|
| 1 | +#ifndef STAN_MATH_PRIM_PROB_BETA_NEG_BINOMIAL_LCDF_HPP |
| 2 | +#define STAN_MATH_PRIM_PROB_BETA_NEG_BINOMIAL_LCDF_HPP |
| 3 | + |
| 4 | +#include <stan/math/prim/meta.hpp> |
| 5 | +#include <stan/math/prim/err.hpp> |
| 6 | +#include <stan/math/prim/fun/constants.hpp> |
| 7 | +#include <stan/math/prim/fun/digamma.hpp> |
| 8 | +#include <stan/math/prim/fun/hypergeometric_3F2.hpp> |
| 9 | +#include <stan/math/prim/fun/grad_F32.hpp> |
| 10 | +#include <stan/math/prim/fun/lbeta.hpp> |
| 11 | +#include <stan/math/prim/fun/lgamma.hpp> |
| 12 | +#include <stan/math/prim/fun/max_size.hpp> |
| 13 | +#include <stan/math/prim/fun/scalar_seq_view.hpp> |
| 14 | +#include <stan/math/prim/fun/size.hpp> |
| 15 | +#include <stan/math/prim/fun/size_zero.hpp> |
| 16 | +#include <stan/math/prim/functor/partials_propagator.hpp> |
| 17 | +#include <cmath> |
| 18 | + |
| 19 | +namespace stan { |
| 20 | +namespace math { |
| 21 | + |
| 22 | +/** \ingroup prob_dists |
| 23 | + * Returns the log CDF of the Beta-Negative Binomial distribution with given |
| 24 | + * number of successes, prior success, and prior failure parameters. |
| 25 | + * Given containers of matching sizes, returns the log sum of probabilities. |
| 26 | + * |
| 27 | + * @tparam T_n type of failure parameter |
| 28 | + * @tparam T_r type of number of successes parameter |
| 29 | + * @tparam T_alpha type of prior success parameter |
| 30 | + * @tparam T_beta type of prior failure parameter |
| 31 | + * |
| 32 | + * @param n failure parameter |
| 33 | + * @param r Number of successes parameter |
| 34 | + * @param alpha prior success parameter |
| 35 | + * @param beta prior failure parameter |
| 36 | + * @param precision precision for `grad_F32`, default \f$10^{-8}\f$ |
| 37 | + * @param max_steps max iteration allowed for `grad_F32`, default \f$10^{8}\f$ |
| 38 | + * @return log probability or log sum of probabilities |
| 39 | + * @throw std::domain_error if r, alpha, or beta fails to be positive |
| 40 | + * @throw std::invalid_argument if container sizes mismatch |
| 41 | + */ |
| 42 | +template <typename T_n, typename T_r, typename T_alpha, typename T_beta> |
| 43 | +inline return_type_t<T_r, T_alpha, T_beta> beta_neg_binomial_lcdf( |
| 44 | + const T_n& n, const T_r& r, const T_alpha& alpha, const T_beta& beta, |
| 45 | + const double precision = 1e-8, const int max_steps = 1e8) { |
| 46 | + static constexpr const char* function = "beta_neg_binomial_lcdf"; |
| 47 | + check_consistent_sizes( |
| 48 | + function, "Failures variable", n, "Number of successes parameter", r, |
| 49 | + "Prior success parameter", alpha, "Prior failure parameter", beta); |
| 50 | + if (size_zero(n, r, alpha, beta)) { |
| 51 | + return 0; |
| 52 | + } |
| 53 | + |
| 54 | + using T_r_ref = ref_type_t<T_r>; |
| 55 | + T_r_ref r_ref = r; |
| 56 | + using T_alpha_ref = ref_type_t<T_alpha>; |
| 57 | + T_alpha_ref alpha_ref = alpha; |
| 58 | + using T_beta_ref = ref_type_t<T_beta>; |
| 59 | + T_beta_ref beta_ref = beta; |
| 60 | + check_positive_finite(function, "Number of successes parameter", r_ref); |
| 61 | + check_positive_finite(function, "Prior success parameter", alpha_ref); |
| 62 | + check_positive_finite(function, "Prior failure parameter", beta_ref); |
| 63 | + |
| 64 | + scalar_seq_view<T_n> n_vec(n); |
| 65 | + scalar_seq_view<T_r_ref> r_vec(r_ref); |
| 66 | + scalar_seq_view<T_alpha_ref> alpha_vec(alpha_ref); |
| 67 | + scalar_seq_view<T_beta_ref> beta_vec(beta_ref); |
| 68 | + int size_n = stan::math::size(n); |
| 69 | + size_t max_size_seq_view = max_size(n, r, alpha, beta); |
| 70 | + |
| 71 | + // Explicit return for extreme values |
| 72 | + // The gradients are technically ill-defined, but treated as zero |
