Prepared for submission to JCAP

Marked correlation functions in

perturbation theory
arXiv:1911.06362v1 [astro-ph.CO] 14 Nov 2019

Alejandro Aviles,a,b Kazuya Koyama,c Jorge L. Cervantes-Cota,b

Hans A. Winther,d,c Baojiu Lie
a Consejo Nacional de Ciencia y Tecnologı́a, Av. Insurgentes Sur 1582, Colonia Crédito
Constructor, Del. Benito Jurez, 03940, Ciudad de México, México
b Departamento de Fı́sica, Instituto Nacional de Investigaciones Nucleares, Apartado Postal

18-1027, Col. Escandón, Ciudad de México, 11801, México.

c Institute of Cosmology and Gravitation, University of Portsmouth, Portsmouth, PO1 3FX,

UK
d Institute of Theoretical Astrophysics, University of Oslo, 0315 Oslo, Norway
e Institute for Computational Cosmology, Department of Physics, Durham University, South

Road, Durham DH1 3LE, UK

E-mail: [email protected]

Abstract. We develop perturbation theory approaches to model the marked correlation

function constructed to up-weight low density regions of the Universe, which might help
distinguish modified gravity models from the standard cosmology model based on general
relativity. Working within Convolution Lagrangian Perturbation Theory, we obtain that
weighted correlation functions are expressible as double convolution integrals that we ap-
proximate using a combination of Eulerian and Lagrangian schemes. We find that different
approaches agree within 1% on quasi non-linear scales. Compared with N-body simulations,
the perturbation theory is found to provide accurate predictions for the marked correlation
function of dark matter fields, dark matter halos as well as Halo Occupation Distribution
galaxies down to 30 Mpc/h. These analytic approaches help to understand the degeneracy
between the mark and the galaxy bias and find a way to maximize the differences among
various cosmological models.

Keywords: large scale structure formation. perturbation theory. modified gravity.

Contents

1 Introduction 1

2 Up-weighting low density regions with marked correlation functions 4

3 Perturbative treatment of the marked correlation function in Eulerian

space 4

4 Perturbative treatment of the marked correlation function in Lagrangian

space 7
4.1 The mean mark 7
4.2 The marked correlation function in CLPT 8
4.3 Approximating the CLPT marked correlation function 11

5 Degeneracies between the mark and the bias 13

6 Comparing to simulations 15
6.1 Matter particles 16
6.2 Halos and mock galaxies 18

7 Conclusions 21

A Brief review of LPT for MG 23

A.1 Hu-Sawicki model 25

B Constructing the marked correlation function 25

B.1 The CLPT marked correlation function 26
B.2 Approximating the CLPT marked correlation function 28

C k and q functions 30

D Adding curvature and tidal bias 31

1 Introduction

In studying the clustering of objects in the sky, the most important statistics is the two-point
correlation function, or its counterpart in Fourier space, the power spectrum. The reason
of this is the nearly Gaussian nature of linear fluctuations in the early stages of the Uni-
verse’s evolution. This method of neglecting nonlinear evolution has proved to be extremely
successful for explaining the anisotropies of the cosmic microwave background radiation [1].
However, nonlinearities are inherent to gravitational instability, and they dominate the evo-
lution of small scales at low redshift, such that the incorporation of higher-than-leading order
contributions to matter fluctuations has been critical for the analytical understanding, and
the construction of theoretical templates, of the processes that yield the structures we ob-
serve nowadays [2]. Despite the success of perturbation theory (PT), the improvement over
linear theory is only important at the edge of the linear regime, named the mildly nonlinear

–1–
regime. Soon after nonlinear evolution is completely onset, field fluctuations become badly
approximated by PT and their predictions become non-trustable. At this point, the direct,
brute force and computationally expensive approach of N-body simulations is the most useful
tool to study the clustering of dark matter and tracers, leading to the “true” solution of the
problem. Another promising route to study the highly non-Gaussian processes is given by
effective field theory, where small scales physics are integrated out of the theory and their
impact over the large scale incorporated under a set of free parameters to be determined by
observations [3].
A different, complementary route to the study of large-scale structure formation, which
is the subject of study in this work, relies on statistics that by construction consider mainly
linear fields, such that nonlinearities become subdominant at any time of the matter clus-
tering. To this end we construct weighted correlation functions that up-weights low density
regions in the Universe [4]. The process to do so is to define an algebraic function – the
mark function – that assigns a mark to each object under consideration (it could be any
tracer of the underlying matter density, or even the dark matter field itself), which thereafter
weights the point process on which we are interested to compute statistics; for definiteness,
the two-point statistics in configuration space. Thereafter, a marked correlation function
[4–11] is constructed by factorizing the clustering of “unmarked” tracers and the mean mark
obtained by averaging the mark of each of the sampled objects. In this way one isolates the
effects of the mark, and focuses only on the clustering of marks.
Marked statistics have by now a long history [5]; they have been used to assign proper-
ties to objects, such as the luminosity, color, and morphology of galaxies [8, 9], and to break
degeneracies between Halo Occupation Distribution (HOD) and cosmologies that arise be-
cause one is usually able to redistribute the galaxies by compensating the halo mass function
in order to obtain the same two-point correlation function [11].
On the other hand, since the discovery of the accelerated expansion of the Universe
[12, 13], tons of models that modify Einstein gravitational theory in the infrared, generically
called modified gravity (MG), were constructed in order to explain the speeding up in the
background expansion rate, as an alternative to the ΛCDM model, see refs. [14–18] among
many others. The main difficulty of constructing such models is, of course, to maintain
the success of general relativity (GR). In particular, in order to not spoil Newtonian grav-
ity in describing a wide variety of astronomical observations, MG models rely on nonlinear
mechanisms that effectively screen their associated fifth-forces in high density or strong grav-
itational potential regions [19–21]. In such a way, MG can have an important impact on
the cosmological scales and low density regions, while complying with observations at higher
densities; see e.g., refs. [22–24] for recent reviews. Motivated by these nonlinear screening
effects, Ref. [4] proposed the use of marked correlation functions that up-weight low-density,
close-to-linear regions of space; which were further studied in refs. [25–28] with MG numer-
ical simulations. The main and first step is to mark the observed objects with a function,
the “mark function” m, that smoothly under-represents them as they reside in regions with
higher and higher densities. In this way, the objects with larger assigned marks (or larger
weights) are enhanced in statistics, where the screening mechanism becomes less important,
and the MG effects, hopefully, may be captured more neatly.
In this article, building upon the work of [4], we will use first linear standard per-
turbation theory, but given that the “standard” correlation function is better modeled by
Lagrangian perturbation theory (LPT), we will develop the marked correlation function the-
ory in this frame, more precisely with Convolution-LPT (CLPT) [29], as well. Contrary

–2–
to standard perturbation theory in MG [30], the MG LPT for matter and tracers in large
scale structure formation has been developed until recently [31–36] and applied to hybrid
N-body/analytical treatments in refs. [32, 37, 38]. However, the methods developed on the
different schemes can be translated by means of a set of kernel identifications between La-
grangian and Eulerian frames [33].
During the development of the CLPT marked correlation functions theory we will clar-
ify some points in the work of ref. [4]. Namely, the relation of the mark Taylor coefficients,
C, with the resummed expansion parameters B. Furthermore, the exact CLPT of weighted
correlation functions leads to a double three-dimensional integral convolution of matter fluc-
tuations and biased tracers [see eq. (4.14) in section 4.2], that is reduced to a single convo-
lution in [4] by identifying Eulerian coordinates of linear fields with Lagrangian coordinates.
This is a consequence, and a drawback, of the use of LPT, because it relies on evolving initial
yet linear matter and tracers densities, but, on the other hand, the marks are assigned at
the moment of observation with the use of already evolved Eulerian densities. And hence,
either the use of mixing Eulerian-Lagrangian methods, or the reduction to pure Lagrangian
methods as in ref. [4], is unavoidable. Therefore, we also propose a method by which some
ingredients are kept exact within CLPT, while others use of resummations that bring them
in a form close to SPT. When comparing all the used methods, in this work, we find that
they differ by less than the 1%, even for methods that do not include loop corrections. This
shows that the use of linear theory, which is simpler for modelling and coding, is reliable for
some applications of marked statistics.
We will exemplify our findings with the representative Hu-Sawicki f (R) gravity model
[39]; specifically we will use the F6, F5 and F4 models, that are introduced in appendix
A.1. Our findings for GR make use of the exact kernels in ΛCDM, and not the widely used
static-approximation, EdS kernels. For our analytical findings we make use of a modified
version of the publicly available code MGPT,1 that computes kernels and integrates a set of
functions necessary to build the correlation functions and power spectra for biased tracers at
1-loop in MG; see ref. [33] for details. All our results are shown at redshift z = 0.5 and we
use WMAP 9yr best fit parameters given by Ωb = 0.046, Ωm = 0.281, h = 0.697, σ8 = 0.82
and ns = 0.971 [40].
The rest of this work is organized as follows: In section 2 we define the mark correlation
function and tracers of the dark matter field, in section 3 we formally introduce the mark,
explain the renormalized bias and establish the relationship of the mark Taylor expansion
parameters (C’s) and the resummed expansion parameters (B’s) to be able to compute the
Euleran correlation of weighted tracers and its corresponding marked correlation function. In
section 4 we compute the marked correlation function in CLPT, and in order to integrate it
we use an approximation in section 4.3. In section 5 we discuss the effects and degeneracies
of bias and mark parameters. In section 6 we compare our analytical results to ΛCDM
and Hu-Sawicki f (R) models making use of Extended LEnsing PHysics using ANalaytic
ray Tracing (elephant) simulations [41]. In section 7 we further discuss and conclude on
our perturbative approach to marks. Finally, in the appendices we added complementary
material on the general theory of LPT for MG (appendix A), cumbersome formulae of the
CLPT formalism to 1-loop and approximations (appendix B), the computation of the different
statistics needed in the main text formulae (appendix C), and finally we put forward a way
to add curvature and tidal bias to the formalism (appendix D).

1
https://github.com/cosmoinin/MGPT

–3–
2 Up-weighting low density regions with marked correlation functions

A marked correlation function (mCF) [4, 5, 7–11] is defined as the sum of pairs of objects
separated by a distance r, weighted by the ratio of the mark function value to the mean mark
mi /m̄ at each point and divided by the number of pairs n(r):
X mi mj
M(r) = . (2.1)
n(r)m̄2
ij|rij =r

That is, it is a 2-point statistics of the clustering of marks. We choose the mark to be a
function of the environmental matter density in which such objects reside, smoothed over
a scale that we take to be larger than the size of the objects. We should note that in real
applications, one rarely has the dark matter densities at hand, so we will relax this assumption
later so that our formalism can be applied to halos and galaxy mocks. In viable chameleon
modified gravity theories, high energy density regions are screened and these models reduce
to General Relativity. Hence, the effects of MG are expected to be more pronounced in low
density regions. For this reason marked functions that enhance low density regions have been
considered in the MG literature. We write such a relation as

m(ρ) = G[δR (x)], (2.2)

with δR (x, t) = dx0 WR (|x − x0 |/R)δ(x0 ) a smoothed matter overdensity and R the smooth-
R

ing scale. The mark function (hereafter, the White-mark)

 p
1 + ρ∗
m= (2.3)
1 + ρ∗ + δR
is proposed in ref. [4], with p and ρ∗ dimensionless parameters chosen in order to up-weight
low density regions. Other marks have been also proposed; see refs. [25–27].
The objects we consider are tracers of the dark matter field δ. Matter and tracers X
overdensities are related through the bias function2

1 + δX (x) = Fx [δ(x)]. (2.4)

It is known that a well defined bias expansion within chameleon theories must contain higher-
order derivative operators (∇2 δ, ∇4 δ, and so on); see e.g. section 8.3 of [43] and ref. [33].
Hence, we will introduce curvature bias, as well as tidal bias, later in appendix D. In order
for some quantities to be well defined, we also smooth matter overdensities with a scale RΛ
that we take to be ideally RΛ  R, hence we simply denote δ = δΛ ; at the end, in numerical
calculations we set RΛ = 0.

3 Perturbative treatment of the marked correlation function in Eulerian

space

For a perturbative treatment we expand the mark in a power series of the smoothed over-
density
1 2
m(δR ; Ci ) = C0 + C1 δR + C2 δR + ··· (3.1)
2
2
This relation has a stochastic nature, that can be made explicit by writing Fx [δ(x); x] [42]. Dependencies
on other bias operators can be also introduced as arguments in F .

