1

I'm trying to prove the Corollary$1$ in a good answer about the Generalized Inclusion-Exclusion Principle, which I'm studying these days: https://math.stackexchange.com/a/362516/390226, for easy reading I use some of the definitions there. (Apologies for not mention this when I posted!)


Let $\{S(i)\}_{i=1}^m$ be a collection of sets from a finite universe, and $A$ a subset of $\{1,2,\dots,m\}$, then $N(j)$ defined as \begin{align}N(j)=\sum_{|A|=j}\left|\bigcap_{i\in A} S(i)\right|\end{align}

What I want to get is the number of elements in at most $k$ of $S(i)$. My progress:
(Edit for note: The $\color{red}{\textrm{red}}$ part is what I did wrong!) \begin{align} \textrm{Atmost(k)}&=\sum_{l=0}^k\textrm{Exact(l)}\\ &=\sum_{l=0}^k\left[\sum_{j=l}^m(-1)^{j-l}\binom{j}{l}N(j)\right]\\ &=\sum_{j=0}^m(-1)^j\sum_{l=0}^{\color{red}{j+[j>k](-j-1)}}(-1)^l\binom{j}{l}N(j)\\ &=\sum_{j=0}^m(-1)^j\sum_{l=0}^{j}(-1)^l(-1)^{k-l}\binom{-1}{k-l}\binom{j}{l}N(j)\\ &=\sum_{j=0}^m(-1)^{j-k}\binom{j-1}{k}N(j)\tag{*}\\ &{\large=^?}\sum_{j=k+1}^m(-1)^{j-k}\binom{j-1}{k}N(j) \end{align}

The problem of $\textrm{(*)}$ is that it has to compute all $N(j), 0\le j\le m$, but the following line seems counter-intuitive since its first term is $$(-1)^1\binom{k}{k}N(k+1),$$

is there any wrong in my proof?

  • 1
    I seem to recall seeing those formulas (at least k, at most k, exactly k) in Feller's Introduction to Probability Theory and its Applications, Vol. 1. The derivation there is probably very clear and readable. – bof Sep 06 '20 at 09:32

2 Answers2

2

If I’m not mistaken, your final result is actually the negative of $\operatorname{Atleast}(k+1)$; adding the size of the universe to it would give you $\operatorname{Atmost}(k)$. Since $N(0)$ is the size of the universe, your final result is missing an $N(0)$ term.

The problem seems to arise in going from the third line to the fourth. The third line can be written

$$\sum_{j=0}^m(-1)^j\sum_{\ell=0}^{\min\{j,k\}}(-1)^\ell\binom{j}\ell N(j)\;,$$

and painstakingly reversing the order of summation makes it

$$\sum_{j=0}^k(-1)^j\sum_{\ell=0}^j(-1)^\ell\binom{j}\ell N(j)+\color{red}{\sum_{j=k+1}^m(-1)^j\sum_{\ell=0}^k\binom{j}\ell N(j)}\,,$$

where the red term ultimately reduces to your final line.

The black term still has an $N(0)$ term when $j=0$, but it disappears when one rewrites $\sum_{\ell=0}^j(-1)^\ell\binom{j}\ell$ as $\binom{j-1}j$, which is equivalent to your manipulation. The problem is that here we’re taking the alternating sum across the entire $j$-th row of Pascal’s triangle, and while the sum is $0$ for all $j\ge 1$, for $j=0$ it’s $1$. Thus, the black term actually reduces to the missing $N(0)$, and your final result should be

$$N(0)+\sum_{j=k+1}^m(-1)^{j-k}\binom{j-1}k N(j)\,.$$

Brian M. Scott
  • 616,228
2

In the third equation of $\text{Atmost}(k)$, $$\require{cancel} \begin{align} &\sum_{l=0}^k\sum_{j=l}^m(-1)^{j-l}\binom{j}{l}N(j)\\ &=\sum_{j=0}^m\sum_{l=0}^{\color{#C00}{\min(j,k)}}(-1)^{j-l}\binom{j}{l}N(j)\tag1\\ &=\sum_{j=0}^k\cancelto{[j=0]}{\sum_{l=0}^j(-1)^{j-l}\binom{j}{l}}N(j)+\sum_{j=k+1}^m\color{#090}{\sum_{l=0}^k}(-1)^{j-k}\color{#090}{\binom{-1}{k-l}\binom{j}{l}}N(j)\tag2\\[3pt] &=N(0)+\sum_{j=k+1}^m(-1)^{j-k}\color{#090}{\binom{j-1}{k}}N(j)\tag3 \end{align} $$ Explanation:
$(1)$: change the order of the summation
$(2)$: in the left sum $j\le k$ so limit of the inner sum is $j$
$\phantom{\text{(2):}}$ in the left sum $j\gt k$ so limit of the inner sum is $k$
$(3)$: in the left sum, the inner sum is $(1-1)^j=[j=0]$
$\phantom{\text{(3):}}$ in the right sum, we apply Vandermonde's Identity

As is marked in red in $(1)$, the upper index is $\min(j,k)$. The upper index in your answer leaves out all terms for $j\gt k$, but the next line seems to ignore this, and, without justification, puts the upper index at $j$, which makes it identical to line $(5)$ below. From there, you get the correct answer. The next line then removes all terms for $j\le k$, which is almost okay, since $\binom{j-1}{k}=0$ for all $j$ except $j=0$. Since your answer leaves out the $j=0$ term, your answer is short by $N(0)$.

However, an easier approach might be $$ \begin{align} \sum_{l=0}^k\sum_{j=l}^m(-1)^{j-l}\binom{j}{l}N(j) &=\sum_{l=0}^m\sum_{j=l}^m(-1)^{j-k}\binom{-1}{k-l}\binom{j}{l}N(j)\tag4\\ &=\sum_{j=0}^m\sum_{l=0}^j(-1)^{j-k}\binom{-1}{k-l}\binom{j}{l}N(j)\tag5\\ &=\sum_{j=0}^m(-1)^{j-k}\binom{j-1}{k}N(j)\tag6 \end{align} $$ Explanation:
$(4)$: $(-1)^{j-k}\binom{-1}{k-l}=(-1)^{j-l}[l\le k]$ so we can extend the outer sum to $m$
$(5)$: change the order of summation (easier since we extended the outer sum)
$(6)$: apply Vandermonde's Identity

Note that $(6)$ is the same as $(3)$; in $(6)$, the coefficient of $N(0)$ is $1$, and for $1\le j\le k$ the coefficient of $N(j)$ is $0$.

robjohn
  • 345,667
  • Is it possible to further simplify the (6)? I think it will take some time to compute all $N(j)$ from $j=0\to m$. Let me study your answer now... Thank you! – linear_combinatori_probabi Sep 06 '20 at 09:17
  • 1
    Other than noting that $\binom{j-1}{k}=0$ for $1\le j\le k$ (so that $N(j)$ does not need to be computed in that range), I don't see a way to simplify it. – robjohn Sep 06 '20 at 09:35