Abstract
We consider an interpolatory quadrature formula having as nodes the zeros of the nth degree Chebyshev polynomial of the second kind, on which the Fejér formula of the second kind is based, and the additional points \(\pm \tau _{c}=\pm \cos \frac{\pi }{2(n+1)}\). The new formula is shown to have positive weights given by explicit formulae. Furthermore, we determine the precise degree of exactness, and we obtain optimal error bounds for this formula either by Peano kernel methods or by Hilbert space techniques for analytic functions and \(1\le n\le 40\). In addition, the convergence of the quadrature formula is shown not only for Riemann integrable functions on \([-1,1]\), but also for functions having monotonic singularities at \(\pm \)1. The new formula has essentially the same rate of convergence as, and it is therefore an alternative to, the well-known Clenshaw-Curtis formula.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
One of the most widely used quadrature formulae is the Clenshaw-Curtis formula
where
are the zeros of the nth degree Chebyshev polynomial \(U_{n}\) of the second kind. Formula (1.1) has all weights positive and expressed by explicit formulae, while its precise degree of exactness is \(d^{*}=n+1\) if n is even and \(d^{*}=n+2\) if n is odd, i.e., \(R_{n}^{*}(f)=0\) for all \(f\in \mathbb {P}_{d^{*}}\) (the space of polynomials with real coefficients and degree at most \(d^{*}\)). Moreover, in view of its performance in practice, the Clenshaw-Curtis formula is sometimes compared favorably to the well-known Gauss formula (see [17]).
All of the above made Hasegawa and Sugiura to try improving the error behavior of the Clenshaw-Curtis formula, by adding to (1.1) the nodes
which lie in the intervals \((\tau _{1},1)\) and \((-1,\tau _{n})\), respectively. That way, they obtained the so-called corrected Clenshaw-Curtis formula
The new formula has all weights, except for \(\bar{w}_{c}^{*}\) when \(n\ge 2\), positive and explicitly expressed, while its degree of exactness is \(n+3\) if n is even and \(n+4\) if n is odd. In addition, the convergence rate of formula (1.4) is better than that of any formula in the Clenshaw-Curtis family (see [9], in particular, Theorem 2 and Remark 1; a detailed description of all interpolatory quadrature formulae with Chebyshev abscissae of any of the four kinds is given in [13]).
The corresponding open-type Clenshaw-Curtis formula is the so-called Fejér formula of the second kind or Filippi formula
having all weights positive and explicitly expressed and degree of exactness \(n-1\) if n is even and n if n is odd. This formula is known to share common properties with the Clenshaw-Curtis formula. An important such property is that formula (1.5) forms a nested set of quadrature formulae, i.e., the nodes of the n-point formula are among those of the \((2n+1)\)-point formula, and the same property is enjoyed by formulae (1.1) and (1.4). This makes all these formulae appropriate for adaptive or cubature integration schemes.
Motivated by the work of Hasegawa and Sugiura in [9], we introduce a corrected Fejér formula of the second kind, by adding to formula (1.5) the nodes \(\pm \tau _{c}\) in (1.3), thus obtaining
The new formula is shown to have all weights positive and given by explicit formulae, and precise degree of exactness \(n+1\) if n is even and \(n+2\) if n is odd. This, together with the convergence of the formula for Riemann integrable functions on \([-1,1]\), is the subject of Sect. 2. Section 3 is devoted to the error term of the formula. First, we obtain optimal error bounds by Peano kernel methods, thus concluding that formula (1.6) has essentially the same rate of convergence as the Clenshaw-Curtis formula. Then, using Hilbert space techniques, we compute the norm of the error functional, which leads to error bounds for analytic functions when \(1\le n\le 40\). In Sect. 4, we prove the convergence of formula (1.6) for functions having a monotonic singularity at one or both endpoints of \([-1,1]\). This property, also satisfied by the Fejér formula of the second kind (cf. [5]), is an advantage of formula (1.6) over the Clenshaw-Curtis formula and its corrected version (1.4), both of which cannot even be applied on functions with singularities at \(\pm \)1. In addition, as expected, formula (1.6) retains the nested quadrature formulae property satisfied by formulae (1.1), (1.4) and (1.5). All this together with its rate of convergence make formula (1.6) an alternative to the Clenshaw-Curtis formula. The paper concludes in Sect. 5, with some numerical examples.
