Queueing Systems 19 (1995) 41-62 41

Diffusion approximation to a queueing system with

time-dependent arrival and service rates
Antonio Di Crescenzo
Dipartirnento di Matematica e Applicazioni, Universith di Napoli "Federico H",
Via Cintia, 80126 Naples, Italy

Amelia G. Nobile
Dipartimento di Informatica e Applicazioni, Universith di Salerno, Via S. Allende,
84081 Baronissi (SA), Italy
Received 30 June 1993; revised 3 August 1994

A time nonhomogeneous diffusion approximation to a single server-single

queue service system is obtained. Under various assumptions on the time-
dependent functions appearing in the infinitesimal moments, transient and steady-
state behaviour are analyzed. In particular, a diffusion approximation characterized
by space-linear and time-varying moments is studied. The density of the busy period
and the probability for the busy period to terminate are obtained. Finally, estimates
of the goodness of the diffusion approximation are given.
Keywords: Adaptive queueing systems; Feller process; busy period; first-passage-
time.

1. Introduction

Various types of approximations of queueing systems are present in the litera-

ture. In particular, a relevant effort has been devoted to diffusion approximations,
both in the case of single-station queues (see, for instance, H e y m a n [13], Kleinrock
[16]) and in the case o f queueing networks (see, for instance, Reiman [21], D u d a [3],
Gelenbe and Mitrani [8], Kobayashi [18, 19]), in order to obtain exact or improved
descriptions of otherwise intractable discrete systems. However, the literature is
scarce and fragmentary with respect to investigations on time-varying queueing
systems that are suitable to describe m a n y realistic situations. Even though a
direct analysis of a time-dependent M / M / 1 queue is due to Keller [15] and
Massey [20], in general the diffusion approximation appears to be necessary and
useful for the study o f time-varying queueing systems. F o r instance, Falin [5] has
shown that the virtual waiting time o f the M(t)/G/1/oo queueing system with
periodic Poisson arrivals can be approximated by a Wiener process with a
reflecting barrier at zero. Also Giorno et al. [9, 10] have obtained some diffusion
approximations to describe queueing systems characterized by time-dependent

9 J.C. Baltzer AG, Science Publishers

42 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueingsystem

arrival and service rates. Hereafter, we propose a new model and indicate how
several non-stationary instances deserving interest may be quantitatively analyzed.
In section 2 we shall analyze the queueing system with arrival rate An(t) =
[oz(t)/e +/3(t)]n + ,,/ol(t)/e and service rate # n ( t ) = o~(t)n/e, where c~(t) and /3(t)
are suitable time-varying functions, and investigate the discrete Markov process
{N(t); t ~> 0} describing the number of customers in the system at time t. A con-
tinuous approximation starting from the process {N*(t) = N(t)e; t ~> 0} will be
indicated in section 3. As e ,l 0, such an approximation leads to a time non-
homogeneous diffusion process {X(t); t >~ 0} characterized by drift Al(x,t)=
/3(t)x+7oe(t) and infinitesimal variance A2(x, t)= 2a(t)x. The approximating
process X(t) will be discussed in section 4. In particular, we shall determine the tran-
sition probability density function (p.d.f.) of X(t) and the first-passage-time p.d.f.
through the state zero, which is useful to approximate the busy period density.
The steady-state behaviour of the approximating process X(t) will then be investi-
gated under the assumption that the queueing system is characterized by periodic
arrival and service rates with equal periods. Since comparisons pointing out the
goodness of the continuous approximation will finally be provided in section 5.
For instance, the mean values of N(t) and X(t)/e will be seen to be equal, while
their variances become asymptotically equal as e ,L 0.

2. A queueing system with varying rates

Let denote by {N(t); t 7> 0} the discrete Markov process having state-space
{0, 1,2,...} that describes the number of customers in the system. We denote by
An(t) and #n(t), respectively, the arrival and service rates at time t. For
n = 0, 1,... and t ~> 0 we assume that

Pr{N(t + A t ) - N(t)= 1 [ N ( t ) = n } = An(t)/Xt +o(/Xt),

P r { N ( t + At) - N(t) = - l l N ( t ) = n} = #n(t)At+ o(At),

Pr{N(t + At) - N(t) = 0 [ N(t) = n} = 1 -[An(t) + #n(t)]At q- o(At),

Pr{lN(t + s N(t)l > 1 I N ( t ) = n } = o ( s

where An(t) > 0 for n = 0 , 1 , . . . , # n ( t ) > 0 for n = 1,2,... and # 0 ( t ) = 0 . The

queueing system we are going to study is characterized by independent arrivals
and services. Further, we assume that arrival and service rates depend linearly on
the number of customers:

(2.1)
 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system 43

with e > 0, 7 > 0 and with a(t) and/3(t) continuous b o u n d e d functions such that
a(t) > 0 and a(t)/e +/3(t) > 0 for all t >~ 0. We also assume that J'~ a(~-) dr = c~.
The chosen form of the arrival rate/~n (t) is suitable to model situations where
customers are encouraged to join the queue when this is growing, which, for
instance, occurs during certain sales times. In this case the length of the queue is
an attraction rather than a deterrent. The assumed service r a t e ~n(t) expresses a
service mechanism where a single server adjusts its service rate instantaneously to
meet the d e m a n d of the customers in the system. Alternatively, this can be
viewed as a service mechanism with "infinite servers". In other words, as soon as
a customer joins the system, a new server with service intensity a(t)/e comes into
action. We remark that the presence of time-dependent terms in the rates (2.1) is
useful to describe realistic situations in which the arrival and service intensities
are subject to vary in time.
The parameter e appearing in rates (2.1) can be considered as a measure of the
size of a(t). It plays a crucial role in the approximating procedure indicated in
section 3.
We note that the queueing system with rates (2.1) can be viewed as a time-
varying extension of the " M o d e l D " in Conolly [2], whose arrival and service
rates are An = A(n + 1) and ~n = #n, with ), and # positive parameters.
Time-dependent queueing systems with rates

