International Journal of Applied Engineering Research, ISSN 0973-4562 Volume 13, Number 9 (2018) pp.

6751–6759
© Research India Publications, http://www.ripublication.com

Convergence Analysis of Modified Triangular and

Triangular Splitting Method for the Solution of Regularized Linear
System-Circulant Matrices
Convergence Analysis of MTTS

Malla Reddy Perati

Department of Mathematics, Kakatiya University, Warangal, T.S., India.

Ramesh Renikunta
Department of Mathematics, Kakatiya Institute of Technology & Science, Warangal, T.S., India.

Rajaiah Dasari
Department of Mathematics, Osmanaia University, Hyderabad, T.S., India.

Rajkumar L.P
Department of Mathematics, Kakatiya University, Warangal, T.S., India.

Abstract: In this paper, the homogeneous system π Q = 0 INTRODUCTION

is transformed to the non homogeneous regularized lin-
ear system Ax = b by introducing small perturbation The investigation of the fractal characteristics of traffic pro-
, and proved that the matrix A = QT +  I is positive cesses comes from the presence of burstiness over several
definite for  > 0. The steady state probability vector time scales. This leads to the introduction of the concept of
π of an irreducible circulant rate matrix Q is computed, self-similarity in teletraffic engineering emphasizing nat-
and also obtained the condition for the convergence of ural length of bursts. A traffic model is a mathematical
unique iterative solution by Modified Triangular and Tri- description of specific traffic type represented in terms
angular Splitting (MTTS) method proposed as in the cases of stochastic matrix. Therefore it is important to cap-
of Traingualr and Triangular Splitting (TTS), Triangular ture packet arrival flows and describe them with suitable
and Skew-symmetric Splitting (TSS). Moreover, we prove stochastic matrices. The use of stochastic rate matrix is
some properties of circulant matrices. From the numerical of interest in wide range of applications of performance
results, we conclude that the steady state probability vector measures of queueing systems, neural networks, dynamical
of proposed method converges rapidly to unique solution systems.
compare to TTS, and Jacobi methods.
For computing these performance measures, it is the key
importance to find out the steady state vector of the pertient
Keywords: Self-Similarity, Circulant stochastic matri- linear system. Hence, the solution of the linear system is
ces, Steady state probability vector, TTS Method, MTTS feasible and unique. Recently, there is a large amount of
Method, Convergence analysis. work devoted to solve the pertinent linear systems. The
steady state vector of the linear system is obtained by direct
or iterative methods [27], [17], [23], [4]. The well-known
 International Journal of Applied Engineering Research, ISSN 0973-4562 Volume 13, Number 9 (2018) pp. 6751–6759
© Research India Publications, http://www.ripublication.com

