In the classic Coupon Collector's problem, it is well known that the time $T$ necessary to complete a set of $n$ randomly-picked coupons satisfies $E[T] \sim n \ln n $,$Var(T) \sim n^2$, and $\Pr(T > n \ln n + cn) < e^{-c}$.

This upper bound is better than the one given by the Chebyshev inequality, which would be roughly $1/c^2$.

My question is: is there a corresponding better-than-Chebyshev lower bound for $T$? (e.g., something like $\Pr(T < n \ln n - cn) < e^{-c}$ ) ?

  • $\begingroup$ An obvious lower bound is $\Pr(T<n) = 0$, but I guess you're aware of that... $\endgroup$
    – onestop
    Mar 2 '11 at 8:56

I'm providing this as a second answer since the analysis is completely elementary and provides exactly the desired result.

Proposition For $c > 0$ and $n \geq 1$, $$ \mathbb{P}(T < n \log n - c n ) < e^{-c} \>. $$

The idea behind the proof is simple:

  1. Represent the time until all coupons are collected as $T = \sum_{i=1}^n T_i$, where $T_i$ is the time that the $i$th (heretofore) unique coupon is collected. The $T_i$ are geometric random variables with mean times of $\frac{n}{n-i+1}$.
  2. Apply a version of the Chernoff bound and simplify.


For any $t$ and any $s > 0$, we have that $$ \mathbb{P}(T < t) = \mathbb{P}( e^{-s T} > e^{-s t} ) \leq e^{s t} \mathbb{E} e^{-s T} \> . $$

Since $T = \sum_i T_i$ and the $T_i$ are independent, we can write $$ \mathbb{E} e^{-s T} = \prod_{i=1}^n \mathbb{E} e^{- s T_i} $$

Now since $T_i$ is geometric, let's say with probability of success $p_i$, then a simple calculation shows $$ \mathbb{E} e^{-s T_i} = \frac{p_i}{e^s - 1 + p_i} . $$

The $p_i$ for our problem are $p_1 = 1$, $p_2 = 1 - 1/n$, $p_3 = 1 - 2/n$, etc. Hence, $$ \prod_{i=1}^n \mathbb{E} e^{-s T_i} = \prod_{i=1}^n \frac{i/n}{e^s - 1 + i/n}. $$

Let's choose $s = 1/n$ and $t = n \log n - c n$ for some $c > 0$. Then $$ e^{s t} = n e^{-c} $$ and $e^s = e^{1/n} \geq 1 + 1/n$, yielding $$ \prod_{i=1}^n \frac{i/n}{e^s - 1 + i/n} \leq \prod_{i=1}^n \frac{i}{i+1} = \frac{1}{n+1} \> . $$

Putting this together, we get that $$ P(T < n \log n - c n) \leq \frac{n}{n+1} e^{-c} < e^{-c} $$

as desired.

  • $\begingroup$ That's very nice and just what the doctor ordered. Thank you. $\endgroup$
    – David
    Mar 6 '11 at 5:00
  • $\begingroup$ @David, just curious: What is the intended application? $\endgroup$
    – cardinal
    Mar 6 '11 at 5:01
  • $\begingroup$ Long story. I'm trying to prove a lower bound for the mixing time of a Markov chain that I've cooked up in order to analyze the running time of an algorithm I'm interested in - which turns out to reduce to lower-bounding the c.collector problem. BTW, I had been playing around with trying to find exactly this kind of Chernoff-style bound, but hadn't figured out how to get rid of that product in $i$. Good call choosing $s = 1/n$ :-). $\endgroup$
    – David
    Mar 6 '11 at 5:23
  • $\begingroup$ @David, $s = 1/n$, while almost certainly suboptimal, seemed like the obvious thing to try since that gave $e^{s t} = n e^{-c}$, which is the same term as the one obtained in the derivation for the upper bound. $\endgroup$
    – cardinal
    Mar 6 '11 at 17:48
  • 1
    $\begingroup$ Request: The proof I've given above is my own. I worked on it out of enjoyment, since the problem intrigued me. However, I make no claim to novelty. Indeed, I cannot imagine that a similar proof using a similar technique doesn't already exist in the literature. If anyone knows of a reference, please post it as a comment here. I would be very interested to know of one. $\endgroup$
    – cardinal
    Mar 7 '11 at 4:26

Although @cardinal has already given an answer that gives precisely the bound I was looking for, I have found a similar Chernoff-style argument that can give a stronger bound:

Proposition: $$ Pr (T \leq n \log n - c n) \leq \exp(- \frac{3c^2}{\pi^2} ) \> . $$ (this is stronger for $c > \frac{\pi^2}{3}$ )


As in @cardinal's answer, we can use the fact that $T$ is a sum of independent geometric random variables $T_i$ with success probabilities $p_i = 1 - i/n$. It follows that $E[T_i] = 1/p_i$ and $E[T] = \sum_{i=1}^{n} E[T_i] = n \sum_{i=1}^n \frac{1}{i}\geq n \log n$.