An(t ) = C~rn(t) + 7k(t), #n(t) = /3rn(t) + 6k(t) (2.2)

have been proposed in Giorno et al. [10]. There it has been shown that the discrete
stochastic process denoting the n u m b e r of customers in the system can be approxi-
m a t e d by a time n o n - h o m o g e n e o u s diffusion process. We stress that the queueing
system with rates (2.1) considered in the present paper is fundamentally different
from that of Giorno et al. [10]. Only in the very special instance when we set
/3(t) = 0 in (2.1) and ~ = / 3 = 1, rn(t ) = na(t)/e, k(t) = a(t)/e and 6 = 0 in (2.2)
our model identifies with that of Giorno et al. [10]. As far as the analysis of the
models is concerned, in Giorno et al. [10] a diffusion approximation is obtained
and analyzed when rn(t) = nr(t) and 7 = 5 in (2.2), i.e. in a situation that is funda-
mentally different from the one expressed by our model.
Assume that at initial time t = 0 the n u m b e r of customers in the system is j,
i.e. Pr{N(0) = j} = 1. The transient behaviour of the queueing system N(t) is then
described by the probabilities

pn(t) = Pr{N(t) = n IN(0) = j } (n, j = 0, 1 , . . . ;t >/0).

M a k i n g use of assumptions (2.1) we obtain

a(t) A t p l (t) + o(At), (2.3a)

44 A. Di Crescenzo, A.G. Nobile/Diffusionapproximation to a queueingsystem

pn(t + A t ) = ( [ a _ ~ + 3(t)] (n --1) + ~/-~)Atpn_l(t )

+a(t) (n+ l)Atpn+l(t)+o(At) (n=l,2,...). (2.3b)

Hence,
e d
a(t) dt po(t) = - 7po(t) +pl(t), (2.4a)

e d
a(t) dtPn(t)=-- { [ 1 + a-~J3(t)e](n-1) +'7} pn-l(t)- { [ 2 + ~3(t)e] n +'7} pn(t)

+ ( n + l ) p n + t ( t ) (n=1,2,...), (2.4b)

which has to be solved with initial condition

pn(O)=6n, j= { a i f n = j , (2.5)
0 ifn Cj.

Let
oo

G(z,t) = Z pn(t)2
n=0

be the probability generating function (p.g.f.) of N(t). From (2.4) and (2.5) it
follows that G(z, t) is a solution of

~'(t) Ot G(z't) + (z- 1) 1 - 1 +3(t)e]zla(t)J J ~zG(Z,t)

0 = (z- 1)TG(z,t) (2.6)

with the following initial and boundary conditions:

G(z,O) =z j, G(1,t) = 1. (2.7)

Although a closed,form expression of G(z, t) does not appear to be achievable,

nevertheless some information about the transient behaviour of the queue N(t)
can be obtained, as shown hereafter.
 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system 45

PROPOSITION 2.1
The conditional mean and variance of N(t) are given by

E[N(t) I N ( 0 ) = j ] - '7 r + j " (2.8a)

~(t) ~(t)'
"T r ~- 2 j ~b(t) 7 ~b(t)
Var[N(t) IN(0) = j ] - e2 r ) e qO2(t) e ~(t)

t
1-~o(t) "T 1 J
+ j ~2(t ) +-2 e ~2(t) C~(T)~2(T) aT, (2.8b)
0
with
t

r = I ~(T)~(T) d~. (2.9)

0

Proof
Expressions (2.8) can be obtained by means of eqs. (2.4) and condition (2.5)
or, equivalently, by means of (2.6) and (2.7). Indeed, setting

Kr(t ) = E[Nr(t) IN(O)=j] (r = 1,2),

we have
d g 1(t) - fl(t)K1 (t) = 7---~,
~-~ oL(t) (2.10a)

d K2(t) - 2flK2(t) =7~ + [/3(t) + 2(7 + l)~-~] (2.10b)

Solving eqs. (2.10), with initial conditions Kr(O) = jr (r = 1,2), we are finally led to
expressions (2.8). []

Let us now analyse the steady-state behaviourof N(t).

PROPOSITION 2.2
If
,imP/'/
t~o~a(t----~= - u ( 0 < u < !) , (2.11)

then the stationary distribution of N(t) exists and is given by

qn--- lim
t-~
pn(t)= F(n+')'/(1-ue))
F(-r/(1-ue))
1-ue) n
n! (ue)'~/(1-~c) (n = 0,1,. ..). (2.12)
46 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system

Proof
Performing the substitution
t

(2.13)
0
eq. (2.6) becomes

OG[z'O-l(u)]+(z-1)(1-{
l O u /~[19-1(U)] }z)~.~G[z,O-I(u)]
: (z-- 1)'Ta[z,O-l(u)].

