betaFeedback

WorksPhysics and Astronomy

Cubic Galileon Gravity on Nonlinear Scales with Hi-COLA for LSST Year 1

Abstract

Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST) is in a prime position to test Modified Gravity (MG) with Weak Lensing (WL) and Large Scale Structure (LSS). However, many nonlinear MG emulators in the literature enforce a ΛΛCDM background by design, missing where data suggest the most interesting deviations lie. We present an emulator for the 3D nonlinear matter power spectrum in Cubic Galileon gravity, a Horndeski theory with a single additional MG parameter, fϕf_ϕ, that modifies the background, growth, and lensing. Trained on Hi-COLA simulations with Gaussian Process Regression (GPR), the emulator achieves ≲2%\lesssim 2\% accuracy for k≤1.2 h Mpc−1k \leq 1.2\ h\,\text{Mpc}^{-1} at the fiducial cosmology relative to Hi-COLA, and ≲5%\lesssim 5\% relative to full N-body EFT-RAMSES simulations. Additionally, conservative scale cuts excluding angular bins affected by baryonic physics enable unbiased parameter estimation in LSST Year 1 weak lensing and galaxy clustering (3×23\times2pt) forecasts, even when validated against simulated data based on higher-resolution N-body simulations. This demonstrates that, within a simplified LSST Y1 setup and the adopted scale cuts, Hi-COLA-based emulation is sufficient to develop inference infrastructure with Stage IV precision for Horndeski gravity models previously unmodeled on nonlinear scales.

Introduction

In the standard cosmological model, Λ\LambdaCDM, the dark energy responsible for the Universe’s accelerated expansion is described by a cosmological constant – an assumption that can explain a suite of cosmological observations, including Type Ia supernovae (SN Ia) distances [1], anisotropies of the cosmic microwave background [2], and data from a range of Weak Lensing and Large Scale Structure surveys [3]. Nonetheless, the cosmological constant comes with significant theoretical challenges, and recent Baryon Acoustic Oscillation (BAO) results from the Dark Energy Spectroscopic Instrument (DESI) [4] and multi-probe constraints from the full six-year Dark Energy Survey (DES Y6) [5] are in tension with the standard cosmological model’s picture of dark energy. We need to test extensions to Λ\LambdaCDM to assess whether these tensions can be explained by new physics, or whether we must instead keep looking for misspecified modelling in the systematics sector of the model.

Driven by theoretical and observational limitations of GR, a well-motivated class of extensions can be found in modified gravity theories [6, 7, 8]. It is possible that the observed acceleration is in fact a signature of a departure from General Relativity (GR) rather than the presence of an additional dark energy fluid. In some modified gravity models, an effective dark energy equation of state w(a)w(a) that evolves with scale factor aa can arise naturally, and can even cross the so-called phantom divide, the boundary at w=−1w=-1 between the quintessence (w>−1w>-1) and phantom (w<−1w<-1) regimes [9, 10, 11, 12, 13, 14, 15, 16, 17, 18]. In this sense, modified gravity provides a physically richer and more predictive explanation for time-evolving cosmic acceleration than a DE fluid. For example, many models belonging to the Horndeski class alter the cosmological background as well as the growth of structure and lensing of light [19], including with the Hi-COLA code used in this work [20]. This scenario exposes a gap in the existing MG modelling landscape: Hu-Sawicki f(R)f(R) [21] and nDGP gravity [22], the two theories of gravity for which nonlinear simulation and emulation infrastructure is most developed (see e.g., link textFORGE-Emu, link textSesame, link textFofrFittingFunction, link texte-MANTIS and BRIDGE [23]), are conventionally implemented with a cosmological background fixed to that of Λ\LambdaCDM, precisely where much of the currently interesting deviation from the standard cosmological model lies. There is therefore a clear need to develop nonlinear modelling tools for theories that modify the background expansion as well as the growth of structure.

Weak lensing (WL) and Large Scale Structure (LSS) are ideally suited to test deviations to our standard cosmological model, especially beyond modifications to the background evolution. They are exceptionally rich probes because they encode information across a range of redshifts, acting complementarily to probes of the cosmic microwave background (CMB) anisotropies. The galaxy density field and the weak lensing shear field jointly constrain the lensing of light, the clustering of matter, and the background evolution as a function of time. For example, it is standard to use combined constraints from the two-point correlation functions generated by these fields: galaxy clustering, galaxy-galaxy lensing and cosmic shear (jointly called the 3×2pt3 \times2\mathrm{pt} datavector) [24, 25, 26], though beyond-2pt statistics of these fields exist and have been used for cosmological inference [27, 28, 29, 30, 31, 32, 33]. In the near future, the first cosmological data from the Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST) [34] and Euclid [35] will deliver competitive constraints on gravity [36, 37, 38, 39, 40].

In order to model accurately the two-point statistics we can extract from these surveys, we need to compute the 3D matter power spectrum on nonlinear scales. This can be done far into the nonlinear regime by using N-body simulations. For the purpose of repeated model evaluation during parameter inference, these unfortunately remain prohibitively computationally expensive; therefore emulators (fast, approximate surrogate modelling tools, usually based on machine learning methods) have been proposed, developed and validated in the literature as a means of quickly replicating the results of these simulations in a fraction of a second [41]. Emulators are an important example of the novel modelling tools required to incorporate small-scale information into the exploration both of the standard cosmological model (see EuclidEmulator2 and BACC0Emu) and of beyond-Λ\LambdaCDM models.

While summary statistics in Hu-Sawicki f(R)f(R) and nDGP gravity have been extensively emulated, very few other MG theories have N-body simulation infrastructure readily available, which hinders the production of training data sets on which emulators are built. New tools (which include approximate N-body simulations and theoretical prescriptions) like Hi-COLA1 [42] are making it less computationally expensive to estimate non-linear MG power spectra for a broader parameter space, opening up the possibility of systematically building emulators for a wide range of alternative theories of gravity. A range of beyond-Λ\LambdaCDM emulators have been built, often based on approximate nonlinear modelling schemes such as the halo model reaction [43] (see e.g., ReACT-based emulators2 and [44]), or COLA (see e.g., MGemu and [45]). For the latter, [46] provides a framework for building these emulators and additionally introduces nDGP-emu. In this paper, we take a significant step in this direction, introducing and validating the first nonlinear matter power spectrum emulator for Cubic Galileon (CG) gravity, accurate enough to meet LSST Year 1 (‘Y1’ hereafter) requirements. Here, we present a worked example of how to build a nonlinear emulator and carry out an LSST Y1 forecast for a theory of gravity that does not enforce a Λ\LambdaCDM background by construction. Cubic Galileon gravity [47, 48, 49] presents a compelling test case. It has some clear advantages over f(R)f(R) and nDGP gravity: unlike these theories, the specific version of Cubic Galileon gravity we targeted when building this emulator does not incorporate extraneous dark energy components. Although the model includes a cosmological constant, its background evolution can deviate from Λ\LambdaCDM, avoiding the fine-tuning of an additional dark energy component at the level of the field equations and resulting in a more theory-driven than a phenomenology-driven framework. Still, we must be clear that the Cubic Galileon model we consider fares relatively poorly against current observations. It cannot produce a dark energy equation of state that crosses the phantom divide, it is observationally ruled out for fϕ=1f_\phi=1 [50] (where fϕf_\phi is the free parameter of the gravity theory), and it is disfavoured by joint constraints from Planck, DESI BAO, DES SNe, galaxy-ISW cross-correlations and void force profile measurements [20]. Nevertheless, this model is a good example of a minimal extension to Λ\LambdaCDM that simultaneously modifies the background evolution, the clustering of matter, and the lensing of light. This establishes a template that can be extended to other Horndeski theories with similar freedom in their background evolution.

This paper is organized as follows. In § 2 we introduce Cubic Galileon gravity and the specific model we consider in this paper (with tracker solution on the background), and then describe the Hi-COLA code used to generate simulations that will form the training set for our emulator. In § 3 we outline the generation of our training and validation data and the evaluation of the matter power spectrum boosts. In § 4 we describe the construction of the emulator, outlining both the single-snapshot emulation and the interpolation in redshift. In § 5 we introduce the assumptions and modelling choices for the forecasting infrastructure used in this paper when analysing simulated LSST Y1 3×23 \times2pt data. In § 6 we present our results: we validate the emulator against Hi-COLA simulations and against a higher-resolution EFT-RAMES N-body simulation [51] (EFT-RAMES is a modified gravity simulation code derived from ECOSMOG [52, 53]), and assess its performance as a forecasting tool through a simulated LSST Y1 3×23 \times2pt parameter inference analysis. Finally, we conclude in § 7.

Theory and Simulations

Cubic Galileon and the Tracker Solution

Cubic Galileon gravity (without a cosmological constant) has been widely explored [47, 48, 49, 54, 55] and strongly constrained observationally [56, 57], with full N-body simulations available in the literature [58, 59, 60]. Here, we consider the Cubic Galileon Lagrangian with the following form:

L=MP22R+MK~2k1X−MG~3Msg31X□ϕ−MP2Λ,(1)\mathcal{L} = \frac{M_{\mathrm{P}}^{2}}{2}R + M_{\widetilde{K}}^{2}k_{1}X - M_{\widetilde{G}_{3}}M_{\mathrm{s}}g_{31}X\Box\phi- M_{\mathrm{P}}^{2}\Lambda, \tag*{(1)}

where RR is the Ricci scalar; MPM_{\mathrm{P}} is the Planck mass; MsM_{\mathrm{s}}, MK~M_{\widetilde{K}} and MG~3M_{\widetilde{G}_{3}} are the mass scales associated with the scalar field and with the Horndeski functions K~\widetilde{K} and G~3\widetilde{G}_{3}; and X=−∂μϕ∂μϕ2X=-\frac{\partial_{\mu}\phi\partial^{\mu}\phi}{2} is the dimensionless kinetic term of the dimensionless scalar field ϕ\phi. k1k_{1} and g31g_{31} are dimensionless constants, and we have also included a cosmological constant term Λ\Lambda. If assume all the mass scales are the same (see Appendix A.§ 2 in [42]) we can find the following equations of motion at the background level:

Ωϕ=k1ϕ′26+g31E2ϕ′3E′E=k1ϕ′24−g31E4ϕ′44(E′Eϕ′+ϕ′′)−Ωr−3ΩΛ+32ϕ′′=−3g31E′ϕ′2+3k1ϕ′+9g31E2ϕ′2k1+6g31E2ϕ′−E′Eϕ′(2)\begin{aligned} \Omega_{\phi} &= \frac{k_1\phi'^2}{6} + g_{31}E^2\phi'^3 \\ \frac{E'}{E} &= \frac{k_1\phi'^2}{4} - \frac{g_{31}E^4\phi'^4}{4}\left(\frac{E'}{E}\phi' + \phi''\right) - \frac{\Omega_r - 3\Omega_{\Lambda} + 3}{2} \\ \phi'' &= -\frac{3g_{31}E'\phi'^2 + 3k_1\phi' + 9g_{31}E^2\phi'^2}{k_1 + 6g_{31}E^2\phi'} - \frac{E'}{E}\phi' \tag*{(2)} \end{aligned}

where we define the dimensionless Hubble rate E=H/H0E = H/H_0, Ωϕ\Omega_{\phi} is the energy contribution of the scalar field (with ΩDE≡ΩΛ+Ωϕ\Omega_{\mathrm{DE}} \equiv\Omega_{\Lambda} + \Omega_{\phi}), all other quantities have their standard meanings and the prime denotes the derivative with respect to the dimensionless time variable xx, defined as the logarithm of the scale factor x=ln⁡(a)x = \ln(a).

The rescaling symmetry of the Cubic Galileon [61, 62, 63] gives one the freedom to fix either the initial conditions for EE and ϕ′\phi', or the values of the model constants. For example, one can avoid having to deduce a suitable initial condition3 for the scalar field derivative ϕ′\phi' by rescaling it by its present day value. If we denote quantities that are rescaled under this ‘today’ convention with a TT subscript, the initial condition for the Hubble rate and scalar field derivative becomes E0=1E_0 = 1 and ϕT0′=1\phi'_{T0} = 1 respectively, and accordingly one can redefine the model constants k1T=k1(ϕ0′)2k_{1T} = k_1(\phi'_0)^2 and g31T=g31(ϕ0′)3g_{31T} = g_{31}(\phi'_0)^3 such that the background equations of motion remain invariant. This is the convention we have chosen to adopt for this work (see [42] for a more in-depth exploration of the ‘today’ formalism).

Cubic Galileon gravity has a set of stable solutions with de Sitter (dS) attractors, that asymptotically approach a universe dominated by a constant vacuum energy at late times. Regardless of their initial conditions, these solutions converge onto a common trajectory in their background evolution, known as the tracker solution [64]. We set up our emulator to produce power spectra for these solutions only. In the ‘today’ formalism, the tracker solution exists on the hypersurface

k1T=6fϕ(Ωm0+Ωr0−1)(3)k_{1T} = 6f_{\phi}(\Omega_{m0} + \Omega_{r0} - 1) \tag*{(3)}

of our parameter space, where fϕf_{\phi} is defined as the fraction of the dark energy component due to the scalar field to the total dark energy today (fϕ=Ωϕ0Ωϕ0+ΩΛ0f_{\phi} = \frac{\Omega_{\phi0}}{\Omega_{\phi0}+\Omega_{\Lambda0}}). The initial conditions on EE and ϕT′\phi'_T also mean that the two constants k1Tk_{1T} and g31Tg_{31T} are dependent, so knowing the value of one is sufficient to find the other (through equations 2.2). Because we consider only tracker solutions for our emulator, we therefore recover a formalism of the theory described by a single modified gravity parameter fϕf_{\phi}. By varying fϕf_{\phi}, we get a different background history in the recent universe compared to that we expect from Λ\LambdaCDM, as well as a different coupling to matter in the Poisson equation and linear growth equation, commonly parameterised by the factor μ\mu [65, 66, 67, 68] (defined in § 5 below). Choosing to restrict to the tracker solution limits generality, but we make it for practical reasons. Off the tracker, k1Tk_{1T} and g31Tg_{31T} are no longer related by Eq. (3). The background then depends on a second free parameter, with a correspondingly larger training set, along with requiring stability checks across the extended space. The tracker family is the minimal one-parameter extension of Λ\LambdaCDM within this model, and it is therefore the natural first target for the emulation and inference infrastructure developed here. Refs. [69, 70, 71] confronted the general (non-tracker) solutions of the Galileon model with cosmological data.

