You are on page 1of 34

SIAM J. MATH. ANAL.

c 2018 Society for Industrial and Applied Mathematics


Vol. 50, No. 1, pp. 557–590

RELATIVE ENTROPY, WEAK-STRONG UNIQUENESS, AND


Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

CONDITIONAL REGULARITY FOR A COMPRESSIBLE


OLDROYD–B MODEL∗
YONG LU† AND ZHIFEI ZHANG‡

Abstract. We consider the compressible Oldroyd–B model derived in [J. W. Barrett, Y. Lu,
and E. Süli, Comm. Math. Sci., 15 (2017), pp. 1265–1323], where the existence of global-in-time
finite-energy weak solutions was shown in a two-dimensional setting. In this paper, we first state a
local well-posedness result for this compressible Oldroyd–B model. In the two-dimensional setting,
we give a (refined) blow-up criterion involving only the upper bound of the fluid density. We then
show that if the initial fluid density and polymer number density admit a positive lower bound,
the weak solution coincides with the strong one as long as the latter exists. Moreover, if the fluid
density of a weak solution issuing from regular initial data admits a finite upper bound, this weak
solution is indeed a strong one; this can be seen as a corollary of the refined blow-up criterion and
the weak-strong uniqueness.

Key words. compressible Oldroyd–B model, relative entropy inequality, weak-strong unique-
ness, conditional regularity

AMS subject classifications. 35Q35, 76N10, 76A05

DOI. 10.1137/17M1128654

1. Introduction. The incompressible Oldroyd–B model is a macroscopic model


involving only macroscopic quantities, such as the velocity, the pressure, and the
stress. It is known that from the incompressible Navier–Stokes–Fokker–Planck system,
which is a micro-macro model describing incompressible dilute polymeric fluids, one
can derive, at least formally, the Oldroyd–B model; see [26].
A similar derivation can be performed in the compressible setting. Indeed, in
[3], a compressible Oldroyd–B model was derived as a macroscopic closure of the
compressible Navier–Stokes–Fokker–Planck equations studied in a series of papers by
Barrett and Süli [7, 6, 4, 5, 8].
We recall the compressible Oldroyd–B model derived in [3]. Let Ω ⊂ Rd be
a bounded open domain with a C 2,β boundary (also called a C 2,β domain), with
β ∈ (0, 1), and let d ∈ {2, 3}. We consider the following compressible Oldroyd–B
model posed in the time-space cylinder (0, T ) × Ω:

(1.1) ∂t % + divx (%u) = 0,


(1.2)
∂t (%u)+divx (%u ⊗ u)+∇x p(%) − (µ∆x u+ν∇x divx u) = divx T − (kLη+z η 2 ) I +% f ,


(1.3) ∂t η + divx (ηu) = ε∆x η,


k 1
∂t T + Divx (u T) − ∇x u T + T ∇T

(1.4) x u = ε∆x T + ηI− T,
2λ 2λ
∗ Received by the editors May 3, 2017; accepted for publication (in revised form) October 31,

2017; published electronically January 30, 2018.


http://www.siam.org/journals/sima/50-1/M112865.html
Funding: The work of the first author was partially supported by project ANR JCJC BORDS
funded by l’ANR of France. The work of the second author was partially supported by NSF of China
under grant 11425103.
† Department of Mathematics, Nanjing University, 210093 Nanjing, China (lvyong@amss.ac.cn).
‡ School of Mathematical Science, Peking University, 100871 Beijing, China (zfzhang@math.pku.

edu.cn).
557

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


558 YONG LU AND ZHIFEI ZHANG

where the pressure p and the density % of the solvent are assumed to be related by
the typical power law relation,
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(1.5) p(%) = a%γ , a > 0, γ > 1.

The term µ∆x u + ν∇x divx u corresponds to divx S(∇x u), where S(∇x u) is the
Newtonian stress tensor defined by

∇x u + ∇T
 
S xu 1
(1.6) S(∇x u) = µ − (divx u)I + µB (divx u)I,
2 d

where µS > 0 and µB ≥ 0 are the shear and bulk viscosity coefficients, respectively.
Indeed, a direct calculation gives

µS µS µS
 
divx S(∇x u) = ∆x u + µB + − ∇x divx u = µ∆x u + ν∇x divx u,
2 2 d

with

µS µS µS
µ := > 0, ν := µB + − ≥ 0.
2 2 d
The velocity gradient matrix is defined as

(∇x u)1≤i,j≤d = (∂xj ui )1≤i,j≤d .

The symmetric matrix function T = (Ti,j ), 1 ≤ i, j ≤ d, defined on (0, T ) × Ω, is


the extra stress tensor, and the notation Divx (u T) is defined by

(Divx (u T))i,j = divx (u Ti,j ), 1 ≤ i, j ≤ d.

The meanings of the various quantities and parameters appearing in (1.1)–(1.4) were
introduced in the derivation of the model in [3]. In particular, the parameters ε, k, λ
are all positive numbers, whereas z ≥ 0 and L ≥ 0 with z + L > 0.
The polymer number density η is a nonnegative scalar function defined as the
integral of the probability density function ψ, which is governed by the Fokker–Planck
equation, in the conformation vector, which is a microscopic variable in the modeling
of dilute polymer chains. The term q(η) := kLη + zη 2 in the momentum equation
(1.2) can be considered as the polymer pressure, compared to the fluid pressure p(%).
The equations (1.1)–(1.4) are supplemented by proper initial conditions for %, u,
η, and T, and the following boundary conditions are imposed:

(1.7) u = 0 on (0, T ) × ∂Ω,


(1.8) ∂n η = 0 on (0, T ) × ∂Ω,
(1.9) ∂n T = 0 on (0, T ) × ∂Ω.

Here ∂n := n · ∇x , where n is the outer unit normal vector on the boundary ∂Ω. The
external force f is assumed to be, at least, in L∞ ((0, T ) × Ω; Rd ).
There are stress diffusion terms ε∆x η and ε∆x T in our model. Such spatial
stress diffusions are indeed allowed in some models of complex fluids, such as the
creeping flow regime, as pointed out in [12]. Also in the modeling of the compressible
Navier–Stokes–Fokker–Planck system arising in the kinetic theory of dilute polymeric

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 559

fluids, where polymer chains are immersed in a barotropic, compressible, isothermal,


viscous Newtonian solvent, Barrett and Süli [7] observed the presence of the center-
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

of-mass diffusion term ε∆x ψ, where ψ is the probability density function depending
on both microscopic and macroscopic variables; as a result, its macroscopic closure
(the compressible Oldroyd–B model) contains such diffusion terms.
The incompressible Oldroyd–B model has continually attracted the attention of
mathematicians. The local-in-time well-posedness, as well as the global-in-time well-
posedness with small data, in various spaces is known, thanks to the contributions of
Renardy [30], Guillopé and Saut [20, 21], and Fernández-Cara, Guillén, and Ortega
[17]. Concerning the global-in-time existence of solutions with large data, in the
corotational derivative setting where, in the system for the extra stress tensor, the
velocity gradient is replaced by its antisymmetric part in one of the terms, Lions and
Masmoudi [27] showed the global-in-time existence of weak solutions with large initial
data. In the presence of stress diffusion, when a regularizing Laplacian term is present
in the extra stress tensor, Barrett and Boyaval [2] showed the global-in-time existence
of large data weak solutions in a two-dimensional setting. Also in the presence of
stress diffusion and in the two-dimensional setting, Constantin and Kliegl [12] proved
the global existence of strong solutions with large initial data, which can be seen as an
extension of the global well-posedness theory for the two-dimensional incompressible
Navier–Stokes equations.
Many fundamental problems for the incompressible Oldroyd–B model are still
open, such as the global-in-time existence of large data solutions—even weak ones—
both in the two-dimensional and the three-dimensional setting without the stress
diffusion. Even with stress diffusion, the global-in-time existence of large data
solutions, strong or weak, is still open in the three-dimensional setting. This is
somewhat expected, since the well-posedness of the incompressible Navier–Stokes
equations is a well-known open problem.
Even less is known concerning the compressible Oldroyd–B models. Let us
mention some mathematical results for compressible viscoelastic models, which have
been the subject of active research in recent years. The existence and uniqueness
of local strong solutions and the existence of global solutions near an equilibrium
for macroscopic models of three-dimensional compressible viscoelastic fluids were
considered in [29, 23, 24, 25]. In particular, Fang and Zi [13] proved the existence
of a unique local-in-time strong solution to a compressible Oldroyd–B model and
established a blow-up criterion for strong solutions. In [3], the existence of global-
in-time weak solutions in a two-dimensional setting for the compressible Oldroyd–B
model (1.1)–(1.9) was shown.
We remark that many of the compressible Oldroyd–B type models considered
before, as in [13, 25], are modifications of the incompressible Oldroyd–B model by
replacing the incompressible Navier–Stokes equations with the compressible ones.
The model considered in this paper is derived from a micro-macro model for dilute
polymeric fluids and may have a stronger physical grounding. Moreover, this fully
macroscopic model derived from a micro-macro model coincides with the formal
derivation of the incompressible Oldroyd–B model from the incompressible Navier–
Stokes–Fokker–Planck equations.

2. Main results. In this section, we state our main results. We first recall the
result shown in [3] concerning the global-in-time existence of weak solutions. We state
a theorem concerning the local-in-time well-posedness of strong solutions and a blow-
up criterion. In the two-dimensional setting, we give a refined blow-up criterion result

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


560 YONG LU AND ZHIFEI ZHANG

where only the L∞ bound of the fluid density is needed. We then show a weak-strong
uniqueness result by using the relative entropy method. As a corollary, this offers us
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

a conditional regularity theorem.

2.1. Global-in-time finite-energy weak solutions. We state our basic


hypotheses on the initial data:

%(0, ·) = %0 (·) with %0 ≥ 0 a.e. in Ω, %0 ∈ Lγ (Ω),


u(0, ·) = u0 (·) ∈ Lr (Ω; Rd ) for some r ≥ 2γ 0 such that %0 |u0 |2 ∈ L1 (Ω),
(
(2.1) η0 ∈ L2 (Ω) if z > 0,
η(0, ·) = η0 (·) with η0 ≥ 0 a.e. in Ω,
η0 log η0 ∈ L1 (Ω) if z = 0,
T(0, ·) = T0 (·) with T0 = TT
0 ≥ 0 a.e. in Ω, T0 ∈ L2 (Ω; Rd×d ).

A related weak solution is defined as follows.

Definition 2.1. Let T > 0, and let Ω ⊂ Rd be a bounded C 2,β domain with
0 < β < 1. We say that (%, u, η, T) is a finite-energy weak solution in (0, T ) × Ω
to the system of equations (1.1)–(1.9), supplemented by the initial data (2.1), if the
following hold:
• % ≥ 0 a.e. in (0, T ) × Ω, % ∈ Cw ([0, T ]; Lγ (Ω)), u ∈ L2 (0, T ; W01,2 (Ω; Rd )),

%u ∈ Cw ([0, T ]; L γ+1 (Ω; Rd )), %|u|2 ∈ L∞ (0, T ; L1 (Ω)),
η ≥ 0 a.e. in (0, T ) × Ω,
η ∈ Cw ([0, T ]; L2 (Ω)) ∩ L2 (0, T ; W 1,2 (Ω))
(
if z > 0,
1
η log η ∈ L∞ (0, T ; L1 (Ω)), η ∈ L2 (0, T ; W 1,2 (Ω)), η ∈ Cw ([0, T ]; L1 (Ω) if z = 0,
2

T = TT ≥ 0 a.e. in (0, T )×Ω, T ∈ Cw ([0, T ]; L2 (Ω; Rd×d ))∩L2 (0, T ; W 1,2 (Ω; Rd×d )).

• For any t ∈ (0, T ) and any test function φ ∈ C ∞ ([0, T ] × Ω), one has

Z tZ Z Z
%∂t φ + %u · ∇x φ dx dt0 =
 
(2.2) %(t, ·)φ(t, ·) dx − %0 φ(0, ·) dx,
0 Ω Ω Ω

(2.3)
Z tZ Z Z
η∂t φ+ηu·∇x φ−ε∇x η·∇x φ dx dt0 =
 
η(t, ·)φ(t, ·) dx− η0 φ(0, ·) dx.
0 Ω Ω Ω

• For any t ∈ (0, T ) and any test function ϕ ∈ C ∞ ([0, T ]; Cc∞ (Ω; Rd )), one has
(2.4)
Z tZ

%u · ∂t ϕ + (%u ⊗ u) : ∇x ϕ + p(%) divx ϕ
0 Ω
+ kLη + z η 2 divx ϕ − S(∇x u) : ∇x ϕ dx dt0
 
Z tZ Z Z
0
 
= T : ∇x ϕ − % f · ϕ dx dt + %u(t, ·) · ϕ(t, ·) dx − %0 u0 · ϕ(0, ·) dx.
0 Ω Ω Ω

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 561

• For any t ∈ (0, T ) and any test function Y ∈ C ∞ ([0, T ] × Ω; Rd×d ), one has
(2.5)
Z tZ
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

0
T : ∂t Y + (u T) :: ∇x Y + ∇x u T + T ∇T
  
x u : Y − ε∇x T :: ∇x Y dx dt
0 Ω
Z tZ  
k 1
= − η tr (Y) + T : Y dx dt0
0 Ω 2λ 2λ
Z Z
+ T(t, ·) : Y(t, ·) dx − T0 : Y(0, ·) dx.
Ω Ω

• The continuity equation holds in the sense of renormalized solutions: for any
b ∈ Cb1 [0, ∞),

(2.6) ∂t b(%) + divx (b(%)u) + (b0 (%)% − b(%)) divx u = 0 in D0 ((0, T ) × Ω).

