STABLE NEARLY SELF-SIMILAR BLOWUP OF THE 2D BOUSSINESQ

AND 3D EULER EQUATIONS WITH SMOOTH DATA

JIAJIE CHEN AND THOMAS Y. HOU

Abstract. Inspired by the numerical evidence of a potential 3D Euler singularity [54, 55],
we prove finite time blowup of the 2D Boussinesq and 3D axisymmetric Euler equations with
smooth initial data of finite energy and boundary. There are several essential difficulties in
arXiv:2210.07191v2 [math.AP] 19 Oct 2022

proving finite time blowup of 3D Euler with smooth initial data. One of the essential diffi-
culties is to control a number of nonlocal terms that do not seem to offer any damping effect.
Another essential difficulty is that the strong advection normal to the boundary introduces a
large growth factor for the perturbation if we use weighted L2 estimates. We overcome this
difficulty by using a combination of a weighted L∞ norm and a weighted C 1/2 norm, and
develop sharp functional inequalities using the symmetry properties of the kernels and some
techniques from optimal transport. Moreover we decompose the linearized operator into a
leading order operator plus a finite rank operator. The leading order operator is designed in
such a way that we can obtain sharp stability estimates. The contribution from the finite rank
operator can be captured by an auxiliary variable and its contribution to linear stability can
be estimated by constructing approximate solution in space-time. This enables us to establish
nonlinear stability of the approximate self-similar profile and prove stable nearly self-similar
blowup of the 2D Boussinesq and 3D Euler equations with smooth initial data and boundary.

1. Introduction
The question whether the 3D incompressible Euler equations can develop a finite time singu-
larity from smooth initial data of finite energy is one of the most outstanding open questions in
the theory of nonlinear partial differential equations and fluid dynamics. The main difficulty is
due to the presence of the vortex stretching term in the vorticity equation:
(1.1) ω t + u · ∇ω = ω · ∇u,
where ω = ∇ × u is the vorticity vector of the fluid, and u is related to ω via the Biot-
Savart law. The velocity gradient ∇u formally has the same scaling as vorticity ω. Thus the
vortex stretching term has a nonlocal quadratic nonlinearity in terms of vorticity. However, the
nonlocal nature of the vortex stretching term can lead to dynamic depletion of the nonlinear
vortex stretching, which could prevent a finite time blowup, see e.g. [18, 26, 46]. The interested
readers may consult the excellent surveys [17, 35, 44, 49, 56] and the references therein.
In this paper, we study the blowup of the axisymmetric Euer equations with smooth initial
data and boundary. Denote by ω θ , uθ and φθ the angular vorticity, angular velocity, and angular
stream function, respectively. The 3D axisymmetric Euler equations are given below:
ωθ ωθ ωθ 1
(1.2) ∂t (ruθ ) + ur (ruθ )r + uz (ruθ )z = 0, ) + ur ( )r + uz ( )z = 4 ∂z ((ruθ )2 ),
∂t (
r r r r
where the radial velocity ur and the axial velocity uθ are given by the Biot-Savart law:
1 1 1
(1.3) − (∂rr + ∂r + ∂zz )φθ + 2 φθ = ω θ , ur = −φθz , uz = φθr + φθ ,
r r r
with the no-flow boundary condition φθ (1, z) = 0 on the solid boundary r = 1 and a periodic
boundary condition in z. For 3D Euler blowup that occurs at the boundary r = 1, we know that
the axisymmetric Euler equations have the same scaling properties as those of the 2D Boussinesq
equations [56]. Thus, we also study the 2D Boussinesq equations on the upper half space:
(1.4) ωt + u · ∇ω = θx ,
(1.5) θt + u · ∇θ = 0,

Date: October 20, 2022.

1
2 JIAJIE CHEN AND THOMAS Y. HOU

where the velocity field u = (u, v)T : R2+ × [0, T ) → R2+ is determined via the Biot-Savart law
(1.6) − ∆φ = ω, u = −φy , v = φx ,
where φ is the stream function with the no-flow boundary condition φ(x, 0) = 0 at y = 0. By
making the change of variables θ̃ , (ruθ )2 , ω̃ = ω θ /r, we can see that θ̃ and ω̃ satisfy the 2D
Boussinesq equations up to the leading order for r ≥ r0 > 0.
The main result of this paper is a rigorous proof of the nearly self-similar blowup of the 2D
Boussinesq and the 3D axisymmetric Euler equations with smooth initial data and boundary.
The blowup mechanism of the 2D Boussinesq equations with boundary is essentially the same as
that of the 3D axisymmetric incompressible Euler equations with boundary. Our work is inspired
by the computation of Luo-Hou [54, 55] in which they presented some convincing numerical
evidence that the 3D axisymmetric Euler equations with smooth initial data and boundary
develop a potential finite time singularity. Inspired by the recent breakthrough of Elgindi [27]
(see also [28]) on the blowup of the axisymmetric Euler equations without swirl for C 1,α initial
velocity, we have proved asymptotically self-similar blowup of the 2D Boussinesq equations and
the nearly self-similar blowup of the 3D axisymmetric Euler equations with C 1,α velocity and
boundary in [12]. The blowup analysis presented in [12] takes advantage of the C 1,α velocity in an
essential way and does not generalize to prove the Hou-Luo blowup scenario with smooth initial
data. The results presented in this paper provide the first rigorous proof of stable nearly self-
similar blowup of the 2D Boussinesq and 3D Euler equations with smooth data and boundary.
1.1. Some main ingredients in our analysis. We follow a general strategy that we have
established in our previous works [12–14]. We first reformulate the problem of studying the
finite time blowup of the 2D Boussinesq and 3D Euler equations as the problem of proving
nonlinear stability of an approximate steady state of the dynamic rescaling formulation. A
very important first step is to construct an approximate steady state of the dynamic rescaling
formulation with sufficiently small residual errors. We achieve this by decomposing the solution
into a semi-analytic part that captures the far field behavior of the solution and a numerically
computed part that has compact support. The approximate steady state gives an approximate
self-similar profile. See more discussions in Section 5. We remark that there has been some
recent exciting development of using a physics-informed neural network (PINN) to construct an
approximate steady state of the 2D Boussinesq equations, see [75].
Establishing linear stability of the approximate steady state is the most crucial step in our
blowup analysis. One essential difficulty is that the advection normal to the boundary for
smooth initial data introduces a large growth factor if we use weighted L2 energy estimates
similar to [12–14, 27], see more discussions in Section 2. Such difficulty is absent in the analysis
of 1D models [13, 14] and in the case of C 1,α initial velocity [12] since we gain a small factor
α for the advection term. To overcome the destabilizing effect due to advection normal to the
boundary, we choose a weighted L∞ norm, which allows us to extract the maximal amount of
damping from the local terms without suffering from the destabilizing effect due to advection
normal to the boundary. In order to close the energy estimates, we use a combination of the
weighted L∞ norm and the weighted C 1/2 norm.
To estimate the nonlocal terms effectively, we derive sharp C 1/2 estimates for ∇u using the
symmetry properties of the kernels and some techniques from optimal transport. We note that
novel functional inequalities on similar Biot-Savart laws have played a crucial role in [27, 48].
The authors identify the main term in the Biot-Savart law, which has a much simpler structure,
and use it to control the velocity effectively. The sharp Hölder estimates play a similar role
in our work. We decompose the Biot-Savart law into two parts. The main terms capture the
most singular part of the Biot-Savart law, and the remaining terms are more regular. We can
generalize these estimates to the weighted sharp C 1/2 estimate since the commutators between
the singular weights and the velocity operators are more regular. Compared with [27, 48], our
remaining terms are not small and cannot be controlled by a-priori estimates.
We decompose the linearized equations into three parts. The first part is the main part
and differs from the linearized equations by a finite rank operator K. We choose K to capture
the main contributions from the remaining terms in the Biot-Savart law and estimate K in
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 3

the second part. The stability analysis for the first part can be established using the new
functional inequalities. The second part accounts for the contributions from K to the linearized
equations. Thanks to the finite rank structure of K, we perform space-time estimates with
computer assistance to obtain sharp stability estimates for the second part. The third part is
small and can be treated as a small perturbation. Note that the first part is not small since the
linearized operator is not compact. By integrating these estimates, we establish linear stability.
The main results of this paper are stated by the two informal theorems below. The more
precise and stronger statement of Theorem 1 will be given by Theorem 3 in Section 2 and the
precise statement of Theorem 2 will be given Theorem 4 in Section 6.
Theorem 1. There is a family of smooth initial data (θ0 , ω0 ) with θ0 being even and ω0 being
odd, such that the solution of the Boussinesq equations (1.4)-(1.6) develops a singularity in
finite time T < +∞. The velocity field u0 has finite energy. The blowup solution (θ(t), ω(t))
is nearly self-similar in the sense that (θ(t), ω(t)) with suitable dynamic rescaling is close to
an approximate blowup profile (θ̄, ω̄) up to the blowup time. Moreover, the blowup is stable for
initial data (θ0 , ω0 ) close to (θ̄, ω̄) in some weighted L∞ and C 1/2 norm.
Theorem 2. Consider the 3D axisymmetric Euler equations in the cylinder r, z ∈ [0, 1] × T.
The solution of the 3D Euler equations (1.2)-(1.3) develops a nearly self-similar blowup (in the
sense described in Theorem 1) in finite time for some smooth initial data ω0θ , uθ0 supported away
from the axis r = 0. The initial velocity field has finite energy, uθ0 and ω0θ are odd and periodic in
z. The blowup is stable for initial data (uθ0 , ω0θ ) that are close to the approximate blowup profile
(ūθ , ω̄ θ ) after proper rescaling subject to some constraint on the initial support size.
1.2. A novel framework of analysis with computer assistance. One of our main contri-
butions is to introduce a novel framework of analysis that enables us to obtain sharp stability
estimates by combining sharp functional inequalities, energy estimates, and approximate space-
time solutions constructed numerically with rigorous error control. Such errors are treated as
small perturbations in the energy estimates. Here we give a high level description of the linear
stability analysis using this new framework of analysis. We use the 2D Boussinesq equations
as an example. The same analysis also applies to 3D Euler with minor modifications. More
discussions and motivation will be provided in Section 2. Let ω̄, θ̄ be an approximate steady
state. We denote W = (ω, θx , θy ) and decompose W = W + W f with W = (ω̄, θ̄x , θ̄y ). We further
denote by L the linearized operator around W that governs the perturbation W f in the dynamic
rescaling formulation (see Section 2.5),
(1.7) ft = L(W
W f ),
where we have neglected the contributions from the third part, the nonlinear terms and the
residual error in the above linearized equation. We note that the coefficients of L depend on the
approximate steady state W . We decompose the linearized operator L into the main part L0
plus a finite rank perturbation K, i.e
(1.8) L = L0 + K.
The leading order operator L0 is constructed in such way that we can obtain sharp stability
estimates and extract damping effect by using sharp functional inequalities.
We first perform energy estimates to extract the damping effect from the leading order oper-
ator L0 in the weighted L∞ norm. By choosing appropriate singular weights, we can establish
stability associated with L0 using the weighted L∞ norm with a weak coupling to the weighted
C 1/2 norm. Obtaining sharp energy estimates in the weighted C 1/2 norm is more challenging.
To obtain a sharp bound in the weighted C 1/2 norm, we exploit the symmetry properties of
the kernels and apply optimal transport along the horizontal and the vertical directions sep-
arately to obtain relatively sharp C 1/2 estimates. Since the contribution from the horizontal
direction dominates that from the vertical direction, our strategy gives a relatively tight bound
in the weighted C 1/2 norm.
To estimate the perturbed operator K, we perform space-time estimates with computer as-
sistance. We use the following toy model to illustrate the main ideas by considering K as a
4 JIAJIE CHEN AND THOMAS Y. HOU

rank-one operator K(W f ) = a(x)P (W f ) for some nonlocal bounded linear operator P satisfying
f ) is constant in space and is only a function of time; (ii) kP (W
the properties (i) P (W f )k ≤ ckW
f k.
Given initial data W0 , we decompose (1.7) as follows
f

f1 (t) = L0 W
∂t W f1 , W
f1 (0) = W
f0 ,
(1.9)
f2 (t) = LW
∂t W f2 + a(x)P (W
f1 (t)), W
f2 (0) = 0.

It is easy to see that W

f=W f1 + W
f2 solves (1.7) with initial data W f0 since L = L0 + a(x)P by
our assumption. By construction, the leading operator L0 has the desired structure that enables
us to obtain sharp stability estimates. The second part W f2 is driven by the rank-one forcing
term a(x)P (W f1 (t)). Using Duhamel’s principle, the fact that P (Wf1 (t)) is constant in space, we
yield
Z t
(1.10) W2 (t) =
f f1 (s))eL(t−s) a(x)ds.
P (W
0

Since W f1 (t) = eL0 (t) W

f0 decays in L∞ (ϕ) (ϕ is a singular weight), we can control P (W f1 (s)).
L(t) ∞ ∞
If e a(x) decays in L (ϕ) for large t, we can show that W2 (t) also decays in L (ϕ) and
f
establish stability estimate of W f2 .
A crucial idea in the estimate of K is that we bridge the energy estimates and numerical
PDEs via an approximate solution in space and time. To see this, we note that eL(t) a(x) is
equivalent to solving the linear evolution equation vt = L(v) with initial data v0 = a(x). Due
to the rapid decay of the linearized equation, we only need to solve this initial value problem
using a numerical scheme up to a modest time T1 and pad the solution beyond T1 to be zero.
The residual error of the numerical PDE in space and time is treated as a small perturbation in
the energy estimates. We then construct an approximate solution by interpolating the solution
in time by a cubic polynomial. The stability property of W f1 allows us to control the numerical
L(t)
error in computing e a(x) and obtain sharp stability estimates for W
f2 .
We remark that we only need to modify the linearized operator by a finite rank operator in
a small sector near the boundary where we have the smallest amount of damping. We also add
a few more terms to approximate the regular part of the solution in the bulk of the domain.
The amount of damping improves significantly as we move away from the boundary and in the
far field. Moreover, due to the decay of the far field solution W , we only need to construct a
finite rank operator K in a finite size domain within this narrow sector. The rank of K that
we use is less than 50. To estimate the contributions from the finite rank operator, we need to
solve the linear PDE in space-time with a number of initial data, which can be implemented in
full parallel. We have not explored the full potential of our method. If we try to perturb the
linearized operator by a rank-N operator, we just need to solve the linearized equation with N
initial data, which can be implemented very efficiently using a powerful parallel cluster.
Our stability analysis uses some quantities involving the approximate steady state. The grid
values of these quantities are available with rigorous error bounds using interval arithmetic. Since
we use piecewise polynomials to approximate the steady state, we can obtain rigorous bounds
for the high order derivatives of the steady state. Such bounds in turn provide rigorous bounds
for lower order derivatives, the pointwise values and various integrals involving the approximate
steady state by using standard numerical analysis. The verification involves several case studies.
We will present all the essential tools for verification in the Appendix and also in later sections,
but will defer the actual verification steps of various quantities to the supplementary material.
To pass from the 2D Boussinesq equations to the 3D axisymmetric Euler equations, we follow
the same ideas presented in our previous work [12] by controlling the support of the vorticity
to be in a small region that is close to the boundary and does not intersect the symmetry axis.
The asymptotic scaling properties of the Biot-Savart kernels are exactly the same as those of
the Biot-Savart kernels for the 2D Boussinesq equations up to some asymptotically small terms
after making appropriate changes of variables. We will provide some additional estimates to
control these asymptotically small terms and prove the blowup of the 3D Euler equations.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 5

The stability analysis presented in this paper can be substantially simplified if we have access
to a very powerful parallel cluster. If the rank of K is sufficiently large and the residual errors can
be made arbitrarily small, we can treat the third part of the linearized operator, the nonlinear
terms and the residual errors as small perturbations. Thus, we just need to focus on establishing
linear stability of L0 . Once we establish linear stability, nonlinear stability is relatively easy to
establish since the residual errors are very small. The a priori estimates of the perturbations in
the weighted L∞ and C 1/2 norms make it easier to establish energy estimates in a higher order
norm, e.g. weighted C 3/2 norm. As in [12–14, 27], by differentiating the perturbed equation for
f in time and performing C 3/2 estimates, one can further prove the existence of the asymptot-
W
ically self-similar blowup profile. This provides a very powerful tool to give a constructive proof
of stable finite time blowup for a large class of nonlinear PDEs with smooth initial data.
1.3. Comparison between our method of analysis and the topological argument. Our
method of analysis shares some similarity with the recently developed blowup analysis using a
topological argument, see e.g. [58, 62–64]. In the topological argument, one also constructs a
compact perturbation operator K to the linearized operator L. After subtracting the compact
perturbation operator from the linearized operator, one can establish linear stability of the lead-
ing operator L0 in some Hilbert space. The compact perturbation operator can be approximated
by a finite rank operator. Typically, this compact perturbation operator contains a number of
unstable directions. A topological argument is used to show that there exists a stable trajectory
leading to a finite time blowup, which avoids a finite number of unstable directions. This method
has been used to prove unstable blowup of several nonlinear PDEs with great success.
The main difference between our method of analysis and the topological argument is in the
way we estimate the finite rank operator K. First of all, in our framework, we do not require
the energy space to be a Hilbert space. The main innovation of our approach is that we develop
a constructive method of analysis to establish stability of the finite rank operator by solving a
finite number of decoupled linear PDEs in space-time with rigorous error control. This allows
to establish linear stability of the original linearized operator L. In comparison, a typical
topological argument may only allow one to establish stability of the leading order operator L0
at the expenses of creating finitely many unstable directions induced by the finite rank operator
K. Such method is ideal if we expect to have unstable blowup. However, if we expect to have
stable blowup as in the Hou-Luo blowup scenario, proving blowup with finitely many unstable
directions for the Hou-Luo blowup scenario using a topological argument is not satisfactory.
1.4. Review of literature. There has been a lot of effort in studying 3D Euler singularities
using various simplified models. Several 1D models, including the Constantin-Lax-Majda (CLM)
model [19], the De Gregorio (DG) model [24,25], the gCLM model [68] and the Hou-Li model [45],
have been introduced to study the effect of advection and vortex stretching in the 3D Euler
equations. Singularity formation from smooth initial data has been established for the CLM
model in [19], for the DG model in [14], and for the gCLM model with various parameters
in [5, 8, 10, 14, 29, 31]. In the viscous case, singularity formation of the gCLM model with some
parameters has been established in [8,70]. In [15], the authors proved the blowup of the Hou-Luo
model proposed in [55]. In [13], Chen-Hou-Huang proved the asymptotically self-similar blowup
of the Hou-Luo model by extending the method of analysis established for the finite time blowup
of the De Gregorio model by the same authors in [14].
In [16,38,39,50], the authors proposed several simplified models to study the Hou-Luo blowup
scenario [54, 55] and established finite time blowup of these models. In these works, the velocity
is determined by a simplified Biot-Savart law in a form similar to the key lemma in the seminal
work of Kiselev-Sverak [48]. In [30, 32], Elgindi and Jeong proved finite time blowup for the 2D
Boussinesq and 3D axisymmetric Euler equations in a domain with a corner using C̊ 0,α data.
There has been some interesting recent results on the potential instability of the Euler blowup
solutions, see [51, 72]. In a recent paper [11], we showed that the blowup solutions of the 2D
Boussinesq and 3D Euler equations with C 1α velocity considered in [12, 27] are also unstable
using the notion of stability introduced in [51, 72]. These two seemingly contradictory results
reflect the difference of the two approaches in studying the stability of 3D Euler blowup solutions.
6 JIAJIE CHEN AND THOMAS Y. HOU

The linear stability analysis in [51,72] is performed by directly linearizing the 3D Euler equations
around a blowup solution in the original variables. It does not take into account the changes
in the blowup time and the blowup exponent due to the change of the initial condition. Such
information has been used in establishing the nonlinear stability of the blowup profile in [12,27].
In [42,43], Hou and Huang reported a potential two-scale traveling singularity for the 3D Euler
equations in the interior domain. Inspired by the work reported in [42,43], Hou discovered a new
class of potential Euler singularity at the origin whose scaling properties are compatible with
those of the Navier-Stokes equations [41]. Moreover, the Navier-Stokes equations with the same
smooth initial data develop potentially singular solutions at the origin with maximum vorticity
increasing by a factor of 107 [40]. Various blowup criteria have been applied to provide strong
support for this potentially singular behavior of the Navier-Stokes equations.
The rest of the paper is organized as follows. Sections 2 – 5 will be devoted to the blowup
analysis for the 2D Boussinesq equations and Section 6 will be devoted to the blowup analysis
for 3D Euler equations. In Section 2, we provide detailed discussions and some key ingredients in
establishing linear stability of an approximate profile using various simplified models. In Section
3, we show how to obtain sharp Hölder estimates using optimal transport. Section 4 is devoted
to energy estimates and Section 5 is devoted to the construction of an approximate self-similar
profile using the dynamic rescaling formulation. In Section 7, we discuss the construction of the
approximate space-time solution to the linearized operator L. Finally we show how to estimate
the L∞ and Hölder norms of the velocity in the regular case in Section 8. Some technical
estimates and derivations are deferred to the Appendix.

2. Linear stability analysis and the main ideas

In this section, we will outline the main ingredients in our stability analysis. We will mainly
focus on the 2D Boussinesq equations since the analysis for the 3D Euler is very similar with
minor modifications, see Section 6. As in [12–14], we will use the dynamic rescaling formulation
for the 2D Boussinesq equations in an essential way. The most essential part of our analysis lies
in the linear stability. We need to use a number of techniques to extract the damping effect from
the linearized operator around the approximate steady state of the dynamic rescaling equations
and obtain sharp estimates of various nonlocal terms. Since the damping coefficients we obtain
are relatively small, we need to construct an approximate steady state with a very small residual
error. This is extremely challenging since the solution is supported on the upper half plane with
a slowly decaying tail in the far field. We use analytic estimates and numerical analysis with
rigorous error control to verify that the residual error is small in the energy norm. See more
detailed discussions in Section 5 and the supplementary material.
Passing from linear stability to nonlinear stability is relatively easier since the perturbation is
quite small due to the small residual error. Yet we need to verify various inequalities involving
the approximate steady state using the interval arithmetic [36,67,69] and numerical analysis with
computer assistance. The most essential part of the linear stability analysis can be established
based on the grid point values of the approximate steady state. The reader who is not interested
in the rigorous verification can skip the verification process in the supplementary material.

2.1. Notations and operators. The upper bar notation is reserved for the approximate steady
state, e.g. ω̄, θ̄.
We introduce the notations for the nonlinear terms
(2.1) N1 = −u·∇ω+cω ω, N2 = −u·∇η−ux η−vx ξ+2cω η, N3 = −u·∇ξ−uy η−vy ξ+2cω ξ.
Without specification, Ni depends on (ω, η, ξ). Given the approximate steady state ω̄, θ̄, c̄l , c̄ω ,
we denote by F i and F̄ω , F̄θ the residual error
F̄ω = −(c̄l x + ū) · ∇ω̄ + θ̄x + c̄ω ω̄, F̄θ = −(c̄l x + ū) · ∇θ̄ + c̄θ θ̄,
(2.2)
F 1 , F̄ω , F 2 , ∂x F̄θ , F 3 , ∂y F̄θ .
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 7

Notations. Denote by Cxα , Cyα the partial Hölder seminorms

|ω(x) − ω(z)| |ω(x) − ω(z)|
(2.3) [ω]Cxα (D) , sup , [ω]Cyα (D) , sup .
x,z∈D,x2 =z2 |x1 − z1 |α x,z∈D,x1 =z1 |x2 − z2 |α
Given a weight g(h) : R2 → R+ that is −α-homogeneous , i.e. g(λh1 , λh2 ) = λ−α g(h), e.g.,
g(h) = |h|−α , we define the weighted Hölder seminorm
(2.4) ||ω||Cgα (D) = sup |(ω(x) − ω(z))g(x − z)|.
x,z∈D

We will mostly use D = R2++ .

In this case, we drop D to simplify the notations.
We define hf, gi the inner product in R2++
Z
(2.5) hf, gi = f (x)g(x)dx.
R2++

2.2. Dynamic rescaling formulation. Following [12–14], we consider the dynamic rescaling
formulation of the 2D Boussinesq equations. Let ω(x, t), θ(x, t), u(x, t) be the solutions of (1.4)-
(1.6). Then it is easy to show that
ω̃(x, τ ) = Cω (τ )ω(Cl (τ )x, t(τ )), θ̃(x, τ ) = Cθ (τ )θ(Cl (τ )x, t(τ )),
(2.6) −1
ũ(x, τ ) = Cω (τ )Cl (τ ) u(Cl (τ )x, t(τ )),
are the solutions to the dynamic rescaling equations
(2.7) ω̃τ (x, τ ) + (cl (τ )x + ũ) · ∇ω̃ = cω (τ )ω̃ + θ̃x , θ̃τ (x, τ ) + (cl (τ )x + ũ) · ∇θ̃ = cθ θ̃,
where ũ = (ũ, ṽ)T = ∇⊥ (−∆)−1 ω̃, x = (x, y)T ,
Z τ  Z τ  Z τ 
(2.8) Cω (τ ) = exp cω (s)dτ , Cl (τ ) = exp −cl (s)ds , Cθ = exp cθ (s)dτ ,
0 0 0
Rτ
t(τ ) = 0
Cω (τ )dτ and the rescaling parameters cl (τ ), cθ (τ ), cω (τ ) satisfy [12]
(2.9) cθ (τ ) = cl (τ ) + 2cω (τ ).
We have the freedom to choose the time-dependent scaling parameters cl (τ ) and cω (τ ) accord-
ing to some normalization conditions. These two free scaling parameters are related to the fact
that Boussinesq equations have scaling-invariant property with two parameters. The 3D Euler
equations enjoy the same property. See [12]. After we determine the normalization conditions
for cl (τ ) and cω (τ ), the dynamic rescaling equation is completely determined and the solution
of the dynamic rescaling equation is equivalent to that of the original equation using the scaling
relationship described in (2.6)-(2.8), as long as cl (τ ) and cω (τ ) remain finite.
We remark that the dynamic rescaling formulation was introduced in [52, 60] to study the
self-similar blowup of the nonlinear Schrödinger equations. This formulation is also called the
modulation technique in the literature and has been developed by Merle, Raphael, Martel, Zaag
and others. It has been a very effective tool to analyze the formation of singularities for many
problems like the nonlinear Schrödinger equation [47,61], compressible Euler equations [2,3], the
nonlinear wave equation [66], the nonlinear heat equation [65], the generalized KdV equation [57],
and other dispersive problems. Recently, this method has been applied to study singularity
formation in incompressible fluids [12, 27] and related models [8–10, 14].
To simplify our presentation, we still use t to denote the rescaled time in (2.7) and simplify
ω̃, θ̃ as ω, θ
ωt + (cl x + u) · ∇ω = θx + cω ω,
(2.10)
θt + (cl x + u) · ∇θ = cθ θ.
Following [13], we impose the following normalization conditions on cω , cl
θxx (0) 1
(2.11) cl = 2 , cω = cl + ux (0), cθ = cl + 2cω .
ωx (0) 2
8 JIAJIE CHEN AND THOMAS Y. HOU

For smooth data, these two normalization conditions play the role of enforcing
(2.12) θxx (t, 0) = θxx (0, 0), ωx (t, 0) = ωx (0, 0)
for all time. In fact, we can derive the ODEs of θxx (t, 0) and ωx (t, 0)
d d
ωx (t, 0) = (cω − cl − ux (0))ωx (t, 0) + θxx (t, 0), θxx (t, 0) = (cθ − 2(cl + ux (0)))θxx (t, 0),
dt dt
where we use v|y=0 = 0, vx (t, 0) = 0. Under the conditions (2.11), the right hand sides vanish.

2.3. Main Result. In this section, we state our main result for the 2D Boussinesq equations.
We first introduce some notations and define our energy. Let ψi , ϕi , ψi,g , gi be the singular
weights defined in (C.1), (C.2), (C.3), and µ1 , µ2 , τ1 , τ2 , µ4 be the parameters chosen in (C.4).
We define the energy E on three variables f1 , f2 , f3 as follows
P1 = max ||fi ϕi ||∞ , P2 = τ1−1 max(||f1 ψ1 ||C 1/2 , µ1 ||f2 ψ2 ||C 1/2 , µ2 ||f3 ψ3 ||C 1/2 )
1≤i≤3 g1 g2 g3

(2.13) P3 = τ2 max(µ2 ||f1 ϕg1 ||∞ , ||f2 ϕg,2 ||∞ , ||f3 ϕg,3 ||∞ ),
1 1 1
P4 = max( |cω (ω)|, |f2,xy (0)|, |f1,xy (0)|), E = max(P1 , P2 , P3 , P4 ).
65 10 5
where ux (f )(0) = − π4 R++ y|y|
1 y2
R
4 f (y)dy.
2

Theorem 3. Let (θ̄, ω̄, c̄l , c̄ω ) be the approximate self-similar profile constructed in Section 5
and E∗ = 5 · 10−6 . For even initial data θ0 and odd ω0 of (2.10) with a small perturbation to
(θ̄, ω̄) with E(ω0 − ω̄, θ0,x − θ̄x , θ̄0,y − θ̄y ) < E∗ , we have
(2.14) ||ω − ω̄||L∞ , ||θx − θ̄x ||L∞ , ||θy − θ̄y ||∞ < 103 E∗ , |ux (t, 0) − ūx (0)|, |c̄ω − cω | < 100E∗
for all time. In particular, we can choose smooth initial data ω0 , θ0 ∈ Cc∞ in this class with
finite energy ||u0 ||L2 < +∞ such that the solution to the physical equations (1.4)-(1.6) with these
initial data blows up in finite time T .
Remark 2.1. In our analysis, we decompose the nonlinear operator that governs the evolution
of the perturbation into two parts. Correspondingly, we decompose the perturbation W f =
(ω − ω̄, θx − θ̄x , θy − θ̄y ) into two components, W = W1 + W2 , see Section 2.10.3 for the precise
f f f
decomposition. At the linear level, the evolution of W f1 is governed by a leading order operator
L0 that enjoys sharp energy estimates after we subtract a finite rank operator. The perturbation
W
f2 mainly captures the contributions from the finite rank perturbation operator. One advantage
of using such decomposition is that W f1 is only weakly coupled to W
f2 through the nonlinear terms
and the residual errors, which are small. We can estimate the contributions of W f2 to the linear
stability by solving a finite number of linear PDEs in space and time and estimate the space-time
residual errors via energy estimates. The stability of the leading order operator L0 enables us to
control the residual errors arising from solving the linear PDEs in space and time rigorously. A
crucial step in proving our main theorem is to establish the estimate E4 (W f1 ) < E∗ with energy
E4 defined in (4.59). Note that W1 = W at t = 0 and the energy E4 agrees with E at t = 0.
f f
See Sections 2.10, 4, and (4.8.5) for more discussions.
We will follow the framework in [12–14] to establish finite time blowup. It consists of the
following steps: (a) construct approximate steady state to (2.10) with small residual error in
suitable functional spaces; (b) perform linear and nonlinear stability analysis around the ap-
proximate steady state; (c) choose small perturbation in the energy space to obtain smooth
initial data with finite energy and obtain finite time blowup using a rescaling argument and the
stability results.
In the remaining of this section, we will outline some main ingredients in our blowup analysis
by using a number of simplified models to illustrate and motivate the main ideas behind our
method of analysis.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 9

2.4. Basic properties of the approximate steady state. Following the ideas in [13,14], we
construct the approximate steady state (ω̄, θ̄, c̄ω , c̄l ) of the dynamic rescaling equations (2.10),
(2.11) by solving them numerically for a long enough time. In Figure 1, we plot the approximate
steady state ω̄, θ̄x . We plot the variable θ̄x rather than θ̄ since θ̄ grows in the far-field. Given
the approximate steady state, we construct the numerical stream function φ̄N by solving the
Poisson equations. Then we can derive the residual (2.2) up to the error in solving the Poisson
equations. In Figure 2, we plot the piecewise rigorous bound of the weighted L∞ (ϕ1 ) norm
of F̄1 and the L∞ (ϕ2 ) norm of F̄2 . We remark that ϕ1 , ϕ2 are very singular near x = 0 with
leading order |x|−2.9 , 0.385|x|−3 . This is why the weighted L∞ (ϕ1 ) norm of F̄1 and the L∞ (ϕ2 )
norm of F̄2 are relatively large near the origin.
We observe that unweighted errors of F̄1 , F̄2 are very small near the origin, less than 2 · 10−12
since we use a uniform fine grid near the origin. In Figure 3, we plot the piecewise rigorous
bound of the unweighted L∞ norm of F̄1 and the L∞ norm of F̄2 in a local domain near the
origin. Note that the unweighted L∞ norm of F̄1 and the L∞ norm of F̄2 increase as |x|
increases. On the other hand, since the singular weights have a very mild growth rate in the
far field of order O(|x|1/16 ), the contributions from these weighted errors to the energy E3 scale
like 0.165 ∗ |x|1/16 kF̄i kL∞ (i = 1, 2). Thus the weighted L∞ norm of F̄1 and the L∞ norm of
F̄2 only amplify the unweighted norms very mildly in the far field. Since the damping effect
is stronger in the far field, we are able to close the energy estimates with the residual errors
that we obtain. We defer the details of numerical computation to Section 5. Here, we list some
important properties of the approximate steady state.

Figure 1. Approximate steady state in the near-field. Left figure: profile ω̄;
right figure: θ̄x .

Exponents. The exponents and the velocity near the origin satisfy
(2.15) c̄l ≈ 3.00649898, c̄ω ≈ −1.02942516, ūx (0) ≈ −2.532674, v̄x (0) ≈ 0.
We remark that the ratio c̄l /c̄ω ≈ −2.9205600 is very close to the one reported by Hou-Luo
[54, 55].
Regularity and representation. The variables ω̄, ψ̄ are odd in x and θ̄ is even in x. Denote
by φ the stream function. One should not confuse the stream function φ with singular weights
ϕ1 , ϕ2 , etc. The approximate steady state (ω̄, θ̄, φ̄) is represented by piecewise fifth order
polynomials ω̄2 , θ̄2 , φ̄2 supported in [0, D1 ]2 with D1 ≈ 1015 , and semi-analytic parts ω̄1 , θ̄1 , φ̄1
that capture the far-field behavior of the solutions
ω̄ = ω̄1 + ω̄2 , θ̄ = θ̄1 + θ̄2 , φ̄ = φ̄1 + φ̄2 .
4,1
See (5.2). In particular, we have ω̄, θ̄, φ̄ ∈ C . The solution enjoys the decay rate
α 1+2α
ω̄ ∼ r , θ̄ ∼ r , α ≈ c̄ω /c̄l .
10 JIAJIE CHEN AND THOMAS Y. HOU

Figure 2. Weighted residual errors of the approximate steady state. Left

figure: piecewise rigorous L∞ (ϕ1 ) bound of F1 in the ω equation. Right figure:
piecewise rigorous L∞ (ϕ2 ) bound of F2 in the θx equation
.

Figure 3. Unweighted residual errors of the approximate steady state. Left

figure: piecewise rigorous L∞ bound of F1 in the ω equation. Right figure:
piecewise rigorous L∞ bound of F2 in the θx equation.

Anisotropic. The solutions θ̄ and ω̄ are anisotropic in the sense that the y-derivative of the
profile is much smaller than the x-derivative, especially in the near field:

(2.16) |θ̄y | < c3 |θ̄x |, c3 ≈ 0.16, |ω̄y | < c4 |ω̄x |, c4 ≈ 0.23,

for (x, y) ∈ [0, 1]2 . Note that the anisotropic property of the solutions has been observed for the
C 1,α singular solution [12].
The advection. The advection in (2.10) satisfies the following important inequalities

c̄l x + ū(x, y) ≥ c1 x, c1 ≈ 0.47, c̄l y + v̄(x, y) ≥ c2 y, c2 ≈ 3,

for all x, y ∈ R2++ . For x, y ∈ R2++ near the origin, we have

c̄l x + ū(x, y) ≈ 0.47x, cl y + v̄(x, y) ≈ 5.54y.

 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 11

2.5. Linearized equations. Linearizing (2.10) around (ω̄, θ̄, ū, c̄l , c̄ω ), we yield
ωt = −(c̄l x + ū) · ∇ω + θx + c̄ω ω − u · ∇ω̄ + cω ω̄ + F̄ω + N (ω),
(2.17)
θt = −(c̄l x + ū) · ∇θ + c̄θ θ + cθ θ̄ − u · ∇θ̄ + F̄θ + N (θ), u = ∇⊥ (−∆)−1 ω,
where F̄ω , F̄θ are the residual errors (2.2), and N (ω), N (θ) are the nonlinear terms
(2.18) N (ω) = −u · ∇ω + cω ω = N1 , N (θ) = −u · ∇θ + cθ θ,
where we have used the notation N1 (2.1) and the following normalization conditions for the
perturbing cl , cω from (2.11)
(2.19) cω = ux (0), cl ≡ 0, cθ = cl + 2cω .
Since ω, ∇θ have similar regularity, we study the system of (ω, θx , θy ) and denote
(2.20) η = θx , ξ = θy .
Taking derivatives on the θ equation in (2.17) and using the notations (2.1), (2.2), we obtain
∂t η = −(c̄l x + ū) · ∇η + (2c̄ω − ūx )η − v̄x ξ − ux · ∇θ̄ − u · ∇θ̄x + 2cω θ̄x + N2 + F 2 ,
(2.21)
∂t ξ = −(c̄l x + ū) · ∇ξ + (2c̄ω + ūx )ξ − ūy η − uy · ∇θ̄ − u · ∇θ̄y + 2cω θ̄y + N3 + F 3 ,
where we have used cθ = cl + 2cω . Due to the normalzation conditions (2.12) and the odd
symmetries of θx , ω we have the following vanishing conditions near the origin
(2.22) ω = O(|x|2 ), θx = O(|x|2 ), θy = O(|x|2 ).
Analyzing the linear stability of the above system is extremelly challenging since it contains
several nonlocal terms, which are not small. We remark that numerical evidence of linear
stability of the above system has been reported by Liu [53].
2.6. Main terms of the system. Firstly, we identify the main terms in the linearized equations
(2.17),(2.21). As we will see later in the derivation of the damping terms, e.g., Section 2.7.2, we
have larger damping factors away from the boundary. Moreover, the solution (ω̄, ∇θ̄) decays in
the far-field. Thus the most difficult region for the analysis is a sector ΣS near the boundary,
e.g. (x, y) : |(x, y)| ≤ 2, y/x ≤ 0.1.
2.6.1. Ansiotropy of the solution. Since the solutions are anisotropic (2.16) in the near field, the
coefficients involving the y-derivative of the solution, e.g., ω̄y , θ̄y , θ̄xy , are relatively small.
For the C 1,α singular solution [12], the perturbation of θ is also anisotropic in the sense that
θy enjoys much better estimate than that of θx since the damping term in the equation of θy is
much larger. This also holds true in (2.17), (2.21) due to the flow structure: compression in the
x-direction and outward flow in the y-direction. Indeed, since ūx (0) ≈ −2.5 near the origin and
c̄ω ≈ −1 (2.15) , we have
(2c̄ω − ūx )η ≈ 0.5η, (2c̄ω + ūx )ξ ≈ −5.5ξ.
Therefore, these terms contribute to a growing term in the equation of η and a large damping
term in the ξ equation. As a result, ξ enjoys much better stability estimate than η.
2.6.2. Weak coupling. Note that v̄x ≈ 0 near 0 (2.15) and v̄(x, 0) = 0 due to the boundary
condition, v̄x is quite small in the near field and near the boundary. We will show that ξ enjoys
better estimates than η using the observation in Section 2.6.1. Therefore, the term v̄x ξ in the
η equation (2.21) is small in ΣS . As a result, ξ is weakly coupled to the equation of η in the
most difficult region of the analysis. This coupling structure between η and ξ is consistent with
that of the C 1,α singular solution in [12], where v̄x ξ is treated as a lower order term in the η
equation.
Using the above analysis and dropping the smaller terms and the ξ equation, we identify the
main terms in the linear part of the system (2.17)
ωt = −(c̄l x + ū) · ∇ω + η + c̄ω ω − uω̄x + cω ω̄ + Rω ,
(2.23)
∂t η = −(c̄l x + ū) · ∇η + (2c̄ω − ūx )η − ux θ̄x − uθ̄xx + 2cω θ̄x + Rη ,
12 JIAJIE CHEN AND THOMAS Y. HOU

where Rω , Rη denote the remaining terms in the equations. The above system is very similar
to that in the Hou-Luo model [13]. Moreover, near the boundary, the coefficients ω̄, θ̄ are
very similar. The contribution of the lower order linear terms to the energy estimate is small
compared to the main term, e.g. about or less than 1/10 of that in the near field.

2.7. The local parts and functional spaces. To understand the linear stability, we first
focus on the local terms in the main system (2.23)
∂t ω + (c̄l x + ū) · ∇ω = c̄ω ω + η + Rω,loc ,
(2.24)
∂t η + (c̄l x + ū) · ∇η = (2̄cω − ūx )η + Rη,loc ,
and design the functional spaces for stability analysis, where Rω,loc , Rη,loc denote other terms
in (2.17), (2.21), (2.23).
Following [12–14], we will perform weighted energy estimate in some suitable space X and
derive the damping terms in the weighted energy estimate from the above local terms, especially
the advection term (c̄l x + ū) · ∇f . See Section 2 in [14] for an example. The principle of choosing
the appropriate energy space X is the following [13]. Firstly, the local part of the linearized
equations should be stable in space X. Secondly, we can estimate the nonlocal terms in X
effectively.
In [12–14], the linear stability analysis is based on some weighted L2 spaces with singular
weight near the origin. In these works, one of the major advantages of the weighted L2 space over
other weighted Sobolev spaces is that one can explore nonlocal quadratic cancellation among
the local and nonlocal terms to obtain sharp estimates of the nonlocal terms. In [13, 14], an
additional advantage is that one can use the L2 isometry of the Hilbert transform to control the
nonlocal terms effectively. We note that the Riesz transform ∇u = ∇∇⊥ (−∆)−1 ω in the 2D
Boussinesq (2.17) also enjoys several sharp L2 estimate, e.g. ||∂xy (−∆)−1 ω||L2 ≤ ||ω||L2 . Yet,
we will use a toy example to show that a weighted L2 space may not be suitable for the stability
analysis due to the y-advection (c̄l y + v̄)∂y f .

