What is an Answer?

Author(s): Herbert S. Wilf

Source: The American Mathematical Monthly, Vol. 89, No. 5 (May, 1982), pp. 289-292
Published by: Mathematical Association of America
Stable URL: http://www.jstor.org/stable/2321713 .
Accessed: 21/01/2015 04:38

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 [email protected].

Mathematical Association of America is collaborating with JSTOR to digitize, preserve and extend access to
The American Mathematical Monthly.

http://www.jstor.org

This content downloaded from 128.235.251.160 on Wed, 21 Jan 2015 04:38:41 AM

All use subject to JSTOR Terms and Conditions
 WHAT IS AN ANSWER?

HERBERT S. WILF
ofMathematics,
Department University Philadelphia,
ofPennsylvania, PA 19104

In manybranchesofpuremathematics itcan be surprisinglyhardto recognize whena question

has,in fact,been answered. A clearcutproofof a theorem or thediscovery of a counterexample
leaves no doubtin the reader'smindthata solutionhas been found.But when an "explicit
solution"to a problemis given,it may happen that more workis needed to evaluatethat
" solution,"in a particular case,thanexhaustively to examineall of thepossibilitiesdirectlyfrom
the originalformulation of theproblem.In such a situation, otherthingsbeingequal, we may
justifiablyquestionwhether theproblemhas in factbeen solved.
Examplesof thissortcan turnup anywhere, but herewe will concentrate on problemsin
combinatorial mathematics, specificallythoseofthetype"how many- arethere?"Suchenumera-
tionproblemslie at theheartof thesubject,and it is important to be able to recognizesolutions
whentheyappear.The point,of course,is thatsometimes the"answer"is presented as a formula
thatis so messyand long,and so fulloffactorials and signalternations and whatnot, thatwe may
feelthatthediseasewas preferable to thecure.
An answerto suchan enumeration questionmaybe givenbymeansof a generating function,a
recurrence relation, orbyan explicitformula. Each oftheseis,in essence,just an algorithm forthe
computation of thecountingsequencethatis to be determined.
How do we judge theusefulness of such answers?Obviouslywe mightbe able to do many
thingswiththeanswer,suchas to makeasymptotic estimates,to discovercongruence to
relations,
delightin its elegance,and so forth.We're goingto restrict attention hereto the appraisalof
solutionsfromthepointofviewof howeasilytheyallowus to calculatethenumberof objectsin
thesetthatis beingstudied.
The qualityof such-formulas shouldtherefore be judgedby theusual combination of esthetic
and quantitative benchmarks thatareused on algorithms. In particular, thequantitative criterion
is thecomputational complexity: theamountofworkrequiredto getan answer.We suggesthere
thatthesamecriterion shouldbe appliedto enumeration formulas. We willsee thata corollary of
thisattitudeis thatour decisionas to whatconstitutes an answermaybe time-dependent: as
fasteralgorithms forlistingtheobjectsbecomeavailable,a proposedformulaforcountingthe
objectswillhaveto be comparably fasterto evaluate.
For concreteness, supposethatforeach integern > 0 thereis a set Sn thatwe wantto count.
Letf(n) = I Sn I (thecardinalityof Sn),foreach n.
Supposefurther thata certainformula has been found,say
f(n) = Formula(n) (n = 1,2 ...) (1)
in whichFormula(n)mayinvolvevariousmultiplesummation overvarioussets
signsextending
and variouscomplicated summands, etc.
In orderto evaluatethe "answer"(1), let's look at the competition.To insuremy own
immortalityin thesubject,I am nowgoingto showyoua simpleformulathat"answers"all such
questionsat once.If you'reready,then,hereit is:
f(n) = E 1. (2)
Sn

Well,anyway, thesummandis elegant,evenif therangeof summation is a bituntidy.

