betaFeedback

WorksPhysics and Astronomy

Stratonovich approximation for halo bias: An efficient framework within excursion set theory

Abstract

We develop an efficient framework for calculating halo multiplicity functions and bias parameters in the excursion-set approach based on the previously presented Hertz and Stratonovich approximations for the first-crossing distribution. The framework enables efficient evaluation of first-crossing distributions for a broad class of moving barriers and filter functions, from which halo biases can be calculated within the peak-background split formalism. We compare halo multiplicity functions and biases obtained from these approximations with those computed from first-crossing distributions obtained from numerically generated ensembles of excursion-set trajectories. We find close agreement between predictions based on the Stratonovich approximation and numerical results across a range of barrier parameters and filter choices, demonstrating that the approximation captures the impact of non-Markovian correlations on first-crossing statistics while substantially reducing the computational cost compared to trajectory-based calculations. Overall, this framework provides a flexible and computationally efficient tool for analyzing halo statistics in cosmological applications.

Introduction

The wealth of observational data from modern galaxy surveys provides unique opportunities for testing different cosmological models and identifying those that best describe our universe. For this purpose, galaxy-halo connection models (e.g. Wechsler & Tinker (2018)) can be used to connect the observed properties and clustering of galaxies to the properties and spatial distribution of their host dark matter halos, whose abundance and clustering are modeled within an underlying cosmological framework. Commonly used methods model the abundance and clustering of dark matter halos at a given redshift in terms of the (statistical) properties of the dark matter density fluctuation field that is linearly evolved from an initial state (e.g. Press & Schechter (1974); Bardeen et al. (1986); Peacock & Heavens (1990); Bond et al. (1991); Mo & White (1996); Sheth & Tormen (1999, 2002); Zentner (2007); Maggiore & Riotto (2010a,b); Ma et al. (2011); Corasaniti & Achitouv (2011a,b); Paranjape & Sheth (2012b); Paranjape et al. (2013a); Lapi & Danese (2014); Castorina et al. (2016); Nikakhtar & Baghram (2017)). In this context, the cosmological model dependence enters, in part, through the statistics of the initial density fluctuation field and its evolution.

One of the most widely used frameworks among such approaches is the excursion set theory (EST) (Bond et al. (1991); Lacey & Cole (1993)). The basic implementation of the EST involves calculating the smoothed linear density contrast, δ(R)\delta(R), at different smoothing scales RR, around a given point in Lagrangian space. Since there is a monotonic relationship between RR and the variance of the smoothed density contrast, S≡⟨δ(R)2⟩S \equiv\langle\delta(R)^2\rangle, it is common to map δ(R)\delta(R) to δ(S)\delta(S) for calculations. In this manner, as one recursively reduces the smoothing scale RR (or equivalently increases SS), the corresponding smoothed density contrast traces out a “trajectory” in the S−δ(S)S-\delta(S) plane. This process can then be repeated for numerous other randomly chosen points in the Lagrangian space, creating an ensemble of trajectories which can then be analyzed as a stochastic process.

In the EST formalism, the first up-crossing of a positive threshold value (the collapse barrier) by a trajectory at a given scale SS, corresponding to a smoothing scale R(S)R(S), is associated with the collapse of the matter enclosed within the corresponding Lagrangian volume into a virialized halo in Eulerian space. The mass of the Eulerian halo can then be related to the mass of the Lagrangian “protohalo” region, M(R)M(R), which is determined by the smoothing scale R(S)R(S) and the filter function used to smooth the density contrast. The simplest collapse barrier is a constant barrier, B=δcB = \delta_c, motivated by the spherical collapse model (Press & Schechter (1974); Bond et al. (1991)). A more physically motivated barrier is the Sheth-Mo-Tormen (SMT) (Sheth et al. (2001)) moving barrier, which depends on SS and is motivated by an ellipsoidal collapse model that accounts for tidal effects. There are also other types of barriers, such as stochastic barriers that partially incorporate the complex nature of halo formation by introducing variability/scatter in the collapse threshold (Maggiore & Riotto (2010b)).

The first crossing distribution (FCD), f(S)f(S), is therefore a key function in the EST analyses as f(S)dSf(S)dS determines the fraction of trajectories that up-cross a given barrier for the first time within the interval (S,S+dS)(S,S+dS). Within the EST framework, this fraction can be interpreted as the fraction of matter that has collapsed into halos of the corresponding mass interval, obtained by mapping SS to the halo mass scale MM. It is clear that halo abundance quantities, such as the halo mass function which describes the comoving number density of halos per unit mass, are proportional to f(S)f(S) and can be calculated from that (e.g. Bond et al. (1991)).

Another key quantity in cosmological studies is halo bias, which characterizes the clustering of halos relative to the underlying matter density field (e.g. Kaiser (1984); Cole & Kaiser (1989); Mo & White (1996); Sheth & Tormen (1999); Tinker et al. (2010); Corasaniti & Achitouv (2011a); Paranjape & Sheth (2012b); Lazeyras et al. (2016a,b); Desjacques et al. (2018); Castro et al. (2024)). In general, halo bias also exhibits scale dependence and non-locality, particularly beyond the linear regime (e.g. Catelan et al. (1998); Smith et al. (2007); McDonald & Roy (2009); Matsubara (2011); Baldauf et al. (2012); Chan et al. (2012); Paranjape & Sheth (2012a); Musso et al. (2012); Paranjape et al. (2013a); Sheth et al. (2013); Paranjape et al. (2013b); Castorina & Sheth (2013); Modi et al. (2017); Lazeyras & Schmidt (2018); Desjacques et al. (2018)).

One of the widely used approaches to predicting halo bias parameters on large scales is the peak-background split (PBS) formalism. The underlying idea is that long-wavelength density perturbations modulate the abundance of halos by effectively modifying the background density in which small-scale density fluctuations evolve and eventually collapse to form halos ([32] (Kaiser 1984; ); Bardeen et al. 1986; ); Cole & Kaiser 1989; ); Catelan et al. 1998; ); Sheth & Tormen 1999)). In particular, when combined with the excursion set theory, the PBS formalism provides a convenient framework for calculating Lagrangian halo bias parameters at different orders from the FCD. It is worth noting that the PBS formalism yields accurate predictions when implemented exactly (Desjacques et al. (2018)), as demonstrated by Lazeyras et al. (2016b).

The closed-form expression of the first crossing distribution is not known for arbitrary barrier and filter functions. Exceptions include the case of Markovian trajectories (e.g. Bond et al. (1991); Sheth (1998)), i.e. trajectories with uncorrelated steps (walks), and completely correlated trajectories (Paranjape et al. (2012)). Assuming Gaussian linear density fluctuations, Markovian trajectories can be generated by applying a “sharp-kk” smoothing filter, which corresponds to a Top-Hat filter in Fourier space. However, in more realistic scenarios, where filters like real-space Top-Hat, Gaussian and other ones are used, the trajectories become non-Markovian (e.g. Peacock & Heavens (1990); Bond et al. (1991)). The problem becomes even more complicated when a generic barrier is considered in conjunction with such trajectories to make the analyses consistent with more realistic descriptions of the halo formation process.

Given the complexity of non-Markovian stochastic processes with generic barriers, various methods and frameworks have been proposed for approximating the first crossing distribution (e.g. Peacock & Heavens (1990); Maggiore & Riotto (2010a,b); Ma et al. (2011); Corasaniti & Achitouv (2011b,a); Paranjape et al. (2012); Musso & Sheth (2012); Farahi & Benson (2013); Musso & Sheth (2014a); Nikakhtar et al. (2018); Verza et al. (2024)). In this work, we adopt the Hertz and Stratonovich approximations introduced by Nikakhtar et al. (2018) to develop a computationally efficient framework for estimating FCDs for generic barriers, which are subsequently used to calculate halo bias parameters within the PBS formalism. We validate the framework by comparing its predictions with those obtained from a numerical evaluation of FCDs, based on large ensembles of excursion set trajectories generated using the Cholesky method (Nikakhtar et al. (2018)).

This paper is organized as follows: § 2 reviews the Hertz and Stratonovich approximations and uses them to develop an efficient framework for calculating FCDs for generic barriers and smoothing filters. § 3 then discusses how halo bias parameters can be efficiently calculated within the PBS formalism. We subsequently use the framework to compute halo multiplicity functions and biases based on the Hertz and Stratonovich approximations, considering different smoothing filters and barrier. We then compare these predictions with their numerical counterparts. Finally, we conclude this work in § 4.

First crossing distribution in Hertz and Stratonovich approximations

The FCD, f(S)f(S), can be obtained by calculating the fraction of trajectories that up-cross a barrier for the first time within the interval (S,S+dS)(S, S+\mathrm{d}S). However, a useful first step towards computing the FCD would be to first calculate the fraction of trajectories that up-cross the barrier at that interval, regardless of whether they have had previous up-crossings. For this purpose, following the arguments in (Musso & Sheth (2012); Baghram et al. (2019)), for a moving barrier B(S)B(S), one can compute the probability that a trajectory, δ(S)\delta(S), satisfies the following conditions: δ(S)≤B(S)\delta(S) \le B(S) and δ(S+dS)>B(S+dS)\delta(S+\mathrm{d}S) > B(S+\mathrm{d}S). The second condition can be expressed as: δ(S)>B(S)+(B′(S)−η(S))dS\delta(S) > B(S) + (B'(S) - \eta(S))\mathrm{d}S, where B′≡dB/dSB' \equiv\mathrm{d}B/\mathrm{d}S and η≡dδ/dS\eta\equiv\mathrm{d}\delta/\mathrm{d}S. Assuming that trajectories start from δ0=0\delta_0 = 0 and η0=0\eta_0 = 0 at S0=0S_0 = 0, the probability that a trajectory up-crosses the barrier within the interval (S,S+dS)(S, S+\mathrm{d}S) can then be written as:

∫B(S)+(B′(S)−ηS)dSB(S)dδS P(δS,ηS,S~)≃P(BS,ηS,S)(ηS−BS′)dS,(1)\int_{B(S)+(B'(S)-\eta_S)\mathrm{d}S}^{B(S)} \mathrm{d}\delta_S\,\mathcal{P}(\delta_S,\eta_S,\widetilde{S}) \simeq\mathcal{P}(B_S,\eta_S,S)(\eta_S-B'_S)\mathrm{d}S, \tag*{(1)}

where XS≡X(S)X_S \equiv X(S) with X=(B,B′,δ,η)X = (B, B', \delta, \eta). Assuming that linear matter density fluctuations are Gaussian, δS\delta_S and ηS\eta_S obey Gaussian statistics with ⟨δS⟩=⟨ηS⟩=0\langle\delta_S\rangle= \langle\eta_S\rangle= 0 and P\mathcal{P} is a joint Gaussian distribution. Now, one can calculate the “rate” of up-crossing the barrier at SS, n1(BS,S)n_1(B_S,S), which is the probability that a trajectory up-crosses the barrier B(S)B(S) per unit SS, by dividing the result in (1) by dS\mathrm{d}S and integrating over all slopes that are larger than the slope of the barrier, i.e. ηS>BS′\eta_S > B'_S. In other words:

n1(BS,S)=∫BS′∞dηS (ηS−BS′)P(BS,ηS,S).(2)n_1(B_S,S) = \int_{B'_S}^{\infty} \mathrm{d}\eta_S\,(\eta_S-B'_S)\mathcal{P}(B_S,\eta_S,S). \tag*{(2)}

The next step in calculating the FCD is to subtract from n1n_1 all contributions corresponding to trajectories that have up-crossed the barrier before SS (see e.g. Verzichethaguina et al. (2006); Musso & Sheth (2014b); Nikakhtar et al. (2018); Baghram et al. (2019)). For this purpose, it is useful to define rate functions for multiple up-crossings, np(BSp,Sp;…;BS1,S1)n_p(B_{S_p},S_p;\ldots;B_{S_1},S_1), which correspond to the joint probability that a trajectory up-crosses the barrier at (S1,S1+dS1),…,(Sp,Sp+dSp)(S_1,S_1+\mathrm{d}S_1),\ldots,(S_p,S_p+\mathrm{d}S_p). These rate functions can be expressed by generalizing Eq. (2):

np(BSp,Sp;…;BS1,S1)=∫BS1′∞dηS1 (ηS1−BS1′)⋯∫BSp′∞dηSp (ηSp−BSp′)×P(BSp,ηSp,Sp;…;BS1,ηS1,S1),(3)\begin{aligned} n_p(B_{S_p},S_p;\ldots;B_{S_1},S_1) &= \int_{B'_{S_1}}^\infty\mathrm{d}\eta_{S_1}\,(\eta_{S_1}-B'_{S_1})\cdots\int_{B'_{S_p}}^\infty\mathrm{d}\eta_{S_p}\,(\eta_{S_p}-B'_{S_p})\\ &\quad\times\mathcal{P}(B_{S_p},\eta_{S_p},S_p;\ldots;B_{S_1},\eta_{S_1},S_1), \tag*{(3)} \end{aligned}

where P\mathcal{P} is a multivariate Gaussian distribution.

To calculate the rate functions, one needs to evaluate auto/cross correlations of δ\delta and η\eta fields. For the δ\delta correlations we have:

⟨δiδj⟩=⟨δSiδSj⟩=⟨δ(Ri(Si))δ(Rj(Sj))⟩=∫0∞dk k2Pm(k)2π2W(kRi)W(kRj),(4)\langle\delta_i\delta_j\rangle= \langle\delta_{S_i}\delta_{S_j}\rangle= \langle\delta(R_i(S_i))\delta(R_j(S_j))\rangle= \int_0^\infty\mathrm{d}k\,\frac{k^2P_m(k)}{2\pi^2}W(kR_i)W(kR_j), \tag*{(4)}

where Pm(k)P_{\mathrm{m}}(k) is the linear matter power spectrum, RR denotes the smoothing scale and WW is the Fourier transform of the smoothing filter function. The variance SiS_i is obtained as the special case of this correlation function in the limit i=ji=j. From this relation, one can obtain a monotonic mapping between SS and RR, and vice versa, since Si=⟨δiδi⟩=⟨δ(Ri)⟩2S_i=\langle\delta_i\delta_i\rangle=\langle\delta(R_i)\rangle^2. The correlations involving the slope variable, η\eta, can also be evaluated using the following relations:

⟨ηiδj⟩≡⟨ηSiδSj⟩=dRidSi⟨dδ(Ri(Si))dRiδ(Rj(Sj))⟩(5)\langle\eta_i\delta_j\rangle\equiv\langle\eta_{S_i}\delta_{S_j}\rangle=\frac{\mathrm{d}R_i}{\mathrm{d}S_i}\left\langle\frac{\mathrm{d}\delta(R_i(S_i))}{\mathrm{d}R_i}\delta(R_j(S_j))\right\rangle \tag*{(5)}
⟨ηiηj⟩≡⟨ηSiηSj⟩=dRidSidRjdSj⟨dδ(Ri(Si))dRidδ(Rj(Sj))dRj⟩(6)\langle\eta_i\eta_j\rangle\equiv\langle\eta_{S_i}\eta_{S_j}\rangle=\frac{\mathrm{d}R_i}{\mathrm{d}S_i}\frac{\mathrm{d}R_j}{\mathrm{d}S_j}\left\langle\frac{\mathrm{d}\delta(R_i(S_i))}{\mathrm{d}R_i}\frac{\mathrm{d}\delta(R_j(S_j))}{\mathrm{d}R_j}\right\rangle \tag*{(6)}

which lead to

⟨ηiδj⟩=dRidSi[∫0∞dk k2Pm(k)2π2dW(kRi)dRiW(kRj)](7)\langle\eta_i\delta_j\rangle=\frac{\mathrm{d}R_i}{\mathrm{d}S_i}\left[\int_0^\infty\mathrm{d}k\,\frac{k^2P_{\mathrm{m}}(k)}{2\pi^2}\frac{\mathrm{d}W(kR_i)}{\mathrm{d}R_i}W(kR_j)\right] \tag*{(7)}
⟨ηiηj⟩=dRidSidRjdSj[∫0∞dk k2Pm(k)2π2dW(kRi)dRidW(kRj)dRj].(8)\langle\eta_i\eta_j\rangle=\frac{\mathrm{d}R_i}{\mathrm{d}S_i}\frac{\mathrm{d}R_j}{\mathrm{d}S_j}\left[\int_0^\infty\mathrm{d}k\,\frac{k^2P_{\mathrm{m}}(k)}{2\pi^2}\frac{\mathrm{d}W(kR_i)}{\mathrm{d}R_i}\frac{\mathrm{d}W(kR_j)}{\mathrm{d}R_j}\right]. \tag*{(8)}

Having obtained the rate functions at different orders, corresponding to different numbers of up-crossings, one can obtain the FCD from n1n_1 using the inclusion-exclusion principle (Baghram et al. (2019)). The final result is an infinite sum whose first term is n1n_1 and the subsequent terms involve integrals of rate functions at different orders. This sum can be expressed as:

f(S)=∑p=0∞(−1)pp!∫0SdS1⋯∫0SdSp np+1(BS,S;BSp,Sp;…;BS1,S1).(9)f(S)=\sum_{p=0}^{\infty}\frac{(-1)^p}{p!}\int_0^S\mathrm{d}S_1\cdots\int_0^S\mathrm{d}S_p\,n_{p+1}(B_S,S;B_{S_p},S_p;\ldots;B_{S_1},S_1). \tag*{(9)}

It is generally difficult to evaluate this expression exactly. However, the sum can be approximated by treating the contribution of the first few terms exactly and approximating the contributions of the higher order terms. In this work, we use two approximations of this kind, referred to as the Hertz and Stratonovich approximations, which were presented in Nikakhtar et al. (2018). To discuss these further, it is useful to express the FCD in the following way (Verechtchaguina et al. (2006)):

f(S)=ψ′(S)e−ψ(S).(10)f(S)=\psi'(S)e^{-\psi(S)}. \tag*{(10)}

with ψ′(S)\psi'(S) denoting a derivative with respect to SS and

ψ(S)≡∑p=1∞(−1)p+1p!∫0SdS1⋯∫0SdSp gp(Sp,…,S1),(11)\psi(S)\equiv\sum_{p=1}^{\infty}\frac{(-1)^{p+1}}{p!}\int_0^S\mathrm{d}S_1\cdots\int_0^S\mathrm{d}S_p\,g_p(S_p,\ldots,S_1), \tag*{(11)}

where gpg_p’s are cumulant functions. For instance, in our case with generic barriers, we have g1(S1)=n1(BS1,S1)g_1(S_1)=n_1(B_{S_1},S_1), g2(S2,S1)=n2(BS2,S2;BS1,S1)−n1(BS2,S2)n1(BS1,S1)g_2(S_2,S_1)=n_2(B_{S_2},S_2;B_{S_1},S_1)-n_1(B_{S_2},S_2)n_1(B_{S_1},S_1), etc.

