Chapter 7

Measuring Power spectrum and Bispectrum

In this chapter, we provide a practical guide for measuring power spectrum and
bispectrum by using Fast Fourier Transforms. First, we show the method of calculating the
power spectrum and bispectrum from N-body simulations, where input density contrast field
is defined in a well defined cubic box. Then, we move to the power spectrum estimation
for general cases and show the way we implement the FKP estimator (Feldman et al.,
1994) with Fast Fourier Transform. For those two cases, we make special emphases on the
normalization coefficients, the shot noise correction and the window function.

7.1

Power spectrum and Bispectrum from N-body simulation

Estimating power spectrum and bispectrum from the N-body simulation data is

less complicated as N-body simulations have 1) the cubic box, 2) the constant mean number
density. We divide the general procedure of measuring power spectrum and bispectrum
from N-body simulation by following five steps:
(1) Distributing particles onto the regular grid
(2) Fourier transformation
(3) Estimating power spectrum and bispectrum
(4) Deconvolving window function
(5) Subtracting shot noise
7.1.1

From particle to grid

In order to apply the Fast Fourier Transform technique, we have to assign the density

field onto each point in the regular grid. The way we distribute a particle to the nearby
grid points is called a particle distribution scheme. For a given distribution scheme, we

209

can define an associated shape function, which quantifies how a quantity (mass, number,
luminosity, etc) of particle is distributed. After this process, the sampling we made from the
particle distribution is not a mere sampling of the underlying density field, but a sampling
convolved with the window function of particle distribution scheme. In this section, we
shall review the three distribution schemes which are widely used in practice: NearestGrid-Point, Cloud-In-Cloud, Triangular-Shape-Cloud. Although we use the particle number
density as a representative example below, one can use the same equation for calculating
mass weighted power spectrum, luminosity weighted power spectrum, etc, by multiplying
the apropos quantity (mass, luminosity) of each particle.
Let us consider the case with Np particles (e.g. dark matters, halos, galaxies) in a
N-body simulation box. The particle number density is given by
n0 (x) =

Np
!
i=1

D (x xi ),

(7.1)

where xi is the position of i-th particle. Then, the particle number assignment can be
formulated by the convolution as
n(x) =

"

d3 x! n0 (x! )W (x x! ),

(7.2)

where W (x) is the window function which quantifies how much of this particle number
density is distributed to a grid point separated by x. We sample the continuous number
density n(x) by the regular grids of size N 3 :
"
ns (xp ) n(xp ) =
d3 x! n(x! )W (xp x! ).

(7.3)

Therefore, the sampled density contrast, defined as s (x) ns (x)/

n 1 is given by

the convolution of the real density contrast and the window function,
m (x) = [ " W ] (x),

(7.4)

and Fourier transformation of the sampled density contrast is

m (k) = (k)W (k).

(7.5)

This procedure of convolving with window function can be think of as following.

First, we define the cloud shape function (or point spreading function) S(x! ) of a particle.
The shape function can be uniquely determined for given distribution scheme such that the

210

fraction of the particle number of the particle at x assigned to the grid point xp is given by
integrating the shape function within the cubic cell surrounding xp (Hockney & Eastwood,
1988, p. 142). That is, the number of a particle at x assigned to grid point xp n(x xp )
is given by

n(x xp ) =

"

|x! xp |i <H/2

d3 x! S(x! x).

(7.6)

In one-dimensional case, the window function can be written in terms of the cloud
shape function as
W (x xp ) = Wp (x) =

1
H

where H = L/Ngrid is the separation of grids.

1
1/2
T(x) =

"

xp +H/2

xp H/2

S(x! x)dx! ,

(7.7)

Using the top-hat function, T(x),

if |x| < 1/2
if |x| = 1/2 ,
if otherwise

(7.8)

equation (7.7) can be also written as a convolution of top-hat function and cloud function
as:
W (x) =

1
H

"

&

x!
H

'

S(x! x)dx!

(7.9)

There are three most widely used distribution (window) functions.

7.1.1.1

NGP
Nearest Grid Point (NGP) scheme assigns particles to their nearest grid points.

Therefore, the number density changes discontinuously when particles cross cell boundaries.
The one dimensional window function for NGP is proportional to the top-hat function

1/H
if |x| < H/2
1 (x)
1/(2H)
if |x| = H/2 ,
WN GP (x) T
=
(7.10)

H
H
0
if otherwise
and its point spreading function is the Dirac delta function as
(x)
1 (x)
1 (x)
1 (x)
1
= T
D (x) = T
D
.
T
H
H
H
H
H
H
H
H

The Fourier Transformation of the top-hat function is the sinc function.

& '
sin (k/2)
k
T(k) =
= sinc
k/2
2
211

(7.11)

(7.12)

Proof.
T(k) =

"

1/2

eikx dx =

1/2

eik/2 eik/2
sin(k/2)
=
ik
k/2

(7.13)

Therefore, the Fourier Transformation of the NGP window function is

&
'
&
'
Hk
k
WN GP (k) = FT[T](Hk) = sinc
= sinc
2
2kN
where, kN = /H is the Nyquist frequency. I use the similarity theorem of the Fourier
transformation.
7.1.1.2

CIC
Cloud In Cell (CIC) assignment is the first order distribution scheme which uni-

formly distributes the particle with top-hat spreading function. In other words, the cloud
shape function is given by

1 (x)
T
.
H
H

(7.14)

1 (x)
1 (x)
T
,
T
H
H
H
H

(7.15)

SCIC (x) =
Therefore, the window function is
WCIC (x) =

and its Fourier Transformation can be simply obtained by the convolution theorem:
&
'
k
WCIC (k) = WN GP (k)2 = sinc2
.
(7.16)
2kN
The explicit expression of the CIC window function is
*
1
1 |x|/H
if |x| < H
WCIC (x) =
.
0
otherwise
H
7.1.1.3

(7.17)

TSC
Triangular Shaped Cloud (TSC) scheme is the second order distribution scheme.

Its point spreading function is triangular, as the convolution of the two first order (CIC)
functions:
ST SC (x) =

1 (x)
1 (x)
T
T
.
H
H
H
H
212

(7.18)

Therefore, the window function is

WT SC (x) =

1 (x)
1 (x)
1 (x)
T
T
,
T
H
H
H
H
H
H

(7.19)

and its Fourier Transformation is given by

WT SC (k) = WN GP (k) = sinc

The explicit expression of TSC scheme is
3 , x -2

1 41 ( 3 H |x| )2
WT SC (x) =
H
H
2 2
0
7.1.1.4

&

k
2kN

if |x|

'

(7.20)

H
2

if H2 |x|
otherwise

3H
2

(7.21)

3D window function
As we use the regular cubic grid, the three dimensional window function is simply

given as the multiplication of three one dimensional window functions.

