Operators

See Architecture → Operators for the model.

Bases: ABC

Abstract linear operator A with an adjoint Aᴴ.

Subclasses implement :meth:_forward and :meth:_adjoint. Declare in_shape / out_shape (tuples, with None for free axes) to get automatic shape checking; leave them None to skip it.

H property

The adjoint operator Aᴴ. A.H.H is A.

__matmul__(other)

(A @ B)(x) == A(B(x)).

adjoint(y)

Apply Aᴴ to y.

dot_test(in_shape=None, out_shape=None, *, n_trials=3, rtol=1e-05, seed=0)

Verify the adjoint via ⟨A x, y⟩ ≈ ⟨x, Aᴴ y⟩ on random complex data.

NumPy-only diagnostic. Provide concrete in_shape / out_shape if the operator declares free (None) axes.

forward(x)

Apply A to x.

normal(x)

Apply AᴴA to x (override :meth:_normal for a fast path).

Bases: LinearOperator

I. Useful as a neutral element and in tests.

Bases: LinearOperator

c * A for a (possibly complex) scalar c.

Bases: LinearOperator

A chain A₁ A₂ … Aₙ applied right-to-left; adjoint reverses the chain.

Bases: LinearOperator

Wraps an operator to swap forward and adjoint.

Bases: LinearOperator

Wrap base, skipping the public input/output shape validation.

Some operators (e.g. :class:~unimri.operators.fourier.FourierOperator and :class:~unimri.operators.nufft.NUFFTOperator) declare the shape for a single image but their _forward/_adjoint also accept a batched input with a leading axis (e.g. coils) -- that is how FourierOperator(...) @ SensitivityOperator(maps) becomes a full SENSE encoding operator. The public :meth:forward/:meth:adjoint/:meth:normal would reject the batched shape; this wrapper calls the unchecked internals directly, for use once you know the batching is intentional (e.g. inside a solver such as :func:unimri.optimization.conjugate_gradient).

Bases: LinearOperator

Centered, unnormalized DFT over image_shape.

Parameters:

Name	Type	Description	Default
image_shape	tuple[int, ...]	The (single-instance) image shape, e.g. (ny, nx) or (nz, ny, nx).	required
axes	tuple[int, ...] | None	Which axes to transform, as negative indices. Defaults to the last len(image_shape) axes, so a leading batch axis (coils, ...) is left untouched.	None

Bases: LinearOperator

FINUFFT-backed non-uniform FFT for a fixed trajectory and image shape.

Parameters:

Name	Type	Description	Default
trajectory	Trajectory	The k-space sample locations. 2-D or 3-D.	required
image_shape	tuple[int, ...]	Target image grid, in NumPy axis order -- (ny, nx) for 2-D, (nz, ny, nx) for 3-D. Trajectory row d pairs with image axis d (so (kz, ky, kx) for a 3-D image).	required
eps	float	FINUFFT relative tolerance.	_EPS

Bases: LinearOperator

Multiply/combine by per-coil sensitivity maps.

Parameters:

Name	Type	Description	Default
sensitivity	Array	Complex array of shape (n_coils, *image_shape).	required