In this paper
AbstractIntroductionReferencesWorksPhysics and Astronomy
Primordial non-Gaussianity as a large-scale cosmic-web bias
Abstract
Primordial non-Gaussianity is a unique probe of the physics of inflation, but its large-scale-structure signature is difficult to measure because it appears on the largest, most cosmic-variance-limited modes. The standard view of local-type primordial non-Gaussianity is isotropic: it induces a scale-dependent correction to tracer bias, , on large scales. Here we show that the same large-scale response can be reformulated as an anisotropic, nonlocal environmental-bias effect. Motivated by the spectral cosmic-web hierarchy of Kitaura and Sinigaglia, we define a Hessian field , whose trace obeys and therefore carries the same enhancement that appears in the Dalal et al. scale-dependent-bias term. Diagonalizing defines four large-scale web environments, analogous to knots, filaments, sheets, and voids, although operating on scales much larger than the conventional nonlinear cosmic web. We assign one weight to each environment and test the construction with the AbacusPNG suite of local-PNG -body simulations, using . We show that these four weights reproduce the large-scale rise of the tracer bias. We further show that existing halo catalogs without PNG can be post-processed to imprint an effective signature. Beyond mock generation, the resulting cosmic-web response motivates an observational statistic based on changes in tracer occupation of large-scale environments relative to a matched Gaussian reference, complementary to the conventional low- power spectrum.
Introduction
The statistics of primordial fluctuations [12, 13, 14, 15, 16] encode information about the dynamics of inflation [1, 2]. In the simplest single-field slow-roll scenarios, primordial fluctuations are close to Gaussian, and the squeezed-limit bispectrum is suppressed by a consistency relation [3, 4, 5]. Conversely, a detection of local-type primordial non-Gaussianity, currently tightly constrained by the CMB but still a major target for large-scale-structure surveys [6, 7],
at a level significantly above the single-field expectation would point to additional fields, nontrivial dynamics, or nonstandard initial conditions.
Large-scale structure offers an especially promising probe because local PNG couples long- and short-wavelength modes. This coupling changes the abundance and clustering of biased tracers, producing the well-known scale-dependent bias [8, 9, 10, 11], which motivates power-spectrum, multitracer, correlation-function, and bispectrum strategies in current and forthcoming surveys
The statistics of primordial fluctuations [12, 13, 14, 15, 16] encode information about the dynamics of inflation. In the simplest single-field slow-roll scenarios, primordial fluctuations are close to Gaussian, and the squeezed-limit bispectrum is suppressed by a consistency relation [3, 4, 5]. Conversely, a detection of local-type primordial non-Gaussianity, currently tightly constrained by the CMB but still a major target for large-scale-structure surveys [6, 7].
where
On sufficiently large scales, , and therefore . This signature is powerful but challenging: it hinges on the largest modes, where finite-volume fluctuations, survey systematics, and relativistic projection effects are most severe. A measurement based on environmental number-density contrasts offers a different compression of the survey: rather than isolating the PNG response through a small set of low- band powers, it compares tracer occupation across spatially distributed regions. The impact of survey selection can then be modeled and tested through environment-conditioned counts and matched random catalogues. This does not render the web insensitive to contaminated long-wavelength modes, but the resulting statistic need not respond to systematics in the same way as the conventional power spectrum; the two are potentially complementary and can be combined.
Here we propose a complementary interpretation. Rather than viewing Eq. (2) purely as an isotropic correction to the tracer bias, we recast it as a large-scale nonlocal environmental-bias effect. The key observation is that the operator
has Fourier-space kernel
up to sign convention. Thus, relative to the density field, carries precisely the large-scale enhancement characteristic of local PNG bias. The underlying local-PNG ensemble remains statistically isotropic; the Hessian eigenvectors and environment classes provide a spatially resolved geometrical representation of its nonlocal bias response, rather than a preferred primordial direction.
