Prove the ratio of convergence implies $\frac{U_{h}(f)-U_{h/b^2}(f)}{U_{h/b^2}(f)-U_{h/b^4}(f)} \approx b^k$

40 Views Asked by At

I need to prove the following statement:

Given a $U_h(f)$ which is an approximation with a composite quadrature of an integral $I(f)$, with $h$ being the length of the subintervals and a ratio of convergence $O(h^k)$. Then

$$\frac{U_{h}(f)-U_{h/b^2}(f)}{U_{h/b^2}(f)-U_{h/b^4}(f)} \approx b^k$$

is true.

I'm pretty lost here. I don't know how what kind of information does $O(h^k)$ gives to help prove the statement. Or if I would decompose the left hand side to get to the right hand side. Thanks for the help.

1

There are 1 best solutions below

1
On

In reality, you do not have enough information to make any progress towards your goal. You need a stronger assumption. Specifically, there must exists an asymptotic error expansion of the form $$I(f) - U_h(f) = \alpha h^k + O(h^q), \quad h \rightarrow 0, \quad h > 0$$ where $\alpha \not = 0$ is a constant independent of $h$ and $0 < k < q$. Now let $\rho>0$. By replacing $h$ with $\rho h$ we have $$I - U_{\rho h} = \alpha \rho^k h^k + O(h^q),\quad h \rightarrow 0, \quad h > 0.$$ It follows that $$U_h - U_{\rho h} = \alpha(\rho^k - 1)h^k + O(h^q), \quad h\rightarrow 0, \quad h > 0.$$ By once again replacing $h$ with $\rho h$ we have $$ U_{\rho h} - U_{\rho^2h} = \alpha (\rho^k-1) \rho^k h^k + O(h^q), \quad h\rightarrow 0, \quad h > 0.$$ It follows that $$ \frac{U_h - U_{\rho h}}{U_{\rho h} - U_{\rho^2h}} = \frac{\alpha(\rho^k - 1)h^k + O(h^q)}{\alpha (\rho^k-1) \rho^k h^k + O(h^q)} = \frac{1 + O(h^{q-k})}{\rho^k + O(h^{q-k})} \approx \rho^{-k}$$ A common choice of $\rho$ is $\rho = \frac{1}{2}$ which corresponds to halving the length of each subinterval. This allows us to recycle the function values that have already been computed. In this case $$ \frac{U_h - U_{h/2}}{U_{h/2} - U_{h/4}} \approx 2^k.$$ Your left-hand side corresponds to the choice of $\rho = b^{-2}$ and your right-hand side should have been $b^{2k}$ rather than $b^k$ but this is not a serious issue.


In the case of the trapezoidal rule and a sufficiently smooth integrand the existence of a suitable asymptotic error expansion follows from the Euler-Maclaurian summation formula and $$(k,q) = (2,4).$$ If the integrand is not smooth, then you may still have an asymptotic error expansion, but the exponents $k$ and $q$ are not necessarily integers. Be mindful of the fact that you will never see the numbers $U_h$, only the computed value $\hat{U}_h$ which is affected by rounding errors and subtractive cancellation is an issue when computing $U_h - U_{\rho h}$ for small values of $h$. It is possible to say substantially more about the behavior of your fraction and the impact of rounding errors, but that is is outside the scope of your question.