Preprint typeset in JHEP style - PAPER VERSION TIFR/TH/08-40

The Incompressible Non-Relativistic Navier-Stokes

Equation from Gravity
arXiv:0810.1545v3 [hep-th] 20 Jul 2009

Sayantani Bhattacharyyaa∗, Shiraz Minwalla a†

and Spenta R. Wadiaa‡
a
Department of Theoretical Physics,Tata Institute of Fundamental Research,
Homi Bhabha Rd, Mumbai 400005, India

Abstract: We note that the equations of relativistic hydrodynamics reduce to the incompressible
Navier-Stokes equations in a particular scaling limit. In this limit boundary metric fluctuations of
the underlying relativistic system turn into a forcing function identical to the action of a background
electromagnetic field on the effectively charged fluid. We demonstrate that special conformal symme-
tries of the parent relativistic theory descend to ‘accelerated boost’ symmetries of the Navier-Stokes
equations, uncovering a conformal symmetry structure of these equations. Applying our scaling limit
to holographically induced fluid dynamics, we find gravity dual descriptions of an arbitrary solution
of the forced non-relativistic incompressible Navier-Stokes equations. In the holographic context we
also find a simple forced steady state shear solution to the Navier-Stokes equations, and demonstrate
that this solution turns unstable at high enough Reynolds numbers, indicating a possible eventual
transition to turbulence.

Keywords: .

∗ [email protected]
† [email protected]
‡ [email protected]
Contents
1. Introduction and Discussion 1

2. Scaling the Navier-Stokes Equation 5

2.1 The Navier-Stoke Equation as a universal limit of fluid dynamics 5
2.2 A scaling limit 6
2.3 Charged Fluids 8
2.4 Reduction of Cauchy Data 8

3. Symmetries 9
3.1 Dilatations 9
3.2 Special Conformal Transformations 9
3.3 The Full Symmetry Group 10

4. The gravity solution dual to fluid dynamics in the scaling regime 12

4.1 The dual bulk metric 12
4.2 Comments on the bulk metric 13

5. Simple Steady State Shear flows at arbitrary Reynolds Number 14

A. Stability Analysis 15
A.1 Setting up the Eigenvalue Equations 16
A.2 Qualitative nature of the solution 17
A.3 Numerical Evaluation of the frequency 18
A.4 Perturbation theory at small Reynolds numbers 20

B. A simple flow on a sphere 20

1. Introduction and Discussion

The flow of non-relativistic fluids is described by the Navier-Stokes equations

~v˙ + ~v .∇~v = −∇P

~ + ν∇2~v + f~
(1.1)
~ v=0
∇.~

where ~v is the fluid velocity, P the fluid pressure, ν the shear viscosity and f~ an externally specified
forcing function. Although these equations describe a wide variety of natural phenomena (see e.g. [1]
) and have been intensively studied for almost two centuries, their extremely rich phenomenology
remains very poorly understood. In particular, most fluid flows go turbulent at high Reynolds number,
i.e. in the regime in which the viscous fluid term is negligible compared to the nonlinear convective term
in (1.1). Although turbulent flows appear complicated and statistical in nature, it has been suggested

–1–
(see e.g. [2] ) that these flows are in fact governed by a new and simple universal mathematical structure
analogous to a fixed point of the renormalization group flow equations. A completely new angle on
fluid dynamics could well be needed in order to uncover such a structure.

Recent investigations [3–13], within the framework of the AdS/CFT correspondence of string
theory [14] have revealed an initially surprising relationship between the vacuum equations of Einstein
gravity in an asymptotically locally AdSd+1 space and the equations of hydrodynamics in d dimensions.
More concretely these papers study a class of regular, long wavelength locally asymptotically AdSd+1
solutions to the vacuum Einstein equations with a negative cosmological constant. These solutions
are shown to be in one to one correspondence with solutions of the d dimensional hydrodynamical
equations ∇µ T µν = 0. In the last equation the stress tensor T µν is a holographically determined
functional of a d dimensional fluid velocity uµ and temperature T . In the long distance limit under
consideration it is appropriate to expand the stress tensor in a power series in the boundary derivatives
of the velocity and temperature fields. Schematically

∞
X
T µν = T d−n Tnµν (1.2)
n=0

where T is the local fluid temperature and Tnµν is a local function of the fluid velocity and temperature
of nth order in spacetime derivatives. The expressions for Tnµν for n ≤ 2 have been explicitly deter-
mined in the references cited above (see [13] for the most general result) and constitute a relativistic
generalization of the incompressible Navier-Stokes stress tensor. In summary, classical asymptotically
AdSd+1 gravity is ‘dual’ to relativistic generalizations of the Navier-Stokes equations at long distance
and time scales.

Many theoretical and experimental investigations of fluid dynamics study the actual incompressible
Navier-Stokes equations (1.1). It is consequently of interest to find a dual description of the Navier-
Stokes equations (1.1) themselves rather than their their relativistic generalizations. This may be
simply achieved by taking the appropriate limit of the results of [4–13], and this limit is the topic of
our note.

In order to find the gravitational description dual to (1.1), we adopt a two step procedure. First,
purely at the level of fluid dynamics we make a straightforward and possibly well known observation.
We note that the non-relativistic incompressible Navier-Stokes equations (1.1) are the precise and
universal outcome of a particular combined scaling limit (one in which we scale to long distances, long
times, low speeds and low amplitudes in a coordinated fashion) applied to any reasonable relativistic
equations of hydrodynamics, i.e. the hydrodynamics a relativistic fluid with any reasonable equation
of state1 .The equations of fluid dynamics become non-relativistic at low speeds for the usual reason,
and also become effectively incompressible, as we go to velocities much lower than the speed of sound
(see for instance [1]).

To be more specific we show that the equations of hydrodynamics reduce, in a precise fashion, to

1 For instance the fluid that is dual to gravity studied below whose equation of state is dictated by conformal invariance.

–2–
the incompressible non-relativistic Navier-Stokes equations (1.1) under the limit
1
δx ∼
Tǫ
1
δt ∼
T ǫ2
(1.3)
vi ∼ ǫ
δP ∼ T d ǫ2
ǫ→0

