See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/355664877

Optimal cooling of an internally heated disc

Preprint · October 2021

CITATIONS READS
0 13

1 author:

Ian Tobasco
University of Illinois at Chicago
20 PUBLICATIONS   198 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Ian Tobasco on 28 October 2021.

The user has requested enhancement of the downloaded file.

 OPTIMAL COOLING OF AN INTERNALLY HEATED DISC

IAN TOBASCO

To Charlie Doering, who taught us how to bound.

arXiv:2110.13291v1 [math.AP] 25 Oct 2021

Abstract. Motivated by the search for sharp bounds on turbulent heat transfer as well as the design of
optimal heat exchangers, we consider incompressible flows that most efficiently cool an internally heated
disc. Heat enters via a distributed source, is passively advected and diffused, and exits through the boundary
at a fixed temperature. We seek an advecting flow to optimize this exchange. Previous work on energy-
constrained cooling with a constant source has conjectured that global optimizers should resemble convection
rolls; we prove one-sided bounds on energy-constrained cooling corresponding to, but not resolving, this
conjecture. In the case of an enstrophy constraint, our results are more complete: we construct a family
of self-similar, tree-like “branching flows” whose cooling is within a logarithm of globally optimal. These
results hold for general space- and time-dependent sources that add more heat than they remove. Our main
technical tool is a non-local Dirichlet-like variational principle for bounding solutions of the inhomogeneous
advection-diffusion equation with a divergence-free velocity.

1. Introduction
Imagine a fluid in a container that is heated from within, and whose temperature is fixed at its boundary.
How should the fluid flow so that the container cools as quickly as possible? This question arises, for
instance, in the design of optimal heat exchangers, whose complicated shapes and resulting flows facilitate
heat transfer at rates far beyond diffusion [1, 7, 19, 40, 46]. More generally, the problem is related to the
ongoing search for sharp bounds on turbulent heat transfer in a variety of settings, including internally
heated [2, 3, 5, 20, 21, 31] as well as buoyancy-driven convection [15, 18, 45, 47], amongst others. While any
bound must explain why the rate of heat transfer associated to a given flow cannot exceed a certain amount,
sharp bounds are distinguished in that their values are attained by particular flows—flows that, as a result,
are extremal. Finding sharp bounds on convective heat transfer, or at least bounds that are asymptotically
sharp with respect to their scaling, has remained a widely open problem in our understanding of turbulence
since the pioneering works of Howard and Busse [6, 24] and Doering and Constantin [13] on Rayleigh-Bénard
convection (for a recent summary of the state of the art, see [11]).
A widely used technique for proving bounds on fluid dynamical quantities is the “background method”
introduced by Doering and Constantin [10, 12, 13], which takes the form of a convex variational problem
whose optimal value is the bound. Whether or not this method and its descendants (e.g., [26, 39]) produce
sharp bounds on convection remains largely unclear, leading some to ask whether other information beyond
the background method might be used, e.g., via the non-quadratic auxiliary functionals of [8, 37, 41, 43],
the conjecture that global optimizers of certain chaotic evolutions are steady [23], or the maximum principle
which is known to improve some bounds [36, 38]. On the other hand, recent progress with the background
method has led to asymptotically sharp bounds for internally heated convection with a steady and perfectly
balanced source [33]. Due to the balanced nature of the source, which puts in just as much heat as it takes
away, the overall rate of heat transfer in the bound turns out to be much larger than the usual “ultimate
scaling” law; it is nonetheless asymptotically sharp and is saturated by a bulk, convection roll-like flow.
Motivated by these questions, we consider an internally heated disc with a general space- and time-
dependent source, subject to an assumption that more heat is added than is removed. Although temperature
will be governed in our setup by the usual advection-diffusion equation, we will not impose a momentum
equation directly on the velocity, but will instead replace it with a constraint on the mean enstrophy it
implies. Optimizing with respect to this constraint leads to a self-similar, tree-like “branching flow” whose
cooling we prove is within a logarithm of globally optimal. While the a priori bound part of this result

Date: October 27, 2021.

1
can be achieved using a variant of the background method, we prefer to take a different approach which
highlights the variational structure of the advection-diffusion equation (more on this later).
Setting up the problem in detail, we let D ⊂ R2 be a disc with radius R > 0 and consider a temperature
field T (x, t) that is passively advected and diffused according to a divergence-free velocity field u(x, t) we
take to be under our control. At the disc’s boundary ∂D we set the temperature to be a constant, say T = 0,
and let f (x, t) be the internal source. Altogether, T solves the inhomogeneous advection-diffusion equation
(
∂t T + u · ∇T = κ∆T + f in D
(1.1)
T =0 at ∂D
where κ > 0 is the thermal diffusivity. We leave the value of T at the initial time t = 0 unspecified, as in the
long run it is lost to diffusion. To measure cooling efficiency, we use the mean-square temperature gradient1
ˆ τˆ
1
|∇T |2 = lim |∇T (x, t)|2 dxdt


τ →∞ τ |D| 0 D
whose value depends on u. The notation |D| stands for the area of the disc. Besides the divergence-free
constraint
∇ · u = 0 in D,
we enforce the no-penetration boundary conditions
u · n̂ = 0 at ∂D
where n̂ is the outwards pointing unit

normal.
|2 amongst all velocities u with a given value of mean enstrophy

Given this setup, we minimize |∇T
|∇u|2 , or mean kinetic energy |u|2 . These “enstrophy-” and “energy-constrained optimal cooling prob-



lems” were originally posed in [32] and analyzed further in [25]. Actually, those papers focused on the special
case of a constant heat source, e.g., f = 1, and were also mostly concerned with the energy-constrained
problem. Here we treat a general source f (x, t) which may vary both in space and in time, and may even be
negative in some places and at some times allowing for temporary sinks. Before coming to our results, we
pause to discuss other possible objective functionals as well as the practical meaning of “optimizing over u”.
Various minimization problems have been proposed to optimize cooling. In [32], the authors minimize
the mean temperature hT i subject to an energy-constraint and a constant positive source. In [25], the same
1/p
minimization is studied alongside a variety of others involving Lp -based quantities, namely, hT p i and
max T for p = ∞. Multiplying the advection-diffusion equation (1.1) by T and integrating by parts shows
that
κ |∇T |2 = hf T i .



If f is constant, minimizing hT i is the same as minimizing |∇T |2 . More generally, if f is bounded and


uniformly positive, meaning that C ≥ f ≥ c for some fixed c, C ∈ (0, ∞), then
c hT i ≤ κ |∇T |2 ≤ C hT i



so that the minimizations

give comparable results. For general f , such a simple relationship does not hold.
Our choice to minimize |∇T |2 is partially out of mathematical convenience, but also because it provides a
more direct assessment of long-term diffusive transport—the rate-limiting step in any heat exchange.
Turning to our choice to treat the velocity directly as a control, perhaps it is more reasonable from the
viewpoint of applications to think of controlling an applied force f (x, t). One may then seek to optimize
cooling subject

to a constraint on the power used. Or, one might limit the complexity of the force, e.g., by
constraining |∇f |2 . In any case, the velocity u selected by f solves the forced, incompressible Navier-Stokes
equations
(1.2) ∂t u + u · ∇u = −∇p + ν∆u + f in D
where p(x, t) and ν > 0 are the pressure and fluid viscosity. Dotting (1.2) by u, integrating by parts, and
remembering that ∇ · u = 0 yields the fundamental “balance law”
ν |∇u|2 = hf · ui .



1All limits are understood in the limit superior sense to ensure they are well-defined.
2
Thus, a constraint on the mean power hf · ui is the same as a constraint on the mean enstrophy |∇u|2 .



Having found an optimal u, we can simply read off from (1.2) the corresponding optimal f . This is not to
say that every optimal velocity is dynamically stable—in fact, in related problems involving extremal orbits
the opposite situation holds [22, 30]. Nevertheless, to get started, we set aside the momentum equation and
focus solely on optimizing advection-diffusion in what follows.
variables, let κ = R = 1. The value of the mean enstrophy |∇u|2 , or


Switching to non-dimensional

2
the mean energy |u| when we discuss it, will be called the Peclét number P e as either quantity sets the
relative strength of advection to diffusion. (In dimensional variables, P e = U R/κ with U
being the imposed
velocity scale.) In the advective limit where P e → ∞, one expects to be able to drive |∇T |2 → 0 using
a well-chosen sequence of velocities. The question is: at what
optimal rate can this convergence occur?
2

Our main result is a set of upper and lower bounds on min |∇T | subject to the enstrophy constraint
|∇u|2 = P e2 , which identify the optimal cooling rate up to a possible logarithm in P e. Let λ1 be the first

Dirichlet eigenvalue of −∆ on D.
Theorem 1.1. Let f (x, t) satisfy
ˆ τ
1
e−λ1 ((τ −t)∧t) ||f (·, t)||L2 (D) dt = 0

2
|f | + |∇f |2 + |∇∇f |2 < ∞

lim √ and
τ →∞ τ 0
and assume hf i > 0. There exist positive constants C(f ), C ′ (f ), and c(f ) depending only on f such that
1 log4/3 P e
|∇T |2 ≤ C ′


C ≤ min
P e2/3 u(x,t) P e2/3
h|∇u|2 i=P e2
u·n̂=0 at ∂D
whenever P e ≥ c(f ).
Remark 1.1.

The first assumption concerns the possibility that

f grows as t → ∞, and limits that growth
such that |∇T |2 remains finite. Without it, one can make |∇T |2 = ∞ for all u, e.g., with f = a(t)ϕ1 (x)


and where ϕ1 ∈ H01 (D) has −∆ϕ1 = λ1 ϕ

1 . The remaining assumptions on f enter into our bounds on C
′ 2 ′ 2 2 2

and C : we achieve C & hf i and C . |f | + |∇f | + |∇∇f | . For the dependence of c, see Section 2
where we prove the lower bound.
Remark 1.2. The same result holds if the no-penetration boundary conditions u · n̂ = 0 are replaced by the
more restrictive no-slip conditions u = 0 at ∂D. This is because we use no-slip velocities to prove the upper
bound. That the lower bound goes through to the no-slip case is clear.
Let us briefly discuss the strategy of our proof. On the one hand, we must explain why no enstrophy-
constrained velocity can lower |∇T |2 significantly beyond P e−2/3 . At the same time, we must construct

a sequence of admissible velocities to cool within a logarithm of this bound. The challenge is to find a way
of computing heat transfer that at the same time shows how it can be optimized. We follow the approach
of our previous articles on maximizing transport across an externally heated fluid layer [14, 42,
44], the
 key
difference being the presence of the internal source f (x, t). We show it is possible to bound |∇T |2 , and
in fact to compute it exactly in the steady case, by a pair of variational problems. These problems remind
of Dirichlet’s principle for Poisson’s equation, but are non-local due to the mixed character of the operator
u · ∇ − ∆. That there exists a non-local Dirichlet-like principle for the effective

diffusivity of a periodic or
random fluid flow is well-known in homogenization [4, 17]. Our idea is that |∇T |2 should behave like an


effective diffusivity upon optimization. This is very much in the spirit of the general connection between
homogenization and optimal design [27, 28, 29].
The bulk of our work goes into constructing the “branching flows” behind the upper bound in Theorem
1.1. In general, we envision an unsteady, tree-like, multi-scale flow whose features refine from the bulk to
the boundary. To achieve it, we piece together a family of convection roll-like flows, where the number of
rolls in the azimuthal direction θ depends on the radial coordinate r. A similar, albeit steady, branching
flow was used in [14, 44] to prove nearly sharp bounds on optimal transport through a layer. Later on, a
beautiful and fully three-dimensional branching-type flow was found numerically via gradient ascent [34].
Each of these may be regarded as a more refined version of the “multi-α” flows of Busse [6], which do not
enforce the full advection-diffusion equation but rather hinge upon various balance laws it implies.
3
 Branching may be anticipated as a mechanism for optimal heat transfer by asking what it takes to
efficiently guide Brownian particles from the boundary to the bulk. Imagine a cloud of particles is released
at r = 1 and is to be transported inwards, say with√ (non-dimensional) speed ∼ U . In time τ , the particles
diffuse in the θ-direction by a typical distance ∼ τ . In the same time, they advect in the r-direction by a
distance ∼ U · τ . With the goal of not losing too many particles to poorly directed advection, we suggest
using streamlines whose azimuthal scale ℓ(r) matches the width of the Brownian cloud:
r
1−r
(1.3) ℓ(r) ∼ for 0 < r < 1.
U
Two parameters emerge, namely the speed U and the boundary layer width δbl where the streamlines must
finally turn around. By the enstrophy constraint,
ˆ 1−rbl  2
U 1
P e2 = |∇u|2 ∼ rdr ∼ U 3 log


