Abstract
We consider the numerical solution of boundary value problems for general neutral functional differential equations. The problems are restated in an abstract form and, then, a general discretization of the abstract form is introduced and a convergence analysis of this discretization is developed.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Let V be the space of the continuous functions \([ a,b]\rightarrow \mathbb {R}^{d}\) and let U be a Banach space of integrable functions \([a,b] \rightarrow \mathbb {R}^{d}\). We deal with the numerical solution of the functional differential equation boundary value problem (BVP)
where the functionals \(F:[a,b] \times V\times U \times \mathbb {R} ^{d_{0}}\rightarrow \mathbb {R}^{d}\) and \(B:V\times U\times \mathbb {R} ^{d_{0}}\rightarrow \mathbb {R}^{d}\times \mathbb {R}^{d_0}\) are given and the pair \((y,p)\in V\times \mathbb {R}^{d_0}\) is unknown.
The reason to include \(p\in \mathbb {R}^{d_0}\) as an unknown of the problem (1) is that, in many real applications, there are parameters to be determined along with the solution y. For example, the determination of periodic solutions for an autonomous functional differential equation reduces to a BVP, where the unknown period of the periodic solution appears as a parameter.
The general functional differential equation
in (1) includes the two particular and important cases of differential equations with deviating arguments
and integro-differential equations
In order to restate (3) and (4) in the form (2), it is necessary to have
for (3) and
for (4). However, one often encounters Eq. (3), or Eq. (4), where the condition (5), or the condition (6), is not fulfilled. For example, this happens when some \(\theta _r(t)\) or \(\vartheta _s(t)\) in (3), or one of \(\alpha (t)\) and \(\beta (t)\) in (4), has the form \(t\pm \tau \), where \(\tau >0\). For such equations, we need to specify the solution y and its derivative \(y^{\prime }\) outside the interval [a, b] by the side condition
where \(\phi ,\varphi :(-\infty ,a) \cup ( b,+\infty )\rightarrow \mathbb {R}^d \) are given functions (of course, it makes sense take \(\varphi =\phi ^\prime \)). Then, the equation can be still restated in the form (2) by incorporating the side condition in the functional F: we write the Eq. (3) as
where
and the Eq. (4) as
where
Observe that the side condition (7) is considered as a part of the functional differential equation (2), not as a boundary condition.
We recall here that a particular and important case of the Eq. (3) is given by delay differential equations, where
and two particular and important cases of the Eq. (4) are given by Fredholm integro-differential equations, where
and Volterra integro-differential equations, where
We also remark that the general form (2) includes integro-differential equations
which cannot be seen as Eq. (4).
The general boundary condition
in (1) includes the classical boundary condition
and the more general multipoint boundary condition
where \(t_{i}\in (a,b),\,i=1,\ldots ,q\), and integral boundary condition
Note that, in general, boundary conditions for first order differential equation do not involve the derivative \(y^\prime \). In this paper, we consider boundary conditions involving \(y^\prime \) since our theory can deal with this situation without any further complication.
By following the usual terminology, the functional differential equation (2) can be called neutral since, in general, the values \( F(t,y,y^{\prime },p)\) depend on \(y^\prime \).
Of course, the neutral equation (2) also includes the case where the values of F do not depend on \(y^\prime \). In this case, we say that the Eq. (2) is non-neutral. Similarly, we say that the boundary condition (8) is non-neutral if the values of B do not depend on \(y^\prime \) and that the BVP (1) is non-neutral if both (2) and (8) are non-neutral.
1.1 Numerical literature on BVPs for functional differential equations and aim of the paper
The papers dealing with the numerical solution of functional differential equation BVPs, apart from [13, 30] described below, address special cases of the problem (1). Table 1 collects such papers according to the special case considered.
Instead, the papers [13, 30] deal with BVPs for general non-neutral second order functional differential equations. The paper [13] deals with BVPs
where \(\mathcal {F}\) is an operator acting on y, and considers a discretization of the second derivative by a central difference. As a consequence, a method of order two is obtained. The paper [30] deals with BVPs
where \(\mathcal {F}\) is an operator acting on y and \(y^\prime \), and uses special continuous (dense output) methods for second order differential equations. Such methods can reach an arbitrarily high order, if \(\mathcal {F}\left( y,y^{\prime }\right) \) is independent of \(y^\prime \), and have order two at most, otherwise.
Regarding the theoretical (non-numerical) literature on BVPs for functional differential equations, among many papers, we mention here the monograph [26], which contains a collection of articles dealing with many aspects of the theory of such problems, and the book [2], which considers only the non-neutral case.
Aim of the present paper is to study the numerical solution of problems (1). We restate such problems in an abstract form and, then, we introduce a general type of discretization of the abstract form and develop a convergence analysis of this discretization.
The general discretization studied in this paper includes the two particular discretizations of the problem (1) given by collocation method and the Fourier series method. However, in order to avoid having a very long paper, here we do not deal with these two discretizations. They are the subject of the papers [41, 42] and the forthcoming papers [43, 44]. The present paper contains the theoretical bases for the numerical solution of functional differential equations BVPs.
When compared to the current literature, the research started in this paper and continued in [41–44] contains the following advances.
-
1.
The general form (1) of BVP has not ever been studied in literature, even confining to the non-neutral case. Also the abstract form and the general type of discretization considered in this paper are a novelty.
-
2.
By confining to the non-neutral case, we consider a more general situation than that dealt in the papers [13, 30]. Moreover, the methods that we introduce have arbitrarily high order of convergence, unlike the methods in [13, 30] which have order two only.
-
3.
The study of the numerical solution of BVPs for neutral differential equations with deviating arguments is at a seminal stage. On this subject, there are only the three papers [5, 14, 17], where [17] introduces and proves the convergence of a method of order one and [5, 14] are experimental works without convergence proofs. Moreover, regarding BVPs for neutral integro-differential equations, the literature is confined to Volterra and Fredholm equations. In our research, we introduce methods for general neutral functional differential equations BVPs of arbitrarily high order of convergence.
The plan of the paper is the following. In Sect. 2, we introduce the abstract form of the problem (1). In Sect. 3, we introduce the general type of discretization used for the abstract form. In Sect. 4, we analyze the convergence of this general discretization. In Sect. 5, we specialize the results obtained in Sect. 4, in preparation for their application in [41–44] to the problem (1) discretized by the collocation method and the Fourier series method.
1.2 Notations
We finish this section giving a list of conventions and notations used throughout the paper.
-
The norm of a space Y is denoted by \(\Vert \ \cdot \ \Vert _Y\).
-
Cartesian product spaces are equipped with the norm given by the sum of the norms of the factor spaces.
-
In the space Y, the closed ball of center \(y\in Y\) and radius \(r\ge 0\) is denoted by \(\overline{B}(y,r)\).
-
The identity operator of a space Y is denoted by \(I_{Y}\).
-
The norm of a bounded linear operator L from the space Y to the space Z is denoted by \(\Vert L \Vert \), without any reference to the domain Y and the codomain Z.
-
The Fréchet-derivative of the operator A at the point y is denoted by DA(y) .
-
For an operator \(A:Y\rightarrow Z_1\times \cdots \times Z_k\), we define the operators \( A_{Z_i}:Y\rightarrow Z_i\), \(i=1,\ldots ,k\), by
$$\begin{aligned} A(y)=(A_{Z_1}(y),\ldots ,A_{Z_k}(y)),\quad y\in Y, \end{aligned}$$and call them the components of A.
