Abstract
This paper fills an important gap in the convergence analysis of collocation solutions in spaces of continuous piecewise polynomials for Volterra integral equations of the second kind. Our analysis is then extended to Volterra functional integral equations of the second kind with constant delays.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
The convergence analysis of piecewise polynomial collocation solutions for Volterra integral equations (VIEs) of the second kind,
with continuous kernel K(t, s) is now largely well understood; see [4–6] and, especially, the surveys [1, 3]. However, there has remained an important gap in the convergence analysis of collocation solutions for the second-kind VIE (1.1): it concerns the convergence/divergence of piecewise polynomial collocation solutions for (1.1) that are globally continuous on I and correspond to collocation points that do not include the points of the underlying mesh \(I_h\).
It is the aim of this paper to close this gap and to employ the gained insight to establish the analogous convergence analysis for the Volterra functional integral equation (VFIE) with constant delay \(\tau > 0\),
The outline of the paper is as follows. In Sects. 2 and 3 we state our main results on the convergence of globally continuous piecewise polynomial collocation solutions for the second-kind VIE (1.1) and the second-kind VFIE (1.2); their proofs are given in Sects. 4 and 5. In Sect. 6, we use a number of examples to illustrate the validity of our results on the attainable order of these collocation solutions. Section 7 concludes with a concluding remark.
2 Continuous collocation solutions for second kind VIEs
2.1 Meshes and collocation spaces
Let \(I_h:=\left\{ t_{n} := nh: \; n = 0,1, \ldots , N \;\, (t_{N} = T)\right\} \) be a given mesh on \(I = [0,T]\), with \(\sigma _n := [t_n,t_{n+1}]\) and mesh diameter \(h = T/N\). We seek a collocation solution \(u_h\) for (1.1) in the space
where \(\pi _m\) denotes the space of all (real) polynomials of degree not exceeding m. For a prescribed set of collocation points
\(u_h\) is defined by the collocation equation
with \(u_h(0)=g(0)\).
Consequently the collocation polynomial can be written as (see [2])
where \(U_{n,i}:=u_h'(t_{n,i}), t_{n,i}:=t_n+c_ih\) and the polynomials
denote the Lagrange fundamental polynomials with respect to the (distinct) collocation parameters \(\{c_i\}\).
Integrating (2.3), we obtain
where \(\beta _j(s):=\int _0^sL_j(v)dv\).
Therefore, at \(t=t_{n,i}\),
where \(a_{ij}:=\beta _j(c_i)\).
Denote \(A:=(a_{ij})_{m\times m}\), \(e:=(1,\ldots , 1)^T\), \(G_n:=(g(t_{n,1}),\ldots , g(t_{n,m}))^T\), \(U_n:=(U_{n,1},\ldots ,U_{n,m})^T\), \(B_n^{(l)}:=\left( \int _0^{1}K(t_{n,i},t_l+sh)\beta _j(s)ds\right) \; (0\le l\le N-1)\), \(B_n:=\left( \int _0^{c_i}K(t_{n,i},t_n+sh)\beta _j(s)ds\right) \), \(C_n^{(l)}:=diag\left( \int _0^{1}K(t_{n,i},t_l+sh)ds\right) \; (0\le l\le N-1)\), and \(C_n:=diag\left( \int _0^{c_i}K(t_{n,i},t_n+sh)ds\right) \), we have
where \(I_m\) denotes the identity in \(L(\mathbb {R}^m)\).
If \(g \in C(I)\) and \(K \in C(D) \; (D := \{(t,s): \; 0 \le s \le t \le T\})\), (2.6) determines a unique \(u_h \in S_m^{(0)}(I_h)\) for all sufficiently small mesh diameters, say \(h \in (0,\bar{h})\). However, the resulting collocation solution will not converge uniformly on I to the exact solution of (1.1) for every choice of the collocation parameters \(\{c_i\}\): while the convergence statement
holds whenever \(c_1 > 0\) and \(c_m = 1\) (cf. [3], and [7, 8]), this will in general no longer remain true when \(c_m < 1\). If (2.7) holds, the order of convergence will not be the same for all \(\{c_i\}\).
