The R Project SVN R

/*
 *  Mathlib : A C Library of Special Functions
 *  Copyright (C) 1998 Ross Ihaka
 *  Copyright (C) 2000--2011 The R Core Team
 *  Copyright (C) 2004--2009 The R Foundation
 *  based on AS 91 (C) 1979 Royal Statistical Society
 *
 *  This program is free software; you can redistribute it and/or modify
 *  it under the terms of the GNU General Public License as published by
 *  the Free Software Foundation; either version 2 of the License, or
 *  (at your option) any later version.
 *
 *  This program is distributed in the hope that it will be useful, but
 *  WITHOUT ANY WARRANTY; without even the implied warranty of
 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 *  General Public License for more details.
 *
 *  You should have received a copy of the GNU General Public License
 *  along with this program; if not, a copy is available at
 *  http://www.r-project.org/Licenses/
 *
 *  DESCRIPTION
 *
 *  Compute the quantile function of the gamma distribution.
 *
 *  NOTES
 *
 *  This function is based on the Applied Statistics
 *  Algorithm AS 91 ("ppchi2") and via pgamma(.) AS 239.
 *
 *  R core improvements:
 *  o  lower_tail, log_p
 *      o  non-trivial result for p outside [0.000002, 0.999998]
 *  o  p ~ 1 no longer gives +Inf; final Newton step(s)
 *
 *  REFERENCES
 *
 *  Best, D. J. and D. E. Roberts (1975).
 *  Percentage Points of the Chi-Squared Distribution.
 *  Applied Statistics 24, page 385.  */

#include "nmath.h"
#include "dpq.h"

#ifdef DEBUG_qgamma
# define DEBUG_q
#endif

attribute_hidden
double qchisq_appr(double p, double nu, double g /* = log Gamma(nu/2) */,
           int lower_tail, int log_p, double tol /* EPS1 */)
{
#define C7  4.67
#define C8  6.66
#define C9  6.73
#define C10 13.32

    double alpha, a, c, ch, p1;
    double p2, q, t, x;

    /* test arguments and initialise */

#ifdef IEEE_754
    if (ISNAN(p) || ISNAN(nu))
    return p + nu;
#endif
    R_Q_P01_check(p);
    if (nu <= 0) ML_ERR_return_NAN;

    alpha = 0.5 * nu;/* = [pq]gamma() shape */
    c = alpha-1;

    if(nu < (-1.24)*(p1 = R_DT_log(p))) {   /* for small chi-squared */
    /* log(alpha) + g = log(alpha) + log(gamma(alpha)) =
     *        = log(alpha*gamma(alpha)) = lgamma(alpha+1) suffers from
     *  catastrophic cancellation when alpha << 1
     */
    double lgam1pa = (alpha < 0.5) ? lgamma1p(alpha) : (log(alpha) + g);
    ch = exp((lgam1pa + p1)/alpha + M_LN2);
#ifdef DEBUG_qgamma
    REprintf(" small chi-sq., ch0 = %g\n", ch);
#endif

    } else if(nu > 0.32) {  /*  using Wilson and Hilferty estimate */

    x = qnorm(p, 0, 1, lower_tail, log_p);
    p1 = 2./(9*nu);
    ch = nu*pow(x*sqrt(p1) + 1-p1, 3);

#ifdef DEBUG_qgamma
    REprintf(" nu > .32: Wilson-Hilferty; x = %7g\n", x);
#endif
    /* approximation for p tending to 1: */
    if( ch > 2.2*nu + 6 )
        ch = -2*(R_DT_Clog(p) - c*log(0.5*ch) + g);

    } else { /* "small nu" : 1.24*(-log(p)) <= nu <= 0.32 */

    ch = 0.4;
    a = R_DT_Clog(p) + g + c*M_LN2;
#ifdef DEBUG_qgamma
    REprintf(" nu <= .32: a = %7g\n", a);
#endif
    do {
        q = ch;
        p1 = 1. / (1+ch*(C7+ch));
        p2 = ch*(C9+ch*(C8+ch));
        t = -0.5 +(C7+2*ch)*p1 - (C9+ch*(C10+3*ch))/p2;
        ch -= (1- exp(a+0.5*ch)*p2*p1)/t;
    } while(fabs(q - ch) > tol * fabs(ch));
    }

    return ch;
}

double qgamma(double p, double alpha, double scale, int lower_tail, int log_p)
/*          shape = alpha */
{
#define EPS1 1e-2
#define EPS2 5e-7/* final precision of AS 91 */
#define EPS_N 1e-15/* precision of Newton step / iterations */
#define LN_EPS -36.043653389117156 /* = log(.Machine$double.eps) iff IEEE_754 */

