WorksPhysics and Astronomy
FORGE'd in Fragments: Pop III star formation in a Milky Way-like progenitor halo
Abstract
We present a radiation-magnetohydrodynamic (RMHD) cosmological zoom-in simulation of Pop III star formation, drawing initial conditions from a Milky Way-like progenitor halo at redshift . We follow the collapse of primordial gas over 10 orders of magnitude, through to the formation of an accretion disk that becomes gravitationally unstable kyr before the first protostar's formation. Prominent spiral arms then develop and fragment vigorously, producing sink formation events over the first kyr post-collapse evolution, with a total stellar mass reaching by kyr, the end of our simulation. The disk remains magnetically supercritical throughout: the mass-to-flux ratio falls from - to by kyr, as the inflow brings in magnetic flux. The magnetic energy in the disk remains sub-dominant throughout, but grows by orders of magnitude after the first star forms. The disk starts sub-virial and reaches virial equilibrium within kyr. The total stellar mass of the population grows as , confirming the prediction of accelerating accretion in self-gravitating turbulent collapse, at rates of -. The star formation efficiency reaches within the disk (after kyr; at kyr) and within a pc sphere. We find a top heavy IMF () with a high-mass slope . The cosmological environment sets the boundary conditions of the disk: an envelope delivers to the outer edge of a disk with a radius of AU; the disk spreads and, in our simulation, its fragmentation sites move outward, from to ,AU, until it has been depleted.
Introduction
The first generation of stars, Population III (Pop III), is thought to have formed from metal-free primordial gas in dark-matter minihalos at redshifts – [2, 11]. Because molecular hydrogen (H) was the only efficient gas coolant at these epochs, the primordial gas could cool to a temperature floor of K at densities cm, before reaching local thermodynamic equilibrium in its ro-vibrational levels [11, 49]. The Jeans mass at this point, M, is orders of magnitude larger than that of present-day star-forming cores, leading to the early expectation that Pop III stars were predominantly massive and formed in isolation within each minihalo [2, 11, 117]. These early simulations bridged cosmological collapse down to protostellar densities ( cm), resolving the characteristic stages of primordial collapse: H three-body formation heating at cm, collision-induced emission cooling at cm, and a second collapse driven by H dissociation at cm [71, 117, 115].
This picture of isolated, monolithic Pop III star formation was altered as a result of simulations that continued to follow the system after the initial protostar formed. The infalling gas inevitably carries angular momentum, seeding a self-gravitating accretion disk prone to gravitational instability and fragmentation, so that a protostellar cluster arises instead of a single object [99, 15, 26, 100, 115].
Corresponding author: Vasilii Pustovoit The protostellar cluster is shaped by competitive accretion ([7]), fragmentation-induced starvation ([78, 80]), protostellar mergers, and dynamical ejections, with 31–43% of the surviving sinks on escaping trajectories ([97, 98]).
Turbulence plays a fundamental role in shaping fragmentation and the resulting mass distribution. Cosmological infall into dark-matter potential wells drives supersonic turbulence in Pop III minihalos, with Mach numbers –5 that increase with halo mass ([14] (M.-Y. Ho et al. 2026)). K. M. J. Wollenberg et al. [115] showed that the degree of initial turbulence strongly affects the number of fragments and the resulting mass distribution, with higher initial turbulent velocities producing more fragments and more stochastic outcomes. J. Prieto et al. [79] showed that turbulent fragmentation in a larger, halo at produces gravitationally unstable clumps, proposing that infall-driven turbulence may be the dominant fragmentation mechanism when halo masses exceed the canonical minihalo scale.
Magnetic fields can further modify Pop III star formation, though their net effect remains debated: even trace seed fields are amplified during collapse to 3–100% of equipartition ([29]) and can become dynamically important within yr of protostar formation ([88, 30, 60]). Strong fields can suppress disk fragmentation ([30, 101]), but fragmentation persists, if reduced, in other runs ([81, 86]), and early on magnetic fields limit protostellar masses more effectively than radiation ([89, 90]). The outcome depends on initial field strength, geometry, resolution, run duration ([49]) and even the numerical MHD scheme ([112]).
The Pop III initial mass function (IMF) inferred from these simulations is generally thought to be top-heavy, spanning to several hundred , with single stars in some halos and multiple systems in others ([107]). When radiative feedback is included, accretion onto a single protostar can be halted at ([43]), though episodic accretion from a fragmenting disk can produce intermittent feedback and final masses from to several hundred ([42]). The large statistical survey of S. Hirano et al. [31] across cosmological halos confirms that the Pop III IMF is not universal: the outcome depends on halo mass, angular momentum, and baryon streaming velocity.
While many Pop III simulations now use cosmological initial conditions ([26, 98, 56, 31, 64]), simulations that simultaneously combine cosmological ICs, full radiation-MHD, and resolved individual protostellar formation remain few. Y. J. Meziani et al. [64] use the FORGE’d in FIRE framework ([41]) to study Pop III star formation in an UFD (Ultra Faint Dwarf) galaxy progenitor, the m09 halo ([113]) ( at ), where the first Pop III star forms at in a halo of . A minimal-embedded in a strong overdensity destined to grow to Milky Way (MW) mass may experience qualitatively different infall rates, tidal torques, and angular momentum injection compared to one in a quieter environment destined to remain a dwarf galaxy. Therefore, in this paper, we zoom-in to the formation of the first stars in a massive halo of (above the atomic-cooling threshold at ) that eventually grows to MW mass at .
We perform a cosmological radiation-magnetohydrodynamic (RMHD) zoom-in simulation of Pop III star formation in the MW-progenitor halo ([39, 113]), adopting the cosmology and the comoving Mpc parent box size of the AGORA project ([47]). We use the FORGE’d in FIRE framework ([41]), combining FIRE-3 ([40]) large-scale physics with STARFORGE ([27]) protostellar physics, and evolve the system for kyr after the first protostar forms, reaching a gas mass resolution of –. We study the thermodynamic collapse, disk formation and fragmentation, magnetic field amplification, and protostellar mass assembly in this environment, providing the first cosmological RMHD simulation that resolves individual Pop III protostellar formation in a MW-progenitor halo. Figure 1 shows a multi-scale view of the simulation at the end of the run (at kyr).

Figure 1. Multi-scale density projection of the simulation at kyr, illustrating the dynamic range from the cosmic web (box size of kpc) to the center of the protostellar disk (resolving AU scales). White boxes mark the region shown in the next panel. Yellow star symbols mark sink particles, with size of the marker scaling with the final mass of the star. The progenitor halo has a dark matter mass of at the first collapse, .
The rest of the paper is organized as follows. § 2 describes the simulation setup, physics modules, refinement strategy, and analysis methods. § 3 presents the disk morphology, thermodynamic and energetic evolution, disk stability, star formation history, accretion rates, protostellar mass function, and magnetic field structure. § 5 compares our results to previous Pop III simulations. § 6 presents caveats of our simulation. We state our conclusions in § 7.
Simulation and Methods
Simulation details
For this paper we utilize the FORGE’d in FIRE framework of P. F. Hopkins et al. [41], extended to the primordial star-formation regime similar to Y. J. Meziani et al. [64]. We employ FIRE-3 ([40]) physics modules at low resolution (, ) and STARFORGE ([27]) physics to resolve individual protostellar formation. We briefly review these below and refer the reader to M. Y. Grudić et al. [27]; P. F. Hopkins et al. [40, 41]. Numerical parameters not specified below are identical to those described in the above references.
Code and gravity solver
The governing equations solved by the code are the ideal MHD equations in conservative form (P. F. Hopkins & M. J. Raives [38]):
where the state vector of conserved quantities and the flux tensor are
Here is the mass density, the velocity, the magnetic field (in Gaussian units with absorbed), the total specific energy, the total (thermal plus magnetic) pressure, and the identity tensor. includes source terms for self-gravity and the Powell 8-wave and Dedner divergence-cleaning terms (P. F. Hopkins & M. J. Raives [38]) that maintain .
The FORGE’d in FIRE framework is based on GIZMO Meshless Finite Mass (MFM) code (P. F. Hopkins 2015). Gravity is solved with the TreePM algorithm [3, 94], which combines a particle-mesh grid for large-scale forces with an oct-tree for short-range forces [4], as in V. Springel [95]. Gas particles use adaptive Lagrangian gravitational softening matching the hydrodynamic kernel radius. Sink particles use a fixed Plummer-equivalent softening AU (gravitational kernel radius AU). Direct -body summation is used for all sink–sink and sink–gas gravitational interactions, with direct summation applied for sink–sink separations below 1000 AU to accurately resolve close binaries and dynamical ejections [27].
Initial Conditions
We use initial conditions (ICs) developed as part of the FIRE collaboration [39], specifically the m12f halo from the Latte suite of MW-mass zoom-in simulations [113]. The m12f halo has a virial mass at and was selected as an isolated halo of this mass, without regard to its formation history [113]. It is one of the four Latte halos (with m12i, m12m and m12b) that are most similar to the MW across a range of properties [113]. At the redshift of first collapse in our simulation (), CDM part of the progenitor halo has a mass of $2 \times 10^{7},\mathrm{M}_{\odot}$. This is above the atomic-cooling threshold at this redshift: the virial temperature is K, whereas K corresponds to at , and the halo is times more massive than the minimum mass for H cooling, a few [116, 54]. As in other atomic-cooling halos, the dense gas still cools through H, whose formation there is largely insensitive to external UV (S. P. Oh & Z. Haiman 2002): later work finds that H cooling in such halos is suppressed only when the Lyman–Werner (11.2–13.6 eV) background exceeds a critical intensity. Expressed through , the specific intensity in units of , this critical value is –300 for soft spectra and for realistic spectra of metal-poor galaxies, and up to – for metal-free stellar spectra [87, 106]. This is far above the external UV background in our run, which is negligible at $z \approx 21 {xref}(subsubsection 2.2.3), and gas becomes fully molecular through three-body H$_2$ formation here by $n \sim 10^{11}$–$10^{12}\,\mathrm{cm}^{-3}$ {xref}(Figure 4); {citet}T. Abel et al. [2]. The late, massive first collapse is most plausibly a consequence of the parent-simulation resolution. With $m_{\mathrm{DM}} \approx 3.5 \times 10^{4}\,\mathrm{M}_{\odot}$, a $10^{6}\,\mathrm{M}_{\odot}$ minihalo contains only $\sim 30$ DM particles, while $\gtrsim 500$ dark-matter particles are needed to resolve the gas content of a halo {citep}[54], i.e. $\gtrsim 2 \times 10^{7}\,\mathrm{M}_{\odot}$ here. Earlier H$_2$-cooling minihalos in this region are therefore unresolved, and star formation can only begin once the halo is this massive. The hyper-Lagrangian refinement ({xref}subsection 2.1.3`) does not change this: it is switched on only at the restart, at , after the parent run has already set the epoch and the host of the first collapse; it splits only gas particles, so a minihalo would still be resolved by DM particles; and its Jeans criterion is triggered only once the gas is already dense and cold (–100 at a few hundred K for the parent gas particle mass), so it can follow a collapse but not start one in an unresolved halo.
We adopt pre-generated ICs at in a 60 Mpc/ comoving cosmological box. This large box contains a zoom-in region of comoving size Mpc/, within which the FIRE galaxy grows by . Only the zoom-in region contains baryonic particles and high-resolution dark matter (DM) particles; the remainder of the volume is populated with low-resolution DM particles. As in the FIRE-2 Latte simulations, we adopt cosmological parameters from WMAP-7 [51]: , , , , , and .
The full simulation snapshot contains million CDM particles and FIRE-resolution gas particles whose smoothing lengths ( kpc) far exceed the protostellar disk scale. We extract spatial cutouts centered on the most-massive sink particle with a physical radius of 1 kpc (22 ckpc). The cutout sphere retains only the gas, high-resolution DM, and sink particles; we then split gas particles as described below, but leave the DM particles unsplit. All analysis presented in this manuscript is of this cutout simulation.
Over the kyr of post-collapse evolution studied here, the sound-crossing time across the cutout radius ( kpc) is Myr for K (), more than three orders of magnitude longer than our integration time. The large-scale structure outside the cutout therefore has no causal influence on the disk dynamics on these timescales, and the cutout fully captures the relevant physics.
Refinement strategy
We first evolve the full cosmological simulation with FIRE-3 physics until first stellar population (SSP) particle forms. We then restart from the preceding snapshot with an additional “hyper-Lagrangian” Jeans refinement layer, applied globally; [41] use a similar “hyper-refinement” layer centred on the central black hole.
We refine gas cells (splitting into pairs of equal-mass daughter cells) whenever their mass exceeds a fraction of the local Jeans mass:
where is the Jeans mass of gas at K and adopted in GIZMO, and is the gas temperature for a mean molecular weight of 1.23. Only gas well above the cosmic mean density is refined, and cells are not split below . We use , corresponding to 100 resolution elements per Jeans mass and satisfying the J. K. Truelove et al. [109] criterion, since we resolve cells per Jeans length. This is sufficient: Lagrangian methods need only particle widths per Jeans length to avoid artificial fragmentation, and particle splitting in GIZMO’s MFM does not seed artificial fragmentation (spurious bound clumps that would go on to form sink particles), its effect on the outcome lying within the run-to-run scatter [46 and references therein]. It is, however, below the cells per Jeans length that grid codes need to capture small-scale dynamo amplification [21].
Figure 2. Top panel: Radial profile of gas particle mass (resolution), centered around the most massive sink. Scatter plot shows particle distribution at the time of first sink formation; solid line and shaded region correspond to the mean and ±1σ for radial profiles across all snapshots after first sink formation. Horizontal lines correspond to 0.5 of previous resolution region, since our particle splitting algorithm is subdividing parent particles in 2 for each new level of split. Bottom panel: Same as top panel, but for radial profile of equivalent cell size .
No equivalent criterion has been calibrated for MFM: only magnetorotational instability (MRI) quality factors [18] and particle-number convergence tests of gravitoturbulent dynamos [17] exist. We therefore do not claim that a small-scale dynamo is resolved here (§ 6).
Only gas particles get split, DM particles stay at the same resolution. This is a fine approximation because dark matter’s potential is smooth at these scales, and splitting of these particles should not change the results of our simulation after the restart. It does matter for the epoch of the first collapse, which is set before the restart by DM halos resolved with (subsubsection 2.1.2).
At our highest refinement level, the gas mass resolution is and the maximum spatial resolution is in most snapshots, reaching briefly, sufficient to resolve densities up to –. This spatial resolution applies only at the highest densities: at 15 kyr the median disk particle has (the depleted late-time disk gas has low density, which is also why it is resolved by cells per Jeans length: the median of the disk particles at 15 kyr, with 10th–90th percentiles of 26 and 160), so the disk scale height is covered by only a few resolution elements at late times.