By adding a cosmological constant and imposing the tracker solution on the background, we can build an emulator for a viable theory of gravity with a single additional parameter compared to Λ\LambdaCDM, whose power spectra can be emulated continuously up to the Λ\LambdaCDM limit. Throughout this paper, we make use of two cosmological models, defined as follows. The first is the standard Λ\LambdaCDM model built on GR (which includes cold dark matter, radiation including massless neutrinos, and baryons). The second is the Cubic Galileon model (based on CG gravity) for which we build our emulator, with the background fixed to the tracker solution described above, and with Λ\LambdaCDM initial conditions. The emulated power spectra we use to compute the boost come from Hi-COLA, which does not fully include the effects of radiation directly in its forward evolution. We partly compensate for the effect of radiation by using a backscaling approach (described in more detail in [42]). For the forecasts in § 6.2, the effects of radiation and baryons are included only insofar as they impact the Λ\LambdaCDM power spectrum, which is multiplied by the CG boost (defined in equation (4)) to compute the CG power spectrum; we do not expect the boost predictions to be significantly affected by these additional parameters. Massive neutrinos are not included in the present analysis. Because their imprint on the power spectrum cannot in general be decoupled from that of modified gravity (see Fig. 7 of [72]), a joint treatment is well-motivated, though on the mildly nonlinear scales probed by our forecasts we expect the impact to be modest. A treatment of massive neutrinos will be implemented in the next version of Hi-COLA (in prep.), so future extensions of this emulator to models whose background evolution can cross the phantom divide, including to Asymptotic Cubic Galileons (ACG, [20]) or Rolling Galileons [73], will be able to include massive neutrinos self-consistently.

The Hi-COLA code

Horndeski-in-COLA (Hi-COLA4) [42] is a hybrid particle-mesh (PM) code, which utilises the Co-moving Lagrangian Acceleration (COLA) [74] scheme to simulate the motion of dark matter particles under gravity more rapidly than traditional NN-body codes. To give a summary of COLA’s key ideas: it takes advantage of the known analytic solutions for the large-scale clustering of matter through Lagrangian Perturbation Theory (LPT). The PM code solves for the residual displacement between the (second-order) LPT observer and the particles, by performing a boost to the frame of observers that move according to LPT. The resulting numerical solution of the particle geodesic equation converges faster than if the geodesic equation was solved in the initial rest frame of the particles. This allows for simulations that can run with fewer time steps while remaining accurate in the mildly nonlinear regime. Hi-COLA adapts COLA for use in modified gravity by including the effects of the scalar field on the expansion history and the linear growth factors. In the case of the latter, this leads to a modification of the growth factor equations, as detailed in Section 4.2 of [42]. For theories like the Cubic Galileon, which feature scale-independent growth factors, solving these equations remains a computationally inexpensive exercise (see [75] for a discussion on dealing with the computational expense of this exercise for theories with scale-dependent growth, such as f(R)f(R)).

Hi-COLA further includes the nonlinear impacts of the scalar field on structure formation through a fifth force (building on [76, 77]), making use of the quasi-static approximation (QSA) and weak field approximation [78]. Hi-COLA uses the screening factor approach of [79]. This work showed that the force modification factor derived in a spherically symmetric configuration can be imported to modify the inter-particle forces in a general simulation, with a minimal loss of accuracy at scales k<1h Mpc−1k < 1 h\,\mathrm{Mpc}^{-1}. This leads to a force expression that can be computed without solving the scalar Klein-Gordon equation, the latter of which is one of the reasons why MG is slower to simulate than GR (see [79] for a discussion).

Where Hi-COLA differs from some of its contemporaries such as MG-PICOLA [80], COLA-FML [81], MG-evolution [82] and the f(R)f(R) COLA code of [75] is the use of symbolic manipulation to enable a construction of the background equations of motion and approximate fifth force for a generic choice of Horndeski functions (K,G3,G4)(K, G_3, G_4). Thus, Hi-COLA is designed as a code that is immediately applicable to a wide class of Horndeski theories, rather than specifically addressing select members of the class; and consistently solves for theories from their Lagrangian formulation, instead of relying on approximate parametrisations of the theory space. Currently, Hi-COLA has been used to study the clustering phenomenology of Horndeski theories with Vainshtein screening [42] and K-mouflage gravity [83, 84], as well as finding viable models with phantom-crossing behaviour [20, 73].

Data Generation for Emulator Development and Validation

Emulator Training Set: experimental design

In this work we run simulations for both the Λ\LambdaCDM model and the Cubic Galileon model described in § 2.1. We will be emulating the Cubic Galileon boost, defined as:

B(z,k,p)=PδCuGal(z,k,p)PδΛCDM(z,k,p)(4)B(z,k,\mathbf{p}) = \frac{P_{\delta}^{\mathrm{CuGal}}(z,k,\mathbf{p})}{P_{\delta}^{\Lambda\mathrm{CDM}}(z,k,\mathbf{p})} \tag*{(4)}

where PδP_\delta is the 3D matter power spectrum computed with Hi-COLA and the labels stand for the Cubic Galileon or Λ\LambdaCDM model. This quantity will be computed as a function of wavenumber kk, redshift zz and a set of parameters p={Ωm0,ns,As,h,fϕ}\mathbf{p} = \{\Omega_{m0}, n_s, A_s, h, f_\phi\}, over the ranges:

0.275≤Ωm0=Ωc0+Ωb0≤0.330.85≤ns≤1.11.45×10−9≤As≤3.3×10−90.61≤h≤0.730.02≤fϕ≤1(5)\begin{aligned} 0.275 &\leq\Omega_{m0} = \Omega_{c0} + \Omega_{b0} \leq0.33 \\ 0.85 &\leq n_s \leq1.1 \\ 1.45 \times10^{-9} &\leq A_s \leq3.3 \times10^{-9} \\ 0.61 &\leq h \leq0.73 \\ 0.02 &\leq f_\phi\leq1 \tag*{(5)} \end{aligned}

where Ωm0\Omega_{m0} is the ratio of today’s total matter density to today’s critical density of the Universe and hh is the dimensionless Hubble constant defined by H0=100h km s−1Mpc−1H_0 = 100 h\,\mathrm{km}\,\mathrm{s}^{-1}\mathrm{Mpc}^{-1}, with Hubble parameter today, H0H_0, in km s−1Mpc−1\mathrm{km}\,\mathrm{s}^{-1}\mathrm{Mpc}^{-1}. Additionally, AsA_s is the primordial super-horizon amplitude of curvature perturbations at k=0.05h Mpc−1k = 0.05 h\,\mathrm{Mpc}^{-1}, nsn_s is the primordial spectral index, and fϕf_\phi is the additional Cubic Galileon parameter described in § 2.1. We compute the boost with constant ωr=Ωrh2=4.18×10−5\omega_r = \Omega_r h^2 = 4.18 \times10^{-5}, equivalent to assuming a constant TCMB=2.725KT_{\mathrm{CMB}} = 2.725\mathrm{K} and massless neutrinos with neff=3.044n_{\mathrm{eff}} = 3.044. We also fix the physical baryon density ωb=Ωbh2=0.0223\omega_b = \Omega_b h^2 = 0.0223, assuming the boost is insensitive to its variation. This is motivated by Figure 10 in [85], which shows that the Λ\LambdaCDM power spectrum is largely unaffected by ωb\omega_b (assuming tight priors on this parameter from CMB constraints). We extend this assumption to a beyond-Λ\LambdaCDM model, and justify this extrapolation in part because modifications are expected to be smaller for the boost than for the power spectrum (see e.g. Figure 1 in [86]). Selection of the ranges of parameters depends on the desired emulator-accuracy requirements and the coverage required for exploration during the Bayesian calibration process. Our cosmological parameter ranges are consistent with those of Coyote Universe [85] and Mira-Titan Universe simulation suites [87], where the precision goals of CMB and large-scale-structure surveys guide the choices. The modified-gravity parameter, by contrast, is sampled over the full range permitted by the theory. This extends beyond the limit imposed by current constraints on Cubic Galileon gravity, but it is standard practice to train MG emulators over a wide MG parameter range. For example, Hu-Sawicki f(R)f(R) is constrained in Refs. [88, 89] to log⁡10∣fR0∣≲−8\log_{10}|f_{R0}| \lesssim-8 at the 95% C.L., but emulators for this theory have been trained up to fR0∼10−4f_{R0} \sim10^{-4}[86, 90, 91]. The reason is that an emulator is used for more than deriving constraints for the specific modified gravity theory, for example serving in pipeline validation and as a building block for alternative modelling schemes [92, 93].

Figure 1. Visualization of the training boosts as a function of kk, zz and fϕf_\phi. Left: Plot of the training boost (defined in 3.1) as a function of kk at z=0z = 0 for our training data. The colour describes variation in the parameter fϕf_\phi. Right: Plot of the training boost as a function of kk and zz for our training data (at our lowest value of fϕf_\phi, fϕ=0.02f_\phi= 0.02). The colour describes variation in redshift (for clarity, we don’t display the redshifts for z≥1z \ge1, opting for dotted black lines). This plot shows that fϕf_\phi is sampled densely enough close to the Λ\LambdaCDM prediction (B=1B = 1).

Figure 10. Posterior constraints from an LSST Y1 3×2pt3 \times2\mathrm{pt} forecast using mock data generated from Hi-COLA for a validation point located near the fiducial cosmology within the LHS (with fϕ=0.99f_{\phi}=0.99). The theoretical model uses the Cubic Galileon emulator. No significant bias is observed in either cosmological or nuisance parameters.

Our sampling strategy involves finding these parameters at 50 points spaced uniformly in parameter space using a 5D Symmetric Latin Hypercube (SLH) [85] sample. Alternative Space-Filling Sampling (SFS) methods include low-discrepancy sequences (e.g., Sobol’s sequence). These have the advantage of not requiring a predetermined training sample size; thus, increasing the size of the training set does not require generating an entirely new sample. We opt for an SLH sample because, although it is difficult to determine an appropriate training sample size in advance, a very similar emulation setup was used in [91] for comparable purposes, offering a reasonable starting point for subsequent validation in our case. These SLH sample training points for the emulator are displayed in blue in Figure 3.

Figure 3. 50 parameter combinations used for our second validation set (in red) compared to the 50 parameter combinations used in the emulator development (in blue).

Emulator Training Set: evaluating the boosts

We compute the power spectra using the Hi-COLA software, following the specifications from Table 1, for both the Λ\LambdaCDM and the Cubic Galileon models with the same initial random seed. In order to account to some degree for the effect of radiation in the simulations, the initial conditions were generated from power spectra at z=0z = 0 that include a radiation component T=TCMBT = T_{\mathrm{CMB}}, which were then back-scaled to z=ziz = z_i using growth factors under the assumption of zero background radiation. We assume that the GR and Cubic Galileon initial conditions are the same at z=49z = 49, and use a Λ\LambdaCDM matter power spectrum to generate the initial fields for both of these models.

Box Size, LL400 Mpc/hh
Number of particles, NpartN_{\mathrm{part}}5123512^3
IC grid size, NmeshN_{\mathrm{mesh}}5123512^3
Grid size for computing particle mesh forces, NF,meshN_{\mathrm{F,mesh}}(3×512)3(3 \times512)^3
Number of time steps, NtN_t51
Initial Redshift, ziz_i49
Final Redshift, zfz_f0

Table 1. Specifications for the Hi-COLA training simulation suite.

For each node in our training set we run two simulations with initial modes exactly out of phase - that is, with matching amplitudes but phase-reversed initial conditions [94]. We do this in order to efficiently remove finite-volume effects from power spectrum estimates [95]. We validate this approach by comparing to the average over 10 random realizations and selecting the initial condition seed that (averaged between the two phase-reversed realizations) minimizes the error. The resulting boost averaged over the paired realizations:

B(z,k,p)=Pδ,Phase 1CuGal(z,k,p)+Pδ,Phase 2CuGal(z,k,p)Pδ,Phase 1ΛCDM(z,k,p)+Pδ,Phase 2ΛCDM(z,k,p)(6)B(z,k,\mathbf{p}) = \frac{P_{\delta,\mathrm{Phase\ 1}}^{\mathrm{CuGal}}(z,k,\mathbf{p}) + P_{\delta,\mathrm{Phase\ 2}}^{\mathrm{CuGal}}(z,k,\mathbf{p})}{P_{\delta,\mathrm{Phase\ 1}}^{\Lambda\mathrm{CDM}}(z,k,\mathbf{p}) + P_{\delta,\mathrm{Phase\ 2}}^{\Lambda\mathrm{CDM}}(z,k,\mathbf{p})} \tag*{(6)}

is within 2% of the mean over 10 realizations while requiring only twice the computing time of a single run.

The boosts were computed at 51 different redshifts in the interval 0≤z≤490 \le z \le49 linearly spaced in the scale factor aa, and for 768 logarithmically spaced wavevectors over the range 5×10−2 h Mpc−1≤k≤10 h Mpc−15 \times10^{-2}\,h\,\mathrm{Mpc}^{-1} \le k \le10\,h\,\mathrm{Mpc}^{-1}. The Hi-COLA power spectrum boosts have been validated to within 2.5% of high-fidelity N-body simulations for k≲1.2 h Mpc−1k \lesssim1.2\,h\,\mathrm{Mpc}^{-1} [42] for some Cubic Galileon models – the boost (and therefore the emulator) is at most reliable up to these scales. The validation boosts are shown in Figure 1. These qualitatively reproduce the characteristic behaviour expected from Cubic Galileon models: a plateau at small kk consistent with the linear prediction of scale-independent growth, a bump on quasi-linear scales arising from the enhanced collapse of structures whose fifth force has not yet been suppressed, and a turn-over at higher kk, where the Vainshtein screening mechanism causes the boost to tend towards the Λ\LambdaCDM prediction. All these features become more pronounced with increasing fϕf_\phi, reflecting the stronger modification to gravity as the scalar field contribution increases.

Finally, to reduce the noise arising from the finite simulation volume and resolution, the boost is smoothed (in wavevector kk) prior to emulator training. Following standard practice (e.g. [91, 96]), we use a Savitzky–Golay filter [97] with window length 13, polynomial order 3, preventing the emulator from learning the noisy features of the data and improving its accuracy.

Emulator Development

Having generated the suite of smoothed boosts described in § 3.1, we now construct an emulator that interpolates the boost B(z,k,p)B(z,k,\mathbf{p}) across the five-dimensional parameter space p={Ωm0,ns,As,h,fϕ}\mathbf{p} = \{\Omega_{\mathrm{m}0}, n_{\mathrm{s}}, A_{\mathrm{s}}, h, f_{\phi}\} at any redshift within the simulated ranges. Our emulator follows the approach pioneered by the CosmicEmu and Coyote Universe emulation program [85, 98, 99]. For each simulation snapshot, we compress the boost into a small set of orthogonal basis functions and fit Gaussian processes to the corresponding weights. Predictions at intermediate redshifts are then obtained by interpolation between snapshots.

Single-snapshot emulation

We first describe the emulator at a fixed simulation snapshot, where the boosts form a matrix of Nsim=50N_{\mathrm{sim}} = 50 simulated curves, each sampled on Nk=768N_k = 768 logarithmically spaced wavevectors. Rather than emulating the boost at every kk-bin independently, we compress this output onto a small set of orthogonal basis functions via a singular value decomposition (SVD). The smoothed boosts are centred and standardised, and we represent the boost as a truncated sum of basis vectors ϕm(k)\phi_m(k) and associated weights wm(p)w_m(\mathbf{p}),

B(k,p)≃Bˉ(k)+∑m=1nPCϕm(k)wm(p),(7)B(k,\mathbf{p}) \simeq\bar{B}(k) + \sum_{m=1}^{n_{\mathrm{PC}}} \phi_m(k)w_m(\mathbf{p}), \tag*{(7)}

where Bˉ(k)\bar{B}(k) is the mean boost over the training set and the number of retained components nPCn_{\mathrm{PC}} is chosen to capture 99.9% of the variance. Because the boost is a smooth, monotonic function of the parameters (see Figure 1 for the variation with fϕf_{\phi}), only a handful of components are required (between 1 and 4 depending on the redshift). The orthogonality of the basis vectors ϕm(k)\phi_m(k) also allows the weights wm(p)w_m(\mathbf{p}) to be treated independently for fitting Gaussian Process (GP) models.

