JWST Pixels Hid Four Close Little Red Dot Pairs

Four close pairs of Little Red Dots sit 1 to 7 kiloparsecs apart in James Webb Space Telescope images of the early universe, a new study finds. Older catalogs treated some of them as single objects because the sources failed a compactness cut.

Takumi S. Tanaka, a University of Tokyo graduate student at the Kavli Institute for the Physics and Mathematics of the Universe, with Professor John D. Silverman and colleagues, published the result on August 31. If each dot holds a growing black hole, the pairs are plausible precursors of mergers.

Compactness Cuts Hid Close Pairs as Single Dots

Little Red Dots, or LRDs, are compact, extremely red sources that JWST began finding in 2022. Most searches picked them with a red color cut and a compactness test meant to keep extended galaxies out of the sample.

A common color cut is m_F277W minus m_F444W greater than 1.5. Hollis B. Akins and colleagues, who are coauthors here, also required that more than half the F444W light inside 0.5 arcsec sit inside 0.2 arcsec. That ratio favors isolated “dots.”

A neighbor closer than about 0.5 arcsec, or about 2 to 3 kpc at redshift 5 to 7, can dilute that ratio and scramble the blended color, especially if the companion is bluer. Tanaka’s group built a pixel-by-pixel color selection method and dropped the compactness demand so close pairs could survive.

HOW THE OLD CUTS MISSED PAIRS

• Color blend: Aperture photometry mixes two sources and can push a true red core below the F277W minus F444W cut of 1.5.
• Compactness fail: Light from a neighbor inside 0.5 arcsec lowers the 0.2-to-0.5 arcsec flux ratio below 0.5, so the object is tossed as extended.
• Pixel rescue: Coloring each pixel, then keeping clumps of at least ten adjacent LRD-like pixels, lets two nearby cores count as two objects.

Machine-learning cleaning removed diffraction spikes and saturated stars, after which visual inspection of the cleaned map turned up four dual candidates. No system held more than two LRDs. Catalogs built to find a single compact dot will, by design, throw out the systems that look like two dots about to meet.

Four COSMOS-Web Pairs Sit 1 to 7 kpc Apart

The pairs come from COSMOS-Web, JWST’s 255 hour Cycle 1 treasury program led by Jeyhan Kartaltepe and Caitlin Casey. Tanaka’s team modeled the search field as 0.54 square degrees. In each pair the brighter core is labeled #1 and the fainter core #2. One kiloparsec is 3,260 light-years.

THE FOUR DUAL LRD CANDIDATES

Those physical distances run from about 5,350 light-years for CW-B2-4383 to about 24,000 light-years for CW-A4-16093, still a fraction of the Milky Way’s roughly 100,000-light-year disk. The light left those systems 12.5 to 12.8 billion years ago.

Matching H-Alpha Lines Lock Two Pairs at One Redshift

COSMOS-3D slitless spectroscopy in F444W covers CW-B2-4383 and CW-A4-16093. Each component shows a line at the same observed wavelength. Tanaka’s group rejects C IV, Mg II, and Paschen lines because those redshifts clash with the photometry, and they reject H-beta because the usual [O III] doublet is absent, leaving H-alpha.

For CW-B2-4383 the two H-alpha redshifts are 5.8222 and 5.8219, a line-of-sight speed offset of 13 km/s. For CW-A4-16093 they are 5.4639 and 5.4649, an offset of 47 km/s. Both offsets sit well below 100 km/s, which the paper finds smaller than the escape speed of a typical dark-matter halo that would host an LRD, so the two spectroscopic pairs are likely bound.

The brighter cores show broad H-alpha. Single-Gaussian full widths are 1,118 km/s in CW-B2-4383 #1 and 1,123 km/s in CW-A4-16093 #1, in line with the LRD hallmark of widths above 1,000 km/s. The fainter cores are weaker; limited signal may hide a broad component.