Now (2) is unacceptablein politesocietyas an answer,despiteitselegantappearance,because
f(n) thatwe
of thequestion,and it does notgiveus a tool forcalculating
it is just a restatement

In additionto his otheraccomplishments

(see thisvolume,
p. 4) theauthoris currently
one of theassociate
editorsofthisMONTHLY.
289

This content downloaded from 128.235.251.160 on Wed, 21 Jan 2015 04:38:41 AM

All use subject to JSTOR Terms and Conditions
290 HERBERT S. WILF [May
didn'thavebefore.It does illustrate an important point,though:thereis no counting problemfor
whicha formula does notalreadyexist,namely(2).
A firstcriterionforevaluatingan "answer,"then,mightbe that"Formula(n)" shouldbe an
improvement overtheinsightful contribution (2).
Whatis an improvement? As notedpreviously, we areconsidering a formula as an invitationto
computenumbers;as an algorithm, ifyouwill,thatdescribesa sequenceof stepsthatwilllead to
an answer,and itwillbe usefuliflesseffort is requiredto use it thanto use someotherformula or
algorithm.
Whatis effort? Now we'reapproaching firmground.The theoryof computational complexity
has been developedrapidlyin recentyearsand thereis no shortageof rigorousstandardsfor
measuring computing effort. Usuallywe adoptsomeindivisible unitsoflabor,suchas multiplica-
tionsor divisionsof numbersin a certainsize range,or bit operations, or function evaluations,
etc.,and thenwe expressthecomplexity of thecomputation bycounting howmanyunitsof each
kindoflaborneedto be performed in orderto getthejob done.
How do we comparethecomplexity of evaluating Formula(n)with(2)? Equation(2) calls for
thestraightforward listingalgorithm: produceall of themembersof theset Sn and countthem.
Hence thecomplexity of (2) is equal to theamountof computational effortthatis requiredto
produceall members of Sn.
Next,recallthatwe needa littlemoregenerality in theformatof the"answer."Formulasare
notalwaysthewaysolutionsaregiven.Two otherusefulmethods, forexample,arethemethodof
generating functions,and theuse of recurrence relations.So, in all cases,we will considerthe
computational complexity of usingwhatevertool is givenforthepurposeof computing f(n),
whether it be a formula, a recurrence relation, or other.
The functions thatwillbe comparedare,therefore:
Count(n)= the complexity of the algorithm forcalculating f(n), whetherit be givenby a
formula, an algorithm, et cetera,and
List(n) = thecomplexity ofproducingall of themembers of thesetSn,one at a time,by the
speediestknownmethod,and countingthem.
DEFINITION 1: We willsaythata solutionof a counting if
problemis effective
li Count(n) =
n -oo List(n)
Whatwe are sayingis thata formulaor whatever
is an effective
solutionof a problemif the
involvedin countingthemembersof Sn by meansof thatformulais asymptotically
effort small
comparedto theeffort all ofthemembers
ofconstructing ofSn,bythebest-known and
algorithm,
countingthem.
EXAMPLE 1: In thetheory ofnumbers itis oftensaid (see [5,p. 344]) thatthereis "no formula
forfr(n)(thenumberofprimesless thanor equal to n)." Now,ofcourse,thereis sucha formula,
namelytheone givenby (2) above.In thelightof thedefinition, though, we can givetheprecise
meaningof the above statement in the form"thereis no effective solutionto thequestionof
enumerating theprimesless thanor equal to n."
How good would a formulaor otheralgorithm for7r(n)have to be in orderto qualifyas
To listall oftheprimesbetween1 and n we can use one ofthefastformsof thesieveof
effective?
Eratosthenes,suchas [4]. That algorithm operatesin timeKn/loglogn. Hence a "formula"for
7r(n) wouldhaveto be fasterthanthatin orderto satisfy theconditionof definition 1.
Again,in complexity theoryone speaksof computational problemsas being"easy" or "hard"
depending on whether thecomplexity offinding a solutionis oris notofpolynomial growth in the
lengthof the inputbit string.This distinction was firstmade by JackEdmonds,and it is
responsibleformuchof theexplosivegrowth in complexity theorytoday.
To takea leaffromthatbook,then,a formula "solves"an enumeration problemiftheformula
yieldsnumbersafter"easy" calculations.To make this assumptionmeaningful we will now
restrict
attentionto thoseproblemsin whichf(n)growsfasterthananypolynomial in n. We will