Large-scale web operator. Following the spectral cosmic-web hierarchy introduced in [17], we define the symmetric tensor
where is the density field smoothed on a large scale . The tensor is a Hessian-like nonlocal object. Its eigenvalues,
define a four-class partition of space. In analogy with standard cosmic-web classifications, we label cells as knot-like, filament-like, sheet-like, or void-like according to the number of eigenvalues above a threshold :
where is the Heaviside step function. Because is constructed from , these environments are not the usual nonlinear cosmic-web structures. They are instead large-scale, nonlocal environments whose trace reproduces the scaling relevant for local PNG.
To extend the reconstruction beyond the largest scales, we also include the standard tidal-field tensor
whose trace is the smoothed density field, . In Fourier space,
up to the same sign convention as in Eq. (5). The tidal tensor therefore encodes the anisotropic large-scale environment without the additional enhancement carried by . The two tensors probe complementary contributions: captures the nonlocal PNG-like modulation, while captures residual tidal dependence that becomes important toward higher wavenumbers.
We classify into the same four environments,
using the number of tidal eigenvalues above a threshold. This introduces a second set of four weights,
analogous to the Hessian weights but associated with the tidal operator rather than with the Hessian operator.
Environmental weighting. The Hessian-weighted tracer field is
where is the indicator field of the environment defined by . In other words,
The corresponding tidal-field extension is
where denotes the indicator field of the environment defined by . This factorized form keeps the model compact: instead of fitting sixteen joint Hessian–tidal weights [18], we use only four Hessian weights and four tidal weights.
For each classification we also measure the four environment number densities,
and the corresponding fractions
The four numbers , , , quantify how the tracer population is distributed among knots, filaments, sheets, and voids. They are used to normalize the weighted catalog so that the mean tracer abundance is preserved. For the joint Hessian–tidal weighting this gives
where is the joint fraction of tracers classified as by and as by . When the two operators are fitted separately, the corresponding one-dimensional normalizations are
These are response weights for transforming a fixed Gaussian reference catalogue, rather than absolute tracer occupation weights in the Gaussian cosmic web. Their common value at is a convention of this transformation. The unweighted number densities of Gaussian tracers in knots, filaments, sheets, and voids need not be equal, and generally are not. An observational estimator must therefore compare its measured environment-dependent weights with the corresponding Gaussian baseline, not test equality of raw environmental occupations.
The weighted overdensities are defined by
We determine , and when included , by fitting the large-scale ratio
or, equivalently, by matching the non-Gaussian response
The target model is
The environment-dependent formulation further suggests a direct observable for future surveys. Let denote survey-calibrated tracer occupation or relative-bias weights, with Gaussian reference . Unlike the response factors defined above, need not have equal components. At leading order, the PNG-specific contrast obeys
where is calibrated from simulations with known . In the independent, normalized-contrast subspace, an inverse-covariance estimator can be constructed as
where denotes the inverse on the independent contrast subspace. This construction is a proposed follow-up estimator, not a measurement of its eventual precision. In particular, it does not by itself remove the degeneracy, and no universal sensitivity gain follows from the relation .
Simulation test. We test this construction with the AbacusPNG suite of local-PNG -body simulations [19]; related local-PNG simulation work has emphasized the mass dependence of halo clustering and accurate treatment of initial conditions and aliasing [20, 21]. We use the simulations with , analysing the snapshot, which allow us to test both the sign and amplitude dependence of the web response. For each catalog, we compute the smoothed density field, construct , classify the grid into the four large-scale web regions, and assign the corresponding weight to halos. We then measure the auto- and cross-power spectra of the weighted and unweighted halo catalogs.
