_{1}

^{*}

Explicit Exact and Approximate Inverse Preconditioners for solving complex linear systems are introduced. A class of general iterative methods of second order is presented and the selection of iterative parameters is discussed. The second order iterative methods behave quite similar to first order methods and the development of efficient preconditioners for solving the original linear system is a decisive factor for making the second order iterative methods superior to the first order iterative methods. Adaptive preconditioned Conjugate Gradient methods using explicit approximate preconditioners for solving efficiently large sparse systems of algebraic equations are also presented. The generalized Approximate Inverse Matrix techniques can be efficiently used in conjunction with explicit iterative schemes leading to effective composite semi-direct solution methods for solving large linear systems of algebraic equations.

During the last decades, considerable research effort has been directed to the solution of complex linear and nonlinear systems of algebraic equation by using a class of iterative methods. This class includes the conjugate gradient method and its hybrid multi-variants. The conjugate gradient method originally introduced by Hestenes and Stiefel [

Certain theoretical issues on this subject, such as 1) the stability and conditions of correctness of approximate system matrix in the form of product of triangular matrix factors, and 2) the convergence analysis of the iterative methods and the quantitative evaluation of the convergence rate of the iterative schemes 4, are of particular interest for the choice of the most efficient methods for systems with predetermined properties.

In the framework of this research work and in order to substantiate the main motivation for the efficient usage of the second order iterative method for solving nonlinear systems, the following basic related questions will be considered: 1) are the second order iterative methods comparable (or even superior) to first order iterative methods for solving nonlinear systems? A survey of related research work will be given; 2) are second order iterative schemes preferable than the first order iterative schemes for solving very large complex computational problems? The computational complexity per iterative step will be also examined.

The structure of this research paper is as follows: in Section 2, several advantages of second order iterative methods in comparison with the first order iterative methods for solving nonlinear systems are presented. In Section 3, a class of general iterative methods of second order is described, while in Section 4, certain explicit iterative schemes and approximate inverse preconditioners are introduced. In Section 5, exact and approximate inverse matrix algorithmic techniques are introduced, while in Section 6, some aspects of stability and correctness of incomplete factorization methods are presented. Finally, in Section 7, the convergence analysis and quantitative evaluation of convergence rate of incomplete factorization methods are discussed.

In recent years, considerable research effort has been focused in the topic of second order iterative methods. In order to substantiate our motivation in our research study, we present a synoptic survey of related computational methods developed on the subject. Several early iterative methods of second order with fixed parameters or variable parameters have been extensively studied [

A class of (optimal) second degree iterative methods for accelerating basic linear stationary methods of first degree with real eigenvalues has been presented [

A three-step iterative method for solving nonlinear equations by using Steffensen’s method in the third step having eight order convergence has been recently presented. This method requires a small number of calculations and does not require calculation of derivative in the third iterative step. A two-step iterative method for solving nonlinear equations, a modified Noor’s method without computing the second derivatives and with fourth order convergence has been presented [

x n + 1 * = G x n + K 1 (1)

and the resulting second order iterative method

x n + 1 = x n + a ( x n − x n − 1 ) + b ( x n + 1 * − x n ) , (2)

where x n + 1 * denotes the first-order iteration, and a and b are real accelerating parameters effecting the convergence rate. For the consistency and convergence of the second order iterative methods the following statements hold:

Preposition 1: If the 1st-order iterative method converges to the exact solution, then the 2nd-order method will converge to the same solution for any values of a ≠ 0 and b ≠ 0 .

Preposition 2: The iterative matrix of the 2nd-order method is known. A necessary and sufficient condition that the iterative method converges for all initial conditions is that, if the spectral radius of matrix G is minimized then the convergence rate is maximized [

It is known that due to the extremely large sizes of power grids, IR drop analysis has become a computationally challenging problem in terms of memory usage and runtime, and second-order iterative algorithms that can significantly reduce the runtime have been proposed. Specifically, the main problems include the following:

1) Very large-scale simulations (millions of elements in power grids) and run time is slow.

