Glossary

This page defines the terms the documentation, the code and the products use, in the sense Hebog gives them. Where a term has a different meaning in PyBDSF, Rapthor or LSMTool, the last section maps their name onto Hebog's. Each table links to the page where the term matters most.

Images and headers

Term	Meaning
FITS	The file format of the input image and the output catalogue. An image file holds a two-dimensional array of pixel values and a header of keyword cards that describe them. See the input header contract.
Header card	One keyword, value and comment in a FITS header, such as BUNIT = 'Jy/beam'. Hebog reads the cards listed in the input header contract and refuses a card that does not hold the kind of value it should.
WCS	World Coordinate System: the header's mapping from pixel position to sky position. Hebog accepts an ICRS or FK5 J2000 celestial WCS and publishes every position in ICRS.
ICRS, FK5 J2000	The two sky coordinate frames Hebog accepts. ICRS is the current standard frame; FK5 J2000 is the older frame many imagers still write, which differs from ICRS by a few tens of milliarcseconds.
Right ascension, declination	The sky's longitude and latitude, in degrees. Hebog publishes right ascension in [0, 360) and declination in [-90, 90].
Pixel	One array sample. Arrays are indexed (y, x), row first; a pixel position written (x, y) is a coordinate for a WCS or a catalogue.
Plane	One two-dimensional image. Hebog analyses one plane; a FITS file with frequency or Stokes axes is accepted only when each has length one.
Stokes I	Total intensity, the only polarisation Hebog measures. A STOKES axis must select it.
Reference frequency	The frequency, in hertz, at which every published flux is defined. Hebog reads it from the image's frequency axis, RESTFRQ or RESTFREQ, or from the request when the header has none.
MFS image	A multi-frequency-synthesis image: one plane made from a whole band, reported at one reference frequency. Hebog's catalogue is an MFS catalogue with no spectral fit.
Jy/beam	Jansky per beam, the brightness unit of the input and of peak fluxes. A jansky is a unit of flux density; dividing by the beam makes it a brightness. Integrated fluxes are in Jy.
Restoring beam	The Gaussian blur every deconvolved radio image carries, written as BMAJ, BMIN and BPA in degrees. It sets the finest detail the image shows and the area that converts Jy/beam to Jy. Hebog applies one beam to the whole image and refuses one wider than 10 pixels.
FWHM	Full width at half maximum, the width used for the beam and for every fitted or deconvolved ellipse. The Gaussian sigma is the FWHM divided by about 2.355.
Beam area	The solid angle of the restoring beam, π × BMAJ × BMIN / (4 ln 2). In pixels it is about 1.13 FWHM² for a circular beam, which is why a minimum island of seven pixels is a fraction of a beam five pixels wide.
Position angle	The orientation of an ellipse's major axis, in degrees east of north, reported in [0, 180).
Primary beam	The telescope's sensitivity pattern across the field, highest at the pointing centre and falling towards the edge. It is unrelated to the restoring beam.
Primary-beam-corrected image	The image after division by the primary beam, so that fluxes are intrinsic but the noise rises towards the edge of the field. Rapthor calls it the true-sky image; it is an estimate, not truth.
Flat-noise image	The image before primary-beam correction. Its noise is roughly uniform across the field, so Rapthor uses it to estimate noise, and its fluxes are attenuated by the primary beam. Rapthor also calls it the apparent-sky image.
Sector	The part of the sky one Rapthor imaging run covers. Rapthor images each sector into a true-sky and a flat-noise image and runs the source finder once per sector.
Invalid pixel	A pixel that takes no part in any estimate or measurement: NaN or infinite, a stored integer equal to BLANK, or any pixel of a 3×3 square holding one repeated value, such as zero padding. The published RMS is NaN there. See invalid pixels.
Supplied metadata	A unit, beam or reference frequency given on the request with SuppliedImageMetadata because the header omits it. A supplied value never overrides a header value, and the diagnostics record it.
Cut-out	A rectangular region cut from a larger image, for example with Astropy's Cutout2D. Images above the size limit need one.

