Separate Universe Framework Overview

• Separate Universe Framework is a method that re-expresses an inhomogeneous universe as locally homogeneous patches with shifted background parameters.
• It enables precise calibration of halo bias and response functions by linking long-wavelength overdensities with small-scale halo statistics.
• The framework underpins techniques in inflation, modified gravity, and quantum cosmology, offering insights into isotropic, anisotropic, and stochastic dynamics.

Searching arXiv for the cited paper and closely related separate-universe work to ground the article in current arXiv records. arXiv search: (Li et al., 2015) separate universe halo bias The separate universe framework is a long-wavelength approximation in which an inhomogeneous universe is re-expressed as an ensemble of locally homogeneous cosmologies, each patch evolving as if it were its own Friedmann–Lemaître–Robertson–Walker universe with suitably shifted background parameters. In large-scale structure, a uniform overdensity can be absorbed into a modified background density, expansion history, and curvature, enabling direct calibration of response functions such as halo bias and power-spectrum responses. In inflationary cosmology, the same logic underlies the gradient expansion, the $\delta N$ formalism, stochastic inflation, and soft-limit relations for correlation functions. Across these settings, the central question is not whether a long mode exists, but under what conditions its effect is exhausted by a background redefinition rather than by residual gradient, anisotropic, or non-adiabatic dynamics (Li et al., 2015, Wagner et al., 2014, Holland, 2024).

1. Foundational statement and geometric content

In its standard form, the framework begins from a long-wavelength perturbation whose scale is much larger than the small-scale observables of interest. For a uniform matter overdensity $\delta_L$, one defines a local matter density

$\bar\rho_{mW} = \bar\rho_m (1+\delta_L),$

and adjusts the local expansion history so that small-scale structure evolves as it would in a true FRW model with shifted parameters. In flat $\Lambda$CDM, working to first order in $\delta_b$, the shifts can be written as

$$\frac{\delta h}{h} = -\frac{5\Omega_m}{6}\frac{\delta_b}{D}, \qquad \frac{\delta\Omega_m}{\Omega_m} = \frac{\delta\Omega_\Lambda}{\Omega_\Lambda} = -\delta\Omega_K = -2\frac{\delta h}{h},$$

with the local scale factor on equal-proper-time slices satisfying

$a_W \simeq a\left(1-\frac{\delta_b}{3}\right).$

This is the background-rescaling content of separate-universe simulations in large-scale structure (Li et al., 2015).

The same idea admits a fully geometric formulation. In a scalar-perturbed FRW spacetime, one can define a local Hubble parameter $H_W$ from the volume expansion of the normal congruence and a local curvature $K_W$ from the perturbed spatial Ricci scalar. The separate-universe ansatz is valid when the local curvature is conserved along freely falling worldlines, which in comoving gauge reduces to the requirement that the lapse perturbation be much smaller than the curvature potential,

$|\xi| \ll |{\cal R}|.$

Under this condition, the local patch is indistinguishable, in its angle average, from a curved FRW cosmology with scale factor $\delta_L$0 and curvature $\delta_L$1 (Hu et al., 2016).

Two complementary GR results sharpen this statement. First, for a spherical compensated tophat of arbitrary amplitude and radius in $\delta_L$2CDM, the separate-universe conjecture can be proved exactly: the overdense region evolves as a distinct curved FLRW universe with its own scale factor and spatial curvature (Dai et al., 2015). Second, for arbitrary long-wavelength scalar perturbations, Conformal Fermi Coordinates isolate the isotropic effect as a local curvature shift while the anisotropic part is captured exactly by a Newtonian tidal field. This makes explicit that the isotropic and trace-free sectors play different roles in the framework (Dai et al., 2015).

The scope of validity is not unlimited. In GR with multiple fluids, the framework is restricted to scales larger than the sound horizons of all fluid components. When pressure gradients or anisotropic stress are relevant, the long mode is no longer exhausted by a pure background rescaling, and additional dynamical structure must be retained (Dai et al., 2015).

2. Response formalism and halo-bias calibration

In halo statistics, the framework turns the peak-background split into a direct numerical calibration. If $\delta_L$3 is the comoving number density of halos of mass $\delta_L$4 in a universe with shifted mean density $\delta_L$5, the Lagrangian response bias is defined by

$\delta_L$6

evaluated at fixed comoving volume, with Eulerian bias

$\delta_L$7

The corresponding large-scale clustering definition is

$\delta_L$8

The separate-universe consistency relation states that, in $\delta_L$9CDM with Gaussian initial conditions,

$\bar\rho_{mW} = \bar\rho_m (1+\delta_L),$0

The physical content is that any effect by which a long mode changes halo abundance must appear both in the one-point response and in the two-point halo–matter cross-correlation in the $\bar\rho_{mW} = \bar\rho_m (1+\delta_L),$1 limit (Li et al., 2015).

