[Rd] pgamma discontinuity (PR#7307)

terra at gnome.org terra at gnome.org
Tue Jan 11 19:57:45 CET 2005


FYI,

this version of pgamma appears to be quite good and certainly a vast
improvement over current R code.  (I last looked at 2.0.1)  Apart from
type naming, this is what I have been using for Gnumeric 1.4.1.

This could be included into R as-is, but you might want to benefit from
making logcf, log1pmx, lgamma1p, and possibly logspace_add/logspace_sub
available to other parts of R.

Comments welcome.

Morten




#ifndef GNUMERIC_VERSION
/* Gnumeric already has these.  */

#define SQR(x) ((x)*(x))
static const double scalefactor = SQR(SQR(SQR(4294967296.0)));
#undef SQR

static double
logcf (double x, double i, double d)
{
	double c1 = 2 * d;
	double c2 = i + d;
	double c4 = c2 + d;
	double a1 = c2;
	double b1 = i * (c2 - i * x);
	double b2 = d * d * x;
	double a2 = c4 * c2 - b2;
	const double cfVSmall = 1.0e-14;

#if 0
	assert (i > 0);
	assert (d >= 0);
#endif

	b2 = c4 * b1 - i * b2;

	while (fabs (a2 * b1 - a1 * b2) > fabs (cfVSmall * b1 * b2)) {
		double c3 = c2*c2*x;
		c2 += d;
		c4 += d;
		a1 = c4 * a2 - c3 * a1;
		b1 = c4 * b2 - c3 * b1;

		c3 = c1 * c1 * x;
		c1 += d;
		c4 += d;
		a2 = c4 * a1 - c3 * a2;
		b2 = c4 * b1 - c3 * b2;

		if (fabs (b2) > scalefactor) {
			a1 *= 1 / scalefactor;
			b1 *= 1 / scalefactor;
			a2 *= 1 / scalefactor;
			b2 *= 1 / scalefactor;
		} else if (fabs (b2) < 1 / scalefactor) {
			a1 *= scalefactor;
			b1 *= scalefactor;
			a2 *= scalefactor;
			b2 *= scalefactor;
		}
	}

	return a2 / b2;
}

/* Accurate calculation of log(1+x)-x, particularly for small x.  */
static double
log1pmx (double x)
{
	static const double minLog1Value = -0.79149064;
	static const double two = 2;

	if (fabs (x) < 1.0e-2) {
		double term = x / (2 + x);
		double y = term * term;
		return term * ((((two / 9 * y + two / 7) * y + two / 5) * y + two / 3) * y - x);
	} else if (x < minLog1Value || x > 1) {
		return log1p (x) - x;
	} else {
		double term = x / (2 + x);
		double y = term * term;
		return term * (2 * y * logcf (y, 3, 2) - x);
	}
}


/* Calculates log(gamma(a+1) accurately for for small a (0 < a & a < 0.5). */
static double
lgamma1p (double a)
{
     const double eulers_const =  0.5772156649015328606065120900824024;

