| 1 |
747 |
jeremybenn |
// Copyright 2010 The Go Authors. All rights reserved.
|
| 2 |
|
|
// Use of this source code is governed by a BSD-style
|
| 3 |
|
|
// license that can be found in the LICENSE file.
|
| 4 |
|
|
|
| 5 |
|
|
package math
|
| 6 |
|
|
|
| 7 |
|
|
// The original C code, the long comment, and the constants
|
| 8 |
|
|
// below are from FreeBSD's /usr/src/lib/msun/src/s_log1p.c
|
| 9 |
|
|
// and came with this notice. The go code is a simplified
|
| 10 |
|
|
// version of the original C.
|
| 11 |
|
|
//
|
| 12 |
|
|
// ====================================================
|
| 13 |
|
|
// Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
|
| 14 |
|
|
//
|
| 15 |
|
|
// Developed at SunPro, a Sun Microsystems, Inc. business.
|
| 16 |
|
|
// Permission to use, copy, modify, and distribute this
|
| 17 |
|
|
// software is freely granted, provided that this notice
|
| 18 |
|
|
// is preserved.
|
| 19 |
|
|
// ====================================================
|
| 20 |
|
|
//
|
| 21 |
|
|
//
|
| 22 |
|
|
// double log1p(double x)
|
| 23 |
|
|
//
|
| 24 |
|
|
// Method :
|
| 25 |
|
|
// 1. Argument Reduction: find k and f such that
|
| 26 |
|
|
// 1+x = 2**k * (1+f),
|
| 27 |
|
|
// where sqrt(2)/2 < 1+f < sqrt(2) .
|
| 28 |
|
|
//
|
| 29 |
|
|
// Note. If k=0, then f=x is exact. However, if k!=0, then f
|
| 30 |
|
|
// may not be representable exactly. In that case, a correction
|
| 31 |
|
|
// term is need. Let u=1+x rounded. Let c = (1+x)-u, then
|
| 32 |
|
|
// log(1+x) - log(u) ~ c/u. Thus, we proceed to compute log(u),
|
| 33 |
|
|
// and add back the correction term c/u.
|
| 34 |
|
|
// (Note: when x > 2**53, one can simply return log(x))
|
| 35 |
|
|
//
|
| 36 |
|
|
// 2. Approximation of log1p(f).
|
| 37 |
|
|
// Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s)
|
| 38 |
|
|
// = 2s + 2/3 s**3 + 2/5 s**5 + .....,
|
| 39 |
|
|
// = 2s + s*R
|
| 40 |
|
|
// We use a special Reme algorithm on [0,0.1716] to generate
|
| 41 |
|
|
// a polynomial of degree 14 to approximate R The maximum error
|
| 42 |
|
|
// of this polynomial approximation is bounded by 2**-58.45. In
|
| 43 |
|
|
// other words,
|
| 44 |
|
|
// 2 4 6 8 10 12 14
|
| 45 |
|
|
// R(z) ~ Lp1*s +Lp2*s +Lp3*s +Lp4*s +Lp5*s +Lp6*s +Lp7*s
|
| 46 |
|
|
// (the values of Lp1 to Lp7 are listed in the program)
|
| 47 |
|
|
// and
|
| 48 |
|
|
// | 2 14 | -58.45
|
| 49 |
|
|
// | Lp1*s +...+Lp7*s - R(z) | <= 2
|
| 50 |
|
|
// | |
|
| 51 |
|
|
// Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2.
|
| 52 |
|
|
// In order to guarantee error in log below 1ulp, we compute log
|
| 53 |
|
|
// by
|
| 54 |
|
|
// log1p(f) = f - (hfsq - s*(hfsq+R)).
|
| 55 |
|
|
//
|
| 56 |
|
|
// 3. Finally, log1p(x) = k*ln2 + log1p(f).
|
| 57 |
|
|
// = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo)))
|
| 58 |
|
|
// Here ln2 is split into two floating point number:
|
| 59 |
|
|
// ln2_hi + ln2_lo,
|
| 60 |
|
|
// where n*ln2_hi is always exact for |n| < 2000.
|
| 61 |
|
|
//
|
| 62 |
|
|
// Special cases:
|
| 63 |
|
|
// log1p(x) is NaN with signal if x < -1 (including -INF) ;
|
| 64 |
|
|
// log1p(+INF) is +INF; log1p(-1) is -INF with signal;
|
| 65 |
|
|
// log1p(NaN) is that NaN with no signal.