Figure 2 shows the radial profile of the length scale of gas particles . It exhibits the refinement profile, described in subsubsection 2.1.3 at all times, except the earliest one. Initial profile at (0.5 kyrs before the first sink formation) is limited by the refinement scale inside of 100 AU radius (the center was not yet refined to the highest resolution). But by the time of first sink formation, the refinement reaches highest resolution that we expect from our scheme, (around or a few AU) so the inner star forming disk region is fully refined.
Physics we model
Magnetic fields
We integrate the ideal MHD equations as described in P. F. Hopkins & M. J. Raives [38], using the constrained-gradient method and divergence cleaning P. F. Hopkins [34] for divergence control, which has been validated for magnetorotational instability dynamics P. F. Hopkins et al. [41].
We include Spitzer–Braginskii thermal conduction and viscosity via P. F. Hopkins [35]. For metal-free gas at the temperatures (– K) and free-electron fractions relevant here, the non-ideal MHD timescales (Ohmic, Hall and ambipolar diffusion) exceed the local dynamical timescales by many orders of magnitude [64], so ideal MHD dominates throughout the simulation. Spitzer conduction and Braginskii viscosity are also negligible: even for fully ionized gas at K and , the conduction and viscous timescales across 100 AU, with the Spitzer thermal or momentum diffusivity, are yr, and the gas here is almost neutral.
The initial magnetic field is a trace cosmological seed field of G (physical), imposed at the start of the parent simulation (), i.e. G, or G comoving, uniform in magnitude and direction, amplified self-consistently by gravitational compression, differential rotation, and (possibly) turbulent dynamo. The seed lies about two orders of magnitude below the present-day field strengths of – G (comoving) that D. Koh et al. [50] consider, and below the CMB upper limits on primordial fields they quote, G for inflationary and G for causal fields; it therefore does not test the possibility that a strong primordial field delays or prevents the collapse. The original orientation is , i.e. , in the cosmological box, but the protostellar disk is oriented randomly in the simulation.
Radiation
We solve radiation transport with the M1 moments method (C. D. Levermore 1984), as implemented in GIZMO and described in P. F. Hopkins & M. Y. Grudić [36]; P. F. Hopkins et al. [37]; M. Y. Grudić et al. [27]; P. F. Hopkins et al. [41] across five frequency bands: H-ionizing ( eV), Lyman–Werner (LW)/photo-electric (8.0–13.6 eV), near-UV (NUV; 3.4–8.0 eV), optical–NIR (0.41–3.4 eV), and an adaptive-wavelength FIR band ( eV) that tracks the evolving dust emission temperature (P. F. Hopkins et al. [40]). Gas, dust, and radiation temperatures are evolved separately with appropriate coupling terms, allowing self-consistent treatment of both optically thin and optically thick regimes. We use a reduced speed of light , i.e. km s, much lower than the adopted for most runs in P. F. Hopkins et al. [41]. Tests in galaxy simulations find small effects for km s, with slightly weaker radiative feedback at lower values down to 300 km s (P. F. Hopkins et al. [37]), and km s suffices to follow expanding ionization fronts (M. Y. Grudić et al. [27]). A radiation-hydrodynamic simulation of Pop III star formation that adopts the same found no significant change in its results for between and , with the fastest outflows at only km s (J. Park et al. [76]). Our value is therefore at the lower edge of the galaxy-scale range and well above the ionization-front limit, but we have not tested the sensitivity of our results to . Fast jets would be the feature most likely to be affected, but the jets of this run do not propagate as intended because of an implementation error (subsection 2.2.4). No ionization fronts are followed either: the Strömgren radii of the sinks are below the resolution, so their ionizing photons are absorbed within a cell (§ 6).
The Lyman-Werner (LW) band is combined with the photo-electric (PE) heating band into a single radiation frequency bin. This follows the five-band scheme of P. F. Hopkins et al. [37] and P. F. Hopkins et al. [41], in which a single far-UV band serves both purposes. Our gas is metal-free and dust-free, so there is no photo-electric heating, and the band acts only as the Lyman–Werner band. H photo-dissociation by LW photons is accounted for through the H destruction term in Equation (4).
Thermochemistry
We solve gas thermodynamics and chemistry as in P. F. Hopkins et al. [40]. The following cooling functions are employed and are relevant to our setup: (atomic H cooling); (HD cooling, D. Galli & F. Palla [22]); (H cooling); , (H interactions with e, H, H, He below and above the critical density at which the H levels reach LTE, combined per molecule as (P. F. Hopkins et al. [40])). The network does not include H line opacity; in GIZMO, optically thick cooling is only limited by a gray slab photosphere cap). The adiabatic index accounts for H, He, free electrons and H (A. C. Boley et al. [6]). We follow non-equilibrium H and He chemical networks. Default FIRE-3 thermochemistry includes cooling due to metal lines, but we do not evolve our simulation past the point of supernova-driven metal enrichment, thus making all metal cooling contributions effectively zero (as expected in the Pop III formation regime).
Since we simulate metal-free primordial gas, the dust-to-gas ratio is zero and the dust term in the H formation equation is absent. H formation proceeds exclusively via gas-phase channels: the two-body H and H pathways at low densities ( cm) and the three-body reaction () at high densities. Molecular fraction of H evolves in every cell according to (eq. 6 from P. F. Hopkins et al. [40]), with the dust formation term set to zero:
where the gas-phase, three-body, photodissociation, cosmic-ray and collisional terms and their coefficients are defined in Appendix § 10.
The chemical energy of H (4.48 eV per molecule) is not tracked in the gas energy equation: neither the heating released when H forms nor the cooling when it is collisionally dissociated is included. This matters below the densities we resolve. Three-body H formation heating is the dominant heat source in primordial gas at – cm (M. J. Turk et al. [110]), and in real gas dissociation cooling thermostats gas heated above K, which our network cannot do. The FIRE-3 network is therefore a reduced, not a complete, primordial network (see subsection 3.1 and § 6). In addition, the protostellar evolution tracks of S. S. R. Offner et al. [69] self-consistently evolve the stellar radius without requiring the gas to track dissociation energy explicitly.
The meta-galactic UV background is that of C.-A. Faucher-Giguère [20]; we expect it to be negligible at compared to local protostellar radiation resolved in the simulation. self-shielding against Lyman-Werner photons is included using the shielding function of N. Y. Gnedin & B. T. Draine [25] (their Eq. 3), applied with a local Sobolev column density , with turbulent line broadening accounted for following a Burgers velocity spectrum.
The simulation does not include a cosmic-ray (CR) module: CRs are neither transported nor injected by stellar feedback (which would in any case be negligible over kyr without supernovae). The thermochemistry instead uses GIZMO’s default for runs without such a module, a uniform and constant CR energy density , the local interstellar value inferred from Voyager 1 data (A. C. Cummings et al. [16]), which corresponds to an ionization rate . 200 times the cosmic mean baryon density and is fully on above times the mean ( at ), which includes all the gas analyzed here. Without attenuation, it enters the network (Equation (4)), the free-electron fraction and the gas heating. We discuss its effect in § 6.
Stellar formation and evolution
We resolve the formation of individual protostars in the densest region of our simulation using the STARFORGE (M. Y. Grudić et al. [27]) formation and evolution modules. A gas particle becomes a sink (protostar) particle when it meets the full STARFORGE formation criteria: the cell must be (i) locally self-gravitating (virial parameter ), (ii) Jeans unstable, (iii) in a converging flow (), (iv) experiencing compressive tides in all directions, (v) a local maximum of the density field, and (vi) not within the accretion radius of a pre-existing sink (M. Y. Grudić et al. [27]). For brevity we refer to these protostellar particles as “sinks” throughout this paper. When a gas particle is converted into a sink, the magnetic field of that gas is deleted.
Once formed, the sink particles accrete the bound gas whose circularized orbit falls within the sink radius. Accreted gas is first added to the sink’s internal reservoir, which represents the unresolved envelope and disk inside the sink radius. Mass flows from the reservoir onto the protostar on a smoothed accretion timescale, and when jets are enabled a fraction of it is diverted into a jet reservoir and returned to the gas as outflow (M. Y. Grudić et al. [27]). In our run the reservoirs hold of the mass of most sinks at 15 kyr (at most 2.4%), so the sink masses we quote are effectively stellar masses. Each sink evolves along protostellar tracks following S. S. R. Offner et al. [69], explicitly tracking the stellar radius, luminosity, and effective temperature as a function of the instantaneous mass and accretion rate.
Jets are the least tested part of this feedback. Most Pop III simulations omit them, and the few that include them treat them as mass-removal prescriptions rather than resolved outflows (Y. J. Meziani et al. [64]). Whether Pop III protostars can launch jets at all depends on whether a trace magnetic field is amplified enough for the mass-to-flux ratio to permit it; Y. J. Meziani et al. [64] find that the ratios around a Pop III protostar fall well within the range required for strong jet launching. Protostellar jets are launched with a mass-loading proportional to the accretion rate onto the protostar and a launch velocity of 0.3 times the Keplerian speed at the protostellar surface (M. Y. Grudić et al. [27]). The jets module is active in the present run, but the jet material is coupled to the gas closer to the protostars than intended. STARFORGE launches jets by spawning new gas cells near the sink (M. Y. Grudić et al. [27]), and in the code version we use these jet cells are merged with the neighboring ambient gas cells too readily. Their mass and momentum are therefore spread into the surrounding gas near the protostar, and the jets propagate less far than intended. Because the internal structure of the jets is not resolved in any case, this amounts to a choice of where the jet feedback is deposited rather than an error in the injected mass or momentum. The same merging is more serious for stellar-wind bubbles, where the contact discontinuity between the shocked wind and the ambient gas must be followed, but STARFORGE winds come only from main-sequence stars (M. Y. Grudić et al. [27]), and none of our protostars reaches the main sequence (subsection 3.5). We defer the analysis of jet feedback to future work.
Postprocessing methods
Disk identification
To compute velocity dispersion, local Toomre , and other disk properties, we identify the gas particles belonging to the disk at each snapshot. We center on the densest gas particle before the first sink forms and on the most massive sink afterwards, and rotate into the frame of the total angular momentum within . Particles are selected by cuts on cylindrical radius, aspect ratio (), density () and rotation (), and the selection is extended to its bounding cylinder. The disk radius is the 90th-percentile cylindrical radius of the disk members. The full algorithm and tests of its robustness are given in Appendix § 11.
Figure 3 shows the gas surface density of the disk and its surroundings at the six epochs used throughout the paper (, 0, 1.3, 3.6, 9.5 and 15 kyr relative to the formation of the first sink). Each epoch has a face-on panel, seen along the angular momentum axis of the procedure above, and a narrower edge-on panel below it that shows the vertical structure. All panels are centerd on the most massive sink, span 4100 AU on a side, and share one logarithmic color scale (– M pc); the bar in each face-on panel marks 1000 AU, and stars mark the sinks inside the image box. Before the first sink forms ( kyr) the gas is a smooth, centrally concentrated core with no sinks. From kyr onward the disk develops spiral arms along which the sinks that formed by fragmentation are scattered, rather than gathering in one central cluster (subsection 3.4). By 15 kyr the disk is depleted, and only a few sinks and a faint spiral remain in the field.

Figure 3. Face-on and edge-on surface density projections at six epochs spanning the kyr after the formation of the first sink. Each panel shows a AU region centerd on the most massive sink. Star symbols mark sink particles within the image box. The sinks are scattered throughout the disk rather than concentrated in a single cluster, tracing the spiral arms.
Results
In this section we present results from our cosmological RMHD simulation of Pop III star formation. We organize the discussion as follows: subsection 3.1 discusses the thermodynamics and chemistry of the collapse; subsection 3.2 describes the structure and kinematics of the pre-stellar envelope and shows that its collapse proceeds outside-in; subsection 3.3 characterizes the disk structure and energetics; subsection 3.4 covers its gravitational stability; subsection 3.5 covers the star formation history, accretion rates, and protostellar mass function; and subsection 3.6 describes the magnetic field evolution.
Thermodynamics
The thermal and chemical state of the gas reflects the competition between compressional heating and radiative cooling. The temperature–density phase trajectory is therefore diagnostic of which physical process dominates at each stage of collapse, while the H fraction shows how much of the main coolant is available. We present phase space plots in Figure 4 for three representative epochs, pre-stellar, first sink formation, and post-first sink formation. Disk particles are identified by the algorithm described in 2.3.1; disk particles are green, particles in the cutout but not in the disk are red, while particles outside the cutout region are shown in blue.
Figure 4. Phase diagram (temperature and H fraction vs. number density) at three representative epochs: pre-stellar ( kyr, left), first sink formation (, middle), and post-first-sink ( kyr, right). Green points correspond to disk-identified particles, red points correspond to points in our cutout simulation that have not been identified as disk, and blue points correspond to points in the full-box simulation right before the cutout was applied (only included in the leftmost plots). for most of particles in the low-density region (blue particles), which therefore do not appear in the bottom-left panel. The curves in the temperature panels show the theoretical – tracks for primordial () and metal-enriched () gas from R. S. Klessen & S. C. O. Glover [49] (KG23); the shaded bands indicate the median standard deviation of the fits over the time intervals (middle) and (right), respectively. The spread of the fit is narrow at early times and broadens at later times. KG23 curve follows similar curve from K. Omukai [71] closely.
Before the first star forms ( kyr; Figure 4, left panels), the gas occupies a narrow track spanning – and K. The low temperature reflects the fact that at densities , the molecular radiative cooling , while the adiabatic heating , and cooling dominates. Above the H levels are in LTE, so that , with the result that heating dominates. Our network follows the same scaling, because at high density it uses the LTE cooling rate per molecule (section 2.2.3).
The H fraction (lower panel) rises modestly from at to at , driven primarily by two-body formation channels (H and H pathways). The increase in H fraction enhances the cooling of the gas, since H is a much more efficient coolant than atomic H.
The low-density (blue) particles in the left column (; see the caption) lie outside our region of interest; the dynamical time scale is long in that region, so the thermodynamic state of the gas does not change significantly over the time interval shown in the figure.
By (Figure 4, middle panels), the gas has collapsed to much higher densities () and the thermal track has extended significantly. At intermediate densities (–), compressional heating outpaces H line cooling, which is saturated at the LTE value, and the temperature rises steadily through K. The H fraction increases steeply in this regime as three-body formation reactions (3H H + H) become efficient at , converting the gas to fully molecular () by –. Note that unlike in the pre-stellar regime, we see an increase in temperature associated with an increase in H fraction. With the levels in LTE, increasing H adds little cooling, so it cannot offset the compressional heating of the continuing infall. In real primordial gas the H lines also become optically thick at these densities (E. Ripamonti & T. Abel [82]), but our network does not model line opacity (subsubsection 2.2.3).