From the assumed divergence of the integral of a(t), one has limu~ ~ 0 -1 (u) = c~;
hence, passing to the limit as u--, oo and recalling (2.11) the stationary p.g.f.
G(z) - limt__,~ G(z, t) - l i m u ~ a[z,O-l(u)] is seen to satisfy

d G ( z ) = 1 - (1L ue)z G(z)'

G(1)= 1.
Hence, the stationary p.g.f, is given by

G(z) =
[1 ue
- (1 - ue)zJ
.]v,(1-~,,) ( 0 < z < 1 - 1 u-------~) " (2.14)

Probabilities (2.12) then follow after an expansion of (2.14) as a power series. []

Remark
After the equilibrium regime has been established, the number N of
customers in the system has a negative binomial distribution. Hence, under the
assumption (2.11), stationary mean value, variance and third order moment
about the origin are given by

E(N)= lim
t-~
E[N(t)[N(O)= j ] = 7__,
/~s
(2.15a)

V a r ( N ) = lim Var[N(t) lN(O) = j ] = 7 (2.15b)

?
E(N3) = t~limE [ N 3 ( t ) [ N ( O ) = j ] = ~ - ~ [ ( ? + 1) 2 + 7 + 1 - ue], (2.15c)

respectively. []
 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system 47

The meaning of the parameter u is made evident by Proposition 2.2. Indeed,

from (2.12) the stationary probability for the queue to be empty is seen to be

qo = (ue) ~'/(1-"~),

which goes to zero as u tends to zero, while from (2.15a) the mean length E(N) of
the queue tends to infinity. Therefore, condition u-~ 0 leads to congestion of
queueing system in the asymptotic regime.
We remark that i f 7 = 1 - ue, with 0 < '7 < l/e, then the stationary distribu-
tion (2.12) is of geometric type, i.e. qn = ue(1 - ue) n. Note that the Model D having
rates An = ),(n + 1) and #, = #n exhibits the same asymptotic behaviour. Indeed, if
), < #, its stationary distribution is also geometric, i.e. q, = (1 - A/#) (A/#)'.

ANALYSISOF A SPECIALCASE
We shall now consider a special case. We point out that even though this is
also a special case of the model proposed in Giorno et al. [10], the following math-
ematical analysis is not present therein.
The appearance of the two different functions of time a(t) and/3(t) in the
rates (2.1) is the main source of difficulty in the study of the transient behaviour
of the queueing system. If the ratio of these functions is a constant, say

/3(t) (p < !), (2.16)

P- o4t)

the transient description of N(t) can be obtained, as shown in the forthcoming

proposition. Note that p < 1/e is implied by conditions a(t) > 0 and a(t)/e +
/3(t) > 0. Under the assumption (2.16) the rates become

A,(t) = a(t----~)[(1 - pe)n + 7], #,(t) = __a(t)n. (2.17)

e e

PROPOSITION2.3
If (2.16) holds, the transient probabilities of N(t) are given by

[H(t)]J min(n,j)(~) [)- H ( / ) ]

pn(t) = [1 + (1 - pe)H(t)] j+7/(1-p~) k=O H(t) J

r(n-k+j+7/(1-pe))[ (1--p_e)H(t) 1"-k (2.18)

• V(j 7-7/--0 - p-~))~nT~).v 1 + (1 -- pe)H(t)J '
48 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system

where
1 - e -p`~ 1
H(t)= ---F , p<~,p#0, (2.19)
o(t), p : o,
and
t

(2.20)
0

Proof
Under the assumption (2.16), making use of transformation (2.13) eq. (2.6)
becomes:

O G[z,O-l(u)] + ( z - 1)[1 - ( 1 - pe)z] O G[z,O-l(u)] = ( z - 1)~/G[z,O-l(u)],

Ou

which is to be solved with conditions G(z, O) = z j and G[1, 0-1(u)] = 1. This is the
analogue of the equation leading to the p.g.f, of a linear birth-and-death process
in the presence of immigration, with rates

An=(1-pe)n+7, #n = n.

Hence, the solution is (cf., for instance, Ricciardi [22])

[(1 - z)H(t) + z]J

G(z,t) = [(1 - z ) ( 1 - pe)H(t) + 1] j + 7 / ( 1 - p ~ ) (0 < z < 1),

with H(t) given in (2.19). Probabilities (2.18) are then obtained by expanding the
generating function G(z, t) as a power series. []

Finally, we remark that assumption (2.16) with 0 < p < 1/e implies (2.11).
Hence, in this case Proposition 2.2 implies the existence of the stationary distri-
bution for the queueing system with rates (2.17).

3. The diffusion approximation

We shall now construct a diffusion approximation starting from the process

N(t) with rates (2.1). It should be mentioned that a scaling similar to the one we are
going to describe has been previously used by Keller [15] and Massey [20] in the
analysis of the M ( t ) / M ( t ) / 1 queue and by Knessl et al. [17] in the analysis of a
state-dependent M / G ~ 1 queueing system.
 A. Di Crescenzo, A.G. Nobile/Diffusionapproximation to a queueingsystem 49

First of all, let us consider the scaled process N,* (t) = N(t)e, where e is a posi-
tive constant. Clearly, {N*(t); t ~> 0} is a discrete Markov process having state
space {0, e, 2e,...}. The transition rates of N[(t) are easily obtained from those
of N(t). Indeed, from (2.1) and for n = 0, 1,... we have

Pr{N~*(t + At) -- N[(t) = e I N~*(t) = ne}

= l {~+3(t)]ne+7o~(t)}At+o(At), (3.1a)

Pr{N*(t+At)-N*(t)=-elN;(t)=ne}=lc~(t)neAt+o(At), (3.1b)
c

Pr{N~*(t + At) -- N[(t) = OIN[(t ) = ne}

= l - l {I2~(-@te)+fl(t)]ne+?a'(t)}At+o(At), (3.1c)

