Mujic Λ \Lambda : Initial Condition Reconstruction on Realistic Redshift Surveys with Augmented Optimization
Chenze Dong, Benjamin Horowitz, Adrian E. Bayer, Khee-Gan Lee · 2026
Read Mujic Λ \Lambda : Initial Condition Reconstruction on Realistic Redshift Surveys with Augmented Optimization by Chenze Dong, Benjamin Horowitz, Adrian E. Bayer, Khee-Gan Lee (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-26823.txt.
Introduction
Dong and three colleagues introduce MujicΛ, a framework for rebuilding the universe's starting conditions from galaxy redshift surveys. It strengthens the L-BFGS optimizer with a projection step and rank-order matching, forcing the rebuilt starting field to stay Gaussian and robust to survey gaps. Tested on a mock catalog from the Millennium simulation, it recovers the true density field down to the model's resolution and broadly recovers the cosmic web's structure. The method also speeds field-level inference for coming large surveys.
Catalog record
| CREATOR | Chenze Dong, Benjamin Horowitz, Adrian E. Bayer, Khee-Gan Lee |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 63417 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.26823
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Mujic Λ \Lambda : Initial Condition Reconstruction on Realistic Redshift Surveys with Augmented Optimization Chenze Dong, Benjamin Horowitz, Adrian E. Bayer, Khee-Gan Lee Abstract In this paper, we introduce Mujic Λ \Lambda ( M apping the U niverse with J ax-based I nitial C ondition Reconstr 𝚲 \mathbf{\Lambda} ction), an optimization-based framework for reconstructing initial conditions from realistic galaxy spectroscopic redshift surveys. Unlike standard optimization-based approaches, Mujic Λ \Lambda augments the L-BFGS algorithm with a projection operator and rank-order matching to enforce Gaussianity of the initial conditions and substantially improve robustness to incomplete survey geometries. We validate Mujic Λ \Lambda on a mock lightcone catalog derived from semi-analytic models applied to the Millennium simulation. We construct a differentiable forward model that incorporates a fast particle-mesh simulation at megaparsec resolution and a comprehensive treatment of observational effects and survey incompleteness. Mujic Λ \Lambda reaches good agreement with the true density field down to the scale of the forward model, while maintaining consistency with the Gaussian prior through the projection step. It also broadly recovers the cosmic web classification, underscoring its value for deciphering environmental information in galaxy evolution studies. Beyond its key role in next-generation constrained simulations, the methodology offers a practical way to generate initial guesses and speed up field-level inference, especially for upcoming large-scale galaxy surveys. I Introduction Large-scale structure forms a “cosmic web” organized into knots, filaments, sheets, and voids. Since being observed ( 55 ; 37 ; 24 ) in early galaxy redshift surveys and analyzed in numerical simulations ( 16 ) , the cosmic web has become a key interface between cosmology and galaxy evolution. The cosmic web is widely studied from both perspectives of observation ( 112 ; 32 ; 62 ; 76 ; 23 ; 6 ) and simulation ( 19 ; 72 ; 78 ; 93 ; 106 ; 89 ) , revealing its importance at the crossroads of cosmology ( 13 ; 15 ; 11 ; 104 ; 101 ; 105 ) and galaxy evolution ( 21 ; 78 ; 35 ; 63 ) . Despite the great importance of the cosmic web for these studies, unambiguously identifying the cosmic web remains a challenging task. In practice, extracting the underlying cosmic web from observables is non-trivial due to its requirement for full 3D information, in contrast to realistic observational data that might be limited by systematics. Many cosmic web finding algorithms (e.g., 34 ; 41 ; 20 ) rely on the density field from simulations, which greatly restricts their application to realistic astrophysical data. Even for those methods that allow galaxies as biased tracers ( 95 ; 103 ) , the incompleteness and selection function of the galaxy survey remain challenging to incorporate into the algorithm. One must therefore take into account observational systematics to obtain a robust cosmic web. Density reconstruction via forward-modeling partly addresses this issue since it explicitly connects the density field to observed tracers and incorporates survey selection functions. Bayesian sampling frameworks, such as Elucid ( 107 ; 108 ; 109 ) , ARGO ( 4 ) , and BORG ( 52 ) , have demonstrated significant success in reconstructing the density field. 