http://www.perlmonks.org?node_id=482056


in reply to Re^3: Algorithm for cancelling common factors between two lists of multiplicands
in thread Algorithm for cancelling common factors between two lists of multiplicands

Sure. For the following 2x2:
  X Y  
A 989 9,400 10,389
B 43,300 2.400 45,700
  44,289 11,800 56,089

The formula comes out to

(44,289! 11,800!) (10,389! 45,700!) ----------------------------------- 56,089! 989! 9,400! 43,300! 11,800! 2,400!

Which infinite precision will calculate, but it will be quite slow. And remember, in order to determine if the result is significant, there are 11,000 more of these calculations to perform and these numbers are still relatively small. And, theoretically at least, the FET can be applied to more than a 2x2 matrix.


Examine what is said, not who speaks -- Silence betokens consent -- Love the truth but pardon error.
Lingua non convalesco, consenesco et abolesco. -- Rule 1 has a caveat! -- Who broke the cabal?
"Science is about questioning the status quo. Questioning authority".
The "good enough" maybe good enough for the now, and perfection maybe unobtainable, but that should not preclude us from striving for perfection, when time, circumstance or desire allow.

Replies are listed 'Best First'.
Re^5: Algorithm for cancelling common factors between two lists of multiplicands
by tmoertel (Chaplain) on Aug 08, 2005 at 23:16 UTC
    For this matrix, I can compute the exact Pcutoff in about 1 second (on a 1.6-GHz Celeron laptop). How long does the brute-force approach take?
    [thor@arinmir fishers-exact-test]$ cat ex1.dat 989 9400 43300 2400 [thor@arinmir fishers-exact-test]$ time ./fet < ex1.dat > /dev/null real 0m1.007s user 0m0.991s sys 0m0.012s

      The results would be interesting?


      Examine what is said, not who speaks -- Silence betokens consent -- Love the truth but pardon error.
      Lingua non convalesco, consenesco et abolesco. -- Rule 1 has a caveat! -- Who broke the cabal?
      "Science is about questioning the status quo. Questioning authority".
      The "good enough" maybe good enough for the now, and perfection maybe unobtainable, but that should not preclude us from striving for perfection, when time, circumstance or desire allow.
Re^5: Algorithm for cancelling common factors between two lists of multiplicands
by tmoertel (Chaplain) on Aug 08, 2005 at 23:31 UTC
    Quick question: If the sample size is so large, is there a reason you aren't using the Chi-square test? My understanding is Fisher's Exact Test may be preferred when the sample size isn't large enough to reasonably support the large-sample approximation for the Chi-square test. Since you have a large sample, why not take the easy road?
      ... is there a reason you aren't using the Chi-square test?

      I'm not trying to solve the sample problem. That was just an example I found on the web and used as a test.

      I set out to solve the problem of performing the FET using Perl. First I did it with Math::Pari, but which gives a result of 8.070604647867604097576877668E-7030 in 26ms, but I was unsure about the accuracy. It also imposes a binary dependancy.

      #! perl -slw use strict; use Benchmark::Timer; use List::Util qw[ sum reduce ]; use Math::Pari qw[ factorial ]; $a=$b; sub product{ reduce{ $a *= $b } 1, @_ } sub FishersExactTest { my @data = @_; return unless @data == 4; my @C = ( sum( @data[ 0, 2 ] ), sum( @data[ 1, 3 ] ) ); my @R = ( sum( @data[ 0, 1 ] ), sum( @data[ 2, 3 ] ) ); my $N = sum @C; my $dividend = product map{ factorial $_ } grep $_, @R, @C; my $divisor = product map{ factorial $_ } grep $_, $N, @data; return $dividend / $divisor; } my $T = new Benchmark::Timer; $T->start( '' ); print FishersExactTest 989, 9400, 43300, 2400;; $T->stop( '' ); $T->report; __END__ P:\test>MP-FET.pl 8.070604647867604097576877668E-7030 1 trial of _default ( 25.852ms total), 25.852ms/trial

      So, then I coded it using Math::BigFloat

      #! perl -slw use strict; use Benchmark::Timer; use List::Util qw[ reduce ]; use Math::BigFloat; $a=$b; sub product{ reduce{ $a *= $b } 1, @_ } sub sum{ reduce{ $a += $b } 0, @_ } sub FishersExactTest { my @data = map{ Math::BigFloat->new( $_ ) } @_; return unless @data == 4; my @C = ( sum( @data[ 0, 2 ] ), sum( @data[ 1, 3 ] ) ); my @R = ( sum( @data[ 0, 1 ] ), sum( @data[ 2, 3 ] ) ); my $N = sum @C; my $dividend = product map{ $_->bfac } grep $_, @R, @C; my $divisor = product map{ $_->bfac } grep $_, $N, @data; return $dividend / $divisor; } my $T = new Benchmark::Timer; $T->start( '' ); print FishersExactTest 989, 9400, 43300, 2400;; $T->stop( '' ); $T->report;

      But that ran for 20 minutes without producing any output before I killed it (I've set it running again now, and my machines fan has been thrashing at full speed for the last 25 minutes).

