The Dance of the Little Red Dots: Spectroscopy of Two Blue Companions at z = 3.9 and 5.5 Reveals Young, Metal-Poor Starbursts Merging with Their LRDs within a Few Hundred Myr

abstract · abstract

The Little Red Dots (LRDs) are compact, red sources, abundant at $4 \lesssim z \lesssim 9$ and likely powered by massive black holes. Photometric studies have found blue companions around them, but none has yet been confirmed spectroscopically. This Letter presents the only two LRD-companion pairs with spectroscopy of both members in the literature: "Waltz" at $z = 3.9$ and "Tango" at $z = 5.5$, with projected separations of $3.6$ and $2.5$ kpc. Both companions lie at the redshift of their LRD, within $132 \, \rm km \, s^{-1}$ and $57 \, \rm km \, s^{-1}$, and are physically identical, but $524$ Myr apart: starbursts $\lesssim 50$ Myr old, metal-poor ($0.2 \, \rm Z_\odot$), dust-free, dynamically cold, and as bright as the LRDs themselves in the rest-frame UV. The LRD in Waltz hosts a $\approx 3 \times 10^{6} \, \rm M_\odot$ black hole accreting at half the Eddington rate, a factor of a few above the birth mass of a direct collapse black hole, and is embedded in gas as metal-poor as its companion. Orbital modeling shows that Tango is bound in $81\%$ of the allowed orbits and that Waltz is bound if its host halo exceeds $5 \times 10^{10} \, \rm M_\odot$. Both blue companions then merge with their respective LRD within a few hundred Myr. We conclude that we are observing what heavy seeding requires: a young black hole beside the metal-poor starburst that the synchronized-pair scenario predicts, in the last steps of a cosmic dance that ends in a merger, and in a galaxy seeded by its black hole.

S4: IV Summary and Conclusions / block 1 · html

This study was motivated by observations [ 95 , 4 , 7 ] and a statistical analysis by [ 87 ] , which found that the brightest LRDs host blue companions within 2 2 kpc, 2.39 2.39 times more often than matched galaxies (significant at 3.3 ​ σ 3.3\sigma ), and that those companions are bluer than the companions of ordinary galaxies. That excess can be interpreted as the geometry of the “synchronized pair” [ 27 , 109 , 91 ] , in which a young star-forming halo irradiates a pristine neighbor and enables the direct collapse of a heavy black hole seed. Photometry, however, left open questions that only spectroscopy can address. For example, whether (i) the companions are physical, (ii) their blue light is starlight or radiation reprocessed from the LRD, and (iii) the pairs are gravitationally bound.

S4: IV Summary and Conclusions / block 1 · html

Both companions lie at the redshift of their LRD, offset by 132 132 and 57 ​ km ​ s − 1 57\,\rm km\,s^{-1} at projected separations of 3.6 3.6 and 2.5 2.5 kpc.

S4: IV Summary and Conclusions / block 1 · html

Tango is bound in 81 % 81\% of the orbital realizations; Waltz is bound if its host halo exceeds 5 × 10 10 ​ M ⊙ 5\times 10^{10}\,{\rm M_{\odot}} , so that its binding is a dynamical constraint on an LRD host. Once bound, the companions merge with their LRDs within a few hundred Myr (modes of 22 22 Myr for Tango and 231 231 Myr for Waltz).

