Roll Until Prime

[Chuck]

If I start with a score of zero and start rolling an ordinary six sided die, adding the rolls to my score until the score is a prime number, what is my expected average score?

That is, if I get a 2, 3, or 5 on the first roll then I'm done after one roll. If I roll 6,3,6,2 then my score is 17.

After 100 million or so computer simulations the average is hovering around 8½, in the 8.499+ to 8.500+ range. But it can't be exactly 8½, can it?

[Zag]

It definitely is not going to converge on a nice, round number. I suspect that it's just a coincidence that it is converging on something so close to one. The reason it won't is that primes don't make a predictable pattern, so there's no chance of terms that cancel away when you extend an infinite series.

I'll bet that in your sample, the highest you've ever reached is in the thousands, but there is always going to be a tiny chance that your game will extend into huge numbers. At that point, prime numbers become wildly unpredictable.

[Thok]

It's probably not anything interesting.

Out of curiosity, what's the distribution of the stopping prime? This would be a good debug check as well, since there's a small but nontrivial chance of integer overflow issues, depending on how the language you are using handles large numbers.

[Chuck]

I didn't put in a stopping point since the probability of a really high score seemed negligible. Of the 270 million simulations I ran, the highest score was 307. I didn't track the distributions. I probably should have.

[Chuck]

Rolling until I get a square gives an average of about 24.78 and a high of 3136 in 50 million simulations. That's not very interesting.

[Elethiomel]

The medians might be more interesting. At least if you want to turn it into a betting game.

[Thok]

It occurs to me that once can do an estimate for what the mean should be. You've got roughly a (5/6)^(n-1)*(1/6) chance of stopping at the nth prime (assuming because of the dice rolls you have a 1/6 chance of hitting a given number: this isn't true but it's not a horrible first estimate), and the nth prime is roughly n/ln(n). Wolfram Alpha claims that sum (5/6)^(n-1)*(n/6ln(n)) converges and estimates it as approximately a bit over 3.

If I do the same logic for n^2, Wolfram Alpha suggests the sum is around 66. (Also as long as you do something growing slower than exponential growth, that should converge.)

[Chuck]

I think I have a 2/7 chance of hitting any given number. Simulation confirms this.

[Thok]

[Chuck]

The median for primes is 5. This isn't surprising since more than half the scores will be 2, 3, or 5. The mode is also 5.

[esme]

1) You should do your simulations with dice with fewer sides.

2) The probability argument will give very good results if you do exact calculations for primes up to, say, 11 and *then* use the probability argument (with 2/7) for the rest. For example, there isn't a probability of 2/7 to hit 2, which shifts the average upwards.

3) How do you know that 24.78 is not interesting?
_________________
Mundus vult decipi, ergo decipiatur.

[Chuck]

I tried using various sized dice and pairs of dice but none of the results looked any more interesting that 24.78 which I know isn't interesting because I'm not interested in it.

I also tried using powers of two instead of primes but the program produced a wide range of averages and eventually hung up because it missed too many low powers of two and it takes a lot of die rolls to reach high ones.

[lostdummy]

[Chuck]

I didn't really think it would be exactly 8½, but that would have been interesting.

It really gets interesting when stopping on a power of 2.

[lostdummy]

It is probably hard to find setup that result in integer number, let alone power of 2 integer number ;p

Initially what I found interesting in this problem was not result, but how to get to result.

Anyway, we can consider as separate problem : " find setup (end condition, die size) that result in integer value, preferably in power of 2" ;p

[lostdummy]

And ... I found one ;p

If you use 7-sided die, and end condition is that number is divisible by 4, then average expected value is 16.

There are few similar solutions with 3-sided and 7-sided dice, and end condition is numbers divisible by 2^n . It is interesting that it only works with those dice size ( if I do not count impossible dice sizes).

*EDIT*

In fact, I find that any combination of these will result in power of 2 result:
- die size of 2^a-1 , a>=2
- end condition of 2^b, b>=1

And it appears on wiki that only real dice with '2^n-1' size are 3-sided and 7-sided.

BTW, that same end condition (divisible by X) have many combinations of die and X that result in integer value - its only 2^n results that are rare.

I also tried few other types of 'end conditions', but had no luck in even finding integer results. One that come close is "use 8-sided die, and finish when last digit in current sum is 1" - that result in average sum value of 40.0000001... not exactly integer, but close.