Repository navigation
Feature request: column/row_sum_to_zero_matrix, sum_to_zero_row_vector #3354
Description
Activity
These would be reasonably easy to implement in Stan given the math implementations we already have.
But, it would be reasonably easy to do this yourself now too:parameters { vector[N-1] raw; } transformed parameters { row_vector[N] sums_to_zero = sum_to_zero_jacobian(raw)'; }
The matrix versions would be essentially the same but would need a loop.
The only differences between this and what we would supply built-in if we did are:
- Extra IO ('raw' would also get written out)
- User provided initialization is a bit more annoying, since an initial value would need to be specified for
raw, not forsums_to_zero.
If we had this framework a few versions ago we may not have added row/column stochastic, since they're similarly easy to do yourself. Doubly stochastic matrices or our current sum_to_zero_matrix that have the constraint in both axes are a different transform that cannot easily be built out of smaller parts
Hi Brian,
Thanks for the explanation. I can see the logic about removing the row/column stochastic matrices and just using a transform in the transformed parameters block. The downside of course is that it's more verbose and required two objects, one in parameters and one in transformed, but I'll leave those choices up to you guys. Thanks again!
Matt
I can see the logic about removing the row/column stochastic matrices and just using a transform in the transformed parameters block.
I think this was just an explanation for why we didn't add these transforms. I think they could be useful and would be useful to add, because it's both error prone and tedious and also not the best performance to write this all yourself in transformed parameters (or as local variables in the transformed parameter blocks).
It would help if we had these, which I think we should also add:
matrix to_matrix(array[] vector x); matrix to_matrix(array[] row_vector u); array[] vector to_vector_array(matrix y); array[] row_vector to_row_vector_array(matrix v);
Hey Bob,
The second function
matrix to_matrix(array[] row_vector u);
I'm pretty sure exists and works. This is another reason by
sum_to_zero_row_vectorwould be useful because you could at least create a matrix of sum to zero rows with one function.
Hi everyone,
The new
sum_to_zero_vectorand*_stochastic_matrixtypes are really useful for my work. The currentsum_to_zero_matrixensures both row and column sums are 0. Is it possible to getsum_to_zero_matrixtypes where only the rows or columns sum to zero, like thestochastic_matrixtypes?Additionally, I think a
sum_to_zero_row_vectoralso makes sense!Cheers,
Matt