Newton-Raphson termination criteria for large problems

243 Views Asked by At

When using Newton-Raphson to solve a system of equations of the form: $$\mathbf{r}(\mathbf{x})=\mathbf{0}, \quad \mathbf{r}=[r_1, r_2,...,r_N], \quad \mathbf{x}=[x_1, x_2,...,x_N]$$ The termination criterion is usually based on the euclidean norm of $\mathbf{r}$: $$ \sqrt{r_1^2+r_2^2+...+r_N^2} \le tolerance $$ As far as I understand, this criterion guarantees that the largest error for a single equation, $r_i$, will always be smaller than the desired tolerance. However, I feel like such a criterion is 'unfair' when the number of equations increases. An extreme example would be that of a system where $N=1$ (a single equation) vs a system where $N \gg 1$.

For $N=1$ we can assume that the value of $r_1$ at the point of termination will be close to the set tolerance.

For $N \gg 1$ my intuition is that each $r_i$ will be significantly smaller than the tolerance, since they are all added up to produce a value close to the tolerance. This means that more iterations are needed, although the set tolerance stays the same.

Then, my question would be: Is there a termination criterion that will provide solutions with the same order of accuracy, regardless of the number of the equations?

1

There are 1 best solutions below

0
On

Yes. Replace the Euclidian norm with the infinity norm and terminate when $$\|\mathbf{r}\|_\infty \leq \tau$$ where $\tau$ is the tolerance.