–4–
with Ci = G(i) [0], the ith derivative of G[δR ] evaluated at δR = 0. These parameters are
intended to enhance low density regions where screenings are not very efficient and the
effects of MG might be detected. For the White-mark in eq. (2.3) one obtains C0 = 1,
C1 = −p/(1 + ρ∗ ), C2 = p(p + 1)/(1 + ρ∗ ). The fewer parameters are needed to model
the mark, the better the convergence will be. We notice that C1 < 0 enhances low density
regions.
It is well known that the correlation function is better modeled by LPT (see e.g.
ref. [44]), however to gain insight and as a warm-up for Lagrangian calculations, in this
section we compute the marked correlation function in Eulerian space. The mean mark is
given by mark weighted by the tracers density field,

m̄ = hG[δR (x)](1 + δX (x))i

D C2 2  c2 E
= C0 + C1 δR (x) + δR (x) + · · · c0 + c1 δ(x) + δ 2 (x) + · · ·
2 2
2 1 2 1 2
= c0 C0 + c1 C1 σR + c0 C2 σRR + C0 c2 σ + · · · , (3.2)
2 2
(i)
with ci = Fx [0], and zero-lag correlators defined as

σ 2 ≡ h(δ(0))2 i, 2
σR ≡ hδR (0)δ(0)i, 2
σRR ≡ h(δR (0))2 i. (3.3)

Below, we use also the correlation and cross-correlation functions

ξ(r) ≡ hδ(x)δ(x + r)i, ξR (r) ≡ hδR (x)δ(x + r)i, ξRR (r) ≡ hδR (x)δR (x + r)i. (3.4)

Now, it is convenient to write eq. (3.2) in terms of renormalized bias parameters [45]
Z
E dλ −λ2 σ2 /2 ˜
bn = e Fx (λ)(iλ)n , (3.5)
2π

where F̃x (λ) is the Fourier transform of Fx with spectral parameter λ (dual to δ). This
definition of bias is more commonly used in LPT, so we are using the label “E” to distin-
guish Eulerian from Lagrangian biases. Equation (3.5) leads to the precise resummation of
bare bias
P∞parameters cn that yields the renormalized bias parameters, becoming related by
bE
n = c
k=0 n+2k σ 2k /(2k k!) [46, 47]. Analogously, we introduce the “resummed” expansion

parameters Bn as
Bn∗
Z
∗ dΛ −Λ2 σ2 /2
Bn = ∗ with Bn = e RR G̃(Λ)(iΛ)n , (3.6)
B0 2π

where G̃(Λ) is the Fourier transform of G,Pand Λ is a spectral parameter, dual to δR , and
find that Bn∗ and Cn are related by Bn∗ = ∞ 2k k
k=0 Cn+2k σRR /(2 k!). Therefore, the expansion
parameters and the Taylor coefficients of the mark function relate as
P∞ 2k k
2 k=0 Cn+2k σRR /(2 k!)
Bn (Cn , σRR ) = P ∞ 2k k
. (3.7)
k=0 C2k σRR /(2 k!)

Inserting the bE
n and Bn parameters in eq. (3.2) we get

m̄ = B0∗ 1 + bE 2
 
1 B1 σR + · · · , (3.8)

–5–
which is the mean mark reported in ref. [4]. We have defined the expansion parameters in
this way because the factor B0∗ is absorbed by the mean mark and we do not have to carry it
in all expressions for the marked correlation function. Moreover, typically C0 = 1, and one
smooths over large regions, such that σRR 2  1 and B ' B ∗ ' C .
n n n
We compute the correlation of weighted (by the mark function m) tracers separated by
a distance r = |x2 − x1 |,



hG1 Fx,1 G2 Fx,2 i ≡ hG[δR (x1 )] 1 + δX (x1 ) G[δR (x2 )] 1 + δX (x2 ) i
Z
dλ1 dλ2 dΛ1 dΛ2 i(λ1 δ1 +Λ1 δR,1 +λ2 δ2 +Λ2 δR,2 )
= he iF̃x (λ1 )F̃x (λ2 )G̃(Λ1 )G̃(Λ2 )
(2π)4
Z
dλ1 dλ2 dΛ1 dΛ2 1 2 2 2 1 2 2 2
= 4
F̃x (λ1 )F̃x (λ2 )G̃(Λ1 )G̃(Λ2 )e− 2 (λ1 +λ2 )σ − 2 (Λ1 +Λ2 )σRR
(2π)
h i
2
× 1 − (λ1 Λ1 + λ2 Λ2 )σR − λ1 λ2 ξ(r) − Λ1 Λ2 ξRR (r) − (λ1 Λ2 + λ2 Λ1 )ξR (r) + · · ·
h i
2 E 2 2 E
= m̄ 1 + (b1 ) ξ(r) + B1 ξRR (r) + 2b1 B1 ξR (r) + · · · , (3.9)

where the ellipsis denote second order terms in ξ,R,RR (r). In the above equation Gi and Fx,i
denote G[δR (xi )] and Fx [δ(xi )] and we have shifted to Fourier space (δ → λ, δR → Λ) in the
second equality. In the third equality we used the cumulant expansion theorem,
" ∞ #
X iN
heiX i = exp hX N ic , (3.10)
N!
N =1

and expanded out of the exponential all terms but those containing σ 2 and σRR2 , such that

we can use eqs. (3.5) and (3.6) to get biases from spectral parameters, as we did in the last
equality. The mCF is obtained by multiplying the above equation by m̄−2 and by taking the
ratio to the correlation function of tracers. We obtain
1 + (bE 2 2 E
1 ) ξ(r) + B1 ξRR (r) + 2b1 B1 ξR (r) + · · · 1 + W (r)
ME (r) = E 2
≡ , (3.11)
1 + (b1 ) ξ(r) + · · · 1 + ξX (r)
such that the weighted correlation function is given by
1
1 + W (r) = hG1 Fx,1 G2 Fx,2 i. (3.12)
m̄2
We emphasize that we are weighting the tracer overdensities Fx with the mark computed
using the total matter smoothed overdensities. We notice that the zero-lag correlators σ 2
and σRR2 do not appear in the mCF as is guaranteed because we are using renormalized b
and B parameters. Meanwhile, the cross-covariance σR 2 is canceled out by the mean mark

squared appearing in the definition of the mCF in eq. (2.1). We remark that a process of
renormalization of the Taylor expansion coefficients C is not strictly necessary because the
scale R is not arbitrary, but is chosen from the beginning to assing the mark to the tracers.
However, we have proceeded by doing so, in order to have a simpler structure of equations
that do not carry the variances σRR2 and matches the notation of ref. [4].
As mentioned previously, in real applications, we have to use a mark computed by the
number density of galaxies. An advantage of using the renormalized B parameters is that
the effect of this reassignment can be included simply by re-scaling the B parameters. For
the application to dark matter halos and galaxies, we will treat them as free parameters and
we will fit these parameters from simulations.

–6–
4 Perturbative treatment of the marked correlation function in Lagrangian
space
In Lagrangian space one considers regions of space, at an initial early time with spatial
coordinates q, that give rise to tracers. We relate matter and tracers overdensities, in an
analogous way to eq. (2.4), by
1 + δX (q) = F [δ(q)]. (4.1)
The function Fx , introduced in eq. (2.4), and F are not simply related, but assuming number
conservation of tracers, (1 + δX (x))d3 x = (1 + δX (q))d3 q, one obtains
d3 k
Z Z Z
3 ik·(x−q) dλ
Fx [δ(x)] = 3
d qe F̃ (λ)eiλδ(q)−ik·Ψ(q,t) , (4.2)
(2π) 2π
with Ψ(q) the Lagrangian displacement vector, that maps Lagrangian coordinates q to Eu-
lerian coordinates x as x(q, t) = q + Ψ(q, t). Equivalently to eq. (3.5), we introduce the
renormalized Lagrangian local biases [45] with
Z
dλ −λ2 σ2 /2
bn = e F̃ (λ)(iλ)n . (4.3)
2π
In this section we evolve initially biased tracers with overdensity δX (q) using Convolution
Lagrangian perturbation theory and thereafter assign them a mark m[δR (x)] at the moment
of observation, where R is the physical size of the (Eulerian) region that hosts the objects and
that is chosen from the beginning. The use of the Lagrangian approach has some advantages
with respect to the Eulerian. In the first place it is well known that the two-point correlation
function is poorly modeled within the Eulerian approach, particularly at the BAO peak
position; second, the (renormalized) Lagrangian bias parameters are obtained through the
peak background split prescription [42, 48–50], and hence are physically appealing. But the
price to pay is that the mCF is cumbersome to compute, and one must evaluate several
6-dimensional integrals [eq. (4.14) below]. However, we will assume that the smoothing scale
is large, R > 1/kNL , and that the mark efficiently up-weights regions of low density, such
that we can deal with the smoothed matter overdensity δR as a linear field; moreover, in
subsection 4.3 we propose a model to approximate the mCF, and we compare our results to
ref. [4], finding a good agreement between both methods.

4.1 The mean mark

The mean mark can be computed in Lagrangian space as
d3 k 3 −ik·q dΛ dλ
Z Z
m̄ = hG[δR (0)]Fx [δX (0)]i = d q e G̃(Λ)F̃ (λ)heiΛδR (0)+iλδ(q)−ik·Ψ(q,t) i,
(2π)3 2π 2π
(4.4)
where we used eq. (4.2) to relate Lagrangian and Eulerian tracer fluctuations. We now use
the cumulant expansion theorem [eq. (3.10)] with X = ΛδR (0) + λδ(q) − k · Ψ(q, t). We have
1 1 1 1
− hX 2 ic = − Λ2 σRR
2
− λ2 σ 2 − λΛξR (q) + Λki UiR (q) − k 2 σΨ
2
, (4.5)
2 2 2 2
with the dispersion of Lagrangian displacements
Z ∞
2 1 1
σΨ (t) = δij hΨi (0)Ψj (0)ic = 2 dpPL (p, t), (4.6)
3 6π 0

–7–
where PL (k) is the linear matter power spectrum, and

d3 p ip·q pi
Z
R
Ui (q) = hδR (0)Ψi (q)ic = e W̃R (p)PL (p), (4.7)
(2π)3 p2

where the second equality is valid to linear order. Expanding correlators at finite separation
and using eqs. (3.6) and (4.3) we have

d3 k
Z Z h i
∗ 3 −ik·q −k2 σΨ2 /2 R
m̄ = B0 d qe e 1 + b1 B 1 ξ R (q) − iB 1 ki Ui (q) + · · ·
(2π)3
" #
Z
d3k
2 2
= B0∗ 1 + (1 + b1 )B1 e−k σΨ /2 W̃R (k)PL (k) + · · ·
(2π)3
" #
= B0∗ 1 + (1 + b1 )B1 σRσ
2
Ψ
+ ··· , (4.8)

where we defined
d3 k ik·r −k2 σ2 /2
Z
2
ξRσΨ (r) ≡ e e Ψ W̃R (k)PL (k), σRσ ≡ ξRσΨ (0). (4.9)
(2π)3 Ψ

For large R, or at early times, σΨ  R and we get m̄ = B0∗ 1 + (1 + b1 )B1 σR 2 , which is

 

eq. (3.8) with the proper identification bE

1 = 1 + b1 . At late times, however, σΨ grows and
becomes important, making the mean mark evolve at a slower rate. Physically, the result
of eq. (4.8) follows from the fact that we are correlating a linearly evolved smoothed matter
2 2
density field ∼ W̃R δ with a Lagrangian evolved field ∼ e−k σΨ /2 δ.3

4.2 The marked correlation function in CLPT

The weighted correlation function of biased tracers (weighted by the mark function over the
mean mark), is given by eqs. (3.12). Using eq. (4.2) to substitute Eulerian by Lagrangian
tracers, and transforming G[δR (x1,2 )] to Fourier space, we have
Z 3
d k1 d3 k2
Z Z
3 3 dλ1 dλ2 dΛ1 dΛ2
hG1 Fx,1 G2 Fx,2 i = 3 3
d q1 d q2 F̃ (λ1 )F̃ (λ2 )G̃(Λ1 )G̃(Λ2 )
(2π) (2π) 2π 2π 2π 2π
eik1 ·(x1 −q1 ) eik2 ·(x2 −q2 ) hei(λ1 δ(q1 )+λ2 δ(q2 )+Λ1 δR (x1 )+Λ2 δR (x2 )) i. (4.10)

As a standard practice we use the cumulant expansion theorem and expand out of the ex-
ponential all terms containing spectral parameters, with the exception of variances σ 2 and
2 , and linear terms in the Lagrangian displacements. Performing the λ and Λ integrations
σRR
with the aid of eqs. (3.6) and (4.3) we obtain
Z 3
d k1 d3 k2
Z
hG1 Fx,1 G2 Fx,2 i = d3 q1 d3 q2 eik1 ·(x1 −q1 ) eik2 ·(x2 −q2 )
(2π)3 (2π)3
1 2 σ2 1 i j L
e− 2 (k1 +k2 ) Ψ e 2 k1 k2 Aij (q) (1 + I) (4.11)

where the “1” gives the Zeldovich Approximation standard correlation function and the
function I = I(x1 , x2 , q1 , q2 , k1 , k2 ) contains the biased and marked densities and correlations
2 2
3
The Zeldovich approximation matter power spectrum is roughly PZA (k) ≈ e−k σΨ
PL (k) [51].

