1 , where n is the death rate.) n (n) denote the rst passage time down from state i to state i ? 1 in BD process n Let Ti with state space f0; 1; : : :; ng, where n > i. Let Fi(n) (t) be the cdf of Ti(n). By the construction above, we can make sample-path comparisons as in Whitt 36] to deduce that a stochastic order relation holds as we change n, i.e.,

Ti(n)
by which we mean that

(n+1) st Ti

for n i ;

(5.4) (5.5)

Fi(n) (t) Fi(n+1) (t) for all t :

Hence, Fi(n) (t) decreases to a limit Fi (t) as n ! 1, which need be a proper cdf. We then apply the following result to deduce that the limiting cdf Fi is CM.

Proposition 5.1. If fFn : n 1g is a sequence of CM cdf's on 0; 1) such that Fn (t) ! G(t) c as n ! 1 for each t, then the limit G is CM. If Fn (0) ! 1, then G is proper.
c Proof. Recall that the ccdf's Fn can be expressed as c Fn (t) =
Z

e?xt dHn(x)

(5.6)

11

c for cdf's Hn by the CM property, so that the ccdf Fn (t) can be identi ed with the LaplaceStieltjes transform of the cdf Hn evaluated at s = t for t real and positive. Hence we can apply the continuity theorem for Laplace transforms on p. 431 of Feller 18]. It is also easy to directly construct CFs representing the LTs of rst passage times down with an in nite state space. Let i and i denote the birth and death rates in state i, respectively. ^ Let fi (s) denote the LT of the pdf of the rst passage time from state i to state i ? 1. By considering the rst transition, we obtain the recursion

f^i(s) =
=

i+ i i + i +s i

i+ i + i+ i +s ^ ^ i fi+1 (s)fi (s) i+ i +s

i+ i

i+ i i+ i +s

f^i+1 (s)f^i (s)

(5.7)

from which we obtain

f^i(s) =

Iterating on (5.8) produces the CF

i is ?

i ^ i fi+1 (s)

(5.8)

f^i (s) = ? 1 1 i ? k k s ; (5.9) k= + + i k k which directly has the form of a real J fraction (J for Jacobi), which can be shown to be equivalent to an S fraction; see Baker and Graves-Morris 12], Wall 35] and Dupuis and
Guillemin 16]. However, the CM property implies the equivalence, so that we do not need to construct one. The analysis above shows that rst passage times to neighboring states have CM pdf's, so that by Section 4 their LTs have CF representations that are S fractions. This S fraction representation provides a basis for computing the LT, which in turn can be used to calculate the more general rst-passage-time pdf's and cdf's by numerical transform inversion, using (5.3). Given the CM transform representation (4.4), if the mixing cdf H has all moments nite, then the CF (5.9) is convergent if and only if the moments of H uniquely determine H by Theorem 4.1. As noted at the end of Section 4, numerically, convergence for all s (except negative real s) can be veri ed by considering the case of a single real s. Then there is convergence if and only if the gap between odd and even approximants i.e., w2n+1 (s) ? w2n(s) in (4.12) decreases to 0 as n increases. In queueing applications the rst passage time of greatest interest is T10, which corresponds to the busy period. Interesting special cases are the M/M/1, M/M/s, M/M/1, M/M/s/0 and 12

M/M/s/r systems. It is interesting that the mixing and spectral cdf's have a continuous spectrum (interval of support) in the M/M/1 case, see Abate and Whitt 2], but have a countably in nite spectrum (support) in the M/M/1 case; see the next section. Alternative methods for computing the rst-passage-time LTs in the M/M/s/0 case were recently discussed in Abate and Whitt 7].

6. The M/M/1 Busy Period

In this section we apply the CF for BD rst-passage-time LTs in Section 5 to calculate the busy-period ccdf in the M/M/1 queue. We apply the Fourier-series method with Euler summation, the algorithm EULER in 5], to numerically invert the LT after we calculate the required LT values. The busy period is the time between an arrival to an empty system and the epoch when the system becomes empty again. It is the rst passage time T10. The LT of the busy cycle (busy period plus subsequent independent idle period) is given in (2) on p. 210 of Takacs 34], from which the transform of the busy period itself is easily obtained. Let the arrival rate be and the service rate be 1. Let b(t) be the probability density function (pdf) of the busy period and let ^(s) its Laplace transform. From 34], we obtain b ^(s) b
Z

1
0

e?st b(t)dt =

+ s ? e?

