Dense LU factorization and its error analysis

Transcription

1 Dense LU factorization and its error analysis Laura Grigori INRIA and LJLL, UPMC February 2016

2 Plan Basis of floating point arithmetic and stability analysis Notation, results, proofs taken from [N.J.Higham, 2002] Direct methods of factorization LU factorization Error analysis of LU factorization - main results Block LU factorization 2 of 33

3 Plan Basis of floating point arithmetic and stability analysis Notation, results, proofs taken from [N.J.Higham, 2002] Direct methods of factorization 3 of 33

4 Norms and other notations A F = n i=1 j=1 n a ij 2 A 2 = σ max (A) n A = max 1 i n a ij A 1 = max 1 j n j=1 n a ij i=1 Inequalities x y and A B hold componentwise. 4 of 33

5 Floating point arithmetic The machine precision or unit roundoff is u The maximum relative error for a given rounding procedure u is of order 10 8 in single precision, in double precision Another definition: the smallest number that added to one gives a result different from one The evaluation involving basic arithmetic operations +,,, / in floating point satisfies fl(x op y) = (x op y)(1 + δ), δ u 5 of 33

6 Relative error Given a real number x and its approximation ˆx, the absolute error and the relative errors are E abs (ˆx) = x ˆx, E rel (ˆx) = x ˆx x (1) The relative error is scale independent Some examples, outline the difference with correct significant digits x = , ˆx = , E rel (ˆx) = x = , ˆx = , E rel (ˆx) = When x is a vector, the componentwise relative error is max i x i ˆx i x i 6 of 33

7 Backward and Forward errors Consider y = f (x) a scalar function of a real scalar variable and ŷ its approximation. Ideally we would like the forward error E rel (ŷ) u Instead we focus on the backward error, For what set of data we have solved the problem? that is we look for min x such that ŷ = f (x + x) 7 of 33

8 Condition number Assume f is twice continuously differentiable, then ŷ y = f (x + x) f (x) = f (x) x + ( ŷ y xf ) (x) x = y f (x) x + O(( x)2 ) f (x + τ x) ( x) 2, τ (0, 1) 2! The condition number is c(x) = xf (x) f (x) Rule of thumb When consistently defined, we have forward error condition number backward error 8 of 33

9 Preliminaries Lemma (Lemma 3.1 in [N.J.Higham, 2002]) If δ i u and ρ i = ±1 for i = 1 : n, and nu < 1, then n (1 + δ i ) ρ i = 1 + Θ n, Θ n nu 1 nu = γ n i=1 Other notations γ n = cnu 1 cnu 9 of 33

10 Inner product in floating point arithmetic Consider computing s n = x T y, with an evaluation from left to right. We denote different errors as 1 + δ i 1 ± δ ŝ 1 = fl(x 1 y 1 ) = x 1 y 1 (1 ± δ) ŝ 2 = fl(ŝ 1 + x 2 y 2 ) = (ŝ 1 + x 2 y 2 (1 ± δ))(1 ± δ) = x 1 y 1 (1 ± δ) 2 + x 2 y 2 (1 ± δ) 2. ŝ n = x 1 y 1 (1 ± δ) n + x 2 y 2 (1 ± δ) n + x 3 y 3 (1 ± δ) n x n y n (1 ± δ) 2 After applying the previous lemma, we obtain ŝ n = x 1 y 1 (1 + Θ n ) + x 2 y 2 (1 + Θ n) x n y n (1 + Θ 2 ) 10 of 33

11 Inner product in FP arithmetic - error bounds We obtain the following backward and forward errors ŝ n = x 1 y 1 (1 + Θ n ) + x 2 y 2 (1 + Θ n) x n y n (1 + Θ 2 ) fl(x T y) = (x + x) T y = x T (y + y), x γ n x, y γ n y, x T y fl(x T y) n γ n x i y i = γ n x T y i=1 High relative accuracy is obtained when computing x T x No guarantee of high accuracy when x T y << x T y 11 of 33

12 Plan Basis of floating point arithmetic and stability analysis Direct methods of factorization LU factorization Block LU factorization 12 of 33

13 Algebra of the LU factorization LU factorization Compute the factorization PA = LU Example Given the matrix Let M 1 = A = , M 1 A = of 33

