Suppose that we have a 6-sided unfair dice, where rolling a 1 is twice as likely as rolling any other number, and the other numbers have the same likelihood. What is the expected number of rolls to get each value at least once?

Thus, $$p(1) = 2/7\qquad p(2) = p(3) = \cdots = p(6) = 1/7$$

I understand how to approach this when the probabilities are the same, as it's just the Coupon collector's problem, but throwing in a non-uniform probability distribution throws me off. I know there is a general solution for this problem, but I can't seem to get any intuition as to why it's the case.

  • $\begingroup$ I think the full solut9nn should be $\frac{7}{2} + \frac{7}{5}+ \frac{7}{4}+\frac{7}{3}+\frac{7}{2}+ 7$ $\endgroup$ – Alex May 20 '16 at 18:00
  • 2
    $\begingroup$ @Alex, that is the expected length of time to roll a one, and then only considering rolls after a one is rolled see each other number at least once. Numbers other than a $1$ are able to be rolled before the one is rolled, and potentially need not be searched for after a one is rolled. $\endgroup$ – JMoravitz May 20 '16 at 18:02
  • $\begingroup$ I have a solution, it is inelegant, and involves the use of an $11$ state markov chain. I expect there to be a cleaner solution, but will begin work on writing it up. $\endgroup$ – JMoravitz May 20 '16 at 18:05
  • $\begingroup$ @JMoravitz I have the answer using an expression from the following paper which solves it generally: combinatorics.org/ojs/index.php/eljc/article/view/v20i2p33/pdf However, my hope was to find a cleaner more elegant solution. $\endgroup$ – Andrew Mascioli May 20 '16 at 18:07
  • 1
    $\begingroup$ I'm a bit perplexed. You both expressed hope for a cleaner, more elegant solution; I provided one, and now you accepted the $11$-state Markov chain without even commenting on my answer -- is there something wrong with it? (This of course with great respect for @JMoravitz's contributions to the site, which are often very elegant.) $\endgroup$ – joriki May 21 '16 at 5:26

Consider the die as a $7$-sided die with two $1$s and wait for all sides to appear. With probability $\frac27$, you unnecessarily waited for the second $1$ for an expected number of $7$ rolls at the end. Correcting for that yields


  • $\begingroup$ Simple and clever! +1 $\endgroup$ – Markus Scheuer May 21 '16 at 5:43
  • $\begingroup$ Very nice solution! +1 $\endgroup$ – drhab May 21 '16 at 7:04

We describe the scenario using an eleven state markov chain. (tedious, I know...if all six die results have different probabilities though, could go upwards of $2^6$ states though, so at least we could trim it down to only eleven...)

Here is a transition diagram.

unfair die couponcollector transition diagram

To give names to states, I will refer to them as $(yes,k)$ for if $1$ has been seen and $k$ not-1 numbers have been rolled, and $(no,k)$ for if $1$ has not been seen and $k$ not-1 numbers have been rolled.

The transition matrix with order of rows/columns as $(yes,5),(yes,0),(yes,1),\dots,(yes,4),(no,1),(no,2),\dots,(no,5)$

$$\begin{bmatrix} 1&0&0&0&0&1/7&0&0&0&0&2/7\\ 0&2/7&0&0&0&0&0&0&0&0&0\\ 0&5/7&3/7&0&0&0&2/7&0&0&0&0\\ 0&0&4/7&4/7&0&0&0&2/7&0&0&0\\ 0&0&0&3/7&5/7&0&0&0&2/7&0&0\\ 0&0&0&0&2/7&6/7&0&0&0&2/7&0\\ 0&0&0&0&0&0&1/7&0&0&0&0\\ 0&0&0&0&0&0&4/7&2/7&0&0&0\\ 0&0&0&0&0&0&0&3/7&3/7&0&0\\ 0&0&0&0&0&0&0&0&2/7&4/7&0\\ 0&0&0&0&0&0&0&0&0&1/7&5/7\end{bmatrix}$$

Following properties of absorbing markov chains, we look to the fundamental matrix $(I-R)^{-1}$

Using a calculator,

enter image description here

enter image description here

Now, since we start in state $(yes,0)$ with probability $2/7$ and start in state $(no,1)$ with probability $5/7$, multiplying our fundamental matrix by the corresponding vector, we get:

$(7/5+7/4+7/3+7/2+7)2/7 + (7/12+7/5+14/5+98/15+7/6+14/15+7/10+7/15+7/30)5/7 = \frac{303}{20}$

Since this calculation assumed however that we started in one of those two states however, the initial roll must also be considered, so our final expected time is found by adding one to it, giving $\frac{323}{20} = 16.15$ rolls on average.

As a side note, the method employed is the same as that of dr.hab, however with the use of technology and the language of matrices, we can conveniently write and compute the result.


Say $T_k$ is the time it takes to get from the $(k-1)$-th to the $k$-th new roll, and $C_k$ is the value of the $k$-th new roll.

You can calculate $\mathbb{E}[T_k]$ by $\mathbb{E}[\mathbb{E}[T_k | C_1,...,C_{k-1}]]$ instead. For example,

$$\mathbb{E}[T_2] = \mathbb{E}[\mathbb{E}[T_2 | C_1]] = \sum_{j=1}^6 \mathbb{E}[T_2 | C_1 = j] \mathbb{P}(C_1 = j) = \frac{7}{5} \cdot \frac{2}{7} + 5 \cdot \frac{7}{6} \cdot \frac{1}{7} = \frac{37}{30}.$$

This is worked out in general on page $9$ and $10$ of this link.

  • $\begingroup$ The link was very helpful! Thanks! $\endgroup$ – Andrew Mascioli May 20 '16 at 19:13

For $k=0,1,2,3,4,5$ let $S_{k}$ denote the status where $1$ has not been rolled yet and exactly $k$ of the other numbers have been rolled.

For $k=0,1,2,3,4,5$ let $T_{k}$ denote the status where $1$ has been rolled and exactly $k$ of the other numbers have been rolled.

Let $\mu_{k}$ denote the expectation of rolls still to go starting in status $S_{k}$.

Let $\nu_{k}$ denote the expectation of rolls still to go starting in status $T_{k}$.

To be found is $\mu_0$ and we have the following relations:

$$\nu_{5}=0\tag1$$ $$\mu_{5}=1+\frac{2}{7}\nu_{5}+\frac{5}{7}\mu_{5}=1+\frac{5}{7}\mu_{5}\tag2$$

for $k=0,1,2,3,4$:



Note that (2) leads to $\mu_5=3.5$. Substituting this together with $\nu_5=0$ in (3) and (4) gives you for $k=4$ two equalities that enable you to find $\nu_4$ and $\mu_4$. This proces can be repeated till you arrive at values for $\nu_0$ and (our final goal) $\mu_0$.


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.