where δx is the spatial length scale, δt the temporal scale while v i and δP represent estimates of
the magnitude of velocity and pressure fluctuations about an ambient configuration of equilibrium
fluid at rest. The rough contours of this choice of scaling are quite intuitive. It is clear we have to
scale to long distances to be in the fluid dynamical regime. The dispersion relation for shear waves,
ω = iνk 2 suggests that time intervals should scale like spatial intervals squared2 . Scalings of distances
and time intervals determine the scaling law for velocities. Finally the pressure variations are scaled
appropriately to ensure that they cannot accelerate the fluid into velocities outside this scaling limit.
As we have explained, the scaling (1.3) of the equations of relativistic fluid mechanics leads to
the Navier-Stokes equations (1.1). It follows that this scaling operation is a symmetry of the same
equations. This is easily directly verified. In particular, if the fields v i (x, t) and p(x, t) obey the
unforced Navier-Stokes equations, then the rescaled fields ǫv i (ǫx, ǫ2 t) and ǫ2 p(ǫx, ǫ2 t) also obey the
same equations. Consequently, the scaling operation described above is a symmetry of the unforced
Navier-Stokes equations.
As we have described above, the Navier-Stokes equations may be obtained as the scaling limit
of any reasonable relativistic equations of fluid dynamics. In describing the connection with gravity,
in most of the rest of this note, we will take the parent fluid dynamical theory to be conformal. It
is natural to wonder how much of the full relativistic conformal group descends to a symmetry of
the Navier-Stokes equations, and in what form it does so. It is obvious that the relativistic Poincare
group descends to the Galilean symmetry group of the Navier-Stokes equations. In the next section
we explain that the scaling symmetry operation described in the previous paragraph is loosely related
to the dilatation operator of the parent relativistic conformal theory. 3 Further we demonstrate that
all spatial special conformal transformations also descend to exact symmetries of the Navier-Stokes
equations. After our scaling these transformations effectively turn out to be the ‘boost’ to a uniformly
accelerated frame; the inertial forces one has to deal with when working in a non-inertial frame are
compensated for by a shift of the pressure. These spatial special conformal transformations, together
with the Galilean group and the scaling symmetry described above, form the (d + 2)(d + 1)/2 − 1
dimensional symmetry algebra of the Navier-Stokes equations. We list the commutation relations of
this algebra4 , and the action of its generators on the velocity fields, in detail in the next section.
The conformal symmetry algebra described above is just subset of the full infinte dimensonal
symmetry algebra of the Navier Stokes equations [16] (see also [17–19] for other related work) . The
additional generators of the full symmetry algebra are very easy to describe; they consist of boosts to
2 The scaling to arbitrarily low velocities projects out sound waves with dispersion relation ω ∝ k.
3 This loose ‘descent’ is analogous to the relation of the dilatation operator of the Schroedinger group to the generator
of scale transformations of the massless Klein Gordon equation.
4 The nonrelativistic conformal symmetry group of the Navier-Stokes equations includes the contraction of spatial

conformal generators Ki but does not include a generator that descends from the temporal conformal generator K0 .
Our algebra is distinct from the Schroedinger group studied for example in [15]

–3–
a reference frame whose velocity is homogeneous in space but an arbitrary function of time. Just as in
our discussion in the paragraph above, the pseudo force from such a frame change may be cancelled
by an appropriate shift in pressure, and so is a symmetry of the Navier Stokes equations. 5
The scaling limit described above admits an interesting generalization. Consider the equations of
fluid dynamics on a base manifold Gµν = gµν + Hµν where Hµν is small. By taking all terms that
depend on Hµν to the RHS, ∇µ T µν = 0 reduces effectively to the equations fluid dynamics on the
base space gµν forced by an Hµν dependent forcing function. If we combine the scaling described in
the paragraph above with the H00 , Hij = O(ǫ2 ) and Hi0 = ǫAi (xi , t), the effective resultant forcing
function survives and is finite in the ǫ → 0 limit. It turns out that the effective scaled forcing function
depends only on Ai (xi , t) and has a very simple form. It is precisely the force applied on a charged
fluid by effectively a background electromagnetic potential A0 = 0, Ai = Ai (xi , t). Consequently the
‘magnetohydrodynamical’ Navier-Stokes equations (i.e. the Navier-Stokes equations with a forcing
function from an arbitrary background electromagnetic field) follows as a universal result of a scaling
of the equations of relativistic hydrodynamics with small metric fluctuations.
We now return to the duality between gravity and fluid dynamics. We apply the scaling limit
described in the previous paragraphs to the equations of fluid dynamics that are holographically dual
to gravity. This procedure gives us an asymptotically locally AdSd+1 gravity dual to any solutions of
the magnetohydrodynamical Navier-Stokes equations. The resultant metric describes small - but non
linearly propagating fluctuations about a uniform black brane. We present the explicit form of the
resultant bulk metric in section 4 below.
It follows from our discussion that every solution to the Navier-Stokes equations (1.1) is also
a scaling limit of a solution to Einstein’s equations with a negative cosmological constant with one
important caveat. The proviso is that many actual solutions of the Navier-Stokes equations describe
fluids subject to hard wall type boundary conditions, and we do not (yet?) understand how to
generate gravitational duals of these boundary conditions. However the qualitative effects of boundary
conditions may easily be mimicked by appropriate forcing functions which are completely in our
hands. In particular, while several experiments that study fully developed steady state turbulence do
so in fluids with hard wall boundary conditions, there should be no barrier to setting up the same
phenomenon for a fluid with no boundaries (e.g. on Rd or on a compact manifold) with the appropriate
forcing function. In order to make this expectation concrete we have identified one forcing function
(of a likely infinite plethora of possibilities) applied to a fluid on Rd , whose steady state end flow we
expect to be turbulent at asymptotically high Reynolds numbers. In the rest of this introduction we
describe this forcing function, and the reason we expect the flow it generates to be turbulent.
Consider the forcing function Ax = αeiω0 t+iko y + cc, together with Ai = 0 for i 6= x acting on (for
concreteness) a fluid, on R2,1 whose spatial sections are parameterized by the Cartesian coordinates
x and y. This gauge fields sets up a time and y dependent electric field in the x direction, together
with a magnetic field in the plane. It is very easy to find one exact solution to the equations of fluid
dynamics subject to this forcing function. On this solution vx is proportional to Ax and vy = 0.
This solutions describes a fluid driven into motion in the direction of an applied electric field (the x
direction), while Lorentz forces from the magnetic field, in the y direction, are balanced by pressure
gradients. All nonlinear terms in the Navier-Stokes equations vanish when evaluated on our solution.
However nonlinear terms have an important effect on the dynamics of small fluctuations about our
solution at high Reynolds number. In particular, in Appendix A below we demonstrate that nonlinear
5 We thank J. Maldacena and R. Gopakumar for suggesting a symmetry enhancement along these lines, and thank

J. Maldacena for drwaing our attention to [16], see especially the top of pg 68. R. Gopakumar and collaborators are
currently further studying this algebra and its extensions.

–4–
terms drive some fluctuation modes (with momentum in the x direction) unstable at high enough
Reynolds numbers. In order to explain the possible significance of this instability, it is useful to recall
the usual situation with fluids at high Reynolds numbers.
As we have described earlier in our introduction, fluid flows at high Reynolds numbers are very
rich, and of potential theoretical interest. Nonetheless, in every highly symmetrical situation, there
exists a simple ‘laminar’ solution, that preserves all the symmetries of the problem, at every value
of the Reynolds number. This solution is often simple to determine analytically, and certainly shows
none of the fascinating phenomenology of turbulence. However in interesting situations this solution
becomes ‘tachyonic’ i.e. goes unstable to linear fluctuations that break some of the symmetries of the
problem, above a critical Reynolds numbers. The ‘end point’ of this ‘tachyon condensation’ typically
has richer dynamical behavior than the original solution itself. As the Reynolds number is further
increased, further instabilities are usually triggered, and at arbitrarily high Reynolds numbers flow is
turbulent.
We suspect that the dynamical pattern described in the previous paragraph applies to the exact
solution described in this paper. As we currently have no theoretical tools to predict the onset of
turbulence in any fluid flow this is necessarily a guess, but one that we believe is natural, given the
results of our stability analysis. This guess suggests that the stable steady state solution to AdS4
gravity, with the effective gauge field Ax described above, is dual to a turbulent fluid flow at high
Reynolds numbers. Of course the particular situation described above is only one of a plethora of
possibilities. We describe this solution in detail in Section 5 and Appendix A below only in order to
have one concrete example of a gravitational set up that is likely to be dual to a turbulent fluid flow;
not because we think that our particular is distinguished in any way.
We believe it is likely that the fluid - gravity map will lead to interesting new insights on the nature
of solutions of Einstein gravity in the presence of a horizon6 . It also does not seem impossible that
gravitational techniques and methods will prove useful in bringing new insights into the investigation
of fascinating fluid phenomena like turbulence. In particular, it would be fruitful to understand the
Kolmogorov laws on well-developed turbulence and their modification within the gravity framework.
The symmetry algebra (3.5) may also throw new light on these issues. We leave further investigation
of these issues to future work.
Note Added : While we were completing this paper we received the preprint [20] which has
substantial overlap with Section 2 of this paper.

