[SciPy-User] random variable for truncated multivariate normal and t distributions

[josef.pktd at gmail.com]

On Thu, Jun 2, 2011 at 7:11 PM, Sturla Molden <sturla at molden.no> wrote:
> Den 02.06.2011 18:54, skrev josef.pktd at gmail.com:
>>
>> If these are the alternatives, then I will stick with rejection sampling.
>> I'm not starting to learn the implementation details of simulating
>> with MCMC, Metropolis-Hastings or Gibbs, and leave it to the pymc
>> developers and to Wes.
>
> Metropolis-Hastings is a form of rejection sampling.
>
> It's just a way to reduce the number of rejections, particularly when the sample space is large.
>
>
>
>
>> rtmvnorm has a big Warning label about the Gibbs sampler, although,
>> for MonteCarlo integration, any serial correlation in the sampler
>> won't be very relevant.
>
> You will get serial correlation with MCMC, but remember they are still
> samples from the stationary distribution of the Markov chain. You can
> still use these samples to compute mean, standard deviation, KDE,
> numerical integrals, etc.

(a late reply, since I was looking at this again.)

Thanks Sturla,

I never thought of this distinction in the usage.
Until now I was mostly interested in simulating stochastic processes,
time series, or random samples for regression. In these cases I cannot
have any spurious serial correlation.
But now that I need more Monte Carlo integration, this will be useful.

Josef



>
>
> Sturla
>
>
>
>
> _______________________________________________
> SciPy-User mailing list
> SciPy-User at scipy.org
> http://mail.scipy.org/mailman/listinfo/scipy-user
>