2
$\begingroup$

The $n$-th Gauss-Laguerre quadrature scheme aims to approximate integral of exponentially decreassing function over $[0 ; \infty[$ by a finite sum, according to: $$ \int _0 ^{+ \infty} e^{-x} f(x) \ dx \approx \sum _{i = 1} ^n \omega_i f(x_i) \ , $$ where $x_1$, $\cdots$, $x_n$ are the roots of the $n$-th Laguerre polynomial $L_n$ and the weights $\omega_1$, $\cdots$, $\omega_n$ are chosen according to $\omega_i = \dfrac{1}{x_i \big ( L_n'(x_i) \big )^2}$, $1 \leq i \leq n$.

$\bullet$ Let us denotes by $E_n(f)$ the error of the quadrature: $E_n(f) = \displaystyle { \int _0 ^{+ \infty} e^{-x} f(x) \ dx } - \displaystyle { \sum _{i = 1} ^n \omega_i f(x_i) } $. The general estimation of the error of gaussian quadratures, specialized in Gauss-Laguerre scheme, is the following:

For all $n \in \mathbb{N}$, there exists $\xi \in ]0 ; \infty[$ such that $E_n(f) = \dfrac{n!^2}{(2n)!} f^{(2n)}(\xi)$.

Unfortunately, this is unusable in many case, since we know nothing on this $\xi$. Therefore, we shall consider functions $f$ with derivatives satisfyng $||f^{2n}||_{\infty, \mathbb{R}^+} = \mathcal{o}\left ( \dfrac{1}{n!^2} \right )$... This is a bit restrictive...

$\bullet$ We also know the Uspensky theorem:

$E_n(f) \underset{n \longrightarrow + \infty}{\longrightarrow} 0$ for functions $f$ satisfying $|f(x)| \leq c \dfrac{e^x}{x^{1 + \rho}}$ for large $x >> 1$, where 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)

But, this do not gives explicit result on the convergence speed to $0$, nor a upper bound of $E_n(f)$ going to $0$.

$\bullet$ So, what is the most precise result about the error term? In particular, how can we know which degree $n$ should we use to find a numerical approximation of $\displaystyle { \int _0 ^{+ \infty} e^{-x + \sqrt{x}} \ dx}$.

$\endgroup$
2
  • $\begingroup$ Maybe this helps ? $\endgroup$
    – user111
    Mar 15, 2021 at 17:49
  • $\begingroup$ Unfortunately, not really... I was aware of the result of Mastroianni and Monegato, but I'm not able to use it... Especially to compute the infinite norm $E_n(f, \omega)$... $\endgroup$ Mar 16, 2021 at 14:51

1 Answer 1

2
$\begingroup$

The function $f(x)=e^{-x+\sqrt{x}}$ belongs to the space $C_{0}^{3}[0,\infty)$ defined, for $q\geq p\geq0$, by $$C _ { p } ^ { q } [ 0,\infty ) : = \{ f \in C ^ { p } [ 0,\infty ) \cap C ^ { q } ( 0,\infty ),~x ^ { i } f ^ { ( p + i ) } ( x ) \in C [ 0,\infty ),~i=1,\ldots,q-p\}.$$ According to the result given in [1], the error rate behaves like \begin{equation}\label{rate-quad} \mathcal{O}(n^{-1})E_{n-1}(\Phi^{(3)}(x),e^{-x/2})=\mathcal{O}(n^{-1})E_{n-1}(\Phi^{(3)}(2x),e^{-x}), \end{equation} where $ \Phi ( x ) : = x ^ { 3} f ( x )=x^{3}e^{-x+\sqrt{x}}$, and $E _ { n } ( f ; w )$ is the rate of weighted polynomial approximation, $$E _ { n } ( f ; w ) : = \inf_ { p _ { n } } \| w (f - p _ { n })\| _ { \infty ,[ 0,\infty ) }.$$ Here, $\Phi^{(3)}(x)=P_{6}(\sqrt{x})e^{-x+\sqrt{x}}$, where $P_{6}$ is a polynomial of degree $6$.

