# Binomial distribution where successes increases the number of trials

I'm trying to calculate the probability of $$k$$ successes when the number of trials ($$n$$) increases by $$6$$ each time a success occurs (starting at $$20$$):

$$P(X=k) = \ ? \qquad\text{where}\qquad p=.05, \quad n=20.$$

I can't just punch in $$n=20+6k$$ because that includes impossible scenarios like $$X>0$$ with no successes in the first $$20$$ trials. The run ends one you've reached the $$(n+6k-k)$$th loss.

Edit: Apologies for the lack of clarity. Hopefully these steps will help

1. Begin run
2. Perform $$20$$ trials
3. Perform check to determine # of successes
4. Perform $$6$$ additional trials for each success ($$2$$ successes = $$12$$ add'l trials, totaling $$32$$ trials)
5. Perform check to determine # of successes since previous check
6. Perform $$6$$ additional trials for each success since previous check
7. Repeat steps $$5$$ and $$6$$ until # of successes since previous check is $$0$$
8. End run

See below for an attempted solution and tell me why it's wrong

$$\frac{20}{20 + 6k} \binom{20 + 6k}{k} \frac{19^{20+5k}}{20^{20+6k}}$$

• The rules aren't clear. Are you saying that you start with $n=20$ trials, but if you get $k$ successes amongst those you then get an additional $6k$ trials? Does the $+6$ property work for those extra ones as well?
– lulu
Commented Dec 17, 2023 at 22:24
• Regardless of the details, I suggest simulating this. With $p=.05$, it's unlikely you'll get more than a couple of successes so the process ought to terminate fairly quickly. But numerical results would be good to see.
– lulu
Commented Dec 17, 2023 at 22:25
• I suspect you can get some mileage out of treating each die roll (I'm going to speak of rolling a d$20$ just to maintain intuition) as a separate trial, and each success (a roll of $20$) yields six additional rolls in the same trial. See if you can create a generating function for a single trial, and then the final result can be obtained by raising that to the $20$th power. I'll think about it some more. Commented Dec 18, 2023 at 17:51
• It turns out that that each success spawning as many as six further trials creates a problem: an unsolvable sextic equation in $T(z)$ (the generating function for a single trial). So this approach leads to an algorithm for generating low-degree terms, but no analytic expression, even in the transform domain. It could be done for fewer spawn (certainly two, maybe three or four), but not six, at least as far as I can tell. Commented Dec 18, 2023 at 18:51
• @BrianTung Basically, you can use Lagrange inversion to get a closed form for the coefficients of the inverse function of the sextic equation without actually solving the equation. This leads to a simple expression for $P(X=k)$, given in my answer (though I didn't use your approach; I shortcutted all of that by using Bertrand's ballot theorem). Commented Dec 18, 2023 at 20:59

Answer: $$P(X=k)=\frac{n}{n+6k}\binom{n+6k}{k}p^k(1-p)^{n+5k}$$

This is exactly what you guessed in your answer (upon setting $$n=20$$ and $$p=1/20$$).

Proof: Here is an equivalent phrasing of the problem. You maintain a counter, $$N$$, representing the number of trials remaining at any point of the process. Initially, $$N=n$$. Then,

• Each time there is a failure, decrease $$N$$ by one.

• Each time there is a success, increase $$N$$ by five. Each success grants six additional trials, and we subtract one from six to account for the trial containing the success.

This continues until we hit $$N=0$$, at which point the process stops. Please convince yourself that this is equivalent to list of steps you stated.

You want to find the probability of exactly $$k$$ successes. This implies that there were $$n + 5k$$ failures. To do this, we count up the number of valid sequences of length $$n+6k$$ which consist of $$k$$ successes and $$n+5k$$ failures, and multiply this number by the probability of each sequence occurring. The probability of such a sequence is $$p^{k}(1-p)^{n+5k}.$$ All that remains is to count the number of valid sequences. A sequence is valid if and only if it consists of $$k$$ successes and $$n+5k$$ failures, and if the running total of trials remaining, $$N$$, never reaches zero until the very end. In other words, for each $$t\in \{0,1,\dots,n+6k-1\}$$, if we let $$k_t$$ be the number of successes up to time $$t$$, then it must be true that $$n+5k_t - (t-k_t)>0 \qquad \text{for all t\in \{0,1,\dots,n+6k-1\}}.$$ Let us rearrange this inequality as follows: $$n+5k-(t-k_t)>5(k-k_t)$$ If we interpret this inequality in words, we are saying that for all $$t$$, the number of failures AFTER time $$t$$ is more than $$5$$ times the number of successes AFTER time $$t$$. Note that $$n+5k-(t-k_t)$$ is the number of failures after time $$t$$, and $$t-k_t$$ is the number of successes after time $$t$$.

Therefore, if we reverse a valid sequence, we see that each initial segment of the reversed sequence has more than five times as many failures as successes. This shows that counting valid sequences is exactly solved by the following generalization of Bertrand's ballot problem.

Theorem Let $$r\ge 1$$ be an integer. Suppose in an election, candidate $$A$$ receives $$a$$ votes, and candidate $$B$$ receives $$b$$ votes, where $$a\ge rb$$. The number of ways to arrange these $$a+b$$ ballots in a sequence such that the running number of votes for $$A$$ is more than $$r$$ times the running number of votes for $$B$$ at all times is $$\frac{a-rb}{a+b}\binom{a+b}{a}.$$

Applying this theorem with $$a=n+5k$$, $$b=k$$, and $$r=5$$, we see that the number of valid sequences is $$\frac{n}{n+6k} \binom{n+6k}{k}$$. Therefore, multiplying this by the probability of each sequence, $$p^k(1-p)^{n+5k}$$, we arrive at the answer advertised at the beginning.