This content downloaded from 128.235.251.160 on Wed, 21 Jan 2015 04:38:41 AM

All use subject to JSTOR Terms and Conditions
1982] WHAT IS AN ANSWER? 291

problems.
(pot frolynomial)
call thistheclassof VqT
Moreprecisely, then,we propose
DEFINITION 2: We willsaythata problemin theclass VqT has beenp-solvedifCount(n)is of
polynomial growth as n -* oo.
Respectively,then,we mayspeakofan enumerative problemas being"unsolved,""effectively
solved,"or "p-solved."
to notethattheonlyinputto thecomputation
It is interesting off(n) is n itself,and so the
lengthoftheinputbitstring is about log n bits.To ask forpolynomialgrowth in n is thereforeto
allow exponentially rapid growthaccordingto the customary standardsof complexity theory.
Nonetheless, thisweakercriterion seemsquitestemenough,as we willsee fromtheexamples.
EXAMPLE 2: considerthenumberof permutations
For quitea clearcutsituation, of n letters
thathave no fixed points.The well-known,and veryelegant, is
solution that
f(n) is equal to the
that
nearestintegerto n!/e, or equivalently,

f(n)= 2 (_l)' t (n= 1,2,...). (3)

i=O J-
and second,
This showsfirstof all thatwe are indeeddealingwitha problemin theclass VqT,
thatthisformulareallyis an answerbecausewe can computefromit in polynomial time,i.e.,the
problemis p-solved.
EXAMPLE 3: Next, here's an examplefromgraph theory.If we ask for the numberof
unlabelledgraphson n vertices, we quicklyfindourselvesusingtheFrobenius-Bumside theoryof
group actionon a set(e.g.,[1]).This theorycounts the classes
equivalence of a set thatis actedon
by a groupofpermutations, in termsof thesetsof elementsthatare fixedby thepermutations in
thegroup.
Obviously,if a certainpermutation has s(i) cyclesof lengthi, thens(l) + 2s(2) + 3s(3)
+ * = n is a partition of theintegern. The answersthatemergefromtheFrobenius-Bumside
lemma(or fromthe moregeneraltheoryof P6lya [2]) dependo;nlyon thecycles,and so they
dependonly on the partitionsof n. Typically,then,an answerthatcomes fromthistheory
involvesa sumof a more-or-less elementaryfunction extendedoverall partitions of n.
For instance,thenumberof nonisomorphic unlabelledgraphson n vertices is exactly([2])
1 _ _ _ _ n
_ _ _ _ _
f( nn)= n lS(I)s(l)!2s(2)s (2)! ...

where
n

g(7r) = (1/2) gcd(idj),s)(i)(si(j) s

s(2k + 1)}

and ?r: n = s(l) + 2s(2) + 3s(3) + * runsthrough thepartitions of n.

Evaluationofg, fora givenpartition of n, is clearlyofpolynomial The numberof
complexity.
termsin thesum(4) is p(n), thenumberof partitions of n, and thisis wellknownto growlike
A exp(BFni)/n.Thus the evaluationof the formula(4) is not a polynomial-time job, since
Count(n)growsfaster thanA exp(Kvn) for some constant K.
behaviorof thenumberof graphson n verticesis
How big is f(n) itself?The asymptotic
(n)
f(n) -2 2/n! (n -oo).