Figure 1 shows the central result. The unweighted Gaussian catalog has an approximately scale-independent large-scale bias, while the local-PNG catalogs exhibit the expected rise or suppression toward small , depending on the sign of . The web-weighted halo catalog reproduces this response without modifying the particle dynamics or regenerating the initial conditions. As illustrated by the comparison with the Dalal et al. leading-order reference in Figure 1, the calibrated cosmic-web response tracks the -body-measured PNG bias more closely across the finite wavenumber range displayed. The method thereby models more than the asymptotic dependence: it captures the measured scale dependence through a small set of environmental occupations. The agreement demonstrates that the leading local-PNG signature can be represented as a nonlocal large-scale environmental modulation. The power-spectrum comparison is the calibration and validation demonstrated here; independent realizations, intermediate values, and PNG-specific squeezed bispectra provide further discriminating tests. Their interest should not be confused with an intrinsic instability of nonlocal bias models to reproduce higher-order statistics. Figure 2 shows that the emulated PNG power spectra are statistically consistent with the corresponding -body measurements, with deviations remaining within the cosmic-variance uncertainty. The weights can also be interpreted operationally as a cosmic-web-dependent selection function. Since only relative weights affect the environmental modulation, we normalize them to the largest weight,
Figure 1. Large-scale bias emulation (blue circles) for different configurations vs reference calculations (orange squares) and Dalal et al. formula (dashed curves). The dark grey band marks the range dominated by the Hessian emulation, while the light grey band marks the range where the tidal correction improves the agreement. The horizontal bars in the ratio panels indicate . Error bars are based on 20 Abacus simulations.
Figure 2. Shown are the emulated (dashed lines) and the corresponding reference (solid lines) power-spectra for different values. Shaded areas correspond to 20 Abacus simulations.
A PNG-like catalog can then be generated by downsampling objects in each web class with probability . In this way, a single high-resolution -body simulation can be converted into a family of effective PNG catalogs with different values and signs of , by changing only the large-scale cosmic-web-dependent selection probabilities.
Weight hierarchy. The fitted Hessian weights are not arbitrary. We find a robust hierarchy controlled by the sign of :
whereas
This ordering has a simple interpretation. Positive local PNG enhances the abundance of biased tracers in regions selected by the corresponding sign of the long-wavelength PNG modulation. In the large-scale web basis, this modulation is captured preferentially by knot-like environments. Negative reverses the modulation, assigning the largest relative weights to void-like environments. This results is fully consistent with the well-known enhancements of knots in positive fnl and of voids in negative fnl scenarios [8].
Table 1 shows the fitted weights for positive and negative . The monotonic ordering of is a compact, interpretable summary of the PNG response. The rows quantify the additional improvement obtained from the tidal-field tensor. While the Hessian weights dominate the largest-scale PNG response, the tidal weights absorb residual anisotropic environmental dependence and improve the agreement at higher . The four environment number-density fractions enter through the normalization of the weighted catalog, ensuring that the reweighted sample preserves the original mean abundance. The cross-power spectrum between the -based
| Config. | Knot | Filament | Sheet | Void | |
| 100 | 1.0000 | 0.9779 0.0030 | 0.9555 0.0031 | 0.9283 0.0038 | |
| 0.8217 0.1377 | 0.9835 0.0986 | 1.0000 | 0.9489 0.0751 | ||
| 30 | 1.0000 | 0.9945 0.0011 | 0.9867 0.0011 | 0.9790 0.0022 | |
| 0.8867 0.0533 | 0.9695 0.0283 | 1.0000 | 0.9677 0.0185 | ||
| 0.9758 0.0021 | 0.9835 0.0022 | 0.9898 0.0021 | 1.0000 | ||
| 0.9734 0.0502 | 0.9814 0.0254 | 1.0000 | 0.9978 0.0200 | ||
| 0.9230 0.0036 | 0.9494 0.0033 | 0.9698 0.0040 | 1.0000 | ||
| 0.9156 0.1208 | 0.9741 0.0614 | 0.9853 0.0510 | 1.0000 |
Table 1. Final results for the four Hessian-environment weights , associated with the PNG-like response, and the four tidal-environment weights , associated with the tidal-field correction. The environment number-density fractions for knots, filaments, sheets, and voids are used to normalize the weighted catalog so that the mean tracer abundance is preserved. For all configurations, the average Gaussian bias is , measured from the ABACUS simulations.
web field and the halo-density response induced by primordial non-Gaussianity reveals a clear large-scale correlation, as shown in Figure 3. The same correspondence is visually apparent in the field-level comparison shown in Figure 4.
