I have random sets of $N$ random 32-bit strings, where all bits are i.i.d. with $\mathbb{P}(0) = \mathbb{P}(1) = 1/2$. Define
$\ \ \ \ $weight( 32-bit x ) = number of 1 bits in x, i.e. Hamming distance to 0
$\ \ \ \ $minweight( set $X$ ) = min$_{\text{x} \in X}$ weight( x )

How does the average minweight( $X$ ) vary with $N$, the size of $X$ ?

My $N$ is small enough that it doesn't matter whether $X$ has duplicates, or not.
A back-of-the envelope estimate would be fine.

(A harder question is to analyze the algorithm for bit-string-nearest-neighbour-searching on stackoverflow, but that seems rather specialized. If anyone could suggest another forum or reference, I'd appreciate it.)

  • $\begingroup$ @cardinal, thanks. I had thought of that -- weight( 32 bits ) ~ Normal( 16 +- 2.8 ) -- but then for N = a million thought that E normal^N would be inaccurate; in fact I get 2.2 +- .5 vs. Monte-Carlo ~ 3.5 bits, dunno. A direct, intuitive way would be nice. $\endgroup$
    – denis
    Apr 20 '12 at 16:10

Given the small length of the strings of interest, providing a (computable) exact solution is straightforward.

As noted in the comments, if $X$ is a random bit-string of length 32, then the weight $$ W = \mathrm{weight}(X) \>, $$ the Hamming distance with respect to zero, is just the sum of the digits and so is binomial with parameters $n = 32$ and $p = 1/2$.

Distribution of the minimum

Now, recall that if $W_1,\ldots,W_n$ is a random sample of size $n$ from distribution $F$, then $$ \renewcommand{\Pr}{\mathbb P} \Pr\left( \min_{1\leq i \leq n} W_i > w\right) = (1 - F(w))^n \>. $$

In our case, each $W_i$ is discrete and so it suffices to consider the probability of each atom.

Let $\newcommand{\Wo}{W_{(1)}}\Wo$ denote the minimum of the sample of size $n$. Then, $\Wo \in \{0,\ldots,32\}$, and so $$ \Pr(\Wo = w) = \Pr(\Wo > w - 1) - \Pr(\Wo > w) = (1-F(w-1))^n - (1-F(w))^n \>. $$

Plugging in $F(w) = 2^{-32} \sum_{k=0}^w {32 \choose k}$ will yield the desired probabilities.

Note that this results extends easily to the case of general $n$ and $p$, as well.

A note on the expected value

The expected value can readily be computed by the standard formula $$ \mathbb E \Wo = \sum_{w=0}^{32} w \Pr(\Wo = w) \>, $$ though to get a rough estimate, we can note that $$ \mathbb E \Wo \leq 32 \Pr(\Wo > 0) = 32 (1 - 2^{-32})^n \>, $$ so the expected value converges to zero geometrically fast.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.