2. Scaling the Navier-Stokes Equation

2.1 The Navier-Stoke Equation as a universal limit of fluid dynamics

In this section we will display a scaling limit 7 that reduces fluid dynamical equations of the form
∇µ T µν = 0, to the incompressible non-relativistic Navier-Stokes equations usually studied in fluid
dynamics text books (see e.g. [1]).

6 For instance, consider a black brane in asymptotically flat space. At sufficiently low energies, a beam of gravitons

shot at this brane perturbs the non normalizable boundary conditions of the effectively AdS near horizon region of the
brane. It may be possible to choose this perturbation to drive the gravity in the near horizon AdS region turbulent.
7 This section was worked out in collaboration with J. Maldacena.

–5–
 The equations of relativistic fluid dynamics are

∇µ T µν = 0
T µν = ρuµ uν + P P µν − 2ησ µν − ζθP µν + . . .
P µν = g µν + uµ uν , (2.1)
 
µν µα νβ gαβ
σ =P P ∇α uβ + ∇β uα − ∂.u , θ = ∇β uβ
d−1
Here P is the pressure, ρ the energy density, η the shear viscosity, ζ the bulk viscosity of the fluid,
and gαβ the metric of the space on which the fluid propagates. Each of the fluid quantities listed
above may be regarded as a function of the fluid temperature - and also of chemical potentials if the
fluid is charged. The . . . in the equation above refer to terms of second or higher order in spacetime
derivatives.
If the fluid is charged, then we must supplement the equation (2.1) with an equation of charge
conservation for every conserved charge. Below, we comment briefly on the scaling limit of this
equation.

2.2 A scaling limit

Now let us study the motion of a fluid on a metric of the form Gµν = gµν + Hµν . We assume that the
background metric gµν has the form

gµν dxµ dxν = −dt2 + gij dxi dxj . (2.2)

(this is simply a choice of coordinate system, for a large class of metrics) while the small fluctuation
Hµν is completely arbitrary. As we have explained in the introduction, we intend to view the fluid
flow on the space with metric Gµν as an effectively forced flow on the space with metric gµν . In order
to do this we need to map the velocity field ũµ on the space Gµν to a velocity field on gµν . This map,
which must respect the requirement that u2 = ũ2 = −1, may be chosen in a natural fashion if one
thinks of the velocity field as generated by the path of ‘particles’ through spacetime. We now describe
this in more detail.
Let the fluid velocity on the space Gµν be given by ũµ where
1  ~
ũµ = √ 1, V (2.3)
V2

where V~ is a d − 1 spatial vector with components V i , V α is a d component object with components

~ ) and V 2 is Gαβ V α V β (the indices α, β run over d spacetime indices while the indices i, j run
(1, V
over the d − 1 spatial indices). Expanding ũµ to first order in Hαβ we have

ũµ = uµ + δuµ + . . .
1 
~

uµ = p 1, V
1 − gij V i V j (2.4)
uα uβ Hαβ
δuµ = −uµ
2
where uµ is the d velocity of the fluid referred to the metric gµν . All terms in δuµ above depend on
the fluctuation metric Hαβ ; below we will take all these terms to the RHS of (1.1) and view them as
contributions to the effective forcing function.

–6–
 One obvious solution to the equations of fluid dynamics on the space with metric gµν is simply
the fluid at rest with constant pressure P0 and density ρ0 . We now turn to a particular kind of small
amplitude and long distance fluctuation about this uniform fluid at rest and with pressure P0 and
energy density ρ0 on a manifold that is ‘close’ to gµν . More specifically we set

H00 = ǫ2 h00 (ǫxi , ǫ2 t)

H0i = ǫAi (ǫxi , ǫ2 t)
Hij = ǫ2 hij (ǫxi , ǫ2 t)
(2.5)
V i = ǫv i (ǫxi , ǫ2 t)
P − P0
= ǫ2 p(ǫxi , ǫ2 t)
ρ0 + P0
and take ǫ to be arbitrarily small. Although we have explicitly listed only the scaling of the pressure
P above, the energy density ρ and viscosity ν also scale in a similar fashion (this is consistent with
the fact that they are all functions of the same underlying variables). We have normalized pressure
fluctuations by ρ0 + P0 rather than P0 for future convenience.
We will now examine what happens to the Navier-Stokes equations under this scaling. Let us first
start with the 0 or temporal component of these equations. It is easy to check that

∇µ T µ0 = ǫ2 ρe ∇i v i + O(ǫ4 )


(2.6)
ρe = ρ0 + P0

Consequently, in the limit of small ǫ this equation reduces simply to ∇i v i = 0, where ∇i is the covariant
derivative with respect to the purely spatial metric gij .
Let us now turn to the spatial Navier-Stokes equations. After some calculation we find
" ! #
∇j v i + ∇i v j ~ v
µi 3 i i µ ij ∇.~ ~ v − f + O(ǫ5 )
i


∇µ T = ǫ ρe ∇ p + ρe ∇µ v v − 2η∇j −g − ζ∇i ∇.~
2 d−1
√
∂j ( gAi v j )
 
∂i h00
f i = ρe − ∂0 Ai − √ + v j i
∂ Aj
2 g
(2.7)
√
∂i ( gv i )
where v µ = (1, v i ). Using the equation ∇i v i = √
g = 0, the coefficient of ǫ3 in the equation above
may be simplified to

∂ i h00
∇i p + ∂0 v i + ~v .∇v i − ν ∇2 v i + Rji v j = − ∂0 Ai + Fji v j (2.8)
2
where Fij = ∂i Aj − ∂j Ai is the field strength for the vector field Ai and ν = η/ρe is the ‘kinematical
viscosity’ of the fluid. 8 (2.8) takes a somewhat simpler form in terms of slightly redefined variables.
Let
Ai = ai + ∇i χ
8 Note also that we have used conventions in which

[∇ρ , ∇σ ]Aν = Aβ Rβ
νρσ , Rβ
αθβ = Rαθ

With these conventions the scalar curvature of a unit d sphere is d(d − 1) and the Ricci tensor of a unit sphere is given
by Rij = (d − 1)gij .

–7–
where χ is chosen to ensure that ∇i ai = 0. This is the usual split of a gauge field into its pure curl
and pure divergence parts. Note that

fij ≡ ∂i aj − ∂j ai = Fij .

We also define the effective pressure

1
pe = p − h00 + χ̇
2
in terms of which (2.8) reduces to

∇i pe + ∂0 vi + ~v .∇vi − ν ∇2 vi + Rij v j = −∂0 ai − v j fji

(2.9)

(2.9) is precisely the Navier-Stokes equation with forcing function generated by an effective background
electromagnetic field (with a0 = 0 and spatial vector ai ) on the effectively charged fluid.

2.3 Charged Fluids

If the fluid under study carries an extra set of conserved charges, it obeys an extra set of conservation
equations of the form ∇µ J µ = 0. It is easy to verify that these equations all reduce to the condition
that the fluid velocity is divergence free in the scaling limit under study in this section. Consequently
charged fluids obey the same equations as uncharged fluids in the scaling limit under study in this
subsection.