2) Memory inefficiency (1 million nodes and trillion elements in matrix)

3) Trade-off between runtime and predetermined accuracy

4) Power delivery issues (increased complexity of VLSI circuits, increased power (current) consumption, decreasing supply voltage, reduced noise margin, increased gate delay)

5) Modelling and analysis of power grid network must be accurate and power grid networks tend to be very large [

Typical applications include also a large class of initial-boundary value problems of general form in 3 space dimensions with strong nonlinearities:

u t + ∑ i = 1 N a i ( x , t , u ) u x i + b ( x , t , u ) = ε ∑ i = 1 N u x i u x i (2a)

where the positive perturbation parameter tends to zero [

The discrete analogues of Equations (1) (2) lead to the solution of the general linear system

A u = s (2b)

where the coefficient matrix A is a large sparse unsymmetric real (n × n) matrix of irregular structure.

Computational complexity of algorithms is an important subject in which considerable research has been focused in the last decades [

An interesting topic in the framework of comparison of the number of flops per iteration to be performed for several classes of solution methods, i.e. 1) direct second order methods (based on direct matrix decomposition), 2) iterative first order methods and 3) iterative second order methods, for the cases of dense and sparse problems reveals the following:

1) In first order iterative methods, the number of flops per iteration is generally much smaller than that of second order methods, while the former generally need many iterations to converge (maybe few hundreds up to few thousands) and also have difficulties to obtain highly accurate optimal solutions.

2) The second order iterative methods usually converge quickly (few tens iterations) to highly accurate solutions.

3) The first order iterative methods overall are considered to be much superior to second order methods based on direct factorization in solving large scale SDP problems. However, the latter methods in solving linear systems may be prohibitively expensive (even impossible) due to high number of flops required and the high amount of memory space needed for storing the coefficient matrix.

4) The computational complexity of second order iterative methods has the following characteristics:

a) There is no need for computing and/or store the coefficient matrix

b) The computational work and storage requirements of inner iterations are very comparable to that of iterations of log-barrier method

c) The total number of inner iterations performed by iterative second order method depends on the choice of preconditioner used in the iterative algorithm of the original system.

Note that a logarithmic barrier term, which emphasizes the improvement of poor quality elements, solves the constrained optimization problem using the gradient of the objective function [

Conclusively, the second order iterative methods, as far as the inner iterations are concerned, behave quite similar to first order methods. Furthermore, the development of efficient preconditioners for solving the original linear system is a decisive factor for making the second order iterative methods superior to the first order iterative methods.

In this section, a class of iterative methods of second order for solving large sparse linear systems of the form Au = b is presented and explicit preconditioned methods for approximating the solution of complex computational problems are given. Apart of the extensive research work that has been done for solving linear systems by using second order iterative methods as indicated in section 2, it is worthwhile to mention that the efficient usage of second order iterative schemes for solving nonlinear systems has been reported by Lipitakis (1990) and these iterative schemes are originated from the E-algorithm, a modified version of the well-known algorithm of Euclid, written in the form of a general second order iterative scheme [

Let us consider a class of non-stationary iterative methods of second order the form:

u i + 1 = π i u i + τ i u i − 1 + δ i , i ≥ 0 (3)

where π i , τ i are real parameters, the so-called E-parameters [

These iterative schemes with appropriate selection of the parameters can be used in conjunction with various explicit preconditioned methods, such as explicit Richardson, Chebychev, fractional step iterative method, Grank-Nickolson, multiple explicit Jacobi, explicit preconditioned Conjugate Gradients, etc. for solving complex computational problems, such as 3D-initial boundary value problems. It is known that the corresponding approximate factorization algorithms and the approximate inverse algorithms are tediously complicated, especially in the three-dimension space with complicated boundary conditions in irregular domains.

Let us consider an unsymmetric large sparse linear system Au = b. Then, the choice of E-parameters:

π i = [ I − H − 1 A ] , τ i = 0 , δ i = H − 1 b (4)

where the non-singular matrix H is chosen such that the computational work required for solving the linear system Hu = s (where s is given) is considerably small compared to that required to solve the system Au = b, leading to the iterative procedure,

u i + 1 = [ I − H − 1 A ] u i + H − 1 b , i ≥ 0 (5)

Note that several choices of matrix H lead to well-known iterative methods, i.e. Richardson, Block Jacobi, Gauss-Seidel, SOR, SSOR, etc.

