Air-Terminal Queues under Time-Dependent Conditions Author(s): Bernard O. Koopman Reviewed work(s): Source: Operations Research, Vol.

20, No. 6 (Nov. - Dec., 1972), pp. 1089-1114 Published by: INFORMS Stable URL: http://www.jstor.org/stable/169300 . Accessed: 30/01/2012 22:39
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Operations Research.

http://www.jstor.org

Air-Terminal

Queues under TimeConditions

Dependent

Bernard 0. Koopman Massachusetts ArthurD. Little,Inc., Cambridge, (ReceivedJune 10, 1971) The queues formed by aircraft in stacks awaiting landing clearance have usually been treated either by machine simulation, or analytically as stochastic processes with time-independent transitionprobabilities (possessing stationary solutions). In contrast to such methods, the present paperregards the queuedeveloping process in question as strongly time-dependent, often with a diurnal (24-hour) periodicity. The formulation and treatment are entirely analytic and make use of machines only to solve the equations for the probabilities, by economical deterministic steps, using the coefficients as given in tabular form. Time-varying Poisson arrivals are assumed, and also an upper limit to queue length. Two laws of servicing are used: Poisson and fixed service time; these extremes are found to lead to numerically close results in the realistic case. This situation contrasts with the much cruder approximationof deterministic flow models. The stochastic equations belong to well studied types of differential or difference equations. When the coefficients have a 24-hour period, so does just one solution, all others approachingit. Actual airport statistics are made the basis of certain revealing computations. A perturbationmethod for treating multiple queues is outlined. The concrete results are exhibited as graphs.

IS BECOMING increasingly clear, the chief bottlenecks to air transport capacity at the present time are to be found at the air terminals: aircraft wishing to land or to take off are delayed (or even diverted) by the saturation of the facilities for these purposes. In order to improve this situation, and to plan for preventing its increasing occurrence in the future, it is necessary to study it in such quantitative detail that its various causes can be understood, and the effects of various proposed improvements predicted. The most conspicuous bottlenecks at air terminals are the lengthening lines of aircraft waiting for service. They may be queues of aircraft on the ground awaiting clearance of the runway to take off; or they may be stacked aircraft awaiting the opportunity to land. Such waiting lines need not all be physically assembled close to the runway: aircraft held at the passenger terminal when the takeoff queue is too long, and even those delayed at their port of origin until the stacks at their destination have diminished enough to warrant their admission, can either be regarded as part of the queues or, in a more detailed analysis, as queues feeding into the runway queues and stacks. Clearly the quantitative study of these waiting-line situations is a necessary condition for the improvement of capacity and the reduction of delays. Curiously enough, the great bulk of published work on these matters (queues, waiting lines,
1089

AS

1090

0. Bernard Koopman

input-output models, etc.) is of very little applicability to the real situation at air terminals. The reasons for this can be made clear by examining three typical cases in which the inadequacy of the conventional methods can be shown. The first case is in the use of simplified deterministic assumptions of the arrival and departure at air terminals, described as if these motions resemble the flow of liquids through pipes and reservoirs, the flow rate modeling the mean arrival or throughput rate of aircraft. While such models give basic guidelines to the process, they are grossly inadequate to the understanding of many of its essential features, which are caused by the high degree of randomness in the arrivals, departures, and processing times. Their inaccuracy will be shown in the graphs to follow. The second case of inadequacy of so many published analytical methods for handling the present class of problem applies to those that acknowledge the random factors, and correctly regard the queues as stochastic processes, but which almost inevitably base their treatment on the assumption of constant input conditions, in contrast to reality, in which the conditions vary strongly with the time. Clearly there is no real problem of capacity during normal operation between midnight and 4:00 AM; but the demands for service may increase to values that cannot all be accommodated during the morning and afternoon rush hours. Therefore treatments of queues that ignore this time dependence are of little value to the problems of air-terminal capacity. The third case in which frequently used methods show serious limitations concerns Monte Carlo simulation. One issue here is that of the cost-effectiveness of the method. The typical situation is the following: A decision is to be taken with regard to installing various choices of equipment, and their cost must be balanced against the possible degrees of service improvement. Obviously, there is no technical problem if the latter is clear and easy to measure quantitatively. The technical problem arises when this is not the case, and when the effects of the installations considered enter in a complex and often indirect way into the quality of transportation service that the equipment under study might produce. Even a theoretically perfect stimulation would only give, on one run, exactly what would be shown by the operation of the airport during one day; and one such observation would not be accepted as conclusive because of its inability to separate the actual benefits of the proposed equipment from random effects. Evidently a sufficient number of repetitions ('runs') of the simulation must be made in order to have statistical significance and thus give a really firm basis for a decision. When the number of runs needed for such a purpose is calculated on statistical principles, it is found to be so great that the very poor cost per given degree of reliability of the method becomes apparent. Roughly stated, the machine time for Monte Carlo simulation increases with the square of the reliable precision, but not much faster than its logarithm with analytic calculation. Comparisons of this and other published methods will be outlined in Appendix III. For these various reasons it has become necessary to return to the philosophy of realistic mathematical formulations and theorems, using computers for computation only. Most of this paper concerns the single queue, which may be thought of as a waiting line of stacked aircraft. The statistics of arrivals follow a Poisson law, but with a rate X that is a strongly varying function of time, according to statistics observed at actual airports. The law of 'service' (landing) has been difficult to settle on. It is clearly intermediate between a fixed service time and a Poisson one. By

Air-TerminalQueues

1091

examining the actual behavior of queues under both assumptions in considerable numerical detail, the results, which are graphed below, show a remarkable degree of closeness. Since this is true under the wide ranges of conditions of practice, the results are correspondingly insensitive to these service-time assumptions. Therefore, it is legitimate to adopt the simplest one in each study. Actually, we have used the Poisson service assumption in our subsequent work. One further assumption has been made throughout that may seem unfamiliar to most standard queue theory, namely, that a maximum upper limit m is imposed on the queue. This is realistic: counting stack spaces available at the J. F. Kennedy airport gives a value of about 60; many other important airports have considerably fewer spaces. By this assumption artificial theoretical difficulties-and absurdities-are avoided (they are the analytical reflection of the noncompactness of queuespace). Moreover, the very practical question of numbers of aircraft diverted under various conditions can be examined numerically. Since the queue conditions (the transition probabilities) are strongly time dependent, no steady-state solution can exist. On the other hand, when a diurnal periodicity of T= 24 hours is present, it is a consequence of classical theorems that one and only one solution (i.e., set of probabilities) is periodic and has T as a period, and that all other solutions approach it exponentially with increase of time. This is true for any service-time assumption provided the same periodicity is present. This periodic solution plays the role of the steady state one in the time-independent case. Most of our specific results are numerical and are exhibited in graphical form. We conclude by outlining an approach to the much more complicated multiple queues. Here, following the example of modern mathematical treatments of composite physical systems, a method of perturbations is used. It applies to the result of taking marginal sums in the double-queue equations, so as to reduce them to single queues, each of which is perturbed by the other.

