Re: Re: bug in RandomChoice if weight is zero?
- To: mathgroup at smc.vnet.net
- Subject: [mg106056] Re: [mg106027] Re: [mg105976] bug in RandomChoice if weight is zero?
- From: Leonid Shifrin <lshifr at gmail.com>
- Date: Thu, 31 Dec 2009 03:17:45 -0500 (EST)
- References: <200912290617.BAA02582@smc.vnet.net>
Nothing, but as I said, issuing diagnostic message for this particular case
seems to me like a point solution, given that there may be extremely small
non-zero weights expressed analytically (as a rational for example). I may
be wrong of course.
On Wed, Dec 30, 2009 at 8:29 AM, DrMajorBob <btreat1 at austin.rr.com> wrote:
> Generally, what you ask for should be possible once you give a specific
>> interpretation of what you mean by zero
>>
>
> What's ambiguous about the Integer 0?
>
> Bobby
>
>
> On Wed, 30 Dec 2009 03:17:21 -0600, Leonid Shifrin <lshifr at gmail.com>
> wrote:
>
> Hi Matthias,
>>
>> presumably this is because RandomSample never samples any given element
>> more
>> than once - in that case it has no other choice, regardless of the weight.
>> Perhaps, the error message might be added for exact symbolic zero
>> appearing
>> somewhere in the wieights for the case where the size of the desired
>> sample
>> is larger than the number of non-zero weights (non-zero means
>> UnsameQ[weight,0]). It is not clear however whether it is worth it given
>> the overhead of such an analysis. Besides, this will be a point solution
>> anyway, since some expressions that might be used for weights may be in
>> fact
>> identically equal to zero but just not simplified to zero - these cases
>> would not be caught by such an analysis.
>>
>> OTOH, it is hard to make a formal (not based on syntax) distinction
>> between
>> zero and some very small weight close to zero (which may be expressed not
>> necessarily with machine precision). This of course depends on the
>> implementation of the underlying algorithm used to do the sampling -
>> whether
>> or not it treats weights smaller than machine epsilon as zero. From the
>> point of view of the abstract interface of RandomSample, there is no
>> reason
>> why it should do that, even if such events are highly unlikely to be
>> sampled.
>>
>> Generally, what you ask for should be possible once you give a specific
>> interpretation of what you mean by zero, but arguably this would add more
>> ambiguities/restrictions than it is worth for a kind of a general-purpose
>> function that RandomSample is. After all, you can always wrap RandomSample
>> in your own wrapper function which will implement this functionality in
>> the
>> way you want it (consistent with your definition of zero), while
>> preserving
>> the syntax of the original RandomSample.
>>
>> Regards,
>> Leonid
>>
>>
>>
>>
>>
>> On Tue, Dec 29, 2009 at 9:17 AM, Matthias Greiff <greiff at mac.com> wrote:
>>
>> If I use RandomChoice with weights of zero the corresponding elements
>>> will
>>> never be selected.
>>>
>>> In[62]:= RandomChoice[{0, 0, 3} -> {3, 2, 1}, 5]
>>> Out[62]= {1, 1, 1, 1, 1}
>>>
>>> Why is it not the same if I use RandomSample?
>>> The following command
>>>
>>> RandomSample[{1, 2, 0} -> Range[3], 2]
>>>
>>> returns either {1,2} or {2,1}. But when I change the command to sample
>>> size
>>> 3 I get the following.
>>>
>>> In[60]:= RandomSample[{1, 2, 0} -> Range[3], 3]
>>> Out[60]= {2, 1, 3}
>>>
>>> Why is the third element selected? Shouldn't Mathematica return an error
>>> message because the sample size is larger than the population size?
>>>
>>> Appreciate your answers.
>>>
>>> Merry Christmas!
>>>
>>> Matthias
>>>
>>>
>>>
>
> --
> DrMajorBob at yahoo.com
>
- References:
- bug in RandomChoice if weight is zero?
- From: Matthias Greiff <greiff@mac.com>
- bug in RandomChoice if weight is zero?