Operator Examples and Concepts

Linear operators supply the operation

\[ y \leftarrow A x . \]

They are used by CG, GMRES, Lanczos, and exponential-action routines.

Storage Concepts

Storage concepts are defined in include/core/concepts.hpp:

• num::VectorLike
• num::MutableVectorLike
• num::ContiguousVectorLike
• num::DenseMatrixLike
• num::MutableDenseMatrixLike
• num::ContiguousDenseMatrixLike

Operator concepts are defined in include/operator/concepts.hpp:

template<class Op, class X = Vector, class Y = Vector>
concept LinearOperator =
    VectorLike<X> && MutableVectorLike<Y> &&
    requires(const Op& A, const X& x, Y& y) {
        { A.rows() } -> std::convertible_to<idx>;
        { A.cols() } -> std::convertible_to<idx>;
        { A.apply(x, y) };
    };

Dense Operator

num::Matrix A(3, 3, 0.0);
fill_spd_matrix(A);
 
num::operators::DenseOp Aop(A);
static_assert(num::operators::LinearOperator<decltype(Aop)>);
 
num::Vector x(3, 0.0);
num::SolverResult info = num::cg(Aop, b, x);

Sparse Operator

num::SparseMatrix A = num::SparseMatrix::from_triplets(
    n, n, rows, cols, values);
 
num::operators::SparseOp Aop(A);
num::SolverResult info = num::gmres(Aop, b, x, 1e-10, 200);

Callable Operator

auto Aop = num::operators::make_op(
    [N](const num::Vector& x, num::Vector& y) {
        apply_stencil(x, y, N);
    },
    N * N);
 
num::SolverResult info = num::cg(Aop, b, x, 1e-8, 1000);

Lanczos

auto eig = num::lanczos(Aop, 20, 1e-10, 100);

Exponential Action

num::Vector y;
num::expv(Aop, v, t, y);

Use operators for algorithms that only require products with \(A\). Use direct factorizations when the assembled matrix and repeated solves justify it.