Noise and background

Term	Meaning
Background	The slowly varying level of the image away from sources, estimated separately from the noise and subtracted before detection. The background setting is estimated (Hebog's spatial estimate) or zero (nothing subtracted, as PyBDSF's mean_map="zero").
RMS	Root mean square: the local noise level, in Jy/beam, estimated from sigma-clipped pixels in windows and interpolated to every pixel. Hebog publishes it as rms.fits.
Sigma clipping	Repeatedly discarding the samples in a window that lie more than a set number of standard deviations from the rest, so that sources do not inflate the noise estimate. Hebog clips at three standard deviations.
σ, signal-to-noise ratio	Thresholds are multiples of the local RMS above the local background: a 5σ pixel is five times the local RMS above the background.
Residual	The background-subtracted image, image − background, on which Hebog detects, grows islands and measures. The residual of a fit is what is left after subtracting fitted Gaussians as well; the text says which is meant.
Normalized residual	The residual divided by the local RMS, so that each pixel reads in σ. Internal only; thresholds apply to it.
Coarse grid	The background and RMS grid of 150-pixel windows every 50 pixels, the same as PyBDSF's rms_box. Its values are interpolated to every pixel.
Fine grid	The RMS grid of 35-pixel windows every 7 pixels, the same as PyBDSF's rms_box_bright. The continuum profile measures it across the whole image on images at least 150 pixels on a side (local noise), and both profiles measure it around bright sources.
Protected pixel, protected window	A pixel that belongs to a candidate source is protected: it is excluded from the noise statistics, and a window that overlaps protected pixels is dropped.
Clean window	A noise window that overlaps no protected pixel and measured a noise above the noise floor. A dropped window takes the RMS of its nearest clean window, within the reach the RMS reference states.
Bright region	The area around sources at 75σ or more, where imaging artefacts make the noise change quickly. There the fine grid replaces the coarse RMS wherever it reads higher. Bright candidates close together chain into one region, estimated in bounded blocks.
Noise floor	The smallest spread a window can measure: the window's brightest absolute value times 2⁻²³, single precision's resolution there. A window under it measures no noise and is dropped.
Unavailable RMS	The state of a run in which no pixel has a finite positive RMS. The RMS product is all NaN and marked unavailable, the catalogue is empty and the mask is zero. It is not evidence of an empty sky.
Flat-noise RMS	The RMS of a sector's flat-noise image, published as flat-noise-rms.fits when the request names that image. Nothing else is measured on it.

Detection

Term	Meaning
Detection threshold	detection_threshold_sigma: the signal-to-noise ratio an island must contain to be accepted. PyBDSF's thresh_pix.
Island threshold	island_threshold_sigma: the lower level to which an accepted island grows. PyBDSF's thresh_isl. It must be below the detection threshold.
Seed	A valid pixel at or above the detection threshold. An island needs a seed, or multiscale evidence over a minimum area in beams.
Eight-connected	Pixels joined through edges or corners. Islands and mask regions are eight-connected.
Island	A connected region of the published mask, and one row of the ISLANDS table. It can hold several sources, and a source can span several islands.
Minimum island size	minimum_island_pixels: the fewest pixels an island may have. PyBDSF derives its minpix_isl from the beam; Hebog requires it explicitly, and the Rapthor adapter applies PyBDSF's rule.
Matched filter	A beam-shaped smoothing filter whose response is highest where the image holds beam-sized emission. Hebog evaluates a bank of them to decide whether faint emission is significant.
À trous wavelet, B3 spline	A multiscale transform that separates an image into planes of increasing scale by repeated smoothing with a B3-spline kernel, without downsampling ("with holes"). Hebog uses scales 1 to 3 of the residual's transform as detection evidence; PyBDSF fits Gaussians to such planes.
Scale	One wavelet plane. Scale 1 holds structure about the size of the beam, and each later scale doubles that, so scale 3 holds structure about four beams wide.
Multiscale support	Pixels a detection gains from the filtered images rather than from its own thresholded pixels. It is kept only where it persists at an adjacent scale, and is assigned to the nearest detection along the emission within half a beam.
Persistence	A multiscale feature's appearance at two adjacent scales. Support that does not persist is dropped.
Support, footprint	The pixels a detection owns: its thresholded pixels plus any multiscale support assigned to it. Publication support is the part published in the mask.
Owner	The detection that owns a pixel. Every mask pixel has exactly one owner, decided by the detection, not by the tile that computed it.
Parent	One retained detection, with its support, as deblending receives it. Deblending splits a parent into components; one too large for the bounded deblend stays one component and is recorded as deferred.
Mask	source-mask.fits: 1 where a pixel is publication support, 0 elsewhere. Its connected regions are the catalogue's islands. It is not a CLEAN mask and not a photometric aperture.

