2
$\begingroup$

The well known identity $\sum_{j=0}^n (-1)^j {n \choose j} j^k =0$ when $k\lt n$ can be viewed as a particular case, applied to a uniform distribution, of a more general relation, defined for a general distribution $\{p_i\}, 1\le i\le n, \sum_{i=1}^np_i=1$, for $k \lt n$: $$\sum_{j=0}^n (-1)^j \sum_{|J|=j} P_J^k = 0$$ where $\sum_{|J|=j}$ is the sum over all the subsets of size j of $\{1..n\}$ , and $P_J=\sum_{j\in J}p_j$. This more general relation can be proven by proving an intermediate result, then the general relation follows. The intermediate result is that, for any distribution, $\sum_{|J|=j} P_J^k$ can be expressed as a weighted sum of binomial coefficients $\sum_{u=1}^k {n-k \choose j-u} \alpha_{k,u}$, where $\alpha_{k,u}$ are independent of j and $\alpha_{k,k}=1$. For example, one can show quite easily that: $\sum_{|J|=j} P_J = {n-1 \choose j-1}$ , i.e., the same as a uniform distribution, and that: $\sum_{|J|=j} P_J^2 = {n-2 \choose j-2} + {n-2 \choose j-1} \sum_{i=1}^n p_i^2 $. However the rest of the proof based on induction on the exponent is quite tedious, several pages long. Is there another proof of this relation ?

If an analytic form exists in the general case, when applied to a uniform probability, the result should be such that: $\alpha_{k,u}n^k=\sum_{i=1}^nS(k,i)i!{n \choose i}{k-i \choose u-i}$ where S is the Stirling number of 2nd order. This can be proven with the well-known relation of Stirling numbers $j^k=\sum_{i=0}^jS(k,i)i!{j \choose i}$ and Chu-Vandermonde identity.

Proof by induction can be found in: https://www.researchgate.net/publication/314243075_On_Von_Schelling_Formula_for_the_Generalized_Coupon_Collector_Problem

  • 0
    Hint: Let $\left[q\right]$ be the set $\left\{1,2,\ldots,q\right\}$ for each $q \in \mathbb{N}$. Then, $P_J^k = \sum\limits_{f : \left[k\right] \to \left[n\right]; \ f\left(\left[k\right]\right) \subseteq J} \prod\limits_{i\in\left[k\right]} p_{f\left(i\right)}$. Now you need to check that for each $f : \left[k\right] \to \left[n\right]$, the terms containing the corresponding $\prod\limits_{i\in\left[k\right]} p_{f\left(i\right)}$ cancel. To prove this, observe that $f\left(\left[k\right]\right)$ is a proper subset of $\left[n\right]$, since its size is $\leq k < n$.2017-02-09
  • 0
    can you detail your hint for k=1 or k=2 for example ? Thanks2017-02-13
  • 0
    I've turned my hint into a proof that's now on my website. See my answer for details.2017-02-15
  • 0
    Trivial remark: Your sum $\sum_{u=1}^k$ should start at $u=0$, to adapt to the case when $k=0$.2017-02-16
  • 0
    This is correct. However it implies handling additional weigths: alpha(k,0)=0 for k>0. Also in my application case, k is necessarily positive.2017-02-17

3 Answers 3

1

The following calculation is likely to be based on the wrong interpretation of the question since the OP informs us that this is a difficult combinatorial identity.

The sum in question is given by

$$k! [z^k] \prod_{q=1}^n (1-\exp(p_q z)).$$

Now observe that $1-\exp(p_q z)$ starts at $z$ namely with $-p_q z - p_q^2 z^2/2-\cdots$ so the product starts at $[z^n]$ and hence the coefficients on $[z^k]$ where $k\lt n$ are zero, which is the claim.

