A Julia package for defining and working with linear maps, also known as linear transformations or linear operators acting on vectors. The only requirement for a LinearMap is that it can act on a vector (by multiplication) efficiently.
-
Fully julia v0.6 compatible; dropped compatibility for previous versions of Julia.
-
LinearMap
is now the name of the abstract type on top (used to beAbstractLinearMap
because you could not have a function/constructor with the same name as an abstract type in older julia versions) -
Specifying the
eltype
of a function to be used as linear map should now use the constructorLinearMap{T}(f, ...)
.
Install with the package manager, i.e. Pkg.add("LinearMaps")
.
Several iterative linear algebra methods such as linear solvers or eigensolvers only require an efficient evaluation of the matrix vector product, where the concept of a matrix can be formalized / generalized to a linear map (or linear operator in the special case of a square matrix).
The LinearMaps package provides the following functionality:
-
A
LinearMap
type that shares with theAbstractMatrix
type that it responds to the functionssize
,eltype
,isreal
,issymmetric
,ishermitian
andisposdef
,transpose
andctranspose
and multiplication with a vector using both*
or the in-place versionA_mul_B!
. Depending on the subtype, alsoAt_mul_B
,At_mul_B!
,Ac_mul_B
andAc_mul_B!
are supported. Linear algebra functions that uses duck-typing for its arguments can handleLinearMap
objects similar toAbstractMatrix
objects, provided that they can be written using the above methods. UnlikeAbstractMatrix
types,LinearMap
objects cannot be indexed, neither usinggetindex
orsetindex!
. -
A single method
LinearMap
function that acts as a general purpose constructor (though it only an abstract type) and allows to construct linear map objects from functions, or to wrap objects of typeAbstractMatrix
orLinearMap
. The latter functionality is useful to (re)define the properties (isreal
,issymmetric
,ishermitian
,isposdef
) of the existing matrix or linear map. -
A framework for combining objects of type
LinearMap
and of typeAbstractMatrix
using linear combinations, transposition and composition, where the linear map resulting from these operations is never explicitly evaluated but only its matrix vector product is defined (i.e. lazy evaluation). The matrix vector product is written to minimize memory allocation by using a minimal number of temporary vectors. There is full support for the in-place versionA_mul_B!
, which should be preferred for higher efficiency in critical algorithms. In addition, it tries to recognize the properties of combinations of linear maps. In particular, compositions such asA'*A
for arbitraryA
or evenA'*B*C*B'*A
with arbitraryA
andB
and positive definiteC
are recognized as being positive definite and hermitian. In case a certain property of the resultingLinearMap
object is not correctly inferred, theLinearMap
method can be called to redefine the properties.
-
LinearMap
General purpose method to construct
LinearMap
objects of specific types, as described in the Types section belowLinearMap{T}(A::AbstractMatrix[; isreal::Bool, issymmetric::Bool, ishermitian::Bool, isposdef::Bool]) LinearMap{T}(A::LinearMap[; isreal::Bool, issym::Bool, ishermitian::Bool, isposdef::Bool])
Create a
WrappedMap
object that will respond to the methodsisreal
,issymmetric
,ishermitian
,isposdef
with the values provided by the keyword arguments, and toeltype
with the valueT
. The default values correspond to the result of calling these methods on the argumentA
; in particular{T}
does not need to be specified and is set aseltype(A)
. This allows to use anAbstractMatrix
within theLinearMap
framework and to redefine the properties of an existingLinearMap
.LinearMap{T}(f, [fc = nothing], M::Int, [N::Int = M]; ismutating::Bool, issymmetric::Bool, ishermitian::Bool, isposdef::Bool])
Create a
FunctionMap
instance that wraps an object describing the action of the linear map on a vector as a function call. Here,f
can be a function or any other object for which either the callf(src::AbstractVector) -> dest::AbstractVector
(whenismutating = false
) orf(dest::AbstractVector,src::AbstractVector) -> dest
(whenismutating = true
) is supported. The value ofismutating
can be spefified, by default its value is guessed by looking at the number of arguments of the first method in the method list off
.A second function or object can optionally be provided that implements the action of the adjoint (transposed) linear map. Here, it is always assumed that this represents the conjugate transpose, though this is of course equivalent to the normal transpose for real linear maps. Furthermore, the conjugate transpose also enables the use of
At_mul_B(!)