We have training inputs {pi}\{\mathbf{p}_i\} and weights w={wm(pi)}\mathbf{w} = \{w_m(\mathbf{p}_i)\}. Then for each ww in w\mathbf{w}, the joint distribution of the training weights and the weight w∗w_* at a new parameter point p∗\mathbf{p}_* is taken to be Gaussian,

p(w,w∗)=N(0,[KK∗TK∗K∗∗]),(8)p(w,w_*) = \mathcal{N}\left(\mathbf{0}, \begin{bmatrix} \mathbf{K} & \mathbf{K}_*^{T} \\ \mathbf{K}_* & K_{**} \end{bmatrix}\right), \tag*{(8)}

where K\mathbf{K} is the covariance of the training points, with entries Kij=k(pi,pj)K_{ij} = k(\mathbf{p}_i,\mathbf{p}_j) given by the kernel evaluated on each pair of training inputs. The term K∗\mathbf{K}_* holds the covariances between the test point and the training points, with entries k(p∗,pi)k(\mathbf{p}_*,\mathbf{p}_i). The term K∗∗=k(p∗,p∗)K_{**} = k(\mathbf{p}_*,\mathbf{p}_*) is the prior variance of the test point.

Conditioning on the training data yields the predictive mean and variance at p∗\mathbf{p}_*,

μ(w∗)=K∗K−1w,(9)\mu(w_*) = \mathbf{K}_*\mathbf{K}^{-1}\mathbf{w}, \tag*{(9)}
σ2(w∗)=K∗∗−K∗K−1K∗T,(10)\sigma^2(w_*) = K_{**} - \mathbf{K}_*\mathbf{K}^{-1}\mathbf{K}_*^{T}, \tag*{(10)}

The predictive mean μ(w∗)\mu(w_*) is a weighted combination of the training weights, with the weighting set by how strongly the test point correlates with each training point through the kernel. The predictive variance σ2(w∗)\sigma^2(w_*) is smallest where the test point lies close to the training nodes and grows in sparsely sampled regions of parameter space.

The kernel function k(p,p′)k(\mathbf{p},\mathbf{p}') encodes how smoothly the weights vary across parameter space. We adopt a squared-exponential kernel for each weight, so that the correlation between two parameter points decays smoothly with their separation. Each kernel has a separate correlation length for each of the five input dimensions corresponding to the cosmology parameters, and a small nugget term:

k(p,p′)=τSE2exp⁡(−∥p−p′∥22l2)+τnoise2I(11)k(\mathbf{p},\mathbf{p}') = \tau_{\mathrm{SE}}^{2}\exp\left(-\frac{\lVert\mathbf{p}-\mathbf{p}'\rVert^{2}}{2l^{2}}\right) + \tau_{\mathrm{noise}}^{2}\mathbb{I} \tag*{(11)}

where I\mathbb{I} is the identity matrix. This form is appropriate because the boost responds smoothly to all five parameters across the ranges in eq. (5). The kernel hyperparameters are inferred within the Bayesian framework of [100] as implemented in SEPIA [101], sampling the posterior with a Markov Chain Monte Carlo of 1000 steps. The trained GPs are then combined with the basis vectors of eq. (7) to predict the boost and its uncertainty at any parameter point within the snapshot.

Redshift interpolation

The procedure above is repeated independently at each of the 51 simulation snapshots, giving a series of 51 separate emulators spanning 0≤z≤490 \leq z \leq49. We treat redshift separately, opting to not include it as a sixth dimension in the Latin hypercube design, or compress it via principal component analysis (like kk). This has a practical advantage: every simulation outputs the boost on the same fixed set of redshifts, so the training data shares a common time grid across all nodes, instead of each node sampling a different redshift as would follow from folding zz into the experimental design. It is also well motivated, since the boost varies smoothly with redshift. Hence, a full GP in that direction would be unnecessary.

To evaluate the boost at an arbitrary redshift zz between two stored snapshots z1z_1 and z2z_2, we deploy the single-redshift emulators at the two nearest snapshots and linearly interpolate between them,

B(z,k,p)=B(z2,k,p)+[B(z1,k,p)−B(z2,k,p)]z−z2z1−z2,(12)B(z,k,\mathbf{p}) = B(z_2,k,\mathbf{p}) + \left[B(z_1,k,\mathbf{p}) - B(z_2,k,\mathbf{p})\right]\frac{z-z_2}{z_1-z_2}, \tag*{(12)}

propagating the per-snapshot emulator errors through the same linear weighting. Validations of the redshift interpolation are shown in § 6, where the emulator is deployed at redshifts not used in the emulator training.

Modelling and Forecasting Methods

Having constructed the Cubic Galileon emulator, we now turn to its principal application in this work, namely as a forecasting tool for constraints on Cubic Galileon gravity with upcoming photometric survey data. The emulator’s speed makes it possible to model nonlinear modified gravity predictions directly within a Bayesian inference pipeline, a task that would be computationally prohibitive with simulations alone. In this section we describe the inference pipeline used in § 6.2, where we produce forecasts for the constraints LSST Y1 will place on the modified gravity parameter ϕ\phi, and additionally test the emulator’s accuracy for a range of scale cuts.

We consider the 3×2pt3 \times2\mathrm{pt} datavector, which has become a standard choice for photometric galaxy surveys [24, 25]. It combines three sets of two-point correlation functions: the auto-correlation of the galaxy density field (galaxy clustering), the auto-correlation of the weak lensing shear field (cosmic shear), and the cross-correlation between the two (galaxy-galaxy lensing). The 3×2pt3 \times2\mathrm{pt} datavector is particularly well suited to tests of gravity. Galaxy clustering is sensitive to the Newtonian potential (Φ\Phi), so to modifications of the Poisson equation, which can be parametrized through the function μ\mu:

k2Φ=−4πGNμ(a,k)a2ρmδ(13)k^2\Phi= -4\pi G_{\mathrm{N}}\mu(a,k)a^2\rho_m\delta \tag*{(13)}

while lensing responds to modifications to the equation governing the Weyl potential (Φ+\Phi_+) and therefore to Σ\Sigma:

k2Φ+=−4πGNΣ(a,k)a2ρmδ(14)k^2\Phi_+ = -4\pi G_{\mathrm{N}}\Sigma(a,k)a^2\rho_m\delta \tag*{(14)}

where GNG_{\mathrm{N}} is Newton’s constant, ρm\rho_m is the background density of matter and δ\delta is the comoving matter density contrast. μ\mu and Σ\Sigma are dimensionless phenomenological functions equal to unity in Λ\LambdaCDM. Because Σ\Sigma affects the photon geodesics only, it will affect the equations describing lensing, but not those describing galaxy clustering. These equations are true under the quasi-static approximation [102], neglecting anisotropic shear, and on linear scales. We can extend these definitions on nonlinear scales by re-defining the functions μ\mu and Σ\Sigma accordingly (e.g. [103]), but a full treatment is beyond the scope of this paper. Our forecasts make the simplifying assumption that Σ\Sigma takes its linear form throughout, but we note here that this is not guaranteed to be a valid approximation for the weak lensing angular power spectrum on nonlinear scales, and should be revisited in a realistic analysis. Additionally, the approximation does not affect the internal consistency of our bias tests, as the same treatment of Σ\Sigma enters both the mock datavectors and the model. We expect that modifications to posterior biases will therefore be small.

Additionally, this 3×2pt3 \times2\mathrm{pt} tracer combination allows a better handle over systematics. For example, the combination of lensing and clustering amplitudes alleviates the degeneracy between cosmological parameters and linear galaxy bias. In addition, the full 3×2pt3 \times2\mathrm{pt} datavector contains significant information from the non-linear regime, where screening mechanisms are expected to operate and where departures from GR acquire scale dependence. Modelling this datavector, however, requires a good handle over observational systematics such as intrinsic alignments, photometric redshift calibration, galaxy bias and baryonic feedback, especially on nonlinear scales [104].

In what follows, we describe the construction of the simulated data vector, our modelling choices and scale cuts, and the likelihood and sampling setup, before forecasting the resulting constraints on the modified gravity parameter fϕf_\phi and the cosmological parameters.

Priors and Parameter Values. Table 2 summarizes the fiducial values and priors used in our 3×2pt3 \times2\mathrm{pt} analysis for cosmology (pco\mathbf{p}_{\mathrm{co}}), modified gravity (fϕf_\phi), and nuisance (pnu\mathbf{p}_{\mathrm{nu}}) parameters used in our analysis. Nuisance parameters (bib^i, the linear galaxy bias parameters for each lens redshift bin i∈{1,…,ng}i \in\{1,\ldots,n_g\}, where ngn_g is the total number of lens bins) are taken from LSST Y1 forecasts [104] (with priors from [105]), and flat priors are imposed on cosmological parameters due to emulator limitations (as described in eq. (5)). The only informative prior is the joint Gaussian prior on ωb\omega_b and nsn_s. We introduce CMB-informed priors, without adopting the full Planck likelihood. Our prior comes from the marginalized posterior distributions for the Planck TT,TE,EE+lowE+lensing data using the late-time DE parametrization (outlined in Equations 56a and 56b in [2], see Fig. 31). These priors are motivated by the weak constraints that cosmic shear alone places on nsn_s and ωb\omega_b [106, 107], and by the expectation that any realistic MG analysis will incorporate CMB information to break degeneracies. The choice of informative prior is not unique: CMB priors are also commonly imposed on AsA_s, or alternatively less constraining BBN-only priors on ωb\omega_b can be adopted5. Since our aim here is to demonstrate a preliminary use case for the emulator rather than to characterise biases and evaluate degeneracies for a range of choices of priors and scale cuts, we leave further investigations along this line of inquiry to future work. The procedure to extract this prior is as follows. We marginalise the posterior over all parameters except ωb\omega_b and nsn_s, and then extract a joint covariance for these two parameters under the assumption that their joint posterior is Gaussian. Finally, we shift the joint mean to match the fiducial values of ωb\omega_b and nsn_s.

Simulated data. We will compute the correlation functions between two observable fields: cosmic shear and galaxy clustering. Because LSST is a photometric survey, galaxies are grouped into tomographic redshift bins based on their photometric redshift estimates (nsn_s bins for the source galaxies sample, and ngn_g bins for lens galaxies sample), and statistical correlations are then computed between the observable fields within these bins. Our 3×2pt datavector is the combination of three sets of angular power spectra (i.e. 2-point statistics)

D={Cδδii(ℓ),Cκδij(ℓ),Cκκij(ℓ)},\mathbf{D} = \{C_{\delta\delta}^{ii}(\ell), C_{\kappa\delta}^{ij}(\ell), C_{\kappa\kappa}^{ij}(\ell)\},

where:

  • Cδδii(ℓ)C_{\delta\delta}^{ii}(\ell) describes the correlations in the galaxy density contrast from galaxies in ngn_g different lens redshift bins, with i∈{1,…,ng}i \in\{1,\ldots,n_g\}, in projected Fourier mode ℓ\ell. We only include the autocorrelation in redshift bins.

  • Cκκij(ℓ)C_{\kappa\kappa}^{ij}(\ell) describes the correlations in the galaxy shear in nsn_s different source redshift bins, with i,j∈{1,…,ns}i,j \in\{1,\ldots,n_s\}, in projected Fourier mode ℓ\ell.

  • Cκδij(ℓ)C_{\kappa\delta}^{ij}(\ell) describes the cross-correlation of galaxy shear in the ithi^{\mathrm{th}} source galaxy bin with the galaxy density contrast of the jthj^{\mathrm{th}} foreground lens galaxy bin, with i∈{1,…,ns}i \in\{1,\ldots,n_s\} and j∈{1,…,ng}j \in\{1,\ldots,n_g\}, in projected Fourier mode ℓ\ell.

We compute the angular power spectrum by using the Cubic Galileon emulator boost in combination with the Core Cosmology Library6 (CCL) [108], which uses the halofit transformation of [109] to estimate the nonlinear Λ\LambdaCDM 3D matter power spectrum. The angular power spectrum between redshift bin ii of observable AA and redshift bin jj of observable BB at multipole ℓ\ell, under the Limber approximation [110] is:

CABij(ℓ)=∫dχqAi(χ)qBj(χ)χ2PCuGal(ℓ+12χ,z(χ)).(15)C_{AB}^{ij}(\ell) = \int\mathrm{d}\chi\frac{q_A^i(\chi)q_B^j(\chi)}{\chi^2}P_{\mathrm{CuGal}}\left(\frac{\ell+\frac{1}{2}}{\chi},z(\chi)\right). \tag*{(15)}

The tracers AB are either cosmic shear (κκ\kappa\kappa), galaxy-galaxy clustering (δδ\delta\delta, labeled GC) or their cross correlation (δκ\delta\kappa, labeled XC), χ\chi is the comoving radial distance and ii and jj denote redshift bin indices. qκi(χ)q_\kappa^i(\chi) and qδi(χ)q_\delta^i(\chi) are line-of-sight kernels, which weigh the power spectrum across all distances along the line of sight. These are given by:

qκi(χ)=3H02ΩmχΣCuGal(z(χ))2a(χ)∫χχhdχ′(χ′−χχ′)nsi(z(χ′))dzdχ′,q_\kappa^i(\chi) = \frac{3H_0^2\Omega_m\chi\Sigma_{\mathrm{CuGal}}(z(\chi))}{2a(\chi)}\int_\chi^{\chi_h}\mathrm{d}\chi'\left(\frac{\chi'-\chi}{\chi'}\right)n_s^i(z(\chi'))\frac{\mathrm{d}z}{\mathrm{d}\chi'},
qδi(χ)=bingi(z(χ))dzdχ.(16)q_\delta^i(\chi) = b^i n_g^i(z(\chi))\frac{\mathrm{d}z}{\mathrm{d}\chi}. \tag*{(16)}

where aa is the scale factor and bib^{i} is a constant galaxy bias for each lens redshift bin (that can in theory be scale- as well as redshift-dependent). The function ΣCuGal(z)\Sigma_{\mathrm{CuGal}}(z) describes deviations from GR in the lensing potential, and ΣCuGal(z)=μCuGal(z)\Sigma_{\mathrm{CuGal}}(z) = \mu_{\mathrm{CuGal}}(z) [57, 111, 112]. § 2.2 of [42] confirms that this equality holds on nonlinear scales. As mentioned above, in this paper we make the further assumption that Σ=μlinear\Sigma= \mu_{\mathrm{linear}}. ns,gi(z)n^{i}_{s,g}(z) denote the redshift distributions of the different source/lens redshift bins, normalized to unity (see Figure 2). Following [104], we assume a parametric redshift distribution for our overall underlying ‘true’ source and lens distributions:

Figure 2. The normalized galaxy number density for the source (solid) and lens (dashed) distributions for the 5 tomographic redshift bins from low-zz (darkest) to high-zz (lightest).

dns,gdz∝z2exp⁡[−(z/z0)α].(17)\frac{\mathrm{d}n_{s,g}}{\mathrm{d}z} \propto z^{2}\exp\left[-\left(z/z_{0}\right)^{\alpha}\right]. \tag*{(17)}

The source sample is separated in 5 bins with equal number counts, and with {z0,α}={0.13,0.78}\{z_{0}, \alpha\} = \{0.13, 0.78\} and with photo-zz scatter σz=0.05(1+z)\sigma_{z} = 0.05(1+z). The lens sample is separated in 5 bins separated by 0.2 in photo-zz in the redshift interval 0.2≤z≤1.20.2 \le z \le1.2, with {z0,α}={0.26,0.94}\{z_{0}, \alpha\} = \{0.26, 0.94\} and with photo-zz scatter σz=0.03(1+z)\sigma_{z} = 0.03(1+z). Our model for this follows [104] in approximating the effect of photo-zz uncertainty without simulating a more complicated treatment of photo-zz errors.

