Académique Documents
Professionnel Documents
Culture Documents
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.
Royal Statistical Society and Wiley are collaborating with JSTOR to digitize, preserve and extend access to Journal of the
Royal Statistical Society. Series B (Methodological).
http://www.jstor.org
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
Journalof the Royal StatisticalSociety
SERIES B (METHODOLOGICAL)
By DAVID G. KENDALL
MagdalenCollege,Oxford
[Received January,1951]
[Read beforethe RESEARCH SECTION of the ROYAL STATISTICAL SOCIETY, March 21st, 1951,
Professor M. G. KENDALL in the Chair]
Summary
THEpaperopenswitha generalreviewofsomepointsin congestion theory,and continueswitha
simplifiedaccountofthePollaczek-Khintchine "equilibrium"theoryforthesingle-counterqueue
fedbyan inputofthePoissontypeand associatedwitha generalservice-time distribution. It is
pointedout thatalthoughthestochastic processdescribingthefluctuations
in queue-sizeis not
(ingeneral)Markovian,itis possibletoworkinsteadwithan enumerableMarkovchainifattention
is directed
to theepochsat whichindividualcustomersdepart(theseepochsforming a sequenceof
regeneration
points).
Theergodicproperties ofthechainareinvestigated
withtheaid ofFeller'stheory ofrecurrent
events; it is foundto be irreducible
and aperiodic,and the further classification
of the states
dependson theValueofthetraffic intensity,
p (measuredin erlangs). Whenp < 1 thestatesare
all ergodicand thesystemultimately settlesdown to a regimeof statistical equilibrium,
the
transition-probabilities
converging to the values givenby the equilibrium solution. When
p > 1 the statesare all transient,
and whenp has thecriticalvalue of unitythe statesare all
recurrent-and-null.Thepapercloseswithsomefurther generalcomments, witha re-interpretation
of Borel's theoryof "busy periods" in terms of the Galton-Watson "branchingprocess" of
stochasticpopulationtheoryand witha sketchof an argumentleading to the distributionof the
lengthin time of a busy period.
1. Introduction.-Thebest way of opening this discussion on the Theory of Queues would
have been to presenta short but balanced surveyof the literature. I had initiallythis aim in
mind,but soon foundmyselfunequal to thetask. It is not so muchtheextentas theinaccessibility
of the literaturewhichmakes the difficulty;many of the most importantarticleshave appeared
in technologicaljournals,and a UniversityLibraryis not thebestplace in whichto look forthem.
Instead, therefore,I shall discuss one or two simple problemswhich are characteristicof the
subject,and it is myhope that the deficienciesof the provisionallist of referenceswill be made
good by those takingpart in the subsequentdiscussion. If Fellows will let me have the details
of those papers which have been omitted,it will then be possible to compile a more adequate
bibliographyfor publicationin the Journal.
The Theory of Queues has a special appeal for the mathematicianinterestedin stochastic
processes because it providesa simple example whichis both (i) stationaryand (ii) (in general)
not Markovian. Apart fromYule's autoregressiveschemesthereare not many otherexamples
whichpossess properties(i) and (ii) and are at the same timeso easy to analyse. Thus one may
expect in studyingthe queue to gain insightinto other stochasticphenomena,and to acquire a
valuable facilityin the handlingof the relevanttechniques.
A furtherattractivefeatureof the theoryis the quite astonishingrange of its applications. I
do not propose to mentionmantyof themnow because I am sure that theirvarietywill become
VOL. XII. NO. 2. M
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
152 KENDALL-SomeProblemsin theTheoryof Queues [No. 2,
apparentfromthesubsequent discussion,butit wouldbe unforgiveable notto mention thatthe
firstmajorstudyof congestion problemsseemsto havebeenthatundertaken byA. K. Erlang*
in 1908at the suggestion of F. Johannsen (himselfa pioneercontributor to the subject),and
undertheauspicesoftheCopenhagen TelephoneCompany(whoseassociation withinvestigations
in the mathematical theoryof probability continuesto-day,to our greatbenefit).My own
interestin thesubjectarosefroma correspondence withF. J. Anscombeand J. Howlettabout
queuesoftaxisin stationyardsand ofcustomers in retailshops,and so forthemostpartI shall
use theterminology appropriateto thesehomelyexamples.
Let me beginwitha verysimpleproblem, onewhichalwaysturnsup in thediscussion ifit is
notdealtwithat theoutset. Supposethattaxisand (single)travellers arriveat a queueing-point
inindependent Poissonstreams,theexpected numbers ofarrivals
perunitoftimebeingrespectively
axand P. Starting fromzerothesize of thequeueat timet willbe q = m - n, wherem and n
are respectivelyPoissonvariableswithmeanvaluescxtand Pt. A positiveq meansthatthereis
a rankoftaxis,and a negativeq meansthatthereis a queueoftravellers;whenq is zerono one
is thereat all and supplyand demandhavereacheda momentary balance. If z = P thenthe
expectedvalueofq is permanently equal to zeroand onemightexpectthesystem to remainin a
balancedstate,butthevarianceof q is 2oct, withthetime. The
and thisincreasesindefinitely
differencebetweentwo independent Poissonvariableshas been discussedby Irwin(1937) and
by Skellam(1946)and theirresultsin thepresentnotationshowthatthechanceof therebeing
a positiveor negativeor zero "queue" notgreaterin absolutevaluethanQ is
e2-at IIil(2xt) + 1
quo Q -V(4ntact)' 1
whichtendsto zerowhent tendsto infinityforeveryfixedQ, however large. Thissimpleexample
displaysa behaviourcharacteristicof manystochastic phenomena;whensupplyand demand
exactlybalance,theexpectation-values ofthevariablesare liableto tella misleading
story.
I turnnowto an abstractdescription ofa simplequeueingsystem ofmuchwiderimportance.
Hereoneis concerned, notwiththematching ofsupplyanddemand, butwitha seriesofcustomers
demanding serviceat a singlecounterand waitingin turnto be served. ("Customer"is used
sense; in practiceit might,
herein a technical forexample,haveto be equatedwith"aircraft".)
To completethespecification one mustgivea carefulaccountof:-
(a) the input-process,
(b) the queue-discipline,
(c) theservice-mechanism.