1. THE CASE OF RANDOM ARRIVAL AND SERVICE TIMES

IT IS PERMITTED to assume that there exists an interval of time h which is long

enough for an aircraft to land in (if cleared to do so), but so short that no significant change in arrival rates or other general conditions of the terminal occur during this period h. It is of the order of a minute or two, whereas the time for diurnal changes in the terminal would be at least half an hour. Under these conditions, the usual assumption of random arrivals takes the form that the probability of an arrival during the interval (t, t+h) is, to terms of first order in h) Xh. In the present case, it is desirable to put into evidence the fact that it may depend on the number n already in the queue (e.g., no admission to a full queue), and also the fact that it depends on the time of day t. Therefore, we write X=Rn(t) ('right-transition'). This will correspond to our slightly modified Poisson law. The question of the statistical law of servicing (landing) has been a difficult one to answer. The more 'realistic' the law, the more complicated a mathematical problem it produces. There is, however, good reason to assume that it is intermediate

1092

Bernard 0. Koopman

between a Poisson law, as representing the extreme of randomness and a given constant service time, as representing the opposite extreme. In the present section the former will be assumed, and we shall write .h=L-L(t)h for the probability, to terms of the first order in h, that an aircraft be removed from the queue because of its landing during the interval (t, t+h). In the succeeding section, the constantservice-time case will be dealt with, after which the consequences of the two will be compared. From the usual probability reasoning (using the habitually unstated independence feature of the Poisson law) we derive the following equations for the probabilities Pn(t) that at the epoch t there be n aircraft in the queue (the accent denoting the time derivative). We use the relative slowness and smoothness of change of the probability quantities on the relatively short interval h as justification for replacing such increments as Pn(t+h)-Pn(t) by the linear approximation Pn'(t)h, etc. We thus obtain
Pn f(t)

=Rn-1 (t)Pn-1 (t) -[R. (t) +Ln (t)]Pn (t) +Ln+l (t)Pn+l (t) I
,

(1.1)

which is true for n = 1, 2,

m-1;

whereas for n = 0 we have

Po (t) = -Ro (t) Po (t) +Li (t) P1 (t), and for n = m, the maximum number allowed in the queue, Pm'(t) = Rm-i (t)Pm-1 (t) -Lm (t)Pm (t).

(1.2)

(1.3)

Equations (1.1), (1.2), and (1.3) form a system of m+1 ordinary linear homogeneous differential equations of the first order, of a type studied in mathematical physics for centuries and whose general properties are thoroughly known. In particular, given any initial values Pn (0), there will be one and only one solution. If the initial values are probabilities (nonnegative and adding up to unity) the same will be true of the corresponding solution. Finally, if the coefficients are periodic
in t, i.e., if

Rn(t+ T) = Rn(t),

L, (t+ T) =Ln (t),

(1.4)

there will be one and only one set of probabilities Pn (t) satisfying the equations and having this period T: P (t+ T) = Pn(t). This unique solution is approached exponentially, with increase in time, by every other system of probability solutions. It plays, therefore, the role of the steady-state solution that exists only when all the transition probabilities Rn(t) and 4n (t) are independent of time. (See Appendix I for a mathematical discussion and references.) It remains to obtain the solution numerically for various actual or representative values of the transition probabilities, which in general (particularly X) have to be given in tabular form. This requires computer routines of a standard type for solving differential equations of the present sort. Actually, the quantities of practical significance are the expected value i (t) of the number in the queue at the times t, its standard deviation, the cumulative number of aircraft 'turned away' (not admitted because the stack is full: they may be diverted or held on the ground at their port of origin); and, finally, the expected waiting time of an aircraft that joins the queue at the epoch t. All these come out of the computer program and are discussed in Section 3 below, computer methods in Section 4.

Air-TerminalQueues
2. THE CASE OF FIXED SERVICE TIME

1093

THE POISSONLAWof service time assumed in the last section represents the extreme

of randomness-intuitively, as well as in the precise sense of information theory: Of all the probability densities p (t) on the positive time axis and of given expected duration of service-, the exponential or Poisson jump density Me-t (p = l/j) is the one of minimum 'information' f p (t)logp (t) dt. Its graph has a spiked maximum at t = 0, decaying monotonically with increasing t. (See Appendix I.) At the opposite extreme is the situation in which the duration of service is a known constant c, the distribution being the Dirac delta function (3(t-c), infinitely peaked at t = c. In this case, = c and the above informational integral is infinite. Evidently any 'real-life' situation will be intermediate between the Poisson and the fixed-time case. This causes practical difficulties in formulation and computation, since it is next to impossible to fix on just what intermediate should be assumed, and since even the simplest suggested formulations greatly complicate the subsequent computational processes. Surprisingly enough, the numerical difference between the results of the two extreme service time assumptions can be neglected for all practical purposes over the ranges of practical importance. The demonstration of this insensitivity is the purpose of the present section. Since the 'real-life' cases are intermediate between the results of the two extremes, we do not have to worry about their more exact formulation. Since we are assuming that, as soon as an aircraft in the stack has landed, the next one enters the pattern and lands after c units of time, and so on, we can think of a sequence of points t1, t2, . . . marked on the time axis, all spaced at distances c apart, and representing the epochs of landing. We may go further, and confine our attention to the queue at these points: at each one, if the number in the queue had previously been positive, just one will be removed; otherwise, of course, zero. On the other hand, during the time c elapsed between two successive epochs t,_1 and ts, the probability that exactly k new aircraft arrive at the queue with the desire of joining it is given by the Poisson expression eC (cX)k/k! If there is space for these when t=t,, they will be admitted at this precise epoch. If not, i.e., if k>m-n, where m is the number of queue spaces and n that of those in it at epoch ts, then m-n will be admitted, the rest definitively turned away (held on the ground or assigned to different terminals, etc.). These rather rigid simplified sets of rules, which are not inconsistent with the phenomena under study, allow the writing down of simple recurrence relations (difference equations) determining the evolution of the system. One could generalize them by assuming a known change in the fixed service interval from epoch to epoch. This would make the spacings of the t, values no longer equal; but with slight and quite obvious changes, and insignificant increase in computational difficulties, the methods would be the same as those used below. If t is the time elapsed since an arbitrarily fixed origin t = 0, and if c goes into t just s times: s = [t/c], our epochs are t, = cs. Instead of the Pn (t) used in the last section, we let pnS be the probability that there be exactly n aircraft in the queue at the sth epoch (i.e., at t8). To write the law of addition to the queue, we have

1094

Bernard0. Koopman

Here and in (2.2), we declutter our formulas by writing a, ui instead of as, us,i for the probability of k 'applications for admission' between the (s-1)st and sth epochs. The only complication is when there are more applications than there are spaces, whereupon the remaining m-n spaces are all filled. The probability of this happening is the sum of the probabilities of (2.1) for all k ? mr-n; it is un,, where"
tk

i-

-aleai! =-l-_

a' e-ali !

(x-'e-x/k !) dx =(k,

a)/r(k).

Recalling our simplifying convention of regarding all queue-length changes as occurring at the precise epochs denoted by the integers s, with landing occurring before admission, recurrence equations are easily obtained. Consider first the case of p+l. The event of no aircraft in the queue at s+ 1 can occur in just two possible (and mutually exclusive) ways: by having none at s and none applying for admission between s and s-+ 1; or else by having one at s (which will be removed at s-+ 1), and none applying, as before. Compound and total probability show that
Po
+s1

-Po e + plse

p -sa

F -sa.

Next, let n be any positive integer less than m. The event of probability pa't can occur in the following n+ 2 different ways: no aircraft at s and n applications between s and s-+ 1; one aircraft at s and n applications; and, so on; finally, n+-1 aircraft at s and no application between s and s+ 1. From this we obtain the first m equations of the system of m+r1 equations (2.3). The last equation is obtained by similar reasoning, but is modified by the 'spill-over' possibility: to fill a queue at s-+ 1 when at s there are k members, any number of arrivals equal to or greater than m-k+l will work; the probability of this is the Um-k+l:
Po+'= 8s+1
pi=

(posI~pis)e)e-a p p (psjpls)ae (pos+pis)

I a+p2sea

pS+l=

(a2/2!)e-F+p2sae

aIp3

ea

(2.3)
pn Pm=
= (poF+pls)

(an/n!)e-a+p2s[a"-l/

(n-1)
t+PM U.