Pr{lN:(t+At)-N2(t)l>elN~*(t)=ne}=o(At ), (3.1d)

where we have set At = Ae2, with A > 0. Assuming that Pr{N~*(0) = j e } = 1, the
transient behaviour of N~*(t) is described by the probabilities

p~(t) = Pr{N~*(t) = ne [ N~(0) = j e }

=Pr{ne<<.N~(t)<(n+l)elN[(O)=je} (n,j=O, 1,...;t>~O). (3.2)

Making use of assumptions (3. I) we obtain the following system of equations:

p~(t + A t ) = [ 1 - 3 , ~ At] p~(t) + ~(-~t

e ) Atp*~(t) + o(At), (3.3a)

pn(t+At)= {I~ +fl(t)](n--1)c+7o~(t)}Atp~_l(t)

1
+- (n+l)eAtpn+l(t)+o(At) (n=1,2,...). (3.3b)

Let now introduce a continuous Markov process {X(t); t ~> 0} having state space
[0, ec) such that when e is close to zero, probabilities (3.2) are near to

f(ne, t lye)e --- Pr{ne ~< X(t) < (n + 1)e IX(0) =je} (3.4)
50 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system

and such that the correspondence between (3.2) and (3.4) becomes increasingly
better as ~ approaches zero. Under such an assumption, as is customary we substi-
tute pn(t) withf(ne, t[jr in eqs. (3.3). After setting x = ne and x 0 = jc, we obtain

li~{f(x,t-t-At'xo)-Ii-([~(~te)+fl(t)]x+~/o~(t)}~-]f(x, tlXo)

f ( x , t + At I x0)= { [~ fl(t)] ( x - e ) + T o ~ ( t ) } ~ f ( x - e , tl Xo)

+ (1-{I2~+fl(t)]x+va(t)}-~)f(x, tlx0)

+a(t)
e (x + e)-~ f(x + e'tlx~ + ~ (x > 0).

Expandingf as Taylor series we then have

0
lim e-~f(x, tlXo) + [fl(t)x +',/a(t)]f(x, tlxo)
x+o
0 s (92 Go(At) "~
Ox[O~(t)xf(x, tlXo)] - ~ Ox---5[a(t)xf(x, tlXo)] + ~ + o(e)j = o, (3.5a)

0
O f(x'tlx~ - Ox{[3(t)x +3'a(t)]f(x'tlx~
1 02
+ ~ ~5x2[{[2o~(t) + fl(t)e]x + ",/a(t)e}f(x, t[ Xo)] + o(At)
At +~ (x > 0). (3.5b)

Recalling that At = Ar 2 and passing to the limit as e + 0, eqs. (3.5) ultimately give

lim{
lO
[fl(t)x + "ya(t)]f (x, t IXo) - 2-~x [2a( t)x f (x, t [Xo)] = 0,
} (3.6a)
x~0l.

0
O f ( x , t I Xo) -- Ox {[fl(t)x + "ya(t)]f(x, t Ix0)}
1 02
+ ~ ~x 2 [2ol(t)xf(x, t[ x0)]. (3.6b)
 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system 51

Equation (3.6b) is the F o k k e r - P l a n c k equation for a temporally n o n - h o m o g e n e o u s

diffusion process X(t), while eq. (3.6a) expresses a zero-flux condition in the state
x = 0. Therefore, X(t) is confined to the interval [0, ec). N o t e that initial condition
(2.5) is replaced by
lim f (x, t[ Xo) = 5(x - Xo), (3.7)
t.Lo

where 5(- ) is the Dirac delta-function.

The above sketched procedure indicates that the discrete process
N~ (t) = N(t)e converges to the continuous diffusion process X(t). Hence, the diffu-
sion process X(t) can be used to calculate quantities that are suitable to describe
N(t) when e is very small. Indeed, for small

Pr{N(t) < n IN(0) =j} - Pr{X(t) < ne IX(0) = j { } . (3.8)

Recalling that x = ne and x0 = je, the approximation is expected to improve as e
goes to zero and as n and j grow larger.

4. Analysis of the limiting diffusion process

Hereafter we shall analyse the process X(t) obtained in the previous section.
This is a time-non-homogeneous diffusion process in [0, co), characterized by the
following drift and infinitesimal variance:

Al(X , t) =/3(t)x + 7a(t),

(4.1)
Az(x,t) = 2a(t)x.

D u e to the assumptions of section 2, we have '7 > 0 and, for t >~ 0, a(t) and/3(t) are
continuous b o u n d e d functions, with a(t) > 0 and f0~ a 0 - ) d'r = oo. As the para-
meter e does not play any role in the description of X(t), here we shall neglect the
previously introduced inequality a(t)/e +/3(t) > 0 for E > 0.
We shall now obtain transition p.d.f., m e a n and variance of X(t). A n asymp-
totic analysis will be performed both in the case when the ratio of/3(t) and a(t) is a
constant as t ~ co and when such functions are periodic. Particular care will be
paid to the first-passage time of X(t) t h r o u g h the b o u n d a r y x -- 0. Its p.d.f, and
the probability of ultimate passage will also be obtained in closed form. We stress
that these functions are of particular interest, as they can be used to understand
the behaviour of the otherwise untractable queue length process N(t). In par-
ticular, the first-passage time t h r o u g h x = 0 is the analogue of the busy period of
N(t).
To obtain the transition p . d . f . f ( x , t [ x0) of X(t), we recall that it is solution of
52 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system

Fokker-Planck equation (3.6b), with the zero-flux condition at x = 0 given by

(3.6a) and the initial condition (3.7).

