PHAS0067: Advanced Physical Cosmology

13 Relativistic Structure Formation

()

While the Newtonian limit is helpful for many applications, to fully understand the origin and evolution of structure in the universe we need to turn to GR, our best theory of gravity. We can then show where Newtonian approximations are valid and derive new results for the cases where the Newtonian approximations fail.

In this section we’ll look at everything required to understand how perturbations evolve from deep in the inflationary era through to the present-day universe. Having done that, we’ll need a final ingredient – quantum mechanics – which we’ll cover in Section 14 before putting everything together.

13.1 Setting up relativistic perturbation theory; the curse of gauge

Relativistic perturbation theory supposes that the metric of spacetime can be decomposed into a background piece g¯μν and a perturbation δgμν:

ds2=(g¯μν+δgμν)dxμdxν. (337)

Just as we did in the Newtonian case, we take g¯μν to be the solution to Einstein’s equations for a homogeneous universe, then require δgμν to be small and find its approximate solution.

📝 Exercise 13A Show that, at linear order in δgμν, the inverse metric gμν may be written gμν=g¯μν-δgμν (338) where δgμν=g¯μαg¯νβδgαβ. Note particularly the sign change in δg when raising its indices (omitting the minus sign is a common source of errors in perturbation theory derivations).

Relative to the Newtonian case, relativistic perturbation theory suffers from added complexity because:

  1. 1.

    There are, at first glance, ten variables in the metric perturbation δgμν (compared to just the single Newtonian potential Φ).

  2. 2.

    There are conceptual complications arising from gauge freedom, which in turn arises from the lack of a preferred Newtonian frame.

Gauge freedom2626 26 The word “gauge” is unhelpfully cryptic, and apparently derives from the problem of railway gauge (i.e. width), as follows: there’s nothing intrinsically better about a railway with rails separated by 4ft 1in rather than 4ft. But to make a coherent railway system, everyone has to choose the same width. Similarly, there’s nothing necessarily better about one “gauge choice” over another in physics, but the choice still needs to be made consistently throughout a given calculation. refers to the fact that two apparently different solutions can, in fact, describe the same physical behaviour. Electromagnetism is the most familiar theory with gauge freedom: for example, the electrostatic potential ϕ can be replaced with ϕ+ϕ0 for some constant ϕ0 – and there is no physical difference in the behaviour of the system described.

Figure 14: An illustration of gauge ambiguity in general relativity. Here I’ve taken a picture where there’s an overall trend in the lightness, but also some small fluctuations superimposed. Should these be decomposed as a straight gradient plus perturbations (upper row) or as a curved gradient plus slightly different perturbations (lower row)? Either is valid, so we just have to make a choice and stick to it for analysis.

In general relativistic perturbation theory, gauge effects arise because there is more than one way to define a background (homogeneous) metric g¯μν for any given perturbed (inhomogeneous) metric gμν. Figure 14 tries to make this clearer. How should the actual picture be decomposed into a smooth background and perturbations? There is more than one way to do it. In Newtonian perturbation theory, the choice is very natural because Newton’s theory contains a notion of absolute time. We average along slices of fixed time to get the “background”. But in Einstein’s theories, there is no definitive time to reference, so this option is no longer open.

It might seem like the situation is hopeless, but that’s not really the case. We just have to be extra careful about being self-consistent in developing our perturbation theory and distinguishing physical predictions (which must be independent of the gauge choice) from purely mathematical statements (which may be gauge-dependendent). There are essentially three ways to do this:

  1. (i)

    1+3 gauge-invariant covariant perturbation theory constructs the entire dynamics of the spacetime in terms of physically observable quantities from the outset. This is very satisfying and beautiful, but is quite unwieldy for most practical purposes. We won’t use it for this course.2727 27 But if you’re interested, take a look at these very well-written lecture notes that pursue the approach: http://arxiv.org/pdf/gr-qc/9812046v5.pdf.

  2. (ii)

    Gauge-invariant perturbation theory (sometimes known as “Bardeen” perturbation theory) starts from gauge-dependent quantities, identifies combinations of these variables that are gauge-independent, then works with those combinations alone. This is closely related to the actual option we’ll pursue which is (iii):

  3. (iii)

    Gauge-fixed perturbation theory decides from the outset to work in a specified gauge. This is simplest, but comes with dangers. For example, one can easily over-interpret the resulting gauge-specific equations and come to misleading conclusions. With sufficient care this is not a huge concern and we will not particularly discuss it further. Just bear in mind that, for complete safety, one should always construct physically observable quantities. The density contrast δ, for example, is not strictly observable. An example of mapping the perturbation variables onto something observable will be given when we discuss the cosmic microwave background in Section 15.

One possible approach to gauge fixing is to realise that there is, after all, a uniquely sensible definition of cosmic time – specifically, the time measured by clocks that are freefalling. The gauge implied by this choice is known as the synchronous gauge2828 28 Usually the freefalling clocks are assumed to coincide with the motion of dark matter on large scales, in which case strictly we have the comoving synchronous gauge. To implement the idea, we have to choose a coordinate system where our chosen family of free-falling particles have x˙μ=uμ=(1,0,0,0) everywhere, meaning that their spatial coordinates never change:

😇 Exercise 13B Suppose the background metric g¯μν is given by equation (23), i.e. the standard flat FRW metric in cartesian coordinates. Show, by writing the geodesic equation (38) for dxμ/dλ=uμ to first order in δg, that the perturbation to the metric in synchronous gauge must obey δg˙00=0 and 2δg˙0i-δg00,i=0. (339) The simplest solution is δg00=δg0i=0.