2.4 Reduction of Cauchy Data

Returning to the study of uncharged fluids, the Cauchy data9 of the parent relativistic fluid dynamical
equations consists of d real functions of space; the value of the pressure field and the value of the d − 1
independent velocity fields on an initial time slice. As the fluid dynamical equations are of first order
in time, the Cauchy data of the problem does not include the time derivatives of all these fields.
Now let us examine the Cauchy data of the incompressible Navier-Stokes equations. Note that
the spatial divergence of (2.8)

∇2 pe = −∇i v j ∇j v i − v i v j Rij + ∇i −νRji + fji v j



(2.10)

determines the pressure of the fluid as a function of the fluid velocity (only the velocity - not its time
derivatives ). In other words the independent data in the fluid is simply given by d − 2 real functions
that parameterize an arbitrary divergence free velocity field.
It follows that two of the degrees of freedom of the equations of relativistic fluid dynamics are
lost on taking the scaling limit of the previous subsection. These two degrees of freedom are simply
the fluctuations of the pressure and the divergence of the velocity. At the linearized level these two
degrees of freedom combine together in sound mode fluctuations. (Note that the relativistic dispersion
of sound implies that a sound mode has twice as much data as a shear mode.). Consequently the
reduction of Cauchy data in the scaling limit of this paper follows simply a nonlinear restatement of
the observation that sound waves are projected out in our scaling limit.

9 We thank S. Trivedi for very useful discussions on the topic of this subsection.

–8–
3. Symmetries
As we have noted above, the Navier-Stokes equations may be obtained by applying the appropriate
scalings to the equations of relativistic hydrodynamics with an arbitrary equation of state. The
parent hydrodynamical system may in particular be chosen to be conformally invariant. It is natural
to wonder whether either the dilatation or special conformal transformations of the parent theory
descend to a symmetry of the the Navier-Stokes equations. In this section we answer this question in
the affirmative. 10

3.1 Dilatations
µ
Let us first explain how this works for dilatations. Dilatations consist of a diffeomorphism (x′ )µ = xλ ,
µ
T ′ = T , (u′ )µ = uλ (g ′ )µν = λ2 gµν compounded with the Weyl transformation x̃µ = (x′ )µ , g̃µν =
(g′ )µν
λ2 , ũµ = λ(u′ )µ , T̃ = λT . In sum

xµ
x̃µ = , g̃µν = gµν , ũµ = uµ , T̃ = λT.
λ
This action on coordinates, temperatures and fields, is a symmetry of the equations of conformal
relativistic fluid dynamics. While this operation commutes with our scaling limit, it is not by itself
a symmetry of the Navier-Stokes equations. This is because the kinematical viscosity ν, regarded as
a parameter in the Navier-Stokes equations, is proportional to T1 and so changes under the variable
transformation described above. The Navier-Stokes equations are however simple enough to allow
dilatation transformations listed above to be modified into a true symmetry of the Navier-Stokes
equations by ‘absorbing’ the transformation of ν into ‘anomalous’ transformations of time and velocity.
The result of this procedure is simply (1.3) with which we started this note.

3.2 Special Conformal Transformations

The situation is more straightforward for special conformal transformations. The scaling law for the
velocity and temperature fields, under a special conformal transformation, may also be obtained by
compounding a diffeomorphism with the appropriate Weyl transformation. Restricting to infinitesimal
conformal transformations, we find 11 12

δxµ = −2c.xxµ + x2 cµ
δuµ = −2 [xµ cν − xν cµ ] uν − δxν ∂ν uµ (3.2)
ν
δT = 2c.xT − δx ∂ν T

Now note that special conformal transformations induce an additive shift on the temperature
fluctuation, δT , proportional to x.cT0 where T0 is the temperature of the background. In order that
10 Thissection was worked out in collaboration with R. Loganayagam.
11 Inorder to verify the covariance of local equations under these symmetry transformations, it is sufficient to omit
the terms proportional to δxµ ∂µ but instead to transform all derivatives according to the rule
` ´ ˆ ˜
δ ∂β = 2 cβ x.∂ − xβ c.∂ + x.c∂β

12 Under this transformation, the shift in a conformal stress tensor is given by

δT µν = 2d(c.x)T µν + 2(xλ cµ − xµ cλ )Tλν + 2(xλ cν − xν cλ )Tλµ − δxλ ∂λ T µν (3.1)
Upon accounting also for the shift in the derivatives, it is easy to convince oneselves that the equation of energy
momentum conservation, for any identically traceless stress tensor, is invariant under special conformal transformations.

–9–
this shift respect the ǫ2 scaling of δT we are required to scale c0 ∝ ǫ4 and ci ∝ ǫ3 . Imposing this
scaling and retaining terms only to leading order in the velocity expansion (3.3) reduces to13

δt = 0
δxi = −t2 ci
(3.3)
δv i = −2ci t + t2 cj ∂j v i
δT = 2(−c0 t + ci xi )T + t2 cj ∂j T

It is clear that the symmetry generated by c0 acts trivially (it does not act on coordinates or
velocities, but merely generates a shift, linear in time, of the pressure; more about this below). However
the symmetries generated by ci act nontrivially, and are directly verified to be symmetries of the
Navier-Stokes equations. 14

3.3 The Full Symmetry Group

We have thus discovered that the Navier-Stokes equations enjoy invariance under a conformal sym-
metry group. The generators of this group are the dilatation D, special conformal symmetries Ki
Gallilian boosts Bi , the generator of time translations (energy) H, momenta Pi and spatial rotations
Mij . The action of these generators on velocity fields is given by

Dv j = (−2t∂t − xm ∂m − 1) v j
Ki v j = − 2tδij + t2 ∂i v j
Bi v j =δij − t∂i v j
(3.4)
Hv j = − ∂t v j
Pi v j = − ∂i v j
Mik v j =δij v k − δkj v i − (xk ∂i − xi ∂k )v j

13 As above, in verifying the covariance of equations under these transformation, it is sufficient to omit the terms

proportional to t2 cj ∂j above, but instead to transform derivatives according to

δ (∂t ) = 2tci ∂i δ (∂i ) = 0.

δp
14 Under the transformations listed in (3.3) we have δpe = dP = 2c.x so that δ∂i pe = 2ci . Further δ(v̇i +v.∇vi ) = −2ci
0
and the viscous term is unchanged. Adding all terms together we have a symmetry of the equations.

– 10 –
 The commutation relations between these various generators is given by

[D, Ki ] = − 3Ki
[D, Bi ] = − Bi
[D, H] =2H
[D, Pi ] =Pi
[D, Mij ] =0
[Mij , Pk ] = − δik Pj + δjk Pi
[Mij , Kk ] = − δik Kj + δjk Ki (3.5)
[Mij , Bk ] = − δik Bj + δjk Bi
[Mij , H] =0
[Ki , Pj ] =0
[Ki , Bj ] =0
[Ki , H] = − 2Bi
[H, Bj ] = − Pi

The symmetry algebra listed in (3.5) may presumably be obtained from the appropriate contraction
of the parent symmetry algebra SO(d, 2). A related study is currently under progress in [21].
In addition to the symmetries listed above, the Navier-Stokes equations have an infinite dimen-
sional group of trivial symmetries, under which the pressure is simply shifted by an arbitrary function
of time. These are symmetries of the equation because only gradients of the pressure enter the Navier-
Stokes equations, and they are trivial because the pressure is not really an independent variable of the
Navier-Stokes equations, which may in fact be eliminated by taking the curl of those equations. For
this reason one need not keep track of the action of symmetry generators on the pressure. However it
is not difficult to do so and we find

Dpe = (−2t∂t − xm ∂m − 2) pe
Ki pe =2xi + t2 ∂i pe
Bi pe = − t∂i pe
(3.6)
Hpe = − ∂t pe
Pi pe = − ∂i pe
Mik pe = − (xk ∂i − xi ∂k )pe

This action of the symmetry generators on the pressure field do not quite yield the commutation
relations (3.5), but instead have additional terms on the RHS corresponding to generators of the
trivial symmetries referred to above (i.e. generators that shift the pressure field by a function of time).
Spatial derivatives of the pressure field, however, honestly transform according to the algebra (3.5).
Consequently, the symmetry algebra (3.5) is not represented on the pressure field itself, but on its
spatial derivatives. This is vaguely reminiscent of the fact that the two dimensional conformal group
has well defined action on all derivatives of massless scalar fields, but not on the field itself.