Define now new variables $S_i : = T_i - E[T_i]$, and $S : = \sum_i S_i$. We can then write $$ \Pr (T \leq n \log n - c n) \leq \Pr (T \leq E[T] - c n) = \Pr (S \leq - c n) $$ $$ = \Pr\left(\exp(-s S ) \geq \exp( s cn) \right) \leq e^{-s c n} E\left[ e^{-s S} \right] $$

Computing the averages, we have

$$ E[e^{-s S}] = \prod_i E[e^{-s S_i}] = \prod_i \frac{e^{s / p_i} } {1 + \frac{1}{p_i} (e^s -1)} \leq e^{\frac{1}{2}s^2\sum_i p_i^{-2}} $$ where the inequality follows from the facts that $e^s - 1\geq s$ and also $\frac{e^z}{1+z}\leq e^{\frac{1}{2}z^2}$ for $z\geq 0$.

Thus, since $\sum_i p_i ^{-2} = n^2 \sum_{i=1}^{n-1} \frac{1}{i^2} \leq n^2 \pi^2/6$, we can write \begin{align*} \Pr( T \leq n \log n - c n ) \leq e^{\frac{1}{12} (n \pi s)^2 - s c n}. \end{align*}

Minimizing over $s>0$, we finally obtain $$ \Pr( T \leq n\log n -cn ) \leq e^{-\frac{3 c^2 }{\pi^2}} $$

  • 1
    $\begingroup$ (+1) Modulo a couple of minor typos, this is nice. Expanding around something close to the mean as you've done often works better. I'm not surprised to see the higher order convergence in light of the asymptotic results. Now, if you show a similar such upper bound, that proves $(T-n\log n)/n$ is subexponential in the terminology of Vershynin, which has many implications regarding measure concentration. $\endgroup$
    – cardinal
    Mar 7 '11 at 18:47
  • 1
    $\begingroup$ The argument doesn't seem to generalize directly to the upper bound. Exchanging $c$ for $-c$ (and $s$ for $-s$), one can follow the same steps up to the point of calculating $E[e^{sS}] \leq \prod_i \frac{e^{-s/p_i}}{1 - \frac{s}{p_i}}$. At this point, however, the best I can do is to use $\frac{e^{-z}}{1-z} \leq \exp( \frac{z^2}{2(1-z)} )$, which still leaves $$ E[e^{sS}] \leq e^{\frac{1}{2} s^2\sum_i \frac{p_i^2}{(1-s/p_i)}}$$ and I don't know what to do with this $\endgroup$
    – David
    Mar 8 '11 at 3:27
  • 2
    $\begingroup$ Interestingly enough, though, the entire argument (for the lower bound) seems to work not only for the coupon collector problem, but for any sum of non-identical, independent geometric variables with bounded variance. Specifically: given $T = \sum_i T_i$, where each $T_i$ is an independent GV with success probability $p_i$, and where $\sum_i p_i^{-2} \leq A < \infty$, then $$ \Pr ( T \leq E[T] - a ) \leq e^{-\frac{a^2}{2A}}$$ $\endgroup$
    – David
    Mar 8 '11 at 3:35

Important Note: I've decided to remove the proof I gave originally in this answer. It was longer, more computational, used bigger hammers, and proved a weaker result as compared to the other proof I've given. All around, an inferior approach (in my view). If you're really interested, I suppose you can look at the edits.

The asymptotic results that I originally quoted and which are still found below in this answer do show that as $n \to \infty$ we can do a bit better than the bound proved in the other answer, which holds for all $n$.

The following asymptotic results hold

$$ \mathbb{P}(T > n \log n + c n ) \to 1 - e^{-e^{-c}} $$


$$ \mathbb{P}(T \leq n \log n - c n ) \to e^{-e^c} \>. $$

The constant $c \in \mathbb{R}$ and the limits are taken as $n \to \infty$. Note that, though they're separated into two results, they're pretty much the same result since $c$ is not constrained to be nonnegative in either case.

See, e.g., Motwani and Raghavan, Randomized Algorithms, pp. 60--63 for a proof.

Also: David kindly provides a proof for his stated upper bound in the comments to this answer.