What’s rather nice about this conclusion is that already we have reduced the number of degrees of freedom in the perturbed metric from 10 to 6 by removing the time-time and space-time parts. This eliminates part of the extra complexity of GR over Newtonian perturbation theory, but we can go even further.

13.2 Scalar-vector-tensor decomposition

After adopting the synchronous gauge, the GR gravitational field is represented by a symmetric 3×3 matrix with 6 independent components at each point in space. But Newtonian theory has just one number – the potential – at each point in space. What is all this extra complexity doing, and how can we tame it?

It helps to work in Fourier space from the outset. That is, we take each slice of the universe at fixed conformal time η and express the metric perturbation δgab(η,𝒙) in terms of a superposition of fourier modes in the background comoving coordinates:

δgab(η,𝒙)=d3𝒌(2π)3/2hab(η,𝒌)ei𝒌𝒙a2. (340)

Note that I’ve included a factor of a2 which makes some of the later equations a bit simpler. In the new view, the metric perturbations hab are a function of fourier wavemode 𝒌 and conformal time η. To get evolution equations for hab we will always consider just one mode at a time, i.e. a fixed 𝒌; because we are considering linear perturbations, the solution for the general case will just be the sum over all wavemodes. Just as with δgμν, in synchronous gauge h00=h0i=0, so only the spatial part hab need be considered.

So, let’s consider the behaviour of one fourier mode. Take it along the z axis, so that 𝒌=(0,0,k) to be definite; obviously the coordinate system can be rotated to make this the case for any single initial 𝒌. There is still a residual freedom in the orientation of the coordinate system: we can rotate around the z axis without changing 𝒌, so whatever conclusions we draw about the physics of hab must be left invariant when we do so. We start by classifying the different 3×3 symmetric matrices according to their transformation under this rotation:

Exercise 13C Write down a 3×3 rotation matrix 𝐑(θ) for angle θ around the z axis. Recalling that any matrix 𝐌 transforms as 𝐑𝐌𝐑, calculate the transformation under 𝐑(θ) of the following: 𝐌S1 =(100010000), 𝐌S2 =(000000001), 𝐌V1 =(001000100), 𝐌V2 =(000001010), 𝐌T1 =(010100000), 𝐌T2 =(1000-10000). (341) Convince yourself also that any symmetric matrix 𝐡 can be written as ihi𝐌i.

You should have found in the exercise above that the S1 and S2 matrices above do not change under a rotation around the z axis, which is the direction of propagation of our example wave. So waves proportional to S1 or S2 are called scalar degrees of freedom. Like you’d expect for vectors, V1 and V2 matrices return to their original form once θ reaches 2π, and for this reason they are called vector degrees of freedom. The T1 and T2 matrices return to their original form once θ reaches π, and so they’re neither vectors nor scalars; they are called tensor degrees of freedom.

Our metric perturbation hab for a single mode 𝒌 can always be uniquely written as a superposition of all six 𝐌 matrices above. By extension, we can uniquely decompose a metric into scalar, vector and tensor parts.

Still considering just a single mode, we can begin to see a glimmer of how the GR perturbation theory connects to the Newtonian perturbation theory:

  • The scalar part (2 degrees of freedom per mode) drives structure formation and is common to GR and Newtonian theories;

  • The vector part (2 degrees of freedom per mode) separates out more naturally in GR, but is also present in Newtonian theory in the form of fluid vorticity;

  • The tensor part (2 degrees of freedom per mode) is unique to GR and in fact corresponds to gravitational waves which do not appear in the Newtonian theory.

It’s important that all three parts evolve independently at linear order in perturbation theory, i.e. they do not affect each other. This can be explicitly demonstrated by deriving equations of motion for the superposition of all three types, though it’s actually a matter of logical consistency: all terms in an equation must transform in the same way under a rotation if the underlying physical law does not have a preferred direction. So the linear equations for the three types of perturbation simply must be independent of each other. By a similar argument, the classification of mode types must be gauge-invariant.

Having gauge-fixed, to work out how a given perturbation will evolve in GR is very similar to the Newtonian analysis. One expands all the evolution equations to first order in the perturbed quantities. The difficulty now is the sheer complexity of that manipulation; remember that the Einstein equations dicate the evolution and these involve the Ricci scalar and tensor, in turn calculated from the Riemann tensor, which itself is calculated from the Christoffel symbols, which have to first be calculated from derivatives of the perturbed metric…

Luckily this exercise only has to be done once to give you the evolution equations. In this course we’ll simply quote the results for the Einstein tensor without going through the intermediate derivations.

13.3 Sub- and super-horizon modes

Before we derive equations that reveal the behaviour of perturbations in the Universe, it is helpful to discuss one piece of terminology: sub- and super-horizon.

In Section 1 we introduced the idea of two objects being either in causal contact or causally disconnected, according to whether their comoving distance is less than or greater than the comoving horizon respectively. When objects are causally disconnected, they cannot affect each other’s behaviour.

We can therefore anticipate that the evolution of a Fourier mode in the Universe will depend greatly on its wavelength and classify the modes accordingly:

  • If the wavelength is very large compared to the causal horizon, the mode is called “super-horizon”.

  • If the wavelength is very small compared to the comoving horizon, the mode is called “sub-horizon”.

Note that the horizon expands with time (at least, if we temporarily ignore inflation), and therefore a mode that is super-horizon at one moment may become sub-horizon at a later time.