2 The quadrature formula
We begin by recalling explicit formulae for the weights of formula (1.5),
or
where \([\,\cdot \,]\) denotes the integer part of a real number (cf. [13, Eqs. (2.8)–(2.10) with \(i=2\)]).
We now turn to the study of formula (1.6). Let \(I(f)=\int _{-1}^{1}f(t)dt\) and \(\bar{Q}_{n}(f)=\bar{w}_{c}^{(+)}f(\tau _{c})+ \sum _{\nu =1}^{n}\bar{w}_{\nu }f(\tau _{\nu })\)+\(\bar{w}_{c}^{(-)}f(-\tau _{c})\).
Theorem 2.1
Consider the quadrature formula (1.6).
-
(a)
The weights \(\bar{w}_{\nu }\) and \(\bar{w}_{c}^{(+)},\ \bar{w}_{c}^{(-)}\) are given by
$$\begin{aligned} \bar{w}_{\nu }=w_{\nu }+\frac{{2\sin ^{2}\theta _{\nu }\cos 2[(n\!+\!1)/2]\theta _{\nu }}}{{(2[n/2]+1)(2[(n+1)/2]+1)\sin (\theta _{\nu }+\theta _{c})\sin (\theta _{\nu }-\theta _{c}})},\ \ \nu =1,2,\ldots , n, \end{aligned}$$(2.2)$$\begin{aligned} \bar{w}_{c}^{(+)}=\bar{w}_{c}^{(-)}=\left\{ \begin{array}{ll} \displaystyle \frac{\sin \theta _{c}}{n+1},&{}n\ \mathrm{even},\\ \displaystyle \frac{(n+1)\tan \theta _{c}}{n(n+2)}, &{}n\ \mathrm{odd}. \end{array}\right. \end{aligned}$$(2.3)In addition, the \(\bar{w}_{\nu },\ \nu =1,2,\ldots ,n\), are all positive.
-
(b)
The quadrature formula has precise degree of exactness \(\bar{d}=n+1\) if n is even and \(\bar{d}=n+2\) if n is odd.
-
(c)
There holds \(\lim _{n\rightarrow \infty }\bar{Q}_{n}(f)=I(f)\) for all functions f that are Riemann integrable on \([-1,1]\).
Proof
(a) As formula (1.6) is precise for polynomials of degree \(n+1\), setting \(f(t)=(t^{2}-\tau _{c}^{2})U_{n}(t)/(t-\tau _{\nu })\), we have
that is,
The nth degree Chebyshev polynomial of the second kind \(U_{n}\) can be represented by
and satisfies the three-term recurrence relation
Then, by means of (2.6) and
(cf. [11, Eq. (2.46)]), we compute
Also, using (2.5), we calculate \(U_{n}'(\cos \theta )\), and setting \(\theta =\theta _{\nu }\), we get
Finally, from (1.2)–(1.3), by the double-angle formula and the formula for the difference of cosines, we get
Now, inserting (2.8)–(2.10) into (2.4), we obtain, after an elementary computation, (2.2).
Furthermore, by symmetry, \(\bar{w}_{c}^{(+)}=\bar{w}_{c}^{(-)}\), hence, setting \(f(t)=(t+\tau _{c})U_{n}(t)\) in formula (1.6), we have
where, as in (2.8),
which inserted, together with (2.12), into (2.11), yields (2.3).