1
0

exp(?sx + e?x )dx

?1

(6.1)

Computation directly with (6.1) is inconvenient because of the integral. There is a quite substantial literature related to the M/G/1 busy period associated with type II particle counters and coverage problems; e.g., see Hall 25]. In particular, Stadje 33] shows that the complementary cdf (ccdf) of an M/G/1 busy period has the form

B c (t)

1
t

X b(u)du = ?1 c n(t); t 0 ;

where c n (t) is a pdf, the n-fold convolution of a pdf with

n=1

(6.2)

c(t) = H c (t)C c (t); t 0 ; H c (t) is the service-time ccdf and C c (t) is the ccdf of c(t), which has the form C c (t) = exp
Thus, in the M/M/1 case,

(6.3)

t
0

H c (u)du :

(6.4)

c(t) = e? expf?t + e?tg; t 0 :

13

(6.5)

While (6.2) and (6.5) provide interesting structural information, they do not seem so useful for computation. In contrast, Dupuis and Guillemin 16] show that (6.1) can be given a series representation which is numerically useful, see p. 61 of 16], in particular,

es ^(s) = + s ? b (1 + s ^(s)) ;
where ^(s) =
1 X
n=1 (s + n)n! n

(6.6) (6.7)

:
1 X
n?1

From (6.6) and (6.7), it is easy to get the rst two moments
n=1 n(n!) For example, for = 1 we get m1 = 1:72, m2 = 7:17 and c2 (m2 ? m2)=m2 = 1:42. It is 1 1 2( ), approaches 1 both interesting that the squared coe cient of variation as a function of , c

m1 = (e ? 1)= and m2 = 2e

(6.8)

as ! 0 and as ! 1. Moreover, c2( ) rst increases and then decreases, so that there is a maximizing c2( ), in particular, max 2:97 and c2 (2:97) ' 1:73. A rough exponential approximation based on matching the rst two moments is

(6.9) B c (t) c2 2 1 exp(?2m1t=m2 ) : + Approximation (6.9) is supported by the fact that B c (t) is asymptotically exponential as t ! 1; see p. 62 of Dupuis and Guillemin 16]. Dupuis and Guillemin also show that the busy-period pdf is a countably in nite mixture of exponentials. In addition, they determine the CF ^(s) = ? 1 1 ? n ; b n=1 s + + n

(6.10)

which we can also obtain as a special case of (5.9) by letting n = and n = n. Additional insight into the M/M/1 transient behavior is provided by Preater 31]. The values of the Laplace transform ^(s) are easily computed by either the series (6.7) or the b CF (6.10). For the argument s needed in the numerical inversion, and for other representative s, we found that 30 terms of each was su cient to produce 20-digit precision in the transform values. We then applied the Fourier-series method with Euler summation from Section 1 of 5] using parameters A = 15 log 10, m = n = 25 to compute values of the ccdf B c (t). (Hence a sum of m + n +1 = 51 terms had to be computed with s values (A +2k i)=2t, k = 0; 1; : : :; 50.) 14

Sample results are displayed in Table 2. The two di erent methods agreed at least to the 15 decimal places displayed. From the last three entries of Table 2, we obtain an estimate of the exponential asymptotics for the tail probabilities

Bc (t) 0:71200 exp(?0:450265t) as t ! 1 :

(6.11)

In contrast, approximation (6.9) yields B c (t) 0:826e?0:480t. In this case we did not actually need CFs, because the series representation is e ective. We used the series representation to con rm the e ectiveness of the CF. The situation is di erent if

t 0:1 1:0 3:0 5:0 10:0 20:0 30:0 40:0 50:0

B c (t) 0:909; 084; 551; 819; 689 0:490; 128; 803; 420; 172 0:185; 450; 685; 115; 345 0:074; 977; 023; 964; 783 0:007; 888; 690; 353; 660 0:000; 087; 403; 419; 314 0:000; 000; 968; 394; 367 0:000; 000; 010; 729; 416 0:000; 000; 000; 118; 878