!]ea+F. .F+ps

iea

(PO'+P1)Um+P2sUm-t1+

It is emphasized that the quantities a and ui will in general depend on the epoch s, and for explicitness should be written as as and usi. Equations (2.3) form a system of recursive or 'difference' equations, m+1 in number, for obtaining p8+1in terms of p ,, and thus successively in terms of any given initial pn? (n = 0, 1, *-, m). Four facts stand out at once: First, if the given initial values are nonnegative, so are all-those obtained by recurrence; this is because the coefficients are nonnegative. Second, if the initial values add up to unity, so do all subsequent sets of values; to see this we have but to add all the equations (2.3), obtaining by virtue of (2.2) the relation
po
s +1

+P.1-+1+...

s+

~ +p"m =+Po8+pS+**

+Pm8.

Third, all values are uniquely determined-and numerically computableby multiplications and additions--in terms of the given initial values: if these are proba-

Air-TerminalQueues

1095

bilities (nonnegative and adding up to unity), the same will be true of the computed values. The fourth fact implied by (2.3) results from the vanishing of the determinant of the (m+ 1) X (m+r1) matrix M of coefficients of the m+ 1 quantities ps on the right in (2.3)-this determinant having its first two columns equal. By the elementary theorems of linear dependence, equations (2.3) will be inconsistent unless the rank of the 'augmented matrix' JlpS'l, Ml' in (2.3) is the same as that of M; the augmented matrix is formed by adjoining the column of quantities on the left to M. It will be shown in Appendix II that the rank of M is m, and that the augmented matrix will have the same rank if and only if the quantities ps'+l satisfy the following equation (in addition to having a unit sum):
Dmp
+-. + (-1) m-lDipsll + ( _ 1) m-a =O -+j-+Dm2p2 -DmipD1P + (- 1)ma+(-tm(1, etc. Dm=am-l/(m-1) !am-2/(m-2)!+ )

Note that the quantities a and Dm in these equations are functions of the epoch s or t, according to (2.1). This last result means that every set of probabilities p'+1 that evolves from initially given probabilities by the process expressed in (2.3) must satisfy (2.4), and that, if the initial probabilities pn0 had not satisfied this relation, they could not have evolved from any previous ones through such a process. This is important in the case of periodic arrival rates X(t) = X(t+ T) and when we wish to find a periodic solution starting from selected initial values: a necessary (but not sufficient) condition for periodicity is that the initial values satisfy (2.4). Both in numerical computation and in applying the general theory, it is better to replace (2.3) by a slightly modified system having m rather than m+ 1 unknowns and a nonsingular matrix of coefficients. For this purpose, we add the first equation in (2.3) to the second and then introduce the symbol
pi, =oP,+p1% (s=O, 1, 2,
* )
*, pS;

(2.5)
Pnm).

whereupon we obtain m recurrence equations in the m unknowns (PIS' p2S Suppose that these m quantities are written as a single-column matrix the square matrix of coefficients be
(1+a)e-a
(a2/2!)e-a e-a
aeon

and let

0
e-a

0
0

...
...

0
0

A=
(am-i/(m-1)
Ua

=A (a) =A (a,) =As.

(2.6)

(am-2/(m-2)
UM-1

!)e-

...
. -a

Then our recurrence equations may be written in matrix form as ps?1= Asp. The solution by recurrence is then (2.7) p = A -A 2 .AIp0. The fact that A is nonsingular and, indeed, that its determinant is given by the formula detA = detA (a) = am-le(n-l)a/ (m-1)! (2.8) is proved in Appendix II. Therefore, the initial values, the elements of p0, may be chosen without restriction other than nonnegativeness and unit sum. Consequently, there is a fundamental system of m independent solutions of (2.7), in terms

1096

Bernard0. Koopman

of which all other solutions may be expressed as linear combinations. Such a fundamental system can be obtained by writing down the unit mXm matrix I (units in the principal diagonal; all other elements zero) and regarding each column in I as a set of initial probabilities p0-the kth column corresponding to an initial queue length of just k aircraft. The resulting m independent solutions form the m columns in the matrix product
pS

= A8

lAs-2

A1.

(2.9)

This is shown by the application of (2.7) to each of the columns of initial values p0 in matrix I. Since the determinant of PS is the product of those of the A matrices, which are not zero, the same is true of that of PS. Therefore, it represents m independent solutions, of which every other solution is a linear combination, the coefficients being indeed the initial probabilities of the latter. Once the values of pi', p22 etc., are obtained, we find po' and pi' from equations (2.4) and (2.5). It is usually sufficient to ignore po' and find po'+ in terms of 1 -p0 po' using (2.5) and the first equation in (2.3); then p' We conclude by stating the theorem, whose proof will be found in Appendix I. THEOREM. If the arrival rate X(t) is periodic in t: X(t+ T) = X(t), one and only one solution pS will be periodic, and will have T as its period, p?S+ = ps (S = T/c = [T/c]). Every othersolution will approachthis one exponentially with increase in t (i.e., of s). Note that we have assumed that the service time c goes exactly into T (24 hours). This represents no material limitation of the result.

3. COMPARISON OF THE NUMERICAL RESULTS

IN ORDER TO show the practical consequences of the treatments given in Sections 1 and 2, in quantitative terms and under realistic orders of time variability, we have used actual FAA statistics.[2 These were gathered in 1968 during one month's operations at J. F. Kennedy and LaGuardia airports, and show the average activity-landings plus departures-at each time of day during 24 hours. While these are approximately twice the arrival rates, we have chosen them as the X(t) functions in our hypothetical Terminals A (J. F. Kennedy data) and B (LaGuardia data). Their relative diurnal variations can be regarded as realistic, and they seem to be the only available data of this sort gathered precisely and in quantity. Moreover, they provide good examples of heavy enough traffic demands to cause real problems of airport capacity. We have made the further rather artificial assumption that the maximum number admitted to the queue is m = 25. This is too small to be realistic for the largest terminals, but it illustrates rather dramatically the problems of capacity under different service and arrival rates. In the absence of firm day-by-day data on landing or servicing rates, three representative averages are used as constant values of ,u (and 1/c) throughout the day: 45, 55, and 70 landings per hour (i.e., when there are aircraft in the queue). [] This is all in the case of a single queue. The problem becomes much more complicated when the same runway system is used for both landing and takeoff. General approximate methods for dealing with multiple queues will be examined in Section 5.

Air-TerminalQueues

1097

_~~~~~~~__ ___

/>

C~

L.O

_~~~~~3V
<>
O

A
<>

1VI8
<>
<s)

_f0:()
<> <>
t

==== t~~~~~~~~~~~~~~~~~~ == =E=Z

<>
:n

SW 2

row

Ace

<>
N

<

< ice~~~~~~~r _~~~~~~~~~~~~i

31vS

1v188t

AlnOH-(}<

1098

Bernard 0. Koopman

A simple and commonly used method of writing schedules, and of calculating the way in which the terminal can be expected to handle its loads at various times of day is based on examining rates-of arrivals, departures, landings, etc. This method gains its simplicity by ignoring the random features of these events, and by regarding all quantities as constantly equal to their expected values. With this assumption, the phenomena become quantitatively identical with those of liquid flow, corresponding in the present case to the following model: Liquid is flowing at the rate
25.00______

20.00

15.00

EH

q
0.0

10.00

5.001

4.00

8.00

12.00

16.00
TIME IN HOURS

20.00

24.00

28.00

Fig. 2. The expectedlength of the queue for terminalB for ,u 45 arrivalsper

with the same data. X(t) units of volume per hour into an open reservoir of capacity m, provided with an exit tube at its base, out of which any liquid in the reservoir flows at the rate , units per hour. When the inflow exceeds the outflow, the volume in the reservoir increases, spilling over the top when its limit m is exceeded. The calculation of the volume in the reservoir at any time t--the model of the expected queue length-is immediate: If it starts at zero, it remains zero as long as ,u AX(t). As soon as a time t1 arrives when X(t) becomes greater than ,4, the volume builds up according to the equation

Air-TerminalQueues

1099

This continues until it either exceeds m or drops to zero, whereupon the appropriate modification of the formula is made. We shall call this the deterministicflow model of the time-dependent queue, and include its consequences in our comparisons.[4] Figure 1 shows the graphs of the arrival functions A(t) for Terminals A and B. Their extreme variability with time should sound the alarm to any attempt to apply queuing theory based on constant transition probabilities, steady-state solutions, and so on.

