3
$\begingroup$

The $n$th Gauss-Laguerre quadrature aims to approximate integral $$\int_{\mathbb{R}_+} f(x) \exp(-x)$$ by the sum $$\sum_{i=1}^n f(x_i) w_i$$ where $x_1,...,x_n$ are the roots of the $n$th Laguerre polynomial $L_n$, and the weights $w_1,...,w_n$ are chosen according to $w_i=\frac{1}{x_i (L_n'(x_i))^2}$.

The intuition is that if $f$ is polynomial of degree at most $2n-1$ then the approximation is exact; In general, the approximation error (as in any Gauss quadrature) is known to be given by $$E_n(f) = \frac{(n!)^2}{(2n)!} f^{(2n)}(\xi)$$ for some $\xi \in (0,\infty)$. My question is simply the following:

Is there function $f$ that is smooth in $(0,\infty)$, such that the the approximation error $E_n(f)$ does not go to zero as $n\to \infty$?

My feeling is that if $f$ derivative grows very fast, or, say, $f$ has infinite derivative at the end point 0, e.g., $f(x)=x \log x$, then maybe $\xi$ is very close to zero and the error will not improve as $n$ grows, but I am no expert in this field. Any comment will be welcomed.

$\endgroup$
6
  • $\begingroup$ You need some extra condition to rule out placing bumps at the largest roots of the Laguerre polynomials. $\endgroup$ Mar 24, 2014 at 6:37
  • $\begingroup$ Douglas, I meant for a fixed function $f$ and let $n$ goes to infinity. So $f$ cannot depend on $n$. $\endgroup$
    – gondolier
    Mar 24, 2014 at 7:24
  • 1
    $\begingroup$ The Runge function $x\mapsto \frac{1}{1+25x^2}$ is the standard example of hard-to-approximate function for the Gauss-Legendre quadrature in the interval $[-1,1]$ (which is exact for integrating polynomials of degree up to $2n-1$). My first attempt would be trying a change of variable of some form to take that interval into $[0,\infty]$ and Legendre into Laguerre, but I am not an expert in quadrature myself. $\endgroup$ Mar 24, 2014 at 7:57
  • 1
    $\begingroup$ I didn't suggest that $f$ should depend on $n$. I meant that you can easily construct $f$ by placing infinitely many bumps, one at the largest root of the $n$th polynomial, so that the $n$th degree Gauss-Laguerre quadrature estimate is off by $1$, for example. Since the largest roots have no limit point you can put a smooth bump there to adjust the $n$th degree estimate arbitrarily without affecting lower estimates. $\endgroup$ Mar 24, 2014 at 16:10
  • 1
    $\begingroup$ @FedericoPoloni, The Runge function seems to be an example for equidistant case instead of Gauss quadrature. $\endgroup$
    – gondolier
    Mar 24, 2014 at 21:11

2 Answers 2

2
$\begingroup$

This is to answer the question that the OP asked in the comment, about convergence of the Gauss-Laguerre quadrature when $f(x)=x\log x$.

Uspensky has shown that the Gauss-Laguerre quadrature rule for computing integrals of the form $$\int_{0}^{\infty}x^{\alpha}e^{-x}f(x)dx$$ converges, as $n\to\infty$, if $f$ satisfies the following growth at infinity, $$|f(x)|\leq c x^{-\alpha-1-\rho}e^{x},\qquad|x|\text{ large},$$ for some $\rho>0$, see

J. V. Uspenksy, On the convergence of quadrature formulas related to an infinite interval, Trans. Amer. Math. Soc. 30 (1928), 542-559.

Thus, the G-L quadrature indeed converges in the case of $f(x)=x\log x$. Actually, even the rate of convergence can be estimated.

The case where $f(x)$ is not smooth at the origin, as for $x\log x$, was studied in

Mastroianni G., Monegato G., Convergence of product integration rules over $(0,\infty)$ for functions with weak singularities at the origin. Math. Comp. 64, (1995), 237-249.