The Hertz approximation

The idea behind the Hertz approximation (Hertz (1909); Nikakhtar et al. (2018)) (see also Musso & Sheth (2014b)) is that the up-crossings at different points are independent of each other. Under this assumption, the multivariate probability distribution in Eq. (3) can be factorized as P(BSp,ηSp,Sp;…;BS1,ηS1,S1)≈P(BSp,ηSp,Sp)…P(BS1,ηS1,S1)P(B_{S_p},\eta_{S_p},S_p;\ldots;B_{S_1},\eta_{S_1},S_1)\approx P(B_{S_p},\eta_{S_p},S_p)\ldots P(B_{S_1},\eta_{S_1},S_1). Consequently, this results in the factorization of the rate functions. Therefore, from Eq. (3), the rate functions in this approximation can be written as:

np(BSp,Sp;…;BS1,S1)≈∫BS1′∞dηS1(ηS1−BS1′)P(BS1,ηS1,S1)⋯∫BSp′∞dηSp(ηSp−BSp′)P(BSp,ηSp,Sp)=n1(BS1,S1)…n1(BSp,Sp).(12)\begin{aligned} n_p(B_{S_p},S_p;\ldots;B_{S_1},S_1)&\approx\int_{B'_{S_1}}^\infty\mathrm{d}\eta_{S_1}(\eta_{S_1}-B'_{S_1})P(B_{S_1},\eta_{S_1},S_1)\\ &\quad\cdots\int_{B'_{S_p}}^\infty\mathrm{d}\eta_{S_p}(\eta_{S_p}-B'_{S_p})P(B_{S_p},\eta_{S_p},S_p)\\ &=n_1(B_{S_1},S_1)\ldots n_1(B_{S_p},S_p). \tag*{(12)} \end{aligned}

Now, this result can be inserted into Eq. (9) to obtain the FCD in the Hertz approximation:

fHertz(S)=∑p=0∞(−1)pp!n1(BS,S)∫0SdS1n1(BS1,S1)⋯∫0SdSpn1(BSp,Sp).(13)f_{\mathrm{Hertz}}(S)=\sum_{p=0}^{\infty}\frac{(-1)^p}{p!}n_1(B_S,S)\int_0^S\mathrm{d}S_1n_1(B_{S_1},S_1)\cdots\int_0^S\mathrm{d}S_pn_1(B_{S_p},S_p). \tag*{(13)}

Since the integration limits of each integral in this expression are 00 and SS, the integrals are identical and depend only on SS. In other words:

∫0SdSin1(BSi,Si)=∫0SdSjn1(BSj,Sj)=n1(<S).(14)\int_0^S\mathrm{d}S_i n_1(B_{S_i},S_i)=\int_0^S\mathrm{d}S_j n_1(B_{S_j},S_j)=n_1(<S). \tag*{(14)}

Therefore, Eq. (13) can be re-expressed as:

fHertz(S)=n1(BS,S)e−n1(<S),(15)f_{\mathrm{Hertz}}(S)=n_1(B_S,S)e^{-n_1(<S)}, \tag*{(15)}

which, in comparison with Eq. (10), implies that:

ψHertz=n1(<S)=∫0SdS1n1(BS1,S1).(16)\psi_{\mathrm{Hertz}}=n_1(<S)=\int_0^S\mathrm{d}S_1n_1(B_{S_1},S_1). \tag*{(16)}

Considering Eq. (11), this approximation implies that all cumulant functions vanish except g1g_1.

Stratonovich approximation

The Hertz approximation was the first step toward estimating the sum in Eq. (9). In terms of the cumulant functions, it includes only the contribution of the first cumulant, while approximating higher order terms as zero. The Stratonovich approximation (Stratonovich (1967); Nikakhtar et al. (2018)) improves upon this estimate by incorporating the contributions of the first two terms exactly and approximating the higher order terms based on these two terms. This approximation can be applied to the cumulant functions as:

gp(Sp,…,S1)≈(−1)p−1(p−1)!n1(BSp,Sp)…n1(BS1,S1){R(S1,S2)R(S1,S3)…R(S1,Sp)}sym,(17)g_p(S_p,\ldots,S_1)\approx(-1)^{p-1}(p-1)!n_1(B_{S_p},S_p)\ldots n_1(B_{S_1},S_1)\left\{R(S_1,S_2)R(S_1,S_3)\ldots R(S_1,S_p)\right\}_{\mathrm{sym}}, \tag*{(17)}

where

R(Si,Sj)=1−n2(BSi,Si;BSj,Sj)n1(BSi,Si)n1(BSj,Sj).(18)R(S_i,S_j)=1-\frac{n_2(B_{S_i},S_i;B_{S_j},S_j)}{n_1(B_{S_i},S_i)n_1(B_{S_j},S_j)}. \tag*{(18)}

This expression can then be inserted into Eq. (11) and the sum can be evaluated, leading to the Stratonovich approximation for the ψ\psi function (Nikakhtar et al. (2018)):

ψStr(S)=−∫0SdS1n1(BS1,S1)ln⁡[1−n1(<S,S1)]n1(<S,S1),(19)\psi_{\mathrm{Str}}(S)=-\int_0^S\mathrm{d}S_1n_1(B_{S_1},S_1)\frac{\ln[1-n_1(<S,S_1)]}{n_1(<S,S_1)}, \tag*{(19)}

where

n1(<S,S1)=∫0SdS2R(S1,S2)n1(BS2,S2).(20)n_1(<S,S_1)=\int_0^S\mathrm{d}S_2R(S_1,S_2)n_1(B_{S_2},S_2). \tag*{(20)}

Finally, to find the FCD in Stratonovich approximation, fStr(S)f_{\mathrm{Str}}(S), one can use Eq. (10) with ψStr(S)\psi_{\mathrm{Str}}(S).

Calculation of the first crossing distribution

Following the previous parts, the two key functions for calculating the FCD in Hertz and Stratonovich approximations are n1n_1 and n2n_2 rate functions which can be obtained from Eq. (3). To find closed-form expressions for them, we make use of the fact that PP in Eq. (3) is a multivariate Gaussian distribution. Hereafter, for simplicity in writing, we drop the explicit dependence of Gaussian distributions on SS parameters: P(BSp,ηSp,Sp;…;BS1,ηS1,S1)=P(BSp,ηSp,…,BS1,ηS1)P(B_{S_p},\eta_{S_p},S_p;\ldots;B_{S_1},\eta_{S_1},S_1)=P(B_{S_p},\eta_{S_p},\ldots,B_{S_1},\eta_{S_1}).

For n1n_1, one can first write P(BS,ηS)=P(ηS∣BS)P(BS)P(B_S,\eta_S)=P(\eta_S|B_S)P(B_S). Considering the properties of conditional Gaussian distributions and also using S=⟨δS2⟩S=\langle\delta_S^2\rangle and ⟨δS⟩=⟨ηS⟩=0\langle\delta_S\rangle=\langle\eta_S\rangle=0, one obtains

n1(BS,S)=∫BS′∞dηS(ηS−BS′)P(ηS∣BS)P(BS),(21)n_1(B_S,S)=\int_{B'_S}^{\infty}d\eta_S(\eta_S-B'_S)P(\eta_S|B_S)P(B_S), \tag*{(21)}

where

P(BS)=12πSe−BS22S,P(ηS∣BS)=12πσc2e−(ηS−μc)22σc2.(22)P(B_S)=\frac{1}{\sqrt{2\pi S}}e^{-\frac{B_S^2}{2S}},\qquad P(\eta_S|B_S)=\frac{1}{\sqrt{2\pi\sigma_c^2}}e^{-\frac{(\eta_S-\mu_c)^2}{2\sigma_c^2}}. \tag*{(22)}

with:

μc≡⟨δSηS⟩SBS,σc2≡⟨ηS2⟩−⟨δSηS⟩2S.(23)\mu_c\equiv\frac{\langle\delta_S\eta_S\rangle}{S}B_S,\qquad\sigma_c^2\equiv\langle\eta_S^2\rangle-\frac{\langle\delta_S\eta_S\rangle^2}{S}. \tag*{(23)}

The integral in Eq. (21) can be done analytically. Starting from Eqs. (21) and (22), since BS′B'_S is usually a well-behaved function, especially in ranges that are relevant for cosmological analyses, we can apply the change of variable ηS−BS′=ξS\eta_S-B'_S=\xi_S at each SS. Therefore, we need to integrate:

n1(BS,S)=P(BS)∫0∞dξSξSe−(ξS+BS′−μc)22σc22πσc2≡P(BS)×IS.(24)n_1(B_S,S)=P(B_S)\int_0^\infty d\xi_S\xi_S\frac{e^{-\frac{(\xi_S+B'_S-\mu_c)^2}{2\sigma_c^2}}}{\sqrt{2\pi\sigma_c^2}}\equiv P(B_S)\times I_S. \tag*{(24)}

Now, defining μ~c≡μc−BS′\tilde{\mu}_c\equiv\mu_c-B'_S, ISI_S can be written as:

IS=σc22πσc2∂∂μ~c∫0∞dξSe−(ξS−μ~c)22σc2+μ~c2πσc2∫0∞dξSe−(ξS−μ~c)22σc2.(25)I_S=\frac{\sigma_c^2}{\sqrt{2\pi\sigma_c^2}}\frac{\partial}{\partial\tilde{\mu}_c}\int_0^\infty d\xi_S e^{-\frac{(\xi_S-\tilde{\mu}_c)^2}{2\sigma_c^2}}+\frac{\tilde{\mu}_c}{\sqrt{2\pi\sigma_c^2}}\int_0^\infty d\xi_S e^{-\frac{(\xi_S-\tilde{\mu}_c)^2}{2\sigma_c^2}}. \tag*{(25)}

One can now apply another change of variables, t=(ξS−μ~c)/(2σc)t=(\xi_S-\tilde{\mu}_c)/(\sqrt{2}\sigma_c) and re-write the integrals as:

IS=(2σc)[σc22πσc2∂∂μ~c∫−μ~c2σc∞dte−t2+μ~c2πσc2∫−μ~c2σc∞dte−t2].(26)I_S=(\sqrt{2}\sigma_c)\left[\frac{\sigma_c^2}{\sqrt{2\pi\sigma_c^2}}\frac{\partial}{\partial\tilde{\mu}_c}\int_{-\frac{\tilde{\mu}_c}{\sqrt{2}\sigma_c}}^\infty dt e^{-t^2}+\frac{\tilde{\mu}_c}{\sqrt{2\pi\sigma_c^2}}\int_{-\frac{\tilde{\mu}_c}{\sqrt{2}\sigma_c}}^\infty dt e^{-t^2}\right]. \tag*{(26)}

The first integral can then be written as:

σc22πσc2∂∂β∫−β∞dte−t2=σc2πe−β2,(27)\frac{\sigma_c^2}{\sqrt{2\pi\sigma_c^2}}\frac{\partial}{\partial\beta}\int_{-\beta}^\infty dt e^{-t^2}=\frac{\sigma_c}{\sqrt{2\pi}}e^{-\beta^2}, \tag*{(27)}

where β≡μ~c2σc\beta\equiv\frac{\tilde{\mu}_c}{\sqrt{2}\sigma_c}. The second integral can also be written as:

μ~c(2σc)2πσc2(∫−μ~c2σc0dte−t2+∫0∞dte−t2)=μ~c2(1−erf⁡(−μ~c2σc)),(28)\frac{\tilde{\mu}_c(\sqrt{2}\sigma_c)}{\sqrt{2\pi\sigma_c^2}}\left(\int_{-\frac{\tilde{\mu}_c}{\sqrt{2}\sigma_c}}^0dt e^{-t^2}+\int_0^\infty dt e^{-t^2}\right)=\frac{\tilde{\mu}_c}{2}\left(1-\operatorname{erf}\left(-\frac{\tilde{\mu}_c}{\sqrt{2}\sigma_c}\right)\right), \tag*{(28)}

where we have used the definition of the error function. Now, using erf⁡(−x)=−erf⁡(x)\operatorname{erf}(-x)=-\operatorname{erf}(x) and combining this result with (27) we find the following relation for ISI_S:

IS=σc2πe−μ~c22σc2+μ~c2(1+erf⁡(μ~c2σc)),(29)I_S=\frac{\sigma_c}{\sqrt{2\pi}}e^{-\frac{\tilde{\mu}_c^2}{2\sigma_c^2}}+\frac{\tilde{\mu}_c}{2}\left(1+\operatorname{erf}\left(\frac{\tilde{\mu}_c}{\sqrt{2}\sigma_c}\right)\right), \tag*{(29)}

which, according to Eq. (24) and the definition of μ~c\tilde{\mu}_c, leads to the following relation for the n1n_1 rate function:

n1(BS,S)=12πSe−BS22S{σc2πe−(μc−BS′)22σc2+(μc−BS′)2[1+erf⁡(μc−BS′2σc)]}.(30)n_1(B_S,S)=\frac{1}{\sqrt{2\pi S}}e^{-\frac{B_S^2}{2S}}\left\{\frac{\sigma_c}{\sqrt{2\pi}}e^{-\frac{(\mu_c-B'_S)^2}{2\sigma_c^2}}+\frac{(\mu_c-B'_S)}{2}\left[1+\operatorname{erf}\left(\frac{\mu_c-B'_S}{\sqrt{2}\sigma_c}\right)\right]\right\}. \tag*{(30)}

The next step is to calculate the n2n_2 rate function, which is more involved. However, as will be shown below, the final expressions are compact and straightforward to evaluate. From Eq. (3), we need to consider the multivariate Gaussian distribution function P(BS1,ηS1,BS2,ηS2)P(B_{S_1},\eta_{S_1},B_{S_2},\eta_{S_2}) which can be expressed as P(ηS1,ηS2∣BS1,BS2)P(BS1,BS2)P(\eta_{S_1},\eta_{S_2}|B_{S_1},B_{S_2})P(B_{S_1},B_{S_2}) with P(ηS1,ηS2∣BS1,BS2)P(\eta_{S_1},\eta_{S_2}|B_{S_1},B_{S_2}) being a conditional bivariate Gaussian distribution. This means that we can write the expression for n2n_2 as:

n2(BS1,S1;BS2,S2)=P(BS1,BS2)×I12(BS1,S1;BS2,S2),(31)n_2(B_{S_1},S_1;B_{S_2},S_2)=P(B_{S_1},B_{S_2})\times I_{12}(B_{S_1},S_1;B_{S_2},S_2), \tag*{(31)}

where

I12(BS1,S1;BS2,S2)=∫BS1′∞dηS1(ηS1−BS1′)∫BS2′∞dηS2(ηS2−BS2′)P(ηS1,ηS2∣BS1,BS2).(32)I_{12}(B_{S_1},S_1;B_{S_2},S_2)=\int_{B'_{S_1}}^\infty d\eta_{S_1}(\eta_{S_1}-B'_{S_1})\int_{B'_{S_2}}^\infty d\eta_{S_2}(\eta_{S_2}-B'_{S_2})P(\eta_{S_1},\eta_{S_2}|B_{S_1},B_{S_2}). \tag*{(32)}

The joint Gaussian distribution in Eq. (31) has a relatively simple expression. It can be obtained from the following relation:

P(BS1,BS2)=1(2π)2det⁡Cexp⁡[−12(BS1BS2)C−1(BS1BS2)](33)P(B_{S_1},B_{S_2})=\frac{1}{\sqrt{(2\pi)^2\det C}}\exp\left[-\frac{1}{2}\begin{pmatrix}B_{S_1}&B_{S_2}\end{pmatrix}C^{-1}\begin{pmatrix}B_{S_1}\\B_{S_2}\end{pmatrix}\right] \tag*{(33)}

with

C=(S1⟨δS1δS2⟩⟨δS2δS1⟩S2).(34)C=\begin{pmatrix}S_1&\langle\delta_{S_1}\delta_{S_2}\rangle\\\langle\delta_{S_2}\delta_{S_1}\rangle&S_2\end{pmatrix}. \tag*{(34)}

The exponent of the bivariate Gaussian can then be expanded to yield:

P(BS1,BS2)=12πS1S2−S122exp⁡[−12BS22S1+BS12S2−2BS1BS2S12S1S2−S122],(35)P(B_{S_1},B_{S_2})=\frac{1}{2\pi\sqrt{S_1S_2-S_{12}^2}}\exp\left[-\frac{1}{2}\frac{B_{S_2}^2S_1+B_{S_1}^2S_2-2B_{S_1}B_{S_2}S_{12}}{S_1S_2-S_{12}^2}\right], \tag*{(35)}

where S12≡⟨δS1δS2⟩=⟨δS2δS1⟩S_{12}\equiv\langle\delta_{S_1}\delta_{S_2}\rangle=\langle\delta_{S_2}\delta_{S_1}\rangle.

Since P(ηS1,ηS2∣BS1,BS2)P(\eta_{S_1},\eta_{S_2}|B_{S_1},B_{S_2}) in Eq. (32) is a conditional bivariate Gaussian distribution, we first briefly review the properties of conditional Gaussian distributions. Consider a vector of Gaussian random variables, x=(xA,xB)T\mathbf{x}=(\mathbf{x}_A,\mathbf{x}_B)^{\mathsf{T}}, where xA∼N(μA,ΣA)\mathbf{x}_A\sim\mathcal{N}(\boldsymbol{\mu}_A,\Sigma_A) and xB∼N(μB,ΣB)\mathbf{x}_B\sim\mathcal{N}(\boldsymbol{\mu}_B,\Sigma_B) are themselves vectors of Gaussian random variables. Then, one can write x∼N(μ,Σ)\mathbf{x}\sim\mathcal{N}(\boldsymbol{\mu},\Sigma) with:

μ=(μAμB),Σ=(ΣAAΣABΣBAΣBB).(36)\boldsymbol{\mu}=\begin{pmatrix}\boldsymbol{\mu}_A\\\boldsymbol{\mu}_B\end{pmatrix},\qquad\Sigma=\begin{pmatrix}\Sigma_{AA}&\Sigma_{AB}\\\Sigma_{BA}&\Sigma_{BB}\end{pmatrix}. \tag*{(36)}

For such vectors, the conditional distribution P(xA∣xB)P(\mathbf{x}_A|\mathbf{x}_B) is also a Gaussian, P(xA∣xB)=N(μ,Σ)P(\mathbf{x}_A|\mathbf{x}_B)=\mathcal{N}(\boldsymbol{\mu},\Sigma), with the following mean and variance:

μ=μA+ΣABΣBB−1(xB−μB),(37)\boldsymbol{\mu}=\boldsymbol{\mu}_A+\Sigma_{AB}\Sigma_{BB}^{-1}(\mathbf{x}_B-\boldsymbol{\mu}_B), \tag*{(37)}
Σ=ΣAA−ΣABΣBB−1ΣBA.(38)\Sigma=\Sigma_{AA}-\Sigma_{AB}\Sigma_{BB}^{-1}\Sigma_{BA}. \tag*{(38)}

For the conditional bivariate Gaussian distribution relevant to our calculations, P(ηS1,ηS2∣BS1,BS2)P(\eta_{S_1},\eta_{S_2}\mid B_{S_1},B_{S_2}), we have x=(ηS1,ηS2,δS1=BS1,δS2=BS2)T\mathbf{x}=(\eta_{S_1},\eta_{S_2},\delta_{S_1}=B_{S_1},\delta_{S_2}=B_{S_2})^{\mathsf{T}}, xA=(ηS1,ηS2)T\mathbf{x}_A=(\eta_{S_1},\eta_{S_2})^{\mathsf{T}} and xB=(δS1=BS1,δS2=BS2)T\mathbf{x}_B=(\delta_{S_1}=B_{S_1},\delta_{S_2}=B_{S_2})^{\mathsf{T}}. Therefore, using Eqs. (37) and (38) for P(xA∣xB)P(\mathbf{x}_A\mid\mathbf{x}_B), one can write:

P(ηS1,ηS2∣BS1,BS2)=12πdet⁡Σexp⁡[−12(η−μ)TΣ−1(η−μ)](39)P(\eta_{S_1},\eta_{S_2}\mid B_{S_1},B_{S_2})=\frac{1}{2\pi\sqrt{\det\boldsymbol{\Sigma}}}\exp\left[-\frac{1}{2}(\boldsymbol{\eta}-\boldsymbol{\mu})^{\mathsf{T}}\boldsymbol{\Sigma}^{-1}(\boldsymbol{\eta}-\boldsymbol{\mu})\right] \tag*{(39)}

where η=(ηS1,ηS2)T\boldsymbol{\eta}=(\eta_{S_1},\eta_{S_2})^{\mathsf{T}} and, using ⟨ηSj⟩=⟨δSj⟩=0\langle\eta_{S_j}\rangle=\langle\delta_{S_j}\rangle=0,

