MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

The coupon collector's problem is a problem in probability theory that states the following (from wikipedia):

Suppose that there are $n$ coupons, from which coupons are being collected with replacement. What is the probability that more than $t$ sample trials are needed to collect all $n$ coupons?

A generalization of this problem was proposed by Newmann & Shepp, by requiring that $k$ samples of each coupon be collected. The answer to this is known.

I, however, need to calculate the answer to an even further generalization, which is:

How many sample trials are needed to collect at least $k$ coupons of $m$ different types?

Any help or a point in the right direction would be greatly appreciated.

(Edit#3): removed the misleading example.

(Edit #2): This problem can be stated simply as a balls-and-bins problem. If we have $n$ bins, how many balls need to be thrown so that at least $m$ bins have at least $k$ balls?

share|cite|improve this question
Maybe this question is confusing since it is not clear to me why we are not just considering n to be the number of coupons in the set m and using 1/N as the probability of drawing a specific coupon when N is the original number of coupons. Follow the proof of the generalization putting 1/N whenever 1/n was used. – Orange Oct 12 '10 at 19:41
I added an edit that hopefully explains the difference. – Herman Oct 12 '10 at 20:00
Your Edit #2 appears to be a different problem: as stated, you want the first $m$ bins to have at least $k$ balls, right? – JBL Oct 12 '10 at 21:53
I see I might have stated it badly earlier, the 2nd edit is the problem I actually wish to solve. So my original question should actually read How many sample trials are needed to collect at least $m$ coupons at least $k$ times? - and by that I mean any $m$ coupons. My example is misleading as well, sorry about that. The question is much more easily stated as a ball-and-bins problem, as I later realized. – Herman Oct 12 '10 at 21:59
You should delete the example. It's very misleading. – Andrew D. King Oct 12 '10 at 22:32

Perhaps this is naive, but shouldn't the expected number of trials be $M/m * T$ where $T$ is the expected number of trials when $M = m$ (i.e.: in the case worked out by Newmann & Shepp)? If half the coupons are useless to you, wouldn't it just take twice as long to collect the ones you wanted?

share|cite|improve this answer
This looks right for the question as originally stated, but the restated version is harder. – JBL Oct 12 '10 at 22:07
Ah. I see. Originally, it seemed that there was a fixed subset of size $m$ that was of interest. Now, it's any $m$-subset. – mhum Oct 12 '10 at 22:44

If you want the expected value, one answer is $n E[S_{(m)}]$, where $S_{(m)}$ is the $m$th order statistic of a sample of $n$ gamma$(k,1)$ random variables. While this expression may not have a simple closed form, you may be able to get a decent-sized approximate answer from the literature on moments of order statistics. (Edit: This appears to be the case, even when comparing with the known asymptotic expression for the case $m=n$. See discussion at end.)

Here's the argument for $n E[S_{(m)}]$: Take a Poisson process $P$ with rate 1 and interarrival times $Z_1, Z_2, \ldots$. Let each event in the process $P$ have probability of $1/n$ of being the first kind of coupon, probability $1/n$ of being the second kind of coupon, and so forth. By the decomposition property of Poisson processes, we can then model the arrival of coupon type $i$ as a Poisson process $P_i$ with rate $1/n$, and the $P_i$'s are independent. Denote the time until process $P_i$ obtains $k$ coupons by $T_i$. Then $T_{i}$ has a gamma$(k,1/n)$ distribution. The waiting time until $m$ processes have obtained $k$ coupons is the $m$th order statistic $T_{(m)}$ of the iid random variables $T_1, T_2, \ldots, T_n$. Let $N_m$ denote the total number of events in the processes at time $T_{(m)}$. Thus $N_m$ is the random variable the OP is interested in. We have

$$T_{(m)} = \sum_{r=1}^{N_m} Z_r.$$

Since $N_m$ and the $Z_r$'s are independent, and the $Z_r$ are iid exponential(1), we have

$$E[T_{(m)}] = E\left[E\left[\sum_{r=1}^{N_m} Z_r \bigg| N_m \right] \right] = E\left[\sum_{r=1}^{N_m} E[Z_r] \right] = E\left[N_m \right].$$

By scaling properties of the gamma distribution, $T_i = n S_i$, where $S_i$ has a gamma$(k,1)$ distribution. Thus $T_{(m)} = n S_{(m)}$, and so $E\left[N_m \right] = n E[S_{(m)}]$.

For more on this idea, see Lars Holt's paper "On the birthday, collectors', occupancy, and other classical urn problems," International Statistical Review 54(1) (1986), 15-27.

(ADDED: Looked up literature on moments of order statistics.)

David and Nagaraja's text Order Statistics (pp. 91-92) implies the bound $$n P^{-1}\left(k,\frac{m-1}{n}\right) \leq n E[S_{(m)}] \leq n P^{-1}\left(k,\frac{m}{n}\right),$$ where $P(k,x)$ is the regularized incomplete gamma function.

Some software programs can invert $P$ for you numerically. Trying a few examples, it appears that the bounds given by David and Nagaraja can be quite tight. For example, taking $n$ = 100,000, $m$ = 50,000, and $k$ = 25,000, the two bounds give estimates (via Mathematica) around $2.5 \times 10^9$, and the difference between the two estimates is about 400. More extreme values for $k$ and $m$ give results that are not as good, but even values as extreme as $m$ = 10, $k$ = 4 with $n$ = 100,000 still yield a relative error of less than 3%. Depending on the precision you need, this might be good enough.

Moreover, these bounds seem to give better results for $m \approx n$ versus using the asymptotic expression for the case $m = n$ given in Flajolet and Sedgewick's Analytic Combinatorics as an estimate. The latter has error $o(n)$ and appears to be for fixed $k$. If $k$ is small, the asymptotic estimate is within or is quite close to the David and Nagaraja bounds. However, for large enough $k$ (say, on the order of $n$) the error in the asymptotic is on the order of the size of estimate, and the asymptotic expression can even produce a negative expected value estimate. In contrast, the bounds from the order statistics approach appear to get tighter when $k$ is on the order of $n$.

(Caution: There are two versions of the regularized incomplete gamma function: the lower one $P$ that we want with bounds from $0$ to $x$, and the upper one $Q$ with bounds from $x$ to $\infty$. Some software programs use the upper one.)

share|cite|improve this answer

Possibly this is helpful:

Extended Pigeonhole Principle: If $nk+1$ objects are placed in $n$ boxes, then one of the boxes must contain at least $k+1$ objects.

Combine this with the probabilistic view of the Pigeonhole Principle briefly explained here: and maybe that's what you're looking for?

share|cite|improve this answer

Your Answer


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.