The numerical implementation in the halo-bias calibration study uses 32 realizations of $\bar\rho_{mW} = \bar\rho_m (1+\delta_L),$2 with $\bar\rho_{mW} = \bar\rho_m (1+\delta_L),$3 particles for $\bar\rho_{mW} = \bar\rho_m (1+\delta_L),$4, with matched initial random phases in each $\bar\rho_{mW} = \bar\rho_m (1+\delta_L),$5 pair, plus 25 realizations of $\bar\rho_{mW} = \bar\rho_m (1+\delta_L),$6 with $\bar\rho_{mW} = \bar\rho_m (1+\delta_L),$7 to improve statistics for the rarest halos. Halos are identified with a spherical-overdensity finder at $\bar\rho_{mW} = \bar\rho_m (1+\delta_L),$8 relative to the global mean; halos with fewer than 100 particles are discarded, and results are reported for halos with at least 400 particles. Mass functions and threshold shifts are extracted without binning by penalized-spline smoothing, while clustering bias is measured by assigning halos and matter to a $\bar\rho_{mW} = \bar\rho_m (1+\delta_L),$9 grid, computing $\Lambda$0 and $\Lambda$1 with FFTs, and fitting over $\Lambda$2 with $\Lambda$3 (Li et al., 2015).

A key practical device is abundance matching between separate-universe pairs with $\Lambda$4 and $\Lambda$5. One adjusts the threshold $\Lambda$6 so that

$\Lambda$7

defines the threshold shift $\Lambda$8, and infers cumulative and then differential response biases. This avoids direct numerical differentiation of noisy mass functions and makes the calibration efficient even for rare, highly biased halos (Li et al., 2015).

The principal result is a quantitative verification of the consistency relation. The response bias and the clustering bias agree at the $\Lambda$9–$\delta_b$0 level for average halo biases $\delta_b$1, with no statistically significant deviations at the $\delta_b$2–$\delta_b$3 level out to $\delta_b$4. By contrast, halo bias inferred from the universal mass-function approximation is inaccurate at the $\delta_b$5 level or more. The framework therefore provides an efficient calibration of linear halo bias in $\delta_b$6CDM even for highly biased rare halos, and any observational violation of the consistency relation would indicate new physics, for example in the dark matter, dark energy, or primordial non-Gaussianity sectors (Li et al., 2015).

3. Real and fake separate universes, modified gravity, and curvature mappings

When additional species possess non-gravitational forces, the long mode cannot always be absorbed into a standard FRW background with constant curvature. In the extension developed for dynamical dark energy and massive neutrinos, one still matches the local Hubble rate $\delta_b$7 through the acceleration equation, but the character of the local curvature depends on the Jeans or free-streaming scale. Above the Jeans scale, $\delta_b$8, the construction is a “real” separate universe: the local Friedmann equation contains real energy densities and a constant $\delta_b$9. Below the Jeans scale, $$\frac{\delta h}{h} = -\frac{5\Omega_m}{6}\frac{\delta_b}{D}, \qquad \frac{\delta\Omega_m}{\Omega_m} = \frac{\delta\Omega_\Lambda}{\Omega_\Lambda} = -\delta\Omega_K = -2\frac{\delta h}{h},$$0, pressure or anisotropic-stress gradients generate relative flows and the curvature evolves; this can be represented by a “fake” density component $$\frac{\delta h}{h} = -\frac{5\Omega_m}{6}\frac{\delta_b}{D}, \qquad \frac{\delta\Omega_m}{\Omega_m} = \frac{\delta\Omega_\Lambda}{\Omega_\Lambda} = -\delta\Omega_K = -2\frac{\delta h}{h},$$1 or, equivalently, by an effective curvature equation of state $$\frac{\delta h}{h} = -\frac{5\Omega_m}{6}\frac{\delta_b}{D}, \qquad \frac{\delta\Omega_m}{\Omega_m} = \frac{\delta\Omega_\Lambda}{\Omega_\Lambda} = -\delta\Omega_K = -2\frac{\delta h}{h},$$2 (Hu et al., 2016).