It is useful to formulate what we mean by sub- and super-horizon in mathematical terms. Since the conformal time η can be regarded as the comoving horizon scale, we might say that sub-horizon means the comoving wavelength λ obeys λη; conversely super-horizon could mean λη. Because we more typically work with k=2π/λ, we normally reformulate this as:

Sub-horizon: |kη|1Super-horizon: |kη|1. (342)

You may have noticed that, technically, a factor of 2π has gone missing in the conversion between wavelength λ and wavenumber k; this has been absorbed into the idea of ”much larger” and ”much smaller”, where factors 6 are supposed to be negligible. Of course there is a broad “in between” range of scales where kη is comparable to unity; in that case, the mode is neither sub- nor super-horizon.

There’s another complication in interpreting the above definitions. If inflation took place, the true comoving horizon is vastly greater than η would suggest – recall we set the zero point of η to be the beginning of radiation domination for convenience. But if we were to try setting the zero-point of η more carefully, it’s not likely to be what we care about; things being in contact a long time ago is unlikely to mean they are physically influencing each other today.

The way around both problems is to think back to Section 2 where we defined the idea of the Hubble radius (aH)-1 – this is the comoving distance that is in recent causal contact. We can redefine our notion of sub-horizon and super-horizon to look at this ’recent communication’ scale rather than the true horizon:

Sub-horizon: k/(aH)1Super-horizon: k/(aH)1. (343)

Now refer back to Table 2. There we saw that, in radiation domination, matter domination or even during inflation, aH|η-1|. Again working on the principle that factors of proportionality can be lost in the definition of ‘much greater than’, conditions (343) are actually equivalent to (342), despite being motivated by a subtly different comparison.

The upshot: we can use conditions (343) and (342) interchangably as a good translation of the intuitive meaning of sub- and super-horizon limits. The simplicity of these conditions hides away sloppiness in the language, but that sloppiness is rarely a problem in practice.

Practically speaking, the behaviour of perturbations that we care about does change quite significantly depending on the magnitude of |kη|. The time at which |kη|=1 is known as horizon crossing – or horizon entry, assuming the mode is coming into the horizon (as in the radiation or matter domianted universes). But recall that during inflation, modes are transported out of the Hubble radius; in this case the transition goes in the opposite direction and the value of η for which |kη|=1 is known as horizon exit.

13.4 Evolution equations for tensor perturbations

Tensor perturbations, also known as gravitational waves, have no equivalent in Newtonian theory. We start with them because their evolution is actually far simpler to calculate than for the scalar perturbations.

A gravitational wave of comoving wavenumber k propagating along the z^ direction has metric perturbation of the form

δgabhab(η)a2=A(η)eikzMaba2, (344)

where Mij is some fixed linear combination of 𝐌T1 and 𝐌T2 from above – like a “template” perturbation to the metric, which is then modulated in time and space by A(η) and eikz to form the final perturbation δgab. Inserting this perturbation into the Einstein equations, we obtain

Einstein00: 8πGδT00 =δG00=0 (345a)
Einsteini0: 8πGδTi0 =δGi0=0 (345b)
Einsteinji: 8πGδTji =δGji=-1a2(A′′+2A+k2A)Mji (345c)

where primes denote derivatives with respect to conformal time η (recall /η=a/t) and =a/a=aH is the conformal Hubble parameter.

These equations provide us with an immediate example of how perturbations of different types can’t mix. We have started with a tensor perturbation, so the time-time and time-space components of the Einstein equation (which transform as scalars and vectors respectively under rotations around the z^ axis) have to vanish. Assuming the matter content of the universe to be a perfect fluid, the tensor part of δTji also vanishes at linear order, so the entire Einstein equations boil down to the background evolution plus the new equation governing gravitational waves:

A′′+2A+k2A=0. (346)

The solution of this equation depends on the dominant content of the universe through (η). For radiation domination, =aH=1/η (Table 2) and you can verify that the solution is

A1ηsinkη or 1ηcoskη (radiation domination). (347)

Since ηa, the overall behaviour described by equation (347) is of an oscillation with a decaying amplitude a-1. For matter domination, =2/η and we have the more complicated solution

Acoskη+kηsinkηη3 or sinkη-kηcoskηη3 (matter domination). (348)

For sub-horizon modes kη1 [see Eq. (342)], and the behaviour is just like that in the radiation-dominated universe – oscillation with a decaying amplitude a-1 (recalling that ηa1/2 in matter domination).

Figure 15: The solution for the gravitational wave amplitude A as a function of kη; the right panel is a zoom-in of the shaded region in the left panel. For η<0 and η>0 respectively I’ve plotted the inflationary and radiation-dominated solutions. These match smoothly at η=0, which is the time of reheating. During inflation, initially sub-horizon fluctuations oscillate then freeze out to a fixed value at kη-1 (left red line) when the horizon shrinks below their wavelength. Later, as the horizon shrinks again so that kη+1 (right red line), the mode is “unfrozen” and starts to decay and oscillate. Note that the physical time in between the two red lines can be the entire age of the universe for those gravitational waves just re-entering the horizon today!

During inflation, =-1/η (η is negative and approaches 0 from below as a). The solution is then

Acoskη+kηsinkη or kηcoskη-sinkη (slow-roll inflation). (349)

The second of these two solutions decays as η0, so is not relevant to the universe after inflation ends. But the first tends to a constant, which means that primordial gravitational waves can be generated and survive through to the present day. By matching the first solution of equation (349) onto the first solution of (347), we obtain a history of a gravitational wave that is generated during inflation (η<0) and survives through to the present-day universe (η>0) – see Figure 15.

13.5 Conformal Newtonian gauge for scalar perturbations

We are about to look at scalar perturbations, which are the most important for structure formation in our Universe. However it will be convenient to do that not in synchronous gauge but in conformal Newtonian gauge (sometimes just called Newtonian gauge), because it will make the connection to Newtonian perturbation theory clearer.

