Lectures on Cosmological Correlations

Cosmological Correlations
Kavli Asian Winter School
Jan 2023
Daniel Baumann
University of Amsterdam &
National Taiwan University

Motivation
By measuring cosmological correlations, we learn both about the evolution
of the universe and its initial conditions:
380 000 years
13.8 billion years

The correlations in the CMB temperature anisotropies have revealed a great
deal about the geometry and composition of the universe:
neutrinos
geometry, dark energy
dark matter
baryons
90◦
18◦
0
1000
2000
3000
4000
5000
6000
Power
[µK
2
]
2◦
0.2◦
0.1◦
0.07◦
Angular separation
Motivation

10−4
10−3
10−2
10−1
100
101
Separation [Mpc−1
]
100
101
102
103
104
Power
[Mpc
3
]
Correlations are also observed in the large-scale structure of the universe.
The observed matter power spectrum is
Motivation
matter density

Under relatively mild assumptions, the observed correlations can be traced
back to primordial correlations on the reheating surface:
Motivation
We have learned two interesting facts about these initial conditions:
1) The fluctuations were created before the hot Big Bang.
2) During a phase of time-translation invariance (= inﬂation).

The study of cosmological correlators is also of conceptual interest:
Motivation
Rd,1
dSd+1 AdSd+1
Our understanding of quantum field theory in de Sitter space (cosmology)
is still rather underdeveloped. Opportunity for you to make progress!

OUTLINE:
Cosmological
Correlations
Wavefunction
Approach
Cosmological
Bootstrap
LECTURE NOTES:
written together with Austin Joyce.
(Lecture 1) (Lectures 2+3) (Lectures 3+4)
https://github.com/ddbaumann/cosmo-correlators

Large scales Small scales
CMB Power Spectrum
2 10 30
0
1000
2000
3000
4000
5000
6000
Power
[µK
2
]
500 1000 1500 2000 2500
Multipole
90◦
18◦ 1◦
0.2◦
0.1◦
0.07◦
Angular separation

10 100 1000
Multipole Figure courtesy of Mathew Madhavacheril

2 10 30
0
1000
2000
3000
4000
5000
6000
Power
[µK
2
]
500 1000 1500 2000 2500
Multipole
90◦
18◦ 1◦
0.2◦
0.1◦
0.07◦
Angular separation
What created the features in the power spectrum?
CMB Power Spectrum

Photon-Baryon Fluid
At early times, photons and baryons (mostly protons and electrons) are
strongly coupled and act as a single fluid:
Dark matter
Baryons
Photons
The photon pressure prevents the collapse of density fluctuations.
This allows for sound waves (like for density fluctuations in air).

Cosmic Sound Waves
Decomposing the sound waves into Fourier modes, we have
Time at recombination
Sound speed cs ⇡ 1/
p
3
The wave amplitudes are random variables, with
hAki = hBki = 0
This is why we don’t see a wavelike pattern in the CMB map.
To see the sound waves, we need to look at the statistics of the map.
⇢
⇢
= Ak cos[csk⌧⇤] + Bk sin[csk⌧⇤] 4
T
T

The two-point function in Fourier space is
h T(k) T(k0
)i =
h
PA(k) cos2
(csk⌧⇤) + 2CAB(k) sin(csk⌧⇤) cos(csk⌧⇤)
+ PB(k) sin2
(csk⌧⇤)
i
(2⇡)3
(k + k0
)
If the sound waves were created by a random source, we would have
CAB(k) = 0 PA(k) = PB(k) ⌘ P(k)
and hence the two-point function becomes
h T(k) T(k0
)i = P(k) (2⇡)3
(k + k0
)
Where are the oscillations?
CMB Power Spectrum

Coherent Phases
To explain the oscillations in the CMB power spectrum, we have to assume
that the sound waves do not have random phases:
T(k) = Ak cos[csk⌧⇤]
Oscillations in the
CMB power spectrum
h T(k) T(k0
)i0
= PA(k) cos2
[csk⌧⇤]
What created the coherent phases in the primordial sound waves?

Before the Big Bang
Phase coherence is created if the fluctuations were produced before the hot
Big Bang. Let’s see why.
Going back in time, all fluctuations have wavelengths > Hubble scale.
The solution on these large scales is
⇢
⇢
= Ck + Dk a 3
If the fluctuations were created long before the hot Big Bang, then only
the constant growing mode survived.
This explains the phase coherence of the primordial sound waves
⇢
⇢
= Ck ! Ak cos
⇥
csk⌧
⇤

Before the Big Bang
bounce
inflation
This suggests two options: rapid expansion or slow contraction?

Scale-Invariance
0.94 0.95 0.96 0.97 0.98 0.99 1
0
0.05
0.1
0.15
m2φ2
R2
1 − cos(φ/f)
ns
(r)
The primordial fluctuations are approximately scale-invariant:
This suggests that the fluctuations were created in a phase of approximate
time-translation invariance (= inflation).
ns < 1
P⇣(k)
k
Cl
l

Inflation
During inflation, the expansion rate is nearly constant:
Exponential expansion:
Quasi-de Sitter geometry
H(t) ⌘
1
a
da
dt
⇡ const
a(t) ⇡ exp(Ht)
Inflation needs to end, so there has to be some time-dependence:
" ⌘
Ḣ
H2
⌧ 1
In slow-roll inflation, this is generated by a field rolling down a flat potential.
However, there are other ways to achieve inflation.

EFT of Inflation
Inflation is a symmetry breaking phenomenon:
⇡(t, x)
ij(t, x)
H(t) (t)
The low-energy EFT is parameterized by two massless fields:
• Goldstone boson
of broken time translations
• Graviton
= (t + ⇡) ¯(t)
Creminelli et al. [2006]
Cheung et al. [2008]

EFT of Inflation
⇡(t, x)
ij(t, x)
H(t) (t)
Creminelli et al. [2006]
gij = a2
e2⇣
[e ]ij
In comoving gauge, the Goldstone boson gets eaten by the metric:
⇣ = H⇡
Curvature perturbation
Inflation is a symmetry breaking phenomenon:

The Goldstone Lagrangian is Creminelli et al. [2006]
M2
plḢ( )2 +
n=2
M4
n
n!
2 ˙ + ( )2 n
M4
n
L = + · · ·
Slow-roll inflation Higher-derivative corrections
(@ )2n
1
2 (@ )2
V ( )
EFT of Inflation
The fluctuations can have a small sound speed and large interactions:
L⇡ =
M2
pl|Ḣ|
c2
s
h
˙
⇡2
c2
s(@i⇡)2
(1 c2
s) ˙
⇡(@µ⇡)2
+ · · ·
i
nonlinearly realized symmetry

Quantum Fluctuations
Quantum fluctuations of the Goldstone produced the density fluctuations in
the early universe:
Let us derive this.
10 32
s 370 000 yrs 1 billion yrs

Quantum Fluctuations
We start from the quadratic Goldstone action
where a(⌧) ⇡ (H⌧) 1
(de Sitter).
S =
Z
d4
x
p
g M2
PlḢ gµ⌫
@µ⇡@⌫⇡
=
Z
d⌧d3
x a2
(⌧) M2
Pl|Ḣ|
⇣
(⇡0
)2
(r⇡)2
⌘
S =
1
2
Z
d⌧d3
x
✓
(u0
)2
(ru)2
+
a00
a
u2
◆
It is convenient to introduce the canonically normalized field
u ⌘ a(⌧)
q
2MPl|Ḣ| ⇡
so that
Time-dependent mass
Minkowski

Classical Dynamics
The equation of motion then is
u00
k +
✓
k2 2
⌧2
◆
uk = 0
uk(⌧) =
c±
p
2k
e±ik⌧
• At early times, k⌧ 1 , this becomes
u00
k + k2
uk ⇡ 0
Harmonic
oscillator
Growing mode
• At late times, k⌧ ! 0 , this becomes
u00
k
2
⌧2
uk ⇡ 0 uk(⌧) = c1⌧ 1
+ c2 ⌧2
Fixed by harmonic oscillator

