3 TWO NEW DIRECT METHODS FOR SOLVING LINEAR ALGEBRAIC EQUATIONS BASED ON ROW-ORTHOGONALIZATION∗ 4 QIN SHU† , SILIANG ZHAO‡ , AND YUNXIU YANG§ 5 6 7 8 9 10 11 12 13 14 Abstract. Two new direct methods for solving the linear algebraic equation Ax = b are proposed, and they include four steps. The first step is to convert A into a row-orthogonal matrix R using elementary row operations, i.e., PAx = Rx . The second step is to left-multiply b by the converting matrix P , which is Pb = b 0 . The third step is to reduce R to diagonal form, i.e., RR T (R T )−1 x = b 0 where D = RR T should be a diagonal matrix. The fourth step is to obtain the solution using x = R T D −1 b 0 . Compared with the existing direct methods, the clear advantages of our methods are that 1) the singularity of A can be easily judged through the row-orthogonal matrix R or the diagonal matrix D, 2) our methods outperform other direct methods regardless of A is singular or nonsingular, and 3) our methods can quickly calculate an upper bound of the relative error. The results also hold for the complex domain. 1 2 15 Key words. direct method, row orthogonalization, ill-conditioned equation 16 AMS subject classifications. 65F05, 65F25, 65G99 17 1. Introduction. 18 19 1.1. Origin of the problem. Consider the nth-order linear algebraic equation as follows: 20 (1.1) 21 22 23 where A ∈ R n×n and x , b ∈ R n×1 . Then to solve x , there are two types of methods that can be used. 1) Direct methods. This class of methods attempts to solve the problem using a finite sequence of operations. In the absence of rounding errors, direct methods would deliver an exact analytic solution. Examples of direct methods include Gaussian elimination and diagonal pivoting [4]. 2) Iterative methods. An iterative method is a mathematical procedure that uses an initial point to generate a sequence of improving approximate solutions for a class of problems in which the n-th approximation is derived from the previous approximations. Examples of iterative methods include the Jacobi method, the Gauss-Seidel method, the SOR (successive over relaxation)method, the Newton-Raphson method and the CG (conjugate gradient)method. As for the iterative methods, convergence is an unavoidable critical problem that may lead to the failure to calculate an acceptable solution. The termination criterion is also a key to an iterative methods specific implementation. To transform the original approximating operator to a better conditioned operator, a class of preconditioned iterative methods has been developed, such as PCG (preconditioned conjugate gradient)method. 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 Ax = b ∗ Submitted to the editors 26 April 2019. of Electrical Engineering and Information, Sichuan University, Chengdu, Sichuan, China.(shuqin@scu.edu.cn). ‡ School of Electrical Engineering and Information, Sichuan University, Chengdu, Sichuan, China.(361799391@qq.com). § South-West Institute of Technical Physics, Chengdu, Sichusn, China.(yangyang judy@126.com). † School 1 This manuscript is for review purposes only. 2 QIN SHU, SILIANG ZHAO, YUNXIU YANG 40 41 42 This paper mainly addresses direct methods. For comparison, we introduced a recursive method that has some similarities with our methods (see in 1.2). All of these methods can be applied into the process of algorithms in various fields. 43 44 45 1.2. Summary of the direct methods. According to the classifications of Abaffy and Spedikatto [1] and Duff and Reid [5], the direct methods for solving linear systems can be classified as follows. 