At kyr (Figure 4, right panels), the temperature range in the lower density part of the phase diagram has broadened substantially. At higher densities, the gas has separated into 2 parts: a branch with K and a hot branch at K, which we interpret as the result of shock-heated and adiabatically compressed gas. This hot branch is capped at K, where Lyman- cooling of atomic hydrogen becomes efficient (S. P. Oh & Z. Haiman 2002; J. H. Wise & T. Abel [114]). We note that while our thermochemical model does not couple H dissociation to thermodynamics, the terms in Equation (4) can still drive partial dissociation of H.
At higher densities attained at later times, H is formed primarily by three-body interactions. This results in constantly increasing fraction of molecular hydrogen in the disk with time.
Pre-stellar envelope and collapse
Here we describe the mass distribution and infall rate of the pre-stellar envelope (subsubsection 3.2.1) and its kinematics (subsubsection 3.2.2), and then compare both with self-similar collapse solutions to show that the simulated collapse is driven from the outside in (subsubsection 3.2.3). This contrasts with the inside-out expansion wave picture of F. H. Shu [91], which applies only when a core is marginally critical.
Mass distribution and infall rate
The density profiles of all epochs coincide outside AU, so only the inner region changes with time (Figure 5, top panel). Between and AU the envelope follows (one fit to all epochs). Inside AU the profile changes with time: it is shallower before the first sink forms, when the center is not yet fully refined, and it becomes irregular once the spiral arms and the sinks dominate the azimuthal average.
Figure 5. Top panel: Radial density profiles through simulation time. The x-axis begins at AU, corresponding to the resolution limit of the simulation. The dashed line is a single power law fitted to all epochs together between and AU, where the profiles coincide: . Top-middle panel: Enclosed gas mass in a sphere vs radius of the sphere. The fit is roughly . Bottom-middle panel: the infall rate into a sphere of radius , , counting only infalling gas () on the surface of the sphere of radius . At later times the inner region gains mass slower than at the earlier times (mostly close to the time of the first sink forming). There is a characteristic “flattening” at smaller radii inside of the disk before the first star has formed. Outer shells gain mass faster at all times—a characteristic of “outside-in” collapse. Bottom panel: evolution of radial profile of kinematic virial parameter with time, calculated with the energies defined in subsubsection 3.3.1 for the gas and sinks inside a sphere of radius (dark matter and stars included in the potential). Each curve starts at the resolution limit of its epoch. Inside AU, .
The enclosed gas mass (Figure 5, top-middle panel) follows overall, a mean density slope of over the whole range, and the curves of all epochs again differ only at small radii. The infall rate through each sphere (Figure 5, bottom-middle panel) is nearly independent of radius, at between and AU, at all epochs, and rises as further out, to at AU. A flux that does not change with radius means that the gas does not pile up on its way in: the inner envelope is a quasi-steady flow fed from larger scales.
The virial parameter (Figure 5, bottom panel) is the ratio of the energies defined for the disk in subsubsection 3.3.1 (Equation 5), but calculated for all the gas and sinks inside a sphere
of radius instead of the disk particles only (Figure 8). The ratio is before the first sink forms and afterwards between and AU, so the gas is close to virial balance in the collapsing region outside of the disk.
Figure 8. Top panel: Time evolution of the main energy components computed for disk-identified particles (the disk radius is taken as the 90th-percentile cylindrical radius of disk-identified particles, see subsection 3.3): rotational kinetic energy (yellow), turbulent kinetic energy (cyan), total kinetic energy (green), gravitational potential energy (red), thermal energy (orange), and magnetic energy (magenta). Bottom panel: pure kinematic virial ratio (black); virial ratio with added thermal component (orange) , and with added thermal and magnetic components (purple), . The dashed red line marks virial equilibrium (). The disk starts sub-virial and reaches equilibrium by kyr. Including the thermal energy matters mainly at early times, when is comparable to the kinetic terms; at late times the disk is supported mainly by rotation and turbulence. Magnetic energy provides negligible additional support to the virial ratio.
Gas kinematics
Figure 6 shows radial profiles of the gas velocity components, the turbulent velocity magnitude and dispersion, and the Keplerian speed of the enclosed mass. We plot the radial velocity, (negated so that positive values indicate radially inward flow), to simplify comparison. All quantities are mass-weighted averages in logarithmically spaced cylindrical bins about the disk axis. Here and are the mean radial and azimuthal velocities of a bin and is the radial velocity dispersion. The turbulent velocity is the mean speed after subtracting the bin’s mean radial and azimuthal flow, and is the sound speed.
Figure 6. Radial profiles of inward radial velocity , azimuthal velocity , sound speed , magnitude of turbulent velocity , radial/azimuthal/vertical turbulent velocity dispersion , Keplerian speed at each radius, computed from the enclosed mass. Bright lines show functions with spline smoothing, dim lines show actual data. grows with time, as well as . Radial infall () speeds up until the bulk star formation events (1.3–3.6 kyr), and slows down afterwards. Disk rotation () is constantly speeding up, leveling off outside of the disk ( AU). , , track each other very closely, but decouples from the other two by the kyr time outside of the disk. Turbulence is supersonic at all times, tracking Keplerian speed closely for most times outside of the disk, but falling below it inside of the disk. Dot-dashed lines correspond to virial radius and virial velocity of the halo in the and axes respectively. The turnover seen in the innermost bins of these wide profiles is a numerical artifact of plotting data in the cutout region – it does not exist in Figure 7, which uses the full simulation data. This is not a physical feature and thus is excluded from the interpretation.
The general trend is that , , increase with time everywhere. is negative and thus gas is mostly infalling for the whole simulation, with increasing until stabilizing at around 1.3–3.6 kyr (bulk star formation events), and dropping after stars are fully formed. rises until 1.3 kyr, after which it stabilizes. It is clear that the turbulence starts off as sonic or sub-sonic, but becomes greatly supersonic once the sink forms. The turbulent velocity dispersion increases inward, tracking the increase in gravitational potential. Because the collapse compresses the gas faster than the turbulence can decay, the turbulence is continuously re-driven by the contraction itself rather than by stellar feedback (B. Robertson & P. Goldreich [84]); feedback only begins to contribute once the first protostar forms.
Figure 7 shows the same inflow over the full simulation volume: beyond the disk, rises with radius overall. Here is the radius inside which the mean total density (gas, dark matter, stars and sinks) equals , with the spherical-collapse overdensity and (the fit for a flat universe of G. L. Bryan & M. L. Norman [12], is the mass inside it, and . Between and AU, where the gas dominates the enclosed mass, it stays close to the local Keplerian speed (, i.e. of the local free-fall speed ), so this envelope collapses at close to free fall. Further out drops below , to near AU, and reaches at the virial radius, about half the free-fall speed of the whole halo. Beyond the virial radius rises steeply as the enclosed mass takes in the surrounding structure, while stays at . This extends the outside-in signature seen in the cutout to the full range of scales feeding the core. Inside AU, falls below as rotation takes over, and within the disk ( AU) the flow becomes predominantly subsonic. The deceleration is gradual rather than abrupt (subsubsection 3.3.2).
Figure 7. Radial inflow profile over the full simulation volume, out to AU (far beyond the virial radius), at four epochs (, 0, 1.3, and 3.6 kyr): the mean inflow velocity (solid) and the Keplerian speed from the total enclosed mass, including dark matter (dashed), both in the color of each epoch, and in black beyond AU, where the epochs are combined. Beyond AU the curves of the four epochs coincide, because the enclosed mass and the inflow there barely change over these 4 kyr, with the virial velocity and the virial radius marked. Bottom: enclosed gas mass . Unlike the wide kinematics profiles of Figure 6, this plot reaches the scale of the halo. The few bins inside 200 AU with are not shown on the logarithmic axis.
Outside-in collapse
The infall-rate profile (Figure 5, bottom-middle panel), flat inside and rising outward, is the opposite of an inside-out collapse, where approaches a constant at small radii but falls as at large radii (F. H. Shu [91]; D. W. Murray et al. [66]), and so is a signature of outside-in collapse. It is what the model of self-gravitating turbulent collapse of N. W. Murray & P. Chang [67] predicts, with independent of radius inside the sphere of influence of the central mass (their Eq. 45), and what D. W. Murray et al. [66] find in their simulations, where the accretion rate is either nearly flat or increasing with radius. Our plateau extends over a much wider range of radii and our density slope () differs from their inside the sphere of influence, so we regard the comparison to their solution as qualitative.
We find that the outer layers gain mass faster (Figure 5, bottom-middle panel). This contrasts with the inside-out model of F. H. Shu [91] and is qualitatively similar to the runaway, Larson–Penston-type collapse (R. B. Larson [55]; M. V. Penston [77]). Our envelope slope, between and AU, is close to the of the isothermal Larson–Penston solution, although the correspondence remains approximate because our gas is not isothermal. Supersonic infall alone does not distinguish the two pictures, since the interior of the Shu expansion wave also falls in supersonically. The envelope allows to differentiate between the two cases. The envelope mass flux, , is well above the Shu rate and comparable to the Larson–Penston values.
The comparison above concerns the collapse dynamics only: the core is not a Bonnor–Ebert sphere (R. Ebert [19]; W. B. Bonnor [8]) since (i) the core is never in hydrostatic equilibrium: it is assembled continuously by supersonic inflow from the halo (Figure 6, Figure 7) and its virial parameter is between and AU at kyr (Figure 5, bottom panel). The gas is therefore near virial balance, but because it is still infalling there is no static equilibrium state that later becomes unstable. (ii) The gas is not isothermal: across the pre-stellar core the temperature changes between and K as H cooling saturates and compressional heating takes over (subsection 3.1). (iii) There is no confining external pressure: the core merges smoothly into the infalling halo envelope instead of ending at a pressure boundary. Rotation, turbulence and magnetic fields also provide support that the Bonnor–Ebert model does not include. The Larson–Penston solution remains a useful comparison because it does not require an equilibrium starting state: R. B. Larson [55] found that the inner density and velocity profiles approach the same self-similar form for various initial conditions, provided the collapse time exceeds the sound-crossing time.
Disk evolution
Once the accretion flow acquires sufficient angular momentum to form a centrifugal barrier, a rotationally supported disk grows around the central protostar. The disk radius (subsubsection 2.3.1) stays at AU throughout the run, close to . The disk gas mass grows from just before the first sink forms to at kyr, and then falls to at 9.5 kyr and at 15 kyr as it is converted into stars. Its mass-weighted thickness, , is 0.23. The disk is thus comparable in size to the pc ( AU) disk of Y. J. Meziani et al. [64], but unlike that Toomre-stable disk, here the disk fragments (subsection 3.4). The subsequent evolution of the system is governed by the competition between gravitational instability (which fragments the outer disk into secondary protostars) and the supply of mass and angular momentum by infall from the envelope; the magnetic energy stays well below the thermal and kinetic energies (subsection 3.6).
Energetics
In this section, we consider the interplay of the gravitational, kinetic, thermal and magnetic energies for the particles identified in the disk. We compute the following energy components: potential , rotational , turbulent , thermal , and magnetic , each summed over the gas particles identified as disk members (subsubsection 2.3.1); the explicit expressions are given in Equation 5 below.
where is the mass of particle , its total velocity vector, , are bulk velocity profile fits for the azimuthal and radial components in each annulus (both functions of only), is the magnetic field vector, is the density, and is the internal energy.
Figure 8 shows the time evolution of the main energy components computed for disk-identified particles. Throughout the simulation, dominates over all other components, confirming that the disk is gravitationally bound. and are the main support terms and track each other to within a factor of , with slightly exceeding . starts comparable to the kinetic energies at early times but becomes subdominant at late times as the gravitational potential well deepens faster than the gas heats.
The magnetic energy, , grows over the kyr simulation, from erg at kyr to erg at 15 kyr. This growth is consistent with a combination of flux-freezing compression during the collapse, field-line winding by differential rotation, and potentially both the small-scale turbulent dynamo and the large-scale mean-field dynamo ([88]), but we cannot separate these channels (subsection 3.6).
Despite this rapid amplification, remains below throughout the simulation, though the gap narrows at late times – consistent with the declining mass-to-flux ratios discussed in subsection 3.6.
The bottom panel shows the virial ratio . At early times (), the virial ratio is , indicating a sub-virial, contracting state, i.e. the system has insufficient kinetic energy to balance gravity. The virial ratio rises as the disk forms and spins up, reaching by kyr. From 2 kyr onward, the virial ratio oscillates around unity, indicating that the disk has achieved and maintains approximate virial equilibrium through a balance of rotational support, turbulent pressure, and gravitational collapse. The virial ratio never approaches the unbound threshold of 2, except towards the end of our run, confirming that the system remains gravitationally bound throughout.
Disk morphology and evolution
As the disk accretes mass and angular momentum, its morphology undergoes a characteristic sequence: an initially smooth, centrally concentrated structure gives way to spiral arms as the Toomre parameter drops below 1 (the disk becomes unstable), and those arms subsequently fragment into secondary sinks. Face-on and edge-on surface density projections at six representative epochs spanning the kyr after the formation of the first sink () are shown in Figure 3 (subsubsection 2.3.1).
The disk does not begin as a rotationally supported structure. At the pre-stellar epoch lies well below the local Keplerian speed at every radius in Figure 6, so the centrifugal support of the collapsing core is weak and the infall is close to radial. Rotational support is acquired only gradually: rises monotonically with time while falls after the bulk star formation events, and only then does approach the threshold used for disk membership (subsubsection 2.3.1) over an extended range of radii. The disk we identify is therefore assembled from an initially slowly rotating, predominantly infalling core. The disk starts thick, with aspect ratio at during the pre-stellar epoch ( kyr), drops to a minimum of around the first bulk star formation event ( kyr) as the disk settles and gains rotational support, and then rises again at later times, to at 15 kyr. This late-time increase is most plausibly a consequence of the disk being disrupted: as gas is converted into stars it is removed from the disk, leaving material that is less rotationally supported and vertically more extended.
Disk stability and fragmentation
In this section, we analyze the stability properties of the disk. Whether a self-gravitating disk fragments into bound clumps, or instead disperses the accumulated mass through spiral-arm shocks and turbulent transport depends on the ratio of stabilizing pressures (thermal, turbulent) to the destabilizing self-gravity of the gas. The Toomre parameter, defined in Equation (10) ([108]), quantifies this balance locally: marks regions where perturbations on the disk scale grow faster than orbital shear can wind them up. We track the spatial and temporal evolution of to identify when and where fragmentation occurs, and how the velocity dispersion stirred up by the forming sinks subsequently self-regulates the disk back toward marginal stability.