The simplesthypothesis about the inputis one whichstatesthatthe customers arrive"at
random'(i.e.,in a Poissonprocess),thenumberofarrivalsin timet beinga Poissonvariableof
tla, say. The time-interval
expectation u betweentwo consecutivearrivalswill thenhave the
distribution
negative-exponential
eu/a dula (O < u < oo), . . . . * (2)
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] KENDALL-Some Problemsin theTheoryof Queues 153
thereare several(sayN) counters,and it is onlyin thesecircumstancesthatmuchvariety in the
of thequeue-discipline
specification becomespossible. The customers maythenforma single
queue,themanat theheadmovingas soonas possibleto a vacantcounter. Alternatively (though
thereare yetotherpossibilities)
each customer on arrivalmaybe handeda ticketcarrying his
serialnumberm and directedto join thequeueformed up in frontofthejth counter,
where
m-j (mod N). . (3)
This meansthatthetotalinputis dividedamongtheseveralsub-queuesin a relatively simple
manner, and it is thenpossible(ifthetotalinputis Poissonian)to discussthebehaviour of one
of thesub-queues in isolationfromtherestofthesystem, itsowninput-processbeingof a more
generaltype(see?5). Therule(3) is rather butitsuseintroduces
artificial, a considerable
simplifica-
tionintothemathematical theory.In thispaper,however, exceptwherethecontraryis explicitly
stated,I shallbe concerned onlywithqueuesformed in frontofa singlecounter(N = 1).
Lastly,theservice-mechanism: letv denotethetimewhichelapseswhilea particular customer
is beingserved; I shallcall thistheservice-time, althoughin telephone-traffic
theory it is more
usual to call it theholding-time: It is naturalto supposethatthe successiveservice-timesare
statistically
independent ofoneanother andoftheinput,andthateachenjoysthesameprobability
distribution
dB(v) (O < v < so), . . (4)
say. Thereare twospecialcases of importance:
(i) (negative-exponential
service-time)
dB(v) e-vIb dv/b . . (5)
(ii) (constant
service-time)
B(v) = O (v <b); B(v) = 1 (v > b). . (6)
In eithercase b is themeanvalueof theservice-time;I shallreservethesymbolb forthis-use
alone,and I shallalwaysassumethatb is finite
and positive. Bothof(i) and(ii) wereconsidered
byErlang,and-healso employed an intermediatehypothesisinwhichtheservice-time distribution
has thex2-form
e- vk" dv (0 < v < o0.). (7)
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
154 KENDALL-SomeProblemsin theTheoryof Queues [No. 2,
of X(t) makesall information about its past historyirrelevant to a predictionof its future
behaviour.Sucha processmaybecomenon-Markovian if partof theinformation containedin
X(t)is suppressed, fora knowledge ofthepreviousbehaviour mayenablesomeofthesuppressed
information about the presentstateto be recovered.(The importance of thisfacthas been
pointedout byBartlett (1950); see also Ketidall(1950).) The congestion process(witha Pois-
sonianinputand a singlecounter)is Markovianifthepresentstateof theprocessis described
by thepair of randomvariablesq and vo,whereq is theinstantaneous queue-sizeand vois the
expended service-time of thecustomer at thehead of thequeue,butin general, it ceasesto be
Markovianifthestateof theprocessis measuredbythequeue-size alone.* The onlyexception
to thisstatement occurswhentheservice-time has a negative-exponentialdistribution;a special
and quitewell-known characteristic
property ofthisdistribution thenensuresthata knowledge of
theexpendedservice-time vois ofno predictive value.
Even witha generaldistribution of service-time,however, thestochastic processdescribing
thefluctuationsin queue-sizeis of thetypewhichBartlett and I (1951)havecalledregenerative.
"Regenerative process"is our (free)translation of Palm'sexpression, "Prozessemit begrenzter
Nachwirkung", and a stochastic processwillbe said to be ofthistypeifand onlyifit possesses
"regeneration points"(Palm: "Gleichgewichtspunkte"). An epochis a pointof regeneration
ifa knowledge ofthestateoftheprocessat thatparticular epochhas thecharacteristic Markovian
consequence thata statement of thepasthistory of theprocesslosesall itspredictive value. A
Markovprocess,therefore, is precisely
a processforwhichevery epochis a pointofregeneration.
In analysing a regenerative processtheidentification of its regenerationpointsis of funda-
mentalimportance, and a classificationof existingtechniques fromthispointof viewis of con-
siderableinterest(see,forexample,Bartlett and Kendall,1951; see also Ramakrishnan, 1951).
Theseideas have,of course,beencurrent in an implicitformformanyyears,butthefirstclear
formulation appearsto havebeenthatgivenin an important monograph byPalm(1943). More
recentlyFeller(1949b,1950)has developeda closelyrelatedtheoryof processesadmitting
"recurrent events".
For thecongestion process,whentheinputis Poissonianand whenthestateofthesystem is
measured bytheinstantaneous queue-size,theregeneration pointsaretheepochsat whichcusto-
mersleave,together withtheepochsat whichthecounterhappensto be free. An appreciation
ofthisfactwillthrowa gooddeal oflighton thesubsequent argumentt>s.
2. An elementary of theformulafor themeanwaiting-time
derivation in statisticalequilibrium.-
Supposethata queueformed in frontof a singlecounter is fedbya random(Poissonian)input,
p (= b/a)is lessthanunity,
and supposethatthetraffic-intensity so thattheinputdoesnotsaturate
the system.The epochsat whichthe customers leave are pointsof regeneration.Consider
such an epoch,and let q be the size of queue whichthe departing customerleavesbehind
him(thisnumber doesnotincludehimself, butit doesincludethepersonnextto be servedin his
place; of courseq couldbe zero). Let thenextpersonto be servedhavea service-time' v,and
duringthistimesupposethatr newcustomers arrive. Thenconditionally r is a Poissonvariable
of meanvaluev/a,and v has theservice-time distribution. When,thenextpersonleavesletq'
be thesize ofthequeuewhichhe leavesbehindhim. Thenin statistical equilibrium (ifsuchan
equilibrium solutioncan exist)the randomvariablesq and q' musthave the same marginal
theirmeanand meansquarevaluesmustbe equal.
and in particular
distribution,
The variablesq and q' are relatedbythefundamental formula
q' = max (q- 1,)?r, * (8)
whichit is moreconvenient
to writeas
q' = q -l+I . . . . . (9)
where8 = 8(q) is zeroforall non-zeroq, and 8(0)= 1. It is important
to notethefollowing
consequencesofthedefinition ofthefunction8(q):
82 = a and q(l-8)=q. . . (10)
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] KENDALL-Some Problemsin theTheoryof Queues 155
b Var (v) + b2
a 2a(a - b) . . .(2
It is temptingto thinkthat the firstexpressionfor E[q] may be valid for more generalinput-
processes,butthisseemsto be a fallacy; the argumentbreaksdown because r and q will not then
be statisticallyindependent(some furtherremarks on more general input-processeswill be
found in ?5).
From (12) it is quite easy to pass to the expressionfor the mean waiting-time.