–8–
of non-linear Lagrangian displacements. A complete expression for the function I up to 1-
loop in fluctuations and third order in bias expansion is given in eq. (B.1) of appendix B.
The matrix [29]
d3 p
Z

pi pj
ALij (q) = 3
2 − eip·q − e−ip·q PL (p) (4.12)
(2π) p4
is the leading order piece of the correlation of displacement fields h∆i ∆j ic , with ∆i = Ψi (q2 )−
Ψi (q1 ). We now redefine variables
1
q = q2 − q1 , Q = (q1 + q2 ),
2
1
r = x2 − x1 , R = (x1 + x2 ),
2
1
ka = (k2 − k1 ), kb = k1 + k2 , (4.13)
2
in eq. (4.11), and perform Gaussian integrations in ka and kb (see e.g. appendix C of [29])
to get the weighted correlation function in CLPT
Z 3 − 1 (r−q)T A−1 (r−q) Z 3 − 1 (R−Q)T C−1 (R−Q) 
1 d qe 2 L d Qe 2 
1 + W (r) = 2 1 + I , (4.14)
m̄ (2π)3/2 |AL |1/2 (2π)3/2 |C|1/2
with matrix
2 1
Cij (q) = σΨ δij − AL (q). (4.15)
4 ij
The complete expression for the function I(r, q, R, Q) is given in eq. (B.8). We notice that
if 1 + I is not a function of Q, it can be pulled out of the Q integral and due to that
Cij depends only on q, the integration over Q gives 1. This is the case of the “standard”
correlation function in CLPT, where by statistical homogeneity we can shift all arguments
of the fields inside the correlator in eq. (4.11) by a vector −q1 and, since the arguments x1,2
are not present, we obtain I = I(q), reducing the double Gaussian convolution in eq. (4.14),
to a single three-dimensional convolution.
In figure 1 we show contour plots for the two Gaussian kernels appearing in eq. (4.14).
GK-q is the kernel of the q-integral and GK-Q the kernel of the Q-integral. These are plotted
as a function of rk (Rk ) and r⊥ (R⊥ ), the components of r (R) parallel and perpendicular
to the Lagrangian coordinate q (Q) with Q = q = 100 Mpc/h fixed.4 We notice that
both kernels have their maximum value at Q = q = 100 Mpc/h, but the GK-Q is more
sharply peaked because the determinants of the correlation matrix comply with |C| < |AL |.
This observation suggests to replace the Gaussian kernel GK-Q by a Dirac Delta function
δD (R − Q), allowing us to perform the d3 Q integral to obtain
Z 3 − 1 (r−q)T A−1 (r−q) 
d qe 2 L 1 
hG1 Fx,1 G2 Fx,2 i = 1 + J (r, (r ± q)) , (4.16)
(2π)3/2 |AL |1/2 2
with J given in eq. (B.20). Now, by noticing from figure 1 that the largest contribution to
the above integral is given by q = r, we replace the argument 21 (r ± q) by q or 0, either if
± = + or −. Following these lines, in appendix B we arrive at
d3 q
Z
− 12 (r−q)T A−1
 
W16 L (r−q) 1 + J 0 (q, r)
1+W (r) = e (4.17)
(2π)3/2 |AL |1/2
4
Reference [52] shows that GK-q gives the probability of finding two dark matter particles separated by a
distance r, given that they were separated by a distance q at an earlier time.

–9–
 120

GK -q

110

r∥ ,R∥ [Mpc/h]
GK -Q
100

90

80

-20 -10 0 10 20
r⊥ ,R⊥ [Mpc/h]

Figure 1. Gaussian kernels in eq. (4.14). GK-q is the kernel of the q-integral (black curves) and
GK-Q the kernel of the Q-integral (red curves). Dashed and solid curves show the regions that enclose
the 68% and 95% of the volume respectively. A similar plot can be found in [52].

where the label “0” in the J function indicates that we have factorized zero-lag correlators
σR that are canceled out by the square of the mean mark. To third order in fluctuations and
bias expansion, this function is given by
1 1
1 + J 0 (q, r) = 1 − Aloop Gij + Wijk Γijk + b21 ξL − 2b1 Ui gi − (b2 + b21 )Ui Uj Gij
2 ij 6
− 2b1 b2 ξL Ui gi + B1 ξRR − 2B1 UiR gi − (B2 + B12 )UiR UjR Gij − 2B1 B2 ξRR UiR gi
2

+ 2b1 B1 ξR − 4b1 B1 Ui UjR Gij − 2b21 B1 ξL UiR g1 − 2B12 b1 ξRR Ui gi

− 2(b2 + b21 )B1 ξR Ui gi − 2(B2 + B12 )b1 ξR UiR gi − b1 A1000 0010
ij Gij − B1 Aij Gij
− b2 Ui2000 gi − B2 Ui0020 gi − b21 Ui1100 gi − B12 Ui0011 gi − 2B1 b1 (Ui1010 + Ui1001 )gi . (4.18)
Hence, eqs. (4.17) and (4.18) give the higher order generalization of the linear expression
found in [4]; hereafter we will refer to this perturbative model as W16, and to the linear
pieces as the Zeldovich Approximation (ZA) mCF. In the above equation the r dependence
appears only through the tensors gi = (A−1 −1
L )ij (qj − rj ), Gij = (AL )ij − gi gj and Γijk =
(A−1
L ){ij gk} − gi gj gk , while ξ, U , A and W are functions of q only. U (q) functions are defined
as
mnpq(r) p q (r)
Ui = hδ m (q1 )δ n (q2 )δR (q1 )δR (q2 )∆i ic . (4.19)
The numbers m, n, p, q denote powers, and (r) is the perturbation theory order of the dis-
placement fields. U ≡ U 1000 and U R ≡ U 0010 . The density fields are linear and are assumed
Gaussian, thus only a few U functions do not vanish. Similarly we have defined
mnpq(rs) p q (r) (s)
Aij ≡ hδ m (q1 )δ n (q2 )δR (q1 )δR (q2 )∆i ∆j ic , (4.20)
0000(11) 0000(22) 0000(13)
and AL
ij = Aij , Aloop
ij = Aij +2Aij . Also, we have used the 3-point correlations
(2) (1) (1)
Wijk = h∆i ∆j ∆k ic + cyclic perm. (4.21)

– 10 –
Expressions for some of these U , A and W functions are known in the literature (see refs. [29,
53] for ΛCDM and ref. [33] for MG), and in appendix C we show explicit expression for the
rest. The mCF is
1 + W W16 (r; bi , Bi )
MW16 (r; bi , Bi ) = CLPT (r; b )
, (4.22)
1 + ξX i
while
CLPT
ξX (r, b1 , b2 ) = W W16 (r; b1 , b2 , B1 = 0, B2 = 0) (4.23)
is the known biased tracers correlation function in CLPT [29] with the resummation of
ref. [54].
We notice that the approach developed in this subsection and in appendix B to arrive
to eq. (4.17) is equivalent to substituting x1,2 by q1,2 in the arguments of the fields inside
the correlator of eq. (4.10) from the beginning. For example, one of the contributions to
the function I that contains both Lagrangian ond Eulerian coordinates is ξR (x2 − q1 ). As
long as the region over which the Gaussian kernels GK-q and GK-Q are non-negligible is
smaller than the support of the smoothing kernel WR , we can approximate ξR (|x2 − q1 |) '
ξR (q) + α(σΨ2 /R2 )∇2 ξ (q) + · · · , with α a constant depending on the particular spherically
R
symmetric kernel WR . If we drop terms of order σΨ 2 /R2 from the contributions to the I

function we arrive to eq. (4.17).

However, a simple substitution of λδ(q1 )+ΛδR (x1 ) → λδ(q1 )+ΛδR (q1 ) in the integrand
of eq. (4.10), is not justified even for linear density fields because the smoothed procedure
is non local, and all matter particles at xa within the region |xa − x1 | < R contribute. So,
the above substitution requires that about the same particles contribute to the smoothing in
the initial slice, |qa − q1 | < R. Hence in addition to the linear evolution of δR we have to
impose that |Ψa − Ψ1 | ∼ σΨ < R. Clearly, to have a linear evolution for δR requires a large
smoothing scale R, but the latter condition is independent and in general more restrictive.
In the next subsection, we will discuss a different approximation to handle the double
convolution without using the substitution of x1,2 by q1,2 .

4.3 Approximating the CLPT marked correlation function

Solving numerically the double convolution in eq. (4.14) is challenging and we do not attempt
it in this work. Instead, in this subsection we put forward an approximation method that
splits the function I in eq. (4.11) into pieces containing q, pieces containing combinations of
qi and xj , and pieces containing r = x2 − x1 :
I = Ir (r) + Iq (q) + Ir,q (r, q). (4.24)
The first and second functions can be integrated by the standard methods of CLPT. For
example, the term ξRR (r) ∈ Ir yields B12 ξRR (r)(1 + ξZA (r)) ∈ 1 + W . The function Iq (q)
gives 1+ξXCLPT (r). For the terms containing both r and q in their arguments we approximate
i i

1 i j L 1
e 2 k1 k2 Aij (q) ≈ 1 + k1i k2j AL
ij (q), (4.25)
2
and solve
d3 k1 d3 k2
Z Z
1 2 2
hG1 Fx,1 G2 Fx,2 i r,q
= d3 q1 d3 q2 eik1 ·(x1 −q1 ) eik2 ·(x2 −q2 ) e− 2 (k1 +k2 ) σΨ
(2π)3 (2π)3
 
1 i j L
× Ir,q 1 + k1 k2 Aij (q) . (4.26)
2

– 11 –
This approximation allows us to write 1 + W as a sum of Fourier transforms of known scalar
functions constructed out of 2- and 3-point correlation functions of Lagrangian displacements.
We obtain (see appendix B.2)
CLPT
1 + W = 1 + ξX (r) + B12 ξRR (r)(1 + ξZA (r)) + 2b1 B1 x̄ξR (r) + 2B1 x̄UR (r)
+ (B2 + B12 )x̄UR UR + 2B1 B2 ξRR (r)x̄UR (r) + 4b1 B1 x̄U UR (r) + 2b21 B1 x̄ξUR (r)
+ 2B12 b1 ξRR (r)xU (r) + 2(b2 + b21 )B1 x̄ξR U (r) + 2(B2 + B12 )b1 x̄ξR UR (r)
+ B1 x̄A0010 (r) + B2 x̄U 0020 (r) + B12 x̄U 0011 (r) + 2B1 b1 x̄U 0110 (r) + 2B1 b1 x̄U 1010 (r). (4.27)

The dominant x̄ functions above are

[1,1] [1,−1]
x̄ξR (r) = ξRσ2 (r) + ξRσ2 (r)ξL (r) − ξRσ2 (r)ξL (r), (4.28)
Ψ Ψ Ψ

loop 4 2 [2,0] [2,0]

x̄UR (r) = ξRσ2 (r) + ξRσ 2 (r) + ξRσ2 (r)ξL (r), + ξRσ2 (r)ξL (r)
Ψ Ψ 3 Ψ 3 Ψ
[1,1] [1,−1] [1,−1] [1,1]
− ξRσ2 (r)ξL (r) − ξRσ2 (r)ξL (r), (4.29)
Ψ Ψ

where Z ∞
[m,n] 1 2 σ 2 /2
ξRσ2 (r) = dp p2+m e−p Ψ W̃R (p)PL (p)jn (pr), (4.30)
Ψ 2π 2 0
are generalizations of the correlation function (see e.g. [55, 56]). The rest of the x̄ functions
are given in appendix B.2. Equivalently to eq. (4.22), the mCF is

1 + W (r; bi , Bi )
M(r; bi , Bi ) = CLPT (r; b )
. (4.31)
1 + ξX i

Figure 2 shows the mCF using different analytical methods for the MG Hu-Sawicki
F5 model: solid black curve is the one presented in this subsection (“This Work”, from
now on); dashed blue the method of eq. (4.17), which is that introduced in ref. [4] plus
1-loop contributions; in dot-dashed red, the Zeldovich Approximation, which in this work
means the linear model of ref. [4]; and, in long-dashed brown, the Eulerian linear model of
eq. (3.11). We are using Lagrangian local biases b1 = 1.2 and b2 = 0.2, and mark parameters
B1 = −0.7, B2 = B12 . In the lower panel we show the relative difference between the models
and the ZA. The agreement between them is very good, even for the linear Eulerian theory,
below 1% for scales above ∼ 20 Mpc/h. The relative differences between the CLPT standard
correlation function with and without loop contributions is shown in the lower panel of figure 2
(red dotted curve); by comparing it with the dashed blue curve (which shows the relative
differences of the mCF W16+1-loop with the ZA) we confirm that mCFs which efficiently
enhance low density regions in the sky are indeed more linear than the standard correlation
function.
To summarise, “This Work” method is a mixture of Lagrangian and Eulerian schemes
that splits the function I according to its arguments. The pieces depending only on co-
ordinates q and on coordinates x are treated exactly (within CLPT), and hence more ap-
pealing than the method of [4], which relies on the substitution of x1,2 by q1,2 . Mean-
while, the Ir,q piece relies on the approximation of eq. (4.25) which is not well justified since
1 i j L 2 ˆ ˆ 2
2 k1 k2 Aij (q → ∞) → σΨ k1 k2 (k1 · k2 ), and σΨ is not small. However, this is not as catas-
trophic as it appears because the oscillatory nature of the integrand in eq. (4.26) gives much
more weight to low values of k1 and k2 . On the other hand, a mixture of Lagrangian and

– 12 –
 1.00

0.95
W16 + loops
0.90

0.85 This work

ℳ(r )
0.80 Eulerian

0.75 ZA
0.70

0.65

10
-1 |ξCLPT/ξZA-1|

-2
10
|ratio -1|

-3
10

-4
10

-5
10
10 20 50 100
r [Mpc/h]

Figure 2. Comparison of perturbative models of sections 3, 4.2, and 4.3. The upper panel shows
the marked correlation functions with parameters b1 = 1.2, b2 = 0.2, B1 = −0.7, B2 = B12 , for
the different methods: solid black curve is the one presented in this subsection [eq. (4.27)]; dashed
blue the method of eq. (4.17), which is the method introduced in ref. [4] plus 1-loop contributions;
in dot-dashed red, the ZA; and in long-dashed brown, the Eulerian linear model of eq. (3.11). The
lower panel shows the relative differences with respect to the ZA. The gray dot-dashed horizontal line
denotes the 1% differences. The red dotted curve in the lower panel shows the relative differences
between the CLPT 1-loop and ZA standard correlation functions.