Components and fitting

Term	Meaning
Deblending	Splitting a parent into one region per significant peak, so that blended sources are fitted separately. Two peaks stay apart when the fainter lies at least 1.5σ above the pass between them. See compact deblending.
Pass, saddle	The highest level at which one connected part of an island still holds two peaks; the lowest point on the ridge between them.
Deblended region	The pixels that rise to one peak after deblending. It initialises one Gaussian fit; it is not yet a component or a source.
Gaussian component	One elliptical Gaussian whose fit was admitted, and one row of the GAUSSIAN_COMPONENTS table. Say "Gaussian component", because component alone can mean a clean or sky-model component.
Moments	The brightness-weighted centroid and covariance of a region's pixels. They initialise the fit and are never published as a fitted shape.
Fit parent	A group of components whose fitting regions touch, fitted jointly so that a neighbour's light is modelled rather than absorbed. A chain of more than sixteen components, or wider than the read budget, is fitted island by island instead.
Joint fit	One least-squares solve for every component of a fit parent at once, flagged joint-gaussian-fit.
Free model, beam-constrained model	The two Gaussian models a fit chooses between: a free ellipse, or one whose axes and angle are fixed to the restoring beam. A component takes the beam when its free fit is not significantly extended, reaches a bound or is ill conditioned.
Admitted fit	A fit that converged, stayed within its bounds, is well conditioned, leaves acceptable residuals and has usable uncertainties. Only admitted fits become Gaussian rows; a rejected fit keeps the detection and records why.
Fit at bound	A fit whose solution reached a configured limit on position, size or amplitude, flagged fit-at-bound; the component then takes the beam.
Well conditioned	A fit whose parameters the data can tell apart. A component that owns a few pixels along a line cannot fix its own width, so the fit is ill conditioned until that component takes the beam.
Residual feature	Positive emission left after subtracting a fit's Gaussians, grown from a detection-threshold seed and touching the fitted support. It makes a fit fit-model-inadequate, or joins components into one source.
Beam deconvolution	Subtracting the restoring beam's covariance from a fitted ellipse's to estimate the source's intrinsic size. See compact astrometry.
Resolved, unresolved, major-axis-only	The three deconvolution states: two significant intrinsic axes, none, or only the major axis. Only an unresolved source publishes a DC_Maj of zero in the Rapthor view.
Extension significance	Whether a fit is measurably larger than the beam: the log ratio of integrated to peak flux must exceed five times its uncertainty. A geometrically resolved fit that fails the test is unresolved, flagged extension-not-significant.
Correlated noise	Noise whose neighbouring pixels are not independent, because the beam smooths them. Hebog's published errors allow for it; its fit weights pixels by their RMS without modelling the correlation.
Deferred	The status of work Hebog did not do because an object exceeded a bound: a parent too large to deblend or fit stays one detection with deferred dispositions and no Gaussian. It is never silently dropped.

