A PRIORI ESTIMATES FOR THE LINEARIZED RELATIVISTIC EULER

EQUATIONS WITH A PHYSICAL VACUUM BOUNDARY AND AN IDEAL

GAS EQUATION OF STATE

BRIAN B. LUCZAK∗,#
arXiv:2411.13726v1 [math.AP] 20 Nov 2024

Abstract. In this paper, we will provide a result on the relativistic Euler equations for an ideal
gas equation of state and a physical vacuum boundary. More specifically, we will prove a priori
estimates for the linearized system in weighted Sobolev spaces. Our focus will be on choosing the
correct thermodynamic variables, developing a weighted book-keeping scheme, and then proving
energy estimates for the linearized system.

Keywords: Relativistic Euler equations, physical vacuum boundary, ideal gas equation of state.
Mathematics Subject Classification (2020): Primary: 35Q75; Secondary: 35Q35, 35Q31,
35R37,

Contents
1. Introduction 2
1.1. Physical vacuum boundary with a barotropic equation of state 3
1.2. The case of an ideal gas 4
1.3. Main Result 5
1.4. The linearized system in terms of entropy and sound speed (squared) 6
1.5. Function spaces and Energies 9
1.6. The basic energy estimate 10
2. Creating a Book-keeping scheme 15
2.1. Defining order for free-boundary terms 17
2.2. Solving for time derivatives 20
3. Estimates for the transport energy 25
3.1. Entropy estimates 25
3.2. Vorticity estimates 26
4. Elliptic and Div-Curl Estimates 31
4.1. Elliptic estimates for r̃ 31
4.2. Elliptic estimates for ũ 34
4.3. Higher order commutators with the convective derivative 37
4.4. Commutators with weighted spatial derivatives 39
5. Energy Equivalence 46
6. Higher order wave energy estimates 48
7. Main Theorem 52
8. Appendix for Lemma 2.14 53
References 56

∗
Vanderbilt University, Nashville, TN, USA. brian.b.luczak@vanderbilt.edu.
#
BBL gratefully acknowledges support from NSF grant DMS-2107701,

1
2 A priori estimates for the linearized system

1. Introduction
The starting point for our work is the relativistic Euler equations which describe the motion
of a relativistic fluid in a Minkowski background (see [3] for historical context). The equations of
motion are given by

uµ ∂µ ϱ + (p + ϱ)∂µ uµ = 0, (1.1a)
µ α αµ
(p + ϱ)u ∂µ u + Π ∂µ p = 0, (1.1b)
µ ν
gµν u u = −1, (1.1c)

where ϱ is the fluid’s (energy) density, u is the fluid’s (four-)velocity, p is the fluid’s pressure given
by an equation of state (whose choice depends on the nature of the fluid, see below), g is the
Minkowski metric1, and Παβ := gαβ + uα uβ is the projection onto the space orthogonal to u. Above
and throughout, we employ standard Cartesian coordinates {xα }3α=0 with t := x0 denoting a time
coordinate and x := (x1 , x2 , x3 ) denoting spatial coordinates, the sum convention is adopted, Greek
indices vary from 0 to 3 and Latin indices from 1 to 3, and indices are raised and lowered with g.

Remark 1.1. We note that (1.1a) is the conservation of energy for the fluid, (1.1b) is the conser-
vation of momentum, and (1.1c) is a normalization condition where the velocity is assumed to be
a forward time-like vector field (and this constraint is propagated by the flow).

We are interested in the case where the fluid is confined within a domain that is not fixed but
is allowed to move with the fluid motion, such as, e.g., the motion of a star. Fluids of this type
are called free-boundary fluids. We further consider the situation where the pressure and density
vanish on the boundary which is often referred to as the gas case. The liquid case where the density
does not vanish on the boundary is a different problem.
Denote the region containing the fluid at time t by Ωt . Then, Ωt can be described as

Ωt = {(τ, x) ∈ R1+3 | τ = t, ϱ(t, x) > 0}.

Equations (1.1) hold in the spacetime region

[
D := {t} × Ωt , (1.2)
0≤t<T

for some T > 0, known as the moving domain. The fluid’s free-boundary at time t is defined as
Γt := ∂Ωt , and the fluid’s free boundary, also called the moving boundary, free interface, or vacuum
boundary, is defined as
[ [
Γ := {t} × Γt = {t} × ∂Ωt .
0≤t<T 0≤t<T

Sometimes we also call Ωt and Γt the moving domain and the free boundary, respectively. In free-
boundary problems, understanding the dynamics of Γ is crucial, as it is the fact that Γt moves with
time that distinguishes such problems from a standard initial-boundary value problem where the
boundary of the domain is fixed. Note that, according to the foregoing, we have

p = ϱ = 0 on Γ. (1.3)

1This problem can be studied for a general background metric, but the Minkowski metric already contains all the
important mathematical features. Coupling to Einstein’s equations, on the other hand, is an entirely different (and
much harder) problem.
 Luczak 3

1.1. Physical vacuum boundary with a barotropic equation of state. Before discussing
our work for a non-barotropic equation of state, let us consider the barotropic setting which will
motivate our discussion. Consider the equation of state:
p = p(ϱ) = ϱκ+1 , κ > 0, (1.4)
where κ is constant. Equation (1.1b) then becomes
(ϱκ+1 + ϱ)uµ ∂µ uα + c2s Παµ ∂µ ϱ = 0, (1.5)
where
c2s := p′ (ϱ) = (κ + 1)ϱκ (1.6)
is the fluid’s sound speed [3] and the second equality in (1.6) is valid for the equation of state (1.4).
Since ϱ vanishes on the free-boundary, we also have
c2s Γ
= 0. (1.7)
Remark 1.2. Observe that (1.7) follows immediately from (1.3) and (1.6). But typically one would
still impose (1.7) for other equations of state than (1.4) when (1.3) holds, since sound waves should
not be allowed to propagate into vacuum across a region where the medium (the density) vanishes.
The decay rate of c2s near the free-boundary plays a key role in this problem. One needs to
consider decay rates that allow for Γt to move with a bounded non-zero acceleration. We have from
(1.5) and (1.6) that the fluid’s (four-)acceleration is
(κ + 1)ϱκ−1 αµ
aα := uµ ∂µ uα = − Π ∂µ ϱ.
1 + ϱκ
Since ϱ ∼ 0 near the free-boundary, we have that 1 + ϱ = O(1) so, near the boundary,
aα ∼ ϱκ−1 Παµ ∂µ ϱ ∼ ∂ϱκ ∼ ∂c2s ,
using ∂ to denote a generic spacetime derivative. At this point, we need to make some assumption
on the decay rate of c2s . In view of the finite speed propagation property, away from the boundary
the motion of the fluid is essentially the same as in the case without a free-boundary. Thus, a
natural scale in this problem which allows us to separate the bulk and boundary behaviors is the
distance to the boundary:
d ≡ d(t, x) = distance from x to Γt .
Therefore, a natural assumption to make on c2s is that it decays like a power of d, i.e., c2s ∼ dβ for
some β > 0. Under this assumption, the acceleration becomes
a ∼ ∂dβ = βdβ−1 ∂d ∼ dβ−1 ,
where we used that ∂d = O(1). From this, we see that

 0,
 β > 1,
a|Γt = ∞, 0 < β < 1,

finite ̸= 0, β = 1.


Hence, we expect a realistic dynamics only in the case β = 1. Therefore, we assume that
c2s (t, x) ≈ d(t, x) for x near Γt , (1.8)
i.e., c2s is comparable to the distance to the boundary. Condition (1.8) is known as the physical
vacuum boundary condition. One should understand (1.8) as a constraint, i.e., something that
is imposed on the initial data and then propagated by the flow. The local well-posedness of the
relativistic Euler equations with (1.4) as equation of state and initial data satisfying the physical
vacuum boundary condition was obtained in [4] using Eulerian coordinates. Additionally, a-priori
4 A priori estimates for the linearized system

estimates for the same problem with equation of state (1.4) were obtained in [6] using Lagrangian
coordinates. In [8], the authors proved a priori estimates using an equation of state where the
pressure is assumed to be a power law of the fluid’s baryon density n (see (1.10) below).
1.2. The case of an ideal gas. We would like to extend the results of [4] to the case of a non-
barotropic equation of state. In addition, we would like, if possible, to treat situations that are of
direct relevance to physicists working in numerical simulations of star evolution. A typical equation
of state used by physicists in their simulations of stars is that of an ideal gas (see [15] Section 2.4),
p = p(n, ε) = nε(γ − 1), (1.9)
where n is the fluid’s baryon density, ε is the fluid’s internal energy, and γ > 1 is a constant. We
recall that n satisfies the equation
uµ ∂µ n + n∂µ uµ = 0, (1.10)
and ε is defined through the relation
ϱ = n(1 + ε). (1.11)
Equation (1.10) is known as the conservation of baryon density [3]. This provides a closed system
in terms of the unknowns (ε, n, uµ ).
Since ε ≥ 0, equation (1.11) implies that n also vanishes on the free boundary. In addition,
since we must have ε = 0 in a vacuum, we also need to impose that ε = 0 on Γt . We need to
find conditions that play a similar role to the physical vacuum boundary condition (1.8) employed
for barotropic fluids. As in Section 1.1, we try to determine such conditions by ensuring a finite
non-zero boundary acceleration.
We now adopt (ε, n, u) as primitive variables. In this situation, using (1.9), equation (1.1b)
becomes
1 + εγ µ
nu ∂µ uα + εΠαµ ∂µ n + nΠαµ ∂µ ε = 0. (1.12)
γ−1
Inspired by the discussion of Section 1.1, we suppose that
n ∼ dβ , ε ∼ dσ , β > 0, σ > 0.
Using this Ansatz into equation (1.12) and proceeding as in Section 1.1, we find
a ∼ dσ−1 .
An interesting observation is that dβ cancels out and we are left with

 0,
 σ > 1,
a|Γt = ∞, 0 < σ < 1,

finite ̸= 0, σ = 1.


Therefore, we conclude that in the case of an equation of state of an ideal gas given by (1.9), the
physical vacuum boundary conditions should be
ε(t, x) ≈ d(t, x) and n(t, x) ≈ (d(t, x))β , β > 0, for x near Γt . (1.13)
Using (1.9) and (1.11), in terms of (ε, n, u), equations (1.1) and (1.10) read
uµ ∂µ ε + (γ − 1)ε∂µ uµ = 0, (1.14a)
1 + εγ µ
nu ∂µ uα + εΠαµ ∂µ n + nΠαµ ∂µ ε = 0, (1.14b)
γ−1
uµ ∂µ n + n∂µ uµ = 0, (1.14c)
µ ν
gµν u u = −1. (1.14d)
 Luczak 5

Remark 1.3. We observe that (1.13) is what we obtained solely from an analysis of the boundary
acceleration. However, from the thermodynamic relations and the equation of state (see (1.22)
below), we have that the entropy s is given by
1  ε 
s= log γ−1 + s0 .
γ−1 n
where s0 is a constant. From the physical boundary conditions, near the free boundary we would
have that  
1 d
s∼ log + s0 ∼ log(d1−β(γ−1) ) + s0 .
γ−1 d(γ−1)β
Thus, as we approach the free boundary, d → 0 and s would behave like
 1
 −∞
 , for β < γ−1
1
s → finite value , for β = γ−1 (1.15)
1

∞ , for β > γ−1 .

1
Thus, β = γ−1 is a natural condition for the decay rate of n in our problem. However, in what
follows, our analysis will be solely with respect to the variables s, u, and r, where r is a multiple of
c2s that we introduce in Section 1.4. We make this remark to indicate that there is a natural choice
for the physical vacuum boundary condition in the case of an ideal gas.
1.3. Main Result. Here, we summarize our main result which is based on the linearized version of
system (1.14) with variables s, u and r (where r is defined in (1.27) below). See (1.35) below for a
rewriting of (1.14) with our new variables, and (1.36) for the full linearized system. We remark that
Theorem 1.4 below is a statement amount the linearized problem, where the non-linear variables
(s, r, u) serve as background data.
Theorem 1.4 (Sobolev estimates for the linearized system). Let (s, r, u) be a smooth solution
to (1.35) that exists on some time interval [0, T ], and for which the physical vacuum boundary
condition (1.13) holds. Let (s̃0 , r̃0 , ũ0 ) be initial data to system (1.36). Then, there exists a constant
C depending only on s, r, u, and T such that, if (s̃, r̃, ũ) ∈ C ∞ (D) is a solution to (1.36) on [0, T ],
then
∥(s̃, r̃, ũ)∥H2k (Ωt ) ≲ C∥(s̃0 , r̃0 , ũ0 )∥H2k (Ω0 ) (1.16)
where H2k (Ωt ) is defined in (1.46).
See Theorem 7.1 in Section 7 for a proof of our main result.
Remark 1.5. In Theorem 1.4, we assume smooth solutions for simplicity. Our quantitative bounds
depend only on finite regularity norms as indicated in Theorem 1.4, and in similar theorems.
Remark 1.6. We note that free boundary fluids have been studied in a variety of contexts since
the 1930s, see [18, 14, 17] for such examples. For more recent papers, we refer the reader to
[1, 11, 16, 10, 9, 19, 20, 5, 8]. For recent work in free-boundary problems from relativity, we refer
the reader to [12, 13]. See [3] and [4] for a review of recent developments in mathematical aspects
of relativistic fluids.
Remark 1.7. (Schematic notation for derivatives). We will use the following schematic notation
when discussing derivatives in the arguments that follow. We will use the symbol
• ∂ to denote a generic spatial derivative, i.e. ∂1 , ∂2 , or ∂3 , and
• ∂ to denote a generic spacetime derivative, i.e. ∂t , ∂1 , ∂2 , or ∂3 .
In view of subsection 2.2, it is important to note that there is significant overlap between these
two symbols. Namely, one can use our primary system (1.35) to solve for the time derivative of a
given quantity in terms of the spatial derivatives in a way that, as it turns out, does not introduce
problematic terms into our estimates.
6 A priori estimates for the linearized system

1.4. The linearized system in terms of entropy and sound speed (squared). Our goal is
to rewrite (1.14) so that we can apply energy estimate techniques. For our purposes, it is more
effective to consider the unknowns (s, r, uµ ) where s denotes the entropy per particle, r is a multiple
of the sound speed (squared), and u is the usual (four)-velocity. We will often abuse language and
simply refer to r as the sound speed (squared) or just the sound speed.
We begin by rewriting (1.14) in terms of p, and then introducing our new variables which will
be used in the remainder of the work.
First, using (1.9) and (1.14a), we observe that
uµ ∂µ p = (γ − 1)nuµ ∂µ ε + (γ − 1)εuµ ∂µ n
= −(γ − 1)2 nε∂µ uµ − (γ − 1)nε∂µ uµ (1.17)
µ
= −γp∂µ u .
Then, rewriting (1.14b) in terms of ε and p, our system (1.14) takes the form
uµ ∂µ ε + (γ − 1)ε∂µ uµ = 0 (1.18a)
1 + εγ µ
pu ∂µ uα + εΠαµ ∂µ p = 0 (1.18b)
γ−1
uµ ∂µ p + γp∂µ uµ = 0. (1.18c)
In the following lines, we will make use of several identities (1.19) (1.20), and (1.21), which can
be found in Section 2 of [15].
Recall that the enthalpy per particle h is defined by
p+ϱ nε(γ − 1) + n(1 + ε))
h= = = εγ + 1 (1.19)
n n
which has been simplified using (1.9) and (1.11). Using the well known thermodynamic relation
ndh − dp = nθds, (1.20)
where θ is the temperature for the fluid, and the fact that
p = nθ (1.21)
for an ideal fluid, we can solve for s by
nγdε − (γ − 1)(ndε + εdn) = nε(γ − 1)ds
1  ε  (1.22)
=⇒ s = log γ−1 + s0
γ−1 n
where s0 is a constant (and we will often set s0 = 0 for convenience). Additionally, a quick
computation (such as in [3]) shows that (1.14c) along with the thermodynamic relations implies
uµ ∂µ s = 0. (1.23)
which is the conservation of entropy along flow lines. Moreover, (1.22) allows us to solve for ε in
ε
terms of s and p. We see that e(γ−1)s = nγ−1 1
= γ−1 · npγ , which leads to
γ−1
s
e γ γ−1
ε= γ−1 p γ (1.24)
(γ − 1) γ

γ−1
Now, substituting for ε in (1.18b) and mulitplying by (γ − 1) γ , we get
γ−1 γ−1
s
C1 puµ ∂µ uα + e γ p γ Παµ ∂µ p = 0 (1.25)
where
1 γ−1
s γ−1 γ
C1 = 1 +e γ p γ .
(γ − 1) γ γ−1
 Luczak 7

Note that C1 is O(1) as we approach the free boundary since p → 0.

Before proceeding, we can quickly verify that the acceleration of the free boundary is O(1) using
(1.13). Near the free boundary, ε ∼ d where d is the distance to the boundary. Using (1.24), we
γ
get p ∼ d γ−1 . In this case,
 γ − 1 1
−1 γ
aα = uµ ∂µ uα ∼ p γ ∂p ∼ d γ−1 d γ−1 = O(1). (1.26)
γ−1
Because, d ∼ p γ , this motivates the definition of our new variable
γ−1
r := p γ (1.27)
with the idea that r ∼ d and r will serve as our weight in the energy for the free boundary problem.
dp
After computing the sound speed squared given by c2s = dϱ |s , we further note that r is comparable
to the sound speed (squared) for the fluid and it will play a similar role as in the work from [4].
Using (1.18c) above, we can quickly verify that
γ − 1 − γ1 µ
uµ ∂µ r = p u ∂µ p = −(γ − 1)r∂µ uµ (1.28)
γ
and thus (1.18c) becomes
uµ ∂µ r + (γ − 1)r∂µ uµ = 0 (1.29)
µ
In order to complete the system in terms of (s, r, u ) we must further rewrite the second equation
1
γ
(1.25). Observing that ∂µ p = γ−1 p γ ∂µ r, we get
γ−1
s − γ1 γ−1
s γ
C1 uµ ∂µ uα + e γ p Παµ ∂µ p = C1 uµ ∂µ uα + e γ Παµ ∂µ r = 0 (1.30)
γ−1
Then, calling
γ − 1 − γ−1 s
C2 := C1 ·
e γ , (1.31)
γ
we get the form of the system in terms of the unknowns (s, r, uµ ):
uµ ∂µ s = 0 (1.32a)
1 αµ
uµ ∂µ uα + Π ∂µ r = 0 (1.32b)
C2
uµ ∂µ r + (γ − 1)r∂µ uµ = 0. (1.32c)
γ−1
(γ−1) γ
where C2 = γ−1 + r. For simplicity, let’s define
γe γ s

γ−1
(γ − 1) γ
Γ := Γ(s) = γ−1 (1.33)
s
γe γ

so that C2 = Γ(s) + r.
Finally, using the convective derivative which we denote as
Dt = uµ ∂µ , (1.34)
our new system in terms of the unknowns (s, r, uα ) takes the form
Dt s = 0 (1.35a)
µ
Dt r + (γ − 1)r∂µ u = 0 (1.35b)
1
Dt uα + Παµ ∂µ r = 0 (1.35c)
Γ+r
gαβ uα uβ = −1 (1.35d)
8 A priori estimates for the linearized system

where Γ = Γ(s) is defined according to (1.33).

Now that the form of our system is complete, we will compute the full linearized equations using
the standard procedure with linearized variables. Consider a one-parameter family of solutions
{sτ , rτ , uτ }τ for the main system (1.35) such that (sτ , rτ , uτ )|τ =0 = (s, r, u). We can now formally
d
define the function δ = dτ |τ =0 and linearized variables (s̃, r̃, ũ) := (δs, δr, δu). These variables will
solve the linearized system that we now compute by taking δ of the equations in (1.35). After a
computation, we can write the linearized equations in the form:
Dt s̃ = f (1.36a)
µ µ
Dt r̃ + ũ ∂µ r + (γ − 1)r∂µ ũ = g (1.36b)
1
Dt ũα + Παµ ∂µ r̃ = hα (1.36c)
Γ+r
ũα uα = 0 (1.36d)
where Γ is defined according to (1.33), and
f = −ũµ ∂µ s
g = −(γ − 1)r̃∂µ uµ
(1.37)
α µ 1
α Γ′ 1
h = −ũ ∂µ u − (ũα uµ + uα ũµ )∂µ r + 2
s̃Παµ ∂µ r + r̃Παµ ∂µ r
Γ+r (Γ + r) (Γ + r)2
Remark 1.8. We note that the terms in blue will later combine to form a perfect derivative
that will assist in our energy estimates. Additionally, each of the terms on the RHS contains
undifferentiated (s̃, r̃, ũ) with coefficients of the form (∂s, ∂r, ∂u).
To clarify the exposition, we will be making the following assumption which can be used when
handling the relativistic Euler equations with a physical vacuum boundary (see [4] and [6] for
similar examples).
Assumption 1.9 (Uniform smallness of r). Without loss of generality, we will assume that for
each t ∈ [0, T ], r is uniformly small in Ωt :
1
∥r∥L∞ (Ωt ) ≤ ε̂ ≪
2
where ε̂ is a small positive constant that we fix for the remainder of the paper. By ‘≪ 12 ’, we
mean, in particular, that there exists a large positive constant C such that C ε̂ < 12 . We will fix this
constant C for the remainder of the paper.
Recall from the physical vacuum boundary conditions and (1.27) that this assumption will be
verified in a small neighborhood of the free boundary Γ. Due to finite speed of propagation, the
behavior of the fluid away from the boundary is non-degenerate and a priori estimates in Sobolev
spaces can be handled using methods from the standard relativistic Euler equations. The removal
of Assumption 1.9 requires a partition of unity argument in which one must separate the fluid into
its bulk behavior away from the free boundary and its behavior near the free boundary. By using
Assumption 1.9, we are able to focus on the essential difficulties posed by the degenerate nature of
the free boundary where ∥r∥L∞ (Ωt ) ≤ ε̂ ≪ 21 .
Remark 1.10. In what follows, we will often refer to the ‘smallness’ of r near the free boundary
which will allow certain terms to make much smaller contributions. We will use ε̂ when appropriate
to reference Assumption 1.9.
Remark 1.11. Additionally, when handling these ‘small’ terms (see Remark 2.6 in Section 6), we
will often use the Cauchy-Schwarz inequality with ε. Thus, it is useful to fix the interplay between
ε̂ and ε so as to not cause any confusion. Given that ε̂ ≪ 12 is fixed by Assumption 1.9, and we
 Luczak 9

1
also have C−1 < 12 where C is the fixed positive constant from Assumption 1.9. We will use ε in
applications of the Cauchy-inequality-with-ε to be a small positive constant such that
1 1
0< <ε< (1.38)
C −1 2
Note that the previous definitions make the following inequalities true
1 1 1
0 < ε̂ < < <ε<
2C C −1 2
  (1.39)
1
1+ ε̂ < C ε̂
ε
Remark 1.12 (The symbol ‘≲’). Throughout our work, the symbol ≲ will be used in the usual
fashion, i.e.
A ≲ B ⇐⇒ A ≤ DB
where D is a constant depending on the the fixed data of the problem. In our case, D will depend on
the background solution (s, r, u) along with γ and T . By slight abuse of notation, we will sometimes
write A ≲ DB or A ≲ D(∂ l (s, r, u))B if we want to highlight, for example, that D depends on l
derivatives of our fixed data (s, r, u). This is also done to assist the reader in following our estimates
where these quantities are often removed from an integral with the L∞ norm.
1.5. Function spaces and Energies. Here, we define the function spaces and energies that will
play a key role in our work. We denote the weighted L2 spaces with weight h as L2 (h) and equip
them with the norm Z
∥f ∥2L2 (h) = h|f |2 dx. (1.40)
Ωt
We will assume throughout that r is a positive function on Ωt which vanishes simply on the free
boundary Γt . Thus, r will be comparable to the distance to Γt .
For pairs of functions in our system defined on Ωt , we will use the base Hilbert space
2−γ 1
H = L2 (r γ−1 ) × L2 (r γ−1 ) (1.41)
with the usual norm depending only on the weight r. However, we will often use an equivalent
norm which is compatible with our energies in the problem. Let Gαβ be the following two form
Gαβ := gαβ + 2uα uβ (1.42)
which, by [6], has been shown to be a Riemannian metric in spacetime.
Then, we define the norm
Z  
2
2−γ 1 2 2
∥(r̃, ũ)∥He = r γ−1 r̃ + (Γ + r)r|ũ| dx (1.43)
Ωt γ−1
where |ũ|2 = |ũ|2G = Gαβ ũα ũβ ≥ 0.
Remark 1.13. We observe that the two norms ∥ · ∥H and ∥ · ∥He are equivalent on the space of
pairs (r̃, ũ) due to a key fact about the weights appearing in the energy. For x sufficiently close to
∂Ωt , we have that r → 0, s approaches a finite value Cs by (1.15), and the weight
γ−1
(γ − 1) γ
(Γ + r) → γ−1 >0 (1.44)
Cs
γe γ

which is a finite positive constant bounded away from 0 (as γ > 1 is fixed, and Cs fixed). Hence,
∥ · ∥H and ∥ · ∥He are equivalent near the free boundary.
10 A priori estimates for the linearized system

We also define the higher order weighted Sobolev spaces H j,σ with integer j ≥ 0 and σ > − 12 to
be the space of all distributions in Ωt whose norm
X
∥f ∥2H j,σ = ∥rσ ∂ α f ∥2L2 (1.45)
|α|≤j

is finite. For higher regularity, we will use the following higher order Sobolev spaces where the
powers of r are linked to the number of derivatives. Similar to [4], we will define higher order
function spaces H2k on triplets (s̃, r̃, ũ) in Ωt with the following norm
2k k
2−γ 2−γ 2−γ
+ 12 +a α +a α + 12 +a α
X X
∥(s̃, r̃, ũ)∥2H2k = ∥r 2(γ−1) ∂ s̃∥2L2 + ∥r 2(γ−1) ∂ r̃∥2L2 + ∥r 2(γ−1) ∂ ũ∥2L2 (1.46)
|α|=0 a=0
|α|−a≤k

where we note that the r weight in the r̃ norm is 21 less than the weights on s̃ and ũ (which is a
direct consequence of the powers of r in (1.36)). Additionally, these higher order function spaces are
based around an even number of derivatives due to the underlying wave-like operator that governs
the second order evolution of (1.36). Taking Dt of (1.36) illustrates this underlying operator which
is a variable coefficient version of Dt2 − r∆.
Remark 1.14. Similar to [4] and [7], we can use embedding theorems to show that the H2k norm
1 2−γ 1
2k, 2(γ−1) +k 2k, 2(γ−1) +k 2k, 2(γ−1) +k
is equivalent to the H ×H ×H norm.