91 proposed an alternative to Bayesian inference, using MAP (maximum a posteriori) for density reconstruction and a Laplacian approximation at the MAP solution for cosmological parameter estimation (see also 43 ). Recent progress in Lyman- \alpha tomography ( 66 ; 67 ; 69 ; 68 ; 83 ; 47 ) also enables the z>2 Lyman- \alpha forest absorption to act as a tracer of the large-scale structure free from the galaxy selection function since foreground absorption is typically separated by hundreds of Megaparsecs from the background sources and thus has little correlation with the latters’ position. The Lyman- \alpha forest tracers have motivated the optimization-based density reconstruction code TARDIS ( 48 ; 50 ) following the MAP approach by 91 , which recovers the cosmic web with good accuracy. Optimization-based density reconstruction has also been shown to robustly improve BAO reconstruction in galaxy surveys ( 9 ) . These developments reflect a broader trend of cosmological analysis toward the field-level, in which inference operates directly on the full data vector rather than on summary statistics. However, it is difficult to quantify the reliability of the reconstructions and their inferred cosmic web. Structure formation models typically rely on strong simplifying assumptions when describing the tracer distribution to keep the computation efficient. At the same time, the incompleteness of input tracers renders the reconstruction intrinsically under-constrained. In Bayesian sampling-based approaches, such as MCMC ( 53 ; 52 ; 10 ; 84 ; 1 ; 92 ) , these difficulties are mitigated by drawing many samples, providing built-in uncertainty quantification at the expense of incurring larger computational costs, while relying on informative cosmological priors to regularize the inference. However, sampling-based methods are not immune to model misspecification: if the forward model is inaccurate, the posterior itself will be biased, regardless of how thoroughly it is sampled. For optimization-based methods, the failure mode is more acute. Because the optimizer seeks a single point estimate, model misspecification and incompleteness can drive it towards reward hacking , exploiting deficiencies in the forward model to reduce the loss function while deviating from the true underlying linear field — either in the recovered cosmological parameters or in the reconstructed field (in particular giving a non-Gaussian field). Without continuous regularization from the sampling process, an unconstrained optimizer can diverge catastrophically from the prior in regions poorly constrained by data. This mismatch leaves clear signatures in both the recovered power spectra and the field-level features ( 48 ; 18 , e.g.,) . The objective of this study is to develop a novel optimization-based density reconstruction algorithm, denoted as Mujic \Lambda , which is capable of explicitly maintaining the Gaussianity of the initial conditions. Instead of the commonly used Maximum a posteriori (MAP) approach, Mujic \Lambda maximizes the likelihood under the constraints based on the prior. This approach acts as a workaround for cases in which the prior overwhelms the likelihood, thereby dominating the posterior and leading MAP estimation to unreliable results. The algorithm features a projection optimizer and a rank-order matching routine; both are crucial for correcting the deviations from the Gaussianity of the initial conditions. These features differentiate Mujic \Lambda from previous approaches and significantly improve the stability and physical fidelity of the reconstruction. A major motivation for developing Mujic \Lambda is the ongoing Subaru Prime Focus Spectrograph (PFS) Subaru Strategic Program (SSP) survey ( 102 ) . The PFS-CO (cosmology) component is focused on mapping [OII] emission line galaxies within the redshift range of 0.8<z<2.4 across an area of 1400\deg^{2} , providing an ideal dataset for density reconstruction. In parallel, the PFS-GE (galaxy evolution; 36 ) component is designed to observe a variety of galaxy types over 0.7\lesssim z\lesssim 6.7 , enabling joint reconstruction using both galaxies over the entire redshift range as well as Lyman- \alpha forest tracers at z\sim 2-3 . Mujic \Lambda is particularly suitable for these applications due to its Gaussianity-preserving formulation and robustness to complicated selection functions. This manuscript is structured as follows. We start by introducing the data and methods used in Section II , including the