Consider the unsymmetric linear system Au = b, a family of ellipses of centred with foci d ± c and assuming that the E-parameters are selected as

π i = [ 1 − β ι − a i A ] , τ ι = − β ι , δ ι = a i b , i ≥ 0 (6)

where the acceleration parameters a i , β ι are defined as

a i = [ d − ( c / 2 ) 2 a i − 1 ] − 1 , β i = d a i − 1 , i ≥ 1 (7)

with

a 1 = 2 d / ( 2 d 2 − c 2 ) , β 1 = d a 0 − 1 , a 0 = 1 / d , β 0 = 0 (8)

The iterative method (3) then can be equivalently written as a second order non-stationary iterative scheme (known as Chebychev iterative method):

u i + 1 = ( 1 + β i − a i A ) u i − β i u i − 1 + a i b , i ≥ 1 (9)

or equivalently

u i + 1 = u i + a i r ι + β i ( u i − u i − 1 ) , i ≥ 1 (10)

where r is the residual factor r = b − A u , and u 0 , u 1 are arbitrary chosen initial values.

The simplified choice of E-parameters

π i = [ 1 − β − a A ] , τ ι = − β , δ ι = a b , i ≥ 0 (11)

leads to second-order iterative scheme (known as the second-order Richardson’s method):

u i + 1 = [ u i + a r i + β ( u i − u i − 1 ) ] , i ≥ 1 (12)

where the parameters α, β are the asymptotic values of a i , β i [

Let us consider the approximate factorization A ≈ L s U s , where L s and U s are sparse decomposition factors [

Note that the effectiveness of explicit iterative methods is mainly related to the fact that the exact inverse of the sparse matrix A, although is full, exhibits a similar “fuzzy” structure as A, i.e. the largest elements are clustered around the principal diagonal and main diagonals [

The selection of the E-parameters

π i = [ 1 + b − a M r A ] , τ i = − β , δ i = a M r b , i ≥ 0 (13)

and

π i = [ 1 + β i − a i M r A ] , τ i = − β i , δ i = a i M r b , i ≥ 0 (14)

where M r is an approximate form of the inverse of A and α, β and a i , β i are preconditioned parameters and sequences of parameters respectively, leads to the following explicit iterative schemes:

u i + 1 = ( 1 + β − α M r A ) u i − β u i − 1 + a M r b , i ≥ 1 (15)

u i + 1 = ( 1 + β − α i M r A ) u i − β i u i − 1 + a i M r b , i ≥ 1 (16)

which are known as the Explicit second order Richardson and Explicit Chebychev methods respectively [

In these explicit iterative schemes several approximate forms of the inverse can be effectively used by retaining only δ U and δ L ( δ U ∈ [ 2 , n − 1 ] , δ L ∈ [ 1 , n ] ) diagonals in the upper and lower triangular parts of M r , the remaining elements being just not computed at all [

A class of optimized approximate inverse variants can be obtained by considering a (near) optimized choice of the approximate inverse M depends on the selection of related parameters, i.e. fill-in parameters r_{1}, r_{2}, retention parameters δl_{1}, δl_{2} and entropy-adaptivity-uncertainty (EAU) parameters [

where M r 1 = m − 1 , r 2 = p − 1 δ l 1 , δ l 2 of sub-class I is a banded form of the exact inverse retaining δ l 1 , δ l 2 elements along each row and column respectively, while its elements are equal to the corresponding elements of the exact inverse. The term M S r 1 = m − 1 , r 2 = p − 1 δ l 1 , δ l 2 of sub-class II is a banded form of M, retaining only δ l 1 , δ l 2 elements along each row and column during the computational procedure of the approximate inverse and under certain hypotheses can be considered as a good approximation of the original inverse, while the entries of the approximate inverse in sub-class III have been retained after computing M* ( r 1 < m − 1 and r 2 < p − 1 ) and are less accurate than the corresponding entries of M * r 1 , r 2 δ l 1 , δ l 2 .Finally, in sub-class IV the elements of the approximate inverse can be computed.

Note that the generalized Approximate Inverse Matrix (AIM) techniques can be efficiently used in conjunction with explicit iterative schemes leading to effective composite semi-direct solution methods solving large linear systems of algebraic equations on multiprocessor systems.