Classical Dynamics
The equation of motion then is
• The complete solution is
uk(⌧) = c1
1
p
2k
✓
1 +
i
k⌧
◆
eik⌧
+ c2
1
p
2k
✓
1
i
k⌧
◆
e ik⌧
where the constants will be fixed by quantum initial conditions.
• The curvature perturbation becomes a constant at late times:
⇣ = H⇡ / a 1
u
/ ⌧ ⇥ ⌧ 1
= const
u00
k +
✓
k2 2
⌧2
◆
uk = 0

Canonical Quantization
S =
1
2
Z
d⌧d3
x
✓
(u0
)2
(ru)2
+
a00
a
u2
◆
Quantization starts from the action of the canonically normalized field:
• Define the conjugate momentum: p = L/ u0
= u0
• Promote fields to operators:
• Impose canonical commutation relations:
u, p ! û, p̂
[û(⌧, x), p̂(⌧, x0
)] = i~ (x x0
)
• Define the mode expansion of the field operator:
û(⌧, x) =
Z
d3
k
(2⇡)3
⇣
u⇤
k(⌧) âk + uk(⌧) â†
k
⌘
eik·x
where uk(⌧) is a solution to the classical equation of motion and
[âk, â†
k0 ] = (2⇡)3
(k k0
)
Creation
Annihilation

• Define the vacuum state: âk|0i = 0
• The expectation value of the Hamiltonian in this state is minimized
for the Bunch-Davies mode function:
Canonical Quantization
• The two-point function of the field operator (in that state) is
uk(⌧) =
1
p
2k
✓
1 +
i
k⌧
◆
eik⌧
h0|ûk(⌧)ûk0 (⌧)|0i = h0|
⇣
u⇤
k(⌧) âk + uk(⌧) â†
k
⌘ ⇣
u⇤
k0 (⌧) âk0 + uk0 (⌧) â†
k0
⌘
|0i
= (2⇡)3
(k + k0
) |uk(⌧)|2
Power spectrum
positive frequency

P⇣(k, ⌧) ⌘
H2
2M2
Pl|Ḣ|
|uk(⌧)|2
a2(⌧)
=
H2
2M2
Pl"
1
k3
(1 + k2
⌧2
)
k⌧!0
!
H2
2M2
Pl"
1
k3
Power Spectrum
Using the Bunch-Davies mode function, we obtain the following power spectrum
for the primordial curvature perturbations:
• The k 3
scaling is the characteristic of a scale-invariant spectrum.
• During inflation, both H(t) "(t)
and depend on time, which leads to a small
scale-dependence
k3
2⇡2
P⇣(k) = As
✓
k
k0
◆ns 1
• Observations find ns = 0.9667 ± 0.0040 .
⌘ 2
⇣(k)

Summary
Time
Scales
Quantum
Classical
k
H 1
Cl
l
2
⇣(k)

Gaussianity
The primordial fluctuations were highly Gaussian (as expected for the ground
state of a harmonic oscillator):
Pr[ ]
2
FNL 3/2
10 3
The universe is more
Gaussian than flat.
So far, we have only studied the free theory.
Interactions during inflation can lead to non-Gaussianity.

Primordial Non-Gaussianity
The main diagnostic of primordial non-Gaussianity is the bispectrum:
• The amplitude of the non-Gaussianity is defined as the size of the
bispectrum in the equilateral configuration:
k1
k2
k3
Shape
h⇣k1
⇣k2
⇣k3
i = (2⇡)3
(k1 + k2 + k3)
(2⇡2
)2
(k1k2k3)2
B⇣(k1, k2, k3)
FNL(k) ⌘
5
18
B⇣(k, k, k)
3
⇣(k)

Shapes of Non-Gaussianity
ϕ̇3
χ
ϕ
e−ikη
• The shape of the non-Gaussianity contains a lot of information about
the microphysics of inflation:

Equilateral Non-Gaussianity
The Goldstone mode during inflation can have large self-interactions:
nonlinearly realized symmetry
L⇡ =
M2
pl|Ḣ|
c2
s
h
˙
⇡2
c2
s(@i⇡)2
(1 c2
s) ˙
⇡(@µ⇡)2
+ · · ·
i
What would measuring (or constraining) equilateral NG teach us about the
physics of inflation?
Correlations are largest when the fluctuations have comparable wavelengths:
equilateral NG: B˙
⇡3 /
k1k2k3
(k1 + k2 + k3)3
FNL / c 2
s

Minimal models of inflation are characterized by three energy scales:
H Mpl
f⇡

Superhorizon Subhorizon
H
• This sets the energy scale of the experiment.
• At the Hubble scale, modes exit the horizon and freeze out.

Freeze-out Einstein gravity
H Mpl
• This sets the amplitude of tensor ﬂuctuations:
• At the Planck scale, gravitational interactions become strong.
No B-modes
2
=
2
⇡2
✓
H
Mpl
◆2
< 0.1 2
⇣ ⇡ 10 10
> 105
H

Background
Freeze-out Fluctuations
H Mpl
f⇡
= 58H
• At the symmetry breaking scale, time translations are broken.
• This sets the amplitude of scalar ﬂuctuations:
2
⇣ =
1
4⇡2
✓
H
f⇡
◆4
⇡ 2.2 ⇥ 10 9
f4
⇡ = M2
pl|Ḣ|cs ! ˙2
(for SR)

Closer inspection may reveal additional scales:
Strongly coupled
Weakly coupled
H Mpl
f⇡
⇤
> 5H
• At the strong coupling scale, interactions become strongly coupled.
• This sets the amplitude of interactions:
No PNG
Fequil
NL ⇠
H2
⇤2
< 0.01

DB, Green, Lee and Porto [2015]
Grall and Mellville [2020]
Asking for the theory to be weakly coupled up to the symmetry breaking
scale implies a critical value for the sound speed:
slow-roll
non-slow-roll
Higgs
Technicolor
1
cs
0.02 0.31
Hard to measure!
⇤
f⇡
⇤ f⇡
Weakly coupled
UV completion
Strongly coupled
UV completion
Fequil
NL ⇠ ⇣

Folded Non-Gaussianity
Excited initial states (negative frequency modes) lead to enhanced correlations
when two wave vectors become collinear:
folded NG: B /
1
k1 + k2 k3
e ik3⌧
• The inflationary expansion dilutes any pre-inflationary particles.
In minimal scenarios, we therefore don’t expect to see folded NG.
• Models with continuous particle production during inflation can
still give a signal in the folded limit.
Flauger, Green and Porto [2013]
Flauger, Mirbabayi, Senatore and Silverstein [2017]
Green and Porto [2020]

Squeezed Non-Gaussianity
In single-field inflation, correlations must vanish in the squeezed limit:
Maldacena [2003]
Creminelli and Zaldarriaga [2004]
k1
k2
k3
lim
k1!1
⇠ 0
Chen and Wang [2009]
DB and Green [2011]
Noumi, Yamaguchi andYokoyama [2013]
Arkani-Hamed and Maldacena [2015]
Lee, DB and Pimentel [2016]
DB, Goon, Lee and Pimentel [2017]
Kumar and Sundrum [2018]
Jazayeri and Renaux-Petel [2022]
Pimentel and Wang [2022]
The signal in the squeezed limit therefore acts as a particle detector.

Cosmological Collider Physics
ii l;!
,
µ1...µs
Massive particles are spontaneously
created in an expanding spacetime.
time
space

ii l;!
,
time
space
However, they cannot be directly observed at late times.
µ1...µJ

time
space
ii l;!
,
¯
Instead, they decay into light fields.
These correlated decays create distinct higher-order
correlations in the inflationary perturbations.

µ1...µJ
Mpl
H
1014
GeV 1018
GeV
⇣
This allows us to probe the particle spectrum at inflationary energies:
This can be many orders of magnitudes above the scales probed by
terrestrial colliders.