To derive and understand this gauge, let’s think a bit more about the scalar modes discussed above. Unlike the vector and tensor matrices, where for example 𝐌V1𝐌V2 under a suitable rotation — and so V1 and V2 are essentially equivalent — the 𝐌S1 and 𝐌S2 matricies cannot be transformed into each other. There are genuinely two separate scalar degrees of freedom in a GR gravitational field, compared with just one in Newtonian physics. Where has this extra freedom come from?

One way to think about it is that Newtonian time is absolute whereas GR time can be distorted, providing the extra degree of freedom. This is obscured in the synchronous gauge where we’ve removed the time freedom by anchoring everything to free-falling particles. The definition of the conformal Newtonian gauge is that the two scalar degrees of freedom appear as

ds2=a2(η)[(1+2Φ(𝒌,η)ei𝒌𝒙)dη2-(1-2Ψ(𝒌,η)ei𝒌𝒙)δijdxidxj] (350)

again for one Fourier mode (in general there will be a superposition of many such modes). One advantage of the Newtonian gauge is that, provided only scalar modes are present, the space is locally isotropic. This contrasts with the synchronous case where there is a preferred direction defined by the freefalling test particles.

We will shortly see that for perfect fluids and scalar fields, the two potentials Φ and Ψ are actually equal; a single Newtonian potential re-emerges. This is why the Newtonian gauge is so useful and for our analysis of scalar perturbations we will stick to it.

Just in case it’s not yet clear: the same physics can be expressed in the synchronous gauge. The perturbation to the dη2 part of the metric will apparently disappear, but the physics encoded in Φ will turn up elsewhere in the metric. We don’t have time to do it in this course, but this relationship between gauges can be explicitly written down without too much difficulty. One uses a coordinate transformation in the perturbed universe to change the implied mapping onto the homogeneous background. The new perturbations are then given in terms of the old ones plus a correction due to the coordinate transformation. An extensive discussion of the relationship between Newtonian and synchronous gauges can also be found at http://arxiv.org/pdf/astro-ph/9401007v1.pdf

13.6 Scalar perturbations

Let’s now look at scalar peturbations in the conformal Newtonian gauge. These are much more messy, mainly because they couple to the matter in a non-trivial way. We get three equations from the time-time, space-time, and space-space Einstein equations:

Einstein00: 4πGa2δT00 =-k2Ψ-3(Ψ+Φ) (351a)
Einsteini0: 4πGa2δTi0 =iki(Ψ+Φ) (351b)
Einsteinji: 4πGa2δTji =-[Ψ′′+(2Ψ+Φ)+(2+2)Φ-12k2(Φ-Ψ)]δij-12(Φ-Ψ)kikj (351c)

Don’t panic! These nasty-looking equations will boil back down to something manageable faster than you might imagine. Before moving on, let’s just recap exactly what we are looking at above. In each case the right-hand-side is the perturbation to (half) the Einstein tensor Gβα/2, calculated through the long process described at the end of Section 2, and the left-hand side is the perturbation to (4πG times) the energy-momentum tensor Tβα which has not yet been expanded into more useful quantities like density and so on (we’ll do this below). Be aware that both sides are gauge-dependent, and in particular the right-hand-side assumes the conformal Newtonian gauge.

13.7 Evolution of scalar perturbations in a perfect fluid

Thankfully, calculating the perturbation to the energy-momentum tensor is much easier than for the Einstein tensor. Let’s do it now for a perfect fluid. We’ll then be ready to boil things down to a single equation that tells us about the evolution of perturbations, just as equation (320) does for the Newtonian theory.

😇 Exercise 13D Let’s calculate the perturbation to the energy-momentum tensor. First remind yourself (Exercise 1) that for a homogeneous universe, the background 4-velocity U¯α must have components U¯α=(a-1,0,0,0) and hence that U¯α=(a,0,0,0). Next show that, to linear order, δU0=-a2δU0. (Hint: remember both Uα and U¯α should be unit-normalised, and do not allow yourself to assume δUα can have its index lowered by a metric — it can’t, because it’s the difference between tensors that live in spaces with different metrics.) Recall that a perfect fluid has energy-momentum tensor given by equation (89). Use the result you derived above to show that, to first order, δT00 =δρ; (352a) δTi0 =1a(ρ¯+P¯)δUi; (352b) δTji =-δPδji. (352c)

With these in hand, we can put everything together to produce the perturbation equations for fluids in a relativistic universe. First, combining eqs. (351c) and (352c), we immediately notice that any off-diagonal terms (ij) must vanish. This means

kikj(Φ-Ψ)=0       (ij). (353)

We are not interested in the case 𝒌=0 (which corresponds to a homogeneous perturbation that can be absorbed into a change in the background). So we can read off Φ=Ψ, which immediately makes everything simpler because we have only one potential to worry about.

Putting together the full set of Einstein equations we now get:

-k2Φ-3(Φ+Φ) =4πGa2δρ; (354a)
iki(Φ+Φ) =4πGa(ρ¯+P¯)δUi; (354b)
Φ′′+3Φ+(2+2)Φ =4πGa2δP. (354c)

To keep things as simple as possible, let’s assume we have a single fluid2929 29 We will discuss generalisation to multiple fluids in Section 2 with equation of state P=wρ. Then we also have δP=wδρ and can combine equations (354a) and (354c) to give:

Φ′′+3(1+w)Φ+wk2Φ+[2+(1+3w)2]Φ=0. (355)

The term in square brackets vanishes:

😇 Exercise 13E Verify that 2+(1+3w)2=0 for this case, for example by starting from the Friedmann equation (recalling also that ρ¯a-3(1+w)).