.
0 ℓ(r) δ bl
p
Thinking of an isotropic roll-based boundary layer, we take δbl /U ∼ δbl giving U ∼ 1/δbl. This determines
all parameters via the scalings
P e2/3 log1/3 P e
(1.4) U∼ and δbl ∼ .
log1/3 P e P e2/3
Even if our thought experiment on how advection can effectively “hug” diffusion suggests a particular way
of branching, it is far from a proof

of Theorem 1.1, or even of its upper bound. To achieve it, we apply our
variational bounds to estimate |∇T |2 on a general roll-based branching construction. By optimizing over


all admissible “scale functions” ℓ(r), we discover the very same scalings as in (1.3) and (1.4). In fact, we see
no way out of these and the logarithmically corrected bound they imply: we conjecture that
log4/3 P e
|∇T |2 = C(f ) ·


min + ··· as P e → ∞
u(x,t) P e2/3
h|∇u|2 i=P e2
u·n̂=0 at ∂D

with the dots representing asymptotically negligible terms. This is for two-dimensional cooling; the data in
[34] suggest a pure power law scaling for optimal transport across a three-dimensional layer. Maybe there is
more room for chaperoning particles in three dimensions versus two? In any case, proving the analog of our
conjecture for a fluid layer would establish a new bound on the longstanding problem of Rayleigh-Bénard
convection, as noted in [14]. This bound would state that N u . Ra1/2 / log2 Ra with N u and Ra being the
Nusselt and Rayleigh numbers (the relevant measures of transport and driving there).
Stochastic analysis of optimal cooling is the focus of [25] where the authors estimate, amongst other
things, the mean temperature of steady convection roll-like flows in an internally heated fluid
layer.
 Using
large deviations-based bounds, they prove that min hT i ≤ C(δ) Ploge1−δPe
for any δ > 0, where |u|2 = P e2 .
Showing the conjectured optimal scaling
1
|∇T |2 = C(f ) ·


min + · · · as P e → ∞
u(x,t) Pe
h|u|2 i=P e2
u·n̂=0 at ∂D

put forth for f = 1 in [32] remains an open challenge, the lower bound part of which seems to require a new
idea. The variational bounds on cooling we achieve on the way to Theorem 1.1 are strong enough to show
the upper bound part, namely min hT i ≤ C(f )/P e, and to likewise improve the estimate from [25]. The
same scaling was found using matched asymptotic analysis in [32], but without a bound on the error terms.
See Proposition 5.1 for a full statement and proof of our upper bound on energy-constrained cooling.
This paper is organized as follows. We begin in Section 2 with our variational bounds on cooling. At the
end of that section, we choose a background-like test function to prove the lower bound in Theorem 1.1. The
rest of the paper is devoted to upper bounds. First, we discuss the steady problem in Section 3 to explain
our general strategy for finding designs. Then, in Section 4 we construct our branching flows to prove the
rest of Theorem 1.1. We end in Section 5 with our upper bound on energy-constrained cooling.
4
1.1. Notation. We write X . Y if X ≤ cY for a fixed numerical constant c > 0, i.e., one that is independent
of all parameters. We write X .a Y if X ≤ cY where c = c(a), and X ≪ Y if X Y → 0 in a relevant limit.
We abbreviate X ∧ Y = min {X, Y } and X ∨ Y = max {X, Y }.
We often conflate a point x with its polar coordinates (r, θ). The unit vectors êr = x/|x| and êθ = ê⊥
r ,
where (·)⊥ is a counterclockwise rotation by π/2. Given a function ϕ(x), its average on the disc D is
1
ˆ
ϕ= ϕ(x) dx
D |D| D
where |D| is the disc’s area. By Dr we mean the concentric disc of radius r. Restricting ϕ to ∂Dr gives a
function of θ, whose average and L1 -norm are
ˆ 2π ˆ 2π
1
ϕ(r) = ϕ(r, θ) dθ and ||ϕ||L1θ (r) = |ϕ(r, θ)| dθ.
2π 0 0

Given a function ϕ(x, t), we write

τ τ
1 1
ˆ ˆ
hϕi = lim sup ϕ = lim sup ϕ(x, t) dxdt
τ →∞ 0 D τ →∞ τ |D| 0 D

for its (limit superior) space and long-time average.

The Sobolev spaces L2 (D) and H 1 (D) are defined as usual, using the norms
sˆ sˆ
||ϕ||L2 (D) = |ϕ|2 dx and ||ϕ||H 1 (D) = |ϕ|2 + |∇ϕ|2 dx.
D D

−1
We write H (D) for the dual of H01 (D), 1
the latter indicating H -functions on D with zero trace at ∂D. It
will be convenient to normalize their duality bracket (·, ·) by |D|, and in particular we take

(∇ · m, ϕ) = − m · ∇ϕ
D

for m ∈ L2 (D; R2 ) and ϕ ∈ H01 (D). We use various mixed spaces such as L2 ((0, ∞); H01 (D)) and its local-
in-time version L2loc ((0, ∞); H01 (D)). For definitions, see a text on partial differential equations (PDE) such
as [16].

1.2. Acknowledgements. We thank Charlie Doering for many inspiring discussions, and David Goluskin
for noting the relevance of balanced versus unbalanced heating. This work was supported by National Science
Foundation Award DMS-2025000.

2. Variational bounds on cooling

We begin with a general method for bounding |∇T |2 from above and below. While it is easy to show



using an integration by parts that

|∇T |2 . |f |2 ,



(2.1)
improving this upper bound and finding a corresponding lower bound is not so simple. Incidentally, it follows
from (2.1) and the linearity of the PDE in T that |∇T |2 does not depend on the exact choice of its initial



data, so long as it belongs to L2 (D). To obtain better bounds, we invoke a certain symmetrizing change of
variables that couples T to a second “temperature” field arising from its adjoint PDE. The resulting bounds
are sharp in the steady case where u and f do not depend on time.
Define the admissible set
 
2 1 2 −1 1
A = θ ∈ Lloc ((0, ∞); H0 (D)) : ∂t θ ∈ Lloc ((0, ∞); H (D)), lim sup √ ||θ(·, τ )||L (D) < ∞
2
τ →∞ τ
and let ∆−1 : H −1 (D) → H01 (D) be the inverse Laplacian with zero Dirichlet boundary conditions. Recall
|∇ϕ|2
´
λ1 = min ´D .
ϕ∈H01 (D) D
|ϕ|2
5
Proposition 2.1. Let u(x, t) be weakly divergence-free and have

2
|u| < ∞
and let f (x, t) satisfy
τ
1
ˆ
lim √ e−λ1 ((τ −t)∧t) ||f (·, t)||L2 (D) dt = 0.
τ →∞ τ 0
Any weak solution T (x, t) of (
∂t T + u · ∇T = ∆T + f in D
T =0 at ∂D
with initial data T (·, 0) ∈ L2 (D) must obey the bounds
D 2 E
 D 2 E
2ξf − |∇ξ|2 − ∇∆−1 (∂t + u · ∇)ξ  ≤ |∇T |2 ≤ |∇η|2 + ∇∆−1 [(∂t + u · ∇)η − f ]
 

for all ξ, η ∈ A.
Remark 2.1. The statement that T is a weak solution deserves to be clarified, especially as u · ∇T is at
first glance only in L1 (D), a.e. in time, and so would appear not to belong to H −1 (D) = (H01 (D))′ . (Two
dimensions is critical for Sobolev embedding.) However, u and ∇T are divergence- and curl-free, and this is
enough to compensate for their lack of regularity.
To see why, introduce a stream function ψ ∈ H 1 (D) with u = ∇⊥ ψ, and note that u · ∇T is the Jacobian
determinant of the mapping x 7→ (ψ, T ). Therefore, by the estimate of Coifman, Lions, Meyer, and Semmes
on Jacobian determinants [9] (see also [35]),
ˆ 
 u · ∇T ϕ dx . ||u||L2 (D) ||∇T ||L2 (D) ||∇ϕ||L2 (D) ∀ ϕ ∈ Cc∞ (D).
 
 
D
Letting (·, ·) denote the (normalized) duality bracket between H −1 (D) and H01 (D), the previous inequality
extends to say that
(2.2) |(u · ∇T, ϕ)| . ||u||L2 (D) ||∇T ||L2 (D) ||∇ϕ||L2 (D) ∀ ϕ ∈ H01 (D).
Thus, u · ∇T ∈ H −1 (D) at a.e. time. Clearly, the same holds for ∆T and f .
At this point, the definition of T as a weak solution can be carried out as usual, by requiring that
T ∈ L2loc ((0, ∞); H01 (D)) and ∂t T ∈ L2loc ((0, ∞); H −1 (D))
and enforcing the PDE as an equality on H −1 (D). Existence and uniqueness of weak solutions with initial
data in L2 (D) follows (see, e.g., [16]).
Remark 2.2. A particular consequence of the previous remark that will be used throughout the proof of
Proposition 2.1 is that
(u · ∇T, T ) = 0 for a.e. t
if u and T are as in the proposition. This follows from a smooth approximation argument and the fact that
T (·, t) ∈ H01 (D) a.e. in t. Indeed, if T (·, t) is smooth and compactly supported in D then
1
(u · ∇T, T ) = u · ∇|T |2 dx = 0
2 D
by the divergence theorem. The bilinear form T 7→ (u · ∇T, T ) is continuous on H01 (D), per (2.2). The
claimed identity now follows from the density of smooth and compactly supported functions in H01 (D).
Remark 2.3. In the case that u and f are steady, the bounds from the proposition become sharp and are
achieved by steady test functions η(x) and ξ(x). That is,
2 2 2 2
|∇T |2 = min |∇η| + ∇∆−1 [u · ∇η − f ] dx = max 2ξf − |∇ξ| − ∇∆−1 u · ∇ξ  dx.

  
1
η∈H0 (D) 1 ξ∈H0 (D)
D D
The crucial
point is that, when u and f are steady, the above optimizations in η and ξ not only provide
bounds on |∇T |2 , but also turn out to be “strongly” dual in that their optimal values are the same (an
easy consequence of their Euler-Lagrange equations). See [14] for the full proof of a similar duality arising
for steady heat transport through a fluid layer.
6
Remark 2.4.