2.2 The main convergence results
Theorem 2.1
Assume that \(g\in C^{m+2}(I), K\in C^{m+2}(D)\), and \(u_h\in S_{m}^{(0)}(I_h)\) is the collocation solution for the second-kind Volterra integral equation (1.1) defined by the collocation equation (2.2) whose underlying meshes have mesh diameters \(h < \bar{h}\). Then (2.7) holds if, and only if, the collocation parameters \(\{c_i\}\) satisfy the condition
The corresponding attainable global order of convergence is given by
where the constant C depends on the collocation parameters \(\{c_i\}\) but not on h.
3 Continuous collocation solutions for second kind VIEs with constant delay
3.1 Meshes and collocation spaces
It is well known (see for example [2, Ch. 4]) that the constant delay \(\tau > 0\) in (1.2) induces the primary discontinuity points \(\xi _{\mu } = \mu \tau \; (\mu \ge 0)\) at which the regularity of the solution u(t) is, at least for small values of \(\mu \), lower than it is in \((\xi _{\mu }, \xi _{\mu +1})\). Thus, the collocation solution \(u_h \in S_m^{(0)}(I_h)\) will attain an order of global convergence equal to that for VIEs (1.1) without delay only if the underlying mesh \(I_h\) includes these primary discontinuity points. Assuming for ease of notation that \(T = \xi _{M+1}\) for some \(M \ge 1\), we choose this so-called constrained mesh to be
where \(h = \tau /N\). We set \(\sigma _n^{(\mu )}:=[t_n^{(\mu )},t_{n+1}^{(\mu )}]\). The solution u of (1.2) will be approximated by the collocation solution
using collocation points
corresponding to prescribed collocation parameters \(\{c_i\}\) with \(0 < c_1 < \cdots < c_m \le 1\). Hence, the collocation equation for the subinterval \(\sigma _n^{(\mu )}\) is
If \(\mu = 0\), the values of \(u_h\) at \(t\in [-\tau , 0]\) are determined by the given initial function, i.e., \(u_h(t)=\varphi (t)\).
3.2 The main convergence results
Theorem 3.1
Assume that \(g\in C^{m+2}(I), K\in C^{m+2}(D), \varphi \in C^{m+1}[-\tau , 0]\), and let \(u_h\in S_{m}^{(0)}(I_h)\) be the collocation solution for the second-kind VFIE (1.2) determined by the collocation equation (3.3), using constrained meshes \(I_h\) of the form (3.1). Then \(u_h\) converges uniformly on I to the solution u of (1.2) if, and only if, the collocation parameters in (3.2) satisfy the condition
The resulting attainable global order of convergence is then given by
where the constant C depends on the collocation parameters \(\{c_i\}\) but not on h.
Remark 3.1
The convergence results for the case \(0 < c_1 < \cdots < c_m = 1\) follow trivially from the proof of Theorem 2.1 (see also [8] and [3]). A similar conclusion holds in the case of second-kind VFIEs (Theorem 3.1 and its proof).
Remark 3.2
The case \(\rho _m= 1\) can happen only when m is even, but as Theorem 2.1 and Theorem 3.1 describe, this case leads to a reduction of the order of convergence.
4 Proof of Theorem 2.1
We assume that \(c_m<1\).
According to the theory of Lagrange interpolation we may write
where the Peano remainder term and Peano kernel (see [2]) are given by
and
Here, \((v-z)_{+}^{m-1}:=0\) for \(v<z\) and \((v-z)_{+}^{m-1}:=(v-z)^{m-1}\) for \(v\ge z\).
Integration of (4.1) leads to
where \(R_{m,n}(s):={\int _0^sR_{m,n}^1(v)dv}\).
We first consider the case of constant kernel \(K(t, s)\equiv 1\). This case already contains all important ideas.
