Asteroids & fast movers
Rate 35–14000″/hr
ROCKS THAT WON'T STAY PUT.
Everything else on these pages sits still and changes brightness. An asteroid does the opposite: its brightness is dull and constant, and what gives it away is that it is somewhere else than it was. But how it gives itself away depends entirely on speed — a slow rock is a dot that shifts between visits, while a fast one smears into a line inside a single exposure. Same object, two completely different detection problems, two completely different models. The instrument below is invented. The physics is not.
NOCTIS ARRAY does not exist. Asteroids do — and this page explains them the way a real time-domain difference-imaging pipeline actually has to think about them.
Leftovers from a solar system that never finished assembling.
Asteroids are the planetesimals that never accreted into a planet — mostly rock and metal, mostly small, and grouped less by what they are made of than by where their orbits put them relative to us.
They are not interesting because of what they look like. They are interesting because they are somewhere new every hour.
The main beltBetween Mars and Jupiter.
The overwhelming majority of catalogued asteroids orbit two to three and a half astronomical units from the Sun, in a belt whose structure was carved by Jupiter's gravity — complete with resonance gaps swept clean of anything that tried to settle there. Seen from Earth they are far away and moving slowly across the sky, typically a few tens of arcseconds per hour, which for imaging purposes means they behave like ordinary point sources that happen to be in a slightly different place each night.
Near-Earth objectsClose, and therefore fast.
A near-Earth asteroid has been nudged — usually by a resonance, often over millions of years — onto an orbit that brings it inside 1.3 AU of the Sun. Being close does not make it physically special; it makes it apparently fast. Parallax and proximity mean an NEA at close approach can cross thousands of arcseconds per hour, and a handful reach five figures. That single number is what changes everything about how it is detected.
Why anyone bothersTwo unrelated reasons.
Scientifically, asteroids are unmelted samples of the early solar system — their composition records conditions that planetary interiors destroyed. Practically, the near-Earth population is the one class of astronomical object that can arrive here, which is why surveys are funded to catalogue it and why finding a genuinely uncatalogued fast mover is a meaningful result rather than a curiosity.
Exposure time multiplied by apparent speed. That is the whole distinction.
Nothing about the rock changes. The shutter is open for a fixed number of seconds, and whatever angle the object sweeps during those seconds is what lands on the detector. Below about a pixel of motion it is a point; well above that it is a line.
A main-belt asteroid moves less than an arcsecond during a 30-second exposure. A close-approach NEA can move nearly a full arcminute.
The slow caseA point source that relocates.
At roughly 35 arcseconds per hour, a main-belt asteroid drifts about 0.3 arcseconds during a 30-second exposure — a fraction of a single pixel on a survey camera, and far less than the atmospheric blur already smearing every star in the frame. It records as a perfectly ordinary point source. The only thing that betrays it is comparing two visits: the stars are where they were, and this one is not.
The fast caseA line drawn inside one frame.
At several thousand arcseconds per hour the same 30 seconds smears the object across tens of pixels. It stops being a point source entirely — its flux is spread along a trail, so it is fainter per pixel than a stationary object of the same total brightness, and every measurement tuned for point sources starts to under-report it. But the trail itself is a signature no star can imitate, visible in a single exposure with no second visit required.
Why it forces two pipelinesDifferent shape, different question.
Finding a dot that moved needs several exposures, precise astrometry and a linking step that decides which dots belong to the same rock. Finding a streak needs one exposure and a classifier that recognises an elongated shape and can tell a real trail from a cosmic ray or a subtraction artefact. These share almost no machinery, which is exactly why the project below runs them as two independent branches.
The same rock, at three different apparent speeds, as the detector actually records it:
Both branches are built, and one of them has recovered a real near-Earth asteroid.
The slow branch reaches asteroids through the same stamp classifier that handles stationary objects — "asteroid" is one of its four output classes, and on a fair labelled evaluation it was the model's most reliable call at 0.89, better even than supernovae. Across our own harvested data it has typed 27 detections, 14 distinct sources, as point-source asteroids.
The fast branch is the larger machine. Elongated detections are cut as 144×144 pixel stamps straight from the difference image and pushed through a pretrained three-gate cascade — real-versus-bogus, then keep-versus-ditch (the actual cosmic-ray filter), then short-versus-long, where short means asteroid and long means satellite. Across 666,811 detections in 503 exposures that cascade has returned 1,057 short near-Earth-asteroid candidates, alongside 2,343 cosmic rays, 16 satellites and 419,160 rejections.