Toomre-Q maps and radial profiles
To characterize non-thermal motions in the disk, we decompose particle velocities into radial and azimuthal components in the disk-plane coordinate system. We subdivide the disk into concentric cylindrical annuli and compute the mass-weighted mean velocity in each annulus, yielding radial profiles and . The residual (streaming-subtracted) velocity for each particle is . The velocity dispersion is computed as the root-mean-square of within each annulus.
We calculate the Toomre parameter as:
where is the sound speed, is the radial velocity dispersion, the epicyclic frequency, computed numerically from the angular velocity profile as (so for Keplerian rotation), and the surface density. In hot disks , while in N-body systems (like a galactic disk) . thus includes both thermal and turbulent contributions, since supersonic turbulence provides additional effective pressure support. We measure , , and in radial annuli and construct a pixel-by-pixel map in the same way; the estimators are given in Appendix § 12.
Figure 9 (top) shows face-on maps of the local . At the pre-stellar epoch ( kyr), the gas is broadly unstable with throughout most of the region. By , the central region near the first sink begins to stabilize, while the surroundings remain marginally unstable. At kyr, a clear boundary emerges between a stabilized inner region and an unstable outer envelope. By the time spiral arms have developed ( kyr), a ring of material traces the spiral arms, while the inner disk is stabilized () by the strong velocity dispersion generated near the central sinks.
Figure 9. Top: Face-on maps of the local Toomre parameter at six epochs. The black contour marks . Bottom: Evolution of radial profile of Toomre (heatmap), and sink formations, mergers, and radial positions’ evolution (lines). At early times the gas is globally unstable (); at late times the inner disk is strongly stabilized while the outer spiral arms remain near or below . Radial profile of Toomre at six epochs. The horizontal dashed line marks . At early times the entire region has ; at late times a radial gradient develops, with at AU, falling to at AU. Heavy sinks are the 10 most massive sinks at kyr (sinks that have merged away by then do not count); all other sinks are light and drawn in gray. In the maps, the size of the sink markers scales with sink mass. In the bottom panel, each heavy sink’s trajectory and formation marker have the same color as that sink in Figure 12 and Figure 13. Most massive sink always stays at by definition of the center of our disk.
The radial Toomre profile (Figure 10) quantifies its evolution. At the pre-stellar epoch ( kyr), is below unity inside AU () and close to unity farther out, reflecting a globally unstable collapsing core. As the first sink forms and accretes, the inner rises: by kyr (green), throughout, consistent with a marginally stable state.
Figure 10. Toomre radial profiles time evolution. The disk starts with inside and near the edge, but rises with disk evolution, since gas is removed from the disk because of star formation events. Center of the disk is extremely stable () by the 15 kyr mark, preventing any further star formation in the middle.
By (brown), is mostly stable in the whole disk, with the central region (, where ) being the most stable. At later times a radial gradient develops, with the inner disk more stable (); the outer spiral arms seen at earlier epochs remain near or below . By the latest epochs the disk () is Toomre-stable throughout, and the profile remains near or above unity out to .
What sets Q: thermal versus turbulent support
The shape of these profiles follows directly from the disk kinematics of Figure 6. Using and for the disk mass enclosed at , and eliminating through , Equation (10) can be rearranged into Equation (11):
where as above and . The balance between these two terms in changes with time. Near marginal stability (), thermal pressure dominates (). The radial velocity dispersion only becomes comparable to at and dominates afterwards.
The at early times is easy to explain from Equation (11): at that point the first two terms are (turbulence is sonic/subsonic, and rotational support is weak), and , making the last term 1. The product is well below unity at every radius, which is why the core starts out globally unstable.
After the first sinks form, all three factors from Equation (11) move upward. and (mostly tracing ) grow monotonically. The inner rises slightly as is growing from the infalling gas; rises steadily as the disk spins up; and the stellar mass grows quadratically while the disk gas is depleted by star formation, so climbs by more than an order of magnitude. The extreme central values ( inside ) are dominated by the last of these, since there the enclosed mass is almost entirely stellar and very little gas is left. The outer disk stays near because it remains gas-dominated and its turbulent velocities stay a roughly fixed fraction of the local orbital speed. The stabilization of the disk is therefore kinematic and structural rather than thermal: turbulence rises relative to the orbital speed, the disk spins up, and gas is converted into stars.
Fragmentation, heating fronts, and self-regulation
Figure 9 (bottom) shows the space–time evolution of the azimuthally-averaged Toomre radial profile. The disk begins in a predominantly unstable state (, red) at all radii. The first sink forms at the center, and within the contour (black line) begins to propagate outward from the center, tracing a wave of stabilization driven by heating of the gas, which we speculate is caused by the first sink. A second stabilization wave starts at , together with the formation of a second sink near the center. The leading edge of the first wave, times the sound speed of the gas ahead of it. Both fronts are driven by a rise in the sound speed. We speculate that this gas is heated by the first sink.
The second wave coincides with the formation of the second sink, but the second sink doesn’t start producing radiation until the ”stability wave” has already began traveling through the disk. We speculate that the second wave is also driven mainly by the first sink. Because is then already throughout the inner 1000 AU, the contour of this wave sweeps the whole disk within , while the maximum of moves outward at . Only sinks forming at the center produce such fronts, because later sinks form in gas that has already been heated. In the second wave the gas stays at for a further , while at 150 AU the molecular fraction falls from to 0 at . This is consistent with collisional dissociation by atomic hydrogen in the network, which at these temperatures destroys in much less than 50 yr but, unlike in real gas, does not absorb energy (subsection 2.2.3).
The early stabilization waves are therefore thermally driven. After the inner disk stays stable because accretion onto sinks depletes its gas, while falls back to farther out as gas accumulates. At a spiral arm fragments into stellar seeds within ; many merge soon after (see [27], for the merging criteria), leaving stars. Locking gas into stars lowers the disk mass and stabilizes it.
The outer disk reaches by , coinciding with the second and last bulk star formation event, which converts most of the available disk gas into stars. In the bottom panel of Figure 9, each successive event occurs farther from the center because the inner gas has already been depleted, so stability propagates inside-out.
Star Formation and the protostellar mass function
In this section, we analyze the star formation outcome of the system by quantifying the total stellar mass growth, the instantaneous star formation efficiency within the disk, and the shape of the proto-stellar mass function (PMF). Because the jet feedback in this run is deposited close to the protostars and does not propagate as far as intended (subsubsection 2.2.4), the stellar masses, accretion rates and star formation efficiencies below represent upper limits (§ 6).
Total stellar mass growth and accretion rates
The top panel of Figure 11 shows the evolution of gas, stellar (sink), and total mass within a 0.1 pc sphere centered on the most massive star in the disk. The total stellar mass (orange, solid line) reaches by 15 kyr, growing as (dashed red line). This is consistent with the scaling predicted by the theory of collapse in self-gravitating turbulent fluids (N. W. Murray & P. Chang [67]) and confirmed in high-resolution AMR simulations by E. J. Lee et al. [57]; D. W. Murray et al. [66]. The density inside the stellar sphere of influence evolves to a time-independent attractor , which, combined with infall at a fixed fraction of the Keplerian speed, yields a mass accretion rate and thus .
Figure 11. Top panel: time evolution of the stellar mass (all sinks; orange), disk gas mass (cyan), gas within a 0.1 pc sphere, (navy), and total mass (, dark gray; , light gray). The best-fit power law is shown as a dashed red line. Top-middle panel: three net rates of change between consecutive snapshots: the growth rate of the total sink mass, (magenta); the rate at which the sphere loses gas, (navy); and the rate of change of the disk gas mass, (cyan). Dim lines show the rates at each snapshot, bright lines the same rates smoothed. Where the smoothed rate is negative, its absolute value is shown dotted. These are changes of the mass inside a region, not fluxes through a surface, so is not measured at the disk edge; it is the infall minus the accretion onto sinks. The infall rate through a sphere of radius is shown in Figure 5. Bottom-middle panel: sphere of influence radius (; orange). Bottom panel: star formation efficiency within the disk, (green solid), and within the 0.1 pc sphere, (green dashed), and within the whole halo, (black dash-dotted), where is the gas inside the virial radius (0.46 kpc, physical), held constant because it changes by less than 0.01% over the run, together with the star formation efficiency per free-fall time (blue), , where and is the free-fall time at the mean density of gas and sinks inside the sphere.
The disk gas mass (cyan) grows as infall from the envelope outweighs gas consumed by sink accretion (values in subsection 3.3). Both the infall rate and the total stellar accretion rate are of order – over this period, with infall initially the larger of the two (Figure 5; top-middle panel of Figure 11). It then declines as the star formation efficiency rises and the envelope supply is depleted. The gas mass within the 0.1 pc sphere (navy) rises from to over the first kyr, declines to by kyr as accretion drains it, and then rises again to by 15 kyr as infalling gas enters the sphere. The reason for this is a large amount of infalling gas suddenly entering the sphere, resulting in a positive (i.e. more gas is entering the system than is being converted to stellar mass). This sudden rise in the gas mass also “ignites” the stellar accretion again after its slow-down, driven by the depletion of initial accessible reservoirs in the original disk, and one can see this “bounce-back” in the top and bottom panels of Figure 11: grows slowly, with increasing faster after , while stays roughly constant.
Figure 11 (top-middle panel) shows the corresponding mass rates. The stellar accretion rate (magenta) fluctuates around at early times, rising to at late times. The median accretion rate is , and the disk-wide gas infall rate is . These rates are for 600 K gas, an order of magnitude above the Shu rate (F. H. Shu [91]). They are comparable to the Larson–Penston rates ( in the envelope at core formation, after it; C. F. McKee & E. C. Ostriker [62]) and to the found by A. Stacy et al. [99], as expected from the envelope kinematics (§ 3).
Figure 11 (third panel) shows the sphere of influence radius , defined as the radius where the enclosed gas mass equals the total stellar mass (). This grows from AU at 1 kyr to AU by 10 kyr. It drops slightly with the new gas entering the sphere. The growth of quantifies the outward expansion of the region from which the central cluster can draw its gas: as grows superlinearly, the scale on which the cluster potential dominates the gravitational field expands at AU kyr, roughly as . This is faster than the predicted for a self-similar envelope (N. W. Murray & P. Chang [67], Eq. 42), which would give only AU at 10 kyr. The difference is consistent with the finite disk reservoir being consumed rather than continuously replenished, since the gas reigniting star formation after 10 kyrs only had time to enter the 0.1 pc sphere and did not yet fall into the disk.
Star formation efficiency
We now examine the star formation efficiency, SFE, and the efficiency per free-fall time, . Both are instantaneous quantities, evaluated at each snapshot for a region (the disk, the 0.1 pc sphere or the whole halo), and shown in the bottom panel of Figure 11:
where , and are the gas and stellar (sink) masses in , is the rate at which the stellar mass in grows, and is the free-fall time at the mean density of gas and sinks in .
The SFE remains below unity throughout our run. Within the disk, SFE rises steeply and saturates at by 10 kyr. Within the 0.1 pc sphere, however, SFE only reaches a maximum of by 10 kyr. Over the whole halo the efficiency is only at 15 kyr, where is the gas inside the virial radius (0.46 kpc, physical), which stays constant to better than 0.01% over the run. This is not the case for ,
which exceeds far above unity, and exhibits a sensitive dependence on the star formation rate and ongoing gas accretion. While is only in Galactic clouds and galaxies from the local universe to high redshift (M. R. Krumholz et al. [53]), it can exceed unity locally when a star forming region is being fed by multiple filaments or streamers (see e.g. C. D. Matzner & P. H. Jumper [61]). This is indeed the case here. Between 4 and 10 kyr the median (and before and after); Appendix § 13 shows that this is what the stellar accretion rate of implies for the free-fall time of a Larson–Penston-like envelope.
Protostellar mass function
The PMF uses the current masses of the 32 sinks present at kyr, not their maximum masses. Most low-mass sinks merged early, so only one sink lies below and the PMF at 15 kyr is not skewed by them.
Figure 12 shows the protostellar mass function (PMF) of the sinks at kyr, compared to the present-day P. Kroupa [52] and G. Chabrier [13] initial mass functions (IMFs). The high-mass slope of the PMF is well fitted by a power law with much shallower than the Salpeter slope () and the Kroupa high-mass slope (). A slope of means equal numbers of sinks per logarithmic mass interval, i.e. a flat . We compare this slope with other Pop III simulations in subsection 4.5.
Figure 12. Protostellar mass function of the 32 sinks present at kyr (blue histogram), compared to the G. Chabrier [13] system IMF (dashed) and the P. Kroupa [52] IMF (dotted), both scaled to the number of sinks above . The blue dashed line is the power-law fit above , . Star symbols mark individual sinks: the position is the sink mass and the height is its formation time along the arrow ( at the bottom, 15 kyr at the top), not a value of . Star size grows with mass. Heavy sinks, the 10 most massive sinks at kyr, are colored and light sinks are gray; each heavy sink has the same color as in Figure 9 and Figure 13 (bottom). The gray shaded region lies below the completeness limit ().
The low-mass end of the distribution must be read with care. Because the PMF is a 15 kyr snapshot, many of the lowest-mass sinks have only recently formed or have been ejected from the disk, and have not yet had time to accrete; they populate the gray region below the completeness limit rather than a genuine low-mass tail. The only sink below that limit, of , formed at kyr and was ejected before it could accrete. In Figure 12 the star symbols form two rows, the sinks of the two fragmentation bursts at and kyr, and ten of the 13 sinks below have moved beyond 2500 AU from the most massive sink, after which they barely accrete. None of the sinks have reached the main sequence by the end of our simulation.
In the protostellar model ([69, 27]) all 18 sinks above are in the shell deuterium-burning phase at 15 kyr, the ten sinks of burn deuterium in their cores at fixed core temperature, three of have not yet ignited deuterium, and the ejected sink is still a pre-collapse first core. None has reached the main sequence.
Individual accretion histories and runaway
Figure 13 shows integrated accretion history of individual sinks. A lot of low-mass () sinks stop accreting relatively early because they get ejected out of the disk, thus limiting the amount of gas, available to them. The accretion histories in Figure 13 (with the times at which sinks move beyond 2500 AU marked as ticks at the end of the lines) shows that accretion is active for most of the time for most of the sinks until they leave the disk.
Figure 13. Evolution of masses of individual sinks. Dashed line represents ideal law, which is being followed for total mass of all the sinks (black line). At earlier times, before the first bulk star formation event at 2-3 kyr, all of stellar mass is concentrated in the 2 central sinks (top blue lines). These massive sinks reach mass of around by the 15 kyr mark, with their accretion slowing down compared to less massive sinks, which still accrete rapidly. Small vertical ticks mark the time at which each sink first moves more than 2500 AU from the most massive sink (“far” sinks). The mass history for these far sinks is drawn fainter because their mass doesn’t change by more than 3% afterwards. The lower panel shows the time evolution of the number of all formed sinks, including those that merged later (black, cumulative), of the sinks present at that time (blue), of far sinks (red, cumulative), and of gravitationally unbound sinks (green): sinks beyond 2500 AU from the center of mass of the central cluster with , where includes the gas and the other sinks. A sink is counted as unbound (or bound again) only after the criterion has held for 10 consecutive snapshots ( kyr), with an reference (P. S. Li et al. 2018). Heavy sinks, the 10 most massive sinks at kyr, are colored with the same color as in Figure 9 (bottom) and Figure 12; all other (light) sinks are gray.
