Linear algebra functions in Julia are largely implemented by calling functions from LAPACK. Sparse factorizations call functions from SuiteSparse.
Matrix multiplication
Matrix division using a polyalgorithm. For input matrices A and B, the result X is such that A*X == B when A is square. The solver that is used depends upon the structure of A. A direct solver is used for upper- or lower triangular A. For Hermitian A (equivalent to symmetric A for non-complex A) the BunchKaufman factorization is used. Otherwise an LU factorization is used. For rectangular A the result is the minimum-norm least squares solution computed by reducing A to bidiagonal form and solving the bidiagonal least squares problem. For sparse, square A the LU factorization (from UMFPACK) is used.
Compute the dot product
Compute the cross product of two 3-vectors
Compute the norm of a Vector or a Matrix
Compute the reduced row echelon form of the matrix A.
Compute a convenient factorization (including LU, Cholesky, Bunch Kaufman, Triangular) of A, based upon the type of the input matrix. The return value can then be reused for efficient solving of multiple systems. For example: A=factorize(A); x=A\\b; y=A\\C.
factorize! is the same as factorize(), but saves space by overwriting the input A, instead of creating a copy.
Compute the LU factorization of A, such that A[p,:] = L*U.
Compute the LU factorization of A, returning an LU object for dense A or an UmfpackLU object for sparse A. The individual components of the factorization F can be accesed by indexing: F[:L], F[:U], and F[:P] (permutation matrix) or F[:p] (permutation vector). An UmfpackLU object has additional components F[:q] (the left permutation vector) and Rs the vector of scaling factors. The following functions are available for both LU and UmfpackLU objects: size, \ and det. For LU there is also an inv method. The sparse LU factorization is such that L*U is equal to``scale(Rs,A)[p,q]``.
lufact! is the same as lufact(), but saves space by overwriting the input A, instead of creating a copy. For sparse A the nzval field is not overwritten but the index fields, colptr and rowval are decremented in place, converting from 1-based indices to 0-based indices.
Compute Cholesky factorization of a symmetric positive-definite matrix A and return the matrix F. If LU is L (Lower), A = L*L'. If LU is U (Upper), A = R'*R.
Compute the Cholesky factorization of a dense symmetric positive-definite matrix A and return a Cholesky object. LU may be ‘L’ for using the lower part or ‘U’ for the upper part. The default is to use ‘U’. The triangular matrix can be obtained from the factorization F with: F[:L] and F[:U]. The following functions are available for Cholesky objects: size, \, inv, det. A LAPACK.PosDefException error is thrown in case the matrix is not positive definite.
Compute the sparse Cholesky factorization of a sparse matrix A. If A is Hermitian its Cholesky factor is determined. If A is not Hermitian the Cholesky factor of A*A' is determined. A fill-reducing permutation is used. Methods for size, solve, \, findn_nzs, diag, det and logdet. One of the solve methods includes an integer argument that can be used to solve systems involving parts of the factorization only. The optional boolean argument, ll determines whether the factorization returned is of the A[p,p] = L*L' form, where L is lower triangular or A[p,p] = scale(L,D)*L' form where L is unit lower triangular and D is a non-negative vector. The default is LDL.
cholfact! is the same as cholfact(), but saves space by overwriting the input A, instead of creating a copy.
Compute the pivoted Cholesky factorization of a symmetric positive semi-definite matrix A and return a CholeskyPivoted object. LU may be ‘L’ for using the lower part or ‘U’ for the upper part. The default is to use ‘U’. The triangular factors contained in the factorization F can be obtained with F[:L] and F[:U], whereas the permutation can be obtained with F[:P] or F[:p]. The following functions are available for CholeskyPivoted objects: size, \, inv, det. A LAPACK.RankDeficientException error is thrown in case the matrix is rank deficient.
cholpfact! is the same as cholpfact, but saves space by overwriting the input A, instead of creating a copy.