Krylov subspace methods, and some preconditioning tech- 

n
and σi = 0. Each row and column sum is zero. Hence
niques in [10], [25], [8], [5], [2], [3], [6], [9], [15], [19]. The i=1
triangular and skew-symmetric (TSS) and triangular and the matrix Q is a doubly stochastic matrix.
symmetric splitting iteration methods has been developed M-Matrix. Any matrix A ∈ Rn×n of the form A = s I −
and discussed for solving positive-definite linear systems B , s > 0, B ≥ 0 is called an M-matrix if s ≥ ρ(B ). If
on stochastic matrices [22], [29], and it is difficult to esti- s > ρ(B ) then A is non-singular M-matrix, otherwise A
mate the optimal parameter. Moreover, the triangular and is singular M-matrix.
triangular splitting iteration method is used for finding the
Definition 1. The matrix A = [aij ] ∈ Rn×n is said to be
solution of the pertinent system [31]. Hence, in this paper, 
we proposed new approach MTTS for determining the diagonal dominant (DD), if |aii | ≥ |aij |, i ∈ N , and A
i=j 
steady state vector of given system. Moreover, we find out is called strictly diagonal dominant if |aii | > |aij |, i ∈
the optimal parameter α , and convergence criteria of circu- i=j
lant stochastic rate matrix. While finding the convergence N , where, N = {1, 2, 3, · · · n}.
condition of the iterative solution, we obtain the value of
that contraction factor α. In the earlier papers there is no Definition 2. A non-symmetric matrix A ∈ Rn×n is
fixed value for the contraction factor α and which is not positive definite if its symmetric part is positive definite.
suitable for long run systems. In this proposed method, we The MTTS iteration method cannot be directly applied
obtain the fixed value for the contraction factor α. to find the steady state vector of the linear system wherein
The rest of the paper is organized as follows: In section 2, the coefficient matrix is circulant stochastic rate matrix.
the regularized preconditioned linear system of circulant Hence, we consider a regularized linear system as follows
rate matrix is considered and some properties of circulant [29]. Consider the homogeneous linear system π Q = 0,
matrices are discussed. In section 3, the MTTS iteration where, Q = I − P and P is circulant stochastic probability
method discussed to solve the regularized linear system. matrix. Taking transpose on both sides of said homoge-
In section 4, the convergence criteria for MTTS method neous system, we obtain, QT π T = 0, and rewritten as
discussed. In section 5, we demonstrate accuracy of the Āx = 0, where x = π T , and Ā = QT has zero row
proposed method by means of numerical results. Finally, and column sum, positive diagonal entries, non-positive
conclusions are made in section 6. off diagonal entries and has the one dimensional null
space is singular. Hence, we modified the given homoge-
neous system of equations into regularized linear system
by introducing small perturbation . Then there exists a non
REGULARIZED PRECONDITIONED negative constant  > 0 such that the preconditioned linear
LINEAR SYSTEM AND SOME system is
PROPERTIES OF CIRCULANT MATRICES
Ax = (Ā +  I )x = (QT +  I )x = en = b. (2)
In this section, we define circulant matrix and some chre-
where en is a vector given by en = [0, 0, . . . , 1]T . The
matistics of the regularized and positive definite circulant
steady-state probability distribution vector is then obtained
matrices.
by normalizing the solution x. For proving the regularized
Consider the circulant rate matrix
preconditioned matrix A is positive define, the matrix must
⎡ ⎤ be non-symmetric. If Q is circulant matrix with distinct
σ1 σ2 σ3 ... σn
⎢ σn σ1 σ2 ... σn−1 ⎥ elements of order n then Q is non-symmetric for n ≥ 3. If
⎢ ⎥
⎢σn−1 σn σn−2 ⎥ the matrix Q is non-symmetric apparently the matrix A is
Q=⎢ σ1 ... ⎥
⎢ .. .. ⎥ non-symmetric.
⎣ ... ... ... . . ⎦
By the definition, for proving the matrix Q is positive
σ2 σ3 σ4 ... σ1
definite, it is enough to prove that symmetric part of the
= (σ1 , σ2 , . . . , σn )(say) matrix Q is positive definite. i.e. Q+2Q is positive definite.
T

Lemma 3. If  is non-negative matrix and i is the row

for σ1 > 0, σi ≤ 0, 2 ≤ i ≤ n (1) sum of  then ρ() = i (A).

6752
 International Journal of Applied Engineering Research, ISSN 0973-4562 Volume 13, Number 9 (2018) pp. 6751–6759
© Research India Publications, http://www.ripublication.com

Since A+2A =  I + Q+2Q ,

T T
Proof. Let
⎡ σ2 +σn σ3 +σn−1 σ2 +σn ⎤ = ( + σ1 ) − ,
0 ...
⎢
2 2 2
⎥ and ρ() = σ2 + σ3 + σ4 + . . . + σn ,
⎢ σ2 +σn σ2 +σn σ3 +σn−1 ⎥
= σ1 .
⎢ 2 0 ... ⎥
⎢ 2 2
⎥ ⇒ (σ1 + ) > ρ().
⎢ σ4 +σn−2 ⎥
 = ⎢ σ3 +σ2 n−1 σ2 +σn
0 ... ⎥ ≥ 0.
⎢ 2 2 ⎥ A is positive definite.
⎢ ⎥
⎢ .. .. .. .. .. ⎥ On the other hand, we assume that A is positive definite.
⎣ . . . . . ⎦
σ2 +σn σ3 +σn−1 σ4 +σn−2 We complete the theorem along the lines of the theorem 1
... 0
[18]. If A is positive definite then A+2A = (σ1 +)I −, is
2 2 2 T

be a non-negative circulant matrix. positive definite. For any set of indices N ⊆ {1, 2, . . . .n},

σkj < σ1 +  for k = 1, 2, . . . .n
⇒ min rA (i, j ) = max rA (i, j ) j =N
i=j i=j
⇒ (σ1 + ) > ρ().