     /* coeffs[i] holds (zeta(i+2)-1)/(i+2)  */
     static const double coeffs[40] = {
	  0.3224670334241132182362075833230126e-0,
	  0.6735230105319809513324605383715000e-1,
	  0.2058080842778454787900092413529198e-1,
	  0.7385551028673985266273097291406834e-2,
	  0.2890510330741523285752988298486755e-2,
	  0.1192753911703260977113935692828109e-2,
	  0.5096695247430424223356548135815582e-3,
	  0.2231547584535793797614188036013401e-3,
	  0.9945751278180853371459589003190170e-4,
	  0.4492623673813314170020750240635786e-4,
	  0.2050721277567069155316650397830591e-4,
	  0.9439488275268395903987425104415055e-5,
	  0.4374866789907487804181793223952411e-5,
	  0.2039215753801366236781900709670839e-5,
	  0.9551412130407419832857179772951265e-6,
	  0.4492469198764566043294290331193655e-6,
	  0.2120718480555466586923135901077628e-6,
	  0.1004322482396809960872083050053344e-6,
	  0.4769810169363980565760193417246730e-7,
	  0.2271109460894316491031998116062124e-7,
	  0.1083865921489695409107491757968159e-7,
	  0.5183475041970046655121248647057669e-8,
	  0.2483674543802478317185008663991718e-8,
	  0.1192140140586091207442548202774640e-8,
	  0.5731367241678862013330194857961011e-9,
	  0.2759522885124233145178149692816341e-9,
	  0.1330476437424448948149715720858008e-9,
	  0.6422964563838100022082448087644648e-10,
	  0.3104424774732227276239215783404066e-10,
	  0.1502138408075414217093301048780668e-10,
	  0.7275974480239079662504549924814047e-11,
	  0.3527742476575915083615072228655483e-11,
	  0.1711991790559617908601084114443031e-11,
	  0.8315385841420284819798357793954418e-12,
	  0.4042200525289440065536008957032895e-12,
	  0.1966475631096616490411045679010286e-12,
	  0.9573630387838555763782200936508615e-13,
	  0.4664076026428374224576492565974577e-13,
	  0.2273736960065972320633279596737272e-13,
	  0.1109139947083452201658320007192334e-13
     };
     const int N = sizeof (coeffs) / sizeof (coeffs[0]);

     const double c =  0.2273736845824652515226821577978691e-12;  /* zeta(N+2)-1 */
     double lgam;
     int i;

     if (fabs (a) >= 0.5)
	  return lgammafn (a + 1);

     /* Abramowitz & Stegun 6.1.33 */
     /* http://functions.wolfram.com/06.11.06.0008.01 */
     lgam = c * logcf (-a / 2, N + 2, 1);
     for (i = N - 1; i >= 0; i--)
	  lgam = coeffs[i] - a * lgam;

     return (a * lgam - eulers_const) * a - log1pmx (a);
}

#endif /* GNUMERIC_VERSION */


/*
 * Compute the log of a sum from logs of terms, i.e.,
 *
 *     log (exp (logx) + exp (logy))
 *
 * without causing overflows and without throwing away large handfuls
 * of accuracy.
 */
static double
logspace_add (double logx, double logy)
{
     return fmax2 (logx, logy) + log1p (exp (-fabs (logx - logy)));
}


/*
 * Compute the log of a difference from logs of terms, i.e.,
 *
 *     log (exp (logx) - exp (logy))
 *
 * without causing overflows and without throwing away large handfuls
 * of accuracy.
 */
static double
logspace_sub (double logx, double logy)
{
     return logx + log1p (-exp (logy - logx));
}


static double
dpois_wrap (double x_plus_1, double lambda, int give_log)
{
#if 0
     printf ("x+1=%.14g  lambda=%.14g\n", x_plus_1, lambda);
#endif

     if (x_plus_1 > 1)
	  return dpois_raw (x_plus_1 - 1, lambda, give_log);
     else {
 	  double d = dpois_raw (x_plus_1, lambda, give_log);
	  return give_log
	       ? d + log (x_plus_1 / lambda)
	       : d * (x_plus_1 / lambda);
     }
}

/*
 * Abramowitz and Stegun 6.5.29 [right]
 */
static double
pgamma_smallx (double x, double alph, int lower_tail, int log_p)
{
     double sum = 0, c = alph, n = 0, term;

#if 0
     printf ("x:%.14g  alph:%.14g\n", x, alph);
#endif

     /*
      * Relative to 6.5.29 all terms have been multiplied by alph
      * and the first, thus being 1, is omitted.
      */

     do {
	  n++;
	  c *= -x / n;
	  term = c / (alph + n);
	  sum += term;
     } while (fabs (term) > DBL_EPSILON * fabs (sum));

     if (lower_tail) {
	  double f1 = log_p ? log1p (sum) : 1 + sum;
	  double f2;
	  if (alph > 1) {
	       f2 = dpois_raw (alph, x, log_p);
	       f2 = log_p ? f2 + x : f2 * exp (x);
	  } else if (log_p)
	       f2 = alph * log (x) - lgamma1p (alph);
	  else
	       f2 = pow (x, alph) / exp (lgamma1p (alph));

