YITP-SB-13-44

Losing Forward Momentum Holographically

Koushik Balasubramanian and Christopher P. Herzog
C. N. Yang Institute for Theoretical Physics
Department of Physics and Astronomy
Stony Brook University, Stony Brook, NY 11794
Abstract
We present a numerical scheme for solving Einsteins Equations in the presence of a
negative cosmological constant and an event horizon with planar topology. Our scheme
allows for the introduction of a particular metric source at the conformal boundary.
Such a spacetime has a dual holographic description in terms of a strongly interacting
quantum eld theory at nonzero temperature. By introducing a sinusoidal static metric
source that breaks translation invariance, we study momentum relaxation in the eld
theory. In the long wavelength limit, our results are consistent with the uid-gravity
correspondence and relativistic hydrodynamics. In the small amplitude limit, our results
are consistent with the memory function prediction for the momentum relaxation rate.
Our numerical scheme allows us to study momentum relaxation outside these two limits
as well.
a
r
X
i
v
:
1
3
1
2
.
4
9
5
3
v
2


[
h
e
p
-
t
h
]


2
8

J
a
n

2
0
1
4
1 Introduction
Through the AdS/CFT correspondence [1, 2, 3], pure Einstein gravity with a negative
cosmological constant describes a sector of many strongly interacting, scale invariant quan-
tum eld theories (QFTs). The gravitational description allows one to compute proper-
ties of these QFTs at nonzero temperature and density, perhaps providing some insight
into strongly interacting real world systems in nuclear and condensed matter physics (see
[4, 5, 6, 7] for reviews). Traditionally, the AdS/CFT approach has focused on systems at
or near equilibrium. Near equilibrium, we can make a further renement into two cases: a)
linear response where the departure from equilibrium itself is small and the system responds
linearly and b) hydrodynamics where the departure happens over long times and distances
such that any space-time point can be treated as if it were in local equilibrium.
1
Assuming a
particle interpretation, long here means compared to scattering times and lengths. Frontier
questions are then those for which linear response and hydrodynamics both fail, for which
the departure from equilibrium is large and happens on short time scales and distances,
The near equilibrium AdS/CFT approach has been quite valuable because strong inter-
actions alone make it dicult to use traditional eld theoretic techniques, but we would like
to go further and address some of the many interesting questions surrounding far from equi-
librium behavior. What does turbulence look like when the driving happens on short time
scales and wavelengths? How do shock waves behave in the limit when the shock becomes
very thin? How does uid ow around a rough object when the scale of the roughness is of
order or smaller than the mean free path? More generally, when precisely do hydrodynamics
and linear response cease to be valid and what should replace them? AdS/CFT combined
with numerical relativity provide methods to address these questions.
In this paper, we shall examine the case where the eld theory is pushed away from
equilibrium by a source for g
tt
, the time-time component of the metric. More precisely,
we will be solving the Einsteins equations that follow from the following four dimensional
action:
S =
1
2
2
_
d
4
x

g (R 2) , (1)
where the cosmological constant < 0. These equations admit a large family of solutions
which have an asymptotic region that approaches four dimensional anti-de Sitter space.
This AdS
4
region in turn possesses a conformal boundary. We will work in the Poincare
1
Note that there are some examples where a hydrodynamic description seems to be valid even when the
local thermal equilibrium assumption does not apply for instance, hydroization [8, 9].
1
patch where this conformal boundary, in the absence of our g
tt
source, is conformal to three
dimensional Minkowski space. Through the AdS/CFT correspondence, the four dimensional
spacetime is dual to a large class of three dimensional strongly interacting conformal eld
theories. The metric g
tt
source in the eld theory acts in gravity as a boundary condition for
the metric at the conformal boundary. When examining linearized perturbations around
a translation invariant equilibrium state, the system has more symmetry, and Einsteins
equations reduce to ordinary dierential equations. In our far from equilibrium context,
there will be no such simplication.
We plan for this paper to be the rst of a series. Here we detail our methods and provide
a simple application. In the future, we will look at some more involved applications. In the
limit in which the eld theory has a hydrodynamic description, the metric source allows us
to recreate many classic uid mechanics experiments. By introducing a time dependent,
random g
tt
, we can drive a turbulent state. We would like to know if such a driven state
satises the Kolmogorov scaling law. Some partial evidence for this scaling was found in
the decaying, non-driven turbulence of [10].
2
While we do not know how to introduce
boundaries with no-slip conditions in gravity, g
tt
allows us to do the next best thing. For
example, we can make a numerical wind tunnel where we construct a region with a dierent
value of g
tt
and then drag it through the uid at a velocity of our choosing. For supersonic
ows, such an experiment would be a worthwhile generalization of [13, 14] where the source
was eectively point-like and gravity was linearized. It would also be interesting to nd the
dual gravity description of Stokes drag of an object for a ow with low Reynolds number.
The application we focus on in this paper is very similar in spirit to the Stokes drag
problem. Starting with uniform subsonic uid ow, we measure the momentum loss in the
presence of a metric source g
tt
= 1 cos(kx). By making g
tt
depend only on x, our
gravity simulation becomes eectively 2+1 instead of 3+1 dimensional, allowing us to get
results more quickly.
3
In the hydrodynamic regime (k/T 1), it is possible to obtain a heuristic estimate of
the momentum relaxation time scale using an analogy with Stokes formula for drag. Note
that an object moving in a viscous uid in the absence of external body forces will come
to rest due to the eect of viscous drag. When viewed from the rest frame of the object,
the ow relaxes to zero velocity after a certain time. Using Stokes formula for the drag
2
See also refs. [11, 12] for related work on turbulence in 2+1 dimensional relativistic ows.
3
The related problem of a statonary ow over a metric bump was considered in [15].
2
coecient C
d
, we write
C
d

1
Re


UL
where Re is the Reynolds number, U is the characteristic ow velocity, L is the character-
istic length scale of the object, is the viscosity and is the mass density of the uid. Note
that the drag coecient for a sphere is 24/Re. Now to estimate the lattice induced momen-
tum relaxation time scale , we choose the object to be a lattice with spacing L = 2/k.
Note that the ow decelerates to zero velocity in a time scale due to presence of the
obstacle; hence the characteristic velocity is U = L/. Assembling the estimates yields
1

1
C
d
k
2

.
In Stokes formula, the drag coecient C
d
is a dimensionless quantity which depends on the
shape or aspect ratio of the object. In the case of a lattice, we expect C
d
to be a function
of the lattice strength .
In this paper, we obtain the relaxation time scale for ows using relativistic hydro-
dynamic simulations and numerical general relativity. We show that for small k/T, both
hydrodynamic and gravity results agree with the above heuristic argument. When is
small, the system is near equilibrium where we check our results using linear response (more
specically the memory function method [16, 17, 18]). These cross checks give us faith in
the robustness and accuracy of our code as we move toward addressing more challenging
problems.
Additionally, our simple experiment may have some interesting condensed matter ap-
plications. As mentioned above, much eort has gone into trying to apply AdS/CFT to
certain strongly correlated condensed matter systems. However, simple AdS/CFT models
typically possess translation symmetry which leads to zero DC resistivity
DC
, inconvenient
for example if one wants to explain the linear rise with temperature in
DC
of a high T
c
superconductor. As a result, much eort has gone into investigating the eects of transla-
tion symmetry breaking in these models (see for example [17, 19] for early papers). Most
of these papers break translation invariance using an external gauge eld (i.e. a spatially
dependent chemical potential) or a source for a scalar. Our system is in some sense simpler
and more economical as the translation breaking is provided by the metric itself, requiring
the addition of fewer extra tuneable parameters. Our code allows us to study the region
where both and k are large and standard techniques fail. (More precisely, we mean 1
and 1 k/T.) We also provide an analytical technique to compute the relaxation time
scale for large and small k using perturbative techniques in the hydrodynamic descrip-
3
tion. This perturbative result agrees with the numerics within numerical error. To our
knowledge, the regime of large has not been explored before in the literature. Though
we focus on subsonic ows in the present paper, it is possible to use the current numerical
framework to explore ows with large velocities as well.
The rest of the paper is summarized as follows. In section 2, we will describe our
numerical scheme for solving Einsteins equations in a spacetime that is asymptotically
AdS
4
and that possesses an event horizon with planar topology. In section 3, we describe
our numerical scheme for solving the relativistic hydrodynamic equations. Section 4 contains
the results of our simple experiment. In appendix A, we give more details how the numerical
codes were implemented and cross checked. Appendix B describes how to determine the
momentum relaxation rate from hydrodynamics in the small limit for arbitrary constant
ow velocity.
2 Numerical Gravity
In this section, we describe the method used for solving Einsteins equations in the presence
of a negative cosmological constant. Our metric ansatz was inspired by ref. [20], and our
computational scheme is very similar to those described in refs. [9, 20, 21]. (See also ref. [22].)
We use a characteristic scheme [20] combined with spectral methods [23, 24, 25]. For
hyperbolic partial dierential equations (PDEs), a characteristic evolution scheme reduces
the PDEs to ordinary dierential equations (ODEs) along the characteristics. In numerical
relativity, the characteristic formalism is based on a choice of frame of reference where
outgoing null cones evolve from an initial null cone.
1. Bondi-Sachs null coordinate system
Our characteristic formalism in relativity is based on a geometry that is described by the
Bondi-Sachs metric:
ds
2
=
_
e
2
V z
h
AB
U
A
U
B
z
2
_
dt
2

2e
2
z
2
dt dz
2h
AB
U
B
z
2
dt dx
A
+
h
AB
z
2
dx
A
dx
B
. (2)
We would like for this metric to describe a four dimensional space-time. Thus A, B = 1, 2.
We will x det h = e
4(t,x,y,z)
by choosing
h = e
2
_
_
e

cosh sinh
sinh e

cosh
_
_
. (3)
This coordinate system (up to the redenition z = 1/r) is described on p 32 of [20].
4
Vacuum Einsteins equations in the presence of a negative cosmological constant < 0
are
G