2.7.1. A toy model for the local term. To understand the behavior of the local terms in (2.24),
we approximate the system (2.24) near the origin by the following model in R++
2

ωt + (a1 x∂x + a2 y∂y )ω = −ω + η,

(2.25)
ηt + (a1 x∂x + a2 y∂y )η = a3 η, a1 = 0.5, a2 = 5.5, a3 = 0.5,
with ω, η being odd in x, where we have used (2.15) to obtain approximations
c̄l x + ū ≈ (c̄l + ūx (0))x ≈ 0.5x, c̄l y + v̄ ≈ 5.5y, 2c̄ω − ūx (0) ≈ 0.5.
Weighted L2 spaces. The first attempt for stability analysis is to perform weighted L2 estimate
with singular weight ϕ = x−α y −β and some α, β > 0 following [12–14]. Due to the symmetry
in x, we focus on D = R++ 2 . For the η equation, using integration by parts and a direct
computation yield
Z Z Z 
1 d 1 1 
η2 ϕ = η(a3 η − a1 x∂x η − a2 y∂y η)ϕ = a3 ϕ + a1 (xϕ)x + a2 (yϕ)y η 2 .
2 dt D D D 2 2
Since (xϕ)x = (1 − α)ϕ, (yϕ)y = (1 − β)ϕ, we get
Z Z
1 d 1 1
η 2 ϕ = a(α, β) η 2 ϕ, a = a3 + a1 (1 − α) + a2 (1 − β).
2 dt D D 2 2
Notice that ω, η do not vanish near the boundary y = 0, we need to choose β < 1 so that the
energy is well-defined. This implies that the y-advection a2 y∂y f contributes to a growing factor
in the energy estimate. Moreover, we cannot choose β sufficiently close to 1 in our stability
analysis of (2.17),(2.21) since we will need to estimate the weighted norm of the nonlocal term,
e.g. ux θ̄x in (2.21). If β is close to 1, since ux does not vanishes near y = 0, we expect a very
poor estimate: ||ux ϕ1/2 ||2 ≤ C(1 − β)−1/2 ||ωϕ1/2 ||2 .
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 13

Since a2 = 11 1
2 is much larger than a1 = 2 , to obtain a damping factor a(α, β) < 0, we need
to pick α sufficiently large. For example, if β = 0, to obtain a(α, β) ≤ 0, we need
1 1 11
+ (1 − α) + ≤ 0, α ≥ 14.
2 4 4
This means that we need to choose a very singular weight near the origin. Yet, the perturbation
ω, η in (2.17), (2.21) does not vanishes to such a high order near the origin (2.22). Another
type of weight is ϕ = x−α y −β (x2 + y 2 )−γ/2 . Similar computation yields that for β = 0, to get
a damping factor, α + γ need to be sufficiently large.
Thus, the y-advection in (2.17), (2.21) can contribute to a large growing factor to the energy
estimate. This is a new difficulty that is absent in [12–14].
A potential L2 based approach to derive the damping term is to perform sufficiently high
order H k estimate. Taking a partial derivative ∂xi ∂yj plays a role similar to a singular weight
x−i y −j . Yet, this approach can lead to many more terms in the system (2.17), (2.21), e.g.
∂xi (ux θ̄x ), which can be difficult to control. Moreover, constructing an approximate steady state
with a small residual error in H k with k ≥ 14 is extremelly challenging.
Weighted L∞ space. Instead of weighted L2 spaces, one can try weighted Lp spaces. In
particular, we can use weighted L∞ estimates to take advantage of the transport structure.
Suppose that ϕ = r−γ , r = (x2 + y 2 )1/2 . Multiplying the η equation with ϕ and a direct
calculation yield
(2.26) ∂t (ηϕ) + (a1 x∂x + a2 y∂y )(ηϕ) = (a3 ϕ + a1 x∂x ϕ + a2 y∂y ϕ)η , a(γ)ηϕ,
Since x∂x ϕ = −γx2 r−γ−2 , y∂y ϕ = −γy 2 r−γ−2 and a2 ≥ a1 , we get
a1 x∂x ϕ + a2 y∂y a1 x2 + a2 y 2
(2.27) a(γ) = a3 + = a3 − γ ≤ a3 − a1 γ.
ϕ x2 + y 2
Since a1 = 0.5, a2 = 5.5, a3 = 0.5, to obtain a damping factor a(γ) ≤ 0, we can choose
γ ≥ 1. Notice that for the system (2.17), (2.21), ω, η vanish at least quadratically near 0 (2.22).
Therefore, we can choose γ ≥ 2 to derive the damping terms in the η equation.
For the system in (2.25), performing L∞ estimate with weight ϕ = r−γ with γ > 1 on both
equations, we get
d d
(2.28) ||ωϕ||∞ ≤ (−1 − a1 γ)||ωϕ||∞ + ||ηϕ||∞ , ||ηϕ||∞ ≤ (a3 − a1 γ)||ηϕ||∞ .
dt dt
It is easy to further obtain that max(||ωϕ||∞ , ||ηϕ||∞ ) decays exponentially fast.
From (2.27), since a2 is much larger than a1 , as the ratio λ = y/x increases, we get a much
larger damping factor
a1 + a2 λ2 1 1 1 11
a(γ, λ) = a3 − γ , a(γ, 0) = a3 − a1 γ = − γ, a(γ, ∞) = a3 − γa2 = − γ .
1 + λ2 2 2 2 2

Weighted Hölder estimate. There is one drawback of applying the weighted L∞ estimate
to (2.17), (2.21). Since (2.21) contains ∇u = ∇∇⊥ (−∆)−1 ω and the Riesz transform is not
bounded from L∞ → L∞ , we cannot close the estimates using the weighted L∞ space. To
overcome this difficulty, we perform weighted Hölder C α estimates. We will use the following
simple identity repeatedly.

Lemma 2.2. Suppose that f satisfies

(2.29) ∂t f + b(x) · ∇f = c(x)f (x) + R, x ∈ R2+ .
Given some weights g(x1 , x2 ) even in x1 , x2 and ϕ, we denote the operator δ and function F
b · ∇ϕ
δ(p)(x, z) = p(x) − p(z), F (x, z, t) = δ(f ϕ)(x, z)g(x − z), d(x) = c(x) + , x, z ∈ R2+ .
ϕ
14 JIAJIE CHEN AND THOMAS Y. HOU

Then we have
(b(x) − b(z)) · (∇g)(x − z)
∂t F + (b(x) · ∇x + b(z) · ∇z )F = (d(x) + )F
(2.30) g(x − z)
+(d(x) − d(z))g(x − z)(f ϕ)(z) + δ(Rϕ)(x, z)g(x − z) .
The proof is based on a direct calculation and is deferred to Appendix A.1. We treat the first
term on the right hand side of (2.30) as a damping term. The term b·∇ϕϕ in d(x) is the damping
term from the singular weight ϕ(x). In (2.31) below, we show that the term
(b(x) − b(z)) · (∇g)(x − z)
F
g(x − z)
has a negative coefficient and is also a damping term. It comes from the Hölder function g.
Next, we apply the computation in Lemma 2.2 to the η equation in (2.25). Denote
ϕ2 = |x|−γ2 , g(h) = |h|−α , h ∈ R2 , b(x) = (a1 x1 , a2 x2 ), F = ((ηϕ2 )(x) − (ηϕ2 )(z))g(x − z)
for x = (x1 , x2 ), z = (z1 , z2 ) ∈ R2+ . Using the identity (2.27) and definitions of g, b, we get
b(x) · ∇ϕ2 h2i h2
d(x) = a3 + = a(γ2 ), hi ∂i g = −α 2+α = −α i2 g,
ϕ2 |h| |h|
b(x) − b(z) = (a1 (x1 − z1 ), a2 (x2 − z2 )).
(b(x)−b(z))·(∇g)(x−z)
Thus, we obtain that g(x−z) F is a damping term
(b(x) − b(z)) · (∇g)(x − z) a1 (x1 − z1 )2 + a2 (x2 − z2 )2
(2.31) F = −α F , e(α, x, z)F,
g(x − z) |x − z|2
where
a1 (x1 − z1 )2 + a2 (x2 − z2 )2
e(α, x, z) = −α .
|x − z|2
Using Lemma 2.2 with R = 0, we yield
∂t F + (b(x)∇x + b(z)∇z )F = (a(γ2 )(x) + e(α, x, z))F + (a(γ2 )(x) − a(γ2 )(z))g(x − z)(ηϕ2 )(z).
Clearly, we have e(α, x, z) ≤ −αa1 . For the last term in the above equation, from definition
(2.27), d(x) = a(γ2 )(x) is not in C α . Yet, we can estimate I4 = (a(γ2 )(x)−a(γ2 )(z))g(x−z)|z|α .
In fact, since a(γ2 )(x)|x|α , |x|α ∈ C α , we can rewrite I4 as follows
I4 = (a(γ2 )(x)|x|α − a(γ2 )(z)|z|α )g(|x − z|) + a(γ2 )(x)(|z|α − |x|α )g(|x − z|),
which is bounded. By combining the L∞ (|x|−γ ) estimate of η with γ = γ2 + α and ϕ2 |x|−α =
|x|−γ2 −α = |x|−γ , we can control the last term
|(a(γ2 )(x) − a(γ2 )(z))g(x − z)ηϕ2 )(z)| ≤ ||η|z|−γ ||∞ · ||(a(γ2 )(x) − a(γ2 )(z))g(x − z)|z|α ||∞ .
Recall a(γ2 ) ≤ a3 − a1 γ2 , d(α) ≤ −a1 α. From the above estimate, we obtain
d
||F ||L∞ (x,z) ≤ (a3 −a1 (γ2 +α))||F ||L∞ (x,z) +||η|z|−γ ||∞ ·||(a(γ2 )(x)−a(γ2 )(z))g(x−z)|z|α ||∞ .
dt
The above estimate provides a weighted C α estimate for η. Since a1 = a3 = 1/2, by choosing
γ2 + α = γ > 1 and combining the above esimate and (2.28), we can establish the stability
estimate for the model problem in a combination of weighted L∞ and C α spaces.
2.7.2. Anisotropy of the flow and the most difficult scenario. The system (2.17), (2.21) is much
more complicated than the model problem (2.25) since it involves variables oefficients and several
nonlocal terms. Similar to [12–14], we will design the weight as linear combination of different
powers |x|−αi to take into account the behavior in the near field and the far field.
From the above analysis of the model problem, (2.27), and (2.31), we see that the estimate
is anisotropic in x and y. In the near field, from (2.27), if y/x is not small, we get a much
larger damping term. Therefore, the most difficult region for the estimate is y/x = 0, which
corresponds to the boundary.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 15

From (2.31), we also get a much larger damping factor if |x2 − z2 |/|x1 − z1 | is large. This
implies that the Hölder estimate in y direction enjoys much better estimates than those in the
x direction. Therefore, the most difficult part of the estimate lies in the Hölder estimate in the
horizontal estimate. We will exploit the structure of the flow in designing the energy functional
and stability analysis. See Section 4.1.1.
2.7.3. Functional spaces and energy. Motivated by the above analysis, we will design the func-
tional spaces X as a combiniation of weighted L∞ and C α spaces with α = 21 .
Denote f1 = ω, f2 = η, f3 = ξ. In the weighted L∞ estimate, we choose singular weights
ϕi , i = 1, 2, 3 and will estimate fi ϕi in L∞ . In the weighted Hölder estimate, we will choose
singular weights ψ1 (x), ψ2 (x), ψ3 (x) even in x1 and −1/2-homogeneous functions gi (h) equivalent
to |h|−1/2 and estimate ||fi ψi ||Cgα (2.4). This quantity is equivalent to the Hölder seminorm
i
||fi ψi ||C 1/2 .
From the discussion in Section 2.6.1, [fi ]C 1/2 enjoys better estimate than that of [fi ]C 1/2 .
y x
To exploit this property, we choose anisotropic weight gi (h) such that g(0, 1) is stronger than
1
g(1, 0), e.g. g(h) = (|h1 |+0.5|h 2 |)
1/2 . We refer more details to Section 4.

2.7.4. Estimates of the Hölder norm in R2++ . To simplify our energy estimate, using the sym-
metry of fi in the x direction, we only perform Hölder estimates in the first quadrant. This
allows us to control
(fi ψi )(x1 , x2 ) − (fi ψi )(z1 , x2 ) (fi ψi )(x1 , x2 ) − (fi ψi )(x1 , z2 )
Ix = , Iy = ,
|x1 − z1 |1/2 |x2 − z2 |1/2
for x = (x1 , x2 ), (z1 , x2 ), (x1 , z2 ) ∈ R2++ . Due to the symmetry in the x direction, we have
|Iy (x1 , x2 , z2 )| = |Iy (−x1 , x2 , z2 )| and obtain the Hölder estimate in the y direction, i.e. control
of Iy for all pair (x1 , x2 ), (x1 , z2 ) ∈ R2 . The above quantities do not control the Hölder estimate
in the x direction for all pair x, z. It only allows us to control the difference when (x1 , x2 ), (z1 , x2 )
are in the same quadrant since |Ix (x1 , z1 , x2 )| = |Ix (−x1 , −z1 , x2 )|.
However, to control the Hölder norm of ∇u(ω) in R2++ , since ∇u(ω) is nonlocal, we need
to control the Hölder norm of ω in the whole R2+ . To further control Ix (ω) with x1 z1 < 0, we
consider x1 < 0 < z1 . In this case, since ωψ is odd and |x1 − z1 | = |x1 | + |z1 | we have
(2.32)
|(ωψ1 )(−x1 , x2 ) + (ωψ1 )(z1 , x2 )|  2|(ωψ )(−x , x )| 2|(ωψ )(z , x )|  |x |1/2 + |z |1/2
1 1 2 1 1 2 1 1
≤ max ,
|x1 − z1 |1/2 |2x1 |1/2 |2z1 |1/2 (2|x1 | + 2|z1 |)1/2
 2|(ωψ )(−x , x )| 2|(ωψ )(z , x )| 
1 1 2 1 1 2
≤ max 1/2
, 1/2
,
|2x1 | |2z1 |
where we have used the Cauchy-Schwarz inequality in the last inequality. Therefore, it suffies
to further control ||ωψ1 |x1 |−1/2 ||L∞ , which will be done by performing a L∞ estimate.
We remark that we do not need to estimate ||ηψ2 |x1 |−1/2 ||L∞ and ||ξψ3 |x1 |−1/2 ||L∞ .
Remark 2.3. In fact, from the Hölder estimate in R2++ , we can also control Ix (x1 , z1 , x2 ) for
x1 < 0 < z1 and x2 = z2 . Using the fact that |Ix (x1 , 0, x2 )| = (|ωψ1 (x)|)|x1 |−1/2 and the above
inequality, for x1 < 0 < z1 , we get
√
|Ix (x1 , z1 , x2 )| ≤ max 2|Ix (x1 , 0, x2 )|.
x
√
However, we need to pay a factor 2. For this reason, we estimate ||ηψ2 |x1 |−1/2 ||L∞ directly.

Several weighted L∞ estimates. To establish the linear stability, we will first choose weights
ϕi that has faster decay, e.g. ϕi ∼ |x|−β with large β for large |x|. This allows us to obtain
a larger damping factor. For example, from (2.27), a(γ) is larger if γ is larger. To close the
nonlinear estimate, we need to control ||∇u||L∞ , ||ω||L∞ , ||∇θ||L∞ . Since the weight ϕi is weaker
than O(1) in the far-field, we cannot control these L∞ quantities using ||fi ϕi ||L∞ , ||fi ψi ||C 1/2 .
Thus, we further choose weights ϕg,i growing in the far-field and estimate fi ϕg,i . We also need
to estimate ||ωψ1 |x1 |−1/2 ||L∞ , as discussed in Section 2.7.4.
16 JIAJIE CHEN AND THOMAS Y. HOU

Note that similar idea of obtaining weighted estimate in the stability analysis through a few
intermediate steps has been used in our previous work [12].
The functional spaces of the perturbation. From the above discussions, we will estimate
the following quantities in R2++
||fi ϕi ||L∞ , ||fi ψi ||C 1/2 , ||fi ϕg,i ||L∞ , ||ωψ1 |x1 |−1/2 ||L∞ ,
gi

where f1 = ω, f2 = η, f3 = ξ, and || · ||C 1/2 is defined in (2.4). In view of Lemma A.2, we will
g
design our final energy E in the form of max(µi Fi ) with Fi being one of the above quantities
and some weights µi . We will determine the weights later. See more details in Section 4.
2.7.5. Vanishing order of the perturbation. From (2.22), the perturbation ω, η, ξ vanishes quadrat-
ically near x = 0. To obtain larger damping factors, from the model problem (2.25) and (2.27),
we can choose a larger γ. We will decompose the perturbation fi into two parts
fi = fi,1 + fi,2 ,
where fi,1 captures the main part of f1 and vanishes to the order O(|x|3 ) near x = 0, and
fi,2 accounts for the contribution from some finite rank operators. For example, if we choose
ω2 = ωxy (0)xyχ(x, y) for some cutoff function χ ∈ Cc∞ with χ = 1 near x = 0, then ω1 =
ω − ω2 = O(|x|3 ) near 0. In this problem, cubic vanishing order is good enough for our stability
analysis. See more discussions in Section 2.10.4 and (2.75).
We will perform energy estimates on fi,1 and use space-time estimates for fi,2 .
2.8. Estimate the nonlocal terms ∇u. In the stability analysis, we need to estimate the
nonlocal terms u, ∇u in (2.17), (2.21). One disadvantage of performing energy estimates in the
new functional space is that we lose the L2 isometry property for the nonlocal velocity: e.g.
||∇u||2 = ||ω||2 , ||∇ux ||2 = ||ωx ||2 , which can be obtained using the elliptic equation −∆ψ = ω
and integration by parts. Although we have standard C α estimates for the Riesz transform
∇∇⊥ (−∆)−1 ω, the constants usually are not given explicitly, and they are not sharp enough
for our purposes.
Strategies. Our strategies to estimate the nonlocal terms are the following. For the nonlocal
terms that are of order ω, e.g., ∇u = ∇∇⊥ (−∆)−1 , and their localized version (3.4), we develop
sharp functional inequalities to estimate them. Due to the localization of the kernel, we will gain
a small factor in the weighted L∞ estimate. For the more regular nonlocal terms, e.g., u, we
will approximate them by finite rank operators Fu . We further develop a crucial decomposition
of the equations (2.17), (2.21) that enable us to estimate the contribution of these finite rank
operators separately using computer assistance. See Section 2.9.
To obtain sharp C α estimates for ∇u, we have a crucial observation that we can use tech-
niques from optimal transport. This is another reason why we choose a weighted C 1/2 space
in the energy estimates. Note that optimal transport has been applied to establish many sharp
functional inequalities and study functional inequalities in details, e.g., the reverse Brascamp-
Lieb inequality [1], the Sobolev and Gagliardo-Nirenberg inequalities [20], the isoperimetric
inequalities [34]. See also the excellent books [73, 74] for more details.
We focus on ux from ux θ̄x in the main system (2.23). This term is the most difficult nonlocal
term to estimate since other nonlocal terms in (2.24) are related to u = (u, v) and are more
regular. We note that in the leading order system for the C 1,α singular solution [12], ux θ̄x
is also the main nonlocal term. The coefficients of other nonlocal terms involving u, ∇u, e.g.
−uω̄x , −uθ̄xx , vx θ̄y , contain a small factor α, which makes the estimates much easier.
Denote by ux (x, a, b) the localized version of ux
Z
1 s1 s2
(2.33) u x (x, a, b) , − P.V. K1 (x − y)W (y)dy, K1 (s) = ,
π |x1 −y1 |≤a,|x2 −y2 |≤b |s|4
where W is an odd extension of ω in y from R+ 2
2 to R (3.3).
We decompose ux into two parts ux = ux,S + ux,R with ux,S (x) = ux (x, a, b) for some
a, b > 0. The main term ux,S captures the most singular part of ux in the Biot-Savart law, and
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 17

the remaining term ux,R is more regular. Using the odd symmetry property of K(s) in s1 , s2
and some techniques from optimal transport, we establish sharp estimates for the singular term
ux,S in Lemma 3.1 uniformly in a, b. Similarly, we have established a sharp estimate of ux in the
1/2
Cy seminorm and the estimates of uy , vx in the Hölder seminorms. To estimate the regular
part ux,R , we follow the previous strategy.

2.8.1. Weighted estimates. In the stability analysis, we need to estimate the weighted L∞ and
C α norm of the nonlocal terms. We focus on estimating the Hölder norm of ux ψ. We observe
that the commutator
[ux (·, a, b), ψ](ω) , ux (ω)(x, a, b)ψ(x) − ux (ωψ)(x, a, b)
Z
1
=− K1 (x − y)W (y)(ψ(x) − ψ(y))dy
π |x1 −y1 |≤a,|x2 −y2 |≤b
is more regular. Therefore, we have the decomposition
(2.34) ux (ω)(x, a, b)ψ = ux (ωψ)(x, a, b) + [ux (·, a, b), ψ](ω).
For the first term on the right hand side, we can apply the sharp Hölder estimate in Section 3.
Given that ω is in some weighted L∞ space, since K1 (x − y)(ψ(x) − ψ(y) has a singularity of
1
order |x−y| , the second term is log-Lipschitz and is more regular than the first term. Therefore,
we can estimate its C 1/2 seminorm by the weighted L∞ norm of ω. In particular, if a and b are
small, we obtain a small factor of order (max(a, b))1/2 in this estimate.
To understand if we can obtain linear stability using a combination of weighted L∞ and
1/2
C space, we consider the Hölder estimate where |x − z| is very small in Section 2.8.2. To
understand our overall strategy of the energy estimate, we will study a model problem in Section
2.8.3, which also motivates the localization of the velocity.

2.8.2. The singular scenario. We rewrite (2.17), (2.21) and keep the local terms and the nonlocal
terms that are of the same order of ω, ∇θ
ωt = −(c̄l x + ū) · ∇ω + c̄ω ω + η + R1 ,
(2.35)
ηt = −(c̄l x + ū) · ∇η + (2c̄ω − ūx )η − (ux − ûx )θ̄x − (vx − v̂x )θ̄y + R2 .
Here ûx , v̂x approximate ux , vx near 0:
(2.36) ûx = (ux (0) + Cux (x)K00 )χ(x), v̂x = Cvx (x)K00 χ
The above approximation can be obtained by Taylor expansions and satisfies
(2.37) ux − ûx = O(|x|5/2 ), vx − v̂x = O(|x|5/2 ),
for ω = O(|x|5/2 ) near x = 0. The above vanishing order can be justified for the perturbation ω
in our energy class. The functions Cux (x), Cvx (x) and operators K00 are defined in (2.80) and
(2.79), and χ is some compactly supported cutoff function with χ = 1 near x = 0. The above
approximation is a simplification of the approximation of the velocity in Section 2.11. For ω
vanishes O(|x|3 ) near x = 0, the new term ûx θ̄x + v̂x θ̄y in (2.35) can be seen as a finite rank
operator. We will use the ideas to be introduced in Sections 2.9 and 2.10 to decompose the
system and perturb the equations by some finite rank operators.
We assume that the remaining terms R1 , R2 have vanishing order O(|x|3 ) near x = 0. For
initial perturbation with vanishing order ω0 , η0 = O(|x|3 ) near x = 0, using (2.37), we obtain
that these vanishing conditions can be preserved. Note that, in general, the solution to the
original system (2.17), (2.21) only vanishes quadratically (2.22), and does not satisfy cubic
vanishing conditions. We will perform a correction near x = 0 so that ω, η = O(|x|3 ) near x = 0.
See Sections 2.7.5 and 2.10.
In the Hölder estimate with x, z close enough, we will show below that (2.35) can be seen as
the leading order system. We also move the term v̄x ξ in the η equation to the remaining term
due to the weak coupling discussed in Section 2.6.2.
18 JIAJIE CHEN AND THOMAS Y. HOU

Goal of the estimates and heuristic. In the following weighted Hölder estimates, we will
show that if |x − z| is sufficiently small with x2 = z2 , we can treat the nonlocal terms ux −
ûx , vx − v̂x as perturbations with size 2.55[ωψ1 ]C 1/2 , 2.53||(ωψ1 )g1 ||L∞ , respectively, using the
x
sharp estimates in Lemmas 3.1 and 3.4.
This can be understood heuristically. Since x2 = z2 and |x1 − z1 | is small, we can interprete
the following estimates as taking a half derivative D on (2.39). If D applies to a regular term,
which is Lipschitz, we almost get 0. If D acts on the nonlocal terms, e.g. ux − ûx , since ûx is
more regular and [ux ]C 1/2 can be bounded using Lemma 3.1, we treat it as 2.55[ωψ1 ]C 1/2 . If D
x x
applies to the local term, we can simply apply the derivations in Section 2.7.1.
Following [12–14], we design the singular weights by choosing different powers to take care
of the behaviors in the near field and the far field. We use the weights (C.1) for the Hölder
estimate, which satisfies
(2.38) ψ1 ∼ |x|−2 , ψ2 ∼ p1 |x|−5/2
near x = 0 for some parameter p1 .
Next, we perform the weighted C 1/2 estimate. Denote
b̄(x) = (c̄l x + ū, c̄l y + v̄).

Derivations for the local terms. Applying Lemma 2.2, we get

∂t δ(ωψ1 )g1 + (b̄(x) · ∇x + b̄(z) · ∇z )(δ(ωψ1 )g1 ) = ē1 (x, z)δ(ωψ1 )g1
+ δ(ηψ1 )g1 (x − z) + δ(d¯1 )g1 (x − z)(ωψ1 )(z) + δ(R1 ψ1 )g1 (x − z),
(2.39)
∂t δ(ηψ2 )g2 + (b̄(x) · ∇x + b̄(z) · ∇z )(δ(ηψ2 )g2 ) = ē2 (x, z)δ(ηψ2 )g2
− δ((ux − ûx )θ̄x ψ2 + (vx − v̂x )θ̄y ψ2 )g2 + δ(d¯2 )g2 (x − z)(ηψ2 )(z) + δ(R2 ψ2 )g2 (x − z)
where g1 = g1 (x − z), δf = f (x) − f (z) for a function f ,
b̄ · ∇ψ1 (b̄(x) − b̄(z))(∇g1 )(x − z)
d¯1 = c̄ω + , ē1 (x, z) = d¯1 (x) + ,
ψ1 g1 (x − z)
(2.40)
b̄ · ∇ψ2 (b̄(x) − b̄(z))(∇g2 )(x − z)
d¯2 = 2c̄ω − ūx + , ē2 (x, z) = d¯2 + .
ψ2 g2 (x − z)
We consider the difficult scenario discussed in Section 2.7.2, i.e. for x, z with x2 = z2 , and the
most singular scenario where x and z are sufficiently close. We will discuss how to handle the
more regular case where |x − z| is not small later. We also neglect the nonlinear terms Nω , Nθ
and the residual F̄ω , F̄θ .
Estimate the nonlocal terms. Suppose that ωϕ1 , ηϕ2 ∈ L∞ and δ(ωψ1 )g1 , δ(ηψ2 )g2 ∈ L∞ .
We consider |x − z| sufficiently small compared to min(1, |x|) with x2 = z2 . To estimate δ2 ((ux −
ψ2
ûx )θ̄x ψ2 )g2 , we introduce A(x) = θ̄x ψ1
and rewrite the difference as follows

δ((ux − ûx )θ̄x ψ2 )g2 = δ((ux − ûx )ψ1 A(x))g2

= (δ((ux − ûx )ψ1 )A(x) + (ux − ûx )(z)ψ1 (z)δ(A))g2 , I + II.
The second term II can be seen as a regular term since the half derivative acts on the
ψ2
coefficient. Indeed, if x is away from the origin, the coefficient θ̄x ψ1
is Lipschitz. Since |x − z| is
sufficiently small and (ux − ûx )ψ1 is bounded (2.37), (2.38), in this case, we get |II| ≈ 0. For x
close to the origin, using the fact that θ̄x is odd and smooth, (2.37), and (2.38), we obtain that
ψ2 ψ2
(2.41) θ̄x ∼ x|x|−1/2 x near 0, θ̄x ∈ C 1/2 .
ψ1 ψ1
Since |x − z|, |x| are small, we get (ux − ûx )(z)ψ1 (z) = O(|z|1/2 ) near z = 0, which is also small.
Thus, we get
|(ux − ûx )(z)ψ1 (z)δ(A))g2 | . |(ux − ûx )(z)ψ1 (z)|||A||C 1/2 . |z|1/2 (||ωϕ1 ||L∞ + ||ωψ1 ||C 1/2 ) ≈ 0.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 19

Next, we focus on I. For x away from the origin, the weight ψ is nonsingular and |A(x)| . 1.
Recall that the commutator [ux , ψ1 ](ω) is more regular and ûx (2.36) is regular. Applying
the above argument to estimate the regular term, Lemma 3.1 to estimate ux (ωψ1 ), and using
g2 (x1 − z1 , 0) = |x1 − z1 |−1/2 , we get
(2.42) I ≈ δ(ux (ω)ψ1 )g2 A(x) ≈ δ(ux (ωψ1 ))g2 A(x),
which can be bounded by
ψ2
2.55[ωψ1 ]C 1/2 |θ̄x
(x)|.
xψ1
In the above approximation, we used the same argument as that for II to show the the estimate
of some regular term is small. The estimate is more involved for x near the origin due to
the singular weight. In this case, we need to use the scaling symmetry of ux and decompose
ux = ux (x, a, b) + uR
x with size of the singular region (a, b) proportional to x. We can obtain the
following estimate
|δ((ux − ûx )ψ1 )| · |x1 − z1 |−1/2 ≤ C1 (x, z)[ωψ1 ]C 1/2 + C2 (x, z)||ωϕ1 ||L∞
x

with C1 (x, z) close to 2.55 and C2 (x, z) uniformly bounded for small x, z. The most singular part
in δ((ux − ûx )ψ1 ) can be estimated using Lemma 3.1. The regular part can be controlled using
the weighted L∞ norm of ω. These constants C1 , C2 can be estimated effectively by computing
suitable integrals, and we will discuss them in details in Section 8. Now since A(z) . |z|1/2
(2.41), in the case near the origin, we have better estimate
|I| . |x|1/2 (C1 (x, z)[ωψ1 ]C 1/2 + C2 (x, z)||ωϕ1 ||L∞ ) ≈ 0.
x

Note that if we fix x 6= 0 and then consider |z1 − x1 | → 0, it is easy to justify the above
estimate by using the argument around (2.42). Similarly, we have
ψ2 ψ2
δ(vx θ̄y ψ2 )g2 ≈ δ(vx (ωψ1 ))|x1 − z1 |−1/2 θ̄y , m, m ≤ 2.53||δ(ωψ1 )g1 ||L∞ θ̄y (x) ,
ψ1 ψ1
where the constant 2.53 comes from applying Lemma 3.4 by optimizing the parameter τ and
using
1
[ω]C 1/2 ≤ ||ωψ1 ||C 1/2 , [ω]C 1/2 ≤ ||ωψ1 ||C 1/2 ,
x g1 y 1.718 g1

to further simplify the upper bound in Lemma 3.4, where || · ||C 1/2 is defined in (2.4).
g

Estimate the regular and remaining terms. For |x − z| sufficiently small, the more regular
term vanishes in this estimate. For example, for uω̄x in (2.17), in our energy estimate for the
Boussinesq equations, we will approximate u using some finite rank operators û similar to (2.36)
and estimate (u − û)ω̄x , which vanishes cubically near x = 0. See Sections 2.10 and 2.11. We
can control the log-Lipschitz norm or C 4/5 norm of u in some weighted space using ||ωϕ1 ||L∞ .
Thus, for |x − z| sufficiently small, we get
|((u − û)ω̄x ψ1 )(x) − ((u − û)ω̄x ψ1 )(z)|
≤ C(x, z)||ωϕ1 ||L∞ ,
|x − z|1/2
with C(x, z) → 0 if we fix x and take |z − x| → 0. The same idea applies to other regular terms
in (2.39), e.g.
δ(d¯1 )g1 (x − z)(ωψ1 ), δ(d¯2 )g2 (x − z)(ηψ2 ).
For δ(ηψ1 )g1 in the ω equation (2.39), we get
ψ1 ψ1 ψ1
δ(ψ1 θx )g1 = δ( θx ψ2 )g1 ≈ (x)δ(θx ψ2 )|x1 − z1 |−1/2 , m1 , |m1 | ≤ | (x)|[θx ψ2 ]C 1/2 .
ψ2 ψ2 ψ2 x

We remark that, if z1 = x1 , all the above approximations become equality since the difference
δ(f )|x − z|−1/2 becomes 0 for f being locally Lipschitz around x.
20 JIAJIE CHEN AND THOMAS Y. HOU

Summarize the estimates. For x2 = z2 with |x1 − z1 | sufficiently small, the damping terms
ē1 , ē2 in (2.39) can be simplified as
1 b̄1 (x) − b̄1 (z) 1 1
ēi (x) = d¯i (x) − ≈ d¯i (x) − ∂1 b̄1 (x) = d¯i (x) − (c̄l + ūx (x)).
2 x1 − z1 2 2
Therefore, we derive
∂t δ(ωψ1 )g1 + (b̄(x) · ∇x + b̄(z) · ∇z )δ(ωψ1 )g1 = ē1 δ(ωψ1 )g1 + B1 (x, z) + R3 ,
∂t µ1 δ(ηψ2 )g2 + (b̄(x) · ∇x + b̄(z) · ∇z )µ1 δ(ηψ2 )g2 = ē2 µ1 δ(ηψ2 )g2 + µ1 B2 (x, z) + µ1 R4 ,
where Bi denotes the linear terms in (2.39) except for v̄x ξ, and R3 , R4 denote the remaining
terms coming from the nonlinear terms, the residual, and the terms related to ξ. Here, we have
multiplied the η equation with an additional parameter
(2.43) µ1 = 0.668.
The role of µ1 is to change the weights between δ(ωψ1 )g1 , δ(ηψ2 )g2 in the energy estimate so
that the damping term dominates. See Lemma A.2. When x2 = z2 with |x1 − z1 | → 0, from the
above estimate, we have
(2.44)
ψ1 ψ2
|B1 (x, x)| ≤ (x) µ1 [θx ψ2 ]C 1/2 , µ1 |B2 (x, x)| ≤ µ1 (2.55|θ̄x | + 2.53|θ̄y |) (x)||ωψ1 ||C 1/2 .
µ1 ψ2 x ψ1 g

In view of (A.6) and Lemma A.2, if in each equation, the coefficient of the damping term
is larger than that of the bad term, then we can obtain stability. Indeed, these conditions are
satisfied: we have
ψ1 ψ2
(2.45) ēi + Si ≤ −ci , S1 = (x) , S2 = µ1 (2.55|θ̄x | + 2.53|θ̄y |) (x),
µ1 ψ2 ψ1
for some c1 , c2 > 0. In Figure 4, we plot the grid point values of S1 , S2 defined in (2.45), which
are the upper bounds of B1 , µ1 B2 in (2.44), and −ē1 , −ē2 on the boundary y = 0 for small x.
Note that the estimates away from the boundary and for large x are much better due to the
larger damping from b̄·∇ψψi
i
and the decay of the profile. See Section 2.7.2.

2.5

1.5

0.5

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
x

Figure 4. Coefficients ē1 , ē2 of the damping terms and the estimates of the
bad terms B1 , B2

For |x − z| sufficiently small but with other ratio |x2 − z2 |/|x1 − z1 |, from Sectin 2.7.2, we
can obtain a larger damping term from δb̄·∇ggi
i
in (2.40) and obtain better estimates. Using the
sharp functional inequalities in Section 3, we can also verify that the damping terms dominate.
The remaining step is to show that when |x − z| is not small, these conditions are also
satisfied. We remark that for larger |x − z|, the constants in the Hölder estimates improve. See
the discussion below Remark 2.5 and Lemmas 3.1-3.4. We will show later that the constants in
the Hölder estimates for the bad terms do not increase too much for |x − z| not too small.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 21

Remark 2.4. We remark that to obtain the stability condition (2.45), one can choose weights
similar to ψ1 , ψ2 with other parameters, or other weights. In view of the model and derivations
in Section 2.7.1, the gap −ci − (ēi + Si ) is larger if one chooses weight |x|γ with smaller γ.
In ψ1 , ψ2 (C.1), the power with the largest exponent is |x|1/6 , which leads to a smaller gap.
Note that to close the nonlinear estimates, we need to control the L∞ norm of ω, η, ξ, ∇u. We
choose this growing weight so that we have a stronger control of the solution in the far-field,
which leads to smaller constants in the nonlinear estimates and makes it easier to control the
nonlinear estimates.
2.8.3. A model problem for localized velocity and energy estimate. We consider the following
model problem to illustrate the ideas of our overall energy estimate and motivate the localization
of velocity (2.33)
(2.46) ωt = −d(x)ω + a(x)ux (ω, ε)(x).
Here ux (ω, ε) denotes the localized velocity ux (ω, a, b) (2.33) with b = a = ε. We assume
(2.47) − d(x) + m|a(x)| ≤ −c1 < 0, −d(x) ≤ −c2 < 0, d(x), |a(x)|, ||d||C 1/2 , ||a||C 1/2 ≤ c3 ,
for some constant c1 , c2 , c3 > 0, where m = max(C1 (∞), 21 C1 (∞) + C2 (∞)) ≤ 2.64. The
constants C1 (·), C2 (·) are defined in Lemmas 3.1, 3.3. From Lemmas 3.1, 3.3, we have
(2.48) ||ux (x, a, b)||C 1/2 ≤ m||ω||C 1/2 ,
uniformly for any a, b. The first condition in (2.47) corresponds to (2.45) and means that the
damping term dominates in the Hölder estimate if |x − y| is very small.
In (2.46), we remove the transport terms. Since in the weighted energy estimate, it contributes
to the damping terms (see Sections 2.7.1, 2.8.2), in this model, we incorporate this effect in the
damping terms −d(x)ω for simplicity. The nonlocal term ux (ω, ε) models other nonlocal terms
∇u in (2.17), (2.21). We remove the more regular nonlocal terms in (2.17), (2.21), e.g. u, cω ,
and the nonsingular part ux,R = ux − ux (ω, ε) , which will be estimated using the methods in
Section 2.9 and 2.10.
We argue that if ε is small enough, we can establish the stability analysis. In the following
estimate, C will be some absolute constant independent of the above parameters. For L∞
estimate, using the definition of ux (ω, ε) (2.33) and the Hölder norm of ω, we get
d
(2.49) |ux (ω, ε)(x)| ≤ Cε1/2 ||ω||C 1/2 , ||ω||L∞ ≤ −c2 ||ω||L∞ + Cc3 ε1/2 ||ω||C 1/2 .
dt
The above L∞ estimate is almost closed since we have a small parameter ε. For any x, z, denote
δ(f )(x, z) = f (x) − f (z). In the Hölder estimate, using (2.47), and a direct calculation, we yield
δ(dω)(x, z) δ(ω) d(x) − d(z) d(x) − d(z)
− = −d(x) − ω(z), ω(z) ≤ c3 ||ω||L∞ .
|x − z|1/2 |x − z|1/2 |x − z|1/2 |x − z|1/2
For a(x)ux (x, ε), using (2.47), (2.48), and the above L∞ estimate on ux , we obtain
δ(aux (ω, ε)) δ(ux (ω, ε)) a(x) − a(z)
≤ a(x) + |ux (ω, ε)(z)| ≤ |a(x)|m||ω||C 1/2 +Cc3 ε1/2 ||ω||C 1/2 .
|x − z|1/2 |x − z|1/2 |x − z|1/2
It follows
(2.50)
δω δω
∂t 1/2
= −d(x) + B(x, z), |B(x, z)| ≤ (|a(x)m| + Cc3 ε1/2 )||ω||C 1/2 + c3 ||ω||L∞ .
|x − z| |x − z|1/2
δω δω
Multiplying both sides by |x−z|1/2
and taking | |x−z|1/2 | ≤ ||ω||C 1/2 , we derive

1 (δω)2 (δω)2
∂t ≤ −(d(x) − |a(x)|m − Cc3 ε1/2 )
2 |x − z| |x − z|
(δω)2
+ (|a(x)|m + Cc3 ε1/2 )(||ω||2C 1/2 − ) + c3 ||ω||L∞ ||ω||C 1/2 .
|x − z|
22 JIAJIE CHEN AND THOMAS Y. HOU

c1
Suppose that ε is small enough such that Cc3 ε1/2 < 2 . From (2.47), we have d(x) − |a(x)|m −
2
Cc3 ε1/2
≥ c1
2 . Evaluating at the maximizer (x∗ , z∗ ) and using ||ω||2C 1/2 − (δω(x ∗ ,z∗ ))
|x∗ −z∗ | = 0, we get
(2.51)
1 d c1 d c1
||ω||2C 1/2 ≤ − ||ω||2C 1/2 + c3 ||ω||L∞ ||ω||C 1/2 , ||ω||C 1/2 ≤ − ||ω||C 1/2 + c3 ||ω||L∞ .
2 dt 2 dt 2
∞
Combining the above estimate and the L estimate (2.49), for ε small enough such that
c2 c1
Cc3 ε1/2 c3 < ,
2
using Lemma A.1 with n = 2 and (F1 , F2 ) = (||ω||L∞ , ||ω||C 1/2 ), we obtain that
s s
d Cc3 ε1/2 c1 /2 Cε1/2 c1 /2
(2.52) E ≤ −λE, E = max(||ω||L∞ , τ ||ω||C 1/2 ), τ = = ,
dt c3 c2 c2
for some λ > 0. The above inequality for ε is from (A.2), and we have chosen τ = µ2 offered in
Lemma A.1. We prove the desired stability.
Interpretation of the estimates. From the estimate (2.52), we see that the weight of ||ω||C 1/2
in the energy is smaller due to the factor ε1/2 . Since the L∞ estimate (2.49) is almost closed,
we can formally treat ||ω||L∞ as an a-priori estimate. We choose the weight τ ∼ ε1/4 so that
c3 ||ω||L∞ and Cc3 ε1/4 ||ω||C 1/2 in (2.51) are balanced. Then for ε small enough, we can close the
Hölder estimate. In our energy estimate of (2.17), (2.21), we will approximate the regular terms
so that we can establish the L∞ estimate with a small cost of the Hölder norm of ω similar to
(2.49).
For the localized velocity, we show in Lemmas 3.1-3.4 that the contants in the Hölder estimates
decrease if the distance |x − z| increases. This improvement allows us to show that the worst
scenario for the Hölder stability estimate occurs when |x − z| is small.
Remark 2.5. Our energy estimate to be presented later follows ideas similar to those mentioned
above. Yet, since it involves variable coefficients and several singular weights, the estimates are
more technical. We will also track the constants and choosing the weight, e.g. τ in (2.52), much
more carefully so that we do not need to choose ε to be too small, or approximate the regular
terms using finite rank operators with a very high rank to get a small approximation error in a
suitable norm. This will reduce our computation cost significantly. See Sections 2.9 and 2.10.
The choice of the parameters in the above example are only used to illustrate the ideas.
2.9. A toy model with a nonlocal term. In this section, we use a model problem to illus-
trate the ideas of stability analysis of a linearized equation perturbed from a simpler linearized
equation. In Section 2.10, we will apply these ideas to decompose the system (2.17), (2.21) and
estimate the more regular nonlocal terms in (2.17), (2.21), e.g. u · ∇ω̄.
We consider the following model
(2.53) ft = −b(x) · ∇f + d(x)f + a(x)P (f ) , Lf
in R++
2 , where the functions b(x), d(x), a(x) are given and
Z
bi (x) ≥ ci xi , ci > 0, P (f ) = f gdx,
R++
2

for some time-independent function g. The terms b(x)·∇f, d(x)f model the local terms in (2.17),
(2.21), and the rank one operator a(x)P (f ) models the nonlocal terms. We want to understand
the long time behavior and the stability of the above model.
If the nonlocal term a(x)P (f ) is absent, we can follow the derivations in Section 2.7.1 to
perform stability analysis. We assume that
(2.54) ft = −b(x) · ∇f + d(x)f , L0 f
is linearly stable in L∞ (ϕ) with some singular weight ϕ. Can we use this information to study
(2.53)? Given initial data f0 , a natural attempt is to use the Duhamel principle to represent
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 23

the solution to (2.53)

Z t Z t
L0 t L0 (t−s) L0 t
(2.55) f (t) = e f0 + e (a(x)P (f (s)))ds = e f0 + P (f (s))eL0 (t−s) a(x)ds,
0 0
where eL0 t g represents the solution to (2.54) at time t from initial data g, and we have used the
fact that P (f (s)) is constant in space in the second equality. Although we know that eL0 t f0 and
eL0 (t−s) a(x) decay in L∞ (ϕ) due to the stability, if a(x) is not small, the nonlocal term P (f (s))
and the contribution from the integral can lead to the growth of f (t). In this case, it is not clear
if f (t) decays in some norm for large time.
Another attempt is to project f onto some space Y orthogonal to g(x) or a(x) so that the
contribution from the nonlocal term is 0 in Y . However, since our estimate is L∞ -based rather
than L2 -based, the orthogonal structure and projection are not compatible with these estimates.
2.9.1. Rank-one perturbation. Since (2.53) is a rank one perturbation of (2.54), we expect that
the solution f to (2.53) can be decomposed as f = f1 + f2 , where the main part f1 has behavior
similar to the solution of (2.54) and f2 can be seen as a low rank correction. In particular, we
draw inspiration from the Sherman-Morrison formula [71] in linear algebra. Given an invertible
square matrix A ∈ Rn×n and column vectors u, v ∈ Rn , the matrix A + uv T is invertible if and
only if 1 + v T A−1 u 6= 0. In this case,
A−1 uv T A−1
(2.56) (A + uv T )−1 = A−1 − .
1 + v T A−1 u
This formula shows that to check the invertibility of the perturbed matrix, it suffices to test
the action of A−1 on these perturbed directions u, v. Now, we generalize this idea to study
(2.53) from (2.54). Given initial data f0 , we decompose (2.53) as follows
∂t f1 (t) = L0 f1 , f1 (0) = f0 ,
(2.57)
∂t f2 (t) = Lf2 + a(x)P (f1 (t)), f2 (0) = 0.
It is easy to see that f1 + f2 solves (2.53) with initial data f0 . The first part f1 satisfies (2.54).
The second part f2 is driven by the rank-one forcing term a(x)P (f1 (t)). Using the Duhamel
principle, the fact that P (f1 (t)) is constant in space, and derivations similar to (2.55), we yield
Z t
(2.58) f2 (t) = P (f1 (s))eL(t−s) a(x)ds.
0
Since f1 (t) = eL0 t f0 decays in L∞ (ϕ), we can control P (f1 (s)). If eLt a(x) decays in L∞ (ϕ)
for large t, we can show that f2 (t) also decays in L∞ (ϕ) and establish stability estimate of f2 .
Remark 2.6. The initial data f0 can be an arbitrary function in some weighted space. By
choosing zero initial data for f2 and using the fact that P (f1 (s)) is a scalar, we can solve f2 for
an arbitrary forcing coefficient P (f1 (s)).
Let us compare the above decomposition with (2.56). The operators L0 , L and rank-one
perturbation a(x)P (f ) are analogs of A, A + uv T and uv T . The stability of L is analogous to
the invertibility of A + uv T . Testing the stability of L from initial data a(x) or the decay of
eL(t−s) a(x) is similar to testing A−1 on the directions u, v so that 1 + v T A−1 v 6= 0.
Similar idea also appears in the celebrated T (1) [22] and T (b) [23, 59] theorems in Harmonic
analysis. Roughly, it states that for a linear operator T associated with a standard kernel K,
proving the L2 boundedness of T reduces to testing T on 1 or on some function b. In our model
problem, using energy estimate to establish the stability of (2.54) is similar to extracting certain
properties of T from a standard kernel. Further, testing the operators L on the perturbations
to obtain its stability is similar to testing T on 1 or b to obtain the L2 boundedness of T .
2.9.2. Constructing approximate solution to f2 and a perturbed operator. In this section, we
discuss how to verify the decay of eLt a. Though the operator L and function a are given, it is
difficult to prove decay estimates of eLt a in the weighted norm analytically since L involves a
nonlocal term. In our application to the Boussinesq system, analyzing eLt a analytically is even
more challenging since the operator L and the nonlocal terms are much more complicated.
24 JIAJIE CHEN AND THOMAS Y. HOU

An alternative approach is to solve (2.53) numerically from initial data a(x) to obtain an
approximate solution ĝ(t, x). Then by showing the error eLt a − ĝ(t, x) is small in the weighted
norm with a singular weight and verifying the decay of ĝ(t), we obtain the decay estimates of
eL a. The difficulty lies in estimating the error in the weighted norm rigorously. The standard
a-priori error estimate provides the following estimate
|eLt a(x) − ĝ(t, x)| ≤ C1 (hm + k n )eC2 t ,
for some constants C1 , C2 depending on a(x) and L, where h is the mesh size, k is the time
step in the computation, and m, n relate to the order of the numerical scheme. However, the
constants are not easy to estimate and can be quite large, and the above estimate does not
account the round off error. Since the solution can have a slow decay in R2+ , our computation
domain can be very large. As a result, choosing a very small mesh size h in the computation
may not be practical. Moreover, t is not small since we want to obtain decay estimates of eLt a
for suitably large t, e.g., t ≥ 10. The above estimate is not practical.
Instead, we seek a-posteriori error estimate. Firstly, we solve (2.53) numerically and obtain
a numerical solution ĝ(tk , x) at time tk , which is represented by piecewise polynomials and thus
defined globally in x. Then we interpolate the solution ĝ(tk , x) in time t using piecewise cubic
polynomials to obtain solution ĝ(t, x) defined on [0, T ] × R+ 2 . Then we introduce the residual
error and a residual operator R related to the nonlocal term P (f1 (t)) in (2.57)
e(t, x) , (∂t − L)ĝ, e0 (x) = ĝ(0) − g0 , g0 = a(x)
(2.59) Z t
R(f1 , t) = P (f1 (t))e0 (x) + P (f1 (s))e(t − s, x)ds.
0

Since ĝ is defined everywhere in space and time, we can estimate e(t, x) and e0 (x).
Now, using the approximate solution ĝ(t, x) for eLt g0 , we can construct the approximate
solution to the f2 -equation in (2.57) based on (2.58)
Z t
(2.60) fˆ2 (t) = P (f1 (s))ĝ(t − s)ds.
0

By definition and (2.59), we have

Z t
ˆ
(∂t − L)f2 = P (f1 (t))ĝ(0) + P (f1 (s))(∂t − L)ĝ(t − s)ds
0
Z t
= P (f1 (t))(a(x) + e0 ) + P (f1 (s))e(t − s, x)ds = P (f1 (t))a(x) + R(f1 , t).
0

If the error e(t, x) and e0 are small, we can show that the norm of the residual operator is small
in some suitable functional space
(2.61) ||R(f1 , t)||X ≤ ε||f1 ||X , ε << 1.