Remark 1.15. We remark that the H e norm will be used to control convective derivatives of r̃ and
2k ). Meanwhile, ũ
ũ which satisfy a system of wave equations (see (1.47) for the definition of Ewave
will be decomposed into its vorticity part which, alongside s̃, satisfies a transport equation that we
2k
can estimate directly (see (1.48) for the definition of Etransport ).
Here, we introduce energies that will be used in the higher order analysis. First, we have the
wave energy
k
∥(Dt2j r̃, Dt2j ũ)∥2He
X
2k
Ewave (s̃, r̃, ũ) = (1.47)
j=0

recalling (1.43). Note that the definition of H̃ guarantees that |Dt2j ũ|2 = |Dt2j ũ|2G ≥ 0.
Additionally, we will use the transport energy
2k
Etransport (s̃, r̃, ũ) = ∥ω̂∥2 2k−1,k+ 1 + ∥s̃∥2 2k,k+ 1 (1.48)
H 2(γ−1) H 2(γ−1)

where ω̂ is the reduced linearized vorticity defined in (3.15), and this will be addressed in Section
3.2. Then, the final linearized energy is given by
E 2k (s̃, r̃, ũ) = Ewave
2k 2k
(s̃, r̃, ũ) + Etransport (s̃, r̃, ũ). (1.49)

Remark 1.16. Although both components of Etransport 2k satisfy transport equations, we remark
that they will be used slightly differently in the arguments that follow. In particular, the estimate
for ω̂ will be used to complete div-curl estimates for ũ, whereas s̃ can be estimated directly with
the H2k norm from above and below as in Section 3.1.
1.6. The basic energy estimate. In this section, we will prove a basic energy estimate for the
quantities s̃, r̃, and ũ in weighted L2 spaces. Although the higher order estimates in Sobolev spaces
require additional techniques, this section serves as a starting point in our analysis. As discussed in
Section 1.5, the s̃ equation will later be treated separately (see section 3.1) as it satisfies a transport
equation as opposed to the wave equations satisfied by r̃ and ũ (see Section 6).
 Luczak 11
1
After computing the linearized equations (1.36), multiply (1.36a) by r γ−1 s̃, the second equation
2−γ 1
1
(1.36b) by γ−1 r γ−1 r̃ and contract the third equation (1.36c) with (Γ+r)r γ−1 Gαβ ũβ . Using (1.35d),
(1.36d), and identities like ũα Παµ = ũµ = Gαβ ũβ Παµ , we get
1 γ−1
1 1
r Dt s̃2 = −r γ−1 s̃ũµ ∂µ s (1.50a)
2
1 2−γ 1 2−γ 1 2−γ
r γ−1 Dt r̃2 + r γ−1 r̃ũµ ∂µ r + r γ−1 r̃∂µ ũµ = −r̃2 r γ−1 ∂µ uµ (1.50b)
2(γ − 1) γ−1
Γ + r γ−1
1 1 1 1
r Dt |ũ|2 + r γ−1 ũµ ∂µ r̃ = −(Γ + r)r γ−1 ũα ũµ ∂µ uα − r γ−1 |ũ|2 uµ ∂µ r
2
(1.50c)
Γ′ γ−1
1 1 1
+ r s̃ũµ ∂µ r + r γ−1 r̃ũµ ∂µ r.
Γ+r Γ+r
Thus, after adding the equations together, the blue terms combine to form a perfect derivative:
1 2−γ 1 1 1
r γ−1 r̃ũµ ∂µ r + r γ−1 r̃∂µ ũµ + r γ−1 ũµ ∂µ r̃ = ∂µ (r γ−1 r̃ũµ ) (1.51)
γ−1
which will be handled using integration be parts. Combining the equations together, we get
1 γ−1
1 1 2−γ Γ + r γ−1
1 1
r Dt s̃2 + r γ−1 Dt r̃2 + r Dt |ũ|2 + ∂µ (r γ−1 r̃ũµ )
2 2(γ − 1) 2
1
= −r γ−1 s̃ũµ ∂µ s
2−γ
− r̃2 r γ−1 ∂µ uµ
1 1
− (Γ + r)r γ−1 ũα ũµ ∂µ uα − r γ−1 |ũ|2 uµ ∂µ r
Γ′ γ−1
1 1 1
+ r s̃ũµ ∂µ r + r γ−1 r̃ũµ ∂µ r.
Γ+r Γ+r
(1.52)
Next, we integrate with respect to x over Ωt and use the moving domain formula (noting that the
i
fluid particles on the boundary move with velocity uu0 ):
Z Z Z  i
d u
f dx = ∂t f dx + ∂i f 0 dx
dt Ωt Ωt Ωt u
Z Z  i (1.53)
1 u
= D f dx +
0 t
f ∂i dx.
Ωt u Ωt u0
For the blue terms, we have r = 0 on ∂Ωt , so integrating by parts yields
Z Z Z
1 1 1
∂µ (r γ−1 r̃ũµ ) dx = ∂i (r γ−1 r̃ũi ) dx + ∂t (r γ−1 r̃ũ0 ) dx
Ωt Ωt
Z ZΩt
1 1
=− r γ−1 r̃ũi νi dS + ∂t (r γ−1 r̃ũ0 ) dx (1.54)
Z ∂Ωt Ωt
1
= ∂t (r γ−1 r̃ũ0 ) dx
Ωt

For the energy terms, we plan to apply (1.53) to the function

1 1 1 2−γ Γ + r γ−1
1
f (t, x) = r γ−1 s̃2 + r γ−1 r̃2 + r |ũ|2 (1.55)
2 2(γ − 1) 2
12 A priori estimates for the linearized system

First, recall using the r equation that Dt r = −(γ − 1)r∂µ uµ . We then get
1 1 1 2−γ Γ + r γ−1 1
Dt f (t, x) = r γ−1 Dt s̃2 + r γ−1 Dt r̃2 + r Dt |ũ|2
2 2(γ − 1) 2
1 1 2 − γ 2−γ 1 1
− r γ−1 ∂µ uµ s̃2 − r γ−1 ∂µ uµ r̃2 − r γ−1 (Γ + γr)∂µ uµ |ũ|2
2 2(γ − 1) 2
1
= −∂µ (r γ−1 r̃ũµ )
1
− r γ−1 s̃ũµ ∂µ s
2−γ
(1.56)
− r̃2 r γ−1 ∂µ uµ
1 1
− (Γ + r)r γ−1 ũα ũµ ∂µ uα − r γ−1 |ũ|2 uµ ∂µ r
Γ′ γ−1 1 1 1
+ r s̃ũµ ∂µ r + r γ−1 r̃ũµ ∂µ r
Γ+r Γ+r
1 1 2 − γ 2−γ 1 1
− r γ−1 ∂µ uµ s̃2 − r γ−1 ∂µ uµ r̃2 − r γ−1 (Γ + γr)∂µ uµ |ũ|2
2 2(γ − 1) 2
using the fact that Dt Γ = 0 and computing several more expressions.
After gathering terms, we apply (1.53) to f (t, x) and integrate over Ωt . This yields
Z Z  i
d u
f (t, x) dx = f (t, x)∂i 0
dx
dt Ωt Ωt u
Z  
1 1 γ−1 1
2 γ 2−γ 1 1
− 0
r s̃ + r γ−1 r̃ + r γ−1 (Γ + γr)|ũ| ∂µ uµ dx
2 2
u 2 2(γ − 1) 2
ZΩt
1 1
− 0
∂t (r γ−1 r̃ũ0 ) dx
u
ZΩt (1.57)
1 γ−1 1
µ
− 0
r s̃ũ ∂µ s dx
Ωt u
Z
1 1 1 1
− 0
(Γ + r)r γ−1 ũα ũµ ∂µ uα + 0 r γ−1 |ũ|2 uµ ∂µ r dx
Ω u u
Z t ′
1 Γ 1
γ−1 s̃ũµ ∂ r +
1 1 1
γ−1 r̃ũµ ∂ rdx
+ 0
r µ r µ
Ωt u Γ + r u0 Γ + r
Thus, we define the squared base energy (k = 0) to be
Z  
0 2 1 2 1 2−γ 1 2 2 2
E (s̃, r̃, ũ)[t] := ∥(r̃, ũ)∥H̃ + ∥s̃∥ 0, 1 = r γ−1 r̃ + (Γ + r)r|ũ| + rs̃ dx.
2 H 2(γ−1) 2 Ωt γ−1
(1.58)
Remark 1.17. Observe the correspondence between (1.58) and the higher order energy (1.49)
where we have a wave part and a transport part. We only make the distinction here between s̃
and (r̃, ũ) to highlight the differences in the higher order estimates. In Section 3.1, we can simply
differentiate the s̃ equation with spatial derivatives ∂ 2k . However, in Section 6, we must differentiate
the r̃ and ũ equations with the convective derivative Dt .
This will allow us to prove the following basic energy estimate.
Proposition 1.18 (Basic Energy Inequality). Let (s, r, u) be a solution to (1.35) that exists on
some time interval [0, T ]. Assume that s, r, u and their first order derivatives are bounded in the
L∞ (Ωt ) norm for each t ∈ [0, T ] and r vanishes simply on the free boundary. Then, the following
 Luczak 13

estimate holds for solutions (s̃, r̃, ũ) to (1.36):

Z t 
0 0
E [t] ≲ E [0] exp C(∥∂s, ∂r, ∂u, s, r, u∥L∞ (Ωτ ) ) dτ (1.59)
0

where E 0 [t] is given by (1.58) and C is a function depending on the L∞ (Ωτ ) norms of s, r, u and
their first order derivatives.

Before giving the proof, we provide the following remarks which are critical in understanding the
methods used throughout this paper.

Remark 1.19 (The symbol ‘≲’). Throughout our work, the symbol ≲ will be used in the usual
fashion, i.e.

A ≲ B ⇐⇒ A ≤ DB

where D is a constant depending on the the fixed data of the problem. In our case, D will depend
on the background solution (s, r, u) along with γ and T .

Remark 1.20. In view of (1.44), we will often use inequalities of the following form when handling
weights that appear in our integrals:
Z Z
1
2 1 1
r γ−1 |ũ| dx ≤ (Γ + r)r γ−1 |ũ|2 dx ≲ E 0 (1.60)
Ωt Γ+r L∞ (Ωt ) Ωt

1
where we note that Γ+r tends to a finite positive constant depending on γ when we are sufficiently
close to the boundary (see Remark 1.13). When we refer to an expression consisting of s, r, u and γ
as O(1) near the free boundary, we simply are referring to the fact that when r is sufficiently small,
this weight tends to a finite constant bounded away from zero. Thus, it can be absorbed into the
≲ symbol without having much effect on the primary estimate.

Remark 1.21. Additionally, there are many situations in which we would like to estimate ex-
pressions involving spacetime derivatives ∂µ over the spatial region Ωt . However, this can be done
with the help of section 2.2. The primary idea is that we can use system (1.35) directly to solve
for ∂t (s, r, u) and ∂t (s̃, r̃, ũ) in terms of the spatial part ∂(s, r, u) and ∂(s̃, r̃, ũ) respectively, plus
additional terms that can be estimated. A key identity is

uj j
ũ0 = ũ (1.61)
u0

which follows from the orthogonality of u and ũ in (1.36d). With section 2.2 taken for granted, one
can treat ∂µ as a generic first order spatial derivative ∂ in the following proof with the understanding
that all time derivatives will be eventually written in terms of purely spatial derivatives.

Here, we will provide the proof of Proposition 1.18.

Proof. µ
 i Using Remarks 1.20 and 1.21, the first two terms in (1.57) on the RHS containing ∂µ u and
∂i uu0 can be quite easily estimated using E 0 and an expression depending on the L∞ (Ωt ) norms
14 A priori estimates for the linearized system

of u and ∂u. We arrive at the inequality

d 0
E ≲ C0 (∥∂u∥L∞ (Ωt ) , ∥u∥L∞ (Ωt ) )E 0
dt Z
1 1
− 0
∂t (r γ−1 r̃ũ0 ) dx
u
ZΩt
1 γ−1 1
− 0
r s̃ũµ ∂µ s dx
u
ZΩt (1.62)
1 1 1 1
− 0
(Γ + r)r γ−1 ũα ũµ ∂µ uα + 0 r γ−1 |ũ|2 uµ ∂µ r dx
Ω u u
Z t ′
1 Γ 1
γ−1 s̃ũµ ∂ r +
1 1 1
γ−1 r̃ũµ ∂ rdx
+ 0
r µ r µ
Ωt u Γ + r u0 Γ + r
:= ∥(∂u, u)∥L∞ (Ωt ) E 0 + I1 + I2 + I3 + I4
where each lower order term I1 , ..., I4 will be estimated in the following way.
The first term I1 is unique in that we plan to combine it with the LHS of our inequality after
integrating in time. We leave this term for now and will return to it later.
For I2 , we estimate it using the Cauchy Schwartz inequality. We have
Z Z  1 Z 1
1 1 2 1 2
0 −1 µ 0 −1 2 2
|I2 | ≤ r γ−1 (u ) |s̃ũ ∂µ s| ≤ ∥∂s∥L∞ (Ωt ) ∥(u ) ∥L∞ (Ωτ ) r γ−1 |s̃| r γ−1 |ũ|
Ωt Ωt Ωt
0
≲ C2 (∥∂s∥L∞ (Ωt ) , ∥u∥L∞ (Ωτ ) )E .
(1.63)

where we used remarks 1.20-1.21, and the fact that G is a Riemannian metric, hence |ũ|2δ ≲ |ũ|2G
where δ is the Euclidean metric in spacetime.
For I3 , we simplify Dt r = −(γ − 1)r∂µ uµ and get a similar inequality
Z Z
1 1
−1
|I3 | ≤ (Γ + r)r γ−1 (u )0 2 µ
|ũ| |∂µ u | dx + r γ−1 (u0 )−1 |ũ|2 (γ − 1)r|∂µ uµ | dx
Ωt Ωt (1.64)
≲ C3 (∥∂u∥L∞ (Ωt ) , ∥u∥L∞ (Ωτ ) )E 0
where we used Remark 1.20.
For I4 , we observe that Γ′ (s) = − γ−1
γ Γ(s) using (1.33), and we can apply similar arguments
along with the Cauchy-Schwarz inequality to get
 
γ−1
Z Z
1 1 1 1
|I4 | ≤ (u0 )−1 r γ−1 |s̃ũµ ∂µ r| dx + r γ−1 (u0 )−1 |r̃ũµ ∂µ r| dx
Ωt γ Γ+r Ωt Γ+r (1.65)
0
≲ C4 (∥∂r∥L∞ (Ωt ) , ∥u∥L∞ (Ωτ ) )E .
Thus, for the energy defined in (1.58) with t replaced by τ , we combine (1.63)-(1.65) and I1 to
get the inequality
Z
d 0 1
E ≲ C(∥∂s∥, ∥∂r∥, ∥∂u∥, ∥u∥)E + ∥u∥ ∂τ (r γ−1 r̃ũ0 ) dx (1.66)
dτ Ωτ

where the constant depends only on γ and the norms are on L∞ (Ωτ ). Integrating in time from 0
to t ∈ [0, T ], this implies
Z t Z tZ
1
0 0 0
E [t] ≲ E [0] + C(∥∂s∥, ∥∂r∥, ∥∂u∥, ∥u∥)E [τ ] dτ + ∂τ (r γ−1 r̃ũ0 ) dxdτ (1.67)
0 0 Ωτ
 Luczak 15

where the constant depends on γ and the L∞ norm of u which is bounded on [0, T ]. Then, from a
quick computation using the moving domain formula, we get
Z tZ Z t Z Z i

1
0 d 1
0
1
0 u
∂τ (r γ−1 r̃ũ ) dxdτ = r γ−1 r̃ũ dxdτ − ∂i (r γ−1 r̃ũ 0 ) dxdτ
0 Ωτ 0 dτ Ωτ Ωτ u
Z t Z
d 1
= r γ−1 r̃ũ0 dxdτ
dτ Ωτ
Z0 Z
1 1
0
= (r(t, x)) γ−1 r̃(t, x)ũ (t, x) dx − (r(0, x)) γ−1 r̃(0, x)ũ0 (0, x) dx
Ωt Ω0
(1.68)
where we integrate by parts for one of the terms and used the fact that r vanishes on the free
boundary. Thus, we can substitute this into our inequality to get
Z t
0 0
E [t] ≲ E [0] + C(∥∂s∥, ∥∂r∥, ∥∂u∥, ∥u∥)E 0 [τ ] dτ
0
Z Z (1.69)
1 1
0 0
+ (r(t, x)) γ−1 r̃(t, x)ũ (t, x) dx − (r(0, x)) γ−1 r̃(0, x)ũ (0, x) dx
Ωt Ω0
where the constant depends on γ and ∥u∥L∞ (Ωτ ) which is bounded.
Then, for x sufficiently close to the boundary, we have estimates of the form
Z Z  1 Z 1
1 1 2 1 2
0 2 2
r γ−1 r̃ũ dx ≤ r γ−1 |r̃| r γ−1 |ũ|
Ωt Ωt Ωt
 1 Z 1
(1.70)
Z
2−γ 2 1 2
1/2 2 2
≲ ∥r∥L∞ (Ωτ ) r γ−1 |r̃| r γ−1 |ũ|
Ωt Ωt
1/2 0
≲ ε̂ E [t]
where we applied Assumption 1.9 and used the smallness of r so that ε̂1/2 ≪ √1 < 1. A similar
2
estimate also holds for the Ω0 term. Thus, we can move the ε̂1/2 E 0 [t]
term to the LHS which will
yield
 Z t 
0 1 0 1/2 0
E [t] ≲ E [0](1 + ε̂ ) + (∥∂s∥L∞ (Ωτ ) + ∥∂r∥L∞ (Ωτ ) + ∥∂u∥L∞ (Ωτ ) )E [τ ] dτ
1 − ε̂1/2 0
(1.71)
Thus, by Gronwall’s inequality, we have
Z t 
0 0
E [t] ≲ E [0] exp C(∥∂s∥, ∥∂r∥, ∥∂u∥, ∥s∥, ∥r∥, ∥u∥) dτ (1.72)
0
with a constant depending only on γ and our distance to the free boundary. This completes our
proof of the basic energy estimate.
□

2. Creating a Book-keeping scheme

In works such as [4], a book-keeping scheme was created using the equations’ scaling law for the
leading order dynamics near the free boundary. However, no such scaling seems available for (1.36);
thus, we will need to develop several key ideas around the notion of order which is first defined in
Definition 2.5. After differentiating the equations, we plan to create a book-keeping scheme which
accounts for the number of derivatives and the powers of our free boundary weight r. Along the
way, any derivatives of the background quantities s, r, u will serve as L∞ coefficients for each of the
terms that we will encounter.
16 A priori estimates for the linearized system

Using system (1.36), we record the computations when taking Dt2k of the equations (where Dt0
is the identity). Since the s̃ equation will be treated separately in Section 3.1, we put only the r̃
and ũ equations here for convenience. Writing the equations in commutator notation and keeping
only the higher order terms on the LHS, we get the following system:

     
Dt Dt2k r̃ + Dt2k ũµ ∂µ r + (γ − 1)r∂µ Dt2k ũµ = B2k (2.1a)
  1  
Dt Dt2k ũα + Παµ ∂µ Dt2k r̃ = C2k
α
(2.1b)
Γ+r
where
2k−1
X  

B2k = Dt2k g − Dti ũµ ∂µ Dt2k−i r − Dt2k−i r∂µ Dti ũµ
i=0
2k−1
X 2k
X
− Dti ũµ [Dt2k−i , ∂µ ]r − (γ − 1) Dt2k−i r[Dti , ∂µ ]ũµ (2.2)
i=0 i=1
2k−1   2k  
X 1 X 1
α
Dt2k hα Dt2k−i Παµ ∂µ Dti r̃ − Dt2k−i αµ
[Dti , ∂µ ]r̃


C2k = − Π
Γ+r Γ+r
i=0 i=1

and g and hα are defined in (1.37). Additionally, we removed terms from the summation where the
commuator was zero. Each of the terms in (2.2) will be included in the higher order wave equation
estimates in Section 6.
Using several lemmas from [4] on embedding theorems combined with Assumption 1.9, we are
able to prove the following lemmas which motivate our definition of order defined in Remark 2.6.
Lemma 2.1. Suppose that σ ≥ 0, r vanishes simply on the free boundary, and f ∈ H k,σ . Then
1
f ∈ H k,σ+ 2 , and more specifically
∥f ∥2 1 ≤ ε̂∥f ∥2H k,σ (Ωt )
H k,σ+ 2 (Ωt )

where the integrals are interpreted with respect to Remark 1.10.

Lemma 2.2 (Free Boundary Embedding Lemma). Let H l,m be the corresponding Sobolev space for
the free-boundary problem with weight r that vanishes simply near the free boundary, and l, m ≥ 0.
If k ≥ 0 is provided, then
1
H 2k,k+ 2 ⊂ H l,m
1
provided l ≤ 2k and m − l + k − 2 ≥ 0.