Accordingly, the equation starts to look much more tractable:

Φ′′+3(1+w)Φ+wk2Φ=0. (356)

Note the similarity to equation (346) for the amplitude of a gravitational wave; although there are some key differences in the specific factors that appear, you might (rightly) expect that a similar story to the one presented in Figure 15 can be assembled for scalar perturbations. We’ll put off examining this in detail until Section 14.

Relationship between Newtonian and GR evolution

Equation (356) doesn’t on the face of it look much like the Newtonian expression (321), but they actually agree for the case where pressure is insignificant, w=0, as the next exercise shows:

Exercise 13F Take the Newtonian expression (321) from earlier, and transform it into conformal time η to show that δ′′+δ+(wk2-322)δ=0. (357) Recall that cs2=w and we are only working in a flat, matter-dominated universe so you can apply the Friedmann equation for that case. Now consider only cold dark matter, so w=0. Use the relationship (324) between δ and the Newtonian potential ϕN to show that, ϕN′′+3ϕN=0. (358) Note that this is equivalent to the GR result for the evolution of Φ when w=0, equation (356).

When w0, the derivation above doesn’t go through and in general the evolution of Newtonian and GR cosmologies is different. For example, radiation has w=1/3; the particles are moving around at the speed of light so relativistic effects are important. Obviously, we then prefer the GR equations over the Newtonian ones.

What’s nice about the situation is that – when talking about the evolution of cold dark matter – we can just take the Newtonian perturbation results and know that they’re justified by the GR derivation.

Enthusiasts only: subtleties in the relationship between Newtonian and GR evolution

Non-examinable: Let’s discuss two subtelties about the result above. We’ve shown that, for a CDM universe, the linear-order Newtonian potential ϕN evolves identically to the linear-order GR potential Φ in the Newtonian gauge. But the Poisson equation (324) does not directly apply… rather, the correct linear-order overdensity is given by solving equation (354a). It’s like a Poisson equation with the extra term -3(Φ+Φ). The evolution of the potential is fine, as we’ve seen, but if the Poisson equation has been modified does the density do the same as implied by the Newtonian equations?

The answer is ‘yes’. There are two ways to understand why:

  1. 1.

    (The simpler but incomplete answer) Provided that 2Φ and Φ are both much less than k2Φ, we can ignore them. This implies that we require k22 which is nothing more than saying the wave under consideration must have a wavelength much smaller than the horizon. You can verify by using the equations of motion that Φk2Φ in this case too. Bottom line: provided the perturbation is sufficiently inside the horizon, we can pretend the Newtonian relationship between potential and density continues to hold.

  2. 2.

    (The more complicated but complete answer) Equation (354b) tells us that the offending term can be rewritten:

    3(Φ+Φ)=12πG(ρ¯+P¯)aiδUikik2 (359)

    Instead of regarding the “correction” to the Poisson equation as part of the gravitational field, we can therefore take it to the other side of equation (354a) and regard it as a correction to the density perturbation:

    -k2Φ=4πGa2[δρ+3(ρ¯+P¯)iδUikiak2]. (360)

    Be clear: we haven’t done anything magical (yet). We’ve just substituted one part of the Einstein equations into another part, which has the effect of changing the original garbage related to curvature into some new garbage related to energy-momentum.

    Here’s the clever bit: remember that δρ is a gauge-dependent quantity. Depending on how we choose the background, δρ can look quite different. It turns out that

    δρsynchronous=δρ+3(ρ¯+P¯)iδUikiak2, (361)

    where the left hand side is the synchronous gauge density perturbation and the right hand side is calculated in the Newtonian gauge3030 30 This equation comes from understanding the gauge transformation in the way that we very briefly described at the end of Section 5. For more information see the discussion leading up to Dodelson equation 5.79, which is the same as our (361). .

    We can therefore write

    -k2Φ=4πGa2δρsynchronous (362)

    and we have recovered something that looks identical to the Newtonian Poisson equation! It relates metric perturbations in the Newtonian gauge to density perturbations in the synchronous gauge. So, in synchronous gauge (and synchronous gauge only), CDM linear-order density perturbations evolve exactly as the Newtonian equations predict.

Let’s finally look at a potential point of pedagogical confusion. The Newtonian equations of motion are derived assuming that the gravitational field Φ is generated instantaneously by the matter distribution at any time. There are no delays while the information travels. But Einstein’s general relativity is a causal theory in which nothing travels faster than the speed of light, not even gravity. How, then, can the two theories possibly agree, especially on super-horizon scales? The answer is that the behaviour of the linear-order solution for CDM is determined without the need for communication across the universe; it is responding purely to the local overdensity, irrespective of what any other part of the universe is doing. Mathematically, equation (356) has no k-dependence for w=0, and therefore is a purely local evolution. This is reconciled with the fact that gravity is a long-range force by energy-momentum conservation which implies there can be “no surprises” of sudden changes in the distant universe (at least if we insist on the entire contents of the universe being slowly-moving CDM)3131 31 Sometimes you hear people assert something like “according to GR, if the Sun disappeared, it would take 8 minutes before the Earth flew out of its current orbit”. This is an over-simplification to the point of being wrong. Because energy-momentum conservation is built into the theory, GR cannot tell you what happens if mass suddenly disappears. Newtonian physics says that the Earth would instantaneously fly out of orbit; GR just asserts that the Sun disappearing is physically impossible and says no more. Building in “no surprises” means that the behaviour for CDM is locally predictable despite the long-range forces involved, reconciling causality with action-at-a-distance..

