The Kuiper belt opens: CCDs, moving-object software and the bright end

Eris was in the images on 21 October 2003. The Palomar wide-field camera had photographed it three times over three hours, and the software had looked at it and moved on. The main search accepted objects moving at least 1.5 arcseconds per hour across the sky. Eris was moving at 1.42. Fourteen months later, on 5 January 2005, a reprocessing of the old data with a slower motion search turned it up. At $R = 18.38$ it was more than two magnitudes brighter than the survey’s limit (Brown et al. 2005), and it would lead, the following year, to Pluto’s reclassification.

Nothing about the photons had changed between 2003 and 2005. What changed was a number in a configuration file, and the willingness to look at the ten to twenty times more false candidates that the lower number let through. That is the central lesson of the period this chapter covers. Between 1992 and 2010 the discovery of distant bodies became an interaction between detectors, observing cadence, software and follow-up. The underlying observation stayed simple: an unresolved body changes position against the stars. The hard questions were how slow a motion to accept, how faint a source to trust, how much sky to cover, and how to keep an unusual orbit alive through the months of recovery that turn a moving dot into a member of a population.

Three survey goals organise the story. The deep CCD imaging of David Jewitt and Jane Luu established that the belt existed and measured its ordinary members. The Caltech and Palomar programme of Michael Brown, Chadwick Trujillo and David Rabinowitz searched huge areas for the rare bright members, and found Sedna, Eris and Makemake. The Deep Ecliptic Survey (DES, not to be confused with the later Dark Energy Survey) and the Canada–France Ecliptic Plane Survey (CFEPS) set out to collect orbital samples whose selection was well enough understood to test models of how the Solar System formed.

An empty region?

Pluto was found in 1930 and Charon in 1978 (Chapter 4). Theorists had argued for a reservoir of small bodies beyond Neptune, but no one had seen one. Jewitt and Luu began a systematic search in 1987 with an observational question: was the apparent emptiness real depletion, or just faintness? Their retrospective describes searches on several telescopes with photographic plates and CCDs that found nothing, before the Mauna Kea programme that succeeded (Jewitt & Luu 1995).

The reason earlier searches had failed is a matter of physics. For a body seen by reflected sunlight,

where $p$ is the geometric albedo, $D$ the diameter, $r$ and $\Delta$ the distances from the Sun and observer in au, $\alpha$ the phase angle, $\Phi$ the phase function and $H$ the absolute magnitude. At opposition for a distant body $r \approx \Delta$, so the flux falls as $r^{-4}$. Doubling the distance costs $10\log_{10}2 \approx 3.01$ magnitudes (Chapter 2). A hypothetical body of 21st magnitude at 20 au would be 24th at 40 au. Tombaugh’s plates reached 17, Palomar–Leiden 20.5. The belt was beyond the reach of photographic plates for all but its very largest members. No search algorithm restores photons that were never collected.

The CCD changed the arithmetic. It records a far larger fraction of the incoming photons than a photographic emulsion, and its response is linear, so images can be calibrated, subtracted and added. On the 2.2-m telescope described below, four 900-second CCD exposures reached about $R = 25$. The price was a tiny field: a 2048-pixel chip at 0.22 arcseconds per pixel covers about 7.5 arcminutes on a side, so a deep search could only sample small patches of sky.

Physics: brightness, motion and trails

The rate of motion at opposition, derived in Chapter 4, is

for a circular, prograde orbit near the ecliptic. The numerator includes the body’s own orbital motion as well as Earth’s reflex motion. At large $r$ it tends to $148/r$, often rounded to $150/r$. The formula appears in the QB1 discovery paper and in later surveys (Jewitt & Luu 1993; Trujillo et al. 2001). It is not a universal distance estimator: eccentric, inclined or off-opposition bodies depart from it.

Over a single night, a constant-velocity model $\mathbf x(t) = \mathbf x_0 + \mathbf v(t - t_0)$ is accurate because the night is short, not because the orbit is circular. During one exposure of length $t_{\rm exp}$ the image trails by $\ell = \mu\,t_{\rm exp}$. At 40 au, $\mu \approx 3.2$ arcseconds per hour, so a 15-minute exposure trails by 0.8 arcseconds, about one seeing width on Mauna Kea. Exposure lengths in the early surveys were chosen to keep the trail within the seeing disk. Longer baselines separate slow motion from centroid errors, but eventually require curvature, parallax and changing geometry to be modelled.

1992 QB1

On 30 August 1992 (UT), Jewitt and Luu were observing with the University of Hawaii 2.2-m telescope on Mauna Kea. The discovery paper gives seeing of about 0.8 arcseconds, a Tektronix 2048 × 2048 CCD with 0.22-arcsecond pixels, and 900-second exposures through a Mould $R$ filter. Each field received four images over at least two hours, tracked at the sidereal rate, near opposition. They searched the images in real time. One faint point moved, consistently, by the right amount (Jewitt & Luu 1993).

