6 1 0 2 n a J Identities for partial Bell polynomials de- 6 rived from identities for weighted integer com- ] O positions C . h t Steffen Eger a m [ Abstract. Wediscussclosed-form formulas forthe(n,k)-thpartialBell 1 polynomialsderivedinCvijovi´c[2].WeshowthatpartialBellpolynomi- v als arespecial cases of weighted integer compositions, anddemonstrate 7 how the identities for partial Bell polynomials easily follow from more 4 general identities for weighted integer compositions. We also provide 3 short and elegant probabilistic proofs of the latter, in terms of sums 1 of discrete integer-valued random variables. Finally, we outline further 0 identities for the partial Bell polynomials. . 1 Mathematics Subject Classification (2010). Primary 11P81; Secondary 0 60C05. 6 1 Keywords.Integercomposition,PartialBellPolynomial,Sumofdiscrete : random variables. v i X r a 1. Introduction In a recent note, Cvijovi´c [2] has derived three new identities — which we list in Equations (5), (6), and (7) below — for the (n,k)-th partial Bell polynomials in the variables x ,x ,...,x , which are defined as1 1 2 n−k+1 B (x ,x ,...,x )= n,k 1 2 n−k+1 n! x1 ℓ1··· xn−k+1 ℓn−k+1, (1) Xℓ1!···ℓn−k+1!(cid:16)1!(cid:17) (cid:18)(n−k+1)!(cid:19) The final publication is available at Springer via http://dx.doi.org/10.1007/s00010-015-0338-2 1AnotherwaytodefinethepartialBellpolynomialsisbytheformalpowerseriesexpansion 1 xmtm k = Bn,k(x1,x2,...,xn−k+1)tn. This wouldalsoimmediatelylead k!(cid:16)Pm≥1 m! (cid:17) Pn≥k n! toCorollary1belowifwedefinedthesubsequentconceptofweightedintegercompositions k by the formal power series expansion f(m)tm = k tn, with notation (cid:16)Pm≥0 (cid:17) Pn≥0(cid:0)n(cid:1)f asexplicated below. 2 Eger wherethesummationisoverallsolutionsinnonnegativeintegersℓ ,...,ℓ 1 n−k+1 of ℓ +2ℓ +···+(n−k+1)ℓ = n and ℓ +ℓ +···+ℓ = k. 1 2 n−k+1 1 2 n−k+1 Such identities may be important for the efficient evaluation of the partial Bellpolynomials,ashintedatinCvijovi´c[2],whoalsoindicates applications ofthepolynomials.Importantly,theygeneralizetheStirlingnumbersS(n,k) of the second kind, since S(n,k)=B (1,...,1). n,k The purpose of the present note is to show that these identities are special cases of identities for weighted integer compositions and to outline very short proofs for the more general identities, based on sums of discrete random variables. Our main point is to compile (identical or very similar) resultsdevelopedwithinheterogeneouscommunitiesofresearch,andtocom- binetheseresults.Finally,inSection4,wepresentthreeadditionalidentities for the partial Bell polynomials. Throughout, we write B as a shorthand n,k for B (x ,...,x ), unless explicit reference to the indeterminates is n,k 1 n−k+1 critical. 2. Preliminaries on weighted integer compositions An integer composition of a nonnegative integer n is a k-tuple (π ,...,π ), 1 k for k ≥ 1, of nonnegative integers such that π + ··· + π = n. We call 1 k k the number of parts. Note that order of part matters, and this distin- guishes integer compositions from integer partitions. Now, to generalize, we may consider f-colored integer compositions, where each possible part size s ∈ N = {0,1,2,...