Analogous relationships can be derived for various explicit CG-type methods as indicated in the following. The generalized alternating group explicit (AGE) iterative methods, based on the known ADI methods [

A = D + G 1 + G 2 (18)

where D is a non-negative diagonal matrix and D, G 1 , G 2 satisfy the conditions

1) The matrices ( G i + θ D + p I ) , i = 1 , 2 are non-singular for any θ ≥ 0 , ρ > 0 ,

2) The system

x = G 1 − 1 v and y = G 2 − 1 z (19)

can be easily solved explicitly for any vectors v and z. Note that this statement holds only to certain model problems. Then, the following generalized AGE scheme can be obtained

( G i + p i + 1 I ) u i + 1 / 2 = ( p i + 1 I − G 2 ) u i + b (20)

( G i + p i + 1 I ) u i + 1 = ( G 2 − ( I − ω ) p i + 1 I ) u i + ( 2 − ω ) p i + 1 u i + 1 / 2 (21)

where ω is a non-negative acceleration parameter [

In an analogous manner can be derived the explicit preconditioned methods 1) Richardson + AGE method and 2) Chebychev semi-iterative method + AGE method [

The multiple explicit Jacobi (μ-EJ) method and its several parametric forms, provided that their spectral radius ρ satisfies the corresponding convergence condition, in combination with the Lanczos economized Chebychev polynomials of degree μ, has proved to be effective for solving elliptic boundary-value problems in parallel processors [

Let us assume that the E-parameters are selected as follows:

π i = p i + 2 ( 1 + γ i + 2 M μ A ) , τ ι = 1 − p i + 1 , δ i = p i + 1 γ i + 1 M μ b , i ≥ 0 (22)

where

p i + 1 = 1 + ( a i / a i − 1 ) β ι , γ i + 1 = α i / p i + 1 , i = 0 , 1 , 2 , ⋯ , k − 1 (23)

with k the smallest integer such that r k = b − A u k = 0 , p i = 1 and a i , β i are the scalar quantities as defined in the original CG paper [

The iterative scheme (21) then becomes

u i + 1 = p i + 1 ( 1 + γ i + 1 M μ A ) u i + ( 1 − p i + 1 ) u i − 1 + p i + 1 γ i + 1 M μ b , i ≥ 1 (24)

or equivalently the second order explicit preconditioned CG scheme can be obtained as

u i + 1 = u i − 1 + p i + 1 ( γ i + 1 M μ r i + u i − u i − 1 ) , i ≥ 1 (25)

In the following, the selection of E-parameters for various explicit preconditioned methods is presented.

The proposed explicit iterative methods can be used for solving large sparse linear systems and the E-iterative schemes can generate useful explicit iterative schemes of higher order with suitable selection of E-parameters for solving a wide class of very large sparse linear systems in multiprocessor systems.

Typical applications include a large class of initial-boundary value problems of general form in three space dimensions with strong nonlinearities

u t + ∑ i = 1 N a i ( x , t , u ) u x i + b ( x , t , u ) = ε ∑ i = 1 N u x i u x i (25a)

where the positive perturbation parameter tends to zero [

The discrete analogues of Equation (25a) lead to the solution of the general linear system

A u = s (25b)

where the coefficient matrix A is a large sparse unsymmetric real (n × n) matrix of irregular structure.

In this section an iterative solver of second order for solving linear and nonlinear systems of irregular structure is presented in pseudo algorithmic form:

Algorithm SOIS-1 (EXSOM, SEIS, U-1, U0, NMAX, Π, R, Δ, U)

Purpose: This algorithm describes the general selection of an iterative solver of second order for solving linear and nonlinear nonsymmetric systems of irregular structure and the selection of corresponding parameters.

Input: the computational module EXSOM selecting the explicit iterative solver SEIS and corresponding parameters, the initial values u − 1 , u 0 , max number of iterations NMAX.

Output: final explicit iterative solver SEIS and approximate iterative solution U.

Computational Procedure:

step 1: Call the explicit iterative module EXSOM and determine the explicit iterative solver to be used

step 2: Select initial vectors u − 1 and u 0 and set the max number of iterations NMAX

step 3: Select the corresponding E-parameters π, r and correction term δ

step 4: For i = 0 , 1 , 2 , ⋯

step 5: Compute approximate vector u i + 1 as u i + 1 = π i u i + r i u i − 1 + δ i

// the E-parameters π, r and correction term δ are selected from the corresponding explicit iterative scheme of computational module EXSOM //

step 6: If the convergence criterion is satisfied and the max number of iteration NMAX is hold,

then go to step 7,

else go to step 4 and

continue

step 7: Print the explicit iterative solver SEIS

step 8: Form approximate solution u

The explicit computational module EXSOM can be used for solving an explicit iterative solver as follows:

module EXSOM (ERI, ECH, GAGE, RAGE, CHAGE, GFSI, MEJ, EPCGSO, SEIS)

Purpose: this module determines the explicit iterative solver that will be used for solving the given linear nonsymmetric system of irregular structure