  • $\begingroup$ Yes, it holds for every fixed $n$. A (very simple) proof can be found, for instance in Levin, Peres and Wilmer's book Markov Chains and Mixing Times, Proposition 2.4. The proof doesn't work for the lower bound, though. $\endgroup$
    – David
    Mar 3 '11 at 4:41
  • 1
    $\begingroup$ In fact, I might as well transcribe the proof here: "Let $A_i$ be the event that the $i$-th [coupon] type does not appear among the first $n \log n + cn$ coupons drawn. Observe first that $P(\tau >n\log n+cn )=P (\cup_{i} A_i ) \leq \sum_i P(A_i)$. Since each trial has probability $1 − n^{−1}$ of not drawing coupon $i$ and the trials are independent, the right-hand side above is bounded above by $\sum_i (1 - 1/n)^{n \log n + cn} \leq n \exp(\frac{n \log n + cn}{n} ) = e^{-c}$, proving (2.7)." $\endgroup$
    – David
    Mar 3 '11 at 5:17
  • $\begingroup$ @David, that's nice and simple enough. I quickly played with expanding the inclusion-exclusion formula out by another term, but didn't get anywhere quickly and haven't had time to look at it further. The event $\{T < t_n\}$ is equivalent to the event that no coupons are left after $t_n$ trials. There should be a martingale associated with that. Did you try Hoeffding's inequality on the (presumed) associated martingale? The asymptotic result suggest strong measure concentration. $\endgroup$
    – cardinal
    Mar 3 '11 at 18:48
  • $\begingroup$ @David, there's a sign flip in your proof above, but I'm sure that's obvious to other readers too. $\endgroup$
    – cardinal
    Mar 3 '11 at 18:50
  • $\begingroup$ @David, please see my other posted answer to your question. The method is different than the upper bound you give, but the tools employed are nearly as elementary, in contrast to the answer I gave here. $\endgroup$
    – cardinal
    Mar 5 '11 at 14:40

Benjamin Doerr gives (in the chapter " Analyzing Randomized Search Heuristics: Tools from Probability Theory" in the book "Theory of Randomized Search Heuristics", see the link for an online PDF) a somewhat simple proof of

Proposition Let $T$ be the stopping time of the coupon collection process. Then $\Pr[T\le (1-\epsilon)(n-1)\ln n]\le e^{-n^{\epsilon}}$.

This seems to give the desired asymptotics (from @cardinal's second answer), but with the advantage of being true for all $n$ and $\epsilon$.

Here is a proof sketch.

Proof Sketch: Let $X_i$ be the event that the $i$-th coupon is collected in the first $t$ draws. Thus, $\Pr[X_i=1]=(1-1/n)^t$. The key fact is that the $X_i$ are negatively correlated, for any $I\subseteq[n]$, $\Pr[\forall i\in I, X_i=1]\le\prod_{i\in I}\Pr[X_i=1]$. Intuitively, this is fairly clear, as knowing that the $i$-th coupon in the first $t$ draws would make it less likely that the $j$-th coupon is also drawn in the first $t$ draws.

One can prove the claim, but enlarging the set $I$ by 1 at each step. Then it reduces to showing that $\Pr[\forall i\in I, X_i=1|X_j=1]\le\Pr[\forall i\in I,X_i=1]$, for $j\notin I$. Equivalently, by averaging, it reduces to showing that $\Pr[\forall i\in I, X_i=1|X_j=0]\ge\Pr[\forall i\in I,X_i=1]$. Doerr only gives an intuitive argument for this. One avenue to a proof is as follows. One can observe that conditioned on the $j$ coupon coming after all of the coupons in $I$, that the probability of drawing a new coupon from $I$ after drawing $k$ so far is now $\frac{|I|-k}{n-1}$, instead of the previous $\frac{|I|-k}{n}$. So decomposing the time to collect all coupons as a sum of geometric random variables, we can see that conditioning on the $j$-coupon coming after $I$ increases the success probabilities, and thus doing the conditioning only makes it more likely to collect the coupons earlier (by stochastic dominance: each geometric random variable is increased, in terms of stochastic dominance, by the conditioning, and this dominance can then be applied to the sum).

Given this negative correlation, it follows that $\Pr[T\le (1-\epsilon)(n-1)\ln n]\le (1-(1-1/n)^t)^n$, which gives the desired bound with $t=(1-\epsilon)(n-1)\ln n$.

Note added in proof: The link above is outdated. A new version of this result with a complete proof (that is, including the negative-correlation statement) can be found in Theorem 1.9.3 in B. Doerr. Probabilistic Tools for the Analysis of Randomized Optimization Heuristics. In B. Doerr and F. Neumann, editors, Theory of Evolutionary Computation: Recent Developments in Discrete Optimization, pages 1-87. Springer, 2020.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.