Introduction
A wealth of cosmological information is encoded in the manner in which matter is clustered across the universe. On the very largest observable scales, this can offer a unique window to constrain the primordial universe and test fundamental models of gravity. While the complexities of non-linear clustering are largely mitigated at these scales, mapping the universe over such large distances on our past light-cone introduces a different set of observational and theoretical challenges. Specifically, as we probe larger and larger scales, phenomena such as dynamical and light-cone relativistic effects (McDonald 2009; Yoo 2010; Challinor & Lewis 2011; Bonvin & Durrer 2011; Tram et al. 2016; Villa & Rampf 2016), survey window effects (e.g. Feldman et al. 1994; Wilson et al. 2017), and wide-separation corrections (Yoo & Seljak 2015; Reimberg et al. 2016; Beutler & Di Dio 2020; Elkhashab et al. 2022; Noorikuhani & Scoccimarro 2023; Addis et al. 2025) to our statistics become increasingly relevant.
Ongoing Stage IV galaxy surveys, such as the Dark Energy Spectroscopic Instrument (DESI) (DESI Collaboration et al. [49 ] ), Euclid (Euclid Collaboration et al. 2025), SPHEREx (Doré et al. 2014) and the Nancy Grace Roman Space Telescope (formerly WFIRST) (Spergel et al. 2013), along with future planned missions like SIR-MOS (Wirth et al. 2026), MegaMapper (Schlegel et al. 2022) and WST (Mainieri et al. 2024), will not only allow us to access these larger scales but will also do so with increased statistical power. It is therefore important that we are confident about the systematics and uncertainties in our Large-Scale Structure (LSS) analyses.
A major motivation for these surveys is to constrain primordial non-Gaussianity (PNG), a key probe of the physics of the early universe (Achúcarro et al. 2022). Local-type (Salopek & Bond 1990; Komatsu & Spergel 2001) PNG is considered a key indicator of multi-field inflation, while the equilateral (Creminelli et al. 2006) and orthogonal (Senatore et al. 2010) shapes largely constrain the nature of single-field inflation models; together, these three shapes can therefore distinguish between different inflationary scenarios. Leading constraints, for all three shapes, derive from the Planck cosmic microwave background (CMB) data (Planck Collaboration 2018). et al. 2020b), but LSS constraints can, in theory, surpass those of the CMB with access to a far greater number of Fourier modes. However, one has to isolate this PNG signal from the proverbial haystack of late-time non-Gaussianity generated by the non-linear evolution of matter, and as such it is critical that we understand this late-time background noise if we are to measure PNG in LSS data.
The PNG signal enters into higher-order correlators, which are sensitive to non-Gaussian signals, the lowest order of which is the bispectrum (or 3-point function, its configuration-space counterpart). However, it also leaves an imprint in the galaxy power spectrum through its modification to halo bias. In the case of local-type PNG this leads to a scale-dependent correction which is a key target of power spectrum analyses on large scales (Dalal et al. 2008; Slosar et al. 2008; Matarrese & Verde 2008; Afshordi & Tolley 2008), but for non-local shapes of PNG, higher-order statistics are necessary (e.g. Cabass et al. 2022b). Next-generation galaxy surveys should provide constraints competitive with those from Planck (Planck Collaboration et al. [124 ] ), with O ( 1 ) O(1) O ( 1 ) uncertainties for local-type non-Gaussianity (Giannantonio et al. 2012; Font-Ribera et al. 2014; Amendola et al. 2018; Mueller et al. 2019; Karagiannis et al. 2020, 2021; Bull et al. 2024).
Additionally, access to these surveys should open up novel observables; while detections of local projection effects have so far been limited to observations of galaxy clusters (Wojtak et al. 2011; Sadeh et al. 2015), previous studies have forecasted that local relativistic projection effects should be detectable in the odd-parity multipoles of the multi-tracer 2-point function (Bonvin et al. 2016; Gaztanaga et al. 2017; Montano & Camera 2024a,b; Novara et al. 2025; Montano et al. 2026a) and single-tracer 3-point function (Maartens et al. 2020; Noorikuhani & Scoccimarro 2023; Addis et al. 2025; Rossiter et al. 2025; Montano et al. 2026a; Rossiter et al. 2026) for Stage IV and V galaxy surveys. Due to their sensitivity to the metric potentials, Φ \Phi Φ and Ψ \Psi Ψ , these corrections can be used to construct novel tests of gravity on the largest scales (Lombriser et al. 2013; Bonvin & Fleury 2018; Bonvin et al. 2020; Castello et al. 2022; Sobral Blanco & Bonvin 2023; Castello et al. 2025; Camera et al. 2026).
These relativistic effects generate signals that share the same scale-dependence as that sourced by local-type PNG (Bruni et al. 2012; Camera et al. 2015; Maartens et al. 2021a). Consequently, neglecting these effects in an analysis can bias the measurement of local-type f N L f_\mathrm{NL} f NL , or produce a false detection, both in the 2-point (Wang et al. 2020a; Guedezounme et al. 2025; Addis et al. 2026) and 3-point functions (Umeh et al. 2017; Di Dio et al. 2017; Rossiter et al. 2025; Addis et al. 2026). Ignoring these contributions can lead to other biases on our inference, such as on cosmological parameters (Addis et al. 2026) and the power spectrum turnover scale (Dube et al. 2026). More recent work has also considered the uncertainty in the relativistic biases, the magnification bias, Q Q Q , and the evolution bias, b e b_e b e , which, through their degeneracy with local-type f N L f_\mathrm{NL} f NL , increases its uncertainty (Wang et al. 2020a; Addis et al. 2026).
Multi-tracer methods, whereby one correlates different samples, have been proposed to enable one to measure the leading-order ( H / k ) (\mathcal{H}/k) ( H / k ) odd-parity relativistic effects in the cross-power spectrum (McDonald 2009; Bonvin 2014). Bright-faint splits, in which a single galaxy population is split by luminosity into two non-overlapping samples, allow such a measurement within a single survey (Bonvin et al. 2016; Gaztanaga et al. 2017; Montano & Camera 2024a,b; Novara et al. 2025; Montano et al. 2026a). Furthermore, as these approaches enhance our knowledge of the luminosity function (e.g. Blanco et al. 2024), they can also strengthen our constraining power on f N L f_\mathrm{NL} f NL (Addis et al. 2026). Additionally, by cross-correlating highly biased samples one can boost the PNG signal (Seljak 2009; Barreira & Krause 2023; Karagiannis et al. 2024). Therefore, multi-tracer approaches can improve LSS constraints on both relativistic effects and PNG.
The 3D Fourier power spectrum and bispectrum (or their configuration-space counterparts) are common bases for this type of analysis. These statistics retain full 3D information and allow for the isolation of leading-order relativistic effects in the odd-parity multipoles. However, one also has to consider the impact of wide-separation effects; these arise when considering large scales on our past light-cone and exhibit the same scale dependence as the signals of interest. Consequently, a degeneracy exists between these effects, PNG, and relativistic corrections (Noorikuhani & Scoccimarro 2023; Pardede et al. 2023; Addis et al. 2025; Guedezounme et al. 2025; Addis et al. 2026; Wen et al. 2026).
On large scales, clustering statistics therefore present a complex picture, with several degeneracies between effects that are not yet part of standard LSS analysis pipelines. Robustly accounting for these effects both in the power spectrum and bispectrum is critical not only for an unbiased analysis but also for recovering the true uncertainties on our parameters of interest. Therefore, understanding when these effects are important or when they can be neglected is useful for future work in this space.
This work builds directly on Addis et al. (2025) and Addis et al. (2026), hereafter Paper I and Paper II, and presents comprehensive joint power spectrum and bispectrum forecasts that focus on relativistic and wide-separation effects, their impact on local, equilateral, and orthogonal PNG, and the detectability of the local relativistic effects themselves, for ongoing and future surveys. We consider the uncertainty in our relativistic biases and the impact it has on our measurement of PNG and relativistic effects; in particular, we show that the odd-parity multipoles break most of the resulting degeneracy with local-type f N L f_\mathrm{NL} f NL , leaving the magnification bias as the limiting nuisance parameter, and we quantify how precisely it must be determined independently of the clustering. We include integrated effects in the power spectrum using the results from Paper II but leave their computation in the bispectrum to future work.
Alternative bases to the 3D Fourier summary statistics are potentially beneficial on ultra-large scales, such as the spherical Fourier–Bessel basis (Fisher et al. 1995; Heavens & Taylor 1995) or the total angular momentum basis (Dai et al. 2012), which retain both the full small-scale redshift information and the complete wide-separation information without the need for a series expansion. Indeed, several works have investigated local-type PNG and relativistic corrections in the spherical Fourier–Bessel basis due to these properties (Yoo & Desjacques 2013; Wang et al. 2020b; Wen et al. 2025, 2026; Lukens & Jeong 2026); however, challenges remain due to its computational complexity, and its extension to higher-order statistics is non-trivial (e.g. Benabou et al. 2024).
In this work we use the term relativistic effects broadly, to encompass integrated (e.g. lensing and ISW), local relativistic projection (e.g. relativistic Doppler and gravitational redshift) and “dynamical” relativistic effects that arise in second-order perturbation theory from non-linear evolution (these are also non-integrated and therefore local). We use the term light-cone effects to include both relativistic effects and wide-separation (wide-angle and radial evolution) corrections.
The paper is laid out as follows: § 2 defines our estimators and focuses on the PNG contributions to them, including how we model the PNG biases. § 3 outlines the key forecast method and details, along with the specific choices and assumptions that we have made. § 4 extends the covariance calculation of Paper II to the bispectrum and defines our covariance matrix. Results focused on PNG and relativistic effects are presented in § 5 and § 6 , respectively. Further details about the survey and bias modelling choices we make are included in the appendices.
All of our results are produced and implemented using the public Python code CosmoWAP ([link](https://github.com/CosmologicalSurveyData/ CosmoWAP)) (Addis et al. 2025, 2026; Addis 2026) with companion Mathematica notebooks included in MATHWAP (link ) (Addis 2025).
Theory and Observables
This section briefly describes the theoretical modelling we use in our forecasts. First, we define our choice of estimators and their connection to what we compute theoretically. § 2.1.1 outlines the PNG contributions to the power spectrum and bispectrum, including through the modification to the halo bias (§ 2.1.2 ), and § 2.2 the corrections to the same kernels from local relativistic projection effects. Finally, § 2.3 describes how we connect our halo bias quantities to the galaxy biases we use in our forecasts.
In this work we consider the observable summary statistics that are measured using the standard FFT estimators: the “Yamamoto” estimator (Feldman et al. 1994; Beutler et al. 2014; Bianchi et al. 2015; Scoccimarro 2015) for the power spectrum,
P ^ ℓ a b ( k ) = 2 ℓ + 1 N k ∫ S 1 d 3 q 1 ( 2 π ) 3 ∫ S 2 d 3 q 2 δ D ( q 1 + q 2 ) ∫ x 1 , x 2 e − i ( q 1 ⋅ x 1 + q 2 ⋅ x 2 ) Δ W a ( x 1 ) Δ W b ( x 2 ) L ℓ ( q ^ 1 ⋅ d ^ ) , (1) \hat{P}^{ab}_{\ell}(k) = \frac{2\ell+1}{N_k}\int_{\mathcal{S}_1}\frac{\mathrm{d}^3q_1}{(2\pi)^3}\int_{\mathcal{S}_2}\mathrm{d}^3q_2\,\delta_{\mathrm{D}}(\mathbf{q}_1+\mathbf{q}_2)\int_{\mathbf{x}_1,\mathbf{x}_2}e^{-\mathrm{i}(\mathbf{q}_1\cdot\mathbf{x}_1+\mathbf{q}_2\cdot\mathbf{x}_2)}\Delta^a_W(\mathbf{x}_1)\Delta^b_W(\mathbf{x}_2)L_\ell(\hat{\mathbf{q}}_1\cdot\hat{\mathbf{d}}), \tag*{(1)} P ^ ℓ ab ( k ) = N k 2 ℓ + 1 ∫ S 1 ( 2 π ) 3 d 3 q 1 ∫ S 2 d 3 q 2 δ D ( q 1 + q 2 ) ∫ x 1 , x 2 e − i ( q 1 ⋅ x 1 + q 2 ⋅ x 2 ) Δ W a ( x 1 ) Δ W b ( x 2 ) L ℓ ( q ^ 1 ⋅ d ^ ) , (1)
and the “Scoccimarro” estimator2 for the bispectrum (Scoccimarro 2015),
B ^ ℓ a b c ( k 1 , k 2 , k 3 ) = 2 ℓ + 1 N 123 T k 1 3 ∫ S 1 d 3 q 1 ( 2 π ) 3 ∫ S 2 d 3 q 2 ∫ S 3 d 3 q 3 δ D ( q 123 ) ∫ x 1 , x 2 , x 3 e − i ( q 1 ⋅ x 1 + q 2 ⋅ x 2 + q 3 ⋅ x 3 ) Δ W a ( x 1 ) Δ W b ( x 2 ) Δ W c ( x 3 ) L ℓ ( q ^ 1 ⋅ d ^ ) , (2) \hat{B}^{abc}_{\ell}(k_1,k_2,k_3) = \frac{2\ell+1}{N^T_{123}k_1^3}\int_{\mathcal{S}_1}\frac{\mathrm{d}^3q_1}{(2\pi)^3}\int_{\mathcal{S}_2}\mathrm{d}^3q_2\int_{\mathcal{S}_3}\mathrm{d}^3q_3\,\delta_{\mathrm{D}}(\mathbf{q}_{123})\int_{\mathbf{x}_1,\mathbf{x}_2,\mathbf{x}_3}e^{-\mathrm{i}(\mathbf{q}_1\cdot\mathbf{x}_1+\mathbf{q}_2\cdot\mathbf{x}_2+\mathbf{q}_3\cdot\mathbf{x}_3)}\Delta^a_W(\mathbf{x}_1)\Delta^b_W(\mathbf{x}_2)\Delta^c_W(\mathbf{x}_3)L_\ell(\hat{\mathbf{q}}_1\cdot\hat{\mathbf{d}}), \tag*{(2)} B ^ ℓ ab c ( k 1 , k 2 , k 3 ) = N 123 T k 1 3 2 ℓ + 1 ∫ S 1 ( 2 π ) 3 d 3 q 1 ∫ S 2 d 3 q 2 ∫ S 3 d 3 q 3 δ D ( q 123 ) ∫ x 1 , x 2 , x 3 e − i ( q 1 ⋅ x 1 + q 2 ⋅ x 2 + q 3 ⋅ x 3 ) Δ W a ( x 1 ) Δ W b ( x 2 ) Δ W c ( x 3 ) L ℓ ( q ^ 1 ⋅ d ^ ) , (2)
where Δ W ( x ) = Δ ( x ) W ( x ) \Delta_W(\mathbf{x}) = \Delta(\mathbf{x})W(\mathbf{x}) Δ W ( x ) = Δ ( x ) W ( x ) represents the windowed field for a given survey window function, W ( x ) W(\mathbf{x}) W ( x ) . We normalise such that ⟨ Δ ( k ) Δ ( k ′ ) ⟩ = ( 2 π ) 3 δ D ( k + k ′ ) P ( k ) \langle\Delta(\mathbf{k})\Delta(\mathbf{k}')\rangle= (2\pi)^3\delta_{\mathrm{D}}(\mathbf{k}+\mathbf{k}')P(k) ⟨ Δ ( k ) Δ ( k ′ )⟩ = ( 2 π ) 3 δ D ( k + k ′ ) P ( k ) , with δ D ( 0 ) = V / ( 2 π ) 3 = k f − 3 \delta_{\mathrm{D}}(0)=V/(2\pi)^3=k_{\mathrm{f}}^{-3} δ D ( 0 ) = V / ( 2 π ) 3 = k f − 3 for survey volume V V V , and N k N_k N k and N 123 T N^T_{123} N 123 T denote the number of fundamental modes and triangles in each bin. In practice, the integrals are evaluated as discrete sums over Fourier modes, ∫ S i d 3 q i F ( q i ) → k f 3 ∑ q i ∈ S i F ( q i ) \int_{\mathcal{S}_i}\mathrm{d}^3q_i\,F(\mathbf{q}_i)\rightarrow k_{\mathrm{f}}^3\sum_{\mathbf{q}_i\in\mathcal{S}_i}F(\mathbf{q}_i) ∫ S i d 3 q i F ( q i ) → k f 3 ∑ q i ∈ S i F ( q i ) , over thin k k k -space shells, S i \mathcal{S}_i S i , of width Δ k \Delta k Δ k , centred at k i k_i k i , such that S i ≡ S ( k i ∣ Δ k ) \mathcal{S}_i\equiv\mathcal{S}(k_i|\Delta k) S i ≡ S ( k i ∣Δ k ) is the region of k k k -magnitudes contained by a given k i k_i k i -bin, k i − Δ k / 2 ≤ k < k i + Δ k / 2 k_i-\Delta k/2\leq k<k_i+\Delta k/2 k i − Δ k /2 ≤ k < k i + Δ k /2 .
Local statistics
In order to compute what these estimators measure in perturbation theory, it is helpful to introduce the ‘local’ power spectrum and bispectrum ([77 ] (Zaroubi & Hoffman 1996; Scoccimarro 2015; Castorina & White 2018)), defined for a region of space described by a single line of sight (LOS), d \mathbf{d} d . For a given LOS, the local power spectrum is
P l o c a b ( k , μ ; d ) = ∫ x 12 e − i k ⋅ x 12 ⟨ Δ a ( x 1 ) Δ b ( x 2 ) ⟩ , (3) P^{ab}_{\mathrm{loc}}(k,\mu;\mathbf{d})=\int_{\mathbf{x}_{12}}e^{-\mathrm{i}\mathbf{k}\cdot\mathbf{x}_{12}}\langle\Delta^a(\mathbf{x}_1)\Delta^b(\mathbf{x}_2)\rangle, \tag*{(3)} P loc ab ( k , μ ; d ) = ∫ x 12 e − i k ⋅ x 12 ⟨ Δ a ( x 1 ) Δ b ( x 2 )⟩ , (3)
where the LOS orientation is defined as μ = k ^ ⋅ d ^ \mu=\hat{\mathbf{k}}\cdot\hat{\mathbf{d}} μ = k ^ ⋅ d ^ and x 12 = x 1 − x 2 \mathbf{x}_{12}=\mathbf{x}_1-\mathbf{x}_2 x 12 = x 1 − x 2 . Similarly, for the local bispectrum,
B l o c a b c ( k 1 , k 2 , k 3 , μ 1 , ϕ 12 ; d ) = ∫ d 3 x 13 d 3 x 23 e − i ( k 1 ⋅ x 13 + k 2 ⋅ x 23 ) ⟨ Δ a ( x 1 ) Δ b ( x 2 ) Δ c ( x 3 ) ⟩ , (4) B^{abc}_{\mathrm{loc}}(k_1,k_2,k_3,\mu_1,\phi_{12};\mathbf{d})=\int\mathrm{d}^3x_{13}\,\mathrm{d}^3x_{23}\,e^{-\mathrm{i}(\mathbf{k}_1\cdot\mathbf{x}_{13}+\mathbf{k}_2\cdot\mathbf{x}_{23})}\langle\Delta^a(\mathbf{x}_1)\Delta^b(\mathbf{x}_2)\Delta^c(\mathbf{x}_3)\rangle, \tag*{(4)} B loc ab c ( k 1 , k 2 , k 3 , μ 1 , ϕ 12 ; d ) = ∫ d 3 x 13 d 3 x 23 e − i ( k 1 ⋅ x 13 + k 2 ⋅ x 23 ) ⟨ Δ a ( x 1 ) Δ b ( x 2 ) Δ c ( x 3 )⟩ , (4)
where the LOS orientation is defined by μ 1 = k ^ 1 ⋅ d ^ \mu_1=\hat{\mathbf{k}}_1\cdot\hat{\mathbf{d}} μ 1 = k ^ 1 ⋅ d ^ and ϕ 12 \phi_{12} ϕ 12 is the azimuthal angle such that μ 2 = μ 1 cos θ + 1 − μ 1 2 sin θ cos ϕ 12 \mu_2=\mu_1\cos\theta+\sqrt{1-\mu_1^2}\sin\theta\cos\phi_{12} μ 2 = μ 1 cos θ + 1 − μ 1 2 sin θ cos ϕ 12 , where cos θ = k ^ 1 ⋅ k ^ 2 \cos\theta=\hat{\mathbf{k}}_1\cdot\hat{\mathbf{k}}_2 cos θ = k ^ 1 ⋅ k ^ 2 (see Scoccimarro et al. 1999 for details on bispectrum geometry).
These configuration-space density fields can then be expressed as the Fourier transforms of the Fourier-space kernels and linear density field, and we can contract the Dirac deltas after computing our wide-separation expansions. By computing these contributions directly in Cartesian Fourier space, this framework enables the consistent inclusion of both wide-separation corrections and integrated relativistic effects (Noorikuhani & Scoccimarro 2023; Addis et al. 2025; Guedezounme et al. 2025; Addis et al. 2026).
If we consider the ‘true’ unwindowed theory3 then we can straightforwardly relate these local quantities to the survey-averaged statistics one measures with the FFT estimators (Scoccimarro 2015). The LOS orientation can be projected onto a Legendre basis such that we obtain the local multipoles, which are then related to the survey-averaged statistics by averaging over the survey volume. Thus, the survey-averaged unwindowed power spectrum can be written as
P ^ ℓ a b ( k ) = 2 ℓ + 1 2 ∫ V d 3 d V ∫ − 1 1 d μ P l o c a b ( k , μ ; d ) L ℓ ( μ ) , (5) \hat{P}^{ab}_{\ell}(k)=\frac{2\ell+1}{2}\int_V\frac{\mathrm{d}^3d}{V}\int_{-1}^{1}\mathrm{d}\mu\,P^{ab}_{\mathrm{loc}}(k,\mu;\mathbf{d})L_\ell(\mu), \tag*{(5)} P ^ ℓ ab ( k ) = 2 2 ℓ + 1 ∫ V V d 3 d ∫ − 1 1 d μ P loc ab ( k , μ ; d ) L ℓ ( μ ) , (5)
and similarly for the ‘Scoccimarro’ bispectrum multipole estimator,
B ^ ℓ a b c ( k 1 , k 2 , k 3 ) = 2 ℓ + 1 2 ∫ V d 3 d V ∫ − 1 1 d μ 1 ∫ 0 2 π d ϕ 12 B l o c a b c ( k 1 , k 2 , k 3 , μ 1 , ϕ 12 ; d ) L ℓ ( μ 1 ) . (6) \hat{B}^{abc}_{\ell}(k_1,k_2,k_3)=\frac{2\ell+1}{2}\int_V\frac{\mathrm{d}^3d}{V}\int_{-1}^{1}\mathrm{d}\mu_1\int_0^{2\pi}\mathrm{d}\phi_{12}\,B^{abc}_{\mathrm{loc}}(k_1,k_2,k_3,\mu_1,\phi_{12};\mathbf{d})L_\ell(\mu_1). \tag*{(6)} B ^ ℓ ab c ( k 1 , k 2 , k 3 ) = 2 2 ℓ + 1 ∫ V V d 3 d ∫ − 1 1 d μ 1 ∫ 0 2 π d ϕ 12 B loc ab c ( k 1 , k 2 , k 3 , μ 1 , ϕ 12 ; d ) L ℓ ( μ 1 ) . (6)
Below, we give the tree-level kernels at zeroth order in the wide-separation expansion: the PNG contributions in § 2.1.1 and 2.1.2 and the local relativistic projection effects in § 2.2 . We do not reproduce the derivations for the wide-separation corrections, or the integrated (lensing, ISW and time-delay) contributions that we include in the power spectrum; for these we adopt the formalism of Paper II for the power spectrum and Paper I for the bispectrum, both of which are implemented in the Python package CosmoWAP.
PNG contributions
PNG generates non-zero higher-order correlators in the primordial curvature perturbations; the primordial bispectrum B Φ ( k 1 , k 2 , k 3 ) B_{\Phi}(k_1,k_2,k_3) B Φ ( k 1 , k 2 , k 3 ) , which can be related to the primordial curvature bispectrum, can be linearly extrapolated to source the PNG contribution to the late-time matter density bispectrum in real space,
B P N G ( k 1 , k 2 , k 3 , z ) = M ( k 1 , z ) M ( k 2 , z ) M ( k 3 , z ) B Φ ( k 1 , k 2 , k 3 ) , (7) B_{\mathrm{PNG}}(k_1,k_2,k_3,z)=M(k_1,z)M(k_2,z)M(k_3,z)B_{\Phi}(k_1,k_2,k_3), \tag*{(7)} B PNG ( k 1 , k 2 , k 3 , z ) = M ( k 1 , z ) M ( k 2 , z ) M ( k 3 , z ) B Φ ( k 1 , k 2 , k 3 ) , (7)
using the transfer function
M ( k , z ) = 2 c 2 k 2 D ( z ) 3 Ω m H 0 2 T ( k ) , (8) M(k,z)=\frac{2c^2k^2D(z)}{3\Omega_{\mathrm{m}}H_0^2}T(k), \tag*{(8)} M ( k , z ) = 3 Ω m H 0 2 2 c 2 k 2 D ( z ) T ( k ) , (8)
which relates the Bardeen power spectrum to the late-time matter power spectrum
P ( k , z ) = M 2 ( k , z ) P Φ ( k ) . (9) P(k,z)=M^2(k,z)P_{\Phi}(k). \tag*{(9)} P ( k , z ) = M 2 ( k , z ) P Φ ( k ) . (9)
Here, D ( z ) D(z) D ( z ) and T ( k ) T(k) T ( k ) are the linear growth factor and the matter transfer function, respectively.
The primordial bispectrum templates, for the local, equilateral, and orthogonal shapes considered here, are defined as:
B Φ l o c ( k 1 , k 2 , k 3 ) = 2 f N L l o c [ P Φ ( k 1 ) P Φ ( k 2 ) + 2 p e r m s ] , (10) B_{\Phi}^{\mathrm{loc}}(k_1,k_2,k_3)=2f_{\mathrm{NL}}^{\mathrm{loc}}\left[P_{\Phi}(k_1)P_{\Phi}(k_2)+2\ \mathrm{perms}\right], \tag*{(10)} B Φ loc ( k 1 , k 2 , k 3 ) = 2 f NL loc [ P Φ ( k 1 ) P Φ ( k 2 ) + 2 perms ] , (10)
B Φ e q ( k 1 , k 2 , k 3 ) = 6 f N L e q [ − ( P Φ ( k 1 ) P Φ ( k 2 ) + 2 p e r m s ) − 2 ( P Φ ( k 1 ) P Φ ( k 2 ) P Φ ( k 3 ) ) 2 / 3 + ( P Φ 1 / 3 ( k 1 ) P Φ 2 / 3 ( k 2 ) P Φ ( k 3 ) + 5 p e r m s ) ] , (11) \begin{aligned}
B_{\Phi}^{\mathrm{eq}}(k_1,k_2,k_3)={}&6f_{\mathrm{NL}}^{\mathrm{eq}}[-(P_{\Phi}(k_1)P_{\Phi}(k_2)+2\ \mathrm{perms})\\
&-2(P_{\Phi}(k_1)P_{\Phi}(k_2)P_{\Phi}(k_3))^{2/3}\\
&+(P_{\Phi}^{1/3}(k_1)P_{\Phi}^{2/3}(k_2)P_{\Phi}(k_3)+5\ \mathrm{perms})], \tag*{(11)}
\end{aligned} B Φ eq ( k 1 , k 2 , k 3 ) = 6 f NL eq [ − ( P Φ ( k 1 ) P Φ ( k 2 ) + 2 perms ) − 2 ( P Φ ( k 1 ) P Φ ( k 2 ) P Φ ( k 3 ) ) 2/3 + ( P Φ 1/3 ( k 1 ) P Φ 2/3 ( k 2 ) P Φ ( k 3 ) + 5 perms )] , (11)
B Φ o r t h ( k 1 , k 2 , k 3 ) = 6 f N L o r t h [ P Φ ( k 1 ) P Φ ( k 2 ) P Φ ( k 3 ) ] 2 / 3 [ ( 1 + q ) Γ 3 k 1 k 2 k 3 − q Γ 2 ( k 1 k 2 k 3 ) 2 / 3 ] , (12) B_{\Phi}^{\mathrm{orth}}(k_1,k_2,k_3)=6f_{\mathrm{NL}}^{\mathrm{orth}}\left[P_{\Phi}(k_1)P_{\Phi}(k_2)P_{\Phi}(k_3)\right]^{2/3}\left[(1+q)\frac{\Gamma^3}{k_1k_2k_3}-q\frac{\Gamma^2}{(k_1k_2k_3)^{2/3}}\right], \tag*{(12)} B Φ orth ( k 1 , k 2 , k 3 ) = 6 f NL orth [ P Φ ( k 1 ) P Φ ( k 2 ) P Φ ( k 3 ) ] 2/3 [ ( 1 + q ) k 1 k 2 k 3 Γ 3 − q ( k 1 k 2 k 3 ) 2/3 Γ 2 ] , (12)
where k T = k 1 + k 2 + k 3 k_{\mathrm{T}}=k_1+k_2+k_3 k T = k 1 + k 2 + k 3 , Δ = ( k T − 2 k 1 ) ( k T − 2 k 2 ) ( k T − 2 k 3 ) \Delta=(k_{\mathrm{T}}-2k_1)(k_{\mathrm{T}}-2k_2)(k_{\mathrm{T}}-2k_3) Δ = ( k T − 2 k 1 ) ( k T − 2 k 2 ) ( k T − 2 k 3 ) , Γ = 2 3 ( k 1 k 2 + k 2 k 3 + k 3 k 1 ) − 1 3 ( k 1 2 + k 2 2 + k 3 2 ) \Gamma=\frac{2}{3}(k_1k_2+k_2k_3+k_3k_1)-\frac{1}{3}(k_1^2+k_2^2+k_3^2) Γ = 3 2 ( k 1 k 2 + k 2 k 3 + k 3 k 1 ) − 3 1 ( k 1 2 + k 2 2 + k 3 2 ) and q = 27 / [ − 21 + 743 / ( 7 ( 20 π 2 − 193 ) ) ] ≈ 8.53 q=27/[-21+743/(7(20\pi^2-193))]\approx8.53 q = 27/ [ − 21 + 743/ ( 7 ( 20 π 2 − 193 ))] ≈ 8.53 . This is the orthogonal-LSS approximation of Senatore et al. (2010, appendix B), which, unlike the standard (CMB) orthogonal template, has the squeezed limit ∝ P Φ ( k 1 ) P Φ ( k 3 ) ( k 3 / k 1 ) 2 \propto P_{\Phi}(k_1)P_{\Phi}(k_3)(k_3/k_1)^2 ∝ P Φ ( k 1 ) P Φ ( k 3 ) ( k 3 / k 1 ) 2 , as the equilateral template does, rather than as k 3 / k 1 k_3/k_1 k 3 / k 1 . The two templates agree on equilateral triangles, and the orthogonal-LSS one is used in LSS analyses (e.g. Cabass et al. (2022b)), since the squeezed limit sets the scale-dependent bias (§ 2.1.2 ). Local-type PNG peaks in squeezed configurations (k 3 ≪ k 2 ≈ k 1 k_3\ll k_2\approx k_1 k 3 ≪ k 2 ≈ k 1 ), the equilateral and orthogonal types in equilateral ones (k 3 ≈ k 2 ≈ k 1 k_3\approx k_2\approx k_1 k 3 ≈ k 2 ≈ k 1 ), and the orthogonal type also in folded ones (k 1 ≈ 2 k 2 ≈ 2 k 3 k_1\approx2k_2\approx2k_3 k 1 ≈ 2 k 2 ≈ 2 k 3 ).