W (x) = W (x1 )W (x2 )W (x3 )
Therefore, its Fourier transformation is
.
&
'
&
'
&
'/p
k1
k2
k3
W (k) = sinc
sinc
sinc
,
2kN
2kN
2kN

(7.22)

(7.23)

where p = 1, 2, 3 for NGP, CIC and TSC, respectively.

7.1.2

Power spectrum and bispectrum: the estimators

The power spectrum and bispectrum are defined as
&(k1 )(k2 )' = (2)3 P (k1 ) D (k1 + k2 )
3

&(k1 )(k2 )(k3 )' = (2) B(k1 , k2 , k3 ) (k1 + k2 + k3 ).

(7.24)
(7.25)

Note that both power spectrum and bispectrum are real because of the parity invariance
and the reality of the configuration space n-point correlation function. (See, appendix A of
Smith et al. (2008b))
With this definition, 8 is
"
"
d3 k
dk 2
2
82 =
P
(k)|W
(kR)|
=
(k)|W (kR)|2
(2)3
k
213

with R = 8Mpc/h, and 2 (k) = P (k)k 3 /2 2 is the dimensionless power spectrum. One can
check the normalization of the power spectrum by calculating 8 . In some literature, the
definition of power spectrum and bispectrum does not contain (2)3 , and we have to drop
(2)3 in the integration measure of 82 equation above.
For the normalization of bispectrum, one can check the value of the reduced bispectrum Q(k1 , k2 , k3 ) which is defined as
Q(k1 , k2 , k3 ) =

B(k1 , k2 , k3 )
.
P (k1 )P (k2 ) + P (k2 )P (k3 ) + P (k3 )P (k1 )

On large scales where so called tree-level bispectrum model works, the reduced bispectrum
is 4/7 for the equilateral configuration.
In this section, we shall find the proper normalization to the power spectrum and
bispectrum estimators which use the unnormalized Fast Fourier Transformation (FFT) such
as FFTW. For denote the unnormalized discrete Fourier transform result by superscript
FFTW.
First, we nondimensionalize the Dirac delta function in the definition of power
spectrum and bispectrum. Using the property of delta function,
!
1
D (x xi ),
D (f (x)) =
! (x )|
|f
i
x zeros
i

we can express the delta function in Fourier space in terms of the Kronecker delta of integer
triplet nk . We denote the Kronecker delta for such integer triplet as K (nk ).
0 1
0
1
nKi = 3 K (nk )
D (kp ) = D (kF nk ) =
D (kF ni ) =
k
k
F
F
i
i
7.1.2.1

Power spectrum estimator: direct sampling

FFTW output is the unnormalized DFT, which is related to the sampled Fourier

space density field as

F F T W (nk ) =

(nr )ei2nk nr /N =

nr

!
rp

(rp )eikp rp =

(kp )
.
H3

(7.26)

From the definition of power spectrum

&(k1 )(k2 )' = H 6 & F F T W (n1 ) F F T W (n2 )' = (2)3 P (k1 ) D (k1 + k2 )
=

(2)3 K
(n1 + n2 )P (kF n1 ),
kF3

214

(7.27)

we find the normalization for the power spectrum from DFT as

P (kF n1 ) =

H 6 kF3 F F T W
&
(n1 ) F F T W (n1 ), '
(2)3

(7.28)

where V is the volume of survey, N is number of one-dimensional grid, H 3 = V /N 3 and

kF3 = (2)3 /V . Therefore, the final estimator for power spectrum is

!
22 3
V 122 F F T W
V
1
F
F
T
W
2
P (kF n1 ) = 6
(n1 )2 = 6
|
(nk )| ,
N
N
Nk
1

(7.29)

|nk n1 | 2

where we sum over all Fourier modes within k1 kF /2 < |k| < k1 + kF /2 to estimate the

power spectrum at k = k1 = kF n1 .
7.1.2.2

Bispectrum estimator: direct sampling

From the definition of the bispectrum, we also find

&(k1 )(k2 )(k3 )' = H 9 & F F T W (n1 ) F F T W (n2 ) F F T W (n3 )'

= (2)3 B(k1 , k2 , k3 ) D (k123 ) =

(2)3 D
(n123 )B(k1 , k2 , k3 ).
kF3

(7.30)

Therefore, the estimator of bispectrum from DFT is

B(kF n1 , kF n2 , kF n3 )
H 9 kF3 F F T W
&
(n1 ) F F T W (n2 ) F F T W (n3 )' D (n123 )
(2)3
V2
= 9 & F F T W (n1 ) F F T W (n2 ) F F T W (n3 )' D (n123 )
N 8
9
!
V2
1
FFTW
FFTW
FFTW
= 9

(m1 )
(m2 )
(m3 ) ,
N
Ntriangle
=

(7.31)

mTri123

where Tri123 is the set of {m1 , m2 , m3 } whose magnitude satisfies |mi ni | 1/2 and three

vectors form a triangle, i.e. m1 + m2 + m3 = 0.

7.1.2.3

Estimating power spectrum II

There is another way of estimating the power spectrum. We slightly change the

previous estimator as
V FFTW
&
(n1 ) F F T W (n2 )' K (n1 + n2 )
N6 :
:
FFTW
(m1 ) F F T W (m2 ) K (m1 + m2 )
V
m1 %n1
m2 %n1
:
:
.
= 6
K
N
m1 %n1
m2 %n1 (m1 + m2 )

P (kF n1 ) =

215

(7.32)

Where, mi ( ni strictly means that |mi ni | < s/2, with k = skF . Note that
"
"
! !
1
4k12 k
K (m1 + m2 ) =
d3 q1
d3 q2 D (q1 + q2 ) (
= 4sn21
3
V
k
f
k
k
F
1
2
m %n m %n
1

is the total number of independent k-modes inside the spherical shell of radius k and width
k = skF . We shall use the total number of k modes here instead of the actual number of k
modes, because when we do the inverse DFT later, both (k) and (k) shall be summed
over.
Using the orthonormality of DFT
1 ! 2inr np /N 2inr nq /N
K (np + nq ) = 3
e
e
,
N n

