I wrote a thing about multivariate normal sampling in #rstats, and why you would be well-advised to prefer mvtnorm::rmvnorm() over MASS::mvrnorm() Now. Let us never speak of this again - ThreadSky

djnavarro.net • 15 days ago

I wrote a thing about multivariate normal sampling in #rstats, and why you would be well-advised to prefer mvtnorm::rmvnorm() over MASS::mvrnorm()

Now. Let us never speak of this again

Comments

tiddlydump.bsky.social•15 days ago

There's gold in them footnotes!

dgbassani.bsky.social•15 days ago

Indeed

alexpghayes.com•13 days ago

Living for the footnotes. Really nice post!

jcsalterego.bsky.social•15 days ago

@dopestatistics.bsky.social

tjmahr.com•14 days ago

"using whatever fancy matrix decomposition trick is still capable of bringing love into in your withered heart"... this is wonderful

djnavarro.net•14 days ago

that’s my favourite line too 🫠

vincentab.bsky.social•15 days ago

This was super fun and informative. Thanks for writing and posting!

bwiernik.bsky.social•15 days ago

Whenever there are essays about a topic, Danielle’s are the most fun!

ojala.bsky.social•15 days ago

You transported me back to 1996-7 when I was implementing multivariate simulations using Cholesky in Fortran 90 for my PhD. My supervisor told me it was the default way when fluffing around in quantitative genetics.

averageatheist.bsky.social•14 days ago

Ah, for the good old days when Mersenne Twister was the state of the art.

djnavarro.net•14 days ago

I mean, as far as I know MT numbers are still fine, but oh boy it’s horrifying how downstream floating point errors can mess up what you do with them sometimes 😬

jverissimo.bsky.social•14 days ago

Interesting... Tried mgcv::rmvn() and mvnfast::rmvn(). Both seems to produce the same random numbers (with and without the perturbations in the cov matrix).

drandreahoward.bsky.social•12 days ago

📌

akjackson.bsky.social•15 days ago

I had a brilliant colleague (he calculated the water wave for the Bikini Atoll tests on a slide rule) who wrote some 2D finite difference code. I mentioned that he must have done a lot of double precision math. Hurt, he replied "Oh no. If you know what you're doing, single precision is enough".

dsmccormick.bsky.social•14 days ago

And I thought of the great Jon Bon Neumann quote: “Anyone who attempts to generate random numbers by deterministic means is, of course, living in a state of sin.”

dsmccormick.bsky.social•14 days ago

That was a complete tour de force of yak shaving.

strengejacke.bsky.social•14 days ago

How to credit your transformer/sampler functions when I'd like to use it in an R package?

djnavarro.net•14 days ago

Anything is fine, but also you don’t need to credit me at all, feel free to reuse as you like. The only thing I’d say is that if you’re planning on using it in a production context please be careful. I haven’t tested it any further than I did in the post

strengejacke.bsky.social•14 days ago

But such a long post comes close to an equivalent of > 95% test coverage :-)
Anyway, thanks, I'll try the code and see if it works well! Maybe it can indeed replace the (few) instances where we currently use `MASS::mvrnorm()` (and if so, I'd link to your post)

richarddmorey.bsky.social•14 days ago

Is it wrong to want to run Wilkinson (1965) through the NotebookLM podcast generator to hear them excidedly talking about matrix decompositions?

djnavarro.net•14 days ago

it would be delightful

richarddmorey.bsky.social•14 days ago

econmaett.bsky.social•15 days ago

📌

abumirchi.com•14 days ago

I'm only half-way through, but I don't believe I saw the Eigen values of the two different covariance matrices. Were close to zero?

abumirchi.com•14 days ago

Ah never mind. I'm now down to condition numbers don't matter which, I must say I was kind of surprised by.

djnavarro.net•14 days ago

I found that surprising too. Even after reading Wilkinsons warnings about condition numbers I really would have expected that the cases where MASS produces the strange behaviour would tend to have higher condition numbers but… nope, apparently not. It’s weird

abumirchi.com•14 days ago

Honestly I was sure the eigenvalues were causing problems but they were all pretty large. I started using MASS back when I started using R back in 2015. I never really thought about it. I'm going to see if I can switch over. It's all about that 20 package CRAN limit.
https://github.com/john-harrold/ubiquity/issues/27

abumirchi.com•14 days ago

By the way have you brought this up with Brian Ripley?

djnavarro.net•14 days ago

lol, i have had exactly one interaction with Brian Ripley in my life and it is an experience i am very keen never to have again

Comments

Posting Rules

Reply