Distribution tests - vector parameters

Developers
1

I’m working on writing distribution tests for the ordered logistic distribution, but all of the other tests have scalar input parameters, for example, normal_test.hpp:

    vector<double> param(3);

    param[0] = 0;  // y
    param[1] = 0;  // mu
    param[2] = 1;  // sigma
    parameters.push_back(param);
    log_prob.push_back(-0.918938533204672669541);  // expected log_prob

Whereas I need param[3] to be a vector of doubles.

Is there a different way that I should be parameterising this?

2

What do the existing tests look like? And what about tests for things like multivariate normal, or the Dirichlet, both of which involve vector-only parameters.

3

All of the existing tests are for the univariate distributions. Looking at test/prob:

  • bernoulli
  • beta
  • beta_binomial
  • beta_proportion
  • cauchy
  • chi_square
  • double_exponential
  • exp_mod_normal
  • exponential
  • frechet
  • gamma
  • gumbel
  • hypergeometric
  • inv_chi_square
  • inv_gamma
  • logistic
  • lognormal
  • neg_binomial
  • neg_binomial_2
  • normal
  • normal_sufficient
  • pareto
  • pareto_type_2
  • poisson
  • rayleigh
  • scaled_inv_chi_square
  • skew_normal
  • std_normal
  • student_t
  • uniform binomial
  • von_mises
  • weibull
  • wiener

So there aren’t any distributions with vector parameters being tested

4

There were some that Rob Trangucci wrote; I think they were vectorized.

We hadn’t been able to write a generic framework for these pre C++11. I think we can actually start making a lot of progress now, but that may not help you test things in the short term. I’d start by writing tests manually. And then trying to derive requirements from that.

5

You need to look in test/unit/math/prim/mat/prob/

They’re all being done by hand, which is a mess. I’d suggest at the very least using my new testing framework.