The Mystery of 355/113

http://davidbau.com/archives/2010/03/14/the_mystery_of_355113.html

by ColinWright • 14 years ago

192 65 14 years ago

fredley2• 14 years ago

I wrote a quick script to walk through Q, spitting out fractions with high "quality". Here are some early values:

  22/7: 3.42928833728
  355/113: 3.20195874233
  710/226: 2.79251047296

A quick search through 1/1 - 100,000/100,000 shows no other interesting values. My algorithm is extremely naive though (brute force), a faster search could potentially be used to find other high quality values much more quickly.

Update - A slightly optimized algorithm, and nothing special found with denominators up to the hundreds of millions. I know this means nothing wrt actual mathematics though...

Jtsummers4• 14 years ago

By your output you're performing some redundant checks. 355/113 == 710/226. Not sure how you're performing the search, but removing those may speed things up.

fredley2• 14 years ago

I think the cost of removing them (checking every value to see if it's a multiple of another) is greater than the benefit of removing them. My search terminated though (with no interesting finds) - my fractional values were getting too small and I got a ZeroDivisionError.

Someone2• 14 years ago

http://en.wikipedia.org/wiki/Continued_fraction#Best_rationa... will teach you how to find the best approximations in no time.

If you understand that, read http://blog.wolfram.com/2011/06/30/all-rational-approximatio... :-)

daurnimator2• 14 years ago

I had a quick go too:

    num     den     difference to pi
    1	1	2.1415926535898
    3	1	0.14159265358979
    22	7	0.0012644892673497
    333	106	8.3219627529107e-05
    355	113	2.6676418940497e-07
    103993	33102	5.7789062424263e-10
    104348	33215	3.3162805834763e-10
    208341	66317	1.2235634727631e-10
    312689	99532	2.914335439641e-11
    833719	265381	8.7152507433075e-12
    1146408	364913	1.6107115641262e-12
    4272943	1360120	4.0412118096356e-13
    5419351	1725033	2.2204460492503e-14
    80143857	25510582	4.4408920985006e-16

Lua script to generate (dirty):

    local max = 100000000
    local c = {}
    local min , j = math.huge
    for i=1,max do
        local a = math.pi-math.abs(math.cos(i))
        if a < min then
            min = a
            j = i
            table.insert ( c , j )
        end
    end

    local min , j = math.huge
    for k=1,#c do
        for i=1,max do
            local a = math.abs(math.pi-c[k]/i)
            if a < min then
                min = a
                j = i
                print(c[k],j,min)
            end
        end
    end

ColinWright6• 14 years ago

It's not actually the difference to pi that matters, it's the "quality" of the approximation, as described in the article. You can always get better approximations by using larger denominators, it's just that some approximations are, given the size of the denominator, unreasonably good.

355/113 is one of those.

drostie2• 14 years ago

The numbers that you're producing can be better produced by the following Python code:

    from fractions import Fraction as frac
    def continued_fraction(x, n=10):
        """Represent x as a continued fraction out to n places."""
        last = int(x)
        out = []
        for i in range(n):
            x = 1/(x - last)
            last = int(x)
            out.append(last)
        return out

    def bras(x, n=10):
        """Compute the best rational approximations to x."""
        base = int(x)
        nums = continued_fraction(x, n)
        S = lambda depth, i = 0: \
            frac(1, nums[i]) if depth == 0 else 1 / (nums[i] + S(depth - 1, i + 1))
        return ', '.join(str(x) for x in (base,) + tuple(base + S(k) for k in range(n)))

    from decimal import Decimal
    print(bras(Decimal('3.1415926535897932384626433832795028841971693993')))
    # 3, 22/7, 333/106, 355/113, 103993/33102, 104348/33215, 208341/66317,
    # 312689/99532, 833719/265381, 1146408/364913, 4272943/1360120

These are the best rational approximations to pi, which are determined by the continued fraction of pi:

    pi ~=  3 + 1/(7 + 1/(15 + ...))
    pi = 3 + [7, 15, 1, 292, 1, 1, 1, 2, 1, 3, 1, 14, 2, 1, 1, 2, 2, 2, 2, ...]

Large numbers are opportunities to "grab a lot of extra digits" and thus the large qualities come when we truncate after 7, 15, 292, and 14.

You may wonder what the "most irrational" number is, in the sense that all of its best-rational-approximations are of low quality. That distinction belongs to:

    phi = (1 + Decimal(5).sqrt()) / 2 # golden ratio
    continued_fraction(phi) # [1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
    bras(phi) # 1, 2, 3/2, 5/3, 8/5, 13/8, 21/13, 34/21, 55/34, 89/55, 144/89

Those last numbers you may recognize from your intro to programming as the Fibonaccis.

