Say you have some finite, non-empty sets $S_1 \ldots S_n$. You are not able to inspect the contents of these sets directly, or even produce their unions or intersections; the only way you can learn about them is through the function $P$, defined as: $$ P(S_i, S_j)=\frac{|S_i \cap S_j|}{|S_i|} $$

In other words, $P$ tells you what fraction of $S_i$ is also contained in $S_j$—i.e., it captures precision.

What I'd ideally like to be able to do is compute $$1 - P\left(S_n, \bigcup_{i=1}^{n-1} S_i\right),$$ AKA how much of $S_n$ is novel with respect to $S_1 \ldots S_{n-1}$.

Of course, this is impossible given only the pairwise $P$ function; an exact result would also need to account for $k$-way intersections between the various sets.

The question, though, is whether it's possible to compute some nontrivial bounds on this value. Intuitively, it seems like all these pairwise intersection sizes should give you a decent amount of information about size of the union of $S_1 \ldots S_{n-1}$. What are the tightest bounds available on the expression above?

Relatedly, does it change anything if we are able to access $|S_i|$? For example, I've worked out (tentatively) that in the case of $n=3$, $$ P(S_3, S_1 \cup S_2) \geq \frac{|S_3 \cap S_1| + |S_3 \cap S_2| - \min(|S_1 \cap S_2|, \max(|S_1 \cap S_3|, |S_2 \cap S_3|))}{|S_3|} $$

And given the $|S_i|$s, all of these values should be computable from the $P(S_i, S_j)$. But I'm still not sure how to generalize this to larger $n$, and in particular whether the bound would become tighter or looser as $n$ grows.

  • $\begingroup$ You didn't write so explicitly, but it seems that all sets $S_i$ are finite. If so, perhaps "translating" this problem into more conventional probabilistic terms would be helpful. For example, your function $P(S_i,S_j)$ is typically written $P(S_j | S_i)$ (notice the reversed order of $S_i$ and $S_j$) and called the conditional probability of $S_j$ given $S_i$. The problem would be more general, then, but also perhaps more recognizable to some. $\endgroup$ – Menachem May 20 '20 at 7:16
  • $\begingroup$ Also, I think that the standard inclusion-exclusion principle would give you $P(S_3, S_1 \cup S_2) = \frac{|S_3\cap S_1| + |S_3 \cap S_2| - |S_3 \cap S_1 \cap S_2|}{|S_3|}$, instead of the inequality you suggested. $\endgroup$ – Menachem May 20 '20 at 7:38
  • $\begingroup$ @Menachem Yes, that is indeed the exact formula for $P(S_3, S_1 \cup S_2)$. But again, I can't compute the last term there exactly, because that would require knowing the three-way intersection, which I don't have. The point of the formula I gave above is to bound the size of the three-way intersection in terms of two-way intersections. (Note the $\geq$.) $\endgroup$ – duncanka May 29 '20 at 18:22
  • $\begingroup$ And good point about casting this in terms of probabilities. It doesn't actually have a sensible probabilistic interpretation in my case, but the math is indeed the same. This question is on the back burner for me right now, but when (if) it becomes significant again I'll revise accordingly. Thanks for the suggestion! $\endgroup$ – duncanka May 29 '20 at 18:25

I'm not a set expert, so there may be a tighter bound. But one nontrivial bound (if you can assume all sets are the same size, if not then I'm not sure about a possible bound) is $$1 - \sum_{i=1}^{n-1} P(S_n,S_i)$$

As an example, assume there are 3 sets. Set 1 and Set 2 have a 1/3 overlap and Set 1 and Set 3 have a 1/4 overlap. in the worst case, Set 2 and Set 3 are disjoint, so the novel part of Set 1 must be equal to or less than $1- (\frac{1}{4} + \frac{1}{3})$. $(1-\frac{7}{12}) = \frac{5}{12}$. So $\frac{1}{3}\leq 1 - P(S_1,\bigcup_{i=2}^3 S_i) \leq \frac{5}{12}$ for this example. Since the overlap cannot be negative, any value less than 0 would be calculated as 0 or greater.

You could also calculate a tighter bound recursively. If you know that $P(S_2,S_3)$ is .9, then even if $P(S_1, S_3)$ is greater than .1, it can only add at most .1 to the value of $P(S_1, S_2)$.

Sorry if the bound isn't tight enough, but hopefully this bound will help. And it shows there is a nontrivial solution, if an additional assumption is made.

  • $\begingroup$ Thanks, Zack! Will think on this; in the meantime, a quick typo correction: I think you meant n-1 at the top of the summation. $\endgroup$ – duncanka May 19 '20 at 21:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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