Behaviour in radiation domination

We’ve already finished discussing the behaviour of perturbations during matter domination, since we just saw how the Newtonian results (Section 12) all carry over into a full relativistic analysis.

However, let’s look at the GR result (356) in the other important limit for the early universe: radiation-domination3232 32 Note that we are continuing to assume zero anisotropic stress so that Φ=Ψ. This is only assured by photons scattering regularly off baryons, so even in the limit that the mass of the universe is completely made up by photons, we still need a trace of baryons for the results in this section to be valid. See Section 15 for a discussion of why “tight coupling” is a good assumption. Without this effect, anisotropic stresses would develop which tend to further suppress structure formation. , when w=1/3. We now have (recalling that =1/η during this phase):

Φ′′+4ηΦ+k23Φ=0, (363)

which has the solution

Φ=1X2[C1(sinXX-cosX)+C2(cosXX+sinX)], (364)

where X=kη/3 and there are two arbitrary constants C1 and C2 (which are in general a function of k – do remember we are still talking about just one mode here). As η0, the second term diverges; so for realistic initial conditions with near-homogeneity in the early universe we require C2=0.

Exercise 13G Set C2=0, then verify that the solution (364) does satisfy the equation of motion (363). Taylor expand the solution (still with C2=0) around X=0 to 𝒪(X2). Hence argue that while the mode is super-horizon (kη1) the potential is approximately constant, while in the sub-horizon limit (kη1) the potential fluctuates η-2coskη/3.

The behaviour uncovered in the exercise above is absolutely critical to understanding the cosmic microwave background (and large scale structure in the present-day universe too) so let’s repeat it: Sub-horizon structure cannot grow during radiation domination, but instead oscillates and decays.

Before moving on, let’s finally consider how the solution for the potential Φ is related to the radiation overdensity δγ. We will stick with considering the synchronous gauge density perturbation because the Poisson equation takes the Newtonian form (362), which (after substituting δργ=ρ¯γδγ and inserting the time-dependence of ρ¯γ) tells us that δγη2Φ. Consequently the limiting behaviours for the radiation overdensity are η2 on super-horizon scales and sinkη/3 on sub-horizon scales.

You might like to look once more at the Meszaros effect (Section 3) to remind yourself that the dark matter density contrast is unable to grow significantly on sub-horizon scales while the (dominant) radiation component is generating this fluctuating potential. However on super-horizon scales, the matter and radiation overdensities have to trace each other, δmδγη2, because it is impossible to change their ratio without moving particles beyond the horizon. We will revisit this issue when we talk about the cosmic microwave background shortly.

13.8 The continuity and Euler equations

Recall that, in addition to the Einstein equations, there is also a conservation equation for energy-momentum μTνμ=0 (Section 2).

If the Universe contains just one fluid, the conservation equation does not add any information to the Einstein equations, since the conservation is already built in geometrically. However, if we have more than one fluid and the fluids do not interact with each other (except through gravity) the fluids individually satisfy energy-momentum conservation. In this case, there is additional information available by looking at the conservation equation for each individual fluid. As we’ve already seen, dark matter, dark energy, baryons and photons are all independent for most of the history of the Universe.

On top of that, we previously discussed that energy-momentum conservation encodes some fluid dynamics familiar from Newtonian physics: the conservation and Euler equations (see Section 2). For this reason, even if we have only one fluid, the conservation equation can provide intuitive insight into the behaviour of GR. So let’s examine its implications.

In Section 2 we showed how energy-momentum conservation implies the continuity equation in the homogeneous limit (taking ν=0). Now that we have perturbed fluid variables and a perturbed metric (which enters through the covariant derviative), at first order we obtain (after simplification) an additional equation, here expressed in Newtonian gauge:

δρ=-3(δρ+δP)-(ρ¯+P¯)(ikiδUi-3Ψ). (365)

In the unperturbed universe, the spatial components of energy-momentum conservation vanished identically. However, when applied to the perturbed universe at linear order one gets

δUi=-δUi-P¯δUi+ikiδPP¯+ρ¯-ikiΦ, (366)

which is the relativistic version of the Euler equation.

We won’t apply these immediately, since for the moment we can get everything we need from the Einstein equations. But we’ll return to them when discussing the CMB in Section 15, where it is essential to consider multiple fluids.

13.9 Scalar perturbations in scalar fields

Before the radiation-domination era, there may have been a period of slow-roll inflation (Section 3). Assuming this is driven by a single scalar field, we can continue to trace our perturbations back in time to their origin by looking at how they behave in this non-fluid setting. As when dealing with scalar field/slow-roll cosmology before, we set 8πG=1 in this section.

Start from the energy-momentum tensor for the scalar field, equation (243), then write φ=φ¯(η)+δφ(𝒙,η); finally expand all expressions to first order. This exercise generates

δT00 =dVdφδφ; (367a)
δTi0 =ikia2φ¯δφ; (367b)
δTji =[dVdφδφ+2Φa2φ¯2+2a2φ¯δφ]δji. (367c)

We can get the equation of evolution for the scalar field by applying the Euler-Lagrange equations (232) to the Lagrangian (229), giving us

-a2d2Vdφ2|φ¯δφ=-4Φφ¯-2Φφ¯′′+k2δφ+2δφ-φ¯Φ+3φ¯Ψ+δφ′′, (368)

where we’ve already made the substitution 2-k2 (i.e. δφei𝒌𝒙) by assuming we’re dealing with just one k-mode at a time, just as in the fluid case. We now need to understand the relationship between Φ, Ψ and φ to make further progress. In fact, for the sub-horizon case, we’ll be able to show that Φ and Ψ are negligible as follows.