• For a.e. t ∈ (0, T ), the following energy inequality holds:


(2.7)Z  
1 2 a γ 2
 1
%|u| + % + kL(η log η + 1) + z η + tr (T) (t, ·)dx
Ω 2 γ−1 2
Z tZ Z tZ
1 1
+ 2ε 2kL|∇x η 2 |2 + z |∇x η|2 dx dt0 + tr (T) dx dt0
0 Ω 4λ 0 Ω
Z tZ
2
+ µ |∇x u| + ν|divx u|2 dx dt0
Z 0 Ω 
1 a  1
≤ %0 |u0 |2 + %γ0 + kL(η0 log η0 + 1) + z η02 + tr (T0 ) dx
Ω 2 γ−1 2
Z tZ Z tZ
kd
+ % f · u dx dt0 + η dx dt0 .
0 Ω 4λ 0 Ω

We recall the associated result concerning the existence of large data global-in-
time finite-energy weak solutions, which can be obtained by summarizing Theorems
11.2 and 12.1 of [3].
Theorem 2.2. Let γ > 1, and let Ω ⊂ R2 be a bounded C 2,β domain with β ∈
(0, 1). Assume that the parameters ε, k, λ are all positive numbers and that z ≥ 0,
L ≥ 0 with z + L > 0. Then for any T > 0, there exists a finite-energy weak solution
(%, u, η, T) in the sense of Definition 2.1 with initial data (2.1). Moreover, the extra
stress tensor T satisfies the bound
Z Z tZ Z tZ
0 1
2
|T(t, ·)| dx + ε 2
|∇x T| dx dt + |T|2 dx dt0 ≤ C(t, kT0 kL2 (Ω) , E0 )
Ω 0 Ω 4λ 0 Ω

for a.e. t ∈ (0, T ), where E0 is given by


Z  
1 a  1
E0 := %0 |u0 |2 + %γ0 + kL(η0 log η0 + 1) + z η02 + tr (T0 ) dx.
Ω 2 γ−1 2

2.2. Local well-posedness and blow-up criterion. We now state a result


concerning the local-in-time existence of strong solutions. By strong solution we
mean a weak solution that satisfies (1.1)–(1.9) a.e. in the space-time cylinder under
consideration.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


562 YONG LU AND ZHIFEI ZHANG

Theorem 2.3. Let γ > 1, and let Ω ⊂ Rd be a bounded C 2,β domain with β ∈
(0, 1). Assume the parameters ε, k, λ are all positive numbers, whereas z ≥ 0 and
L ≥ 0 with z+L > 0. We assume the external force f ∈ W 1,2 ((0, ∞)×Ω). In addition
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

to the assumption on the initial data in (2.1), we assume that


(2.8) %0 ∈ W 1,6 (Ω), η0 ∈ Wn2,2 (Ω), T0 ∈ Wn2,2 (Ω; Rd×d ), u0 ∈ W01,2 ∩ W 2,2 (Ω; Rd ),
where the notation Wn2,2 (Ω) := {f ∈ W 2,2 (Ω) : ∂n f = 0 on ∂Ω}. Suppose that there
holds

(2.9) − (µ∆x u0 + ν∇x divx u0 ) + ∇x p(%0 ) − divx T0 + ∇x (kLη0 + z η02 ) = %0 g
for some g ∈ L2 (Ω; Rd ). Thus, there exists a unique strong solution (%, u, η, T) to
(1.1)–(1.9) with a maximal existence time T∗ ∈ (0, ∞] such that
(2.10)
% ≥ 0, % ∈ C([0, T∗ ), W 1,6 (Ω)),
u ∈ C([0, T∗ ), W01,2 ∩ W 2,2 (Ω; Rd )) ∩ L2loc ([0, T∗ ); W 2,r (Ω; Rd )),
η ≥ 0, T = TT ≥ 0, (η, T) ∈ C([0, T∗ ), Wn2,2 ) ∩ L2loc ([0, T∗ ); W 3,2 )(Ω; R × Rd×d ),
where r = 6 when d = 3 and r ∈ (1, ∞) is arbitrary when d = 2.
If T∗ < ∞, then there is a finite-time blow-up at T∗ in the sense that

(2.11) lim sup k%kL∞ ((0,T )×Ω) + kηkL∞ ((0,T )×Ω) + kTkL2 (0,T ;L∞ (Ω;Rd×d )) = ∞.
T →T∗

Remark 2.4. This local-in-time well-posedness result for strong solutions is


inspired by the study [11] for compressible Navier–Stokes equations. If the initial
density is additionally assumed to have a positive lower bound, condition (2.9) is
automatically satisfied. In such a setting, local-in-time well-posedness can also be
obtained by employing the method presented in [31, 34].
Remark 2.5. Theorem 2.3 is given in a similar manner as Theorems 1.1 and 1.2 in
[13] and can be proved similarly. In fact, there are extra diffusion terms in our model
compared to the model considered in [13], and this makes the analysis even easier.
Hence, we omit the proof of this theorem. The regularity assumption on the initial
data may not be optimal and can be relaxed accordingly by employing the argument
in [11].
In the two-dimensional setting, we offer the following refined blow-up criterion.
Theorem 2.6. Let d = 2, and let (%, u, η, T) be the strong solution obtained in
Theorem 2.3 to (1.1)–(1.9) with a maximal existence time T∗ ∈ (0, ∞]. If T∗ < ∞,
there holds
(2.12) lim sup k%kL∞ ((0,T )×Ω) = ∞.
T →T∗

Remark 2.7. The blow-up criterion (2.11), which is reproduced from [13], is
inspired by the related studies for the compressible Navier–Stokes equations in [32, 33]
and for the incompressible Oldroyd–B model in [10]. Our refined criterion in Theorem
2.6 coincides with those in [32, 33], where only the upper bound of the fluid density
is needed. Such a refinement crucially depends on the two-dimensional setting and
the presence of the diffusion terms in T and η. This setting allows us to obtain
improved estimates for T and η that are uniform in time (see Propositions 4.1 and
4.2). In the three-dimensional setting, it is not known whether one can achieve such
an improvement.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 563

2.3. Weak-strong uniqueness and conditional regularity. In the two-


dimensional setting, we show the following weak-strong uniqueness result, provided
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

the fluid density and polymer number density admit positive lower bounds.
Theorem 2.8. Let d = 2. Let (%, u, η, T) be a finite-energy weak solution obtained
in Theorem 2.2, and let (%̃, ũ, η̃, T̃) be the strong solution obtained in Theorem 2.3 with
the same initial data satisfying the assumptions stated in Theorem 2.3. If, in addition,
the initial data satisfy

(2.13) inf %0 > 0, inf η0 > 0,


Ω Ω

then there holds

(2.14) (%, u, η, T) = (%̃, ũ, η̃, T̃) in [0, T∗ ) × Ω.

Finally, as a corollary of Theorems 2.6 and 2.8, we have the following conditional
regularity result for finite-energy weak solutions.
Theorem 2.9. Let d = 2. Let (%, u, η, T) be a finite-energy weak solution obtained
in Theorem 2.2, with initial data satisfying the assumptions stated in Theorem 2.3 and
the additional lower bound constraints (2.13). If, for some T > 0, there holds the upper
bound

(2.15) sup % < ∞,


(0,T )×Ω

then the weak solution (%, u, η, T) is actually a strong solution satisfying the estimates
(2.10) over the time interval [0, T ].
The rest of the paper is devoted to the proofs of Theorems 2.6, 2.8, and 2.9. In
section 3, we recall some necessary lemmas. Theorems 2.6, 2.8, and 2.9 are proved in
sections 4 and 6.
Throughout the paper, C denotes some uniform constant whose value may
differ from line to line. In what follows, to avoid notation complications we
sometimes use Lr (0, T ; X(Ω)) to denote the scalar function space Lr (0, T ; X(Ω)),
the vector valued function space Lr (0, T ; X(Ω; Rd )), or the matrix valued function
space Lr (0, T ; X(Ω; Rd×d )) if there is no danger of confusion.
3. Preliminaries. In this section we recall some technical tools that will be used
later. The first one concerns the Dirichlet problem for the Lamé system.
Lemma 3.1. Let µ > 0 and ν ≥ 0, and let G ⊂ Rd be a bounded C 2 domain. Let
u be the unique weak solution to
(
Lu := −µ∆x u − ν∇x divx u = f in G,
(3.1)
u=0 on ∂G

for some f ∈ Lr (G; Rd ) with 1 < r < ∞.


(i) There holds the following estimate for some constant C = C(µ, r, d, G),
depending only on µ, r, d, and the C 2 norm of G:

kukW 1,r ∩W 2,r (G) ≤ C(µ, r, d, G)kf kLr (G) .


0

(ii). If, moreover, f = divx F with F = (Fi,j )1≤i,j≤d ∈ Ls (G), 1 < s < ∞, then

kukW 1,s (G) ≤ C(µ, s, d, G)kF kLs (G)


0

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


564 YONG LU AND ZHIFEI ZHANG

for some constant C(µ, s, d, G) depending only on µ, s, d, and the C 1 norm


of G.
k
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(iii) If, moreover, for any 1 ≤ i, j ≤ d, Fi,j = divx Hi,j with Hi,j = (Hi,j )1≤k≤d ∈
t
L (G), 1 < t < ∞, and Hi,j · n = 0 on ∂G, then

kukLt (G) ≤ C(µ, t0 , d, G)kHkLt (G) ,

where the constant C(µ, t0 , d, G) is the same as in (i), where t0 is such that
1/t + 1/t0 = 1.
Remark 3.2. In the statement of (iii), it is meaningful to demand the normal
trace Hi,j · n = 0 on ∂G in the sense of distributions, due to the fact that Hi,j =
k
(Hi,j )1≤k≤d ∈ Lt (G) and divx Hi,j = Fi,j ∈ Ls (G) for some s, t ∈ (1, ∞). We refer
the reader to [28, section 3.2] and [15, section 10.3] for detailed descriptions.
Proof. The results in (i) and (ii) in Lemma 3.1 are collected from classical
estimates for linear elliptic systems; see, for example, [1]. The result in (iii) can
be proved by a duality argument by using the estimate in (i). Indeed, for any
0 0 0
ϕ ∈ Lt (G; Rd ), we let v ∈ W01,t ∩ W 2,t (G; Rd ) be the unique solution to
(
Lv = ϕ in G,
(3.2)
v=0 on ∂G,

satisfying, by the estimate in (i),

(3.3) kvkW 1,t0 ∩W 2,t0 (G) ≤ C(µ, t0 , d, G)kϕkLt0 (G) .


0

Then, by using the boundary conditions on v and Hi,j , we have


(3.4)
Z Z Z Z Z
u · ϕ dx = u · Lv dx = Lu · v dx = divx divx H · v dx = H : ∇2x v dx
G G G G G
≤ kHkLt (G) k∇2x vkLt0 (G) ≤ C(µ, t0 , d, G)kHkLt (G) kϕkLt0 (G) .
0
This is true for any ϕ ∈ Lt (G; Rd ). We then obtain the estimate in (iii).
In what follows, we will use L−1 (f ) to denote the solution u to (3.1).
We then recall the following classical Gagliardo–Nirenberg inequality.
Lemma 3.3. Let G ⊂ Rd be a bounded Lipschitz domain; then, for any r ∈ [2, ∞)
if d = 2, and r ∈ [2, 6] if d = 3, one has, for any v ∈ W 1,2 (G), that
 
1 1
kvkLr (G) ≤ C(r, d, G)kvk1−θ
L2 (G) kvk θ
W 1,2 (G) , θ := d − .
2 r
Now we recall some regularity results for parabolic Neumann problems. We first
introduce fractional-order Sobolev spaces. Let G be the whole space Rd or a bounded
Lipschitz domain in Rd . For any k ∈ N, β ∈ (0, 1), and s ∈ [1, ∞), we define

W k+β,s (G) := v ∈ W k,s (G) : kvkW k+β,s (G) < ∞ ,




where
X Z Z |∂ α v(x) − ∂ α v(y)|s  1s
kvkW k+β,s (G) := kvkW k,s (G) + dx dy .
G G |x − y|d+βs
|α|=k

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 565

The following classical results are taken from section 7.6.1 in [28]. Consider the
parabolic initial-boundary value problem,
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(3.5)
∂t ρ − ε ∆x ρ = h in (0, T ) × G; ρ(0, ·) = ρ0 in G; ∂n ρ = 0 in (0, T ) × ∂G.
Here ε > 0, ρ0 and h are known functions, and ρ is the unknown solution. The first
regularity result of relevance to us here is the following lemma.
Lemma 3.4. Let 0 < β < 1, 1 < p, q < ∞, and let G ⊂ Rd be a bounded C 2,β
domain with β ∈ (0, 1),
2
2− p ,q
ρ0 ∈ Wn , h ∈ Lp (0, T ; Lq (G)),
2− 2 ,q
where Wn p is the completion of the linear space {v ∈ C ∞ (G) : ∂n v|∂G = 0} with
2
respect to the norm of W 2− p ,q (G). Then there exists a unique function ρ satisfying
2
ρ ∈ Lp (0, T ; W 2,q (G)) ∩ C([0, T ]; W 2− p ,q (G)), ∂t ρ ∈ Lp (0, T ; Lq (G))
solving (3.5) in (0, T ) × G; in addition, ρ satisfies the Neumann boundary condition
in (3.5) in the sense of the normal trace, which is well defined since ∆x ρ ∈
Lp (0, T ; Lq (G)). Moreover,
1
ε1− p kρk 2− 2 ,q + k∂t ρkLp (0,T ;Lq (G)) + εkρkLp (0,T ;W 2,q (G))
L∞ (0,T ;W p (G))

1
≤ C(p, q, G) ε1− p kρ0 k 2− p2 ,q
 
+ khkLp (0,T ;Lq (G)) .
W (G)

The second result concerns parabolic problems with a divergence-form source


term h = divx g.
Lemma 3.5. Let 0 < β < 1, 1 < p, q < ∞, and let G ⊂ Rd be a bounded C 2,β
domain with β ∈ (0, 1),
ρ0 ∈ Lq (G), g ∈ Lp (0, T ; Lq (G; Rd )).
Then, there exists a unique solution ρ ∈ Lp (0, T ; W 1,q (G))∩C([0, T ]; Lq (G)) satisfying
Z Z Z
d
%(0, ·) = %0 (·) a.e. in G, ρ φ dx+ε ∇x ρ·∇x φ dx = − g·∇x φ dx in D0 (0, T )
dt G G G

for any φ ∈ C ∞ (G). Moreover,


1
ε1− p kρkL∞ (0,T ;Lq (G)) +εk∇x ρkLp (0,T ;Lq (G;Rd ))
1
≤ C(p, q, G) ε1− p kρ0 kLq (G) +kgkLp (0,T ;Lq (G;Rd )) .
 

Finally, we recall the Bogovskiı̆ operator, whose construction can be found in [9]
and in Chapter III of Galdi’s book [18].
Lemma 3.6. Let 1 < p < ∞, and let G ⊂ Rd be a bounded Lipschitz domain. Let
Lp0 (G)be the space of all Lp (G) functions with zero mean value. Then there exists a
linear operator BG from Lp0 (G) to W01,p (G; Rd ) such that for any ρ ∈ Lp0 (G) one has
divx BG (ρ) = ρ in G; kBG (ρ)kW 1,p (G;Rd ) ≤ C(p, d, G) kρkLp (G) .
0

If, in addition, ρ = divx g for some g ∈ Lq (G; Rd ), 1 < q < ∞, g · n = 0 on ∂G,


then the following inequality holds:
kBG (ρ)kLq (G;Rd ) ≤ C(p, d, G) kgkLq (G;Rd ) .