Sources and the catalogue

Term	Meaning
Source	One or more components that Hebog associates as a single object in the image, and one row of the SOURCES table. Not necessarily one galaxy.
Association	The continuum profile's decision about which components form one source, from overlapping Gaussians, resolved arcs and loops, and shared residual emission. Each merge records its evidence in the diagnostics. The compact profile skips it.
Association evidence	The record of one merge: its reason (directional-fwhm-overlap, resolved-loop, resolved-open-arc or persistent-residual), the scales involved and the members.
Segment	The region a source is measured in: its support, expanded by a bounded aperture of 1.5 beams. Catalogue rows are measured one segment at a time.
Association aperture	A source's own non-overlapping region of the image. Each observable pixel contributes to at most one source's aperture.
ASSOCIATION_APERTURE_FLUX	The signed sum of background-subtracted pixels in a source's aperture, divided by the beam area, in Jy. Empty when the sum is not positive. Observed flux inside the footprint; no fit contributes to it.
INTEGRATED_FLUX	For a Gaussian component, the integral of its fitted Gaussian over the whole plane. For a continuum source, the sum of its components' integrals, PyBDSF's definition, or the aperture flux when no fit was admitted. See where the two source fluxes part.
PEAK_FLUX	A fitted Gaussian's peak amplitude in Jy/beam. A source of one fitted component publishes its component's; any other source publishes its brightest background-subtracted pixel.
Position	A single-component source publishes its Gaussian's fitted position and errors. Any other source publishes a centroid of its detection footprint, which can lie between two peaks or inside a ring, with no position errors.
Catalogue row	One row of the ISLANDS, SOURCES or GAUSSIAN_COMPONENTS table. A detection without a row still has a disposition in the diagnostics.
Quality flag	One of the comma-separated labels in QUALITY_FLAGS, such as resolved or aperture-flux-without-fitted-component. Readers must keep flags they do not recognise. See interpreting quality flags.
Disposition	The diagnostics' record of what happened to one component or source: measured, unavailable or deferred, the estimator, the reason and whether a row was published. Every detected identity has one.
Estimator	How a measured quantity was obtained: original-pixel-gaussian-model for a component, summed-fitted-component-flux or source-owned-signed-aperture for a source.
Profile	The reviewed set of algorithm settings a run uses: continuum (default, with association and local noise) or compact (one source per Gaussian, no association). Everything but the thresholds and island sizes is fixed by the profile.
Reviewed science profile	The packaged settings behind the profiles, reviewed_continuum_profile.json, whose SHA-256 the provenance records as scientific_profile_sha256.
Configuration qualification	The diagnostics label development-unqualified (the 5σ, 3σ, seven-pixel example settings) or custom-unqualified (anything else). Both mean Hebog is not qualified for survey use.
Product	One output file with its record: role, path, byte count, SHA-256, media type, schema version and scientific status. The readers verify the record before parsing.
Bundle	The output directory of four products, five with a flat-noise RMS, published together in one rename.
Scientific status	valid, or unavailable for an RMS product with no estimate. It says whether the file holds the quantity its role names.
Provenance	The diagnostics' identities of the run: SHA-256 of the input bytes, the configuration, the reviewed profile and the implementation, plus any supplied metadata.
Scientific composition	The implementation identity: a hash over every Hebog module the finder imports and every packaged resource. Two runs with the same input, configuration and composition give the same products.
Run ID	The caller's name for one run, carried into the catalogue and diagnostics. The same request with the same run ID returns its finished products instead of running again.
Schema version	The integer version of a record or file format. Readers reject versions they do not support; nothing is migrated before 1.0.