– 11 –
4. The gravity solution dual to fluid dynamics in the scaling regime
4.1 The dual bulk metric
As we have described above, any solution of the incompressible non-relativistic Navier-Stokes equations
solves the equations of fluid dynamics dual to gravity up to O(ǫ3 ), under the scaling listed in the
previous section. Now in [9, 10, 13] the equations of fluid dynamics are obtained as the Einstein
constraint equation of a bulk asymptotically locally AdSd+1 space. It is thus natural to define the
gravitational dual to a solution of the Navier-Stokes equations as any small fluctuation about a black
brane background that solves all of Einstein’s equations (constraint as well as dynamical) to cubic
order in ǫ. Adopting this definition, it is easy to read off the bulk metric dual to the Navier-Stokes
equations from an appropriate scaling of the bulk metric of [9, 10, 13].
In computing the bulk metric upto O(ǫ3 ) in the sense described above, it turns out that terms
from only the zeroth order and the first order in the derivative expansion of the gravitational solutions
of [9, 10, 13] are relevant. The metric up to first order in derivative has the following form

ds2 = ds20 + ds21

where
1
ds20 = −2uµ dxµ dr + uµ uν dxµ dxν + r2 gµν dxµ dxν
bd rd−2
ds21 = −2ruν uα ∇ ¯ α uµ dxµ dxν + 2 r ∇ ¯ α uα uµ uν dxµ dxν + 2br2 F (br) σµν dxµ dxν



d−1
1 ¯
1
¯ ν uµ + ¯ α uµ + uµ uα ∇ ¯ α uν − 1 ¯ α uα (uµ uν + gµν ) (4.1)
uν uα ∇





σµν = ∇µ uν + ∇ ∇
2 2 d−1
Z ∞  d−1 
y −1
F (x) = dy
x y(y d − 1)
d
b= = b0 + δb
4πT
T = T0 + δT

Here ∇¯ denotes the covariant derivative with respect to the full boundary metric which is equal to
a background gµν plus perturbation Hµν . T0 is the temperature of the background blackbrane. The
terms that appear in the metric involve covariant derivatives of the d velocity uµ with respect to the
full boundary metric. These terms can be expressed as covariant derivatives of the d − 1 velocity vi
and the metric perturbation Ai = H0i with respect to spatial part of the background metric gij .

¯ i uj = ∇i vj + O(ǫ4 )
∇
¯ i u0 + ∇
∇ ¯ 0 ui = ∂0 (vi + Ai ) − 1 ∂i h00 − 1 ∂i (vj v j ) − v j Fij + O(ǫ4 )
2 2
∇¯ µ uµ = ∇j v j + O(ǫ4 )
(4.2)
¯ µ u0 = O(ǫ4 )
uµ ∇
¯ µ ui = ∂0 (vi + Ai ) − 1 ∂i h00 + (v j ∇j )vi − v j Fij + O(ǫ4 )
uµ ∇
2
Fij = ∂i Aj − ∂j Ai

Here ∇ denotes the covariant derivative with respect to gij . The raising and lowering of the i, j
indices are also with respect to the metric gij . To simplify the expression of σµν in (4.2) the constraint

– 12 –
∇i v i = 0 has been used. Using these expressions the derivative part of the metric can be written as

ds21 = b0 r2 F (b0 r) (∇i vj + ∇j vi ) dxi dxj − 2b0 r2 F (b0 r)v j (∇i vj + ∇j vi ) dt dxi
(4.3)
 
1
+ 2r ∂0 (vi + Ai ) − ∂i h00 − v j Fij + (v j ∇j )vi dt dxi + O(ǫ4 )
2

Here the first term is of order ǫ2 and the last two terms are of order ǫ3 . Since from the constraint
equations ∇i v i = 0, there is no contribution from the scalar sector. The zeroth order metric can also
be expanded in powers of ǫ. It turns out that to solve Einstein equation up to order ǫ3 ,it is sufficient
to expand the zeroth order metric up to order ǫ2 in the fluctuations.
1
ds20 = dt2 + r2 −dt2 + gij dxi dxj + 2dt dr


bd0 rd−2
2
− (Ai + vi ) dt dxi − 2 (Ai + vi ) dxi dr
bd0 rd−2
1 1
−d δb + vj v j − h00 dt2 + d−2 (Ai + vi ) (Aj + vj ) dxi dxj − −vj v j + h00 dt dr



+ d+1
b0 rd−2 r
(4.4)

Here the first line is of order ǫ0 , the second line is of order ǫ1 and the third is of order ǫ2 . The full
metric is given by the sum of (4.4) and (4.3); and solves Einsteins equations to O(ǫ3 ) provided the
velocity and temperature fields above obey the incompressible Navier-Stokes equations (1.1). These
equations imply in particular that

∇2 T
 1
= −∇i v j ∇j v i − v i v j Rij + ∇i −νRji + Fji v j + ∇2 h00 − ∂0 (∇.A)

(4.5)
T0 2

an equation that determines δT (and hence δb in (4.4)) as a spatially nonlocal but temporally ultralocal
functional of the velocity fields v i . It follows that though the bulk metric at xµ is determined locally as
a function of temperatures and velocities at xµ , it is not determined locally as a function of velocities
at xµ . This non locality is a consequence of the infinite speed of sound (and consequent action at a
distance) in our scaling limit.

4.2 Comments on the bulk metric

Several comments on the bulk metric (4.3) and (4.4) are in order. Let us first spell out some termi-
nology. We refer to the parts of the bulk metric that are proportional to a linear combination of dt2
or i (dxi )2 as its scalar components. Terms in the metric proportional to dtdxi are called its vector
P

components, while terms proportional to dxi dxj are referred to as its tensor components.
Now the derivative expansion of [4–13] solves the Einstein equations in a multi step process. In the
first step the dynamical Einstein equations are used to determine the r dependence of the bulk metric
as a function of boundary data, in each of the sectors described above. The requirement of regularity
of the future horizon is then used to determine the boundary data of the tensor sector in terms of the
boundary data in the vector and scalar sectors (roughly the fluid temperature and velocity).
In the scaling limit described in the previous subsection, the tensor sector of the bulk metric is
particularly simple. It is given by
1
b0 r2 F (b0 r) (∇i vj + ∇j vi ) dxi dxj + (Ai + vi ) (Aj + vj ) dxi dxj . (4.6)
rd−2

