I have some data, which looks like: curr_A, curr_B, curr_C, curr_D, curr_E, days, end_1, end_2, end_3 1, 0, 0, 0, 0, 1, 0.3, 0.7, 0.0 1, 0, 0, 0, 0, 2, 0.2, 0.6, 0.2 My input variables are curr_A...curr_E and days. While my outp…

There are three possible issues here. First, the softmax function is only defined for vectors, not row vectors. So if you want to end up with a row vector, then you need to transpose the inside of the function call and then transpose the result: pred[n] = softmax((X[n] * C + (X[n] * days[n]) * B)')…

The cause of the specific error seems to be that the dirichlet_lpdf has type real, so I’m guessing if you initialize real pred[K] this error would go away. A simplex type variable should not be needed for a Dirichlet prior, but maybe it’s recommended as it would enforce the correct constraint on th…

Hello Jacob. Premise: I did not study your model well, I am just referring to the dirichlet part. Few points, you cannot transform any parameter in a non-linear way (e.g. softmax, log etc…) without Jacobian correction When you tranform any variable with N degrees of freedom with softmax, you get…

[image] andrjohns: A rough implementation of this is: for(n in 1:N) { matrix[K,K] jacobian; for(i in 1:K) { for(j in 1:K) { jacobian[i,j] = pred[n][i] * (ident[i,j] - pred[n][j]); } } target += log_determinant(jacobian); } Nice! what do you mean by rou…

Oh just that it might not be the most efficient implementation with the nested loops

[image] andrjohns: Putting this all together (into untested code, so no promises), you end up with I woudl like to point out that the issue with indeterminability remains. Unless I am missing something.

Hi @stemangiola , what would you recommend as an alternative to Dirichlet distribution? I considered it, because unlike categorical, or multinomial, it should handle proportions better. If this premise is faulty, please advise. My data is built in terms of proportions, so in your example of softmax(…

[image] Jacob_Moore: ipt * C + (ipt * days[n]) * B is X a matrix of simplexes? Which variable includes end_1, end_2, end_3? So, just at looking at the formula I see softmax(real_unconstrained * real_unconstrained + (real_unconstrained * int) * real_unconstrained ) For sure C and B are u…

How to properly use Simplex with Dirichlet likelihood?

Modeling

stemangiola November 2, 2020, 7:21am 7

I see, rough was not about correctness (this would make the aswer to one post of mine Jacobian of softmax tansformation of a (n-1 degrees of freedom) unbounded parameter).

Topic		Replies	Views
Simplex Regression Modeling	11	2739	December 14, 2017
Prior for Simplex, more informative than Dirichlet Modeling	9	710	February 26, 2024
Informative prior for simplex Modeling	10	309	December 1, 2025
Following up on several discussions of simplex adjustments and ragged arrays of simplexes Modeling techniques , specification , user-defined-functions , constraint-transform	13	353	December 10, 2024
Log-Simplex Constraints Modeling jacobian-adjustment	17	383	June 26, 2025

How to properly use Simplex with Dirichlet likelihood?

Related topics