} may come in f(s) different colors, whereby f : N → N. For example, the integer n = 4 has nine distinct f-colored integer composi- tions, for f(0) = 2, f(1) = f(2) = 1, f(3) = f(4) = f(5) = ··· = 0, with k =3 parts, namely, (2,2,0),(2,2,0∗),(2,1,1),(1,2,1),(2,0,2),(2,0∗,2),(1,1,2),(0,2,2),(0∗,2,2), whereweuse astarsuperscripttodistinguishthe twodifferentcolorsofpart size 0. For weighted integer compositions we let, more generally, f(s) be an arbitrary real number (or even a value in a commutative ring), the weight of partsize s. Coloredinteger compositions have been discussed in [6, 9, 10], and weighted integer compositions have been under review in Eger [4], but have been investigated as early as Hoggatt and Lind’s [7] work. Let k denote the number of f-weighted integer compositions of n n f with k pa(cid:0)rts(cid:1), when the range of f is N, and let k denote the total weight n f of all f-weighted integer compositions of n with(cid:0)k(cid:1)parts, when the range of f is the set of reals R. We then have the following theorem. Identities for partial Bell polynomials 3 Theorem 1. Let k,n ≥ 0 be integers, and let f : N → R be arbitrary. Then the following identities hold. k = f(π )···f(π ), (2) (cid:18)n(cid:19) 1 k f X k k = f(0)ℓ0···f(n)ℓn, (3) (cid:18)n(cid:19) (cid:18)ℓ ,ℓ ,...,ℓ (cid:19) f X 0 1 n where the sum in (2) is over all solutions in nonnegative integers π ,...,π 1 k of π +···+π = n, and the sum in (3) is over all solutions in nonnegative 1 k integers ℓ ,...,ℓ of ℓ +···+ℓ =k and 0ℓ +1ℓ +···+nℓ =n. Finally, 0 n 0 n 0 1 n k = k! denote the multinomial coefficients. ℓ0,ℓ1,...,ℓn ℓ0!ℓ1!···ℓn! (cid:0) (cid:1) The proofofTheorem1 is simple. Identity (2) is a direct applicationof the definition of k and (3) follows, combinatorially, from (2) by rewriting n f the summationo(cid:0)ve(cid:1)rinteger compositionsinto asummationoverintegerpar- titions andthen adjustingeachterminthe sumappropriately(inparticular, the multinomial coefficients accountfor distributing the parts in partitions). Alternatively, when k is defined as coefficient of a certain polynomial, n f then (3) can also be(cid:0)ar(cid:1)rived at via application of the multinomial theorem (see [4]) orthe formulaofFaa` diBruno[8]forthe higherorderderivativesof composite functions (see [7]). Interpreting f(s), for s∈N, as indeterminates, Theorem 1 identity (3) immediatelyimpliesthatf-weightedintegercompositions‘generalize’partial Bell polynomials. Corollary 1. Let k,n≥0 be integers. Then: k! k B (x ,x ,...,x )= , (4) n! n,k 1 2 n−k+1 (cid:18)n(cid:19) f whereby f(0)=f(n−k+2)=f(n−k+3)=···=0, x f(s)= s, for s∈{1,2,...,n−k+1}. s! Corollary1 has been established in [1] in the more particular setting of restrictedinteger compositions,or,classical‘multinomial/extendedbinomial coefficients’, where, in our notation, f(s)=1 for all s∈{1,...,q}, for some positive integer q. This yielded the conclusion that partial Bell polynomials ‘generalize’restrictedintegercompositionsinsofarasB (1!,2!,...,q!,0,...)= n,k n! k , where f is the indicator function on {1,...,q}. k! n f (cid:0) (cid:1)Representation (2) in Theorem 1 is very useful on its own since it cap- tures an equivalence between f-weighted integer compositions and distri- butions of sums of independent and identically distributed (i.i.d.) discrete random variables as explicated in the following lemma. Lemma1. Letk,n≥0beintegersandletf :N→Rbearbitrary.Thenthere existi.i.d.nonnegativeinteger-valuedrandomvariablesX ,...,X suchthat 1 k 4 Eger k is given as the distribution of the sum of X ,...,X (times a suitable n f 1 k (cid:0)no(cid:1)rmalization factor). Conversely, the distribution P [X +···+X = n] of arbitrary i.i.d. f 1 k nonnegative integer-valued random variables X ,...,X with common dis- 1 k tribution function f is given by k . n f (cid:0) (cid:1) Proof. Consider k . When f is zero almost everywhere, i.e., f(s) = 0 for n f all s > x, for