Eulerian schemes has a disadvantage in describing the BAO peaks in the correlation func-
tion. In section 6, we will verify various approaches to model the marked correlation function
shown in figure 2 using the measurements from N-body simulations.

5 Degeneracies between the mark and the bias

As discussed above, a mark that enhances low-density regions typically has B1 < 0, meaning
that these terms contribute by lowering the mCF, the smaller B1 the larger the effect. In the
left panel of figure 3 we show this effect on the mCF for matter (bi = 0) for the gravitational
models GR, F6, F5 and F4. We notice that for models that depart more from ΛCDM, the
effect is more pronounced, which is the consequence of a enhanced clustering in f (R) than
in GR. This behavior has been observed recently in simulations [25].
In the right panel of figure 3, we plot the mCF for biased tracers with linear and second
order biases b1 = −0.3, 0, 0.5 , 1.5, b2 = 0, and mark parameters fixed to B1 = −1, B2 = 0.
Contrary to what may be naively expected, the effect of a greater large-scale bias is similar
to making B1 more negative, bringing down even more the mCF. This is because larger bias
increases functions W and ξX of the mCF, while B1 only affects W .
On the other hand, it has been shown that in MG models that produce more clustering
the linear local bias parameters have smaller values than those obtained in ΛCDM [33, 57].
This can be understood because typically local biases depend on two parameters: the variance
of linear fluctuations in a given region of Lagrangian radius R∗ , σ 2 (M ), and the density

– 13 –
 1.0
1.00

B1 =-0.2
0.95
0.9

0.90

GR b1 =-0.3
0.8
ℳ(r)

ℳ(r)
0.85
F6 b1 =0

F5 b1 =0.5
0.80 B1 =-1 0.7
F4 b1 =1.5

0.75

0.6
0.70

10 20 50 10 20 50

r [Mpc/h] r [Mpc/h]

Figure 3. Marked correlation function for matter for GR, F6, F5 and F4 models. In the left panel
we consider matter (b1 = b2 = 0) and set B1 = −1, B2 = 0 (lower curves) and B1 = −0.2, B2 = 0
(upper curves). The right panel uses F5 model with fixed B1 = −1, B2 = 0, b2 = 0 and show the
results with Lagrangian local biases b1 = −0.3, 0, 0.5 , 1.5 from top to bottom.

3.0 2.0

LCDM 1.5 LCDM

2.5
F6 F6
F5 1.0 F5
2.0 F4 F4
1 + b1

b2

0.5

1.5
0.0

1.0
-0.5

0.5 -1.0

12 13 14 15 12 13 14 15

Log10 [(M/M⊙ h -1 )] Log10 [(M/M⊙ h -1 )]

Figure 4. Large scale bias bLS = 1 + b1 and second order local Lagrangian bias b2 binned over halo
mass intervals, for GR, F6, F5 and F4 models.

threshold for collapse δc (M ), with M the mass enclosed by a spherical region of radius R∗ .
The former is bigger in MG than in GR because there is a major clustering, PLMG > PLGR ,
while the latter is smaller in MG because of the extra attractive fifth-force; note also that
a violation of Birkhoff theorem in MG implies the density threshold for spherical collapse
becomes mass dependent. Schematically, these two quantities appear in the combination
νc = δc (M )/σ 2 (M ), hence one expects to obtain bMG < bGR . In figure 4 we show the linear
and second order local biases for halos obtained with the Sheth-Tormen like prescription of
[33] and averaged over mass bins as in refs. [35, 45]. We interpret figure 4 as a more rapid
relaxation of linear bias in MG than in GR due to a more efficient formation of massive halos.
Although this method to obtain biases in chameleon MG theories from the peak-
background-split prescription gives reasonable values for the local bias parameters [33, 35],
we will find later in section 6 that it is not sufficiently accurate for our purposes on the

– 14 –
 1.00

0.95 Matter

0.90

0.85 GR

ℳ(r)
F6
0.80
F5

0.75 F4

0.70 Halos (14<Log10 M <14.5)

0.65

10 20 50 100
r [Mpc/h]

Figure 5. Marked correlation function for matter and tracers for models ΛCDM, F6, F5 and F4
at redshift z = 0. We fix B1 = −1, B2 = 0 and take the bias values from figure 4 for the interval
1014 < M/M < 1014.5 : bGR 1 = 1.15, bF6 F5 F4
1 = 1, b1 = 0.75, b1 = 0.5. The upper curves are for matter
and the lower curves for halos.

marked correlation function, and hence we will be forced to fit the biases directly from the
simulations. Therefore, figure 4 should be taken mainly as indicative, to gain intuition on
the behaviour of the relative bias values in different gravitational theories.
In figure 5 we show plots for matter and halos with masses in the interval 1014 <
M/M < 1014.5 . In this case we find that while for matter MG mCFs lie below that of
GR, for tracers the opposite happens, the halos MG mCFs are brought above that of GR.
This effect of inversion of the trends for the mCFs for tracers and matter can be interpreted
by considering the mean mark, m̄ ≈ (1 + b1 )B1 σR 2 , which shows that for the unbiased case
MG GR MG GR
m̄matter < m̄matter , because σR > σR and B1 is negative. However, if the differences in
linear local bias are sufficiently large they yield m̄MG GR
tracers > m̄tracers .

6 Comparing to simulations

The simulation suite used in this paper is the Extended LEnsing PHysics using ANalaytic
ray Tracing (elephant) simulations [58] run with the ECOSMOG code [59]. This suite contains
five realisations of the initial conditions and for each realisation we have one simulation of
ΛCDM (GR) together with three simulations of the Hu-Sawicki f (R) model with parameters
fR0 = −10−4 (F4), fR0 = −10−5 (F5) and fR0 = −10−6 (F6). It also contains galaxy mock
catalogs that were made with a Halo Occupation Distribution method. The HOD parameters
for the ΛCDM model are the best-fit parameter values from the CMASS data [60]. For the
f (R) models, we tune the HOD parameters so that we reproduce the correlation function
in the ΛCDM model. The simulations were run in a box of size L = 1024 Mpc/h with
N = 10243 particles and the cosmological parameters used to make the initial conditions
were Ωb = 0.046, ΩΛ = 0.719, Ωm = 0.281, h = 0.697, σ8 = 0.82 and ns = 0.971, and for our
analysis we choose snapshots at redshift z = 0.5
To compute the mark for each of our tracers (dark matter particles, halos or mock
galaxies) we binned the particles/halos/mock galaxies to a grid with gridsize of 20 Mpc/h

– 15 –
 1.00
Matter
0.99

0.98 GR

ℳ(r)
0.97 F6

F5
0.96
F4
0.95

0.94

1.000

0.998
ℳ/ℳZA,GR

0.996

0.994

0.992

0.990
20 40 60 80 100
r [h-1 Mpc]

Figure 6. The top panel shows the marked correlation functions for matter particles in GR and F4
HS gravity and the bottom panel F4, F5, F6, and GR ratios to the ZA in GR.

(corresponding to a N = 523 grid) using a Nearest Grid Point (NGP) assignment scheme to
get an estimate for the density for which the mark depends on.
The correlation functions were computed using the Correlation Utilities and Two-point
Estimates (CUTE) code5 [61]. We computed both standard (ξ(r)) and weighted (W (r)) two-
point correlation functions, the latter using the mark as a weight in CUTE. From this the
marked correlation function follows simply as M(r) = 1+W (r)
1+ξ(r) .
We consider the White-mark with ρ∗ = 10, p = 7, corresponding to coefficients C0 = 1,
C1 = −0.64, C2 = 0.46 in the Taylor expansion of the mark function in eq. (3.1). In our
analytical models we smooth the matter fields that assign the mark with a top-hat filter WR
of radius R = 10 Mpc/h.

6.1 Matter particles

In this subsection we confront different analytical methods against matter data, since in
this case the mCF does not suffer from biasing-marking degeneracies and we can observe
more neatly the effects of the weights. The expansion parameters B are obtained from
eq. (3.7): B1GR = −0.6495, B1F6 = −0.6497, B1F5 = −0.6505, B1F4 = −0.6522, B2GR = 0.4628,
B2F6 = 0.4502, B2F5 = 0.4495, B2F4 = 0.4480; which are numerically very close to C1 and C2
2
values because C0 = 1 and the variances σRR are small.
In the top panel of figure 6 we show the analytical mCFs for GR and F4 obtained
with the perturbative method presented in section 4.3 [eqs. (4.27, 4.31)], showing that they
follow reasonably well the trends of the matter mCF extracted from simulations. The bars
correspond to the root-mean-squared (RMS) error of the five simulation boxes available for
each model. In the bottom panel, we show the ratios of the mCF for each model to the
Zel’dovich (ZA) model in GR – by ZA we mean the original model in ref. [4], that is, the
5
The code can be found at https://github.com/damonge/CUTE

– 16 –
 1.002 GR F6

1.000

0.998
ℳ(r)/ℳZA (r)

0.996

W16 + 1-loop
0.994
This Work

0.992

0.990

0.988

1.002 F5 F4

1.000

0.998
ℳ(r)/ℳZA (r)

0.996

ZA (C params)
0.994
Eulerian

0.992

0.990

0.988
20 40 60 80 20 40 60 80
r [Mpc/h] r [Mpc/h]

Figure 7. Ratios of different mCFs analytical models to the ZA mCF for GR, F6, F5 and F4.
We show the W16 model plus 1-loop corrections of eq. (4.22) (dashed blue); the model of this work
[eqs. (4.27,4.31)] (solid black); the linear Eulerian model of eq. (3.11) (long-dashed brown); and the
ZA but using the Taylor expansion parameters of the mark function C in eq. (3.1) (dot-dashed green),
instead of the resummed expansion parameters B.

linear piece of eq. (4.22). We notice that the data for different gravitational theories are very
close to each other, and the RMS errors, although small, are large enough such that one
cannot distinguish between the models. Hence, PT may not help to differentiate between
gravitational theories with matter mCFs. One may try other values for parameters ρ∗ and
p in the White-mark, but as was shown in [25], the differences of mCFs between different
models remain almost the same. In the upcoming subsections we will see that this situation
changes drastically by using tracers instead of matter. Nevertheless, the matter mCF data
still enables us to compare the different perturbative methods, which is the main objective
of this section.
In figure 7 we show the ratios of the different analytical methods to the ZA model.
On each panel the dashed blue curves show the results with the W16 model plus 1-loop
corrections of eq. (4.22); solid black, the method presented in this work [eqs. (4.27,4.31)]
(“This Work”); long-dashed brown, the linear Eulerian model of eq. (3.11); and dot-dashed
green, the ZA model, but instead of using the resummed B expansion parameters we use the
Taylor coefficients C. The differences between the analytical methods are apparent but small,
being lesser than the 1%. At scales r > 40 Mpc/h, all analytic models are indistinguishable

– 17 –
 300 500

Log10M > 12.5 Log10M > 13.5

250
bF4
LS =1.52 400
bF4
LS =2.64
[(Mpc/h)2 ]

[(Mpc/h)2 ]
200

300
bF5
LS =1.66
bF5
LS =2.64

150
r 2ξ(r ) + shift

r 2ξ(r ) + shift
200

bF6 bF6
LS =2.78
100 LS =1.66

100
50 bGR
LS =2.80
bGR
LS =1.72

0 0

20 40 60 80 100 120 20 40 60 80 100 120

r [Mpc/h] r [Mpc/h]

Figure 8. Halo correlation functions ξh (r). From bottom to top, we show GR, F6, F5, and F4
models. Left panel is for halo masses 12.5 < log10 [M/(M h−1 )] < 15, and right panel for 13.5 <
log10 [M/(M h−1 )] < 15. The horizontal dashed gray lines denote the values ξh (r) = 0 for each
model, which have been shifted for visualization purposes. The analytical curves are given by ξhModel =
(1 + bModel
1 )2 ξm
Model
, where the matter correlation functions are computed with the ZA. Notice that
the Eulerian linear bias (or Large-Scales bias) is bLS = 1 + b1 .

and they are within the errors of the simulation data. At smaller scales, the method “This
Work” outperforms the other perturbative approaches and captures pretty well the trend of
the data all the way up to the smoothing scale R = 10 Mpc/h, but it is not accurate enough
to lie inside the RMS errors of the data.

6.2 Halos and mock galaxies

Halos are identified from the elephant suite of simulations using the publicly available
code ROCKSTAR [62]6 that uses a hierarchical refinement of friends-of-friends groups in phase-
space. The particles unbinding procedure used by ROCKSTAR is gravitational model dependent,
however in refs. [41, 63] it is shown that for f (R) theories one can neglect MG effects and
use the standard approaches implemented in halo finders.
We test our perturbative method against the halos mCF. We notice that the mark
function is assigned by taking as argument the number density of halos, and not the smoothed
dark matter density as in eq. (2.3). That is, we consider m = m(δR h ) with

nh (x0 ) − n̄h
Z
1
h
δR (x) = d3 x0 , (6.1)
VΩx Ωx n̄h

where the region of integration Ωx corresponds to the the cubic cells of volume VΩx used to
assign the mark at the position x. nh (x) is the number density of halos with masses in a given
interval and n̄h the average number density of such halos. The effect of this reassignment
is to redefine the mark function, and hence the expansion parameters B. In such a case,
we cannot use the values of the B parameters obtained in the previous subsection, and we
6
https://bitbucket.org/gfcstanford/rockstar

– 18 –
 1.00 Log10 M > 12.5 1.00 Log10 M > 13.5

0.98
GR 0.95 GR
0.96
F6 F6
ℳ(r)

ℳ(r)
0.94
F5 0.90 F5

0.92 F4 F4

0.85
0.90

1.010 1.010
ℳ/ℳGR

ℳ/ℳGR
1.005 1.005

1.000 1.000

0.995 0.995
30 60 110 30 60 110
r [h-1 Mpc] r [h-1 Mpc]

Figure 9. Halo marked correlation functions. Left panel is for halo masses 12.5 <
log10 [M/(M h−1 )] < 15, and right panel for 13.5 < log10 [M/(M h−1 )] < 15.

