8 Structure Formation: 8.1 Inhomogeneity

8 Structure Formation
Up to this point we have discussed the universe in terms of a homogeneous and isotropic model
(which we shall now refer to as the “unperturbed” or the “background” universe). Clearly
the universe is today rather inhomogeneous. By structure formation we mean the generation
and evolution of this inhomogeneity. We are here interested in distance scales from galaxy size
to the size of the whole observable universe. The structure is manifested in the existence of
galaxies and in their uneven distribution, their clustering. This is the obvious inhomogeneity,
but we understand it reflects a density inhomogeneity also in other, nonluminous, components
of the universe, especially the cold dark matter. The structure has formed by gravitational
amplification of a small primordial inhomogeneity. There are thus two parts to the theory of
structure formation:
1) The generation of this primordial inhomogeneity, “the seeds of galaxies”. This is the more
speculative part of structure formation theory. We cannot claim that we know how this
primordial inhomogeneity came about, but we have a good candidate scenario, inflation,
whose predictions agree with the present observational data, and can be tested more
thoroughly by future observations. In inflation, the structure originates from quantum
fluctuations of the inflaton field ϕ near the time the scale in question exits the horizon.
2) The growth of this small inhomogeneity into the present observable structure of the uni-
verse. This part is less speculative, since we have a well established theory of gravity,
general relativity. However, there is uncertainty in this part too, since we do not know the
precise nature of the dominant components to the energy density of the universe, the dark
matter and the dark energy. The gravitational growth depends on the equations of state
and the streaming lengths (particle mean free path between interactions) of these density
components. Besides gravity, the growth is affected by pressure forces.
We shall do the second part first. But before that we discuss statistical measures of inho-
mogeneity: correlation functions and power spectra.
8.1 Inhomogeneity
We write all our inhomogeneous quantities as a sum of a homogeneous background value, and
a perturbation, the deviation from the background value. For example, for energy density and
pressure we write
ρ(t, x) = ρ̄(t) + δρ(t, x)

p(t, x) = p̄(t) + δp(t, x) , (1)
where ρ̄ and p̄ are the background density and pressure, x is the comoving 3D space coordinate,
and δρ and δp are the density and pressure perturbations. We further define the relative density
perturbation
δρ(t, x)
δ(t, x) ≡ . (2)
ρ̄(t)
Since ρ ≥ 0, necessarily δ ≥ −1. These quantities can be defined separately for different
components to the energy density, e.g., matter, radiation, and dark energy. Perturbations in
dark energy are expected to be small, and if it is just vacuum energy, it has no perturbations.
When we discuss the later history of the universe, the main interest is in the matter density
perturbation,
δρm (t, x)
δm (t, x) ≡ , (3)
ρ̄m (t)
25
8 STRUCTURE FORMATION 26
and then we will often write just δ for δm .

We do the split into the background and perturbation so that the background is equal to the
mean (volume average) of the full quantity. An important question is, whether the ρ̄(t) and p̄(t)
defined this way correspond to a (homogeneous and isotropic) solution of General Relativity,
i.e., an FRW universe. We expect the exact answer to be negative, since GR is a nonlinear
theory, so that perturbations affect the evolution of the mean. This effect is called backreaction.
However, if the perturbations are small, we can make an approximation, where we drop from
our equations all those terms which contain a product of two or more perturbations, as these
are “higher-order” small. The resulting approximate theory is called first-order perturbation
theory or linear perturbation theory. As the second name implies, the theory is now linear in
the perturbations, meaning that the effect of overdensities cancel the effect of underdensities on,
e.g., the average expansion rate. In this case the mean values evolve just like they would in the
absence of perturbations.
While the perturbations at large scales have remained small, during the later history of
the universe the perturbations have grown large at smaller scales. How big is the effect of
backreaction, is an open research question in cosmology, since the calculations are difficult, but
a common view is that the effect is small compared to the present accuracy of observations.
For this course, we adopt this view, and assume that the background universe simultaneously
represents an FRW universe (“the universe we would have if we did not have the perturbations”)
and the mean values of the quantities in the true universe at each time t.
Moreover, in Cosmology II we shall (mostly) assume that the background universe is flat
(K = 0).
8.1.1 Statistical homogeneity and isotropy

We assume that the origin of the perturbations is some random process in the early universe.
Thus over- (δ > 0) and underdensities (δ < 0) occur at randomly determined locations and we
cannot expect to theoretically predict the values of δ(t, x) for particular locations x. Instead,
we can expect theory to predict statistical properties of the inhomogeneity field δ(t, x). The
statistical properties are typically defined as averages of some quantities. We will deal with two
kind of averages: volume average and ensemble average; the ensemble average is a theoretical
concept, whereas the volume average is more observationally oriented.
We denote the volume average of some quantity f (x) with the overbar, f¯, and it is defined
as Z
¯ 1
f≡ d3 xf (x) . (4)
V V
The integration volume V in question will depend on the situation.
For the ensemble average we assume that our universe is just one of an ensemble of an infinite
number of possible universes (realizations of the random process) that could have resulted from
the random process producing the initial perturbations. To know the random process, means to
know the probability distribution Prob(γ) of the quantities γ produced by it. (At this stage we
use the abstract notation of γ to denote the infinite number of these quantities. They could be
the generated initial density perturbations at all locations, δ(x), or the corresponding Fourier
coefficients δk . We will be more explicit later.) The ensemble average of a quantity f depending
on these quantities γ as f (γ) is denoted by hf i and defined as the (possibly infinite-dimensional)
integral Z
hf i ≡ dγProb(γ)f (γ) . (5)
Here f could be, e.g., the value of ρ(x) at some location x. The ensemble average is also called
the expectation value. Thus the ensemble represents a probability distribution of universes. A
cosmological theory predicts such a probability distribution, but it does not predict in which
realization from this distribution we live in. Thus the theoretical properties of the universe
we will discuss (e.g., statistical homogeneity and isotropy, and ergodicity, see below) will be
properties of this ensemble.
We now make the assumption that, although the universe is inhomogeneous, it is statistically
homogeneous and isotropic. This is the second version of the Cosmological Principle. Statistical
homogeneity means that the expectation value hf (x)i must be the same at all x, and thus we
can write it as hf i. Statistical isotropy means that for quantities which involve a direction,
the statistical properties are independent of the direction. For example, for vector quantities
v, all directions must be equally probable. This implies that hvi = 0. The assumption of
statistical homogeneity and isotropy is justified by inflation: inflation makes the background
universe homogeneous and isotropic so that the external conditions for quantum fluctuations
are everywhere the same.
If the theoretical properties of the universe are those of an ensemble, and we can only
observe one universe from that ensemble, how can we compare theory and observation? It
seems reasonable that the statistics we get by comparing different parts of the universe should
be similar to the statistics of a given part of the universe over different realizations, i.e., that
they provide a fair sample of the probability distribution. This is called ergodicity. Fields f (x)
that satisfy
f¯ = hf i (6)
for an infinite volume V (for f¯) and an arbitrary location x (for hf i) are called ergodic. We
assume that cosmological perturbations are ergodic. The equality does not hold for a finite
volume V ; the difference is called sample variance or cosmic variance. The larger the volume,
the smaller is the difference. Since cosmological theory predicts hf i, whereas observations probe
f¯ for a limited volume, cosmic variance limits how accurately we can compare theory with
observations.1
8.1.2 Density autocorrelation function

From ergodicity,
hρi = ρ̄ ⇒ hδρi = 0 and hδi = 0 . (7)
Thus we cannot use hδi as a measure of the inhomogeneity. Instead we can use the square of δ,
which is necessarily nonnegative everywhere, so it cannot average out like δ did. Its expectation
value
hδρ2 i
hδ2 i = (8)
ρ̄2
is the variance of the density perturbation, and the square root of the variance,
p
δrms ≡ hδ2 i (9)
the root-mean-square (rms) density perturbation, is a typical expected absolute value of δ at an
arbitrary location.2 It tells us about how strong the inhomogeneity is, but nothing about the
shapes or sizes of the inhomogeneities. To get more information, we introduce the correlation
function ξ.
We define the density (2-point) autocorrelation function (often called just correlation func-
tion) as
ξ(x1 , x2 ) ≡ hδ(x1 )δ(x2 )i . (10)
1
Another notation I will use for volume average is fb, for smaller volumes, e.g., the volume observed in a galaxy
survey. I try to reserve f¯ for situations where we can assume f¯ = hf i, whereas cosmic variance is the difference
between fb and hf i.
2
In other words, δrms is the standard deviation of ρ/ρ̄.
It is positive if the density perturbation is expected to have the same sign at both x1 and x2 ,
and negative for an overdensity at one and underdensity at the other. Thus it probes how
density perturbations at different locations are correlated with each other. Due to statistical
homogeneity, ξ(x1 , x2 ) can only depend on the difference r ≡ x2 − x1 , so we redefine ξ as
ξ(r) ≡ hδ(x)δ(x + r)i . (11)
From statistical isotropy, ξ(r) is independent of direction, i.e., spherically symmetric (isotropic),
ξ(r) = ξ(r) . (12)
We will have use for both the 3D, ξ(r), and 1D, ξ(r), versions. The correlation function is large
and positive for r smaller than the size of a typical over- or underdense region, and becomes
small for larger distances.
The correlation function at zero separation gives the variance of the density perturbation,
hδ2 i ≡ hδ(x)δ(x)i ≡ ξ(0) . (13)

b for a single realization as a volume average,
We can also define a correlation function ξ(r)
Z
b 1
ξ(r) ≡ d3 x δ(x)δ(x + r) . (14)
V
Integrating over r and assuming periodic boundary conditions3 we get the integral constraint
Z Z Z Z
3 b 1 3 3 1
d r ξ(r) = d rd x δ(x)δ(x + r) = d x δ(x) d3 r δ(x + r) = 0 ,
3
(16)
V V
b
since the latter integral is δ̄ = 0. Since ξ(r) = hξ(r)i the integral constraint applies to it likewise.
Therefore ξ(r) must become negative at some point, so that at such a distance from an overdense
region we are more likely to find an underdense region. Going to ever larger distances, ξ as a
function of r may oscillate around zero, the oscillation becoming ever smaller in amplitude. Most
of the interest in ξ(r) is for the small r within the initial positive region.
8.1.3 Fourier space

The evolution of perturbations is best discussed in Fourier space. For mathematical convenience,
we assume the observable part of the universe lies within a fiducial cubic box, volume V = L3 ,
with periodic boundary conditions. This box is assumed to be much larger than the region of
interest, so that these boundary conditions should have no effect. Since the infinite universe is
now periodic, the volume average over the infinite universe will be equal to the volume average
over the fiducial box. Thus also the ergodicity assumption requires the fiducial volume to be
large, so that it can provide a fair sample of the ensemble. We can now expand any function of
space f (x) as a Fourier series X
f (x) = fk eik·x , (17)
k
3
The other option is not to use periodic boundary conditions but to understand the integral in (14) to go over
only those x, for which both x and x + r ∈ V . This is what we have to do when V refers to an actual survey.
The double integral in (16) then goes over all pairs (x, y) in V and can be written as
Z Z
1
d3 x δ(x) d3 y δ(y) = 0 · 0 . (15)
V V V
where the wave vectors k = (k1 , k2 , k3 ) take values

2π
ki = ni , ni = 0, ±1, ±2, . . . (18)
L
The Fourier coefficients fk are obtained as
Z
1
fk = f (x)e−ik·x d3 x . (19)
V V
If f (x) is a perturbation so that its mean value vanishes, then the term k = 0 does not occur.
The Fourier coefficients are complex numbers even though we are dealing with real quantities
f (x). From the reality of f (x) follows that
f−k = fk∗ . (20)
The Fourier expansion works only if the background universe is flat, although it can be used
as an approximation in open and closed universes,4 if the region of interest is much smaller than
the curvature radius.
The separation of neighboring ki values is ∆ki = 2π/L, so we can write
X 3 Z
ik·x L 1
f (x) = fk e ∆k1 ∆k2 ∆k3 ≈ f (k)eik·x d3 k , (21)
2π (2π)3
k
where
f (k) ≡ L3 fk . (22)
replacing the Fourier series with the Fourier integral. The size of the Fourier coefficients depends
on the fiducial volume V – increasing V tends to make the fk smaller to compensate for the
denser sampling of k in Fourier space.
In the limit V → ∞, the approximation in (21) becomes exact, and we have the Fourier
transform pair
Z
1
f (x) = f (k)eik·x d3 k
(2π)3
Z
f (k) = f (x)e−ik·x d3 x . (23)
Note that this assumes that the integrals converge, which requires that f (x) → 0 for |x| → ∞.
Thus we use only the Fourier series for, e.g., δ(x), but for, e.g., the correlation function ξ(x) the
Fourier transform is appropriate.
Even with a finite V we can use the Fourier integral as an approximation. Often it is
conceptually simpler to work first with the Fourier series (so that one can, e.g., use the Kronecker
delta δkk′ instead of the Dirac delta function δD (k − k′ )), replacing it with the integral in the
end, when it needs to be calculated. The recipe for going from the series to the integral is
3 X Z
2π
→ d3 k
L
k
L3 f k → f (k) (24)
3
L 3
δkk′ → δD (k − k′ ) .
2π
so that, e.g., Z
X 1
ik·x
fk e → f (k)eik·x d3 k . (25)
(2π)3
k
4
An exact treatment in open and closed universes requires expansion in terms of suitable other functions
instead of the plane waves eik·x .
8.1.4 Power spectrum

We now expand the density perturbation as a Fourier series
X
δ(x) = δk eik·x , (26)
k
with Z
1
δk = δ(x)e−ik·x d3 x (27)
V V
and δ−k = δk∗ . Note that
hδ(x)i = 0 ⇒ hδk i = 0 . (28)
In analogy with the correlation function ξ(x, x′ ), we may ask what is the corresponding
correlation in Fourier space, hδk∗ δk′ i. Note that due to the mathematics of complex numbers,
correlations of Fourier coefficients are defined with the complex conjugate ∗ . This way the
correlation of δk with itself, hδk∗ δk i = h|δk |2 i is a real (and nonnegative) quantity, the expectation
value of the absolute value (modulus) of δk squared, i.e., the variance of δk . Calculating
Z Z
∗ 1 3 ik·x ′ ′
hδk δk′ i = 2
d xe d3 x′ e−ik ·x hδ(x)δ(x′ )i
V
Z Z
1 3 ik·x ′
= 2
d xe d3 re−ik ·(x+r) hδ(x)δ(x + r)i
V
Z Z
1 3 −ik′ ·r ′
= 2
d re ξ(r) d3 xei(k−k )·x
V
Z
1 1
= δkk′ d3 re−ik·r ξ(r) ≡ δkk′ P (k) , (29)
V V
where we used hδ(x)δ(x + r)i = ξ(r), i.e., independent of x, which results from statistical
homogeneity, and the orthogonality of plane waves
Z
′
d3 xei(k−k )·x = V δkk′ → (2π)3 δD
3
(k − k′ ) . (30)
Note that here δkk′ is the Kronecker delta, 1 for k = k′ , 0 otherwise – nothing to do with the
density perturbation! In the limit V → ∞ we get the Dirac delta function δD 3 (k − k′ ).
Written in terms of δ(k) = V δk , the result (29) reads as
hδ(k)∗ δ(k′ )i = V δkk′ P (k) → (2π)3 δD

3
(k − k′ )P (k) , (31)
Thus, from statistical homogeneity follows that the Fourier coefficients δk are uncorrelated.
The quantity Z
2
P (k) ≡ V h|δk | i = d3 r e−ik·r ξ(r) , (32)
which gives the variance of δk , is called the power spectrum of δ(x). Since the correlation function
→ 0 for large distances, we can replace the integration volume V in (32) with an infinite volume.
We see that the power spectrum is the 3D Fourier transform of ξ(r), and therefore also
Z
1
ξ(r) = d3 k eik·r P (k) . (33)
(2π)3
Unlike the correlation function, the power spectrum P (k) is positive everywhere. Perturbations
at large distance scales are more commonly discussed in terms of P (k) than ξ(r).
From statistical isotropy
ξ(r) = ξ(r) ⇒ P (k) = P (k) (34)
(the 3D Fourier transform of a spherically symmetric function is also spherically symmetric),

so that the variance of δk depends only on the magnitude k of the wave vector k, i.e., on
the corresponding distance scale. Using spherical coordinates and doing the angular integrals
we obtain (exercise) the relation between the 1D correlation function ξ(r) and the 1D power
spectrum P (k),
Z ∞
sin kr
P (k) = ξ(r) 4πr 2 dr
0 kr
Z ∞
1 sin kr
ξ(r) = 3
P (k) 4πk2 dk , (35)
(2π) 0 kr
For the density variance we get
Z ∞ Z ∞ Z ∞
2 1 2 1 3
hδ i ≡ ξ(0) = P (k)4πk dk = k P (k)d ln k ≡ P(k)d ln k . (36)
(2π)3 0 2π 2 0 −∞
where we have defined

k3
P(k) ≡ P (k) . (37)
2π 2
Another common notation for P(k) is ∆2 (k). The word “power spectrum” is used to refer to
both P (k) and P(k). Of these two, P(k) has the more obvious physical meaning: it gives the
contribution of a logarithmic interval of scales, i.e., from k to ek, to the density variance. P(k)
is dimensionless, whereas P (k) has the dimension of Mpc3 (when discussing observed values, it
is usually given in units of h−3 Mpc3 as distance determinations are proportional to the Hubble
constant).
b as the volume average of δ(x)δ(x + r), i.e., integrate x over the box V with
Exercise: Define ξ(r)
periodic boundary conditions, and show that
Z
b V
ξ(r) = d3 k |δk |2 eik·r , (38)
(2π)3
for a single realization. Note that here we do not need any statistical assumptions (like statistical
homogeneity or ergodicity). Contrast this result with (33).
8.1.5 Scales of interest and window functions

In (36) we integrated over all scales, from the infinitely large (k = 0 and ln k = −∞) to the
infinitely small (k = ∞ and ln k = ∞) to get the density variance. Perhaps this is not really
what we want. The average matter density today is 3 × 10−27 kg/m3 . The density of the Earth
is 5.5 × 103 kg/m3 and that of an atomic nucleus 2 × 1017 kg/m3 , corresponding to δ ≈ 2 × 1030
and δ ≈ 1044 . Probing the density of the universe at such small scales finds a huge variance in
it, but this is no longer the topic of cosmology - we are not interested here in planetary science
or nuclear physics.
Even the study of the structure of individual galaxies is not considered to belong to cos-
mology, so the smallest (comoving) scale of cosmological interest, at least when we discuss the
present universe,5 is that of a typical separation between neighboring galaxies, of the order of
1 Mpc.
5
In early universe cosmology we may study events, or possible events, related to also smaller comoving scales.
To exclude scales smaller than R (r < R or k > R−1 ) we filter the density field with a
window function. This can be done in k-space or x-space.
The filtering in x-space is done by convolution. We introduce a (usually spherically sym-
metric) window function W (r) such that
Z
d3 r W (r) = 1 (39)
(normalization) and W ∼ 0 for |r| ≫ R and define the filtered density field
Z
δ(x, R) ≡ (δ ∗ W )(x) ≡ d3 x′ δ(x′ )W (x′ − x) . (40)
The simplest window function is the top-hat window function

4π 3 −1
WT (r) ≡ R for |r| ≤ R (41)
3
and WT (r) = 0 elsewhere, i.e., δ(x) is filtered by replacing it with its mean value within the
distance R. Mathematically more convenient is the Gaussian window function
1 1 2 /R2
WG (r) ≡ e− 2 |r| . (42)
(2π)3/2 R3
By the convolution theorem, the filtering in Fourier space becomes just multiplication:
δ(k, R) = δ(k)W (k) , (43)
where W (k) is the Fourier transform of the window function. For WT and WG we have (exer-
cise)
3(sin kR − kR cos kR)