som(cid:0)e(cid:1)x ∈ N, then let F¯ denote the sum f(s′) and let s′∈N g(s) = f(s) for all s ∈ N. Consider the i.i.d. random vPariables X ,...,X F¯ 1 k with common distribution function g, i.e., P[X = s] = g(s). By definition, i the distribution of the sum X +···+X is given as 1 k P [X +···+X =n]= P[X =π ]···P[X =π ] g 1 k 1 1 k k π1+·X··+πk=n = g(π )···g(π ) 1 k π1+·X··+πk=n 1 1 k = f(π )···f(π )= . F¯n 1 k F¯n(cid:18)n(cid:19) π1+·X··+πk=n f When f is not zero almost everywhere, then note that k = k =F¯ˆnP [Y +···+Y =n], (cid:18)n(cid:19) (cid:18)n(cid:19) gˆ 1 k f fˆ whereby fˆ(s) = f(s) for s ≤ n and fˆ(s) = 0 for s > n, and Y ,...,Y are 1 k i.i.d.randomvariableswithcommondistributionfunctiongˆ(s)= fˆ(s),where F¯ˆ F¯ˆ = fˆ(s′). s′∈N PConversely, the distribution of the sum of i.i.d. nonnegative integer- valued random variables X ,...,X with common distribution function f is 1 k given by, as above, P [X +···+X =n]= P[X =π ]···P[X =π ] f 1 k 1 1 k k π1+·X··+πk=n k = f(π )···f(π )= , 1 k (cid:18)n(cid:19) π1+·X··+πk=n f applying Theorem 1 in the last equality. (cid:3) Lemma 1 has appearedin [4] and,in the specialcase when f is the dis- crete uniform measure, in [1]. Lemma 1 allows us to prove properties of the weightedintegercompositions k byreferringtopropertiesofthe distribu- n f tion ofthe sum of discrete rand(cid:0)om(cid:1) variables,whichis oftentimes convenient, as we shall see below. Identities for partial Bell polynomials 5 3. Identities and proofs The three identities for partial Bell polynomials that Cvijovi´c [2] introduces are the following: 1 1 n−k n n+1 B = (k+1)− x B , (5) n,k x n−k (cid:18)α(cid:19)(cid:20) α+1(cid:21) α+1 n−α,k 1 αX=1 n k !k ! n B = 1 2 B B , (6) n,k1+k2 (k +k )! (cid:18)α(cid:19) α,k1 n−α,k2 1 2 αX=0 1 n−1 αk−1−1 n α B = ··· ··· k−1 x ···x x . n,k+1 (k+1)! (cid:18)α (cid:19) (cid:18) α (cid:19) n−α1 αk−1−αk αk αX1=k αXk=1 1 k (7) As shown in Cvijovi´c [2], (7) easily follows inductively from (6). Therefore, we concentrate on the identities (5) and (6). Throughout, we let k,n be nonnegative integers and f be a function f : N → R. As we will see now, identity (6) may be seen as a special case of the convolution formula for the sum of discrete random variables. Lemma 2. Let n, k and k be nonnegative integers. Then: 1 2 k +k k k 1 2 = 1 2 . (cid:18) n (cid:19) (cid:18)x(cid:19) (cid:18)y(cid:19) f x+Xy=n f f Proof. By our previousdiscussion, it suffices to provethe lemma for sums of i.i.d. integer-valued random variables X ,...,X ,Y ,...,Y . Now, 1 k1 1 k2 P [(X +···+X )+(Y +···+Y )=n] f 1 k1 1 k2 = P [X +···+X =x]P [Y +···+Y =y] f 1 k1 f 1 k2 x+Xy=n by the discrete convolution formula for discrete random variables. Toformallycompletetheproof,2 weapplyLemma1(or,moreprecisely, its proof), leading to: k1+k2 = k1+k2 =F¯ˆnP [X +···+X =n] (cid:18) n (cid:19) (cid:18) n (cid:19) gˆ 1 k1+k2 f fˆ =F¯ˆn P [X +···+X =x]P [Y +···+Y =y] gˆ 1 k1 gˆ 1 k2 x+Xy=n = F¯ˆxP [X +···+X =x]F¯ˆyP [Y +···+Y =y] gˆ 1 k1 gˆ 1 k2 x+Xy=n k k k k = 1 2 = 1 2 . (cid:18)x(cid:19) (cid:18)y(cid:19) (cid:18)x(cid:19) (cid:18)y(cid:19) x+Xy=n fˆ fˆ x+Xy=n f f (cid:3) 2Weomitthisstraightforwardstepinallsubsequent proofs. 6 Eger Proof of identity (6). Using Lemma 2 and Corollary 1, we obtain that n! k +k n! k k B = 1 2 = 1 2 n,k1+k2 (k +k )!