PJ (cos ✓)
The signal depends on the masses and spins of the new particles:
lim
k1!0
h⇣k1
⇣k2
⇣k3
i0
/
8
>
>
>
<
>
>
>
:
✓
k1
k3
◆
m < 3
2 H
✓
k1
k3
◆3/2
cos

µ ln
k1
k3
m > 3
2 H
k1
k2
k3
✓
spin
/

Observational Constraints
non-perturbative
gravitational
floor
FNL
1
10 7
ruled out by
Planck
Planck has ruled out three orders of magnitude.
The theoretically interesting regime of non-Gaussianity spans about seven
orders of magnitude:
window of opportunity
There is still room for new particles to leave their mark.

So far, I have focussed mostly on the practical motivations for studying
cosmological correlators.
In the rest of these lectures, I will introduce three diﬀerent ways to compute
primordial correlation functions:
• In-In Formalism
• Wavefunction Approach
• Cosmological Bootstrap
I will leads us to more conceptual questions about the theory of quantum
fields in cosmological spacetimes.
Lectures
Notes

O(t)
U+
U
h0|
h0|
In-In Formalism
The main observables in cosmology are spatial correlation functions:
These correlators are usually computed in the in-in formalism:
h⌦| (t, x1) (t, x2) · · · (t, xN ) |⌦i
time
Weinberg [2005]
which leads to the following master equation
hO(t)i = h0| T̄ei
R t
1
dt00
Hint(t00
)
O(t) Te i
R t
1
dt0
Hint(t0
)
|0i

hO(t)i = h0| T̄ei
R t
1
dt00
Hint(t00
)
O(t) Te i
R t
1
dt0
Hint(t0
)
|0i
In-In Formalism
Diagrammatically, this can be written as
Giddings and Sloth [2010]
Wightman
Feynman
Wightman
I++ = I =
I + =
I+ =
In the lecture notes, we apply this to many examples (see Section 3).
Here, we will instead follow the alternative wavefunction approach.

Wavefunction Approach
We first define a wavefunction for the late-time fluctuations:
(x)
(t, x)
and then use it to compute the boundary correlations
[ ] ⌘ h |⌦i
h (x1) . . . (xN )i =
Z
D (x1) . . . (xN ) | [ ]|2
The wavefunction has nicer analytic properties than the correlation functions
and it also plays a central role in the cosmological bootstrap.

OUTLINE:
Wavefunction
of the Universe
Warmup
in QM
Flat-Space
Wavefunction
De Sitter
Wavefunction
REFERENCE:
DB and Joyce, Lectures on Cosmological Correlations

N (k) = (2⇡)3
(k1 + · · · + kN ) N (k) .
Wavefunction Coeﬃcients
For small fluctuations, we expand the wavefunction as
Wavefunction coeﬃcient
• By translation invariance, we have
[ ] = exp
X
N
Z
d3
k1 . . . d3
kN
(2⇡)3N N (k) k1
. . . kN
!
• We will also write
N (k) = hO1 . . . ON i0
,
where Oa ⌘ Oka are dual operators.
This will just be notation and (so far) does not have a deeper holographic
meaning (as in dS/CFT).

Relation to In-In Correlators
In perturbation theory, the correlation functions
are related to the wavefunction coeﬃcients:
h 1 . . . N i =
Z
D 1 . . . N | [ ]|2
h 1 2i =
1
2RehO1O2i
h 1 2 3 4i =
RehO1O2O3O4i
8
Q4
a=1 RehOaOai4
+
hO1O2XihXO3O4i + c.c.
8RehXXi
Q4
a=1 RehOaOai4
h 1 2 3i =
RehO1O2O3i
4
Q3
a=1 RehOaOai
Wavefunction coeﬃcients therefore contain the same information, but are
easier to compute and have nicer analytic properties.

Computing the Wavefunction
The wavefunction has the following path integral representation:
[ ] =
Z
(0) =
( 1) = 0
D eiS[ ]
Action
Boundary conditions:
(t = 1(1 i")) = 0
selects the Bunch-Davies vacuum.
For tree-level processes, we can evaluate this in a saddle-point approximation:
Classical solution
[ ] ⇡ eiS[ cl]
with cl(0) = and cl( 1) = 0.
To find the wavefunction, we therefore need to find the classical solution for
the bulk field with the correct boundary conditions.

Harmonic Oscillator
Consider our old friend the simple harmonic oscillator
S[ ] =
Z
dt
✓
1
2
˙ 2 1
2
!2 2
◆
• The classical solution (with the correct boundary conditions) is
cl(t) = ei!t
• The on-shell action becomes
S[ cl] =
Z t⇤
ti
dt

1
2
@t( ˙ cl cl)
1
2
cl (¨cl + !2
cl)
| {z }
= 0
=
1
2
˙ cl cl
t=t⇤
=
i!
2
2

Harmonic Oscillator
• The wavefunction then is
[ ] ⇡ exp(iS[ cl]) = exp
⇣ !
2
2
⌘
• The quantum variance of the oscillator therefore is
h 2
i =
1
2!
• In QFT, the same result applies to each Fourier mode
where !k =
p
k2 + m2.
h k ki0
=
1
2!k
,
Gaussian

Time-Dependent Oscillator
To make this more interesting, consider a time-dependent oscillator
S[ ] =
Z
dt
✓
1
2
A(t) ˙ 2 1
2
B(t) 2
◆
• The classical solution is
cl(t) = K(t) , with
K(0) = 1
K( 1) ⇠ ei!t
• The on-shell action becomes
S[ cl] =
Z t⇤
ti
dt

1
2
@t(A ˙ cl cl)
1
2
cl (@t(A ˙ cl) + B cl)
| {z }
= 0
=
1
2
A ˙ cl cl
t=t⇤
=
1
2
A 2
@t log K
t=t⇤

• The wavefunction then is
which implies
Time-Dependent Oscillator
[ ] ⇡ exp(iS[ cl]) = exp
✓
i
2
(A@t log K)
⇤
2
◆
,
| [ ]|2
= exp Im(A@t log K) ⇤
2
h 2
i =
1
2 Im(A@t log K) ⇤

Free Field in de Sitter
The action of a massless ﬁeld in de Sitter is that of a time-dependent oscillator:
S =
Z
d⌘d3
x a2
(⌘)
h
( 0
)2
(r )2
i
=
1
2
Z
d⌘
d3
k
(2⇡)3

1
(H⌘)2
0
k
0
k
k2
(H⌘)2 k k
• The classical solution is
and hence
cl(⌘) = K(⌘) , with
K(⌘) = (1 ik⌘)eik⌘
log K(⌘) = log(1 ik⌘) + ik⌘

• Using our result for the variance of a time-dependent oscillator, we then get
h k ki0
=
H2
2k3
• The result for a massive field is derived in the lecture notes.
which we derived in the last lecture using canonical quantization.