WT (k) =
(kR)3
1 2
WG (k) = e− 2 (kR) . (44)
We can also define the k-space top-hat window function
Wk (k) ≡ 1 for k ≤ 1/R (45)
and Wk (k) = 0 elsewhere. In x-space this becomes (exercise)

1 sin y − y cos y
Wk (r) = , where y ≡ |r|/R . (46)
2π 2 R3 y3
The variance of the filtered density field (Exercise: derive the second equalities of both
expressions)
Z
1
σ 2 (R) ≡ hδ(x, R)2 i = d3 k P (k)|W (k)|2
(2π)3
Z Z
1 V
b2 (R) ≡
σ d3 x δ(x, R)2 = d3 k |δk |2 |W (k)|2 . (47)
V (2π)3
is a measure of the inhomogeneity at scale R. For the k-space top-hat window this becomes
simply
Z R−1 Z − ln R
1
σ 2 (R) = 4πk 2
P (k)dk = P(k)d ln k . (48)
(2π)3 0 −∞
One may also ask, whether scales larger than the observed universe (the lower limit k = 0
or ln k = −∞ in the k integrals) are relevant, since we cannot observe the inhomogeneity at
such scales. Due to such very-large-scale inhomogeneities, the average density in the observed
universe may deviate from the average density of the entire universe. Inhomogeneities at scales
somewhat larger than the observed universe could appear as an anisotropy in the observed
universe. The importance of such large scales depends on how strong the inhomogeneities at
these scales are, i.e., how the power spectrum behaves as k → 0. The present understanding,
supported by observations, is that the contribution of such large scales is small.
8.1.6 Power-law spectra

We have observational information and theoretical predictions for ξ(r) and P (k) for a wide
range of scales. (We will discuss the theory in detail later.) For certain intervals, they can be
approximated by a power-law form,
ξ(r) ∝ r −γ or P (k) ∝ kn . (49)
When plotted on a log-log scale, such functions appear as straight lines with slope −γ and n.
The proportionality constant can be given in terms of a reference scale. For ξ(r) we usually
choose the scale r0 where ξ(r0 ) = 1, so that
−γ
r
ξ(r) = . (50)
r0
For P (k) we may write

n n+3
2 k 2 k
P (k) = A or P(k) = A , (51)
kp kp
p
where
p kp is called a pivot scale (whose choice depends on the application) and A ≡ P (kp ) or
P(kp ) is the amplitude of the power spectrum at the pivot scale.
We define the spectral index n(k) as
d ln P
n(k) ≡ . (52)
d ln k
It gives the slope of P (k) on a log-log plot. For a power-law P (k), n(k) = const = n. We can
study power-law ξ(r) and P (k) as a playground to get a feeling what different values of the
spectral index mean, and, e.g., how γ and n are related.6
The Fourier transform of a power law is a power law. For the correlation function of (50) we
get (exercise)
4π (2 − γ)π
P (k) = 3
Γ(2 − γ) sin (kr0 )γ
k 2
2 (2 − γ)π
P(k) = Γ(2 − γ) sin (kr0 )γ (53)
π 2
for 1 < γ < 2 or 2 < γ < 3, and
2π 2
P (k) = (kr0 )2
k3
P(k) = (kr0 )2 (54)
6
In reality the spectral index is very different at small scales than at large scales. Observationally, for small
scales, γ ∼ 1.8, and for large scales, n ∼ 1. We discuss this later.
for γ = 2. Thus
n = γ −3 for 1 < γ < 3, i.e., −2 < n < 0 . (55)
The variance
Z ∞ Z ∞
2 dk 1 n+3 ∞
hδ i = ξ(0) = P(k) ∝ kn+2 dk = k 0
for n 6= −3 (56)
0 k 0 n+3
diverges at small scales (high k) for n ≥ −3 and at large scales (low k) for n ≤ −3. We cure the
small scale divergence with filtering as discussed in Sec. 8.1.5.
Exercise: For a power-law spectrum and a Gaussian window function, show that

2 1 n+3
σ (R) = Γ P(R−1 ) . (57)
2 2
8.1.7 Galaxy 2-point correlation function

The most obvious way to try to measure the cosmological density perturbations is to observe
the spatial distribution of galaxies. We treat individual galaxies as mathematical points, so
that each galaxy has a comoving coordinate value x. We define the galaxy 2-point correlation
function ξg (r) as the excess probability of finding a galaxy at separation r from another galaxy:
dP ≡ n̄ [1 + ξg (r)] dV (58)
where n̄ is the mean galaxy number density, dV is a volume element that is a separation r away
from a chosen reference galaxy, and dP is the probability that there is a galaxy within dV . (Here
dV is assumed so small that there is at most one galaxy in it.)
If the galaxy number density n(x) faithfully traces the underlying matter density, so that
δn δρm
δg ≡ =δ≡ , (59)
n̄ ρ̄m
then ξg becomes equal to the matter density autocorrelation function ξ: The probability of
finding a galaxy in volume dV1 at a random location x is
dP1 = hn(x)idV1 = hn̄ + δn(x)idV1 = n̄dV1 . (60)
The probability of finding a galaxy pair at x and x + r is
dP12 = hn(x)n(x + r)idV1 dV2 = n̄2 h[1 + δ(x)][1 + δ(x + r)]idV1 dV2
= n̄2 [1 + hδ(x)i + hδ(x + r)i + hδ(x)δ(x + r)i] dV1 dV2
= n̄2 [1 + hδ(x)δ(x + r)i] dV1 dV2 , (61)
since hδ(x)i = hδ(x + r)i = 0. Dividing dP12 with dP1 we get the probability dP2 of finding the
second galaxy once we have found the first one
dP2 = n̄ [1 + hδ(x)δ(x + r)i] dV2 = n̄ [1 + ξ(r)] dV2 . (62)
Thus ξg = ξ.
It is probable that the galaxy number density does not trace the matter density faithfully,
since galaxy formation is likely to be more efficient in high-density regions. This is called bias.
Specifically the bias, or galaxy bias bg , is defined as the ratio
δg
bg ≡ ⇒ ξg = b2g ξ , (63)
δm
where the expectation is that bg > 1. In principle the bias could depend on the scale k, the time
t (or redshift z), and/or the strength of the density perturbation δm . The simplest treatment of
bias is to assume bg is a constant over the observationally relevant ranges of these quantities.
The bias will depend on the type of tracer (all galaxies, specific types of galaxies, galaxy
clusters) and is typically larger for more massive objects.
8.2 Newtonian perturbation theory

We shall now study the evolution of perturbations during the history of the universe. Initially
the perturbations were small and we restrict the quantitative treatment to that part of the
evolution when they remained small (for large scales, this extends to the present time and the
future). This allows us to use first-order perturbation theory, where we drop from our equations
all those terms which contain a product of two or more perturbations (as these products are even
smaller). The remaining equations will then contain only terms which are either zeroth order,
i.e., contain only background quantities, or first order, i.e., contain exactly one perturbation. If
we kept only the zeroth order parts, we would be back to the equations of the homogenous and
isotropic universe. Subtracting these from our equations we arrive at the perturbation equations
where every term is first-order in the perturbation quantities, i.e., it is a linear equation for
them. This makes the equations easy to handle, we can, e.g., Fourier transform them.
As we discovered in our discussion of inflation, the different cosmological distance scales
first exit the horizon during inflation, then enter the horizon during various epochs of the later
history. Matter perturbations at subhorizon scales, i.e., after horizon entry, can be treated with
Newtonian perturbation theory, but scales which are close to horizon size or superhorizon require
relativistic perturbation theory, which is based on general relativity.
The Newtonian equations for (perfect gas)7 fluid dynamics with gravity are
∂ρ
+ ∇r · (ρu) = 0 (64)
∂t′
∂u 1
′
+ (u · ∇r )u + ∇r p + ∇r Φ̃ = 0 (65)
∂t ρ
∇2r Φ̃ = 4πGρ (66)
Here ρ is the mass density, p is the pressure, and u is the flow velocity of the fluid. We write Φ̃
for the Newtonian gravitational potential, since we want to reserve Φ for its perturbation. The
subscript r in ∇r emphasizes that the space derivatives are taken with respect to the Newtonian
space coordinate r (instead of a comoving coordinate). Although the Newtonian time coordinate
t′ is equal to the cosmic time coordinate t, we need to make a distinction between t′ and t in
partial derivatives as will become clear soon.
The first equation is the law of mass conservation. The second equation is called the Euler
equation, and it is just “F = ma” for a fluid element, whose mass is ρdV . Here the acceleration
of a fluid element is not given by ∂u/∂t′ which just tells how the velocity field changes at a given
position, but by du/dt′ , where
d ∂
′
≡ ′ + (u · ∇r ) (67)
dt ∂t
is the convective time derivative, which follows the fluid element as it moves. The two other
terms give the forces due to pressure gradient and gravitational field.
We can apply Newtonian physics if:
1) Distance scales considered are ≪ the scale of curvature of spacetime (given by the Hubble
length in cosmology8 )
2) The fluid flow is nonrelativistic, u ≪ c ≡ 1.
3) We are considering nonrelativistic matter, |p| ≪ ρ

7
perfect gas = no internal friction ⇒ pressure is isotropic
8
As discussed in Chapter 3, the spacetime curvature has two distance scales, the Hubble length H −1 and
the curvature radius Rcurv ≡ a|K|−1/2 . From observations we know that the curvature radius is larger than the
Hubble length (at all times of interest), possibly infinite.
The last condition corresponds to particle velocities being nonrelativistic, if the matter is made
out of particles. Although the pressure is small compared to mass density, the pressure gradient
can be important if the pressure varies at small scales.
Note: Energy density and mass density. In Newtonian gravity, the source of gravity is mass
density ρm , not energy density ρ. For nonrelativistic matter, the kinetic energies of particles are negligible
compared to their masses, and thus so is the energy density compared to mass density, if we don’t count
the rest energy in it. The Newtonian equations for mass density and energy density are
∂ρm
+ ∇r · (ρm u) = 0 (68)
∂t′
∂ρu
+ ∇r · (ρu u) + p∇r · u = 0, (69)
∂t′
where ∇r · u gives the rate of change in the volume of the fluid element and p∇r · u is the work done
by pressure. In Newtonian physics, rest energy (mass) is not included in the energy density. Eq. (69)
applies whether we include it or not. Define total energy density as
ρ ≡ ρm + ρu ,
where ρu is the Newtonian energy density and ρm is the mass density. Adding Eqs. (68) and (69) gives
∂ρ
+ ∇r · (ρu) + p∇r · u = 0 . (70)
∂t′
For nonrelativistic matter ρu ≪ ρm and p ≪ ρm . We can thus drop the last term in (70) and ignore the
distinction between mass density and total energy density.
A homogeneously expanding fluid,
ρ = ρ(t0 )a−3 (71)

ȧ
u = r (72)
a
2πG 2
Φ̃ = ρr (73)
3
is a solution to these equations (exercise), with a condition to the function a(t) giving the
expansion law. It is the Newtonian version of the matter-dominated Friedmann model. Writing
H(t) ≡ ȧ/a we find that the homogeneous solution satisfies
ρ̇ + 3Hρ = 0 , (74)
and the condition for a(t) (from the exercise) can be written as
ä 4πG
= Ḣ + H 2 = − ρ. (75)
a 3
You should recognize these equations as the energy-continuity equation and the second Fried-
mann equation for a matter-dominated FRW universe.9 The result for Φ̃, Eq. (73), has no
relativistic counterpart, the whole concept of gravitational potential does not exist in relativity
(except in special cases; like here in perturbation theory, where we introduce potentials related
to perturbations).
9
The freedom of choosing the initial value of the expansion rate leaves the connection between H and ρ open
up to a constant. This constant has the same effect on the time evolution of a(t) as the curvature constant K in
the first Friedmann equation, but of course in the Newtonian treatment it is not interpreted as curvature, and it
does not otherwise have the same physical effects. We shall (unless otherwise noted) choose this constant so that
the background solution matches the flat FRW universe. Then we have
8πG 3 2
H2 = ρ or 4πGρ = H . (76)
3 2
8.2.1 Comoving coordinates

Introduce now a new (comoving) coordinate system (t, x) which is related to the Newtonian
coordinate system (t′ , r) by
t′ = t r = a(t)x . (77)
Thus the time coordinate is the same in both coordinate systems, but we need to distinguish
between the partial derivatives ∂/∂t and ∂/∂t′ , since in the first x is kept constant and in the
second r is kept constant. Relate now the partial derivatives:
∂ ∂t′ ∂ X ∂ri ∂ ∂ X ∂ ∂
= + = + ȧxi = + Hr · ∇r
∂t ∂t ∂t′ ∂t ∂ri ∂t′ ∂ri ∂t′
i
∂ ∂t′ ∂ X ∂rj ∂ X ∂ ∂
= ′
+ = δij a = a ⇒ ∇x = a∇r . (78)
∂xi ∂xi ∂t ∂xi ∂rj ∂rj ∂ri
j j
Thus
∂ ∂ 1
′
= − Hx · ∇x and ∇r = ∇x . (79)
∂t ∂t a
(Later we will work exclusively in the comoving coordinates and write just ∇ for ∇x . The
“original” coordinates r are just an artifact of the Newtonian approach and do not appear in
relativistic perturbation theory.)
8.2.2 The perturbation

Now, consider a small perturbation, so that
ρ(t′ , r) = ρ̄(t) + δρ(t′ , r) (80)

′ ′
p(t , r) = p̄(t) + δp(t , r) (81)
′ ′
u(t , r) = H(t)r + v(t , r) (82)
2πG 2
Φ̃(t′ , r) = ρ̄r + Φ(t′ , r) , (83)
3
where ρ̄, p̄, and H denote homogeneous background quantities (solutions of the background, or
zeroth-order, equations) and δρ, δp, v, Φ are small inhomogeneous perturbations.
Inserting these into the Eqs. (64,65,66) and subtracting the homogeneous equations (73,74,75)
we get (exercise) the perturbation equations
∂δρ
+ 3Hδρ + Hr · ∇r δρ + ρ̄∇r · v = 0 (84)
∂t′
∂v 1
+ Hv + Hr · ∇r v + ∇r δp + ∇r Φ = 0 (85)
∂t′ ρ̄
∇2r Φ = 4πGδρ . (86)
In terms of the comoving coordinates these become (exercise):

∂δρ ρ̄
+ 3Hδρ + ∇x · v = 0 (87)
∂t a
∂v 1 1
+ Hv + ∇x δp + ∇x Φ = 0 (88)
∂t aρ̄ a
∇2x Φ = 4πGa2 δρ . (89)
In terms of the relative density perturbation δ ≡ δρ/ρ̄ we have δρ = ρ̄ · δ and

∂δρ ∂δ
= ρ̄˙ · δ + ρ̄ where ρ̄˙ · δ = −3H ρ̄ δ , (90)
∂t ∂t
and we can write

∂v 1 ∂
+ Hv = (av) (91)
∂t a ∂t
so that the set of perturbation equations becomes
∂δ 1
+ ∇x · v = 0 (92)
∂t a
∂ 1
(av) + ∇x δρ + ∇x Φ = 0 (93)
∂t ρ̄
∇2x Φ = 4πGa2 ρ̄δ (94)
Finally, we Fourier expand the perturbations,

X
δ(t, x) = δk (t)eik·x etc. (95)
k
In Fourier space the perturbation equations become

ik · vk
δ̇k + = 0 (96)
a
d δpk
(avk ) + ik + ikΦk = 0 (97)
dt ρ̄
a 2
Φk = −4πG ρ̄δk . (98)
k
Solving the evolution of the perturbations is a two-step process:
1) Solve the background equations to obtain the functions a(t), H(t), and ρ̄(t). After this,
these are known functions in the perturbation equations.
2) Solve the perturbation equations.
8.2.3 Vector and scalar perturbations

We now divide the velocity perturbation field v(t, r) into its rotational (solenoidal, divergence-
free) and irrotational (curl-free) parts,
v = v⊥ + vk , (99)
where ∇ · v⊥ = 0 and ∇ × vk = 0. For Fourier components this simply means that k · v⊥k = 0
and k × vkk = 0. That is, we divide vk into the components perpendicular and parallel to the
wave vector k. The parallel part we can write in terms of a scalar function v, whose Fourier
components vk are given by
vkk ≡ vk k̂ , (100)
where k̂ denotes the unit vector in the k direction.
We can now take the perpendicular and parallel parts of Eq. (97),
d
(av⊥k ) = 0 (101)
dt
d δpk
(avk ) + ik + ikΦk = 0 . (102)
dt ρ̄
We see that the rotational part of the velocity perturbation has a simple time evolution,
v⊥ ∝ a−1 , (103)
i.e., it decays from whatever initial value it had, inversely proportional to the scale factor.
The other perturbation equations involve only the irrotational part of the velocity perturba-
tion. Thus we can divide the total perturbation into two parts, commonly called the vector and
scalar perturbations, which evolve independent of each other:
1) The vector perturbation: v⊥ .
2) The scalar perturbation: δ, δp, v, Φ, which are all coupled to each other.
The vector perturbations are thus not related to the density perturbations, or the structure
of the universe. Also, any primordial vector perturbation should become rather small as the
universe expands, at least while first-order perturbation theory applies.10 They are thus not very
important, and we shall have no more to say about them. The rest of our discussion focuses on
the scalar perturbations.
8.2.4 The equations for scalar perturbations

We summarize here the equations for scalar perturbations:
ikvk a
δ̇k + = 0 ⇒ vk = i δ̇k (104)
a k
d δpk
(avk ) + ik + ikΦk = 0 (105)
dt ρ̄
a 2
Φk = −4πG ρ̄δk . (106)
k
Inserting vk from (104) and Φk from (106) into (105) we get
k2 δpk
δ̈k + 2H δ̇k = − + 4πGρ̄δk . (107)
a2 ρ̄
8.2.5 Adiabatic and entropy perturbations

Suppose the equation of state is barotropic,
p = p(ρ) (108)
i.e., pressure is uniquely determined by the energy density. Then the perturbations δp and δρ
are necessarily related by the derivative dp/dρ of this function p(ρ),
dp dp
p = p̄ + δp = p̄(ρ̄) + (ρ̄)δρ ⇒ δp = δρ .
dρ dρ
The time derivatives of the background quantities p̄ and ρ̄ are related by this same derivative,
dp̄ dp dρ̄ dp
p̄˙ = = (ρ̄) = ρ̄˙ .
dt dρ dt dρ
Assuming this derivative dp/dρ is nonnegative, we call its square root the speed of sound
s
dp
cs ≡ . (109)
dρ
10
Thus we end up with an irrotational velocity field. The rotational motion (e.g., rotation of galaxies) which
is common in the present universe at small scales has arisen from higher-order effects from the primordial scalar
perturbations, not from the primordial vector perturbations.
Figure 1: For adiabatic perturbations, the conditions in the perturbed universe (right) at (t1 , x)
equal conditions in the (homogeneous) background universe (left) at some time t1 + δt(x).
(We shall indeed find that sound waves propagate at this speed.) We thus have the relation
δp p̄˙
= = c2s .
δρ ρ̄˙
In general, when p may depend on other variables besides ρ, the speed of sound in a fluid is
given by
∂p
c2s = (110)
∂ρ S
where the subscript S indicates that the derivative is taken so that the entropy of the fluid
element is kept constant. Since the background universe expands adiabatically (meaning that
there is no entropy production), we have that

p̄˙ ∂p
= = c2s . (111)
ρ̄˙ ∂ρ S
Perturbations with the property

δp p̄˙
= (112)
δρ ρ̄˙
are called adiabatic perturbations in cosmology.
If p = p(ρ), perturbations are necessarily adiabatic. In the general case the perturbations
may or may not be adiabatic. In the latter case, the perturbation can be divided into an
adiabatic component and an entropy perturbation. An entropy perturbation is a perturbation
in the entropy-per-particle ratio.
For adiabatic perturbations we thus have
p̄˙
δp = c2s δρ = δρ . (113)
ρ̄˙
Adiabatic perturbations have the property that the local state of matter (determined here by
the quantities p and ρ) at some spacetime point (t, x) of the perturbed universe is the same
as in the background universe at some slightly different time t + δt, this time difference being
different for different locations x. See Fig. 1.
Thus we can view adiabatic perturbations as some parts of the universe being “ahead” and
others “behind” in the evolution.
Adiabatic perturbations are the simplest kind of perturbations. Single-field inflation pro-
duces adiabatic perturbations, since perturbations in all quantities are proportional to a pertur-
bation δϕ in a single scalar quantity, the inflaton field.
Adiabatic perturbations stay adiabatic while they are outside horizon, but may develop
entropy perturbations when they enter the horizon. This happens for many-component fluids
(discussed a little later).
Present observational data is consistent with the primordial (i.e., before horizon entry) per-
turbations being adiabatic.
8.2.6 Adiabatic perturbations in matter

