Inconsistency in vectorisation of multivariate distributions

[robjhyndman]

library(distributional)
mu <- c(0, 0)
Sigma <- diag(2)
dist <- distributional::dist_multivariate_normal(list(mu), list(Sigma))

density(dist, at = c(0,0))
#> [[1]]
#> [1] 0.1591549
density(dist, at = matrix(c(0,0), nrow=1))
#> [1] 0.1591549

^{Created on 2024-07-31 with reprex v2.1.1}

[mitchelloharawild]

I think the first option here should error, since the dimensions of at do not match the dimension of the distribution. What do you think?

The intended vectorisation is described in https://github.com/mitchelloharawild/distributional/issues/52#issuecomment-939891904, although more formal descriptions of how distributional vectorises distributions/operations is needed (#107).

[robjhyndman]

I was thinking a vector should be interpreted as a 1 row matrix, and then the dimensions do match. This is how mvtnorm::dmvnorm() works.

[mitchelloharawild]

For me it's a question of if that default mapping of dimensions is appropriate, and I'm leaning toward being strict and not having a default dimension for univariate inputs -> multivariate distributions (essentially requiring matrix inputs)