To ensure our forecasts are not biased by the specific nonlinear model assumptions built into our Λ\LambdaCDM matter power spectrum predictions, we construct all power spectra consistently. Both for the simulated datavector and for the model predictions, rather than using the nonlinear Cubic Galileon power spectrum directly (from the simulations), we compute the Cubic Galileon power spectrum as the product of its boost and a Λ\LambdaCDM model baseline:

PCuGal=BCuGal×PΛCDM,CCL.(18)P_{\mathrm{CuGal}} = B_{\mathrm{CuGal}} \times P_{\Lambda\mathrm{CDM},\mathrm{CCL}}. \tag*{(18)}

This approach isolates the MG signal and avoids importing modelling differences from the GR calculation between CCL (halofit) and the N-body codes (both Hi-COLA and EFT-RAMSES) into our MG forecasts. The boost BB is either the one predicted from the emulator or from a Hi-COLA or EFT-RAMSES simulation. The extrapolation procedure for the latter is described in § 6.2.3. For the former, we extrapolate beyond z=12z = 12 to higher zz by assuming that the power spectrum is close to GR (so we set B(k,z≥12)=1B(k, z \ge12) = 1; this is a good approximation because the MG boost is always within 0.3% of 1 at those redshifts). Additionally, if k≥5 h Mpc−1k \ge5\,h\,\mathrm{Mpc}^{-1}, we use a power-law extrapolation for the boost: B(k)=Ak−p+1B(k) = Ak^{-p} + 1, with p≥0p \ge0 always (for a similar approach, see e.g., [107]).

Since the χ2\chi^2-based scale cuts of § 6.2 require an input datavector, before applying scale cuts we model 20 logarithmically spaced ℓ\ell bins over 20≤ℓ≤1500020 \le\ell\le15000, following [104] to match the DESC Science Requirements Document datavector. In practice these high-ℓ\ell scales do not enter our analysis, as the baryonic cuts restrict us to ℓ\ell of a few hundred and our less conservative cuts to ≤1500\le1500.

For galaxy clustering, we only consider the lens bins auto-correlations Cδgδgii(ℓ)C_{\delta_g\delta_g}^{ii}(\ell). Cross-correlations between lens bins do not contribute significantly to the signal, which rather than carrying genuine cosmological information is dominated by photo-zz outlier systematics [113]. For galaxy-galaxy lensing correlations, we only consider lens and source bin combinations where the source bins are at a higher redshift than the lens bins (allowing for at most a 10% overlap). The selected bin pairs are (nsi,ngj)={(3,1),(4,1),(5,1),(4,2),(5,2),(5,3),(5,4)}(n_s^i,n_g^j) = \{(3,1),(4,1),(5,1),(4,2),(5,2),(5,3),(5,4)\} [104].

Baryonic scale cuts. While all the power spectra presented in this work are dark matter only (DMO), a real datavector will be affected by baryonic physics, in particular by the impact of galaxy formation processes (e.g., AGN feedback, cooling, star formation) on the distribution of matter [114]. To produce more realistic and conservative forecasts, we account for this lack of baryonic physics modelling in our angular power spectra by applying scale cuts in ℓ\ell. We adopt this approach because baryonic modelling in modified gravity remains relatively poorly understood. Baryonification schemes are available7 and have been applied in modified gravity forecasts, typically as an additional baryonic boost to DMO power spectra [45]. Investigation of the degeneracies between the baryonification parameters and the modified gravity parameters are being undertaken in order to ensure these tools can be used in a realistic MG analysis. [72] first showed that, in a halo model context, MG and baryons can be modelled independently, and the limitations of this approximation have been tested for specific modified gravity theories (see e.g. [115]). Still, scale cuts are likely to be adopted in practice for early LSST MG analyses. The use of less conservative approaches is possible (for example retaining smaller scales via the BNT transform, see [116, 117]) and applying it when forecasting Cubic Galileon gravity constraints with 3×2pt3 \times2\mathrm{pt} data is a direction of future work.

We therefore describe now our procedure for cutting scales from the 3×2pt3 \times2\mathrm{pt} analysis that are sensitive to baryonic physics. There will be a maximum wavevector beyond which nonlinear bias results in significant deviations in the power spectrum from the linear bias, which we call kmax⁡k_{\max}. We fix this scale to kmax⁡=0.2h Mpc−1k_{\max}=0.2h\,\mathrm{Mpc}^{-1}, a slightly more conservative value than the one suggested in [104]. We note here that the linear galaxy bias model is expected to deliver unbiased (sub-percent) constraints only for k≲0.1 h/Mpck\lesssim0.1\,h/\mathrm{Mpc} [118]. We retain kmax⁡=0.2 h Mpc−1k_{\max}=0.2\,h\,\mathrm{Mpc}^{-1} as a compromise between this limit and the value adopted in the DESC SRD. Unbiased recovery at this kmax⁡k_{\max} implies unbiased recovery for any more conservative choice, such as kmax⁡=0.1 h Mpc−1k_{\max}=0.1\,h\,\mathrm{Mpc}^{-1}. We find ℓmax⁡\ell_{\max} using the following approximation (see e.g. [119]):

ℓmax⁡=kmax⁡χ(⟨z⟩)−0.5(19)\ell_{\max}=k_{\max}\chi(\langle z\rangle)-0.5 \tag*{(19)}

where ⟨z⟩\langle z\rangle the mean redshift of the relevant lens bin.

We use this method to implement scale cuts for Cδδij(ℓ)C_{\delta\delta}^{ij}(\ell) and Cδκij(ℓ)C_{\delta\kappa}^{ij}(\ell), but not for Cκκij(ℓ)C_{\kappa\kappa}^{ij}(\ell), given Equation (19) only works for narrow kernels and the lensing kernel is broad. For the latter, we will instead use a method based on [120]. This involves comparing two synthetic, noiseless datavectors for a DMO and baryonic-rescaled Λ\LambdaCDM universe computed at a fiducial cosmology:

Pbaryonic(pfid,k,z)=POWLSsim(k,z)PDMOsim(k,z)×PDMO(pfid,k,z)P_{\mathrm{baryonic}}(\mathbf{p}^{\mathrm{fid}},k,z)=\frac{P_{\mathrm{OWLS}}^{\mathrm{sim}}(k,z)}{P_{\mathrm{DMO}}^{\mathrm{sim}}(k,z)}\times P_{\mathrm{DMO}}(\mathbf{p}^{\mathrm{fid}},k,z)

where the power spectra come from the OWLS simulations [121], PDMOsim(k,z)P_{\mathrm{DMO}}^{\mathrm{sim}}(k,z) for DMO and POWLSsim(k,z)P_{\mathrm{OWLS}}^{\mathrm{sim}}(k,z) with AGN feedback only, matching the baryonic scenario adopted in the DES Y3 analysis [122]. After computing the shear angular power spectrum datavectors using Equation (15), we compute:

Δχ2=(DbaryonicΛCDM−DDMOΛCDM)TC−1(DbaryonicΛCDM−DDMOΛCDM)(20)\Delta\chi^2=\left(\mathbf{D}_{\mathrm{baryonic}}^{\Lambda\mathrm{CDM}}-\mathbf{D}_{\mathrm{DMO}}^{\Lambda\mathrm{CDM}}\right)^T\mathbf{C}^{-1}\left(\mathbf{D}_{\mathrm{baryonic}}^{\Lambda\mathrm{CDM}}-\mathbf{D}_{\mathrm{DMO}}^{\Lambda\mathrm{CDM}}\right) \tag*{(20)}

where C\mathbf{C} is the data covariance matrix described further below. We then identify the point that contributes the most to Δχ2\Delta\chi^2, remove it and repeat the process until Δχ2≤1\Delta\chi^2\leq1.

The criteria of Eq. (20), when enforced on the full 3×2pt3\times2\mathrm{pt} datavector up to ℓ=15000\ell=15000, retains counterintuitively the two largest ℓ\ell bins of the C11(ℓ)C^{11}(\ell) datavector for galaxy-galaxy lensing. While we expect the greatest difference in nonlinear vs linear modelling at the smallest scales, this condition is allowed by the criteria due to the role of the covariance matrix, potentially in combination with inaccurate modelling of the power spectrum for the largest ℓ\ell values. We therefore decide to additionally cut these points in our analysis; we enforce this by removing all ℓ≥ℓmax⁡Δχ2\ell\geq\ell_{\max}^{\Delta\chi^2} during the iteration.

Linear scale cuts. For comparison with the expected constraints of a forecast based on linear-only power spectrum modelling, we will implement the following scale cuts on the 3×2pt3\times2\mathrm{pt} datavector based on [120]:

Δχ2=(DNLΛCDM−DlinΛCDM)TC−1(DNLΛCDM−DlinΛCDM)≤1.\Delta\chi^2=\left(\mathbf{D}_{\mathrm{NL}}^{\Lambda\mathrm{CDM}}-\mathbf{D}_{\mathrm{lin}}^{\Lambda\mathrm{CDM}}\right)^T\mathbf{C}^{-1}\left(\mathbf{D}_{\mathrm{NL}}^{\Lambda\mathrm{CDM}}-\mathbf{D}_{\mathrm{lin}}^{\Lambda\mathrm{CDM}}\right)\leq1.

We do this by finding the point in the full 3×2pt3\times2\mathrm{pt} datavector that contributes the most to Δχ2\Delta\chi^2, removing it, and repeating the process until Δχ2≤1\Delta\chi^2\leq1.

Likelihood formalism. We describe here the general method for computing the likelihood in our simulated analysis. We first compute our mock data vector for some cosmological parameters (labeled ‘fid’ – unless otherwise specified, see Table 2 for the fiducial values) and some choice of fϕf_\phi:

D={Cδgδgii,NL(ℓ,pcofid,pnufid,fϕ),Cκδgij,NL(ℓ,pcofid,pnufid,fϕ),Cκκij,NL(ℓ,pcofid,pnufid,fϕ)}.\mathbf{D} = \{C_{\delta_g\delta_g}^{ii,\mathrm{NL}}(\ell,\mathbf{p}_{\mathrm{co}}^{\mathrm{fid}},\mathbf{p}_{\mathrm{nu}}^{\mathrm{fid}},f_\phi), C_{\kappa\delta_g}^{ij,\mathrm{NL}}(\ell,\mathbf{p}_{\mathrm{co}}^{\mathrm{fid}},\mathbf{p}_{\mathrm{nu}}^{\mathrm{fid}},f_\phi), C_{\kappa\kappa}^{ij,\mathrm{NL}}(\ell,\mathbf{p}_{\mathrm{co}}^{\mathrm{fid}},\mathbf{p}_{\mathrm{nu}}^{\mathrm{fid}},f_\phi)\}.

We do this by using a 3D matter power spectrum interpolated from either Hi-COLA (studied in § 6.2.2) or EFT-RAMSES simulations (studied in § 6.2.3). We then find our model datavector as a function of model parameters p\mathbf{p}:

M(p)={Cδgδgii,NL(ℓ,p),Cκδgij,NL(ℓ,p),Cκκij,NL(ℓ,p)}\mathbf{M}(\mathbf{p}) = \{C_{\delta_g\delta_g}^{ii,\mathrm{NL}}(\ell,\mathbf{p}), C_{\kappa\delta_g}^{ij,\mathrm{NL}}(\ell,\mathbf{p}), C_{\kappa\kappa}^{ij,\mathrm{NL}}(\ell,\mathbf{p})\}

using the 3D matter power spectrum prediction from the emulator. For chains that implement linear scale cuts, we use the 3D linear matter power spectrum instead. We then model the likelihood as Gaussian in the datavector and as a function of the parameters p\mathbf{p}:

ln⁡L(D∣p)∝−12[(D−M(p))TC−1(D−M(p))](21)\ln\mathcal{L}(\mathbf{D}|\mathbf{p}) \propto-\frac{1}{2}\left[(\mathbf{D}-\mathbf{M}(\mathbf{p}))^{T}\mathbf{C}^{-1}(\mathbf{D}-\mathbf{M}(\mathbf{p}))\right] \tag*{(21)}

where C is the full non-Gaussian Y1 3×2pt covariance8 described in [104] (created following the numerical prescription from [123]). While this covariance is pre-computed at different cosmological parameter values (pcofid,SRD={Ωm=0.3156,σ8=0.831,ns=0.9645,Ωb=0.0492,h=0.6727}\mathbf{p}_{\mathrm{co}}^{\mathrm{fid,SRD}} = \{\Omega_m = 0.3156, \sigma_8 = 0.831, n_s = 0.9645, \Omega_b = 0.0492, h = 0.6727\}) [104] to our fiducial ones [2], this should not significantly affect results (see [124]).

We sample the posterior defined by Equation (21) and the priors of Table 2 over the full cosmological, modified gravity, and nuisance parameter space in § 6.2 below. The machinery described in this section allows us to perform a simulated analysis, by generating D\mathbf{D} from an independent Hi-COLA or EFT-RAMSES simulation, while modelling M(p)\mathbf{M}(\mathbf{p}) with our emulator. Any offset between the recovered and input parameters directly measures the bias induced by emulator at the level of cosmological inference. At the same time, the resulting posteriors give an estimate of the constraining power of an LSST Y1 3 × 2pt analysis on fϕf_\phi. In the next section we present the results of these forecasts, alongside direct comparisons of the emulated boosts against validation Hi-COLA and EFT-RAMSES simulations.

Results

The emulator’s accuracy is limited by a range of factors. Firstly, approximations in the underlying simulations used to produce the training set introduce error: the finite simulation box size affects large-scale accuracy, and in order to reduce computational cost we use lower-resolution COLA simulations instead of full N-body, but in doing so decrease small-scale accuracy. Secondly, errors arise during emulator development due to the limited number of training samples, and the methods used for training data selection and interpolation.