Consider now adiabatic perturbations of a non-relativistic single-component fluid. The equation
for the density perturbation is now

c2s k2
δ̈k + 2H δ̇k + − 4πGρ̄ δk = 0 . (114)
a2
I shall call this the Jeans equation11 (although Jeans considered a static, not an expanding fluid).
This is a second-order differential equation from which we can solve the time evolution of
the Fourier amplitudes δk (t) of the perturbation. Before solving this equation we need to first
find the background solution which gives the functions a(t), H(t) = ȧ/a, and ρ̄(t).
The nature of the solution to Eq. (114) depends on the sign of the factor in the brackets.
The first term in the brackets is due to pressure gradients. Pressure tries to resist compression,
so if this term dominates, we get an oscillating solution, standing density (sound) waves. The
second term in the brackets is due to gravity. If this term dominates, the perturbations grow.
The wavenumber for which the terms are equal,
√ r
a 4πGρ̄ 31
kJ = = H, (115)
cs 2 cs
is called the Jeans wave number, and the corresponding wavelength
r
2π 2 −1
λJ = = 2πcs H (116)
kJ 3
the Jeans length. In the latter equalities we assumed that the background solution is the flat
FRW universe, so that
4πGρ̄ = 32 H 2 . (117)
For nonrelativistic matter cs ≪ 1, so that the Jeans length is much smaller than the Hubble
length, kJ ≫ H. Thus we can apply Newtonian theory for scales both larger and smaller than
the Jeans length.
For scales much smaller that the Jeans length, k ≫ kJ , we can approximate the Jeans
equation by
c2 k2
δ̈k + 2H δ̇k + s 2 δk = 0 . (118)
a
The solutions are oscillating, i.e., we get sound waves. The exact solutions of (118) are Bessel
functions, but for small scales we can make a further approximation by first ignoring the middle
term (which is smaller than the other two) and the time-dependence of a and cs to get that
δk (t) ∼ e±iωt , where ω = cs k/a. These oscillations are damped by the 2H δ̇k term, so the
amplitude of the oscillations decreases with time. There is no growth of structure for sub-Jeans
scales.
Exercise: Sound waves. For short-wavelength modes k ≫ kJ , density perturbations in the matter-
dominated universe satisfy (118). Switch to conformal time, dη = dt/a, and solve δk (η) for the Ωm = 1,
11
In the literature, there is usually no name given to this equation, but the terms Jeans length etc. are standard.
ΩΛ = 0 cosmology, assuming cs = const . How does the amplitude and frequency of the oscillations
change with time and scale factor? (Hint: The solutions are spherical Bessel functions.)
For scales much longer than the Jeans length (but still subhorizon), H ≪ k ≪ kJ , we
can approximate the Jeans equation by
δ̈k + 2H δ̇k − 4πGρ̄δk = 0 . (119)
We dropped the pressure gradient term, which means that this equation applies also to nona-
diabatic perturbations for scales where pressure gradients can be ignored. Note that Eq. (119)
is the same for all k, i.e., there is no k-dependence in the coefficients. This means that the
equation applies also in coordinate space, i.e. for δ(x), as long as we ignore contributions from
scales that do not satisfy H ≪ k ≪ kJ .
For a matter-dominated universe, the background solution is a ∝ t2/3 , so that
ȧ 2
H= = (120)
a 3t
and
8πG 4 1
ρ̄ = H 2 = 2 ⇒ ρ̄ = , (121)
3 9t 6πGt2
so the Jeans equation becomes
4 2
δ̈k + δ̇k − 2 δk = 0 . (122)
3t 3t
The general solution is
δk (t) = b1 t2/3 + b2 t−1 . (123)
The first term is the growing mode and the second term the decaying mode. After some time
the decaying mode has died out, and the perturbation grows
δ ∝ t2/3 ∝ a . (124)
Thus density perturbations in matter grow proportional to the scale factor.

From Eq. (98) we have that
Φ ∝ a2 ρ̄δ ∝ a2 a−3 a = const.
The gravitational potential perturbation is constant in time during the matter-dominated era.
8.2.7 Many fluid components

Assume now that the “cosmic fluid” contains several components i (different types of matter
or energy) which do not interact with each other, except gravitationally. This means that
each component sees only its own pressure12 , and that the components can have different flow
velocities. Then the Newtonian equations for each component i are
∂ρi
+ ∇r · (ρi ui ) = 0 (125)
∂t′
∂ui 1
+ (ui · ∇r )ui + ∇r pi + ∇r Φ̃ = 0 (126)
∂t′ ρi
∇2r Φ̃ = 4πGρ , (127)
12
In standard cosmology, we actually have just one component, the baryon-photon fluid, which sees its own
pressure, and the other components do not see even their own pressure (neutrinos after decoupling) or do not
even have pressure (cold dark matter). But we shall first do this general treatment, and do the application to
standard cosmology later.
P
where ρ = ρi . Note that there is only one gravitational potential Φ̃, due to the total density,
and this way the different components do interact gravitationally.
We again have the homogeneous solution, where now each component has to satisfy
ρ̇i + 3Hρi = 0 , (128)
and the expansion law

4πG
Ḣ + H 2 = − ρ, (129)
3
is determined by the total density.
We can now introduce the density, pressure, and velocity perturbations for each component
separately,
ρi (t′ , r) = ρ̄i (t) + δρi (t′ , r) (130)

′ ′
pi (t , r) = p̄i (t) + δpi (t , r) (131)
′ ′
ui (t , r) = H(t)r + vi (t , r) , (132)
but there is only one gravitational potential perturbation,

2πG 2
Φ̃(t′ , r) = ρ̄r + Φ(t′ , r) . (133)
3
Following the earlier procedure, we obtain the perturbation equations for the fluid compo-
nents,
∂
δρi + 3Hδρi + Hr · ∇r δρi + ρ̄i ∇ · vi = 0 (134)
∂t′
∂ 1
′
vi + Hvi + Hr · ∇r vi + ∇r δpi + ∇r Φ = 0 (135)
∂t ρ̄i
∇2r Φ = 4πGδρ (136)
in Newtonian coordinate space, and

ik · vik
δ̇ik + = 0 (137)
a
d δpik
(avik ) + ik + ikΦk = 0 (138)
dt ρ̄i
a2 X
Φk = −4πG ρ̄i δik (139)
k2
P
in comoving Fourier space. Here δρ = δρi and
δρi
δi ≡ . (140)
ρ̄i
Separating out the scalar perturbations we finally get
k2 δpik
δ̈ik + 2H δ̇ik = − + 4πGδρk , (141)
a2 ρ̄i
where X
δρk = ρ̄j δjk . (142)
j
8.2.8 Radiation
Since radiation is a relativistic form of energy, we cannot apply the preceding Newtonian dis-
cussion to perturbations in radiation. However, the qualitative results are similar.
The equation of state for radiation is p = ρ/3, and the speed of sound in a radiation fluid is
given by
dp 1
c2s = = .
dρ 3
Thus the Jeans length for radiation is comparable to the Hubble length, and the subhorizon
scales are also sub-Jeans scales for radiation. Thus for subhorizon radiation perturbations we
only get oscillatory solutions. During the radiation-dominated epoch they are not damped by
expansion, but the oscillation amplitude stays roughly constant.
Relativistic perturbations in non-expanding space. While the full treatment of relativistic
perturbations is beyond the level of this course, we can obtain the limit where we ignore the effect of
expansion by combining special relativity and the Newtonian limit of general relativity. Special relativistic
fluid dynamics follows from the energy-momentum continuity equation
∂T µν
≡ ∂ν T µν ≡ T µν,ν = 0 . (143)
∂xν
For a perfect fluid
T µν = (ρ + p)uµ uν + pg µν , (144)
µν µ
where the metric is now that of Minkowski space, g = diag(−1, 1, 1, 1). The 4-velocity u is related to
the 3-velocity ~v = v i by
uµ = (γ, γv) , (145)
√
where γ = 1/ 1 − v 2 .
By contracting the energy tensor T µν with the 4-velocity uµ we obtain uν T µν
,ν = 0, which gives
(ρuµ ),µ + puµ,µ = 0 , (146)
the energy continuity equation. Subtracting uν times this from (143) we get the special relativistic Euler
equation
(ρ + p)uµ uν,µ + (g µν + uµ uν )p,µ = 0 , (147)
where
uµ uν,µ ≡ aν (148)
is the 4-acceleration.
For small velocities, v ≪ 1, we can approximate γ ≈ 1, so that
uµ ≈ (1, v) (149)
and (146),(147) become

∂ρ
+ ∇ · (ρv) = −p∇ · v
∂t
∂
(ρ + p) + v · ∇ ~v = −∇p − v(v · ∇p) ≈ −∇p . (150)
∂t
In the Newtonian limit of general relativity, but without the assumption p ≪ ρ, the passive gravita-
tional mass density is given by ρ + p, so that the gravitational force on a volume element of fluid is given
by −(ρ + p)∇Φ and the active gravitational mass density by ρ + 3p, so that the gravitational potential
is given by
∇2 Φ = 4πG(ρ + 3p) . (151)
Thus the Euler equation with gravity becomes

∂
(ρ + p) + v · ∇ v = −∇p − (ρ + p)∇Φ . (152)
∂t
For several fluid components, not interacting with each other except gravitationally, the fluid equa-
tions become thus
∂ρi
+ ∇ · (ρi vi ) = −pi ∇ · vi
∂t
∂
(ρi + pi ) + vi · ∇ vi = −∇pi − (ρi + pi )∇Φ (153)
∂t
X
∇2 Φ = 4πG (ρi + 3pi ) .
i
For perturbations ρi = ρ̄i + δρi = ρ̄i (1 + δi ), pi = p̄i + δpi , where the background density and pressure
are now constant both in space and time, we get to first order in perturbations
∂δi
= −(1 + wi )∇ · vi
∂t
∂vi
(ρ̄i + p̄i ) = −∇δpi − (ρ̄i + p̄i )∇Φ (154)
∂t
X
∇2 Φ = 4πG (ρ̄i δi + 3δpi ) .
i
For Fourier components this becomes
δ̇ik = −ik(1 + wi )k · vik

(ρ̄i + p̄i )v̇ik = −ikδpik − ik(ρ̄i + p̄i )Φk (155)
−4πG X
Φk = (ρ̄i δik + 3δpik ) .
k2 i
For vector perturbations the second equation gives
v̇i⊥k = 0 ⇒ vi⊥k = const , (156)
and for scalar perturbations the first and second equations become
δ̇ik = −i(1 + wi )kvik

δpik
v̇ik = −ik − ikΦk , (157)
ρ̄i + p̄i
from which we get (note that ẇi = 0, since wi ≡ ρ̄i /p̄i ) the Jeans equation as
δpik
δ̈ik + k 2 + k 2 (1 + wi )Φk = 0 . (158)
ρ̄i
8.2.9 Adiabatic and entropy perturbations again

The simplest inflation models predict that the primordial perturbations are adiabatic. This
means that locally the perturbed universe at some (t, x) looks like the background universe at
some time t + δt(x). See Sec. 8.2.5.

)  δpi p̄˙ i

 =
˙
δρi (x) = ρ̄i δt(x) δρi ρ̄˙ i
⇒ (159)
δpi (x) = p̄˙ i δt(x) 
 δρ ρ̄˙ δi ρ̄˙ i ρ̄j
 i = i ⇒ =
δρj ρ̄˙ j δj ρ̄i ρ̄˙ j
If there is no energy transfer between the fluid components at the background level, the energy
continuity equation is satisfied by them separately,
ρ̄˙ i = −3H(ρ̄i + p̄i ) ≡ −3H(1 + wi )ρ̄i , (160)

where wi ≡ p̄i /ρ̄i . Thus for adiabatic perturbations,
δi δj
= (161)
1 + wi 1 + wj
(which is thus related to ρ̄i ∝ a−(1+wi ) ). For matter components wi ≈ 0, and for radiation
components wi = 31 . Thus, for adiabatic perturbations, all matter components have the same
perturbation
δi = δm
and all radiation perturbations have likewise
4
δi = δr = δm .
3
We can define a relative entropy perturbation13 between two components

δρi δρj δi δj
Sij ≡ −3H − = − (162)
˙ρ̄i ˙ρ̄j 1 + wi 1 + wj
to describe a deviation from the adiabatic case. The relative entropy perturbation is a pertur-
bation in the ratio of the number densities of the two species. For a nonrelativistic species
δρi δni
ρi = mi ni ⇒ δρi = mi δni and δi ≡ = , (163)
ρ̄i n̄i
whereas for an ultrarelativistic species (µ ≪ T and m ≪ T )
δTi
ρi ∝ Ti4 ⇒ δρi = ρ̄i · 4
Ti
δT i
ni ∝ Ti3 ⇒ δni = n̄i · 3
Ti
δρi 4 δni
⇒ δi ≡ = . (164)
ρ̄i 3 n̄i
For both cases
δni
δi = (1 + wi ) . (165)
n̄i
Thus
δni δnj δ(ni /nj )
Sij = − = . (166)
n̄i n̄j n̄i /n̄j
Even if perturbations are initially adiabatic, relative entropy perturbation may develop inside
the horizon. We shall encounter such a case in Sec. 8.3.4.
8.2.10 The effect of a homogeneous component

The energy density of the real universe consists of several components. In many cases it is
reasonable to ignore the perturbations in some components (since they are relatively small in
the scales of interest). We call such components smooth and we can add them together into a
single smooth component ρs = ρ̄s .
13
There is a connection to entropy/particle of the different components, but we need not concern ourselves with
it now. It is not central to this concept, and it is perhaps somewhat unfortunate that it has become customary,
for historical reasons, to use the word “entropy” for these perturbations.
Consider the case where we have perturbations in a nonrelativistic (“matter”) component

ρm , and the other components are smooth. Then
ρ = ρm + ρs (167)
but
δρ = δρm ≡ ρ̄m δ . (168)
We write just δ for δρm /ρ̄m , since there is no other density perturbation, but note that now
δ 6= δρ/ρ̄ (beware of this trap!).
Assuming adiabatic perturbations, we have then from Eq. (141) that
2 2
cs k
δ̈k + 2H δ̇k + − 4πGρ̄m δk = 0 . (169)
a2
The difference from Eq. (114) is that now the background energy density in the “gravity” term
still contains only the matter component ρ̄m , but the expansion law, a(t) and H(t) comes from
the full background energy density ρ̄ = ρ̄m + ρ̄s .
Newtonian perturbation theory can be applied even with the presence of relativistic energy
components, like radiation and dark energy, as long as they can be considered as smooth com-
ponents and their perturbations can be ignored. Then they contribute only to the background
solution. In this case we have to calculate the background solution using general relativity, i.e.,
the background solution is a FRW universe, but the perturbation equations are the Newtonian
perturbation equations. We can also consider a non-flat (open or closed) FRW universe, as long
as we only apply perturbation theory to scales much shorter than the curvature radius (and
the Hubble length). Thus the background quantities are to be solved from the Friedmann and
energy continuity equations
K 8πG
H2 + = ρ (170)
a2 3
4πG
Ḣ + H 2 = − (ρ + 3p) (171)
3
ρ̇ = −3H(ρ + p) . (172)
Example: Matter perturbations in flat vacuum-dominated universe. Consider the case

where ρs = ρvac ≫ ρm and matter is approximated as pressureless (we do not then have to make a
separate adiabaticity assumption, since the pressure term does not appear). Then the Jeans equation
becomes
δ̈k + 2H δ̇k − 4πGρ̄m δk = 0 . (173)
To estimate the relative order of magnitude of the three terms it is better to divide the equation by H 2 ,
4πGρ̄m
H −2 δ̈k + 2H −1 δ̇k − δk = 0 , (174)
H2
so that the Hubble time H −1 provides the time scale for the time derivatives. Now
8πG 8πG
H2 = ρcr ≈ ρvac = const (175)
3 3
and in the last term δk is multiplied with 23 ρ̄m /ρvac ≪ 1, so that we can drop the last term and approxi-
mate the Jeans equation by
δ̈k + 2H δ̇k = 0 . (176)
We see immediately that δk = const is a solution. For the other solution, solve first δ̇k :
dδ̇k dδ̇k
= −2H δ̇k ⇒ = −2Hdt , (177)
dt δ̇k
whose solution is ln δ̇k = −2Ht + const or δ̇k = Ce−2Ht . Integrating this gives
δk = Ae−2Ht + B , (178)
with a constant term and an exponentially decaying term. Thus in a vacuum-dominated universe matter
perturbations stay constant (after the decaying term has died out); or to be more precise and referring
to the original equation (174), the relative change in δk in a Hubble time is of order ρ̄m /ρvac ≪ 1
We shall do this calculation more accurately later, including the transition from matter
domination to vacuum domination. The main lesson now is that the increased expansion rate
due to the presence of a smooth component slows down the growth of perturbations.
Exercise: Find the solution for the Jeans equation for pressureless matter perturbations when a)
the energy density is dominated by a smooth radiation component b) when there is no other energy
component, but the universe has the open geometry (K < 0) and is curvature dominated, considering
only scales ≪ curvature radius.
8.3 Perturbations at subhorizon scales in the real universe

