...
Run Format

Source file src/math/erf.go

Documentation: math

  // Copyright 2010 The Go Authors. All rights reserved.
  // Use of this source code is governed by a BSD-style
  // license that can be found in the LICENSE file.
  
  package math
  
  /*
  	Floating-point error function and complementary error function.
  */
  
  // The original C code and the long comment below are
  // from FreeBSD's /usr/src/lib/msun/src/s_erf.c and
  // came with this notice. The go code is a simplified
  // version of the original C.
  //
  // ====================================================
  // Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
  //
  // Developed at SunPro, a Sun Microsystems, Inc. business.
  // Permission to use, copy, modify, and distribute this
  // software is freely granted, provided that this notice
  // is preserved.
  // ====================================================
  //
  //
  // double erf(double x)
  // double erfc(double x)
  //                           x
  //                    2      |\
  //     erf(x)  =  ---------  | exp(-t*t)dt
  //                 sqrt(pi) \|
  //                           0
  //
  //     erfc(x) =  1-erf(x)
  //  Note that
  //              erf(-x) = -erf(x)
  //              erfc(-x) = 2 - erfc(x)
  //
  // Method:
  //      1. For |x| in [0, 0.84375]
  //          erf(x)  = x + x*R(x**2)
  //          erfc(x) = 1 - erf(x)           if x in [-.84375,0.25]
  //                  = 0.5 + ((0.5-x)-x*R)  if x in [0.25,0.84375]
  //         where R = P/Q where P is an odd poly of degree 8 and
  //         Q is an odd poly of degree 10.
  //                                               -57.90
  //                      | R - (erf(x)-x)/x | <= 2
  //
  //
  //         Remark. The formula is derived by noting
  //          erf(x) = (2/sqrt(pi))*(x - x**3/3 + x**5/10 - x**7/42 + ....)
  //         and that
  //          2/sqrt(pi) = 1.128379167095512573896158903121545171688
  //         is close to one. The interval is chosen because the fix
  //         point of erf(x) is near 0.6174 (i.e., erf(x)=x when x is
  //         near 0.6174), and by some experiment, 0.84375 is chosen to
  //         guarantee the error is less than one ulp for erf.
  //
  //      2. For |x| in [0.84375,1.25], let s = |x| - 1, and
  //         c = 0.84506291151 rounded to single (24 bits)
  //              erf(x)  = sign(x) * (c  + P1(s)/Q1(s))
  //              erfc(x) = (1-c)  - P1(s)/Q1(s) if x > 0
  //                        1+(c+P1(s)/Q1(s))    if x < 0
  //              |P1/Q1 - (erf(|x|)-c)| <= 2**-59.06
  //         Remark: here we use the taylor series expansion at x=1.
  //              erf(1+s) = erf(1) + s*Poly(s)
  //                       = 0.845.. + P1(s)/Q1(s)
  //         That is, we use rational approximation to approximate
  //                      erf(1+s) - (c = (single)0.84506291151)
  //         Note that |P1/Q1|< 0.078 for x in [0.84375,1.25]
  //         where
  //              P1(s) = degree 6 poly in s
  //              Q1(s) = degree 6 poly in s
  //
  //      3. For x in [1.25,1/0.35(~2.857143)],
  //              erfc(x) = (1/x)*exp(-x*x-0.5625+R1/S1)
  //              erf(x)  = 1 - erfc(x)
  //         where
  //              R1(z) = degree 7 poly in z, (z=1/x**2)
  //              S1(z) = degree 8 poly in z
  //
  //      4. For x in [1/0.35,28]
  //              erfc(x) = (1/x)*exp(-x*x-0.5625+R2/S2) if x > 0
  //                      = 2.0 - (1/x)*exp(-x*x-0.5625+R2/S2) if -6<x<0
  //                      = 2.0 - tiny            (if x <= -6)
  //              erf(x)  = sign(x)*(1.0 - erfc(x)) if x < 6, else
  //              erf(x)  = sign(x)*(1.0 - tiny)
  //         where
  //              R2(z) = degree 6 poly in z, (z=1/x**2)
  //              S2(z) = degree 7 poly in z
  //
  //      Note1:
  //         To compute exp(-x*x-0.5625+R/S), let s be a single
  //         precision number and s := x; then
  //              -x*x = -s*s + (s-x)*(s+x)
  //              exp(-x*x-0.5626+R/S) =
  //                      exp(-s*s-0.5625)*exp((s-x)*(s+x)+R/S);
  //      Note2:
  //         Here 4 and 5 make use of the asymptotic series
  //                        exp(-x*x)
  //              erfc(x) ~ ---------- * ( 1 + Poly(1/x**2) )
  //                        x*sqrt(pi)
  //         We use rational approximation to approximate
  //              g(s)=f(1/x**2) = log(erfc(x)*x) - x*x + 0.5625
  //         Here is the error bound for R1/S1 and R2/S2
  //              |R1/S1 - f(x)|  < 2**(-62.57)
  //              |R2/S2 - f(x)|  < 2**(-61.52)
  //
  //      5. For inf > x >= 28
  //              erf(x)  = sign(x) *(1 - tiny)  (raise inexact)
  //              erfc(x) = tiny*tiny (raise underflow) if x > 0
  //                      = 2 - tiny if x<0
  //
  //      7. Special case:
  //              erf(0)  = 0, erf(inf)  = 1, erf(-inf) = -1,
  //              erfc(0) = 1, erfc(inf) = 0, erfc(-inf) = 2,
  //              erfc/erf(NaN) is NaN
  