Anharmonic Oscillator
Consider the following anharmonic oscillator
S[ ] =
Z
dt
✓
1
2
˙ 2 1
2
!2 2 1
3
g 3
◆
• The classical equation of motion is
¨cl + !2
cl = g 2
cl
• A formal solution is
where
K(t) = ei!t
,
G(t, t0
) =
1
2!
⇣
e i!(t t0
)
✓(t t0
) + ei!(t t0
)
✓(t0
t) ei!(t+t0
)
⌘
.
cl(t) = K(t) + i
Z
dt0
G(t, t0
) g 2
cl(t0
)
Feynman propagator

The terms linear in cancel.
S[ cl] =
i!
2
2 g
2
Z
dt ei!t 2
cl
+
Z
dt

1
2
✓
ei!t
ig
Z
dt0
G(t, t0
) 2
cl(t0
)
◆ ⇣
g 2
cl(t)
⌘ g
3
3
cl .
• Deriving the on-shell action is now a bit more subtle:
S[ cl] =
Z t⇤
ti
dt

1
2
@t( cl
˙ cl)
1
2
cl(¨cl + !2
cl)
g
3
3
cl
=
1
2
cl
˙ cl
t=t⇤
+
Z
dt

1
2
cl(¨cl + !2
cl)
g
3
3
cl .
Since limt!0 @tG(t, t0
) = iei!t0
6= 0, the boundary term is
1
2
cl
˙ cl
t=t⇤
=
1
2
✓
i! ig
Z
dt0
iei!t0
2
cl(t0
)
◆
=
i!
2
2 g
2
Z
dt0
ei!t0
2
cl(t0
) .
The action then becomes

• The final on-shell action is
S[ cl] =
i!
2
2 g
3
Z
dt 3
cl(t)
ig2
2
Z
dt dt0
G(t, t0
) 2
cl(t0
) 2
cl(t)
• To evaluate this, we write the classical solution as
where
(0)
(t) = ei!t
,
(1)
(t) = i
Z
dt0
G(t, t0
)
⇣
2
e2i!t0
⌘
=
2
3!2
e2i!t
ei!t
.
• With this, the wavefunction becomes
[ ] ⇡ eiS[ cl]
= exp
✓
!
2
2 g
9!
3
+
g2
72!3
4
+ · · ·
◆
• From this, we can compute h 3
i, h 4
i, etc.
cl(t) =
X
n
gn (n)
(t),

Interacting Fields
Back to field theory:
S[ ] =
Z
d4
x
p
g
✓
1
2
gµ⌫
@µ @⌫
1
2
m2 2 1
3
g 3
◆
The analysis is similar to that of the anharmonic oscillator (see lecture notes).
In the interest of time, we jump directly to Feynman rules for WF coeﬃcients:
Kk(t)
Bulk-to-bulk propagator
for every internal line
G(kI; t, t0
)
Bulk-to-boundary propagator
for every external line
ig
Z
dt a3
(t)
Time integral for
every vertex

The wavefunction is an interesting object even in flat space.
• The propagators are the same as for the harmonic oscillator:
• All correlators can be evaluated at
We will compute the simplest tree-level correlators in this theory.
• We will study the following simple toy model
Flat-Space Wavefunction

Contact Diagrams
The three-point wavefunction coeﬃcient is
= ig
Z 0
1
dt Kk1
(t)Kk2
(t)Kk3
(t)
= ig
Z 0
1
dt ei(k1+k2+k3)t
=
g
(k1 + k2 + k3)
.
This is easily generalized to N-point wavefunction coeﬃcients:
We will have more to say
about this singularity.

Exchange Diagrams
The four-point wavefunction coeﬃcient in

Exchange Diagrams
The final answer is remarkably simple and has an interesting singularity
structure:
Energy entering
the right vertex
Energy entering
the left vertex
Total energy
We will have more to say about this later.

Recursion Relation
More complicated graphs can be computed in the same way, but the complexity
of the time integrals increases quickly with the number of internal lines.
Fortunately, there is a powerful recursion relation that allows complex graphs
to be constructed from simpler building blocks:
Arkani-Hamed, Benincasa and Postnikov [2017]
This provides a simple algorithm to produce the result for any tree graph.

Consider a massive scalar in de Sitter space.
• The action of is
• The equation of motion is
• The general solution of this equation is
Hankel function

Consider a massive scalar in de Sitter space.
• The action of is
• The equation of motion is
• For , the field is conformally coupled and
Harmonic
oscillator
We will often consider this special case, because the wavefunction can
then be computed analytically.

Propagators
For a generic field, with mode function , the propagators are
In general, these are given by Hankel functions which leads to complicated
integrals in the wavefunction calculations.
Our explicit calculations will therefore be for conformally coupled scalars.
We return to the general case in the cosmological bootstrap.

Contact Diagrams
The three-point WF coeﬃcient is
Product of
Hankel functions

Contact Diagrams
For conformally coupled scalars, we get
where
The corresponding correlator is

Contact Diagrams
The four-point WF coeﬃcient for a contact interaction of conformally coupled
scalars is
Since the interaction preserves the conformal symmetry of the free theory,
this is the same result as in flat space (modulo the prefactor).

Exchange Diagrams
The four-point WF coeﬃcient corresponding to an exchange diagram
(of conformally coupled scalars) is

Exchange Diagrams
We see that this four-point function can be written as an integral of the
flat-space result:
This integral can be performed analytically to give

A Challenge
So far, we have only computed the correlators of conformally coupled scalars.
Consider now the exchange of a generic massive scalar:
Conformally
coupled scalars
Massive scalar
In general, the time integrals cannot be performed analytically.
We need a diﬀerent approach.

?
t = 0
If inflation is correct, then the reheating surface is the future boundary of an
approximate de Sitter spacetime:
Can we bootstrap these boundary correlators directly?
• Conceptual advantage: focuses directly on observables.
• Practical advantage: simpliﬁes calculations.
end of inflation
hot big bang

| f i
h i |
+
?
The bootstrap perspective has been very influential for scattering amplitudes:
• Practical advantage: simpliﬁes calculations.
• Conceptual advantage: reveals hidden structures.
There has be enormous progress in bypassing Feynman diagram expansions to
write down on-shell amplitudes directly:

t = 0
+
?
I will describe recent progress in developing a similar bootstrap approach for
cosmological correlators:
Goal: Develop an understanding of cosmological correlators that parallels our
understanding of flat-space scattering amplitudes.

OUTLINE:
REFERENCES:
Bootstrapping
Tools
The Bootstrap
Philosophy
Examples
DB, Green, Joyce, Pajer, Pimentel, Sleight and Taronna,
Snowmass White Paper: The Cosmological Bootstrap [2203.08121]
DB and Joyce, Lectures on Cosmological Correlations

How we usually make predictions:
Lagrangians
equations of motion
spacetime evolution
Feynman diagrams
Physical Principles Observables
locality, causality,
unitarity, symmetries
Models
Works well if we have a well-established theory (Standard Model, GR, …)
and many observables.
The Conventional Approach

The Conventional Approach
Although conceptually straightforward, this has some drawbacks:
• It involves unphysical gauge degrees of freedom.
• Relation between Lagrangians and observables is many-to-one.
• Even at tree level, the computation of cosmological correlators
involves complicated time integrals:
Hard to compute!
• Fundamental principles (e.g. locality and unitarity) are obscured.
• To derive nonperturbative correlators and constraints from the UV
completion, we need a deeper understanding.
N =
X
Diagrams
Y
Vertices
Z
dt
! ✓
External
propagators
◆ ✓
Internal
propagators
◆ ✓
Vertex
factors
◆

Physical Principles Observables
Models
In the bootstrap approach, we cut out the middle man and go directly from
fundamental principles to physical observables:
This is particularly relevant when we have many theories (inflation, BSM, …)
and few observables.
The Bootstrap Method

S-matrix Bootstrap
• What stable particles exist?
fixed by locality and
Lorentz symmetry
=
g2
PJ (cos ✓)
s M2
locality
unitarity
No Lagrangian is required to derive this.
Basic principles allow only a small menu of possibilities.
• What interactions are allowed?
Lorentz
=
✓
h12i3
h13ih23i
◆J

Consistent factorization is very constraining for massless particles:
µ1...µS
S = { 0 , 1
2 , 1 , 3
2 , 2 }
Only consistent for spins
YM
GR
SUSY
Benincasa and Cachazo [2007]
McGady and Rodina [2013]
S-matrix Bootstrap

The modern amplitudes program has been very successful:
1. New Computational Techniques 2. New Mathematical Structures
3. New Relations Between Theories
• Recursion relations
• Generalized unitarity
• Soft theorems
• Positive geometries
• Amplituhedrons
• Associahedrons
• Color-kinematics duality
• BCJ double copy
4. New Constraints on QFTs
• Positivity bounds
• EFThedron
5. New Applications
• Gravitational wave physics
• Cosmology
A Success Story

Recently, there has been significant progress in bootstrapping correlators using
physical consistency conditions on the boundary:
symmetries, locality,
causality, unitarity, …
Cosmological Bootstrap
In this part of the lectures, I will review these developments.

Singularities
Symmetries Unitarity
Bootstrapping Tools

SO(D, 1) : [JAB, JCD] = ⌘BCJAD ⌘ACJBD + ⌘ADJBC ⌘BDJAC
⌘ABXA
XB
= X2
0 + X2
1 + · · · + X2
D = H 2
D-dimensional de Sitter space can be represented as an hyperboloid in
(D+1)-dimensional Minkowski space:
de Sitter radius
The symmetries of de Sitter then correspond to higher-dimensional Lorentz
transformations:
In a specific time slicing, we define
Dilatations
D̂ ⌘ JD0
Pi ⌘ JDi J0i
Ki ⌘ JDi + J0i
Translations
Boosts
De Sitter Symmetries
Rotations

⌘ABXA
XB
= X2
0 + X2
1 + · · · + X2
D = H 2
D-dimensional de Sitter space can be represented as an hyperboloid in
(D+1)-dimensional Minkowski space:
de Sitter radius
The symmetries of de Sitter then correspond to higher-dimensional Lorentz
transformations:
The de Sitter algebra then becomes
[Jij, Jkl] = jkJil ikJjl + ilJjk jlJik , [D̂, Pi] = Pi ,
[Jij, Kl] = jlKi ilKj , [D̂, Ki] = Ki ,
[Jij, Pl] = jlPi ilPj , [Ki, Pj] = 2 ijD̂ 2Jij .
SO(D, 1) : [JAB, JCD] = ⌘BCJAD ⌘ACJBD + ⌘ADJBC ⌘BDJAC

ds2
=
1
H2⌘2
d⌘2
+ dx2
We will work in the flat-slicing, where the de Sitter metric is
and the symmetry generators (Killing vectors) become
Pi = @i D̂ = ⌘@⌘ xi
@i
Jij = xi@j xj@i Ki = 2xi⌘@⌘ +
⇣
2xj
xi + (⌘2
x2
) j
i
⌘
@j
• Dilatations and boosts correspond to the following finite transformations:
⌦ ⌘ 1 + 2(b · x) + b2
(x2
⌘2
)
where
⌘ 7! ⌘
x 7! x
⌘ 7! ⌘/⌦
x 7!
⇥
x + (x2
⌘2
)b
⇤
/⌦
• At late times, the boosts become special conformal transformations (SCTs).
• Late-time correlators are therefore constrained by conformal symmetry.

Unitarity Representations
Physical states are described by unitary representations of the de Sitter group.
These representations are labelled by the eigenvalues of the Casimir operators:
C2 ⌘
1
2
JAB
JAB = D̂(d D̂) + Pi
Ki +
1
2
Jij
Jij
C4 ⌘ JABJBC
JCDJDA
Spin
Scaling
dimension
| , Ji
Unitarity puts constraints on the allowed scaling dimensions.
For scalars, we have
0 1 2 3 4
Re∆
Im∆
Principal series
Complementary series

Bulk Fields
Consider a massive scalar field. Its equation of motion can be written as an
eigenvalue equation for the quadratic Casimir:
Using C2 = (d ) , we get
and
C2 = ⌘2
✓
@2
@⌘2
d 1
⌘
@
@⌘
r2
◆
=
m2
H2
These representations can be realized by bulk fields.
0 1 2 3 4
Re∆
Im∆
Heavy fields:
Light fields:
⌘
d
2
r
d2
4
m2
H2
m/H > d/2
m/H < d/2
¯ = d .

Boundary Fluctuations
Our main interest are the fluctuations on the late-time boundary.
At late times, the solution for the bulk field is
(x, ⌘ ! 0) = (x) ⌘ + ¯(x) ⌘d
Boundary field profile
For light fields, the first term of the asymptotic solution dominates.
The de Sitter isometries then act as
Pi = @i
Jij = (xi@j xj@i)
D = ( + xi
@i)
Ki =
h
2xi +
⇣
2xj
xi x2 j
i
⌘
@j
i
⌘@⌘ 7!
Transformations of a primary operator of weight in a CFT.

Conformal Ward Identities
The boundary correlators are constrained by the conformal symmetry:
N
X
a=1
h 1 · · · a · · · N i = 0
where a stand for any of the field transformations.
• Translations and rotations require that the correlator is only a
function of the separations |xa xb| .
• Dilatation and SCT invariance imply
0 =
N
X
a=1
✓
a + xj
a
@
@xj
a
◆
h 1 · · · a · · · N i
0 =
N
X
a=1
✓
2 axi
a + 2xi
axj
a
@
@xj
a
x2
a
@
@xa,i
◆
h 1 · · · a · · · N i

0 =
N
X
a=1
✓
a + xj
a
@
@xj
a
◆
h 1 · · · a · · · N i
0 =
N
X
a=1
✓
2 axi
a + 2xi
axj
a
@
@xj
a
x2
a
@
@xa,i
◆
h 1 · · · a · · · N i
Exercise: Show that for two- and three-point functions the conformal Ward
identities are solved by
h 1 2i =
1
x2 1
12
1, 2
h 1 2 3i =
c123
x t 2 3
12 x t 2 1
23 x t 2 2
31
where a ⌘ a(xa) , xab ⌘ |xa xb| and t ⌘
X
a .

N
X
a=1
Z
dd
x1 · · · dd
xN N (x) (x1) · · · (xa) · · · (xN )
Similar Ward identities can be derived for the wavefunction coeﬃcients.
Consider
[ ] = exp
✓
· · · +
Z
dd
x1 · · · dd
xN N (x) (x1) · · · (xN ) + · · ·
◆
N
X
a=1
Z
dd
x1 · · · dd
xN N (x) (x1) · · ·
⇣
a xj
a@xj
a
⌘
(xa) · · · (xN )
=
N
X
a=1
Z
dd
x1 · · · dd
xN
h⇣
(d a) + xj
a@xj
a
⌘
N (x)
i
(x1) · · · (xN )
Integrate by parts, so that the generator of the transformation acts on the
wavefunction coeﬃcient. For example, for the dilatation, we have

Under dilatations, the wavefunction coeﬃcients therefore transform as
N (x) =
N
X
a=1
⇣
(d a) + xj
a@xj
a
⌘
N (x)
A similar derivation applies to SCTs.
Imposing N = 0 leads to the conformal Ward identity for WF coeﬃcients:
0 =
N
X
a=1
✓
(d a) + xj
a
@
@xj
a
◆
hO1 · · · Oa · · · ON i
0 =
N
X
a=1
✓
2(d a)xi
a + 2xi
axj
a
@
@xj
a
x2
a
@
@xa,i
◆
hO1 · · · Oa · · · ON i
N ⌘ hO1 · · · ON i
where .
Same as before, except the dual operators have weight ¯ a = d a .

In cosmology, we need these Ward identities in Fourier space.
Although the correlators themselves are hard to Fourier transform, the
symmetry generators are not. This gives
Since hO1 · · · ON i = (2⇡)d (d)
(k1 + · · · + kN ) hO1 · · · ON i0
attention to the momentum derivatives hitting the delta function.
, we need to pay
This gives an extra term in the dilatation Ward identity, so that
"
d +
N
X
a=1
D̂a
#
hO1 · · · ON i0
(k) = 0
N
X
a=1
Ki
ahO1 · · · ON i0
(k) = 0
D̂Ok =
h
+ ki
@ki
i
Ok
KiOk =
h
2 @ki ki@kj @kj + 2kj
@kj @ki
i
Ok

t d(N 1)
t d ⇥ N
t ⌘
P
a a
"
d +
N
X
a=1
D̂a
#
hO1 · · · ON i0
(k) = 0
N
X
a=1
Ki
ahO1 · · · ON i0
(k) = 0
• An N-point function has scaling dimension in position space.
• The momentum-space correlator then has dimension .
• After stripping oﬀ the delta function, the dimension is .
This fixes the overall momentum scaling of the correlator, allowing us to make
an ansatz that automatically solves the dilatation Ward identity.
All the juice is then in the SCT Ward identity.