= σ2 + σ3 + σ4 + . . . + σn ,
⇒ σ2 + σ3 + σ4 + . . . + σn ≤ ρ() Corollary 5. [18] Let D = ρI −A be a positive definite M-
≤ σ2 + σ3 + σ4 + . . . + σn , matrix of order n, A = (aij , i = 1, . . . , n, j = 1, . . . , n).
n  n
⇒ ρ() = σ2 + σ3 + σ4 + . . . + σn . Then aij < ρn; aij + aj i < 2ρ.
i=1 j =1


Corollary 6. Let A be a positive definite M-matrix of order
Theorem 4. If Q ∈ Rn×n is a circulant stochastic rate n. Then (σ1 + ) > n2 (σ2 + σ3 + σ4 + . . . + σn ).
matrix given in (1) then there exists  > 0 such that the
matrix A = Ā +  I = QT +  I is positive definite iff Proof. Along the lines of the corollary 5, 2(σ2 + σ3 + σ4 +
(σ1 + ) > ρ(). . . . + σn ) < n(σ1 + ),
⇒ σ1 +  > n2 (σ2 + σ3 + σ4 + . . . + σn ).
Proof. For proving the matrix A is positive definite it is For proving the given system possesses an unique solution
enough to prove that the symmetric part of the matrix A+2A
T
here, we prove some properties of circulant matrices. These
is positive definite. properties are used in the convergence analysis of regular-
Consider the Q given by (1). Now, ized linear system to get the unique solution.
⎡ ⎤ n
σ2 +σn σ3 +σn−1 σ2 +σn Let i (A) = |aij | be non diagonal row sum of a
σ1 + 2 
2 2
... 2 j =1,j =i
⎢ ⎥
⎢ σ2 +σn σ2 +σn σ3 +σn−1 ⎥ matrix A = [aij ] ∈ Rn×n .
⎢ 2 σ1 + 2 . . . ⎥
A + AT ⎢ 2 2
⎥ The spectral radius ρ(A) of the non-negative matrix is
⎢ σ3 +σn−1 σ2 +σn σ4 +σn−2 ⎥
=⎢ 2 σ + 
. . . ⎥ obtained by Gersogorin theorem [28],
2 ⎢ 2 1 2 2 ⎥
⎢ ⎥
⎢ .. .. .. .. .. ⎥
.
⎣ . . . . ⎦
σ2 +σn σ3 +σn−1 σ4 +σn−2 1
2 2 2
. . . σ1 + 2 
rA (i, j ) = {aii + ajj + [(aii − ajj )2 + 4i (A)j (A)] 2 }
1

2
= (σ1 + )I − ,

where for i  = j , 1 ≤ i, j ≤ n, and improved Perron�Frobenius

⎡ σ2 +σn σ3 +σn−1 σ2 +σn
⎤
0 ... result [28]
2 2 2
⎢ ⎥
⎢ σ2 +σn σ2 +σn σ3 +σn−1 ⎥
⎢ 2 0 ... ⎥
⎢ 2 2
⎥ min rA (i, j ) ≤ max rA (i, j ).
⎢ σ4 +σn−2 ⎥
 = ⎢ σ3 +σ2 n−1 σ2 +σn
0 ... ⎥ ≥ 0. i=j i=j
⎢ 2 2 ⎥
⎢ ⎥
⎢ .. .. .. .. .. ⎥
⎣ . . . . . ⎦
σ2 +σn σ3 +σn−1 σ4 +σn−2

2 2 2
... 0
Let ρ() denote the spectral radius of the non-negative Lemma 7. If Q ∈ Rn×n is a circulant stochastic rate
matrix . matrix then Q is diagonally dominant matrix.

6753
 International Journal of Applied Engineering Research, ISSN 0973-4562 Volume 13, Number 9 (2018) pp. 6751–6759
© Research India Publications, http://www.ripublication.com

Proof. From the equation (2), we have Since A is positive definite with σ1 +  is positive, then
⎡ ⎤ sgn(det A) is positive.

σ1 σ2 σ3 . . . σn
⎢ σn σ1 σ2 . . . σn−1 ⎥ 
n
⎢ ⎥ Lemma 11. [14] Let αi > 0(i = 1, 2, . . . , n) and αi ≥
⎢σn−1 σn σ1 . . . σn−2 ⎥
Q=⎢ ⎥ i=1
⎢ .. ⎥ 1. Let aij > 0, (i = 1, 2, . . . , n), then
⎣ . . . . . . . . . ... . ⎦
σ2 σ3 σ4 ... σ1

n
α α αn
a1j1 a2j2 · · · anj
is a circulant stochastic rate matrix. j =1
n
⇒ σi = 0, ⎛ ⎞α1 ⎛ ⎞α2 ⎛ ⎞αn
i=1
n

n
n

n ≤⎝ a1j ⎠ ⎝ a2j ⎠ · · · ⎝ anj ⎠ .
⇒ σ1 = σi = 0, σi > 0,
j =1 j =1 j =1
i=2
n
⇒ σ1 ≥ σi . Theorem 12. Let P = [aij ], for i = 1, 2, 3, . . . , n be a
i=1
circulant stochastic probability matrix such that aij > 0
⇒ Q is diagonally dominant matrix.

