SciPy keywords

Back to the backend overview

About this backend

The scipy backend builds operators as SciPy CSR matrices. build_op and get_ops return labeled factories that construct these matrices when called. Pass the factories to public ed, xas, and rixs; the primary Hamiltonian’s .backend label selects SciPy. The backend calls each supplied factory at calculation start, including repeated arguments. See Operator factories and solver dispatch.

The backend-specific ed_scipy, xas_scipy, and rixs_scipy functions also require factories. These can return NumPy arrays, SciPy sparse matrices, or a matrix-free LinearOperator. ED uses SciPy’s LOBPCG eigensolver; XAS and RIXS use EDRIXS’s Lanczos routines, and the RIXS correction-vector step uses SciPy’s GMRES solver.

XAS and RIXS use the lowest retained initial energy and the first incident energy as internal references. Assuming the incident grid is near resonance, this centers the intermediate Hamiltonian; RIXS also centers the final Hamiltonian on the initial energy. Dense and sparse diagonals are shifted before matrix-vector multiplication, without changing the input operators. The public energy grids and returned RIXS pole energies retain their original reference. Spectra are evaluated before restoring the pole energies. For a custom LinearOperator, shifting is necessarily applied after its matrix-vector product and cannot recover precision lost inside that product.

Options are specific to the stage where they are supplied.

build_op and get_ops

Option	Default	Description
tol	1e-10	Discard operator coefficients whose magnitude does not exceed this threshold.

Here tol controls operator truncation. It is separate from the eigensolver tolerance accepted by ed.

ed

Option	Default	Description
blocksize	None	LOBPCG block size; defaults to the requested number of eigenpairs.
tol	1e-10	LOBPCG convergence tolerance.
maxiter	1000	Maximum LOBPCG iterations.
seed	None	Seed used to generate the initial LOBPCG block.
initial_guess	None	Initial LOBPCG block with shape (dimension, blocksize).
suppress_lobpcg_warnings	True	Suppress SciPy nonconvergence warnings; results are still returned.

xas

Option	Default	Description
nkryl	200	Maximum intermediate-state Lanczos dimension.

rixs

Option	Default	Description
nkryl	200	Maximum final-state Lanczos dimension.
linsys_tol	1e-09	GMRES relative convergence tolerance (rtol; tol in older SciPy).
linsys_maxiter	50000	Maximum GMRES restart cycles.
linsys_restart	200	GMRES restart interval.

linsys_tol sets GMRES’s relative tolerance, passed as rtol (or tol in older SciPy). With the rtol API, EDRIXS sets the absolute tolerance atol to zero.