On the
 other hand, if f is allowed to depend on time, equality need not hold between our
bounds on |∇T |2 . This is because one cannot rule out the possibility that
τ τ
lim inf |∇T |2 < lim sup |∇T |2
τ →∞ 0 D τ →∞ 0 D
in general while, as will become clear in the proof below, our bounds actually hold on these limit inferior
and limit superior long-time averages (see (2.6) and (2.7)). We do not know if our bounds are sharp when
f (x) is steady and u(x, t) is unsteady, though we guess the answer is “no”.
Proof. Consider the formally adjoint pair of problems
(
(±∂t ± u · ∇)T± = ∆T± + f in D
,
T± = 0 at ∂D

the + version of which is solved by the given temperature T . As the value of |∇T |2 does not depend on



its initial data, we let T (·, 0) = 0 and refer to this version of the temperature as T+ in the rest of this proof.
Fixing an arbitrary time τ > 0 which we will eventually take to ∞, let T− be the unique weak solution of
the − problem with T− (·, τ ) = 0. (The definition of “weak solution” is standard; see Remark 2.1.) A quick
argument using integration by parts and Gronwall’s inequality shows that
ˆ τ
√
||T+ (·, τ )||L2 (D) ∨ ||T− (·, 0)||L2 (D) ≤ e−λ1 ((τ −t)∧t) ||f (·, t)|L2 (D) dt ≪ τ as τ → ∞
0
q
by our hypothesis on f . Briefly, one notes for Z = ǫ + ||T+ ||2L2 (D) that

d
Z + λ1 Z ≤ λ1 ǫ1/2 + ||f ||L2 (D)
dt
for all ǫ > 0. Using the integrating factor e−λ1 (τ −t) and taking ǫ → 0 yields the + version of the inequality.
The − version follows by reversing time.
Having defined T± , we now change variables to the pair
1 1
η= (T+ − T− ) and ξ = (T+ + T− )
2 2
and note they satisfy

(∂t + u · ∇)η = ∆ξ + f
 in D
(2.3) (∂t + u · ∇)ξ = ∆η in D

η=ξ=0 at ∂D


weakly for t ∈ (0, τ ). It follows from our previous bounds on T± that

√
(2.4) ||ξ(·, 0)||L2 (D) + ||ξ(·, τ )||L2 (D) + ||η(·, 0)||L2 (D) + ||η(·, τ )||L2 (D) ≪ τ as τ → ∞.
ffl τ ffl
Introduce the notation h·iτ = 0 D · dxdt for the truncated space- and time-average. Testing the second
PDE in (2.3) against ξ, we get that
τ τ
h∇η · ∇ξiτ = (−∆η, ξ) dt = (−(∂t + u · ∇)ξ, ξ) dt
0 0
τ  
d 1
= − |ξ|2 dx + (u · ∇ξ, ξ) dt
0 dt 2 D
1
= |ξ(x, 0)|2 − |ξ(x, τ )|2 dx
2τ D
where again (·, ·) is the duality bracket normalized by |D|. That u · ∇ξ ∈ H −1 (D) at a.e. time follows from
the assumed mean-square integrability of u and the statement that it is weakly divergence-free; see Remark
2.1 and Remark 2.2 for the proof that (u · ∇ξ, ξ) = 0. Combined with (2.4), the conclusion is that
h∇η · ∇ξiτ → 0 as τ → ∞.
7
Hence, T = T+ = (η + ξ)/2 obeys
|∇T+ |2 τ = |∇η|2 + |∇ξ|2 τ + oτ (1)



(2.5)
= |∇η|2 + |∇∆−1 [(∂t + u · ∇)η − f ] |2 τ + oτ (1)



with oτ (1) denoting a quantity that vanishes as τ → ∞. This way of writing the mean-square temperature
gradient leads to the upper bound from the claim.
To see why, let η̃ ∈ A and consider the difference
D 2 E D 2 E
Aτ := |∇η̃|2 + ∇∆−1 [(∂t + u · ∇)η̃ − f ] − |∇η|2 + ∇∆−1 [(∂t + u · ∇)η − f ]
 
.
τ τ
2
We claim it is non-negative up to a term that vanishes as τ → ∞. Since | · | is convex,
1
Aτ ≥ ∇η · ∇(η̃ − η) + ∇ξ · ∇∆−1 (∂t + u · ∇)(η̃ − η) τ


2  τ 
= ∇η · ∇(η̃ − η) dx − ((∂t + u · ∇)(η̃ − η), ξ) dt
0 D
τ  
1
= (−∆η + (∂t + u · ∇)ξ, η̃ − η) dt − (η̃ − η)ξ |t=τ
t=0 ≥ −oτ (1)
0 τ D
again by (2.4) and the growth condition in our definition of A. Hence,
D 2 E
|∇T+ |2 τ ≤ |∇η̃|2 + ∇∆−1 [(∂t + u · ∇)η̃ − f ]

 
(2.6) + oτ (1)
τ
for all η̃ ∈ A. This bound is actually a bit better than the one from the claim (however, see Remark 2.4).
Returning to the PDEs in (2.3), we now test the second one against η and integrate by parts, again using
the abbreviation oτ (1) for terms that limit to zero as τ → ∞. The result is that
τ τ
|∇η|2 τ =


(−∆η, η) dt = (−(∂t + u · ∇)ξ, η) dt
0 0
τ τ
= ((∂t + u · ∇)η, ξ) dt + oτ (1) = (∆ξ + f, ξ) dt + oτ (1)

0 0
= −|∇ξ|2 + f ξ τ + oτ (1)


where in the second line we applied the first PDE in (2.3). Combined with (2.5), this proves that
|∇T+ |2 τ = hf ξiτ + oτ (1)



and that
|∇T+ |2 τ = 2f ξ − |∇η|2 − |∇ξ|2 τ + oτ (1)




= 2f ξ − |∇ξ|2 − |∇∆−1 (∂t + u · ∇)ξ|2 τ + oτ (1).



We are ready to deduce the lower bound from the claim.

Let ξ˜ ∈ A and call
D E
˜ 2 − |∇∆−1 (∂t + u · ∇)ξ|
Bτ := 2f ξ − |∇ξ|2 − |∇∆−1 (∂t + u · ∇)ξ|2 τ − 2f ξ̃ − |∇ξ| ˜2 .


τ

As before, we use the convexity of | · |2 to write that

1 D E
Bτ ≥ −f (ξ˜ − ξ) + ∇ξ · ∇(ξ˜ − ξ) + ∇η · ∇∆−1 (∂t + u · ∇)(ξ˜ − ξ)
2 τ
τ  
= −f (ξ˜ − ξ) + ∇ξ · ∇(ξ˜ − ξ) dx − ((∂t + u · ∇)(ξ˜ − ξ), η) dt
0 D
τ  
˜ 1 ˜
= (−f − ∆ξ + (∂t + u · ∇)η, ξ − ξ) dt − (ξ − ξ)η |t=τ
t=0 ≥ −oτ (1).
0 τ D
In the last step we applied (2.4) to handle the terms at t = 0 and t = τ . Thus,
D E
|∇T+ |2 τ ≥ 2f ξ̃ − |∇ξ|˜ 2 − |∇∆−1 (∂t + u · ∇)ξ|
˜ 2 − oτ (1)


(2.7)
τ
8
for all ξ˜ ∈ A. Again, the result is a bit better than the lower bound from the claim. 
As a first application of our variational bounds on cooling, we obtain the lower bound from Theorem 1.1
which applies to all velocities with a specified mean enstrophy.
Proof of the lower bound from Theorem 1.1. The lower bound half of Proposition 2.1 applied with a steady
test function ξ ∈ H01 (D) shows that
2 hf ξi ≤ |∇T |2 + |∇ξ|2 + |∇∆−1 ∇ · (uξ)|2 .




Scaling ξ → λξ and optimizing over λ ∈ R, we now write that

2
hf ξi ≤ |∇T |2 |∇ξ|2 + |∇∆−1 ∇ · (uξ)|2 .



(2.8)
This bound holds for all ξ ∈ H01 (D), and we proceed to make a choice.
Given any δ ∈ (0, 1), define
ξδ = χδ (r)
where χδ (r) is a smooth, radial “cutoff function” that goes from zero to one across a small boundary layer.
Precisely, χδ ∈ Cc∞ ([0, 1)) satisfies
0 ≤ χδ (r) ≤ 1 ∀ r ∈ [0, 1) and χδ (r) = 1 ∀ r ≤ 1 − δ
and has
1
||χ′δ ||L∞ ([0,1)) .
δ
with a constant independent of δ. Note that
h|f |2 i h1r>1−δ i
p p
|hf i| ≤ |hf ξδ i| + |hf (ξδ − 1)i| ≤ |hf ξδ i| +
≤ |hf ξδ i| + Cδ.
Hence, there exists δ0 (f ) > 0 such that
1
|hf i| ≤ |hf ξδ i| ∀ δ ∈ (0, δ0 ].
2
Similarly,
1 1
|∇ξδ |2 = |χ′δ (r)|2 . 1r>1−δ .


2
D D δ δ
for all δ.
At this point, we have dealt with each of the terms in the bound (2.8) except for the non-local one
involving the L2 -orthogonal projection ∇∆−1 ∇. For this, observe that
ˆ ˆ
|∇∆−1 ∇ · v|2 dx = max 1
2∇θ · v − |∇θ|2 dx
D θ∈H0 (D) D
2
for all v ∈ L (D). So, the estimate
|∇∆−1 ∇ · (uξδ )|2 . δ 2 |∇u|2



(2.9)
follows from its steady version
ˆ ˆ
∇θ · uξδ . |∇θ|2 + δ 2 |∇u|2 ∀ θ ∈ H01 (D)
D D
which we now establish at a.e. time. Since u is divergence-free and ξδ = χδ (r) depends only on r,
ˆ  ˆ  ˆ 1
θu · êr  |χ′δ | rdr
     
 ∇θ · uξδ  =  θu · ∇ξδ  .
   
D D 1−δ
1  
. θu · êr  dr.
1−δ
ffl 2π
Note · = 0
· dθ gives the θ-average, as explained in Section 1.1. Continuing, we claim that
(2.10) |θu · êr | . (1 − r)||∇θ||L2 (D) ||∇u||L2 (D) for a.e. r ∈ (0, 1).
9
In fact, (2.10) follows by a more or less standard argument involving the Cauchy–Schwarz inequality applied
d 2 d d
to dr θ and dr (u · êr )2 , and then again to dr θu · êr , all of which are derivatives of quantities vanishing at
r = 1. (For the complete details in an analogous fluid layer setting, see [14, Lemma 2.6].) Applying (2.10),
we get that
1   1
θu · êr  dr . (1 − r) dr · ||∇θ||L2 (D) ||∇u||L2 (D) . δ||∇θ||L2 (D) ||∇u||L2 (D)
1−δ 1−δ
ˆ
. |∇θ|2 + δ 2 |∇u|2 .
D

As θ ∈ H01 (D) was arbitrary, the estimate (2.9) is proved.

Summing up, we have shown via (2.8) and the test function ξδ = χδ (r) that
 
2  1
hf i . |∇T |2 + δ 2 |∇u|2



∀ δ ∈ (0, δ0 (f )].
δ

From this, the bound

 
1

2 2 1/3
hf i . |∇T |2


∨ |∇u|
δ0
1/3 −1/3
follows. Indeed when δ0−1 ≤ |∇u|2 the choice δ = |∇u|2

is allowed, otherwise we take δ = δ0 .

2 −3
This proves the lower bound in Theorem 1.1 with C & hf i and c = δ0 . 

The reader familiar with the background method may wonder whether it also leads to the lower bound in
Theorem 1.1. Indeed it does, the key point being that (2.9) verifies the spectral constraint. For more on the
connection between our symmetrization-based bounds and those of the background method, see [14, 42].

3. Optimal steady flows for enstrophy-constrained cooling

The rest of this paper is about upper bounds. In this section and the next, we prove the one from Theorem
1.1 on enstrophy-constrained cooling. First, to warm-up, we study the steady version of the problem:

|∇T |2


(3.1) min
u(x)
|∇u|2 =P e2
ffl
D
u=0 at ∂D

for a given steady source f (x). Then, we return to the unsteady problem in Section 4. Note we use no-slip
boundary conditions from now on. Evidently, this is compatible with our goal of proving an upper bound
on optimal no-penetration flows, as minimizing over no-slip flows can only increase the result.
When u and f are steady, the upper bound from Proposition 2.1 leads with a little effort to the double
minimization
1
(3.2) min |∇∆−1 (u · ∇η − f ) |2 + |∇η|2 |∇u|2 .
u(x),η(x) D P e2 D D
u=0,η=0 at ∂D

Its optimal value is the same as that of the original problem

p (3.1), and its optimizers (u, η) yield solutions
2
to it under the rescaling u → λP e u with λP e = P e/ h|∇u| i. After explaining this in Section 3.1, we go
on in Section 3.2 to show how to achieve

u · ∇η ≈ f in H −1 (D)

with convection roll-based flows. The estimates given there on the non-local part of (3.2) will come in handy
when we discuss branching flows.
10
3.1. A change of variables. First, we show that the problems in (3.1) and (3.2) are the same. Applying
the upper bound from Proposition 2.1—here we use Remark 2.3 as it gives the better result in the steady
case—we see that
|∇T |2 = min F (u, η), where F (u, η) = |∇∆−1 (u · ∇η − f ) |2 + |∇η|2 .