This “real-and-fake” distinction is not merely terminological. It encodes the fact that the response of small-scale observables becomes scale-dependent and temporally nonlocal when the long mode crosses a species-dependent propagation scale. In this setting, responses such as the nonlinear matter power spectrum, halo abundance, halo bias, and squeezed $$\frac{\delta h}{h} = -\frac{5\Omega_m}{6}\frac{\delta_b}{D}, \qquad \frac{\delta\Omega_m}{\Omega_m} = \frac{\delta\Omega_\Lambda}{\Omega_\Lambda} = -\delta\Omega_K = -2\frac{\delta h}{h},$$3-point functions acquire a nontrivial dependence on the full time history of $$\frac{\delta h}{h} = -\frac{5\Omega_m}{6}\frac{\delta_b}{D}, \qquad \frac{\delta\Omega_m}{\Omega_m} = \frac{\delta\Omega_\Lambda}{\Omega_\Lambda} = -\delta\Omega_K = -2\frac{\delta h}{h},$$4 rather than only on an instantaneous overdensity (Hu et al., 2016).

A more theory-agnostic criterion is provided by the metric-based formulation beyond GR. There, the separate-universe ansatz holds whenever the effective stress energy, defined geometrically from the Einstein tensor, comoves with freely falling synchronous observers so that the local curvature is conserved. Operationally, one requires the comoving-gauge lapse perturbation to be negligible relative to the curvature perturbation. Under that condition, one can run small-scale $$\frac{\delta h}{h} = -\frac{5\Omega_m}{6}\frac{\delta_b}{D}, \qquad \frac{\delta\Omega_m}{\Omega_m} = \frac{\delta\Omega_\Lambda}{\Omega_\Lambda} = -\delta\Omega_K = -2\frac{\delta h}{h},$$5-body or hydrodynamic simulations in a background with $$\frac{\delta h}{h} = -\frac{5\Omega_m}{6}\frac{\delta_b}{D}, \qquad \frac{\delta\Omega_m}{\Omega_m} = \frac{\delta\Omega_\Lambda}{\Omega_\Lambda} = -\delta\Omega_K = -2\frac{\delta h}{h},$$6 replacing $$\frac{\delta h}{h} = -\frac{5\Omega_m}{6}\frac{\delta_b}{D}, \qquad \frac{\delta\Omega_m}{\Omega_m} = \frac{\delta\Omega_\Lambda}{\Omega_\Lambda} = -\delta\Omega_K = -2\frac{\delta h}{h},$$7 even in modified gravity theories; failure of small-scale observables to follow this prediction would signal genuine new physics such as fifth forces or environment-dependent screening (Hu et al., 2016).

The same logic can be turned around and used as an approximation scheme for global spatial curvature. In the non-flat $$\frac{\delta h}{h} = -\frac{5\Omega_m}{6}\frac{\delta_b}{D}, \qquad \frac{\delta\Omega_m}{\Omega_m} = \frac{\delta\Omega_\Lambda}{\Omega_\Lambda} = -\delta\Omega_K = -2\frac{\delta h}{h},$$8CDM application, a curved model with $$\frac{\delta h}{h} = -\frac{5\Omega_m}{6}\frac{\delta_b}{D}, \qquad \frac{\delta\Omega_m}{\Omega_m} = \frac{\delta\Omega_\Lambda}{\Omega_\Lambda} = -\delta\Omega_K = -2\frac{\delta h}{h},$$9 is mapped to a flat separate-universe model threaded by a long-wavelength density perturbation $a_W \simeq a\left(1-\frac{\delta_b}{3}\right).$0. The mapping is fixed by

$a_W \simeq a\left(1-\frac{\delta_b}{3}\right).$1

together with

$a_W \simeq a\left(1-\frac{\delta_b}{3}\right).$2

and the associated shifts in $a_W \simeq a\left(1-\frac{\delta_b}{3}\right).$3 and $a_W \simeq a\left(1-\frac{\delta_b}{3}\right).$4. The method predicts the nonlinear matter power spectrum for $a_W \simeq a\left(1-\frac{\delta_b}{3}\right).$5 up to $a_W \simeq a\left(1-\frac{\delta_b}{3}\right).$6 over $a_W \simeq a\left(1-\frac{\delta_b}{3}\right).$7 to fractional accuracy within $a_W \simeq a\left(1-\frac{\delta_b}{3}\right).$8, with the normalized power-spectrum response to $a_W \simeq a\left(1-\frac{\delta_b}{3}\right).$9 well approximated by the response to the Hubble parameter $H_W$0 within the flat model (Terasawa et al., 2022).