	  return log_p ? f1 + f2 : f1 * f2;
     } else {
	  double lf2 = alph * log (x) - lgamma1p (alph);
#if 0
	  printf ("1:%.14g  2:%.14g\n", alph * log (x), lgamma1p (alph));
	  printf ("sum=%.14g  log(1+sum)=%.14g  lf2=%.14g\n", sum, log1p (sum), lf2);
#endif
	  if (log_p)
	       return R_Log1_Exp (log1p (sum) + lf2);
	  else {
	       double f1m1 = sum;
	       double f2m1 = expm1 (lf2);
	       return -(f1m1 + f2m1 + f1m1 * f2m1);
	  }
     }
}

static double
pd_upper_series (double x, double y, int log_p)
{
     double term = x / y;
     double sum = term;

     do {
	  y++;
	  term *= x / y;
	  sum += term;
     } while (term > sum * DBL_EPSILON);

     return log_p ? log (sum) : sum;
}

static double
pd_lower_cf (double i, double d)
{
     double f = 0, of;

     double a1 = 0;
     double b1 = 1;
     double a2 = i;
     double b2 = d;
     double c1 = 0;
     double c2 = a2;
     double c3;
     double c4 = b2;

     while (1) {
	  c1++;
	  c2--;
	  c3 = c1 * c2;
	  c4 += 2;
	  a1 = c4 * a2 + c3 * a1;
	  b1 = c4 * b2 + c3 * b1;

	  c1++;
	  c2--;
	  c3 = c1 * c2;
	  c4 += 2;
	  a2 = c4 * a1 + c3 * a2;
	  b2 = c4 * b1 + c3 * b2;

	  if (b2 > scalefactor) {
	       a1 = a1 / scalefactor;
	       b1 = b1 / scalefactor;
	       a2 = a2 / scalefactor;
	       b2 = b2 / scalefactor;
	  }

	  if (b2 != 0) {
	       of = f;
	       f = a2 / b2;
	       if (fabs (f - of) <= DBL_EPSILON * fmin2 (1.0, fabs (f)))
		    return f;
	  }
     }
}

static double
pd_lower_series (double lambda, double y)
{
     double term = 1, sum = 0;

     while (y >= 1 && term > sum * DBL_EPSILON) {
	  term *= y / lambda;
	  sum += term;
	  y--;
     }

     if (y != floor (y)) {
	  /*
	   * The series does not converge as the terms start getting
	   * bigger (besides flipping sign) for y < -lambda.
	   */
	  double f = pd_lower_cf (y, lambda + 1 - y);
	  sum += term * f;
     }

     return sum;
}

/*
 * Asymptotic expansion to calculate the probability that poisson variate
 * has value <= x.
 */
static double
ppois_asymp (double x, double lambda,
	     int lower_tail, int log_p)
{
     static const double coef15 = 1/12.0;
     static const double coef25 = 1/288.0;
     static const double coef35 = -139/51840.0;
     static const double coef45 = -571/2488320.0;
     static const double coef55 = 163879/209018880.0;
     static const double coef65 =  5246819/75246796800.0;
     static const double coef75 = -534703531/902961561600.0;
     static const double coef1 = 2/3.0;
     static const double coef2 = -4/135.0;
     static const double coef3 = 8/2835.0;
     static const double coef4 = 16/8505.0;
     static const double coef5 = -8992/12629925.0;
     static const double coef6 = -334144/492567075.0;
     static const double coef7 = 698752/1477701225.0;
     static const double two = 2;

     double dfm, pt,s2pt,res1,res2,elfb,term;
     double ig2,ig3,ig4,ig5,ig6,ig7,ig25,ig35,ig45,ig55,ig65,ig75;
     double f, np, nd;

     dfm = lambda - x;
     pt = -x * log1pmx (dfm / x);
     s2pt = sqrt (2 * pt);
     if (dfm < 0) s2pt = -s2pt;

     ig2 = 1.0 + pt;
     term = pt * pt * 0.5;
     ig3 = ig2 + term;
     term *= pt / 3;
     ig4 = ig3 + term;
     term *= pt / 4;
     ig5 = ig4 + term;
     term *= pt / 5;
     ig6 = ig5 + term;
     term *= pt / 6;
     ig7 = ig6 + term;