#define MAXIT 1000/* was 20 */

#define pMIN 1e-100   /* was 0.000002 = 2e-6 */
#define pMAX (1-1e-14)/* was (1-1e-12) and 0.999998 = 1 - 2e-6 */

    const static double
    i420  = 1./ 420.,
    i2520 = 1./ 2520.,
    i5040 = 1./ 5040;

    double p_, a, b, c, g, ch, ch0, p1;
    double p2, q, s1, s2, s3, s4, s5, s6, t, x;
    int i, max_it_Newton = 1;

    /* test arguments and initialise */

#ifdef IEEE_754
    if (ISNAN(p) || ISNAN(alpha) || ISNAN(scale))
    return p + alpha + scale;
#endif
    R_Q_P01_boundaries(p, 0., ML_POSINF);

    if (alpha < 0 || scale <= 0) ML_ERR_return_NAN;

    if (alpha == 0) /* all mass at 0 : */ return 0.;

    if (alpha < 1e-10) {
    /* Warning seems unnecessary now: */
#ifdef _DO_WARN_qgamma_
    MATHLIB_WARNING("value of shape (%g) is extremely small: results may be unreliable", alpha);
#endif
    max_it_Newton = 7;/* may still be increased below */
    }

    p_ = R_DT_qIv(p);/* lower_tail prob (in any case) */

#ifdef DEBUG_qgamma
    REprintf("qgamma(p=%7g, alpha=%7g, scale=%7g, l.t.=%2d, log_p=%2d): ",
         p,alpha,scale, lower_tail, log_p);
#endif
    g = lgammafn(alpha);/* log Gamma(v/2) */

    /*----- Phase I : Starting Approximation */
    ch = qchisq_appr(p, /* nu= 'df' =  */ 2*alpha, /* lgamma(nu/2)= */ g,
             lower_tail, log_p, /* tol= */ EPS1);
    if(!R_FINITE(ch)) {
    /* forget about all iterations! */
    max_it_Newton = 0; goto END;
    }
    if(ch < EPS2) {/* Corrected according to AS 91; MM, May 25, 1999 */
    max_it_Newton = 20;
    goto END;/* and do Newton steps */
    }

    /* FIXME: This (cutoff to {0, +Inf}) is far from optimal
     * -----  when log_p or !lower_tail, but NOT doing it can be even worse */
    if(p_ > pMAX || p_ < pMIN) {
    /* did return ML_POSINF or 0.;  much better: */
    max_it_Newton = 20;
    goto END;/* and do Newton steps */
    }

#ifdef DEBUG_qgamma
    REprintf("\t==> ch = %10g:", ch);
#endif

/*----- Phase II: Iteration
 *  Call pgamma() [AS 239]  and calculate seven term taylor series
 */
    c = alpha-1;
    s6 = (120+c*(346+127*c)) * i5040; /* used below, is "const" */

    ch0 = ch;/* save initial approx. */
    for(i=1; i <= MAXIT; i++ ) {
    q = ch;
    p1 = 0.5*ch;
    p2 = p_ - pgamma_raw(p1, alpha, /*lower_tail*/TRUE, /*log_p*/FALSE);
#ifdef DEBUG_qgamma
    if(i == 1) REprintf(" Ph.II iter; ch=%g, p2=%g\n", ch, p2);
    if(i >= 2) REprintf("     it=%d,  ch=%g, p2=%g\n", i, ch, p2);
#endif
#ifdef IEEE_754
    if(!R_FINITE(p2) || ch <= 0)
#else
    if(errno != 0 || ch <= 0)
#endif
        { ch = ch0; max_it_Newton = 27; goto END; }/*was  return ML_NAN;*/

    t = p2*exp(alpha*M_LN2+g+p1-c*log(ch));
    b = t/ch;
    a = 0.5*t - b*c;
    s1 = (210+ a*(140+a*(105+a*(84+a*(70+60*a))))) * i420;
    s2 = (420+ a*(735+a*(966+a*(1141+1278*a)))) * i2520;
    s3 = (210+ a*(462+a*(707+932*a))) * i2520;
    s4 = (252+ a*(672+1182*a) + c*(294+a*(889+1740*a))) * i5040;
    s5 = (84+2264*a + c*(1175+606*a)) * i2520;