,,

00.
Lii

W~~~~~~~~~~~~~~~ 0I

W/

0DXt

C) I0
0
0-

Fi.3the

thTuu

execedlegt of wit

B erminal n =u=4

o 55,epciey

for h

arrivalcsolto

lineei perw hourl t /che solid linve, iase for the constan service time; cnd the dashed flown the dotrrated linetsafo the service raumteon It= 1/ad fornthe ranomi(oison

3,the expecatedvauresis t isl 'ofrtet lengthe other qusefue show ine tatnd on Figs.w pheroi swolution55crseciel.thein 45efcistaaband fomr chase ethen of Trina Bland the wit curve eahe figureaiythree sent the dotteronegive areshsetbekown: by detrmnisi the

between our curves shows that-in practical relevance-we are not too badly off. In fact, on turning to Fig. 4, which shows the standard deviation of the queue

1100

Bernard 0. Koopman

length under each hypothesis and with 4 = 55, one sees that either one leads to a value decidedly in excess of the width of the band of indeterminacy. Figures 5 and 6 show the probability of a full queue P. (t) and an empty queue Po (t) under the three hypotheses in Terminal B with ,u= 55. We have carried out many further
10
8
0

4.00

8.00

12.00

16.00
Time in Hours

20.00

24.00

28.00

Fig. 4. The standarddeviationof the numberin the queue for terminalB for 55 arrivals per hour= 1/c. The solid line is for constant service time, the dashedline is for Poissonservice,and the dotted line is for deterministic flow.
=

computations by these methods; their results all show this relative insensitivity to the servicing hypothesis. [5' Returning to Figs. 2 and 3, in which the consequences of the deterministic flow model are plotted in the dotted lines, it is obvious that, since the random features
1.0 ..00.0.0.

0.8

0.6

0.4

0.2

0.0 4.00

8.00

12:00

16.00
Time-in Hours

20.00

24.00

28.00

Fig. 5. The probabilityof an empty queue for the conditionsof Fig. 4. of the actual events are ignored, the results are reliable only to the extent that they approximate to the band of indeterminacy between the dashed and the solid curves. The latter curve is, as it should be, closer to the dotted one. The facts brought out here graphically are a comment on the degree of validity of the schedules and forecasts that are based on assumptions logically equivalent to the deterministic flow model.

Air-TerminalQueues
4. PRACTICAL EXAMPLES OF THE NUMERICAL RESULTS

1101

THE INSENSITIVITY OF its numerical consequences to the landing-time assumptions having been established, in the sense of the preceding section, the option is presented of basing the equations of evolution on the particular assumption that leads to the simplest treatment. It soon becomes clear that the Poisson service-time assumption leads to the most economical calculations, as well as the simplest formulation and application of theory. Accordingly, this assumption will henceforth be made and the equations (1.1), (1.2), (1.3) will be used.

0.8
1.0 0.6

~0.4 CLE.
0 .2

0.2 4.00 8.00 112.00 16.00 Time in Hours 20.00 24.00 28.00

Fig. 6. The probabilityof a full queue for the conditionsof Fig. 4. The quantities of practical significance are the following, expressed as functions of the time. A. The expectedvalue and standarddeviationof the queue length. Pm(t)and Po(t) of full and empty queue. B. The probabilities C. The cumulativeexpectednumberturnedaway fromthe queue. D. The expectedwaitingtime from the time t of joiningthe queue. The initial values required to complete the determination of these quantities are chosen in whatever manner seems most relevant to the practical matters considered. When the general conditions of the terminal have a diurnal periodicity, it is usually appropriate to employ the unique periodic solution, which can be found by the relaxation method of Appendix I. It is emphasized that the quantities A-D depend on the solution {Pn(t)} by simple equations, and that an easy adjunction to the program for computing the latter (i.e., solving the equations of evolution) will provide the values of the former. The evaluation of C and D is given in Appendix I. The present methods have been applied to many computations of this sort for the use of the Federal Aviation Authorityi'] The arrival values of Terminals A and B shown in Fig. 1 have served as examples, as well as the three constant values -=45, 55, 70 of the preceding section. In addition, a time-dependent ,u showing a sharp dip was used: = 55 from early morning to 1500; 1= 25 between 1500 and 1700; p=55 from then on. This assumption was used to show the applicability of the methods to the strongly time-dependent service rate as well as arrival rate, and

1102

Bernard 0. Koopman

also to show how it could be used to trace through the effect on capacity of a sudden temporary shutdown of an air terminal, due, for example, to a temporary degradation of local weather. Such a 'dip' can be represented as a continuous function 1u(t), coinciding with the step function having the two stated values (55 and 25) except within a microsecond of the two discontinuity points of the latter: neither observations nor computers could reveal their difference. Clearly, the assumption of the arrival values X(t) equal to the total rate of events at J. F. Kennedy or LaGuardia, and handled by a single-landing runway and under the restriction of 25 in the stack, represents an extreme situation of heavy arrival and lack of stacking space: it does not now face the two airports named, but illustrates what realistic planning must foresee. Moreover, the computations reveal the dramatic improvements secured by increases of service (landing) rates of about 20 per cent. Thus, in the case of Terminal B, the total daily number of aircraft turned away fell from about 100 when y =45 to 25 when ,u=55; and to zero when = 70. In these three cases, the waiting times of an aircraft joining the queue during the rush hour period were 30, 20, and 5 minutes, respectively. Thus, even with the unreasonable overloads assumed, the situation can be brought to a very satisfactory one if the landing rate can be increased to the capability of landing one and a sixth aircraft per minute on the average. In these cases, the numbers turned away as derived from the deterministic flow model are, respectively, 76, 6, and 0. These over-optimistic figures can be taken either as evidence of the misleading nature of the model, or as objectives that might be approached by flow control. For the numerical integration of the differential equations (1.1)-(1.3), the fourth-order RUNGE-KUTTA method was used, with a step size of two-tenths of an hour (cf. Anthony Ralson and Herbert S. Wilf, Mathematical Methods for Digital Computers, Vol. I, Chap. 9, Wiley, New York, 1960).

5. THE MULTIPLE QUEUE

WHEN A SINGLE runway system is shared by arriving and departing aircraft, it is necessary to use two integers to describe the state of the system at each epoch t. Let Pm, (t) be the probability that at t there are m aircraft in the air stack awaiting the opportunity to land and n in the ground queue awaiting takeoff. Suppose as before that these two waiting lines are restricted in length to the numbers M and N; then O<MM, On<N. (5.1)

The Poisson assumption of arrivals may still be made; however, we shall allow for the possibility of controlled arrivals (decreasing the rate of granting clearances to leave for the terminal as its queue lengths increase), and shall write, when the lengths are m and n, Rm,,(t) dt=prob{arrival into stack during t, t+dt} R2 (t) dt = prob{arrival into ground queue during t, t+dt}. For servicing (landing or taking off) we may assume, according to the results of (5.2)

Air-TerminalQueues

1103

Section 3, a Poisson law, and we write L.,, (t) dtt=prob leaving stack during t, t+dt5, L, (t) dt = prob leaving ground queue during t, t+dt 3 (

By virtue of the restrictions (5.1), these transition-probability coefficients have to satisfy the equations
Rn(t)0=

for m> M,

R, n(t)= 0

for n?N.

(5.4)

Similarly, by the principle ex nihilo nihil trahere,we must write

Finally, from the constraint implied by (5.1)

under any one of the cases m<0, n <0, m>M, n>N.