for i = 1, 2, 3, . . . , n, then,
Corollary 8. If A ∈ Rn×n is regularized circulant matrix

n
α α
then A is strictly diagonally dominant. αn
a1j1 a2j2 · · · anj
j =1
Proof. Let A = QT +  I ⎛ ⎞α1 ⎛ ⎞α2 ⎛ ⎞αn

n

n
n
⎡ ⎤ ≤⎝ a1j ⎠ ⎝ a2j ⎠ · · · ⎝ anj ⎠ ≤ 1.
σ1 +  σ2 σ3 ... σn
j =1 j =1 j =1
⎢ σn σ1 +  σ2 ... σn−1 ⎥
⎢ ⎥
⎢ σ1 +  ⎥ Proof. From Lemma 11, we have,
A = ⎢ σn−1 σn ... σn−2 ⎥,
⎢ .. .. ⎥
⎣ ... ... ... . . ⎦

n
α α αn
σ2 σ3 σ4 ... σ1 +  a1j1 a2j2 · · · anj
j =1
Since Q is circulant stochastic matrix, then ⎛ ⎞α1 ⎛ ⎞α2 ⎛ ⎞αn
n
⇒ σi = 0 with σ1 > 0 and σi < 0 for 2 < i < n,
n
n
n
i=1 ≤⎝ a1j ⎠ ⎝ a2j ⎠ · · · ⎝ anj ⎠
σ1 = σ2 + σ3 + σ4 + . . . + σn . j =1 j =1 j =1
⇒ σ1 +  > σ2 + σ3 + σ4 + . . . + σn ,
n 
n
⇒ σ1 +  > σi , Since P is doubly stochastic matrix, a1j = i for i =
i=2 j =1
⇒ A is strictly diagonal dominant matrix.
1, 2, 3, . . . , n.

n
Lemma 9. [21] If A is diagonally dominant and non- ⇒
α α αn
a1j1 a2j2 · · · anj
α α
≤ 1 1 2 2 · · · αnn = 1.1...1 (n times) = 1,

singular then sgn(det A) = sgn( aii ). j =1
i

n
α α αn
Theorem 10. If A is circulant matrix with diagonally ⇒ a1j1 a2j2 · · · anj
dominant and positive definite then sgn(det A) is positive. j =1
⎛ ⎞α1 ⎛ ⎞α2 ⎛ ⎞αn

n

n
n
Proof. From corollary 8, ≤⎝ a1j ⎠ ⎝ a2j ⎠ · · · ⎝ anj ⎠ ≤ 1.
j =1 j =1 j =1
A = QT +  I is a diagonally dominant.


⇒ sgn(det A) = sgn( aii ),
i
Corollary 13. Let P = [aij ], for i = 1, 2, 3, · · · n be a
= sgn(σ1 + .σ1 + ....σ1 + ), circulant stochastic probability matrix such that aij > 0
α α αn
= sgn((σ1 + )n ), for i = 1, 2, 3, · · · n, then a1j1 a2j2 · · · anj ≤ n1 .

6754
 International Journal of Applied Engineering Research, ISSN 0973-4562 Volume 13, Number 9 (2018) pp. 6751–6759
© Research India Publications, http://www.ripublication.com

THE MTTS ITERATION METHOD i.e. T ∈ Rn×n is a lower triangular matrix with its diago-

nal elements positive, and T ∈ Rn×n is a upper triangular
In this section, we find the steady state probability vector matrix with its diagonal elements positive and its off diag-
and convergence analysis of the regularized linear sys- onal elements negative. We can establish the following
tem (2) by applying the MTTS iteration method. Splitting MTTS iteration method for the regularized linear system
iteration methods for (2) require efficient splitting of the (2).
regularized matrix A. The matrix A can also be split into Given an initial guess x (0) ∈ Rn , consider the iterative
its Triangular and Triangular parts as scheme
A = [£ + D] + [U ], (3) 
(α I + T )x (k+1/2) = (α I − T )x (k) + b,

where ⎡ ⎤ (α I + T )x (k+1) = (α I − T )x (k+1/2) + b, (5)
σ1 0 0 . . . 0
⎢ 0 σ1 0 0 0⎥ for k = 0, 1, 2, . . . , where α is a given positive constant.
⎢ ⎥
⎢
D = ⎢ 0 0 σ1 0 0⎥ The above iterative scheme can be written as x (k+1) =
⎥,
⎣ 0 0 0 σ1 0 ⎦
M(α)x (k) + N (α)b, for k = 0, 1, 2, . . . , where
0 0 0 0 σ1 
⎡ ⎤ M(α) = (α I + T )−1 (α I − T )(α I + T )−1 (α I − T  ),
2
0 0 ... 0
⎢σ2 (6)
⎢