The matter bispectrum is also sourced by non-linear gravitational evolution, and as such one can write the local multi-tracer galaxy bispectrum as the standard tree-level contribution plus the intrinsic PNG contribution4
\begin{aligned}
B_{\mathrm{loc}}^{abc}(k_1,k_2,k_3,\mu_1,\phi_{12})={}&\mathcal{K}_1^a(\mathbf{k}_1,d)\mathcal{K}_1^b(\mathbf{k}_2,d)\mathcal{K}_1^c(\mathbf{k}_3,d)\\
&\times D(d)^3B_{\mathrm{PNG}}(k_1,k_2,k_3)\\
&+2\left[\mathcal{K}_1^a(\mathbf{k}_1,d)\mathcal{K}_1^b(\mathbf{k}_2,d)\mathcal{K}_2^c(-\mathbf{k}_1,-\mathbf{k}_2,d)\\
&\qquad\times D(d)^4P(k_1)P(k_2)+\mathrm{perms}\right] \tag*{(13)}
\end{aligned}
where D ( d ) D(d) D ( d ) is the linear growth factor expressed as a function of comoving distance. Additionally, the local power spectrum in linear perturbation theory is
P l o c a b ( k , μ ) = K 1 a ( k , d ) K 1 b ( − k , d ) D ( d ) 2 P ( k ) . (14) P_{\mathrm{loc}}^{ab}(k,\mu)=\mathcal{K}_1^a(k,d)\mathcal{K}_1^b(-k,d)D(d)^2P(k). \tag*{(14)} P loc ab ( k , μ ) = K 1 a ( k , d ) K 1 b ( − k , d ) D ( d ) 2 P ( k ) . (14)
The second kernel is evaluated at − k -k − k , since ⟨ δ ( k ) δ ( k ′ ) ⟩ ∝ δ D ( k + k ′ ) \langle\delta(\mathbf{k})\delta(\mathbf{k}')\rangle\propto\delta_{\mathrm{D}}(\mathbf{k}+\mathbf{k}') ⟨ δ ( k ) δ ( k ′ )⟩ ∝ δ D ( k + k ′ ) , and similarly the second-order kernel in Eq. (13) , which is contracted with the linear fields at k 1 k_1 k 1 and k 2 k_2 k 2 , is evaluated at ( − k 1 , − k 2 ) (-k_1,-k_2) ( − k 1 , − k 2 ) . The redshift-space kernels K 1 \mathcal{K}_1 K 1 and K 2 \mathcal{K}_2 K 2 , which also contain PNG contributions through the modification to the dark matter halo bias ((Baldauf et al. 2011; Tellarini et al. 2016)), can be expressed, up to linear order in f N L f_{\mathrm{NL}} f NL and dropping the explicit redshift dependence, as
K 1 a ( q i , d ) = b 1 a + f μ i 2 + f N L b Ψ a q α M ( q ) (15) \mathcal{K}_1^a(q_i,d)=b_1^a+f\mu_i^2+\frac{f_{\mathrm{NL}}b_{\Psi}^aq^\alpha}{M(q)} \tag*{(15)} K 1 a ( q i , d ) = b 1 a + f μ i 2 + M ( q ) f NL b Ψ a q α (15)
K 2 a ( q i , q j , d ) = b 1 a F 2 ( q i , q j ) + f ( q ^ i j ⋅ d ^ ) 2 G 2 ( q i , q j ) + b 2 a 2 + b Γ 2 S 2 ( q i , q j ) + f 2 ( q i j ⋅ d ^ ) [ μ i q i K 1 a ( q i , d ) + μ j q j K 1 a ( q j , d ) ] + f N L [ b Ψ a + b Ψ a N 2 ( q i , q j ) ] q i α 2 M ( q i ) + [ b Ψ a + b Ψ a N 2 ( q i , q j ) ] q j α 2 M ( q j ) . (16) \begin{aligned}
\mathcal{K}_2^a(q_i,q_j,d)={}&b_1^aF_2(q_i,q_j)+f(\hat{q}_{ij}\cdot\hat{d})^2G_2(q_i,q_j)\\
&+\frac{b_2^a}{2}+b_{\Gamma_2}S_2(q_i,q_j)\\
&+\frac{f}{2}(q_{ij}\cdot\hat{d})\left[\frac{\mu_i}{q_i}\mathcal{K}_1^a(q_i,d)+\frac{\mu_j}{q_j}\mathcal{K}_1^a(q_j,d)\right]\\
&+f_{\mathrm{NL}}\left[b_{\Psi}^a+b_{\Psi}^{a}N_2(q_i,q_j)\right]\frac{q_i^\alpha}{2M(q_i)}\\
&+\left[b_{\Psi}^a+b_{\Psi}^{a}N_2(q_i,q_j)\right]\frac{q_j^\alpha}{2M(q_j)}. \tag*{(16)}
\end{aligned} K 2 a ( q i , q j , d ) = b 1 a F 2 ( q i , q j ) + f ( q ^ ij ⋅ d ^ ) 2 G 2 ( q i , q j ) + 2 b 2 a + b Γ 2 S 2 ( q i , q j ) + 2 f ( q ij ⋅ d ^ ) [ q i μ i K 1 a ( q i , d ) + q j μ j K 1 a ( q j , d ) ] + f NL [ b Ψ a + b Ψ a N 2 ( q i , q j ) ] 2 M ( q i ) q i α + [ b Ψ a + b Ψ a N 2 ( q i , q j ) ] 2 M ( q j ) q j α . (16)
where μ i = q ^ i ⋅ d ^ \mu_i=\hat{q}_i\cdot\hat{d} μ i = q ^ i ⋅ d ^ , α = 0 \alpha=0 α = 0 for the local PNG case, and α = 2 \alpha=2 α = 2 for both the equilateral and orthogonal shapes. We also define the shorthand notation q i j = q i + q j q_{ij}=q_i+q_j q ij = q i + q j . Further, f f f is the linear growth rate, N 2 = ( q 1 ⋅ q 2 ) / q 1 2 N_2=(q_1\cdot q_2)/q_1^2 N 2 = ( q 1 ⋅ q 2 ) / q 1 2 , F 2 F_2 F 2 and G 2 G_2 G 2 are the second-order symmetric standard perturbation theory kernels (see Bernardeau et al. (2002) for a review), while b Γ 2 b_{\Gamma_2} b Γ 2 is the tidal bias and S 2 ( q 1 , q 2 ) = ( q ^ 1 ⋅ q ^ 2 ) 2 − 1 S_2(q_1,q_2)=(\hat{q}_1\cdot\hat{q}_2)^2-1 S 2 ( q 1 , q 2 ) = ( q ^ 1 ⋅ q ^ 2 ) 2 − 1 (McDonald & Roy 2009).
PNG biases
We model the PNG biases, b Ψ b_{\Psi} b Ψ and b Ψ ϕ b_{\Psi\phi} b Ψ ϕ , using results from the peak-background split (PBS) formalism (Mo et al. 1997; Slosar et al. 2008; Schmidt et al. 2013). The bias, b Ψ b_{\Psi} b Ψ , which results from PNG modulating the amplitude of small-scale fluctuations, can be defined by the logarithmic derivative of the halo number density, n h n_h n h , with respect to the amplitude of the smoothed density field, σ 8 \sigma_8 σ 8 ,
b Ψ ≡ 2 ∂ ln n h ∂ ln σ 8 . (17) b_{\Psi}\equiv2\frac{\partial\ln n_h}{\partial\ln\sigma_8}. \tag*{(17)} b Ψ ≡ 2 ∂ ln σ 8 ∂ ln n h . (17)
For a universal mass function we can derive analytic expressions for the non-Gaussian bias parameters (Schmidt et al. 2013; Desjacques et al. 2018),
b Ψ ( M , z ) = A [ 2 δ c ( b 1 − 1 ) + 4 ( d ln σ R , − α 2 d ln σ R 2 − 1 ) σ R , − α 2 σ R 2 ] (18) b_{\Psi}(M,z)=A\left[2\delta_c(b_1-1)+4\left(\frac{\mathrm{d}\ln\sigma_{R,-\alpha}^2}{\mathrm{d}\ln\sigma_R^2}-1\right)\frac{\sigma_{R,-\alpha}^2}{\sigma_R^2}\right] \tag*{(18)} b Ψ ( M , z ) = A [ 2 δ c ( b 1 − 1 ) + 4 ( d ln σ R 2 d ln σ R , − α 2 − 1 ) σ R 2 σ R , − α 2 ] (18)
and (Giannantonio & Porciani 2010; Karagiannis et al. 2018)
b Ψ ϕ ( M , z ) = 2 A [ δ c ( b 2 − 8 21 ( b 1 − 1 ) ) + ( b 1 − 1 ) ( 2 d ln σ R , − α 2 d ln σ R 2 − 3 ) σ R , − α 2 σ R 2 ] + b Ψ (19) \begin{aligned}
b_{\Psi\phi}(M,z)=2A\bigg[&\delta_c\left(b_2-\frac{8}{21}(b_1-1)\right)\\
&+(b_1-1)\left(2\frac{\mathrm{d}\ln\sigma_{R,-\alpha}^2}{\mathrm{d}\ln\sigma_R^2}-3\right)\frac{\sigma_{R,-\alpha}^2}{\sigma_R^2}\bigg]+b_{\Psi} \tag*{(19)}
\end{aligned} b Ψ ϕ ( M , z ) = 2 A [ δ c ( b 2 − 21 8 ( b 1 − 1 ) ) + ( b 1 − 1 ) ( 2 d ln σ R 2 d ln σ R , − α 2 − 3 ) σ R 2 σ R , − α 2 ] + b Ψ (19)
where M M M is halo mass, δ c = 1.686 \delta_c = 1.686 δ c = 1.686 is the critical overdensity for spherical collapse,
σ R , n 2 = 1 2 π 2 ∫ d k k 2 + n W R ( k ) 2 P m ( k ) , \sigma^2_{R,n} = \frac{1}{2\pi^2}\int dk\,k^{2+n}W_R(k)^2P_m(k), σ R , n 2 = 2 π 2 1 ∫ d k k 2 + n W R ( k ) 2 P m ( k ) ,
and W R W_R W R is a spherical top hat filter. For local PNG, α = 0 \alpha= 0 α = 0 , A = 1 A = 1 A = 1 , and thus the biases reduce to their well-known results. For equilateral PNG, A = 3 A = 3 A = 3 (Schmidt & Kamionkowski 2010; Giannantonio et al. 2012), and for the orthogonal-LSS template A = 3 − 5 q / 9 ≈ − 1.74 A = 3 - 5q/9 \approx-1.74 A = 3 − 5 q /9 ≈ − 1.74 , from its squeezed limit at k 1 = k 2 k_1 = k_2 k 1 = k 2 , taken as for the equilateral template (Schmidt & Kamionkowski 2010, section V)5 . For the non-local cases, σ R , − α \sigma_{R,-\alpha} σ R , − α can be computed for a given halo mass function (HMF). How we connect these halo biases to the galaxy biases we use in our forecasts is described in § 2.3 .
Local relativistic effects
The local relativistic projection effects add terms suppressed by powers of H / k \mathcal{H}/k H / k , where H \mathcal{H} H is the conformal Hubble rate, so that K n → K n + K n , G R \mathcal{K}_n \to\mathcal{K}_n + \mathcal{K}_{n,\mathrm{GR}} K n → K n + K n , GR . At first order,
K 1 , G R a ( q i , d ) = i μ i f γ 1 a H q i + γ 2 a H 2 q i 2 , (20) \mathcal{K}^{a}_{1,\mathrm{GR}}(\mathbf{q}_i,\mathbf{d}) = i\mu_i f\gamma^a_1\frac{\mathcal{H}}{q_i} + \gamma^a_2\frac{\mathcal{H}^2}{q_i^2}, \tag*{(20)} K 1 , GR a ( q i , d ) = i μ i f γ 1 a q i H + γ 2 a q i 2 H 2 , (20)
where the first term is the Doppler contribution and the second collects those of the velocity potential and the gravitational potentials. In terms of the evolution and magnification biases of the sample, γ 1 = b e − 2 Q − 2 ( 1 − Q ) / ( x H ) − H ′ / H 2 \gamma_1 = b_e - 2Q - 2(1-Q)/(x\mathcal{H}) - \mathcal{H}'/\mathcal{H}^2 γ 1 = b e − 2 Q − 2 ( 1 − Q ) / ( x H ) − H ′ / H 2 and γ 2 = f ( 3 − b e ) + 3 2 Ω m ( 2 − f − 2 Q + γ 1 ) \gamma_2 = f(3-b_e) + \frac{3}{2}\Omega_m(2-f-2Q+\gamma_1) γ 2 = f ( 3 − b e ) + 2 3 Ω m ( 2 − f − 2 Q + γ 1 ) (e.g. Maartens et al. 2021a), where x x x is the comoving distance, the prime denotes a conformal-time derivative and Ω m \Omega_m Ω m is evaluated at the redshift of the sample. Substituting Eq. (54) expresses them in terms of the count slope, n e v o l n_{\mathrm{evol}} n evol , and Q Q Q , the relativistic nuisance parameters of our forecasts (§ 3 ): for a sample selected at a fixed flux,
γ 1 = − ( 1 + n e v o l + 4 Q ) , (21) \gamma_1 = -(1+n_{\mathrm{evol}}+4Q), \tag*{(21)} γ 1 = − ( 1 + n evol + 4 Q ) , (21)
γ 2 = f [ 4 + n e v o l + 2 Q − 2 ( 1 − Q ) x H − H ′ H 2 ] + 3 2 Ω m ( 1 − f − n e v o l − 6 Q ) . (22) \gamma_2 = f\left[4+n_{\mathrm{evol}}+2Q-\frac{2(1-Q)}{x\mathcal{H}}-\frac{\mathcal{H}'}{\mathcal{H}^2}\right]+\frac{3}{2}\Omega_m(1-f-n_{\mathrm{evol}}-6Q). \tag*{(22)} γ 2 = f [ 4 + n evol + 2 Q − x H 2 ( 1 − Q ) − H 2 H ′ ] + 2 3 Ω m ( 1 − f − n evol − 6 Q ) . (22)
A K-correction, or a bright-faint split at a fixed bright fraction, changes only the relation between n e v o l n_{\mathrm{evol}} n evol and b e b_e b e (Appendices A1 and A3), and so adds terms proportional to Q Q Q to Eq. (18) but leaves the expressions in terms of b e b_e b e unchanged.
The second-order kernel is longer and contains relativistic terms up to fourth order in H / k \mathcal{H}/k H / k , with coefficients that depend on b 1 b_1 b 1 , b e b_e b e , Q Q Q and their derivatives with respect to time and luminosity (Jolicoeur et al. 2018; we use the form of Maartens et al. 2021a). Since the fields are real in configuration space, every kernel satisfies K n ( − q 1 , … , − q n , d ) = K n ∗ ( q 1 , … , q n , d ) \mathcal{K}_n(-\mathbf{q}_1,\ldots,-\mathbf{q}_n,\mathbf{d})=\mathcal{K}_n^*(\mathbf{q}_1,\ldots,\mathbf{q}_n,\mathbf{d}) K n ( − q 1 , … , − q n , d ) = K n ∗ ( q 1 , … , q n , d ) : the Newtonian terms and the relativistic terms of even order in H / k \mathcal{H}/k H / k are real and even in the wavevectors, while those of odd order, such as the Doppler term in Eq. (20) , are imaginary and odd.
Galaxy bias
To connect halo-based quantities to galaxy-based quantities one needs a prescription for how galaxies populate haloes. One way of doing this is to use a Halo Occupation Distribution (HOD) (Cooray & Sheth 2002) to describe the mean number of galaxies per halo, ⟨ N ( M , z ) ⟩ \langle N(M,z)\rangle ⟨ N ( M , z )⟩ . Thus, the galaxy bias values can be obtained by a weighted integral of the halo bias, b i h ( M , z ) b_i^h(M,z) b i h ( M , z ) , above a minimum halo mass, M min M_{\min} M m i n ,
b i ( z ) = ∫ M min ( z ) ∞ d ln M n h ( M , z ) b i h ( M , z ) ⟨ N ( M , z ) ⟩ n ˉ g ( z ) , (23) b_i(z)=\frac{\int_{M_{\min}(z)}^\infty d\ln M\,n_h(M,z)b_i^h(M,z)\langle N(M,z)\rangle}{\bar n_g(z)}, \tag*{(23)} b i ( z ) = n ˉ g ( z ) ∫ M m i n ( z ) ∞ d ln M n h ( M , z ) b i h ( M , z ) ⟨ N ( M , z )⟩ , (23)
where n h ( M , z ) n_h(M,z) n h ( M , z ) is the mean number density of haloes for a given mass and n ˉ g ( z ) \bar n_g(z) n ˉ g ( z ) is the mean galaxy density. The mean galaxy density can be computed from the integral
n ˉ g ( z ) = ∫ M min ( z ) ∞ d ln M n h ( M , z ) ⟨ N ( M , z ) ⟩ . (24) \bar n_g(z)=\int_{M_{\min}(z)}^\infty d\ln M\,n_h(M,z)\langle N(M,z)\rangle. \tag*{(24)} n ˉ g ( z ) = ∫ M m i n ( z ) ∞ d ln M n h ( M , z ) ⟨ N ( M , z )⟩ . (24)
The HOD models we use in our forecasts are outlined in Appendix B.
Forecast Details
This section outlines the details of our data vector, forecast set-up and the specific fiducial parameters that we choose. The full covariance matrix is detailed in § 4 .
Throughout this study, we assume a fiducial cosmology from the Planck results (Planck Collaboration et al. [123 ] ), and in all our forecasts, for both PNG and the local relativistic effects, we marginalise over the cosmological parameters with Planck constraints as priors; in the Fisher forecasts these are Ω m \Omega_m Ω m , h h h , n s n_s n s and ln ( 10 10 A s ) \ln(10^{10}A_s) ln ( 1 0 10 A s ) . We assume a fiducial value of f N L = 0 f_{\mathrm{NL}}=0 f NL = 0 for all PNG shapes and forecast for each PNG shape independently. For multipoles beyond the monopole we assume an endpoint LOS, as enforced by the FFT estimators, specifically d = x 1 \mathbf{d}=\mathbf{x}_1 d = x 1 for the power spectrum and d = x 3 \mathbf{d}=\mathbf{x}_3 d = x 3 for the bispectrum. This choice is relevant for wide-separation type corrections including the integrated effects beyond the Limber approximation (see Paper II). In the monopole, where we are free to choose our LOS as it only denotes the point about which we do our wide-separation expansion, we adopt a midpoint LOS for the power spectrum and a centre-of-mass LOS for the bispectrum. This choice has minimal impact on our forecasts; see Paper I and Paper II for more details.
Bias nuisance parameters
We marginalise over the bias parameters in each redshift bin, namely the linear bias, b 1 b_1 b 1 , and, where we include relativistic effects, the magnification bias, Q Q Q , and the count slope, n e v o l n_{\mathrm{evol}} n evol , which we use in place of the evolution bias, b e b_e b e .6 Whenever the bispectrum is included, we also marginalise over the quadratic and tidal biases, b 2 b_2 b 2 and b Γ 2 b_{\Gamma_2} b Γ 2 , in each redshift bin (and for each sample of a bright-faint split) with no prior; as they enter only the bispectrum, they are not varied in the power spectrum analyses. The fiducial values of b 2 b_2 b 2 come from the HOD (§ 2.3 ) and we use the local Lagrangian relation, b Γ 2 = − 2 7 ( b 1 − 1 ) b_{\Gamma_2}=-\frac{2}{7}(b_1-1) b Γ 2 = − 7 2 ( b 1 − 1 ) , for the tidal bias. The fiducial bias model is described in Appendix B. We consider two limiting treatments of the relativistic biases, either holding them at their fiducial values (‘fixed’), which assumes that they are known exactly, or marginalising over them in each redshift bin with no prior (‘free’), giving 3 and 5 nuisance parameters per bin, respectively, for a single tracer when the bispectrum is included. The two use the same signal and covariance.
Priors on the bias parameters
For the Euclid-like Hα \alpha α survey we also consider Gaussian priors on n e v o l n_{\mathrm{evol}} n evol and Q Q Q forward-modelled from the uncertainty in its luminosity function, following Wang et al. (2020a), as an example of a sample-motivated prior; the method is described in Appendix A4. We discuss how else these biases can be determined, including from the survey catalogue itself, in § 5.2 .
We do not, however, account for uncertainty in the PNG biases, b Ψ b_{\Psi} b Ψ and b Ψ δ b_{\Psi\delta} b Ψ δ , and hold them fixed in our forecasts. Under the assumption of universality of the mass function we can derive expressions for the PNG biases for a given HOD and linear bias for each shape using a peak-background split approach (Slosar et al. 2008; Matarrese & Verde 2008; Dalal et al. 2008; Giannantonio & Porciani 2010; Baldauf et al. 2011). While extensive results from N N N -body simulations have shown PNG biases to depart from their universal mass function expectations (Grossi et al. 2009; Barreira et al. 2020; Lazeyras et al. 2023; Gutiérrez Adame et al. 2024; Fondi et al. 2024), and one could account for this in our forecasts by introducing modifications to the PNG biases along with further nuisance parameters, we leave a discussion of non-universal estimates of the bias parameters for the different PNG shapes to future work. We expect this to mainly affect the uncertainty on the determination of f N L f_{\mathrm{NL}} f NL , but to have less impact on the core focus of this work – the degeneracy of f N L f_{\mathrm{NL}} f NL with light-cone effects (since it does not change the k k k -dependence). Breaking the degeneracy between the PNG biases and f N L f_{\mathrm{NL}} f NL is non-trivial without strict priors on b Ψ b_{\Psi} b Ψ (Barreira et al. 2020; Barreira 2022; Moradinezhad Dizgah et al. 2021; Euclid Collaboration et al. 2026), such as from simulations calibrated to the selected sample (Fondi et al. 2024) or from the redshift evolution of its number counts ([115 ] ), which is related to the evolution bias (Appendix A2); this motivates the inclusion not only of the bispectrum but also of multi-tracer approaches (e.g. Barreira & Krause 2023).
Data vector
Our full data vector, D D D , is a sum over redshift bins7 . The data vector in each bin, D z D_z D z , then contains a given selection of power spectrum and bispectrum multipoles:
D z = { P ℓ i , … , P ℓ n , B ℓ i , … , B ℓ m } . (25) D_z = \{P_{\ell_i}, \ldots, P_{\ell_n}, B_{\ell_i}, \ldots, B_{\ell_m}\}. \tag*{(25)} D z = { P ℓ i , … , P ℓ n , B ℓ i , … , B ℓ m } . (25)
For our single-tracer analysis, we include even multipoles up to the hexadecapole, ℓ = 0 , 2 , 4 \ell= 0, 2, 4 ℓ = 0 , 2 , 4 , for the power spectrum, and up to the quadrupole for the bispectrum, ℓ = 0 , 2 \ell= 0, 2 ℓ = 0 , 2 , and can also include the non-zero bispectrum dipole, ℓ = 1 \ell= 1 ℓ = 1 . In the full multi-tracer analysis, the data vector for each power spectrum multipole consists of three combinations, P ℓ i e v e n = { P ℓ i B B , P ℓ i B F , P ℓ i F F } P_{\ell_i}^{\mathrm{even}} = \{P_{\ell_i}^{BB}, P_{\ell_i}^{BF}, P_{\ell_i}^{FF}\} P ℓ i even = { P ℓ i B B , P ℓ i B F , P ℓ i F F } from the bright, B B B , and faint, F F F , samples. In the case of the bright-faint split for the bispectrum, each multipole consists of four independent combinations, B ℓ i e v e n = { B ℓ i B B B , B ℓ i B B F , B ℓ i B F F , B ℓ i F F F } B_{\ell_i}^{\mathrm{even}} = \{B_{\ell_i}^{BBB}, B_{\ell_i}^{BBF}, B_{\ell_i}^{BFF}, B_{\ell_i}^{FFF}\} B ℓ i even = { B ℓ i B B B , B ℓ i B B F , B ℓ i B F F , B ℓ i F F F } . Additionally, in the multi-tracer case one can now also consider the odd multipoles of the power spectrum, (ℓ = 1 , 3 \ell= 1, 3 ℓ = 1 , 3 ); however, the auto-correlations are still zero and therefore we only include the cross-correlations for the odd multipoles, P ℓ i o d d = P ℓ i B F P_{\ell_i}^{\mathrm{odd}} = P_{\ell_i}^{BF} P ℓ i odd = P ℓ i B F .
Methods
To explore the constraints and biases on the parameters of interest, we either sample the posteriors with a Markov Chain Monte Carlo (MCMC) pipeline or use Fisher forecasts8 . These methods are described in Section 3 of Paper II and in Addis (2026); we summarise the key points below. The MCMCs are implemented in CosmoWAP using COBAYA (Neal 2005; Lewis & Bridle 2002; Lewis 2013; Torrado & Lewis 2021) where we speed up the sampling over cosmological parameters by using CosmoPower (Spurio Mancini et al. 2022) emulators for the linear power spectrum. For all chains the tolerance for convergence is set by the Gelman–Rubin parameter R − 1 = 0.001 R-1 = 0.001 R − 1 = 0.001 .
Multi-tracer Multipole Covariance
While computing covariances from mock catalogues offers many advantages, it presents several key challenges on ultra-large scales. Most notably, the need for large-volume mocks significantly increases
For each multipole, we use k k k -bins of width Δ k = 4 k f \Delta k = 4k_{\mathrm{f}} Δ k = 4 k f , with the largest scale set by the fundamental frequency of the survey, k f = 2 π / V 1 / 3 k_{\mathrm{f}} = 2\pi/V^{1/3} k f = 2 π / V 1/3 , and impose a constant cut-off at k max = 0.15 k_{\max} = 0.15 k m a x = 0.15 [h/Mpc] for the power spectrum and k max = 0.1 k_{\max} = 0.1 k m a x = 0.1 [h/Mpc] for the bispectrum. We also impose an effective k min k_{\min} k m i n by excluding every k k k -bin that contains Fourier modes with k < 2 π / χ ( z min ) k < 2\pi/\chi(z_{\min}) k < 2 π / χ ( z m i n ) , where χ \chi χ is the comoving distance and z min z_{\min} z m i n is the minimum redshift of the redshift bin, as above these scales we expect the perturbative wide-separation expansion to have broken down, which is particularly relevant for low-redshift surveys.
Each estimator averages over the Fourier modes in its bin, so we average the model in the same way rather than evaluating it at the bin centre. For the power spectrum, the multipoles, their derivatives and C P P C_{PP} C P P are averaged over the modes of each k k k -shell. For the bispectrum we follow Oddo et al. (2020) and use the effective triangle of each bin, whose sides are the mean longest, middle and shortest sides of its closed triangles, weighted by their number; as the estimator sums over every labelling of the sides in the same shell, isosceles and equilateral bins are averaged over the corresponding relabellings. The bispectrum, its derivatives and C P P P C^{PPP} C P P P are evaluated on this triangle, with the triangle count of the full bin (Eq. (34) ), while the non-Gaussian covariance terms use the bispectrum averaged over the triangles of each bin (§ 4.3 ). Averaging the bispectrum and its derivatives over the triangles of each bin instead changes our single-tracer Euclid -like constraints on f N L f_{\mathrm{NL}} f NL and on the amplitude of the local relativistic effects by ≲ 1 % \lesssim1\% ≲ 1% .
For simplicity, in the bispectrum we restrict the analysis to tri-angular configurations satisfying k 1 ≥ k 2 ≥ k 3 k_1 \ge k_2 \ge k_3 k 1 ≥ k 2 ≥ k 3 , although beyond the monopole of the bispectrum is no longer symmetric under cyclic permutations of k 1 , k 2 , k 3 k_1, k_2, k_3 k 1 , k 2 , k 3 . computational costs, and incorporating the relativistic effects that become relevant at these scales remains non-trivial. As a result, analytical covariance computations are essential for probing the information content in this large, linear regime. Here, we introduce a novel calculation of the multi-tracer complex bispectrum multipole covariance that includes the odd-parity multipoles, together with its leading non-Gaussian contributions and the power spectrum–bispectrum cross-covariance.
One can define the full covariance for a combined power spectrum and bispectrum analysis as
C = ( C P P C P B C B P C B B ) (26) \mathbf{C} =
\begin{pmatrix}
C_{PP} & C_{PB} \\
C_{BP} & C_{BB}
\end{pmatrix} \tag*{(26)} C = ( C P P C B P C P B C B B ) (26)
This is defined for a single redshift bin and we assume each redshift bin is independent.
The multi-tracer power spectrum covariance including odd-parity multipoles is detailed for the first time in Paper II with the inclusion of integrated effects and wide-separation corrections (see also Montano et al. 2026b for a similar derivation of just the local relativistic terms); here we extend this computation to the bispectrum. As in that work, we take the power spectrum covariance, C P P C_{PP} C P P , to be Gaussian. For the bispectrum, we compute the Gaussian covariance and model the leading non-Gaussian contributions, which also correlate the power spectrum and bispectrum through C P B C_{PB} C P B , in a squeezed-limit approximation (Section 4.2). We neglect wide-separation corrections here, though they are computed for the power spectrum covariance in Paper II. Unlike Paper II, we also neglect integrated relativistic contributions to the covariance, as these first require a treatment of the radial window (Section 4.1.1). We also neglect super-sample covariance from modes larger than the survey (Chan et al. 2018).
Multi-tracer bispectrum multipole covariance
For a given redshift bin, we can express the bispectrum-bispectrum covariance matrix as a block matrix for each multipole pair
C B B = ( C [ B ℓ i , B ℓ i ] ⋯ C [ B ℓ i , B ℓ n ] ⋮ ⋱ ⋮ C [ B ℓ n , B ℓ i ] ⋯ C [ B ℓ n , B ℓ n ] ) (27) C_{BB} =
\begin{pmatrix}
C[B_{\ell_i},B_{\ell_i}] & \cdots& C[B_{\ell_i},B_{\ell_n}] \\
\vdots& \ddots& \vdots\\
C[B_{\ell_n},B_{\ell_i}] & \cdots& C[B_{\ell_n},B_{\ell_n}]
\end{pmatrix} \tag*{(27)} C B B = C [ B ℓ i , B ℓ i ] ⋮ C [ B ℓ n , B ℓ i ] ⋯ ⋱ ⋯ C [ B ℓ i , B ℓ n ] ⋮ C [ B ℓ n , B ℓ n ] (27)
where each cross/auto-multipole covariance block is a 4 × 4 4 \times4 4 × 4 sub-matrix. Writing C a b c , d e f ≡ C [ B a b c , B d e f ] C^{abc,def} \equiv C[B^{abc},B^{def}] C ab c , d e f ≡ C [ B ab c , B d e f ] for brevity, for two distinct populations, X X X and Y Y Y , the sub-matrix can be written as
C ( B i , B j ) = ( C X X X , X X X C X X X , X X Y C X X X , X Y Y C X X X , Y Y Y C X X Y , X X X C X X Y , X X Y C X X Y , X Y Y C X X Y , Y Y Y C X Y Y , X X X C X Y Y , X X Y C X Y Y , X Y Y C X Y Y , Y Y Y C Y Y Y , X X X C Y Y Y , X X Y C Y Y Y , X Y Y C Y Y Y , Y Y Y ) (28) C(B_i,B_j) =
\begin{pmatrix}
C^{XXX,XXX} & C^{XXX,XXY} & C^{XXX,XYY} & C^{XXX,YYY} \\
C^{XXY,XXX} & C^{XXY,XXY} & C^{XXY,XYY} & C^{XXY,YYY} \\
C^{XYY,XXX} & C^{XYY,XXY} & C^{XYY,XYY} & C^{XYY,YYY} \\
C^{YYY,XXX} & C^{YYY,XXY} & C^{YYY,XYY} & C^{YYY,YYY}
\end{pmatrix} \tag*{(28)} C ( B i , B j ) = C X X X , X X X C X X Y , X X X C X Y Y , X X X C Y Y Y , X X X C X X X , X X Y C X X Y , X X Y C X Y Y , X X Y C Y Y Y , X X Y C X X X , X Y Y C X X Y , X Y Y C X Y Y , X Y Y C Y Y Y , X Y Y C X X X , Y Y Y C X X Y , Y Y Y C X Y Y , Y Y Y C Y Y Y , Y Y Y (28)
where the tracers are listed in the order of the sides, k 1 ≥ k 2 ≥ k 3 k_1 \ge k_2 \ge k_3 k 1 ≥ k 2 ≥ k 3 , so that the second tracer occupies the shortest sides.