S4: IV Summary and Conclusions / block 1 · html

If the seed formed at z ≳ 10 z\gtrsim 10 , its halo has had a billion years to grow, and we expect a mass of 5 × 10 10 ​ M ⊙ 5\times 10^{10}\,{\rm M_{\odot}} at z = 3.9 z=3.9 . If instead the seed formed in the last 100 − 200 100-200 Myr, its halo would have just crossed the atomic-cooling threshold, 2 − 4 × 10 8 ​ M ⊙ 2-4\times 10^{8}\,{\rm M_{\odot}} at z = 4 − 6 z=4-6 , and could not have reached 5 × 10 10 ​ M ⊙ 5\times 10^{10}\,{\rm M_{\odot}} since. For Waltz, the two readings thus part ways: if the pair is bound, the seed is old; if the system is young, the pair is unbound, which its 48 % 48\% bound probability allows. An old seed must then have grown slowly. Reaching 3 × 10 6 ​ M ⊙ 3\times 10^{6}\,{\rm M_{\odot}} takes 100 − 520 100-520 Myr at the present rate for seeds of 10 6 − 10 4 ​ M ⊙ 10^{6}-10^{4}\,{\rm M_{\odot}} , against the ∼ 1 \sim 1 Gyr elapsed since z ≈ 10 z\approx 10 , so the black hole was active for 10 − 50 % 10-50\% of its life, as the intermittent, radiation-regulated accretion of DCBHs predicts [ 82 ] . Tango is less constraining: binding requires only 10 9 ​ M ⊙ 10^{9}\,{\rm M_{\odot}} inside 2.5 kpc (Table 1 ), which an atomic-cooling halo alone could not supply but ∼ 10 9 ​ M ⊙ \sim 10^{9}\,{\rm M_{\odot}} of baryons in the LRD host could, so both interpretations remain open there. The young reading, where it survives, requires a near-pristine atomic-cooling halo beside a star-forming one at z = 4 − 6 z=4-6 .

S1: I Introduction / block 1 · html

We describe the observations and the measurements in § II , and present our results in § III . In § IV , we discuss what these systems imply for the nature of LRDs. Throughout, we adopt the Planck18 cosmology [ 89 ] , where 1 ′′ 1^{\prime\prime} corresponds to 7.15 7.15 and 6.13 6.13 proper kpc at z ≈ 3.9 z\approx 3.9 and z ≈ 5.5 z\approx 5.5 , respectively.

S3: III Results / block 1 · html

We present the two pairs, Tango and Waltz, in parallel throughout this Section. Table 1 collects all measured quantities and their uncertainties for both systems, while the main text discusses only the relevant numbers.

S4: IV Summary and Conclusions / block 1 · html

This Letter presents the only two LRD–companion pairs reported in the literature with spectra of both members: Waltz, at z = 3.936 z=3.936 , observed with the NIRSpec MSA, and Tango, at z = 5.4816 z=5.4816 , observed with the NIRSpec IFU. Our key findings are the following.

S4: IV Summary and Conclusions / block 1 · html

The two companions are identical, but separated by 524 524 Myr of cosmic time: starbursts ≲ 50 \lesssim 50 Myr old, metal-poor ( 0.2 ​ Z ⊙ 0.2\,\rm Z_{\odot} ), dust-free, dynamically cold ( σ int = 20 − 65 ​ km ​ s − 1 \sigma_{\rm int}=20-65\,\rm km\,s^{-1} ), of 10 7 − 10 8 ​ M ⊙ 10^{7}-10^{8}\,{\rm M_{\odot}} , and as bright as their LRDs in the rest-frame UV.

S4: IV Summary and Conclusions / block 1 · html

The LRD does not power the companion’s light.

S4: IV Lyman-Werner Radiation from the Companions / block 1 · html

The above estimates for the LW radiation field do not account for self-shielding effects [ 134 ] due to a column of gas between the blue and red components. We test whether this affects our conclusions by estimating the timescale for molecular hydrogen dissociation in this column. For a red component exposed to an intrinsic LW radiation field J 21 , LW J_{\rm 21,LW} produced by its companion(s), the corresponding LW photon production rate of the companions can be written as N ˙ LW = 3.5 × 10 48 ​ ( 4 ​ π ) ​ ( d eff / kpc ) 2 ​ J 21 , LW \dot{N}_{\rm LW}{=}3.5\times 10^{48}(4\pi)(d_{\rm eff}/\rm kpc)^{2}J_{\rm 21,LW} photons s -1 . Assuming the LW photons dissociate H 2 \rm H_{2} with 10% efficiency, we can estimate the propagation of the H 2 dissociation front through a sphere with radius d eff d_{\rm eff} , following equation (2) of Sullivan et al. [117] , as

S6: VI Conclusions / block 1 · html

In this work, we have identified candidate UV-bright companions to a substantial fraction of LRDs (43% increasing to 72% when including tentative companions), assumed a physical connection between the components, and used their luminosities and projected separations to infer the local Lyman-Werner radiation fields experienced by the compact red sources. The inferred radiation fields are sufficiently intense to plausibly suppress molecular cooling and enable rapid collapse to very high densities, in line with established theoretical models. While this agreement does not constitute a proof of the scenario, it demonstrates that the picture is internally consistent and observationally viable.