Equation (367a) substituted into (351a) is the Poisson-like equation that will give us the relationship we need once suitably simplified. Just as with the fluid case, we first spot that (367c) is diagonal (zero for ij) which implies that Φ=Ψ. Then we can use the 0i equations from equation (367b) substituted into (351b) to re-express the Ψ+Φ term in (351a), giving us

-k2Φ=12(dVdφa2+3φ¯)δφ, (369)

remembering that we are setting 8πG=1 when dealing with scalar fields. Next, we recall that the background evolves according to

φ¯′′+2φ¯+a2dVdφ=0, (370)

which you showed3333 33 It can also be derived directly from the Euler-Lagrange equations written in conformal time η instead of clock time t in an exercise leading to equation (251) (remember that aH). This allows us to simplify (369) further to

k2Φ=12(φ¯′′-φ¯)δφ=φ¯¨a22δφ, (371)

where the last step has re-expressed the conformal time derivatives as physical time derivatives: the double overdot denotes d2/dt2 as normal.

😇 Exercise 13H Show that, in terms of the flatness parameters ϵV and ηV introduced in Section 3, the previous equation may be expressed as k2Φ=1222ϵV(ηV-ϵV)δφ. (372)

The exercise above shows that the gravitational potential is suppressed both by the slow-roll factors and, on sub-horizon scales, by a factor 2/k2. For perturbations within the horizon and even somewhat outside, we can consequently ignore the gravitational effects. We discard all such terms from equation (368) and one additional term which is a2δφd2V/dφ2=3ηVδφ2. Finally, since we assume we are in a period of inflation with w=-1, we can replace with -1/η (Table 2). The equation of motion for the perturbations then boils down to something much more palatable:

δφ′′-2ηδφ+k2δφ0 (373)

Solutions to this equation are relatively easy to find; in fact we’ve already stated them in (349) for the gravitational wave, where the e.o.m. was given by equation (346). Note the e.o.m. is identical once you replace Aδφ and =-1/η. So, one of the solutions (349) decays as a (i.e. as inflation proceeds it is erased) but the other is more interesting:

δφ=C[cos(kη)+kηsin(kη)], (374)

for arbitrary constant C. There is a plot of this function in Figure 15 (though for scalar field perturbations you should only look at the part for η<0, as we’ll need to handle the post-inflationary universe differently).

Exercise 13I Verify that equation (374) solves equation (373) and show that as a, δφC.

This tendency of δφ to “freeze out” at a fixed value is in contrast to its behaviour on strongly sub-horizon scales where the mode oscillates and decays, just like we saw with gravitational waves.

We are getting very close to being able to predict how inflation seeds structure in the universe. However, the behaviour described by equation 374 is only valid up to the mildly super-horizon regime – i.e. until the increasing 2/k2 overcomes the suppression by the slow-roll factors. Eventually the potential (372) must become large and we need a different approximation to understand the super-horizon evolution.

13.10 Dealing with super-horizon modes

In the section above, we threw away a number of terms to make progress. However for super-horizon modes and as slow-roll inflation comes to an end, the resulting approximation is no longer valid and we need to look again.

Before we do so, there is an additional consideration. When inflation ends, a radiation-dominated universe is generated through reheating. This is a complex, still poorly-understood process that takes the energy from the scalar field and dumps it into particles from the standard model and dark sectors. How can we relate perturbations before reheating to perturbations afterwards?

One might imagine that keeping track of fluctuations as an unknown physical process proceeds would be impossible, or at least require numerical integration. But we are saved by two neat properties:

  1. 1.

    all scales which will later be of cosmological interest are super-horizon during reheating – we can treat the universe as a series of separate homogeneous patches evolving independently of each other;

  2. 2.

    the slow-roll trajectory is an attractor (Section 3). It removes any slight differences between the patches by the time we hit reheating, so each patch goes through an identical process.

Working on super-horizon scales because of the first consideration, the second consideration turns out to imply the following:

η=0 where -Ψ+δρ3(ρ¯+P¯). (375)

It looks like I’ve plucked this out of thin air, so let’s first verify that it is true and then explain why. We will make use of the continuity and Euler equations given3434 34 Note that, despite being couched in terms of fluid variables, these are just as valid for a scalar field. The conservation equations from which they derive are implied by the Einstein equations so must hold true with the appropriate definitions of ρ=φ˙2+V(φ) etc (see Section 9). in Section 8.

😇 Exercise 13J Use the perturbed continuity equation (365) together with the background continuity equation (96) to show that equation (375) is true in the super-horizon limit k0. You can assume35 that δP=P¯/ρ¯δρ.
3535footnotetext: Property (2) above is what implies δP=P¯/ρ¯δρ. We are saying that there is a unique physical trajectory that each patch runs through, so P(ρ) is a unique function.

But what is and why is it conserved? Some people are satisfied by the proof in Exercise 10; but if you would like to deepen your understanding of where comes from, read on (otherwise skip to Section 11).

Suppose we have some solution a(t) that describes the universe running through slow-roll, reheating, and then into radiation domination. Consider a completely homogeneous universe and plot the scalefactor as a function of time (upper left panel of Figure 16). During slow-roll inflation, the scalefactor expands exp(Ht). Then there is a period where slow-roll ends and the universe re-heats; we don’t really know what happens in detail. Finally, the energy in the field is dumped into relativistic particles and radiation domination begins, so the scalefactor starts expanding t1/2.