400
Log10M > 12.5
Log10M > 13.5
60
F4

50 F4 300
[(Mpc/h)2 ]

[(Mpc/h)2 ]

40 F5

F5 200
r 2W (r ) + shift

r 2W (r ) + shift

30

F6
20 F6
100

10
GR
GR
0 0

20 40 60 80 100 120 20 40 60 80 100 120

r [Mpc/h] r [Mpc/h]

Figure 10. Halo weighted correlation functions Wh (r). From bottom to top we show GR, F6, F5,
and F4 models. Left panel is for halo masses in the interval 12.5 < log10 [M/(M h−1 )] < 15, and
right panel for 13.5 < log10 [M/(M h−1 )] < 15. The horizontal dashed gray lines denote the zeros for
each model, which have been shifted for visualization purposes. The analytical curves are given by
the ZA with the bias parameters, b, and the expansion parameters, B, those obtained from the fitting
to the halo correlation functions and to the marked correlation functions, respectively, as it was done
in figs. 8 and 9. Dotted curves correspond to the “This Work” method and the solid curves to the
W16 + 1-loop.

have to treat them as free parameters that should be fitted from the simulations. To do
this, we first fit the large-scale, Eulerian bias bLS = 1 + b1 to the correlation function using

– 19 –
the ZA, such that ξh (r) = b2LS ξZA (r). We use two interval of masses: log10 M > 12.5 and
log10 M > 13.5, with M the halo mass in units of M h−1 . In section 5 we discussed that the
effect of biasing on the mCF is to bringing it down, such that as more different the biases
are, the more the mCF differs. Because of this, we choose a large halo mass interval (> 13.5)
that we expect to give larger different biases for the different gravitational theories. However,
we have a relatively small number of halos in this interval, and hence large statistical errors
in the data. Therefore we also use an interval with a moderate lower mass (> 12.5). A
histogram with the number of halos over mass intervals in the elephant suite of simulations
is presented in figure 2 of ref. [26].
The fittings to the halo correlation functions ξh (r) are shown in figure 8, providing the
large scale values which are reported in the left and right panels of that figure, corresponding
to the two mass intervals considered. As expected, the linear local biases are larger for more
massive halos, while the error bars also increase because we have many fewer halos in the
interval log10 M > 13.5 than in the interval log10 M > 12.5. Also, we confirm numerically
that bMG < bGR , as explained in section 5. However, we do not obtain larger differences in
the bias of MG and GR for more massive halos as predicted in the peak-background split
formalism. This may be an indication of a breakdown of bias models that assume conservation
of tracers [33, 35, 64]: for example, it is known that in F4 fewer small mass halos survive
because of a faster merging rate than in GR [26]. We also note that the number of available
halos in the interval log10 M > 13.5 is quite limited due to the limited size of the simulations
so the scatter in the correlation function is large.
With the estimation of the biases at hand, we fit the mark function expansion parameters
B directly to the mCF, obtaining B1GR = −1.20, B1F6 = B1F5 = −1.22, B1F4 = −1.25, for
the interval log10 M > 12.5; and B1GR = −1.43, B1F6 = B1F5 = −1.45, B1F4 = −1.55, for
log10 M > 13.5. We plot these results in figure 9, where the top panels show the mCF and
the bottom panels the ratio of the mCF of the different gravitational theories to the GR
case. Our fittings are very good for intermediate scales 30 < r < 90 Mpc/h. At larger scales,
particularly at the BAO peak, we have a considerable discrepancy. There are two sources
of errors, the first one is due to that the used simulations underestimate the BAO peak in
the correlation function, see figure 8 or refs. [35, 36]; and the second because our analytical
model fails at the BAO scale since our method mixes Euleran and Lagrangian correlation
functions (see discussion below). We have checked that varying B2 and b2 over reasonable
intervals has little impact on the results so we let them fixed to B2 = 0.5 and b2 = 0.2. Also,
the computation of the marked correlation function using the method W16+1-loop gives as
accurate results as those shown in figure 9, but using slightly different parameters B values.
To check consistency in our fittings, in figure 10 we show the weighted correlation func-
tion Wh (r) for each model and for each of the halo mass intervals. Dotted curves correspond
to the “This Work” method and the solid curves to the W16+1-loop. Both methods perform
poorly at scales r < 30 Mpc/h, while in the interval 30 < r < 80 Mpc/h the “This Work”
method works slightly better. However, at larger scales the method W16+1-loop is superior
and it draws correctly the BAO peak. This can be understood analytically by taking the
large scale limit for each model,7

W This Work (r) ' b2LS ξZA (r) + B12 ξRR (r) + 2bLS B1 ξRσ2 (r), (6.2)
Ψ
W16 2
W (r) ' (bLS + B1 ) ξZA (r), (6.3)
7
See section 3.4 of ref. [29] for the large scale limit of the CLPT correlation function.

– 20 –
 1.00
HODs

0.95

GR

ℳ(r)
F4
0.90

0.85

1.006

1.004
ℳ/ℳGR

1.002

1.000

0.998

0.996
20 40 60 80 100
r [h-1 Mpc]

Figure 11. mCFs of HOD galaxy mocks for F4 and GR models. The linear local biases are fitted by
using the ZA correlation function and the data from the simulations, giving bLS
1 = 2.1 for all models.
The B parameters are B1GR = −1.10 and B1F4 = −1.05.

where in the second equality we have neglected the differences between ξZA and its smoothed
versions (8th and 12th terms on the right hand side of eq. (4.18)) which are smaller than
the differences between the ZA and linear Eulerian correlation functions. Hence, W16 is able
to capture well the BAO peak because from the very beginning it neglects the differences
between Eulerian and Lagrangian coordinates, while our method fails because the contribu-
tion of ξRσ2 (being negative for B1 < 0) competes with the ξRR (r) and ξZA (r) correlation
Ψ
functions.
Finally, we consider the mCF for mock galaxies. For the modified gravity models
the HOD parameters were fitted as to give rise to the same observed two-point correlation
function as in ΛCDM. To test our model to data we proceed exactly as we did for halos above.
We first fit the linear local bias, obtaining bLS
1 = 2.1 for all models. This is expected of course,
because by construction the correlation functions of galaxies should be equal and the matter
correlation functions are indistinguishable at large scales. Thereafter, we fit directly to the
mCF data. The results are shown in figure 11, where we only plot the GR and F4 models;
the corresponding mCFs for F5 and F6 lie somewhere in between them. Given that the
correlation function is tuned to be the same in different models, the difference of the mCF
is much smaller compared with that for dark matter halos. This implies that the mCF with
this choice of the mark is not sufficient to detect the difference between these models. To
enhance the difference between the models, we need to consider different marks from White’s
[25–27].

7 Conclusions

In this work we have studied marked correlation functions that up-weights low density regions
in the Universe [4]. The idea behind marked statistics is to assign a value (the mark) to each

– 21 –
entity in a point process and perform statistics over the resulting weighted objects. The
mark can be anything that can be quantified as an astronomical property to be conveniently
contrasted such as luminosity, color, morphology, etc.. A way to suppress nonlinearities in
statistics is to choose a mark that gives more weight to objects that reside in low density
regions, and that smoothly fades away as one moves to higher and higher density regions.
This marking technique may be particularly important for testing general relativity with
cosmological probes, since modified gravity models, tailored to provide cosmic acceleration,
often rely on screening mechanisms to hide their impact on high density environments, and
the screening switches off in regions that are more depleted of matter. Therefore, the effects
of MG are more pronounced in low density regions. It is then natural to search for such
marked statistics, as proposed by Martin White in ref. [4].
Throughout the paper, we have compared our results with those of ref. [4], clarifying
some of its issues and generalizing its perturbative model (W16) to include higher than linear
order fluctuations. We have used renormalized bias parameters and a complete resummation
of the mark function’s Taylor coefficients in order to write the two-point statistics with no
zero-lag correlators. First, we have found an expression for the mCF in linear Eulerian the-
ory, which is the natural frame to assign marks. However, given that the correlation function
is well known to be better described with LPT, we have moved to Lagrangian space, with
initially Lagrangian biased tracers, and compute the mCF in the CLPT resummation scheme.
This is given by a six dimensional double convolution integral, eq. (4.14). We did not proceed
to directly integrate this complicated integral, but derived two different approximations from
it: the one obtained in ref. [4] plus its loop corrections (W16+loops), and a second that we
called throughout as the “This Work” model. Both methods have advantages and disad-
vantages that we summarize in the following. In the Lagrangian formulations, one evolves
initial, yet linear, fields and assign them the mark at the moment of observation with Eulerian
environmental densities. This mix of Eulerian-Lagrangian schemes is difficult to overcome
unless one is able to integrate the above-mentioned double convolution. The method W16
from the beginning approximates the Eulerian coordinates to their corresponding Lagrangian
coordinates, reducing the expression of eq. (4.14) to a single convolution that can be numeri-
cally treated with the known methods of CLPT. Meanwhile, the “This Work” method splits
the integral into three pieces, one containing only Lagrangian coordinates, one Eulerian, and
the third a mix of both. The first two pieces are integrated exactly, but the third, which
contains both Lagrangian and Eulerian coordinates, uses the expansion of Lagrangian dis-
placements of eq. (4.25). When comparing both methods, together with the linear Eulerian
and the ZA (the linear contributions to W16; that is, the original model of ref. [4]), we show
that all models give quite similar results, with differences below 1% for scales > 20 Mpc/h
and below 0.1% for > 50 Mpc/h (see figure 2); for comparison, the difference between the
ZA and 1-loop CLPT correlation functions is about the 1% at scales above r = 20 Mpc/h.
This shows the highly linear nature of mCFs that up-weights low density regions of space.
At smaller scales, however, the results are unreliable because our perturbative methods are
limited by the smoothing scale of the matter (or tracers) densities used to assign the mark,
which should be large.
We discussed the effects of bias on the mCF, showing that they were quite degenerated
with the mark itself. As the mark parameter B1 is more negative, or as larger the value of the
linear local bias is, the more the mCF reduces its amplitude. However, as known from earlier
works, MG models generally have smaller biases due to a more efficient clustering. Then,
while for matter particles the MG mCFs are smaller than that of GR, for biased tracers the

– 22 –
opposite may happen.
We tested our analytical methods against the elephant suite of MG N-body simula-
tions. First, we showed results for dark matter. Since this case is free of bias, we can observe
the effects of marking dark matter particles more clearly and it allows us to compare the
performance of the different perturbative models. The four PT methods we tested show dif-
ferent behaviors, but all lie within the 1% accuracy. However, at smaller scales, < 40 Mpc/h,
the method “This Work” outperforms the other perturbative approaches and captures pretty
well the trend of the data all the way down to the smoothing scale R = 10 Mpc/h, but it is
not accurate enough to lie inside the error bars of the simulation data. Finally, we compared
our PT prediction for biased tracers, dark matter halos and HOD galaxy catalogs, finding
good agreement between theory and simulations for scales r > 30 Mpc/h. However, we no-
tice that at large scales, around the BAO position, the W16 method performs better than
“This Work” because of the mixing of Eulerian and ZA correlation functions in our proposed
method.
In this work we formally developed the perturbative theory of marks using dark matter
and tracers for general gravity models, and applied to ΛCDM and modified gravity models
to enhance differences among models that eventually could serve to discriminate them with
the use of simulations and observations.

Acknowledgments

We would like to thank Martin White and Georgios Valogiannis for useful discussions and
suggestions. A.A. and J.L.C.C. acknowledge support by Conacyt project 283151. K.K. is sup-
ported by the European Research Council through 646702 (CosTesGrav) and the UK Science
and Technologies Facilities Council grants ST/N000668/1 and ST/S000550/1. B.L. acknowl-
edges supports by the European Research Council via an ERC Starting Grant (ERC-StG-
716532-PUNCA) and the UK Science and Technology Facilities Council (STFC) via Consol-
idated Grant No. ST/L00075X/1. The simulations described in this work used the DiRAC
Data Centric system at Durham University, operated by the Institute for Computational Cos-
mology on behalf of the STFC DiRAC HPC Facility (www.dirac.ac.uk). This equipment
was funded by BIS National E-infrastructure capital grant ST/K00042X/1, STFC capital
grants ST/H008519/1, ST/K00087X/1, STFC DiRAC Operations grant ST/K003267/1 and
Durham University. DiRAC is part of the National E-Infrastructure.