By (2.4) and (4.2), the collocation error \(e_h:=u-u_h\) on \([t_n, t_{n+1}]\) may be written as
where \(\varepsilon _{n,i}:=u'(t_{n,i})-u_h'(t_{n,i})\). Particularly,
By (1.1)–(2.2) and using (4.3), it can be shown that
where \(\gamma _j(s):=\int _0^s\beta _j(v)dv, \;\; b_{ij}:=\int _0^{c_i}\beta _j(s)ds=\gamma _j(c_i), \) and \( \tilde{R}_{m,n}(c_i):=\sum \nolimits _{l=0}^{n-1}h\int _0^{1}R_{m,l}(s)ds+h\int _0^{c_i}R_{m,n}(s)ds. \) So by (4.4) and (4.5), we have
By the standard technique used by Brunner (see [2]), rewriting (4.6) with n replaced by \(n-1\) and with \(i= m\) and subtract it from (4.6), we find
This can be written in the more concise form
with obvious meaning of \(R_{m,{n}}\), and with \(C:=diag(c_1, \ldots , c_m)\), \(\varepsilon _n:=(\varepsilon _{n,1}, \ldots , \varepsilon _{n,m})^T\), \(B:=(b_{ij})_{m\times m}\), \(\gamma :=(\gamma _1(1),\ldots , \gamma _m(1))^T\).
Since \(e_h\) is continuous in I, and hence at the mesh points, by (4.3) and \(e_h(0)=0\), we also have the relation (see [2, (1.1.27)])
where \(b_j:=\int _0^1L_j(s)ds\) and \(b^T:=(b_1,\ldots ,b_m)\).
Substituting (4.8) into (4.7), we have
This equation can be written in the form
or
with obvious meaning of \(\tilde{R}_{m,{n}},\bar{R}_{m,{n}}\) and D.
Since
the rank of the matrix \(ee_m^TA-eb^T\) is one, implying that the rank of the matrix \(A^{-1}(ee_m^TA-eb^T)\) is also one. This means that this matrix has exactly one nonzero eigenvalue. Setting \(A^{-1}:=(v_{ij})_{m\times m}\), we have
and the nonzero eigenvalue is
By Proposition 3.8 and Theorem 3.10 of [5], we know that the stability function \(R(z) = P(z)/Q(z)\) of the collocation method has the value \(R(\infty )=1-b^TA^{-1}e\), where Q(z) and P(z) are the polynomials
with
Therefore,
that is, the only nonzero eigenvalue of \(A^{-1}(ee_m^TA-eb^T)\) is
Therefore \(A^{-1}\left( ee_m^TA-eb^T\right) \) is diagonalizable and there exists a nonsingular matrix T such that
Multiplying (4.10) by \(T^{-1}\) and setting \(Z_n:=T^{-1}\varepsilon _{n}\), we obtain
We consider the following three cases:
Case I \(-1< \rho _m<1\)
Using standard techniques of error estimation for collocation solutions of VIEs (see [2, 5, 7]), we know that there exists a constant \(C_{1}\), such that
It follows from (4.6) that there exist constants \(C_{2}\) and \(C_{3}\) such that
and hence by the discrete Gronwall inequality (see [2]), there exists a constant \(C_{4}\), such that
Case II \(\rho _m=-1\)
Rewriting (4.11) with n replaced by \(n-1\) and subtract it from (4.11), we have
Therefore,
Since \(\bar{R}_{m,{n}}-\bar{R}_{m,{n-1}}=O(h)\) for \(u\in C^{m+2}\), we define \(r_{m,n}:=\frac{\bar{R}_{m,{n}}-\bar{R}_{m,{n-1}}}{h}\) and set
We may then write
The eigenvalues of the matrix G are \(\underbrace{1, 1,\ldots , 1}_m; -1, \underbrace{0, \ldots , 0}_{m-1}\). The eigenvalue 1 of multiplicity m has m linearly independent eigenvectors, while to the eigenvalue 0 of multiplicity \(m-1\) there correspond \(m-1\) linearly independent eigenvectors. Therefore, G is diagonalizable, and there exists a nonsingular matrix P such that
Defining \(Y_n:=P^{-1}X_n\) we obtain
Similar to [7] we can assert that there exist constants \(C_5, C_6\) so that
An induction argument then leads to
and we can then show that there exists a constant \(C_7\) such that,
By (4.6), and as in Case I, there exists hence a constant \(C_8\) so that
Case III \(\rho _m=1\)
Here, the eigenvalues of G defined in Case II are \(\underbrace{1, 1,\ldots , 1}_m; 1, \underbrace{0, \ldots , 0}_{m-1}\), where now the eigenvalue 1 of multiplicity \(m+1\) also has m linearly independent eigenvectors. This means that G is not diagonalizable, but there exists a nonsingular matrix Q, such that
Defining \(\bar{\varLambda }:=Q^{-1}GQ, \bar{Y}_n:=Q^{-1}X_n\) and recalling (4.16) we obtain
An induction argument yields
and thus there exist constants \(C_9, C_{10}\) such that
It is easily to check that
Therefore, there exists a constant \(C_{11}\) such that
and an argument analogous to the one employed in the analysis of Case I shows that there exists a constant \(C_{12}\) such that
Obviously, the collocation solution \(u_h\) is divergent if \(|\rho _m|>1\). The proof is completed by recalling (2.3), (4.1) and (4.3).
