Take the 2-minute tour ×
MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

GIVEN: Positive integers $n,m,L$ and probabilities $p_1, p_2, \ldots, p_n$.

GOAL: Choose $L$ size-$m$ subsets $S_1, S_2, \ldots, S_L$ of $\{1,2,\ldots,n\}$ to maximize $\displaystyle \mathbb{E}[ \max_{1\leq \ell \leq L} |S \cap S_\ell| ]$, where the expectation is taken over (all $2^n$) $S \subseteq \{1,2,\ldots,n\}$ using the pdf $f$ defined as \[ f(S) = \left( \prod_{i \in S} p_i \right) \left( \prod_{i \not\in S} (1-p_i) \right)\ . \]

(Equivalently, $S$ is chosen by examining each $i \in \{1,2,\ldots,n\}$ independently, choosing $i$ to be in $S$ with probability $p_i$.)

Note: the problem is trivial unless $m < n$ and $L < \binom{n}{m}$.

A nice, closed-form expression for $\displaystyle \mathbb{E}[ \max_{1\leq \ell \leq L} |S \cap S_\ell| ]$ (in terms of $n,m,L$ and the $p_i$ and the $S_\ell$) would be a good start.

EDIT:

WLOG, $p_1 \geq p_2 \geq \cdots \geq p_n$.

Special cases:

  • $L=1$: $S_1 = \{ 1, 2, \ldots, m \}$ maximizes $\displaystyle \mathbb{E}[ \max_{1\leq \ell \leq L} |S \cap S_\ell| ]$.

  • $m=1$: Denote by $x_i$ the single element of $S_i$ (i.e., $S_i = \{ x_i \}$) and $X = \{ x_1, x_2, \ldots, x_L \} = \bigcup S_i$. $\displaystyle \mathbb{E}[ \max_{1\leq \ell \leq L} |S \cap S_\ell| ] = \sum_{S \cap X \neq \emptyset} 1 \cdot f(S) = 1 - \sum_{S \subseteq \{ 1, 2, \ldots, n \} \setminus X} f(S) = 1 - \prod_{i \in X} (1-p_i)$, which is maximized by the greedy solution of $S_i = \{ i \}$ (for $1 \leq i \leq L$).

Here's how I examined a few more special cases:

This example ($(n,m,L)=(4,1,2)$) falls under the already discussed $m=1$ case, but I present it to hopefully make the presentation of subsequent examples clearer:

$\binom{\binom{n}{m}}{L} = \binom{\binom{4}{1}}{2} = 6$ possibilities for $(S_1,S_2)$:

1100 1ooo,o1oo,  p1+p2        -  p1p2
1010 1ooo,oo1o,  p1+   p3     -       p1p3
1001 1ooo,ooo1,  p1+      p4  -            p1p4
0110 o1oo,oo1o,     p2+p3     -                 p2p3
0101 o1oo,ooo1,     p2+   p4  -                      p2p4
0011 oo1o,ooo1,        p3+p4  -                           p3p4
  • Column 1: the sum of the subsequent $L=2$ columns (I used "o" for "0" in those for readability)
  • Column 2: represents $S_1$ (the 1st bit is 1 iff $1 \in S$ (0 otherwise), the 2nd bit is 1 iff $2 \in S$, etc.)
  • Column 3: represents $S_2$
  • Column 4: $\displaystyle \mathbb{E}[ \max_{1\leq \ell \leq L} |S \cap S_\ell| ]$ (formatted to put the degree-1 terms all together first, then the degree-2 terms, etc., with spacing so you can look up and down and see which other rows share those terms).

$\binom{\binom{n}{m}}{L} = \binom{\binom{4}{2}}{2} = 15$ possibilities for $(S_1,S_2)$:

2110 11oo,1o1o, p1+p2+p3     -                 p2p3
2101 11oo,1oo1, p1+p2+   p4  -                      p2p4
1210 11oo,o11o, p1+p2+p3     -       p1p3
1201 11oo,o1o1, p1+p2+   p4  -            p1p4
2011 1o1o,1oo1, p1+   p3+p4  -                           p3p4
1120 1o1o,o11o, p1+p2+p3     -  p1p2
1021 1o1o,oo11, p1+   p3+p4  -            p1p4
1102 1oo1,o1o1, p1+p2+   p4  -  p1p2
1012 1oo1,oo11, p1+   p3+p4  -       p1p3
0211 011o,o1o1,    p2+p3+p4  -                           p3p4
0121 011o,oo11,    p2+p3+p4  -                      p2p4
0112 01o1,oo11,    p2+p3+p4  -                 p2p3

1111 11oo,oo11, p1+p2+p3+p4  -       p1p3-p1p4-p2p3-p2p4       +  p1p2p3+p1p2p4+p1p3p4+p2p3p4  -  2p1p2p3p4
1111 1o1o,o1o1, p1+p2+p3+p4  -  p1p2-     p1p4-p2p3-     p3p4  +  p1p2p3+p1p2p4+p1p3p4+p2p3p4  -  2p1p2p3p4
1111 1oo1,o11o, p1+p2+p3+p4  -  p1p2-p1p3-          p2p4-p3p4  +  p1p2p3+p1p2p4+p1p3p4+p2p3p4  -  2p1p2p3p4

I separated the rows into two groups; two rows are in the same group if their first-column entries are permutations of each other.