This two-tracer expression is equivalent to the covariance matrices in Karagiannis et al. (2024). The details of the calculation for a given covariance matrix element are given in the next section.
Covariance matrix element
In this section we compute the covariance of the multi-tracer bispectrum multipoles for the Scoccimarro estimator. The covariance matrix
matrix is defined for complex variables as
C [ B ℓ i a b c , B ℓ j d e f ] ( k , q ) = ⟨ B ^ ℓ i a b c ( k 1 , k 2 , k 3 ) B ^ ℓ j d e f ( q 1 , q 2 , q 3 ) ∗ ⟩ − ⟨ B ^ ℓ i a b c ( k 1 , k 2 , k 3 ) ⟩ ⟨ B ^ ℓ j d e f ( q 1 , q 2 , q 3 ) ∗ ⟩ . (29) C[B_{\ell_i}^{abc},B_{\ell_j}^{def}](\mathbf{k},\mathbf{q}) =
\left\langle\hat{B}_{\ell_i}^{abc}(\mathbf{k}_1,\mathbf{k}_2,\mathbf{k}_3)\hat{B}_{\ell_j}^{def}(\mathbf{q}_1,\mathbf{q}_2,\mathbf{q}_3)^* \right\rangle
-
\left\langle\hat{B}_{\ell_i}^{abc}(\mathbf{k}_1,\mathbf{k}_2,\mathbf{k}_3)\right\rangle
\left\langle\hat{B}_{\ell_j}^{def}(\mathbf{q}_1,\mathbf{q}_2,\mathbf{q}_3)^* \right\rangle. \tag*{(29)} C [ B ℓ i ab c , B ℓ j d e f ] ( k , q ) = ⟨ B ^ ℓ i ab c ( k 1 , k 2 , k 3 ) B ^ ℓ j d e f ( q 1 , q 2 , q 3 ) ∗ ⟩ − ⟨ B ^ ℓ i ab c ( k 1 , k 2 , k 3 ) ⟩ ⟨ B ^ ℓ j d e f ( q 1 , q 2 , q 3 ) ∗ ⟩ . (29)
For convenience, we define the function,
F ℓ i ( k ) = ∫ x e − i k ⋅ x L ℓ i ( k ^ ⋅ x ^ ) Δ ( x ) , (30) F_{\ell_i}(\mathbf{k}) = \int_{\mathbf{x}} e^{-i\mathbf{k}\cdot\mathbf{x}}\mathcal{L}_{\ell_i}(\hat{\mathbf{k}}\cdot\hat{\mathbf{x}})\Delta(\mathbf{x}), \tag*{(30)} F ℓ i ( k ) = ∫ x e − i k ⋅ x L ℓ i ( k ^ ⋅ x ^ ) Δ ( x ) , (30)
as well as the shortened integral notation
∫ q ^ 1 = ∫ S 1 d 3 q 1 (31) \int_{\hat{q}_1} = \int_{S_1} \mathrm{d}^3q_1 \tag*{(31)} ∫ q ^ 1 = ∫ S 1 d 3 q 1 (31)
such that, using k f 3 = ( 2 π ) 3 / V k_f^3=(2\pi)^3/V k f 3 = ( 2 π ) 3 / V to write the estimator normalisation as 1 / ( V N 123 T k f 6 ) 1/(VN_{123}^T k_f^6) 1/ ( V N 123 T k f 6 ) , the bispectrum-bispectrum multipole matrix element can be expressed, ignoring weights and windows, in terms of a 6-point function
C [ B ℓ i a b c , B ℓ j d e f ] ( k , q ) = ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) V 2 k f 6 N 123 T N q T ∫ k ^ 1 , k ^ 2 , k ^ 3 , q ^ 1 , q ^ 2 , q ^ 3 δ D ( k 123 ) δ D ( q 123 ) [ ⟨ F ℓ i a ( k 1 ) F 0 b ( k 2 ) F 0 c ( k 3 ) F ℓ j d ( q 1 ) ∗ F 0 e ( q 2 ) ∗ F 0 f ( q 3 ) ∗ ⟩ − ⟨ F ℓ i a ( k 1 ) F 0 b ( k 2 ) F 0 c ( k 3 ) ⟩ ⟨ F ℓ j d ( q 1 ) ∗ F 0 e ( q 2 ) ∗ F 0 f ( q 3 ) ∗ ⟩ ] . (32) C[B_{\ell_i}^{abc},B_{\ell_j}^{def}](\mathbf{k},\mathbf{q}) =
\frac{(2\ell_i+1)(2\ell_j+1)}{V^2 k_f^6 N_{123}^T N_q^T}
\int_{\hat{k}_1,\hat{k}_2,\hat{k}_3,\hat{q}_1,\hat{q}_2,\hat{q}_3}
\delta_D(k_{123})\delta_D(q_{123})
\left[
\left\langle F_{\ell_i}^a(\mathbf{k}_1)F_0^b(\mathbf{k}_2)F_0^c(\mathbf{k}_3)F_{\ell_j}^d(\mathbf{q}_1)^*F_0^e(\mathbf{q}_2)^*F_0^f(\mathbf{q}_3)^* \right\rangle
-
\left\langle F_{\ell_i}^a(\mathbf{k}_1)F_0^b(\mathbf{k}_2)F_0^c(\mathbf{k}_3)\right\rangle
\left\langle F_{\ell_j}^d(\mathbf{q}_1)^*F_0^e(\mathbf{q}_2)^*F_0^f(\mathbf{q}_3)^*\right\rangle
\right]. \tag*{(32)} C [ B ℓ i ab c , B ℓ j d e f ] ( k , q ) = V 2 k f 6 N 123 T N q T ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) ∫ k ^ 1 , k ^ 2 , k ^ 3 , q ^ 1 , q ^ 2 , q ^ 3 δ D ( k 123 ) δ D ( q 123 ) [ ⟨ F ℓ i a ( k 1 ) F 0 b ( k 2 ) F 0 c ( k 3 ) F ℓ j d ( q 1 ) ∗ F 0 e ( q 2 ) ∗ F 0 f ( q 3 ) ∗ ⟩ − ⟨ F ℓ i a ( k 1 ) F 0 b ( k 2 ) F 0 c ( k 3 ) ⟩ ⟨ F ℓ j d ( q 1 ) ∗ F 0 e ( q 2 ) ∗ F 0 f ( q 3 ) ∗ ⟩ ] . (32)
where N 123 T N_{123}^T N 123 T and N q T N_q^T N q T are the numbers of triangles in the two bins. In the continuum limit, Δ k ≪ k \Delta k \ll k Δ k ≪ k , we can re-express the integrals over spherical shells as angular integrals
1 N 123 T k f 6 ∫ q ^ 1 , q ^ 2 , q ^ 3 δ D ( q 123 ) = ∫ d Ω 4 π , (33) \frac{1}{N_{123}^T k_f^6}\int_{\hat{q}_1,\hat{q}_2,\hat{q}_3}\delta_D(q_{123}) = \int\frac{\mathrm{d}\Omega}{4\pi}, \tag*{(33)} N 123 T k f 6 1 ∫ q ^ 1 , q ^ 2 , q ^ 3 δ D ( q 123 ) = ∫ 4 π d Ω , (33)
where N 123 T = 8 π 2 k 1 k 2 k 3 ( Δ k ) 3 β / k f 6 N_{123}^T = 8\pi^2 k_1 k_2 k_3(\Delta k)^3\beta/k_f^6 N 123 T = 8 π 2 k 1 k 2 k 3 ( Δ k ) 3 β / k f 6 is the number of fundamental triangles in the bin. Here β \beta β is the fraction of the bin S 1 × S 2 × S 3 S_1 \times S_2 \times S_3 S 1 × S 2 × S 3 satisfying the closure condition q 1 ≤ q 2 + q 3 q_1 \le q_2+q_3 q 1 ≤ q 2 + q 3 . Writing q i = k i + Δ k δ i q_i=k_i+\Delta k\,\delta_i q i = k i + Δ k δ i , so that δ i ∈ [ − 1 / 2 , 1 / 2 ] \delta_i\in[-1/2,1/2] δ i ∈ [ − 1/2 , 1/2 ] spans the S i \mathit{S}_i S i -th shell, this is
β = 1 k 1 k 2 k 3 ∫ − 1 / 2 1 / 2 d δ 1 d δ 2 d δ 3 ∏ i = 1 3 ( k i + Δ k δ i ) × Θ ( u − δ 1 + δ 2 + δ 3 ) , (34) \beta=
\frac{1}{k_1k_2k_3}
\int_{-1/2}^{1/2}\mathrm{d}\delta_1\mathrm{d}\delta_2\mathrm{d}\delta_3
\prod_{i=1}^{3}(k_i+\Delta k\,\delta_i)
\times\Theta(u-\delta_1+\delta_2+\delta_3), \tag*{(34)} β = k 1 k 2 k 3 1 ∫ − 1/2 1/2 d δ 1 d δ 2 d δ 3 i = 1 ∏ 3 ( k i + Δ k δ i ) × Θ ( u − δ 1 + δ 2 + δ 3 ) , (34)
where Θ \Theta Θ is the Heaviside step function and u ≡ ( k 2 + k 3 − k 1 ) / Δ k u\equiv(k_2+k_3-k_1)/\Delta k u ≡ ( k 2 + k 3 − k 1 ) /Δ k measures the distance from the closure boundary in units of the bin width. The radial integrals contribute ∫ S i q i d q i = k i Δ k \int_{S_i}q_i\mathrm{d}q_i=k_i\Delta k ∫ S i q i d q i = k i Δ k exactly, so the prefactor in N 123 T N_{123}^T N 123 T receives no further correction in Δ k \Delta k Δ k and the entire bin-width dependence resides in β \beta β . Since ∣ δ 1 − δ 2 − δ 3 ∣ ≤ 3 / 2 \lvert\delta_1-\delta_2-\delta_3\rvert\le3/2 ∣ δ 1 − δ 2 − δ 3 ∣ ≤ 3/2 , the step function is unity across the whole bin for u ≥ 3 / 2 u\ge3/2 u ≥ 3/2 and vanishes across it for u ≤ − 3 / 2 u\le-3/2 u ≤ − 3/2 , giving β = 1 \beta=1 β = 1 and β = 0 \beta=0 β = 0 respectively; only bins straddling the closure boundary need be evaluated. In the thin-bin limit, Δ k / k → 0 \Delta k/k\to0 Δ k / k → 0 , Eq. (34) recovers the familiar β = 1 / 2 \beta=1/2 β = 1/2 for folded triangles, k 1 = k 2 + k 3 k_1=k_2+k_3 k 1 = k 2 + k 3 . At the bin width adopted here, Δ k = 4 k f \Delta k=4k_f Δ k = 4 k f , these configurations instead have 0.53 ≲ β ≲ 0.60 0.53\lesssim\beta\lesssim0.60 0.53 ≲ β ≲ 0.60 , and we therefore use Eq. (34) throughout.
The 6-point function in Eq. (32) decomposes into products of connected correlators: three power spectra (P P PP P P ), two bispectra (B B BB B B ), a power spectrum and a trispectrum (P T PT P T ), and the connected 6-point function. Of the ten ways of splitting the six fields into two triplets, the one that separates the two estimators cancels the disconnected product in Eq. (32) , and any power spectrum of two sides of same triangle vanishes, as k 1 + k 2 = 0 \boldsymbol{k}_1+\boldsymbol{k}_2=\mathbf{0} k 1 + k 2 = 0 would require k 3 = 0 \boldsymbol{k}_3=\mathbf{0} k 3 = 0 . Every remaining term therefore contracts fields of both triangles, and neglecting the connected 6-point function, which is subleading on the scales of interest ((Biagetti et al. 2022; Salvalaggio et al. 2024)), the covariance is
C = C P P P + C B B + C P T , (35) C=C^{PPP}+C^{BB}+C^{PT}, \tag*{(35)} C = C P P P + C B B + C P T , (35)
with six, nine and nine terms, respectively. The Gaussian term, C P P P C^{PPP} C P P P , pairs every side of one triangle with a side of the other, so it is non-zero only when the two bins coincide, N k T ≡ N q T = N 123 T N_k^T\equiv N_q^T=N_{123}^T N k T ≡ N q T = N 123 T , whereas C B B C^{BB} C B B and C P T C^{PT} C P T link the two triangles through a single shared mode; we compute the former here and return to the latter in § 4.3 . With the assumption that we have non-overlapping samples ((Smith 2009)), the Gaussian term can be written as
C P P P [ B ^ ℓ i a b c , B ^ ℓ j d e f ] ( k , q ) = ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) ( V N 123 T k f 6 ) ∫ k ^ 1 , k ^ 2 , k ^ 3 , q ^ 1 , q ^ 2 , q ^ 3 × δ D ( k 123 ) δ D ( q 123 ) × { ⟨ F ℓ i a ( k 1 ) F ℓ j d ( q 1 ) ∗ ⟩ [ ⟨ F 0 b ( k 2 ) F 0 e ( q 2 ) ∗ ⟩ ⟨ F 0 c ( k 3 ) F 0 f ( q 3 ) ∗ ⟩ + ⟨ F 0 b ( k 2 ) F 0 f ( q 3 ) ∗ ⟩ ⟨ F 0 c ( k 3 ) F 0 e ( q 2 ) ∗ ⟩ ] + ⟨ F ℓ i a ( k 1 ) F 0 e ( q 2 ) ∗ ⟩ [ ⟨ F 0 b ( k 2 ) F ℓ j d ( q 1 ) ∗ ⟩ ⟨ F 0 c ( k 3 ) F 0 f ( q 3 ) ∗ ⟩ + ⟨ F 0 b ( k 2 ) F 0 f ( q 3 ) ∗ ⟩ ⟨ F 0 c ( k 3 ) F ℓ j d ( q 1 ) ∗ ⟩ ] + ⟨ F ℓ i a ( k 1 ) F 0 f ( q 3 ) ∗ ⟩ [ ⟨ F 0 b ( k 2 ) F 0 e ( q 2 ) ∗ ⟩ ⟨ F 0 c ( k 3 ) F ℓ j d ( q 1 ) ∗ ⟩ + ⟨ F 0 b ( k 2 ) F ℓ j d ( q 1 ) ∗ ⟩ ⟨ F 0 c ( k 3 ) F 0 e ( q 2 ) ∗ ⟩ ] } . (36) \begin{aligned}
C^{PPP}[\hat{B}_{\ell_i}^{abc},\hat{B}_{\ell_j}^{def}](\boldsymbol{k},\boldsymbol{q})={}&\frac{(2\ell_i+1)(2\ell_j+1)}{(V N_{123}^T k_f^6)}\int_{\hat{k}_1,\hat{k}_2,\hat{k}_3,\hat{q}_1,\hat{q}_2,\hat{q}_3}\\
&\times\delta_D(\boldsymbol{k}_{123})\delta_D(\boldsymbol{q}_{123})\\
&\times\Bigl\{\left\langle F_{\ell_i}^{a}(\boldsymbol{k}_1)F_{\ell_j}^{d}(\boldsymbol{q}_1)^*\right\rangle\left[\left\langle F_0^b(\boldsymbol{k}_2)F_0^e(\boldsymbol{q}_2)^*\right\rangle\left\langle F_0^c(\boldsymbol{k}_3)F_0^f(\boldsymbol{q}_3)^*\right\rangle\right.\\
&\qquad\left.+\left\langle F_0^b(\boldsymbol{k}_2)F_0^f(\boldsymbol{q}_3)^*\right\rangle\left\langle F_0^c(\boldsymbol{k}_3)F_0^e(\boldsymbol{q}_2)^*\right\rangle\right]\\
&+\left\langle F_{\ell_i}^{a}(\boldsymbol{k}_1)F_0^e(\boldsymbol{q}_2)^*\right\rangle\left[\left\langle F_0^b(\boldsymbol{k}_2)F_{\ell_j}^{d}(\boldsymbol{q}_1)^*\right\rangle\left\langle F_0^c(\boldsymbol{k}_3)F_0^f(\boldsymbol{q}_3)^*\right\rangle\right.\\
&\qquad\left.+\left\langle F_0^b(\boldsymbol{k}_2)F_0^f(\boldsymbol{q}_3)^*\right\rangle\left\langle F_0^c(\boldsymbol{k}_3)F_{\ell_j}^{d}(\boldsymbol{q}_1)^*\right\rangle\right]\\
&+\left\langle F_{\ell_i}^{a}(\boldsymbol{k}_1)F_0^f(\boldsymbol{q}_3)^*\right\rangle\left[\left\langle F_0^b(\boldsymbol{k}_2)F_0^e(\boldsymbol{q}_2)^*\right\rangle\left\langle F_0^c(\boldsymbol{k}_3)F_{\ell_j}^{d}(\boldsymbol{q}_1)^*\right\rangle\right.\\
&\qquad\left.+\left\langle F_0^b(\boldsymbol{k}_2)F_{\ell_j}^{d}(\boldsymbol{q}_1)^*\right\rangle\left\langle F_0^c(\boldsymbol{k}_3)F_0^e(\boldsymbol{q}_2)^*\right\rangle\right]\Bigr\}. \tag*{(36)}
\end{aligned} C P P P [ B ^ ℓ i ab c , B ^ ℓ j d e f ] ( k , q ) = ( V N 123 T k f 6 ) ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) ∫ k ^ 1 , k ^ 2 , k ^ 3 , q ^ 1 , q ^ 2 , q ^ 3 × δ D ( k 123 ) δ D ( q 123 ) × { ⟨ F ℓ i a ( k 1 ) F ℓ j d ( q 1 ) ∗ ⟩ [ ⟨ F 0 b ( k 2 ) F 0 e ( q 2 ) ∗ ⟩ ⟨ F 0 c ( k 3 ) F 0 f ( q 3 ) ∗ ⟩ + ⟨ F 0 b ( k 2 ) F 0 f ( q 3 ) ∗ ⟩ ⟨ F 0 c ( k 3 ) F 0 e ( q 2 ) ∗ ⟩ ] + ⟨ F ℓ i a ( k 1 ) F 0 e ( q 2 ) ∗ ⟩ [ ⟨ F 0 b ( k 2 ) F ℓ j d ( q 1 ) ∗ ⟩ ⟨ F 0 c ( k 3 ) F 0 f ( q 3 ) ∗ ⟩ + ⟨ F 0 b ( k 2 ) F 0 f ( q 3 ) ∗ ⟩ ⟨ F 0 c ( k 3 ) F ℓ j d ( q 1 ) ∗ ⟩ ] + ⟨ F ℓ i a ( k 1 ) F 0 f ( q 3 ) ∗ ⟩ [ ⟨ F 0 b ( k 2 ) F 0 e ( q 2 ) ∗ ⟩ ⟨ F 0 c ( k 3 ) F ℓ j d ( q 1 ) ∗ ⟩ + ⟨ F 0 b ( k 2 ) F ℓ j d ( q 1 ) ∗ ⟩ ⟨ F 0 c ( k 3 ) F 0 e ( q 2 ) ∗ ⟩ ] } . (36)
In the next section, we restrict the calculation to the zeroth-order term in the wide-separation expansion (equivalent to assuming the plane-parallel constant-redshift limit) for simplicity, as one would need to define wide-separation corrections on our local 6-point function (see Paper II for details of this calculation in the power spectrum multipole covariance). We also neglect integrated effects in the covariance. In principle they can be included in the Gaussian term, provided they are defined with respect to the LOS of the full local 6-point function (see Paper II for relevant discussion of the 4-point function in the power spectrum covariance). However, without a radial window function the lensing contribution stays correlated along the LOS out to the observer, so in P ( k ) P(\boldsymbol{k}) P ( k ) it is concentrated in a narrow spike at μ ≃ 0 \mu\simeq0 μ ≃ 0 . This spike integrates to a finite contribution to the multipoles of the signal, but the covariance, as a product of power spectra, strongly overweights it. In practice, the redshift bin limits the LOS separations that contribute, which spreads this spike over a width set by the inverse radial depth of the bin. We therefore leave integrated contributions to the covariance to future work, alongside the window convolution.
Gaussian plane-parallel constant-redshift covariance
At zeroth order in the wide-separation expansion, as in the power spectrum case, we can remove the Legendre polynomial from the integral
F ℓ i a ( k ) = L ℓ i ( k ⋅ x ^ ^ ) F a ( k ) , (37) F_{\ell_i}^{a}(\boldsymbol{k})=\mathcal{L}_{\ell_i}(\hat{\boldsymbol{k}\cdot\hat{\boldsymbol{x}}})F^{a}(\boldsymbol{k}), \tag*{(37)} F ℓ i a ( k ) = L ℓ i ( k ⋅ x ^ ^ ) F a ( k ) , (37)
such that the covariance matrix element can be expressed in terms of power spectrum contributions,
C P P P [ B ^ ℓ i a b c , B ^ ℓ j d e f ] ( k , q ) = ( 2 π ) 9 ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) ( V N 123 T k f 6 ) 2 × ∫ k ^ 1 , k ^ 2 , k ^ 3 , q ^ 1 , q ^ 2 , q ^ 3 δ D ( k 123 ) δ D ( q 123 ) × L ℓ i ( k ^ 1 ⋅ d ^ ) L ℓ j ( q ^ 1 ⋅ d ^ ) [ P ^ a d ( k 1 , μ 1 ) δ D ( k 1 − q 1 ) × [ P ^ b e ( k 2 , μ 2 ) P ^ c f ( k 3 , μ 3 ) δ D ( k 2 − q 2 ) δ D ( k 3 − q 3 ) + P ^ b f ( k 2 , μ 2 ) P ^ c e ( k 3 , μ 3 ) δ D ( k 2 − q 3 ) δ D ( k 3 − q 2 ) ] + P ^ a e ( k 1 , μ 1 ) δ D ( k 1 − q 2 ) × [ P ^ b d ( k 2 , μ 2 ) P ^ c f ( k 3 , μ 3 ) δ D ( k 2 − q 1 ) δ D ( k 3 − q 3 ) + P ^ b f ( k 2 , μ 2 ) P ^ c d ( k 3 , μ 3 ) δ D ( k 2 − q 3 ) δ D ( k 3 − q 1 ) ] + P ^ a f ( k 1 , μ 1 ) δ D ( k 1 − q 3 ) × [ P ^ b e ( k 2 , μ 2 ) P ^ c d ( k 3 , μ 3 ) δ D ( k 2 − q 2 ) δ D ( k 3 − q 1 ) + P ^ b d ( k 2 , μ 2 ) P ^ c e ( k 3 , μ 3 ) δ D ( k 2 − q 1 ) δ D ( k 3 − q 2 ) ] } , (38) \begin{aligned}
C^{PPP}[\hat{B}_{\ell_i}^{abc},\hat{B}_{\ell_j}^{def}](\boldsymbol{k},\boldsymbol{q})={}&\frac{(2\pi)^9(2\ell_i+1)(2\ell_j+1)}{(V N_{123}^T k_f^6)^2}\\
&\times\int_{\hat{k}_1,\hat{k}_2,\hat{k}_3,\hat{q}_1,\hat{q}_2,\hat{q}_3}\delta_D(\boldsymbol{k}_{123})\delta_D(\boldsymbol{q}_{123})\\
&\times\mathcal{L}_{\ell_i}(\hat{\boldsymbol{k}}_1\cdot\hat{\boldsymbol{d}})\mathcal{L}_{\ell_j}(\hat{\boldsymbol{q}}_1\cdot\hat{\boldsymbol{d}})\Bigl[\hat{P}^{ad}(k_1,\mu_1)\delta_D(\boldsymbol{k}_1-\boldsymbol{q}_1)\\
&\qquad\times\bigl[\hat{P}^{be}(k_2,\mu_2)\hat{P}^{cf}(k_3,\mu_3)\delta_D(\boldsymbol{k}_2-\boldsymbol{q}_2)\delta_D(\boldsymbol{k}_3-\boldsymbol{q}_3)\\
&\qquad\qquad+\hat{P}^{bf}(k_2,\mu_2)\hat{P}^{ce}(k_3,\mu_3)\delta_D(\boldsymbol{k}_2-\boldsymbol{q}_3)\delta_D(\boldsymbol{k}_3-\boldsymbol{q}_2)\bigr]\\
&+\hat{P}^{ae}(k_1,\mu_1)\delta_D(\boldsymbol{k}_1-\boldsymbol{q}_2)\\
&\qquad\times\bigl[\hat{P}^{bd}(k_2,\mu_2)\hat{P}^{cf}(k_3,\mu_3)\delta_D(\boldsymbol{k}_2-\boldsymbol{q}_1)\delta_D(\boldsymbol{k}_3-\boldsymbol{q}_3)\\
&\qquad\qquad+\hat{P}^{bf}(k_2,\mu_2)\hat{P}^{cd}(k_3,\mu_3)\delta_D(\boldsymbol{k}_2-\boldsymbol{q}_3)\delta_D(\boldsymbol{k}_3-\boldsymbol{q}_1)\bigr]\\
&+\hat{P}^{af}(k_1,\mu_1)\delta_D(\boldsymbol{k}_1-\boldsymbol{q}_3)\\
&\qquad\times\bigl[\hat{P}^{be}(k_2,\mu_2)\hat{P}^{cd}(k_3,\mu_3)\delta_D(\boldsymbol{k}_2-\boldsymbol{q}_2)\delta_D(\boldsymbol{k}_3-\boldsymbol{q}_1)\\
&\qquad\qquad+\hat{P}^{bd}(k_2,\mu_2)\hat{P}^{ce}(k_3,\mu_3)\delta_D(\boldsymbol{k}_2-\boldsymbol{q}_1)\delta_D(\boldsymbol{k}_3-\boldsymbol{q}_2)\bigr]\Bigr\}, \tag*{(38)}
\end{aligned} C P P P [ B ^ ℓ i ab c , B ^ ℓ j d e f ] ( k , q ) = ( V N 123 T k f 6 ) 2 ( 2 π ) 9 ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) × ∫ k ^ 1 , k ^ 2 , k ^ 3 , q ^ 1 , q ^ 2 , q ^ 3 δ D ( k 123 ) δ D ( q 123 ) × L ℓ i ( k ^ 1 ⋅ d ^ ) L ℓ j ( q ^ 1 ⋅ d ^ ) [ P ^ a d ( k 1 , μ 1 ) δ D ( k 1 − q 1 ) × [ P ^ b e ( k 2 , μ 2 ) P ^ c f ( k 3 , μ 3 ) δ D ( k 2 − q 2 ) δ D ( k 3 − q 3 ) + P ^ b f ( k 2 , μ 2 ) P ^ ce ( k 3 , μ 3 ) δ D ( k 2 − q 3 ) δ D ( k 3 − q 2 ) ] + P ^ a e ( k 1 , μ 1 ) δ D ( k 1 − q 2 ) × [ P ^ b d ( k 2 , μ 2 ) P ^ c f ( k 3 , μ 3 ) δ D ( k 2 − q 1 ) δ D ( k 3 − q 3 ) + P ^ b f ( k 2 , μ 2 ) P ^ c d ( k 3 , μ 3 ) δ D ( k 2 − q 3 ) δ D ( k 3 − q 1 ) ] + P ^ a f ( k 1 , μ 1 ) δ D ( k 1 − q 3 ) × [ P ^ b e ( k 2 , μ 2 ) P ^ c d ( k 3 , μ 3 ) δ D ( k 2 − q 2 ) δ D ( k 3 − q 1 ) + P ^ b d ( k 2 , μ 2 ) P ^ ce ( k 3 , μ 3 ) δ D ( k 2 − q 1 ) δ D ( k 3 − q 2 ) ] } , (38)
where each Wick contraction gives ⟨ F a ( k i ) F b ( q j ) ∗ ⟩ = ( 2 π ) 3 δ D ( k i − q j ) P ^ a b ( k i ) \langle F^a(\boldsymbol{k}_i)F^b(\boldsymbol{q}_j)^*\rangle=(2\pi)^3\delta_D(\boldsymbol{k}_i-\boldsymbol{q}_j)\hat{P}^{ab}(k_i) ⟨ F a ( k i ) F b ( q j ) ∗ ⟩ = ( 2 π ) 3 δ D ( k i − q j ) P ^ ab ( k i ) , and the power spectrum includes the shot-noise contribution in the auto-correlation case, assuming we have non-overlapping samples,
P ^ a b ( k , d ) = P l o c a b ( k , d ) + δ a , b K n a , (39) \hat{P}^{ab}(k,\boldsymbol{d})=P_{\mathrm{loc}}^{ab}(k,\boldsymbol{d})+\frac{\delta_{a,b}^{K}}{n_a}, \tag*{(39)} P ^ ab ( k , d ) = P loc ab ( k , d ) + n a δ a , b K , (39)
where n a n_a n a is the number density of our given sample. P a b ( k , d ) P^{ab}(k,\boldsymbol{d}) P ab ( k , d ) can in principle contain all relevant contributions to the power spectrum, but here we include only the local relativistic corrections (§ 4.1.1 ). Also, note the relative sign change compared to the standard bispectrum covariance literature ((Ivanov et al. 2023; Rizzo et al. 2023; Salvalaggio et al. 2024)), which arises as this is the covariance of a complex variable.