Of the 62 sinks formed during the run, 30 (48%) were lost to mergers at low mass: STARFORGE merges two sinks when they are bound with a semimajor axis below the sink radius and the lighter one has less than , to remove spurious sinks that form next to existing ones ([27]).
In our simulation, 14 of the 32 surviving sinks are far ( AU) by 15 kyr. Applying the escape criterion of A. Stacy & V. Bromm [97], (lower panel of Figure 13), only 7 of the 32 survivors are unbound at 15 kyr (). All of them have , as expected for dynamical ejection, which preferentially removes the lowest-mass members of a few-body system.
The unbound count peaks at 8 at kyr, and these sinks are on wide orbits within the envelope rather than escaping. Far sinks typically retain close to their mass at ejection.
Figure 13 also shows that the 2 heaviest sinks are the first 2 sinks that formed in the simulation, both in the
center of the disk. At the same figure it can be seen that the accretion in the center of the disk (heaviest 2 sinks) mostly halts, which is a consequence of depletion of gas reservoir. This gas depletion is shown by a drop in stellar accretion rate in Figure 11. The most massive sink reaches 25 by 15 kyr, but its accretion has slowed to over the final kiloyear as the local gas reservoir is depleted. The “bounce-back”, described in detail in subsection 3.5, seems to mostly target the non-central sinks, since we still observe their growth after kyr.
The second-most massive sink (23 ) is at AU and its accretion has stalled entirely by 15 kyr. It is a far sink, but it is gravitationally bound to the central group of five sinks ( within 200 AU of the primary). Between 12 and 15 kyr its two-body orbit has , a semi-major axis of AU, an eccentricity of and a period of kyr. It is a wide, eccentric companion near apocenter, and it returns to AU at kyr in the extended run. Its accretion stalls because its orbit lies largely outside the gas disk.
Competitive accretion in a cluster environment would preferentially feed the most massive objects (I. A. Bonnell et al. [7]), flattening the mass function slope. To distinguish competitive accretion from simple disk fragmentation, we show in Figure 14 (top) the accretion rate against the current mass for all sinks, with the binned median, compared with the scalings of tidal-lobe (), Bondi–Hoyle () competitive accretion (I. A. Bonnell et al. [7]), accretion from a turbulent core (; C. F. McKee & J. C. Tan [63]) and inertial inflow ( independent of ; P. Padoan et al. [74]). Above the median rises roughly as , between the tidal-lobe and Bondi–Hoyle cases. Discarding the per cent of rates below , from sinks that have almost stopped accreting, flattens the slope to , between the turbulent-core (1/2) and tidal-lobe (2/3) values; the data cannot distinguish between these two. Below it falls with mass, because newly formed sinks accrete fastest just after formation.
Figure 14. Top: accretion rate against current mass for all sinks (gray) with the binned median (black), compared with tidal-lobe () and Bondi–Hoyle () competitive accretion (I. A. Bonnell et al. [7]), turbulent-core accretion (; C. F. McKee & J. C. Tan [63]) and inertial inflow ( independent of ; P. Padoan et al. [74]), normalized to the median (line colors as in the legend of the bottom panel). Bottom: mass of each sink against time since its formation, with the integrated form of each scaling anchored to the median track at 1 kyr. Individual sinks grow at most linearly in time, with no sign of the runaway growth of Bondi–Hoyle accretion.
This ensemble relation mixes sinks of different ages, however: the most massive sinks are the oldest, central ones with the best access to gas. The bottom panel therefore tests the scalings on individual sinks, comparing their mass growth since formation with the integrated form of each model. Individual sinks grow at most linearly in time, close to the constant-rate (inertial-inflow) track and far slower than the runaway growth expected from Bondi–Hoyle accretion. We find no evidence of runaway competitive accretion: each sink grows at a roughly constant or declining rate set by its local gas supply, and growth is limited by competition for a shared, finite reservoir.
Magnetic fields
In this section, we analyze the evolution of the magnetic field as the disk collapses and fragments. The magnetic fields are amplified during the collapse, and within the disk once it forms. We characterize the resulting field morphology and its evolution through the vertical () and toroidal () components.
Field morphology and topology
Figure 15–Figure 16 show face-on and edge-on thin slices of the toroidal and vertical components of the magnetic field at six epochs. In the pre-stellar epoch, shows large-scale coherent structure with a preferred sign, aligned with the imposed initial field orientation. As the disk forms and develops spiral structure, is amplified in the disk midplane and develops sign reversals that trace the spiral arms, with field strengths reaching G at late times.

Figure 15. Thin face-on () and edge-on () slices through the most massive sink, with equal to 5% of the panel width ( AU for the 4100 AU panels) and the field mass-weighted across the slice thickness, of the toroidal magnetic field component at six epochs, with in-plane magnetic field lines overlaid (face-on panels: , ; edge-on panels: , ). The sign pattern of is present before the first sink forms and is inherited from the nearly uniform seed field; differential rotation winds it into the two-armed spiral visible by kyr.

Figure 16. Thin face-on and edge-on slices, as in Figure 15, of the vertical magnetic field component at six epochs. By the late epochs shown here, most of the stars in the simulation have already formed, so the field structure reflects the post-formation disk rather than the initial collapse.
The toroidal component (Figure 15) has a two-tone appearance before the first sink forms ( kyr, top left panel): roughly half of the disk plane has , shown as red, and the other , shown as blue, with the two colors inhabiting clearly distinct regions. Differential rotation winds the two regions into a two-armed spiral pattern of alternating color by kyr, with the field gaining in strength close to the center of the spiral.
By kyr, the toroidal field component reaches strengths of G over a region of size 1000 AU. The field winding due to the differential rotation of the disk has produced a field that is primarily toroidal. This is the expected geometry: winding an in-plane seed produces a toroidal field that is symmetric about the midplane ([10 Sec. 11.5.2]). By late times (15 kyr), the toroidal field in the inner disk is predominantly negative (in our chosen reference frame, where the disk rotates with ).
The alternating sign of between the two arms is similar to the well known in-plane bisymmetric spiral (BSS) configuration seen in galactic magnetic fields; the field reverses direction from one arm to the next across a neutral line in the interarm region ([92, 5, 73]). In our simulation, the reversal is inherited from our choice of a uniform initial magnetic field, i.e., a field with an azimuthal pattern. As long as differential rotation and collapse remain approximately axisymmetric, they amplify the field while preserving the initial pattern [83, 9]. The approximate axisymmetry of the flow holds until kyr.
At later times the spiral arms and the lopsided potential of the off-center sink cluster make the flow non-axisymmetric; the induction term transfers field from the pattern into other azimuthal modes, including an axisymmetric () net toroidal component. We see only weak evidence of growth in the radial field, discussed below, suggesting that this early-time pattern results from linear amplification of the seed field rather than arising from a dynamo [93 Secs. 4.2, 5.1].
W.-T. Kim & E. C. Ostriker [48] showed that magnetic fields modify the growth of non-axisymmetric structure in self-gravitating disks when the fields are dynamically important (). Here throughout (subsubsection 3.3.1), so on energetic
grounds the field is not expected to shape the spiral arms strongly, although only a run without magnetic fields can confirm this. The field does not remain wound up: the ratio of toroidal to radial field stays at rather than growing with the shear, the net toroidal flux has a definite sign set by the more massive spiral arm, and the – relation steepens from to . We cannot say whether this is anisotropic compression, winding or a small-scale dynamo; the evidence and the amplification channels are discussed in Appendix § 14.
Mass-to-flux ratio
We define mass-to-flux ratio per annulus as:
where, in Figure 18, is the mass-weighted mean of in the annulus. This measures the unsigned flux.
Figure 18. Dimensionless mass-to-flux ratio as a function of radius at six epochs. Solid lines use the net vertical flux (volume-weighted signed in each annulus), which sets the magnetic support; dashed lines use the unsigned flux (mass-weighted ), which also counts field reversals. The dotted line marks the critical value . The net-flux ratio falls from – at early times to at late times, whereas the unsigned ratio drops below unity at some inner radii by 9.5 kyr as the field becomes tangled.
We also compute from the volume-weighted signed mean of in each annulus, and its cumulative version . The mass-to-flux ratio (Figure 18) evolves significantly over time.
At early times (), at all radii, indicating an extremely magnetically supercritical state in which gravity overwhelms magnetic support by more than two orders of magnitude. The mass-to-flux ratio then declines steadily as the unsigned vertical field grows relative to : by (red), has dropped to –. By (purple), exhibits large radial scatter, with values ranging from to and dipping below the critical value () at some inner radii. At the latest epoch (, brown), significant portions of the inner disk () have in this unsigned measure.
The net-flux ratio tells a different story. At , within 2500 AU, consistent with the unsigned values because the field is still coherent. By 3.6 kyr the unsigned ratio has dropped to – in individual annuli, while – in the same annuli. At 9.5 kyr and 15 kyr, and , respectively. In individual annuli (solid lines in Figure 18) the net ratio never falls below , reached in the innermost annuli at 9.5 kyr, where little gas is left. The unsigned flux inside 2500 AU exceeds the net flux by a factor of at early times and at late times.
The net ratio (solid lines in Figure 18) falls by one to two orders of magnitude as time increases. In ideal MHD, neither compression nor winding can do this: compression conserves the mass-to-flux ratio of a flux tube, and winding converts into without changing . The decline instead comes from the flux budget of the disk region. Between and 15 kyr the net flux inside 2500 AU grows by a factor of as the inflow drags field inward, while the gas mass there falls by a factor of as is accreted onto sinks. Sink accretion removes the accreted gas together with its field, but the accreted gas is the densest, and since with (Figure 17) the flux per unit mass, , is smallest there. Accretion therefore removes mass much more efficiently than flux. (In the sink implementations of S. Hirano & M. N. Machida [30] and A. Stacy et al. [101] accretion removes the mass but not the flux, which lowers even more strongly.)
Figure 17. Evolution of phase plot of magnitude of magnetic field vs number density . Solid lines correspond to mass-weighted median profiles at different times, with the shaded regions spanning the 16th–84th percentiles. corresponds to the exponent in power-law fit . spans 3 orders of magnitude at the point of first sink formation, and up to 4 orders of magnitude at later times. The best-fit power-law slope rises from at the pre-stellar epoch to (first sink), 0.69 (early disk), 0.76 (GI onset), 0.92 (late time) and 1.00 (extended), i.e. from (flux-freezing compression) toward at late times; each fitted slope is given in the legend and drawn as a dotted line.
The much larger growth of the unsigned flux requires some other growth mechanism, for example, vertical motions that tilt the strong toroidal field into of both signs, as at spiral-arm shocks, and possibly a turbulent dynamo.
Comparison to previous studies
The collapse and subsequent disk fragmentation picture in this work is consistent with the broad expectations of Pop III star formation theory. In this section, we compare our results with other works and highlight differences.
Role of the large-scale environment
The properties of the first collapsing core are set by the large-scale accretion history of the host minihalo, which in turn depends on where that halo sits within the cosmic web. A halo embedded in a stronger overdensity accretes faster and with more angular momentum than one in a quieter region of the cosmic web, biasing the collapse toward a more massive, rotationally supported structure. Our halo has a mass of when the first star forms, which is above the atomic-cooling threshold. The halo is a progenitor of a MW-mass system (with at ) and sits in a large-scale overdensity, unlike the field minihalos studied in most prior work.
Our halo is a typical Pop III host in this kind of environment: in the MEGATRON simulations [103] of a MW progenitor, the very first Pop III stars form in minihalos, but most form above the atomic-cooling threshold, in halos of –, once a Lyman–Werner background is established [103].
The most directly comparable simulation, Y. J. Meziani et al. [64], targets the dwarf galaxy progenitor halo, and finds a single Pop III protostar growing to over 31 kyr – in stark contrast to our -fragment cluster. The two runs differ in more than environment, however: that star forms at in a minihalo of (), with times finer DM resolution, but similar baryonic resolution, and in a disk that stays Toomre-stable () with jets active from the start. We speculate that the difference in the outcome is partially dictated by the environment in which the gravitational collapse is happening.
The overdense filamentary infall geometry of the MW-progenitor naturally seeds higher angular momentum, more complex disk structure, and stronger turbulence than the quieter m09 environment. M.-Y. Ho et al. (2026) showed that infall into dark-matter potential wells drives supersonic turbulence (–) in 15 minihalos, with increasing with halo mass.
In our simulation, at – pc, the velocity dispersion about the mean infall gives at all times, similar to the value seen in the slightly less massive () halo in M.-Y. Ho et al. (2026). Within the disk in our simulation, before kyr and after the first fragmentation burst (Figure 6).
Fragmentation and sink counts
Comparing sink counts across simulations is difficult due to differences in the sink formation and accretion algorithm used, the spread in integration times, or the stochastic nature of large simulations. Early simulations with cosmological initial conditions found 4–5 protostars within yr in one halo [15] and of order ten per halo after 1 kyr in five halos [26]. Single-halo integrations over kyr found 5–8 sinks [100], whereas A. Stacy et al. [98] formed 102 sinks within 2 kyr, and O. Jaura et al. [45] formed 40–150 sinks over 20 kyr. The radiation-hydrodynamic cosmological zoom-ins of M. A. Latif et al. [56] across five minihalos produced up to 23 stars per halo (55 total), with masses – and a slope . The systematic study of K. M. J. Wollenberg et al. [115] with 45 controlled simulations (varying rotation and turbulence) found 15 to sinks (mean 30–59 per setup) within kyr, without sink mergers. Thus our sink formation events over the first kyr fall within the range of other high-resolution runs, confirming that fragmentation is highly stochastic and sensitive to initial conditions.
Several factors contribute to our fragment count. First, our 15 kyr integration is substantially longer than most published runs at similar resolution in cosmological environment, allowing many cycles of instability, collapse, and re-accretion. Second, cosmological initial conditions from the MW progenitor provide asymmetric, turbulent infall that continually replenishes the outer disk and seeds new unstable structures.
Of the 62 sinks formed, 30 (48%) were lost to mergers, and 14 of the 32 survivors (44%) have moved more than 2500 AU from the primary, of which 7 ( of the survivors) are unbound by the criterion. The merger fraction is in line with [97] (59%) and [98] (64%); the unbound fraction ( of survivors) is about half of their escaping fractions (43% and 31% of survivors), while the fraction beyond 2500 AU (44%) is comparable. Other works report merger fractions of –50% depending on the merging criterion [115] and ejection fractions of up to [56].
