Linear Algebra

Standard Functions

Linear algebra functions in Julia are largely implemented by calling functions from LAPACK. Sparse factorizations call functions from SuiteSparse.

*(A, B)

Matrix multiplication

\(A, B)

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 a pivoted QR factorization of A and a rank estimate of A based on the R factor.

When A is sparse, a similar polyalgorithm is used. For indefinite matrices, the LDLt factorization does not use pivoting during the numerical factorization and therefore the procedure can fail even for invertible matrices.

dot(x, y)
(x, y)

Compute the dot product. For complex vectors, the first vector is conjugated.

vecdot(x, y)

For any iterable containers x and y (including arrays of any dimension) of numbers (or any element type for which dot is defined), compute the Euclidean dot product (the sum of dot(x[i],y[i])) as if they were vectors.

cross(x, y)
×(x, y)

Compute the cross product of two 3-vectors.

factorize(A)

Compute a convenient factorization (including LU, Cholesky, Bunch-Kaufman, LowerTriangular, UpperTriangular) 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.

full(F)

Reconstruct the matrix A from the factorization F=factorize(A).

lu(A) → L, U, p

Compute the LU factorization of A, such that A[p,:] = L*U.

lufact(A[, pivot=Val{true}]) → F

Compute the LU factorization of A. The return type of F depends on the type of A. In most cases, if A is a subtype S of AbstractMatrix with an element type T supporting +, -, * and / the return type is LU{T,S{T}}. If pivoting is chosen (default) the element type should also support abs and <. When A is sparse and have element of type Float32, Float64, Complex{Float32}, or Complex{Float64} the return type is UmfpackLU. Some examples are shown in the table below.

Type of input A Type of output F Relationship between F and A
Matrix() LU F[:L]*F[:U] == A[F[:p], :]
Tridiagonal() LU{T,Tridiagonal{T}} F[:L]*F[:U] == A[F[:p], :]
SparseMatrixCSC() UmfpackLU F[:L]*F[:U] == (F[:Rs] .* A)[F[:p], F[:q]]

The individual components of the factorization F can be accessed by indexing:

Component Description LU LU{T,Tridiagonal{T}} UmfpackLU
F[:L] L (lower triangular) part of LU
F[:U] U (upper triangular) part of LU
F[:p] (right) permutation Vector
F[:P] (right) permutation Matrix  
F[:q] left permutation Vector    
F[:Rs] Vector of scaling factors    
F[:(:)] (L,U,p,q,Rs) components    
Supported function LU LU{T,Tridiagonal{T}} UmfpackLU
/    
\
cond  
det
logdet  
logabsdet  
size  
lufact!(A) → LU

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.

chol(A[, LU]) → F

Compute the Cholesky factorization of a symmetric positive definite matrix A and return the matrix F. If LU is Val{:U} (Upper), F is of type UpperTriangular and A = F'*F. If LU is Val{:L} (Lower), F is of type LowerTriangular and A = F*F'. LU defaults to Val{:U}.

cholfact(A, [LU=:U[,pivot=Val{false}]][;tol=-1.0]) → Cholesky

Compute the Cholesky factorization of a dense symmetric positive (semi)definite matrix A and return either a Cholesky if pivot==Val{false} or CholeskyPivoted if pivot==Val{true}. 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. For CholeskyPivoted there is also defined a rank. If pivot==Val{false} a PosDefException exception is thrown in case the matrix is not positive definite. The argument tol determines the tolerance for determining the rank. For negative values, the tolerance is the machine precision.

cholfact(A; shift=0, perm=Int[]) → CHOLMOD.Factor

Compute the Cholesky factorization of a sparse positive definite matrix A. A fill-reducing permutation is used. F = cholfact(A) is most frequently used to solve systems of equations with F\b, but also the methods diag, det, logdet are defined for F. You can also extract individual factors from F, using F[:L]. However, since pivoting is on by default, the factorization is internally represented as A == P'*L*L'*P with a permutation matrix P; using just L without accounting for P will give incorrect answers. To include the effects of permutation, it’s typically preferable to extact “combined” factors like PtL = F[:PtL] (the equivalent of P'*L) and LtP = F[:UP] (the equivalent of L'*P).

Setting optional shift keyword argument computes the factorization of A+shift*I instead of A. If the perm argument is nonempty, it should be a permutation of 1:size(A,1) giving the ordering to use (instead of CHOLMOD’s default AMD ordering).