S3.SS6: III.6 Ly α \alpha Emission Estimate / block 1 · html

where we assume an escape fraction f esc = 5 % f_{\rm esc}=5\% following [ 100 ] and a characteristic duration of ionizing photon production t age = 30 t_{\mathrm{age}}=30\, Myr according to the median stellar age from our SED fitting. At the level of this order-of-magnitude estimate, the resulting bubble size depends only weakly on these parameters. We calculate n ¯ H \bar{n}_{\rm H} according to the mean cosmological hydrogen number density at redshift z z . The total ionizing photon production rate N ˙ ion \dot{N}_{\rm ion} is estimated by summing the contributions from all galaxies within the overdensity, inferred from their M UV M_{\rm UV} and β \beta using the prescription of Mason and Gronke [63] . This estimate should be regarded as a lower limit, as our sample includes only relatively bright galaxies, while faint sources are expected to contribute additional ionizing photons.

S3.SS1: III.1 Measurement / block 2 · html

m the galaxies and their subcomponents, and because we adopt a different SED modeling framework. The photometric redshifts inferred with Prospector are slightly lower than EAZY results due to the difference in Ly α \alpha modeling.

S1: I Introduction / block 1 · html

In this Letter, we compare a sample of 253 253 LRDs to a carefully matched sample of galaxies, constructed as described in § II . Our results are described in § III and discussed in § IV . Throughout this Letter, we adopt a flat Λ \Lambda CDM cosmology, with H 0 = 70 ​ km ​ s − 1 ​ Mpc − 1 H_{0}=70\rm\,km\,s^{-1}\,Mpc^{-1} and Ω m = 0.3 \Omega_{m}=0.3 ; across the redshift range of our sample, 1 ′′ 1^{\prime\prime} corresponds to 4.6 − 7.0 4.6-7.0 kpc of projected separation.

S3: III Results / block 1 · html

Figure 2: Cumulative fraction of the brightest LRD quartile hosting a blue companion as a function of projected separation (red; shading is the 1 ​ σ 1\sigma confidence interval) and their Tier-1 control galaxies (gray). Both curves are empty below ≈ 0.7 \approx 0.7 kpc: the 0 ​ . ′′ ​ 15 0\farcs 15 resolution floor corresponds to ≈ 0.7 \approx 0.7 kpc at z = 8.5 z=8.5 , comparable to the smallest separation recorded in the sample, 0.713 0.713 kpc. The companion fraction rises steeply at < 2 <2 kpc, where 18 / 65 18/65 ( 27.7 % 27.7\% ) of the brightest quartile host a companion, a × 2.39 \times 2.39 excess over controls, significant at 3.3 ​ σ 3.3\sigma . Beyond the close-pair window, the two populations accumulate companions at comparable rates, and the cumulative ratio declines to × 1.79 \times 1.79 by 5 5 kpc. When measured in the 2 − 5 2-5 kpc annulus against Tier-2 controls, the ratio is × 1.04 \times 1.04 : at larger separations, the surroundings of the bright LRDs are indistinguishable from those of ordinary compact galaxies of the same brightness.

S5: 5 Conclusions / block 1 · html

Empirical M BH M_{\rm BH} measurements from broad H α \alpha emission ( FWHM ≳ 4000 \mathrm{FWHM}\gtrsim 4000 km s -1 ) place our [NeV]-detected QGs on the local M BH M_{\rm BH} – M ⋆ M_{\star} scaling relation. The bolometric luminosities inferred from [NeV] and [O III ] reach quasar-like values ( L bol = 10 45 − 46 L_{\rm bol}=10^{45-46} erg s -1 ), with Eddington ratios of λ Edd ≈ 10 − 50 \lambda_{\rm Edd}\approx 10-50 % and black hole accretion rates of a few M ⊙ M_{\odot} yr -1 , comparable to or exceeding the residual star formation rates in the most extreme cases.