|
| 66 |
|
|
//
|
| 67 |
|
|
// Accuracy:
|
| 68 |
|
|
// according to an error analysis, the error is always less than
|
| 69 |
|
|
// 1 ulp (unit in the last place).
|
| 70 |
|
|
//
|
| 71 |
|
|
// Constants:
|
| 72 |
|
|
// The hexadecimal values are the intended ones for the following
|
| 73 |
|
|
// constants. The decimal values may be used, provided that the
|
| 74 |
|
|
// compiler will convert from decimal to binary accurately enough
|
| 75 |
|
|
// to produce the hexadecimal values shown.
|
| 76 |
|
|
//
|
| 77 |
|
|
// Note: Assuming log() return accurate answer, the following
|
| 78 |
|
|
// algorithm can be used to compute log1p(x) to within a few ULP:
|
| 79 |
|
|
//
|
| 80 |
|
|
// u = 1+x;
|
| 81 |
|
|
// if(u==1.0) return x ; else
|
| 82 |
|
|
// return log(u)*(x/(u-1.0));
|
| 83 |
|
|
//
|
| 84 |
|
|
// See HP-15C Advanced Functions Handbook, p.193.
|
| 85 |
|
|
|
| 86 |
|
|
// Log1p returns the natural logarithm of 1 plus its argument x.
|
| 87 |
|
|
// It is more accurate than Log(1 + x) when x is near zero.
|
| 88 |
|
|
//
|
| 89 |
|
|
// Special cases are:
|
| 90 |
|
|
// Log1p(+Inf) = +Inf
|
| 91 |
|
|
// Log1p(±0) = ±0
|
| 92 |
|
|
// Log1p(-1) = -Inf
|
| 93 |
|
|
// Log1p(x < -1) = NaN
|
| 94 |
|
|
// Log1p(NaN) = NaN
|
| 95 |
|
|
|
| 96 |
|
|
//extern log1p
|
| 97 |
|
|
func libc_log1p(float64) float64
|
| 98 |
|
|
|
| 99 |
|
|
func Log1p(x float64) float64 {
|
| 100 |
|
|
return libc_log1p(x)
|
| 101 |
|
|
}
|
| 102 |
|
|
|
| 103 |
|
|
func log1p(x float64) float64 {
|
| 104 |
|
|
const (
|
| 105 |
|
|
Sqrt2M1 = 4.142135623730950488017e-01 // Sqrt(2)-1 = 0x3fda827999fcef34
|
| 106 |
|
|
Sqrt2HalfM1 = -2.928932188134524755992e-01 // Sqrt(2)/2-1 = 0xbfd2bec333018866
|
| 107 |
|
|
Small = 1.0 / (1 << 29) // 2**-29 = 0x3e20000000000000
|
| 108 |
|
|
Tiny = 1.0 / (1 << 54) // 2**-54
|
| 109 |
|
|
Two53 = 1 << 53 // 2**53
|
| 110 |
|
|
Ln2Hi = 6.93147180369123816490e-01 // 3fe62e42fee00000
|
| 111 |
|
|
Ln2Lo = 1.90821492927058770002e-10 // 3dea39ef35793c76
|
| 112 |
|
|
Lp1 = 6.666666666666735130e-01 // 3FE5555555555593
|
| 113 |
|
|
Lp2 = 3.999999999940941908e-01 // 3FD999999997FA04
|
| 114 |
|
|
Lp3 = 2.857142874366239149e-01 // 3FD2492494229359
|
| 115 |
|
|
Lp4 = 2.222219843214978396e-01 // 3FCC71C51D8E78AF
|
| 116 |
|
|
Lp5 = 1.818357216161805012e-01 // 3FC7466496CB03DE
|
| 117 |
|
|
Lp6 = 1.531383769920937332e-01 // 3FC39A09D078C69F
|
| 118 |
|
|
Lp7 = 1.479819860511658591e-01 // 3FC2F112DF3E5244
|
| 119 |
|
|
)
|
| 120 |
|
|
|
| 121 |
|
|
// special cases
|
| 122 |
|
|
switch {
|
| 123 |
|
|
case x < -1 || IsNaN(x): // includes -Inf
|
| 124 |
|
|
return NaN()
