Eccentricity of an ellipse

Percentage Accurate: 76.7% → 100.0%
Time: 52.4s
Alternatives: 2
Speedup: 4.8×

Specification

?
\[\left(0 \leq b \land b \leq a\right) \land a \leq 1\]
\[\sqrt{\left|\frac{a \cdot a - b \cdot b}{a \cdot a}\right|} \]
(FPCore (a b)
  :precision binary64
  :pre (and (and (<= 0.0 b) (<= b a)) (<= a 1.0))
  (sqrt (fabs (/ (- (* a a) (* b b)) (* a a)))))
double code(double a, double b) {
	return sqrt(fabs((((a * a) - (b * b)) / (a * a))));
}

real(8) function code(a, b)
use fmin_fmax_functions
    real(8), intent (in) :: a
    real(8), intent (in) :: b
    code = sqrt(abs((((a * a) - (b * b)) / (a * a))))
end function

public static double code(double a, double b) {
	return Math.sqrt(Math.abs((((a * a) - (b * b)) / (a * a))));
}

def code(a, b):
	return math.sqrt(math.fabs((((a * a) - (b * b)) / (a * a))))

function code(a, b)
	return sqrt(abs(Float64(Float64(Float64(a * a) - Float64(b * b)) / Float64(a * a))))
end

function tmp = code(a, b)
	tmp = sqrt(abs((((a * a) - (b * b)) / (a * a))));
end

code[a_, b_] := N[Sqrt[N[Abs[N[(N[(N[(a * a), $MachinePrecision] - N[(b * b), $MachinePrecision]), $MachinePrecision] / N[(a * a), $MachinePrecision]), $MachinePrecision]], $MachinePrecision]], $MachinePrecision]

f(a, b):
	a in [-inf, 1],
	b in [0, +inf]
code: THEORY
BEGIN
f(a, b: real): real =
	sqrt((abs((((a * a) - (b * b)) / (a * a)))))
END code
\sqrt{\left|\frac{a \cdot a - b \cdot b}{a \cdot a}\right|}

Local Percentage Accuracy vs ?

The average percentage accuracy by input value. Horizontal axis shows value of an input variable; the variable is choosen in the title. Vertical axis is accuracy; higher is better. Red represent the original program, while blue represents Herbie's suggestion. These can be toggled with buttons below the plot. The line is an average while dots represent individual samples.

Accuracy vs Speed?

Herbie found 2 alternatives:

AlternativeAccuracySpeedup
The accuracy (vertical axis) and speed (horizontal axis) of each alternatives. Up and to the right is better. The red square shows the initial program, and each blue circle shows an alternative.The line shows the best available speed-accuracy tradeoffs.

Initial Program: 76.7% accurate, 1.0× speedup?

\[\left(0 \leq b \land b \leq a\right) \land a \leq 1\]
\[\sqrt{\left|\frac{a \cdot a - b \cdot b}{a \cdot a}\right|} \]
(FPCore (a b)
  :precision binary64
  :pre (and (and (<= 0.0 b) (<= b a)) (<= a 1.0))
  (sqrt (fabs (/ (- (* a a) (* b b)) (* a a)))))
double code(double a, double b) {
	return sqrt(fabs((((a * a) - (b * b)) / (a * a))));
}

real(8) function code(a, b)
use fmin_fmax_functions
    real(8), intent (in) :: a
    real(8), intent (in) :: b
    code = sqrt(abs((((a * a) - (b * b)) / (a * a))))
end function

public static double code(double a, double b) {
	return Math.sqrt(Math.abs((((a * a) - (b * b)) / (a * a))));
}

def code(a, b):
	return math.sqrt(math.fabs((((a * a) - (b * b)) / (a * a))))

function code(a, b)
	return sqrt(abs(Float64(Float64(Float64(a * a) - Float64(b * b)) / Float64(a * a))))
end

function tmp = code(a, b)
	tmp = sqrt(abs((((a * a) - (b * b)) / (a * a))));
end

code[a_, b_] := N[Sqrt[N[Abs[N[(N[(N[(a * a), $MachinePrecision] - N[(b * b), $MachinePrecision]), $MachinePrecision] / N[(a * a), $MachinePrecision]), $MachinePrecision]], $MachinePrecision]], $MachinePrecision]

f(a, b):
	a in [-inf, 1],
	b in [0, +inf]
code: THEORY
BEGIN
f(a, b: real): real =
	sqrt((abs((((a * a) - (b * b)) / (a * a)))))
END code
\sqrt{\left|\frac{a \cdot a - b \cdot b}{a \cdot a}\right|}