KingLeo5• 14 years ago

I think that code must be doing something wrong. 355/113 = 710/226 but you have different values for both

darkstalker• 14 years ago

355/113 has good enough precision to aproximate pi on single precision floating point

ajross• 14 years ago

Not quite, but close. I get the delta as one part in 2^21.8. The mantissa of a IEEE float is 24 bits (including the leading one).

cperciva2• 14 years ago

FWIW, all of the convergents from the continued fraction expansion of a number have "quality" greater than 2. Conversely, the Thue–Siegel–Roth theorem says that any algebraic number and any epsilon > 0, the algebraic number has only finitely many approximations with quality more than 2 + epsilon; so the "quality" values of successive best approximations are a series which converges to 2 from above.

This was how the first known transcendental numbers -- the Liouville numbers -- were constructed: They have approximations with unbounded quality, thus they cannot be algebraic. In practice, however, this isn't a very useful method: "Almost all" transcendental numbers also obey the Thue–Siegel–Roth theorem.

michaelf• 14 years ago

Is it also the case that all transcendental numbers have approximations of unbounded quality? Or only that Liouville numbers have approximations of unbounded quality, and therefore cannot be algebraic?

cperciva2• 14 years ago

Liouville numbers have approximations of unbounded quality and thus cannot be algebraic. "Most" transcendental numbers only have finitely many approximations of quality higher than 2 + epsilon for any epsilon.

Xcelerate2• 14 years ago

Pi has the continued fraction expansion:

3; 7, 15, 1, 292, 1, ...

That 355/113 is one of the best possible rational approximations to PI, and it occurs right after the 292. In general, a best rational approximation cut off at a very large term in the CF expansion will provide more accuracy than you would think.

walrus• 14 years ago

355/113 is actually [3,7,15,1]. The reason it's particularly good compared to its length is because it falls right before the 292 term (since 1/292 is a much smaller contribution than 1/1).

Xcelerate2• 14 years ago

Ah, indeed you are correct. I had that wrong.

caf• 14 years ago

So, as noted in one of the comments to the original article, the question of why 355/113 is particularly good can be restated as a question of why the 292 term in that expansion is so large.

coreygoodie• 14 years ago

Okay, I'll bite - why does this matter?

ColinWright6• 14 years ago

http://www.smbc-comics.com/index.php?db=comics&id=2674

ufo• 14 years ago

You are kind of stretching though. Why do we have to assume that 355/113 is "magical" in order to advance mathematics:

cheshirecat• 14 years ago

This is called Diophantine approximation.

http://en.wikipedia.org/wiki/Diophantine_approximation

p.s. Take a look at the "Liouville's result" part.

rplnt• 14 years ago

> and it causes various oddities elsewhere in math.

Could someone elaborate? First, explain the one example from the article (cos(355)) and perhaps show some other oddities?

_delirium• 14 years ago

If p/q is a close approximation to π, then cos(p), taken in radians, will be close to +1 or -1, because it implies that p is close to an integer multiple of π, and therefore cos(p) is close to the cosine of an integer multiple of π:

   p/q =~ π 
   p =~ qπ
   cos(p) =~ cos(qπ)

In turn, the cosine of an integer multiple of π is +1 or -1 because the period of cosine (as a trigonometric function) is 2π radians, with maxima at 0, 2π, 4π, ... and minima at π, 3π, 5π, ...

StefanKarpinski• 14 years ago

There is no mystery here. If you pick a random real number and find the best quality approximation of it — as calculated by log_b 1/|x-a/b| — then the median quality of the best such approximations within a tolerance of 1/1024 is around 3.19. So with respect to the measure this article is talking about, pi is completely normal and 355/113 is a completely unexceptional best approximation with a denominator of at most nine binary digits (about 3 decimal digits). 355/113 isn't even the best quality approximation of pi by the given criterion — that would be 22/7, which, as mentioned in the article, has a quality measure of 3.429288337281781.

Julia code to figure this out can be found here:

https://gist.github.com/3170899

fragsworth• 14 years ago

These kinds of articles about pi, or numbers, where they try to summon up some mystery and wonder around it, come across to me as not very genuine. I get the impression it gets upvoted because people want to feel like they're smarter, or more geeky and hip for having looked at it.

ChuckMcM• 14 years ago

I don't know, I've met folks for whom math is so magical they do nothing but talk about it all the time. I can recognize that passion in other pursuits as well, from knitting to physics.

I find a close correlation between people who play with numbers to people who like to solve puzzles.

shin_lao• 14 years ago

Or perhaps math are fun and people upvote a post they liked.

raverbashing3• 14 years ago

One of the most "mind boggling" areas of math is number theory

Math dealing with real/complex numbers has been vastly explored, and there are very powerful tools, and computers are very good at these kind of problems (Calculus, etc)

With number theory, progress is much harder, things work in a completely different logic (think for example that 5+3 can be 0 for example)

One of the places where these mysterious sides of math comes together is this: http://en.wikipedia.org/wiki/Riemann_zeta_function

sampo2• 14 years ago

Maybe there is a little mystery.

Enumerate all the fractions with at most 3-digit denominator (0, 1, 1/2, 1/3, 2/3, 1/4, ..., 998/999) between 0 and 1. Now what is the probability that amongst them we find an approximation to a random real that is at least as good as 355/113 is to pi?