(7.33)

we transform the estimator as

P (kF n1 )
1
1 ! ! ! FFTW
V

(m1 ) F F T W (m2 )e2inr m1 /N e2inr m2 /N

= 6
N 4sn21 N 3 n m %n m %n
r
1
1
2
1
<;
<
;
!
1
1 ! ! FFTW
V
FFTW
2inr m2 /N
2inr m1 /N

(m1 )e

(m2 )e
= 6
N 4sn21 N 3 n m %n
m %n
r

(7.34)

Lets define mi (nr ) as

ni (nr ) =

F F T W (m)e2inr m/N

m%ni

In practice, ni (nr ) can be calculated by applying fftw dft c2r to the array whose values are
F F T W (m) when |m ni | < s/2 and otherwise zero. Note that ni (nr ) is real.
Proof. One can break the summation into two part, one with positive mx index, and another
with negative mx index.
!
ni (nr ) =
F F T W (m)e2inr m/N
m%ni

m%ni ,mx >0

m%ni ,mx >0

m%ni ,mx >0

=
>
F F T W (m)e2inr m/N + F F T W (m)e2inr m/N

=
>
FFTW
F F T W (m)e2inr m/N + [ ]
(m)e2inr m/N
=
>
F F T W (m)e2inr m/N + c.c. R

In third line, we used the reality of (r), then (k) = (k).

216

Then, we find the second estimator for the power spectrum as

P (kF n1 ) =

V
1
1 !
n1 (nr )2 .
2
6
N 4sn1 N 3 n

(7.35)

This method takes more time than the first estimator which uses the direct sampling method.
However, real strength of this method is apparent when we calculate the higher order
polyspectra, e.g. Bispectrum and Trispectrum etc, because we do not have to explicitly
sum up all the possible triangles (Bispectrum) or rectangles (Trispectrum), etc.
7.1.2.4

Estimating Bispectrum II
We can similarly construct the Bispectrum estimator. The direct sampling estima-

tor we derived before is given by

B(kF n1 , kF n2 , kF n3 ) =

V 2 FFTW
&
(n1 ) F F T W (n2 ) F F T W (n3 )' K (n1 + n2 + n3 ), (7.36)
N9

where we estimate the ensemble average by summing over the all possible triangles with
side of ki skF /2 < qi < ki + skF /2. Therefore, the estimator can be recasted as
B(kF n1 , kF n2 , kF n3 )
:
:
:
F F T W (m1 ) F F T W (m2 ) F F T W (m3 ) K (m123 )
V2
:
= 9 m1 %n1 m2 %n2 :m3 %n3 :
. (7.37)
K
N
m1 %n1
m2 %n2
m3 %n3 (m123 )

Again, mi ( ni means |mi ni | < s/2. The denominator is the number of possible triangles

with side of (k1 , k2 , k3 ) :

!

D (m123 ) =

m1 %n1 m2 %n2 m3 %n3

1
Vf2

"

d3 q1
k1

"

k2

d3 q2

"

k3

d3 q3 D (q123 )

8 2 k1 k2 k3 (k)3
= 8 2 s3 n1 n2 n3
kF6

(7.38)

By using the normalization of DFT,

D (np + nq + ns ) =

1 ! 2inr np /N 2inr nq /N 2inr ns /N

e
e
e
,
N3 n
r

and following the exactly same procedure of finding the second estimator for power spectrum,
we find the second bispectrum estimator
&
'
V2
1
1 !
B(kF n1 , kF n2 , kF n3 ) = 9
n1 (nr )n2 (nr )n3 (nr ).
2
3
N
8 s n1 n2 n3 N 3 n
r

217

(7.39)

7.1.2.5

Counting the number of triangles

In the previous section, we use the integral approximation to estimate the number

of triangles. However, this approximation is broken down when three vectors are parallel to
each other, i.e. when
k1 = k2 = k3 .
Also, it is not an accurate approximation for the triangle including the large scale modes.
Therefore, in this section, we present the way we calculate the actual number of triangles
for given triangular configurations.
We shall use the exactly same trick as you used in the previous section for estimating
bispectrum. Since it is a trivial normalization factor, the number of triangles are the same
as the bispectrum for the unit density contrast. i.e.
! ! !
Ntri (n1 , n2 , n3 ) =
K (m1 + m2 + m3 )
m1 %n1 m2 %n2 m3 %n3

1 ! ! ! ! 2inr m1 /N 2inr m2 /N 2inr m3 /N

e
e
e
N 3 n m %n m %n m %n
r
1
1
2
2
3
3
<
<;
<;
;
!
!
1 ! ! 2inr m1 /N
2inr m2 /N
2inr m3 /N
e
e
e
= 3
N n m %n
m2 %n2
m3 %n3
r
1
1
1 !
3
In1 (nr )In2 (nr )In3 (nr ),
N n

where
Ini (nr )

e2inr mi /N .

m1 %ni

The function Ini (nr ) is the inverse Fourier transformation of the function in k space, which
has the value unity within the shell of |mi ni | < s/2, otherwise zero.
By calculating the number of triangles, and compare the result with the exact
calculation, we find that the logical size of the 1D Fourier grid has to be at least three times
as large as that of the maximum wavenumber for which we want to estimate the bispectrum.
That is, the array size of calculating Ini (nr ), bnmesh, has to satisfy
bnmesh > 3(s nkmax),

(7.40)

where s and nkmax are the bin size, and maximum wavenumber in the unit of the fundamental frequency. This is to avoid the fictitious increasing of the number of triangles (when
calculating the number of triangles) or power (when calculating bispectrum) due to the
aliasing effect.

