# 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 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{ZZRingElem} 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 $R$.

source
sparse_rowMethod
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 $J = (i_j, x_j)_j$.

source
sparse_rowMethod
sparse_row(R::NCRing, 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 $V = (x_j)_j$.

source

### 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 $a_j$.

source
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 $(aA + bB, cA + dB)$.

source

### 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 $R$.

source

### Maximum, minimum and 2-norm

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

Returns the smallest entry of $A$.

source
  minimum(A::RelNumFieldOrderIdeal) -> AbsNumFieldOrderIdeal{AbsSimpleNumField, AbsSimpleNumFieldElem}
minimum(A::RelNumFieldOrderIdeal) -> RelNumFieldOrderIdeal

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

source

### Functionality for integral sparse rows

liftMethod
lift(A::SRow{zzModRingElem}) -> SRow{ZZRingElem}

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

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

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

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

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

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

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

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

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

source

### Conversion to/from julia and AbstractAlgebra types

VectorMethod
Vector(a::SMat{T}, n::Int) -> Vector{T}

The first n entries of a, as a julia vector.

source
sparse_rowMethod
sparse_row(A::MatElem)

Convert A to a sparse row. nrows(A) == 1 must hold.

source
dense_rowMethod
dense_row(r::SRow, n::Int)

Convert r[1:n] to a dense row, that is an AbstractAlgebra matrix.

source

## 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{ZZRingElem} is the type for sparse matrices over the integers.

In contrast 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, n::Int, m::Int) -> SMat

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

source

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 $A$.

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

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

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

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

source

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.

source

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 $A$.

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

Vertically joins $A$ and $B$.

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

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

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

Horizontally concatenates $A$ and $B$.

source

(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 $R$.

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

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

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

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

source
block_diagonal_matrixMethod
block_diagonal_matrix(xs::Vector{SMat})

Return the block diagonal matrix with the matrices in xs on the diagonal. Requires all blocks to have the same base ring.

source

Slices:

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

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

source

Transpose:

### 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.

source
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.

source
nnzMethod
nnz(A::SMat) -> Int

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

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

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

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

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

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

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.

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

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

source
reduce_fullMethod
reduce_full(A::SMat{ZZRingElem}, g::SRow{ZZRingElem},
with_transform = Val(false)) -> SRow{ZZRingElem}, 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 with_transform is set to Val(true), then additionally an array of transformations is returned.

source
hnf!Method
hnf!(A::SMat{ZZRingElem})

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

source
hnfMethod
hnf(A::SMat{ZZRingElem}) -> SMat{ZZRingElem}

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

source
snfMethod
snf(A::SMat{ZZRingElem})

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

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

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

source
is_diagonalMethod
is_diagonal(A::SMat) -> Bool

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

source
detMethod
det(A::SMat{ZZRingElem})

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

source
det_mcMethod
det_mc(A::SMat{ZZRingElem})

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

source
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.

source
saturateMethod
saturate(A::SMat{ZZRingElem}) -> SMat{ZZRingElem}

Computes the saturation of $A$, that is, a basis for $\mathbf{Q}\otimes M \cap \mathbf{Z}^n$, 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.

source
hnf_kannan_bachemMethod
hnf_kannan_bachem(A::SMat{ZZRingElem}) -> SMat{ZZRingElem}

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

source
diagonal_formMethod
diagonal_form(A::SMat{ZZRingElem}) -> SMat{ZZRingElem}

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

source

### Manipulation/ Access

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

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

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

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

source
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 $b$.

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

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

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

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

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

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

source
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 $A$.

source
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 $A$.

source
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 $A$.

source
diagonalMethod
diagonal(A::SMat) -> ZZRingElem[]

The diagonal elements of $A$ in an array.

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

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

source
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.

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

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.

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

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

source
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 $A$.

source

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 $A$.

source

### Arithmetic

Matrices support the usual operations as well

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

Various products:

*Method
*(A::SMat{T}, b::AbstractVector{T}) -> Vector{T}

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

source
*Method
*(A::SMat{T}, b::AbstractMatrix{T}) -> Matrix{T}

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

source
*Method
*(A::SMat{T}, b::MatElem{T}) -> MatElem

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

source
*Method
*(A::SRow, B::SMat) -> SRow

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

source
dotMethod
dot(x::SRow{T}, A::SMat{T}, y::SRow{T}) where T -> T

Return the generalized dot product dot(x, A*y).

source
dotMethod
dot(x::MatrixElem{T}, A::SMat{T}, y::MatrixElem{T}) where T -> T

Return the generalized dot product dot(x, A*y).

source
dotMethod
dot(x::AbstractVector{T}, A::SMat{T}, y::AbstractVector{T}) where T -> T

Return the generalized dot product dot(x, A*y).

source

Other:

sparseMethod
sparse(A::SMat) -> SparseMatrixCSC

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

source
ZZMatrixMethod
ZZMatrix(A::SMat{ZZRingElem})

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

source
ZZMatrixMethod
ZZMatrix(A::SMat{T}) where {T <: Integer}

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

source
MatrixMethod
Matrix(A::SMat{T}) -> Matrix{T}

The same matrix, but as a julia matrix.

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

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

source