(cid:18) n (cid:19) (k +k )! (cid:18)x(cid:19) (cid:18)y(cid:19) 1 2 f 1 2 x+Xy=n f f n! k ! k ! = 1 B 2 B (k +k )! x! x,k1 y! y,k2 1 2 x+Xy=n k !k ! n! = 1 2 B B (k +k )! x!y! x,k1 y,k2 1 2 x+Xy=n k !k ! n = 1 2 B B . (k +k )! (cid:18)x(cid:19) x,k1 y,k2 1 2 x+Xy=n (cid:3) The proof of the following identity for weighted integer compositions which directly entails identity (5) is a straightforwardvariation of the proof outlinedinDePril[3],whoconsiderssumsofinteger-valuedrandomvariables. Lemma 3. Let k,n ≥ 0 be integers and let f : N → R be arbitrary with f(0)=0 and f(1)6=0. Then: k 1 n+1 k = k+1− (s+1)f(s+1) . (cid:18)n(cid:19) f(1)(n−k) (cid:18) s+1(cid:19) (cid:18)n−s(cid:19) f Xs≥1 f Proof. Let X ,...,X be i.i.d. nonnegative integer-valued random vari- 1 k+1 ables, with distribution function f. Consider the conditional expectation E[X |X +...+X = n+1]. Due to identical distribution of the vari- 1 1 k+1 ablesandlinearityofE[·|·],itfollowsthatE[X |X +...+X =n+1]= 1 1 k+1 E[X1+...+Xk+1|X1+...+Xk+1=n+1] = n+1.Thus,E[k+1X −1|X +...+X = k+1 k+1 n+1 1 1 k+1 n+1]=0, which means that n k+1 P [X =s,X +···+X =n+1] 0= ( s−1) f 1 1 k+1 n+1 P [X +...+X =n+1] Xs=1 f 1 k+1 n 1 k+1 = ( s−1)P[X =s]·P [X +···+X =n+1−s] k+1 n+1 1 f 2 k+1 n+1 f Xs=1 (cid:0) (cid:1) n 1 k+1 k = ( s−1)f(s) , k+1 n+1 (cid:18)n+1−s(cid:19) n+1 f Xs=1 f (cid:0) (cid:1) so that rewriting and shifting indices lead to the required expression. (cid:3) We omit the proof of identity (5) since it is a simple application of Lemma 3. Identities for partial Bell polynomials 7 4. More identities for the partial Bell polynomials We mention the following three additional identities for partial Bell polyno- mials, 1 n B = x B , (8) n,k k (cid:18)α(cid:19) α n−α,k−1 αX≥0 n−1 B = x B , (9) n,k (cid:18)α−1(cid:19) α n−α,k−1 αX≥1 n B (x ,...,x )= xαB (0,x ,...,x ). (10) n,k 1 n−k+1 (cid:18)α(cid:19) 1 n−α,k−α 2 n−k+1 αX≥0 (note that (8) is a special case of (6)) which straightforwardly follow from the following corresponding identities for weighted integer compositions, k k−1 = f(s) , (11) (cid:18)n(cid:19) (cid:18)n−s(cid:19) f Xs≥0 f k k k−1 = sf(s) , (12) (cid:18)n(cid:19) n (cid:18)n−s(cid:19) f Xs≥1 f k k k−i = f(r)i . (13) (cid:18)n(cid:19) (cid:18)i(cid:19)(cid:18)n−ri(cid:19) f Xi≥0 f˜ In Equation (13), f˜(r) = 0 and f˜(s) = f(s) for all s 6= r (note that (10) is a special case of (13) in which r = 1). Proofs of identities (11) to (13) can, e.g., be found in Fahssi [5] and Eger [4], who also give further identities for weighted integer compositions. References [1] Belbachir,H.,Bouroubi,S.,Khelladi,A.:Connectionbetweenordinarymultino- mials, Fibonacci numbers, Bell polynomials and discrete uniform distribution. Ann.Math. Inform. 35, 21–30 (2008) [2] Cvijovi´c, D.: New identities for the partial Bell polynomials. Applied Mathe- matical Letters 24, 1544–1547 (2011) [3] DePril, N.:Recursionsfor convolutionsof arithmeticdistributions. Astin Bull. 15, 135–139 (1985) [4] Eger,S.:Restrictedweighted integercompositions andextendedbinomial coef- ficients. J. Integer Seq.16 (2013) [5] Fahssi, N.-E.: The polynomial triangles revisited. Preprint available at http://arxiv.org/abs/1202.0228 (2012) [6] Guo, Y.-H.: Somen-color compositions. J. Integer Seq.15 (2012) [7] Hoggatt, V.E., Lind, D.A.: Compositions and Fibonacci Numbers. Fibonacci Quart. 7, 253–266 (1969) [8] Johnson, W.P.: The curious history of Fa`a di Bruno’s formula. Amer. Math. Monthly 109, 217–234 (2002) 8 Eger [9] Mansour, T., Shattuck, M.: A statistic on n-color compositions and related sequences. Proc. Indian Acad. Sci. (Math. Sci.) 124, 127–140 (2014) [10] ShapcottC.:C-colorcompositionsandpalindromes,FibonacciQuart.50,297– 303 (2012) Steffen Eger Robert Mayer Str. 10 P.O. Box 154 D-60325 Frankfurt am Main Germany e-mail: [email protected]