The object moved 2.6 arcseconds per hour west and 1.1 south, about 2.8 arcseconds per hour in total. Follow-up on 31 August and 1 September confirmed the motion. The absence of detectable diurnal parallax helped rule out a nearby asteroid that happened to mimic a distant body’s rate. Brian Marsden at the Minor Planet Center computed orbits, and observations through 25 December placed the body about 41 au from the Sun. The formula above gives about 3.1 arcseconds per hour at 41 au, within about 10% of the measured rate. Its brightness was $R = 22.8 \pm 0.2$, well above the survey limit. The preliminary orbit had $a = 44.4$ au, $e = 0.11$ and $i = 2.2^\circ$, values that later observations refined (Jewitt & Luu 1993). The object is now named Albion.

At publication the survey had covered 0.7 square degrees to about $R = 25$. A note added in proof recorded a second object, 1993 FW, found on 28 March 1993. One object could be a curiosity. Two meant a population.

Counting from a handful

The 1995 paper describes the method precisely enough to reproduce (Jewitt & Luu 1995). Each field, about 448 arcseconds wide, received four 900-second exposures. The first three were consecutive, about 1,050 seconds apart, and the fourth extended the baseline beyond two hours. Small dithers of about 20 arcseconds let them build flat fields from the night sky. The images were flattened, sky-subtracted with a spline fit, aligned on stars, and blinked by eye at the telescope.

for each four-exposure field:
    subtract bias and sky; flat-field from dithered night-sky frames
    align the images on stars
    blink the registered images at matched display stretch
    mark point-like sources with coherent slow displacement
    check flux, direction and constant rate in every exposure
    measure positions; schedule further nights; send to the MPC

The selection demanded linear constant-velocity motion, a stellar image, consistent brightness and a rate below 10 arcseconds per hour. There was no explicit lower rate limit and no directional selection, although the objects found did all move westward as expected near opposition. The paper favoured the human eye over the software of the time near the photometric limit, a contemporary operational judgement, not evidence against automated detection.

The calibration was blinded. Artificial moving objects with the right PSF and a rate of 3.2 arcseconds per hour were added with IRAF’s ARTDATA package, and the observer did not know where they were. Recovery was reliable to about $R = 24.5$ and zero at $R = 25.2$, and a linear ramp between the two served as the selection function. The homogeneous deep survey covered 1.2 square degrees and found seven objects. The injections measured sensitivity at one rate. They did not map how efficiency depended on rate or crowding.

Mosaics and machines

A new dynamical class

The next step combined a larger field with automated detection. On 15 October 1996 the UH 2.2-m with an eight-chip mosaic camera, binned to 0.41-arcsecond pixels and covering 0.09 square degrees, imaged a body of $R \approx 20.9$ after about 25 square degrees of searching. Its first short arc fitted a plutino, a body in the 3:2 resonance with Neptune like Pluto. December observations ruled that out. The eventual orbit, $a \approx 84$ au, $e \approx 0.58$, $i \approx 24^\circ$, showed a population reaching far beyond the belt sampled so far (Luu et al. 1997). This was 1996 TL66, a member of what is now called the scattered disk.

The detection algorithm found it, but the discovery of a new class required something else: the willingness to keep following an object whose orbit did not match the preferred interpretation. A survey that recovered only objects whose ephemerides came from a default orbit would have lost it.

Triplets at CFHT

Trujillo, Jewitt and Luu’s 1999–2000 survey at the Canada–France–Hawaii Telescope shows the mature automated method (Trujillo et al. 2001). The CFH12K mosaic had 12,288 × 8,192 pixels of 0.206 arcseconds, about 0.330 square degrees per field and about 200 MB per image. Each field was observed three times with 180-second exposures about an hour apart. Fields were placed at ecliptic latitudes of 0, 10 and 20 degrees, so that the survey sampled inclinations rather than assuming the ecliptic was representative.

After overscan subtraction, flat-fielding and bad-pixel masking, the Moving Object Detection Software (MODS), written by Trujillo and Jewitt, searched the triplets. The paper documents automated processing, artificial-source calibration and near-real-time candidates but not every internal data structure. A reconstruction of the general catalogue-association method is:

detect sources separately in each of the three exposures
register the catalogues; remove sources with stationary counterparts
index the remaining detections by sky position
for each plausible (first, last) pair:
    infer the velocity; require 1 <= rate <= 10 arcsec/h
    predict the middle-epoch position
    look for a detection there with consistent flux and shape
inspect survivors by eye; begin follow-up promptly

Enumerating all triplets of $n$ sources per epoch costs $O(n^3)$. Enumerating first–last pairs and querying an indexed middle catalogue costs about $O(n^2\log n + K)$, with $K$ the number of matches returned, and much less when a tight rate window limits the pairs. Bright stars and galaxies can dominate the practical cost of both image processing and association. These are implementation estimates, not the authors’ benchmarks.

The rate window of 1–10 arcseconds per hour corresponds to about 10–140 au on the opposition formula. Efficiency was uniform with rate inside that tested range. MODS ran on two Sun Ultra 10 workstations. The team took about 35 field triplets and 20 GB of raw data a night and processed them fast enough for prompt recovery. The survey found 86 KBOs in about 73 square degrees, two of them previously known. The paper gives no CPU-hour or FLOP count, and none should be invented (Trujillo et al. 2001).