μ=DC−1(BS1BS2),Σ=E−DC−1DT(40)\boldsymbol{\mu}=DC^{-1}\begin{pmatrix}B_{S_1}\\B_{S_2}\end{pmatrix},\qquad\boldsymbol{\Sigma}=E-DC^{-1}D^{\mathsf{T}} \tag*{(40)}

with

C=(S1⟨δS1δS2⟩⟨δS2δS1⟩S2)=(S1S12S21S2),(41)C=\begin{pmatrix}S_1&\langle\delta_{S_1}\delta_{S_2}\rangle\\\langle\delta_{S_2}\delta_{S_1}\rangle&S_2\end{pmatrix}=\begin{pmatrix}S_1&S_{12}\\S_{21}&S_2\end{pmatrix}, \tag*{(41)}
D=(⟨ηS1δS1⟩⟨ηS1δS2⟩⟨ηS2δS1⟩⟨ηS2δS2⟩)=(ed11ed12ed21ed22),(42)D=\begin{pmatrix}\langle\eta_{S_1}\delta_{S_1}\rangle&\langle\eta_{S_1}\delta_{S_2}\rangle\\\langle\eta_{S_2}\delta_{S_1}\rangle&\langle\eta_{S_2}\delta_{S_2}\rangle\end{pmatrix}=\begin{pmatrix}\mathrm{ed}_{11}&\mathrm{ed}_{12}\\\mathrm{ed}_{21}&\mathrm{ed}_{22}\end{pmatrix}, \tag*{(42)}
E=(⟨ηS1ηS1⟩⟨ηS1ηS2⟩⟨ηS2ηS1⟩⟨ηS2ηS2⟩)=(e1e12e21e2).(43)E=\begin{pmatrix}\langle\eta_{S_1}\eta_{S_1}\rangle&\langle\eta_{S_1}\eta_{S_2}\rangle\\\langle\eta_{S_2}\eta_{S_1}\rangle&\langle\eta_{S_2}\eta_{S_2}\rangle\end{pmatrix}=\begin{pmatrix}e_1&e_{12}\\e_{21}&e_2\end{pmatrix}. \tag*{(43)}

In these expressions, we also have S12=S21S_{12}=S_{21} and e12=e21e_{12}=e_{21}. From these matrices, one can obtain the elements of μ\boldsymbol{\mu} and Σ\boldsymbol{\Sigma} introduced in (40). The final results are:

μ1=ed11(BS2S12−BS1S2)+ed12(BS1S12−BS2S1)S122−S1S2,(44)\mu_1=\frac{\mathrm{ed}_{11}(B_{S_2}S_{12}-B_{S_1}S_2)+\mathrm{ed}_{12}(B_{S_1}S_{12}-B_{S_2}S_1)}{S_{12}^2-S_1S_2}, \tag*{(44)}
μ2=ed21(BS2S12−BS1S2)+ed22(BS1S12−BS2S1)S122−S1S2,(45)\mu_2=\frac{\mathrm{ed}_{21}(B_{S_2}S_{12}-B_{S_1}S_2)+\mathrm{ed}_{22}(B_{S_1}S_{12}-B_{S_2}S_1)}{S_{12}^2-S_1S_2}, \tag*{(45)}
Σ11=ed122S1−2ed11ed12S12+e12S122+ed112S2−e12S1S2S122−S1S2,(46)\Sigma_{11}=\frac{\mathrm{ed}_{12}^{2}S_1-2\mathrm{ed}_{11}\mathrm{ed}_{12}S_{12}+e_1^{2}S_{12}^{2}+\mathrm{ed}_{11}^{2}S_2-e_1^{2}S_1S_2}{S_{12}^{2}-S_1S_2}, \tag*{(46)}
Σ22=ed222S1−2ed21ed22S12+e22S122+ed212S2−e22S1S2S122−S1S2,(47)\Sigma_{22}=\frac{\mathrm{ed}_{22}^{2}S_1-2\mathrm{ed}_{21}\mathrm{ed}_{22}S_{12}+e_2^{2}S_{12}^{2}+\mathrm{ed}_{21}^{2}S_2-e_2^{2}S_1S_2}{S_{12}^{2}-S_1S_2}, \tag*{(47)}
Σ12=Σ21=1S122−S1S2(ed12(ed22S1−ed21S12)+ed11(ed21S2−ed22S12)+e12(S122−S1S2)).(48)\Sigma_{12}=\Sigma_{21}=\frac{1}{S_{12}^{2}-S_1S_2}\left(\mathrm{ed}_{12}(\mathrm{ed}_{22}S_1-\mathrm{ed}_{21}S_{12})+\mathrm{ed}_{11}(\mathrm{ed}_{21}S_2-\mathrm{ed}_{22}S_{12})+e_{12}(S_{12}^{2}-S_1S_2)\right). \tag*{(48)}

These equations provide everything needed to fully determine the conditional bivariate Gaussian distribution in Eq. (39). The next step is obtaining a closed-form expression for the I12I_{12} function. From Eqs. (32) and (39), one can re-write the I12I_{12} integral as:

I12=∫BS1′∞dηS1(ηS1−BS1′)∫BS2′∞dηS2(ηS2−BS2′)(12πΣ11Σ22−Σ122exp⁡[−Σ11Δη22+Σ22Δη12−2Σ12Δη1Δη22(Σ122−Σ11Σ22)]),(49)I_{12}=\int_{B'_{S_1}}^{\infty}d\eta_{S_1}(\eta_{S_1}-B'_{S_1})\int_{B'_{S_2}}^{\infty}d\eta_{S_2}(\eta_{S_2}-B'_{S_2})\left(\frac{1}{2\pi\sqrt{\Sigma_{11}\Sigma_{22}-\Sigma_{12}^{2}}}\exp\left[-\frac{\Sigma_{11}\Delta\eta_2^2+\Sigma_{22}\Delta\eta_1^2-2\Sigma_{12}\Delta\eta_1\Delta\eta_2}{2(\Sigma_{12}^{2}-\Sigma_{11}\Sigma_{22})}\right]\right), \tag*{(49)}

where Δηj≡ηSj−μj\Delta\eta_j\equiv\eta_{S_j}-\mu_j. By applying the following change of variables:

u=Δη1Σ11v=Δη2Σ22,(50)u=\frac{\Delta\eta_1}{\sqrt{\Sigma_{11}}}\qquad v=\frac{\Delta\eta_2}{\sqrt{\Sigma_{22}}}, \tag*{(50)}

and defining:

σ1≡Σ11σ2≡Σ22ρ≡Σ12σ1σ2,(51)\sigma_1\equiv\sqrt{\Sigma_{11}}\qquad\sigma_2\equiv\sqrt{\Sigma_{22}}\qquad\rho\equiv\frac{\Sigma_{12}}{\sigma_1\sigma_2}, \tag*{(51)}

after some algebra, the integral takes the following form:

I12=∫BS1′−μ1σ1∞du(uσ1+μ1−BS1′)∫BS2′−μ2σ2∞dv(vσ2+μ2−BS2′)(12π1−ρ2exp⁡[−u2+v2−2uvρ2(1−ρ2)]).(52)I_{12}=\int_{\frac{B'_{S_1}-\mu_1}{\sigma_1}}^{\infty}du(u\sigma_1+\mu_1-B'_{S_1})\int_{\frac{B'_{S_2}-\mu_2}{\sigma_2}}^{\infty}dv(v\sigma_2+\mu_2-B'_{S_2})\left(\frac{1}{2\pi\sqrt{1-\rho^2}}\exp\left[-\frac{u^2+v^2-2uv\rho}{2(1-\rho^2)}\right]\right). \tag*{(52)}

Now, we define

a1≡BS1′−μ1σ1,a2≡BS2′−μ2σ2,(53)a_1\equiv\frac{B'_{S_1}-\mu_1}{\sigma_1},\qquad a_2\equiv\frac{B'_{S_2}-\mu_2}{\sigma_2}, \tag*{(53)}

and the integral is further simplified as

I12=σ1σ2∫a1∞du(u−a1)∫a2∞dv(v−a2)(12π1−ρ2exp⁡[−u2+v2−2uvρ2(1−ρ2)])(54)I_{12}=\sigma_1\sigma_2\int_{a_1}^{\infty}du(u-a_1)\int_{a_2}^{\infty}dv(v-a_2)\left(\frac{1}{2\pi\sqrt{1-\rho^2}}\exp\left[-\frac{u^2+v^2-2uv\rho}{2(1-\rho^2)}\right]\right) \tag*{(54)}

≡σ1σ2J12(a1,a2;ρ)\equiv\sigma_1\sigma_2J_{12}(a_1,a_2;\rho).

Note that the expression inside the parentheses is the probability density function (PDF) of a standard bivariate Gaussian distribution, which is a well-known function in statistics. Referring to this expression as ϕ2(u,v;ρ)\phi_2(u,v;\rho), the integrand of J12(a1,a2;ρ)J_{12}(a_1,a_2;\rho) can be expanded:

J12(a1,a2;ρ)=∫a1∞du∫a2∞dv(uv)ϕ2(u,v;ρ)−a2∫a1∞du∫a2∞dv uϕ2(u,v;ρ)−a1∫a1∞du∫a2∞dv vϕ2(u,v;ρ)+a1a2∫a1∞du∫a2∞dv ϕ2(u,v;ρ).(55)\begin{aligned} J_{12}(a_1,a_2;\rho)={}&\int_{a_1}^{\infty}du\int_{a_2}^{\infty}dv(uv)\phi_2(u,v;\rho)\\ &-a_2\int_{a_1}^{\infty}du\int_{a_2}^{\infty}dv\,u\phi_2(u,v;\rho)-a_1\int_{a_1}^{\infty}du\int_{a_2}^{\infty}dv\,v\phi_2(u,v;\rho)\\ &+a_1a_2\int_{a_1}^{\infty}du\int_{a_2}^{\infty}dv\,\phi_2(u,v;\rho). \tag*{(55)} \end{aligned}

Here, the last integral gives the probability P2(u>a1,v>a2)P_2(u>a_1,v>a_2). Similarly, the first three integrals correspond to different “moments” of the standard bivariate Gaussian PDF over the desired region. Such moments have already been calculated in the mathematical literature. Here we use the results of Rosenbaum (1961), which, after some algebra, lead to the following expression for the first three terms:

−a1ϕ(a2)Φ(ρa2−a11−ρ2)−a2ϕ(a1)Φ(ρa1−a21−ρ2)+1−ρ22πϕ(a12+a22−2a1a2ρ1−ρ2)+ρ∫a1∞du∫a2∞dvϕ2(u,v;ρ),(56)\begin{aligned} -a_1\phi(a_2)\Phi\left(\frac{\rho a_2-a_1}{\sqrt{1-\rho^2}}\right)-a_2\phi(a_1)\Phi\left(\frac{\rho a_1-a_2}{\sqrt{1-\rho^2}}\right) \\ &+\frac{\sqrt{1-\rho^2}}{\sqrt{2\pi}}\phi\left(\frac{\sqrt{a_1^2+a_2^2-2a_1a_2\rho}}{\sqrt{1-\rho^2}}\right)+\rho\int_{a_1}^{\infty}du\int_{a_2}^{\infty}dv\phi_2(u,v;\rho), \tag*{(56)} \end{aligned}

where ϕ\phi and Φ\Phi are, respectively, the PDF and cumulative distribution function (CDF) of a one-dimensional standard Gaussian distribution. The probability integral P2(u>a1,v>a2)P_2(u>a_1,v>a_2) appears here again in the last term. This can be expressed as:

P2(u>a1,v>a2)=1−P2(u≤a1)−P2(v≤a2)+P2(u≤a1,v≤a2).(57)P_2(u>a_1,v>a_2)=1-P_2(u\leq a_1)-P_2(v\leq a_2)+P_2(u\leq a_1,v\leq a_2). \tag*{(57)}

Therefore, for the integral we have:

∫a1∞du∫a2∞dvϕ2(u,v;ρ)=P2(u>a1,v>a2)=1−Φ(a1)−Φ(a2)+Φ2(a1,a2;ρ).(58)\begin{aligned} \int_{a_1}^{\infty}du\int_{a_2}^{\infty}dv\phi_2(u,v;\rho)=P_2(u>a_1,v>a_2) \\ &=1-\Phi(a_1)-\Phi(a_2)+\Phi_2(a_1,a_2;\rho). \tag*{(58)} \end{aligned}

where Φ2(a1,a2;ρ)\Phi_{2}(a_{1},a_{2};\rho) is the CDF of a standard bivariate Gaussian distribution with correlation coefficient ρ\rho. From this expression and (56) for the first three terms of J12J_{12}, one can obtain the final expression as:

J12(a1,a2;ρ)=−a1ϕ(a2)Φ(ρa2−a11−ρ2)−a2ϕ(a1)Φ(ρa1−a21−ρ2)+1−ρ22πϕ(a12+a22−2a1a2ρ1−ρ2)+(ρ+a1a2)(1−Φ(a1)−Φ(a2)+Φ2(a1,a2;ρ)).(59)J_{12}(a_{1},a_{2};\rho) = -a_{1}\phi(a_{2})\Phi\left(\frac{\rho a_{2}-a_{1}}{\sqrt{1-\rho^{2}}}\right)-a_{2}\phi(a_{1})\Phi\left(\frac{\rho a_{1}-a_{2}}{\sqrt{1-\rho^{2}}}\right) +\frac{\sqrt{1-\rho^{2}}}{\sqrt{2\pi}}\phi\left(\frac{\sqrt{a_{1}^{2}+a_{2}^{2}-2a_{1}a_{2}\rho}}{\sqrt{1-\rho^{2}}}\right) +(\rho+a_{1}a_{2})\left(1-\Phi(a_{1})-\Phi(a_{2})+\Phi_{2}(a_{1},a_{2};\rho)\right). \tag*{(59)}

Finally, since we had defined I12≡σ1σ2J12(a1,a2;ρ)I_{12}\equiv\sigma_{1}\sigma_{2}J_{12}(a_{1},a_{2};\rho), this results in:

I12(BS1,S1;BS2,S2)=σ1σ2{−a1ϕ(a2)Φ(ρa2−a11−ρ2)−a2ϕ(a1)Φ(ρa1−a21−ρ2)+1−ρ22πϕ(a12+a22−2a1a2ρ1−ρ2)+(ρ+a1a2)(1−Φ(a1)−Φ(a2)+Φ2(a1,a2;ρ))}.(60)I_{12}(B_{S_{1}},S_{1};B_{S_{2}},S_{2}) = \sigma_{1}\sigma_{2}\left\{-a_{1}\phi(a_{2})\Phi\left(\frac{\rho a_{2}-a_{1}}{\sqrt{1-\rho^{2}}}\right)-a_{2}\phi(a_{1})\Phi\left(\frac{\rho a_{1}-a_{2}}{\sqrt{1-\rho^{2}}}\right) +\frac{\sqrt{1-\rho^{2}}}{\sqrt{2\pi}}\phi\left(\frac{\sqrt{a_{1}^{2}+a_{2}^{2}-2a_{1}a_{2}\rho}}{\sqrt{1-\rho^{2}}}\right) +(\rho+a_{1}a_{2})\left(1-\Phi(a_{1})-\Phi(a_{2})+\Phi_{2}(a_{1},a_{2};\rho)\right)\right\}. \tag*{(60)}

based on which the rate function n2n_{2} was presented in Eq. (31). We recall that ϕ\phi and Φ\Phi denote, respectively, the PDF and CDF of a one-dimensional standard Gaussian distribution and Φ2(a1,a2;ρ)\Phi_{2}(a_{1},a_{2};\rho) denotes the CDF of a standard bivariate Gaussian distribution. All these functions are well-known and have been implemented in most standard computational packages. The five parameters in this expression are:

σ1≡Σ11,σ2≡Σ22,ρ≡Σ12σ1σ2,(61)\sigma_{1}\equiv\sqrt{\Sigma_{11}},\quad\sigma_{2}\equiv\sqrt{\Sigma_{22}},\quad\rho\equiv\frac{\Sigma_{12}}{\sigma_{1}\sigma_{2}}, \tag*{(61)}
a1≡BS1′−μ1σ1,a2≡BS2′−μ2σ2,(62)a_{1}\equiv\frac{B'_{S_{1}}-\mu_{1}}{\sigma_{1}},\quad a_{2}\equiv\frac{B'_{S_{2}}-\mu_{2}}{\sigma_{2}}, \tag*{(62)}

where Σ11\Sigma_{11}, Σ22\Sigma_{22}, Σ12\Sigma_{12}, μ1\mu_{1} and μ2\mu_{2} can be found in Eqs (44) - (48). These are functions that, in general, depend on ⟨δSiδSj⟩\langle\delta_{S_{i}}\delta_{S_{j}}\rangle, ⟨ηSiηSj⟩\langle\eta_{S_{i}}\eta_{S_{j}}\rangle, ⟨ηSiδSj⟩\langle\eta_{S_{i}}\delta_{S_{j}}\rangle and BSiB_{S_{i}} where each of ii and jj can be 1, 2.

Summary of the calculation steps

We are now able to calculate the FCD based on the Hertz and Stratonovich approximations for any given smoothing filter, barrier function, and set of cosmological parameters. The first step is to calculate the correlation functions from Eqs (4), (7) and (8) for a given filter function. The next step is to calculate the n1n_{1} rate function using Eq. (30). Next, the n2n_{2} rate function can be calculated using Eq. (31), where the expressions appearing in this equation are given by Eqs (35) and (60). Given the forms of these expressions, all steps are computationally straightforward and can be performed efficiently. Finally, after calculating the n1n_{1} and n2n_{2} rates, one can use Eqs (16) and (19) to obtain the ψ\psi functions in Hertz and Stratonovich approximations, respectively, and calculate the corresponding FCDs from Eq. (10). All these steps are also numerically straightforward and computationally efficient.

Halo multiplicity function and bias approximations

After calculating the FCD for a given barrier and filter function following the steps outlined in the previous subsection, one can readily determine the corresponding halo multiplicity function — a quantity widely used to calculate halo abundances. Halo multiplicity function is usually expressed as νf(ν)\nu f(\nu), where ν≡δc/S\nu\equiv\delta_{c}/\sqrt{S} denotes the peak height parameter and δc\delta_{c} is the critical overdensity for collapse (halo formation threshold). The function f(ν)f(\nu) can be calculated from the FCD using the relation f(S)dS=−f(ν)dνf(S)dS=-f(\nu)d\nu, which implies that f(ν)f(\nu) is a normalized distribution with an interpretation similar to that of f(S)f(S), but for ν\nu intervals. Considering the relation between ν\nu and SS, one obtains

f(ν)=2δc2ν3f(S(ν)),(63)f(\nu)=\frac{2\delta_{c}^{2}}{\nu^{3}}f(S(\nu)), \tag*{(63)}

where S(ν)=δc2/ν2S(\nu)=\delta_{c}^{2}/\nu^{2}.

One motivation for using f(ν)f(\nu) instead of f(S)f(S) is the fact that early models predicted the multiplicity function to be universal, i.e. absorbing the cosmology and redshift dependence into the peak height parameter ν\nu (e.g. Press & Schechter (1974); Sheth & Tormen (1999); Jenkins et al. (2001)). However, later studies based on higher-resolution N-body simulations demonstrated that universality is not exact (e.g. Warren et al. (2006); Reed et al. (2007); Tinker et al. (2008); Crocce et al. (2010)). Furthermore, EST-based predictions for f(ν)f(\nu) for arbitrary choices of smoothing filter and barrier functions are not guaranteed to be universal either. Nonetheless, using f(ν)f(\nu) is still useful as it is near-universal in many cases and, more generally, can provide an efficient framework for modeling departures from universality (e.g. Despali et al. (2016)). The relation between the halo multiplicity function and the halo mass function, the comoving number density of a halos per unit log MM interval, can be written as:

dndlog⁡M=ρˉMf(S)∣dSdlog⁡M∣=ρˉM(νf(ν))dln⁡νdlog⁡M,(64)\frac{dn}{d\log M}=\frac{\bar{\rho}}{M}f(S)\left|\frac{dS}{d\log M}\right|=\frac{\bar{\rho}}{M}\left(\nu f(\nu)\right)\frac{d\ln\nu}{d\log M}, \tag*{(64)}

where ρˉ\bar{\rho} denotes the average comoving matter density.