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


566 YONG LU AND ZHIFEI ZHANG

4. A refined blow-up criterion. This section is devoted to the proof of


Theorem 2.6. Let d = 2, and let (%, u, η, T) be the strong solution given in Theorem
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

2.3, with T∗ the corresponding maximal existence time. By a contradiction argument,


to prove Theorem 2.6 it is sufficient to show that if

(4.1) T∗ < ∞, lim sup k%kL∞ ((0,T )×Ω) < ∞,
T →T∗

then the following uniform estimates in η and T hold:



(4.2) lim sup kηkL∞ ((0,T )×Ω) + kTkL2 ((0,T ),L∞ (Ω;R2×2 )) < ∞.
T →T∗

Let T1 ∈ (0, T∗ ) be close to T∗ , to be determined later. By Theorem 2.3, there


holds

(4.3) k%kL∞ (0,T1 ;W 1,6 (Ω)) + k(u, η, T)kL∞ (0,T1 ;W 2,2 (Ω;R2 ×R×R2×2 )) ≤ C(T1 , T∗ ) < ∞.

We remark that the constant C(T1 , T∗ ) may be unbounded as T1 → T∗ .


By Sobolev embedding we have k%kL∞ (Ω) ≤ Ck%kW 1,6 (Ω) , and by noting (4.3),
the assumption (4.1) is equivalent to the following assumption:

(4.4) k%kL∞ ((0,T∗ )×Ω) < ∞.

Now, starting from (4.4), we show our desired estimate (4.2) step by step in the rest
of this section. Some ideas are based on the methods in [22, 32], while new technical
difficulties arising from the terms in η and T need to be resolved.
In the rest of this section, the constant C depends only on the initial data and
the quantity k%kL∞ ((0,T∗ )×Ω) , which is assumed to be bounded for contradiction.
4.1. A priori estimates. In this section and section 4.2, we give some estimates
that are uniform over the time interval (0, T∗ ); in particular, these estimates hold
without assuming the condition (4.4).
We briefly recall the a priori energy estimates, and we refer the reader to section
3 in [3] for the details of the derivation for a slightly modified model.
For any time t ∈ (0, T∗ ), calculating
Z t Z Z 
1
(1.2) · u dx + tr(1.4) dx dt0
0 Ω 2 Ω

implies the energy inequality (2.7). In fact, here an energy equality is obtained due
to the smoothness of the solution. Then, applying Gronwall’s inequality gives the
following inclusions:
(4.5)
% ∈ L∞ (0, T∗ ; Lγ (Ω)), u ∈ L2 (0, T∗ ; W01,2 (Ω; R2 )), %|u|2 ∈ L∞ (0, T∗ ; L1 (Ω)),
η ∈ L∞ (0, T∗ ; L2 (Ω)) ∩ L2 (0, T∗ ; W 1,2 (Ω)) if z > 0,
(
1
η log η ∈ L∞ (0, T∗ ; L1 (Ω)), η 2 ∈ L2 (0, T∗ ; W 1,2 (Ω)) if z = 0,
tr(T) ∈ L∞ (0, T∗ ; L1 (Ω)).

We then take the inner product of (1.4) with T and integrate over Ω. Direct

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 567

calculation implies
Z Z Z
1 d 1
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

|T|2 dx + ε |∇x T|2 dx + |T|2 dx


2 dt Ω Ω 2λ Ω
Z Z Z
k
∇x u T + T ∇T

=− Divx (u T) : T dx + x u : T dx + η tr (T) dx
Ω Ω 2λ Ω
k2
Z Z
1
≤ 3 k∇x ukL2 (Ω;R2×2 ) kTk2L4 (Ω;R2×2 ) + |T|2 dx + η 2 dx ∀ t ∈ (0, T∗ ).
4λ Ω 2λ Ω

Applying the Gagliardo–Nirenberg inequality recalled in Lemma 3.3 in the case


of d = 2, we have, over the time interval (0, T∗ ), that

kTk2L4 (Ω) ≤ C kTkL2 (Ω) kTkW 1,2 (Ω) ≤ C kTkL2 (Ω) kTkL2 (Ω) + k∇x TkL2 (Ω) .


This implies
1 ε
3 k∇x ukL2 (Ω) kTk2L4 (Ω) ≤ C k∇x uk2L2 (Ω) kTk2L2 (Ω) + kTk2L2 (Ω) + k∇x Tk2L2 (Ω) .
8λ 2

We thus obtain for any t ∈ (0, T∗ )


Z Z tZ Z tZ
1
0
2
|T| (t, ·) dx + ε 2
|∇x T| dx dt + |T|2 dx dt0
Ω 0 Ω 04λ Ω
(4.6) Z t
k2 t
Z Z Z
≤ |T0 |2 dx + C k∇x uk2L2 (Ω) kTk2L2 (Ω) dt0 + η 2 dx dt0 .
Ω 0 λ 0 Ω

If z > 0, we have from (4.5) that kηkL∞ (0,T∗ ;L2 (Ω)) ≤ C. We thus deduce from
(4.6) by using Gronwall’s inequality that
Z tZ
(4.7) kT(t, ·)k2L2 (Ω) + ε |∇x T|2 dx dt0 ≤ C(E0 , kT0 kL2 (Ω) ) ∀ t ∈ (0, T∗ ).
0 Ω

We now consider the case z = 0, L > 0. By (4.5), there holds


1
(4.8) kη log ηkL∞ (0,T∗ ;L1 (Ω)) + k∇x η 2 kL2 (0,T∗ ;L2 (Ω)) ≤ C.

Then by (4.8), we have


Z Z
1 1 1 1
|∇x η| dx = |2η 2 ∇x η 2 | dx ≤ 2 kη 2 kL2 (Ω) k∇x η 2 kL2 (Ω) ≤ C.
Ω Ω

As d = 2, the Sobolev embedding of W 1,1 (Ω) into L2 (Ω) gives

kηkL2 (0,T∗ ;L2 (Ω)) ≤ C.

Then, by (4.6) and Gronwall’s inequality, we obtain the same estimate as in (4.7).
Hence, there holds the uniform estimates

(4.9) kTkL∞ (0,T∗ ;L2 (Ω;R2×2 )) + kTkL2 (0,T∗ ;W 1,2 (Ω;R2×2 )) ≤ C.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


568 YONG LU AND ZHIFEI ZHANG

4.2. Higher order estimates for η, T. Based on Sobolev embedding, the


interpolation between Lebesgue spaces, and the repeated application of Lemma 3.5,
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

we show the following higher order estimates for the polymer number density.
Proposition 4.1. For any r ∈ (1, ∞), there holds

(4.10) kηkL∞ (0,T∗ ;Lr (Ω)) + kηkL2 (0,T∗ ;W 1,r (Ω)) ≤ C.

Proof. The proof is similar to that of Proposition 12.2 in [3]. We have


shown in section 4.1 that u ∈ L2 (0, T∗ ; W 1,2 (Ω, R2 )) and η ∈ L∞ (0, T∗ ; L1 (Ω)) ∩
1
L2 (0, T∗ ; L2 (Ω)). By Sobolev embedding W 1,2 (Ω) ,→ L δ (Ω), for any δ ∈ (0, 1), we
have that
δ δ
ηu ∈ L2− 2 (0, T∗ ; L1 (Ω; R2 )) ∩ L1 (0, T∗ ; L2− 2 (Ω; R2 )) ,→ L1+c(δ) (0, T∗ ; L2−δ (Ω; R2 ))

for any δ ∈ (0, 1) and some c(δ) > 0. By observing that η0 ∈ W 2,2 (Ω) ⊂ L∞ (Ω), we
can apply Lemma 3.5 to deduce that

η ∈ L∞ (0, T∗ ; L2−δ (Ω))∩L1+c(δ) (0, T∗ ; W 1,2−δ (Ω)) for any δ ∈ (0, 1) and some c(δ) > 0.

This implies furthermore that

ηu ∈ L2−δ (0, T∗ ; L2 (Ω; R2 )) ∩ L2 (0, T∗ ; L2−δ (Ω; R2 )) for any δ ∈ (0, 1).

Applying Lemma 3.5 once more gives

η ∈ L∞ (0, T∗ ; L2 (Ω))∩L2−δ (0, T∗ ; W 1,2 (Ω))∩L2 (0, T∗ ; W 1,2−δ (Ω)) for any δ ∈ (0, 1).

This implies
1
ηu ∈ L2 (0, T∗ ; L2−δ (Ω; R2 ))∩L1+c(δ) (0, T∗ ; L δ (Ω; R2 )) for any δ ∈ (0, 1) and some c(δ) > 0.

Again by Lemma 3.5 we deduce

η ∈ L∞ (0, T∗ ; Lr (Ω))∩L1+c(r) (0, T∗ ; W 1,r (Ω)) for any r ∈ (1, ∞) and some c(r) > 0.

Again by the bound on u and Sobolev embedding, we have

ηu ∈ L2 (0, T∗ ; Lr (Ω; R2 )) for any r ∈ (1, ∞).

Finally, one more application of Lemma 3.5 implies (4.10).


Similarly, for the extra stress tensor, we have the following improved estimates.
Proposition 4.2. For any r ∈ (1, ∞), there holds

(4.11) kTkL∞ (0,T∗ ;Lr (Ω)) + kTkL2 (0,T∗ ;W 1,r (Ω)) ≤ C.

Proof. We rewrite the equation for T as the following equation for each of its
components Ti,j , 1 ≤ i, j ≤ 2:
k 1
(4.12) ∂t Ti,j − ε∆x Ti,j = −divx (u Ti,j ) + ∇x u T + T ∇T

x u i,j + ηδi,j − Ti,j .
2λ 2λ
Let
k 1
hi,j := ∇x u T + T ∇T

x u i,j + ηδi,j − Ti,j .
2λ 2λ

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 569

By (4.5) and (4.9), Sobolev embedding gives

∇x u T + T ∇T 2 1 1+c(δ)
(0, T∗ ; L2−δ (Ω))

x u i,j ∈ L (0, T∗ ; L (Ω)) ∩ L
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(4.13)
for any δ ∈ (0, 1) and some c(δ) > 0.

Then by the estimates in (4.9), (4.10), and (4.13), we have


(4.14)
hi,j ∈ L2 (0, T∗ ; L1 (Ω)) ∩ L1+c(δ) (0, T∗ ; L2−δ (Ω)) for any δ ∈ (0, 1) and some c(δ) > 0.

By applying the Bogovskiı̆ operator (Lemma 3.6), there exists an

Hi,j ∈ L1+c(δ) (0, T∗ ; W01,2−δ (Ω; R2 ))

such that
Z
1
(4.15) divx Hi,j = hi,j − hhi,j i with hhi,j i(t) := hi,j (t, x)dx ∈ L2 (0, T∗ ).
|Ω| Ω

By Sobolev embedding, there holds

(4.16) kHi,j k 1 ≤ C for any δ ∈ (0, 1) and some c(δ) > 0.


L1+c(δ) (0,T∗ ;L δ (Ω))

Thus, we can rewrite (4.12) as

(4.17) ∂t Ti,j − ε∆x Ti,j = divx (−u Ti,j + Hi,j ) + hhi,j i.

Let T̃ < T∗ , where T̃ can be arbitrarily close to T∗ . Recall that there holds the
estimates in (4.3) over the time interval (0, T̃ ], so that the right-hand side of (4.17)
has good integrability, for example, in L∞ (0, T̃ ; Lr (Ω)) for any r ∈ (1, ∞). Now we
consider (4.17) in the restricted time interval (0, T̃ ).
(1) (2)
Let Ti,j and Ti,j be the solutions to the equations

(1) (1)
(4.18) ∂t Ti,j − ε∆x Ti,j = divx (−u Ti,j + Hi,j )

and
(2) (2)
(4.19) ∂t Ti,j − ε∆x Ti,j = hhi,j i,

respectively, subject to the Neumann boundary conditions


(1) (2)
∂n Ti,j = ∂n Ti,j = 0 on (0, T̃ ) × ∂Ω.

(1) (2)
The initial data for Ti,j and Ti,j are taken to be in Wn2,2 (Ω; Rd×d ) satisfying

(1) (2)
Ti,j (0) + Ti,j (0) = Ti,j (0) in Ω,

(1) (2)
where a simple choice is to take Ti,j (0) = Ti,j (0), Ti,j (0) = 0.
By the estimates in (4.3), the right-hand sides of (4.18) and (4.19) are both
in L∞ (0, T̃ ; Lr (Ω)) for any r ∈ (1, ∞). Hence, we can apply Lemma 3.4 to show the
existence, uniqueness, and maximal regularity for their solutions over (0, T̃ ). However,
some related estimates may blow up as T̃ → T∗ . In what follows, we will show some

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


570 YONG LU AND ZHIFEI ZHANG

estimates that are uniform as T̃ → T∗ , which implies our desired estimates (4.11) over
the time interval (0, T∗ ).
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

A direct consequence of the uniqueness of solutions to the parabolic initial-


boundary value problem (3.5) is that, for any T̃ < T∗ ,
(1) (2)
(4.20) Ti,j + Ti,j = Ti,j a.e. in (0, T̃ ) × Ω.
By (4.5) and (4.9), Sobolev embedding gives
(4.21)
1 1
u ∈ L2 (0, T∗ ; L δ (Ω)), T ∈ L2+c(δ) (0, T∗ ; L δ (Ω)) for any δ ∈ (0, 1) and some c(δ) > 0.
Thus,
1
(4.22) u Ti,j ∈ L1+c(δ) (0, T∗ ; L δ (Ω)) for any δ ∈ (0, 1) and some c(δ) > 0.
This allows us to apply Lemma 3.5, together with the estimates (4.16) and (4.22),
(1)
and with the initial regularity Ti,j (0) ∈ W 2,2 (Ω; R2×2 ) ⊂ L∞ (Ω; R2×2 ), to obtain
(4.23)
(1) (1)
kTi,j k ∞ 1 +kTi,j k 1+c(δ) 1, 1 ≤ C for any δ ∈ (0, 1) and some c(δ) > 0,
L (0,T̃ ;L (Ω))
δ L δ(0,T̃ ;W (Ω))

where the constant C depends only on the stress diffusion coefficient ε, the C 2,β
(1)
norm of ∂Ω, the initial datum norm kTi,j (0)kW 2,2 (Ω) , and the norm of the source
term k − u Ti,j + Hi,j k 1+c(δ) 1 . In particular, C is independent of T̃ and is
L δ (0,T∗ ;L (Ω))
certainly uniformly bounded as T̃ → T∗ .
(2)
For Ti,j , by the facts that hhi,j i ∈ L2 (0, T∗ ) depending only on the time variable
(2)
t and that the initial regularity Ti,j (0) ∈ W 2,2 (Ω) ⊂ W 1,r (Ω) for any 1 < r < ∞,
applying Lemma 3.4 implies
(2) (2)
(4.24) kTi,j kL∞ (0,T̃ ;W 1,r (Ω)) + kTi,j kL2 (0,T̃ ;W 2,r (Ω)) ≤ C for any 1 < r < ∞,

where the constant C is independent of T̃ and is certainly uniformly bounded as


T̃ → T∗ .
Thus, by the uniform in time estimates in (4.23) and (4.24) and by (4.20), we can
pass T̃ → T∗ to deduce that
(4.25) kTi,j kL∞ (0,T∗ ;Lr (Ω)) ≤ C for any 1 < r < ∞.
Again by (4.21), together with the new estimate (4.25), we deduce that
1
u Ti,j ∈ L2 (0, T∗ ; L δ (Ω)), ∇x u T + T ∇T 2 2−δ

x u i,j ∈ L (0, T∗ ; L (Ω)) for any δ ∈ (0, 1).