The efficiency measurement is the part that matters most. The paper fitted

with plateau $\epsilon_{\max} = 0.83$, half-plateau magnitude $m_{50} = 23.74$ and width $w = 0.48$ overall. Note that $m_{50}$ is where efficiency falls to half the plateau, 0.415, not to 50%. Split by seeing, the fits differ sharply.

Good, medium and poor seeing (median FWHM 0.76, 0.90 and 1.07 arcseconds) gave half-plateau magnitudes of 24.01, 23.64 and 23.35, and widths of 0.29, 0.38 and 0.47 magnitudes (Trujillo et al. 2001). Two-thirds of a magnitude of depth came and went with the weather, on the same telescope and camera. For background-limited point sources this is expected: the noise in a PSF-sized aperture scales with the PSF width, so a 40% wider image costs about 0.37 magnitudes on that argument alone. A single “limiting magnitude” hides this. The plateau of 0.83 is also worth noticing: even bright objects were missed 17% of the time, through chip gaps, masks, bright stars and blends.

Going deep: pencil beams and shift-and-stack

The alternative to covering more sky was to stare longer at a small patch. Brett Gladman, JJ Kavelaars, Philip Nicholson, Thomas Loredo and Joseph Burns did this at Palomar and CFHT in pencil-beam surveys published in 1998 (Gladman et al. 1998). They registered many exposures on the stars, then shifted and combined them along hypothesised object motions, so that light from a body moving at the trial rate piled up in one place while stars smeared out.

For each trial rate and direction they formed two combinations: a sum, which preserves flux, and a robust combination that rejects the highest value at each pixel and takes the median of the rest, which suppresses cosmic rays and residual stars. Rates were spaced at about one-third of the seeing per hour. For Palomar data they used three directions, 18, 23 and 28 degrees from the equator, enough for the targeted inclination range given the observing geometry. They blinked adjacent trial-rate images: a real object brightens and fades as the trial rate passes through its true value.

A representative Palomar field on 6 September 1997 combined 27 exposures of 480 seconds over a 5.80-hour baseline, covering 0.025 square degrees, and reached a 50% recovery magnitude of $R = 25.6 \pm 0.2$. The survey found five objects and estimated about 90 objects per square degree brighter than $R \approx 25.9$ (Gladman et al. 1998).

The cost is in the trials. With $N_v$ trial velocities, $J$ images and $P$ pixels, direct shift-and-add costs $O(N_vJP)$, and the median adds a per-pixel selection cost. The faintest pencil-beam detections were also hard to recover for orbits, so they served as a luminosity-function sample, not an orbital census. The methods and their accelerations are the subject of Chapter 17, Chapter 18 and Chapter 21.

The Deep Ecliptic Survey: red and cyan

The Deep Ecliptic Survey used the Mosaic cameras on the 4-m Mayall and Blanco telescopes and began successful searching in October 1998 (Millis et al. 2002). It was led in its first methods paper by Robert Millis, Marc Buie, Lawrence Wasserman, James Elliot, Susan Kern and R. Mark Wagner. The original cadence was three 300-second exposures about two hours apart. In 2000 the team switched to two exposures per field to cover more area, with a third usually taken on an adjacent night, and later used routine exposures of four minutes (Buie et al. 2003). Images were binned 2 × 2 to 0.52-arcsecond pixels to cut readout time and storage. A broad $VR$ filter, centred near 6080 Å and about 2230 Å wide, collected as many photons as possible. Fields lay within 6.5 degrees of the ecliptic and generally within 30 degrees of opposition.

Source lists were made at about $3\sigma$ above sky. Stationary sources were removed, and software paired candidates with rates up to 5 arcseconds per hour within 30 degrees of the ecliptic direction. Then people inspected every image pair, not just the software’s candidates. The first exposure went into the red channel of a colour display and the second into green and blue. Stationary sources appeared white. A moving object appeared as a red dot next to a cyan one. Two people scanned each field independently, followed by a final review of slow movers and any confirming frames (Millis et al. 2002).

The distinction between the software gate and the survey’s acceptance matters. The authors describe visual sensitivity from about 0.5 to 200 arcseconds per hour and tried to report moving sources at any rate. Calling DES a strict $\mu \lt 5''\,{\rm h}^{-1}$ search would be wrong. Equally, visual recognition across that range is not a published uniform efficiency.

A two-image search gets little rejection from linearity, since two positions always define a line. DES made up for that with human inspection and second-night confirmation. Orbit work started with families of Väisälä orbits (a classic short-arc method that assumes the body is at perihelion at one of the observations) and then used Bernstein–Khushalani fits for recovery planning. Classification integrated orbits for 10 million years, rather than assigning resonances from semimajor axis alone (Elliot et al. 2005).

DES depth numbers need their definitions. The first paper found an image-level limit near 23.9 in the white filter, defined by a Gaussian source whose peak was $2\sigma$ above sky, a peak-pixel criterion, not an integrated signal-to-noise of two. The later survey-wide analysis reported a mean 50% sensitivity of $VR = 22.5$, obtained by fitting population and sensitivity parameters together, about 1.43 magnitudes brighter than the image-level limits, and explicitly provisional pending better calibration (Elliot et al. 2005). Through 2003 the survey covered about 550 square degrees in 88 allocated nights, with 655 observing hours. It reported 382 designated discoveries, 196 of them secure enough for dynamical classification, and documented 240 more undesignated candidates.