= 0 . (4)
With our metric ansatz, the system is under-determined. By redening the radial coordi-
nate, we can eliminate either or . Motivated by the form of the boosted black brane
metrics, we will eventually partially x the gauge by taking
(t, x, y, z) =
1
4
ln
_
1 + 2z
3

3
(t, x, y)
_
. (5)
The remaining gauge degrees of freedom in
3
will be used to place an apparent horizon at
z = 1, allowing us to truncate the region z > 1 and solve Einsteins equations on a constant
radial domain 0 < z < 1. (We choose the radial coordinate such that z = 0 corresponds to
the conformal boundary.)
4
2. Apparent horizon
The fact that light cannot escape an apparent horizon classically allows us to restrict the
computational domain to the region outside the apparent horizon. Since the null rays on
the apparent horizon are all converging, information behind the horizon is not propagated
into the computational domain. We use a coordinate system where the apparent horizon
is located at a xed radial coordinate. Spectral methods are much simpler to implement
when the domain is a box of xed size independent of time.
Before we proceed to the equations of motion, we would like to derive the constraints
imposed by xing the apparent horizon at z = 1. Consider a vector eld

associated to
the tangents of a congruence of null geodesics. In other words, consider a vector eld that
satises the conditions

= 0 and

;
= 0. Given the conditions on

, the expansion
is given by =

;
. An apparent horizon corresponds to a surface where = 0.
We parametrize

in terms of the gradient of a hypersurface S = 0 orthogonal to the

geodesic congruence and an arbitrary rescaling function R:

= RS
,
. Without knowing
the precise form of S, we can solve for S
;t
using the null condition

= 0. Note that the

null condition also implies that S
;
S
;
= 0. The geodesic condition

;
= 0 implies the
orthogonality S
;
R
;
= 0, allowing us to solve for R
;t
.
It turns out that we need one and one hypersurface equation S for each instant of
time t, so lets make them functions of t: [t] and S[t]. To nd the apparent horizon at
4
At this point our numerical strategy diers somewhat from [9, 21]. Those authors chose to x instead.
5
an instant of time t
0
, we choose S[t
0
] such that S[t
0
](t
0
, x, y, z) = z 1. Note that more
generally S[t
0
](t, x, y, z) ,= z 1. Substituting S[t] into , we conclude that
z
2
(z

1)V D
A
U
A
2

z=1
= 0 , (6)
where

=
z
and =
t
.
Note that a trapped surface need not always exist. Even if it exists, it may not be
simply connected which is not very helpful for restricting the computational domain. In
certain cases, the characteristic evolution could encounter caustics before they reach the
apparent horizon. Some methods to deal with specic caustics (such as point caustics) in the
characteristic formalism are known [26]. In the present paper, we work with smooth initial
and boundary conditions that do not lead to formation of caustics within the domain. It is
not clear if caustics can always be avoided in the characteristic formalism for asymptotically
AdS solutions, but it seems unlikely.
We will now present the equations of motion with the above choice of coordinates and
gauge.
3. Equations of motion
Einsteins equations for the Bondi-Sachs metric have a convenient nested structure that
allows for ecient numerical integration. In the characteristic formalism, , U
A
and V on
a time slice can be determined if the values of , , and on that time slice are known. We
will also need to assume knowledge of z = 0 boundary values of and U
A
. As we have not
yet discussed a z = 0 expansion of the solution, we save a detailed discussed of boundary
conditions for later.
The scheme rst determines on a constant t slice. From the zz component of Einsteins
equations, we nd
G
zz
=
4
z
(1 +z

1
2
_
4(

)
2
+ (

)
2
cosh
2
+ (

)
2

, (7)
where f

z
f. This rst order dierential equation is solved for after specifying the
boundary conditions on at z = 0.
From G
Az
, we obtain
G
Az
=
z
2
e
2
2

1
z
2
e
2
(z
2
e
2

A
)

+
1
2
h
BC
D
C
h

AB
2D
A

, (8)
where we have dened

A
z
2
e
2()
h
AB
(U
B
)

. (9)
6
This dierential equation is solved in two steps, rst for
A
and then for U
A
, and requires
boundary data for U
A
.
From the combination h
AB
G
AB
we nd
h
AB
G
AB
= 4e
2(+)
z
2
_
e
2
z
2
d
t

1
2
e
4
h
AB
(U
A
)

(U
B
)

2
z
2
(10)
+!2e
2
D
A
e
2
h
AB
D
B
e

+z
4
e
2(+)
D
A
e
2
_
e
4
U
A
z
4
_

.
We have dened
d
t

z
2
2
(z

1)V .
Here ! is the Ricci scalar computed from the 2 2 metric h
AB
and D
A
is the associated
covariant derivative. This equation is solved for d
t
by using the apparent horizon condition
(6) as a boundary condition at z = 1.
The remaining two linearly independent combinations of R
AB
g
AB
allow one to solve
for
d
t
=
z
3
2
V

, (11)
d
t
=


z
3
2
V

. (12)
To wit, we have
G
AB
= e
2
_
ze

_
e

z
d
t

h
AB
_

1
2
e
2
h
AB
tr[h

(d
t

h)] (13)
2e

D
A
D
B
e

1
2
e
2
h
AC
h
BD
(U
C
)

(U
D
)

+
1
2
_
h
AB
z
2
_

D U
e
2
(

h
C(A
)

(D
C
U
B)
D
B)
U
C
) +h
C(A
D
B)
(U
C
)

+ 2
_

1
z
_
D
(A
U
B)
+h

AB
d
t
+ (D
C
h

AB
)U
C
+
_
1
2
e
2
!+ 2z
2
e

_
e

z
2
d
t

z
2
e
2
_
h
AB
_
,
where we have dened the normalized spatial metric

h
AB
e
2
h
AB
. The two dierential
equations are solved with Dirichlet like boundary conditions at z = 0.
The horizon value of V can be obtained from the equation G
z
t
+U
A
G
z
A
. Using (6) and
(7), this equation reduces to the following elliptic equation on the apparent horizon z = 1:
G
z
t
+U
A
G
z
A
[
z=1
=
1
2
D
2
V
1
2
_
e
2
U

+ 2D
_
DV (14)

_
d
t

+
1
2
D U U D

_
e
2
V +e
2()
(d
t

h
AB
)D
A
U
B
+
1
4
e
2(2)
tr[(d
t

h) (d
t

h)]
+
1
2
e
2
_
(D
A
U
B
)(D
A
U
B
) + (D
A
U
B
)(D
B
U
A
) (D U)
2

[
z=1
.
7
Note that this elliptic equation also plays a key role in the integration strategy described
in [21].
Once the horizon value of V is obtained, at the horizon can be deduced from the
apparent horizon condition (6). Having xed the radial dependence of through a gauge
choice, we can compute everywhere. The functional form of V can then be reconstructed
from the denition of d
t
. With V in hand, and

can be computed from the denitions
of d
t
and d
t
.
Modulo boundary conditions which we will discuss momentarily, given ,

, and , we
can compute , , and on the next time slice. This whole process can then in principle
be iterated, carrying the solution forward an arbitrary number of time steps.
4. Boundary expansion
We make the gauge choice (5) for . Let us also assume that the sources for , , and U
A
vanish. Near the conformal boundary z = 0, we nd the expansions
V =
1
z
3
(V
0
e
2
0
+V
2
z
2
+V
3
z
3
+O(z
4
)) , (15)
=
0
+
3
z
3
+O(z
6
) , (16)
U
A
= U
A
1
z +U
A
3
z
3
+O(z
4
) , (17)
=
3
z
3
+O(z
4
) , (18)
=
3
z
3
+O(z
4
) . (19)
The function
0
(t, x, y) is a source term. The functions V
3
(t, x, y),
3
(t, x, y),
3
(t, x, y),
and U
A
3
(t, x, y) are integration constants which determine the stress tensor in the dual eld
theory. We nd that
V
2
= (
2
x
+
2
y
)e
2
0
,
3
=
1
2

3
, U
A
1
=
A
e
2
0
. (20)
The (t) components of Einsteins equations each require V
0
= /3. In addition to the
source term
0
, our boundary data consists of ve parameters: V
3
,
3
,
3
, and U
A
3
. Five
parameters are exactly what is needed to describe a traceless stress energy tensor. Dening
the boundary stress tensor in the usual way [27] as
T
ab
= lim
z0

V
0
z
_
K
ab
(K + 2
_
V
0
)
ab
+
1

V
0
_
!
ab

1
2
!
ab
__
, (21)
8
where
ab
is the induced metric on a slice of constant z, K
ab
is the associated extrinsic trace,
and !
ab
is the three dimensional Ricci tensor. We nd that:
T
tt
= V
0
V
3
e
2
0
2e
4
0
V
2
0

3
, (22)
T
tA
=
3
2
V
0
U
A
3
+e
2
0

A
[e
4
0
(
C

0
)] , (23)
T
xx
=
1
2
e
2
0
V
3

3
2
V
0

3
+
e
6
0
V
0
_

x
((
x

0
)e
4
0
)
y
((
y

0
)e
4
0
)
_
(24)
V
0

3
,
T
xy
=
3
2
V
0

3
+
e
6
0
V
0
_

y
((
x

0
)e
4
0
) +
x
((
y

0
)e
4
0
)
_
, (25)
T
yy
=
1
2
e
2
0
V
3
+
3
2
V
0

e
6
0
V
0
_

x
((
x

0
)e
4
0
)
y
((
y

0
)e
4
0
)
_
(26)
V
0

3
.
Note that
ab
T
ab
= 0.
Given the derivation of (21) from a variational principle, we are guaranteed that this
stress tensor is covariantly conserved on the boundary z = 0, i.e.
a
T
ab
= 0. (With
0
= 0,
the boundary is at and we have the stronger condition
a
T
ab
= 0.) These conservation
conditions impose the following dierential relations on the ve parameters:
2
t
(e
2
0
V
3
) = 3V
0
e
2
0