|
| 125 |
|
|
case x == -1:
|
| 126 |
|
|
return Inf(-1)
|
| 127 |
|
|
case IsInf(x, 1):
|
| 128 |
|
|
return Inf(1)
|
| 129 |
|
|
}
|
| 130 |
|
|
|
| 131 |
|
|
absx := x
|
| 132 |
|
|
if absx < 0 {
|
| 133 |
|
|
absx = -absx
|
| 134 |
|
|
}
|
| 135 |
|
|
|
| 136 |
|
|
var f float64
|
| 137 |
|
|
var iu uint64
|
| 138 |
|
|
k := 1
|
| 139 |
|
|
if absx < Sqrt2M1 { // |x| < Sqrt(2)-1
|
| 140 |
|
|
if absx < Small { // |x| < 2**-29
|
| 141 |
|
|
if absx < Tiny { // |x| < 2**-54
|
| 142 |
|
|
return x
|
| 143 |
|
|
}
|
| 144 |
|
|
return x - x*x*0.5
|
| 145 |
|
|
}
|
| 146 |
|
|
if x > Sqrt2HalfM1 { // Sqrt(2)/2-1 < x
|
| 147 |
|
|
// (Sqrt(2)/2-1) < x < (Sqrt(2)-1)
|
| 148 |
|
|
k = 0
|
| 149 |
|
|
f = x
|
| 150 |
|
|
iu = 1
|
| 151 |
|
|
}
|
| 152 |
|
|
}
|
| 153 |
|
|
var c float64
|
| 154 |
|
|
if k != 0 {
|
| 155 |
|
|
var u float64
|
| 156 |
|
|
if absx < Two53 { // 1<<53
|
| 157 |
|
|
u = 1.0 + x
|
| 158 |
|
|
iu = Float64bits(u)
|
| 159 |
|
|
k = int((iu >> 52) - 1023)
|
| 160 |
|
|
if k > 0 {
|
| 161 |
|
|
c = 1.0 - (u - x)
|
| 162 |
|
|
} else {
|
| 163 |
|
|
c = x - (u - 1.0) // correction term
|
| 164 |
|
|
c /= u
|
| 165 |
|
|
}
|
| 166 |
|
|
} else {
|
| 167 |
|
|
u = x
|
| 168 |
|
|
iu = Float64bits(u)
|
| 169 |
|
|
k = int((iu >> 52) - 1023)
|
| 170 |
|
|
c = 0
|
| 171 |
|
|
}
|
| 172 |
|
|
iu &= 0x000fffffffffffff
|
| 173 |
|
|
if iu < 0x0006a09e667f3bcd { // mantissa of Sqrt(2)
|
| 174 |
|
|
u = Float64frombits(iu | 0x3ff0000000000000) // normalize u
|
| 175 |
|
|
} else {
|
| 176 |
|
|
k += 1
|
| 177 |
|
|
u = Float64frombits(iu | 0x3fe0000000000000) // normalize u/2
|
| 178 |
|
|
iu = (0x0010000000000000 - iu) >> 2
|
| 179 |
|
|
}
|
| 180 |
|
|
f = u - 1.0 // Sqrt(2)/2 < u < Sqrt(2)
|
| 181 |
|
|
}
|
| 182 |
|
|
hfsq := 0.5 * f * f
|
| 183 |
|
|
var s, R, z float64
|
| 184 |
|
|
if iu == 0 { // |f| < 2**-20
|
| 185 |
|
|
if f == 0 {
|
| 186 |
|
|
if k == 0 {
|
| 187 |
|
|
return 0
|
| 188 |
|
|
} else {
|
| 189 |
|
|
c += float64(k) * Ln2Lo
|
| 190 |
|
|
return float64(k)*Ln2Hi + c
|
| 191 |
|
|
}
|
| 192 |
|
|
}
|
| 193 |
|
|
R = hfsq * (1.0 - 0.66666666666666666*f) // avoid division
|
| 194 |
|
|
if k == 0 {
|
| 195 |
|
|
return f - R
|
| 196 |
|
|
}
|
| 197 |
|
|
return float64(k)*Ln2Hi - ((R - (float64(k)*Ln2Lo + c)) - f)
|
| 198 |
|
|
}
|
| 199 |
|
|
s = f / (2.0 + f)
|
| 200 |
|
|
z = s * s
|
| 201 |
|
|
R = z * (Lp1 + z*(Lp2+z*(Lp3+z*(Lp4+z*(Lp5+z*(Lp6+z*Lp7))))))
|
| 202 |
|
|
if k == 0 {
|
| 203 |
|
|
return f - (hfsq - s*(hfsq+R))
|
| 204 |
|
|
}
|
| 205 |
|
|
return float64(k)*Ln2Hi - ((hfsq - (s*(hfsq+R) + (float64(k)*Ln2Lo + c))) - f)
|
| 206 |
|
|
}
|