1) In the first class of methods, the initial system is transformed into a system with smaller ranks in every step, and they include Gaussian elimination and diagonal pivoting [4]. In addition, there is another important class of methods named direct projection methods [2] that have been proved to be equal to Gaussian elimination [13]. 2) The second category includes methods for converting the matrix of the original system into another matrix. Solving the linear algebraic equations with the other matrix is easy and can be accomplished, for instance, by reducing the matrix to a triangular or diagonal form. Moreover, the accuracy of the regular direct methods is very limited, and they even fail to solve the ill-conditioned algebraic equations that are often encountered in engineering. In 1971, based on Kahans [7] generalization of a pivot that include 2 × 2 principal submatrices, Bunch [4] provided a diagonal pivoting method by selecting the pivot as being 1 × 1 or 2 × 2 when A is symmetric. In 1992, Yu et al. [12] proposed a recursive algorithm combining the Gram-Schmidt technique and the Kaczmarz iterative method; in the work, the authors decomposed the matrix A into LQ, where L, Q is a lower triangular matrix and row-orthogonal matrix, respectively. Then the Kaczmarz projection method is used to calculate the ultimate solution x , but it requires a long computing time for the decomposition of LQ, and the numerical error is big. In 1996, based on the LQ decomposition method, Kim et al. [8] presented a unique method that could improve the accuracy of the solution by enhancing the condition of the ill-conditioned matrix. In this paper, we converted the matrix A into a row-orthogonal matrix, and then the row-orthogonal matrix was transformed into a diagonal matrix. Therefore we name our methods the Orthogonal Elimination (OE) method and the Improved Orthogonal Elimination (IOE) method, respectively. 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 2. Two new methods. 73 74 75 2.1. The principle of new methods. First, we convert the matrix A into the row-orthogonal matrix R using elementary row operations, such as the Gram-Schmidt technique. Let P be the converting matrix. Then, 76 (2.1) 77 Second, we perform the same row operations as A for b, which is 78 (2.2) 79 With 2.1 and 2.2, we can obtain Rx = b. Third, suppose that x = R T x 0 . Then, 80 (2.3) 81 Finally, we can obtain the solution x as 82 (2.4) P Ax = Rx P b = b0 RRT x0 = b0 x = RT (RRT )−1 b0 This manuscript is for review purposes only. TWO NEW DIRECT METHODS BASED ON ROW-ORTHOGONALIZATION 3 83 84 85 From the above, compared with the Gauss-Jordan elimination method, no back substitution is needed in our new methods, so the solution is more stable. The following is the specific implementation of our new methods. 86 87 2.2. Orthogonal Elimination(OE) method. Here is the Orthogonal Elimination(OE) methods specific implementation. Let A = [ã1 , ã2 , . . . , ãn ]T , b = (b1 , b2 , . . . , bn )T , and x = (x1 , x2 , . . . , xn )T . Then we can obviously obtain the equations ã1 x = b1 , ã2 x = b2 , . . . , ãn x = bn . Then, we convert A and b into R = [r̃1 , r̃2 , . . . , r̃n ]T and b 0 = (b10 , b20 , . . . , bn0 )T , respectively, by performing the same Gram-Schmidt technique. r̃i and b 0 i can be calculated as follows. When i = 1 , and r̃1 = ã1 , we compute r̃i (i ≥ 2) as 88 89 90 91 92 93 (2.5) r̃ = ãi − i−1 X j =1 94 95 hãi , r̃j i r̃j hr̃j , r̃j i Then, we right-multiply by x on the both sides of equation (2.5) as (2.6) r̃i x = ãi x − i−1 X j =1 96 hãi , r̃j i r̃j x hr̃j , r̃j i With Ax = b and Rx = b 0 , we obtain 97 98 (2.7a) ãi x = bi 99 (2.7b) r̃i x = bi0 100 101 Then, we substitute equations (2.7a) and (2.7b) into equation (2.6), where b10 = b1 . (2.8) bi0 = bi − i−1 X j =1 102 103 104 105 hãi , r̃j i ˜0 b hr̃j , r̃j i j According to the recursions of equations (2.6) and (2.8), we obtain Rx = b 0 (2.9) Suppose that x = R T x 0 . Then, RR T x 0 = b 0 (2.10) where RR T should be a diagonal matrix in rd(1,1) 0 0 r d(2,2) RR T = . 