η(x) D D
η=0 at ∂D

Evidently, minimizing |∇T |2 is the same as minimizing F . A simple change of variables removes the



enstrophy constraint. Consider the substitutions

qffl
Pe D
|∇ũ|2
(3.3) u = qffl ũ and η = η̃
|∇ũ|2 Pe
D

where ũ is not allowed to be identically zero. Evidently,

|∇u|2 = P e2 ∀ ũ
D
while
1
F (u, η) = |∇∆−1 (ũ · ∇η̃ − f )|2 + |∇ũ|2 · |∇η̃|2 .
D P e2 D D
Thus, the two minimization problems
1
|∇T |2 |∇∆−1 (ũ · ∇η̃ − f )|2 + |∇ũ|2 · |∇η̃|2


min and min
u(x) ũ(x),η̃(x) D P e2 D D
|∇u|2 =P e2 ũ=0,η̃=0 at ∂D
ffl
D
u=0 at ∂D

are one and the same. Their optimal values agree, and their optimizers are related via (3.3).
A similar change of variables can be done for the unsteady problem, with the result being a bound rather
than an equivalence. Setting p
Pe h|∇ũ|2 i
u= p ũ and η = η̃
2
h|∇ũ| i Pe
into the upper bound from Proposition 2.1 shows that
* +
|∇ũ|2



2
  −1
2 2
|∇T | ≤ min ∇∆ [(∂t + ũ · ∇)η̃ − f ] +
 |∇η̃|
η̃(x,t) P e2
η̃=0 at ∂D

where T (x, t) is a temperature field associated to the unsteady velocity u(x, t). This is the starting point of
Section 4.

3.2. Steady advection. We now return to steady enstrophy-constrained cooling, and ask what it takes
to drive its non-local term to zero. More specifically, given a steady source f (x), what does it take for a
velocity–test function pair (u(x), η(x)) to make
u · ∇η ≈ f in H −1 (D)
where the notation means that D |∇∆−1 (u · ∇η − f ) |2 is small? Guided by the divergence theorem and
ffl
our usual assumption that ∇ · u = 0, we see that any successful pair must also achieve
ˆ ˆ
uη · êr ≈ f
∂U U
for U ⊂ D. The following result makes this intuition precise.
Introduce the notation Dr for the disc of radius r > 0 centered at the origin of D, and again let
ˆ 2π
1
ϕ(r) = ϕ(r, θ) dθ
2π 0
denote the average over its boundary ∂Dr . As noted in Section 1.1, we will allow ourselves to conflate a
point x with its polar coordinates (r, θ). Any integral over ρ is done with respect to the radial coordinate.
11
Lemma 3.1. Let f ∈ L2 (D) and suppose that (u, η) ∈ H 1 (D; R2 )×H 1 (D) where u is divergence-free. Then,
ˆ 1
1
|∇∆−1 (u · ∇η − f )|2 dx = |uη · êr − F |2 rdr + Q(uη − gêr )
D 2π 0
where
r
1 1
ˆ ˆ
F (r) = f (x) dx, g(r, θ) = ρf (ρ, θ) dρ,
2πr Dr r 0
1
Q(v) = min | − ∂θ ϕ + v · êr − v · êr |2 + |∂r ϕ + v · êθ |2 dx.
1
ϕ∈H (D) D r
Remark 3.1. A particular choice of test function we use often below is
ˆ θ
ϕ = r∂θ−1 (v · êr ) = r v · êr (r, θ′ ) − v · êr (r) dθ′ .
0
It sets the first integral in Q to zero, giving the bound

∂r r∂ −1 (v · êr ) + v · êθ 2 dx.



Q(v) ≤ θ
D

Proof. We need the fact that

ˆ ˆ ˆ
(3.4) |∇∆−1 ∇ · v|2 = min
2
|m + v|2 = min
1
|∇⊥ ϕ + v|2
D m∈L (D) D ϕ∈H (D) D
∇·m=0

for any v ∈ L2 (D). A quick proof of it goes as follows: let ζ ∈ H01 (D) satisfy ∆ζ = ∇ · v in D, and note
that ∇ζ is L2 -orthogonal to divergence-free m, including ∇ζ − v. Hence,
ˆ ˆ ˆ ˆ ˆ
2 2 2 2
|m + v| = |m + v − ∇ζ| + |∇ζ| ≥ |∇ζ| = |∇∆−1 ∇ · v|2
D D D D D

and equality holds for m = ∇ζ − v. This proves the first part of (3.4), and the rest of it follows from the
usual representation of a divergence-free vector field m as the perpendicular gradient of a streamfunction ϕ.
Now using (3.4) and the definition of g, which satisfies ∂r (rg) = rf , write that
ˆ ˆ ˆ
|∇∆−1 (u · ∇η − f ) |2 = |∇∆−1 ∇ · (uη − gêr ) |2 = min |m + v|2
D D ∇·m=0 D
where
v = uη − gêr .
In the first step we used that ∇ · (gêr ) = f , which is clear from its expression in polar coordinates. Observe
that a vector field a(r)êr is L2 -orthogonal to any divergence-free m. Indeed, writing
1
m = ∇⊥ ϕ = − ∂θ ϕêr + ∂r ϕêθ
r
where ϕ is 2π-periodic in θ,
1 ˆ 2π 
1
ˆ ˆ
m · a(r)êr = − ∂θ ϕ dθ a(r) rdr = 0.
D 0 0 r
2
This prompts the L -orthogonal decomposition
v = v · êr êr + w
where w is the remainder. By orthogonality,

min |m + v|2 = min |m + w|2 + |v · êr êr |2

∇·m=0 D ∇·m=0 D D
ˆ 1
= Q(uη − gêr ) + |uη · êr − g|2 rdr
0
12
as · yields functions of r alone. Since
2π  ˆ r 
1 1 1
ˆ ˆ
g(r) = ρf dρ dθ = f = F (r)
2π 0 r 0 2πr Dr
the result is proved. 

We proceed to construct pairs (u, η) which are not necessarily admissible for the minimization in (3.2),
but nevertheless do a good´ job at achieving´ u · ∇η ≈ f in H −1 (D) for a given f ∈ L2 (D). We continue to
1 r
use the functions F = 2πr Dr
f and g = 1r 0 ρf from Lemma 3.1. Given n ∈ N, define the streamfunction
√
2
(3.5) ψ(x) = rg(r, θ)Ψ(θ) where Ψ(θ) = cos(nθ)
n
whose velocity is
(3.6) u = ∇⊥ ψ = − (∂θ gΨ + gΨ′ ) êr + rf Ψêθ .
Likewise, define the test function
(3.7) η(x) = −Ψ′ (θ).
Lemma 3.2. Let f ∈ L2 (D). The velocity–test function pair (u, η) given in (3.5)-(3.7) satisfies
1
ˆ ˆ
(3.8) |∇∆−1 (u · ∇η − f )|2 dx . 2 |f |2 + |∇f |2 dx.
D n D
In particular, we have the estimates

uη · êr − F  . 1
ˆ
 
(3.9) |∇f | dx,
n Dr
ˆ 
1 |f |
(3.10) ||∂r (r∂θ−1 (uη · êr − g))||L1θ . + |∇f | dx + r||f ||L1θ + r2 ||∇f ||L1θ ,
n Dr r
r
(3.11) ||uη · êθ ||L1θ . ||f ||L1θ
n
for a.e. r > 0. Furthermore, the streamfunction ψ and velocity u obey
1 r
ˆ r
ρ|f | ρ|∇f | r
ˆ
(3.12) |ψ| . ρ|f | dρ, |u · êr | . + dρ, |u · êθ | . |f |,
n r n n
ˆ r0 0
nρ|f | nρ|∇f | ρ|∇∇f |
(3.13) |∇u| . + + dρ + |f | + r|∇f |
0 r2 r n
and the test function η obeys
n
(3.14) |η| . 1 and |∇η| .
r
for a.e. θ ∈ [0, 2π] and r > 0. The constants implicit in these estimates are independent of all parameters.
Proof. We start at the bottom of the claim and work backwards. The estimates in (3.14) are clear ´ r given the
formula for η. To prove (3.12) and (3.13), we require the following inequalities involving g = 1r 0 ρf :
1 r 1 r
ˆ ˆ ˆ r
(3.15) |g| ≤ ρ|f | dρ, |∂r g| ≤ 2 ρ|f | dρ + |f |, |∂θ g| ≤ ρ|∇f | dρ
r 0 r 0 0
1 r
ˆ ˆ r ˆ r
(3.16) |∂rθ g| ≤ ρ|∇f | dρ + r|∇f |, |∂θθ g| ≤ r ρ|∇∇f | dρ + ρ|∇f | dρ.
r 0 0 0

Now by the definitions of ψ and u,

r 1 r
|ψ| . |g|, |u · êr | . |∂θ g| + |g|, |u · êθ | . |f |
n n n
13
and (3.12) is proved. Next, compute the gradient
∇u = −∇êr (∂θ gΨ + gΨ′ ) − êr ⊗ ∇ (∂θ gΨ + gΨ′ )
+ ∇êθ (rf Ψ) + êθ ⊗ ∇ (rf Ψ)
1
= −êθ ⊗ êθ (∂θ gΨ + gΨ′ ) − êr ⊗ ∇ (∂θ gΨ + gΨ′ )
r
− êr ⊗ êθ f Ψ + êθ ⊗ ∇ (rf Ψ)
where
1
∇ (∂θ gΨ + gΨ′ ) = êr (∂rθ gΨ + ∂r gΨ′ ) + êθ (∂θθ gΨ + 2∂θ gΨ′ + gΨ′′ ) ,
r
∇ (rf Ψ) = êr (f + r∂r f ) Ψ + êθ (∂θ f Ψ + f Ψ′ ) .
Hence,
1 1
|∇u| ≤ |∂θ gΨ + gΨ′ )| + |∂rθ gΨ + ∂r gΨ′ | + |∂θθ gΨ + 2∂θ gΨ′ + gΨ′′ |
r r
+ |f Ψ| + |(f + r∂r f ) Ψ| + |∂θ f Ψ + f Ψ′ |
n
. (|g| + |∂θ g|) + |f | + |∂r g|
 r 
1 1 1 r
+ |∂r f | + |∂θ f | + |∂rθ g| + 2 |∂θθ g| .
r r r n
Combined with (3.15) and (3.16), this yields the desired estimates in (3.13).
It remains to show the first half of the claim. We require the inequalities
1 1
ˆ ˆ
(3.17) ||g||L1θ ≤ |f | dx, ||∂r g||L1θ ≤ 2 |f | dx + ||f ||L1θ ,
r Dr r Dr
1
ˆ ˆ
(3.18) ||∂θ g||L1θ ≤ |∇f | dx, ||∂rθ g||L1θ ≤ |∇f | dx + r||∇f ||L1θ
Dr r Dr
which follow from (3.15) and (3.16) by integration. Going back to the definitions,
uη − gêr = (∂θ gΨ + gΨ′ ) Ψ′ êr − rf ΨΨ′ êθ − gêr
= ∂θ gΨΨ′ + g (Ψ′ )2 − 1 êr − rf ΨΨ′ êθ .