Compute the QR factorization of A such that A = Q*R. Also see qrfact. The default is to compute a thin factorization. Note that R is not extended with zeros when the full Q is requested.
Computes the QR factorization of A and returns a QR type, which is a Factorization F consisting of an orthogonal matrix F[:Q] and a triangular matrix F[:R]. The following functions are available for QR objects: size, \. The orthogonal matrix Q=F[:Q] is a QRPackedQ type which has the * operator overloaded to support efficient multiplication by Q and Q'. Multiplication with respect to either thin or full Q is allowed, i.e. both F[:Q]*F[:R] and F[:Q]*A are supported. A QRPackedQ matrix can be converted into a regular matrix with full.
qrfact! is the same as qrfact(), but saves space by overwriting the input A, instead of creating a copy.
Computes the QR factorization of A with pivoting, such that A[:,p] = Q*R, Also see qrpfact. The default is to compute a thin factorization.
Computes the QR factorization of A with pivoting and returns a QRPivoted object, which is a Factorization F consisting of an orthogonal matrix F[:Q], a triangular matrix F[:R], and a permutation F[:p] (or its matrix representation F[:P]). The following functions are available for QRPivoted objects: size, \. The orthogonal matrix Q=F[:Q] is a QRPivotedQ type which has the * operator overloaded to support efficient multiplication by Q and Q'. Multiplication with respect to either the thin or full Q is allowed, i.e. both F[:Q]*F[:R] and F[:Q]*A are supperted. A QRPivotedQ matrix can be converted into a regular matrix with full.
qrpfact! is the same as qrpfact(), but saves space by overwriting the input A, instead of creating a copy.
Compute the Bunch Kaufman factorization of a real symmetric or complex Hermitian matrix A and return a BunchKaufman object. The following functions are available for BunchKaufman objects: size, \, inv, issym, ishermitian.
bkfact! is the same as bkfact(), but saves space by overwriting the input A, instead of creating a copy.
Compute the matrix square root of A. If B = sqrtm(A), then B*B == A within roundoff error.
Compute eigenvalues and eigenvectors of A
Compute generalized eigenvalues and vectors of A and B
Returns the eigenvalues of A.
Returns the largest eigenvalue of A.
Returns the smallest eigenvalue of A.
Returns the eigenvectors of A.
For SymTridiagonal matrices, if the optional vector of eigenvalues eigvals is specified, returns the specific corresponding eigenvectors.
Compute the eigenvalue decomposition of A and return an Eigen object. If F is the factorization object, the eigenvalues can be accessed with F[:values] and the eigenvectors with F[:vectors]. The following functions are available for Eigen objects: inv, det.
Compute the generalized eigenvalue decomposition of A and B and return an GeneralizedEigen object. If F is the factorization object, the eigenvalues can be accessed with F[:values] and the eigenvectors with F[:vectors].
eigfact! is the same as eigfact(), but saves space by overwriting the input A (and B), instead of creating a copy.
Compute the Hessenberg decomposition of A and return a Hessenberg object. If F is the factorization object, the unitary matrix can be accessed with F[:Q] and the Hessenberg matrix with F[:H]. When Q is extracted, the resulting type is the HessenbergQ object, and may be converted to a regular matrix with full.
hessfact! is the same as hessfact(), but saves space by overwriting the input A, instead of creating a copy.
Computes the Schur factorization of the matrix A. The (quasi) triangular Schur factor can be obtained from the Schur object F with either F[:Schur] or F[:T] and the unitary/orthogonal Schur vectors can be obtained with F[:vectors] or F[:Z] such that A=F[:vectors]*F[:Schur]*F[:vectors]'. The eigenvalues of A can be obtained with F[:values].