A
(e
2
0
U
A
3
) + 2e
2
0

A
[e
4
0
(
C

0
)]
4
3
, (27)
3
t
(e
2
0
U
1
3
) = e
4
0

x
(V
3
e
4
0
) 3V
0

x
(
3
e
2
0
) 3V
0

y
(
3
e
2
0
) (28)
6
3

x
e
2
0
2e
2
0

3
,
3
t
(e
2
0
U
2
3
) = e
4
0

y
(V
3
e
4
0
) + 3V
0

y
(
3
e
2
0
) 3V
0

x
(
3
e
2
0
) (29)
6
3

y
e
2
0
2e
2
0

3
.
These dierential relations can also be obtained by systematically solving the equations
of motion order by order in z. For example, considering the (t) along with the (Az)
components of Einsteins equations yields the three relations above in addition to conditions
on V
4
and U
A
4
.
The six bulk equations of motion G
zz
, G
Az
, and G
AB
that we use are a subset of the
ten Einsteins equations. It is an interesting exercise to see how our integration scheme
above guarantees that the four equations G
t
are also satised. With some eort, one can
establish that G
tz
is a linear combination of G
zz
, G
Az
, G
AB
and their derivatives. Although
we do not use G
tt
and G
tA
in the bulk, we do use them to set boundary conditions. In
particular, we use them to propagate the boundary values of U
A
and also to derive the
9
elliptic equation (14) used to set the value of V at the horizon. That G
tt
and G
tA
are
satised everywhere then follows from a Bianchi identity, as we now argue. Let us dene
H

= R

1
2
Rg

.
The contracted Bianchi identity implies that

= 0 =
1

gH

) +

= 0
is satised as an algebraic identity. Now H
z
z
and H
z
A
depend linearly on G
tt
and G
tA
respectively. Additionally, H
t
z
, H
t
A
, H
B
z
and H
B
A
are independent of G
tt
and G
tA
. Thus
if G
tt
and G
tA
are satised at some point in z, integrating the Bianchi identity in the z
direction, they must be satised everywhere in the interval 0 z 1. The way we set the
boundary conditions for U
A
guarantee that G
tA
= 0 are satised at z = 0 so they must be
satised everywhere. The way we set boundary conditions for V then guarantees that G
tt
is satised at z = 1. Thus G
tt
= 0 everywhere. A way of monitoring the accuracy of our
integration scheme is to check how well G
tt
is satised at z = 0, in other words to monitor
(27).
5. Marching orders
We are now ready to specify precisely what elds we numerically integrate and which
boundary conditions we apply. We dene new functions with subscript s.
=
0

z
3
2

3
+z
4

s
,
U
A
= z
A
(e
2
0
) +z
2
U
A
s
,
A
=
2
z
2

0
+
A
s
,
V =
1
z
3
(V
0
e
2
0
+z
2
V
s
) , d
t
=
e
2
0
2z
+
z
2
(
2
x
+
2
y
)e
2
0
+z
2
e
2
d
t

s
,
= z
2

s
, = z
2

s
, d
t
= zd
t

s
, d
t
= zd
t

s
.
The denition of the subscripted s functions while somewhat arbitrary is guided by some
underlying principles. At a minimum, we are required to subtract singular terms from the
metric functions so that the boundary conditions are well behaved at z = 0. It is then
convenient to rescale the subtracted metric functions by powers of z such that the stress
tensor can be extracted without trying to compute a high order numerical derivative of the
solutions.
We then numerically integrate to nd the subscripted s functions. We impose the
10
following boundary conditions on these functions at the singular point z = 0:

s
= 0 , (30)

A
s
= 3e
2
0
U
A
3
, (31)

z
U
A
s
= U
A
3
, (32)
d
t

s
= 0 , d
t

s
= 0 . (33)
In integrating the d
t
equation, we are faced with a choice. We can either apply the Dirichlet
condition
d
t

s
=
1
2
V
3

3
4
e
2
0

3
,
at z = 0 or the apparent horizon Dirichlet condition (6) at z = 1. We choose the latter
as it allows us not to propagate the boundary value of V
3
forward in time. In integrating
the U
A
s
and
s
equations, we were also faced with a choice. We could have applied Dirich-
let conditions U
A
s
= 0 and
s
= 0 at z = 0 instead. However, we nd in general that
in discretizing the dierential operator that we need then to invert, Neumann boundary
conditions produce matrices with a lower condition number.
To have a well dened Cauchy problem, we also need to give initial conditions. In
this case, a set of good Cauchy data is provided by bulk data for and and boundary
data for
3
and U
A
. As our choice of initial conditions will be guided by a particular
hydrodynamics problem, let us postpone a discussion until after we have reviewed some
facts about relativistic conformal hydrodynamics.
We relegate to appendix A the precise numerical details of our algorithm.
3 Numerical Relativistic Conformal Hydrodynamics
Thermal eld theories generically admit a hydrodynamic description of their long wave-
length, low frequency modes provided the wavelengths are long compared to the mean free
path of the particles and the frequencies are small compared to inverse scattering times.
In particular, eld theories with gravity duals admit such a description. The goal of this
section is to write down hydrodynamic equations whose numerical solutions can be com-
pared with the numerical solutions of the gravity model in the same low frequency, long
wavelength regime.
In our case, the eld theory dual to our gravity model is both relativistic and conformal
which puts some additional constraints on the constitutive relations for the stress tensor.
The only scale in our eld theory is the temperature T. Thus a hydrodynamic description
11
will be valid when the typical wavelength 1/T and the typical frequency T. We
follow [28] in our description. We assume the stress tensor has the form
T

= ( +p)u

+pg

(34)
where we dene

recursively
5
in a gradient expansion

_
(D)

+
d + 1
d

( u)
_
+
_
R

(d 1)u

_
+

+
3

. (35)
Conformality implies tracelessness of T

which in turn yields a relationship = d p between

the energy density and pressure p in d spatial dimensions. To unpack these expressions,
we need a number of subsidiary denitions. We are working with a metric with mostly plus
signature. The four velocity u

has norm u

= 1. (In the uid rest frame in Minkowski

space u

= (1, 0, . . . , 0).) The derivative D u

. The vorticity is

1
2

) , (36)
where we have dened a projector onto a subspace orthogonal to the four velocity:

+u

. (37)
The shear stress tensor is

. (38)
The angular brackets on the indices indicate projection onto traceless tensors orthogonal
to the four velocity:
A

1
2

(A

+A

)
1
d

. (39)
Note that with these denitions, both

and

are traceless and orthogonal to the four

velocity
u

= u

= 0 ,

= 0 .
5
The implicit denition of

makes our formulation of the hydrodynamic equations Israel-Stewart like.

Formally, higher than second order gradient corrections are present in the denition of

. However, if one
wanted a third or higher order accurate formulation, additional terms should be added to the denition of

. Note some second order terms are necessary in order for stability of the numerics.
12
In 2+1 space time dimensions (even in curved space), it can be shown that the coecients
of
1
and
3
vanish. Thus, we are left with the four transport coecients ,

, , and
2
.
We will be interested in what follows in a metric of the form ds
2
= g(t, x, y)dt
2
+dx
2
+dy
2
.
In this case, the Weyl curvature vanishes and we can forget about as well. For a uid
dual to pure Einstein gravity in 3+1 dimensions, was rst computed in [29],

in [30],
and
2
in [31]. Assuming a normalization of the energy density where
=
_
4T
3
_
3
, p =
1
2
_
4T
3
_
3
(40)
(such that 2p = 0), we have
=
1
2
_
4T
3
_
2
,

=
1
36
_
4T
3
_
_

3 9 log 3 + 18
_
,

2
=
1
36
_
4T
3
_
_

3 9 log 3
_
.
For the numerics, we will use the relations

= 0 and u

= 0 to eliminate all of the

components of

except for
xx

yy
B and
xy
. The three conservation conditions

= 0 along with the recursive denitions of B and

xy
become the ve dierential
equations used to propagate the ve variables T, u
1
, u
2
, B, and
xy
. A similar numerical
scheme was described and implemented in ref. [32].
3.1 Boosted Black Brane Background with g
tt
and general
The relationship between hydrodynamics and gravity can be made arbitrarily precise [33] in
the sense of an asymptotic series. In principle, a solution to the hydrodynamic equations of
motion at a given order in the gradient expansion can be integrated in the radial direction
of the gravity spacetime to provide a solution to Einsteins equations, accurate at the same
order in a gradient expansion. Here, we review how to perform this matching at the zeroth
level in the gradient expansion for ideal hydrodynamics where the eects of viscosity and
the other transport coecients can be ignored. In practice, this matching is important for
us as it allows us to choose the same initial conditions in our gravity and hydrodynamic
simulations.
The matching makes use of the boosted black brane metric. This metric has a dual
interpretation as a uid moving with four velocity u

and temperature T. The line element

13
for the boosted black brane metric is
ds
2
=
1
r
2
[2u

dx

dr f(br)u

dx

dx

+P

dx

dx

] , (41)
P

= g

+u

, g

+ (1 g)
0

0
, (42)
f(r) = 1 r
3
. (43)
The temperature is given by the relation T = 3b/4. The horizon is at r = 1/b; the
conformal boundary at r = 0. When b and u

are constant, this solution is an exact

solution of Einsteins equations. More generally, the metric will only be accurate at zeroth
order in gradients. Note that we have allowed for an arbitrary g
tt
= g.
We would like to put this metric into our null coordinate system. To that end, we
consider a coordinate transformation x

= y

(r) where for the moment we do not alter

r. The vector is chosen such that the g
r
components of the transformed metric vanish.
The line element of the transformed metric takes the form
ds
2
=
1
r
2
_
_
g

+ (br)
3
u

_
dy

dy

2g
1/2
_
f(br) + (u
0
)
2
(br)
3
g
dt dr
_
. (44)
In our coordinate system, we have specied the functional form of the spatial determi-
nant:
e
4(z)
z
4
=
1
r
4
_
f(br) (br)
3
u
0
u
0
_
, (45)
which denes r as a function of z. In the z coordinate system, we place the horizon at z = 1
which relates (1) to b and u
0
:
e
2(1)
= b
2
u
0
g
1/2
. (46)
We can reconstruct from the g
tr
component of the transformed metric and the deriva-
tive dr/dz:
e
2
=

g
1 +u
A
u
A
(br)
3
z
2
r
2
dr
dz
. (47)
The other dening functions of the transformed metric are straightforward to reconstruct
e