  const (
  	erx = 8.45062911510467529297e-01 // 0x3FEB0AC160000000
  	// Coefficients for approximation to  erf in [0, 0.84375]
  	efx  = 1.28379167095512586316e-01  // 0x3FC06EBA8214DB69
  	efx8 = 1.02703333676410069053e+00  // 0x3FF06EBA8214DB69
  	pp0  = 1.28379167095512558561e-01  // 0x3FC06EBA8214DB68
  	pp1  = -3.25042107247001499370e-01 // 0xBFD4CD7D691CB913
  	pp2  = -2.84817495755985104766e-02 // 0xBF9D2A51DBD7194F
  	pp3  = -5.77027029648944159157e-03 // 0xBF77A291236668E4
  	pp4  = -2.37630166566501626084e-05 // 0xBEF8EAD6120016AC
  	qq1  = 3.97917223959155352819e-01  // 0x3FD97779CDDADC09
  	qq2  = 6.50222499887672944485e-02  // 0x3FB0A54C5536CEBA
  	qq3  = 5.08130628187576562776e-03  // 0x3F74D022C4D36B0F
  	qq4  = 1.32494738004321644526e-04  // 0x3F215DC9221C1A10
  	qq5  = -3.96022827877536812320e-06 // 0xBED09C4342A26120
  	// Coefficients for approximation to  erf  in [0.84375, 1.25]
  	pa0 = -2.36211856075265944077e-03 // 0xBF6359B8BEF77538
  	pa1 = 4.14856118683748331666e-01  // 0x3FDA8D00AD92B34D
  	pa2 = -3.72207876035701323847e-01 // 0xBFD7D240FBB8C3F1
  	pa3 = 3.18346619901161753674e-01  // 0x3FD45FCA805120E4
  	pa4 = -1.10894694282396677476e-01 // 0xBFBC63983D3E28EC
  	pa5 = 3.54783043256182359371e-02  // 0x3FA22A36599795EB
  	pa6 = -2.16637559486879084300e-03 // 0xBF61BF380A96073F
  	qa1 = 1.06420880400844228286e-01  // 0x3FBB3E6618EEE323
  	qa2 = 5.40397917702171048937e-01  // 0x3FE14AF092EB6F33
  	qa3 = 7.18286544141962662868e-02  // 0x3FB2635CD99FE9A7
  	qa4 = 1.26171219808761642112e-01  // 0x3FC02660E763351F
  	qa5 = 1.36370839120290507362e-02  // 0x3F8BEDC26B51DD1C
  	qa6 = 1.19844998467991074170e-02  // 0x3F888B545735151D
  	// Coefficients for approximation to  erfc in [1.25, 1/0.35]
  	ra0 = -9.86494403484714822705e-03 // 0xBF843412600D6435
  	ra1 = -6.93858572707181764372e-01 // 0xBFE63416E4BA7360
  	ra2 = -1.05586262253232909814e+01 // 0xC0251E0441B0E726
  	ra3 = -6.23753324503260060396e+01 // 0xC04F300AE4CBA38D
  	ra4 = -1.62396669462573470355e+02 // 0xC0644CB184282266
  	ra5 = -1.84605092906711035994e+02 // 0xC067135CEBCCABB2
  	ra6 = -8.12874355063065934246e+01 // 0xC054526557E4D2F2
  	ra7 = -9.81432934416914548592e+00 // 0xC023A0EFC69AC25C
  	sa1 = 1.96512716674392571292e+01  // 0x4033A6B9BD707687
  	sa2 = 1.37657754143519042600e+02  // 0x4061350C526AE721
  	sa3 = 4.34565877475229228821e+02  // 0x407B290DD58A1A71
  	sa4 = 6.45387271733267880336e+02  // 0x40842B1921EC2868
  	sa5 = 4.29008140027567833386e+02  // 0x407AD02157700314
  	sa6 = 1.08635005541779435134e+02  // 0x405B28A3EE48AE2C
  	sa7 = 6.57024977031928170135e+00  // 0x401A47EF8E484A93
  	sa8 = -6.04244152148580987438e-02 // 0xBFAEEFF2EE749A62
  	// Coefficients for approximation to  erfc in [1/.35, 28]
  	rb0 = -9.86494292470009928597e-03 // 0xBF84341239E86F4A
  	rb1 = -7.99283237680523006574e-01 // 0xBFE993BA70C285DE
  	rb2 = -1.77579549177547519889e+01 // 0xC031C209555F995A
  	rb3 = -1.60636384855821916062e+02 // 0xC064145D43C5ED98
  	rb4 = -6.37566443368389627722e+02 // 0xC083EC881375F228
  	rb5 = -1.02509513161107724954e+03 // 0xC09004616A2E5992
  	rb6 = -4.83519191608651397019e+02 // 0xC07E384E9BDC383F
  	sb1 = 3.03380607434824582924e+01  // 0x403E568B261D5190
  	sb2 = 3.25792512996573918826e+02  // 0x40745CAE221B9F0A
  	sb3 = 1.53672958608443695994e+03  // 0x409802EB189D5118
  	sb4 = 3.19985821950859553908e+03  // 0x40A8FFB7688C246A
  	sb5 = 2.55305040643316442583e+03  // 0x40A3F219CEDF3BE6
  	sb6 = 4.74528541206955367215e+02  // 0x407DA874E79FE763
  	sb7 = -2.24409524465858183362e+01 // 0xC03670E242712D62
  )
  