Now, we modify the decomposition (2.57) as follows (f = f1 + fˆ2 )

∂t f1 = L0 f1 − R(f1 , t), f1 (0) = f0 ,
(2.62)
∂t fˆ2 = Lfˆ2 + a(x)P (f1 (t)) + R(f1 , t).
We remark that the solution (2.60) constructed by the numerical solution ĝ solves the second
equation exactly. Now, due to the smallness (2.60), we can apply the stability estimate of L0
and treat R(f1 , t) as perturbation to obtain stability estimate of f1 .
Remark 2.7. We should compare the original problem (2.53) with (2.62). By performing the
decomposition (2.57),(2.62), constructing an approximating solution ĝ to eLt a(x), and testing
its decay, we replace a difficult nonlocal term in (2.53) by a small error term R(f1 , t) in (2.62)
that can be treated as a small perturbation. Moreover, we do not need to assume any specific
form about the rank-one operator a(x)P (f1 (t)).
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 25

2.10. Finite rank perturbations to the linearized operators. We generalize the idea in
the previous section to the Boussinesq equations. Using the ideas in Section 2.9, we can modify
the operators in (2.17), (2.21) by a finite rank operator K with rank N by testing these operators
on N suitable functions. Then we perform energy estimate on L−K. These finite rank operators
approximate the contributions from the more regular terms in (2.17), (2.21), e.g., u · ∇ω̄, u · ∇θ̄x ,
which we neglect in Section 2.8. The more regular terms can be seen as compact operators of ω
in some suitable weighted spaces.
Since we will perform weighted estimates with singular weights near x = 0, it is useful
to rewrite (2.17),(2.21) such that each term has the right vanishing order. We introduce the
following notations [12, 13]
(2.63) ũ = u − ux (0)x, ṽ , v − vy (0)y = v + ux (0)y.
Since ωx (0) = 0 (2.22), ωx (0) = −∆ψx (0), ψ(x, 0) = 0, and ψ is odd in x, we yield
ψ = O(|x|4 ), ũ = O(|x|3 ), ṽ = O(|x|3 ), ∇ũ = O(|x|2 ),
for perturbations that are regular enough. Recall cω = ux (0) (2.19). Using ũx = ux −ux (0), ũy =
uy , ṽx = vx , ṽy = vy + ux (0),
−u · ∇ω̄ + cω ω̄ = −ũ · ∇ω̄ + cω (ω̄ − xω̄x + y ω̄y ),
−u · ∇θ̄x − ux · θ̄ + 2cω θ̄x = −ũ · ∇θ̄x − ũx · ∇θ̄ + cω (θ̄x − xθ̄xx + y θ̄xy ),
−u · ∇θ̄y − uy · ∇θ̄ + 2cω θ̄y = −ũ · ∇θ̄y − ũy · ∇θ̄ + cω (θ̄y − xθ̄xy + y θ̄yy ),
and denoting
(2.64) f¯c ω ,1 = ω̄ − xω̄x + y ω̄y , f¯cω ,2 = θ̄x − xθ̄xx + y θ̄xy , f¯cω ,3 = 3θ̄y − xθ̄xy + y θ̄yy ,
we can rewrite (2.17), (2.21) as follows
∂t ω = −(c̄l x + ū) · ∇ω + c̄ω ω + η − ũ · ∇ω̄ + cω f¯cω ,1 + N1 + F 1 , L1 + N1 + F 1 ,
∂t η = −(c̄l x + ū) · ∇η + (2c̄ω − ūx )η − v̄x ξ − ũx · ∇θ̄ − ũ · ∇θ̄x + cω f¯c ,2 ω

(2.65) + ∂x (Nθ + F̄θ ) , L2 + N2 + F 2 ,

∂t ξ = −(c̄l x + ū) · ∇ξ + (2c̄ω + ūx )ξ − ūy η − ũy · ∇θ̄ − ũ · ∇θ̄y + cω f¯cω ,3
+ ∂y (Nθ + F̄θ ) , L3 + N3 + F 3 .
The nonlocal terms cω f¯cω ,i , −ũ · ∇f for f = ω̄, θ̄x , θ̄y , and ∇ũR , the nonsingular part of the
integral, are more regular than ω. We will choose finite rank operators to approximate them.
2.10.1. Correction near the origin. We discuss in Section 2.7.5 that to obtain better stability
factors, we choose more singular weights for the stability analysis. We consider the following
corrections
K1i (ω) , cω (ω)f¯c ,1 , ω

N F1 (ω, η, ξ) = (cω ωxy (0) + ∂xy F 1 (0))fχ,1 , fχ,1 , xyχ,

(2.66) N F2 (ω, η, ξ) = (cω ηxy (0) + ∂xy F 2 (0))fχ,2 , fχ,2 , xyχ,
x2
N F3 (ω, η, ξ) = (cω ξxx (0) + ∂xx F 3 (0))fχ,3 , χ, fχ,3 ,
2
where χ is some cutoff function with χ = 1 near x = 0 defined later. The operator K1i is
correction to the linear part, and N Fi is correction to the nonlinear term, and the residual in
(2.65), respectively. After subtracting K1i and N Fi from (2.65), the resulting equations preserve
the vanishing conditions ω, η, ξ = O(|x|3 ).
We can derive the ODE for ωxy (0), θxxy (0) using (2.10)
d
ωxy (0) = (−2cl + cω )ωxy (0) − ωx2 (0) + θxxy (0),
(2.67) dt
d
θxxy (0) = (−2cl + 2cω − ux (0))θxxy (0) − 2ωx (0)θxx (0).
dt
26 JIAJIE CHEN AND THOMAS Y. HOU

Since ωx (0), θxx (0) are preserved (2.12), to estimate ωxy (0), θxxy (0), using (2.67), we only need
to control cl , cω , ux (0) rather than some higher order norm of ω, θ, e.g. ||ω||C 2 , ||θ||C 3 .

2.10.2. Approximation of the velocity. For f = u, v, ux , uy , vx , vy , we will construct in (2.89),

ˆ
(2.82), (2.88) in Section 2.11 the finite rank approximations fˆ for f˜ so that we get smaller
ˆ
constants C in the weighted estimate of f˜− f˜ using the energy ||ωϕ||L∞ , ||ωψ1 ||C 1/2 , ||ωψ1 ||C 1/2 .
x y
We remark that for these operators, we do not have

i j
∂xi ∂yj û = ∂\ i j
∂xi ∂yj v̂ = ∂\
x ∂y u, x ∂y v,

for i + j = 1. These approximations contribute to the following finite rank operators

(2.68) ˆ · ∇ω̄,
K21 = −ũ ˆ x · ∇θ̄ − ũ
K22 = −ũ ¯ x,
ˆ · ∇θ ˆ y · ∇θ̄ − ũ
K23 = −ũ ˆ · ∇θ̄y ,

which are designed to capture the contributions from the regular nonlocal terms. We defer the
construction of K2i to Section 2.11.
The linearized operators in (2.65) perturbed by these finite rank operators are similar to the
one in Section 2.8.3, which we can estimate following the ideas in Section 2.8.3. This stability
estimate serves as the role of stability estimate of L0 in the model problem in Section 2.9. Then
we further test the original operators in (2.65) on these finite rank operators using the ideas in
Section 2.9 and next section to close the stability analysis.

2.10.3. Decomposition of the system. Denote W1 = (ω1 , η1 , ξ1 ), W2 = (ω2 , η2 , ξ2 ). Recall the

notations (2.1) and (2.2). Following Section 2.9 and (2.57), we decompose (2.65) as follows

∂t W1,i = (Li − K1i − K2i )W1 + Ni (W1 + W2 ) + F i − N F (W1 + W2 ),

(2.69) ∂t W2,i = Li W2 + K1i (W1 ) + K2i (W1 ) + N Fi (W1 + W2 ),
W1 |t=0 = (ω0 , η0 , ξ0 ), W2 |t=0 = (0, 0, 0),

with ω0 , η0 , ξ0 being the initial perturbation with vanishing order O(|x|3 ).

Firstly, we show that W = W1 + W2 solves (2.65) with initial data = (ω0 , η0 , ξ0 ). A direct
calculation yields

∂t (W1 + W2 ) = Li (W1 + W2 ) + Ni (W1 + W2 ) + F i ,

which are the same equations as (2.65). Since W1 + W2 has initial data (ω0 , η0 , ξ0 ), W1 + W2
solves (2.65) with the given initial data. Using the definitions (2.66) and a Taylor expansion
near x = 0, we obtain that the vanishing conditions ω1 , η1 , ξ1 = O(|x|3 ) are preserved.

Remark 2.8. Although W1,2 + W2,2 = θx , W1,3 + W2,3 = θy , since the finite rank operators
Kij we choose do not satisfy similar partial derivative relations, the solution to (2.69) does not
satisfy ∂y Wi,2 = ∂x Wi,3 for i = 1 or i = 2.

Let us motivate the decomposition (2.69). At the linear level, we choose finite rank operators
K1i , K2i to approximate Li . Then Li − K1i − K2i , Li serve as the L0 , L operators in the model
problem (2.54), (2.53), respectively. The decomposition of the solutions W1 , W2 is similar to
(2.57). Since we want to perform energy estimate on W1 using more singular weights, we correct
the nonlinear terms and the forcing terms in the first equation in (2.69). Although N Fi (W1 +W2 )
involves nonlinear factors, e.g. cω (ω1 + ω2 )∂xy (ω1 + ω2 )(0), since these factors are constant in
space, we can still apply Duhamel’s formula in (2.55) to N Fi (W1 + W2 ), i.e.,
Z t   Z t
e L(t−s) ¯
cω ∂xy (ω1 (s) + ω2 (s))(0)f ds = cω ∂xy (ω1 (s) + ω2 (s))(0)eL(t−s) f¯ds,
0 0

and obtain the formula of W2 in (2.69).

 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 27

Avoiding the loss of derivatives. Note that in the equation of W1 in (2.69), it contains the
nonlinear terms u(W1 + W2 ) · ∇(W1 + W2 ). In general, the term u · ∇W2 can lead to loss of
derivatives. We overcome this difficulty by constructing smoother W2 . Note that W2 in (2.69)
is driven by the forcing terms of the following forms
X
(2.70) ai (W1 , W2 )(t)fi
1≤i≤N

for some N , time-dependent scalars (independent of x) ai (W1 , W2 )(t), and time-independent

functions fi , e.g. cω (W1 )f¯cω ,i in (2.66). By choosing smoother fi in the approximation, we can
obtain solution W2 smooth enough for our energy estimates.

2.10.4. Constructing the approximate solution of W2 and modifying the decomposition. Follow-
ing the ideas in Section 2.9.2, instead of solving the W2 equations in (2.69) exactly, we solve
them with an error term. Assume that we have the following representations for the operators
X
~ 1j (W1 ) + K
K ~ 2j (W1 ) = ai (W1 )(t)F ij ,
1≤i≤n1

where K~ i· = (Ki1 , Ki2 , Ki3 ), F̄i (x) = (f¯i,1 (x), f¯i,2 (x), f¯i,3 (x)) : R2+ → R × R × R, and ai (W1 )(t)
is some linear functional on W1 . For example, the formula (2.66) can be written as
a(W1 )(t)(f¯cω ,1 , f¯cω ,2 , f¯cω ,3 ), a(W1 )(t) = cω (W1 ) = ux (W1 (t))(0),
where we have used (2.19) for cω . Recall the operators N Fi and functions fχ,1 from (2.66).
Writing (2.66) as vectors, we have
N F (ω) = (cω ∂xy ω(0) + ∂xy F 1 (0))(fχ,1 , 0, 0) + (cω ∂xy η(0) + ∂xy F 2 (0))(0, fχ,2 , 0)
(2.71)
X
+ (cω ∂xx ξ(0) + ∂xx F 3 (0))(0, 0, fχ,3 ) , anl,i (ω(t))Fχ,i ,
1≤i≤3

where
(2.72)
anl,· = (cω ∂xy ω(0) + ∂xy F 1 (0), cω ∂xy η(0) + ∂xy F 2 (0), cω ∂xx ξ(0) + ∂xx F 3 (0)), Fχ,i = fχ,i ei ,
and e1 , e2 , e3 are the standard basis for R3 . Denote by F̂i (t, x) and F̂χ,i (t, x) the approximation
of eLt F̄i and eLt F̄χ,i . Following (2.59) and (2.60), and using the idea in (2.70), we construct the
approximate solution to W2 in (2.69) as follows
XZ t XZ t
(2.73) Ŵ2 (t) = ai (W1 (s))F̂i (t − s)ds + anl,i (W1 (s) + Ŵ2 (s))F̂χ,i (t − s)ds.
i≤n1 0 i≤3 0

We introduce the residual operator

n1 
X Z t 
Rl (W1 ) , ai (W1 (t))(F̂i (0) − F̄i ) + ai (W1 (s))((∂t − L)F̂i )(t − s)ds ,
i=1 0
3  t
(2.74) X Z 
Rnl (W ) , anl,i (W (t))(F̂χ,i (0) − F̄χ,i ) + anl,i (W (s))((∂t − L)F̂χ,i )(t − s)ds ,
i=1 0

R(W1 , Ŵ2 ) , Rl (W1 ) + Rnl (W1 + Ŵ2 ),

where Rl , Rnl denote the linear and the nonlinear parts, respectively. Note that R(W1 , Ŵ2 )(x)
is a vector in R3 .
Remark 2.9. We remark that given W1 and F̂i , F̂χ,i , the solution Ŵ2 (2.73) is not completely
determined since the second part depends on Ŵ2 . At the linear level, the solution Ŵ2 (2.73)
is determined. Since the second part depends on Ŵ2 nonlinearly, we will show that it is small
relative to the linear part using a bootstrap argument. Therefore, we can still use (2.73) to
estimate Ŵ2 .
28 JIAJIE CHEN AND THOMAS Y. HOU

Similar to (2.62), using the above operators, we modify the decomposition (2.69) as follows
∂t W1,i = (Li − K1i − K2i )(W1 ) + Ni (W1 + Ŵ2 ) + F i − N Fi (W1 + Ŵ2 ) − Ri (W1 , Ŵ2 ),
(2.75) ∂t Ŵ2,i = Li Ŵ2 + K1i (W1 ) + K2i (W1 ) + N Fi (W1 + Ŵ2 ) + Ri (W1 , Ŵ2 ),
W1 |t=0 = (ω0 , η0 , ξ0 ), Ŵ2 |t=0 = (0, 0, 0),
where R = (R1 , R2 , R3 ). The above decomposition is a nonlinear generalization of (2.62). We
solve the Ŵ2 equation using the formula (2.73) exactly. It is easy to see that W1 + Ŵ2 solves
(2.65) from initial data (ω0 , η0 , ξ0 ). If the error in (2.74), e.g. F̂i (0) − F̂i , (∂t − L)F̂i , is small,
we expect that the following estimates for R:
||R(W1 , Ŵ2 )||X ≤ ε(||W1 ||X + ||W1 + Ŵ2 ||2X + ε̄)
in some suitable weighted space X with very small ε, ε̄, where the second and the third terms
come from the estimate of anl,i (W1 + Ŵ2 ) defined in (2.71). Since F i is the residual error of the
profile, for i + j = 2, ∂xi ∂yj F(0) is small and contributes to the small factor ε̄. Thus, the residual
operator R can be treated as a small perturbation in (2.75). In particular, at the linear level,
Ŵ2 does not appear in the W1 equation.
We will construct approximate solution F̂i and F̂χ,i so that the error vanishes cubically near
x = 0:
(2.76) F̂i (0) − F̄i , (∂t − L)F̂i = O(|x|3 ), F̂χ,i (0) − F̄χ,i , (∂t − L)F̂χi = O(|x|3 ).
We will discuss how to construct these approximate solution and estimate these errors in the
weighted functional spaces for the energy estimate rigorously in Section 7.
For initial perturbation ω0 , η0 , ξ0 = O(|x|3 ), from the definitions (2.66) and the above vanish-
ing order of the error, we obtain that the vanishing conditions ω1 , η1 , ξ1 = O(|x|3 ) are preserved.
Thus, we can perform energy estimates on W1 using singular weights of order |x|−3 near x = 0.
See Section 2.7.5 for more discussions on the vanishing order. We will perform the energy
estimates in Section 4.
Remark 2.10. Since F̂i is the numerical solution to ∂t Fi = LFi , the above error terms (2.76)
are small. Since the solution is smooth enough, in principle, by choosing a high order numerical
scheme with sufficiently small mesh size and timestep, one can make the error term to be
arbitrarily small. In such a case, the residual operators in (2.74), (2.75) are also sufficiently
small compared to the perturbation W1 , W .
2.11. ApproximatingR the regular part of the velocity. We want to construct a finite rank
approximation K(ω)of R2 Kf (x − y)ω(y)dy, where Kf is the kernel for ∂xi ∂yj (−∆)−1 ω, i + j ≤ 2,
so that we can estimate
Z
(2.77) | Kf (x − y)ω(y)dy − K(ω)| ≤ C1 (x)||ωϕ1 ||L∞ + C2 (x)||ωψ1 ||C 1/2 + C3 (x)||ωψ1 ||C 1/2 ,
x y
R2
with small constant C1 (x), C2 (x), C3 (x) for some given weights, where the Hölder seminorms
1/2 1/2
Cx , Cy are defined in (2.3).
Since Kf (z) is smooth away from z = 0, a natural approach is to approximate the nonsingular
part of K(x − y) by interpolating K(x − y) on finite many points xi :
X
K(x − y)1|x−y|≥ε ≈ χi (x)K(xi − y)1|xi −y|≥ε ,
1≤i≤n

where χi is some cutoff function localized to xi . The above right hand sides lead to the finite
rank operator
Xn Z
(2.78) K(ω) = χi (x) K(xi − y)1|xi −y|≥ε ω(y)dy.
i=1

We will construct the bulk part of the approximation in Section 2.11.2 based on (2.78). Due
¯ x in ∇u · ∇θx , these
to the decay of the coefficients of u, ∇u in (2.65), e.g. ∇ω̄ in u · ∇ω̄, ∇θ
nonlocal terms are small for large |x|. Thus, we only need to approximate u, ∇u for x in the
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 29

near field, especially for x close to the boundary due to the anisotropy of the flow. See Section
2.7.2.
Regularity of the velocity. For u = ∇⊥ (−∆)−1 ω, given ω in some weighted space L∞ (ϕ), u is
log-Lipschitz. Thus we can approximate u(x) in C β ∩L∞ for any β < 1 by interpolating discrete
points u(xi ), i = 1, 2, ..., n with n sufficiently large. The C β ∩ L∞ norm of the approximation
error can be bounded by c||ωϕ||∞ with a small constant c. Similarly, for ∇u = ∇∇⊥ (−∆)−1 ω,
given ω in some weighted L∞ space, the nonsingular part of ∇u, K(z)1|z|≥ε ∗ ω, is Lipschitz.
Thus we can approximate it in C 1/2 ∩ L∞ . Since ∇u = ∇∇⊥ (−∆)−1 ω is not bounded from L∞
to L∞ , for the singular part of ∇u, K(z)1|z|≤ε ∗ ω, we need to use the Hölder regularity of ω to
control it. These motivate (2.77).

2.11.1. Approximation near 0. Since we will perform weighted energy estimates with singular
weights and the velocity u, ∇u do not vanish near x = 0 with high order, we first approximate
u, ∇u by its leading order behavior at x = 0.
In our energy estimate, we consider perturbation ω with vanishing order O(|x|2+α ) for some
α > 0 near x = 0. Recall −∆ψ = ω and u = ∇⊥ ψ. Using Taylor expansion and
0 = ωx (0) = −ψxxx (0) − ψxyy (0), 0 = ωxy (0) = −ψxxxy (0) − ψxyyy (0),
we get
1 1
ψ(x, y) = ψxy (0)xy+ (ψxxxy (0)x3 y+ψxyyy (0)xy 3 )+h.o.t. = ψxy (0)xy+ ψxxxy (0)(x3 y−xy 3 )+h.o.t.
6 6
We can represent ψxxxy (0) as an integral of ω
(2.79)
24y1 y2 (y12 − y22 )
Z Z
2 1
ψxxxy (0) = ω(y)K00 (y)dy, K00 (y) , , K00 (ω) , K00 (y)ω(y)dy.
π R2++ |y|8 π R2++

For ω = O(|x|2+α ) with a suitable decay, the above integral is well-defined. By definition, we
have ψxxxy (0) = 2K00 . Here and below, we use K00 as a short hand notation for K00 (ω). Note
that ux (0) = −ψxy (0). Using the above formulas, near 0, the leading order term for ∇u and u
are given by
1 1
u = −ψy = ux (0)x − ( x3 − xy 2 )K00 + h.o.t., v = ψx = −ux (0)y + (x2 y − y 3 )K00 + h.o.t.,
3 3
ux = ux (0) − (x2 − y 2 )K00 + h.o.t., vx = 2xyK00 + h.o.t., uy = −ψyy = −2xyK00 + h.o.t.
By introducing
Cu0 = x, Cv0 = −y, Cux 0 = 1, Cuy 0 = Cvx 0 = 0,
1 3 1
(2.80) Cu = −( x − x2 y), Cv = x2 y − y 3 ,
3 3
Cux = −(x2 − y 2 ), Cvx = 2xy, Cuy = −2xy,
we can rewrite the above leading order formulas as
(2.81) f (x, y) = ux (0)Cf 0 (x, y) + K00 Cf (x, y) + h.o.t., f = u, v, ux , vx , uy .
For vy , we will use vy = −ux to estimate it. We will localize the above leading order terms
to construct the approximation term near 0 in the next subsection.

2.11.2. Approximation along the boundary. Let χ be the cutoff function constructed in (F.1)
and χ̃ = 1 − χ. They satisfy
χ(x) = 0, χ̃(x) = 1, x ≤ 0, χ(x) = 1, χ̃(x) = 0, x ≥ 1.
Given 0 < x0 < x1 < ... < xn < xn+1 and y0 > 0, we construct the cutoff functions
x − x0 y − y0  x−x
i−1 x − xi  y−y
0
χ0 = χ̃( )χ̃( ), χi = χ( )1x≤xi + χ̃( )1x≥xi χ̃( ), 1 ≤ i ≤ n.
x1 − x0 y0 xi − xi−1 xi+1 − xi y0
30 JIAJIE CHEN AND THOMAS Y. HOU

We impose the cutoff function χ̃( y−y

y0 ) so that χi is supported near y = 0. By definition, for
0

x ≤ xn , y ≤ y0 , we have
X
χi (x, y) = 1.
i≤n

We want to approximate u, ∇u such that the remainders u−uapp vanish near x = 0 with high
order. See Section 2.11.1. To preserve these vanishing orders in the approximations and obtain
smoother approximations, we consider the following approximation, which modifies (2.78)
(2.82)
 X 1 
fˆ1 (x, y) , Cf 0 (x, y)ux (0)+fˆ10 (x, y), fˆ10 , Cf (x, y) K00 χ0 (x, y)+ χi (x, y)fˆ(xi , 0) ,
Cf (xi , 0)
1≤i≤n

where
Z
(2.83) fˆ(xi , 0) = Kf ((xi , 0) − y)ω(y)dy − Cf 0 (xi , 0)ux (0),
y∈R2 ,max(|y1 −xi ,y2 |)≥ti

and Kf is the kernel for f = u, v, ux , vx , uy , and functions Cf 0 and Cf are the coefficients of the
leading order approximations of ∂xi ∂yj ψ near x = 0. See (2.81) and (2.80). We add the functions
Cf (x, y) in (2.82) so that fˆ1 has the same vanishing order as that of f .
We construct the above approximations along the boundary for u, ux . For v, vx , uy , since
the associated coefficients are relatively small, e.g. ω̄y in v ω̄y and θ̄y in vx θ̄y (2.65), we only
construct the approximation term Cf (x, y)K00 χ0 (x, y) near 0. Now, by definition, we have

f (xi , 0) − fˆ1 (xi , 0) = f (xi , 0) − Cf 0 (xi , 0)ux (0) − fˆ(xi , 0), 0≤i≤n
(2.84)
f (x, y) − fˆ1 (x, y) = f (x, y) − Cf 0 (x, y)ux (0) − Cf (x, y)K00 , for x ≤ x0 , y ≤ y0 ,

The first identity shows that fˆ1 is an interpolation of the non-singular part of Kf ∗ ω, which
is similar to (2.78). Here, we consider a weighted version of (2.78) with weight Cf (x) so that
the approximation has the right vanishing order near x = 0. The second identity shows that
near x = 0, the approximation fˆ1 captures the leading order behavior of f near x = 0 (2.81).
Thus, fˆ1 can approximate f near the points (0, 0), (xi , 0), 0 ≤ i ≤ n.

2.11.3. Approximation in the far-field. To improve the far-field estimate, instead of using ux (0)
to approximate u, v, ∂u (2.83), we use the truncated version of ux (0)
Z
4 y1 y2
In , − ω(y)dy.
π max(y1 ,y2 )≥Rn |y|4
Note that the above approximation is similar to the leading order term of the velocity derived
in [48]. For f = ux , the leading order part of the kernel K(s) = − π1 s|s|
1 s2
4 with symmetrization is

given by
(2.85)
4 y1 y2
K sym (x, y) = K(x − y) + K(x + y) − K(x1 − y1 , x2 + y2 ) − K(x1 + y1 , x2 − y2 ) = − + l.o.t.
π |y|4
for max |yi | ≥ C max |xi | with large C. For f = u, v, vy , using a similar argument, we can show
that the leading order part of the associated kernel Kf is
4 y1 y2
(2.86) Kfsym = −Cf 0 + l.o.t.
π |y|4
for max |yi | ≥ C max |xi | with large C, where Cf 0 is defined in (2.80). When |y| is small and |x|
is large, the function − π4 y|y|
1 y2
4 does not approximate K
sym
(x, y) well, and thus we truncate it.
Note that the operator
Z
y1 y2
(2.87) − 4 ω(y)dy
max |yi |≥C max |xi | |y|
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 31

does not have a finite rank due to the hard cutoff function 1|y|≥C|x| . To approximate it by a
finite rank operators, we approximate 1|y|≥C|x| by a smooth cutoff function
X
g(x, y) = 1|y1 |∨|y2 |≥Ri χi (x)
i

such that χi (x) is localized to the domain with |x| comparable Ri . Then for x close to Ri , we
obtain g(x, y) ≈ 1|y1 |∨|y2 |≥Ri ≈ 1|y1 |∨|y2 |≥|x| .
More specifically, given R0 < R2 < .. < Rm , we construct cutoff functions as follows
x − Ri y − Ri x − Ri−1 y − Ri−1
χR
i (x, y) = χ̃( )χ̃( ) − χ̃( )χ̃( ), 1 ≤ i ≤ m − 1,
Ri+1 − Ri Ri+1 − Ri Ri − Ri−1 Ri − Ri−1
x − Rm−1 y − Rm−1
χR
m (x, y) = 1 − χ̃( )χ̃( ).
Rm − Rm−1 Rm − Rm−1
By definition, χRi is supported in the annulus
2 2 R
P [0, Ri+1 ] \[0, Ri−1 ] , 1 ≤ i ≤ m − 1, and χm = 1
for max(x, y) ≥ Rm . Moreover, we have i≤m χR i (x, y) = 1 for max(x, y) ≥ R1 . Now, we
construct the second approximation
 X 
(2.88) fˆ2 (x, y) = Cf 0 (x, y)(1 − χtot (x, y)) χR
i (x, y)(Ii − ux (0)) ,
1≤i≤m

where χtot (x, y) is the sum of the cutoff functions for the first approximation in Section 2.11.2:
X
χtot (x, y) = χi (x, y).
0≤i≤n

For χtot (x, y) = 0 and x = Ri , y ∈ [0, Ri ], from (2.83), we get fˆ1 = Cf 0 (x)ux (0), fˆ2 = Cf 0 (Ii −
ux (0)),
Z
ˆ ˆ
f − f1 − f2 = f − Cf 0 ux (0) − Cf 0 (Ii − ux (0)) = (Kf (x − y)ω(y)dy − Cf 0 Ii
Z
4 y1 y2
= (Kfsym (x, y) + Cf 0 (x) 4 1max(|y1 |,|y2 |)≥Ri )ω(y)dy,
R2++ π |y|

where Kfsym is the symmetrized kernel. Therefore, for large |x|, the approximation fˆ1 + fˆ2 can
be seen as a smooth interpolation of the main term related to (2.86), (2.87) with C = 1. In
view of the leading order structure (2.85), (2.86), we can obtain a better estimate of the above
integral than the original integral without approximation.
Remark 2.11. We remark that the lower order term in (2.85), (2.86) is about |x|/|y| of the
main term, which is not small if |y| and |x| are comparable. As a result, the above finite rank
approximation (2.88) can only approximate part of the integral in f . Nevertheless, it allows us
to obtain a better estimate of f − fˆ1 − fˆ2 than f , which is sufficient for our purpose.
Combining (2.82), (2.88), we construct the following approximation for f = u, ∇u and for
f˜ = f − Cf 0 ux (0)
ˆ
(2.89) fˆ(ω) , fˆ1 (ω) + fˆ2 (ω), f˜ , fˆ(ω) − Cf 0 ux (0) = fˆ10 (ω) + fˆ2 (ω),
ˆ
where the notations ũ, ṽ are introduced in (2.63). Clearly, fˆ, f˜ is a finite rank operator of ω.
We list the parameters xi , ti , Ri in the above approximations in Appendix C.2. To obtain
sharp estimate of the constants in (2.77) and u − û, ∇u − ∇u d in the energy estimates, which
are important for us to reduce the number of approximation terms and obtain sharp energy
estimates, we will estimate the integrals with computer assistance. We will discuss them in
details in Section 8.
We remark that we do not need to construct too many approximaton terms. The total number
of approximation terms we choose is less than 50.
32 JIAJIE CHEN AND THOMAS Y. HOU

3. Sharp Hölder estimate via optimal transport

In this section, we derive the sharp Hölder C 1/2 estimate for ∇u using the symmetry proper-
ties of the kernels and some techniques from optimal transport. We note that novel functional
inequalities on similar Biot-Savart laws have played a crucial role in the important works [27,48].
Those estimates enable the authors to control the velocity effectively. The sharp Hölder esti-
mates play a similar role in our work.
To obtain the Hölder estimate of ∇u in R++ 2 , the natural approach is to estimate ∇u(x) −
∇u(z) for all pairs x, z ∈ R2++ , which has a dimension of 4. It is very difficult to obtain a sharp
estimate since the kernel in ∇u(x) − ∇u(z) for arbitrary x, z has a complicated sign structure
and destroys some symmetry properties of the kernels in ∇∇(−∆)−1 . Instead, we will estimate
1/2 1/2
the Cx and Cy seminorms (2.3) due to the following important observations. Firstly, the
linearized operators (2.17), (2.21) are anisotropic in x and y. See Section 2.7.2. We have much
larger damping factors along the y direction. Therefore, a sharp Hölder estimate of ∇u in the
1/2 1/2
x direction, i.e. [∇u]C 1/2 is much more important. Secondly, if we estimate the Cx or Cy
x
seminorm (2.3), where we assume x1 = z1 or x2 = z2 , we reduce the dimension of (x, z) from 4
to 3. Moreover, the kernel in ∇u(x) − ∇u(z) enjoys better symmetry properties and the sign
properties are much simpler. These properties allow us to reduce estimating the 2D integral
into estimating many 1D integrals. After we estimate [∇u]C 1/2 , [∇u]C 1/2 , using the triangle
x y
inequality, we can obtain the estimate of ||∇u||C 1/2 .
The kernels associated with ∇u are given by
y1 y2 1 y12 − y22 1
(3.1) K1 (y) , , K2 (y) , , G(y) = − log |y|,
|y|4 2 |y|4 2
where π1 G is the Green function of −∆ in R2 . Note that ∂1 ∂2 G = K1 , ∂12 G = K2 .
Denote by Ki,s the symmetrized kernel
(3.2) Ki,s (x, y) , Ki (x − y) − Ki (x1 + y1 , x2 − y2 ) − Ki (x1 − y1 , x2 + y2 ) + Ki (x + y).
Consider the odd extension of ω in y from R+
2 to R2

(3.3) W (y) = ω(y) for y2 ≥ 0, W (y) = −ω(y1 , −y2 ) for y2 < 0.

W is odd in both y1 and y2 variables. Clearly, ux can be written as
Z Z
1 1
ux (x) = − K1 (x − y)W (y)dy = − ω(y)K1,s (x, y)dy.
π R2 π R++
2

For any a, b1 , b2 > 0, we consider the localized velocity

2(x1 − y1 )(x2 − y2 )
Z
1
ux (x, a, b1 , b2 ) , − P.V. ω(y)dy,
2π |x1 −y1 |≤a, −b1 ≤x2 −y2 ≤b2 |x − y|2
(x1 − y1 )2 − (x2 − y2 )2
Z
1 ω(x)
(3.4) uy (x, a) , P.V. 4
ω(y)dy + ,
2π |x1 −y1 |≤a,|x2 −y2 |≤a |x − y| 2
(x1 − y1 )2 − (x2 − y2 )2
Z
1 ω(x)
vx (x, a) , P.V. 4
ω(y)dy − .
2π |x1 −y1 |≤a,|x2 −y2 |≤a |x − y| 2
In particular, if b1 = b2 = b, we simplify ux (x, a, b1 , b2 ) as ux (x, a, b); if b1 = b2 = a, we further
simplify ux (x, a, b1 , b2 ) as ux (x, a).
Denote by dx (f, x, z), dy (f, x, z) the Hölder difference
|f (x) − f (z)| |f (x) − f (z)|
(3.5) dx (f, x, z) , , dy (f, x, z) , .
|x1 − z1 |1/2 |x2 − z2 |1/2
3.1. The Hölder estimates of the velocity. We have the following important estimates for
∇u. We will discuss the ideas in Section 3.2 and the proof in Sections 3.3, 3.4, and Appendix B.
b
We localize the velocity in (3.4) to obtain improvement of the constant C1 ( |x−z| ) when |x − z|
is large.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 33

Lemma 3.1 (Estimate of ||ux ||C 1/2 ). For any b1 , b2 > 0, a ≥ 12 |x1 − z1 |, x = (x1 , x2 ), z =
x
(z1 , x2 ) ∈ R+
2 , we have
 
|ux (x, a, b1 , b2 ) − ux (z, a, b1 , b2 )| b
≤ C 1 ||ω||C 1/2 ,
|x − z|1/2 |x − z| x

where b = max(b1 , b2 ) and C1 (a) is an increasing function given by

Z b Z ∞
4
C1 (b) = ds2 |T (s1 , s2 ) − s1 |1/2 ∆(s1 , s2 )ds1 ,
π 0 f (s2 )
(s1 + 1/2)s2 (s1 − 1/2)s2
∆(s1 , s2 ) = − .
((s1 + 1/2)2 + s22 )2 ((s1 − 1/2)2 + s22 )2
Here, f (s2 ) is the unique solution in [0, ∞) satisfying ∆(f (s2 ), s2 ) = 0 and T (s1 , s2 ) is the
unique solution in [0, f (s2 )) that solves
Z s1
∆(s1 , s2 )ds1 = 0,
T (s1 ,s2 )

for s1 > f (s2 ). In particular, T (s1 , s2 ) can be obtained explicitly by solving a cubic equation and
C1 (b) ≤ 2.55 for any b > 0.
If a < 21 |x1 − z1 |, the singular region is small. We can simply apply the triangle inequality to
estimate each term.
Remark 3.2. The above Lemma can be further generalized to the localized velocity
ω ∗ K1 (s)1−a1 ≤s≤a2 ,−b1 ≤s2 ≤b2 , i.e., we do not need a1 = a2 in (3.4). The proof follows from the
same argument. Yet, we will only use the special case a1 = a2 in our later estimates.
The upper bounds in the following Lemmas involve [ω]C 1/2 and [ω]C 1/2 . We will further
x y
bound it using the energy norm. To obtain the sharp estimate, we will optimize the choice of τ .
Lemma 3.3 ( Estimate of ||ux ||C 1/2 ). For any a, b ≥ 1, x = (x1 , x2 ), z = (x1 , z2 ) ∈ R+
2 , we have
y

|ux (x, a, b) − ux (z, a, b)| 1 a a

≤ C1 ( )||ω||C 1/2 + C2 ( )||ω||C 1/2 ,
|x − z|1/2 2 |x − z| y |x − z| x

where C1 (a) is defined in the previous Lemma and C2 (a) is given by

√ Z aZ ∞
2 1/2 y1 (1/2 − y2 ) y1 (1/2 + y2 )
C2 (a) = y1 2 2 2
+ 2 dy.
π 0 0 (y1 + (1/2 − y2 ) ) (y1 + (1/2 + y2 )2 )2
4.26
In particular, C2 (a) ≤ π .

Next we estimate the other kernel. We remark that for uy (x, a) and vx (y, a), the estimates
are different due to the local term related to ω (3.4).
1/2
Lemma 3.4 (Cx estimate of uy , vx ). For any a ≥ 1, x = (x1 , x2 ), z = (z1 , x2 ) ∈ R+
2 and
τ > 0, we have
|vx (x, a) − vx (z, a)| 1 1/2
1/2
≤ C1 (τ ) max(||ω||C 1/2 , τ −1 ||ω||Cy ),
|x − z| π x

|uy (x, a) − uy (z, a)| 1 1/2

1/2
≤ C2 (τ ) max(||ω||C 1/2 , τ −1 ||ω||Cy ),
|x − z| π x

for some constant C1 (τ ), C2 (τ ) > 0, where C1 (0.589) = 7.94 and C2 (0.589) = 5.32. The above
estimates can be improved for a smaller window size a and the corresponding constant can be
computed.
In the proof of the above Lemma, we provide the upper bounds for C1 (τ ), C2 (τ ), which can
be computed. Although the estimates are equivalent for different τ , we choose τ according to
the weight g1 in Hölder seminorm [ωψ1 ]C 1/2 . In practice, we choose τ = g1 (0, 1)/g1 (1, 0) which
g1
is close to 0.589.
34 JIAJIE CHEN AND THOMAS Y. HOU

1/2
In general, the localized Cy estimate does not hold for uy due to the presence of the boundary
and the discontinuity of W cross y = 0. Thus, we consider the estimate without localizing the
kernel.
1/2
Lemma 3.5 ( Cy estimate of uy , vx ). For x = (x1 , x2 ), z = (x1 , z2 ) ∈ R+
2 and any τ > 0, we
have
|vx (x, ∞) − vx (z, ∞)| C3 (τ )
1/2
≤ max(||ω||C 1/2 , τ −1 ||ω||C 1/2 ),
|x − z| π x y

|uy (x, ∞) − uy (z, ∞)| C4 (τ )

≤ max(||ω||C 1/2 , τ −1 ||ω||C 1/2 ),
|x − z|1/2 π x y

for some constant C3 (τ ), C4 (τ ) > 0. We have C3 (0.589) = 8.14, C4 (0.589) = 8.14.

3.2. Connection to optimal transport and ideas of the proof. A key observation is that
the Hölder estimate is related to an optimal transport problem. We illustrate the ideas by
proving a sharp Hölder estimate of the Hilbert transform. The Hilbert transform can be seen
as an approximation of ux (ω), which is exact if ω(x, y) is constant in y [15, 54].
1
We estimate |x−z| 1/2 |Hf (x) − Hf (z)| by ||f ||Ċ 1/2 . Due to translation and scaling symmetry,

we can assume x = 1, z = −1 without loss of generality. Then we need to estimate

Z Z
1 1 1 2 1
(3.6) S = Hf (1) − Hf (−1) = ( + )f (y)dy = f (y)dy.
π R 1−y 1+y π R 1 − y2
2
R
The kernel k(y) = 1−y 2 is positive on (−1, 1) and negative for |y| > 1, and satisfies k(y)dy = 0.
Denote k ± (y) = max(±k(y), 0). An estimate of S using sup |f|x−z| (x)−f (z)|
1/2 is equivalent to esti-
mating the transportation cost of moving the positive region of k(y) with measure k + (y)dy to
its negative region with measure k − (y)dy with distant function c(x, y) = |x − y|1/2 .
For example, if k(y) = δ1 (y) + δ2 (y) − δ3 (y) − δ4 (y), where δa (x) is the Dirac function centered
at a, then we get
Z √ √
| k(y)f (y)dy| = |f (1) + f (2) − f (3) − f (4)| ≤ |f (2) − f (3)| + |f (1) − f (4)| ≤ ( 1 + 3)||f ||C 1/2 .