CFEPS: a survey designed to be interpreted

CFEPS treated discovery, recovery, calibration and modelling as one experiment (Jones et al. 2006; Petit et al. 2011). Its camera was CFHT’s MegaCam: 36 CCDs, 0.187-arcsecond pixels and a field of about 0.96 by 0.94 degrees. A preliminary survey began in 2002. The main-survey cadence had discovery images of 70 seconds in $g'$, about 50 minutes apart, a 90-second “nailing” exposure a few days later, check-up exposures of 180 seconds in $i'$ about two months later, and 110-second $r'$ recovery exposures a year later.

The key design choice was to arrange fields in contiguous blocks and revisit the blocks, not individual predicted positions. If recovery depends on an ephemeris computed from an assumed orbit, an object with a surprising orbit falls outside the recovery images and quietly disappears from the sample, while ordinary objects survive. The survey then becomes biased toward the orbits it expected. Revisiting the whole block keeps unusual bodies in view.

Detection used the Moving Object Pipeline (MOP), which ran two independent source finders, SExtractor and a wavelet method. Each stream removed stationary sources and looked for moving triplets. Only candidates found by both were shown to an operator. The preliminary survey searched rates of 1–6 arcseconds per hour within 23 degrees of the ecliptic direction. Requiring both extractors reduced contamination but could reject real sources that either one missed (Jones et al. 2006).

The calibration was thorough. The presurvey implanted 42,286 artificial objects, about 150 per chip, from magnitude 21 to 25, and fitted

with $A = 0.906$, $R_c = 24.20$, $W_1 = 0.615$ and $W_2 = 0.309$. Absolute 50% recovery was at about $R = 23.9$, and efficiency was zero outside the rate window.

The final release characterised 321 square degrees. Block limits ran from about $g = 23.5$ to 24.4 (AB), defined at 40% efficiency, because below that level the authors found that human and automated acceptance diverged. Fainter discoveries were reported but excluded from the quantitative sample. Rate windows varied by block: 1.7–5.1 arcseconds per hour in one, 0.4–7.3 in another (Petit et al. 2011). Of 231 discoveries, 196 lay inside the characterised limits, and 169 of those, about 86%, were tracked to precise orbits with semimajor-axis uncertainties below 0.1%. The tracking fraction was modelled as

within the characterised range only.

Those nested numbers carry the chapter’s statistical lesson. A survey’s scientific sample is not its discovery list. It is the subset whose detection and tracking probabilities are known. The CFEPS project estimated that follow-up took about five times the observing resources of discovery (CFEPS).

With measured efficiencies, the expected number of detections under a population model with parameters $\theta$ can be written

where $\mathbf o$ are orbital elements, $c$ colour, $W_k$ the sky coverage of block $k$, $\epsilon_k$ the detection efficiency as a function of magnitude, rate $\mu$, direction $\psi$ and colour, and $f_{\rm track}$ the recovery probability. A Poisson likelihood is $-\Lambda + \sum_i \log\lambda(z_i\mid\theta)$ over the detected objects $z_i$. CFEPS implemented this as a forward survey simulator, drawing synthetic orbits, propagating them to the observation dates, testing footprints and rate windows, computing brightness, accepting with the block’s efficiency and tracking probability, and comparing the simulated detections with the real ones. That simulator’s successor is used by later surveys (OSSOS SurveySimulator; Chapter 16).

One relation from the luminosity-function literature is worth keeping. If the cumulative sky density of objects brighter than $m$ is $\Sigma(\lt m) = 10^{\alpha(m - m_0)}$, and all objects share an albedo and a distance distribution independent of size, then a size distribution $dN/dD \propto D^{-q}$ implies $q = 5\alpha + 1$. Breaking those assumptions changes the interpretation, and early disagreements came partly from treating cumulative points as independent measurements and from ignoring uncertainty in efficiency (Gladman et al. 1998; Trujillo et al. 2001).

The bright end: Palomar

The Caltech Wide Area Sky Survey

The deep surveys sampled the faint, numerous members of the belt. The rare large ones, each a few per thousand square degrees at best, needed a different design: very wide coverage at modest depth. In late November 2001, Chadwick Trujillo and Michael Brown began the Caltech Wide Area Sky Survey with the robotised 48-inch Samuel Oschin Schmidt at Palomar (the telescope Gehrels and Kowal had used, Chapter 4) and the NEAT camera (Trujillo & Brown 2003).

That camera had three 4080 × 4080 CCDs at about one arcsecond per pixel, a nominal footprint of 3.85 square degrees and an average useful tessellated coverage of 2.75 square degrees per exposure. A typical clear night observed 12 fields three times at intervals of about 1.5 hours. Median seeing was 3.1 arcseconds. A modified version of MODS searched the images with a rate window of 1–10 arcseconds per hour, which the paper translated into 11.4–136 au. Its outputs took about two hours of visual review per good night’s 33 square degrees. By mid-April 2003 the survey had covered 5,108 square degrees and detected 28 KBOs and four Centaurs, 19 of them new, including the large body Quaoar. The paper estimated a provisional depth of $R = 20.7$ from the median brightness of its detections, acknowledging that its photometry was not fully characterised.