For the Gaussian part of the covariance distinct triangle bins are uncorrelated. Since the two triangles lie in the same bin, a term of Eq. (38) survives only if it pairs each side q i q_i q i with a side of k k k in the same shell, so each surviving term carries the tracers of the second triangle, ( d , e , f ) (d,e,f) ( d , e , f ) , in some order ( d ′ , e ′ , f ′ ) (d',e',f') ( d ′ , e ′ , f ′ ) on the sides k 1 k_1 k 1 , k 2 k_2 k 2 and k 3 k_3 k 3 . Integrating over the q i q_i q i then sets q 1 q_1 q 1 to the side carrying d d d , which moves the Legendre polynomial of the second triangle onto that side, and leaves δ D ( k 123 ) 2 = δ D ( 0 ) δ D ( k 123 ) \delta_D(k_{123})^2=\delta_D(0)\delta_D(k_{123}) δ D ( k 123 ) 2 = δ D ( 0 ) δ D ( k 123 ) , with δ D ( 0 ) = V / ( 2 π ) 3 \delta_D(0)=V/(2\pi)^3 δ D ( 0 ) = V / ( 2 π ) 3 . Using the continuum limit, Eq. (33) , and k f 3 = ( 2 π ) 3 / V k_f^3=(2\pi)^3/V k f 3 = ( 2 π ) 3 / V , the covariance then reduces to an angular integral,
C P P P [ B ^ ℓ i a b c , B ^ ℓ j d e f ] ( k ) = ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) V N 123 T × ∑ ( d ′ , e ′ , f ′ ) ∫ d Ω 4 π L ℓ i ( μ 1 ) L ℓ j ( μ d ) × P ^ a d ′ ( k 1 , μ 1 ) P ^ b e ′ ( k 2 , μ 2 ) P ^ c f ′ ( k 3 , μ 3 ) , (40) \begin{aligned}
C^{PPP}[\hat{B}_{\ell_i}^{abc},\hat{B}_{\ell_j}^{def}](\boldsymbol{k})={}&\frac{(2\ell_i+1)(2\ell_j+1)V}{N_{123}^T}\\
&\times\sum_{(d',e',f')}\int\frac{\mathrm{d}\Omega}{4\pi}\mathcal{L}_{\ell_i}(\mu_1)\mathcal{L}_{\ell_j}(\mu_d)\\
&\times\hat{P}^{ad'}(k_1,\mu_1)\hat{P}^{be'}(k_2,\mu_2)\hat{P}^{cf'}(k_3,\mu_3), \tag*{(40)}
\end{aligned} C P P P [ B ^ ℓ i ab c , B ^ ℓ j d e f ] ( k ) = N 123 T ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) V × ( d ′ , e ′ , f ′ ) ∑ ∫ 4 π d Ω L ℓ i ( μ 1 ) L ℓ j ( μ d ) × P ^ a d ′ ( k 1 , μ 1 ) P ^ b e ′ ( k 2 , μ 2 ) P ^ c f ′ ( k 3 , μ 3 ) , (40)
where the sum runs over the orderings of ( d , e , f ) (d,e,f) ( d , e , f ) that only exchange tracers between sides in the same shell, and μ d \mu_d μ d is the cosine of the side carrying d d d . For scalene triangles only ( d , e , f ) (d,e,f) ( d , e , f ) survives; k 2 = k 3 k_2=k_3 k 2 = k 3 adds six. For a single tracer the ( d , f , e ) (d,f,e) ( d , f , e ) term equals the first, and for ℓ i = ℓ j = 0 \ell_i=\ell_j=0 ℓ i = ℓ j = 0 every term does, recovering the triangle counting factor s B = 1 , 2 , 6 s_B=1,2,6 s B = 1 , 2 , 6 . The ( e , d , f ) (e,d,f) ( e , d , f ) term, however, moves the Legendre weight to μ 2 \mu_2 μ 2 , so the higher multipoles of equilateral and k 1 = k 2 k_1=k_2 k 1 = k 2 triangles are not simply rescaled by s B s_B s B (Rizzo et al. 2023; Salvalaggio et al. 2024).
The Gaussian part of the multi-tracer bispectrum-bispectrum covariance, C B B C_{BB} C B B , is therefore block-diagonal in triangle bins, with each block constructed of 4 × 4 4\times4 4 × 4 sub-matrices, see Eq. (28) , for each multipole pair. This is shown for ℓ = 0 , 1 , 2 \ell=0,1,2 ℓ = 0 , 1 , 2 in Figure 1 for equilateral and scalene configurations; the non-Gaussian terms of § 4.3 then couple different triangle bins. The imaginary part of each 4 × 4 4\times4 4 × 4 sub-matrix (right panels of Figure 1 ) is non-zero only for odd–even cross-multipole pairs, and the even–odd multipole pairs behave such that the matrix is Hermitian and positive semi-definite.
Figure 1. Real (left) and imaginary (right) parts of the Gaussian multipole covariance, C B B P P P C^{PPP}_{BB} C B B P P P , of a single triangle bin for a multi-tracer analysis. The top row is for a scalene configuration with k k k -bins centred at k 1 = 0.049 k_1 = 0.049 k 1 = 0.049 , k 2 = 0.036 k_2 = 0.036 k 2 = 0.036 , k 3 = 0.024 k_3 = 0.024 k 3 = 0.024 [h / M p c h/\mathrm{Mpc} h / Mpc ] and the bottom row represents an equilateral configuration k 1 = k 2 = k 3 = 0.024 k_1 = k_2 = k_3 = 0.024 k 1 = k 2 = k 3 = 0.024 [h / M p c h/\mathrm{Mpc} h / Mpc ]. We assume a bright-faint split for a Euclid-like survey with F s = 3 × 10 − 16 F_s = 3 \times10^{-16} F s = 3 × 1 0 − 16 [e r g c m − 2 s − 1 \mathrm{erg}\,\mathrm{cm}^{-2}\,\mathrm{s}^{-1} erg cm − 2 s − 1 ], for a redshift bin (1 < z < 1.2 1 < z < 1.2 1 < z < 1.2 ), including the local relativistic contributions to the covariance. White entries vanish. For the equilateral configuration the single-tracer dipoles, B 1 B B B B^{BBB}_1 B 1 B B B and B 1 F F F B^{FFF}_1 B 1 F F F , are strongly suppressed, as the estimator sums over every labelling of the sides and μ 1 + μ 2 + μ 3 = 0 \mu_1 + \mu_2 + \mu_3 = 0 μ 1 + μ 2 + μ 3 = 0 for an exactly equilateral triangle; they are non-zero only through the spread of the triangles across the bin (§ 3.2 ). For the equivalent power spectrum plot, see Figure 9 of Paper II.
Non-Gaussian contributions
For the bispectrum, the Gaussian covariance overestimates the information content (Chan & Blot 2017; Byun et al. 2017), and in squeezed configurations the non-Gaussian contributions can exceed the Gaussian one (Barreira 2019; Biagetti et al. 2022). Beyond the Gaussian term, the B B BB B B and P T PT P T terms of Eq. (35) are the leading contributions to the bispectrum covariance, and for a joint analysis the power spectrum and bispectrum are correlated at the same order through a term with one power spectrum and one bispectrum (P B PB P B ) (Chan & Blot 2017; Sugiyama et al. 2020; Biagetti et al. 2022; Salvalaggio et al. 2024). Whereas C P P P C^{PPP} C P P P pairs every side of one triangle with a side of the other, these terms link the two estimators through a single shared mode, so they correlate different configurations whenever these share a k k k -shell. As for C P P P C^{PPP} C P P P , we derive each term from Eq. (32) , or its 5-point analogue for P B PB P B , and evaluate it in the plane-parallel constant-redshift limit. The power spectra and bispectra in these terms include the local relativistic contributions, but we neglect the integrated ones.
B B BB B B term
Each of the nine B B BB B B permutations connects two fields of one triangle with one field of the other, such that
C B B [ B ^ ℓ i a b c , B ^ ℓ j d e f ] ( k , q ) = ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) V 2 k 1 2 N k T N q T ∫ k 1 , k 2 , k 3 , q ^ 1 , q ^ 2 , q ^ 3 × δ D ( k 123 ) δ D ( q 123 ) ⟨ F ℓ i a ( k 1 ) F 0 e ( q 2 ) F 0 f ( q 3 ) ∗ × F 0 b ( k 2 ) F 0 c ( k 3 ) F ℓ j d ( q 1 ) ∗ ⟩ + 8 p e r m s . (41) C^{BB}\left[\hat{B}^{abc}_{\ell_i},\hat{B}^{def}_{\ell_j}\right](\boldsymbol{k},\boldsymbol{q}) =
\frac{(2\ell_i+1)(2\ell_j+1)}{V^2 k_1^2 N_k^T N_q^T}
\int_{k_1,k_2,k_3,\hat{q}_1,\hat{q}_2,\hat{q}_3}
\times\delta_D(\boldsymbol{k}_{123})\delta_D(\boldsymbol{q}_{123})
\left\langle F^{a}_{\ell_i}(\boldsymbol{k}_1)F^e_0(\boldsymbol{q}_2)F^f_0(\boldsymbol{q}_3)^*
\times F^b_0(\boldsymbol{k}_2)F^c_0(\boldsymbol{k}_3)F^d_{\ell_j}(\boldsymbol{q}_1)^* \right\rangle+ 8\ {\rm perms}. \tag*{(41)} C B B [ B ^ ℓ i ab c , B ^ ℓ j d e f ] ( k , q ) = V 2 k 1 2 N k T N q T ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) ∫ k 1 , k 2 , k 3 , q ^ 1 , q ^ 2 , q ^ 3 × δ D ( k 123 ) δ D ( q 123 ) ⟨ F ℓ i a ( k 1 ) F 0 e ( q 2 ) F 0 f ( q 3 ) ∗ × F 0 b ( k 2 ) F 0 c ( k 3 ) F ℓ j d ( q 1 ) ∗ ⟩ + 8 perms . (41)
In the plane-parallel limit each 3-point function gives a bispectrum, e.g. ⟨ F a ( k 1 ) F e ( q 2 ) F f ( q 3 ) ∗ ⟩ = ( 2 π ) 3 δ D ( k 1 − q 2 − q 3 ) B ^ a e f ( k 1 , − q 2 , − q 3 ) \langle F^a(\boldsymbol{k}_1)F^e(\boldsymbol{q}_2)F^f(\boldsymbol{q}_3)^*\rangle=(2\pi)^3\delta_D(\boldsymbol{k}_1-\boldsymbol{q}_2-\boldsymbol{q}_3)\hat{B}^{aef}(\boldsymbol{k}_1,-\boldsymbol{q}_2,-\boldsymbol{q}_3) ⟨ F a ( k 1 ) F e ( q 2 ) F f ( q 3 ) ∗ ⟩ = ( 2 π ) 3 δ D ( k 1 − q 2 − q 3 ) B ^ a e f ( k 1 , − q 2 , − q 3 ) , so that
C B B [ B ^ ℓ i a b c , B ^ ℓ j d e f ] ( k , q ) = ( 2 π ) 6 ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) V 2 k 1 2 N k T N q T × ∫ k 1 , k 2 , k 3 , q ^ 1 , q ^ 2 , q ^ 3 δ D ( k 123 ) δ D ( q 123 ) × L ℓ i ( k ^ 1 ⋅ d ^ ) L ℓ j ( q ^ 1 ⋅ d ^ ) [ δ D ( k 1 − q 2 − q 3 ) δ D ( k 2 + k 3 − q 1 ) × B ^ a e f ( k 1 , − q 2 , − q 3 ) B ^ d b c ( − q 1 , k 2 , k 3 ) + 8 p e r m s ] . (42) C^{BB}\left[\hat{B}^{abc}_{\ell_i},\hat{B}^{def}_{\ell_j}\right](\boldsymbol{k},\boldsymbol{q}) =
\frac{(2\pi)^6(2\ell_i+1)(2\ell_j+1)}{V^2 k_1^2 N_k^T N_q^T}
\times\int_{k_1,k_2,k_3,\hat{q}_1,\hat{q}_2,\hat{q}_3}
\delta_D(\boldsymbol{k}_{123})\delta_D(\boldsymbol{q}_{123})
\times\mathcal{L}_{\ell_i}(\hat{\boldsymbol{k}}_1\cdot\hat{\boldsymbol{d}})\mathcal{L}_{\ell_j}(\hat{\boldsymbol{q}}_1\cdot\hat{\boldsymbol{d}})
\left[\delta_D(\boldsymbol{k}_1-\boldsymbol{q}_2-\boldsymbol{q}_3)\delta_D(\boldsymbol{k}_2+\boldsymbol{k}_3-\boldsymbol{q}_1)
\times\hat{B}^{aef}(\boldsymbol{k}_1,-\boldsymbol{q}_2,-\boldsymbol{q}_3)\hat{B}^{dbc}(-\boldsymbol{q}_1,\boldsymbol{k}_2,\boldsymbol{k}_3)+8\ {\rm perms}\right]. \tag*{(42)} C B B [ B ^ ℓ i ab c , B ^ ℓ j d e f ] ( k , q ) = V 2 k 1 2 N k T N q T ( 2 π ) 6 ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) × ∫ k 1 , k 2 , k 3 , q ^ 1 , q ^ 2 , q ^ 3 δ D ( k 123 ) δ D ( q 123 ) × L ℓ i ( k ^ 1 ⋅ d ^ ) L ℓ j ( q ^ 1 ⋅ d ^ ) [ δ D ( k 1 − q 2 − q 3 ) δ D ( k 2 + k 3 − q 1 ) × B ^ a e f ( k 1 , − q 2 , − q 3 ) B ^ d b c ( − q 1 , k 2 , k 3 ) + 8 perms ] . (42)
where, as for P ^ a b \hat{P}^{ab} P ^ ab in Eq. (39) , the bispectrum includes shot noise for non-overlapping samples. Since the shot noise of each estimator is subtracted, only galaxies shared between the two estimators contribute, so the single field from the other estimator, here k 1 \boldsymbol{k}_1 k 1 in B ^ a e f ( k 1 , − q 2 , − q 3 ) \hat{B}^{aef}(\boldsymbol{k}_1,-\boldsymbol{q}_2,-\boldsymbol{q}_3) B ^ a e f ( k 1 , − q 2 , − q 3 ) , must coincide with one of the other two (Sugiyama et al. 2020). With this side first,
B ^ a b c ( k 1 , k 2 , k 3 ) = B ^ a b c + δ a b K n a P c a ( k 3 ) + δ a c K n a P b a ( k 2 ) , (43) \hat{B}^{abc}(\boldsymbol{k}_1,\boldsymbol{k}_2,\boldsymbol{k}_3)=\hat{B}^{abc}+\frac{\delta^K_{ab}}{n_a}P^{ca}(\boldsymbol{k}_3)+\frac{\delta^K_{ac}}{n_a}P^{ba}(\boldsymbol{k}_2), \tag*{(43)} B ^ ab c ( k 1 , k 2 , k 3 ) = B ^ ab c + n a δ ab K P c a ( k 3 ) + n a δ a c K P ba ( k 2 ) , (43)
without the δ b c K P a b ( k 1 ) / n b \delta^K_{bc}P^{ab}(\boldsymbol{k}_1)/n_b δ b c K P ab ( k 1 ) / n b and δ a c K / n a 2 \delta^K_{ac}/n_a^2 δ a c K / n a 2 terms of the shot noise of a single bispectrum estimator. Once both triangles close, the two Dirac deltas in the bracket each set q 1 = − k 1 \boldsymbol{q}_1=-\boldsymbol{k}_1 q 1 = − k 1 , so one gives δ D ( 0 ) = V / ( 2 π ) 3 \delta_D(0)=V/(2\pi)^3 δ D ( 0 ) = V / ( 2 π ) 3 , and since Δ ( x ) \Delta(\boldsymbol{x}) Δ ( x ) is real, B ^ a e f ( k 1 , − q 2 , − q 3 ) = B ^ a e f ( q 1 , q 2 , q 3 ) ∗ \hat{B}^{aef}(\boldsymbol{k}_1,-\boldsymbol{q}_2,-\boldsymbol{q}_3)=\hat{B}^{aef}(\boldsymbol{q}_1,\boldsymbol{q}_2,\boldsymbol{q}_3)^* B ^ a e f ( k 1 , − q 2 , − q 3 ) = B ^ a e f ( q 1 , q 2 , q 3 ) ∗ . The term is therefore B ^ a b c ( k 1 , k 2 , k 3 ) B ^ a e f ( q 1 , q 2 , q 3 ) ∗ \hat{B}^{abc}(\boldsymbol{k}_1,\boldsymbol{k}_2,\boldsymbol{k}_3)\hat{B}^{aef}(\boldsymbol{q}_1,\boldsymbol{q}_2,\boldsymbol{q}_3)^* B ^ ab c ( k 1 , k 2 , k 3 ) B ^ a e f ( q 1 , q 2 , q 3 ) ∗ , such that each triangle keeps its own shape but the tracers on the shared side are exchanged, and it is non-zero only if k 1 k_1 k 1 and q 1 q_1 q 1 lie in the same shell. For a fixed shared mode, the integral over the rest of each triangle counts the triangles of its bin that contain the mode, on average N k T / N k 1 N_k^T/N_{k_1} N k T / N k 1 and N q T / N q 1 N_q^T/N_{q_1} N q T / N q 1 , and averages the bispectrum over their total count; it here N k = 4 π k 2 Δ k [ 1 + ( Δ k / k ) 2 / 12 ] / k f 3 N_k=4\pi k^2\Delta k[1+(\Delta k/k)^2/12]/k_f^3 N k = 4 π k 2 Δ k [ 1 + ( Δ k / k ) 2 /12 ] / k f 3 is the number of modes in the shell, which reduces to the familiar 4 π k 2 Δ k / k f 3 4\pi k^2\Delta k/k_f^3 4 π k 2 Δ k / k f 3 for thin shells. These counts cancel the normalisation, and the plane-parallel element is
C B B [ B ^ ℓ i a b c , B ^ ℓ j d e f ] ( k , q ) = ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) [ δ k 1 q 1 K N k 1 ∫ d Ω 4 π ∫ 0 2 π d ϕ 12 ′ 2 π L ℓ i ( μ 1 ) L ℓ j ( − μ 1 ) × B ^ d b c ( k 1 , k 2 , k 3 ) B ^ a e f ( − k 1 , q 2 , q 3 ) ∗ + 8 p e r m s ] . (44) C^{BB}\left[\hat{B}^{abc}_{\ell_i},\hat{B}^{def}_{\ell_j}\right](\boldsymbol{k},\boldsymbol{q}) =
(2\ell_i+1)(2\ell_j+1)
\left[\frac{\delta^K_{k_1q_1}}{N_{k_1}}
\int\frac{\mathrm{d}\Omega}{4\pi}\int_0^{2\pi}\frac{\mathrm{d}\phi'_{12}}{2\pi}
\mathcal{L}_{\ell_i}(\mu_1)\mathcal{L}_{\ell_j}(-\mu_1)
\times\hat{B}^{dbc}(\boldsymbol{k}_1,\boldsymbol{k}_2,\boldsymbol{k}_3)\hat{B}^{aef}(-\boldsymbol{k}_1,\boldsymbol{q}_2,\boldsymbol{q}_3)^*+8\ {\rm perms}\right]. \tag*{(44)} C B B [ B ^ ℓ i ab c , B ^ ℓ j d e f ] ( k , q ) = ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) [ N k 1 δ k 1 q 1 K ∫ 4 π d Ω ∫ 0 2 π 2 π d ϕ 12 ′ L ℓ i ( μ 1 ) L ℓ j ( − μ 1 ) × B ^ d b c ( k 1 , k 2 , k 3 ) B ^ a e f ( − k 1 , q 2 , q 3 ) ∗ + 8 perms ] . (44)
where, as in Eq. (40) , ∫ d Ω / 4 π \int\mathrm{d}\Omega/4\pi ∫ d Ω/4 π averages over the orientations of the k \boldsymbol{k} k triangle, parametrised by μ 1 \mu_1 μ 1 and the azimuthal angle ϕ 12 \phi_{12} ϕ 12 of k 2 \boldsymbol{k}_2 k 2 about k 1 \boldsymbol{k}_1 k 1 (Eq. (4)), and ϕ 12 ′ \phi'_{12} ϕ 12 ′ is the corresponding angle of the q \boldsymbol{q} q triangle about its first side, now fixed to q 1 = − k 1 \boldsymbol{q}_1=-\boldsymbol{k}_1 q 1 = − k 1 . The permutations run over the nine pairings of a side k m k_m k m with a side q n q_n q n in the same shell, with the Legendre polynomials remaining on k 1 k_1 k 1 and q 1 q_1 q 1 ; the pairing shown places both on the shared mode, so that, as for C P P P C^{PPP} C P P P , they depend only on μ 1 \mu_1 μ 1 .
Using the average counts assumes that the triangles of each bin are spread evenly over the modes of the shell. This holds for closed bins, but not for (nearly) flattened bins, where the closure condition cuts through the shell (Eq. (30)), so that the number of triangles containing a mode varies across its width (Biagetti et al. 2022). We therefore evaluate the shared mode at three Gauss–Legendre nodes across the width of its shell, weighting each triangle by its closure fraction there, which is sufficient for the mode counting of flattened bins to converge. For the same reason, at each node we average the bispectra over the triangles of the bin whose shared side lies there, rather than using the effective triangle of the whole bin (Section 3.2), and the same averaged bispectra enter C P B C_{PB} C P B (Eq. (46)). Evaluated on the bin centres instead, these terms could be overestimated by a factor of a few for flattened, squeezed bins.
For ℓ i = ℓ j = 0 \ell_i=\ell_j=0 ℓ i = ℓ j = 0 in real space Eq. (44) reduces to the familiar B ^ B ^ / N k \hat{B}\hat{B}/N_k B ^ B ^ / N k per shared side (Biagetti et al. 2022). However, since B ^ ( − k 1 , q 2 , q 3 ) = B ^ ( k 1 , k 2 , k 3 ) \hat{B}(-\boldsymbol{k}_1,\boldsymbol{q}_2,\boldsymbol{q}_3)=\hat{B}(\boldsymbol{k}_1,\boldsymbol{k}_2,\boldsymbol{k}_3) B ^ ( − k 1 , q 2 , q 3 ) = B ^ ( k 1 , k 2 , k 3 ) , the average over ϕ 12 ′ \phi'_{12} ϕ 12 ′ of B ^ a e f ( − k 1 , q 2 , q 3 ) \hat{B}^{aef}(-\boldsymbol{k}_1,\boldsymbol{q}_2,\boldsymbol{q}_3) B ^ a e f ( − k 1 , q 2 , q 3 ) is the complex conjugate of that with q 1 = + k 1 \boldsymbol{q}_1=+\boldsymbol{k}_1 q 1 = + k 1 , while L ℓ j ( − μ 1 ) = ( − 1 ) ℓ j L ℓ j ( μ 1 ) \mathcal{L}_{\ell_j}(-\mu_1)=(-1)^{\ell_j}\mathcal{L}_{\ell_j}(\mu_1) L ℓ j ( − μ 1 ) = ( − 1 ) ℓ j L ℓ j ( μ 1 ) . The B B BB B B term alone is therefore indefinite: for odd multipoles it can contribute a negative variance and can spuriously increase its signal-to-noise.
P T PT P T term
The nine P T PT P T terms instead contract one field of each triangle into a power spectrum and the remaining four fields into a trispectrum. Taking the same pairing of sides as in Eq. (41) , i.e. ⟨ F ℓ i a ( k 1 ) F ℓ j d ( q 1 ) F 0 b ( k 2 ) F 0 c ( k 3 ) F 0 e ( q 2 ) F 0 f ( q 3 ) ⟩ \langle F^{a}_{\ell_i}(\boldsymbol{k}_1)F^{d}_{\ell_j}(\boldsymbol{q}_1)F^{b}_{0}(\boldsymbol{k}_2)F^{c}_{0}(\boldsymbol{k}_3)F^{e}_{0}(\boldsymbol{q}_2)F^{f}_{0}(\boldsymbol{q}_3)\rangle ⟨ F ℓ i a ( k 1 ) F ℓ j d ( q 1 ) F 0 b ( k 2 ) F 0 c ( k 3 ) F 0 e ( q 2 ) F 0 f ( q 3 )⟩ , the plane-parallel element is
C P T [ B ^ ℓ i a b c , B ^ ℓ j d e f ] ( k , q ) = ( 2 π ) 6 ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) V 2 k f 12 N k T N q T × ∫ k 1 , k 2 , k 3 , q ^ 1 , q ^ 2 , q ^ 3 δ D ( k 123 ) δ D ( q 123 ) × L ℓ i ( k ^ 1 ⋅ d ^ ) L ℓ j ( q ^ 1 ⋅ d ^ ) [ δ D ( k 1 − q 1 ) δ D ( k 2 + k 3 − q 2 − q 3 ) × P ^ a d ( k 1 ) T b c e f ( k 2 , k 3 , − q 2 , − q 3 ) + 8 p e r m s ] . (45) \begin{aligned}
C^{PT}\left[\hat{B}^{abc}_{\ell_i},\hat{B}^{def}_{\ell_j}\right](\boldsymbol{k},\boldsymbol{q}) ={}& \frac{(2\pi)^6(2\ell_i+1)(2\ell_j+1)}{V^2 k_f^{12}N^T_{\boldsymbol{k}}N^T_{\boldsymbol{q}}}\\
&\times\int_{k_1,k_2,k_3,\hat{\boldsymbol{q}}_1,\hat{\boldsymbol{q}}_2,\hat{\boldsymbol{q}}_3}
\delta_D(\boldsymbol{k}_{123})\delta_D(\boldsymbol{q}_{123})\\
&\times\mathcal{L}_{\ell_i}(\hat{\boldsymbol{k}}_1\cdot\hat{\boldsymbol{d}})\mathcal{L}_{\ell_j}(\hat{\boldsymbol{q}}_1\cdot\hat{\boldsymbol{d}})\left[\delta_D(\boldsymbol{k}_1-\boldsymbol{q}_1)\delta_D(\boldsymbol{k}_2+\boldsymbol{k}_3-\boldsymbol{q}_2-\boldsymbol{q}_3)\right.\\
&\left.\times\hat{P}^{ad}(\boldsymbol{k}_1)T^{bcef}(\boldsymbol{k}_2,\boldsymbol{k}_3,-\boldsymbol{q}_2,-\boldsymbol{q}_3)+8\ \mathrm{perms}\right]. \tag*{(45)}
\end{aligned} C P T [ B ^ ℓ i ab c , B ^ ℓ j d e f ] ( k , q ) = V 2 k f 12 N k T N q T ( 2 π ) 6 ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) × ∫ k 1 , k 2 , k 3 , q ^ 1 , q ^ 2 , q ^ 3 δ D ( k 123 ) δ D ( q 123 ) × L ℓ i ( k ^ 1 ⋅ d ^ ) L ℓ j ( q ^ 1 ⋅ d ^ ) [ δ D ( k 1 − q 1 ) δ D ( k 2 + k 3 − q 2 − q 3 ) × P ^ a d ( k 1 ) T b ce f ( k 2 , k 3 , − q 2 , − q 3 ) + 8 perms ] . (45)
The shared side is now parallel, q 1 = + k 1 \boldsymbol{q}_1 = +\boldsymbol{k}_1 q 1 = + k 1 , and the trispectrum is in the collapsed configuration k 2 + k 3 = − k 1 \boldsymbol{k}_2 + \boldsymbol{k}_3 = -\boldsymbol{k}_1 k 2 + k 3 = − k 1 , whose diagonal is the shared mode. In general this does not factorise over the two triangles,
but it does when the shared mode is soft compared to the other sides, as for squeezed triangles that share their shortest side.
We therefore model the B B BB B B and P T PT P T terms together in this limit. On the shared mode, each triangle is the product of the field on its shared side, Δ a ( k 1 ) \Delta^a(\boldsymbol{k}_1) Δ a ( k 1 ) , and its other two fields, which we denote γ b c ( − k 1 ) \gamma^{bc}(-\boldsymbol{k}_1) γ b c ( − k 1 ) . The B B BB B B term contracts the shared field of each triangle with the other two fields of the other, and the P T PT P T term contracts the two shared fields with each other and the remaining fields with each other. In the squeezed limit, the other two fields respond linearly to the shared mode (Barreira 2019). Writing Δ X = Z X δ L + ϵ X \Delta^X = Z^X\delta_L + \epsilon_X Δ X = Z X δ L + ϵ X , where Z X Z^X Z X is the linear kernel of tracer X X X and ϵ X \epsilon_X ϵ X its Poisson noise, with ⟨ ϵ X ϵ X ′ ⟩ ′ = δ X X ′ K / n X \langle\epsilon_X\epsilon_{X'}\rangle' = \delta^K_{XX'}/n_X ⟨ ϵ X ϵ X ′ ⟩ ′ = δ X X ′ K / n X (the prime denotes a correlator without its factor of ( 2 π ) 3 δ D (2\pi)^3\delta_D ( 2 π ) 3 δ D ), we take
γ b c ( − k 1 ) = R b c P L ( k 1 ) δ L ( − k 1 ) + ∑ X n X s X b c ϵ X ( − k 1 ) , (46) \gamma^{bc}(-\boldsymbol{k}_1) = \frac{R^{bc}}{P_L(k_1)}\delta_L(-\boldsymbol{k}_1) + \sum_X n_X s_X^{bc}\epsilon_X(-\boldsymbol{k}_1), \tag*{(46)} γ b c ( − k 1 ) = P L ( k 1 ) R b c δ L ( − k 1 ) + X ∑ n X s X b c ϵ X ( − k 1 ) , (46)
where R b c R^{bc} R b c is the response of the pair to the linear density on the shared mode, multiplied by P L ( k 1 ) P_L(k_1) P L ( k 1 ) , and s X b c = B ^ X b c − B X b c s_X^{bc} = \hat{B}^{Xbc} - B^{Xbc} s X b c = B ^ X b c − B X b c is the shot noise of Eq. (43) with X X X on the shared side. This is the minimal format that gives ⟨ γ b c ( − k 1 ) Δ X ( k 1 ) ⟩ ′ = Z X R b c + s X b c \langle\gamma^{bc}(-\boldsymbol{k}_1)\Delta^X(\boldsymbol{k}_1)\rangle' = Z^X R^{bc} + s_X^{bc} ⟨ γ b c ( − k 1 ) Δ X ( k 1 ) ⟩ ′ = Z X R b c + s X b c , i.e. the bispectrum with any tracer X X X on the shared side, B ^ X b c \hat{B}^{Xbc} B ^ X b c , in the squeezed limit for its clustering part and exactly for its shot noise. The P T PT P T contraction then gives
P ^ a d ( k 1 ) T b c e f → P ^ a d ( k 1 ) [ R b c R e f ∗ P L ( k 1 ) + ∑ X n X s X b c s X e f ∗ ] , (47) \widehat{P}^{ad}(\boldsymbol{k}_1)T^{bcef}\to\widehat{P}^{ad}(\boldsymbol{k}_1)\left[\frac{R^{bc}R^{ef*}}{P_L(k_1)}+\sum_X n_Xs_X^{bc}s_X^{ef*}\right], \tag*{(47)} P a d ( k 1 ) T b ce f → P a d ( k 1 ) [ P L ( k 1 ) R b c R e f ∗ + X ∑ n X s X b c s X e f ∗ ] , (47)
whose clustering part is the collapsed exchange trispectrum. The shot noise part is a model rather than a derivation, as the shot noise of the trispectrum has no separable collapsed limit; without the shot noise of the bispectra it reduces to [ 1 + δ a d K / ( n a P a a ) ] B a b c B d e f ∗ [1+\delta^{K}_{ad}/(n_aP^{aa})]B^{abc}B^{def*} [ 1 + δ a d K / ( n a P aa )] B ab c B d e f ∗ , as kept in C P T C^{PT} C P T by Salvalaggio et al. (2024).