Execution

Term	Meaning
Executor	The object that runs Hebog's tasks: SerialExecutor in the calling thread, ThreadExecutor on a pool of threads in one process, or DaskExecutor on a Dask client the caller owns. All three give the same products. See the executor contract.
Serial reference	SerialExecutor, the deterministic implementation every other executor and every tiling must reproduce byte for byte.
Driver	The process that called find_sources. It plans tiles, submits tasks, reduces their small results and writes the products; it holds no image-sized array.
Worker	A process or thread that runs tasks: a Dask worker, a thread of the pool, or the calling thread itself under the serial executor.
Task	One call of a function on one batch, submitted to the executor. Its arguments and result are small serializable records, never arrays the size of the image.
Batch	The unit of work one task receives: several tiles, or up to 256 objects whose reads fit a budget. Batching keeps the number of tasks proportional to tiles and objects, not pixels.
Tile	One rectangle of the image's grid, the unit of tiled work. Background and RMS use 128-pixel tile cores; every other stage uses 2,048-pixel cores, so an image up to 2,048 pixels is one tile there.
Core	The pixels one tile owns and writes. Cores do not overlap and cover the image exactly once.
Halo	The read-only margin around a core that a stage needs to compute the core's edge exactly, such as a filter's width. Each stage declares the smallest halo that makes its output exact.
Stage	One step of the finder in its own module under hebog.stages: background and detection, multiscale, support, publication, objects, association, sources, islands and catalogue rows, run in that order by composition.py.
Pass	One of the four groups of stages ADR-008 defines, A (noise), B (detection), C (support) and D (objects). A pass ends only where a global reduction must finish before any later decision.
Round	One map_batches call on the executor: one function over one set of batches, with one declared memory. A stage runs one or more rounds.
Boundary summary	The small record a tile task returns about its core's edges, such as the labels on each edge and each label's pixel count, bounding box and peak, so that objects crossing tile edges can be joined without sending pixels.
Reconciliation	Joining tile summaries into one global answer, such as one label for a region that crosses several tiles, through a pairwise tree fixed by tile order so that the result does not depend on completion order.
Invariance	The guarantee that tile shape, partition origin, worker count, task order and retries do not change the products. Tested by comparing one tile against an eleven-by-eleven grid and Serial against Thread and Dask.
Read budget	maximum_batch_read_pixels: the most pixels one object task may read. An object whose window exceeds it is measured from the cores that hold it instead.
Wide object	An object whose window exceeds the read budget: a filament crossing the image, say. The driver reduces it from the pixels its cores return, the one memory term that grows with an object rather than a tile.
Declared memory	The TaskRequirement every round states for its tasks: the measured cost per read pixel times the largest read, with headroom. The executor refuses a task larger than a worker admits and runs only as many at once as fit.
Capacity	ExecutorCapacity: the budget the caller admits, as worker count, threads per worker, tasks in flight and memory per worker. A DaskExecutor without one reads the cluster's current size.
Resource names	The names of the Dask worker resources, chosen by the caller, under which a DaskExecutor requests each task's declared memory and one core, so that Dask places tasks only where that much is free.
Analysis memory	What find_sources itself holds in the calling process, growing with the sources found, about 40 KB each, and with the image through the background's noise grids. hebog.analysis_memory_bytes states it for the caller to reserve, from the image's shape and the upper estimate of the source count that hebog.expected_source_count makes from the image.
Generation	One run's set of intermediate Zarr planes. It is complete only once an immutable marker names exactly one validated chunk for every plane and tile.
Chunk	One stored piece of a Zarr plane, one per tile core. A chunk is a storage unit; a tile is a work unit.
Staging directory	The hidden .<output name>.<random> directory beside the output in which a run builds its planes and products. A finished run removes it; a killed run leaves it with an owner record, which the next run to the same output reclaims once the owner has provably stopped.
Atomic publication	Writing the whole bundle into the staging directory, validating it, then moving it into place with one rename, so that a partial result is never visible.
Secede	A Dask worker task stepping out of its thread slot while it waits on other tasks, so that a waiting analysis does not block the tasks it submitted. Planned for Hebog's analysis inside a Rapthor task.

Rapthor and sky models

Term	Meaning
Rapthor	The LOFAR calibration and imaging pipeline that is Hebog's first consumer. It runs on Prefect and Dask and uses PyBDSF today.
Prefect	The workflow framework Rapthor's main branch runs on. Rapthor's filter step is one Prefect task per sector.
PyBDSF	The Python Blob Detector and Source Finder, the finder Rapthor uses and the reference Hebog is compared against. Agreement with it shows compatibility, not truth.
LSMTool	The LOFAR sky-model tool. Its filter_sources keeps and groups sky-model components by the source finder's mask.
filter_skymodel	Rapthor's step that runs the source finder on a sector and then filters the sky model by the mask. See the Rapthor source-finding contract.
Sky model	The list of known sources, with positions, fluxes and spectral models, that a calibration pipeline predicts the visibilities from. WSClean writes one as a component list.
Sky-model component	One row of a sky model. Not a Gaussian component fitted by the source finder.
Clean component	One component the deconvolver (WSClean) placed while cleaning. Also not a fitted Gaussian component.
Patch	A group of sky-model components calibrated as one direction. LSMTool groups surviving components into patches by the mask's islands.
Source-filtering mask	What Rapthor and LSMTool do with Hebog's mask: keep the sky-model components that lie on detected emission. The mask is not a CLEAN mask, which controls deconvolution.
Adapter	hebog.adapters.find_sector_sources: the module that runs one sector through find_sources and writes Rapthor's file names and the eight-column catalogue view, importing no Rapthor, Prefect or LSMTool code.
Eight-column view	The Source_id, RA, DEC, Isl_Total_flux, Total_flux, DC_Maj, E_RA and E_DEC table Rapthor reads. See the Rapthor catalogue view.
Fallback	Rapthor's planned behaviour of running PyBDSF when Hebog raises a typed SourceFinderError. A refused sector leaves none of Hebog's files behind.
Dual run	Running both finders on one sector and comparing their products with compare_sector_products before choosing one set.