8.3.1 Horizon entry
Newtonian perturbation theory is valid only at subhorizon scales, k ≫ H, or k−1 ≪ H−1 . During
“normal”, decelerating expansion, i.e., after inflation but before the recent onset of dark energy
domination, scales are entering the horizon. Short scales enter first, large scales enter later.
We have not yet studied what happens to perturbations outside the horizon (for that we need
(general) relativistic perturbation theory, to be discussed somewhat later). So, for the present
discussion, whatever values the perturbation amplitudes δk have soon after horizon entry, are
to be taken as an initial condition, the primordial perturbation14 . Observations actually suggest
that different scales enter the horizon with approximately equal perturbation amplitude, whose
magnitude is characterized by the number15 few × 10−5 .
The history of the different scales after horizon entry, and thus their present perturbation
amplitude, depends on at what epoch they enter. The scales which enter during transitions
between epochs are thus special scales which should characterize the present structure of the
universe. Such important scales are the scale (exercise)
−1
keq = (Heq )−1 ∼ 13.7 Ω−1
m h
−2 −1
Mpc ≡ 13.7 ωm Mpc , (179)
which enters at the time teq of matter-radiation equality, and the scale
−1/2
−1 −1 −1/2 Ωr
kdec = (Hdec ) ∼ 91 Ωm1+ (1 + zdec ) h−1 Mpc
Ωm
−1/2
−1/2 ωr
≡ 91 ωm 1+ (1 + zdec ) Mpc , (180)
ωm
which enters at the time tdec (zdec = 1090) of photon decoupling. Here ωr = 4.18 × 10−5 includes
relativistic neutrinos, since the result above only requires them to be relativistic at tdec . For
ΩΛ = 0.7, Ωm = 0.3, h = 0.7, these scales are
−1
keq = 65 h−1 Mpc = 93 Mpc
−1
kdec = 145 h−1 Mpc = 207 Mpc . (181)
The smallest “cosmological” scale is that corresponding to a typical distance between galaxies,
about 1 Mpc.16 This scale entered during the radiation-dominated epoch (well after Big Bang
nucleosynthesis).
The scale corresponding to the present “horizon” (i.e. Hubble length) is
k0−1 = (H0 )−1 = 2998 h−1 Mpc ∼ 4300 Mpc . (182)
Because of the acceleration due to dark energy, this scale is actually exiting now, and there are
scales, somewhat larger than this, that have briefly entered, and then exited again in the recent
past. The horizon entry is not to be taken as an instantaneous process, so these scales were
14
We shall later redefine primordial perturbation to refer to the perturbations at the epoch when all cosmologi-
cally interesting scales were well outside the horizon, which is the standard meaning of this concept in cosmology.
15
Although in coordinate space the relative density perturbation δ(x) is a dimensionless number, the Fourier
quantity δk is not. The size of δk is characterized by the dimensionless value P(k)1/2 .
16
In the present universe, structure at smaller scales has been completely messed up by galaxy formation, so
that it bears little relation to the primordial perturbations at these scales. However, observations of the high-
redshift universe, especially so-called Lyman-α observations (absorption spectra of high-z quasars, which reveal
distant gas clouds along the line of sight), can reveal these structures when they are closer to their primordial
state. With such observations, the “cosmological” range of scales can be extended down to ∼ 0.1 Mpc.
never really subhorizon enough for the Newtonian theory to apply to them. Thus we shall just
consider scales k−1 < k0−1 . The largest observable scales, of the order of k0−1 , are essentially at
their “primordial” amplitude now.
We shall now discuss the evolution of the perturbations at these scales (k−1 < k0−1 ) after
horizon entry, using the Newtonian perturbation theory presented in the previous section.
8.3.2 Composition of the real universe

The present understanding is that there are five components to the energy density of the universe,
1. cold dark matter (c)
2. baryonic matter (b)
3. photons (γ)
4. neutrinos (ν)
5. dark energy (d)
(during the time of interest for this section, i.e., from some time after BBN until the present).
Thus
ρ = ρc + ρb + ργ + ρν +ρd . (183)
| {z } | {z }
ρm ρr
(Note that ρc here is the CDM density, not the critical density, for which we write ρcr .) Baryons
and photons interact with each other until t = tdec , so for t < tdec they have to be discussed as
a single component,
ρbγ = ρb + ργ . (184)
The other components do not interact with each other, except gravitationally, during the time
of interest. The fluid description of Sec. 8.2 can only be applied to components whose particle
mean free paths are shorter than the scales of interest. After decoupling, photons “free stream”
and cannot be discussed as a fluid. On the other hand, the photon component becomes then
rather homogeneous quite soon, so we can approximate it as a “smooth” component17 . The
same applies to neutrinos for the whole time since the BBN epoch, until the neutrinos become
nonrelativistic. After neutrinos become nonrelativistic, they should be treated as matter (hot
dark matter), not radiation. According to observations, the neutrino masses are small enough,
not to have a major impact on structure formation. Thus we shall here approximate neutrinos
as a smooth radiation component. Dark energy is believed to be relatively smooth. If it is a
cosmological constant (vacuum energy) then it is perfectly homogeneous.
The discussion in Sec. 8.2 applies to the case, where ρ can be divided into two components,
ρ = ρm + ρs , (185)
where the perturbation is only in the matter component ρm and ρs = ρ̄s is homogeneous.
For perturbations in radiation components and dark energy the Newtonian treatment is not
enough. Unfortunately, we do not have quite this two-component case here. Based on the above
discussion, a reasonable approximation is given by a separation into three components:
t < tdec : ρ = ρc + ρbγ + ρs (ρs = ρν + ρd ) (186)

t > tdec : ρ = ρc + ρb + ρs (ρs = ργ + ρν + ρd ) . (187)
17
As long as we are interested in density perturbations only. When we are interested in the CMB anisotropy,
the momentum distribution of these photons becomes the focus of our attention.
After decoupling, both ρc and ρb are matter-like (p ≪ ρ) and we’ll discuss in Sec. 8.3.4 how this
case is handled. Before decoupling, the situation is more difficult, since ρbγ is not matter-like, the
pressure provided by the photons is large. Here we shall be satisfied with a crude approximation
for this period.
The most difficult period is that close to decoupling, where the photon mean free path λγ
is growing rapidly. The fluid description, which we are here using for the perturbations, applies
only to scales ≫ λγ , whereas the photons are smooth only for scales ≪ λγ . Thus this period
can be treated properly only with large numerical “Boltzmann” codes, such as CMBFAST or
CAMB.
8.3.3 CDM density perturbations

Cold dark matter is the dominant structure-forming component in the universe (dark energy
dominates the energy density at late times, but does not form structure, or, if it does, these
structures are very weak, not far from homogeneous). Observations indicate that ρb ∼ 0.2ρc .
Thus we get a reasonable approximation for the behavior of the CDM perturbations by ignoring
the baryon component and equating
ρm ≈ ρc .
The CDM is pressureless, and thus the CDM sound speed is zero, and so is the CDM Jeans
length. Thus, for CDM, all scales are larger than the Jeans scale, and we don’t get an oscillatory
behavior. Instead, perturbations grow at all scales. On the other hand, as we shall discuss in
Sec. 8.3.4, perturbations in ρbγ oscillate before decoupling. Therefore the perturbations in ρbγ
will be smaller than those in ρc , and we can make a (crude) approximation where we treat ρbγ
as a homogeneous component before decoupling. This is important, since although ρb ≪ ρc , this
is not true for ρbγ at earlier, radiation-dominated, times. At decoupling ρb < ργ < ρc . Before
matter-radiation equality, there is an epoch when ρc < ργ , but δρc > δρbγ . For simplicity, we
now approximate
ρ = ρm + ρr + ρd (188)
where ρm = ρc and ρr = ργ + ρν is a smooth component (ρν truly smooth, ργ truly smooth
after decoupling, and (crudely) approximated as smooth before decoupling). We have ignored
baryons, since they are a subdominant part of ρbγ before decoupling, and a subdominant matter
component after decoupling. Likewise, ρd is also smooth, and becomes important only close to
present times.
We can now study the growth of CDM perturbations even during the radiation-dominated
period, as the radiation-component is taken as smooth and affects only the expansion rate. We
can study it all the way from horizon entry to the present time, or until the perturbations
become nonlinear (δc = δρc /ρ̄c ∼ 1).
We get the equation for the CDM perturbation from Eq. (169) by setting cs = 0 (or rather,
δp = 0; we need not invoke the assumption of adiabaticity, since CDM is pressureless),
δ̈k + 2H δ̇k − 4πGρ̄m δk = 0 . (189)
Note that the equation is the same for all k and therefore it applies also in the coordinate space,
i.e., for δ(x). To simplify notation, we drop the subscript k .
We now assume a flat universe, and ignore the ρd component (a good approximation at early
times18 ), so that the Friedmann equation is
2
2 ȧ 8πG
H = = ρ̄ ,
a 3
18
For a flat universe (or an open/closed universe without dark energy) we can also do the late times (see
Sec. 8.3.5). Then dark energy (or curvature) must be included in the background evolution, but radiation can be
ignored.
where ρ̄ = ρ̄m + ρ̄r and ρ̄m ∝ a−3 and ρ̄r ∝ a−4 .

A useful trick is to study this as a function of a instead of t or η. We define a new time
coordinate,
a ρ̄m
y≡ = . (190)
aeq ρ̄r
(y = 1 at t = teq .) Now
y 3 y
4πGρ̄m = 4πG ρ̄ = H2 (191)
y+1 2y+1
and Eq. (189) becomes
3 y
δ̈ + 2H δ̇ − H2 = 0 . (192)
2y+1
Performing the change of variables from t to y (Exercise; you may need the 2nd Friedmann
equation), we arrive at the equation (where ′ ≡ d/dy)
2 + 3y ′ 3
δ′′ + δ − δ = 0, (193)
2y(1 + y) 2y(1 + y)
known as the Meszaros equation.

It has two solutions, one growing, the other one decaying. The growing solution is

3y 3 a
δ = δprim 1+ = δprim 1+ . (194)
2 2 aeq
We see that the perturbation remains frozen to its primordial value, δ ≈ δprim , during the
radiation-dominated period. By t = teq , it has grown to δ = 52 δprim .
During the matter-dominated period, y ≫ 1, the CDM perturbation grows proportional to
the scale factor,
δ ∝ y ∝ a ∝ t2/3 . (195)
In reality, for the case of adiabatic primordial perturbations, there is an additional logarith-
mic growth factor ∼ ln(k/keq ) the CDM perturbations get from the gravitational effect (ignored
in the above) of the oscillating radiation perturbation during the radiation-dominated epoch.
To get this boost the CDM perturbations must initially be in the same direction (positive or
negative) as the radiation perturbations, which is the case for adiabatic primordial perturbations:
For adiabatic primordial perturbations, the baryon, CDM, and radiation perturbations at
are related at horizon entry as δc = δb = 34 δγ . Consider scales that enter during the radiation-
dominated epoch (t < teq < tdec ). The gravitational effect is dominated initially by the radiation
perturbations, which begin to oscillate after horizon entry; the baryon perturbations will oscillate
with them until tdec . CDM on the other hand, does not see the radiation pressure responsible
for the oscillation, it sees only the gravitational effect of the baryon-photon fluid. In the first
phase of the oscillation period δc is of the same sign as δbγ so δbγ adds to the gravitational pull
to increase δc and since at first δρbγ > δρc , this additional pull is larger than that of CDM itself,
leading to a much faster growth of δc (which otherwise would grow very little during the radiation
domination). The flow of CDM is accelerated towards CDM overdensities. In the next phase
of the oscillation, the sign of δbγ reverses, and now the pull of δρbγ on CDM is in the opposite
direction, and will slow down the flow of CDM towards overdensities. But this is not enough
to reverse the CDM flow before the sign of δbγ changes again and begins to accelerate CDM
again towards CDM overdensities. Thus the effect of the radiation oscillations is to increase δc
stepwise, one step for each oscillation period. As the ρ̄γ /ρ̄c ratio decreases the relative increases
per step decrease; but this effect keeps adding steps until tdec . The smaller the scale (the higher
the k) the more steps there are between horizon entry (tk ) and tdec , and the larger the first
steps. The calculation of this effect is too complicated for this course (I do it in Cosmological
Perturbation Theory) but for k ≫ keq the effect is a boost by a factor ∼ 8 ln(keq /6k), so that
(194) is modified to

3 a k
δc ≈ δprim 1 + 8 ln for k ≫ 6keq and t > tdec (196)
2 aeq 6keq
(for k < 6keq ) the logarithm is negative; this approximate result does not apply for such large
scales).
8.3.4 Baryon density perturbations

Although CDM is the dominant matter component in the universe, we cannot directly see it.
The main method to observe the density perturbations today is to study the distribution of
galaxies. But the part of galaxies that we can see is baryonic. Thus to compare the theory
of structure formation to observations, we need to study how perturbations in the baryonic
component evolves.
We define the baryon Jeans length as λJ = 2πkJ−1 , where
cs
kJ−1 = √ , (197)
a 4πGρ̄b
and cs is the speed of sound for baryons (i.e., in the baryon-photon fluid before decoupling,
and in the baryon fluid after decoupling). This definition compares baryon pressure to baryon
gravity, so it addresses the question whether baryonic density perturbations can grow under
their own gravity. This is not the question we face in reality, since at early times baryons
were coupled to photons, and after decoupling the gravity of CDM perturbations dominates.
The baryon Jeans length can still be used for order-of-magnitude estimates on at what scales
the baryon perturbations can grow (and for the argument that we cannot match observations
without CDM).
In general,
2 ∂p
cs = , (198)
∂ρ σ
where σ refers to constant entropy per baryon. Since in our case the entropy is completely
dominated by photons,
4π 2 3 2π 4
sbγ ∼ sγ = T = nγ , (199)
45 45ζ(3)
we have
sbγ sγ 2π 4 nγ 1
σ≡ ∼ = ≈ 3.6016 , (200)
nb nb 45ζ(3) nb η
where η is the baryon-to-photon ratio.
We find the speed of sound by varying ρbγ and pbγ adiabatically, (i.e., keeping σ, the en-
tropy/baryon constant), which in this case means keeping η constant. Now
2ζ(3) 3 δT
ρb = mnb = mηnγ = mη T ⇒ δρb = ρ̄b · 3
π2 T
π2 4 δT
ργ = T ⇒ δργ = ρ̄γ · 4
15 T
π2 4 δT 4 δT
pγ = T ⇒ δpγ = p̄γ · 4 = ρ̄γ · .
45 T 3 T
Since pb ≪ pγ ⇒ δpb ≪ δpγ , we get

4
δp δpγ 3 ρ̄γ 1 1
c2s = = = = . (201)
δρ δργ + δρb 4ρ̄γ + 3ρ̄b 3 1 + 34 ρ̄ρ̄b
γ
This was a calculation of the speed of sound, which one gets by varying the pressure and den-
sity adiabatically. It is independent of whether the actual perturbations we study are adiabatic
or not.
This result, Eq. (201), applies before decoupling. As we go back in time, ρ̄b /ρ̄γ → 0 and
2
cs → 1/3. As we approach decoupling, ρ̄b becomes comparable to (but still smaller than) ρ̄γ
and the speed of sound falls, but not by a large factor.
Newtonian perturbation theory applies only to subhorizon scales. The ratio of the (comoving)
baryon Jeans length
2πcs
λJ = √
a 4πGρ̄b
to the comoving Hubble length
1
H−1 = q
a 8πG3 ρ̄
is r
λJ 2ρ̄
= HλJ = 2π cs .
H−1 3ρ̄b
Thus we see that before decoupling the baryon Jeans length is comparable to the Hubble length,
and thus all scales for which our present discussion applies are sub-Jeans. Therefore, if baryon
perturbations are adiabatic19 , they oscillate before decoupling20 .
After decoupling, the baryon component sees just its own pressure. This component is now
a gas of hydrogen and helium. This gas is monatomic for the epoch we are now interested in.
Hydrogen forms molecules only later. For a non-relativistic monatomic gas,
5Tb
c2s = , (202)
3m
where we can take m ≈ 1 GeV, since hydrogen dominates. Down until z ∼ 100, residual
free electrons maintain enough interaction between the baryon and photon components to keep
Tb ≈ Tγ . After that the baryon temperature falls faster,
Tb ∝ (1 + z)2 whereas Tγ ∝ 1 + z (203)
(as shown in an exercise in Chapter 4). For example, at 1 + z = 1000, soon after decoupling,
Tb = 2725 K = 0.2348 eV and the speed of sound is cs = 5930 m/s. The baryon density is
ρ̄b = Ωb (1 + z)3 ρcr = ωb (1 + z)3 1.88 × 10−26 kg/m3 , and we get for the Jeans length
√
πcs
λJ = (1 + z) √ (204)
Gρ̄b
19
If there is an initial baryon entropy perturbation, i.e., a perturbation in baryon density without an accom-
panying radiation perturbation, it will initially begin to grow in the same manner as a CDM perturbation, since
the pressure perturbation provided by the photons is missing. (Such a baryon entropy perturbation corresponds
to a perturbation in the baryon-photon ratio η.) But as the movement of baryons drags the photons with them,
a radiation perturbation is generated, and the baryon perturbation begins to oscillate around its initial value
(instead of oscillating around zero).
20
We have not calculated this exactly, since all our calculations have been idealized, i.e., we have used per-
turbation theory which applies only to matter-dominated perturbations, and here we have ignored the CDM
component. But this qualitative feature will hold also in the exact calculation, and this will be enough for us
now.
that soon after decoupling

−1/2
λJ (1 + z = 1000) = ωb 0.96 × 103 pc = η10 0.016 Mpc ∼ 0.095 Mpc , (205)
where η10 ≡ 1010 η = 274 ωb or ωb = 0.00365 η10 , and the last number is for η10 ∼ 6.
We define the baryon Jeans mass
π
MJ ≡ ρ̄b0 λ3J (206)
6
as the mass of baryonic matter within a sphere whose diameter is λJ . Note that since λJ is
defined as a comoving distance, we must use here the present (mean) baryon density ρ̄b0 . At
−1/2 −1/2
1 + z = 1000, the baryon Jeans mass is ωb 1.3 × 105 M⊙ = η10 2.1 × 106 M⊙ ∼ 9 × 105 M⊙
for η10 ∼ 6. This corresponds to the mass of a globular cluster and is much less than the mass of
a galaxy. Thus, for our purposes, the baryonic component is pressureless after decoupling, i.e.,
baryon pressure can be ignored in the evolution of perturbations at cosmological scales (greater
than ∼ 1 Mpc). (The pressure cannot be ignored for smaller scale physics like the formation of
individual galaxies.)
After decoupling, the evolution of the baryon density perturbation is governed by the grav-
itational effect of the dominant matter component, the CDM.
We now have the situation of Sec. 8.2.10, except that we have two matter components,
ρ = ρc + ρb + ρs , (207)
where we approximate ρs = ργ + ρν + ρd as homogeneous. With the help of Sec. 8.2.7, the

discussion is easy to generalize for the present case.
We can ignore the pressure of both ρb and ρc . Therefore their perturbation equations are
δ̈c + 2H δ̇c = 4πGρ̄m δ (208)

δ̈b + 2H δ̇b = 4πGρ̄m δ (209)
where ρ̄m = ρ̄c + ρ̄b is the total background matter density and
δρc + δρb
δ= (210)
ρ̄c + ρ̄b
is the total matter density perturbation.
We can now define the baryon-CDM entropy perturbation,
Scb ≡ δc − δb , (211)
which expresses how the perturbations in the two components deviate from each other. Sub-
tracting Eq. (209) from (208) we get an equation for this entropy perturbation,
S̈cb + 2H Ṡcb = 0 . (212)
We assume that the primordial perturbations were adiabatic, so that we had δb = δc , i.e,
Scb = 0 at horizon entry. For large scales, which enter the horizon after decoupling, an Scb never
develops, so the evolution of the baryon perturbations is the same as CDM perturbations.
But for scales which enter before decoupling, an Scb develops because the baryon perturba-
tions are then coupled to the photon perturbations, whereas the CDM perturbations are not.
After decoupling, δb ≪ δc , since δc has been growing, while δb has been oscillating. The initial
condition for Eqs. (208,209,212) is then Scb ∼ δc (“initial” time here being the time of decoupling
Figure 2: Evolution of the CDM and baryon density perturbations after horizon entry (at t = tk ). The
figure is just schematic; the upper part is to be understood as having a ∼ logarithmic scale; the difference
δc − δb stays roughly constant, but the fractional difference becomes negligible as both δc and δb grow by
a large factor.
tdec ). During the matter-dominated epoch, when a ∝ t2/3 , so that H = 2/3t, the solution for
Scb is
Scb = A + Bt−1/3 , (213)
whereas for δc it is, neglecting the effect of baryons on it, from Eq. (123),
δc = Ct2/3 + Dt−1 ∼ Ct2/3 . (214)
We call the first term the “growing” and the second term the “decaying” mode (although for
Scb the “growing” mode is actually just constant). For δc the growing and decaying modes have
been growing and decaying since horizon entry, so we can now drop the decaying part of δc .
To work out the precise initial conditions, we would need to work out the behavior of Scb
during decoupling. However, we really only need to assume that initially there is no strong
cancellation between the growing and decaying modes, so that Scb = δc − δb either shrinks or
stays roughly constant near the initial value of δc . While δc grows by a large factor, δb must
follow it to keep the difference close to the initial small value of δc , so that δb /δc → 1.
Thus the baryon density contrast δb grows to match the CDM density contrast δc (see Fig. 2),
and we have eventually δb = δc = δ to high accuracy.
The baryon density perturbation begins to grow only after tdec . Before decoupling the radi-
ation pressure prevents it. Without CDM it would grow only as δb ∝ a ∝ t2/3 after decoupling
(during the matter-dominated period; the growth stops when the universe becomes dark energy
dominated). Thus it would have grown at most by the factor a0 /adec = 1 + zdec ∼ 1100 after de-
coupling. In the anisotropy of the CMB we observe the baryon density perturbations at t = tdec .
They are too small (about 10−4 ) for a growth factor of 1100 to give the present observed large
scale structure21 .
With CDM this problem was solved. The CDM perturbations begin to grow earlier, at
t ∼ teq , and by t = tdec they are much larger than the baryon perturbations. After decoupling
21
This assumes adiabatic primordial perturbations, since we are seeing δγ , not δb . For a time, primordial baryon
entropy perturbations Sbγ = δb − 34 δγ were considered a possible explanation, but more accurate observations
have ruled this model out.
Figure 3: A figure summarizing the evolution of perturbations at different subhorizon scales. The
baryon Jeans length kJ−1 drops precipitously at decoupling so that all cosmological scales became super-
Jeans after decoupling, whereas all subhorizon scales were sub-Jeans before decoupling. The wavy lines
symbolize the oscillation of baryon perturbations before decoupling, and the opening pair of lines around
them symbolize the ∝ a growth of CDM perturbations after teq . There is also an additional weaker
(logarithmic) growth of CDM perturbations between horizon entry and teq .
the baryons have lost the support from photon pressure and fall into the CDM gravitational
potential wells, catching up with the CDM perturbations.
This allows the baryon perturbations to be small at t = tdec and to grow after that by much
more than the factor 103 , matching observations. This is one of the reasons we are convinced
that CDM exists.22
The whole subhorizon evolution history of all the different cosmological scales of perturba-
tions is summarized by Fig. 3.
8.3.5 Late-time growth in the ΛCDM model