The above estimate can be interpreted as moving the mass from 2 to 3 and 1 to 4 with
cost function |x − y|1/2 ||f ||C 1/2 . Using the language of optimal transport, to obtain sharp
estimate of S (3.6), we are seeking a measurable map T such that T #k+ dy = k− dy, where
(T #µ)(A) = µ(T (A)) for a measurable set A, and the following cost
Z
C(T ) = |T (y) − y|1/2 k(y)dy||f ||C 1/2
k(y)≥0
is as small as possible. Based on the above discussion, we have the following transportation
lemma, which will be used repeatedly in the Hölder estimate.
Lemma 3.6 (Transportation Lemma). Suppose that there exists c ∈ (a, b) such that f < 0 on
Rb 1/2
(a, c), f > 0 on (c, b), f |x − c|1/2 ∈ L1loc with a f (x)dx = 0 . For g ∈ Cx (a, b), we have
Z b Z b Z c
f (x)g(x)dx ≤ |f (x)||x − T (x)|1/2 dx[g]C 1/2 = |f (x)||x − T (x)|1/2 dx[g]C 1/2 ,
x x
a c a
R T (x)
where T (x) solves f (s)ds = 0.
Rx
The estimate of f g in terms of [g]C 1/2 can be interpreted as transporting the negative part
x
of f to its positive part with distant function c(x, y) = |x − y|1/2 .
Proof. Firstly, we want to understand how to construct the map T . Note that for x1 < x2 <
x3 < x4 , we have
Z
| (δx1 + δx2 − δx3 − δx4 )g(x)dx| = |g(x1 ) + g(x2 ) − g(x3 ) − g(x4 )|

≤ min(|x1 − x3 |1/2 + |x2 − x4 |1/2 , |x1 − x4 |1/2 + |x2 − x3 |1/2 )[g]C 1/2
x

=(|x1 − x4 |1/2 + |x2 − x3 |1/2 )[g]C 1/2 .

x
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 35

The above estimate indicates that to find an optimal map T moving (δx1 +δx2 )dx to (δx3 +δx4 )dx
with cost |x − y|1/2 , we should choose T (x1 ) = x4 , T (x2 ) = x3 , which implies that T (x) is
decreasing in x. Due to conservation of mass and the sign properties of f , a natural construction
of T : (a, c) → (c, b) is given by
Z T (x)
(3.7) f (x)dx = 0,
x
0
for x < c, which implies T (x)f (T (x)) = f (x) for smooth f . The idea of the above map is to
move the mass in the positive region to its closest negative region that has not been occupied
due to the monotonicity of T . Using a change of variable y = T (x) : (a, c) → (c, b], we get
Z b Z c Z c
0
f (x)g(x)dx = − f (T (x))g(T (x))T (x)dx = − f (x)g(T (x))dx.
c a a
It follows
Z b Z c Z c
| f g| = | f (x)(g(T (x)) − g(x))dx| ≤ |f (x)||T (x) − x|1/2 dx[g]C 1/2 .
x
a a a
Rx
Similarly, we can define T : (c, b) → (a, c) by solving T (x) f (s)ds = 0, which is also equivalent
to (3.7). The first inequality in Lemma 3.6 follows from the same argument. 
3.2.1. C 1/2 estimate of the Hilbert transform. Now, we apply Lemma 3.6 with f = k(y), g = f (y)
to estimate (3.6). For any y > 0, we construct the transportation map T (y) by solving
Z T (x) Z T (x)
1
0= k(y)dy = dy = 0,
x x 1 − y2
which implies
x+1 T +1 1
= , T (x) = ,
1−x T −1 x
1
where we have used T (x) > 1 if x < 1 and T (x) < 1 if x > 1 due to the sign of 1−y 2 . This map

also applies to y < 0. Using this map, we yield

Z Z
2 2
|S| = | k(y)(f (y) − f (T (y)))dy + k(y)(f (y) − f (T (y)))dy|
π y>1 π y<−1
|y 2 − 1|1/2
Z Z
4 1/2 4 1
≤ |k(y)||y − T (y)| dy||f ||C 1/2 = dy||f ||C 1/2
π y>1 π y>1 |y 2 − 1| |y|1/2
Z
4 1
= ||f ||C 1/2 dy = C||f ||C 1/2 .
π y>1 |y 2 − 1|1/2 y 1/2
The above formula is an example of the derivation in Lemma 3.6. The constant C can be
estimated. It follows
S C C 7.4
(3.8) 1/2
≤ √ ||f ||C 1/2 , √ ≈ ≈ 2.37.
|x − z| 2 2 π
C
Since x, z are arbitrary, we yield ||Hf ||C 1/2 ≤ √
2
||f ||C 1/2 .

Remark 3.7. Note that the Hilbert transform satisfies H(Hf ) = −f . It implies that the sharp
constant in ||Hf ||C 1/2 ≤ C||f ||C 1/2 satisfies C ≥ 1. Thus, the above estimate does not overesti-
mate too much.
In the following subsections, we prove Lemmas 3.1, 3.3 for ux , which is the most singular
nonlocal term in (2.23). The proofs of Lemmas 3.4, 3.5 are similar but technical, which are
deferred to Appendix B.
To apply Lemma 3.6 to the Hölder estimate of ∇u, which involves 2D integrals, we need two
steps. Firstly, we identify the sign of the kernel K in the integral of ∇u(x) − ∇u(z). Next,
we fix a variable in the 2D integral in one direction, e.g. fix x = a, and then apply Lemma
3.6 to estimate the 1D integral in the other direction, e.g., on the line {(a, y) : y ∈ R}. One
may generalize Lemma 3.6 to 2D and construct the 2D optimal transport map directly. Yet,
36 JIAJIE CHEN AND THOMAS Y. HOU

the domain where the kernel K is positive or negative is complicated. To avoid this difficulty,
we build the 2D transport map using the 1D Lemma 3.6 repeatedly. The odd symmetry of the
kernel K1 (s) in s1 enables us to apply this approach to obtain sharp estimate of ux effectively.
See Remark 3.8.
1/2
3.3. Estimate of [ux ]C 1/2 . In the Cx estimate of ux , we have x2 = z2 . In the case without
x
localization of the kernel, using the scaling symmetry and translation invariance, we only need
to estimate the following
Z
1 1 1
(3.9) ux ( , x2 ) − ux (− , x2 ) = − P.V. K(s)W (s1 , x2 − s2 )ds
2 2 π R2

for any x2 , where W is an odd extension of ω from R+ 2

2 to R , and K(s) is given by

1 1 (s1 + 21 )s2 (s1 − 21 )s2

(3.10) K(s) = K1 (s1 + , s2 ) − K1 (s1 − , s2 ) = − .
2 2 ((s1 + 21 )2 + s22 )2 ((s1 − 21 )2 + s22 )2
Since K(s) is odd in s2 , we consider s2 ≥ 0 without loss of generality. We will only use
the Hölder seminorm of W , [W ]C 1/2 , to estimate the above quantity. Note that [W ]C 1/2 (R2 ) =
x x
[ω]C 1/2 (R2 ) . Without loss of generality, we can assume that x2 = 0.
x +
A direct calculation yields
s2 ∆1 (s1 , s2 ) 1 1 1
K(s) = 1 2 , ∆1 = s42 − 2s21 s22 − 3s41 + s21 + s22 + .
((s1 + 2) + s22 )2 ((s1 − 12 )2 + s22 )2 2 2 16
For fixed s2 , ∆1 (s1 , s2 ) = 0 implies
 1 − 2s2 + p16s4 + 4s2 + 1 1/2
2 2 2
(3.11) s1 = f (s2 ) = 2 .
6
Moreover, for s1 , s2 ≥ 0, it is easy to see that ∆1 (s1 , s2 ) ≥ 0 if and only if
(3.12) K(s1 , s2 ) ≥ 0 for s1 ∈ [0, f (s2 )], K(s1 , s2 ) ≤ 0 for s1 ≥ f (s2 ).
Note that the sign changes if s2 ≤ 0 since K is odd in s2 . Since K1 is odd in s1 , we get
Z ∞ Z ∞
1 1
K(s1 , s2 )ds1 = K1 (s1 + , s2 ) − K1 (s1 − , s2 )ds1
0 0 2 2
Z ∞ Z ∞ Z 1/2
= K1 (s1 , s2 )ds1 − K(s1 , s2 )ds1 = − K1 (s1 , s2 )ds1 = 0.
1/2 −1/2 −1/2

To estimate the integral in (3.9), we first fix s2 and then apply Lemma 3.6 to estimate
(3.13) Z Z Z
I(s2 ) = K(s1 , s2 )W (s1 , x2 −s2 )ds1 = ( + )K(s1 , s2 )W (s1 , x2 −s2 )ds1 , I− (s2 )+I+ (s2 ).
R R− R+

We do so for the following reason. Near the singularity, from Taylor expansion of (3.11):
s1 = 21 + O(s42 ), the curve Γ separating the positive and negative region of K(s) is close to a
straight line in the vertical direction. Similar to the idea below (3.7), an effective plan in 2D
should move the mass in the positive region to its closest possible negative region that has not
been occupied. Based on this idea, we expect that an effective 2D transport plan (x, y) → T (x, y)
is orthogonal to the curve Γ and thus almost parallel to the x direction.
Remark 3.8. The fact that near the singularity s = (± 12 , 0), the curve Γ is almost vertical is
due to the odd symmetry of K1 (s1 , s2 ) in s1 . In fact, from (3.10), for s close to ( 21 , 0), we have
K(s) ≈ −K1 (s1 − 12 , s2 ), whose sign is determined by sgn(s1 − 21 ).
Since K(s1 , s2 ) is even in s1 , we can estimate I+ (s2 ), I− (s2 ) in the same way. To apply
Lemma 3.6, we first construct T (·, s2 ) on [0, ∞) by solving
Z s1
(3.14) K(t, s2 )dt = 0.
T (s1 ,s2 )
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 37

We will show later that this equation has a unique solution of T on [0, ∞] for s1 > 0. Then
applying Lemma 3.6 to I+ (s2 ) and using [W ]C 1/2 = [ω]C 1/2 , we get
x x
Z ∞
|I+ (s2 )| ≤ [ω]C 1/2 K(s1 , s2 )|T (s1 , s2 ) − s1 |1/2 ds1 , M (s2 ).
x
f (s2 )

Since K(s1 , s2 ) is even in s1 , the estimate of I− (s2 ) in (3.13) is the same: I− (s2 ) ≤ M (s2 ).
Since K(s) is odd in s2 , from (3.14), we get T (s1 , −s2 ) = T (s1 , s2 ). Therefore, the estimate of
I(s1 , s2 ) is the same as I(s1 , −s2 ): |I(s1 , s2 )| ≤ 2M (|s2 |). Integrating the estimate of I(s) over
s2 , we yield
Z Z ∞ Z ∞
1 4
− P.V. K(s)W (s1 , x2 − s2 )ds ≤ [ω]C 1/2 ds2 K(s1 , s2 )|T (s1 , s2 ) − s1 |1/2 ds1 ,
π R2 π x
0 f (s2 )

which along with (3.9) prove Lemma 3.1 in the case of a = b1 = b2 = ∞.

xy 1 y
3.3.1. Formula of T . From (3.10) and (x2 +y 2 )2 = − 2 ∂x x2 +y 2 , equation (3.14) is equivalent to

 1 1   1 1 
s2 − = s2 −
(T + 12 )2 + s22 (T − 12 )2 + s22 (s1 + 12 )2 + s22 (s1 − 12 )2 + s22
where we have simplified T (s1 , s2 ) as T . For s2 6= 0, expanding the identity yields
(3.15) 0 = −1 − 8T s1 + 16T s1 (T 2 + T s1 + s21 ) − 8s22 (1 − 4s1 T + 2s22 ).
The above equation is cubic in T , and thus can be solved explicitly. In Appendix B.3, for
fixed s2 and s1 > 0, we show that it has a unique solution on [[0, ∞) using the discriminant
of the cubic solution. We also study the properties of T (s1 , s2 ), which allows us to estimate a
sharp bound for C1 (b) in Lemma 3.1.
Remark 3.9. In the special case where ω(x, y) is constant in y, we have ux (ω)(x, 0) = Hω(x),
which has been observed in [15, 54]. Thus, the optimal constant in Lemma 3.1 must be larger
√ transform (3.8). Here, the upper C1 (b) ≤ 2.55 is very close to that the
than that of the Hilbert
Hilbert transform C/ 2 ≈ 2.37 (3.8), which reflects the effectiveness of the optimal transport
approach to estimate the sharp constant.
3.3.2. Localized estimate of ux . Next, we estimate ux (x, a, b1 , b2 ) − ux (z, a, b1 , b2 ) with x2 = z2
using [W ]C 1/2 . The estimate consists of following steps. Firstly, we identify the sign of the kernel
x
similar to those between (3.9) and (3.11). Secondly, we construct the transportation map along
the x direction and derive the transportation cost. Thirdly, we compare the transportation cost
in the case of kernel localization with that without kernel localization using the properties of
the transportation maps, and show that the cost with kernel localization is smaller than that
without using kernel localization.
Without loss of generality, we assume x1 = 21 , z1 = − 12 , x2 = 0. Denote
Ia , [−a, a], Ib = [−b1 , b2 ], Q , Ia × Ib , b = max(b1 , b2 ),
1
Since we assume a ≥ 2 |x1 − z1 | in Lemma 3.1, we have
(3.16) a ≥ 1/2.
The kernel associated with ux ( 21 , x2 ) − ux (− 21 , x2 ) (3.4) becomes
(s1 + 21 )s2
 (s1 − 12 )s2 
Ka,b (s1 , s2 ) =1s2 ∈Ib 1 1 − 1 1
((s1 + 21 )2 + s22 )2 s1 + 2 ∈Ia ((s1 − 21 )2 + s22 )2 s1 − 2 ∈Ia
(3.17)
=((K1 (s1 + 1/2, s2 ) − K1 (s1 − 1/2, s2 ))1s+1/2∈Q
− K1 (s1 − 1/2, s2 )(1s−1/2∈Q − 1s1 +1/2∈Q )).
1
Since a ≥ 2 and K1 (s) is odd in s1 , for fixed s2 , we have
Z ∞ Z a Z a Z 1/2 Z 0
Ka,b (s1 , s2 )ds1 = ( − )K1 (s)ds1 = − K1 (s)ds1 = 0, Ka,b (s1 , s2 )ds1 = 0.
0 1/2 −1/2 −1/2 −∞
38 JIAJIE CHEN AND THOMAS Y. HOU

Similar to the case without localization, for each s2 , we consider the transportation from the
positive part of Ka,b to its negative part. Firstly, we identify the sign of Ka,b . We restrict to
s2 ∈ [−b1 , b2 ] and s2 6= 0 since otherwise Ka,b = 0. We focus on s1 , s2 ≥ 0 and the estimate for
s1 < 0 or s2 < 0 is the same. Since a > 12 , we always have

(3.18) s1 ± 1/2 > −a, for s1 ≥ 0.

Thus, for s1 ≥ 0, we can neglect the constraint s1 ± 21 ≥ −a in the localization in (3.17).

Case 1: a ∈ (1/2, 1]. Clearly, Ka,b (s1 , s2 ) > 0 for s1 < 21 since both kernels in (3.17) are
non-negative. For s1 ≥ 12 , since a ≤ 1, we get

1
Ka,b (s1 , s2 ) = −K1 (s1 − , s2 )1s−1/2∈Q ≤ 0.
2
In this case, we denote sc (s2 ) = 12 .
Case 2: a ∈ (1, f (s2 ) + 12 ). Recall f (s2 ) from (3.11). For s1 > a − 1
2 > 12 , we get

1
Ka,b = −K1 (s1 − , s2 )1s−1/2∈Q ≤ 0.
2
1 1
For s1 ≤ a − 2 < f (s2 ), using (3.12) and s1 + 2 ≤ a, we obtain

1 1
Ka,b = K1 (s1 + , s2 ) − K1 (s1 − , s2 ) ≥ 0.
2 2
We denote sc (s2 ) = a − 21 .
Case 3: a ≥ f (s2 ) + 12 . For s1 < f (s2 ), using (3.12) and s1 ± 1
2 < a, we get

1 1
Ka,b = K1 (s1 + , s2 ) − K1 (s1 − , s2 ) ≥ 0.
2 2
For s1 ≥ f (s2 ) > 12 , since K1 (s1 + 12 , s2 ) − K1 (s1 − 21 , s2 ) ≤ 0 (3.12) and

1s−1/2∈Q − 1s+1/2∈Q = 1s1 −1/2≤a − 1s+1/2≤a ≥ 0,

we get
Ka,b ≤ −K1 (s1 − 1/2, s2 )(1s−1/2∈Q − 1s+1/2∈Q )) ≤ 0.
We denote sc (s2 ) = f (s2 ). In summary, for fixed s2 , we define
1 1 1 1
sc (s2 ) = , if a ∈ ( , 1], sc (s2 ) = a − , if a ∈ (1, f (s2 ) + ),
(3.19) 2 2 2 2
sc (s2 ) = f (s2 ), if a ≥ f (s2 ) + 1/2,

which satisfies

(3.20) Ka,b (s1 , s2 ) ≥ 0, s1 ∈ [0, sc ], Ka,b (s1 , s2 ) ≤ 0, s1 ∈ [sc , ∞], sc (s2 ) ≤ f (s2 ),
1
where the last inequality follows from the definition of sc and f (s2 ) ≥ 2 (3.11).
In each case i = 1, 2, 3, we construct the transport map by solving
Z s1
1
(3.21) Ka,b (x, s2 )dx = 0, Ti ≤ a + .
Ti (s1 ,s2 ) 2

We add the restriction Ti ≤ a + 12 since Ka,b (s) = 0 for s1 > a + 21 by definition (3.17).
Applying Lemma 3.6 in the s1 direction and using Ka,b (s) = 0 for |s2 | ≥ b, we yield
Z Z b Z sc (s2 )
Ii , Ka,b (s1 , s2 )ω(s1 , −s2 )ds ≤ |Ka,b (s)||Ti (s) − s1 |1/2 ds · [ω]C 1/2 .
x
s1 ≥0,s2 ≥0 0 0
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 39

3.3.3. Comparison of the cost. Next, we show that the cost can be bounded uniformly by the
cost of the case without localization
Z bZ ∞
(3.22) Ii ≤ |K(s)||T (s) − s1 |1/2 ds · [ω]C 1/2 ,
x
0 f (s2 )

where T is defined in (3.14). It suffices to prove

Z sc (s2 ) Z f (s2 )
1/2
(3.23) Ji , |Ka,b (s)||Ti (s) − s1 | ds ≤ |K(s)||T (s) − s1 |1/2 ds
0 0

for any s2 . We focus on |s2 | ≤ a and s2 6= 0. The intuition behind the above inequality is
that if the mass is localized, we should get “cheaper” transportation cost than the case without
localization since the transportation distance is shorter. To justify these heuristics, we compare
the kernels and will prove
(3.24) |Ka,b (s)| ≤ |K(s)|, s1 ∈ [0, sc (s2 )],
and use (3.14) and (3.21) to compare Ti and T
(3.25) s1 ≤ sc (s2 ) ≤ Ti (s) ≤ T (s), s1 ∈ [0, sc (s2 )]
and thus Ti (s) − s1 ≤ T (s) − s1 . Clearly, inequality (3.23) follows from (3.24) and (3.25).
Compare the kernels. From (3.20) and (3.12), since sc (s2 ) ≤ f (s2 ), we get Ka,b (s), K(s) ≥ 0
for s1 ∈ [0, sc (s2 )]. Hence, for fixed s2 ∈ [−b1 , b2 ], (3.24) is equivalent to
1 1
0 ≤ K(s) − Ka,b (s) = K1 (s1 + , s2 )(1 − 1s1 +1/2∈Ia ) − K1 (s1 − , s2 )(1 − 1s1 −1/2∈Ia ) , I.
2 2
From the definition of (3.19) and (3.18), for s1 ∈ [0, sc (s2 )], we have
s1 ± 1/2 ≥ −a, s1 − 1/2 ≤ a, 1 − 1s1 −1/2∈Ia = 0,
which along with K1 (s1 + 21 , s2 ) ≥ 0 (3.17) implies (3.24)
1
I = K1 (s1 + , s2 )(1 − 1s1 +1/2∈Ia ) ≥ 0.
2
Remark 3.10. In the above derivations, we consider s2 ≥ 0. If s2 ≤ 0, one needs to track the
sign to prove inequality (3.24).

Compare the maps. To prove (3.25), our idea is to use the equations (3.14), (3.21) and the
sign of the kernels Ka,b , K to compare Ti and T .
We fix s2 > 0 in the following derivations. To simplify the notation, we simplify T (s1 , s2 )
as T (s1 ) in some places. Since Ti , T (3.14), (3.21) are decreasing and sc (s2 ) is a fixed point for
Ti (·, s2 ), for s1 ≤ sc (s2 ), we get
(3.26) Ti (s1 , s2 ) ≥ Ti (sc (s2 ), s2 ) = sc (s2 ), T (s1 , s2 ) ≥ T (f (s2 ), s2 ) = f (s2 ) ≥ sc (s2 ).
Moreover, from (3.14), (3.21), we have
(3.27) Ti (Ti (s1 )) = s1 , T (T (s1 )) = s1 .
Denote
1 1 − 1
(3.28) K ± = K(s1 ± , s2 ), +
Ka,b = K1 (s1 + , s2 )1s1 + 21 ≤a , Ka,b = K1 (s1 − , s2 )1s1 − 12 ≤a .
2 2 2
We remark that K − is not non-negative but K + is positive. By definition, we have
−
+
Ka,b = Ka,b − Ka,b , K = K + − K −.
(3.29)
K + (s) ≥ 0, s1 , s2 ≥ 0, K − (s) ≥ 0, s1 ≥ 1/2, s2 ≥ 0.
Next, we study each case in the order of 3, 2, 1 to prove (3.25).
40 JIAJIE CHEN AND THOMAS Y. HOU

Case 3: a ≥ f (s2 ) + 12 . In this case, recall sc (s2 ) = f (s2 ) from (3.19).

For s1 ≤ a − 21 , we get Ka,b = K (3.17). Hence, equations (3.14) and (3.21) are the same for
s1 ≤ a − 1/2, and we get
1 1
(3.30) s1 ∈ [T (a − ), a − ].
T3 (s1 , s2 ) = T (s1 , s2 ),
2 2
It follows (3.25) for s1 ∈ [T (a − 1/2), f (s2 )]. We recall that from (3.26), a − 1/2 ≥ f (s2 ) and
T (a − 1/2) = T3 (a − 1/2), we have
(3.31) T (a − 1/2) ≤ T (f (s2 )) = f (s2 ) ≤ a − 1/2, T (s1 ), T3 (s1 ) ≥ a − 1/2, s1 ≤ T (a − 1/2).
Next, we compare T (s1 ), T3 (s1 ) for s1 < T (a − 1/2) ≤ f (s2 ). From (3.14),(3.21), and T (a −
1/2) = T3 (a − 1/2) ≤ a − 1/2, we have
Z a−1/2 Z a−1/2 Z a−1/2
K(s)ds1 = Ka,b (s)ds1 = Ka,b (s)ds1 = 0.
T (a−1/2) T3 (a−1/2) T (a−1/2)

Moreover, from (3.17) and (3.21), we have

−
Ka,b (t, s2 ) = −Ka,b (t, s2 ) = −K − (t, s2 ), t ∈ [a − 1/2, T3 (s1 )] ⊂ [1 − 1/2, a + 1/2],
Ka,b (t, s2 ) = K(t, s2 ), t ≤ T (a − 1/2) ≤ a − 1/2.
Plugging the above identities in (3.14), (3.21) for s1 ≤ T (a − 1/2), we yield
Z T (s1 ) Z T (a−1/2) Z T
0= K(t, s2 )dt = K(t, s2 ) + K(t, s2 )dt,
s1 s1 a−1/2
Z T3 (s1 ) Z T (a−1/2) Z T3 (s1 )
0= Ka,b (t, s2 )dt = K(t, s2 ) − K − (t, s2 )dt.
s1 s1 a−1/2

Note that K = K + − K − . Calculating the difference between the two identities yields
Z T (s1 ) Z T3 (s1 ) Z T (s1 ) Z T3 (s1 )
0= (K + − K − )ds + K− = K + ds1 + K − ds1 .
a−1/2 a−1/2 a−1/2 T (s1 )

Recall s2 6= 0 and from (3.31), we obtain T3 (s1 ), T (s1 ) ≥ a − 1/2 ≥ 1/2. From (3.29), we
yield K + > 0 and K− > 0 for s1 ≥ min(a − 1/2, T3 , T ). Since T is decreasing and
T (s1 ) ≥ T (T (a − 1/2)) = a − 1/2, s1 ≤ T (a − 1/2),
the first integral is non-negative. We prove T (s1 ) ≥ T3 (s1 ) for s1 ≤ T (a − 1/2), which along
with (3.30) implies (3.25).
The proof in the case 1,2 is completely similar.
Case 2: a ∈ (1, f (s2 ) + 21 ). Recall sc (s2 ) = a − 21 ≤ f (s2 ) from (3.19) and (3.26). For any
s1 ≤ a − 21 ≤ f (s2 ), using (3.14), (3.21), and an argument similar to that in case 3, we yield
Z T2 (s1 ) Z a−1/2 Z T2 (s1 )
0= Ka,b (t, s2 )dt = K(t, s2 )dt − K − (t, s2 )dt,
s1 s1 a−1/2
Z T (s1 ) Z a−1/2 Z T (s1 )
0= K(t, s2 )dt = K(t, s2 )dt + (K + − K − )(t, s2 )dt,
s1 s1 a−1/2
−
where we have used Ka,b (t, s2 ) = −K (t, s2 ) for t ≥ a−1/2 (3.17), (3.28) in the second equality.
Comparing the difference between two identities yields
Z T (s1 ) Z T2 (s1 ) Z T (s1 ) Z T2 (s1 )
0= K + − K − (t, s2 )dt + K − (t, s2 )dt = K + (t, s2 )dt + K − (t, s2 )dt.
a−1/2 a−1/2 a−1/2 T (s1 )

Recall from (3.26) that T (s1 ), T2 (s1 ) ≥ sc (s2 ) = a − 1/2 for s1 ≤ a − 1/2. For s2 6= 0
and s1 ≥ min(T, T2 , a − 1/2) = a − 1/2 > 1/2, we have K − > 0, K + > 0 (3.29). We obtain
T (s1 ) ≥ T2 (s1 ), which implies (3.25).
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 41

1
Case 1: a ∈ (1/2, 1]. In this case, sc (s2 ) = 2 < f (s2 ). From (3.17), we yield
0 ≤ Ka,b = 1s1 +1/2≤a K1 (s1 + 1/2, s2 ) − K1 (s1 − 1/2, s2 )
≤ K1 (s1 + 1/2, s2 ) − K1 (s1 − 1/2, s2 ) = K, s1 ∈ [0, 1/2],
− −
Ka,b = −K1 (s1 − 1/2, s2 ) = −K (s1 , s2 ), K (s1 , s2 ) ≥ 0, s1 ∈ [1/2, a + 1/2].
For any s1 < 21 and s2 6= 0, from (3.26), we get T1 (s1 ) ≥ 1/2, T (s1 ) ≥ f (s2 ) > 1/2. Using
(3.14), (3.21) and the above estimates for Ka,b , we yield
Z 1/2 Z T1 (s1 ) Z 1/2 Z T
−
0= Ka,b (t, s2 )dt − K (t, s2 )dt = K(t, s2 )dt + (K + − K − )(t, s2 )dt.
s1 1/2 s1 1/2
It follows
Z 1/2 Z T Z T1
0= (K − Ka,b )(t, s2 )dt + K + +
K − (t, s2 )dt , II1 + II2 + II3 .
s1 1/2 T

From (3.24) and Ka,b , K > 0 on t ∈ [0, sc (s2 )] = [0, 1/2], we get II1 ≥ 0. Note that
K − , K + > 0 for s1 > 1/2 (3.12), (3.20). Since T1 , T > 1/2, we must obtain T (s1 ) ≥ T1 (s1 ),
which implies (3.25).
We have proved (3.25) in all three cases, which implies |T (s) − s1 | ≥ |Ti (s) − s1 |. Combining
this estimate and (3.24), we prove (3.22) and conclude the proof of Lemma 3.1.
3.4. Estimate of [ux ]C 1/2 . Firstly, we note that b1 = b2 = b and x1 = z1 in this case.
y
Without loss of generality, we assume z2 = m + 1/2, x2 = m − 1/2 and x1 = y1 = 0 for some
m ≥ 1/2. We have
Z
1  
ux (z) − ux (x) = W (y) Ka,b (y1 , y2 − (m − 1/2)) − Ka,b (y1 , y2 − (m + 1/2)) dy,
π R2
where W is the odd extension of ω in R2 (3.3). Note that W is not Hölder in the y-direction
1/2
near y2 = 0, we cannot use the same method as that in the estimate of [ux ]Cx . On the other
1/2
hand, since W ∈ Cy (R × [m, ∞)), we can apply the previous method to obtain
Z
1 1
W (y)(Ka,b (y1 , y2 − (m − 1/2)) − Ka,b (y1 , y2 − (m + 1/2))dy) ≤ C1 (a)||ω||C 1/2 .
π y2 ≥m 2 y

Rotating the coordinate by 90 degree, we obtain the case studied in Section 3.3.
It remains to estimate
Z
1
I(b) = W (y)(Ka,b (y1 , y2 − (m − 1/2)) − Ka,b (y1 , y2 − (m + 1/2))dy)
π y2 ≤m
Z
1
= W (y1 , y2 + m)(Ka,b (y1 , y2 + 1/2) − Ka,b (y1 , y2 − 1/2)dy).
π y2 ≤0
Since W is not Hölder continuous across y = 0, we use [W ]C 1/2 to control I. Our idea is to
x
compare the integral I(b) with the case b = ∞, I(∞). To do so, we need a monotonicity Lemma.
Lemma 3.11. Suppose f, f g ∈ L1 and g ≥ 0 is monotone increasing on [0, ∞]. For any
0 ≤ k ≤ b ≤ c, we have
Z b+k Z c−k Z c+k
|f (x − k)|g(x)dx ≤ |f (x − k) − f (x + k)|g(x)dx + |f (x − k)|g(x)dx.
b−k b−k c−k

Proof. Denote by R, L the right and the left hand side of the above inequality, respectively. We
have
Z c−k   Z c+k Z b+k
R−L≥ |f (x − k)| − |f (x + k)| g(x)dx + |f (x − k)|g(x)dx − |f (x − k)|g(x)dx
b−k c−k b−k
Z c+k Z c−k Z c
= |f (x − k)|g(x)dx − |f (x + k)|g(x)dx = |f (x)|(g(x + k) − g(x − k))dx.
b+k b−k b
Since g is increasing on [0, ∞), we prove R ≥ L. 
42 JIAJIE CHEN AND THOMAS Y. HOU

Now, we are in a position to estimate I. Since Ka,b (y1 , y2 ) is odd in y1 , we yield

Z
1 p
|I| ≤ 2y1 Ka,b (y1 , y2 + 1/2) − Ka,b (y1 , y2 − 1/2) dy · [ω]C 1/2 .
π y2 ≤0,y1 ≥0 x

For a fixed y1 with |y1 | ≤ a and b ≥ 1/2, using the definition of Ka,b (3.4), the odd symmetry
Ka,b (y1 , y2 + 1/2) − Ka,b (y1 , y2 − 1/2 in y2 , and Lemma 3.11 with k = 1/2 and c = ∞, we get
Z
|Ka,b (y1 , y2 + 1/2) − Ka,b (y1 , y2 − 1/2)|dy2
y ≤0
Z 2
= |Ka,b (y1 , y2 + 1/2) − Ka,b (y1 , y2 − 1/2)|dy2
y2 ≥0
Z b−1/2 Z b+1/2
= |K1 (y1 , y2 + 1/2) − K1 (y1 , y2 − 1/2)|dy2 + |K1 (y1 , y2 − 1/2)|dy2
0 b−1/2
Z ∞
≤ |K1 (y1 , y2 + 1/2) − K1 (y1 , y2 − 1/2)|dy2 .
0

Since Ka,b (y) = 0 for |y1 | ≥ a, integrating the above inequality in y1 from 0 to a, we prove
1 a ∞p
Z Z
|I| ≤ 2y1 |K1 (y1 , y2 + 1/2) − K1 (y1 , y2 − 1/2)|dy · [ω]C 1/2 .
π 0 0 x

4. Energy estimates
Recall the decomposition (2.75) in Section 2.10. In this section, we perform energy estimates
of W1 following the ideas and some derivations in Sections 2.7.1, 2.8.2, 2.8.3. The goal of the
energy estimates is to control several weighted L∞ norms of ω, η, ξ and their weighted Hölder
norms and establish the estimates (A.6) for the coefficients in the estimates. The condition (A.6)
means that the damping term is stronger than the bad terms. Then we can further establish
stability using the stability Lemma A.2.

4.1. The main equation. After choosing a suitable approximation for the velocity u and using
the approach described in Section 2.10, the main equations (2.75) for ω1 , η1 , ξ1 read
∂t ω1 + (c̄l x + ū + u(ω)) · ∇ω1 = c̄ω ω1 + η1 − uA (ω1 ) · ∇ω̄,
(4.1) ∂t η1 + (c̄l x + ū + u(ω)) · ∇η1 = (2c̄ω − ūx )η1 − v̄x ξ1 − uA · ∇θ̄x − ux,A · ∇θ̄,
∂t ξ1 + (c̄l x + ū + u(ω)) · ∇ξ1 = (2c̄ω − v̄x )ξ1 − ūy η1 − uA · ∇θ̄y − uy,A · ∇θ̄,
˜ defined
where ū = (ū, v̄), and uA (ω1 ) is the velocity after subtracting the approximation term û
in (2.82), (2.88), (2.89), (2.68)
(4.2) ˜ ),
uA (f ) , ũ(f ) − û(f ˜ x (f ),
ux,A , ũx (f ) − û ˜ y (f ).
uy,A , ũy (f ) − û
Note that we do not have ∂x uA = ux,A since we choose the approximations for u, ux , uy sepa-
rately. Similarly, we do not have ∂y η1 = ∂x ξ1 . We make these finite rank perturbations to the
velocity and the linearized equations by subtracting K2i (2.68) from L in (2.75). We also remove
the cω f¯χ,i terms (2.66) in (2.65) by subtracting K1i from L in (2.75). We adopt the notation
from (2.75)
(4.3) W1,1 = ω1 , W1,2 = η1 , W1,3 = ξ1 .
Since we will perform weighted L∞ and C 1/2 estimate on ω1 , η1 , ξ1 , we also include the
advection terms u(ω) · ∇f from the nonlinear part in the above system. At this stage, we have
dropped the remaining term Ri , N F (W1 + W2 ), part of the nonlinear terms Ni , the error term
F i in (2.75) to simplify the presentation.
For initial perturbations that satisfy ω1,0 , η1,0 , ξ1,0 = O(|x|3 ), the system (2.75) preserve these
vanishing orders. See more discussions in Section 2.10.4.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 43

We introduce Td (ρ), di (ϕ) to denote the coefficients of the damping terms and b(x) to denote
the coefficient of the advection
 
b(x) = c̄l x + ū + u, Td (ρ) = ρ−1 (c̄l x + ū + u) · ∇ρ = ρ−1 (b · ∇ρ),
(4.4)
d1 (ρ) = Td (ρ) + c̄ω , d2 (ρ) = Td (ρ) + 2c̄ω − ūx , d3 (ρ) = Td (ρ) + 2c̄ω + ūx .
The terms di (ρ) appear naturally in the weighted L∞ (ρ) estimates of W1,i . See below (4.6).
In the equation of W1,i , we treat the terms other than the local terms of W1,i in (4.1) as bad
terms
B1 (x) , η1 − uA (ω1 ) · ∇ω̄, B2 (x) , −v̄x ξ1 − uA · ∇θ̄x − ux,A · ∇θ̄,
(4.5)
B3 (x) , −ūy η1 − uA · ∇θ̄y − uy,A · ∇θ̄.
With the above notations, we can simplify (4.1) as follows
∂t W1,i + b · ∇W1,i = di (1)W1,i + Bi ,
where di (1) acts on constant function 1 and Td (1) = 0. The weighted quantity enjoys the
following estimate
(4.6) ∂t (W1,i ρ) + b · ∇(W1,i ρ) = di (ρ)W1,i ρ + Bi ρ.
We choose the following weights for the weighted C 1/2 estimate
ψ1 = p11 |x|−2 + p12 |x|−1/2 + p13 |x|−1/6 ,
(4.7)
ψi = pi1 |x|−5/2 + pi2 |x|−1 + pi3 |x|−1/2 + pi4 |x|1/6 , i = 2, 3,
where pij are given in (C.1). The above weights can be determined by the analysis of the singular
scenario in Section 2.8.2, where we consider the Hölder estimate for any pair x, z ∈ R++
2 with
x2 = z2 and |x − z| being sufficiently small.
Remark 4.1. The reader should not confuse the weights ψi with the notation for the stream
function φ = (−∆)−1 ω. In this paper, we rarely use the stream function.
From Section 2.8.2, we know that in the scenario when x2 = z2 and |x−z| is sufficiently small,
we have enough damping to obtain the stability estimate. See (2.45) and (A.6) in Lemma A.2.
To estimate the more regular case when |x − z| is not small, we need to control the weighted
L∞ norm of ω1 , η1 , ξ1 . We will follow the ideas in Sections 2.8.3 2.10 to first show that the
weighted L∞ estimates with suitable weights are almost close. We then combine the L∞ and
C 1/2 estimates to close the stability estimate. We will show that in the more regular case when
|x − z| is not small, the damping factor in the Hölder estimate, i.e., λ in (A.6), is similar or even
larger than c1 , c2 in (2.45). Therefore, from c1 , c2 in (2.45), we can get a good estimate of the
stability factor λ∗ ≈ c1 , c2 in our overall energgy estimates based on (A.6) and Lemma A.2
(4.8) λ∗ ∈ [0.06, 0.12].

4.1.1. Guidelines of choosing the Hölder weights. To choose the parameters in the above weights
(C.1), we first choose different powers so that we can control the solution in the near-field and
the far-field. Then we choose the coefficients pij such that we can obtain the damping terms
from the local parts following the derivations in the weighted Hölder estimates in Section 2.7.1.
Next, we use the estimates in Section 2.8.2 and treat the nonlocal terms as bad terms (4.5).
We further optimize the coefficients so that we can obtain (2.45) with c1 , c2 as large as possible.
These ideas are similar to those presented in [13], and we refer to [13] for more discussions.
After we determine the weight ψi , we further determine the weights gi in the Hölder norm
||(W1,i ψi (x) − W1,i ψi (z))gi (x − z)||∞ by studying the Hölder estimates with |x − z| being suf-
ficiently small. In this case, similar to the analysis in Section 2.8.2, the more regular terms
vanish. We use Lemma 3.1-Lemma 3.5 and the triangle inequality
g(x − z)|f (x) − f (z)| ≤ g(x − z)(|f (x1 , x2 ) − f (z1 , x2 )| + |f (z1 , x2 ) − f (z1 , z2 )|)
44 JIAJIE CHEN AND THOMAS Y. HOU

to estimate the Hölder norm of ∇u. In general, applying the triangle inequality in the Hölder
estimate leads to a larger constant. For example, if |f (x1 , a) − f (z1 , a)| ≤ A|x1 − z1 |1/2 for any
a, x1 , z1 and |f (a, x2 ) − f (a, z2 )| ≤ A|x2 − z2 |1/2 for any a, x2 , z2 , then
1/2 1/2
|f (x) − f (z)| h + h2
≤ A 12 , hi = xi − zi .
|x − z|1/2 |h1 + h22 |1/4
√ √
The upper bound can be 2A if h1 = h2 , which leads to an additional factor 2. One way to
avoid this overestimate is to choose
(4.9) g(h1 , h2 ) = (|h1 |1/2 + c|h2 |1/2 )−1 .
for some constant c in the weighted Hölder estimate. In the above example, one can choose c = 1
and obtain g(x − z)|f (x) − f (z)| ≤ A for any x, z. However, in the weighted Hölder estimate,
we want to obtain a damping factor from the weight g
dg = g(x − z)−1 (b(x) − b(z)) · (∇g)(x − z) < 0, b(x) = c̄l x + ū + u.
See (2.30). Yet, for the weight (4.9), dg can be positive and even unbounded since (∂1 g)(0, h2 ) =
∞ and b1 (x) − b1 (z) 6= 0 when x1 = z1 . To overcome this difficulty, we perturb the weight (4.9)
by considering
p p
(4.10) gi (h) = qi0 (h)gi0 (1, 0)−1 , gi0 (h) = ( |h1 | + qi1 |h2 | + qi3 |h2 | + qi2 |h1 |)−1 ,
for some small qi1 , qi2 . We divide the factor gi0 (1, 0) to normalize gi (1, 0) = 1. To exploit the
anisotropy of the flow (see Section 2.7.2), we choose qi3 < 1. The parameters qij are given in
(C.1).
We remark that we still have a larger constant when we estimate g(x − z)(∇u(x) − ∇u(z))
for general (x, z) than the case x1 = z1 or x2 = z2 . Yet, since we also gain more damping from
the above dg when |x2 − z2 |/|x1 − z1 | is not too small, we can still show that the damping term
dominates other nonlocal terms.
For η and ξ, we choose the same weight: g3 (h) = g2 (h). To determine the parameters gij , we
first find x ∈ R++
2 such that we have the least damping in case when |x − z| is sufficiently small
with x2 = z2 . That is, we find x∗ such that the left hand side of (2.45) achieves the maximum
at x∗ . Then at such point, we perform the Hölder estimates with other ratio |z2 − x2 |/|z1 − x1 |
and keeping |x − z| small. We choose qij so that the damping factor is larger than or close to
the one in the case of x2 = z2 .
4.2. Ideas of estimating the nonlocal terms. In the energy estimates, we need to per-
form weighted L∞ and Hölder estimates on the velocity uA , uA,x , uA,y (4.2) given that ω1 ϕ1 ∈
L∞ , ω1 ψ1 ∈ C 1/2 for some weights ϕ1 , ψ1 . For f = uA , (∇u)A , it can be written as
Z
I(f )(x) = Kf (x, y)Ω1 (y)dy,
R2
for some kernel Kf , where Ω1 is the odd extension of ω1 in y from R2+ to R2 (3.3). In the
case without the approximation terms, the formulas of ∇u are given in (3.4). For f = uA , the
kernel involves ∇⊥ log |y| and has a singularity of order |x|−1 , which is locally integrable. To
obtain a sharp weighted estimate of uA with some singular weight ρ, since Ω1 is odd in y1 , y2 ,
we symmetrize the kernel and then apply the L∞ estimate
Kfsym (x, y) = Kf (x, y) − Kf (x, −y1 , y2 ) − Kf (x, y1 , −y2 ) + Kf (x, y),
Z Z
(4.11) |ρ(x)I(u)(x)| = ρ(x) Kusym ω1 (y)dy ≤ ρ(x) |Kusym |ϕ−1
1 (y)dy · ||ω1 ϕ1 ||L∞
R2++ R2++

, ρ(x)C(u, x)||ω1 ϕ1 ||L∞ ,

where C(u, x) denotes the last integral on the second line. The above estimate is sharp in the
sense that for a fixed x, the equality can be achieved if ωϕ1 (y) = Csgn(K sym (x, y)) for some
constant C. For a given weight ϕ1 , the constant C(u, x) is independent of ω1 and is an integral
of some explicit function. We can estimate it effectively for all x using the scaling symmetry of
the kernel and numerical computation with rigorous error estimates.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 45

For f = (∇u)A , the kernel has a singularity of order |x|−2 , which is not integrable near the
singularity. We decompose the integral I(f ) into the nonsingular part (NS) and the singular
part (S) with singular region R centered around x with radius r(x)
I(f ) = IN S (f ) + IS (f ), R(x) = {y : max |xi − yi | ≤ r(x)}.
i=1,2
(4.12) Z Z
IN S (f ) = Kf (x, y)Ω1 (y)dy, IS (f ) = Kf (x, y)Ω1 (y)dy.
Rc R
∞
In the weighted L estimate of (∇u)A , for IN S , we use the above idea and ||ωϕ1 ||L∞ to
control it. For the singular part, we further decompose it using the identity related to the
commutator, e.g., (2.34). We apply the above L∞ estimate (4.11) to the regular term. The
singular term related to ∇u(ωψ1 ) is estimated using ||ωψ1 ||C 1/2 . For example, we have the
following estimate
Z Z
s1 s2 s1 s2
4
(ωψ1 )(x − s)ds = 4
((ωψ1 )(x − s) − (ωψ1 )(x1 + s1 , x2 − s2 )ds
|s1 |,|s2 |≤τ |s| 0≤s1 ≤s2 ,|s2 |≤τ |s|

≤Cτ 1/2 [ωψ1 ]C 1/2 ,

x

where C is some constant related to the kernel and is independent of τ . In short, we can estimate
ρI(f ) with some singular weight as follows
(4.13) |ρ(x)I(f )(x)| ≤ C1 (x)||ωϕ1 ||L∞ + C2 (x)[ωψ1 ]C 1/2 + C3 (x)[ωψ1 ]C 1/2 ,
x y

for some constant Ci (x). We will further bound the right hand side using the energy.
The weighted Hölder estimate is more involved. For (∇u)A , we again decompose it into the
regular part and the singular part. For the singular part, we will use the sharp Hölder estimates
in Lemma 3.1-Lemma 3.5. For the nonsingular part, it is locally Lipschitz. We can estimate
its Lipschitz norm by computing suitble integrals and using ideas similar to the above. The
estimate for uA is easier since it is more regular. We refer the details to Section 8.