1 + (u
1
)
2
(br)
3
1 + (u
2
)
2
(br)
3
, sinh = u
1
u
2
e
2(z)
b
3
rz
2
, (48)
U
A
=
u
A
u
0
(br)
3
1 +u
A
u
A
(br)
3
, V =

gf(br)
_
1 +u
A
u
A
(br)
3
1
z
3
dz
dr
. (49)
The relation (45) suggests a natural choice for , namely
=
1
4
log(1 +u
A
u
A
z
3
) + log b . (50)
14
Allowing us to make the identications
rb = z , 2
3
= u
A
u
A
. (51)
With our gauge choice for , these relations reduce to the simpler
e
2
=
b

g
_
1 +u
A
u
A
z
3
, (52)
e
2
=
1 + (u
1
)
2
z
3
1 + (u
2
)
2
z
3
, sinh =
u
1
u
2
z
3
_
1 +u
A
u
A
z
3
, (53)
U
A
=
u
A
u
0
z
3
1 +u
A
u
A
z
3
, V =
bf(z)

g
z
3
_
1 +u
A
u
A
z
3
. (54)
Seting b = 1 recovers the gauge choice (5) described above.
4 A Simple Experiment
For our simple experiment, we start with a constant uid ow in the x direction at time
t = 0 and time-time component of the metric of the form
g
tt
= (1 + cos(kx)e
m/t
). (55)
At time t = 0, the metric reduces to Minkowski space while for t m, the uid experiences a
roughly constant sinusoidal potential in the x-direction. As the potential breaks translation
invariance, we expect that the uid velocity will eventually relax to zero. A nonzero m is
used solely to increase the stability of the numerical simulations. In our analytic estimates,
we assume m = 0. We expect these estimates to be valid for times t m.
4.1 Analytic Estimates of Momentum Relaxation
Before entering a discussion of our numerical simulations, let us begin with three analytic
estimates of the momentum relaxation rate. The rst is valid in the regime where hydrody-
namics and linear response are both valid, the second when linear response is to be trusted,
and the third when hydrodynamics is valid.
The technique we shall use in both the rst and second cases is called the memory
function formalism and relies on the validity of linear response. In other words, the metric
source must be small 1. (Although not necessary, we will also assume the uid velocity
is small. We relax this assumption in appendix B.) This method was rst used in a holo-
graphic context by ref. [19]. Later uses include refs. [17, 18]. The method is described in
detail in the book [16].
15
In this method, we break translation invariance by adding the following perturbation to
the action

_
O(x)e
ikx
dx , (56)
where O(x) is an arbitrary operator. The memory function formalism can be used to deduce
the relaxation time:
1

=

2
k
2
+p
_
lim
0
ImG
OO
R
(, k)

=0
_
, (57)
where G
OO
R
(, k) is the Fourier transform of the retarded Greens function. The method
is perturbative, and one may worry about the running of with energy scale. In our
experiment, in order for this method to be reliable, we need to remain small at the scale
set by the temperature T. (In the limit T 0, this restriction would mean that O(x)
should be marginal or irrelevant.) In our case, we are perturbing the metric component g
tt
which couples to the energy density T
tt
/2, a marginal operator.
6
Thus we need to compute
the retarded Greens function for the energy density G

R
(, k).
1. Hydrodynamic regime
We rst compute this Greens function purely in the hydrodynamic limit where k T.
This Greens function has a universal form (see for example [34]):
G

R
=
k
2
( +p)
k
2
(c
2
s
+i)
2
+ , (58)
where c
2
s
= p/ is the speed of sound squared while the damping constant is
=
1
+p
_
2(d 1)
d
+
_
.
As it is simple to do, we have kept the spatial dimension d arbitrary and restored the bulk
viscosity , which will not be present for a conformal uid. Plugging this universal form
into the expression for the relaxation time yields
1

=

2
k
2
4sTc
4
s
_
2(d 1)
d
+
_
. (59)
The speed of sound appearing in the denominator implies that this relaxation eect will
disappear in the incompressible limit where c
s
and sound waves can be neglected.
For a cosine perturbation instead of the e
ikx
dependence considered here, we multiply the
result by a factor of one half.
6
The factor of two comes from the canonical denition of the stress tensor.
16
2. Short wavelength limit
Next we compute the Greens function directly from gravity. The result should be applicable
away from the hydrodynamic limit, where k is no longer necessarily small. We follow the
gauge invariant formulation of ref. [35] where a corresponding calculation was performed
for a ve dimensional spacetime. We begin instead with the black brane metric for a four
dimensional space-time:
ds
2
=
1
z
2
_
f(z)dt
2
+dx
2
+dy
2
+
dz
2
f(z)
_
,
where f(z) = 1 z
3
. Thus, the horizon of our black hole is at z = 1 and we will be
measuring everything in units of the horizon radius. (At the end of the day, we can restore
the temperature dependence by performing the rescalings 3/4T and k 3k/4T.)
We consider small dieomorphisms of the form g

+
;
+
;
where we restrict

(z)e
it+ikx
. We nd that the following linear combination of metric uctuations,
again restricted to have the form h

= h

(z)e
it+ikx
, is invariant under such gauge
transformations:
Z = z
2
_
k
2
h
tt
+ 2kh
tx
+
2
h
xx
+
_
k
2
f(z)
2

z
2
k
2
f

(z)
_
h
yy
_
. (60)
This gauge invariant combination satises the second order linear dierential equation
Z

+
_
2
z

f

f
+
6k
2
z
2
h(, k; z)
_
Z

+
_

2
f
2

k
2
f
+
9k
2
z
(3k
2
4
2
)f
+
36k
2
(k
2

2
)z
(3k
2
4
2
)h(, k; z)
_
Z = 0 , (61)
where we have dened h(, k; z) k
2
(4 +z
3
) + 4
2
.
To deduce the retarded Greens function, at the horizon z = 1, we enforce ingoing
boundary conditions
Z = (1 z)
i/3
(C(, k) +O(1 z)) . (62)
At the conformal boundary z = 0, we nd a series expansion of the solution
Z(z) = A(, k)(1 +O(z
2
)) +B(, k)(z
3
+O(z
5
)) . (63)
Up to real contact terms, the retarded Greens function can then be extracted from the
ratio B/A. There is an overall normalization missing from this Greens function which we
can choose to gain agreement with the hydrodynamic result (59) along with the choice (40).
We will see that in units where T = 3/4, we need
G

R
(, k) =
3
2
B(, k)
A(, k)
. (64)
17
We are interested in only the imaginary part of the Greens function for which it is
enough to know the absolute values [A[ and [C[. To see why, it simplies the algebra to put
the dierential equation in Schrodinger form. We dene a new wave function such that
Z =
z

f
(4
2
+k
2
(4 +z
3
)) , (65)
In this case, the dierential equation reduces to

+V = 0 where the potential is given

by
V =
2
z
2
+
k
2
_
18z
4
+h(, k; z)
_
(1 z
3
)h(, k; z)
+
18k
4
z
4
h(, k; z)
2

4z
2

2
+ 9z
6
4z
2
(1 z
3
)
2
. (66)
We now nd boundary behaviors
(z) =
_
1
z
+. . .
_

A+ (z
2
+. . .)

B and (z) (1 z)
i/3+1/2

C (67)
at z = 0 and z = 1 respectively. The ratio

B/

A = B/A is invariant. Given complex
conjugate solutions and

for real , their corresponding Wronskian must be constant.

In other words, there is a conserved probability current. This conservation condition implies
Im
B
A
=
[

C[
2
[

A[
2

9
. (68)
Figure 1 is a numerical determination lim
0
ImG

R
(, k)/2. For small k, the value
is very close to 1 while for large k, the Greens function is exponentially damped. We can
recover these two limits analytically. For small k, the dierential equation can be solved in
a hydrodynamic expansion where k, 1. The answer is
Z(z) = Cf(z)
i/3
_
1 +
k
2
f(z)
4
2
3k
2

4i
3
+. . .
_
. (69)
Note the leading f(z)
i/3
factor enforces ingoing boundary conditions at the horizon z = 1.
From this answer it follows that
lim
k0
lim
0
Im
B(, k)
A(, k)
=
4
3
. (70)
Next, we can solve this dierential equation in a WKB limit when k 1. As described
in [36], the imaginary part of the stress tensor Greens function is given by the tunneling
probability through the potential V . In the large k limit, this potential reduces to k
2
/f.
Taking

A = 1, the WKB connection formulae give that

C exp
_
k
_
1
0
dz
_
f(z)
_
= exp
_
k

(4/3)
(5/6)
_
. (71)
18
0 2 4 6 8 10
2.3k
8
7
6
5
4
3
2
1
0
l
o
g

I
m
(
G
(

)
)
2

Figure 1: The variation of lim

0
ImG()
2
as a function of the lattice wave number. The
black curve shows the result obtained from solving (61) numerically, i.e. linearized gravity.
The dotted red line shows the approximate behavior of the Greens function for large values
of k. The markers , , , and show the values obtained by solving the full nonlinear
gravity equations for k = /50, 4/50, 5/50, 6/50 and 20/50 in the linearized regime
( = 0.2). Note that k is expressed in units where 3/(4T) = 1. The x-axis is rescaled so
that the red line has slope minus one.
Numerically, the factor multiplying k in the exponent is approximately 1.40. The imaginary
part of the energy density Greens function should then scale as [

