High-Low Hijinks Extra Credit

Your friend is playing an epic game of “high-low” and has made it incredibly far, having racked up a huge number of points. Given this information, and only this information, what is the probability that your friend wins the next round of the game?

Let's define this problem as the following probabilistic statement, what is the conditional probability that $N \geq n+1$ given that $N \geq n,$ when $n$ is large? That is, find $$\lambda = \lim_{n \to \infty} \mathbb{P} \left \{ N \geq n+1 \mid N \geq n \right\} = \lim_{n\to \infty} \frac{\mathbb{P} \{ N \geq n+1 \}}{\mathbb{P} \{ N \geq n \} } = \lim_{n\to\infty} \frac{p_{n+1}}{p_n}.$$

One method that we can do is to just compute successive values of $p_n,$ as per the following table:

$n$	$p_n$
1	$\frac{3}{4}$
2	$\frac{13}{24}$
3	$\frac{25}{64}$
4	$\frac{541}{1920}$
5	$\frac{1561}{7680}$
6	$\frac{47293}{322560}$
7	$\frac{36389}{344064}$
8	$\frac{7087261}{92897280}$
9	$\frac{34082521}{619315200}$
10	$\frac{1622632573}{40874803200}$

Even by this point, it seems that the ratio of $p_{n+1} / p_n$ approaches roughly $0.72134752\dots,$ but let's see if we can make sense of this otherwise mysterious limit.

Let's focus on the recursion $h_{n+1}(t) = \int_0^1 I(t,x) h_n(x) \,dx$ that we saw in the Classic problem. Let's assume that we have $h_{n+1} (x) \approx \lambda h_n (x),$ for large values of $n \gg 1.$ Then the recursion formula would settle down to something like $$\lambda h^*(t) = \int_0^1 I(t,x) h^*(x) \,dx = \begin{cases} \int_t^1 h^*(x)\,dx, &\text{if $t \lt \frac{1}{2};$}\\ \int_0^t h^*(x) \,dx, &\text{if $t \gt \frac{1}{2};$}\end{cases}.$$ Differentiating both sides of the equation, we get $$\lambda \frac{d}{dt} h^*(t) = \begin{cases} -h^*(t), &\text{if $t \lt \frac{1}{2};$}\\ h^*(t), &\text{if $t \gt \frac{1}{2}.$}\end{cases}.$$ Let's focus on when $t \gt \frac{1}{2},$ in which case we have $\lambda \frac{d}{dt} h^*(t) = h^*(t),$ so we would have $h^*(t) = C \exp \left( \frac{t}{\lambda} \right)$ for $t \gt \frac{1}{2},$ with $h^*(t) = h^*(1-t) = C \exp \left( \frac{ 1-t}{\lambda} \right)$ for $t \lt \frac{1}{2}.$ In this case, plugging back in, if our recursion formula still holds for $t \gt \frac{1}{2},$ we would need to have \begin{align*}\lambda C e^{t/\lambda} &= \int_0^\frac{1}{2} C \exp \left( \frac{1-t}{\lambda} \right)\,dt + \int_\frac{1}{2}^t C\exp \left(\frac{t}{\lambda}\right) \,dt \\ &= \lambda C \left( e^{1/\lambda} - e^{1/(2\lambda)} \right) + \lambda C \left( e^{t/\lambda} - e^{1/(2\lambda)} \right) \\ &= \lambda C e^{t/\lambda} + \lambda C \left( e^{1/\lambda} - 2 e^{1/(2\lambda)} \right).\end{align*} Therefore, we must have either $\lambda = 0$ or $e^{1/\lambda} - 2 e^{1/(2\lambda)} = 0$ in order tof the recursion formula to hold, which implies that the conditional probability of winning the next round in a long game of high-low is $$\lambda = \lim_{n \to \infty} \frac{p_{n+1}}{p_n} = \frac{1}{2\ln 2} \approx 0.721347520444\dots.$$