Core Concepts¶

This page explains the key ideas behind unite’s design. Understanding these concepts will make the tutorials and guides much easier to follow. Please reference the relevant usage sections for detailed examples, explanation, and API documentation.

Key Assumptions¶

unite makes several scientific and practical assumptions. Understanding these upfront will help you assess whether the package is appropriate for your data and avoid common pitfalls.

Same-Source Assumption¶

The model is evaluated independently for each disperser/spectrum, but all spectra are assumed to observe the same physical source with the same intrinsic properties (redshift, line widths, line fluxes). This is what enables simultaneous fitting across multiple gratings.

This assumption may not hold if:

The source varies between observations — e.g., time-variable AGN or transients observed at different epochs.
Different slit positions or fiber placements sample different spatial regions of an extended source, leading to different line ratios or kinematics.
Different aperture sizes capture different fractions of the source flux, biasing relative line strengths.

For JWST/NIRSpec MSA observations, objects observed with the same mask configuration are typically observed simultaneously across all gratings, so the same-source assumption is usually valid. Calibration tokens (FluxScale, RScale) can absorb some inter-disperser differences in flux calibration or resolution, but they cannot account for fundamentally different source properties.

Gaussian Line Spread Function¶

The spectral LSF is modelled as a Gaussian at every pixel, with FWHM determined by the disperser’s resolution curve \(R(\lambda)\):

\[\mathrm{lsf\_fwhm}(\lambda) = \frac{\lambda}{R(\lambda)}\]

For profiles with a Gaussian component (Gaussian, PseudoVoigt, SEMG, GaussHermite, SplitNormal), the intrinsic and LSF widths are added in quadrature:

\[\mathrm{fwhm\_total} = \sqrt{\mathrm{fwhm\_intrinsic}^2 + \mathrm{lsf\_fwhm}^2}\]

For purely Lorentzian components (Cauchy, and the Lorentzian part of PseudoVoigt), the LSF is not convolved into the Lorentzian width. This means a “Cauchy” profile in unite is effectively a Voigt-like profile (Lorentzian convolved with the Gaussian LSF), which is physically appropriate since the instrumental broadening is always present.

Pixel Integration¶

unite provides two integration modes, selectable via integration_mode on build():

Analytic mode (default) integrates each line profile over pixel bins using its CDF:

Exact for emission lines — no discretisation error from summing sub-pixel samples
Fast — one CDF evaluation per pixel edge, independent of line width
Robust for undersampled data — critical for NIRSpec PRISM where lines can be narrower than a single pixel
Approximate for absorption lines — each profile is integrated independently before the nonlinear transmission exp(-τ·φ) is applied. This computes exp(-τ·∫φ) rather than ∫F·exp(-τ·φ), which is accurate when the absorber is well-resolved but introduces an approximation for unresolved or marginally resolved lines.

Quadrature mode evaluates the full composed model — emission, absorption, and continuum together — at Gauss-Legendre sub-pixel nodes and integrates via weighted sum:

Exact pixel integration for both emission and absorption — properly computes ∫F(λ)·exp(-τ·φ(λ)) dλ over each pixel, avoiding the analytic-mode approximation
Slower — requires n_nodes (default 7) profile evaluations per pixel instead of one
Recommended when tau-parametrized lines are unresolved or marginally resolved, or when mixing emission and absorption at similar wavelengths

# Analytic (default) — fast, exact for emission lines
model_fn, args = builder.build(integration_mode='analytic')

# Quadrature — correct pixel integration for absorption lines
model_fn, args = builder.build(integration_mode='quadrature', n_nodes=7)

The continuum is evaluated at pixel centers in both modes, since it varies slowly enough that sub-pixel variation is negligible.

LSF Pre-Convolution of Absorption Profiles¶

Warning

This approximation applies in both integration modes.

In both analytic and quadrature modes, the absorption profile φ(λ) used in exp(-τ·φ) is the LSF-convolved profile, not the intrinsic one. The physically correct observable is:

\[\mathrm{obs}(\lambda) = \mathrm{LSF} \otimes \bigl[F(\lambda) \cdot e^{-\tau\,\phi_\mathrm{intrinsic}(\lambda)}\bigr]\]

Computing this requires convolving the nonlinear product over the full multi-pixel LSF kernel, which is not currently supported. Instead, the code evaluates:

\[\mathrm{obs}(\lambda) \approx F(\lambda) \cdot e^{-\tau\,\phi_\mathrm{LSF}(\lambda)}\]

When the approximation is accurate:

Resolved absorbers (intrinsic FWHM ≫ LSF FWHM): φ_LSF ≈ φ_intrinsic, so the two expressions agree regardless of optical depth.
Optically thin lines (τ ≪ 1): the integrand is linear in φ, convolution distributes, and both approaches give identical first-order results. Errors are O(τ²).

When it matters:

For unresolved, optically thick absorbers — e.g. narrow ISM absorbers or stellar Balmer absorption observed at moderate spectral resolution — the LSF-broadened profile has a lower peak than the intrinsic profile. The code therefore underestimates the absorption depth for a given τ, so inferred τ values will be biased high and the curve of growth will be misrepresented.

Quadrature mode fixes only the pixel-integration approximation (analytic mode’s exp(-τ·∫φ) vs. ∫exp(-τ·φ)); it does not address this LSF pre-convolution issue.