Photometry Alone Still Carries Two More Pairs

CW-B5-15958 and CW-A6-19978 lack grism spectra. Their photometric redshifts still match inside each pair, and both sit on 1 to 2 kpc scales. CW-B5-15958 lies behind a foreground type-1 AGN, CID-643, which the team subtracted before measuring the two red cores. Several of the systems also show a faint off-center blob or a companion galaxy that is not itself an LRD.

Random Alignments Fail a 10,000-Mock Test

Two red dots can sit near each other on the sky and still lie at different distances. Tanaka’s group tested that with the cleaned parent sample of 829 LRD-like objects, kept at signal-to-noise above 10 in F444W after brown-dwarf and dusty-galaxy cuts.

THE MOCK-SKY TEST

• Four close pairs: The chance of drawing four pairs inside 1.2 arcsec from a random map is 7 in 100,000, or 3.8 sigma.
• Three tight pairs: The chance of three pairs inside 0.4 arcsec is 4 in 1 million, or 4.5 sigma.
• Mock size: Each of 10,000 random catalogs had 829 sources scattered across a 0.54-square-degree square.
• Clustering excess: The LRD angular auto-correlation sits about 20 to 30 times above a power-law fit to JWST active galactic nuclei measured on 10 to 100 arcsec scales.

On wide scales the observed pair histogram matches the random mocks, so LRDs are not globally clumped in this field. The excess is confined to kilo-parsec separations. Other JWST programs have already found dual active nuclei at a rate of roughly 10 to 30 percent, but dual LRDs had not turned up until compactness was relaxed. After this paper was first posted, R. M. Mérida and colleagues in 2025 and H. Yanagisawa and colleagues in 2026 reported photometric close LRD pairs in lensed fields.

What Little Red Dots Are

Little Red Dots emerge around 600 million years after the Big Bang and fall off around 1.5 billion years, according to a NASA and ESA/Webb compilation from CEERS, JADES, NGDEEP, and related surveys. Spectra from the RUBIES program suggest many are accreting black holes, though the class is still mixed.

They show a V-shaped spectrum, blue in rest-frame ultraviolet and red in rest-frame optical, with the valley near the Balmer limit. Broad Balmer lines sit with a Balmer break and, often, narrow absorption. X-ray, radio, and hot-dust torus light are weak or missing, which is why they do not behave like ordinary quasars.

WHERE EXPERTS DISAGREE

• Black hole star: A young black hole sits in a dense gas envelope, n_H above about 10⁹ cm^-3, and the envelope, not a giant stellar bulge, makes the red optical light.
• Line-width mass: Broad H-alpha treated as a single-epoch AGN line gives M_BH,Hα of 10^7.1 to 10^7.3 solar masses for the two bright spectroscopic cores, an order of magnitude above the envelope fit.
• Envelope mass: Blackbody fits at Eddington accretion give M_BH,BB of 10^5.3 to 10^6.3 solar masses, including 10^5.59 for CW-B2-4383 #1 and 10^6.30 for CW-A4-16093 #1.

Tanaka’s group does not pick a mass scale. Either way, the pairs point to black holes of about 10⁵ to 10⁷ solar masses sharing kilo-parsec real estate while the universe was a little over a billion years old. A stacked COSMOS-Web study of 217 LRDs has also found faint extended light on ~200-parsec scales, so some “dots” sit inside tiny host galaxies rather than in empty space.

LISA Could Hear These Mergers if the Black Holes Coalesce

Galaxy mergers can drive gas inward and feed a central black hole. If both members of an LRD pair host a growing hole, the holes themselves may later merge. Tanaka’s group writes that such events belong in the band of ESA’s Laser Interferometer Space Antenna, planned as the first space-based gravitational wave observatory, with a launch around 2035.

Our sample is likely to represent precursors of mergers between LRDs, and such mergers may be one of the mechanisms that can drive the rapid growth of SMBHs in their early evolutionary stages.