2 The abstract form
We assume that the functional F in (1) is such that \(F(\cdot ,v,u,\beta ) \in U\), for any \((v,u,\beta )\in V\times U\times \mathbb {R}^{d_{0}}\). This assumption permits to introduce the operator \(\mathcal {F}:V\times U\times \mathbb {R}^{d_{0}}\rightarrow U\) given by
and write the functional differential equation (2) as
which is an equation in the space U.
The BVP (1) is now restated in abstract form, where we use the derivative \(y^{\prime }\), instead of y, as the actual unknown of (1).
Consider the very simple linear differential equation
where \(u\in U\) is given and \(v\in V\) is the unknown. Each solution of this equation is determined by a parameter \(\alpha \in \mathbb {R}^d\). Thus, we introduce a linear operator \(\mathcal {G}:U\times \mathbb {R} ^{d}\rightarrow V\) such that, for any \(u\in U\),
By following the usual terminology of the differential equations, the linear operator \(\mathcal {G}\) can be called a Green operator for the Eq. (11). Of course, examples of a Green operator for (11) are
where \(c\in [a,b]\).
The abstract form of the BVP (1) is based on the interpretation of the Eq. (11) as
In other words, we replace the derivative operator with the Green operator.
Once a Green operator for (11) is given, the abstract form is introduced by defining what we mean for a solution of (1). Let \((y,p) \in V\times \mathbb {R}^{d_{0}}\). We say that (y, p) is a solution of (1) if \(y=\mathcal {G}(u,\alpha ) \) for some \(u\in U\) and \(\alpha \in \mathbb {R}^{d}\) such that
Hence, we reach the following abstract form of the problem (1).
PAF (Problem in Abstract Form). Given:
-
a normed space \(\mathbb {V}\) and Banach spaces \(\mathbb {U}\), \(\mathbb {A}\) and \(\mathbb {B}\);
-
operators \(\mathfrak {F}:\mathbb {V}\times \mathbb {U}\times \mathbb {B}\rightarrow \mathbb {U}\) and \(\mathfrak {B}:\mathbb {V}\times \mathbb {U}\times \mathbb {B}\rightarrow \mathbb {A}\times \mathbb {B}\);
-
a linear operator \(\mathfrak {G}:\mathbb {U}\times \mathbb {A}\rightarrow \mathbb {V}\);
find a pair \((v,\beta )\in \mathbb {V}\times \mathbb {B}\) such that \(v=\mathfrak {G}(u,\alpha )\) for some \(u\in \mathbb {U}\) and \(\alpha \in \mathbb {A}\) satisfying
Clearly, the BVPs (1) are in the form PAF with \(\mathbb {V}=V\), \(\mathbb {U}=U\), \(\mathbb {A}=\mathbb {R}^d\), \(\mathbb {B}=\mathbb {R}^{d_0}\), \(\mathfrak {F}=\mathcal {F}\), \(\mathfrak {B}=B\) and \(\mathfrak {G}=\mathcal {G}\).
2.1 Other instances of the abstract form
Besides the BVPs (1), PAF includes other types of BVPs.
For example, consider a second order problem (not restated as a first order problem)
where \(F:[a,b]\times V\times \mathbb {R}^{d_0}\rightarrow \mathbb {R}^d\) and \(B:V\times \mathbb {R}^{d_0}\rightarrow \mathbb {R}^d\times \mathbb {R}^d\times \mathbb {R}^{d_0}\), V being, as above, the space of the continuous function \([a,b]\rightarrow \mathbb {R}^d\).
Let U be a Banach space of integrable function \([a,b]\rightarrow \mathbb {R}^d\) and consider the differential equation
where \(u\in U\) is given and \(v\in V\) is the unknown. A Green operator for (13) is the linear operator \(\mathcal {G}:U\times \mathbb {R}^d\times \mathbb {R}^d\rightarrow V\) given by
i.e.
Under the assumption that \(F(\cdot ,v,\beta )\in U\) for any \((v,\beta )\in V\times \mathbb {R}^{d_0}\), the problem (12) can be restated in the form PAF by introducing the operator \(\mathcal {F}:V\times \mathbb {R}^{d_0}\rightarrow U\) given by
The BVPs (12) are in the form PAF with \(\mathbb {V}=V\), \(\mathbb {U}=U\), \(\mathbb {A}=\mathbb {R}^d\times \mathbb {R}^d\), \(\mathbb {B}=\mathbb {R}^{d_0}\), \(\mathfrak {F}=\mathcal {F}\), \(\mathfrak {B}=B\) and \(\mathfrak {G}=\mathcal {G}\) (of course, here U, \(\mathcal {F}\), B and \(\mathcal {G}\) are those defined for the problems (12)).
Clearly, PAF also includes BVPs for general differential equations obtained by replacing the second derivative \(y^{\prime \prime }(t)\) in (12) with a general linear differentiation operator.
PAF even includes BVPs for partial functional differential equations. In fact, consider the problem
where \(\Omega \) is an open set of \(\mathbb {R}^d\) with boundary \(\partial \Omega \), \(\Delta \) is the Laplacian operator, \(F:\Omega \times V\rightarrow \mathbb {R}\), with V the space of the continuous function \(\overline{\Omega }=\Omega \cup \partial \Omega \rightarrow \mathbb {R}\), and \(B:V\rightarrow A\), with A the space of the continuous function \(\partial \Omega \rightarrow \mathbb {R}\).
Given a Banach space U of integrable functions \(\overline{\Omega }\rightarrow \mathbb {R}\), we consider the differential equation
where \(u\in U\) is given and \(v\in V\) is unknown. A Green operator for (15) is the linear operator \(\mathcal {G}:U\times A\rightarrow V\) given by
Under the assumption that \(F(\cdot ,v)\in U\) for any \(v\in V\), the problem (14) can be restated in the form PAF (without the space \(\mathbb {B}\)), by introducing the operator \(\mathcal {F}:V\rightarrow U\) given by
The BVPs (14) are in the form PAF with \(\mathbb {V}=V\), \(\mathbb {U}=U\), \(\mathbb {A}=A\), \(\mathfrak {F}=\mathcal {F}\), \(\mathfrak {B}=B\) and \(\mathfrak {G}=\mathcal {G}\).
Note that BVPs (14) have the space \(\mathbb {A}=A\) of infinite dimension. However, in the particular and important case of Dirichlet boundary conditions
where \(g\in A\), we can consider \(\mathbb {V}=\left\{ v\in V:v|_{\partial \Omega }\in \mathrm {span}(g)\right\} \), instead of \(\mathbb {V}=V\), and \(\mathbb {A}=\mathrm {span}(g)\), instead of \(\mathbb {A}=A\), where \(\mathrm {span}(g)=\{kg:k\in \mathbb {R}\}\), so to have the space \(\mathbb {A}\) of finite dimension.
2.2 The abstract form as a fixed point problem
From now on we consider the problem PAF with \(\mathbb {A}\) and \(\mathbb {B}\) of finite dimension, rather than its particular instance (1). By introducing the product Banach space
PAF can be seen as the search for fixed points of the operator \(\varPhi :X\rightarrow X\) given by
We have that \((v,\beta ) \in \mathbb {V}\times \mathbb {B}\) is a solution of PAF if and only if \(v=\mathcal {G}(u,\alpha ) \) for some fixed point \((u,\alpha ,\beta ) \in X\) of \(\varPhi \).
Regarding the operators \(\mathfrak {F}\) and \(\mathfrak {B}\) and the linear operator \(\mathfrak {G}\), we do the following assumptions.