In the following, we prove the results for general, non-constant kernels. Now, by (4.4), (1.1)–(2.2) and using (4.3), we obtain
By the standard technique used by Brunner (see [2]), rewriting (4.23) with n replaced by \(n-1\) and with \(i= m\) and subtract it from (4.23), we find
where \(\tilde{R}_{m,n}(c_i):=-R_{m,{n}}(c_i)+R_{m,{n-1}}(c_m)+h\int _0^{1}K(t_{n,i}, t_{n-1}+sh)R_{m,{n-1}}(s)ds +\sum \nolimits _{l=0}^{n-2}h\int _0^{1}[K(t_{n,i}, t_l+sh)-K(t_{n-1,m}, t_l+sh)]R_{m,l}(s)ds +h\int _0^{c_i}K(t_{n,i}, t_n+sh)R_{m,n}(s)ds-h\int _0^{c_m}K(t_{n-1,m}, t_{n-1}+sh)R_{m,n-1}(s)ds\), \(\xi _{{n,i}}\in (t_{n-1,m}, t_{n,i})\).
This can be written in the more concise form
with obvious meaning of \(R^{(1)}_{m,{n}}\), and with \(\bar{C}_n^{(l)}:=diag (\int _0^{1}K'_1(\xi _{n,i}, t_l+sh)ds)\; (0\le l\le N-1)\) and \(\bar{B}_n^{(l)}:= (\int _0^{1}K'_1(\xi _{n,i}, t_l+sh)\beta _j(s)ds)\; (0\le l\le N-1)\).
Substituting (4.8) into (4.24), we have
with obvious meaning of \(\bar{\tilde{R}}_{m,{n}}\). This equation can be written in the form
or
with obvious meaning of \(\bar{\bar{R}}_{m,{n}}\) and \(\tilde{D}_n\).
Comparison (4.10) of the case \(K(t,s)\equiv 1\) with (4.26), and similar to the proof of the case \(K(t,s)\equiv 1\), we can obtain now (4.11) becomes
We also consider the following three cases:
Case I \(-1< \rho _m<1\)
By the same technique of the case \(K(t,s)\equiv 1\), we can prove that there exists a constant \(\tilde{C}_{4}\), such that
Case II \(\rho _m=-1\)
Rewriting (4.27) with n replaced by \(n-1\) and subtract it from (4.27), we find
Notice that \(\tilde{D}_{n}-\tilde{D}_{n-1}=O(h)\), therefore,
Now (4.17) becomes
Similar to the case of \(K(t,s)\equiv 1\), we can assert that there exist constants \(\tilde{C}_5, \tilde{C}'_5, \tilde{C}_6\) so that
An induction argument then leads to
Therefore, by the discrete Gronwall inequality (see [2]), we can get that there exists a constant \(\tilde{C}_7\) such that,
and similar to the case of \(K(t,s)\equiv 1\), we can then show that there exists a constant \(\tilde{C}_8\) so that
Case III \(\rho _m=1\)
Using the technique of [7], we write the collocation approximation \(u_h\) and the exact solution in the form
and
where \(\eta _n, \eta '_n\in (t_n, t_{n+1})\).
where \(\hat{R}_{n}(s):=\frac{u^{(m)}(\eta '_n)-u_h^{(m)}(\eta _n)}{m!}\prod \limits _{i=1}^m(s-c_i)\).
Now, by (1.1)–(2.2) and using (4.35), we obtain
with obvious meanings of \(\bar{\hat{R}}_{n}(s)\).