C[
2
. In this WKB analysis,
we have swept under the rug subtleties associated with the quadratic singularities in the
potential at z = 0 and z = 1 and also the fact that there is a classical turning point in the
potential at z
0
< 1 an O(k
2
) distance from z = 1. These subtleties produce 1/k
2
corrections
to the tunneling amplitude. In our numerics, we have not been able to get to large enough
values of k to see the scaling (71) although we do see exponentially damped behavior with
a slightly smaller exponent (2.3 instead of 2.80) in Figure 1.
19
3. Small uid velocity limit for arbitrary
When is large, it is not possible to use linear response theory to compute the relaxation
time scale. It is possible to obtain an expression for the relaxation time scale by studying
linearized uctuations around the exact solution of second order hydrodynamics, to which
the system must relax eventually. We can check that the following is a solution of the
hydrodynamic equations
7
(with ,= 0):
u
1
= u
2
= 0 ; T =
T
0
_
g(x)
. (72)
We expect that for any arbitrary initial conditions, the relativistic ow will relax to the
above solution in the presence of viscosity. Note that the parameter T
0
appearing in the
solution is related to the spatial average of the nal steady state temperature as follows

T
f
=
1
L
_
L/2
L/2
Tdx =
_
2T
0

1
_
K
_
2
1
_
(73)
where K() is the quarter period of the Jacobi elliptic function and
2
is the elliptic modulus.
We will make use of this expression when we present the numerical results.
In the limit where the ow velocity is small, the temperature and velocity proles can
be expanded as
T =
T
0
_
g(x)
+ T
1
(x, t) +O(
2
) , (74)
u
1
= v
1
(x, t) +O(
2
) , (75)
where is a small number parametrizing the speed of the ow.
We now specialize to relativistic conformal hydrodynamics at rst order in the gradient
expansion and take the metric source to be of the form g = 1+ cos kx. The hydrodynamic
equations linearize in the small ow velocity regime, allowing us to make a separation of
variables ansatz v(x, t) = e
k
2
t
v
1
(kx) and T
1
(x, t) = e
k
2
t
T
1
(kx). The resulting pair of
ordinary dierential equations in x are then
0 = v

1
(kx) +
sin kx
2(1 + cos kx)
v
1
(kx)
2kT
1
(kx)
T
0
, (76)
0 = v

1
(kx)
3
0
2
0
k
T

1
(kx) +
3
0
sin kx
4
0
k(1 + cos kx)
T
1
(kx) (77)
+
12
0
T
0
5
0

2
4(
0
3
0
T
0
) cos kx +
0

2
cos 2kx
8
0
(1 + cos kx)
2
v
1
(kx) ,
7
Interestingly, this same solution is used as a starting point to nd the non-dissipative transport coe-
cients in a Lagrangian formulation of hydrodynamics in refs. [37, 38].
20
where we have parametrized the energy density and viscosity as =
0
T
3
and =
0
T
2
respectively.
We solve this pair of equations to leading order in k. It is pointless to keep terms
beyond leading order as we have already thrown out the second order corrections to the
hydro equations. The solution is
v
1
(y) = v
0
_
1 + cos y +O(k
2
) , (78)
T
1
(y) = k
(3
0
T
0
2
0
)y + 4
0

1
2
tan
1
__
1
1+
tan
y
2
_

1 + cos y
+O(k
3
) . (79)
In order for T
1
to be a continuous function of y, we must have that
=
2
0
3
0
T
0
_
1
_
1
2
_
. (80)
The momentum relaxation time scale is given by
1
= (k
2
). When > 1, the metric
changes signature and hence the momentum relaxation time scale becomes complex. In the
limit where 1, we recover the result (59) obtained using linear response theory.
4.2 Numerical Results
In this subsection, we study momentum relaxation numerically in the presence of the metric
source (55) using (i) gravity equations in an asymptotically AdS
4
space-time and (ii) second
order hydrodynamics equations. The intial conditions for , , and U
A
3
are obtained from
the boosted black brane metric in 3.1 with u
A
= (0.2, 0). In the following, we study the
dependence of relaxation time on the lattice parameters , m and k for a xed box size
L = 100 3/(4T
i
) where T
i
is the initial temperature. In future discussions, k, T, 1/x,
1/t and m are expressed in units where 3/(4T
i
) = 1.
1. Relaxation time scale for large lattice spacing and small
First we will show that at small k, the results obtained from gravity and hydrodynamic
simulations agree. Figs. 2a and 3a show plots of logT
tx
for dierent values of k (with
3k/4T 1 ) obtained using hydrodynamics and gravity respectively. In these examples,
we assume that = 0.2, m = 20, and the number of spatial grid points used for the simu-
lation is N = 101. Note that the time is rescaled by a factor of
2
k
2
/8T
i
, which is the
inverse relaxation time scale obtained analytically using linear response theory for hydro-
dynamics. T
i
is the initial temperature. The slope of the plot of logT
tx
is approximately
-1.0 initially. The reference line has slope -1.0 which is the linear response theory result.
21
We conclude that the results of the full non-linear hydrodynamics and gravity simulations
agree with the linear response theory.
However, the plots show deviations from the linear response theory at late times. The
small deviations in the hydrodynamic computation arise from the slow variation of temper-
ature (105). (As we review in appendix B, energy conservation relates the decrease in ow
velocity to an increase in temperature.) Now let us look at Fig. 2b. In this plot, time is
scaled by a factor of
2
k
2
/8T
0
, where T
0
is computed from the nal temperature using
(73). This rescaling allows us to see that at late times the full non-linear hydrodynamics
agrees with the result in (80) when is small.
In the gravity simulations, the nal temperature is computed using the boundary stress
tensor as follows. First we compute the energy density as the eigenvalue associated with
the time-like eigenvector of the boundary stress tensor: T

= u

, where u

is time-
like. We can then compute the temperature
8
assuming the equation of state = (4T/3)
3
.
It is clear from Fig. 3a that the deviations from hydrodynamic linear response theory
at late times in the gravity simulations are larger than in the hydrodynamic simulations.
As we did in the hydrodynamic plot, we can get better agreement at late times by rescaling
by T
0
instead of T
i
(Fig. 3b). The deviations in gravity then become greater at early times.
The deviations we are seeing come from gradient corrections to hydrodynamics. We can
infer from Fig. 1 that the relaxation time computed using linear response theory in gravity
diers from hydrodynamic linear response theory when k is not close to zero. In fact, one
can check that the relaxation time scales computed from Fig. 3a agree with Fig. 1.
These results are treated as checks on our numerical results. We present additional
checks on the numerics in appendix A. In particular, we show that in the small k limit, the
stress tensors computed from hydrodynamic and gravity simulations are in good agreement.
At late times, the system approaches the equilibrium solution described in (72). Figs.
2 and 3 show that the momentum relaxes to zero at late times. Furthermore, Fig. 4 shows
that the temperature prole at late times agrees with the expression in (72). In particular,
the dierence between the numerically computed value of temperature and the analytic
expression decreases as time evolves, and the gradients in T become smaller.
9
8
The temperature can also be computed from the surface gravity of the apparent horizon. The surface
gravity is given by = l

n where ldx

= dt and ndx

= dr. This denition assumes that the

apparent horizon is a Killing horizon. However, in non-static spacetimes the apparent horizon need not be
a Killing horizon, leading to an ambiguity in the denition.
9
It is tempting to interpret the deviations of the gravity solution from (72) in the third panel of Fig. 4
as gradient corrections. However, the deviations are comparable to our numerical error.
22
a)
0 1 2 3 4 5 6 7 8
k
2

2
t
8T
i
8
7
6
5
4
3
2
1
0
l
o
g

T
t
x
(
t
)
T
t
x
(
0
)

k =

50
k =
4
50
k =
5
50
k =
6
50
Ref.
b)
0 1 2 3 4 5 6 7 8
k
2

2
t
8T
0
8
7
6
5
4
3
2
1
0
l
o
g

T
t
x
(
t
)
T
t
x
(
t

k =

50
k =
4
50
k =
5
50
k =
6
50
Ref.
Figure 2: A plot of logT
tx
as a function of scaled time for dierent values of k, obtained
from hydrodynamic simulations. The analytical expression for relaxation time computed in
the previous section corresponds to the reference line with slope -1.0. In (b), values of t

are chosen such that the lines agree at late times.

23
a)
0 1 2 3 4 5 6
k
2

2
t
8T
i
6
5
4
3
2
1
0
l
o
g

T
t
x
(
t
)
T
t
x
(
0
)

k =

50
k =
4
50
k =
5
50
k =
6
50
Ref.
b)
0 1 2 3 4 5 6
k
2

2
t
8T
0
6
5
4
3
2
1
0
l
o
g

T
t
x
(
t
)
T
t
x
(
t

k =

50
k =
4
50
k =
5
50
k =
6
50
Ref.
Figure 3: A plot of logT
tx
as a function of scaled time for dierent values of k, obtained
from gravity simulations. The analytical expression for relaxation time computed in the
previous section corresponds to the reference line with slope -1.0. In (b), values of t

are
chosen such that the lines agree at late times. The simulations were run for 25 10
4
time
steps with t = 0.05.
24
40 20 0 20 40
0.6
0.4
0.2
0.0
0.2
0.4
0.6
0.8

T
10
3
t = 2000
40 20 0 20 40
0.6
0.4
0.2
0.0
0.2
0.4
0.6
0.8

T
10
3
t = 4000
40 20 0 20 40
x
0.6
0.4
0.2
0.0
0.2
0.4
0.6
0.8

T
10
3
t = 10000
Figure 4: A plot of the dierence in the numerically computed temperature and the
exact analytical expression for temperature in (72) at t = 2000, 4000 and 10,000. The
lattice wavenumber is k = 4/50. The red curve corresponds to the result obtained from
hydrodynamic simulations and the black curve corresponds to gravity simulations. We use
the nal value of mean temperature computed from gravity and hydrodynamic simulations
to compute T
0
.
2. Relaxation time scale for small lattice spacing and small
In the large k regime, the results from hydrodynamic simulations do not match the results
from gravity. As expected, the gradient expansion breaks down when the wave number
k T. Fig. 5 shows that the relaxation time scales computed from hydrodynamics and
gravity are dierent when k = 20/50. The rest of the parameters are the same as those in
the previous subsection.
While gravity does not agree with hydrodynamics in this limit, the relaxation time scales
computed using the full gravity simulation and its linearized counterpart do agree. In Fig.
5, the solid black and dashed blue lines have very similar slope. The solid black line was
computed from the full gravity simulation, while the dashed blue line was computed from
25
the low frequency limit of ImG