Remark 2.3. By following the proof, we also have that natural corollary that H 2k,k ⊂ H l,m
provided m − l + k ≥ 0.
By combining the above lemmas we have the following useful corollary:
Corollary 2.4 (Free-Boundary trading derivatives for weight). Suppose that f ∈ H j+1,σ+1 and
j, σ ≥ 0. Then, we have
∥f ∥H j,σ ≲ ε̂∥f ∥H j+1,σ (2.3)
where ε̂ is the small positive constant defined in Assumption 1.9.
Proof. From [4], we see that ∥f ∥H j,σ ≲ ∥f ∥H j+1,σ+1 with a constant depending on σ. Then, we can
simply pull out a power of r in the L∞ (Ωt ) norm and use Assumption 1.9. □
 Luczak 17

2.1. Defining order for free-boundary terms.

Definition 2.5 (H2k -critical, subcritical, and supercritical terms). Let m, l ∈ N. In view of Lemma
2.2, any terms of the form
rm ∂ l ũ, rm ∂ l s̃
or
rm ∂ l r̃
with 0 ≤ l ≤ 2k and m ≥ 0 will be called H2k -critical provided that m−l+k− 21 = 0, or respectively
m − l + k = 0. Additionally, terms such that m − l + k − 12 > 0 or respectively m − l + k > 0
will be called H2k -subcritical. Similarly, terms where m − l + k − 21 < 0, m − l + k < 0 will be
denoted H2k -supercritical. In view of subsection 2.2, a similar classification can also be used for
the terms rm ∂ l ũ, rm ∂ l s̃, and rm ∂ l r̃.
Remark 2.6. We can also refer to the H2k -order of a given term by computing the value O :=
m − l + k − 12 (or O := m − l + k for terms involving r̃) and the sign of O will determine if the
term is subcritical or supercritical. As an example, a term of the form rk ∂ 2k ũ will be called H2k -
supercritical of order − 12 . The value O can be interpreted as a indicator for how well that term can
be estimated using the H2k norm. Critical terms have the exact number of derivatives and powers
of r to be estimated, whereas supercrtical terms are lacking key additional powers of r. Also, note
that the order O depends on a particular derivative level, i.e. what value of k we are using in order
to estimate terms using the H2k norm. Also, note that that the weights in the H2k norm depend
not only on k, but also on γ > 1 which is a fixed constant. For clarity, one can interpret this notion
of order as in the simplified setting where γ = 2 since our starting Lemma 2.2 is built around the
1
function spaces H 2k,k+ 2 and H 2k,k . In that case, the H2k -order of a term T can be interpreted as
a statement about the powers of r present/needed in the term T in order for the inequality
∥T ∥2L2 ≲ ∥(s̃, r̃, ũ)∥2H2k ≈ ∥s̃∥2 1 + ∥r̃∥2H 2k,k + ∥ũ∥2 1
H 2k,k+ 2 H 2k,k+ 2
to hold, and the symbol ≈ is used to represent the norm equivalence from Remark 1.14. For
2−γ 2−γ
2k,k+ 1 + 2k,k+ 2(γ−1)
arbitrary γ > 1, We will instead use the function spaces H 2 2(γ−1) and H . In this
2k
setting, the H -order for a term T is a statement about the inequality
∥T ∥2 2−γ ≲ ∥(s̃, r̃, ũ)∥2H2k ≈ ∥s̃∥2 2k,
2−γ 1
+k+ 2
+ ∥r̃∥2 2k,
2−γ
+k
+ ∥ũ∥2 2k,
2−γ 1
+k+ 2
(2.4)
L2 (r γ−1 ) H 2(γ−1) H 2(γ−1) H 2(γ−1)

which we summarize:
• Suppose that T is H2k subcritical. We claim that inequality (2.4) holds as well as
∥T ∥2 2−γ ≲ ε̂∥(s̃, r̃, ũ)∥2H2k (2.5)
L2 (r γ−1 )

using the previous Lemmas (also, note that (2.5) implies (2.4)). To see this fact, let’s assume
that T consists of r̃ terms so that T = rm ∂ l r̃ (although ũ and s̃ terms are similar). By
Definition 2.5, we have that m − l + k > 0. If l = 2k, then we must have m > k (extra power
of r) and we can apply Lemma 2.1 to get the desired inequality (2.5) with ε̂. If l < 2k, i.e.
l = 2k − b with 0 < b ≤ 2k, then we must have m > k − b (and m = 0 for b > k). We would
get
∥rm ∂ 2k−b r̃∥2 0, 2−γ ≲ ∥r̃∥ 2k−b,m+ 2−γ ≤ ∥r̃∥ 2k,m+b+ 2−γ . (2.6)
2(γ−1) H 2(γ−1) H 2(γ−1)
H
Then, since m + b > k we can apply Lemma 2.1 to finish the proof of (2.5).
• If T is H2k critical, then (2.4) holds by Lemma 2.2, but (2.5) in general does not since we
do not have any additional powers of r to spare.
18 A priori estimates for the linearized system

• If T is H2k supercritical with order O < 0, then (2.4) does not hold (and subsequently (2.5)
also does not), but (2.4) is true with T replaced by T̂ = r−O T as T is lacking key powers of
r. To see this fact, we simply observe that the order of the term T̂ will be (m−O)+l−k = 0
which is critical.
Lemma 2.7 (Sum of H2k orders). Suppose that T1 and T2 are free boundary terms (i.e. T1 and
T2 are of the form rm ∂ l (ũ, r̃, s̃) for some m ≥ 0 and 0 ≤ l ≤ 2k) that have H2k orders O1 and O2
respectively. If O1 + O2 ≥ 0, then ∥T1 T2 ∥L1 ≲ ∥(s̃, r̃, ũ)∥2H2k , i.e the H2k order of a product is the
sum of its orders.
Proof. Without loss of generality, the proof can be obtained by splitting into cases:
(1) T1 = rm1 ∂ l1 ũ and T2 = rm2 ∂ l2 ũ, m1 , m2 ≥ 0, 0 ≤ l1 , l2 ≤ 2k
m l m
(2) T1 = r ∂ r̃ and T2 = r ∂ r̃,
1 1 2 l 2 m1 , m2 ≥ 0, 0 ≤ l1 , l2 ≤ 2k
(3) T1 = rm1 ∂ l1 r̃ and T2 = rm2 ∂ l2 ũ, m1 , m2 ≥ 0, 0 ≤ l1 , l2 ≤ 2k
and correctly distributing the powers of r as needed.
□
Remark 2.8. We note that Lemma 2.7 is especially useful in energy estimates where we would
like to compute the order of a product of two terms, but we are estimating in L1 instead of L2 and
utilizing the Cauchy-Schwarz inequality. Using Lemma 2.7, we can extend the notion of order to a
product of terms T1 T2 by replacing inequalities (2.4) and (2.5) with
∥T1 T2 ∥ 2−γ ≲ ∥(s̃, r̃, ũ)∥2H2k (2.7)
L1 (r γ−1 )

∥T1 T2 ∥ 2−γ ≲ ε̂∥(s̃, r̃, ũ)∥2H2k (2.8)

L1 (r γ−1 )

With this new definition for products, Lemma 2.7 states that if the T1 has H2k -order O1 and T2
has H2k -order O2 , then T1 T2 has H2k -order O1 + O2 provided that O1 + O2 ≥ 0. When referring to
the order of products T1 T2 , we will always compute the order of T1 and T2 separately before using
Lemma 2.7 to make a statement about the order of T1 T2 .
Remark 2.9 (Definition of the ≃ symbol). Now that we have described the notion of order,
we will often use the symbol ‘≃’ to mean the following. Suppose that Ã(ũ, r̃, s̃) and B̃(ũ, r̃, s̃)
are expressions involving powers of r and derivatives (∂, ∂, Dt ) of our linearized variables. Using
Remark 2.6, suppose that à and B̃ both contain terms of order O at worst (i.e. there are no terms
with order strictly more negative than O). Then, we will write
à ≃ B̃ ⇐⇒ C1 (∂ 2k+1 (u, r, s), u, s)à = C2 (∂ 2k+1 (u, r, s), u, s)B̃ + P (2.9)
where C1 and C2 are expressions containing (u, s) and up to 2k + 1 derivatives of (u, r, s), and P
contains terms with order J > O (i.e. strictly better terms). Also, note that C1 and C2 must not
contain undifferentiated r, since powers of r would contribute to the calculated order for à and
B̃. For illustrative purposes (and by slight abuse of notation), we will sometimes write A ≃ B + P
instead of just à ≃ B̃ if we want to highlight the terms in P and calculate their order which will
be strictly greater that B̃.
As a final remark, we stated that ∂(r, s, u) or ∂(r, s, u) does not, in general, contribute to the
order of given term. However, there is one exception which can be seen by analyzing (1.35b) above.
For situations where we have Dt r, we see that a full power of r is obtained since Dt r = −(γ−1)r∂µ uµ
and powers of r are needed for calculating order. Thus, it can be useful to add the fact that
Dt r ≃ r (2.10)
even though in (2.9) we used the symbol ≃ when referring to terms involving (ũ, r̃, s̃).
 Luczak 19

It remains to show several lemmas and commutator identities. We will ultimately use our scheme
to simplify the B2k and C2kα terms that appear in (2.2). Then, we can estimate these terms using
2k
the H norm in Section 6. The proof of these lemmas is a routine application of the commutator
identities:
[∂, DtN ]ϕ = [∂, Dt ](DtN −1 ϕ) + Dt ([∂, DtN −1 ]ϕ)
(2.11)
[Dt , ∂ N ]ϕ = [Dt , ∂]∂ N −1 ϕ + ∂ [Dt , ∂ N −1 ]ϕ

Lemma 2.10 (First Commutator Lemma). Let N ∈ N. For the convective derivative Dt = uµ ∂µ ,
generic spatial derivative ∂, and ϕ sufficiently smooth, the following identities hold:
[∂, Dt ]ϕ = (∂uν )∂ν ϕ
[∂, Dt2 ]ϕ = 2(∂uν )∂ν (Dt ϕ) + [Dt (∂uν ) − (∂uµ )(∂µ uν )]∂ν ϕ
... (2.12)
N
X −1
[∂, DtN ]ϕ ≃ Ciν ∂ν (Dti ϕ)
i=0

where Ciν := Ci (∂u, Dt u) for i ∈ {0, ..., N − 1} are coefficients depending on the spatial and con-
vective derivatives of u.

Remark 2.11. We note that Lemma 2.10 is used extensively in Theorem 6.2 in order to simplify
the commutator terms from (2.2). For a discussion of the commutators used in elliptic estimates,
see Section 4.

We also have an important commutator lemma which will be used in our higher order estimates.

Lemma 2.12 (Second Commutator Lemma). Let N ∈ N. For the convective derivative Dt = uµ ∂µ ,
generic N -th order spatial derivative ∂ N , and ϕ sufficiently smooth, the following identities hold:
[Dt , ∂]ϕ = −(∂uν )∂ν ϕ
[Dt , ∂ 2 ]ϕ = −2(∂uν )∂ν (∂ϕ) − (∂ 2 uν )∂ν ϕ
... (2.13)
N
X −1
[Dt , ∂ N ]ϕ ≃ Biν ∂ν (∂ i ϕ)
i=0

where Biν := Bi (∂u) for i ∈ {0, ..., N − 1} are coefficients depending on the spatial derivatives of u.

Let’s continue the book-keeping process for higher order convective derivatives Dti s̃, Dti r̃, and
Dti ũ.The following Lemma can be taken with Section 2.2 and remark 2.15 in mind that we
preview here. Namely, we use a generic spacetime derivative with the understanding that time
derivatives will eventually be solved for in terms of spatial derivatives plus additional terms that
are more sub-critical (see Lemma 2.14 and Remark 2.15 below). For now, we will write Lemma
2.13 using the generic spacetime derivative symbol ∂ from Remark 1.7.

Lemma 2.13 (Book-keeping Dti derivatives). The following simplifications hold when counting
derivatives and powers of r. For i = 1, we are left with the terms
Dt s̃ ≃ ũ
Dt r̃ ≃ r∂ ũ + ũ (2.14)
Dt ũ ≃ ∂ r̃ + ũ + s̃
20 A priori estimates for the linearized system

where we only ignored ∂s, ∂r, ∂u, s, u, and Π, and we listed only Hi critical or supercritical terms.
Moreover, for i ≥ 2, we get the following

Dti s̃ ≃ Dti−1 ũ
(P
i/2 l l+i/2
i l=0 r ∂
Dt r̃ ≃ P(i+1)/2
r̃ , if i even
r∂l l+(i−1)/2 ũ , if i odd
l=0 (2.15)
( Pi/2 i/2−1
rl ∂ l+i/2 ũ + j=0 rj ∂ j+i/2 r̃
P
, if i even
Dti ũ ≃ Pl=0 (i−1)/2 l l+(i+1)/2
l=0 r∂ r̃ + rl ∂ l+(i−1)/2 (ũ + s̃) , if i odd

Proof. This lemma is proved inductively using the fact that Dt r ≃ r, Lemma 2.10, and our as-
sumptions on key terms that can be ignored at each step.
□

**************************************

2.2. Solving for time derivatives. There are many instances in which we would like to estimate
the spacetime derivatives ∂µ of a particular quantity using only the spatial derivatives ∂. To do
so, we can can solve for the time derivatives in terms of spatial derivatives by using system (1.35)
directly and the orthogonality identity (1.61).

Lemma 2.14 (Solving for time derivatives). The following computations and simplifications hold
under the book-keeping scheme from Section 2:
 Luczak 21

(u0 )2 = 1 + uj uj
uj
ũ0 = 0 ũj ≃ ũ
u
uj uj
∂i ũ = 0 ∂i ũj + ũj ∂i
0
≃ ∂ ũ + ũ
u u0
ui
∂t s = − 0 ∂i s
u
ui ui ũ0 − u0 ũi
∂t s̃ = − 0 ∂i s̃ + ∂i s ≃ ∂s̃ + ũ
u (u0 )2
∂t u0 = C1i ∂i u0 + C2i ∂i r + C3 r∂i ui
∂t ũ0 = C1i ∂i ũ0 + C2i ∂i r̃ + C3 r∂i ũi + C3 r̃∂i ui + C
fi ∂ u0 + C
1 i
f3 r∂i ui
fi ∂ r + C
2 i
≃ ∂ ũ + ∂ r̃ + r∂ ũ + (ũ + s̃ + r̃) + r(ũ + s̃ + r̃) + r2 (ũ + s̃ + r̃)
ui γ−1 γ−1
∂t r = − ∂i r − r∂i ui − r[∂t u0 ]
u0 u0 u0
ui γ−1 γ−1
∂t r̃ = − 0 ∂i r̃ − r∂i ũi − r[∂t ũ0 ]
u u0 u0
ui ũ0 − u0 ũi
   
(γ − 1) 0 γ−1 i (γ − 1) 0 γ−1
+ ∂i r + ũ r − r̃ ∂i u + ũ r − r̃ [∂t u0 ]
(u0 )2 (u0 )2 u0 (u0 )2 u0
≃ ∂ r̃ + r(∂ ũ + ∂ r̃) + (ũ + r̃) + r(ũ + s̃ + r̃) + r2 (ũ + s̃ + r̃) + r3 (ũ + s̃ + r̃)
ui 1 1
∂t uj = − 0
∂i uj − 0
Πj0 [∂t r] − Πji ∂i r
u (Γ + r)u (Γ + r)u0
ui 1 1
∂t ũj = − 0 ∂i ũj − 0
Πj0 [∂t r̃] − Πji ∂i r̃
u (Γ + r)u (Γ + r)u0
 j ′
ui ũ0 − u0 ũi u (Γ s̃ + r̃) + (Γ + r)ũj

j
+ ∂i u + [∂t r]
(u0 )2 (Γ + r)2
(δ + uj ui )[(Γ + r)ũ0 + u0 (Γ′ s̃ + r̃)] − (Γ + r)u0 (ũj ui + uj ũi )
 ji 
+ [∂i r]
(Γ + r)2 (u0 )2
≃ ∂ ũ + ∂ r̃ + r(∂ ũ + ∂ r̃) + (ũ + s̃ + r̃) + r(ũ + s̃ + r̃) + r2 (ũ + s̃ + r̃) + r3 (ũ + s̃ + r̃)
(2.16)
22 A priori estimates for the linearized system

with
(γ − 1)((u0 )2 − 1)
a1 = u0 − r
(Γ + r)u0
(γ − 1)((u0 )2 − 1)
ã1 = ũ0 − r̃
(Γ + r)u0
2(Γ + r)(γ − 1)u0 ũ0 − (γ − 1)((u0 )2 − 1)[(Γ′ s̃ + r̃)u0 + (Γ + r)ũ0 ]
 
−r
(Γ + r)2 (u0 )2
≃ ũ + r̃ + r(ũ + s̃ + r̃)
ui
C1i : = −
a1
fi = ui ã1 − a1 ũi
C 1 ≃ ũ + r̃ + r(ũ + s̃ + r̃)
(a1 )2
1
C2i : = Ci
(Γ + r)u0 1
1 ′ 0 0
fi =
C Cfi − [(Γ s̃ + r̃)u + (Γ + r)ũ ] C i ≃ ũ + s̃ + r̃ + r(ũ + s̃ + r̃)
2 1
(Γ + r)u0 1 (Γ + r)2 (u0 )2
(γ − 1)((u0 )2 − 1) i
C3 : = C2
ui
(γ − 1)((u0 )2 − 1) fi ui 2(γ − 1)ũ0 − (γ − 1)((u0 )2 − 1)ũi i
C
f3 = C2 + C2 ≃ ũ + s̃ + r̃ + r(ũ + s̃ + r̃)
ui (ui )2
(2.17)
where, namely, we only ignored ∂s, ∂r, ∂u, s, u and O(1) coefficients in the simplfication ≃.
Remark 2.15. In general, we have that
∂t s̃ ≃ ∂s̃ + (terms which are more subcritical than ∂s̃)
∂t ũ ≃ ∂ ũ + (terms which are more subcritical than ∂ ũ) (2.18)
∂t r̃ ≃ ∂ r̃ + (terms which are more subcritical than ∂ r̃)
with coefficient expressions that have at most one spatial derivative of s, r, or u. The point is that
any time derivative of our linearized variables (s̃, r̃, ũ) can be written in terms of spatial derivatives
(with good coefficients) plus additional subcritical terms that can be easily estimated. When
estimating terms such as in Sections 3.2 and 6, we will often apply Lemma 2.14 proactively so as
to not introduce unnecessary subcritical terms that obscure the argument. Thus, going forward,
one can treat the spacetime expressions below with the generic spatial derivative symbol ∂
∂µ s̃ ≃ ∂s̃
∂µ r̃ ≃ ∂ r̃
∂µ ũν ≃ ∂ ũ
from the point of view of their H2k order, as opposed to Remark 1.7 where ∂ would be used.
Recall that when referring to the order of a free boundary term, we are doing so with respect
to Definition 2.5 and Remark 2.6 where terms with order O = 0 are denoted as critical, O < 0 are
supercritical, and O > 0 are subcritical. For example, it is crucial in our expression (2.18) for ∂t r̃
that the parentheses do not contain terms of the form ∂ ũ. This is because, at a given level say
k = 1, the H2 order of a term ∂ r̃ is 0 (i.e. critical), but the order of a term like ∂ ũ is − 21 (i.e.
supercritical) with order strictly worse. If this were the case, we could not reliably argue that we
could replace ∂t r̃ with ∂ r̃ since we would be introducing terms that require additional powers of r
in order to be estimated with the H2k norm.
 Luczak 23

Proof. We leave computation details for Lemma 2.14 in the Appendix. □

To finish this section, we prove a useful Lemma which expands on the notion of order O to each
of our key operations on free boundary terms. This Lemma is especially helpful in proving higher
order analogs of our elliptic and div-curl estimates using induction. The last proposition in Lemma
2.16 is helpful in jumping from the base case H2 to the general case H2k .
Lemma 2.16 (The orders of each operation). Let k ∈ N be fixed. Suppose that T is a free boundary
term with H2k order O (i.e. T is of the form rm ∂ l (s̃, r̃, ũ) for some m, l ∈ N, m ≥ 0, 0 ≤ l ≤ 2k, ).
Then the following holds:
(1) the term rT has H2k order O + 1
(2) the term ∂T has H2k order O − 1
(3) the term Dt T has a minimum H2k order O − 12
(4) If K ∈ N with K ≥ k, then T has H 2K order O + K − k.
Proof. Since the case with s̃ is identical to ũ, it suffices to consider the following cases for our free
boundary term T :
(a) T1 = rm1 ∂ l1 r̃ which has H2k order m1 − l1 + k = O
(b) T2 = rm2 ∂ l2 ũ which has H2k order m2 − l2 + k − 12 = O
For Part 1, observe that for cases (a) and (b), we simply have an extra power of r
 
rT1 = r rm1 ∂ l1 r̃ = rm1 +1 ∂ l1 r̃ =⇒ rT1 has H2k order m1 + 1 − l1 + k = O + 1
  1 (2.19)
rT2 = r rm2 ∂ l2 ũ = rm2 +1 ∂ l2 ũ =⇒ rT2 has H2k order m + 1 − l + k − = O + 1
2
For Part 2, we see that there are two sub-cases when m1 , m2 ≥ 1 and m1 = 0 = m2 . If m1 = 0 = m2 ,
then
∂T1 = ∂(∂ l1 r̃) = ∂ l1 +1 r̃ =⇒ ∂T1 has H2k order m1 − (l1 + 1) + k = O − 1
1 (2.20)
∂T2 = ∂(∂ l2 ũ) = ∂ l2 +1 ũ =⇒ ∂T2 has H2k order m2 − (l2 + 1) + k − = O − 1
2
If m1 , m2 ≥ 1, then we get
∂T1 = ∂(rm1 ∂ l1 r̃) = m1 rm1 −1 ∂ l1 r̃ + rm1 ∂ l1 +1 r̃ =⇒ ∂T1 has H2k order O − 1
(2.21)
∂T2 = ∂(rm2 ∂ l2 ũ) = m2 rm2 −1 ∂ l2 ũ + rm2 ∂ l2 +1 ũ =⇒ ∂T2 has H2k order O − 1
since both terms have H2k order O − 1, noting that we either lose a power of r or gain a derivative
(and both operations decrease the order from O to O − 1).
For Part 3, we will first consider when l1 = 0 = l2 . For case with T1 , we have
Dt T1 = Dt (rm1 r̃) = m1 rm1 −1 (Dt r)r̃ + rm1 Dt r̃
≃ m1 rm1 r̃ + rm1 (r∂ ũ + ũ) (2.22)
= rm1 +1 ∂ ũ + rm1 ũ + m1 rm1 r̃
where we used (2.10) and Lemma 2.13. Then, computing the H2k order (and noting that we now
have some terms with ũ instead of r̃), we see that the first two terms have order O − 21 , while the
last one has order O. Thus, Dt T1 has a minimum H2k order O − 21 (i.e. it has order O − 12 at
worst). For T2 , we get a similar computation using (2.10) and Lemma 2.13
Dt T2 = Dt (rm2 ũ) = m2 rm2 −1 (Dt r)ũ + rm2 Dt ũ
≃ m2 rm2 ũ + rm2 (∂ r̃ + ũ + s̃) (2.23)
m2 m2 m2 m2
=r ∂ r̃ + r ũ + m2 r ũ + r s̃
24 A priori estimates for the linearized system

where the first term has order m2 − 1 + k = (m2 − 0 + k − 12 ) − 12 = O − 12 and the remaining terms
have order O.
For Part 3, it remains to consider when l1 , l2 ≥ 1. We will additionally make use of Lemma 2.12.
For the T1 case, we have
Dt T1 = Dt (rm1 ∂ l1 r̃) ≃ rm1 Dt (∂ l1 r̃) + m1 rm1 ∂ l1 r̃
≃ rm1 ∂ l1 (Dt r̃) + rm1 [Dt , ∂ l ]r̃ + m1 rm1 ∂ l1 r̃
1 −1
lX
!
m1 l1 m1
≃ r ∂ (r∂ ũ + ũ) + r Bi ∂ν (∂ r̃) + m1 rm1 ∂ l1 r̃
ν i

i=0
1 −1
lX
!
≃ rm1 ∂ l1 ũ + rm1 ∂r∂ l1 ũ + rm1 +1 ∂ l1 +1 ũ + rm1 ∂ i+1 r̃ + m1 rm1 ∂ l1 r̃
i=0
(2.24)
where we note that ∂ l1 (r∂ ũ) only produces critical terms when all ∂ l1 derivatives hit ∂ ũ, or one
derivative hits r and ∂ l1 −1 derivatives hit ∂ ũ. For the summation terms, we make use of Remak
2.15 to simplify ∂ν ∂ i r̃ as ∂ i+1 r̃. Then, counting the H2k order for each of the terms, we see that the
first three have order O − 21 , whereas the remaining terms have order O at worst (with additional
terms in the summation having order O + l1 − i with 0 ≤ i < l1 ).
For the T2 case with l1 , l2 ≥ 1, we perform a similar computation and keep track of whether the
terms have r̃ or ũ when computing order
Dt T2 = Dt (rm2 ∂ l2 ũ) ≃ rm2 Dt (∂ l2 ũ) + m2 rm2 ∂ l2 ũ
≃ rm2 ∂ l2 (Dt ũ) + rm2 [Dt , ∂ l ]ũ + m2 rm2 ∂ l2 ũ
2 −1
lX
!
m2 l2 m2
≃r ∂ (∂ r̃ + ũ + s̃) + r Biν ∂ν (∂ i ũ) + m2 rm2 ∂ l2 ũ (2.25)
i=0
1 −1
lX
!
m2 l2 +1 m2 l2 m2 l2 m2 i+1
≃r ∂ r̃ + r ∂ ũ + r ∂ s̃ + r ∂ ũ + m2 rm2 ∂ l2 ũ
i=0

where the first term has order m2 − (l2 + 1) + k = (m2 − l2 + k − 21 ) − 21 = O − 12 , and the
H2k
remaining terms have order O at worst. Thus Dt T only produces terms that have a minimum order
of O − 12 , and all other terms have strictly O order or better at the H2k level. This completes the
proof of Part 3.
Lastly, we will provide the proof of Part 4. In the previous parts, each operation changed the
number of powers of r, the number of derivatives, or whether we are using r̃ or ũ and each of these
has an effect on the computed order. In Part 4, we are only seeing what happens if we raise the
level from k to some K ≥ k, and the proof can be seen by noticing that for T1 ,
m1 − l1 + K = (m1 − l1 + k) + K − k = O + K − k
and for T2 ,
1 1
m2 − l2 + K −
= (m2 − l2 + k − ) + K − k = O + K − k.
2 2
This completes the proof of Part 4, and the proof of Lemma 2.16 is complete. □
Remark 2.17. A natural corollary of this Lemma is the following. Suppose that T has H2k -order
O. Then, Dt2 T has H2k -order O − 1, but this implies that Dt2 T has H2(k+1) -order O (if we increase
2(k−1)
the level k by 1). Hence, if a given term is H2 subcritical (with k = 1), applying Dt (for k ≥ 2)
will keep the term H2k -subcritical at the corresponding k level.
 Luczak 25