Seems to be about 15%, so not much mystery here. But we allow denominators as large as 999 because we use the decimal system. In hexadecimal it would be at most 2-digit denominators, up to ff = 255.

In this case, the probability is about 1%. Much more mystery. So it comes down to, how large a set of denominators we think 113 is drawn from, and whether living in a 1 in 100 universe is enough to be mysterious in this case.

Sampled by this python code (takes minutes to run): https://gist.github.com/3172978

zeroonetwothree• 14 years ago

And in base-128 it would be only denominators up to 127, which is an even more amazing 0.2% ;)

Jtsummers4• 14 years ago

So just a note, you've got a great deal of repetition in your search. 0.5, for instance, turns up once for every even denominator. That's going to skew your percentages a bit.

CoreDumpling• 14 years ago

In Chinese the fraction 355/113 has been known as the 密率 ("detailed ratio") and 22/7 as the 约率 ("approximate ratio"). It is documented in the Book of Sui, volume 16:

宋末，南徐州從事史祖沖之，更開密法，以圓徑一億為一丈，圓周盈數三丈一尺四寸一分五厘九毫二秒七忽，朒數三丈一尺四寸一分五厘九毫二秒六忽，正數在盈朒二限之間。密率，圓徑一百一十三，圓周三百五十五。約率，圓徑七，週二十二。 [1]

While it suggests that the value was calculated around the end of the Liu Song Dynasty (~479 CE) by Zu Chongzhi to have an upper bound of 3.1415927 and lower bound of 3.1415926, the 密法 method is not explained.

It has been thought that Zu may have used a "method of averaging days" to calculate an intermediate value between two prior known approximations of 22/7 and 157/50 [2]. Not a very satisfying result for anyone who wants to find something special in the numbers 355 and 113.

[1] https://zh.wikisource.org/wiki/%E9%9A%8B%E6%9B%B8/%E5%8D%B71...

[2] http://books.google.com/books?id=QlbzjN_5pDoC&lpg=PA34&#...

criveros• 14 years ago

I just saw all the numbers and bailed.

K2h3• 14 years ago

if one wanted to find the 'best' converging sequence, you might go about it by brute force. Seems lots of people beat me to it. http://pastebin.com/iujCgUSu

  my $lasterror = 3;
  for (1..10000){ my $d = $_;
      for (3..40000){ my $n = $_;
         if (abs(($n/$d) - pi) < $lasterror){
         $lasterror = abs(($n/$d) - pi);
         print "\n $n / $d: error $lasterror";}}}
  
  #  output
  # 3 / 1: error 0.141592653589793
  # 13 / 4: error 0.108407346410207
  ...
  # 355 / 113: error 2.66764189404967e-007

eridius2• 14 years ago

That's not a converging series. That's just a sequence of ratios that have successively smaller errors. But there's nothing to relate each ration to the next one.

K2h3• 14 years ago

correct - one of observations in the article was that many of the series identified so far lack the magical 355/113 - I was curious what a converging series would look like, not the equation that describes such a series.

I would be interested in identifying this next super series in terms of an equation but given the few minutes on break, and that total it is unlikely someone as bad at math, such as myself, could come up with this series in a few minutes - I opted for the fun programming route.

bikenaga• 14 years ago

355/113 is a convergent of the continued fraction for pi. The convergents of a continued fraction (for an irrational number) give the best rational approximations to the number. I'm teaching this stuff this summer; here's a link to my notes on this:

http://www.millersville.edu/~bikenaga/number-theory/approxim...

For instance, 278/125 is the best rational approximation to the cube root of 11 having denominator less than 155. The proof is pretty easy --- you don't have to do it by trial and error.

As for why 355/113 doesn't come up in other expressions for pi --- I'm not sure why it should. The fact that two expressions converge to pi is no reason that I can see to expect that they should involve the same integer coefficients, say. There are infinitely many integers. There are infinitely many expressions for pi.

emmelaich• 14 years ago

Very interesting stuff.

Lovers of pi might like Eve Andersson's Pi Land: http://www.eveandersson.com/pi/

As for sequences, I particularly liked pimasterfromhk's.

peteri• 14 years ago

Very useful when performing integer mathematics... First met it in Brodies Starting Forth I always remember it as 113355 (split it in half)

Having gone away to check my memory I notice that he also provides: 1068966896 / 340262731 which is suppose to be accurate to E-18

lkbm• 14 years ago

I've never been compelled to use any fractional approximation of pi. Either I'm doing it in my head and pi is ~3, I'm doing it on paper and pi is ~3.14, or I'm using a calculator and pi is as many digits as I feel like.

We already have the decimal representation memorized for more digits than this approximates, so unless your diameter is given to you in 113ths, is there any practical application here?

abc_lisper• 14 years ago

Try fourth root of (9^2 + 19^2/22)

The Mystery of 355/113

Related Stories