( = 3)
hO1O2i0
= Ak2 1 d
1 1, 2
1 = 2
hO1O2i0
/ k 1+ 2 d
1
Two-Point Functions
Consider the two-point function of two arbitrary scalar operators.
• The SCT Ward identity then forces , and hence
• A massless field in d = 3 dimensions has
h 1 2i0
=
1
2 RehO1O2i0
/
1
k3
as expected for a scale-invariant power spectrum.
• An ansatz that solves the dilatation Ward identity is

hO1O2O3i0
= k 1+ 2+ 3 2d
3 Ĝ(p, q)
Consider the three-point function of three arbitrary scalar operators.
• An ansatz that solves the dilatation Ward identity is
Three-Point Functions
where p ⌘ k1/k3 and q ⌘ k2/k3.
• The SCT Ward identity then implies two diﬀerential equations for
which are solved by
Ĝ(p, q),
hO1O2O3i0
= Appell F4
/
Z 1
0
dx x
d
2 1
K⌫1 (k1x)K⌫2 (k2x)K⌫3 (k3x)

where u ⌘ k3/(k1 + k2) and w ⌘ k3/(k1 k2).
hO1O2X3i0
= k 2
3 Ĝ(u, w)
Three-Point Functions
It will be useful to work out more explicitly the case of two conformally
coupled scalars and a generic scalar in 3 dimensions:
where u ⌘ u2
(1 u2
)@2
u 2u3
@u and µ2
+ 1
4 = ( 2)( 1).
Ĝ(u, w) = Ĝ(u), with
• The SCT Ward identity then implies
• This is solved by

u +
✓
µ2
+
1
4
◆
Ĝ(u) = 0
Ĝ+(u) / P0
iµ 1
2
(u 1
) Ĝ (u) / P0
iµ 1
2
( u 1
)
and .
associated Legendre

Four-Point Functions
Finally, we consider the four-point function of conformally coupled scalars:
hO1O2O3O4i0
=
kI
k1
k2
k4
k3
=
1
kI
F̂(u, v)
where u ⌘ kI/(k1 + k2) and v ⌘ kI/(k3 + k4).
• The SCT Ward identity then implies
where u ⌘ u2
(1 u2
)@2
u 2u3
@u.
• To solve this equation requires additional physical input.
Arkani-Hamed, DB, Lee
and Pimentel [2018]

Much of the physics of scattering
amplitudes is fixed by their singularities.
Singularities also play an important role in
constraining the structure of cosmological correlators.

=
g2
(k12 + k34)(k12 + kI)(k34 + kI)
⌘
g2
EELER
= ig
Z 0
1
dt eiEt
=
g
E
Flat-Space Wavefunction
In the last lecture, we computed wavefunction coeﬃcients in flat space.
Let us look back at some of these results.
Correlators are singular when energy is conserved.
= g2
Z 0
1
dt1dt2 eik12t1
GkI
(t1, t2) eik34t2

Total Energy Singularity
Every correlator has a singularity at vanishing total energy:
• The residue of the total energy singularity is the corresponding amplitude.
Raju [2012]
Maldacena and Pimentel [2011]
N =
AN =
⇠
Z 0
1
dt eiEt
f(t) =
AN
E
+ · · ·
⇠
Z 1
1
dt eiEt
f(t) = MN (E)

Total Energy Singularity
• This is easy to confirm for the tree-exchange diagram:
lim
E!0
g2
EELER
=
1
E
g2
(k12 + kI)( k12 + kI)
=
1
E
g2
s
=
A4
E
• The total energy singularity arises from the early-time limit of the
integration (where the boundary is infinitely far away):
where s ⌘ k2
12 k2
I is the Mandelstam invariant.
N ⇠
Z 0
1
dt eiEt
f(t) =
AN
E
+ · · ·
This explains intuitively why this singularity reproduces flat-space physics.
• The singularity can also be seen in the flat-space recursion relation for the
wavefunction. Arkani-Hamed, Benincasa and Postnikov [2017]

=
AL ⇥ ˜R
EL
+ · · ·
˜R ⌘
1
2kI
⇣
R(k34 kI) R(k34 + kI)
⌘
Partial Energy Singularities
Exchange diagrams lead to additional singularities:
lim
EL!0
• On the pole, the correlator factorizes into a product of an amplitude and a
shifted correlator:
• The singularity arises from the early-time limit of the integration, where the
bulk-to-bulk propagator factorizes:
GkI
(tL, tR)
tL! 1
!
eikI tL
2kI
⇣
e ikI tR
e+ikI tR
⌘

Partial Energy Singularities
This pattern generalizes to arbitrary graphs.
=
AL ⇥ ˜R
EL
+ · · ·
Arkani-Hamed, Benincasa and Postnikov [2017]
DB, Chen, Duaso Pueyo, Joyce, Lee and Pimentel [2021]
Salcedo, Lee, Melville and Pajer [2022]
• The wavefunction is singular when the energy flowing into any connected
subgraph vanishes:
lim
EL!0
• For loops, the poles can become branch points.

Generalization to De Sitter
The wavefunction in de Sitter also has singularities when the energy flowing
into any connected subgraphs vanishes:
• These singularities again arise from the early-time limit of the integration,
where the bulk-to-bulk propagator factorizes.
• In these singular limits, the correlator factorizes into a product of an
amplitude and a shifted correlator:
˜R ⌘ P(kI)
⇣
R(ka, kI) R(ka, kI)
⌘
Power spectrum of
exchanged field
lim
EL!0 =
AL ⇥ ˜R
EP
L
+ · · ·

Singularities as Input
1) Ward identities
2) Unitarity
The singularities of the wavefunction only arise at unphysical kinematics.
Nevertheless, they control the structure of cosmological correlators.
• Amplitudes are the building blocks of cosmological correlators:
=
(EL)p
(ER)q
• To construct the full correlator, we need to connect its singularities.
We will present two ways of doing this:

Unitarity is a fundamental feature of quantum mechanics.
However, the constraints of bulk unitary on cosmological
correlators were understood only very recently.
Goodhew, Jazayeri and Pajer [2020]
Meltzer and Sivaramakrishnan [2020]

T T†
= iT†
T
S-matrix Optical Theorem
A consistent theory must have a unitary S-matrix:
Writing S = 1 + iT, this implies
Sandwiching this between the states hf| and |ii, and inserting a complete
set of states, we get
S†
S = 1
hf|T|ii hi|T|fi⇤
= i
X
X
Z
d⇧X hf|T†
|XihX|T|ii
Using hf|T|ii ⌘ (2⇡)4
(pi pf )A(i ! f), we obtain
2Im A(i ! f) =
X
X
Z
d⇧X (2⇡)4
(pi pX ) A(i ! X)A⇤
(f ! X)
The right-hand side can be written in terms of the total cross section, giving
the familiar form of the optical theorem.

Writing Û = 1 + V̂ , this implies
Cosmological Optical Theorem
Recently, an analogous result was derived for cosmological correlators.
Recall that [ ] = h |Û|0i, where
Û = T exp
✓
i
Z ⌘⇤
1
d⌘ Ĥint
◆
Probabilities, | (⌘)|2
, are conserved if U†
U = 1.
V̂ + V̂ †
= V̂ †
V̂
Sandwiching this between and inserting a complete set of states,
we get
h | and |0i,
h |V̂ |0i + h |V̂ †
|0i =
Z
DX h |V̂ †
|XihX|V̂ |0i
Cosmological
Optical Theorem
To use this, we need to relate the matrix elements to wavefunction coeﬃcients.