Figure 3. We show the cross correlation of the averaged Hessian-Web over a wide range of eigenvalue thresholds and the difference field with .

Figure 4. Two-dimensional slices of the dark matter distribution and its large-scale web classification. Bottom panels: Two-dimensional sample slices of the dark matter overdensity field, with side length 2000 and grid size , extracted from ABACUS mock catalogs at . The left panel displays a simulation with primordial non-Gaussianity implemented in the initial conditions with , while the right panel shows the corresponding Gaussian catalog with the same seed. Top panels: Cosmic-web classifications obtained from the Hessian operator. The structured appearance results from averaging over different thresholds, ranging from to . The matter distribution is segmented into four large-scale environments: knots, filaments, sheets, and voids. The difference field is scaled to enhance the contrast, we also applied a gaussian filter .
Post-processing existing catalogs. An immediate application is to imprint an effective PNG signature on existing Gaussian catalogs. Given a catalog and its density field, one computes , assigns each object to a large-scale web class, and applies the weights in Eq. (13). When extending the reconstruction to higher wavenumbers, one additionally computes and applies the joint weighting in Eq. (15). This operation changes the large-scale clustering while preserving the small-scale catalog structure, positions, velocities, and selection function. It therefore provides a fast emulator for PNG-like responses in mock catalogs.
The essential advantage of applying these weights to a high-resolution Gaussian -body catalogue is that the detailed halo phase-space distribution and the higher-order clustering already encoded by the expensive reference simulation remain the starting point of the construction. The PNG signal is imposed by adjusting the relative tracer number densities within a few nonlocal web environments, rather than by replacing the resolved catalogue with a new, substantially lower-resolution evolution. These changes in occupation do modify the -point statistics of the resulting tracer sample; they do not erase the information on halo positions, velocities and internal structure supplied by the reference simulation. If the weights are implemented by stochastic subsampling, the associated sampling noise must be included in any covariance analysis.
The present operation does not regenerate PNG-dependent merger histories or a fully evolved PNG halo mass function [22]; it is not intended to do so. This does not imply a loss of the reference simulation’s validated higher-order clustering information. Indeed, the DESI mock challenge of Balaguera-Antolínez et al. [23] validated halo abundance, two- and three-point statistics, and relevant covariance structure against detailed -body references. Coloma-Nadal et al. [18] further demonstrated accurate bispectra from hierarchical cosmic-web bias without fitting that observable directly. These tests underpin our DESI HOLI-mocks modelling programme and contradict any general assertion that cosmic-web-based bias modelling is intrinsically inaccurate at higher order. The newly introduced PNG weighting changes the higher-order statistics of the output sample, as a PNG response should; direct validation of its PNG-specific squeezed bispectrum and covariance is complementary to the already demonstrated large-scale halo–matter bias response.
A promising direction for future work is to combine this framework with machine-learning techniques to learn an optimal mapping between high-fidelity -body simulations and computationally inexpensive PNG mock catalogues. Such a multi-fidelity approach could learn the dependence of nonlocal environmental bias on , redshift, tracer population, and cosmic-web morphology, while retaining the detailed halo dynamics and higher-order clustering information of the reference simulations. This would extend the present low-dimensional weighting scheme towards a flexible, physically interpretable emulator, calibrated and validated against high-fidelity simulations, for the efficient production of large ensembles of PNG mocks.
Discussion. We have shown that the leading large-scale signature of local primordial non-Gaussianity can be expressed as a nonlocal cosmic-web bias. The trace of the Hessian is , and therefore has the same large-scale scaling as the Dalal et al. scale-dependent-bias correction. Diagonalizing this Hessian yields four large-scale environments. Weighting tracers by these environments reproduces the PNG-induced rise of the bias and reveals a sign-dependent hierarchy among knot-, filament-, sheet-, and void-like regions.