To assess these effects, we conduct two kinds of tests. In the first, we compare the emulator’s predictions with those from Hi-COLA simulations that were not used in training. We do so by comparing the boost predictions directly (in § 6.1 below, where we compare with results from two additional sets of respectively 25 and 50 simulations), or by running simulated forecasts for 3×2pt LSST Y1 constraints based on Hi-COLA simulations (in § 6.2). This helps us measure the accuracy of the emulator across parameter space and its

ParameterFiducial ValuePriors
Ωm\Omega_m0.321U(0.275,0.33)\mathcal{U}(0.275, 0.33)
hh0.6688U(0.61,0.73)\mathcal{U}(0.61, 0.73)
pco\mathbf{p}_{\mathrm{co}} AsA_s2.092×10−92.092 \times10^{-9}U(1.45×10−9,3.3×10−9)\mathcal{U}(1.45 \times10^{-9}, 3.3 \times10^{-9})
wbw_b0.022383N([0.0223830.9626],[2.0381×10−53.6116×10−73.6116×10−72.5265×10−8])\mathcal{N}\left(\begin{bmatrix}0.022383\\0.9626\end{bmatrix},\begin{bmatrix}2.0381 \times10^{-5} & 3.6116 \times10^{-7}\\3.6116 \times10^{-7} & 2.5265 \times10^{-8}\end{bmatrix}\right)
nsn_s0.9626
pnu\mathbf{p}_{\mathrm{nu}} b1b_11.562362U(0.8,3.0)\mathcal{U}(0.8, 3.0)
b2b_21.732963U(0.8,3.0)\mathcal{U}(0.8, 3.0)
b3b_31.913252U(0.8,3.0)\mathcal{U}(0.8, 3.0)
b4b_42.100644U(0.8,3.0)\mathcal{U}(0.8, 3.0)
b5b_52.293210U(0.8,3.0)\mathcal{U}(0.8, 3.0)
fϕf_\phi1.0U(0.0,1.0)\mathcal{U}(0.0, 1.0)

caption: Priors and fiducial (i.e. simulated universe) parameters (pfid\mathbf{p}^{\mathrm{fid}}) for the cosmological, nuisance and modified gravity parameters in our simulated analysis. We quote the lower and upper limits of flat priors and the mean and standard deviation of Gaussian priors. The priors for the standard scale cuts (“Baseline Prior distributions”) match those in [105].

:::

effectiveness as a forecasting tool for LSST Y1, relative to Hi-COLA simulations of the same kind it was trained on. In the second test, we compare the emulator’s performance against a higher resolution EFT-RAMSES N-body simulation to evaluate its overall effectiveness, by running simulated forecasts for 3×23 \times2pt LSST Y1 constraints (in § 6.2.3 below)9. This allows us to assess the performance of the emulator by separately investigating errors from both the COLA simulations and the emulator construction.

Emulator Accuracy and Validation: comparison with Hi-COLA

Here we introduce the Hi-COLA simulated datasets used for validation.

Dataset 1: varying fϕf_\phi only We fixed the cosmological parameters to the fiducial values in Table 2 and sampled fϕf_\phi uniformly between 0 and 1, choosing values that lie between those used in the emulator’s training set (in red, compared with the training-set values in blue). We also increased the number of redshift timesteps to 100, both to improve the accuracy of the simulations and to assess the emulator’s accuracy at redshifts other than those it was trained on.

Dataset 2: sampling the centre and edges The second validation dataset samples 50 points in parameter space, chosen to give representative coverage of both the centre and the edges of the parameter hypervolume. This allows us to assess the performance both in the region most important for parameter estimation (near the fiducial parameter values) and in the areas we expect the emulator will perform the worst (the edges). The hyperparameter space spanned by these points (in red) as compared to the training dataset (in blue) is displayed in Figure 3.

We now assess the accuracy of the emulator by comparing its boost predictions to the boost from Hi-COLA simulations across the two validation datasets.

Figure 4 shows the comparison for the first validation set, in which only the parameter fϕf_{\phi} is varied while all other cosmological parameters are fixed to our fiducial cosmology. The agreement improves as fϕf_{\phi} decreases, i.e. as the GR limit is approached and the boosts become progressively flatter. This behaviour is common to emulators of this type [46, 91]. The emulator predictions closely follow the Hi-COLA boosts at z=0z = 0. The corresponding relative errors are shown in the lower panel. We find that the emulator achieves ∼1%\sim1\% accuracy for k≲0.9 h Mpc−1k \lesssim0.9\,h\,\mathrm{Mpc}^{-1} (with less than 2%2\% error for 3σ3\sigma of the validation sample for k≲1.2 h Mpc−1k \lesssim1.2\,h\,\mathrm{Mpc}^{-1}), despite the validation data not being smoothed with the Savitzky-Golay filter (unlike for the training set). When looking at the emulator performance over the full range of redshifts (see Figure 5), we additionally find that the emulator remains accurate at <3%< 3\% for k≲1.2h Mpc−1k \lesssim1.2h\,\mathrm{Mpc}^{-1}. This demonstrates that the emulator generalizes well across variations in fϕf_{\phi} in regions of parameter space we expect to sample densely during parameter inference.

Figure 4. Top panel: Boosts predicted by the Cubic Galileon emulator (dashed lines) and the Hi-COLA boosts (solid lines) at z=0z = 0 for the validation dataset 1. Bottom panel: relative error between the emulator prediction and the Hi-COLA boosts at z=0z = 0 for the same validation set. In gray, we also show boost residuals at the remaining fϕf_{\phi} values for which there was validation data. The emulator is trained on a set smoothed with a Savitzky–Golay filter (window length 13, polynomial order 3), while the validation set was not smoothed. Errors remain ∼1%\sim1\% for k≲0.9h Mpc−1k \lesssim0.9h\,\mathrm{Mpc}^{-1} and under 2%2\% for k≲1.2h Mpc−1k \lesssim1.2h\,\mathrm{Mpc}^{-1}.

Figure 5. Top panel: Boosts predicted by the Cubic Galileon emulator (dashed lines) and the Hi-COLA boosts (solid lines) for the first validation set described in § 6 with fϕ=1f_\phi= 1 at a range of redshifts up to z=0.29z = 0.29. Bottom panel: relative error between the emulator prediction and the Hi-COLA boosts for the same validation set and redshift range. In gray, we show boost residuals at the remaining redshifts for which there was validation data (which includes redshifts up to z=49z = 49). The emulator is trained on a set smoothed with a Savitzky–Golay filter (window length 13, polynomial order 3), while the validation set was not smoothed.

Figure 6 presents the results for the second validation set, which samples both the centre and edges of the parameter space. As expected, the errors increase toward the edges of the parameter volume, where the training data are sparser. Nevertheless, the performance remains good: 68% of the predictions exhibit sub-percent accuracy (<0.5%< 0.5\%), 95% of the predictions have errors below 2.5%, and 99.7% below 5% for k<1.2h Mpc−1k < 1.2h\,\mathrm{Mpc}^{-1} at z=0z = 0. We also assess the emulator’s performance across the full redshift range (see for example Figure 7). Naively, one might expect the accuracy to improve monotonically with redshift, since the boost approaches unity as zz increases, and the modified gravity signal tends to the Λ\LambdaCDM initial conditions. In practice, this improvement doesn’t occur at intermediate redshifts: up to z∼0.7z \sim0.7, the errors are comparable to those at z=0z = 0. This is partly attributable to the validation power spectra being computed with a larger number of timesteps than the emulator was trained on so that the expected improvement in accuracy is balanced out by the redshift interpolation error. We conclude that the emulator reproduces the boost well across redshift with a constant accuracy at low-to-intermediate zz, beyond which the accuracy increases with redshift. Overall, 68% of the predictions have an error ≲1.2%\lesssim1.2\%, 95% below 2.5%, and 99.7% below ∼5%\sim5\% for k<1.2h Mpc−1k < 1.2h\,\mathrm{Mpc}^{-1}.

Figure 6. Top panel: Boosts predicted by the Cubic Galileon emulator (dashed lines) and the Hi-COLA boosts (solid lines) at z=0z = 0 for the validation set shown in Figure 3. The clustering into two groups (large vs. small boost) reflects the underlying validation values of fϕf_{\phi}, which lie close to either 0 or 1 (see Figure 3). Bottom panel: relative error between the emulator prediction and the Hi-COLA boosts at z=0z = 0 for the same validation set. The emulator is trained on a set smoothed with a Savitzky–Golay filter (window length 13, polynomial order 3), while the validation set was not smoothed. 68% of the emulated data has sub-percent errors, 95% has an error <2.5%< 2.5\%, and 99.7% at <5%< 5\% for k<1.2h Mpc−1k < 1.2h\,\mathrm{Mpc}^{-1}.

Figure 7. Top panel: Boosts predicted by the Cubic Galileon emulator (dashed lines) and the Hi-COLA boosts (solid lines) for a point in the second validation set with {Ωm=0.284,Ωb=0.0600,h=0.610,ns=0.871,As=1.479×10−09,fϕ=0.974}\{\Omega_m = 0.284, \Omega_b = 0.0600, h = 0.610, n_s = 0.871, A_s = 1.479 \times10^{-09}, f_\phi= 0.974\}, for a range of redshifts up to z=49z = 49. Bottom panel: relative error between the emulator prediction and the Hi-COLA boosts for the same validation set and redshift range. In gray, we show boost residuals at the remaining redshifts for which there was validation data. The emulator is trained on a set smoothed with a Savitzky–Golay filter (window length 13, polynomial order 3), while the validation set was not smoothed.

Overall, these results indicate that the emulator reliably reproduces Hi-COLA boosts across the relevant parameter space, with highest accuracy near the centre, and significant but sparse degradation toward the boundaries. This reduction in accuracy occurs primarily in sparsely sampled and physically unrealistic regions of parameter space, and is not expected to significantly impact cosmological parameter inference. In particular, we show that the emulator is accurate relative to Hi-COLA simulations to enable unbiased parameter estimation for LSST Y1 3×2pt3 \times2\mathrm{pt} data with realistic scale cuts for physically motivated parameter values (see § 6.2.2 below).

Parameter inference for Stage IV-like data

We now forecast the constraints that LSST Y1 3×2pt3 \times2\mathrm{pt} data could place on the Cubic Galileon model, using our emulator to model the modified gravity matter power spectrum, and propagating its predictions through to the observable two-point statistics. It takes between 2.5 and 4 hours on 76 cores to run an individual chain. We have described the methods and forecasting choices for this in § 5.

Null test: forecasts for LSST Y1 relative to a CDM universe

We expect a well-behaved pipeline should return an upper limit on the modified gravity parameter rather than a spurious detection when constraining a simulated universe with fϕ=0f_\phi= 0 (i.e. a Λ\LambdaCDM universe, to which we expect a realistic scenario of a future LSST Y1 3×23 \times2pt analysis to be close). A preference for fϕ>0f_\phi> 0 in a GR universe would indicate model misspecification, for example from residual emulator inaccuracies being absorbed into the modified gravity sector. We therefore perform a null test, following the approach of recent Stage IV forecasts for f(R)f(R) gravity [125]. We generate a noiseless synthetic 3×23 \times2pt datavector in which the matter power spectrum is computed for GR using CCL. We then analyse it with the Cubic Galileon emulator, using the same sampler settings, baryonic scale cuts and maximum wavevector for galaxy clustering, kmax⁡=0.2 h Mpc−1k_{\max} = 0.2\,h\,\mathrm{Mpc}^{-1}, as in our baseline analysis (§ 5). We also impose Planck priors on ωb\omega_b and nsn_s.

Figure 8 shows the resulting posteriors. The recovered cosmological and nuisance parameters are consistent with their input values, and we find no preference for a departure from GR. The marginalised posterior on fϕf_\phi peaks at the Λ\LambdaCDM boundary, fϕ=0f_\phi= 0, and we obtain upper limits of

Figure 8. Posterior constraints from an LSST Y1 3 × 2pt forecast using mock data generated in GR, analysed with the Cubic Galileon emulator and Planck priors. The recovered cosmological and nuisance parameters are unbiased, and the posterior on fϕf_\phi is consistent with zero, giving fϕ<0.202f_\phi< 0.202 at 95% CL.

fϕ<0.112  (68% CL),fϕ<0.202  (95% CL).(22)f_\phi< 0.112\; (68\% \ \mathrm{CL}), \qquad f_\phi< 0.202\; (95\% \ \mathrm{CL}). \tag*{(22)}

Our pipeline therefore passes the null test. We note here that these bounds depend on our choice of prior on fϕf_{\phi}, the Planck prior on nsn_s and ωb\omega_b, and our systematic modelling choices, and because of the latter two they are optimistic for our particular scale cut choice.

Forecasts for LSST Y1 relative to a Hi-COLA simulated universe

We sample our posterior parameter space using the nautilus [126] sampler, with minimum effective sample size of 20000 and a maximum live set evidence fraction of 0.01. Figure 9 shows the posteriors for our analysis with a simulated noiseless datavector which uses a matter power spectrum generated with Hi-COLA, for a validation point with fϕ≈0.16f_{\phi} \approx0.16 (a more realistic and as of yet unconstrained value of fϕf_{\phi}), with a maximum wavevector for galaxy clustering angular power spectra of kmax⁡=0.2h Mpc−1k_{\max} = 0.2h\,\mathrm{Mpc}^{-1}. The emulator produces unbiased constraints on the cosmological and modified gravity parameters for both the baryonic cuts described in § 5 and cuts at ℓmax⁡=1500\ell_{\max} = 1500, and increases the significance of the detection in fϕf_{\phi} relative to linear-only modelling from 1.9σ1.9\sigma to 2.6σ2.6\sigma.

Figure 9. Posterior constraints from an LSST Y1 3×23 \times2pt forecast using mock data generated from Hi-COLA for a validation point located near the fiducial cosmology within the LHS (with fϕ≈0.16f_\phi\approx0.16). The theoretical model uses the Cubic Galileon emulator. No significant bias is observed in either cosmological or nuisance parameters.

We also test the robustness of the emulator under less conservative scale cuts for a stronger modification to Λ\LambdaCDM. Because we are assessing the bias introduced by the emulator, we design a worst-case scenario by changing two things at once relative to the previous test. We increase fϕf_{\phi}, increasing the MG signature and the error on the boost, and we relax the scale cuts, introducing scales where the emulator inaccuracies are largest. Figure 10 shows the posteriors for our analysis with a simulated noiseless datavector which uses a Hi-COLA matter power spectrum with fϕ=0.99f_{\phi} = 0.99, with a maximum wavevector for galaxy clustering angular power spectra of kmax⁡=0.2h Mpc−1k_{\max} = 0.2h\,\mathrm{Mpc}^{-1}. Both with baryonic cuts for shear, and with uniform scale cuts for shear of ℓmax⁡fk=1500\ell_{\max}^{\mathrm{fk}} = 1500, we recover unbiased constraints in our cosmological and nuisance parameters. The emulator remains reliable for a large departure from Λ\LambdaCDM and under more aggressive scale cuts than our baseline analysis. We caution that this conclusion depends on our scale cut scheme. A nulling scheme such as the BNT transform [116], which retains information from smaller angular scales, could expose inaccuracies in the forecast masked by the decreased statistical significance of the detection after conservative scale cuts. In that case, the unbiased recovery reported here would not necessarily hold.