Example: Four-Point Exchange
As a concrete example, we consider the four-point exchange diagram:
To extract a constraint on the WF coeﬃcient from the COT, we let be the
| i
h{ka}|V̂ |0i + h{ka}|V̂ †
|0i =
X
X
Z Y
b2X
d3
qb
(2⇡)3
h{ka}|V̂ †
|{qb}ih{qb}|V̂ |0i
where V̂ = i
Z ⌘⇤
1
d⌘ Ĥint(⌘)
Z ⌘⇤
1
d⌘
Z ⌘
1
d⌘0
Ĥint(⌘)Ĥint(⌘0
) + · · · .
momentum eigenstate |k1, k2, k3, k4i ⌘ |{ka}i:
Our task is to relate the matrix elements in this expression to WF coeﬃcients.
2

h{ka}|V̂ |0i + h{ka}|V̂ †
|0i =
X
X
Z Y
b2X
d3
qb
(2⇡)3
h{ka}|V̂ †
|{qb}ih{qb}|V̂ |0i
• Substituting the mode expansions
ˆk(⌘) = f⇤
k (⌘) âk + fk(⌘) â k
ˆk(⌘) = ⇤
k(⌘) b̂k + k(⌘) b̂ k
the ﬁrst term on the left-hand side becomes
h{ka}|V̂ |0i = g2
Z
d⌘d⌘0
a4
(⌘)a4
(⌘0
) GF (kI; ⌘, ⌘0
)
4
Y
a=1
fka
(⌘)
where GF is the Feynman propagator.

h{ka}|V̂ |0i + h{ka}|V̂ †
|0i =
X
X
Z Y
b2X
d3
qb
(2⇡)3
h{ka}|V̂ †
|{qb}ih{qb}|V̂ |0i
GF (k; ⌘, ⌘0
) = G(k; ⌘, ⌘0
) +
⇤
k(⌘⇤)
k(⌘⇤)
k(⌘) k(⌘0
)
we get
h{ka}|V̂ |0i =
⇣
4(ka, kI) 3(k12, kI) 3(k34, kI) P (kI)
⌘ 4
Y
a=1
fka
(⌘⇤)
!
.
h{ka}|V̂ †
|0i =
⇣⇥
4( ka, kI)
⇤⇤
3(k12, kI) 3(k34, kI) P (kI)
⌘ 4
Y
a=1
fka
(⌘⇤)
!
• A similar analysis for the second term leads to
Analytic
continuation
Substituting

h{ka}|V̂ |0i + h{ka}|V̂ †
|0i =
X
X
Z Y
b2X
d3
qb
(2⇡)3
h{ka}|V̂ †
|{qb}ih{qb}|V̂ |0i
• Finally, the third term leads to
RHS =
h
3(k12, kI) 3(k34, kI) + 3(k12, kI) 3(k34, kI)
i
P (kI)
4
Y
a=1
fka
(⌘⇤)
!
• Putting it all together, we get
4(ka, kI) +
⇥
4( ka, kI)
⇤⇤
= P (kI) ˜3(k12, kI) ˜3(k34, kI)
where ˜3(k12, kI) ⌘ 3(k12, kI) 3(k12, kI).

Cutting Rules
For more complicated graphs, the explicit evaluation of the COT quickly
becomes cumbersome. A simpler alternative are cutting rules.
We start with some analytic properties of the propagators defining the
perturbative wavefunction:
• A useful concept is the discontinuity of a function
• Both propagators decay at past infinity, which we can enforce by giving
the norm k a negative imaginary part: k2
= rei✓
, with ✓ 2 ( 2⇡, 0).
Disc[f(k)] ⌘ f(k) f(e i⇡
k)
• The discontinuity of the bulk-to-bulk propagator is
Disc[G(k; ⌘, ⌘0
)] = P(k) Disc[K(k, ⌘)] ⇥ Disc[K(k, ⌘0
)]

DisckI
⇥
4(ka, kI)
⇤
=
Cutting Rules
Consider, the four-point exchange diagram
4(ka, kI) =
Taking the discontinuity (with respect to ) corresponds to cutting the
internal line:
kI
P (k)
Hence, we find
DisckI
⇥
4(ka, kI)
⇤
= P (kI) DisckI
⇥
3(k12, kI)
⇤
⇥ DisckI
⇥
3(k34, kI)
⇤
which is equivalent to the result derived from the COT.

Cutting Rules
This procedure generalizes to systematic cutting rules:
These are the analog of the Cutkosky cutting rules for the S-matrix.
Disc
⇥
N
⇤
=
X
all possible cuts
More details can be found in the lecture notes and in the relevant literature.
Meltzer and Sivaramakrishnan [2020]
Caspedes, Davis and Melville [2020]
Goodhew, Jazayeri, Lee and Pajer [2021]
DB, Chen, Duaso Pueyo, Joyce, Lee and Pimentel [2021]
Melville and Pajer [2021]
Sleight and Taronna [2021]
Jazayeri, Pajer and Stefanyszyn [2021]
Meltzer [2021]
Benincasa [2022]

Unitarity as Input
• The discontinuity of a WF coeﬃcient factorizes into lower-point objects.
• This contains slightly more information that the energy singularities:
• These unitarity cuts have become an important bootstrapping tool.
lim
EL!0
Disc
= lim
EL!0
h
N (ka, kI) N (ka, kI)
i
= lim
EL!0
N (ka, kI)
has no singularities in EL
The cutting rule fixes the entire Laurent series in EL.

Dispersive
Integrals
Massive
Exchange
FRW
Correlators
Examples

Exchange Four-Point Function
We are now ready to return to the challenge of massive particle exchange.
Conformally
coupled scalars
Massive scalar
= g2
Z
d⌘
⌘2
Z
d⌘0
⌘02
eik12⌘
eik34⌘0
G(kI; ⌘, ⌘00
)
hO1O2O3O4i0
=
In general, the time integrals cannot be performed analytically.

Conformal Ward Identity
Recall that the s-channel correlator can be written as
hO1O2O3O4i0
=
kI
k1
k2
k4
k3
=
1
kI
F̂(u, v)
where u ⌘ kI/(k1 + k2) and v ⌘ kI/(k3 + k4).
We have seen that the conformal symmetry implies the following
diﬀerential equation
where u ⌘ u2
(1 u2
)@2
u 2u3
@u.
We will classify solutions to this equation by their singularities.

=
X
n
cn(u, v)
E2n+1
Contact Solutions
The simplest solutions correspond to contact interactions:
F̂c =
Only total energy
singularities
F̂c,0 =
kI
E
=
uv
u + v
4
• For , we have .
• Higher-derivative bulk interactions, lead to F̂c,n = n
uF̂c,0 .
F̂c,1 = 2
✓
kI
E
◆3
1 + uv
uv
,
F̂c,2 = 4
✓
kI
E
◆5
u2
+ v2
+ uv(3u2
+ 3v2
4) 6(uv)2
6(uv)3
(uv)3
.
Fixed by symmetry
This can get relatively complex, but everything is fixed by symmetry.

Exchange Solutions
For tree exchange, we try
( u + M2
)F̂e = Ĉ
( v + M2
)F̂e = Ĉ
=
For amplitudes, the analog of this is
This is an algebraic relation, while for correlators we have a diﬀerential equation.
=

Exchange Solutions
Using the simplest contact interaction as a source, we have
Before solving this equation, let us look at its singularities.
• Flat-space limit:
• Folded limit:
This singularity should be absent
in the Bunch-Davies vacuum.
• Factorization limit:
• Collapsed limit:
This non-analyticity is a key
signature of particle production.

