torchsky.map.operators.continuous_response¶

Adapt an offset-only unit-square PixelIntegratedResponse.

Both position tensors must be in detector-pixel coordinates. Wavelength does not affect this reference response; no further pixel integral is applied. Arbitrary footprints require a different callable.

response: PixelIntegratedResponse

__call__(
    detector_xy: Tensor,
    source_xy: Tensor,
    wavelength_angstrom: Tensor,
    *,
    detector_footprint: Tensor | None,
) -> Tensor

Transport sampled source measures into native detector intensities.

Forward accepts arbitrary prefix axes followed by [N,L] and returns the same prefix followed by [P]. Its discrete matrix is A[p,n,l] = R[p,n,l] * source_measure[n] * wavelength_measure[l] * source_spectral_weight[n,l]. The optional last factor defaults to one and accepts [L] or [N,L]. It weights each source spectrum before transport; it never mixes rendered output bands. Separate operators represent bands.

source_measure is explicit: use ones for point-source spectral flux density, or source-node solid-angle quadrature weights (sr) for spectral radiance per sr. These are source areas, never detector-pixel areas. Use wavelength_measure in Angstrom for densities per Angstrom, or ones for already bin-integrated spectral flux. The response already includes the detector footprint integral exactly once. Output units are input units times these two measures and the response's calibration units. Electron conversion (area, duration, QE) must be declared by the caller; no implicit conversion from energy density or photon density is made.

Source and detector positions, measures and response are fixed when taking the Euclidean adjoint. They and response parameters retain their gradients when differentiating forward/adjoint. mask=True keeps detector rows; removed rows are zero without renormalizing retained flux or PSF tails.

interaction_pairs[E,2] optionally gives detector/source index pairs. Response evaluation costs O(E*L), with edge and wavelength chunk limits. Every forward/adjoint call also validates tensor values and graph bounds and checks graph uniqueness with torch.unique; this validation is not restricted to the response chunk limits or a one-time construction cost. The required interaction_support string declares why omitted edges are zero in the represented response (e.g. exact compact support, or a calibrated core whose wings are transported separately). The declaration is provenance, not a support proof. Never truncate an infinite-tail response implicitly. Duplicate pairs are rejected. The graph can represent radiance mesh nodes and does not require a source catalogue.

Without a graph all pairs are evaluated: time is O(PNL), a bounded reference/fallback rather than a survey-scale algorithm. Three chunk limits bound each evaluated response block, avoiding a dense global matrix. Autograd may retain block graphs; this is not a bounded training memory claim. Quadrature refinement and support verification remain the caller's responsibility. This operator does not certify wavefront fidelity.

detector_chunk_size: int = 128

detector_footprint: Tensor | None = None

detector_xy: Tensor

interaction_pairs: Tensor | None = None

interaction_support: str | None = None

mask: Tensor | None = None

pair_chunk_size: int = 1024

response: NativeSpectralPixelResponse

source_chunk_size: int = 128

source_measure: Tensor

source_spectral_weight: Tensor | None = None

source_xy: Tensor

wavelength_angstrom: Tensor

wavelength_chunk_size: int = 16

wavelength_measure: Tensor

__post_init__() -> None

_block(
    pixels: slice, sources: slice, waves: slice
) -> Tensor

_pair_block(pairs: Tensor, waves: slice) -> Tensor

_validate() -> None

adjoint(detector_values: Tensor) -> Tensor

forward(intensity: Tensor) -> Tensor