107 (2.11) .. .. . 106 0 0 exact arithmetic. ··· 0 ··· 0 .. . 0 ··· rd(n,n) From matrix multiplication theorem, x 0 can be computed as 0 b1 rd(1,1) b02 rd(2,2) 0 109 (2.12) x = .. . 0 bn rd(n,n) 108 This manuscript is for review purposes only. 4 QIN SHU, SILIANG ZHAO, YUNXIU YANG Finally, we obtain x by substituting x 0 into x = R T x 0 . Next, we combine equations (2.5), (2.8), (2.11), and (2.12), the new direct method, the Orthogonal Elimination(OE) algorithm, can be given as follows. 113 Algorithm 2.1 Orthogonal Elimination(OE) 110 111 112 114 115 116 117 118 119 120 121 122 123 set r̃1 = ã1 , b10 = b1 for i = 1 ; i < n; i + + do for j = 1 ; j < i − 1 ; j + + do Pi−1 hã ,r̃ i r̃i = ãi − j =1 hr̃ji ,r̃jj i r̃j Pi−1 hã ,r̃ i bi0 = bi − j =1 hr̃ji ,r̃jj i b˜j0 end for end for rd(1,1) 0 0 r d(2,2) Compute RR T = . .. . . . ··· ··· .. . ··· 0 0 0 b1 rd(1,1) b02 rd(2,2) 0 Compute x = .. . 0 bn rd(n,n) Finally, x = R T x 0 0 0 0 rd(n,n) 2.3. Improved Orthogonal Elimination(IOE) method. From the analysis of equations (2.5) and (2.8), it can be seen that the value of hr̃j , r̃j i may become increasingly smaller as the recursion continues until the numerical reliability is lost. In view of this situation, we proposed an improved algorithm. The specific process is as follows. 129 We multiply hr̃j , r̃j i (j = 1 , 2 , . . . , i − 1 ) on both sides of equations (2.5) and (2.8) 130 in order to improve the accuracy of the calculation when the denominator is getting 131 increasingly closer to zero. 124 125 126 127 128 132 (2.13a) r̃i i−1 Y hr̃j , r̃j i = ãi j =1 i−1 Y hr̃j , r̃j i − i−1 X i−1 Y hr̃k , r̃k i hãi , r̃j i r̃j j =1 k =1 ,k 6=j j =1 133 134 (2.13b) bi0 i−1 Y j =1 hr̃j , r̃j i = bi0 i−1 Y j =1 hr̃j , r̃j i − i−1 X i−1 Y hr̃k , r̃k i hãi , r̃j i bj0 j =1 k =1 ,k 6=j Qi−1 Qi−1 Suppose that r̃i0 = r̃i j =1 hr̃j , r̃j i, bi00 = bi0 j =1 hr̃j , r̃j i, R 0 = [r̃10 , r̃20 , . . . , r̃n0 ]T , 00 T 136 and b 00 = [b100 , b200 , . . . , bn ] . According to the vector orthogonal theorem, R 0 is still 137 a row orthogonal matrix after the above transformation. Therefore, 0 rd(1,1) 0 ··· 0 0 0 rd(2,2) ··· 0 R 0 R 0T = . 138 (2.14) . . .. .. .. 0 0 0 0 · · · rd(n,n) 135 139 By referring to equations (2.9), (2.10) and (2.12), similar formula can be obtained: This manuscript is for review purposes only. TWO NEW DIRECT METHODS BASED ON ROW-ORTHOGONALIZATION 5 140 (2.15a) R0x = D (2.15b) R 0 R 0T x 00 = D (2.15c) . 0 d1 rd(1,1) . 0 d2 rd(2,2) x 00 = .. . . 0 dn rd(n,n) 147 (2.15d) x = R 0T x 00 148 149 150 151 152 The problem that the denominators hr̃j , r̃j i in equations 2.5 and 2.8 tend to zero is effectively solved by the above procedures. Q We will show this improvement in 5. i−1 However, in actual computations, the value of j =1 hr̃j , r̃j i tending to infinity is also possible as the recursion continues. Given this, we can combine the OE method with the IOE method. That is, if the Euclidean norm of hr̃j , r̃j i is less than 1, we use the IOE method(IOE). Conversely, we employ the OE method. In particular, the value of hr̃j , r̃j i is equal to 0 for some j ∈ [2 , n] when the matrix A of the system is singular. Therefore, regularization is needed in equation (2.15c), i.e. . 0 b001 rd(1,1) +ε . 00 0 b2 rd(2,2) +ε 00 (2.16) x = .. . . 