4.2.1. Scaling symmetry and rescaled integral. In the above computation of the integrals, e.g.,
(4.11), there are two singularities. Firstly, theRweight ρ(x) is singular near 0, which can amplify
the error in the computation of the integral R+ |Kusym (x, y)ρ(y)−1 |dy significantly. Secondly,
2
the kernel Ku (x, y) is singular near y = x. In the case of ∇u, the associated kernel is singular
of order 2. If there are only a few x, one can design a mesh that is adapted to the singularity
y = x and then apply the standard quadrature rule. However, it is very difficult to apply this
method to compute the integrals for all x. A crucial observation is that the kernel K(x, y) enjoys
scaling symmetry, which enables us to restrict the singularity x in a finite domain away from 0
by choosing suitable rescaling.
Denote fλ (x) , f (λx). We consider the kernels about ∇u, which are singular of order
2 and satisfy K(λx, λy) = λ−2 K(x, y). For λ to be chosen, applying a change of variables
y = λŷ, x = λx̂, we get
Z Z Z
ρ(x) K sym (x, y)ω(y)dy = ρ(λx̂) K sym (λx̂, λŷ)ω(λŷ)λ2 dŷ = ρλ (x̂) K sym (x̂, ŷ)ωλ (ŷ)dŷ.
R++
2 R++
2 R++
2

Now, applying the L∞ estimates, we obtain

Z Z
|ρ(x) K sym (x, y)ω(y)dy| ≤ ||ωλ ϕ1,λ ||L∞ ρλ (x̂) |K sym (x̂, ŷ)|ϕ1,λ (ŷ)−1 dŷ.
R++
2 R++
2

Note that ||ωλ ϕλ ||L∞ = ||ωϕ||L∞ . Hence, to establish the estimate, it suffices to compute the
rescaled integral. The advantage of the above integral compared to the one without rescaling
is that the integral is singular at the rescaled point x̂, which can be restricted to some finite
domain by choosing suitable rescaling parameter λ. As a result, we can design an adaptive mesh
which is dense in the O(1) region to compute the integrals and we do not need to remesh in
the computation of integrals with different x̂. In addition, x̂ can be chosen to be away from 0,
e.g. |x̂|  1, so that ρλ (x̂) is not singular in x̂. For example, we can write |x|−2 = λ−2 |x̂|−2 by
46 JIAJIE CHEN AND THOMAS Y. HOU

choosing λ = |x|/|x̂| with |x̂|  1. The above rescaling argument enables us to overcome the
difficulties caused by the singularities in our computation. We refer more details to Section 8.

4.3. Weighted L∞ estimate with decaying weights. Recall the discussion of several weighted
L∞ estimates after Remark 2.3. We first perform energy estimate with decaying weights in the
weighted L∞ estimate so that we have more damping in the energy estimates. See the weighted
L∞ estimate in the model problem in Section 2.7.1 for more motivations. We choose the follow-
ing weights

ϕ1 = (p41 |x|−2.4 + p42 |x|−1/2 )|x1 |−1/2 + p43 |x|−1/6 , ϕ4 = ψ1 |x1 |−1/2 ,
(4.14)
ϕi−3 = (pi1 |x|−5/2 + pi2 |x|−3/2 + pi3 |x|−1/6 )|x1 |−1/2 + pi4 |x|−1/4 + pi5 |x|1/7 , i = 5, 6,

for the weighted L∞ estimate with parameters pij given in (C.2). We apply ψ1 , ϕ1 for ω, ψ2 , ϕ2
for η, and ψ3 , ϕ3 for ξ. We will use ϕ4 in Section 4.5.2 for an additional weighted L∞ estimate
of ω1 . We will discuss the ideas of choosing ϕi in Section 4.3.3.
Using the weights ϕ1 , ψ1 , we can estimate the constants in the weighted estimate of uA , (∇u)A
Cij,k in (4.11), (4.13) using the ideas in Section 4.2

(4.15) |ρij fij (x)| ≤ Cij,1 (x)||ωϕ1 ||∞ + Cij,2 (x)[ωψ1 ]C 1/2 + Cij,3 (x)[ωψ1 ]C 1/2 ,
x y

where f01 = uA , f10 = vA , f11 = ux,A , etc. We use these indices since u = −∂y φ, v = ∂x φ, ux =
−∂x ∂y φ etc, where φ is the stream function. We add the weight ρij to capture the vanishing
order and decays of uA , ∇uA

(4.16) ρ10 = ρ01 = |x|−3 + |x|−7/6 , ρij = ψ1 , i + j = 2.

For (∇u)A , we choose ρij = ψ1 , i + j = 2, since we need to estimate (∇u)A ψ1 using the Hölder
norm of ω1 ψ1 and ∇u and ω1 are of the same order. To control uA , we do not need to use the
Hölder seminorm and have

(4.17) Cij,2 (x) = Cij,3 (x) = 0, i + j = 1.

4.3.1. Piecewise upper bounds. In practice, we discretize a very large domain [0, D]2 in R++ 2
using the mesh y in Section 5 to compute the profiles. Using the method in Section 8, we can
estimate uA , ∇uA uniformly in each grid [yi , yi+1 ] × [yj , yj+1 ]. In particular, Cij,1 , Cij,2 , Cij,3 in
the upper bound (4.14) are piecewise constants. Then we can track these bounds using n × n
/ [0, D]2 is straightforward since the coefficients of the
matrices. The estimate in the far-field x ∈
nonlocal terms in (4.1) have fast decay. The same ideas apply to all other estimates.
Operators and functions. To simplify the notations, we introduce some operators and func-
tions. We define

Tu (f )(x) = C01,1 |fx | + C10,1 |fy |, C(ux , i) , C11,i |θ̄x | + C20,i |θ̄y |,
(4.18)
C(uy , i) , C02,i |θ̄x | + C11,i |θ̄y |, C(f, a) , C(f, 1)a1 + C(f, 2)a2 + C(f, 3)a3 ,

for µ ∈ R3 and f = ux or f = uy . We will use Tu for the estimate of the uA · ∇f , C(ux , i) for
uA,x · ∇θ̄, and C(uy , i) for uA,y · ∇θ̄. Note that C(f, µ) is linear in µ.
Following the derivations in the weighted L∞ estimate in Section 2.7.1 and using (4.6), we
have the following estimates for W1,i = ω1 , η1 , ξ1 (4.3) in (4.1)

(4.19) ∂t (W1,i ϕi ) + (c̄l x + ū + u) · ∇(W1,i ϕi ) = −di (ϕi )W1,i ϕi + Bi (x)ϕi ,

where we have used the operators (4.4) di (·) to denote the coefficient of the damping terms, and
Bi (x) are the bad terms defined in (4.5). We estimate Bi (x) directly using the above estimates
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 47

for the nonlocal terms

ϕ1 ϕ1
|B1 (x)| ≤ ||η1 ϕ2 ||∞ + Tu (ω̄)||ω1 ϕ1 ||∞ ,
ϕ2 ρ10
ϕ2 ϕ2
|B2 (x)| ≤ |v̄x |||ξ1 ϕ3 ||∞ + Tu (θ̄x )||ω1 ϕ1 ||∞
ϕ3 ρ10
ϕ2
(4.20) + (C(ux , 1)||ω1 ϕ1 ||∞ + C(ux , 2)[ω1 ψ1 ]C 1/2 + C(ux , 3)[ω1 ψ1 ]C 1/2 ),
ψ1 x y

ϕ3 ϕ3
|B3 (x)| ≤ |ūy |||η1 ϕ2 ||∞ + Tu (θ̄y )||ω1 ϕ1 ||∞
ϕ2 ρ10
ϕ3
+ (C(uy , 1)||ω1 ϕ1 ||∞ + C(uy , 2)[ω1 ψ1 ]C 1/2 + C(uy , 3)[ω1 ψ1 ]C 1/2 ).
ψ1 x y

4.3.2. Weights between the L∞ norm and the Hölder norm. We cannot close the L∞ estimate
since the estimate of ∇uA involves the Hölder seminorm of ω1 . If we neglect the Hölder semi-
norm, we indeed obtain stability for the L∞ estimate with energy maxi ||W1,i ϕi ||∞ by checking
the condition (A.6). For example, for the η equation, we have
ϕ2 ϕ2 ϕ2
−d2 (ϕ2 ) − |v̄x | − Tu (θ̄x ) − C(ux , 1) ≥ λ > 0,
ϕ3 ρ10 ψ1
for some λ. To close our weighted L∞ and Hölder estimate using Lemma A.2, we need to choose
the weights µi among different norms such that (A.6) holds.
Recall the weighted Hölder norm from (2.4). We introduce the first energy
(4.21) E1 (t) = max(max ||W1,i ϕi ||∞ , τ1−1 ||ω1 ψ1 ||C 1/2 ).
i g1

Since g1 (h1 , 0) = |h1 |−1/2 , g1 (0, h2 ) = g1 (0, 1)|h2 |−1/2 (4.10), we obtain
E1 (t) ≥ τ1−1 ||ω1 ψ1 ||C 1/2 ≥ τ1−1 [ω1 ψ1 ]C 1/2 , τ1−1 g1 (0, 1)[ω1 ψ1 ]C 1/2 .
g1 x y

Using E1 (t) and the notation (4.18), we can simplify the estimate (4.15) for ∇uA as follows
(4.22)
|ρij fij | ≤ (C(fij , 1) + C(fij , 2)τ1 + C(fij , 3)τ1 g1 (0, 1)−1 )E1 = E1 · C(fij , (1, τ1 , τ1 g1 (0, 1)−1 )).
Similarly, the bound in Bi can be simplified as follows
C(f, 2)[ω1 ψ1 ]C 1/2 + C(f, 3)[ω1 ψ1 ]C 1/2 ) ≤ τ1 C(f, (0, 1, g1 (0, 1)−1 ))E1 (t), f = ux , uy ,
x y

where C(f, µ) is defined in (4.18). The constraint (A.6) for the η equation becomes
 ϕ2 ϕ2 ϕ2  ϕ2
(4.23) − d2 (ϕ2 ) − |v̄x | − Tu (θ̄x ) − C(ux , 1) − τ1 C(ux , (0, 1, g1 (0, 1)−1 )) ≥ λ.
ϕ3 ρ10 ψ1 ψ1
Similarly, we have another constraint for τ1 from the estimate of ξ1 . We want to obtain an
overall stability factor λ∗ (4.8) and thus choose λ ≈ λ̄∗ . We choose the largest τ1 such that the
inequality (4.23) and a similar inequality for ξ1 hold. The idea to choose large τ1 (or small τ1−1 )
is similar to that in (2.52) for the model problem, where the weight τ for the Hölder norm is
small. We choose the largest τ1 so that in the Hölder estimate for τ1−1 ω1 ψ1 , we have the small
factor τ1−1 associated with the weighted L∞ norm maxi ||W1,i ϕi ||∞ in (A.6). In our estimate,
we can choose
(4.24) τ1 = 5.
Although τ1 is not very large, it is enough for us to show that the estimate of the more
regular case in the Hölder estimate, i.e. |x − z| is not very small, is similar to or even better
than that in the singular case when |x − z| is small. There are two reasons. Firstly, we get
the above small factor τ1−1 when we estimate the more regular terms using the weighted L∞
norm. Secondly, as |x − z| increases, due to our localized estimates in Lemmas 3.1-3.4, the
constants in the estimates of the nonlocal terms decrease. Note that from (4.23), choosing a
larger τ1 requires better estimates on the nonlocal terms, e.g., smaller C(ux , i). For this reason,
we need to approximate the nonlocal terms u with finite rank operators with a higher rank,
which increases the computation cost. Due to this consideration, we choose a moderate τ1 .
48 JIAJIE CHEN AND THOMAS Y. HOU

Due to the anisotropy of the flow (see Section 2.7.2), we have a larger damping factors di
away from the boundary. Due to the decay of the solution, the coefficients in the estimates of
the nonlocal terms become smaller. Thus, we can focus on the near-field and the boundary to
test the conditions (A.6) for given weights and the approximations of u, ∇u that we choose in
Section 2.11.
In Figure 5, we plot the coefficients of the damping terms, e.g. −d2 (ϕ2 ) (4.23), and the
estimates of the bad terms, i.e. the sum of the terms with negative sign (4.23), and the remaining
damping factors (the left hand side of (4.23)) on the grid points along the boundary. The ξ1
variable enjoys a much better estimates near the boundary, so we do not plot it.
We choose the approximation terms û, ∇u d for u, ∇u along the boundary in Section 2.11 such
that the weighted estimates of u − û, ∇u − ∇u d are small. Near the center of the approximation
terms, xi in (2.82), we have better estimates of the bad terms. In Figure 5, the points xi are
the local minimum of the blue dashed curve. Since the coefficients of the nonlocal terms (4.1),
e.g., ω̄x , θ̄x , decay for large x, we do not need to construct the approximations for large x.
We choose ϕ1 slightly weaker than ϕ2 near the origin (4.14) such that ϕ1 /ϕ2 ||η1 ϕ2 ||∞ is
small, and we can obtain large stability factors for both ω and η, which are larger than 0.7. This
allows us to control a larger weighted residual error near the origin.
In Figure 6, we plot the stability factors, e.g., the left hand side of (4.23), in the weighted
L∞ estimates of ω1 , η1 . Due to the anisotropy of the flow, the damping terms and the stability
factors are much larger if the angle y/x is large. See Section 2.7.2.
Note that we only use these plots to visualize the estimates. To justify the inequalities (4.23),
we follow the methods in [13,14] and derive piecewise bounds of different functions based on the
estimates of the approximate steady state and the weights in Appendices E, C.3.

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 5 10 15 20 25 30 35 40

x x

Figure 5. Weighted L∞ estimates with slowly decay weights. Left figure:

estimates near 0, x ∈ [0, 1.8]; Right figure: estimates in a larger domain, x ∈
[0, 35]. The red curves shows the coefficient of the damping term −d1 (ϕ1 ), and
the estimate of B1 ; the blue curves are for −d2 (ϕ2 ), B2 in the estimates of η1 .
The green and the black curves are the stability factors in the estimate of ω1
and η1 .

4.3.3. Order of choosing the parameters. We have discussed how to choose the Hölder estimate
in Section 4.1.1. For ϕi , we first choose the weight ϕ1 for ω1 consisting of different powers to
take into account the vanishing order of ω1 near 0 and its decay in the far field. We add the
power |x1 |−1/2 in ϕ1 , ϕ2 , ϕ3 (4.14) since we need to control ||ω1 |x1 |−1/2 ψ1 ||∞ for the Hölder
estimate. See Section 2.7.4. In the estimate of ω1 |x1 |−1/2 ψ1 , we need to control η1 |x1 |−1/2 ψ1
and other weighted quantities with weights singular at x1 = 0. Thus, we add the weight |x1 |−1/2
in ϕi . We adjust the parameters in ϕ1 so that we have a good damping factor d1 (x) from the
local term for ω1 . Then we can estimate the nonlocal terms and the constants (4.15). Once we
obtain the estimates for ∇uA , uA , we choose the exponents of different powers in ϕ2 and adjust
the parameters so that we have better stability factors in the weighted L∞ estimate and choose
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 49

Figure 6. Weighted L∞ estimates in the near-field. Left: stability factors in

the estimate of ω1 . Right: stability factor in the estimate of η1 .

a larger τ1 (4.23). Since the equations of ξ1 and η1 are similar, we choose the same combination
of powers in ϕ2 and ϕ3 (4.14). Moreover, since ξ1 is weakly coupled with ω1 and η1 (see Section
2.6.2), we determine the parameters in ϕ3 after we obtain ϕ1 , ϕ2 .

4.4. Weighted Hölder estimates. Recall the weights ψi for ω1 , η1 , ξ1 (C.1) and gi (h) (4.10)
in the weighted C 1/2 estimates, the notation W1,i (4.3), and the simplified equation (4.6). The
goal of the weighted Hölder estimate is to control ||W1,i ψi ||C 1/2 , where || · ||C 1/2 is defined in
gi gi

(2.4), which along with the weighted L∞ estimate, we can control the second energy

(4.25) E2 (t) , max(E1 (t), τ1−1 max(||W1,1 ψ1 ||C 1/2 , µ1 ||W1,2 ψ2 ||C 1/2 , µ2 ||W1,3 ψ3 ||C 1/2 )
g1 g2 g3

for the weights µ1 , µ2 determined by analyzing the most singular scenario in Section 2.8.2 (2.43).
They are given in (C.4). In fact, these factors can be absorbed in the definition of ψ2 , ψ3 . We
have normalized the coefficient of the most singular power in ψ1 to be 1.
Following the derivations in the weighted Hölder estimates in Section 2.7.1 and using (4.6)
and Lemma 2.2, we derive the following for W1,i ψi and any x, z ∈ R2++
 
∂t Hi + (b(x) · ∇x + b(z) · ∇z )Hi = (di (ψi )W1,i ψi )(x) − (di (ψi )W1,i ψi )(z) gi (x − z)
(4.26)  
+ dg,i Hi + (Bi ψi )(x) − (Bi ψi )(z) gi (x − z) , I1 + I2 + I3 , Ri ,

where b(x) is the coefficient of the advection (4.4), dgi is the damping factor from gi in the
Hölder estimate, and Ji , Hi are given below
(4.27)
(b(x) − b(z)) · (∇gi )(x − z)
Ji , W1,i ψi , Hi (x, z) = (Ji (x) − Ji (z))gi (x − z), dg,i , .
gi (x − z)
The factor di is defined in (4.4)

(4.28) d1 (ψ1 ) = Td (ψ1 ) + c̄ω , d2 (ψ2 ) = Td (ψ2 ) + 2c̄ω − ūx , d3 (ψ3 ) = Td (ψ3 ) + 2c̄ω + ūx ,

and Bi is the bad term defined in (4.5).

We note that the second term I2 in (4.26) is already a damping term. See Section 2.7.1 and
discussion below Lemma 2.2. To further simplify the notation, we introduce

(4.29) ai (x) = di (ψi )(x), δ(f )(x, z) = f (x) − f (z).

We also drop the dependence of δ(f ) on x, z when there is no confusion. Using the above
notation we get

(4.30) δ(f g) = (f g)(x) − (f g)(z) = f (x)δ(g) + δ(f )g(z).

50 JIAJIE CHEN AND THOMAS Y. HOU

4.4.1. Estimate the explicit coefficients. In the Hölder estimates, we need to estimate (p̄q)(x) −
(p̄q)(z))g(x − z) for some coefficient p̄, perturbation q, e.g. q = ω1 , η1 , and some weight g,
e.g. g = gi . The coefficient p̄ depends on the weights ψi , ϕi and the approximate steady state
only. In particular, p̄ is quite smooth in a local region. We can estimate g(x − z)δ(p̄) effectively
using the method in Appendix G. Note that the approximate steady state, the singular weights
and their derivatives can be estimated effectively using the method in Appendices E, C.3. In
1/2 1/2
practice, we estimate the piecewise Cx and Cy seminorms of p̄(x), and then use the triangle
inequality to obtain the estimate of g(x − z)δ(p̄). Namely, given x, z, we have
|p(x1 , x2 ) − p(z1 , z2 )| ≤ |p(x1 , x2 ) − p(z1 , x2 )| + |p(z1 , x2 ) − p(z1 , z2 )|
(4.31)
≤ |x1 − z1 |1/2 A(x1 , x2 , z1 ) + |x2 − z2 |1/2 B(z1 , x2 , z2 ),
for some constants A, B. We discretize the domain R2++ using the same mesh y0 < y1 < .., yn in
our computation for the profile in Section 5 and estimate these constants for x ∈ Q1 , z ∈ Q2 for
different grids uniformly. These piecewise Hölder estimates of a function can be established using
the methods in Appendices G.6, G.7. Therefore, we can track the piecewise bounds A(x1 , x2 , z1 )
for x1 , x2 , z1 in each cube Ii × Ij × Ik , Ii = [yi , yi+1 ] using n × n × n matrices.
In general, such an estimate has some overestimates. Yet, since the problem is anisotropic in
the x and y directions, in the worst case scenario where |x2 − z2 | is much smaller than |x1 − z1 |,
this simple estimate is effective. See also Section 4.1.1.
Although the weights ψi , ϕi are singular near x = 0, from the estimates in the most singular
scenario in Section 2.8.2 (see Figure 4), we have better estimates near x = 0. Thus, the more
challenging part of our estimates comes from the region where x is away from 0, e.g. x around
0.5. In such a case, we can simply treat the weights ψi , ϕi as smooth functions.
Now, using (4.30), we obtain
P , δ(p̄q)g(x − z) = (p̄(x)δ(q) + δ(p̄)q(z))g(x − z) , P1 + P2 .
The second term is more regular. We can use the weighted L∞ norm of q to control it. For the
first term, we bound it using the weighted Hölder (semi)norm. Below, we discuss different cases.
In all cases, the estimate of P2 is much smaller than that of P1 . Moreover, we have another
decomposition
P = (p̄(z)δ(q) + δ(p̄)q(x))g(x − z) , P̃1 + P̃2 .
For x, z not sufficiently close, these two decompositions lead to two different estimates. We
will optimize these two estimates.

4.4.2. Estimate of I1 . Recall I1 from (4.26). Note that Hi = δ(Ji )gi (x − z) is the energy we
want to control. We have
I1 = δ(ai Ji )gi (x − z) = (ai (x)δ(Ji ) + δ(ai )Ji (z))gi (x − z)
(4.32)
= ai (x)Hi + δ(ai )gi (x − z)Ji (z) , I11 + I12 .
The first term is a damping term. Note that we can control Ji (x) using the weighted L∞ norm
in the energy E1 (4.21)
ψi (z) ψi (z)
|Ji (z)| = |(W1,i ψi )(z)| ≤ ||W1,i ϕi ||∞ ≤ E1 .
ϕi (z) ϕi (z)
Since ai is a given function with an explicit expression, we follow Section 4.4.1 and estimate
δ(ai )gi (x−z) using the method in Appendix G. In particular, when |x−z| is small, δ(ai )gi (x−z)
is very small. It follows
ψi (z)
(4.33) |I12 | ≤ |δ(ai )gi (x − z)| E1 .
ϕi (z)
Similarly, we can also define I˜11 = ai (z)Hi and I˜12 = δ(ai )gi (x − z)Ji (x) and obtain
ψi (x)
(4.34) I1 = I˜11 + I˜12 , I˜11 = ai (z)Hi , |I˜12 | ≤ |δ(ai )gi (x − z)| E1 .
ϕi (x)
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 51

We choose one of the above estimates according to the relative size of the following terms
ψi (z) ψi (x)
(4.35) m1 = ai (x) + τ1−1 |δ(ai )gi (x − z)| , m2 = ai (z) + τ1−1 |δ(ai )gi (x − z)| ,
ϕi (z) ϕi (x)
where τ1 is the ratio between the L∞ and C 1/2 norm in the energy. We use the decomposition
(4.32) and its estimates if m1 is smaller. We choose (4.34) if m2 is smaller. We remark that
this optimization is similar to maximize the left hand side of (A.6) (the sign is different), and
allows us to obtain a better stability factor. We will use similar optimizations several times to
get better stability factors. Following the discussions and ideas in Section 4.4.1, we can track
the piecewise bounds of the above functions and estimates, e.g. ai , τ1−1 |δ(ai )gi (x − z)| ψi (z)
ϕi (z) .

4.4.3. Estimate of I3 . Recall I3 from (4.26) and Bi from (4.5). The term Bi involves both the
local term and nonlocal terms. We treat them as bad terms and estimate them separately.
Estimate of the local part. We focus on η1 in B1 . Other terms −v̄x ξ1 in B2 , −ūy η1 in B3
(4.5) can be estimated similarly. Note that the weights are different for ω1 , η1 . We rewrite the
difference as follows
ψ1  ψ1 ψ1 
δ(η1 ψ1 )g1 (x − z) = δ(η1 ψ2 )g1 (x − z) = δ(η1 ψ2 ) (x) + (η1 ψ2 )(z)δ( ) g1 (x − z) , P1 + P2 .
ψ2 ψ2 ψ2
ψ1
The term P2 is more regular. We follow Section 4.4.1 to estimate δ( ψ2
)g1 (x − z). Using the
∞
weighted L norm of η1 and the energy E2 (4.25), we obtain
ψ2 (z) ψ1 ψ2 (z) ψ1
(4.36) |P2 | ≤ ||η1 ϕ2 ||∞ δ( )g1 (x − z) ≤ E2 δ( )g1 (x − z) .
ϕ2 (z) ψ2 ϕ2 (z) ψ2
Following Section 4.4.1, we can track the piecewise bound of the coefficient in the above upper
bound. For P1 , we have
ψ1 g1 (x − z) ψ1 g1 (x − z)
|P1 | ≤ (x) |δ(η1 ψ2 )g2 (x − z)| ≤ (x) ||η1 ψ2 ||C 1/2
ψ2 g2 (x − z) ψ2 g2 (x − z) g2
(4.37)
ψ1 g1 (x − z)
≤ (x) E2 τ1 µ−1
2 ,
ψ2 g2 (x − z)
where we have used the energy E2 (4.25) in the last inequality. We note that in the estimate of
τ −1 ||ω1 ψ1 ||C 1/2 , we have the term τ1−1 P1 . The weight τ1−1 cancels τ1 in the above upper bound.
g2

Note that g1 and g2 are equivalent to |h|−1/2 and homogeneous of order −1/2. The quantity
g1 (x−z)
g2 (x−z) only depends on the ratio between x1 − z1 , x2 − z2 . We also track this ratio.
For large |x − z|, we have a trivial estimate
ψ1 ψ1 ψ1 ψ1
|δ(η1 ψ1 )g1 (x − z)| ≤ ||η1 ϕ2 ||∞ ( (x) + (z))g1 (x − z) ≤ E2 ( (x) + (z))g1 (x − z).
ϕ2 ϕ2 ϕ2 ϕ2

1/2
Estimate of the nonlocal part. To control the nonlocal terms in Bi , we use the sharp Cx
1/2
and Cy estimates in Section 3 for the most singular part and the estimates in Section 8 for
the more regular part. We focus on the estimate of −ux,A θ̄x in B2 (4.5), which contributes to
the largest part in the estimate. Firstly, we have
ψ2  ψ 
2
 ψ 
2
δ(ux,A θ̄x ψ2 ) = δ(ux,A ψ1 θ̄x ) = δ(ux,A ψ1 ) θ̄x (x) + ux,A (z)ψ1 δ θ̄x , P3 + P4 .
ψ1 ψ1 ψ1
 
ψ2
The term P4 is more regular. For ux,A (z), we use the estimate in Section 4.3. For δ θ̄x ψ 1
,
the term θ̄x ψ
ψ1 has vanishing order |x|
2 1/2
near x = 0 and it is in C 1/2 . We follow Section 4.4.1
to estimate it. In particular, we have
(4.38) |P4 | ≤ E1 (C1 (x, z)|x1 − z1 |1/2 + C2 (x, z)|x2 − z2 |1/2 ),
for some functions Ci (x, z) depending on the weights and the approximate profile. See Section
4.3.1. Again, we can obtain piecewise upper bound of these functions.
52 JIAJIE CHEN AND THOMAS Y. HOU

For P3 , using the triangle inequality, we have

|ux,A ψ1 (x) − ux,A ψ1 (z)| = |ux,A ψ1 (x1 , x2 ) − ux,A ψ1 (z1 , x2 )| + |ux,A ψ1 (z1 , x2 ) − ux,A ψ1 (z1 , z2 )|.
1/2 1/2
Applying the Cx and Cy estimates in Section 3, 8 to P51 , P52 , respectively, we obtain
(4.39)
1/2
|ux,A ψ1 (x) − ux,A ψ1 (z)| ≤ C3 (x, z)|x1 − z1 |1/2 + C4 (x, z)|x1 − z1 |1/2 max(τ1 ||ω1 ϕ1 ||∞ , [ω1 ψ1 ]Cg )
1

≤ (C3 (x, z)|x1 − z1 |1/2 + C4 (x, z)|x1 − z1 |1/2 )τ1−1 E2 ,

for some constants C3 , C4 only depending on the weight and the approximate profile. We remark
that the constants C3 , C4 are very close to the constants provided by the sharp Hölder estimates
in Section 3. Again, we can obtain these piecewise upper bounds and track them carefully. See
Section 4.3.1.
When |x − z| is not small, we can apply the triangle inequality and the L∞ estimate of ux,A
in Section 4.3 to obtain another bound.
In practice, we only need to apply the above Hölder estimate when |x| and |z| are comparable
and |x − z| is very small, e.g. |x| ≤ |z| ≤ 1.2|x|. Beyond such a range, the L∞ estimate already
provides a better estimate.
4.4.4. Summarize the estimates. In summary, for the right hand sides in (4.26), when x, z are
close, we can obtain the following estimates
Ri = (dg,i (x, z) + ai (p))Hi + B̃i , B̃i = Iˆ12 + δ(Bi ψi )gi (x − z),
where B̃i is the combination of the term Iˆ12 (4.32) or (4.34) and I3 , and we can estimate it as
follows,
 gi (x − z) 
(4.40) |B̃i | ≤ (|x1 − z1 |1/2 A1 (x, z) + |x2 − z2 |1/2 A2 (x, z))gi (x − z) + A3 (x, z) E2 ,
gki (x − z)
with (p, Iˆ12 ) = (x, I12 ) or (z, I˜12 ) depending on the size of m1 , m2 in (4.35), (k1 , k2 , k3 ) = (2, 3, 2).
The term A3 (x, z) ggki (x−z)
(x−z) comes from the estimate of the local terms, e.g. (4.37), the term
i
|x1 − z1 |1/2 A1 (x, z) + |x2 − z2 |1/2 A2 (x, z) bounds the estimate of the nonlocal terms, e.g. (4.39),
and the regular terms such as (4.33), (4.36), (4.38) using the method (4.31). We remark that we
can obtain piecewise upper bounds of the coefficients Ai in the above estimates and can track
them using matrices. See Sections 4.3.1, 4.4.1.
Checking the stability conditions. According to Lemma A.2, to obtain linear stability, we
need to check the conditions (A.6). We use the following method to check such a condition.
We discretize a large domain [0, D]2 into small grid cells Qij = Ii × Ij using the same mesh
y0 < y1 < .. < yn as that in Section 5. We have piecewise bounds for the coefficients Ai (4.40)
in each pair of grid cells x ∈ Qij , z ∈ Qkl .
Firstly, we fix the locations of x, z to some grid cells: x ∈ Q1 , z ∈ Q2 . Then we can derive the
bound Ai (x, z). We still need to control the bound for different x − z. One observation is that
given Ai is constant in Q1 × Q2 , the upper bound in (4.40) is 0−homogeneous in x − z. Thus, we
only need to further consider the ratio between r1 = x1 − z1 , r2 = x2 − z2 . Similar considerations
apply to the damping factors dg,i (4.26). We consider four different cases depending on the sign
of r1 /r2 and the size between |r1 |, |r2 |. We focus on the r1 /r2 > 0 and r2 ≤ r1 to illustrate
the ideas. In such a case, we can normalize r1 = 1 and 0 ≤ r2 ≤ 1. Now the problem reduces
to checking the inequality in 1D. Since these functions have monotone properties, e.g. gi (1, r2 )
is decreasing in r2 , these inequalities can be checked by partitioning r2 ∈ [0, 1] into smaller
intervals r2 ∈ [bi , bi+1 ], 0 = b0 < b1 < ... < bN = 1.
Note that when |x − z| is far away, we will have much better estimate due to the improvement
from the sharp Hölder estimates in Lemmas 3.1-3.4. In practice, for x ∈ Qi1 ,i2 z ∈ Qj1 ,j2 with
max(|i1 − j1 |, |i2 − j2 |) ≥ 20, we already have much better stability factors.
In Figure 7, we plot the estimates for the grid point along the boundary with r2 /r1 = 0. Here,
we consider x ∈ Ii × I0 , z ∈ Ij × I0 , where Ii = [yi , yi+1 ] is a small interval. This corresponds
to the case where we have the smallest damping. Other cases with small j − i are similar
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 53

and the estimate is better. The estimate of the bad terms in the η1 equation is very close to
the one in the most singular scenario based on the sharp inequalities. In some cases, we have
better estimates since |x − z| is far away and the improvement of constants for the localized
velocity from Lemmas 3.1-3.4. For larger ratio |r2 /r1 |, we have larger stability factors due to
the anisotropy of the flow. See Section 2.7.2. We have designed the weights gi to exploit this
properties.

3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 5 10 15 20 25 30

x x

Figure 7. Weighted Hölder estimates. Left figure: estimates near 0, x ∈

[0, 1.8]; Right figure: estimates in a larger domain, x ∈ [0, 30] The red curves
shows the coefficient of the damping term for δ(W1,1 ψ1 )g1 (x − z) and the esti-
mate of the bad terms; the blue curves are for the Hölder estimate of W1,2 ψ2 .
The green curve is the same estimated as that in the most singular scenario
based on the sharp inequalities. The magenta and the black curves are the sta-
bility factors in the Hölder estimate for ω1 , η1 . The stability factors are larger
than 0.08.

In Figure 8, we consider x ∈ Ii × I0 , z ∈ Ij × I0 with j − i = 10. The stability factor for η1

shown by the black curve becomes much larger and is larger than 0.3.

3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 5 10 15 20 25 30

x x

Figure 8. Weighted Hölder estimates with larger |x − z|

We refer the full details of the verification to the upcoming supplementary material.

4.5. Additional Weighted L∞ estimates.

4.5.1. Weighted L∞ estimates with growing weights. To close the nonlinear estimate in (2.65),
(2.75), we need to control ||ω||∞ , ||∇θ||∞ , ||W1,i ||∞ . The quantities we estimate in the previous
sections only allow us to control ω1 in the near field since the singular weight ϕ1 decays in the
far field, see (4.14). To obtain better control of W1,i in the far-field, we perform similar weighted
54 JIAJIE CHEN AND THOMAS Y. HOU

L∞ energy estimates on W1,i ϕgi with a different weight ϕgi that grows faster than ϕi in the
far-field, i = 1, 2, 3.
We choose the following weights ϕgi
(4.41)
ϕg1 = ϕ1 + p71 |x|1/16 , ϕg2 = ϕ2 + p81 |x|1/4 ,
ϕg3 = (p31 |x|−5/2 + p32 |x|−3/2 + p33 |x|−1/6 )|x1 |−1/2 + p34 |x|−1/4 + p91 |x|1/7 + p92 |x|1/4 .
The subscript “g” is short for “grow”. For ϕg1 , ϕi2 , we simply add the powers |x|1/16 , |x|1/4 ,
which grow faster than ϕ1 , ϕ2 , respectively. For ϕg3 , we also add the power |x|1/4 . In addition,
we modify the coefficients of |x|1/7 in ϕ3 (4.14). Other parts of ϕg3 are the same as ϕ3 in (4.14).
The parameters are given in (C.3). Since we can close the weighted L∞ and C 1/2 estimates at
the linear level, we can establish the stability estimate for W1,i ϕgi for ϕgi close to ϕi .
We remark that to determine these parameters, especially p71 , we first use the estimate of the
nonlocal terms in (4.15), (4.22) and then perform energy estimates similar to those in Section
4.3. By checking the stability condition for ||W1,i ϕgi ||∞ , we can determine the value of p71 . To
improve the estimates, we further use ||ω1 ϕg1 ||∞ and the Hölder norm of ω1 ψ1 to derive another
estimate of uA , ∇uA
(4.42) |ρij fij (x)| ≤ Cgij,1 (x)||ωϕg1 ||∞ + Cgij,2 (x)[ωψ1 ]C 1/2 + Cgij,3 (x)[ωψ1 ]C 1/2 ,
x y

where g is short for “grow”, f01 = uA , f10 = vA , etc similar to those in (4.15). Similar to (4.17),
we have
(4.43) Cgij,2 (x) = Cgij,3 (x) = 0, i + j = 1.
For some weight parameters τ2 , µ4 to be determined, we consider a new energy
 
(4.44) E3 (t) = max E2 (t), τ2 max(µ2 ||ω1 ϕg1 ||∞ , ||η1 ϕg,2 ||∞ , ||ξ1 ϕg,3 ||∞ ) .

where E2 (t) is defined in (4.25). The parameters τ2 , µ4 are chosen in (C.4).

With these weight parameters τ2 , µ4 , we can simplify the above bound similar to (4.22)
|ρij fij (x)| ≤ (Cgij,1 (x)(τ2 µ4 )−1 + Cgij,2 (x)τ1 + Cgij,3 (x)τ1 g1 (0, 1)−1 )E3 (t).
Combining the above estimate and (4.22) and using E3 ≥ E1 , we yield
(4.45)
|ρij fij (x)| ≤ Cgij (τ2 µ4 )(x)E3 (t)
Cgij (τ )(x) = min(Cgij,1 (x)τ −1 + Cgij,2 (x)τ1 + Cgij,3 (x)τ1 g1 (0, 1)−1 , C(fij , [1, τ1 , τ1 g1 (0, 1)−1 ])).
Here, since τ1 has been chosen, Cgij only depends on the weight τ2 µ4 . Since ϕg1 grows faster
than ϕ1 in the far field, the above estimates provide better estimates of uA , ∇uA in the far-field.
Now, following the derivations and estimates in Section 4.3, we obtain
(4.46) ∂t (W1,i ϕgi ) + (c̄l x + ū + u) · ∇(W1,i ϕgi ) = −dgi (x)W1,i ϕgi + Bgi (x),
where the damping terms di are given by
(4.47)
dg1 (x) = Td (ϕg1 ) + c̄ω , dg2 (x) = Td (ϕg2 ) + 2c̄ω − ūx , dg3 (x) = Td (ϕg3 ) + 2c̄ω + ūx ,
and Td is defined in (4.4). To estimate the local terms, e.g. η in (4.1), we have
|ϕg1 η| ≤ ϕg1 (ϕ−1 −1 −1 −1 −1 −1
2 ||η1 ϕ2 ||∞ ∧ τ2 ϕg2 ||τ2 ηϕg2 ||∞ ) ≤ ϕg1 (ϕ2 ∧ τ2 ϕg2 )E3 (t),

where a ∧ b = min(a, b). Similarly, we have

|ϕg2 v̄x ξ1 | ≤ ϕg2 |v̄x |(ϕ−1 −1 −1
3 ∧ τ2 ϕg3 )E3 , |ϕg3 ūy η1 | ≤ ϕg3 |ūy |(ϕ−1 −1 −1
2 ∧ τ2 ϕg2 )E3 .

For the nonlocal terms uA , ∇uA , we apply (4.45). To simplify the notation, we introduce an
operator Tu,g similar to Tu in (4.18)
(4.48) Tu,g (f, τ ) = Cg01 (τ )|fx | + Cg01 (τ )|fy |
to control uA · ∇f .
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 55

We obtain the following estimate for the remaining terms

|Bgi (x)| ≤ Agi (x)E3 (t), i = 1, 2, 3,
ϕg1
Ag1 = ϕg1 (ϕ−1 −1 −1
2 ∧ τ2 ϕg2 ) + Tu,g (ω̄, τ2 µ4 ),
ρ01
(4.49) ϕg2 ϕg2
Ag2 = ϕg2 |v̄x |(ϕ−1 −1 −1
3 ∧ τ2 ϕg3 ) + Tu,g (θ̄x , τ2 µ4 ) + (Cg11 |θ̄x | + Cg20 |θ̄y |),
ρ10 ψ1
ϕg3 ϕg3
Ag3 = ϕg3 |ūy |(ϕ−1 −1 −1
2 ∧ τ2 ϕg2 ) + Tu,g (θ̄y , τ2 µ4 ) + (Cg02 |θ̄x | + Cg11 |θ̄y |).
ρ10 ψ1
In the above estimates, Cgij depends on τ2 µ4 (4.45). Now, the inequality (A.6) for |W1,i ϕgi ||∞
with weights (τ2 µ4 , τ2 , τ2 ) reads
(4.50) − dg1 (x) − µ2 τ2 Ag1 ≥ λ, −dg2 (x) − τ2 Ag2 ≥ λ, −dg3 (x) − τ2 Ag3 ≥ λ,
where we choose λ to be the same as that in (4.23). The above constraints involve the param-
eters p71 , p81 , p91 , p92 , µ, τ2 . The parameter p71 has been chosen before we compute (4.42). See
the paragraph above (4.42). We choose the remaining parameters p81 , p91 , p92 , µ2 , τ2 such that
p81 τ2 , τ2 µ4 are as large as possible under the above constraints. From the definitions (4.41) and
(4.44), this allows us to obtain a stronger control of the solution in the far-field using the energy.
It is easy to see that µ4 τ2 Ag1 and τ2 Agi are increasing in τ2 . For example, in τ2 Ag2 , we have
τ2 ϕg2 |v̄x |(ϕ−1 −1 −1 −1 −1
3 ∧ τ2 ϕg3 ) = ϕg2 |v̄x |(τ2 ϕ3 ∧ ϕg3 ).

Moreover, for larger pij , the left hand side of (4.50) usually becomes smaller. In fact, the
damping term di becomes smaller since the weight of the largest power |x|γ increases and |x|γ
contributes less to the damping term if γ is large. See (2.27) for the computation of |x|γ with
larger γ . Moreover, if we neglect τ2−1 ϕ−1
gi part in Agi in (4.49), other quantities are increasing
with respect to pij due to the weights ϕgi (4.41). Due to these monotonicity properties, it is
not difficult to find a set of reasonable parameters (pij , µ4 ) and τ2 .

4.5.2. Weighted L∞ estimate related to the Hölder norm. Recall from Section 2.7.4 that to
control the Hölder norm of ω1 in R2+ , which is used to control ||∇u(ω1 )||C 1/2 , we need to further
−1/2
estimate ||ω1 ψ1 x1 ||L∞ . To simplify the notation, we use ϕ4 = ψ1 |x1 |−1/2 from (4.14). Using
the existing estimates and derivations in Section 4.3, including (4.15), (4.19)-(4.20), we obtain
∂t (ω1 ϕ4 ) + (c̄l x + ū + u) · ∇(ω1 ϕ4 ) = −d4 (x)ω1 ϕ4 + B1 (x)ϕ4 ,
where the damping term d4 is given below and B1 ϕ4 defined in (4.5) satisfies the following
estimate
ϕ4 ϕ4
d4 (x) , d1 (ϕ4 ) = Td (ϕ4 ) + c̄ω , |B1 (x)ϕ4 | ≤ ||η1 ϕ2 ||∞ + Tu (ω̄)||ω1 ϕ1 ||∞ .
ϕ2 ρ10
The operator Td , d1 (·) is defined in (4.4). We want to include ||ω1 ϕ4 ||∞ in a new energy E4 with
E4 (t) ≥ τ1−1 [ω]C 1/2 .
x

We have the factor τ1−1 since the previous energy E3 already controls τ1−1 [ω]C 1/2 (R2 . In view
x ++ )
of the estimate (2.32) and Section 2.7.4, we define
√
E4 (t) = max(E3 (t), τ1−1 2||ωϕ4 ||∞ ).
From (2.32) and E3 ≥ τ1−1 [ω]C 1/2 (R2 , we obtain the above inequality for E4 (t). We do not use
x ++ )
a larger weight for ||ωϕ4 ||∞ since it is not necessary,
√ and it is easier to establish the condition
(A.6) for ||ωϕ4 ||∞ with a smaller weight µi = τ1−1 2, which reads
√
2 ϕ4 ϕ4
(4.51) − d4 − ( + Tu (ω̄)) ≥ λ.
τ1 ϕ2 ρ10
We choose the same λ as that in (4.23).
56 JIAJIE CHEN AND THOMAS Y. HOU

4.6. Estimate of some linear functionals. In the previous sections, we have performed the
weighted L∞ and C 1/2 estimates on W1,i for the main equations (4.1) and established the
stability estimates provided that (4.23), (4.50), (4.51) hold. To close the energy estimates of
(2.75), we need to futher estimate the residual operators R (2.74). The error part related to
the approximate solution constructed numerically, e.g. F̂i (0) − F̄i , will be estimated in Section
7. To control R, we need to control the functional ai (W ) and anl,i (W ).
For the linear functional ai (W ), we have two types. The first type is cω (ω1 ) from K1i (ω1 )
(2.66). The second type is from Ki2 (ω1 ) (2.68) for the approximation of ũ, ∇u ˜ (2.88), (2.82),
2
(2.89). For anl,i defined in (2.72), we need to control cω (W1 + Ŵ2 ) and ∂ (W1 + Ŵ2 )(0). For
the second type of term, we will represent it as
Z
ω(y)p(y)dy
R2++

for some function p(y) that has a fast decay, e.g. it has a decay rate faster than |x|−3 . Then we
estimate it directly using the norms ||ωϕ1 ||∞ , ||ωϕg1 ||∞ in the energy. See (4.55) for an example.
Next, we consider how to estimate cω (ω1 ), cω (ω1 + ω̂2 ), ∂ 2 (W1 + Ŵ2 )(0).

