25

The first round begins with a $\frac1{10}$ chance of you winning. If you win, congratulations! Once you have won, the simulation is complete. If not, you enter a second round. This second round gives you a $\frac1{100}$ chance of winning. If you win, you stop. If you lose again, you are entered into a third round where you have a $\frac1{1000}$ chance of winning. This process continues instantaneously forever. How likely are you to win this multi-stage lottery?

Some arguments:

  • Argument A: You will win every time because you have infinite tries.

  • Argument B: You will win $\frac19$ times because as you approach infinite tries, you approach a probability of winning of $0$.

  • Argument C: Other.

Who's asking? An enthusiastic high-school freshman with little understanding of probabilities and even less of advanced mathematical notation.

  • 2
    (off-topic) Puzzles like this where you have to deal with uncountably infinitely many possible outcomes (and maybe some slightly more advanced ones) make a better motivation for definition of measure space than the dry "Lebesgue measure is the completion of Peano–Jordan measure" – user202729 Mar 06 '25 at 17:00

1 Answers1

44

The probability that you lose at every stage in the process is a product of probabilities: $$ \frac{9}{10} \cdot \frac{99}{100} \cdot \frac{999}{1000} \cdot \frac{9999}{10000} \cdots $$ or, in infinite product notation, $$\prod_{k=1}^\infty \left(1 - \frac1{10^k}\right).$$ The answer we want is $1$ minus this infinite product.

This infinite product does not have a closed form in terms of familiar functions. (Mathematica "simplifies" the infinite product for me to the $q$-Pochhammer symbol $(\frac1{10}; \frac1{10})_{\infty}$, but that's less of a simplification and more of "giving this infinite product we can't evaluate a name".)

Approximately, $1 - (\frac1{10}; \frac1{10})_{\infty} \approx 0.1099899$, so the answer of $\frac19$ is very close, even though it's not correct.


To see why the answer is close to $\frac19$, imagine that at the very start of the lottery, you were given a ticket for every stage of the lottery, and - crucially - promised that at most one of those tickets is a winning ticket. In this case, the probabilities $\frac1{10}, \frac1{100}, \frac1{1000}, \dots$ would be probabilities of disjoint events: that is, there would be no overlap between them. We could then simply add them together for a final answer of $$\frac1{10} + \frac1{100} + \frac1{1000} + \cdots = 0.111\dots = \frac19.$$ Unfortunately, the actual lottery is a bit less nice than this; there are some outcomes where multiple of these tickets could be winning tickets, so some of the wins are "wasted". This brings the probability down from $\frac19$ by a small amount.


Even though we don't have a formula for the answer better than the unhelpful "$1 - (\frac1{10}; \frac1{10})_{\infty}$", we can actually compute many decimal digits of this number easily! (Credit to Don Hatch in the comments for helping me notice this.) This comes down to a powerful result known as Euler's pentagonal theorem: the identity $$\prod_{k=1}^\infty (1 - x^n) = 1 + \sum_{k=1}^\infty (-1)^k (x^{k(3k+1)/2} + x^{k(3k-1)/2}).$$ In our case, if we set $x = \frac1{10}$, we get $$1 - \prod_{k=1}^\infty \left(1 - \frac1{10^k}\right) = \sum_{k=1}^\infty (-1)^{k+1} \Big(10^{-k(3k+1)/2} + 10^{-k(3k-1)/2}\Big).$$ If we take the first few terms of this infinite sum, it looks like $$10^{-1} + 10^{-2} - 10^{-5} - 10^{-7} + 10^{-12} + 10^{-15} - 10^{-22} - 10^{-26} + \dots$$ which are in fact adjustments to individual digits past the decimal, increasingly far apart in the decimal expansion.

It takes a bit of work to turn this into something practical, but here is the pattern for the decimal digits of the final answer. We can divide the digits into blocks of lengths $4, 7, 10, 13, 16, \dots$, so that:

  1. The first block is $1\,0\,99$.
  2. The second block is $8\,99\,0000$.
  3. The third block is $1\,000\,999999$.
  4. The fourth block is $8\,9999\,00000000$.
  5. In general, for odd $k$, the $k^{\text{th}}$ block is $1\, \underbrace{00\dots00}_k\, \underbrace{99\dots99}_{2k}$...
  6. ...and for even $k$, the $k^{\text{th}}$ block is $8\,\underbrace{99\dots99}_k\, \underbrace{00\dots00}_{2k}$.

This lets us very quickly write down many many digits of the probability of winning the multi-stage lottery!

Misha Lavrov
  • 159,700
  • 6
    Actually $\frac{1}{9}$ is the expected number of wins, assuming that we keep playing forever regardless of wins and losses. The probability of winning is slightly lower precisely because it is possible to win multiple times. – Adayah Mar 05 '25 at 09:35
  • @Adayah That's also a good way to phrase it! – Misha Lavrov Mar 05 '25 at 16:01
  • 4
    The decimal expansion is fun: 0.10998990000100099999989999000000001000009999999999899999900000000000010000000999999999999998999999990000000000000000100000000099999999999999999989999999999000000000000000000001000000000009999999999999999999999899999999999900000000000000000000000010000000000000999999999999999999999999998999999999999990000000000000000000000000000100000000000000099999999999999999999999999999989999999999999999000000000000000000000000000000001... – Don Hatch Mar 06 '25 at 02:20
  • 2
    @DonHatch And it turns out there's a nice explanation for the patterns you see there! Thank you for pointing it out; I've added more detail about it to my answer. – Misha Lavrov Mar 06 '25 at 05:56
  • @Adayah: Indeed, your point about expectation is relatively straightforward to make fully precise, as an argument that $P(\text{win}) \leq \frac{1}{9}$: the expected number of wins is $\frac{1}{9}$ by linearity of expectation and summing the geometric series, and the probably of winning at least once is certainly $\leq$ the expected number of wins. (You probably had this in mind already, but I think it’s worth pointing out explicitly.) – Peter LeFanu Lumsdaine Mar 07 '25 at 16:12