PROPOSITION 4.1
For x > 0, x0 > 0 and t > 0 the transition p.d.f, of X(t) is given by

f(x, tlXo) = ~---~expt -~ j [~qo(t)J L~_l k %b(t) ) , (4.2)

where ~(t) and ~(t) have been defined in (2.9) and where I ~ _ l ( 2 y ) =
~-0 Y'Y-I+2k/(I'(7 q - k ) ' k ! ) denotes the modified Bessel function of first kind
(cf. Gradshteyn and Ryzhik [12]).

Proof
The transformation (cf. Capocelli and Ricciardi [1])

x
t= t t = W(t), x'o = x o ,
(4.3)
f(x, t l Xo) = ~(t)fl(x ', t' I

changes eq. (3.6b) and condition (3.6a) into

0
~)-Tf(, ' x' t' I x ~ ~ t' I x ~ ) + 210~02 {2x' f'(x', t' Ix~)} (4.4)

and

lim {~/f
x'+0/ ' (x', t' l X'o) -~ 10@ [2x'f'(x', t' Ix~)] } =0, (4.5)

respectively. Furthermore, initial condition (3.7) is changed into

lim f'(x', t' i X~o)= 6(x' - X~o). (4.6)

el0

Equation (4.4) is the Fokker-Planck equation for the time-homogeneous Feller

diffusion process (cf. Feller [6, 7]) having drift and infinitesimal variance

Bl(X') : 7, B2(x') = 2x'. (4.7)

Equation (4.5) expresses a zero-flux condition on the boundary x' = 0 (cf. Feller [7]
or Karlin and Taylor [14]). Hence, the diffusion interval of the process (4.7) is
[0, ec). If 0 < 7 < 1 then x f = 0 is a regular boundary. In this case condition (4.5)
 A. Di Crescenzo, A.G. Nobile/Diffusionapproximation to a queueingsystem 53

implies that the regular boundary x ~ = 0 is a reflecting state. Further, if'7 >~ 1 then
x I = 0 is an entrance boundary. In both cases the transition p.d.f, for x ~, x~ ~> 0 is
given by (cf. Giorno et al. [11]):

f'(x',t'lx~o) = ~exp{ x' +--x~~ J ~,-~oj['x"~('~-l)/2I'~-~(2x~-~~ ). (4.8)

Making use of (4.8), from transformation (4.3) we finally obtain (4.2). []

The nature of the boundary x t = 0 of the process (4.7) is determined by the

drift Bl(X t) = "7. Indeed, recalling that "7 is positive, the state x ~ = 0 is a regular
reflecting boundary if 0 < "7 < 1, while it is an entrance boundary if '7 ~> 1. This
classification is reflected on the nature of boundary x = 0 of the process (4.1).
Such a nature also emerges by analyzing the transition p.d.f, f(x, t Ix o) when x
approaches zero. Indeed, from (4.2) and from the properties of the Bessel function
one obtains

if0 < '7 < 1,

x0
lximf(x,
~o
tlXo)= ~(t) e x p ' - ~--~
r [
if 7 = 1,

0 i f T > 1.

Let us now introduce the Laplace transform of the transition p.d.f.f(x, t I x0):
(3O

'b(z, tlXo) = I e-ZXf(x't[x~ (Re z > 0).

0

F r o m (4.2) one obtains (cf. Erd61yi et al. [4], p. 197, no. 18):

9 (z, tlXo)= E l + z ~ ( t ) ] -'~

~--~J f -XoZ ]
exp~,qo(,] T z~b(,) ~'
(4.9)

from which the conditional moments of X(t) can be obtained. In particular, the
conditional mean and variance are specified by the following proposition of
straightforward proof.

PROPOSITION 4.2
The mean and the variance of X(t) conditional upon X(O) = x0 are given by

,(t), x0
E[X(t)]X(O) = x0] = qo(t) ' (4.10a)
54 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system

Var[X(t)[X(O) = Xo] = ^, - ~b2(t)

-~+ ,, r
LXo ~---~) ~
(4.lOb)

with ~p(t) and ~(t) defined in (2.9).

Note that results (4.10) are in agreement with those of (2.8) as e J, 0. We shall
now analyze two different situations leading to the asymptotic behaviour of X(t).

PROPOSITION 4.3
If
lim /3(t) _ u (u > 0), (4.11)

then the diffusion process X(t) admits the steady-state p.d.f.

X7 - 1

W(x) - lim f ( x , t I Xo) = - - u ~ e -"x (x > 0). (4.12)

r(v)

Proof
Recalling (2.9), from (4.11) one has

,.
l i m ~ ( t ) = lim~b(t)=c~, nm~=u.

Hence, from (4.9) we obtain the Laplace transform of the steady-state p.d.f.

9 (z) - lim r t[ x0) = (4.13)

which is the moment generating function of (4.12). []

It should be pointed out that by virtue of Propositions 2.2 and 4.3, condition
(2.11) is sufficient for the existence of the stationary distribution of N(t) and of the
steady-state p.d.f, of X(t). Moreover, as we shall see in section 5, they are in agree-
ment as e $ 0.

Remark
Since (4.12) is a gamma-type p.d.f., asymptotic mean, variance and third
order moment of the random variable X describing the equilibrium regime are
given by:

E(X) = lim E[X(t) IX(O ) = Xo] = _7, (4.14a)

t ----r ~ L/
 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system 55

Var(X) = t--,~limVar[X(t)IX(O) = x0] = ~22, (4.14b)

E(X3) = t~lim E[X3(t) l X(O)= x~ g[(~/ + l)2 +'/ + l]' (4.14c)

respectively. []

PROPOSITION 4.4
If a(t) and /3(t) are periodic functions of period T and satisfying
fl(t) < 0 < a(t) for all t/> 0, then for t > 0 the diffusion process X(t) admits of
the asymptotic (time-dependent) gamma-type density:

X'7- 1
W(x,t) = l i m f ( x , t + n T I x o ) = r----(~ [u(t)]'~ e -~'(Ox (x > 0), (4.15)

where
- 9(0

is positive for t ~> O.

