The Number of Partitions of a Set

Author(s): Gian-Carlo Rota

Reviewed work(s):
Source: The American Mathematical Monthly, Vol. 71, No. 5 (May, 1964), pp. 498-504
Published by: Mathematical Association of America
Stable URL: http://www.jstor.org/stable/2312585 .
Accessed: 08/01/2013 00:33

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 on Tue, 8 Jan 2013 00:33:31 AM

All use subject to JSTOR Terms and Conditions
498 THE NUMBER OF PARTITIONS OF A SET [May

leads to
=rn 1
n= ? 1
Anon-1 s g fom

+
A?,n-2 =3 x

Ann3=15( +
2)+(n

An,n--6=105(
+ )+15 f+ )+ 5(

but thereseemsto be no simplegeneralform.

References
1. E. T. Bell,GeneralizedStirlingtransforms ofsequences,Amer.J.Math.,62 (1940) 717-724.
2. L. Carlitz,On theproductoftwoLaguerrepolynomials, J. LondonMath. Soc., 36 (1961)
399-402.
3. of the reciprocalof Jo(x),Arch.Math.,6 (1955) 121-127.
, The coefficients
4. H. W. Gould,A newconvolution formula and someneworthogonal relationsforinversion
ofseries,Duke Math.J.,29 (1962) 393-404.
5. , A seriestransformation for findingconvolutionidentities,Duke Math. J., 28
(1961) 193-202.
6. I. Lah, Eine neueArtvon Zahlen,ihreEigenschaften und Anwendung in der mathema-
tischenStatistik,Mitteilungsbl. Math. Statist.,7 (1955) 203-212.
7. AlfredRenyi,Some remarkson the theoryof trees,Publ. Math. Inst. Hung.Acad. Sci.,
4 (1959) 73-85.
8. JohnRiordan,An introduction to combinatorial analysis,Wiley,New York,1958.
9. , Enumeration of lineargraphsformappingsof finitesets,Ann.Math. Statist.,33
(1962) 178-185.
10. HerbertJohnRyser,Combinatorialmathematics, Carus Math. MonographNo. 14,
New York,1963.

THE NUMBER OF PARTITIONS OF A SET

GIAN-CARLO ROTA, MassachusettsInstituteofTechnology
Let S be a finitenonemptyset with n elements.A partitionof S is a familyof
disjoint subsets of S called "blocks" whose union is S. The number Bn of dis-
tinct partitionsof S has been the object of several arithmeticaland combina-
torialinvestigations.The earliestoccurrencein printof these numbershas never
been traced; as expected, the numbers have been attributedto Euler, but an
explicit referenceto Euler has not been given, and Bell [7] doubts that it can

This content downloaded on Tue, 8 Jan 2013 00:33:31 AM

All use subject to JSTOR Terms and Conditions
1964] THE NUMBER OF PARTITIONS OF A SET 499