218

7.1.3

Deconvolution
Now, we have the estimators for the power spectrum and the bispectrum. However,

as we have employed the distribution scheme, the power spectrum and the bispectrum we
would measure with those estimators are not the same as the power/bi-spectrum of the
real density contrast, but the power/bi-spectrum of density contrast convolved with the
window function. Therefore, the power spectrum and bispectrum we estimate will show the
artificial power suppression on small scales. Therefore, we have to deconvolve the window
function due to the particle distribution scheme in order to estimate the power spectrum
and bispectrum of the true density contrast.
7.1.3.1

Deconvolving only window function

First, as we know the exact shape of the window function in Fourier space, we can

simply divide the resulting density contrast in Fourier space by the window function. That
is, we deconvolve each k mode of density contrast as
(k) =

m (k)
,
W (k)

or, deconvolve the estimated power spectrum by

2 m 22
'
&
'
&
'/2p
.
&
2 (k) 2
2 = P m (k1 , k2 , k3 ) sinc k1 sinc k2 sinc k3
,
P (k) = 22
W (k) 2
2kN
2kN
2kN

(7.41)

(7.42)

for k < kN . Again, p = 1, 2, 3 for NGP, CIC and TSC scheme, respectively. Here, superscript
m denote the measured quantity.
7.1.3.2

Deconvolving window function and aliasing

When we want to extract the power spectrum for k ( kN 1 , we have to take the

alias effect into account. As we show in Appendix A.2.1, the Fourier counterpart of the
sampled data are the aliased sum of Fourier transformation of the continuous function. i.e.
!
2
P m (k) =
|W (k + 2kN n)| P (k + 2kN n)
(7.43)
n

For the case of uniformly distributed random particles, the power spectrum is simply
a constant (Poisson shot noise), 1/n, where n is the number density. In that case, the
1 Actually, it may not happen for most of N-body simulations because of the Force/Mass resolution is
usually too poor to use the power spectrum for such a high wavenumber.

219

convolved power spectrum becomes

P m (k) =

1!
2
|W (k + 2kN n)| .
n n

(7.44)

Let us calculate the convolved power spectrum analytically for p = 1, 2, 3 cases:

;
<
3
!
0
!
2
2
|W (k + 2kN n)| =
W (ki + 2kN n)
n=

i=1

)
(ki + 2kN n)
)2p

2kN (ki + 2kN n)

2p
3

0
! sin
=

(
n=

i=1

2kN

&
' !
3

0
2p ki
sin
=

(
2kN n= ki
i=1

2kN

)2p
+ n)

(7.45)

Here, the infinite summation can be calculated by using following identity (Jing, 2005). (so
called Glaishers series 2 )

M=

=
a >2
1
=
csc
(a + M d)2
d
d

(7.46)

NGP (p=1)
Lets put d = .

1
= csc2 a
(a
+
n)2
n=
Therefore, for p = 1,
!
n

|W (k + 2kN n)|2 =

3
0

sin2

i=1

&

ki
2kN

'

csc2

(7.47)

&

ki
2kN

'

= 1,

(7.48)

and sum over infinite aliasing effect of power suppression ends up giving rise to the original
power.
CIC (p=2)
Differentiate both side with respect to a twice, we get

6
= 2(3 2 sin2 a) csc4 a
(a
+
n)4
n=

(7.49)

2 F. Soddy, Proceedings of the Royal Society of London, Series A, Mathematical and Physical Sciences,
Vol. 182, No. 989. (Dec. 16.1943), pp. 113-129

220

Therefore, the aliased sum becomes

!
n

3
0

&

sin4

ki
2kN

'.

2
sin2
3
i=1
&
'/
3 .
0
2
ki
=
1 sin2
.
3
2kN
i=1

|W (k + 2kN n)| =

&

ki
2kN

'/

csc4

&

ki
2kN

'
(7.50)

TSC (p=3)
Differentiate the both side with a four times, we get

120
= (120 120 sin2 a + 16 sin4 a) csc6 a
6
(a
+
n)
n=

(7.51)

Therefore,
!
n

3
0

|W (k + 2kN n)|

'.
&
'
&
'/
&
'
ki
ki
2
ki
ki
1 sin2
+
csc6
sin4
2kN
2kN
15
2kN
2kN
i=1
.
&
'
&
'/
3
0
ki
2
ki
1 sin2
+
.
=
sin4
2k
15
2k
N
N
i=1
=

sin6

&

(7.52)

When the measured power spectrum include the both true power spectrum and
Poisson shot noise, we have to cure it iteratively, See, e.g. Jing (2005). However, as the
nonlinear galaxy power spectrum is dominated by the constant term P0 and large k plateau
of nonlinear bias terms, Pb2 , Pb22 , we may simply apply the same aliased window function
as for the constant power spectrum. In practice, we calculated both window-corrected and
window-alias-corrected power spectrum, and only use the wavenumber ranges where those
two power spectra agree with each other.
7.1.4

Poisson shot noise

Finally, we have to subtract the Poisson shot noise from the deconvolved power

spectrum and bispectrum. As a general terminology, shot noise refers to the self-particle
contribution to the statistics. We call it Poisson shot noise, as it appears when we think
of a realization of galaxy (matter) distribution as a Poisson sampling of underlying smooth
galaxy (matter) density contrast field. Poisson shot noise appears whenever we calculate
the n-point function from a discrete set of objects, like a galaxy, dark matter, etc.

221

7.1.4.1

Poisson sampling and underlying density field

Suppose we have a point process n(r) which is a Poisson sample of some contin-

uous stochastic field 1 + (r). That is, the probability that an infinitesimal volume element
V contains an object is n
(r) [1 + (r)] V .
Following Peebles (1980, 36) we describe the process by dividing the space into the

infinitesimal micro-cells of volume V which has a occupation numbers ni = 0 or 1. That

is, the statistical average of self-correlator for a given cell is
E 3F E 2F
ni = ni = &ni ' = n
(ri )Vi ,

(7.53)

and the correlator for different cells are given by the underlying density contrast as
&ni nj 'i(=j
=
n(ri )
n(rj )Vi Vj [1 + &(ri )(rj )']

(7.54)

&ni nj nk 'i(=j(=k
=
n(ri )
n(rj )
n(rk )Vi Vj Vk
[1 + &(ri )(rj )' + &(rj )(rk )' + &(ri )(rk )' + &(ri )(rj )(rk )'].
7.1.4.2

(7.55)

Power spectrum and Bispectrum of discrete particles

We follow the derivation given in Feldman et al. (1994). Consider the expectation

value of

"

d3 r

"

d3 r! g(r, r! )n(r)n(r! ),

for an arbitrary function g(r, r! ). By using the infinitesimal micro-cells, the expectation
value becomes
G"

H
"
d3 r d3 r! g(r, r! )n(r)n(r! )
"
"
= d3 r d3 r! g(r, r! ) &n(r)n(r! )'
!
g(ri , rj ) &ni nj '
=

(7.56)

i,j

!
i(=j

=
=

"

"

Vi Vj g(ri , rj )
n(ri )
n(rj ) [1 + &(ri )(rj )'] +

d3 r
d3 r

"

"

d3 r! g(r, r! )
n(r)
n(r! ) [1 + &(r)(r! )'] +

Vi g(ri , ri )
n(ri )

i=j

"

d3 rg(r, r)
n(r)

I
J
d3 r! g(r, r! ) n
(r)
n(r! ) [1 + &(r)(r! )'] + n
(r) D (r r! ) .
222

As this equation holds for an arbitrary function g(r, r! ), comparing the first and last line
of the equation above, we find the relation between the two point correlator of the discrete
number density and that of underlying density contrast:
&n(r)n(r! )' = n
(r)
n(r! ) [1 + &(r)(r! )'] + n
(r) D (r r! ).