Table 2: The M/M/1 busy-period ccdf B c (t) computed by the Fourier-series method to 10?15 precision based on computing the LT ^(s) computed to 20-digit precision by both the series, b (6.6) and (6.7), and the continued fraction (6.10).

we consider the excursion time above some level c, i.e., the rst-passage-time Tc+1;c . Guillemin and Simonian 24, p. 870] show that the Laplace transform of the excursion time in the M/M/1 system can be represented as the ratio of two Kummer functions, i.e.,

c 1 M (s; c + s + 2 ) f^c (s) = c + + + s M (s; c + s + 1;; ) ; 1

where M is the Kummer function

(6.12)

M (a; b; z) =

n=0 (b)n

1 X (a)n z n

n! ;

(6.13)

with (x)n being the Pochhammer symbol, i.e., (x)0 = 1 and (x)n = x(x + 1) : : : (x + n ? 1) = ?(x + n)=?(x), where ?(x) is the gamma function. In this case, we know of no alternative for computation of f^c+1 (s) to the CF

? f^c+1 (s) = ? 1 1 s + (n + c + 1) 1 ; n=1 +n+c+

15

(6.14)

which again follows immediately from (5.9). Giullemin and Simonian 24] also prove that the scaled excursion time cTc+1;c in the M/M/1 system with arrival rate c and individual service rate 1 converges to the busy period T10 in an M/M/1 system with arrival rate and service rate 1 as c ! 1. We can establish additional results. We can make a stochastic comparison by noting that the scaled M/M/1 system above level c is equivalent to a BD process with constant birth rate k = and death rates k (c) = (c + k)=c. Since k (c) decreases to 1 as c increases, we can apply 36] again to conclude that the variables cTc+1;c increase stochastically in c, i.e.,

cTc+1;c

st

(c + 1)Tc+2;c+1

(6.15)

for all c, as well as converge in distribution as c ! 1. Moreover, we can show that the entire scaled BD process above c converges to the M/M/1 queue-length process in the sense of weak convergence on function space. Similar observations are made by Preater 31]. The M/G/1 busy-period LT is known to satisfy the Kendall functional equation ^(s) = g (s + ? ^(s)) ; b ^ b (6.16)

where g (s) is the service-time LT. For M/M/1, g (s) = (1 + s)?1 , so that (6.16) becomes ^ ^ ^(s) = (1 + s + ? ^(s))?1 ; b b which upon iteration gives the CF (6.17)

? ^(s) = ? 1 1 b n=1 s + + 1 ;
as in (5.9). Of course, we can solve (6.17) to get ^(s) = 1 1 + + s ? p(1 + + s)2 ? 4 b 2

(6.18)

(6.19)

If we scale the M/M/1 busy-period pdf to have mean 1 and squared coe cient of variation c2 = v then the LT has the form ^(s) = 1 s + v ? p(s + v )2 ? (v 2 ? 1) ; b v?1 which can be written as a functional equation ^(s) = b (6.20)

: ? b s + v ? v?1 ^(s) 2
16

(v + 1)=2

(6.21)

From (6.21), we obtain the J fraction However, from (2.26) on p. 161 of 2] and (94.23) on p. 376 of Wall 35], we see that the LT ^(s) b has an S fraction of the form (4.10) with a1 = a2 = 1, a2k?1 = (v ? 1)=2 and a2k = (v + 1)=2 for all k 2. We now apply numerical inversion to evaluate the quality of the approximation for various values of c. Since the M/M/1 busy period pdf does not have a pure-exponential tail, whereas the M/M/1 busy period has a pure exponential tail, we are led to expect some discrepancies for large values. The numerical results are displayed in Table 3. Consistent with our expectations, when t is small, the limiting M/M/1 values are good approximations when c is not large. However, as t increases, c needs to increase too in order for the approximation to be good. time 1 3 10 50 100 200 400
(v ?1) ^(s) = ?2 1 ? 4 : b v ? 1 n=1 s + v
2