In the same model, the B B BB B B term of Eq. (44) pairs B ^ d b c = Z d d R b c + s b c \widehat{B}^{dbc}=Z_d^{d}R^{bc}+s^{bc} B d b c = Z d d R b c + s b c , the k \boldsymbol{k} k triangle with the other triangle’s tracer on its shared side, with its counterpart for the q \boldsymbol{q} q triangle. Its shot noise is therefore the exact exchange of Eq. (43) , while its clustering part takes the exchange in the squeezed limit, B d b c → ( Z d d / Z a d ) B a b c B^{dbc}\to(Z_d^{d}/Z_a^{d})B^{abc} B d b c → ( Z d d / Z a d ) B ab c , whose ratios cancel in the product. The integrals for the P T PT P T term reduce as for Eq. (44) , but with q 1 = + k 1 \boldsymbol{q}_1=+\boldsymbol{k}_1 q 1 = + k 1 , so that both Legendre polynomials are evaluated at + μ 1 +\mu_1 + μ 1 , and the two terms combine into
C B B + C P T ≃ ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) [ δ k 1 q 1 K N k 1 ∫ d Ω 4 π ∫ 0 2 π d ϕ 12 ′ 2 π × L ℓ i ( μ 1 ) L ℓ j ( μ 1 ) { P ^ a d ( k 1 ) [ R b c R e f ∗ P L ( k 1 ) + ∑ X n X s X b c s X e f ∗ ] × B ^ d b c ( k 1 , k 2 , k 3 ) B ^ a e f ( − k 1 , q 2 , q 3 ) ∗ } + 8 p e r m s ] , (48) \begin{aligned}
C^{BB}+C^{PT}\simeq{}&(2\ell_i+1)(2\ell_j+1)\left[\frac{\delta^{K}_{k_1q_1}}{N_{k_1}}\int\frac{\mathrm{d}\Omega}{4\pi}\int_0^{2\pi}\frac{\mathrm{d}\phi'_{12}}{2\pi}\right.\\
&\times L_{\ell_i}(\mu_1)L_{\ell_j}(\mu_1)\left\{\widehat{P}^{ad}(\boldsymbol{k}_1)\left[\frac{R^{bc}R^{ef*}}{P_L(k_1)}+\sum_Xn_Xs_X^{bc}s_X^{ef*}\right]\right.\\
&\left.\left.\times\widehat{B}^{dbc}(\boldsymbol{k}_1,\boldsymbol{k}_2,\boldsymbol{k}_3)\widehat{B}^{aef}(-\boldsymbol{k}_1,\boldsymbol{q}_2,\boldsymbol{q}_3)^*\right\}+8\ \mathrm{perms}\right], \tag*{(48)}
\end{aligned} C B B + C P T ≃ ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) [ N k 1 δ k 1 q 1 K ∫ 4 π d Ω ∫ 0 2 π 2 π d ϕ 12 ′ × L ℓ i ( μ 1 ) L ℓ j ( μ 1 ) { P a d ( k 1 ) [ P L ( k 1 ) R b c R e f ∗ + X ∑ n X s X b c s X e f ∗ ] × B d b c ( k 1 , k 2 , k 3 ) B a e f ( − k 1 , q 2 , q 3 ) ∗ } + 8 perms ] , (48)
where R b c R^{bc} R b c and s X b c s_X^{bc} s X b c are evaluated on the k \boldsymbol{k} k triangle, R e f R^{ef} R e f and s X e f s_X^{ef} s X e f on the q \boldsymbol{q} q triangle with q 1 = + k 1 \boldsymbol{q}_1=+\boldsymbol{k}_1 q 1 = + k 1 , and in each term q 2 \boldsymbol{q}_2 q 2 and q 3 \boldsymbol{q}_3 q 3 follow from q 1 \boldsymbol{q}_1 q 1 and ϕ 12 ′ \phi'_{12} ϕ 12 ′ . At fixed k 1 \boldsymbol{k}_1 k 1 the integrals over ϕ 12 \phi_{12} ϕ 12 and ϕ 12 ′ \phi'_{12} ϕ 12 ′ factorise, so each triangle enters only through azimuthal averages about its shared side.
Since the B B BB B B and P T PT P T terms are the two contractions of products of the jointly Gaussian variables of Eq. (46) , their sum is the covariance of these products and is therefore positive semi-definite for any number of tracers, even though the B B BB B B term alone is not. We therefore use this separable form for all nine pairings, extending the squeezed-limit approximation to configurations where the shared mode is not soft. With the full tree-level trispectrum instead, the P T PT P T term on its own need not be positive semi-definite, so we use it only to test the separable form (Section 4.2.4). Only the tracers of the two triangles enter, so the block of two single-tracer triangles, X X X XXX X X X , reduces to the single-tracer covariance of X X X .
For the real-space monopole without shot noise the two clustering parts are equal, recovering the C P T ≃ C B B C^{PT}\simeq C^{BB} C P T ≃ C B B approximation used for squeezed configurations (Biagetti et al. (2022); Salvalaggio et al. (2024)). For odd multipoles, however, the P T PT P T term cancels the B B BB B B term rather than doubling it: for a real bispectrum (the Newtonian limit) the clustering parts of Eq. (48) cancel, leaving only shot noise, so these multipoles, which carry the leading local relativistic signal, are largely insensitive to these terms (Figure 2 ).
Figure 2. Diagonal of the single-tracer bispectrum multipole covariance including the non-Gaussian terms of Eq. (48) (solid), and with the B B BB B B term alone (dashed), relative to the Gaussian term, for the monopole (top), dipole and quadrupole of each triangle bin. The bottom panel shows the sides of each triangle: the triangles are ordered by k 1 k_1 k 1 , then k 2 k_2 k 2 , then k 3 k_3 k 3 , with the dark (faint) dashed lines marking where k 1 k_1 k 1 (k 2 k_2 k 2 ) changes. The grey bands mark flattened configurations, k 1 = k 2 + k 3 k_1 = k_2 + k_3 k 1 = k 2 + k 3 at the bin centres. We assume a Euclid-like survey with a redshift bin (1 < z < 1.2 1 < z < 1.2 1 < z < 1.2 ) and bins of width Δ k = 4 k f \Delta k = 4k_{\mathrm{f}} Δ k = 4 k f up to k max = 0.1 k_{\max} = 0.1 k m a x = 0.1 [h / M p c h/\mathrm{Mpc} h / Mpc ], including the local relativistic contributions to the covariance.
Figure 2 shows the effect of these terms on the variance of the single-tracer bispectrum multipoles. For the monopole and quadrupole, they are concentrated in (nearly) flattened configurations, k 1 ≃ k 2 + k 3 k_1\simeq k_2+k_3 k 1 ≃ k 2 + k 3 , where two sides are aligned and the bispectrum is largest relative to the power spectra, and are largest for those that are also squeezed, as their long-wavelength side lies in a shell with few modes, each shared by many triangles: there they reach a significant fraction of the Gaussian variance. Here the P T PT P T term roughly doubles the B B BB B B term, close to the C P T ≃ C B B C^{PT}\simeq C^{BB} C P T ≃ C B B limit. For the dipole, the B B BB B B term alone reduces the variance of the same configurations, whereas with the P T PT P T term it remains very close to Gaussian.
Power spectrum–bispectrum cross-covariance
For a joint analysis we also require C P B C_{PB} C P B . Writing the normalisation of Eq. (1) as 1 / ( V N k f k f 3 ) 1/(V N_{k_f}k_f^3) 1/ ( V N k f k f 3 ) , the analogue of Eq. (32) is a 5-point function,
C [ P ^ ℓ i a b , B ^ ℓ j d e f ] ( k , q ) = ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) V 2 k f 2 N k N q T ∫ k ^ 1 , k ^ 2 , q ^ 1 , q ^ 2 , q ^ 3 δ D ( k 12 ) δ D ( q 123 ) × [ ⟨ F ℓ i a ( k 1 ) F 0 b ( k 2 ) F ℓ j d ( q 1 ) F 0 e ( q 2 ) F 0 f ( q 3 ) ∗ ⟩ − ⟨ F ℓ i a ( k 1 ) F 0 b ( k 2 ) ⟩ ⟨ F ℓ j d ( q 1 ) F 0 e ( q 2 ) F 0 f ( q 3 ) ∗ ⟩ ] , (49) \begin{aligned}
C[\widehat{P}^{ab}_{\ell_i},\widehat{B}^{def}_{\ell_j}](\boldsymbol{k},\boldsymbol{q})=\frac{(2\ell_i+1)(2\ell_j+1)}{V^2k_f^2N_kN_q^T}\int_{\widehat{k}_1,\widehat{k}_2,\widehat{q}_1,\widehat{q}_2,\widehat{q}_3}\delta_D(k_{12})\delta_D(q_{123}) \\
&\times\left[\langle F^{a}_{\ell_i}(\boldsymbol{k}_1)F^b_0(\boldsymbol{k}_2)F^d_{\ell_j}(\boldsymbol{q}_1)F^e_0(\boldsymbol{q}_2)F^f_0(\boldsymbol{q}_3)^*\rangle-\langle F^{a}_{\ell_i}(\boldsymbol{k}_1)F^b_0(\boldsymbol{k}_2)\rangle\langle F^d_{\ell_j}(\boldsymbol{q}_1)F^e_0(\boldsymbol{q}_2)F^f_0(\boldsymbol{q}_3)^*\rangle\right], \tag*{(49)}
\end{aligned} C [ P ℓ i ab , B ℓ j d e f ] ( k , q ) = V 2 k f 2 N k N q T ( 2 ℓ i + 1 ) ( 2 ℓ j + 1 ) ∫ k 1 , k 2 , q 1 , q 2 , q 3 δ D ( k 12 ) δ D ( q 123 ) × [ ⟨ F ℓ i a ( k 1 ) F 0 b ( k 2 ) F ℓ j d ( q 1 ) F 0 e ( q 2 ) F 0 f ( q 3 ) ∗ ⟩ − ⟨ F ℓ i a ( k 1 ) F 0 b ( k 2 )⟩ ⟨ F ℓ j d ( q 1 ) F 0 e ( q 2 ) F 0 f ( q 3 ) ∗ ⟩ ] , (49)
where k 1 k_1 k 1 and k 2 k_2 k 2 both run over the shell of k k k . Of the ten ways of splitting the five fields into a pair and a triplet, the one that separates the two estimators cancels the disconnected product, and a power spectrum of two sides of the triangle vanishes as before. Neglecting the connected (tetraspectrum) contribution, which is subleading on the scales of interest (Biagetti et al. (2022); Salvalaggio et al. (2024)), each of the six remaining terms contracts one field of P ^ \widehat{P} P with a side of the triangle and closes a bispectrum with the other field and the remaining two sides, e.g. ⟨ F ℓ i a ( k 1 ) F ℓ j d ( q 1 ) ∗ ⟩ ⟨ F 0 b ( k 2 ) F 0 e ( q 2 ) ∗ F 0 f ( q 3 ) ∗ ⟩ \langle F^{a}_{\ell_i}(\boldsymbol{k}_1)F^{d}_{\ell_j}(\boldsymbol{q}_1)^*\rangle\langle F^b_0(\boldsymbol{k}_2)F^e_0(\boldsymbol{q}_2)^*F^f_0(\boldsymbol{q}_3)^*\rangle ⟨ F ℓ i a ( k 1 ) F ℓ j d ( q 1 ) ∗ ⟩ ⟨ F 0 b ( k 2 ) F 0 e ( q 2 ) ∗ F 0 f ( q 3 ) ∗ ⟩ . In the plane-parallel limit this sets q 1 = k 1 \boldsymbol{q}_1=\boldsymbol{k}_1 q 1 = k 1 and gives P ^ a d ( k 1 ) B ^ b e f ( q 1 , q 2 , q 3 ) ∗ \widehat{P}^{ad}(\boldsymbol{k}_1)\widehat{B}^{bef}(\boldsymbol{q}_1,\boldsymbol{q}_2,\boldsymbol{q}_3)^* P a d ( k 1 ) B b e f ( q 1 , q 2 , q 3 ) ∗ , so, in Eq. (44) , the other tracer of P ^ \widehat{P} P takes the place of the shared side. Reducing the integrals as before,
\begin{aligned}
C[\widehat{P}^{ab}_{\ell_i},\widehat{B}^{def}_{\ell_j}](\boldsymbol{k},\boldsymbol{q})={}&(2\ell_i+1)(2\ell_j+1)\left[\frac{\delta^{K}_{kq_1}}{N_k}\int\frac{\mathrm{d}\Omega}{4\pi}\right.\\
&\times L_{\ell_i}(\mu_1)L_{\ell_j}(\mu_1)\left(\widehat{P}^{ad}(\boldsymbol{k}_1)\widehat{B}^{bef}(\boldsymbol{k}_1,\boldsymbol{q}_2,\boldsymbol{q}_3)^*\\
&\left.\qquad+(-1)^{\ell_i}\widehat{P}^{bd}(\boldsymbol{k}_1)\widehat{B}^{aef}(\boldsymbol{k}_1,\boldsymbol{q}_2,\boldsymbol{q}_3)^*\right)+2\ \mathrm{perms}\right], \tag*{(50)}
\end{aligned}
where ∫ d Ω / 4 π \int\mathrm{d}\Omega/4\pi ∫ d Ω/4 π now averages over the orientations of the q \boldsymbol{q} q triangle, with its first side along the power spectrum mode, q 1 = k 1 q_1=k_1 q 1 = k 1 . The two terms correspond to the two fields of P ^ \widehat{P} P , with k 1 → − k 1 k_1\to-k_1 k 1 → − k 1 in the second giving the factor ( − 1 ) ℓ i (-1)^{\ell_i} ( − 1 ) ℓ i , and the permutations to the three sides of the triangle. For ℓ i = ℓ j = 0 \ell_i=\ell_j=0 ℓ i = ℓ j = 0 in real space this reduces to 2 P ^ B ^ / N k 2\widehat{P}\widehat{B}/N_k 2 P B / N k per shared side (Biagetti et al. (2022)). Here the mode count is exact, as the triangles on each mode of a shell sum to N q T N_q^T N q T , and N k N_k N k is the exact number of modes in the shell rather than its thin-shell approximation.
For consistency with Eq. (48) we take the exact exchange of the clustering part in the squeezed limit, B b e f → ( Z b b / Z d b ) B d e f B^{bef}\to(Z_b^b/Z_d^b)B^{def} B b e f → ( Z b b / Z d b ) B d e f , keeping the exact exchange in the shot noise, so that B ^ b e f = Z b b R e f + s b e f \widehat{B}^{bef}=Z_b^bR^{ef}+s_b^{ef} B b e f = Z b b R e f + s b e f is the correlator of Eq. (46) that also enters the B B BB B B term. With this choice, Eq. (50) follows from the same model as Eq. (48) and the Gaussian C P P C_{PP} C P P , so the joint covariance of the power spectrum and bispectrum is consistent. This requires C P P C_{PP} C P P to be averaged over the modes of each shell (§ 3.2 ) on the same nodes, with the same exact mode count. This matters for multi-tracer analyses: we find that using only the triangles’ own tracers in the shot noise of the P T PT P T term, or the exact tracer exchange for the clustering part of Eq. (50) , can make the joint covariance indefinite, as can the thin-shell mode count at the bin centre in C P P C_{PP} C P P and C P B C_{PB} C P B .
Away from the diagonal, these terms correlate the triangle bins that share a k k k -shell, and C P B C_{PB} C P B correlates each triangle with the power spectrum in the shells of its sides, as shown by the correlation matrix of the joint data vector in Figure C2 . As for the Gaussian term, the real part couples only multipoles of the same parity and the imaginary part only odd–even pairs, so the bispectrum dipole is correlated with the even multipoles and the power spectrum only through the imaginary part, which arises from the local relativistic contributions to the bispectrum and is small. These terms are implemented in CosmoWAP in covariances.py, with the tree-level comparison in the class PTTreeCovBk, and the tests described here are in test_bb_cov.py.9
Figure C2. Real (top) and imaginary (bottom) parts of the correlation matrix, C i j / C i i C j j C_{ij}/\sqrt{C_{ii}C_{jj}} C ij / C ii C j j , of the joint single-tracer data vector of the power spectrum (ℓ = 0 , 2 \ell= 0, 2 ℓ = 0 , 2 ) and bispectrum (ℓ = 0 , 1 , 2 \ell= 0, 1, 2 ℓ = 0 , 1 , 2 ) multipoles, including the non-Gaussian terms of Eqs. (48) and (50) , for the same survey and redshift bin as Figure 2 , with k max = 0.1 k_{\max} = 0.1 k m a x = 0.1 [h / M p c h/\mathrm{Mpc} h / Mpc ] for both. Within each block, the rows run over the k k k -bins of the power spectrum, or over the triangle bins in the order of Figure 2 . The colour scales are logarithmic, and linear below 10 − 3 10^{-3} 1 0 − 3 (top) and 10 − 4 10^{-4} 1 0 − 4 (bottom), with the imaginary part on a much smaller scale. The imaginary part is antisymmetric and non-zero only between multipoles of opposite parity, so the bispectrum dipole is correlated with the even multipoles and the power spectrum only through it, and only weakly (see § 4.3 ). For visual clarity we use wider bins here, Δ k = 6 k f \Delta k = 6k_{\mathrm{f}} Δ k = 6 k f .
After the azimuthal integrals, Eqs. (48) and (50) depend on the configurations only through μ 1 \mu_1 μ 1 of the shared mode, so on a Gauss–Legendre grid in μ 1 \mu_1 μ 1 the non-Gaussian covariance is a low-rank product, C N G = W Ξ W † C^{\mathrm{NG}} = W\Xi W^\dagger C NG = W Ξ W † . Here W W W has a fixed number of columns per k k k -shell, built from the azimuthally averaged power spectra, bispectra and their shot noise, and Ξ \Xi Ξ holds the quadrature weights and couples columns only within a shell. The rank is therefore independent of the number of triangles, and the inverse of the full covariance follows from the Woodbury identity in terms of the inverse of the Gaussian covariance, which is block-diagonal in k k k -bins and triangles, without forming the dense matrix.
We have verified Eqs. (44) , (48) and (50) against direct sums over the orientations of the triangles and power spectrum modes, and the mode counting of Eq. (44) against a count of the triangle pairs that share a Fourier mode on a lattice. To test the separable form of Eq. (47) , we have also computed the P T PT P T term with the full tree-level trispectrum, for which the s s s channel, the exchange of the shared mode, is exactly separable, but the t t t and u u u channels and the T 3111 T_{3111} T 3111 contribution depend on the relative orientation of the two triangles and must be integrated directly. The two give consistent signal-to-noise for both the Newtonian and the local relativistic bispectrum.
Results I: PNG
This section presents the results of the forecasts on PNG, with § 5.1 covering the bias in the determination of f N L f_{\mathrm{NL}} f NL if one neglects relativistic and wide-separation effects. Including these effects introduces the relativistic biases, n e v o l n_{\mathrm{evol}} n evol and Q Q Q , as additional nuisance parameters, so in § 5.2 we consider how well they are constrained by the clustering data themselves and how they can be determined independently. Then, in § 5.3 we investigate the constraining power on f N L f_{\mathrm{NL}} f NL from single-tracer and multi-tracer forecasts and the impact that the uncertainty introduced by relativistic effects has on the inference. In particular, we focus on the underestimation of the true parameter covariance when one neglects this uncertainty.
We constrain each shape separately, holding the amplitudes of the other two at zero. We give the joint equilateral and orthogonal constraints for comparison in § 5.5 . We focus primarily on a Euclid -like Hα \alpha α survey, but also consider a SPHEREx -like survey and a MegaMapper-like Stage V Lyman-break galaxy (LBG) survey; details of the survey modelling are given in Appendix B. For the SPHEREx -like survey we use only the sample with the most precise redshifts, which is partly why our constraints are weaker than published SPHEREx forecasts (Doré et al. 2014; Heinrich et al. 2024) that combine all five redshift-accuracy samples (see Section B2.4). We do not consider the DESI-like BGS survey here since, in our forecasts, it lacks the volume for competitive constraints on PNG.
In § 5.1 we sample the posterior on local-type f N L f_{\mathrm{NL}} f NL with MCMCs, and elsewhere use Fisher forecasts (§ 3 ).
Systematic bias
The forecasted parameter constraints for a ‘naive’ analysis that neglects relativistic and wide-separation effects are shown in the ellipses in Figure 3 for a joint power spectrum and bispectrum analysis. For the Euclid -like survey the inferred value is shifted to f N L l o c = 17. 2 − 9.4 + 9.2 f_{\mathrm{NL}}^{\mathrm{loc}} = 17.2^{+9.2}_{-9.4} f NL loc = 17. 2 − 9.4 + 9.2 , a ∼ 2 σ \sim2\sigma ∼ 2 σ shift from the true fiducial value of f N L = 0 f_{\mathrm{NL}} = 0 f NL = 0 considered in this forecast.
Figure 3. Forecasted marginalised and joint 1 and 2 σ \sigma σ parameter covariances for a Euclid -like Hα \alpha α survey (blue), a SPHEREx -like survey (green) and a MegaMapper-like LBG survey (red), if we neglect integrated, local relativistic and wide-separation effects, for a joint power spectrum and bispectrum analysis with priors on the cosmological parameters from the Planck 2018 results. b 1 b_1 b 1 , b 2 b_2 b 2 and b Γ 2 b_{\Gamma_2} b Γ 2 are marginalised in each redshift bin.
This bias is comparable to the Euclid error bar but negligible for SPHEREx , f N L l o c = 0. 5 − 7.7 + 7.7 f_{\mathrm{NL}}^{\mathrm{loc}} = 0.5^{+7.7}_{-7.7} f NL loc = 0. 5 − 7.7 + 7.7 , whose lower redshift coverage and magnification bias suppress the lensing contribution. It dominates the error budget, however, for the MegaMapper-like LBG survey constraints shown in the red posteriors. This arises for two main reasons, firstly due to the improved constraining power, and secondly due to the greater dominance of the lensing effect at higher redshifts. Thus, we forecast that one would ‘detect’ a positive value at high significance, f N L l o c = 22. 8 − 0.9 + 0.9 f_{\mathrm{NL}}^{\mathrm{loc}} = 22.8^{+0.9}_{-0.9} f NL loc = 22. 8 − 0.9 + 0.9 , a ∼ 25 σ \sim25\sigma ∼ 25 σ shift10 , generated purely by neglecting relativistic effects despite there being no fiducial PNG signal.
As in Paper II, local relativistic effects shift f N L l o c f_{\mathrm{NL}}^{\mathrm{loc}} f NL loc towards negative values, but this is countered by a larger positive shift from integrated lensing: if we only neglect the local relativistic and wide-separation terms (Figure C1 ), the shift is small and negative for Euclid and MegaMapper, ∼ 0.1 σ \sim0.1\sigma ∼ 0.1 σ and ∼ 0.4 σ \sim0.4\sigma ∼ 0.4 σ , and negligible for SPHEREx , while the shifts in the non-local shapes are negligible and those in the cosmological parameters are ∼ 0.2 σ \sim0.2\sigma ∼ 0.2 σ .
Figure C1. As Figure 3 , but with the integrated (lensing, ISW and time-delay) contributions omitted from the data, so that the shifts arise only from neglecting the local relativistic and wide-separation effects. All shifts are ≲ 0.4 σ \lesssim0.4\sigma ≲ 0.4 σ , so the biases in Figure 3 are driven by the integrated terms.
Additionally, for MegaMapper, the inference on the orthogonal shape is also biased, by ∼ 3 σ \sim3\sigma ∼ 3 σ , f N L o r t h = 26 ± 9 f_{\mathrm{NL}}^{\mathrm{orth}} = 26 \pm9 f NL orth = 26 ± 9 , while the equilateral shape by ∼ 1 σ \sim1\sigma ∼ 1 σ , f N L e q = − 6 ± 6 f_{\mathrm{NL}}^{\mathrm{eq}} = -6 \pm6 f NL eq = − 6 ± 6 .11 For Euclid these shifts are forecasted to be well within the survey error budget, f N L e q ≈ − 9 ± 34 f_{\mathrm{NL}}^{\mathrm{eq}} \approx-9 \pm34 f NL eq ≈ − 9 ± 34 and f N L o r t h ≈ 16 ± 40 f_{\mathrm{NL}}^{\mathrm{orth}} \approx16 \pm40 f NL orth ≈ 16 ± 40 . Neglecting lensing also biases the inference of the cosmological parameters for the MegaMapper-like case, despite the priors from Planck : by ∼ 3 σ \sim3\sigma ∼ 3 σ for Ω m \Omega_{\mathrm{m}} Ω m , h h h and n s n_{\mathrm{s}} n s , and ∼ 10 σ \sim10\sigma ∼ 10 σ for ln ( 10 10 A s ) \ln(10^{10}A_{\mathrm{s}}) ln ( 1 0 10 A s ) . For Euclid the largest shift is ∼ 1.2 σ \sim1.2\sigma ∼ 1.2 σ , in ln ( 10 10 A s ) \ln(10^{10}A_{\mathrm{s}}) ln ( 1 0 10 A s ) . Consistently including light-cone effects removes this bias; however, it introduces additional nuisance parameters in the form of the relativistic biases, n e v o l n_{\mathrm{evol}} n evol and Q Q Q . Before looking at the resulting constraints on f N L f_{\mathrm{NL}} f NL we first consider how well these bias parameters can be determined.
Information on the relativistic biases
Once relativistic effects are included in the model, n e v o l n_{\mathrm{evol}} n evol and Q Q Q become nuisance parameters in each redshift bin, and the uncertainty on these biases propagates into any parameter that is degenerate with the relativistic signal – such as local-type f N L f_{\mathrm{NL}} f NL (§ 5.3 ) and the amplitude of the local relativistic effects themselves (§ 6 ). Throughout, we bracket this with the ‘fixed’ and ‘free’ cases of § 3.1 . Where a real analysis lies between the two depends on how well the biases can be determined independently of the clustering data, so here we consider both what the clustering data themselves say about n e v o l n_{\mathrm{evol}} n evol and Q Q Q and how the biases can be determined independently of them.
In the even multipoles the relativistic biases enter only at O ( H / k ) 2 \mathcal{O}(\mathcal{H}/k)^2 O ( H / k ) 2 , where they are degenerate with each other and the k − 2 k^{-2} k − 2 scaling of local-type PNG, while Q Q Q additionally enters through the lensing contribution. The leading-order O ( H / k ) \mathcal{O}(\mathcal{H}/k) O ( H / k ) relativistic terms, which are sensitive to b e b_{\mathrm{e}} b e (or n e v o l n_{\mathrm{evol}} n evol ) and Q Q Q (through Eq. (54) ), instead enter into the odd-parity multipoles: the bispectrum dipole (and octupole) for a single tracer, and the odd cross-power spectrum multipoles for a multi-tracer analysis. The odd multipoles are therefore the natural way to constrain these biases from the clustering data itself; in particular b e b_{\mathrm{e}} b e (n e v o l n_{\mathrm{evol}} n evol ) as the magnification bias is already partially constrained by the lensing signal.12
Figure 4 shows the marginalised constraints on n e v o l n_{\mathrm{evol}} n evol and Q Q Q from a joint power spectrum and bispectrum analysis of the Euclid -like survey, including the odd multipoles, together with the luminosity function priors discussed below. The clustering data barely constrain the count slope for a single tracer, σ ( n e v o l ) ≈ 16 \sigma(n_{\mathrm{evol}}) \approx16 σ ( n evol ) ≈ 16 –26 in each redshift bin, compared with fiducial values ∣ n e v o l ∣ ≈ 2.5 |n_{\mathrm{evol}}| \approx2.5 ∣ n evol ∣ ≈ 2.5 –5. The magnification bias is much better determined, with σ ( Q ) \sigma(Q) σ ( Q ) falling from ≈ 1 \approx1 ≈ 1 at z ≈ 1 z \approx1 z ≈ 1 to ≈ 0.3 \approx0.3 ≈ 0.3 at z ≈ 1.7 z \approx1.7 z ≈ 1.7 , as the lensing signal grows with redshift. As we show in § 5.3 , the weak constraint on n e v o l n_{\mathrm{evol}} n evol matters little for local-type f N L f_{\mathrm{NL}} f NL once the bispectrum dipole is included, since the remaining degeneracy is then with Q Q Q .
Figure 4. Constraints on the count slope n e v o l n_{\mathrm{evol}} n evol (Eq. (54) ; top) and the magnification bias (bottom) for a Euclid -like Hα \alpha α survey, for a single-tracer analysis (left) and a bright-faint split at a fixed flux, for which each sample has its own n e v o l n_{\mathrm{evol}} n evol (right). The shaded bands show the 1 and 2 σ 2\sigma 2 σ priors forward-modelled from the uncertainty in the ‘Model 3’ luminosity function (Appendix A4). The error bars show the marginalised 1 σ 1\sigma 1 σ Fisher constraints from a joint power spectrum and bispectrum analysis without luminosity function priors, including the bispectrum dipole (and, for the bright-faint split, the odd cross-power spectrum multipoles), centred on the fiducial value in each redshift bin. For the bright-faint split, the bright and faint error bars are offset slightly in redshift for clarity.
The picture is similar for the bright-faint split (right-hand panels of Figure 4 ). The single-bin posteriors in Figure C5 (in the appendix) show that the clustering data mainly constrain combinations of the bright and faint biases, with strong positive correlations between n e v o l B n_{\mathrm{evol}}^{B} n evol B and n e v o l F n_{\mathrm{evol}}^{F} n evol F and between Q B Q^{B} Q B and Q F Q^{F} Q F . This is consistent with the odd-parity cross-power being sourced by the difference between these biases.
Figure C5. As Figure C4 , but for a bright-faint split at a fixed flux, so that each sample has its own n e v o l n_{\mathrm{evol}} n evol . The joint constraints include the bispectrum dipole and the odd cross-power spectrum multipoles.
Figure C4. Constraints on the relativistic biases for a single-tracer Euclid -like Hα \alpha α survey from an MCMC for a single redshift bin of width Δ z = 0.18 \Delta z = 0.18 Δ z = 0.18 centred at z = 1.35 z = 1.35 z = 1.35 , with 1 and 2σ \sigma σ contours. Unlike the Fisher constraints in Figure 4 , these capture the non-Gaussianity of the posteriors. The joint power spectrum and bispectrum constraints, including the bispectrum dipole, are shown in red and the priors forward-modelled from the ‘Model 3’ luminosity function in blue.
Both biases can, however, be potentially determined independently of the clustering. Unlike the evolution bias, n e v o l n_{\mathrm{evol}} n evol and Q Q Q are properties of the observed sample at the survey flux cut (Appendix A1), so they can be measured from the catalogue itself, without assuming a form for the luminosity function. The count slope is the logarithmic slope of the redshift distribution of the sample, and can be measured by binning the catalogue in redshift. Its precision is set less by the number of galaxies than by how well the redshift dependence of the completeness, purity and redshift success rate of the sample is known, since any unmodelled trend in these enters the measured slope directly. In addition, Eq. (54) assumes that the sample is defined by a single flux cut: any further selection adds terms to it, so that n e v o l n_{\mathrm{evol}} n evol and Q Q Q no longer determine b e b_{\mathrm{e}} b e on their own ([115 ] (Sullivan & Seljak 2025)). For the BOSS luminous red galaxies, for example, Sullivan & Seljak (2025) find that the other criteria used to select the samples affect the redshift evolution of their number density more than the magnification correction does. The magnification bias is the response of the number counts to a small change in the flux (or magnitude) cut, and can be measured by perturbing the fluxes in the catalogue and re-applying the full sample selection (Zhou et al. 2023; Elvin-Poole et al. 2023; Wenzl et al. 2024), or, where the selection depends on more than the flux, by injecting synthetic sources into the survey images (Everett et al. 2022; Elvin-Poole et al. 2023). The latter may be better suited to the Hα \alpha α sample, for which the efficiency of the redshift measurement depends on the signal-to-noise ratio of the emission line rather than being set by a sharp flux cut (e.g. Euclid Collaboration et al. 2023). For current samples, finite-difference estimates of Q Q Q reach per-cent-level statistical precision (Zhou et al. 2023; Wenzl et al. 2024).
To consider sample-motivated priors on the relativistic biases in our forecasts, we use existing measurements of the luminosity function of the target population, propagated into priors on n e v o l n_{\mathrm{evol}} n evol and Q Q Q , following Wang et al. (2020a). We use the Euclid-like sample as an example, with priors forward-modelled from the uncertainty in the ‘Model 3’ Hα \alpha α luminosity function of Pozzetti et al.
(2016) (Appendix A4), shown by the shaded bands in Figure 4 . These priors, σ ( n e v o l ) ≈ 1 \sigma(n_{\mathrm{evol}}) \approx1 σ ( n evol ) ≈ 1 –2 2 2 and σ ( Q ) ≈ 0.3 \sigma(Q) \approx0.3 σ ( Q ) ≈ 0.3 in each redshift bin, are tighter than the clustering constraints on n e v o l n_{\mathrm{evol}} n evol by an order of magnitude or more, and comparable to those on Q Q Q at z ≳ 1.5 z \gtrsim1.5 z ≳ 1.5 . However, they include only the statistical uncertainty on the parameters of a single model. In Appendix A4 we show that different luminosity function models that give the same number density agree closely on n e v o l n_{\mathrm{evol}} n evol , since matching the number density largely fixes the redshift distribution, but differ in Q Q Q by several times its statistical uncertainty. The model dependence of such priors therefore enters mainly through Q Q Q and thus the actual results will be dependent on the true magnification bias of the sample.