be foundin Euler's work.The propertiesof these numbersare periodicallybeing

rediscovered,as recentlyas 1962 (cf. [13]). Following Eric Temple Bell, we
shall call them the exponentialnumbers.Bell [4, 5, 6, 7], used the notation e(;
on the otherhand, Jacques Touchard [29 and 30] used a. to celebrate the birth
of his daughterAnn; Becker and Riordan [3] used B. in honorof Bell. We shall
followtheirchoice.
A great many problemsof enumerationcan be interpretedas countingthe
numberof partitionsof a finiteset; forexample, the numberof rhymeschemes
forn verses, the numberof ways of distributingn distinctthingsinto n boxes
(emptyboxes permitted),the numberof equivalence relationsamong n elements
(cf. [8]), the numberof decompositionsof an integerinto coprimefactorswhen
n distinctprimesare concerned (cf. Bell [7]), the numberof permutationsof n
elementswith orderedcycles (cf. Riordan [27], page 77 ff.),the numberof Borel
fieldsover a set of n elements(cf. Binet and Szekeres [8]), etc., etc. Exponential
numbers occur frequentlyin probability,and their theoryis closely related to
that of the Poisson-Charlierpolynomials (see below).
Several explicitexpressionsforthe exponentialnumbersare known,and can
be foundin [2, 3, 5, 6, 10, 13, 14, 15, 16, 22, 25, 29, 30, 32]. One of the simplest
ways of describingthe sequence Bn is by its exponentialgeneratingfunction
(1) co Bn -
-xn = eeX-,
n-0 n!I
where we have set Bo= 1 by convention. All known explicitformulas,however,
except the one we shall derive, rely to a greater or lesser degree upon direct
enumerationsleading to nonimmediaterecursionsforthe sequence Bn.
In this note we shall give a new formulafor the exponential numbers (for-
mula (4) below) which differsfromthe previous ones in that it relies least upon
directcountingarguments,and whichhingesinstead upon some elementarycon-
siderationsof a "functorial"nature. It is the author's convictionthat formula
(4), which we derive below, is the natural descriptionof the exponential num-
bers. The basic idea is a general one, and can be applied to a variety of other
combinatorialinvestigations.We shall see that it easily leads to quick deriva-
tions of the propertiesof the Bn.
Consideran auxiliary finiteset U having u elements,u>0. We shall examine
the structureof the set US offunctionswith domain S, a set with n elements,
and range a subset of U. The basic fact is that there are un distinctsuch func-
tions, as is evidenced by the most elementaryof countingarguments.We shall
now examine this set of functionsin greaterdetail.
To every functionf: S-> U thereis naturallyassociated a partition7rof the
set S, called the kerneloff, definedas follows.Two elementsa and b of S are
to belong to the same block of 7r,if and only iff(a) =f(b).
How many distinctfunctionsare therewith a given kernelw? This question
is easily answered. Indeed, let N(wx)denote the numberof distinctblocks of the
partition 7r.A functionhaving kernel w must take distinct values on distinct

This content downloaded on Tue, 8 Jan 2013 00:33:31 AM

All use subject to JSTOR Terms and Conditions
500 THE NUMBER OF PARTITIONS OF A SET [May

blocks of 7r.Thus, such a functiontakes altogetherN(ir) distinctvalues, and the

number of distinct such functionsequals the number of one-to-onefunctions
froma set of N(r) elementsto the set U. Again, it is well known that such a
numberis u(u -1) (u - N(r) + 1) =(U)(, and this expressionis called the
factorialpowerof the numberu, with exponent N(wx).
Now, every functionhas a unique kernel. Thereforewe have the following
identity,valid forall integersu>O:
(2) X(U)N (X) = Un,

wherethe sum on the leftrangesover all partitionsof7rthe set S.

We now come to the main idea. Let V be the vector space over the reals
consisting of all polynomials in the single variable u. Any sequence of poly-
nomials of degrees0, 1, 2, * * *, is a basis for this vectorspace, in particular,the
sequence (u)o =1, (u)1, (U)2, (u)3, * - . Since a linear functional L on V is
uniquely determinedby assigning the values it takes on an arbitrarybasis,
thereexists a unique linear functionalL on V such that
L(1) = 1, L((U)) = 1, k = 1, 2> 3,
ApplyingL to both sides of (2) we obtain
(3) E L((u)N(r)) = L(un);

but, by the definitionof L, the leftside simplifiesto a sum of as many ones as

there are partitionsof the set S. In other words, (3) simplifiesto
(4) Bn = L(un).
This formulais the explicit expressionfor the exponential numbers which
we wanted to establish. Let us see now how it can be used to derive the other
propertiesof the exponentialnumbers.
We begin by derivingthe recursionformulaforthe numbersBn,

(5) Bn+l = )Bk.

k==O k
Now, since u(u- X= (u)X+l, we have L(u(u- 1)n)= 1 =L((u)n). Since the
polynomials 1, (u)n for n = 2, 3, * * * form a basis for the vector space V, it
followsfromthe linearityof L, that
(6) L(up(u - 1)) = L(p (u))
forevery polynomialp. In particular,forp(u) = (u+1)n we obtain
L(un+l) = L((u + 1)n),
but this is preciselyformula(5), as we wanted to show.
Note that identity(6) forall polynomialsp togetherwith the initial condi-