We now turn into proving the positivity of \(\bar{w}_{\nu },\ \nu =1,2,\ldots ,n\). First of all, by symmetry,
hence, we only need to prove the positivity of \(\bar{w}_{\nu }\) for \(\nu =1,2,\ldots ,[(n+1)/2]\). Furthermore, as \(w_{\nu }>0\) (cf. [13, Sect. 2.1]) and
by (2.2), \(\bar{w}_{\nu }>0\) for \(\nu \) even. It therefore remains to prove the positivity of \(\bar{w}_{\nu }\) for \(\nu \) odd. Let first n be even. Then, by the second equation in (2.1), in view of (2.14),
and, by virtue of
(proved by a partial fraction decomposition of the left-hand side), the formula for the product of sines and the fact that \(2\theta _{c}=\theta _{1}\) (cf. (1.2)–(1.3)), we get, after a simple computation,
If, on the other hand, n is odd, starting from the first equation in (2.1) and proceeding in a like manner, we obtain
thus concluding the proof.
(b) Let n be even. First of all, formula (1.6) has degree of exactness at least \(n+1\). Furthermore, by a repeated application of (2.6), and in view of (2.7), we compute
proving that formula (1.6) has precise degree of exactness \(n+1\).
Similarly, for n odd, the degree of exactness is at least \(n+2\), and as
it is precisely \(n+2\).
(c) This, by a well-known result of Fejér (cf. [4, Satz 1]), is an immediate consequence of the positivity of the weights \(\bar{w}_{c}^{(+)},\ \bar{w}_{\nu }, \ \nu =1,2,\ldots ,n\), and \(\bar{w}_{c}^{(-)}\). \(\square \)
3 The error term of the quadrature formula
Our error estimates for formula (1.6) are of two different types. Optimal error bounds, by Peano kernel methods, for functions that are sufficiently smooth; and error bounds, by Hilbert space techniques, for analytic functions.
3.1 Peano kernel error bounds
Given that formula (1.6) has degree of exactness \(\bar{d}\), for \(f\in C^{\bar{d}+1}[-1,1]\), we have
where \(\bar{K}_{\bar{d}}\) is the \(\bar{d}\)th Peano kernel. From (3.1), we immediately derive
If, in addition, \(\bar{K}_{\bar{d}}\) does not change sign on \([-1,1]\), formula (1.6) is called definite; in particular, positive definite if \(\bar{K}_{\bar{d}}\ge 0\), and negative definite if \(\bar{K}_{\bar{d}}\le 0\). In this case, (3.1), by the Mean Value Theorem for integrals, gives
(cf. [3, Sect. 4.3]).
The derivation of the error bounds will be based on the following lemma, which precedes our results.
Lemma 3.1
([10, Lemma B]) Let \(g\in C^{m+s}[a,b],\ m\ge 1,\ s\ge 0\), have the zeros \(t_{\nu },\ 1\le \nu \le m+s\). For a \(k,\ 1\le k\le m\), assume that the polynomial \(q_{k}(t)=\Pi _{i=1}^{k}(t-t_{i})\) has only simple zeros. Then there exist functions \(r_{i}\in C^{k+s-1}[a,b],\ 1\le i\le k\), such that
Each \(r_{i}\) has \(k+s-1\) zeros, especially, the \(t_{\nu },\ m+1\le \nu \le m+s\), are zeros of \(r_{i}\). In addition, there exist \(\xi _{i}=\xi _{i}(t)\in [a,b], \ 1\le i\le k\), such that
Theorem 3.2
Consider the quadrature formula (1.6). There holds, for n even and \(f\in C^{n+2}[-1,1]\),
and, for \(n(odd)\ge 3\) and \(f\in C^{n+3}[-1,1]\),
On the other hand, if \(n=1\), the quadrature formula is positive definite, and
Proof
Let first n be even. As formula (1.6) is interpolatory, having degree of exactness \(n+1\), there holds
