• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 #ifndef JEMALLOC_ENABLE_INLINE
2 double	ln_gamma(double x);
3 double	i_gamma(double x, double p, double ln_gamma_p);
4 double	pt_norm(double p);
5 double	pt_chi2(double p, double df, double ln_gamma_df_2);
6 double	pt_gamma(double p, double shape, double scale, double ln_gamma_shape);
7 #endif
8 
9 #if (defined(JEMALLOC_ENABLE_INLINE) || defined(MATH_C_))
10 /*
11  * Compute the natural log of Gamma(x), accurate to 10 decimal places.
12  *
13  * This implementation is based on:
14  *
15  *   Pike, M.C., I.D. Hill (1966) Algorithm 291: Logarithm of Gamma function
16  *   [S14].  Communications of the ACM 9(9):684.
17  */
18 JEMALLOC_INLINE double
ln_gamma(double x)19 ln_gamma(double x)
20 {
21 	double f, z;
22 
23 	assert(x > 0.0);
24 
25 	if (x < 7.0) {
26 		f = 1.0;
27 		z = x;
28 		while (z < 7.0) {
29 			f *= z;
30 			z += 1.0;
31 		}
32 		x = z;
33 		f = -log(f);
34 	} else
35 		f = 0.0;
36 
37 	z = 1.0 / (x * x);
38 
39 	return (f + (x-0.5) * log(x) - x + 0.918938533204673 +
40 	    (((-0.000595238095238 * z + 0.000793650793651) * z -
41 	    0.002777777777778) * z + 0.083333333333333) / x);
42 }
43 
44 /*
45  * Compute the incomplete Gamma ratio for [0..x], where p is the shape
46  * parameter, and ln_gamma_p is ln_gamma(p).
47  *
48  * This implementation is based on:
49  *
50  *   Bhattacharjee, G.P. (1970) Algorithm AS 32: The incomplete Gamma integral.
51  *   Applied Statistics 19:285-287.
52  */
53 JEMALLOC_INLINE double
i_gamma(double x,double p,double ln_gamma_p)54 i_gamma(double x, double p, double ln_gamma_p)
55 {
56 	double acu, factor, oflo, gin, term, rn, a, b, an, dif;
57 	double pn[6];
58 	unsigned i;
59 
60 	assert(p > 0.0);
61 	assert(x >= 0.0);
62 
63 	if (x == 0.0)
64 		return (0.0);
65 
66 	acu = 1.0e-10;
67 	oflo = 1.0e30;
68 	gin = 0.0;
69 	factor = exp(p * log(x) - x - ln_gamma_p);
70 
71 	if (x <= 1.0 || x < p) {
72 		/* Calculation by series expansion. */
73 		gin = 1.0;
74 		term = 1.0;
75 		rn = p;
76 
77 		while (true) {
78 			rn += 1.0;
79 			term *= x / rn;
80 			gin += term;
81 			if (term <= acu) {
82 				gin *= factor / p;
83 				return (gin);
84 			}
85 		}
86 	} else {
87 		/* Calculation by continued fraction. */
88 		a = 1.0 - p;
89 		b = a + x + 1.0;
90 		term = 0.0;
91 		pn[0] = 1.0;
92 		pn[1] = x;
93 		pn[2] = x + 1.0;
94 		pn[3] = x * b;
95 		gin = pn[2] / pn[3];
96 
97 		while (true) {
98 			a += 1.0;
99 			b += 2.0;
100 			term += 1.0;
101 			an = a * term;
102 			for (i = 0; i < 2; i++)
103 				pn[i+4] = b * pn[i+2] - an * pn[i];
104 			if (pn[5] != 0.0) {
105 				rn = pn[4] / pn[5];
106 				dif = fabs(gin - rn);
107 				if (dif <= acu && dif <= acu * rn) {
108 					gin = 1.0 - factor * gin;
109 					return (gin);
110 				}
111 				gin = rn;
112 			}
113 			for (i = 0; i < 4; i++)
114 				pn[i] = pn[i+2];
115 
116 			if (fabs(pn[4]) >= oflo) {
117 				for (i = 0; i < 4; i++)
118 					pn[i] /= oflo;
119 			}
120 		}
121 	}
122 }
123 
124 /*
125  * Given a value p in [0..1] of the lower tail area of the normal distribution,
126  * compute the limit on the definite integral from [-inf..z] that satisfies p,
127  * accurate to 16 decimal places.
