Float is float [Software]

posted by PharmCat  – Russia, 2020-12-18 20:53 (272 d 20:28 ago) – Posting: # 22150
Views: 3,653

(edited by PharmCat on 2020-12-18 22:02)

Hi all,

This is very trivial story... let's take

a = 0.1

next float is:

julia> nextfloat(0.1)
0.10000000000000002


so

julia> nextfloat(0.1) - 0.1
1.3877787807814457e-17


float numbers is discrete... you can't use 0.10000000000000001 ...

you can use rational numbers:

julia> 5//10 - 4//10 - 1//10 == -4//10 - 1//10 + 5//10
true


or some packages for working with big decimals

julia> using Decimals
julia> decimal(0.5)-decimal(0.4)-decimal(0.1) == decimal(-0.4)-decimal(0.1)+decimal(0.5)
true


I think in R it can be done... and never use Excel for cacl...

edit: or you can do float(big(1//10)) if you just need more precision...

edit2:

» which is just “noise”

you should be very careful with eps() because this is not an indicator of accuracy

you can use something like:

a = 1e-70
b = 2e-70
b - a


end it will work fine, problems begans whe you do somesing like: 2.0 + 1e-70 , but you can still use 2.0 * 1e-70 without problems ...

make sure not to sum numbers with difference in eps() times ...

for comparation

all.equal(target, current, tolerance = sqrt(.Machine$double.eps))

can be used, but for small numbers you should adjust tolerance

Complete thread:

Activity
 Admin contact
21,687 posts in 4,534 threads, 1,542 registered users;
online 8 (1 registered, 7 guests [including 4 identified bots]).
Forum time: Friday 18:21 CEST (Europe/Vienna)

Genius is that which forces
the inertia of humanity to learn.    Henri Bergson

The Bioequivalence and Bioavailability Forum is hosted by
BEBAC Ing. Helmut Schütz
HTML5