  • 0
    Agreed this is a much more straigthforward way to prove the generalization of the identity. I was more interested in an alternative version of proof of what I call intermediate result. I would really appreciate a reference to the OP.2017-02-07
0

It appears that what the OP is seeking is the following formula, which follows by inspection, namely that when $j\ge k$ we have

$$\sum_{|J|=j} P_J^k = \sum_{q=1}^k \sum_{|J|=j} \sum_{Q\subseteq J, \; |Q|=q} k! [z^k] \prod_{r\in Q} (\exp(rz)-1) \\ = k! [z^k] \sum_{q=1}^k \sum_{|J|=j} \sum_{Q\subseteq J, \; |Q|=q} \prod_{r\in Q} (\exp(rz)-1) \\ = k! [z^k] \sum_{q=1}^k \sum_{|J|=j} [u^q] \prod_{r\in J} (1+ u(\exp(rz)-1)) \\ = k! [z^k] \sum_{q=1}^k [v^j] [u^q] \prod_{r\in P} (1+ v(1+ u(\exp(rz)-1))) \\ = k! [z^k] \sum_{q=1}^k [v^j] [u^q] \prod_{r\in P} (1+ v + uv(\exp(rz)-1)) \\ = k! [z^k] \sum_{q=1}^k [v^j] [u^q] \prod_{r\in P} (1+ v -uv + uv\exp(rz)).$$

This is

$$k! [z^k] \sum_{q=1}^k [v^j] [u^q] (1+v-uv)^n \prod_{r\in P} \left(1 + \frac{uv}{1+v-uv} \exp(rz)\right) \\ = k! [z^k] \sum_{q=1}^k [v^j] [u^q] (1+v-uv)^n \sum_{Q\subseteq P} \left( \frac{uv}{1+v-uv}\right)^{|Q|} \prod_{q\in Q} \exp(qz) \\ = \sum_{q=1}^k [v^j] [u^q] (1+v-uv)^n \sum_{Q\subseteq P} \left( \frac{uv}{1+v-uv}\right)^{|Q|} \left(\sum_{l\in Q} l\right)^k.$$

This becomes

$$\sum_{q=1}^k [v^j] [u^q] \sum_{Q\subseteq P} (uv)^{|Q|} (1+v-uv)^{n-|Q|} \left(\sum_{l\in Q} l\right)^k \\ = \sum_{Q\subseteq P, \; |Q|\le j} \left(\sum_{l\in Q} l\right)^k [v^{j-|Q|}] \sum_{q=|Q|}^k [u^{q-|Q|}] (1+v-uv)^{n-|Q|} \\ = \sum_{Q\subseteq P, \; |Q|\le j} \left(\sum_{l\in Q} l\right)^k [v^{j-|Q|}] \sum_{q=|Q|}^k {n-|Q|\choose q-|Q|} (-1)^{q-|Q|} v^{q-|Q|} (1+v)^{n-q} \\ = \sum_{Q\subseteq P, \; |Q|\le j} \left(\sum_{l\in Q} l\right)^k \sum_{q=|Q|}^k {n-|Q|\choose q-|Q|} (-1)^{q-|Q|} [v^{j-q}] (1+v)^{n-q} \\ = \sum_{Q\subseteq P, \; |Q|\le j} \left(\sum_{l\in Q} l\right)^k \sum_{q=|Q|}^k {n-|Q|\choose q-|Q|} (-1)^{q-|Q|} {n-q\choose j-q}.$$

Note that the subsets with $j\ge |Q|\gt k$ do not contribute because the range of the second inner sum is up to $k.$ We re-write one more time to isolate the dependency on $j$ and get

$$\bbox[5px,border:2px solid #00A000]{ \sum_{q=1}^k {n-q\choose j-q} \sum_{Q\subseteq P, \; 1\le |Q|\le q} {n-|Q|\choose q-|Q|} (-1)^{q-|Q|} \left(\sum_{l\in Q} l\right)^k.}$$

Here we have started $q$ at one because the empty set does not contribute, with the power sum being zero. It thus appears that the queried constants are given by

$$\alpha_{k,q} = \sum_{Q\subseteq P, \; 1\le |Q|\le q} {n-|Q|\choose q-|Q|} (-1)^{q-|Q|} \left(\sum_{l\in Q} l\right)^k.$$

