Sparse linear algebra

Introduction

This chapter deals with sparse linear algebra over commutative rings and fields.

Sparse linear algebra, that is, linear algebra with sparse matrices, plays an important role in various algorithms in algebraic number theory. For example, it is one of the key ingredients in the computation of class groups and discrete logarithms using index calculus methods.

Sparse rows

Building blocks for sparse matrices are sparse rows, which are modelled by objects of type \texttt{SRow}. More precisely, the type is of parametrized form objects of type SRow. More precisely, the type is of parametrized form SRow{T}, where T is the element type of the base ring $R$. For example, SRow{fmpz} is the type for sparse rows over the integers.

It is important to note that sparse rows do not have a fixed number of columns, that is, they represent elements of $\{ (x_i)_i \in R^{\mathbb{N}} \mid x_i = 0 \text{ for almost all }i\}$. In particular any two sparse rows over the same base ring can be added.

Creation

sparse_rowMethod
sparse_row(R::Ring, J::Vector{Tuple{Int, T}}) -> SRow{T}

Constructs the sparse row $(a_i)_i$ with $a_{i_j} = x_j$, where $J = (i_j, x_j)_j$. The elements $x_i$ must belong to the ring $R$.

sparse_rowMethod
sparse_row(R::Ring, J::Vector{Tuple{Int, T}}) -> SRow{T}

Constructs the sparse row $(a_i)_i$ with $a_{i_j} = x_j$, where $J = (i_j, x_j)_j$. The elements $x_i$ must belong to the ring $R$.

sparse_row(R::Ring, J::Vector{Tuple{Int, Int}}) -> SRow

Constructs the sparse row $(a_i)_i$ over $R$ with $a_{i_j} = x_j$, where $J = (i_j, x_j)_j$.

sparse_rowMethod
sparse_row(R::Ring, J::Vector{Int}, V::Vector{T}) -> SRow{T}

Constructs the sparse row $(a_i)_i$ over $R$ with $a_{i_j} = x_j$, where $J = (i_j)_j$ and $V = (x_j)_j$.

Basic operations

Rows support the usual operations:

  • +, -, ==
  • multiplication by scalars
  • div, divexact
getindexMethod
getindex(A::SRow, j::Int) -> RingElem

Given a sparse row $(a_i)_{i}$ and an index $j$ return $a_j$.

add_scaled_rowMethod
add_scaled_row(A::SRow{T}, B::SRow{T}, c::T) -> SRow{T}

Returns the row $c A + B$.

add_scaled_rowMethod
add_scaled_row(A::SRow{T}, B::SRow{T}, c::T) -> SRow{T}

Returns the row $c A + B$.

transform_rowMethod
transform_row(A::SRow{T}, B::SRow{T}, i::Int, j::Int, a::T, b::T, c::T, d::T)

Returns the tuple $(aA + bB, cA + dB)$.

lengthMethod
length(A::SRow)

Returns the number of nonzero entries of $A$.

Change of base ring

change_base_ringMethod
change_base_ring(R::Ring, A::SRow) -> SRow

Create a new sparse row by coercing all elements into the ring $R$.

Maximum, minimum and 2-norm

maximumMethod
maximum(A::SRow{T}) -> T

Returns the largest entry of $A$.

maximumMethod
maximum(A::SRow{T}) -> T

Returns the largest entry of $A$.

minimumMethod
  minimum(A::NfRelOrdIdl) -> NfOrdIdl
  minimum(A::NfRelOrdIdl) -> NfRelOrdIdl

Returns the ideal $A \cap O$ where $O$ is the maximal order of the coefficient ideals of $A$.

minimum(A::SRow{T}) -> T

Returns the smallest entry of $A$.

minimumMethod
minimum(A::SRow{T}) -> T

Returns the smallest entry of $A$.

norm2Method
norm2(A::SRow{T} -> T

Returns $A \cdot A^t$.

Functionality for integral sparse rows

liftMethod
lift(A::SRow{nmod}) -> SRow{fmpz}

Return the sparse row obtained by lifting all entries in $A$.

mod!Method
mod!(A::SRow{fmpz}, n::fmpz) -> SRow{fmpz}

Inplace reduction of all entries of $A$ modulo $n$ to the positive residue system.

mod_sym!Method
mod_sym!(A::SRow{fmpz}, n::fmpz) -> SRow{fmpz}

Inplace reduction of all entries of $A$ modulo $n$ to the symmetric residue system.

mod_sym!Method
mod_sym!(A::SRow{fmpz}, n::Integer) -> SRow{fmpz}

Inplace reduction of all entries of $A$ modulo $n$ to the symmetric residue system.

maximumMethod
maximum(abs, A::SRow{fmpz}) -> fmpz

Returns the largest, in absolute value, entry of $A$.