where \(\bar{r}_{n}(f;\cdot )\) is the error of the interpolation based on the \(n+2\) points \(\tau _{\nu },\ \nu =1,2,\ldots ,n\), and \(\pm \tau _{c}\). Assuming that \(f\in C^{n+2}[-1,1]\), the same is true for \(\bar{r}_{n}(f;\cdot )\). Since, in addition, \(\bar{r}_{n}(f;\tau _{\nu })=0, \ \nu =1,2,\ldots ,n\), and \(\bar{r}_{n}(f;\pm \tau _{c})=0\), we can apply Lemma 3.1 with \(g(\cdot )=\bar{r}_{n}(f;\cdot ),\ m=n,\ s=2\) and \([a,b]=[-1,1]\). Setting \(k=2\) and \(q_{2}(t)=(t-\tau _{1})(t-\tau _{n})=(t-\tau _{1})(t+\tau _{1})=t^{2}-\tau _{1}^{2}\) (cf. (2.13)), (3.3) gives
where \(r_{i}\in C^{3}[-1,1],\ i=1,2\), each \(r_{i}\) has three zeros, in particular, \(r_{i}(\pm \tau _{c})=0\), and there exist \(\xi _{i}=\xi _{i}(t)\in [-1,1],\ i=1,2\), such that
(cf. (3.4)). Now, let the functions \(h_{i},\ i=1,2,3\), be defined by
and
whence
Then, inserting (3.7) into (3.6), and applying, in view of (3.9) and (3.11)–(3.12), integration by parts, we get
from which, there follows
Now, from (3.9) and (3.10), using \(T'_{n+1}=(n+1)U_{n}\), where \(T_{n+1}\) is the \((n+1)\)th degree Chebyshev polynomial of the first kind (cf. [11, Eq. (2.48)]) and
(cf. [11, Eq. (2.43)]), we compute
hence, we find
Furthermore, as \(r_{i},\ i=1,2\), has three zeros, among them \(\pm \tau _{c}\), by Rolle’s Theorem, \(r'_{i}\) and \(r''_{i},\ i=1,2\), have two and one zeros, respectively, and let \(t'_{i},\ i=1,2\), be one of the zeros of \(r'_{i}\) and \(t''_{i},\ i=1,2\), be the zero of \(r''_{i}\). Then, by the Mean Value Theorem, applied first to \(r_{i},\ i=1,2\), on \([\tau _{c},1]\),
then to \(r'_{i},\ i=1,2\), between \(t'_{i}\) and \(\zeta _{i}\),
and finally to \(r''_{i},\ i=1,2\), between \(t''_{i}\) and \(\zeta '_{i}\),
which combined, together with (1.3) and (3.8), give, in view of the double-angle formula for cosines,
where the estimate for \(r_{i}(-1)\) is derived by the same steps. In a like manner,
Finally, from (1.2), by virtue of \(\cos \theta \ge 1-2\theta /\pi ,\ 0\le \theta \le \pi /2\), we get
Now, inserting (3.8) and (3.15)–(3.22) into (3.13), taking into account that n is even, we obtain, after an elementary computation, \((3.5_{e})\).
We next turn to the case of \(n(\mathrm{odd})\ge 3\). As in this case formula (1.6) has degree of exactness \(n+2\), we consider the even part of the interpolation error \(\bar{r}_{n}(f;\cdot )\) defined by
First of all, a simple change of variables shows that
hence,
(cf. (3.6)). As \(\bar{r}_{n}(f;0)=\bar{r}_{n}(f;\tau _{(n+1)/2})=0\), we have \(\bar{r}_{n,e}(f;0)=0\); and as \(\bar{r}_{n,e}(f;\cdot )\) is an even function, there holds \(\bar{r}_{n,e}^{\prime }(f;0)=0\). Consequently, \(\bar{r}_{n,e}(f;\cdot )\) has \(n+3\) zeros, the \(\tau _{\nu },\ \nu =1,2,\ldots ,n\), and the \(\pm \tau _{c}\), where \(\tau _{(n+1)/2}=0\) is a double zero. Therefore, assuming that \(f\in C^{n+3}[-1,1]\), the same is true for \(\bar{r}_{n,e}(f;\cdot )\), and we can apply Lemma 3.1 with \(g(\cdot )=\bar{r}_{n,e}(f;\cdot ),\ m=n,\ s=3\) and \([a,b]=[-1,1]\). Choosing the same k and \(q_{k}\) as in the case of n even, \(\bar{r}_{n,e}(f;\cdot )\) has the representation (3.7), except that here each \(r_{i}\in C^{4}[-1,1]\), it has four zeros, among which \(\pm \tau _{c}\), and there exist \(\xi _{i}=\xi _{i}(t)\in [-1,1]\) such that
Then, we proceed as in the case of n even, by defining the \(h_{i},\ i=1,2,3\), in (3.9)–(3.12) and
We get
Now, from (3.24), in view of (3.14), an elaborate computation gives
hence,
Furthermore, as in the case of n even,
Now, inserting (3.15)–(3.18), (3.22), (3.23) and (3.26)–(3.29) into (3.25), taking into account that n is odd, we obtain, after an elementary computation, \((3.5_{o})\).