The paper also recorded a warning. About 60 apparently promising very slow candidates all turned out, after more imaging, to be background objects (Trujillo & Brown 2003). At the slow end, the sky itself generates false movers: seeing changes, refraction, blends, imperfect centroids and sources that flicker above and below the detection threshold all produce small apparent displacements of things that do not move.

In summer 2003 the survey switched to the QUEST camera described by Charles Baltay and collaborators: 112 CCDs and about 161 megapixels, with 9.6 square degrees of active area, used by the distant-body survey with an effective field of 8.3 square degrees (Baltay et al. 2007). The camera could drift-scan or point and track. The Kuiper-belt survey used repeated pointed images. Brown’s own account describes a bank of ten computers for the Caltech search (Brown, planetlila). The QUEST team separately reported about 100 GB per night for point-and-track observing with 60-second exposures, and eight hours on 28 nodes per night for the Yale photometric pipeline. That benchmark belongs to the general QUEST processing, not to the moving-object search, and the two should not be combined into a single discovery compute budget.

Sedna

On 14 November 2003, Brown, Trujillo and Rabinowitz imaged an object of $R = 20.7$ that moved 4.6 arcseconds in 3.1 hours (Brown et al. 2004). That is 1.48 arcseconds per hour, which on the opposition formula corresponds to a circular orbit near 90 au (a derived estimate). The body was the most distant known in the Solar System.

A few weeks of new observations could not say what kind of orbit it had. A short arc of a very distant body fixes its position and angular velocity well and its radial velocity hardly at all, so orbits of very different eccentricities fit equally well. The orbit was established by going back. Archival detections from 2001 to 2003 were added one at a time, each shrinking the uncertainty region for the next earlier epoch. A 2002 detection favoured the unconstrained least-squares orbit over the bound-energy preference built into the Bernstein–Khushalani fitter, and after it was included the two methods agreed. The paper estimated the probability of an accidental match by counting unrelated sources in the surrounding image regions, and rejected a possible 1991 detection whose coincidence probability was about 3% while accepting much cleaner later matches. The published solution gave a heliocentric distance of $90.32 \pm 0.02$ au, $a = 480 \pm 40$ au, $e = 0.84 \pm 0.01$ and a perihelion near 76 au (Brown et al. 2004).

A perihelion of 76 au puts Sedna far beyond Neptune’s influence. Detecting it was a matter of a moving dot. Establishing what made it remarkable needed the archive. (The 2005 Eris paper describes Sedna’s motion as 1.75 arcseconds per hour, which does not match the 4.6 arcseconds in 3.1 hours of the discovery preprint. The two values are reported here as given.) What placed Sedna on its orbit is still debated, and is the subject of Chapter 8.

Eris and the 1.5-arcsecond cut

Eris was imaged on 21 October 2003 and recognised on 5 January 2005. Brown’s account corrects an initially reported 8 January (Brown, planetlila). The main search of that period rejected rates below 1.5 arcseconds per hour. Eris moved at 1.42 (Brown et al. 2005).

The cut was not arbitrary. At a seeing of two to three arcseconds, a body moving 1.42 arcseconds per hour shifts only about 2.1 arcseconds between consecutive images 1.5 hours apart. That is less than the width of its image. Each adjacent pair of detections barely moves at all, and the stationary sky produces many apparent movers of that size. A low rate cut kept the candidate list manageable.

The slow reprocessing used the geometry that the figure shows. It paired the first and third images across the full three-hour baseline at rates of 1–2 arcseconds per hour, then checked that the middle image agreed, without requiring a significant displacement in each adjacent pair. Candidates within 2 arcseconds of a USNO catalogue star were discarded. This slow search produced 10–20 times the false positives of the fast one, about 1,200 candidates a month for human review. In the two years of data examined for the discovery paper, exactly two real slow objects survived: Eris and Sedna (Brown et al. 2005). That ratio, thousands of false candidates per real one, measures operational contamination. It is not a Gaussian false-alarm probability.

Makemake and Haumea

Makemake’s discovery observations are dated 31 March 2005 in the Minor Planet Electronic Circular that announced it, at $R = 16.8$–16.9 and an initial inclination near 29 degrees (MPEC 2005-O42).

Haumea needs care. The Caltech team independently identified and studied it, but MPEC 2005-O36 records observations from the Sierra Nevada Observatory in March 2003, including a discovery-marked measurement on 7 March, by a team that included Francisco Aceituno, Pablo Santos-Sanz and José Luis Ortiz (MPEC 2005-O36). NASA’s current account says that two teams claim the discovery, and that the IAU record lists the Sierra Nevada Observatory without an official discoverer (NASA, Haumea). The credit should not be assigned to either team alone.

Pluto reclassified

The Kuiper belt made Pluto one member of a large population, and Eris made a decision unavoidable. In 2006 the International Astronomical Union defined a planet of the Solar System as a body that orbits the Sun, is massive enough for self-gravity to make it nearly round, and has cleared the neighbourhood around its orbit. Pluto became a dwarf planet because of the third, dynamical, criterion (NASA, What Is a Planet?). Merely crossing another orbit is not the reason, since stable resonances allow crossing orbits.