(6.22)

Scaled M/M/1 c = 1000 c = 10,000 :48583276 :48930439 :48965265 :24519851 :25005028 :25053923 :08584378 :09192578 :09255093 :00617291 :00983952 :01029354 :00054713 :00173518 :00193941 :00000571 :00010209 :00013611 :00000000 :00000067 :00000143

c = 100

M/M/1 c=1 :48969135 :25059360 :09262058 :01034506 :00196341 :00014051 :00000156

Table 3: The ccdf B c (t) for the scaled busy period c Tc+1;c in an M/M/1 system with arrival rate c and service rate 1 for three values of c.

7. Beta Mixtures of Exponentials

In Abate and Whitt 8] we studied a class of pdf's obtained by taking beta mixtures of exponentials (BMEs), i.e.,

v(p; q; t) =

Z 1

where b(p; q ; y ) is the standard beta pdf, i.e., ?(p + (7.2) b(p; q; y) = ?(p)?(q)) yp?1 (1 ? y)q?1 ; 0 y 1 ; q ?(x) is the gamma function and p > 0 and q > 0. We observed that the BME pdf v (p; q ; t) has Laplace transform

y ?1 e?t=y b(p; q; y)dy ;

(7.1)

v(p; q; s) ^

e?st v(p; q; t)dt = 2 F1 (1; p; p + q; ?s) ;

17

(7.3)

where 2 F1 (a; b; c; z ) is the Gauss hypergeometric function. In Section 2 of 8] and in 9], we showed that the BME pdf and its Laplace transform have Laguerre-series representations, which can be used for numerical calculations. For the BME LT, we obtain a closed-form expression for the Laguerre coe cients, so that 1 1 X (q )n s n v(p; q; s) = 1 + s (p + q) 1 + s ; ^ (7.4) n n=0 where (x)n is again the Pochhammer symbol. Here we observe that continued fractions tends to be more e ective for this example. The CF representation for the Gauss hypergeometric function in (7.3) was found by Gauss in 1812 by exploiting recursions; see p. 88 of Erdelyi 17]. The CF can also be constructed by the QD scheme mentioned at the end of Section 2 from the series representation of 2F1 ; see p. 533 of Henrici 26]. For 2 F1 (1; p; p + q ; ?s), the elements can be taken to be bn = 1, n 1, a1 = 1, (p + n p+ 2)s a2n = (p + q + 2? 1)(3)(p q + n ?n ? 2) ; n 1 ; n? +q+2 (7.5) n(q + n ? 1)s a2n+1 = (p + q + 2n ? 2)(p + q + 2n ? 1) ; n 1 : For the special case p = 1=2 and q = 3=2, the BME pdf coincides with the re ected ^ Brownian motion (RBM) rst-moment pdf, i.e., h1 (s) = v (1=2; 3=2; s); e.g., see 8]. In that ^ ^ case, a2n = a2n+1 = 1=4 for all n 1. The special role of the RBM rst-moment LT h1 (s) can be explained from the fact that it is the unique xed point of the exponential mixture operator; see p. 93 of 6]. The CF associated with the BME transform tends to converge rapidly. This can be seen from the fact that an k=4 (independent of n) as n ! 1 in the S fraction for v(p; q ; ks). ^ Table 4 displays the rst seven coe cients for the S fractions associated with the LTs of four distributions. One is an exponential; two are BME's; and the last is an exponential mixture of exponentials with pdf Z 1 x?1 e?t=x e?x dx : (7.6) f (t) =
0

The exponential LT is a simple rational function, so it has a simple terminating CF. The two BME examples have an = O(1) as n ! 1, so the CF's are easily calculated. The exponential mixture of exponentials has an = O(n) as N ! 1, so the CF can be calculated with some e ort. To illustrate the computational advantages of continued fractions over the series representation (7.4), we consider the number of terms (in units of 10) required to compute the 18

exponential exponential BME mixture of ?1 v (1; 1; 2s) v(1=2; 3=2; 4s) exponentials coe cient (1 + s) ^ ^ a1 1 1 1 1 a2 1 1 1 1 a3 0 1=3 1 1 a4 0 2=3 1 2 a5 0 2=5 1 2 a6 0 3=5 1 3 a7 0 3=7 1 3
Table 4: The coe cients of S fractions associated with Laplace transforms of four di erent pdf's.