-
A \(\varvec{\mathfrak {F}}\varvec{\mathfrak {B}}\) (Assumption \(\varvec{\mathfrak {F}}\varvec{\mathfrak {B}}\)). The operators \(\mathfrak {F}\) and \(\mathfrak {B}\) are Fréchet-differentiable at any point \((v_{0},u_{0},\beta _{0})\in \mathbb {V}\times \mathbb {U}\times \mathbb {B}\).
-
A \(\varvec{\mathfrak {G}}\). The linear operator \(\mathfrak {G}\) is bounded.
Since A \(\varvec{\mathfrak {F}\mathfrak {B}}\) and A \(\varvec{\mathfrak {G}}\) hold, \(\varPhi \) is Fréchet-differentiable at any point \(x_{0}=(u_{0},\alpha _{0},\beta _{0}) \in X\) and the Fréchet-derivative \(D\varPhi \left( x_{0}\right) \) is given by
where \(v_{0}=\mathfrak {G}\left( u_{0},\alpha _{0}\right) \) and \(v=\mathfrak {G} (u,\alpha ) \).
3 Discretization of the abstract form
Our aim is to numerically solve PAF and, in this section, we describe an its quite general discretization. In the following, the positive integer K denotes the level of the discretization: the larger K, the higher is the “quality” of the discretization.
There are two types of discretizations involved in the numerical solution of PAF, that we call secondary discretization and primary discretization.
3.1 The secondary discretization
Consider the BVP (1). In some cases, the values of the functional F cannot be exactly computed. For example, in case of integro-differential equation (4), F involves an integral which has to be replaced with a quadrature rule. Therefore, for any positive integer K, we have to replace F with a suitable functional \(F_{K}:[a,b] \times V\times U\times \mathbb {R}^{d_0}\rightarrow \mathbb {R}^d\), whose values can be exactly computed. If the values of F can be exactly computed, as in case of differential equations with deviating arguments (3), we consider \(F_{K}=F\).
As done for F, we require \(F_K(\cdot ,v,u,\beta ) \in U\), for any positive integer K and \((v,u,\beta )\in V\times U\times \mathbb {R}^{d_{0}}\). Hence, for any positive integer K, we can replace the operator \(\mathcal {F}\) with the operator \(\mathcal {F} _{K}:V\times U\times \mathbb {R} ^{d_0}\rightarrow U\) given by
Analogously, for any positive integer K, we replace the functional B with a suitable functional \(B_{K}:V\times U\times \mathbb {R}^{d_0}\rightarrow \mathbb {R}^{d}\times \mathbb {R}^{d_0}\), whose values can be exactly computed. If the values of B can be exactly computed, as in case of multipoint boundary conditions (9), we consider \(B_K=B\).
The secondary discretization of PAF consists in replacing, for any positive integer K, the operators \(\mathfrak {F}\) and \(\mathfrak {B}\) with operators \(\mathfrak {F}_K:\mathbb {V}\times \mathbb {U}\times \mathbb {B}\rightarrow \mathbb {U}\) and \(\mathfrak {B}_K:\mathbb {V}\times \mathbb {U}\times \mathbb {B}\rightarrow \mathbb {A}\times \mathbb {B}\), respectively, whose values can be exactly computed. As done for \(\mathfrak {F}\) and \(\mathfrak {B}\), we assume what follows.
-
A \(\varvec{\mathfrak {F}_K}\varvec{\mathfrak {B}_K}\). For any positive integer K, the operators \(\mathfrak {F}_K\) and \(\mathfrak {B}_K\) are Fréchet-differentiable at any point \((v_{0},u_{0},\beta _{0})\in \mathbb {V}\times \mathbb {U}\times \mathbb {B}\).
For any positive integer K, the operator \(\varPhi \) is then replaced with the operator \(\varPhi _K:X\rightarrow X\) given by
Since A \(\varvec{\mathfrak {F}_K\mathfrak {B}_K}\) and A \(\varvec{\mathfrak {G}}\) hold, \(\varPhi _K\) is Fréchet-differentiable at any point \( x_{0}=(u_{0},\alpha _{0},\beta _{0})\in X\) and \(D\varPhi _K \left( x_{0}\right) \) is given by
where \(v_{0}=\mathfrak {G}\left( u_{0},\alpha _{0}\right) \) and \(v=\mathfrak {G} (u,\alpha ) \).
3.2 The primary discretization
The primary discretization consists in the discretization of the space X into a finite dimensional space and of the operator \(\varPhi \), actually replaced with \(\varPhi _K\) by the secondary discretization, into an operator acting on this finite dimensional space.
Let K be a positive integer (level of discretization). Given a finite dimensional space \(\widehat{\mathbb {U}}_{K}\) and linear bounded operators \(\pi _{K}:\widehat{\mathbb {U}}_{K}\rightarrow \mathbb {U}\) and \(\rho _{K}:\mathbb {U}\rightarrow \widehat{\mathbb {U}}_{K}\), called prolongation to \(\mathbb {U}\) and restriction to \(\widehat{\mathbb {U}}_{K}\), respectively, we consider the finite-dimensional product space
and the linear bounded operators \(P_K:\widehat{X}_K\rightarrow X\) and \( R_K:X\rightarrow \widehat{X}_K\) defined by
and
Note that if the spaces \(\mathbb {A}\) and \(\mathbb {B}\) were not finite-dimensional, restrictions and prolongations also for these spaces had to be introduced.
The finite-dimensional space \(\widehat{X}_K\) is considered as the discretization of level K of X and the operator
is considered as the discretization of level K of \(\varPhi \). We have
Given a fixed point \(\widehat{x}_K^*=\left( \widehat{u}_{K}^{*},\alpha _K^{*},\beta _K^{*}\right) \in \widehat{X}_K\) of \(\widehat{\varPhi }_K\) , which can be found by a standard numerical method for solving nonlinear systems of algebraic equations, we consider
as an approximation of a fixed point of \(\varPhi \) and
as an approximation of a solution of PAF.
The papers [41–44], for the particular instance of PAF given by a BVP (1), deal with two types of primary discretization falling in the previous abstract general description, namely the collocation method and the Fourier series method.
Remark 1
Note that, unlike the operators \(\mathfrak {F}\) and \(\mathfrak {B}\), we do not replace the linear operator \(\mathfrak {G}\) with an approximation \(\mathfrak {G}_K\), whose values can be exactly computed. The reason for this is that we assume, as it happens for the primary discretizations dealt in [41–44], the possibility to compute exactly \(\mathfrak {G}(u,\alpha )\) for any \(u\in \pi _K(\widehat{\mathbb {U}}_K)\) and \(\alpha \in \mathbb {A}\) (see (17)).
4 Convergence analysis
Let \(x^{*}=(u^{*},\alpha ^{*},\beta ^{*})\) be a fixed point of \(\varPhi \) and let \((v^*,\beta ^*)\), where \(v^*=\mathfrak {G} (u^*,\alpha ^*)\), be the relevant solution of PAF.
We set \(D^*\varPhi :=D\varPhi (x^*)\) and we make the following two assumptions regarding \(x^*\).
