Costica Mustata
“Tiberiu Popoviciu” Institute of Numerical Analysis, Romanian Academy, Romania
Adrian Muresan
“Tiberiu Popoviciu” Institute of Numerical Analysis, Romanian Academy, Romania
Radu Mustata
Babes-Bolyai University, Romania
Keywords
Paper coordinates
C. Mustăţa, A. Mureşan and R. Mustăţa, The approximation by spline functions of the solutions of a singularly perturbed bilocal problem, Rev. Anal. Numér. Théor. Approx. 27 (1998).
[1] N. Adžié, Domain decomposition for spectral approximation of lhe layer solution, Univ. u Novom Sadu, Zb. Rad. Prirod.-Mat. Fak., Ser. Mat. 24, I (1994), pp. 347-357.
[2] P.Blaga and Gh.Micula, Polynomial natural spline of even degree, Studia Univ.”Babeş-Bolyai”, Mathematica 38, 2 (1993), pp. 31-40.
[3] W. Heinrichs, A stabilized Multidomain approach for singular perturbations Problems, Journal of Scientific Computing 7 (1992), https://doi.org/10.1007/bf01059944
[4] P. W. Hemker, A Numerical Study of Stiff Two-Point Boundary Problems, Mathematical Centre Tracts 80, Amsterdam, 1977.
[5] H. B. Keller, Numerical Methods for Two-point Boundary-value Problems, Blaisdell Publ. Comp., 1968.
[6] C. Mustăţa, Approximation by spline functions of the solution of a linear bilocal problem, Rev. Anal. Numér. Theor. approx. 26 (1997), 1-2, pp. 137-148.
Paper (preprint) in HTML form
1998-Mustata-The approximation by spline functions of the solution-Jnaat
THE APPROXIMATION BY SPLINE FUNCTIONS OF THE SOLUTION OF A SINGULARLY PERTURBED BILOCAL PROBLEM
C. MUSTÃTA, A. C. MUREŞAN AND R. MUSTĂTA
The singularly perturbed bilocal problems admit exact solutions having both slowly and rapidly varying parts. There are thin transition layers where the solution can jump abruptly, having as effect strong oscillations of the approximate solutions obtained by the method of centered differences, spectral methods, etc.
We define a class of spline functions of degree 5 which are appropriate for these problems and obtain sufficiently good approximate solutions, with attenuated oscillations on the subintervals where the exact solution jumps rapidly.
Let n >= 3,n inNn \geq 3, n \in \mathbb{N}, and let
a division of the real axis.
Denote by O_(s)(Delta_(n))\mathscr{O}_{s}\left(\Delta_{n}\right) the set of functions s:RrarrRs: \mathbb{R} \rightarrow \mathbb{R} verifying the conditions:
where P_(m)\mathscr{P}_{m} denotes the set of polynomials of degree mm.
As concerns the behavior of functions in this class, one can prove
THEOREM 1. Every function s inP_(5)(Delta_(n))s \in \mathscr{P}_{5}\left(\Delta_{n}\right) can be written in the form
s(t)=sum_(i=0)^(3)A_(i)t^(i)+sum_(k=0)^(n)a_(k)(t-t_(k))_(+)^(5),quad t inR,s(t)=\sum_{i=0}^{3} A_{i} t^{i}+\sum_{k=0}^{n} a_{k}\left(t-t_{k}\right)_{+}^{5}, \quad t \in \mathbb{R},
and (t-t_(k))_(+)={[0," if "t < t_(k)","],[t-t_(k)," if "t >= t_(k)","k=0","1","dots","n]:}\left(t-t_{k}\right)_{+}= \begin{cases}0 & \text { if } t<t_{k}, \\ t-t_{k} & \text { if } t \geq t_{k}, k=0,1, \ldots, n\end{cases}
Proof. Let s inS_(5)(Delta_(n))s \in \mathscr{S}_{5}\left(\Delta_{n}\right). By definition s^((4))(t)=0s^{(4)}(t)=0 for all t >= bt \geq b so that s^((4))(t)=sum_(k=0)^(n)a_(k)(t-t_(k))=0s^{(4)}(t)=\sum_{k=0}^{n} a_{k}\left(t-t_{k}\right)=0, for all t >= bt \geq b, showing that sum_(k=0)^(n)a_(k)=0\sum_{k=0}^{n} a_{k}=0 and sum_(k=0)^(n)a_(k)t_(k)=0\sum_{k=0}^{n} a_{k} t_{k}=0.