Brightness is a discovery observable, not a size measurement. The 2005 discovery paper argued that Eris must be larger than Pluto. On 6 November 2010 a stellar occultation, the passage of Eris in front of a star, gave a radius of $1163 \pm 6$ km and a very high albedo. The same paper states that Eris is 27% more massive than Pluto (Sicardy et al. 2011). Optical brightness fixes $pD^2$. Separating $p$ from $D$ took an occultation (Chapter 25), as Hubble imaging had done for Quaoar, whose albedo also proved higher than the canonical dark value (Brown & Trujillo 2004).

Two nights instead of one: Palomar 2007–2008

The cure for the Eris problem is a longer baseline. Megan Schwamb, Brown, Rabinowitz and Darin Ragozzine extended the Palomar cadence to two nights (Schwamb et al. 2010). From 8 May 2007 to 27 September 2008 they covered 11,786 square degrees, mostly within 30 degrees of the ecliptic, to a mean image limit of $R = 21.3$. Each field received two 240-second exposures about an hour apart on each of two nights, the second normally the next night and never more than four nights later. Fields lay within 42 degrees of opposition and avoided Galactic latitudes below 15 degrees.

The pipeline, written in IDL, ran SExtractor on each CCD. Typical detection settings required four contiguous pixels above $1.2\sigma$, and deeper auxiliary catalogues with three pixels above $1.1\sigma$ identified faint stationary sources. These are pixel thresholds, not integrated signal-to-noise. The median relative astrometric error was 0.076 arcseconds, much smaller than the 0.4-arcsecond median absolute error, and relative precision is what matters for detecting motion. Stationary sources were rejected by comparing the two nights. Pairs within a night could move up to 14.4 arcseconds per hour. Linked pairs from the two nights had to be separated by more than 4.38 arcseconds, show the expected retrograde direction, and agree in flux and velocity. Bernstein–Khushalani orbit fits then had to satisfy $\chi^2 \lt 25$ with a barycentric distance between 15 and 1,000 au. The nominal slow reach was about 0.2 arcseconds per hour, more than seven times slower than the Eris-era cut.

The team inspected 39,110 candidates by eye, about 200 a night, and found 52 KBOs and Centaurs, 25 of them new, with 96% of the detected KBOs eventually observed at more than one opposition. Sedna was recovered. No other Sedna-like body with a perihelion beyond 45 au turned up.

The efficiency was measured from known objects. The bright-object plateau was only about 0.66, with a rollover at $V = 21.5$ and a width of 0.05 magnitudes. Of 19 known objects brighter than $V = 21$ in the footprint, 13 were found, a 51–89% one-sigma range. Eris itself was missed: it lay in a gap between CCDs. Separately, of 781,763 synthetic distant orbits passed through the orbit-fitting stage, about 5% failed it, mostly because the fit looked like an asteroid’s. That tests one stage. It does not mean the survey was 95% complete, since flux losses, blends, gaps and masks enter separately (Schwamb et al. 2010).

Fitting orbits to short arcs

The orbit fitter used by DES, Sedna’s discoverers and the Palomar distant survey was written by Gary Bernstein and Bharat Khushalani (Bernstein & Khushalani 2000). It uses a parameterisation suited to distant bodies observed over short arcs: angular position, angular velocity, inverse distance and a line-of-sight velocity term, in which the poorly constrained combinations are explicit. It propagates covariances and treats the degeneracies of short arcs specially, including an optional prior that the orbit be bound. The original implementation is portable C, with programs fit_radec, abg_to_aei and predict. The authors report about one second to fit 275 simulated Pluto observations spanning 14 years, and a few seconds for all of the roughly 300 KBOs then known, on a contemporary desktop.

That timing makes a point about where the costs were. Orbit fitting was cheap. Candidate generation, human review and the observing time needed to build a useful arc were the expensive parts. The bound-orbit prior also illustrates the danger of regularisation. It stabilises a fit that the data do not constrain, and can bias the result, as the Sedna precovery showed. Chapter 13 develops orbit determination for modern surveys.

Landmarks

What the era left unsolved

The slow end is dominated by the stationary sky. Seeing changes, refraction, blends, centroiding errors and threshold flicker turn stationary sources into apparent movers. The early Palomar slow candidates all failed, and the Eris search accepted a large increase in contamination. Removing every candidate near a catalogued star controls contamination but throws away real bodies that pass near stars, and that loss depends on where in the sky one looks.

Depth does not give an orbit. The faintest pencil-beam detections were hard to recover. A same-night rate says much about distance near opposition and very little about radial velocity. Bound-orbit priors regularise the fit and can bias it.

Recovery loss changes the population. Wrong ephemerides, weather, lack of telescope time, changing brightness and detector gaps decide whether a detection becomes an orbit. CFEPS’s block revisits reduced the dependence on an assumed orbit and then measured the remaining magnitude dependence. That is a methodological improvement, not proof of an unbiased sample.