The function calls the C library CHOLMOD and many other functions from the library are wrapped but not exported.

cholfact!(A [,LU=:U [,pivot=Val{false}]][;tol=-1.0]) → Cholesky

cholfact! is the same as cholfact(), but saves space by overwriting the input A, instead of creating a copy. cholfact! can also reuse the symbolic factorization from a different matrix F with the same structure when used as: cholfact!(F::CholmodFactor, A).

ldltfact(A) → LDLtFactorization

Compute a factorization of a positive definite matrix A such that A=L*Diagonal(d)*L' where L is a unit lower triangular matrix and d is a vector with non-negative elements.

ldltfact(A; shift=0, perm=Int[]) → CHOLMOD.Factor

Compute the LDLt factorization of a sparse symmetric or Hermitian matrix A. A fill-reducing permutation is used. F = ldltfact(A) is most frequently used to solve systems of equations with F\b, but also the methods diag, det, logdet are defined for F. You can also extract individual factors from F, using F[:L]. However, since pivoting is on by default, the factorization is internally represented as A == P'*L*D*L'*P with a permutation matrix P; using just L without accounting for P will give incorrect answers. To include the effects of permutation, it’s typically preferable to extact “combined” factors like PtL = F[:PtL] (the equivalent of P'*L) and LtP = F[:UP] (the equivalent of L'*P). The complete list of supported factors is :L, :PtL, :D, :UP, :U, :LD, :DU, :PtLD, :DUP.

Setting optional shift keyword argument computes the factorization of A+shift*I instead of A. If the perm argument is nonempty, it should be a permutation of 1:size(A,1) giving the ordering to use (instead of CHOLMOD’s default AMD ordering).

The function calls the C library CHOLMOD and many other functions from the library are wrapped but not exported.

qr(A[, pivot=Val{false}][;thin=true]) → Q, R, [p]

Compute the (pivoted) QR factorization of A such that either A = Q*R or A[:,p] = 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.

qrfact(A[, pivot=Val{false}]) → F

Computes the QR factorization of A. The return type of F depends on the element type of A and whether pivoting is specified (with pivot==Val{true}).

Return type eltype(A) pivot Relationship between F and A
QR not BlasFloat either A==F[:Q]*F[:R]
QRCompactWY BlasFloat Val{false} A==F[:Q]*F[:R]
QRPivoted BlasFloat Val{true} A[:,F[:p]]==F[:Q]*F[:R]

BlasFloat refers to any of: Float32, Float64, Complex64 or Complex128.

The individual components of the factorization F can be accessed by indexing:

Component Description QR QRCompactWY QRPivoted
F[:Q] Q (orthogonal/unitary) part of QR ✓ (QRPackedQ) ✓ (QRCompactWYQ) ✓ (QRPackedQ)
F[:R] R (upper right triangular) part of QR
F[:p] pivot Vector    
F[:P] (pivot) permutation Matrix    

The following functions are available for the QR objects: size, \. When A is rectangular, \ will return a least squares solution and if the solution is not unique, the one with smallest norm is returned.

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 Q matrix can be converted into a regular matrix with full() which has a named argument thin.

Note

qrfact returns multiple types because LAPACK uses several representations that minimize the memory storage requirements of products of Householder elementary reflectors, so that the Q and R matrices can be stored compactly rather as two separate dense matrices.

The data contained in QR or QRPivoted can be used to construct the QRPackedQ type, which is a compact representation of the rotation matrix:

\[Q = \prod_{i=1}^{\min(m,n)} (I - \tau_i v_i v_i^T)\]

where \(\tau_i\) is the scale factor and \(v_i\) is the projection vector associated with the \(i^{th}\) Householder elementary reflector.

The data contained in QRCompactWY can be used to construct the QRCompactWYQ type, which is a compact representation of the rotation matrix

\[Q = I + Y T Y^T\]

where Y is \(m \times r\) lower trapezoidal and T is \(r \times r\) upper triangular. The compact WY representation [Schreiber1989] is not to be confused with the older, WY representation [Bischof1987]. (The LAPACK documentation uses V in lieu of Y.)