2
0 . . . 0⎥ ⎥
⎢ σ 3 σ2 
0 ⎥
£=⎢ 2
. . . ⎥, is the iteration matrix of the MTTS iteration method, and
⎢. . . . ⎥ 
⎣ .. .. .. . . . .. ⎦ N (α) = 2α(α I + T )−1 (α I + T )−1 . If ρ(M(α)) < 1 then
σn σn−1 σn−2 . . . 2 the iterative method MTTS is convergent.
and ⎡ ⎤
σn σn−1 . . . σ2 Lemma 14. [2] If T be a triangular matrix with positive
2
⎢0  σn . . . σ 3 ⎥ diagonal elements then T is positive definite.
⎢ 2 ⎥
⎢0 0 
. . . σ4 ⎥
U =⎢ 2 ⎥. 
⎢. . .. .. .. ⎥ Lemma 15. [2] If T be a triangular matrix with positive
⎣ .. .. . . .⎦ 
diagonal elements then T is positive definite matrix.
0 0 0 . . . 2
Here, D is a diagonal matrix with the diagonal elements of Theorem 16. [2] Let W (α) = (α I − T )(α I + T )−1 .
the matrix A and £, U are lower, upper triangular matrices If T ∈ Rn×n is a positive-definite matrix, then we have
of A respectively of regularized linear system. W (α) 2 < 1, ∀α > 0.

Let T = £ + D and T = U , then the splitting   
 Corollory 17. [2] Let W (α) = (α I − T )(α I + T )−1 .
A=T +T , (4) 
If T ∈ Rn×n is a positive-definite matrix, then we have

is called the Modified Triangular and Triangular splitting W (α) 2 < 1, ∀α > 0.
(MTTS) method for circulant stochastic matrices.
Where Theorem 18. Let A ∈ Rn×n be the regularized matrix
⎡ ⎤
σ1 + 2 0 0 ... 0 defined in (2), and splitting into triangular and triangu-
⎢ σ2 σ1 + 2 0 ... 0 ⎥ lar matrices given in (4). Then the spectral radius of the
⎢ ⎥
⎢ σ3 σ1 + 2 . . .

0 ⎥ iterative matrix M(α) < 1.
T =⎢ σ2 ⎥,
⎢ . . . . . ⎥
⎣ . . .
. .
. . . . ⎦
.
Proof. Let A ∈ Rn×n be the regularized matrix defined in
σn σn−1 σn−2 ... σ1 + 
2 (2), and splitting into the form (4). Let M(α) be the itera-
and ⎡ ⎤ tion matrix given in (6). Then the iteration matrix M(α) is
σn σn−1 . . .
2
σ2 similar to the matrix M(α) = (α I − T )(α I + T )−1 (α I −
⎢0  σn . . . σ3 ⎥   
⎢ 2 ⎥ T )(α I + T )−1 = W (α)W (α), where W (α) = (α I −
 ⎢0 0 
σ4 ⎥   
T =⎢ 2
... ⎥. T )(α I + T )−1 , and W (α) = (α I − T )(α I + T )−1 .
⎢. . . .. .. ⎥
⎣ .. .. .. . .⎦

Since the triangular matrices T and T of the regu-
0 0 0 ... 
2
larized preconditioned matrix A given in (3) are positive