Nominal coverage hides holes. Eris fell in a chip gap in the 2007–2008 survey despite being one of the brightest bodies in its footprint. Synthetic geometry tests in that survey found losses as large as about 10% in some latitude ranges, separate from source-detection efficiency.

Population fits can absorb calibration errors. DES fitted luminosity function and sensitivity together and found a large offset from its nominal frame limits. CFEPS emphasised measured efficiency and a conservative cutoff. Neither licenses a single luminosity-function slope across all dynamical classes, magnitudes and albedos. Part of the problem is identifiability, not just error bars.

Sizes need other observations. Optical detection fixes $pD^2$. Thermal measurements, occultations, resolved imaging and satellites each supply different constraints (Chapter 24, Chapter 25).

The lessons persist. A 2025 discovery of a body nicknamed Ammonite, with a secure perihelion of $65.9 \pm 0.2$ au from an arc of about ten years, repeated Sedna’s pattern of detection followed by archival orbit reconstruction (Chen et al. 2025). A September 2026 JWST paper reported 27 TNOs to a 40% threshold of $m_{\rm F150W2} = 28.8$ in 0.05 square degrees, found with shifted stacking and machine-assisted vetting, from observations made in 2023 (Eduardo et al. 2026). Its infrared magnitudes cannot be placed on the same scale as the $R$ depths above without assuming a colour. What carries over is multi-epoch confirmation, source injection and characterised selection. The modern surveys are the subject of Chapter 6.

Where the room is

The chief methodological gap this era exposed is incomplete accounting across the whole chain from photons to an orbit-classified sample. A detector with excellent performance at the source level can still produce biased science if its catalogue gates or recovery scheduling quietly reject the unusual cases. With that in mind, several concrete opportunities stand out. All are assessments, not measured results.

Keep the information until motion is tested. Every survey in this chapter except the pencil beams thresholded single images before testing motion. A modern reimplementation should carry PSF-filtered flux and variance through the motion test instead of requiring a significant detection in each exposure. Ideally that gains $1.25\log_{10}J$ magnitudes for $J$ exposures, before paying for the extra trials. A GPU can compute each image’s matched-filter map once and sample it along many tracks. For the QUEST camera, a four-image set is about 1.29 GB of raw 16-bit pixels (an estimate from the 161-megapixel format), a convenient unit for one modern accelerator, although keeping several filtered versions multiplies the memory. Direct stacking still costs $O(PJN_v)$, but calibration, filtering and interpolation can be fused to cut memory traffic (Chapter 40).

Reuse neighbouring trajectory sums. For a velocity domain of area $A_v$, a baseline $T$, a seeing width $\theta$ and an allowed fractional mismatch $\eta$, the number of trial velocities needed is roughly

As a rough illustration, a one-dimensional search from 1 to 10 arcseconds per hour over a three-hour baseline, with a two-arcsecond image and $\eta = 0.5$, needs about $9\times3/(0.5\times2) = 27$ rate steps in each direction searched. Lengthening the baseline removes the minimum-rate barrier that cost Eris, but the template count grows as $T^2$. Tree sums, hierarchical combination and fast transforms share partial sums among nearby tracks (Chapter 21, Chapter 39). They must be judged at equal recovery efficiency as a function of brightness, rate, seeing, crowding and distance from the edges of the rate window. For bright-source triplet linking, a dense trajectory transform may cost more than a sparse catalogue search. GPUs pay off where many faint-track hypotheses share pixels, or where injection tests require the pipeline to be run many times.

Spend rejection effort where the false sky is. The Eris case suggests a cascade: a broad motion hypothesis over the full baseline, forced photometry at the predicted middle-epoch position, and then a joint scene-plus-trajectory fit for the few survivors. Replacing a hard veto near catalogued stars with a model of a stationary plus a moving source would recover chance overlaps. A cheap score can rank, but a calibrated score should decide. Learned artefact classifiers can cut the human load, but injected moving sources must pass through every stage including the classifier, and validation must separate fields and conditions and include objects near stars, chip edges and unusual rates. A classifier that confidently rejects low rates can reproduce the Eris miss on a larger scale (Chapter 22).

Optimise the whole experiment. Observing strategy can be optimised for expected information about orbital class per unit of exposure time, rewarding recoverable arcs and constraints on distance and radial velocity, and spreading probability over plausible orbits instead of one best fit. CFEPS’s block revisits are a tested alternative to brittle targeted follow-up. Archive indexing can supply long baselines cheaply where old exposures exist, as Sedna showed.

Which limits are hard. Photon statistics, illumination geometry, finite resolution and unobserved epochs are physical. Readout overhead, memory traffic, source-list thresholds, restricted motion banks, brittle masks and the capacity of human review are engineering. Confusion lies between them: a foreground star adds irreducible shot noise, but a better scene model avoids discarding the location entirely. A chip gap holds no recoverable flux, so only scheduling or archival coverage helps there, not faster arithmetic. Chapter 42 compares these limits across the book.

