1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205
|
/* actuar: Actuarial Functions and Heavy Tailed Distributions
*
* Functions to compute density, cumulative distribution and quantile
* functions, and to simulate random variates for the Poisson-inverse
* gaussian distribution. See ../R/PoissonInverseGaussian.R for
* details.
*
* We work with the density expressed as
*
* p(x) = sqrt(1/phi) sqrt(1/(pi/2)) exp(1/(phi mu))/x!
* * [sqrt(2 phi (1 + (2 phi mu^2)^(-1)))]^(-(x - 0.5))
* * bessel_k(sqrt(2/phi (1 + (2 phi mu^2)^(-1))), x - 0.5)
*
* or, is essence,
*
* p(x) = A exp(1/(phi mu))/x! B^(-y) bessel_k(B/phi, y)
*
* The limiting case mu = Inf is handled "automatically" with terms
* going to zero when mu is Inf. Specific code not worth it since the
* function should rarely be evaluated with mu = Inf in practice.
*
* AUTHOR: Vincent Goulet <vincent.goulet@act.ulaval.ca>
*/
#include <R.h>
#include <Rmath.h>
#include "locale.h"
#include "dpq.h"
#include "actuar.h"
double dpoisinvgauss_raw(double x, double mu, double phi, int give_log)
{
/* Here assume that x is integer, 0 < x < Inf, mu > 0, 0 < phi < Inf */
int i;
double p, pi1m, pi2m;
double twophi = phi + phi;
/* limiting case mu = Inf with simpler recursive formulas */
if (!R_FINITE(mu))
{
p = -sqrt(2/phi); /* log p[0] */
if (x == 0.0)
return ACT_D_exp(p);
pi2m = exp(p); /* p[i - 2] = p[0]*/
p = p - (M_LN2 + log(phi))/2; /* log p[1] */
if (x == 1.0)
return ACT_D_exp(p);
pi1m = exp(p); /* p[i - 1] = p[1] */
for (i = 2; i <= x; i++)
{
p = (1 - 1.5/i) * pi1m + pi2m/twophi/(i * (i - 1));
pi2m = pi1m;
pi1m = p;
}
return ACT_D_val(p);
}
/* else: "standard" case with mu < Inf */
double A, B;
double mu2 = mu * mu;
double twophimu2 = twophi * mu2;
p = (1.0 - sqrt(1.0 + twophimu2))/phi/mu; /* log p[0] */
if (x == 0.0)
return ACT_D_exp(p);
pi2m = exp(p); /* p[i - 2] = p[0]*/
p = log(mu) + p - log1p(twophimu2)/2.0; /* log p[1] */
if (x == 1.0)
return ACT_D_exp(p);
pi1m = exp(p); /* p[i - 1] = p[1] */
A = 1.0/(1.0 + 1.0/twophimu2); /* constant in first term */
B = mu2/(1.0 + twophimu2); /* constant in second term */
for (i = 2; i <= x; i++)
{
p = A * (1 - 1.5/i) * pi1m + (B * pi2m)/(i * (i - 1));
pi2m = pi1m;
pi1m = p;
}
return ACT_D_val(p);
}
double dpoisinvgauss(double x, double mu, double phi, int give_log)
{
#ifdef IEEE_754
if (ISNAN(x) || ISNAN(mu) || ISNAN(phi))
return x + mu + phi;
#endif
if (mu <= 0.0 || phi <= 0.0)
return R_NaN;
ACT_D_nonint_check(x);
if (!R_FINITE(x) || x < 0.0)
return ACT_D__0;
/* limiting case phi = Inf */
if (!R_FINITE(phi))
return (x == 0) ? ACT_D__1 : ACT_D__0;
return dpoisinvgauss_raw(x, mu, phi, give_log);
}
/* For ppoisinvgauss(), there does not seem to be algorithms much
* more elaborate than successive computations of the probabilities.
*/
double ppoisinvgauss(double q, double mu, double phi, int lower_tail, int log_p)
{
#ifdef IEEE_754
if (ISNAN(q) || ISNAN(mu) || ISNAN(phi))
return q + mu + phi;
#endif
if (mu <= 0.0 || phi <= 0.0)
return R_NaN;
if (q < 0)
return ACT_DT_0;
/* limiting case phi = Inf */
if (!R_FINITE(phi))
return ACT_DT_1;
if (!R_FINITE(q))
return ACT_DT_1;
int x;
double s = 0;
for (x = 0; x <= q; x++)
s += dpoisinvgauss_raw(x, mu, phi, /*give_log*/ 0);
return ACT_DT_val(s);
}
/* For qpoisinvgauss() we mostly reuse the code from qnbinom() et al.
* of R sources. From src/nmath/qnbinom.c:
*
* METHOD
*
* Uses the Cornish-Fisher Expansion to include a skewness
* correction to a normal approximation. This gives an
* initial value which never seems to be off by more than
* 1 or 2. A search is then conducted of values close to
* this initial start point.
*
* For the limiting case mu = Inf (that has no finite moments), we
* use instead the quantile of an inverse chi-square distribution as
* starting point.
*/
#define _thisDIST_ poisinvgauss
#define _dist_PARS_DECL_ double mu, double phi
#define _dist_PARS_ mu, phi
#include "qDiscrete_search.h" /* do_search() et al. */
double qpoisinvgauss(double p, double mu, double phi, int lower_tail, int log_p)
{
#ifdef IEEE_754
if (ISNAN(p) || ISNAN(mu) || ISNAN(phi))
return p + mu + phi;
#endif
if (mu <= 0.0 || phi <= 0.0)
return R_NaN;
/* limiting case phi = Inf */
if (!R_FINITE(phi))
return 0.0;
ACT_Q_P01_boundaries(p, 0, R_PosInf);
double
phim2 = phi * mu * mu,
sigma2 = phim2 * mu + mu,
sigma = sqrt(sigma2),
gamma = (3 * phim2 * sigma2 + mu)/sigma2/sigma;
/* q_DISCRETE_01_CHECKS(); */
/* limiting case mu = Inf -> inverse chi-square as starting point*/
/* other cases -> Cornish-Fisher as usual */
double z, y;
if (!R_FINITE(mu))
y = ACT_forceint(1/phi/qchisq(p, 1, lower_tail, log_p));
else
{
z = qnorm(p, 0., 1., lower_tail, log_p);
y = ACT_forceint(mu + sigma * (z + gamma * (z*z - 1) / 6));
}
q_DISCRETE_BODY();
}
double rpoisinvgauss(double mu, double phi)
{
if (mu <= 0.0 || phi <= 0.0)
return R_NaN;
return rpois(rinvgauss(mu, phi));
}
|