Averaging the êr -component in θ and using that F = g, there follows

  2π 2π
   
′
g (Ψ′ )2 − 1 
    

uη · êr − F  = uη · êr − g  ≤  ∂ gΨΨ
θ +


0 0
1  2π
 
g (Ψ′ )2 − 1 


. ||∂θ g||L1θ + 
n 0

since |Ψ| . 1/n and |Ψ′ | . 1. Introduce the operator ∂θ−1 defined by
ˆ θ
−1
∂θ ϕ(r, θ) = ϕ(r, θ′ ) − ϕ(r) dθ′
0

and observe that (Ψ′ )2 = 1. Hence,

 2π   2π 
′ 2
∂θ g∂θ−1 (Ψ′ )2 

  

 g (Ψ ) − 1 =
 
0 0
1
(3.19) . ||∂θ g||L1θ ||∂θ−1 (Ψ′ )2 ||L∞ . ||∂θ g||L1θ
θ
n
and (3.9) is proved.
Continuing with (3.10), write that
|∂r (r∂θ−1 (uη · êr − g))| ≤ |∂θ−1 (uη · êr − g)| + |r∂r ∂θ−1 (uη · êr − g)|
= |∂θ−1 ∂θ gΨΨ′ + g (Ψ′ )2 − 1 | + r|∂r ∂θ−1 ∂θ gΨΨ′ + g (Ψ′ )2 − 1 |.

 


14
Of course,
1 1
|∂θ−1 [∂θ gΨΨ′ ] | . ||∂θ g||L1θ and |∂r ∂θ−1 [gΨΨ′ ] | . ||∂rθ g||L1θ .
n n
Using the operator ∂θ−1 again, we see that
ˆ  ˆ
 −1   θ   2π 
′ 2 ′ 2 ′ 2




∂ g (Ψ ) − 1 ≤ 
 
g (Ψ ) − 1  + 
 
g (Ψ ) − 1  ,
θ  0  0
ˆ  ˆ

  θ
  2π

∂r ∂ −1 g (Ψ′ )2 − 1  ≤  ′ 2 ′ 2
 

θ ∂r g (Ψ ) − 1  +  ∂r g (Ψ ) − 1 .
 0   0 

The last two terms on the righthand side are controlled by the inequality (3.19) and one just like it with ∂r g
in place of g. To deal with the first two terms on the righthand side above, we require the inequalities
ˆ
θ
 1
′ 2
(3.20) g (Ψ ) − 1 . ||g||L1θ + ||∂θ g||L1θ ,

n

0 1
Lθ
ˆ
θ
 1
′ 2
(3.21) ∂r g (Ψ ) − 1 . ||∂r g||L1θ + ||∂rθ g||L1θ .

n

0 1
Lθ

To prove them, divide [0, 2π) into the disjoint intervals Ij = [2πj/n, 2π(j + 1)/n) indexed by j = 0, . . . , n − 1,
and note that (Ψ′ )2 − 1 is 2π
n -periodic and averages to zero over each Ij . Choosing k such that θ ∈ Ik , write
that
ˆ  k−1 ˆ !  ˆ 
 θ
 X 
  θ

′ 2 ′ 2 ′ 2
g (Ψ ) − 1  ≤ g− g (Ψ ) − 1  +  g (Ψ ) − 1 
 
 
 0   Ij Ij   2πk/n 
j=0
 
k−1 ˆ   ˆ θ
X 1
. g − g + |g| . ||∂θ g||L1θ + ||g||L1θ (Ik ) .
 
n

j=0 Ij Ij 2πk/n
 

Remembering that k = k(θ) and integrating this bound over θ ∈ [0, 2π) yields (3.20). The proof of the
inequality from (3.21) is much the same. Altogether, we have shown that
−1   1
∂θ gΨΨ′ + g (Ψ′ )2 − 1 L1 . ||g||L1θ + ||∂θ g||L1θ


(3.22) ∂
θ ,
θ n
∂r ∂ −1 ∂θ gΨΨ′ + g (Ψ′ )2 − 1 1 . ||∂r g||L1 + ||∂rθ g||L1 1

  
(3.23) θ Lθ θ θ n
which when combined with the inequalities in (3.17) and (3.18) lead to the estimate (3.10) in the claim. The
estimate (3.11) follows from what we have already proved (namely, (3.12) and (3.14)).
Finally, we prove (3.8). By Lemma 3.1 and the remark appearing immediately after,
ˆ ˆ 1 ˆ ˆ
−1 2 2 −1

2
|∇∆ (u · ∇η − f )| . |uη · êr − F | rdr + |∂r r∂θ (uη · êr − g) | + |uη · êθ |2 .
D 0 D D
Applying (3.9)-(3.11) along with Jensen’s inequality gives the bound
ˆ 1 ˆ 1 2 ˆ 
r 1
ˆ
|uη · êr − F |2 rdr . 2
|∇f | 2
rdr . 2
|∇f |2
0 0 n Dr n D
as well as the bounds
1 ˆ 
1
ˆ ˆ ˆ
|∂r r∂θ−1 (uη · êr − g) |2 . |f |2 + r2 |∇f |2 + r2 ||f ||2L2 + r4 ||∇f ||2L2


rdr
D 0 n2 Dr Dr
θ θ

1
ˆ
. |f |2 + |∇f |2
n2 D
and
1
r2
 
1
ˆ ˆ ˆ
|uη · êθ |2 . ||f ||2L2 rdr . |f |2 .
D 0 n2 θ n2 D
15
The proof is complete. 
As a quick application of this last result, we show that the velocities in (3.6) have finite enstrophy if f is
sufficiently regular. By the pointwise estimate from (3.13) and Jensen’s inequality,
ˆ r 2 2
n 2 ρ2 ρ2 r

n ρ
|∇u|2 . |f | 2
+ |∇f | 2
+ |∇∇f | 2
dρ
0 r3 r n2
+ |f |2 + r2 |∇f |2 .
So,
2π
n 2 ρ2 ρ2 r 2
ˆ ˆ ˆ  
2 2 2 2 2 2
|∇u| dx . |f (ρ, θ)| + n ρ |∇f (ρ, θ)| + 2 |∇∇f (ρ, θ)| drdρdθ
D 0 0≤ρ≤r≤1 r2 n
ˆ
+ |f |2 + r2 |∇f |2 dx
D
r
ˆ
. n2 |f |2 + n2 r|∇f |2 + |∇∇f |2 dx.
D n2
4. Unsteady branching flows for enstrophy-constrained cooling
Section 3 considered the steady optimal cooling problem, and explained how to find “approximate H −1 -
solutions” to the corresponding advection equation u · ∇η = f . We now return to the original unsteady
setting of Theorem 1.1 and prove our upper bound on
|∇T |2


min
u(x,t)
h|∇u|2 i=P e2
u=0 at ∂D

for a general source f (x, t) as in the theorem. To do this, we construct a family of well-chosen branching
flows {uP e } whose temperatures {TP e } satisfy
 log4/3 P e
|∇TP e |2 . |f |2 + |∇f |2 + |∇∇f |2 |∇uP e |2 = P e2 .




with
P e2/3
Section 4.1 starts by defining a general family of branching flows based on convection rolls. Section 4.2
estimates their cooling and Section 4.3 optimizes over their free parameters. The upper bound from Theorem
1.1 is finally proved at the end of this section.
Picking up where we left off in Section 3, recall the upper bound
* +
|∇u|2


2 −1 2 2


min |∇T | ≤ min |∇∆ [(∂t + u · ∇)η − f ] | + |∇η|
u(x,t) u(x,t),η(x,t) P e2
h|∇u| i=P e
2 2 u=0,η=0 at ∂D
u=0 at ∂D
where on the righthand side the magnitude of u is unconstrained. p This bound follows from Proposition 2.1
and the change of variables (u, η) → (λP e u, λ−1 2
P e η) with λP e = P e/ h|∇u| i, as explained in Section 3.1. A
special case occurs for a steady test function η(x): the temperature field T associated to λP e u obeys
1

|∇T |2 ≤ |∇∆−1 [u · ∇η − f ] |2 + |∇u|2 |∇η|2


 
Pe 2
D
for all η ∈ H01 (D). We proceed to define our branching flows.
4.1. Branching flows. By a branching flow u(x, t) and a corresponding (steady) test function η(x), we
mean a divergence-free velocity field
Xn
u = ∇⊥ ψ where ψ(x, t) = χk (r)ψk (x, t)
k=1
and the scalar function
n
X
η(x) = χk (r)ηk (x)
k=1
16
where {χk }, {ψk }, and {ηk } are as follows. Let
r
1
ˆ
(4.1) F (r, t) = f (x, t) dx and g(r, θ, t) = ρf (ρ, θ, t) dρ
2πr Dr 0
and set
√ θ
ψk = rgΨk (θ), ηk = −Ψ′k (θ), and Ψk (θ) =
2lk cos( ) for k = 1, . . . , n.
lk
The parameters {lk−1 } ⊂ N and n ∈ N are free, and will eventually be optimized when it comes time to prove
Theorem 1.1. The functions {χk } are described in the paragraph after the next.
To help organize the discussion, we always assume that
(4.2) l1 > l2 > · · · > ln .
We label the largest and smallest scales as
lbulk = l1 and lbl = ln
noting that they occur in the bulk of the disc and in a boundary layer near r = 1, respectively. The individual
velocities
uk = ∇⊥ ψk = − (∂θ gΨk + gΨ′k ) êr + rf Ψk êθ
are simply unsteady versions of the ones occurring in our prior discussion of roll-like flows, and so are
governed by the estimates in Lemma 3.2. The individual test functions
ηk = −Ψ′k
are also like those in the lemma.
The new ingredients are the functions {χk } which we use to interpolate between the individual building
blocks listed above. We use a family of smooth and compactly supported “cutoff functions” defined via a
choice of points {rk } satisfying
1
(4.3) < r1 < r2 < · · · < rn < 1,
2
with
r1 = rbulk and rn = rbl .
These functions can be quite general, but to fix ideas we let
(4.4) supp χ1 ⊂ (0, r2 ), supp χn ⊂ (rn−1 , 1), and supp χk ⊂ (rk−1 , rk+1 ) for k = 2, . . . , n − 1.
We also let
n
X
(4.5) (χk (r))2 = 1 ∀ r ∈ (0, rbl )
k=1
and take
(4.6) χk χk′ 6= 0 ⇐⇒ |k − k ′ | ≤ 1.
This last condition simplifies the calculation of products such as uη.
To specify our cutoff functions further, introduce the lengths
δbulk = 1 − rbulk , δbl = 1 − rbl , and δk = rk+1 − rk for k = 1, . . . , n − 1
and, following the pattern, call
δn = δbl .
Let
1 1
(4.7) |χk | ∨ |χk+1 | . 1, |χ′k | ∨ |χ′k+1 | . , |χ′′k | ∨ |χ′′k+1 | . ∀ r ∈ (rk , rk+1 )
δk δk2
for k = 1, . . . , n − 1, and let
1 1
(4.8) |χn | . 1, |χ′n | . , |χ′′n | . 2 ∀ r ∈ (rbl , 1).
δbl δbl
The constants implicit in these hypotheses are independent of all parameters.
17
 The reader looking for specific cutoff functions should consult [14, Sect. 5.1]. There, we describe a similar
branching flow in a fluid layer, the z-coordinate of which is analogous to r. The exact choice of these functions
does not affect the scaling of our bounds with respect to P e, but does affect their prefactors.

