hypergeometric distribution - expected number of draws until k successes are drawn

Hypergeometric distribution describes the probability of k successes in n draws from population of N with K successes. But what if I want expected number of draws until k successes are drawn from the same population?

• Wider question: in a population of N is 1 ill person. At each moment of time, two people meet. If one of them is ill and the other healthy, the healthy gets ill. What is the expected time count until at least n people are ill? Jan 7 '18 at 11:32
• So, in you case you need to apply the following trick: you keep drawing until $k$ successes, you last draw (the $k$-th one) is a success, the remaining $k-1$ are scattered across $n-1$ draws. Or $P(n \text{ draws until }k\text{ successes})=P(\text{last draw is success})\cdot P(n-1 \text{ draws have }k-1\text{ successes})$ Jan 7 '18 at 11:43
• okay but the expected value of the second factor is still dependent on n, specifically it is (n-1)K/N Jan 7 '18 at 11:59
• As per my comment $$P(n \text{ draws until }k\text{ successes})=P(\text{last draw is success})\cdot P(n-1 \text{ draws have }k-1\text{ successes})$$ $P(n-1 \text{ draws have }k-1\text{ successes})$ is the standard hyper geometric probability $$P(n-1 \text{ draws have }k-1\text{ successes})=\frac{\binom{K}{k-1}\binom{N-K}{n-k+1}}{\binom{N}{n-1}}$$ and (everything is without replacement) $$P(\text{last draw is success})=\frac{1}{K-k+1}$$ And $$E[X]=\sum\limits_{n\geq k}nP(n \text{ draws until }k\text{ successes})$$ Jan 7 '18 at 12:03
• ah! you can use the infinite sum to get the expected value, thanks! if you submit this as an aswer, I will gladly accept it. Jan 7 '18 at 12:07

As per my comments $$P(n \text{ draws until }k\text{ successes})=P(\text{last draw is success})\cdot P(n-1 \text{ draws have }k-1\text{ successes})$$

$P(n-1 \text{ draws have }k-1\text{ successes})$ is the standard hyper geometric probability $$P(n-1 \text{ draws have }k-1\text{ successes})=\frac{\binom{K}{k-1}\binom{N-K}{n-1-k+1}}{\binom{N}{n-1}}=\frac{\binom{K}{k-1}\binom{N-K}{n-k}}{\binom{N}{n-1}}$$ and (everything is without replacement) $$P(\text{last draw is success})=\frac{K-k+1}{N-n+1}$$ $E[X]$ is a function of $k$, also considering $\binom{n}{k}=\binom{n}{n-k}$ $$\color{red}{E[X]=}\sum\limits_{n\geq k}nP(n \text{ draws until }k\text{ successes})=\sum\limits_{n\geq k}n\frac{\binom{K}{k-1}\binom{N-K}{n-k}}{\binom{N}{n-1}}\frac{K-k+1}{N-n+1}=\\ \sum\limits_{n\geq k}n\frac{\binom{K}{K-k+1}\binom{N-K}{n-k}}{\binom{N}{N-n+1}}\frac{K-k+1}{N-n+1}=\\ \binom{K}{K-k+1}(K-k+1)\sum\limits_{n\geq k}n\frac{\binom{N-K}{n-k}}{\binom{N}{N-n+1}(N-n+1)}=\\ \frac{K!}{(K-k+1)!(k-1)!}(K-k+1)\sum\limits_{n\geq k}n\frac{\binom{N-K}{n-k}}{\frac{N!}{(N-n+1)!(n-1)!}(N-n+1)}=\\ \frac{K!}{(K-k)!(k-1)!}\sum\limits_{n\geq k}n\frac{\binom{N-K}{n-k}}{\frac{N!}{(N-n)!(n-1)!}}=\\ k\frac{K!}{(K-k)!k!}\sum\limits_{n\geq k}\frac{\binom{N-K}{n-k}}{\frac{N!}{(N-n)!n!}}= \color{red}{k\binom{K}{k}\sum\limits_{n\geq k}\frac{\binom{N-K}{n-k}}{\binom{N}{n}}}$$

I have tried the following Python $3$ code

from scipy.special import comb
from decimal import *

N = 10000
K = 1000

k = 1000

res = Decimal(k * comb(K, k, exact=True))
sum = Decimal('0.0')
one = Decimal('1.0')

def combinations(b, a):
if (b >= a):
return comb(b, a, exact=True)
else:
return 0

for n in range(k, N - K + k + 1):
sum = sum + (Decimal(combinations(N-K, n-k)) * one) / Decimal(combinations(N, n))

res = res * sum
print(res)


producing

k=1     E[X]=9.991008991008991008991008979
k=10    E[X]=99.91008991008991008991008991
k=100   E[X]=999.1008991008991008991008979
k=1000  E[X]=9991.008991008991008991008992


all within 8 minutes (altogether, so $\approx 2$ minutes per run) on my MacBook Pro. And

N = 10000 K=5000 k=3000 E[X]=5999.400119976004799040191959


took less than $2$ minutes. Larger values of $N$ take longer to compute (E.g. $N=100000$ I had to stop the program after 1 hour).

Although, it looks like empirically $$E[X]\approx k\cdot \frac{N}{K}$$

• yes, I had a very similar code, although my CPU isn't as powerful. Considering I would have to evaluate E[X] for all K in 1...1000, it would take hours, but the principle works. Jan 9 '18 at 20:02
• I think, it is worth focusing on proving the empirical result. Jan 9 '18 at 20:11