Input: the explicit iterative solvers ERI, ECH, GAGE, RAGE, CHAGE, GFSI, MEJ, EPCGSO, IS integer (=1, 2, 3, 4, 5, 6, 7, 8) is the number of explicit iterative solver to be used

Output: The selected explicit iterative solver SEIS = IS

Computational Procedure:

step 1: Consider one of the following available explicit iterative solvers and set the corresponding value of IS

step 1.1: If IS==1,

then the explicit Richardson method ERI is to be used;

return

step 1.2: If else IS==2,

then the explicit Chebychev ECH is to be used

return

step 1.3: If else IS==3,

then the generalized AGE method GAGE is to be used;

return

step 1.4: If else IS==4,

then the Richardson + AGE method RAGE is to be used;

return

step 1.5: If else IS==5,

then the Chebychev + AGE method CHAGE is to be used;

return

step 1.6: If else IS==6,

then the general fractional step iteration GFSI is to be used;

return

step 1.7: If else IS==7,

then the multiple explicit Jacobi MEJ is to be used;

return

step 1.8: If else IS==8,

then the explicit preconditioned CG of second order is to be used

return

step 2: return the selected explicit iterative solver (SEIS)

Additional explicit iterative solvers can be used in module EXSOM for solving the given linear system. The algorithm SOIS-1 can be used as a general explicit iterative solver of second order for solving linear and nonlinear non-symmetric systems of irregular structure by selecting the corresponding parameters.

The selection of an efficient approximate inverse preconditioner for solving explicitly complex computational problems is an interesting research topic of critical importance. Let us assume that a non-singular large sparse unsymmetric matrix of irregular structure can be factorized as A = LU (

In the following, a class of adaptive exact and approximate inverse solvers based on exact/approximate LU decompositions and approximate inverse methods is described. Let us assume the general linear system

A u = s (25b)

where the coefficient matrix A is a large sparse real (n × n) matrix of irregular structure.

The structure of A is shown in the following diagram:

Note that such regular matrix structures occur only for certain model type problems.

Let us consider, the factorization A = LU and an approximate factorization of the coefficient matrix A,

A ≈ L s U s (27)

where L s and U s , are lower and upper sparse triangular matrices of irregular structures of semi-bandwidths m and p retaining r 1 and r 2 fill-in terms respectively. The decomposition factors L s and U s are banded matrices with l_{1} and l_{2} the numbers of diagonals retained in semi-bandwidths m and p respectively, of the following form:

The elements of the decomposition factors L_{s} and U_{s} can be obtained from the algorithmic procedure FELUBOT [

A class of exact and approximate inverse matrix techniques can be considered containing several sub-classes of approximate inverses according to memory requirements, computational work, accuracy, as indicated in the following scheme:

Let us assume that M r 1 , r 2 , a non-singular (n × n) matrix, is an approximate inverse of A, i.e. M r 1 , r 2 = { M i , j } r 1 , r 2 , i , j ∈ [ 1 , n ] . Note that if r 1 = m − 1 and r 2 = p − 1 non-zero elements have been retained in the corresponding decomposition factors, then M r 1 , r 2 = M , where M is the exact inverse of A. The elements of M can be determined by solving recursively the systems

M L = U − 1 and U M = L − 1 (28)

having main disadvantages, i.e. high storage requirements and computational work involved particularly in the case of solving very large unsymmetric linear systems.

A class of approximate inverses M δ l 1 , δ l 2 can be obtained by retaining only δl_{1} and δl_{2} diagonals in the lower and upper triangular parts of inverse respectively, the remaining elements being just not computed at all. Optimized forms of this algorithm are particularly effective for solving banded sparse FE systems of very large order, i.e. [ δ l 1 + δ l 2 ] > n / 2 or in the case of narrow-banded sparse FE systems of very large order, i.e. [ δ l 1 + δ l 2 ] ≪ n / 2 . Then an explicit approximate inverse preconditioner can be described in the following adaptive algorithmic form.

The elements of the exact and approximate inverse of a given unsymmetric and irregular structure can be obtained as follows and the explicit banded approximate inverse algorithm can be described by the following algorithmic procedure in pseudocode form:

Algorithm EBAIM-1 (A, n, εEM, r_{1}, r_{2}, m, p, l_{1}, l_{2}, δl_{1}, δl_{2}, M)

Purpose: This algorithm computes the elements of the exact inverse of a given real unsymmetric (n × n) matrix of irregular structure arising in FE/FD discretization of elliptic and parabolic boundary value in three space dimensions. The algorithm can also compute approximate inverse matrices of the given coefficient matrix.