(5.6)

To write the equations of evolution, we observe that, to terms of higher order than the first in dt, the only possible transitions out of the state (m, n) are the

(rm-I, n)
*

(Mr-I, n)
~L

m,n n

mn

Lf .(m, n-I)

msn
Figure 7

man
(mi,

n-l1)

following, with the indicated transition probability coefficients shown in Fig. 7. There is, of course, the possibility of no transitions during t, t+dt, the probability of which is
1 (Rmn+R,'n L ?Lmn?+L, a,) dt.

By the usual type of reasoning, we express the probability that the system be in the state (m, n) at the epoch t+dt in terms of the probabilities of all the events that could lead to this state, starting from the possible ones at t. The usual manipulations lead to the following system of (M+ 1) X (N+ 1) differential equations:
dPmn (t)/dt (=Rmn +R ,n+Lmn+Lm,n)Pm, (t)

(t)(57 (t) +Rm-lanPm-1,n +R'n lpm~n-l

These equations will be modified at the boundaries of the rectangle of states (m, n) by the replacement of certain of the terms by zero, in accordance with (5.1)- (5.6). Evidently the number of equations goes up geometrically. with the number of = queues; if as in Sections 1 and 2 the queue-size limits are AL N = 25, there will be 262= 676 equations. Clearly, straightforward numerical computation would not

1104

Bernard 0. Koopman

be economical (and Monte Carlo simulation far worse). Therefore, a directed method of approximation must be used. Before entering into it, some simple conclusions may be drawn. The form of the equations (5.7) is that of (M+ 1) X (N+ 1) linear differential equations of the first order, homogeneous, and with coefficients given functions of the independent variable. Furthermore, they are stochastic equations: nondiagonal coefficients nonnegative; and each column [i.e., for each (m, n)] adding up to zero. Therefore, they possess a unique solution for each set of initial values; and if the latter are probabilities (nonnegative and adding up to unity), the same will be true of the former. Finally, if the coefficients are periodic functions of the time of period T =24 hours, the same will be true of just one of the solutions, all others approaching it with increase of time. These facts are all proved along the, lines of the same ones in the cases of Section 1. (See Appendix I.) We now reduce equations (5.7) by the method of 'marginal sums.' On setting

Pm(t)= Eon=u Pm,n(t),

(5.8)

it is realized that Pm (t) represents the probability that at time t there are exactly m aircraft in the air stack-nothing stated with regard to how many may be in the ground queue. If we had summed over m instead of n, we would have obtained the corresponding probability for n aircraft in the ground queue at t; and, of course, the methods we shall illustrate in the case of the air stack will be applicable to the ground queue. Thinking of equations (5.7) written for a fixed m but for all n = 0, 1, * , N, and then added, we find, after using (5.1)- (5.6) and making certain rather evident

the formal manipulations, followingequationsfor Pm(t):

dPm /dt= RJmiPmi(t)- (R. (t) m+Lm)Pm +LmnPm+i (t) (t),

(5.9)

Rm= n-U Rm~n[Pm~n(t)/Pm(t)]* Lm-n.-O L.,n[P.,n(t)/Pm(tA] (5.10) In this reduction,use is madeof the fact 'thatthe pair of sums
=N t 'V-n -Lan=u R1,n mn(t), En=u
n-N

Rm 'n-lpm~n-lt

cancel, since, in view of (5.2) and (5.4), the last term in the formerand the first term in the latter are zero; the others are equal and oppositein sign. A similar cancellationtakes place for the sums
E-

LmnPmn(t),

En-u

Lmn+iPmn+l (t).

It is now necessary to interpret the expressionLmand Rmgiven by (5.10). Clearlythe ratio Pm,n (t)/Pm (t) is a conditionalprobability:the probabilitythat, at time t, there aren aircraftin the groundqueue,given that the numberin the air is m. Therefore,Lmis a 'weightedmean':strictly, the conditionalexpectedvalue of landing rate, given that the numberin the stack is m. Similarly,Rmis the conditionalexpectedvalue of the arrivalrate, given m in the stack.
In formulas (to terms of order At), LmAt=prob[a landing during (t, t+At),

given that numberawaitinglandingis m but that nothing is given concerning the numberwaitingon the ground]. SimilarlyforRmAt arrivals. and Evidently Lm, Rm depend, not only on the time t, but on the whole solution

Air-TerminalQueues

1105

Pm,,n(t) (for all m, n, t), and therefore, strictly speaking, (5.9) cannot be solved without fully solving the original equation (5.7). However, while Lm, Rmmay have this degree of complication, it is not necessary to find them exactly, but only at the level of approximation consistent with that in which the other quantities are known. Simple reasoning can indicate their order of values: during dull periods, there is little interference between the two uses of the landing strips; then Lm,= A. During busy periods, when both X and X' are large, turns have to be taken and, on the average, the same number will take off as will land. Since the mean times taken for these operations are 1/il and 1/1A,in every interval 14A'+1/1A there will be one landing on the average, at the rate Lm. Therefore Lm is the harmonic mean: 1/Lm = 1/A+ 11/s. What this means is that, if the operating procedure aims at avoiding a greater average accumulation of aircraft in one queue than in the other, the same number, on the average, will have to be serviced per hour: the expected rates will be the same as if they are serviced in strict alternation, giving the above harmonic mean for both Lm and for Lm,. This assumes that, at each time of day t, the arrival rates X(t) and X'(t) are approximately equal. In the contrary case, an easy argument shows that we must take the weighted harmonic mean, and use the formula = X (t)/Lm'. X (t) /Lm= X (t) /I+ X'(t) / (5.11)

To embody these ideas in a definite procedure, we have to decide what to regard as 'dull periods' and 'busy periods.' One way of doing this is to borrow some of the notions of the deterministic flow model of Section 3, and to regard a dull period as one in which the total inflow X+X' is less than the total outflow given by (5.11). This requirement reduces to (5.12) X/4+ X'//' < 1. When (5.12) is valid, we write Lm=pA; when invalid, we apply (5.11). We emphasize that the above paragraph is merely one example among many other possibilities, reflecting different queue disciplines.

6. CONCLUSIONS

Two GENERAL CONCLUSIONS follow from the facts set forth in this paper. The first is that the quantitative results derived from the single-air-terminal waiting line in the general time-dependent case, while highly sensitive to the servicing rate, are remarkably insensitive to the statistical assumptions concerning the law of service. This validates the use of the Poisson simplification, but not the ignoring of the random nature of the waiting line problem, since the deterministic flow model to which this would inevitably lead is, as we have seen, but a crude approximation to the realistic situation. The second conclusion is that the double queue (takeoff and landing) at an air terminal can be handled analytically on the basis of Poisson assumptions with variable parameters, writing the exact equations and then taking marginal sums, whereby the study is reduced to that of the single queue. In the latter, both the service rates and incoming rates are time dependent. But the numerical computations are perfectly tractable at moderate computing-time expense.