The following is proved. For $q\geq p\geq0$, let $$C _ { p } ^ { q } [ 0,\infty ) : = \{ f \in C ^ { p } [ 0,\infty ) \cap C ^ { q } ( 0,\infty ),~x ^ { i } f ^ { ( p + i ) } ( x ) \in C [ 0,\infty ),~i=1,\ldots,q-p\}.$$ Then, the remainder term $R_{n}(f)$ of the Gauss-Laguerre quadrature satisfies $$ |R_{n}(f)|=\begin{cases} \mathcal O(n^{-q/2}) E_{n-p-1}(\Phi^{(q)},e^{-x/2}) & \text{if }q\leq 2p+1, \\ \mathcal O(n^{-p-1}\log n) E_{n-p-1}(\Phi^{(q)},e^{-x/2}) & \text{if }q= 2p+2, \\ \mathcal O(n^{-p-1}) E_{n-p-1}(\Phi^{(q)},e^{-x/2}) & \text{if }q\geq 2p+3, \end{cases} $$ where $ \Phi ( x ) : = x ^ { q - p } f ( x ) \in C ^ { q } [ 0,\infty ) $, and $ E _ { n } ( f ; w ) : = \operatorname{inf} _ { p _ { n } } \| w \left[ f - p _ { n } \right] \| _ { \infty ,[ 0,\infty ) } $.

Hence, for the function $f(x)=x\log x\in C_{0}^{3}[0,\infty)$, $$\Phi(x)=x^{4}\log x,\qquad\Phi^{(3)}(x)=24x\log x+26x,$$ and thus $$|R_{n}(f)|=\mathcal O(n^{-1})E_{n-1}(x\log x, e^{-x/2}).$$ Since, it is also known that $E_{n-1}(x\log x, e^{-x/2})=\mathcal O(n^{-1})$ for $n$ large, one finally gets for the function $f(x)=x\log x$, $$|R_{n}(f)|=\mathcal O(n^{-2}), \qquad n\text{ large}.$$

$\endgroup$
1
$\begingroup$

I think you misunderstand the error term in gauss-laguerre quadrature !

that is $$E=\frac{2^{2n+1}{(n!)}^4}{(2n+1)(2n!)^3}{f^{(2n)}},$$

see : http://mathworld.wolfram.com/Legendre-GaussQuadrature.html.

Hence the condition is:

$$\frac{2^{2n+1}{(n!)}^4}{(2n+1)(2n!)^3}\ge\frac{1}{\epsilon}$$

where $\epsilon=f^{(2n+1)}=C_{2n}$ is the first coefficient ! So, \begin{align*} \frac{2^{2n+1}{(n!)}^4}{(2n+1)(2n!)^3} & \ge\frac{1}{f(x)-f(a)-\frac{f^{(1)}{}(a)}{1!}(x-a)-\frac{f^{(2)}{(a)}}{2!}(x-a)^{2}-......}\frac{(x-a)^{2n+1}}{(2n+1)!} \\ & =\frac{2}{f(x)-C_{1}-\frac{C_{1}+C_{2}}{1!}x-\frac{C_{1}+C_{2}+C_{3}}{2!}x^{2}-......}\frac{x^{2n+1}}{(2n+1)!}=C_{2n}, \end{align*} and thus $$\frac{2}{\sum{C_{i}}-C_{1}-\frac{C_{1}+C_{2}}{1!}-\frac{C_{1}+C_{2}+C_{3}}{2!}-......}\frac{1}{(2n+1)!}\le\frac{2^{2n+1}{(n!)}^4}{(2n+1)(2n!)^3},$$

with $C_{2n}\to0$.

$\endgroup$
1
  • $\begingroup$ This is the error for the Gauss-Legendre quadrature, not the Gauss-Laguerre one. $\endgroup$ Jun 27, 2023 at 14:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.