virgil.likelihood

Model to evaluate.

Data to compare with.

Error-inflation terms, vis_scale, phi_scale, vis_error_rel and phi_error (see inflated_errors), and the widths of the data's gains, vis_gain_<group> (see OIData.with_gains), closure-phase offsets, phi_offset_<group> (see OIData.with_closure_offsets), the wavelength scale, wavel_scale and wavel_offset (see OIData.with_wavelength_scale), and the North angle, north_angle (degrees; see OIData.with_north_angle).

One dimensionless residual per independent observable (n_independent of them), in the order of flatten_data; correlated closure phases are replaced by their whitened independent combinations, followed by one periodic penalty residual per closure phase, so the length is n_residuals.

Model to evaluate.

Data to compare with.

If True, return -inf when is_physical is false, e.g. for a negative flux or a rim whose brightness goes negative. This works inside jax.jit, so samplers can use it as a hard prior boundary.

Error-inflation terms, e.g. fitted as nuisance parameters: vis_scale and phi_scale multiply the uncertainties, and vis_error_rel (relative to the model visibility) and phi_error (radians) are added in quadrature (see inflated_errors). The Gaussian normalization uses the inflated errors. For data with gains (OIData.with_gains), vis_gain_<group> sets the width of a group of gains, which are marginalized: the normalization then includes the log-determinant of the visibility covariance. phi_offset_<group> does the same for closure-phase offsets (OIData.with_closure_offsets). wavel_scale and wavel_offset evaluate the model at corrected wavelengths (see OIData.with_wavelength_scale), and north_angle on a rotated sky (see OIData.with_north_angle).

Data whose uncertainties are inflated.

Model vector, e.g. from OIData.model.

Extra visibility error, as a fraction of the model visibility observable (e.g. of the model V² for squared visibilities): a calibration error.

Extra phase error in radians.

Factors multiplying the visibility and phase uncertainties.

Factors multiplying the visibility and phase uncertainties.

Extra absolute visibility error, in the units of the observable.

What vis_error_rel is relative to: the model (default, the right choice for a fitted term, as it does not reward the model for low data points) or the data (as error floors are).

Add the terms in quadrature (default), or take the largest (a floor, as with_error_floor does; both use _utils.inflate_errors).

Uncertainties matching flatten_data.

E.g. the best fit.

Data with extra spectra (extras=("flux",) and the like).

Per kind ("flux", "nflux", "corrflux"): mean (n_group, p) and cov (n_group, p, p) of the weights, whose first is the scale k multiplying the model's template (normalized to a mean of 1 per group); and scale, the scale of total_spectrum itself (k over the template's normalization), for "flux" and "corrflux": data ≈ scale × model. groups gives each sample's group.

{site: (prior, datasets, term)}, datasets being the indices of the datasets the term applies to; prior is the function for a tied term.

Values of the model parameters.

List of parameter names.

Template model whose parameters at the dot-separated paths params are replaced by values, or a class/callable called as model(**dict(zip(params, values))) (see build_model).

Object containing the data to be fitted.

Error terms and reject_unphysical, passed to model_loglike.

Log-likelihood value.

Either a template model whose leaves at the paths in priors are sampled, or a function called with the sampled values as keyword arguments that returns a SourceModel. A function lets you sample parameters that are not leaves of the model, such as a separation and position angle, or one inclination shared by two components (see build_model). As for fit, the function may return a list of models, one per dataset in data, sharing parameters: for example binaries at two epochs with one flux ratio and a position each. Model i is compared with dataset i, and regularizers act on the first model only.

Mapping from parameter path (e.g. "comp.flux") or function argument name to prior; each key is also used as the numpyro sample-site name. Priors on fluxes (keys named flux or ending in .flux) must have non-negative support. An angle (degrees) with an AngleVector prior is sampled as a 2-D vector at the site "<path>_vec", with the angle recorded as the deterministic site "<path>": there is no wrap boundary at 0°/360°.

Data whose Gaussian log likelihood is added with numpyro.factor. May be () when likelihoods holds all the data.

Log-prior terms on the model, e.g. a Centroid prior, added with numpyro.factor. Only genuine prior densities (probabilistic) are allowed: penalties such as maximum entropy are for fit.

Priors on error-inflation terms (vis_scale, phi_scale, vis_error_rel, phi_error; see inflated_errors) and on gain widths (vis_gain_<group>; see OIData.with_gains), on closure-offset widths (phi_offset_<group>; see OIData.with_closure_offsets), on the wavelength scale (wavel_scale, wavel_offset; see OIData.with_wavelength_scale) and on the North angle (north_angle, degrees; see OIData.with_north_angle), sampled as sites "noise.<term>". A list gives each dataset its own terms, as sites "noise[i].<term>". An entry may instead be a function of the dict of sampled values (keyed like priors), recorded as a deterministic site: a term tied to parameters in priors, as in a hierarchical model where each epoch's error scale is drawn from a population with fitted hyperparameters (see hierarchical_scales). A tied term's log_prior(values), if it has one, is added once (site "noise_prior"). Priors. As for fit: the default (Jeffreys) prior of these scale parameters is log-uniform on stated bounds that must contain the plausible values.

Extra data terms, as for fit: callables of the sampled values (a dict keyed like priors) returning whitened residuals, such as PositionData.term(orbit_fn) or RVData.term(params_fn). Term i is added with numpyro.factor as site "likelihood_<i>". Those two built-in terms add their full normalized Gaussian log density, like the OIData terms; a plain callable adds -0.5 * sum(r**2) only.

Fixed error terms and reject_unphysical, passed to model_loglike.

Zero-argument numpyro model, e.g. for numpyro.infer.NUTS.

The numpyro model, e.g. from numpyro_model.

One dict of values per chain, keyed by sample site as init_to_value takes them (e.g. FitResult.values, which holds the "<path>_vec" sites of angle vectors). Sites missing from a dict start uniformly at random, as init_to_value does.

Random key for those missing sites (default PRNGKey(0)).

Unconstrained parameters by site, with a leading axis of len(starts) (none for a single start). Pass them as mcmc.run(key, init_params=...) with num_chains=len(starts).

Posterior samples, as equal-length 1D arrays keyed by parameter name or path (e.g. from mcmc.get_samples()).

Template model or class, as for :func:loglike.

Data defining the observables.

Which keys of samples to use (default: all of them).

vis_mean, vis_std, phi_mean and phi_std: the mean and standard deviation over the samples of each visibility and phase observable.