much improved chi_squared_quantile, incremental update

release/4.3a0
Varun Agrawal 2023-10-20 07:00:34 -04:00
parent a46a78d413
commit c8a0cdc543
3 changed files with 94 additions and 441 deletions

View File

@ -36,7 +36,7 @@ namespace gtsam {
* Equivalent to chi2inv in Matlab.
*/
static double Chi2inv(const double alpha, const size_t dofs) {
return chi_squared_quantile(dofs, alpha);
return internal::chi_squared_quantile(dofs, alpha);
}
/* ************************************************************************* */

View File

@ -11,480 +11,133 @@
/**
* @file chiSquaredInverse.h
* @brief Helper functions for use with the GncOptimizer
* @brief This file contains an implementation of the Chi Squared inverse
* function, which is implemented similar to Boost with additional template
* parameter helpers.
*
* A lot of this code has been picked up from
* https://www.boost.org/doc/libs/1_83_0/boost/math/special_functions/detail/igamma_inverse.hpp
* https://www.boost.org/doc/libs/1_83_0/boost/math/tools/roots.hpp
*
* @author Varun Agrawal
*/
#pragma once
#include <gtsam/nonlinear/internal/gamma.h>
#include <gtsam/nonlinear/internal/halley.h>
#include <gtsam/nonlinear/internal/utils.h>
#include <algorithm>
// TODO(Varun) remove
#include <boost/math/distributions/gamma.hpp>
namespace gtsam {
/// Template type for numeric limits
template <class T>
using LIM = std::numeric_limits<T>;
template <typename T>
using return_t =
typename std::conditional<std::is_integral<T>::value, double, T>::type;
/// Get common type amongst all arguments
template <typename... T>
using common_t = typename std::common_type<T...>::type;
/// Helper template for finding common return type
template <typename... T>
using common_return_t = return_t<common_t<T...>>;
/// Check if integer is odd
constexpr bool is_odd(const long long int x) noexcept { return (x & 1U) != 0; }
/// Templated check for NaN
template <typename T>
constexpr bool is_nan(const T x) noexcept {
return x != x;
}
/// @brief Gauss-Legendre quadrature: 50 points
static const long double gauss_legendre_50_points[50] = {
-0.03109833832718887611232898966595L, 0.03109833832718887611232898966595L,
-0.09317470156008614085445037763960L, 0.09317470156008614085445037763960L,
-0.15489058999814590207162862094111L, 0.15489058999814590207162862094111L,
-0.21600723687604175684728453261710L, 0.21600723687604175684728453261710L,
-0.27628819377953199032764527852113L, 0.27628819377953199032764527852113L,
-0.33550024541943735683698825729107L, 0.33550024541943735683698825729107L,
-0.39341431189756512739422925382382L, 0.39341431189756512739422925382382L,
-0.44980633497403878914713146777838L, 0.44980633497403878914713146777838L,
-0.50445814490746420165145913184914L, 0.50445814490746420165145913184914L,
-0.55715830451465005431552290962580L, 0.55715830451465005431552290962580L,
-0.60770292718495023918038179639183L, 0.60770292718495023918038179639183L,
-0.65589646568543936078162486400368L, 0.65589646568543936078162486400368L,
-0.70155246870682225108954625788366L, 0.70155246870682225108954625788366L,
-0.74449430222606853826053625268219L, 0.74449430222606853826053625268219L,
-0.78455583290039926390530519634099L, 0.78455583290039926390530519634099L,
-0.82158207085933594835625411087394L, 0.82158207085933594835625411087394L,
-0.85542976942994608461136264393476L, 0.85542976942994608461136264393476L,
-0.88596797952361304863754098246675L, 0.88596797952361304863754098246675L,
-0.91307855665579189308973564277166L, 0.91307855665579189308973564277166L,
-0.93665661894487793378087494727250L, 0.93665661894487793378087494727250L,
-0.95661095524280794299774564415662L, 0.95661095524280794299774564415662L,
-0.97286438510669207371334410460625L, 0.97286438510669207371334410460625L,
-0.98535408404800588230900962563249L, 0.98535408404800588230900962563249L,
-0.99403196943209071258510820042069L, 0.99403196943209071258510820042069L,
-0.99886640442007105018545944497422L, 0.99886640442007105018545944497422L};
/// @brief Gauss-Legendre quadrature: 50 weights
static const long double gauss_legendre_50_weights[50] = {
0.06217661665534726232103310736061L, 0.06217661665534726232103310736061L,
0.06193606742068324338408750978083L, 0.06193606742068324338408750978083L,
0.06145589959031666375640678608392L, 0.06145589959031666375640678608392L,
0.06073797084177021603175001538481L, 0.06073797084177021603175001538481L,
0.05978505870426545750957640531259L, 0.05978505870426545750957640531259L,
0.05860084981322244583512243663085L, 0.05860084981322244583512243663085L,
0.05718992564772838372302931506599L, 0.05718992564772838372302931506599L,
0.05555774480621251762356742561227L, 0.05555774480621251762356742561227L,
0.05371062188899624652345879725566L, 0.05371062188899624652345879725566L,
0.05165570306958113848990529584010L, 0.05165570306958113848990529584010L,
0.04940093844946631492124358075143L, 0.04940093844946631492124358075143L,
0.04695505130394843296563301363499L, 0.04695505130394843296563301363499L,
0.04432750433880327549202228683039L, 0.04432750433880327549202228683039L,
0.04152846309014769742241197896407L, 0.04152846309014769742241197896407L,
0.03856875661258767524477015023639L, 0.03856875661258767524477015023639L,
0.03545983561514615416073461100098L, 0.03545983561514615416073461100098L,
0.03221372822357801664816582732300L, 0.03221372822357801664816582732300L,
0.02884299358053519802990637311323L, 0.02884299358053519802990637311323L,
0.02536067357001239044019487838544L, 0.02536067357001239044019487838544L,
0.02178024317012479298159206906269L, 0.02178024317012479298159206906269L,
0.01811556071348939035125994342235L, 0.01811556071348939035125994342235L,
0.01438082276148557441937890892732L, 0.01438082276148557441937890892732L,
0.01059054838365096926356968149924L, 0.01059054838365096926356968149924L,
0.00675979919574540150277887817799L, 0.00675979919574540150277887817799L,
0.00290862255315514095840072434286L, 0.00290862255315514095840072434286L};
namespace internal {
/// 50 point Gauss-Legendre quadrature
template <typename T>
constexpr T incomplete_gamma_quad_inp_vals(const T lb, const T ub,
const int counter) noexcept {
return (ub - lb) * gauss_legendre_50_points[counter] / T(2) +
(ub + lb) / T(2);
}
template <typename T>
constexpr T incomplete_gamma_quad_weight_vals(const T lb, const T ub,
const int counter) noexcept {
return (ub - lb) * gauss_legendre_50_weights[counter] / T(2);
}
template <typename T>
constexpr T incomplete_gamma_quad_fn(const T x, const T a,
const T lg_term) noexcept {
return exp(-x + (a - T(1)) * log(x) - lg_term);
}
template <typename T>
constexpr T incomplete_gamma_quad_recur(const T lb, const T ub, const T a,
const T lg_term,
const int counter) noexcept {
T val = incomplete_gamma_quad_fn(
incomplete_gamma_quad_inp_vals(lb, ub, counter), a, lg_term) *
incomplete_gamma_quad_weight_vals(lb, ub, counter);
if (counter < 49) {
return val + incomplete_gamma_quad_recur(lb, ub, a, lg_term, counter + 1);
} else {
return val;
}
}
template <typename T>
constexpr T incomplete_gamma_quad_lb(const T a, const T z) noexcept {
// break integration into ranges
return (a > T(1000) ? std::max(T(0), std::min(z, a) - 11 * sqrt(a))
: a > T(800) ? std::max(T(0), std::min(z, a) - 11 * sqrt(a))
: a > T(500) ? std::max(T(0), std::min(z, a) - 10 * sqrt(a))
: a > T(300) ? std::max(T(0), std::min(z, a) - 10 * sqrt(a))
: a > T(100) ? std::max(T(0), std::min(z, a) - 9 * sqrt(a))
: a > T(90) ? std::max(T(0), std::min(z, a) - 9 * sqrt(a))
: a > T(70) ? std::max(T(0), std::min(z, a) - 8 * sqrt(a))
: a > T(50) ? std::max(T(0), std::min(z, a) - 7 * sqrt(a))
: a > T(40) ? std::max(T(0), std::min(z, a) - 6 * sqrt(a))
: a > T(30) ? std::max(T(0), std::min(z, a) - 5 * sqrt(a))
: std::max(T(0), std::min(z, a) - 4 * sqrt(a)));
}
template <typename T>
constexpr T incomplete_gamma_quad_ub(const T a, const T z) noexcept {
return (a > T(1000) ? std::min(z, a + 10 * sqrt(a))
: a > T(800) ? std::min(z, a + 10 * sqrt(a))
: a > T(500) ? std::min(z, a + 9 * sqrt(a))
: a > T(300) ? std::min(z, a + 9 * sqrt(a))
: a > T(100) ? std::min(z, a + 8 * sqrt(a))
: a > T(90) ? std::min(z, a + 8 * sqrt(a))
: a > T(70) ? std::min(z, a + 7 * sqrt(a))
: a > T(50) ? std::min(z, a + 6 * sqrt(a))
: std::min(z, a + 5 * sqrt(a)));
}
template <typename T>
constexpr T incomplete_gamma_quad(const T a, const T z) noexcept {
return incomplete_gamma_quad_recur(incomplete_gamma_quad_lb(a, z),
incomplete_gamma_quad_ub(a, z), a,
lgamma(a), 0);
}
// reverse cf expansion
// see: https://functions.wolfram.com/GammaBetaErf/Gamma2/10/0003/
template <typename T>
constexpr T incomplete_gamma_cf_2_recur(const T a, const T z,
const int depth) noexcept {
T val = 1 + (depth - 1) * 2 - a + z;
if (depth < 100) {
return val +
depth * (a - depth) / incomplete_gamma_cf_2_recur(a, z, depth + 1);
} else {
return val;
}
}
template <typename T>
constexpr T incomplete_gamma_cf_2(
const T a,
const T z) noexcept { // lower (regularized) incomplete gamma function
return (T(1.0) - exp(a * log(z) - z - lgamma(a)) /
incomplete_gamma_cf_2_recur(a, z, 1));
}
// cf expansion
// see: http://functions.wolfram.com/GammaBetaErf/Gamma2/10/0009/
template <typename T>
constexpr T incomplete_gamma_cf_1_coef(const T a, const T z,
const int depth) noexcept {
if (is_odd(depth)) {
return -(a - 1 + T(depth + 1) / T(2)) * z;
} else {
return T(depth) / T(2) * z;
}
}
template <typename T>
constexpr T incomplete_gamma_cf_1_recur(const T a, const T z,
const int depth) noexcept {
T val = a + depth - 1;
if (depth < 55) {
return val + incomplete_gamma_cf_1_coef(a, z, depth) /
incomplete_gamma_cf_1_recur(a, z, depth + 1);
} else {
return val;
}
}
/// lower (regularized) incomplete gamma function
template <typename T>
constexpr T incomplete_gamma_cf_1(const T a, const T z) noexcept {
return exp(a * log(z) - z - lgamma(a)) / incomplete_gamma_cf_1_recur(a, z, 1);
}
/// Perform NaN check
template <typename T>
constexpr T incomplete_gamma_check(const T a, const T z) noexcept {
if (is_nan(a) || is_nan(z)) {
return LIM<T>::quiet_NaN();
} else if (a < T(0)) {
return LIM<T>::quiet_NaN();
} else if (LIM<T>::min() > z) {
return T(0);
} else if (LIM<T>::min() > a) {
return T(1);
} else if (a < T(10) && z - a < T(10)) { // cf or quadrature
return incomplete_gamma_cf_1(a, z);
} else if (a < T(10) || z / a > T(3)) {
return incomplete_gamma_cf_2(a, z);
} else {
return incomplete_gamma_quad(a, z);
}
}
template <typename T1, typename T2, typename TC = common_return_t<T1, T2>>
constexpr TC incomplete_gamma_type_check(const T1 a, const T2 p) noexcept {
return incomplete_gamma_check(static_cast<TC>(a), static_cast<TC>(p));
}
} // namespace internal
/**
* Compile-time regularized lower incomplete gamma function
*
* @param a a real-valued, non-negative input.
* @param x a real-valued, non-negative input.
*
* @return the regularized lower incomplete gamma function evaluated at (\c a,
* \c x), \f[ \frac{\gamma(a,x)}{\Gamma(a)} = \frac{1}{\Gamma(a)} \int_0^x
* t^{a-1} \exp(-t) dt \f] When \c a is not too large, the value is computed
* using the continued fraction representation of the upper incomplete gamma
* function, \f$ \Gamma(a,x) \f$, using \f[ \Gamma(a,x) = \Gamma(a) -
* \dfrac{x^a\exp(-x)}{a - \dfrac{ax}{a + 1 + \dfrac{x}{a + 2 - \dfrac{(a+1)x}{a
* + 3 + \dfrac{2x}{a + 4 - \ddots}}}}} \f] where \f$ \gamma(a,x) \f$ and \f$
* \Gamma(a,x) \f$ are connected via \f[ \frac{\gamma(a,x)}{\Gamma(a)} +
* \frac{\Gamma(a,x)}{\Gamma(a)} = 1 \f] When \f$ a > 10 \f$, a 50-point
* Gauss-Legendre quadrature scheme is employed.
*/
template <typename T1, typename T2>
constexpr common_return_t<T1, T2> incomplete_gamma(const T1 a,
const T2 x) noexcept {
return internal::incomplete_gamma_type_check(a, x);
}
namespace internal {
template <typename T>
constexpr T incomplete_gamma_inv_decision(const T value, const T a, const T p,
const T direc, const T lg_val,
const int iter_count) noexcept;
//
// initial value for Halley's method
template <typename T>
constexpr T incomplete_gamma_inv_t_val_1(const T p) noexcept { // a > 1.0
return (p > T(0.5) ? sqrt(-2 * log(T(1) - p)) : sqrt(-2 * log(p)));
}
template <typename T>
constexpr T incomplete_gamma_inv_t_val_2(const T a) noexcept { // a <= 1.0
return (T(1) - T(0.253) * a - T(0.12) * a * a);
}
//
template <typename T>
constexpr T incomplete_gamma_inv_initial_val_1_int_begin(
const T t_val) noexcept { // internal for a > 1.0
return (t_val -
(T(2.515517L) + T(0.802853L) * t_val + T(0.010328L) * t_val * t_val) /
(T(1) + T(1.432788L) * t_val + T(0.189269L) * t_val * t_val +
T(0.001308L) * t_val * t_val * t_val));
}
template <typename T>
constexpr T incomplete_gamma_inv_initial_val_1_int_end(
const T value_inp, const T a) noexcept { // internal for a > 1.0
return std::max(
T(1E-04), a * pow(T(1) - T(1) / (9 * a) - value_inp / (3 * sqrt(a)), 3));
}
template <typename T>
constexpr T incomplete_gamma_inv_initial_val_1(
const T a, const T t_val, const T sgn_term) noexcept { // a > 1.0
return incomplete_gamma_inv_initial_val_1_int_end(
sgn_term * incomplete_gamma_inv_initial_val_1_int_begin(t_val), a);
}
template <typename T>
constexpr T incomplete_gamma_inv_initial_val_2(
const T a, const T p, const T t_val) noexcept { // a <= 1.0
if (p < t_val) {
return pow(p / t_val, T(1) / a);
} else {
return T(1) - log(T(1) - (p - t_val) / (T(1) - t_val));
}
}
// Initial value
template <typename T>
constexpr T incomplete_gamma_inv_initial_val(const T a, const T p) noexcept {
if (a > T(1)) {
return incomplete_gamma_inv_initial_val_1(
a, incomplete_gamma_inv_t_val_1(p), p > T(0.5) ? T(-1) : T(1));
} else {
return incomplete_gamma_inv_initial_val_2(a, p,
incomplete_gamma_inv_t_val_2(a));
}
}
//
// Halley recursion
template <typename T>
constexpr T incomplete_gamma_inv_err_val(
const T value, const T a, const T p) noexcept { // err_val = f(x)
return (incomplete_gamma(a, value) - p);
}
template <typename T>
constexpr T incomplete_gamma_inv_deriv_1(
const T value, const T a,
const T lg_val) noexcept { // derivative of the incomplete gamma function
// w.r.t. x
return (exp(-value + (a - T(1)) * log(value) - lg_val));
}
template <typename T>
constexpr T incomplete_gamma_inv_deriv_2(
const T value, const T a,
const T deriv_1) noexcept { // second derivative of the incomplete gamma
// function w.r.t. x
return (deriv_1 * ((a - T(1)) / value - T(1)));
}
template <typename T>
constexpr T incomplete_gamma_inv_ratio_val_1(const T value, const T a,
const T p,
const T deriv_1) noexcept {
return (incomplete_gamma_inv_err_val(value, a, p) / deriv_1);
}
template <typename T>
constexpr T incomplete_gamma_inv_ratio_val_2(const T value, const T a,
const T deriv_1) noexcept {
return (incomplete_gamma_inv_deriv_2(value, a, deriv_1) / deriv_1);
}
template <typename T>
constexpr T incomplete_gamma_inv_halley(const T ratio_val_1,
const T ratio_val_2) noexcept {
return (ratio_val_1 /
std::max(T(0.8), std::min(T(1.2), T(1) - T(0.5) * ratio_val_1 *
ratio_val_2)));
}
template <typename T>
constexpr T incomplete_gamma_inv_recur(const T value, const T a, const T p,
const T deriv_1, const T lg_val,
const int iter_count) noexcept {
return incomplete_gamma_inv_decision(
value, a, p,
incomplete_gamma_inv_halley(
incomplete_gamma_inv_ratio_val_1(value, a, p, deriv_1),
incomplete_gamma_inv_ratio_val_2(value, a, deriv_1)),
lg_val, iter_count);
}
template <typename T>
constexpr T incomplete_gamma_inv_decision(const T value, const T a, const T p,
const T direc, const T lg_val,
const int iter_count) noexcept {
constexpr int INCML_GAMMA_INV_MAX_ITER = 35;
if (iter_count <= INCML_GAMMA_INV_MAX_ITER) {
return incomplete_gamma_inv_recur(
value - direc, a, p, incomplete_gamma_inv_deriv_1(value, a, lg_val),
lg_val, iter_count + 1);
} else {
return value - direc;
}
}
template <typename T>
constexpr T incomplete_gamma_inv_begin(const T initial_val, const T a,
const T p, const T lg_val) noexcept {
return incomplete_gamma_inv_recur(
initial_val, a, p, incomplete_gamma_inv_deriv_1(initial_val, a, lg_val),
lg_val, 1);
}
template <typename T>
constexpr T incomplete_gamma_inv_check(const T a, const T p) noexcept {
T gamma_p_inv_imp(const T a, const T p) {
if (is_nan(a) || is_nan(p)) {
return LIM<T>::quiet_NaN();
} else if (LIM<T>::min() > p) {
return T(0);
} else if (p > T(1)) {
return LIM<T>::quiet_NaN();
} else if (LIM<T>::min() > fabs(T(1) - p)) {
return LIM<T>::infinity();
} else if (LIM<T>::min() > a) {
return T(0);
} else {
return incomplete_gamma_inv_begin(incomplete_gamma_inv_initial_val(a, p), a,
p, lgamma(a));
if (a <= T(0)) {
throw std::runtime_error(
"Argument a in the incomplete gamma function inverse must be >= 0.");
}
} else if (p < T(0) || p > T(1)) {
throw std::runtime_error(
"Probability must be in the range [0,1] in the incomplete gamma "
"function inverse.");
} else if (p == T(0)) {
return 0;
}
}
template <typename T1, typename T2, typename TC = common_return_t<T1, T2>>
constexpr TC incomplete_gamma_inv_type_check(const T1 a, const T2 p) noexcept {
return incomplete_gamma_inv_check(static_cast<TC>(a), static_cast<TC>(p));
}
// TODO
// Get an initial guess (https://dl.acm.org/doi/abs/10.1145/22721.23109)
// T guess = find_inverse_gamma<T>(a, p, 1 - p);
bool has_10_digits = false;
boost::math::policies::policy<> pol;
T guess = boost::math::detail::find_inverse_gamma<T>(a, p, 1 - p, pol,
&has_10_digits);
} // namespace internal
T lower = LIM<T>::min();
if (guess <= lower) {
guess = LIM<T>::min();
}
// TODO
// Number of Halley iterations
// The default used in Boost is 200
// uint_fast16_t max_iter = 200;
// The number of digits to converge to.
