Likelihood terms
A likelihood term is any object passed in the likelihoods argument of
fit, variational and
numpyro_model, or returned by make_term(mask) in
cross_validate. There is no base class: fit and
crossval test for attributes, and the term does only what its attributes
say. The built-in terms are
PositionData.term,
RVData.term,
ComplexVisData.term and
ToeplitzVisData.
values below is the dict of constrained parameter values, keyed like
priors.
Required
| Attribute | Contract |
|---|---|
term(values) |
Whitened residuals, an array of any shape. The term's χ² is the sum of their squares, and its size is the size of the array. Not needed when the term has chi2_and_log_norm, but such a term must then define n_data. |
Optional
| Attribute | Read by | Effect |
|---|---|---|
has_log_norm, log_norm(values) |
fit, crossval |
If has_log_norm is true, log_norm(values) is added to ½χ² (the Gaussian normalization, for a fitted jitter or a prior Jacobian). A term with one has no least-squares form: use method="lbfgs" or "adam". |
residuals_and_log_norm(values) |
fit |
(residuals, log_norm) in one evaluation; used instead of term(values) plus log_norm when the normalization needs the residuals (a profiled scale). |
chi2_and_log_norm(values) |
fit |
A scalar-χ² term: (χ², log_norm) with no residual vector. Needs the scalar-χ² attributes below. It has no least-squares form either. |
chi2_constant, loss_constant |
fit |
Constants left out of chi2_and_log_norm and of the loss; added back to the reported χ² and loss. Scalar-χ² terms only. |
curvature_params(values), curvature_loss(params) |
fit (the L-BFGS/Adam preconditioner), gauss_newton_diag |
The curvature of a scalar-χ² term, through a function of a few parameters. Required by scalar-χ² terms whenever fit preconditions. gauss_newton_mass does not accept scalar-χ² terms. |
n_data |
fit |
The number of data points, for the per-point loss scale; otherwise the size of term(start). |
resolve(start, least_squares) |
fit |
Returns (term, form): the term fit should actually use at start, and a note of which form ran (a radio term may swap in its Toeplitz form). |
derived(values) |
fit |
A dict of extra quantities (not parameters) added to info["derived"]. |
arrays |
kkt_report |
A dict with "u" and "v" (in wavelengths), so the band-limit diagnostic sees the term's longest baseline. |
loglike(values) |
numpyro_model |
The term's own normalized log density, added with numpyro.factor. Without it, the factor is \(-\tfrac12\sum r^2\) of term(values) only: log_norm is ignored, so a term with a log_norm (or a scalar χ²) must define loglike to be sampled correctly. |
Cross-validation
cross_validate calls only term(values) and, if present, has_log_norm
and log_norm(values). Its make_term(mask) must return a term that uses
the points where mask is true and gives zero residuals elsewhere, without
changing array shapes. See Cross-validation.