Proof
Recalling (2.9) and the above assumptions on c~(t) and/3(t) we have

lim qo(t + nT) = lim •(t + nT) = ec, lira ~(t + nT) _ u(t).
n --+ c c n--*~ n--*~ ~(t + nT)

Hence, setting t = t + nT in (4.2) and passing to the limit as n ~ co, the asymptotic
density (4.15) is obtained. []

Remark
Asymptotic mean and variance immediately follow from (4.15):

nlim E [ X ( t + n T ) lX(O) = x0] _ 3'

u(t) (4.16a)

7 )9
nlirnc~Var[X(t + nT)IX(0) = x0] - u2(t (4.16b)
[]

It has been mentioned elsewhere (cf. Giorno et al. [10]) that for certain
queueing systems the assumption of constant or time-monotonic rates is not appro-
priate as it does not reflect the possibility of daily or seasonal changes in the arrival
56 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system

and service rates that typically characterize systems such as large shared compu-
tational resources. Proposition 4.4 is then a contribution towards the analysis of
the approximating continuous queueing systems under the effect of such time-
dependent arrival and service rates.
We now point out that in the quoted paper (Giorno et al. [10, section 4]) it
was also analyzed a time-non-homogeneous diffusion process that arises as an
approximation of the queueing system having time-varying rates (2.2) with
rn(t ) = nr(t) and ',/= ~5.These rates are linearly dependent on the number of custo-
mers, as in model (2.1) considered here. However, the two discrete processes are
relevantly different and cannot be transformed into each other. Indeed, in the
quoted paper the approximating process is a time-non-homogeneous extension of
the Ornstein-Uhlenbeck process, while the process of the present paper is an exten-
sion of the Feller process. We point out that in section 5 of Giorno et al. [10] the
asymptotic behaviour of the approximating process is also studied in the case of
periodic rates. However, in the asymptotic density W(x, t) and its moments the
parameters of the approximated queueing system are not present in a simple or
explicit manner. On the contrary, here expressions (4.15) and (4.16) exhibit the
very same periodic terms appearing in the rates (2.1). This makes the present
results much more useful for statistical purposes.

FIRST-PASSAGE-TIME T H R O U G H THE ZERO STATE

As noted at the beginning of this section, the first-passage-time of X(t)

through the state zero can be used to approximate the busy period of N(t). Let
us then introduce the first-passage time (FPT) of X(t) through x = 0, conditional
on X(0) = x0 > 0, i.e. the random variable

Tx0,0 = i n f { t : X ( t ) = 0 } X(0)=x0,
t~>0

and let
0
g(0, t[ x0) = ~Pr{Tx0,0 < t}

be its p.d.f. Note that when 7 >~ 1 the state x' = 0 of the diffusion process (4.7) is
an entrance boundary. Hence, due to transformation (4.3), also the boundary
x = 0 of the diffusion process (4.1) is entrance. Roughly speaking, the sample
paths of such process can start from the boundary, but quickly move to the interior
of the state-space never to return to the boundary. In this case one cannot define the
first-passage-time through x = 0. Hence, we shall focus our attention on the deter-
mination of the FPT p.d.f, through x = 0 when 0 < 7 < 1. Such density is obtained
by making use of the relationship between the first-passage-time p.d.f.'s of the diffu-
sion processes with infinitesimal moments (4.1) and (4.7).
 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system 57

PROPOSITION 4.5
If 0 < "), < 1, for all initial states x 0 > 0 the first-passage-time p.d.f, through
x -- 0 is given

~(t) ~(t) [ x_~]'-~ f xo l

g(O'tlx~ -1-'(1 -7) ~b(t) exp~-~--~f (t > 0). (4.17)

Furthermore, the ultimate first-passage probability Pxo = Pr{Txo,O < or is

if lim r = c~,
t ---+ o o

Pxo : j'g(O, tlxo)dt = r(1-%Xo/r (4.18)

< 1 if lim ~p(t) = % < e~,
o r(1 -'r) l ----~ O0

where
O(3

r(a,x) = Iza-le-Zdz

denotes the incomplete gamma function.

Proof
Due to transformation (4.3), the first-passage-time p.d.f.'s of the diffusion
processes with infinitesimal moments (4.1) and (4.7) are related by

x , = x~(t), t' = r xot = xo,

(4.19)
g(O, t I Xo) = a(t)~(t) g'(O, t' l x~o).

When 0 < 7 < 1 the first-passage-time p.d.f, of the diffusion process (4.7) through
the regular reflecting boundary x' = 0 is (cf. Giorno et al. [11]):

g'(O,t'lXo)-F(1-7) t' exp - 7 " (4.20)

Hence, making use of (4.20) in (4.19) one immediately obtains (4.17), from which
(4.18) follows. []

F r o m (4.18) we see that for all initial states x0 > 0 the ultimate first-passage
probability Pxo is unity iff l i m t ~ ~b(t) = oo. Note that such condition is surely
satisfied if -a(t) </3(t) ~< 0 for all t >~ 0. Furthermore, if the stationary condition
(4.11) holds, from Proposition 4.3 we have l i m t _ ~ ~b(t) = ~ , so that the prob-
ability Px0 is again unity.
It should be mentioned that the determination of closed-form expressions for
58 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system

first-passage-time p.d.f.'s is in general very hard. As far as the process X(t) is

concerned, the unique solvable case consists of the first passage problem through
the zero boundary, which leads to density (4.17).

