Traditional methods of solving equations of mathematical physics, such as finite element method, are to find a solution in the entire domain of definition. Corresponding methods of computational error analysis consist in defining the norm for difference between the exact and approximate solutions also in the whole domain. Meanwhile, in a number of applications, researchers are interested not in the solution as a whole, but only in its certain values or functionals named as goal-oriented or output ones. For example, in air flow around the body, engineers are interested in lift and drag, not in the solution at every point in space, especially away from the constructed body [1, 2]. In such cases, one would be interested in the calculation precision of these output functionals, and not of the entire solution. Moreover, with appropriate triangulation in the finite element method one can achieve a significant increase in the accuracy of the required functionals without increasing the computational time for the problem as a whole [2, 3].
One of the interesting tasks consists in computation of the solution values along the part of boundary with Neumann boundary condition. Such situation occurs when the flux (heat flux, for instance) at body surface equals to zero, and we need to know the resulting temperature at this surface.
In this article we consider a one-dimensional model problem with Neumann boundary condition at one of boundary points and try to find the solution value in the same point. For solving we use cubic Hermitian finite element method. Unexpectedly the ultrahigh order of convergence was obtained, which is not guaranteed by the finite element theory. Therefore the paper is devoted to the theoretical justification of this beneficial effect with the help of adjoint operators.
II. Formulations of problem
Functions are smooth enough for further considerations. Here and after any constant is supposed to be positive and independent of meshsizes Introduce the functional space where last equality is valid in the sense of traces. We use the standard notation for functional spaces that can be found for instance in [4, 5].
In the weak form [4, 11] this problem is reformulated as follows:
As it was mentioned above, in many real problems of mathematical physics the main important result is not the solution as a whole, but some its functionals. In this paper for above formulated problem we take the output functional
as an example and show that it is computed with better order of accuracy than the solution as a whole.
3. Finite element method with cubic Hermitian elements
In accordance with handbooks [4, 11] introduce cubic basis functions of first type:
and second type:
Due to Galerkin approach we get the following formulation:
with the vector of unknowns
where sign (…)T means transposition. This vector is found from the system of linear algebraic equations [6, 10]
This property is well-known as Galerkin orthogonality property.
The accuracy of approximate solution is known from handbooks [4, 11]:
Therefore we can guarantee only
4. A numerical result
Now consider the following concrete problem:
And then solve it by cubic finite element method that has been introduced before.
Numerical result for problem (11).
‒ 9.11 × 10-10
‒1.75 × 10-11
‒3.03 × 10-13
Now let us prove it for more general problems.
5. The proof of the higher convergence order for this type of output functional
For smooth coefficients this weak formulation is equivalent to the following problem:
Its sufficiently smooth coefficients imply a sufficiently smooth solution. Like for the original problem (4), Galerkin approach in this case gives the following weak formulation:
Therefore, we get the discrete system for adjoint problem
defines the approximate solution
Theorem. For sufficiently smooth data of problem (1)-(2) we get
Proof. From (12), (7), (13) we get
By standard way we get inequality
Since both problems have sufficiently smooth solutions, then
It implies that
Thus, we indeed get the sixth order of convergence for this output functional.
By virtue of the duality theory we prove the higher order of convergence than that is following from standard theory of Hermitian finite element method.
Note, that this effect became possible for cubic finite elements. For linear ones we should not get higher order of accuracy. Indeed, repeating this way of justification we should get only
It is not unusual for this type of elements for which we have typical estimate
Shaydurov V., Liu T., Zheng Z.: Four-stage computational technology with adaptive numerical methods for computational aerodynamics. // American Institute of Physics. Conference Proceedings. 2012. – Vol. 1484. – P. 42-48.
ADIGMA.A European Initiative on the Development of Adaptive Higher-Order Variational Methods for Aerospace Applications. // Notes on Numerical Fluid Mechanics and Multidisciplinary Design. Vol. 113. - Berlin: Springer, 2010. – 498 p.
Bangerth W., Rannacher R.: Adaptive Finite Element Methods for Differential Equations. – Berlin: Birkhäuser, 2003. – 204 p.
Ciarlet P.G.: The Finite Element Method for Elliptic Problems. – New York: North Holland, 1978. – 530 p.
Сьярле Ф.: Метод конечных элементов для эллиптических задач. – М.: Мир, 1980. – 512 с.
Даутов Р.З., Карчевский М.М.: Ведение в теорию метода конечных элементов. – Казань: Казанский госуниверситет. – 2004. – 239 с.
Стренг Г., Фикс Дж.: Теория метода конечных элементов. – М.: Мир. – 1977. – 351 с.
Шайдуров В.В.: Многосеточные методы конечных элементов. – М.: Наука. – 1989. – 288 с.
Shaidurov V.: Multigrid Methods for Finite Elements. – Berlin: Springer. – 1995. – 300 p.
Strang G., Fix G.J.: An Analysis of the Finite Element Method. – New York: Prentice-Hall. – 1973. – 306 p.
Zhiming Chen, Haijun Wu: Selected Topics in Finite Element Methods. – Beijing: Science Press. – 2010. – 162 p.