Rewriting (4.36) with n replaced by \(n-1\) and \(i=m\) and subtract it from (4.36), we can get
Denoting \(E_n:=(e_h(t_{n,1}), \ldots , e_h(t_{n,m}))^T\) and noticing that \(\bar{\hat{R}}_{n}(s)-\bar{\hat{R}}_{n-1}(s)=O(h)\), we can rewrite (4.37) as the more concise form
Since
the rank of \(ee_m^T\) is 1, and the unique nonzero eigenvalue is 1, so \(ee_m^T\) is diagonalizable and there exists a nonsingular matrix \(\tilde{T}\), such that
Denote \(\tilde{E}_n=\tilde{T}^{-1}E_n\). Then (4.38) becomes
Therefore, there exist constants \(\bar{C}_5, \bar{C}'_5\) and \(\bar{C}_6\), such that
Similar to Case II, we can get that there exists a constant \(\bar{C}_7\) such that
By (4.35), we can then get there exists a constant \(\bar{C}_8\) such that
Obviously, the collocation solution \(u_h\) is divergent also if \(|\rho _m|>1\). The proof is completed by recalling (2.3), (4.1) and (4.3).
5 Proof of Theorems 3.1
On \(\sigma _n^{(\mu )}:=(t_n^{(\mu )}, t_{n+1}^{(\mu )}]\), the derivative \(u_h'\) of the collocation solution has the local Lagrange representation,
where \(U_{n,i}^{(\mu )}:=u_h'(t_{n,i}^{(\mu )})\). Upon integration of (5.1) we obtain
Since on each subinterval \(\sigma _{n}^{(\mu )}\) the exact solution of the delay VIE (1.2) is in \(C^{m+2}\), we may write
where the Peano remainder term is given by
Integration of (5.3) leads to
with \(R^{(\mu )}_{m,n}(s):=\int _0^sR_{m,n}^{(1, \mu )}(v)dv\).
For ease of notation we will again assume that \(K(t,s)\equiv 1\), and we can extend to the proof to the non-constant kernel by the same technique as the proof of Theorem 2.1.
By (5.2) and (5.4), the collocation error \(e_h:=u-u_h\) on \(\bar{\sigma _n}^{(\mu )}:=[t_n^{(\mu )}, t_{n+1}^{(\mu )}]\) can be written as
where \(\varepsilon _{n,i}^{(\mu )}=u'(t_{n,i}^{(\mu )})-u_h'(t_{n,i}^{(\mu )})\). Particularly,
By (1.2)–(3.3) and using (5.5), we have
where \( \tilde{R}_{m,n}^{(\mu )}(c_i) = \sum \nolimits _{l=n+1}^{N-1}h\int _0^{1}R^{(\mu -1)}_{m,l}(s)ds +h\int _{c_i}^1R^{(\mu -1)}_{m,n}(s)ds +\sum \nolimits _{l=0}^{n-1}h\int _0^{1} R_{m,l}^{(\mu )}(s)ds +h\int _0^{c_i}R_{m,n}^{(\mu )}(s)ds. \) Thus, by (5.5) and (5.7), we have
Rewriting (5.8) with n replaced by \(n-1\) and with \(i=m\) and subtract it from (5.8), we can obtain
By (5.5), the continuity of \(e_h\) on I, and \(e_h(0)=0\), we find by induction
For \(\mu =0\), it follows from (5.9) and (5.10) that we can write the error equation in the form
or
with obvious meaning of \(\tilde{R}_{m,n}^{(0)}\).
The proof of Theorem 2.1 reveals that \(\varepsilon _{n}^{(0)}\) converges if, and only if
and that there exist constant \(C_{1}^{(0)}\), such that
Equation (5.8) implies that there exist constants \(C_{2}^{(0)}\) and \(C_{3}^{(0)}\) such that
and thus the discrete Gronwall inequality (see [2]) guarantees the existence of a constant \(C_{4}^{(0)}\) for which
holds. Assume that for \(\nu =1,\ldots , \mu -1\), \(\varepsilon _{n}^{(\nu )}\) converges if, and only if
and that there exist constants \(C_{1}^{(\nu )}\) and \(C_{4}^{(\nu )}\) such that
and
with obvious meaning of \(R_{m,n}^{(\mu )}\).
In the remaining part of the proof we consider the following three cases.
