bigdecimal with lower precision that Float
I'm not sure if I've misunderstood the bigdecimal class but in the following example, I only get 12 significant digits using bigdecimal while using Float, I get a correct value with 17 significant digits.
# using floats 101/0.9163472602589686 # 110.22022368622177 (OK: floating point computation) # using bigdecimal a = BigDecimal('101'); a.precs # [9, 18] b = BigDecimal('0.9163472602589686'); b.precs # [18, 27] c = a/b; c.precs # [18, 36] (OK: I understand that c is computed with 18 significant digits) c.to_s # "0.110220223686e3" (Mmm: I see only 12 significant digits) c - BigDecimal('0.110220223686e3') # 0.0 (Looks like c only stores 12 significant digits and not 18)
Using the Rational class, I've seen that the value I'm expecting is about:
BigDecimal.new(Rational(101/Rational('0.9163472602589686')), 25) # 0.1102202236862217746799312e3
Updated by karatedog (Földes László) over 1 year ago
That is the same problem as here: https://bugs.ruby-lang.org/issues/8826
#/ is the same method as #quo (according to documentation both methods are defined in 'bigdecimal.c' at line 1281). Currently you can divide a bigdecimal by using #/, #quo and #div but I don't really understand the design behind these methods (on a "which should do what" level).
#div accepts a precision argument, while #quo does not. Without precision argument #div returns Fixnum even if its first argument is a Float, it even returns Fixnum if both divisor and dividend are Float..
Thus far I don't know any method that could be able to calculate a division AND set the proper precision on the result. What you can do is to manually set precision by using #div. If you set the precision to the same amount as the divisor, you will not miss any significant digits, the drawback is that you will see a lot of digit repetition for most of the numbers.
(1019 is a long prime, its reciprocal has 1018 significant digits)