-
A \(\varvec{x^*}\) 1. There exist \(r_0>0\) and \(L\ge 0\) such that
$$\begin{aligned} \Vert D\varPhi (x)-D^*\varPhi \Vert \le L\Vert x-x^*\Vert _X,\quad x\in \overline{B} (x^*,r_0). \end{aligned}$$ -
A \(\varvec{x^*}\) 2. The linear bounded operator \(I_X-D^*\varPhi \) is invertible, i.e. for any \((u_0,\alpha _0,\beta _0)\) \(\in X\) the linear problem
$$\begin{aligned} \left\{ \begin{array}{l} u=D^*\mathfrak {F}(\mathfrak {G}(u,\alpha ),u,\beta )+u_0 \\ D^*\mathfrak {B}(\mathfrak {G}(u,\alpha ),u,\beta )=(\alpha _0,\beta _0), \end{array} \right. \end{aligned}$$where \(D^*\mathfrak {F}:=D\mathfrak {F}(v^*,u^*,\beta ^*)\) and \( D^*\mathfrak {B}:=D\mathfrak {B}(v^*,u^*,\beta ^*)\), has a unique solution \( (u,\alpha ,\beta )\in X\).
Observe that A \(\varvec{x^*}\) 2 says that \(x^*\) is a simple zero of \(I_X-\varPhi \) and implies that \(x^*\) is an isolated fixed point of \(\varPhi \).
In this section, we study how \(x^*\) and \((v^*,\beta ^*)\) can be approximated by the approximations (18) and (19), respectively, obtained by some fixed point \(\widehat{x}_K^*\) of \(\widehat{\varPhi }_K\).
Our analysis is based on studying of how \(x^{*}\) is approximated by fixed points of the operator
Unlike \(\widehat{\varPhi }_K=R_K\varPhi _KP_K\), this operator has the advantage to be defined on the space X as \(\varPhi \).
Clearly, the operator \(P_KR_K\varPhi _K\) is Fréchet-differentiable at any point \(x_{0}\in X\) and its Fréchet-derivative at \(x_0\) is \( P_KR_KD\varPhi _K(x_0)\), where \(D\varPhi _K(x_0)\) is given in (16). We set \( D^*\varPhi _K:=D\varPhi _K(x^*)\).
For notational convenience, we also introduce the operator
whose zeros are the fixed points of \(P_KR_K\varPhi _K\). Note that \(\varPsi _K\) is Fréchet-differentiable at any point \(x_0\in X\) and
We set
Since we consider the operator \(P_KR_K\varPhi _K\) as an approximation of the operator \(\varPhi \), it is expected that \(\varPsi _Kx^*\) has a small norm. We call \(\varPsi _Kx^*\) the consistency error.
Now, we introduce the following two stability conditions.
CS1 (Condition Stability 1) There exist \(r_1>0\) and, for any positive integer K, \( L_K\ge 0\) such that
(compare with A \(\varvec{x^*}\) 1).
CS2. There exists a positive integer \(K_{2}\) such that, for any positive integer \(K\ge K_{2}\), \(D^*\varPsi _{K}\) is invertible and
where
with \(r_1\) and \(L_K\) given in CS1.
By using the Lemma 1 on the zeros of Fréchet-differentiable operators given in Appendix, we obtain the next theorem.
Theorem 1
Let CS1 and CS2 hold. Then, there exists a positive integer \(\overline{K}\) such that, for any positive integer \( K\ge \overline{K}\), \(P_KR_K\varPhi _{K}\) has a unique fixed point \(x_K^{*}\) in \(\overline{B}\left( x^{*},r_{2}(K)\right) \) and
holds. Moreover, we have the expansion
where
Here, \(L_K\) is defined in CS1 and \(r_2(K)\) is defined in CS2.
Proof
The proof is an application of the Lemma 1 with \(Y=X\), \( A=\varPsi _K \) and \(y^*=x^*\). Note that \(Y=X\) is a Banach space since \(\mathbb {U}\), \(\mathbb {A}\) and \(\mathbb {B}\) are Banach spaces. For \(K\ge K_2\), where \(K_2\) is defined in CS2, it is immediate to verify that
where the quantity q(r) is defined in Lemma 1. Now, let \( \overline{K}\ge K_2\) be such that, for \(K\ge \overline{K}\),
(recall (22)). For \(K\ge \overline{K}\), we have
and so, since (71) in Lemma 1 is fulfilled for \( r=r_2(K) \), \(\varPsi _K\) has a unique zero \(x_K^{*}\) in \( \overline{B}\left( x^{*},r_{2}(K)\right) \) and (23) holds by (72).
As for the second part of the theorem, for \(K\ge \overline{K}\), take
in the second part of Lemma 1. Since \(r\le r_2(K)\) holds (recall (26)), we have
and
and so the condition (71) is fulfilled. Then, we obtain (24) by (73) and
by (74). Now, since
holds, we have (25). \(\square \)
Next result is a consequence of the Theorem 1 and says how \( x^*\) and \((v^*,\beta ^*)\) can be approximated by (18) and (19), respectively.
Theorem 2
Let CS1 and CS2 hold. Then, there exists a positive integer \(\widehat{K}\) such that, for any positive integer \(K\ge \widehat{K},\) the operator \(\widehat{\varPhi }_K\) has a fixed point \(\widehat{x}_K^*\) and
and
where \(\delta _K\) is defined in (24) and satisfies (25), hold. Moreover, if \(\widehat{x}_K\) is a fixed point of \(\widehat{\varPhi }_K\) different from \(\widehat{x}^*_K\), then
and
Here, \(r_2(K)\) is defined in CS2. Finally, regarding the approximation \((v_K^*,\beta _K^*)\) of \((v^*,\beta ^*)\), we have
Proof
By recalling Theorem 1, for \(K\ge \overline{K}\), let \(x_K^*\) be the unique fixed point of \(P_KR_K\varPhi _K\) in \(\overline{B}(x^*,r_2(K))\) . It is immediate to verify that \(\widehat{x}_K^*=R_K\varPhi _Kx_K^{*}\) is a fixed point of \(\widehat{\varPhi }_K\). Moreover, we have
Therefore, (27) and (28) follow by (23) and (24) in Theorem 1, respectively.
Now, we prove the second part. Let \(\widehat{x}_K\) be a fixed point of \( \widehat{\varPhi }_K\) different from \(\widehat{x}^*_K\). It is immediate to verify that \(P_K\widehat{x}_K\) is a fixed point of \(P_KR_K\varPhi _K\). Since \(x^*_K\) is the unique fixed point of \(P_KR_K\varPhi _K\) in \(\overline{B} \left( x^*,r_2(K)\right) \), we have (29). As for the inequality (30), observe that
Since (22) and (23) hold, we have
Hence, there exists \(\widehat{K}\ge \overline{K}\) such that, for \(K \ge \widehat{K}\), we have
and then
and
Finally, the estimate (31) is obtained by
\(\square \)
In the next subsection, we will give an estimate of the error of the approximation \(( v^*_K,\beta ^*_K)\), better than (31) in some situations.
Remark 2
Regarding the consistency error \(\varPsi _K x^*\), which appears in (27) and (31), we have
where we have separated the contributions of the primary and secondary discretizations. If only a primary discretization is used, i.e. \(\varPhi _K=\varPhi \), then
Remark 3
Suppose there exists a sequence \(\{\widehat{x}_K\}\) of fixed points of \( \widehat{\varPhi }_K\) such that \(\widehat{x}_K\) is eventually different from \(\widehat{x}_K^*\). By (29) and (30), we obtain
and
respectively. Note that, by (33), (27) and (22), we have
The estimates (33)–(35) give informations on how much the fixed point \(\widehat{x}^*_K\) is isolated from other fixed points of \(\widehat{\varPhi }_K\).