Suppose thata departingcustomerleaves q customersbehind him and let his own waiting-and
service-times
be respectivelyw and v. Then q is the numberof arrivals in a total time w + v,
and so
E[q] = {E[w] + E[v]}/a. . * * * . (13)
From this it will be found that
E[w] p {I + Var (v/b)}, .(14)
E[v] 2(1 - p)
and thisis equivalentto thePollaczek-Khintchine
formula.
It is convenientto expresstheresultin thisway,becausetheratioon theleft-handsideis a
useful"figure of demerit"forthesystem(it is theratioof themeantimespentwaitingto the
meanservice-time waitedfor). If themeanservice-time is keptconstantand if thefrequency
ofcallsforserviceremains thesamethenthevaluesofb and a willbe fixed,
and in thesecircum-
stancesit followsfrom(14) thatmaximumefficiency
willbe obtainedifand onlyif thereis no varia-
tionin theservice-time.
Witha negative-exponential
distribution
of service-times
theratio(14)
is equal to twice the minimumpossible value, and so in this sense a systemcharacterizedby a
negative-exponential distributionis workingat only 50 per cent.efficiency.
service-time
For a fixed
formof service-time on theotherhand,the"congestion-ratio"
distribution, (14)
is a constantmultipleof p/(l - p), and the situationcan then only be improvedby a reduction
of the trafficintensityp (i.e., by a reductionin the mean service-timeor by a reductionin the
frequencyof calls for service). This implies a cQrrespondingincrease in the fractionof time
(1 - p) duringwhichthe counteris unused,*so thatwhilea reductionin the value of p increases
in one respect,it automaticallyreducesit in another.
the efficiency
In practicethesesimpleprinciples
will not alwaysapply,or at leasttheirworking.
maybe
obscuredby a numberof second-ordereffects. For example,in some circumstancesa reduction
in the mean service-timemay resultin a proportionateincreasein the ntumberof calls forservice,
so that p would tend to stay constant. When this situationobtains the only way of improving
the system;is to-modifytheformof the service-timedistribution(in the directionof complete
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
156 KENDALL-Some Problems
intheTheory
ofQueues [No. 2,
at themeanvalue),butthesimpletheory
concentration would,of course,cease to applyif the
customer
disgruntled marched backto theendofthequeue.
immediately
Othersecond-ordereffects
whichprovidean automaticcontrolreducingcongestion below
thetheoretical
valueare:
(a) theprematuredepartureofcustomers
whobecometiredofwaiting;
(b) thediscouraging
effect
ofa longqueueon an incoming
customer;
to servemorerapidlya customerwho has a long queue waiting
(c) the tendency
behindhim.
Similarremarks applywhentheotherapplicationsarebeingconsidered, although thereare often
instructive ofdetail.
differences
If one.continues
to assumethatthesystem can maintainitselfin statistical
equilibrium,then
itispossibletodeveloptheprecedingargumentina fairly
obviouswayso as toobtainthestationary
distributions forthewaiting-time
w of an arbitrary
customer and thequeueq whichhe leaves
behindhim. The generating functionfortheq-distribution
is
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] KENDALL-Some Problemsin theTheoryof Queues 157
- 7'Cia
Paj = T P ia 7CaS
I E J= 7'Cij
7'Cia7CaS * (20)
a=0 a=O a=O
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
158 KENDALL-SOmeProblemsin theTheoryof Queues [No. 2,
pathological behaviour are not realized, and to establish this fact two furtherresultswill be
needed.
(i)' The transition-probabilities
pn,j convergeto the limitsrrj in the ordinarysense if it
happens thatall the diagonal elementsof thep-matrixare positive.
(ii)' The systemwill be non-dissipative(so thatthe 7r-matrix will have unitrow-sums)
if
co
y- jPij < i . . . . . . (21)
j=O
foreveryi.
The firstof theseresultsis relativelya "deep" one; it followseasilyfromthetheorygivenin
Feller'gbookon notingthat(becauseofthestatedcondition) everystatein theMarkovchainis
necessarily "aperiodic".* (I owethishelpful remark to Mr.F. G. Foster.) The usefulcriterion,
(ii)' fornon-dissipation
is dueto Foster(1951); inwords,hiscondition requiresthattheexpected
increase'in is neverto be positive. In thesamepaperhe has givenone or twoother
state-label
conditions sufficient
fornon-dissipation,and I havegeneralized (ii)' to givea criterion
invariant
undera changeof state-labelling (Kendall,1951). It shouldbe observedthatneither of (i)' and
(ii)' makesanystatement aboutnecessary conditions.
The matrix{pij} defining the Markovchainassociatedwiththe congestion processhas a
particularlysimpleformt; it is
/0 1 2 3 4
0 ko k1 k4 k3 k4
O ko kL k2 k3 k4
2 ko kL k2 k3
3 . ko kl k2
4 . ko k
where
kr--E g e-v(a()dB(v) (r= 0, 1,2,. . .) . . . (22)
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] KENDALL-Some Problemsin theTheoryof Queues 159
It hasjustbeenshownthatthelimits7ii existin theordinary
sense; I shallnextshowthat
eitherall are zero, or else all are positive. For let 7ab = 0. Then since
pnf+M+Nab > pMai pnN,b
.pN . . (23)
it followsthat
Tab > PMai
> i
Nb . . . (24)
foreveryM and N. Now forgivena, b, i andj it is possibleto chooseM and N so thatthe
in (24) are bothpositive,and thusthevanishing
p-factors of -ab mustimplythevanishing of
wilforeveryi and j.
Becauseof (19) and becausethek's forma probability
distribution
witha finite
meanvalue
p, thethreepower-series
K(z) _- E k, zr
r=O
00 a,
I -,K(z)
L(z) 1 1 zK(z) - Zr k
-z r=O s=r+l
and
00
H i(z) v Zi 7ri (iE= O, 1, 2, .. . . . (25)
=0
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
160 Problemsin theTheoryof Queues
KENDALL-Somne [No. 2,
a0 ha Pai = hi (j = 0, 1, 2, . . . (33)
a=O
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] KENDALL-SomeProblemsin theTheoryof Queues 161
fromwhichitfollowsthat
oo n
E ha paj = hj
a=o
foreveryn. Accordingly
N n
hi < Y Pai + Y ha
a=o a>N
N n
< Y- Pai +
a=0
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
162 KENDALL-Some Problems
in theTheory
of Queues lNo. 2,
Theirreducibility of thechainmeansessentially that,givenanyorderedpairof states(which
neednotbe distinct), thereis alwaysa positiveprobability thatthesystem willmovefromthe
firstto thesecondstatein somefinite number ofsteps,and itimpliesthateverystateofthesystem
is of thesame"type"(see Feller'sbook,p. 322). It willtherefore be sufficient
to explainthe
situation (and to provethetheorem) in relationto thezero-statecharacterized bya zeroqueue-
size and a freecounter.