R
, described in Section 4.1.2. Equivalently, in Fig. 1, the
star lies very nearly on the black curve. The star was computed from the value of the slope
of the solid black line in Fig. 5, while the black curve in Fig. 1 was computed from the
Greens function. This agreement is a non-trivial check of the gravity code.
0 1 2 3 4 5
k
2

2
t
8T
i
5
4
3
2
1
0
l
o
g

T
t
x
(
t
)
T
t
x
(
0
)

Gravity
Hydro
Ref GR
Ref Hydro
Figure 5: A plot of logT
tx
as a function of scaled time when k = 20/50, obtained from
hydrodynamic and gravity simulations. In the gravity simulations we choose the size of
the time step t = 0.002. The slopes of the reference lines (dashed and dotted lines) are
computed using linear response theory.
3. Dependence of relaxation time scale on
Here, we consider what happens in our simulations when leaves the linear response regime.
In (80), we showed that at low wave numbers, the momentum relaxation time scale is given
by
1

=
2
0
k
2
3
0
T
0
_
1
_
1
2
_
=

0
k
2
f()
3
0
T
0
.
This expression agrees with the linear response theory computation for small . To the
extent to which f() ,=

2
2
, Fig. 6 shows deviation from linear response theory for larger . In
particular, the results from numerical hydrodynamics and gravity show that the relaxation
time scale agrees with the low-velocity approximation result obtained analytically in (80),
within numerical error.
26
a)
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
k
2
f()t
8T
0
1.5
1.0
0.5
0.0
l
o
g

T
t
x
(
t
)
T
t
x
(
t

= 0.2
= 0.3
= 0.4
= 0.5
Ref.
b)
0.0 0.2 0.4 0.6 0.8 1.0
k
2
f()t
8T
0
1.0
0.8
0.6
0.4
0.2
0.0
0.2
l
o
g

T
t
x
(
t
)
T
t
x
(
t

= 0.2
= 0.3
= 0.4
Ref.
Figure 6: A plot of logT
tx
as a function of scaled time for dierent values of , obtained
from a) hydrodynamic simulations and b) gravity. We have chosen k = /50. The sim-
ulations were run for 25 10
4
time steps with t = 10
2
. When = 0.4, the gravity
simulations become less stable with m = 20. As we were only interested in the late time
behavior, we were able to improve stability by choosing m = 100 for the = 0.4 case.
27
The regime where and k are both large is intractable analytically. A hydrodynamic
description is not useful here as the gradient expansion (or continuum approximation) is
not valid. Moreover, nonlinearities in are presumably important. In Fig. 7, we show
numerical results for = 0.4 and = 0.3 (velocity is 0.2). We choose m = 1000 to
improve numerical stability. (There may be an issue with the development of caustics in
the spacetime for smaller values of m and larger values of .) Fig. 7 shows a plot of logT
tx

versus time. This plot shows the power of our techniques to get at an otherwise inaccessible
regime. Interestingly, momentum relaxation does not appear to be exponential in time for
the parameter regime explored.
Moving forward, it should be possible to obtain an analog of the low-velocity approxi-
mation (80) in gravity; however, the steady state solution (to which the gravity equations
relax) is not known analytically when k is large. One might be able to get further insight
into this regime by numerically constructing the steady state regime and looking at small
uctuations about it. Note that for large k, we cannot use the hydrodynamic denitions
of s and T. Perhaps one could use the properties of the apparent horizon to dene these
quantities.
0 20 40 60 80 100
k
2
f()t
8T
0
10
8
6
4
2
0
l
o
g

T
t
x
(
t
)
T
t
x
(
0
)

= 0.2
= 0.3
= 0.4
Ref.
Figure 7: A plot of logT
tx
as a function of scaled time for dierent values of , obtained
from gravity simulations. We have chosen k = 20/50. To improve stability, we chose
t = 0.005 for = 0.4, t = 0.01 for the other two cases and m = 1000 for all cases. Note
f(0.4)k
2
m/8T
0
44. The reference curve is the prediction from hydrodynamics.
28
5 Discussion and Outlook
We have presented a numerical technique to study lattice-induced momentum relaxation in
theories with gravity duals. We have used the characteristic formulation to solve Einsteins
equations in the presence of a negative cosmological constant. We infer that the charac-
teristic evolution scheme can be used reliably, without encountering caustics, for studying
gravitational dynamics in AdS space even when metric sources are turned on. However, the
numerical method becomes less stable when the strength of the source (lattice strength)
increases. While we have focused on momentum in this paper, one could also consider heat
transport with our formalism.
10
We performed a number of cross checks of our code. The gravity and hydrodynamic
simulations agree with each other when the lattice wave number k is small, i.e. in the
hydrodynamic regime. The relaxation time extracted from these simulations agrees with
an exact analytical expression we obtained, valid for all and small k. As is arbitrary,
this agreement is not limited to the regime of linear response. The simulations approach a
steady state equilibrium temperature distribution predicted by hydrodynamics. Next, for
large k and small , where hydrodynamics is not valid but linear response is, the momentum
relaxation time scale extracted from our gravity simulations agrees with the results obtained
from linearized gravity. Given our cross checks, we can hopefully trust our numerical results
in the regime of large k and large , where neither linear response nor hydrodynamics is
valid. Fig. 7 presents some simulations of momentum relaxation in this scaling regime.
In hydrodynamics, it is straightforward to obtain analytically the solution to which
the ow relaxes at late times. The answer is (72). On the gravity side, in principle it is
straightforward to obtain this solution numerically by letting our simulations run for a long
time. While for small k this approach is feasible, we nd that for large k, because of the
exponential damping of the relaxation rate, we have to wait a prohibitively long time for
the simulation to nish. An interesting project for the future would be to obtain the large
k, steady-state solution directly using a dierent numerical approach. This large k limit is
also the small temperature limit because of the relatively few scales in our problem. The low
temperature limit might conceivably be somewhat more interesting for condensed matter
applications. Of course, it would also be interesting to couple gravity to a complex scalar
and gauge eld. One could then study more elaborate systems involving nonzero charge
density and superuid phases [40, 41].
10
See ref. [39] for related work on this front.
29
In the immediate future, we would like to use the numerical framework presented here
to study more intricate ows involving shock waves, turbulent eddies, etc. In particular, it
would be interesting to understand if the gradient expansion breaks down in such ows.
Acknowledgments
We would like to thank P. Chesler and L Yae for discussion and for collaboration during
the early stages of this project. We thank K. Jensen for comments on the manuscript. We
also thank the Whiteley Center, where this project was started, for hospitality. This work
was supported in part by the National Science Foundation under Grants No. PHY-0844827
and PHY-1316617. C. H. also thanks the Sloan Foundation for partial support.
A Numerical Details and Cross Checks
1. Overview of the numerical scheme
In this appendix, we provide details of the numerical simulations. We begin by describing
briey pseudo spectral (collocation) methods.
11
We begin by describing briey pseudo
spectral (collocation) methods. Detailed descriptions of spectral methods can be found in
[24, 25]. The basic idea of a spectral method is to expand the variables or elds in terms
of basis functions
n
(x), for example Fourier series or Chebyshev polynomials, that satisfy
some orthogonality relation and are appropriate for imposing boundary conditions. In a
pseudo spectral method, a eld u(x) and its gradients are evaluated at discrete collocation
points x x
0
, x
1
, . . . , x
N1
. The number of collocation points x
j
is the same as the
number of basis functions
k
:
u
N
(x
j
, t)
N1

n=0

n
(x
j
)a
k
(t),
m
x
u
N
(x, t)
N1

n=0
d
m

n
(x)
dx
m
a
k
(t), (81)
Using orthogonality among the
n
, one can then express gradients of the elds in terms of
linear combinations of the u
N
(x
j
, t):

x
u
N
(x
i
, t) =
N1

j=0
D
(1)
ij
u
N
(x
j
, t),
2
x
u
N
(x
i
, t) =
N1

j=0
D
(2)
ij
u
N
(x
j
, t), . . .
where D
(1)
N
, D
(2)
N
, . . . are N N derivative matrices.
11
We would like to thank P. Chesler and L. Yae for convincing us to take a spectral method approach
to this problem.
30
In this paper, our PDEs are all schematically of the form
t
u = /(u,
x
u,
2
x
u, . . . ).
To a rst approximation, we solve such an equation by rst replacing elds and spatial
derivative operators with their discrete versions u
N
, D
(1)
N
, D
(2)
N
, etc., and then integrating
in time using a Runge-Kutta or Adams-Bashforth scheme. The relativistic hydrodynamic
equations take precisely this form, treating u as a ve dimensional vector with components
(T, u
x
, u
y
, B,
xy
). We work in a box with periodic boundary conditions, and thus we use
Fourier basis functions
n
(x) = e
inx
to compute the derivatives appearing on the RHS
of the hydrodynamic equations. We use the third order Adams-Bashforth technique to
integrate in time. The starting values at t = 0, t, 2t are computed using Runge-Kutta
integration. The error in this method of integration is O
_
t
3
_
. As the equations already
contain viscous terms, we do not need to add articial viscosity or use ltering to stabilize
the code.
We will now proceed to describe the details of the gravity simulations. We use a coor-
dinate system such that the apparent horizon is at z = 1 and the boundary is at z = 0. We
discretize the holographic direction using a Chebyshev grid and the boundary spatial direc-
tions using a Fourier grid. As described in the text, the characteristic method allows us to
write Einsteins equation in a form that has a nested structure. The discretized Einsteins
equations take the following form;
_
D
z
+
4
z
I
_

s
= S

(z,
s
,
s
,
s
) (82)
(D
z
)
A
s
= S

A (z,
s
,
s
,
s
,
s
) (83)
_
D
z
+
2
z
I
_
U
A
s
= S
U
A
_
z,
s
,
s
,
s
,
s
,
A
s
_
(84)
(D
z
) d
t