S5: 5 Conclusions / block 1 · html

Semi-analytical models ( Shark , Gaea ) and zoom-in simulations ( Black Dawn with the Mistral feedback model) broadly reproduce the observed trend of declining AGN activity with time since quenching, consistent with the physical picture emerging from the observations. At peak post-quenching activity, all models produce Eddington ratios of ≈ 10 − 50 \approx 10-50 % and bolometric luminosities broadly consistent with our observed values.

Sx1: Results / block 2 · html

r} during which the blue peak Lyman- α \alpha photons are still visible, corresponding to a 90 % 90\% decline of the photo-ionisation from its “quasar phase” equilibrium value. Regardless of the exact values assumed above, the timescale to observe the double-peaks in COLA1 and NEPLA4 after a quasar phase is extremely short, of the order of a Myr or less.

Sx1: Results / block 1 · html

We now conclude by discussing the impact of episodic BH growth on the visibility of Lyman- α \alpha in the Reionisation Era. In the HEROES survey of ultra-luminous LAEs from which COLA1 and NEPLA4 were discovered Taylor et al. (2023) , the fraction of observed double-peaks is 3.5 − 1.7 + 3.4 % 3.5^{+3.4}_{-1.7}\% . Low-redshift studies, where the IGM is not absorbing the blue-peak photons, show that the intrinsic fraction of double-peaked LAEs is around ∼ 30 % \sim 30\% Kulas et al. (2012) ; Vitte et al. (2025) . Therefore the fraction of ultra-luminous Lyman- α \alpha emitters whose visibility is temporarily boosted by a previous AGN/quasar accretion phase is ∼ 11 − 3 + 6 % \sim 11^{+6}_{-3}\% . This is a conservative estimate where we have not considered longer timescales where the photo-ionisation would not enable blue-peak photons to escape but the large ionised bubble would still diminish the impact of IGM absorption and boost the Lyman- α \alpha red peak transmission.

S5: V Summary / block 1 · html

Similar to the luminosity ratio, the ratio of BH mass to nuclear stellar mass at the end of LRD phase also converges to a narrow range at each redshift, with M BH / M ⋆ , nuc ≃ 0.1 M_{\rm BH}/M_{\star,\rm nuc}\simeq 0.1 for 5 < z < 10 5<z<10 . Toward lower redshifts ( z < 1 z<1 ), the ratio shifts increasingly toward stellar dominance and eventually approaches the local value of ∼ 0.1 % \sim 0.1\% [ 82 , 132 , 44 , e.g.,] .

S2.SS1: II.1 Rapidly accreting envelope onto a BH / block 1 · html

To maintain a radiation-dominated convective envelope against disruption by AGN feedback, the mass inflow rate from the parent halo must far exceed the Eddington rate (see discussion in Section III.2 ), defined as M ˙ Edd ≡ L Edd / ( η rad ​ c 2 ) \dot{M}_{\rm Edd}\equiv L_{\rm Edd}/(\eta_{\rm rad}c^{2}) with a 10% radiative efficiency η rad = 0.1 \eta_{\rm rad}=0.1 . Such high inflow rates are plausible in protogalactic nuclei hosted by massive dark matter halos with masses of M h M_{\rm h} , in which the baryonic accretion rate with a cosmic baryon fraction of f b = 0.16 f_{\rm b}=0.16 , can reach

S5: 5 Conclusions / block 1 · html

We photometrically identify 22 extremely metal-poor galaxy candidates with Z < 1 % ​ Z ⊙ Z<1\%~\mathrm{Z}_{\odot} at 2.5 < z < 6.5 2.5<z<6.5 , made possible by the extensive deep JADES NIRCam medium-band imaging. EMPGs are identified based off strong photometric boosts by H α \alpha , yet correspondingly weak boosts by [ O iii ] + H β \beta , likely requiring extremely low metallicity to explain the lack of [ O iii ] emission, resulting in characteristically red m [ O ​ III ] − m H ​ α m_{\mathrm{[O\ III]}}-m_{\mathrm{H\alpha}} colours in the wide-band filters covering [ O iii ] and H α \alpha . Medium bands greatly aid in the confident identification of EMPG candidates, also yielding tighter constraints on the inferred gas-phase metallicities. The medium bands disentangle the line and continuum emission, provide constraints on the rest-frame optical/NIR continuum level, establish the presence of Balmer jumps, and tighten the photometric redshift constraints.