4.1 The simple case
Let us introduce the space
and the linear operator \(\varLambda :X\rightarrow Z\) given by
Clearly, the linear operator \(\varLambda \) is bounded and
holds.
In this subsection, we consider the situation where, for any \( x\in X\), we can factorize \(D\varPhi (x) \) as
and, for any positive integer K, \(D\varPhi _K(x)\) as
where \(\varSigma (x),\varSigma _K(x):Z\rightarrow X\) are linear bounded operators. We call this situation the simple case.
In the following, we set \(\varSigma ^*:=\varSigma (x^*)\) and \(\varSigma _K^*:=\varSigma _K(x^*)\).
Note that the simple case holds if \(\mathfrak {F}(v,u,\beta )=\mathfrak {F}(v,\beta )\), \(\mathfrak {F}_K(v,u,\beta )=\mathfrak {F}_K(v,\beta )\), \(\mathfrak {B}(v,u,\beta )=\mathfrak {B}(v,\beta )\) and \(\mathfrak {B}_K(v,u,\beta )=\mathfrak {B}_K(v,\beta )\). In fact, for \(x_0=(u_0,\alpha _0,\beta _0)\in X\), factorizations (36) and (37) hold with \(\varSigma (x_0),\varSigma _K(x_0):Z\rightarrow X\) given by
where \(D^0\mathfrak {F}:=D\mathfrak {F}(v_0,\beta _0)\), \(D^0\mathfrak {F}_K:=D \mathfrak {F}_K(v_0,\beta _0)\), \(D^0 \mathfrak {B}:=D\mathfrak {B}(v_0,\beta _0)\) and \(D^0 \mathfrak {B}_K\) \(:=D\mathfrak {B}_K(v_0,\beta _0)\), with \(v_0=\mathfrak {G}(u_0,\alpha _0)\).
Therefore, the simple case holds for the particular instance of PAF given by a non-neutral BVP (1). As it is shown in [43], the simple case can hold also in case of BVP for neutral functional differential equations. In particular, it holds in case of BVPs for neutral integro-differential equation (4) and non-neutral boundary conditions.
Now, we present two theorems for the simple case. The first result is a condition under which the invertibility of the linear bounded operator \(D^*\varPsi _K\), and the uniform boundedness with respect to K of the norm of its inverse, are guaranteed. We recall that \(D^*\varPsi _K\) is defined in (20) and the norm of its inverse appears in CS2 and in the error estimates of Theorem 2.
Theorem 3
Assume the simple case. If
then there exists a positive integer \(K_{2}\) such that, for any positive integer \(K\ge K_{2}\), \(D^*\varPsi _{K}\) is invertible and
Note that the previous theorem implicitly requires the invertibility of \(I_X-D^*\varPhi \), which is assumed in A \(\varvec{x^*}\) 2.
Proof
By recalling (20), we have
The theorem now follows by an application of the Banach perturbation lemma. \(\square \)
Note that in (38) we can take advantage of the fact that the error operator \( P_KR_K\varSigma _{K}^*-\varSigma ^*\) is applied to elements that have been regularized by the operator \(\varLambda \).
Remark 4
By separating the contributions of the primary and secondary discretizations, we have that (38) holds if
and
The second result for the simple case is an estimate of the error of \(\left( v^{*}_K,\beta ^{*}_K\right) \) different from (31).
Theorem 4
Let CS1 and CS2 hold. Assume the simple case. We have
where
and \(\Vert \delta _K\Vert _X\) satisfies (25).
Proof
Of course, all that is stated in Theorem 2 holds. Now, consider the expansion (28):
where we set
We have
From (20), we obtain
and then
Now, since \(D^*\varPhi _K=\varSigma _K^*\varLambda \) holds, we have
with \(\varXi _K\) defined in (40), and the estimate (39) follows by (41). \(\square \)
This result is indeed useful since it can happen, as it is illustrated in [41] in case of the collocation method (version finite element method), that \(\Vert \varLambda \varPsi _{K}x^{*}\Vert _{Z}\) has an order of convergence to zero, as \(K\rightarrow \infty \), higher than \(\Vert \varPsi _{K}x^{*}\Vert _{X}\).
Remark 5
Regarding the regularized consistency error \(\varLambda \varPsi _Kx^*\), we have
where we have separated the contributions of the primary and secondary discretizations. If only a primary discretization is used, i.e. \(\varPhi _K=\varPhi \), then
4.2 Invertibility of \(D^*\varPhi _K\)
In the previous subsection, in Theorem 3, it has been presented a condition under which the invertibility of \(D^*\varPhi \) is guaranteed in the simple case. In this subsection, we study the invertibility of \(D^*\varPhi \) in the general case.
We consider a splitting
of \(D^*\varPhi \), where \(\varGamma ^*:X\rightarrow X\) and \(\varSigma ^*:Z \rightarrow X\) are linear bounded operators. (Recall that Z and \(\varLambda \) have been introduced at the beginning of the previous subsection). Similarly, for any positive integer K, we consider a splitting
of \(D^*\varPhi _K\), where \(\varGamma _K^*:X\rightarrow X\) and \( \varSigma _K^*:Z\rightarrow X\) are linear bounded operators.
Note that in the simple case, described in the previous subsection, we have splittings (42) and (43) with \(\varGamma ^*=\varGamma _K^*=0\).
In this subsection, by using splittings (42) and (43), we give a theorem concerning the invertibility of \(D^*\varPsi _K\) and the norm of its inverse. This theorem is an extension of the Theorem 3 (which is valid only for the simple case) and it is based on the Lemma 2 in Appendix.
Theorem 5
Assume that there exist a splitting (42) such that \(I_X-\varGamma ^*\) is invertible and, for any positive integer K , a splitting (43) such that \(I_X-P_KR_K\varGamma _K^*\) is invertible.
If
and
then there exists a positive integer \(K_{2}\) such that, for any positive integer \(K\ge K_{2}\), \(D^*\varPsi _{K}\) is invertible and
Proof
The proof is an application of the Lemma 2 with \(Y=X\), \(A=I_X-D^*\varPhi \) (recall A \(\varvec{x^*}\) 2), \(B=I_X-\varGamma ^*\), \(C=-\varSigma ^*\varLambda \) and, for any positive integer K, \(A_K=D^*\varPsi _K=I_X-P_KR_KD^*\varPhi _K\), \(B_K=I_X-P_KR_K\varGamma _K^*\) and \( C_K=-P_KR_K\varSigma _K^*\varLambda \). \(\square \)
The previous theorem reduces the invertibility of \(D^*\varPsi _{K}=I_X-P_KR_KD^*\varPhi _K\) to the invertibility of \(I_X-P_KR_K\varGamma _K^*\). Note that in (44) and (45), we can take advantage of the fact that the error operators \(P_KR_K\varGamma _{K}^*-\varGamma ^*\) and \( P_KR_K\varSigma _{K}^*-\varSigma ^*\) are applied to elements that have been regularized by \(\varLambda \).
Remark 6
By separating the contributions of the primary and secondary discretizations, we have that (44) holds if
and
and (45) holds if
and
4.3 The nilpotency case and the splitting case
In view of Theorem 5, it remains to study the invertibility of \(I_X-P_KR_K\varGamma _K^*\). To this aim, we consider the situation where there exist a positive integer c, a splitting (42) such that \((\varGamma ^*)^c=0\) and, for any positive integer K, a splitting (43) such that \((P_KR_K\varGamma _K^ *)^c=0\). We call this situation the nilpotency case.