Case I \(-1 < \rho _m <1\)
By assumption (5.14) and (5.15), we obtain from (5.16)
Proceeding as in the proof of Theorem 2.1 we see that there exists a constant \(C_{1}^{(\mu )}\) such that
and hence, by (5.8) and the discrete Gronwall lemma (see [2]), there exist constants \(C_{2}^{(\mu )},\, C_{3}^{(\mu )}\) and \(C_{4}^{(\mu )}\) for which
and
are true.
Case II \(\rho _m=-1\)
Rewriting (5.16) with n replaced by \(n-1\) and with \(i=m\) and subtract it from (5.16), and by (5.10), we can obtain
with obvious meaning of \(\bar{R}_{m,n}^{(\mu )}\). Therefore,
Obviously, the inverse of the coefficient matrix is \(,\) so that by assumption (5.14) we obtain
where the meaning of \(\tilde{\bar{R}}_{m,n}^{(\mu )}\) is clear. Since the eigenvalues of the matrix on the right-hand side are \(\underbrace{1, 1,\ldots , 1}_m; -1, \underbrace{0, \ldots , 0}_{m-1}\), we may use an argument similar to the one in Case II for Theorem 2.1 to establish the existence of constants \(C_{5}^{(\mu )}\) and \(C_{6}^{(\mu )}\) such that
Case III \(\rho _m=1\)
Using the technique of [7], on \(\sigma _n^{(\mu )}:=(t_n^{(\mu )}, t_{n+1}^{(\mu )}]\), we write the collocation approximation \(u_h\) and the exact solution in the form
and
where \(\eta ^{(\mu )}_n, (\eta ^{(\mu )}_n)'\in (t^{(\mu )}_n, t^{(\mu )}_{n+1})\).
where \(\hat{R}^{(\mu )}_{n}(s):=\frac{u^{(m)}((\eta ^{(\mu )}_n)')-u_h^{(m)}(\eta ^{(\mu )}_n)}{m!}\prod \nolimits _{i=1}^m(s-c_i)\). Therefore, the proof is completed by resorting to the finial argument in the proof of Theorem 2.1 and the mathematical induction.
6 Numerical examples
In this section, we present two examples to illustrate the foregoing convergence results. For the numerical solution of (1.1) and (1.2), we choose \(m=1\), \(m=2\) and \(m=3\). For \(m=1\) we use \(c_1=\frac{1}{3}, 0.49, 0.5, 0.8, 1\) respectively, and \(\rho _m=-2, -\frac{51}{49}, -1, -\frac{1}{4}, 0\) respectively. For \(m=2\) we use the Gauss collocation parameters, \(c_1=\frac{3-\sqrt{3}}{6}, c_2=\frac{3+\sqrt{3}}{6}\); the Radau IIA collocation parameters, \(c_1=\frac{1}{3}, c_2=1\); and three sets of arbitrary collocation parameters, \(c_1=\frac{1}{2}, c_2=1; c_1=\frac{1}{3}, c_2=\frac{2}{3}; c_1=\frac{1}{6}, c_2=0.82\) respectively, and \(\rho _m=1, 0, 0, 1, \frac{45}{41}\) respectively. For \(m=3\) we use the Gauss collocation parameters, \(c_1=\frac{5-\sqrt{15}}{10}, c_2=\frac{1}{2}, c_3=\frac{5+\sqrt{15}}{10}\); the Radau IIA collocation parameters, \(c_1=\frac{4-\sqrt{6}}{10}, c_2=\frac{4+\sqrt{6}}{10}, c_3=1\); and three sets of arbitrary collocation parameters, \(c_1=\frac{1}{2}, c_2=\frac{2}{3}, c_3=1; c_1=\frac{1}{3}, c_2=\frac{1}{2}, c_3=\frac{2}{3}; c_1=\frac{1}{4}, c_2=\frac{1}{2}, c_3=0.7\) respectively, and \(\rho _m=-1, 0, 0, -1, \frac{9}{7}\) respectively. In Tables 1, 2, 3, 4, 5 and 6 we list the absolute errors for the five collocation parameters and for \(m=1\), \(m=2\) or \(m=3\).
Example 6.1
In (1.1) let \(K(t, s)=e^{t-s}\) and with g(t) such that the exact solution is \(u(t)=e^{-t}\).