By properties of the Bogovskiı̆ operator and Sobolev embedding, we have

−u Ti,j + Hi,j ∈ L2 (0, T∗ ; Lr (Ω)) for any 1 < r < ∞.


This allows us to apply Lemma 3.5 to obtain
(1) (1)
(4.26) kTi,j kL∞ (0,T̃ ;Lr (Ω)) + kTi,j kL2 (0,T̃ ;W 1,r (Ω)) ≤ C for any 1 < r < ∞,

where again the constant C is independent of T̃ .


Hence, by (4.20), (4.24), and (4.26) and by passing T̃ → T∗ , we deduce the
estimate (4.11) and complete the proof.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 571

Remark 4.3. From the proof, we see that Propositions 4.1 and 4.2 hold for the
case z = 0, L > 0. We remark that the estimates shown in sections 4.1 and 4.2
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

depend only on the initial data; more precisely, they depend only on the norms given
in (2.1) and the norm k(η0 , T0 )kW 2,2 (Ω) ; in particular, the estimates are independent
of k%kL∞ ((0,T∗ )×Ω) .
4.3. Estimates for %|u|α with α > 2. The goal of this section is to prove, by
assuming (4.4), that
(4.27) k%|u|α kL∞ (0,T∗ ;L1 (Ω)) ≤ C < ∞ for some α > 2.
As the derivation of the a priori energy estimates in section 4.1, the idea to show
(4.27) is done by multiplying (1.2) by α|u|α−2 u for some α > 2 and integrating over
Ω. By the fact that ∂|u|α = α|u|α−2 u · ∂u and integration by parts, one can obtain
Z Z Z
α−2 α−2 d
∂t (%u) · α|u| u dx + divx (%u ⊗ u) · α|u| u dx = %|u|α dx.
Ω Ω dt Ω
By repeatedly using ∂|u|α = α|u|α−2 u · ∂u and integration by parts, through direct
calculation we deduce that
Z Z
d α
α|u|α−2 µ|∇x u|2 + ν|divx u|2 dx

%|u| dx +
dt Ω
Z Ω 
2
+ α(α − 2) µ|u|α−2 ∇x |u| + ν(divx u)|u|α−3 u · (∇x |u|) dx

(4.28) Z Z
=α p(%)divx (|u|α−2 u) dx + α (kLη + z η 2 )divx (|u|α−2 u) dx

Z ZΩ
−α T : ∇x (|u|α−2 u) dx + α %f · |u|α−2 u dx.
Ω Ω

Observing that |∇x |u|| = ||u|−1 u · ∇x u| ≤ |∇x u| and taking 2 < α ≤ 3 close to
2 such that (α − 2)ν ≤ µ2 1 implies
Z Z
1
(α − 2) ν(divx u)|u|α−3 u · (∇x |u|) dx ≤ µ|u|α−2 |∇x u|2 dx.
Ω 2 Ω

We thus deduce from (4.28) that for any t ∈ (0, T∗ ),


(4.29)
α t
Z Z Z Z
α α−2 2 0
%|u| dx + µ|u| |∇x u| dx dt ≤ %0 |u0 |α dx
Ω 2 0 Ω Ω
Z tZ Z tZ
+α p(%)divx (|u|α−2 u) dx dt0 + α (kLη + z η 2 )divx (|u|α−2 u) dx dt0
0 Ω 0 Ω
Z tZ Z tZ
−α T : ∇x (|u|α−2 u) dx dt0 + α %f · |u|α−2 u dx dt0 .
0 Ω 0 Ω
α−2
Note the fact that |divx (|u| u)| ≤ (α−1)|u|α−2 |∇x u|. Then Hölder’s inequality
implies that
(4.30)
Z tZ Z tZ
α
0
p(%)divx (|u|α−2
u) dx dt ≤ C(k%kL∞ ((0,T∗ )×Ω) ) %|u|α + |∇x u| 2 dx dt0 ,
0 Ω 0 Ω

1 By applying the Cauchy–Schwarz inequality and more precise analysis, this condition can be

relaxed to (α − 2)ν < 4µ. If there holds 8µ − ν > 0, one may choose some α > 3 in this step. We
remark that the condition 8µ − ν > 0 appears in the study of related 3D problems, for instance, in
[33].

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


572 YONG LU AND ZHIFEI ZHANG

α
where the integral related to |∇x u| 2 is uniformly bounded in t ∈ (0, T∗ ) as long as
α ≤ 4.
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

We then calculate
Z tZ
(kLη + z η 2 )divx (|u|α−2 u) dx dt0
(4.31) 0 Ω
≤ Ckη + η 2 kL∞ (0,t;L4 (Ω)) k∇x ukL2 ((0,t)×Ω) k|u|α−2 kL2 (0,t;L4 (Ω)) .
By (4.5), Proposition 4.1, and Sobolev embedding, the quantity on the right-hand
side of (4.31) is uniformly bounded in t ∈ (0, T∗ ) as long as 2(α − 2) ≤ 2, which is
equivalent to α ≤ 3.
By Proposition 4.2 and the Sobolev embedding inequality, we have
(4.32)
Z tZ
− T : ∇x (|u|α−2 u) dx dt0 ≤ CkTkL∞ (0,t;L4 (Ω) k∇x ukL2 ((0,t)×Ω) k|u|α−2 kL2 (0,t;L4 (Ω)) ,
0 Ω

which is uniformly bounded in t ∈ (0, T∗ ) provided that 2(α − 2) ≤ 2 ⇔ α ≤ 3.


Similarly,
Z tZ Z tZ
(4.33) %f · |u|α−2 u dx dt0 ≤ C %|u|α dx dt0 + C.
0 Ω 0 Ω

Summing up the estimates (4.29)–(4.33), Gronwall’s inequality gives our desired


estimate in (4.27).
4.4. Improved estimates for ∇x u. We consider v% , vη , vτ that solve the
following Dirichlet problems for the Lamé system:
(
− µ∆x v% − ν∇x divx v% = ∇x p(%) in Ω,
v% = 0 on ∂Ω,
(
− µ∆x vη − ν∇x divx vη = ∇x (kLη + zη 2 ) in Ω,
(4.34)
vη = 0 on ∂Ω,
(
− µ∆x vτ − ν∇x divx vτ = −divx T in Ω,
vτ = 0 on ∂Ω.

By the notation introduced below Lemma 3.1, we write


v% = L−1 (∇x p(%)), vη = L−1 (∇x (kLη + zη 2 )), vτ = L−1 (−divx T).
Since L−1 is a linear operator independent of the time variable t, we have, as long as
the calculation makes sense, that

∂t L−1 (v) = L−1 (∂t v).


By (4.4) and Propositions 4.1 and 4.2, the application of Lemma 3.1 gives
(4.35)
v% ∈ L∞ (0, T∗ ; W01,r (Ω; R2 )), vη ∈ L∞ (0, T∗ ; W01,r (Ω; R2 )) ∩ L2 (0, T∗ ; W 2,r (Ω; R2 )),
vτ ∈ L∞ (0, T∗ ; W01,r (Ω; R2 )) ∩ L2 (0, T∗ ; W 2,r (Ω; R2 )) for any r ∈ (1, ∞).
We then introduce
(4.36) w := u − v, v := (v% + vη + vτ )

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 573

that solves over (0, T∗ ) × Ω the system

(4.37) %∂t w − µ∆x w − ν∇x divx w = −%u · ∇x u − %∂t v,


Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

with no-slip boundary condition

(4.38) w = 0 on (0, T∗ ) × ∂Ω.

The main goal of this section is to prove the following proposition, which is
inspired by Proposition 3.2 in [32]. The proof here is more difficult and technical due
to the presence of the extra terms involving η and T.
Proposition 4.4. Under the assumption (4.4), we have for some T1 ∈ (0, T∗ )
that

w ∈ L∞ (0, T∗ ; W01,2 (Ω; R2 )) ∩ L2 (0, T∗ ; W 1,r (Ω; R2 )) ∩ L2 (T1 , T∗ ; W 2,2 (Ω; R2 ))

for any r ∈ (1, ∞).


Proof. By (4.3) and (4.35), there holds, for any T1 ∈ (0, T∗ ), that

w ∈ L∞ (0, T1 ; W01,2 ) ∩ L2 (0, T1 ; W 1,r )(Ω; R2 ) for any r ∈ (1, ∞).

Thus, it is sufficient to prove for some T1 ∈ (0, T∗ ), which shall be fixed later on close
to T∗ , that
(4.39)
w ∈ L∞ (T1 , T∗ ; W01,2 )∩L2 (T1 , T∗ ; W 1,r )∩L2 (T1 , T∗ ; W 2,2 )(Ω; R2 ) for any r ∈ (1, ∞).

From (4.37), we deduce by direct calculation that, for any t ∈ (0, T∗ ),


Z Z Z
1 d
%|∂t w|2 dx + µ|∇x w|2 dx ≤ %|∂t w| |u · ∇x u − ∂t v| dx
Ω 2 dt Ω Ω
Z Z Z
1
≤ %|∂t w|2 dx + %|u · ∇x u|2 dx + %|∂t v|2 dx.
2 Ω Ω Ω

This gives, for any T1 ∈ (0, T∗ ) and any t ∈ (T1 , T∗ ), that


(4.40)
Z Z tZ
µ|∇x w|2 (t, ·) dx + %|∂t w|2 dx dt0
Ω T1 Ω
Z Z t Z Z t Z
0
≤ 2
µ|∇x w| (T1 , ·) dx + 2 2
%|u · ∇x u| dx dt + 2 %|∂t v|2 dx dt0 .
Ω T1 Ω T1 Ω

We need to estimate the last two terms in (4.40). For the penultimate term,
by (4.27), Hölder’s inequality, Young’s inequality, and the Gagliardo–Nirenberg
inequality, we have, for some α > 2, that
(4.41)
√ 1
k %|u · ∇x u|kL2 (T1 ,t;L2 (Ω)) ≤ Ck% α |u|kL∞ (T1 ,t;Lα (Ω)) k∇x uk 2 2α
L (T1 ,t;L α−2 (Ω))
 
≤ C 1 + k∇x wk 2 2α
L (T1 ,t;L α−2 (Ω))
 
≤ C 1 + k∇x wkθL2 (T1 ,t;L2 (Ω)) k∇2x wk1−θ
L2 (T1 ,t;L2 (Ω))

≤ C + Cδ k∇x wkL2 (T1 ,t;L2 (Ω)) + δk∇2x wkL2 (T1 ,t;L2 (Ω))

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


574 YONG LU AND ZHIFEI ZHANG

for some θ ∈ (0, 1) determined by α, for any δ > 0, and for some Cδ > 0.
We now estimate the L2 norm of ∂t v = ∂t v% + ∂t vη + ∂t vτ . For ∂t v% , we have
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

∂t v% = ∂t L−1 (∇x p(%)) = L−1 (∇x (p0 (%)∂t %)) = −L−1 (∇x (p0 (%)divx (%u)))
= −L−1 (∇x divx (p(%)u)) − L−1 (∇x ((p0 (%)% − p(%))divx u)).
By (4.4), applying Lemma 3.1 gives
(4.42) k∂t v% kL2 (0,t;L2 (Ω)) ≤ Ck∇x ukL2 (0,t;L2 (Ω)) ≤ C.
It is much more complicated to estimate ∂t vη . By the equation in η, we have

∂t vη = ∂t L−1 (∇x (kLη + zη 2 )) = kLL−1 (∇x (∂t η)) + 2zL−1 (∇x (η∂t η))
= kLL−1 (∇x (−divx (ηu) + ε∆x η)) + 2zL−1 (∇x (−ηdivx (ηu) + εη∆x η))
= kLL−1 (∇x (−divx (ηu + ε∇x η)))
+ zL−1 (∇x (−divx (η 2 u))) + zL−1 (∇x (−η 2 divx u)) + 2zεL−1 (∇x (η∆x η)).
By Lemma 3.1, (4.5), Proposition 4.1, and Sobolev embedding, we have

kL−1 (∇x (−divx (ηu + ε∇x η)))kL2 (0,t;L2 (Ω)) ≤ Ckηu + ∇x ηkL2 (0,t;L2 (Ω))
≤ CkηkL∞ (0,t;L4 (Ω)) kukL2 (0,t;L4 (Ω)) + CkηkL2 (0,t;W 1,2 (Ω)) ≤ C.
Similarly,

kL−1 (∇x (−divx (η 2 u)))kL2 (0,t;L2 (Ω)) ≤ Ckη 2 ukL2 (0,t;L2 (Ω)) ≤ C,
kL−1 (∇x (−η 2 divx u))kL2 (0,t;L2 (Ω)) ≤ Ckη 2 divx uk 3 ≤ C.
L2 (0,t;L 2 (Ω))

Then for ∂t vη it is left to estimate L−1 (∇x (η∆x η)), which is the most difficult term
to estimate. Observing the fact that
1 1
(∆x η 2 − |∇x η|2 ) = divx (η∇x η) − |∇x η|2
η∆x η =
2 2
and applying Lemma 3.1 and Proposition 4.1 gives that

kL−1 (∇x (η∆x η))kL2 (T1 ,t;L2 (Ω))


1
≤ kL−1 (∇x divx (η∇x η))kL2 (T1 ,t;L2 (Ω)) + kL−1 (∇x (|∇x η|2 ))kL2 (T1 ,t;L2 (Ω))
2
≤ Ckη∇x ηkL2 (T1 ,t;L2 (Ω)) + Ck|∇x η|2 k 2 3
L (T1 ,t;L (Ω))
2

≤C+ Ck∇x ηk2L4 (T1 ,t;L3 (Ω)) .