     term = pt * (two / 3);
     ig25 = 1.0 + term;
     term *= pt * (two / 5);
     ig35 = ig25 + term;
     term *= pt * (two / 7);
     ig45 = ig35 + term;
     term *= pt * (two / 9);
     ig55 = ig45 + term;
     term *= pt * (two / 11);
     ig65 = ig55 + term;
     term *= pt * (two / 13);
     ig75 = ig65 + term;

     elfb = ((((((coef75/x + coef65)/x + coef55)/x + coef45)/x + coef35)/x + coef25)/x + coef15) + x;
     res1 = ((((((ig7*coef7/x + ig6*coef6)/x + ig5*coef5)/x + ig4*coef4)/x + ig3*coef3)/x + ig2*coef2)/x + coef1)*sqrt(x);
     res2 = ((((((ig75*coef75/x + ig65*coef65)/x + ig55*coef55)/x + ig45*coef45)/x + ig35*coef35)/x + ig25*coef25)/x + coef15)*s2pt;

     if (!lower_tail) elfb = -elfb;
     f = (res1 + res2) / elfb;

     np = pnorm (s2pt, 0.0, 1.0, !lower_tail, log_p);
     nd = dnorm (s2pt, 0.0, 1.0, log_p);

#if 0
     printf ("f=%.14g  np=%.14g  nd=%.14g  f*nd=%.14g\n", f, np, nd, f * nd);
#endif

     if (log_p)
	  return (f >= 0)
	       ? logspace_add (np, log (fabs (f)) + nd)
	       : logspace_sub (np, log (fabs (f)) + nd);
     else
	  return np + f * nd;
}


static double
pgamma_raw (double x, double alph, int lower_tail, int log_p)
{
     double res;

#if 0
     printf ("x=%.14g  alph=%.14g  low=%d  log=%d\n", x, alph, lower_tail, log_p);
#endif

     if (x < 1) {
	  res = pgamma_smallx (x, alph, lower_tail, log_p);
     } else if (x <= alph - 1 && x < 0.8 * (alph + 50)) {
	  double sum = pd_upper_series (x, alph, log_p);
	  double d = dpois_wrap (alph, x, log_p);

	  if (!lower_tail)
	       res = log_p
		    ? R_Log1_Exp (d + sum)
		    : 1 - d * sum;
	  else
	       res = log_p ? sum + d : sum * d;
     } else if (alph - 1 < x && alph < 0.8 * (x + 50)) {
	  double sum;
	  double d = dpois_wrap (alph, x, log_p);

	  if (alph < 1) {
	       double f = pd_lower_cf (alph, x - (alph - 1))
		    * x / alph;
	       sum = log_p ? log (f) : f;
	  } else {
	       sum = pd_lower_series (x, alph - 1);
	       sum = log_p ? log1p (sum) : 1 + sum;
	  }

	  if (!lower_tail)
	       res = log_p ? sum + d : sum * d;
	  else
	       res = log_p
		    ? R_Log1_Exp (d + sum)
		    : 1 - d * sum;
     } else {
	  res = ppois_asymp (alph - 1, x, !lower_tail, log_p);
     }

     /*
      * We lose a fair amount of accuracy to underflow in the cases
      * where the final result is very close to DBL_MIN.  In those
      * cases, simply redo via log space.
      */
     if (!log_p && res < DBL_MIN / DBL_EPSILON)
	  return exp (pgamma_raw (x, alph, lower_tail, 1));
     else
	  return res;
}


double pgamma(double x, double alph, double scale, int lower_tail, int log_p)
{
#ifdef IEEE_754
     if (ISNAN(x) || ISNAN(alph) || ISNAN(scale))
	  return x + alph + scale;
#endif
     if(alph <= 0. || scale <= 0.)
	  ML_ERR_return_NAN;
     if (x <= 0.)
	  return R_DT_0;
     x /= scale;
#ifdef IEEE_754
     if (ISNAN(x)) /* eg. original x = scale = +Inf */
	  return x;
#endif

     return pgamma_raw (x, alph, lower_tail, log_p);
}



More information about the R-devel mailing list