Stellar mass growth and accretion rates
The total stellar mass assembled by a given time is the time-integrated sum of the accretion rates of all active sinks, so it reflects both the efficiency of angular momentum transport (which determines how quickly gas drains inward through the disk) and the number of competing accretion regions (which determines how the common reservoir is partitioned among the protostars). A large sink count spreading the accretion over many objects can produce a similar total mass to a single rapidly accreting protostar while leaving a completely different PMF. We compare our integrated mass growth and accretion rates to prior work to disentangle these effects.
The total stellar mass of at 15 kyr is broadly consistent with canonical Pop III accretion rates. A. Stacy et al. [99] found onto the star-disk system over 5 kyr, similar to our time-averaged rate, and a main sink at 5 kyr growing as . Single-star growth in other runs is similarly sub-linear: P. Sharda & S. H. Menon [89] find . Our scaling is consistent with that found by D. W. Murray et al. [66] and P. S. Li et al. (2018), and is a property of the whole cluster, not of individual sinks, which indeed grow sub-linearly (Figure 13).
The POPSICLE simulations of P. Sharda & S. H. Menon [89] produce a single most-massive protostar of within 5 kyr in their RMHD run; our total mass of is distributed across surviving sinks at 15 kyr, implying a mean sink mass of , consistent with fragmentation-induced starvation redistributing mass from the most massive object to the cluster population ([78, 80, 89]). K. Sugimura et al. [104] and K. Sugimura et al. [105] produced massive Pop III binaries with individual masses at separations AU. In their runs the first protostar’s disk fragments, most fragments merge in few-body encounters, and only 2–4 stars survive. They accrete from a shared circumbinary disk, which drives the mass ratio towards unity and widens the orbit, because the accreted gas has high angular momentum. Radiative feedback ends accretion after yr, at final mass ratios – and separations of – AU (K. Sugimura et al. [105]).
Our two most massive sinks ( and , ) also form a bound pair. After kyr, the secondary orbits the central group on a wide, eccentric orbit ( AU, ), with a separation similar to that of the K. Sugimura et al. [104] binary at yr, when it has widened from AU to a few AU. Unlike their pair, our secondary has stopped accreting because its orbit carries it outside the gas disk, and the separation was set by few-body dynamics rather than by circumbinary accretion. Without radiative feedback, and with a different sink radius and merger criterion, our pair is therefore at most a qualitative analogue of their early binary phase.
Star formation efficiency
The star formation efficiency we find (SFE within the disk and within a pc sphere at 15 kyr) is high compared to present-day values (–), but these are typically measured for entire giant molecular clouds rather than individual protostellar disks, so this comparison is only qualitative. Other Pop III simulations measure the efficiency across the halo. Over the whole halo our efficiency is at 15 kyr (Figure 11), below the – of S. Hirano et al. [31], who divide the core mass by the gas mass of the virial halo.
Protostellar mass function and runaway fraction
A top-heavy, roughly log-flat PMF is commonly found in numerical studies of Pop III star formation, although most of them, like ours, follow only the first – yr. Several authors attribute it to fragmentation together with mergers and ejections (e.g. [98, 56]). Two effects favor a top-heavy PMF: the primordial gas temperature ( K) sets a large Jeans mass at the fragmentation scale, seeding comparatively massive protostars ([11]), and low-mass objects are preferentially ejected or starved, which flattens the slope at the low-mass end ([115]).
Our protostellar mass function (PMF) at 15 kyr has a high-mass slope , shallower than the Salpeter slope () but broadly consistent with other Pop III simulations that find top-heavy, log-flat distributions. A. Stacy & V. Bromm [97] find and M. A. Latif et al. [56] find across five halos with RHD; A. Stacy et al. [98] and T. H. Greif et al. [26] both report roughly log-flat mass functions extending from sub-solar to ( in A. Stacy et al. [98]). These slopes drift with time and resolution ([98, 80]).
We emphasize that our PMF is a snapshot at 15 kyr, not a final IMF: all sinks are still actively accreting, no radiative feedback has terminated accretion, and the population will be further shaped by continued mergers, ejections, and eventual UV feedback ([43, 42, 98]).
The two most massive sinks, a bound wide pair, have each reached by 15 kyr, about half of the mass at which accretion was halted by feedback in the cloud of T. Hosokawa et al. [43]. At late times, their accretion rates flatten due to local gas depletion in the inner disk. The outer sinks, by contrast, continue to grow by accreting from the surrounding envelope, so the cluster mass budget shifts progressively outward.
The Pop III IMF is not necessarily universal: its slope and peak can be modified by the large-scale environment, the initial turbulence and angular momentum, and the strength and geometry of the magnetic field, so different halos may produce different mass distributions ([115, 30, 31, 89]). With the same physics setup, albeit with a smaller halo mass, Y. J. Meziani et al. [64] form just a single protostar. This indicates that the large-scale environment is a key driver of the Pop III PMF. Y. J. Meziani et al. [65] simulate an ultra-faint dwarf progenitor from to (about 500 Myr) and form 56 stars across five minihalos. Although the slope of the IMF for all the stars taken together is , their most massive, first collapsing halo has a slope .
Magnetic field evolution
The amplification of primordial seed fields during the collapse and disk phase is one of the most uncertain aspects of Pop III star formation: the initial field strength is poorly constrained, and the relevant dynamo mechanisms (turbulent small-scale, large-scale differential rotation) operate on overlapping timescales. We compare the evolution of our mass-to-flux ratio and magnetic energy to MHD simulations that resolve the disk phase, to assess whether the rapid amplification we observe is generic or a consequence of our particular choice of seed field and numerical method.
Our net-flux mass-to-flux ratio falls from – at early times to by 15 kyr (subsection 3.6), and the magnetic energy grows by orders of magnitude. Despite this amplification, disk fragmentation in our simulation remains vigorous until kyr, in contrast to the strong suppression found in some MHD simulations. M. N. Machida et al. [60] systematically varied initial field strength from to G and found that even the weakest field suppresses fragmentation and drives protostellar mergers within the first yr, once it has been amplified (weak-field models fragment first and merge afterwards). A. Stacy et al. [101] found only two secondary sinks in cosmological MHD runs. A key difference is that our simulation starts from a cosmologically plausible highly supercritical configuration (). We have not measured the magnetic torques directly. How much the field changes the fragmentation and the disk evolution can only be measured against an otherwise identical run without magnetic fields, which we leave to future work.
K. E. Sadanari et al. [85] found that ambipolar diffusion is dynamically subdominant in primordial gas (diffusion heating always smaller than compressional heating), broadly consistent with our use of ideal MHD (Y. J. Meziani et al. [64]). K. E. Sadanari et al. [86] similarly found in primordial disk simulations that turbulent magnetic fields reduce but do not eliminate fragmentation, consistent with the vigorous fragmentation of our magnetized disk.
Caveats
Several limitations of this calculation should be kept in mind when reading the results above.
Cosmic-ray background. The cosmic-ray ionization and dissociation rate entering the H network is left at its default, MW-like value rather than being scaled to a primordial environment. Cosmic rays destroy H, but the free electrons they supply also increase H and HD formation through the H channel. The net effect of the MW-like rate is to make the K gas that contracts slowly before runaway collapse more molecular and colder, and the lowest-density gas hotter, than it would be at the much lower rate expected at (D. Nakauchi et al. [68]; J. A. Hummel et al. [44]). Across a wide range of rates, the thermal path of the collapsing gas converges to that without cosmic rays once runaway collapse begins (J. A. Hummel et al. [44]). The cosmic-ray rate is therefore unlikely to cause the offset from the primordial – track at (subsection 3.1), which is more likely set by the reduced chemistry network (see below).
Chemistry network. The primordial chemistry is followed with the network described in subsection 2.2.3 rather than with a dedicated package such as GRACKLE, and H dissociation is not coupled self-consistently to the radiation transport. The network also omits the chemical heating and cooling associated with H formation and dissociation, and ties HD to H by a fixed ratio. The three-body rate is a single prescription, whereas published fits are uncertain by nearly an order of magnitude and can change the disk accretion and fragmentation (M. J. Turk et al. [111]). Our gas is 0.4 dex colder than the primordial track, which lowers the Jeans mass by a factor of . Raising the temperature to the primordial track (a factor of 1.8–3.2) would increase by during the first fragmentation burst, when thermal pressure supplies most of the support (subsection 3.4), and by after kyr, when turbulence dominates. The disk would still fragment, but this bias would make our sink count an upper estimate, and the initial fragment masses may be low by a factor of a few. The total stellar mass is less affected, because it is set by the envelope inflow at densities below those where three-body H formation heating matters. The molecular fractions and temperatures we quote should therefore be treated as indicative rather than converged.
Magnetic field limitations. We cannot establish whether a small-scale turbulent dynamo is resolved in this calculation. The – relation and the field normalization (subsection 3.6) cannot separate anisotropic compression, winding and dynamo amplification, and a resolved dynamo could add to the field strengths we report. In addition to that, the current STARFORGE implementation removes the magnetic field of a gas particle when a sink particle is spawned, which is not a realistic prescription.
Parent-simulation resolution. The epoch () and host (, an atomic-cooling halo) of the first collapse are set by the parent FIRE-resolution run, before the hyper-refinement is switched on. With , H-cooling minihalos of – are resolved by only DM particles, and the gas-only Jeans refinement cannot compensate for this. In MW-progenitor regions the very first Pop III stars form earlier, in minihalos, but most form in atomic-cooling halos like ours once a Lyman–Werner background is established (§ 5; A. Storck et al. [103]). Our results therefore describe the first collapse in an atomic-cooling halo, and the progenitor minihalos, their Lyman-Werner radiation and any pre-enrichment are missing.
Protostellar radii and accretion luminosity. The protostellar model (S. S. R. Offner et al. [69]) gives our central sinks radii of only while they grow from to during the first 1.5 kyr. Pop III protostars accreting at – are instead swollen to at –, (S. W. Stahler et al. [102], Eq. 21b), as in the protostellar tracks of K. Omukai & F. Palla [72]. Since the accretion luminosity scales as , it, and with it the early heating fronts of subsection 3.4, is likely overestimated by a factor of .
Integration time. We follow only kyr after the first protostar forms. All the sinks are still accreting when the calculation ends, no object has reached the main sequence (all sinks above are in the shell deuterium-burning phase of the protostellar model; subsection 3.5), and the ionizing feedback that would eventually terminate accretion has not had time to act. The mass function we report is protostellar, and the final stellar masses could differ substantially. Fragmentation in the disk has stopped by kyr, while the central sinks keep accreting, so in the near term the high-mass end should grow further. Over longer times the mass function is expected to steepen, as in the population run of Y. J. Meziani et al. [65] (subsection 4.5).
Protostellar jets. The jets are launched with the intended mass loading but couple to the gas close to the protostars, so they propagate less far than intended and entrain and clear little of the infalling envelope (sub-subsection 2.2.4). Jets matter for the stellar masses: in STARFORGE simulations of present-day clouds they reduce the stellar masses by about an order of magnitude, both by removing accreted gas and by disrupting the accretion flow onto the protostars (D. Guszejnov et al. [28]). In the FORGE’d in FIRE Pop III protostar of Y. J. Meziani et al. [64], jets regulate the accretion and the final mass of the star. With jets that couple only close to the star, the part of the accreted gas they remove is still taken out, but the disruption of the larger-scale accretion flow is underestimated, so our protostellar masses, accretion rates and star formation efficiency are therefore overestimates, and the PMF may be shifted to lower masses.
Conclusions
We have presented a radiation-magnetohydrodynamic cosmological zoom-in simulation of Pop III star formation, following the collapse of primordial gas in the FIRE MW-progenitor halo from cosmological initial conditions through the formation and evolution of a protostellar accretion disk over kyr. Our main conclusions are as follows:
Vigorous disk fragmentation. The accretion disk becomes gravitationally unstable within kyr of the first protostar’s formation, developing spiral arms that fragment into sink formation events over the first kyr, producing a total stellar mass .
Sub-virial start to virial equilibrium. The disk starts in a sub-virial state () and reaches virial equilibrium within kyr, thereafter self-regulating around , robustly across disk identification criteria.
Magnetic field amplification and tangling. Magnetic energy grows by orders of magnitude. The net-flux to flux ratio falls from – to , because the inflow brings in flux while the dense gas accreted onto sinks carries little flux per unit mass, while the unsigned vertical flux grows times more than the net flux. The field becomes tangled into mixed polarity, with reversing along the spiral arms.
Accelerating accretion and high SFE. The total stellar mass grows as at accretion rates of –, an order of magnitude above the Shu rate and close to the Larson-Penston rate of an outside-in envelope, which falls in at a nearly constant . The star formation efficiency reaches within the disk (after kyr; at kyr) and within a pc sphere.
Top-heavy protostellar mass function. The protostellar mass function has a high-mass slope , i.e. roughly flat in , much shallower than the Salpeter/Kroupa value () and similar to other Pop III simulations (–). Individual sinks grow at most linearly in time, with no sign of runaway (Bondi–Hoyle) competitive accretion; the most massive sinks are the oldest, central ones. The second most massive sink is on a wide, bound, eccentric orbit.
The cosmological envelope sets the mass supply. The cosmological environment sets the mass supply, its duration and its angular momentum: a Larson-Penston-like inflow of through AU for most of the run (falling to after kyr), whose centrifugal radius stays at AU throughout, with a broad spread.
Fragmentation moves outward. In our simulation the fragmentation sites move outward: the first burst ( kyr, sinks) forms at AU and the second ( kyr, sinks) at AU, outside where most of the infall lands. The gas mass of the disk moves outward with them, the sinks empty the inner disk while the outer disk keeps gaining mass.
The results presented here are for a single realization of collapse in a massive halo at , and are therefore subject to that caveat. Pop III star formation is highly stochastic, so the fragment counts and the shape of the protostellar mass function carry an uncertainty that a single run cannot quantify. We leave a population study of Pop III collapse from a large-scale cosmological environment to future work.