Notes: the coefficient of 2 in the degree-4 terms of the last three rows seems interesting to me. The sign of a term is the negation of $(-1)$ to the degree of the term.

$\binom{\binom{n}{m}}{L} = \binom{\binom{4}{2}}{3} = 20$ possibilities for $(S_1,S_2,S_3)$:

3111  11oo,1o1o,1oo1,  p1+p2+p3+p4   -                  p2p3-p2p4-p3p4   +                        p2p3p4
1311  11oo,o11o,o1o1,  p1+p2+p3+p4   -        p1p3-p1p4-          p3p4   +                 p1p3p4
1131  1o1o,o11o,oo11,  p1+p2+p3+p4   -   p1p2-     p1p4-     p2p4        +          p1p2p4
1113  1oo1,o1o1,oo11,  p1+p2+p3+p4   -   p1p2-p1p3-     p2p3             +   p1p2p3

2220  11oo,1o1o,o11o,  p1+p2+p3                                          -   p1p2p3
2202  11oo,1oo1,o1o1,  p1+p2+   p4                                       -          p1p2p4
2022  1o1o,1oo1,oo11,  p1+   p3+p4                                       -                 p1p3p4
0222  o11o,o1o1,oo11,     p2+p3+p4                                       -                        p2p3p4

2211  11oo,1o1o,o1o1,  p1+p2+p3+p4   -             p1p4-p2p3-     p3p4   +                 p1p3p4+p2p3p4   -   p1p2p3p4
2121  11oo,1o1o,oo11,  p1+p2+p3+p4   -             p1p4-p2p3-p2p4        +          p1p2p4+       p2p3p4   -   p1p2p3p4
2211  11oo,1oo1,o11o,  p1+p2+p3+p4   -        p1p3-          p2p4-p3p4   +                 p1p3p4+p2p3p4   -   p1p2p3p4
2112  11oo,1oo1,oo11,  p1+p2+p3+p4   -        p1p3-     p2p3–p2p4        +   p1p2p3+              p2p3p4   -   p1p2p3p4
1221  11oo,o11o,oo11,  p1+p2+p3+p4   -        p1p3-p1p4-     p2p4        +          p1p2p4+p1p3p4          -   p1p2p3p4
1212  11oo,o1o1,oo11,  p1+p2+p3+p4   -        p1p3–p1p4-p2p3             +   p1p2p3+       p1p3p4          -   p1p2p3p4
2121  1o1o,1oo1,o11o,  p1+p2+p3+p4   -   p1p2-               p2p4–p3p4   +          p1p2p4+       p2p3p4   -   p1p2p3p4
2112  1o1o,1oo1,o1o1,  p1+p2+p3+p4   -   p1p2-          p2p3-     p3p4   +   p1p2p3+              p2p3p4   -   p1p2p3p4
1221  1o1o,o11o,o1o1,  p1+p2+p3+p4   -   p1p2-     p1p4-          p3p4   +          p1p2p4+p1p3p4          -   p1p2p3p4
1122  1o1o,o1o1,oo11,  p1+p2+p3+p4   -   p1p2-     p1p4-p2p3             +   p1p2p3+p1p2p4                 -   p1p2p3p4
1212  1oo1,o11o,o1o1,  p1+p2+p3+p4   -   p1p2-p1p3-               p3p4   +   p1p2p3+       p1p3p4          -   p1p2p3p4
1122  1oo1,o11o,oo11,  p1+p2+p3+p4   -   p1p2-p1p3-          p2p4        +   p1p2p3+p1p2p4                 -   p1p2p3p4

Note: the sign of a term is the negation of $(-1)$ to the parity of the degree of the term only in the first and third block of rows — the degree-3 terms in the 4 rows in the middle block are all negative.

share|improve this question
    
Presumably you have tried some special cases? or are you just hoping some here will just know the answer? –  Yemon Choi Jan 26 '10 at 7:09
    
While it would be great it someone knew the answer, realistically I was hoping at least for some insight/suggestions in how to proceed and/or related results/concepts. I have played with some special cases; I could add the details to the original post, but, as I'm new, I wasn't sure if that was appropriate. (I assume they don't belong in comments, if only because of the character limit.) –  Jay Howard Jan 26 '10 at 8:08
add comment

1 Answer

If you instead ask for a random subset of a fixed size, rather than a random subset of a random size, then being able to maximize the expected maximum size of the intersection would let you solve covering design problems.

"A $(v,k,t)$-covering design is a collection of $k$-element subsets, called blocks, of ${1,2,...,v}$, such that any $t$-element subset is contained in at least one block."

That means the complements of the blocks, of size $v-k$, have maximum intersection of size $v-k$ with each set of size $v-t$.

Let's return to subsets of random size. If you had an oracle which gave you the exact expected value, then setting the $p_i$ equal and close to 1, you can recover the minimum number of $t$-element subsets the complements don't cover as long as the complements cover each set of size at most $t-1$, hence whether a covering design is possible, which appears hard to determine.

So, it doesn't look like a closed form expression is feasible, and you should look at covering designs.

share|improve this answer
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.