Sparse matrices

Let $R$ be a commutative ring. Sparse matrices with base ring $R$ are modelled by objects of type SMat. More precisely, the type is of parametrized form SRow{T}, where T is the element type of the base ring. For example, SMat{fmpz} is the type for sparse matrices over the integers.

In constrast to sparse rows, sparse matrices have a fixed number of rows and columns, that is, they represent elements of the matrices space $\mathrm{Mat}_{n\times m}(R)$. Internally, sparse matrices are implemented as an array of sparse rows. As a consequence, unlike their dense counterparts, sparse matrices have a mutable number of rows and it is very performant to add additional rows.

Construction

sparse_matrixMethod
sparse_matrix(R::Ring) -> SMat

Return an empty sparse matrix with base ring $R$.

Sparse matrices can also be created from dense matrices as well as from julia arrays:

sparse_matrixMethod
sparse_matrix(A::MatElem; keepzrows::Bool = true)

Constructs the sparse matrix corresponding to the dense matrix $A$. If keepzrows is false, then the constructor will drop any zero row of $A$.

sparse_matrixMethod
sparse_matrix(R::Ring, A::Matrix{T}) -> SMat

Constructs the sparse matrix over $R$ corresponding to $A$.

sparse_matrixMethod
sparse_matrix(R::Ring, A::Matrix{T}) -> SMat

Constructs the sparse matrix over $R$ corresponding to $A$.

The normal way however, is to add rows:

push!Method
push!(A::SMat{T}, B::SRow{T}) where T

Appends the sparse row B to A.

Sparse matrices can also be concatenated to form larger ones:

vcat!Method
vcat!(A::SMat, B::SMat) -> SMat

Vertically joins $A$ and $B$ inplace, that is, the rows of $B$ are appended to $A$.

vcatMethod
vcat(A::SMat, B::SMat) -> SMat

Vertically joins $A$ and $B$.

hcat!Method
hcat!(A::SMat, B::SMat) -> SMat

Horizontally concatenates $A$ and $B$, inplace, changing $A$.

hcatMethod
hcat(A::SMat, B::SMat) -> SMat

Horizontally concatenates $A$ and $B$.

(Normal julia $cat$ is also supported)

There are special constructors:

identity_matrixMethod
identity_matrix(::Type{SMat}, R::Ring, n::Int)

Return a sparse $n$ times $n$ identity matrix over $R$.

zero_matrixMethod
zero_matrix(::Type{SMat}, R::Ring, n::Int)

Return a sparse $n$ times $n$ zero matrix over $R$.

zero_matrixMethod
zero_matrix(::Type{SMat}, R::Ring, n::Int, m::Int)

Return a sparse $n$ times $m$ zero matrix over $R$.

Slices:

subMethod
sub(A::SMat, r::UnitRange, c::UnitRange) -> SMat

Return the submatrix of $A$, where the rows correspond to $r$ and the columns correspond to $c$.

Transpose:

transposeMethod
transpose(A::SMat) -> SMat

Returns the transpose of $A$.

Elementary Properties

sparsityMethod
sparsity(A::SMat) -> Float64

Return the sparsity of A, that is, the number of zero-valued elements divided by the number of all elements.

densityMethod
density(A::SMat) -> Float64

Return the density of A, that is, the number of nonzero-valued elements divided by the number of all elements.

nnzMethod
nnz(A::SMat) -> Int

Return the number of non-zero entries of $A$.

nrowsMethod
nrows(A::SMat) -> Int

Return the number of rows of $A$.

ncolsMethod
ncols(A::SMat) -> Int

Return the number of columns of $A$.

isoneMethod
isone(A::SMat)

Tests if $A$ is an identity matrix.

iszeroMethod
iszero(A::SMat)

Tests if $A$ is a zero matrix.

isupper_triangularMethod
isupper_triangular(A::SMat)

Returns true if and only if $A$ is upper (right) triangular.

maximumMethod
maximum(A::SMat{T}) -> T

Finds the largest entry of $A$.

minimumMethod
minimum(A::SMat{T}) -> T

Finds the smallest entry of $A$.

maximumMethod
maximum(abs, A::SMat{fmpz}) -> fmpz

Finds the largest, in absolute value, entry of $A$.

elementary_divisorsMethod
elementary_divisors(A::SMat{fmpz}) -> Vector{fmpz}

The elementary divisors of $A$, i.e. the diagonal elements of the Smith normal form of $A$.

