in reply to Add numbers in log domain [solved]

well, suppose that x >= y, and that we are using the natural base: 
log(x + y) = log(x * (1 + y/x)) = log(x) + log(1 + y/x)

y/x = exp(log(y/x)) = exp(log(y) - log(x))

log(x + y) = log(x) + log(1 + exp(log(y) - log(x)))

[download]

And that formula can be easily calculated using regular perl scalars because log(y) - log(x) is in [-log(x), 0] and exp(log(y) - log(x)) is in (0, 1].

Replies are listed 'Best First'.
Re^2: Add numbers in log domain
by BrowserUk (Patriarch) on Jun 04, 2012 at 12:47 UTC

    Spot on! (Though you do need to know the arbitrary base B):

    sub log2p5{ log( $_[0] ) / log( 2.5 ) };;

$X = log2p5( 10 ); 
$Y = log2p5( 15 ); 
$Z = log2p5( 10 + 15 ); 
print $X, $Y, $Z;;
2.51294159473206 2.95544864408182 3.51294159473206

print 2.51294159473206 + log2p5( 1 + 2.5**( 2.95544864408182 - 2.51294
+159473206 ) );;
3.51294159473206

    [download]
    With the rise and rise of 'Social' network sites: 'Computers are making people easier to use everyday'
    Examine what is said, not who speaks -- Silence betokens consent -- Love the truth but pardon error.
    "Science is about questioning the status quo. Questioning authority".
    In the absence of evidence, opinion is indistinguishable from prejudice.

    The start of some sanity?
[reply]
[d/l]
      yes, but it is easy to convert from any base to natural, perform the computation and then convert back to the original base.
[reply]
Re^2: Add numbers in log domain
by ambrus (Abbot) on Jun 04, 2012 at 12:34 UTC

    Yes, something like that should work. Take the Math::Libm module, and use the expm1 and log1p functions for this. I won't figure out the correct formula now.

    Update 2012-06-10: though this won't help OP, future people who stumble this question by searching might want to look at the Math::BigApprox module.

[reply]
Re^2: Add numbers in log domain
by pRaNaV (Novice) on Jun 05, 2012 at 07:20 UTC

    Thanks! That will save one exponentiation step. Earlier I had to do base^x, base^y, now only one exponentiation step is required (exp(log(y) - log(x))). Btw, range for log(y) - log(x) is (-inf 0]

    As I came to learn from another source where they CREATE log table for this particular problem (function logmath_init() in http://cmusphinx.svn.sourceforge.net/viewvc/cmusphinx/trunk/sphinxbase/src/libsphinxbase/util/logmath.c?revision=11275&view=markup)

    It's like this - first set (log(y) - log(x)) = 0; then compute base^(1 + exp(log(y) - log(x))

    then decrement log(y) - log(x) by 1 (equivalent to dividing by base) and recomputing base^(1 + exp(log(y) - log(x))

    this is done untill we reach the MOST negative number. Using this table one can back-compute any log(x+y) for that particular base.

    Thanks again for your help.

[reply]

In Section Seekers of Perl Wisdom