Good rational approximations of pi

Golfscript, 71 70 69 chars

2\!:^2^..292^15.2/3]{(.)2/.9>+{\+.((}*;.}do;;]-1%{^0@{2$*+\}/"/"\n}/;

(Assumes that you don't pass it anything on stdin)

I don't want to hear any more whinging by people who don't have built-in constants for pi. I don't even have floating point numbers!

See http://en.wikipedia.org/wiki/Continued_fraction#Best_rational_approximations for the background.

# No input, so the stack contains ""
2\!:^2^..292^15.2/3]
# ^ is used to store 1 because that saves a char by allowing the elimination of whitespace
# Otherwise straightforward: stack now contains [2 1 2 1 1 1 292 1 15 7 3]
# Pi as a continued fraction is 3+1/(7+1/(15+1/(...)))
# If you reverse the array now on the stack you get the first 10 continuants followed by 2
# (rather than 3)
# That's a little hack to avoid passing the denominator 1000000

{
    # Stack holds: ... [c_n c_{n-1} ... c_0]
    (.)2/.9>+
    # Stack holds ... [c_{n-1} ... c_0] c_n (1+c_n)/2+((1+c_n)/2 > 9 ? 1 : 0)
    # (1+c_n)/2 > 9 is an ad-hoc approximation of the "half rule"
    # which works in this case but not in general
    # Let k = (1+c_n)/2+((1+c_n)/2 > 9 ? 1 : 0)
    # We execute the next block k times
    {
        # ... [c_{n-1} ... c_0] z
        \+.((
        # ... [z c_{n-1} ... c_0] [c_{n-1} ... c_0] z-1
    }*
    # So we now have ... [c_n c_{n-1} ... c_0] [(c_n)-1 c_{n-1} ... c_0] ...
    #                    [(c_n)-k+1 c_{n-1} ... c_0] [c_{n-1} ... c_0] c_n-k
    ;
    # Go round the loop until the array runs out
    .
}do

# Stack now contains all the solutions as CFs in reverse order, plus two surplus:
# [2 1 2 1 1 1 292 1 15 7 3] [1 2 1 1 1 292 1 15 7 3] ... [6 3] [5 3] [4 3] [3] [2] []
# Ditch the two surplus ones, bundle everything up in an array, and reverse it
;;]-1%

# For each CF...
{
    # Stack holds ... [(c_n)-j c_{n-1} ... c_0]
    # We now need to convert the CF into a rational in canonical form
    # We unwind from the inside out starting with (c_n)-j + 1/infinity,
    # representing infinity as 1/0
    ^0@
    # ... 1 0 [c_n-j c_{n-1} ... c_0]
    # Loop over the terms of the CF
    {
        # ... numerator denominator term-of-CF
        2$*+\
        # ... (term-of-CF * numerator + denominator) numerator
    }/

    # Presentation
    "/"\n
    # ... numerator "/" denominator newline
}/

# Pop that final newline to avoid a trailing blank line which isn't in the spec
;

Mathematica, 67 63

This isn't going to be fast, but I believe it is technically correct.

Round[π,1/Range@1*^6]//.x_:>First/@Split[x,#2≥#&@@Abs[π-{##}]&]

Round[π, x] gives the closest fraction to π in steps of x. This is "listable" so Round[π,1/Range@1*^6] does this for all fractions down to 1/10^6 in order. The resulting list with many "bad" rational approximations is then repeatedly (//.) processed by removing any elements which are farther from π than the preceding one.

Perl, 77 chars

$e=$p=atan2 0,-1;($f=abs$p-($==$p*$_+.5)/$_)<$e&&($e=$f,say"$=/$_")for 1..1e6

A minor challenge is that Perl doesn't have a built-in π constant available, so I first had to calculate it as atan2(0,-1). I'm sure this will be beaten by languages more suited for the job, but it's not bad for a language mainly designed for text processing.