Variable-directions method

An iterative method (cf. Iteration methods) for solving systems of linear or non-linear equations arising in difference or projection-difference methods for the approximate solution of, for example, boundary value problems for partial differential equations of elliptic type.

Let, for example, there be two spatial variables and a sequence of square grids with step and nodes , where is a vector with integer components. Let be the set of nodes at which one seeks a solution to a difference or projection-difference problem written in the form of the operator equation

in the Euclidean space . The latter can be identified with the space of functions given at the nodes of ; the dimension of coincides with the number of points from .

Let

(1)

be linear operators mapping into . Among the operators (1) there are such for which the non-zero coefficients in (1) correspond only to shift vectors having . Such operators are called one-dimensional, acting with respect to , and they are denoted by ; similarly, for shift vectors having , one defines the one-dimensional operators acting on .

The systems of equations

split up into individual subsystems, each of which links only the values of at nodes lying on separate horizontal (for ) or vertical (for ) lines in the grid. The method is characterized by the use of intertwining operators for of the form

The solution to the system

(2)

then amounts to the successive solution of the two systems

(3)

(4)

in which one first solves the separate subsystems on the horizontal lines in the grid (in the case (3)) and then varies the directions and solves the subsystems on the vertical lines (in the case (4)). Usually, the operators are taken such that only arithmetic operations are involved in solving (3) and (4), and consequently also (2). Therefore, each iteration in the method, of the form

(5)

requires arithmetic operations, where the superscript corresponds to the number of the iteration.

The most effective results are obtained for the commutative case, where is a self-adjoint positive-definite operator, while the operators are self-adjoint and commute with . In that case, for any the error in the initial approximation can be reduced in norm by a factor in iterations. The commutative case will be encountered only for boundary value problems in which the variables can be separated, and therefore, the region should be a rectangle. The most common case of the method (5) for the equation

is the method

where is the identity operator.

One also uses approaches based on different variational principles, supplementing the approach in which the iteration parameters are chosen to minimize the norm of the operator for passing from the zero-th iteration to an iteration of a given index.

The method is frequently used as an internal iterative process in two-step iteration methods based on operators that are equivalent in spectrum and suitable for use with variable coefficients and non-linear problems.

References

[1]	D.W. Peaceman, H.H. Rachford, "The numerical solution of parabolic and elliptic differential equations" SIAM J. , 3 : 1 (1955) pp. 28–41
[2]	E.G. D'yakonov, "Iterative methods for solving difference analogues of boundary value problems for equations of elliptic type" , Kiev (1970) (In Russian)
[3]	N.N. Yanenko, "The method of fractional steps; the solution of problems of mathematical physics in several variables" , Springer (1971) (Translated from Russian)
[4]	A.A. Samarskii, "Theorie der Differenzverfahren" , Akad. Verlagsgesell. Geest u. Portig K.-D. (1984) (Translated from Russian)
[5]	G.I. Marchuk, "Methods of numerical mathematics" , Springer (1982) (Translated from Russian)