BME transform v (1; 1; s) to obtain 16-digit precision. The numerical values for several s are ^ displayed in Table 5. As noted in Section 2 of 9], when we apply the Fourier-series method of numerical inversion of Laplace transforms incorporating Euler summation, we typically need to compute transform values at about 40 values of s = u + iv , with u 15=t and v k =t for k = 1; 2; : : :; 40. For the series in (7.4), the worst case (making js=(1 + s)j close to 1) is k = 40. Then the required number of terms for the series (7.4) is approximately n 4:4t?1 104 . In contrast, Table 5 shows that the continued-fraction method is much more e cient. complex number continued s series fraction 1 50 30 3 130 40 5 190 50 9 330 60 15 540 70 20 710 80 25 890 90 1 + 5i 630 60 2 + 10i 1470 70 15 + 125i 44,000 250 150 + 1250i 440,000 710
Table 5: The number of terms needed to compute the BME transform v (1; 1; s) to 16-digit ^ precision by the series in (7.4) and the continued fraction in (2.7) and (7.5) for several values of s.

19

8. Distributions Associated with BMEs

Our ability to calculate BME LTs enables us to calculate LTs of several other important related distributions. First, in 8] we also considered a second beta mixture of exponentials, denoted by B2 ME, which is constructed by using the beta pdf of the second kind, i.e.,

v2(p; q; t)
where

1
0

y ?1 e?t=y b2(p; q; y)dy; t 0 ;

(8.1) (8.2)

Unlike the BME ccdf, the B2ME ccdf has a long tail, i.e.,

?(p + b2 (p; q; t) ?(p)?(q) yp?1 (1 + y)?(p+q) ; y 0 : q)

q V c (p; q; t) ?(pp+tq ) e?t and V2c (p; q; t) ?(pp+tq) as t ! 1 : ?( ) ?( ) q

By Theorem 1.6 of 8], the ccdf's are related simply by

(8.3)

V2c (p; q; t) = etV c (p; q; t); t 0 :

Thus, from (1.22) and (5.6) of 8], the LTs are related by 1 ^ v2 (p; q; s) = s ? 1 (sv(p; q; s ? 1) ? 1) : ^ Thus, the CF for v (p; q; s ? 1) yields the LT v2 (p; q ; s). ^ ^ A third BME ccdf, denoted by B3 ME, is the dual ccdf as de ned in Section 4, i.e.,

(8.4)

(8.5)

V3c (p; q; t) =

1
0

e?txb(p; q; y)dy; t 0 ;

(8.6)

for b(p; q ; y ) in (7.2), with Laplace transform ^ V3c (p; q; s) = s?1 v(p; q; s?1) : ^ (8.7)

The B3 ME LT is easily computed from the BME LT via (8.7). In Section 7 of 8] we considered gamma mixtures of exponentials (GME) pdf's as limits of BME pdf's. Directly the GME pdf can be represented as

f (p; t) =

1
0

p?1e?y p 2 y?1 e?t=y y ?(p) dy = ?(p) t(p?1)=2 Kp?1(2 t) :

(8.8)

In 8] we noted that the LT of f (p; t) is

f^(p; s) = s?p e1=s?(1 ? p; s?1 ) ;

20

(8.9)

where ?(a; z ) is the incomplete Gamma function. ^ In 8] we noted that f (p; s) = lim q!1 v(p; q ; qs). Hence, we can let q ! 1 in the CF for ^ 2 F1 (1; p; p + q ; ?qs) in (7.5) to obtain a CF representation of f (p; s) of the form (4.10) with a2n = n + p ? 1 and a2n+1 = n. Now consider the dual ccdf associated with the pdf f (p; t), obtained as a mean-1 exponential random variable divided by a gamma random variable, i.e.,

Gc (p; t) =

1
0

e?t=y y pe ?(p) dy = (1 + t)?p ; +1

?1=y

(8.10)