141 142 143 144 145 146 153 154 155 156 157 b00n 158 159 160 0 rd(n,n) +ε where 0 < ε 1 . Until now, the steps of the improved method can be detailed as follows. Algorithm 2.2 Orthogonal Elimination(IOE) 169 set r̃1 = ã1 , b10 = b1 for i = 1 ; i < n; i + + do for j = 1 ; j < i − 1 ; j + + do if kr̃j k2 < 1 then set r̃i = hr̃j , r̃j i r̃i = hr̃j , r̃j i ãi − hãi , r̃j i r̃j set bi0 = hr̃j , r̃j i bi0 = hr̃j , r̃j i bi − hãi , r̃j i b˜j0 else Pi−1 hã ,r̃ i set r̃ = ãi − j =1 hr̃ji ,r̃jj i r̃j Pi−1 hãi ,r̃j i ˜0 set b 0 = bi − b 170 171 172 end if end for end for 161 162 163 164 165 166 167 168 i j =1 hr̃j ,r̃j i j This manuscript is for review purposes only. 6 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 QIN SHU, SILIANG ZHAO, YUNXIU YANG Compute RR T rd(1,1) 0 = . .. 0 rd(2,2) .. . ··· ··· .. . 0 0 ··· . 0 0 b r +ε 1 . d(1,1) 0 0 +ε b2 rd(2,2) Compute x 0 = .. . . 0 b0n rd(n,n) +ε 0 0 0 rd(n,n) Finally, x = R T x 0 3. Singularity judgment of matrix A. According to the equations 2.1 and 2.2, PAx = Rx = Pb = b 0 where P ∈ R n×n is a lower triangular matrix that is obtained by the elementary row operations. Since the elementary operations do not change the rank of a matrix [6], rank (PA) = rank (A), i.e., rank (R) = rank (A). According to the fundamental equalities involving the rank, we know that rank (RR T ) = rank (R), and thus rank (D) = rank (R) = rank (A), where D = RR T is a diagonal matrix. Then, the singularity of A can be judged by analyzing the row-orthogonal matrix R or diagonal matrix D. If A is singular, let the row-vector a~i be linearly dependent with the first i − 1 linearly independent row-vectors. Without the rounding error, the row-orthogonal result of a~i would be a zero vector. Therefore, we can judge taht A is singular when the Euclidean norm of a row of R tends to zero. Correspondingly, there is a diagonal element that tends to zero in D. 4. Error analysis of OE and IOE. Inspired by the work of Bunch [3], [4] and Wilkinson [9], [10], [11]. we presented expressions of the error and the relative error. 190 We consider x0 as the exact solution of the system Ax = b, According to equations 191 2.1 and 2.2, its easy to know that Rx0 = b 0 and −1 192 (4.1) x = b0 RR T R T 188 189 which implies that x = R T (RR T )−1 b 0 . Finally, we substitute the equation Rx0 = b 0 into equation 4.1, and we obtain the computational expression of x . −1 195 (4.2) x = R T RR T Rx0 193 194 201 Note that the elements of RR T were calculated with finite precision, and thus RR would no longer be strictly diagonal due to the rounding error. Nevertheless, we still treat this variable as a diagonal matrix and define it as diag RR T . If desired, the same regularization as was mentioned in 2.3 is also taken into account and the error is given by h −1 i (4.3) x0 − x = I − R T diag RR T + εI R x0 202 where 0 < ε 1 . The relative error that is usually a concern in engineering will be 203 (4.4) 196 197 198 199 200 204 205 T −1 kx0 − x k2 ≤ I − R T diag RR T + εI R kx0 k2 F Through the analysis of the above formulas, the closer RR T is to diagonal matrix, the smaller the error that was caused by the rounding error will be. Without This manuscript is for review purposes only. TWO NEW DIRECT METHODS BASED ON ROW-ORTHOGONALIZATION 7 −1 the rounding error, the value of the error expression [I − R T diag RR T R]x0 should be zero. Numerical experiments (see section 6)proved the validity of relative error expression. As a result, equation (4.4) can be a useful estimation before the 209 calculation. 206 207 208 210 5. Comparisons of operation counts. 5.1. The operation counts of the OE method and the upper bound of the IOE’s operation counts. 213 • Step 1. The operation counts of the r̃i (i = 2 , 3 , . . . , n) is shown in Table 1 (OE) and Table 2 (IOEs upper bound). 211 212 Table 1 Operation counts of the OE to calculate r̃i (i = 2 , 3 , . . . , n) The i-th row orthogonal vector Number of multiplications Number of additions 2 3 ... i ... n 3n + 0 × 2n + 1 3n + 1 × 2n + 2 ... 3n + (i − 2 ) × 2n + i − 1 ... 