Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

I am interested in making a certain statement on the probability distribution of "drawing a sample with $d$ distinct elements, when the sample is of size $k$ and is drawn (with replacement) from a set containing $N$ distinct elements ". Here is an example (just to avoid confusion):

--EXAMPLE--

$N=\{A,B,C\}$; $k=2$

The number of possible samples of length $2$ when drawing a random sample (with replacement) from $N$ is given by $3^{2}=9$. More specifically these samples are:

$\{A,A\}$,$\{B,B\}$,$\{C,C\}$ (1 distinct element in each sample)

$\{A,B\}$,$\{B,A\}$,$\{A,C\}$,$\{C,A\}$,$\{B,C\}$,$\{C,B\}$ (2 distinct elements in each sample)

So the probability of drawing 1 distinct element is 3/9, and the probability of drawing 2 distinct elements is 3/9. Any ideas?

--END OF EXAMPLE--

In general the probability is given by:

$Prob(d|k,N)=(\frac{1}{N^{k}})\frac{N!}{(N-d)!}S(k,d)$

where $S(k,d)$ are Stirling numbers of the second kind.

I would now like to make a statement on what happens if $k$ or $N$ changes. $Intuitively$, if $N$ or $k$ increase, it will become $less$ likely that you draw a sample with a $low$ number of distinct elements and $more$ likely that you draw a sample with a $high$ number of distinct items (probability mass is shifting towards high values of $d$).

If we therefore consider the sign of the change in probability (that is the sign of $Prob(d|k+1,N)-Prob(d|k,N)$), there should be some critical value $d^*$ such that if $d<d^*$ then the sign is negative, whereas if $d>d^*$ then the sign if positive. There would therefore only be $one$ sign change (this is what I ultimately want to prove).

$\textbf{Question}$: How do I go about proving this $formally$? One thing I cannot seem to get my head around is how to deal with the Stirling numbers.

Clearly if $d=1$,then $S(k,1)=1$ so $Prob(1|k,N)=N^{-k}$. For this case it is easy to see what happens if $N$ or $k$ change, but for the subsequent cases it is less obvious. One thing I have been trying is to exploit the recurrence relationship of the Stirling numbers: $S(k+1,d)=d*S(k,d)+S(k,d-1)$, but without any success. Using this approach I do already know that the expected number of distinct items is equal to:

$E[d]=N-\frac{(N-1)^{k}}{N^{k-1}}$

Which is clearly increasing in $k$ and $N$. This does however not suffice to show the above statement.

share|improve this question
    
I think you're trying to show the probabilities are unimodal, and maybe using that word as part of a websearch will turn up something useful for you. –  Gerry Myerson Aug 9 '12 at 7:22
    
Dear Gerry, thanks for the suggestion. $Prob(d|k,N)$ can indeed be shown to be (strongly) unimodal in $d$. This essentially follows from the log-concavity in $d$ of $S(k,d)$ and $\frac{1}{(N-d)!}$. However, what I want to show is what happens if $k$ or $N$ change. This new probability distribution will again be (strongly) unimodal. I now want to show that the difference in probability of these two distributions only has one sign change. –  Abe Doe Aug 9 '12 at 13:26
    
EDIT: Graphically the difference distribution (i.e. $Prob(d|k+1,N)-Prob(d|k,N)$) does not appear to be unimodal in $d$. –  Abe Doe Aug 9 '12 at 13:39

1 Answer 1

up vote 0 down vote accepted

An increase in $N$

Given that: $$Prob(d|k,N)=(\frac{1}{N^{k}})\frac{N!}{(N-d)!}S(k,d)$$

We have that $$Prob(d|k,N+1)\geq Prob(d|k,N)$$ if and only if $$\frac{\left( N+1\right) !}{\left( N+1\right) ^{k}\left( N+1-d\right) !}\geq\frac{N!}{N^{k}\left( N-d\right) !}$$

Solving for $d$ yields

$$d \geq d_{N}^{\ast}\left( k\right) =\left( N+1\right) \left( 1-\left( \frac{N}{N+1}\right) ^{k}\right) ,\text{ where }1<d_{N}^{\ast}\left( k\right) <N+1$$

So there would be only one sign change (i.e. at $d_{N}^{\ast}$)

An increase in $k$

Similarly we have that $$Prob(d|k+1,N)\geq Prob(d|k,N)$$ if and only if

$$\frac{1}{N^{k+1}}S\left( k+1,d\right)\geq\frac{1}{N^{k}}S\left(k,d\right)$$ which reduces to

$$\frac{1}{N} \geq R\left( k,d\right) =\frac{S\left( k,d\right) }{S\left( k+1,d\right) }$$

Now note that the LHS of the inequality is positive, weakly smaller than $1 $, and independent of $d$. The RHS is strictly decreasing in $d$ [See Theorem 3.2 in Sibuya (1988)]; It also holds that $R\left( k,1\right) =1$ $\ $and $R\left( k,k+1\right) =0$. Consequently there exists a unique $1\leq d_{k}^{\ast}\left( N\right) \leq k+1$ such that $\frac{1}{N}\geq R\left( k+1,d\right) $ if $m\geq d_{k}% ^{\ast}\left( N\right) $.

So there would be only one sign change (i.e. at $d_{k}^{\ast}$)

Reference:

SIBUYA, M. "Log-Concavity of Stirling Numbers and Unimodality of Stirling Numbers", Ann. Ins. Statist. Math., Vol. 40, Nr. 4 (1988), pp. 693-713.

share|improve this answer

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.