4.6.1. Controlling cω , ωxy (0) and θxxy (0). Denote

(4.52) ϕM 1 (x) , max(ϕ1 (x), τ2 µ4 ϕg1 (x)), ϕM i (x) , max(ϕi (x), τ2 ϕgi (x)), i = 2, 3.
From the definition of the energy (4.21), (4.44), we have the pointwise control
(4.53) |W1,i (x)| ≤ ϕ−1
M i (x)E3 (t), W1,1 = ω1 , Wi,2 = η1 , Wi,3 = ξ1 .
Denote
y1 y2
(4.54) f∗ (y) = .
|y|4
Recall the inner product (2.5). Controlling the normalization factor (2.19)
Z
4 y1 y2 4
cω (ω) = ux (0) = − 4
ω(y)dy = − hω, f∗ i
π R++
2
|y| π
is difficult since the integrand y|y|
1 y2
4 decays slowly (it is not L
1
integrable) and our weight for ω1
is very weak in the far-field. See (4.41) and (4.44). Using (4.53) and a direct estimate, we get
Z
4 y1 y2 −1
(4.55) |cω (ω1 )| ≤ E3 ϕ (y)dy = C1 E3 ,
π R2++ |y|4 M 1
where C1 is about 170 − 200 and is very large. Although this estimate and constant only enter
the energy estimates via the lower order and small terms, e.g. the residual operators and the
nonlinear terms, it requires us to obtain a smaller residual error in the computation. To ease
the computation burden, we seek a more effective estimate based on the evolution of cω (ω1 ).

4.6.2. Controlling of cω (ω1 ). Following [12, 13], we perform the estimates based on the ODE of
cω . Using the main equations (4.1) and dropping the nonlinear and error terms, we can derive
the evolution of hω1 , f∗ i, hη1 , f∗ i
d
hω1 , f∗ i = c̄ω hω1 , f∗ i − h(c̄l x + ū) · ∇ω̄, f∗ i − huA · ∇ω1 , f∗ i + hη1 , f∗ i,
dt
(4.56) d
hη1 , f∗ i = 2c̄ω hη1 , f∗ i − h(c̄l x + ū) · ∇η1 , f∗ i − hūx , η1 f∗ i − hv̄x , ξ1 f∗ i
dt
− huA · ∇θ̄x + ux,A · ∇θ̄, f∗ i,
where h·, ·i defined in (2.5) is the standard inner product on R2++ . Here, we derive the evolution
of hη1 , ϕ∗ i to control it in the first equation. Similar ideas have been used in [12, 13]. Using
integration by parts, we get
Z Z Z
− (c̄l y + ū) · ∇g(y)f∗ (y)dy = g(y)∇ · ((c̄l y + ū)f∗ (y))dy = g(y)ū · ∇f∗ (y)dy,
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 57

where we have used ∇ · (yf∗ (y)) = 0, which is an algebraic property of f∗ (4.54), and ∇ · (ūf∗ ) =
ū · ∇f∗ .
The first terms on the right hand side are damping terms since c̄ω ≈ −1. The advantage of
the above ODE system is that the integrands, e.g. g(y)ū · ∇f∗ , have faster decay than g(y)ϕ∗
since ū grows sublinearly.
For the nonlocal terms involving uA , ∇uA , we apply the estimates (4.45). For the local terms
other than hω1 , f∗ i, hη1 , f∗ i, we use (4.53) to estimate them. Performing energy estimates, we
can estimate hω1 , ϕ∗ i and hη1 , ϕ∗ i.
Improvement. We can further improve the above estimate by decomposing
hω1 , f∗ i = hω1 , χf∗ i + hω1 , (1 − χ)f∗ i, hη1 , f∗ i = hη1 , χf∗ i + hη1 , (1 − χ)f∗ i,
where χ is a smooth cutoff function supported away from the origin. We derive the ODEs
for hω1 , χf∗ i, hη1 , χf∗ i similar to (4.56), and perform energy estimates on these terms. For the
second part hω1 , (1 − χ)f∗ i with integrand supported near 0, we estimate it directly using (4.53)
Z
| ω1 (1 − χ)f∗ | ≤ E3 ||(1 − χ)f∗ ϕ−1
1 ||L1 .

The motivation for the above decomposition is that the R estimate via the ODE system is more
effective to the control the far-field part of the integral ω1 f∗ since the integrand in the ODE
system has faster decay. For the integrand near 0, since our weight ϕ1 is singular and large near
0, we can control it effectively using a direct estimate.
y−ν3
We choose χ(x, y) = χe ( x−ν
ν3 )χe ( ν3 ) with a suitable support size ν3 , where χe is the cutoff
3

function defined in (F.1). Then we perform energy estimate on the above system and show that
1
the condition (A.6) is satisfied by choosing a weight 85 for cω (ω1 ).

4.6.3. Controlling cω (ω). Recall that W1 + Ŵ2 = (ω, η, ξ) is the solution to (2.65). We use
similar ideas to estimate cω (ω). Using (2.65) and focusing on the linear terms, we derive the
ODEs for hω, f∗ i, hη, f∗ i
d
hω, f∗ i = c̄ω hω, f∗ i + cω hf¯cω ,1 , f∗ i − h(c̄l x + ū) · ∇ω, f∗ i − hũ · ∇ω̄, f∗ i + hη, f∗ i,
dt
(4.57) d
hη, f∗ i = 2c̄ω hη, f∗ i + cω hf¯cω ,2 , f∗ i − h(c̄l x + ū) · ∇η, f∗ i − hūx , ηf∗ i − hv̄x , ξf∗ i
dt
− hũ · ∇θ̄x + ũx · ∇θ̄, f∗ i.
Different from (4.56), the above system contains extra terms cω hfcω i , f∗ i, i = 1, 2. For the ω
equation, this term is an additional damping term
4
cω hfcω 1 , f∗ i = − hω, f∗ i · hfcω i , f∗ i ≈ −2.5hω, f∗ i.
π
Thus, we have a better estimate for cω (ω) than cω (ω1 ). For the nonlocal terms in (4.57), we
decompose them into three parts. For example, we have
ˆ 1 ) + ũ(ω̂2 )) · ∇ω̄, f∗ i , I1 + I2 + I3 .
hũ · ∇ω̄, f∗ i = h(ũA (ω1 ) + ũ(ω
For I1 , the estimate is similar to that in Section 4.6.2. For I2 , we can rewrite it as an integral
of ω1 since X
˜ 1) =
û(ω hω1 , fi igi
1≤i≤n
for some functions fi , gi and n. It follows
X X
˜ 1 ), ω̄x f∗ i =
hû(ω hω1 , fi i · hgi , ω̄x f∗ i = hω1 , āi fi i, āi = hgi , ω̄x f∗ i.
1≤i≤n 1≤i≤n

We can apply (4.53) to further estimate it. The advantage of the above estimate is that we
exploit the cancellations among āi fi by performing an argument similar to integration by parts
and rewriting an integral of the nonlocal terms to the local term. For I3 , we use the formula of
ω̂2 to obtain its sharp estimate.
58 JIAJIE CHEN AND THOMAS Y. HOU

For the local terms in (4.57) except for hω, f∗ i, hη, f∗ i, we decompose it into the integrals of
(ω1 , η1 ) and of (ω̂2 , η̂2 ), and then use (4.53) and the formulas of ω̂2 , η̂2 (2.73) to estimate them.
We refer to (2.75) in Section 2.10.4 for the decomposition of ω = ω1 + ω̂2 and η = η1 + η̂2 .
1
Using the above estimates, we can choose a weight 65 for cω (ω).

4.6.4. Controlling ωxy (0), ηxy (0), ξxx (0). To control ωxy (0), ηxy (0), ξxx (0), we first note that
ηxy (0) = ξxx (0) = θxxy (0) since the solution to (2.65) satisfies η = θx , ξ = θy (2.20).
Recall the ODEs for ωxy (0), θxxy (0) in (2.67). Linearizing it around the approximate steady
state and using the normalization conditions (2.19), (2.22), we yield the equations for the per-
turbations
d
ωxy (0) = (−2c̄l + c̄ω )ωxy (0) + θxxy (0) + cω ω̄xy (0) + cω ωxy (0) + ∂¯xy F1 (0),
(4.58) dt
d
θxxy (0) = (−2c̄l + 2c̄ω − ūx (0))θxxy (0) + cω θ̄xxy (0) + cω θxxy (0) + ∂¯xy F2 (0),
dt
where F̄i is defined in (2.2).
Note that the matrix involving ωxy (0), θxxy (0) has negative eigenvalues. After we obtain the
ODE for cω , we can diagonalize the above ODE system and estimate ωxy (0), θxxy (0). Using the
1
above ODEs, we can choose a weight 10 for θxxy (0), and 51 for ωxy (0).
Then we define our final energy E4 (t) to close the estimates
 
E4 (t) , max E3 (t), µ5 |cω (W1 )|, µ6 |cω (ω)|, µ7 |θxxy (0)|, µ8 |ωxy (0)| ,
(4.59) 1 1 1 1
µ5 = , µ6 = , µ7 = , µ8 = .
85 65 10 5
where the energy E3 is defined in (4.44). See also (4.21), (4.25).

4.7. Estimate Ŵ2 and the residual operator. Using the estimates in the previous section
on the time dependent factor, we can control the residual error (2.74). To illustrate the ideas of
the estimate, we want to control the error R and estimate the solution
Xn Z t Xn Z t
(4.60) R= bi (s)R̂i (t − s)ds, G = bi (s)Ĝi (t − s)ds,
i=1 0 i=1 0

for some time-dependent factor bi (s) depending on W1 and W1 + Ŵ2 , where R̂i is the residual
error in solving a specific equation, e.g., R̂i = (∂t − L)F̂i in (2.74), and Ĝi is the numerical
approximation of some solution, e.g. Ĝi = F̂i in (2.73). The difficult part is to control the
time integral part in (2.73), (2.74). Thus, we focus on the above variables. The time-dependent
factor bi (s) can be controlled using the energy estimate
|bi (s)| ≤ ci E4 (s).
Since we will use a bootstrap argument to show that E4 (s) is below some threshold E∗ for all
time, under such an assumption we further have
(4.61) |bi (s)| ≤ ci E∗
for some threshold E∗ to be determined.
We remark that R̂i , Ĝi only depend on the numerical construction. From Section 7, since we
construct Ĝi such that it is piecewise cubic in time, the residual R̂i (t − s) is also piecewise cubic
in time.

4.7.1. Linearity and estimate R and G. To estimate R, G (4.60), we have a crucial observation
that both R and G are linear in the numerical solution Ĝi and residual R̂i .
To control the weighted norm of R, G, we only need to control the piecewise derivatives
bound. Using linearity in Ĝi , R̂i , and (4.61), for α, β ≥ 0, we have
XZ t X Z t
(4.62) |∂1α ∂2β R(t, x)| = bi (s)∂1α ∂2β R̂i (t − s)ds ≤ ci E∗ |∂1α ∂βj R̂i (s)|ds.
i≤n 0 i≤n 0
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 59

Rt
In the above summation, the integral only depends on |∂1α ∂2β Ri (s)| and 0 |∂1α ∂2β Ri (s)|ds
is monotone in time t. Moreover, we completely decouple the numerical solution R̂i and the
time-depend factor bi (t). Since R̂i is a piecewise cubic polynomial in time, we have
X
R̂i (tm + s) = R̂i,m,j sj , s ∈ [0, tm+1 − tm ]
l≤3

in each time interval Im = [tm , tm+1 ] in our construction in Section 7 for some time-independent
residual error R̂i,m,j . The main term is given by Ri,m,0 . Other terms have a small factor s that
is bounded by the time step. Therefore, using the triangle inequality, we yield
X
|∂1α ∂2β Ri (tm + s)| ≤ |∂1α ∂2β R̂i,m,j |sj .
j≤3

Plugging the above estimates in (4.62), we can bound |∂1α ∂2β R(t, x)| in terms of the linear
combination of |∂1α ∂2β R̂i,m,j |,
Z tM X X Z τ
|∂1α ∂2β R̂i (s)|ds ≤ |∂1α ∂2β R̂i,m,j | sj ds,
0 j≤3 i≤M −1 0

where we have used uniform time step tj+1 − tj = τ . Therefore, to control |∂1α ∂2β R(t, x)|, we
only need to track the sum of the discrete residual error |∂1α ∂2β R̂i,m,j | over m.
As a result, once we have an estimate of the piecewise derivatives bound of R̂i,m,j , we can
assemble them to yield the bound for R. We can also control the weighted piecewise bounds
of R using the associated bounds for R̂i,m,j . Then we can apply the methods in Section G to
estimate the weighted L∞ and C 1/2 norm of R rigorously. We can also apply the same ideas to
estimate G in (4.60).
Remark 4.2. Using linearity and theR triangle inequality, we can assemble the estimates for
t
R from the estimates of each mode 0 bi (s)R̂i (t − s)ds. In practice, this means that we can
implement the above estimate for each individual mode completely in parallel.
Although we estimate G (4.60) by applying the triangle inequality and combining the estimate
of different modes, such an estimate does not lead to a constant of O(n) since the different
solutions are large in different regions. In fact, when we construct the approximation terms
for the velocity in Section 8, we apply some partition of unity. The coefficients of different
approximation terms are large in different regions. These coefficients are the initial conditions
for the approximate solution (2.73). We can exploit these properties in the above estimates and
do not obtain a large constant when we combine the estimates of different modes. The same
reasoning applies to the estimate of the residual operator.

4.7.2. Estimate the nonlocal parts. In the estimate, we also need to control the nonlocal terms
of G, e.g. u(Ŵ2 ) appears in the nonlinear terms in (2.69). The ideas and derivations in Section
4.7.1 generalizes to linear operator K. We focus on the estimate of u(G) in (4.60). We can
decompose the velocity as the numerical approximation and the error ε̂i ,
(4.63) u(Ĝi (s)) = u((−∆)φ̂N
i (s)) + u(ε̂i ), ε̂i = Ĝi (s) − (−∆)φ̂N
i (s),

where φ̂N is our numerical solution for the Poisson equation −∆φ̂N = Ĝi , and εi is the error.
Since Ĝi is piecewise cubic in time, we can also construct the associated φ̂N i cubic in time. The
first part is our numerical solution for the velocity u, which is local in φ̂N
i
⊥ N
u((−∆)φ̂N
i ) = ∇ φ̂i .

Applying u to (4.60) and using linearity, we yield

XZ t XZ t XZ t
u(G) = bi (s)u(Ĝi (t − s))ds = bi (s)∇⊥ φ̂N
i (t − s)ds + bi (s)u(ε̂i )ds , I + II.
i≤n 0 i≤n 0 i≤n 0
60 JIAJIE CHEN AND THOMAS Y. HOU

The first part is the main term, which can be estimated using the method in Section 4.7.1.
To control II, using linearity, we rewrite it as follows
XZ t 
II = u bi (s)ε̂i ds .
i≤n 0

We can estimate the velocity by applying the functional inequalities

||u(f )||X ≤ CXY ||f ||Y ,
between two spaces X, Y . Thus to control II, we only need to control the error
XZ t
bi (s)ε̂i ds,
i≤n 0

which is small. Since ε̂i depends on Ĝi , φ̂N i locally (4.63), again, we can apply the method in
Section 4.7.1 to control it. This allows us to control the velocity of G effectively. We also apply
these ideas to estimate the error in solving the Poisson equations in Section 7.
Using the above methods and the methods in Section 7, Appendix G , we can control the
residual operator (2.74), which is much smaller than the damping factors we obtain in the
weighted L∞ and Hölder estimates. The largest term in the residual operator is associated to
the term cω (W1 )f¯cω ,i , i = 1, 2, 3 from (2.66) since we only have a poor estimate for cω (W1 )
|cω (W1 )| ≤ 85E4 (t),
which leads to a large constant c1 = 85 in (4.61). On the other hand, since the functions
f¯cω ,i , i = 1, 2, 3, are smooth, we have a very small residual error (∂t − L)F̂1 , i.e. R̂i in (4.60).
Remark 4.3. We remark that we only need to construct the approximate solution Ĝi (s) in a
finite time T1 since the solution decays over time. See Section 7.6 for more discussions.
For a domain D near the origin, the residual operator enjoys the estimate
max ||Rl,i (W1 )ϕi ||L∞ (D ≤ 0.02E4 (t),
i≤3

where Rl,i is the i−th component of Rl (2.74). Note that the damping factors in the weighted L∞
estimate and Hölder estimate are large. See Figure 5, 7 for an illustration. The weighted estimate
of the residual operator in the bulk of R++
2 is much better since the weight is nonsingular. The
nonlinear part Rnl (2.74) is much smaller.
4.8. Nonlinear estimates. The nonlinear stability estimate of W1 = (ω1 , η1 , ξ1 ) in (2.75) is
relatively simple after we obtain the linear stability. Using the estimates established in Sections
4.3 and 4.5.1, including the L∞ estimates of ∇u, ω1 , η1 , ξ1 , we can control the nonlinear parts
associated to W1 in (2.75). For Ŵ2 , it depends on W1 and can be estimated using the methods
in Section 4.7. We remark that to close the nonlinear stability, we need to check the condition
(A.14).
Recall the nonlinear operators N from (2.1) and N F from (2.71). We focus on the estimate
of the nonlinear part N1 (W1 + Ŵ2 ) − N F1 (W1 + Ŵ2 ) in the ω equation (2.75), which takes the
form
(4.64) J0 , −u(ω) · ∇ω + cω (ω)ω − cω ωxy (0)f¯, f¯ = f¯c ,1 ,
ω

where ω = W1,1 + Ŵ2,1 , and f¯cω ,i is defined in (2.64). To simplify the notation, we introduce
the above f¯. Note that in the weighted L∞ (ϕ1 ) estimate of ω1 , the transport term u(ω) · ∇ω1
leads to a nonlinear term
ϕ−1
1 (u(ω) · ∇ϕ1 )ω1 .
See (4.6) and (4.4). Note that we do not need to estimate the advection in the weighted L∞ (ϕ1 )
estimate
u(ω) · ∇(ω1 ϕ1 ).
Thus, instead of estimating J0 , we need to estimate the following J
(4.65) J = −u(ω) · ∇ω2 + cω (ω)ω − cω ωxy (0)f¯ + ϕ−1 (u(ω) · ∇ϕ1 )ω1 ,
1
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 61

which vanishes cubically near the origin. We have a similar term in the Hölder estimate of W1 .
Next, we focus on the weighted L∞ estimate. The Hölder estimate is similar but contains
more terms. The constants in our estimate of the nonlinear term in the Hölder estimate are
actually smaller due to the factor τ1−1 in the energy E2 (4.25). Our goal is to show that the
nonlinear term can be all bounded by
(4.66) 12000E42 (t),
in the estimate of each term in the energy E4 (4.59), e.g. τ −1 ||ω1 ψ1 ||C 1/2 and µ5 cω (W1 ) where
g1
E4 is defined in (4.59). The above estimate of Ni (x) can be significantly improved in the region
away from the origin. We refine the estimate in the region where the stability factors in the
linear estimate are not too large.
4.8.1. The main term in the near-field. Recall from the discussion in Section 2.10.4 that ω =
ω1 + ω̂2 and ω1 = O(|x|3 ) near 0. Therefore, we have ωxy (0) = ∂xy ω̂20 . We decompose (4.65) as
follows
J = −u(ω) · ∇ω̂2 + cω (ω)ω − cω ωxy (0)f¯ + ϕ−1
1 (u(ω) · ∇ϕ1 )ω1
   
= − u(ω) · ∇ω̂2 + cω ω̂2 − cω ω̂2,xy (0)f¯ + cω ω1 + ϕ−1
1 (u(ω) · ∇ϕ1 )ω1 , J1 + J2 .

The term in our estimate that has the largest constant is J1 due to the poor estimate of cω (ω)
and cω (W1 ) in terms of the energy E4 (4.59).
(4.67) |cω (ω)| ≤ µ6 E4 , |cω (W1 )| ≤ µ5 E4 .
From (4.61) and Section 4.7, in our estimate of ω̂2 , we need to pay a large constant c1 = µ5 . In
the nonlinear estimate of cω ω̂2 , we have a large constant µ5 µ6 , which causes the poor bound in
the near field.
In comparison, for ω1 , by definition, we have
||ω1 ϕ1 ||L∞ ≤ E4 , |cω ω1 ϕ1 | ≤ µ6 E42 ,
where the constant is much smaller. Similarly, the velocity with approximation uA (4.2) also
has size 1. In fact, we have developed many inequalities to show that it is small.
Therefore, we should mainly focus on J1 . Near the origin, we further decompose J1 as follows
J1 = ũ(ω) · ∇ω̂2 + cω (ω)(ω̂2 − xω̂2,x + y ω̂2,y − cω ω̂2,xy (0)f¯)
= ũ(ω1 ) · ∇ω̂2 + ũ(ω2 ) · ∇ω̂2 + cω (ω)(ω̂2 − xω̂2,x + y ω̂2,y − cω ω̂2,xy (0)f¯) , J11 + J12 + J13 ,
where we have used the notation (2.63). The first term is small relative to the second and the
third term. The term ũ(ω1 ) can be estimated similar to that of uA (ω1 ) and the constant in the
estimate of ũ(ω1 ) is of order 1. The reason why we have much better estimate of ũ(ω1 ) than
u(ω1 ) or cω (ω1 ) is that the term (cω x1 , −cω x2 ) captures the slow decay part in the kernel of u.
See the estimate of the kernel in (4.19), (4.20). Since our weights ϕ1 , ϕ1,g for ω1 is strong in the
near-field and the kernel in ũ(ω1 ) has a fast decay, we have a much better estimate for ũ(ω1 )
for x in the near-field.
For the second and the third terms, we can further decompose ω̂2 according to different
modes. The dominated term in Ŵ2 is generated by cω (W1 )f¯cω ,i in (2.66)
Z t
I= cω (W1 )(s)F̂1 (t − s)ds.
0

Using the method in Section 4.7, we can estimate J12 (ω̂2 )+J13 for ω̂2 generated by cω (W1 )f¯cω ,i .
Such term is the main term in the nonlinear estimate in the near-field. Other terms are much
smaller since they do not involve the product of two large constants. In the left figure in Figure
9, we plot the piecewise bounds of (J12 (ω̂2 ) + J13 )ϕ1 . In the right figure, we plot the estimate
of a similar term in the η1 equation. In both cases, such a term is bounded by 6200E42 . We
have a large bound since we multiply the constants µ5 µ6 (4.67). Since the approximate solution
decays, the estimate becomes much better for large x. Other terms in the estimate are much
smaller since it only involves one of these two constants. In particular, we can establish the
62 JIAJIE CHEN AND THOMAS Y. HOU

conservative bounds in the near field. We can obtain a uniform estimate over time since the
solution decays in time and we only construct F̂i in a finite time. See Section 7.6.
For other weighted L∞ or the Hölder estimates, the main terms in the near-field are also
determined by the term generated by cω (W1 )f¯cω ,i in Ŵ2 .

Figure 9. Weighted L∞ estimate of the main terms in the near field. Left
figure: estimate in the ω−equation; right figure: estimate in the η− equation.

4.8.2. The main term in the far-field. Due to the decay of Ŵ2 in the far-field, the nonlinear
terms involving Ŵ2 are small. The main terms in the far-field are from W1 . We focus on a
typical term ux (ω1 )η1 in the nonlinear terms in N2 (2.1). To estimate such a term, we write it
as follows
−ux (ω1 )η1 = −ux,A (ω1 )η1 + ûx (ω1 )η1 ,
where the approximation ûx is constructed in (2.89), (2.82), (2.88). In the far-field, we only
have Cf 0 (x, y)ux (ω1 )(0) in the first approximation (2.82) and the second approximation term
(2.88). For ûx , the term is not large

||Cf 0 (x, y)ux (ω1 )(0)η1 ϕ2 ||∞ ≤ E4 µ5 .

The second approximation term involves Ii − ux (0), which is an integral of ω1 in the near-field.
It can be estimated directly using the pointwise control (4.53).
Since our weight ϕg,1 grows slowly |x|1/16 and we choose a small parameter τ2 µ4 in the
energy E3 (4.44), the leading order behavior in the upper bound (4.53) is (τ2 µ4 )−1 |x|−1/16 with
(τ2 µ4 )−1 ≈ 64. As a result, we have weak control of ω1 and the velocity ∇uA . In particular, we
have the estimate
|uA,x | < 1000E4 ,
using the estimates in Sections 4.3, 4.5.1. It follows

||ϕ1 ux,A (ω1 )η1 ||∞ < ||ux,A ||∞ · ||η1 ϕ1 ||∞ < 1000E42 .

Such a constant is still much smaller than that in (4.66). We remark that in the far-field, we have
much larger stability factors in the linear estimates in Sections 4.3-4.5.1 since the coefficients of
the nonlocal term decay. For example, in the weighted L∞ estimate, for large x, the stability
factors for η1 is larger than 1. Thus, controlling the nonlinear terms in the far-field is relatively
simple.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 63

4.8.3. Estimate near the region with small stability factor. In the weighted L∞ estimate and the
Hölder estimate, we have a stability factor with a minimum about 0.08. Notice that we have a
small stability factor only in the bulk region, x away from the origin and in the near-field. We
need to discuss a few cases more carefully.
In the weighted L∞ , we can have a large constant in the nonlinear estimates, and the stability
factor is not very large for |x| ≤ 0.2, especially for x around 0.1. In such a region, we have a
stability factor ≥ 0.13. We need a tighter bound for the constant in the nonlinear estimate.
In particular, the nonliner term can be bounded by 8000E42 . Thus, to choose the boostrap
threshold and close the argument, we need to consider the constraint
8000E∗2 < 0.13E∗
according to (A.14). For the threshold E∗ = 5 · 10−6 we will choose, the left hand side is 0.04E∗
and the inequality is valid.
We need to check a few cases in a similar way. Note that other cases enjoy similar or better
estimates. We refer the discussion to the upcoming supplementary material.
4.8.4. Estimate of the error terms. Note that the residual error in the weighted norm is very
small in the bulk region. In the weighted L∞ (ϕ1 ) norm, it is smaller than 2 · 10−8 in the near-
field and near the origin. See Figure 2. Such an error is very small compared to the damping
factor. For example, for a bootstrap threshold E∗ = 5 cot 10−6 , we have 2 · 10−8 = 0.004E∗ ,
which is very small compared to the stability terms, e.g. λE∗ , with λ = 0.08.
Near the origin (the first 3 grid points), we have a relative large residual error, which is
about 4 · 10−7 in the weighted L∞ (ϕ1 ) norm. See Figure 2. On the other hand, we design our
weights such that we have relative large damping factors. In particular, the stability factors in
the L∞ (ϕ1 ) estimate of ω1 is larger than 0.5. At the bootstrap threshold E∗ = 5 · 10−6 , we have
0.5 · 5 · 10−6 = 2.5 · 10−6
which is much larger than the weighted residual error near the origin. Similar consideration
applies to the estimates of η1 , ξ1 .
In the Hölder estimate, we have large damping factors near the origin. Thus, it is easy to
check the stability conditions (A.14).
The estimate of the nonlinear terms and the error terms in the estimate of the linear functional
in Section 4.6.3 is simpler since the functional such as cω (W1 ), ωxy (0) is constant in space.
For the error in solving the Poisson equation, i.e. the error between φ̄N and (−∆)−1 ω̄, where
φ̄ is our numerical approximation of the stream function (−∆)−1 ω̄, we use the method in
N

Section 7 to show that it is small enough.

Remark 4.4. We remark that an advantage of the stability condition in (A.14) is that it depends
on the estimate locally. In our estimates, we have small stability factors in the region away from
the origin, which is different from the region with large weighted residual error. We remark
that we do not have the situation in which the weighted residual error is large, constants in the
nonlinear estimates are large, and the stability factors are small. This allows us to close the
nonlinear estimates using Lemma A.3.
4.8.5. Nonlinear stability and finite time blowup. To close the nonlinear estimates, we choose
the threshold E∗ = 5 · 10−6 for the bootstrap argument in Lemma A.3. Using Lemma A.3, we
can obtain that if the initial perturbation satisfies
E4 (ω1 (0), η1 (0), ξ1 (0)) < E∗ ,
then we have
E4 (ω1 (t), η1 (t), ξ1 (t)) < E∗ ,
for all time t > 0. With the estimates of W1 , we can control W̄2 using the estimates in Sections
4.6, 4.7. In particular, we can obtain
||W1,i + Ŵ2,i ||∞ < 1000E∗ , |cω (ω)| < 100E∗ .
Recall the normalization condition (2.19). We also have |ux (0)| = |cω | < 100E∗ .
64 JIAJIE CHEN AND THOMAS Y. HOU

Moreover, since we choose 0 initial condition for Ŵ2 , we have W1 = (ω, η, ξ) = (ω, θx , θy ) at
the initial time. Therefore, we prove the estimates in Theorem 3 . Passing from the stability
analysis to finite time blowup follows the standard rescaling argument [12], [14].

5. Construction of an approximate steady state

Following our previous works with Huang on the De Gregorio model [14] and the Hou-Luo
model [13], we construct the approximate steady state to the dynamic rescaling equations (2.10)
with the normalization conditions (2.11) by solving (2.10) numerically for a long enough time.
The residual error is estimated a-posteriori and incorporated in the energy estimate as a small
error term. It is extremely challenging to obtain an approximate steady state with a sufficiently
small residual error, e.g. of order 10−7 , since the solution is supported on the whole R2+ with
a slowly decaying tail in the far-field, e.g., ω(t, x) ∼ |x|−1/3 for large x. See (5.1). If we solve
(2.10) in a very large domain to capture the far-field behavior of the solution, we have to deal
with the relatively large round-off errors in the computation. To overcome these difficulties, we
follow [13] to use a combination of numerical computation and a semi-analytic construction.

5.1. Far-field asymptotics. Let (r, β) be the polar coordinate in R+ 2 2 1/2

2 : r = (x + y ) ,β =
arctan(y/x). It has be observed in [13] that the approximate steady state (2.10) enjoys the
following asymptotics
cω 1
(5.1) ω(r, β) ∼ g1 (β)rα , θ(r, β) ∼ g2 (β)r1+2α , α= < 0, α ≈ − ,
cl 3
in the far-field for some angular profiles g1 (β), g2 (β), under the mild assumption that ω decays
for large |x|, cl > 0, and cω > 0. These conditions are satisfied by the blowup solutions [54, 55].
In fact, if ω decays for large |x|, the velocity u = ∇⊥ (−∆)−1 ω has a sublinear growth: u(x)
r →
0 as r → ∞. Note that x · ∇ = r∂r . Passing to the polar coordinate (r, β), r = |x|, β = arctan xx12
and dropping the lower order terms, we yield
cl r∂r ω(r, β) = cω ω + θx + l.o.t., cl r∂r θ(r, β) = (2cω + cl )θ + l.o.t..
Assume that ω(r, b) = rk g1 (β), θ(r, β) = rl g2 (β). Using the above equations and matching the
power, we obtain the asymptotic relation (5.1). Thus, we represent the approximate steady
state as follows
(5.2) ω̄ = ω̄1 + ω̄2 , θ̄ = θ̄1 + θ̄2 , ω̄1 = χ(r)rα g1 (β), θ̄1 = χ(r)r1+2α g2 (β),
where χ(r) is the radial cut-off function defined in (F.4). The crucial first part is constructed
semi-analytically, and it captures the far-field asymptotic behavior of the approximate steady
state. The second part has a much faster decaying rate, and we construct it using numerical
computation with a piecewise sixth order B-spline.

5.2. Angular profiles and the representation. Due to symmetry in x, we compute (2.10)
in a domain [0, L]2 with L ≈ 1013 . Since θ(t, x, y) vanishes quadratically on x = 0, instead of
using θ in our computation, we consider ζ(t, x, y) = x1 θ(t, x, y). Then ζ is odd in x, and its
equation can be derived by dividing the θ equation by x.
In the case without semi-analytic part, we represent the numerical solution (ω, ρ) using a
piecewise 6th order B-spline in x and y, e.g.
X
(5.3) ω(t, x, y) = ai,j Bi (x)Bj (y)
i,j

where Bi (x) is the B-spline basis (E.2). For ψ, we represent it using a piecewise B-spline with
additional weight ρp (y) vanishing on the boundary y = 0 to enforce the no-flow boundary
condition ψ(x, 0) = 0. See more details about the representation in Appendix E.1. Note that
similar representations based on piecewise B-splines have been used in [54]. Given the grid
point values of ω, we obtain the coefficents of the variable ω by solving the linear equations
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 65

(5.3) for (x, y) on the grid and using suitable extrapolation in the far-field. After we obtain the
coefficients ai,j , we compute the derivatives of ω using the basis functions
X
∂xi ∂yj ω(t, x, y) = ai,j ∂xi Bi (x)∂yj Bj (y).
i,j

Similar consideration applies to ζ. We solve the Poisson equations

(5.4) − ∆ψ = ω
using B-spline based finite element method. After we obtain the B-spline coefficients for ψ, we
compute its derivatives by taking derivatives on the basis functions.
In the temporal variable, we use a second order Runge-Kutta method to update the PDE.
To construct the decomposition in (5.2), firstly, we obtain the exponent α1 and construct
the angular profile and the semi-analytic part ω̄1 , θ̄1 in (5.2). Then, using ω̄1 , θ̄1 , we refine the
construction ω̄2 , θ̄2 in (5.2).

5.2.1. Fitting the angular profile and the exponent. We need to find the angular profiles in the
semi-analytic parts in (5.2). Firstly, we solve (2.10) numerically using the above method without
the semi-analytic part, i.e. ω̄1 = 0, θ̄1 = 0, to obtain an approximate steady state (ω̄, ζ̄), θ̄ = xζ̄.
Using the ansatz in (5.2) and fitting the angular part of the far-field of r−α1 ω1 , r−1−2α1 θ̄ =
r−2α1 cos β · ζ̄ with exponent α1 = c̄c̄ωl (5.1), we find the following approximate profiles

a11 β̃(1 + a15 β̃ 2 ) π a21 cos2 β(1 + a25 sin β)

g10 (β) = , β̃ = −β, g20 (β) = ,
2 2/3 2
(β̃ + a12 ) + a13 β̃ + a14 2 (cos2 β + a22 )2/3 + a23 + a24 cos2 β
for some parameters aij . We have the factor π2 − β since ω is odd in x and g10 (β) is odd with
respect to β = π/2. Similarly, we add the factor cos2 β in g20 (β) since θ(x, 0) = 0 and θx (x, 0) is
odd in x. After we find the above analytic formulas, we further approximate the above profiles
by piecewise 8th order B-splines (E.2) Bi with k = 8
X
gj (β) = bji Bi (x), j = 1, 2,
1≤i≤n

for some coefficients bji . We further use the B-spline to represent the angular profiles for the
following reason. To verify that the approximate steady state (ω̄, θ̄) has a small residual error,
we need to estimate the high order derivatives of ω̄, θ̄, e.g. 6-th order. However, the high order
derivatives of the above analytic forms are very complicated, and are difficult to estimate. For
solution represented by piecewise polynomials, we have a systematic approach to estimate it.
In 1D, the estimate follows our previous work with Huang [13, 14]. Once we obtain gi (β), we
construct the semi-analytic part
(5.5) ω̄10 = χ(r)rᾱ1 g1 (β), θ̄10 = χ(r)r1+2ᾱ2 g2 (β), ζ̄10 = θ̄10 x−1 .
To compute the semi-analytic part of the stream function, we follow the ideas outlined in [13].
Given the asymptotic behavior of ω̄ in (5.5), the far-field asymptotic behavior of φ = (−∆)−1 ω̄10
is r2+ᾱ1 f (β) for some profile f (β). We construct f (β) by solving
−∆(r2+ᾱ1 f (β)) = rᾱ1 g1 (β)
with boundary condition f (0) = f (π/2) = 0 due to the Dirichlet boundary condition ψ(x, 0) = 0
and the odd symmetry for the solution ω. In the polar coordinate, the above equation is
equivalent to
(5.6) (−∂β2 − (2 + ᾱ1 )2 )f (β) = g1 (β), f (0) = f (π/2) = 0.
We represent f (β) using a weighted 8th order B-spline and solve the above elliptic equations
using the finite element method. Then, we construct the semi-analytic part for ψ as follows
(5.7) ψ̄10 = χ(r)r2+ᾱ1 f (β).
66 JIAJIE CHEN AND THOMAS Y. HOU

5.2.2. Refinement. We use the semi-analytic profile (5.5) to capture the far-field contribution of
ω̄, ζ̄. Note that in this step, we do not update the angular profile nor the exponent in (5.5).
Given the grid point values of ω(t, x, y), we first update the constant c(t) such that c(t)ω̄10
best approximate ω(t, x, y) in the far-field. Then we represent ω2 (t, x, y) = ω(t, x, y) − c(t)ω̄10
using the B-spline (5.3). In other words, we interpolate the grid point values using the represen-
tation c(t)ω̄10 + ω2 (t, x, y), where ω2 is a piecewise polynomial in the compact domain. Similar
consideration applies to ζ. To update the stream functions φ, we use c(t)φ̄10 to capture the
far-field of φ and then construct the near-field part by solving
(5.8) − ∆(φ2 + c(t)φ̄10 ) = ω2 + c(t)ω̄10 , or − ∆φ2 = ω2 + c(t)(ω̄10 + ∆φ̄10 ).
Then the stream function is represented as φ2 + c(t)φ̄10 .
Let us motivate the above decomposition to construct the stream function over (5.4). If we
use (5.4), the source term ω has a slow decay rα1 ≈ r−1/3 . Since the domain is very large, we
have to use an adaptive mesh to discretize the domain, which leads to a poor condition number of
the stiffness matrix in (5.4). Thus, solving (5.4) can have a significant round-off error. In (5.4),
since the semi-analytic part c(t)ω̄10 captures the asymptotic behavior of ω(t, x, y), ω2 is much
smaller than ω in the far-field. By definition of ω̄10 , φ̄10 (5.5)-(5.7), the far-field of ω̄10 + ∆φ̄10
is about εr−1/3 with a small constant ε. Hence, the far-field of the source term in (5.8) is much
smaller than ω(t, x), which enables us to overcome the significant round-off error. We remark
that similar technique has been used in the Hou-Luo model [13] to overcome the significant
round-off errors. The above decomposition is a generalization of the method in [13] to 2D. We
refer to [13] for the more motivations and the difficulties caused by the round-off error.
After we obtain the stream function, we can update the PDE using the second order Runge-
Kutta method. We stop the computation at time t∗ if the residual error on the grid points is
about the round-off error. Then we finalize the semi-analytic part in (5.2) as
(5.9) ω̄1 = c̄1 ω̄10 , θ̄1 = c̄2 θ̄10 , φ̄1 = c̄3 φ̄10 ,
where c̄1 ω̄10 , c̄2 ζ̄10 best approximate ω(t∗ , x, y), ζ(t∗ , x, y) in the far-field, respectively. We con-
struct ω̄2 , θ̄2 = xζ̄2 in (5.2) by interpolating the grid point values of ω − ω̄1 , θ − θ̄1 and applying
a low-pass filter to the solution to reduce the round off error.
We refer more details of the computation to the upcoming supplementary material.
In Appendix E, we estimate the derivatives of the approximate steady state rigorously, which
will be used to verify the residual error.

5.3. Rigorous verification of the numerical values. In this subsection, we describe some
main ideas how we use computer-assisted proofs in our stability analysis.
(1) As we discussed in the Introduction and Section 2, the most challenging and essential
part in the proof is the weighted L∞ and weighted C 1/2 linear stability analysis established in
Section 4, since there is no small parameter and the linearized equations are complicated.
(2) The weighted L∞ linear stability estimates can be seen as a-priori estimates on the
perturbation, and we proceed to perform C 1/2 estimates in a similar manner and establish the
nonlinear energy estimate for some energy E(t) of the perturbation
d 2
(5.10) E ≤ CE 3 − λE 2 + εE.
dt
Here, −λE 2 with λ > 0 comes from the linear stability, CE 3 with some constant C(ω̄, θ̄) > 0
controls the nonlinear terms, and ε is the weighted norm of the residual error of the approximate
steady state. To close the bootstrap argument E(t) < E ∗ with some threshold E ∗ > 0, a
sufficient condition is that ε < ε∗ = λ2 /(4C), which provides an upper bound on the required
accuracy of the approximate steady state.
The essential parts of the estimates in (1), (2) are established based on the grid point values of
(ω̄, θ̄) constructed using a moderate fine mesh. These parts do not involve the lengthy rigorous
verification in the supplementary material. These estimates already provide a strong evidence
of nonlinear stability.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 67

A significant difference from this step and step (1) is that we have a small parameter ε. As
long as ε is sufficiently small, thanks to the damping term −λE 2 established in step (1), we can
afford a large constant C(ω̄, θ̄) in the estimate of the nonlinear terms and close the nonlinear
estimates. We can complete all the nonlinear estimates in this step.
(3) We described how to construct an approximate steady state with residual error below a
required level ε∗ in the previous subsections. To achieve the desired accuracy, we need to solve
the dynamic rescaling equations for a sufficiently long time using a fine mesh. The difficulty of
the construction depends on the target accuracy ε∗ . Here, the mesh size plays a role similar to a
small parameter that we can use. In practice, the profile (ω̄1 , θ̄1 ) constructed using a moderate
fine mesh Ω1 is close to the one (ω2 , θ2 ) constructed using a finer mesh Ω2 with higher accuracy.
As a result, the constants C(ω̄, θ̄) and λ that we estimate in (5.10) using different approximate
steady states (ωi , θi ) are nearly the same. This refinement procedure allows us to obtain an
approximate steady state, based on which we close the nonlinear estimates (5.10). We refer
more discussion of this philosophy to [14].
(4) Finally, we follow the standard procedure to perform rigorous verification on the estimates
to pass from the grid point value to its continuous counterpart. In the verification step, we can
evaluate the approximate steady state on a much finer mesh Ω3 with many more grid points so
that they almost capture the whole behavior of the solution. Then, we use the regularity of the
solution to pass from finite grid points to the whole upper half plane. In this procedure, the
mesh size in Ω3 plays a role similar to a small parameter that we can exploit.
In summary, in steps (2)-(4), we can take advantage of a small parameter which can be either
the small error or the small mesh size, while there is no small parameter in step (1). Though
these three steps could be technical, they are relatively standard from the viewpoint of analysis.
We remark that the approach of computer-assisted proof has played an important role in the
analysis of many PDE problems, especially in computing explicit tight bounds of complicated
(singular) integrals [7, 21, 37] or bounding the norms of linear operators [4, 33]. We refer to [36]
for an excellent survey on computer-assisted proofs in establishing rigorous analysis for PDEs.
Our approach to establish stability analysis with computer assistance is different from existing
computer-assisted approach, e.g. [6], where the stability is established by numerically tracking
the spectrum of a given operator and quantifying the spectral gap. The key difference between
their approach and ours is that we do not use direct computation to quantify the spectral gap
of the linearized operator since the linearized operator in our case is not compact.

6. Finite time blowup of 3D axisymmetric Euler equations with solid boundary

In this section, we prove the finite blowup of the axisymmetric Euler equations with smooth
initial data and boundary. We will follow the same proof strategy as in our previous work [12].
We first review the setup of the problem. In Section 6.1, we reformulate the 3D Euler equations
using the dynamic rescaling formulation and discuss the connection between the 3D Euler and 2D
Boussinesq; see e.g. [56]. In Section 6.2, we establish the localized elliptic estimates. In Section
6.4, we will construct initial data and control the support of the solution under some bootstrap
assumptions. With these estimates, the rest of the proof follows essentially the nonlinear stability
analysis of the 2D Boussinesq equations and we will sketch the part of the analysis that is different
from the 2D Boussinesq equations.
Notations. In this section, we use x1 , x2 , x3 to denote the Cartesian coordinates in R3 , and
q
(6.1) r = x21 + x22 , z = x3 , ϑ = arctan(x2 /x1 )

to denote the cylindrical coordinates.