(7.57)

The 2nd term in Equation (7.57) is called a Poisson shot noise, as the Dirac delta function
manifests its identity as a self-particle contribution. The shot noise term n
(r) D (r r ! ) can

also be understood as follows. Because the presence of galaxies obeys the Poisson statistics, for a single position r with the mean number density n
(r), its variance is 2 [n(r)]

&n(r)n(r)' n
2 (r) = n
(r).

By using equation (7.57), the correlation function of galaxy density contrast measured from the discrete samples of galaxies is given by
n (r r! )

&n (r)n (r! )'

G&
'&
'H
n(r) n
(r)
n(r! ) n
(r! )
=
n
(r)
n
(r! )
!
!
&n(r)n(r )' n
(r)
n(r )
=
n
(r)
n(r! )
D (r r! )
= &(r)(r! )' +
,
n
(r)

(7.58)

where we denote the density contrast of the discrete Poisson sample as

n (r)

n(r) n
(r)
.
n
(r)

The Fourier transform of Equation (7.58) yields following relation in Fourier space:
"
E
F E
F
1 ir(k+k! )
!
!
n (k)n (k ) = (k)(k ) + d3 r
e
.
(7.59)
n
(r)
For N-body simulation, as the mean number density over the simulation box are
constant, n
(r) n
; the equation (7.59) reduces to
E
F E
F (2)3 D !
(k + k).
n (k)n (k! ) = (k)(k! ) +
n

(7.60)

Therefore, we get the formula for Poisson shot noise.

Pn (k) = P (k) +

223

1
.
n

(7.61)

7.1.4.3

Bispectrum of discrete particles

Similarly, we calculate the bispectrum of discrete particles from the expectation

value of

"

d3 r1

"

d3 r2

"

d3 r3 g(r1 , r2 , r3 )n(r1 )n(r2 )n(r3 )

for an arbitrary function g(r1 , r2 , r3 ). By using the infinitesimal micro-cells, the expectation
value becomes
G"

H
"
"
d3 r1 d3 r2 d3 r3 g(r1 , r2 , r3 )n(r1 )n(r2 )n(r3 )
"
"
"
= d3 r1 d3 r2 d3 r3 g(r1 , r2 , r3 ) &n(r1 )n(r2 )n(r3 )'
!
=
g(ri , rj , rk ) &ni nj nk '
i,j,k

i(=j(=k

Vi Vj Vk gijk n
in
j n
k [1 + &i j ' + &j k ' + &k i ' + &i j k ']

i=j(=k

"

Vi Vk giik n
in
k [1 + &i k '] + (2 cyclic) +

Vi giii n
i

i=j=k

"
"
d3 r1 d3 r2 d3 r3 g(r1 , r2 , r3 )
*
n
(r1 )
n(r2 )
n(r3 ) [1 + &(r1 )(r2 )(r3 )' + &(r1 )(r2 )' + (2 cyclic)]
+n
(r1 )
n(r2 ) [1 + &(r1 )(r2 )'] D (r1 r3 ) + (2 cyclic)
K
+n
(r1 ) D (r1 r2 ) D (r1 r3 ) ,

(7.62)

where we use subscript to mark the coordinate in the fourth and fifth line. For example,
gijk g(ri , rj , rk ). Again, as the equation (7.62) has to hold for arbitrary g(r1 , r2 , r3 ), we

find

&n(r1 )n(r2 )n(r3 )'

=
n(r1 )
n(r2 )
n(r3 ) [1 + &(r1 )(r2 )(r3 )' + &(r1 )(r2 )' + (2 cyclic)]
+n
(r1 )
n(r2 ) [1 + &(r1 )(r2 )'] D (r1 r3 ) + (2 cyclic)
+n
(r1 ) D (r1 r2 ) D (r1 r3 ).

We calculate the three-point correlation function of discrete particles by

'&
'&
'H
G&
n(r2 )
n(r3 )
n(r1 )
1
1
1
&n (r1 )n (r2 )n (r3 )' =
n
(r1 )
n
(r2 )
n
(r3 )
&n(r1 )n(r2 )n(r3 )' &n(r1 )n(r2 )'
=

+ (2 cyclic) + 2.
n
(r1 )
n(r2 )
n(r3 )
n
(r1 )
n(r2 )

224

(7.63)

(7.64)

Using equation (7.57) and equation (7.63), we find

&n (r1 )n (r2 )n (r3 )'
= &(r1 )(r2 )(r3 )' +
+

D (r1 r3 )
&(r1 )(r2 )' + (2 cyclic)
n
(r3 )

1
D (r1 r2 ) D (r1 r3 ),
n
(r2 )
n(r3 )

(7.65)

and the Fourier transform of equation (7.65) yields,

&n (k1 )n (k2 )n (k3 )' = &(k1 )(k2 )(k3 )'
"
"
1
+ d3 r1 d3 r2
&(r1 )(r2 )' eir1 (k1 +k3 )ir2 k2
n
(r1 )
"
1
eir1 k123 .
(7.66)
+ d3 r1 2
n
(r1 )
Especially, when the mean number density does not vary in time, the three point function
in k space becomes
&n (k1 )n (k2 )n (k3 )'
= &(k1 )(k2 )(k3 )' +

1
(2)3 D
&(k13 )(k2 )' + (2 cyclic) +
(k123 ),
n

n
2

and, the bispectrum of discrete Poisson samples is reduced to

*
K
1
1
P (k1 ) + P (k2 ) + P (k3 ) + 2 .
Bn (k1 , k2 ), k3 ) = B(k1 , k2 , k3 ) +
n

7.2

(7.67)

(7.68)

Power spectrum from Galaxy surveys

For the real galaxy survey, we have to take into account the spatial variance of

the mean number density. In that case, we have to weight galaxies differently depending
on the mean number density in order to optimize the variance of power spectrum. The
most popular weighting is given from Feldman et al. (1994, FKP). In this section, we first
review the FKP estimator, and derive the optimal weighting function for power spectrum
measurement. Then, we show the practical implementation of FKP estimator for galaxy
surveys by using Discrete Fourier Transformation.