Takumi S. Tanaka and colleagues, Publications of the Astronomical Society of Japan

LISA’s three spacecraft will fly in a triangle about 50 million km from Earth, with laser arms about 2.5 million km long and a four-year prime mission. The LISA Consortium says the observatory will catch coalescences of black holes from 10,000 to 10 million solar masses as early as redshift ~10, including holes that are not shining as quasars. Comparing future detection rates with the two LRD mass estimates would test which mass scale is right. A raw dual fraction of four pairs in 829 objects is 0.5%, with Poisson errors of +0.4 and -0.2 percentage points, but the team warns that fraction likely changes with luminosity and completeness.

The COSMOS-Web Catalog Grows Once Dots Can Have Neighbors

Akins’s aperture catalog on the same color cut found 434 COSMOS-Web LRDs. Tanaka’s pixel sample, which keeps close companions and fainter cores, reaches 829 objects, approximately twice as large. Paper II, listed as in preparation, will treat that parent catalog in full.

The extra objects are not a rounding error. They are the LRDs that looked messy because something sat next door. If mergers help black holes gain mass in the first billion years, those messy systems are the ones that matter, and they were the ones the compactness cut was built to discard.

Two Pairs Still Need Deeper Spectra

CW-B5-15958 and CW-A6-19978 still rest on photometry. Even the two grism pairs show only a single line, and the fainter cores lack a clear broad component. The paper calls for deeper spectroscopy before anyone treats the dual fraction as a merger rate.

WHAT WE KNOW

• Four candidates: Two have matching H-alpha at z = 5.822 and z = 5.464; two have consistent photometric redshifts near 6.4.
• Not random: Four pairs inside 1.2 arcsec are a 7-in-100,000 fluke on a random sky of 829 sources.

WHAT IS UNCONFIRMED

• True black hole binaries: Envelope masses and H-alpha masses still differ by about a factor of ten to a hundred.
• Merger rate: The 0.5% dual fraction is uncorrected for luminosity and completeness, so it is not yet a gravitational-wave forecast.

Tanaka’s group says the next step is more spectra and a wider area. Until those data arrive, the four pairs sit in the pixels as systems the isolated-dot cut was never going to count.

Frequently Asked Questions

When Did Astronomers First Notice Little Red Dots?

JWST’s first science images in 2022 were already sprinkled with tiny red sources; Jorryt Matthee and colleagues named the class “Little Red Dots” in 2024. The objects were unknown to Hubble because they are faint in optical light and bright at 2 to 5 microns, where Webb’s NIRCam is built to work.

What Does a Typical Little Red Dot Spectrum Look Like?

The rest-frame energy distribution is V-shaped, with a blue ultraviolet side, a red optical side, and a trough near the Balmer limit. Many also show a Balmer break and narrow Balmer absorption on top of broad emission, while X-ray, radio, and warm-dust torus light stay weak or absent compared with ordinary quasars of similar optical brightness.

What Is COSMOS-Web Designed to Do?

COSMOS-Web is JWST program 1727, a Cycle 1 treasury survey whose three stated goals are mapping reionization on large scales, taking a census of massive galaxies in the first two billion years, and measuring how stellar mass sits inside dark-matter halos. The team expects the mosaic to contain about a million galaxies across cosmic time, which is why it is now the widest hunting ground for rare LRD pairs.

Which Black Hole Mergers Can LISA Hear?

LISA is built to record coalescences of black holes from 10,000 to 10 million solar masses, a range that overlaps both LRD mass estimates in this paper, and to do so across most of cosmic time, including epochs near redshift 10. Ground detectors such as LIGO and Virgo hear stellar-mass holes at much higher frequencies and cannot reach this mass window.

Did the Search Find Any Triple Little Red Dots?

No. Tanaka’s visual inspection found many LRDs sitting next to ordinary galaxies, but zero systems with three or more LRD-like cores. The dual sample is four pairs, and the paper treats that ceiling as a real outcome of the search rather than a cut they imposed.