1
$\begingroup$

Good evening. I would like to show the following statement :

Let $n>1$, and $E,F$ two subsets of $[\![ 1,n ]\!]$ randomly and independently chosen such that $|E| = |F| = \lfloor\sqrt{ n }\rfloor + 1$. We have : $$\mathbb{P}(E \cap F \neq \varnothing) \geqslant 1 - e^{-\frac{7}{10}}$$

I tried to used the logarithm, i tried to use the convexity inequality $1-x \leqslant e^{-x}$ (true for every real numbers), but nothing worked. I didn't tried $\ln(1-x) \leqslant -x$ for the moment.

Why $e^{-\frac{7}{10}}$ is here ?!

The text says that this is a rewording of birthday paradox, I don't even see why.


My biggest attempt :

Proof : It is equivalent to show that : $$\mathbb{P}(E \cap F = \varnothing) \leqslant e^{-\frac{7}{10}}$$

Let $k := \lfloor \sqrt{n} \rfloor + 1$. Without loss of genrality we can suppose that $E = [\![1, k ]\!]$. We have : $$\mathbb{P}(E \cap F = \varnothing) = \frac{\binom{n-k}{k}}{\binom{n}{k}} = \frac{(n-k)! (n-k)!}{n! (n-2k)!} = \frac{(n-k) \dots (n-2k+1)}{n(n-1) \dots (n-k+1)}$$

We see that :

$\displaystyle\frac{n-k}{n} = 1 - \displaystyle\frac{k}{n} \leqslant e^{-\frac{k}{n}}$

$\displaystyle\frac{n-k-1}{n-1} = 1 - \displaystyle\frac{k}{n-1} \leqslant e^{-\frac{k}{n-1}}$

$\dots$

$\displaystyle\frac{n-2k+1}{n-k+1} = 1 - \displaystyle\frac{k}{n-k+1} \leqslant e^{-\frac{k}{n-k+1}}$


Thus : $$\mathbb{P}(E \cap F = \varnothing) \leqslant \prod\limits_{i=0}^{k-1} e^{-\frac{k}{n-i}} = e^{-k\sum\limits_{i=0}^{k-1} \frac{1}{n-i}}$$

As a consequence : $$\mathbb{P}(E \cap F = \varnothing) \leqslant e^{-(\lfloor \sqrt{n} \rfloor + 1) \sum\limits_{i=0}^{\lfloor \sqrt{n} \rfloor} \frac{1}{n-i}}$$


Can I say that : $$\sum\limits_{i=0}^{\lfloor \sqrt{n} \rfloor} \frac{1}{n-i} \leqslant \int_0^{\lfloor \sqrt{n} \rfloor + 1} \frac{\mathrm{d}x}{n-x} = \ln\left| \frac{n}{n -\lfloor \sqrt{n} \rfloor -1 } \right|$$

And then work with that or it's totally illegal.

$\endgroup$
8
  • $\begingroup$ Yes sorry, I correct it $\endgroup$
    – LexLarn
    Commented May 13 at 17:54
  • $\begingroup$ Hi! Where does this question come from? Is it supposed to be exactly true (doubtful), asymptotically true, approximately true, or something? ETA: Oops, sorry, missed the inequality. I'm guessing now it's a hard lower bound. But it would still be good to know where this comes from. $\endgroup$
    – Brian Tung
    Commented May 13 at 18:44
  • $\begingroup$ Just to answer one piece: $1-e^{-7/10}$ is in there because it's very close to $1/2$ ($\approx 0.503$). $\endgroup$
    – mjqxxxx
    Commented May 13 at 18:45
  • 1
    $\begingroup$ Hello ! It comes from this french document (proposition 4) : agreg.org/data/uploads/textes/public2015-A3.pdf Ok I see the interest of $e^{-\frac{7}{10}}$ now. $\endgroup$
    – LexLarn
    Commented May 13 at 18:51
  • 1
    $\begingroup$ I wonder, if you produced that LaTeX, why not simply type it in here, rather than pasting the resulting image? $\endgroup$
    – Brian Tung
    Commented May 13 at 18:57

1 Answer 1

1
$\begingroup$

It's related to the birthday paradox (not really a paradox, but still...) because it shows that if you have $2(\lfloor \sqrt{365} \rfloor + 1) = 40$ people, the probability that two of them have the same birthday is at least $1/2$. To see this, split them into two equal groups, $E$ and $F$. If there are duplicate birthdays within $E$ or $F$ alone, we certainly have duplicate birthdays in the full group. Otherwise (i.e., conditioned on all birthdays within $E$ and within $F$ being unique), there are duplicate birthdays exactly when $E \cap F \neq \emptyset$, which occurs (once you've proved the theorem above) with probability at least $1 - e^{-7/10} \approx 0.503 > 1/2$. Therefore, if $p$ is the probability of duplicate birthdays within $E$ or $F$ alone, the probability of duplicate birthdays in the full group of $40$ people is $> p + (1/2)(1-p) = 1/2 + p/2 \ge 1/2$.

$\endgroup$
4
  • $\begingroup$ Thank you for this explanation, this is clearer for me on that part ! $\endgroup$
    – LexLarn
    Commented May 13 at 19:01
  • $\begingroup$ I like this interpretation. It's clearly related to the birthday paradox (I might add that tag), but I had trouble articulating exactly what that relationship was. But I thought the OP's main question was how to prove the proposition. $\endgroup$
    – Brian Tung
    Commented May 13 at 19:02
  • 1
    $\begingroup$ But the usual cutoff for the birthday problem is $23$, not $40$. And $\lfloor{\sqrt{365}}\rfloor + 1 = 20$, so I'm confused. $\endgroup$ Commented May 13 at 19:02
  • $\begingroup$ Sure... this just gives an upper bound of $40$ for the usual birthday problem. (The factor of $2$ comes from the need to divide the people into two equal groups, $E$ and $F$, before applying the proven bound.) $\endgroup$
    – mjqxxxx
    Commented May 13 at 20:41

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .