| 1 |
// (C) Copyright Nick Thompson 2018. |
| 2 |
// Use, modification and distribution are subject to the |
| 3 |
// Boost Software License, Version 1.0. (See accompanying file |
| 4 |
// LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt) |
| 5 |
|
| 6 |
#ifndef BOOST_MATH_TOOLS_BIVARIATE_STATISTICS_HPP |
| 7 |
#define BOOST_MATH_TOOLS_BIVARIATE_STATISTICS_HPP |
| 8 |
|
| 9 |
#include <iterator> |
| 10 |
#include <tuple> |
| 11 |
#include <limits> |
| 12 |
#include <boost/math/tools/assert.hpp> |
| 13 |
#include <boost/math/tools/header_deprecated.hpp> |
| 14 |
|
| 15 |
BOOST_MATH_HEADER_DEPRECATED("<boost/math/statistics/bivariate_statistics.hpp>"); |
| 16 |
|
| 17 |
namespace boost{ namespace math{ namespace tools { |
| 18 |
|
| 19 |
template<class Container> |
| 20 |
auto means_and_covariance(Container const & u, Container const & v) |
| 21 |
{ |
| 22 |
using Real = typename Container::value_type; |
| 23 |
using std::size; |
| 24 |
BOOST_MATH_ASSERT_MSG(size(u) == size(v), "The size of each vector must be the same to compute covariance."); |
| 25 |
BOOST_MATH_ASSERT_MSG(size(u) > 0, "Computing covariance requires at least one sample."); |
| 26 |
|
| 27 |
// See Equation III.9 of "Numerically Stable, Single-Pass, Parallel Statistics Algorithms", Bennet et al. |
| 28 |
Real cov = 0; |
| 29 |
Real mu_u = u[0]; |
| 30 |
Real mu_v = v[0]; |
| 31 |
|
| 32 |
for(size_t i = 1; i < size(u); ++i) |
| 33 |
{ |
| 34 |
Real u_tmp = (u[i] - mu_u)/(i+1); |
| 35 |
Real v_tmp = v[i] - mu_v; |
| 36 |
cov += i*u_tmp*v_tmp; |
| 37 |
mu_u = mu_u + u_tmp; |
| 38 |
mu_v = mu_v + v_tmp/(i+1); |
| 39 |
} |
| 40 |
|
| 41 |
return std::make_tuple(mu_u, mu_v, cov/size(u)); |
| 42 |
} |
| 43 |
|
| 44 |
template<class Container> |
| 45 |
auto covariance(Container const & u, Container const & v) |
| 46 |
{ |
| 47 |
auto [mu_u, mu_v, cov] = boost::math::tools::means_and_covariance(u, v); |
| 48 |
return cov; |
| 49 |
} |
| 50 |
|
| 51 |
template<class Container> |
| 52 |
auto correlation_coefficient(Container const & u, Container const & v) |
| 53 |
{ |
| 54 |
using Real = typename Container::value_type; |
| 55 |
using std::size; |
| 56 |
BOOST_MATH_ASSERT_MSG(size(u) == size(v), "The size of each vector must be the same to compute covariance."); |
| 57 |
BOOST_MATH_ASSERT_MSG(size(u) > 0, "Computing covariance requires at least two samples."); |
| 58 |
|
| 59 |
Real cov = 0; |
| 60 |
Real mu_u = u[0]; |
| 61 |
Real mu_v = v[0]; |
| 62 |
Real Qu = 0; |
| 63 |
Real Qv = 0; |
| 64 |
|
| 65 |
for(size_t i = 1; i < size(u); ++i) |
| 66 |
{ |
| 67 |
Real u_tmp = u[i] - mu_u; |
| 68 |
Real v_tmp = v[i] - mu_v; |
| 69 |
Qu = Qu + (i*u_tmp*u_tmp)/(i+1); |
| 70 |
Qv = Qv + (i*v_tmp*v_tmp)/(i+1); |
| 71 |
cov += i*u_tmp*v_tmp/(i+1); |
| 72 |
mu_u = mu_u + u_tmp/(i+1); |
| 73 |
mu_v = mu_v + v_tmp/(i+1); |
| 74 |
} |
| 75 |
|
| 76 |
// If one dataset is constant, then they have no correlation: |
| 77 |
// See https://stats.stackexchange.com/questions/23676/normalized-correlation-with-a-constant-vector |
| 78 |
// Thanks to zbjornson for pointing this out. |
| 79 |
if (Qu == 0 || Qv == 0) |
| 80 |
{ |
| 81 |
return std::numeric_limits<Real>::quiet_NaN(); |
| 82 |
} |
| 83 |
|
| 84 |
// Make sure rho in [-1, 1], even in the presence of numerical noise. |
| 85 |
Real rho = cov/sqrt(Qu*Qv); |
| 86 |
if (rho > 1) { |
| 87 |
rho = 1; |
| 88 |
} |
| 89 |
if (rho < -1) { |
| 90 |
rho = -1; |
| 91 |
} |
| 92 |
return rho; |
| 93 |
} |
| 94 |
|
| 95 |
}}} |
| 96 |
#endif |