A Brief review of LPT for MG

The LPT for MG was constructed in ref. [31], in this appendix we give a brief review of that
work. In LPT one considers overdensities δ(q) at an initial time tini sufficiently early such that
δ is still linear at all scales of interest of the problem at hand. q denote the space coordinates
at this initial time and the Lagrangian displacement vector Ψ relates them to Eulerian
coordinates defined at later times: x(q, t) = q+Ψ(q, t), such that x(q, tini P) = q. (n)
Perturbation
theory expands the Lagrangian displacement vector fields as Ψi (k) = ∞ n=0 Ψ , and each
term is written in a Fourier space Taylor series in matter linear overdensities as
n
!
d3 ki
Z
(n) i Y (n)
Ψ = (2π)3 δD (k − k1 − · · · − kn )Li (k1 , ..., kn )δL (k1 ) · · · δL (kn ), (A.1)
n! (2π)3
i=1

– 23 –
with L(n) the Lagrangian kernels at order n. In MG new scales are introduced, such that the
linear growth function D+ (k, t) is the fastest growing solution to the linearized fluid equation
(T̂ − A(k))D+ (k, t) = 0, (A.2)
d2 d
where T̂ ≡ dt2
+ 2H dt [65], and we introduced
2β 2 k 2
 
A(k, t) = A0 1 + 2 , A0 = 4πGρ̄, (A.3)
k + m2 a2
with ρ̄ the background matter density. In general β and m are time and scale dependent and
can be considered as parametrizations for unknown MG theories [66, 67]. Or, otherwise, they
can be obtained directly from a specific gravitational model. In scalar tensor theories, β gives
the strength of the fifth-force and 1/m its range. In models with a chameleon screening, the
mass m depends on the environmental density becoming large in high density regions.
Since
Lagrangian displacements and matter fluctuations are related by δ(x) = 1 − J(q) J −1 (q),
with Jij = δij + Ψi,j , at first order we obtain the Zeldovich solution
(1) ki
Ψi (k, t) = i D+ (k, t)δL (k, t0 ), (A.4)
k2
(1)
so we can read the first order Lagrangian kernel, Li (k) = ki /k 2 . To second order,
(k1 · k2 )2
 
(2) 3 ki
Li (k1 , k2 ) = A(k1 , k2 ) − B(k1 , k2 ) , (A.5)
7 k2 k12 k22
with k = k1 + k2 , and
(2) (2)
7DA (k1 , k2 ) 7DB (k1 , k2 )
A(k1 , k2 ) = , B(k1 , k2 ) = , (A.6)
3D+ (k1 )D+ (k2 ) 3D+ (k1 )D+ (k2 )
and the second order growth functions, D(2) , are the solutions to equations [31–33]
"
(2) k1 · k 2 k1 · k 2
(T̂ − A(k))DA = A(k) + (A(k) − A(k1 )) 2 + (A(k) − A(k2 ))
k2 k12
#
2A0 2 k 2
 
M2 (k1 k2 )
− 2
D+ (k1 )D+ (k2 ), (A.7)
3 a 6Π(k)Π(k1 )Π(k2 )
h i
(2)
(T̂ − A(k))DB = A(k1 ) + A(k2 ) − A(k) D+ (k1 )D+ (k2 ), (A.8)

with appropriate initial conditions. As it is common, we have used Π(k) ≡ (k 2 +m2 a2 )/6β 2 a2 .
The function M2 in eq. (A.7) is the first coefficient in a Fourier space Taylor expansion of
the non-linear piece of the potential in the Klein-Gordon equation of the scalar field that
mediates the fifth-force [30], so it is responsible for the screening mechanism that drives the
theory to GR in high density regions. Expressions for the third order growth are large and
not displayed here; see ref. [31].
We stress out that the LPT formalism can be used to obtain statistics in Fourier space as
the 2-point correlation function, but also to obtain the power spectrum in SPT, for example,
the kernel F2 is given by [33]
(k1 · k2 )2 k1 · k2 1
   
1 3 1 3 1
F2 (k1 , k2 ) = + A + − B + + , (A.9)
2 14 2 14 k21 k22 2 k12 k22
which reduces to the well-known kernel in EdS, since in that case A = B = 1.

– 24 –
A.1 Hu-Sawicki model

The Hu-Sawicki model [39] is a particular realisation of f (R) gravity that is able to evade the
strong constraints coming from local test of gravity and still give rise to interesting observable
signatures on cosmological scales. We are here considering the case where the index n = 1
so the model has only one free parameter fR0 . Taking this parameter to zero we recover
GR. The three choices for the parameters we are considering in this paper (F4, F5 and F6)
are such that they lie in the region around where the best constraints lie today. The F4
model corresponds to |fR0 | = 10−4 (is in tension with local experiments), the F5 model to
|fR0 | = 10−5 (agrees with most experiments and observations, but are in tension with others)
and the F6 model to |fR0 | = 10−6 (which is still allowed).
this model the functions describing the first order LPT are given by β 2 = 1/6 and
In p
m(a) = M1 (a)/3 where

3 H02 (Ωm0 a−3 + 4ΩΛ )3

M1 (a) = . (A.10)
2 |fR0 | (Ωm0 + 4ΩΛ )2

The function M2 (a) that enters at second order in LPT is given by

9 H02 (Ωm0 a−3 + 4ΩΛ )5

M2 (a) = . (A.11)
4 |fR0 |2 (Ωm0 + 4ΩΛ )4

For a complete description of the model and f (R) gravity in general see refs. [22, 23, 39] and
for more information about the LPT equations in this model see ref. [31].

B Constructing the marked correlation function

Starting from eq. (4.10) we use the standard methods of CLPT [29] and the cumulant ex-
pansion theorem to get biases from spectral parameters with the aid of eqs. (3.6) and (4.3).
We obtain eq. (4.11) with 1 + I = 1 + I 0 + I|zero-lag , where

1 i
1 + I 0 = 1 + k1i k2i Aloop
ij − W̄ + b21 ξ(|q2 − q1 |) + 2b1 B1 ξR (|x2 − q1 |) + B12 ξRR (|x2 − x1 |)
2 6
− ib1 (k1i − k2i )Ui (q2 − q1 ) − 2iB1 k1i UiR (x2 − q1 ) − ib1 b2 (k1i − k2i )ξ(|q2 − q1 |)Ui (q2 − q1 )
− 2ib21 B1 k1i ξ(|q2 − q1 |)UiR (x2 − q1 ) − i(b21 + b2 )B1 (k1i − k2i )ξR (|x2 − q1 |)Ui (q2 − q1 )
− 2iB1 B2 k1i ξR (|x2 − q1 |)UiR (x2 − q1 )
− iB12 b1 (k1i − k2i )ξRR (|x2 − x1 |)Ui (q2 − q1 ) − 2iB1 B2 k1i ξRR (|x2 − x1 |)UiR (x2 − q1 )
1
− b2 (k1i k1j + k2i k2j ) + b21 (k1i k2j + k2i k1j ) Ui (q2 − q1 )Uj (q2 − q1 )

+
2
− 2B1 b1 (k1i − k2i )k1j Ui (q2 − q1 )UjR (x2 − q1 ) − (B12 + B2 )k1i k1j UiR (x2 − q1 )UjR (x2 − q1 )
+ ib2 Z 2000 + iB2 Z 0020 + ib21 Z 1100 + iB12 Z 0011 + 2ib1 B1 Z 1001 + 2ib1 B1 Z 1010
− b1 Ā1000 − B1 Ā0010 , (B.1)

– 25 –
which is valid to 1-loop and up to third order in bias expansion (counting b1 and B1 as linear
and b2 and B2 as second order). We have defined

Z pqrs = hδ p (q1 )δ q (q2 )δR

r s
(x1 )δR (x2 )(−k1i Ψi (q1 ) − k2i Ψi (q2 ))ic , (B.2)
Ā pqrs p
= hδ (q1 )δ q r
(q2 )δR s
(x1 )δR (x2 )(−k1i Ψi (q1 ) − k2i Ψi (q2 ))(−k1j Ψj (q1 ) − k2j Ψj (q2 ))ic ,
(B.3)
W̄ = h(−k1i Ψi (q1 ) − k2i Ψi (q2 ))(−k1j Ψj (q1 ) − k2j Ψj (q2 ))(−k1k Ψk (q1 ) − k2k Ψk (q2 ))ic ,
(B.4)

and as usual [29]

d3 p p·(qa −qb ) pi
Z
Ui (qa − qb ) ≡ hδ(qa )Ψi (qb )i = −i e PL (p), (B.5)
(2π)3 p2
d3 p p·(xa −qb ) pi
Z
R
Ui (xa − qb ) ≡ hδR (xa )Ψi (qb )i = −i e PL (p)W̃ (p). (B.6)
(2π)3 p2

In deriving eq. (B.1) we have used repeated times that the integral in eq. (4.10) is invariant
under the interchange (k1 , x1 , q1 ) ←→ (k2 , x2 , q2 ). Besides I0 , there are also terms depending
on x1 −q1 (or x2 −q2 ) that lead to zero-lag correlators σRσ 2 after integration, and ultimately
Ψ
−2
cancel out with the factor m̄ in the mCF, as it will become more clear in subsection B.2,
hence we omit to explicitly write all of them. To linear order in PL , these are

I|zero-lag = 2b1 B1 ξR (|x1 − q1 |) − 2iB1 k1i UiR (x1 − q1 ) + Non-linear. (B.7)

B.1 The CLPT marked correlation function

Using the variables transformation (q1 , q2 , x1 , x2 , k1 , k2 ) −→ (q, Q, r, R, ka , kb ), given in
eqs. (4.13), into eq. (4.11), we analytically perform two Gaussian integrations, one for d3 ka
and the other for d3 kb , to obtain eq. (4.14) with
1 1
1 + I 0 = 1 − Aloop 2 2
ij Gij − Γijk Wijk + b1 ξ(q) + 2b1 B1 ξR (z) + B1 ξRR (r) − 2b1 gi Ui (q)
2 6
− 2B1 gi UiR (z) + B1 giC UiR (z) − 2b1 b2 ξ(q)Ui (q)gi + b21 B1 ξ(q)UiR (z)giC − 2b21 B1 ξ(q)UiR (z)gi
− 2(b21 + b2 )B1 ξR (z)Ui (q)gi − 2B1 B2 ξR (z)UiR (z)gi + B1 B2 ξR (z)UiR (z)giC
− 2B12 b1 ξRR (r)Ui (q)gi − 2B1 B2 ξRR (r)UiR (z)gi + B1 B2 ξRR (|x2 − x1 |)UiR (z)giC
1
− (b21 + b2 )Ui (q)Uj (q)Gij − 4B1 b1 Ui (q)UjR (z)Gij − B1 b1 Ui (q)UjR (z)giC gi
2
1
− (B1 + B2 )Ui (z)Uj (z)Gij − (B1 + B2 )(−gi gj + GC
2 R R 2 C C
)U R (z)UjR (z)
4 ij i
+ b2 gi Ui ,2000 + B2 gi Ui ,0020 + b21 gi Ui ,1100 + B12 gi Ui ,0011 + 2b1 B1 gi Ui ,1001
1 1
+ 2b1 B1 gi Ui ,1010 − B2 giC Ui⊕,0020 − B12 giC Ui⊕,0011 − b1 B1 giC Ui⊕,1001 − b1 B1 giC Ui⊕,1010
2 2
1 ⊕⊕,0010
+ b1 Gij A1000
ij + B1 Gij Aij ,0010 − B1 GC ij Aij − B1 gi gjC A⊕
ij
,0010
(B.8)
4
where z = x2 − q1 = R − Q + 12 r + 12 q and
−1 −1 −1 C
giC = Cij (Qj − Rj ), GC C C
ij = Cij − gi gj , ΓC C C C
ijk = C{ij gk} − gi gj gk , (B.9)

– 26 –
where indices in between brackets {· · · } are cyclically summed, and

∆i ≡ Ψi (q2 ) − Ψi (q1 ) = ∆i , ∆⊕
i ≡ Ψi (q2 ) + Ψi (q1 ), (B.10)

Ui⊕,pqrs = hδ p (q1 )δ q (q2 )δR

r s
(x1 )δR (x2 )∆⊕
i ic (B.11)
,pqrs
Ui = hδ p (q1 )δ q (q2 )δR
r s
(x1 )δR (x2 )∆i ic = Uipqrs (B.12)
A⊕⊕,pqrs
ij = p q
hδ (q1 )δ (q2 )δR r s
(x1 )δR (x2 )∆⊕ ⊕
i ∆j ic (B.13)
A⊕
ij
,pqrs
= hδ p (q1 )δ q (q2 )δR
r s
(x1 )δR (x2 )∆⊕
i ∆j ic (B.14)
Aij ,pqrs = hδ p (q1 )δ q (q2 )δR
r s
(x1 )δR (x2 )∆i ∆j ic = Apqrs
ij (B.15)
Wijk = h∆i ∆j ∆k ic (B.16)

Note that if the functions ξ, U , A, W , in eq. (B.8) depend only on q and r, we obtain the
standard CLPT approach with only one convolution integral, this is because the following
integrals hold
1 T −1 1 T −1
d3 Qe− 2 (R−Q) C (R−Q) d3 Qe− 2 (R−Q) C (R−Q) C
Z Z
= 1, gi = 0, (B.17)
(2π)3/2 |C|1/2 (2π)3/2 |C|1/2
1 T −1 1 T −1
d3 Qe− 2 (R−Q) C (R−Q) C d3 Qe− 2 (R−Q) C (R−Q) C
Z Z
Gij = 0, Γijk , = 0, (B.18)
(2π)3/2 |C|1/2 (2π)3/2 |C|1/2

which are a consequence that the matrix Cij is a function of q only. Indeed we have already
used the above integrals to derive eq. (B.8), where we omitted to write terms, like for example
⊕ ⊕ ⊕
ΓC
ijk h∆i ∆j ∆k ic , that vanish when integrated over Q. Moreover, if we evaluate eq. (B.8) in
R = Q, we have giC = GC C
ij = Γijk = 0, this is what we do in section 4.2, where we approximate

1 T −1
e− 2 (R−Q) C (R−Q)
≈ δD (R − Q), (B.19)
(2π)3/2 |C|1/2

to obtain eq. (4.16) with

1 1
1 + J 0 = 1 − Aloop 2 2
ij Gij − Γijk Wijk + b1 ξ(q) + 2b1 B1 ξR (y) + B1 ξRR (r)
2 6
− 2b1 gi Ui (q) − 2b2 gi UiR (y) − 2b1 b2 ξ(q)Ui (q)gi − 2b21 B1 ξ(q)UiR (y)gi
− 2(b21 + b2 )B1 ξR (y)Ui (q)gi − 2B1 B2 ξR (y)UiR (y)gi
− 2B12 b1 ξRR (r)Ui (q)gi − 2B1 B2 ξRR (r)UiR (y)gi − (b21 + b2 )Ui (q)Uj (q)Gij
− 4B1 b1 Ui (q)UjR (y)Gij − (B12 + B2 )UiR (y)UjR (y)Gij
+ b1 Gij Aij,1000 + B1 Gij Aij,0010 + b2 gi Ui ,2000
+ B2 gi Ui ,0020
+ b21 gi Ui ,1100

,0011 ,1001 ,1010

+ B12 gi Ui + 2b1 B1 gi Ui + 2b1 B1 gi Ui (B.20)

with y = 12 r + 21 q. If we further substitute y → q, in the arguments of functions ξ(y), U (y),

A(y) we obtain eq. (4.17).