solve_dixon_sfMethod
solve_dixon_sf(A::SMat{fmpz}, b::SRow{fmpz}, is_int::Bool = false) -> SRow{fmpz}, fmpz
solve_dixon_sf(A::SMat{fmpz}, B::SMat{fmpz}, is_int::Bool = false) -> SMat{fmpz}, fmpz

For a sparse square matrix $A$ of full rank and a sparse matrix (row), find a sparse matrix (row) $x$ and an integer $d$ s.th. $x A = bd$ holds. The algorithm is a Dixon-based linear p-adic lifting method. If \code{is_int} is given, then $d$ is assumed to be $1$. In this case rational reconstruction is avoided.

hadamard_bound2Method
hadamard_bound2(A::SMat{T}) -> T

The square of the product of the norms of the rows of $A$.

echelon_with_transformMethod
echelon_with_transform(A::SMat{nmod}) -> SMat, SMat

Find a unimodular matrix $T$ and an upper-triangular $E$ s.th. $TA = E$ holds.

reduce_fullMethod
reduce_full(A::SMat{fmpz}, g::SRow{fmpz},
                      trafo = Val{false}) -> SRow{fmpz}, Vector{Int}

Reduces $g$ modulo $A$ and assumes that $A$ is upper triangular.

The second return value is the array of pivot elements of $A$ that changed.

If trafo is set to Val{true}, then additionally an array of transformations is returned.

hnf!Method
hnf!(A::SMat{fmpz})

Inplace transform of $A$ into upper right Hermite normal form.

hnfMethod
hnf(A::SMat{fmpz}) -> SMat{fmpz}

Return the upper right Hermite normal form of $A$.

snfMethod
snf(A::SMat{fmpz})

The Smith normal form (snf) of $A$.

hnf_extend!Method
hnf_extend!(A::SMat{fmpz}, b::SMat{fmpz}, offset::Int = 0) -> SMat{fmpz}

Given a matrix $A$ in HNF, extend this to get the HNF of the concatenation with $b$.

is_diagonalMethod
is_diagonal(A::SMat) -> Bool

True iff only the i-th entry in the i-th row is non-zero.

detMethod
det(A::SMat{fmpz})

The determinant of $A$ using a modular algorithm. Uses the dense (nmod_mat) determinant on $A$ for various primes $p$.

det_mcMethod
det_mc(A::SMat{fmpz})

Computes the determinant of $A$ using a LasVegas style algorithm, i.e. the result is not proven to be correct. Uses the dense (nmod_mat) determinant on $A$ for various primes $p$.

valence_mcMethod
valence_mc{T}(A::SMat{T}; extra_prime = 2, trans = Vector{SMatSLP_add_row{T}}()) -> T

Uses a Monte-Carlo algorithm to compute the valence of $A$. The valence is the valence of the minimal polynomial $f$ of $transpose(A)*A$, thus the last non-zero coefficient, typically $f(0)$.

The valence is computed modulo various primes until the computation stabilises for extra_prime many.

trans, if given, is a SLP (straight-line-program) in GL(n, Z). Then the valence of trans * $A$ is computed instead.

saturateMethod
saturate(A::SMat{fmpz}) -> SMat{fmpz}

Computes the saturation of $A$, that is, a basis for $\mathbf{Q}\otimes M \meet \mathbf{Z}^n$, where $M$ is the row span of $A$ and $n$ the number of rows of $A$.

Equivalently, return $TA$ for an invertible rational matrix $T$, such that $TA$ is integral and the elementary divisors of $TA$ are all trivial.

hnf_kannan_bachemMethod
hnf_kannan_bachem(A::SMat{fmpz}) -> SMat{fmpz}

Compute the Hermite normal form of $A$ using the Kannan-Bachem algorithm.

diagonal_formMethod
diagonal_form(A::SMat{fmpz}) -> SMat{fmpz}

A matrix $D$ that is diagonal and obtained via unimodular row and column operations. Like a snf without the divisibility condition.