3n + (n − 2 ) × 2n + n − 2 2n − 2 + 1 × (n − 1 ) + 1 × n 2n − 1 + 2 × (n − 1 ) + 2 × n ... 2n − 2 + (i − 1 ) × (n − 1 ) + (i − 1 ) × n ... 2n − 2 + (n − 1 ) × (n − 1 ) + (n − 1 ) × n 214 Table 2 Operation counts(upper bound) of the IOE to calculate r̃i (i = 2 , 3 , . . . , n) The i-th row orthogonal vector Number of multiplications Number of additions 2 3 ... i ... n 4n 3 × (2n + 1 ) ... i × (2n + 1 ) ... n × (2n + 1 ) 2 × (n − 1 ) + n 3 × (n − 1 ) + 2n ... i × (n − 1 ) + (i − 1 ) × n ... n × (n − 1 ) + (n − 1 ) × n 223 224 225 226 According to Table 1 and Table 2, the operation counts when we calculate r̃i are 3 2 as follows. For the OE method, the number of multiplications is n − n 2 −2n + 1 /2 , and the number of additions is n 3 + n 2 2 − 7n/2 + 2 . For the IOE methods upper bound, the number of multiplications is n 3 + 3n 2 2 − 3n/2 − 3 , and the number 3 2 of additions is n − n 2 − 3n/2 + 1 . • Step 2. When we calculate bi0 , the number of multiplications and additions is i − 1 , respectively. Then, when i ranges from 2 to n, the number of multiplications and additions both are n(n − 1 )/2 . • Step 3. To acquire equation (2.7b), equation (2.8) and x = R T x 0 , the numbers of calculation that are required are as follows. The number of multiplications is 2n 2 + n, and the number of additions is 2n(n − 1 ). In summary, the total numbers of this papers algorithm are shown in Table 3. 227 228 229 230 5.2. Operation counts of the LQ decomposition with kaczmarz iteration [12]. An arbitrary nonsingular coefficient matrix A can be decomposed into the following form, A = LQ. Here, Q = (ã10 , ã20 , . . . , ãn0 )T is a row-orthogonal matrix and ãi0 can be calculated using the Gram-Schmidt technique as follows. 215 216 217 218 219 220 221 222 This manuscript is for review purposes only. 8 QIN SHU, SILIANG ZHAO, YUNXIU YANG Table 3 Operation counts of this papers algorithm Method Number of multiplications Orthogonal Elimination 3n 1 n + 2n − + 2 2 3 2 n + 4n − n − 3 Improvement of Orthogonal Elimination(upper bound) 3 2 Pi−1 hãi ,ãj0 i 0 j =1 hã 0 ,ã 0 i ãj where i ∈ [1 , n]. In addition, L is j j 232 matrix that can be calculated as below, 1 0 0 ··· 0 hã2 ,ã10 i 0 0 1 0 ··· 0 hã1 ,ã1 i ã ,ã 0 0 hã , ã i 3 h 3 1i 2 1 ··· 0 233 L= hã10 ,ã10 i hã20 , ã20 i . .. .. .. .. . . . . . . hãn ,ã 0 i 0 hãn ,ã20 i hãn ,ã30 i hãn ,ãn−1 i 1 · · · 0 0 0 0 0 0 0 0 hã1 ,ã1 i hã2 ,ã2 i hã3 ,ã3 i hãn−1 ,ãn−1 i 231 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 ã 0 = ãi − Number of additions n 3 + 3n 2 − 6n + 2 n 3 + 2n 2 − 4n + 1 a lower-triangular 0 0 0 .. . 1 Then, the solution can be obtained by two steps. • Step 1. We solve the equation Ly = b to obtain vector ỹ = (y1 , y2 , . . . , yn )T , Pi−1 hãi ,ã 0 i where yi = bi − j =1 ã 0 ,ãj0 ỹj0 (i = 1 , 2 , 3 , . . . , n). h j ji • Step 2. Then, we solve the equation Qx = y to find the final solution. Since matrix Q is row-orthogonal, this equation can be solved using Kaczmarz’s method ã 0 as follows: x0 is arbitrary, x̃i−1 = ã 0 ,ãi 0 (yi − hãi , x̃i−1 i) and we set the number of h i ii iteration steps to n. According to the above two steps, the operation counts of ai0 are the same as that shown in Table 1, i.e., the numberof multiplications is n 3 − n 2 2 − 2n + 1 /2 and the number of additions is n 3 + n 2 2 − 7n/2 + 2 . Then, the operation counts to ob 3 2 tain yi and x̃i are the followding: the number of multiplications is (n +n ) 2 +2n +1 and the number of additions is (3n 2 + 5n − 4 ) 2 . In other words, the total operation counts of the LQ decomposition with the Kaczmarz iteration algorithm [12] are as follows: the number of multiplications is 3n 3 2 + 3 /2 and the number of additions is n 3 + 2n 2 − n. 5.3. Operation counts of the Gauss-Jordan elimination with the partial pivoting. The Gauss-Jordan elimination with the partial pivoting method is mainly divided into three steps, including finding the pivot, the elimination on process and the substitution process. • Step 1. Partial pivoting requires one to bring the largest element in the first column of the reduced matrix into the leading diagonal position. Therefore, the number of comparisons is n(n − 1 )/2 . • Step 2. The operation counts of the elimination process are shown in Table 4. 257 From table 4, we see that the number of multiplications is (2n 3 + 3n 2 − 5n) 6 , 259 and the number of additions is (n 3 − n) 3 . 260 • Step 3. The operation counts of back substitution process are as follows: the 261 number of multiplications is n(n + 1 )/2 and the number of additions is n(n − 1 )/2 . 258 This manuscript is for review purposes only. TWO NEW DIRECT METHODS BASED ON ROW-ORTHOGONALIZATION 9 Table 4 Operation counts of the elimination process of the Gauss-Jordan elimination method k-th row orthogonal vector Number of multiplications Number of additions 1 2 ... k ... n-1 (n + 1 )(n − 1 ) n(n − 2 ) ... (n − k + 2 )(n − k ) ... 3 ×1 n(n − 1 ) (n − 1 )(n − 2 ) ... (n − k + 1 )(n − k ) ... 2 ×1 Finally, the total operation counts for the Gauss-Jordan elimination method are as follows: the number of multiplications is n 3 3 + n 2 − n/3 and the number of 264 additions is n 3 3 + n 2 2 − 5n/6 . 262 263 265 266 267 268 5.4. Operation counts of Bunchs diagonal pivoting method [4]. Based on P(j ) Bunch’s idea [4], let denote the summation over those indices i , where 1 6 i 6 n, such that if A(i) exists, then it yields a j × j pivot, where j = 1 or 2 . Let p be the number of 1 × 1 pivots that is used. Then, the operation counts for solving Ax = b are shown in Table 5. Table 5 Operation counts of the diagonal pivoting method Operations multiplications additions Exact Upper bound P(2 ) 3n 2 5n n3 + − +p+3 i 6 2 3 3 P(2 ) 13n n + n2 − +p+3 i 6 6 9n 2 n n3 + − 6 4 6 5n 2 5n 1 n3 + − + 6 4 3 4 269 270 271 272 Through the comparison, the operation counts of all the abovementioned methods are on the order of n 3 , but the operation counts of our methods are slightly larger than those of the other methods, since the LQ decomposition with Kaczmarz iteration [12] is the most complex. For clarity, we summarized the comparison below in Table 6. Table 6 Comparison of the methods for solving Ax = b Method Orthogonal Elimination Improvement of Orthogonal Elimination(upper bound) LQ decomposition with the Kaczmarz iteration Elimination(partial pivot) Diagonal pivoting(upper bound) Number of multiplications 3n 1 n + 2n − + 2 2 3 2 n + 4n − n − 3 3 3 3n + 2 2 3 n n + n2 − 3 3 3 2 n 9n n + − 6 4 6 3 2 Number of additions n 3 + 3n 2 − 6n + 2 n 3 + 2n 2 − 4n + 1 n 3 + 2n 2 − n n3 n2 5n + − 3 2 6 3 2 n 5n 5n 1 + − + 6 4 3 4 273 274 275 6. Experimental results. To further demonstrate the validity of our methods in some computational situations, especially in the case of ill-conditioned equations This manuscript is for review purposes only. 10 276 277 278 279 280 281 282 QIN SHU, SILIANG ZHAO, YUNXIU YANG which are often encountered in scientific calculation, we introduced the following three examples. Problem 1. Solve the ill-conditioned linear equations Ax = b, where 23.0 5.0 7.0 6.0 5.0 32.0 7.0 10.0 8.0 7.0 A= 6.0 8.0 10.0 9.0 , b = 33.0 31.0 5.0 7.0 9.0 10.0 Obviously, the theoretical solution is x = (1 .0 , 1 .0 , 1 .0 , 1 .0 )T , and the conditional number cond (A) = kAk2 A−1 2 is approximately 2984 when we only take the integer part. Then, the comparison results(with 4 significant digits) of the OE and IOE with the Gauss-Jordan elimination(partial pivot) method are listed in Table 7. Table 7 Comparison of the results of this paper