LAPACK is a freely available package which provides high-quality functions for solving linear algebra problems such as linear equations, least-square problems and eigenvalues. It relies on the BLAS (Basic Linear Algebra Subprograms) for low-level operations. For more information, please refer to the "LAPACK Users' Guide", 3rd ed., Anderson, E. et al., Society for Industrial and Applied Mathematics, Philadelphia (USA), 1999, ISBN 0-89871-447-8. You can download the source code of LAPACK from http://www.netlib.org.
LAPACK functions are not integrated in LME, but rather provided as replacements and additions to the LME core functions. While it does not change the way you use the functions, this approach offers more flexibility for future improvements and permits to keep LME light for applications where memory is limited. Currently, depending on the platform, the LME functions based on LAPACK weight between 700 and 800 kilobytes, more than the core of LME itself; if new functions are implemented, or if better versions become available, it will be possible to replace the LAPACK extension without changing LME itself or the whole application.
Currently, only a subset of LAPACK is available for LME. The functions have been chosen from the set of double-precision subroutines and usually apply to general real or complex matrices.
x * y M1 * M2 M * x
x*y multiplies the operands together. Operands can be scalars (plain arithmetic product), matrices (matrix product), or mixed scalar and matrix.
2 * 3 6 [1,2;3,4] * [3;5] 13 29 [3 4] * 2 6 8
Matrix left division.
a \ b a \ B A \ B
a\b, where a is a scalar, divides the second operand by the first operand. If the second operand is a matrix, each element is divided by a.
In the case where the left operand A is a matrix, A\B solves the set of linear equations A*x=B. If B is an m-by-n matrix, the n columns are handled separately and the solution x has also n columns. The solution x is inv(A)*B if A is square; otherwise, it is a least-square solution.
[1,2;3,4] \ [2;6] 2 0 [1;2] \ [1.1;2.1] 1.06 [1,2] \ 1 0.2 0.4
dgesv, zgesv (square matrices); dgelss, zgelss (non-square matrices)
Matrix right division.
a / b A / b A / B
a/b, where b is a scalar, divides the first operand by the second operand. If the first operand is a matrix, each element is divided by b.
In the case where the right operand B is a matrix, A/B solves the set of linear equations A=x*B. If A is an m-by-n matrix, the m rows are handled separately and the solution x has also m rows. The solution x is A*inv(B) if B is square; otherwise, it is a least-square solution.
[2,6] / [1,3;2,4] 2 0 [1.1,2.1] / [1,2] 1.06 1 / [1;2] 0.2 0.4
dgesv, zgesv (square matrices); dgelss, zgelss (non-square matrices)
Diagonal similarity transform for balancing a matrix.
B = balance(X) (T, B) = balance(X)
balance(X) applies a diagonal similarity transform to the square matrix X to make the rows and columns as close in norm as possible. Balancing may reduce the 1-norm of the matrix, and improves the accuracy of the computed eigenvalues and/or eigenvectors.
balance returns the balanced matrix B which has the same eigenvalues and singular values as X, and optionally the scaling matrix T such that T\X*T=B.
A = [1,2e6;3e-6,4]; (T,B) = balance(A) T = 1e6 0 0 1 B = 1 2 3 4 eig(A)-eig(B) 0 0
C = chol(X)
If a square matrix X is symmetric (or hermitian) and positive definite, it can be decomposed into the following product:
where C is an upper triangular matrix.
The part of X below the main diagonal is not used, because X is assumed to be symmetric or hermitian. An error occurs if X is not positive definite.
C = chol([5,3;3,8]) C = 2.2361 1.3416 0 2.4900 C'*C 5 3 3 8
d = det(X)
det(X) is the determinant of the square matrix M, which is 0 (up to the rounding errors) if M is singular. The function rank is a numerically more robust test for singularity.
det([1,2;3,4]) -2 det([1,2;1,2]) 0
Eigenvalues and eigenvectors.
e = eig(A) (V,D) = eig(A) e = eig(A,B) (V,D) = eig(A,B)
eig(A) returns the vector of eigenvalues of the square matrix A.
(V,D) = eig(A) returns a diagonal matrix D of eigenvalues and a matrix V whose columns are the corresponding eigenvectors. They are such that A*V = V*D.
eig(A,B) returns the generalized eigenvalues and eigenvectors of square matrices A and B, such that A*V = B*V*D.