In addition to the multiplicity function and halo mass function, one can also estimate the halo bias from the FCD for a given collapse barrier, B(S,δc)B(S,\delta_{c})1, and a chosen smoothing filter. In general, the parameters of the barrier, including those determining its shape, can be calibrated against N-body simulations (e.g. Sheth et al. (2001); Verza et al. (2024)). Here, we focus on the calculation of halo bias using the PBS formalism (see the references in Section 1). Given that halo bias can be interpreted as the response of halo abundance to a long-wavelength background density fluctuation, and considering the abundance formula in Eq. (64), one can calculate the Lagrangian bias parameters within the EST framework by studying the conditional function f(S∣δe,Se)f(S|\delta_{e},S_{e}), which is the FCD for trajectories conditioned on crossing the point (Se,δe)(S_{e},\delta_{e}), where δe\delta_{e} is the long-wavelength background density fluctuation and SeS_{e} its variance. In fact, the Lagrangian halo density contrast can be written as

1+δh(L)=f(S∣Se,δe)f(S∣0,0),(65)1+\delta_{h}^{(\mathrm{L})}=\frac{f(S|S_{e},\delta_{e})}{f(S|0,0)}, \tag*{(65)}

where f(S∣0,0)=f(S)f(S|0,0)=f(S) is the unconditional FCD. At large scales where Se≪1S_{e}\ll1, one can obtain local Lagrangian bias parameters by expanding the conditional FCD in terms of δℓ\delta_\ell. In other words:

δh(L)=∑n=1∞bn(L)n!δℓn=1f(S)∑n=1∞1n!∂nf(S∣0,δℓ)∂δℓn∣δℓ=0δℓn(66)\delta_h^{(\mathrm{L})} = \sum_{n=1}^{\infty} \frac{b_n^{(\mathrm{L})}}{n!}\delta_\ell^n = \frac{1}{f(S)}\sum_{n=1}^{\infty}\frac{1}{n!}\frac{\partial^n f(S|0,\delta_\ell)}{\partial\delta_\ell^n}\bigg|_{\delta_\ell=0}\delta_\ell^n \tag*{(66)}

which leads to

bn(L)=1f(S)∂nf(S∣0,δℓ)∂δℓn∣δℓ=0.(67)b_n^{(\mathrm{L})}=\frac{1}{f(S)}\frac{\partial^n f(S|0,\delta_\ell)}{\partial\delta_\ell^n}\bigg|_{\delta_\ell=0}. \tag*{(67)}

These Lagrangian bias parameters can then be transformed into Eulerian bias parameters at different orders using appropriate transformation relations (e.g. Catelan et al. (1998); Matsubara (2011); Abidi & Baldauf (2018)).

For uncorrelated (Markovian) steps, the conditional FCD in Eq. (65) can be replaced with an unconditional FCD with appropriate shifts in SS and the barrier (e.g. Paranjape & Sheth (2012a)). In fact, for Markovian excursion set trajectories that first up-cross a barrier of the general form, B(S)=Bc+Bs(S)B(S)=B_{\mathrm{c}}+B_{\mathrm{s}}(S), the conditional FCD, f(S∣Sℓ,δℓ)f(S|S_\ell,\delta_\ell), can be replaced with an unconditional FCD, f(S~)f(\widetilde{S}) which is obtained from the FCD of trajectories for a barrier modified as B(S~+Sℓ)−δℓB(\widetilde{S}+S_\ell)-\delta_\ell with S~=S−Sℓ\widetilde{S}=S-S_\ell (here we assume that no trajectories, or only a negligible number of trajectories, have crossed the barrier at s<Sℓs<S_\ell). At the large scale limit (Sℓ≪1S_\ell\ll1) for Sℓ≪SS_\ell\ll S, f(S~)f(\widetilde{S}) can be simply approximated as f(S)f(S) for a barrier shifted as B(S)→B(S)−δℓB(S)\rightarrow B(S)-\delta_\ell. This implies that in the Markovian case, the local bias parameters at large scales can be obtained from an equation similar to Eq. (67), but with the conditional FCD replaced by the unconditional one described above and the derivatives taken with respect to −Bc-B_{\mathrm{c}} instead of δℓ\delta_\ell. In other words, for uncorrelated trajectories, the large scale Lagrangian local bias parameters can be attributed to the response of the f(S)f(S) function to variations of the BcB_{\mathrm{c}}.

The situation, however, becomes different for correlated (non-Markovian) trajectories. Following (Paranjape & Sheth (2012a); Musso et al. (2012); Desjacques et al. (2018)), for non-Markovian trajectories, the large scale (Sℓ≪1S_\ell\ll1) Lagrangian bias parameters can be well approximated as:

bn(L)=(S×Sℓ)nbn0(L).(68)b_n^{(\mathrm{L})}=\left(\frac{S_\times}{S_\ell}\right)^n b_{n0}^{(\mathrm{L})}. \tag*{(68)}

Here, S×≡⟨δ(S)δℓ(Sℓ)⟩S_\times\equiv\langle\delta(S)\delta_\ell(S_\ell)\rangle and also Sℓ=⟨δℓ2⟩S_\ell=\langle\delta_\ell^2\rangle can be calculated from Eq. (4) for Lagrangian radii RR (corresponding to SS) and Rℓ≫RR_\ell\gg R (corresponding to SℓS_\ell). The ratio S×/SℓS_\times/S_\ell is an O(1)O(1) quantity which becomes equal to 1 for Markovian trajectories (sharp-k filter). The bn0(L)b_{n0}^{(\mathrm{L})} parameters, which are distinguished from other bias parameters that are multiplied by scale-dependent quantities that become negligible at Sℓ≪1S_\ell\ll1 limit, can be obtained from

bn0(L)=(−1)nf(S)∂nf(S)∂Bcn.(69)b_{n0}^{(\mathrm{L})}=\frac{(-1)^n}{f(S)}\frac{\partial^n f(S)}{\partial B_{\mathrm{c}}^n}. \tag*{(69)}

This relation is motivated by the first order bias relation for non-Markovian trajectories in the presence of a moving barrier derived in Paranjape & Sheth (2012a), as well as by the bias relations at different orders for non-Markovian trajectories in the presence of a constant barrier derived in Musso et al. (2012); Desjacques et al. (2018). Eqs. (68) and (69) recover the correct limits both for a constant barrier and when non-Markovian contributions become negligible.

Finally it is worth noting that these biases are real space bias parameters. In general, scale dependent contributions to bias parameters give rise to kk-dependent terms in their Fourier space relations. The large scale limit of the halo bias in Fourier space can then be recovered in the limit k≪1 Mpc−1k\ll1\,\mathrm{Mpc}^{-1}. Considering the full kk-dependent expression for the halo bias in Fourier space in Musso et al. (2012), at linear order one can write: δh(L)(k)≃b10(L)W(kR)δlin(L)(k)\delta_h^{(\mathrm{L})}(k)\simeq b_{10}^{(\mathrm{L})}W(kR)\delta_{\mathrm{lin}}^{(\mathrm{L})}(k) in the limit k≪1 Mpc−1k\ll1\,\mathrm{Mpc}^{-1}. This implies that b10(L)b_{10}^{(\mathrm{L})} can be interpreted as the Fourier space Lagrangian linear bias parameters on large scales.

The framework provided in this work for calculation of the FCD (summarized in subsection 2.4) can be used to efficiently calculate the bn0(L)b_{n0}^{(\mathrm{L})} parameters and obtain the corresponding Lagrangian bias parameters from Eq. (68). The procedure for obtaining the bn0(L)b_{n0}^{(\mathrm{L})} parameters from Eq. (69) can be described as follows:

  • For a given barrier of the form B(S)=Bc+Bs(S)B(S)=B_{\mathrm{c}}+B_{\mathrm{s}}(S) and a given filter function, one can first calculate the unconditional FCDs for a range of values of BcB_{\mathrm{c}} around the desired fiducial value. This provides, at each SS, a range of values of f(S)f(S) corresponding to different values of BcB_{\mathrm{c}} at and around the fiducial value.

  • After calculating the corresponding f(S)f(S) functions, one can evaluate their derivatives with respect to BcB_{\mathrm{c}} at any SS using straightforward numerical procedures, and obtain the corresponding bn0(L)b_{n0}^{(\mathrm{L})} values from Eq. (69).

  • Then, the results can be inserted into Eq. (68) to obtain the corresponding bias parameters.

This procedure can be used to obtain the large scale bias parameters for general choices of the barrier and filter functions. However, in some cases, these parameters can be calculated from more universal expressions. Following the discussions and results in Desjacques et al. (2018) on the derivation of the Lagrangian bias parameters based on the exact implementation of the PBS formalism through the separate-universe approach, under certain conditions and assumptions, one can find:

bn(L)=(−ν)n(δc)n1νf(ν)dn(νf(ν))dνn.(70)b_{n}^{(\mathrm{L})}=\frac{(-\nu)^n}{(\delta_{\mathrm{c}})^n}\frac{1}{\nu f(\nu)}\frac{\mathrm{d}^n(\nu f(\nu))}{\mathrm{d}\nu^n}. \tag*{(70)}

One of the key assumptions underlying the derivation of this relation is the universality of the multiplicity function, νf(ν)\nu f(\nu). That is, all the cosmological dependence is assumed to be encapsulated in the peak height parameter ν\nu, with no additional dependence on the critical density, δc\delta_{\mathrm{c}}. In addition, although this relation is derived for a constant barrier, it can also be extended to a moving barrier, B(S)B(S), provided that the barrier is assumed to remain unchanged in the presence of a long-wavelength fluctuation (Desjacques et al. (2018)). It should, however, be noted that this is, in general, an approximation since physical effects governing the behavior of a moving barrier (e.g. tidal effects) can potentially be influenced by a background fluctuation.

Testing the framework

In this part, we compare predictions for mass functions, multiplicity functions and biases based on the Hertz and Stratonovich approximations with those calculated based on numerically generated excursion-set trajectories. For the bias parameters, one can use Eqs. (68) and (69) for a given filter and barrier functions and calculate the parameters as summarized above. However, since our aim is to assess the performance of the Hertz and Stratonovich approximations by comparing their predictions with results from numerically generated large ensembles of excursion set trajectories (Monte Carlo-based results), we will be sufficient for our purpose to use the bias relation from Eq. (70). This also has the advantage of allowing us to generate Monte Carlo-based results for comparisons much more efficiently. However, in general, for non-Markovian trajectories, the multiplicity function is not guaranteed to be universal. Nevertheless, Eq. (70) should provide a reasonable approximation for a broad class of barriers. This is particularly likely to be the case for barriers for which the quantity B(S)/SB(S)/\sqrt{S}, on which the FCD depends strongly, can be expressed entirely as a function of ν=δc/S\nu=\delta_{\mathrm{c}}/\sqrt{S}.