^ which is a Pareto distribution. Since Gc (s) = s?1 f (s?1 ), from (8.9) we obtain ^ Gc (p; s) = sp?1 es ?(1 ? p; s) ; (8.11)

which can be shown to be the Laplace transform of (1 + t)?p from p. 21 of Oberhettinger and Badii 30]. Again the CF can be used to compute the transform values. By the duality, we see ^ that Gc (p; s) has a CF of the form (4.11) with a2n = (n + p ? 1) and a2n+1 = n. Since the pdf ^ f (p; t) in (8.8) and ccdf Gc (t) in (8.10) are CM, it is natural to consider computations of the Laplace transforms by continued fractions.

9. Other Examples
We conclude with a few other examples of pdf's whose LTs can be e ectively computed via CFs. These examples show that the CF elements can have remarkably simple structure.

Example 9.1. First Bell pdf. As in Example 6.1 of 9], consider the rst Bell pdf with
Laplace transform
X f^(s) = (e?1 =k!)(1 + sk)?1 ;

which has moments mn = n!b(n), where

k=1

(9.1)

b(n)

1 n ?1 X k e k=1 k!

(9.2)

is the nth Bell number. By (9.1), the pdf is a countably in nite mixture of exponentials. The LT values can be e ectively computed from the series (9.1), but they also can be e ectively computed from CFs. The PD algorithm applied to these moments yields a CF of the form (3.8) with a2n+1 = n and a2n = 1. Since a2n+1 = O(n), computation of the CF is possible but not easy. 21

Laplace transform

Example 9.2. Second Bell pdf. As in Example 6.2 of 9], consider the second Bell pdf with
f^(s) =
1 X
k=1

2?(k+1) (1 + sk)?1 ;

(9.3)

which has moments mn = n!~(n), where b

1 n ~(n) = X k+1 b k k=0 2

(9.4)

is the nth ordered Bell number. This second Bell pdf is also a countably in nite mixture opf exponentials. Again the LT values can be computed from either the series (9.3) or a CF. The PD algorithm applied to these moments yields a CF of the form (3.8) with a2n+1 = 2n and a2n = n. Again an = O(n), so that computation of the CF is possible but not easy.
mixture of inverse Gaussian (EMIG) pdf scaled to have mean 1 and squared coe cient of variation c2 > 1. The pdf thus has explicit LT

Example 9.3. EMIGs. As in 10] and Section 8 of 6], consider the two-parameter exponential
c2 ? 1

f^(s) =
Z 1

c2 ? 2 + 1 + 2(c2 ? 1)s
1 1?y
r

(9.5)

From Theorem 4.1 of 10], the LT in (9.5) has integral representation

f^(s) =

0 1 + 2(c2 ? 1)sy

y (1 +

c2 ? 1 2 ? 2)2 ? 1]y ) dy (c

(9.6)

By applying (94.22) on p. 375 of Wall 35], we see that the LT has the CF representation (4.10) with a1 = a2 = 1 and ak = (c2 ? 1)=2 for all k 3. Since the LT is given explicitly in (9.5), the CF is not needed for computation. We display the CF because it has a remarkably simple form. For the special case c2 = 3, ak = 1 for all ^ k 1. The case c2 = 3 corresponds to the RBM rst-moment LT ^ 1(s). The CF for h1 (s) also h can be obtained from the property that it is the unique xed point of the exponential mixture operator, i.e., 1 ^ ; (9.7) h1 (s) = 1 + s^ 1 (s) h see Section 7 of 6].

