Logistic function from Lakshay Garg

Average Error: 29.6 → 0.3

Time: 2.6s

Precision: binary64

Math

\[\frac{2}{1 + e^{-2 \cdot x}} - 1 \]

↓

\[\begin{array}{l} t_0 := \mathsf{expm1}\left(\log 2 - \mathsf{log1p}\left(e^{-2 \cdot x}\right)\right)\\ \mathbf{if}\;-2 \cdot x \leq -913142.5470619951:\\ \;\;\;\;t_0\\ \mathbf{elif}\;-2 \cdot x \leq 0.0008985694067275785:\\ \;\;\;\;\mathsf{fma}\left({x}^{3}, -0.3333333333333333, \mathsf{fma}\left(0.13333333333333333, {x}^{5}, x\right)\right)\\ \mathbf{else}:\\ \;\;\;\;t_0\\ \end{array} \]

FPCore

(FPCore (x y) :precision binary64 (- (/ 2.0 (+ 1.0 (exp (* -2.0 x)))) 1.0))

↓

(FPCore (x y)
 :precision binary64
 (let* ((t_0 (expm1 (- (log 2.0) (log1p (exp (* -2.0 x)))))))
   (if (<= (* -2.0 x) -913142.5470619951)
     t_0
     (if (<= (* -2.0 x) 0.0008985694067275785)
       (fma
        (pow x 3.0)
        -0.3333333333333333
        (fma 0.13333333333333333 (pow x 5.0) x))
       t_0))))

C

double code(double x, double y) {
	return (2.0 / (1.0 + exp((-2.0 * x)))) - 1.0;
}

↓

double code(double x, double y) {
	double t_0 = expm1((log(2.0) - log1p(exp((-2.0 * x)))));
	double tmp;
	if ((-2.0 * x) <= -913142.5470619951) {
		tmp = t_0;
	} else if ((-2.0 * x) <= 0.0008985694067275785) {
		tmp = fma(pow(x, 3.0), -0.3333333333333333, fma(0.13333333333333333, pow(x, 5.0), x));
	} else {
		tmp = t_0;
	}
	return tmp;
}

Julia

function code(x, y)
	return Float64(Float64(2.0 / Float64(1.0 + exp(Float64(-2.0 * x)))) - 1.0)
end

↓

function code(x, y)
	t_0 = expm1(Float64(log(2.0) - log1p(exp(Float64(-2.0 * x)))))
	tmp = 0.0
	if (Float64(-2.0 * x) <= -913142.5470619951)
		tmp = t_0;
	elseif (Float64(-2.0 * x) <= 0.0008985694067275785)
		tmp = fma((x ^ 3.0), -0.3333333333333333, fma(0.13333333333333333, (x ^ 5.0), x));
	else
		tmp = t_0;
	end
	return tmp
end

Wolfram

code[x_, y_] := N[(N[(2.0 / N[(1.0 + N[Exp[N[(-2.0 * x), $MachinePrecision]], $MachinePrecision]), $MachinePrecision]), $MachinePrecision] - 1.0), $MachinePrecision]

↓

code[x_, y_] := Block[{t$95$0 = N[(Exp[N[(N[Log[2.0], $MachinePrecision] - N[Log[1 + N[Exp[N[(-2.0 * x), $MachinePrecision]], $MachinePrecision]], $MachinePrecision]), $MachinePrecision]] - 1), $MachinePrecision]}, If[LessEqual[N[(-2.0 * x), $MachinePrecision], -913142.5470619951], t$95$0, If[LessEqual[N[(-2.0 * x), $MachinePrecision], 0.0008985694067275785], N[(N[Power[x, 3.0], $MachinePrecision] * -0.3333333333333333 + N[(0.13333333333333333 * N[Power[x, 5.0], $MachinePrecision] + x), $MachinePrecision]), $MachinePrecision], t$95$0]]]

Error

Bits error versus x

Bits error versus y

Derivation

1. Split input into 2 regimes

2. if (*.f64 -2 x) < -913142.54706199514 or 8.98569406727578479e-4 < (*.f64 -2 x)

   1. Initial program 0.0

      \[\frac{2}{1 + e^{-2 \cdot x}} - 1 \]

   2. Applied egg-rr 0.0

      \[\leadsto \color{blue}{\mathsf{expm1}\left(\log 2 - \mathsf{log1p}\left(e^{-2 \cdot x}\right)\right)} \]

   if -913142.54706199514 < (*.f64 -2 x) < 8.98569406727578479e-4

   1. Initial program 58.5

      \[\frac{2}{1 + e^{-2 \cdot x}} - 1 \]

   2. Taylor expanded in x around 0 0.6

      \[\leadsto \color{blue}{\left(0.13333333333333333 \cdot {x}^{5} + x\right) - 0.3333333333333333 \cdot {x}^{3}} \]

   3. Simplified 0.6

      \[\leadsto \color{blue}{\mathsf{fma}\left({x}^{3}, -0.3333333333333333, \mathsf{fma}\left(0.13333333333333333, {x}^{5}, x\right)\right)} \]
3. Recombined 2 regimes into one program.

4. Final simplification 0.3

   \[\leadsto \begin{array}{l} \mathbf{if}\;-2 \cdot x \leq -913142.5470619951:\\ \;\;\;\;\mathsf{expm1}\left(\log 2 - \mathsf{log1p}\left(e^{-2 \cdot x}\right)\right)\\ \mathbf{elif}\;-2 \cdot x \leq 0.0008985694067275785:\\ \;\;\;\;\mathsf{fma}\left({x}^{3}, -0.3333333333333333, \mathsf{fma}\left(0.13333333333333333, {x}^{5}, x\right)\right)\\ \mathbf{else}:\\ \;\;\;\;\mathsf{expm1}\left(\log 2 - \mathsf{log1p}\left(e^{-2 \cdot x}\right)\right)\\ \end{array} \]

Reproduce

herbie shell --seed 2022133 
(FPCore (x y)
  :name "Logistic function from Lakshay Garg"
  :precision binary64
  (- (/ 2.0 (+ 1.0 (exp (* -2.0 x)))) 1.0))