The following source contains four proofs of the theorem above, along with some information about the history of the problem:

Renault, Marc. “Four Proofs of the Ballot Theorem.” Mathematics Magazine, vol. 80, no. 5, 2007, pp. 345–52. JSTOR, http://www.jstor.org/stable/27643060.

• (+1) This is great to know! I took back my skepticism. Just for fun, I also verified that your result matches numerical computation and correctly demonstrates the phase-transition of the probability $\mathbf{P}(X<\infty)$ as a function of $p$. Commented Dec 19, 2023 at 4:58
• It is always a pleasure to read your answers. Thank you for the detailed description and the reference to the Renault paper.
– user
Commented Dec 19, 2023 at 20:23

Let $$g(m,k)$$ be the probability of $$k$$ successes after $$m$$ sets of six, without yet having reached the end. I think the following recursion holds. $$g(0,k)={20\choose k}p^k(1-p)^{20-k}\text{ if }k\gt0\\ g(m,k)=0 \text{ if }k\le m\\ g(m,k)=\sum_{i=0}^6g(m-1,k-i){6\choose i}p^i(1-p)^{6-i}\text{ if }k\gt m$$

Then $$Pr(X=k)=g(k-1,k)(1-p)^6$$.
The recursion is easy enough to calculate for small $$k$$ and specific $$p$$, but I can't see how to get a general formula.

Thank you to the person who pointed out the flaw in my first approach. After some thinking, I believe this functions.

Lets try to make all of our trials the same, so that this can be a random walk. Then, we set $$X_i \thicksim Bin(2,0.05)$$, which is a set of 2 trials. Therefore, no matter what, we want at least 5 $$X_i$$. Now, we want this to be a random walk that counts how many trials we get to until we hit 0 remaining. So, if we start at 20, and let $$Y_i$$ be the the $$i^{th}$$ step on our walk. Then, The p.m.f. of $$Y_i$$ would be: $$P_{Y_i} = \begin{cases} -2 & \text{if X_i = 0} \\ 6 & \text{if X_i = 1} \\ 12 & \text{if X_i = 2} \end{cases}$$ IF we add all of the $$Y_i$$, we would find that it models the number of trials we have remaining. To find the $$\mathbb{P}(X = k)$$, we would want $$\mathbb{P}(\sum_{i=1}^{5+3k}Y_i+20 = 0)$$ Under the condition that our sum never crosses below $$0$$.

• You are right! I misunderstood the problem and assumed we were guaranteed $20+6k$ trials, when it looks like we need less than that. Commented Dec 17, 2023 at 23:44

I interpret the problem to intend that (for example) in order to have $$~k = 2~$$ successes (in what would have to be $$~32~$$ trials), you must have at least one success in the first $$~20~$$ trials, and at least two successes in the first $$~26~$$ trials.

So, when $$~k = 2,~$$ trials 27 through 32, inclusive must all fail.

$$\text{Let} ~~f(n,r) ~~\text{denote} ~~\binom{n}{r}(0.05)^r(0.95)^{n-r}.$$

In order to explicitly compute the probability of exactly $$~k = 2~$$ successes, you first reserve the factor $$~f(6,0)~$$ which refers to the necessary failures on trials 27 through 32, inclusive. Then, either

• Compute $$~f(26,2)~$$ and subtract from this $$~[f(20,0) \times f(6,2)].$$
This complementary approach signifies that both success must occur in the first $$~26~$$ trials but you must reject the situation where both successes occurred after trial 20.

• Compute $$~[f(20,2) \times f(6,0)] + [f(20,1) \times f(6,1)].$$
This direct approach signifies that either both successes occurred on or before trial 20, or exactly one of the two successes occurred on or before trial 20.

I used $$~k = 2~$$ as an example to facilitate describing the gymnastics that will be required, for larger values of $$~k,~$$ regardless of whether you use the complementary or direct approach.

For example, consider $$~k = 5.~$$ Then, there will be $$~20 + (6 \times 5) = 50~$$ trials, and you must have:

• No more than $$~5~$$ successes ever.

• At least 5 successes on or before trial 44.

• At least 4 successes on or before trial 38.

• At least 3 successes on or before trial 32.

• At least 2 successes on or before trial 26.

• At least 1 success on or before trial 20.

Edit
I also consider recursion to be no walk in the park here. For example, let $$~P(k)~$$ denote the probability of exactly $$~k~$$ successes, and consider trying to use the computation of $$~P(2)~$$ to compute $$~P(3).~$$

The constraints on $$~P(3)~$$ are very similar to the constraints on $$~P(2).~$$ However, the change to the constraints are:

• The maximum number of allowable successes is $$~3,~$$ rather than $$~2.$$

• You have the added constraint that you must have at least $$~3~$$ successes on or before move $$~32.$$

The difficulty in attempting to use $$~P(2)~$$ to assist in computing $$~P(3)~$$ is that $$~P(2)~$$ (in effect) has the constraint that you must have exactly $$~2~$$ successes on or before trial 26, while $$~P(3)~$$ allows the third success to occur anywhere in the range of trials 1 through 32.

• More specifically, It should be at least 1 and at most 5 success on or before trials 20, and 0 successes from trials 21-44, and so on. Commented Dec 17, 2023 at 23:45
• @RobertMurray Good catch. I edited my answer. Commented Dec 17, 2023 at 23:47