The aperiodicity of thechainhas alreadybeennoted; in orderto explainthebasis of the
further it is necessaryto introducetheprobability
classification fnij thatthe system (initially
in statei) willmoveintostatej forthefirsttimewiththenthstep(in similarly fntithe
defining
initialresidencein thestatei is to be disregarded).Despiteits obviousintuitive relationship
withthecomponents of thep-matrix it is usefulto havea purelyarithmeticaldefinitionoffnij,
and a recursive one is providedbythepairof statements
fii-_fli =pii and fnl+ =- Piaf nai . . . (36)
a =rj
wherethesummation is to be performed
overeveryvaluesavej ofthedummy a. It is then
suffix
a simplematter
to showthat
0 < 6 1,. . . . . (37)
and that
00
F,j E f n,j1< I (38)
n=1
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] KENDALL-SomeProblemsin theTheoryof Queues 163
and
O < Fio < I for every i, when p>I . . . (42)
and the proofsof (41) and (42) will be givenin a moment.* The resultsneeded to completethe
proofof Theorem4 followfrom(41) and (42) on puttingi = 0, and on notingthat all statesin
the chain are necessarilyof the same type.
Feller's theorygives a completeand elegantaccount of the statesin a Markov chain and it
enables one to surveyexhaustivelyall the possible types of behaviour. It does not, however,
provideone witha diagnosticprocedure,and when one is faced withthe problemof identifying
the haturesof the statesin some specificchain,it is usually necessaryto employsome more or
less ad hoc method. I shall now describe a systematictechnique for distinguishingbetween
transientand recurrentstateswhichis relatedto a methodof matrix-modification suggestedin a
similarcontextby Bartlett(1950); it seemslikelyto be of fairlygeneralapplication.
I shall presentthe method in a formwhich can be applied to any Markov chain with an
enumerableinfinity of states,and it will be necessaryto introducea new item of notation. I
shall writecpyijfor the n-steptransition-probabilityin the modifiedMarkov chain in whichthe
zero-state has been made "completelyabsorbing",all transitionsout of the zero-state being
forbidden. When n = 1 the 9-matrixthereforediffersfromthep-matrixonly in the elementsin
the top row, and it is, in fact,
, 0 1 2 3
0 1
Each termin the last series is non-negative(from(38)) and so Foo will be equal to unityif and
onlyif
Poa (1 - Fao) = 0
foreach a > 1.
* The proofof (41) terminates thissectionwillbe concernedwith
withequation(45), and thereafter
theproofof (42).
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
164 KENDALL-SomeProblemsin theTheoryof Queues [No. 2,
Theorem
6.-In anyenumerable
Markovchain
Fio = (i> 1),
wherethesymbolon therightis an ultimate
transition-probability Markov
forthemodified
chainin whichthezero-state
is "completelyabsorbing".
The twotheorems together enableone to determine
the(recurrent natureofa given
or transient)
a singlecolumnofthe7t-matrix
statebyevaluating fortheappropriately
modifiedchain.Theorem
6 is an immediate
consequence of theidentity
n
I ft`O = 9'io (i > 1), . . . . . (44)
m= I
whichis triviallytruewhenn = 1. Assumingthetruthof (44) forsomeparticular
value of
n > 1, itstruthin generalcan be established
byobserving
that
n+1 n
Z fm iO = fio + v Z Piaftao
m=1 mn=1 a*O
1 ko k, k2 k3 k4
2 . ko k, k2 k3
3 . . ko k, k2
4 . . ko k,
and a littleconsideration
of thiswillmakeit clearthat9'%jO is a monotone-increasing
functionof
n foreachi. Moreover, thezero-stateis accessiblewithpositiveprobability fromeveryother
statein somefinite numberof steps,and so ?n"o iS positiveifn is sufficiently
large; thisshows
thatTc*iois positiveforeveryvalueof i (in particular,
Tc*oo= 1). Now the7*-matrix is idem-
potent,and so (foreveryi)'
7 iO = 7* iO+ z Tia 7*aO,
a*O
fromwhichit followsthattheTc*-matrix
is identically in theextreme
zerosavefortheelements
left-handcolumn (all of whichare positive). So far no appeal has been made to the value of the
trafficintensity,
p.
It is nowto be observed
thattherow-means
ofthe9-matrix
are
00
y- 1Po,= 0
J=0
and
00
v j ij = i(-p) > 1),
o
,-o
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
19511 KENDALL-Some Problems
intheTheory
ofQueues 165
so thatFoster'ssufficient
conditionfornon-dissipation
is satisfied
if p < 1. Withthislimitation
on thetraffic
intensity, therow-sums
therefore, of the7*-matrix willbe equal to unityand so*
(byan appealto Theorems 5 and 6)
and it is clearthatFoster'scriterion
willbe of no assistance
herebecauseit does notprovidea
necessaryconditionfor non-dissipation. In orderto obtaintheresult(46) a ratherdifferent
technique is required.
Let a new("condensed")Markovchainbe defined withtheone-steptransition-probabilities
/0 1 2 3
0 1
1 9l110 111 9112 1l13
The first
taskis to showthat
< 4)io
9tPnio (n 1, 2, 3,. . .) * (48)
truewheni = 0. Whenn = 1, thetruthof (48) is
foreveryvalueof i. The resultis trivially
a simpleconsequence of themonotone-increasingcharacterof y'ljo (forincreasing
n); suppose
therefore
that(48) is trueforsomeparticular valueof n > 1. Then,ifi > 1,
< = 4)n+1iOs
Y- tIiaQJ%O
a=O
and thiscompletes
theinductive
proofof theinequality.
NextI shallprovethat
00
Z Za pnia = Zi [K(z)/z]n . . . . . (49).
a=O
when I z I < 1 and i > n, whereK(z) is the power-seriesdefinedin (25). The resultis trivially
* This could have been inferred fromgeneraltheorems about "cascade processes"givenby Foster
(1951); see, in particular,
theremarkassociatedwithhis equation(10').
t It is easy to show by inductionthat cpnj = pn,, (all i > n), and so 4,j = pii (i > 1).
