描述
Consider the function
f: x -> sqrt(1 + x) - 1 at x = 1e-15.
We get: f(x) = 4.44089209850062616e-16
or something around that, depending on the language.
This function involves the subtraction of a pair of similar numbers when x is near 0 and the results are significantly erroneous in this region. Using pow instead of sqrt doesn’t give better results.
A “good” answer is 4.99999999999999875… * 1e-16.
Can you modify f(x) to give a good approximation of f(x) in the neigbourhood of 0?
分析
当两个非常接近0的浮点数做减法时,会出现结果不精确的情况。我们可以通过将减法转换为其它运算解决这个问题。
由:
(
1
+
x
+
1
)
×
(
1
+
x
−
1
)
=
x
(\sqrt{1+x}+1)\times(\sqrt{1+x}-1)=x
(1+x+1)×(1+x−1)=x
得到:
1
+
x
−
1
=
x
1
+
x
+
1
\sqrt{1+x}-1=\frac{x}{\sqrt{1+x}+1}
1+x−1=1+x+1x
实现
import "math"
func F(x float64) float64 {
return x / (math.Sqrt(float64(1 + x)) + 1)
}