Remark 7
If the nilpotency case holds, then \(I_X-\varGamma ^*\) is invertible and, for any positive integer K, \(I_X-P_KR_K\varGamma _K^*\) is invertible and
holds.
On the other hand, we call the splitting case the more general situation (including the nilpotency case) described in the premise in Theorem 5, namely there exist a splittings (42) such that \(I_X-\varGamma ^*\) is invertible and, for any positive integer K , a splitting (43) such that \(I_X-P_KR_K\varGamma _K^*\) is invertible.
4.4 The operators \(P_KR_K\) and \(P_KR_K-I_X\)
Regarding the primary discretization, the previous subsections have shown that the role played by the linear operators \(P_{K}R_{K}:X\rightarrow X\) and \(P_{K}R_{K}-I_{X}:X\rightarrow X\) given by
and
is crucial. In this subsection, we list some simple facts about them to be used in the next section.
We have
and, for a linear bounded operator \(A:X\rightarrow X\),
where \(A_\mathbb {U}\) is the \(\mathbb {U}\)-component of A defined in Sect. 1.2.
Moreover, note that
and
where \(x^*=(u^*,\alpha ^*,\beta ^*)\) is the fixed point of \(\varPhi \) and
can be called the consistency error of the primary discretization (see Remark 2).
5 Specialization of the convergence results
In the convergence analysis presented above, we have considered the general situation where, beside a primary discretization, also a secondary discretization is introduced. This means that approximations \(\mathfrak {F}_K\) of \(\mathfrak {F}\) and \(\mathfrak {B}_K\) of \(\mathfrak {B}\) are used and then the operator \(\varPhi \) is actually replaced by \(\varPhi _K\). However, in the papers [41, 44], where the results of this paper are specialized to the problem (1) for two particular primary discretizations, we do not consider a secondary discretization, in order to avoid giving results with too many assumptions and details.
As previously remarked, in case of the problem (1), approximations \(\mathfrak {F}_K=\mathcal {F}_K\) of \(\mathfrak {F}=\mathcal {F}\) and \(\mathfrak {B}_K=B_K\) of \(\mathfrak {B}=B\) are used for integro-differential equation (4) and integral boundary conditions (10), respectively, where the involved integrals are approximated by quadrature rules. Convergence results, when quadrature rules are used in integro-differential equations BVPs, can be deduced from the general theory given above and they are addressed in [43]. We also remark that the choice of considering the exact computation of integrals in integro-differential equations BVPs is adopted in the papers [21, 22, 25, 27].
The case of BVPs (1) for differential equations with deviating arguments (3) and multipoint boundary conditions (9), where a secondary discretization is not necessary, is dealt in [42].
In this section, for the situation where only a primary discretization is used (i.e., for any positive integer K, we have \(\mathfrak {F}_K=\mathfrak {F} \) and \(\mathfrak {B}_K=\mathfrak {B}\) and then \(\varPhi _K=\varPhi \)), we give two convergence theorems for the problem PAF, less abstract than Theorem 2. The first is for the simple case and the second is for the splitting case, which includes the nilpotency case. Such theorems are used in [41, 44], in case of the problem (1), for the two particular primary discretizations given by the collocation method and the Fourier series method.
As already remarked, the simple case holds for non-neutral BVPs (1) and for BVPs given by neutral integro-differential equation (4) and non-neutral boundary conditions. Moreover, for the collocation method and the Fourier series method, the nilpotency case holds for BVPs given by neutral differential equations with deviating arguments (3) and non-neutral boundary conditions, whenever the neutral deviating arguments \(\vartheta _s\), \(s=1,\ldots ,l\), are such that
or
for some \(\tau >0\). This is shown in [42, 44].
Below, since we are considering only a primary discretization, we have that:
-
the simple case reduces to the sole factorization (36);
-
the splitting case uses the sole splitting (42) and requires the invertibility of \(I_X-\varGamma ^*\) and, for any positive integer K, of \(I_X-P_KR_K\varGamma ^*\);
-
the nilpotency case uses the sole splitting (42) and requires \(\left( \varGamma ^*\right) ^c=0\) and, for any positive integer K, \((P_KR_K\varGamma ^*)^c=0\), for some positive integer c.
Moreover, we use diffusely the notation \(A_\mathbb {U}\) of the \(\mathbb {U}\)-component of an operator A introduced in Sect. 1.2. Finally, we remark that the quantities \(\lambda _K\) and \(\Vert e_K^*\Vert _\mathbb {U}\) (see (49)–(51)) play a crucial role. In particular, we have
(see Remark 2).
5.1 The simple case
Here is the theorem for the simple case. Before to present it, we introduce the following condition, which is formulated only for the simple case.
-
CSC (Condition Simple Case) There exist \(r_{2}>0\) and, for any positive integer K, \(\sigma _K\ge 0\) such that
$$\begin{aligned} \Vert (\pi _K\rho _K-I_\mathbb {U})(D\varPhi (x)-D^*\varPhi )_{\mathbb {U}}\Vert&=\Vert (\pi _K\rho _K-I_\mathbb {U})((\varSigma (x)-\varSigma ^*)\varLambda )_\mathbb {U}\Vert \\&\le \sigma _K\Vert x-x^*\Vert _X,\quad x\in \overline{B}(x^*,r_2), \end{aligned}$$and
$$\begin{aligned} \sigma _K=O(1),\quad K\rightarrow \infty . \end{aligned}$$
Theorem 6
Assume that only a primary discretization is used. Moreover, assume the simple case,
and
(One has to read the lower row after \(\{,\) instead of the upper one, if CSC holds). Then, there exists a positive integer \(\widehat{K}\) such that, for any positive integer \(K\ge \widehat{K}\), \(\widehat{\varPhi }_K\) has a fixed point \(\widehat{x}_K^*\) such that
Moreover, for the approximation \((v_K^*,\beta _K^*)\) of \((v^*,\beta ^*)\), we have the two estimates
and
Finally, suppose there exists a sequence \(\{\widehat{x}_K\}\) of fixed points of \( \widehat{\varPhi }_K\) such that \(\widehat{x}_K\) is eventually different from \(\widehat{x}^*_K\). Then
and
Note that, in CSC and (53), the error operator \(\pi _K\rho _K-I_\mathbb {U}\) is applied to elements regularized by means of \(\varLambda \).
Proof
Since A \(\varvec{x^*}\) 1 holds, CS1 is fulfilled with a constant
Now, we show that CS2 holds. By (53), (48), Remark 4 and Theorem 3, we obtain that there exists a positive integer \(K_{2}\) such that, for any positive integer \(K\ge K_{2}\), \( D^*\varPsi _{K}\) is invertible and
Now, we have
By (61), (60), (52) and (54), we conclude that CS2 is fulfilled.
Then, Theorem 2 says that there exists a positive integer \( \widehat{K}\) such that, for any positive integer \(K\ge \widehat{K}\), \( \widehat{\varPhi }_K\) has a fixed point \(\widehat{x}_K^*\) such that (55) and (56) hold: see (60) and (52). Moreover, by Remark 3, we obtain (58) and (59): see (61).