[Bischof1987](1, 2) C Bischof and C Van Loan, “The WY representation for products of Householder matrices”, SIAM J Sci Stat Comput 8 (1987), s2-s13. doi:10.1137/0908009
[Schreiber1989]R Schreiber and C Van Loan, “A storage-efficient WY representation for products of Householder transformations”, SIAM J Sci Stat Comput 10 (1989), 53-57. doi:10.1137/0910005
qrfact(A) → SPQR.Factorization

Compute the QR factorization of a sparse matrix A. A fill-reducing permutation is used. The main application of this type is to solve least squares problems with \. The function calls the C library SPQR and a few additional functions from the library are wrapped but not exported.

qrfact!(A[, pivot=Val{false}])

qrfact! is the same as qrfact() when A is a subtype of StridedMatrix, but saves space by overwriting the input A, instead of creating a copy.

full(QRCompactWYQ[, thin=true]) → Matrix

Converts an orthogonal or unitary matrix stored as a QRCompactWYQ object, i.e. in the compact WY format [Bischof1987], to a dense matrix.

Optionally takes a thin Boolean argument, which if true omits the columns that span the rows of R in the QR factorization that are zero. The resulting matrix is the Q in a thin QR factorization (sometimes called the reduced QR factorization). If false, returns a Q that spans all rows of R in its corresponding QR factorization.

bkfact(A) → BunchKaufman

Compute the Bunch-Kaufman [Bunch1977] 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.

[Bunch1977]J R Bunch and L Kaufman, Some stable methods for calculating inertia and solving symmetric linear systems, Mathematics of Computation 31:137 (1977), 163-179. url.
bkfact!(A) → BunchKaufman

bkfact! is the same as bkfact(), but saves space by overwriting the input A, instead of creating a copy.

eig(A,[irange,][vl,][vu,][permute=true,][scale=true]) → D, V

Computes eigenvalues and eigenvectors of A. See eigfact() for details on the balance keyword argument.

julia> eig([1.0 0.0 0.0; 0.0 3.0 0.0; 0.0 0.0 18.0])
([1.0,3.0,18.0],
3x3 Array{Float64,2}:
 1.0  0.0  0.0
 0.0  1.0  0.0
 0.0  0.0  1.0)

eig is a wrapper around eigfact(), extracting all parts of the factorization to a tuple; where possible, using eigfact() is recommended.

eig(A, B) → D, V

Computes generalized eigenvalues and vectors of A with respect to B.

eig is a wrapper around eigfact(), extracting all parts of the factorization to a tuple; where possible, using eigfact() is recommended.

eigvals(A,[irange,][vl,][vu])

Returns the eigenvalues of A. If A is Symmetric, Hermitian or SymTridiagonal, it is possible to calculate only a subset of the eigenvalues by specifying either a UnitRange irange covering indices of the sorted eigenvalues, or a pair vl and vu for the lower and upper boundaries of the eigenvalues.

For general non-symmetric matrices it is possible to specify how the matrix is balanced before the eigenvector calculation. The option permute=true permutes the matrix to become closer to upper triangular, and scale=true scales the matrix by its diagonal elements to make rows and columns more equal in norm. The default is true for both options.

eigmax(A)

Returns the largest eigenvalue of A.

eigmin(A)

Returns the smallest eigenvalue of A.

eigvecs(A, [eigvals,][permute=true,][scale=true]) → Matrix

Returns a matrix M whose columns are the eigenvectors of A. (The kth eigenvector can be obtained from the slice M[:, k].) The permute and scale keywords are the same as for eigfact().

For SymTridiagonal matrices, if the optional vector of eigenvalues eigvals is specified, returns the specific corresponding eigenvectors.

eigfact(A,[irange,][vl,][vu,][permute=true,][scale=true]) → Eigen

Computes the eigenvalue decomposition of A, returning an Eigen factorization object F which contains the eigenvalues in F[:values] and the eigenvectors in the columns of the matrix F[:vectors]. (The kth eigenvector can be obtained from the slice F[:vectors][:, k].)

The following functions are available for Eigen objects: inv, det.

If A is Symmetric, Hermitian or SymTridiagonal, it is possible to calculate only a subset of the eigenvalues by specifying either a UnitRange irange covering indices of the sorted eigenvalues or a pair vl and vu for the lower and upper boundaries of the eigenvalues.