S5: 5 Conclusions / block 1 · html

From SED-fitting with close-to-pristine models, we infer our EMPG candidates to be very low-mass dwarfs (median M ∗ = 10 6.7 ​ M ⊙ M_{*}=10^{6.7}~\mathrm{M}_{\odot} ), constituting the extreme low metallicity tail of the mass–metallicity relation. The EMPGs are starburst galaxies, characterised by strong Balmer jumps and high H α \alpha equivalent widths, placed well above the star-forming main sequence. They are > 1 >1 dex more metal-poor than would be expected given their stellar mass and star formation rate, being near-pristine young starburst galaxies, representing an extreme departure from the slowly-evolving equilibrium framework of the fundamental metallicity relation. The EMPG candidates are generally very UV-faint, with median M UV = − 16.6 M_{\mathrm{UV}}=-16.6 , brought into view thanks to the exceptionally deep JADES imaging, though the rare, brighter tail would be detectable with shallower imaging from wider programmes. The EMPG candidates are extremely efficient producers of ionising photons, with median log ⁡ ( ξ ion , obs / ( Hz ​ erg − 1 ) ) = 26.0 \log\,(\xi_{\mathrm{ion,obs}}/\mathrm{(Hz\ erg^{-1})})=26.0 . To power this extreme emission, they possibly harbour top-heavy stellar populations, their exceptionally high ξ ion , obs \xi_{\mathrm{ion,obs}} compatible with being nebular-dominated, indicating a possibly dominant nebular continuum in the rest-frame UV. Despite their UV faintness, the EMPG candidates have intermediate H α \alpha luminosities (median L H ​ α = 10 41.4 L_{\mathrm{H\alpha}}=10^{41.4} erg s -1 ), favouring their detection in deep JWST spectroscopy. EMPGs are currently rare even in the deep JADES imaging, constituting ∼ 0.04 ​ – ​ 0.6 % {\sim}0.04\textrm{--}0.6\% of 2.5 < z < 6.5 2.5<z<6.5 galaxies at − 19 < M UV < − 16 -19<M_{\mathrm{UV}}<-16 . We anticipate that even deeper imaging (likely boosted by strong gravitational lensing) will reveal progressively more EMPG candidates.

S1: I Introduction / block 1 · html

The discovery of broad ( ≳ 1000 \gtrsim 1000 km s -1 ) H α \alpha emission provided a concrete clue for the origin of LRDs [ 44 , 57 , 73 ] . Interpretations invoking accreting black holes quickly came to dominate their discussion [ 57 , 71 , 39 ] . If LRDs are powered by accreting black holes, they are remarkably distinct from other known black hole populations. LRDs rarely have detected X-rays [ 6 , 55 ] , many exhibit Balmer absorption even in H α \alpha [ 73 ] , lack the rest-frame mid-IR colors typical of a dusty torus [ 85 ] , occupy lower mass halos than standard quasars [ 87 ] , and have inferred black-hole-to-stellar-mass ratios approaching unity [ 31 , 71 ] . Individually, these properties can each be explained as black holes forming before the bulk of stars by accreting at super-Eddington rates [ 66 ] with their accretion disks enshrouded by dense neutral gas cocoons [ 51 , 78 ] . Since accreting black holes are expected to evolve strongly with redshift, the extreme accretion physics powering the LRD population was viewed as a natural extension of early black hole formation.

S3.SS1: III.1 The present-day shape of the LRD mass function / block 1 · html

The resulting LRD mass function at z ∼ 7 z\sim 7 is shown as the solid gold curve in Figure 3 . We normalize by the total number of objects to emphasize its shape and enable a comparison to local objects. The inferred LRD initial mass function has a classic Schechter function shape: an exponential cut-off above log \log (M ∗ / M ⊙ M_{\odot} ) ∼ 6 \sim 6 and a steep power-law slope extending to lower masses.