A = [1,2;3,4]; B = [2,1;3,3]; eig(A) 5.3723 -0.3723 (V,D) = eig(A,B) V = -0.5486 -1 -1 0.8229 D = 1.2153 0 0 -0.5486 A*V,B*V*D ans = -2.5486 0.6458 -5.6458 0.2915 ans = -2.5486 0.6458 -5.6458 0.2915
dgeev, zgeev for eigenvalues and eigenvectors; dgegv, zgegv for generalized eigenvalues and eigenvectors
(P,H) = hess(A) H = hess(A)
hess(A) reduces the square matrix A A to the upper Hessenberg form H using an orthogonal similarity transformation P'*H*P=A. The result H is zero below the first subdiagonal and has the same eigenvalues as A.
(P,H)=hess([1,2,3;4,5,6;7,8,9]) P = 1 0 0 0 -0.4961 -0.8682 0 -0.8682 0.4961 H = 1 -3.597 -0.2481 -8.0623 14.0462 2.8308 0 0.8308 -4.6154e-2
dgehrd, zgehrd; dorghr, zunghr for computing P
R = inv(X)
inv(X) returns the inverse of the square matrix X. X must not be singular; otherwise, its elements are infinite.
To solve a set of linear of equations, the operator \ is more efficient.
inv([1,2;3,4]) -2 1 1.5 -0.5
L = logm(X)
logm(A) returns the square matrix logarithm of A, the inverse of the matrix exponential. The matrix logarithm does not always exist.
L = logm([1,2;3,4]) L = -0.3504+2.3911j 0.9294-1.0938j 1.394-1.6406j 1.0436+0.7505j expm(L) 1-4.4409e-16j 2-6.1062e-16j 3-9.4369e-16j 4
(L,U,P) = lu(X) (L2,U) = lu(X) M = lu(X)
(L,U,P)=lu(X) factorizes the square matrix X such that P*X=L*U, where L is a lower-triangular square matrix, U is an upper-triangular square matrix, and P is a permutation square matrix (a real matrix where each line and each column has a single 1 element and zeros elsewhere).
(L2,U)=lu(X) factorizes the square matrix X such that X=L2*U, where L2=P\L.
M=lu(X) yields a square matrix M whose upper triangular part is U and whose lower triangular part (below the main diagonal) is L without the diagonal.
X = [1,2,3;4,5,6;7,8,8]; (L,U,P) = lu(X) L = 1 0 0 0.143 1 0 0.571 0.5 1 U = 7 8 8 0 0.857 1.857 0 0 0.5 P = 0 0 1 1 0 0 0 1 0 P*X-L*U ans = 0 0 0 0 0 0 0 0 0
Z = null(A)
null(A) returns a matrix Z whose columns are an orthonormal basis for the null space of m-by-n matrix A. Z has n-rank(A) columns, which are the last right singular values of A (that is, those corresponding to the negligible singular values).
Without input argument, null gives the null value (the unique value of the special null type, not related to linear algebra).
null([1,2,3;1,2,4;1,2,5]) -0.8944 0.4472 8.0581e-17
Q = orth(A)
orth(A) returns a matrix Q whose columns are an orthonormal basis for the range of those of matrix A. Q has rank(A) columns, which are the first left singular vectors of A (that is, those corresponding to the largest singular values).
orth([1,2,3;1,2,4;1,2,5]) -0.4609 0.788 -0.5704 8.9369e-2 -0.6798 -0.6092
R = pinv(A) R = pinv(A,tol)
pinv(A) returns the pseudo-inverse of matrix A, i.e. a matrix B such that B*A*B=B and A*B*A=A. For a full-rank square matrix A, pinv(A) is the same as inv(A).
pinv is based on the singular value decomposition; all values below the tolerance times the largest singular value are neglected. The default tolerance is the maximum dimension times eps; another value may be supplied to pinv as a second parameter.
pinv([1,2;3,4]) -2 1 1.5 -0.5 B = pinv([1;2]) B = 0.2 0.4 [1;2] * B * [1;2] 1 2 B * [1;2] * B 0.2 0.4
(Q, R, E) = qr(A) (Q, R) = qr(A) (Qe, Re, e) = qr(A, false) (Qe, Re) = qr(A, false)
With three output arguments, qr(A) computes the QR decomposition of matrix A with column pivoting, i.e. a square unitary matrix Q and an upper triangular matrix R such that A*E=Q*R. With two output arguments, qr(A) computes the QR decomposition without pivoting, such that A=Q*R.
With a second input argument with the value false, if A has m rows and n columns with m>n, qr produces an m-by-n Q and an n-by-n R. Bottom rows of zeros of R, and the corresponding columns of Q, are discarded. With column pivoting, the third output argument e is a permutation vector: A(:,e)=Q*R.