We benefit from the parabolic equation (1.3) in η, by Lemma 3.5, to obtain the
control

k∇x ηkL4 (T1 ,t;L3 (Ω)) ≤ CkηukL4 (T1 ,t;L3 (Ω)) ≤ CkηkL∞ (T1 ,t;L12 (Ω)) kukL4 (T1 ,t;L4 (Ω))
1 1
≤ CkukL2 ∞ (T1 ,t;L2 (Ω)) k∇x ukL2 2 (T1 ,t;L2 (Ω)) ,
where we used the Gagliardo–Nirenberg inequality. Therefore,
Z tZ
%|∂t vη |2 dx dt0 ≤ Ck∂t vη k2L2 (T1 ,t;L2 (Ω))
(4.43) T1 Ω
≤ C + Ckuk2L∞ (T1 ,t;L2 (Ω)) k∇x uk2L2 (T1 ,t;L2 (Ω)) .

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 575

We remark that there is no control for kuk2L∞ (T1 ,t;L2 (Ω)) so far. We will see later on
that this term can be absorbed into a positive term on the left-hand side by using the
smallness of k∇x uk2L2 (T1 ,t;L2 (Ω)) when T1 is close to T∗ .
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

For ∂t vτ , direct calculation gives

∂t vτ = ∂t L−1 (−divx T) = L−1 divx (−∂t T)


 
−1 T
 k 1
= −L divx − Divx (u T) + ∇x u T + T ∇x u + ε∆x T + ηI− T .
2λ 2λ

Then by (4.5), Lemma 3.1, and Propositions 4.1 and 4.2, we obtain
(4.44) 
k∂t vτ kL2 (0,t;L2 (Ω)) ≤ C k|∇x u||T|k 2 4 + k(|u||T|, ∇x T, T, η)kL2 (0,t;L2 (Ω))
L (0,t;L 3 (Ω))

≤ C kukL2 (0,t;L4 (Ω)) kTkL∞ (0,t;L4 (Ω)) + k∇x ukL2 (0,t;L2 (Ω)) kTkL∞ (0,t;L4 (Ω)) + 1 ≤ C.

Using the estimates in (4.41), (4.42), (4.43), and (4.44) in (4.40) implies
(4.45)
Z Z tZ Z
2 0
µ|∇x w| (t) dx+ %|∂t w| dx dt ≤ µ|∇x w|2 (T1 ) dx+δk∇2x wkL2 (T1 ,t;L2 (Ω))
2
Ω T1 Ω Ω
+ Cδ k∇x wkL2 (T1 ,t;L2 (Ω)) + Ckuk2L∞ (T1 ,t;L2 (Ω)) k∇x uk2L2 (T1 ,t;L2 (Ω)) + C.

By rewriting (4.37) as the Lamé system

(4.46) − µ∆x w − ν∇x divx w = −%∂t w − %u · ∇x u − %∂t v,

supplemented with the no-slip boundary condition (4.38), we can apply Lemma 3.1
to obtain

(4.47) k∇2x wkL2 (T1 ,t;L2 (Ω)) ≤ Ck%∂t w + %u · ∇x u + %∂t vkL2 (T1 ,t;L2 (Ω)) .

Again by the estimates in (4.41), (4.42), (4.43), and (4.44), we deduce from (4.47)
that

k∇2x wkL2 (T1 ,t;L2 (Ω)) ≤ Ck %∂t wkkL2 (T1 ,t;L2 (Ω)) + Cδ k∇x wkL2 (T1 ,t;L2 (Ω))
(4.48)
+ Cδk∇2x wkL2 (T1 ,t;L2 (Ω)) + Ckuk2L∞ (T1 ,t;L2 (Ω)) k∇x uk2L2 (T1 ,t;L2 (Ω)) + C.

By choosing δ > 0 small enough such that Cδ ≤ 1/2, we deduce from (4.48) that

k∇2x wkL2 (T1 ,t;L2 (Ω)) ≤ Ck %∂t wkkL2 (T1 ,t;L2 (Ω)) + Ck∇x wkL2 (T1 ,t;L2 (Ω))
(4.49)
+ Ckuk2L∞ (T1 ,t;L2 (Ω)) k∇x uk2L2 (T1 ,t;L2 (Ω)) + C.

Then plugging (4.49) into (4.45) and choosing δ > 0 sufficiently small (fixed) implies

1 t
Z Z Z Z
2 2 0
µ|∇x w| (t, ·) dx + %|∂t w| dx dt ≤ µ|∇x w|2 (T1 , ·) dx
Ω 2 T1 Ω Ω
(4.50) Z tZ
+C |∇x w|2 dx dt0 + Ckuk2L∞ (T1 ,t;L2 (Ω)) k∇x uk2L2 (T1 ,t;L2 (Ω)) + C.
T1 Ω

By (4.35) and Poincaré’s inequality, we have



(4.51) kukL∞ (T1 ,t;L2 (Ω)) ≤ Ck∇x ukL∞ (T1 ,t;L2 (Ω)) ≤ C 1 + k∇x wkL∞ (T1 ,t;L2 (Ω)) .

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


576 YONG LU AND ZHIFEI ZHANG

Combining (4.51) with (4.50), we obtain for any t ∈ (T1 , T∗ ) that


Z Z t Z Z
1
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

µ|∇x w|2 (t, ·) dx + %|∂t w|2 dx dt0 ≤ µ|∇x w|2 (T1 , ·) dx


Ω 2 T1 Ω Ω
(4.52) Z t Z
+C |∇x w|2 dx dt0 + Ck∇wk2L∞ (T1 ,t;L2 (Ω)) k∇x uk2L2 (T1 ,t;L2 (Ω)) + C.
T1 Ω

This implies that the nonnegative quantity

ξ(t) := k∇x wk2L∞ (T1 ,t;L2 (Ω))

satisfies, for any t ∈ (T1 , T∗ ),


Z t
(4.53) ξ(t) ≤ ξ(T1 ) + C ξ(t0 )dt0 + Cξ(t)k∇x uk2L2 (T1 ,T∗ ;L2 (Ω)) + C.
T1

Since k∇x ukL2 (0,T∗ ;L2 (Ω)) < ∞, there holds

k∇x ukL2 (T1 ,T∗ ;L2 (Ω)) → 0 as T1 → T∗ .

Thus, by choosing T1 ∈ (0, T∗ ) close to T∗ such that Ck∇x uk2L2 (T1 ,T∗ ;L2 (Ω)) ≤ 1/2, we
deduce from (4.53) that
Z t
ξ(t) ≤ 2ξ(T1 ) + C ξ(t0 )dt0 + C.
T1

Gronwall’s inequality implies for any t ∈ (T1 , T∗ ) that

ξ(t) ≤ eC(t−T1 ) 2ξ(T1 ) + C =⇒ k∇x wkL∞ (T1 ,t;L2 (Ω)) ≤ C.



(4.54)

Combining the estimates in (4.52) and (4.54) gives, for any t ∈ (T1 , T∗ ), that

(4.55) k∇x wkL∞ (T1 ,t;L2 (Ω)) + k %∂t wkkL2 (T1 ,t;L2 (Ω)) ≤ C.

Combining (4.55) with (4.48) and (4.51), we obtain

(4.56) k∇2x wkL2 (T1 ,t;L2 (Ω)) ≤ C for any t ∈ [T1 , T∗ ).

By (4.55) and (4.56), using Sobolev embedding, we thus obtain our desired
estimates in (4.39). The proof is then completed.
By the estimates in (4.3) and (4.35), a direct corollary of Proposition 4.4 is the
following.
Proposition 4.5. Under the assumption (4.4), we have

u ∈ L∞ (0, T∗ ; W01,2 ) ∩ L2 (0, T∗ ; W 1,r ) for any r ∈ (1, ∞).

4.5. End of the proof. We are now ready to prove the L∞ bound on η and T.
We rewrite (1.3) as

∂t η − ε∆x η = −∇x η · u − ηdivx u.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 577

By Propositions 4.1 and 4.5, we have for any r ∈ (1, ∞) that

k − ∇x η · u − ηdivx ukL2 (0,T∗ ;Lr (Ω))


Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

≤ k∇x ηkL2 (0,T∗ ;L2r (Ω)) kukL∞ (0,T∗ ;L2r (Ω))


+ kdivx ukL2 (0,T∗ ;L2r (Ω)) kηkL∞ (0,T∗ ;L2r (Ω)) ≤ C.
This allows us to apply Lemma 3.4 with p = 2, q = r to deduce that

(4.57) kηkL∞ (0,T∗ ;W 1,r (Ω)) + k∂t ηkL2 (0,T∗ ;Lr (Ω)) + kηkL2 (0,T∗ ;W 2,r (Ω)) ≤ C.

This implies, by choosing r > 2 and the Sobolev embedding theorem, that

kηkL∞ (0,T∗ ;L∞ (Ω)) ≤ C < ∞,

while for T, a similar argument implies

kTkL∞ (0,T∗ ;L∞ (Ω)) ≤ C < ∞.

We have thus obtained our desired estimate (4.2) and finished the proof of
Theorem 2.6.
5. Relative entropy. To prove the weak-strong uniqueness stated in Theorem
2.8, in the same spirit as the study in [19, 14, 16] for the compressible Navier–
Stokes equations, we introduce a proper relative entropy and build a relative entropy
inequality, for which a consequence is the weak-strong uniqueness through tedious
analysis.
First, we introduce a suitable relative entropy for our compressible Oldroyd–B
model. Based on the relative entropy used in [19, 14] for the compressible Navier–
Stokes equations, some modifications related to the additional terms in η, T need to
be done. The modifications are not a direct result of analyzing the a priori energy
estimate (2.7). For example, the term tr(T) on the left-hand side of the energy
estimate (2.7) has a sign due to the positive definite property of T, but tr(T − T̃) has
no sign given two positive definite matrices T and T̃. We will see later that we do not
include T − T̃ in our definition of the relative entropy.
For notational convenience, we define
a γ
s , G(s) := kLs log s + z s2 ∀s ∈ [0, ∞)

(5.1) H(s) :=
γ−1
satisfying
(5.2)
H 0 (s)s − H(s) = p(s), G0 (s)s − G(s) = q(s), H 00 (s) = p0 (s)/s, G00 (s) = q 0 (s)/s,

where p(s) = asγ and q(s) = kLs + z s2 denote the fluid pressure and the polymer
pressure functions, respectively
Now we introduce the relative entropy. Let (%, u, η, T) be a finite-energy weak
solution in the sense of Definition 2.1 and obtained in Theorem 2.2. Let %̃, ũ, η̃, T̃ be
the so-called relative functions which have sufficient regularity. Define the following
two relative entropies:
(5.3) Z
1
E1 (t) = E1 (%, u, %̃, ũ)(t) := %|u − ũ|2 + (H(%) − H(%̃) − H 0 (%̃)(% − %̃)) (t, ·) dx,
Ω 2
Z
E2 (t) = E2 (η, η̃)(t) := (G(η) − G(η̃) − G0 (η̃)(η − η̃)) (t, ·) dx.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


578 YONG LU AND ZHIFEI ZHANG

Remark 5.1. The relative entropy E1 is the same as that in [19, 14]. The new
one, E2 , is built in a similar manner. We remark that the extra stress tensor T is
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

not included in the relative entropies. One reason, explained above, is that tr (T − T̃)
has no sign. Another reason is that we do not want the remainder term R in the
relative entropy inequality, shown later in Proposition 5.3, to become too massy. This
is enough to show the weak-strong uniqueness. Indeed, as we shall see in section 6,
based on the relative entropy inequality obtained in this section, together with an
L2 -type estimate for T − T̃, we can derive the weak-strong uniqueness.
We prove some properties that we will use for the quantities appearing in the
relative entropies.
Lemma 5.2. There exist δ > 0, c > 0 depending only on a and γ such that for
any %, %̃ ≥ 0,
(
0 c%̃γ−2 (% − %̃)2 if δ %̃ ≤ % ≤ δ −1 %̃,
(5.4) H(%) − H(%̃) − H (%̃)(% − %̃) ≥
c max{%γ , %̃γ } otherwise.

For any η, η̃ ≥ 0, there holds

kL(η − η̃)2

if η ≤ 2η̃,



(5.5) G(η) − G(η̃) − G0 (η̃)(η − η̃) ≥ 2z(η − η̃)2 + 2η̃
 kLη

if η ≥ 2η̃.

4
Proof. We recall that a > 0, γ > 1. We use the definition of H in (5.1) to obtain
a aγ γ−1
(5.6) H(%) − H(%̃) − H 0 (%̃)(% − %̃) = %γ − %̃ % + a%̃γ .
γ−1 γ−1
Let 0 < δ ≤ 1/2. We suppose that % ≤ δ %̃. Then
a
H(%) − H(%̃) − H 0 (%̃)(% − %̃) = a%̃γ + %γ − γ %̃γ−1 %

γ−1
(5.7)  
γ a γ γ−1
 γ 1 γ
≥ a%̃ + (δ %̃) − γ %̃ (δ %̃) = a%̃ 1 + (δ − δγ) ,
γ−1 γ−1

where we have used the fact that the function f (%) := %γ − γ %̃γ−1 % is decreasing for
% ∈ [0, %̃]. The limit limδ→0 (δ γ − δγ) = 0 implies that there exists some δ ∈ (0, 12 )
depending only on γ such that
1 1
1+ (δ γ − δγ) ≥ .
γ−1 2
Combining this with (5.7), we obtain

(5.8) H(%) − H(%̃) − H 0 (%̃)(% − %̃) ≥ a%̃γ /2 if % ≤ δ %̃.