At late times, dark energy begins to accelerate the expansion, which will slow down the growth
of density perturbations. In the ΛCDM model dark energy is just a constant vacuum energy, so
it has no perturbations and thus affects just the background. The perturbations are in CDM
and baryons, and we can ignore the pressure term in the Jeans equation, since at such small
scales where baryon pressure gradients would be important, first-order perturbation theory is
not valid anyway at late times. Thus we are facing a similar calculation as we did in Sec. 8.3.3,
the solution of Eq. (189),
δ̈k + 2H δ̇k − 4πGρ̄m δk = 0 , (215)
22
Historically, the above situation became clear in the 1980’s when the upper limits to CMB anisotropy (which
was finally discovered by COBE in 1992) became tighter and tighter. By today we have accurate detailed
measurements of the structure of the CMB anisotropy which are compared to detailed calculations including the
CDM so the argument is raised to a different level—instead of comparing just two numbers we are now comparing
entire power spectra (to be discussed later).
with δb = δc = δ, but instead of radiation we have now vacuum energy contributing to the
background solution, which is the Concordance Model discussed in Cosmology I (Chapter 3):
1/3 p
Ωm
a(t) = sinh2/3 3
2 ΩΛ H0 t . (216)
ΩΛ
The Hubble parameter is given by
p
H = H0 Ωm a−3 + ΩΛ . (217)
Again, it is better to use the scale factor as time coordinate. The difference in the power
of a in the behavior of the two density components is now 3 instead of 1, which makes the
calculation more difficult. We follow here Dodelson[5]. After the change of variable from t to a,
(215) becomes (exercise)
′
H 3 3Ωm H0 2
δ′′ + + δ′ − δ = 0, (218)
H a 2a5 H
where ′ ≡ d/da. The decaying solution is

p
δ ∝ H ∝ Ωm a−3 + ΩΛ (219)
and the growing solution is

Z a p Z a
dx −3 + Ω
x3/2 dx
δ ∝ H ∝ Ω m a Λ 3/2 (220)
H 3 x3 ΩΛ 3
1+ Ωm x
The effect of changing the lower limit of integration can be incorporated in the decaying solution;
so we can set the lower limit to 0. (Equation (218) is valid in general for matter perturbations
with an additional smooth background component. The first forms of the solutions (219) and
(220) are valid when the smooth component is vacuum energy or negative curvature.)
In the limit a ≪ 1, or rather, ΩΛ ≪ Ωm a−3 , the decaying solution becomes
δ ∝ a−3/2 ∝ t−1 (221)
and the growing solution becomes

δ ∝ a ∝ t2/3 (222)
the familiar results for the matter-dominated universe from Sec. 8.2.6. We can ignore the
decaying mode, since it has become completely negligible when the vacuum energy begins to
have an effect.
To fix the proportionality coefficient in the growing mode, we write it as
Z
1/2 a x3/2 dx
δ = A Ωm a−3 + ΩΛ 3/2 (223)
0 ΩΛ 3
1 + Ωm x
and note that in the limit ΩΛ ≪ Ωm a−3 it becomes

Z a
2
δ ≈ AΩ1/2
m a −3/2
x3/2 dx = Ω1/2
m Aa . (224)
0 5
At a = a0 = 1 this would give
2 1/2 5 −1/2
Ω A ≡ δ̃ ⇒ A= Ω δ̃ , (225)
5 m 2 m
where we have defined δ̃ as the value δ would have “now”23 if there were no vacuum energy, i.e.,
the universe had stayed matter dominated.
Thus we write (223) as
1/2 Z a
5 −3 ΩΛ x3/2 dx
δ = δ̃ a + 3/2 . (226)
2 Ωm 0 ΩΛ 3
1+ Ωm x
Unfortunately, the integral in (226) does not give an elementary function. (I think it is a so-
called hypergeometric function, which does not give much useful information compared to just
integrating (226) numerically.) We can see that at late (future) times, when a ≫ 1, there is very
little growth, since the factor outside the integral approaches a constant and for any a1 ≫ 1 and
a2 ≫ 1, the contribution to the integral,
Z a2 Z
x3/2 dx Ωm 3/2 a2 −3 1 Ωm 3/2 −2
3/2 ≈ x dx = a1 − a−22 (227)
ΩΛ 2 ΩΛ
a1
1 + ΩΩm
Λ 3
x a1
is very small.
It turns out that the integral can be done if we extend it to the infinite future (exercise) :
As a → ∞,
Z
5 −3 ΩΛ 1/2 ∞ x3/2 dx 5 Ωm 1/3
δ → δ(∞) ≡ 2 δ̃ a + 3/2 = 6 δ̃ Ω B( 56 , 23 ) , (228)
Ωm 0 ΩΛ 3 Λ
1 + Ωm x
where Z 1
Γ(p)Γ(q)
B(p, q) ≡ tp−1 (1 − t)q−1 dt = (229)
0 Γ(p + q)
is the beta function and
5 2
B 6, 3 ≈ 1.725 . (230)
Thus the perturbations “freeze”, i.e., approach a final value

Ωm 1/3
δ(∞) = 1.437 δ̃ . (231)
ΩΛ
which for Ωm = 0.3, ΩΛ = 0.7 gives
δ(∞) = 1.084 δ̃ , (232)
i.e., the perturbations will never become much stronger than what they in the matter-dominated
model would be already “now”. To get the present density perturbation δ(a = 1) one has to do
(226) numerically. This is done in Fig. 4, from which one can read that δ(a = 1) ≈ 0.78 δ̃.
For perturbations that entered horizon well before matter-radiation equality teq , we have
from (196) that
3 k
δ̃ ≈ δprim 1 + 8 ln , (233)
2aeq 6keq
23
Note that we defined “now” as a = a0 = 1, not as t = t0 ; or in more physical terms as T = T0 = 2.725 K.
The comparison situation ( ˜ ) we have in mind is that the early universe (where vacuum energy has no effect)
is the same as in the ΛCDM model, but there is no vacuum energy to accelerate the expansion at late times,
so that by “now” the expansion rate, i.e., H0 is smaller than we observe in reality. The present matter density
1/2
ρm0 = (3/8πG)Ωm H02 is the same as in the ΛCDM model, but Ω̃m = 1, so H̃0 = Ωm H0 . The age of the
−1 2 −1/2 −1
universe is t̃0 = 3 H̃0 = 3 Ωm H0 , which for h = 0.7 and Ωm = 0.3 gives t̃0 = 17.0 × 109 years, instead of the
2
t0 = 13.5 × 109 years of the ΛCDM model.

assuming that this is still ≪ 1 so that first-order perturbation theory remains valid, and re-
membering that this was an approximate result that ignored the effect of the baryon-photon
oscillations on CDM during the radiation-dominated epoch. For larger scales, and for what δprim
is, we need the remaining sections of this chapter.
For Ωm = 0.3, ΩΛ = 0.7, h = 0.7, we have keq −1 = 65 h−1 Mpc and a
eq = 1/3603. Equa-
−1 −1
tion (233) gives then for the scale k = 8 h Mpc,
δ̃ ≈ 13 000δprim and δ(a = 1) ≈ 10 000δprim . (234)
Observationally, the variance of the top-hat-filtered density field of the galaxy distribution today
is ≈ 1 at this scale. Because of the galaxy bias bg , the corresponding variance for the matter
distribution is less by factor b−2
g , but still not far from 1, meaning that the linear perturbation
theory approximation is beginning to break.
8.3.6 Growth function

Inside the horizon, during the matter- and dark-energy-dominated epochs the linear growth of
perturbations is independent of scale (once the decaying mode has died out and ignoring the
subcosmological scales where pressure gradients have a role). Thus it can be described by a
function that depends on time (or scale factor, or redshift) only, called the growth function,
δ(a)
D(a) ≡ (235)
δref
where δ(a) is the density perturbation (δk or δ(x); D(a) is the same function for any k or x)
when scale factor is a and δref is it at some reference time. The choice of reference time fixes
the normalization of D. We define the growth rate
d ln D d ln δ a dδ
f≡ = = , (236)
d ln a d ln a δ da
which is independent of this normalization.
For the ΛCDM model of Sec. 8.3.5, we get from (226) (exercise)
!
1 5 δ̃
f (a) = a −3 (237)
1 + ΩΩΛ a3 2 δ 2
m
It turns out that a good approximation for the growth rate is
f (a) ≈ Ωm (a)γ , where γ = 0.55 , (238)
where γ is called the growth index. (This result assumes General Relativity, and the measurement
of the growth index from galaxy surveys is a way of testing gravity theory.) We plot D, f , and
the approximation (238) for ΛCDM in Fig. 4.
Figure 4: The growth function D(a) (blue, with normalization δref = δ̃), growth rate f (a) (red) and the
approximation (238) (red, dashed) for ΛCDM with Ωm = 0.3.
8.4 Relativistic perturbation theory

For scales comparable to, or larger than the Hubble scale, Newtonian perturbation theory does
not apply, because we can no more ignore the curvature of spacetime. Therefore we need to
use (general) relativistic perturbation theory. Instead of the Newtonian equations of gravity
and fluid mechanics, the fundamental equation is now the Einstein equation of general relativity
(GR). We assume a background solution, which is homogeneous and isotropic, i.e., a solution of
the Friedmann equations, and study small perturbations around it. This particular choice of the
background solution means that we are doing a particular version of relativistic perturbation
theory, called cosmological perturbation theory.
The evolution of the perturbations while they are well outside the horizon is simple, but the
mathematical machinery needed for its description is complicated. This is due to the coordinate
freedom of general relativity. For the background solution we had a special coordinate system
(time slicing) of choice, the one where the t = const slices are homogeneous. The perturbed
universe is no more homogeneous, it is just ”close to homogeneous”, and therefore we no more
have a unique choice for the coordinate system. We should now choose a coordinate system
where the universe is close to homogeneous on the time slices, but there are many different
possibilities for such slicing. This freedom of choosing the coordinate system in the perturbed
universe is called gauge freedom, and a particular choice is called a gauge.24 The most important
part of the choice of gauge is the choice of the time coordinate, because it determines the slicing
of the spacetime into t = const slices, ”universe at time t”. Sometimes the term ‘gauge’ is used
to refer only to this slicing.
Because the perturbations are defined in terms of the chosen coordinate system, they look
different in different gauges. We can, for example, choose the gauge so that the perturbation in
one scalar quantity, e.g., proper energy density, disappears, by choosing the ρ = const 3-surfaces
as the time slices (this is called ”the uniform energy density gauge”).
24
If you are familiar with gauge field theories, like electrodynamics, the concept of ‘gauge’ may look different
here. The mathematical similarity appears when the perturbation equations are developed. In relativistic per-
turbation theory gauge has this geometric origin (this is where the use of the word “gauge” comes from), unlike
electrodynamics.
The true nature of gravitation is spacetime curvature, so perturbations should be described

in terms of curvature.
We leave the actual development of cosmological perturbation theory to a more advanced
course, and just summarize here some basic concepts and results.
In the Newtonian theory gravity was represented by a single function, the gravitational
potential Φ. In GR, gravity is manifested in the geometry of spacetime, described in terms of
the metric. Thus in addition to the density, pressure, and velocity perturbations, we have a
perturbation in the metric. The perturbed metric tensor is
gµν = ḡµν + δgµν . (239)
For the background metric, ḡµν , we choose that of the flat Friedmann-Robertson-Walker
universe,
ds2 = ḡµν dxµ dxν = −dt2 + a(t)2 (dx2 + dy 2 + dz 2 ) (240)
The restriction to the flat case is an important simplification, because it allows us to Fourier
expand our perturbations in terms of plane waves.25 Fortunately the real universe appears to
be flat, or at least close to it. And earlier it was even flatter. Inflation predicts a flat universe.
For the metric perturbation, we have now 10 functions δgµν (t, x). So there appears to be ten
degrees of freedom. Four of them are not physical degrees of freedom, since they just correspond
to our freedom in choosing the four coordinates. So there are 6 real degrees of freedom.
Two of these metric degrees of freedom couple to density and pressure perturbations and
the irrotational velocity perturbation. These are the scalar perturbations. Two couple to the
rotational velocity perturbation to make up the vector perturbations. The remaining two are not
coupled to the cosmic fluid at all26 , and are called tensor perturbations. They are gravitational
waves, which do not exist in Newtonian theory.
The vector perturbations decay in time, and are not produced by inflation, so they are the
least interesting. Although the tensor perturbations also are not related to growth of structure,
they are produced in inflation and affect the cosmic microwave background anisotropy and
polarization. Different inflation models produce tensor perturbations with different amplitudes
and spectral indices (to be explained later), so they are an important diagnostic of inflation. No
tensor perturbations have been detected in the CMB so far, but they could be detected in the
future with more sensitive instruments if their amplitude is large enough.27
Since the three kinds of perturbations evolve independently of each other, they can be stud-
ied separately. We shall first concentrate on the scalar perturbations, returning to the tensor
perturbations later.
8.4.1 Gauges for scalar perturbations

Consider now scalar perturbations. The gauges discussed in the following assume scalar pertur-
bations.
The perturbations appear different in the different gauges. When needed, we use superscripts
to indicate in which gauge the quantity is defined: C for the comoving gauge and N for the
Newtonian gauge. Some other gauges are the synchronous gauge (S), spatially flat gauge (Q),
and the uniform energy density gauge (U).
25
In the Newtonian case this restriction was not necessary, and we could apply it to any Friedmann model,
as there is no curvature of spacetime in the Newtonian view, and only the expansion law a(t) of the Friedmann
model is used. The Newtonian theory of course is only valid for small scales were the curvature can indeed be
ignored.
26
This is true in first-order perturbation theory in the perfect fluid approximation, but not in general.
27
Typically, large-field inflation models produce tensor perturbations with much larger amplitude than small-
field inflation models. In the latter case they are likely to be too small to be detectable.
There are two common ways to specify a gauge, i.e., the choice of coordinate system in the
perturbed universe:
• A statement about the relation of the coordinate system to the fluid perturbation. This
will lead to some condition on the metric perturbations.
• A statement about the metric perturbations. This will then lead to some condition on the
coordinate system.
The two gauges (C and N) we shall refer to in the following, give an example of each.
The comoving gauge is defined so that the space coordinate lines x = const follow fluid flow
lines, and the time slice, the t = const hypersurface is orthogonal to them. Thus the velocity
perturbation is zero in this gauge,
vC = 0 . (241)
The conformal-Newtonian gauge, also called the longitudinal gauge, or the zero-shear gauge,
and sometimes, for short, just the Newtonian gauge, is defined by requiring the metric to be of
the form
ds2 = −(1 + 2Φ)dt2 + a2 (1 − 2Ψ)(dx2 + dy 2 + dz 2 ) . (242)
This means that we require
δg0i = 0, δg11 = δg22 = δg33 , and gij = 0 for i 6= j . (243)
(This is possible for scalar perturbations). The two metric perturbations, Φ(t, x) and Ψ(t, x) are
called Bardeen potentials.28 Φ is also called the Newtonian potential, since in the Newtonian limit
(k ≫ H and p ≪ ρ), it becomes equal to the Newtonian gravitational potential perturbation.
Thus we can use the same symbol for it. Ψ is also called the Newtonian curvature perturbation,
because it determines the curvature of the 3-dimensional t = const subspaces, which are flat in
the unperturbed universe (since it is the flat FRW universe).
It turns out that the difference Φ − Ψ is caused only by anisotropic stress (or anisotropic
pressure). We shall here consider only the case of a perfect fluid. For a perfect fluid the pressure
(or stress) is necessary isotropic. Thus we have only a single metric perturbation29
Ψ=Φ (244)
The density perturbations in these two gauges become equal in the limit k ≫ H, and we can
then identify them with the “usual” density perturbation δ of Newtonian theory.
8.4.2 Evolution at superhorizon scales