For \(n=1\), formula (1.6) has the form
with degree of exactness 3 and 3rd Peano kernel
As \(\bar{K}_{3}(t)\ge 0,\ -1\le t\le 1\), the formula is positive definite, hence,
(cf. (3.2)), where, from (3.31), in view of (3.30), we get
thus obtaining (3.5\(_1\)). \(\square \)
Remark 3.1
We have, in view of (2.15\(_{e}\))–(2.15\(_{o}\)), for n even,
and for n odd,
which, compared to (\(3.5_{ e}\)) and (\(3.5_{ o}\)), respectively, show that our bounds are optimal.
Furthermore, as in both (\(3.5_{ e}\)) and (\(3.5_{ o}\)), the quantity in the braces is of order \(O(n^{-3})\), the rate of convergence of formula (1.6) is the same as that of the Clenshaw-Curtis formula (cf. [2, Theorem 2]), which is also confirmed numerically in Example 5.1.
Remark 3.2
Given that formula (1.5) is definite (cf. [1]), one could ask the same question for formula (1.6), in which case we could obtain results analogous to (3.5\(_{1}\)). A few calculations, for small values of n, indicate that the answer to this question could be affirmative, although further investigations would be needed. Note, however, that, even though proving the definiteness of formula (1.6) would be quite cumbersome, requiring a substantial effort (cf. [1]), it will not essentially improve the results of Theorem 3.2.
3.2 Hilbert space error bounds
Another estimate for the error term of formula (1.6) can be obtained by a Hilbert space technique proposed by Hämmerlin (cf. [8]). Assuming that f is a single-valued holomorphic function in the disk \(C_{r}=\{z\in \mathbb {C}:|z|<r\},\ r>1\), then it can be written as
Define
which is a seminorm in the space
Then, it can easily be shown that \(\bar{R}_{n}(\cdot )\) is a continuous linear functional in \((X_{r},|\cdot |_{r})\), and its norm is given by
while, in case that
one can derive the representation
(cf. [15, Sect. 2]). Consequently, for \(f\in X_{R}\),
and optimizing the right-hand side of (3.35) as a function of r, we get
Another estimate can be obtained if \(|f|_{r}\) is estimated by \(\max _{|z|=r}|f(z)|\), which exists at least for \(r<R\) (cf. [15, Eq. (2.9)]), giving
The latter can also be derived by a contour integration technique on circular contours (cf. [6]).
Therefore, in order to compute the norm of \(\bar{R}_{n}\) by (3.34), we first need to examine the validity of (3.33). First of all, by Theorem 2.1(b),
Then, we can prove
Lemma 3.3
The error term of the quadrature formula (1.6) satisfies
where \(\bar{k}_{n}\ge [(n+1)/2]+1\) is a constant.
Proof
Setting \(f(t)=t^{2l}\) in formula (1.6), we have
and, as \(\lim _{l\rightarrow \infty }(2l+1)\tau _{c}^{2l}=0\), our assertion follows. \(\square \)
From the last part in (3.39), we can find the constant \(\bar{k}_{n}\). This was done for \(1\le n\le 40\), and the values are given in Table 1. We have also examined numerically and found that \(\bar{R}_{n}(t^{2l})>0\) for all \([(n+1)/2]+1\le l\le \bar{k}_{n}-1\) and \(1\le n\le 40\). Putting everything together, we conclude
Furthermore, from (1.6), there follows, by symmetry, that
which, combined with (3.38) and (3.40), gives
Interestingly enough, by (3.32e)–(3.32o),
i.e., \(\bar{R}_{n}(t^{2l})>0\) is theoretically confirmed when \(l=[(n+1)/2]+1\) for all \(n\ge 1\). This together with our numerical findings suggest the following
Conjecture 3.4
The error term of the quadrature formula (1.6) satisfies
We are now in a position to compute \(\Vert \bar{R}_{n}\Vert \).