5. Comparison of the processes

In the present section we shall make use of the mathematical results outlined
in the foregoing to discuss the problem of the "goodness" of the diffusion approxi-
mation to the queueing system N(t).
As seen in (3.8), when c is near zero the process X(t) provides a good approxi-
mation to N(t), i.e. it provides information both on transient and asymptotic phases
of the queueing system N(t). More precisely, N~*(t) = N(t)c converges to X(t) as e
goes to 0; to pinpoint the goodness of the approximation we shall now compare the
transient and the asymptotic behaviours of N(t) and X(t)/c. First of all, from (2.8a)
and (4.10a) we are led to

E[X(t) X(O)=j] _ 7 7(t) + J.j__"

E[N(t) IN(0) ~j] ~
j
The expected values of N(t) and X(t)/r conditional upon an arbitrarily preassigned
statej at time 0, are thus found to be identical in the transient phase. A confirmation
that the continuous approximation improves as e tends to 0 arises by comparing the
variances of the two processes. Indeed, from (2.8b) and (4.10b) we obtain

limVar[X(t)/e I X(O)/e = j] = 1 .
~o Var[N(t)[N(O) = j ]

Let us now assume that (2.11) holds; in this case both processes N(t) and X(t)
admit of stationary distributions, as shown by Propositions 2.2 and 4.3. Let us
consider the stationary distribution of N(t) as r goes to zero and n --+ cx~, with
ne = x. F r o m (2.12), by making use o f F ( n + 7 ) / n ! ~ n"y-1 as n ~ ~ (cf. 8.328.2
pag. 937 of Gradshteyn and Ryzhik [12]), we have

n -1 (1 -
q" ~ v(-//(l -

nJ F-~
XT- 1
_ _ lj"f e-UXs
 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system 59

Table 1
For v = 3' = 1 and e = 0.1 the steady-state distribution functions Fu(n) and Fx(ne) are listed together
with the relative difference A(n) -- [Fx(ne) - FN(n)]/FN(n) for various integer values of n.

n FN(n) Fx(ne) A(n)

1 0.077426 0.095163 0.229072
2 0.154853 0.181269 0.170591
3 0.228408 0.259182 0.134733
4 0.297059 0.329680 0.109812
5 0.360562 0.393469 0.091268
6 0.418984 0.451188 0.076863
7 0.472538 0.503415 0.065343
8 0.521501 0.550671 0.055934
9 0.566180 0.593430 0.048129
10 0.606888 0.632120 0.041577
15 0.761151 0.766870 0.020651
20 0.855777 0.864665 0.010385
25 0.913255 0.917915 0.005102
30 0.947967 0.950213 0.002369
35 0.968851 0.969802 0.000982
40 0.981382 0.981684 0.000308
45 0.988885 0.988891 0.000005
50 0.993371 0.993262 -0.000110
75 0.999505 0.999447 -0.000058
100 0.999963 0.999954 -0.000010

D u e to (4.12) this implies

qn ~ W(x)e =_ W(ne)e,

in a g r e e m e n t with (3.8) as t ---+ oo. A f u r t h e r c o n f i r m a t i o n o f the g o o d n e s s o f the

a p p r o x i m a t i n g p r o c e d u r e follows b y c o m p a r i n g the m o m e n t s o f the processes
N(t) a n d X ( t ) / e after the e q u i l i b r i u m regime has b e e n attained. I n d e e d , if limit
(2.11) holds, f r o m (2.15) a n d (4.14) a s y m p t o t i c m e a n s a n d variances o f these two
processes are identical:

Var() -

L e t us also c o m p a r e the expressions o f the t h i r d - o r d e r m o m e n t s o f N a n d X/e.

F r o m (2.15c) a n d (4.14c) we h a v e

E ( X 3 / e 3) (7+1)2+7+1
E(N3) (~' + 1) 2 + 7 + 1 - ue
60 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system

Table 2
Same as in table 1, with e = 0.01.
n FN(n) Fx(ne) A(n)
1 0.009545 0,009950 0.042395
2 0.019091 0.019801 0.037209
3 0.028589 0.029554 0.033780
4 0.038023 0.039211 0.031228
5 0.047387 0.048771 0.029201
10 0.093044 0.095163 0.022770
20 0.178346 0.181269 0.016390
30 0.255902 0.259182 0.012816
40 0.326284 0.329680 0.010408
50 0.390099 0.393469 0.008639
60 0.447930 0.451188 0.007274
70 0.500320 0.503415 0.006185
80 0.547769 0.550671 0.005297
90 0.590736 0.593430 0.004561
100 0.629638 0.632120 0.003943
150 0.775345 0.776870 0.001966
200 0.863807 0.864665 0.000993
300 0.949994 0.950213 0.000230
400 0.981653 0.981684 0.000031
500 0.993272 0.993262 -0.000010

Clearly, when e $ 0 the ratio E(X3/e3)/E(N 3) tends to 1 and hence E ( N 3)

E(X3/r This stresses that the approximation of N(t)e to X(t) improves as
tends to 0. A more powerful result can also be obtained. Indeed, from (2.14) and
(4.13)

lim E(e-ZX/') - lim ~b(z/e) = (1 ze-Z) "r. (5.1)

,~o E(e. ~s) - , l o G(e -~)

As z $ 0, limit (5.1) gives

E(e-ZX/~) ~ 1.
E(e -zN)

This implies that if e is near O, the asymptotic moments of the approximating diffu-
sion process X(t) are close to those of the queueing system N(t):

E(N k ) ~ E ( X k ) (k=1,2, .).

s ""
 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system 61

Table 3
Same as in table 1, with e = 0.001.