It remains to prove (57). Since
(see Remark 5, (47) and (51)) and
(see (40) with \(\varSigma ^*_K=\varSigma ^*\)) and
(see (25)), we obtain (57) by Theorem 4. \(\square \)
5.2 The splitting case
Now, we give the theorem for the splitting case. If the splitting case holds, we set, for any positive integer K,
If the nilpotency case holds, then, by Remark 7, we have
Theorem 7
Assume that only a primary discretization is used. Moreover, assume the splitting case,
and
Then, there exists a positive integer \(\widehat{K}\) such that, for any positive integer \(K\ge \widehat{K}\), \(\widehat{\varPhi }_K\) has a fixed point \( \widehat{x}_K^*\) such that
Moreover, for the approximation \((v_K^*,\beta _K^*)\) of \((v^*,\beta ^*)\), we have the estimate
Finally, suppose there exists a sequence \(\{\widehat{x}_K\}\) of fixed points of \(\widehat{\varPhi }_K\) such that \(\widehat{x}_K\) is eventually different from \(\widehat{x}^*_K\). Then
and
Proof
Since A \(\varvec{x^*}\) 1 holds, CS1 is fulfilled with
Now, we show that CS2 holds. By (62), (63), (48), Remark 6 and Theorem 5, we obtain that there exists a positive integer \(K_{2}\) such that, for any positive integer \(K\ge K_{2}\), \(D^*\varPsi _{K}\) is invertible and
Now, we have
By (70), (69), (52) and (64), we conclude that \(\mathbf {CS2}\) is fulfilled.
Then, Theorem 2 says that there exists a positive integer \( \widehat{K}\) such that, for any positive integer \(K\ge \widehat{K}\), \( \widehat{\varPhi }_K\) has a fixed point \(\widehat{x}_K^*\) such that (65) and (66) hold: see (69) and (52). Moreover, by Remark 3 we obtain (67) and (68): see (70). \(\square \)
6 Conclusions
In this paper we have studied the numerical solution of PAF, introduced in Sect. 2, in case of spaces \(\mathbb {A}\) and \(\mathbb {B}\) of finite dimension. PAF has been discretized by a primary discretization and a secondary discretization, as explained in Sect. 3. A convergence analysis has been carried out in Sect. 4. In Sect. 4, we have also addressed the two particular situations of the simple case and the splitting case (which includes the nilpotency case). In Sect. 5, under the assumption that only a primary discretization is used, the convergence results have been specialized to these two particular situations.
The functional differential equation BVP (1) is a particular instance of PAF. The results of Sect. 5 are applied in [41, 44] to this particular instance for the two particular primary discretizations given by the collocation method and the Fourier series method. The present paper provides the theoretical basis for the analysis of such methods.
We finish observing that PAF also includes BVPs for partial functional differential equations, as it has been illustrated in Sect. 2.1. Apart from the possible infinite-dimensionality of the space \(\mathbb {A}\), a numerical study of such problems in the context of PAF has to take into account the use of approximations \(\mathfrak {G}_K\) of the linear operator \(\mathfrak {G}\).
References
Abell, K.A., Elmer, C.E., Humpries, A.R., Van Vleck, E.K.: Computation of mixed type functional differential boundary value problems. SIAM J. Appl. Dyn. Syst. 4(3), 755–781 (2005)
Azbelev, N.V., Maksimov, V.K., Rakhmatullina L.F.: Introduction to the theory of functional differential equations: methods and applications. In: Contemporary Mathematics and its Applications, vol 3. Hindawi Publishing Corporation (2007)
Bader, G.: Solving boundary value problems for functional differential equations by collocation. In: Ascher, U.M., Russel, R.D. (eds.) Numerical Boundary Value ODEs, Progress in Scientific Computing, vol. 5, pp. 227–243. Birkhauser, Boston (1985)
Bakke, V.L., Jackiewicz, Z.: The numerical solution of boundary value problems for differential equations with state-dependent delays. Appl. Math. 34, 1–17 (1989)
Barton, D.A.W., Krauskopf, B., Wilson, R.E.: Collocation schemes for periodic solutions of neutral delay differential equations. J. Differ. Equ. Appl. 12(11), 1087–1101 (2006)
Bellen, A.: The collocation method for the numerical approximation of the periodic solution of functional differential equations. Computing 23, 55–66 (1978)
Bellen, A.: Monotone methods for periodic solutions of second order scalar functional differential equations. Numer. Math. 42, 15–30 (1983)
Bellen, A.: A Runge–Kutta–Nystrom method for delay differential equations. In: Ascher, U.M., Russel, R.D. (eds.) Numerical Boundary Value ODEs, Progress in Scientific Computing, vol. 5, pp. 271–283. Birkhauser, Boston (1985)
Bellen, A., Zennaro, M.: A collocation method for boundary value problems of differential equations with functional arguments. Computing 32, 307–318 (1984)
Burkowski, F.J., Cowan, D.D.: The numerical derivation of a periodic solution of a second order differential difference equation. SIAM J. Numer. Anal. 10, 489–495 (1973)
Chi, H., Bell, J., Hassard, B.: Numerical solution of a nonlinear advance-delay-differential equation from nerve conduction theory. J. Math. Biol. 24, 583–601 (1986)
Chocholaty, P., Slahor, L.: A numerical method to boundary value problems for second order delay differential equations. Numer. Math. 33, 69–75 (1979)
Cryer, C. W.: The numerical solution of boundary value problems for second order functional differential equations by finite differences. Numer. Math. 20, 288–299 (1972/73)
De Luca, J., Humphries, T., Rodrigues, S.B.: Finite element boundary value integration of Wheeler–Feynman electrodynamics. J. Comput. Appl. Math. 236, 3319–3337 (2012)
Engelborghs, K., Luzyanina, T., Hout, K.J.I., Roose, D.: Collocation methods for the computation of periodic solutions of delay differential equations. SIAM J. Sci. Comput. 22(5), 1593–1609 (2000)
Engelborghs, K., Luzyanina, T., Roose, D.: Numerical bifurcation analysis of delay differential equations. J. Comput. Appl. Math. 125, 265–275 (2000)
Engelborghs, K., Doedel, E.J.: Convergence of a boundary value difference equation for computing periodic solutions of neutral delay differential equations. J. Differ. Equ. Appl. 7, 927–940 (2001)
Engelborghs, K., Doedel, E.J.: Stability of piecewise polynomial collocation for computing periodic solution of delay differential equations. Numer. Math. 91, 627–648 (2002)
Ford, N.J., Lumb, P.M.: Mixed-type functional differential equations: a numerical approach. J. Comput. Appl. Math. 229, 471–479 (2009)
Ford, N.J., Lumb, P., Lima, P.M., Teodoro, M.F.: The numerical solution of forward-backward differential equations: decomposition and related issues. J. Comput. Appl. Math. 234, 2745–2756 (2010)
Ganesh, M., Spence, A.: Orthogonal collocation for a nonlinear integro-differential equation. IMA J. Numer. Anal. 18, 191–206 (1998)
Ganesh, M., Sloan, I.H.: Optimal order spline methods for nonlinear differential and integro-differential equations. Appl. Numer. Math. 29, 445–478 (1999)
Garey, L.E., Gladwin, C.J.: Numerical methods for second order Volterra integro-differential equations with two point boundary conditions. Util. Math. 35, 103–109 (1989)
Garey, L.E., Shaw, R.E.: Solving VIDEs with two-point boundary values and naturally auxiliary conditions. Int. Math. J. 2(5), 433–443 (2002)