with the elimination(partial pivot) method(4 significant digits) 283 284 285 286 287 288 289 290 Solution OE IOE Elimination(partial pivot) Exact solution x1 x2 x3 x4 1 .011 0 .9948 0 .9958 0 .9998 1 .011 0 .9948 0 .9958 0 .9998 1 .039 0 .9764 0 .9898 1 .006 1 .000 1 .000 1 .000 1 .000 From Table 7, we can conclude that in the case of 4 significant digits, the results of our methods are the same, and the relative errors of our methods and the GaussJordan elimination method(partial pivot) are 0 .6454 % and 2 .355 %, respectively. Apparently, the accuracy of our methods is better than that of the Gauss-Jordan elimination(partial pivot). Problem 2. Let A be a Hilbert matrix as follows: 1 1/2 ··· 1/n 1/2 1/3 ··· 1/(n + 1) 1 A= . , i.e., A(i , j ) = . .. . . . . i +j −1 . . . . 1/n 1/(n + 1) · · · 1/(2n − 1) i , j ∈ [1 , n] 297 We select the exact solution x = (1 , 1 , . . . , 1 )T , and b = Ax . Then, we take the order n of the equation from 5 to 50 with the adjacent interval 5, and the comparison results of the relative error of our my methods with the three other methods are presented in Table 8. Note that since the accuracy of the LQ decomposition [12] is affected by the initial iteration point xini , we ran it 10 times and took the average value. In the table, the upper error bound denotes the value of equation −1 −1 I − R T diag RR T R or I − R T diag RR T + εI R . 298 299 300 301 302 Table 8 shows that the regular Gauss-Jordan elimination(partial pivot) method and diagonal pivoting method can not find any acceptable numerical solution when n > 10 . By applying the OE and IOE methods, more accurate numerical solution can be obtained for any n < 50 than those provided by the LQ decomposition [12], and the IOE is more stable than the OE as the order increases. In addition, the upper 291 292 293 294 295 296 F This manuscript is for review purposes only. F TWO NEW DIRECT METHODS BASED ON ROW-ORTHOGONALIZATION 11 Table 8 Comparison of the relative error of this paper with three other methods 303 304 305 Unknowns OE(%) Upper error bound(%) IOE(%) Upper error bound(%) LQ [12](%) Elimination(partial pivot)(%) Diagonal pivoting(%) 5 10 15 20 25 30 35 40 45 50 2.16 × 10−07 1.74 × 10−01 2.53 × 10+00 9.55 × 10+00 7.54 × 10+00 4.40 × 10+01 6.72 × 10+01 8.42 × 10+01 2.65 × 10+01 1.99 × 10+02 1.09 × 10−05 3.46 × 10+02 8.46 × 10+03 1.34 × 10+03 1.75 × 10+03 2.33 × 10+03 2.80 × 10+03 3.27 × 10+03 3.78 × 10+03 4.31 × 10+03 2.46 × 10−01 2.13 × 10+00 4.65 × 10+00 7.28 × 10+00 9.84 × 10+00 1.23 × 10+01 1.45 × 10+01 1.66 × 10+01 1.86 × 10+01 2.05 × 10+01 9.98 × 10+01 2.24 × 10+02 3.16 × 10+02 3.87 × 10+02 4.47 × 10+02 5.00 × 10+02 5.48 × 10+02 5.92 × 10+02 6.32 × 10+02 6.71 × 10+02 2.15 × 10+01 2.88 × 10+01 3.46 × 10+01 3.52 × 10+01 3.69 × 10+01 3.73 × 10+01 4.02 × 10+01 3.95 × 10+01 4.12 × 10+01 4.46 × 10+01 3.49 × 10−11 4.21 × 10−04 1.72 × 10+03 5.93 × 10+03 1.29 × 10+05 3.26 × 10+04 8.10 × 10+04 2.35 × 10+05 1.11 × 10+05 6.68 × 10+04 2.92 × 10−10 1.46 × 10−02 4.63 × 10+02 2.82 × 10+03 1.42 × 10+04 1.20 × 10+04 2.65 × 10+07 8.82 × 10+03 9.96 × 10+03 1.90 × 10+04 bound of the error analysis expression that is obtained in chapter 4 is proved to be valid. Problem 3. Consider the ill-conditioned linear equations Ax = b, where A = (aij )n×n , b = [b1 , b2 , b3 , . . . , bn ]T 306 ( 1 aij = 1 + p2 bi = 307 n X i 6= j i=j aik i, j = 1, 2, 3, . . . , n i = 1, 2, 3, . . . , n k =1 Let p = 0 .5 × 10 −10 and randomly generate the real solution between [0 .5 , 1 .5 ]. Then, take a value for the order n of the equation from 100 to 500 with the adjacent 310 interval 100 . Table 9 shows the comparison results of the relative error of our method 311 with the three other methods. We also ran the LQ decomposition [12] 10 times and 312 took the average. 