Probability that a Bernoulli random walk is always at least its mean?

96 Views Asked by At

Let $(X_i)_{i\le n}\sim B(p)$ be a length $n$ sequence of iid. Bernoulli random variables. That is $\Pr[X_i=1]=p$ and $\Pr[X_i=0] = 1-p$. Let $S_k = \sum_{i\le k} X_i$ be the sum of the first $k$ variables. What is the probability that each $S_k \ge p k$? That is the probability that the random walk $S_k-pk$ stays non-negative?

Clearly, the answer must be at most $p$, since the first step has to succeed. That is $X_1=1$. For $p<<1/k$, this would also be a lower bound, since we would only expect one of the $X_i$ to ever be positive.

From comparison with Brownian motion, we might expect the probability tot be around $1/\sqrt{n}$.

I wonder if one could prove that these are indeed the two cases to look at. That is the probability that for all $k\le n$, $S_k \ge pk$ is $\sim p/\sqrt{k}$. It seems like a problem that has been well studied, but I couldn't find anything in e.g. Feller except for $p=1/2$.

A very related question is requiring that $S_k \ge \lfloor pk\rfloor$. It appears that here the success probability is always at least $1/\sqrt k$. That makes sense since the above version requires $X_1=1$ and then continues very similarly to the floored version.

Update: I realized I could plot the probability for various $p$ and $k$. As @antkam says the plot is discontinuous roughly at $p=1/n$ for all integers $1<n<k$. However it seems to be pretty smooth asymptotically. It also seems like a lower bound of $C p/\sqrt{k}$ should be achievable for some constant $0<C<1$, while an upper bound would have to go through (1,1). Probability of surviving 200 steps Here thee blue line is the true probability (as calculated by dynamic programming) while the orange line is $p/\sqrt{k}$.