Input: given matrix A; n order of A; submatrices F, H, Γ, Z; parameter εEM (indicating the exact inverse or approximate inverse matrix), s m, p; l_{1} and l_{2} numbers of diagonals retained in semi-bandwidths m and p respectively, δl_{1} and δl_{2} widths of bands in A

Output: elements μ_{i}_{,j} of the exact inverse M

Computational Procedure:

Step 0: Read the value of adaptive parameter εEM

//for the appropriate value of adaptive parameter εEM the algorithm computes the exact inverse matrix or approximate inverse matrices//

Call module exactmode-1(εEM)

//If the module exactmode-1is activated then the algorithm EBAIM-1 computes the exact inverse of a given unsymmetric matrix of irregular structure using an exact LU factorization, otherwise the algorithm can be used for computing an approximate inverse matrix of the given coefficient matrix//

step 1: Let r l 1 = r 1 + l 1 ; r l 2 = r 2 + l 2 ; r l 11 = r l 1 − 1 ; r l 21 = r l 2 − 1 ; m r 1 = m − r 1 ; m l 1 = m + l 1 ; p r 2 = p − r 2 ; p l 2 = p + l 2 ; n m r 1 = n − m + r 1 ; n p r 2 = n − p + r 2

step 2: For i = n:1

step 3: For j = i : max ( 1 , i − δ l + 1 )

step 4: If j > n m r 1 then

step 5: If i = j then

step 6: If i = n then

step 7: μ n , n = 1

step 8: else

step 9: μ i , j = 1 − g j μ i , j + 1

step 10: μ i , j = ω n − β j μ i , j + 1

step 11: else

step 12: μ i , j = − g j μ i , j + 1

step 13: μ i , j = − β j μ i , j + 1

step 14: else

step 15: If j > n p r 2 and j ≤ n m r 1 then

step 16: If i = j then

step17: μ i , j = 1 − g j μ i , j + 1 − ∑ k = 0 n m r 1 − j h r l 11 − k , j + k + 1 − r 1 μ i , j + m r 1 + k

step18: μ i , j = ω 1 − β j μ i , j + 1 − ∑ k = 0 n m r 1 − j γ r l 11 − k , j + k + 1 − r 1 μ i , j + m r 1 + k

step 19: else

step 20: μ i , j = − g j μ i , j + 1 − ∑ k = 0 n m r 1 − j h r l 11 − k , j + k + 1 − r 1 μ i , j + m r 1 + k

step 21: μ i , j = − β j μ i , j + 1 − ∑ k = 0 n m r 1 − j γ r l 11 − k , j + k + 1 − r 1 μ i , j + m r 1 + k

step 22: else

step 23: If j ≥ r l 1 and j ≤ n p r 2 then

step 24: If i = j then

step 25: μ i , j = 1 − g j μ i , j + 1 − ∑ k = 0 n m r 1 − j h r l 11 − k , j + k + 1 − r 1 μ i , j + m r 1 + k − ∑ k = 0 n p r 2 − j f r l 12 − k , j + k + 1 − r 2 μ i , j + p r 2 + k

Step 26: μ i , j = ω 1 − β j μ i , j + 1 − ∑ k = 0 n m r 1 − j γ r l 11 − k , j + k + 1 − r 1 μ i , j + m r 1 + k − ∑ k = 0 n p r 2 − j z r l l 2 − k , j + k + l − r 2 μ i , j + p r 2 + k

Step 27: else

Step 28: μ i , j = − g j μ i , j + 1 − ∑ k = 0 n m r 1 − j h r l 11 − k , j + k + 1 − r 1 μ i , j + m r 1 + k − ∑ k = 0 n p r 2 − j f r l 12 − k , j + k + 1 − r 2 μ i , j + p r 2 + k

Step 29: μ i , j = − β j μ i , j + 1 − ∑ k = 0 n m r 1 − j γ r l 11 − k , j + k + 1 − r 1 μ i , j + m r 1 + k − ∑ k = 0 n p r 2 − j z r l l 2 − k , j + k + l − r 2 μ i , j + p r 2 + k

Step 30: else

Step 31: If i = j then

Step 32: If i = 1 then

Step 33: μ 1 , 1 = 1 − g 1 μ 1 , 2 − ∑ k = 1 l 1 h 1 , k μ 1 , m + k - 1 − ∑ k = 1 l 2 f 1 , k μ 1 , p + k - 1

