You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@commons.apache.org by Sébastien Brisard <se...@m4x.org> on 2011/12/31 15:50:44 UTC
[math] Reliability of Well1024a
Hi,
In rev 1226041, I committed a patch proposed by Christian (see
MATH-692). Unfortunately, I didn't notice that this patch causes the
failure of Well1024aTest. Since this morning, I've been looking into
that failure. I don't think that any of Christian's proposed
modifications is to be incriminated. Rather, I'm wondering whether
Well1024a is reliable. Indeed, varying the seed in
Well1024aTest.makeGenerator() causes failure of various unit tests
- SEED = 100 causes testNextPoissonConsistency() to fail,
- SEED = 1000 causes testNextIntPositiveRange(),
testNextLongNegativeRange() and testNextLongPositiveRange() to fail,
while *all* tests pass with SEED = 1001. I think this probability of
failure is well above the 0.001 threshold of the chi-square test.
I'm not very familiar with this part of CM, and would very much like
to know what you think.
NOTE: in rev 1226096, I set the SEED to 1001, so as to make Gump stop
complaining. Obivously, this is a very dirty trick.
Sébastien
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
For additional commands, e-mail: dev-help@commons.apache.org
Re: [math] Reliability of Well1024a
Posted by Sébastien Brisard <se...@m4x.org>.
2011/12/31 Luc Maisonobe <Lu...@free.fr>:
> Le 31/12/2011 15:50, Sébastien Brisard a écrit :
>> Hi,
>> In rev 1226041, I committed a patch proposed by Christian (see
>> MATH-692). Unfortunately, I didn't notice that this patch causes the
>> failure of Well1024aTest. Since this morning, I've been looking into
>> that failure. I don't think that any of Christian's proposed
>> modifications is to be incriminated. Rather, I'm wondering whether
>> Well1024a is reliable. Indeed, varying the seed in
>> Well1024aTest.makeGenerator() causes failure of various unit tests
>> - SEED = 100 causes testNextPoissonConsistency() to fail,
>> - SEED = 1000 causes testNextIntPositiveRange(),
>> testNextLongNegativeRange() and testNextLongPositiveRange() to fail,
>> while *all* tests pass with SEED = 1001. I think this probability of
>> failure is well above the 0.001 threshold of the chi-square test.
>>
>> I'm not very familiar with this part of CM, and would very much like
>> to know what you think.
>
> I am sorry not to have any time to look at this these days.
> Well 1024 has been checked with respect to the reference C
> implementation though. Perhaps we should check it again.
>
> Luc
>
I'm sorry, Luc, perhaps I should have made a less peremptory
statement. I'm not implying this implementation is faulty, I'm merely
suggesting we should look into it again. No offense meant.
Two points are worth noting
- it seems to me that regardless of the seed, all test pass if you
increase the size of the sample (from 1000 to 10000).
- Also, I think that the tests which fail have recently been added,
as part of MATH-724.
Sébastien
>>
>> NOTE: in rev 1226096, I set the SEED to 1001, so as to make Gump stop
>> complaining. Obivously, this is a very dirty trick.
>>
>> Sébastien
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
>> For additional commands, e-mail: dev-help@commons.apache.org
>>
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
> For additional commands, e-mail: dev-help@commons.apache.org
>
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
For additional commands, e-mail: dev-help@commons.apache.org
Re: [math] Reliability of Well1024a
Posted by Luc Maisonobe <Lu...@free.fr>.
Le 31/12/2011 15:50, Sébastien Brisard a écrit :
> Hi,
> In rev 1226041, I committed a patch proposed by Christian (see
> MATH-692). Unfortunately, I didn't notice that this patch causes the
> failure of Well1024aTest. Since this morning, I've been looking into
> that failure. I don't think that any of Christian's proposed
> modifications is to be incriminated. Rather, I'm wondering whether
> Well1024a is reliable. Indeed, varying the seed in
> Well1024aTest.makeGenerator() causes failure of various unit tests
> - SEED = 100 causes testNextPoissonConsistency() to fail,
> - SEED = 1000 causes testNextIntPositiveRange(),
> testNextLongNegativeRange() and testNextLongPositiveRange() to fail,
> while *all* tests pass with SEED = 1001. I think this probability of
> failure is well above the 0.001 threshold of the chi-square test.
>
> I'm not very familiar with this part of CM, and would very much like
> to know what you think.
I am sorry not to have any time to look at this these days.
Well 1024 has been checked with respect to the reference C
implementation though. Perhaps we should check it again.
Luc
>
> NOTE: in rev 1226096, I set the SEED to 1001, so as to make Gump stop
> complaining. Obivously, this is a very dirty trick.
>
> Sébastien
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
> For additional commands, e-mail: dev-help@commons.apache.org
>
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
For additional commands, e-mail: dev-help@commons.apache.org