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
166 KENDALL-Some Problems
intheTheory
of Queues [No.2,
valueof n > 1. Then
thatit is trueforsomeparticular
truewhenn = 1; supposetherefore
00 00 00
ZaZpn4-1. = Izp (pn# E za-P
a=O =O a=O
E Z# I?n ij E Zs-' ks
,#=1 s=O
=z Z#qpn,#
[K(Z)/Z] = Zi [K(Z)lz]n+1
,6=0
is definsd
If thesequence{_7n} recursively
bythestatements
o1= lo=Plo = ko and N + = K(i)n) . . . . (51)
so that0 < iN < 1 foreveryn, then
4Jni = (0') * * * . (52)
)nAlO = y
'h,a qnaO = - (Nn)a ?q= [K(Nn)]=
0 (In+])i (i >1),
a=O a=O
stepwhichshowsthat(52) is trueforeveryn.
an inductive
FromtheresuLtsof thelastthreeparagraphs it followsthat
no < ' nio = (,'nYt * * (53)
and fromtheaperiodicity in thecondensed
of thezero-state chain(%00is positive!)it is known
that4n%otendsto a limitas n tendsto infinity.It followsthatthesequence{,Jn} mustconverge
to a limitn (say)satisfying
0 < ?< 1 and K(n~)= .
processes.*It is
familiarfromthetheoryof branching
One can now adapt an argument
knownthat
0 < ko = K(O) < 1,
and that
K'(1 - 0) = p > 1,
and so theequationK(x) = x has precisely one rooti in theinterval 0 < x < 1 in additionto
the root at x = 1. It followsthateitherr or X = 1, and it is not difficult
= to show thatthe
secondalternative is impossible.For
= K(O) < K(,) =
Thus,byan inductive
argument,
TN <
? forall n, andso n < < 1. Thisexcludesthepossibility
that I = 1, and so = .
* See,rorexample,
pp.223-7ofFeller'sbook.
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] KENDALL-SomeProblemsin theTheoryof Queues 167
thelastresultwith(53), it followsthat
On combining
O < Fio = r*io < it < 1 i>1)
whichwas to be proved.
5. Someclosingremarks.-InthispaperI havehad verylittleto say aboutthemuchmore
difficultproblems associatedwiththe"compound"queueformed in frontof N counters,where
N > 1. A greatmanywriters (fromErlangonwards)havediscussed questionsofthiskind,and
particular reference shouldbe madeto theseriesofpapersbyPollaczek(ofwhichthatpublished
in theAnnalesde l'Institut HenriPoincare(1946c)is especiallyuseful;it includesa summary of
Pollaczek'searlierwork). Theruleof queue-discipline associated with(3) wouldinmostapplica-
tionsrepresent a veryinefficient
procedure, butpresumably itis notaltogether to some
irrelevant
of thetelephone trafficapplications wherethequeue-control maybe effected by an automatic
switching in theliterature.It also providesan
device; it has receiveda good deal of attention
interesting linkwiththetheory ofthesimplequeuefedbya "regenerating" input.
If the service-times are distributedin the negative-exponentialdistribution(5) and if the
intervals betweenarrivalsare statisticallyindependent,any one such"arrival-interval" u being
distributed according to an arbitrarylaw
dA(u) (O< u<
u o)
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
168 KENDALL-SomeProblemsin theTheoryof Queues [No. 2,
theresultsgivenbyBellcan be onlyapproximations to thetruth,thoughtheywillbe exactifthe
inputis Poissonian.The difficulty can be statedin termsof regeneration-points (as above),or
alternativelymyobjectioncan be illustrated by explaining whyI now maintain(afterhaving
beenmistaken aboutthisforsometime)thatthearguments givenherein ?2 can notbe extended
to coverbotha generalinputand a generalservice-time. If thesearguments are re-examined it
willbe seenthata criticalstepis theobservation that-
(a) thequeue-size following thedeparture ofa customer,
(b) thenumber ofpersonsarriving duringthenextcompleted service-time,
are statistically
independent.Thiswillbe trueiftheinputis Poissonian, butwithmoregeneral
typesof inputone wouldexpecta statistical dependence between(a) and (b) whichwouldhave
tobe takenintoaccount. I expectto betoldthatnothing morethanan approximation is required
in theaircraft problem, butI thinknonethelessthatthedifficulty I havesketched is a realone,
and worthy of study.
Beforeleavingthequestionofa moregeneralinputI shouldliketo drawattention to a paper
by Kolmogorov(1931),whoconsidered themany-counter problemwitha negative-exponential
service-time and a variablePoissonianinput. In thisand in similarproblemstreatedin an
important memoirbyPalm(1943)thewholecharacter ofthesystem is allowedto varywiththe
time; forexample, thetraffic
intensity maybe suddenly raisedorsubjected toa periodicvariation.
Yetanother typeofinput-process resultsfromtheassumption thatthepopulation ofcustomers
is finite
and equalin number to M, say. If theycall forservicein a randommanner theinput-
process(whileq customers arewaiting or beingserved)willbe Poissonian, butwilldependon the
valueofq, theexpected number ofarrivals perunittimebeing(1 - q/M)/a. Whenever a customer
arrivesor departsthetraffic intensitychangesdiscontinuously, and theinputceasesaltogether
forso longas q has itsmaximum valueM. The situations ofthissortwhichoccurin telephony
are fullydiscussedin thebooksbyFry(1928)and Brockmeyer et al. (1948). A relatedproblem
occursin connection withtherepairto industrial machinery;forthissee Khintchine (1933),
Wright etal. (1936),Palm(1947),Feller(1950)andAshcroft (1950).
The differentpeopleassociatedwitha queueingsystem willassessitsmeritsand demerits in
differentways. The customer willbe mostconcerned withthewaiting-time, themanagement
maypay particular attentionto thequeue-size, and thepersonoperating theservice-mechanism
maybe mostinterested inthelengths ofslackandbusyperiods. It hasbeenseenthatinstatistical
equilibrium one shouldexpectthecounter to be unoccupied foran averagefraction (1 - p) ofa
giventime-interval (wherep is thetrafficintensityinerlangs),butitis possibleto saya gooddeal
morethanthis. If thecounteris freeat t = 0 and a customer arrivesat thatinstant, and ifthe
counterremainsoccupieduntilt = T (whena customer departsand thereis no one to takehis
place),I shallcall theinterval(0, T) a busyperiod. Busyperiodsalternate withslackperiods,*
andwithaPoissonian inputitcanreadily be seenthata slackperiodstarting at t 0= willterminate
at theepocht = T', wheretherandomvariableT' has thenegative-exponential distribution
e-T'IadT'Ia (O < T'< oo). (54)
ofthelengthofa busyperiodis rather
The distribution morecomplicated, however.
An analysisof thebehaviourof a simplequeue fromthisstandpoint was givenin a very
paperby Borel in 1942. He assumeda Poissonianinputand a fixedservice-time,
interesting
andhe obtaineda mostelegantexplicitformula forthedistribution ofthelength ofa busyperiod
in termsQfthenumberof participating customers.Thereis no difficulty in extending
his dis-
cussionto covera generalservice-time as soon as it is realizedthattheproblemis
distribution
closelyrelatedto a famousone in thetheoryof stochastic
population growth.