This content downloaded on Tue, 8 Jan 2013 00:33:31 AM

All use subject to JSTOR Terms and Conditions
1964] THE NUMBER OF PARTITIONS OF A SET 501

tion L(1) = 1 completelycharacterizesthe linear functionalL, as definedby (4),

since the argumentby whichwe have established (6) is reversible.We shall now
use this factin establishingthe generatingfunction(1) forthe exponentialnum-
bers. To this end, let gn/n!be the nth coefficient
in the Taylor series expansion
of eex1:
E -Xn = ee-1.
nO0n!
There exists a unique linear functionalM on V such that M(un) =gn, and
it will sufficeto prove that L = M, to conclude that gn= Bn. Now,
eex-1 =

whereM(ezu)is defined
as
?? M(un)
n=O n!

Differentiating,
we get

(7) exee-l = M( - ex) =H(exu)

whence M(e&(u+l)) = M(uexu). Expanding the functionsez(u+l)and ezu into Taylor

series in the variable x and comparingterms,we obtain M((u+1)n) = M(un+l).
But,sincethe polynomials un forma basis forV, thisimpliesat onceprop-
erty(6). Hence M=L.
Note thatdifferentiating
underM, as we have donein (7), does notrequire
any continuity ofthefunctional
properties M: it is "purelyformal."
There is another,moreamusingderivationof the generatingfunctiondi-
rectlyfrom(4), whichgoes as follows:

E _Xn E ( )xn_L(eux).
L
nvo n nO n!
Now,set ell I +v, and expand(1 +v) u by the binomialtheorem:
CIO
B(
.,
_Xn = L((l + v)u) = Lt Eu)0 n =
(1 8Vn) L(u
L_____n_ vn Av
11s I! n-0 n! n=o n
= e= ee1-i q.e.d.
In this derivation,it may at firstseem puzzling(as suggestedby R. D.
Schafer)that infinitesums have been commutedwithL, withoutdiscussing
any continuity properties ofL. The puzzleis solvedas soonas it is noticedthat
all appearancesof the symbolL can be completely eliminated, and the whole
derivationamountsto the proofof an infinite sequenceof identitiesrelating
thecoefficientsoftwoTaylorseries.The use ofL is just a speedyway ofestab-
lishingtheseidentities.

This content downloaded on Tue, 8 Jan 2013 00:33:31 AM

All use subject to JSTOR Terms and Conditions
502 THE NUMBER OF PARTITIONS OF A SET [May

Next, we shall establish the remarkableformulaof Dobinski [14]:

1 / 2n 3n 4n
(8) Bn+1 ln+ _+ 2+ - +

We begin by noticing that the exponential series e= ?-O 1/k! can be

triviallyrewrittenas e = k=o (k)n/k!,wheren is any nonnegativeinteger.In view
of the definitionof the linear functionalL, this gives
I1 (k)n
L((n)n) =-E - k!
e k=O

Using again the fact that the polynomials(u)n forma basis forthe vector space
V, and that the functionalL is linear,we inferat once that
1 p(k)
(9) L(p(,)) = -
e k=O k!
forany polynomialp. Dobinski's formulanow followsby settingp(u) = un.
Dobinski's formulais particularlysuited to the computationof Bn forlarge
n, by an application of the Euler-Maclaurin summation formula (cf. [16] and
[25]).
Identity (9) establishes an importantpropertyof the linear functionalL,
namely,that it is positivedefiniteon the half-line[0, oo). We can thereforede-
fine a sequence of orthogonalpolynomialsrelative to L, and the propertiesof
classical systemsof orthogonalpolynomials (cf. Szego [28]) will apply to this
set. Such a set of polynomials,we shall now prove, is
n /n\
(10) h1i(U) = E(-l)* ( n-k,

wherewe use Touchard's notation hnfrom [30].