We quantify how well the relativistic biases must be known to approach the ‘fixed’ constraints in § 5.4.1 for local-type f N L f_{\mathrm{NL}} f NL and in § 6 for the amplitude of the local relativistic effects.
Constraints
The forecast constraints on f N L f_{\mathrm{NL}} f NL for each shape are summarised in Table 1 . We compare the ‘fixed’ and ‘free’ treatments of the relativistic biases (§ 3.1 ) with, for the Euclid -like survey, the luminosity function priors of Appendix A4. The ‘fixed’ constraint approximates the error of an analysis that neglects relativistic corrections. The headline result is that marginalising over the relativistic biases increases σ ( f N L l o c ) \sigma(f_{\mathrm{NL}}^{\mathrm{loc}}) σ ( f NL loc ) by up to ∼ 60 % \sim60\% ∼ 60% (for the Euclid -like power spectrum), so that an analysis that neglects relativistic corrections would underestimate its error on local-type f N L f_{\mathrm{NL}} f NL by the same factor and hence overstate the significance of any detection. Below we first consider the information from the bispectrum and the bright-faint split with the biases fixed (§ 5.3.1 ), then the cost of marginalising over them and how much of it the clustering data themselves recover (§ 5.4 ), and finally the impact of external priors (§ 5.4.1 ).
Table 1. Forecast 1 σ 1\sigma 1 σ constraints on f N L f_{\mathrm{NL}} f NL for each shape, with n e v o l n_{\mathrm{evol}} n evol and Q Q Q free (no prior), with the luminosity function (LF) prior of Appendix A4, and held fixed at their fiducial values. ST denotes a single tracer and MT a bright-faint split with a fixed bright fraction n B / n T = 0.2 n_{\mathrm{B}}/n_{\mathrm{T}}=0.2 n B / n T = 0.2 . All multipoles are included, as in Figure 5 . We do not forward model LF priors for the MegaMapper-like sample. Each shape is constrained independently, with the other two held at zero.
Figure 5. Constraints on local-type f N L f_{\mathrm{NL}} f NL for a Euclid-like Hα \alpha α survey (left) and a MegaMapper-like LBG survey (right), for a power spectrum (P P P ) and a joint power spectrum and bispectrum (P + B P+B P + B ) analysis. The filled bars show the single-tracer constraints and the dashed unfilled bars a bright-faint split with a fixed bright fraction, n B / n T = 0.2 n_{\mathrm{B}}/n_{\mathrm{T}}=0.2 n B / n T = 0.2 , in every redshift bin (so that both samples share the same n e v o l n_{\mathrm{evol}} n evol ). In both cases n e v o l n_{\mathrm{evol}} n evol and Q Q Q are marginalised over in each redshift bin with no luminosity function prior, while the black dash-dotted lines show the single-tracer constraint with them held at their fiducial values. All multipoles are included: ℓ = 0 , 2 , 4 \ell=0,2,4 ℓ = 0 , 2 , 4 for the power spectrum (plus the odd cross-multipoles ℓ = 1 , 3 \ell=1,3 ℓ = 1 , 3 for the bright-faint split) and ℓ = 0 , 1 , 2 \ell=0,1,2 ℓ = 0 , 1 , 2 for the bispectrum. The horizontal axes are linear in 1 / σ 1/\sigma 1/ σ .
Information from the bispectrum and the bright-faint split
With n e v o l n_{\mathrm{evol}} n evol and Q Q Q fixed, adding the bispectrum to a single-tracer analysis of the Euclid -like survey improves the constraint on local-type f N L f_{\mathrm{NL}} f NL by only ∼ 2 % \sim2\% ∼ 2% (9.36 → 9.17 9.36 \rightarrow9.17 9.36 → 9.17 ). For the MegaMapper-like survey the gain is similarly small, ∼ 4 % \sim4\% ∼ 4% (1.12 → 1.08 1.12 \rightarrow1.08 1.12 → 1.08 ). The local-type signal lies predominantly in squeezed configurations, where triangles sharing a long mode are strongly correlated with each other and with the power spectrum (Biagetti et al. 2022; Ssalalila et al. 2024); these triangles largely measure the same scale-dependent bias as the power spectrum13 . This is consistent with the modest gain from the bispectrum in BOSS ((Cabass et al. 2022a)); raising the bispectrum k max k_{\max} k m a x to 0.15 h / M p c 0.15\,h/\mathrm{Mpc} 0.15 h / Mpc increases the gain only to ∼ 7 % \sim7\% ∼ 7% .
With the relativistic biases fixed, the gain from a bright-faint split on local-type f N L f_{\mathrm{NL}} f NL comes from the cancellation of sample variance between the two samples (Seljak (2009)), which depends on how much their ratios of PNG to linear bias, b Ψ / b 1 b_{\Psi}/b_1 b Ψ / b 1 , differ. For the peak-background split biases (§ 2.1.2 ) this ratio tends to 2 δ c 2\delta_{\mathrm{c}} 2 δ c for highly biased tracers, so the bright and faint samples differ little: by ≲ 10 % \lesssim10\% ≲ 10% for the Euclid-like sample, and from ∼ 20 % \sim20\% ∼ 20% at z ≈ 2.3 z\approx2.3 z ≈ 2.3 to ∼ 2 % \sim2\% ∼ 2% at z ≈ 4.9 z\approx4.9 z ≈ 4.9 for the MegaMapper-like sample. The split therefore changes the local-type constraint by ≲ 1 % \lesssim1\% ≲ 1% for the Euclid-like survey and by ∼ 5 % \sim5\% ∼ 5% for the MegaMapper-like survey (Table 1 ).
The dependence of the joint bright-faint split constraints on the bright fraction is shown for the MegaMapper-like sample in Figure 6 , where all three shapes favour small bright samples. The local-type constraint depends only weakly on the split, loosening slowly as the bright sample grows, from σ ( f N L l o c ) ≈ 1.12 \sigma(f_{\mathrm{NL}}^{\mathrm{loc}})\approx1.12 σ ( f NL loc ) ≈ 1.12 for n B / n T ≈ 0.15 n_{\mathrm{B}}/n_{\mathrm{T}}\approx0.15 n B / n T ≈ 0.15 –0.3 0.3 0.3 to 1.19 1.19 1.19 at n B / n T = 0.9 n_{\mathrm{B}}/n_{\mathrm{T}}=0.9 n B / n T = 0.9 , and is below the single-tracer constraint for every fraction we consider. The equilateral and orthogonal constraints are tightest at n B / n T ≈ 0.1 n_{\mathrm{B}}/n_{\mathrm{T}}\approx0.1 n B / n T ≈ 0.1 –0.15 0.15 0.15 and degrade by ∼ 85 % \sim85\% ∼ 85% and ∼ 75 % \sim75\% ∼ 75% , respectively, by n B / n T = 0.9 n_{\mathrm{B}}/n_{\mathrm{T}}=0.9 n B / n T = 0.9 . We therefore use n B / n T = 0.2 n_{\mathrm{B}}/n_{\mathrm{T}}=0.2 n B / n T = 0.2 unless otherwise stated, which is within 2 % 2\% 2% of the best constraint for all shapes.
Figure 6. Dependence of the bright-faint split constraints on the bright fraction, n B / n T n_{\mathrm{B}}/n_{\mathrm{T}} n B / n T , held fixed at every redshift, for local (left), equilateral (centre) and orthogonal (right) type f N L f_{\mathrm{NL}} f NL , for a joint power spectrum and bispectrum analysis of a MegaMapper-like LBG survey. The coloured lines show the bright-faint split and the black horizontal lines the single-tracer constraints. Solid lines marginalise over n e v o l n_{\mathrm{evol}} n evol and Q Q Q in each redshift bin with an uninformative prior (with both samples of the split sharing n e v o l n_{\mathrm{evol}} n evol ), while dashed lines hold them at their fiducial values. All multipoles are included, as in Figure 5 .
Degeneracy with the relativistic biases
The cost of marginalising over the relativistic biases is broken down by analysis for the Euclid-like survey in Figure 7 . It is largest for the single-tracer power spectrum, for which σ ( f N L l o c ) \sigma(f_{\mathrm{NL}}^{\mathrm{loc}}) σ ( f NL loc ) increases by ∼ 60 % \sim60\% ∼ 60% (9.4 → 14.8 9.4\rightarrow14.8 9.4 → 14.8 ), as the even multipoles n e v o l n_{\mathrm{evol}} n evol and Q Q Q enter at their k − 2 k^{-2} k − 2 with the same k − 2 k^{-2} k − 2 scaling as local-type PNG (§ 5.2 ). The odd-parity multipoles add no PNG signal14 , but they are sensitive to the leading-order H / k \mathcal{H}/k H / k relativistic terms and so help to break the degeneracy between f N L f_{\mathrm{NL}} f NL and the relativistic biases.
Figure 7. Forecast 1 σ 1\sigma 1 σ constraints on local-type f N L f_{\mathrm{NL}} f NL from each analysis, for a Euclid-like Hα survey. The orange points marginalise over n e v o l n_{\mathrm{evol}} n evol and Q Q Q with no prior, the blue points add the priors forward-modelled from the ‘Model 3’ luminosity function (Appendix A4), and the black ticks mark the ‘fixed’ case, so that each grey line spans the full cost of marginalising over the relativistic biases. The numbers on the right give σ ( f N L ) \sigma(f_{\mathrm{NL}}) σ ( f NL ) without → \to → with the luminosity function prior, with the ‘fixed’ case in brackets. The bright-faint splits use either a fixed split flux, F s = 3 × 10 − 16 e r g c m − 2 s − 1 F_{\mathrm{s}} = 3 \times10^{-16}\,\mathrm{erg}\,\mathrm{cm}^{-2}\mathrm{s}^{-1} F s = 3 × 1 0 − 16 erg cm − 2 s − 1 , with independent n e v o l n_{\mathrm{evol}} n evol for each sample, or a fixed bright fraction n B / n T = 0.2 n_{\mathrm{B}}/n_{\mathrm{T}} = 0.2 n B / n T = 0.2 at every redshift, for which both samples share the same n e v o l n_{\mathrm{evol}} n evol . Unless stated otherwise all multipoles, including the odd-parity ones, are used.
The even bispectrum multipoles partially weaken the degeneracy (14.8 → 13.8 14.8\rightarrow13.8 14.8 → 13.8 ), and the non-zero bispectrum dipole generated by local relativistic projection effects tightens the constraint by a further ∼ 23 % \sim23\% ∼ 23% (to 10.6 10.6 10.6 ), leaving it ∼ 16 % \sim16\% ∼ 16% above the ‘fixed’ case. The dipole does this by breaking the degeneracy between n e v o l n_{\mathrm{evol}} n evol and f N L f_{\mathrm{NL}} f NL (Figure C3 ), even though n e v o l n_{\mathrm{evol}} n evol itself remains poorly constrained by the clustering (σ ( n e v o l ) ≈ 16 \sigma(n_{\mathrm{evol}})\approx16 σ ( n evol ) ≈ 16 –26 26 26 ; Figure 4 ), so that the remaining loss is almost entirely due to the magnification bias (Figure 8 ), which is still constrained mainly through the lensing contribution.
Figure 8. Forecast 1 σ 1\sigma 1 σ constraint on local-type f N L f_{\mathrm{NL}} f NL as a function of the widths of Gaussian priors on n e v o l n_{\mathrm{evol}} n evol and on the fractional magnification bias, σ ( Q ) / Q \sigma(Q)/Q σ ( Q ) / Q , for a single-tracer Euclid -like Hα \alpha α survey (top) and MegaMapper -like LBG survey (bottom), from a joint power spectrum (ℓ = 0 , 2 , 4 \ell= 0, 2, 4 ℓ = 0 , 2 , 4 ) and bispectrum analysis with the even bispectrum multipoles only (ℓ = 0 , 2 \ell= 0, 2 ℓ = 0 , 2 ; left) and with the dipole included (ℓ = 0 , 1 , 2 \ell= 0, 1, 2 ℓ = 0 , 1 , 2 ; right). Each survey has its own colour scale. With n e v o l n_{\mathrm{evol}} n evol and Q Q Q held at their fiducial values the constraints are σ ( f N L l o c ) = 9.2 \sigma(f_{\mathrm{NL}}^{\mathrm{loc}}) = 9.2 σ ( f NL loc ) = 9.2 (Euclid ) and 1.08 (MegaMapper ), compared with 13.8 and 10.6 (Euclid , left and right) and 1.28 and 1.21 (MegaMapper ) with both free. The same prior width is applied in every redshift bin, with Q Q Q evaluated at the centre of each bin, and added to the Fisher matrix, which includes the Planck prior and marginalises over b 1 b_1 b 1 , b 2 b_2 b 2 and b Γ 2 b_{\Gamma_2} b Γ 2 in each bin, on an 80 × 80 80 \times80 80 × 80 logarithmically spaced grid. The crosses mark the priors forward-modelled from the ‘Model 3’ luminosity function (Appendix A4), placed at their mean width over the redshift bins. L-shaped contours indicate that a prior on either bias alone breaks the degeneracy with f N L f_{\mathrm{NL}} f NL , while horizontal contours indicate that only Q Q Q matters; including the bispectrum dipole makes the contours close to horizontal (see text).
Figure C3. Fisher constraints on local-type f N L f_{\mathrm{NL}} f NL and on the relativistic biases, n e v o l n_{\mathrm{evol}} n evol and Q Q Q , in the redshift bin centred at z = 1.35 z = 1.35 z = 1.35 , for a single-tracer Euclid -like Hα \alpha α survey, with 1 and 2σ \sigma σ contours. The data vectors are the power spectrum (ℓ = 0 , 2 , 4 \ell= 0, 2, 4 ℓ = 0 , 2 , 4 ; blue) and the joint power spectrum and bispectrum without (green) and with (red) the bispectrum dipole. All other parameters are marginalised over. Without the dipole, f N L f_{\mathrm{NL}} f NL is anticorrelated with both biases. The dipole effectively removes the correlation with n e v o l n_{\mathrm{evol}} n evol and narrows its constraint by a factor of ∼ 4 \sim4 ∼ 4 .
A bright-faint split doubles the per-bin linear and magnification biases, and the count slope too if the split is at a fixed flux, so its additional information must outweigh these extra nuisance parameters. For a split at a fixed flux, for which each sample has its own n e v o l n_{\mathrm{evol}} n evol , the power spectrum constraint barely improves on the single tracer (14.8 → \to → 14.1), whereas for a fixed bright fraction, for which both samples share a single n e v o l n_{\mathrm{evol}} n evol (Appendix A3), it improves by ∼ 25 % \sim25\% ∼ 25% (to 11.1). Once the bispectrum dipole is included, however, either split adds little (∼ 1 % \sim1\% ∼ 1% ), as the dipole has already broken the degeneracy with n e v o l n_{\mathrm{evol}} n evol .
Priors on the relativistic biases
Applying the luminosity function priors for the Euclid-like survey (blue points in Figure 7 ) improves the constraints on local-type f N L f_{\mathrm{NL}} f NL almost to the ‘fixed’ limit. Since these priors are tighter than the clustering constraints on the relativistic biases (Figure 4 ), they leave little for the odd-parity multipoles to add. However, these depend on the luminosity function model (§ 5.2 ), so the blue points should potentially be regarded as optimistic.
Figure 8 shows how precisely the relativistic biases must be known to recover the ‘fixed’ constraint on local-type f N L f_{\mathrm{NL}} f NL . For the Euclid -like power spectrum (and similarly for a joint analysis with the even multipoles) the two biases are strongly degenerate: with both free, σ ( f N L l o c ) \sigma(f_{\mathrm{NL}}^{\mathrm{loc}}) σ ( f NL loc ) is 58% larger than with both fixed, but knowing either one removes most of this loss (15% with n e v o l n_{\mathrm{evol}} n evol fixed, 3% with Q Q Q fixed). However, including the bispectrum dipole makes the constraint almost insensitive to n e v o l n_{\mathrm{evol}} n evol : with the dipole the loss with both free drops to 16%, almost all of it due to Q Q Q , and fixing Q Q Q alone brings the constraint to within 0.2% of the fixed case. The luminosity function priors constrain n e v o l n_{\mathrm{evol}} n evol well enough to break the degeneracy in every case, leaving a ∼ 4 \sim4 ∼ 4 –5% degradation set by σ ( Q ) / Q ≃ 10 \sigma(Q)/Q \simeq10 σ ( Q ) / Q ≃ 10 –14%; reaching 1% would require σ ( Q ) / Q ≲ 4 % \sigma(Q)/Q \lesssim4\% σ ( Q ) / Q ≲ 4% in each redshift bin, which is of the same order as the statistical precision of current finite-difference estimates of Q Q Q from the catalogue (§ 5.2 ; Zhou et al. 2023; Wenzl et al. 2024).
For the MegaMapper-like sample the power spectrum constraint degrades by 19%. The joint constraint degrades by 19% (12% with the dipole) and, with n e v o l n_{\mathrm{evol}} n evol fixed, requires σ ( Q ) / Q ≲ 0.8 % \sigma(Q)/Q \lesssim0.8\% σ ( Q ) / Q ≲ 0.8% to come within 1%. The large lensing contribution at these redshifts constrains Q Q Q from the clustering itself, to ≈ 13 % \approx13\% ≈ 13% in the lowest redshift bin and ≈ 2 % \approx2\% ≈ 2% above z ≈ 3.4 z \approx3.4 z ≈ 3.4 . A prior of σ ( Q ) / Q ≈ 10 % \sigma(Q)/Q \approx10\% σ ( Q ) / Q ≈ 10% , comparable to the luminosity function priors for the Euclid -like sample, therefore tightens the joint constraint by only ≲ 1 % \lesssim1\% ≲ 1% , and a prior on Q Q Q helps appreciably only once it reaches the per-cent level, where it begins to compete with the clustering constraint.
Finally, since the odd multipoles constrain n e v o l n_{\mathrm{evol}} n evol and Q Q Q directly from the clustering (§ 5.2 ), they also provide a consistency test of any external determination of the biases.
Non-local PNG
Equilateral and orthogonal PNG are far less degenerate with relativistic effects than the local shape: the scale-dependent bias in Eq. (13a) scales as k − 2 / T ( k ) k^{-2}/T(k) k − 2 / T ( k ) , i.e. as k − 2 k^{-2} k − 2 for both shapes, compared with the scaling shared by local-type PNG and the leading even-parity ( H / k ) 2 (\mathcal{H}/k)^2 ( H / k ) 2 relativistic terms. The conclusions of the local case therefore carry over but are far less pronounced. For the single-tracer joint forecasts15 , neglecting relativistic effects leads one to underestimate the errors on f N L e q f_{\mathrm{NL}}^{\mathrm{eq}} f NL eq and f N L o r t h f_{\mathrm{NL}}^{\mathrm{orth}} f NL orth by only ≲ 2 % \lesssim2\% ≲ 2% , and in contrast to the local case, priors on the relativistic biases make little difference (≲ 1 % \lesssim1\% ≲ 1% ) for either the single-tracer or bright-faint split analysis in a Euclid -like survey. This remains true for the MegaMapper-like sample, for which marginalising over them loosens the joint constraints on both shapes by ≲ 1 % \lesssim1\% ≲ 1% (Table 1 ), even though neglecting relativistic effects biases the orthogonal shape by ∼ 3 σ \sim3\sigma ∼ 3 σ (§ 5.1 ).
The bright-faint split gives a large gain for the MegaMapper-like sample, improving σ ( f N L e q ) \sigma(f_{\mathrm{NL}}^{\mathrm{eq}}) σ ( f NL eq ) from 5.8 to 2.9 and σ ( f N L o r t h ) \sigma(f_{\mathrm{NL}}^{\mathrm{orth}}) σ ( f NL orth ) from 9.3 to 4.9 (Table 1 ), but for a Euclid -like survey it changes the joint constraints by ≲ 2 % \lesssim2\% ≲ 2% , giving σ ( f N L e q ) ≃ 34 \sigma(f_{\mathrm{NL}}^{\mathrm{eq}}) \simeq34 σ ( f NL eq ) ≃ 34 and σ ( f N L o r t h ) ≃ 40 \sigma(f_{\mathrm{NL}}^{\mathrm{orth}}) \simeq40 σ ( f NL orth ) ≃ 40 with or without the split. For reference, the current Planck constraints are σ ( f N L e q ) = 47 \sigma(f_{\mathrm{NL}}^{\mathrm{eq}}) = 47 σ ( f NL eq ) = 47 and σ ( f N L o r t h ) = 24 \sigma(f_{\mathrm{NL}}^{\mathrm{orth}}) = 24 σ ( f NL orth ) = 24 ([124 ] ): the Euclid -like survey would improve on the equilateral one by a factor of ∼ 1.4 \sim1.4 ∼ 1.4 , but its orthogonal constraint is ∼ 1.7 \sim1.7 ∼ 1.7 times weaker than Planck ’s. Fitting the equilateral and orthogonal shapes jointly loosens the single-tracer constraints, to σ ( f N L e q ) ≃ 40 \sigma(f_{\mathrm{NL}}^{\mathrm{eq}}) \simeq40 σ ( f NL eq ) ≃ 40 and σ ( f N L o r t h ) ≃ 47 \sigma(f_{\mathrm{NL}}^{\mathrm{orth}}) \simeq47 σ ( f NL orth ) ≃ 47 for the Euclid -like survey (correlation coefficient + 0.52 +0.52 + 0.52 ) and to 13.1 and 20.9 for the MegaMapper-like survey (+ 0.90 +0.90 + 0.90 ), as the two shapes are then separated only by the bispectrum, and their correlation depends on the uncertain ratio of their PNG biases. We caution that these constraints derive mainly from bispectrum configurations near k m a x k_{\mathrm{max}} k max , and are therefore the results most sensitive to our tree-level modelling.
Results II: Local Relativistic Effects
This section focuses on forecasts for the detection of local relativistic projection effects. For forecasts on the detectability of integrated relativistic effects in the power spectrum see Appendix H of Paper II. We group all the local relativistic effects together and quantify their contribution through a global amplitude parameter, α G R \alpha_{\mathrm{GR}} α GR , with a fiducial value of unity. We first consider the impact of wide-separation corrections on this analysis, and then the constraining power of multi-tracer and joint power spectrum and bispectrum analyses, particularly once the uncertainty in the relativistic biases is included. In all of these forecasts we marginalise over the same set of parameters as in the PNG analyses with Planck priors (§ 3 ) and with f N L = 0 f_{\mathrm{NL}} = 0 f NL = 0 held fixed for every shape.
The odd multipoles are particularly sensitive to local relativistic effects, as this is where the leading-order H / k \mathcal{H}/k H / k signal enters. One can access this signal in our standard summary statistics either through the bispectrum, which has a non-zero dipole (and octupole) outside the equilateral limit, or through multi-tracer approaches.
We find that single-tracer approaches (including the bispectrum dipole) are insufficient for confident detections of local relativistic effects, and we therefore focus mainly on bright-faint split approaches.
Biases
Figure 10 shows the constraints on α G R \alpha_{\mathrm{GR}} α GR if one ignores wide-separation type corrections in the bright-faint split analysis. The impact of ignoring wide-separation effects is twofold: first, the unaccounted-for signal biases the inference on the local relativistic effects, in both the odd and even multipoles, owing to their shared k k k -dependence. This is particularly prominent for the lower-redshift BGS survey, for which the inferred value of α G R \alpha_{\mathrm{GR}} α GR is shifted by ≈ 9 σ \approx9\sigma ≈ 9 σ in the power spectrum analysis and ≈ 10 σ \approx10\sigma ≈ 10 σ in the joint analysis, due to a larger wide-angle contribution. At higher redshifts the radial evolution is the dominant part of the wide-separation corrections and this biases the inference for the Euclid -like survey by ≈ 3 σ \approx3\sigma ≈ 3 σ and the MegaMapper-like survey by less than 1 σ 1\sigma 1 σ . Second, the wide-separation terms themselves contain local relativistic signal; the first-order wide-separation corrections to the odd-parity H / k \mathcal{H}/k H / k local relativistic term generate a relevant ∼ ( H / k ) 2 \sim(\mathcal{H}/k)^2 ∼ ( H / k ) 2 contribution (see § 3 of Paper I for a discussion of the scale and redshift dependence of the wide-separation corrections). Including them tightens the constraint on α G R \alpha_{\mathrm{GR}} α GR by ∼ 15 % \sim15\% ∼ 15% for BGS and MegaMapper and by only ∼ 1 % \sim1\% ∼ 1% for Euclid . This contribution is included in the constraints discussed in the next section.
Figure 10. Forecasted 1 σ 1\sigma 1 σ constraints on the amplitude of local relativistic effects, α G R \alpha_{\mathrm{GR}} α GR , for a DESI-like BGS survey (purple), a Euclid-like Hα \alpha α survey (blue) and a MegaMapper-like LBG survey (red), if we neglect wide-separation corrections beyond the zeroth order of the wide-separation expansion. The triangles show a multi-tracer (bright-faint) joint power spectrum and bispectrum analysis, and the squares a multi-tracer power spectrum analysis, as in Figure 11 . The relativistic biases, n e v o l n_{\mathrm{evol}} n evol and Q Q Q , are held at their fiducial values, and each sample is split at a fixed bright fraction, n B / n T = 0.2 n_{\mathrm{B}}/n_{\mathrm{T}}=0.2 n B / n T = 0.2 for the DESI-like BGS and MegaMapper-like samples and n B / n T = 0.4 n_{\mathrm{B}}/n_{\mathrm{T}}=0.4 n B / n T = 0.4 for the Euclid-like sample, as in Figure 11 . The numbers on the right give the shift of each from the fiducial value, α G R = 1 \alpha_{\mathrm{GR}}=1 α GR = 1 (dotted line), in units of its 1 σ 1\sigma 1 σ error.
Constraints
The forecasted cumulative marginalised SNR on the amplitude of local relativistic effects, α G R \alpha_{\mathrm{GR}} α GR , is plotted in Figure 11 and listed in Table 2 for the DESI-like BGS, Euclid-like and MegaMapper-like surveys16 . In Figure 11 , n e v o l n_{\mathrm{evol}} n evol and Q Q Q are held at their fiducial values.
Table 2. Forecast SNR on the amplitude of local relativistic effects, α G R \alpha_{\mathrm{GR}} α GR , over all redshift bins, with n e v o l n_{\mathrm{evol}} n evol and Q Q Q marginalised over in each redshift bin with no prior (free) and held at their fiducial values (fixed), for the data vectors of Figure 11 . ST denotes a single tracer (ℓ = 0 , 2 , 4 \ell= 0, 2, 4 ℓ = 0 , 2 , 4 for the power spectrum and ℓ = 0 , 1 , 2 \ell= 0, 1, 2 ℓ = 0 , 1 , 2 for the bispectrum) and MT a bright-faint split at a fixed bright fraction, n B / n T = 0.2 n_B/n_T = 0.2 n B / n T = 0.2 for the DESI-like BGS and MegaMapper-like samples and 0.4 for the Euclid-like sample, which also includes the odd cross-power spectrum multipoles.
Figure 11. Forecasted cumulative marginalised SNR on the amplitude of local relativistic effects, α G R \alpha_{\rm GR} α GR , for a DESI-like BGS survey (purple), a Euclid -like Hα \alpha α survey (blue) and a MegaMapper-like LBG survey (red), over redshift bins, for different data vectors, with n e v o l n_{\rm evol} n evol and Q Q Q held at their fiducial values. The DESI-like BGS and MegaMapper-like samples are split at a fixed bright fraction, n B / n T = 0.2 n_{\rm B}/n_{\rm T}=0.2 n B / n T = 0.2 , and the Euclid -like sample at n B / n T = 0.4 n_{\rm B}/n_{\rm T}=0.4 n B / n T = 0.4 : the SNR depends only weakly on the bright fraction around these values, which give the highest SNR of the fractions we consider for the Euclid -like sample and are within 1% and 2% of it for the DESI-like BGS and MegaMapper-like samples, respectively (see Figure 13 ). For the DESI-like BGS survey, the shaded bands extend each curve to the SNR without the wide-separation-based minimum k k k -cut, and therefore give some indication of the additional signal on scales not accessible to the 3D Fourier statistics; for the other two surveys the cut removes no k k k -bins.
With n e v o l n_{\mathrm{evol}} n evol and Q Q Q fixed, we forecast that local relativistic effects should be detectable at high significance for a DESI-like BGS survey with a bright-faint split, with a cumulative SNR of ≃ 12.5 \simeq12.5 ≃ 12.5 for the multi-tracer power spectrum and ≃ 12.6 \simeq12.6 ≃ 12.6 for the joint analysis. For the Euclid-like and MegaMapper-like surveys the multi-tracer power spectrum gives an SNR of ≃ 1.5 \simeq1.5 ≃ 1.5 and ≃ 2 \simeq2 ≃ 2 , respectively, rising to ≃ 1.6 \simeq1.6 ≃ 1.6 and ≃ 2.2 \simeq2.2 ≃ 2.2 with the multi-tracer bispectrum, so that local relativistic effects are at best marginally detectable.17
Figure 9. Constraints on equilateral (left) and orthogonal (right) type f N L f_{\mathrm{NL}} f NL for a Euclid-like Hα \alpha α survey (top) and a MegaMapper-like LBG survey (bottom). The filled bars show the single-tracer constraints and the dashed unfilled bars a bright-faint split with a fixed bright fraction, n B / n T = 0.2 n_{\mathrm{B}}/n_{\mathrm{T}}=0.2 n B / n T = 0.2 , in every redshift bin (so that both samples share the same n e v o l n_{\mathrm{evol}} n evol ). In both cases n e v o l n_{\mathrm{evol}} n evol and Q Q Q are marginalised over in each redshift bin with no luminosity function prior, while the black dash-dotted lines show the single-tracer constraint with them held at their fiducial values. All multipoles are included: ℓ = 0 , 2 , 4 \ell=0,2,4 ℓ = 0 , 2 , 4 for the power spectrum (plus the odd cross-multipoles ℓ = 1 , 3 \ell=1,3 ℓ = 1 , 3 for the bright-faint split) and ℓ = 0 , 1 , 2 \ell=0,1,2 ℓ = 0 , 1 , 2 for the bispectrum. The horizontal axes are linear in 1 / σ 1/\sigma 1/ σ .
The single-tracer joint analysis gives an SNR below 1 for all three surveys (≈ 0.2 \approx0.2 ≈ 0.2 for BGS and ≈ 0.6 \approx0.6 ≈ 0.6 for the other two), and as such we forecast that bright-faint split approaches are far more important than the bispectrum dipole for detecting local relativistic effects.
This result is contrary to much of the existing literature, which forecasts a detectable bispectrum dipole in single-tracer analyses for both Euclid and DESI BGS (Maartens et al. 2020; Noorikuhani & Scoccimarro 2023; Addis et al. 2025; Rossiter et al. 2025). Part of this difference arises from a sign error in the tree-level bispectrum of Umeh et al. (2017): the second-order kernel contracted with the linear fields at k 1 \mathbf{k}_1 k 1 and k 2 \mathbf{k}_2 k 2 must be evaluated at ( − k 1 , − k 2 ) (-\mathbf{k}_1,-\mathbf{k}_2) ( − k 1 , − k 2 ) (Eq. (13) ), i.e. complex conjugated (§ 2.2 ). This leaves the Newtonian bispectrum unchanged but, rather than simply flipping the sign of the imaginary part, changes the relative sign of the first- and second-order relativistic contributions to both its imaginary and real parts. In our forecasts the correction strengthens cancellations between these terms and so reduces the SNR of the local relativistic contribution; differences in the survey modelling, covariance and nuisance parameters also contribute.
