dvats
commented
5 years ago Hi Luiz,
Thanks for your email. I am glad you've found the package to be useful.
You are absolutely right, in that the problem is mainly of numerical instability, and can easily be avoided by replacing the negative eigenvalues with a small number. This is exactly what I have now done in the GitHub version of the package (I had the code written up, I just hadn't pushed it through yet). I will make more changes to the usability of the package over the summer break, and update the CRAN version as well. Till then, I hope this fix works for you.
Kind regards, Dootika
On Thu, May 2, 2019 at 8:19 PM Luiz Max F. Carvalho notifications@github.com<mailto:notifications@github.com> wrote:
First of all, thanks for this excellent package. What follows is not a complaint but a doubt. I'm trying to compute mESS for a particular MCMC run and I keep getting the warning message: You either need more samples or x is not full column rank, which is mcse.multi trying to tell me something is off with the covariance matrix. When I compute the eigen values, I get something like this:
eigen(cc, only.values = TRUE) $values [1] 1.831821e+03 3.489175e+02 6.867367e+01 4.431231e+01 1.894525e+01 1.736171e+01 3.760838e-02 2.132381e-02 1.203988e-02 [10] 2.598515e-03 2.441643e-03 6.132745e-04 5.422644e-04 4.971468e-04 4.429832e-04 4.231605e-04 4.144702e-04 3.477708e-04 [19] 2.920540e-04 2.866941e-04 2.245915e-08 2.896050e-09 5.610856e-18 2.122652e-18 3.862830e-19 1.115808e-19 -9.999266e-19 [28] -3.026382e-18 -7.703938e-18
So you can see that the eigen values are technically negative, but very close to zero. My question is then this: would it be OK if mcse.multi just replaced those values with zeros and carried on? I'm aware that zero/and or eigen values are a sign of perfect correlations in the matrix, but this is not the case here, it's just a matter of estimating accuracy, I guess. I have tried creating a much larger mcmc run by re-sampling the matrix rows (after burn-in/warmup) but I still get the same thing. Also tried removing a bunch of columns (parameters) but it mostly doesn't work.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/statvats/mcmcse/issues/1, or mute the threadhttps://github.com/notifications/unsubscribe-auth/ABSZGSBRDXETSAPVRS4UOE3PTM5D5ANCNFSM4HKFBWEA.
-- Dootika Vats NSF Postdoctoral Fellow Department of Statistics University of Warwick http://warwick.ac.uk/dvats
maxbiostat
First of all, thanks for this excellent package. What follows is not a complaint but a doubt. I'm trying to compute
mESSfor a particular MCMC run and I keep getting the warning message:You either need more samples or x is not full column rank, which ismcse.multitrying to tell me something is off with the covariance matrix. When I compute the eigen values, I get something like this:So you can see that the eigen values are technically negative, but very close to zero. My question is then this: would it be OK if
mcse.multijust replaced those values with zeros and carried on? I'm aware that zero/and or eigen values are a sign of perfect correlations in the matrix, but this is not the case here, it's just a matter of estimating accuracy, I guess. I have tried creating a much larger mcmc run by re-sampling the matrix rows (after burn-in/warmup) but I still get the same thing. Also tried removing a bunch of columns (parameters) but it mostly doesn't work.