1106

Bernard 0. Koopman

Regarding the second conclusion, it should be noted that the same type of method can be used for the study of different queues occurring at different terminals between which aircraft travel. The simplification by marginal sums allows the effects of all queues upon a given queue to be studied as effects on the input coefficientsand these can be simplified. This will lead to the direction of the quantitative treatment of the many-air-terminal complex. Our closing remark, which concerns the philosophy of every quantitative treatment of this and every repetitive massive operation, regards the validity of the assumptions concerning the transitions from state to state with increase in time. However the system is treated-by whatever methods of analysis, computer simulation, and such like-it is always assumed that at each epoch the state of the system can be described in such detail that the probabilities of the transitions into the possible states in the succeeding epochs are determined quantities. It is immaterial in principle whether these are introduced in formulas, as we have done, or by a mechanical selection employing a table of random numbers. In other words, the transitions are treated as if they are Markovian. 11 The examination of any large operation of this sort, as for example an air terminal, casts serious doubt on such a simplified assumption. To bring out the point concretely, suppose that our air terminal at some given time has 13 aircraft in the air stack and 17 in the ground waiting line, and that a space opens up on the landing strip. Will permission be given to a landing or to a departing aircraft to use this space? Clearly this can depend on other factors than the mere numbers (13 and 17) of aircraft in the two waiting lines. For example, if the policy-the 'queue discipline'-is to take care of a group of planes in one line together first, and then to switch to the other, and if this is being done for the landing ones, the next few spaces may be given to the latter category; a later group of departing ones will have their turn. Similarly for other factors affecting the situation; and as a general matter, the mere numbers m and n in the air and ground queues do not in themselves determine the 'state of the system' completely enough to allow the prediction of the next state into which it goes-or even its transition probabilities. There is, however, another way of looking at the situation. Suppose that our air terminal is considered, not just on one occasion, but on a very large number of different occasions under similar conditions of weather, time of day, etc. Suppose that from this ensemble of samples a subensemble is drawn, for example, the cases in which the numbers in the two queues are m= 13 and n = 17. In this subensemble, a certain fraction would have the next open space assigned to the landing aircraft, while in the complementary fraction, to a departing one. In other words, this crosssampling in the ensemble will give perfectly definite probabilities of evolution; and these are thus determinedas functions of the numbers m and n (at the stated time t): the evolution is Markovian, although with variable transition probabilities. Such an idealized system with transition probabilities defined in this manner will be called a 'cross-cut' model. It will not tell us very reliably what the probabilities of evolution of the two queues are on some particular day, in which a host of special items of information are at hand that can supplement our information and alter the probabilities obtained by merely picking at random from our ensemble. But it can be highly useful as an indicator of the effects of various proposals for improving capacity and delay-on the average and in the long run. It will, for example, guide

Air-TerminalQueues

1107

policy in the choice of improved systems for handling traffic congestion, or optimal operating rules-and thus further the basic objectives of operations research.

APPENDIX

APPLICATION

OF STANDARD

MATHEMATICAL

METHODS

this appendix is to bring together various classical and recent mathematical facts and methods needed to justify and amplify the developments carried out in this paper in its practical orientation.
THE OBJECT OF

The differential equations of evolution of Sections 1 and 5 are ordinary stochastic equations, for which the existence and uniqueness of probability solutions for any given set of
The standard methods are complicated by initial values is examined in the literature.[7] In the cases of present concern, a very simple treattheir objective of extreme generality. ment suffices, which we outline using the formulas of Section 1; the methods and results are altogether similar when applied to those of Section 5. Our equations (1.1), (1.2), (1.3) form a system of m+1 first-order linear homogeneous The classical elementary treatment by differential equations with continuous coefficients. successive approximations shows that, for any set of given initial values Pn(O), one and only one set of solutions Pn(t) exists. By summing the equations (1.1) over n, it is found that the derivative of the sum EnPn(t) is zero. Hence, if for the initial values this sum is unity, the same will be true for every later value of t. The only complication is in showing This is accomplished by that when the initial values are nonnegative, so are the solutions. process just mentioned to a modified system of equaapplying the successive-approximation but in which the coefficients are all nonnegative. The tions, equivalent to (1.1)-(1.3), usual formulas show that then all the approximants are nonnegative; and similarly for their limit. The method of obtaining this equivalent system is to replace Pn(t) by the new dependent variable

Q. (t) = P.(t) expt-

[R.(t)+L.(t)]

dtl,

I1

Substitution into (1.1)-(1.3) shows that the whose nonnegativeness implies that of Pn(t). The remaining terms have Qn(t) satisfy a similar system but without the negative terms. modified but simple and nonnegative coefficients that can be written explicitly. In the examination of periodicity, we assume the periodicity of the coefficients (1.4). is positive And we assume, furthermore, that every coefficient appearing in (1.1)-(1.3)

throughout the interval 0?<t < T or at least is positive throughoutsome positive subinterval of
this interval. Let Pi,,(t) Then the method of succession approximationsE71 as conducted above shows

that every probability solution P,(t) is positive for t = T.

Pi n(O)=1

or 0 according as n=i

or n~i.

(1.2)

The set of these solutions corresponding to the values i =0, 1, * *, m forms a fundamental system of solutions, in terms of which every given solution P, (t) can be expressed as a homogeneous linear combination with constant coefficients

1108

0. Bernard Koopman
This classical result is a

Evidently {Ci} is a set of probabilities, the initial values of Pn(t). consequence of the classical uniqueness theorem.[71

Evidently a solution Pn(t) will be periodic with the period T of the coefficients of (1.1)(1.3) if and only if (1.4) P. (T) = P,, (?). This is a further consequence of the uniqueness theorem, since any solution satisfying (1.4) will continue to do so, and with the same initial conditions, when translated by the amount T: Pn(t)-*Pn(t+T). On writing in (1.3) and (1.4) Pi,n(T) =Ai and Pn(O) =Cn, our condition takes the form (n=O, 1, m) (1.5) D=o CiAi,.=Cn. Thus C: {Co,C1, , Cm} must be an eigenvector corresponding to the unit eigenvalue of the matrix A = IIA II. The corresponding results in the case of the difference equations of Section 2 are obtained more directly: the existence and uniqueness of the probability solution is established by recurrence, as noted in that section. On adding the assumption of the positiveness of the elements of the matrix A (except for the upper right-hand triangle of zeros) for m epochs, it follows that their products in (2.7) will have all elements positive, once s>m; since the number S of epochs in the 24-hour period T is always greater than m, we see that the matrix .* A1, which takes the place of the matrix of the same designation in the last A =A_1A2 two paragraphs, is positive; and, when the transition probabilities in the difference equations are periodic, this is evidently true of A. Hence, all the conditions of the preceding paragraph apply to this case also. In each case, our square matrix A is a stochastic matrix (nonnegative elements, having unit row sums). Moreover, every element is greaterthan zero. It follows by an elementary theorem[8] that a unique eigenvector C exists satisfying (1.5), and that every other solution Pj, (sT) will approach Cn exponentially (sometimes described as 'geometrically') as s-oo0 through integral values. It follows by the simplest mathematical reasoning that every solution, say Pif,(t), will approach the whole periodic solution uniformly at the exponential rate. A practical consequence of the last fact stated is that, in order to obtain the periodic solution numerically, it is quite unnecessary to compute all (m+1)2 elements of A: by following just one solution until it practically reproduces its m+1 values at the two ends of a Tinterval, one is assured of having the periodic solution. This is often called the relaxation method. With a judicious choice of initial conditions, the work may be greatly simplified. Thus in our calculations we started with a zero-length queue [Po(O)= 1] at a time of day of minimum arrivals (very small X). In the case of Terminal B, this essentially gave the periodic solution within 24 hours; with Terminal A it did not quite return after a single day. The above considerations show how Items A and B of Section 4 are calculated. To find the expected rate of turning arrivals away at epoch t, we multiply the probability of a full queue Pm(t) by the expected rate by which arrivals exceed departures. In the interval (t, t+h) the probability of one arrival is X(t)h and of one departure is u(t)h, to quantities of higher order in h; the three events are independent. Therefore the excess rate of arrivals is A(t) -g(t) when this is positive; zero otherwise. By numerical integration, the cumulative values are obtained, thus giving C. In the case of the fixed service time, a similar type of reasoning is used, but the numerical manipulations are far more complicated; in view, however, of the little difference between the results of this and the Poisson case, they are unlikely to be needed. Of course the deterministic flow case verges on the trivial, but it requires a table of values (or a graph) of the integral f X(t) dt. The last item, D, the expected time in the queue taken between joining it and landing, was calculated in the case of a constant g: if the expected number in the queue is n(t) at the

Air-TerminalQueues

1109

time of joining it and if the expected 'servicing' time for one aircraft is 1/,u (or c), the product n(t)/,u [or cn(t)] is the answer. When there is so long a wait that the rate of service changes, the above formula has to be modified: the difficulty is not in the probability reasoning but in knowing what time-dependent probabilities to use in describing the actual situation. Again, the deterministic flow model is easy to deal with. A final matter to be explained here is the precise background of the statement of the first paragraph of Section 2 that the exponential (or Poisson jump) probability density Po(t) = depot is the 'most random' law of withdrawal from the queue, in comparison with any other density, p(t), also defined for to 0 and having the same mean 17/. This is an immediate result of modern information theory. [9] Using the expression z log z in its usual sense when z >0 and as its limit, zero, when z =0, it is shown to be strictly convex. The information content (negative entropy) in a distribution such as p(x) is defined as

Go

p (t)logp (t) dt.