For general nonsymmetric matrices it is possible to specify how the matrix is balanced before the eigenvector calculation. The option permute=true permutes the matrix to become closer to upper triangular, and scale=true scales the matrix by its diagonal elements to make rows and columns more equal in norm. The default is true for both options.

eigfact(A, B) → GeneralizedEigen

Computes the generalized eigenvalue decomposition of A and B, returning a GeneralizedEigen factorization object F which contains the generalized eigenvalues in F[:values] and the generalized eigenvectors in the columns of the matrix F[:vectors]. (The kth generalized eigenvector can be obtained from the slice F[:vectors][:, k].)

eigfact!(A[, B])

Same as eigfact(), but saves space by overwriting the input A (and B), instead of creating a copy.

hessfact(A)

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!(A)

hessfact! is the same as hessfact(), but saves space by overwriting the input A, instead of creating a copy.

schurfact(A) → Schur

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].

schurfact!(A)

Computes the Schur factorization of A, overwriting A in the process. See schurfact()

schur(A) → Schur[:T], Schur[:Z], Schur[:values]

See schurfact()

ordschur(Q, T, select) → Schur

Reorders the Schur factorization of a real matrix A=Q*T*Q' according to the logical array select returning a Schur object F. The selected eigenvalues appear in the leading diagonal of F[:Schur] and the the corresponding leading columns of F[:vectors] form an orthonormal basis of the corresponding right invariant subspace. A complex conjugate pair of eigenvalues must be either both included or excluded via select.

ordschur!(Q, T, select) → Schur

Reorders the Schur factorization of a real matrix A=Q*T*Q', overwriting Q and T in the process. See ordschur()

ordschur(S, select) → Schur

Reorders the Schur factorization S of type Schur.

ordschur!(S, select) → Schur

Reorders the Schur factorization S of type Schur, overwriting S in the process. See ordschur()

schurfact(A, B) → GeneralizedSchur

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].

schur(A, B) → GeneralizedSchur[:S], GeneralizedSchur[:T], GeneralizedSchur[:Q], GeneralizedSchur[:Z]

See schurfact()

ordschur(S, T, Q, Z, select) → GeneralizedSchur

Reorders the Generalized Schur factorization of a matrix (A, B) = (Q*S*Z^{H}, Q*T*Z^{H}) according to the logical array select and returns a GeneralizedSchur object GS. The selected eigenvalues appear in the leading diagonal of both (GS[:S], GS[:T]) and the left and right unitary/orthogonal Schur vectors are also reordered such that (A, B) = GS[:Q]*(GS[:S], GS[:T])*GS[:Z]^{H} still holds and the generalized eigenvalues of A and B can still be obtained with GS[:alpha]./GS[:beta].

ordschur!(S, T, Q, Z, select) → GeneralizedSchur

Reorders the Generalized Schur factorization of a matrix by overwriting the matrices (S, T, Q, Z) in the process. See ordschur().

ordschur(GS, select) → GeneralizedSchur

Reorders the Generalized Schur factorization of a Generalized Schur object. See ordschur().

ordschur!(GS, select) → GeneralizedSchur

Reorders the Generalized Schur factorization of a Generalized Schur object by overwriting the object with the new factorization. See ordschur().

svdfact(A[, thin=true]) → SVD

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!(A[, thin=true]) → SVD

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.

svd(A[, thin=true]) → U, S, V

Wrapper around svdfact extracting all parts the factorization to a tuple. Direct use of svdfact is therefore generally more efficient. Computes 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. The default is to produce a thin decomposition.

svdvals(A)

Returns the singular values of A.

svdvals!(A)

Returns the singular values of A, while saving space by overwriting the input.

svdfact(A, B) → GeneralizedSVD

Compute the generalized SVD of A and B, returning a GeneralizedSVD Factorization object F, such that A = F[:U]*F[:D1]*F[:R0]*F[:Q]' and B = F[:V]*F[:D2]*F[:R0]*F[:Q]'.

svd(A, B) → U, V, Q, D1, D2, R0

Wrapper around svdfact extracting all parts the factorization to a tuple. Direct use of svdfact is therefore generally more efficient. The function returns 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'.

svdvals(A, B)

Return only the singular values from the generalized singular value decomposition of A and B.

triu(M)

Upper triangle of a matrix.

triu(M, k)

Returns the upper triangle of M starting from the kth superdiagonal.

triu!(M)

Upper triangle of a matrix, overwriting M in the process.

triu!(M, k)

Returns the upper triangle of M starting from the kth superdiagonal, overwriting M in the process.

tril(M)

Lower triangle of a matrix.

tril(M, k)

Returns the lower triangle of M starting from the kth superdiagonal.

tril!(M)

Lower triangle of a matrix, overwriting M in the process.

tril!(M, k)

Returns the lower triangle of M starting from the kth superdiagonal, overwriting M in the process.

diagind(M[, k])

A Range giving the indices of the kth diagonal of the matrix M.

diag(M[, k])

The kth diagonal of a matrix, as a vector. Use diagm to construct a diagonal matrix.

diagm(v[, k])

Construct a diagonal matrix and place v on the kth diagonal.

scale(A, b)
scale(b, A)

Scale an array A by a scalar b, returning a new array.