4.2. Upper bounds on branching flows. Having defined our branching flows, we now estimate their
cooling using the results of Section 2. We write our estimate in terms of a continuously varying version of
the parameters {lk } and {rk }, given by
rk+1 − r r − rk
(4.9) ℓ(r) = lk + lk+1 for r ∈ [rk , rk+1 ]
rk+1 − rk rk+1 − rk
and k = 1, . . . , n − 1. By construction, ℓ(rk ) = lk for each k. Recall δk = rk+1 − rk .
Proposition 4.1. Let {u} be a family of branching flows as defined in Section 4.1, whose parameters {lk }nk=1
and {rk }nk=1 obey
(4.10) |lk+1 − lk | ∼ lk+1 ∼ lk and δk+1 ∼ δk for k = 1, . . . , n − 1
(4.11) l k . δk for k = 1, . . . , n
with fixed numerical prefactors. Define the rescaled velocities
Pe
λP e u with λP e = p
h|∇u|2 i
and let their temperature fields TP e solve
(
∂t T + λP e u · ∇T = ∆T + f in D
T =0 at ∂D

weakly with arbitrary L2 -initial data. The estimate

" ˆ rbl  ˆ rbl 2 #

2
 2 ′ 2 1 1 1 δbl
|∇TP e | . C0 (f ) lbulk + (ℓ (r)) dr + δbl + 2 + dr + 2
rbulk P e2 lbulk rbulk (ℓ(r))
2 lbl

holds with
C0 (f ) = |f |2 + |∇f |2 + |∇∇f |2



and a numerical prefactor depending only on the ones from (4.10) and (4.11).
By Proposition 2.1, we already know that the rescaled velocities λP e u achieve
1

|∇TP e |2 ≤ |∇∆−1 [u · ∇η − f ] |2 + |∇u|2 |∇η|2 ∀ η ∈ H01 (D).


 
Pe 2
D

To prove Proposition 4.1, we now plug in the test functions η from Section 4.1, and estimate the ensuing
mess of terms. We

handle the “advection term” involving u · ∇η − f in Section 4.2.1 and the “enstrophy
term” involving |∇u|2 in Section 4.2.2.


4.2.1. The advection term. Averaging the result of Lemma 3.1 shows that
|∇∆−1 (u · ∇η − f ) |2 ≤ |u · êr η − F |2 + hQ(uη − gêr )i




where for the reader’s convenience we repeat the definition of the quadratic form:
1
(4.12) Q(v) = min | − ∂θ ϕ + v · êr − v · êr |2 + |∂r ϕ + v · êθ |2 dx.
1
ϕ∈H (D) D r
We bound the first average in Lemma 4.1 and the second one in Lemma 4.2.
Lemma 4.1. Every branching flow–test function pair (u, η) defined in Section 4.1 satisfies
2
|u · êr η − F |2 . lbulk |∇f |2 + δbl |f |2 .





18
Proof. It suffices to prove the steady analog of the result at a.e. time, i.e.,

|u · êr η − F |2 . lbulk
2
|∇f |2 + δbl |f |2 .
D D D

From the definitions of u and η and our assumption (4.6),

X
uη · êr = χk χk′ uk ηk′ · êr
|k−k′ |≤1

so that  
 n
X  X n X
2 
χ2k uk ηk · êr − F  +
   
uη · êr − χk F  ≤ |χk χk′ | uk ηk′ · êr  .

 
k=1 k=1 |k−k′ |≤1
k6=k′

The first term is handled by (3.9), which shows that

ˆ
 
uk ηk · êr − F  . lk |∇f |.
Dr

To control the second term, write using the L2 -orthogonality of Ψ′k (θ) and Ψ′k′ (θ) that
  2π
 
′ ′ ′ 
 
uk ηk′ · êr  =  ∂ θ gΨ Ψ
k k′ + gΨ Ψ
k k′ 

0
 2π   2π 
′  −1 ′ ′ 
   
≤ ∂θ gΨk Ψk′  +   ∂θ g∂θ (Ψk Ψk′ )
0 0
ˆ
. ||∂θ g||L1θ (lk ∨ lk′ ) . (lk ∨ lk′ ) |∇f |
Dr

where in the last step we used (3.18). Adding up,

 n
 ˆ
 X  X
2 
uη · êr − χk F  . |χk χk′ |(lk ∨ lk′ ) |∇f |

 
k=1′ |k−k |≤1 Dr

for r > 0.
Continuing, we have by the triangle and Jensen’s inequality that
ˆ 1  2
ˆ ˆ 1 X n Xn 
|uη · êr − F |2 dx . |uη · êr − χ2k F |2 rdr + 1 − χ2k  |F |2 rdr
 
D 0 k=1 0 
k=1

ˆ 1  n
2
X ˆ 1 ˆ X 
2 2 3 2
. |χk χk′ | (lk ∨ lk′ ) r |∇f | + 1 − χk  |F |2 r
2

|k−k′ |≤1 0 Dr 0  
k=1
ˆ 1
2
. lbulk |∇f |2 + δbl |F |2 r.
D rbl
1
´
Recalling the definition of F = 2πr Dr f from (4.1), we use Jensen’s inequality once more to get that
1 1 ˆ  ˆ
2 2
|F | r . |f | r . |f |2 .
rbl rbl Dr D

The result follows. 

Lemma 4.2. Under the assumptions in (4.10), the branching flow–test function pairs (u, η) from Section
4.1 satisfy
 ˆ rbl 2

2 ′ 2 lbl
2
|f | + |∇f |2 + δbl |f |2


hQ(uη − gêr )i . lbulk + (ℓ (r)) dr +
rbulk δbl
with a constant depending only on those in (4.10).
19
Proof. Again we argue a.e. in time. Begin with the bound
n n
ˆ !
X X
Q(uη − gêr ) . Q(uη − χ2k gêr ) + | χ2k − 1 g|2
k=1 D k=1
Xn ˆ
. Q(uη − χ2k gêr ) + δbl |f |2
k=1 D

where in the first stepffl we applied the definition (4.12) of the quadratic form Q, and in the second step we
r
used the formula g = 0 ρf from (4.1) along with Jensen’s inequality and (4.5). Calling
n
X
v = uη − χ2k gêr
k=1

and using
ˆ θ
ϕ = r∂θ−1 (v · êr ) = r v · êr (r, θ′ ) − v · êr (r) dθ′
0

as in Remark 3.1, we deduce that

|∂r r∂θ−1 (v · êr ) + v · êθ |2

Q(v) ≤
D

. |∂θ−1 (v · êr )|2 + r2 |∂r ∂θ−1 (v · êr )|2 + |v · êθ |2 .

D D D

We estimate these integrals one-by-one.

For the first, note using the definitions from Section 4.1 and in particular (4.6) that
n
X X
∂θ−1 (v · êr ) = ∂θ−1 (uη · êr − χ2k g) = χk χk′ ∂θ−1 [uk ηk′ · êr − δkk′ g]
k=1 |k−k′ |≤1
X
= χk χk′ ∂θ−1 [∂θ gΨk Ψ′k′ + g (Ψ′k Ψ′k′ − δkk′ )]
|k−k′ |≤1

where δkk′ is the Kronecker delta function (one if k = k ′ , zero otherwise). Arguing as in the proof of Lemma
3.2—see the paragraph leading up to (3.22)—we write that
 
(4.13) ||∂θ−1 [∂θ gΨk Ψ′k′ + g (Ψ′k Ψ′k′ − δkk′ )] ||L1θ . (lk ∨ lk′ ) ||g||L1θ + ||∂θ g||L1θ
 ˆ 
1
ˆ
. (lk ∨ lk′ ) |f | + |∇f |
r Dr Dr

where in the last step we used (3.17) and (3.18). Squaring and integrating, there follows
ˆ X ˆ 1 ˆ ˆ 
|∂θ−1 (v · êr )|2 . |χk χk′ |2 (lk ∨ lk′ )2 |f |2 + r2 |∇f |2 rdr
D |k−k′ |≤1 0 Dr Dr
ˆ
2
. lbulk |f |2 + |∇f |2 .
D

Continuing with the second integral, write using the product rule that
X
∂r ∂θ−1 (v · êr ) = (χk χk′ )′ ∂θ−1 [∂θ gΨk Ψ′k′ + g (Ψ′k Ψ′k′ − δkk′ )]
|k−k′ |≤1
X
+ χk χk′ ∂r ∂θ−1 [∂θ gΨk Ψ′k′ + g (Ψ′k Ψ′k′ − δkk′ )]
|k−k′ |≤1

20
and note the estimate
(4.14)
 
||∂r ∂θ−1 [∂θ gΨk Ψ′k′ + g (Ψ′k Ψ′k′ − δkk′ )] ||L1θ . (lk ∨ lk′ ) ||∂r g||L1θ + ||∂rθ g||L1θ
 ˆ 
1 1
ˆ
. (lk ∨ lk′ ) 2 |f | + ||f ||L1θ + |∇f | + r||∇f ||L1θ
r Dr r Dr
holds in addition to (4.13). Its proof is essentially the same, and again we point the reader to the paragraph
leading up to (3.23) for the details. Using both (4.13) and (4.14), we get that
ˆ X ˆ 1  ˆ ˆ 
2 −1 2 ′ 2 2 2 2 4 2
r |∂r ∂θ (v · êr )| . |(χk χk′ ) | (lk ∨ lk′ ) r |f | + r |∇f | rdr
D |k−k′ |≤1 0 Dr Dr

X ˆ 1 ˆ ˆ 
2 2 2 2 2 2
+ |χk χ | (lk ∨ l )
k′ k′ |f | + r |∇f | + r ||f ||2L2 +r 4
||∇f ||2L2 rdr
θ θ
|k−k′ |≤1 0 Dr Dr
ˆ rbl 2
ˆ
′ 2 lbl 2
. (ℓ (r)) dr + + lbulk |f |2 + |∇f |2 .
rbulk δbl D
Note we used the definition of ℓ and our hypotheses that |lk+1 − lk | ∼ lk+1 ∼ lk and δk+1 ∼ δk to bring in ℓ′ .
Finally, since
X
v · êθ = uη · êθ = (χ′k ψk + χk uk · êθ )χk′ ηk′
|k−k′ |≤1
X
=− χ′k χk′ rgΨk Ψ′k′ + χk χk′ rf Ψk Ψ′k′
|k−k′ |≤1

we get by a completely analogous argument that

ˆ X ˆ 1 
|v · êθ |2 . |χ′k χk′ |2 lk2 r2 ||g||2L2 + |χk χk′ |2 lk2 r2 ||f ||2L2 rdr
θ θ
D |k−k′ |≤1 0
ˆ rbl 2
ˆ
lbl
. (ℓ′ (r))2 dr +
2
+ lbulk |f |2 .
rbulk δbl D
ffl r
In the last step, we applied Jensen’s inequality with g = 0 ρf . Adding up the estimates and averaging in
time proves the result. 
Combining Lemma 4.1 and Lemma 4.2 with our hypothesis that lbl . δbl from (4.11) proves the first part
of the estimate in Proposition 4.1.
4.2.2. The enstrophy term. We now estimate the gradients of u and η.
Lemma 4.3. Under the assumptions in (4.10) and (4.11), the branching flow–test function pairs (u, η) from
Section 4.1 satisfy
 ˆ rbl 
2 1 1 δbl
2
|f | + |∇f |2 + |∇∇f |2 ,

 
|∇u| . 2 + 2
dr + 2
lbulk r (ℓ(r)) l bl
ˆ rblbulk
1 1 δ bl
|∇η|2 . 2 +


2
dr + 2 .
lbulk rbulk (ℓ(r)) lbl
The constants implicit in these estimates depend only on those in (4.10) and (4.11).
Proof. We start with the formulas
Xn
u= χk uk + χ′k ψk êθ ,
k=1
Xn
∇u = χk ∇uk + χ′k (uk ⊗ êr − êθ ⊗ u⊥ ′′
k ) + χk ψk êθ ⊗ êr ,
k=1
21
which follow from the definitions in Section 4.1 and the fact that u⊥ ⊥
k = −∇ψk as (·) is a counterclockwise
rotation by π/2. Looking back at the estimates in (3.12) and (3.13), we see that
n
X
|∇u| . |χk ||∇uk | + |χ′k ||uk | + |χ′′k ||ψk |
k=1
n  
X A1 A2
. |χk | + + rlk A3 + |f | + r|∇f |
rlk lk
k=1
+ |χ′k | (A1 + rlk A2 + rlk |f |) + |χ′′k |rlk A1
where
1 r 1 r 1 r
ˆ ˆ ˆ
A1 = ρ|f | dρ, A2 = ρ|∇f | dρ, A3 = ρ|∇∇f | dρ.
r 0 r 0 r 0
Squaring and integrating, there follows
n ˆ 1  
1 1
ˆ X
2 2 2 2 2 2 2 2 2 2
|∇u| . |χk | ||A1 ||L2 + 2 ||A2 ||L2 + r lk ||A3 ||L2 + ||f ||L2 + r ||∇f ||L2 rdr
D r2 lk2 θ lk θ θ θ θ
k=1 0
Xn ˆ 1   X n ˆ 1
+ |χ′k |2 ||A1 ||2L2 + r2 lk2 ||A2 ||2L2 + r2 lk2 ||f ||2L2 rdr + |χ′′k |2 r2 lk2 ||A1 ||2L2 rdr
θ θ θ θ
k=1 0 k=1 0