Example 6.2
Consider (1.2) with \(K(t, s)=e^{t-s}, \;\tau =1\) and \(\phi (t)=1\), and with g(t) such that the exact solution is \(u(t)=\cos t\) for \(t\ge 0\).
From Tables 1, 2, 3, 4, 5 and 6, we can see that the numerical results are consistent with our theoretical analysis.
In practical applications one will rarely use collocation space \(S_m^{(0)}(I_h)\) with \(m > 3\), since \(m = 3\) yields the global convergence order \(p = m+1 = 4\) and very small absolute errors already for moderately small stepsizes.
7 Concluding remark
As we mentioned in Sect. 1, the main purpose of this paper was to close a gap in previous convergence analyses of continuous piecewise polynomial collocation solutions for second-kind Volterra integral equations. While such globally continuous collocation approximations may occasionally be desirable (for example in VFIEs with non-vanishing delays), their accuracy is in general inferior to the one obtained by using discontinuous piecewise polynomials (at essentially the same computational cost).
References
Brunner, H.: The use of splines in the numerical solution of differential and Volterra integral equations. In: Dubuc, S., Deslauriers, G. (eds.) Spline Functions and the Theory of Wavelets (Montreal 1996), CRM Proceedings and Lecture Notes, vol. 18, pp. 15–31. American Mathematical Society, Providence (1999)
Brunner, H.: Collocation Methods for Volterra Integral and Related Functional Differential Equations. Cambridge University Press, Cambridge (2004)
Brunner, H.: On the divergence of collocation solutions in smooth piecewise polynomial spaces for Volterra integral equations. BIT 44, 631–650 (2004)
El Tom, M.E.A.: On the numerical stability of spline function approximations to the solutions of Volterra integral equations of the second kind. BIT 14, 136–143 (1974)
Hairer, E., Lubich, Ch., N\(\phi \)rsett, S.P.: Order of convergence of cne-step methods for Volterra integral equations of the second kind. SIAM J. Numer. Anal. 20, 569–579 (1983)
Hung, H.S.: The numerical solution of differential and integral equations by spline functions, MRC Tech. Summary Report, No. 1053, University of Wisconsin, Madison (1970)
Kauthen, J.P., Brunner, H.: Continuous collocation approximations to solutions of first kind Volterra equations. Math. Comput. 66, 1441–1459 (1997)
N\(\phi \)rsett, S.P.: Splines and collocation for ordinary initial value problems. In: Singh, S.P. (ed.) Approximation Theory and Spline Functions (St. John’s, Nfld., 1983), NATO Adv. Sci. Inst. Ser. C Math. Phys., vol. 136, pp. 397–417. Reidel, Dordrecht (1984)
Acknowledgments
The research of Hermann Brunner was supported by the Hong Kong Research Grants Council (GRF Grant HKBU 200207, 200210), and the Natural Sciences and Engineering Research Council of Canada (NSERC Discovery Grant No. 9406). The research of Hui Liang was supported by the National Nature Science Foundation of China (No. 11101130), the Research Fund of the Heilongjiang Provincial Education Department for the Academic Backbone of the Excellent Young People (No. 1254G044), Science and Technology Innovation Team in Higher Education Institutions of Heilongjiang Province (No. 2014TD005), the Heilongjiang University Science Funds for Distinguished Young Scholars (No. JCL201303), the Natural Science Foundation of Heilongjiang Province (No. A201211). Part of the work of the first author was carried out while she was a Visiting Research Scholar at Hong Kong Baptist University (March 2010 and August 2011); she gratefully acknowledges the hospitality extended to her by HKBU’s Department of Mathematics. Especially, she is also thankful to Professor Tao Tang and Professor Hermann Brunner for their invitation to visit HKBU. The authors thank the anonymous referee for his/her careful reading of the manuscript and for the valuable comments and suggestions. They greatly improved the presentation of the results.
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Anne Kværnø.
Rights and permissions
About this article
Cite this article
Liang, H., Brunner, H. On the convergence of collocation solutions in continuous piecewise polynomial spaces for Volterra integral equations. Bit Numer Math 56, 1339–1367 (2016). https://doi.org/10.1007/s10543-016-0609-x
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10543-016-0609-x
Keywords
- Volterra integral equations
- Collocation solutions
- Continuous piecewise polynomials
- Convergence
- Volterra functional integral equations with constant delays