Manipulation/ Access

getindexMethod
getindex(A::SMat, i::Int, j::Int)

Given a sparse matrix $A = (a_{ij})_{i, j}$, return the entry $a_{ij}$.

getindexMethod
getindex(A::SMat, i::Int) -> SRow

Given a sparse matrix $A$ and an index $i$, return the $i$-th row of $A$.

setindex!Method
setindex!(A::SMat, b::SRow, i::Int)

Given a sparse matrix $A$, a sparse row $b$ and an index $i$, set the $i$-th row of $A$ equal to $b$.

swap_rows!Method
swap_rows!(A::SMat{T}, i::Int, j::Int)

Swap the $i$-th and $j$-th row of $A$ inplace.

swap_cols!Method
swap_cols!(A::SMat, i::Int, j::Int)

Swap the $i$-th and $j$-th column of $A$ inplace.

scale_row!Method
scale_row!(A::SMat{T}, i::Int, c::T)

Multiply the $i$-th row of $A$ by $c$ inplace.

add_scaled_col!Method
add_scaled_col!(A::SMat{T}, i::Int, j::Int, c::T)

Add $c$ times the $i$-th column to the $j$-th column of $A$ inplace, that is, $A_j \rightarrow A_j + c \cdot A_i$, where $(A_i)_i$ denote the columns of $A$.

add_scaled_row!Method
add_scaled_row!(A::SMat{T}, i::Int, j::Int, c::T)

Add $c$ times the $i$-th row to the $j$-th row of $A$ inplace, that is, $A_j \rightarrow A_j + c \cdot A_i$, where $(A_i)_i$ denote the rows of $A$.

transform_row!Method
transform_row!(A::SMat{T}, i::Int, j::Int, a::T, b::T, c::T, d::T)

Applies the transformation $(A_i, A_j) \rightarrow (aA_i + bA_j, cA_i + dA_j)$ to $A$, where $(A_i)_i$ are the rows of $A$.

diagonalMethod
diagonal(A::SMat) -> fmpz[]

The diagonal elements of $A$ in an array.

reverse_rows!Method
reverse_rows!(A::SMat)

Inplace inversion of the rows of $A$.

mod_sym!Method
mod_sym!(A::SMat{fmpz}, n::fmpz)

Inplace reduction of all entries of $A$ modulo $n$ to the symmetric residue system.

find_row_starting_withMethod
find_row_starting_with(A::SMat, p::Int) -> Int

Tries to find the index $i$ such that $A_{i,p} \neq 0$ and $A_{i, p-j} = 0$ for all $j > 1$. It is assumed that $A$ is upper triangular. If such an index does not exist, find the smallest index larger.

reduceMethod
reduce(A::SMat{fmpz}, g::SRow{fmpz}, m::fmpz) -> SRow{fmpz}

Given an upper triangular matrix $A$ over the integers, a sparse row $g$ and an integer $m$, this function reduces $g$ modulo $A$ and returns $g$ modulo $m$ with respect to the symmetric residue system.

reduceMethod
reduce(A::SMat{fmpz}, g::SRow{fmpz}) -> SRow{fmpz}

Given an upper triangular matrix $A$ over a field and a sparse row $g$, this function reduces $g$ modulo $A$.

reduceMethod
reduce(A::SMat{T}, g::SRow{T}) -> SRow{T}

Given an upper triangular matrix $A$ over a field and a sparse row $g$, this function reduces $g$ modulo $A$.

rand_rowMethod
rand_row(A::SMat) -> SRow

Return a random row of the sparse matrix $A$.

Changing of the ring:

map_entriesMethod
map_entries(f, A::SMat) -> SMat

Given a sparse matrix $A$ and a callable object $f$, this function will construct a new sparse matrix by applying $f$ to all elements of $A$.

change_base_ringMethod
change_base_ring(R::Ring, A::SMat)

Create a new sparse matrix by coercing all elements into the ring $R$.

Arithmetic

Matrices support the usual operations as well

  • +, -, ==
  • div, divexact by scalars
  • multiplication by scalars

Various products:

mulMethod
mul(A::SMat{T}, b::AbstractVector{T}) -> Vector{T}

Return the product $A \cdot b$ as a dense vector.

mulMethod
mul(A::SMat{T}, b::AbstractMatrix{T}) -> Matrix{T}

Return the product $A \cdot b$ as a dense array.

mulMethod
mul(A::SMat{T}, b::MatElem{T}) -> MatElem

Return the product $A \cdot b$ as a dense matrix.

mulMethod
mul(A::SRow, B::SMat) -> SRow

Return the product $A\cdot B$ as a sparse row.

Other:

sparseMethod
sparse(A::SMat) -> SparseMatrixCSC

The same matrix, but as a sparse matrix of julia type SparseMatrixCSC.

fmpz_matMethod
fmpz_mat(A::SMat{fmpz})

The same matrix $A$, but as an fmpz_mat.

fmpz_matMethod
fmpz_mat(A::SMat{T}) where {T <: Integer}

The same matrix $A$, but as an fmpz_mat. Requires a conversion from the base ring of $A$ to $\mathbb ZZ$.

ArrayMethod
Array(A::SMat{T}) -> Matrix{T}

The same matrix, but as a two-dimensional julia array.