// This is an arbitrary number,
// but Boost does more sophisticated things
// using the first derivative.
// unsigned digits = 40;
// // Perform Halley iteration for root-finding to get a more refined answer
// guess = halley_iterate(gamma_p_inverse_func<T>(a, p, false), guess, lower,
// LIM<T>::max(), digits, max_iter);
unsigned digits =
boost::math::policies::digits<T, boost::math::policies::policy<>>();
if (digits < 30) {
digits *= 2;
digits /= 3;
} else {
digits /= 2;
digits -= 1;
}
if ((a < T(0.125)) && (fabs(boost::math::gamma_p_derivative(a, guess, pol)) >
1 / sqrt(boost::math::tools::epsilon<T>())))
digits =
boost::math::policies::digits<T, boost::math::policies::policy<>>() - 2;
//
// Go ahead and iterate:
//
std::uintmax_t max_iter = boost::math::policies::get_max_root_iterations<
boost::math::policies::policy<>>();
guess = boost::math::tools::halley_iterate(
boost::math::detail::gamma_p_inverse_func<
T, boost::math::policies::policy<>>(a, p, false),
guess, lower, boost::math::tools::max_value<T>(), digits, max_iter);
if (guess == lower) {
throw std::runtime_error(
"Expected result known to be non-zero, but is smaller than the "
"smallest available number.");
}
return guess;
}
/**
* Compile-time inverse incomplete gamma function
* Compile-time check for inverse incomplete gamma function
*
* @param a a real-valued, non-negative input.
* @param p a real-valued input with values in the unit-interval.
*
* @return Computes the inverse incomplete gamma function, a value \f$ x \f$
* such that \f[ f(x) := \frac{\gamma(a,x)}{\Gamma(a)} - p \f] equal to zero,
* for a given \c p. GCE-Math finds this root using Halley's method: \f[ x_{n+1}
* = x_n - \frac{f(x_n)/f'(x_n)}{1 - 0.5 \frac{f(x_n)}{f'(x_n)}
* \frac{f''(x_n)}{f'(x_n)} } \f] where \f[ \frac{\partial}{\partial x}
* \left(\frac{\gamma(a,x)}{\Gamma(a)}\right) = \frac{1}{\Gamma(a)} x^{a-1}
* \exp(-x) \f] \f[ \frac{\partial^2}{\partial x^2}
* \left(\frac{\gamma(a,x)}{\Gamma(a)}\right) = \frac{1}{\Gamma(a)} x^{a-1}
* \exp(-x) \left( \frac{a-1}{x} - 1 \right) \f]
*/
template <typename T1, typename T2>
constexpr common_return_t<T1, T2> incomplete_gamma_inv(const T1 a,
const T2 p) noexcept {
return internal::incomplete_gamma_inv_type_check(a, p);
return internal::gamma_p_inv_imp(static_cast<common_return_t<T1, T2>>(a),
static_cast<common_return_t<T1, T2>>(p));
}
/**
* @brief Compute the quantile function of the Chi squared distribution.
* @brief Compute the quantile function of the Chi-Squared distribution.
*
* @param dofs Degrees of freedom
* @param alpha Quantile value
* @return constexpr double
* @return double
*/
constexpr double chi_squared_quantile(const size_t dofs, const double alpha) {
double chi_squared_quantile(const double dofs, const double alpha) {
// The quantile function of the Chi-squared distribution is the quantile of
// the specific (inverse) incomplete Gamma distribution
return 2 * incomplete_gamma_inv(dofs * 0.5, alpha);
return 2 * incomplete_gamma_inv(dofs / 2, alpha);
}
} // namespace internal
} // namespace gtsam

View File

@ -24,7 +24,7 @@ using namespace gtsam;
/* ************************************************************************* */
TEST(ChiSquaredInverse, ChiSqInv) {
double barcSq = chi_squared_quantile(2, 0.99);
double barcSq = internal::chi_squared_quantile(2, 0.99);
EXPECT_DOUBLES_EQUAL(9.21034, barcSq, 1e-5);
}