When the perturbations are outside the horizon (meaning that the wavelength of the Fourier
mode we are considering is much longer than the Hubble length), very little happens to them,
and we can find quantities which remain constant for superhorizon scales. Such a quantity is the
(comoving) curvature perturbation R(x), which describes how curved is the t = const slice in
28
Warning: The sign conventions for Ψ differ, and many authors call them Ψ and Φ instead.
29
In reality, neutrinos develop anisotropic pressure after neutrino decoupling. Therefore the two Bardeen
potentials actually differ from each other by about 10 % between the times of neutrino decoupling and matter-
radiation equality. After the universe becomes matter-dominated, the neutrinos become unimportant, and Ψ and
Φ rapidly approach each other. The same happens to photons after photon decoupling, but the universe is then
already matter-dominated, so they do not cause a significant Ψ − Φ difference.
the comoving gauge.30 For adiabatic perturbations, the curvature perturbation R stays constant
in time outside the horizon.
Using gauge transformation equations R can be related to the metric in the Newtonian
gauge. The result is
5 + 3w 2
R = − Φ− H −1 Φ̇ , (247)
3 + 3w 3 + 3w
where w ≡ p̄/ρ̄.
Because Rk stays constant while k ≪ H, it is a very useful quantity for “carrying” the
perturbations from their generation at horizon exit during inflation to horizon entry at later
times. We now define the primordial perturbation to refer to the perturbation at the epoch
when it is well outside the horizon. For adiabatic perturbations, the primordial perturbation
is completely characterized by the set of these constant values Rk . We shall later discuss how
the primordial perturbation is generated by inflation, and how these superhorizon values Rk are
determined by it.
However, we would like to describe the perturbation in more “familiar” terms, the gravita-
tional potential perturbation Φ and the density perturbation δ. When Rk remains constant this
turns out to be easy. Eq. (247) can be written as a differential equation for Φk ,
2 −1 5 + 3w
H Φ̇k + Φk = −(1 + w)Rk . (248)
3 3
During any period, when also w = const, the solution of this equation is
3 + 3w
Φk = − Rk + a decaying part . (249)
5 + 3w
Thus, after w has stayed constant for some time, the Bardeen potential has settled to the
constant value
3 + 3w
Φk = − Rk (w = const ) . (250)
5 + 3w
In particular, we have the relations
2
Φ k = − Rk (rad.dom, w = 31 ) (251)
3
3
Φ k = − Rk (mat.dom, w = 0) . (252)
5
After the potential has entered the horizon, we can use the Newtonian perturbation theory
result, Eq. (106), which gives the density perturbation as
2
k Φk 2 k 2 2 k 2
δk = − = − Φk = − Φk , (253)
a 4πGρ̄ 3 aH 3 H
30
Technically, R is defined in terms of the trace of the space part of the comoving gauge metric perturbation
(−Ψ is the corresponding quantity in the Newtonian gauge), and it is related to the scalar curvature (3) RC of the
comoving gauge time slice (the (3) reminds us that we are considering a 3-dimensional subspace, and the C refers
to the comoving gauge) so that
(3) C
R = −4a−2 ∇2 R . (245)
For Fourier components we have then that
1 a 2 (3) C
Rk ≡ Rk . (246)
4 k
Another similar quantity is the (uniform-density-gauge) curvature perturbation ζ that is defined the same way,
but for the uniform-density-gauge time slice. For superhorizon scales they are equal, R = ζ (in the limit k ≪ H).
where we used the background relation
8πG 3
H2 = ρ̄ ⇒ 4πGρ̄ = H 2 . (254)
3 2
The problem is to get Φk from its superhorizon epoch where it is constant (as long as
w = const) through the horizon entry to its subhorizon epoch where it evolves according to
Newtonian theory. For scales k which enter while the universe is matter dominated, this is easy,
since in this case Φk stays constant the whole time (until dark energy becomes important).
Thus we can relate the constant values of Φk , and the corresponding subhorizon density
perturbations δk during the matter-dominated epoch to the primordial perturbations Rk by
3
Φk = − Rk (mat.dom)
5
(255)
2 k 2 2 k 2 1
δk = − Φk = Rk ∝ ∝ t2/3 ∝ a
3 H 5 H (aH)2
Note that by Rk we refer always to the constant primordial value, when we use it in equations,
like (255), that give other quantities at later times.
For perturbations which enter during the radiation-dominated epoch, the potential Φk does
not stay constant. We learned earlier, that in this case the density perturbations oscillate with
roughly constant amplitude, which means that the amplitude for the potential Φ must decay
∝ a2 ρ̄ ∝ a−2 . This oscillation applies to the baryon-photon fluid, whereas the CDM density
perturbations grow slowly. After the universe becomes matter dominated, it is these CDM
perturbations that matter.
We shall now make a crude estimate how the amplitudes of these smaller-scale perturbations
during the matter-dominated epoch are related to the primordial perturbations. These pertur-
bations enter during the radiation-dominated epoch. Assume that the relation Φk = − 23 Rk
holds all the way to horizon entry (k = H). Assume then that the Newtonian relation (253)
holds already. Then

2 k 2 2 4
δk ≈ − Φk = − Φk ≈ Rk (256)
3 H 3 9
at horizon entry. The universe is now radiation-dominated, and therefore δrk = δk . We are
assuming primordial adiabatic perturbations and therefore the adiabatic relations δc = 34 δr ,
δγ = δr hold at superhorizon scales. Assume that these relations hold until horizon entry. After
that δγk begins to oscillate, whereas δck grows slowly. Thus we have that at horizon entry
3 1
δck ≈ δk ≈ Rk . (257)
4 3
Ignoring the slow growth of δc we get that δck stays at this value until the universe becomes
matter-dominated at t = teq , after which we can approximate δk ≈ δck and δk begins to grow
according to the matter-dominated law, ∝ 1/H2 .
Thus
1
δk (teq ) ≈ Rk (258)
3
and
1 Heq 2 1 keq 2
δk (t) ≈ Rk = Rk for t > teq , (259)
3 H 3 H
as long as the universe stays matter dominated.
8.4.3 Transfer function

For large scales (k ≪ keq ) which enter the horizon during the matter-dominated epoch, we got

2 k 2
δk (t) = Rk (k ≪ keq ) , (260)
5 H
for as long as the universe stays matter dominated.
This is a simple result, and we use this as a reference for the more complicated result at
smaller scales. That is, we define a transfer function T (k, t) so that

2 k 2
δk (t) = T (k, t)Rk (261)
5 H
where Rk refers to the primordial perturbation. Thus by definition T (k, t) = 1 for k ≪ keq .31
Using the rough estimate from the previous subsection we get that

5 keq 2
T (k, t) ≈ (262)
6 k
during the matter-dominated epoch, where we can drop the factor 65 , since this is anyway just
a rough estimate.
Once we are well into the matter-dominated era, perturbations at all scales grow ∝ a ∝
1/(aH)2 and the transfer function becomes independent of time,32
T (k) = 1 k ≪ keq

keq 2
T (k) ∼ k ≫ keq (263)
k
A more accurate calculation, including the gravitational effect of baryon-photon oscillations
on the CDM perturbations, and assuming adiabatic primordial perturbations, adds a logarithmic
growth factor and gives

keq 2 k
T (k) ≈ 12 ln k ≫ keq (264)
k 6keq
Note that that logarithm is negative for k < 6keq ; the equation is not supposed to apply yet for
this low k.
According to present understanding, the universe becomes dark energy dominated as we
approach the present time. The equation-of-state parameter w begins to decrease (becomes
negative) and therefore Φ begins to change again. The growth of the density perturbations is
slowed down as we saw in Sec. 8.3.5. The effect is not very big since the universe has expanded
by less than a factor of 2 after the onset of dark energy domination, but it is important in
detailed matching of observations and theory.
Our calculation of the growth of structure has been just a rough approximation. A detailed
calculation including all relevant effects has to be done numerically. There are publicly available
computer programs (such as CMBFAST and CAMB) that do this (you give your favorite values
for the cosmological parameters as input). The exact result can be given in form of the transfer
function T (k) we defined above. The main corrections to our simple results, (263) and (264),
include:
31
With the given definition for T (k, t), this holds for t ≪ t0 , i.e, before we entered the present dark-energy-
dominated epoch.
32
We shall later define other transfer functions, but this is the transfer function T (k) of structure formation
theory. It relates the perturbations inside the horizon during the matter-dominated epoch to the primordial
perturbations, and it is independent of time.
1) The transition from k ≪ keq behavior to k ≫ keq behavior is, of course, smooth.
2) The effect of baryon acoustic oscillations (i.e., the oscillations of δbγ before decoupling,
which leave a trace in δb ) shows up as a small-amplitude wavy pattern in the k > keq part
of the transfer function, since different modes k were at a different phase of the oscillation
when that ended around tdec .
We have calculated everything using linear perturbation theory. This breaks down when
the perturbations become large, δ(x) ∼ 1. We say that the perturbation becomes nonlinear.
This has happened for the smaller scales, k−1 < 10 Mpc by now. When the perturbation
becomes nonlinear, i.e., an overdense region becomes significantly denser (say, twice as dense)
as the average density of the universe, it collapses rapidly, and forms a gravitationally bound
structure, e.g. a galaxy or a cluster of galaxies. Further collapse is prevented by the angular
momentum of the structure. Galaxies in a cluster and stars (and CDM particles) in a galaxy
orbit around the center of mass of the bound structure.
8.4.4 Tensor perturbations

In addition to scalar and vector perturbations, in general relativistic perturbation theory we have
tensor perturbations. They have the nice property that we do not have to worry about different
gauges, since they are gauge invariant in the sense that, if we first do a gauge transformation and
then separate out the scalar, vector, and tensor parts, the tensor part has remained unchanged.
These are perturbations of the metric that for one Fourier mode take the form

ds2 = −dt2 + a(t)2 (1 + h)dx2 + (1 − h)dy 2 + dz 2

= a(η)2 −dη 2 + (1 + h)dx2 + (1 − h)dy 2 + dz 2 (265)
where
h = hk (t)eikz (266)
is the perturbation and η is conformal time. In (265) we have chosen the z axis in the direction
of the wave vector, so that k = k k̂ and k·x = kz. Since the metric is a real quantity, in (265) and
(271) h should be interpreted as the real part of h; like one should always do when one makes
physical interpretations for a single Fourier mode. Remember that when one sums over Fourier
components the imaginary parts of hk (t)eikz + h−k (t)e−ikz cancel since h−k = h∗k , and thus the
imaginary parts have no physical significance, they are just a mathematical convenience.
The effect of the tensor perturbation is to stretch space in one direction (here x if h is
positive) and compress it in the other direction (here y) orthogonal to the wave vector of the
Fourier mode. In (265) we also chose the orientation of the x and y axes so that they correspond
to these stretch/compress directions. But of course the perturbation could be oriented differently.
We get the other possibilities by rotating the pattern around the wave vector k by some angle
ϕ, which is mathematically equivalent to rotating the coordinate system by angle −ϕ.
In matrix form the metric is
 
−1
 1+h 
[gµν ] = a2 

 (267)
1−h 
1
After rotation by ϕ around the z axis it becomes
   
1 −1 1
 cos ϕ − sin ϕ   1+h  cos ϕ sin ϕ 
[gµν ] = a2 
 sin ϕ cos ϕ



 
 (268)
1−h − sin ϕ cos ϕ 
1 1 1
√
Rotation by 45◦ , i.e., cos ϕ = sin ϕ = 1/ 2, gives
 
−1
 1 h 
[gµν ] = a2 


 (269)
h 1
1
We call (267) the + mode and (269) the × mode. An arbitrary orientation of the stretch/compress
pattern can be obtained as a linear combination of these two modes, so that the general form
of the tensor perturbation is
 
−1
 1 + h+ h× 
[gµν ] = a2 

 (270)
h× 1 − h+ 
1
or

ds2 = −dt2 + a(t)2 (1 + h+ )dx2 + 2h× dxdy + (1 − h+ )dy 2 + dz 2

= a(η)2 −dη 2 + (1 + h+ )dx2 + 2h× dxdy + (1 − h+ )dy 2 + dz 2 (271)
for a Fourier mode in the z direction. Thus we have two Fourier amplitudes h+k (t) and h×k (t)
for each wave vector k. In the following we mostly write just h(t) to represent an arbitrary such
mode.
The evolution equation for h(t),
2
k
ḧ + 3H ḣ + h=0 ⇔ H −2 ḧ + 3H −1 ḣ + (k/H)2 h = 0 , (272)
a
can be obtained from the Einstein equation. This derivation is beyond the level of this course,
but the equation has a simple and plausible form: it is the wave equation with a damping term
3H ḣ; the wave velocity is the speed of light = 1.
For superhorizon scales we can ignore the last term, and we get h = const as a solution
and another solution where ḣ ≡ dh/dt ∝ a−3 so it also approaches a constant. Thus tensor
perturbations remain essentially constant outside the horizon.
For evolution inside the horizon we get oscillatory solutions and then it is better to work
with conformal time. The h(η) evolution equation is
h′′ + 2Hh′ + k2 h = 0 ⇔ H−2 h′′ + 2H−1 h′ + (k/H)2 h = 0 , (273)
where ′ ≡ d/dη. If we first ignore the middle term, we get solutions of the form h ∝ e±ikη , where
− represents a wave moving in the k direction and + in the −k direction. These are gravitational
waves. They propagate at the speed of light and they are transverse waves. During one half-
period of the wave oscillation, space is stretched in one direction orthogonal to the direction of
propagation, and compressed in the other orthogonal direction. During the next half-period the
opposite happens. The amplitude of the stretching is given by h, meaning that the maximum
stretching is by factor 1 + |h| and the maximum compression is by factor 1 − |h|.
The middle term in (273) represents the damping of gravitational terms due to the expansion
of the universe. Write
h(η) = A(η)e−ikη (274)
and insert this into (273) to get
A′′ + 2HA′ − 2ik(A′ + HA) = 0 . (275)

For k ≫ H, the part 2ik(A′ + HA) dominates the left-hand side, and we get
a′ 1
A′ + HA = A′ + A = (aA)′ = 0 ⇒ aA = const ⇒ A ∝ a−1 . (276)
a a
Thus gravitational waves are damped inside the horizon as a−1 independent of the expansion
law.
For simple expansion laws one can also solve Eq. (273) exactly, covering also horizon en-
try/exit. These solutions are Bessel functions.
8.5 Nonlinear growth

When δ grows the evolution becomes nonlinear, requiring a more complicated discussion. One
can get further with higher-order perturbation theory, or what is called the Zeldovich approx-
imation, but eventually one has to resort to numerical simulations. We shall not discuss these
in this course. The spherically symmetric special case can be done analytically by basing it on
solutions for FRW universes with different densities. We do it below for an overdensity in a flat
matter-dominated background universe.
8.5.1 Closed Friedmann model

In Cosmology I we derived the expansion law for the closed (Ω > 1) matter-dominated FRW
universe. It cannot be given in closed form as a(t), but can be given in terms of an auxiliary
variable, the development angle ψ, as
Ωi Ω(ψ)
a(ψ) = ai (1 − cos ψ) = a(ψ) (1 − cos ψ)
2(Ωi − 1) 2[Ω(ψ) − 1]
Ωi Ω(ψ)
t(ψ) = Hi−1 (ψ − sin ψ) = H(ψ)−1 (ψ − sin ψ) , (277)
2(Ωi − 1)3/2 2[Ω(ψ) − 1]3/2
where ai , Ωi , and Hi are the scale factor, density parameter, and Hubble parameter at some
reference time ti (usually chosen as the present time t0 , but below we will instead choose ti to
be some early time, when Ω is still very close to 1). In the second forms we took advantage of
the fact that we can choose ti to be any time during the development and replaced it with the
“current” time. See Fig. 5 for the shape of a(t). This curve is called a cycloid. (It is the path
made by a point at the rim of a wheel.) From (277) we solve
2
Ω(ψ) = . (278)
1 + cos ψ
Calculating da/dt = da/dψ × dψ/dt we find (exercise)
(Ωi − 1)3/2 sin ψ

H(ψ) = 2Hi . (279)
Ωi (1 − cos ψ)2
The matter density is given by

3
ai (Ωi − 1)3
ρ(ψ) = ρi = 8ρi . (280)
a(ψ) Ω3i (1 − cos ψ)3
The scale factor reaches a maximum ata (and the density a minimum) at the “turnaround”
time tta , when ψ = π, so that
Ωi π −1 Ωi (Ωi − 1)3
ata = ai , tta = Hi , and ρ(tta ) = ρi . (281)
Ωi − 1 2 (Ωi − 1)3/2 Ω3i
At this point H = 0 and then the universe begins to shrink. Since
3Ωi Hi2 3π
ρi = we have ρ(tta ) = . (282)
8πG 32Gt2ta
The universe collapses at tcoll = 2tta , when ψ = 2π and a = 0 again.

14
12
10
scale factor
0
0 10 20 30 40 50
time
Figure 5: The expansion law for the flat matter-dominated universe (blue) and for closed matter-
dominated universes with different initial values Ωi > 1 for the density parameter. Both axes are linear,
the units are arbitrary.
8.5.2 Spherical collapse

The expansion law (277) will hold also for a spherically symmetric overdense region within a
flat (Ω = 1) matter-dominated FRW universe. Denote the quantities for this flat background
universe by ā, H̄, ρ̄. (Time t is the same for both solutions and Ω̄ = 1, so we don’t need notations
for them.) The background universe has
2
8πG 2 1
H̄ 2 = ρ̄ = ⇒ ρ̄ = (283)
3 3t 6πGt2
Thus we see that at tta , the density of the overdense region is
9π 2
ρ(tta ) =
ρ̄(tta ) ≈ 5.5517ρ̄(tta ) , (284)
16
i.e., at the turnaround time the density contrast has the value
9π 2
− 1 ≈ 4.4417 .
δta = (285)
16
Until then the overdense region has been expanding, although slower than the surrounding
background universe. At turnaround the overdense region begins to shrink (in terms of proper
distance).
The preceding applies both for an overdense region with homogeneous density and for one
with a spherically symmetric density profile. In the latter case, we have to apply it separately
for each spherical shell, and the density ρ refers, not to the density of the shell, but to the mean
density within the shell, as it is the total mass within the shell that is responsible for the gravity
affecting the expansion or contraction of the shell. To avoid shell crossing the density profile
has to decrease outward, so that outer shells do not collapse before inner shells.33
33
We should also include in our model an underdense region around our overdense region so that their combined
mean density equals that of the background universe, so as not to affect the evolution of the surroundings.
In linear perturbation theory, which applies when δ ≪ 1, density perturbations in the flat
matter-dominated universe grow as
δlin ∝ a ∝ t2/3 . (286)
When the density contrast δ becomes large it begins to grow faster. Compare now the linear
growth law to the above result for δ at turnaround.
The initial density contrast δi is given by ρi = (1 + δi )ρ̄i . On the other hand
8πG 8πG
H̄i2 = ρ̄i and Hi2 = Ω i ρi (287)
3 3
so that
Hi2 H2
1 + δi = Ωi or at any time 1+δ =Ω . (288)
H̄i2 H̄ 2
Thus the density contrast is not simply given by Ω− Ω̄ = Ω−1, since also the Hubble parameters
are different for the two solutions. We can sort out the separate contributions from Ωi − 1 and
(Hi /H̄i )2 at an early time when Ω − 1 ≪ 1 and ψ ≪ 1, by expanding Ω, H and H̄ from (278),
(279) and (283&277) in terms of ψ (exercise) to get
Hi2
Ωi ≈ 1+ 41 ψi2 and 1 2
≈ 1− 10 ψ ⇒ 3 2
1+δi ≈ 1+ 20 ψ ⇒ δi ≈ 35 (Ωi −1) . (289)
H̄i2
We can now give the linear prediction for the density contrast at turnaround time34 :
2/3 2/3 2/3
lin āta tta 3π δi 3 3π
δta = δi = δi ≈ ≈ ≈ 1.0624 , (290)
āi ti 4 Ωi − 1 5 4
where we approximated
π −1 1
tta ≈ H̄i and ti = 23 H̄1−1 . (291)
2 (Ωi − 1)3/2
Thus we conclude that density perturbations begin to collapse when the linear prediction is
δ ∼ 1, at which time the true density perturbation is already over 4 times stronger.
The collapse is completed at tcoll = 2tta , when the linear prediction gives
lin
δcoll = 22/3 δta
lin
≈ 1.6865 . (292)
The above special case can be extended to the situation where the background universe is a
closed or open Friedmann model (i.e., a matter-dominated FRW universe), and to the ΛCDM
model, with more complicated math.
8.5.3 Without spherical symmetry

I suppose these idealized cases would lead to a supermassive black hole at the center of symmetry
(for perturbations at cosmological scales, for a smaller scale perturbation we might end up
with a star). In reality overdensities are never exactly spherically symmetric. The deviation
from spherical symmetry increases as the collapse progresses. For an ellipsoidal overdensity the
flattest direction collapses first leading first to a “Zeldovich pancake”, and the second flattest
next leading then to an elongated structure. In the situation where the density refers to a number
density of galaxies instead of a smooth continuous density, the galaxies will pass the center point
at various distances (instead of colliding at the center as in the perfectly spherically symmetric
34
Note that Kolb&Turner[6], p. 328, misses the factor 3/5 .
case), after which they will move away from the center and will be decelerated, eventually falling
back in and ending up orbiting the center, forming a cluster of galaxies.
For the real universe the different distance scales are in a different stage of the collapse. The
largest distance scales are still “falling in”, leading to flattened structures at the largest scales
and elongated structures, “filaments”, at somewhat smaller scales. These structures surround
rounder underdense regions, “voids”. Smaller scales have already collapsed into galaxy clusters.
8.6 Perturbations during inflation

So far we have developed perturbation theory describing the substance filling the universe in
fluid terms, i.e., giving the perturbations in terms of δρ and δp. During inflation the universe is
dominated by a scalar field, the inflaton ϕ, so it is better to give the perturbation directly as a
perturbation in the inflaton field,
ϕ(t, x) = ϕ̄(t) + δϕ(t, x) . (293)
8.6.1 Evolution of inflaton perturbations