Theorem 3.5
Consider the quadrature formula (1.6). For \(1\le n \le 40\), we have
Proof
Let \(1\le n \le 40\). Then, in view of (3.41) (cf. (3.33)), \(\Vert \bar{R}_{n}\Vert \) is given by (3.34). Writing
splitting the integral on the right-hand side in two, and using (2.6), we get
By
(cf. [14, Proposition 2.2(i), Eq. (2.9)]), and (2.7), we find
4 Convergence of the quadrature formula for functions with singularities
We show that formula (1.6) converges, not only for Riemann integrable functions on \([-1,1]\), but also for functions having monotonic singularities at \(\pm 1\).
Following the notation in [5], we denote by \(M[-1,1)\) the class of functions f that are continuous on the half-open interval \([-1,1)\), monotonic in some neighborhood of 1, and such that \(\lim _{x\rightarrow 1^{-}}\int _{-1}^{x}f(t)dt\) exists. The classes \(M(-1,1]\) and \(M(-1,1)\) are defined analogously, while M stands for the union of all three classes.
Let, in the quadrature formula (1.6), \(\overline{\tau }_{1}=\tau _{c},\ \overline{\tau }_{\nu }={\tau }_{\nu -1},\ \nu =2,3,\ldots ,n+1,\ \overline{\tau }_{n+2}=-\tau _{c}\), and, accordingly, \(\overline{w}_{1}=\bar{w}_{c}^{(+)},\ \overline{w}_{\nu }=\bar{w}_{\nu -1},\ \nu =2,3,\ldots ,n+1,\ \overline{w}_{n+2}=\bar{w}_{c}^{(-)}\). Furthermore, as in Sect. 2, \(I(f)=\int _{-1}^{1}f(t)dt\) and \(\bar{Q}_{n}(f)=\bar{w}_{c}^{(+)}f(\tau _{c})+ \sum _{\nu =1}^{n}\bar{w}_{\nu }\) \(f(\tau _{\nu })\)+\(\bar{w}_{c}^{(-)}f(-\tau _{c})=\sum _{\nu =1}^{n+2}\overline{w}_{\nu }f(\overline{\tau }_{\nu })\), while \(\overline{\tau }_{0}=1\). For \(f\in M[-1,1)\), we have
if the following two conditions are satisfied:
-
(i)
\(\lim _{n\rightarrow \infty }\bar{Q}_{n}(g)=I(g)\) for all \(g\in C[-1,1]\).
-
(ii)
There exist constants \(c>0,\ \delta >0\) such that \(|\overline{w}_{\nu }|\le c\,(\overline{\tau }_{\nu -1}-\overline{\tau }_{\nu })\) for all sufficiently large n and for all \(\nu \ge 1\) such that \(1-\delta \le \overline{\tau }_{\nu }\le 1\).
As formula (1.6) is symmetric (cf. (2.3) and (2.13)), conditions (i) and (ii) also imply (4.1) for all \(f\in M(-1,1]\) or \(f\in M(-1,1)\), thus, for all \(f\in M\) (see [16, Sect. 4, Lemma 4.1] ).
Our results are summarized in the following
Theorem 4.1
Consider the quadrature formula (1.6). Then (4.1) holds for all \(f\in M\).
Proof
By what was said previously, it suffices to satisfy conditions (i) and (ii).
The first has already been proved in Theorem 2.1(c).
Regarding condition (ii), we shall show that
for all \(n\ge 1\) and \(\nu =1,2,\ldots ,[(n+2)/2]\).
