Hi all, Indeed there is a problem with the Halton function under the 64 bit architecture. For those who want to take a look, I put two csv files here http://dutangc.free.fr/pub/
I will ask Diethelm, the author of this Fortran code, to see what can be done. Christophe Le 15 sept. 2009 à 18:53, Anirban Mukherjee a écrit : > I did not compile R. I used the Leopard installer from > http://r.research.att.com/ > which installs both the 64 bit and the 32 bit frame works and apps. > > There may be a problem with the gap between the two, but it seemed > like the xcode version recommended for Leopard was the same as the > xcode version for Snow Leopard, and I have the Fortran compiler from > the same website. I don't get any errors when compiling the package. > > Let me know if I can help. I am not a Fortran programmer, so > unfortunately cannot be helpful there. Thanks! > > Best, > A > > On Sep 15, 2009, at 12:40 PM, Christophe Dutang wrote: > >> Ok so you mistype pseudo instead of quasi. My Email was useless... I >> will get snow Leopard in the coming days and will try to reproduce >> your problem. Did you compile R 64 yourself? >> >> iPhone.fan >> >> Le 15 sept. 2009 à 18:25, Anirban Mukherjee <[email protected]> a >> écrit : >> >> > Sorry, what I should have said was Halton numbers are quasi-random, >> > and not pseudo-random. Quasi-random is the technically appropriate >> > terminology. >> > >> > Halton sequences are low discrepancy: the subsequence looks/smells >> > random. Hence, they are often used in quasi monte carlo >> simulations. >> > To be precise, there is only 1 Halton sequence for a particular >> prime. >> > Repeated calls to Halton should return the same numbers. The first >> > column is the Halton sequence for 2. the second for 3, the third >> for 5 >> > and so on using the first M primes (for M columns). (You can also >> > scramble the sequence to avoid this.) >> > >> > I am using them to integrate over a multivariate normal space. If >> you >> > take 1000 random draws, then sum f() over the draws is the >> expectation >> > of f(). If f() is very non-linear (and/or multi-variate) then even >> > with large N, its often hard to get a good integral. With quasi- >> random >> > draws, the integration is better for the same N. (One uses the >> inverse >> > distribution function.) For an example, you can look at Train's >> paper >> > (page 4 and 5 have a good explanation) at: >> > >> > http://elsa.berkeley.edu/wp/train0899.pdf >> > >> > In the context of simulated maximum likelihood estimation, such >> > integrals are very common. Of course true randomness has its own >> > place/importance: its just that quasi-random numbers can be very >> > useful in certain contexts. >> > >> > Regards, >> > Anirban >> > >> > PS: qnorm(halton()) gets around the problem of the random deviates >> > not working. >> > >> > On Tue, Sep 15, 2009 at 11:37 AM, David Winsemius >> > <[email protected]> wrote: >> >> >> >> On Sep 15, 2009, at 11:10 AM, Anirban Mukherjee wrote: >> >> >> >>> Thanks everyone for your replies. Particularly David. >> >>> >> >>> The numbers are pseudo-random. Repeated calls should/would give >> the >> >>> same output. >> >> >> >> As I said, this package is not one with which I have experience. >> It >> >> has _not_ however the case that repeated calls to (typical?) >> random >> >> number functions give the same output when called repeatedly: >> >> >> >> > rnorm(10) >> >> [1] -0.8740195 2.1827411 -0.1473012 -1.4406262 0.1820631 >> >> -1.3151244 -0.4813703 0.8177692 >> >> [9] 0.2076117 1.8697418 >> >> > rnorm(10) >> >> [1] -0.7725731 0.8696742 -0.4907099 0.1561859 0.5913528 >> >> -0.8441891 0.2285653 -0.1231755 >> >> [9] 0.5190459 -0.7803617 >> >> > rnorm(10) >> >> [1] -0.9585881 -0.0458582 1.1967342 0.6421980 -0.5290280 >> >> -1.0735112 0.6346301 0.2685760 >> >> [9] 1.5767800 1.0864515 >> >> > rnorm(10) >> >> [1] -0.60400852 -0.06611533 1.00787048 1.48289305 0.54658888 >> >> -0.67630052 0.52664127 -0.36449997 >> >> [9] 0.88039397 0.56929333 >> >> >> >> I cannot imagine a situation where one would _want_ the output >> to be >> >> the same on repeated calls unless one reset a seed. Unless >> perhaps I >> >> am not understanding the meaning of "random" in the financial >> domain? >> >> >> >> -- >> >> David >> >> >> >>> Currently, Halton works fine when used to just get the >> >>> Halton sequence, but the random deviates call is not working in >> 64 >> >>> bit >> >>> R. For now, I will generate the numbers in 32 bit R, save them >> and >> >>> then load them back in when using 64 bit R. The package >> maintainers >> >>> can look at it if/when they get a chance and/or access to 64 >> bit R. >> >>> >> >>> Thanks! >> >>> >> >>> Best, >> >>> Anirban >> >>> >> >>> On Tue, Sep 15, 2009 at 9:01 AM, David Winsemius <[email protected] >> >>>> wrote: >> >>>> I get very different output from the two versions of Mac OSX R >> as >> >>>> well. The 32 bit version puts out a histogram that has an >> expected, >> >>>> almost symmetric unimodal distribution. The 64 bit version >> >>>> created a >> >>>> bimodal distribution with one large mode near 0 and another >> smaller >> >>>> mode near 10E+37. Postcript output attached. >> >>>> >> >>> >> >>> >> >>> >> >>> -- >> >>> Anirban Mukherjee | Assistant Professor, Marketing | LKCSB, SMU >> >>> 5062 School of Business, 50 Stamford Road, Singapore 178899 | >> >>> +65-6828-1932 >> >>> >> >>> _______________________________________________ >> >>> R-SIG-Mac mailing list >> >>> [email protected] >> >>> https://stat.ethz.ch/mailman/listinfo/r-sig-mac >> >> >> >> David Winsemius, MD >> >> Heritage Laboratories >> >> West Hartford, CT >> >> >> >> >> > >> > >> > >> > -- >> > Anirban Mukherjee | Assistant Professor, Marketing | LKCSB, SMU >> > 5062 School of Business, 50 Stamford Road, Singapore 178899 | >> +65-6828-1932 >> > >> > _______________________________________________ >> > R-SIG-Mac mailing list >> > [email protected] >> > https://stat.ethz.ch/mailman/listinfo/r-sig-mac >> > -- Christophe Dutang Ph.D. student at ISFA, Lyon, France website: http://dutangc.free.fr [[alternative HTML version deleted]]
_______________________________________________ R-SIG-Mac mailing list [email protected] https://stat.ethz.ch/mailman/listinfo/r-sig-mac
