| Re: random musings [message #53121] |
Tue, 20 March 2007 01:53 |
Paolo Grigis
Messages: 171
Registered: December 2003
|
Senior Member |
|
|
Christopher Thom wrote:
> Quoth Wayne Landsman:
>
>> On Mar 19, 4:30 pm, Christopher Thom <c...@oddjob.uchicago.edu> wrote:
>>> Hi all,
>>>
>>> this morning I started writing some monte carlo code, in which I'm
>>> generating random arrays, and scaling/shifting them to fit the (gaussian)
>>> distribution parameters I need. One thing is not clear in the online help:
>>> Which random function is "better"? Should I be using randomn() or
>>> randomu()?
>> I believe these functions are identical. In the ancient days, IDL had
>> two types of randomn number generators: RANDOMU, for a uniform
>> distribution, and RANDOMN for a normal distribution. When addition
>> distributions (e.g. binomial, Poisson, Gamma) were added (in V5.0 I
>> believe), RSI chose against adding additional RANDOM* functions, and
>> instead made all the distributions available from either RANDOMU or
>> RANDOMN.
>
> Hmmmm. If this is true, then the documentation is definitely in need of an
> update. Randomn() supposedly uses the "box muller method", while randomu()
> uses one from an ACM paper that is cited, with a "bays durham shuffle
> added", giving something "similar to ran1()" from NR in C section 7.1.
> If they are the same thing, it is certainly not clear from the docs. :-(
In most implementations, uniformly distributed random numbers are generated
first, and then some trick is used to transform these into the distribution
you want (e.g. normal in this case). Therefore, "ran1()" (or whatever flavor
thereof they use internally) produces the uniform samples, an then the box
muller transformation is applied to these values to get the normally
distributed samples.
Ciao,
Paolo
>
> OTOH, the interfaces seem very similar...they all provide the same flags
> (/gamma, /normal etc), which certainly points towards a common algorithm.
>
> Oh. I just did the obvious thing -- tried both with the same seed:
>
> IDL> x=randomn(180479L, 5,/normal)
> IDL> y=randomu(180479L, 5,/normal)
> IDL> print,x,y
> 2.01146 -1.55186 -0.0893551 -0.627972 -0.780082
> 2.01146 -1.55186 -0.0893551 -0.627972 -0.780082
>
> Surely different algorithms would return different randomness for the same
> seed.
>
> Problem solved...
>
> cheers
> chris
|
|
|
|
| Re: random musings [message #53124 is a reply to message #53121] |
Mon, 19 March 2007 21:41  |
Christopher Thom
Messages: 66
Registered: October 2006
|
Member |
|
|
Quoth Wayne Landsman:
> On Mar 19, 4:30 pm, Christopher Thom <c...@oddjob.uchicago.edu> wrote:
>> Hi all,
>>
>> this morning I started writing some monte carlo code, in which I'm
>> generating random arrays, and scaling/shifting them to fit the (gaussian)
>> distribution parameters I need. One thing is not clear in the online help:
>> Which random function is "better"? Should I be using randomn() or
>> randomu()?
>
> I believe these functions are identical. In the ancient days, IDL had
> two types of randomn number generators: RANDOMU, for a uniform
> distribution, and RANDOMN for a normal distribution. When addition
> distributions (e.g. binomial, Poisson, Gamma) were added (in V5.0 I
> believe), RSI chose against adding additional RANDOM* functions, and
> instead made all the distributions available from either RANDOMU or
> RANDOMN.
Hmmmm. If this is true, then the documentation is definitely in need of an
update. Randomn() supposedly uses the "box muller method", while randomu()
uses one from an ACM paper that is cited, with a "bays durham shuffle
added", giving something "similar to ran1()" from NR in C section 7.1.
If they are the same thing, it is certainly not clear from the docs. :-(
OTOH, the interfaces seem very similar...they all provide the same flags
(/gamma, /normal etc), which certainly points towards a common algorithm.
Oh. I just did the obvious thing -- tried both with the same seed:
IDL> x=randomn(180479L, 5,/normal)
IDL> y=randomu(180479L, 5,/normal)
IDL> print,x,y
2.01146 -1.55186 -0.0893551 -0.627972 -0.780082
2.01146 -1.55186 -0.0893551 -0.627972 -0.780082
Surely different algorithms would return different randomness for the same
seed.
Problem solved...
cheers
chris
|
|
|
|
| Re: random musings [message #53125 is a reply to message #53124] |
Mon, 19 March 2007 13:58 |
news.verizon.net
Messages: 47
Registered: August 2003
|
Member |
|
|
On Mar 19, 4:30 pm, Christopher Thom <c...@oddjob.uchicago.edu> wrote:
> Hi all,
>
> this morning I started writing some monte carlo code, in which I'm
> generating random arrays, and scaling/shifting them to fit the (gaussian)
> distribution parameters I need. One thing is not clear in the online help:
> Which random function is "better"? Should I be using randomn() or
> randomu()?
I believe these functions are identical. In the ancient days, IDL
had two types of randomn number generators: RANDOMU, for a uniform
distribution, and RANDOMN for a normal distribution. When addition
distributions (e.g. binomial, Poisson, Gamma) were added (in V5.0 I
believe), RSI chose against adding additional RANDOM* functions, and
instead made all the distributions available from either RANDOMU or
RANDOMN.
This is one of those case where IDL suffers from having a long
history.
--Wayne
|
|
|
|
| Re: random musings [message #53127 is a reply to message #53125] |
Mon, 19 March 2007 13:37 |
news.qwest.net
Messages: 137
Registered: September 2005
|
Senior Member |
|
|
"Christopher Thom" <cthom@oddjob.uchicago.edu> wrote in message
news:Pine.SOC.4.64.0703191516310.28339@oddjob.uchicago.edu.. .
> Hi all,
>
> this morning I started writing some monte carlo code, in which I'm
> generating random arrays, and scaling/shifting them to fit the (gaussian)
> distribution parameters I need. One thing is not clear in the online help:
> Which random function is "better"? Should I be using randomn() or
> randomu()? The help is explicit about the algorithms used, but there is
> comparison between the randomness of the two functions. And since both
> functions accept a /normal flag, it's unclear as to which is the better
> choice.
>
> Is there any common wisdom that makes this choice obvious?
>
> cheers
> chris
That's a good question. I have always used randomn for my purposes.
If you really need a good RNG, check out:
http://sprng.cs.fsu.edu/Version2.0/users-guide.html
You can read more about RNG and Monte Carlo and the
tests they perform in the user's guide. They have a PC linux port
available.
Cheers,
bob
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
