symmetry log of sum of exp

Average Accuracy: 53.5% → 98.3%

Time: 15.7s

Precision: binary64

Cost: 32448

\[ \begin{array}{c}[a, b] = \mathsf{sort}([a, b])\\ \end{array} \]

Math

\[\log \left(e^{a} + e^{b}\right) \]

↓

\[\mathsf{log1p}\left(e^{a}\right) + \frac{b}{\mathsf{expm1}\left(\mathsf{log1p}\left(e^{a} + 1\right)\right)} \]

FPCore

(FPCore (a b) :precision binary64 (log (+ (exp a) (exp b))))

↓

(FPCore (a b)
 :precision binary64
 (+ (log1p (exp a)) (/ b (expm1 (log1p (+ (exp a) 1.0))))))

C

double code(double a, double b) {
	return log((exp(a) + exp(b)));
}

↓

double code(double a, double b) {
	return log1p(exp(a)) + (b / expm1(log1p((exp(a) + 1.0))));
}

Java

public static double code(double a, double b) {
	return Math.log((Math.exp(a) + Math.exp(b)));
}

↓

public static double code(double a, double b) {
	return Math.log1p(Math.exp(a)) + (b / Math.expm1(Math.log1p((Math.exp(a) + 1.0))));
}

Python

def code(a, b):
	return math.log((math.exp(a) + math.exp(b)))

↓

def code(a, b):
	return math.log1p(math.exp(a)) + (b / math.expm1(math.log1p((math.exp(a) + 1.0))))

Julia

function code(a, b)
	return log(Float64(exp(a) + exp(b)))
end

↓

function code(a, b)
	return Float64(log1p(exp(a)) + Float64(b / expm1(log1p(Float64(exp(a) + 1.0)))))
end

Wolfram

code[a_, b_] := N[Log[N[(N[Exp[a], $MachinePrecision] + N[Exp[b], $MachinePrecision]), $MachinePrecision]], $MachinePrecision]

↓

code[a_, b_] := N[(N[Log[1 + N[Exp[a], $MachinePrecision]], $MachinePrecision] + N[(b / N[(Exp[N[Log[1 + N[(N[Exp[a], $MachinePrecision] + 1.0), $MachinePrecision]], $MachinePrecision]] - 1), $MachinePrecision]), $MachinePrecision]), $MachinePrecision]

Derivation

1. Initial program 53.5%

   \[\log \left(e^{a} + e^{b}\right) \]

2. Taylor expanded in b around 0 98.1%

   \[\leadsto \color{blue}{\log \left(1 + e^{a}\right) + \frac{b}{1 + e^{a}}} \]

3. Simplified 98.3%

   \[\leadsto \color{blue}{\mathsf{log1p}\left(e^{a}\right) + \frac{b}{1 + e^{a}}} \]
   Proof

   [Start] 98.1	\[ \log \left(1 + e^{a}\right) + \frac{b}{1 + e^{a}} \]
   log1p-def [=>] 98.3	\[ \color{blue}{\mathsf{log1p}\left(e^{a}\right)} + \frac{b}{1 + e^{a}} \]

4. Applied egg-rr 98.3%

   \[\leadsto \mathsf{log1p}\left(e^{a}\right) + \frac{b}{\color{blue}{\mathsf{expm1}\left(\mathsf{log1p}\left(1 + e^{a}\right)\right)}} \]
   Proof

   [Start] 98.3	\[ \mathsf{log1p}\left(e^{a}\right) + \frac{b}{1 + e^{a}} \]
   expm1-log1p-u [=>] 98.3	\[ \mathsf{log1p}\left(e^{a}\right) + \frac{b}{\color{blue}{\mathsf{expm1}\left(\mathsf{log1p}\left(1 + e^{a}\right)\right)}} \]

5. Final simplification 98.3%

   \[\leadsto \mathsf{log1p}\left(e^{a}\right) + \frac{b}{\mathsf{expm1}\left(\mathsf{log1p}\left(e^{a} + 1\right)\right)} \]

Alternatives

Alternative 1
Accuracy	98.3%
Cost	19648

\[\mathsf{log1p}\left(e^{a}\right) + \frac{b}{e^{a} + 1} \]

Alternative 2
Accuracy	97.8%
Cost	19392

\[\mathsf{log1p}\left(e^{a} + \mathsf{expm1}\left(b\right)\right) \]

Alternative 3
Accuracy	95.7%
Cost	12992

\[\mathsf{log1p}\left(e^{a} + b\right) \]

Alternative 4
Accuracy	50.3%
Cost	12864

\[\mathsf{log1p}\left(e^{a}\right) \]

Alternative 5
Accuracy	49.0%
Cost	6720

\[b \cdot 0.5 + \log 2 \]

Alternative 6
Accuracy	48.7%
Cost	6592

\[\log \left(b + 2\right) \]

Alternative 7
Accuracy	48.8%
Cost	6592

\[\mathsf{log1p}\left(b + 1\right) \]

Alternative 8
Accuracy	48.3%
Cost	6464

\[\log 2 \]

Alternative 9
Accuracy	2.6%
Cost	192

\[a \cdot 0.5 \]

Reproduce

herbie shell --seed 2023151 
(FPCore (a b)
  :name "symmetry log of sum of exp"
  :precision binary64
  (log (+ (exp a) (exp b))))