Forecasts for LSST Y1 relative to full N-body simulations

The forecast in § 6.2.2 isolates the emulator construction error and demonstrates that the emulator recovers unbiased parameters when the mock data are generated with Hi-COLA. It characterises the error in the emulator relative to the Hi-COLA simulations it was trained on, but it doesn’t capture the error inherent in the COLA approximation, which trades small-scale accuracy for speed ([42] finds that Hi-COLA simulations are valid at the 2.5% level down to kmax⁡=1.2h Mpc−1k_{\max} = 1.2h\,\mathrm{Mpc}^{-1} at z=0z = 0). To quantify this contribution to the total error budget of the Cubic Galileon emulator, we compare Hi-COLA with a high-resolution N-body simulation run with EFT-RAMSES [51], an adaptive mesh refinement code for modified gravity built on RAMSES. To do this, we use the power spectra for a Cubic Galileon model on the tracker solution with fϕ=1f_\phi= 1 and Ωm=0.3138\Omega_m = 0.3138, Ωb=0.0493\Omega_b = 0.0493, h=0.6736h = 0.6736, ns=0.9649n_s = 0.9649, σ8=0.8143\sigma_8 = 0.8143 for two phase-reversed simulations, together with matched GR runs sharing the same initial conditions, so that the simulated boost is computed consistently with the emulator. While Cubic Galileon without a cosmological constant has been ruled out experimentally, fϕ=1f_\phi= 1 the maximal departure from Λ\LambdaCDM in our parameter range, ensuring we consider the most stringent test of the COLA approximation. Here, we apply a more demanding test than that in § 6.2.2, by analysing a mock datavector built from the full N-body EFT-RAMSES simulation, so that the COLA approximation error now enters as a potential source of bias on the inferred parameters.

We construct the simulated datavector from the EFT-RAMSES Cubic Galileon power spectrum. The boosts and residuals for the Hi-COLA simulation relative to the EFT-RAMSES simulation and the emulator are displayed in Figure 11. First, we estimate the theoretical Cubic Galileon boost from the average of two realizations with phase-shifted ICs, B(k,z)=(P1CG+P2CG)/(P1ΛCDM+P2ΛCDM)B(k,z) = (P^{\mathrm{CG}}_1+P^{\mathrm{CG}}_2)/(P^{\Lambda\mathrm{CDM}}_1+P^{\Lambda\mathrm{CDM}}_2), and smooth it with the same Savitzky–Golay filter (window length 13, polynomial order 3) applied to the training set. Outside the simulated range, we follow a similar extrapolation procedure as for the emulator: for z>2z > 2 we replace the boost with its linear theory prediction until z=12z = 12, and then set the boost to 1. For k≥5h Mpc−1k \ge5h\,\mathrm{Mpc}^{-1} we use the power-law form B(k)=Ak−p+1B(k) = Ak^{-p} + 1 with p≥0p \ge0, ensuring the boost decays smoothly towards unity on small scales. The boost is then multiplied by a Λ\LambdaCDM nonlinear matter power spectrum estimated with halofit (to ensure consistency with our model vector) to build the 3×23 \times2pt simulated datavector through Equation (15). We then estimate the datavector as outlined in § 5.

Figure 11. Top panel: the EFT-RAMSES boost, the Hi-COLA boost, and the emulator prediction at z=0z = 0 for fϕ=1f_{\phi} = 1. Bottom panel: relative error of Hi-COLA vs EFT-RAMSES, and Hi-COLA vs the emulator prediction.

Using the EFT-RAMSES 3×23 \times2pt vector as our simulated data, we carry out the same nautilus-sampled 3×23 \times2pt inference as in § 6.2.1 and § 6.2.2, using the Cubic Galileon emulator as the theoretical model. Figure 12 shows the posterior distribution for the parameters in this analysis. We can see that the cosmological parameters are significantly biased when using ℓmax⁡WL=1500\ell_{\max}^{\mathrm{WL}} = 1500. However, we note that fϕf_\phi itself remains consistent with its fiducial value in this case. This is likely an artefact of the prior boundary: the fiducial fϕ=1f_\phi= 1 sits at the upper edge of the emulator’s training range, so any bias towards larger fϕf_\phi is truncated by the prior rather than being absorbed into that parameter.

Figure 12. Posterior constraints from an LSST Y1 3×2pt forecast using mock data generated from an EFT–RAMSES N-body simulation near the fiducial cosmology (with fϕ=1f_\phi= 1). The theoretical model uses the Cubic Galileon emulator. No significant bias is observed when scale cuts are baryonic (as defined in § 5), but significant (>2σ> 2\sigma) bias is introduced for less conservative scale cuts methods (ℓmax⁡κκ=1500\ell_{\max}^{\kappa\kappa} = 1500).

We now want to understand if the bias is due to a breakdown of the COLA approximation, or a loss of emulator accuracy at small scales. The total emulator error relative to full N-body propagated to parameter constraints comes from two contributions: the emulator-vs-Hi-COLA error (which affects the results described in § 6.2.2), and the Hi-COLA-vsN-body (i.e. COLA approximation) error. We have shown in § 6.2.2 that when the mock data are generated from Hi-COLA at fϕ=0.99f_{\phi}=0.99 and analysed at a fixed clustering cut of ℓmax⁡=1500\ell_{\max}=1500, the emulator recovers the input cosmology and fϕf_{\phi} without significant bias for physical values of the cosmological parameters (Figure 10, in green). When the mock data are instead generated from the EFT-RAMSES N-body simulation and analysed at the same fixed ℓmax⁡=1500\ell_{\max}=1500, the recovered cosmology is biased at >2σ>2\sigma (Figure 12, in green). Taken together, these results show that the emulator cannot reproduce the behaviour of the N-body simulation at these scales because it inherits the Hi-COLA small-scale behaviour. These results therefore show that the dominant limitation is not the emulator interpolation but the COLA approximation on small scales.

When the same EFT-RAMSES mock data are analysed with the more conservative baryonic scale cuts (eq. (19), with kmax⁡=0.2 h Mpc−1k_{\max}=0.2\ h\ \mathrm{Mpc}^{-1}), the emulator recovers an unbiased cosmology during parameter inference (Figure 12, in orange). The scales at which the COLA approximation breaks down lie beyond those retained after baryonic cuts, so the emulator is fit for purpose for LSST Y1 under a realistic but conservative analysis. For a less conservative analysis, we advise to test the emulator against N-body simulations beforehand for a more realistic setup (including systematics like intrinsic alignment modelling and a more realistic photo-zz uncertainty), which we expect might be more forgiving than this low-parameter simplified forecast.

Overall, we show that the baryonic scale cuts required to control for the systematic of baryonic feedback also suppress the COLA-induced bias. Using less conservative scale cuts (keeping scales up to ℓ=1500\ell= 1500) exposes the COLA small-scale approximation error. The emulator is therefore sufficient for unbiased Cubic Galileon parameter inference with LSST Y1 3×2pt3 \times2\mathrm{pt} data if conservative scale cuts are applied.

Conclusions

We created an emulator over 5 parameters, Ωm0\Omega_{\mathrm{m0}}, nsn_{\mathrm{s}}, AsA_{\mathrm{s}}, hh and modified gravity parameter fϕf_{\phi}, for the boost of a Cubic Galileon model whose background follows the tracker solution, trained on Hi-COLA simulations. We validated this emulator against both Hi-COLA and full N-body (EFT-RAMSES) simulations, comparing the predicted boosts directly as well as forecasting constraints for a simulated LSST Y1 3×2pt3 \times2\mathrm{pt} analysis. The emulator and forecasting code are publicly available1011.

The emulator reliably reproduces Hi-COLA boosts, with highest accuracy near the centre of the parameter hypervolume (∼1%\sim1\% accuracy for k≲0.9h Mpc−1k \lesssim0.9h\,\mathrm{Mpc}^{-1}), and larger occasional degradation toward the boundaries, where the training data are sparser (with most errors remaining below 5%5\% for k<1.2h Mpc−1k < 1.2h\,\mathrm{Mpc}^{-1}). The emulator is accurate enough relative to Hi-COLA simulations to enable unbiased parameter estimation for LSST Y1 3×2pt3 \times2\mathrm{pt} data, for physically motivated parameter values, even under optimistic scale cuts (ℓκκ≤1500\ell^{\kappa\kappa} \le1500 and kGC,XC≤0.2h Mpc−1k^{\mathrm{GC,XC}} \le0.2h\,\mathrm{Mpc}^{-1}).

We also find that the boost deviates more substantially when compared against a high-resolution N-body (EFT–RAMSES) simulation, which we treat as the benchmark in our analysis. The Hi–COLA and emulator boosts differ from the EFT–RAMSES boost by ∼5%\sim5\% at k≈1.2 h Mpc−1k \approx1.2\,h\,\mathrm{Mpc}^{-1}. This shows that the COLA approximation rather than the emulator interpolation is the dominant source of error on small scales. When EFT–RAMSES mock data are analysed with the optimistic scale cuts above, the inferred cosmological parameters are biased at ≥2σ\ge2\sigma. When more conservative baryonic scale cuts are applied (eq. (19), with kmax⁡=0.2h Mpc−1k_{\max} = 0.2h\,\mathrm{Mpc}^{-1}), the emulator recovers an unbiased cosmology during parameter inference for 3×2pt LSST Y1-like data. The emulator can therefore produce unbiased posteriors for our Cubic Galileon model with LSST Y1-like 3×2pt data under conservative baryonic cuts. We stress that the accuracy demonstrated here applies to the specific LSST Y1 3×2pt setup and scale cuts we adopted. It should not be read as validation for Stage IV analyses. in general.

More specifically, this conclusion depends on the scale cut scheme: a nulling approach such as the BNT transform [116], which retains information from smaller angular scales, could expose inaccuracies masked by our conservative cuts. We also note that, since baryonic feedback in MG remains poorly understood, we used DMO power spectra modelling, with scale cuts which impose a strict removal of ℓ\ell bins affected by baryonic physics, rather than introducing additional parameters for baryonic physics modelling. Finally, our forecast is a simplified, low-dimensional setup that neglects systematics such as intrinsic alignments and adopts a simplified photo-zz treatment. We expect this last simplification to be conservative as far as the emulator is concerned: a more complete analysis with additional nuisance parameters would widen the posteriors and introduce degeneracies that could further mitigate the modelling biases. We recommend that these forecasts be repeated for the specific analysis setup of interest, with the full parameter set and the chosen scale cuts scheme, to confirm the emulator’s accuracy under those specifications.

This Cubic Galileon model is a good target for emulator development with Hi-COLA, as it provides an alternative theoretically motivated model to Hu-Sawicki f(R)f(R) and nDGP, and it modifies the cosmological background. It is, however, of limited interest as an alternative to dynamical dark energy: it has been ruled out observationally at fϕ=1f_{\phi}=1, and it does not cross the phantom divide. Future work will therefore aim to extend the emulated parameter space to Horndeski theories that cross the phantom divide, addressing more directly the MG modelling required to address the dynamical dark energy signals now emerging from the data. Additionally, in future emulation with Hi-COLA, we will improve on the current techniques by using a larger training set spanning wider parameter ranges (for example in Ωm0\Omega_{m0}), additionally provide LSST Y10 forecasts, and aim to introduce less conservative scale cuts for baryonic physics modelling (e.g. [127] or [116]), potentially introducing propagation of the EFT–RAMSES-to-Hi-COLA error into the covariance matrix.

More broadly, this work demonstrates that the combination of Hi-COLA and emulation is a viable and fast pipeline for under-explored Horndeski theories towards developing nonlinear modelling tools useful for Stage IV analyses, providing a route to systematically test a much wider class of gravity models than the simulation infrastructure has so far allowed.

Acknowledgements

This work has undergone internal review by the LSST Dark Energy Science Collaboration; the internal reviewers were Sankarshana Srinivasan and Georgios Valogiannis. CMAZ is supported by the Philip Robinson Cosmology PhD studentship. ASG is supported by the European Union (ERC StG, LSS BeyondAverage, 101075919). TB is supported by ERC Starting Grant SHADE (grant no. StG 945752) and by a Royal Society University Research Fellowship (grant no.URF\R\231006). CDL is supported by the Science and Technology Facilities Council (STFC) (grant No. UKRI1172). We acknowledge travel support provided by STFC for UK participation in the Rubin LSST, through grant ST/X001334/1.

Work at Argonne National Laboratory was supported by the U.S. Department of Energy, Office of High Energy Physics. Argonne, a U.S. Department of Energy Office of Science Laboratory, is operated by UChicago Argonne LLC under contract no. DE-AC02-06CH11357. Part of this work was done on Improv, a CPU system located at the Laboratory Computing Resource Center (LCRC) of Argonne National Laboratory.

The DESC acknowledges ongoing support from the Institut National de Physique Nucléaire et de Physique des Particules in France; the Science & Technology Facilities Council in the United Kingdom; and the Department of Energy and the LSST Discovery Alliance in the United States. DESC uses resources of the IN2P3 Computing Center (CC-IN2P3–Lyon/Villeurbanne - France) funded by the Centre National de la Recherche Scientifique; the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231; STFC DiRAC HPC Facilities, funded by UK BEIS National E-infrastructure capital grants; and the UK particle physics grid, supported by the GridPP Collaboration. This work was performed in part under DOE Contract DE-AC02-76SF00515.

For the purpose of open access, the author has applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version arising from this submission.

Author Contributions. CMAZ ran emulator training and validation simulation suites, performed the analysis tied to § 3, § 5, and § 6 (including writing and running the forecasting code) and wrote a significant portion of the text of this paper. ASG worked on the modelling of the on-tracker Cubic Galileons, assisted with the model selection for creating the emulator training and validation sets, and assisted with the running of Hi-COLA. ASG also assisted with the measurements of 2-point matter power spectra, and worked on the analysis and comparison of results with the EFT-RAMSES code. ASG contributed to the writing and revising of the manuscript. NR developed the emulator, played a major role in designing the simulation suite and the forecasting analysis, wrote § 4, and contributed to writing and revising the rest of the manuscript. TB worked on the modelling of the on-tracker Cubic Galileons, advised on running Hi-COLA and on the simulation specifications, provided scientific and technical guidance particularly on the work in § 2, § 3, and § 6, and contributed to writing and revising the manuscript. BF advised on running Hi-COLA and on the simulation specifications, and provided guidance on the emulator design, on the COLA code, on the modeling of the on-tracker Cubic Galileons, and on the comparison with N-body simulations. DJ contributed to the early development of the emulation infrastructure and to the project’s inception. KH helped conceive and design the project, assisted with planning the preliminary simulation suite and developing the initial version of the emulator, and contributed to revising the manuscript. CDL provided scientific and technical guidance particularly with regards to the work of § 3, § 5, and § 6, and contributed to writing and revising the paper text. BL provided access to the EFT-RAMSES simulation, gave scientific and technical guidance, and contributed to writing and revising the manuscript. NŠ developed tools for redshift distributions and tomographic binning; DESC Builder.