Let thecustomer whosearrivalinitiates
thebusyperiodbe calledthe"ancestor";he consti-
tutesthezero-order "generation".Duringhisservice-time letnl further customersarrive;'they
constitutethe first-order
generation.Duringtheirtotalservice-time let n2 further
customers
arriveto formthesecond-order generation,and so on. The busyperiodterminates whenthe
"family"becomesextinct;tWLe lengthofa busyperiodas measured byBorelis thetotalnumber
* A rather distinction
similar between occursin Ashcroft's
"clear"and"repair"periods discussion
(1950)ofthemachine-servicing
problem.
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] KENDALL-Some Problemsin theTheoryof Queue.s 169
Cpr+?(S,s) + 1 a } (58)
9cp= ts + 1 9t . . . . . . (59)
fortheLaplacetransform,
cp(S) E[e-ST], . . . . (60)
oftheoveralllengthT ofthebusyperiod.
of thedistribution
* See, forexample,Bartlett
(1949),Kendall(1949),Feller(1950)and Harris(1951)forcontemporary
reviewsof thisproblemand forfutther references
to theliterature.
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
170 KENDALL-Some Problemsin the Theoryof Queues [No. 2,
1
e-st i(t) = (R(s) > 1) (63)
Jo' t =s + VI(s2-1I)
one obtains
fromwhich,aftera fewtransformations,
e-('+ (b
0)TIb1 TIVp ) (O < T < oo) **..(64)
in theinterval0 < i < 1. Now 4"1ois the probabilitythatthe busyperiod will come to an end
withthe (n - I)st generation,if not sooner,: so that i is indeedthelimitingprobabilityrequired.
The busy-perioddistribution(64) (for the negative-exponential case) is chieflyof interestin
the regionof large values of T/b,and thenif p is not too small one can use the asymptotic
approximation,
prob{T/b> } 2/' p34 e (68)
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] KENDALL-Some Problemsin theTheoryof Queues 171
Traffic p
Intensity, of
ValueexceededbyT/b withprobability
(erlangs) 0.1 0.01 0.001
References
(a) General.-The followingpapersare referredto in thetextbut theyare not directlyrelevant
to congestiontheory.
BARTLETT, M. S. (1949),"Someevolutionary stochastic processes",J.R.Statist.Soc. (B), 11,211.
- (1950),"Recurrence times",Nature(London),165,727.
- and KENDALL, D. G. (1951),"On theuse of thecharacteristic functional in the analysisof some
stochastic processesoccurring in physics and biology",Proc.Cambridge Phil.Soc.,47, 65.
DOEBLIN, W. (1939), "Sur deux problemes de M. Kolmogoroff concernant les chainesdenombrables",
Bull.Soc. Math.de France,66, 1.
DOETSCH,G. (1947),TabellenzurLaplace-Transformation undAnleitung zumGebrauch.Berlin: Springer.
DOOB, J.L. (1942),"Topicsin thetheory of Markoff chains",Trans.American Math.Soc., 52, 37.
(1945),"Markoff chains-denumerable case", ibid.,58, 455.
ERLANG, A. K. (1929),Problem15 in Matematisk Tidsskrift,B, p. 36. (This is Erlang'sversionof the
problemoftheextinction ofsurnaTnes.)
FELLER, W. (1949b),"Fluctuation theoryof recurrent events",Trans.American Math.Soc., 67, 98.
FOSTER,F. G. (1951),"Markoff chainswithan enumerable numberof statesand a class of cascadepro-
cesses",Proc.Cambridge Phil.Soc., 47, 77.
GOOD, I. J.(1948),"The number ofindividuals in a cascadeprocess",ibid.,45, 360.
HARRIS, T. E. (1948),"Branching processes", Ann.Math.Statist.,19,474.
- (1951),"Some mathematical modelsforbranching processes",Proc. SecondBerkeleySymposium
on Math.Statigtics andProbability.
IRWIN, J.0. (1937),"Thefrequency-distribution ofthedifference between twoindependent variatesfollowing
thesamePoissondistribution", J.R.Statist.Soc., 100,415.
KENDALL, D. G. (1948),"On theroleof variablegeneration timein thedevelopment of a stochastic birth
process",Biometrika, 35, 316.
- (1949),"Stochasticprocessesand populationgrowth", J.R.Statist.Soc., (B) 11,230.
(1950),"Randomfluctuations in theaoe-distributionofa population whosedevelopment is controlled
bythesimplebirth-and-death process",ibid.(B), 12,278.
- (1951),"On non-dissipative Markoff chainswithan enumerable of states",Proc.Cambridge
infinity
Phil.Soc.,47, 633.
KOLMOGOROV, A. (1936), "Anfangsgriinde der Theorieder Markoffschen Kettenmitunendlichvielen
m6glichen Zustanden",Matem.Sbornik.(N.S.), 1, 607.
(1937),Bull.Univ.d'Etata Moscou(A), 1, 1(Russian). (According to Feller,thiscontains a complete
exposition oftheresultsstatedin thepreceding paper.)
RAMAKRISHNAN, A. (1951),"Some simplestochastic processes",J.R.Statist.Soc. (B), 13, 131.
SKELLAM, J.G. (1946),"Thefrequency ofthedifference
distribution between twoPoissonvariatesbelonging
to differentpopulations", ibid.,109,296.
STEFFENSEN, J.F. (1930),"Om Sandsynligheden forat Afkommet uddor",Matematisk Tidsskrift,B, 19.
- (1933),"Deux problemesdu calculdes probabilites", Ann.Inst.H. Poincare, 3, 319.
YOSIDA,K., and KATUKANI, S. (1940),"Markoff processwithan enumerable infinitenumberof possible
states",Japanese J.ofMaths.,16,47.
(b) The Theoryof Queues.-In preparingthislist I have aimed at completenessbut thereare
probablymanyomissions. (Papers concernedonly withthe probabilityof loss in "busy-signal"
telephone systemshave been deliberatelyomitted.) The most comprehensivereviews of the
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
172 KENDALL-Some Problemsin theTheoryof Queues [No. 2,
subjectare those by Brockmeyeret al. (1948), Fry (1928) and Pollaczek (1946c). A briefintro-
ductionfromthemodernpoint of view has been givenby Feller (1949a), and a number of special
problemsare discussedin his book (1950).
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] Discussionto Mr. Kendall'sPaper 173
whichwere
(added inproof).-The followingis a list of thereferences
(c) Additionalreferences
sent to me afterthe meetingin answerto my requestin ?1, togetherwithdetails of one or two
paperswhichhave appeared whilethepresentpaper was in thepress.
systems
BAILEY,N. T. J. (1952),"A studyof queuesand appointment inhospitalout-patient
departments,
withspecialreference to waiting-times",J. R. Statist.Soc. (A) (in thepress).