3. Estimates for the transport energy

2k
3.1. Entropy estimates. Recall the definition of Etransport in (1.48) which consists of both the
entropy and the vorticity. Our goal in this section is to prove estimates for the entropy, whereas
the voriticity will be handled separately in Section 3.2. Instead of applying Dt2k to the s̃ equation,
we can simply take ∂ 2k and estimate s̃ using the H2k norm directly. This can be done because s̃
2k
satisfies a transport equation instead of a wave equation that would require the intermediate Ewave
energy. We have the following proposition which is routine by envoking out previous Lemmas on
the order of terms.
Proposition 3.1 (Entropy Estimates).
d
∥s̃∥2 2k,k+ 1 ≲ B2 ∥(s̃, r̃, ũ)∥2H2k (3.1)
dt H 2(γ−1)

∞
where B2 depends on L norms of up to 2k + 1 derivatives of (s, r, u).
Proof. Using equation (1.36a) Lemma 2.12 for the commutator [Dt , ∂ 2k ], we get the following:
Dt (∂ 2k s̃) = [Dt , ∂ 2k ]s̃ + ∂ 2k f
2k−1
X
= Biν ∂ν (∂ i s̃) + ∂ 2k (ũµ ∂µ s)
i=0 (3.2)
2k
X
≃ Bi ∂ i s̃ + ∂ 2k (ũµ ∂µ s)
i=0
where we recall the definition of f from (1.37), and we used Remark 2.15 to combine ∂ν ∂ i s̃ and
1
2k+ γ−1
write everything in terms of spatial derivatives. Then, multipliying the equation by r ∂ 2k s̃,
we get
2k
!
1 2k+ γ−11
2k+ 1 X 1
2k+ γ−1
r Dt (|∂ 2k s̃|2 ) ≃ r γ−1 ∂ 2k s̃ Bi ∂ i s̃ + r ∂ 2k s̃∂ 2k (ũµ ∂µ s)
2
i=0
2k
!
1 1
k+ 2(γ−1) 2k 2 1 2k+ γ−1 1
2k 2 2k+ γ−1 1
2k
X
i 2k+ γ−11
Dt (|r ∂ s̃| ) ≃ − Dt (r )|∂ s̃| + r ∂ s̃ Bi ∂ s̃ + r ∂ 2k s̃∂ 2k (ũµ ∂µ s)
2 2
i=0
2k
  !
1 2k+ γ−1 1 1
2k+ γ−1 2k
X
≃ k(γ − 1) + r |∂ 2k s̃|2 + r ∂ s̃ Bi ∂ i s̃
2
i=0
1
2k+ γ−1
+r ∂ 2k s̃∂ 2k (ũµ ∂µ s)
(3.3)
Integrating, this leads to:
Z Z  
d1 k+ 1 1 1 1
2k+ γ−1
|r 2(γ−1) ∂ 2k s̃|2 dx ≃ 0
k(γ − 1) + r |∂ 2k s̃|2 dx
dt 2 Ωt Ωt u 2
2k
Z !
1
2k+ γ−1
X
+ r ∂ 2k s̃ i
Bi ∂ s̃ dx
Ωt i=0
Z
1 2k+ γ−1
1 (3.4)
+ 0
r ∂ 2k s̃∂ 2k (ũµ ∂µ s) dx
Ωt u
Z  i
1 k+ 1 u
+ |r 2(γ−1) ∂ 2k s̃|2 ∂i dx
2 Ωt u0
= J1 + J2 + J3 + J4 .
26 A priori estimates for the linearized system

For the J1 term, we note that u and its derivatives have been absorbed in ≃ symbol (for example
1
2k+ γ−1
when we simplifying Dt (r )). Thus, we can remove any derivatives of u and quickly estimate
Z 
1
k+ 2(γ−1)
|J1 | ≲ ∥∂u∥L∞ (Ωt ) |r ∂ s̃| dx ≲ ∥∂u∥L∞ (Ωt ) ∥(s̃, r̃, ũ)∥2H2k
2k 2
(3.5)
Ωt

H2k
recallingthedefinition of from (1.46), and Remark 1.14. The estimate for J4 follows similarly
ui
since ∂i u0 is an expression depending on up to one spatial derivative of u. For the J2 and J3
terms, we simply use Lemmas 2.2 and 2.14 as needed while keeping track of subcritical terms.
□
3.2. Vorticity estimates. We also plan to derive similar estimates for the vorticity which will be
used in connection with the elliptic estimates. We recall from [2] that the relativistic vorticity is
given by the following
ωαβ = ∂α vβ − ∂β vα . (3.6)
where v is the enthalpy current defined by
γ−1
Γ+r (γ − 1) γ − γ−1 s
vα = huα , h= , Γ = Γ(s) = e γ (3.7)
Γ γ
Similar to [15], we get an expression for uα ωαβ using the following steps. From the fact that
v vα = −h2 , we get
α

−v α ∂β vα = h∂β h. (3.8)
From (1.35c) rewritten in terms of v, we get
1 α 1
2
v ∂α vβ = − ∂β r (3.9)
h Γ+r
Then, we can compute the following expression
1 α r γ−1
(u ∂α vβ − uα ∂β vα ) = ∂β s (3.10)
h (Γ + r) γ
Multiplying by h, we get
r
uα ωαβ = (γ − 1)∂β s (3.11)
γΓ
which agrees with [15] as
r
ε= (3.12)
γΓ
after checking (1.24). To find the evolution equation satisfied by ωαβ , we can first compute that
 
1 γ−1
∂α ε = ∂α r + r∂α s .
γΓ γ
Using similar computations as [15] and applying (3.12), we have
γ−1
uµ ∂µ ωαβ + ∂α uµ ωµβ + ∂β uµ ωαµ = (∂α r∂β s − ∂β r∂α s). (3.13)
γΓ
which is the evolution equation for the vorticity ωαβ in terms of the unknowns (s, r, u).
Next, we introduce the full linearized voricity ω̃αβ by
gβ ) − ∂β (hu
ω̃αβ = ∂α (hu gα )
= ∂α (hũβ ) − ∂β (hũα ) + ∂α (h̃uβ ) − ∂β (h̃uα ) (3.14)
:= ω̂αβ + ω αβ
where
ω̂αβ := ∂α (hũβ ) − ∂β (hũα ) (3.15)
 Luczak 27

will be called the reduced linearized vorticity and ω αβ is the h̃-scaled vorticity.
The plan is to derive estimates for the full linearized vorticity ω̃, and then relate this back to
the reduced linearized vorticity ω̂ by ∥ω̂∥ ≲ ∥ω̃∥ + ∥ω∥. Our reasoning is a matter of convenience,
since estimates for ω̃ are more readily available in view of (3.17). Additionally, ω̂ better resembles
curl ũ in connection with our elliptic estimates (see Section 4 and Lemma 4.4).
Throughout this section, we will use the following facts similar to the bookkeeping system from
Lemma 2.13. We put the order simplifications here for reference

ω αβ ≃ ∂ h̃ + h̃
h̃ ≃ r̃ + rs̃ (3.16)
∂α h̃ ≃ ∂ r̃ + r∂s̃ + r̃ + s̃ + rs̃

where we also make use of Lemma 2.14 in the expression for ∂α h̃, i.e. ∂t h̃ will have a slightly
different expression, just with more subcritical terms coming from ∂t s̃ ∼ ∂s̃ and ∂t r̃ ∼ ∂ r̃.
From linearizing (3.13), we have the evolution equation for ω̃ given by

uµ ∂µ ω̃αβ + ∂α uµ ω̃µβ + ∂β uµ ω̃αµ = −ũµ ∂µ ωαβ − ∂α ũµ ωµβ − ∂β ũµ ωαµ

(γ − 1)2
+ s̃(∂α r∂β s − ∂β r∂α s) (3.17)
γ2Γ
γ−1
+ (∂α r̃∂β s + ∂α r∂β s̃ − ∂β r̃∂α s − ∂β r∂α s̃)
γΓ

Remark 3.2. For the following lemma, recall Remark 1.14 where we have by embedding theorems
1 1
2k,k+ 2(γ−1) 2k,k+ 2(γ−1) − 12 1
2k,k+ 2(γ−1)
that H2k ∼ H ×H ×H where the weight on r̃ is less by 12 .

Lemma 3.3 (Estimates for ω̃). For a solution ω̃ to (3.17) on an interval [0, T ] defined in this
section, we have the following estimate

d
∥ω̃∥2 2k−1,k+ 1 ≲ C∥(s̃, r̃, ũ)∥2H2k (3.18)
dt H 2(γ−1)

where C depends on the L∞ (Ωt ) norms of up to 2k + 1 derivatives of s, r, and u.

Proof. We begin by fixing a spatial multiindex l such that |l| ≤ 2k − 1. We recall the notation
|∂ l ω̃|2G = Gαγ Gβδ ∂ l ω̃ γδ ∂ l ω̃αβ (where l is not summed over) and G is the Riemannian metric defined
1
2k+ (γ−1)
in (1.42). Then, we apply ∂ l to (3.17) and contract against the expression r Gαγ Gβδ ∂ l ω̃ γδ .
When contracting, we will use the quick observation that

1
2k+ (γ−1) 1 2k+ (γ−1)
1 1
2k+ (γ−1) 1
r Gαγ Gβδ ∂ l ω̃ γδ uµ ∂µ ∂ l ω̃αβ = r Dt (|∂ l ω̃|2G ) − r Dt (Gαγ Gβδ )∂ l ω̃ γδ ∂ l ω̃αβ
2 2
1 1
2k+ (γ−1) 1 1 1
2k+ (γ−1)
= Dt (r |∂ l ω̃|2G ) − (2k + )r ∂µ uµ |∂ l ω̃|2G
2 2 γ−1
1
2k+ (γ−1) 1
−r Dt (Gαγ Gβδ )∂ l ω̃ γδ ∂ l ω̃αβ .
2
(3.19)

and only the first term in (3.19) will be kept on the LHS. This leads to the following long expression:
28 A priori estimates for the linearized system

1 1
2k+ (γ−1) 1 1 1
2k+ (γ−1)
Dt (r |∂ l ω̃|2G ) = (2k + )r ∂µ uµ |∂ l ω̃|2G
2 2 γ−1
1
2k+ (γ−1) 1
+r Dt (Gαγ Gβδ )∂ l ω̃ γδ ∂ l ω̃αβ
2
1
2k+ (γ−1)
−r Gαγ Gβδ ∂ l ω̃ γδ ∂ l (∂α uµ ω̃µβ + ∂β uµ ω̃αµ )
1
2k+ (γ−1)
−r Gαγ Gβδ ∂ l ω̃ γδ ∂ l (ũµ ∂µ ωαβ )
1
2k+ (γ−1)
−r Gαγ Gβδ ∂ l ω̃ γδ ∂ l (∂α ũµ ωµβ )
1
2k+ (γ−1)
−r Gαγ Gβδ ∂ l ω̃ γδ ∂ l (∂β ũµ ωαµ )
(γ − 1)2
 
1
2k+ (γ−1)
+r Gαγ Gβδ ∂ l ω̃ γδ ∂ l s̃(∂ r∂
α β s − ∂β αr∂ s)
γ2Γ
 
α β l γδ l γ − 1
1
2k+ (γ−1)
+r Gγ Gδ ∂ ω̃ ∂ (∂α r̃∂β s + ∂α r∂β s̃ − ∂β r̃∂α s − ∂β r∂α s̃)
γΓ
8
X
:= Ri
i=1
(3.20)
where each term on the RHS of (3.20) will be handled separately. Then, using the function
1
2k+ (γ−1)
f (x, t) = r |∂ l ω̃|2G , we have:
Z Z  i Z 8
d u 1 X
f dx = f ∂i dx + Ri dx (3.21)
dt Ωt Ωt u0 Ωt u
0
i=1

We plan to estimate each of the terms on the RHS of (3.21) by the H2k norm which is also used in
Theorem 6.2. This is done by analyzing the H2k order (see Remark 2.6) and making use of Lemma
2.14 as needed for any time derivatives that appear.
We start by observing that the first term on the RHS of (3.21) is handled in an identical way as
the term J4 from (3.4) in the entropy estimates. We can use (3.14) and (3.16) to get the additional
comment that
ω̃ := ω̂ + ω ≃ ∂ ũ + ∂ r̃ + r∂s̃ + s̃ + r̃ + rs̃ (3.22)
noting that, from an H2k order perspective, ω̃ contains terms with the same order as ∂ ũ or better.
These terms primarily come from the reduced linearized vorticity ω̂, whereas ω ∼ ∂ h̃ + h̃ and h̃
contains terms that are like r̃ at worst.
To finish estimating the first term on the RHS of (3.21), we have
Z  i Z
1
2k+ (γ−1) u 1
2k+ (γ−1)
r l 2
|∂ ω̃|G ∂i 0
dx ≲ ∥∂u∥ L∞ (Ω )
t
r Gαγ Gβδ ∂ l ω̃ γδ ∂ l ω̃αβ dx
Ωt u
ZΩt
2−γ 1 2
≲ ∥∂u∥L∞ (Ωt ) r γ−1 rk+ 2 ∂ l (∂ ũ + ∂ r̃ + r∂s̃ + s̃ + r̃ + rs̃) dx
Ωt
(3.23)
where we applied (3.22). The point is that if we greatly expand the expression Gαγ Gβδ ∂ l ω̃ γδ ∂ l ω̃αβ ,
only the terms with the worst H2k order need to be checked, since terms that are more subcritical
have plenty of powers of r to be estimated using the H2k norm. In this integral, the terms with
1
the worst order are rk+ 2 ∂ l+1 ũ. Since |l| ≤ 2k − 1, we see that the order of this term is exactly 0
(critical) when |l| = 2k − 1 and subcritical when |l| < 2k − 1. Thus, it can be estimated using the
 Luczak 29

2
H2k norm. For any cross terms in the expansion of ∂ l (∂ ũ + ∂ r̃ + r∂s̃ + s̃ + r̃ + rs̃) , we simply
apply Lemma 2.7 where the order of a product of is the sum of the orders of each term. For
k+ 1 k+ 1
example, in the product |(r 2(γ−1) ∂ l+1 ũ)(r 2(γ−1) ∂ l+1 r̃)|, we see that the order (i.e. when γ = 2)
is 0 + 1/2 = 1/2 ≥ 0 at worst when |l| = 2k − 1, so Lemma 2.7 implies that
Z
k+ 1 k+ 1
|(r 2(γ−1) ∂ l+1 ũ)(r 2(γ−1) ∂ l+1 r̃)| dx ≲ ∥(s̃, r̃, ũ)∥2H2k (3.24)
Ωt
2
All other cross terms in ∂ l (∂ ũ + ∂ r̃ + r∂s̃ + s̃ + r̃ + rs̃) are handled using Lemma 2.7. Thus, we
get
Z  i
2−γ
k+ 12 l 2 u
r γ−1 |r ∂ ω̃|G ∂i dx ≲ C0 ∥(s̃, r̃, ũ)∥2H2k (3.25)
Ωt u0
where C0 depends on the L∞ (Ωt ) norms of up to 2k derivatives of s, r, and u.
We will proceed through each remaining term on the RHS of (3.21). For the integrals containing
R1 and R2 , the estimate is quite similar to the first term since derivatives of u will be removed,
and Dt (Gαγ Gβδ ) = Dt ((gγα + 2uα uγ )(gδβ + 2uβ uδ )) is another expression containing derivatives of u.
We have
Z  
2−γ 1 1
|R1 | ≲ ∥∂u∥L∞ (Ωt ) r γ−1 k+ |rk+ 2 ∂ l ω̃|2G dx ≲ C1 ∥(s̃, r̃, ũ)∥2H2k (3.26)
Ωt 2(γ − 1)
and
Z
1
2k+ (γ−1) 1
|R2 | ≲ r |Dt (Gαγ Gβδ )∂ l ω̃ γδ ∂ l ω̃αβ | dx
Ωt 2
Z
2−γ 2
≲ ∥∂u∥L∞ (Ωt )
1
r γ−1 rk+ 2 ∂ l (∂ ũ + ∂ r̃ + r∂s̃ + s̃ + r̃ + rs̃) dx (3.27)
Ωt
2
≲ C2 ∥(s̃, r̃, ũ)∥H2k
where we handled it in the same way as (3.23).
For the R3 integral, distributing ∂ l (∂α uµ ω̃µβ + ∂β uµ ω̃αµ ) leads to terms that contain ∂ l ω̃ at
worst along with up to 2k derivatives of u. We can write this as
|l|
X
l µ
∂ (∂α u ω̃µβ + ∂β u ω̃αµ ) ≃ µ
∂ n ω̃∂ |l|−n+1 u (3.28)
n=0

so that, upon estimating, we will have

Z |l|
2−γ X 1 1
2k
|R3 | ≲ ∥∂ u∥L∞ (Ωt ) r γ−1 |rk+ 2 ∂ l ω̃||rk+ 2 ∂ n ω̃| dx ≲ C3 ∥(s̃, r̃, ũ)∥2H2k (3.29)
Ωt n=0

since the worst term in the summation only contains 2k − 1 derivatives of ω̃.
For the R4 integral, we have ∂ l (ũµ ∂µ ωαβ ) which will lead to terms with ∂ 2k−1 ũ at worst and
coefficients containing up to 2k + 1 derivatives of u, r, and s coming from ∂ l (∂µ ωαβ ). Since we
have previously estimated up to ∂ 2k ũ, the terms in R4 will be more subcritical (and thus easier to
estimate) than R1 through R3 which all contain 2k − 1 derivatives of ω̃. We can write it as
Z
2−γ 1 1
2k+1
|R4 | ≲ ∥∂ (u, r, s)∥L∞ (Ωt ) r γ−1 |(rk+ 2 ∂ l ω̃)(rk+ 2 ∂ l ũ)| dx
Ωt (3.30)
2
≲ C4 ∥(s̃, r̃, ũ)∥H2k
For the R5 and R6 integrals, the argument is quite similar since we get up to 2k derivatives of ũ
and up to 2k derivatives of ω.
30 A priori estimates for the linearized system

For the R7 , integral, we will have terms containing up to 2k −1 derivatives of s̃ and 2k derivatives
of r and s. However, we know from computing H2k orders that ∂ 2k−1 s̃ has the same order as ∂ 2k−1 ũ,
so it can be estimated in the same way as R4 .
Finally, for the R8 integral, we will have terms containing up to 2k derivatives of s̃ and r̃, and up
to 2k derivatives of s and r. However, we know that the order of these terms will be comparable
to ∂ 2k ũ at worst (and in fact, terms with ∂ 2k r̃ are more subcritical by an order value of 1/2).
Thus, by estimating each of the terms in (3.21) separately and removing expressions involving
L∞ norms of up to 2k + 1 derivatives of s, r, and u, we combine the above estimates with (3.21) to
get Z
d 2−γ 1
r γ−1 |rk+ 2 ∂ l ω̃|2G dx ≲ C∥(s̃, r̃, ũ)∥2H2k (3.31)
dt Ωt
where C depends on up to 2k + 1 derivatives of s, r, and u. Summing over the multiindex l achieves
the desired vorticity estimate using the H2k norm.
□
Lemma 3.3 leads nicely into the next theorem, and we note how this theorem compares with the
transport energy defined in (1.48).
Theorem 3.4 (Linearized vorticity, Transport energy estimates). The following estimate holds for
the linearized vorticity defined in (3.14).
Z t
∥ω̂∥2 2k−1,k+ 1 ≲ ∥ω̃0 ∥2 2k−1,k+ 1 +ε∥(s̃, r̃, ũ)∥2H2k (Ωt ) + C∥(s̃, r̃, ũ)∥2H2k (Ωτ ) dτ (3.32)
H 2(γ−1) (Ω ) H 2(γ−1) (Ω )
t 0 0

where ω̃0 is the initial linearized vorticity and C is from Lemma 3.3.
Proof. We plan to use Lemma 3.3 and the following identity ω̂ = ω̃ − ω to split the estimate:
∥ω̂∥2 2k−1,k+ 1 ≲ ∥ω̃∥2 2k−1,k+ 1 + ∥ω∥2 2k−1,k+ 1 (3.33)
H 2(γ−1) (Ω ) H 2(γ−1) (Ω ) H 2(γ−1) (Ω )
t t t

Now, from the definition of ω and the simplifications outlined in (3.16), we get the following when
applying a multiindex l with |l| ≤ 2k − 1:
∂ l ω ≃ ∂ l (∂(r̃ + rs̃) + r̃ + rs̃)
(3.34)
≃ ∂ l+1 r̃ + r∂ l+1 s̃ + ∂ l s̃
where the additional terms are even more subcritical than the ones listed. Then, using the
1
2k+ (γ−1)
L2 (r ) norm, we have
Z Z Z
2−γ 1 2−γ 1 2−γ 1
k+ l 2 k+ l+1 2
r γ−1 |r 2 ∂ ω|G dx ≲ r γ−1 |r 2 ∂ r̃| dx + r γ−1 |rk+ 2 +1 ∂ l+1 s̃|2 dx
Ωt Ωt Ωt
Z
2−γ 1
+ r γ−1 |rk+ 2 ∂ l s̃|2 dx
Ωt
Z
2−γ
Z
2−γ
(3.35)
1
k l+1 2
≲ ε̂ r γ−1 |r ∂ r̃| dx + ε̂ r γ−1 |rk+ 2 ∂ l+1 s̃|2 dx
Ωt Ωt
Z
2−γ 1
k−1+ l 2
+ ε̂ r γ−1 |r 2 ∂ s̃| dx
Ωt

where we used the smallness of r from Remark 1.10. Summing over |l| ≤ 2k − 1 and counting the
1 1
order for each term (where the worst terms take the form rk ∂ 2k r̃, rk+ 2 ∂ 2k ũ, and rk+ 2 −1 ∂ 2k−1 s̃),
we get the estimate
∥ω∥2 2k−1,k+ 1 ≲ ε̂∥(s̃, r̃, ũ)∥2H2k (3.36)
H 2(γ−1)
 Luczak 31

By integrating Lemma 3.3, we have

Z t
∥ω̃∥2 2k−1,k+ 1 ≲ ∥ω̃0 ∥2 2k−1,k+ 1 + C∥(s̃, r̃, ũ)∥2H2k (Ωτ ) dτ (3.37)
H 2(γ−1) (Ω 2(γ−1) (Ω
t) H 0) 0

where ω̃0 can be solved for using the initial data for the linearized equation (s̃0 , r̃0 , ũ0 ). Combining
the above equations together yields the desired result. □

Remark 3.5. Theorem 3.4 will be used in our main result, Theorem 7.1.