Step 34: μ 1 , 1 = ω 1 − β 1 μ 1 , 2 − ∑ k = 1 l 1 γ 1 , k μ 1 , m + k − 1 − ∑ k = 1 l 2 z 1 , k μ 1 , p + k − 1

Step 35: else

Step 36: μ i , j = 1 − g j μ i , j + 1 − ∑ k = j + 1 − r 1 l 1 h j , k μ i , m + k − 1 − ∑ k = 1 j − 1 h j − k , l 1 + k μ i , m l 1 + k − 1

Step 37: μ i , j = ω 1 − β j μ i , j + 1 − ∑ k = j + 1 − r 1 l 1 γ j , k μ i , m + k − 1 − ∑ k = 1 j − 1 γ j − k , l 1 + k μ i , m l 1 + k − 1 − ∑ k = 1 j − 1 z j − k , l 2 + k μ i , p l 2 + k − 1 − ∑ k = j + 1 − r 2 l 2 z j , k μ i , p + k − 1

Step 38: else

Step 39: μ i , j = − g j μ i , j + 1 − ∑ k = j + 1 − r 1 l 1 h j , k μ i , m + k − 1 − ∑ k = 1 j − 1 h j − k , l 1 + k μ i , m l 1 + k − 1 − ∑ k = 1 j − 1 z j − k , l 2 + k μ i , p l 2 + k − 1 − ∑ k = j + 1 − r 2 l 2 z j , k μ i , p + k − 1

Step 40: μ i , j = − β j μ i , j + 1 − ∑ k = i + 1 − r 1 l 1 γ j , k μ i , m + k − 1 − ∑ k = 1 j − 1 γ j − k , l 1 + k μ i , m l 1 + k − 1 − ∑ k = 1 j − 1 z j − k , l 2 + k μ i , p l 2 + k − 1 − ∑ k = j + 1 − r 2 l 2 z j , k μ i , p + k − 1

Step 41: For j = i − 1 to max [ 1 , i − δ l + 1 ]

Step 42: μ i , i = μ i , j

Step 43: Print out the inverse elements μ i , i

Step 44: End

Note that if the computational module exactmode-1 is activated, as indicated below, then the algorithm EBAIM-1 computes the elements of the exact inverse of a given unsymmetric matrix of irregular structure using an exact LU factorization.

Module exactmode-1 (εEM)

/If this module is activated (εEM = 1), then the algorithm EBAIM-1 computes the exact inverse of a given unsymmetric matrix of irregular structure using an exact LU factorization, otherwise the algorithm can be used for computing an approximate inverse matrix of the given coefficient matrix/

If εEM = 1 then

Set r 1 = p − 1 ; r 2 = m − 1

Set l 1 = n − 2 ; l 2 = n − p

Set δ l 1 = n − 2 ; δ l 2 = n − 2

else

exit

end module exactmode-1

The computational work of the EBAIM-1 algorithm is ≈ O [ n ( δ l 1 + δ l 2 ) ( r 1 + r 2 + l 1 + l 2 + 2 ) ] multiplications, while the memory requirements are (n × n) words. In the case of very large systems, the memory requirements could be prohibitively high and the usage of efficient memory requirements of approximate inverse matrices is desirable.

The preconditioned PCG method can solve the problem min ‖ b − A R − 1 x ‖ , where R is the sparse, non-singular QR factor, while the preconditioned CGLS method can solve the equations:

M = R T R , R − T A T A R − 1 u = R − T A T b and u = R x . (29)

Note that the factor Q cannot be stored, while the only additional computational work is solving the two equations R T w = v and R z = w . All the factorization processes are numerically stable. It should be pointed out that the sparse preconditioner M^{*} used in the modified PCG method is the approximate inverse of factor R, which using is closely related with the so-called mrTIGO method.