128  *
129  * This implementation is based on:
130  *
131  *   Wichura, M.J. (1988) Algorithm AS 241: The percentage points of the normal
132  *   distribution.  Applied Statistics 37(3):477-484.
133  */
134 JEMALLOC_INLINE double
pt_norm(double p)135 pt_norm(double p)
136 {
137 	double q, r, ret;
138 
139 	assert(p > 0.0 && p < 1.0);
140 
141 	q = p - 0.5;
142 	if (fabs(q) <= 0.425) {
143 		/* p close to 1/2. */
144 		r = 0.180625 - q * q;
145 		return (q * (((((((2.5090809287301226727e3 * r +
146 		    3.3430575583588128105e4) * r + 6.7265770927008700853e4) * r
147 		    + 4.5921953931549871457e4) * r + 1.3731693765509461125e4) *
148 		    r + 1.9715909503065514427e3) * r + 1.3314166789178437745e2)
149 		    * r + 3.3871328727963666080e0) /
150 		    (((((((5.2264952788528545610e3 * r +
151 		    2.8729085735721942674e4) * r + 3.9307895800092710610e4) * r
152 		    + 2.1213794301586595867e4) * r + 5.3941960214247511077e3) *
153 		    r + 6.8718700749205790830e2) * r + 4.2313330701600911252e1)
154 		    * r + 1.0));
155 	} else {
156 		if (q < 0.0)
157 			r = p;
158 		else
159 			r = 1.0 - p;
160 		assert(r > 0.0);
161 
162 		r = sqrt(-log(r));
163 		if (r <= 5.0) {
164 			/* p neither close to 1/2 nor 0 or 1. */
165 			r -= 1.6;
166 			ret = ((((((((7.74545014278341407640e-4 * r +
167 			    2.27238449892691845833e-2) * r +
168 			    2.41780725177450611770e-1) * r +
169 			    1.27045825245236838258e0) * r +
170 			    3.64784832476320460504e0) * r +
171 			    5.76949722146069140550e0) * r +
172 			    4.63033784615654529590e0) * r +
173 			    1.42343711074968357734e0) /
174 			    (((((((1.05075007164441684324e-9 * r +
175 			    5.47593808499534494600e-4) * r +
176 			    1.51986665636164571966e-2)
177 			    * r + 1.48103976427480074590e-1) * r +
178 			    6.89767334985100004550e-1) * r +
179 			    1.67638483018380384940e0) * r +
180 			    2.05319162663775882187e0) * r + 1.0));
181 		} else {
182 			/* p near 0 or 1. */
183 			r -= 5.0;
184 			ret = ((((((((2.01033439929228813265e-7 * r +
185 			    2.71155556874348757815e-5) * r +
186 			    1.24266094738807843860e-3) * r +
187 			    2.65321895265761230930e-2) * r +
188 			    2.96560571828504891230e-1) * r +
189 			    1.78482653991729133580e0) * r +
190 			    5.46378491116411436990e0) * r +
191 			    6.65790464350110377720e0) /
192 			    (((((((2.04426310338993978564e-15 * r +
193 			    1.42151175831644588870e-7) * r +
194 			    1.84631831751005468180e-5) * r +
195 			    7.86869131145613259100e-4) * r +
196 			    1.48753612908506148525e-2) * r +
197 			    1.36929880922735805310e-1) * r +
198 			    5.99832206555887937690e-1)
199 			    * r + 1.0));
200 		}
201 		if (q < 0.0)
202 			ret = -ret;
203 		return (ret);
204 	}
205 }
206 
207 /*
208  * Given a value p in [0..1] of the lower tail area of the Chi^2 distribution
209  * with df degrees of freedom, where ln_gamma_df_2 is ln_gamma(df/2.0), compute
210  * the upper limit on the definite integral from [0..z] that satisfies p,
211  * accurate to 12 decimal places.
212  *
213  * This implementation is based on:
214  *
215  *   Best, D.J., D.E. Roberts (1975) Algorithm AS 91: The percentage points of
216  *   the Chi^2 distribution.  Applied Statistics 24(3):385-388.
217  *
218  *   Shea, B.L. (1991) Algorithm AS R85: A remark on AS 91: The percentage
219  *   points of the Chi^2 distribution.  Applied Statistics 40(1):233-235.