Because everything starts with a slow-roll attractor phase, there are only a very limited range of alternative histories a~(t) which could fit the same physics. In fact if we want to keep everything else constant (e.g. the time t of reheating), the only possibility is to rescale, a~=ae for constant (centre panel of Figure 16). Note that for small rescalings 1, we have a~a(1+). We will shortly see that this is exactly the same as I wrote in equation (375).

Figure 16: Understanding the significance of . Left panel: The physics of reheating is complex and model-dependent, so (for example) the evolution of the scalefactor through this time is uncertain. Centre panel: Regardless of the correct physical solution in a homogeneous universe, an equally valid solution is obtained by rescaling the scalefactor aae for constant . Right panel: A fixed scalefactor change specified by aae can be thought of as a combination of aae-Ψ(t) with a time-varying Ψ(t) alongside a compensating time-varying δt shift. This is what happens in the super-horizon solution in the conformal Newtonian gauge.

The constant rescaling might appear, depending on your perspective, as either a time delay or a scalefactor shift, or some combination of both (Figure 16, right panel). Under a time shift tt+δt, the scalefactor maps to aa+a˙δt=a+aHδt. If we now return to perturbation theory with the conformal Newtonian gauge metric (350), the scalefactor shift from the perturbation is described by a(1-Ψ)a at fixed t, so we have

(1++Hδt)a(η)=a~(t) =(1-Ψ)a(t)
=-Ψ-Hδt. (376)

In fact, once we’ve decided on a specific gauge, we don’t get left with a choice about how to decompose into the scalefactor shift Ψ and time shift δη. We arranged our setup so that reheating happens at the same time t in the different patches and at the end of reheating one ends up with identical densities. So the only way of generating a density perturbation is by making use of the time shift, specifically

δρ=ρ¯˙δt=-3H(ρ¯+P¯)δt, (377)

using the conservation equation (146). Consequently we have the final version of our expression for in the conformal Newtonian gauge:

=-Ψ+δρ3(ρ¯+P¯). (378)

which of course is the same definition of from the start of the section.

13.11 Connecting δφ to and to δργ

Horizon Slow-roll era Radiation era Matter era Λ era
Sub-horizon k-1 δφsinkη+kηcoskη (eq 374) Φ negligible (eq 372) Φη-2sinkη/3 (Ex 7) δγsinkη/3 (Sec 7) δmlnt/t0 (Sec 3) Φ constant δmaη2 (Secs 1 + 7) Φa-1 δm constant (Secs 2 + 7)
Super-horizon k-1 Mildly super-horizon: δφ constant Φ remains negligible (Ex. 9) =-δφ/φ¯ also constant (Sec 11) Strongly super-horizon: remains constant. Φ grows to match onto radiation solution (Sec 11) Φ constant (Sec 7) Specifically, Φ=-2/3 (Sec 11) δγη2 δmη2 (Sec 7) As sub-horizon case As sub-horizon case
Table 3: A summary of the behaviours we have uncovered for perturbations during different eras. Overdensity evolution is given for the synchronous gauge, i.e. using Eq. (362) to relate to the Newtonian gauge potential.

The value of is that it allows us to connect the value of δφ as it exits the horizon in slow-roll inflation to the perturbations that will later re-enter the horizon during radiation domination.

During inflation, the density and pressure are specified by equations (246) and (247). Consequently ρ+P=φ˙2 and

δρVφδφ, (379)

where I have used φ˙2V. We also saw in Exercise 9 that Φ=Ψ is negligible provided we consider only mildly super-horizon modes during slow-roll. Consequently, equation 378 becomes

=-Hδφφ˙=-δφφ     (mildly super-horizon, slow-roll inflation), (380)

which is the desired relation between and δφ that we will use as perturbations cross the horizon.

😇 Exercise 13K Show that during radiation domination on super-horizon scales (k22), and assuming that Φ is constant, equation (354a) boils down to -2Φ=δγ. Use this result and the fact that Φ=Ψ to show that Φ=-23     (super-horizon, radiation domination). (381)

Equations (380) and (381) are the take-away point from this discussion: the potential Φ on entry to radiation domination is set by by equation (381); in turn, is obtained from the field fluctuations back in the slow-roll regime using equation (380).

13.12 Summary

At this point it’s helpful to step back and take a look at what the previous sections have established, and where all this is headed.

Overall, this section has been looking at the growth of inhomogeneous perturbations around the homogeneous background universe. The full set of results is summarised in Table 3. Most importantly, we’ve established that:

  • there are three different types of perturbations in relativistic cosmology: scalar, vector and tensor;

  • tensor perturbations are actually gravitational waves, and we discussed their behaviour;

  • vector perturbations are not of interest because they decay rapidly, and scalar perturbations correspond to density waves that generate structure in the Universe;

  • for cold dark matter, densities evolve identically to the Newtonian case – in particular, there is a growing mode with Φ constant and δ growing, δa;

  • more generally, the evolution of perturbations depends on whether they are sub-horizon or super-horizon, with equivalent conditions for this provided by Eqs. (342) and (343);

  • for a universe dominated by radiation (recall this used to be true of our own universe), super-horizon scalar modes have a fixed potential Φ while sub-horizon modes fluctuate and decay (η-2);

  • earlier still, in the slow-roll inflationary era, density perturbations can be characterised by the comoving curvature perturbation which is fixed if the mode is outside the horizon; or, if inside the horizon, it’s more convenient to work directly with the scalar field perturbation δφ, which oscillates and converges to a fixed (non-zero) value as the mode is stretched by inflation.

We have written down a relationship between δφ, and Φ so all these results can be stitched together into a single history for structure in our universe. There is one important missing piece of information: what generates the fluctuations in δφ in the first place? That is our next topic.