浮点数累加误差

Question:

The floating point sum of $x+y$ has a relative error $ϵ<{ϵ}_m$. ${ϵ}_m$ is the machine epsilon.

double compute_sum(double x[], int n)
{
	double S = 0;
	for (int i = 0; i < n; ++i)
	S += x[i];
	return S;
}

Show that the absolute error in the sum computed in the above code is no worse than $(n-1)\cdot {ϵ}_m{\sum }_{i=0}^{n-1}|x_i|$.

Answer:

$$S=x_0+x_1+\dots +x_{n-1}$$

Considering the error $ϵ$ , let:

$$S_0=(x_0+x_1)\cdot (1+{ϵ}_0)$$

$$S_1=(S_0+x_2)\cdot (1+{ϵ}_1)$$

$$S_2=(S_1+x_3)\cdot (1+{ϵ}_2)$$

$$⋮$$

$$S_{n-2}=(S_{n-3}+x_{n-1})\cdot (1+{ϵ}_{n-2})$$

Considering $S_0$ & $S_1$, we have:

$$\begin{array}{rl}{S}_0& =(x_0+x_1)\cdot (1+{ϵ}_0)\\ & =(x_0+x_1)+(x_0+x_1)\cdot {ϵ}_0\end{array}$$

$$\begin{array}{rl}{S}_1& =(S_0+x_2)\cdot (1+{ϵ}_1)\\ & =((x_0+x_1+x_2)+(x_0+x_1)\cdot {ϵ}_0)(1+{ϵ}_1)\\ & =(x_0+x_1+x_2)+(x_0+x_1)\cdot {ϵ}_0+(x_0+x_1+x_2){ϵ}_1+(x_0+x_1)\cdot {ϵ}_0\cdot {ϵ}_1\\ & =(x_0+x_1+x_2)+(x_0+x_1)\cdot {ϵ}_0+(x_0+x_1+x_2){ϵ}_1\end{array}$$

Hence,

$$\begin{array}{rl}{S}_{n-2}-S& =x_0\cdot ({ϵ}_0+\dots +{ϵ}_{n-2})+x_1\cdot ({ϵ}_0+\dots +{ϵ}_{n-2})\\ & +x_2\cdot ({ϵ}_1+\dots +{ϵ}_{n-2})+\dots +x(n-1)\cdot {ϵ}_{n-2}\\ & <|x_0|\cdot ({ϵ}_0+\dots +{ϵ}_{n-2})+|x_1|\cdot A+\dots +|x_{n-1}|\cdot A\\ & <|x_0|\cdot ((n-1){ϵ}_m)+|x_1|\cdot B+\dots +|x_{n-1}|\cdot B\\ & =(n-1)\cdot {ϵ}_m\cdot \sum _{i=0}^{n-1}|x_i|\end{array}$$