      Whilst I was waiting for the BigFloat version, I coded this version which attempts to reduce the size of the problem by eliminating (exactly common) factors:

      sub FishersExactTest2 { my @data = @_; return unless @data == 4; my @C = ( sum( @data[ 0, 2 ] ), sum( @data[ 1, 3 ] ) ); my @R = ( sum( @data[ 0, 1 ] ), sum( @data[ 2, 3 ] ) ); my $N = sum @C; my %dividends; $dividends{ $_ }++ for map{ factors $_ } grep $_, @ +R, @C; my %divisors; $divisors { $_ }++ for map{ factors $_ } grep $_, $ +N, @data; for my $i ( keys %divisors ) { if( exists $dividends{ $i } ) { $divisors{ $i }--, $dividends{ $i }-- while $divisors{ $i } and $dividends{ $i }; delete $divisors { $i } unless $divisors { $i }; delete $dividends{ $i } unless $dividends{ $i }; } } my $dividend = product( map{ ( $_ ) x $dividends{ $_ } } keys %div +idends ); my $divisor = product( map{ ( $_ ) x $divisors { $_ } } keys %div +isors ); return $dividend / $divisor; }

      This works well for values smallish values, but cannot handle the example I gave above (NV overflow).

      It was then I started thinking about how to eliminate more factors from the equation so as to reduce the size of the intermediate terms, and posted my SoPW. I think that hv's solution of expanding all terms to their prime factorizations before performing the cancelling out will be a winner--but I haven't finished coding that yet.


      Examine what is said, not who speaks -- Silence betokens consent -- Love the truth but pardon error.
      Lingua non convalesco, consenesco et abolesco. -- Rule 1 has a caveat! -- Who broke the cabal?
      "Science is about questioning the status quo. Questioning authority".
      The "good enough" maybe good enough for the now, and perfection maybe unobtainable, but that should not preclude us from striving for perfection, when time, circumstance or desire allow.
        My Haskell implementation represents numbers as the ratio of products of ordered integer streams. For example, I represent 3!/(4*5) as (R numerator=[1,2,3] denominator=[4,5]). In this representation, multiplication becomes merging the numerator and denominator streams and then canceling the first stream by the second. In this way I can remove all cancelable original terms in the Pcutoff formula before finally multiplying the terms that remain.
        *FishersExactTest> fac 6 R {numer = [2,3,4,5,6], denom = []} *FishersExactTest> fac 3 R {numer = [2,3], denom = []} *FishersExactTest> fac 6 `rdivide` fac 3 R {numer = [4,5,6], denom = []}
        Here's the example from the MathWorld page:
        *FishersExactTest> rpCutoff [ [5,0], [1,4] ] R {numer = [2,3,4,5], denom = [7,8,9,10]} *FishersExactTest> fromRational . toRatio $ it 2.3809523809523808e-2
        The code:
        module FishersExactTest (pCutoff) where import Data.Ratio import Data.List (transpose) pCutoff = toRatio . rpCutoff rpCutoff rows = facproduct (rs ++ cs) `rdivide` facproduct (n:xs) where rs = map sum rows cs = map sum (transpose rows) n = sum rs xs = concat rows -- cells facproduct = rproduct . map fac fac n | n < 2 = runit | otherwise = R [2..n] [] -- I represent numbers as ratios of products of integer streams -- R [1,2,3] [4,5] === (1 * 2 * 3) / (4 * 5) data Rops = R { numer :: [Int], denom :: [Int] } deriving Show runit = R [] [] -- the number 1 toRatio (R ns ds) = bigProduct ns % bigProduct ds bigProduct = product . map toInteger -- multiplication is merging numerator and denominator streams -- and then canceling the first by the second rtimes (R xns xds) (R yns yds) = uncurry R $ (merge xns yns) `cancel` (merge xds yds) rproduct = foldr rtimes runit -- division is multiplication by the inverse rdivide x (R yns yds) = rtimes x (R yds yns) -- helpers merge (x:xs) (y:ys) | x < y = x : merge xs (y:ys) | otherwise = y : merge (x:xs) ys merge [] ys = ys merge xs [] = xs cancel (x:xs) (y:ys) | x == y = cancel xs ys | x < y = let (xs', ys') = cancel xs (y:ys) in (x:xs', ys') | otherwise = let (xs', ys') = cancel (x:xs) ys in (xs', y:ys') cancel xs ys = (xs, ys)
Re^5: Algorithm for cancelling common factors between two lists of multiplicands
by sk (Curate) on Aug 08, 2005 at 23:05 UTC

      As I understand the FET, it is important that the probabilities add up to 1.


      Examine what is said, not who speaks -- Silence betokens consent -- Love the truth but pardon error.
      Lingua non convalesco, consenesco et abolesco. -- Rule 1 has a caveat! -- Who broke the cabal?
      "Science is about questioning the status quo. Questioning authority".
      The "good enough" maybe good enough for the now, and perfection maybe unobtainable, but that should not preclude us from striving for perfection, when time, circumstance or desire allow.
Re^5: Algorithm for cancelling common factors between two lists of multiplicands
by Limbic~Region (Chancellor) on Aug 10, 2005 at 15:30 UTC
    BrowserUk,
    Once you have a division math problem of factorials that has been reduced by some factoring method (GCD or prime), it can be reduced even further by substraction.
    47! * 1091! ----------- 55! * 1001! (1002 .. 1091) -------------- (48 .. 55)
    This is only a savings when the factorial is being calculated by multiplying all the terms and not by some other approximation method.

    Cheers - L~R