Example 9.4. Weibull pdf's. Consider a random variable Xr with the Weibull ccdf
Frc (t) = exp(?r(t)1=r); t 0 ;
22 (9.8)

see Chapter 20 of Johnson and Kotz 27]. Since Xr1=r is distributed as X1=r, the nth moment of Xr is + (9.9) mn (Xr) E Xrn] = ?(nrnr 1) : r Given the moments, we can construct the CF using the PD algorithm. Here we are interested in the \long-tail" case of r > 1. The ccdf Frc (t) in (9.8) is CM in that case, so that the CF of the LT is an S fraction. It is known that the moments do not determine the distribution in that case. The indeterminateness follows from the Krein condition, i.e., for r > 1, the divergence Z 1 (1 + x2 )?1 log fr (x)dx = ?1 (9.10)
0

fails to hold, where fr is the pdf associated with Fr ; e.g. see Akhiezer 11]. For integer r > 1, we can identify products of random variables (mixtures of distributions) that have the moments of Xr . (We conjecture that the full distributions coincide with the Weibull as well.) For integer r, 1 2 r ? 1 (1) nr mn(Xr) = (rnr)! = r r n n n r n = (1=r)n (2=r)n ((r ? 1)=r)n (1)r ; (1)n (1)n (1)n
n

(9.11)

where as before (x)n is the Pochhammer symbol. However, recall that for 0 < p 1, 0 < q < 1, (1)n = n!, (p)n and (p)n =(p + q )n are the moment sequences of the exponential, gamma and beta pdf's; e.g., see 8]. Let Z (p) be a random variable with a gamma pdf having shape parameter p, as in (8.8), so that Z (1) has a mean-1 exponential pdf, and let Y (p; q) be a random variable having the beta pdf in (7.3) with parameters p and q . Then, for integer r > 1,

# " r r r?1 n] = E Y Z (k=r) = E Y Y (k=r; (r ? k)=r) Y Z (1) Xr k k=1 k=1 k=1

for all n ;

(9.12)

where the random variables on the right are mutually independent, with Zk (1) all distributed as Z (1). For the case r = 2, we can conclude that they have the same distributions, because Frc (t) is CM for r > 1. Hence, d X2 = Z (1=2)Z (1) ; (9.13)
d where = denotes equality in distribution. Since the gamma (1/2) pdf of Z (1=2) is determined by its moments and X2 is CM, (9.13) is justi ed. Moreover, we can apply Theorem 4.1 to conclude that the S fraction associated with the LT Ee?sX converges.
2

23

However, for integer r > 2, we can conclude that the S fraction associated with Ee?sX fails to converge, because the moments mn (Xr )=n! do not determine the mixing cdf H . To illustrate, the rst 7 numerator elements of the CF (4.10) for Ee?sX are given in Table 6. Since an = n ? 1 for X2, we see that convergence takes place, but it is not too rapid. For r = 3 and 4, we see that an fails to be O(n) as n ! 1. Moreover, for r = 3; 4 the CF fails to converge, demonstrating that the moments of the mixing cdf H indeed do not determine H. continued fraction elements
r r

r a1 a2 a3

2 1 1 2 3 4 5 6 3 1 1 9 20 40 61:25 93:15 4 1 1 34 133:82 364:25 736:02 1342:57

Table 6: The rst seven CF numerator elements, for a CF of the form (4.10), of Weibull Laplace transforms for r = 2; 3 and 4.

a4

a5

a6

a7

24

References
1] J. Abate, G. L. Choudhury and W. Whitt, An introduction to numerical transform inversion and its application to probability models. In Advances in Computational Probability, W. Grassman (ed.), Kluwer, Boston, 1998, to appear. 2] J. Abate and W. Whitt, Simple spectral representations for the M/M/1 queue. Queueing Systems 3 (1988), 321{346. 3] J. Abate and W. Whitt, Approximations for the M/M/1 busy-period distribution. Queueing Theory and its Applications, Liber Amicorum for J. W. Cohen, O. J. Boxma and R. Syski (eds.), North-Holland, Amsterdam, 1988, 149{191. 4] J. Abate and W. Whitt, The Fourier-series method for inverting transforms of probability distributions. Queueing Systems 10 (1992), 5{88. 5] J. Abate and W. Whitt, Numerical inversion of Laplace transform of probability distributions. ORSA J. Computing 7 (1995), 36{43. 6] J. Abate and W. Whitt, An operational calculus for probability distributions via Laplace transforms. Adv. Appl. Prob. 28 (1996), 75{113. 7] J. Abate and W. Whitt, Calculating transient characteristics of the Erlang loss model by numerical transform inversion. Stochastic Models 14 (1998), 663{680. 8] J. Abate and W. Whitt, Modeling service-time distributions with non-exponential tails, AT&T Labs, 1998. 9] J. Abate and W. Whitt, Laplace transforms of probability density functions with series representations, AT&T Labs, 1998. 10] J. Abate and W. Whitt, Explicit M/G/1 waiting-time distributions for a class of long-tail service-time distributions. AT&T Labs, 1998. 11] N. I. Akhiezer, The Classical Moment Problem, translated from Russian by N. Kemmer, Oliver and Boyd, Edinburgh, 1965. 12] G. A. Baker and P. Graves-Morris, Pade Approximants, second edition, Encylopedia of Mathematics and its Applications 59, Cambridge University Press, Cambridge, England, 1996. 25

