The method described in the link you cited is based on Floyd's cycle finding algorithm, also known as "the tortoise and the hare" algorithm. This is a general-purpose algorithm for detecting cycles in iterated maps, which I will first describe below.
Specifically, consider the sequence $(x_i)$ defined by $x_i = H(x_{i-1})$ for some map $H$ and some initial value $x_0$. If this sequence is cyclic, then we have some integers $j > 0$ and $k > 0$ such that $x_j = x_{j+k}$, and thus also $x_i = x_{i+nk}$ for all integers $i \ge j$ and $n \ge 0$.
In particular, this holds for $i = nk$ and any integer $n$ satisfying $n \ge j/k$, yielding $x_i = x_{2i}$.
Thus, if the sequence $(x_i)$ is cyclic, there exists an integer $i > 0$ such that $x_i = x_{2i}$. Conversely, the existence of such an integer also clearly implies that the sequence must be cyclic (with a period that evenly divides $i$).
It thus follows that, to detect cycles in the sequence $(x_i)$, it suffices to check whether $x_i = x_{2i}$ for any positive integer $i$. We can do this iteratively using constant space by keeping track of two elements of the sequence, $y = x_i$ and $z = x_{2i}$ (both initialized to $x_0$ at iteration $i = 0$), and, on each successive iteration, updating them as $y \gets H(y)$ and $z \gets H(H(z))$.
If we find such an $i$, we'll know that the sequence $(x_i)$ is cyclic. Now, we have two possibilities: either the initial value $x_0$ is part of the cycle, or it is not. In the latter case, we have $x_0 \ne x_i$ but $$H^{(i)}(x_0) = x_i = x_{2i} = H^{(i)}(x_i),$$ and we've thus found a collision in the $i$-fold iterated map $H^{(i)}$, which in turn implies the existence of a collision in the underlying map $H$.
All that remains to be done is locating the underlying collision. To accomplish that, we rewind the iteration back $i$ steps, so that $y = x_0$ and $z = x_i$, and advance them one step at a time, this time as $y \gets H(y)$ and $z \gets H(z)$ so that, at iteration $j$, we'll always have $y = x_j$ and $z = x_{i+j}$. Since we know that $x_0 \ne x_i$ and $x_i = x_{2i}$, it follows that there must be some $j$ between $0$ and $i-1$ such that $x_j \ne x_{i+j}$ but $x_{j+1} = x_{i+j+1}$, and thus $H(x_j) = H(x_{i+j})$. When we find that $j$ — i.e. when we find the first $y$ and $z$ such that $H(y) = H(z)$ — we stop; that's the collision we've been looking for.
This algorithm only requires space for storing a fixed number of values: $x_0$, $y$ and $z$. How much time does it take? Well, if $j$ and $k$ are the lowest positive integers satisfying $x_j = x_{j+k}$, then Floyd's cycle finding algorithm will take $i = k \lceil j/k \rceil < k(j/k + 1) = j + k $ steps (each involving three evaluations of $H$) to detect the cycle, and then $j$ further steps (involving two evaluations of $H$) to locate the collision, for a total of up to $5j + 3k \le 5(j+k)$ evaluations of $H$.
Now, if $H$ is a random function on an $m$-element set, then, by the birthday paradox, the expected number of steps $\mathbb E[j+k]$ before the first collision is $O(\sqrt{m})$. Thus, the expected runtime of the collision-finding algorithm described above is also $O(\sqrt{m})$.