Exchange Solutions
Using the simplest contact interaction as a source, we have
Before solving this equation, let us look at its singularities.
• Flat-space limit:
• Folded limit:
• Factorization limit:
• Collapsed limit:
Factorization and folded limits uniquely fix the solution.
The remaining singularities become consistency checks.

whose solution is
particle production
EFT expansion
analytic non-analytic
F̂ =
X
n
( 1)n
(n + 1
2 )2 + M2
⇣u
v
⌘n+1
+
⇡
cosh(⇡M)
sin(M log(u/v))
M
The general solution takes a similar form. Arkani-Hamed, DB, Lee and Pimentel [2018]
t ⌘ log(u/v)
Forced Harmonic Oscillator

The correlator is forced to have an oscillatory feature in the physical regime:
BRANCH CUT
OSCILLATIONS
COLLAPSED LIMIT
Oscillations

These oscillations reflect the evolution of the massive particles during inflation:
Time-dependent eﬀects have emerged in the solution of the time-independent
bootstrap constraints.
Particle Production

Centre-of-mass energy (GeV)
Cross
section
(nb)
Four-point
function
Four-point
function
Momentum ratio
Correlation
strength
The oscillatory feature is the analog of a resonance in collider physics:

A =
g2
s M2
PS(cos ✓)
The angular dependence of the signal depends on the spin of the particles.
This is very similar to what
we do in collider physics:
✓
kI / sin[M log kI ] PS(cos ✓)
The frequency of the oscillations depends on the mass of the particles:
Particle Spectroscopy

We have seen that the discontinuity of a wavefunction
coeﬃcient is related to lower-point objects:
For rational correlators, we can “integrate” the
discontinuity to get the full correlator.

Dispersive Integral
By causality, the bulk-to-bulk propagator is an analytic function of k2
that
decays at large k. It can therefore be written as
G(k; ⌘, ⌘0
) =
1
2⇡i
Z 1
0
dq2
q2 k2 + i"
Disc[G(k; ⌘, ⌘0
)]
This can be used to integrate the discontinuities of tree-level WF coeﬃcients.
As an example, we consider the four-point exchange diagram:
Meltzer [2021]
Often this dispersive integral is easier to compute than the bulk time integrals.
=
1
2⇡i
Z 1
0
dq2
q2 k2
I + i"
Discq


kI
k34
k12
q
Using the unitarity cut, this becomes
=
1
2⇡i
Z 1
0
dq2
q2 k2
I + i"
P(q) Discq

Discq

Dispersive Integral
The full correlator can then be written as a sum over residues, using only
three-point data.

Conformally Coupled Scalar
As an example, let us reproduce the result for conformally coupled scalars.
= i log(k1 + k2 + k3)
Discq

• Input: three-point function
• Discontinuity:
=
1
2q
log
✓
k12 q
k12 + q
◆
log
✓
k34 q
k34 + q
◆
= P(q) Discq

Discq

= 4q
Z 1
k12
dx
x2 q2
Z 1
k34
dy
y2 q2

Conformally Coupled Scalar
As an example, let us reproduce the result for conformally coupled scalars.
• Dispersive integral
=
Z 1
k12
dx
Z 1
k34
dy
Z 1
1
dq
⇡i
q2
q2 k2
I
1
x2 q2
1
y2 q2
=
Z 1
k12
dx
Z 1
k34
dy
1
(x + y)(x + kI)(y + kI)
kI
x y
q
Correct!

EFT of Inflation
One of the important features of using unitarity cuts to bootstrap correlators
is that it doesn't rely the interactions being approximately boost-invariant.
Exercise: Consider ˙
⇡3
in the EFT of inflation.
• The three-point wavefunction coeﬃcient is
=
(k1k2k3)2
(k1 + k2 + k3)3
• Using the dispersive integral, show that the exchange 4pt function is
=
(k1k2k3k4)2
E5E3
LE3
R

k2
I
⇣
6E2
LE2
R + 3EELER(EL + ER)
+ E2
(EL + ER)2
+ E3
(EL + ER) + E4
⌘
6E3
LE3
R

We have seen that cosmological correlators satisfy
interesting diﬀerential equations in kinematic space:

u2
(1 u2
)@2
u 2u3
@u +
✓
µ2
+
1
4
◆
=
We derived these equations from conformal symmetry,
but is there a deeper reason for their existence?

In an upcoming paper, we will answer this question for
conformal scalars in a power-law FRW background.
What classes of correlation functions can arise from consistent cosmological
evolution? Surprisingly little is know about this in a systematic fashion.
Arkani-Hamed, DB, Hillman, Joyce,
Lee and Pimentel [to appear]

Conformal Scalars in dS
We have seen that the four-point function of conformal scalars in de Sitter
space can be written as an integral of the flat-space result:
where and .

Conformal Scalars in FRW
An interesting generalization of this integral is
Twist
This describes the four-point function of conformal scalars in FRW backgrounds
with scalar factor
Similar integrals arise for loop amplitudes in dimensional regularization.
In that case, powerful mathematics can be used to evaluate these integrals:
We will show that a similar approach applies to our FRW correlators.
• Twisted cohomology
• Canonical diﬀerential equations

Family of Integrals
To attack the problem, we introduce a family of integrals with the same
singularities (divisors):
where
• Not all of these integrals are independent because of integration-by-
parts and partial fraction identities.
• How many independent master integrals are there?
The answer is provided by a beautiful fact of twisted cohomology.
Aomoto [1975]
Mastrolia and Mizera [2018]

A Beautiful Fact
The number of independent master integrals is equal to the number
of bounded regions defined by the divisors of the integrand.
x1
x2
L1
L2
L3
L4
L5
1
2
3
4

Master Integrals
The vector space of independent master integrals is four-dimensional:
Canonical forms
The basis integrals are not unique, but a convenience choice is

Divide and Conquer
Our original integral of interest is
i.e. it is a sum of all basis integrals. Each basis integral is easier to evaluate!
Key idea:
The basis integral forms a finite dimensional vector space, so taking derivatives
with respect to must lead to coupled diﬀerential equations:
Exterior derivative
Solving this diﬀerential equation turns out to be easy.

Diﬀerential Equation
Explicitly, we find
with

The Solution
More abstractly, we have
This leads to two types of functions:
decoupled
coupled
Power laws
Hypergeometric
In the de Sitter limit, these become the logs and dialogs we found earlier.

The Solution
Imposing appropriate boundary conditions (i.e. singularities), the solution for
our original integral is

Space of Functions
The number of special functions depends on the number of bulk vertices, and
the function have an interesting iterative structure:
Power law
This structure is hidden in bulk perturbation theory.
Just the beginning of a systematic exploration of these mathematical
structures living inside cosmological correlators.

Despite enormous progress in recent years, we are only at the beginning of a
systematic exploration of cosmological correlators:
under-
standing
Many fundamental questions still remain unanswered.
tree level
De Sitter
nonperturbative
Anti-de Sitter
tree level
(some loops)
Minkowski

So far, the cosmological bootstrap has only been applied to individual Feynman
diagrams (which by themselves are unphysical).
However, in the S-matrix bootstrap, the real magic is
found for on-shell diagrams:
What is the on-shell formulation of the cosmological bootstrap?
What magic will it reveal?
• Recursion relations
• Generalized unitarity
• Color-kinematics duality
• Hidden positivity
• …
Beyond Feynman Diagrams

So far, the bootstrap constraints (singularities, unitarity cuts, etc.) are only
formulated in perturbation theory and implemented only at tree level.
In contrast, we understand the conformal bootstrap nonperturbatively:
Recently, some preliminary progress was made to apply unitarity constraints
in de Sitter nonperturbatively, but much work remains.
Hogervorst, Penedones and Vaziri [2021]
Di Pietro, Gorbenko and Komatsu [2021]
A nonperturbative bootstrap will be required if we want to
understand the UV completion of cosmological correlators.
Beyond Perturbation Theory

In quantum gravity, spacetime is an emergent concept. In cosmology, the
notation of “time” breaks down at the Big Bang. What replaces it?
The cosmological bootstrap provides a modest form of emergent time, but
there should be a much more radical approach where space and time are
outputs, not inputs.
For scattering amplitudes, such a reformulation was achieved in
the form of the amplituhedron:
What is the analogous object in cosmology?
How does the Big Bang arise from it?
Beyond Spacetime
Arkani-Hamed and Trnka [2013]

Thank you for your attention!
NTU, Taiwan

Lectures on Cosmological Correlations

More Related Content

What's hot (20)

Similar to Lectures on Cosmological Correlations (20)

More from DanielBaumann11 (8)

Recently uploaded (20)

Lectures on Cosmological Correlations