References

  1. [1]D. M. Scolnic et al. The Complete Light-curve Sample of Spectroscopically Confirmed SNe Ia from Pan-STARRS1 and Cosmological Constraints from the Combined Pantheon Sample. The Astrophysical Journal, 859(2):101, May 2018. 1DOI
  2. [2]N. et al. Aghanim. Planck 2018 results: VI. Cosmological parameters. Astronomy & Astrophysics, 641:A6, September 2020. 1, 11DOI
  3. [3]Shadab Alam et al. Completed SDSS-IV extended Baryon Oscillation Spectroscopic Survey: Cosmological implications from two decades of spectroscopic surveys at the Apache Point Observatory. Phys. Rev. D, 103:083533, Apr 2021. 1DOI
  4. [4]C. Garcia-Quintero et al. Cosmological implications of DESI DR2 BAO measurements in light of the latest ACT DR6 CMB data. Phys. Rev. D, 112:083529, Oct 2025. 1
  5. [5]DES Collaboration et al. Constraints on dynamical dark energy from multiple probes in the full dark energy survey, May 2026. Submitted to PRL. 1
  6. [6]Timothy Clifton, Pedro G. Ferreira, Antonio Padilla, and Constantinos Skordis. Modified gravity and cosmology. Physics Reports, 513(1–3):1–189, March 2012. 1
  7. [7]Kazuya Koyama. Cosmological tests of modified gravity. Reports on Progress in Physics, 79(4):046902, March 2016.DOI
  8. [8]Mustapha Ishak. Testing general relativity in cosmology. Living Reviews in Relativity, 22(1), December 2018. 1
  9. [9]Anton Chudaykin and Martin Kunz. Modified gravity interpretation of the evolving dark energy in light of DESI data, 2024. 1arxiv.org/abs/2407.02558
  10. [10]Zhibang Yao, Gen Ye, and Alessandra Silvestri. A general model for dark energy crossing the phantom divide, 2025.
  11. [11]William J. Wolf, Pedro G. Ferreira, and Carlos García-García. Matching current observational constraints with nonminimally coupled dark energy. Physical Review D, 111(4), February 2025.DOI
  12. [12]William J. Wolf, Carlos García-García, Theodore Anton, and Pedro G. Ferreira. Assessing cosmological evidence for nonminimal coupling. Physical Review Letters, 135(8), August 2025.DOI
  13. [13]Gen Ye, Anton Chudaykin, Camille Bonvin, and Martin Kunz. Late-time reconstruction of non-minimally coupled gravity with a smoothness prior. arXiv e-prints, May 2026. 2605.12415.arxiv.org/abs/2605.12415
  14. [14]William J. Wolf, Pedro G. Ferreira, and Carlos García-García. Cosmological constraints on galileon dark energy with broken shift symmetry. Physical Review D, 113(2), January 2026.DOI
  15. [15]Eric V. Linder. Cosmology after Phantom Crossing by Horndeski Gravity. arXiv e-prints, December 2025. arXiv:2512.03139.arxiv.org/abs/2512.03139
  16. [16]Rodrigo Calderon and Eric V. Linder. Charging Across the Phantom Divide with Modified Gravity. arXiv e-prints, June 2026. Submitted to JCAP.arxiv.org/abs/2605.26259
  17. [17]Shinji Tsujikawa. Crossing the phantom divide in scalar-tensor and vector-tensor theories. Physical Review D, 113(4), February 2026.DOI
  18. [18]Gen Ye, Matteo Martinelli, Bin Hu, and Alessandra Silvestri. Hints of Nonminimally Coupled Gravity in DESI 2024 Baryon Acoustic Oscillation Measurements. Physical Review Letters, 134(18), May 2025. 1DOI
  19. [19]Emilio Bellini and Ignacy Sawicki. Maximal freedom at minimum cost: linear large-scale structure in general modifications of gravity. Journal of Cosmology and Astroparticle Physics, 2014(07):050–050, 2014. 2DOI
  20. [20]Krishna Naidoo, James Hallam, Tessa Baker, and Sergi Sirera. Constraints on Horndeski Gravity with Phantom Crossing. arXiv e-prints, June 2026. 2606.20794. Submitted to MNRAS. 2, 3, 5, 6
  21. [21]Wayne Hu and Ignacy Sawicki. Models of f(R) cosmic acceleration that evade solar system tests. Phys. Rev. D, 76:064004, Sep 2007. 2DOI
  22. [22]Gia Dvali, Gregory Gabadadze, and Massimo Porrati. 4d gravity on a brane in 5d minkowski space. Physics Letters B, 485(1):208–214, 2000. 2
  23. [23]Joachim Harnois-Déraps, Cesar Hernandez-Aguayo, Carolina Cuesta-Lazaro, Christian Arnold, Baojiu Li, Christopher T. Davies, and Yan-Chuan Cai. MGLenS: Modified gravity weak lensing simulations for emulation-based cosmological inference, 2022. 2
  24. [24]D. Sanchez-Cid et al. Dark Energy Survey Year 6 results: Weak lensing and galaxy clustering cosmological analysis framework. Phys. Rev. D, 114(6):063506, 2026. 2, 10
  25. [25]Catherine Heymans et al. Kids-1000 cosmology: Multi-probe weak gravitational lensing and spectroscopic galaxy clustering constraints. Astronomy & Astrophysics, 646:A140, February 2021. 10
  26. [26]Hironao Miyatake et al. Hyper Suprime-Cam Year 3 results: Cosmology from galaxy clustering and weak lensing with HSC and SDSS using the emulator based halo model. PRD, 108(12):123517, December 2023. 2arxiv.org/abs/2304.00704
  27. [27]Tom Charnock, Guilhem Lavaux, and Benjamin D. Wandelt. Automatic physical inference with information maximizing neural networks. Physical Review D, 97(8), April 2018. 2DOI
  28. [28]Alex Gough and Cora Uhlemann. One-point statistics matter in extended cosmologies. Universe, 8(1):55, January 2022.DOI
  29. [29]Joachim Harnois-Déraps, Nicolas Martinet, Tiago Castro, Klaus Dolag, Benjamin Giblin, Catherine Heymans, Hendrik Hildebrandt, and Qianli Xia. Cosmic shear cosmology beyond two-point statistics: a combined peak count and correlation function analysis of DES-Y1. Monthly Notices of the Royal Astronomical Society, 506(2):1623–1650, 09 2021.DOI
  30. [30]Nicolas Martinet et al. KiDS-450: cosmological constraints from weak-lensing peak statistics – ii: Inference from shear peaks using n-body simulations. Monthly Notices of the Royal Astronomical Society, 474(1):712–730, 02 2018.DOI
  31. [31]Nisha Grewal, Joe Zuntz, Tilman Tröster, and Alexandra Amon. Minkowski functionals in joint galaxy clustering & weak lensing analyses. The Open Journal of Astrophysics, 5(1), August 2022.DOI
  32. [32]Georgios Valogiannis and Cora Dvorkin. Towards an optimal estimation of cosmological parameters with the wavelet scattering transform. Physical Review D, 105(10), May 2022.DOI
  33. [33]ChangHoon Hahn and Francisco Villaescusa-Navarro. Constraining Mν with the bispectrum. Part II. The information content of the galaxy bispectrum monopole. Journal of Cosmology and Astroparticle Physics, 2021(04):029, April 2021. 2DOI
  34. [34]Željko Ivezić et al. Lsst: From science drivers to reference design and anticipated data products. The Astrophysical Journal, 873(2):111, March 2019. 2DOI
  35. [35]Euclid Collaboration. Euclid: I. overview of the euclid mission. Astronomy & Astrophysics, 697:A1, April 2025. 2
  36. [36]Mustapha Ishak, Tessa Baker, Philip Bull, Eske M. Pedersen, Jonathan Blazek, Pedro G. Ferreira, C. Danielle Leonard, Weikang Lin, Eric Linder, Kris Pardo, and Georgios Valogiannis. Modified Gravity and Dark Energy models Beyond w(z)CDM Testable by LSST. arXiv e-prints, September 2019. 1905.09687. 2arxiv.org/abs/1905.09687
  37. [37]C. D. Leonard, S. Alam, R. Mandelbaum, M. M. Rau, S. Singh, and C. M. A. and Zanoletti. Toward testing gravity with lsst using E_G. Physical Review D, 113(4), February 2026.DOI
  38. [38]Santiago Casas, Martin Kunz, Matteo Martinelli, and Valeria Pettorino. Linear and non-linear modified gravity forecasts with future surveys. Physics of the Dark Universe, 18:73–104, 2017.arxiv.org/abs/1703.01271
  39. [39]Euclid Collaboration et al. Euclid: Constraining linearly scale-independent modifications of gravity with the spectroscopic and photometric primary probes. Astronomy & Astrophysics, 690:A133, October 2024.
  40. [40]Euclid Collaboration et al. Euclid preparation - CIII. Constraining parameterised models of modifications of gravity with the spectroscopic and photometric primary probes. Astronomy & Astrophysics, 713:A214, 2026. 2
  41. [41]Jonathan Gordon, Bernardo F. de Aguiar, João Rebouças, Guilherme Brando, Felipe Falciano, Vivian Miranda, Kazuya Koyama, and Hans A. Winther. Modeling nonlinear scales with the comoving lagrangian acceleration method: Preparing for lsst yl. Phys. Rev. D, 110:083529, Oct 2024. 2
  42. [42]Bill S. Wright, Ashim Sen Gupta, Tessa Baker, Georgios Valogiannis, and Bartolomeo Fiorini. Hi-cola: fast, approximate simulations of structure formation in horndeski gravity. Journal of Cosmology and Astroparticle Physics, 2023(03):040, March 2023. 2, 4, 5, 6, 8, 13, 22DOI
  43. [43]B Bose, M Tsedrik, J Kennedy, L Lombriser, A Pourtsidou, and A Taylor. Fast and accurate predictions of the non-linear matter power spectrum for general models of dark energy and modified gravity. Monthly Notices of the Royal Astronomical Society, 519(3):4780–4800, 03 2023. 2DOI
  44. [44]Daniela Saadeh, Kazuya Koyama, and Xan Morice-Atkinson. A field-level reaction for screened modified gravity. Mon. Not. Roy. Astron. Soc., 544(4):3634–3646, 2025. 2arxiv.org/abs/2506.18876
  45. [45]Sankarshana Srinivasan, Shreya Prabhu, Kai Lehman, Ajiv Krishnan V., and Jochen Weller. Cosmological gravity on all scales. Part V. MCMC forecasts combining large scale structure and CMB lensing for binned phenomenological modified gravity. JCAP, 07:058, 2026. 2, 14DOI
  46. [46]Bartolomeo Fiorini, Kazuya Koyama, and Tessa Baker. Fast production of cosmological emulators in modified gravity: the matter power spectrum. JCAP, 12:045, 2023. 2, 18
  47. [47]Alberto Nicolis, Riccardo Rattazzi, and Enrico Trincherini. Galileon as a local modification of gravity. Physical Review D, 79(6), March 2009. 3DOI
  48. [48]C. Defayet, G. Esposito-Farèse, and A. Vikman. Covariant galileon. Physical Review D, 79(8), April 2009.DOI
  49. [49]C. Defayet, S. Deser, and G. Esposito-Farèse. Generalized galileons: All scalar models whose curved background extensions maintain second-order field equations and stress tensors. Physical Review D, 80(6), September 2009. 3DOI
  50. [50]Alexandre Barreira, Baojiu Li, Carlton M. Baugh, and Silvia Pascoli. The observational status of Galileon gravity after Planck. JCAP, 2014(8):059–059, August 2014. 3
  51. [51]Nathaniel Woodcock, Sownak Bose, Yunhao Gao, and Baojiu Li. EFT-RAMSES: a code to simulate the effective field theory of dark energy. MNRAS, submitted, 2026. 3, 22
  52. [52]Baojiu Li, Gong-Bo Zhao, Romain Teyssier, and Kazuya Koyama. Ecosmog: an efficient code for simulating modified gravity. Journal of Cosmology and Astroparticle Physics, 2012(01):051–051, January 2012. 3DOI
  53. [53]Baojiu Li, Gong-Bo Zhao, and Kazuya Koyama. Exploring Vainshtein mechanism on adaptively refined meshes. JCAP, 05:023, 2013. 3DOI
  54. [54]Dina Traykova, Emilio Bellini, Pedro G. Ferreira, Carlos García-García, Johannes Noller, and Miguel Zumalacárregui. Theoretical priors in scalar-tensor cosmologies: Shift-symmetric horndeski models. Physical Review D, 104(8), October 2021. 3DOI
  55. [55]Alexandre Barreira, Baojiu Li, Carlton M Baugh, and Silvia Pascoli. Spherical collapse in galileon gravity: fifth force solutions, halo mass function and halo bias. Journal of Cosmology and Astroparticle Physics, 2013(11):056–056, November 2013. 3DOI
  56. [56]Simone Peirone, Noemi Frusciante, Bin Hu, Marco Raveri, and Alessandra Silvestri. Do current cosmological observations rule out all covariant galileons? Physical Review D, 97(6), March 2018. 3DOI
  57. [57]Noemi Frusciante, Simone Peirone, Luís Atayde, and Antonio De Felice. Phenomenology of the generalized cubic covariant galileon model and cosmological bounds. Physical Review D, 101(6), March 2020. 3, 13DOI
  58. [58]Mark Wyman, Elise Jennings, and Marcos Lima. Simulations of galileon modified gravity: Clustering statistics in real and redshift space. Physical Review D, 88(8), October 2013. 3
  59. [59]Alexandre Barreira, Baojiu Li, Wojciech A Hellwing, Carlton M Baugh, and Silvia Pascoli. Nonlinear structure formation in the cubic galileon gravity model. Journal of Cosmology and Astroparticle Physics, 2013(10):027–027, October 2013.DOI
  60. [60]Jiajun Zhang, Bikash R. Dinda, Md. Wali Hossain, Anjan A. Sen, and Wentao Luo. Study of cubic galileon gravity using n-body simulations. Physical Review D, 102(4), August 2020. 3DOI
  61. [61]Antonio De Felice and Shinji Tsujikawa. Cosmology of a covariant galileon field. Physical Review Letters, 105(11), 2010. 4
  62. [62]Antonio De Felice and Shinji Tsujikawa. Generalized galileon cosmology. Physical Review D, 84(12), December 2011.DOI
  63. [63]Savvas Nesseris, Antonio De Felice, and Shinji Tsujikawa. Observational constraints on galileon cosmology. Physical Review D, 82(12), December 2010. 4
  64. [64]Antonio De Felice and Shinji Tsujikawa. Conditions for the cosmological viability of the most general scalar-tensor theories and their applications to extended galileon dark energy models. Journal of Cosmology and Astroparticle Physics, 2012(02):007–007, February 2012. 4DOI
  65. [65]Bhuvnesh Jain and Pengjie Zhang. Observational tests of modified gravity. Physical Review D, 78(6), 2008. 4
  66. [66]Gong-Bo Zhao, Levon Pogosian, Alessandra Silvestri, and Joel Zylberberg. Searching for modified growth patterns with tomographic surveys. Physical Review D, 79(8), April 2009.DOI