308 309 Table 9 Comparison of the relative error of this paper with three other methods Unknowns OE(%) IOE(%) Upper error bound(%) LQ [12](%) Elimination(partial pivot)(%) Diagonal pivoting(%) 100 200 300 400 500 N aN N aN N aN N aN N aN 4.71 × 10+01 4.97 × 10+01 5.10 × 10+01 4.83 × 10+01 4.90 × 10+01 9.95 × 10+02 1.41 × 10+03 1.73 × 10+03 2.00 × 10+03 2.23 × 10+03 1.19 × 10+01 9.40 × 10+01 8.67 × 10+01 7.55 × 10+01 7.52 × 10+01 N aN N aN N aN N aN N aN N aN N aN N aN N aN N aN Obviously, the IOE method has the best performance, while the other three methods failed in their calculations. Furthermore, this result provides Qi−1 an example of a situation in which the OE method can not work anymore as j =1 hr˜j , r˜j i increases. 316 Finally, the Upper error bound is proved to be valid for estimating error range once 317 again. 313 314 315 318 319 320 321 322 323 7. Conclusions. Two new direct methods based on row orthogonalization have been developed to solve the linear algebraic equation Ax = b. Compared with the regular direct method of the Gauss-Jordan elimination(partial pivot) method, diagonal pivoting [4], and the recursive LQ decomposition method presented in [12], our methods for solving the highly ill-conditioned linear algebraic equations outperform the other three methods whether the coefficient matrix A is singular or nonsingular. This manuscript is for review purposes only. 12 QIN SHU, SILIANG ZHAO, YUNXIU YANG A useful upper bound of the relative error is presented, which is an important estimation in engineering calculations. Another meaningful result is that the singularity of matrix A can be easily judged using the row-orthogonal matrix R or the diagonal 327 matrix D. In addition, the row-orthogonal matrix R can be used to find the inverse 328 of nonsingular coefficient matrix A or the pseudo-inverse of the singular coefficient 329 matrix A. 324 325 326 330 REFERENCES 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 [1] J. Abaffy and E. Spedicato, ABS projection algorithms: mathematical techniques for linear and nonlinear equations, Prentice-Hall, Inc., 1989. [2] M. Benzi and C. D. Meyer, A direct projection method for sparse linear systems, SIAM Journal on Scientific Computing, 16 (1995), pp. 1159–1176. [3] J. R. Bunch, Analysis of the diagonal pivoting method, SIAM Journal on Numerical Analysis, 8 (1971), pp. 656–680. [4] J. R. Bunch and B. N. Parlett, Direct methods for solving symmetric indefinite systems of linear equations, SIAM Journal on Numerical Analysis, 8 (1971), pp. 639–655. [5] I. S. Duff and J. K. Reid, A comparison of some methods for the solution of sparse overdetermined systems of linear equations, IMA Journal of Applied Mathematics, 17 (1976), pp. 267–280. [6] R. A. Horn and C. R. Johnson, Matrix analysis, Cambridge university press, 2012. [7] W. Kahan, Numerical linear algebra, Canadian Mathematical Bulletin, 9 (1966), pp. 757–801. [8] H. J. Kim, K. Choi, H. Lee, H. Jung, and S. Hahn, A new algorithm for solving ill conditioned linear systems, IEEE Transactions on Magnetics, 32 (1996), pp. 1373–1376. [9] J. H. Wilkinson, Error analysis of direct methods of matrix inversion, Journal of the ACM (JACM), 8 (1961), pp. 281–330. [10] J. H. Wilkinson, The algebraic eigenvalue problem, vol. 662, Oxford Clarendon, 1965. [11] J. H. Wilkinson, Rounding errors in algebraic processes, Courier Corporation, 1994. [12] X. Yu, N. Loh, and W. Miller, New recursive algorithm for solving linear algebraic equations, Electronics Letters, 28 (1992), pp. 2069–2071. [13] A. I. Zhdanov, A direct sequential method for solving systems of linear algebraic equations, Dokl. Akad. Nauk, 356 (1997), pp. 442–444. This manuscript is for review purposes only.