One example is the constant barrier B=δcB = \delta_c. Another example is the family of SMT barriers, which can be written in the general form (Sheth et al. (2001)):

B(S,δc)=aδc(1+β(Saδc2)γ)(71)B(S,\delta_c) = \sqrt{a}\delta_c \left(1+\beta\left(\frac{S}{a\delta_c^2}\right)^\gamma\right) \tag*{(71)}

where δc\delta_c is usually treated as a fixed value, while aa, β\beta and γ\gamma are model parameters that can be fitted to simulation data. For the FCDs obtained with these classes of barriers, one can obtain f(ν)f(\nu) from Eq. (63) and approximate the bias parameters from Eq. (70). Ultimately, one can obtain Eulerian bias parameters using proper transformations. Here, for comparisons, we focus on the first order Eulerian bias parameter which is obtained from: b1(E)=1+b1(L)b_1^{(E)} = 1 + b_1^{(L)}. The reason we focus on the first order bias is that, besides being the most important contribution at large scales, it is also more suitable for comparisons with our Monte Carlo-based results. This is because even a small numerical fluctuation in the Monte-Carlo based multiplicity functions can get significantly amplified upon differentiation, making comparisons with our predictions for higher order bias terms difficult.

We consider both a constant barrier and the SMT barriers given by Eq. (71), using different values of the parameters aa, β\beta, γ\gamma. For the smoothing filters, we consider Top-Hat, Gaussian and W2W_2 (Musso & Sheth (2021)) filters, which are defined in Fourier space as

WTH(x)=3x3(sin⁡x−xcos⁡x)=3xj1(x),(72)W_{\mathrm{TH}}(x)=\frac{3}{x^3}(\sin x-x\cos x)=\frac{3}{x}j_1(x), \tag*{(72)}
WG(x)=exp⁡(−x22),(73)W_{\mathrm{G}}(x)=\exp\left(-\frac{x^2}{2}\right), \tag*{(73)}
W2(x)=15x2j2(x),(74)W_2(x)=\frac{15}{x^2}j_2(x), \tag*{(74)}

where jnj_n denotes the spherical Bessel function of order nn and x=kRx=kR, with RR being the Lagrangian smoothing scale.

For the numerical trajectory-based calculation of the FCD for a given smoothing filter and barrier function, we first generate logarithmically spaced values of RR and then calculate the corresponding SS values using the chosen filter function. We generate linear matter power spectra using the Core Cosmology Library (CCL)2 (Chisari et al. (2019)). The SS values are calculated over a suitable range, with a resolution sufficient for our comparisons. We then generate a large ensemble of excursion set trajectories (2×1092 \times10^9 for this work) in the SS–δ\delta plane using the Cholesky method (Nikakhtar et al. (2018)).

Next, we identify the locations of the first up-crossings of the trajectories using an appropriate interpolation method to improve the accuracy of the crossing point determinations. After identifying the first up-crossing points using the initial large set of SS values (densely sampled), we consider a new, smaller, set of SS values to calculate the FCD based on the first up-crossing points of the large set. For instance, for a given bin, [Si,Si+1)[S_i,S_{i+1}), the corresponding FCD value is calculated as

f(Smid,i)=NcrossNtotalΔS,(75)f(S_{\mathrm{mid},i})=\frac{N_{\mathrm{cross}}}{N_{\mathrm{total}}\Delta S}, \tag*{(75)}

where Smid,i=(Si+Si+1)/2S_{\mathrm{mid},i}=(S_i+S_{i+1})/2, NcrossN_{\mathrm{cross}} is the number of trajectories that have up-crossed the barrier for the first time within that bin, NtotalN_{\mathrm{total}} is the total number of trajectories and ΔS=Si+1−Si\Delta S=S_{i+1}-S_i. The reason we use the midpoint, Smid,iS_{\mathrm{mid},i}, as the point at which the corresponding FCD is evaluated is that, by definition, Ncross/NtotalN_{\mathrm{cross}}/N_{\mathrm{total}} is equal to the integral of f(S)f(S) over the bin interval and, therefore, f(Smid,i)ΔSf(S_{\mathrm{mid},i})\Delta S provides a better approximation for this integral. Also for comparisons in this work, we consider the following cosmological parameters for a flat Λ\LambdaCDM universe: Ωc=0.25\Omega_{\mathrm{c}}=0.25, Ωb=0.05\Omega_{\mathrm{b}}=0.05, h=0.68h=0.68, ns=0.97n_{\mathrm{s}}=0.97 and σ8=0.82\sigma_8=0.82.

We first compare halo mass functions obtained from Eq. (64) using predicted and numerically computed (trajectory-based) FCDs. Figure 1 shows the halo mass functions predicted from the Stratonovich approximation for the FCD, compared with their counterparts obtained based on numerically computed FCDs from Cholesky-generated trajectories. For this plot, we consider a top-hat filter and an SMT barrier, at two redshifts. Since the filter is top-hat in real space, the halo masses can be calculated from M=ρˉVM=\bar{\rho}V, where VV is the volume of a sphere with Lagrangian radius RR. One of the advantages of the

Figure 1. halo mass functions calculated from Eq. (64), based on first crossing distributions calculated from the Stratonovich approximation (blue solid lines) and numerically generated excursion set trajectories using Cholesky method (red points), at two redshifts. The botton panels show the percent difference between the Stratonovich and numerical results. The range of halo masses are chosen such that the statistical errors of the numerical values are sub-percent.

Stratonovich approximation over more recent approximations for the FCD (e.g. Verza et al. (2024)) is that it maintains consistently high accuracy in the large SS (low mass) regime with relative errors at the sub-percent or O(1%)O(1\%) levels.

Figure 2 shows the multiplicity functions and biases obtained based on Hertz and Stratonovich approximations, as well as those calculated based on numerically generated excursion set trajectories using the three different smoothing filters defined in Eqs. (72)–(74), for a constant barrier B=δc=1.686B = \delta_{\mathrm{c}} = 1.686. For comparison, the figure also shows the multiplicity function and bias obtained from a Markovian model. For the Markovian multiplicity function, we use the result based on the Press–Schechter model (Press & Schechter (1974)),

Figure 2. Halo multiplicity functions and biases obtained from different approaches for different smoothing filters, assuming a constant barrier B=δc=1.686B = \delta_c = 1.686. The predictions are based on the Hertz (green dashed lines) and Stratonovich (blue solid lines) approximations, compared with their numerical counterparts based on numerically generated excursion set trajectories using the Cholesky method (red points) and also Markovian multiplicity function and bias results based on Eqs (76) and (77) (black dotted line).

νfPS(ν)=2πνexp⁡(−ν22)(76)\nu f_{\mathrm{PS}}(\nu) = \sqrt{\frac{2}{\pi}}\nu\exp\left(-\frac{\nu^2}{2}\right) \tag*{(76)}

and for the Markovian bias, we use the Mo and White result (Mo & White [32]),

bMW=1+ν2−1δc.(77)b_{\mathrm{MW}} = 1 + \frac{\nu^2 - 1}{\delta_{\mathrm{c}}}. \tag*{(77)}

It is clear from this figure that non-Markovian effects become increasingly important as ν\nu decreases (or, equivalently, as SS increases, corresponding to smaller halo mass scales), and that the Stratonovich approximation provides the best agreement with the numerical results in capturing such effects. Figure 3 and Figure 4 show the results for SMT barriers with different values of the parameters aa, β\beta and γ\gamma. Again, one can see that the Stratonovich approximation provides the best agreement with the numerical predictions across all filter functions and barrier parameter values.

Figure 3. Halo multiplicity functions and biases obtained from different approaches for different smoothing filters, assuming an SMT moving barrier with the parameters specified in the title.

Figure 4. Similar to Figure 3, but for different values of the SMT barrier parameters.

Regarding the comparison between the Hertz and Stratonovich approximations across all figures, the performance of the Hertz approximation deteriorates as ν\nu decreases. This is expected because smaller values of ν\nu correspond to the larger-SS regime, where the probability of finding trajectories with multiple up-crossings increases. Following the arguments in § 2, the Hertz approximation is expected to perform worse than the Stratonovich approximation in correcting for the contributions of such multiple up-crossings. Regarding the Stratonovich approximation itself, we can see that it performs better at small ν\nu for SMT barriers with larger values of the parameters aa, β\beta and γ\gamma. This is because the probability of multiple up-crossings becomes smaller with such barriers compared to the constant barrier or SMT barriers with smaller parameter values. However, overall, the results across all figures indicate that the Stratonovich approximation performs best based on comparisons with the numerical results and captures the non-Markovian effects most accurately.

Conclusions

In this work, we presented an efficient method for calculating halo multiplicity functions and bias parameters based on the previously presented Hertz and Stratonovich approximations for the first crossing distribution within the excursion set framework. The procedure for calculating the first crossing distribution for a generic barrier and smoothing filter is summarized in Subsection 2.4. In § 3, we discussed how halo bias parameters can, in general, be obtained in a numerically efficient manner using the framework presented in this work. We then showed examples of halo mass functions, multiplicity functions and biases calculated based on the Hertz and Stratonovich approximations, compared with numerical trajectory-based results in Figure 1–Figure 4. Figure 1 (mass function comparisons) corresponds specifically to the Stratonovich approximation. The multiplicity function and bias comparisons considered three different smoothing filters, together with constant and SMT moving barriers with different parameter values. In all cases, the Stratonovich approximation showed the best agreement with the numerical results. Furthermore, the Stratonovich approximation maintains consistently high accuracy even at large values of the parameter SS (corresponding to the low-mass halo regime) where non-Markovian effects become increasingly important.

Given the accuracy of the Stratonovich approximation for calculation of the first crossing distribution, as well as the flexibility and computational efficiency of the framework presented in this work, this framework could provide an efficient tool for training differentiable emulators of halo abundances and bias parameters for use in cosmological analyses. Developing such emulators can be an interesting avenue for future work.

Acknowledgments

The authors acknowledge the use of AI-based tools, including ChatGPT (OpenAI) and Claude (Anthropic), for assistance with improving the clarity and readability of the manuscript. SB is partially supported by the Abdus Salam International Center for Theoretical Physics (ICTP) under the regular associateship scheme. Moreover, SB is partially supported by the Sharif University of Technology Office of Vice President for Research under Grant No. QB050522.

Data Availability

No external datasets were used in this work. All codes developed for the implementation of the framework are available upon request. This paper has been typeset from a TeX/LaTeX file prepared by the author.

References

  1. [32]Mo H. J., White S. D. M., 1996, MNRAS, 282, 347

Paper details

vlnc:W_5zkm3m93
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

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

  • Physics and Astronomy

ChatGPT, Claude

View statement
Looking for related papers…
Contents