(Q,R) = qr([1,2;3,4;5,6]) Q = -0.169 0.8971 0.4082 -0.5071 0.276 -0.8165 -0.8452 -0.345 0.4082 R = -5.9161 -7.4374 0 0.8281 0 0 (Qe,Re) = qr([1,2;3,4;5,6],false) Qe = -0.169 0.8971 -0.5071 0.276 -0.8452 -0.345 Re = -5.9161 -7.4374 0 0.8281
dgeqrf, zgeqrf for decomposition without pivoting; dgeqpf, zgeqpf for decomposition with pivoting; dorgqr, zung2r for computing Q
Generalized Schur factorization.
(S, T, Q, Z) = qz(A, B)
qz(A,B) computes the generalized Schur factorization for square matrices A and B, such that Q*S*Z=A and Q*B*Z=B. For real matrices, the result is real with S block upper-diagonal with 1x1 and 2x2 blocks, and T upper-diagonal. For complex matrices, the result is complex with both S and T upper-diagonal.
(S, T, Q, Z) = qz([1,2;3,4], [5,6;7,8]) S = 5.3043 0.5927 -1.2062 0.2423 T = 13.19 0 0 0.1516 Q = -0.5921 -0.8059 -0.8059 0.5921 Z = -0.6521 -0.7581 0.7581 -0.6521 Q*S*Z 1 2 3 4 Q*T*Z 5 6 7 8
Rank of a matrix.
n = rank(X) n = rank(X, tol)
rank(X) returns the rank of matrix X, i.e. the number of lines or columns linearly independent. To obtain it, the singular values are computed and the number of values significantly larger than 0 is counted. The value below which they are considered to be 0 can be specified with the optional second argument.
Reciprocal condition number.
r = rcond(A)
rcond(A) computes the reciprocal condition number of matrix A, i.e. a number r=1/(norm(A,1)*norm(inv(A),1)). The reciprocal condition number is near 1 for well-conditioned matrices and near 0 for bad-conditioned ones.
rcond([1,1;0,1]) 0.3 rcond([1,1e6;2,1e6]) 5e-7
dlange, zlange for the 1-norm of X; dgecon, zgecon for the rcond of X using its 1-norm
(U,T) = schur(A) T = schur(A)
schur(A) computes the Schur factorization of square matrix A, i.e. a unitary matrix U and a square matrix T (the Schur matrix) such that A=U*T*U'. If A is complex, the Schur matrix is upper triangular, and its diagonal contains the eigenvalues of A; if A is real, the Schur matrix is real upper triangular, except that there may be 2-by-2 blocks on the main diagonal which correspond to the complex eigenvalues of A.
(U,T) = schur([1,2;3,4]) U = 0.416 -0.9094 0.9094 0.416 T = 5.3723 -1 0 -0.3723 eig([1,2;3,4]) 5.3723 -0.3723
Matrix square root.
S = sqrtm(X)
sqrtm(A) returns the square root of the square matrix A, i.e. a matrix S such that S*S=A.
S = sqrtm([1,2;3,4]) S = 0.5537+0.4644j 0.807-0.2124j 1.2104-0.3186j 1.7641+0.1458j S*S 1 2 3 4
Singular value decomposition.
s = svd(X) (U,S,V) = svd(X) (U,S,V) = svd(X,false)
The singular value decomposition (U,S,V) = svd(X) decomposes the m-by-n matrix X such that X = U*S*V', where S is an m-by-n diagonal matrix with decreasing positive diagonal elements (the singular values of X), U is an m-by-m unitary matrix, and V is an n-by-n unitary matrix. The number of non-zero diagonal elements of S (up to rounding errors) gives the rank of X.
When m>n, (U,S,V) = svd(X,false) produces an n-by-n diagonal matrix S and an m-by-n matrix U. The relationship X = U*S*V' still holds.
With one output argument, s = svd(X) returns the vector of singular values.
(U,S,V)=svd([1,2,3;4,5,6]) U = -0.3863 -0.9224 -0.9224 0.3863 S = 9.508 0 0 0 0.7729 0 V = -0.4287 0.806 0.4082 -0.5663 0.1124 -0.8165 -0.7039 -0.5812 0.4082 U*S*V' 1 2 3 4 5 6 (U,S,V)=svd([1,2,3;4,5,6],false) U = -0.3863 -0.9224 -0.9224 0.3863 S = 9.508 0 0 0.7729 V = -0.4287 0.806 -0.5663 0.1124 -0.7039 -0.5812 U*S*V 1.4944 -2.4586 2.4944 3.7929 -7.2784 4.7929