220  */
221 JEMALLOC_INLINE double
pt_chi2(double p,double df,double ln_gamma_df_2)222 pt_chi2(double p, double df, double ln_gamma_df_2)
223 {
224 	double e, aa, xx, c, ch, a, q, p1, p2, t, x, b, s1, s2, s3, s4, s5, s6;
225 	unsigned i;
226 
227 	assert(p >= 0.0 && p < 1.0);
228 	assert(df > 0.0);
229 
230 	e = 5.0e-7;
231 	aa = 0.6931471805;
232 
233 	xx = 0.5 * df;
234 	c = xx - 1.0;
235 
236 	if (df < -1.24 * log(p)) {
237 		/* Starting approximation for small Chi^2. */
238 		ch = pow(p * xx * exp(ln_gamma_df_2 + xx * aa), 1.0 / xx);
239 		if (ch - e < 0.0)
240 			return (ch);
241 	} else {
242 		if (df > 0.32) {
243 			x = pt_norm(p);
244 			/*
245 			 * Starting approximation using Wilson and Hilferty
246 			 * estimate.
247 			 */
248 			p1 = 0.222222 / df;
249 			ch = df * pow(x * sqrt(p1) + 1.0 - p1, 3.0);
250 			/* Starting approximation for p tending to 1. */
251 			if (ch > 2.2 * df + 6.0) {
252 				ch = -2.0 * (log(1.0 - p) - c * log(0.5 * ch) +
253 				    ln_gamma_df_2);
254 			}
255 		} else {
256 			ch = 0.4;
257 			a = log(1.0 - p);
258 			while (true) {
259 				q = ch;
260 				p1 = 1.0 + ch * (4.67 + ch);
261 				p2 = ch * (6.73 + ch * (6.66 + ch));
262 				t = -0.5 + (4.67 + 2.0 * ch) / p1 - (6.73 + ch
263 				    * (13.32 + 3.0 * ch)) / p2;
264 				ch -= (1.0 - exp(a + ln_gamma_df_2 + 0.5 * ch +
265 				    c * aa) * p2 / p1) / t;
266 				if (fabs(q / ch - 1.0) - 0.01 <= 0.0)
267 					break;
268 			}
269 		}
270 	}
271 
272 	for (i = 0; i < 20; i++) {
273 		/* Calculation of seven-term Taylor series. */
274 		q = ch;
275 		p1 = 0.5 * ch;
276 		if (p1 < 0.0)
277 			return (-1.0);
278 		p2 = p - i_gamma(p1, xx, ln_gamma_df_2);
279 		t = p2 * exp(xx * aa + ln_gamma_df_2 + p1 - c * log(ch));
280 		b = t / ch;
281 		a = 0.5 * t - b * c;
282 		s1 = (210.0 + a * (140.0 + a * (105.0 + a * (84.0 + a * (70.0 +
283 		    60.0 * a))))) / 420.0;
284 		s2 = (420.0 + a * (735.0 + a * (966.0 + a * (1141.0 + 1278.0 *
285 		    a)))) / 2520.0;
286 		s3 = (210.0 + a * (462.0 + a * (707.0 + 932.0 * a))) / 2520.0;
287 		s4 = (252.0 + a * (672.0 + 1182.0 * a) + c * (294.0 + a *
288 		    (889.0 + 1740.0 * a))) / 5040.0;
289 		s5 = (84.0 + 264.0 * a + c * (175.0 + 606.0 * a)) / 2520.0;
290 		s6 = (120.0 + c * (346.0 + 127.0 * c)) / 5040.0;
291 		ch += t * (1.0 + 0.5 * t * s1 - b * c * (s1 - b * (s2 - b * (s3
292 		    - b * (s4 - b * (s5 - b * s6))))));
293 		if (fabs(q / ch - 1.0) <= e)
294 			break;
295 	}
296 
297 	return (ch);
298 }
299 
300 /*
301  * Given a value p in [0..1] and Gamma distribution shape and scale parameters,
302  * compute the upper limit on the definite integral from [0..z] that satisfies
303  * p.
304  */
305 JEMALLOC_INLINE double
pt_gamma(double p,double shape,double scale,double ln_gamma_shape)306 pt_gamma(double p, double shape, double scale, double ln_gamma_shape)
307 {
308 
309 	return (pt_chi2(p, shape * 2.0, ln_gamma_shape) * 0.5 * scale);
310 }
311 #endif
312