pop-cosmos: Galaxy size evolution across structural and star-formation classifications in COSMOS-Web

Technical Summary: pop-cosmos: Galaxy size evolution across structural and star-formation classifications in COSMOS-Web

Problem and Motivation
Galaxy sizes are known to correlate with stellar mass and redshift, characterized by size scaling relations. However, the inferred forms of these relations are sensitive to how galaxies are classified. Traditionally, studies separate galaxies by star-formation activity (e.g., specific star-formation rate, sSFR) or by morphological markers (e.g., bulge-to-total ratio, Sérsic index), often treating these classifications as equivalent. A key unresolved question is whether these schemes identify the same structural populations or capture different dimensions of galaxy evolution. Furthermore, previous work has relied on rest-frame color selections (e.g., $UVJ$) which can suffer from contamination by dusty star-forming galaxies, and ground-based imaging which may blur structural details. This paper investigates whether sSFR-based and morphology-based classifications yield quantitatively different size–mass and size–redshift relations, and whether these differences are recoverable only with space-based resolution.

Methodology
The authors combine two distinct datasets for the COSMOS field to construct a sample of 99,369 galaxies:

1. Physical Parameters: Stellar mass ($M_*$), redshift ($z$), and sSFR estimates are derived using the pop-cosmos generative model (Thorp et al. 2025). This model is trained on 26-band COSMOS2020 photometry (Weaver et al. 2022) and utilizes a score-based diffusion model to learn the joint distribution of 16 spectral energy distribution (SED) parameters, avoiding restrictive functional forms.
2. Size and Morphology: Half-light radii ($R_{50}$), Sérsic indices ($n$), and bulge-to-total ratios ($B/T$) are obtained from the COSMOS2025 catalogue (Shuntov et al. 2025), based on high-resolution JWST/NIRCam imaging from the COSMOS-Web survey.

The analysis sample is cross-matched with a 0.5" aperture and subjected to quality cuts to remove stars, blended sources, and unreliable size measurements. The sample is restricted to $0.0 < z < 3.0$.

The authors fit size–mass and size–redshift relations using Bayesian inference with Gaussian likelihoods that account for measurement uncertainties and intrinsic scatter ($\sigma_{int}$).

• Star-forming/Disc-dominated/Late-type populations are fitted with a single-power law: $R_{50} = R_0 (M_*/5\times10^{10} M_\odot)^B$.
• Quiescent/Bulge-dominated/Early-type populations are fitted with a double-power law to capture potential breaks: $R_{50} = \gamma (M_*/M_\odot)^\alpha (1 + M_*/10^\delta M_\odot)^{\beta-\alpha}$.
• Size–Redshift relations are modeled as $R_0 = B_z (1+z_m)^{-\beta_z}$.

Classifications are applied using:

• sSFR: Quiescent ($\log_{10}(\text{sSFR}/\text{yr}^{-1}) \le -11$) vs. Star-forming.
• B/T: Disc-dominated ($B/T \le 0.2$), Intermediate, Bulge-dominated ($B/T \ge 0.6$).
• Sérsic Index: Late-type ($n \le 2.5$) vs. Early-type ($n > 2.5$).

Key Results

1. Quantitative Differences in Scaling Relations: The sSFR and morphology splits produce quantitatively different slopes, intercepts, and intrinsic scatter behaviors.

   • Slopes: Star-forming/disc-dominated systems follow shallower slopes ($B \sim 0.12\text{--}0.27$), while quiescent/bulge-dominated systems exhibit steeper high-mass slopes ($B \sim 0.46\text{--}0.62$).
   • Intrinsic Scatter: Crucially, $\sigma_{int}$ depends on structural morphology but not on sSFR. Disc-dominated systems show lower scatter ($\sigma_{int} \sim 0.13\text{--}0.14$ dex) compared to bulge-dominated systems ($\sigma_{int} \sim 0.18\text{--}0.20$ dex). In contrast, star-forming and quiescent populations show similar scatter ($\sim 0.15$ dex).
   • Breaks: Quiescent and bulge-dominated populations show double-power law breaks, but at different pivot masses. The quiescent pivot mass is $\delta_Q \sim 10^{10.7} M_\odot$, while the bulge-dominated pivot is $\delta_B \sim 10^{11.1} M_\odot$.

2. Decoupling of Quenching and Structural Transformation: The offset between the quiescent pivot mass ($\delta_Q$) and the bulge-dominated pivot mass ($\delta_B$) suggests that star-formation cessation (quenching) and structural transformation (bulge formation) occur on different time-scales. Quenching precedes morphological transformation.

3. Morphology-Dependent Trends Require Space-Based Imaging: The structural trends identified (specifically the correlation between $B/T$, mass, and redshift) are washed out in ground-based imaging (HSC-I) but are clearly recovered in JWST/COSMOS-Web data. Ground-based measurements fail to resolve the structural diversity within sSFR classes, particularly at low redshift where a significant fraction of quiescent galaxies are disc-dominated.

4. AGN Feedback Connection:

   • The quiescent pivot mass ($\sim 10^{10.7} M_\odot$) coincides with the mass scale where AGN (infrared torus) bolometric luminosity fraction peaks in transitioning galaxies and where AGN feedback becomes effective in halos ($M_{200} \sim 10^{12} M_\odot$).
   • The bulge-dominated pivot mass ($\sim 10^{11.1} M_\odot$) coincides with the halo mass ($M_{200} \sim 10^{12.7} M_\odot$) where AGN-driven baryonic redistribution peaks, creating conditions for disc heating and structural transformation.
   • At $z < 0.5$, the double-power law pivot becomes unconstrained for the most massive galaxies, suggesting that in the highest mass halos ($M_{200} \gtrsim 10^{13.5} M_\odot$), AGN feedback may fail to prevent gas re-accretion, leading to a breakdown of the sharp quenching transition.

Significance and Claims
The paper claims that galaxy size scaling relations encode a sequence of physically distinct thresholds rather than a single evolutionary transition. The primary significance lies in demonstrating that:

• Classification Schemes are Not Interchangeable: sSFR, $B/T$, and Sérsic index capture different stages of the "quenching-then-transformation" sequence. Treating them as equivalent obscures the mass-dependent decoupling of these processes.
• Intrinsic Scatter as a Halo Probe: The dependence of intrinsic scatter on morphology (but not sSFR) implies that galaxy size is connected to halo spin primarily through structural configuration (e.g., angular momentum conservation in discs vs. secular processes in bulges), rather than star-formation activity.
• Necessity of Space-Based Resolution: Recovering the morphology-dependent trends and the true structural diversity within sSFR classes requires the resolution provided by space-based imaging (JWST), as ground-based data blurs these distinctions.
• Two-Stage Evolution: The results support a two-stage picture where AGN feedback triggers quenching at a specific mass scale, followed by a delayed structural transformation into bulge-dominated systems, driven by secular evolution and disc heating, with the pace modulated by environment.

The authors conclude that using a population-level forward model (pop-cosmos) combined with high-resolution space-based morphology provides a coherent framework to disentangle these evolutionary pathways.