  • 0
    Hi, interesting, very interesting.2017-02-09
  • 0
    Hi, interesting, very i.. You end up with binomial coeeficients in n-q whereas my constants are in n-k, but important thing is the indendence from j. I have not yet checked that your derivation correctly implies the second condition (constant is 1 when u=k). If I understand well, your j >=k limitation comes from line 13 (numbering your proof) , where exponent of v must be positive or null, am I correct ? My humble suggestion is notation P line 4 be explicit as the reference set, i.e. |P|=n, you need to go to 7 to be sure of it. I'm still struggling to be sure of the 8->9, (reference ?).BR2017-02-09
  • 0
    8->9, I got it.OK2017-02-09
  • 0
    My sunday homework: it's fairly easy to derive (my) alpha coefficients from your expression and indeed, they give the right result for k=1 and k=2 and for the uniform case (with stirling numbers). This seems to indicate that your expression is correct even for j2017-02-13
  • 0
    just figured out that from your decompostion, it is possible to derive an expression giving the sum of the k-power of all j-subsets probabilities as a sum of k-power of all 1..k-subsets probabilities (for j>=k). If n=100, j=50 and k=5, this is 1.27*10^21 less operations to perform. Nice!2017-02-21
  • 0
    I meant: this is 1.27*10^21 times less operations2017-02-23
  • 0
    I kindly ask that you do not contact me personally per email as I cannot and will not answer these. Please post to MSE if you have additional questions. Thank you for your understanding.2017-02-23
  • 0
    Hello,I’m completing a paper/report related to a question I submitted on MSE and that you answered.I would appreciate if you have comments on the part I copy below (rest of paper is not yet completed) and if you agree/disagree with the wordings. Many Thanks and best regards.2017-02-24
  • 0
    During the finalization phase of this document, a major contribution was made by Marko Riedel ) who gave an analytic form of decomposition when j≥k (in fact j>k, the analytic form lending to a tautology for j=k).2017-02-24
0

This is Exercise 6.50 (b) in my Notes on the combinatorial fundamentals of algebra, version of 10 January 2019. (The numbering of the exercises might change, but you can find the version of 10 January 2019 archived on github.)

Note that the proof I give there is really written up "from scratch" (and I spend a while ruminating on trivialities about sums of products because my ring is noncommutative; not sure if you need that generality). You probably will find Lemma 7.222 trivial, and maybe Lemma 7.223 known. Parts (c) and (d) of that exercise give a formula for the sum in the case when $k=n$ (rather than $k). Part (a) gives a formula in the general case, but your mileage may vary about how useful it is.

  • 0
    i agree 18(b) is a proof of the generalisation of the combinatorial identity, however as I said, I'm interested in what I called the intermediate result (decomposition of sum of powers of j-subset probabilites on a basis of binomial coefficients) and computation of the2017-02-16
  • 0
    .. computation of the coefficients. Regards,2017-02-16
  • 0
    Ah! That's indeed a more interesting question. My computations show that $\alpha_{k,u} = \sum\limits_{a \in \operatorname{Par}^{\leq n}} \dbinom{k}{a} \dbinom{k - \ell\left(a\right)}{u - \ell\left(a\right)} m_a$, where I use the following notations: $\mathbb{N} = \left\{0,1,2,\ldots\right\}$. The notation $\dbinom{k}{a}$ is a multinomial coefficient, defined to be $0$ if the sum of the entries of $a$ is not $k$. The number $\ell\left(a\right)$ is the length of a partition $a$. The set $\operatorname{Par}^{\leq n}$ is the set of all partitions of length $\leq n$. The ...2017-02-16
  • 0
    ... polynomial $m_a$ is the $a$-th monomial symmetric polynomial in the $p_1,p_2,\ldots,p_n$; in other words, it is the sum of all monomials whose exponents are a reordering of the entries of $a$. But I'm not sure how useful this is. One consequence (using some symmetric function theory) is that the $\alpha_{k,u}$ are $\mathbb{Z}$-polynomials in the power sums $p_1^q + p_2^q + \cdots + p_n^q$. But I suspect there is a simpler way to prove this.2017-02-16