Let first n be even and \(\nu =2,3,\ldots ,n/2+1\). Then \(\overline{w}_{\nu }=\bar{w}_{\nu -1}\), with \(\bar{w}_{\nu -1}\) given by (2.1) and (2.2). Setting \(\nu -1\) in place of \(\nu \) in (2.1), we have, in view of the cosine series (cf. [7, Eq. 1.444.7]), a partial fraction decomposition in \(2\sum _{k=n/2}^{\infty }1/(4k^{2}-1)=1/(n-1)\) and \(\sin \theta \le \theta ,\ 0\le \theta \le \pi /2\),
hence,
Also, using \(2\theta /\pi \le \sin \theta \le \theta ,\ 0\le \theta \le \pi /2\), we get
which inserted, together with (4.3), into (2.2) with \(\nu -1\) in place of \(\nu \), gives
Moreover,
Now, combining (4.4) and (4.5), we get
hence,
On the other hand, for n even and \(\nu =1\), we have, in view of \(2\theta /\pi \le \sin \theta \le \theta ,\ 0\le \theta \le \pi /2\), from (2.3),
and
which, combined together, yield
Now, from (4.6) and (4.6\(_1\)), we finally obtain
for all n even and \(\nu =1,2,\ldots ,n/2+1\).
In a like manner, we show
for all n odd and \(\nu =1,2,\ldots ,(n+1)/2\).
Putting (\(4.7_{ e}\)) and (\(4.7_{ o}\)) together, we conclude (4.2). \(\square \)
5 Numerical examples
Our examples focus on comparing formula (1.6) with the Clenshaw-Curtis formula, on showing the efficiency of bounds (3.36)–(3.37), and on demonstrating the ability of formula (1.6) to integrate functions with monotonic singularities at one or both endpoints of \([-1,1]\).
Example 5.1
We approximate the integral \(\int _{-1}^{1}f(t)dt\) by means of formula (1.6) or the Clenshaw-Curtis formula (1.1), when f(t) is any one of the four functions \(e^{-t^{2}},\ 1/(1+16t^{2}),\ e^{-1/t^{2}}\) or \(|t|^{3}\), borrowed from [17], where they were used for comparing the Clenshaw-Curtis formula with the Gauss formula. The first function is entire, the second analytic, the third \(C^{\infty }\) and the fourth \(C^{2}\). The modulus of the actual error is given in Table 2. (Numbers in parentheses indicate decimal exponents.) All computations were performed on a SUN Ultra 5 computer in quad precision (machine precision \(1.93\times 10^{-34}\)). Whenever the actual error is close to machine precision, we enter instead “m.p.” (for machine precision).
Our numerical results confirm what was theoretically proved in Theorem 3.2 (cf. Remark 3.1), namely, that formulae (1.6) and (1.1) have the same rate of convergence. Indeed, the actual errors of both formulae for each of our test functions are very close or almost identical to each other.
Example 5.2
We want to approximate the integral
by means of formula (1.6).
The function \(f(z)=\frac{z^{2}}{4+z^{2}}=\sum _{k=0}^{\infty }(-1)^{k}\frac{z^{2k+2}}{2^{2k+2}}\) is holomorphic in \(C_{2}=\{z\in \mathbb {C}:|z|<2 \}\), hence, taking into account that formula (1.6) has degree of exactness \(2[(n+1)/2]+1\), we find
thus, \(f\in X_{2}\). Then, from (3.36),
with \(\Vert \bar{R}_{n}\Vert \) given by (3.42). As, in addition,
we have, from (3.37),
Our results are summarized in Table 3. The value of r, at which the infimum in each of bounds (5.2) and (5.4) was attained, is given in the column headed \(r_{opt}\) and placed immediately before the column of the corresponding bound. In the last column, we give the modulus of the actual error.
Bound (5.2) is quite reasonable, overestimating the actual error by no more than two orders of magnitude. Bound (5.4) is inferior to (5.2), particularly as n increases, because then \(r_{opt}\) approaches 2 and that way increases the value of \(\max _{|z|=r}|f(z)|\) in (5.3). Using bound (5.4) in order to estimate not the error, but the appropriate value of n to be used, yields an overestimation of n by just a few units.