4. Elliptic and Div-Curl Estimates

Our goal is to bound the total energy (1.49) from above and below by the H2k norm (1.46).
In this section, we focus on bounding (1.49) from below, and one of the key ingredients for the
wave part (1.47) is elliptic estimates involving r̃ and the spatial divergence of ũ. The proofs in this
section involve isolating a “good” spatial elliptic operator similar to [4], with the added difficulty in
that we must explicitly split off time derivatives while isolating the critical terms and using Section
2.2. Additionally, we will often make use of Lemma 2.13 where identities with Dt serve as another
method for counting order.
Our analysis will focus on isolating the “good” spatial elliptic operators, as one can readily
compare that the structure of these operators closely resembles the elliptic operators already treated
in [4].

4.1. Elliptic estimates for r̃. Upon taking Dt of equation (1.36b) and applying Theorem 2.13
as needed, we can rewrite Dt2 r̃ in the following way:

Dt2 r̃ ≃ L1 r̃ + additional terms (4.1)

with
γ − 1 µν 1
L1 r̃ := Π (r∂µ ∂ν r̃ + ∂µ r∂ν r̃)
Γ+r γ−1
γ − 1 ij 1
= Π (r∂i ∂j r̃ + ∂i r∂j r̃)
Γ+r γ−1
γ − 1 00 2 1
+ Π (r∂t r̃ + ∂t r∂t r̃)
Γ+r γ−1
(4.2)
2(γ − 1) i0
+ Π (r∂i ∂t r̃)
Γ+r
1
+ Π0j (∂t r∂j r̃)
Γ+r
1
+ Πi0 (∂i r∂t r̃)
Γ+r
serving as the “spacetime” elliptic operator for r̃, and we have split the time derivatives from the
spatial derivatives. To see how the additional terms in (4.1) will be treated, as well as the higher
order elliptic estimates, we refer the reader to section 4.3 where these additional terms are shown
to be subcritical using our book-keeping scheme from Section 2.
It is important to note that since we are estimating the wave energy from below by the H2k
norm, we cannot simply cite section 2.2 for the time derivatives since the structure of the terms is
quite important. Instead, we will make use of the following additional facts from the definition of
32 A priori estimates for the linearized system

Dt :
1 ui
∂t r̃ = D t r̃ − ∂i r̃
u0 u0
1 ui uj ui 1 1
r∂t2 r̃ = 0 2 rDt2 r̃ − 0 2 r∂i ∂j r̃ − 2 0 r∂i ∂t r̃ − 0 2 Dt u0 r∂t r̃ − 0 2 Dt ui r∂i r̃
(u ) (u ) u (u ) (u )
1 uu i j ui i
u u j 1 1
= 0 2 rDt2 r̃ + 0 2 r∂i ∂j r̃ − 2 0 2 r∂i (Dt r̃) + 2 0 ∂i 0 r∂j r̃ − 0 2 Dt u0 r∂t r̃ − 0 2 Dt ui r∂i r̃
(u ) (u ) (u ) u u (u ) (u )
(4.3)
where, by Remark 2.6, the only critical terms (for k = 1) are in red. For the remaining terms, we
can write
ui
 
1 1 1
Π00 ∂t r∂t r̃ = Π00 ∂t r Dt r̃ − ∂ i r̃
Γ+r Γ+r u0 u0
(4.4)
1 00 1 00 i
= Π ∂t rDt r̃− Π ∂t ru ∂i r̃
(Γ + r)u0 (Γ + r)u0
and
uj
 
2(γ − 1) i0 2(γ − 1) i 0 1
Π (r∂i ∂t r̃) = u u r∂i Dt r̃ − 0 ∂j r̃
Γ+r Γ+r u0 u
2(γ − 1) i 2(γ − 1) i j
= ru ∂i (Dt r̃)− ru u ∂i ∂j r̃ (4.5)
Γ+r Γ+r
uj
 
2(γ − 1) i 0 1
+ u u r Dt r̃∂i 0 − ∂i 0 ∂j r̃ .
Γ+r u u
and
uj
 
1 i0 1 i0 1
Π (∂i r∂t r̃) = Π ∂i r Dt r̃ − 0 ∂j r̃
Γ+r Γ+r u0 u
(4.6)
1 1
= 0
Πi0 ∂i rDt r̃− Πi0 ∂i ruj ∂j r̃.
(Γ + r)u (Γ + r)u0
So, if we gather up all the red critical terms (which look like r∂ 2 r̃ or ∂ r̃), we will get precisely,
((u0 )2 − 1)ui uj
 
γ−1
L1 r̃ = Πij + − 2ui j
u r∂i ∂j r̃
Γ+r (u0 )2
((u0 )2 − 1) j
 
1 ij 1 0 j
+ g ∂i r∂j r̃ + u u − u ∂t r∂j r̃
Γ+r Γ+r u0
ui ui uj
 
γ−1 1 2 1 0 1 i
+ rD r̃ − 2 0 2 r∂i (Dt r̃) + 2 0 ∂i 0 r∂j r̃ − 0 2 Dt u r∂t r̃ − 0 2 Dt u r∂i r̃
Γ + r (u0 )2 t (u ) u u (u ) (u )
1
+ Π00 ∂t rDt r̃
(Γ + r)u0
uj
 
2(γ − 1) i 2(γ − 1) i 0 1
+ ru ∂i (Dt r̃) + u u r Dt r̃∂i 0 − ∂i 0 ∂j r̃
Γ+r Γ+r u u
1
+ Πi0 ∂i rDt r̃
(Γ + r)u0
(4.7)
and we can check that each of the black terms is subcritical with the help of Lemma 2.13 and
Remark 2.6. For a discussion of how these subcritical terms are handled, see section 4.3 below.
 Luczak 33

Finally, we use (4.3) to simplify ∂t r in the red critical terms (which splits into a subcritical and a
critical term). Combining, we end up with
γ − 1 ij 1
L1 r̃ = H (r∂i ∂j r̃ + ∂i r∂j r̃)
Γ+r γ−1 (4.8)
+ (black subcritical terms)
where
ui uj
H ij := δ ij −
, (4.9)
(u0 )2
If we set the “good” spatial elliptic operator to be
γ − 1 ij 1
L̃1 r̃ := H (r∂i ∂j r̃ + ∂i r∂j r̃) (4.10)
Γ+r γ−1
then, we get the expression:

L1 r̃ = L̃1 r̃ + (black subcritical terms) (4.11)

where the term in red is critical, and all the black sub-critical terms are given by:
1 1
= Dt ruj ∂j r̃
Γ + r (u0 )2
ui ui uj
 
γ−1 1 2 1 0 1 i
+ rD r̃ − 2 0 2 r∂i (Dt r̃) + 2 0 ∂i 0 r∂j r̃ − 0 2 Dt u r∂t r̃ − 0 2 Dt u r∂i r̃
Γ + r (u0 )2 t (u ) u u (u ) (u )
1
+ Π00 ∂t rDt r̃
(Γ + r)u0
uj
 
2(γ − 1) i 2(γ − 1) i 0 1
+ ru ∂i (Dt r̃) + u u r Dt r̃∂i 0 − ∂i 0 ∂j r̃
Γ+r Γ+r u u
1
+ Πi0 ∂i rDt r̃
(Γ + r)u0
under the book-keeping scheme of Section 2 (where we can use Remark 2.6 and Lemma 2.13 as
needed to see that these terms have order O = 12 at worst). See section 4.3 below for how these
terms are handled.
Lemma 4.1 (Ellptic estimates for r̃). The following estimates hold for r̃ and the “good” spatial
elliptic part L̃1 r̃:
∥r̃∥ 2, 1 + 21 ≲ ∥L̃1 r̃∥ 0, 1 − 12 + ∥r̃∥ 2−γ (4.12)
H 2(γ−1) H 2(γ−1) L2 (r γ−1 )

Proof. After examining the structure of L̃1 r̃, we see that the proof will closely follow [4] by proving
two related inequalities:
∥r̃∥ 2, 1 +1 ≲ ∥L̃1 r̃∥ 0, 1 −1 + ∥r̃∥ 1, 1 −1
H 2(γ−1) 2 H 2(γ−1) 2 H 2(γ−1) 2
(4.13)
∥r̃∥ 1, 1 −1 ≲ ∥L̃1 r̃∥ 0, 1 −1 + ∥r̃∥ 2−γ
H 2(γ−1) 2 H 2(γ−1) 2 L2 (r γ−1 )

Instead of repeating the proof of Lemma 5.3 in [4], we illustrate a few of the primary differences.
1
When derivatives hit the weight Γ+r , we see by (1.33) that
   
1 1 γ−1
∂k = Γ∂ k s − ∂ k r (4.14)
Γ+r (Γ + r)2 γ
which only contains derivatives of r and s. Similarly, when derivatives hit the metric H ij , we get
an expression that only contains derivatives of u (and this metric also appears in [4]).
34 A priori estimates for the linearized system

To show the next inequality in (4.13), we use a similar method as the proof of Corollary 5.5 in
[4]. □

4.2. Elliptic estimates for ũ. Next, we will handle the ũ equation. From (2.1b), we note that
Dt2 ũα will satisfy an equation involving one derivative of the spacetime divergence ∂ν ũν (see (4.18)
for the definition of operator L2 ). In view of Section 2.2, we would like to solve for time derivatives
in terms of spatial derivatives and relate our analysis back to the spatial divergence of ũ which we
will denote by
−−→
div ũ := ∂j ũj (4.15)
In order to estimate our energies from below with the H2k norm, we will need to do elliptic estimates
−−→
involving div ũ. After extracting the spatial part, we will arrive at the “good” operator L̃2 (see
(4.27)) which is properly suited for div-curl estimates. To complete the estimates, we will need to
examine the spatial curl whose components are given by:
−−→
(curl ũ)ij := ∂i ũj − ∂j ũi . (4.16)
We plan to pair L̃2 with the corresponding operator L̃3 (see (4.29) below), and we note that L̃3 is
−−→
roughly one spatial derivative of curl ũ with a weight r.
Our goal is Lemma 4.3, but we will begin by highlighting key steps in the following computation
to show how we correctly isolate the spatial components. Similar to the r̃ equation, we start by
isolating the spacetime elliptic piece of Dt2 ũα which will produce the following upon taking Dt of
(2.1b):
Dt2 ũα ≃ (L2 ũ)α + additional terms (4.17)
where
 
γ − 1 αµ 1
(L2 ũ)α := Π ∂µ (r∂ν ũν ) + ∂µ ũν ∂ν r
Γ+r γ−1
   
γ − 1 αi ν 1 ν γ − 1 α0 ν 1 ν
= Π ∂i (r∂ν ũ ) + ∂i ũ ∂ν r + Π ∂t (r∂ν ũ ) + ∂t ũ ∂ν r
Γ+r γ−1 Γ+r γ−1
(4.18)
and the additional terms will be shown to be subcritical (see section 4.3 for a discussion of these
terms). Then, solving for
1 ui
∂t = 0 Dt − 0 ∂i (4.19)
u u
and gathering up similar terms, we have the useful identity
Παµ ∂µ = Παi ∂i + Πα0 ∂t
1 (4.20)
:= B αi ∂i + Πα0 0 Dt
u
where we defined:
ui
B αi := g αi − g α0 . (4.21)
u0
Further, note the explicit identity
Gαβ B αi B βj = H ij . (4.22)
where Gαβ is defined in (1.42), and H ijis defined in (4.9).
Using these identities, we arrive at the equivalent expression
   
γ − 1 αi 1 γ − 1 α0 1 1
(L2 ũ)α = B ∂i (r∂ν ũν ) + ∂i ũν ∂ν r + Π D t (r∂ν ũν
) + D t ũν
∂ ν r .
Γ+r γ−1 Γ+r u0 γ−1
(4.23)
 Luczak 35

Remark 4.2. If we suppose that k = 1 for computing order (see Remark 2.6), note that the terms
in red have order − 21 which are supercritical, but the terms in black have order 0 which is a strictly
1
better order. In the context of estimating L2 ũα , we will use the norm H 0, 2 in which case we will
have an extra 1/2 power of r in order to safely move these black terms to the other side. Thus, the
terms in black can be treated in a very similar way as (4.11) in the r̃ elliptic estimates.
At this stage, we still have an expression for L2 ũα involving the spacetime divergence ∂ν ũν . To
further isolate the spatial divergence, we make use of the following identities by repeatedly splitting
and removing time derivatives as needed. We apply (1.61) and (4.19) to get
uj uj
∂ν ũν = H jk ∂j ũk + 0 2
Dt ũj + ũj ∂t 0 . (4.24)
(u ) u
Similarly, we can apply the same identities such as (1.61) to replace any instances of ũ0 with a
spatial component in the expression ∂i ũν ∂ν r. Splitting the sum over ν and solving for ∂t r using
(4.19), we arrive at

ul um ũl um ul ũl ul
 
ul
ν k
∂i ũ ∂ν r = ∂i ũ ∂k r + − 0 2 ∂i ũl ∂m r + 0 2 ∂i ũl Dt r − 0 ∂i 0 ∂m r + 0 ∂i 0 Dt r
(u ) (u ) u u u u
 l m l
 (4.25)
ul ũ u ul ũ ul
= H km ∂i ũk ∂m r + 0 2
∂i ũl Dt r − 0 ∂i 0 ∂m r + 0 ∂i 0 Dt r
(u ) u u u u
Plugging these identities into (4.23), we have
 
γ − 1 αi jk 1
(L2 ũ)α = B H ∂i (r∂j ũk ) + ∂j r∂i ũk
Γ+r γ−1
 j 
γ − 1 αi u j uj
+ B ∂i rDt ũj + rũ ∂t 0 (4.26)
Γ+r (u0 )2 u
l m ũl ul
 
1 αi ul l ũ u ul
+ B ∂i ũ Dt r − 0 ∂i 0 ∂m r + 0 ∂i 0 Dt r .
Γ+r (u0 )2 u u u u
If we define the “good” spatial elliptic part for the divergence of ũ to be
 
 α γ − 1 αi jk 1
L̃2 ũ := B H ∂i (r∂j ũk ) + ∂j r∂i ũk , (4.27)
Γ+r γ−1
and observe Remark 4.2, we get the nice expression
 α
(L2 ũ)α = L̃2 ũ + (black critical and subcritical terms) (4.28)
 α
Next, we plan to combine the L̃2 ũ with a corresponding curl term in order to prove the
desired div-curl estimates. Consider the following
 α γ − 1 αi − γ−1
1

1+ 1

L̃3 ũ := B r H ml ∂l r γ−1 (∂m ũi − ∂i ũm ) (4.29)
Γ+r
Lemma 4.3  (Div-Curl
α estimates
 αfor ũ). The following estimates hold for ũ and the “good” spatial
elliptic parts L̃2 ũ and L̃3 ũ :

∥ũ∥ 2, 1 +1 ≲ ∥(L̃2 + L̃3 )ũ∥ 0, 1 + ∥ũ∥ 1 (4.30)

H 2(γ−1) H 2(γ−1) L2 (r γ−1 )

Proof. The proof is completed in two steps similar to Lemma 4.1, and it closely follows Lemma 5.3
in [4] where one should compare with our definitions of L̃2 and L̃3 . Addtionally, we make explicit
use of identity (1.61) since ũα uα = 0, and we refer the reader to Remark 1.21
36 A priori estimates for the linearized system

We start by writing:
Z
1
∥(L̃2 + L̃3 )ũ∥2 0, 1 = r γ−1 Gαβ [(L̃2 + L̃3 )ũ]α [(L̃2 + L̃3 )ũ]β dx (4.31)
H 2(γ−1)
Ωt
Then, using identity (4.22), we get the following collection of terms in the integral:
γ − 1 2 γ−1
   
1
ia jk ln 1 1
− γ−1
 1
1+ γ−1
r H H H r∂i ∂j ũk + ∂i r∂j ũk + ∂j r∂i ũk + r ∂k r (∂j ũi − ∂i ũj )
Γ+r γ−1
 
1 1
− γ−1
 1
1+ γ−1
× r∂a ∂l ũn + ∂a r∂l ũn + ∂l r∂a ũn + r ∂n r (∂l ũa − ∂a ũl )
γ−1
(4.32)
which can be expanded to get the following:
γ − 1 2 γ−1
 
1
r H ia H jk H ln
Γ+r
 
1 γ
× r∂i ∂j ũk + ∂i r∂j ũk + ∂j r∂i ũk + r∂k ∂j ũi − r∂k ∂i ũj + (∂k r∂j ũi − ∂k r∂i ũj )
γ−1 γ−1
 
1 γ
× r∂a ∂l ũn + ∂a r∂l ũn + ∂l r∂a ũn + r∂n ∂l ũa − r∂n ∂a ũl + (∂n r∂l ũa − ∂n r∂a ũl )
γ−1 γ−1
(4.33)
where each of the terms in red take the form r∂ 2 ũ. Then, we group the red terms and integrate by
parts while using the symmetry present in the expression
γ − 1 2 ia jk ln
 
ia,jk,ln
D := H H H . (4.34)
Γ+r
In this form, we can apply the same ideas as [4] where we first prove
∥(L̃2 + L̃3 )ũ∥2 0, 1 ≳ ∥ũ∥2 2, 1 +1
− ∥ũ∥2 1, 1 , (4.35)
H 2(γ−1) H 2(γ−1) H 2(γ−1)

and then integrate by parts with ∂3 ũa to prove the second inequality
∥(L̃2 + L̃3 )ũ∥2 0, 1 ≳ ∥ũ∥2 1, 1 − ∥ũ∥2 1 (4.36)
H 2(γ−1) H 2(γ−1) L2 (r γ−1 )
in which case (4.35) and (4.36) will combine to prove the desired result.
□
For the last lemma in this section, we will ultimately need to connect the spacetime two-form ω̂
−−→
back to curl ũ which is a key quantity in our div-curl estimates for ũ.
Lemma 4.4 (Relating vorticity to spatial curl). The following estimate holds for the reduced
linearized vorticity:
−−→
∥ω̂∥2 2k−1,k+ 1 ≳ ∥curl ũ∥2 2k−1,k+ 1 − ε̂2 ∥ũ∥2 2k,k+ 1 (4.37)
H 2(γ−1) H 2(γ−1) H 2(γ−1)
−−→
where ε̂ is small positive constant defined in Assumption 1.9, and curl ũ is defined in 4.16.
Proof. Recalling the definition of ω̂ in (3.15) which is a spacetime two-form, we observe the fol-
→
−
lowing. First, we define the spatial reduced linearized vorticity ω̂ to be a spatial two-form with
components
ω̂ij = ∂i (hũj ) − ∂j (hũi ) = h(∂i ũj − ∂j ũi ) + ũj ∂i h − ũi ∂j h,
Then, if we denote |·|2δ(3) to be the spatial Euclidean norm (squared), we have
→
− 2
ω̂ (3) := δ ij δ km ω̂ik ω̂jm . (4.38)
δ
 Luczak 37

For a generic multiindex l with |l| ≤ 2k − 1, we claim that the following inequality holds:

→
− 2 2
∂ l ω̂ ≲ ∂ l ω̂ (4.39)
δ (3) G
Indeed, since G is a Riemannian metric on R4 , it equivalent to δ on R4 with constants depending
on u. This gives
2 2
∂ l ω̂ ≳ ∂ l ω̂ = δ αγ δ βδ ∂ l ω̂αβ ∂ l ω̂γδ
G δ (4)
= δ iγ δ βδ ∂ l ω̂iβ ∂ l ω̂γδ + δ 0γ δ βδ ∂ l ω̂0β ∂ l ω̂γδ
= δ ij δ βδ ∂ l ω̂iβ ∂ l ω̂jδ + δ i0 δ βδ ∂ l ω̂iβ ∂ l ω̂0δ + δ βδ ∂ l ω̂0β ∂ l ω̂0δ
(4.40)
= δ ij δ kδ ∂ l ω̂ik ∂ l ω̂jδ + δ ij δ 0δ ∂ l ω̂i0 ∂ l ω̂jδ + δ βδ ∂ l ω̂0β ∂ l ω̂0δ
= δ ij δ km ∂ l ω̂ik ∂ l ω̂jm + δ ij δ k0 ∂ l ω̂ik ∂ l ω̂j0 + δ ij ∂ l ω̂i0 ∂ l ω̂j0 + δ βδ ∂ l ω̂0β ∂ l ω̂0δ
→
− 2
= ∂ l ω̂ (3) + δ ij ∂ l ω̂i0 ∂ l ω̂j0 + δ ij ∂ l ω̂0i ∂ l ω̂0j
δ
where we split indices, the red terms are zero, and the purple term is zero when β = 0 or δ = 0
using the definition of ω̂. Then, using ω̂ij = −ω̂ji , we have
3
l→
− l→
−
2 2 X
ij l l ij l l
∂ ω̂ + δ ∂ ω̂i0 ∂ ω̂j0 + δ ∂ ω̂0i ∂ ω̂0j = ∂ ω̂ + (∂ l ω̂i0 )2
δ (3) δ (3)
i=1 (4.41)
→
− 2
≥ ∂ l ω̂
δ (3)
1
2k+ γ−1
which completes the proof of (4.39). Multiplying (4.39) by the weight r , integrating, and
summing over the multiindex l with |l| ≤ 2k − 1 produces the inequality
→
−
∥ω̂∥2 2k−1,k+ 1 ≳ ∥ ω̂ ∥2 2k−1,k+ 1 . (4.42)
H 2(γ−1) H 2(γ−1)

To finish the proof of Lemma 4.4, we simply observe from (4.16) that
h −−−→ i h −−−→ i
δ ij δ km ∂ l ω̂ik ∂ l ω̂jm = δ ij δ km ∂ l h( curl ũ)ik + ũk ∂i h − ũi ∂k h ∂ l h( curl ũ)jm + ũm ∂j h − ũj ∂m h
−−−→  −−−→ 
≃ δ ij δ km (h∂ l ( curl ũ)ik + additional terms) h∂ l ( curl ũ)jm + additional terms
(4.43)
−−−→
We note that any terms where less than |l| derivatives hit curl ũ will be less critical. For each
additional term, we will always have 2k − 2 or less derivatives of ũ, and thus we can can apply
Corollary 2.4 to gain an extra weight and apply the standard smallness arguments with Cauchy-
Schwarz as needed for cross terms (see Lemma 4.3 where similar cross terms are handled). Moreover,
since h ∼ O(1) near the free boundary, and ∂i h = Γ1 ∂i r + γ−1 γΓ r∂i s does not contribute to the order,
∞
we can pull out derivatives or r and s in the L (Ωt ) norm. Thus, we have
−−−→
∥ω̂∥2 2k−1,k+ 1 ≳ ∥ curl ũ + (smallness terms)∥2 2k−1,k+ 1 . (4.44)
H 2(γ−1) 2(γ−1) H
which produces the desired inequality (4.37) after removing the terms with ε̂. □
4.3. Higher order commutators with the convective derivative. In this subsection, we plan
to adapt the elliptic estimates to higher values of k and show that all lower order and commutator
terms can be treated using the correct weighted Sobolev norm. First, we will prove Lemma 4.5
which handles the additional terms that appear when taking multiple convective derivatives of
our system (1.36) for r̃ and ũ. In Section 4.4, we will focus on commuting with weighted spatial
38 A priori estimates for the linearized system

derivatives where the r̃ equation will be discussed in detail. For ũ, the method is quite similar, and
we state the important result in Lemma 4.14.
Earlier in section 4, recall that there were many “additional terms” or black subcritical terms
that appeared when defining our “good” operators L̃1 , L̃2 , and L̃3 . In the following summarizing
lemma, we will indicate how each of these terms is subcritical, as well as the interaction between
Dt and these operators so that we can eventually prove higher order elliptic estimates.