Hence thecriterionof Definition1 is amplysatisfied:theanswergivenby Frobenius-Bumside's

lemmais a drasticimprovement over(2). This problemis thereforeeffectivelysolved.It is not
p-solved,however,in the senseof definition2, because the formula(4) calls formorethan a
polynomialamountofwork.
questionof whether
Thisraisestheinteresting thereexistsa polynomial-time formulaforthis

This content downloaded from 128.235.251.160 on Wed, 21 Jan 2015 04:38:41 AM

All use subject to JSTOR Terms and Conditions
292 LOUISE E. MOSER [May

problem.Indeed,can one describea reasonableand naturalfamily of combinatorialenumeration

problemsforwhichthereis provablyno polynomial-in-n timeformulaor algorithm to compute
is thereanyrelationship
f(n)? Further, betweentheintractability
of theisomorphism problemfor
unlabelledgraphsand theapparentun-polynomial-ness of thecountingproblem?Thesequestions
are reminiscentof,butnotidenticalto,theconceptof #P-completeness discussedin [3].
In additionto the theoreticalinterestof the questionsof computationalcomplexityof
formulas, thereare practicalalgorithmic consequencestoo. Frequentlywe findthatin orderto
carryout a combinatorial algorithm we need somevaluesof therelevantcountingfunctions. In
[6],forexample,thereis an algorithm thatneedsto knowthenumberof unlabelledgraphson n
verticesbeforeit can beginto do itsjob. OtheralgorithmsneedBell numbers, and so forth.
EXAMPLE 4: We'llconcludewiththeexampleof thepartitionfunctionitself.Sincethesum(4)
extendedoverpartitionswas hard,whataboutthesumof 1 overthepartitions of n?
That one is easy,because we have any numberof simplerecurrence relations,generating
fromwhichthevaluesofp(n) can be rapidlycalculated.For instance
and so forth,
functions,
n
np(n) = I p(n - k)u(k) (n > I,p(0) = 1) (5)
k=1

wherea(k) is thesumof thedivisorsofk. Thisraisesat leastthehopethatthepreviousproblem

mayhave a polynomialtimeformulaalso, and it raisesthequestionof describing theclass of
functionsf of partitionsthathave theproperty thatthefunction
g(n), obtainedby summing f
of n, can be evaluatedin polynomial
overall partitions time.
Addedin proof:Afterreadinga preprint
of thisarticle,Jeffrey Lagariasand AndrewOdlyzkoof Bell
foundan algorithm
Laboratories iT(x) in timeO(x5 +): a true"formula"
thatcomputes forv7i
x).

References
1. N. G. de Bruijn,P6lya'stheoryof counting, Chapter5 in AppliedCombinatorial Mathematics,E. F.
Beckenbach,editor,
Wiley,NewYork,1964.
2. G. P6lya,KombinatorischeAnzahlbestimmun,gen fuirGruppen,
Grapheniundchemische Verbindungen,Acta
Math.,68 (1937)145-254.
3. MichaelR. Gareyand David S. Johnson, Computers and Intractability,
a Guide to the Theoryof
NP-Completeness,Freeman, San Francisco, 1979.
4. PaulPritchard,A sublinQaradditivesieveforfinding Comm.ACM,24 (1) 1981,pp. 18-23.
primenumbers,
5. G. H. HardyandE. M. Wright, An Introduction ofNumbers,
to theTheory 3rded.,Oxford, theClarendon
Press,1954.
6. JohnD. DixonandHerbert S. Wilf,On therandomselection graphs,
ofunlabelled in preparation.

A MATHEMATICAL ANALYSIS OF
TIE GAME OF JAI ALAI
LOUISE E. MOSER
ofMathematics,
Department Hayward,CA 94542
CaliforniaState University,

JaiAlai is a ball gamewhichevolvedduringtheseventeenth

1. Introduction. century in the
Basque provincesof Spain.It is playedin a courtwiththreewallsby eightplayers(singles)or
eightteamsof two players(doubles).The ball, called the pelota,is thrownagainstthe front
playingwallwitha curvedreedbasket,calledthecesta.It can bounceoffthesidewall,theback

LouiseMoserreceived herPh.D. fromtheUniversityof Wisconsin and

in 1970.Her mainarea of research
has been3-dimensional
publication topologyand knottheory. a professor
She is currently State
at California
Hayward,
University, andComputer
wheresheteachescoursesin Mathematics Science.

This content downloaded from 128.235.251.160 on Wed, 21 Jan 2015 04:38:41 AM

All use subject to JSTOR Terms and Conditions