Precovery Methods

How ADAM Precovery recovers an object's historical detections: orbit propagation, on-sky covariance frame selection, and the detection-acceptance gates, plus why it uses a 3-sigma innovation gate and an 800-arcsecond on-sky uncertainty cutoff.

What precovery does

Precovery ("pre-discovery recovery") finds an object's previously unrecognized detections in archived survey imaging. Given an orbit and its covariance, ADAM Precovery propagates the orbit to the time of each historical exposure, predicts where the object should have appeared on the sky, and searches that exposure's archived detections for a match. Recovered detections extend the object's observational arc and sharpen its orbit, often turning a short, uncertain arc into a well-constrained one.

Designation-based searches currently obtain the input orbit and covariance from one of these sources:

  • JPL Small-Body Database (SBDB)
  • ESA Near-Earth Object Coordination Centre (NEOCC)
  • JPL Scout
  • Minor Planet Center (MPC)

How it works

  1. Propagate with uncertainty. The orbit and its 6×6 covariance are propagated to each candidate exposure epoch with an n-body integrator, transporting the covariance (via sigma-point / variant particles) rather than re-sampling it at span boundaries.
  2. Predict position + uncertainty. At each epoch the pipeline produces the predicted on-sky position and its 2D uncertainty ellipse (the projected on-sky covariance).
  3. Select candidate frames. The predicted uncertainty region is rasterized to HEALPix pixels and intersected with each exposure's footprint pixels. Exposures whose footprint overlaps the predicted region, and the detections inside them, become candidates.
  4. Filter detections. Each candidate detection passes through the acceptance gates described below.
  5. Record outcomes. Accepted detections are recoveries; every other candidate is recorded with a fine-grained diagnostic reason (outside the healpixel match, magnitude outlier, innovation-gate rejection, no catalog detection, and so on).

Magnitude prediction

When the submitted orbit includes absolute-magnitude information, precovery also reports a predicted magnitude for each candidate detection. The astrometric prediction is produced from the sigma-point / variant-particle cloud: the pipeline collapses those particles into a weighted mean sky position and an on-sky covariance ellipse. The reported magnitude, however, is evaluated from the nominal particle — the central orbit represented by the nominal sigma point — and then attached to the collapsed astrometric prediction.

This preserves the intended model choice: a central-orbit brightness estimate shown alongside a covariance-aware position and uncertainty. It is not a brightness computed from the post-collapse weighted-mean coordinates. For ordinary well-constrained objects the two choices are effectively indistinguishable, but for extremely uncertain short-arc objects they can differ because brightness depends nonlinearly on observer-object-Sun geometry. In those cases the predicted magnitude should be interpreted as an approximate central-orbit value, not a full distributional photometric summary.

On-sky uncertainty frame filter (800 arcsec)

Before searching an exposure, precovery inspects the predicted on-sky covariance. If the 1σ major axis of the uncertainty ellipse (its along-track length) exceeds 800 arcseconds, the exposure is skipped: the predicted position is too uncertain — a broad covariance tail, common near planetary close approaches or for short-arc orbits — to search productively. Skipping these frames prevents "frame explosion" (an enormous candidate set) and the false positives that come with it.

Why 800 arcsec? The cutoff was selected from a parameter study sweeping the sigma-major threshold over NEOMOD synthetic and Rubin-like detection sets. Tightening the cutoff from 1000 to 800 arcsec cut contamination by roughly 2.3× (a false-positive-like-to-truth ratio of 0.47 down to 0.20) while costing only about 0.2 percentage points of recall (0.608 to 0.607). 800 arcsec is the balanced operating point: it keeps essentially all of the recoverable real detections while sharply suppressing noise, and keeps search cost low. (Frame-level uncertainty filtering uses the sigma-major axis only.)

3σ innovation gate

For each candidate detection that survives frame selection, precovery computes the innovation — the offset between the observed position and the predicted position — and normalizes it by the combined covariance (the predicted-position uncertainty together with the detection's own reported astrometric uncertainty). A detection is accepted only if this normalized innovation distance is within 3σ: a 2D Mahalanobis distance, i.e. the predicted on-sky innovation ellipse.

Why 3σ? For the well-behaved majority of objects, the true detections fall within a few sigma of the prediction, so a 3σ ellipse recovers them. Because that ellipse is small (in studied data the predicted innovation major axis is around 0.3 arcsec at the median), an unrelated source has a vanishingly small chance of landing inside it, so 3σ keeps recall high while keeping false positives very low. Loosening the gate (5σ, 7σ, 10σ) recovers more in pathological regimes but pulls in proportionally more false positives, which are better controlled downstream by requiring multiple consistent detections than by widening the default gate.

The main limitation is objects whose predicted mean is biased — most notably active comets with non-gravitational accelerations (A1/A2/A3) that the ephemeris does not model — where even genuine detections can sit many sigma from the prediction. For those the right remedy is a regime-specific wider gate plus multi-detection consistency, not a looser global default. (The gate also clamps the innovation sigma to a small floor so it is never made artificially tight by optimistic reported uncertainties.)

Datasets & provenance

Precovery searches five survey-detection datasets across ten mapped MPC observatory codes:

  • ATLAS difference-image detections (atlas ): ATLAS survey telescopes at Mauna Loa (T08), Haleakalā (T05), South Africa (M22), and Chile (W68); catalog coverage 2015-06-18 onward.
  • Zwicky Transient Facility alerts (ztf ): Samuel Oschin Telescope at Palomar Observatory (I41); catalog coverage 2018-06-01 onward.
  • NOIRLab Source Catalog DR2 (nsc ): Blanco/DECam at CTIO (W84), Bok/90Prime (V00), and Mayall/Mosaic3 at Kitt Peak (695); catalog coverage 2012-09-23 through 2019-11-16.
  • SkyMapper Southern Survey DR2 (skymapper ): SkyMapper telescope at Siding Spring Observatory (Q55); catalog coverage 2014-03-15 through 2018-03-31.
  • Rubin Observatory / LSST alerts (lsst ): Simonyi Survey Telescope at Vera C. Rubin Observatory (X05); catalog coverage default start 2025-09-01; continuing.
  • Minor Planet Center published observations ( mpc ): Published astrometry from the worldwide MPC station network. This is a non-image reference dataset used for recovery and attribution accounting, not a single telescope archive.

Each source has its own sky coverage, cadence, and quality cuts. See the AIMS data provenance pages for per-source detail on how each dataset is ingested and filtered.