Lemma 4.5 (Commutators between Dt and L̃1 , L̃2 , L̃3 ). Using the Book-keeping scheme of Lemma
2.13, the following identities hold

 
1
Dt2k r̃ ≃ L̃1 (Dt2k−2 r̃) 2k
+ terms with H -order at worst (4.45)
2
 
Dt2k ũα ≃ L̃2 (Dt2k−2 ũα ) + terms with H2k -order 0 at worst (4.46)
 
Dt2k−2 (L˜3 ũ)α ≃ L̃3 (Dt2k−2 ũα ) + terms with H2k -order 0 at worst (4.47)

where the order of the terms has been computed with Remark 2.6. Moreover, [L˜3 , Dt2k−2 ]ũ contains
terms which are critical/subcritical at worst.

Proof. First, we plan to collect “additional terms” that appear in (4.1). We begin by writing
the structure of (4.1) in a manner similar to Lemma 2.13 where derivatives and powers of r are
examined. We get the following:

Dt2 r̃ ≃ L1 r̃ + r∂(s̃ + r̃ + ũ) + r[Dt , ∂]ũ + ũ

(4.48)
≃ L1 r̃ + r∂ ũ + r∂s̃ + ũ + r∂ r̃

where L1 is the operator defined in (4.2). If we account for the black subcritical terms that
appear in (4.11), we will have
1
Dt2 r̃ ≃ L̃1 r̃ + (H2 -order terms at worst) + r∂ ũ + r∂s̃ + ũ + r∂ r̃ (4.49)
2
If we compute the order of each of these terms at the j = 1 level, we see that the blue terms have
order O = 21 > 0 and the purple terms have order O = 1 > 0. By Lemma 2.16 and Remark 2.17,
applying Dt2k−2 to the blue and purple terms will produce corresponding subcritical terms with
H2k -orders O = 12 at worst for level k. Thus, we will take Dt2k−2 of (4.49) to get
1
Dt2k r̃ ≃ Dt2k−2 (L̃1 r̃) + (O = terms at worst)
2 (4.50)
1
= L̃1 (Dt2k−2 r̃) + [Dt2k−2 , L̃1 ]r̃ + (O = terms at worst)
2
To finish the proof of (4.45), it remains to check the commutator [Dt2k−2 , L̃1 ]r̃ which can be done
with the help of Lemma 2.12 and an induction argument. We start by computing the first commu-
tator with the help of Lemma 2.12.
γ − 1 ij 1
[Dt , L̃1 ]ϕ = [Dt , H r∂i ∂j ]ϕ + [Dt , H ij ∂i r∂j ]ϕ
Γ+r Γ+r
≃ [Dt , r∂ 2 ]ϕ + [Dt , ∂]ϕ (4.51)

≃ r∂ 2 ϕ + ∂ϕ + r∂ϕ.
 Luczak 39

recalling that ∂ indicates a generic spatial derivative, and we are ordering terms from the worst to
the best order. We have
[Dt2 , L̃1 ]ϕ = Dt [Dt , L̃1 ]ϕ + [Dt , L̃1 ]Dt ϕ
(4.52)
≃ r∂ 2 (Dt ϕ) + ∂(Dt ϕ) + r∂(Dt ϕ) + r∂ 2 ϕ + ∂ϕ + r∂ϕ
and we also note by Remark 2.17 and when ϕ = r̃, we will get
[Dt2 , L̃1 ]r̃ ≃ r∂ 2 (r∂ ũ + ũ) + ∂(r∂ ũ + ũ) + r∂(r∂ ũ + ũ) + r∂ 2 r̃ + ∂ r̃ + r∂ r̃
(4.53)
≃ (r2 ∂ 3 ũ + r∂ 2 ũ + ∂ ũ) + r∂ 2 r̃ + ∂ r̃ + r∂ r̃
Then since, the “k” value when computing order at this level is 2, (i.e. from Remark 2.6) we see
that each one of these terms has H4 -order 12 , 1, and 2 respectively (and they are all subcritical with
O = 21 at worst). To complete the induction, suppose that for some j ≥ 2, we have [Dt2j−2 , L̃1 ] is
subcritical (with order O = 12 ) at level j. Then, we compute using the commutator identities that
2(j+1)−2
[Dt , L̃1 ]r̃ = [Dt2j , L̃1 ]r̃ = [Dt2 Dt2j−2 , L̃1 ]r̃ = Dt2 [Dt2j−2 , L̃1 ]r + [Dt2 , L̃1 ]Dt2j−2 r̃ (4.54)
Since [Dt2j−2 , L̃1 ] is subcritical at level j, we can conclude using Remark 2.17 that the first term
in (4.54) must have order O = 21 at the j + 1 level. For the second term in (4.54), it is easiest to
compute the order at the j + 1 level by counting derivatives and powers of r. We will have using
Lemma 2.13 that
[Dt2 , L̃1 ]Dt2j−2 r̃ ≃ r∂ 2 (Dt Dt2j−2 r̃) + ∂(Dt Dt2j−2 r̃) + r∂(Dt Dt2j−2 r̃) + r∂ 2 Dt2j−2 r̃ + ∂Dt2j−2 r̃ + r∂Dt2j−2 r̃
≃ r∂ 2 (rj ∂ 2j−1 ũ) + ∂(rj ∂ 2j−1 ũ) + r∂(rj ∂ 2j−1 ũ)
+ r∂ 2 (rj−1 ∂ 2j−2 r̃) + ∂ rj−1 ∂ 2j−2 r̃ + r∂ rj−1 ∂ 2j−2 r̃

(4.55)

noting that there are collection of terms in the simplification for Dt2j−1 r̃ which all have the same
order. Thus, computing the total order at the j + 1 level, we get that each of the terms has order
O = 21 at worst. We have completed the induction and shown that [Dt2k−2 , L̃1 ] has order O = 12
at worst for level k, and this completes the proof of (4.45). The proofs of (4.46) and (4.47) are
similar.
□
Remark 4.6. When we go to estimate Ewave 2k from below, we will observe that these subcritical/
critical terms will not cause any issue. First, consider estimating terms in the r̃ equation which are
subcritical. In Lemma 4.12 below, there terms are ultimately handled with smallness arguments
combined with Remark 2.6. A similar situation also happens for the ũ equation. Although the
expression for Dt2k ũ contains terms which are critical as well (and not just subcritical), we recall
the useful fact that the ũ equation gets a multiplier of order 1/2 when doing energy estimates.
2k
Subsequently, the Ewave energy has an extra r1/2 weight and this will be perfect for applying the
same smallness arguments for these extra terms.
4.4. Commutators with weighted spatial derivatives. Our goal is to extend Lemmas 4.1 and
4.3 to higher order Sobolev spaces. To do so, we will need to compute commutators between our
good elliptic operators L̃1 , L̃2 , L̃3 , and the weighted spatial derivatives rk−j ∂ 2(k−j) .
Let k ∈ N be fixed, and j ∈ N with 1 ≤ j ≤ k. To begin, we will use the following notation:
m := k − j
a,b
L := ra ∂ b (4.56)
r̃2j := Dt2j r̃
40 A priori estimates for the linearized system

and we recall from (4.10) that

γ − 1 ij 1
L̃1 = H (r∂i ∂j + ∂i r∂j )
Γ+r γ−1
We plan to handle L̃1 first, before moving on to L̃2 and L̃3 . By Lemma 4.5, we have that
r̃2j ≃ L̃1 r̃2j−2 + P (4.57)
where P are terms that have an order of 1/2 or better at level j (i.e. at the H2j level). We plan
to ultimately handle these terms using smallness arguments after we apply Lm,2m to (4.57) and
2−γ
0, 2(γ−1)
analyze with the H norm. Applying Lm,2m to both sides (where Lm,2m is defined in (4.56)),
we have
Lm,2m r̃2j ≃ L̃1 Lm,2m r̃2j−2 + [Lm,2m , L̃1 ]r̃2j−2 + Lm,2m P (4.58)
Remark 4.7. Observe what will happen to the order of the terms in P when we apply Lm,2m . At
the start, P has terms with H2j order 21 (at worst). However, by repeatedly applying Lemma 2.16,
we see that ∂ 2m P has H2j -order 21 − 2m. Then, rm ∂ 2m P has H2j -order 12 − 2m + m = 21 − m =
1
2 − (k − j). Then, if we want to calculate the order at level k instead of level j (and note that
k ≥ j), we see by Part 4 in Lemma 2.16 that rm ∂ 2m P must have H2k -order 12 −(k −j)+(k −j) = 21 .
By Remark 4.7, this means that at the H2k level, Lm,2m P will still be subcritical and we plan
to handle these terms with smallness arguments. Thus, let us make use of the following notation:
 
2k 1
PO≥1/2 := terms with H order O ≥ (4.59)
2
which we will use to collect any perturbative/smallness terms along the way that can handled easily
with respect to the H2k norm. Using this notation, we have
Lm,2m r̃2j ≃ L̃1 Lm,2m r̃2j−2 + [Lm,2m , L̃1 ]r̃2j−2 + PO≥1/2 (4.60)
and the bulk of our analysis concerns proper handling of the commutator [Lm,2m , L̃1 ]r̃2j−2 .
Since the commutator is trivial for m = 0, let’s assume that m ≥ 1, and we plan to apply an
induction argument on m. As we will see in the coming pages, the critical terms in the commutator
need to be handled in a particular way where we absorb a critical term into the elliptic operator
L̃1 , and then the rest of the terms are subcritical. Note that if we had arrived at (4.60) with
Lm−1,2(m−2) instead of Lm,2m , this would produce
Lm−1,2(m−1) r̃2j ≃ L̃1 Lm−1,2(m−1) r̃2j−2 + [Lm−1,2(m−1) , L̃1 ]r̃2j−2 + PO≥3/2 (4.61)
where the other terms, at worst, have H2k -order 1 (instead of critical in the m case). For the
inductive hypothesis, we need to assume that Lm−1,2(m−1) “behaves well” when we commute it
with L̃1 in that we have already absorbed the order 1 terms. We will make the following inductive
assumption (recalling from (4.56) that k = m + j):
Assumption 4.8 (Inductive Hypothesis on m − 1).
(1) The terms in [Lm−1,2(m−1) , L̃1 ]r̃2j−2 all have H2(m−1+j) order 21 , except for key order 0 terms
which have already been handled by absorbing into L̃1 Lm−1,2(m−1) r̃2j−2 with the analog of
Proposition 4.10 below.
γ−1 ij
(2) The above assumption also holds for L̃1 replaced by L̃1 + b Γ+r H ∂i r∂j with b ≥ 0.
Remark 4.9. There are several consequences of Assumption 4.8 that we summarize:
(1) The terms in [Lm−1,2(m−1) , L̃1 ]r̃2j−2 have H2(m+j) = H2k order 23 , i.e. they belong to PO≥3/2
(2) The terms in [Lm−1,2(m−1) , L̃1 ]∂ r̃2j−2 belong to PO≥1/2 .
 Luczak 41

(3) Suppose that m ≥ 2. We claim that Assumption 4.8 on the m − 1 case also provides
information on lower order commutators. Indeed, we compute the following using two
different ways:
[r∂ Lm−2,2m−3 , L˜1 ]φ ≃ [Lm−1,2m−2 + Lm−2,2m−3 , L̃1 ]φ = [Lm−1,2m−2 , L̃1 ]φ + [Lm−2,2m−3 , L̃1 ]φ

[r∂ Lm−2,2m−3 , L˜1 ]φ = r∂[Lm−2,2m−3 , L̃1 ]φ + [r∂, L̃1 ]Lm−2,2m−3 φ

(4.62)
Combining, we have
[Lm−1,2m−2 , L̃1 ]φ ≃ r∂[Lm−2,2m−3 , L̃1 ]φ + [r∂, L̃1 ]Lm−2,2m−3 φ + [Lm−2,2m−3 , L̃1 ]φ. (4.63)
Thus, since we know that [Lm−1,2(m−1) , L̃1 ]r̃2j−2 belongs to PO≥3/2 and [Lm−1,2(m−1) , L̃1 ]∂ r̃2j−2
belongs to PO≥1/2 , we can also conclude that [Lm−2,2m−3 , L̃1 ]r̃2j−2 belongs to PO≥3/2 and
[Lm−2,2m−3 , L̃1 ]∂ r̃2j−2 belongs to PO≥1/2 . By repeated use of this argument, a similar
statement also holds for commutators involving Ll,m+l−1 where 1 ≤ l ≤ m − 1.
Now that we have finished discussing the consequences of our inductive hypothesis, we proceed
with the commutator by computing the following
[Lm,2m , L̃1 ]r̃2j−2 = Lm,2m L̃1 r̃2j−2 − L̃1 Lm,2m r̃2j−2 (4.64)
and
 
m,2m m 2m γ − 1 ij 1
L L̃1 r̃2j−2 =r ∂ H (r∂i ∂j r̃2j−2 + ∂i r∂j r̃2j−2 )
Γ+r γ−1
  (4.65)
γ − 1 ij m 2m 1
≃ H r ∂ r∂i ∂j r̃2j−2 + ∂i r∂j r̃2j−2 + PO≥1/2
Γ+r γ−1
γ−1 ij
where we note that any terms where derivatives hit Γ+r H get absorbed into PO≥1/2 (since these
derivatives are not hitting either r or r̃2j−2 which are critical terms at worst). For analyzing the
remaining terms, we continue
  2m
m 2m 1 X 1
r ∂ r∂i ∂j r̃2j−2 + ∂i r∂j r̃2j−2 ≃ rm ∂ l r∂ 2m−l ∂i ∂j r̃2j−2 + ∂ l ∂i r∂ 2m−l ∂j r̃2j−2
γ−1 γ−1
l=0
(4.66)
where we absorbed combinatorial constants with ≃. Now, we plan to separate out any of the terms
that belong to PO≥1/2 from the critical terms which must be handled. When ∂ 2m hits L̃1 , we
observe that the worst terms only occur when 2m derivatives hit r̃2j−2 , or 1 derivative hits r (in the
first term) and 2m − 1 derivatives hit r̃2j−2 . The point is that after the first power of r is removed,
any subsequent derivatives that hit ∂r do not actually contribute to the order. In fact, they are
only taking away derivatives that could have hit r̃2j−2 . Indeed, we can check by computing the
order. When l = 0, we have that
rm r∂ 2m ∂i ∂j r̃2j−2 ≃ rm+1 ∂ 2m+2 (Dt2j−2 r̃) is H2k -critical (O = 0)
1 m (4.67)
r ∂i r∂ 2m ∂j r̃2j−2 ≃ rm ∂ 2m+1 (Dt2j−2 r̃) is H2k -critical (O = 0)
γ−1
using Lemma 2.16. When l = 1, we see that
rm ∂r∂ 2m−1 ∂i ∂j r̃2j−2 ≃ rm ∂ 2m+1 (Dt2j−2 r̃) is H2k -critical (O = 0)
1 m (4.68)
r ∂∂i r∂ 2m−1 ∂j r̃2j−2 ≃ rm ∂ 2m (Dt2j−2 r̃) is H2k -subcritical (O = 1)
γ−1
42 A priori estimates for the linearized system

Thus, in the summation when l ≥ 2, we will get

rm ∂ l r∂ 2m−l ∂i ∂j r̃2j−2 ≃ rm ∂ ≤2m (Dt2j−2 r̃) is H2k -subcritical (O ≥ 1)
1 m l (4.69)
r ∂ ∂i r∂ ≤2m−l ∂j r̃2j−2 ≃ rm ∂ ≤2m−1 (Dt2j−2 r̃) is H2k -subcritical (O ≥ 2)
γ−1
Summarizing our order analysis, we see that many of the terms in the summation will be absorbed
into PO≥1/2 , and we have:
 
m,2m γ − 1 ij m+1 2m 1 m 2m m 2m−1
L L̃1 r̃2j−2 ≃ H r ∂ ∂i ∂j r̃2j−2 + r ∂i r∂ ∂j r̃2j−2 + r ∂r∂ ∂i ∂j r̃2j−2
Γ+r γ−1
+ PO≥1/2
(4.70)
For the other side of the commutator, we have
 
m,2m

γ − 1 ij m 2m 1 m 2m
L̃1 L r̃2j−2 ≃ H r∂i ∂j (r ∂ r̃2j−2 ) + ∂i r∂j (r ∂ r̃2j−2 )
Γ+r γ−1
γ − 1 ij m+1
∂i ∂j ∂ 2m r̃2j−2 + 2r∂i (rm )∂j ∂ 2m r̃2j−2 + r∂i ∂j (rm )∂ 2m r̃2j−2


≃ H r
Γ+r
1
H ij ∂i r rm ∂j ∂ 2m r̃2j−2 + ∂j (rm )∂ 2m r̃2j−2 .



+
Γ+r
(4.71)
Now, if we compute the order of the purple term, we see that
γ − 1 ij
H r∂i ∂j (rm )∂ 2m r̃2j−2 ≃ rm−1 ∂ 2m r̃2j−2 is H2k -subcritical (O ≥ 2)
Γ+r
and thus, we can absorb it into PO≥1/2 . Observing that the blue terms cancel in (4.70) and (4.71),
we have
 
m,2m γ − 1 ij m,2m−1 m,2m 1 m−1,2m
[L , L̃1 ]r̃2j−2 = H ∂rL ∂i ∂j r̃2j−2 − 2m∂i rL ∂j r̃2j−2 − ∂i r∂j rL r̃2j−2
Γ+r γ−1
+ PO≥1/2
(4.72)
where the first two terms take the form Lm,2m+1 r̃2j−2 .
2−γ
0, 2(γ−1)
Then, combining with (4.60), and applying the H norm, we have
∥L̃1 Lm,2m r̃2j−2 ∥ 0,
2−γ ≲ ∥Lm,2m r̃2j ∥ 0,
2−γ + ∥Lm,2m+1 r̃2j−2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1) H 2(γ−1)
m−1,2m
(4.73)
+ ∥L r̃2j−2 ∥ 0,
2−γ + ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)

For the LHS, we can apply Lemma 4.1 to get

∥L̃1 Lm,2m r̃2j−2 ∥ 0,
2−γ ≳ ∥Lm,2m r̃2j−2 ∥ 2,1+
2−γ − ∥Lm,2m r̃2j−2 ∥ 0,
2−γ (4.74)
H 2(γ−1) H 2(γ−1) H 2(γ−1)

Then, adding the blue term to the other side of (4.73), we have
∥Lm,2m r̃2j−2 ∥ 2,1+
2−γ ≲ ∥Lm,2m r̃2j ∥ 0,
2−γ + ∥Lm,2m+1 r̃2j−2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1) H 2(γ−1)
m−1,2m
+ ∥L r̃2j−2 ∥ 0,
2−γ + ∥Lm,2m r̃2j−2 ∥ 0,
2−γ (4.75)
H 2(γ−1) H 2(γ−1)

+ ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1)
 Luczak 43

Now, by adding the following quantity on both sides of our inequality:

m−1
X
∥r̃2j−2 ∥ 0,
2−γ + ∥Ll,m+l r̃2j−2 ∥ 2,1+
2−γ (4.76)
H 2(γ−1) H 2(γ−1)
l=0
we see using the help or Remark 1.14 that the LHS will be equivalent to
∥r̃2j−2 ∥ 2m+2,m+1+
2−γ (4.77)
H 2(γ−1)

For the RHS, we will get the following collection of terms which we number
∥Lm,2m r̃2j ∥ 0,
2−γ + ∥Lm,2m+1 r̃2j−2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)
m−1,2m m,2m
+ ∥L r̃2j−2 ∥ 0,
2−γ + ∥L r̃2j−2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)

m−1
(4.78)
X
∥Ll,m+l r̃2j−2 ∥ 2,1+
2−γ + ∥r̃2j−2 ∥ 0,
2−γ + ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1) H 2(γ−1)
l=0
6
X
:= ∥Lm,2m r̃2j ∥ 0,
2−γ + R6
H 2(γ−1)
a=0
using R1 through R6 .
For the terms R1 , R2 , and R3 , we first observe that
∥Lm,2m+1 r̃2j−2 ∥ 0,
2−γ + ∥Lm−1,2m r̃2j−2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)

+ ∥Lm,2m r̃2j−2 ∥ 0,
2−γ ≲ ∥Lm−1,2m−1 r̃2j−2 ∥ 2,1+
2−γ (4.79)
H 2(γ−1) H 2(γ−1)

= ∥Lm−1,2m−2 ∂ r̃2j−2 ∥ 2,1+

2−γ
H 2(γ−1)

Then, we claim that the following inequality holds:

Proposition 4.10.
∥Lm−1,2m−2 ∂ r̃2j−2 ∥ 2,1+
2−γ ≲ ∥Lm−1,2m−2 ∂ r̃2j ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)
(4.80)
+ ∥Lm−1,2m−2 ∂ r̃2j−2 ∥ 0,
2−γ + ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)

Proof. To prove this proposition, we rely on the smallness of ∂3 r (similar to the work in [4]) as well
as a key step in which we we must absorb a term into L̃1 , creating the updated elliptic operator
L
c̃ . Similar to [4], we rely on localizing in the neighborhood of a boundary point such that
1
′
|∂ ′ r| ≲ A, |∂3 r − 1| ≲ A, |H 3,i | ≲ A (4.81)
where A ≪ 1 is small constant, and primed indices will range over and (i.e. ≃ ∂1 r, ∂2 r). x1 x2 ∂′r
Note that a key assumption is present on the off-diagonal components of H (and a similar simpli-
fication can also be found in [4]). The proof is quite similar, and we only need to show the related
inequalities
∥Lm−1,2m−2 ∂3 r̃2j−2 ∥ 2,1+
2−γ ≲ ∥Lm−1,2m−2 ∂3 r̃2j ∥ 0,
2−γ + ∥Lm−1,2m−2 ∂ ′ r̃2j−2 ∥ 2,1+
2−γ
H 2(γ−1) H 2(γ−1) H 2(γ−1)

+ ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1)
(4.82)
and
∥Lm−1,2m−2 ∂ ′ r̃2j−2 ∥ 2,1+
2−γ ≲ ∥Lm−1,2m−2 ∂ ′ r̃2j ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)
(4.83)
+ ∥Lm−1,2m−2 ∂ ′ r̃2j−2 ∥ 0,
2−γ + ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)
44 A priori estimates for the linearized system

□
Corollary 4.11 (Corollary for Proposition 4.10). Let l ∈ N with 1 ≤ l ≤ m − 1. The following
inequality also holds

∥Ll,m+l−1 ∂ r̃2j−2 ∥ 2,1+

2−γ ≲ ∥Ll,m+l−1 ∂ r̃2j ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)
(4.84)
+ ∥Ll,m+l−1 ∂ r̃2j−2 ∥ 0,
2−γ + ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)

Proof. To prove this corollary, we simply need to follow the proof of Proposition 4.10 combined with
the last consequence in Remark 4.9. Anywhere the inductive hypothesis is used on the commutator
Lm−1,2m−2 , we know that a similar fact holds for Ll,m+l−1 by Remark 4.9. □
Now, we can return to (4.73) - (4.78) and summarize our results for what we have on the LHS
and RHS now that we have estimated the terms R1 , R2 , and R3 using Proposition 4.10. Simplifying
Lm−1,2m−2 ∂ = Lm−1,2m−1 , this yields
∥r̃2j−2 ∥ 2m+2,m+1+
2−γ ≲ ∥Lm,2m r̃2j ∥ 0,
2−γ + ∥Lm−1,2m−1 r̃2j ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1) H 2(γ−1)

m−1
X
+ ∥Lm−1,2m−1 r̃2j−2 ∥ 0,
2−γ + ∥Ll,m+l r̃2j−2 ∥ 2,1+
2−γ (4.85)
H 2(γ−1) H 2(γ−1)
l=0
+ ∥r̃2j−2 ∥ 0,
2−γ + ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)

where only the red terms need to be handled, and we will be keeping R5 and R6 from before on
the RHS. For the first red term, we have
∥Lm−1,2m−1 r̃2j−2 ∥ 0,
2−γ ≲ ∥Lm−1,2m−2 r̃2j−2 ∥ 1,
2−γ
H 2(γ−1) H 2(γ−1)
(4.86)
≲ ∥L̃1 (Lm−1,2m−2 r̃2j−2 )∥ 0,
2−γ
H 2(γ−1)

where we used the elliptic estimate for L̃1 in the last line. Then, we can use (4.8) to commute with
Lm−1,2(m−1) which will only produce additional terms in PO≥1/2 . Finally, we use (4.57) to express
L̃1 r̃2j−2 in terms of r̃2j plus additional P terms. We have
∥Lm−1,2m−1 r̃2j−2 ∥ 0,
2−γ ≲ ∥Lm−1,2m−2 (L̃1 r̃2j−2 ) + PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)

≲ ∥Lm−1,2m−2 (r̃2j + P )∥ 0,
2−γ + ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)
(4.87)
≲ ∥Lm−1,2m−2 r̃2j ∥ 0,
2−γ + ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)
≲ ∥r̃2j ∥ 2m,m+
2−γ + ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)

which is an important estimate for the first red term.