6755
 International Journal of Applied Engineering Research, ISSN 0973-4562 Volume 13, Number 9 (2018) pp. 6751–6759
© Research India Publications, http://www.ripublication.com

definite, then from the lemma 14, and corollory 15, we have Consider

W (α) 2 < 1, W (α) 2 < 1, ∀α > 0.
(α I + Ti )−1 = [(α I + D) + Gi ]−1
∴ ρ(M(α)) = ρ(M(α)) = (α I + D)−1 [I + Gi (α I + D)−1 ]−1 ,
−1   −1
= (αI − T )(αI + T ) (αI − T )(αI + T ) 2

n−1
−1

= W(α)W (α) 2 , = (α I + D) ( − 1)j [Gi (α I + D)−1 ]j .
 j =0
⇒ ρ(M(α)) = ρ(M(α)) = W(α) 2 W (α) 2 < 1. (7)
Now,
Therefore, the MTTS iteration method converges to unique
solution of the regularized linear system (2).
(αI − Ti )(αI + Ti )−1
= (αI − D − Gi )(αI + Ti )−1 ,

n−1
= (αI − D − Gi )(αI + D)−1 ( − 1)j [Gi (αI + D)−1 ]j ,
CONVERGENCE CONDITION FOR MTTS j =0
METHOD ≈ (αI − D)(αI + D) [I − Gi (αI + D)−1 ],
−1

In this section, we find out the optimal parameter and dis-

cuss the inexact triangular and triangular (ITT) iterative by evaluating the above expression, and after ignoring
method by using the krylov subspace method as the inner higher order approximations and taking norm on both sides,
iterative process of the MTTS iteration method for the we get
Markov process [2]. Here, we find out the contraction factor
α on the lines of [2,29]. From the TSS, HSS, PSS iterative (α I − Ti )(α I + Ti )−1 2 ≈ (α I − Di )(α I + Di )−1 2 ,
methods and the theoretical approach in the above cases, W (α) 2 = max {(α − djj )(α + djj )−1 }.
1≤j ≤n
we observed that the iterative solution of (2) by modified
triangular and triangular splitting method of the precon- That is,
ditioned matrix A converge for any contraction factor α .
The iterative solution not only based the contraction factor W (α) 2 = {(α − σ1 )(α + σ1 )−1 }, (8)
√
α, but also based on the parameter . For the fixed value ⇒ W (α) 2 = σ1 σ1 = σ1 . (9)
of the contraction factor α, and different arbitrary values
of the parameter , we get the different solutions of the similarly 
MTTS iteration method for the numerical solutions of the   
W (α) 2 = = .
Markov process. Therefore, for the fast convergence in the 22 2
MTTS iteration method, it is the key importance to choose  
⇒ M(α) 2 = |W (α) 2 .|W (α) 2 = σ1 <1 if σ1 < 1.
the values of . Since the pre conditioned matrix A of the 2
regularized linear system (2) splits into triangular matrices
given (3).
Consider NUMERICAL RESULTS

A = [£ + D] + [U ] = T1 + T1 In this section, we find the optimal parameter α of the cir-

= [U + D] + [£] = T2 + T2 , culant stochastic rate matrix. We examine the effectiveness
of the MTTS method for the numerical solution in the case

where Ti and Ti , (i = 1, 2) are triangular matrices. Along of circulant Markov process and compare it with thatof the
the lines of [29], we find out the contraction factor as fol- Jacobi, TTS iteration methods. Consider the 4×4 circulant
lows: stochastic rate matrix
Let G1 = £ + D and G2 = U + D then Gi (i = 1, 2) are ⎡ ⎤
0.3 −0.1 −0.15 −0.05
lower and upper triangular matrices such that ⎢−0.05 0.3 −0.1 −0.15⎥
Q=⎢ ⎣−0.15 −0.05
⎥.
0.3 −0.1 ⎦
[Gi (α I + D)−1 ]n = [(α I + D)−1 Gi ]n = 0 for i = 1, 2. −0.1 −0.15 −0.05 0.4

6756
 International Journal of Applied Engineering Research, ISSN 0973-4562 Volume 13, Number 9 (2018) pp. 6751–6759
© Research India Publications, http://www.ripublication.com

FIGURE 1. Relative error of the MTTS method for the contraction factor α = 0.3 for different values of .

FIGURE 2. Relative errors of the Jacobi, TTS, and MTTS methods for the contraction factor α = 0.3, and  = 0.05 .

We consider only one case A = [£ + D] + [U ] = T1 + Figure 1 illustrates the results for the case of contrac-

T1 of MTTS splitting method and other methods would tion factor α = 0.3 choosen based on the analysis given in
follow. The initial distribution x (0) = [0, 0, 0, 1]T has been the above section, and different arbitrary values of . From
taken for the scheme (5) to compute relative error where this figure, we conclude that as  value increases, the rel-
absolute value lessthan or equal to 10−10 . Finally, the steady ative error decrease in three methods. Figure 2 depicts the
state probability distribution vector x of the preconditioned results of contraction factor α = 0.3, and  = 0.05, in the
linear system (2) obtained, and results are presented in table cases of MTTS and the existing methods. From the table 1
1. The error Vs number of iterations are presented in the and figures, we conclude that the MTTS iterative solution
Figs. 1–2. converges rapidly than the TTS and Jacobi’s methods.