If A is a matrix and b is a vector, then scale(A,b) scales each column i of A by b[i] (similar to A*diagm(b)), while scale(b,A) scales each row i of A by b[i] (similar to diagm(b)*A), returning a new array.

Note: for large A, scale can be much faster than A .* b or b .* A, due to the use of BLAS.

scale!(A, b)
scale!(b, A)

Scale an array A by a scalar b, similar to scale() but overwriting A in-place.

If A is a matrix and b is a vector, then scale!(A,b) scales each column i of A by b[i] (similar to A*diagm(b)), while scale!(b,A) scales each row i of A by b[i] (similar to diagm(b)*A), again operating in-place on A.

Tridiagonal(dl, d, du)

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().

Bidiagonal(dv, ev, isupper)

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().

SymTridiagonal(d, du)

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().

rank(M)

Compute the rank of a matrix.

norm(A[, p])

Compute the p-norm of a vector or the operator norm of a matrix A, defaulting to the p=2-norm.

For vectors, p can assume any numeric value (even though not all values produce a mathematically valid vector norm). In particular, norm(A, Inf) returns the largest value in abs(A), whereas norm(A, -Inf) returns the smallest.

For matrices, the matrix norm induced by the vector p-norm is used, where valid values of p are 1, 2, or Inf. (Note that for sparse matrices, p=2 is currently not implemented.) Use vecnorm() to compute the Frobenius norm.

vecnorm(A[, p])

For any iterable container A (including arrays of any dimension) of numbers (or any element type for which norm is defined), compute the p-norm (defaulting to p=2) as if A were a vector of the corresponding length.

For example, if A is a matrix and p=2, then this is equivalent to the Frobenius norm.

cond(M[, p])

Condition number of the matrix M, computed using the operator p-norm. Valid values for p are 1, 2 (default), or Inf.

condskeel(M[, x, p])
\[\begin{split}\kappa_S(M, p) & = \left\Vert \left\vert M \right\vert \left\vert M^{-1} \right\vert \right\Vert_p \\ \kappa_S(M, x, p) & = \left\Vert \left\vert M \right\vert \left\vert M^{-1} \right\vert \left\vert x \right\vert \right\Vert_p\end{split}\]

Skeel condition number \(\kappa_S\) of the matrix M, optionally with respect to the vector x, as computed using the operator p-norm. p is Inf by default, if not provided. Valid values for p are 1, 2, or Inf.

This quantity is also known in the literature as the Bauer condition number, relative condition number, or componentwise relative condition number.

trace(M)

Matrix trace

det(M)

Matrix determinant

logdet(M)

Log of matrix determinant. Equivalent to log(det(M)), but may provide increased accuracy and/or speed.

logabsdet(M)

Log of absolute value of determinant of real matrix. Equivalent to (log(abs(det(M))), sign(det(M))), but may provide increased accuracy and/or speed.

inv(M)

Matrix inverse

pinv(M[, tol])

Computes the Moore-Penrose pseudoinverse.

For matrices M with floating point elements, it is convenient to compute the pseudoinverse by inverting only singular values above a given threshold, tol.

The optimal choice of tol varies both with the value of M and the intended application of the pseudoinverse. The default value of tol is eps(real(float(one(eltype(M)))))*maximum(size(A)), which is essentially machine epsilon for the real part of a matrix element multiplied by the larger matrix dimension. For inverting dense ill-conditioned matrices in a least-squares sense, tol = sqrt(eps(real(float(one(eltype(M)))))) is recommended.

For more information, see [8859], [B96], [S84], [KY88].