  67. [67]Levon Pogosian and Alessandra Silvestri. What can cosmology tell us about gravity? constraining horndeski gravity with Σ and μ. Physical Review D, 94(10), November 2016.DOI
  68. [68]Tessa Baker and Philip Bull. Observational signatures of modified gravity on ultra-large scales. The Astrophysical Journal, 811(2):116, 2015. 4DOI
  69. [69]C. Leloup, V. Ruhlmann-Kleider, J. Neveu, and A. de Mattia. Observational status of the galileon model general solution from cosmological data and gravitational waves. Journal of Cosmology and Astroparticle Physics, 2019(05):011–011, May 2019. 5DOI
  70. [70]J. Neveu, V. Ruhlmann-Kleider, P. Astier, M. Besançon, A. Conley, J. Guy, A. Möller, N. Palanque-Delabrouille, and E. Babichev. First experimental constraints on the disformally coupled galileon model. Astronomy and Astrophysics, 569:A90, 2014.DOI
  71. [71]J. Neveu, V. Ruhlmann-Kleider, A. Conley, N. Palanque-Delabrouille, P. Astier, J. Guy, and E. Babichev. Experimental constraints on the uncoupled galileon model from snls3 data and other cosmological probes. Astronomy and Astrophysics, 555:A53, 2013. 5DOI
  72. [72]A. J. Mead, C. Heymans, L. Lombriser, J. A. Peacock, O. I. Steele, and H. A. Winther. Accurate halo-model matter power spectra with dark energy, massive neutrinos and modified gravitational forces. Monthly Notices of the Royal Astronomical Society, 459(2):1468–1488, March 2016. 5, 14DOI
  73. [73]James Hallam, Krishna Naidoo, Sergi Sirera, and Tessa Baker. Rolling Galileons: Evolving Braiding Strength for Viable Dark Energy. July 2026. 5, 6arxiv.org/abs/2607.16395
  74. [74]Svetlin Tassev, Matias Zaldarriaga, and Daniel Eisenstein. Solving Large Scale Structure in Ten Easy Steps with COLA. JCAP, 06:036, 2013. 5DOI
  75. [75]Georgios Valogiannis and Rachel Bean. Efficient simulations of large-scale structure in modified gravity cosmologies with comoving lagrangian acceleration. Physical Review D, 95(10), May 2017. 5, 6DOI
  76. [76]Rampei Kimura, Tsutomu Kobayashi, and Kazuhiro Yamamoto. Vainshtein screening in a cosmological background in the most general second-order scalar-tensor theory. Physical Review D, 85(2):024023, January 2012. 5DOI
  77. [77]Lucas Lombriser. A parametrisation of modified gravity on nonlinear cosmological scales. Journal of Cosmology and Astroparticle Physics, 2016(11):039–039, November 2016. 5DOI
  78. [78]Sergi Sirera, Tessa Baker, James Hallam, and Krishna Naidoo. Master equation for screening in luminal horndeski gravity. Phys. Rev. D, 114:064028, Sep 2026. 6DOI
  79. [79]Hans A. Winther and Pedro G. Ferreira. A Fast Route to Non-Linear Clustering Statistics in Modified Gravity Theories. Physical Review D, 91(12):123507, June 2015. 6
  80. [80]Hans A. Winther, Kazuya Koyama, Marc Manera, Bill S. Wright, and Gong-Bo Zhao. COLA with scale-dependent growth: Applications to screened modified gravity models. JCAP, 08:006, 2017. 6DOI
  81. [81]Guilherme Brando, Bartolomeo Fiorini, Kazuya Koyama, and Hans A. Winther. Enabling matter power spectrum emulation in beyond-ΛCDM cosmologies with COLA. JCAP, 09:051, 2022. 6DOI
  82. [82]Farbod Hassani and Lucas Lombriser. $N$-body simulations for parametrised modified gravity. Monthly Notices of the Royal Astronomical Society, 497(2):1885–1894, September 2020. 6
  83. [83]Ashim Sen Gupta, Bartolomeo Fiorini, and Tessa Baker. K-mouflage at high k: Extending the reach of Hi-COLA. JCAP, 11:052, 2024. 6DOI
  84. [84]B Bose, A Sen Gupta, B Fiorini, G Brando, F Hassani, T Baker, L Lombriser, B Li, C Ruan, C Hernández-Aguayo, L Atayde, and N Frusciante. Matter power spectra in modified gravity: A comparative study of approximations and N-body simulations. Monthly Notices of the Royal Astronomical Society, 536(1):664–683, November 2024. 6arxiv.org/abs/2406.13667
  85. [85]Katrin Heitmann, David Higdon, Martin White, Salman Habib, Brian J. Williams, Earl Lawrence, and Christian Wagner. The coyote universe. ii. cosmological models and precision emulation of the nonlinear matter power spectrum. The Astrophysical Journal, 705(1):156–174, October 2009. 6, 7, 9DOI
  86. [86]Iñigo Sáez-Casares, Yann Rasera, and Baojiu Li. The e-mantis emulator: fast predictions of the non-linear matter power spectrum in f(r)cdm cosmology. Monthly Notices of the Royal Astronomical Society, 527(3):7242–7262, November 2023. 7DOI
  87. [87]Katrin Heitmann, Derek Bingham, Earl Lawrence, Steven Bergner, Salman Habib, David Higdon, Adrian Pope, Rahul Biswas, Hal Finkel, Nicholas Frontiere, and Suman Bhattacharya. The mira–titan universe: Precision predictions for dark energy surveys. The Astrophysical Journal, 820(2):108, March 2016. 7DOI
  88. [88]Harry Desmond and Pedro G. Ferreira. Galaxy morphology rules out astrophysically relevant hu-sawicki f(r) gravity. Phys. Rev. D, 102:104060, Nov 2020. 7DOI
  89. [89]Ricardo G Landim, Harry Desmond, Kazuya Koyama, and Samantha Penny. Testing screened modified gravity with sdss-iv-manga. Monthly Notices of the Royal Astronomical Society, 534(1):349–360, 10 2024. 7DOI
  90. [90]Christian Arnold, Baojiu Li, Benjamin Giblin, Joachim Harnois-Déraps, and Yan-Chuan Cai. FORGE: the f(r)-gravity cosmic emulator project – i. introduction and matter power spectrum emulator. Monthly Notices of the Royal Astronomical Society, 515(3):4161–4175, August 2022. 7DOI
  91. [91]Nesar Ramachandra, Georgios Valogiannis, Mustapha Ishak, and Katrin Heitmann. Matter power spectrum emulator for f(r) modified gravity cosmologies. Physical Review D, 103(12), June 2021. 7, 8, 18
  92. [92]J. and Adamek et al. Euclid preparation: LXII. Simulations and non-linearities beyond lambda cold dark matter. I. numerical methods and validation. Astronomy & Astrophysics, 695:A230, March 2025.
  93. [93]C. M. A. Zanoletti and C. D. Leonard. Principal components for model-agnostic modified gravity with 3x2pt data. Physical Review D, 112(6), 2025.DOI
  94. [94]Gábor Rácz, Alina Kiessling, István Csabai, and István Szapudi. Complementary cosmological simulations. Astronomy & Astrophysics, 672:A59, March 2023.DOI
  95. [95]Andrew Pontzen, An že Slosar, Nina Roth, and Hiranya V. Peiris. Inverted initial conditions: Exploring the growth of cosmic structure and voids. Phys. Rev. D, 93:103519, May 2016.DOI
  96. [96]Renate Mauland, Hans A. Winther, and Cheng-Zong Ruan. Sesame: A power spectrum emulator pipeline for beyond-lambdacdm models. Astronomy & Astrophysics, 685:A156, May 2024.DOI
  97. [97]Abraham Savitzky and Marcel JE Golay. Smoothing and differentiation of data by simplified least squares procedures. Analytical chemistry, 36(8):1627–1639, 1964.
  98. [98]Earl Lawrence, Katrin Heitmann, Martin White, David Higdon, Christian Wagner, Salman Habib, and Brian Williams. The Coyote Universe. III. Simulation Suite and Precision Emulator for the Nonlinear Matter Power Spectrum. ApJ, 713(2):1322–1331, April 2010.DOI
  99. [99]Katrin Heitmann, Earl Lawrence, Juliana Kwan, Salman Habib, and David Higdon. The Coyote Universe Extended: Precision Emulation of the Matter Power Spectrum. ApJ, 780(1):111, January 2014.
  100. [100]Dave Higdon, James Gattiker, Brian Williams, and Maria Rightley. Computer model calibration using high-dimensional output. Journal of the American Statistical Association, 103(482):570–583, 2008.
  101. [101]James Gattiker, Natalie Klein, Grant Hutchings, and Earl Lawrence. lanl/sepia: v1.1, September 2020.
  102. [102]Alessandra Silvestri, Levon Pogosian, and Roman V. Buniy. Practical approach to cosmological perturbations in modified gravity. Physical Review D, 87(10), May 2013.
  103. [103]Daniel B. Thomas. Cosmological gravity on all scales: Simple equations, required conditions, and a framework for modified gravity. Physical Review D, 101(12), 2020.DOI
  104. [104]the LSST Dark Energy Science Collaboration. The LSST Dark Energy Science Collaboration (DESC) Science Requirements Document, 2021.
  105. [105]T. M. C. Abbott et al. Dark Energy Survey Year 3 Results: Constraints on extensions to ΛCDM with weak lensing and galaxy clustering. PRD, 107(8):083504, April 2023.
  106. [106]Alex Hall. Cosmology from weak lensing alone and implications for the hubble tension. Monthly Notices of the Royal Astronomical Society, 505(4):4935–4955, May 2021.DOI
  107. [107]Maria Tsedrik, Benjamin Bose, Pedro Carrilho, Alkistis Pourtsidou, Sefa Pamuk, Santiago Casas, and Julien Lesgourgues. Stage-IV cosmic shear with Modified Gravity and model-independent screening. JCAP, 10:099, 2024.arxiv.org/abs/2404.11508
  108. [108]Nora Elisa Chisari et al. Core cosmology library: Precision cosmological predictions for LSST. The Astrophysical Journal Supplement Series, 242(1):2, may 2019.DOI
  109. [109]Ryuichi Takahashi, Masanori Sato, Takahiro Nishimichi, Atsushi Taruya, and Masamune Oguri. Revising the halofit model for the nonlinear matter power spectrum. The Astrophysical Journal, 761(2):152, December 2012.DOI
  110. [110]D. Nelson Limber. The Analysis of Counts of the Extragalactic Nebulae in Terms of a Fluctuating Density Field. ApJ, 117:134, January 1953.DOI
  111. [111]Simone Peirone, Giampaolo Benevento, Noemi Frusciante, and Shinji Tsujikawa. Cosmological data favor galileon ghost condensate over λcdm. Physical Review D, 100(6), September 2019.DOI
  112. [112]Johannes Noller, Luca Santoni, Enrico Trincherini, and Leonardo G. Trombetta. Scalar-tensor cosmologies without screening. Journal of Cosmology and Astroparticle Physics, 2021(01):045–045, January 2021.
  113. [113]Michael Schneider, Lloyd Knox, Hu Zhan, and Andrew Connolly. Using galaxy two-point correlation functions to determine the redshift distributions of galaxies binned by photometric redshift. The Astrophysical Journal, 651(1):14–23, November 2006.DOI
  114. [114]Nora Elisa Chisari, Alexander J. Mead, Shahab Joudaki, Pedro G. Ferreira, Aurel Schneider, Joseph Mohr, Tilman Tröster, David Alonso, Ian G. McCarthy, Sergio Martin-Alvarez, Julien Devriendt, Adrianne Slyz, and Marcel P. van Daalen. Modelling baryonic feedback for survey cosmology. The Open Journal of Astrophysics, 2(1), 2019.
  115. [115]Christian Arnold, Matteo Leo, and Baojiu Li. Realistic simulations of galaxy formation in f(r) modified gravity. Nature Astronomy, 3(10):945–954, 2019.
  116. [116]Peter L. Taylor, Francis Bernardeeau, and Eric Huff. k-cut cosmic shear: Optimally removing sensitivity to baryonic and nonlinear physics with an application to the dark energy survey year 1 shear data. Physical Review D, 103(4), February 2021.
  117. [117]Sankarshana Srinivasan, Daniel B. Thomas, and Peter L. Taylor. Cosmological gravity on all scales: part iv. 3× 2 pt fisher forecasts for pixelised phenomenological modified gravity. Journal of Cosmology and Astroparticle Physics, 2025(02):071, February 2025.DOI
  118. [118]Ottavia Truttero, Maria Tsedrik, Joe Zuntz, Alkistis Pourtsidou, Nikolina Šarčević, Christos Georgiou, and the LSST Dark Energy Science Collaboration. Balancing bias, baryons, and scale cuts in lsst 3x2pt analysis. arXiv e-prints, 2026. 2606.10679.arxiv.org/abs/2606.10679
  119. [119]Sebastian Grandis, Giovanni Aricò, Aurel Schneider, and Laila Linke. Determining the baryon impact on the matter power spectrum with galaxy clusters. Monthly Notices of the Royal Astronomical Society, 528(3):4379–4392, 03 2024.
  120. [120]C et al. Doux. Dark energy survey year 3 results: cosmological constraints from the analysis of cosmic shear in harmonic space. Monthly Notices of the Royal Astronomical Society, 515(2):1942–1972, July 2022.DOI
  121. [121]Marcel P. van Daalen, Joop Schaye, C. M. Booth, and Claudio Dalla Vecchia. The effects of galaxy formation on the matter power spectrum: a challenge for precision cosmology: Galaxy formation and the matter power spectrum. Monthly Notices of the Royal Astronomical Society, 415(4):3649–3665, July 2011.
  122. [122]C. Abbott, T. M et al. Dark energy survey year 3 results: Constraints on extensions to λcdm with weak lensing and galaxy clustering. Physical Review D, 107(8), April 2023.
  123. [123]Elisabeth Krause and Tim Eifler. cosmolike – cosmological likelihood analyses for photometric galaxy surveys. Monthly Notices of the Royal Astronomical Society, 470(2):2100–2112, May 2017.DOI
  124. [124]Darsh Kodwani, David Alonso, and Pedro Ferreira. The effect on cosmological parameter estimation of a parameter dependent covariance matrix. The Open Journal of Astrophysics, 2(1), March 2019.DOI
  125. [125]K. and Koyama and others for the Euclid collaboration. Euclid preparation: Lxxi. simulations and nonlinearities beyond λcdm. 3. constraints on f(r) models from the photometric primary probes. Astronomy & Astrophysics, 698:A233, 2025.DOI
  126. [126]Johannes U Lange. nautilus: boosting Bayesian importance nested sampling with deep learning. Monthly Notices of the Royal Astronomical Society, 525(2):3181–3194, 08 2023.arxiv.org/abs/2306.16923
  127. [127]Hung-Jin Huang, Tim Eifler, Rachel Mandelbaum, and Scott Dod elson. Modelling baryonic physics in future weak lensing surveys. Monthly Notices of the Royal Astronomical Society, 488(2):1652–1678, 2019. 28DOI

Paper details

vlnc:W_c8h8yea3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents