This paper is concerned with solving nonlinear singularly perturbed boundary value problems. Robust monotone iterates for solving nonlinear difference scheme are constructed. Uniform convergence of the monotone methods is investigated, and convergence rates are estimated. Numerical experiments complement the theoretical results.
1. Introduction
We are interested in numerical solving of two nonlinear singularly perturbed problems of elliptic and parabolic types.
The first one is the elliptic problem
where is a positive parameter, and is sufficiently smooth function. For this problem is singularly perturbed, and the solution has boundary layers near and (see [1] for details).
The second problem is the onedimensional parabolic problem
where is a positive parameter. Under suitable continuity and compatibility conditions on the data, a unique solution of this problem exists. For problem (1.2) is singularly perturbed and has boundary layers near the lateral boundary of (see [2] for details).
In the study of numerical methods for nonlinear singularly perturbed problems, the two major points to be developed are: (i) constructing robust difference schemes (this means that unlike classical schemes, the error does not increase to infinity, but rather remains bounded, as the small parameter approaches zero); (ii) obtaining reliable and efficient computing algorithms for solving nonlinear discrete problems.
Our goal is to construct a uniform numerical method for solving problem (1.1), that is, a numerical method which generates uniformly convergent numerical approximations to the solution. We use a numerical method based on the classical difference scheme and the piecewise uniform mesh of Shishkintype [3]. For solving problem (1.2), we use the implicit difference scheme based on the piecewise uniform mesh in the direction, which converges uniformly [4].
A major point about the nonlinear difference schemes is to obtain reliable and efficient computational methods for computing the solution. The reliability of iterative techniques for solving nonlinear difference schemes can be essentially improved by using componentwise monotone globally convergent iterations. Such methods can be controlled every time. A fruitful method for the treatment of these nonlinear schemes is the method of upper and lower solutions and its associated monotone iterations [5]. Since an initial iteration in the monotone iterative method is either an upper or lower solution, which can be constructed directly from the difference equation without any knowledge of the exact solution, this method simplifies the search for the initial iteration as is often required in the Newton method. In the context of solving systems of nonlinear equations, the monotone iterative method belongs to the class of methods based on convergence under partial ordering (see [5, Chapter 13] for details).
The purpose of this paper is to construct uniformly convergent monotone iterative methods for solving uniformly convergent nonlinear difference schemes.
The structure of the paper is as follows. In Section 2, we prove that the classical difference scheme on the piecewise uniform mesh converges uniformly to the solution of problem (1.1). A robust monotone iterative method for solving the nonlinear difference scheme is constructed. In Section 3, we construct a robust monotone iterative method for solving problem (1.2). In the final Section 4, we present numerical experiments which complement the theoretical results.
2. The Elliptic Problem
The following lemma from [1] contains necessary estimates of the solution to (1.1).
Lemma 2.1.
If is the solution to (1.1), the following estimates hold:
where constant is independent of .
For , the boundary layers appear near and .
2.1. The Nonlinear Difference Scheme
Introduce a nonuniform mesh
For solving (1.1), we use the classical difference scheme
where and . We introduce the linear version of this problem
where . We now formulate a discrete maximum principle for the difference operator and give an estimate of the solution to (2.4).
Lemma 2.2.
(i) If a mesh function satisfies the conditions
then , .
(ii) If , then the following estimate of the solution to (2.4) holds true:
where , .
The proof of the lemma can be found in [6].
2.2. Uniform Convergence on the Piecewise Uniform Mesh
We employ a layeradapted mesh of a piecewise uniform type [3]. The piecewise uniform mesh is formed in the following manner. We divide the interval into three parts , and . Assuming that is divisible by , in the parts , we use the uniform mesh with mesh points, and in the part the uniform mesh with mesh points is in use. The transition points and are determined by
This defines the piecewise uniform mesh. If the parameter is small enough, then the uniform mesh inside of the boundary layers with the step size is fine, and the uniform mesh outside of the boundary layers with the step size is coarse, such that
In the following theorem, we give the convergence property of the difference scheme (2.3).
Theorem 2.3.
The difference scheme (2.3) on the piecewise uniform mesh (2.8) converges uniformly to the solution of (1.1):
where constant is independent of and .
Proof.
Using Green's function of the differential operator on , we represent the exact solution in the form
where the local Green function is given by
and are defined by
Equating the derivatives and , we get the following integraldifference formula:
where here and below we suppress variable in . Representing on and in the forms
the above integraldifference formula can be written as
where the truncation error of the exact solution to (1.1) is defined by
From here, it follows that
From Lemma 2.1, the following estimate on holds:
We estimate the truncation error in (2.17) on the interval . Consider the following three cases: , and . If , then , and taking into account that in (2.18), we have
where here and throughout denotes a generic constant that is independent of and . If , then , . Taking into account that , and , we have
If , then , and we have
Thus,
In a similar way we can estimate on and conclude that
From here and (2.8), we conclude that
From (2.3), (2.15), by the meanvalue theorem, we conclude that satisfies the difference problem
Using the assumption on from (1.1) and (2.24), by (2.6), we prove the theorem.
2.3. The Monotone Iterative Method
In this section, we construct an iterative method for solving the nonlinear difference scheme (2.3) which possesses monotone convergence.
Additionally, we assume that from (1.1) satisfies the twosided constraint
The iterative method is constructed in the following way. Choose an initial mesh function , then the iterative sequence , , is defined by the recurrence formulae
where is the residual of the difference scheme (2.3) on .
We say that is an upper solution of (2.3) if it satisfies the inequalities
Similarly, is called a lower solution if it satisfies the reversed inequalities. Upper and lower solutions satisfy the inequality
Indeed, by the definition of lower and upper solutions and the meanvalue theorem, for we have
where . In view of the maximum principle in Lemma 2.2, we conclude the required inequality.
The following theorem gives the monotone property of the iterative method (2.27).
Theorem 2.4.
Let , be upper and lower solutions of (2.3) and satisfy (2.26). Then the upper sequence generated by (2.27) converges monotonically from above to the unique solution of (2.3), the lower sequence generated by (2.27) converges monotonically from below to :
and the sequences converge at the linear rate .
Proof.
We consider only the case of the upper sequence. If is an upper solution, then from (2.27) we conclude that
From Lemma 2.2, by the maximum principle for the difference operator , it follows that , . Using the meanvalue theorem and the equation for , we represent in the form
where , . Since the mesh function is nonpositive on and taking into account (2.26), we conclude that is an upper solution. By induction on , we obtain that , , , and prove that is a monotonically decreasing sequence of upper solutions.
We now prove that the monotone sequence converges to the solution of (2.3). Similar to (2.33), we obtain
and from (2.27), it follows that satisfies the difference equation
Using (2.26) and (2.6), we have
This proves the convergence of the upper sequence at the linear rate . Now by linearity of the operator and the continuity of , we have also from (2.27) that the mesh function defined by
is the exact solution to (2.3). The uniqueness of the solution to (2.3) follows from estimate (2.6). Indeed, if by contradiction, we assume that there exist two solutions and to (2.3), then by the meanvalue theorem, the difference satisfies the difference problem
By (2.6), which leads to the uniqueness of the solution to (2.3). This proves the theorem.
Consider the following approach for constructing initial upper and lower solutions and . Introduce the difference problems
where from (2.26). Then the functions , are upper and lower solutions, respectively. We check only that is an upper solution. From the maximum principle in Lemma 2.2, it follows that on . Now using the difference equation for and the meanvalue theorem, we have
Since and is nonnegative, we conclude that is an upper solution.
Theorem 2.5.
If the initial upper or lower solution is chosen in the form of (2.39), then the monotone iterative method (2.27) converges uniformly to the solution of the nonlinear difference scheme (2.3)
Proof.
From (2.27), (2.39), and the meanvalue theorem, by (2.6),
From here and estimating from (2.39) by (2.6),
we conclude the estimate on in the form
where is defined in the theorem. From here and (2.36), we conclude that
Using this estimate, we have
Taking into account that as , where is the solution to (2.3), we conclude the theorem.
From Theorems 2.3 and 2.5 we conclude the following theorem.
Theorem 2.6.
Suppose that the initial upper or lower solution is chosen in the form of (2.39). Then the monotone iterative method (2.27) on the piecewise uniform mesh (2.8) converges uniformly to the solution of problem (1.1):
where and constant is independent of and .
3. The Parabolic Problem
3.1. The Nonlinear Difference Scheme
Introduce uniform mesh on
For approximation of problem (1.2), we use the implicit difference scheme
where and are defined in (2.2) and (2.3), respectively. We introduce the linear version of problem (3.2)
We now formulate a discrete maximum principle for the difference operator and give an estimate of the solution to (3.3).
Lemma 3.1.
(i) If a mesh function on a time level satisfies the conditions
then , .
(ii) If , then the following estimate of the solution to (3.3) holds true:
where , .
The proof of the lemma can be found in [6].
3.2. The Monotone Iterative Method
Assume that from (3.2) satisfies the twosided constraint
We consider the following iterative method for solving (3.2). Choose an initial mesh function . On each time level, the iterative sequence , , is defined by the recurrence formulae
where is the residual of the difference scheme (3.2) on .
On a time level , we say that is an upper solution of (3.2) with respect to if it satisfies the inequalities
Similarly, is called a lower solution if it satisfies all the reversed inequalities. Upper and lower solutions satisfy the inequality
This result can be proved in a similar way as for the elliptic problem.
The following theorem gives the monotone property of the iterative method (3.7).
Theorem 3.2.
Assume that satisfies (3.6). Let be given and , be upper and lower solutions of (3.2) corresponding . Then the upper sequence generated by (3.7) converges monotonically from above to the unique solution of the problem
the lower sequence generated by (3.7) converges monotonically from below to and the following inequalities hold
Proof.
We consider only the case of the upper sequence, and the case of the lower sequence can be proved in a similar way.
If is an upper solution, then from (3.7) we conclude that
From Lemma 3.1, it follows that
and from (3.7), it follows that satisfies the boundary conditions.
Using the meanvalue theorem and the equation for from (3.7), we represent in the form
where , . Since the mesh function is nonpositive on and taking into account (3.6), we conclude that is an upper solution to (3.2). By induction on , we obtain that , , , and prove that is a monotonically decreasing sequence of upper solutions.
We now prove that the monotone sequence converges to the solution of (3.2). The sequence is monotonically decreasing and bounded below by , where is any lower solution (3.9). Now by linearity of the operator and the continuity of , we have also from (3.7) that the mesh function defined by
is an exact solution to (3.2). If by contradiction, we assume that there exist two solutions and to (3.2), then by the meanvalue theorem, the difference satisfies the system
By Lemma 3.1, which leads to the uniqueness of the solution to (3.2). This proves the theorem.
Consider the following approach for constructing initial upper and lower solutions and . Introduce the difference problems
The functions , are upper and lower solutions, respectively. This result can be proved in a similar way as for the elliptic problem.
Theorem 3.3.
Let initial upper or lower solution be chosen in the form of (3.17), and let satisfy (3.6). Suppose that on each time level the number of iterates . Then for the monotone iterative methods (3.7), the following estimate on convergence rate holds:
where is the solution to (3.2), , and constant is independent of , and .
Proof.
Similar to (3.14), using the meanvalue theorem and the equation for from (3.7), we have
From here and (3.7), we have
Using (3.5) and (3.6), we have
where is defined in (3.18).