4. Anisotropic separate universes and environment-dependent bias

The isotropic framework absorbs a long-wavelength overdensity into a scalar background shift. Its anisotropic generalization absorbs a uniform, trace-free tidal field into direction-dependent expansion factors. In the principal-axis frame of the tidal tensor, one promotes the global scale factor $H_W$1 to three directional factors $H_W$2 or $H_W$3, and evolves particles in anisotropic comoving coordinates. This yields a controlled implementation of a large-scale tidal field in periodic $H_W$4-body simulations (Stücker et al., 2020, Masaki et al., 2020).

Two response functions then become central. The growth-only tidal response $H_W$5 measures the change of the power spectrum at fixed anisotropic-comoving wavenumber, while the total response

$H_W$6

includes the dilation from coordinate rescaling. Perturbation theory gives $H_W$7 on large scales. In the nonlinear regime, anisotropic separate-universe simulations show that the response does not vanish on small physical scales. One study finds that at $H_W$8 the total tidal response approaches a constant value $H_W$9 for $K_W$0 up to $K_W$1, implying that even the inner regions of haloes are affected by the large-scale tidal field; another measures characteristic scale dependence down to nonlinear scales up to $K_W$2 and validates the implementation against perturbation theory and high-resolution PM simulations (Stücker et al., 2020, Masaki et al., 2020).

This anisotropic response supplies the tidal part of the squeezed bispectrum. Together with the overdensity response measured in isotropic separate-universe simulations, it completely specifies the nonlinear matter bispectrum in the squeezed limit in the response approach (Stücker et al., 2020). The same simulations also directly quantify halo-shape alignments with a large-scale tidal field, finding a clear signal that increases with halo mass (Stücker et al., 2020).

The framework also supports separate-universe calibrations of assembly bias. In one line of work, the response of halo abundances is measured at fixed mass and concentration; the concentration distribution at fixed mass is modeled as nearly lognormal, which yields analytic expressions for concentration-dependent $K_W$3 and $K_W$4 in terms of Hermite polynomials and response coefficients extracted from the simulations. This produces the first calibration of assembly bias in $K_W$5 and reveals a non-universality in the $K_W$6–$K_W$7 relation when halos are split by concentration (Paranjape et al., 2016). In another extension, the local cosmic-web tidal anisotropy $K_W$8 is treated as the conditioning variable. Its distribution at fixed mass is nearly lognormal, and the resulting analytic model reproduces the measured $K_W$9 and $|\xi| \ll |{\cal R}|.$0 to sub-percent precision over $|\xi| \ll |{\cal R}|.$1 and $|\xi| \ll |{\cal R}|.$2, with the calibration of $|\xi| \ll |{\cal R}|.$3 identified as the first demonstration of the dependence of non-linear bias on the local web environment (Ramakrishnan et al., 2020).

5. Inflationary long-wavelength dynamics, soft limits, and stochastic applications

In inflation, the separate-universe framework is the leading term of the gradient expansion. On scales much larger than the Hubble radius, one neglects spatial gradients of the lapse, shift, and scalar fields so that each local Hubble patch evolves as an independent FLRW universe. For multifield models written in ADM variables,

$|\xi| \ll |{\cal R}|.$4

the long-wavelength ordering is $|\xi| \ll |{\cal R}|.$5, $|\xi| \ll |{\cal R}|.$6, and $|\xi| \ll |{\cal R}|.$7, with $|\xi| \ll |{\cal R}|.$8 (Holland, 2024). A more general formulation emphasizes that locality and spatial-diffeomorphism invariance are the structural conditions behind the framework, allowing a generalized $|\xi| \ll |{\cal R}|.$9 formalism even in the presence of large-scale shear and in theories that violate full spacetime diffeomorphism invariance but preserve time reparametrization and spatial diffeomorphisms (Tanaka et al., 2021).

The Hamiltonian formulation makes explicit that separate universes are a phase-space reduction. In the isotropic truncation, one retains only homogeneous variables in each patch and drops anisotropic modes and gradients. For a single-field model this gives canonical variables such as $\delta_L$00 with a reduced scalar constraint; for multifield non-linear sigma models, the isotropic variables are $\delta_L$01, $\delta_L$02, $\delta_L$03, and $\delta_L$04. Comparison with full cosmological perturbation theory shows that the reduced dynamics agrees with the full large-scale dynamics when gradients are sufficiently suppressed and the gauge choice is compatible with the truncation. Uniform-expansion gauge is identified as especially well described by the separate-universe picture in the multifield phase-space analysis (Artigas et al., 2021, Grain et al., 16 Apr 2025).