Computer the Schur factorization of A, overwriting A in the process. See schurfact()
See schurfact
Computes the Generalized Schur (or QZ) factorization of the matrices A and B. The (quasi) triangular Schur factors can be obtained from the Schur object F with F[:S] and F[:T], the left unitary/orthogonal Schur vectors can be obtained with F[:left] or F[:Q] and the right unitary/orthogonal Schur vectors can be obtained with F[:right] or F[:Z] such that A=F[:left]*F[:S]*F[:right]' and B=F[:left]*F[:T]*F[:right]'. The generalized eigenvalues of A and B can be obtained with F[:alpha]./F[:beta].
See schurfact
Compute the Singular Value Decomposition (SVD) of A and return an SVD object. U, S, V and Vt can be obtained from the factorization F with F[:U], F[:S], F[:V] and F[:Vt], such that A = U*diagm(S)*Vt. If thin is true, an economy mode decomposition is returned. The algorithm produces Vt and hence Vt is more efficient to extract than V. The default is to produce a thin decomposition.
svdfact! is the same as svdfact(), but saves space by overwriting the input A, instead of creating a copy. If thin is true, an economy mode decomposition is returned. The default is to produce a thin decomposition.
Compute the SVD of A, returning U, vector S, and V such that A == U*diagm(S)*V'. If thin is true, an economy mode decomposition is returned.
Returns the singular values of A.
Returns the singular values of A, while saving space by overwriting the input.
Compute the generalized SVD of A and B, returning a GeneralizedSVD Factorization object, such that A = U*D1*R0*Q' and B = V*D2*R0*Q'.
Compute the generalized SVD of A and B, returning U, V, Q, D1, D2, and R0 such that A = U*D1*R0*Q' and B = V*D2*R0*Q'.
Return only the singular values from the generalized singular value decomposition of A and B.
Upper triangle of a matrix.
Upper triangle of a matrix, overwriting M in the process.
Lower triangle of a matrix.
Lower triangle of a matrix, overwriting M in the process.
A Range giving the indices of the k-th diagonal of the matrix M.
The k-th diagonal of a matrix, as a vector.
Construct a diagonal matrix and place v on the k-th diagonal.
scale(A::Array, B::Number) scales all values in A with B. Note: In cases where the array is big enough, scale can be much faster than A .* B, due to the use of BLAS.
scale(A::Matrix, B::Vector) is the same as multiplying with a diagonal matrix on the right, and scales the columns of A with the values in B.
scale(A::Vector, B::Matrix) is the same as multiplying with a diagonal matrix on the left, and scales the rows of B with the values in A.
scale!(A,B) overwrites the input array with the scaled result.
symmetrize!(A) converts from the BLAS/LAPACK symmetric storage format, in which only the UL (‘U’pper or ‘L’ower, default ‘U’) triangle is used, to a full symmetric matrix.
Construct a tridiagonal matrix from the lower diagonal, diagonal, and upper diagonal, respectively. The result is of type Tridiagonal and provides efficient specialized linear solvers, but may be converted into a regular matrix with full.
Constructs an upper (isupper=true) or lower (isupper=false) bidiagonal matrix using the given diagonal (dv) and off-diagonal (ev) vectors. The result is of type Bidiagonal and provides efficient specialized linear solvers, but may be converted into a regular matrix with full.
Construct a real-symmetric tridiagonal matrix from the diagonal and upper diagonal, respectively. The result is of type SymTridiagonal and provides efficient specialized eigensolvers, but may be converted into a regular matrix with full.
Construct a matrix in a form suitable for applying the Woodbury matrix identity
Compute the rank of a matrix
Compute the p-norm of a vector or a matrix. p is 2 by default, if not provided. If A is a vector, norm(A, p) computes the p-norm. norm(A, Inf) returns the largest value in abs(A), whereas norm(A, -Inf) returns the smallest. If A is a matrix, valid values for p are 1, 2, or Inf. In order to compute the Frobenius norm, use normfro.
Compute the Frobenius norm of a matrix A.