using some extra conjugation calls on the input and output vector. If no second function is provided, thanAt_mul_B(!)
andAc_mul_B(!)
cannot be used with this linear map, unless it is symmetric or hermitian.M
is the number of rows (length of the output vectors) andN
the number of columns (length of the input vectors). When the latter is not specified,N = M
.Finally, one can specify the
eltype
of the resulting linear map using the type parameterT
. If not specified, a default value ofFloat64
is assumed. Use a complex typeT
if the function represents a complex linear map.In summary, the keyword arguments and their default values are:
ismutating
:false
if the functionf
accepts a single vector argument corresponding to the input, andtrue
if they accept two vector arguments where the first will be mutated so as to contain the result. In both cases, the resultingA::FunctionMap
will support both the mutating as non-mutating matrix vector multiplication. Default value is guessed based on the number of arguments for the first method in the method list off
; it is not possible to usef
andfc
where only one of the two is mutating and the other is not.issymmetric [=false]
: whether the function represents the multiplication with a symmetric matrix. Iftrue
, this will automatically enableA'*x
andA.'*x
.ishermitian [=T<:Real && issymmetric]
: whether the function represents the multiplication with a hermitian matrix. Iftrue
, this will automatically enableA'*x
andA.'*x
.isposdef [=false]
: whether the function represents the multiplication with a positive definite matrix.
-
Base.full(linearmap)
Creates a full matrix representation of the linearmap object, by multiplying it with the successive basis vectors. This is mostly for testing purposes or if you want to have the explicit matrix representation of a linear map for which you only have a function definition (e.g. to be able to use its
(c)transpose
). -
All matrix multiplication methods and the corresponding mutating versions.
None of the types below need to be constructed directly; they arise from performing operations between LinearMap
objects or by calling the LinearMap
constructor described above.
-
LinearMap
Abstract supertype
-
FunctionMap
Type for wrapping an arbitrary function that is supposed to implement the matrix vector product as a
LinearMap
. -
WrappedMap
Type for wrapping an
AbstractMatrix
orLinearMap
and to possible redefine the propertiesisreal
,issymmetric
,ishermitian
andisposdef
. AnAbstractMatrix
will automatically be converted to aWrappedMap
when it is combined with otherAbstractLinearMap
objects via linear combination or composition (multiplication). Note thatWrappedMap(mat1)*WrappedMap(mat2)
will never evaluatemat1*mat2
, since this is more costly then evaluatingmat1*(mat2*x)
and the latter is the only operation that needs to be performed byLinearMap
objects anyway. While the cost of matrix addition is comparable to matrix vector multiplication, this too is not performed explicitly since this would require new storage of the same amount as of the original matrices. -
IdentityMap
Type for representing the identity map of a certain size
M=N
, obtained simply asIdentityMap{T}(M)
,IdentityMap(T,M)=IdentityMap(T,M,N)=IdentityMap(T,(M,N))
or evenIdentityMap(M)=IdentityMap(M,N)=IdentityMap((M,N))
. IfT
is not specified,Bool
is assumed, since operations betweenBool
and any otherNumber
will always be converted to the type of the otherNumber
. IfM!=N
, an error is returned. AnIdentityMap
of the correct size and element type will automatically be created ifLinearMap
objects are combined withI
, Julia's built in identity (UniformScaling
). -
LinearCombination
,CompositeMap
,TransposeMap
andCTransposeMap
Used to add and multiply
LinearMap
objects, don't need to be constructed explicitly.
The LinearMap
object combines well with the iterative eigensolver eigs
, which is the Julia wrapper for Arpack.
using LinearMaps
function leftdiff!(y::AbstractVector, x::AbstractVector) # left difference assuming periodic boundary conditions
N=length(x)
length(y)==N || throw(DimensionMismatch())
@inbounds for i=1:N
y[i]=x[i]-x[mod1(i-1,N)]
end
return y
end
function mrightdiff!(y::AbstractVector, x::AbstractVector) # minus right difference
N=length(x)
length(y)==N || throw(DimensionMismatch())
@inbounds for i=1:N
y[i]=x[i]-x[mod1(i+1,N)]
end
return y
end
D=LinearMap(leftdiff!, mrightdiff!, 100; ismutating=true)
eigs(D'*D;nev=3,which=:SR)