For the multi-tracer analysis, as in the local PNG case, the bispectrum adds little information beyond the power spectrum once the non-Gaussian covariance is included. With the biases fixed, the multi-tracer bispectrum adds ∼ 1 % \sim1\% ∼ 1% to the SNR for BGS, ∼ 8 % \sim8\% ∼ 8% for Euclid and ∼ 11 % \sim11\% ∼ 11% for MegaMapper. With the biases free, however, it raises the SNR by ∼ 15 \sim15 ∼ 15 –75 % 75\% 75% (Table 2 ), as it helps to break the degeneracy of α G R \alpha_{\mathrm{GR}} α GR with the relativistic biases.
If n e v o l n_{\mathrm{evol}} n evol and Q Q Q are instead marginalised over in each redshift bin with no prior, the constraints on α G R \alpha_{\mathrm{GR}} α GR loosen by factors of ∼ 15 \sim15 ∼ 15 and ∼ 13 \sim13 ∼ 13 for the multi-tracer power spectrum and joint analyses of the DESI-like BGS survey, and ∼ 2.6 \sim2.6 ∼ 2.6 and ∼ 1.6 \sim1.6 ∼ 1.6 for the Euclid-like survey. Below we quantify how well the biases must be determined externally in order to detect local relativistic effects.
BGS
The DESI-like BGS survey gives a larger SNR than the Euclid -like and MegaMapper-like samples despite its smaller volume, largely because of its lower redshift coverage. At leading order in H / k \mathcal{H}/k H / k , the relativistic signal of a bright-faint split appears mainly in the dipole of the cross-power spectrum (Eqs. (12) and (17)'),
Im P 1 B F ( k ) = f H k [ ( b 1 B γ 1 B − b 1 F γ 1 F ) + 3 5 f ( γ 1 B − γ 1 F ) ] D 2 P ( k ) , (51) \operatorname{Im} P^{\rm BF}_{1}(k)=f\frac{\mathcal{H}}{k}\left[\left(b^{\rm B}_{1}\gamma^{\rm B}_{1}-b^{\rm F}_{1}\gamma^{\rm F}_{1}\right)+\frac{3}{5}f\left(\gamma^{\rm B}_{1}-\gamma^{\rm F}_{1}\right)\right]D^{2}P(k), \tag*{(51)} Im P 1 BF ( k ) = f k H [ ( b 1 B γ 1 B − b 1 F γ 1 F ) + 5 3 f ( γ 1 B − γ 1 F ) ] D 2 P ( k ) , (51)
which is set by the difference between the Doppler coefficients of the two samples. Grouping the terms of γ 1 \gamma_1 γ 1 (§ 2.2 ) by their origin,
γ 1 = b e − ( 2 x H + H ′ H 2 ) + 2 Q ( 1 x H − 1 ) . (52) \gamma_1=b_{\rm e}-\left(\frac{2}{x\mathcal{H}}+\frac{\mathcal{H}'}{\mathcal{H}^{2}}\right)+2Q\left(\frac{1}{x\mathcal{H}}-1\right). \tag*{(52)} γ 1 = b e − ( x H 2 + H 2 H ′ ) + 2 Q ( x H 1 − 1 ) . (52)
The Doppler shift displaces each galaxy radially, by an amount ∝ v ∥ / H \propto v_{\parallel}/\mathcal{H} ∝ v ∥ / H , at a fixed observed redshift. Each term of Eq. (52) is one consequence of this: the comoving number density is different at the galaxy’s true redshift (evolution), the redshift shell has a different volume (volume), and the galaxy’s angular diameter distance, d A d_A d A ,
and hence its flux at the survey cut, changes (Doppler magnification; Bonvin 2008; Bacon et al. 2014), with 1 / ( x H ) − 1 = d ln d A / d ln ( 1 + z ) 1/(x\mathcal{H})-1=\mathrm{d}\ln d_A/\mathrm{d}\ln(1+z) 1/ ( x H ) − 1 = d ln d A / d ln ( 1 + z ) ).
The volume term depends only on redshift, so it is common to the two samples and enters Eq. (51) only multiplied by b 1 F − b 1 B b^{\rm F}_1-b^{\rm B}_1 b 1 F − b 1 B , whereas Doppler magnification enters through the difference in their magnification biases. Both of its factors, 1 / ( x H ) − 1 1/(x\mathcal{H})-1 1/ ( x H ) − 1 and Q B − Q F Q^{\rm B}-Q^{\rm F} Q B − Q F , are larger for BGS. Since d A ∝ z d_A\propto z d A ∝ z at low redshift, 1 / ( x H ) − 1 ≈ 1 / z 1/(x\mathcal{H})-1\approx1/z 1/ ( x H ) − 1 ≈ 1/ z , which is ≈ 6 \approx6 ≈ 6 in our lowest BGS redshift bin, whereas it vanishes at the maximum of d A d_A d A , z ≈ 1.6 z\approx1.6 z ≈ 1.6 , in the Euclid -like range. The magnification biases differ by Q B − Q F ≈ 0.9 Q^{\rm B}-Q^{\rm F}\approx0.9 Q B − Q F ≈ 0.9 –1.4 1.4 1.4 for BGS, compared with ≈ 0.25 \approx0.25 ≈ 0.25 –0.6 0.6 0.6 for the Euclid -like split. Doppler magnification therefore provides ≈ 95 % \approx95\% ≈ 95% of the BGS dipole in our lowest redshift bin and ≈ 80 % \approx80\% ≈ 80% in the highest, whereas the Euclid -like dipole is an order of magnitude smaller and set mainly by the difference in evolution biases (Figure 12 ). This 1 / z 1/z 1/ z growth is also why Doppler magnification dominates for sources living at low redshift (Bonvin 2008; Bacon et al. 2014; Bonvin et al. 2017), where it could also be detected by correlating BGS counts with galaxy sizes (Ye et al. 2026).
Figure 12. Top: contributions to the count slope, n e v o l n_{\rm evol} n evol (Eq. (54) ), of the DESI-like BGS (left) and Euclid -like Hα \alpha α (right) samples: the volume, d ln ( x 2 / H ) / d ln ( 1 + z ) \mathrm{d}\ln(x^2/\mathcal{H})/\mathrm{d}\ln(1+z) d ln ( x 2 / H ) / d ln ( 1 + z ) , the flux cut, − Q d ln L c / d ln ( 1 + z ) -Q\,\mathrm{d}\ln L_{\rm c}/\mathrm{d}\ln(1+z) − Q d ln L c / d ln ( 1 + z ) , where L c L_{\rm c} L c is the luminosity at the flux cut (including the K-correction for BGS), and the evolution, − b e -b_{\rm e} − b e . The kink in the top-left panel at z = 0.4 z=0.4 z = 0.4 arises because the mean colours of the Smith et al. (2017) colour model, which set the K-correction, are held constant above this redshift. Bottom: the terms of Eq. (52) in the dipole of the cross-power spectrum, Eq. (51) , divided by f ( H / k ) D 2 P f(\mathcal{H}/k)D^2P f ( H / k ) D 2 P , for the bright-faint splits of Figure 11 . The vertical dotted lines mark the centres of the redshift bins.
Removing the 1 / ( x H ) 1/(x\mathcal{H}) 1/ ( x H ) terms of γ 1 \gamma_1 γ 1 and γ 2 \gamma_2 γ 2 at fixed b e b_{\rm e} b e and Q Q Q , together with their counterparts in the second-order kernel, lowers the SNR of the multi-tracer power spectrum on the local relativistic terms alone, without the wide-separation corrections that raise it to ≈ 12.5 \approx12.5 ≈ 12.5 (Table 2 ), from ≈ 11 \approx11 ≈ 11 to ≈ 3 \approx3 ≈ 3 . These terms are not explicit in the count-slope form used in our forecasts (Eq. (18a)), as they are part of n e v o l n_{\rm evol} n evol (Eq. (54) ; Figure 12 , top).
The dependence of the BGS constraints on the split is shown in Figure 13 . For a fixed bright fraction, the SNR on α G R \alpha_{\rm GR} α GR depends only weakly on the split for 0.1 ≲ n B / n T ≲ 0.3 0.1\lesssim n_{\rm B}/n_{\rm T}\lesssim0.3 0.1 ≲ n B / n T ≲ 0.3 , peaking at n B / n T ≈ 0.15 n_{\rm B}/n_{\rm T}\approx0.15 n B / n T ≈ 0.15 , and falls by a factor of ≈ 3.5 \approx3.5 ≈ 3.5 by n B / n T = 0.9 n_{\rm B}/n_{\rm T}=0.9 n B / n T = 0.9 . A fixed bright fraction also outperforms every fixed-magnitude split we consider, by at least ∼ 15 % \sim15\% ∼ 15% , and we therefore split the BGS sample at n B / n T = 0.2 n_B/n_T = 0.2 n B / n T = 0.2 in Figure 10 and Figure 11 , whose SNR is within 1% of that of the best fraction and about a third higher than that of a split at r s = 19.5 r_s = 19.5 r s = 19.5 .
Figure 13. SNR on α G R \alpha_{\rm GR} α GR as a function of the bright fraction for a bright-faint split of a DESI-like BGS survey, for a multi-tracer power spectrum (green) and a multi-tracer joint power spectrum and bispectrum (gold) analysis, with n e v o l n_{\rm evol} n evol and Q Q Q held at their fiducial values, as in Figure 10 and Figure 11 . The solid lines split the sample at a fixed bright fraction, n B / n T n_B/n_T n B / n T , at every redshift, and the dashed lines at a fixed apparent magnitude, r s r_s r s (top axis), placed at the overall bright fraction of the survey, N B / N T N_B/N_T N B / N T , over all redshift bins. Odd multipoles are included.
With the biases free, the SNR at n B / n T = 0.2 n_B/n_T = 0.2 n B / n T = 0.2 falls below 1 (Table 2 ), so a detection requires prior knowledge of the relativistic biases.
Figure 14 shows how well the relativistic biases must be known to recover these constraints, for the BGS sample split at n B / n T = 0.2 n_B/n_T = 0.2 n B / n T = 0.2 . As for local-type f N L f_{\rm NL} f NL with the Euclid-like survey (Figure 8 ), the loss is driven mainly by Q Q Q , but it is far larger: with n e v o l n_{\rm evol} n evol fixed but Q Q Q free in each redshift bin, σ ( α G R ) \sigma(\alpha_{\rm GR}) σ ( α GR ) increases by a factor of ∼ 13 \sim13 ∼ 13 for the joint analysis, whereas with Q Q Q fixed but n e v o l n_{\rm evol} n evol free it increases by only ∼ 70 % \sim70\% ∼ 70% . This reflects the structure of the multi-tracer signal: for a fixed bright fraction the two samples share n e v o l n_{\rm evol} n evol , which therefore enters the leading-order dipole of the cross-power spectrum only multiplied by b 1 B − b 1 F ≈ 0.1 b_1^B-b_1^F \approx0.1 b 1 B − b 1 F ≈ 0.1 –0.4 0.4 0.4 , whereas Q B Q^B Q B and Q F Q^F Q F enter it directly. Coming within 5% of the fixed case requires σ ( Q ) / Q ≲ 4 % \sigma(Q)/Q \lesssim4\% σ ( Q ) / Q ≲ 4% with n e v o l n_{\rm evol} n evol fixed, or σ ( n e v o l ) ≲ 3 \sigma(n_{\rm evol}) \lesssim3 σ ( n evol ) ≲ 3 with Q Q Q fixed, in each redshift bin; given that ∣ n e v o l ∣ ≈ 4 |n_{\rm evol}| \approx4 ∣ n evol ∣ ≈ 4 and ≈ 17 \approx17 ≈ 17 in the upper two redshift bins, this is a much weaker requirement on n e v o l n_{\rm evol} n evol than on Q Q Q . The requirement on Q Q Q is comparable to the expected statistical precision of current finite-difference estimates (§ 5.2 ), so whether it can be met depends mainly on how well their systematic errors are controlled. These requirements are for recovering the fixed constraint, however, and a detection is far less demanding: an SNR of 5 requires the magnification bias to be known only to ≈ 25 % \approx25\% ≈ 25% in each redshift bin, even with n e v o l n_{\rm evol} n evol free.
Figure 14. Forecast SNR on the amplitude of local relativistic effects, α G R \alpha_{\rm GR} α GR , as a function of the widths of Gaussian priors on n e v o l n_{\rm evol} n evol and on the fractional magnification bias, σ ( Q ) / Q \sigma(Q)/Q σ ( Q ) / Q , in each redshift bin, for a DESI-like BGS survey split at a fixed bright fraction, n B / n T = 0.2 n_B/n_T = 0.2 n B / n T = 0.2 , from a joint multi-tracer power spectrum and bispectrum analysis, as in Figure 11 . The thick contour marks an SNR of 5. With n e v o l n_{\rm evol} n evol and Q Q Q held at their fiducial values the SNR is 12.6, compared with 0.94 with both free. The multi-tracer power spectrum alone gives a very similar picture, with SNRs of 12.5 and 0.82 with the biases fixed and free, respectively. The two samples share a single n e v o l n_{\rm evol} n evol in each redshift bin (Appendix A3). The prior on Q Q Q is a fractional error on the magnification biases of the total and bright samples, propagated to the faint sample through Eq. (56) . The prior on n e v o l n_{\rm evol} n evol is absolute, as n e v o l n_{\rm evol} n evol crosses zero at the boundary of the two lowest BGS redshift bins, z ≈ 0.23 z \approx0.23 z ≈ 0.23 : its fiducial values at the centres of the three BGS bins are 4.1, − 3.8 -3.8 − 3.8 and − 17.4 -17.4 − 17.4 .
Finally, the shaded bands for the BGS survey in Figure 11 show the SNR if we do not impose the wide-separation-based minimum k k k -cut of § 3.2 , i.e. if we retain modes with k < 2 π / x ( z min ) k < 2\pi/x(z_{\min}) k < 2 π / x ( z m i n ) . On these scales we keep the same signal model, including the wide-separation corrections, although the wide-separation expansion no longer converges there, so that the top of each band differs from the curve below it only in the k k k -bins it includes. These bands therefore give a rough indication of the signal that is inaccessible to the standard 3D Fourier statistics, but which could in principle be recovered in a basis that retains the full wide-separation information, such as the spherical Fourier–Bessel basis discussed in § 1 . The cut is relevant mainly for BGS (and SPHEREx), owing to their low-redshift coverage. For BGS it removes the first k k k -bin of the lowest redshift bin, yet this lowers the cumulative SNR by ∼ 34 % \sim34\% ∼ 34% for both the multi-tracer power spectrum and the joint multi-tracer analysis (and by ∼ 7 % \sim7\% ∼ 7% for the single-tracer joint analysis), as the relativistic corrections are largest on the largest scales. We note, however, that a more aggressive cut would likely be required in practice, so this should be regarded as a lower limit on the signal lost to the 3D Fourier basis.
Summary and Conclusions
Building on the formalism of Paper I and Paper II for relativistic and wide-separation effects in the bispectrum and power spectrum, implemented in CosmoWAP, this work has presented joint multi-tracer power spectrum and bispectrum forecasts of local, equilateral and orthogonal PNG and of the local relativistic effects themselves, for Stage IV (DESI-like BGS, Euclid-like Hα and SPHEREx-like) and Stage V (MegaMapper-like LBG) surveys. To this end we have derived the multi-tracer bispectrum multipole covariance, including the odd-parity multipoles, the leading non-Gaussian contributions and the cross-covariance with the power spectrum; the non-Gaussian terms matter mainly for flattened, squeezed triangles and largely cancel in the odd multipoles (§ 4 ). We have also modelled the second-order biases and the PNG biases of all three shapes with a peak-background split. We parametrise the relativistic biases by the count slope, n e v o l n_{\rm evol} n evol , and the magnification bias, Q Q Q , which, unlike the evolution bias, are observable properties of the sample.
Before we reiterate our key results, we stress that our forecasts are limited in scope and note the numerous shortcomings – many of which are discussed above, and which we briefly list below for clarity:
We restrict our analysis to tree-level perturbation theory, omitting higher-order non-linear physics and Fingers-of-God damping, and instead apply a minimal k k k -cut. While our forecasts generally focus on large-scale effects, this is less true for non-local PNG, which lacks the steep k k k -dependence of other effects. Furthermore, the bispectrum intrinsically correlates small and large scales, which is especially crucial in the squeezed limit.
Window convolution effects will be relevant for an analysis as they dampen signal on large scales and shift signal into different multipoles. Specifically, the odd-parity multipoles are particularly sensitive to leakage from the window-convolved even-parity signal.
We group all local relativistic effects together and do not attempt to detect individual terms but just use the combined contribution as an indicator for detectability of these effects.
All of our forecast results depend on the bias modelling choices we have made, in particular the linear bias, luminosity function and HOD.
We do not compute integrated contributions to the bispectrum – we could expect these to be relevant, as they are in the power spectrum (see Paper II), but we leave their computation to future work as they are non-trivial to include. The size of these contributions is unclear; neglecting them would likely lead to a bias in the inference on large scales, but how their inclusion affects the constraining power on a particular parameter depends on their shape, scale and redshift dependence and how that is degenerate with the parameter in question.
Luminosity function models that give the same number density agree closely on n e v o l n_{\mathrm{evol}} n evol but differ in Q Q Q by several times its statistical uncertainty. Therefore, if one is to take prior information on these terms from a model of the luminosity function then this must be done carefully, with particular caution taken for the magnification bias.
We ignore the cosmology dependence of our HODs (e.g. Smith et al. 2024) and luminosity functions.
As detailed in § 3 , we omit uncertainties for the PNG bias parameters, b Ψ b_{\Psi} b Ψ and b Ψ δ b_{\Psi\delta} b Ψ δ . We do not attempt to break their degeneracy with f N L f_{\mathrm{NL}} f NL ; instead, we hold them fixed to the universal mass function estimates. Much of the information, including for non-local shapes, comes from these biases, so the constraints should be regarded as optimistic.
Our non-Gaussian covariance is approximate (§ 4.3 ): the P T PT P T term uses a separable squeezed-limit form for all configurations, and the non-Gaussian terms are evaluated in the plane-parallel constant-redshift limit, without wide-separation contributions. We neglect integrated contributions to all of the covariance terms, since without a radial window function the lensing contribution to the power spectrum is sharply peaked at μ = 0 \mu= 0 μ = 0 , and the covariance strongly overweights this peak. We also neglect the connected 6-point function, non-Gaussian contributions to the power spectrum covariance, C P P P C_{PPP} C P P P , and super-sample covariance.
Thus, the ratio of the constraints with and without a proper treatment of these light-cone effects is often more relevant to the reader than their absolute size. With that in mind, our main results are as follows.
Biases from neglecting light-cone effects
We forecast that neglecting light-cone effects would bias the inference on local-type f N L f_{\mathrm{NL}} f NL for a joint power spectrum and bispectrum analysis for both a Euclid-like Hα \alpha α and a MegaMapper-like LBG survey, at ∼ 2 σ \sim2\sigma ∼ 2 σ and ∼ 25 σ \sim25\sigma ∼ 25 σ respectively, which could lead to a false detection of local-type PNG (Figure 3 ). This shift is driven by lensing, which pushes f N L l o c f_{\mathrm{NL}}^{\mathrm{loc}} f NL loc to positive values; neglecting only the local relativistic and wide-separation terms shifts it by ≲ 0.4 σ \lesssim0.4\sigma ≲ 0.4 σ . This result is perhaps straightforward for those used to angular analyses, but is less intuitively understood by the community when considering 3D Fourier statistics. The precise size of the shift depends on the survey window and covariance modelling: a radial window function roughly halves the MegaMapper-like shift, but it remains large and positive in every set-up we have tested. By contrast, the shift is negligible for our SPHEREx-like sample, whose lower redshift coverage and magnification bias suppress the lensing contribution. Since integrated contributions to the bispectrum are not included, the lensing contribution enters these forecasts through the power spectrum alone. For the MegaMapper-like survey, neglecting light-cone effects also biases the orthogonal shape by ∼ 3 σ \sim3\sigma ∼ 3 σ , and the cosmological parameters despite the Planck priors, by up to ∼ 10 σ \sim10\sigma ∼ 10 σ for A s A_s A s , while for the Euclid-like survey the shifts in the non-local shapes remain within the forecast errors. Lensing must therefore be modelled for precision PNG analyses, while the local relativistic terms also matter for the uncertainty on f N L f_{\mathrm{NL}} f NL , through the relativistic biases they introduce (§ 7.2 ).
Neglecting wide-separation corrections also biases the inferred amplitude of local relativistic effects, α G R \alpha_{\mathrm{GR}} α GR . For the bright-faint analyses with n e v o l n_{\mathrm{evol}} n evol and Q Q Q fixed, the shift is ≃ 9 \simeq9 ≃ 9 –10 σ 10\sigma 10 σ for the DESI-like BGS survey, mainly through the wide-angle terms, ≃ 3 σ \simeq3\sigma ≃ 3 σ for the Euclid-like survey and less than 1 σ 1\sigma 1 σ for the MegaMapper-like survey (Figure 10 ). The wide-separation corrections also carry relativistic signal of their own, and including them tightens the constraint on α G R \alpha_{\mathrm{GR}} α GR by ∼ 15 % \sim15\% ∼ 15% for BGS and the MegaMapper-like survey.
Local-type PNG and the relativistic biases
Modelling light-cone effects introduces n e v o l n_{\mathrm{evol}} n evol and Q Q Q as nuisance parameters in each redshift bin. Uncertainty in these biases must therefore be propagated into the PNG constraints, which are otherwise overly optimistic. In the even multipoles both biases enter at O ( H / k ) 2 O(\mathcal{H}/k)^2 O ( H / k ) 2 , with the same k − 2 k^{-2} k − 2 scaling as local-type PNG, so for the Euclid-like survey marginalising over them with no prior increases σ ( f N L l o c ) \sigma(f_{\mathrm{NL}}^{\mathrm{loc}}) σ ( f NL loc ) by ∼ 60 % \sim60\% ∼ 60% for the power spectrum and by ∼ 50 % \sim50\% ∼ 50% for a joint power spectrum and bispectrum analysis of the even multipoles (Figure 7 ).
The odd-parity multipoles, which carry the leading-order H / k \mathcal{H}/k H / k relativistic terms, recover most of this loss: including the bispectrum dipole reduces it to ∼ 15 % \sim15\% ∼ 15% by breaking the degeneracy between n e v o l n_{\mathrm{evol}} n evol and f N L l o c f_{\mathrm{NL}}^{\mathrm{loc}} f NL loc , without requiring a precise measurement of n e v o l n_{\mathrm{evol}} n evol itself. For the power spectrum alone, a bright-faint split at a fixed bright fraction, whose two samples share one n e v o l n_{\mathrm{evol}} n evol , recovers part of the loss, whereas a fixed-flux split, with a count slope for each sample, barely helps; neither adds much once the dipole is included.
The remaining loss is driven by Q Q Q , which is therefore the limiting nuisance parameter (Figure 8 ). For the Euclid-like survey, coming within 1 σ 1\sigma 1 σ of the ‘fixed’ constraint requires σ ( Q ) / Q ≲ 4 % \sigma(Q)/Q \lesssim4\% σ ( Q ) / Q ≲ 4% in each redshift bin, comparable to the statistical precision of current catalogue estimates. Priors forward-modelled from the luminosity function come within ∼ 5 % \sim5\% ∼ 5% , but depend on its functional form, as models with similar number densities disagree on Q Q Q . For the MegaMapper-like survey, the stronger lensing signal constrains Q Q Q from the clustering itself, so the joint constraint degrades by only ∼ 12 % \sim12\% ∼ 12% , and an external prior on Q Q Q helps appreciably only once it reaches the per-cent level.
What the bispectrum and the bright-faint split add for PNG
Both the bispectrum and the bright-faint split bring in further nuisance parameters (b 2 b_2 b 2 and b Γ 2 b_{\Gamma_2} b Γ 2 , and a second set of sample biases), so their additional information must outweigh the cost of marginalising over them. With the relativistic biases fixed, neither gains much on local-type f N L f_{\rm NL} f NL . The bispectrum improves the constraint by only a few per cent, as the squeezed triangles that carry the signal are strongly correlated with each other and with the power spectrum (a Gaussian covariance would give ∼ 12 % \sim12\% ∼ 12% ). The split gains ≲ 6 % \lesssim6\% ≲ 6% , as the bright and faint parts of a single population have similar b Ψ / b 1 b_{\Psi}/b_1 b Ψ / b 1 , which tends to 2 δ c 2\delta_c 2 δ c for highly biased tracers; larger gains would need distinct tracer populations (Seljak 2009). For local-type PNG, the main benefit of the bispectrum is instead its dipole, which breaks the degeneracy with the relativistic biases (§ 7.2 ). The bispectrum may also help to break the degeneracy between b Ψ b_{\Psi} b Ψ and f N L f_{\rm NL} f NL (§ 3.1.1 ).
The bispectrum is essential for competitive constraints on equilateral and orthogonal PNG: the Euclid -like survey gives σ ( f N L e q ) ≈ 34 \sigma(f_{\rm NL}^{\rm eq}) \approx34 σ ( f NL eq ) ≈ 34 and σ ( f N L o r t h ) ≈ 40 \sigma(f_{\rm NL}^{\rm orth}) \approx40 σ ( f NL orth ) ≈ 40 , improving on Planck by a factor of ∼ 1.4 \sim1.4 ∼ 1.4 for the equilateral shape but not for the orthogonal one. These shapes are much less degenerate with the relativistic terms: marginalising over n e v o l n_{\rm evol} n evol and Q Q Q loosens their joint power spectrum and bispectrum constraints by ≲ 2 % \lesssim2\% ≲ 2% for both the Euclid -like and MegaMapper-like surveys. For the MegaMapper-like survey, the split improves the joint power spectrum and bispectrum constraints on both shapes by a factor of ∼ 2 \sim2 ∼ 2 , to σ ( f N L e q ) ≈ 2.9 \sigma(f_{\rm NL}^{\rm eq}) \approx2.9 σ ( f NL eq ) ≈ 2.9 and σ ( f N L o r t h ) ≈ 4.9 \sigma(f_{\rm NL}^{\rm orth}) \approx4.9 σ ( f NL orth ) ≈ 4.9 (Table 1 ). The non-local constraints derive mainly from configurations near k max k_{\max} k m a x and are therefore particularly sensitive to our tree-level modelling.
Detecting local relativistic effects
For the surveys and scales considered, the single-tracer joint analyses give an SNR on α G R \alpha_{\rm GR} α GR below 1, even with n e v o l n_{\rm evol} n evol and Q Q Q fixed. With these biases fixed, a bright-faint split gives an SNR of ≈ 13 \approx13 ≈ 13 for the DESI-like BGS survey, driven mainly by the cross-power spectrum, compared with ≈ 1.5 \approx1.5 ≈ 1.5 –2.2 2.2 2.2 for the Euclid -like and MegaMapper-like surveys (Table 2 ). BGS does better despite its smaller volume mainly because of Doppler magnification, which grows as 1 / z 1/z 1/ z at low redshift, and the larger difference between the magnification biases of its bright and faint samples (Section 6.3). The adopted BGS split at a fixed bright fraction, n B / n T = 0.2 n_{\rm B}/n_{\rm T}=0.2 n B / n T = 0.2 , is near-optimal with the biases fixed and outperforms every fixed-magnitude split we consider (Figure 13 ).
Earlier forecasts, including Paper I, predicted a detectable single-tracer bispectrum dipole. The discrepancy can be linked to a sign error in the construction of the tree-level bispectrum in Umeh et al. (2017): the second-order kernel must be evaluated at − k 1 -\mathbf{k}_1 − k 1 and − k 2 -\mathbf{k}_2 − k 2 , though the kernel itself is unchanged. Correcting this strengthens cancellations between terms and reduces the dipole; the full non-Gaussian covariance and marginalising over nuisance parameters also contribute (§ 6 ).
Marginalising over n e v o l n_{\rm evol} n evol and Q Q Q with uninformative priors lowers the BGS joint SNR to ≈ 0.9 \approx0.9 ≈ 0.9 , so a detection requires independent information on these biases (Figure 14 ). This requirement falls mainly on Q Q Q . For a fixed-fraction cut the two samples share n e v o l n_{\rm evol} n evol , which therefore enters the leading dipole only through the small difference in their linear biases: with Q Q Q fixed, freeing n e v o l n_{\rm evol} n evol lowers the SNR only to ≈ 7 \approx7 ≈ 7 , whereas with Q Q Q free, fixing n e v o l n_{\rm evol} n evol makes almost no difference. An SNR of 5 requires σ ( Q ) / Q ≈ 25 % \sigma(Q)/Q \approx25\% σ ( Q ) / Q ≈ 25% in each redshift bin (even with n e v o l n_{\rm evol} n evol free), while recovering the fixed-bias SNR to within 5% requires σ ( Q ) / Q ≲ 4 % \sigma(Q)/Q \lesssim4\% σ ( Q ) / Q ≲ 4% with n e v o l n_{\rm evol} n evol fixed. The multi-tracer bispectrum helps to break the degeneracies between α G R \alpha_{\rm GR} α GR and the relativistic biases, raising the SNR without priors by ∼ 15 \sim15 ∼ 15 –75 % 75\% 75% , compared with ∼ 1 \sim1 ∼ 1 –11 % 11\% 11% with the biases fixed (Table 2 ), but not enough for a detection. Finally, for the largest scales in the lower redshift bins of the BGS survey one would expect the perturbative wide-separation expansion to break down. Removing these scales lowers the multi-tracer SNR by about a third even with a minimal k k k -cut, as the relativistic corrections are largest on the largest scales. Therefore, to capture the full information one should use alternative bases to the standard 3D Fourier summary statistics.
Outlook
Of the limitations listed above, the window function is the most pressing, both for the odd multipoles and because it must be treated before the integrated terms can be included in the covariance; we also leave the integrated contributions to the bispectrum to future work. Additionally, an alternative basis that retains the full wide-separation information, such as the spherical Fourier–Bessel basis, could in principle recover the full large-scale information where our perturbative wide-separation expansion breaks down.
Our results identify that determining the magnification bias of the sample is a priority for both local-type PNG inference and the detection of local relativistic effects. As properties of the observed sample, n e v o l n_{\rm evol} n evol and Q Q Q can be measured from the catalogue, n e v o l n_{\rm evol} n evol from the slope of the redshift distribution and Q Q Q by perturbing the flux cut and re-applying the sample selection, while the odd multipoles provide a complementary consistency test. For BGS, moderately informative constraints on Q Q Q can enable a significant detection; approaching the full forecast sensitivity requires measurements at the few-per-cent level, together with sufficient knowledge of n e v o l n_{\rm evol} n evol . Establishing these measurements with controlled selection and completeness errors is therefore an important next step.
Acknowledgements
CA is supported by a studentship from the UK Science and Technology Facilities Council (STFC). CC is supported by STFC (UK) grant ST/X000931/1. DK is supported by the Dutch Research Council (NWO) (grant no. OCENW.M.22.307). LLMs were used to improve the formatting and readability of the manuscript as well as debugging, optimisation and testing in CosmoWAP. The authors accept full responsibility for the accuracy and interpretation of the results, and the information presented in the text.