We firstnote that (6) can be rewrittenin more enlighteningformby using
operatornotation.Let E: p(u) ->p(u + 1) be the shiftoperator,let D: p(u) ->p'(u)
be the derivative, and let V: f(x) -f(l) be the linear functionalconsistingin
evaluating a functionat x =1. Then (6) can be rewrittenforany integerk > 0,
by iteration,as
L(Ekp(u)) = L(p(u)VDkXu)
wherewe have used the fact that (u)k = VDkxu.It followsfromlinearityif g is
any polynomial,that
L(g(E)p(u)) L(p(u)Vg(D)xu).
Now set g(x) = (1- x)i, giving g(E) = (-1) ii, the iterated differenceoperator.
For this choice of g, we have evidently Vg(D)xu = (-1)ihj(x). Set p(u) = h7l(u),
and obtain

This content downloaded on Tue, 8 Jan 2013 00:33:31 AM

All use subject to JSTOR Terms and Conditions
1964] THE NUMBER OF PARTITIONS OF A SET 503

(-1) i'+L(Ahn(u)) = L(hn (u)hj(u)).

If j> n, A'h. vanishes identically,provingthe orthogonalityof the polynomials,

and ifj n, we get L(h.(u) 2) = n!, which gives the normalizingfactors.
The polynomialshnare the special case of the Poisson-Charlierpolynomials
(cf. Szego [28], p. 34) obtained by setting a= 1, in Szego's notation. As re-
marked by Touchard [30], they are particularlyuseful forcomputation of the
exponentialnumbersby recursion.Formulas forthe firstseven polynomialsare
given by Touchard [30].
These examples sufficeto give an idea of the use of formula(4), and to sup-
portthe contentionthat thisformulagives the natural definitionof the exponen-
tial numbers. Formula (4) has been suggested by the Blissard calculus tech-
niques so usefulin enumerativeanalysis, (cf. Riordan [27], Ch. 2 Section 4); by
the systematic use of linear functionalswe can give a rigorousfoundationto
this calculus, as well as extendits uses in some directions.We hope to implement
these contentionsin a futurepublication.
This workwas begunundercontractNSF-GP-149,continuedundercontractwiththeOffice
of Naval Research,and concludedwhilethe authorwas a Fellowof the Sloan Foundation.The
authorwishesto thanktherefereeforseveralimprovements in thetext,and forvaluablehistorical
references.
The followingbibliographycontainsall publicationsknownto the authorwhichstudythe
exponentialnumbers.He willgreatlyappreciateanysuggestions ofomittedworks.
References
1. C. A. Aitken,EdinburghMath. Notes,28 (1933) 18-33.
2. F. Anderegg, Problem129,thisMONTHLY, 8 (1901) 54.
3. H. W. Beckerand JohnRiordan,The arithmetic of Bell and Stirlingnumbers, Amer.J.
Math. 70 (1934) 385-394.
4. Eric TempleBell,Exponentialpolynomials, Ann.ofMath.,35 (1934) 258-277.
5. , Exponentialnumbers, Trans.Amer.Math. Soc., 41 (1934) 411-419.
6. , Exponentialnumbers, thisMONTHLY, 41 (1934) 411-419.
7. , The iteratedexponential Anii.of Math.,39 (1938) 539-557.
integers,
8. E. F. Binetand G. Szekeres,On Borelfieldsoverfinitesets,Ann.Math. Stat., 29 (1957)
494-498.
9. GarrettBirkhoff, Latticetheory,Amer.Math. Soc., 1948,rev.ed.
10. Ugo Broggi,Rendiconti dell' IstitutoLombardodi Scienzee Lettere,2ndseries,66 (1933)
196-202.
11. E. Catalan,Note surune equationaux differences finies,J. Math, PuresAppl.,3 (1838)
508-516.
12. ErnestoCesaro,NouvellesAnnalesdes Mathematiques, 4 (1885) 39.
13. MartinCohn,ShimanEven, Karl Menger,Jr.,and PhilipK. Hooper,On thenumberof
partitioningsofa set ofn distinctobjects,thisMONTHLY, 69 (1962) 782-785.
14. G. Dobinski,Grunert's Archiv,61 (1877) 333-336.
15. Mauriced'Ocagne,Sur une classe de nombresremarquables, Amer.J. Math., 9 (1886)
353-380.
16. Leo F. Epstein,A function relatedto theserieseo',J. Math. Phys.,18 (1939) 153-173.
17. P. Epstein,ArchivderMathematikund Physik,8 (1904-05) 329-330.
18. I. M. H. Etherington, Nonassociative powersand a functional equation,Math. Gaz., 21
(1937) 36-39.