14 Algebra of the LU factorization In general I k 1 1 A (k+1) = M k A (k) := m k+1,k 1 A (k), where m n,k 1 M k = I m k e T k, M 1 k = I + m k e T k where e k is the k-th unit vector, ei T m k = 0, i k The factorization can be written as M n 1... M 1 A = A (n) = U 14 of 33

15 Algebra of the LU factorization We obtain A = M M 1 n 1 U = (I + m 1 e1 T )... (I + m n 1 en 1)U T ( ) n 1 = I + m i ei T U = i=1 1 m U = LU m n1 m n of 33

16 The need for pivoting For stability, avoid division by small diagonal elements For example A = (2) has an LU factorization if we permute the rows of matrix A PA = = (3) Partial pivoting allows to bound the multipliers m ik 1 and hence L 1 16 of 33

17 Existence of the LU factorization Theorem Given a full rank matrix A of size m n, m n, the matrix A can be decomposed as A = PLU where P is a permutation matrix of size m m, L is a unit lower triangular matrix of size m n and U is a nonsingular upper triangular matrix of size n n. Proof: simpler proof for the square case. Since A is full rank, there is a permutation P 1 such that P 1 a 11 is nonzero. Write the factorization as ( ) ( ) ( ) a11 A P 1 A = a11 A = 12 A 21 A 22 A 21 /a 11 I 0 A 22 a 1 11 A, 21A 12 where S = A 22 a 1 11 A 21A 12. Since det(a) 0, then det(s) 0. Continue the proof by induction on S. 17 of 33

18 Solving Ax = b by using Gaussian elimination Composed of 4 steps 1. Factor A = PLU, (2/3)n 3 ) flops 2. Compute P T b to solve LUx = P T b 3. Forward substitution: solve Ly = P T b, n 2 flops 4. Backward substitution: solve Ux = y, n 2 flops 18 of 33

19 Algorithm to compute the LU factorization Algorithm for computing the in place LU factorization of a matrix of size n n. #flops = 2n 3 /3 1: for k = 1:n-1 do 2: Let a ik be the element of maximum magnitude in A(k : n, k) 3: Permute row i and row k 4: A(k + 1 : n, k) = A(k + 1 : n, k)/a kk 5: for i = k + 1 : n do 6: for j = k + 1 : n do 7: a ij = a ij a ik a kj 8: end for 9: end for 10: end for 19 of 33

20 Algorithm to compute the LU factorization Left looking approach, pivoting ignored, A of size m n #flops = n 2 m n 3 /3 1: for k = 1:n do 2: for j = k:n do 3: u kj = a kj k 1 i=1 l kiu ij 4: end for 5: for i = k+1:m do 6: l ik = (a ik k 1 j=1 l iju jk )/u kk 7: end for 8: end for 20 of 33

21 Error analysis of the LU factorization Given the first k 1 columns of L and k 1 rows of U were computed, we have a kj = l k1 u 1j l k,k 1 u k 1,j + u kj, j = k : n a ik = l i1 u 1k l ik u kk, i = k + 1 : m The computed elements of ˆL and Û satisfy: a k 1 k kj ˆl ki û ij û kj γ k ˆl ki û ij, j k, i=1 i=1 k k a ik ˆl ij û jk γ k ˆl ij û jk, i > k. j=1 j=1 21 of 33

22 Error analysis of the LU factorization (continued) Theorem (Theorem 9.3 in [N.J.Higham, 2002]) Let A R m n, m n and let ˆL R m n and Û Rn n be its computed LU factors obtained by Gaussian elimination (suppose there was no failure during GE). Then, ˆLÛ = A + A, A γ n ˆL Û. Theorem (Theorem 9.4 in [N.J.Higham, 2002]) Let A R m n, m n and let ˆx be the computed solution to Ax = b obtained by using the computed LU factors of A obtained by Gaussian elimination. Then (A + A)ˆx = b, A γ 3n ˆL Û. 22 of 33

23 Error analysis of Ax = b Theorem (Theorem 9.4 in [N.J.Higham, 2002] continued) Proof. We have the following: Thus (A + A)ˆx = b, A γ 3n ˆL Û. ˆLÛ = A + A, A γn ˆL Û, (ˆL + L)ŷ = b, L γ n ˆL, (Û + U)ˆx = ŷ, U γn Û. b = (ˆL + L)(Û + U)ˆx = (A + A 1 + ˆL U + LÛ + L U)ˆx = (A + A)ˆx, where 2 A (3γ n + γn) ˆL Û γ3n ˆL Û. 23 of 33