The inclusion of the tidal tensor extends the description beyond the leading PNG-like response. In contrast to , whose trace carries the inverse-Laplacian enhancement, describes the standard anisotropic tidal environment. Its four additional weights capture residual environmental dependence and improve the reconstruction toward higher wavenumbers while preserving a compact eight-parameter structure.
This perspective connects three ideas that are usually treated separately: primordial mode coupling, scale-dependent bias, and anisotropic environmental dependence. It also provides a fast way to post-process halo or galaxy mock catalogs into catalogs with an effective large-scale PNG response, while preserving their small-scale structure, positions, velocities, and selection function.
The new interpretation also motivates a complementary observational programme. A standard PNG power-spectrum analysis is particularly exposed to contamination of its lowest Fourier modes. In contrast, a cosmic-web measurement can compress the information into tracer-number-density or relative-bias contrasts across the many knot-, filament-, sheet-, and void-like regions distributed throughout the survey. The observable is not the power at one chosen small but a calibrated change in environmental occupation relative to Gaussian expectations. Consequently, selection effects can be diagnosed and marginalized using environment-conditioned counts, angular and radial splits, and survey randoms propagated through the same web-reconstruction pipeline. The inverse-Laplacian operator itself is sensitive to large-scale contamination, so these are distinct controls rather than a guarantee of immunity to systematics. A particularly promising strategy is a joint analysis of the conventional and the environmental contrasts, with their cross-covariance modelled consistently.
The observational use of this large-scale web response for PNG detection requires a dedicated treatment of survey masks, selection functions, redshift-space distortions, random catalogs, and calibration against mocks. We leave this data-analysis framework to a companion paper.
KPV thanks the IAC for hospitality during his undergraduate internship and bachelor thesis under the supervision of FSK and FS. FSK proposed the idea of this project and acknowledges the Instituto de Astrofísica de Canarias (IAC) for continuous support of the Cosmology with LSS probes activities. The authors thank the Spanish Ministry of Innovation and Science / Agencia Estatal de Investigación for support of the project Big Data of the Cosmic Web (PID2020-120612GB-I00/AEI/10.13039/501100011033), under which this work was conceived and carried out, and the FIRE (Field-level bayesian Inference to Reconstruct the univErse) project, recently funded by a Proyecto de Generación de Conocimiento 2024, PID2024-160504NB-I00. FS acknowledges support from the Institute for Fundamental Physics of the Universe postdoctoral fellowship scheme. FS is also grateful to IAC for hospitality during part of the realization of this project.
References
References
- [1]N. Bartolo, E. Komatsu, S. Matarrese, and A. Riotto, Physics Reports 402, 103 (2004), arXiv:astro-ph/0406398 [astro-ph].DOI
- [2]X. Chen, Advances in Astronomy 2010, 638979 (2010), arXiv:1002.1416 [astro-ph.CO].DOI
- [3]J. Maldacena, Journal of High Energy Physics 2003, 013 (2003), arXiv:astro-ph/0210603 [astro-ph].DOI
- [4]V. Acquaviva, N. Bartolo, S. Matarrese, and A. Riotto, Nuclear Physics B 667, 119 (2003), arXiv:astro-ph/0209156 [astro-ph].DOI
- [5]P. Creminelli and M. Zaldarriaga, Journal of Cosmology and Astroparticle Physics 2004 (10), 006, arXiv:astro-ph/0407059 [astro-ph].DOI
- [6]Planck Collaboration, Astronomy & Astrophysics 641, A9 (2020), arXiv:1905.05697 [astro-ph.CO].