225

7.2.1
7.2.1.1

The FKP estimator

The power spectrum with weighting function w(r)
Let us denote the weighted overdensity F (r) as a weighted overdensity of a galaxies

at position r in the survey:

F (r) w(r) [n(r) n
(r)] = w(r)
n(r)n (r) W (r)n (r).

(7.69)

Here, n
(r) is the mean number density of galaxies expected at position r, and w(r) is a
weighting function which optimizes the variance of the estimated power spectrum. Combining the two effect of the selection function and the weighting defines the window function,
W (r).
By convolution theorem, the Fourier mode of the weighted density field becomes
"
d3 q
F (k) =
W (k q)n (q).
(7.70)
(2)3
The two point correlation function of the weighted density field is
"
"
d3 q
d3 q !
&F (k)F (k! )' =
W (k q)W (k! q! )&n (q)n (q! )'
3
(2)
(2)3
"
d3 q
=
W (k q)W (k! + q)P (q)
(2)3
"
"
"
d3 q !
1 ir(q+q! )
d3 q
!
!
+
W
(k

q)W
(k

q
)
d3 r
e
.
3
3
(2)
(2)
n
(r)
The second term is further simplifies to
"
"
"
d3 q
1 ir(q+q! )
d3 q !
!
!
W
(k

q)W
(k

q
)
d3 r
e
(2)3
(2)3
n
(r)
"
"
!
1 ir(k+k! )
=
d3 rW 2 (r)
e
= d3 rw2 (r)
n(r)eir(k+k ) .
n
(r)
Therefore, the power spectrum of the weighted overdensity can be calculated from
"
"
!
d3 q
!
W
(k

q)W
(k
+
q)P
(q)
+
d3 rw2 (r)
n(r)eir(k+k ) ,
(7.71)
&F (k)F (k! )' =
(2)3
that is,
&F (k)F (k)' =

"

d3 q
2
|W (k q)| P (q) +
(2)3

"

d3 rw2 (r)
n(r).

(7.72)

In Feldman, Kaiser & Peacock (1994, FKP), they first subtract the synthetic catalog, ns (r), generated by the number density of n
(r)/. Also, they divide the magnitude

226

of the window function

W

."

/1/2 ."
d3 rW 2 (r)
=

d3 k
2
|W (k)|
(2)3

/1/2

when defining the weighted density field. That is, the weighted density field in FKP is
F (r)

w(r) [n(r) ns (r)]

W (r)
=
[n (r) s (r)] ,
W
W

where
s (r)

(7.73)

ns (r) n
(r)/
.
n
(r)/

Assuming that the synthetic catalog is completely random, and independent of the galaxy
distribution, i.e.,
&n (r)s (r! )' = 0
&s (r)s (r! )' =

(7.74)
D

(r r )
,
n
(r)

the power spectrum of the weighted density field of FKP is

"
1
3 " d3 q |W (k q)|2
1+
2
P (k) |F (k)| =
P
(q)
+
d3 rw2 (r)
n(r).
(2)3
W2
W2

(7.75)

(7.76)

Note that in FKP, they refer W (k)/W as G(k):

L 3
d r
n(r)w(r)eikr
G(k) ML
N1/2 .
d3 r
n2 (r)w2 (r)

If survey has a typical size of D ( V 1/3 , then the window function W (k)/W will be

a rather compact function with width k 1/D. Therefore, for high enough wavenumber,
|k| , 1/D, we can approximate the power spectrum of weighted overdensity as
P (k) ( P (k) + Pshot
provided that P (k) is a smooth function of k. Here,
L
(1 + ) d3 r
n(r)w2 (r)
L
Pshot
.
3
2
d r
n (r)w2 (r)

is the shot noise power spectrum in the presence of the weighting function. We estimate the
angular averaged (monopole) power spectrum estimation by
"
=
>
1
P (k)
d3 q P (q) Pshot ,
Vk Vk
where Vk is the volume of the shell in Fourier space.

227

(7.77)

7.2.2

The variance of the power spectrum

We estimate the variance of the power spectrum by
G=
"
"
>2 H
1
3
1
2
3

p P (k) P (k)
= 2
d q
d3 q ! P (q) P (q! )
Vk Vk
Vk

(7.78)

with P (q) = P (q) P (q). If F (k) obeys Gaussian statistics, then

1
3
2
P (q) P (q! ) = |&F (q)F (q! )'|

With Equation (7.71), we write the right hand side as

&F (q)F (q! )'

"
"
!
d3 p W (q p)W (q! p)
1+
=
P
(p)
+
d3 rw2 (r)
n(r)eir(qq ) .
(2)3
W2
W2

(7.79)

In FKP paper, the second term is referred as S(k):

L
(1 + ) d3 rw2 (r)
n(r)eikr
L
S(k)
3
2
2
d rw (r)
n (r)

As Equation (7.78) integrate over the same spherical shell, we consider the case when qq! =
q. For |q| , 1/D, we approximate Equation (7.79) as
&F (q)F (q! )'
"
"
!
P (q)
d3 p
1+
!
( 2
W
(q

p)W
(q

p)
+
d3 rw2 (r)
n(r)eir(qq ) ,
W
(2)3
W2

(7.80)

where the first term becomes

"
d3 p
W (q p)W (q! p)
(2)3
"
"
"
!
!
d3 p
3
=
d
r
d3 r! W (r)W (r! )eir(qp) eir (q p)
(2)3
"
"
! !
=
d3 r d3 r! W (r)W (r! ) D (r r! )eirq eir q
"
!
=
d3 rW 2 (r)eir(qq ) .