– 27 –
B.2 Approximating the CLPT marked correlation function

Keeping for the moment only linear terms in the function Ir,q in eq. (4.26) and using eqs. (B.1)
and (B.7) we write at leading order

d3 k1 d3 k2 3 3 ik1 ·(x1 −q1 ) ik2 ·(x2 −q2 ) − 1 (k1 +k2 )2 σ2

Z
hG1 Fx,1 G2 Fx,2 i r,q 3 (B0∗ )2 d q1 d q2 e e e 2 Ψ
(2π)3 (2π)3
h
× 2b1 B1 ξR (|x1 − q1 |) − i2B1 k1i hδR (x1 )Ψi (q1 )i + 2b1 B1 ξR (|x2 − q1 |)
i 1 
− i2B1 k1i hδR (x2 )Ψi (q1 )i + · · · 1 + k1i k2j ALij (q) . (B.21)
2

The first two terms within the brackets are integrated to give 2(B0∗ )2 (1 + b1 )B1 σRσ
2
Ψ
+ ···.
Such terms that have as arguments x1 − q1 are canceled out by the squared of the mean
mark in the mCF. Let us consider, as an example, the last term in eq. (B.21), which is the
more cumbersome

d3 k1 d3 k2 3 3 ik1 ·(x1 −q1 ) ik2 ·(x2 −q2 ) − 1 (k1 +k2 )2 σ2

Z
I8 (r) ≡ d q1 d q2 e e e 2 Ψ
(2π)3 (2π)3
h 1 i
× − ik1` hδR (x2 )Ψ` (q1 )i k1i k2j AL ij (q)
2
Z 3 3
d p1 d p2 −ip1 ·r ip2 ·r − 1 p21 σ2 p` (p1 − p2 )` (p1 − p2 )i pi2
=− e e e 2 Ψ W̃ (k )P (k )P (p ) 1 ,
1 L 1 L 2
(2π)6 p21 p22
(B.22)

where the second equality is obtained after several manipulations and the use of eqs. (4.12)
and (B.6). Using

p`1 (p1 − p2 )` (p1 − p2 )i pi2 p2 p1

2 2 = δij p̂i1 p̂j2 + δij p̂i1 p̂j2 − δij δmn p̂i1 p̂j2 p̂m m
1 p̂2 − 1, (B.23)
p1 p2 p1 p2

and the solid angle integral identities

−i
Z
dΩp̂ eip·r p̂i = j1 (pr)r̂i , (B.24)
4π
Z
1 j1 (pr)
dΩp̂ eip·r p̂i p̂j = δij − j2 (pr)r̂i r̂j , (B.25)
4π pr

we arrive to

4 2 [2,0] [2,0] [1,1] [1,−1] [1,−1] [1,1]

I8 = ξRσΨ (r)ξL (r) + ξRσΨ (r)ξL (r) − ξRσΨ (r)ξL (r) − ξRσΨ (r)ξL (r), (B.26)
3 3

– 28 –
which is the nonlinear contribution to x̄U R in eq. (4.29). Analogous manipulations, now
including also nonlinear pieces in Ip,q , yield all the terms in eq. (4.27):

d3 k ik·r − 1 k2 σ2 RR
Z
x̄UR UR (r) = e e 2 Ψ Q9 (k), (B.27)
(2π)3
d3 k ik·r  RσΨ2
Z
1 RσΨ 2 
x̄U UR (r) = e Q 9 (k) + Q 12 (k) , (B.28)
2 (2π)3
d3 k ik·r RσΨ2
Z
x̄ξUR (r) = e Q12 (k), (B.29)
(2π)3
d3 k ik·r
Z  
RσΨ 2 1 RσΨ2
x̄ξR U (r) = e Q12 (k) − Q13 (k) , (B.30)
(2π)3 2
Z 3
d k ik·r − 1 k2 σ2 RR
x̄ξR UR (r) = e e 2 Ψ Q12 (k), (B.31)
(2π)3
d3 k ik·r  RσR
Z
6 − 12 k2 σΨ
2 RσR

x̄A0010 (r) = e R 3 (k) + e W̃ R (k)R 2 (k) + Q8 (k) , (B.32)
7 (2π)3
d3 k ik·r − 1 k2 σ2 RR
Z
3
x̄U 0020 (r) = e e 2 Ψ Q8 (k), (B.33)
7 (2π)3
d3 k ik·r RRσΨ
Z
3
x̄U 0011 (r) = e R3 (k), (B.34)
7 (2π)3
d3 k ik·r RσΨ
Z
3
x̄U 1010 (r) = e Q8 (k), (B.35)
7 (2π)3
d3 k ik·r  − 1 k2 σ2 R
Z
3 RσΨ

x̄U 0110 (r) = e e 2 ΨR
1+2 (k) + R3 (k) , (B.36)
7 (2π)3

with Q(k) and R(k) functions

d3 p (p · (k − p))2
Z  
QRR
8 (k) = A−B W̃R (p)PL (p)W̃R (|k − p|)PL (|k − p|), (B.37)
(2π)3 p2 |k − p|2

d3 p (p · (k − p))2
Z  
1 2 2
QRσ
8
Ψ
(k) = A−B 2 e− 2 p σΨ W̃R (p)PL (p)PL (|k − p|), (B.38)
(2π)3 p |k − p|2

d3 p (k · p)2
Z  
1 2 2
R3RσΨ (k) = A−B 2 2 e− 2 p σΨ W̃R (p)PL (p)PL (k), (B.39)
(2π)3 p k

d3 p (k · p)2
Z  
1 2 σ2
R3RRσΨ (k) = A−B e− 2 |k−p| Ψ W̃R (p)PL (p)W̃R (k)PL (k), (B.40)
(2π)3 p2 k 2

d3 p (k · p)2 k · (k − p)
Z  
R
R1+2 (k) = A−B 2 2 PL (k)W̃R (p)PL (p), (B.41)
(2π)3 k p |k − p|2

where in R functions we evaluate A, B(k, −p), while in Q functions A, B(p, k − p).

– 29 –
C k and q functions

R(k) and Q(k) functions are the building blocks for power spectra for both LPT and SPT.
They were introduced in [51] for matter statistics, and extended for tracers in ref. [45].
These are constructed by 2- and 3-point correlations of Lagrangian displacements, and are
well known in the literature; we refer the reader to the above references for their expressions
in ΛCDM, and to refs. [31, 33] for MG. In this section we display those functions that are
not presented in the above references but are necessary for the marked correlation functions.
Defining r = k/p, x = k̂ · p̂ and y = 1 + r2 − 2rx:
∞ 1
k3 (r − x)2
 
√
Z Z
QR
8 (k) = 2 drW̃R (kr)PL (kr) dxr A − B 2
PL (k y), (C.1)
4π0 −1 y
Z ∞ Z 1
k3 (r − x)2
 
RR 2 √ √
Q8 (k) = 2 drW̃R (kr)PL (kr) dxr A − B W̃R (k y)PL (k y), (C.2)
4π 0 −1 y
3 Z ∞ Z 1 2 (−r + x)2
 
R k √ r (1 − rx)
Q5 (k) = 2 drW̃R (kr)PL (kr) dxPL (k y) A−B , (C.3)
4π 0 −1 y y

with A and B functions evaluated as the Q functions of the previous appendix.

Correlations functions in CLPT are constructed from a set of q-functions U (q), A(q)
and W (q); see refs. [29, 53, 54] for details. The expressions of these q-functions in MG are
derived in ref. [33], here we write those q-functions not presented in that work. Splitting the
A(q) functions in irreducible components, A(q) = X(q)δij +Y (q)q̂ i q̂ i , we have for A0010 (q) =
A0001 (q)

"
Z ∞
1 1
X 0010 (q) = X 0001 (q) = 2 dk 2W̃R (k)(RI − R2 ) + 3W̃R (k)RI j0 (kq)
2π 0 14
#
R j1 (kq)
− 3(W̃R (k)(RI + 2R2 ) + 2R1+2 + 2QR 5) , (C.4)
kq
Z ∞  
0010 0001 1 3 R
Y (q) = Y (q) = 2 dk − (W̃R (k)(RI + 2R2 ) + 2R1+2 + 2QR
5)
2π 0 14
 
j1 (kq)
× j0 (kq) − 3 . (C.5)
kq

Similarly, U functions can be written as Ui (q) = q̂i U (q) with

Z ∞  
0020 0002 1 3
U (q) = U (q) = 2 dkk − QRR
8 (k)j1 (kq),
2π 0 7
Z ∞  
1010 0101 1 3
U (q) = U (q) = 2 dkk − QR8 (k)j1 (kq),
2π 0 7
Z  
1 6
U 0011 (q) = 2 dkk − R
W̃R (k)R1+2 (k)j1 (kq),
2π 7
Z  
1001 0110 1 3 R
U (q) = U (q) = 2 dkk − (R1+2 (k) + W̃R (k)R1+2 (k))j1 (kq). (C.6)
2π 7

In deriving these equations we made use of the identities (B.24) and (B.25).

– 30 –
 1.00
Top-Hat
0.95

0.90
R=1 Mpc/h

ℳ(r)
0.85
R=10 Mpc/h
0.80
R=20 Mpc/h
0.75

0.70

-1
10

-2
10
| ℳ/ℳR=1 -1|

-3
10

-4
10

10 20 50 100
r [h-1 Mpc]

Figure 12. Marked correlation function for model F5 using a Top-Hat kernel with different smoothing
radius R = 1, R = 10 and R = 20 Mpc/h. The lower panel shows the relative differences with respect
to the R = 1 Mpc/h case. The dot-dashed gray horizontal line shows the 1% offset.

D Adding curvature and tidal bias

The non-trivial degeneracy between biases and the mark makes difficult to estimate the mCF
from first principles. On the other hand, the function W has a more neat degeneracy between
bias and the mark, indeed one gets

W (b1 , b2 , B1 , B2 ) ≈ ξX (b1 + B1 , b2 + 2B1 b1 + B2 ) (D.1)

where the approximation becomes an equality for W W16 in the limit R → 0. The identifica-
tions b1 → b1 + B1 and b2 → b2 + 2B1 b1 + B2 can be deduced from the definitions of bn and
Bn by replacing λ → λ + Λ. In figure 12 we use the F5 model to plot the mCF for different
smoothing radius R = 1, R = 10 and R = 20 Mpc/h, using a Top-Hat kernel. We note that
for sufficiently large scales the results are very robust against changes in R. This observation
allows us to use standard methods to introduce curvature bias into the perturbation theory
of marks by utilising the relation (D.1). Perhaps the simplest method to do so is by adding

2(1 + b1 + B1 )b∇2 δ x∇2 (r) + b2∇2 δ x∇4 (r), (D.2)

to the W function, and

2(1 + b1 )b∇2 δ x∇2 (r) + b2∇2 δ x∇4 (r), (D.3)
to the correlation function of tracers [33], with CLPT
ξX

d3 q
Z
1 T −1
x∇ 2 = 3/2 1/2
e− 2 (r−q) AL (r−q) ∇2 ξ(q), (D.4)
(2π) |AL |
d3 q
Z
1 T −1
x∇4 = 3/2 1/2
e− 2 (r−q) AL (r−q) ∇4 ξ(q). (D.5)
(2π) |AL |

– 31 –
Analogously, this prescription can be used to add tidal bias with the standard methods of
CLPT (see ref. [68] and its extension to MG in ref. [36]): whenever the linear bias parameter,
b1 , appears accompanying the tidal bias in the correlation function ξ, substitute it by b1 + B1
to obtain the tidal bias contributions to the weighted correlation function W .
We have introduced curvature bias for theoretical consistency in MG models. However,
we find that for reasonable values of b∇2 δ its influence on the mCF is negligible above a
smoothing scale R = 10 Mpc/h, which is the scale we use in section 6 to compare the
perturbation theory models to simulations.