Next, the rate of approximation $E_{n}(f,w)$, $w(x)=e^{-x}$, is known from [2] p.112, available here, namely \begin{equation}\label{rate-En} E_{n}(f,e^{-x})\leq C\frac{\sqrt{a_{n}}}{n}\left\|\sqrt{x}f'(x)e^{-x}\right\|_{\infty}, \end{equation} for $f\in W_{1}^{\infty}(e^{-x})$, where $a_{n}\sim n$ is the so-called Mhaskar-Rakhmanov-Saff number, and, for $r\geq1$, $W_{r}^{\infty}(e^{-x})$ is the Sobolev-type space, $$ W_{r}^{\infty}(e^{-x})=\left\{f \in L_{e^{-x}}^{\infty}, f^{(r-1)}\text{ abs. continuous on }(0,\infty)\text { and }\left\|f^{(r)}(x) x^{r/2} e^{-x}\right\|_{\infty}<\infty\right\}, $$ $$ L_{e^{-x}}^{\infty}=\{f\in C((0,\infty)),~\lim_{x\to0,~x\to\infty} f(x)e^{-x}=0\}. $$ Note that the function $f(x)=e^{-2x+\sqrt{2x}}$ belongs to $W_{1}^{\infty}(e^{-x})$, but not to any of the more regular spaces $W_{r}^{\infty}(e^{-x})$, $r\geq2$ (for which the rate of approximation is improved to $(\sqrt{a_{n}}/n)^{r}$).

Finally, taking into account the two previous estimates, we get $\mathcal{O}(n^{-3/2})$ for the error rate of the Gauss-Laguerre quadrature for $f(x)=e^{-x+\sqrt{x}}$.

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

[2] G. Mastroianni, J. Szabados, Polynomial approximation on the real semiaxis with generalized Laguerre weights. Stud. Univ. Babes-Bolyai Math. 52 (2007), nr.4, 105--128.

$\endgroup$
3
  • $\begingroup$ Thanks you for your developped answer. Just for completness, the n-th Mhaskar-Rakhmanov-Saff number $a_n$ associated to the weight $w(x)=e^{-Q(x)}$ is the positive root of the equation $$n=\dfrac{2}{\pi} \displaystyle{\int_0^1\dfrac{a_n tQ'(a_n t)}{\sqrt{1-t^2}}dt}$$ If $Q(x) = |x|^\alpha$, $a_n = \left ( 2^{\alpha - 2} \dfrac{\Gamma(\frac{\alpha}{2})^2}{\Gamma(\alpha)} \right )^{\frac{1}{\alpha}} n^{\frac{1}{\alpha}}$. Here, $\alpha = 1$ and $a_n = \dfrac{\pi}{2}n$. See DS Lubinsky, A Survey of Weighted Approximation for Exponential Weights (arxiv.org/abs/math/0701099), p.11 $\endgroup$ May 11, 2021 at 13:43
  • $\begingroup$ Is the underlying constant of the $\mathcal{O}(n^{-1})$ obtain by applying the result of [1] can be explicitely computed or estimated ? Is the constant C used in Proposition 4.1 of [2] can be explicitely computed or estimated? $\endgroup$ May 11, 2021 at 13:53
  • $\begingroup$ Is the space $L^{\infty}_{e^{-x}}$ well-defined? Shoudn't it be $$L^{\infty}_{e^{-x}} = \{ f \in \mathcal{C}((0,∞)), \underset{x \longrightarrow + \infty}{\text{lim}} f(x) \cdot e^{-x} = 0\} \ ?$$ Otherwise, the fonction $x \longmapsto \Phi^{(3)}(2x)$ do not belong to $W_1^{\infty}(e^{-x})$, according to $\Phi^{(3)}(0) = 6$. $$ $$ I suspect that your definition is actually a valid one for generalized Laguerre weight (see [2], p. 109) $\endgroup$ May 12, 2021 at 7:48

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.