To show that this is greater for any density p than for the exponential density po, we first show that G=f p(t)log[p(t)/po(t)] dt

is strictly positive unless p(t) _po(t). This is because G may be written as theweightedmean of z log z, where z =p(t) /po(t) and the weighting factor is po(t). This fact can be written as p(t)logp(t) dt>
0

i
0

p(t)logpo(t)

dt.

The insertion of the exponential expression for po(t) and the use of the fact that both distributions have the same mean shows that the right-hand member of this inequality is equal to f0 po(t)logpo(t) dt. Thus the exponential distribution is that of least information-highest entropy or random.

APPENDIX II

MATRIX PROPERTIES USED IN SECTION 2

IN ORDER establish equations (2.4) and (2.9), we first examine the m determinants (n TO 1, 2, * ,m) a
a2/2! 1 a 0 1 0 0 ...
*

0 0

0 0

D,, (a)

...
aw1/(n-1)! 1

...
ane2/(n-2) 1.*-

... .

...
... -

...

... a
1

... 1
1

; DiD(a)=1.

The first step in their evaluation is to differentiate Dn(a) when n>1 with respect to a, using the rule of differentiation by rows (cf. below). The derivative of the last row is a row of zeros. The derivative of the next to the last row gives a row equal to the one above it; and this will continue to be true for all rows except the first. Therefore, the only nonvanishing determinant is the one in which the first row is differentiated; and this is evidently Dn-i(a). Therefore, for n>1, Dn'(a) =D1_1(a), whereas D1'(a) =0. Consequently,

1110
for the kth derivative

Bernard Koopman 0.

Dnk(a)==D. -(a),

k<n-1;

Dn (a) =0.

(Ill)

The second step in the evaluation is based on the obvious fact that Dn(a) is a polynomial in a; let it be expanded about a = 0 by the Taylor formula: D. (a) =Dn (O)+D.' (0)a+Dn" (O)a2/2!+ ** +D' 7-1) (O)an- /(n-1)!+ Applying (11.1), together with the obvious fact that D,(0)
=

(- 1)n-1, we obtain, for n!1,

D. (a) =an-1/(n-1)!-an-1/(n-

2)! + -

+(-l)na+

(- 1) 1-1.

(11.2)

It is now easy to establish formula (2.8). Upon the determinant of the matrix A(a) of (2.6), we operate as follows: First, add the first mr-I rows to the last row, which thereby becomes a row of units. Second, take out from each of the first m -1 rows the factor e-a, calling the resulting determinant, which contains no exponential factor, Am(a); so we have detA (a) =e-(m-l)aAnl (a). Third, expand Am(a) in terms of the elements of its first row, obtaining A (a) = D. (a) +Dm (a). -1 On combining these two formulas with (11.2), equation (2.8) follows. We turn now to the proof of (2.4). The augmented matrix llp,?l, (2.3), is the following matrix of m +1 rows and m +2 columns
PS+1 ps0l Pi+
ps~l
[aml/

lf1l, as found from

e-a
aeo
m-l

~
I~-a
[a8l

-~~a em a ae-( ?(
e

0...

e ]

~
ea
..*

-1

pmM

_ 1)!z

-a~/(m2 - t2/( -2

-a e

For consistency of the equations (2.3) it is necessary and sufficient that the p8+l quantities obtained be such that the rank of this matrix be the same as the rank of the square matrix MVI from the one above by removing the first column.t81 To apply this criterion, it is sufficient to apply it to the equivalent modified matrix obtained from the one above by the elementary transformation consisting of adding all the first rn rows to the last, which thereby becomes a row of units. The (m+1)-rowed determinant on the right, equivalent to Ill, vanishes, having its first two columns equal. On the other hand, the lowest m-rowed determinant on the right is not zero, and in fact can be evaluated by first taking out the (mr-1) factors ena from the rows, and then noting that the result is Dm(a). The value is therefore e-"?-5)D(a). Thus the rank of M11 m. is In order that the augmented matrix also have the rank m, it is necessary and sufficient to show that every m+1 rowed determinant in the augmented matrix (or its equivalent) having the nonvanishing r-rowed determinant considered above as a minor vanishes.[",1 Clearly, this is equivalent to the equation obtained by erasing the second column of the (modified) augmented matrix, and setting the resulting (m+1)-rowed determinant equal to zero. On expanding this determinant in terms of the elements of the first column, the desired linear relation (2.4) is obtained. 8 We have already obtained the coefficient of po To find the coefficient of pt+l, we must take (-_ )n times the determinant having the last m columns of the (modified) augmented matrix, but with the single row numbered by n removed. Consider the first n rows of the latter determinant. All but one of the n-rowed determinants in these rows is zero, the exception being the one composed of the first n columns; and we see at once that this is unity. Therefore, the Laplace expansion1"1of the required determinant in terms of the determinants

Air-TerminalQueues

1111

in these first n rows gives, simply, this unit determinant, multiplied by the determinant in the last m-n rows and columns. This latter is evidently Dmn(a). Therefore, the coefficient of p1-1 is (-_)ne-(m-l)aDm-n(a). Finally, the cofactor of the (unit) last element in the first column is (-1) me-ma. On writing the resulting linear relation and dividing out e-(m1)a the required formula (2.4) follows. It may be noted that in this reasoning it has been assumed that, when certain determinants are not identically zero, they are not equal to zero; this assumption causes no restriction in the present applications. Finally, the rule of differentiation by rows (columns) of an n-rowed determinant of variable elements-which gives its derivative as the sum of n determinants, obtained by differentiating the elements of just one of its rows (all n choices)-is an immediate consequence of the definition of a determinant as a sum of products, and the differentiation formula for the product of n variable factors.

APPENDIX III

SIMULATIONAND OTHER METHODS IN AIR-TERMINAL CONGESTION

SINCETHE SUBJECT of queues has become-both in number of problems and in volume of publications-one of the most extensive parts of operations research, our first step in attacking the air-traffic-control problems facing the FAA was to seek appropriate methods in the literature. Our requirements were four-fold: (1) Visibility and soundness of the reasoning, (2) adaptability to strongly time-dependent inputs and environmental conditions, (3) sufficiently precise and reliable results to serve as a basis for practical decisions (when will better ATC equipment be needed? what advantages will improvement U have over improvement V, under various stated conditions? etc.), (4) computability at reasonable cost. As a result, we were led to develop the methods of the present paper. Without attempting a detailed methodological review, a few words may clarify this decision. There is a frequently expressed view suggesting that the choice is between 'analytical' and 'simulation' methods-not only in the case of queues but in a vast number of civil and military problems in operations research. Moreover, current parlance tends to use the term 'analytical' in a highly restricted sense: essentially to obtain an explicit solution by formal manipulations, using the properties of standard functions and standard limiting processes, such as series, integral transforms, and generating functions. Such 'analytical' methods have a restricted applicability, since they require the data (e.g., the coefficients of the functional equations) to be in simple explicit form; and, even when applicable, they too frequently lead to results unreasonably difficult for numerical computation. A useful comparative critique of such restrictedly analytical methods in a problem of timedependent queues closely similar to that of Section 1, together with Monte Carlo methods, has recently been given by LEESE AND BOYD.[12l Actually, there are relatively few publicacations on time-dependent queues: in addition to the well-known paper of GALLIHER AND WHEELER,[131 which uses Monte Carlo simulation, there are others fitting essentially into the 'analytic' category as described above. [14' Actual mathematics has long outgrown the aforementioned restrictions in its analytical handling of the concrete problems of science: It formulates them in precise but general terms; it extracts the maximum practical value from its theorems; and at the stage where numbers are needed, these theoretical preliminaries make possible the optimal use of modern com-