In FKP, this function is called Q(k):

Q(k)

d3 rw2 (r)
n2 (r)eikr
L
.
d3 rw2 (r)
n2 (r)

Therefore, the variance of the power spectrum is

1
3
2
P (q) P (q! )
= |&F (q)F (q! )'|

= |P (q)Q(q) + S(q)|2 .
228

(7.81)

(7.82)

7.2.3

Optimal weighting
When the spherical shell is larger than the coherence length ( 1/D), the double

integration in Equation (7.78) reduces to

"
1
2
p2 (k) =
d3 q |P (k)Q(q) + S(q)| .
Vk
Therefore, the fractional variance of the power is
22
2
"
2
p2 (k)
1
S(q) 22
3 2
=
d q 2Q(q) +
P 2 (k)
Vk
P (k) 2
22
2"
'
&
"
2
2
1
1+
iqr 2
3 2
3
2
2
e
=
d
q
d
rw
(r)
n
(r)
1
+
2 .
2
Vk W 2
n
(r)P (k)
Using Parsevals theorem3 , the equation is further simplified as
'2
&
"
p2 (k)
(2)3
1+
3
4
4
.
=
d
rw
(r)
n
(r)
1
+
P 2 (k)
Vk W 2
n
(r)P (k)

(7.83)

(7.84)

For optimal weighting, we choose the weighting function w(r) which minimizes the variance
in Equation (7.84). First, lets abbreviate the equation as
L 3 4
p2 (k)
d rw (r)f (r)
= ML
N2
P 2 (k)
d3 rw2 (r)g(r)

That is,

f (r) =
g(r) =

&
n
4 (r) 1 +
n
2 (r).

1+
n
(r)P (k)

'2

When we take the variation of w(r) = w0 (r) + w(r), Equation (7.84) becomes
L 3 4
p2 (k)
d rw0 [1 + 4w/w0 ] f
= ML
N2
P 2 (k)
d3 rw2 (1 + 3w/w0 ) g
L 3 04
L
L 3
*
& 3 3
'K
d rw0 f
d rw0 wf
d rw0 wg
L
L
= ML
1
+
4
+

N
2
d3 rw04 f
d3 rw02 g
d3 rw02 g
3 The

proof of Parcevals theorem is given as following.

"!
"2
!
!
!
!
"
"
!
d3 q "" d3 rf (r)eiqr ""
=
d3 q d3 rf (r)eiqr d3 r " f (r" )eiqr
!
!
!
= (2)3 d3 r d3 r " f (r)f (r" )D (r r" ) = (2)3 d3 rf 2 (r)

229

(7.85)

Therefore, the optimal weighting function has to satisfy

L 3 3
L 3
d rw0 wf
d rw0 wg
L
= L 3 2 ,
d3 rw04 f
d rw0 g

(7.86)

whose solution is given by

w02 g/f.

(7.87)

In terms of the mean number density and the power spectrum, the optimal weighting has
to satisfy
w(r)

1
.
n
(r) + (1 + )/P (k)

(7.88)

Finally, the dimensionless optimal weighting function is

w(r) =

1
.
1 + + P (k)
n(r)

(7.89)

When we choose a large number of synthetic sample ( . 1), we recover the result

of the FKP optimal weighting function:

w(r) =

7.3

1
.
1 + P (k)
n(r)

(7.90)

Implementing the FKP estimator

In this section, we show the Discrete Fourier Transform implementation of the FKP

estimator. Let us consider the case when the mean number density depends on the position:
n
(r). In the previous section, we have shown that the optimal weighting function is given
by
w(r) =

1
.
1+n
(r)P (k)

(7.91)

We define the weighted overdensity in the discrete grid point nr as

F (nr ) =
=

(nr )
w(nr )
w(nr )N
[Ng (nr ) Ns (nr )] =
[n (nr ) s (nr )]
W
W
W (nr )H 3
[n (nr ) s (nr )] .
W

(7.92)

Here, Ni (nr ) denote the number density of galaxies (i = g), and synthetic random samples
(i = s) assigned to the grid nr ,
Ni (nr )

"

nr

d3 rni (r)W s (r nr ),

230

(7.93)

where W s (x) is the sampling window function we discussed in Section 7.1.1. Similarly,
(nr ) is the mean number density of galaxies assigned to the grid nr :
N
"
(nr )
N
d3 r
n(r)W s (r nr ).
(7.94)
nr

As synthetic random samples are generated from the mean number density rescaled by 1/,
they are related by
(nr ) = &Ng (nr )' = &Ns (nr )' .
N

(7.95)

Also we define the survey window function on the grid point nr as

W (nr )

(nr )
w(nr )N
.
H3

(7.96)

Fourier transform the weighted overdensity above yields4

FDF T (nk )
!
H3 !
F (nr )ei2nk nr /N =
=
W (nr ) [n (nr ) s (nr )] ei2nk nr /N .
W n
n
r

(7.97)

That is, the Fourier transform is given by the convolution of W (nk ) and n (nk ) s (nk ).
Let us be explicit about the convolution in DFT. For discrete sampling of A(nr ) and B(nr ),

the Fourier transform of its multiplication C(nr ) A(nr )B(nr ) is given by

CDF T (nk )
!
A(nr )B(nr )ei2nk nr /N

nr

!
!
1
1

A(np )ei2np nr /N
B(nq )ei2nq nr /N ei2nk nr /N
=
V n
V n
nr
p
q
;
<
!
!
!
1
i2(np +nq )nr /N i2nk nr /N
= 2
A(np )B(nq )
e
e
V n n
n
!

N3 ! !
A(np )B(nq )nk mN,np +nq
= 2
V n n
p

N !
= 2
A((nk nq )N )B(nq ).
V n

(7.98)

4 For

the normalization of the Discrete Fourier Transform, see Appendix A.

231

Here, m can be any integer, and A((nk )N ) denotes the integer triplet modulated by N .
Using the convolution in DFT, we find the DTF of F becomes
FDF T (nk ) =

1 !
W ((nk nq )N ) [n (nq ) s (nq )] .
WV n

(7.99)

Note that when W (nr ) w

n (constant), then,
W (nk ) = H 3

w
nei2nk nr /N = H 3 w
nN 3 nk ,0

nr

therefore, we reproduce the normalization of the previous section:

FDF T (nk ) =

w
n
[n (nk ) s (nk )] .
W

We estimate the power spectrum by calculating the ensemble average of

&FDF T (nk )FDF T (nk! )'

1 !!
W ((nk nq )N )W ((nk! nq! )N )
W 2V 2 n n
q

q!

&[n (nq ) s (nq )] [n (nq! ) s (nq! )]' .

(7.100)

First, let us evaluate the two point correlation function without the window function for the
discrete grid

&n (nq )n (nq! )'

!!
H6
&n (nr )n (n!r )' ei2nr nq /N ei2nr! nq! /N
nr nr!

!!
nr nr!

&(nr )(nr! )' +

&(nq )(nq! )' + H 3

!
nr

nKr ,nr!