For the final red term, we start by rewriting and then applying Corollary 4.11.
m−1
X m−1
X 
l,m+l−1
∥L ∂ r̃2j−2 ∥ 2,1+ 2−γ ≲ ∥Ll,m+l−1 ∂ r̃2j ∥ 0, 2−γ + ∥Ll,m+l−1 ∂ r̃2j−2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1) H 2(γ−1)
l=0 l=0
m−1
X 
l,m+l l,m+l
≲ ∥L r̃2j ∥ 0,
2−γ + ∥L r̃2j−2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)
l=0
m−1
X
≲ ∥r̃2j ∥ 2m,m+
2−γ + ∥Ll,m+l−1 r̃2j−2 ∥ 1,
2−γ
H 2(γ−1) H 2(γ−1)
l=0
(4.88)
 Luczak 45

where, in the last line, we used an inequality similar to 4.86 for the r̃2j−2 terms. For the remaining
sum, it remains to apply the elliptic estimate for L̃1 and then commute L̃1 with Ll,m+l−1 using
Remark 4.9. We have
m−1
X  m−1
X
l,m+l−1
∥L r̃2j−2 ∥ 1, 2−γ ≲ ∥L̃1 (Ll,m+l−1 r̃2j−2 )∥ 0, 2−γ
H 2(γ−1) H 2(γ−1)
l=0 l=0
m−1
X (4.89)
≲ ∥PO≥1/2 ∥ 0,
2−γ + ∥Ll,m+l−1 (r̃2j + P )∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)
l=0
≲ ∥r̃2j ∥ 2m,m+
2−γ + ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)

where we followed the same argument as (4.87) with Lm−1,2m−1 replaced by Ll,m+l−1 .
Finally combining (4.85) - (4.89), we are now ready to prove our higher order elliptic estimates
for the r̃ equation.
Lemma 4.12 (Higher Order Elliptic estimates for r̃). Under the conditions for Lemma 4.1, the
following inequality holds
k
∥Dt2j r̃∥
X
∥r̃∥ 2k,k+
2−γ ≲ 0,
2−γ + Cε̂ ∥(s̃, r̃, ũ)∥H2k (4.90)
H 2(γ−1) H 2(γ−1)
j=0

where Cε̂ ≪ 1 depends on ε̂.

Proof. Combining (4.85) - (4.89) and summarizing our work in this section, we get the following
inequality:
∥r̃2j−2 ∥ 2m+2,m+1+
2−γ ≲ ∥Lm,2m r̃2j ∥ 0,
2−γ + ∥Lm−1,2m−1 r̃2j ∥ 0,
2−γ + ∥r̃2j ∥ 2m,m+
2−γ
H 2(γ−1) H 2(γ−1) H 2(γ−1) H 2(γ−1)

+ ∥r̃2j−2 ∥ 0,
2−γ + ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1)

≲ ∥r̃2j ∥ 2m,m+
2−γ + ∥r̃2j−2 ∥ 0,
2−γ + ∥PO≥1/2 ∥ 0,
2−γ
H 2(γ−1) H 2(γ−1) H 2(γ−1)

≲ ∥r̃2j ∥ 2m,m+
2−γ + ∥r̃2j−2 ∥ 0,
2−γ + ε̂∥(s̃, r̃, ũ)∥H2k
H 2(γ−1) H 2(γ−1)
(4.91)
which we showed using an induction argument on m ≥ 1, and the PO≥1/2 terms are treated using
(4.59) and Remark 2.6. Then, we recall from (4.56) that m = k − j, so when m = k − 1, we have
j = 1 and
∥r̃∥ 2k,k+
2−γ ≲ ∥Dt2 r̃∥ 2k−2,k−1+
2−γ + ∥r̃∥ 0,
2−γ + ε̂∥(s̃, r̃, ũ)∥H2k (4.92)
H 2(γ−1) H 2(γ−1) H 2(γ−1)

Iterating, we obtain successive inequalities all the way up to ∥Dt2k r̃∥ 0,

2−γ . Combining together,
H 2(γ−1)
we have
k
∥Dt2j r̃∥
X
∥r̃∥ 2k,k+
2−γ ≲ 0,
2−γ + Cε̂ ∥(s̃, r̃, ũ)∥H2k (4.93)
H 2(γ−1) H 2(γ−1)
j=0

where Cε̂ ≪ 1 depends on ε̂. This completes the proof. □

Remark 4.13. We observe that the RHS of the inequality in Lemma 4.12 closely resembles the
2k
higher order wave energy Ewave found in (1.47). After obtaining the matching estimates for Dt2j ũ,
2k
we will have the full Ewave energy on the RHS and we will almost have the proof of Theorem 5.1
on the equivalence between our full energy (1.49) and the H2k norm (1.46).
46 A priori estimates for the linearized system

Next, we apply a similar argument for the ũ equation and the operators L2 with “good” spatial
parts L̃2 and L̃3 . We observe that much of the analysis will be quite similar with L̃1 replaced by
2−γ 2−γ
0, 0, 1 + 0, 1
L̃2 , and H 2(γ−1) replaced by H 2 2(γ−1) = H 2(γ−1) . Following the arguments of Section 4.4, we
will need to also incorporate the L̃3 operator at each step in order to apply Lemma 4.3. We will
state the desired Lemma here for convenience:
Lemma 4.14 (Higher order div-curl estimates for ũ). Under the conditions for Lemma 4.3, the
following inequality holds
 
k
−−→
∥Dt2j ũ∥ 0, 1  + ∥curl ũ∥ 2k−1,k+ 1 + Dε̂ ∥(s̃, r̃, ũ)∥H2k
X
∥ũ∥ 2k,k+ 1 ≲  (4.94)
H 2(γ−1) H 2(γ−1) H 2(γ−1)
j=0

where Dε̂ ≪ 1 depends on ε̂.

Remark 4.15. We observe that the RHS of the inequality in Lemma 4.14 closely resembles the
2k −−→
higher order wave energy Ewave found in (1.47) plus an additional term that depends on curl ũ.
However, in view of Lemma 4.4, this term is precisely bounded by ∥ω̂∥ 2k−1,k+ 1 which is the
H 2(γ−1)
2k
key vorticity piece of Etransport . We note that Lemma 4.14 is used in conjuction with Lemma 4.12
to prove Theorem 5.1 on the equivalence between our full energy (1.49) and the H2k norm (1.46).

5. Energy Equivalence
Theorem 5.1 (Equivalence between E 2k and H2k norms). Let (s̃, r̃, ṽ) be smooth functions in Ω.
If r is positive and uniformly non-degenerate on Γ, then
E 2k (s̃, r̃, ũ) ≈ ∥(s̃, r̃, ũ)∥2H2k
where the equivalence depends on γ and up to 2k derivatives of s, r, and u.
Proof. Let’s begin with the ≲ direction. Using our book-keeping scheme from Section 2, it will
be quite straightforward to track the order of each of the terms as they appear. We plan to use
2−γ
2k, 2(γ−1) + 12 +k
Remark 1.14 in which the H2k norm for (s̃, r̃, ũ) is shown to be equivalent to the H ×
2−γ 2−γ
2k, 2(γ−1) +k 2k, 2(γ−1) + 12 +k
H ×H norm.
2k
We start with Etransport , which we recall as
2k
Etransport (s̃, r̃, ũ) = ∥ω̂∥2 2k−1,
2−γ
+ 1 +k
+ ∥s̃∥2 2k,
2−γ
+ 1 +k (5.1)
H 2(γ−1) 2 H 2(γ−1) 2

First, it is clear from Remark 1.14 that

∥s̃∥2 2k,
2−γ
+ 1 +k
≲ ∥(s̃, r̃, ũ)∥2H2k (5.2)
H 2(γ−1) 2

2k
For the vorticity part of Etransport , we look at the order of each of the terms and observe that
ω̂αβ = ∂α (hũβ ) − ∂β (hũα )
= h(∂α ũβ − ∂β ũα ) + ũβ ∂α h − ũα ∂β h (5.3)
≃ ∂ ũ + ũ
using our book-keeping scheme and recalling that h = Γ+r Γ which is O(1) near the free boundary,
as well as ∂h. We plan to estimate ω̂ using a similar argument as the estimate for ω in (3.36).
Applying a multiindex l with |l| ≤ 2k − 1 and counting the number of derivatives and powers of r,
we have
∂ l ω̂ ≃ ∂ l+1 ũ + ∂ l ũ (5.4)
 Luczak 47

2−γ 2−γ
+2k+1 0, + 1 +k
Applying the L2 (r γ−1 ) = H 2(γ−1) 2 norm, we have
Z Z Z
2−γ 1 2−γ 1 2−γ 1
r γ−1 |rk+ 2 ∂ l ω̂|2G dx ≲ r γ−1 |rk+ 2 ∂ l+1 ũ|2 dx + r γ−1 |rk+ 2 ∂ l ũ|2 dx (5.5)
Ωt Ωt Ωt
2−γ
2k−1, + 1 +k
Summing over l, we obtain the H 2(γ−1) 2 norm, and we will get terms that are only critical
1
k+ 2 2k
at worst, i.e. they will have the form r ∂ ũ modulo coefficients depending on derivatives of s
1
and r, and the second intergral produces subcritical terms of the form rk+ 2 ∂ 2k−1 ũ at worst. Each
of these can easily be estimated via the H2k norm, so we have
∥ω̂∥2 2k−1,
2−γ
+ 1 +k
≲ ∥(s̃, r̃, ũ)∥2H2k . (5.6)
H 2(γ−1) 2

and combining with (5.2) yields

2k
Etransport (s̃, r̃, ũ) ≲ ∥(s̃, r̃, ũ)∥2H2k . (5.7)
2k , we first use Remark 1.13 to simplfy
For Ewave
k
∥(Dt2j r̃, Dt2j ũ)∥2He
X
2k
Ewave (s̃, r̃, ũ) =
j=0
(5.8)
k  
∥Dt2j r̃∥2 ∥Dt2j ũ∥2 0, 2−γ + 1
X
≃ 0,
2−γ +
j=0 H 2(γ−1) H 2(γ−1) 2
2−γ 2−γ 2−γ 2−γ
+1 0, 0, +1
using the fact that norms for H̃, (L2 (r γ−1 ) × L2 (r γ−1 )), and (H 2(γ−1) × H 2(γ−1) 2 ) are equiv-
alent. Then, we make frequent use of Lemma 2.13 which allows us to greatly simplify convective
derivatives. Taking 0 ≤ j ≤ k, we have
j 2

∥Dt2j r̃∥2 0, 2−γ

X
≃ rl ∂ l+j r̃ ≲ ∥r̃∥2 2j,
2−γ
+j
(5.9)
2−γ
H 2(γ−1) l=0
0,
2(γ−1) H 2(γ−1)
H
noting that each term appearing in the summation will be critical at the j level. Then, further
applying our embedding lemmas, we get
∥r̃∥2 2j,
2−γ
+j
≲ ∥r̃∥2 2j+k−j,
2−γ
+j+k−j
H 2(γ−1) H 2(γ−1)
2
= ∥r̃∥ k+j,
2−γ
+k (5.10)
H 2(γ−1)
2
≲ ∥r̃∥ 2k,
2−γ
+k
≲ ∥(s̃, r̃, ũ)∥2H2k
H 2(γ−1)

noting that 0 ≤ j ≤ k and we can apply Corollary 2.4 as needed when j < k. Thus, the r̃ part of
2k
Ewave is bounded by the desired H2k norm. For the ũ part, we can similarly invoke Lemma 2.13
again, where the only difference is that we now have terms involving both ũ and r̃. However, the
extra r1/2 weight is exactly what we need to handle any of the terms that appear:
j j−1 2

∥Dt2j ũ∥2 0, 2−γ + 1

X X
l l+j i i+j
≃ r∂ ũ + r∂ r̃
H 2(γ−1) 2 l=0 i=0 0,
2−γ
+1
2(γ−1) 2
(5.11)
H
≲ ∥(s̃, r̃, ũ)∥2H2k
and we observe that the r̃ terms actually have an extra r1/2 in terms of bounding by the H2k norm.
Combining (5.8)-(5.11) produces
2k
Ewave (s̃, r̃, ũ) ≲ ∥(s̃, r̃, ũ)∥2H2k . (5.12)
Thus, we combine with (5.7) and the ≲ direction for Theorem 5.1 is complete.
48 A priori estimates for the linearized system

For the ≳ direction, we make use of many previously established lemmas. First, by the norm
equivalence from Remark 1.14, and Lemmas 4.12, and 4.14, we have
∥(s̃, r̃, ũ)∥2H2k ≈ ∥r̃∥2 2k,
2−γ
+k
+ ∥ũ∥2 2k,
2−γ
+ 1 +k
+ ∥s̃∥2 2k,
2−γ
+ 1 +k
H 2(γ−1) H 2(γ−1) 2 H 2(γ−1) 2

k  
∥Dt2j r̃∥2 0, 2−γ ∥Dt2j ũ∥2 0, 2−γ + 1
X
≲ +
j=0 H 2(γ−1) H 2(γ−1) 2 (5.13)
−−→
+ ∥curl ũ∥2 2k−1,k+ 1 + ∥s̃∥2 2k,
2−γ
+ 1 +k
H 2(γ−1) 2(γ−1) 2
H
+ (Cε̂2 + Dε̂2 )∥(s̃, r̃, ũ)∥2H2k
2k
where Cε̂ , Dε̂ ≪ 1. Then, we observe by examining Ewave in (1.47) that
k   X k
2j 2 2j
∥(Dt2j r̃, Dt2j ũ)∥2He = Ewave
X
2 2k
∥Dt r̃∥ 0, 2−γ + ∥Dt ũ∥ 0, 2−γ + 1 ≲ (s̃, r̃, ũ) (5.14)
H 2(γ−1) H 2(γ−1) 2
j=0 j=0

as any weights involving Γ can be removed in view of Remark 1.13. Then, combining (5.13), (5.14),
−−→
and applying Lemma 4.4 for the term with curl ũ, we have
∥(s̃, r̃, ũ)∥2H2k ≲ Ewave
2k
(s̃, r̃, ũ) + ∥ω̂∥2 2k−1,
2−γ
+ 1 +k
+ ∥s̃∥2 2k,
2−γ
+ 1 +k
H 2(γ−1) 2 H 2(γ−1) 2

+ ε̂2 ∥ũ∥2 2k,

2−γ
+ 1 +k
+ (Cε̂2 + Dε̂2 )∥(s̃, r̃, ũ)∥2H2k (5.15)
H 2(γ−1) 2

2k 2k
≲ Ewave (s̃, r̃, ũ) + Etransport (s̃, r̃, ũ) + (Cε̂2 + Dε̂2 + ε̂2 )∥(s̃, r̃, ũ)∥2H2k
Absorbing the constants (Cε̂2 + Dε̂2 + ε̂2 ) ≪ 1 into the LHS produces
E 2k (s̃, r̃, ũ) = Ewave
2k 2k
(s̃, r̃, ũ) + Etransport (s̃, r̃, ũ) ≳ ∥(s̃, r̃, ũ)∥2H2k (5.16)
which completes the proof of the ≳ direction. Thus, the proof of Theorem 5.1 is complete. □

6. Higher order wave energy estimates

Using the book-keeping scheme from Section 2, let’s estimate each of the terms on the RHS of
2−γ
1
(2.1) with the following Theorem. Note that the B2k equation will get multiplied by γ−1 r γ−1 Dt2k r̃,
1
α equation will get contracted with (Γ + r)r γ−1 G D 2k ũβ .
and the C2k αβ t
We plan to integrate and estimate using the H2k norm which we recall:
2k k
2−γ 2−γ 2−γ
+ 12 +a α +a α + 12 +a α
X X
∥(s̃, r̃, ũ)∥2H2k = ∥r 2(γ−1) ∂ s̃∥2L2 + ∥r 2(γ−1) ∂ r̃∥2L2 + ∥r 2(γ−1) ∂ ũ∥2L2
|α|=0 a=0
|α|−a≤k

The following book-keeping remark is useful when taking convective derivatives of the many
expressions that appear:
Remark 6.1. Using the fact that Dt (Γ(s)) = 0 from (1.35a), we have the following observations:
   2
1 1 1
Dt =− Dt r ≃ r
Γ+r (Γ + r)2 Γ+r
1
where we applied (2.10). Thus, taking multiple convective derivatives of our weight Γ+r leads to
gaining additional powers of r, and thus makes the terms more subcritical. In view of Remarks 2.5
and 2.9, we will often ignore these terms and use the ≃ symbol so that we can highlight the key
critical/supercritical terms that cause difficulty when estimating.
 Luczak 49

Theorem 6.2 (Higher Order energy estimates). Let (s, r, u) be a solution to (1.35) that exists on
some time interval [0, T ]. Assume that s, r, u and their 2k + 1 order derivatives are bounded in the
L∞ (Ωt ) norm for each t ∈ [0, T ] and r vanishes simply on the free boundary. Then, the following
estimate holds for solutions (s̃, r̃, ũ) to the higher order linearized equations (2.1):
d 2k
E (s̃, r̃, ũ) ≲ B∥(s̃, r̃, ũ)∥2H2k (6.1)
dt wave
where B is a function that depends on up to 2k + 1 derivatives of s, r, and u.