  // Erf returns the error function of x.
  //
  // Special cases are:
  //	Erf(+Inf) = 1
  //	Erf(-Inf) = -1
  //	Erf(NaN) = NaN
  func Erf(x float64) float64 {
  	const (
  		VeryTiny = 2.848094538889218e-306 // 0x0080000000000000
  		Small    = 1.0 / (1 << 28)        // 2**-28
  	)
  	// special cases
  	switch {
  	case IsNaN(x):
  		return NaN()
  	case IsInf(x, 1):
  		return 1
  	case IsInf(x, -1):
  		return -1
  	}
  	sign := false
  	if x < 0 {
  		x = -x
  		sign = true
  	}
  	if x < 0.84375 { // |x| < 0.84375
  		var temp float64
  		if x < Small { // |x| < 2**-28
  			if x < VeryTiny {
  				temp = 0.125 * (8.0*x + efx8*x) // avoid underflow
  			} else {
  				temp = x + efx*x
  			}
  		} else {
  			z := x * x
  			r := pp0 + z*(pp1+z*(pp2+z*(pp3+z*pp4)))
  			s := 1 + z*(qq1+z*(qq2+z*(qq3+z*(qq4+z*qq5))))
  			y := r / s
  			temp = x + x*y
  		}
  		if sign {
  			return -temp
  		}
  		return temp
  	}
  	if x < 1.25 { // 0.84375 <= |x| < 1.25
  		s := x - 1
  		P := pa0 + s*(pa1+s*(pa2+s*(pa3+s*(pa4+s*(pa5+s*pa6)))))
  		Q := 1 + s*(qa1+s*(qa2+s*(qa3+s*(qa4+s*(qa5+s*qa6)))))
  		if sign {
  			return -erx - P/Q
  		}
  		return erx + P/Q
  	}
  	if x >= 6 { // inf > |x| >= 6
  		if sign {
  			return -1
  		}
  		return 1
  	}
  	s := 1 / (x * x)
  	var R, S float64
  	if x < 1/0.35 { // |x| < 1 / 0.35  ~ 2.857143
  		R = ra0 + s*(ra1+s*(ra2+s*(ra3+s*(ra4+s*(ra5+s*(ra6+s*ra7))))))
  		S = 1 + s*(sa1+s*(sa2+s*(sa3+s*(sa4+s*(sa5+s*(sa6+s*(sa7+s*sa8)))))))
  	} else { // |x| >= 1 / 0.35  ~ 2.857143
  		R = rb0 + s*(rb1+s*(rb2+s*(rb3+s*(rb4+s*(rb5+s*rb6)))))
  		S = 1 + s*(sb1+s*(sb2+s*(sb3+s*(sb4+s*(sb5+s*(sb6+s*sb7))))))
  	}
  	z := Float64frombits(Float64bits(x) & 0xffffffff00000000) // pseudo-single (20-bit) precision x
  	r := Exp(-z*z-0.5625) * Exp((z-x)*(z+x)+R/S)
  	if sign {
  		return r/x - 1
  	}
  	return 1 - r/x
  }
  