:= I1 + I2 + I3 .
We bound these three sums in turn, using the estimates
r ˆ
(4.15) ||A1 ||2L2 ≤ ρ 2
||f ||2L2 dρ . |f |2 dx,
θ θ
0 Dr
r ˆ
(4.16) ||A2 ||2L2 ≤ ρ 2
||∇f ||2L2 dρ . |∇f |2 dx,
θ θ
0 Dr
r ˆ
(4.17) ||A3 ||2L2 ≤ ρ2 ||∇∇f ||2L2 dρ . |∇∇f |2 dx.
θ θ
0 Dr
Note these follow from the definitions via Jensen’s inequality.
The first sum I1 involves the cutoff functions {χk } directly. Focusing on r ∈ (rbulk , rbl ) for now, which
are always larger than 1/2 and no larger than 1, we estimate
n ˆ rbl  
X
2 1 2 1 2 2 2 2
|χk | ||A1 ||L2 + 2 ||A2 ||L2 + r lk ||A3 ||L2 rdr
r2 lk2 θ lk θ θ
k=1 rbulk
n ˆ rbl
X |χk |2  2 2 2

. ||A ||
1 L 2 + ||A ||
2 L 2 + ||A ||
3 L 2
lk2 θ θ θ
k=1 rbulk
ˆ rbl
1
ˆ
. 2
dr · |f |2 + |∇f |2 + |∇∇f |2
rbulk (ℓ(r)) D

by our definition of ℓ and the hypothesis that lk+1 ∼ lk . For r ∈ (0, rbulk ), which belong only to the support
of χ1 , we use Fubini’s theorem and the first parts of (4.15)-(4.17) to write that
ˆ rbulk  
1 1
|χ1 |2 ||A ||
1 L2
2
+ ||A ||
2 L2
2
+ r 2 2
l 1 ||A ||
3 L2
2
rdr
0 r2 l12 θ l12 θ θ

ρ2 ρ2 2
ρ2
 
lbulk
ˆ
2 2 2
. 2 2
||f (ρ, ·)||L2 + 2 ||∇f (ρ, ·)||L2 + ||∇∇f (ρ, ·)||L2 dρdr
0≤ρ≤r≤rbulk r lbulk
θ lbulk θ r2 θ

1
ˆ
. 2 |f |2 + |∇f |2 + |∇∇f |2
lbulk D
similarly to what we did at the very end of Section 3. Finally,
ˆ 1  
1 1 δbl
ˆ
2 2 2 2 2 2
|χn | ||A1 || Lθ2 + ||A2 || Lθ2 + r l n ||A3 || Lθ2 rdr . 2 |f |2 + |∇f |2 + |∇∇f |2 .
rbl r2 ln2 ln2 lbl D
22
The bound
n ˆ
X 1   ˆ
|χk |2 ||f ||2L2 + r2 ||∇f ||2L2 rdr . |f |2 + |∇f |2
θ θ
k=1 0 D

is clear. Altogether,
 rbl  ˆ
1 1 δbl
ˆ
I1 . 2 + 2
dr + 2 · |f |2 + |∇f |2 + |∇∇f |2 .
lbulk rbulk (ℓ(r)) lbl D

The second sum I2 involves the derivatives {χ′k }.

This time, we only need to handle r > rbulk . Using the
second parts of (4.15) and (4.16), we get that
Xn ˆ 1  
I2 = |χ′k |2 ||A1 ||2L2 + r2 lk2 ||A2 ||2L2 + r2 lk2 ||f ||2L2 rdr
θ θ θ
k=1 rbulk
n−1
! ˆ
X ˆ rk+1 1
ˆ 1
1
. + 2 · |f |2 + |∇f |2
r k
(δk ∧ δk+1 )2 r bl
δbl D
k=1
ˆ rbl  ˆ
1 δbl
. 2
dr + 2 · |f |2 + |∇f |2 .
rbulk (ℓ(r)) lbl D

The bounds
1 1 1 1
. . . ∀ r ∈ (rk , rk+1 )
δk ∧ δk+1 δk+1 lk+1 ℓ(r)
were used to pass between the second and third lines in the estimate above. They hold by our assumptions
that δk ∼ δk+1 , lk . δk , and lk ∼ lk+1 along with the definition of ℓ.
Finally, we estimate I3 which involves the second derivatives {χ′′k }. Again we use (4.15) to write that
Xn ˆ 1
I3 = |χ′′k |2 r2 lk2 ||A1 ||2L2 rdr
θ
k=1 rbulk
n−1
X ˆ rk+1 ˆ 1 2 ! ˆ
(lk ∨ lk+1 )2 lbl
. 4
+ 4 · |f |2
(δ k ∧ δ k+1 ) δ
k=1 rk rbl bl D
ˆ rbl  ˆ
1 δbl
. 2
dr + 2 · |f |2 .
rbulk (ℓ(r)) l bl D

Adding up the estimates on I1 , I2 , and I3 and averaging in time gives the first part of the claim.
A similar, and much simpler, argument proves the desired bound on ∇η. In particular, by the definition
of η in Section 4.1 and the estimates from (3.14), there holds
n n
X X 1
|∇η| ≤ |χk ||∇ηk | + |χ′k ||ηk | . |χk | + |χ′k |.
rlk
k=1 k=1

Hence
n ˆ 1  
1
ˆ X
2
|∇η| . 2
|χk | 2 2 + |χ′k |2 rdr.
D 0 r lk
k=1

We recognize terms similar to those from our proof of the bound on ∇u, with the difference being that the
terms involving f are now replaced by the number one. The same manipulations apply as before. 

Combining Lemma 4.3 with what we proved in Section 4.2.1 yields the rest of Proposition 4.1.

4.3. Optimal branching flows. Finally, we optimize over our branching flows to prove the upper bound
in Theorem 1.1. Recall from Proposition 4.1 that the (rescaled) velocities λP e u achieve
|∇T |2 . C0 (f ) · M ({lk }nk=1 , {rk }nk=1 , n; P e) and |∇(λP e u)|2 = P e2



(4.18)
23
where
C0 = |f |2 + |∇f |2 + |∇∇f |2 ,


ˆ rbl  ˆ rbl 2
2 1 1 1 δbl
M = lbulk + (ℓ′ (r))2 dr + δbl + 2 + dr + 2 .
rbulk P e2 lbulk rbulk (ℓ(r))
2 lbl
We find it useful to work directly with the continuous “scale function” ℓ(r) introduced in (4.9), and to wait
to enforce the interpolation rule
(4.19) lk = ℓ(rk ) for k = 1, . . . , n
until it comes time to select the parameters {lk }nk=1
and {rk }nk=1 giving the desired branching flows.
Recall l1 = lbulk , ln = lbl , r1 = rbulk , and rn = rbl , and consider the one-dimensional variational problem
ˆ rbl ˆ rbl 2
′ 2 1 1
min (ℓ ) dr + dr
ℓ(r) rbulk P e2 rbulk ℓ
2
ℓ(rbulk )=lbulk
ℓ(rbl )=lbl

suggested by minimizing M . Its solution obeys

ˆ rbl 
′ 2 1 1 1
(ℓ ) ∼ dr .
P e2 rbulk ℓ2 ℓ2
Setting ℓ(1) = 0 and integrating yields
rbl 1/4
√
ˆ
1 1
(4.20) ℓ(r) ∼ dr 1 − r.
P e1/2 rbulk ℓ2
Put another way, √
ℓ(r) = c(P e) 1 − r.
The constant c(P e) and the parameters rbulk , rbl , lbulk , and lbl must be determined. For c, note that
ˆ rbl
1 rbl dr
 
1 1 δbulk
ˆ
2
dr = = log
rbulk ℓ c2 rbulk 1 − r c2 δbl
where δbulk = 1 − rbulk and δbl = 1 − rbl . Setting this into (4.20) and rearranging, we get that
 
1 1/6 δbulk
c(P e) ∼ log .
P e1/3 δbl
By (4.19),
1/2 1/2
δbulk
 
δbl
 
δbulk δbulk
lbulk = ℓ(rbulk ) ∼ log1/6 and lbl = ℓ(rbl ) ∼ log1/6 .
P e1/3 δbl P e1/3 δbl
All that remains is to choose δbulk and δbl . Note the quantity M from (4.18) satisfies
2
 
2 1 1 1 δbl 1 4/3 δbulk
M . lbulk + 4 + δbl + 4 + log
P e2 lbulk P e2 lbl P e2/3 δbl
 
 
1  1 4/3 δbulk 
. δbl + + log .
P e2/3 δ 2 log2/3 δbulk
 
δbl
bulk δbl

Minimizing with
log1/3 P e
δbulk ∼ 1 and δbl ∼
P e2/3
yields the desired bound
log4/3 P e
|∇TP e |2 ≤ C ′ (f ) ·


(4.21)
P e2/3
where C ′ . |f |2 + |∇f |2 + |∇∇f |2 . Up to prefactors, this is the best upper bound achievable by a roll-



based branching flow. To complete the proof of Theorem 1.1, we only need to verify that our choices for
{lk }nk=1 , {rk }nk=1 , and n are actually admissible in our analysis of branching flows.
24
Proof of the upper bound from Theorem 1.1. To be absolutely clear, we now fix
log1/6 P e √ 1
(4.22) ℓ(r) = 1 − r, r ∈ ( , 1)
P e1/3 2
and let n ∈ N satisfy
P e1/3
n ≤ log2 ≤ n + 1.
log1/6 P e
Implicit in this is the requirement that 2 ≤ P e1/3 / log1/6 P e. Define the scales {lk }nk=1 by taking
lbulk
lk = for k = 1, . . . , n
2k−1
−1
where lbulk ∈ N obeys
P e1/3 1 P e1/3
2 1/6
≤ ≤4 .
log P e lbulk log1/6 P e
n
For the points {rk }k=1 , enforce the interpolation rule
lk = ℓ(rk )
which says here that
lbulk log1/6 P e √
(4.23) = 1 − rk for k = 1, . . . , n.
2k−1 P e1/3
At this point, all available choices have been made and we can go ahead with our proof of the desired bound
(4.21). Actually, we already did most of the heavy lifting in the paragraphs above, where we explained how
these choices follow from optimizing the result of Proposition 4.1. All that remains is to verify the hypotheses
from Section 4.1 and Proposition 4.1.
First, we check the assumptions of Section 4.1. Clearly, lk decreases with increasing k per (4.2). And as
r 7→ ℓ(r) is strictly decreasing, the points rk increase with increasing k. Taking k = 1 in (4.23), we get that
1 log1/6 P e √ √
1= 1/3
1 − r1 ≥ 2 1 − r1
lbulk P e
by our choice for lbulk . So, r1 ≥ 3/4 > 1/2 as required in (4.3). Thus, our choices for {lk }, {rk }, and n
constitute a viable branching flow.
Now, we check the hypotheses of Proposition 4.1. Its first one (4.10) requires that |lk+1 − lk | ∼ lk+1 ∼ lk
and δk+1 ∼ δk with fixed numerical constants. The former holds by the dyadic nature of our lk . For the
latter, introduce the inverse ℓ 7→ r(ℓ), which is strictly decreasing, and note that δk = rk+1 − rk obeys
ˆ lk
δk = |r′ (ℓ)| dℓ.
lk+1
′ ′
It suffices to check that |r (ℓk )| ∼ |r (ℓk+1 )|. Differentiating (4.22) implicitly and rearranging gives
P e1/3 √
(4.24) r′ (ℓ) = −2 1 − r.
log1/6 P e
One sees from (4.23) that 1 − rk ∼ 1 − rk+1 with a fixed numerical constant. Hence, (4.10) is proved.
Finally, we verify the second hypothesis (4.11) of Proposition 4.1, which is that lk . δk for each k. Since
δk = rk+1 − rk and lk ∼ |lk+1 − lk |, we must show that
 
 rk+1 − rk 
1.  .
lk+1 − lk 
Referring again to r(ℓ), it suffices to check that 1 . |r′ (ℓ)| for ℓ ∈ (l1 , ln ). Evidently by (4.24),
P e1/3 √ P e1/3 √
|r′ (ℓ)| & 1/6
1−r ≥ 1 − rn ∼ 1.
log P e log1/6 P e
In the last step we used (4.23) with k = n. That lbl ∼ δbl is clear. Theorem 1.1 is proved. 
25
 5. Unsteady roll-like flows for energy-constrained cooling
We close by proving the upper bound on energy-constrained cooling mentioned in the introduction. It is
an open challenge to decide whether this bound is sharp in its scaling with respect to P e, in the advective
limit P e → ∞.
Proposition 5.1. Let f (x, t) satisfy
ˆ τ
1
e−λ1 ((τ −t)∧t) ||f (·, t)||L2 (D) dt = 0