Introduce the notation
where . Using the meanvalue theorem, from (3.2) and (3.19), we conclude that satisfies the problem
where , and we have taken into account that . By (3.5), (3.6), and (3.21),
Using (3.6), (3.17), and the meanvalue theorem, estimate from (3.7) by (3.5),
where is independent of (), and . Thus,
Similarly, from (3.2) and (3.19), it follows that
Using (3.21), by (3.5),
Using (3.17), estimate from (3.7) by (3.5),
where . As follows from Theorem 3.2, the monotone sequences and are bounded from above and below by, respectively, and . Applying (3.5) to problem (3.17) at , we have
where constant is independent of , and . Thus, we prove that is independent of , and . From (3.26) and (3.28), we conclude
By induction on , we prove
where all constants are independent of , and . Taking into account that , we prove the estimate (3.18) with .
In [4], we prove that the difference scheme (3.2) on the piecewise uniform mesh (2.8) converges uniformly to the solution of problem (1.2):
where is the exact solution to (3.2), and constant is independent of , and . From here and Theorem 3.3, we conclude the following theorem.
Theorem 3.4.
Suppose that on each time level the initial upper or lower solution is chosen in the form of (3.17) and . Then the monotone iterative method (3.7) on the piecewise uniform mesh (2.8) converges uniformly to the solution of problem (1.2):
where , and constant is independent of , and .
4. Numerical Experiments
It is found that in all numerical experiments the basic feature of monotone convergence of the upper and lower sequences is observed. In fact, the monotone property of the sequences holds at every mesh point in the domain. This is, of course, to be expected from the analytical consideration.
4.1. The Elliptic Problem
Consider problem (1.1) with . We mention that is the solution of the reduced problem, where . This problem gives , , and initial lower and upper solutions are chosen in the form of (2.39). The stopping criterion for the monotone iterative method (2.27) is
Our numerical experiments show that for and , iteration counts for monotone method (2.27) on the piecewise uniform mesh are independent of and , and equals 12 and 8 for the lower and upper sequences, respectively. These numerical results confirm our theoretical results stated in Theorem 2.5.
In Table 1, we present numbers of iterations for solving the test problem by the Newton iterative method with the initial iterations , . Here is in use, and we denote by an "*" if more than 100 iterations is needed to satisfy the stopping criterion, or if the method diverges. The numerical results indicate that the Newton method cannot be used successfully for this test problem.
Table 1. Numbers of iterations for the Newton iterative method.
4.2. The Parabolic Problem
For the parabolic problem (1.2), we consider the test problem with and . This problem gives , , and the initial lower and upper solutions are chosen in the form of (3.17).
The stopping test for the monotone method (3.7) is defined by
Our numerical experiments show that for and , on each time level the number of iterations for monotone method (3.7) on the piecewise uniform mesh is independent of and and equal 4, 4, and 3 for , respectively. These numerical results confirm our theoretical results stated in Theorem 3.3.
References

Boglaev, I: Approximate solution of a nonlinear boundary value problem with a small parameter for the highestorder differential. USSR Computational Mathematics and Mathematical Physics. 24(6), 30–35 (1984). Publisher Full Text

Boglaev, I: Numerical method for quasilinear parabolic equation with boundary layer. USSR Computational Mathematics and Mathematical Physics. 30, 716–726 (1990)

Miller, JJH, O'Riordan, E, Shishkin, GI: Fitted Numerical Methods for Singular Perturbation Problems: Error Estimates in the Maximum Norm for Linear Problems in One and Two Dimensions,p. xiv+166. World Scientific, Singapore (1996)

Boglaev, I, Hardy, M: Uniform convergence of a weighted average scheme for a nonlinear reactiondiffusion problem. Journal of Computational and Applied Mathematics. 200(2), 705–721 (2007). Publisher Full Text

Ortega, JM, Rheinboldt, WC: Iterative Solution of Nonlinear Equations in Several Variables,p. xx+572. Academic Press, New York, NY, USA (1970)

Samarskii, AA: The Theory of Difference Schemes, Monographs and Textbooks in Pure and Applied Mathematics,p. xviii+761. Marcel Dekker, New York, NY, USA (2001)