– 13 –
Note that even though the velocity vi is scaled to zero in the scaling limit under study in this paper,
the leading order in ǫ metric (4.6) includes a quadratically nonlinear in the velocities. This is a
consequence of the fact that we scale distances to infinity at the same rate as velocities to zero. Had
we simply scaled amplitudes to be small, while keeping distance scales finite, we would have ended up
with only the first of these two terms in (4.6) and the resultant geometry would simply have been the
dual of linearized fluid dynamics.
The Navier-Stokes equations are obtained out of the Einstein constraint equations acting on the
bulk metric described above. The first term in (4.6) gives rise to the viscous term in the Navier-
Stokes equations, while the second term in (4.6) is the origin of the nonlinear convective term in
those equations. It is consequently not surprising that ratio of the first to the second term in (4.6) is
proportional to the Reynolds number Re of the flow: in fact it is of order Re × rd F1(r) . Now F (r) ∼ r1 .
at large r; so that the viscous term in (4.6) appears to dominates the nonlinear term in that expression
when r > (Re)1/d−1 . This appearance is atleast partly a coordinate artifact, as is evidenced by the
fact that the two terms in (4.6) contribute equally to the Einstein constraint equations at every r (this
follows as as these equations are independent of r). 15

5. Simple Steady State Shear flows at arbitrary Reynolds Number

Consider a fluid on Rd−1,1 , subject to an effective forcing
Z
ax = dωdka(k, ω) exp [iky + iωt]
(5.1)
ai = 0, (i 6= x)
Here x and y parameterize orthogonal displacements in the d − 1 spatial directions, and a(k, ω) is any
function.
In the presence of this forcing function, it is easy to verify that the velocity field
a(k, ω)eiky+iωt
Z
vx = − dkdω 2
1 − i νk
ω (5.2)
vi = 0 i 6= x
(together with the pressure field which may be obtained by integrating (2.10) ), is an exact solution
of (2.9).
In order to be concrete let us make the simple choice a(k, ω) = a(−k, −ω) = αδ(k − k0 )δ(ω − ω0 ).
In this special case the forcing function is given by
ax = α exp [ik0 y + iω0 t] + cc
(5.3)
ai = 0, (i 6= x)
and the velocity field is given by
αeik0 y+iω0 t
vx = − νk02
+ cc
1−i ω0 (5.4)
vi = 0 i 6= x
15 The ratio computed above naively suggests that nonlinearities at different effective Reynolds numbers are separated

in the bulk radial coordinate; a result reminiscent of the approximate locality in scale of high Reynolds number fluid
flows. While this interpretation is too quick for the purposes of evaluating the Einstein constraint equation, perhaps
the naive ratio of the two terms in (4.6) is physically relevant for some physically interesting questions. We feel this
deserves further contemplation.

– 14 –
 2
16
This solution is characterized by two dimensionless numbers, a = kαo ν and b = νkω The Reynolds
number for this solution is given by
vL a
R= =√ . (5.5)
ν 1 + b2
and can be made large by taking a to infinity at fixed b. This may be achieved, for instance, by
increasing the strength of the forcing amplitude α keeping everything else fixed.
In Appendix A we outline a linear stability analysis of the special flow described above. We find
that in d = 2 this flow is unstable at large enough Reynolds numbers for a range of the parameter
b. As we have described in the introduction, this suggests that the flow is dynamically interesting,
and likely turbulent, at asymptotically high Reynolds numbers. Assuming this is the case, we have
identified atleast one gravitational system that is dual to steady state turbulence. In some detail,
consider gravity with a negative cosmological constant, subject to the small z boundary condition on
the metric
1
lim ds2 = 2 dz 2 − dt2 + ǫdtdxi ai + dxi dxi


z→0 z
(where i = i . . . d − 1 and ai is the gauge field (5.1)). The steady state solution of this gravitational
system is dual to a turbulent flow of d dimensional fluid √ dynamics whenever the parameters in the
function ai above are chosen to satisfy the inequality a ≫ 1 + b2 .
In Appendix B we have presented a generalization of the solution described in this section to the
forced flow of a fluid on a sphere. In that situation one should obtain the dual of a turbulent fluid in
an spacetime that is asymptotically global AdS.

Acknowledgements
We would like especially to thank J. Maldacena for discussions over which the scaling limit described
in this paper were formulated, collaboration in the initial stages of this project, and useful discussions
and comments throughout its execution. We would also especially like to thank R. Loganayagam
for collaboration on the calculations in Section 3, for useful discussions throughout the execution
of this project and for several detailed comments on the manuscript. We have profited greatly from
discussions with S. Trivedi and have also had fruitful discussions with J.Bhattacharya, R. Gopakumar,
G. Mandal, H. Ooguri and M. Van Raamsdonk. We would like to acknowledge useful discussion with
the students in the TIFR theory room. The work of S.M. was supported in part by Swarnajayanti
Fellowship. The work of S.R.W was supported by J.C.Bose fellowship. We would all also like to
acknowledge our debt to the people of India for their generous and steady support to research in the
basic sciences.

Appendices

A. Stability Analysis
In this appendix we study linear fluctuations about the solution (5.4) for the special case d = 3. In
two spatial dimensions the dual of a divergenceless velocity is the gradient of a scalar, and the field
strength of a gauge field is itself dual to a scalar. In other words we can write
vi = ǫij ∂j χ
(A.1)
fij = ǫij f
16 This special solution may be thought of as living either on a spatial Rd−1 or on a spatial torus, whose y edge is of
2π
size k0
.

– 15 –
 The Navier-Stokes equations (2.9) may be rewritten (after eliminating the pressure by taking the
curl) in terms of these scalars as

−∇2 χ̇ + ν∇4 χ − ǫjm ∇m χ∇j ∇2 χ + f˙ + ǫiq ∇i χ∇q f = 0 (A.2)

In terms of these variables the solution (5.4) takes the form

f0 = −ik0 α exp [ik0 y + iω0 t] + cc

1 αeik0 y+iω0 t (A.3)
χ0 = − × νk2
+ cc
ik0 1 − i ω00

A.1 Setting up the Eigenvalue Equations

Let us now set
 Z Z 
iq0 x iωy+ikx −iq0 x ∗ −iωy−ikx
χ = χ0 + θ e dωdkχq0 (ω, k)e +e dωdkχq0 (ω, k)e

where θ is a small parameter, and solve (A.2) to linear order in θ. Note that while our fluctuation has
a specific frequency (namely q0 ) in the x direction, it is a sum over frequencies in the y direction and
in time. This is necessary as our background breaks translational invariance in the y and t directions,
but preserves this invariance in the x direction. It is easy to work out the linear equations that our
fluctuation coefficients χq0 (ω, k) obey. We find

iω(k 2 + q02 ) + ν(k 2 + q02 )2 χq0 (ω, q) + β(iq0 ) (k − k0 )2 + q02 − k02 + αiq0 k02 χq0 (ω − ω0 , k − k0 )
  


+ β ∗ (iq0 ) (k + k0 )2 + q02 − k02 + α∗ iq0 k02 χq0 (ω + ω0 , k + k0 ) = 0




(A.4)

where
α
β=− νk02
.
1+ iω0

Let us define
fk (γ, κ) = χq0 [ω0 (γ + k), k0 (κ + k)] .
The fluctuation equation (A.4) may be recast in terms of fk as (we will usually omit to write the
functional dependence of fk in the equations that follow)

(κ + k − 1)2 + c2 − 1
   
i
(γ + k)((κ + k)2 + c2 ) + ((κ + k)2 + c2 )2 fk + ica − + 1 fk−1
b 1 − ib
 2 2
 (A.5)
(κ + k + 1) + c − 1
+ ica∗ − + 1 fk+1
1 + ib

The dimensionless quantities a, b and c are parameters in (A.5) (a and b were defined in section 4
while c = kq00 ). γ and κ are variables in this equation; (A.5) is the condition that a matrix acting on
the columns {fk } has a zero eigenvalue. Together with the physical requirement that fk decay at large
|k|; this eigenvalue equation yields an expression for unknown temporal frequency γ as a function of
κ. 17 That is, the equation (A.5) should yield a dispersion relation of the form

γ = γ(κ, a, b, c, n) (A.6)
17 Recall that κ is a real number (in order that the mode in question is well defined at all y) of unit periodicity.

