www.digitalmars.com         C & C++   DMDScript  

digitalmars.D.learn - Should uniform(-real.max, real.max) be inf?

reply Magnus Lie Hetland <magnus hetland.org> writes:
What's the preferred way of generating a random floating-point number 
in the range of a given floating-point type? We have uniform!T() for 
integral types, but nothing similar for floats? And uniform(-real.max, 
real.max) (possibly tweaking the limits) seems to only return inf, 
which isn't terribly helpful.

What's the standard thing to do here? I could just use

  uniform(cast(T) -1, cast(T) 1)*T.max

I guess (for some floating-point type T). Seems to work fine, at least.

Am I missing the obvious way to do it?

-- 
Magnus Lie Hetland
http://hetland.org
Mar 01 2012
next sibling parent Magnus Lie Hetland <magnus hetland.org> writes:
On 2012-03-01 10:52:49 +0000, Magnus Lie Hetland said:


 I could just use
 
   uniform(cast(T) -1, cast(T) 1)*T.max
 
 I guess (for some floating-point type T). Seems to work fine, at least.


Aaactually, not so much. The output here seems to get about the same 
exponent as T.max. Which isn't all that surprising, I guess. (Then 
again, most floating-point numbers *are* pretty large ;-)

So ... any suggestions?

-- 
Magnus Lie Hetland
http://hetland.org
Mar 01 2012
prev sibling parent reply =?UTF-8?B?QWxpIMOHZWhyZWxp?= <acehreli yahoo.com> writes:
On 03/01/2012 02:52 AM, Magnus Lie Hetland wrote:

 What's the preferred way of generating a random floating-point number in
 the range of a given floating-point type? We have uniform!T() for
 integral types, but nothing similar for floats? And uniform(-real.max,
 real.max) (possibly tweaking the limits) seems to only return inf, which
 isn't terribly helpful.

 What's the standard thing to do here? I could just use

 uniform(cast(T) -1, cast(T) 1)*T.max

 I guess (for some floating-point type T). Seems to work fine, at least.

 Am I missing the obvious way to do it?


I recommend reading this page:

   http://dlang.org/d-floating-point.html

Especially the ASCII graph there is very interesting. The number of 
distinct values between T.min_normal and 1 are equal to the distinct 
values between 1 and T.max.

Since there are also sub-normal values between 0 and T.min_normal, it 
may make sense to use the range [T.min_normal, 1) and scale the result 
from there. But I haven't tested whether the distinct values in that 
range are equally distributed.

Ali
Mar 01 2012
next sibling parent Magnus Lie Hetland <magnus hetland.org> writes:
On 2012-03-01 16:34:23 +0000, Ali Çehreli said:


 I recommend reading this page:
 
    http://dlang.org/d-floating-point.html
 


Thanks.


 Especially the ASCII graph there is very interesting. The number of 
 distinct values between T.min_normal and 1 are equal to the distinct 
 values between 1 and T.max
 .


Ah -- right.


 Since there are also sub-normal values between 0 and T.min_normal, it 
 may make sense to use the range [T.min_normal, 1) and scale the result 
 from there.


Hm. Seems reasonable.


 But I haven't tested whether the distinct values in that range are 
 equally distributed.


At the moment, I'm just using this for test-case generation, so 
anything close to a reasonable approximation is fine :)

However: Perhaps it would be useful with a uniform!real() or the like 
in Phobos? Or is that (because of the vagaries of FP) a weird thing to 
do, inviting misunderstandings and odd behavior? Perhaps it's been left 
out for a reason? (Sounds sort of likely ;-)

-- 
Magnus Lie Hetland
http://hetland.org
Mar 02 2012
prev sibling parent Magnus Lie Hetland <magnus hetland.org> writes:
On 2012-03-01 16:34:23 +0000, Ali Çehreli said:


 Since there are also sub-normal values between 0 and T.min_normal, it 
 may make sense to use the range [T.min_normal, 1) and scale the result 
 from there. But I haven't tested whether the distinct values in that 
 range are equally distributed.


I get similar results now that I did when I started with the range from 
-1 to 1, but I'm guessing it's my brain that's a little slow. I was 
perplexed that (for float) almost all the numbers had an exponent of 
38, while only a few had 37, and none had anything else (in my limited 
tests).

Buuuut that's just the problem with (at least my) "common sense" and 
exponentials/logarithms. There are, of course, ten times as many 
numbers with an exponent of 38 as there are with an exponent of 37 (and 
so on, down the line; c.f., the incompressibility lemma etc.).

For testing, I might want some small numbers, too -- perhaps I should 
just generate the mantissa and exponent separately (maybe even throwing 
in some NaNs and Infs etc.) :)

Thanks for your help, though.

-- 
Magnus Lie Hetland
http://hetland.org
Mar 02 2012