H 3n
(nr )

<

ei2nr nq /N ei2nr! nq! /N

1
ei2nr (nq +nq! )/N ,
n
(nr )

(7.101)

where in the second line, we changes the Dirac delta function to the Kronecker delta by
explicitly factoring out the dimensionality (1/H 3 ). From the same procedure, we can also
calculate
&n (nq )s (nq! )'
&s (nq )s (nq! )'

= 0,
= H3

(7.102)
!
nr

232

ei2nr (nq +nq! )/N .

n
(nr )

(7.103)

Putting the result above all together, we finally have

&FDF T (nk )FDF T (nk! )'
1 !!
= 2 2
W ((nk nq )N )W ((nk! nq! )N ) &(nq )(nq! )'
W V n n
q

q!

! 1
1 + !!
+H
ei2nr (nq +nq! )/N .
W ((nk nq )N )W ((nk! nq! )N )
2
2
W V n n
n

(n
)
r
n
3

q!

(7.104)

The first term is the Window function convolved galaxy power spectrum:
1

!!

W 2V 2

nq nq!

W ((nk nq )N )W ((nk! nq! )N )

(2)3
P (nq )nDq ,nq!
kF 1 kF 2 kF 3

1 !
W ((nk nq )N )W ((nk! + nq )N )P (nq ),
W 2V n

(7.105)

and the second term is the shot-noise term:

H9

1 + !!!
W (nr1 )ei2nr1 (nk nq )/N
W 2V 2 n n n
q

q!

W (nr2 )e

r1

i2nr2 (nk! nq! )/N

nr2

!
nr

1
ei2nr (nq +nq! )/N
n
(nr )

1 + ! ! ! W (nr1 )W (nr2 ) i2(nr1 nk +nr2 nk! )/N 3 D

=H
e
N nr1 ,nr N 3 nDr2 ,nr
W 2V 2 n n n
n
(nr )
9

r1

r2

1 + ! W 2 (nr ) i2nr (nk +nk! )/N

=N 6 H 9 2 2
e
W V n n
(nr )
r

=H

+ ! W 2 (nr ) i2nr (nk +nk! )/N

e
.
W2 n n
(nr )

31

(7.106)

Adding up the results, we find that

1

|FDF T (nk )|

1 !
1 + ! W 2 (nr )
2
|W ((nk nq )N )| P (nq ) + H 3
.
2
W V n
W2 n n
(nr )
q

(7.107)

The normalization factor W can be calculated as

<1/2
."
/1/2 ;! 2
2 (nr )
w (nr )N
3
2
.
W
d rW (r)
=
H3
n

(7.108)

Let us check the limiting case when n

(nr ) = n
. For that case, the weighting function

233

is also a constant, w(nr ) = w. Then, the normalization factor becomes

W =

;
!

2 2

w n
H

nr

<1/2

M
N1/2
= N 3 w2 n
2H 3
= V w
n,

and power spectrum can be estimated by

3
1
2
|FDF T (nk )|
=

2 w2 n
2

V 2 w2 n
2 nk ,nq P (nq ) + H 3

nq

(7.109)

1 + ! w2 n
2
2
2
Vw n
n
n

1+
,
= P (nk ) +
n

(7.110)

which are what we expected from the calculation of the previous section.
Finally, we have to correct for the window function due to the number density
distribution, by following the method described in Section 7.1.3.
7.3.1

The estimator
In summary, we estimate the power spectrum as follows.

7.3.1.1

Constant weighting
When we do not employ the weighting function, first calculate
F (nr ) = Ng (nr ) Ns (nr )

(7.111)

and Fourier transform it. Then, the square of the Fourier transform becomes
3
1 1
2
|F
(n
)|
DF
T
k
W2
1 + ! W 2 (nr )
1 !
|W (nk nq )|2 P (nq ) + H 3
,
= 2
W V n
W2 n n
(nr )
q

(7.112)

where the window function is given by

W (nr ) =

(nr )
N
Ns (nr )
(
,
H3
H3

(7.113)

and its Fourier transform is

W (nk ) = H 3

W (nr )ei2nk nr /N .

nr

234

(7.114)

As we distribute particle numbers to regular grid points, FDF T (nk ) has to be deconvolved
by the method we described in 7.1.3.
The normalization factor W 2 is given by
W2 =

1 ! 2
2 ! 2
N (nr ) ( 3
Ns (nr ).
3
H n
H n
r

(7.115)

(nr ) by
Note that we can approximate the average number of galaxies N
"
(nr )
N
d3 r
n(r) ( Ns (nr ).

(7.116)

H3

With this approximation, and n

(nr ) ( Ns (nr )/H 3 the shot noise term may be approximated as

H3

1 + ! W 2 (nr )
( H3
W2 n n
(nr )
r

7.3.1.2

&

1+

':
N (n )
: s2 r .
Ns (nr )

(7.117)

FKP optimal weighting

When estimating power spectrum with a optimal weighting function
w(nr ) =

1
1
H3
(
= 3
,
3
1+n
(r)P (k)
1 + Ns (r)P (k)/H
H + Ns (r)P (k)

(7.118)

first calculate
F (nr ) = w(nr ) [Ng (nr ) Ns (nr )]

(7.119)

and Fourier transform it. Then, the square of the Fourier transform becomes
3
1 1
2
|FDF T (nk )|
2
W
1 !
1 + ! W 2 (nr )
2
= 2
|W (nk nq )| P (nq ) + H 3
,
W V n
W2 n n
(nr )
q

(7.120)

where the window function is given by

W (nr ) =

(nr )
w(nr )N
w(nr )Ns (nr )
Ns (nr )
(
= 3
,
3
3
H
H
H + Ns (nr )P (k)

(7.121)

and its Fourier transform is

W (nk ) = H 3

W (nr )ei2nk nr /N .

nr

235

(7.122)

Then, according to the number distribution scheme, FDF T (nk ) has to be deconvolved by
the method we described in 7.1.3. The normalization factor W 2 is
/2
!.
1 ! 2
Ns (nr )
2 (nr ) ( H 3
W2 = 3
w (nr )N
,
H n
H 3 + Ns (nr )P (k)
n
r

(7.123)

and the shot noise term may be approximated as

1 + ! W 2 (nr )
W2 n n
(nr )
r
8
98
91
&
' !
!
1+
Ns (nr )
Ns2 (nr )
3
(H
.

(H 3 + Ns (nr )P (k))2
(H 3 + Ns (nr )P (k))2
n
n
H3

236

(7.124)