We then consider the case % ≥ δ −1 %̃. By (5.6), we have


 
a γ
H(%) − H(%̃) − H 0 (%̃)(% − %̃) = %γ + a %̃γ − %̃γ−1 %
γ−1 γ−1
 
a γ a
%γ + a (δ%)γ − (δ%)γ−1 % = %γ 1 + (γ − 1)δ γ − γδ γ−1 ,


γ−1 γ−1 γ−1

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 579
γ
where we used the fact that the function g(%̃) := %̃γ − γ−1 %̃γ−1 % is decreasing for
%̃ ∈ [0, %]. The limit limδ→0 (γ − 1)δ γ − γδ γ−1 = 0 implies for some δ > 0 small and
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

determined by γ that
1
1 + (γ − 1)δ γ − γδ γ−1 ≥ .
2

Thus, for such a fixed δ,

a
(5.9) H(%) − H(%̃) − H 0 (%̃)(% − %̃) ≥ %γ if % ≥ δ −1 %̃.
2(γ − 1)

Now we consider the case δ %̃ ≤ % ≤ δ −1 %̃. By Taylor’s formula, direct calculation


gives

H(%) − H(%̃) − H 0 (%̃)(% − %̃) = H 00 (%̂)(% − %̃)2 = aγ %̂γ−2 (% − %̃)2

for some %̂ between % and %̃. Thus,


(5.10)
H(%) − H(%̃) − H 0 (%̃)(% − %̃) ≥ aγ min{δ γ−2 , δ 2−γ }%̃γ−2 (% − %̃)2 if δ %̃ ≤ % ≤ δ −1 %̃.

Summing up the estimates in (5.8), (5.9), and (5.10) gives our desired result (5.4).
By Taylor’s formula, we have

G(η) − G(η̃) − G0 (η̃)(η − η̃) = 2z(η − η̃)2 + kL η̂ −1 (η − η̃)2

for some η̂ between η and η̃. This implies directly our desired estimate (5.5).
We now state the relative entropy inequality in our setting.
Proposition 5.3. Let T > 0, and let Ω ⊂ R2 be a C 2,β domain with β ∈ (0, 1).
Let (%, u, η, T) be a finite-energy weak solution in the sense of Definition 2.1 and
obtained in Theorem 2.2. Let %̃, ũ, η̃ be smooth functions in (t, x) ∈ [0, T ] × Ω with
constraints

ũ = 0 on [0, T ] × ∂Ω, inf %̃ > 0, inf η̃ > 0.


(0,T )×Ω (0,T )×Ω

Then there holds the following relative entropy inequality: for a.e. t ∈ (0, T ],
(5.11)
Z tZ
2
E1 (%, u, %̃, ũ)(t) + E2 (η, η̃)(t) + µ |∇x (u − ũ)| + ν|divx (u − ũ)|2 dx dt0
0 Ω
Z tZ
1 1
+ 2ε 2kL|∇x (η 2 − η̃ 2 )|2 + z |∇x (η − η̃)|2 dx dt0
0 Ω
Z t
≤ E1 (%0 , u0 , %̃0 , ũ0 ) + E2 (η0 , η̃0 ) + R(t0 ) dt0 ,
0

where (%0 , u0 , %̃0 , ũ0 , η0 , η̃0 ) denotes the corresponding initial values and R(t) =

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


580 YONG LU AND ZHIFEI ZHANG

P5
j=1 Rj (t) with
(5.12) Z
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

R1 (t) := %(∂t ũ + u · ∇x ũ) · (ũ − u) dx



Z Z
+ µ∇x ũ : ∇x (ũ − u) + νdivx ũ divx (ũ − u) dx + %f · (u − ũ) dx
ZΩ Z Ω
+ (%̃ − %)∂t H 0 (%̃) + (%̃ũ − %u) · ∇x H 0 (%̃) dx + divx ũ(p(%̃) − p(%)) dx,
ZΩ ZΩ
R2 (t) := (η̃ − η)∂t G0 (η̃) + (η̃ ũ − ηu) · ∇x G0 (η̃) dx + divx ũ(q(η̃) − q(η)) dx,
Ω Ω
Z
1 1 1 1 1 1 1
R3 (t) := −4εkL ∇x η̃ 2 · ∇x (η 2 − η̃ 2 ) + ∇x η 2 · ∇x η̃ 2 (1 − η̃ − 2 η 2 ) dx,
Z Ω
R4 (t) := −2εz ∇x η̃ · ∇x (η − η̃) dx,

Z
R5 (t) := T : ∇x (ũ − u) dx.

Proof of Proposition 5.3. We calculate

Z Z Z Z
1 1 1
(5.13) %|u − ũ|2 dx = %|u|2 dx − %u · ũ dx + %|ũ|2 dx,
Ω 2 Ω 2 Ω Ω 2

where for the second and third terms we have, by taking ũ as a test function in the
weak formulation of the momentum equation (2.4) and taking 21 |ũ|2 as a test function
in the weak formulation of the continuity equation (2.2), that
(5.14)
Z Z Z tZ

%u · ũ dx = %0 u0 · ũ0 dx + %u · ∂t ũ + (%u ⊗ u) : ∇x ũ + p(%) divx ũ
Ω Ω 0 Ω
+ q(η) divx ũ − (µ∇x u : ∇x ũ + νdivx u divx ũ) − T : ∇x ũ + % f · ũ dx dt0


and

Z Z Z tZ
1 1
(5.15) %|ũ|2 dx = %0 |ũ0 |2 dx + %ũ · ∂t ũ + %u · ∇x ũ · ũ dx dt0 .
Ω 2 Ω 2 0 Ω

Similarly, testing the continuity equation by H 0 (%̃) gives

Z Z Z tZ
(5.16) %H 0 (%̃) dx = %0 H 0 (%̃0 ) dx + %∂t H 0 (%̃) + %u · ∇x H 0 (%̃) dx dt0 .
Ω Ω 0 Ω

Plugging (5.14) and (5.15) into (5.13) and using (5.16) and the energy inequality

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 581

(2.7) gives
(5.17)
Z tZ
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Z
1 0 2
2
µ |∇x (u − ũ)| + ν|divx (u − ũ)|2 dx dt0

%|u − ũ| + H(%) − H (%̃)% dx +
Ω 2 0 Ω
Z Z tZ Z tZ
1 1 1
+ G(η)+ tr (T) dx+2ε 2kL|∇x η 2 |2 +z |∇x η|2 dx dt0 + tr (T) dx dt0
Ω 2 0 Ω 4λ 0 Ω
Z Z Z tZ
1 1 k
%0 |u0 − ũ0 |2 + H(%0 )−H 0 (%̃0 )%0 dx+ G(η0 )+ tr (T0 ) dx+ η dx dt0


Ω 2 Ω 2 2λ 0 Ω
Z tZ Z tZ
+ % f · (u − ũ) dx dt0 + µ∇x ũ : ∇x (ũ − u) + νdivx ũ div(ũ − u) dx dt0
0 Ω 0 Ω
Z tZ
%u · (∂t ũ + u · ∇x ũ) + p(%) divx ũ + q(η) divx ũ − T : ∇x ũ dx dt0
 

0 Ω
Z tZ Z tZ
+ %ũ · ∂t ũ + %u · ∇x ũ · ũ dx dt0 − %∂t H 0 (%̃) + %u · ∇x H 0 (%̃) dx dt0 .
0 Ω 0 Ω

By (5.2), we have

%∂t H 0 (%̃) + %u · ∇x H 0 (%̃)


= %̃∂t H 0 (%̃) + %̃ũ · ∇x H 0 (%̃) + (% − %̃)∂t H 0 (%̃) + (%u − %̃ũ) · ∇x H 0 (%̃)
(5.18)
= (% − %̃)∂t H 0 (%̃) + (%u − %̃ũ) · ∇x H 0 (%̃) + ∂t p(%̃) + ũ · ∇x p(%̃)
= (% − %̃)∂t H 0 (%̃) + (%u − %̃ũ) · ∇x H 0 (%̃) + ∂t p(%̃) + divx (p(%̃)ũ) − p(%̃)divũ.

Taking the test function Y = I/2 in the weak formulation (2.5) implies
(5.19)
Z Z tZ Z Z tZ
1 1 0 1 k
tr(T) dx+ tr (T) dx dt = tr (T0 ) dx+ η + T : ∇x u dx dt0 .
Ω 2 4λ 0 Ω Ω 2 0 Ω 2λ
Using (5.18) and (5.19) in (5.17), together with the fact that p(%̃) = H 0 (%̃)%̃−H(%̃),
implies
(5.20)
Z Z tZ
2
E1 (t) + G(η)dx + µ |∇x (u − ũ)| + ν|divx (u − ũ)|2 dx dt0
Ω 0 Ω
Z tZ
1
+ 2ε 2kL|∇x η 2 |2 + z|∇x η|2 dx dt0
0 Ω
Z Z tZ
≤ E1 (0) + G(η0 )dx + % f · (u − ũ) dx dt0
Ω 0 Ω
Z tZ Z tZ
+ µ∇x ũ : ∇x (ũ − u) + νdivx ũ div(ũ − u) dx dt0 − q(η) divx ũ dx dt0
0 Ω 0 Ω
Z tZ Z tZ
p(%̃)−p(%) divx ũ dx dt0 − (% − %̃)∂t H 0 (%̃)+(%u − %̃ũ)·∇x H 0 (%̃) dx dt0
 
+
0 Ω 0 Ω
Z tZ Z tZ
+ %(∂t ũ + u · ∇x ũ) · (ũ − u) dx dt0 + T : ∇x (u − ũ) dx dt0 .
0 Ω 0 Ω
0
Now we include E2 = E2 (η, η̃). We take G (η̃) as a test function in (2.3) to obtain
(5.21)
Z Z Z tZ
0 0
η∂t G0 (η̃)+ηu·∇x G0 (η̃)−ε∇x η·∇x G0 (η̃) dx dt0 .
 
ηG (η̃) dx = η0 G (η̃0 ) dx+
Ω Ω 0 Ω

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


582 YONG LU AND ZHIFEI ZHANG

We deduce from (5.2) that

η∂t G0 (η̃) + ηu · ∇x G0 (η̃)


Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

= η̃∂t G0 (η̃) + η̃ ũ · ∇x G0 (η̃) + (η − η̃)∂t G0 (η̃) + (ηu − η̃ ũ) · ∇x G0 (η̃)


(5.22)
= (η − η̃)∂t G0 (η̃) + (ηu − η̃ ũ) · ∇x G0 (η̃) + ∂t q(η̃) + ũ · ∇x q(η̃)
= (η − η̃)∂t G0 (η̃) + (ηu − η̃ ũ) · ∇x G0 (η̃) + ∂t q(η̃) + divx (q(η̃)ũ) − q(η̃)divũ.

By (5.20), (5.21), and (5.22) and the fact q(η̃) = G0 (η̃)η̃ − G(η̃), we have
(5.23)
Z tZ
2
E1 (t) + E2 (t) + µ |∇x (u − ũ)| + ν|divx (u − ũ)|2 dx dt0
0 Ω
Z tZ
1
+ 2ε 2kL|∇x η 2 |2 + z |∇x η|2 dx dt0
0 Ω
Z tZ
≤ E1 (0) + E2 (0) + % f · (u − ũ) dx dt0
0 Ω
Z tZ
+ µ∇x ũ : ∇x (ũ − u) + νdivx ũ div(ũ − u) dx dt0
0 Ω
Z tZ Z tZ
p(%̃)−p(%) divx ũ dxdt0 − (% − %̃)∂t H 0 (%̃)+(%u − %̃ũ)·∇x H 0 (%̃) dx dt0
 
+
0 Ω 0 Ω
Z tZ Z tZ
0
(η − η̃)∂t G0 (η̃)+(ηu − η̃ ũ)·∇x G0 (η̃) dx dt0
 
+ q(η̃)−q(η) divx ũ dx dt −
0 Ω 0 Ω
Z tZ Z tZ
0
+ %(∂t ũ + u · ∇x ũ) · (ũ − u) dx dt + T : ∇x (u − ũ) dx dt0
0 Ω 0 Ω
Z tZ
+ ε∇x η · ∇x G0 (η̃) dx dt0 .
0 Ω

By (5.1), there holds

(5.24) ∇x η · ∇x G0 (η̃) = ∇x η · (kLη̃ −1 + 2z)∇x η̃ = kLη̃ −1 ∇x η · ∇x η̃ + 2z∇x η · ∇x η̃.

We then calculate
(5.25)
1 1 1 1 1 1 1 1 1 1
4|∇x η 2|2 −η̃ −1 ∇x η·∇x η̃ = 4|∇x (η 2−η̃ 2 )|2+4∇x η̃ 2 ·∇x (η 2−η̃ 2 )+4∇x η 2 ·∇x η̃ 2(1−η 2 η̃ − 2 )

and

(5.26) |∇x η|2 − ∇x η · ∇x η̃ = |∇x (η − η̃)|2 + ∇x η̃ · ∇x (η − η̃).