BENSON, F., and Cox, D. R. (1951),"Theproductivity ofmachines requiring attentionat randomintervals",
J.R.Statist.Soc. (B), 13,65.
BERNSTmN, P. (1941),"How manyautomatics shoulda manrun?"FactoryMaintenance andManagement,
99, 85.
BOUCHMAN, E. N. (1942),"Problemsof congestionin telephony", Akad.Nauk S.S.S.R. ZhurnalPrikl.
Mat. Mech.,6, 247. (Russian;Englishsummary.)
(1947),"The problemof waitingtime",ibid.,11,475. (Russian;Englishsummary.)(See remarks
byW. Fellerin Math.Reviews, 9, 194.)
DOMB,C. (1950),"The statistics ofcorrelatedevents,I", Phil.Mag. (7), 41,969.
FIELD, J.W. (1946),"Machineutilization and economicalassignment", FactotyMaintenance and Manage-
ment,104,288.
GREENSHIELDS, B. D, SCHAPIRO, D., and ERICKSEN, E. L. (1947),Traffic Performance at UrbanStreetInter-
sections,YaleBureauofHighwayTraffic Tech.Report, no. 1 (1-152).
GRoss,M. (1949),"Leistungsvorrechnung bei Mehrmaschinenbedienung", 4, 113.
Textil-Praxis,
JACOBAEUS, C. (1950),"A studyon congestion in linksystems", EricssonTechnics, 1950,no. 48 (pp. 1-68).
JONES,D. (1946),"A simplewayto figure machinedowntime",FactoryMaintenance and Management,
104,118.
KRONIG,R. (1943),"On timelossesin machinery undergoing interruptions", Physica,10,215.
- and MoNDRIA,H..(1943),"On timelossesin machinery uhdergoing interruptions,11",ibid.,10, 331.
PALM,C. (1937oc), "Etudedes delaisd'attente", EricssonTechnics, 1937,no. 2. (Thisis a translation into
FrenchofPalm(1937).)
(1946), Specialnummet fdr Telettafikteknik, Tekn. Medd. fian Kungl. Telegrafstylelsen (pp. 1- 10).
104,118.
des resauxtele-
de la theoriede l'encombrement
POLLACZEK,F. (1942),"Sur quelqueslois asymptotiques
phoniques", Ann. Univ.Lyon (A) (3), 5, 21.
d'operateurs
(1949a),"Application int6grocombinatoires de
multiples
dans la theoriedes integrales
Dirichlet", Ann. Inst. H. Poincare, 11, 113.
- (1949b),"Reductionde differents problemesconcernant la probabilited'attenteau telephone,a la
resolution de systemes ibid.,11, 135.
d'equationsintegrales",
RAFF,M. S. (1951),"Thedistributionofblocksinan uncongested streamofautomobile J.American
traffic",
Statist.Ass.,46, 114.
STOUT,H. P. (1950),"Synchronisation on machineefficiency",
and itseffect J. TextileInst.,41,225.
VANREST,E. D. (1951),"Problemsofevenflowin production" (unpublished).
WEIR,W. F. (1944), "Figuring most economical machineassignment",Factory Maintenanceand Manage-
ment,
102,100.
D. R. (1948),"The effect
WESTGARTH, in theoperationofweavingmachines"(unpublished).
ofcongestion
DISCUSSION ON MR. KENDALL'S PAPER
Mr. D. V. LINDLEY: It is a veryreal pleasureto propose thisvote of thanksfora paper which
reallyis excellent. We have a lot to learn fromit, not only in statisticalmatters,but also on
the subject of how to write a paper.
I should like to raise two points both concerningthe steady state solution, assumingit to
exist. The firstconcernsthecharacterof theprocessas exemplified by whatBartlettand Kendall
have called "regenerationpoints". In thispaper thereis unfortunately no mathematicaldefinition
of thisterm,but in the paper wherethefirstmention,in translation,occurs,the authorsdefineit
as a point towhere,forall t > to
- Distr. {X(t) I X(to)} = Distr. {X(t) I X(X) forall <6 to}.
This is confirmedin the presentpaper.
If thesystemis measuredby theinstantaneousqueue size q(t), thereseem to me to be no points
for which the above equation holds except those whereq(to) = 0. The statementat the.end
of the firstsectionis thereforewrong,unless I have misunderstoodthe concept,and the epochs
at whichcustomersleave are not "regenerationpoints". To see thatthesepointsare in no way
special it sufficesto considerinsteadof the epochs at whichcustomersleave, the epochs at which
customershave been waitingto, for some fixedto. One can carrythroughan analysis exactly
parallelto thatin equations(8)-(14) of thepaper,thoughthealgebrais considerablymorecomplex.
This shows,more than any argumentabout the meaningof regenerationpoints,thatthe epochs
used in this paper are in no way special. In fact,at these epochs one knows q(to), and that it
has just been diminishedby one' This additional informationis vital.
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
174 Discussionto Mr. Kendall'sPaper [No. 2,
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] Discussionto Mr. Kendall'sPaper 175
A simpleheuristic
argument showsthattheasymptotic ifitexists,forthewhole
distribution,
processwillbe in fact
,(ho + hi), (hi + h2),
i ho,
A rigorous butit is worthconsideration.It is rathernicely
proofof thismightbe difficult,
illustratedby the simplequeueingsystemwhichhas both negative-exponential arrival-time-
intervalsandservice-times,
withmeansrespectively of a, b (a beingthegreater).Thisis thecase
discussedby Mr. Kendall,wheretheprocessas.a wholeis Markovian;it is therefore easyto
calculatetheasymptotic distribution,Confiningour attention to thoseepochsat whichjumps
takeplace,we see that,at anytime,thequeue-size, n, has probabilities,
a b
a+b'
a aa+b'
whenn is greater
or equal to 1,of becomingat thenextjumprespectively n - 1,n + 1,withthe
zeroqueue-size
havingprobability oneofbecoming 1atthenextjump. Itisfairly clear,intuitively,
thattheasymptoticdistribution forthewholeprocesswillbe thatforthesimpleMarkovchain
describingthesystemat thissequenceofepochs,marking andarrivals. This
all thedepartures
Markovchainhas thematrix
O 1 0 0 0
OI 0 1 0 0
A= 0 .IL 0 ?x 0
o o 0 ?x
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
176 Discussionto Mr. Kendall'sPaper [No. 2,
PI
;i,,
E F
0) in
o 0E
4-J ~~~~~~~~4-,
1
o
0
LO ~~~~~~~L
0 LA 0 'A
n
X . 0e
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] Discussionto Mr. Kendall'sPaper 177
SystemA SystemB
Frequencies Cumulated Cumulated Frequencies
Queue , K Expected Expected A
size Observed Expected Frequencies Frequencies Expected Observed
(%) ( 0)
0 34 33.3 33.3 33-3 33*3 31
1 30 31-6 64-9 55 6 22-2 17
2 20 18-3 83-2 704 14-8 21
3 11 9.0 92*1 80-2 9*9 10
4 4 4-2 96-3 86-8 6-6 8
5 1 2-0 98-3 91-2 4-4 3
6 0.9 99-2 94-1 2-9 2
7 0 4 99-6 96-1 2*0 2
8 0-2 99-8 97:4 1-3 1
9 0.1 99*9 98-3 0 9 2
10 98-8 06 2
11 992 04 1
Mean
queuesjze 1-2 1-3 2*0 2-2
Thetableshowsthefrequenciesofdifferent forsystems
queue-sizes afterepochs
A, B, observed
ofdepartures,
and comparesthesewiththeirexpected
values. The twocentralcolumnscompare
thecumulated expectedfrequencies
(expressed
as percentages)
forthetwosystems.
