Consider the following experiment.

We throw a three-sided die with sides $1$, $2$ and $3$ infinitely many times. Let $T_i$ denote the outcome of the $i$'th throw. Define $N:=\min\{i:T_i\neq1\}$. Let $X$ be the event that $T_N=2$ and let $Y$ be the event that $T_N=3$.

Some calculation (*) leads to the result that $\mathbb{E}(N)=\mathbb{E}(N|X)=\mathbb{E}(N|Y)=3/2$.

Let $Z$ be the event that $T_i\neq3$ for all $i$. Some calculation (**) leads to the result that $\mathbb{E}(N|Z)=2$.

I find it very unintuitive that $\mathbb{E}(N|X)\neq\mathbb{E}(N|Z)$. Obviously we have $Z\subsetneq X$. However, the information $Z$ gives, which $X$ does not give, intuitively only affects what comes after the $N$'th throw. So how is it possible that the probability distribution of $N$ is different when conditioning on $X$ or $Z$?

(*) We have $\mathbb{P}(X)=\mathbb{P}(Y)$ and $\mathbb{E}(N|X)=\mathbb{E}(N|Y)$ by symmetry. Also notice that $X$ and $Y$ partition the event space, so $\mathbb{P}(X)+\mathbb{P}(Y)=1$, so $\mathbb{P}(X)=\mathbb{P}(Y)=\frac12$. Since $\mathbb{P}(T_i\neq1)=2/3$, we have $\mathbb{E}(N)=3/2$. By the principle of divide and conquer, we have $\mathbb{E}(N)=\mathbb{P}(X)\mathbb{E}(N|X)+\mathbb{P}(Y)\mathbb{E}(N|Y)$, so we find $\mathbb{E}(N)=\mathbb{E}(N|X)=\mathbb{E}(N|Y)=3/2$.

(**) We have $\mathbb{P}(T_i\neq1|Z)=1/2$, so $\mathbb{E}(N|Z)=2/1=2$.

By the way, if you have a suggestion for a better, more specific title, be my guest. I could not come up with a good descriptive title for this very specific question.

  • $\begingroup$ do you mean $Z \subsetneq X$? $\endgroup$ – mathworker21 Aug 31 '19 at 3:47
  • $\begingroup$ Correct, I fixed it $\endgroup$ – SmileyCraft Sep 1 '19 at 15:45

In all honesty, I am confused why you are bringing $X$ into the situation. Why aren't you just saying "The information $Z$ gives only affects what comes after the $N$'th throw, so why isn't $E(N) = E(N | Z)$?" The issue here (as well as in your actual question) is that you have the quantifiers/chronology backwards; $N$ isn't determined yet.

You are of course correct that once we know $N$, then $X$ and $Z$ provide the same information up to the $N^{th}$ throw. However, being in the world of $Z$ significantly affects the (expected) value of $N$. View $E(N | Z)$ as saying "I guarantee you that 3 won't be rolled; how long will you have to wait on average to see a 2?", the (partial) answer to which is obviously "longer than I would have to wait to see a 2 or 3 with no other conditions". I.e., rolling a non-1 on a two-sided die takes longer than rolling a non-1 on a 3-sided die.

To see it more clearly, start from scratch. Consider only a 2-sided die with sides 1,2. Let $T_i$ be the outcome of the $i^{th}$ throw and $N = \min\{i : T_i \not = 1\}$. Let $Z$ be the event that $T_i \not = 2$ for all $i$ (or, if you want to make $Z$ have positive measure, define $Z$ to be the event that $T_i \not = 2$ for $i = 1,2,\dots,10^{100}$). Then $E(N) = 2$ while $E(N | Z) = \infty$ (or some very very large number).

The moral: you shouldn't look at the information provided by an event that is conditioned upon once the random variable has been realized; you must look at how the conditioning first influences the random variable.


I think the main problem is that $P(Z)=0$, that means $Z$ almost surely does not happen. Arguing about expected values under conditons that almost surely do not happen are bound to be counterintutive, they are roughly equivalent to $\frac00$ limit forms in calculus, as both $P(N \cap Z)=0$ and $P(Z)=0$.

  • $\begingroup$ In this case $Z$ basically only reduces the three sided dice to a two sided dice. Not too bad imo. You can also define $Z_n$ as $T_i\neq1$ for all $i\leq n$. Then for any event $\omega$ we have $\lim\mathbb{P}(\omega|Z_n)=\mathbb{P}(\omega|Z)$. This way you can avoid conditioning on an event of zero probability. $\endgroup$ – SmileyCraft Jan 28 '19 at 23:23

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.