– 16 –
where the integer n labels which of the infinitely many solutions to the zero eigenvalue condition we
have chosen to study. A result for γ with a negative imaginary part represents an instability of the
system.
Let us study the large |k| asymptotics of the variables fk in a little more detail. According to
our boundary conditions, at large positive k fk+1 is much smaller than fk−1 . As a consequence (A.5)
implies that  
fk ica k + 2(κ − 1) 1
= × × 1 + O( 2 ) (A.7)
fk−1 1 − ib k 2 (k + 4κ + bi ) k
It follows that, at large k,
 k
ica Γ(k + 2κ − 1)
fk ≈ D(a, b, cκ) i
. (A.8)
i − ib Γ(k + 1)2 Γ(k + 4κ + b + 1)

where D is a constant. This estimate is valid provided that k ≫ 1 and that k 2 ≫ √ ca . Similarly,
1+b2
the behavior of fk at large negative values of k is given by
−k
ica∗

Γ(−k − 2κ − 1)
fk ≈ D′ (a∗ , b, c, κ) i
(A.9)
i + ib Γ(−k + 1)2 Γ(−k − 4κ + b + 1)

where D′ is an independent constant. Note χ is real provided that D(a, b, c, κ) = D′ (a∗ , b, −c, −κ); a
condition that we consequently demand on physical grounds.
We are interested in determining the dispersion relation (A.6) as the condition for the existence of
a solution of (A.5) that achieves both asymptotic conditions (A.8) and (A.9) (this leads to an equation
as, generic solutions of (A.5) that obey (A.8) blow up at large negative k).

A.2 Qualitative nature of the solution

Let us first note that the off diagonal terms in (A.5) are both proportional to a×c. The proportionality
to a reflects the fact that the problem gets strongly coupled at high Reynolds numbers; we will have
a lot more to say about this below. The fact that these terms are proportional to c, however, implies
that fluctuations with no x momentum are governed by the linear Navier-Stokes equations, and so
are always stable. An instability, if it occurs, must necessarily break a symmetry (in this case of x
translations) that the background solution preserves.
In much of the rest of this section we specialize our analysis to κ = 0 and c = 1 (i.e. k0 = q0 ) but
for a reasonably wide range of the parameters a and b. We perform this specialization for convenience
in the numerical analysis that we will describe below; we expect qualitatively similar results at all
values of κ and reasonable nonzero values of c, though we have not explicitly tested this expectation.
Let us start by painting a qualitative picture of physically relevant solutions of (A.5). We start with
the simple limit a → 0. Solutions are given exactly by fk = δk,K with γ = −K +ib(K 2 +1) for K rang-
2
ing over integers. The time dependence of this solution is given by fk (t) = fk (0)e−iKω0 t−bωo (K +1)t ;
it follows that all these modes decay in time, so that our solution is stable against small fluctuations.
a
At small but nonzero values of Reynolds number Re = √1+b 2
the solutions to (A.5) are small
perturbations of the solutions described in the paragraph above. On the K th such solution fk is
nonzero at k = K, but decays rapidly as k moves away from K. In particular fk−K ∝ (Re)|k−K| . 18
The dispersion relation for γ, described in the previous paragraph, is corrected in a power series in
(Re)2 . We will compute the first term in this series below.
18 This decay is visible, for instance, in the asymptotic forms (A.8) and (A.9).

– 17 –
 As the Reynolds number is increased, the width in the distribution of fk (as a function of k
√ 1
centered around k = K) increases. For Re ≫ 1 and Re ≫ n this width is of order k ∼ (Re) 2 . (see
(A.8) and (A.9)).
In summary, the fluctuation mode we study is highly localized about a particular y momentum
1
at small Reynolds number, but consists of a ‘cascade’ or roughly equal superpositions of order (Re) 2
modes at large Reynolds numbers. The fluctuation modes oscillate rather than growing exponentially
in time at small Reynolds numbers; however the stability properties of these modes must be elucidated
by explicit calculation at large Reynolds numbers.

A.3 Numerical Evaluation of the frequency

In this subsection we present the results of our rough numerical evaluation of the frequency γ of the
mode with n = 1 as a function of the Reynolds number, at various values of b. The method we use
is very simple. We start with a value of k large enough for the asymptotic form (A.8) to be valid 19
. We then use the recursion relations (A.5) to evaluate f1 /f0 . We then independently evaluate the
same ratio using (A.9) and recursion relations. Equating these expressions for the same ratio gives us
an equation which we use to solve for γ as a function of all other parameters. Of course this equation
has several solutions. At small Reynolds number we choose the solution that is near to γ = −1 + 2ib
(i.e. the mode with K = 1 in the language of the previous subsection) , and then ‘follow’ this root as
we numerically increase the Reynolds number (in small steps) to large values. We have performed all
our calculations using Mathematica.

Figure 1: The imaginary part of the frequency γ plotted against a, for κ = 0, c = 1 and the root with K = 1.
Note that Im(γ) = 2 at b = 0 and decreases monotonically as |b| is increased.

Let us describe our results in detail at b = 1. In Fig. 1. below we present a plot of the
imaginary part of γ as a function of Reynolds number. In Fig. 2 below we plot the real part of the
same frequency as a a function of Reynolds number. Notice that the imaginary part changes sign
(indicating a transition to instability) at a Reynolds number of order 13.
We have performed the same analysis at several different values of b. At each value of b the
imaginary part of γ turns negative at a particular value of the Reynolds number. In Fig. 3 below we
have plotted this critical Reynolds number as a function of b for a range of values of b.
19 We have performed computations with values of this starting k ranging from 20 to 8. The plots presented below are

generated using the starting value 10. We have verified that our starting value of k is large enough, by checking that
our results are not significantly affected by increasing k.

– 18 –
Figure 2: The real part of the frequency γ plotted against a, for κ = 0, c = 1 and the root with K = 1. Note
that Re(γ) = −1 at b = 0 and decreases monotonically as |b| is increased.

Figure 3: The curve that separates the region of stability (below) from instability (above) of the eigenvalue
at K = 1, plotted as a function of the Reynolds number on the y axis versus b on the x axis. We have used
κ = 0, c = 1 in the calculations that generated this plot.

As our principle aim in this subsection is to establish that the small fluctuations about our solutions
are unstable at high enough Reynolds numbers we have not attempted to carefully estimate the errors
in our numerical calculations. However we think it is unlikely that the errors in, for instance, Fig. 3
exceed a few percent.

As we have described above, at every value of b we start our calculations at small Reynolds
numbers which are then slowly increased. As a check on our numerics, at small we have compared our
numerical results for γ versus k against the predictions of perturbation theory (see the next subsection
for details) at lowest order in (Re)2 ; we find good agreement. In fact at b = 1 and a = 1/100,
the difference between our numerical result and the zeroth order answer matched the prediction of
perturbation theory upto one part in 105 ; a more accurate result than we had expected.