THEOREM 2. Let f:RrarrRf: \mathbb{R} \rightarrow \mathbb{R} be such that
where t_(k)(t_(0)=1,t_(n)=b),k= bar(0,n)t_{k}\left(t_{0}=1, t_{n}=b\right), k=\overline{0, n}, are the knots of the division Delta_(n)\Delta_{n} and alpha,beta,lambda_(k)\alpha, \beta, \lambda_{k}, k= bar(0,n)k=\overline{0, n} are given numbers.
Then there exists a unique function s_(f)inS_(5)(Delta_(n))s_{f} \in \mathscr{S}_{5}\left(\Delta_{n}\right) such that
of n+5n+5 equations with n+5n+5 unknowns A_(0),A_(1),A_(2),A_(3),a_(0),a_(1),dotsa_(n)A_{0}, A_{1}, A_{2}, A_{3}, a_{0}, a_{1}, \ldots a_{n}.
The system (6) has a unique solution if and only if the corresponding homogeneous system (obtained for alpha=beta=0,lambda_(j)=0,j= bar(0,n)\alpha=\beta=0, \lambda_{j}=0, j=\overline{0, n} ) has only the null solution.
If s inP_(5)(Delta_(n))s \in \mathscr{P}_{5}\left(\Delta_{n}\right) verifies the homogeneous conditions (5) (i.e., with alpha=beta=0\alpha=\beta=0, lambda_(k)=0,j= bar(0,n)\lambda_{k}=0, j=\overline{0, n} ), then
where c_(k)=s^((5))(t)|_(I_(k)),k= bar(1,n)c_{k}=\left.s^{(5)}(t)\right|_{I_{k}}, k=\overline{1, n}.
It follows that s^((4))(t)=0s^{(4)}(t)=0 for all t in[a,b]t \in[a, b]. Since the restrictions of ss to the intervals I_(0),I_(n+1)I_{0}, I_{n+1} are in P_(3)\mathscr{P}_{3} and s inC^(4)(R)s \in C^{4}(\mathbb{R}), it results that s^((4))(t)=0s^{(4)}(t)=0 for all t inRt \in \mathbb{R}. Therefore s^('')inPs^{\prime \prime} \in \mathscr{P} and, taking into account the equalities s^('')(t_(k))=0s^{\prime \prime}\left(t_{k}\right)=0, k=0,n,n >= 3k=0, n, n \geq 3 (verified by hypothesis), one obtains s^('')(t)=0s^{\prime \prime}(t)=0 for all t inRt \in \mathbb{R}. Since s(a)=s(b)=0s(a)=s(b)=0, it follows s(t)=0s(t)=0 for all t inRt \in \mathbb{R}, which is equivalent to A_(0)=A_(1)=A_(2)=A_(3)=0A_{0}=A_{1}=A_{2}=A_{3}=0 and a_(k)=0,k= bar(0,n)a_{k}=0, k=\overline{0, n}.