Matrix condition number, computed using the p-norm. p is 2 by default, if not provided. Valid values for p are 1, 2, or Inf.
Matrix trace
Matrix determinant
Log of Matrix determinant. Equivalent to log(det(M)), but may provide increased accuracy and/or speed.
Matrix inverse
Moore-Penrose inverse
Basis for null space of M.
Construct a matrix by repeating the given matrix n times in dimension 1 and m times in dimension 2.
Construct an array by repeating the entries of A. The i-th element of inner specifies the number of times that the individual entries of the i-th dimension of A should be repeated. The i-th element of outer specifies the number of times that a slice along the i-th dimension of A should be repeated.
Kronecker tensor product of two vectors or two matrices.
Determine parameters [a, b] that minimize the squared error between y and a+b*x.
Weighted least-squares linear regression.
Matrix exponential.
Test whether a matrix is symmetric.
Test whether a matrix is positive-definite.
Test whether a matrix is positive-definite, overwriting A in the processes.
Test whether a matrix is lower-triangular.
Test whether a matrix is upper-triangular.
Test whether a matrix is hermitian.
The transpose operator (.').
The conjugate transpose operator (').
eigs returns the nev requested eigenvalues in d, the corresponding Ritz vectors v (only if ritzvec=true), the number of converged eigenvalues nconv, the number of iterations niter and the number of matrix vector multiplications nmult, as well as the final residual vector resid.
peakflops computes the peak flop rate of the computer by using BLAS dgemm. By default, if no arguments are specified, it multiplies a matrix of size n x n, where n = 2000. If the underlying BLAS is using multiple threads, higher flop rates are realized. The number of BLAS threads can be set with blas_set_num_threads(n).
If the keyword argument parallel is set to true, peakflops is run in parallel on all the worker processors. The flop rate of the entire parallel computer is returned. When running in parallel, only 1 BLAS thread is used. The argument n still refers to the size of the problem that is solved on each processor.
This module provides wrappers for some of the BLAS functions for linear algebra. Those BLAS functions that overwrite one of the input arrays have names ending in '!'.
Usually a function has 4 methods defined, one each for Float64, Float32, Complex128 and Complex64 arrays.
Dot product of two vectors consisting of n elements of array X with stride incx and n elements of array Y with stride incy. There are no dot methods for Complex arrays.
The following functions are defined within the Base.LinAlg.BLAS module.
Copy n elements of array X with stride incx to array Y with stride incy. Returns Y.
2-norm of a vector consisting of n elements of array X with stride incx.
sum of the absolute values of the first n elements of array X with stride incx.
Overwrite Y with a*X + Y. Returns Y.
Overwrite X with a*X. Returns X.
Returns a*X.
Rank-k update of the symmetric matrix C as alpha*A*A.' + beta*C or alpha*A.'*A + beta*C according to whether trans is ‘N’ or ‘T’. When uplo is ‘U’ the upper triangle of C is updated (‘L’ for lower triangle). Returns C.
Returns either the upper triangle or the lower triangle, according to uplo (‘U’ or ‘L’), of alpha*A*A.' or alpha*A.'*A, according to trans (‘N’ or ‘T’).
Methods for complex arrays only. Rank-k update of the Hermitian matrix C as alpha*A*A' + beta*C or alpha*A'*A + beta*C according to whether trans is ‘N’ or ‘T’. When uplo is ‘U’ the upper triangle of C is updated (‘L’ for lower triangle). Returns C.
Methods for complex arrays only. Returns either the upper triangle or the lower triangle, according to uplo (‘U’ or ‘L’), of alpha*A*A' or alpha*A'*A, according to trans (‘N’ or ‘T’).
Update vector y as alpha*A*x + beta*y or alpha*A'*x + beta*y according to trans (‘N’ or ‘T’). The matrix A is a general band matrix of dimension m by size(A,2) with kl sub-diagonals and ku super-diagonals. Returns the updated y.