24 Wilkinson s backward error stability result Growth factor g W defined as Note that g W = max i,j,k a k ij max i,j a ij u ij = aij i g W max a ij i,j Theorem (Wilkinson s backward error stability result, see also [N.J.Higham, 2002] for more details) Let A R n n and let ˆx be the computed solution of Ax = b obtained by using GEPP. Then (A + A)ˆx = b, A n 2 γ 3n g W (n) A. 24 of 33

25 The growth factor The LU factorization is backward stable if the growth factor is small (grows linearly with n). For partial pivoting, the growth factor g(n) 2 n 1, and this bound is attainable. In practice it is on the order of n 2/3 n 1/2 Exponential growth factor for Wilkinson matrix A = diag(±1) of 33

26 Experimental results for special matrices Several errror bounds for GEPP, the normwise backward error η and the componentwise backward error w (r = b Ax). η = w = max i r 1 A 1 x 1 + b 1, r i ( A x + b ) i. matrix cond(a,2) g W L 1 cond(u,1) PA LU F A F η w b hadamard 1.0E+0 4.1E+3 4.1E+3 5.3E+5 0.0E+0 3.3E E-15 randsvd 6.7E+7 4.7E+0 9.9E+2 1.4E E E E-15 chebvand 3.8E E+2 2.2E+3 4.8E E E E-16 frank 1.7E E+0 2.0E+0 1.9E E E E-23 hilb 8.0E E+0 3.1E+3 2.2E E E E of 33

27 Block formulation of the LU factorization Partitioning of matrix A of size n n [ ] A11 A A = 12 A 21 A 22 where A 11 is of size b b, A 21 is of size (m b) b, A 12 is of size b (n b) and A 22 is of size (m b) (n b). Block LU algebra The first iteration computes the factorization: [ ] P1 T L11 A = L 21 I n b [ Ib A 1 ] [ ] U11 U 12 I n b The algorithm continues recursively on the trailing matrix A of 33

28 Block LU factorization - the algorithm 1. Compute the LU factorization with partial pivoting of the first block column ( ) ( ) A11 L11 P 1 = U A L Pivot by applying the permutation matrix P T 1 3. Solve the triangular system 4. Update the trailing matrix, Ā = P T 1 A. L 11 U 12 = Ā12 A 1 = Ā 22 L 21 U Compute recursively the block LU factorization of A 1. on the entire matrix, 28 of 33

29 LU Factorization as in ScaLAPACK LU factorization on a P = Pr x Pc grid of processors For ib = 1 to n-1 step b A(ib) = A(ib : n, ib : n) 1. Compute panel factorization find pivot in each column, swap rows 2. Apply all row permutations broadcast pivot information along the rows swap rows at left and right 3. Compute block row of U broadcast right diagonal block of L of current panel 4. Update trailing matrix broadcast right block column of L broadcast down block row of U 29 of 33

30 Cost of LU Factorization in ScaLAPACK LU factorization on a P = Pr x Pc grid of processors For ib = 1 to n-1 step b A(ib) = A(ib : n, ib : n) 1. Compute panel factorization #messages = O(n log 2 P r ) 2. Apply all row permutations #messages = O(n/b(log 2 P r + log 2 P c)) 3. Compute block row of U #messages = O(n/b log 2 P c) 4. Update trailing matrix #messages = O(n/b(log 2 P r + log 2 P c) 30 of 33

31 Cost of parallel block LU Consider that we have a P P grid, block size b ( ) 2/3n 3 γ + n2 b + β n2 log P + P P P α ( 1.5n log P + 3.5n b log P ). 31 of 33

32 Acknowledgement Stability analysis results presented from [N.J.Higham, 2002] Some of the examples taken from [Golub and Van Loan, 1996] 32 of 33

33 References (1) Golub, G. H. and Van Loan, C. F. (1996). Matrix Computations (3rd Ed.). Johns Hopkins University Press, Baltimore, MD, USA. N.J.Higham (2002). Accuracy and Stability of Numerical Algorithms. SIAM, second edition. Schreiber, R. and Loan, C. V. (1989). A storage efficient WY representation for products of Householder transformations. SIAM J. Sci. Stat. Comput., 10(1): of 33