Hey all, I’m seeing a weird issue where my Stan model fails to initialize with a very simple model: parameters { cholesky_factor_corr[100] ccm; } model { ccm ~ lkj_corr_cholesky(1); } because lkj_corr_cholesky(1); on any of the constrained cholesky_factor_corr matrices returns -nan, I think b…

For Cholesky factors that big, it is possible that the diagonal elements result in the square root of a slightly negative number when symbolically they should be slightly positive. If you search for the “onion method” on the Stan-users Google group, there is some code for a more numerically stable w…

-nan results with lkj_corr_cholesky

Developers

bgoodri July 22, 2020, 1:35am 3

onion.stan (1.1 KB) is the 2020 version.

4 Likes

Using the onion method in occupancy hierarchical model

Topic		Replies	Views
Cholesky_factor_corr initializing with invalid values Developers	8	2935	July 14, 2021
Error when specfying initial values for cholesky factor CmdStan	1	376	April 14, 2020
Bad LKJ Initialization in model from "Fast hierarchical Gaussian processes" Modeling specification	2	1483	March 22, 2018
Error transforming variable L: lub_free: Correlation variable is nan, but must be in the interval [-1, 1] General	3	730	April 15, 2019
Sparse sampling for a Cholesky correlation matrix (or modelling correlations among some, but not all parameters) Modeling	15	1309	February 1, 2019

-nan results with lkj_corr_cholesky

Related topics