Repository navigation
Feature request: row_centred_matrix and column_centred_matrix #3361
Description
Activity
This would be a simple contribution, very similar to the row/column stochastic types:
https://github.com/stan-dev/math/blob/develop/stan/math/prim/constraint/stochastic_row_constrain.hpp
https://github.com/stan-dev/math/blob/develop/stan/math/rev/constraint/stochastic_row_constrain.hpp
https://github.com/stan-dev/math/blob/develop/stan/math/prim/constraint/stochastic_row_free.hppIf you need a matrix where the rows sum to zero, then you can do this quite directly with our new built-in transform functions with Jacobians. Here's a user-defined function to do this along with a trivial program with a proper prior to run it.
functions { matrix sum_to_zero_rows_jacobian(array[] vector alpha_raw) { int M = size(alpha_raw); if (M == 0) reject("require M > 0, found M = ", M); int N = rows(alpha_raw[1]) + 1; matrix[M, N] alpha; for (m in 1:M) { alpha[m] = sum_to_zero_jacobian(alpha_raw[m])'; } return alpha; } } data { int<lower=1> M; int<lower=1> N; } parameters { array[M] vector[N - 1] alpha_raw; } transformed parameters { matrix[M, N] alpha = sum_to_zero_rows_jacobian(alpha_raw); // flat over sum-to-zero row matrices } model { to_vector(alpha) ~ normal(0, 1); }It'll get the job done, but it's definitely not as convenient as having a
sum_to_zero_row_matrixtype. It wouldn't be that hard to add the type to the language. A non-optimally-performance-tuned version would just copy this to C++ rather than collapse the autodiff tree (not a really huge savings here). But then there's writing the unit tests, expanding the reference manual chapter on transforms, and the user's guide lists of types. So overall, a bit of work.It would be nice if we had a way to declare a matrix with constraints on the rows or columns. As is, we have to multiply built-in types in a really clunky way.
Also, we'd use the American spellings for consistency :-).
- marked Feature request: column/row_sum_to_zero_matrix, sum_to_zero_row_vector stan#3354 as a duplicate of this issue
on Sep 2, 2026
To facilitate working with the sum-to-zero constraint in multivariate contexts, it would be valuable to be able to specify not only individual
sum_to_zero_vectors but also matrices for which all rows or all columns are constrained to sum to zero. To my knowledge, there is no universally standard name for such a mathematical object (nor to Claude's knowledge, which only produced a list of suggestions), but for the sake of starting a conversation, I would propose something likerow_centered_matrixandcol_centered_matrix(and perhapsrow_centred_matrixandcol_centred_matrixas aliases for spelling-dialect neutrality).The primary use case for me would be hierarchical modelling with correlated multivariate random effects, e.g., random-slopes models or multivariate item-response models. To reduce problems with funnel geometries during sampling, it can be wise to write such models using ‘raw’
parameterswith standard normal priors and then construct the actual parameters in thetransformed parametersblock via matrix multiplication with standard-deviation and correlation-matrix hyperpriors (e.g., the User's Guide chapter on ‘Regression Models’).That's awkward to do using the
sum_to_zero_vectortype alone. TheNraw parameter vectors (let's say of dimensionK) have to be stored asarray[K] sum_to_zero_vector[N]. This structure needs to be pulled apart and restructured asmatrix[K, N]orarray[N] vector[K]intransformed parametersin order to be multiplied by acholesky_factor_corr[K]. In practice, the problem is mostly that this restructuring step is ugly and confusing to read, but in theory, it is also memory inefficient and could cause slowdowns at large model scale.