In Minkowski space the field equation for a scalar field is
ϕ̈ − ∇2 ϕ + V ′ (ϕ) = 0 . (294)
In the flat Friedmann-Robertson-Walker universe (the background universe) the field equation
is
ϕ̈ + 3H ϕ̇ − a−2 ∇2 ϕ + V ′ (ϕ) = 0 . (295)
(Here ∇ = ∇x , i.e., with respect to the comoving coordinates x, and therefore the factor 1/a
appears in front of it.)
We ignore for the moment the perturbation in the spacetime metric and just insert (293)
into Eq. (295),
(ϕ̄ + δϕ)¨ + 3H(ϕ̄ + δϕ)˙ − a−2 ∇2 (ϕ̄ + δϕ) + V ′ (ϕ̄ + δϕ) = 0 . (296)
Here V ′ (ϕ̄ + δϕ) = V ′ (ϕ̄) + V ′′ (ϕ̄)δϕ and ϕ̄(t) is the homogeneous background solution from our
earlier discussion of inflation. Thus ∇2 ϕ̄ = 0, and ϕ̄ satisfies the background equation
ϕ̄¨ + 3H ϕ̄˙ + V ′ (ϕ̄) = 0 . (297)
Subtracting the background equation from the full equation (296) we get the perturbation
equation
δϕ̈ + 3Hδϕ̇ − a−2 ∇2 δϕ + V ′′ (ϕ̄)δϕ = 0 (298)
In Fourier space we have
" #
k 2 2
δϕ̈k + 3Hδϕ̇k + + m (ϕ̄) δϕk = 0 , (299)
a
or " 2 #
−2 −1 k m2
H δϕ̈k + 3H δϕ̇k + + 2 δϕk = 0 , (300)
aH H
where
m2 (ϕ̄) ≡ V ′′ (ϕ̄) . (301)
During inflation, H and m2 change slowly. Thus we make now an approximation where we
treat them as constants. If the slow-roll approximation is valid, m2 ≪ H 2 , since
m2 2 V
′′
= 3MPl = 3η ≪ 1 . (302)
H2 V
Thus we can ignore the m2 /H 2 in Eq. (300)35 . The general solution becomes then
δϕk (t) = Ak wk (t) + Bk wk∗ (t) , (305)
35
The general solution to (299), when H and m2 are constants, is

k k
δϕk (t) = a−3/2 Ak J−ν + Bk Jν , (303)
aH aH
where
k ik
wk (t) = i+ exp . (306)
aH aH
(Exercise: Show that this is a solution of (299) when H = const and m2 = 0.) The time
dependence of (305) is in
a = a(t) ∝ eHt . (307)
Well before horizon exit, k ≫ aH, the argument of the exponent is large. As a(t) increases
the solution oscillates rapidly and its amplitude is damped. After horizon exit, k ≪ aH, the
solution stops oscillating and approaches the constant value i(Ak − Bk ).
We have cheated by ignoring the metric perturbation. We should use GR and write the
curved-spacetime field equation using the perturbed metric. Perturbations in a scalar field
couple only to scalar perturbations, so we need to consider scalar perturbations only. For
example, in the conformal-Newtonian gauge the correct perturbation equation is
" #
k 2
N N
δϕ̈k + 3Hδϕ̇k + + V (ϕ̄) δϕN
′′
k = −2Φ k V (ϕ̄) + Φ̇ k + 3Ψ̇ ˙.
k ϕ̄ (308)
a
That is, there are additional terms which are first order in the metric and zeroth order (back-
ground) in the scalar field ϕ.
Fortunately, it is possible to choose the gauge so that the terms with the metric perturbations
are negligible during inflation36 , and the previous calculation applies in such a gauge. The
comoving gauge is not such a gauge, so a gauge transformation is required to obtain the comoving
gauge curvature perturbation R. Gauge transformations are beyond the scope of these lectures,
but the result is
δϕ
R = −H . (309)
ϕ̄˙
Thus it is clear what we want from inflation. We want to find the inflaton perturbations δϕk
some time after horizon exit. We can use the constant value the solution (305) approaches after
horizon exit. Then Eq. (309) gives us Rk , which remains constant while the scale k is outside
the horizon, and is indeed the primordial Rk discussed in the previous section. And then we
can use the results of Sec. 8.4 to get δk .
We are still missing the initial conditions for the solution (305). These are determined by
quantum fluctuations, which we shall discuss in Sec. 8.6.3. Quantum fluctuations produce the
initial conditions in a random manner, so that we can predict only their statistical properties. It
turns out that the quantum fluctuations are a Gaussian process, a term which specifies certain
statistical properties, which we shall discuss next before returning to the application to inflaton
fluctuations.
8.6.2 Statistical properties of Gaussian perturbations

The statistical (Gaussian) nature of the inflaton perturbations δϕ(x) are inherited later by
other perturbations, which depend linearly on them. Let us therefore discuss a generic Gaussian
where Jν is the Bessel function of order ν and
r
9 m2
ν= − 2. (304)
4 H
With m2 = 0, ν = 32 . Bessel functions of half-integer order are spherical Bessel functions which can be expressed
in terms of trigonometric functions, or e±ikx .
36
One such gauge is the spatially flat gauge Q. For scalar perturbations it is possible to choose the time
coordinate so that the time slices have Euclidean geometry. This leads to the spatially flat gauge. (There are still
perturbations in the spacetime curvature; they show up when one considers the time direction).
perturbation X
g(x) = gk eik·x , (310)
k
where the set of Fourier coefficients {gk } is a result of a statistically homogeneous and isotropic
Gaussian random process. We assume g(x) is real, so that g−k = gk∗ . We write gk in terms of
its real and imaginary part,
gk = αk + iβk . (311)
For Fourier analysis of statistically homogeneous and isotropic random perturbations, see
sections (8.1.1, 8.1.3, 8.1.4), where the probability distribution was treated as unknown. The
new ingredient (in addition to the assumption that the perturbations are small, allowing the
use of first-order perturbation theory, which we introduced in Sec. 8.2), is that the probability
distribution is known to be Gaussian. This means that

1 1 |gk |2
Prob(gk ) = exp −
2πs2k 2 s2k
(312)
1 1 α2k 1 1 βk2
=√ exp − 2 × √ exp − 2 ,
2πsk 2 sk 2πsk 2 sk
i.e., the real and imaginary parts are independent Gaussian random variables37 with equal
variance s2k .
The expectation value of a quantity which depends on gk as f (gk ) is given by
Z
hf (gk )i ≡ f (gk )Prob(gk )dαk dβk , (313)
where the integral is over the complex plane, i.e.,

Z ∞ Z ∞
dαk dβk .
−∞ −∞
We immediately get (exercise) the mean
hgk i = 0 (314)
and variance
h|gk |2 i = 2s2k (315)
of gk .
The distribution has one free parameter, the real positive number sk which gives the width
(determines the variance) of the distribution. From statistical isotropy and homogeneity follows
that sk = s(k) and
hgk∗ gk′ i = 0 for k 6= k′ . (316)
We can combine Eqs. (315) and (316) into a single equation,
δkk′ 2π 2 δkk′
hgk∗ gk′ i = 2δkk′ s2k = δkk′ h|gk |2 i = Pg (k) = Pg (k) , (317)
V V k3
where 3
L V 3
Pg (k) ≡ 4πk3 h|gk |2 i = k h|gk |2 i , (318)
2π 2π 2
37
gk is a complex Gaussian random variable and αk and βk are real Gaussian random variables.
which gives the dependence of the variance of gk on the wave number k, is the power spectrum
of g.
Going back to coordinate space, we find
* +
X X
ik·x
hg(x)i = gk e = hgk ieik·x = 0 (319)
k k
The square of the perturbation can be written as

X X ′
g(x)2 = gk∗ e−ik·x gk′ eik ·x (320)
k k′
since g(x) is real. The typical amplitude of the perturbation is described by the variance, the
expectation value of this square,
X X X 3 X
2 ′ 2π 1
hg(x) i = hgk∗ gk′ iei(k −k)·x
= 2
h|gk | i = 2 2
sk . = Pg (k)
′
L 4πk3
kk k k k
Z 3 Z ∞ Z ∞
1 d k dk
→ Pg (k) = Pg (k) = Pg (k)d ln k . (321)
4π k3 0 k −∞
Note that there is no x-dependence in the result, since this is an expectation value. g(x)2 of
course varies from place to place, but its expectation value from the random process is the same
everywhere—the perturbed universe is statistically homogeneous.
Thus the power spectrum of g gives the contribution of a logarithmic scale interval to the
variance of g(x). For Gaussian perturbations, the power spectrum gives a complete statistical
description. All statistical quantities can be calculated from it.
In practice the integration is not extended all the way from k = 0 to k = ∞. Rather, there
is usually some largest and smallest relevant scale, which introduce natural cutoffs at both ends
of the integral. The largest relevant scale could be the size of the observable universe: The
perturbation g(x) represents a deviation from the background quantity, but the best estimate
we have for the background may be the average taken over the observable universe. Then
perturbations at larger scales contribute to our estimate of the background value instead of
contributing to the perturbation away from it. The smallest relevant scale could be the resolution
of the observational survey considered. For example, density perturbations are observed as
perturbations in the number density of galaxies; such a number density can only be meaningfully
defined at scales larger than the typical separation between galaxies.
It can be shown (under weak assumptions about the power spectrum), that statistically
homogeneous and isotropic Gaussian perturbations are ergodic, so that we do not need to make
a separate assumption of ergodicity.38
An alternative definition for the power spectrum is
Pg (k) ≡ V h|gk |2 i (322)
While this definition is simpler, the result for the variance of g(x) in terms of it and thus the
interpretation is more complicated. Because of the common use of this latter definition, we shall
make reference to both power spectra, and distinguish them by the different typeface. They are
related by
2π 2
Pg (k) = 3 Pg (k) . (323)
k
38
Liddle & Lyth [2], in Sec. 4.3.3, make this claim but do not provide a proof.
8.6.3 Generation of primordial perturbations from inflation

Subhorizon scales during inflation are microscopic39 and therefore quantum effects are important.
Thus we should study the inflaton field using quantum field theory.
This goes beyond the level of this course, so we have relegated the discussion into an ap-
pendix. The basic idea is that for scales that are inside horizon there are quantum fluctuations,
called vacuum fluctuations, in the inflaton field. For a homogeneous inflaton field, the Fourier
amplitudes δϕk of its perturbations would be identically zero, but analogous to a quantum har-
monic oscillator, it is not possible for them to stay there, but instead they fluctuate around this
value.
We saw in Sec. 8.6.1 that the classical solutions to the evolution of δϕk reach a constant
value after horizon exit (in the approximation H = const during horizon exit). The quantum
treatment gives that at this stage we can neglect further quantum fluctuations and treat δϕk
classically—the fluctuations “freeze”.
The final result is that well after horizon exit, k ≪ H, the Fourier amplitudes δϕk have
acquired a power spectrum
2
k3 2 H
Pϕ (k) ≡ V 2 h|δϕk | i = . (324)
2π 2π
After this we can ignore further quantum effects and treat the later evolution of the in-
flaton field, both the background and the perturbation, classically. The effect of the vacuum
fluctuations was to produce “out of nothing” the perturbations δϕk . We can’t predict their
individual values; their production from quantum fluctuations is a random process. We can
only calculate their statistical properties. Closer investigation reveals that this is a Gaussian
random process. All δϕk acquire their values as independent random variables (except for the
reality condition δϕ−k = δϕ∗k ) with a Gaussian probability distribution. Thus all statistical
information is contained in the power spectrum Pϕ (k).
The result (324) was obtained treating H as a constant. However, over long time scales, H
does change. The main purpose of the preceding discussion was to follow the inflaton pertur-
bations through the horizon exit. After the perturbation is well outside the horizon, we switch
to other variables, namely the curvature perturbation Rk , which, unlike δϕk , remains constant
outside the horizon, even though H changes. Therefore we have to use for each scale k a value of
H which is representative for the evolution of that particular scale through the horizon. That is,
we choose the value of H at horizon exit,40 so that aH = k. Thus we write our power spectrum
result as
2
k3 2 H
Pϕ (k) = V 2 h|δϕk | i = , (325)
2π 2π aH=k
to signify that the value of H for each k is to be taken at horizon exit of that particular scale.
Equation (325) is our main result from inflaton fluctuations.
8.6.4 Transfer functions

Since the inflaton fluctuations are assumed to be the origin of structure, all later perturbations
are related to the inflaton perturbations δϕk . As long as all inhomogeneities are small (“per-
39
We later give an upper limit to the inflation energy scale, i.e., V (ϕ) at the time cosmological scales exited
the horizon, V 1/4 < 1.9 × 1016 GeV. From H 2 = V (ϕ)/3MPl 2
we have H < 9 × 1013 GeV or for the Hubble length
−1 −30
H > 2.3 × 10 m. This is a lower limit to the horizon size, but it is not expected to be very many orders of
magnitude larger.
40
One can do a more precise calculation, where one takes into account the evolution of H(t). The result is that
one gets a correction to the amplitude of PR (k), which is first order in slow-roll parameters, and a correction to
its spectral index n which is second order in the slow-roll parameters.
turbations”), the relationship is linear. We can express these linear relationships as transfer
functions T (t, k), e.g.,
gk (t) = Tgϕ (t, k)δϕk (tk ) . (326)
The linearity implies several things:
1. The Fourier coefficient gk depends only on the Fourier coefficient of δϕ corresponding to

the same wave vector k, not on any other k′ .
2. The relationship is linear, so that if δϕk were, e.g., twice as big, then so would gk be.
3. The perturbations of g inherit the Gaussian statistics of δϕ.
We could also define transfer functions relating perturbations at any two different times, t
and t′ , and call them T (t, t′ , k), but here we are referring to the inflaton perturbations at the time
of horizon exit, tk , which is different for different k. Actually, by δϕk (tk ) we mean the constant
value the perturbation approaches after horizon exit in the H = const = Hk approximation.
That the transfer function depends only on the magnitude k results from the fact that
physical laws are isotropic. The transfer function of Eq. (326) will then relate the power spectra
of {gk (t)} and {δϕk (tk )} as
Pg (t, k) = Tgϕ (t, k)2 Pϕ (k) . (327)
The transfer functions thus incorporate all the physics that determines how structure evolves.
For the largest scales, k−1 ≫ 10h−1 Mpc, the perturbations are still small today, and one
needs not go beyond the transfer function. For smaller scales, corresponding to galaxies and
galaxy clusters, the inhomogeneities have become large at late times, and the physics of structure
growth has become nonlinear. This nonlinear evolution is typically studied using large numerical
simulations. Fortunately, the relevant scales are small enough that Newtonian physics is usually
sufficient.
We are now in position to put together all the results we obtained. From Eq. (309)
δϕk
Rk = −H , (328)
ϕ̄˙
so that
Hk
TRϕ (k) = − (329)
˙ k)
ϕ̄(t
and 2 2
H H H
PR (k) = Pϕ (k) = , (330)
ϕ̄˙ ϕ̄˙ 2π H=k
where we used the result (325).
This primordial spectrum is the starting point for calculating structure formation (discussed
already) and the CMB anisotropy (Chapter 9). Thus CMB and large-scale structure observations
can be used to constrain PR together with other cosmological parameters.
8.6.5 Generation of primordial gravitational waves

The quantum fluctuations at subhorizon scales during inflation apply also to the spacetime itself.
We do not yet have a complete theory of quantum gravity, so we do not know how spacetime
behaves in the Planck era. At lower energy scales the spacetime fluctuations are smaller and
for small perturbations around a FRW universe we can use the linearized equations for metric
perturbations, for which quantization is straightforward. In fact, the proper treatment of the
generation of inflaton perturbations, where we include the scalar metric perturbations in the
inflaton perturbation equation (see Eq. 308), contains also the quantum treatment of scalar
metric perturbations.
Likewise, we have quantum fluctuations of tensor metric perturbations during inflation.
These do not couple to density perturbations, but they become classical gravitational waves
after horizon exit. These primordial gravitational waves have an effect on CMB anisotropy and
polarization. √
In the quantum treatment, (MPl / 2)h fluctuates like a scalar field, so that in inflation the
gravitational wave amplitudes h acquire a spectrum
2 2
V 2 H 8 H
Ph (k) ≡ 4 2 k3 h|hk |2 i = 4 2 = 2 (331)
2π MPl 2π H=k MPl 2π H=k
(the factor 4 in this customary definition is related to the way h appears in several places in the
metric and to there being two modes for each k).
The tensor-to-scalar ratio is the ratio of the two primordial spectra (331) and (330),
2
Ph (k) 8 ϕ̄˙
r ≡ = 2 . (332)
PR (k) MPl H H=k
8.7 The primordial spectrum

8.7.1 Primordial spectrum from slow-roll inflation
The final result of the previous section is thus that inflation generates primordial scalar pertur-
bations Rk with the power spectrum
2 2 2
H H 1 H
PR (k) = = . (333)
ϕ̇ 2π H=ak 4π 2 ϕ̇ t=tk
and primordial tensor perturbations with the power spectrum
2
8 H
Ph (k) = 2 . (334)
MPl 2π t=tk
In this section ϕ and ϕ̇ refer to the background values.
Applying the slow-roll equations
V ϕ̇ ′
2 V
H2 = 2 and 3H ϕ̇ = −V ′ ⇒ = −MPl
3MPl H V
these become
1 1 V3 1 1 V
PR (k) = 6 V ′2 = 24π 2 M 4 ε
12π 2 MPl Pl
2 V
Ph (k) = 4 , (335)
3π 2 MPl
where ε is the slow-roll parameter. The tensor-to-scalar ratio is thus
Ph (k)
r ≡ = 16ε . (336)
PR (k)
According to present observational CMB and large-scale structure data, the amplitude of
the primordial power spectrum is about
PR (k)1/2 ≈ 5 × 10−5 (337)
at cosmological scales. This gives a constraint on inflation
1/4
V √ p
≈ 241/4 π 5 × 10−5 MPl ≈ 0.028MPl = 6.8 × 1016 GeV . (338)
ε
The best chance of detecting primordial gravitational waves is based on their effect on CMB.
They have not been observed so far and the present upper limit is about
r < 0.1 ⇒ Ph (k)1/2 < 1.5 × 10−5 and ε < 0.006 . (339)
This implies an upper limit to the inflation energy scale
V 1/4 ≈ ε1/4 0.028MPl < 0.008MPl = 1.9 × 1016 GeV . (340)
Since during inflation, V and V ′ change slowly while a wide range of scales k exit the horizon,
PR (k) and Ph (k) should be slowly varying functions of k. We define the spectral indices ns and
nt of the primordial spectra as
d ln PR
ns (k) − 1 ≡
d ln k
d ln Ph
nt (k) ≡ . (341)
d ln k
(The −1 is in the definition of ns for historical reasons, to match with the definition in terms of
density perturbations, see Sec. 8.7.2.) If the spectral index is independent of k, we say that the
spectrum is scale free. In this case the primordial spectra have the power-law form
ns −1 n t
2 k 2 k
PR (k) = As and Ph (k) = At , (342)
kp kp
where kp is some chosen reference scale, “pivot scale”, and As and At are the amplitudes at this
pivot scale.
If the power spectrum is constant,
P = const. , (343)
corresponding to ns = 1 and nt = 0, we say that the spectrum is scale invariant. A scale-

invariant scalar spectrum is also called the Harrison–Zeldovich spectrum.
If ns 6= 1 or nt 6= 0, the spectrum is called tilted. A tilted spectrum is called red, if ns < 1
(more structure at large scales), and blue if ns > 1 (more structure at small scales).
Using Eqs. (335) and (341) we can calculate the spectral index for slow-roll inflation.
Since P(k) is evaluated from Eqs. (333) and (334) or (335) when k = aH,
d ln k d ln(aH) ȧ Ḣ
= = + = (1 − ε)H ,
dt dt a H
where we used Ḣ = −εH 2 (in the slow-roll approximation) in the last step. Thus
d 1 1 d 1 ϕ̇ d M2 V ′ d ′
2 V d
= = = − Pl ≈ −MPl . (344)
d ln k 1 − ε H dt 1 − ε H dϕ 1 − ε V dϕ V dϕ
Let us first calculate the scale dependence of the slow-roll parameters:
" # " ′ 2 ′′ #
2 ′ 2
dε ′
2 V d MPl V 4 V′ 4 V V
= −MPl = MPl − = 4ε2 − 2εη (345)
d ln k V dϕ 2 V V V V
and, in a similar manner (exercise),
dη
= . . . = 2εη − ξ , (346)
d ln k
where we have defined a third slow-roll parameter
4 V ′ ′′′
ξ ≡ MPl V . (347)
V2
p
The parameter ξ is typically second-order small in the sense that |ξ| is of the same order of
magnitude as ε and η. (Therefore it is sometimes written as ξ 2 , although nothing forces it to be
positive.)
We are now ready to calculate the spectral indices:

1 dPR ε d V 1 dV 1 dε
ns − 1 = = = −
PR d ln k V d ln k ε V d ln k ε d ln k
′
2 V 1 dV
= −MPl · − 4ε + 2η = −6ε + 2η
V V dϕ
′
′ 2
1 dPh 2 V 1 dV 2 V
nt = = −MPl = −MPl = −2ε . (348)
Ph d ln k V V dϕ V
Since ε > 0, the tensor spectrum is necessarily red. (This follows already from (334), since H
is decreasing, or from (335) since V is decreasing.) Slow-roll requires ε ≪ 1 and |η| ≪ 1, so
both spectra are close to scale invariant. For scalar perturbations this is verified by observation.
Based on CMB anisotropy data from the Planck satellite, the Planck Collaboration [4] finds
ns = 0.965 ± 0.004 . (349)
If one were able to measure all three values ns , r, and nt from observations, one could solve
from them the slow-roll parameters ε and η and moreover, check the consistency condition
r
nt = − (350)
8
for single-field slow-roll inflation. This consistency condition is the only truly quantitative
prediction of the inflation scenario (as opposed to some specific inflation model) – all the other
predictions (Ωk very small, ns close to 1 and nt close to 0, primordial perturbations Gaussian)
are of qualitative nature, not a specific number not equal to 0 or 1.
Unfortunately, the existing upper limit to r already means that it will be difficult to ever
determine the spectral index nt with sufficient accuracy to distinguish between nt = −r/8 and
nt = 0. The most sensitive probe to primordial gravitational waves is provided by polarization of
CMB on which they will imprint a characteristic pattern (discussed briefly in the next chapter).
The theoretical limit to detection is r ∼ 10−4 and there are proposals41 for future CMB satellite
missions that could reach r ∼ 10−3 . If r is significantly larger than these detection limits, after
detection one could still measure nt accurately enough to distinguish, say, nt ≈ −1, nt ≈ 0
(which includes the case nt = −r/8), and nt ≈ 1 from each other. There have been other
proposals (other than inflation) for very-early-universe physics, which predict primordial tensor
perturbations that deviate from scale invariance this much or more.
Detection of primordial gravitational waves, i.e., measurement of r, would be enough to
determine ε and η and thus the inflation energy scale from Eq. (338).
One can also calculate the scale-dependence of the spectral index (exercise):
dns
= 16εη − 24ε2 − 2ξ . (351)
d ln k
It is second order in slow-roll parameters, so it’s expected to be even smaller than the deviation
from scale invariance, ns − 1. Planck Collaboration finds it consistent with zero to accuracy
O(10−2 ), as expected.
Cosmologically observable scales have a range of about ∆ ln k ∼ 10. Planck measured the
CMB anisotropy over a range ∆ ln k ∼ 6 (missing the shortest scales, where the CMB is expected
to have negligible anisotropy). Some inflation models have |ns −1|, r, and |dns /d ln k| larger than
the Planck results, while others do not. These observations already ruled out many inflation
models.
Example: Consider the simple inflation model
1 2 2
V (ϕ) = m ϕ .
2
In Chapter 7 we already calculated the slow-roll parameters for this model:
2
MPl
ε=η=2
ϕ2
41
See, e.g., http://www.core-mission.org/
and we immediately see that ξ = 0. Thus

2
MPl
ns = 1 − 6ε + 2η = 1 − 8
ϕ
4
dns MPl
= 16εη − 24ε2 − 2ξ = −32
d ln k ϕ
2
MPl
r = 16ε = 32
ϕ
2
MPl
nt = −2ε = −4
ϕ
To get the numbers out, we need the values of ϕ when the relevant cosmological scales exited the
horizon. The number of inflation e-foldings after that should be about N ∼ 50. We have
Z ϕ Z
1 V 1 ϕ 1
N (ϕ) = 2 ′
dϕ = 2 = 2 ϕ2 − ϕend 2 ,
MPl ϕend V MPl 2 4MPl
2
√
and we estimate ϕend from ε(ϕend ) = 2MPl /ϕend 2 = 1 ⇒ ϕend = 2MPl to get
ϕ2 = ϕend 2 + 4MPl
2 2
N = 2MPl 2
+ 4MPl 2
N ≈ 4MPl N.
Thus 2
MPl 1
=
ϕ 4N
and
2
ns = 1− ≈ 0.96
N
dns 2
= − ≈ −0.0008
d ln k N2
8
r = ≈ 0.16
N
1
nt = − ≈ −0.02
N
We see that this model is ruled out by the observed upper limit r < 0.1.42
8.7.2 Scale invariance of the primordial power spectrum

Inflation predicts and observations give evidence for an almost scale invariant primordial power
spectrum. Let us forget the “almost” for a moment and discuss what it means for the primordial
spectrum to be scale invariant.
The primordial spectrum is something we have at superhorizon scales, where we have dis-
cussed it in terms of the comoving curvature perturbation R, and we are calling it scale invariant,
when
PR (k) = A2s = const. (352)
We would like the spectrum in terms of more familiar concepts like the density perturbation,
but at superhorizon scales the density perturbation is gauge dependent.
For small scales the perturbation spectrum gets modified when the scales enter the horizon,
but for large scales k ≪ keq the spectrum maintains its primordial shape, at least as long as
42
There was enormous excitement in early 2014, when the BICEP2 collaboration[7] claimed to have detected
the effect of primordial gravitational waves with r = 0.20+0.07
−0.05 in CMB polarization, consistent with this inflation
model. However, it turned out that their data was contaminated by polarized emission from dust in our own
galaxy.[8]
the universe stays matter dominated. This allows the discussion of the primordial spectrum
at subhorizon scales, where we can talk about the density perturbations without specifying a
gauge.
From Eq. (255), the gravitational potential and density perturbation are related to the
curvature perturbation as
3
Φ k = − Rk (mat.dom)
5
(353)
2 k 2 2 k 2
δk = − Φk = Rk ,
3 H 5 H
giving
9 9
PΦ (k) = PR (k) = A2s = const (354)
25 25
4
4 k 4 4 k
Pδ (t, k) = PΦ (k) = PR (k)
9 H 25 H
4
4 k
= A2s ∝ t4/3 k4 (355)
25 H
Thus perturbations in the gravitational potential are scale invariant, but perturbations in density
are not. Instead the density perturbation spectrum is steeply rising, meaning that there is much
more structure at small scales than at large scales. Thus the scale invariance refers to the
gravitational aspect of perturbations, which in the Newtonian treatment is described by the
gravitational potential, and in the GR treatment by spacetime curvature.
The relation between density and gravitational potential perturbations reflects the nature of
gravity: A 1% overdense region 100 Mpc across generates a much deeper potential well than a
1% overdense region 10 Mpc across, since the former has 1000 times more mass. Therefore we
need much stronger density perturbations at smaller scales to have an equal contribution to Φ.
However, if we extrapolate Eq. (355) back to horizon entry, k = H, we get
2
2 4 2
δH (k) ≡ “Pδ (k, tk )” ≡ PR (k) = As = const (356)
25 5
Thus for scale-invariant primordial perturbations, density perturbations of all scales enter the
horizon with the same amplitude, δH = (2/5)As ∼ 2 × 10−5 . Since the density perturbation
at the horizon entry is actually a gauge-dependent quantity, and our extension of the above
Newtonian relation up to the horizon scale is not really allowed, this statement should be taken
just qualitatively (hence the quotation marks around the Pδ ). As such, it applies also to the
smaller scales which enter during the radiation-dominated epoch, since the perturbations only
begin to evolve after horizon entry.
What is the deep reason that inflation generates (almost) scale invariant perturbations?
During inflation the universe is almost a de Sitter universe, which has the metric
ds2 = −dt2 + e2Ht (dx2 + dy 2 + dz 2 )
with H = const . In GR we learn that it is an example of a “maximally symmetric spacetime”.
In addition to being homogeneous (in the space directions), it also looks the same at all times.
Therefore, as different scales exit at different times they all obtain the same kind of perturbations.
In terms of the other definition of the power spectrum, P (k) ≡ (2π 2 /k3 )P(k), the relations
(355) for scale-invariant perturbations give
PR (k) ∝ k−3 PR ∝ k−3
(357)
Pδ (k) ∝ k−3 Pδ ∝ k4 PR ∝ kPR ∝ k
For PR (k) ∝ kn−1 we have Pδ (k) ∝ kn . This is the reason for the −1 in the definition of the
spectral index in terms of PR —it was originally defined in terms of Pδ .
8.8 The power spectrum today

8.8.1 Density perturbations
From Eq. (261), the density perturbation spectrum at late times is
4
4 k
Pδ (k) = T (k, t)2 PR (k) (358)
25 H
where, from Eq. (263)
T (k) = 1 for k ≪ keq

keq 2
T (k) ∼ ln k for k ≫ keq .
k
Thus the present-day density power spectrum rises steeply ∝ k4 at large scales, but turns at
∼ keq to become less steep (growing ∼ ln k) at small scales. This is because the growth of density
perturbations was inhibited while the perturbations were inside the horizon during the radiation-
dominated epoch. The ∼ ln k factor comes from the slow growth of CDM perturbations during
this time.
−1 ∼ 100 Mpc.
Thus the structure in the universe appears stronger at smaller scales, down to keq
The ∼ 100 Mpc scale is indeed quite prominent in large scale structure surveys, like the 2dF-
GRS and SDSS galaxy distribution surveys. Towards smaller scales the structure keeps getting
stronger, but now more slowly. However, perturbations are now so large that first-order pertur-
−1
bation theory begins to fail, and that limit is crossed at around k−1 ∼ knl ∼ 10 Mpc. Nonlinear
effects cause the density power spectrum to rise more steeply than calculated by perturbation
theory at scales smaller than this.
The present-day density power spectrum Pδ (k) can be determined observationally from the
distribution of galaxies (Fig. 7). The quantity plotted is usually Pδ (k). It should go as
Pδ (k) ∝ kn for k ≪ keq

n−4
(359)
Pδ (k) ∝ k ln k for k ≫ keq .
See Fig. 8.
8.8.2 Primordial gravitational waves

We found that outside the horizon tensor perturbations remain constant,
hk (t) = hk,prim = const , (360)
whereas inside the horizon they become gravitational waves whose amplitude decays
|hk (t)| ∝ a−1 . (361)
Define the transfer function for gravitational waves
|hk (t0 )|
Th (k) ≡ , (362)
hk,prim
Figure 6: The whole picture of structure formation theory from quantum fluctuations during inflation
to the present-day power spectrum at t0 .
so that the present-day power spectrum of primordial gravitational waves is
Pgrav (k, t0 ) = Th (k)2 Ph (k) . (363)
Make the approximation that the transition from (360) to (361) is instantaneous at horizon
entry defined as
k = H = aH . (364)
Denote these values of a, H, and H by ak , Hk , and Hk . Then
ak
Th (k) = = ak . (365)
a0
The shape of the transfer function is determined by the rate at which different comoving scales
k enter horizon as the universe expands. This is determined by the evolution of the comoving
Hubble distance H−1 .
In the matter-dominated universe
2
a ∝ t2/3 and H = ∝ a−3/2 ⇒ H ∝ a−1/2 . (366)
3t
Make first the approximation that the universe is still matter dominated. Then
−2 −2
ak Hk k
Th (k) = = = (H0 < k < keq ) (367)
a0 H0 a0 H 0
for scales that entered during the matter-dominated epoch.

To correct this result for the effect of dark energy at late times, we note that because of dark
energy, the comoving Hubble distance H−1 = (aH)−1 stopped growing and began to shrink,
Figure 7: Distribution of galaxies according to the Sloan Digital Sky Survey (SDSS). This figure shows
galaxies that are within 2◦ of the equator and closer than 858 Mpc (assuming H0 = 71 km/s/Mpc).
Figure from astro-ph/0310571[9].
10
1
2
k P(k) / 2π
0.1
3
0.01
0.001
0.0001
0.01 0.1 1
-1
k [Mpc ]
1e+05
P(k) [Mpc ]
3
10000
1000
100
0.01 0.1 1
-1
k [Mpc ]
Figure 8: The matter power spectrum from the SDSS obtained using luminous red galaxies [10]. The
top figure shows Pδ (k) and the bottom figure Pδ (k). A Hubble constant value H0 = 71.4 km/s/Mpc has
been assumed for this figure. (These galaxy surveys only obtain the scales up to the Hubble constant,
and therefore the observed Pδ (k) is usually shown in units of h Mpc−1 , so that no value for H0 need
to be assumed.) The black bars are the observations and the red curve is a theoretical fit, from linear
perturbation theory, to the data. The bend in P (k) at keq ∼ 0.01 Mpc−1 is clearly visible in the bottom
figure. Linear perturbation theory fails when P(k) & 1, and therefore the data points do not follow the
theoretical curve to the right of the dashed line (representing an estimate on how far linear theory can
be trusted). Figure by R. Keskitalo.
REFERENCES 91
so that the scale k = H0 is actually exiting now, and it entered at an earlier time t1 when the
expansion was still (barely) matter dominated. Thus the above result for Th (k) should apply
(roughly) at that earlier time:
−2 −2
k k
Th (t1 , k) = = (H0 < k < keq ) (368)
a1 H1 a0 H 0
While the scale k = H0 was inside the horizon, the universe expanded by about a factor of two,
so the correct transfer function is about half of (367).
Exercise: Extend the result (367) to scales k > keq . You can make the approximation where the
transition from radiation-dominated expansion law to matter-dominated expansion law is instantaneous
at teq . (This approximation actually underestimates Th (k > keq ) by a factor that roughly compensates
the overestimation in (367) from ignoring dark energy at late times.)
Gravitational waves were detected for the first time on September 14, 2015 at the LIGO observatory.
These were not primordial gravitational waves; they were caused by a collision of two black holes about 400
Mpc from here, and they were observed only for about 0.2 seconds. The peak amplitude was h ≈ 10−21 .
LIGO is sensitive to frequencies near 100 Hz, and with further refinements it is expected to reach a
sensitivity of h = 10−22 . Assume the primordial tensor perturbations had amplitude h = 10−5 (close to
the upper limit from CMB observations). What is their amplitude today at the 100 Hz frequency?
ESA is planning to launch a space gravitational wave observatory (LISA) in 2034. It would have
similar sensitivity as LIGO, but for frequencies lower by a factor 10−4 . What do you conclude about the
prospect for observing primordial gravitational waves this way?
References
[1] J.A. Peacock: Cosmological Physics (Cambridge University Press 1999), Chapter 16
[2] A.R. Liddle and D.H. Lyth: Cosmological Inflation and Large-Scale Structure (Cambridge
University Press 2000)
[3] Planck Collaboration, Astronomy & Astrophysics 594, A13 (2016), arXiv:1502.01589
[4] Planck Collaboration, Planck 2018 results. VI. Cosmological parameters, arXiv:1807.06209
[5] S. Dodelson: Modern Cosmology (Academic Press 2003), Chapter 7
[6] E.W. Kolb and M.S. Turner: The Early Universe (Addison-Wesley 1990)
[7] P.A.R. Ade et al., Phys. Rev. Lett. 112, 241101 (2014), arXiv:1403.3985
[8] Planck Collaboration, Astronomy & Astrophysics 586, A133 (2016), arXiv:1409.5738
[9] J. Richard Gott III et al., A Map of the Universe, Astrophys. J. 624, 463 (2005), astro-
ph/0310571
[10] M. Tegmark et al., Cosmological Constraints from the SDSS Luminous Red Galaxies, Phys.
Rev. D74, 123507 (2006), astro-ph/0608632

8 Structure Formation: 8.1 Inhomogeneity

Uploaded by

Copyright:

Available Formats

8 Structure Formation: 8.1 Inhomogeneity

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

8 Structure Formation: 8.1 Inhomogeneity

Uploaded by

Copyright:

Available Formats

8 Structure Formation

ρ(t, x) = ρ̄(t) + δρ(t, x)

and then we will often write just δ for δm .

8.1.1 Statistical homogeneity and isotropy

8.1.2 Density autocorrelation function

ξ(r) ≡ hδ(x)δ(x + r)i . (11)

ξ(r) = ξ(r) . (12)

hδ2 i ≡ hδ(x)δ(x)i ≡ ξ(0) . (13)

8.1.3 Fourier space

where the wave vectors k = (k1 , k2 , k3 ) take values

8.1.4 Power spectrum

Written in terms of δ(k) = V δk , the result (29) reads as

hδ(k)∗ δ(k′ )i = V δkk′ P (k) → (2π)3 δD

From statistical isotropy

ξ(r) = ξ(r) ⇒ P (k) = P (k) (34)

(the 3D Fourier transform of a spherically symmetric function is also spherically symmetric),

where we have defined

8.1.5 Scales of interest and window functions

The simplest window function is the top-hat window function

δ(k, R) = δ(k)W (k) , (43)

3(sin kR − kR cos kR)

We can also define the k-space top-hat window function

Wk (k) ≡ 1 for k ≤ 1/R (45)

and Wk (k) = 0 elsewhere. In x-space this becomes (exercise)

8.1.6 Power-law spectra

ξ(r) ∝ r −γ or P (k) ∝ kn . (49)

For P (k) we may write

8.1.7 Galaxy 2-point correlation function

dP1 = hn(x)idV1 = hn̄ + δn(x)idV1 = n̄dV1 . (60)

The probability of finding a galaxy pair at x and x + r is

dP2 = n̄ [1 + hδ(x)δ(x + r)i] dV2 = n̄ [1 + ξ(r)] dV2 . (62)

8.2 Newtonian perturbation theory

2) The fluid flow is nonrelativistic, u ≪ c ≡ 1.

3) We are considering nonrelativistic matter, |p| ≪ ρ

A homogeneously expanding fluid,

ρ = ρ(t0 )a−3 (71)

8.2.1 Comoving coordinates

8.2.2 The perturbation

ρ(t′ , r) = ρ̄(t) + δρ(t′ , r) (80)

In terms of the comoving coordinates these become (exercise):

In terms of the relative density perturbation δ ≡ δρ/ρ̄ we have δρ = ρ̄ · δ and

and we can write

Finally, we Fourier expand the perturbations,

In Fourier space the perturbation equations become

2) Solve the perturbation equations.

8.2.3 Vector and scalar perturbations

8.2.4 The equations for scalar perturbations

8.2.5 Adiabatic and entropy perturbations

Perturbations with the property

8.2.6 Adiabatic perturbations in matter

δ̈k + 2H δ̇k − 4πGρ̄δk = 0 . (119)

Thus density perturbations in matter grow proportional to the scale factor.

Φ ∝ a2 ρ̄δ ∝ a2 a−3 a = const.

8.2.7 Many fluid components

ρ̇i + 3Hρi = 0 , (128)

and the expansion law

ρi (t′ , r) = ρ̄i (t) + δρi (t′ , r) (130)

but there is only one gravitational potential perturbation,

in Newtonian coordinate space, and