    ch += t*(1+0.5*t*s1-b*c*(s1-b*(s2-b*(s3-b*(s4-b*(s5-b*s6))))));
    if(fabs(q - ch) < EPS2*ch)
        goto END;
    if(fabs(q - ch) > 0.1*ch) {/* diverging? -- also forces ch > 0 */
        if(ch < q) ch = 0.9 * q; else ch = 1.1 * q;
    }
    }
/* no convergence in MAXIT iterations -- but we add Newton now... */
#ifdef DEBUG_q
    MATHLIB_WARNING3("qgamma(%g) not converged in %d iterations; rel.ch=%g\n",
             p, MAXIT, ch/fabs(q - ch));
#endif
/* was
 *    ML_ERROR(ME_PRECISION, "qgamma");
 * does nothing in R !*/

END:
/* PR# 2214 :    From: Morten Welinder <terra@diku.dk>, Fri, 25 Oct 2002 16:50
   --------  To: R-bugs@biostat.ku.dk     Subject: qgamma precision

   * With a final Newton step, double accuracy, e.g. for (p= 7e-4; nu= 0.9)
   *
   * Improved (MM): - only if rel.Err > EPS_N (= 1e-15);
   *            - also for lower_tail = FALSE    or log_p = TRUE
   *            - optionally *iterate* Newton
   */
    x = 0.5*scale*ch;
    if(max_it_Newton) {
    /* always use log scale */
    if (!log_p) {
        p = log(p);
        log_p = TRUE;
    }
    if(x == 0) {
        const double _1_p = 1. + 1e-7;
        const double _1_m = 1. - 1e-7;
        x = DBL_MIN;
        p_ = pgamma(x, alpha, scale, lower_tail, log_p);
        if(( lower_tail && p_ > p * _1_p) ||
           (!lower_tail && p_ < p * _1_m))
        return(0.);
        /* else:  continue, using x = DBL_MIN instead of  0  */
    }
    else
        p_ = pgamma(x, alpha, scale, lower_tail, log_p);
    if(p_ == ML_NEGINF) return 0; /* PR#14710 */
    for(i = 1; i <= max_it_Newton; i++) {
        p1 = p_ - p;
#ifdef DEBUG_qgamma
        if(i == 1) REprintf("\n it=%d: p=%g, x = %g, p.=%g; p1=d{p}=%g\n",
                i, p, x, p_, p1);
        if(i >= 2) REprintf("          x{it= %d} = %g, p.=%g, p1=d{p}=%g\n",
                i,    x, p_, p1);
#endif
        if(fabs(p1) < fabs(EPS_N * p))
        break;
        /* else */
        if((g = dgamma(x, alpha, scale, log_p)) == R_D__0) {
#ifdef DEBUG_q
        if(i == 1) REprintf("no final Newton step because dgamma(*)== 0!\n");
#endif
        break;
        }
        /* else :
         * delta x = f(x)/f'(x);
         * if(log_p) f(x) := log P(x) - p; f'(x) = d/dx log P(x) = P' / P
         * ==> f(x)/f'(x) = f*P / P' = f*exp(p_) / P' (since p_ = log P(x))
         */
        t = log_p ? p1*exp(p_ - g) : p1/g ;/* = "delta x" */
        t = lower_tail ? x - t : x + t;
        p_ = pgamma (t, alpha, scale, lower_tail, log_p);
        if (fabs(p_ - p) > fabs(p1) ||
        (i > 1 && fabs(p_ - p) == fabs(p1)) /* <- against flip-flop */) {
        /* no improvement */
#ifdef DEBUG_qgamma
        if(i == 1 && max_it_Newton > 1)
            REprintf("no Newton step done since delta{p} >= last delta\n");
#endif
        break;
        } /* else : */
#ifdef Harmful_notably_if_max_it_Newton_is_1
        /* control step length: this could have started at
           the initial approximation */
        if(t > 1.1*x) t = 1.1*x;
        else if(t < 0.9*x) t = 0.9*x;
#endif
        x = t;
    }
    }

    return x;
}