Example 5.3
We want to approximate the integral
whose integrand has a monotonic singularity at 0. Previously, this example has been employed in [5, Sect. 4], [12, Sect. 5], and [16, Sect. 5], where the integral has been approximated, for various values of a, by means of interpolatory and product type formulae for Chebyshev weights based on the Chebyshev abscissae of any one of the four kinds. Here, we use formula (1.6), appropriately transformed onto the interval [0,1],
where
and where we set \(n-2\) in place of n in order to have an n-point formula. For comparison, we also compute integral (5.5) by means of the Fejér formula of the second kind (1.5),
or the Gauss formula for the Chebyshev weight function of the second kind (cf. [16, Sect. 5]),
The moduli of the actual errors, in units of \(10^{-6}\), are shown in Table 4.
Our numerical results indicate that formula (5.6) is more accurate than formulae (5.7) and, particularly, (5.8) for all values of a, probably because the nodes of the former are distributed closer to the point of singularity. For \(a<0\), all quadrature formulae converge extremely slowly, apparently due to the combined effect of two singularities in the integrand of (5.5), while things improve dramatically as a increases from 0 to 1.
References
Brass, H., Schmeisser, G.: The definiteness of Filippi’s quadrature formulae and related problems. In: Hämmerlin, G. (ed.) Numerische Integration, ISNM 45, pp. 109–119. Birkhäuser, Basel (1979)
Brass, H., Schmeisser, G.: Error estimates for interpolatory quadrature formulae. Numer. Math. 37, 371–386 (1981)
Davis, P.J., Rabinowitz, P.: Methods of Numerical Integration, 2nd edn. Academic Press, San Diego (1984)
Fejér, L.: Mechanische Quadraturen mit positiven Cotesschen Zahlen. Math. Z. 37, 287–309 (1933)
Gautschi, W.: Numerical quadrature in the presence of a singularity. SIAM J. Numer. Anal. 4, 357–362 (1967)
Gautschi, W.: Remainder estimates for analytic functions. In: Espelid, T.O., Genz, A. (eds.) Numerical Integration: Recent Developments, Software and Applications, pp. 133–145. Kluwer Academic Publishers, Dordrecht (1992)
Gradshteyn, I.S., Ryzhik, I.M.: Table of Integrals, Series and Products, 4th edn. Academic Press, San Diego (1980)
Hämmerlin, G.: Fehlerabschätzung bei numerischer Integration nach Gauss. In: Brosowski, B., Martensen, E. (eds.) Methoden und Verfahren der mathematischen Physik, vol. 6, pp. 153–163. Bibliographisches Institut, Mannheim (1972)
Hasegawa, T., Sugiura, H.: Error estimate for a corrected Clenshaw-Curtis quadrature rule. Numer. Math. 130, 135–149 (2015)
Kütz, M.: A note on error estimates for the Clenshaw-Curtis and other interpolatory quadratures. Analysis 4, 45–51 (1984)
Mason, J.C., Handscomb, D.C.: Chebyshev Polynomials. Chapman & Hall/CRC, Boca Raton, FL (2002)
Notaris, S.E.: Interpolatory quadrature formulae with Chebyshev abscissae of the third or fourth kind. J. Comput. Appl. Math. 81, 83–99 (1997)
Notaris, S.E.: Interpolatory quadrature formulae with Chebyshev abscissae. J. Comput. Appl. Math. 133, 507–517 (2001)
Notaris, S.E.: Integral formulas for Chebyshev polynomials and the error term of interpolatory quadrature formulae for analytic functions. Math. Comp. 75, 1217–1231 (2006)
Notaris, S.E.: The error norm of quadrature formulae. Numer. Algorithms 60, 555–578 (2012)
Notaris, S.E.: Product integration rules for Chebyshev weight functions with Chebyshev abscissae. J. Comput. Appl. Math. 257, 180–194 (2014)
Trefethen, L.N.: Is Gauss quadrature better than Clenshaw-Curtis? SIAM Rev. 50, 67–87 (2008)
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
About this article
Cite this article
Notaris, S.E. On a corrected Fejér quadrature formula of the second kind. Numer. Math. 133, 279–302 (2016). https://doi.org/10.1007/s00211-015-0750-5
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00211-015-0750-5