[Numpy-discussion] Random int64 and float64 numbers

Sturla Molden sturla@molden...
Sat Nov 7 13:40:07 CST 2009


David Cournapeau wrote:
> On Fri, Nov 6, 2009 at 6:54 AM, David Goldsmith <d.l.goldsmith@gmail.com> wrote:
>   
>> Interesting thread, which leaves me wondering two things: is it documented
>> somewhere (e.g., at the IEEE site) precisely how many *decimal* mantissae
>> are representable using the 64-bit IEEE standard for float representation
>> (if that makes sense); and are such decimal mantissae uniformly distributed
>>     
>
> They are definitely not uniformly distributed: that's why two numbers
> are close around 1 when they have only a few EPS difference, but
> around 1e100, you have to add quite a few EPS to even get a different
> number at all.
>
> That may be my audio processing background, but I like to think about
> float as numbers which have the same relative precision at any "level"
> - a kind of dB scale. If you want numbers with a fixed number of
> decimals, you need a fixed point representation.
>   

David Godsmith was asking about the mantissae. For a double, that is a 
53 bit signed integer. I.e. you have 52 bit fractional part (bit 0-51), 
11 bit exponent (bit 52-62), and one sign bit (bit 63). The mantissae is 
uniformly distributed like any signed integer. The mantissae of a double 
have 2**53 different integer values: -2**52 to 2**52-1.

But the value of a floating point number is

   value = (-1)**signbit * 2**(exponent - bias) * (1 - fraction)

with bias = 1023 for a double. Thus, floating point numbers are not 
uniformly distributed, but the mantissae is.

For numerical illiterates this might come as a surprise. But in 
numerical mathematics, the resolution is in the number of "significant 
digits", not in "the number of decimals". 101 and .00201 have the same 
numerical precision.

A decimal, on the other hand, can be thought of as a floating point 
number using base-10 instead of base-2 for the exponent:

   value = (-1)**signbit * 10**(exponent - bias) * (1 - fraction)

Decimals and floats are not fundamentally different. There are number 
exactly representable with a decimal that cannot be exactly represented 
with a float. But numerical computation do not become more precise with 
a decimal than a float.



















More information about the NumPy-Discussion mailing list