Hi, I am writing a model where I need to sample N different MVNormal variables, which all have the same shape, but are each parameterized by a different vector of means and covariance matrix. My code right now looks something like this: for n in range(N): pm.MvNormal(f"x_{n}", mu = my_mea…

How big are we talking for N and q? A bunch of independent MvN distributions is equivalent to a single MvN, if you stack up the means and concatenate the covariance matrices block-diagonally. It might get slow if N * q is too huge, though.

Thanks! q is relatively small (< 5), but N is fairly big (several thousand observations). Do you think that is probably too big? If not, is there an efficient way to convert my (N,q,q) tensor into the block-diagonal format?

I wrote this code for doing it a while ago, but maybe kroneker products is a better choice, like this: # Make a set of (N, N) matrices of all zeros with a 1 in the (i,i)th position indicators = [np.zeros((5, 5)) for _ in range(5)] for i in range(5): indicators[i][i, i] = 1 with pm.Model() as m…

[image] jessegrabowski: # pt.einsum when? Whenever someone takes Implement `einsum` equivalent · Issue #57 · pymc-devs/pytensor · GitHub :smiley:

Can the logp of that model be evaluated, or does it trigger: Increase support for batched multivariate distributions · Issue #5383 · pymc-devs/pymc · GitHub ?

It samples without error (the kron one) edit: I think I misunderstood @ricardoV94 ’s question; the model with a single MvN and batch dimension doesn’t sample and runs into the linked issue

Thanks so much! I’ll give this a try and see if it works for my case. :slightly_smiling_face:

Does this mean it accepts a list of covariance matrices for cov? with pm.Model() as model: nu = pm.Truncated("nu", pm.Exponential.dist(lam=1/10), lower=2) sd_dist = pm.Exponential.dist(lam=2, shape=2) chol, corr, stds = pm.LKJCholeskyCov('chol_cov', n=2, eta=1, sd_dist=sd_d…

It accepts an ndarray with dimensions (..., m, m). Not sure what .tolist() does, but if you wrap it in np.asarray and make sure the core dims are the rightmost ones, it will work.

Is it possible to declare multiple MvNormal variables with different covariance matrices in a vectorized way?

Questions v5

ricardoV94 October 22, 2023, 4:01pm 9

If anyone stumbles upon this thread, the MvNormal now accepts arbitrary batched params.

Lkjcorr returning error

Topic		Replies	Views
About set of Multivariate Normal Distributions Questions	2	1196	July 3, 2018
Draw number of MvNormal in once Questions	2	587	April 5, 2019
Multivariate normal with diagonal covariance Questions	5	3119	March 20, 2019
Multivariable observables with error covariance Questions	3	457	October 28, 2018
Matrix \mu in multivariate normal Questions	2	607	May 10, 2019

Is it possible to declare multiple MvNormal variables with different covariance matrices in a vectorized way?

Related topics