Finally, plugging (5.24)–(5.26) into (5.23) gives our desired inequality (5.11). The
proof is completed.
Remark 5.4. By the proof of Proposition 5.3, we see that the regularity con-
straints on (%̃, ũ, η̃) can be relaxed accordingly, as long as all the integrals in (5.11)
make sense.
6. Weak-strong uniqueness. This section is devoted to proving Theorem 2.8.
We shall employ the relative entropy inequality shown in the last section to achieve
our goal. Let (%, u, η, T) be the finite-energy weak solution obtained in Theorem
2.2, and let (%̃, ũ, η̃, T̃) be the strong solution obtained in Theorem 2.3 with the same

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 583

initial data satisfying the assumptions in Theorem 2.3 and the lower bound constraint
(2.13). Then for any T < T∗ , by the continuity equation (1.1), we have
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

RT
inf %̃ ≥ e− 0
kdivx ũ(t)kL∞ (Ω) dt
inf %0 > 0,
[0,T ]×Ω Ω
(6.1) RT
inf η̃ ≥ e− 0
kdivx ũ(t)kL∞ (Ω) dt
inf η0 > 0.
[0,T ]×Ω Ω

Let T < T∗ be arbitrary and fixed. In the rest of this section we restrict t ∈
[0, T ]. Thus, by (6.1), we can choose (%̃, ũ, η̃) as the relative functions in the entropy
inequality (5.11). We then analyze the corresponding right-hand side of (5.11) until
we can use Gronwall-type inequalities to show the relative entropy is identically zero,
which implies that the weak solution and the strong solution are equal. This is done
step by step in the rest of this section.
6.1. A new expression for the remainder. Since (%̃, ũ, η̃, T̃) is the strong
solution to (1.1)–(1.9) satisfying (6.1), we have
(6.2)
∂t ũ + ũ · ∇x ũ + %̃−1 ∇x p(%̃) − %̃−1 (µ∆x ũ + ν∇x divx ũ) = %̃−1 divx T̃ − %̃−1 ∇x q(η̃) + f .
Plugging (6.2) into R1 in (5.12), together with the fact that %̃−1 ∇x p(%̃) = ∇x H 0 (%̃),
gives
Z Z
R1 = %(u − ũ) · ∇x ũ · (ũ − u) dx + (µ∆x ũ + ν∇x divx ũ)(%̃−1 % − 1) · (ũ − u) dx
Z Ω Z Ω
Z
− %∇x H (%̃) · (ũ − u) dx− %%̃ ∇x q(η̃) · (ũ − u) dx+ %%̃−1 divx T̃ · (ũ − u) dx
0 −1

ZΩ Ω
Z Ω

+ (%̃ − %)∂t H 0 (%̃) + (%̃ũ − %u) · ∇x H 0 (%̃) dx + divx ũ(p(%̃) − p(%)) dx.
Ω Ω

By the continuity equation (1.1), we have


(6.3)
− %∇x H 0 (%̃) · (ũ − u) + (%̃ − %)∂t H 0 (%̃) + (%̃ũ − %u) · ∇x H 0 (%̃)
= (%̃ − %) (∂t H 0 (%̃) + ũ · ∇x H 0 (%̃))
= (%̃ − %) [∂t H 0 (%̃) + divx (ũH 0 (%̃)) + (H 00 (%̃)%̃ − H 0 (%̃))divx ũ] − (%̃ − %)H 00 (%̃)%̃divx ũ
= −(%̃ − %)p0 (%̃)divx ũ.
P6
Then we can write R1 := j=1 R1,j with
Z
R1,1 := %(u − ũ) · ∇x ũ · (ũ − u) dx,
ZΩ
R1,2 := (µ∆x ũ + ν∇x divx ũ)%̃−1 (% − %̃) · (ũ − u) dx,
ZΩ
divx ũ p(%̃) − p(%) − p0 (%̃)(%̃ − %) dx,

R1,3 :=
(6.4) ZΩ Z
R1,4 := %̃−1 (%̃ − %)∇x q(η̃) · (ũ − u) dx + %̃−1 (% − %̃)divx T̃ · (ũ − u) dx,
Ω Ω
Z Z
R1,5 := − ∇x q(η̃) · (ũ − u) dx = − η̃∇x G0 (η̃) · (ũ − u) dx,
Ω Ω
Z Z
R1,6 := divx T̃ · (ũ − u) dx = − T̃ : ∇x (ũ − u) dx.
Ω Ω

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


584 YONG LU AND ZHIFEI ZHANG

By (5.12), we have
(6.5) Z Z
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

R1,5 + R2 = (η̃ − η) (∂t G0 (η̃) + ũ · ∇x G0 (η̃)) dx + divx ũ(q(η̃) − q(η)) dx,


Ω Ω
Z
R1,6 + R5 = (T − T̃) : ∇x (ũ − u) dx.

By (1.3) and similar calculations as in (6.3),


(6.6) Z Z
divx ũ q(η̃) − q(η) − q 0 (η̃)(η̃ − η) dx + ε (η̃ − η)η̃ −1 q 0 (η̃)∆x η̃ dx

R1,5 + R2 =
Ω Ω
Z
−1
= R2,1 + ε (η̃ − η)(kLη̃ + 2z)∆x η̃ dx,

where
Z
divx ũ q(η̃) − q(η) − q 0 (η̃)(η̃ − η) dx.

(6.7) R2,1 :=

Thus, direct calculation gives


(6.8) Z
R1,5 + R2 + R3 + R4 = R2,1 + εkL (η̃ − η)η̃ −1 ∆x η̃ dx

Z
1 1 1 1 1 1 1
− 4εkL ∇x η̃ 2 · ∇x (η 2 − η̃ 2 ) + ∇x η 2 · ∇x η̃ 2 (1 − η̃ − 2 η 2 ) dx

Z
1 1 1 1 1 1 1 1
= R2,1 + εkL 4η̃ − 2 (η̃ 2 − η 2 )∇x η̃ 2 · ∇x (η̃ 2 − η 2 ) − η̃ −1 ∆x η̃(η 2 − η̃ 2 )2 dx.

Summarizing the calculations in (6.4)–(6.8), we obtain a new expression for the


remainder R:
(6.9) Z Z
R= %(u − ũ) · ∇x ũ · (ũ − u) dx + (µ∆x ũ + ν∇x divx ũ)%̃−1 (% − %̃) · (ũ − u) dx
Ω Ω
Z Z
0
+ divx ũ p(%̃) − p(%) − p (%̃)(%̃ − %) dx+ divx ũ q(η̃) − q(η) − q 0 (η̃)(η̃ − η) dx
 

ZΩ Z Ω
−1 −1
+ %̃ (%̃ − %)∇x q(η̃) · (ũ − u) dx + %̃ (% − %̃)divx T̃ · (ũ − u) dx
Ω Ω
Z
1 1 1 1 1 1 1 1
+ εkL 4η̃ − 2 (η̃ 2 − η 2 )∇x η̃ 2 · ∇x (η̃ 2 − η 2 ) − η̃ −1 ∆x η̃(η 2 − η̃ 2 )2 dx

Z
+ (T − T̃) : ∇x (ũ − u) dx.

6.2. Estimate for the remainder. We estimate the right-hand side of (6.9)
term by term. For notational convenience, let ζ(t) be a universal nonnegative
integrable function in L1 (0, T ); its value may differ from line to line.
By (2.10) and Sobolev embedding W 2,6 (Ω) ⊂ W 1,∞ (Ω), we have ∇x ũ ∈
L (0, T ; L∞ (Ω)). Thus,
2

Z Z
(6.10) %(u − ũ) · ∇x ũ · (ũ − u) dx ≤ k∇x ũ(t)kL (Ω)
∞ %|u − ũ|2 dx ≤ ζ(t)E1 (t).
Ω Ω

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 585

By the fact that ∇x ũ ∈ L2 (0, T ; L∞ (Ω)) and

p(%̃) − p(%) − p0 (%̃)(%̃ − %) = (γ − 1)−1 H(%̃) − H(%) − H 0 (%̃)(%̃ − %) ,


 
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

q(η̃) − q(η) − q 0 (η̃)(η̃ − η) = 2z(η̃ − η)2 ≤ G(η) − G(η̃) − G0 (η̃)(η − η̃),

we thus have
(6.11)
Z Z
divx ũ p(%̃) − p(%) − p0 (%̃)(%̃ − %) dx + divx ũ q(η̃) − q(η) − q 0 (η̃)(η̃ − η) dx
 
Ω Ω

≤ ζ(t) E1 (t) + E2 (t) .

By (2.10) (or by the argument of proving (4.57) using Lemma 3.4), we have, by
Sobolev embedding, for any r ∈ (2, ∞),

η̃ ∈ L∞ (0, T ; W 1,r (Ω)) ∩ L2 (0, T ; W 2,r (Ω)) ⊂ L∞ (0, T ; L∞ (Ω)) ∩ L2 (0, T ; W 1,∞ (Ω)).

Combining the estimates for η̃ with the lower bound for η̃ in (6.1), we have
(6.12)
Z
1 1 1 1 1 1
εkL 4η̃ − 2 (η̃ 2 − η 2 )∇x η̃ 2 · ∇x (η̃ 2 − η 2 ) dx

1 1 1 1
≤ 2εkLkη̃ −1 (t)kL∞ (Ω) k∇x η̃(t)kL∞ (Ω) k(η̃ 2 − η 2 )(t)kL2 (Ω) k∇x (η̃ 2 − η 2 )(t)kL2 (Ω)
1 1 1 1
≤ 4εkLkη̃ −1 (t)k2L∞ (Ω) k∇x η̃(t)k2L∞ (Ω) k(η̃ 2 −η 2 )(t)k2L2 (Ω) +εkLk∇x (η̃ 2 −η 2 )(t)k2L2 (Ω)
1 1 1 1
≤ ζ(t)k(η̃ 2 − η 2 )(t)k2L2 (Ω) + εkLk∇x (η̃ 2 − η 2 )(t)k2L2 (Ω) .

By (5.5), we have the bound


1 1
(6.13) (η̃ 2 − η 2 )2 ≤ 4 (G(η) − G(η̃) − G0 (η̃)(η − η̃)) ,

which is actually uniform in η, η̃ ∈ (0, ∞). Combining (6.13) with (6.12), we have
(6.14)
Z
1 1 1 1 1 1 1 1
εkL 4η̃ − 2 (η̃ 2 − η 2 )∇x η̃ 2 · ∇x (η̃ 2 − η 2 ) dx ≤ ζ(t)E2 (t) + εkLk∇x (η̃ 2 − η 2 )(t)k2L2 (Ω) .

By (6.13), we have the estimate


(6.15) Z
1 1
εkL −η̃ −1 ∆x η̃(η 2 − η̃ 2 )2 dx

Z Z
1 1 1 1 1 1
≤ εkL η̃ −2 |∇x η̃|2 (η 2 − η̃ 2 )2 dx − εkL 2η̃ −1 ∇x η̃ · ∇(η 2 − η̃ 2 )(η 2 − η̃ 2 ) dx
Ω Ω
Z
1 1
−1 2 2
≤ kη̃ ∇x η̃(t)kL∞ (Ω) εkL(η 2 − η̃ 2 ) dx

1 1 1 1
−1
+ 4εkLkη̃ (t)∇x η̃(t)k2L∞ (Ω) k(η̃ 2 − η 2 )(t)k2L2 (Ω) + εkLk∇x (η̃ 2 − η 2 )(t)k2L2 (Ω)
1 1
≤ ζ(t)E2 (t) + εkLk∇x (η̃ 2 − η 2 )(t)k2L2 (Ω) .

We now consider
Z
(6.16) %̃−1 (%̃ − %)∇x q(η̃) · (ũ − u) dx = I1 + I2 + I3 ,

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


586 YONG LU AND ZHIFEI ZHANG

where, for δ be chosen as in Lemma 5.2,


Z
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

I1 := %̃−1 (%̃ − %)∇x q(η̃) · (ũ − u) dx,


δ %̃≤%≤δ −1 %̃
Z
(6.17) I2 := %̃−1 (%̃ − %)∇x q(η̃) · (ũ − u) dx,
%≥δ −1 %̃
Z
I3 := %̃−1 (%̃ − %)∇x q(η̃) · (ũ − u) dx.
%≤δ %̃

By (5.4), (2.10), the lower bound of %̃ in (6.1), and Sobolev embedding, we have
for some σ > 0 small that

I1 ≤ C(σ)k∇x q(η̃)k2L3 (Ω) k(% − %̃)k2L2 (δ%̃≤%≤δ−1 %̃) + σk(ũ − u)k2L6 (Ω)
(6.18)
Z
µ
≤ ζ(t) H(%) − H(%̃) − H 0 (%̃)(% − %̃) dx + k∇(ũ − u)k2L2 (Ω) .
−1
δ %̃≤%≤δ %̃ 16

Similarly, for I2 we have


(6.19) Z Z Z
I2 ≤ C |∇x q(η̃)|%|ũ − u| dx ≤ C %|∇x q(η̃)|2 dx + C %|ũ − u|2 dx
%≥δ −1 %̃ %≥δ −1 %̃ Ω
Z Z
≤ C|∇x q(η̃)|2L∞ (Ω) %γ dx + C %|ũ − u|2 dx
%≥δ −1 %̃ Ω
Z Z
≤ ζ(t) H(%) − H(%̃) − H 0 (%̃)(% − %̃) dx + C %|ũ − u|2 dx.
%≥δ −1 %̃ Ω

Finally for I3 we have


(6.20) Z
I3 ≤ C |∇x q(η̃)| |(ũ − u)| dx ≤ C(σ)k∇x q(η̃)k2L3 (Ω) k1k2L2 (%≤δ%̃) + σk(ũ − u)k2L6 (Ω)
%≤δ %̃
Z
µ
≤ ζ(t) H(%) − H(%̃) − H 0 (%̃)(% − %̃) dx + k∇(ũ − u)k2L2 (Ω) .
%≤δ %̃ 16

Summing up (6.18)–(6.20), we obtain


Z
µ
(6.21) %̃−1 (%̃ − %)∇x q(η̃) · (ũ − u) dx ≤ ζ(t)E1 (t) + k∇(ũ − u)k2L2 (Ω) .
Ω 8

A similar argument implies


Z
µ
(6.22) %̃−1 (% − %̃)divx T̃ · (ũ − u) dx ≤ ζ(t)E1 (t) + k∇(ũ − u)k2L2 (Ω)
Ω 8

and
(6.23)
Z
µ
(µ∆x ũ + ν∇x divx ũ)%̃−1 (% − %̃) · (ũ − u) dx ≤ ζ(t)E1 (t) + k∇(ũ − u)k2L2 (Ω) .
Ω 8

We remark that, unlike ∇x q(η̃) or divx T̃, we do not have control over the
L2 (0, T ; L∞ (Ω)) norm of ∆x ũ in Theorem 2.3. Thus further steps need to be taken

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 587

concerning the estimate (6.23); precisely, we need to modify the estimate of the
following term compared to (6.19):
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Z
I4 := (µ∆x ũ + ν∇x divx ũ)%̃−1 (% − %̃) · (ũ − u) dx.
%≥δ −1 %̃