s
= S
dt
_
z,
s
,
s
,
s
,
s
, U
A
s
_
(85)
_
D
z
+
1
z
I
_
d
t

s
+C

d
t

s
+C

d
t

s
= S
dt
(. . . , d
t

s
) (86)
_
D
z
+
1
z
I
_
d
t

s
+C

d
t

s
+C

d
t

s
= S
dt
(. . . , d
t

s
) (87)
C
H
xx
D
(2)
x
V
H
+C
H
x
D
(1)
x
V
H
+C
H
0
V
H
= S
V
H
_

H
,
H
,
H
, U
A
H
,
H
, d
t

H
, d
t

H
, d
t

H
_
(88)

s
=
1
z
(d
t
)
s
+
1
2
_
z

s
+ 2
s
_
z
3
V (89)

s
=
1
z
(d
t
)
s
+
1
2
_
z

s
+ 2
s
_
z
3
V (90)

t
= S

_
V
H
, U
A
H
,
H
_
(91)

t
U
A
3
= S
U
A
3
_

3
,
3
,
3
, V
3
, U
A
3
,
0
_
(92)

t
V
3
= S
V
3
_

3
,
3
,
3
, V
3
, U
A
3
,
0
_
(93)
31
Note that time derivatives of X
i
do not appear in the source terms S(X
1
, . . . , X
n
) but
spatial gradients generically will.
As described in section 2, eqs. (82)-(88) are hypersurface equations [20], i.e. we solve
these equations at every time slice and propagate this information to the next time using
the bulk evolution equations (89), (90) and the boundary evolution equations (91) and (92).
The time constraint equations equations (82)-(88) must be satised at t = 0 in order to
have well dened Cauchy data. The Cauchy data consist of initial conditions for (x, z, t =
0), (x, z, t = 0), U
A
3
(x, t = 0) and
3
(x, t = 0). Numerically, the boundary conditions
are applied by replacing the rst row of the matrices appearing on the LHS of the time
constraint equations (82)-(87) and the rst entry of the source terms. Using Fourier series,
periodic boundary conditions for the elliptic eq. (88) are enforced automatically. Moreover,
generically (88) will have no zero modes. To integrate the time evolution equations, we have
used both third order Adams-Bashforth and fourth order Runge-Kutta. As a consequence
of the Bianchi identities (see section 2), the last equation (93) should be satised if all
other equations are satised. We use this Bianchi constraint for monitoring the error, as
we discuss in greater detail later in this section.
Note that the operators appearing on the LHS of equations (82)-(85) are independent
of x and t. These equations can be solved by multiplying the source terms by the inverse of
these operators. It is sucient to compute the inverse of these operators only once, leading
to an improvement in the speed of the code. As mentioned earlier, we choose Neumann
boundary condition for
s
and U
A
s
to minimize the condition numbers of the matrices
appearing on the LHS of (82) and (84). The round-o error arising from the inversion of
operators with lower condition number is less.
There are several unpleasant details involved in implementing this scheme that remain
to be discussed.
2. Factors aecting stability and accuracy of the code
There are many factors that aect the reliability of numerical methods. We discuss some of
the issues that we did and did not encounter and the strategies we employed to make the
code work.
Round-o
We encountered round-o error in trying to evaluate the source terms in the PDEs for
our gravity code. These source terms are large rational and polynomial expressions.
32
One technique that helped reduce error was to place the polynomials in Horner form.
12
Another strategy was judicious use of expm1 and log1p functions that avoid round-o
error in evaluating e
x
1 and log(1+x) respectively when x is close to zero. A typical
situation was a need to evaluate accurately an expression like (e
z
1)/z close to the
conformal boundary z = 0.
Truncation error
Truncation error, or the error associated with the discretization, is typically less of an
issue for spectral methods then for nite volume methods. We found in our simulations
that the truncation error in the z-direction became comparable to machine epsilon
with fewer than 20 collocation points. In the x-direction, we needed in contrast on
the order of 100 points to achieve accuracy at the part per million level.
There is also a truncation error arising from time integration. We found the third order
Adams-Bashforth technique sucient for our computation, resulting in an O
_
t
3
_
error in the constraints. The error can be ameliorated by using a fourth order Runge-
Kutta or Adams-Bashforth method.
Aliasing error
While viscous terms eliminate the need for ltering the hydrodynamic code, aliasing
is a noticeable problem for the gravity code. The standard solution to this problem,
which we have implemented, is a low pass lter. The ltering leads to signicant
improvement in the behavior of constraints and stability of the code. We implement
ltering in real space using matrix multiplication.
For the Fourier grid, we employ a 2/3-rule. The elds take the schematic form
u(x) =

j
h(x, x
j
)u(x
j
)
where,
h(x, x
j
) =
1
N
N/2

n=N/2
1
c
n
e
in(xx
j
)
, with c
n
= 1 +
|n|,N/2
In order to get rid of the higher modes we use a low-pass lter by computing u(x) on
a coarser grid with 1/3 of the modes and 1/3 of the collocation points eliminated:
u( x
i
) =

j
1
N
N/2

n=N/2
1
c
n

_
2[n[
N
_
e
in( x
i
x
j
)
u(x
j
) =

j
T
N
ij
u(x
j
)
12
In Mathematica, the relevant command is HornerForm.
33
where ([n[/N) = 0 for [n[ > N/3 and one otherwise. T
N
is the ltering-matrix.
We then interpolate back to a ner grid using the interpolating matrix J
N
which is
obtained by setting (x) = 1. The de-aliasing matrix is product of J
N
and T
N
.
Our radial ltering for the Chebyshev grid just involves interpolating to a coarser grid
and then back to the original grid, that is, we choose the lter function to be 1 for
Chebyshev methods.
Formation of caustics
As discussed in section 2, the utility of the null-characteristic formulation will break
down in the presence of caustics, and our numerical scheme will fail. Fortunately we
did not encounter caustics for a wide range of parameters. The gravity code is less
stable for very large 0.5 suggesting a possible formation of caustics in this regime.
3. Some checks on the numerical method
It is important that the constraints arising from Bianchi identities (discussed earlier) remain
close to zero. We check the Bianchi constraint by computing V
3
in two dierent ways: (i)
using equations (85), (88) and (91),
13
and (ii) using the boundary equation (93). These
two methods of evaluating V
3
should give the same result up to numerical error. In all
our numerical simulations the constraints remained around 10
6
or 10
7
in the presence of
radial ltering. When radial ltering was turned o the constraints became as large as 10
4
but eventually decreased to 10
6
. Fig. 8 shows that the constraints remain within 10
6
for
some representative cases.
Another non-trivial check of our method is to compare the hydrodynamic stress-tensor
with the stress-tensor obtained using gravity for low values of k (hydrodynamic regime).
Fig. 9 show that the dierence in T
tx
, T
tt
and T
xx
when k = 4/50 is around 10
3
O(k
5
).
4. Computational Platform
Two separate software packages were developed to run the simulations described in this
paper. The rst was developed using Matlab [42] and used the code described in ref. [14] as
a starting point.
14
The second was developed from scratch using Python. The simulations
were run on standard desktop and laptop computers. The longest gravity simulation (1
million time steps) took about 6 hours to complete. The Python code ultimately ran about
13
Recall that =
1
4
log

1 + 2z
3
3

.
14
We thank Paul Chesler for making his code available to us.
34
0 1 2 3 4 5
k
2

2
t
8T
0
10
11
10
10
10
9
10
8
10
7
10
6
10
5

V
3
k =
4
50
, = 0.2
k =
6
50
, = 0.2
k =
20
50
, = 0.4
Figure 8: A plot of V
3
as a function of scaled time for dierent values of k and .
0.006
0.004
0.002
0.000
0.002

T
t
x
0.003
0.002
0.001
0.000
0.001

T
t
t
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
k
2

2
t
8T
i
0.003
0.002
0.001
0.000

T
x
x
Figure 9: Plots of T

obtained from hydrodynamic and gravity simulations with k =

4/50 and = 0.2.
35
ten times faster than the Matlab code although at some point (largely for scal reasons) we
stopped trying to optimize the Matlab code.
B Small momentum relaxation from hydrodynamics
In this section, we present a method using hydrodynamics to compute the relaxation time
scale when the metric perturbation is small but the velocity is arbitrary. This method
does not rely on the memory function formalism, but the results agree with (59) when
the velocity is small. We look for perturbative solutions of the relativistic hydrodynamic
equations with the structure
g
tt
= (1 + 2g
1
cos(kx)) , (94)
T = T
0
(t) + (T
1
e
ikx
+T

1
e
ikx
) + (T
2
e
2ikx
+T

2
e
2ikx
)
2
+O(
3
) , (95)
u
1
= v(t) + (v
1
e
ikx
+v

1
e
ikx
) + (v
2
e
2ikx
+v

2
e
2ikx
)
2
+O(
3
) . (96)
We will allow T
0
(t) and v(t) to be weakly time dependent, i.e.

