# Pseudorandom Generator

In a cryptographic application, two types of (pseudo)random bit streams are needed:

• a stream $$A= a_{1}a_{2}a_{3}\ldots$$ in which $$\Pr[a_{i}=0]=\Pr[a_{i}=1]= 1/2\ \forall i$$ and
• a stream $$B= b_{1}b_{2}b_{3}\ldots$$ in which $$\Pr[b_{i}=0]=2/3; \Pr[b_{i}=1]=1/3\ \forall i$$.

Propose the following constructions:

1. Given a generator $$G_{A}$$ for $$A$$, propose an efficient construction that uses $$G_{A}$$ to generate $$B$$.
2. Given a generator $$G_{B}$$ for $$B$$, propose an efficient construction that uses $$G_{B}$$ to generate $$A$$.

I don't need the solution to this question. I want to know, how to think about this problem to get a solution. I know basic definition of Pseudorandom generator.

• Hint, working for both 1 and 2: devise a method that groups bits of what's available, and makes it's best to output what's needed (as a fallback, output nothing)
– fgrieu
Oct 9, 2020 at 6:37
• There is a nice problem about generating random from biased dice. It has also at least two solution in our site. Oct 9, 2020 at 9:29
• Is it still work if I say output $b_{i}=0$ if 0 or 10 comes and output $b_{i}=1$ if 11 comes ??? @ fgrieu Nov 11, 2020 at 6:58

The hints given refer to unbiasing algorithms, which is the standard answer in cryptography. However, there is a very nice and different solution for the example of generating the $$(2/3,1/3)$$ distributed bits from an unbiased stream which I will mention here. For clarity call the output symbols $$a$$ and $$b$$ with $$p_a=2/3=1-p_b$$ instead of using binary symbols

Expand the probabilities in base 2 to get: $$\frac{2}{3}=0.101010101010\ldots$$ and $$\frac{1}{3}=0.0101010101\ldots$$ which means these probability atoms have dyadic expansions $$\frac{2}{3}=\frac{1}{2}+\frac{1}{8}+\frac{1}{32}+\cdots=\frac{1/2}{1-{1/4}}$$ and $$\frac{1}{3}=\frac{1}{4}+\frac{1}{16}+\frac{1}{64}+\cdots=\frac{1/4}{1-{1/4}}$$ This decomposition yields the following infinite tree which generates the symbols required:

if the unbiased bit is 0 go left, otherwise go right, when an $$a$$ or $$b$$ is emitted start again;

(the first left branch should have the symbol $$a$$).

In Chapter 5 of Cover & Thomas' Elements of Information Theory book it is proved that this procedure is optimal, i.e., it gives a tree of expected minimum length generating this distribution.

Edit: As in the comments by @supercat, if $$p$$ is unknown but the input bits are independent, one can group say $$3-$$tuples into two sets of desired probability ratio, and also try to maximize the probability that a bit is actually output. $$k=3$$ is convenient for this case since the binomial coefficients $$\binom{3}{j}$$ are divisible by 3, if we ignore the first and last coefficient, so the grouping to obtain a $$(1/3,2/3)$$ probability [when symbols are output] becomes possible. Concretely

If you see 100 or 011, output $$b$$. If you see 010, 001, 101, or 110, output $$a.$$ If you see 000 or 111, throw out the bits and try again.

• Another interesting variation is to design a construction that will work for any source which generates bits that are statistically independent, but whose bias may be unknown. To generate a 2/3-1/3 distribution from that, one could take trios of input bits. If 100 or 011, output a. If 010, 001, 101, or 110, output b. If 000 or 111, throw out the bits and try again. Not an optimal approach, but simple to understand, and it will work with any level of bias in the source. Oct 9, 2020 at 23:07
• @supercat, of course, that is the standard approach I alluded to in my answer. Oct 9, 2020 at 23:43