Indeed, we have
Z
I4 ≤ C %|µ∆x ũ + ν∇x divx ũ||ũ − u| dx.
%≥δ −1 %̃

For the case γ ≥ 2, by the lower bound on %̃ in (6.1), Lemma 5.2, and Sobolev
embedding, we have
(6.24) Z
γ
I4 ≤ C % 2 |µ∆x ũ + ν∇x divx ũ||ũ − u| dx
%≥δ −1 %̃
γ
≤ Ck% 2 kL2 (%≥δ−1 %̃) k∇2x ũkL3 (Ω) kũ − ukL6 (Ω)
γ µ
≤ Ck% 2 k2L2 (%≥δ−1 %̃) k∇2x ũk2L3 + k∇x (ũ − u)k2L2 (Ω)
Z 16
2 2 µ
= Ck∇x ũkL3 % dx + k∇x (ũ − u)k2L2 (Ω)
γ
−1
%≥δ %̃ 16
Z
µ
≤ Ck∇2x ũk2L3 H(%) − H(%̃) − H 0 (%̃)(% − %̃) dx + k∇x (ũ − u)k2L2 (Ω) .
−1
%≥δ %̃ 16
For the case 1 < γ ≤ 2, by Lemma 5.2, Sobolev embedding, and the fact that
% ∈ L∞ (0, T ; Lγ (Ω)), we have
(6.25)
I4 ≤ Ck%kLγ (%≥δ−1 %̃) k∇2x ũk 2γ kũ − uk 2γ
L γ−1 (Ω) L γ−1 (Ω)
µ
≤ Ck%k2Lγ (%≥δ−1 %̃) k∇2x ũk2 2γ + k∇x (ũ − u)k2L2 (Ω)
L γ−1 (Ω) 16
γ 2 2 µ
≤ Ck%kLγ (%≥δ−1 %̃) k∇x ũk 2γ + k∇x (ũ − u)k2L2 (Ω)
L γ−1 (Ω) 16
Z
µ
≤ Ck∇2x ũk2 2γ H(%) − H(%̃) − H 0 (%̃)(% − %̃) dx + k∇x (ũ − u)k2L2 (Ω) .
L γ−1 (Ω) %≥δ −1 %̃ 16
Then, by the estimate on ∇2x ũ in (2.10), we deduce from (6.24) and (6.25) that
Z
µ
I4 ≤ ζ(t) H(%) − H(%̃) − H 0 (%̃)(% − %̃) dx + k∇x (ũ − u)k2L2 (Ω)
−1
%≥δ %̃ 16
(6.26)
µ
≤ ζ(t)E1 (t) + k∇(ũ − u)k2L2 (Ω) .
16
Hölder’s inequality implies
Z Z
µ
(6.27) (T − T̃) : ∇ x (ũ − u) dx ≤ C |T − T̃|2 dx + k∇x (ũ − u)k2L2 (Ω) .
Ω Ω 8
Finally, summarizing the estimates in (6.10), (6.11), (6.14), (6.15), (6.21), (6.22),
(6.23), and (6.27), we deduce from (6.9) that
(6.28) Z
µ 1 1
R(t) ≤ ζ(t)(E1 +E2 )(t)+ k∇(ũ−u)k2L2 (Ω) +2εkLk∇x (η̃ 2 −η 2 )k2L2 (Ω) +C |T− T̃|2 dx.
2 Ω
2
R
It is left to deal with Ω |T − T̃| dx.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


588 YONG LU AND ZHIFEI ZHANG

6.3. End of the proof. First, since the initial data are assumed to be regular
enough as in Theorem 2.3, we can employ Propositions 4.1 and 4.2 to obtain better
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

estimates for η and T:

(6.29) (η, T) ∈ L∞ (0, T∗ ; Lr (Ω)) ∩ L2 (0, T∗ ; W 1,r (Ω)) for any r ∈ (1, ∞).

This allows us to take T as a test function in the weak formulation (2.5). We can also
take the strong solution T̃ as a test function in (2.5). Then by using the equation in
T̃, through tedious but rather direct calculations, we obtain
Z Z tZ Z tZ
1 1 0
2
|T − T̃| dx + 2
|T − T̃| dx dt + ε |∇x (T − T̃)|2 dx dt0
Ω 2 2λ 0 Ω 0 Ω
Z tZ
Divx (u − ũ)T + Divx ũ(T − T̃) : (T − T̃) dx dt0
  
=−
0 Ω
Z tZ
k t
Z Z
0
∇x (u − ũ)T+∇T (η − η̃) tr (T − T̃) dx dt0 .
 
+ x ũ(T − T̃) : (T − T̃) dx dt +
0 Ω 2λ 0 Ω

We use (6.29) to get the estimate


(6.30)
Z Z
 
− Divx (u − ũ)T : (T − T̃) dx = − Divx (u − ũ)T + (u − ũ) · ∇x T : (T − T̃) dx
Ω Ω

≤ kdivx (u − ũ)kL2 (Ω) kTkL∞ (Ω) + ku − ũkL6 (Ω) k∇x TkL3 (Ω) kT − T̃kL2 (Ω)
µ
≤ ζ(t)kT − T̃k2L2 (Ω) + k∇x (u − ũ)k2L2 (Ω) .
8
Similarly as in (6.30), we have
Z Z
1
(divx ũ)|T− T̃|2 dx ≤ kdivx ũkL∞ (Ω) kT− T̃k2L2 (Ω) .

− Divx ũ(T− T̃) : (T− T̃) dx = −
Ω 2 Ω
The other terms can be estimated similarly. At last, we arrive at
Z Z tZ
1
|T − T̃|2 dx + ε |∇x (T − T̃)|2 dx dt0
(6.31) Ω 2 0 Ω
k µ
≤ ζ(t)kT − T̃kL2 (Ω) + kη − η̃k2L2 (Ω) + k∇x (u − ũ)k2L2 (Ω) .
2λ 4
Denote
Z
1
(6.32) E(t) := E1 (t) + E2 (t) + |T − T̃|2 (t, ·) dx.
Ω 2

Thus, by the estimates (6.28) and (6.31), by Proposition 5.3, we derive for any t ∈
(0, T ] that
Z tZ
µ 1 1
E(t)+ |∇x (u − ũ)|2 +2εkL|∇x (η̃ 2 −η 2 )|2 +2εz|∇x (η̃−η)|2 +ε|∇x (T− T̃)|2 dx dt0
0 Ω 4
Z t
≤ ζ(t0 )E(t0 ) dt0 for some ζ(t) ∈ L1 (0, T ).
0

Gronwall’s inequality gives E(t) ≡ 0 for t ∈ [0, T ] for any T ∈ (0, T∗ ), which implies
the weak-strong uniqueness (2.14). The proof of Theorem 2.8 is completed.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


COMPRESSIBLE OLDROYD–B MODEL 589

6.4. Conditional regularity. Finally, we prove Theorem 2.9 concerning the


conditional regularity for weak solutions. This is indeed a consequence of the refined
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

blow-up criterion and the weak-strong uniqueness. By Theorem 2.3, we let (%̃, ũ, η̃, T̃)
be the strong solution, with T∗ the maximal existence time issued from the same
initial data as for the weak solution (%, u, η, T).
We first show that T < T∗ . For contradiction we assume that T ≥ T∗ . Then for
any T1 < T∗ , since %̃ and η̃ have positive lower bounds over [0, T1 ] × Ω (see (6.1)), we
can apply Theorem 2.8 to derive that the weak solution coincides with the strong one
over [0, T1 ]. This implies that

sup %̃ = sup % ≤ sup % < ∞,


[0,T1 ]×Ω [0,T1 ]×Ω [0,T )×Ω

where the upper bound is uniform as T1 → T∗ . Thus,

sup %̃ ≤ sup % < ∞.


[0,T∗ )×Ω [0,T )×Ω

This implies, by Theorem 2.6, that T∗ = ∞, which contradicts the assumption that
T ≥ T∗ .
Now we have T < T∗ . We can choose T1 such that T ≤ T1 < T∗ . Since %̃ and η̃
have positive lower bounds over [0, T1 ] × Ω (see (6.1)), we can apply Theorem 2.8 to
deduce that the weak solution coincides with the strong solution over [0, T1 ] ⊃ [0, T ].
The proof of Theorem 2.9 is then completed.
Acknowledgments. The authors gratefully acknowledge the anonymous refer-
ees for their valuable comments.

REFERENCES

[1] S. Agmon, A. Douglis, and L. Nirenberg, Estimates near the boundary for solutions of
elliptic partial differential equations satisfying general boundary conditions, Comm. Pure
Appl. Math., 12 (1959), pp. 623–727.
[2] J. W. Barrett and S. Boyaval, Existence and approximation of a (regularized) Oldroyd–B
model, Math. Models Methods Appl. Sci., 21 (2011), pp. 1783–1837.
[3] J. W. Barrett, Y. Lu, and E. Süli, Existence of large-data finite-energy global weak solutions
to a compressible Oldroyd–B model, Comm. Math. Sci., 15 (2017), pp. 1265–1323.
[4] J. W. Barrett and E. Süli, Existence and equilibration of global weak solutions to kinetic
models for dilute polymers I: Finitely extensible nonlinear bead-spring chains, Math.
Models Methods Appl. Sci., 21 (2011), pp. 1211–1289.
[5] J. W. Barrett and E. Süli, Existence and equilibration of global weak solutions to kinetic
models for dilute polymers II: Hookean-type bead-spring chains, Math. Models Methods
Appl. Sci., 22 (2012), 1150024.
[6] J. W. Barrett and E. Süli, Existence of global weak solutions to finitely extensible nonlinear
bead-spring chain models for dilute polymers with variable density and viscosity, J.
Differential Equations, 253 (2012), pp. 3610–3677.
[7] J. W. Barrett and E. Süli, Existence of global weak solutions to compressible isentropic
finitely extensible nonlinear bead-spring chain models for dilute polymers, Math. Models
Methods Appl. Sci., 26 (2016), pp. 469–568.
[8] J. W. Barrett and E. Süli, Existence of global weak solutions to compressible isentropic
finitely extensible nonlinear bead-spring chain models for dilute polymers: The two-
dimensional case, J. Differential Equations, 261 (2016), pp. 592–626.
[9] M. E. Bogovskiı̆, Solution of some vector analysis problems connected with operators div and
grad, Trudy Sem. S. L. Soboleva, 80 (1980), pp. 5–40 (in Russian).
[10] J.-Y. Chemin and N. Masmoudi, About lifespan of regular solutions of equations related to
viscoelastic fluids, SIAM J. Math. Anal., 33 (2001), pp. 84–112, https://doi.org/10.1137/
S0036141099359317.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.


590 YONG LU AND ZHIFEI ZHANG

[11] J. H. Choe and H. Kim, Strong solutions of the Navier–Stokes equations for isentropic
compressible fluids, J. Differential Equations, 190 (2003), pp. 504–523.
[12] P. Constantin and M. Kliegl, Note on global regularity for two-dimensional Oldroyd–B fluids
Downloaded 02/21/18 to 130.64.11.153. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

with diffusive stress, Arch. Rational Mech. Anal., 206 (2012), pp. 725–740.
[13] D. Fang and R. Zi, Strong solutions of 3D compressible Oldroyd–B fluids, Math. Methods
Appl. Sci., 36 (2013), pp. 1423–1439.
[14] E. Feireisl, B. Jin, and A. Novotný, Relative entropies, suitable weak solutions, and weak-
strong uniqueness for the compressible Navier–Stokes system, J. Math. Fluid Mech., 14
(2012), pp. 717–730.
[15] E. Feireisl and A. Novotný, Singular Limits in Thermodynamics of Viscous Fluids, Adv.
Math. Fluid Mech., Birkhäuser Verlag, Basel, 2009.
[16] E. Feireisl, A. Novotný, and Y. Sun, A regularity criterion for the weak solutions to the
Navier–Stokes–Fourier system, Arch. Rational Mech. Anal., 212 (2014), pp. 219–239.
[17] E. Fernández-Cara, F. Guillén, and R. R. Ortega, Mathematical modeling and analysis
of viscoelastic fluids of the Oldroyd kind, in Handbook of Numerical Analysis, Vol. VIII.
Solution of Equations in Rn (Part 4). Techniques of Scientific Computing (Part 4).
Numerical Methods for Fluids (Part 2), P. G. Ciarlet and J. L. Lions eds., Handbook
of Numerical Analysis VIII. North-Holland, Amsterdam, 2002, pp. 543–661.
[18] G. P. Galdi, An Introduction to the Mathematical Theory of the Navier–Stokes Equations.
Steady-state problems, 2nd ed., Springer Monogr. Math., Springer, New York, 2011.
[19] P. Germain, Weak-strong uniqueness for the isentropic compressible Navier–Stokes system, J.
Math. Fluid Mech., 13 (2010), pp. 137–146.
[20] C. Guillopé and J. C. Saut, Existence results for the flow of viscoelastic fluids with a
differential constitutive law, Nonlinear Anal. Theory Methods Appl., 15 (1990), pp. 849–
869.
[21] C. Guillopé and J. C. Saut, Global existence and one-dimensional nonlinear stability of
shearing motions of viscoelastic fluids of Oldroyd type, RAIRO Math. Model. Numer.
Anal., 24 (1990), pp. 369–401.
[22] D. Hoff, Global solutions of the Navier–Stokes equations for multidimensional compressible
flow with discontinuous initial data, J. Differential Equations, 120 (1995), pp. 215–254.
[23] X. Hu and D. Wang, Strong solutions to the three-dimensional compressible viscoelastic fluids,
J. Differential Equations, 252 (2012), pp. 4027–4067.
[24] X. Hu and G. Wu, Global existence and optimal decay rates for three-dimensional compressible
viscoelastic flows, SIAM J. Math. Anal., 45 (2013), pp. 2815–2833, https://doi.org/10.
1137/120892350.
[25] Z. Lei, Global existence of classical solutions for some Oldroyd–B model via the incompressible
limit, Chinese Ann. Math. Ser. B, 27 (2006), pp. 565–580.
[26] C. Le Bris and T. Lelièvre, Micro-macro models for viscoelastic fluids: Modelling,
mathematics and numerics, Sci. China Math., 55 (2012), pp. 353–384.
[27] P. L. Lions and N. Masmoudi, Global solutions for some Oldroyd models of non-Newtonian
flows, Chin. Ann. Math. Ser. B, 21 (2000), pp. 131–146.
[28] A. Novotný and I. Straškraba, Introduction to the Mathematical Theory of Compressible
Flow, Oxford Lecture Ser. Math. Appl. 27, Oxford University Press, Oxford, UK, 2004.
[29] J. Qian and Z. Zhang, Global well-posedness for compressible viscoelastic fluids near
equilibrium, Arch. Rational Mech. Anal., 198 (2010), pp. 835–868.
[30] M. Renardy, Local existence of solutions of the Dirichlet initial-boundary value problem for
incompressible hypoelastic materials, SIAM J. Math. Anal., 21 (1990), pp. 1369–1385,
https://doi.org/10.1137/0521076.
[31] V. A. Solonnikov, The solvability of the initial-boundary value problem for the equations of
motion of a viscous compressible fluid, J. Soviet Math., 14 (1980), pp. 1120–1133.
[32] Y. Sun and Z. Zhang, A blow-up criterion of strong solution for the 2D compressible Navier–
Stokes equations, Sci. China Math., 54 (2011), pp. 106–116.
[33] Y. Sun, C. Wang, and Z. Zhang, A Beale–Kato–Majda blow-up criterion for the 3-D
compressible Navier–Stokes equations, J. Math. Pures Appl., 95 (2011), pp. 36–47.
[34] A. Valli, Periodic and stationary solutions for compressible Navier–Stokes equations via a
stability method, Ann. Sc. Norm. Sup. Pisa Cl. Sci., 10 (1983), pp. 607–647.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

You might also like