Priors¶

Every token carries a prior distribution that is sampled in the NumPyro model:

from unite import prior

prior.Uniform(low, high)                       # flat prior
prior.TruncatedNormal(loc, scale, low, high)   # Gaussian with hard bounds
prior.Fixed(value)                             # not sampled; held constant

Priors can be dependent on other tokens — for example, constraining a broad FWHM to always exceed a narrow FWHM by at least 150 km/s:

fwhm_narrow = line.FWHM('narrow', prior=prior.Uniform(100, 1000))
fwhm_broad  = line.FWHM('broad',  prior=prior.Uniform(fwhm_narrow + 150, 5000))

See Priors for the full reference on supported priors, dependent priors, and topological sorting.

Line Configuration¶

LineConfiguration is the container for emission line specifications. Lines are added with add_line, specifying the rest-frame center wavelength, kinematic tokens, flux, and (optionally) a profile shape.

lc = line.LineConfiguration()
lc.add_line(
    name='H_alpha',
    center=6563.0 * u.AA,
    redshift=z,
    fwhm_gauss=fwhm,
    flux=line.Flux('Ha_flux', prior=prior.Uniform(0, 10)),
    profile='Gaussian',   # default
)

Multiple calls with the same name create multiple components (e.g., narrow + broad) whose fluxes are summed. See Line Configuration for the full guide including all seven profile shapes, parameter sharing patterns, and merging configurations.

Continuum Configuration¶

ContinuumConfiguration defines wavelength regions with a functional continuum form attached to each. The easiest way to build one is automatically from the line centers:

from unite.continuum import ContinuumConfiguration, Linear

cc = ContinuumConfiguration.from_lines(
    lc.centers,    # wavelength array of line rest-frame centers
    form=Linear(), # continuum form (or string name, e.g. 'Linear')
)

Nine built-in forms are available: Linear, PowerLaw, Polynomial, Chebyshev, BSpline, Bernstein, Blackbody, ModifiedBlackbody, and AttenuatedBlackbody.

Each form’s model parameters (e.g. scale, slope, temperature) receive default priors that can be overridden per-region via ContinuumRegion(params={...}). Sharing the same Parameter instance across regions ties them to a single model parameter — the same token pattern used for emission lines.

See Continuum Configuration for all available functional forms, custom priors, parameter sharing, and the quick-reference table.

Dispersers & Spectra¶

A disperser represents an instrument’s wavelength dispersion and resolution properties. It encodes two calibrations: the resolving power \(R(\lambda)\) (used to compute the instrumental line spread function) and the wavelength dispersion per pixel \(d\lambda/\mathrm{dpix}(\lambda)\). Built-in support is provided for JWST/NIRSpec and SDSS; custom instruments can be configured using generic disperser classes. Optional calibration tokens (RScale, FluxScale, PixOffset) can absorb uncertainties in resolution, flux calibration, and wavelength solution.

Spectra is a container for one or more spectra:

from unite.instrument import nirspec
from unite.spectrum import Spectra

# Configure dispersers
g235h = nirspec.G235H()
g395m = nirspec.G395M()

# Load spectra (NIRSpec example)
spectrum1 = nirspec.NIRSpecSpectrum.from_DJA('g235h.fits', disperser=g235h)
spectrum2 = nirspec.NIRSpecSpectrum.from_DJA('g395m.fits', disperser=g395m)

# Wrap in Spectra container
spectra = Spectra([spectrum1, spectrum2], redshift=5.28)

The redshift argument shifts all line centers to the observed frame before coverage filtering. See Instruments & Spectrum Loading for more on data handling, error scaling, and multi-spectrum fits.

The Prepare → Scale → Build Pipeline¶

Before running the sampler, three preparation steps are needed:

1. `Spectra.prepare()`¶

spectra = Spectra([spectrum], redshift=0.0)
filtered_lines, filtered_cont = spectra.prepare(lc, cc)

This filters out lines and continuum regions not covered by the spectrum’s wavelength range. The result is two new configuration objects containing only the relevant features.

2. `Spectra.compute_scales()`¶

spectra.compute_scales(filtered_lines, filtered_cont, error_scale=True)

This performs two tasks:

Flux normalisation. Estimates characteristic flux scales (line_scale, continuum_scale) so that sampler parameters are near unity. This is important for efficient MCMC sampling — parameters spanning many orders of magnitude lead to poor posterior geometry and slow convergence.

Error scaling (when error_scale=True). Spectral reduction pipelines — including NIRSpec’s — often produce error arrays that over/underestimate the true noise. unite can optionally rescale errorbars to mitigate this effect, though it does not currently take into account correlated noise.

3. `ModelBuilder.build()`¶

builder = model.ModelBuilder(filtered_lines, filtered_cont, spectra)
model_fn, model_args = builder.build()

Returns a NumPyro model function and a ModelArgs dataclass containing all pre-computed matrices, scales, and data arrays.

User Controls the Sampler¶

unite does not bundle a fitting loop. The (model_fn, model_args) tuple can be used with:

# NUTS (recommended for well-behaved posteriors)
from numpyro import infer
kernel = infer.NUTS(model_fn)
mcmc = infer.MCMC(kernel, num_warmup=500, num_samples=1000)
mcmc.run(jax.random.PRNGKey(0), model_args)

# SVI (variational inference — faster but approximate)
guide = infer.autoguide.AutoNormal(model_fn)
svi = infer.SVI(model_fn, guide, infer.optim.Adam(0.01), infer.Trace_ELBO())

Configuration Serialization¶

Every configuration object can be saved to and loaded from YAML:

from unite.config import Configuration

config = Configuration(lines=lc, continuum=cc, dispersers=dc)
config.save('my_fit.yaml')

# Load back — all tokens and sharing relationships are preserved
config2 = Configuration.load('my_fit.yaml')

See Configuration Serialization for the full workflow, YAML format, and sub-configuration serialization.