T
0
v
2
. Let us dene

T
0
=
2

T
0
and

v =
2
v. We rst solve for T
1
and v
1
:
T
1
=
c
2
s
g
1
sT
2
2(c
2
s
sT(1 +v
2
) +iv(isTv +k(1 +v
2
)( +)))
, (97)
v
1
=
g
1
v(1 +v
2
)((c
2
s
1)sT +ikv( +))
2(c
2
s
sT(1 +v
2
) +iv(isTv +k(1 +v
2
)( +)))
. (98)
We have made use of the thermodynamic relation +p = sT where s is the entropy density
and introduced the speed of sound c
2
s
= p/.
We then calculate how well the stress tensor conservation condition is met at second
order in our expansion. The x-dependent pieces will be satised by adjusting T
2
and v
2
accordingly. The x-independent piece will allow us to solve for the time dependence of T
0
(t)
and v(t). Let us rst consider the energy conservation condition. The x-independent piece
at second order (equivalently the spatially averaged piece) is

T
0
=
_
2sTv

v +
s
c
2
s
(1 + (1 +c
2
s
)v
2
)

T
_

2
+. . . (99)
=
d
dt
_
sTv
2
+

. (100)
This equation can be trivially integrated and expresses energy conservation
sTv
2
+ = constant . (101)
36
For momentum conservation, we nd (using the relation for

T from energy conservation)

T
1
=
d
dt
sTv
_
1 +v
2
+ (102)
+
g
2
1
k
2
s
2
T
2
( +)v
2[c
2
s
sT(1 +v
2
) +iv(isTv +k(1 +v
2
)( +))[
2

2
+. . . ,
= sT
_
_
2(1 +v
2
)
1 + (1 +c
2
s
)v
2
1
_

v

1 +v
2
+ (103)
+
g
2
1
k
2
sT( +)v
2[c
2
s
sT(1 +v
2
) +iv(isTv +k(1 +v
2
)( +))[
2
_

2
+. . . .
At leading order in velocity, momentum conservation reduces to

T
1
= sT v +
( +)v
2
_
g
1
k
c
2
s
_
2
+O(v
3
,
3
) . (104)
From equations (101) and (102), it is clear that v decreases slowly over time while T, by
energy conservation, must increase.
In the conformal case, where c
2
s
= 1/2 and = 0, we calculate the form of the decrease
in velocity and increase in temperature. We parametrize the energy density and viscosity
by =
0
T
3
and =
0
T
2
. Thus sT = 3
0
T
3
/2. Assuming rst that T is roughly constant,
we nd that
v v
0
exp
_
4
0
g
2
1
k
2

2
3
0
T
t
_
.
Then from energy conservation it follows that
T
3
T
3
0
v
2
0
+ 2/3
v
2
+ 2/3
, (105)
where T
0
is the initial temperature.
References
[1] J. M. Maldacena, The Large N limit of superconformal eld theories and supergrav-
ity, Adv. Theor. Math. Phys. 2, 231 (1998) [hep-th/9711200].
[2] S. S. Gubser, I. R. Klebanov and A. M. Polyakov, Gauge theory correlators from
noncritical string theory, Phys. Lett. B 428, 105 (1998) [hep-th/9802109].
[3] E. Witten, Anti-de Sitter space and holography, Adv. Theor. Math. Phys. 2, 253
(1998) [hep-th/9802150].
37
[4] S. S. Gubser and A. Karch, From gauge-string duality to strong interactions: A
Pedestrians Guide, Ann. Rev. Nucl. Part. Sci. 59, 145 (2009) [arXiv:0901.0935 [hep-
th]].
[5] J. Casalderrey-Solana, H. Liu, D. Mateos, K. Rajagopal and U. A. Wiedemann,
Gauge/String Duality, Hot QCD and Heavy Ion Collisions, arXiv:1101.0618 [hep-
th].
[6] C. P. Herzog, Lectures on Holographic Superuidity and Superconductivity, J. Phys.
A 42, 343001 (2009) [arXiv:0904.1975 [hep-th]].
[7] S. A. Hartnoll, Lectures on holographic methods for condensed matter physics, Class.
Quant. Grav. 26, 224002 (2009) [arXiv:0903.3246 [hep-th]].
[8] P. M. Chesler and L. G. Yae, Boost invariant ow, black hole formation, and far-
from-equilibrium dynamics in N = 4 supersymmetric Yang-Mills theory, Phys. Rev.
D 82, 026006 (2010) [arXiv:0906.4426 [hep-th]].
[9] P. M. Chesler and L. G. Yae, Holography and colliding gravitational shock waves in
asymptotically AdS
5
spacetime, Phys. Rev. Lett. 106, 021601 (2011) [arXiv:1011.3562
[hep-th]].
[10] A. Adams, P. M. Chesler and H. Liu, Holographic turbulence, arXiv:1307.7267 [hep-
th].
[11] F. Carrasco, L. Lehner, R. C. Myers, O. Reula and A. Singh, Turbulent ows for
relativistic conformal uids in 2+1 dimensions, Phys. Rev. D 86, 126006 (2012)
[arXiv:1210.6702 [hep-th]].
[12] S. R. Green, F. Carrasco and L. Lehner, A Holographic Path to the Turbulent Side
of Gravity, arXiv:1309.7940 [hep-th].
[13] S. S. Gubser, S. S. Pufu and A. Yarom, Sonic booms and diusion wakes gener-
ated by a heavy quark in thermal AdS/CFT, Phys. Rev. Lett. 100, 012301 (2008)
[arXiv:0706.4307 [hep-th]].
[14] P. M. Chesler and L. G. Yae, The Wake of a quark moving through a strongly-coupled
plasma, Phys. Rev. Lett. 99, 152001 (2007) [arXiv:0706.0368 [hep-th]].
38
[15] P. Figueras and T. Wiseman, Stationary holographic plasma quenches and nu-
merical methods for non-Killing horizons, Phys. Rev. Lett. 110, 171602 (2013)
[arXiv:1212.4498 [hep-th]].
[16] D. Forster, Hydrodynamic Fluctuations, Broken Symmetry, and Correlation Func-
tions, Westview Press, 1995.
[17] S. A. Hartnoll and C. P. Herzog, Impure AdS/CFT correspondence, Phys. Rev. D
77, 106009 (2008) [arXiv:0801.1693 [hep-th]].
[18] S. A. Hartnoll and D. M. Hofman, Locally Critical Resistivities from Umklapp Scat-
tering, Phys. Rev. Lett. 108, 241601 (2012) [arXiv:1201.3917 [hep-th]].
[19] S. A. Hartnoll, P. K. Kovtun, M. Muller and S. Sachdev, Theory of the Nernst eect
near quantum phase transitions in condensed matter, and in dyonic black holes, Phys.
Rev. B 76, 144502 (2007) [arXiv:0706.3215 [cond-mat.str-el]].
[20] J. Winicour, Characteristic Evolution and Matching, Living Rev. Relativity, 15
(2012) 2.
[21] P. M. Chesler and L. G. Yae, Numerical solution of gravitational dynamics in asymp-
totically anti-de Sitter spacetimes, arXiv:1309.1439 [hep-th].
[22] Z. Cao and X. He, Generalized Bondi-Sachs equations for characteristic formalism of
numerical relativity, Phys. Rev. D 88, 104002 (2013).
[23] P. Grandclement and Jerome Novak, Spectral Methods for Numerical Relativity,
Living Rev. Relativity, 12 (2009) 1.
[24] John P. Boyd, Chebyshev and Fourier Spectral Methods.
[25] Lloyd N. Trefethen, Spectral Methods in MATLAB, SIAM, 2000.
[26] J. M. Stewart, & H. Friedrich 1982, Royal Society of London Proceedings Series A,
384, 427
[27] V. Balasubramanian and P. Kraus, A Stress tensor for Anti-de Sitter gravity, Com-
mun. Math. Phys. 208, 413 (1999) [hep-th/9902121].
[28] R. Baier, P. Romatschke, D. T. Son, A. O. Starinets and M. A. Stephanov, Relativis-
tic viscous hydrodynamics, conformal invariance, and holography, JHEP 0804, 100
(2008) [arXiv:0712.2451 [hep-th]].
39
[29] C. P. Herzog, The Hydrodynamics of M theory, JHEP 0212, 026 (2002) [hep-
th/0210126].
[30] M. Natsuume and T. Okamura, Causal hydrodynamics of gauge theory plasmas from
AdS/CFT duality, Phys. Rev. D 77, 066014 (2008) [Erratum-ibid. D 78, 089902
(2008)] [arXiv:0712.2916 [hep-th]].
[31] M. Van Raamsdonk, Black Hole Dynamics From Atmospheric Science, JHEP 0805,
106 (2008) [arXiv:0802.3224 [hep-th]].
[32] M. Luzum and P. Romatschke, Conformal Relativistic Viscous Hydrodynamics: Ap-
plications to RHIC results at s(NN)**(1/2) = 200-GeV, Phys. Rev. C 78, 034915
(2008) [Erratum-ibid. C 79, 039903 (2009)] [arXiv:0804.4015 [nucl-th]].
[33] S. Bhattacharyya, V. Hubeny, S. Minwalla and M. Rangamani, Nonlinear Fluid Dy-
namics from Gravity, JHEP 0802, 045 (2008) [arXiv:0712.2456 [hep-th]].
[34] P. Kovtun, Lectures on hydrodynamic uctuations in relativistic theories, J. Phys.
A 45, 473001 (2012) [arXiv:1205.5040 [hep-th]].
[35] P. K. Kovtun and A. O. Starinets, Quasinormal modes and holography, Phys. Rev.
D 72, 086009 (2005) [hep-th/0506184].
[36] D. T. Son and A. O. Starinets, Minkowski space correlators in AdS / CFT correspon-
dence: Recipe and applications, JHEP 0209, 042 (2002) [hep-th/0205051].
[37] N. Banerjee, J. Bhattacharya, S. Bhattacharyya, S. Jain, S. Minwalla and T. Sharma,
Constraints on Fluid Dynamics from Equilibrium Partition Functions, JHEP 1209,
046 (2012) [arXiv:1203.3544 [hep-th]].
[38] K. Jensen, M. Kaminski, P. Kovtun, R. Meyer, A. Ritz and A. Yarom, Towards
hydrodynamics without an entropy current, Phys. Rev. Lett. 109, 101601 (2012)
[arXiv:1203.3556 [hep-th]].
[39] S. Fischetti, D. Marolf and J. E. Santos, AdS owing black funnels: Stationary AdS
black holes with non-Killing horizons and heat transport in the dual CFT, Class.
Quant. Grav. 30, 075001 (2013) [arXiv:1212.4820 [hep-th]].
[40] S. S. Gubser, Breaking an Abelian gauge symmetry near a black hole horizon, Phys.
Rev. D 78, 065034 (2008) [arXiv:0801.2977 [hep-th]].
40
[41] S. A. Hartnoll, C. P. Herzog and G. T. Horowitz, Building a Holographic Supercon-
ductor, Phys. Rev. Lett. 101, 031601 (2008) [arXiv:0803.3295 [hep-th]].
[42] MATLAB R2013a. Natick, Massachusetts: The MathWorks Inc., 2013.
41