1 12

Bernard 0. Koopman

puters. It is in this broadened sense that the present paper is 'analytical.' Thus it assumes no explicit formula for the coefficient X(t), but derives from its periodicity the periodic properties of solutions. Turning next to methods of simulation, the situation has been summarized as follows by Leese and Boyd (loc. cit., p. 1) ". . . the run-to-run variations in the results obtained from Monte Carlo simulations are frequently so large that one cannot obtain reliable estimates of the required quantities without using an extravagant amount of computer time." This is natural: Monte Carlo transforms problems of calculating probability quantities to problems of sampling-with all the limitations and uncertainties of this subject.3151 To give a concrete illustration, suppose that we wish to compare two suggested improved airtraffic-control equipments, U and V, each very expensive in money and installation time and training. If their direct effect can be expressed as changes in the coefficients in the equations examined in this paper, one measure of their benefitswould be in the values of quantities A, B, C, and D of Section 4. Consider A, the expected length n (t) of the queue at various hours of the day and its standard deviation o(t). The Monte Carlo method is easy to describe: the sample number of aircraft moves through the machine for the 24-hour period, moving step by step through the S steps into which the day is divided. In the Runge-Kutta method, about 900 multiplications (the time-consuming items) were needed for each step: 900S in all. In the Monte Carlo, only 6 multiplications per step-only 6S per run! But whereas the former computation gave all the needed results with the accuracy commensurate with the input data, a large enough number R of repetitions of the Monte Carlo run are needed for reliable accuracy: We must find R and then compare 900S with 6RS. Figure 4 indicates that a value of =5 may be expected during the important times of day. If we are to compare queue lengths Nu and Nv with U and V, we may stipulate enough trials to give a standard deviation in these lengths of h for one particular epoch t. This would require an R given by 5/V/? = } or R = 100; this is two-thirds of the machine time for the deterministic calculation; but the figure is deceptive for two reasons. The first reason is that the figure R = 100 is based on knowing that 0f = 5 (approximately); and this was computed deterministically by the same process that did in fact give all the other quantities without simulation. If the value of a had not been known, and we had had to rely on the Monte Carlo process alone, it would have been necessary to obtain the moments A and 0-2 by statistical sampling from a distribution of an unknown type-i.e., without a given 'parent population.' An assumption tending to the optimistic is that the latter is approximately normal. Then the sample variance in estimating the standard deviation by mean residuals, i.e., (M2 - In12), is very nearly 2&f4/R. If a = 5 and we want a sample variance of 1/4,we require that R = 5000, which would take 33 times the machine time for the simulation as for deterministic calculation. The second reason that the optimistic figure R = 100 is deceptive is that what is needed for our decision between U and V is a comparison of expected queue lengths, not just for one epoch, but throughout the busy time of day. Without going into the elaborate study of reliability for two whole sequences of S statistically estimated numbers, it is pretty obvious that the number R of runs would be considerably increased over the above estimates. Finally, no amount of simulation could ever prove the periodicity situation established above by straightforward mathematical reasoning. Simulation can suggest a situation but it takes mathematics to prove it. ACKNOWLEDGMENT
THE WORKupon which this paper is based was done at Arthur D. Little, Inc., for The author is unqiuely responsible for all the Federal Aviation Administration.

Air-TerminalQueues

111 3

views expressed in the paper. The author takes pleasure in acknowledging the help of his colleagues, in particular of ASHA S. KAPADIA, SIMONLISTER, and SUSAN L. LEWIS.

NOTES AND REFERENCES

1. Tables of this Incomplete Gamma function are to be found in Handbook of Mathematical Functions, National Bureau of Standards, Applied Mathematics Series 55. 2. Contained in an unpublished FAA Staff Study of Airport Congestionin 11iajorMetropolitan Areas, made available by the Federal Aviation Administration in connection with contract work by Arthur D. Little, Inc. The statistics were graphed as follows: for each interval between consecutive hours (e.g., 10 and 11) the average number of events was found and plotted as the ordinate of a point whose abscissa was the corresponding midpoint (e.g., 10:30). The resulting 24 points were then joined by straight-line segments. 3. See A Study of Air Traffic ControlSystem Capacity, Arthur D. Little, Inc., Interim Report, Sept.-Aug. 1970, Report No. FAA-RD-70-70 [FAA Contract No. FA70WA-2141], Chapter 6. 4. See preceding reference, Chapter 3. 5. See Analytical Tools for the Study of Airport Congestion, Arthur D. Little, Inc., Interim Report, June 1971 [FAA Contract FA70WA-2141]. 6. B. 0. KOOPMAN, "A Study of the Logical Basis of Combat Simulation," Opns. Res. 18, 855-882 (1970). 7. For the existence, uniqueness, and other properties of linear differential systems, together with examples of the method of successive approximations, see, e g., F. J. MURRAY and K. S. MILLER, Existence Theorems (N. Y. Univ. Press, 1954) Chapter 6; in less complete form, see W. KAPLAN, Ordinary Differential Equations (AddisonWesley, 1961) Chapter 12. 8. See J. L. DOOB,Stochastic Processes (Wiley, New York, N.Y., 1953). The theorem used here is in Chapter V, Section 2, "Case (b)," p. 173. 9. See S. KULLBACK, Information Theory and Statistics (J. Wiley, New York, N.Y., 1969), for a modern treatment, including convexity, Chapter 2, Section 3. 10. See e.g., M. BOCHER, Introduction to Higher Algebra, (Macmillan, New York, N.Y., 1915), Chap. IV, Theorem 1, and Chap V., Theorem 1. 11. See M. BOCHER, loc. cit., Chapter 11.8. 12. E. L. LEESE AND D. W. BOYD, "Numerical Methods of Determining the Transient Behavior of Queues with Variable Arrival Rates," J. Canadian Opns. Res. Soc. 4, 1-13 (1966). 13. H. P. GALLIHER AND R. C. WHEELER, "Nonstationary Queuing Probabilities for Landing Congested Aircraft," Opns. Res. 6, 264-275 (1958). 14. The following publications may be cited in connection with time-varying queues: D. G. KENDALL, "On the Generalized Birth and Death Process," Ann. Math. Stat. 19, 1-15 (1948). A. B. CLARKE, "A Waiting Line Process of Markov Type," Ann. Math. Stat. 27, 452-459 (1956). G. LUCHAK, "The Solution of the Single-Channel Queuing Equations Characterized by a Time-dependent Poisson-Distributed Arrival Rate and a General Class of Holding Times," Opns. Res. 4, 711-732 (1956). ,"Distribution of the Time Required to Reduce to Some Preassigned Level a Single Channel Queue Characterized by a Time-dependent Poisson Distributed Arrival Rate and a General Class of Holding Times," Opns. Res. 5, 205-209 (1957).

1114

Bernard 0. Koopman

L. VON SYDOW, "Some Aspects on the Variations in Traffic Intensity," Teleteknik 1, 58-64 (1957). J. F. C. KINGMAN, "The Effect of Queue Discipline on Waiting Time Variance," Proc. Camb. Phil. Soc. 58, 163-164 (1962). D. G. TAMBOURATZIS, "On the Property of the Variance of the Waiting Time of a Queue," J. Apple.Prob. 3, 702-703 (1968). 15. Cf., e.g., M. G. KENDALL AND A. STUART, The Advanced Theory of Statistics (Hafner, New York, 1958), Vol. 1, Chapters 9, 12, 13.