[8859]Issue 8859, “Fix least squares”, https://github.com/JuliaLang/julia/pull/8859
[B96]Åke Björck, “Numerical Methods for Least Squares Problems”, SIAM Press, Philadelphia, 1996, “Other Titles in Applied Mathematics”, Vol. 51. doi:10.1137/1.9781611971484
[S84]G. W. Stewart, “Rank Degeneracy”, SIAM Journal on Scientific and Statistical Computing, 5(2), 1984, 403-413. doi:10.1137/0905030
[KY88]Konstantinos Konstantinides and Kung Yao, “Statistical analysis of effective singular values in matrix rank determination”, IEEE Transactions on Acoustics, Speech and Signal Processing, 36(5), 1988, 757-763. doi:10.1109/29.1585
nullspace(M)

Basis for nullspace of M.

repmat(A, n, m)

Construct a matrix by repeating the given matrix n times in dimension 1 and m times in dimension 2.

repeat(A, inner = Int[], outer = Int[])

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.

kron(A, B)

Kronecker tensor product of two vectors or two matrices.

blkdiag(A...)

Concatenate matrices block-diagonally. Currently only implemented for sparse matrices.

linreg(x, y) → [a; b]

Linear Regression. Returns a and b such that a+b*x is the closest line to the given points (x,y). In other words, this function determines parameters [a, b] that minimize the squared error between y and a+b*x.

Example:

using PyPlot;
x = float([1:12])
y = [5.5; 6.3; 7.6; 8.8; 10.9; 11.79; 13.48; 15.02; 17.77; 20.81; 22.0; 22.99]
a, b = linreg(x,y) # Linear regression
plot(x, y, "o") # Plot (x,y) points
plot(x, [a+b*i for i in x]) # Plot the line determined by the linear regression
linreg(x, y, w)

Weighted least-squares linear regression.

expm(A)

Compute the matrix exponential of A, defined by

\[e^A = \sum_{n=0}^{\infty} \frac{A^n}{n!}.\]

For symmetric or Hermitian A, an eigendecomposition (eigfact()) is used, otherwise the scaling and squaring algorithm (see [H05]) is chosen.

[H05]Nicholas J. Higham, “The squaring and scaling method for the matrix exponential revisited”, SIAM Journal on Matrix Analysis and Applications, 26(4), 2005, 1179-1193. doi:10.1137/090768539
logm(A)

If A has no negative real eigenvalue, compute the principal matrix logarithm of A, i.e. the unique matrix \(X\) such that \(e^X = A\) and \(-\pi < Im(\lambda) < \pi\) for all the eigenvalues \(\lambda\) of \(X\). If A has nonpositive eigenvalues, a warning is printed and whenever possible a nonprincipal matrix function is returned.

If A is symmetric or Hermitian, its eigendecomposition (eigfact()) is used, if A is triangular an improved version of the inverse scaling and squaring method is employed (see [AH12] and [AHR13]). For general matrices, the complex Schur form (schur()) is computed and the triangular algorithm is used on the triangular factor.

[AH12]Awad H. Al-Mohy and Nicholas J. Higham, “Improved inverse scaling and squaring algorithms for the matrix logarithm”, SIAM Journal on Scientific Computing, 34(4), 2012, C153-C169. doi:10.1137/110852553
[AHR13]Awad H. Al-Mohy, Nicholas J. Higham and Samuel D. Relton, “Computing the Fréchet derivative of the matrix logarithm and estimating the condition number”, SIAM Journal on Scientific Computing, 35(4), 2013, C394-C410. doi:10.1137/120885991
sqrtm(A)

If A has no negative real eigenvalues, compute the principal matrix square root of A, that is the unique matrix \(X\) with eigenvalues having positive real part such that \(X^2 = A\). Otherwise, a nonprincipal square root is returned.

If A is symmetric or Hermitian, its eigendecomposition (eigfact()) is used to compute the square root. Otherwise, the square root is determined by means of the Björck-Hammarling method, which computes the complex Schur form (schur()) and then the complex square root of the triangular factor.

[BH83]Åke Björck and Sven Hammarling, “A Schur method for the square root of a matrix”, Linear Algebra and its Applications, 52-53, 1983, 127-140. doi:10.1016/0024-3795(83)80010-X
lyap(A, C)

Computes the solution X to the continuous Lyapunov equation AX + XA' + C = 0, where no eigenvalue of A has a zero real part and no two eigenvalues are negative complex conjugates of each other.

sylvester(A, B, C)

Computes the solution X to the Sylvester equation AX + XB + C = 0, where A, B and C have compatible dimensions and A and -B have no eigenvalues with equal real part.

issym(A) → Bool

Test whether a matrix is symmetric.

isposdef(A) → Bool

Test whether a matrix is positive definite.

isposdef!(A) → Bool

Test whether a matrix is positive definite, overwriting A in the processes.