COROLLARY 3. a) There exists a system B={s_(0),s_(1),S_(0),S_(1),dots,S_(n)}\mathscr{B}=\left\{s_{0}, s_{1}, S_{0}, S_{1}, \ldots, S_{n}\right\} of functions in S_(5)(Delta_(n))\mathscr{S}_{5}\left(\Delta_{n}\right) verifying the conditions:
b) If f:RrarrRf: \mathbb{R} \rightarrow \mathbb{R} verifies the conditions (4) and s_(f)inO_(5)(Delta_(n))s_{f} \in \mathscr{O}_{5}\left(\Delta_{n}\right) verifies the conditions (5) then
{:(8)s_(f)(t)=s_(0)(t)*f(a)+s_(1)(t)*f(b)+sum_(k=0)^(n)s_(k)(t)*f^('')(t_(k))","quad t inR:}\begin{equation*}
s_{f}(t)=s_{0}(t) \cdot f(a)+s_{1}(t) \cdot f(b)+\sum_{k=0}^{n} s_{k}(t) \cdot f^{\prime \prime}\left(t_{k}\right), \quad t \in \mathbb{R} \tag{8}
\end{equation*}
Remark 1. By Corollary 3, O_(5)(Delta_(n))\mathscr{O}_{5}\left(\Delta_{n}\right) is a real linear space of dimension n+3n+3, and the system B\mathscr{B} is a basis in O_(5)(Delta_(n))\mathscr{O}_{5}\left(\Delta_{n}\right).
THEOREM 4. If s inS_(5)(Delta_(n))nnW_(2,f,D)^(4)(Delta_(n))s \in \mathscr{S}_{5}\left(\Delta_{n}\right) \cap W_{2, f, D}^{4}\left(\Delta_{n}\right) and f inW_(2)^(4)(Delta_(n))f \in W_{2}^{4}\left(\Delta_{n}\right), then
" a) "||s_(f)^((4))||_(2) <= ||g^((4))||_(2)" for all "g inW_(2,f,D)^(4)(Delta_(n))"; "\text { a) }\left\|s_{f}^{(4)}\right\|_{2} \leq\left\|g^{(4)}\right\|_{2} \text { for all } g \in W_{2, f, D}^{4}\left(\Delta_{n}\right) \text {; }
b) ||s_(f)^((4))-f^((4))||_(2) <= ||s^((4))-f^((4))||_(2)\left\|s_{f}^{(4)}-f^{(4)}\right\|_{2} \leq\left\|s^{(4)}-f^{(4)}\right\|_{2} for all s inS_(S)(Delta_(n))s \in \mathscr{S}_{S}\left(\Delta_{n}\right).
COROLLARY 5. If f inW_(2)^(4)(Delta_(n))f \in W_{2}^{4}\left(\Delta_{n}\right) and s_(f)inS_(5)(Delta_(n))s_{f} \in \mathscr{S}_{5}\left(\Delta_{n}\right) is given by (8), then the following relations hold:
1201 Proof. Since (15) holds for every s inS_(5)(Delta_(n))s \in \mathscr{S}_{5}\left(\Delta_{n}\right), one obtains (16) by taxing s-=0s \equiv 0 in (15).
The inequalities (17) and (18) are immediate consequences of the equality
Remark 2. 1^(@)1^{\circ} The property expressed by the inequality (12) is called the minimum norm property. 2^(@)2^{\circ} The property exprimed by the inequality (13) is called the best approximation property.
APPLICATION
Consider the singularly perturbed bilocal problem
(D)
{:[epsiy^('')=f(t,y,y^('))","t in[a","b]","epsi > 0],[y(a)=alpha","y(b)=beta.]:}\begin{gathered}
\varepsilon y^{\prime \prime}=f\left(t, y, y^{\prime}\right), t \in[a, b], \varepsilon>0 \\
y(a)=\alpha, y(b)=\beta .
\end{gathered}
One supposes that the problem (D) has a unique solution.