Acknowledgments
We thank Shyam Menon, James Beattie, Bart Ripperda, Chris Thompson, Piyush Sharda, and several members of CITA for useful discussions, as well as SciNet maintenance and support teams for help with the usage of Trillium supercomputer. Analysis was performed on the Trillium supercomputer at the SciNet HPC Consortium. SciNet is funded by Innovation, Science and Economic Development Canada; the Digital Research Alliance of Canada; the Ontario Research Fund: Research Excellence; and the University of Toronto. VP thanks the CCA at the Flatiron Institute for hospitality while a portion of this research was carried out. The Center for Computational Astrophysics at the Flatiron Institute is supported by the Simons Foundation. The Flatiron Institute is a division of the Simons Foundation. The computations in this work were, in part, run at facilities supported by the Scientific Computing Core at the Flatiron Institute, a division of the Simons Foundation. NM and VP were supported in part by the Natural Sciences & Engineering Research Council of Canada (NSERC) [funding reference number RGPIN-2023-04901]. Cette recherche a été financée par le Conseil de recherches en sciences naturelles et en génie du Canada (CRSNG), [numéros de référence RGPIN-2023-04901]. It was performed in part at Aspen Center for Physics, which is supported by National Science Foundation grant PHY-2210452. CAFG was supported by NSF through grants AST-2108230, AST-2307327 and AST-2606015; by NASA through grants 80NSSC22K0809, 80NSSC22K1124 and 80NSSC24K1224; by STScI through grant JWST-AR-03252.001-A; and by BSF through grant #2024262. Support for P.F.H. was provided by a Simons Investigator grant.
Software: GIZMO, NumPy, SciPy, Matplotlib, Astropy, h5py, yt, Meshoid
Appendix
H2 network coefficients
The rate equation for the H abundance is given in Equation (4). Here is volume-averaged number density of inside of a cell. is the th-order clumping factor. is the gas-phase formation rate coefficient via the H and H pathways [1, 22, 24]. The three-body formation coefficient [40], which is a single prescription that differs from the fits of T. Abel et al. [2] and F. Palla et al. [75], becomes the dominant H formation channel for . is the coefficient associated with LW and ionizing photon dissociation. accounts for cosmic-ray-induced ionization and dissociation (the code uses a fixed cosmic-ray energy density of , giving and in gas denser than times the cosmic mean, a MW-like value; this rate is unconstrained at , where the physically expected value is far lower in the absence of supernovae. Besides destroying H, cosmic-ray ionization supplies free electrons that catalyse H (and HD) formation through the H channel, so a rate this high makes the gas that contracts slowly before runaway collapse more molecular and colder (and can move into onto the HD-cooling track), while CR heating warms the lowest-density gas [96, 68, 44]. These differences are erased as the gas collapses: once runaway collapse begins, its thermal path converges to that without CRs [44]; see § 6. The last sum runs over collisional dissociation contributions from , H, H, H, HeI, HeII, with rate coefficients from [23 Table A1, reactions 7–11, 24–25 and 37–39].
Disk identification
To compute velocity dispersion, local Toomre , and other disk properties, we identify gas particles belonging to the disk at each snapshot using the following procedure:
We define a search radius , chosen because it roughly matches the size of the disk that forms in our simulation. This radius sets the region over which the bulk velocity and angular momentum are measured, and we adopt it as the disk scale throughout this work: the disk we identify below is approximately equal to .
Before any sinks form, the center is placed at the densest gas particle within a refinement region. After the first sink forms, we place the disk center at the most massive sink particle. We find that this approach results in a
center location that is more stable than using than the mass-weighted center-of-mass (COM) of all sinks once secondary sinks form at large separations.
We compute a mass-weighted COM velocity from gas particles within and subtract it from all particle velocities.
We compute the disk normal from the total angular momentum summed over particles within , using the COM-subtracted velocities from the previous step. The sign of is fixed so that to prevent frame-to-frame flipping.
We convert particle positions and velocities onto the disk-plane coordinate system .
We apply geometric and density cuts: AU, (aspect ratio), and .
We apply kinematic cuts: (co-rotating) and , where is the Keplerian speed computed from the cumulative enclosed gas mass plus the total sink mass placed at the origin.
We extend the kinematically selected set to its bounding cylinder (100th percentile in and ) to recover any disk material just outside the strict kinematic boundary.
Particles passing all cuts are labeled as disk members. The disk radius is taken as the 90th-percentile cylindrical radius of disk members.
We tested this selection at nine epochs with eleven variants. The disk mass changes by for aspect ratios of 0.2–0.5, for –0.5, for a center at the center of mass of gas and sinks, and for an iterated disk axis. It depends more strongly on the bounding-cylinder step and on the density threshold: dropping the former, or raising the latter to , lowers by up to a factor of at late times, when the disk is small. The star formation efficiency within the disk is robust to all of these choices ( at 9.5 kyr and 0.94 at 15 kyr).
Toomre-Q estimator
For each snapshot the gas is binned into 20 annular bins. In each annulus we compute mass-weighted face-on density:
with and corresponding to the inner and outer radii of the annulus respectively. The radial velocity dispersion for each gas particle is measured as its deviation from the annulus average :
The angular velocity is measured directly from the rotation profile:
where is the mass-weighted mean azimuthal velocity in annulus and is the bin center. The epicyclic frequency is computed numerically from the measured rotation curve (e.g. [108]),
The Toomre parameter per annulus is then
A pixel-by-pixel face-on map is produced by computing local mass-weighted quantities for all of the relevant parameters:
where , and , which sums over thermal and turbulent contributions.
Star formation efficiency per free-fall time
We make the following assumptions. (i) The pre-stellar envelope is spherically symmetric, with a power-law density profile out to the radius considered. The sound speed is not constant in the envelope: in Figure 6 it rises outward and stays close to a constant fraction of the Keplerian speed, – (– between and AU before the first sink forms). For the enclosed mass is , so and hence are then constant with radius, and for ; gives the found below. We use only through the product , which is set by the measured density normalization, so does not depend on the value of . (ii) The normalization of the profile is fixed by the gas mass inside our 0.1 pc sphere before the first sink forms, which gives below, i.e. between the singular isothermal sphere () and the Larson–Penston value. (iii) The gas mass in the sphere is taken to stay equal to this envelope mass, so that follows from the stellar accretion rate alone. (iv) The free-fall time is that of the gas alone, whereas the measured also includes the sinks in ; the difference is quantified at the end of this appendix.
Approximate the envelope as a power law , written relative to the singular isothermal sphere (SIS) of sound speed as
where for the SIS and for the Larson–Penston envelope (F. H. Shu [91]). The mass and the mean density inside are
so the free-fall time at that mean density is
The mass inside divided by its own free-fall time, i.e. the rate at which this gas would turn into stars if all of it collapsed within one free-fall time, is then the same at every radius,
and the efficiency per free-fall time of a sphere of radius in the envelope is
Before the first sink forms, the envelope holds inside our 0.1 pc sphere, where (Figure 6), i.e. , so that (this is for ; for the SIS it would be ). Between 4 and 10 kyr the stars accrete at ( for ), which gives with the free-fall time of the gas alone. Including the sinks in , as in our definition, shortens and gives the measured median over this period ( before 4 kyr and after 10 kyr).
Magnetic field amplification channels and net toroidal flux
The field does not remain wound up for long; by kyr the field shows little order. Differential rotation, , winds a radial field into a toroidal one that grows linearly in time, (A. Brandenburg & K. Subramanian [10], Sec. 3.3), so after a time the toroidal field should exceed the radial one by a factor .
We measure the ratio of the two in midplane annuli at , 500 and 800 AU as the ratio of their mass-weighted rms values, , which avoids the divergence of where changes sign. This ratio stays at , and the mass-weighted median pitch angle at , at all epochs, whereas winding since would give at 550 AU by 2.8 kyr and later. Winding also brings field lines of opposite polarity together (A. Brandenburg & K. Subramanian [10], Sec. 6.5): the radial wavenumber of the wound field grows as , so reversals are apart. This spacing stays larger than the local resolution until kyr at 300 AU, so numerical reconnection cannot explain the deficit at early times. This suggests that the radial field is instead continually regenerated by the turbulent, gravitationally driven motions of the disk, whose velocity dispersion is comparable to the rotation speed, as in a gravoturbulent dynamo, in which vertical circulation around the spiral arms regenerates the poloidal field (H. Deng et al. [17]), but we have not confirmed this.
The net sign of is not set by the winding, which by symmetry leaves the azimuthal average near zero. It is instead set by the uneven, one-sided mass distribution of the disk, which the star formation traces. The denser, strongly arm is the one in which most of the stars form (see the lower left panel at of the field-morphology figure in subsection 3.6); that arm is also where compression has made the field strongest. Because the field is frozen into the gas, the inflow of this arm carries its negative toroidal flux into the inner disk, while the positive flux of the lighter arm remains at larger radii. Opposite-signed flux brought together reconnects, and the net flux that remains in the inner disk is negative; the sign of the field is simply that carried by the spiral arm that happened to become more massive.
We have verified that the dense gas does carry most of the negative toroidal field: inside 2500 AU the mass-weighted exceeds the volume-weighted one by a factor of –.
The redistribution of magnetic flux, and the reduction of toroidal flux by reconnection, is, however, not the whole story. The net toroidal flux through the meridional plane stays negative out to AU (not shown in the figure), with , instead of being balanced by positive flux at larger radii; we conclude that a net () toroidal field is also being generated, by winding of a net radial field, by the disk’s turbulent motions, or by some other mechanism.
Figure 17 shows the evolution of as a function of for each of the 6 epochs. We fit a power law at each epoch. The power-law exponent rises from at early times, consistent with flux-freezing compression of spherically collapsing gas, toward at late times. The value is expected when compression becomes anisotropic and the field is stretched in the plane of the forming disk. However, kinematic-dynamo calculations give similar slopes, (see the summary in S. Hirano & M. N. Machida [30]), so the – relation alone cannot separate anisotropic compression, winding and dynamo amplification. The magnitude of up to the time the first star forms is consistent with flux freezing of the initial seed field. Assuming flux freezing, the cosmological expansion gives at the mean density at (the redshift when the disk forms), which is what the low-density gas of the parent simulation has. Isotropic compression from the mean baryon density () to then gives (subsection 2.2.1), close to the measured at that density before the first sink forms (Figure 17). Compression alone therefore can account for the field strength at the onset of star formation.
The later steepening towards may come from anisotropic compression, winding or a dynamo.
The alignment of in-plane flow and field lines does not allow us to determine whether a dynamo is acting. The median rises from 0.69 before the first sink forms, which is the value for randomly oriented in-plane vectors (), to 0.94 once the spiral arms are established. In a rotating disk, winding aligns with whatever the field strength. The ratio , the plasma or the Alfvén Mach number would be the appropriate diagnostics; at late times near the center (Figure 8), so the field is not deeply kinematic there. We cannot demonstrate that a small-scale dynamo is resolved, but it is absent (see § 6), and neither the measured nor the flow–field alignment can decide whether compression and shear alone account for the amplification.
This caution is reinforced by the scheme dependence of magnetic amplification reported by M. I. Ugalino et al. [112]. Evolving identical Milky-Way-like galaxy initial conditions with Eulerian (ART, RAMSES) and (quasi-)Lagrangian (AREPO, SWIFT) MHD codes, they found that the two families differ by orders of magnitude in their amplification histories: the lower numerical diffusivity of the Lagrangian schemes allows small-scale motions to persist and the field to grow to near equipartition, whereas the grid-based codes remain close to the shear-driven scaling. Because the meshless finite-mass scheme used here is Lagrangian, the same tendency could inflate the field strengths and lower the mass-to-flux ratios we measure relative to grid-based calculations, although that comparison was made for galactic disks and did not include GIZMO.