2
|f | + |∇f |2 < ∞.

lim √ and
τ →∞ τ 0
There exists a fixed, numerical constant C > 0 such that
C
2
|∇T |2 ≤ · |f | + |∇f |2

 
min
u(x,t) Pe
h|u|2 i=P e2
u=0 at ∂D
whenever P e ≥ 1. The same bound holds using no-penetration conditions u · n̂ = 0 in place of the no-slip
ones u = 0 at ∂D.
Proof. Pure convection roll-like flows occur as the simplest case of our branching flows from Section 4.1,
with n = 1 and using only l1 = lbulk and δ1 = δbl . The rescaled velocities
Pe
λP e u with λP e = p
h|u|2 i
generate temperature fields TP e satisfying
  

2
 2 1 1 δbl
|∇TP e | . C0 (f ) lbulk + δbl + 2 + 2
P e2 lbulk lbulk
so long as lbulk . δbl , where now

C0 (f ) = |f |2 + |∇f |2 .



The proof of this is contained in that of Proposition 4.1, once one notes that the un-scaled velocities from
Section 4.1 obey
2
|u| . 1.
To see this write
u = χ1 u1 + χ′1 ψ1 êθ
in the case n = 1, and use the definitions to get that
lbulk
|u| . 1 + . 1.
δbl
(The proof for n > 1 is the same, but we leave it to the reader as it is not needed here.) The dependence of
C0 on f comes from Lemma 4.1 and Lemma 4.2 and not, in this case, from Lemma 4.3. Optimizing gives
1
δbl ∼ lbulk ∼
P e1/2
and this proves the result. 

References
1. Tabish Alam and Man-Hoe Kim, A comprehensive review on single phase heat transfer enhancement techniques in heat
exchanger applications, Renewable and Sustainable Energy Reviews 81 (2018), 813–839.
2. Ali Arslan, Giovanni Fantuzzi, John Craske, and Andrew Wynn, Bounds for internally heated convection with fixed boundary
heat flux, Journal of Fluid Mechanics 922 (2021), R1.
3. , Bounds on heat transport for convection driven by internal heating, Journal of Fluid Mechanics 919 (2021), A15.
4. Marco Avellaneda and Andrew J. Majda, An integral representation and bounds on the effective diffusivity in passive
advection by laminar and turbulent flows, Comm. Math. Phys. 138 (1991), no. 2, 339–391.
5. Vincent Bouillaut, Simon Lepot, Sébastien Aumaître, and Basile Gallet, Transition to the ultimate regime in a radiatively
driven convection experiment, Journal of Fluid Mechanics 861 (2019), R5.
6. F. H. Busse, On Howard’s upper bound for heat transport by turbulent convection, J. Fluid. Mech. 37 (1969), no. 3, 457–477.
26
 7. Ivan Catton, Conjugate Heat Transfer Within a Heterogeneous Hierarchical Structure, Journal of Heat Transfer 133 (2011),
no. 10, 103001.
8. S. I. Chernyshenko, P. Goulart, D. Huang, and A. Papachristodoulou, Polynomial sum of squares in fluid dynamics: a
review with a look ahead, Philos. Trans. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci. 372 (2014), no. 2020, 20130350, 18.
9. Ronald Coifman, Pierre-Louis Lions, Yves Meyer, and Stephen Semmes, Compensated compactness and Hardy spaces,
Journal de mathématiques pures et appliquées 72 (1993), no. 3, 247–286.
10. Peter Constantin and Charles R. Doering, Variational bounds on energy dissipation in incompressible flows. II. Channel
flow, Phys. Rev. E (3) 51 (1995), no. 4, part A, 3192–3198.
11. Charles R. Doering, Turning up the heat in turbulent thermal convection, Proceedings of the National Academy of Sciences
117 (2020), no. 18, 9671–9673.
12. Charles R. Doering and Peter Constantin, Variational bounds on energy dissipation in incompressible flows: Shear flow,
Phys. Rev. E (3) 49 (1994), no. 5, part A, 4087–4099.
13. , Variational bounds on energy dissipation in incompressible flows. III. Convection, Phys. Rev. E 53 (1996), 5957–
5981.
14. Charles R. Doering and Ian Tobasco, On the optimal design of wall-to-wall heat transport, Communications on Pure and
Applied Mathematics 72 (2019), no. 11, 2385–2448.
15. Theodore D. Drivas, Huy Q. Nguyen, and Camilla Nobili, Bounds on heat flux for Rayleigh-Bénard convection between
Navier-slip fixed-temperature boundaries, preprint arXiv:2109.13205.
16. Lawrence C. Evans, Partial differential equations, second ed., Graduate Studies in Mathematics, vol. 19, American Math-
ematical Society, Providence, RI, 2010. MR 2597943
17. Albert Fannjiang and George Papanicolaou, Convection enhanced diffusion for periodic flows, SIAM J. Appl. Math. 54
(1994), no. 2, 333–408. MR 1265233
18. Giovanni Fantuzzi, Bounds for Rayleigh-Bénard convection between free-slip boundaries with an imposed heat flux, Journal
of Fluid Mechanics 837 (2018), R5.
19. David Geb, Michael Ge, Jonathan Chu, and Ivan Catton, Measuring Transport Coefficients in Heterogeneous and Hierar-
chical Heat Transfer Devices, Journal of Heat Transfer 135 (2013), no. 6, 061101.
20. David Goluskin, Internally heated convection beneath a poor conductor, Journal of Fluid Mechanics 771 (2015), 36–56.
21. , Internally heated convection and Rayleigh-Bénard convection, Springer, 2016.
22. , Bounding averages rigorously using semidefinite programming: mean moments of the Lorenz system, Journal of
Nonlinear Science 28 (2018), no. 2, 621–651.
23. David Goluskin and Giovanni Fantuzzi, Bounds on mean energy in the Kuramoto-Sivashinsky equation computed using
semidefinite programming, 32 (2019), no. 5, 1705–1730.
24. Louis N. Howard, Heat transport by turbulent convection, J. Fluid. Mech. 17 (1963), no. 3, 405–432.
25. Gautam Iyer and Truong-Son Van, Bounds on the heat transfer rate via passive advection, 2021.
26. R. R. Kerswell, Unification of variational principles for turbulent shear flows: the background method of Doering-Constantin
and the mean-fluctuation formulation of Howard-Busse, Phys. D 121 (1998), no. 1-2, 175–192.
27. Robert V Kohn and Gilbert Strang, Optimal design and relaxation of variational problems, I, Communications on pure
and applied mathematics 39 (1986), no. 1, 113–137.
28. , Optimal design and relaxation of variational problems, II, Communications on Pure and Applied Mathematics 39
(1986), no. 2, 139–182.
29. , Optimal design and relaxation of variational problems, III, Communications on Pure and Applied Mathematics
39 (1986), no. 3, 353–377.
30. Mayur V Lakshmi, Giovanni Fantuzzi, Jesús D Fernández-Caballero, Yongyun Hwang, and Sergei I Chernyshenko, Finding
extremal periodic orbits with polynomial optimization, with application to a nine-mode model of shear flow, SIAM Journal
on Applied Dynamical Systems 19 (2020), no. 2, 763–787.
31. Simon Lepot, Sébastien Aumaître, and Basile Gallet, Radiative heating achieves the ultimate regime of thermal convection,
Proceedings of the National Academy of Sciences 115 (2018), no. 36, 8937–8941.
32. F. Marcotte, C. R. Doering, J.-L. Thiffeault, and W. R. Young, Optimal heat transfer and optimal exit times, SIAM J.
Appl. Math. 78 (2018), no. 1, 591–608.
33. Benjamin Miquel, Simon Lepot, Vincent Bouillaut, and Basile Gallet, Convection driven by internal heat sources and sinks:
Heat transport beyond the mixing-length or “ultimate” scaling regime, Phys. Rev. Fluids 4 (2019), 121501.
34. S. Motoki, G. Kawahara, and M. Shimizu, Maximal heat transfer between two parallel plates, J. Fluid Mech. 851 (2018),
R4, 14.
35. Stefan Müller, Hardy space methods for nonlinear partial differential equations, Equadiff 8 (1994), 159–168.
36. C. Nobili and F. Otto, Limitations of the background field method applied to Rayleigh-Bénard convection, J. Math. Phys.
58 (2017), no. 9, 093102, 46.
37. Matthew L Olson, David Goluskin, William W Schultz, and Charles R Doering, Heat transport bounds for a truncated
model of Rayleigh–Bénard convection via polynomial optimization, Physica D: Nonlinear Phenomena 415 (2021), 132748.
38. F. Otto and C. Seis, Rayleigh–Bénard convection: improved bounds on the Nusselt number, J. Math. Phys. 52 (2011),
no. 8, 083702.
39. S. C. Plasting and R. R. Kerswell, Improved upper bound on the energy dissipation rate in plane Couette flow: the full
solution to Busse’s problem and the Constantin-Doering-Hopf problem with one-dimensional background field, J. Fluid
Mech. 477 (2003), 363–379.

27
 40. Hamidreza Rastan, Amir Abdi, Bejan Hamawandi, Monika Ignatowicz, Josua P. Meyer, and Björn Palm, Heat transfer
study of enhanced additively manufactured minichannel heat exchangers, International Journal of Heat and Mass Transfer
161 (2020), 120271.
41. Ricardo M. S. Rosa and Roger M. Temam, Optimal minimax bounds for time and ensemble averages for the incompressible
Navier-Stokes equations, 2021.
42. Andre N. Souza, Ian Tobasco, and Charles R. Doering, Wall-to-wall optimal transport in two dimensions, Journal of Fluid
Mechanics 889 (2020), A34.
43. I. Tobasco, D. Goluskin, and C. R. Doering, Optimal bounds and extremal trajectories for time averages in nonlinear
dynamical systems, Phys. Lett. A 382 (2018), no. 6, 382–386.
44. Ian Tobasco and Charles R. Doering, Optimal wall-to-wall transport by incompressible flows, Phys. Rev. Lett. 118 (2017),
264502.
45. Srikanth Toppaladoddi, Andrew J Wells, Charles R Doering, and John S Wettlaufer, Thermal convection over fractal
surfaces, Journal of Fluid Mechanics 907 (2021).
46. Arel Weisberg, Haim H. Bau, and J.N. Zemel, Analysis of microchannels for integrated cooling, International Journal of
Heat and Mass Transfer 35 (1992), no. 10, 2465–2474.
47. Jared P. Whitehead and Charles R. Doering, Ultimate state of two-dimensional Rayleigh-Bénard convection between free-
slip fixed-temperature boundaries, Phys. Rev. Lett. 106 (2011), 244501.

28

View publication stats