This content downloaded on Tue, 8 Jan 2013 00:33:31 AM

All use subject to JSTOR Terms and Conditions
504 ON THE SPANS OF DERIVATIVES OF POLYNOMIALS [May
19. Glover,Tables of appliedmathematics, Ann Arbor,1923.
20. A. Krug,ArchivderMathematikundPhysik,9 (1905) 189-191.
21. E. Lucas, Theoriedes nombres, vol. 1, GauthierVillars,Paris,1891.
22. N. S. Mendelsohn, Problem4340,thisMONTHLY, 56 (1949) 187.
23. , Applicationsof combinatorial formulaeto generalizations
of Wilson'stheorem,
Canad. J. Math.,1(1947) 328-336.
24. SilvioMinetola,Principiidi AnalisiCombinatoria, Giornaledi Matematiche,45 (1907)
333-366.
25. Leo Moserand Max Wyman,An asymptotic formulaforthe Bell numbers, Trans.Roy.
Soc. Canada, Sect. III, 49 (1955) 49-54.
26. E. Netto,Lehrbuchder Kombinatorik, Teubner,Leipzig,1901.
27. JohnRiordan,An introduction to combinatorialanalysis,Wiley,New York,1958.
28. GaborSzego,Orthogonalpolynomials, Revisededition,Amer.Math. Soc., 1959.
29. JacquesTouchard,Proprietes arithmetiques de certainsnombresr6currents, Ann. Soc.
Sci., Bruxelles,
A 53 (1933) 21-31.
30. , Nombresexponentiels et nombresde Bernoulli,Canad. J. Math.,8 (1956) 305-
320.
31. WilliamAllenWhitworth, Choiceand chance,Cambridge,Deighton,Bell and Co., 1901
(reprinted by Stechert).
32. G. T. Williams,Numbersgeneratedby thefunction eal, thisMONTHLY, 52 (1945) 323-
327.
Additionalreferences are to be foundin the master'sthesisof F. Finlayson,University of
Alberta,1955.

ON THE SPANS OF DERIVATIVES OF POLYNOMIALS

of California,
RAPHAEL M. ROBINSON, University Berkeley
1. Introduction.By the span of a polynomialall of whose roots are real, we
shall mean the differencebetween its largest and smallest roots. We are inter-
ested in the followingproblem: If thespan of a polynomialf(x) of degreen with
real rootsis given,howis thespan ofits k-thderivativemaximized?It will be suffi-
cient to considerpolynomialsf(x)= Xn+? * all of whose roots lie in the inter-
val -1< x< 1. Then all of the roots off(k)(x) lie in the same interval,and we
try to maximize the difference between its largest and smallest roots.
We shall suppose throughoutthat k _ n -2, so that f(k) (x) will have more
than one root. On the otherhand, we see that the problemis trivialifn>2k+2.
For in this case, we may put k+ 1 roots off(x) at each end point x = + 1. Then
f(k)(x) will have a root at each end point, and will thereforehave span 2. If
n>2k+2, some of the roots of f(x) are arbitrary,whereas for n=2k+2 the
span off(k)(x) is maximized only forf(x) = (x- 1)k+1(x+1)k"+.
Thus the nontrivialcases of the problem are those with k+2? n ?2k + 1.
We shall show in Section 2 that in these cases thespan off(k)(x) can be maximized
onlywhenall of therootsoff(x) are at theend pointsx = + 1.

This content downloaded on Tue, 8 Jan 2013 00:33:31 AM

All use subject to JSTOR Terms and Conditions