In multifield inflation, validity requires more than $\delta_L$05. The effective mass matrix

$\delta_L$06

must dominate the gradient terms along both adiabatic and entropic directions. In the adiabatic–entropic basis this implies conditions of the form

$\delta_L$07

or, in the phase-space presentation, wavelengths must exceed both the Hubble radius and the inverse effective masses of the fields (Holland, 2024, Grain et al., 16 Apr 2025). These conditions justify the use of stochastic inflation in wide classes of multifield models and underpin applications to amplified power spectra and primordial black-hole production (Holland, 2024).

Soft-limit correlation functions are a particularly natural output of the formalism. Long modes act as background shifts for short modes, so squeezed and collapsed limits of inflationary correlators can be written in terms of derivatives of hard correlators with respect to background fields. This yields a diagrammatic separate-universe formalism for single- and double-soft limits of the bispectrum and trispectrum in multifield inflation, together with an infinite tower of inequalities generalizing the Suyama–Yamaguchi inequality (Kenton et al., 2016).

The inflationary framework also has important failure modes. For sudden transitions from slow roll to ultra-slow roll, the separate-universe approximation can fail on a finite range of super-Hubble scales. The approximation remains valid piecewise before and after the transition, but residual spatial gradients require a discontinuity in the homogeneous solution at the transition. This has direct implications for the $\delta_L$08 formalism and stochastic inflation, which otherwise miss the transition-induced kick (Jackson et al., 2023). Relatedly, loop corrections in models with a narrow short-scale spike can be organized efficiently in the separate-universe picture: large back-reaction requires both short-scale nonlinearities and long-short couplings that modulate the short-scale power spectrum, whereas in the absence of long-short couplings the effect reduces to incoherent shot noise that is volume-suppressed (Iacconi et al., 2023).

6. Quantum-gravity realizations, bouncing cosmologies, and primordial-black-hole limits

The framework has also been transplanted into quantum cosmology. In loop quantum cosmology, one splits the spatial slice into patches larger than the sound horizon and quantizes each patch with the effective LQC Hamiltonian constraint. Long-wavelength scalar perturbations then obey the same form of evolution equation as in GR but with a modified pump field,

$\delta_L$09

and tensor modes have an analogous dressed quantity $\delta_L$10. For a constant equation of state, the long-wavelength curvature perturbation and tensor modes can be solved analytically across the bounce, and the tensor-to-scalar ratio may be suppressed or amplified by quantum-gravity effects depending on $\delta_L$11. In particular, if the equation of state lies between $\delta_L$12 and $\delta_L$13, the tensor-to-scalar ratio is suppressed during the bounce (Wilson-Ewing, 2015).

In group field theory condensate cosmology, the framework is implemented by a tensor product of condensate states, one per spatial patch, with the patch labels supplied by scalar rods. Under the separate-universe approximation, the patch wavefunctions decouple and each obeys the homogeneous condensate equation of motion. The resulting long-wavelength scalar-perturbation equations agree with the classical separate-universe equations of GR in the classical limit while acquiring quantum-gravity corrections that become important near Planckian curvature (Gerhardt et al., 2018).

A different, older use of “separate universe” concerns the maximum size of positively curved overdense regions and the associated upper bound on primordial-black-hole masses. For a perfect fluid with equation of state $\delta_L$14, a positively curved overdense region can only extend to a maximum proper scale before it closes up as a separate Friedmann 3-sphere. For $\delta_L$15, this separate-universe scale is always of order the cosmological particle horizon, confirming that a primordial black hole cannot be much larger than the particle horizon at formation. For $\delta_L$16, the interpretation changes: a sufficiently large positive-curvature region produces a baby universe rather than a black hole (Carr et al., 2014).

Taken together, these developments show that the separate universe framework is not a single model but a transferable reduction principle. In large-scale structure it is a precision calibration tool for responses, bias parameters, and squeezed limits; in inflation it is the backbone of $\delta_L$17, stochastic methods, and soft-limit expansions; in modified gravity and multicomponent cosmologies it becomes a diagnostic for when background rescaling ceases to be complete; and in quantum cosmology it provides a bridge between homogeneous effective dynamics and long-wavelength perturbations (Li et al., 2015, Holland, 2024, Wilson-Ewing, 2015).