THEOREM 6. If the exact solution yy of the problem (D)(D) belongs to W_(2)^(4)(Delta_(n))W_{2}^{4}\left(\Delta_{n}\right) and s_(y)inS_(5)(Delta_(n))s_{y} \in \mathscr{S}_{5}\left(\Delta_{n}\right) is the function given by ( 8 ), then
Proof. Since y^('')(t_(i))-s_(y)^('')(t_(i))=0,i= bar(0,n)y^{\prime \prime}\left(t_{i}\right)-s_{y}^{\prime \prime}\left(t_{i}\right)=0, i=\overline{0, n}, by Rolle's theorem there exist t_(i)^((1))in(t_(i),t_(i+1)),i= bar(0,n-1)t_{i}^{(1)} \in\left(t_{i}, t_{i+1}\right), i=\overline{0, n-1} such that
Applying again Rolle's theorem, it follows the existence of the points t_(i)^((2))in(t_(i)^((1)),t_(i+1)^((1))),i= bar(0,n-2)t_{i}^{(2)} \in\left(t_{i}^{(1)}, t_{i+1}^{(1)}\right), i=\overline{0, n-2} such that
Since for every t in[a,b]t \in[a, b] there exists i_(0)in{0,1,dots,n-1}i_{0} \in\{0,1, \ldots, n-1\} such that |t-t_(i_(0))^((1))| <= 2||Delta_(n)||\left|t-t_{i_{0}}^{(1)}\right| \leq 2\left\|\Delta_{n}\right\|, one obtains:
Similarly, for every t in[a,b]t \in[a, b], there exists j_(0)in{0,1,dots,n-2}j_{0} \in\{0,1, \ldots, n-2\} such that |t-t_(j_(0))| <= ||Delta_(n)||\left|t-t_{j_{0}}\right| \leq\left\|\Delta_{n}\right\|, implying
showing that (19) holds for k=2k=2.
Taking into account the equalities y(a)-s_(y)(a)=0y(a)-s_{y}(a)=0 and y(b)-s_(y)(b)=0y(b)-s_{y}(b)=0, it follows (by Rolle's theorem) the existence of a point c in(a,b)c \in(a, b) such that y^(')(c)-s_(y)^(')(c)=0y^{\prime}(c)-s_{y}^{\prime}(c)=0. Then, for every t in[a,b]t \in[a, b] one has
This choice is motivated by the fact that the parameter epsi > 0\varepsilon>0 is multiplied by y^('')y^{\prime \prime} and s_(y)s_{y} is determined by the interpolation conditions on the second derivative of yy on the knots of Delta_(n)\Delta_{n}, which give epsiy^('')(t_(i))=epsis_(y)^('')(t_(i)),i= bar(0,n)\varepsilon y^{\prime \prime}\left(t_{i}\right)=\varepsilon s_{y}^{\prime \prime}\left(t_{i}\right), i=\overline{0, n}.
Proposition 8. If the real-valued function f(t,u,v)f(t, u, v) defined on D sub[a,b]xxR^(2)D \subset[a, b] \times \mathbb{R}^{2} has continuous partial derivatives (del f)/(del u),(del f)/(del v)\frac{\partial f}{\partial u}, \frac{\partial f}{\partial v}, then the unknowns u_(i),u_(i)^('),i= bar(0,n)u_{i}, u_{i}^{\prime}, i=\overline{0, n}, can be obtained from the system
have the order O(||Delta_(n)||^(3//2))O\left(\left\|\Delta_{n}\right\|^{3 / 2}\right). qquad\qquad
Eliminating E_(i),E_(i)^('),i= bar(0,n)E_{i}, E_{i}^{\prime}, i=\overline{0, n}, one obtains the system (22). ◻\square
A NUMERICAL EXAMPLE
We consider the following singularly perturbed problem:
which may be regarded as a linearized one-dimensional version of a convectiondominated flow problem.
This problem has a unique solution y(x)=(e^((x+1)/(epsi))-e^((2)/(epsi)))/(1-e^((2)/(epsi)))y(x)=\frac{e^{\frac{x+1}{\varepsilon}}-e^{\frac{2}{\varepsilon}}}{1-e^{\frac{2}{\varepsilon}}}, which displays one boundary layer at the point x=1x=1, of the length O(epsi)O(\varepsilon).