Returns alpha*A*x or alpha*A'*x according to trans (‘N’ or ‘T’). The matrix A is a general band matrix of dimension m by size(A,2) with kl sub-diagonals and ku super-diagonals.
Update vector y as alpha*A*x + beta*y where A is a a symmetric band matrix of order size(A,2) with k super-diagonals stored in the argument A. The storage layout for A is described the reference BLAS module, level-2 BLAS at http://www.netlib.org/lapack/explore-html/.
Returns the updated y.
Returns alpha*A*x where A is a symmetric band matrix of order size(A,2) with k super-diagonals stored in the argument A.
Returns A*x where A is a symmetric band matrix of order size(A,2) with k super-diagonals stored in the argument A.
Update C as alpha*A*B + beta*C or the other three variants according to tA (transpose A) and tB. Returns the updated C.
Returns alpha*A*B or the other three variants according to tA (transpose A) and tB.
Returns alpha*A*B or the other three variants according to tA (transpose A) and tB.
Update the vector y as alpha*A*x + beta*x or alpha*A'x + beta*x according to tA (transpose A). Returns the updated y.
Returns alpha*A*x or alpha*A'x according to tA (transpose A).
Returns A*x or A'x according to tA (transpose A).
Update C as alpha*A*B + beta*C or alpha*B*A + beta*C according to side. A is assumed to be symmetric. Only the ul triangle of A is used. Returns the updated C.
Returns alpha*A*B or alpha*B*A according to side. A is assumed to be symmetric. Only the ul triangle of A is used.
Returns A*B or B*A according to side. A is assumed to be symmetric. Only the ul triangle of A is used.
Returns alpha*A*B or the other three variants according to tA (transpose A) and tB.
Update the vector y as alpha*A*y + beta*y. A is assumed to be symmetric. Only the ul triangle of A is used. Returns the updated y.
Returns alpha*A*x. A is assumed to be symmetric. Only the ul triangle of A is used.
Returns A*x. A is assumed to be symmetric. Only the ul triangle of A is used.
Update B as alpha*A*B or one of the other three variants determined by side (A on left or right) and tA (transpose A). Only the ul triangle of A is used. dA indicates if A is unit-triangular (the diagonal is assumed to be all ones). Returns the updated B.
Returns alpha*A*B or one of the other three variants determined by side (A on left or right) and tA (transpose A). Only the ul triangle of A is used. dA indicates if A is unit-triangular (the diagonal is assumed to be all ones).
Overwrite B with the solution to A*X = alpha*B or one of the other three variants determined by side (A on left or right of X) and tA (transpose A). Only the ul triangle of A is used. dA indicates if A is unit-triangular (the diagonal is assumed to be all ones). Returns the updated B.
Returns the solution to A*X = alpha*B or one of the other three variants determined by side (A on left or right of X) and tA (transpose A). Only the ul triangle of A is used. dA indicates if A is unit-triangular (the diagonal is assumed to be all ones).
Update b as alpha*A*b or one of the other three variants determined by side (A on left or right) and tA (transpose A). Only the ul triangle of A is used. dA indicates if A is unit-triangular (the diagonal is assumed to be all ones). Returns the updated b.
Returns alpha*A*b or one of the other three variants determined by side (A on left or right) and tA (transpose A). Only the ul triangle of A is used. dA indicates if A is unit-triangular (the diagonal is assumed to be all ones).
Overwrite b with the solution to A*X = alpha*b or one of the other three variants determined by side (A on left or right of X) and tA (transpose A). Only the ul triangle of A is used. dA indicates if A is unit-triangular (the diagonal is assumed to be all ones). Returns the updated b.
Returns the solution to A*X = alpha*b or one of the other three variants determined by side (A on left or right of X) and tA (transpose A). Only the ul triangle of A is used. dA indicates if A is unit-triangular (the diagonal is assumed to be all ones).
Set the number of threads the BLAS library should use.