Alternative 1: 100.0% accurate, 1.2× speedup?

\[\left(0 \leq b \land b \leq a\right) \land a \leq 1\]
\[\sqrt{\left|\mathsf{fma}\left(\frac{b}{a}, \frac{b}{a}, -1\right)\right|} \]
(FPCore (a b)
  :precision binary64
  :pre (and (and (<= 0.0 b) (<= b a)) (<= a 1.0))
  (sqrt (fabs (fma (/ b a) (/ b a) -1.0))))
double code(double a, double b) {
	return sqrt(fabs(fma((b / a), (b / a), -1.0)));
}

function code(a, b)
	return sqrt(abs(fma(Float64(b / a), Float64(b / a), -1.0)))
end

code[a_, b_] := N[Sqrt[N[Abs[N[(N[(b / a), $MachinePrecision] * N[(b / a), $MachinePrecision] + -1.0), $MachinePrecision]], $MachinePrecision]], $MachinePrecision]

f(a, b):
	a in [-inf, 1],
	b in [0, +inf]
code: THEORY
BEGIN
f(a, b: real): real =
	sqrt((abs((((b / a) * (b / a)) + (-1)))))
END code
\sqrt{\left|\mathsf{fma}\left(\frac{b}{a}, \frac{b}{a}, -1\right)\right|}
Derivation

  1. Initial program 76.7%

    \[\sqrt{\left|\frac{a \cdot a - b \cdot b}{a \cdot a}\right|} \]

  3. Applied rewrites77.5%

    \[\leadsto \sqrt{\left|\mathsf{fma}\left(b, \frac{b}{a \cdot a}, -1\right)\right|} \]

  5. Applied rewrites100.0%

    \[\leadsto \sqrt{\left|\mathsf{fma}\left(\frac{b}{a}, \frac{b}{a}, -1\right)\right|} \]
  6. Add Preprocessing

Alternative 2: 98.0% accurate, 4.8× speedup?

\[\left(0 \leq b \land b \leq a\right) \land a \leq 1\]
\[\sqrt{\left|1\right|} \]
(FPCore (a b)
  :precision binary64
  :pre (and (and (<= 0.0 b) (<= b a)) (<= a 1.0))
  (sqrt (fabs 1.0)))
double code(double a, double b) {
	return sqrt(fabs(1.0));
}

real(8) function code(a, b)
use fmin_fmax_functions
    real(8), intent (in) :: a
    real(8), intent (in) :: b
    code = sqrt(abs(1.0d0))
end function

public static double code(double a, double b) {
	return Math.sqrt(Math.abs(1.0));
}

def code(a, b):
	return math.sqrt(math.fabs(1.0))

function code(a, b)
	return sqrt(abs(1.0))
end

function tmp = code(a, b)
	tmp = sqrt(abs(1.0));
end

code[a_, b_] := N[Sqrt[N[Abs[1.0], $MachinePrecision]], $MachinePrecision]

f(a, b):
	a in [-inf, 1],
	b in [0, +inf]
code: THEORY
BEGIN
f(a, b: real): real =
	sqrt((abs((1))))
END code
\sqrt{\left|1\right|}
Derivation

  1. Initial program 76.7%

    \[\sqrt{\left|\frac{a \cdot a - b \cdot b}{a \cdot a}\right|} \]

  2. Taylor expanded in a around inf

    \[\leadsto \sqrt{\left|1\right|} \]

  4. Applied rewrites98.0%

    \[\leadsto \sqrt{\left|1\right|} \]
  5. Add Preprocessing

Reproduce

?
herbie shell --seed 2026074 +o generate:egglog
(FPCore (a b)
  :name "Eccentricity of an ellipse"
  :precision binary64
  :pre (and (and (<= 0.0 b) (<= b a)) (<= a 1.0))
  (sqrt (fabs (/ (- (* a a) (* b b)) (* a a)))))