References

1. Baltay, C. et al. (2007). The QUEST Large Area CCD Camera. Publications of the Astronomical Society of the Pacific 119, 1278. arXiv:astro-ph/0702590
2. Bernstein, G., Khushalani, B. (2000). Orbit Fitting and Uncertainties for Kuiper Belt Objects. Astronomical Journal 120, 3323–3332. arXiv:astro-ph/0008348
3. Brown, M. E. (accessed 2026). How was Eris found? (participant account). Caltech web page. web.gps.caltech.edu
4. Brown, M. E., Trujillo, C. A. (2004). Direct Measurement of the Size of the Large Kuiper Belt Object (50000) Quaoar. Astronomical Journal 127, 2413–2417. doi:10.1086/382513
5. Brown, M. E., Trujillo, C., Rabinowitz, D. (2004). Discovery of a Candidate Inner Oort Cloud Planetoid. Astrophysical Journal 617, 645–649. arXiv:astro-ph/0404456
6. Brown, M. E., Trujillo, C. A., Rabinowitz, D. L. (2005). Discovery of a Planetary-sized Object in the Scattered Kuiper Belt. Astrophysical Journal Letters 635, L97–L100. arXiv:astro-ph/0508633
7. Buie, M. W. et al. (2003). Procedures, resources and selected results of the Deep Ecliptic Survey. Earth, Moon, and Planets 92, 113–124. full text
8. CFEPS collaboration (accessed 2026). Canada–France Ecliptic Plane Survey project pages. Project web site. cfeps.net
9. Chen, Y.-T., Lykawka, P. S., Huang, Y. et al. (2025). Discovery and dynamics of a Sedna-like object with a perihelion of 66 au. Nature Astronomy. arXiv:2508.02162
10. Eduardo, M. R., Morgan, A. N., Fraser, W. C. et al. (2026). The Luminosity Function of Ultra-Faint Trans-Neptunian Objects Detected by JWST. arXiv preprint. arXiv:2609.09044
11. Elliot, J. L. et al. (2005). The Deep Ecliptic Survey: A Search for Kuiper Belt Objects and Centaurs. II. Dynamical Classification, the Kuiper Belt Plane, and the Core Population. Astronomical Journal 129, 1117–1162. doi:10.1086/427395
12. Gladman, B., Kavelaars, J. J., Nicholson, P. D., Loredo, T. J., Burns, J. A. (1998). Pencil-Beam Surveys for Faint Trans-Neptunian Objects. Astronomical Journal 116, 2042–2054. arXiv:astro-ph/9806344
13. Jewitt, D., Luu, J. (1993). Discovery of the candidate Kuiper belt object 1992 QB1. Nature 362, 730–732. doi:10.1038/362730a0
14. Jewitt, D. C., Luu, J. X. (1995). The Solar System Beyond Neptune. Astronomical Journal 109, 1867–1876. ADS scan
15. Jones, R. L. et al. (2006). The CFEPS Kuiper Belt Survey: Strategy and presurvey results. Icarus 185, 508–522. arXiv:astro-ph/0510826
16. Luu, J., Marsden, B. G., Jewitt, D., Trujillo, C. A., Hergenrother, C. W., Chen, J., Offutt, W. B. (1997). A new dynamical class of object in the outer Solar System. Nature 387, 573–575. doi:10.1038/42413
17. Millis, R. L., Buie, M. W., Wasserman, L. H., Elliot, J. L., Kern, S. D., Wagner, R. M. (2002). The Deep Ecliptic Survey: A Search for Kuiper Belt Objects and Centaurs. I. Description of Methods and Initial Results. Astronomical Journal 123, 2083–2109. full text
18. Minor Planet Center (2005). MPEC 2005-O36: 2003 EL61. Minor Planet Electronic Circular. minorplanetcenter.net
19. Minor Planet Center (2005). MPEC 2005-O42: 2005 FY9. Minor Planet Electronic Circular. minorplanetcenter.net
20. NASA Science (accessed 2026). Haumea: Facts. NASA. science.nasa.gov
21. NASA Science (accessed 2026). What Is a Planet? (reproducing IAU Resolution B5). NASA. science.nasa.gov
22. OSSOS collaboration (accessed 2026). SurveySimulator. GitHub repository. github.com/OSSOS/SurveySimulator
23. Petit, J.-M. et al. (2011). The Canada-France Ecliptic Plane Survey—Full Data Release: The Orbital Structure of the Kuiper Belt. Astronomical Journal 142, 131. arXiv:1108.4836
24. Schwamb, M. E., Brown, M. E., Rabinowitz, D. L., Ragozzine, D. (2010). Properties of the Distant Kuiper Belt: Results from the Palomar Distant Solar System Survey. Astrophysical Journal 720, 1691–1707. arXiv:1007.2954
25. Sicardy, B. et al. (2011). A Pluto-like radius and a high albedo for the dwarf planet Eris from an occultation. Nature 478, 493–496. doi:10.1038/nature10550
26. Trujillo, C. A., Brown, M. E. (2003). The Caltech Wide Area Sky Survey: Beyond (50000) Quaoar. Earth, Moon, and Planets 92, 99–112. doi:10.1023/B:MOON.0000031929.19729.a1
27. Trujillo, C. A., Jewitt, D. C., Luu, J. X. (2001). Properties of the Trans-Neptunian Belt: Statistics from the CFHT Survey. Astronomical Journal 122, 457–473. arXiv:astro-ph/0104104