Code availability
CosmoWAP is publicly available at https://github.com/craddis1/CosmoWAP and the companion repository MathWAP is available at https://github.com/craddis1/MathWAP . The results in this paper were produced with CosmoWAP v0.10.0 (https://github.com/craddis1/CosmoWAP/tree/v0.10.0 ), which is archived on Zenodo (doi:10.5281/zenodo.24373137 ) and is also available on PyPI. A full description of CosmoWAP is given in Addis (2026). CosmoWAP utilises the public code CLASS (Lesgourgues 2011; Blas et al. 2011), and the following Python packages and libraries: NUMPY (Harris et al. 2020), SCIPY (Virtanen et al. 2020), MATPLOTLIB (Hunter 2007), NUMBA (Lam et al. 2015), CHAIN-CONSUMER (Hinton 2016), COSMOBA (Torrado & Lewis 2021) and COSMOPOWER (Spurio Mancini et al. 2022).
Data Availability
The results in this article are all derived using the public code CosmoWAP.
Bias Modelling
In this section, we outline some details of our bias modelling machinery. First, we provide an overview of the peak-background split (PBS) approach, and then we discuss our luminosity-function-dependent biases and the computation of these biases for our bright and faint samples.
The PBS approach splits the density field into a long-wavelength mode and noisy, short-wavelength fluctuations that sit atop it. In this picture, an overdense, long ‘background’ mode effectively lowers the local threshold for gravitational collapse. The halo clustering bias parameters can be derived from the derivatives of the halo mass function (HMF) with respect to the background density—see Desjacques et al. (2018) and Appendix C of Karagiannis et al. (2018) for a detailed review. Thus, given a cosmology and a theoretical HMF model18 , we can compute the halo bias parameters. These can then be related to the galaxy bias parameters, as detailed in § 2.3 .
Magnification and evolution bias
The magnification bias, Q Q Q , relates to the galaxies lost or gained by a sample at a certain flux or magnitude cut due to perturbations of their observed flux, such as lensing magnification and Doppler effects, while the evolution bias, b e b_e b e , describes the physical change in number density due to formation and mergers. These biases can be expressed as derivatives of the number density and therefore are closely related to the survey luminosity function.
For a flux- (or magnitude-)limited survey, the evolution, b e b_e b e , and magnification, Q Q Q , biases are defined with partial derivatives at the luminosity cut L c L_c L c (Challinor & Lewis 2011; Alonso et al. 2015)
b e ( z , L c ) = − ∂ ln n g ∂ ln ( 1 + z ) ∣ L c , Q ( z , L c ) = − ∂ ln n g ∂ ln L ∣ z , (53) b_e(z,L_c)=-\left.\frac{\partial\ln n_g}{\partial\ln(1+z)}\right|_{L_c},\qquad Q(z,L_c)=-\left.\frac{\partial\ln n_g}{\partial\ln L}\right|_z, \tag*{(53)} b e ( z , L c ) = − ∂ ln ( 1 + z ) ∂ ln n g L c , Q ( z , L c ) = − ∂ ln L ∂ ln n g z , (53)
where n g n_g n g is the comoving number density and x ∣ x \left.\vphantom{x}\right|_x x ∣ x refers to what is being held fixed; see Maartens et al. (2021b) for a more detailed overview.
While Q Q Q can be measured directly from the slope of the source counts at the flux cut, b e b_e b e cannot, as it is defined at fixed luminosity rather than fixed flux. The corresponding observable is the slope of the redshift distribution of the sample, n e v o l ≡ d ln N / d ln ( 1 + z ) n_{\mathrm{evol}}\equiv\mathrm{d}\ln N/\mathrm{d}\ln(1+z) n evol ≡ d ln N / d ln ( 1 + z ) Blanco et al. (2024), where N = χ 2 n g / H N=\chi^2n_g/H N = χ 2 n g / H is the number of sources above the flux cut per unit redshift and solid angle, and H H H is the Hubble rate. Accounting for the evolution of the luminosity cut and of the volume factor, the two are related by (Alonso et al. 2015; Maartens et al. 2021b)
b e = − n e v o l − d ln H d ln ( 1 + z ) + 2 χ H − 2 ( 1 + 1 χ H ) Q , (54) b_e=-n_{\mathrm{evol}}-\frac{\mathrm{d}\ln H}{\mathrm{d}\ln(1+z)}+\frac{2}{\chi H}-2\left(1+\frac{1}{\chi H}\right)Q, \tag*{(54)} b e = − n evol − d ln ( 1 + z ) d ln H + χH 2 − 2 ( 1 + χH 1 ) Q , (54)
where, for surveys that measure fluxes in fixed wavelength bands, the last term also contains a K-correction (Maartens et al. 2021b). We therefore take n e v o l n_{\mathrm{evol}} n evol and Q Q Q as the parameters describing each sample, with b e b_e b e following from Eq. (54) , such that it varies with cosmology through the remaining terms.
Evolution bias and the PNG bias
The PNG bias is the response of the number density to the amplitude of the small-scale fluctuations, b Ψ = 2 ∂ ln n g / ∂ ln σ 8 b_\Psi=2\partial\ln n_g/\partial\ln\sigma_8 b Ψ = 2 ∂ ln n g / ∂ ln σ 8 (Section 2.1.2). At fixed cosmology σ R ∝ D ( z ) \sigma_R\propto D(z) σ R ∝ D ( z ) , the linear growth factor, so the redshift evolution of the number density carries the same response [115 ] (Sullivan & Seljak 2025). Since b e b_e b e is defined at fixed luminosity (Eq. (A1)), with d ln D = − f d ln ( 1 + z ) \mathrm{d}\ln D=-f\,\mathrm{d}\ln(1+z) d ln D = − f d ln ( 1 + z ) it is the sum of this response and the evolution of the population at fixed σ R \sigma_R σ R Nascimento & Dalal [115 ] , so that
b Ψ = 2 b e f − 2 ∂ ln n g ∂ ln D ∣ σ R , L c . (55) b_\Psi=\frac{2b_e}{f}-2\left.\frac{\partial\ln n_g}{\partial\ln D}\right|_{\sigma_R,L_c}. \tag*{(55)} b Ψ = f 2 b e − 2 ∂ ln D ∂ ln n g σ R , L c . (55)
If the last term vanishes, this gives b e = f b Ψ / 2 b_e=f b_\Psi/2 b e = f b Ψ /2 , and so b e = δ c f ( b 1 − 1 ) b_e=\delta_c f(b_1-1) b e = δ c f ( b 1 − 1 ) for a universal mass function (Jeong et al. 2012; Sullivan & Seljak 2025).
Written in terms of n e v o l n_{\mathrm{evol}} n evol and Q Q Q through Eq. (54) , Eq. (55) is the time-evolution estimator of Sullivan & Seljak (2025) and Nascimento & Dalal [115 ] applied to the count slope of a flux-limited sample. The Q Q Q term in Eq. (54) , together with the K-correction, accounts for the drift of the luminosity cut with redshift at fixed flux, which the evolution bias excludes by definition. Sullivan & Seljak (2025) give this correction in terms of Q Q Q without the K-correction (their equation 22), while Nascimento & Dalal [115 ] evaluate it, including the K-correction, by shifting the spectral energy distributions of the galaxies (their equation 13). In our parametrisation it is set by the same Q Q Q that enters the relativistic clustering signal.
The last term in Eq. (55) is the evolution of the population at fixed luminosity that is not driven by the amplitude of the fluctuations, such as the passive fading of stellar populations, feedback and mergers Nascimento & Dalal [115 ] , and, for emission-line samples such as the Euclid-like Hα \alpha α sample, the decline of the cosmic star formation rate. It must be modelled, for example with simulations, for any time-evolution estimate of b Ψ b_\Psi b Ψ , and we do not attempt this here.
Applied to the catalogue, the time-evolution estimator shares the limitation of the catalogue measurement of n e v o l n_{\mathrm{evol}} n evol (Section 5.2): any unmodelled redshift dependence of the completeness of the sample biases it. The leading-order relativistic terms in the odd multipoles instead depend on b e b_e b e and Q Q Q directly, and so offer a route to b Ψ b_\Psi b Ψ through Eq. (55) that does not rely on the redshift distribution. For the Euclid-like survey, however, the clustering constraint on n e v o l n_{\mathrm{evol}} n evol is weak (Section 5.2), and since Q Q Q is much better determined, the constraint on b e b_e b e is similarly weak, so this route would at most provide a consistency check. Throughout our forecasts b Ψ b_\Psi b Ψ is held fixed (Section 3) and Eq. (55) is not imposed but CosmoWAP has been extended to allow for forecasts using it.
Bright-faint split
The treatment of the biases for the bright sample is reasonably straightforward: they can be treated as a single-tracer survey with a more conservative luminosity threshold. The biases for the faint sample are less simple, as one now has to consider both the flux cut and the flux split. For most biases, including the magnification bias, the faint sample bias can be derived from the biases of the total and bright samples (e.g. Ferramacho et al. 2014; Bonvin et al. 2023)
b F = n g T b T − n g B b B n g T − n g B , (56) b^F=\frac{n_g^T b^T-n_g^B b^B}{n_g^T-n_g^B}, \tag*{(56)} b F = n g T − n g B n g T b T − n g B b B , (56)
which, as the faint counts are simply N F = N T − N B N^F=N^T-N^B N F = N T − N B , also holds for n e v o l n_{\mathrm{evol}} n evol . However, relating the evolution bias of the faint sample to n e v o l F n_{\mathrm{evol}}^F n evol F is slightly more involved, as the faint sample is bounded by two luminosity cuts, and we detail this below for a flux-limited survey.
The evolution bias of the faint tracer population is the partial derivative of the faint number density at fixed luminosity cuts
b e F = − ∂ ln ( n g T − n g B ) ∂ ln ( 1 + z ) ∣ L c , L s . (57) b_{\mathrm{e}}^{F}=-\left.\frac{\partial\ln\left(n_{\mathrm{g}}^{T}-n_{\mathrm{g}}^{B}\right)}{\partial\ln(1+z)}\right|_{L_{\mathrm{c}},L_{\mathrm{s}}}. \tag*{(57)} b e F = − ∂ ln ( 1 + z ) ∂ ln ( n g T − n g B ) L c , L s . (57)
As for Eq. (54) , we relate this to the total derivative of the observed faint counts, N F = N T − N B N^{F}=N^{T}-N^{B} N F = N T − N B , which includes the volume factor and the redshift-dependent luminosity cuts. With upper and lower luminosity cuts, this can be written as
b e v o l F = − b e F − d ln H d ln ( 1 + z ) + 2 x H + ∂ ln n g F ∂ ln L c ∣ z , L s d ln L c d ln ( 1 + z ) + ∂ ln n g F ∂ ln L s ∣ z , L c d ln L s d ln ( 1 + z ) . (58) \begin{aligned}
b_{\mathrm{evol}}^{F}=-b_{\mathrm{e}}^{F}-\frac{\mathrm{d}\ln H}{\mathrm{d}\ln(1+z)}+\frac{2}{x\mathcal{H}}+\left.\frac{\partial\ln n_{\mathrm{g}}^{F}}{\partial\ln L_{\mathrm{c}}}\right|_{z,L_{\mathrm{s}}}\frac{\mathrm{d}\ln L_{\mathrm{c}}}{\mathrm{d}\ln(1+z)} \\
&+\left.\frac{\partial\ln n_{\mathrm{g}}^{F}}{\partial\ln L_{\mathrm{s}}}\right|_{z,L_{\mathrm{c}}}\frac{\mathrm{d}\ln L_{\mathrm{s}}}{\mathrm{d}\ln(1+z)}. \tag*{(58)}
\end{aligned} b evol F = − b e F − d ln ( 1 + z ) d ln H + x H 2 + ∂ ln L c ∂ ln n g F z , L s d ln ( 1 + z ) d ln L c + ∂ ln L s ∂ ln n g F z , L c d ln ( 1 + z ) d ln L s . (58)
Using the fact that n g T n_{\mathrm{g}}^{T} n g T and n g B n_{\mathrm{g}}^{B} n g B are independent of L s L_{\mathrm{s}} L s and L c L_{\mathrm{c}} L c respectively, we can rewrite the partial derivatives
∂ ln n g F ∂ ln L c ∣ z , L s = n g T n g F ∂ ln n g T ∂ ln L c ∣ z , L s , \left.\frac{\partial\ln n_{\mathrm{g}}^{F}}{\partial\ln L_{\mathrm{c}}}\right|_{z,L_{\mathrm{s}}}=\left.\frac{n_{\mathrm{g}}^{T}}{n_{\mathrm{g}}^{F}}\frac{\partial\ln n_{\mathrm{g}}^{T}}{\partial\ln L_{\mathrm{c}}}\right|_{z,L_{\mathrm{s}}}, ∂ ln L c ∂ ln n g F z , L s = n g F n g T ∂ ln L c ∂ ln n g T z , L s ,
∂ ln n g F ∂ ln L s ∣ z , L c = − n g B n g F ∂ ln n g B ∂ ln L s ∣ z , L c . (59) \left.\frac{\partial\ln n_{\mathrm{g}}^{F}}{\partial\ln L_{\mathrm{s}}}\right|_{z,L_{\mathrm{c}}}=-\left.\frac{n_{\mathrm{g}}^{B}}{n_{\mathrm{g}}^{F}}\frac{\partial\ln n_{\mathrm{g}}^{B}}{\partial\ln L_{\mathrm{s}}}\right|_{z,L_{\mathrm{c}}}. \tag*{(59)} ∂ ln L s ∂ ln n g F z , L c = − n g F n g B ∂ ln L s ∂ ln n g B z , L c . (59)
For a split at fixed flux, both luminosity cuts evolve at the same rate, d ln L c / d ln ( 1 + z ) = d ln L s / d ln ( 1 + z ) = 2 [ 1 + 1 / ( x H ) ] \mathrm{d}\ln L_{\mathrm{c}}/\mathrm{d}\ln(1+z)=\mathrm{d}\ln L_{\mathrm{s}}/\mathrm{d}\ln(1+z)=2[1+1/(x\mathcal{H})] d ln L c / d ln ( 1 + z ) = d ln L s / d ln ( 1 + z ) = 2 [ 1 + 1/ ( x H )] . Thus, using the definition of the magnification bias in Eq. (53) , the evolution bias of the faint sample is given by Eq. (54) with the count slope and magnification bias of the faint sample, n e v o l F n_{\mathrm{evol}}^{F} n evol F and Q F Q^{F} Q F . As before, for galaxy surveys, like DESI BGS, that measure fluxes in fixed wavelength bands, the magnification bias term also contains a K-correction. If the split is instead chosen to keep a fixed fraction of bright galaxies at each redshift, the split flux varies with redshift, the two cuts no longer evolve at the same rate, and both the bright and faint evolution biases pick up an additional term proportional to Q B d ln F s / d ln ( 1 + z ) Q^{B}\,\mathrm{d}\ln F_{\mathrm{s}}/\mathrm{d}\ln(1+z) Q B d ln F s / d ln ( 1 + z ) (Blanco et al. 2024). For such a split the bright and faint counts are fixed fractions of the total, N B ∝ N F ∝ N T N^{B}\propto N^{F}\propto N^{T} N B ∝ N F ∝ N T , so all three samples share the same count slope, n e v o l n_{\mathrm{evol}} n evol , and their evolution biases differ only through Q B Q^{B} Q B , Q F Q^{F} Q F and this additional term. In our forecasts with a fixed bright fraction we therefore marginalise over a single n e v o l n_{\mathrm{evol}} n evol per redshift bin, shared by the two samples.
Luminosity function priors
The count slope and magnification bias are the logarithmic slopes of the number of sources above the survey flux cut with respect to redshift and luminosity, respectively (Eqs. (53) and (54) ), and are therefore fixed by the survey luminosity function, so existing constraints on the luminosity function can be turned into prior information on n e v o l n_{\mathrm{evol}} n evol and Q Q Q . To propagate the luminosity function uncertainty into these biases, we adopt a Monte Carlo forward model, following Wang et al. (2020a): we draw realisations of the luminosity function parameters from their fitted distributions, compute n e v o l n_{\mathrm{evol}} n evol and Q Q Q at the survey flux cut for each draw, and use the resulting mean and covariance as a Gaussian prior in each redshift bin. For the Euclid -like Hα \alpha α survey we use the ‘Model 3’ luminosity function of Pozzetti et al. (2016), with the best-fit parameters and uncertainties given in their Table 2. Since each draw is a realisation of the full luminosity function, this approach automatically captures the correlation between n e v o l n_{\mathrm{evol}} n evol and Q Q Q , and, for a bright-faint split, between all of the relativistic biases of the two samples, { n e v o l B , Q B , n e v o l F , Q F } \{n_{\mathrm{evol}}^{B},Q^{B},n_{\mathrm{evol}}^{F},Q^{F}\} { n evol B , Q B , n evol F , Q F } or { n e v o l , Q B , Q F } \{n_{\mathrm{evol}},Q^{B},Q^{F}\} { n evol , Q B , Q F } for a fixed bright fraction, for which the two samples share the same count slope; Appendix A3, which all derive from the same underlying luminosity function. For a single tracer, the resulting prior has σ ( n e v o l ) \sigma(n_{\mathrm{evol}}) σ ( n evol ) falling from ≃ 1.9 \simeq1.9 ≃ 1.9 to ≃ 1.0 \simeq1.0 ≃ 1.0 across our redshift bins and σ ( Q ) ≃ 0.3 \sigma(Q)\simeq0.3 σ ( Q ) ≃ 0.3 (10–14% of its fiducial value), with a correlation of ≃ − 0.8 \simeq-0.8 ≃ − 0.8 between the two in each bin.
These priors are only as reliable as the luminosity function model from which they are derived, and their widths include only the statistical uncertainty on the fitted parameters of a single model. With Euclid ’s far greater area and depth, the statistical uncertainty on the luminosity function should shrink considerably, leaving systematics19 and the choice of functional form as the dominant contributions to the error budget. The latter is especially important here: n e v o l n_{\mathrm{evol}} n evol and Q Q Q are derivatives of the luminosity function, so mismodelling of its shape is amplified rather than averaged over. Indeed, Guandalin et al. (2023) showed that quasar luminosity functions that are all consistent with the same data can yield significantly different relativistic biases.
To illustrate this for our Hα \alpha α sample, we forward model n e v o l n_{\mathrm{evol}} n evol and Q Q Q from all three luminosity function models of Pozzetti et al. (2016). To isolate the effect of the shape of the luminosity function, we require that all three models share the same number density at every redshift, which we achieve via a redshift-dependent flux cut for Models 1 and 2 (so that Q Q Q is necessarily evaluated at a different flux cut for each model). The results are shown in Figure A1 . Since matching the number density largely fixes the redshift distribution, the three models agree closely on n e v o l n_{\mathrm{evol}} n evol , to within ≃ 0.3 \simeq0.3 ≃ 0.3 and well inside the ‘Model 3’ prior, apart from a sharp feature near z ≃ 1.3 z\simeq1.3 z ≃ 1.3 that arises from the break in the broken power law of Model 1. Their magnification biases, however, are mutually inconsistent, particularly at higher redshift: by z ≃ 1.8 z\simeq1.8 z ≃ 1.8 Models 1 and 3 differ by Δ Q ≃ 0.75 \Delta Q\simeq0.75 Δ Q ≃ 0.75 , several times the statistical width of the ‘Model 3’ prior, which through Eq. (54) corresponds to Δ b e ≃ 3 \Delta b_{\mathrm{e}}\simeq3 Δ b e ≃ 3 .
Figure A1. Count slope n e v o l n_{\mathrm{evol}} n evol (left) and magnification bias (right) derived from the three Hα luminosity function models of Pozzetti et al. (2016), with the bands showing the 1 σ 1\sigma 1 σ uncertainty forward-modelled from the uncertainties on each model’s fitted parameters. The flux cuts for Models 1 and 2 are chosen as a function of redshift such that all three models give the same comoving number density. The feature in n e v o l n_{\mathrm{evol}} n evol for Model 1 at z ≈ 1.3 z \approx1.3 z ≈ 1.3 arises from its broken power law form.
The model-to-model scatter in Q Q Q therefore exceeds the statistical uncertainty on which our priors are based, and the forecasts with these priors should be read as optimistic. Elkhashab et al. (2022) similarly found that, for a Euclid -like Hα \alpha α survey, the detectability of lensing in the power spectrum monopole depends on the faint-end slope of the luminosity function. We leave a systematic treatment of luminosity function modelling uncertainty to future work.
Survey details
In this appendix we describe the modelling and assumptions for each survey. For MegaMapper and Euclid these set-ups are similar to those described in Paper II; however, we also extend the modelling to the second-order and PNG biases using a peak-background split approach. We outline the HODs we use and then briefly describe the remaining assumptions for each survey.
Halo occupation distributions
In our forecasts we use two different HOD models which are briefly described below:
Two-parameter HOD
The forecasts for the Euclid -like Hα \alpha α , MegaMapper-like, and SPHEREx-like surveys use an HOD, ⟨ N ( M , z ) ⟩ \langle N(M,z)\rangle ⟨ N ( M , z )⟩ , following Yankele vich & Porciani (2019), based on earlier work from Geach et al. (2012); Gonzalez-Perez et al. (2018), with two free parameters, M 0 M_0 M 0 and N H O N_{\mathrm{HO}} N HO . These are fitted to the linear bias and number density models at each redshift, so that both are functions of redshift; M 0 M_0 M 0 can be fitted independently, as the linear bias does not depend on N H O N_{\mathrm{HO}} N HO . See CosmoWAP for implementation details and best-fit parameters for different cosmologies and luminosity cuts.
DESI-like BGS HOD
The DESI-like BGS survey uses a five-parameter HOD model based on Zheng et al. (2005); Zehavi et al. (2011) and implemented by Smith et al. (2017). We adopt the best-fitting HOD parameters of Smith et al. (2024), which are smooth functions of the absolute magnitude threshold in 0.1 M r ^{0.1}M_r 0.1 M r , the r r r band K-corrected to z = 0.1 z = 0.1 z = 0.1 . These were fitted on the z = 0.2 z = 0.2 z = 0.2 snapshot of the AbacusSummit simulations (Maksimova et al. (2021)) to reproduce both the clustering and the number densities of magnitude-threshold samples in the mock catalogue of Smith et al. (2017), which was itself tuned on SDSS and GAMA data. We extend the fits to other redshifts with the method of Smith et al. (2017), following the public hodpy code; at each redshift a target luminosity function, the evolving Schechter fit to GAMA of Loveday et al. (2012), gives the number density of galaxies brighter than the absolute magnitude threshold. The threshold is mapped to the one at z = 0.2 z = 0.2 z = 0.2 with the same target number density and the HOD parameters are taken there, so that the shape of the HOD is fixed at fixed number density; its three mass parameters are then scaled by a common factor until the HOD reproduces the target number density on our halo mass function. The number density, Q Q Q and n e v o l n_{\mathrm{evol}} n evol are therefore those of the target luminosity function, while the HOD sets the clustering biases. Unlike hodpy, we use the GAMA luminosity function at all redshifts rather than blending to an SDSS one below z ≈ 0.15 z \approx0.15 z ≈ 0.15 , as that sharp switch produces a spurious feature in n e v o l n_{\mathrm{evol}} n evol .
Survey specifications
DESI BGS
To forecast for a DESI-like BGS survey we use a three-bin, 15,000 deg2 ^2 2 survey over the redshift range ( 0.1 < z < 0.5 ) (0.1 < z < 0.5) ( 0.1 < z < 0.5 ) . We fix the r r r -band apparent magnitude cut-off to m c = 20.175 m_c = 20.175 m c = 20.175 and use the colour-dependent GAMA K-corrections of Smith et al. (2017), averaged over the colours of their mock galaxies at the magnitude limit, so that a single K-correction is applied at each redshift, to the bright and faint samples alike. For r < 19.8 r < 19.8 r < 19.8 the resulting redshift distribution agrees with the GAMA fit of Smith et al. (2017) to within 3% for 0.1 < z < 0.4 0.1 < z < 0.4 0.1 < z < 0.4 . K-correcting each galaxy by its own colour, as the mocks do, would instead give a Q Q Q that is ∼ 5 \sim5 ∼ 5 –10 % 10\% 10% larger for z ≳ 0.3 z \gtrsim0.3 z ≳ 0.3 , so our relativistic biases are more conservative at these redshifts.
Euclid
We consider a five-bin, 15,000 deg2 ^2 2 Euclid-like survey ( 0.9 < z < 1.8 ) (0.9 < z < 1.8) ( 0.9 < z < 1.8 ) . Along with the two-parameter HOD, we adopt the ‘Model 3’ Hα luminosity function (Pozzetti et al. 2016; Euclid Collaboration et al. 2020) and use a flux cut of F c = 2 × 10 − 16 [ e r g c m − 2 s − 1 ] F_c = 2 \times10^{-16}\,[\mathrm{erg\ cm^{-2}\ s^{-1}}] F c = 2 × 1 0 − 16 [ erg c m − 2 s − 1 ] . We also adopt a phenomenological fit for the linear bias,
b 1 ( x , z ) = a + b ( 1 + z ) e ( 1 + exp ( ( x − c ) d ) ) (60) b_1(x,z) = a + b(1+z)^e(1+\exp((x-c)d)) \tag*{(60)} b 1 ( x , z ) = a + b ( 1 + z ) e ( 1 + exp (( x − c ) d )) (60)
as a function of redshift, z z z , and luminosity, x = log 10 ( L / [ e r g s − 1 ] ) x = \log_{10}(L/[\mathrm{erg\ s^{-1}}]) x = log 10 ( L / [ erg s − 1 ]) . The fitted parameters are taken from Pan et al. (2020), Table 1, their fit to the differential bias of galaxies at luminosity x x x . These fits derive from the semi-analytic galaxy formation model GALFORM (Cole et al. 2000; Gonzalez-Perez et al. 2018), whose cumulative linear galaxy bias above the luminosity cut, L c L_c L c , is then given by Pan et al. (2020), eq. 15)
b 1 ( ≥ L c ) = ∫ L c + ∞ d L b 1 ( L , z ) ϕ ( L , z ) ∫ L c + ∞ d L ϕ ( L , z ) (61) b_1(\geq L_c) = \frac{\int_{L_c}^{+\infty}\mathrm{d}L\ b_1(L,z)\phi(L,z)}{\int_{L_c}^{+\infty}\mathrm{d}L\ \phi(L,z)} \tag*{(61)} b 1 ( ≥ L c ) = ∫ L c + ∞ d L ϕ ( L , z ) ∫ L c + ∞ d L b 1 ( L , z ) ϕ ( L , z ) (61)
where ϕ ( L , z ) \phi(L,z) ϕ ( L , z ) is the Hα luminosity function. By incorporating b 1 b_1 b 1 and the comoving number density, n g n_g n g , derived from the luminosity function, we can constrain the two-parameter HOD to model the remaining bias parameters. See Appendix A4 for how we forward model the uncertainty in the H α \mathrm{H}\alpha H α luminosity function into priors on n e v o l n_{\mathrm{evol}} n evol and Q Q Q .
MegaMapper
For the MegaMapper-like forecasts we consider a ten-bin, Stage V 20,000 deg2 ^2 2 idealised Lyman-break galaxy (LBG) survey over (2 < z < 5 2 < z < 5 2 < z < 5 ). To do this, we follow Wilson & White (2019) and use a UV Schechter luminosity function (Schechter (1976); Ono et al. (2018)) with an apparent magnitude cut, m c = 24.5 m_{\mathrm{c}} = 24.5 m c = 24.5 . For the linear clustering bias, we adopt the redshift and apparent magnitude dependent function from Wilson & White (2019), of the form,
b ( z , m ) = A ( m ) ( 1 + z ) + B ( m ) ( 1 + z ) 2 (62) b(z,m) = A(m)(1+z) + B(m)(1+z)^2 \tag*{(62)} b ( z , m ) = A ( m ) ( 1 + z ) + B ( m ) ( 1 + z ) 2 (62)
where A ( m ) = − 0.98 ( m − 25 ) + 0.11 A(m)=-0.98(m-25)+0.11 A ( m ) = − 0.98 ( m − 25 ) + 0.11 and B ( m ) = 0.12 ( m − 25 ) + 0.17 B(m)=0.12(m-25)+0.17 B ( m ) = 0.12 ( m − 25 ) + 0.17 . As this is a fit to the measured bias of magnitude-limited samples, we evaluate it directly at the survey cut, b 1 ( z ) = b ( z , m c ) b_1(z)=b(z,m_{\mathrm{c}}) b 1 ( z ) = b ( z , m c ) . As in the Euclid -like H α \mathrm{H}\alpha H α survey, we constrain the HOD using this linear bias model and a number density derived from the luminosity function for a given apparent magnitude cut. Absolute magnitudes are obtained with a K-correction, K ( z ) = − 2.5 log 10 ( 1 + z ) K(z)=-2.5\log_{10}(1+z) K ( z ) = − 2.5 log 10 ( 1 + z ) , i.e. the bandwidth term alone with no colour correction, as the observed band tracks the rest-frame UV.
SPHEREx
For the SPHEREx -like forecasts we consider a ten-bin, all-sky survey over (0.1 < z < 4.3 0.1 < z < 4.3 0.1 < z < 4.3 ), retaining 75% of the sky (∼ 31,000 d e g 2 \sim31{,}000\ \mathrm{deg}^2 ∼ 31 , 000 deg 2 ) (Doré et al. (2014)). We use only the highest-accuracy redshift sample, σ z / ( 1 + z ) < 0.003 \sigma_z/(1+z)<0.003 σ z / ( 1 + z ) < 0.003 . This is partly why our constraints on local-type f N L f_{\mathrm{NL}} f NL are weaker than previous SPHEREx forecasts, σ ( f N L ) ≈ 0.9 \sigma(f_{\mathrm{NL}})\approx0.9 σ ( f NL ) ≈ 0.9 from the power spectrum (Doré et al. (2014)) and ≈ 0.7 \approx0.7 ≈ 0.7 from the bispectrum (Heinrich et al. (2024)), which combine all five redshift-accuracy samples as multiple tracers. The power spectrum constraint is driven by the larger samples with less precise redshifts, up to σ z / ( 1 + z ) = 0.2 \sigma_z/(1+z)=0.2 σ z / ( 1 + z ) = 0.2 (Doré et al. (2014)), and even for the bispectrum, keeping only the two most precise samples weakens the constraint by ∼ 90 % \sim90\% ∼ 90% (Heinrich et al. (2024)). The comoving number density and the abundance-matched linear bias of each subsample are taken directly from the public SPHEREx GitHub, and we fit the latter with
b 1 ( z ) = A ( 1 + β z ) γ . (63) b_1(z)=A(1+\beta z)^\gamma. \tag*{(63)} b 1 ( z ) = A ( 1 + β z ) γ . (63)
As for the Euclid -like and MegaMapper-like surveys, we constrain the two-parameter HOD with this linear bias and number density to model the remaining bias parameters. Following Lukens & Jeong (2026), we adopt the 2.4 μ \mu μ m WISE continuum luminosity function of Lake et al. (2018), taking their canonical ‘High z z z Prior’ fit together with an effective flux cut of ν F ν = 2 × 10 − 16 [ e r g c m − 2 s − 1 ] \nu F_\nu= 2 \times10^{-16}\ [\mathrm{erg}\ \mathrm{cm}^{-2}\ \mathrm{s}^{-1}] ν F ν = 2 × 1 0 − 16 [ erg cm − 2 s − 1 ] at 2.4 μ \mu μ m. Since this is a continuum rather than a line luminosity, the limiting luminosity also carries a K-correction; we adopt the same K-correction as in the MegaMapper case and neglect the colour term. We note that this luminosity function is calibrated on z ≲ 1 z \lesssim1 z ≲ 1 data, so its use over the full SPHEREx range is an extrapolation.
Additional Plots
This paper has been typeset from a \TeX/\LaTeX file prepared by the author.