– 19 –
A.4 Perturbation theory at small Reynolds numbers
Let us formally identify the equation (A.5) with free index k with
X
hk|M |ψi = hk|M |mihm|ψi
m

and simultaneously identify

hk|ψi = fk
With these formal identifications, the set of equations (A.5) (for all k) are simply the equation for the
operator M to have a zero eigenvalue. The corresponding eigenvector is specified by the values of fk
on this solution.
Let us write M = M0 + M1 where M0 is a diagonal operator in the basis described above and M1
is an off-diagonal operator, proportional to α. The elements of the matrix M0 and M1 are given by

hfk |M0 |fk i = A(k)

 
i
= (γ + k)((κ + k)2 + c2 ) + ((κ + k)2 + c2 )2 δkj
b
hfk |M1 |fk−1 i = B(k)
(κ + k − 1)2 + c2 − 1 (A.10)
 
= ica − + 1 δk,(j−1)
1 − ib
hfk |M1 |fk+1 i = C(k)δk,(j+1)
(κ + k + 1)2 + c2 − 1
 
= ica∗ − + 1 δk,(j+1)
1 + ib
When α = 0 the eigenvectors of the operator M are simply |Ki with eigenvalue A(K). This
eigenvalue vanishes when
γ0 = −K + i b (κ + K)2 + c2 .
 

We now wish to study how this eigenvector - and the corresponding solution for γ that keeps the
eigenvalue zero - evolves in perturbation theory in α.
To lowest nontrivial order in perturbation theory, the shift in the eigenvalue A(k) is given by
hfk |M1 |fk−1 ihfk−1 |M1 |fk i hfk |M1 |fk+1 ihfk+1 |M1 |fk i
Ã(k) = A(k, ω) + +
A(k) − A(k − 1) A(k) − A(k + 1)
(A.11)
B(k)C(k − 1) C(k)B(k + 1)
= A(k) + +
A(k) − A(k − 1) A(k) − A(k + 1)
In order that the eigenvalue vanish at this order we must have
 
 2 2
 ib B(K)C(K − 1) C(K)B(K + 1)
γ = −K + i b (κ + K) + c − +
(K + κ)2 + c2 A(K − 1) A(K + 1)
where the third term is evaluated at γ = γ0 and we have used the fact that A(K) = 0 at γ = γ0 .

B. A simple flow on a sphere

It is not difficult to generalize the simple laminar flow presented in section 4 above to a shear flow
on a 2 sphere. The corresponding dual gravitational solution to this flow is asymptotic to (slightly
perturbed) global AdS space.

– 20 –
 For concreteness we choose d = 3 and choose the base metric of our space to be the two sphere,
i.e.
ds2 = −dt2 + dθ2 + sin2 θdφ2 (B.1)
Our velocity and forcing function fields will be vector fields on the two sphere, so be briefly pause
to recall some necessary definitions. Recall that the spherical harmonics, Yml form a basis for the
expansion of an arbitrary scalar field on the sphere. On the other hand an arbitrary vector field on the
sphere is given by linear combinations of ∂i Yml and ǫji ∂j Yml where the ǫ symbol includes relevant factors
√
of g. Now, in our problem, both v i and ai are divergenceless. It follows that each of these fields
may be expanded in a sum over only the vector (rather than also the derivative of scalar) spherical
harmonics.
In order to obtain one solution (that can easily be generalized in many ways) to the Navier-Stokes
equations, let
ai = αǫm l
i ∂m Y0 (θ, φ).
20
The velocity configuration
α
vi = − × ǫm l
i ∂m Y0 (θ, φ).
1− i l(l+1)−2
ω

(together with the implied pressure field) yields a steady state, laminar, shear flow. The Reynolds
number of this flow is given by (5.5) with a and b now given by a = kαo ν and b = ν(l(l+1)−2
ω . As for the
solution presented in section 4, we expect this flow to be unstable to small perturbations at fixed b in
the limit of large a (we have not performed the linear stability analysis in this case).

References
[1] L. D. Landau and E. M. Lifshitz, Fluid mechanics. Course of theoretical physics, Oxford: Pergamon
Press, 1959.
[2] A. M. Polyakov, Conformal turbulence, hep-th/9209046.
[3] D. T. Son and A. O. Starinets, Viscosity, Black Holes, and Quantum Field Theory, and references
therein, Ann. Rev. Nucl. Part. Sci. 57 (2007) 95–118, [arXiv:0704.0240].
[4] S. Bhattacharyya, V. E. Hubeny, S. Minwalla, and M. Rangamani, Nonlinear fluid dynamics from
gravity, 0712.2456.
[5] R. Loganayagam, Entropy Current in Conformal Hydrodynamics, JHEP 05 (2008) 087,
[arXiv:0801.3701].
[6] M. Van Raamsdonk, Black Hole Dynamics From Atmospheric Science, JHEP 05 (2008) 106,
[arXiv:0802.3224].
[7] S. Dutta, Higher Derivative Corrections to Locally Black Brane Metrics, JHEP 05 (2008) 082,
[arXiv:0804.2453].
[8] S. Bhattacharyya et. al., Local Fluid Dynamical Entropy from Gravity, arXiv:0803.2526.
[9] S. Bhattacharyya et. al., Forced Fluid Dynamics from Gravity, arXiv:0806.0006.
[10] M. Haack and A. Yarom, Nonlinear viscous hydrodynamics in various dimensions using AdS/CFT,
arXiv:0806.4602.
20 We use that
∇m ∇m ∇i φ = ∇i ∇m ∇m + Rm
i ∇m φ
.

– 21 –
[11] J. Erdmenger, M. Haack, M. Kaminski, and A. Yarom, Fluid dynamics of R-charged black holes,
arXiv:0809.2488.
[12] N. Banerjee et. al., Hydrodynamics from charged black branes, arXiv:0809.2596.
[13] S. Bhattacharyya, R. Loganayagam, I. Mandal, S. Minwalla, and A. Sharma, Conformal Nonlinear Fluid
Dynamics from Gravity in Arbitrary Dimensions, arXiv:0809.4272.
[14] J. M. Maldacena, The large N limit of superconformal field theories and supergravity, Adv. Theor. Math.
Phys. 2 (1998) 231–252, [hep-th/9711200].
[15] D. T. Son, Toward an AdS/cold atoms correspondence: a geometric realization of the Schroedinger
symmetry, Phys. Rev. D78 (2008) 046003, [arXiv:0804.3972].
[16] V. N. Gusyatnikova and V. A. Yumaguzhin, Symmetries and conservation laws of navier-stokes
equations, Acta Applicandae Mathematicae 15 (January, 1989) 65–81.
[17] L. O’Raifeartaigh and V. V. Sreedhar, The maximal kinematical invariance group of fluid dynamics and
explosion-implosion duality, Annals Phys. 293 (2001) 215–227, [hep-th/0007199].
[18] M. Hassaine and P. A. Horvathy, Symmetries of fluid dynamics with polytropic exponent, Phys. Lett.
A279 (2001) 215–222, [hep-th/0009092].
[19] C. Duval, M. Hassaine, and P. A. Horvathy, The geometry of Schródinger symmetry in gravity
background/non-relativistic CFT, Annals Phys. 324 (2009) 1158–1167, [arXiv:0809.3128].
[20] I. Fouxon and Y. Oz, Conformal Field Theory as Microscopic Dynamics of Incompressible Euler and
Navier-Stokes Equations, arXiv:0809.4512.
[21] Gopakumar.R., Biswas.T., and Bagchi.A., Work under progress, .

– 22 –