Hangelbroek, R.J., Kaper, H.G., Leaf, G.K.: Collocation methods for integro-differential equations. SIAM J. Numer. Anal. 14(3), 377–390 (1977)
Henderson, J. (ed.): Boundary Value Problems for Functional-Differential Equations. World Scientific Publishing Co., Inc., River Edge (1995)
Hu, Q.: Interpolation correction for collocation solutions of Fredholm integro-differential equations. Math. Comput. 67, 987–999 (1998)
Jankowski, T.: Convergence of multistep methods for retarded functional differential equations with parameters. Appl. Math. 37(1–4), 227–251 (1990)
Jankowski, T.: Numerical solution of boundary value problems for retarded functional differential equations with a parameter. Nihonkai Math. J. 6(2), 115–128 (1995)
Jankowski, T.: Boundary value problem for systems of functional differential equations. Appl. Math. 47(5), 427–458 (2002)
Kadalbajoo, M.K., Sharma, K.K.: Numerical analysis of boundary value problems for singularly-perturbed differential-difference equations with small-shifts of mixed type. J. Optim. Theory Appl. 115(1), 145–163 (2002)
Kadalbajoo, M.K., Sharma, K.K.: Numerical analysis of singularly perturbed delay differential equations with layer behavior. Appl. Math. Comput. 157, 11–28 (2004)
Kadalbajoo, M.K., Sharma, K.K.: Numerical treatment of boundary value problems for second order singularly perturbed delay differential equations. Comput. Appl. Math. 24(2), 151–172 (2005)
Kadalbajoo, M.K., Sharma, K.K.: Parameter-uniform fitted mesh method for singularly perturbed delay differential equations with layer behavior. Electron. Trans. Numer. Anal. 23, 180–201 (2006)
Kadalbajoo, M.K., Sharma, K.K.: An exponentially fitted finite difference scheme for solving boundary-value problems for singularly-perturbed differential-difference equations: small shifts of mixed type with layer behavior. J. Comput. Anal. Appl. 8, 151–171 (2006)
Krasnosel’skii, M.A., Vainikko, G.M., Zabreiko, P.P., Rutitskii, Y.B., Stetsenko, V.Y.: Approximate Solution of Operator Equations. Walters-Noordhoof publishing, Groningen (1972)
Kurihara, M., Suzuki, T.: Chebyshev approximation for a boundary value problem of differential difference equations. Nonlinear Anal. 47, 3839–3847 (2001)
Liu, X.: Periodic boundary value problems for differential equations with finite delay. Dyn. Syst. Appl. 3(3), 357–367 (1994)
Luzyanina, T., Engelborghs, K., Lust, K., Roose, D.: Computation, continuation and bifurcation analysis of periodic solutions of delay differential equations. Int. J. Bifurc. Chaos 7(11), 2547–2560 (1997)
Luzyanina, T., Engelborghs, K., Roose, D.: Numerical bifurcation analysis of differential equations with state-dependent delay. Int. J. Bifurc. Chaos 11(3), 737–753 (2001)
Maset, S.: The collocation method in the numerical solution of BVPs for neutral functional differential equations. Part I: convergence results (in revision)
Maset, S.: The collocation method in the numerical solution of BVPs for neutral functional differential equations. Part II: differential equations with deviating arguments (in revision)
Maset, S.: The collocation method in the numerical solution of BVPs for integro-differential equations (in preparation)
Maset, S.: The Fourier series method in the numerical solution of BVPs for neutral functional differential equations (in preparation)
Mohsen, A., El-Gamel, M.: A sinc-collocation method for linear Fredholm integro-differential equations. Zeitschrift für Angewandte Mathematik und Physik 58(3), 380–390 (2007)
Parts, I., Pedars, A., Tamme, E.: Piecewise polynomial collocation for Fredholm integro-differential equations with weakly singular kernels. SIAM J. Numer. Anal. 43(5), 1897–1911 (2005)
Pedars, A., Tamme, E.: Spline collocation method for integro-differential equations with weakly singular kernels. J. Comput. Appl. Math. 197(1), 253–269 (2006)
Pedars, A., Tamme, E.: Discrete Galerkin method for Fredholm integro-differential equations with weakly singular kernels. J. Comput. Appl. Math. 213(1), 111–126 (2008)
de Nevers, K., Schmitt, K.: An application of shooting method to boundary value problem for second order delay equations. J. Math. Anal. Appl. 36, 588–597 (1971)
Reddien, G.W., Travis, C.C.: Approximation methods for boundary value problems of differential equations with functional arguments. J. Math. Anal. Appl. 46, 62–74 (1974)
Sakai, M.: Numerical solution of boundary value problems for second order functional differential equations by the use of cubic splines. Mem. Fac. Sci. Kyushu Univ. Ser. A 29(1), 113–122 (1975)
Shinohara, Y., Fujimori, H., Suzuki, T., Kurihara, M.: On a boundary value problem for delay differential equations of population dynamics and Chebyshev approximations. J. Comput. Appl. Math. 201, 348–355 (2007)
Teodoro, F., Lima, P.M., Ford, N.J., Lumb, P.: New approach to the numerical solution of forward–backward equations. Front. Math. China 4, 155–168 (2009)
Verheyden, K., Lust, K.: A Newton-Picard collocation method for periodic solution of delay differential equations. BIT Numer. Math. 45, 605–625 (2005)
Volk, W.: The numerical solution of linear integro-differential equations by projection methods. J. Integral Equ. 9(1), 171–190 (1985)
Volk, W.: The iterated Galerkin method for linear integro-differential equations. J. Comput. Appl. Math. 21(1), 63–74 (1988)
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
Lemma 1
Let Y be a Banach space with norm \(\left\| \ \cdot \ \right\| _{Y}\), let \(A:\Omega \subseteq Y\rightarrow Y\), where \(\Omega \) is open, be a Fréchet-differentiable operator and let \(y^{*}\in \Omega \) such that \(DA\left( y^{*}\right) \) is invertible. For any \(r>0\) such that \( \overline{B}\left( y^{*},r\right) \subseteq \Omega \), define
Now, let \(r>0\) be such that \(\overline{B}\left( y^{*},r\right) \subseteq \Omega \) . If
then A has a unique zero \(\overline{y}^{*}\) in \(\overline{B}\left( y^{*},r\right) \) and
Moreover, we have
where
The proof of the first part is more or less similar to the proof of [36, Lemma 19.1, page 293]. The proof of the second part (73) is clear once the proof of first part is understood.
Lemma 2
Let Y be a Banach space with norm \(\left\| \ \cdot \ \right\| _{Y}\). Let \(A,B,C:Y\rightarrow Y\) be linear bounded operators such that \(A=B+C\) and B is invertible. Let \(\left\{ A_{K}\right\} \), \( \left\{ B_{K}\right\} \) and \(\left\{ C_{K}\right\} \) be sequences of linear bounded operators \(Y\rightarrow Y\) such that, for any positive integer K, \(A_{K}=B_{K}+C_{K} \) and \(B_{K}\) is invertible.
If A is invertible,
and
then there exists a positive integer \(K_{2}\) such that, for any positive integer \(K\ge K_{2}\), \( A_{K}\) is invertible and
Proof
Assume that A is invertible and (75) and (76) hold. For any positive integer K, we have
and then \(A_{K}\) is invertible if \(I_{Y}+B_{K}^{-1}C_{K}\) is invertible. In this case, we have
Now, since
and
is invertible with inverse
the thesis follows by the Banach perturbation Lemma. \(\square \)
Rights and permissions
About this article
Cite this article
Maset, S. An abstract framework in the numerical solution of boundary value problems for neutral functional differential equations. Numer. Math. 133, 525–555 (2016). https://doi.org/10.1007/s00211-015-0754-1
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00211-015-0754-1