Let u be the axi-symmetric velocity and ω = ∇ × u be the vorticity vector. In the cylindrical
coordinates, we have the following representation

u(r, z) = ur (r, z)er + uθ (r, z)eθ + uz (r, z)ez , ω = ω r (r, z)er + ω θ (r, z)eθ + ω z (r, z)ez ,
68 JIAJIE CHEN AND THOMAS Y. HOU

where er , eθ and ez are the standard orthonormal vectors defining the cylindrical coordinates,
x1 x2 x2 x1
er = ( , , 0)T , eθ = ( , − , 0)T , ez = (0, 0, 1)T ,
r r r r
p
2 2
and r = x1 + x2 and z = x3 .
We study the 3D axisymmetric Euler equations in a cylinder D = {(r, z) : r ∈ [0, 1], z ∈
T}, T = R/(2Z) that is periodic in z. The 3D axisymmetric Euler equations are given below:
ωθ ωθ ωθ 1
(6.2) ∂t (ruθ ) + ur (ruθ )r + uz (ruθ )z = 0, ) + ur ( )r + uz ( )z = 4 ∂z ((ruθ )2 ).
∂t (
r r r r
The radial and axial components of the velocity can be recovered from the Biot-Savart law
1 1 1
(6.3) − (∂rr + ∂r + ∂zz )φ̃ + 2 φ̃ = ω θ , ur = −φ̃z , uz = φ̃r + φ̃
r r r
with a no-flow boundary condition on the solid boundary r = 1
(6.4) φ̃(1, z) = 0
and a periodic boundary condition in z.
We consider solution ω θ with odd symmetry in z, which is preserved by the equations dy-
namically. Then φ̃ is also odd in z. Moreover, since φ̃ is 2-periodic in z, we obtain
(6.5) φ̃(r, 2k − 1) = 0. for all k ∈ Z
This setup of the problem is essentially the same as that in [54, 55].
Due to the periodicity in z direction, it suffices to consider the equations in the first period
D1 = {(r, z) : r ∈ [0, 1], |z| ≤ 1}. We have the following pointwise estimate on φ̃, which will be
used to estimate φ̃ away from the supp(ω θ ) in Section 6.2.
Lemma 6.1. Let φ̃ be a solution of (6.3)-(6.4), and ω θ ∈ C α (D1 ) for some α > 0 be odd in z
with supp(ω θ ) ∩ D1 ⊂ {(r, z) : (r − 1)2 + z 2 < 1/4}. For 41 < r ≤ 1, |z| ≤ 1, we have
Z  
|φ̃(r, z)| . |ω θ (r1 , z1 )| 1 + | log((r − r1 )2 + (z − z1 )2 )| r1 dr1 dz1 .
D1

If the initial data u of (6.2)-(6.4) is non-negative, uθ remains non-negative before the blowup,
θ

if it exists. Then, uθ can be uniquely determined by (uθ )2 . We introduce the following variables
(6.6) θ̃ , (ruθ )2 , ω̃ = ω θ /r.
We reformulate (6.2)-(6.4) as
1
∂t θ̃ + ur θ̃r + uz θ̃z = 0, ∂t ω̃ + ur ω̃r + uz ω̃z = θ̃z ,
(6.7) r4
1 1 1
−(∂r2 + ∂r + ∂z2 − 2 )φ̃ = rω̃, φ̃(1, z) = 0, ur = −φ̃z , uz = φ̃ + φ̃r .
r r r
6.1. Dynamic rescaling formulation. We introduce new coordinates (x, y) centered at r =
1, z = 0 and its related polar coordinates
(6.8) x = Cl (τ )−1 z, y = (1 − r)Cl (τ )−1 ,
where Cl (τ ) is defined below (6.11). By definition, we have
(6.9) z = Cl (τ )x, r = 1 − Cl (τ )y.
We consider the following dynamic rescaling formulation centered at r = 1, z = 0
θ(x, y, τ ) = Cθ (τ )θ̃(1 − Cl (τ )y, Cl (τ )x, t(τ )),
(6.10) ω(x, y, τ ) = Cω (τ )ω̃(1 − Cl (τ )y, Cl (τ )x, t(τ )),
φ(x, y, τ ) = Cω (τ )Cl (τ )−2 φ̃(1 − Cl (τ )y, Cl (τ )x, t(τ )),
Rτ
where Cl (τ ), Cθ (τ ), Cω (τ ), t(τ ) are given by Cθ = Cl−1 (0)Cω2 (0) exp 0 cθ (s)dτ ,

(6.11) Z τ  Z τ  Z τ
Cω (τ ) = Cω (0) exp cω (s)dτ , Cl (τ ) = Cl (0) exp −cl (s)ds , t(τ ) = Cω (τ )dτ,
0 0 0
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 69

and the rescaling parameters cl (τ ), cθ (τ ), cω (τ ) satisfy cθ (τ ) = cl (τ ) + 2cω (τ ). We remark that

Cθ (τ ) is determined by Cl , Cω via Cθ = Cω2 Cl−1 . We have this relation due to the same reason
as that of (2.9). We choose (r, z) = (1, 0) as the center of the above transform since the singular
solution is concentrated near this point. Since we rescale the cylinder D1 = {(r, z) : r ≤ 1, |z| ≤
1}, from (6.8), the domain for (x, y) is
(6.12) D̃1 , {(x, y) : |x| ≤ Cl−1 , y ∈ [0, Cl−1 ]}.
We have a minus sign for ∂y
∂y θ = −Cθ Cl (τ )θ̃r , ∂y ω = −Cω Cl (τ )ω̃r , ∂y φ = −Cω Cl (τ )−1 φ̃r .
Let (θ̃, ω̃) be a solution of (6.7). It is easy to show that ω, θ satisfy
1
θt + cl x · ∇θ + (−ur )θy + uz θx = cθ θ, ωt + cl x · ∇ω + (−ur )ωy + uz ωx = cω ω + θx .
r4
The Biot-Savart law in (6.7) depends on the rescaling parameter Cl , τ
1 1 1
−(∂xx + ∂yy )φ + Cl ∂y ψ + 2 Cl2 φ = rω, ur (r, x) = −φx , uz (r, x) = Cl (τ )φ − φy ,
r r r
where r = 1 − Cl (τ )y (6.9). We introduce u = uz , v = −ur . Then, we can further simplify
1 − r4
θt + (cl x + u · ∇)θ = cθ θ, ωt + (cl x + u · ∇)ω = θx + θx ,
(6.13) r4
1 1 1
− (∂xx + ∂yy )φ + Cl ∂y φ + 2 Cl2 φ = rω, u(x, y) = −φy + Cl φ, v = φx ,
r r r
with boundary condition φ(x, 0) ≡ 0. If Cl is extremely small, we expect that the above
equations are essentially the same as the dynamic rescaling formulation (2.10) of the Boussinesq
equations. We look for solutions of (6.13) with the following symmetry
ω(x, y) = −ω(−x, y), θ(x, y) = θ(−x, y).
Obviously, the equations preserve these symmetry properties and thus it suffices to solve (6.13)
on x, y ≥ 0 with boundary condition φ(x, 0) = φ(y, 0) = 0 for the elliptic equation.
We now state a more precise version of Theorem 2 below.
Theorem 4. Let (θ̄ν , ω̄ν , c̄l , c̄ω ) be the approximate self-similar profile constructed in Section
6.4.2 and E∗ = 5 · 10−6 . For even initial data θ0 and odd ω0 of (6.13) compactly supported with
size S(0) to be defined in Definition 6.2 and a small perturbation to (θ̄, ω̄) with
E(ω0 − ω̄, θ0,x − θ̄x , θ̄0,y − θ̄y ) < E∗ ,
where E is defined in (2.13), there exists a constant C(S(0)) depending on such that if the initial
rescaling factor Cl (0) (6.11) satisfies Cl (0) < C(S(0)), we have
(6.14) ||ω − ω̄||L∞ , ||θx − θ̄x ||L∞ , ||θy − θ̄y ||∞ < 103 E∗ , |ux (t, 0) − ūx (0)|, |c̄ω − cω | < 100E∗
for all time. In particular, we can choose smooth initial data ω0 , θ0 ∈ Cc∞ in this class with
finite energy ||u0 ||L2 < +∞ such that the solution to the physical equations (1.4)-(1.6) with these
initial data blows up in finite time T .
We need to choose a small rescaling factor Cl (0) so that the solution in the physical space is
confined in the cylinder, which is not scaling invariant.
6.2. The elliptic estimates. In this subsection, we follow the ideas in [12] to estimate the
elliptic equation with time-dependent coefficients in (6.13). We first estimate φ away from
supp(ω). Then we localize the elliptic equation and perform weighted L∞ and Hölder estimate.
We will show that within the support of ω, θ, the estimates for the velocity are the same as those
in the 2D Boussinesq equations up to a lower order term, which can be made arbitrarily small.
Throughout this Section, we assume that ω(x, y) is odd in x.
Definition 6.2. We define the size of support of (θ, ω) of (6.13)
S(τ ) = essinf{ρ : θ(x, y, τ ) = 0, ω(x, y, τ ) = 0 for x2 + y 2 ≥ ρ2 }.
70 JIAJIE CHEN AND THOMAS Y. HOU

After rescaling the spatial variable, the support of (θ̃, ω̃) of (6.7) satisfies
supp θ̃(t(τ )), supp ω̃(t(τ )) ⊂ {(r, z) : ((r − 1)2 + z 2 )1/2 ≤ Cl (τ )S(τ )}.
We will construct initial data of (6.13) with compact support S(0) < +∞ and follow [12] to
prove that Cl (τ )S(τ ) remains sufficiently small for all τ > 0. We assume S(τ ) > 1 in the
following derivation. Otherwise, we can just add 1 in the above definition. In fact, S(τ ) can be
very large.
Remark 6.3. There are several small parameters Cl (τ ), Cl (τ )S(τ ) in the following estimates. We
will choose Cl (0) to be very small at the final step of the proof. This allows us to prove that
Cl (τ ), Cl (τ )S(τ ) are very small. One can essentially regard Cl (τ ) ≈ 0. Recall the relation (6.9)
about r. In the support of the solution, we have r = 1 − Cl ρ sin(β) ≈ 1. We treat the error
terms in these approximations as small perturbations.
The elliptic equation in (6.13) contains the first order term 1r Cl ∂y ψ, which leads to a few
technical difficulties in the ellilptic estimate. To overcome it, we nultiply the equation with an
integrating factor r1/2 . Using ∂y r1/2 = −Cl r−1/2 /2, ∂yy r1/2 = 14 Cl2 r−3/2 ,
Cl 1
∂yy (φr1/2 ) = r1/2 ∂yy φ + 2∂y φ∂y r1/2 + φ∂yy r1/2 = r1/2 ∂yy φ − ∂y φ + Cl2 r−3/2 φ,
r1/2 4
we can rewrite (6.13) as follows
aCl2 1/2 3
−∆(φr1/2 ) + 2
φr = ωr3/2 , a = .
r 4
Note that within the support of ω, θ, r, r−1 are smooth. Once we obtain the estimate of φr1/2 ,
we can recover the estimate of φ. We rewrite the above equation as follows
aCl2
(6.15) − ∆φ1 = Ω1 + φ1 , Ω1 = ωr3/2 , φ1 = φr1/2 .
r2
Our goal is to show that φ1 and φ enjoy estimates similar to those for (−∆2D )−1 ω, then we
can generalize the analysis for 2D Boussinesq to 3D Euler equations.
6.2.1. Estimate of φ away from the support. To localize the elliptic equations, we first estimate
φ away from the support of the solution. Based on Lemma 6.1, we have the following estimate.
Lemma 6.4. Suppose that the assumptions in Lemma 6.1 hold true. Let S(τ ) be the support
size of ω(τ ), θ(τ ). Assume Cl (τ )S(τ ) < 41 . For any |x| > 2S and β ∈ [0, 1), the solution to
(6.13) satisfies
|φ(x)| . ||ω(1 + |x|β )||L∞ (1 + | log(Cl |x|)|)S 2−β .
Since x is away from the support, the proof follows from the rescaling relation (6.10) and the
estimate in Lemma 6.1 by putting ω(y)(1 + |y|β ) in L∞ . We defer the proof to Appendix D.2.
6.2.2. Localize the elliptic equation. We will take advantage of the fact that Cl (τ )S(τ ) can be
extremely small and localize the elliptic equation. Firstly, we assume that Cl (τ )S(τ ) < 41 . Recall
the relation (6.9) about r. Within the support, we have r = 1 − Cl y ≥ 34 , r−1 . 1.
Let χ(·) : R+
2 → [0, 1] be a smooth cutoff function even in x1 , such that χ(x) = 1 for |x| ≤ 1,
χ(x) = 0 for |x| ≥ 2. It is easy to verify that
(6.16) |∇k χ(x/R)| . R−k 1R≤|x|≤2R ,
for 1 ≤ k ≤ 5. Next, we choose several radii and define the related cutoff function
(6.17) Ri = 4−i Cl−1 , χi (x) , χ(x/Ri ), i ≤ 5.
By definition, we have χi = 1 in the support of χi+1 . Multiplying (6.15) with χi , we obtain
the equation of φ1 χi
aCl2
(6.18) − ∆(φ1 χi ) = Ω1 χi − Z1,i − Z2,i , Z1,i , 2∇φ1 · ∇χi + φ1 ∆χi , Z2,i , φ1 χi ,
r2
with boundary condition
(φ1 χi )(0, y) = 0, (φ1 χi )(x, 0) = 0.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 71

After we localize the elliptic equation, (6.18) can be seen as an elliptic equation in R+ 2 with
compactly supported source term. Since the solution φ1 χi decays for large |x|, it agrees with
the solution defined by the Green function log(|x − y|) in the upper half space:
(6.19) Z Z
1 1
((−∆)−1 f )(x) = − (log |x−y|−log |(x1 −y1 , x2 +y2 )|)f (y)dy = − log |x−y|F (y)dy.
2π R+ 2
2π R2
where F is the odd extension of f from R+
2 to R2 . Similar formula also holds for ∇(−∆)
−1
f
−
Z
1 x i yi
(6.20) ∂i (−∆)−1 f = − F (y)dy.
2π R2 |x − y|2

Ideas of the estimates. We will assume that Ω1 is in a suitable weighted L∞ and Hölder
space. Our goal is to show that the terms on the right hand side of (6.18) except for Ωχi are
very small in such a space. Then we can obtain the estimate for φ1 χi by inverting −∆, which
is similar to that of (−∆)−1 (Ω1 χi ). We will also require that the support satisfies
(6.21) S(τ ) < Ri , or Cl S < 2−i , i≤5
so that Ω1 χi = Ω1 . We will choose Cl S to be sufficiently small.
We need to estimate the L∞ norm of ∇φ1 and its Hölder norm. We will first estimate ∇φ1
for |x| ≤ R2 , and then ∇2 φ1 for |x| ≤ R3 . Once we obtain the estimates of ∇φ1 , ∇2 φ1 , due to
the small parameters on the right hand side of (6.18) and the decay of the solution, we establish
the desired estimate.
We need several weighted estimates of the Laplacian in R+ 2 , which are the consequences of
the estimates of the symmetrized kernel in Appendix D.
Lemma 6.5. Suppose that −∆φ = ω in R++ 2 , ω is odd, and φ satisfies the Dirichlet boundary
condition. For α > 0 and β ∈ (0, 1), we have
|∇φ| .α,β |x| ∧ |x|1−β ||ω(|x|−α + |x|β )||L∞ .
For α ∈ (0, 2), we have
|∇(φ − φxy (0)x1 x2 )| .α,β |x|1+α ||ω(|x|−α + |x|β )||L∞ ,
|φ − φxy (0)x1 x2 | .α,β |x|2+α ||ω(|x|−α + |x|β )||L∞ .
For α ∈ (2, 3], we have
|∇(φ − φxy (0)x1 x2 )| .α,β |x|3 ||ω(|x|−α + |x|β )||L∞ ,
|φ − φxy (0)x1 x2 | .α,β |x|4 ||ω(|x|−α + |x|β )||L∞ .
1
We will mostly use α = 1, 2, 2.9, β = 16 relates to the weight ϕ1 , ϕg,1 for ω1 in the 2D
Boussinesq equations (C.2), (C.3). To obtain the above estimates, we use the kernel related to
(−∆)−1 and the estimates of the symmetrized kernels in Appendix D.1. We only prove the first
estimate and defer the proof of the second and third to Appendix D.2.
Proof. Denote by M = ||ω(|x|−α + |x|β )||∞ . Clearly, we have
(6.22) |ω| ≤ min(|x|α , |x|−β )M.
Clearly, we have Z
zi
∂i φ = Ci Ki (x − y)W (y)dy, Ki (z) =,
|z|2
where W is the odd extension of ω from R+ 2 to R2 . Then W is odd in both x and y. Without
loss of generality, we consider i = 2. For a fixed x, we partition the integral into three regions:
Q1 = {y : |y| ≥ 2|x|}, Q2 = {y : |y − x| ≤ |x|/2}, Q3 = (Q1 ∪ Q2 )c .
In Q1 , symmetrizing the kernel, we need to estimate
Z Z
I1 , K2 (x − y)W (y)dy = (K2 (x1 − y1 , x2 − y2 ) − K2 (x1 + y1 , x2 − y2 ))W (y)dy.
Q1 Q1 ,y1 ≥0
72 JIAJIE CHEN AND THOMAS Y. HOU

Since K2 (z) is odd in z1 and even in z2 , for |y| ≥ 2|x|, we get

|x1 |
|K2 (x1 −y1 , x2 −y2 )−K2 (x1 +y1 , x2 −y2 )| = |K2 (y1 −x1 , x2 −y2 )−K2 (x1 +y1 , x2 −y2 )| ≤ 2 .
|y|2
Using (6.22), we get
Z
|I1 | . M |x1 | |y|−2 min(|y|α , |y|−β )dy . M |x1 | min(1, |x|−β ) . M min(|x|, |x|1−β ).
|y|≥2|x|

In Q2 , since |x − y| ≤ |x|/2, we have |y|  |x|, and min(|x|α , |x|−β )  min(|y|α , |y|−β ). It
follows
Z Z
|K2 (x − y)W (y) ≤ M |x − y|−1 min(|y|α , |y|−β )dy . M min(|x|α , |x|−β )|x|.
Q2 |x−y|≤|x|/2

In Q3 , we have |x|/2 ≤ |x − y| ≤ 3|x|, |y| ≤ 2|x|. Using this estimate and (6.22), we obtain
Z Z Z
−1 −β −1
| K(x − y)W (y)dy| . M α
|x − y| min(|y| , |y| )dy . M |x| min(|y|α , |y|−β )dy
Q3 Q3 |y|≤2|x|
−1 2+α 2−β 1+α
. M |x| min(|x| , |x| ) . M min(|x| , |x|1−β ).
Combining the above estimates, we prove the first estimate in Lemma 6.5. 
6.2.3. Estimate of ∇φ. We have the following estimate of ∇φ1 in |x| ≤ R2 .
Proposition 6.6. Let φ1 be the solution in (6.15). Let α, β ∈ (0, 1). There exists some absolute
1
constant ν1 (α, β) < 32 such that if Cl (τ )(1 + S(τ )) < ν1 , we have
max |∇φ1 |(|x|−1 + |x|−1+β ) . ||Ω1 (|x|−α + |x|β )||∞ .
|x|≤2R2

For 8S ≤ |x| ≤ R1 /2 away from the support of ω, we have an improved estimate

 S 3−β 
(6.23) |∇φ| . ||Ω1 (|x|−α + |x|β )||∞ + |x|1−β
(C l S)2−β
.
|x|2
In the following estimate, since r is sufficiently close to 1 within the support of ω, we can
simply treat Ω1 = ωr3/2 and ω as the same.
Proof. We choose i = 1 in (6.18). Denote
B1 , max |∇φ(|x|−1 + |x|−1+β )|, M1 , ||Ω1 (|x|−α + |x|β )||∞ .
|x|≤2R2

Inverting −∆ and then apply ∇, we obtain

∇(φ1 χ1 ) = ∇(−∆)−1 (Ω1 χ1 ) − ∇(−∆)−1 Z1,1 − ∇(−∆)−1 Z2,1 , I1 + I2 + I3 ,
aCl2
Z1,1 = 2∇φ1 · ∇χ1 + φ1 ∆χi , Z2,1 = φ1 χ i .
r2
Our goal is to prove the following estimate
(6.24) B1 ≤ Cα,β (M1 + (Cl S)2 · B1 ).
Then as long as Cl S is small, we can obtain the bound for B1 .
For I1 , I3 , applying Lemma 6.5, we get
aCl2
|I1 | ≤ min(|x|, |x|1−β )M1 , |I3 | ≤ min(|x|, |x|1−β )||Z2,1 (|x|−1 + |x|β )||∞ , Z2,1
φ1 χ 1 . =
r2
It sufficies to bound the norm of Z2,1 . For |x| ≤ 2S < R2 (6.21), using the definition of M1 ,
φ(0) = 0, and integration, we get
|φ1 | . M1 min(|x|2 , |x|2−β ),
which along with r−1 . 1 within the support of χ1 yields
|Z2,1 |(|x|−1 + |x|β ) . B1 Cl2 min(|x|2 , |x|2−β )(|x|−1 + |x|β ) . B1 Cl2 (|x|2 + |x|) . B1 Cl2 S 2 ,
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 73

For |x| ≥ 2S > 2, using Lemma 6.4, we yield

|Z2,1 (|x|−1 + |x|β )| . Cl2 (1 + | log(Cl |x|)|)S 2−β |x|β 1|x|≤2R1 ||ω(1 + |x|−β )||∞ .
By definition, we have Cl R1 ∈ [0, 1/2] (6.17). Within the support of ω, |ωr3/2 | = |Ω1 | and ω
are equivalent. Hence, we obtain (1 + | log Cl x|)|Cl x|β . 1 and
(6.25) |Z2,1 (|x|−1 + |x|β )| . (Cl S)2−β ||Ω1 (1 + |x|β )||∞ .
Next, we estimate I2 . Since R2 ≤ R41 , for |x| ≤ 2R2 and y ∈ supp(Z1,1 ), we have 2|x| ≤ |y|.
We estimate a typical term in Z1,1 . To use the formula (6.19), (6.20), we extend φ1 , χ1 naturally
from R+ 2
2 to R as an odd, and even function, respectively. For i, j ∈ {1, 2}, using integration by
parts, we get
xi − yi xi − yi
Z Z
−1
J , ∂i (−∆) (∂j φ1 ∂j χ1 ) = C 2
(∂j φ1 ∂j χ1 ) = −C ∂j ( ∂j χ1 )φ1 , J1 + J2 ,
R2 |x − y| R2 |x − y|2
xi − yi xi − yi 2
Z Z
J1 , −C ∂j ∂ χ φ , J2 , −C
2 j 1 1
∂ χ 1 φ1 .
R2 |x − y| R2 |x − y|2 j
Since the singularity x is away from the support of the integrand, the singular integral kernel
is smooth. We estimate the first term with i = 2, j = 1. Estimates of other cases and the
second term are similar. Denote K(z) = z|z| 1 z2
4 . Using the fact that ∂1 χ1 ψ1 is even in y1 and

symmetrizing the kernel in y1 , we get

Z Z
J1 = C K(x−y)∂1 χ1 φ1 (y)dy = C (K(x1 −y1 , x2 −y2 )+K(x1 +y1 , x2 −y2 ))∂1 χ1 φ1 (y)dy,
y1 ≥0

where we have used that ∂1 χ1 φ1 are even in y1 . Since K(z) is odd in z1 and |y| ≥ 2|x| for y in
the support of the integrand, we get
2x1
|K(x1 − y1 , x2 − y2 ) + K(x1 + y1 , x2 − y2 )| = |K(y1 + x1 , x2 − y2 ) − K(y1 − x1 , x2 − y2 )| . .
|y|3
Using Lemma 6.4, (6.16), and (6.17), we get
1 2−β
|∂1 χ1 φ| . S 1R1 ≤|y|≤2R1 M1 .
R1
It follows
|x|S 2−β |x|S 2−β
Z
|J1 | . M1 1R1 ≤|y|≤2R1 |y|−3 dy . M1 .
R1 R12
Using a similar symmetrization argument and the fact that x is away from the singularity of
the kernel when |x| ≤ 2R2 , we can obtain that
|x|S 2−β
(6.26) |∇(−∆)−1 Z1,1 | . M1 . |x|Cl2 S 2−β M1 . min(|x|, |x|1−β )(Cl S)2−β M1 .
R12
Combining the above estimate and using Cl S ≤ 1, we obtain
|∇φ| .α,β min(|x|, |x|1−β )(M1 + (Cl S)2−β B1 ).
Taking the maximum of x over |x| ≤ 2R2 , we prove (6.24), which further implies the desired
result.
Improved estimate. For 8S ≤ |x| ≤ R1 /2, we refine the estimate of I1 and I3 . For I1 , for
z2
y in the support of Ω1 , we have |x| ≥ 2|y|. For K(z) = |z| 2 , using the same symmetrization

argument, we get
Z Z
| K(x − y)Ω1 dy| = (K(x1 − y1 , x2 − y2 ) − K(x1 + y1 , x2 − y2 ))Ω1 dy
R2 y ≥0
Z 1 Z
2y1 −2
. 2
|Ω1 (y)|dy . |x| M1 |y| min(|y|α , |y|−β )dy . M1 |x|−2 S 3−β .
y1 ≥0 |x| |y|≤S
74 JIAJIE CHEN AND THOMAS Y. HOU

The term ∇(−∆)−1 Z1,1 is already estimated in (6.26). For Z2,1 , we estimate the norm
||Z2,1 (|x|−α + |x|β ) again. For |x| ≥ 2S, we have the estimate (6.25). For |x| ≤ 2S, using the
first estimate in Proposition 6.6 we just proved, we obtain
|Z2,1 |(|x|−α + |x|β ) . Cl2 min(|x|2 , |x|2−β )(|x|−α + |x|β )M1 . Cl2 |x|2 . Cl2 S 2 M1 .
It follows
||Z2,1 |(|x|−α + |x|β )||∞ . (Cl S)2−β M1 .
Applying Lemma 6.5 again, we obtain
|∇⊥ (−∆)−1 Z2,1 | . |x|1−β (Cl S)2−β M1 .
Note that for |x| ≤ R1 . Cl−1 , we have
|x|R1−2 . |x|1−β R1−2+β . |x|1−β Cl2−β .
Combining the above estimates and (6.26), we prove (6.23). 
6.2.4. Estimate of ∇2 φ. Based on the estimate in Proposition 6.6 for ∇φ, we further estimate
∇2 φ for |x| ≤ R3 .
Proposition 6.7. Let φ1 be the solution in (6.15). Let β ∈ (0, 1) and α ∈ (0, 1]. There exists
1
some absolute constant ν2 (α, β) < 32 such that if Cl (τ )(1 + S(τ )) < ν2 , we have
|∇2 (φ1 χ3 ) − ∇2 (−∆)−1 Ω1 | .α,β Clβ ||Ω1 (|x|−α + |x|β )||∞ .
In particular, we have
(6.27) ∂xy (φ1 ) = ∂xy φ(0), |∂xy φ(0) − ∂xy (−∆)−1 Ω1 (0)| .α,β Clβ ||Ω1 (|x|−α + |x|β )||∞ .
For Ω1 being the perturbation, we will further bound ∇2 (−∆)−1 Ω1 using the energy defined
in the Boussinesq equation (4.59).
Proof. We consider (6.18) with i = 2. Denote
M1 = ||Ω1 (|x|−α + |x|β )||∞ .
Using (6.18), we have
∇2 (φ1 χ2 ) = ∇2 (−∆)−1 Ω1 − ∇2 (−∆)−1 Z1,2 − ∇2 (−∆)−1 Z2,2 , I1 + I2 + I3 ,
where we have used Ω1 χi = Ω1 by requiring Cl S small. We only need to estimate I2 , I3 . The
estimate of I2 is similar to that in the proof of Proposition 6.6. We consider the typical term
J = ∂12 (−∆)−1 (∂1 φ1 ∂1 χ2 ).
z1 z2
For |x| ≤ R3 , it is away from the support of ∂1 φ1 ∂1 χ2 . Denote K(z) = |z|4 . We have
Z
J = C K(x − y)(∂1 φ1 ∂1 χ2 )(y)dy.

Using Proposition 6.6 and (6.16), we get

(6.28)
|∂1 φ1 ∂1 χ2 | . R2−1 1R2 ≤|y|≤2R2 |∂1 φ1 | . M1 R2−1 R21−β 1R2 ≤|y|≤2R2 . M1 R2−β 1R2 ≤|y|≤2R2 .
Since |x| ≤ R3 ≤ R2 /4 ≤ |y|/4 and |K(x − y)| . |y − x|−2 . |y|−2 , we get
Z Z
−β −β
|J| . M1 R2 |K(x − y)|dy . M1 R2 |y|−2 dy . M1 R2−β .
R2 ≤|y|≤2R2 R2 ≤|y|≤2R2

Similarly, we can obtain

|I2 | . M1 R2−β . M1 Clβ ,
where we have used (6.17) to obtain the last inequality.
aC 2
For I3 , we estimate ∂12 (−∆)−1 Z2,2 . Recall Z2,2 = r l φ1 χ2 from (6.18). Other derivatives
are similar. By definition, we have
Z
∂12 (−∆)−1 Z2,2 = C K(x − y)Z2,2 (y)dy.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 75

For a fixed x, we partition the region of the integral into three parts
Q1 = {y : |y| ≥ 2|x|}, Q2 = {y : |y − x| ≤ |x|/2}, Q3 = (Q1 ∪ Q2 )c .
Applying Proposition 6.6, using (6.16) and |∂i r−1 | . Cl when |r| ≥ 1
2, for |x| ≤ 2R2 , we
obtain
|∂i Z2,2 | . Cl2 (|∂i r−1 φ1 χ2 | + |∂i φ1 χ2 | + |φ1 ∂i χ2 |)
. M1 Cl2 1|x|≤2R2 (Cl |x|2 ∧ |x|2−β + |x| ∧ |x|1−β + (|x|2 ∧ |x|2−β )R2−1 ) . M1 Cl2 R21−β . Cl1+β M1 .
We also have the pointwise estimate
(6.29) |Z2,2 | . Cl2 |φ1 χ2 | . Cl2 min(|x|2 , |x|2−β )1|x|.R2 M1 . M1 Clβ .
Using the above pointwise estimate, for |x| ≤ R3 , we can obtain
Z Z
| K(x − y)Z2,2 (y)dy| . M1 |y|−2 Cl2 min(|y|2 , |y|2−β )dy . Cl2 R22−β M1 . M1 Clβ ,
Q1 2|x|.|y|.R2
Z Z
| K(x − y)Z2,2 (y)dy| . max |∇Z2,2 (y)| |x − y|−1 dy . Cl1+β M1 |x| . M1 Clβ ,
|x−y|≤|x|/2 |x−y|≤|x|/2 |x−y|≤|x|/2

Since Q3 ⊂ {y : |x|/2 ≤ |x − y| ≤ 3|x|}, we get

Z Z Z
| K(x − y)Z2,2 (y)dy| . |x|−2 | Z2,2 (y)dy| . |x|−2 M1 Clβ dy . M1 Clβ .
Q3 Q3 |y|≤4|x|

Combining the above estimates, we prove the desired result.

To obtain (6.27), we simply use r = 1 when x = y = 0 and
∂xy (φ1 ) = ∂xy (φr1/2 )(0) = ∂xy φ(0).


6.2.5. Weighted L∞ and Hölder estimate. Based on Propositions 6.6 and Proposition 6.7, we
show that Z1,i , Z2,i in (6.18) are small in the energy norm. Recall the weights ϕ1 , ϕg,1 , ψ1 and
g1 from (C.1), (C.2), (C.3). Denote
(6.30) ||Ω||X , max(||Ωϕ1 ||L∞ , ||Ωϕg,1 ||L∞ , ||Ωψ1 ||C 1/2 ).
g1

The energy (4.59), (4.44) also includes the norm ||Ω1 ψ1 |x1 |−1/2 ||L∞ , which can be bounded by
||Ωψ1 ||C 1/2 up to some absolute constant. Thus, we do not include it in the above norm.
g1

Recall that ϕ1,g ∼ c|x|1/16 for large |x|. We will fix

1
(6.31) β=
16
in the following estimate.
We want to show that the Z term in (6.18) is small in X. However, Z2,4 only vanishes to
order O(|x|2 ) near x = 0 and is not in the space X since space X involves singular weights
of order |x|−γ with γ ∈ (2, 3]. We need to subtract a rank one correction near x = 0. In the
following estimates, the sizes of Z1,i , Z2,i are very small. The reader can mainly pay attention
to the vanishing order of these terms near x = 0.
Proposition 6.8. Let φ1 be the solution to (6.18). Suppose that Ω1 ∈ X and Cl S < ν2 , where
ν2 is the constant in Proposition 6.7. For |x| ≤ 2, α, β > 0, α < 2, we have
|∇(φ1 − x1 x2 ∂xy φ1 (0))| .α,β |x|1+α ||Ω1 |x|−α + |x|β ||L∞ ,
|φ1 − x1 x2 ∂xy φ1 (0)| .α,β |x|2+α ||Ω1 |x|−α + |x|β ||L∞ .
If Ω1 ∈ X, the vanishing order can be further improved. The weight ϕ1 (C.2) is singular of
order |x|−2.9 near x = 0,
76 JIAJIE CHEN AND THOMAS Y. HOU

Proof. Using (6.18) with i = 4, we get

φ1 χ4 = (−∆)−1 (Ω1 − Z1,4 − Z2,4 ).
Using Lemma 6.5, we only need to prove that
||(Ω1 − Z1,4 − Z2,4 )(|x|−α + |x|β )||∞ . ||Ω1 (|x|−α + |x|β )||∞ .
Since χ1 = 1 in the support of χ4 , the estimate of Z1,4 , Z2,4 follows directly from Proposition
6.6 and its proof. We only consider a typical term. For ∂1 φ1 ∂1 χ4 in Z1,4 , using Proposition 6.6,
we get
|∂1 φ1 ∂1 χ4 |(|x|−α + |x|β ) . ||Ω1 (|x|−α + |x|β )||∞ min(|x|, |x|1−β )R4−1 1|x|R4 (|x|−α + |x|β )
. ||Ω1 (|x|−α + |x|β )||∞ .
We need to require α < 2 since Z2,4 in (6.15) only vanishes to order |x|2 near x = 0. 
Now, we are in a position to show that the Z term in (6.18) with a correction is small in
space X.
Proposition 6.9. Suppose that Ω1 ∈ X and Cl S < ν2 , where ν2 is the constant in Proposition
6.7. We have
||Z1,4 + Z2,4 − ∂xy (0)Z2,4 (−∆κ)||X . Cl S(||Ω1 (|x|−1 + |x|β )||∞ + ||χ3 ∇2 (−∆)−1 Ω1 ||∞ ),
3
where κ = − xy2 χ and χ is some cutoff function supported near x = 0, e.g. (6.16).
We will apply Proposition 6.9 to Ω1 = ωr3/2 with ω ∈ X or ω = ω̄χ(x/ν) with 1 < ν < R4 ,
where ω̄ is the approximate steady state for the 2D Boussinesq equation. In both cases, we can
further bound the right hand side as follows
||Ω1 (|x|−1 + |x|β )||∞ + ||χ3 ∇2 (−∆)−1 Ω1 ||∞ . ||Ω1 ||X ,
(6.32)
||ω̄ν r3/2 (|x|−1 + |x|β )||∞ + ||χ3 ∇2 (−∆)−1 (ω̄χν )||∞ . 1.
The estimates of ∇2 (−∆)−1 Ω1 in both inequalities follow from standard interpolation in-
equalities. The first inequality also follows from the sharp estimate in Section 8.
Proof. Let φ1 be the solution of (6.18). Denote
M = ||Ω1 (|x|−1 + |x|β )||∞ + ||χ3 ∇2 (−∆)−1 Ω1 ||∞ .
Now, using Proposition 6.7, we yield
|∇2 φ1 (x)| . M,
1
for |x| ≤ R3 . Note that the norm of Ω1 in Propositions 6.6 and 6.7 with β = 16 can be bounded
by M .
Estimate of Z1,4 . Firstly, we estimate Z1,4 (6.18). For |x|  R4 , from (6.17), we have Cl R4  1.
Using Lemma 6.4 and (6.23) in Propposition 6.6, we obtain
|φ(x)| . S 2−β M, |∇φ(x)| . |x|1−β M (Cl S)2−β ,
where we have used S . |x|, |x|Cl . 1 to simplify the upper bound in (6.23). Using the above
estimate and R4−1  Cl , we obtain the pointwise estimate
|Z1,4 | . 1|x|R4 (R4−1 |∇φ1 | + R4−2 |φ1 |) . M 1|x|R4 (R4−1 |x|1−β (Cl S)2−β + S 2−β R4−2 )
. Clβ (Cl S)2−β M 1|x|R4 .
Using Lemma 6.4 and Proposition 6.6, we obtain
|∇Z1,4 | . 1|x|R4 (|∇2 φ1 ||∇χ4 | + |∇φ1 ||∇2 χ4 | + |φ1 ||∇3 χ4 |)
. 1|x|R4 M (R1−3 |x|2−β + R1−2 |x|1−β + R4−1 ) . 1|x|R4 M R4−1 . 1|x|R4 M Cl .
Note that the weight ϕ1 , ϕg,1 involves the weight |x1 |−1/2 which is singular along x1 = 0. For
any power γ ∈ [−3, β], we have
|Z1,4 |x|γ | . Clβ |x|γ (Cl S)2−β 1|x|R4 M . (Cl S)2−β 1|x|R4 M,
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 77

where we have used ϕ1 . ϕg,1 and ϕg,1  |x|β for large x.

For any power γ ∈ [−5/2, 0) and x1 ≤ 1, we have
|Z1,4 |x|γ |x1 |−1/2 | . max |∇Z1,4 (z, x2 )| · |x|γ . R4−1+γ M . Cl M.
|z|≤1

To estimate the Hölder norm of Z1,4 ψ1 , following similar estimates, we obtain

|Z1,4 ψ1 | . (Cl S)2−β M, |∇(Z1,4 ψ1 )| . (Cl + (Cl S)2−β )M.
We can obtain better estimates due to the decay of ψ1 (x) for large x, see (C.1). But we do
not need this extra smallness. Using the above estimates and embedding inequalities, we obtain
|Z1,4 ϕ1 | . |Z1,4 ϕg,1 | . (Cl + (Cl S)2−β )M, ||Z1,4 ψ||C 1/2 . (Cl + (Cl S)2−β )M.
gi

Estimate of Z2,4 . Recall

aCl2
Z2,4 = φ1 χ4 .
r
Clearly, we have ∂xy Z1,4 (0) = aCl2 ∂xy φ1 (0). We perform the following decomposition
aCl2 aC 2
Z2,4 − ∂xy Z2,4 (0)(−∆κ) = (φ1 − ∂xy φ1 (0)(−∆κ)) + ∂xy φ1 (0) l (1 − r)(−∆κ) , I + II.
r r
From the definition of κ in Proposition 6.9, for |x| ≤ 1, we have
−∆κ = x1 x2 , (1 − r) = Cl x2 .
Thus II = O(x1 x22 ) near x = 0. Using Proposition 6.7 for ∂xy ψ(0) and the fact that II is
suported near x = 0, we get
||II||X . Cl3 M.
Applying Propositions 6.6 and 6.8 with α = 1 and we obtain
(6.33) |∇I| . Cl2 min(|x|2 , |x|1−β )M, |I| . Cl2 x1 min(|x|2 , |x|1−β )M,
where to obtain the second bound Cl2 x1 |x|1−β , we have used Proposition 6.6 and integrated the
estimate for ∂1 φ in x1 to estimate φ1 . Note that if the derivative acts on r−1 , we get |∇r−1 | . Cl
and then use Cl |x| . 1 to remove a growing power |x|.
For large |x| ≥ 8S, the correction vanishes (−∆κ) = 0. Using Lemma 6.4 and the improved
estimate (6.23), for 8S ≤ |x| ≤ R1 /2 . Cl−1 , we have |x|1−β (Cl S)2−β . Cl S 2−β . S 1−β and
|φ1 | . (1 + | log(Cl S)|)S 2−β M, |∇(φ1 /r)| . |∇φ1 | + Cl |φ1 | . S 1−β M,
where we have used Cl S(1 + | log |Cl S||) . 1 to absorb the logarithm factor. It follows
|I| . Cl2 (1 + | log(Cl S)|)S 2−β M . (Cl S)1−β (1 + | log(Cl S)|)Cl1+β SM . Cl1+β SM,
(6.34)
|∇I| . Cl2 S 1−β M,
for |x| ∈ [8S, R1 /2]. Therefore, for any γ1 ∈ [−3, β], combining the above estimates and (6.33)
and using |x| ≤ R1 . Cl−1 within the support of I, we have
(6.35) |x|γ |I| . Cl SM.
Next, we bound |I||x|γ2 |x1 |−1/2 for γ2 ∈ [−5/2, 0]. If |x1 | ≥ 1, it follows from the above
bound. If |x1 | ≤ 1, using (6.33), we obtain
|x|γ2 |x1 |−1/2 |I| . |x|γ2 Cl2 |x1 |1/2 min(|x|2 , |x|1−β )M . Cl2 min(1, |x|1−β )M . Cl1+β M.
The above estimates imply
||Iϕ1 ||∞ . ||Iϕg,1 ||∞ . (Cl + Cl S)M . Cl S||Ω1 ||X , ||Iψ1 ||∞ . Cl SM.
Note that ψ1 . |x|−2 + 1. For the Hölder estimate of I, using (6.33)-(6.35), we get
|∇(Iψ1 )| . |∇I|ψ1 + |I∇ψ1 | . Cl2 min(|x|2 , |x|1−β )(|x|−2 + 1)||Ω1 ||X + Cl SM
. (Cl2 (1 + |x|1−β ) + Cl S)M . Cl SM.
78 JIAJIE CHEN AND THOMAS Y. HOU

In the last inequality, we have used Cl1−β |x|1−β . 1 within the support of I and Cl  1 < S.
Using embedding inequalities, we obtain the Hölder estimate of Iψ1 . We conclude
||Z2,4 − ∂xy Z1,4 (0)(−∆)κ||X . ||I||X + ||II||X . Cl SM.
Since Cl S . 1, Cl . 1, combining the estimate of Z1,4 , Z2,4 , we complete the proof. 

6.3. Main terms for the stream function and velocity. Based on Proposition 6.9, we
rewrite (6.18) with i = 4 as follows
−∆(φ1 χ4 + aCl2 φxy (0)κ) = Ω1 χ4 − Z1,4 − (Z2,4 − ∂xy Z2,4 (0)(−∆κ)),
where we have used (6.27) and
∂xy Z2,4 (0) = aCl2 ∂xy φ1 (0) = aCl2 ∂xy φ(0).
Recall the definitions of φ1 , Ω1 from (6.18). We introduce
Ψ = φr1/2 χ4 + aCl2 φxy (0)κ, Ω = ωr3/2 χ4 − Z1,4 − (Z2,4 − ∂xy Z2,4 (0)(−∆κ)),
(6.36)
Ψ2 = −aCl2 φxy (0)κ.
Then we obtain
(6.37) − ∆Ψ = Ω, φr1/2 χ4 = Ψ + Ψ2 .
Since within the support of Ω, we have r−1 . 1 and |r − 1| . Cl S. Clearly, we have
(6.38) ||Ωρ||∞ ≤ (1 + CCl S)||ωρ||∞ , ||Ωψ1 ||C 1/2 ≤ ||ωψ1 ||C 1/2 + CCl S||ω||X ,
g1 g1

for any weight provided that ωρ ∈ L∞ . Thus, Ω and ω enjoy almost the same estimate. Later,
we will generalize all the estimates for u, v in the 2D Boussinesq equations to Ψ.
From Propositions 6.7, 6.9, the term φxy (0) satisfies
(6.39) |φxy (0)| . ||Ω1 ||X . ||Ω||X , Ψxy (0) = φxy (0).
Therefore, the term aCl2 φxy (0)κ is very small and vanishes to the order |x|4 near x = 0.

6.3.1. Main terms for the velocity. Recall u, v from (6.13). Since we will only need the estimate
of the velocity within the support of the solution, where χi = 1, in the following derivation, we
drop the cutoff functions χi . Firstly, from (6.37), we yield
φr1/2 = Ψ + Ψ2 .
The term Ψ2 is smooth with vanishing order |x|4 , compactly supported, and small. We treat
it as a lower order term and do not expand its derivation below. The velocity depends on the
derivatives of φ. Using ∂x r = 0, ∂y r = −Cl defined in (6.9) (please do not confuse r with
p
x2 + y 2 here), we rewrite ∇φ as follows
Cl
φy = (r1/2 φr−1/2 )y = (Ψr−1/2 )y + (Ψ2 r−1/2 )y = Ψy r−1/2 + Ψ + (Ψ2 r−1/2 )y ,
(6.40) 2r3/2
φx = r−1/2 Ψx + r−1/2 Ψ2,x .
Then using (6.13), we can rewrite u, v as follows
(6.41)
1 Cl 1 1
u = −φy + Cl φ = −Ψy r−1/2 − 3/2 Ψ + 3/2 Cl Ψ − (Ψ2 r−1/2 )y − 3/2 Cl Ψ2 , uM + uR
r 2r r r
where the main term and the remainder are given by
Cl 1
(6.42) uM = −Ψy , uR = −Ψy (r−1/2 − 1) +3/2
Ψ − (Ψ2 r−1/2 )y − 3/2 Cl Ψ2 .
2r r
An important observation is that the second and the third terms cancel each other near the
origin. To see this, we have
1 − r1/2 1−r Cl y
(6.43) r−1/2 − 1 = 1/2
= 1/2 1/2
= 1/2 .
r r (1 + r ) r (1 + r1/2 )
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 79

It follows
(6.44)
Cl
− Ψy (r−1/2 − 1) + Ψ
2r3/2
Cl Cl xy
= − (Ψy − Ψxy (0)x)(r−1/2 − 1) + (Ψ − Ψxy (0)xy) + Ψxy (0)(−x(r−1/2 − 1) + 3/2 ),
2r3/2 2r
Cl Cl xy(1 + r1/2 − 2r)
= − (Ψy − Ψxy (0)x)(r−1/2 − 1) + 3/2 (Ψ − Ψxy (0)xy) + Ψxy (0) .
2r 2r3/2 (1 + r1/2 )
The last term vanishes to the order O(|x|3 ) near x = 0. We treat uR as the remainder since it
vanishes to the order O(|x|3 ) near x = 0 and contain the small factor Cl . Within the support
of the solution, we get Cl |x| ≤ Cl S, which is small.
For v in (6.13), using (6.40) we have
v = φx = Ψx + (r−1/2 − 1)Ψx + r−1/2 Ψ2,x , vM + vR ,
(6.45)
vM = Ψx , vR = (r−1/2 − 1)Ψx + r−1/2 Ψ2,x .
We treat vR as the remainder since it contains the small factor Cl and vanishes to the order |x|2
near x = 0. Within the support, vR has size of order Cl S.