References
- [1]Abel, T., Anninos, P., Zhang, Y., & Norman, M. L. 1997, New Astronomy, 2, 181, doi: 10.1016/S1384-1076(97)00010-9
- [2]Abel, T., Bryan, G. L., & Norman, M. L. 2002, Science, 295, 93, doi: 10.1126/science.1063991DOI
- [3]Bagla, J. S. 2002, Journal of Astrophysics and Astronomy, 23, 185, doi: 10.1007/BF02702282DOI
- [4]Barnes, J., & Hut, P. 1986, Nature, 324, 446–449, doi: 10.1038/324446a0
- [5]Beck, R. 2015, A&A Rv, 24, 4, doi: 10.1007/s00159-015-0084-4DOI
- [6]Boley, A. C., Hartquist, T. W., Durisen, R. H., & Michael, S. 2007, ApJL, 656, L89, doi: 10.1086/512235DOI
- [7]Bonnell, I. A., Bate, M. R., Clarke, C. J., & Pringle, J. E. 2001, MNRAS, 323, 785, doi: 10.1046/j.1365-8711.2001.04270.xDOI
- [8]Bonnor, W. B. 1956, Monthly Notices of the Royal Astronomical Society, 116, 351, doi: 10.1093/mnras/116.3.351
- [9]Brandenburg, A., & Ntormousi, E. 2023, ARA&A, 61, 561, doi: 10.1146/annurev-astro-071221-052807DOI
- [10]Brandenburg, A., & Subramanian, K. 2005, Physics Reports, 417, 1
- [11]Bromm, V., Coppi, P. S., & Larson, R. B. 2002, ApJ, 564, 23, doi: 10.1086/323947DOI
- [12]Bryan, G. L., & Norman, M. L. 1998, The Astrophysical Journal, 495, 80, doi: 10.1086/305262DOI
- [13]Chabrier, G. 2003, PASP, 115, 763, doi: 10.1086/376392DOI
- [14]Chen, K.-J., Ho, M.-Y., & Tung, P.-C. 2025, ApJL, 988, L67, doi: 10.3847/2041-8213/adf18dDOI
- [15]Clark, P. C., Glover, S. C. O., Smith, R. J., et al. 2011, Science, 331, 1040, doi: 10.1126/science.1198027DOI
- [16]Cummings, A. C., Stone, E. C., Heikkila, B. C., et al. 2016, The Astrophysical Journal, 831, 18, doi: 10.3847/0004-637X/831/1/18DOI
- [17]Deng, H., Mayer, L., & Latter, H. T. 2020, The Astrophysical Journal, 891, 154, doi: 10.3847/1538-4357/ab77b2DOI
- [18]Deng, H., Mayer, L., Latter, H., Hopkins, P. F., & Bai, X.-N. 2019, The Astrophysical Journal Supplement Series, 241, 26, doi: 10.3847/1538-4365/ab0957DOI
- [19]Ebert, R. 1955, Zeitschrift für Astrophysik, 37, 217
- [20]Faucher-Giguère, C.-A. 2020, Monthly Notices of the Royal Astronomical Society, 493, 1614, doi: 10.1093/mnras/staa302DOI
- [21]Federrath, C., Sur, S., Schleicher, D. R. G., Banerjee, R., & Klessen, R. S. 2011, The Astrophysical Journal, 731, 62, doi: 10.1088/0004-637X/731/1/62DOI
- [22]Galli, D., & Palla, F. 1998, A&A, 335, 403
- [23]Glover, S. C. O., & Abel, T. 2008, MNRAS, 388, 1627, doi: 10.1111/j.1365-2966.2008.13224.xDOI
- [24]Glover, S. C. O., & Jappsen, A.-K. 2007, The Astrophysical Journal, 666, 1–19, doi: 10.1086/519445DOI
- [25]Gnedin, N. Y., & Draine, B. T. 2014, The Astrophysical Journal, 795, 37, doi: 10.1088/0004-637X/795/1/37DOI
- [26]Greif, T. H., Springel, V., White, S. D. M., et al. 2011, ApJ, 737, 75, doi: 10.1088/0004-637X/737/2/75DOI
- [27]Grudić, M. Y., Guszejnov, D., Hopkins, P. F., Offner, S. S. R., & Faucher-Giguère, C.-A. 2021, Monthly Notices of the Royal Astronomical Society, 506, 2199–2231, doi: 10.1093/mnras/stab1347DOI
- [28]Guszejnov, D., Grudić, M. Y., Offner, S. S. R., et al. 2022, MNRAS, 515, 4929, doi: 10.1093/mnras/stac2060DOI
- [29]Higashi, S., Susa, H., Federrath, C., & Chiaki, G. 2024, ApJ, 962, 158, doi: 10.3847/1538-4357/ad2066DOI
- [30]Hirano, S., & Machida, M. N. 2022, ApJL, 935, L16, doi: 10.3847/2041-8213/ac85e0DOI
- [31]Hirano, S., Shen, Y., Nishijima, S., Sakai, Y., & Umeda, H. 2023, MNRAS, 525, 5737, doi: 10.1093/mnras/stad2693DOI
- [34]Hopkins, P. F. 2016, MNRAS, 462, 576, doi: 10.1093/mnras/stw1578DOI
- [35]Hopkins, P. F. 2017, Monthly Notices of the Royal Astronomical Society, 466, 3387–3405, doi: 10.1093/mnras/stw3306DOI
- [36]Hopkins, P. F., & Grudić, M. Y. 2019, Monthly Notices of the Royal Astronomical Society, 483, 4187–4196, doi: 10.1093/mnras/stz3089
- [37]Hopkins, P. F., Grudić, M. Y., Wetzel, A., et al. 2020, Monthly Notices of the Royal Astronomical Society, 491, 3702–3729, doi: 10.1093/mnras/stz3129DOI
- [38]Hopkins, P. F., & Raives, M. J. 2016, Monthly Notices of the Royal Astronomical Society, 455, 51–88, doi: 10.1093/mnras/stv2180DOI
- [39]Hopkins, P. F., Wetzel, A., Kereš, D., et al. 2018, Monthly Notices of the Royal Astronomical Society, 480, 800–863, doi: 10.1093/mnras/sty1690DOI
- [40]Hopkins, P. F., Wetzel, A., Wheeler, C., et al. 2022, Monthly Notices of the Royal Astronomical Society, 519, 3154–3181, doi: 10.1093/mnras/stac3489DOI
- [41]Hopkins, P. F., Grudic, M. Y., Su, K.-Y., et al. 2024, The Open Journal of Astrophysics, 7, doi: 10.21105/astro.2309.13115DOI
- [42]Hosokawa, T., Hirano, S., Kuiper, R., et al. 2016, ApJ, 824, 119, doi: 10.3847/0004-637X/824/2/119DOI
- [43]Hosokawa, T., Omukai, K., Yoshida, N., & Yorke, H. W. 2011, Science, 334, 1250, doi: 10.1126/science.1207433DOI
- [44]Hummel, J. A., Stacy, A., & Bromm, V. 2016, Monthly Notices of the Royal Astronomical Society, 460, 2432, doi: 10.1093/mnras/stw1127DOI
- [45]Jaura, O., Glover, S. C. O., Wollenberg, K. M. J., et al. 2022, Monthly Notices of the Royal Astronomical Society, 512, 116, doi: 10.1093/mnras/stac487DOI
- [46]Khullar, S., Grudić, M. Y., Matzner, C. D., et al. 2026, arXiv e-prints, arXiv:2610.00513, doi: 10.48550/arXiv.2610.00513DOI
- [47]Kim, J.-h., Abel, T., Agertz, O., et al. 2013, The Astrophysical Journal Supplement Series, 210, 14, doi: 10.1088/0067-0049/210/1/14DOI
- [48]Kim, W.-T., & Ostriker, E. C. 2001, ApJ, 559, 70, doi: 10.1086/322330DOI
- [49]Klessen, R. S., & Glover, S. C. O. 2023, ARA&A, 61, 65, doi: 10.1146/annurev-astro-071221-053453DOI
- [50]Koh, D., Abel, T., & Jedamzik, K. 2021, ApJL, 909, L21, doi: 10.3847/2041-8213/abe8ddDOI
- [51]Komatsu, E., et al. 2011, ApJS, 192, 18, doi: 10.1088/0067-0049/192/2/18DOI
- [52]Kroupa, P. 2001, MNRAS, 322, 231, doi: 10.1046/j.1365-8711.2001.04022.xDOI
- [53]Krumholz, M. R., Dekel, A., & McKee, C. F. 2012, ApJ, 745, 69, doi: 10.1088/0004-637X/745/1/69DOI
- [54]Kulkarni, M., Visbal, E., & Bryan, G. L. 2021, The Astrophysical Journal, 917, 40, doi: 10.3847/1538-4357/ac08a3
- [55]Larson, R. B. 1969, MNRAS, 145, 271, doi: 10.1093/mnras/145.3.271DOI
- [56]Latif, M. A., Whalen, D., & Khochfar, S. 2022, ApJ, 925, 28, doi: 10.3847/1538-4357/ac3916DOI
- [57]Lee, E. J., Chang, P., & Murray, N. W. 2015, ApJ, 800, 49, doi: 10.1088/0004-637X/800/1/49DOI
- [60]Machida, M. N., Hirano, S., & Basu, S. 2025, ApJ, 988, 6, doi: 10.3847/1538-4357/addc56DOI
- [61]Matzner, C. D., & Jumper, P. H. 2015, The Astrophysical Journal, 815, 68, doi: 10.1088/0004-637X/815/1/68
- [62]McKee, C. F., & Ostriker, E. C. 2007, Annual Review of Astronomy and Astrophysics, 45, 565
- [63]McKee, C. F., & Tan, J. C. 2003, ApJ, 585, 850, doi: 10.1086/346149DOI
- [64]Meziani, Y. J., Hopkins, P. F., Grudić, M. Y., et al. 2026a, arXiv e-prints. https://arxiv.org/abs/2602.02953
- [65]Meziani, Y. J., Hopkins, P. F., Gandhi, P. J., et al. 2026b
- [66]Murray, D. W., Chang, P., Murray, N. W., & Pittman, J. 2017, MNRAS, 465, 1316, doi: 10.1093/mnras/stw2796DOI
- [67]Murray, N. W., & Chang, P. 2015, ApJ, 804, 44, doi: 10.1088/0004-637X/804/1/44DOI
- [68]Nakauchi, D., Inayoshi, K., & Omukai, K. 2014, Monthly Notices of the Royal Astronomical Society, 442, 2667, doi: 10.1093/mnras/stu1042DOI
- [69]Offner, S. S. R., Klein, R. I., McKee, C. F., & Krumholz, M. R. 2009, The Astrophysical Journal, 703, 131–149, doi: 10.1088/0004-637X/703/1/131
- [71]Omukai, K. 2000, The Astrophysical Journal, 534, 809, doi: 10.1086/308776
- [72]Omukai, K., & Palla, F. 2003, The Astrophysical Journal, 589, 677
- [73]Oswald, L. S., Weltvrede, P., Posselt, B., et al. 2025, MNRAS, doi: 10.1093/mnras/staf645DOI
- [74]Padoan, P., Pan, L., Juvela, M., Haugbølle, T., & Nordlund, Å. 2020, The Astrophysical Journal, 900, 82
- [75]Palla, F., Salpeter, E. E., & Stahler, S. W. 1983, ApJ, 271, 632, doi: 10.1086/161231DOI
- [76]Park, J., Ricotti, M., & Sugimura, K. 2022, Population III star formation in an X-ray background: III. Periodic radiative feedback and luminosity induced by elliptical orbits, https://arxiv.org/abs/2212.04564
- [77]Penston, M. V. 1969, MNRAS, 144, 425, doi: 10.1093/mnras/144.4.425DOI
- [78]Peters, T., Klessen, R. S., Mac Low, M.-M., & Banerjee, R. 2010, The Astrophysical Journal, 725, 134
- [79]Prieto, J., Padoan, P., Jiménez, R., & Infante, L. 2011, ApJL, 731, L38, doi: 10.1088/2041-8205/731/2/L38DOI
- [80]Prole, L. R., Clark, P. C., Klessen, R. S., & Glover, S. C. O. 2022a, Monthly Notices of the Royal Astronomical Society, 510, 4019, doi: 10.1093/mnras/stab3697
- [81]Prole, L. R., Clark, P. C., Klessen, R. S., Glover, S. C. O., & Pakmor, R. 2022b, MNRAS, 516, 2223, doi: 10.1093/mnras/stac2327DOI
- [82]Ripamonti, E., & Abel, T. 2004, Monthly Notices of the Royal Astronomical Society, 348, 1019, doi: 10.1111/j.1365-2966.2004.07422.x
- [83]Roberts, W. W., & Yuan, C. 1970, ApJ, 161, 887, doi: 10.1086/150592
- [84]Robertson, B., & Goldreich, P. 2012, ApJL, 750, L31, doi: 10.1088/2041-8205/750/2/L31DOI
- [85]Sadanari, K. E., Omukai, K., Sugimura, K., Matsumoto, T., & Tomida, K. 2023, MNRAS, 519, 3076, doi: 10.1093/mnras/stac3724DOI
- [86]Sadanari, K. E., Omukai, K., Sugimura, K., Matsumoto, T., & Tomida, K. 2024, PASJ, 76, 823, doi: 10.1093/pasj/psae051DOI
- [87]Shang, C., Bryan, G. L., & Haiman, Z. 2010, MNRAS, 402, 1249, doi: 10.1111/j.1365-2966.2009.15960.xDOI
- [88]Sharda, P., Federrath, C., Krumholz, M. R., & Schleicher, D. R. G. 2021, MNRAS, 503, 2014, doi: 10.1093/mnras/stab531DOI
- [89]Sharda, P., & Menon, S. H. 2025, MNRAS, 540, 1745, doi: 10.1093/mnras/staf803DOI
- [90]Sharda, P., Menon, S. H., Gerasimov, R., et al. 2025, MNRAS, 541, L1, doi: 10.1093/mnras/slaf043
- [91]Shu, F. H. 1977, The Astrophysical Journal, 214, 488–497, doi: 10.1086/155274
- [92]Sofue, Y., & Fujimoto, M. 1983, ApJ, 265, 722, doi: 10.1086/160718DOI
- [93]Sofue, Y., Fujimoto, M., & Wielebinski, R. 1986, Annual Review of Astronomy and Astrophysics, 24, 459
- [94]Springel, V. 2005, Monthly Notices of the Royal Astronomical Society, 364, 1105, doi: 10.1111/j.1365-2966.2005.09655.xDOI
- [95]Springel, V. 2010, Monthly Notices of the Royal Astronomical Society, 401, 791–851, doi: 10.1111/j.1365-2966.2009.15715.xDOI
- [96]Stacy, A., & Bromm, V. 2007, Monthly Notices of the Royal Astronomical Society, 382, 229, doi: 10.1111/j.1365-2966.2007.12247.xDOI
- [97]Stacy, A., & Bromm, V. 2013, Monthly Notices of the Royal Astronomical Society, 433, 1094–1107, doi: 10.1093/mnras/stt789DOI
- [98]Stacy, A., Bromm, V., & Lee, A. T. 2016, MNRAS, 462, 1307, doi: 10.1093/mnras/stw1728DOI
- [99]Stacy, A., Greif, T. H., & Bromm, V. 2010, MNRAS, 403, 45, doi: 10.1111/j.1365-2966.2009.16113.xDOI
- [100]Stacy, A., Greif, T. H., & Bromm, V. 2012, MNRAS, 422, 290, doi: 10.1111/j.1365-2966.2012.20605.xDOI
- [101]Stacy, A., McKee, C. F., Lee, A. T., Klein, R. I., & Li, P. S. 2022, MNRAS, 511, 5042, doi: 10.1093/mnras/stac372DOI
- [102]Stahler, S. W., Palla, F., & Salpeter, E. E. 1986, The Astrophysical Journal, 302, 590
- [103]Storck, A., Katz, H., Devrient, J., et al. 2026, MNRAS, 548, stag529, doi: 10.1093/mnras/stag529DOI
- [104]Sugimura, K., Matsumoto, T., Hosokawa, T., Hirano, S., & Omukai, K. 2020, ApJL, 892, L14, doi: 10.3847/2041-8213/ab7d37DOI
- [105]Sugimura, K., Matsumoto, T., Hosokawa, T., Hirano, S., & Omukai, K. 2023, ApJ, 959, 17, doi: 10.3847/1538-4357/ad02fcDOI
- [106]Sugimura, K., Omukai, K., & Inoue, A. K. 2014, MNRAS, 445, 544, doi: 10.1093/mnras/stu1778DOI
- [107]Susa, H., Hasegawa, K., & Tominaga, N. 2014, ApJ, 792, 32, doi: 10.1088/0004-637X/792/1/32DOI
- [108]Toomre, A. 1964, The Astrophysical Journal, 139, 1217–1238, doi: 10.1086/147861DOI
- [109]Truelove, J. K., Klein, R. I., McKee, C. F., et al. 1997, ApJL, 489, L179, doi: 10.1086/310975DOI
- [110]Turk, M. J., Clark, P., Glover, S. C. O., et al. 2011a, The Astrophysical Journal, 726, 55
- [111]Turk, M. J., Clark, P., Glover, S. C. O., et al. 2011b, ApJ, 726, 55, doi: 10.1088/0004-637X/726/1/55DOI
- [112]Ugalino, M. I., Diemer, B., Semenov, V. A., et al. 2026, arXiv e-prints. https://arxiv.org/abs/2610.02308
- [113]Wetzel, A., Hayward, C. C., Sanderson, R. E., et al. 2023, ApJS, 265, 44, doi: 10.3847/1538-4365/acb99aDOI
- [114]Wise, J. H., & Abel, T. 2007, The Astrophysical Journal, 665, 899
- [115]Wollenberg, K. M. J., Glover, S. C. O., Clark, P. C., & Klessen, R. S. 2020, MNRAS, 494, 1871, doi: 10.1093/mnras/staa289DOI
- [116]Yoshida, N., Abel, T., Hernquist, L., & Sugiyama, N. 2003, The Astrophysical Journal, 592, 645
- [117]Yoshida, N., Omukai, K., & Hernquist, L. 2008, Science, 321, 669, doi: 10.1126/science.1160259DOI