6757
 International Journal of Applied Engineering Research, ISSN 0973-4562 Volume 13, Number 9 (2018) pp. 6751–6759
© Research India Publications, http://www.ripublication.com

Table 1. Comparison of relative error between Jacobi, TTS, [2] Z.-Z. Bai, G.H. Golub, L.-Z. Lu, J.F. Yin, Block triangu-
and MTTS lar and skew-Hermitian splitting methods for positive-
Number of Jacobi TTS MTTS definite linear systems, SIAM J. Sci. Comput., 20 (2005)
Iterations 844–863.
5 0.1402 0.1008 0.0698
[3] Z.-Z. Bai, G.H. Golub, M.K. Ng, Hermitian and skew-
10 0.0447 0.0195 0.0064
Hermitian splitting methods for non-Hermitian positive
15 0.0181 0.0044 6.3548e-004
20 0.0079 0.0010 6.4030e-005 definite linear systems, SIAM J. Matrix Anal. Appl.,
25 0.0036 2.4746e-004 6.4572e-006 24(2003) 603–626.
30 0.0016 5.8806e-005 6.5124e-007 [4] G. Barker, R.J. Plemmons, Convergent iterations for
35 7.5024e-004 1.3981e-005 6.5681e-008 computing stationary distributions of Markov chains,
40 3.4627e-004 3.3241e-006 6.6244e-009 SIAM J. Algebr. Disc. Meth., 7 (1986) 390–398.
[5] M. Benzi, B. Uar, Product preconditioning for Markov
chain problems, in: A.N. Langville, W.J. Stewart (Eds.),
CONCLUSIONS Proceedings of the 2006 Markov Anniversary Meeting,
Raleigh, NC, Boson Books, Charleston, SC, 2006, 239–
256.
In this paper, we present MTTS splitting iterative method
[6] A. Berman, R. Plemmons, Nonnegative Matrices in
for the regularized matrix of the system of circulant Mathematical Sciences, SIAM, Philadelphia, 1994.
stochastic matrix. We proved that the regularized matrix [7] C. Blondia, The N/G/l finite capacity queue, Commum.
is positive definite under specific condition. Moreover, we Statist.-Stochastic Models, 5(1989) 273–294.
proved that the regularized matrix is diagonally dominant [8] P. Brown, H.F. Walker, GMRES on (nearly) singular
and signature of the determinent of said matrix is positive. systems, SIAM J. Matrix Anal. Appl., 18(1997) 37–51.
We compared the relative error of TTS method, and Jacobi [9] R. Chan, W.K. Ching, Circulant preconditioners for
method with proposed MTTS method. We conclude that stochastic automata networks, Numer. Math., 87(2000)
35–57.
this method unconditionally converges to a unique solu-
[10] R. Bru, R. Cant, J.J. Climent, On M-multisplittings of
tion and the convergence rate is rapid when compared to the singular M-matrices and Markov chains, Numer. Linear
existing methods. Numerical results show that the MTSS Algebra Appl., 5 (1998) 299–311.
method performs very well than the existing Jacobi, and [11] M. Crovella,A. Bestavros, Self-Similarity in World Wide
TTS methods. Web traffic: evidence and possible causes, IEEE/ ACM
Trans. Networking, 5(1997) 835–846.
[12] W. Fischer and K. Meier-Hellsterm, The Markov-
modulated Poisson process (MMPP) Cookbook, Perf.
ACKNOWLEDGMENT Eval., 18(2)(1993) 149–171.
[13] H. Heffes, D. Lucantoni, A Markov modulated charac-
One of the five authors (Rajaiah Dasari) wishes to terization of packetized voice and data traffic and related
acknowledge University Grants Commission(UGC), Gov- statistical multiplexer performance, IEEE J. Sel. Areas
ernment of India for its funding under Dr. D.S. Commun., SAC-4(1986) 856–868.
[14] Kazarinoff, Nicholas D. Analytic inequalities. Courier
Kothari Postdoctoral Fellowship(Normal) scheme (No.F.4-
Corporation, 2014.
2/2006(BSR)/MA/15-16/0026) and authors acknowledge [15] L.A. Krukier, T.S. Martynova, Z.-Z. Bai, Product-type
Department of Science and Technology(DST) for fund- skew-Hermitian triangular splitting iteration methods
ing through Fund for Improvement of S&T Infrastruc- for strongly non-Hermitian positive definite linear sys-
ture(FIST) program sanctioned to the Department of tems, J. Comput. Appl. Math., 232(2009) 3–16.
Mathematics, Kakatiya University, Warangal with File [16] W.E. Leland, M.S. Taqqu, W. Willinger, and W.V. Wil-
No.SR/FST/MSI-101/2014. son, On the Self-Similar Nature of Ethernet Traffic
(Extended version), IEEE / ACM Trans. Networking.,
2(1994) 1–15.
[17] P. Malla Reddy, Chih-How Chang, Shou-Kuo Shao, and
REFERENCES Jingshown Wu, An Efficient Approximate Markovian
Model for Optical Packet Switches Employing Partial
[1] A. Andersen, B. Nielsen, A Markovian approach for Buffer Sharing Mechanism under Self-Similar Traffic
modeling packet traffic with long-range dependence, Input, Proceedings of IEEE HPSR-2007 held at Poly-
IEEE Journal on Selected Areas in Communications, technic University, Brooklyn, New York, USA, May
16(1998) 719–732. 30-June1,ISBN:1-4244-1206-4, IEEE Catalog number:

6758
 International Journal of Applied Engineering Research, ISSN 0973-4562 Volume 13, Number 9 (2018) pp. 6751–6759
© Research India Publications, http://www.ripublication.com

07EX1775C(2007).
[18] Marat Ibragimov, Conditions for Positive Definiteness of
M-Matrices, Linear and Multi Linear Algebra, Vol. 48,
pp. 93–106, 2000.
[19] H. Minc, Nonnegative Matrices, Wiyley, Newyork,
1988.
[20] V.Paxson, S. Floyd, Wide Area Traffic: The Failure
of Poisson Modeling, IEEE/ACM Trans. Networking.,
3(1995) 226–244.
[21] Rafael Bru, Maria T. Gass, Isabel Gimnez, and Jos A.
Scott. The Hadamard Product of a Nonsingular Gen-
eral H-Matrix and Its Inverse Transpose Is Diagonally
Dominant. Journal of Applied Mathematics, 2015.
[22] Rajaiah Dasari, Rajkumar L. P. and Malla Reddy Perati,
Steady state probability vector of positive definite reg-
ularized linear systems of circulant stochastic matrices,
vol. 65, no. 1, pp. 140–155, 2016.
[23] D. Rajaiah, P. Malla Reddy, “Loss Behavior of an
Internet Router with Self-Similar Input Traffic via Frac-
tal Point Process”, Proceedings of IEEE ICON-2013
held at Singapore, December 11-13, 978-1-4799-2084-
6/13(2013).
[24] D. Ranadheer, R. Ramesh, D. Rajaiah, and P. Malla
Reddy, Study of Delay and Loss Behavior of Internet
Router- Markovian, Modelling using Circulant Markov
Modulated Poisson Process(CMMPP), Applied Mathe-
matics, 5(2014), 512–519.
[25] Y. Saad, Iterative Methods for Sparse Linear Systems,
2nd ed. SIAM, Philadelphia, PA, 2003.
[26] S.K. Shao, P. Malla Reddy, M.G. Tsai, H.W. Tsao, and
Jingshown Wu, Generalized variance-based Markovian
fitting for self-similar traffic modeling, IEICE Trans.
Commun., E88-B(4)(2005) 1493–1502.
[27] W. J. Stewart, Introduction to Numerical Solution of
Markov Chains, Princeton University Press, Princeton,
1994.
[28] Te Wang, Hongbin Lv* and Haifeng Sang (2014). Esti-
mations for spectral radius of nonnegative matrices
and the smallest eigenvalue of M-matrices. Journal of
Inequalities and Applications, 2014.
[29] Chun Wen, H. Ting-Zhu, W. Chao, Triangular and skew-
symmetric splitting methods for numerical solutions of
Markov chains, Comp. and Math. With App., 62(2011)
4039–4048.
[30] T. Yoshihara, S. Kasahara, Y. Takahashi, Practical time-
scale fitting of self-similar traffic with Markov mod-
ulated Poisson process, Telecommun. Syst., 17(2001)
185–211.
[31] Zhang, C. Y., Yang, Y. Q., & Sun, Q. (2014). Con-
vergence of TTS iterative method for non-Hermitian
positive definite linear systems. Mathematical Problems
in Engineering, 2014.

6759