13] C. M. Bender and S. A. Orszag, Advanced Mathematical Methods for Scientists and Engineers, McGraw-Hill, New York, 1978. 14] G. Bordes and B. Roehner, Applications of Stieltjes theory for S -fractions to birth and death processes. Adv. Appl. Prob. 15 (1983), 507{530. 15] K. O. Bowman and L. R. Shenton, Continued Fractions in Statistical Applications, Marcel Dekker, New York, 1989. 16] A. Dupuis and F. Guillemin, Etude des Phenomenes Transitoires de Congestion Lies au Multiplexage Statistique en Boucle Ouverte sur un Lien ATM, France Telecom CNET Report, December 1997. 17] A. Erdelyi, Higher Transcendental Functions (Bateman manuscripts), vol. I, McGraw-Hill, New York, 1953. 18] W. Feller, An Introduction to Probability Theory and its Applications, vol. II, second ed., Wiley, New York, 1971. 19] R. G. Gordon, Error bounds in equilibrium statistical mechanics. J. Math. Physics 9 (1968), 655{663. 20] F. Guillemin and D. Pinchon, Continued fraction analysis of the duration of an excursion in an M/M/1 system, J. Appl. Prob., to appear. 21] F. Guillemin and D. Pinchon, On the area swept under the occupation process of an M/M/1 queue in a busy period, France Telecom, 1998. 22] F. Guillemin and D. Pinchon, On a random variable associated with excursions in an M/M/1 system, France Telecom, 1998. 23] F. Guillemin and D. Pinchon, Excursions of birth and death processes, orthogonal polynomials and continued fractions, France Telecom, 1997. 24] F. Guillemin and A. Simonian, Transient characteristics of an M/M/1 system. Adv. Appl. Prob. 27 (1995), 862{888. 25] P. Hall, Introduction to the Theory of Coverage Processes, Wiley, New York, 1988. 26] P. Henrici, Applied and Computational Complex Analysis, vol. II, Wiley, New York, 1977. 26

27] N. L. Johnson and S. Kotz, Distributions in Statistics, Continuous Univariate Distributions-1, Wiley, New York, 1970. 28] W. B. Jones and W. J. Thron, Continued Fractions, Addison Wesley, Reading, MA, 1980. 29] J. Keilson, Markov Chain Models | Rarity and Exponentiality, Springer-Verlag, New York, 1979. 30] F. Oberhettinger and L. Badii, Tables of Laplace Transforms, Springer, New York, 1973. 31] J. Preater, M/M/1 transience revisted. J. Appl. Prob. 34 (1997), 1061{1067. 32] W. H. Press, B. P. Flannery, S. A. Teukolsky and W. T. Vetterling, Numerical Recipes, The Art of Scienti c Computing, Cambridge University Press, Cambridge, England, 1986. 33] W. Stadje, The busy period of the queueing system M/G/1. J. Appl. Prob. 22 (1985), 697{704. 34] L. Takacs, Introduction to the Theory of Queues, Oxford University Press, New York, 1962. 35] H. S. Wall, Analytic Theory of Continued Fractions, Van Nostrand, New York, 1948. 36] W. Whitt, Comparing counting processes and queues. Adv. Appl. Prob. 13 (1981), 207{ 220. 37] J. Wimp, Sequence Transformations and Their Applications, Academic, New York, 1981.

27