| 73 | + for (int i = 0; i < size_n; i++) { |
| 74 | + if (n_vec.val(i) < 0) { |
| 75 | + return negative_infinity(); |
| 76 | + } |
| 77 | + } |
| 78 | + |
| 79 | + using T_partials_return = partials_return_t<T_n, T_r, T_alpha, T_beta>; |
| 80 | + T_partials_return log_cdf(0.0); |
| 81 | + auto ops_partials = make_partials_propagator(r_ref, alpha_ref, beta_ref); |
| 82 | + for (size_t i = 0; i < max_size_seq_view; i++) { |
| 83 | + // Explicit return for extreme values |
| 84 | + // The gradients are technically ill-defined, but treated as zero |
| 85 | + if (n_vec.val(i) == std::numeric_limits<int>::max()) { |
| 86 | + return 0.0; |
| 87 | + } |
| 88 | + auto n_dbl = n_vec.val(i); |
| 89 | + auto r_dbl = r_vec.val(i); |
| 90 | + auto alpha_dbl = alpha_vec.val(i); |
| 91 | + auto beta_dbl = beta_vec.val(i); |
| 92 | + auto b_plus_n = beta_dbl + n_dbl; |
| 93 | + auto r_plus_n = r_dbl + n_dbl; |
| 94 | + auto a_plus_r = alpha_dbl + r_dbl; |
| 95 | + using a_t = return_type_t<decltype(b_plus_n), decltype(r_plus_n)>; |
| 96 | + using b_t = return_type_t<decltype(n_dbl), decltype(a_plus_r), |
| 97 | + decltype(b_plus_n)>; |
| 98 | + auto F = hypergeometric_3F2( |
| 99 | + std::initializer_list<a_t>{1.0, b_plus_n + 1.0, r_plus_n + 1.0}, |
| 100 | + std::initializer_list<b_t>{n_dbl + 2.0, a_plus_r + b_plus_n + 1.0}, |
| 101 | + 1.0); |
| 102 | + auto C = lgamma(r_plus_n + 1.0) + lbeta(a_plus_r, b_plus_n + 1.0) |
| 103 | + - lgamma(r_dbl) - lbeta(alpha_dbl, beta_dbl) - lgamma(n_dbl + 2.0); |
| 104 | + auto ccdf = stan::math::exp(C) * F; |
| 105 | + log_cdf += log1m(ccdf); |
| 106 | + |
| 107 | + if constexpr (!is_constant_all<T_r, T_alpha, T_beta>::value) { |
| 108 | + auto chain_rule_term = -ccdf / (1.0 - ccdf); |
| 109 | + auto digamma_n_r_alpha_beta = digamma(a_plus_r + b_plus_n + 1.0); |
| 110 | + T_partials_return dF[6]; |
| 111 | + grad_F32<false, !is_constant<T_beta>::value, !is_constant_all<T_r>::value, |
| 112 | + false, true, false>(dF, 1.0, b_plus_n + 1.0, r_plus_n + 1.0, |
| 113 | + n_dbl + 2.0, a_plus_r + b_plus_n + 1.0, 1.0, |
| 114 | + precision, max_steps); |
| 115 | + |
| 116 | + if constexpr (!is_constant<T_r>::value || !is_constant<T_alpha>::value) { |
| 117 | + auto digamma_r_alpha = digamma(a_plus_r); |
| 118 | + if constexpr (!is_constant<T_r>::value) { |
| 119 | + auto partial_lccdf = digamma(r_plus_n + 1.0) |
| 120 | + + (digamma_r_alpha - digamma_n_r_alpha_beta) |
| 121 | + + (dF[2] + dF[4]) / F - digamma(r_dbl); |
| 122 | + partials<0>(ops_partials)[i] += partial_lccdf * chain_rule_term; |
| 123 | + } |
| 124 | + if constexpr (!is_constant<T_alpha>::value) { |
| 125 | + auto partial_lccdf = digamma_r_alpha - digamma_n_r_alpha_beta |
| 126 | + + dF[4] / F - digamma(alpha_dbl); |
| 127 | + partials<1>(ops_partials)[i] += partial_lccdf * chain_rule_term; |
| 128 | + } |
| 129 | + } |
| 130 | + |
| 131 | + if constexpr (!is_constant<T_alpha>::value |
| 132 | + || !is_constant<T_beta>::value) { |
| 133 | + auto digamma_alpha_beta = digamma(alpha_dbl + beta_dbl); |
| 134 | + if constexpr (!is_constant<T_alpha>::value) { |
| 135 | + partials<1>(ops_partials)[i] += digamma_alpha_beta * chain_rule_term; |
| 136 | + } |
| 137 | + if constexpr (!is_constant<T_beta>::value) { |
| 138 | + auto partial_lccdf = digamma(b_plus_n + 1.0) - digamma_n_r_alpha_beta |
| 139 | + + (dF[1] + dF[4]) / F |
| 140 | + - (digamma(beta_dbl) - digamma_alpha_beta); |
| 141 | + partials<2>(ops_partials)[i] += partial_lccdf * chain_rule_term; |
| 142 | + } |
| 143 | + } |
| 144 | + } |
| 145 | + } |
| 146 | + |
| 147 | + return ops_partials.build(log_cdf); |
| 148 | +} |
| 149 | + |
| 150 | +} // namespace math |
| 151 | +} // namespace stan |
| 152 | +#endif |
0 commit comments