References

[1] Planck collaboration, N. Aghanim et al., Planck 2018 results. VI. Cosmological parameters,
1807.06209.
[2] F. Bernardeau, S. Colombi, E. Gaztanaga and R. Scoccimarro, Large scale structure of the
universe and cosmological perturbation theory, Phys. Rept. 367 (2002) 1–248,
[astro-ph/0112551].
[3] D. Baumann, A. Nicolis, L. Senatore and M. Zaldarriaga, Cosmological Non-Linearities as an
Effective Fluid, JCAP 1207 (2012) 051, [1004.2488].
[4] M. White, A marked correlation function for constraining modified gravity models, JCAP 1611
(2016) 057, [1609.08632].
[5] C. Beisbart and M. Kerscher, Luminosity- and morphology-dependent clustering of galaxies,
Astrophys. J. 545 (2000) 6, [astro-ph/0003358].
[6] C. Beisbart, M. Kerscher and K. Mecke, Mark correlations: Relating physical properties to
spatial distributions, Lecture Notes in Physics (2002) 358390.
[7] S. Gottloeber, M. Kerscher, A. V. Kravtsov, A. Faltenbacher, A. Klypin and V. Mueller,
Spatial distribution of galactic halos and their merger histories, Astron. Astrophys. 387 (2002)
778, [astro-ph/0203148].
[8] R. K. Sheth, The halo-model description of marked statistics, Mon. Not. Roy. Astron. Soc. 364
(2005) 796, [astro-ph/0511772].
[9] R. K. Sheth, A. J. Connolly and R. Skibba, Marked correlations in galaxy formation models,
Submitted to: Mon. Not. Roy. Astron. Soc. (2005) , [astro-ph/0511773].
[10] R. Skibba, R. K. Sheth, A. J. Connolly and R. Scranton, The luminosity-weighted or ‘marked’
correlation function, Mon. Not. Roy. Astron. Soc. 369 (2006) 68–76, [astro-ph/0512463].
[11] M. White and N. Padmanabhan, Breaking Halo Occupation Degeneracies with Marked
Statistics, Mon. Not. Roy. Astron. Soc. 395 (2009) 2381, [0812.4288].
[12] Supernova Cosmology Project collaboration, S. Perlmutter et al., Measurements of
Omega and Lambda from 42 high redshift supernovae, Astrophys. J. 517 (1999) 565–586,
[astro-ph/9812133].
[13] Supernova Search Team collaboration, A. G. Riess et al., Type Ia supernova discoveries at
z ¿ 1 from the Hubble Space Telescope: Evidence for past deceleration and constraints on dark
energy evolution, Astrophys. J. 607 (2004) 665–687, [astro-ph/0402512].
[14] G. W. Horndeski, Second-order scalar-tensor field equations in a four-dimensional space, Int. J.
Theor. Phys. 10 (1974) 363–384.
[15] G. R. Dvali, G. Gabadadze and M. Porrati, 4-D gravity on a brane in 5-D Minkowski space,
Phys. Lett. B485 (2000) 208–214, [hep-th/0005016].

– 32 –
[16] S. Capozziello, Curvature quintessence, Int. J. Mod. Phys. D11 (2002) 483–492,
[gr-qc/0201033].
[17] S. M. Carroll, V. Duvvuri, M. Trodden and M. S. Turner, Is cosmic speed - up due to new
gravitational physics?, Phys. Rev. D70 (2004) 043528, [astro-ph/0306438].
[18] A. Nicolis, R. Rattazzi and E. Trincherini, The Galileon as a local modification of gravity,
Phys. Rev. D79 (2009) 064036, [0811.2197].
[19] A. I. Vainshtein, To the problem of nonvanishing gravitation mass, Phys. Lett. 39B (1972)
393–394.
[20] J. Khoury and A. Weltman, Chameleon fields: Awaiting surprises for tests of gravity in space,
Phys. Rev. Lett. 93 (2004) 171104, [astro-ph/0309300].
[21] J. Khoury and A. Weltman, Chameleon cosmology, Phys. Rev. D69 (2004) 044026,
[astro-ph/0309411].
[22] T. Clifton, P. G. Ferreira, A. Padilla and C. Skordis, Modified Gravity and Cosmology, Phys.
Rept. 513 (2012) 1–189, [1106.2476].
[23] K. Koyama, Cosmological Tests of Modified Gravity, Rept. Prog. Phys. 79 (2016) 046902,
[1504.04623].
[24] B. Li and K. Koyama, Modified Gravity. WORLD SCIENTIFIC, 2019, 10.1142/11090.
[25] G. Valogiannis and R. Bean, Beyond δ: Tailoring marked statistics to reveal modified gravity,
Phys. Rev. D97 (2018) 023535, [1708.05652].
[26] C. Hernández-Aguayo, C. M. Baugh and B. Li, Marked clustering statistics in f (R) gravity
cosmologies, Mon. Not. Roy. Astron. Soc. 479 (2018) 4824–4835, [1801.08880].
[27] J. Armijo, Y.-C. Cai, N. Padilla, B. Li and J. A. Peacock, Testing modified gravity using a
marked correlation function, Mon. Not. Roy. Astron. Soc. 478 (2018) 3627–3632, [1801.08975].
[28] S. Satpathy, R. A. C. Croft, S. Ho and B. Li, Measurement of marked correlation functions in
SDSS-III Baryon Oscillation Spectroscopic Survey using LOWZ galaxies in Data Release 12,
Mon. Not. Roy. Astron. Soc. 484 (2019) 2148–2165, [1901.01447].
[29] J. Carlson, B. Reid and M. White, Convolution Lagrangian perturbation theory for biased
tracers, Mon. Not. Roy. Astron. Soc. 429 (2013) 1674, [1209.0780].
[30] K. Koyama, A. Taruya and T. Hiramatsu, Non-linear Evolution of Matter Power Spectrum in
Modified Theory of Gravity, Phys. Rev. D79 (2009) 123512, [0902.0618].
[31] A. Aviles and J. L. Cervantes-Cota, Lagrangian perturbation theory for modified gravity, Phys.
Rev. D96 (2017) 123526, [1705.10719].
[32] H. A. Winther, K. Koyama, M. Manera, B. S. Wright and G.-B. Zhao, COLA with
scale-dependent growth: applications to screened modified gravity models, JCAP 1708 (2017)
006, [1703.00879].
[33] A. Aviles, M. A. Rodriguez-Meza, J. De-Santiago and J. L. Cervantes-Cota, Nonlinear
evolution of initially biased tracers in modified gravity, JCAP 1811 (2018) 013, [1809.07713].
[34] A. Aviles, J. L. Cervantes-Cota and D. F. Mota, Screenings in Modified Gravity: a perturbative
approach, Astron. Astrophys. 622 (2019) A62, [1810.02652].
[35] G. Valogiannis and R. Bean, Convolution Lagrangian Perturbation Theory for biased tracers
beyond general relativity, 1901.03763.
[36] G. Valogiannis, R. Bean and A. Aviles, An accurate perturbative approach to redshift space
clustering of biased tracers in modified gravity, 1909.05261.

– 33 –
[37] G. Valogiannis and R. Bean, Efficient simulations of large scale structure in modified gravity
cosmologies with comoving Lagrangian acceleration, Phys. Rev. D95 (2017) 103515,
[1612.06469].
[38] C. Moretti, S. Mozzon, P. Monaco, E. Munari and M. Baldi, Fast numerical method to generate
halo catalogs in modified gravity (part I): second-order Lagrangian Perturbation Theory,
1909.06282.
[39] W. Hu and I. Sawicki, Models of f(R) Cosmic Acceleration that Evade Solar-System Tests,
Phys. Rev. D76 (2007) 064004, [0705.1158].
[40] WMAP collaboration, G. Hinshaw et al., Nine-Year Wilkinson Microwave Anisotropy Probe
(WMAP) Observations: Cosmological Parameter Results, Astrophys. J. Suppl. 208 (2013) 19,
[1212.5226].
[41] M. Cautun, E. Paillas, Y.-C. Cai, S. Bose, J. Armijo, B. Li et al., The
SantiagoHarvardEdinburghDurham void comparison I. SHEDding light on chameleon gravity
tests, Mon. Not. Roy. Astron. Soc. 476 (2018) 3195–3217, [1710.01730].
[42] F. Schmidt, D. Jeong and V. Desjacques, Peak-Background Split, Renormalization, and Galaxy
Clustering, Phys. Rev. D88 (2013) 023515, [1212.0868].
[43] V. Desjacques, D. Jeong and F. Schmidt, Large-Scale Galaxy Bias, Phys. Rept. 733 (2018)
1–193, [1611.09787].
[44] T. Baldauf, M. Mirbabayi, M. Simonovi and M. Zaldarriaga, Equivalence Principle and the
Baryon Acoustic Peak, Phys. Rev. D92 (2015) 043514, [1504.04366].
[45] T. Matsubara, Nonlinear perturbation theory with halo bias and redshift-space distortions via
the Lagrangian picture, Phys. Rev. D78 (2008) 083519, [0807.1733].
[46] A. Aviles, Renormalization of Lagrangian bias via spectral parameters, Phys. Rev. D98 (2018)
083541, [1805.05304].
[47] A. Eggemeier, R. Scoccimarro and R. E. Smith, Bias Loop Corrections to the Galaxy
Bispectrum, Phys. Rev. D99 (2019) 123514, [1812.03208].
[48] N. Kaiser, On the Spatial correlations of Abell clusters, Astrophys. J. 284 (1984) L9–L12.
[49] H. J. Mo, Y. P. Jing and S. D. M. White, High-order correlations of peaks and halos: A Step
toward understanding galaxy biasing, Mon. Not. Roy. Astron. Soc. 284 (1997) 189,
[astro-ph/9603039].
[50] R. K. Sheth and G. Tormen, Large scale bias and the peak background split, Mon. Not. Roy.
Astron. Soc. 308 (1999) 119, [astro-ph/9901122].
[51] T. Matsubara, Resumming Cosmological Perturbations via the Lagrangian Picture: One-loop
Results in Real Space and in Redshift Space, Phys. Rev. D77 (2008) 063530, [0711.2521].
[52] S. Tassev, Lagrangian or Eulerian; Real or Fourier? Not All Approaches to Large-Scale
Structure Are Created Equal, JCAP 1406 (2014) 008, [1311.4884].
[53] Z. Vlah, U. Seljak and T. Baldauf, Lagrangian perturbation theory at one loop order: successes,
failures, and improvements, Phys. Rev. D91 (2015) 023508, [1410.1617].
[54] Z. Vlah, M. White and A. Aviles, A Lagrangian effective field theory, JCAP 1509 (2015) 014,
[1506.05264].
[55] M. Schmittfull, Z. Vlah and P. McDonald, Fast large scale structure perturbation theory using
one-dimensional fast Fourier transforms, Phys. Rev. D93 (2016) 103528, [1603.04405].
[56] Z. Slepian and D. J. Eisenstein, Modelling the large-scale redshift-space 3-point correlation
function of galaxies, Mon. Not. Roy. Astron. Soc. 469 (2017) 2059–2076, [1607.03109].

– 34 –
[57] C. Arnold, P. Fosalba, V. Springel, E. Puchwein and L. Blot, The modified gravity lightcone
simulation project I: Statistics of matter and halo distributions, Mon. Not. Roy. Astron. Soc.
483 (2019) 790–805, [1805.09824].
[58] M. Cautun, E. Paillas, Y.-C. Cai, S. Bose, J. Armijo, B. Li et al., The
Santiago-Harvard-Edinburgh-Durham void comparison - I. SHEDding light on chameleon
gravity tests, MNRAS 476 (May, 2018) 3195–3217, [1710.01730].
[59] B. Li, G.-B. Zhao, R. Teyssier and K. Koyama, ECOSMOG: an Efficient COde for Simulating
MOdified Gravity, JCAP 2012 (Jan, 2012) 051, [1110.1379].
[60] M. Manera, R. Scoccimarro, W. J. Percival, L. Samushia, C. K. McBride, A. J. Ross et al., The
clustering of galaxies in the SDSS-III Baryon Oscillation Spectroscopic Survey: a large sample
of mock galaxy catalogues, MNRAS 428 (Jan, 2013) 1036–1054, [1203.6609].
[61] D. Alonso, CUTE solutions for two-point correlation functions from large cosmological datasets,
arXiv e-prints (Oct, 2012) arXiv:1210.1833, [1210.1833].
[62] P. S. Behroozi, R. H. Wechsler and H.-Y. Wu, The Rockstar Phase-Space Temporal Halo
Finder and the Velocity Offsets of Cluster Cores, Astrophys. J. 762 (2013) 109, [1110.4372].
[63] B. Li and H. Zhao, Structure Formation by the Fifth Force III: Segregation of Baryons and
Dark Matter, Phys. Rev. D81 (2010) 104047, [1001.3152].
[64] L. Lombriser, K. Koyama and B. Li, Halo modelling in chameleon theories, JCAP 1403 (2014)
021, [1312.1292].
[65] T. Matsubara, Recursive Solutions of Lagrangian Perturbation Theory, Phys. Rev. D92 (2015)
023534, [1505.01481].
[66] E. Bertschinger and P. Zukin, Distinguishing Modified Gravity from Dark Energy, Phys. Rev.
D78 (2008) 024015, [0801.2431].
[67] A. Silvestri, L. Pogosian and R. V. Buniy, Practical approach to cosmological perturbations in
modified gravity, Phys. Rev. D87 (2013) 104015, [1302.1193].
[68] Z. Vlah, E. Castorina and M. White, The Gaussian streaming model and convolution
Lagrangian effective field theory, JCAP 1612 (2016) 007, [1609.02908].

– 35 –