A What is the Probability of Sampling a Set with a Certain Size?

[yahastu]

I've come up with a problem that's important for an algorithm I'm developing, and I don't know how to solve it. Wondering if anyone here can help?

I have an initial set S1 of size N.

S2 is created by randomly sampling M samples from S1 with replacement (meaning the same item could be selected multiple times).

Let C = |S2| be the size of S2 (ie, the number of unique elements that get sampled).

For what value of X can we say that C > X with 95% probability?

It would be easy to solve this problem through numerical simulation, but I need a fast analytical solution (or approximate).

Any ideas?

 

[Office_Shredder]

Good approximate solutions probably require knowing what M is relative to N. If M<<N then M is going to be your best guess (you'll never grab the same item twice), if $M\approx N/2$ you're going to get a very different answer.

 

[yahastu]

Good approximate solutions probably require knowing what M is relative to N. If M<<N then M is going to be your best guess (you'll never grab the same item twice), if $M\approx N/2$ you're going to get a very different answer.

I did a numerical simulation, and found that the relationship between M/N and any given percentile of the distribution of C is linear. Using a linear regression I computed that the solution to X above is approx given by:

X = M( 0.922954 - 0.3144(M/N) )

So this is cool, can't beat it in terms of efficiency, although I have absolutely no clue how one would go about deriving this relationship analytically!

 

[Office_Shredder]

Let $e_m$ be the expected number of unique hits after $m$ drops. Then roughly, $e_{m+1}\approx e_m+(1-e_m/N)$. And of course $e_1=1$ so you could solve that recurrence relation numerically and see how it compares...

 

[Stephen Tashi]

Is the relevant random variable $Y$, a sum of independent geometric random variables?

Let $X_i$ be the geometric random variable defined as the number of trials, up to an including the trial that attains the first "success" in a sequence of independent trials where the probability of success on each trial is $\frac{N-i}{N}$

Define $Y = \sum_{i=0}^M X_i$ Does a result $Y = k$ correspond to getting $M$ distinct samples after taking $k$ samples from a population of size $N$ using random sampling with replacement?

 

[Stephen Tashi]

so you could solve that recurrence relation numerically and see how it compares...

Recursion is a good idea. We can apply it to recursively generate the relevant distributions.

Let $N$ be the population size, and let $M$ the number of samples taken. We can imagine the samples taken taken in a sequence. Let $f_M()$ be the probability distribution for the number of distinct populations members in the sample after M samples are taken. So $f_M(k)$ is the probability that after $M$ samples are taken that exactly $k$ distinct members of the population are in the sample. Let $F_M(k)$ be the cumulative distribution for $f$.

Consider computing $F_{M+1}(k+1)$. In order to have $k+1$ or fewer samples after the $M+1$ sample is taken we must be in one of the following mutually exclusive cases after the $M$th sample was taken.

1) In the previous $M$ samples there are $k$ or fewer distinct members in the sample.

2) In the previous $M$ samples there are exactly $k+1$ distinct members in the sample

So we have the recursion relation

$F_{M+1}(k+1) = F_M(k) + (f_M(k+1))(\frac{k+1}{N})\ \$ [eq. 1]

which is based on computing the probability of transitioning between each of the above cases to the situation of having $k+1$ or fewer distinct members after the $M+1$ sample is taken.We have the initial conditions
$f_1(1) = 1, F_1(1) = 1$
$F(k) =1$ for $k \ge M$
$f_M(k) = 0$ for $k > M$ or $k < 1$Examples:

Applying eq. 1 with $M = 1$ we have:

$F_2(1) = F_1(0) + f_1(1) \frac{1}{N} = 0 + (1)(\frac{1}{N}) = \frac{1}{N}$

$F_2(2) = F_1(1) + f_1(2)\frac{2}{N} = 1 + 0 = 1$

which implies $f_2(1) = \frac{1}{N},\ f_2(2) = 1 - \frac{1}{N} = \frac{N-1}{N}$

For $M = 2$ we have:

$F_3(1) = F_2(0) + (f_2(1))(\frac{1}{N}) = 0 + (\frac{1}{N})(\frac{1}{N}) = \frac{1}{N^2}$

$F_3(2) = F_2(1) + (f_2(2))(\frac{2}{N}) = \frac{1}{N} + (\frac{N-1}{N})(\frac{2}{N}) = \frac{3N-2}{N^2}$

$F_3(3) = F_2(2) + f_2(3)(\frac{3}{N}) = 1 + (0)(\frac{3}{N}) = 1$

which implies

$f_3(1) = \frac{1}{N^2}$
$f_3(2) = \frac{3N-2}{N^2} - \frac{1}{N^2} = \frac{2N-2}{N^2}$
$f_3(3) = 1 - \frac{3N-2}{N^2} = \frac{N^2 - 3N + 2}{N^2}$

I don't know if pursuing the symbolic algebra implied by the recursion works out in a nice way. At least the recursion implies a numerical algorithm for generating $F_M$.

 

[Office_Shredder]

Anyway, assuming each step is small, and changing e to f for reasons that will be obvious, so f(m) is the number of buckets filled after m drops

$f_{m+1}-f_m \approx 1-f_m/N$

Turns into a separable differential equation

$df/dm = 1-f/N$.

This is separable,
$df/(1-f/N) = dm$
So,
$-ln(|1-f/N|)= m+C$
For some constant C, which turns into

$1-f/N =Be^{-m}$
For some constant B. Solving for f yields
$f=N-NBe^{-m}$.

We know f(0)=0 since we haven't dropped anything yet so $B=1$. Then $f(N) = N-N/e$. 1-1/e is about 0.632 which matches pretty closely the linear regression value of 0.922954 - 0.3144 from the post above.