istril(A) → Bool

Test whether a matrix is lower triangular.

istriu(A) → Bool

Test whether a matrix is upper triangular.

isdiag(A) → Bool

Test whether a matrix is diagonal.

ishermitian(A) → Bool

Test whether a matrix is Hermitian.

transpose(A)

The transposition operator (.').

transpose!(dest, src)

Transpose array src and store the result in the preallocated array dest, which should have a size corresponding to (size(src,2),size(src,1)). No in-place transposition is supported and unexpected results will happen if src and dest have overlapping memory regions.

ctranspose(A)

The conjugate transposition operator (').

ctranspose!(dest, src)

Conjugate transpose array src and store the result in the preallocated array dest, which should have a size corresponding to (size(src,2),size(src,1)). No in-place transposition is supported and unexpected results will happen if src and dest have overlapping memory regions.

eigs(A, [B, ]; nev=6, which="LM", tol=0.0, maxiter=300, sigma=nothing, ritzvec=true, v0=zeros((0, ))) -> (d, [v, ]nconv, niter, nmult, resid)

Computes eigenvalues d of A using Lanczos or Arnoldi iterations for real symmetric or general nonsymmetric matrices respectively. If B is provided, the generalized eigenproblem is solved.

The following keyword arguments are supported:
  • nev: Number of eigenvalues

  • ncv: Number of Krylov vectors used in the computation; should satisfy nev+1 <= ncv <= n for real symmetric problems and nev+2 <= ncv <= n for other problems, where n is the size of the input matrix A. The default is ncv = max(20,2*nev+1).

    Note that these restrictions limit the input matrix A to be of dimension at least 2.

  • which: type of eigenvalues to compute. See the note below.

    which type of eigenvalues
    :LM eigenvalues of largest magnitude (default)
    :SM eigenvalues of smallest magnitude
    :LR eigenvalues of largest real part
    :SR eigenvalues of smallest real part
    :LI eigenvalues of largest imaginary part (nonsymmetric or complex A only)
    :SI eigenvalues of smallest imaginary part (nonsymmetric or complex A only)
    :BE compute half of the eigenvalues from each end of the spectrum, biased in favor of the high end. (real symmetric A only)
  • tol: tolerance (\(tol \le 0.0\) defaults to DLAMCH('EPS'))

  • maxiter: Maximum number of iterations (default = 300)

  • sigma: Specifies the level shift used in inverse iteration. If nothing (default), defaults to ordinary (forward) iterations. Otherwise, find eigenvalues close to sigma using shift and invert iterations.

  • ritzvec: Returns the Ritz vectors v (eigenvectors) if true

  • v0: starting vector from which to start the iterations

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.

Note

The sigma and which keywords interact: the description of eigenvalues searched for by which do _not_ necessarily refer to the eigenvalues of A, but rather the linear operator constructed by the specification of the iteration mode implied by sigma.

sigma iteration mode which refers to eigenvalues of
nothing ordinary (forward) \(A\)
real or complex inverse with level shift sigma \((A - \sigma I )^{-1}\)
svds(A; nsv=6, ritzvec=true, tol=0.0, maxiter=1000) -> (left_sv, s, right_sv, nconv, niter, nmult, resid)

svds computes largest singular values s of A using Lanczos or Arnoldi iterations. Uses eigs() underneath.

Inputs are:
  • A: Linear operator. It can either subtype of AbstractArray (e.g., sparse matrix) or duck typed. For duck typing A has to support size(A), eltype(A), A * vector and A' * vector.
  • nsv: Number of singular values.
  • ritzvec: Whether to return the left and right singular vectors left_sv and right_sv, default is true. If false the singular vectors are omitted from the output.
  • tol: tolerance, see eigs().
  • maxiter: Maximum number of iterations, see eigs().

Example:

X = sprand(10, 5, 0.2)
svds(X, nsv = 2)
peakflops(n; parallel=false)

peakflops computes the peak flop rate of the computer by using double precision Base.LinAlg.BLAS.gemm!(). 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.