w 5-
0 50
n
(w) of 100customers
FIG. 2.-The graphscomparethesequenceof waiting-times (n = 0 to 99),
havingidentical forthetwosystems,
arrival-times, A, B.
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
178 Discussionto Mr. Kendall'sPaper [No, 2,
a, a, aO 0 0 . . .
a2 a2 a0 aO 0. . .
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] Discussionto Mr. Kendall's Paper 179
and
{nP} nPo-
nP,
nP2
'P2 P2 P2'.
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
180 Discussionto Mr. Kendall'sPaper [No. 2,
Pois-
of note,whenthevehicleflowis completely
A specialcase of thisproblemis worthy
sonian,i.e., (3= 0. We thenhave
- ea -aq-1
w=
q
This case has beenconsidered obtainedforthenumberof
in moredetailand distributions
pedestrianswaitingand thesize of groupscrossing.The distribution function of the delay
hasan interestingformhere; itis notdifficult P(w) ofa delaygreater
to showthattheprobability
thanwsatisfiestherelation
P'(w) = -aqe-aq P(w-).
(0,oc),
Fromthiswe findthatthereis a chancee-aq ofa zerodelay,and thatin theintervals
(Q, 2ac),. . . , the frequencyfunctionis successivelyconstant,and then a linear, quadratic,
. . . function of w. The onlydiscontinuities are at w = 0 and w= a.
I turnnowto questionsof vehicledelayand willmention twoconnected signals.
withtraffic
As faras we know,neither has beensolved,and we shouldbe grateful foranyideason thistype
of problem.
The operationof fixed-time signalscan be reducedto the following
traffic process:
We havealternate "green"and "red" periodswhichhavefixedthoughnotnecessarily equal
lengths,duringwhichvehiclesin a single-channelqueuewithconstant holdingtimeand random
arrivalscan and cannotpassthelights.
The distributionofdelaycausedbythesimpler situationwherethereis justone "red"period
surrounded "green"periodscan be found,butthegeneralproblemis stillunsolved.
by infinite
The secondproblemconcernsvehicle-actuated signals; theiroperationcan be ratherover-
simplifiedas follows:
Vehiclescrossingthedetector pad havea minimum separation(, and each one passingthe
intersection fora periodoc. However,if thereis a vehiclewaitingto
blocksit to cross-traffic
crossand thereis no gap greaterthanoain thestreamof vehiclesentering theintersection,the
waitingvehicleis automaticallygivenrightofwayaftera certainperiody. Garwoodhas given
thedistributionofdelayto an isolatedcrossingvehiclewhen(3= 0, butthegeneralcaseis ofgreat
interestin consideringoptimumsettings of signalcontrollers.
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] Discussionto Mr. Kendall'sPaper 181
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
182 Discussionto Mr. Kendall'sPaper [No. 2,
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
1951] Discussionto Mr. Kendall'sPaper 183
multiplicative)
processes, providedthattheinputis Poissonian. As a matterof factit is also
possibleto deducethedistributionoftheduration ofa busyperiodfromthesametheory.
For letvrbe theprobabilitythata customerwillhaveto waitr elements
oftimeafterarriving
at thecounter,wherethenumberof elementsof timein a timeunitis N (large). Let V(z) =
v0 + vIz + v2z2+ . . .. . Duringeach elementof timethereis a probability 1 - 1/aN that
no newcustomer joinsthequeue,and a probability1/aNthatonenewcustomer joinsthequeue.
Theprobability thatmorethanonecustomer joinsthequeuein oneelement oftimeis negligible.
as formingthetwophases of a specieswithalternating
We can regardcustomersand time-elements
generations.(In a sensemanis sucha species.)
The probability function
generating of one customer
(P.G.F.) forthechildren(customers) is
v(I - I + aN). The P.G.F. for the children(time-elements) of one time-elementis
This resultis equivalentto equation(59). [Put z = e-SIN, g(S) = V(y), V(z) X3(S), so
that V(y) = (- Nlogy). We get
ag
-Nlogy =N log I1a + N)NlogI1-N) S + aI1
Therefore
p= V(y)=- 3(-Nlogy)- (S + 1- )]
it is also possibleto treattheproblemas a gambler's
service-time
For a negative-exponential
ruinproblemin whichdrawscan occuras wellas winsand losses. Thisis a specialcase of the
randomwalkproblem. The lengthofthequeuecan be identified withthecapital- of a gambler
and his opponentmaybe assumedto be infinitely rich. Thisadditivestochastic process,then,
is a specialcase of a multiplicative
process.
The entiretheoryof queuescould be extendedto thecase of different typesof customers,
each typehavinga different distribution
service-time and a different inputdistribution.For
thisgeneralized problemthedistribution of thenumberof customers servedin a busyperiod
and of thedurationof theperiodcouldbe obtainedfrombranching theory withseveralsexes,
all female. Thosewho prefergenerating functions to moreadvancedmethodsmaysee Journ.
Roy.Stat.Soc., 11 (1949),271-3,and theErrataboundin vol. 12, no. 1. It is reasonableto
assumeindependent foreach different
Poissondistributions typeof individual.
Mr. Kendallpointedoutthatlongqueueshavea discouraging effect.It is also possiblethat
a queue,rather thanno queue,has an encouraging becauseofthetendency
effect tojoin queues
without knowing w~hattheyarefor.
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
184 Discussionto Mr. Kendall'sPaper [No.2,
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions
19511 Discussion to Mr. Kendall's paper 185
This content downloaded from 143.167.2.135 on Sun, 04 Oct 2015 20:38:18 UTC
All use subject to JSTOR Terms and Conditions