$BPL$ with polylog random bits is in $L$

Question

Consider a $BPL$ machine (namely, a probabilistic algorithm that uses logspace and polynomially many random bits). It is known (Saks-Zhou) that $BPL \subseteq DSPACE(log^{1.5}(n))$.

My question is about a $BPL$ machine that uses only polylog many bits of randomness. In one of Goldreich's papers, it is mentioned in passing that a language decided by such a $BPL$ machine is actually in deterministic logspace $L$. But I am unable to find any explanations of this remark anywhere.

Why can it be derandomized completely in logspace?

Answer 1

It follows from this PRG of Nisan and Zuckerman. This paper shows that if you have an algorithm that uses space $S$ and only $\mathrm{poly}(S)$ random bits, then the number of random bits can be decreased to $O(S)$.

In particular, in the setting you describe, we have $S = O(\log n)$, so the number of random bits can be reduced to $O(\log n)$. Then, we can derandomize the algorithm completely by enumerating over all possible coin tosses.