Validation and evidence

Term	Meaning
Injected truth	Generated images whose sources' positions, fluxes and shapes are known, so that completeness, reliability and errors can be measured against an answer.
Completeness, reliability	The fraction of true sources recovered, and the fraction of reported detections that are true, for a stated population and matching rule.
Quick science check	just quick-science-check: 18 fixed cases compared with injected truth and pinned PyBDSF master, run on every scientific change. A regression detector, not qualification. See the contribution guide.
Quick benchmark	just quick-benchmark: five timed repetitions of the public finder per case, compared with the previous release and with PyBDSF master.
Traced peak	just traced-peak: tracemalloc's peak allocation in a fresh process, the figure memory gates use. Peak RSS is reported beside it as an envelope only.
Release check	The sequence of quick check, quick benchmark and traced peak run on a release candidate, whose results the progress page summarises.
Equivalence lane	just test-equivalence: find_sources and its stages against frozen PyBDSF products, run in CI.
Contract test	A test in tests/contract/ of a public behaviour, run through every executor. An unimplemented behaviour is a strict expected failure.
Acceptance test	A Given/When/Then test in tests/acceptance/ of a Rapthor-facing scenario.
Anchor	A fixed input the performance and memory figures are measured on, such as the 10,000² LoTSS-DR3 cut-out and the whole 15,402² mosaic.
Performance ladder	The frozen list of image sizes and workloads a release is timed on. See the performance contracts.
Matched benchmark	The deployment gate: Rapthor's complete filter_skymodel step with Hebog's native task against the same step with pinned PyBDSF master, in one container on the same cores and inputs. Measured by scripts/benchmark/measure_filter_skymodel.py (how to).
Dataset role	development for routine work, regression to preserve a decided case, or qualification, held out from tuning. See dataset manifests.
Evidence record	A versioned JSON document from hebog.validation.evidence that binds a measurement to its inputs, software and environment. reviewed only after its protocol and results pass review.
Qualification	Reviewed evidence that a frozen candidate meets the plan's gates on held-out data. No Hebog release has it; passing tests and development comparisons do not establish it.

Naming conventions

• Use (y, x) for array shapes and indices, and names such as x_pixel and y_pixel where an interface takes (x, y).
• Put the frame and unit in a name the type does not carry: ra_deg, beam_fwhm_deg, frequency_hz, flux_jy, flux_jy_per_beam.
• Say detection_threshold and island_threshold; keep thresh_pix and thresh_isl only at the Rapthor boundary.
• Qualify every component as gaussian_component, clean_component or sky_model_component, and never use source, island, Gaussian component, catalogue row and sky-model component interchangeably.
• Say primary_beam_corrected and flat_noise; keep true_sky and apparent_sky where the Rapthor mapping is explicit.
• Write catalogue in Hebog prose and modules; keep catalog in external field names and file names.
• Name executors serial, thread and dask, as the classes are named.
• A chunk is a storage unit and a tile a work unit; do not use one for the other.

PyBDSF, Rapthor and LSMTool names

Their name	Hebog's term
thresh_pix	Detection threshold
thresh_isl	Island threshold
minpix_isl	Minimum island size
mean_map = "zero"	background = "zero"
rms_box	Coarse grid, 150-pixel windows every 50 pixels
rms_box_bright, adaptive_thresh	Fine grid, 35-pixel windows every 7 pixels, around bright regions at 75σ
atrous_do, atrous_jmax	À trous wavelet scales, used as detection and association evidence
flat_noise_image	Flat-noise image
true_sky_image	Primary-beam-corrected image
island_mask, <image>.mask.fits	Mask, used by Rapthor as the source-filtering mask
srl catalogue	The SOURCES table; the eight-column view for Rapthor
gaul catalogue	The GAUSSIAN_COMPONENTS table
Source_id	SOURCE_ID, numbered from zero in the eight-column view
Total_flux	A source's INTEGRATED_FLUX
Isl_Total_flux	An island's INTEGRATED_FLUX
DC_Maj, E_RA, E_DEC	DECONVOLVED_MAJOR, RIGHT_ASCENSION_ERROR, DECLINATION_ERROR
filter_skymodel	Rapthor's step: run the finder, then filter and group the sky model by the mask
source_finder	LSMTool's choice of finder; hebog is reached through Rapthor's native task instead