  // Erfc returns the complementary error function of x.
  //
  // Special cases are:
  //	Erfc(+Inf) = 0
  //	Erfc(-Inf) = 2
  //	Erfc(NaN) = NaN
  func Erfc(x float64) float64 {
  	const Tiny = 1.0 / (1 << 56) // 2**-56
  	// special cases
  	switch {
  	case IsNaN(x):
  		return NaN()
  	case IsInf(x, 1):
  		return 0
  	case IsInf(x, -1):
  		return 2
  	}
  	sign := false
  	if x < 0 {
  		x = -x
  		sign = true
  	}
  	if x < 0.84375 { // |x| < 0.84375
  		var temp float64
  		if x < Tiny { // |x| < 2**-56
  			temp = x
  		} else {
  			z := x * x
  			r := pp0 + z*(pp1+z*(pp2+z*(pp3+z*pp4)))
  			s := 1 + z*(qq1+z*(qq2+z*(qq3+z*(qq4+z*qq5))))
  			y := r / s
  			if x < 0.25 { // |x| < 1/4
  				temp = x + x*y
  			} else {
  				temp = 0.5 + (x*y + (x - 0.5))
  			}
  		}
  		if sign {
  			return 1 + temp
  		}
  		return 1 - temp
  	}
  	if x < 1.25 { // 0.84375 <= |x| < 1.25
  		s := x - 1
  		P := pa0 + s*(pa1+s*(pa2+s*(pa3+s*(pa4+s*(pa5+s*pa6)))))
  		Q := 1 + s*(qa1+s*(qa2+s*(qa3+s*(qa4+s*(qa5+s*qa6)))))
  		if sign {
  			return 1 + erx + P/Q
  		}
  		return 1 - erx - P/Q
  
  	}
  	if x < 28 { // |x| < 28
  		s := 1 / (x * x)
  		var R, S float64
  		if x < 1/0.35 { // |x| < 1 / 0.35 ~ 2.857143
  			R = ra0 + s*(ra1+s*(ra2+s*(ra3+s*(ra4+s*(ra5+s*(ra6+s*ra7))))))
  			S = 1 + s*(sa1+s*(sa2+s*(sa3+s*(sa4+s*(sa5+s*(sa6+s*(sa7+s*sa8)))))))
  		} else { // |x| >= 1 / 0.35 ~ 2.857143
  			if sign && x > 6 {
  				return 2 // x < -6
  			}
  			R = rb0 + s*(rb1+s*(rb2+s*(rb3+s*(rb4+s*(rb5+s*rb6)))))
  			S = 1 + s*(sb1+s*(sb2+s*(sb3+s*(sb4+s*(sb5+s*(sb6+s*sb7))))))
  		}
  		z := Float64frombits(Float64bits(x) & 0xffffffff00000000) // pseudo-single (20-bit) precision x
  		r := Exp(-z*z-0.5625) * Exp((z-x)*(z+x)+R/S)
  		if sign {
  			return 2 - r/x
  		}
  		return r / x
  	}
  	if sign {
  		return 2
  	}
  	return 0
  }
  

View as plain text