[random] infinite loop in uniform_real_distribution

Jérémy Coulon

25 Nov 2013 25 Nov '13

7:16 p.m.

Hello, I'm using Boost 1.54. The following code hangs in infinite loop when generating a single floating point number : #include #include <iostream> #include <limits> int main() { boost::random::mt19937 rng; boost::random::uniform_real_distribution<float> distrib( -std::numeric_limits<float>::max(), std::numeric_limits<float>::max()); std::cout << distrib(rng) << std::endl; return 0; } I can reproduce the bug with both - Linux GCC 4.8 x86_64 and - Windows VisualStudio 2012 32bits It looks like this uniform distribution doesn't work if the range [min,max] is large. From uniform_real_distribution.hpp:61 T result = numerator / divisor * (max_value - min_value) + min_value; Here the operation (max_value - min_value) produces an overflow. Regards, Jeremy

Show replies by date

Brian Budge

25 Nov 25 Nov

9:03 p.m.

New subject: [random] infinite loop in uniform_real_distribution

On Mon, Nov 25, 2013 at 11:16 AM, Jérémy Coulon wrote:

...

Hello,

I'm using Boost 1.54. The following code hangs in infinite loop when generating a single floating point number :

#include #include <iostream> #include <limits>

int main() { boost::random::mt19937 rng; boost::random::uniform_real_distribution<float> distrib( -std::numeric_limits<float>::max(), std::numeric_limits<float>::max());

std::cout << distrib(rng) << std::endl; return 0; }

I can reproduce the bug with both - Linux GCC 4.8 x86_64 and - Windows VisualStudio 2012 32bits

It looks like this uniform distribution doesn't work if the range [min,max] is large. From uniform_real_distribution.hpp:61 T result = numerator / divisor * (max_value - min_value) + min_value;

Here the operation (max_value - min_value) produces an overflow.

Regards, Jeremy

It seems to me that it's not really just "large" that is the issue, but only if you have a range larger than std::numeric_limits<float>::max(). Ideally, you'd not really want to be messing with too very many numbers that are within many orders of magnitude of the max float value, as your precision will be awful. Are you just edge-case testing, or do you have a real reason to generate uniform distributions over this range? Brian

Jérémy Coulon

9:27 p.m.

New subject: [random] infinite loop in uniform_real_distribution

...

It seems to me that it's not really just "large" that is the issue, but only if you have a range larger than std::numeric_limits<float>::max(). Ideally, you'd not really want to be messing with too very many numbers that are within many orders of magnitude of the max float value, as your precision will be awful. Are you just edge-case testing, or do you have a real reason to generate uniform distributions over this range?

Brian The documentation doesn't say anything about the input range of the distribution. So I expect it to work with *any* range. As for the real reason, I need to unit-test a function of my own code

Le 25/11/2013 22:03, Brian Budge a écrit : that takes a float as an argument. I want to test that my function behaves correctly for any given float. Using a few thousands of randomly generated floats and calling my function with that seems a good idea. Jeremy

Brian Budge

10:22 p.m.

New subject: [random] infinite loop in uniform_real_distribution

On Mon, Nov 25, 2013 at 1:27 PM, Jérémy Coulon wrote:

...

Le 25/11/2013 22:03, Brian Budge a écrit :

...
It seems to me that it's not really just "large" that is the issue, but only if you have a range larger than std::numeric_limits<float>::max(). Ideally, you'd not really want to be messing with too very many numbers that are within many orders of magnitude of the max float value, as your precision will be awful. Are you just edge-case testing, or do you have a real reason to generate uniform distributions over this range?

Brian

The documentation doesn't say anything about the input range of the distribution. So I expect it to work with *any* range.

Fair enough :) I'd say that you are right that it should be documented.

...

As for the real reason, I need to unit-test a function of my own code that takes a float as an argument. I want to test that my function behaves correctly for any given float. Using a few thousands of randomly generated floats and calling my function with that seems a good idea.

Sure. I suppose you have other test cases for inf, NaN, and denorm values? How will you know it behaves correctly for the random input (unless of course, you have another implementation that can run in lockstep to verify correctness?). Sorry, I'm sure you're not here looking for criticism of your testing methods :) You can achieve what you want by generating a random number on [-1, 1] and multiplying it by the max float. I don't think that there is a *general* way to arrive at a uniform distribution of a range greater than max float by first generating a random number on [0, 1), so it's unlikely that the library can be "fixed" to achieve what you want. Another method would be to use double precision generation on the max float range, and then convert to float.

...

Jeremy

Brian

Steven Watanabe

10:53 p.m.

New subject: [random] infinite loop in uniform_real_distribution

AMDG On 11/25/2013 02:22 PM, Brian Budge wrote:

...

You can achieve what you want by generating a random number on [-1, 1] and multiplying it by the max float. I don't think that there is a *general* way to arrive at a uniform distribution of a range greater than max float by first generating a random number on [0, 1), so it's unlikely that the library can be "fixed" to achieve what you want.

It's really quite easy. if ( max - min == infinity ) { divisor = 2; min /= divisor; max /= divisor; } else { divisor = 1; } // compute a value in [min, max) as usual result *= divisor;

...

Another method would be to use double precision generation on the max float range, and then convert to float.

In Christ, Steven Watanabe

Brian Budge

26 Nov 26 Nov

12:33 a.m.

New subject: [random] infinite loop in uniform_real_distribution

On Mon, Nov 25, 2013 at 2:53 PM, Steven Watanabe wrote:

...

AMDG

On 11/25/2013 02:22 PM, Brian Budge wrote:

...
You can achieve what you want by generating a random number on [-1, 1] and multiplying it by the max float. I don't think that there is a *general* way to arrive at a uniform distribution of a range greater than max float by first generating a random number on [0, 1), so it's unlikely that the library can be "fixed" to achieve what you want.

It's really quite easy.

if ( max - min == infinity ) { divisor = 2; min /= divisor; max /= divisor; } else { divisor = 1; }

Ah, I see. Nice :) Thanks for suggesting another simple solution. Brian

Jérémy Coulon

6:51 p.m.

New subject: [random] infinite loop in uniform_real_distribution

...

On Mon, Nov 25, 2013 at 2:53 PM, Steven Watanabe wrote:

...
It's really quite easy.

if ( max - min == infinity ) { divisor = 2; min /= divisor; max /= divisor; } else { divisor = 1; }

Thank you Steven for your solution. I just searched on Boost Trac and found this 8 months old ticket: https://svn.boost.org/trac/boost/ticket/8351

It looks like the same bug we are talking about even if it is not really related to #6059. Is there any chance to see it fixed in the next Boost release ? Thank you both for your time. Regards, Jeremy

Brian Budge

12:33 a.m.

New subject: [random] infinite loop in uniform_real_distribution

On Mon, Nov 25, 2013 at 2:53 PM, Steven Watanabe wrote:

...

AMDG

On 11/25/2013 02:22 PM, Brian Budge wrote:

...
You can achieve what you want by generating a random number on [-1, 1] and multiplying it by the max float. I don't think that there is a *general* way to arrive at a uniform distribution of a range greater than max float by first generating a random number on [0, 1), so it's unlikely that the library can be "fixed" to achieve what you want.

It's really quite easy.

if ( max - min == infinity ) { divisor = 2; min /= divisor; max /= divisor; } else { divisor = 1; }

Ah, I see. Nice :) Thanks for suggesting another simple solution. Brian

4002

Age (days ago)

4003

Last active (days ago)

List overview

Download

7 comments

3 participants

participants (3)

Brian Budge
Jérémy Coulon
Steven Watanabe