BLAS Functions

Base.LinAlg.BLAS 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(n, X, incx, Y, incy)

Dot product of two vectors consisting of n elements of array X with stride incx and n elements of array Y with stride incy.

dotu(n, X, incx, Y, incy)

Dot function for two complex vectors.

dotc(n, X, incx, U, incy)

Dot function for two complex vectors conjugating the first vector.

blascopy!(n, X, incx, Y, incy)

Copy n elements of array X with stride incx to array Y with stride incy. Returns Y.

nrm2(n, X, incx)

2-norm of a vector consisting of n elements of array X with stride incx.

asum(n, X, incx)

sum of the absolute values of the first n elements of array X with stride incx.

axpy!(a, X, Y)

Overwrite Y with a*X + Y. Returns Y.

scal!(n, a, X, incx)

Overwrite X with a*X. Returns X.

scal(n, a, X, incx)

Returns a*X.

ger!(alpha, x, y, A)

Rank-1 update of the matrix A with vectors x and y as alpha*x*y' + A.

syr!(uplo, alpha, x, A)

Rank-1 update of the symmetric matrix A with vector x as alpha*x*x.' + A. When uplo is ‘U’ the upper triangle of A is updated (‘L’ for lower triangle). Returns A.

syrk!(uplo, trans, alpha, A, beta, C)

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.

syrk(uplo, trans, alpha, A)

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’).

her!(uplo, alpha, x, A)

Methods for complex arrays only. Rank-1 update of the Hermitian matrix A with vector x as alpha*x*x' + A. When uplo is ‘U’ the upper triangle of A is updated (‘L’ for lower triangle). Returns A.

herk!(uplo, trans, alpha, A, beta, C)

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.

herk(uplo, trans, alpha, A)

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’).

gbmv!(trans, m, kl, ku, alpha, A, x, beta, y)

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.

gbmv(trans, m, kl, ku, alpha, A, x, beta, 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.

sbmv!(uplo, k, alpha, A, x, beta, y)

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.

sbmv(uplo, k, alpha, A, x)

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.

sbmv(uplo, k, A, x)

Returns A*x where A is a symmetric band matrix of order size(A,2) with k super-diagonals stored in the argument A.

gemm!(tA, tB, alpha, A, B, beta, C)

Update C as alpha*A*B + beta*C or the other three variants according to tA (transpose A) and tB. Returns the updated C.

gemm(tA, tB, alpha, A, B)

Returns alpha*A*B or the other three variants according to tA (transpose A) and tB.

gemm(tA, tB, A, B)

Returns A*B or the other three variants according to tA (transpose A) and tB.

gemv!(tA, alpha, A, x, beta, y)

Update the vector y as alpha*A*x + beta*y or alpha*A'x + beta*y according to tA (transpose A). Returns the updated y.

gemv(tA, alpha, A, x)

Returns alpha*A*x or alpha*A'x according to tA (transpose A).

gemv(tA, A, x)

Returns A*x or A'x according to tA (transpose A).

symm!(side, ul, alpha, A, B, beta, C)

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.

symm(side, ul, alpha, A, B)

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.

symm(side, ul, A, B)

Returns A*B or B*A according to side. A is assumed to be symmetric. Only the ul triangle of A is used.

symm(tA, tB, alpha, A, B)

Returns alpha*A*B or the other three variants according to tA (transpose A) and tB.

symv!(ul, alpha, A, x, beta, y)

Update the vector y as alpha*A*x + beta*y. A is assumed to be symmetric. Only the ul triangle of A is used. Returns the updated y.

symv(ul, alpha, A, x)

Returns alpha*A*x. A is assumed to be symmetric. Only the ul triangle of A is used.

symv(ul, A, x)

Returns A*x. A is assumed to be symmetric. Only the ul triangle of A is used.

trmm!(side, ul, tA, dA, alpha, A, B)

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.

trmm(side, ul, tA, dA, alpha, A, 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).

trsm!(side, ul, tA, dA, alpha, A, B)

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.

trsm(side, ul, tA, dA, alpha, A, 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).

trmv!(side, ul, tA, dA, alpha, A, b)

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.

trmv(side, ul, tA, dA, alpha, A, 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).

trsv!(ul, tA, dA, A, b)

Overwrite b with the solution to A*x = b or one of the other two variants determined by tA (transpose A) and ul (triangle of A used). dA indicates if A is unit-triangular (the diagonal is assumed to be all ones). Returns the updated b.

trsv(ul, tA, dA, A, b)

Returns the solution to A*x = b or one of the other two variants determined by tA (transpose A) and ul (triangle of A is used.) dA indicates if A is unit-triangular (the diagonal is assumed to be all ones).

blas_set_num_threads(n)

Set the number of threads the BLAS library should use.

I

An object of type UniformScaling, representing an identity matrix of any size.