Considering the solution y_(r)(x)y_{r}(x) of the reduced problem
{[y_(r)^(')(x)-0","x in[-1","1]],[y_(r)(-1)=1","]:}\left\{\begin{array}{l}
y_{r}^{\prime}(x)-0, x \in[-1,1] \\
y_{r}(-1)=1,
\end{array}\right.
the following estimations holds (see [6]):
|y(x)-y_(r)(x)| <= C(epsi+e^((x-1)/(epsi))),x in[-1,1]\left|y(x)-y_{r}(x)\right| \leq C\left(\varepsilon+e^{\frac{x-1}{\varepsilon}}\right), x \in[-1,1]
where CC denotes an arbitrary constant independent of xx and epsi\varepsilon.
The exact solution is of the form
y(x)=y_(r)(x)+u(x),y(x)=y_{r}(x)+u(x),
where u(x)u(x) will be approximated by
u(x)~~{[0",",x in[-1","1-p epsi]],[v(x)",",x in[1-p epsi","1]","]:}u(x) \approx \begin{cases}0, & x \in[-1,1-p \varepsilon] \\ v(x), & x \in[1-p \varepsilon, 1],\end{cases}
and the function nu(x)\nu(x) is the solution of the following problem:
Thus, we approximate the solution y(x)y(x) by the solution y_(r)(x)y_{r}(x) on the domain [-1,1-p epsi][-1,1-p \varepsilon], so that the error would be O(epsi)O(\varepsilon). In this way we obtain p=ln((1)/(epsi))p=\ln \frac{1}{\varepsilon}. We approximate the solution y(x)y(x) upon the domain [1-p epsi,1][1-p \varepsilon, 1] by y_(r)(x)+v(x)y_{r}(x)+v(x).
We use elements from the space of spline functions O_(5)(Delta_(n))\mathscr{O}_{5}\left(\Delta_{n}\right) in order to approximate v(x)v(x).
In Table 1 we present the error of approximation of solution y(x)y(x) by y_(r)(x)+s_(v)(x)y_{r}(x)+s_{v}(x) upon the domain [1-p epsi,1\varepsilon, 1 ] for different values of epsi\varepsilon and nn. The linear system for determining the spline s_(v)s_{v} is solved by using a direct method.
In Fig. 1 there are displayed the exact solution y(x)y(x) for epsi=10^(-3)\varepsilon=10^{-3} denoted by a continuous curve, and the approximated solution for n=3n=3 denoted by a dotted curve.
ACKNOWLEDGEMENT. The authors would like to thank Dr. C. I. Gheorghiu for several interesting discussions.
REFERENCES
N. Adžié, Domain decomposition for spectral approximation of the layer solution, Univ. u Novom Sadu, Zb. Rad. Prirod.-Mat. Fak., Ser. Mat. 24, 1 (1994), 347-357.
P. Blaga and Gh. Micula, Polynomial natural spline of even degree, Studia Univ. "Babeş-Bolyai", Mathematica 38, 2 (1993), 31-40.
W. Heinrichs, A stabilized Multidomain approach for singular perturbations Problems, Journal of Scientific Computing 7 (1992), 95-125.
P. W. Hemker, A Numerical Study of Stiff Two-Point Boundary Problems, Mathematical Centre Tracts 80, Amsterdam, 1977.
H. B. Keller, Numerical Methods for Two-point Boundary-value Problems, Blaisdell Publ. Comp., 1968.
C. Mustăta, Approximation by spline functions of the solution of a linear bilocal problem, Rev. Anal. Numér. Theor. Approx. 26 (1997) 1-2, 137-148.
Received July 15, 1996.
"Popoviciu" Institute of Numerical Analysis
01-84 P.O. Box 68
3400 Cluj-Napoca
E-mail.tmustata@math.ubbcluj.ro
Received July 15, 1996. "Popoviciu" Institute of Numerical Analysis
01-84 P.O. Box 68
3400 Cluj-Napoca
E-mail.tmustata@math.ubbcluj.ro
| Received July 15, 1996. | "Popoviciu" Institute of Numerical Analysis |
| :--- | :--- |
| | 01-84 P.O. Box 68 |
| | 3400 Cluj-Napoca |
| | E-mail.tmustata@math.ubbcluj.ro |
| | |
| | |