Proof. The proof relies on Section 1.6 and the basic energy estimate in Proposition 1.18. Along
the way, we will make use of Lemmas 2.7, 2.10, 2.13, and 2.16.
2−γ
1
We start multiplying (2.1a) and (2.1b) by their respective multipliers γ−1 r γ−1 Dt2k r̃ and (Γ +
1
r)r γ−1 Gαβ Dt2k ũβ . Then, we compare with (1.50b) and (1.50c) and observe that we will be following
the Proof of Proposition 1.18 with r̃ and ũ replaced by Dt2k r̃ and Dt2k ũ. Integrating over Ωt and
applying the moving domain formula (1.53), it remains to show that all of the terms on the RHS
can be estimated using the H2k norm. More specifically, we plan to show that
Z
2−γ 1
r γ−1 B2k D2k r̃ dx ≲ C1 ∥(s̃, r̃, ũ)∥2H2k (6.2a)
Ωt γ−1 t
Z
2−γ
α
r γ−1 C2k (Γ + r)rGαβ Dt2k ũβ dx ≲ C2 ∥(s̃, r̃, ũ)∥2H2k (6.2b)
Ωt

where B2k and α

C2k are long expressions defined in (2.2). Recall from Remarks 2.6 and 2.8 that when
calculating the order of a given term where γ > 1 is arbitrary, we simply ignore powers of r coming
2−γ 2−γ 2−γ
0, 2(γ−1)
from the weight r γ−1 , since our estimates are built around the weighted space L2 (r γ−1 ) = H .
2−γ
Also, recall Lemma 2.7 and Remark 2.8 where estimates involving L1 (r ) are used extensively.
γ−1

We will begin by showing (6.2a). First, observe that the multiplier Dt2k r̃ is H2k critical since by
Lemma 2.13, we have
Dt2k r̃ ≃ (rk ∂ 2k r̃ + rk−1 ∂ 2k−1 r̃ + ... + ∂ k r̃) + ...
where one can quickly check the order of the terms listed as being H2k critical, and all remaining
terms are subcritical.
For the first term in (6.2a) after expanding B2k using (2.2), we get
Dt2k r̃Dt2k g = Dt2k r̃Dt2k (r̃∂µ uµ )
2k
X (6.3)
= Dt2k r̃ Dti r̃Dt2k−i (∂µ uµ )
i=0

Then, by computing the order of terms in Lemma 2.13, we see that Dti r̃ only contains H2k -critical
2−γ
terms when i = 2k. After integrating in L1 (r γ−1 ), the product can estimated by making use of
Lemma 2.7 since we have the product of two terms which are both critical:
Z 2k
2−γ X
r γ−1 Dt2k r̃ Dti r̃Dt2k−i (∂µ uµ ) dx ≲ C1 ∥(s̃, r̃, ũ)∥2H2k (6.4)
Ωt i=0

where C1 depends on the L∞ norms of up to 2k + 1 derivatives of (s, r, u). Notice that in the
coefficient expression Dt2k−i (∂µ uµ ), we can use Lemma 2.14 to solve for the time derivative of u in
terms of spatial derivatives. Additionally, we can solve for Dt2k−i (∂u) in terms of spatial derivatives
of u with additional powers of r in a similar way as Lemma 2.13 be returning to (1.35c).
50 A priori estimates for the linearized system

For the second term in (6.2a), we get

2k−1
X  

Dt2k r̃ Dti ũµ ∂µ Dt2k−i r − Dt2k−i r∂µ Dti ũµ (6.5)
i=0

where each piece will be handled separately. For the first piece with Dti ũ, we see by Lemma 2.10
that Dti ũ only contains critical terms at worst when i = 2k − 1 , and all other terms are subcritical.
Thus, the product can be estimated using Lemma 2.7 since Dt2k r̃ is also critical:
Z 2k−1
2−γ X  
r γ−1 2k
Dt r̃ Dti ũµ ∂µ Dt2k−i r dx ≲ C2,1 ∥(s̃, r̃, ũ)∥2H2k (6.6)
Ωt i=0

For the second piece, with we see that Dt2k−i r contains one power of r by (2.10), and up
∂µ (Dti ũµ )
to 2k − i derivatives of r and u. Similar to the first piece, it suffices to check when i = 2k − 1 in
which case Dt2k−1 ũ contains terms that are H2k critical, and terms that are subcritical with order
α = 21 . Simplifying the expression when i = 2k − 1 and solving for time derivatives in terms of
spatial derivatives with Lemma 2.14, we will get
  
2k−1 µ 1
Dt r∂µ (Dt ũ ) ≃ r∂ (terms with order α = 0) + terms with order α = (6.7)
2
Then, by applying Lemma 2.16, we see that applying r∂ to any free boundary term will produce a
new term with the same order, i.e. here we have α = 0 and α = 21 respectively. Thus, the second
piece can be estimated using Lemma 2.7
Z 2k−1
2−γ X
2k
Dt2k−i r∂µ Dti ũ dx ≲ C2,2 ∥(s̃, r̃, ũ)∥2H2k


r γ−1 Dt r̃ (6.8)
Ωt i=0

For the third term in (6.2a), the estimate is quite similar to before as Dti ũ only contains critical
terms when i = 2k − 1. By Lemma 2.10, we get
[∂µ , Dt2k−i ]r = C2k−i−1
ν
∂ν (Dt2k−i−1 r) + C2k−i−2
ν
∂ν (Dt2k−i−2 r)
+ ...
(6.9)
+ C1ν ∂ν (Dt r)
+ C0ν ∂ν r
which consists of derivatives of u and r. Then, by Lemma 2.7, the estimate will look like
Z 2k−1
2−γ X
r γ−1 Dt2k r̃ Dti ũµ [Dt2k−i , ∂µ ]r dx ≲ C3 ∥(s̃, r̃, ũ)∥2H2k (6.10)
Ωt i=0

since we have the product of terms that are both H2k -critical at worst.
To estimate the G2k fourth term, we will use the commutator identities from Lemma 2.10:
 
2k 2k i−1
Cjν ∂ν (Dtj ũµ )
X X X
Dt2k r̃ Dt2k−i r[Dti , ∂µ ]ũµ ≃ Dt2k r̃ Dt2k−i r  (6.11)
i=1 i=1 j=0

Now, 0
 the worst possible terms only occur when i = 2k, in which case we will get Dt r = r and
Dt2k r̃ r∂(Dt2k−1 ũ) . This term can be handled in a similar way as (6.7) since we get the product
of terms which are both critical. In fact, when i ≤ 2k − 1, we get that all terms are subcritical and
easily estimated. Thus, we have
Z 2k
2−γ X
r γ−1 2k
Dt r̃ Dt2k−i r[Dti , ∂µ ]ũµ dx ≲ C4 ∥(s̃, r̃, ũ)∥2H2k (6.12)
Ωt i=1
 Luczak 51

Combining the previous inequalities proves (6.2a).

Let’s turn now to the C2k α terms occuring in (6.2b). Recall from (6.2b) that the C α has the
2k
multiplier (Γ + r)rGαβ Dt ũβ . In general, we observe using Lemma 2.13 that the multiplier rDt2k ũβ
2k

has H2k subcritical terms with order 1/2 at worst since

k
X k−1
X
Dt2k ũ ≃ rl ∂ l+k ũ + rj ∂ j+k r̃
l=0 j=0
≃ (α = −1/2 supercritical terms) + (α = 0 critical terms)
k
X k−1
X
=⇒ rDt2k ũ ≃ rl+1 ∂ l+k ũ + rj+1 ∂ j+k r̃
l=0 j=0
≃ (α = 1/2 subcritical terms) + (α = 1 subcritical terms)
For the first term in (6.2b), we get
(Γ + r)rGαβ Dt2k ũβ Dt2k hα (6.13)
and we recall the from (1.37) that
Γ′
 
2k α 2k µ α 1 α µ α µ αµ 1 αµ
Dt h = Dt −ũ ∂µ u − (ũ u + u ũ )∂µ r + s̃Π ∂µ r + r̃Π ∂µ r
Γ+r (Γ + r)2 (Γ + r)2
2k 2k
X 1 X i α µ
≃− Dti ũµ Dt2k−i (∂µ uα ) − Dt (ũ u + uα ũµ )Dt2k−i (∂µ r)
Γ+r
i=0 i=0
2k 2k
Γ′ X 1 X
+ Dti s̃Dt2k−i (Παµ ∂µ r) + Dti r̃Dt2k−i (Παµ ∂µ r)
(Γ + r)2 (Γ + r)2
i=0 i=0
(6.14)
 
1 1
where we recall that Dt2k Γ+r ≃ − (Γ+r) 2k
2 Dt r using Remark 6.1, and this only contributes

additional powers of r. After observing Dt2k hα , we see that it contains terms of the form Dt2k (ũ, s̃, r̃)
which are order − 21 supercritical at worst. However, those terms all get multiplied by rDt2k ũ which
has order 12 . Thus, by Lemma 2.7, these terms are also estimated using the H2k norm and we can
solve for any expressions like Dt2k−i (Παµ ∂µ r) in terms of spatial derivatives of u, r, and s.
For the second term in (6.2b), we get the following:
2k−1  
2k β
X
2k−i 1 αµ
∂µ Dti r̃


(Γ + r)rGαβ Dt ũ Dt Π (6.15)
Γ+r
i=0

Now, when i = 2k − 1, Dt2k−1 r̃ contains terms that have order 12 at worst by Remarks 2.5 and
2.17. This implies that ∂(Dt2k−1 r̃) will contain terms that are order − 21 at worst by Lemma 2.16.
However, the multiplier rDt2k ũhas order 12 , which is just enough to allow us to apply Lemma 2.7.
Thus, we can estimate this term using the H2k norm.
For the third and final term in (6.2b), we can follow the analysis of the second term and use
Lemma 2.10 to get
2k  
2k β
X
2k−i 1 αµ
(Γ + r)rGαβ Dt ũ Dt Π [∂µ , Dti ]r̃
Γ+r
i=1
  (6.16)
2k   X i−1
1
Cjν ∂ν (Dtj r̃)
X
≃ (Γ + r)Gαβ rDt2k ũβ Dt2k−i Παµ 
Γ+r
i=1 j=0
52 A priori estimates for the linearized system

Since the worst terms appear when i = 2k, we see that once again Dt2k−1 r̃ has terms with order
1
2 at worst, thus, ∂(Dt2k−1 r̃) has terms with − 12 order. However, the multiplier has order 12 , and
this allows us to apply Lemma 2.7 as usual. This completes estimate (6.2b), and all higher order
terms on the RHS in (2.1) have been handled. □

7. Main Theorem
Combining the previous sections, we have reached our final result:
Theorem 7.1 (Estimates in H2k ). Let (s, r, u) be a smooth solution to (1.35) that exists on some
time interval [0, T ], and for which the physical vacuum boundary condition (1.13) holds. Let
(s̃0 , r̃0 , ũ0 ) be initial data to system (1.36). Then, there exists a constant C depending only on
s, r, u, and T such that, if (s̃, r̃, ũ) ∈ C ∞ (D) is a solution to (1.36) on [0, T ], then
∥(s̃, r̃, ũ)∥H2k (Ωt ) ≲ C∥(s̃0 , r̃0 , ũ0 )∥H2k (Ω0 ) (7.1)
where H2k (Ωt ) is defined in (1.46).
Proof. Combining Proposition 3.1 and Theorem 6.2, we get the following
 
d 2
∥s̃∥ 2k,k+ 2−γ + 1 + Ewave (s̃, r̃, ũ) ≲ C1 ∥(s̃, r̃, ũ)∥2H2k .
2k
(7.2)
dt H 2(γ−1) 2

where C1 is a constant depending on L∞ (Ωt ) norm of up to 2k + 1 derivatives of s, r, and u. Then,

d
after adding dt ∥ω̂∥2 2k−1,k+ 2−γ + 1 to both sides and recalling (1.48) and (1.49), we have
H 2(γ−1) 2

d 2k d
Etotal (s̃, r̃, ũ) ≲ C1 ∥(s̃, r̃, ũ)∥2H2k + ∥ω̂∥2 2k−1,k+ 2−γ + 1 . (7.3)
dt dt H 2(γ−1) 2

Integrating both sides and using Theorem 5.1 on the energy equivalence, we have
Z t
2 2
∥(s̃, r̃, ũ)∥H2k (Ωt ) −∥(s̃0 , r̃0 , ũ0 )∥H2k (Ω0 ) ≲ C1 ∥(s̃, r̃, ũ)∥2H2k (Ωτ ) dτ +∥ω̂∥2 2k−1,k+ 2−γ +1
. (7.4)
0 H 2(γ−1) 2 (Ωt )
Applying Theorem 3.4 for the ω̂ term and rearranging, we get
Z t
2 2
∥(s̃, r̃, ũ)∥H2k (Ωt ) ≲ ∥(s̃0 , r̃0 , ũ0 )∥H2k (Ω0 ) + C1 ∥(s̃, r̃, ũ)∥2H2k (Ωτ ) dτ
0
Z t
2
+ ∥ω̃0 ∥ 2k−1,k+ 1 + ε̂∥(s̃, r̃, ũ)∥2H2k (Ωt ) + C2 ∥(s̃, r̃, ũ)∥2H2k (Ωτ ) dτ
H 2(γ−1) (Ω )
0 0
(7.5)
Then, combining the initial data ω̃0 with the initial data term ∥(s̃0 , r̃0 , ũ0 )∥2H2k (Ω0 ) , and soaking the
term with ε̂ to the LHS, we arrive at
 Z t 
2 1 2 2
∥(s̃, r̃, ũ)∥H2k (Ωt ) ≲ ∥(s̃0 , r̃0 , ũ0 )∥H2k (Ω0 ) + C3 ∥(s̃, r̃, ũ)∥H2k (Ωτ ) dτ (7.6)
1 − ε̂ 0
Finally, the desired result is a straightforward application of Gronwall’s inequality.
□
 Luczak 53

8. Appendix for Lemma 2.14

Here, we record some of the computations for proving Lemma 2.14. Using (1.35) directly, we
have
ui
∂t s = − ∂i s
u0
ui γ−1 γ−1
∂t r = − 0 ∂i r − 0
r∂t u0 − r∂i ui (8.1)
u u u0
ui 1 1
∂t uα = − 0 ∂i uα − 0
Πα0 ∂t r − Παi ∂i r
u (Γ + r)u (Γ + r)u0
Starting from (8.1), we can further solve for ∂t u0 by plugging in α = 0 into the last equation. If
γ−1
we set a1 = u0 − (Γ+r)u 00
0 rΠ , then

 
1 γ−1 1 1
∂t u0 = −ui ∂i u0 + rΠ 00
∂ i ui
+ Π 00 i
u ∂ i r − Π 0i
∂ i r
a1 (Γ + r)u0 (Γ + r)u0 Γ+r
1 1 (γ − 1)((u0 )2 − 1)
= − ui ∂i u0 − u i
∂ i r + r∂i ui
a1 a1 (Γ + r)u0 a1 (Γ + r)u0
(8.2)
:= C1i ∂i u0 + C2i ∂i r + C3 r∂i ui
≃ ∂u0 + ∂r + r∂u
=⇒ ∂t ũ0 = C1i ∂i ũ0 + C2i ∂i r̃ + C3 r∂i ũi + C3 r̃∂i ui + C
fi ∂ u0 + C
1 i
f3 r∂i ui
fi ∂ r + C
2 i

We have the computations

(γ − 1)((u0 )2 − 1)
a1 = u0 − r (8.3)
(Γ + r)u0
(γ − 1)((u0 )2 − 1)
=⇒ ã1 = ũ0 − r̃ (8.4)
(Γ + r)u0
2(Γ + r)(γ − 1)u0 ũ0 − (γ − 1)((u0 )2 − 1)[(Γ′ s̃ + r̃)u0 + (Γ + r)ũ0 ]
 
−r (8.5)
(Γ + r)2 (u0 )2
and
ui i
fi = u ã1 − a1 ũ
i
C1i := − =⇒ C 1 (8.6)
a1 (a1 )2
1 1 ′ 0 0
C2i := C i
=⇒ fi =
C fi − [(Γ s̃ + r̃)u + (Γ + r)ũ ] C i
C
(Γ + r)u0 1 2
(Γ + r)u0 1 (Γ + r)2 (u0 )2 1

(8.7)
(γ − 1)((u0 )2 − 1) i 0 2
f3 = (γ − 1)((u ) − 1) C
fi
C3 := i
C2 =⇒ C i 2 (8.8)
u u
ui 2(γ − 1)ũ0 − (γ − 1)((u0 )2 − 1)ũi i
+ C2 . (8.9)
(ui )2
Note that after analyzing all of the terms (with k = 1 here), we see that a large number are
subcritical, and we can simplify:
∂t ũ0 ≃ ∂ ũ0 + ∂ r̃ + r∂ ũ + r̃∂u (8.10)
1 1
≃ (H2 super-critical of order − ) + (H2 critical) + (H2 sub-critical of order ) (8.11)
2 2
54 A priori estimates for the linearized system

Similiary, we can simplify ∂t r and ∂t r̃ in the following way:

ui γ−1 γ−1
r∂i ui − r C1i ∂i u0 + C2i ∂i r + C3 r∂i ui


∂t r = −0
∂i r − 0 0
u u u
u i γ−1 γ−1
i 0


=⇒ ∂t r̃ = − 0 ∂i r̃ − r∂ i ũ − r ∂ t ũ
u u0 u0
i 0 0 i
   
u ũ − u ũ (γ − 1) 0 γ−1 i (γ − 1) 0 γ−1
+ ∂i r + ũ r − r̃ ∂i u + ũ r − r̃ ∂t u0
(u0 )2 (u0 )2 u0 (u0 )2 u0
≃ ∂ r̃ + r∂ ũ + r∂ ũ0 + r∂ r̃ + r2 ∂ ũ + (additional subcritical terms)
1 ui 1
∂t uj = − 0
Π j0
[∂ t r] − 0
∂i uj − Πji ∂i r
(Γ + r)u u (Γ + r)u0
≃ −[∂t r] − ∂u − ∂r
(8.12)
The expressions for ∂t s and ∂t s̃ take the easiest form, as
ui
∂t s = − ∂i s
u0
(8.13)
ui ui ũ0 − u0 ũi
∂t s̃ = − 0 ∂i s̃ + ∂i s
u (u0 )2
We also see that (u0 )2 = 1 + uj uj , so
uj j
ũ0 = ũ
u0
uj uj (8.14)
∂i ũ0 = 0 ∂i ũj + ũj ∂i
u u0
Thus, linearizing our expression for ∂t uj , we get

∂t ũj ≃ [∂t r̃] + ∂ ũ + ∂ r̃ (8.15)

2
≃ ∂ ũ + ∂ r̃ + (r∂ ũ + r∂ r̃ + r ∂ ũ + (additional subcritical terms)) (8.16)
Starting from (8.1), we can further solve for ∂t u0 by plugging in α = 0 into the last equation. If
γ−1
we set a1 = u0 − (Γ+r)u 00
0 rΠ , then

 
0 1 i 0 γ−1 00 i 1 00 i 1 0i
∂t u = −u ∂i u + rΠ ∂i u + Π u ∂i r − Π ∂i r
a1 (Γ + r)u0 (Γ + r)u0 Γ+r
1 1 (γ − 1)((u0 )2 − 1)
= − ui ∂i u0 − u i
∂ i r + r∂i ui
a1 a1 (Γ + r)u0 a1 (Γ + r)u0
(8.17)
:= C1i ∂i u0 + C2i ∂i r + C3 r∂i ui
≃ ∂u0 + ∂r + r∂u
=⇒ ∂t ũ0 = C1i ∂i ũ0 + C2i ∂i r̃ + C3 r∂i ũi + C3 r̃∂i ui + C
fi ∂ u0 + C
1 i
f3 r∂i ui
fi ∂ r + C
2 i

We have the computations

(γ − 1)((u0 )2 − 1)
a1 = u0 − r (8.18)
(Γ + r)u0
(γ − 1)((u0 )2 − 1)
=⇒ ã1 = ũ0 − r̃ (8.19)
(Γ + r)u0
 Luczak 55

2(Γ + r)(γ − 1)u0 ũ0 − (γ − 1)((u0 )2 − 1)[(Γ′ s̃ + r̃)u0 + (Γ + r)ũ0 ]

 
−r (8.20)
(Γ + r)2 (u0 )2
and
ui i
fi = u ã1 − a1 ũ
i
C1i := − =⇒ C 1 (8.21)
a1 (a1 )2
1 1 ′ 0 0
C2i := C i
=⇒ fi =
C fi − [(Γ s̃ + r̃)u + (Γ + r)ũ ] C i
C
(Γ + r)u0 1 2
(Γ + r)u0 1 (Γ + r)2 (u0 )2 1

(8.22)
(γ − 1)((u0 )2 − 1) i 0 2
f3 = (γ − 1)((u ) − 1) C
fi
C3 := i
C2 =⇒ C i 2 (8.23)
u u
ui 2(γ − 1)ũ0 − (γ − 1)((u0 )2 − 1)ũi i
+ C2 . (8.24)
(ui )2
Note that after analyzing all of the terms (with k = 1 here), we see that a large number are
subcritical, and we can simplify:
∂t ũ0 ≃ ∂ ũ0 + ∂ r̃ + r∂ ũ + r̃∂u (8.25)
1 1
≃ (H2 super-critical of order − ) + (H2 critical) + (H2 sub-critical of order ) (8.26)
2 2
Similiary, we can simplify ∂t r and ∂t r̃ in the following way:
ui γ−1 i γ−1 i 0 i i


∂t r = −∂ i r − r∂ i u − r C 1 ∂ i u + C 2 ∂ i r + C 3 r∂i u
u0 u0 u0
u i γ−1 γ−1
i 0


=⇒ ∂t r̃ = − 0 ∂i r̃ − r∂ i ũ − r ∂ t ũ
u u0 u0
i 0 0 i
   
u ũ − u ũ (γ − 1) 0 γ−1 i (γ − 1) 0 γ−1
+ ∂i r + ũ r − r̃ ∂i u + ũ r − r̃ ∂t u0
(u0 )2 (u0 )2 u0 (u0 )2 u0
≃ ∂ r̃ + r∂ ũ + r∂ ũ0 + r∂ r̃ + r2 ∂ ũ + (additional subcritical terms)
1 ui 1
∂t uj = − 0
Π j0
[∂ t r] − 0
∂i uj − Πji ∂i r
(Γ + r)u u (Γ + r)u0
≃ −[∂t r] − ∂u − ∂r
(8.27)
The expressions for ∂t s and ∂t s̃ take the easiest form, as
ui
∂t s = − ∂i s
u0
(8.28)
ui ui ũ0 − u0 ũi
∂t s̃ = − 0 ∂i s̃ + ∂i s
u (u0 )2
We also see that (u0 )2 = 1 + uj uj , so
uj j
ũ0 = ũ
u0
uj uj (8.29)
∂i ũ0 = 0 ∂i ũj + ũj ∂i
u u0
Thus, linearizing our expression for ∂t uj , we get
∂t ũj ≃ [∂t r̃] + ∂ ũ + ∂ r̃ (8.30)
2
≃ ∂ ũ + ∂ r̃ + (r∂ ũ + r∂ r̃ + r ∂ ũ + (additional subcritical terms)) (8.31)
56 A priori estimates for the linearized system

References
[1] Luis A. Caffarelli and Avner Friedman. Regularity of the free boundary for the one-dimensional flow of gas in a
porous medium. Amer. J. Math., 101(6):1193–1218, 1979.
[2] Yvonne Choquet-Bruhat. General relativity and the Einstein equations. Oxford Mathematical Monographs. Ox-
ford University Press, Oxford, 2009.
[3] Marcelo M. Disconzi. Recent developments in mathematical aspects of relativistic fluids, 2023.
[4] Marcelo M. Disconzi, Mihaela Ifrim, and Daniel Tataru. The relativistic Euler equations with a physical vacuum
boundary: Hadamard local well-posedness, rough solutions, and continuation criterion. Archive for Rational
Mechanics and Analysis, 2022. 55 Pages.
[5] Marcelo M. Disconzi, Igor Kukavica, and Amjad Tuffaha. A Lagrangian interior regularity result for the incom-
pressible free boundary Euler equation with surface tension. SIAM J. Math. Anal., 51(5):3982–4022, 2019.
[6] Mahir Hadžić, Steve Shkoller, and Jared Speck. A priori estimates for solutions to the relativistic Euler equations
with a moving vacuum boundary. Comm. Partial Differential Equations, 44(10):859–906, 2019.
[7] Mihaela Ifrim and Daniel Tataru. The compressible euler equations in a physical vacuum: A comprehensive
eulerian approach. Annales de l’Institut Henri Poincaré C, Analyse non linéaire, 2020.
[8] Juhi Jang, Philippe G. LeFloch, and Nader Masmoudi. Lagrangian formulation and a priori estimates for rela-
tivistic fluid flows with vacuum. J. Differential Equations, 260(6):5481–5509, 2016.
[9] Philippe G. LeFloch and Seiji Ukai. A symmetrization of the relativistic Euler equations with several spatial
variables. Kinet. Relat. Models, 2(2):275–292, 2009.
[10] Tetu Makino and Seiji Ukai. Local smooth solutions of the relativistic Euler equation. II. Kodai Math. J.,
18(2):365–375, 1995.
[11] Šárka Matušocircu-Nečasová, Mari Okada, and Tetu Makino. Free boundary problem for the equation of spher-
ically symmetric motion of viscous gas. III. Japan J. Indust. Appl. Math., 14(2):199–213, 1997.
[12] Shuang Miao and Sohrab Shahshahani. Well-posedness for the free boundary hard phase model in general
relativity. Advances in Mathematics, 443:109614, 2024.
[13] Shuang Miao, Sohrab Shahshahani, and Sijue Wu. Well-posedness of the free boundary hard phase fluids in
minkowski background and its newtonian limit. arXiv:2003.02987 [math.AP], 2020.
[14] J.R. Oppenheimer and G.M. Volkoff. On Massive neutron cores. Phys. Rev., 55:374–381, 1939.
[15] Luciano Rezzolla and Olindo Zanotti. Relativistic Hydrodynamics. Oxford University Press, New York, 2013.
[16] Jalal Shatah and Chongchun Zeng. Geometry and a priori estimates for free boundary problems of the Euler
equation. Comm. Pure Appl. Math., 61(5):698–744, 2008.
[17] Richard C. Tolman. Effect of imhomogeneity on cosmological models. Proc. Nat. Acad. Sci., 20:169–176, 1934.
[18] Richard C. Tolman. Static solutions of Einstein’s field equations for spheres of fluid. Phys. Rev., 55:364–373,
1939.
[19] Yuri Trakhinin. Local existence for the free boundary problem for nonrelativistic and relativistic compressible
Euler equations with a vacuum boundary condition. Comm. Pure Appl. Math., 62(11):1551–1594, 2009.
[20] Ping Zhang and Zhifei Zhang. On the free boundary problem of three-dimensional incompressible Euler equations.
Comm. Pure Appl. Math., 61(7):877–940, 2008.