In order to compute efficiently the solution of the linear system Ax = b, a modified Preconditioned Conjugate Gradient (mPCG) method in conjunction to the modified rTIGO method [

Algorithm mPCG (A, b, tol, x_{0}, M^{*}, x)

Purpose: a modified PCG method is used for solving a given system of linear equations

Input: A is a symmetric and positive definite coefficient matrix, b is the right hand side vector, tol is the predetermined tolerance, x 0 is the initial guess, M^{*} is the required preconditioner

Output: x the solution vector

Computational Procedure:

Step 1: Given x 0 and preconditioner M^{*}

Step 2: Set r 0 = A x 0 − b

Step 3: Solve M y 0 = r 0

Step 4: Set p 0 = − y 0 , k = 0

Step 5: While r k ≠ 0

Step 6: Compute a step length a k = ( r k T y k ) / ( p k T A p k )

Step 7: Update the approximate solution x k + 1 = x k + a k p k

Step 8: Update the residual r k + 1 = r k + a k A p k

Step 9: Solve M * y k + 1 = r k + 1

Step 10: Compute a gradient correction factor β k + 1 = ( r k + 1 T y k + 1 ) ( r k T y k )

Step 11: Set the new search direction p k + 1 = − y k + 1 + p k + 1 p k

Step 12: κ = κ + 1

Step 13: End (while)

This algorithm requires the additional work that is needed to solve the linear system

M * y ˜ n = r n (29)

once per iteration. Therefore, the preconditioner M^{*} should be chosen such that can be done easily and efficiently.

The preconditioner M^{*} = G that results in a minimal memory use. The storage requirement was the vectors r, x, y, p and the upper triangular matrix G, in the data implementation. The convergence rate of preconditioned CG is independent of the order of equations and the matrix vector products are orthogonal and independent. The preconditioned CG method in not self-correcting and the numerical errors accumulate every round. Therefore, to minimize the numerical errors in the pCG, it was used double precision variables at the cost of memory use. The explicit pCG method of second order can be alternatively used in conjunction with the explicit approximate inverse Mμ^{*} for solving complex computational problems with the appropriate selection of the E-parameters of

The presented second order iterative schemes can be efficiently used for solving 2d and 3d initial and boundary value problems.

A class of general iterative methods of second order is described. Explicit adaptive iterative schemes and exact/approximate inverse preconditioners have been introduced and the selection of iterative parameters has been discussed. Exact and Approximate Inverse Matrix Algorithmic Techniques have been presented. Exact and Approximate Inverse Preconditioners have been described in adaptive algorithmic form. Explicit preconditioned Conjugate Gradients methods of

E-parameters | Explicit Preconditioned Methods | ||
---|---|---|---|

π i _{ } | τ i _{ } | δ i _{ } | |

1 + β − α M r A | − β | α M r b | Explicit Richardson (Lipitakis and Evans, 1987) |

1 + β i − α i M r A | − β i _{ } | α i M r b | Explicit Chebychev (Lipitakis and Evans, 1987) |

( ρ i I − G 1 ) − 1 ( ρ i I − G 2 ) − 1 ^{ } | 0 | 2 ρ i ( ρ i I + G i ) − 1 ^{ } | Generalized AGE method (Lipitakis and Evans, 1981) |

1 + ω i + 1 ( ρ i I + G 1 ) − 1 ( ρ i I + G 2 ) − 1 A | 0 | − ω i + 1 ( ρ i I + G 1 ) − 1 ( ρ i I + G 2 ) − 1 b | Richardson + AGE method (Evans, 1985) |

ω i H i _{ } | 1 − ω i _{ } | 2 ω i ρ i ( ρ i I + G 1 ) − 1 ( ρ i I + G 2 ) − 1 b | Chebychev + AGE method (Evans, 1985) |

( I + θ G 1 ) − 1 ( I + θ G 2 ) − 1 { ( I + θ G 1 ) ( I + θ G 2 ) − η A } | 0 | η b | General fractional step iteration (Marchuk, 1975) |

∏ i = 0 ν − 1 ( I − α M μ A ) | 0 | α i M μ b | Multiple explicit Jacobi (Lipitakis, 1984) |

ρ i + 1 ( 1 + γ i + 1 M μ A ) | 1 − ρ i + 1 | ρ i + 1 γ i + 1 M μ b | Explicit Preconditioned Conjugate Gradients of second order (Lipitakis et al., 2017) |

second order are given. An Adaptive Preconditioned Conjugate Gradient Method using explicit approximate preconditioners has been also presented. Future research work will be focused on the implementation of the presented methods to parallel computer environments and related applications.

The authors declare no conflicts of interest regarding the publication of this paper.

Lipitakis, A.-D. (2020) Explicit Iterative Methods of Second Order and Approximate Inverse Preconditioners for Solving Complex Computational Problems. Applied Mathematics, 11, 307-327. https://doi.org/10.4236/am.2020.114023