Candidates are then linked: detections from different exposures — often different fields, since a fast rock leaves one field between that field's two nightly visits — are fitted to a single constant-rate motion vector, and a track needs at least three consistent points to exist at all. That has produced 85 tracks from 643 member detections.
The flagship result is a blind recovery of 2019 BE5, a real near-Earth asteroid. With no ephemeris supplied to the detection stage at any point, the pipeline detected it, routed it through the cascade, and linked 13 detections spanning two fields and a 70-minute arc into one track — measuring its motion at 6,901 arcseconds per hour against a catalogued 6,664, a ratio of 1.04. It then correctly identified the object as already known.
That is a recovery, not a discovery, and the distinction is the point. 2019 BE5 was discovered by other people in 2019; finding it again blind is how you demonstrate the machine works. This project has made no novel discoveries, and says so everywhere. The honest ceiling is documented too: a fainter target, 2018 VJ10 at roughly magnitude 17.6, was detected and measured correctly — 6.6 arcseconds from truth — but the pretrained real/bogus gate declined it as too marginal. End-to-end recovery works for bright fast movers and stops somewhere between magnitude 15 and 17.6.
Finding a moving object is the easy half. Proving nobody has it already is the hard half.
Most asteroids a survey detects are catalogued. Deciding whether a particular streak is one of them — and refusing to guess when you cannot tell — is what separates a candidate list from a claim.
"I asked and nothing was there" and "I could not ask" are different answers. Collapsing them into one is how an outage becomes a discovery.
-
01
Match on speed, not just position.
Near the ecliptic, any ten-arcminute cone you search will contain slow main-belt asteroids — so "is there a catalogued object near this position" always answers yes, and means nothing. Our filter first restricts the catalogue response to fast movers above 1,000 arcseconds per hour. In the 2019 BE5 field that reduced 108 known objects to exactly one: BE5 itself. Rarity is what makes the match unambiguous.
-
02
A freshly-discovered object has a bad catalogue entry.
An orbit fitted to a short observing arc predicts poorly. The catalogue's position for 2019 BE5 was six arcminutes off and its predicted brightness six magnitudes wrong — but its motion was roughly right. Because a track carries a measured rate, not just a position, our match confirmed identity on speed where position alone would have failed: 6,901 against 6,664 arcseconds per hour.
-
03
Never let a failed lookup mean "new".
The catalogue service is a live external server, and during one sweep it returned server errors. The naive handling — treat no answer as no match — would have marked every candidate in that window a novel discovery. The filter therefore reports three states, not two: known, novel, and unchecked, with "novel" requiring a lookup that actually succeeded.
-
04
One detection flagged "novel" is not a candidate.
This is the most important line on the page, and it was learned the expensive way. In a blind hunt across sixteen visits to unvetted sky, the cascade produced 19 short candidates and the per-detection novelty filter flagged 16 of them as novel. After linking, zero survived — they were fixed positions recurring across time with implied rates far below any real mover, the fingerprint of subtraction artefacts. Without the linker that hunt reports sixteen false discoveries. With it, none.
Proven in both directions. The "novel" verdict is not an untested branch. Running the real 2019 BE5 track past the catalogue twice — once normally, once with BE5 itself hidden from the response, as though it had never been catalogued — flips the verdict from known to novel exactly as it should. Same track, same measurements, only the catalogue's knowledge changed. An actual novel detection is now a question of sweeping sky that contains one, not of untested code.
Fast movers are the least cooperative thing a fixed survey can chase.
Everything that makes a near-Earth asteroid worth finding — speed, proximity, faintness — also makes it awkward for an instrument designed to tile the sky on a schedule.
Trailing lossSpeed costs you brightness.
Spreading a fixed amount of light along a 50-pixel trail leaves far less in any single pixel than concentrating it in one stellar profile. The faster the object, the fainter it appears to every detection threshold tuned for point sources — so the objects most worth catching are systematically the hardest to see.
FragmentationOne rock, several detections.
Segmentation at a low threshold chops a long trail into collinear pieces rather than returning one object. Our own BE5 detections arrived as two or three fragments per exposure, which the linker has to merge before fitting — feed them in raw and the spread within a single exposure reads as impossible motion and poisons the fit.
Cadence geometryTwo visits cannot make three points.
A public survey images a given field roughly twice a night. A track needs three consistent detections. A mover confined to one field therefore cannot produce a track at all — BE5 only reached thirteen detections because it was fast enough to cross several adjacent fields in one night. Where you point matters as much as what you run.
Go see the tracks it actually built.
The array, the cadence and the field notes above are invented. The browsers below are not — they open the same working ZTF pipeline whose 85 tracks and blind 2019 BE5 recovery are described in the panel above.
Back to the overview