n FN(n) Fx(ne) A(n)

1 0.000993 0.001000 0.006435
10 0.009906 0.009950 0.004499
50 0.048628 0.048771 0.002941
100 0.094947 0.095163 0.002275
200 0.180974 0.181269 0.001632
300 0.258852 0.259182 0.001275
400 0.329339 0.329680 0.001035
500 0.393105 0.393469 0.000859
600 0.450862 0.451188 0.000723
700 0.503105 0.503415 0.000615
800 0.550381 0.550671 0.000527
900 0.593161 0.593430 0.000454
1000 0.631873 0.632120 0.000392
1200 0.698599 0.698806 0.000296
1400 0.753234 0.753403 0.000224
1600 0.797967 0.798103 0.000171
1800 0.834593 0.834701 0.000130
2000 0.864579 0.864665 0.000099
2500 0.917870 0.917915 0.000048
3000 0.950191 0.950213 0.000023

SOME N U M E R I C A L RESULTS

In tables 1-3 the evaluated steady-state distribution functions Fu(n ) and

Fx(ne) are listed for some values of n and e. The distribution function Fu(x) is
given by

Fu(x) = lim Pr{N(t) < x l N ( 0 ) = j } -

t -"-* ~
~ qn,
O<~n<x

where the expression of qn appears in (2.12). The distribution function Fx (x) is given
by
x

Fx(x) = t~lim Pr{X(t) < x[X(0) = x0} - I W(y)dy,

0

with W(x) shown in (4.12). Recall that both steady-state distribution functions exist
under condition (2.11). As pointed out, the approximation of X(t) to N(t)c is
expected to improve as e goes to zero. This is confirmed by the data of tables 1-3
showing that the agreement of the distribution functions Fu(n) and Fx(nc)
improves as e decreases and as n increases.
62 A. Di Crescenzo, A.G. Nobile/Diffusion approximation to a queueing system

Acknowledgements

W e gratefully a c k n o w l e d g e the c o n s t r u c t i v e criticism o f the referees o n a n

earlier version o f this paper. This research has b e e n partially s u p p o r t e d b y C N R
a n d b y M U R S T ( 4 0 % funds).

References

[1] R.M. Capocelli and L.M. Ricciardi, On the transformation of diffusion processes into the Feller
process, Math. Biosci. 29 (1976) 219-234.
[2] B. Conolly, Lecture Notes on Queueing Systems (Ellis Horwood, Chichester, 1975).
[3] A. Duda, Diffusion approximations for time-dependent queueing systems, IEEE J. Select. Areas
Commun. SAC-4 (1986) 905-918.
[4] A. Erd61yi, W. Magnus, F. Oberthettinger and F.G. Tricomi, Tables of Integral Transforms, Vol.
1 (McGraw-Hill, New York, 1954).
[5] G.I. Falin, Periodic queues in heavy traffic, Adv. Appl. Prob. 21 (1989) 485-487.
[6] W. Feller, Two singular diffusion processes, Ann. Math. 54 (1951) 173-182.
[7] W. Feller, The parabolic differential equations and the associated semi-groups of transfor-
mations, Ann. Math. 55 (1952) 468-518.
[8] E. Gelenbe and I. Mitrani, Analysis and Synthesis of Computer Systems (Academic Press,
London, 1980) pp. 991-1014.
[9] V. Giorno, A.G. Nobile and L.M. Ricciardi, On some diffusion approximations to queueing
systems, Adv. Appl. Prob. 18 (1986) 991-1014.
[10] V. Giorno, A.G. Nobile and L.M. Ricciardi, On some time non-homogeneous diffusion approxi-
mations to queueing systems, Adv. Appl. Prob. 19 (1987) 974-994.
[11] V. Giorno, A.G. Nobile, L.M. Ricciardi and L. Sacerdote, Some remarks on the Rayleigh
process, J. Appl. Prob. 23 (1986) 398-408.
[12] I.S. Gradshteyn and I.M. Ryzhik, Tables of Integrals, Series and Products (Academic Press, New
York, 1980).
[13] D.P. Heiman, A diffusion model approximation for the GI/G/1 queue in heavy traffic. The Bell
Syst. Tech. J. 54 (1975) 1637-1646.
[14] S. Karlin and H.W. Taylor, A Secound Course in Stochastic Processes (Academic Press, New
York, 1981).
[15] J.B. Keller, Time-dependent queues, SIAM Rev. 24 (1982) 401-412.
[16] L. Kleinrock, Queueing Systems, Vol. II: Computer Applications (Wiley, New York, 1976).
[17] C. Knessl, B.J. Matkowsky, Z. Schuss and C. Tier, Asymptotic analysis of a state-dependent
M/G/1 queueing system, SIAM J. Appl. Math. 46 (1986) 483-505.
[18] H. Kobayashi, Applications of the diffusion approximation to queueing networks I: Equilibrium
queue distributions, J. ACM 21 (1974) 316-328.
[19] H. Kobayashi, Applications of the diffusion approximation to queueing networks II: Nonequi-
librium distributions and applications to computer modeling, J. ACM 21 (1974) 459-469.
[20] W.A. Massey, Asymptotic analysis of the time dependent queue, Math. Oper. Res. 10 (1985)
305-327.
[21] M.I. Reiman, Open queueing networks in heavy traffic, Math. Oper Res. 9 (1984) 441-458,
[22] L.M. Ricciardi, Stochastic population theory: birth and death processes, in: Biomathematics,
Vol. 17: Mathematical Ecology, eds. T.G. Hallam and S.A. Levin (Springer, Berlin, 1986)
pp. 155-190.