- [7]C. T. Byrnes and K.-Y. Choi, Advances in Astronomy 2010, 724525 (2010), arXiv:1002.3110 [astro-ph.CO].DOI
- [8]N. Dalal, O. Doré, D. Huterer, and A. Shirokov, Physical Review D 77, 123514 (2008), arXiv:0710.4560 [astro-ph].DOI
- [9]A. Slosar, C. Hirata, U. Seljak, S. Ho, and N. Padmanabhan, Journal of Cosmology and Astroparticle Physics 2008 (8), 031, arXiv:0805.3580 [astro-ph].DOI
- [10]S. Matarrese and L. Verde, Astrophysical Journal Letters 677, L77 (2008), arXiv:0801.4826 [astro-ph].DOI
- [11]V. Desjacques and U. Seljak, Physical Review D 81, 023006 (2010), arXiv:0907.2257 [astro-ph.CO].DOI
- [12]N. Hamaus, U. Seljak, and V. Desjacques, Physical Review D 84, 083509 (2011), arXiv:1104.2321 [astro-ph.CO].DOI
- [13]D. Karagiannis, A. Lazanu, M. Liguori, A. Raccanelli, N. Bartolo, and L. Verde, Monthly Notices of the Royal Astronomical Society 478, 1341 (2018), arXiv:1801.09280 [astro-ph.CO].DOI
- [14]M. Ballardini, W. L. Matthewson, and R. Maartens, Monthly Notices of the Royal Astronomical Society 489, 1950 (2019), arXiv:1906.04730 [astro-ph.CO].DOI
- [15]W. Riquelme, S. Avila, J. García-Bellido, A. Porredon, I. Ferrero, K. C. Chan, R. Rosenfeld, H. Camacho, A. G. Adame, A. Carnero Rosell, M. Crocce, J. De Vicente, T. Eifler, J. Elvin-Poole, X. Fang, E. Krause, M. Rodriguez Monroy, A. J. Ross, E. Sanchez, and I. Sevilla, Monthly Notices of the Royal Astronomical Society 523, 603 (2023), arXiv:2209.07187 [astro-ph.CO].DOI
- [16]C. Heinrich, O. Doré, and E. Krause, Physical Review D 109, 123511 (2024), arXiv:2311.13082 [astro-ph.CO].
- [17]F.-S. Kitaura and F. Sinigaglia, arXiv e-prints, arXiv:2603.15834 (2026), arXiv:2603.15834 [astro-ph.CO].arxiv.org/abs/2603.15834
- [18]J. M. Coloma-Nadal, F.-S. Kitaura, J. E. García-Farieta, F. Sinigaglia, G. Favole, and D. Forero Sánchez, Journal of Cosmology and Astroparticle Physics 2024 (7), 083, arXiv:2403.19337 [astro-ph.CO].
- [19]B. Hadzhyska, L. H. Garrison, D. J. Eisenstein, and S. Ferraro, Physical Review D 109, 103530 (2024), arXiv:2402.10881 [astro-ph.CO].
- [20]O. Gutiérrez Adame, S. Avila, V. Gonzalez-Perez, G. Yepes, M. Pellejero, M. S. Wang, C.-H. Chuang, Y. Feng, J. Garcia-Bellido, and A. Knebe, Astronomy & Astrophysics 689, A69 (2024), arXiv:2312.12405 [astro-ph.CO].DOI
- [21]A. G. Adame, S. Avila, V. Gonzalez-Perez, O. Hahn, G. Yepes, and M. Manera, Astronomy & Astrophysics 705, A20 (2026), arXiv:2506.06200 [astro-ph].DOI
- [22]R. Scoccimarro, L. Hui, M. Manera, and K. C. Chan, Physical Review D 85, 083002 (2012), arXiv:1108.5512 [astro-ph.CO].DOI
- [23]A. Balaguera-Antolínez, F.-S. Kitaura, S. Alam, C.-H. Chuang, Y. Yu, G. Favole, F. Sinigaglia, C. Zhao, D. Brooks, A. de la Macorra, A. Font-Ribera, S. Gontcho A. Gontcho, K. Honscheid, R. Kehoe, A. Meis-ner, R. Miquel, G. Tarlé, M. Vargas-Magaña, and Z. Zhou, Astronomy & Astrophysics 673, A130 (2023), arXiv:2211.10640 [astro-ph.CO].


