Imagine I have a process with $N$ distinct trials, $(t_1,t_2,t_3,t_4,...,t_N)$, where each trial $t_i$ has its own probability of success $p_i$ and probability of failure $q_i = (1-p_i)$.

After performing a single random instance of each trial, how do we calculate something like a probability distribution for the overall number of successes, or less ambitiously, the mean / median number of expected overall successes?

For example:

Let's say we perform ten distinct trials, $(t_1,t_2,...,t_10)$, a single time per trial, where the probabilities of success for each trial are $(0.3,0.23,0.94,0.1,0.1,0.44,0.1,0.4955,0.64,0.0987)$. What probability distribution do we have for the number of overall successes, or can we at least compute a mean / median?

  • $\begingroup$ The mean has a simple answer. Let $X_i=1$ if we have a success on the $i$-th trial, and $0$ otherwise. Then the number of successes is $X_1+\cdots+X_n$, so the mean is $E(X_1)+\cdots+E(X_n)$. This is $p_1+\cdots+p_n$. $\endgroup$ – André Nicolas Apr 7 '14 at 6:05
  • $\begingroup$ @AndréNicolas Ah, yup that makes sense. Is there something similar we can do for the median? $\endgroup$ – T.S. Apr 7 '14 at 6:09
  • $\begingroup$ I don't think so. The linearity of the mean is one reason for using it as the measure of (average) location. $\endgroup$ – André Nicolas Apr 7 '14 at 6:11
  • $\begingroup$ @AndréNicolas Surely there must be some known algorithm though for computing a median under these circumstances? $\endgroup$ – T.S. Apr 7 '14 at 6:12
  • $\begingroup$ Probably. But not known by me. $\endgroup$ – André Nicolas Apr 7 '14 at 6:13

The probability generating function for $Y = X_1 + \ldots + X_n$ is $$E[ z^Y] = \prod_{i=1}^n E[z^{X_i}] = \prod_{i=1}^n (q_i + p_i z) $$ Expand this and you can read off the probability mass function of $Y$.

Or you can proceed iteratively: if $Y_k = X_1 + \ldots + X_k$, then for $0 \le j \le k$, $$P(Y_k = j) = q_k P(Y_{k-1} = j) + p_k P(Y_{k-1} = j-1)$$

There is no nice expression for the median, however.


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.