Accept parameter arrays that broadcast against the event axis: parameters of shape $(p, 1)$ × events $(n,)$ produce intensities $(p, n)$, and estimator reductions over axis=-1 return one estimator value per parameter set, all on-device. This is useful e.g. for propagating fit uncertainties over bootstrap parameter samples, as done in ComPWA/polarimetry.
The broadcasting convention is backend-agnostic (NumPy, JAX, TensorFlow, Numba). For scalar parameters, behavior is unchanged (sum(x, axis=-1) == sum(x) for 1-D intensities).
Accept parameter arrays that broadcast against the event axis: parameters of shape$(p, 1)$ × events $(n,)$ produce intensities $(p, n)$ , and estimator reductions over
axis=-1return one estimator value per parameter set, all on-device. This is useful e.g. for propagating fit uncertainties over bootstrap parameter samples, as done in ComPWA/polarimetry.The broadcasting convention is backend-agnostic (NumPy, JAX, TensorFlow, Numba). For scalar parameters, behavior is unchanged (
sum(x, axis=-1) == sum(x)for 1-D intensities).