6.3.2. Main terms for the velocity of the approximate steady state. Following [12], we want to
construct the approximate steady state for the 3D Euler (6.13) by trunctating the approximate
steady state (ω̄, θ̄) for the 2D Boussinesq. We need to show that the associated velocity (6.13)
is close to that in the 2D Boussinesq equation ∇⊥ (−∆)−1 ω̄. For ν sufficiently large and to be
chosen, we define
ω̄ν = χ(x/ν)ω̄,
where χ is the cutoff function chosen above (6.16).
To avoid confusion ,we denote φ̄2D = (−∆)−1 ω̄. Using (6.36), (6.37) with ω = ω̄ν constructed
above, and then subtracting (6.36) by −∆φ̄2D = ω̄, we yield
(−∆)(φr1/2 χ4 + aCl2 φxy (0)κ − φ̄2D ) = ω̄χν r3/2 χ4 − ω̄ − Z1,4 − (Z2,4 − ∂xy Z2,4 (0)(−∆)κ).
Applying Proposition 6.9 and (6.32) with Ω1 = ω̄ν , we have the following estimates for the Z
terms
||Z1,4 + (Z2,4 − ∂xy Z2,4 (0)(−∆)κ)||X . Cl S.
Note that the above right hand side only vanishes to the order |x|2 near x = 0. Near x = 0,
we have
3
ω̄χν r3/2 χ4 − ω̄ = ω̄(r3/2 − 1) = − Cl ω̄x (0)xy + O(|x|3 ).
2
We do a correction of the above elliptic equation
 3 
(−∆) φr1/2 χ4 + (aCl2 φxy (0) + ω̄x (0) Cl )κ − φ̄2D
2
3/2 3
=ω̄χν χ4 (r − 1) + Cl ω̄x (0)(−∆κ) + ω̄(χ4 χν − 1) − Z1,4 − (Z2,4 − ∂xy Z2,4 (0)(−∆)κ).
2
The right hand side vanishes near x = 0 to the order |x|3 . Denote
(6.46)
3
Ων,R , ω̄χν χ4 (r3/2 − 1) + Cl ω̄x (0)(−∆κ) + ω̄(χ4 χν − 1) − Z1,4 − (Z2,4 − ∂xy Z2,4 (0)(−∆)κ),
2
3 3
Ψ̄ = φr1/2 χ4 + (aCl2 φxy (0) + ω̄x (0) Cl )κ, Ψ̄2 = −(aCl2 φxy (0) + ω̄x (0) Cl )κ.
2 2
We yield
(6.47) − ∆(Ψ̄ − φ̄2D ) = Ων,R , −∆Ψ̄ = Ωµ,R + ω̄.
1 −1/6
Since ω̄ ∈ C and |ω̄| . min(|x1 |, |x| ), It is easy to obtain
(6.48) ||Ων,R ||X . ||ω̄(χ4 χν − 1)||X + Cl S.
80 JIAJIE CHEN AND THOMAS Y. HOU

By choosing ν sufficiently large and Cl S to be small, we can obtain that Ψ̄ − φ̄2D is very
small.
Similar to (6.42) and (6.45), based on Ψ̄, Ψ̄2 in (6.46) and
φr1/2 = Ψ̄ + Ψ̄2 ,
we can decompose the velocity in (6.13) associated with ω̄ν as follows
(6.49) ū = −Ψ̄y + ūR , v̄ = Ψ̄y + v̄R ,
for |x| ≤ R4 . The formulas of ūR , v̄R are similar to those in (6.42), (6.45) with Ψ, Ψ2 replaced
by Ψ̄, Ψ̄2 . The remaining term vanishes near the origin with order |x|3 and is small.

6.3.3. Estimate of the velocity. Since the main terms in the velocity are given by (−∆)−1 Ω, we
need several weighted estimate of the velocities.
Lemma 6.10. Suppose that Ω ∈ X (6.30) is odd and −∆Ψ = Ω. We have
 1 
(6.50) ∇2 Ψ − Ψxy (0)xy − ∂1112 Ψ(0)(x31 x2 − x1 x32 ) . |x|2.5 ||Ω||X , |∂1112 Ψ(0)| . ||Ω||X ,
6
for |x| ≤ 1 and
∇2 (Ψ − Ψxy (0)xy) . min(|x|2 , 1)||Ω||X .

The formula of ∂1112 Ψ(0) can be written as an integral of Ω and is given in (2.79). Note
that in Section 8, we develop the sharp version of the above estimates with better constants. In
Appendix D.2.3, we present the proof, which also helps to illustrate the ideas for Section 8.
In Sections 3 and 8, for u = ∇⊥ (−∆)−1 ω with ω ∈ X, we develop weighted estimate for
ux − ûx , uy − ûy , vx − v̂x , where the approximation terms are constructed in Section 2.11. In
particular, we obtain
|x − z|−1/2 |(f − fˆ)ψ1 (x) − (f − fˆ)ψ1 (z)| . ||ωψ1 ||C 1/2 + ||ωϕ1 ||∞ . ||ω||X , f = ux , uy , vx .
g1

for x, z with x1 = z1 or x2 = z2 and |x| ≤ |z| ≤ (1 + µ)|x| for some µ ∈ (0, 1). The estimate
up to some absolute constant can be established following the decomposition and argument in
Section 8 and using the asymptotics of the weights ϕ1 , ψ1 (C.1), (C.2). When |z| > (1 + µ)|x|,
the estimate follows from the above Lemma, the triangle inequality, and |ψ1 | . |x|−2 + |x|−1/6 .
Note that near x = 0, f − fˆ agree with the left hand side of (6.50). See (2.80) and (2.82). In
particular, we have
||(f − fˆ)ψ1 ||C 1/2 . ||ω||X .
Note that the approximation terms in (2.82), (2.88) except Cf 0 (x)(−∂12 (−∆)−1 ω)(0), Cf (x)K00 χ0
are supported away from 0 with smooth coefficient. Moreover, the functionals in (2.82), (2.88),
e.g. fˆ(xi , 0), can be bounded by ||ω||X . Using triangle inequality, we yield
 
|| f − Cf 0 (x)(−∂12 (−∆)−1 ω)(0) − Cf (x)K00 ψ1 ||C 1/2 . ||ω||X ,

where χ0 is defined in Section 2.11.2 and supported near x = 0. In summary, we have

Lemma 6.11. Suppose that Ω ∈ X (6.30) is odd and −∆Ψ = Ω. We have
  1
||ψ1 ∂ij (Ψ − Ψxy (0)x1 x2 ) − χ0 ∂1112 Ψ(0)∂ij G(x) ||C 1/2 . ||Ω||X , G(x) , (x31 x2 − x1 x32 ).
6
Note that using the above estimates for ∇2 (−∆)−1 Ω, we can obtain the estimate for ∇(−∆)−1 Ω
by integration, which is more regular.

6.4. Nonlinear stability. We apply the nonlinear stability analysis of the 2D Boussinesq equa-
tions to prove Theorem 4. In Section 6.4.1, we impose the bootstrap assumption on the support
size. In Section 6.4.2, we construct the approximate steady state and impose the normalization
conditions, which are small perturbations to those in the 2D Boussinesq.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 81

6.4.1. Bootstrap assumption on the support size. We fix the exponents α = 1 or α = 2.9, β =
1
16 in Propositions 6.6, 6.7, 6.9. These exponents are related to the singular weights we use.
Then the constants ν1 , ν2 in these propositions are determined. We impose the first bootstrap
assumption: for t ≥ 0, we have
(6.51) Cl (t)(1 + max(S(t), S(0))) < min(ν2 , 4−6 ).
Under the above Bootstrap assumption, the support of ω, θ in D1 does not touch the symmetry
axis and z = ±1, the cutoff functions (6.17) satisfy χi = 1, i ≤ 5 for x in the support, and
the assumptions in Propositions 6.6, 6.7, 6.9. We will choose Cl (0) at the final step, which
guarantees the smallness in (6.51).
6.4.2. Approximate steady state and the normalization condition. Since the rescaled domain D̃1
(6.12) is bounded, we construct the approximate steady state with bounded support. We localize
the approximate steady state ω̄, θ̄ for the 2D Boussinesq constructed in Section 5 to construct
the approximate steady state for (6.13)
(6.52) ω̄0 , χν ω̄, θ̄0 , χν (1 + θ̄),
where χν (x) = χ1 (|x|/ν) is some cutoff function, and χ1 (y) : R → [0, 1] is even in y, χ1 = 1 for
|y| ≤ 1, and χ1 (y) = 0 for |y| ≥ 2. We can choose χ1 = χ̃21 for another smooth cutoff function
1/2
χ̃1 such that χ1 = χ̃1 is smooth. Clearly, from Definition 6.2, the support size of ω̄0 , θ¯0 is 2ν.
We truncate 1 + θ̄ rather than θ̄ so that 1 + θ̄ & 1 and (1 + θ̄)1/2 has the same regularity as θ̄.
This idea follows [12].
Denote β = arctan xy , ρ = x2 + y 2 . Recall the formula (D.13) in Polar coordinate. We have
p

1 1
(6.53) ∂x θ̄0 = χν θ̄x +cos βχ01 (|x|/ν)(1 + θ̄), ∂y θ̄0 = χν θ̄y + sin βχ01 (|x|/ν)(1 + θ̄).
ν ν
To distinguish the notations between the 3D Euler and the 2D Boussinesq equations, we write
φ̄2D = (−∆)−1 ω̄, ū2D = ∇⊥ φ̄2D
for the 2D Boussinesq. Let φ̄ and ū be the stream function and velocity in (6.13) associated
with ω̄0 . We have the leading order terms for u (6.49). See more discussions in Section 6.3.2.
We need to adjust the time-dependent normalization condition for cω (t), cl (t). We impose
the following conditions
θ̄xx (0) 1
(6.54) c̄l = 2 , c̄ω (t) = c̄l + ūx (0), c̄θ (t) = c̄l + 2c̄ω (t)
ω̄x (0) 2
for the approximate steady state ω̄0 , θ̄0 , and
(6.55) cl (t) = 0, cω (t) = ux (t, 0)
for the perturbations, where u(t, 0) is the velocity in (6.13) and is different from −∂y (−∆)−1 ω.
The above conditions are the same as (2.11) and (2.19), and play the same role of enforcing
(2.12). As a result, the perturbation ω, ∇θ satisfies the vanishing condition (2.22)
ω = O(|x|2 ), ∇θ = O(|x|3 )
near x = 0. Since ∇θ̄0 = ∇θ̄, ω̄0 = ω̄ near x = 0, the factor c̄l is the same as that for the 2D
Boussinesq.
We remark that c̄ω (t) is time-dependent since it depends on ūx (0) and the elliptic equation in
(6.13) depends on the rescaling factor Cl . From the estimate in Proposition (6.7), ūx (0) is very
close to −∂xy (−∆)−1 ω̄0 . For ν sufficiently large, comparing the above conditions and (2.11), c̄ω
is very close to cω,2D (2.15) used for the 2D Boussinesq equations in Section 2. From (6.54) and
(2.11), we yield
(6.56) c̄ω − c̄ω,2D = ūx (0) − ūx,2D (0).
Remark 6.12. We will choose ν to be very large relatively to 1. Therefore, we treat ω̄0 ≈ ω̄, θ̄0 ≈
θ̄. Due to these samll factors and using (6.46) and (6.49), we can treat ū ≈ ū2D . From Remark
6.3 and the bootstrap assumption (6.51), we also have Cl ≈ 0, Cl S ≈ 0, r ≈ 1. We treat the
error terms in these approximations as perturbation.
82 JIAJIE CHEN AND THOMAS Y. HOU

6.4.3. Linearized equations. The equations (6.13) are slightly different from (2.10) for the Boussi-
nesq systems. Denote
η = θx , ξ = θy .
Linearizing (6.13) around the approximate steady state, we obtain the equations for the
perturbation (ω, η, ξ), which are similar to (2.17), (2.21)
(6.57)
1
∂t ω = −(c̄l x + ū) · ∇ω + 4 η + c̄ω ω − u · ∇ω̄0 + cω ω̄0 + F̄1 + N1 , L1 (ω, η, ξ) + F̄1 + N1 ,
r
∂t η = −(c̄l x + ū) · ∇η + (2c̄ω − ūx )η − v̄x ξ − ∂x (u · ∇θ̄0 ) + 2cω θ̄0,x + N2 + F 2
, L2 (ω, η, ξ) + F̄2 + N2 ,
∂t ξ = −(c̄l x + ū) · ∇ξ + (2c̄ω − v̄y )ξ − ūy η − ∂y (u · ∇θ̄0 ) + 2cω θ̄0,y + N3 + F 3 ,
, L3 (ω, η, ξ) + N3 + F 3 ,
where
1
F̄1 = −(c̄l x + ū) · ∇ω̄ +
θ̄x + c̄ω ω̄,
r4
and we adopt the notatons Ni , F̄i for other nonlinear terms and the error terms from (2.1),
(2.2). The ω equation is different from the corresponding equation in (2.17) since we have r14 θx
in (6.13). The ξ equation is also different from the corresponding equation in (2.21) since we
do not have the same incompressible conditions ūx + v̄y = 0, ux + vy = 0. We remark that the
velocity u, ∇u in the above system are determined by the elliptic equation in (6.13).
To generalize the analysis of the 2D Boussinesq equations to the 3D Euler equations, we
derive the different terms, which are all of lower orders. In the following derivations, we use
f2D to denote the quantity f used in the 2D Boussinesq. For example, ū2D denote the
approximate steady state for the velocity for 2D Boussinesq. It satisfies ū2D = ∇⊥ (−∆)−1 ω̄.
We introduce the norm Xi related to the energy (4.59)
(6.58) ||f ||Xi , ||f ϕi ||∞ + ||f ϕg,i ||∞ + ||f ψi ||C 1/2 .
gi

Lower order terms in the linearied and nonlinear operator. Using (6.42), we get
(6.59)
Li = LM,i + LR,i ,
1
LM,1 = −(c̄l x + ū) · ∇ω + η − ∇⊥ Ψ · ∇ω̄0 + Ψxy (0)ω̄0 + c̄ω ω, LR,1 = ( 4 − 1)η − uR · ∇ω̄0
r
LM,2 = −(c̄l x + ū) · ∇η + (2c̄ω − ūx )η − v̄x ξ − ∂x (∇⊥ Ψ · ∇θ̄0 ) + 2Ψxy (0)θ̄0,x ,
LR,2 = −∂x (uR · ∇θ̄0 ),
LM,3 = −(c̄l x + ū) · ∇ξ + (2c̄ω − v̄y )ξ − ūy η − ∂y (∇⊥ Ψ · ∇θ̄0 ) + 2Ψxy (0)θ̄0,y ,
LR,3 = −∂y (uR · ∇θ̄0 ).
Note that from (6.39) and (6.42), we have
(6.60) uR = O(|x|3 ), uR,x (0) = 0, Ψxy (0) = φxy (0) = ux (0) = cω .
We will estimate LR,i in Section 6.4.4 and show that it can be bounded by (Cl S + Clβ )E4 (t),
where E4 (t) is the energy norm (4.59) for the 2D Boussinesq.
For the nonlinear terms Ni , we decompose the velocity u into uM and uR similarly. We only
focus on N2 since other terms are decomposed similarly. Using (6.42), (6.45), we have
N2 = NM,2 + NR,2 , NM,2 = −u · ∇η − uM,x η − vM,x ξ + 2uM,x (0)η,
(6.61)
NR,2 = −uR,x η − vR,x ξ.
where we have used uR,x (0) = 0, uM,x (0) = ux,0 = cω (6.60). The estimate of the lower order
terms NR,2 follow the estimates of uR,x in Section 6.4.4 and the fact that η ∈ X2 , ξ ∈ X3 . The
estimate of the product in the space Xi is relatively standard since Xi is some weighted L∞ and
C 1/2 spaces. We do not decompose the transport term since we need to apply the weighted L∞
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 83

and C 1/2 estimate. In the weighted estimate, it leads to the nonlinear term di (ρ)W1,i ρ in (4.6).
The estimate of the lower order terms in di (ρ)W1,i ρ follows the estimate of NR,2 .
Lower order terms in the residual error. Denote
δf = f − f2D , δ ū = ū − ū2D , δc̄ω = c̄ω − c̄ω,2D .
For the residual error, using (6.49) and (2.2), we obtain
(6.62)
F̄i = F̄M,i + F̄R,i ,
1 − r4
F̄M,1 = −(c̄l x + ū2D ) · ∇ω̄0 + c̄ω,2D ω̄0 , FR,1 = −δ ū · ∇ω̄0 + δc̄ω · ω̄0 + θ̄x ,
   r4 
F̄M,i+1 = ∂i − (c̄l x + ū2D ) · ∇θ̄0 + (c̄l + c̄ω,2D )θ̄0 , FR,i = ∂i − δ ū · ∇θ̄0 + 2δc̄ω · θ̄0 .

Note that the profiles ω̄, ∇θ̄ decay and we choose the weights (C.3), (C.2), such that
|ω̄ϕg,1 | . |x|−γ1 , |∇θ̄ϕg,2 | . |x|−γ2
for some γ2 > 0, e.g. γ2 = 81 . Since F̄M,i agrees with the residual error F̄i (2.2) for the 2D
Boussinesq for |x| ≤ ν, where ν is the size of the cutoff function in (6.52), we have
(6.63) ||F̄M,i − F̄i,2D ||Xi . ν −γ
for some γ > 0, e.g. γ = 81 . The Hölder estimate of the tail is even smaller since ψ1 decays.
The error term FR,i does not vanish to the order |x|3 near 0. We use correction similar to
(2.66) and put the correction to the Wc2 equations.

6.4.4. Estimate the lower order terms in the linearized operator. In this section, we estimate
LR,i and show that
(6.64) ||LR,i ϕi ||∞ + ||LR,i ϕg,i ||∞ + ||LR,i ψg,i ||C 1/2 + ||LR,1 |x1 |−1/2 ψ1 ||∞ . (Cl S + Clβ )E4 (t),
gi

under the bootstrap assumption (6.51), for i = 1, 2, 3. Note that by definition of E4 , we have
||Ω||X . E4 (t).
We focus on the estimate of a typical term ∂x uR θ̄0,x .
Estimate of Ψ2 . Recall the formulas of uR from (6.42) and Ψ, Ψ2 from (6.36). The estimates
of the terms involving Ψ2 are simple since
Ψ2 = −αCl2 φxy (0)κ.
Recall the definition of κ from Proposition 6.9. Using (6.39), we yield
(6.65) |∇k Ψ2 | . Cl2 |x|4−k E4 .
For |x| ≤ 2, we have
|∇θ̄0 | . |x|, |∇2 θ̄0 | . 1.
We consider a typical term related to Ψ2 in ∂x uR θ̄0,x , e.g. ∂xy Ψ2 r−1/2 θ̄0,x (6.42). We can
bound it by
|∂xy Ψ2 r−1/2 θ̄0,x ϕ2 | . Cl2 |x|2 |x||x|−3 1|x|≤2 E4 . Cl2 E4 .
Note that for |x| ≤ 1, we have χ = 1,
∂xy Ψ2 r−1/2 θ̄0,x ψ2 = CCl2 r−1/2 y 2 θ̄0,x ψ2 ,
for some absolute constant C, ψ2 ∼ c|x|−5/2 near x = 0, and f = r−1/2 y 2 θ̄0,x ψ2 vanishes to the
order of |x|1/2 near x = 0. For |x| > 1/2, f is smooth and is supported near x = 0. Hence, we
obtain that f is in C 1/2 and
||∂xy Ψ2 r−1/2 θ̄0,x ψ2 ||C 1/2 . Cl2 E4 .
g2

The estimates of other terms related to Ψ2 or Ψ̄2 (6.49) in the residual error, nonlinear terms,
or linear parts related to Ψ2 follow similar estimates since Ψ2 , Ψ̄2 = O(|x|4 ) near x = 0 and
contain the small factor Cl2 . We treat them as lower order terms.
84 JIAJIE CHEN AND THOMAS Y. HOU

Estimate of Ψ − Ψxy (0)xy. Next, we estimate other terms in uR related to Ψ. Recall the de-
composition (6.44). The third term in (6.44) follows an estimate similar to that of Ψ2 performed
above. The first two terms vanish to a higher order near x = 0. We consider the estimate of a
typical term in ∂x uR θ̄0,x related to Ψ:

(6.66) I , ∂x (Ψy − Ψxy (0)x)(r−1/2 − 1)θ̄0,x = (Ψxy − Ψxy (0))(r−1/2 − 1)θ̄0,x .

Note that |r−1/2 − 1| . Cl |x2 | . Cl S within the support of the solution. The weighted L∞
estimate is simple and follows from Lemma 6.10. For example, using θ̄0,x . min(|x1 |, |x|−3/5 ),
we have
|Iϕ2 | . |Iϕg,2 | . 1|x|≤S Cl |x2 | min(|x|2 , 1) min(|x1 |, |x|−3/5 )ϕg,2 ||Ω||X . Cl SE4 ,

where the weights ϕ2 , ϕg,2 are defined in (C.3), (C.2) and satisfy ϕg,2 . |x|1/4 +|x1 |−1/2 (|x|−1/6 +
|x|−5/2 ).
To obtain the Hölder estimate, we use Lemma 6.11. Recall G(x) defined in Lemma 6.11.
Firstly, we rewrite I as follows
I = (Ψxy −Ψxy (0)−χ0 ∂xy G(x)Ψxxxy (0))(r−1/2 −1)θ̄x +χ0 Ψxxxy (0)∂xy G(x)(r−1/2 −1)θ̄x , I1 +I2 .
Denote
Ψxy,A , Ψxy − Ψxy (0) − χ0 ∂xy G(x)Ψxxxy (0).
The estimate of I2 is simple since the coefficient vanishes to O(|x|4 ) and we obtain a small
factor: |r−1/2 − 1| . Cl y . Cl S within the support of the solution. In particular, we have
||I2 ψ2 ||C 1/2 . |Ψxxxy (0)| . E4 .
For I1 , we first rewrite I1 ψ2 as follows
ψ2 −1/2
I1 ψ2 = (Ψxy,A ψ1 ) (r − 1)θ̄x .
ψ1
ψ2 −1/2
From Lemmas 6.10, 6.11, we have Ψxy,A ψ1 ∈ L∞ ∩C 1/2 . The coefficient ψ1 (r −1)θ̄x vanishes
to the order |x|3/2 near x = 0. Since
ψ2 −1/2 ψ2 −1/2
| (r − 1)θ̄x | . Cl S, |∇( (r − 1)θ̄x )| . Cl S,
ψ1 ψ1
we obtain the Hölder estimate for I1 ψ2 . The estimates of other terms in uR are similar. We
establish (6.64).

6.4.5. Estimates of the lower order terms in the residual error. In this section, we estimate the
lower order terms in the residual error (6.62) and show that

(6.67) ||F̄R,i − ci fχ,i ||Xi . Cl S + Clβ + ||Ων,R ||X ,

where the norm Xi is defined in (6.58), ci = ∂xy F̄R,i for i = 1, 2 and c3 = ∂xx F̄R,3 (0), and fχ,i
is defined in (2.66), and Ων,R is defined in (6.46). Using (6.48), we can further bound Ων,R as
follows
||Ων,R ||X . ||ω̄(χ4 χν − 1)||X + Cl S = ||ω̄(χν − 1)||X + Cl S,
where we have used νCl ≤ Cl S ≤ 4−5 in the last inequality to simplify χ4 χν = χν , which can
be done by choosing Cl sufficiently small later.
We focus on the case i = 2, i.e. the estimate of F̄R,2 − c2 fχ,2 . Firstly, from (6.56), we have
δc̄ω = c̄ω − c̄ω,2D = ūx (0) − ūx,2D (0) = δ ūx (0).
Note that ūx (0) + v̄y (0) = 0 for the velocity (6.49). A direct computation yields
c1 = ∂xy F̄R,1 (0) = δc̄ω ω̄xy (0), c2 = δc̄ω θ̄xxy (0), c3 = δc̄ω θ̄xxy (0).
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 85

We can rewrite FR,i (6.62) as follows

I = ∂x (−δ ū · ∇θ̄0 + 2δ ūx (0)θ̄0 ) − c2 fχ,2
= −(δ ū − δ ūx (0)x)θ̄0,xx − ∂x (δ ū − δ ūx (0)x)θ̄0,x − (δv̄ − δv̄y (0)y)θ̄0,xy
− ∂x (δv̄ − δv̄y (0)y)θ̄0,y + δ ūx (0)(θ̄0,x − xθ̄0,xx + y θ̄0,xy − θ̄xxy (0)fχ,2 )
, I1 + I2 + I3 + I4 + I5 .
For I5 , the coefficient is C 2 and has sufficiently fast decay. Moreover, using (6.46), (6.47),
(6.49), and Proposition 6.7, we have
|δc̄ω | = |δ ūx (0)| . ||Ων,R ||X .
Thus, we can obtain
||I3 ||X2 . ||Ων,R ||X .
The estimates of Ij , 1 ≤ j ≤ 4 are similar. We focus on the typical terms in I2
(6.68) I2 = −∂x (δ ū − δ ūx (0)x)θ̄0,x .
Recall Ψ̄ from (6.46) and φ̄2D = (−∆)−1 ω̄. Denote
Ψ̃ = Ψ̄ − φ̄2D .
Recall the formula of ū from (6.49). We have
 
δ ū − δ ūx (0)x = − ∂y (Ψ̄ − φ̄2D ) − ∂xy (Ψ̄ − φ̄2D )(0)x + ūR = −(∂y Ψ̃ − ∂xy Ψ̃(0)x) + ūR .

The formula of the remainder ūR is given by (6.42) with Ψ, Ψ2 replaced by Ψ̄, Ψ̄2 . From (6.46),
we have
−∆Ψ̃ = Ων,R , Ων,R ∈ X.
Then the estimate of
∂x (∂y Ψ̃ − ∂xy Ψ̃(0)x)θ̄0,x
in I2 follows from the estimate of Ψ − Ψxy (0)xy at the end of Section 6.4.4. In particular, we
can obtain
||∂x (∂y Ψ̃ − ∂xy Ψ̃(0)x)θ̄0,x ||X2 . ||Ων,R ||X .
Other terms in Ij , 1 ≤ j ≤ 4 related to Ψ̃ can be estimated similarly.
For the term in I2 (6.68) related to ūR , we have several terms due to the formula (6.42),
(6.44). The term involving Ψ̄2 is simple and its estimate follows from the estimate of Ψ2 in
Section 6.4.4. For other terms, we estimate a typical term
J = ∂x (Ψ̄y − Ψ̄xy (0)x) · (r−1/2 − 1)θ̄0,x .
Since Ψ̄ is close to φ̄2D , we use the decomposition Φ̄ = Φ̃ + φ̄2D and
J = ∂x (Ψ̃y − Ψ̃xy (0)x) · (r−1/2 − 1)θ̄0,x + ∂x (φ̄2D,y − φ̄2D,xy (0)x) · (r−1/2 − 1)θ̄0,x , J1 + J2 .
The term J1 follows from the above estimate. For J2 , we note that φ̄2D satisfies the elliptic
equation −∆φ̄2D = ω̄. From the construction of ω̄ in Section 5, we have ω̄ ∈ C 2 with decays
|ω̄| . min(|x1 |, |x|−1/4 ), |∇ω̄| . min(1, |x|−5/4 ).
To control φ̄, we use elementary embedding inequalities
||∇2 (−∆)ω||L∞ .α,p ||ω||C α + ||ω||Lp , α ∈ (0, 1), p < ∞,
which can be proved by decomposing the domain of the singular integral into the region near
the singularity and away from the singularity, and estimaing them by the C α norm of ω and the
Lp norm of ω separately. In particular, from −∆φ̄2D,x = ω̄x , −∆φ̄2D = ω̄, we obtain
|∇2 φ̄2D,x | . 1, |∇2 φ̄2D | . 1.
For the estimate of φ̄yyy , it involves the integral of ω̄ along the boundary y = 0, which can
be estimated similarly using the regularity and decay of ω̄. Hence, we obtain |∇3 φ̄2D,x | . 1.
86 JIAJIE CHEN AND THOMAS Y. HOU

Now, using the estimate of φ̄, |θ̄0,x | . min(|x1 |, |x|−3/5 ) , and the smallness of |r−1/2 − 1| (6.43)
within the support of the solution, we yield

|J2 | . Cl |x| min(|x|, 1) min(|x|−3/5 , |x1 |)1|x|≤S ,

which vanishes to the order |x|3 near x = 0. It follows the weighted L∞ estimate

|J2 ϕ2 | . Cl S, |J2 ϕg,2 | . Cl S, |J2 ψ2 | . Cl S min(|x|1/2 , 1).

Recall the weight ψ2 from (C.1). Note that ψ2 ∼ |x|−5/2 near x = 0 and ψ2 ∼ |x|1/6 for large
x, we get
|∇(J2 ψ2 )| . Cl S(|x|−1/2 + 1).
Combining the above two estimates, we obtain the C 1/2 estimate of J2 ψ2 . Other terms follow
similar estimates. We prove (6.67).

6.4.6. Comparison between the operators. In this section, we show that the difference between
the main parts of the operators in (6.59), (6.61) and the operators in (2.65) is small. We estimate
the lower order operators in Section 6.4.4, 6.4.5. Here, we only focus on the main terms.
There are three differences between LM,i in (6.59) and Li,2D in (2.65). Firstly, we use ū in
the transport term instead of ū2D = ∇⊥ φ2D . To quantify the difference, we use the elliptic
equations (6.47) and (6.49). We estimate the difference ū − ū2D in Section 6.4.5, and quantify
the smallness. It leads to a difference in the linear stability analysis bounded by

C(Cl S + Clβ + ||Ων,R ||X )E4 ,

where E4 is the energy (4.59) for the perturbation.

The second difference is that we use the truncated profile θ̄0 , ω̄0 in (6.59) rather than the
(θ̄, ω̄) in (2.65). Due to the decay of the profile and that we choose the weights (C.1), (C.2),
(C.3) such that |ω̄ϕg,1 | . |x|−γ , |∇θ̄ϕg,2 | . |x|−γ for large |x| and some γ > 0, e.g. γ = 81 , this
leads to a difference in the linear stability analysis is bounded by

Cν −γ E4 .

Thirdly, for the main term in the velocity uM = ∇⊥ Ψ, the modified stream function Ψ is
obtained from Ω (6.36), (6.37) rather than ω. Due to the equivalence (6.38), this leads to a
difference in the linear stability analysis bounded by

CCl SE4 .

We also refer to Section 6.4.4 for the estimate of the lower order part uR , which is small.
Note that the terms in the nonlinear operators (2.1), (6.61) all involve the nonlocal terms
determined by u. Recall that in the energy estimate of the 2D Boussinesq, we treat the nonlocal
term u2D as a bad term. Using the estimate of the lower order part uR in Section 6.4.4 and the
above reasoning in the estimate of uM , we have a difference in the nonlinear stability bounded
by
(CCl S + Clβ )E42 .

Decomposition of the solution. Note that due to the difference of the operators between
the 3D Euler (6.57) and the 2D Boussinesq (2.65), we need to replace the operators in the
decomposition (2.75) by the operators for the 3D Euler used above. For example, we need to
use the above linearized operator instead of Li,2D in (2.65). As a result, we also need to change
the linearized operator Li,2D in the residual operator (2.74). In Sections 6.4.4 and 6.4.6, we
show that the linearized operator Li and Li,2D are sufficiently close, with a difference bounded
by (Cl S + Clβ + ν −γ )E4 . The residual operator Ri , is also sufficiently close to that for the 2D
Boussinesq equations Ri,2D and their difference is bounded by (Cl S + Clβ + ν −γ )E4 .
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 87

6.5. Nonlinear stability and finite time blowup. For initial perturbation ω, η, ξ in the
energy class E4 (4.59) with E4 (ω, η, ξ) < E∗ , under the boostrap assumption (6.51), we can
perform nonlinear energy estimates similar to those for the 2D Boussinesq equations in Section
4. Combining the estimate (6.63), the estimates in Section 6.4.4, 6.4.5, and the discussion in
Section (6.4.6), we can bound the additonal terms due to the diference between two energy
estimates by
C(Cl S + ν −γ + Clβ + ||Ων,R ||X )(1 + E4 + E42 ) . C(Cl S + ν −γ + Clβ )(1 + E4 + E42 )
where β = 1
16 . Since ω̄ decays and |ω̄ϕg,i | . |x|−γ , using (6.48) and the decay of ω̄, we have
||Ων,R ||X . ||ω̄(χ4 χν − 1)||X + Cl S . (ν −γ + Cl S)E4 ,
for the some γ > 0, e.g. γ = 81 . The Hölder estimate of the tail is even smaller since ψ1 decays.
In summary, the difference can be bounded by
|J| ≤ C1,∗ (ν −γ + Cl S + Clβ )(1 + E4 + E42 ).
For the estimate of cω + c̄ω , from the energy estimate and the definition of E4 (4.59), we have

|cω + c̄ω − c̄ω,2D | ≤ 100E4 + C2,∗ (ν −γ + Cl S + Clβ ),

where cω is chosen in (6.54).
Note that the energy estimates for the 2D Boussinesq equations satisfy the nonlinear stability
conditions (A.14) with some ε0 > 0. Now, we choose ν > ν∗ with ν∗ large enough and a small
δ such that
(6.69) (C1,∗ + C2,∗ )(ν∗ + δ + δ β )(1 + E∗ + E∗2 ) < min(ε0 /4, 10−4 ).
We impose a stronger boostrap assumption than (6.51)
(6.70) Cl (t)(1 + S(t)) < min(δ, ν2 , 4−6 ).
Under the above boostrap assumption and the energy assumption for the W1 part of the
solution (see (2.75))
(6.71) E4 (t) < E∗ ,
using the nonlinear stability estimate for the 2D Boussiesq equations and (6.69), we can continue
the bootstrap assumption for the energy inequality. Moreover, using |ω| . |x|−β E4 from the
L∞ (ϕg,1 ) estimate, we have
|u + ū| . |x|1−γ/2 ,
which decays sublinearly, and the control for the rescaling parameter
1
c̄l > 2, c̄ω + cω < − .
2
Moreover, under the boostrap assumption, for small ν −γ + Cl S + Clβ , we have the control for
the rescaling parameter
1
c̄l > 2, c̄ω + cω < − .
2
Thus, following the argument in [12], under the boostrap assumption, we can control the support
Cl (t)(1 + S(t)) ≤ C(S(0))Cl (0),
for some absolute constant C depending on S(0). Therefore, for any S(0) < +∞, by choos-
ing Cl (0) sufficiently small, the assumption (6.70) is also satisfied. Therefore, the bootstrap
asumption can be continued.
Passing from nonlinear stability to finite time blowup with smooth data ω θ , uθ compactly
supported near (r, z) = (1, 0) follows the argument in [12]. We conclude the proof of Theorems
4, 2.
88 JIAJIE CHEN AND THOMAS Y. HOU

7. Constructing and estimaing the approximate solution to the linearized

equations
Recall the decomposition of the system (2.65) in Sections 2.10.3, 2.10.4. We need to construct
the approximate solutions to eLt F0 for several initial data F̄i , F̄χ,i . In this section, we discuss
how to construct these space-time solutions numerically with the vanishing properties (2.76).
The linearizeq equations associated with L (2.65) read
∂t ω = −(c̄l x + ū) · ∇ω + η + c̄ω ω − u · ∇ω̄ + cω ω̄ = L1 (ω, η, ξ),
(7.1) ∂t η = −(c̄l x + ū) · ∇η + (2c̄ω − ūx )η − v̄x ξ − ux · ∇θ̄ − u · ∇θ̄x + 2cω θ̄x = L2 (ω, η, ξ),
∂t ξ = −(c̄l x + ū) · ∇ξ + (2c̄ω + ūx )ξ − ūy η − uy · ∇θ̄ − u · ∇θ̄y + 2cω θ̄y = L3 (ω, η, ξ),
where cω = ux (0) (2.19). Although η, ξ represent θx , θy in the Boussinesq equations (2.20),
we will consider initial data (ω0 , η0 , ξ0 ) with ∂y η0 6= ∂x ξ0 . Thus, we do not have the relation
∂y η = ∂x ξ and will treat η, ξ as two independent variables.
The solutions ω, η are odd, ξ is even with ξ(0, y) = 0. We consider initial data (ω0 , η0 , ξ0 ) =
O(|x|2 ) near x = 0. Using an argument similar to (2.22) and a direct calculation, we obtain that
these vanishing conditions are preserved
(7.2) ω(t, x), η(t, x), ξ(t, x) = O(|x|2 ).

7.1. A posteriori error estimates: decomposition of errors. Since we cannot solve the
Poisson equation exactly, for the stream function φ̄, φ, we have the natural decomposition
φ̄ = φ̄N + φ̄e , φ = φN + φe ,
where the short hands N, e denote numeric, error, respectively. We use similar notations below
for other nonlocal terms since we cannot construct them exactly. We will construct φ̄N , φN
numerically and explicitly and treat φ̄e , φe as error. The reader should not confuse φN with the
N -th power of φ. We will never use power of φ throughout the paper. Similarly, we denote by
uN , ue the velocities corresponding to φN , φe . For example, we have
uN = −∂y (−∆)−1 φN , ue = −∂y (−∆)−1 φe , cN N
ω = ux (0), ceω = uex (0).
The above decomposition leads to the following decomposition of the operator L
L1 = LN e ē
1 + L1 + L1 , L2 = LN e ē
2 + L2 + L2 , L3 = LN e ē
3 + L3 + L3 ,

LN N N N N
1 = η + c̄ω ω − (c̄l x + ū ) · ∇ω + cω ω̄ − u · ∇ω̄,
Le1 = ceω ω̄ − ue · ∇ω̄, Lē1 = c̄eω ω − ūe · ∇ω,
(7.3) LN N N N N N N N
2 = −(c̄l x + ū ) · ∇η + (2c̄ω − ūx )η − v̄x ξ − ux · ∇θ̄ − u · ∇θ̄x + 2cω θ̄x ,
Le2 = −uex · ∇θ̄ − ue · ∇θ̄x + 2ceω θ̄x , Lē2 = −ūe · ∇η + (2c̄eω − ūex )η − v̄xe ξ,
LN N N N N N N N
3 = −(c̄l x + ū ) · ∇ξ + (2c̄ω − v̄y )ξ − ūy η − uy · ∇θ̄ − u · ∇θ̄y + 2cω θ̄y ,
Le3 = −uey · ∇θ̄ − ue · ∇θ̄y + 2ceω θ̄y , Lē3 = −ūe · ∇ξ + (2c̄eω − v̄ye )ξ − ūey η,
where Lei , Lēi denote the errors from ψ e , ψ̄ e , respectively. These operators depend on ω, η, ξ, and
we drop the dependence in (7.3) to simplify the presentation.

7.2. First correction and the construction of φN . Firstly, we solve (7.1) using the numerical
method outlined in Section 5 to obtain the solution (ωk , ηk , ξk ) at discrete time tk . Since ξ is
even with ξ(0, y) = 0, we write ξ = xζ for an odd function ζ. Then we represent ω, η, ζ using
the piecewise 6-th order B-spline (5.3). We remark that we do not add the semi-analytic part
in this construction for efficiency consideration and that the far-field behaviors of the solutions
are different for different initial data.
According to the normalization condition and (7.2), the solution to (7.1) satisfies ωx (0, t) =
ηx (0, t) = 0. To obtain an approximate solution with this condition, we make the first correction
ωk → ωk − ωk,x (0, 0)χ11 , ηk → ηk − ηk,x (0, 0)χ21 ,
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 89

where χij are cutoff functions defined in (7.6) with χij = x + O(|x|4 ) near 0. We do not modify
ξk since ξk already vanishes quadratically near (0, 0). We remark that the first correction does
not change the second order derivatives of the solution near 0 and cω since
∂xy χ11 (0) = ∂xy χ21 (0) = 0, cω (χ11 ) = −∂xy ψ1 (0) = 0.
Here ψ1 is defined below
xy 2
ψ1 = − κ(x)κ(y),
2
where κ(x) is the cutoff function chosen in (F.5) in Appendix F.2 satisfying κ(x) = 1 + O(|x|4 )
near x = 0, and ψ1 satisfies −∆ψ1 = x + O(|x|4 ). To construct the stream function φN , we first
solve −∆φk = ωk numerically to obtain φN k,1 . Then we correct it as follows

φN N N N
k,1 → φk,1 + ∂xyy φk,1 (0)ψ1 , φk .

This allows us to obtain

∂x (−∆)φN N N
k (0) = −∂x ∆φk,1 (0) + ∂xyy φk,1 (0) = 0,
(7.4)
∆φN 2
k = O(|x| ), ωk − (−∆)φN 2
k = O(|x| ).

We further extend it to Lipschitz continuous solutions W ˆ

c , (ω̂(t), η̂(t), ξ(t)) in time using a
cubic spline interpolation in t. See section 7.5 for more details.
7.3. The second correction. The error
ˆ
(∂t − Li )(ω̂(t), η̂(t), ξ(t))
may not vanish to the order O(|x|3 ), which is a property (2.76) that we require in the energy
estimate. Then we add the second correction
ˆ → ξ(t)
ω̂(t) → ω̂(t) + a1 (t)χ12 , η̂ → η̂ + a2 (t)χ22 , ξ(t) ˆ + a3 (t)χ32
so that the error satisfies
(2) ˆ + a3 (t)χ32 ) = O(|x|3 )
(7.5) εi , (∂t − Li )(ω̂(t) + a1 (t)χ12 , η̂(t) + a2 (t)χ22 , ξ(t)
near x = 0. We use the following functions for this correction
xy 2
χ11 = −∆ψ1 , ψ1 = − κ(x)κ(y), χ21 = xκ(x)κ(y),
(7.6) 2
xy 3 x2
χ12 = −∆ψ2 , ψ2 = − κ(x)κ(y), χ22 = xyκ(x)κ(y), χ32 = κ(x)κ(y),
6 2
where κ(x) is chosen in (F.5). Since κ(x) satisfies κ(x) = 1 + O(|x|4 ) near x = 0, the behaviors
of the above functions near x = 0 are given by
χ11 = y + l.o.t., χ21 = x + l.o.t., χ12 = xy + l.o.t., χ22 = xy + l.o.t., χ32 = x2 /2 + l.o.t.
We choose χ1j = −∆ψj so that its associated velocity ∇⊥ (−∆)−1 χ1j can be obtained explicitly.
For cutoff functions χ1 , χ2 , χ3 with cω (χ1 ) = −∂xy (−∆)−1 χ1 = 0, e.g. χi = χi2 chosen above,
we have the following general formula of Li (a1 (t)χ1 , a2 (t)χ2 , a3 (t)χ3 )
 
L1 (a1 χ1 , a2 χ2 , a3 χ3 ) = a1 (t) − (c̄l x + ū) · ∇χ1 + c̄ω χ1 − u(χ1 ) · ∇ω̄ + a2 (t)χ2 ,
   
L2 (a1 χ1 , a2 χ2 , a3 χ3 ) = a2 (t) − (c̄l x + ū) · ∇χ2 + (2c̄ω − ūx )χ2 − a3 (t)v̄x χ3 − a1 (t) u(χ1 ) · ∇θ̄ ,
   x
L3 (a1 χ1 , a2 χ2 , a3 χ3 ) = a3 (t) − (c̄l x + ū) · ∇χ3 + (2c̄ω + ūx )χ3 − a2 (t)ūy χ2 − a1 (t) u(χ1 ) · ∇θ̄ ,
y

where u(χ1 ) is the velocity associated with χ1 . We want to apply the above formulas to the
second corrections χi2 , i = 1, 2, 3. We use the Hadamard product
(7.7) (A ◦ B)i = Ai Bi ,
to simplify the notation as follows
Li (a ◦ χ) = Corij (x; χ)aj (t).
90 JIAJIE CHEN AND THOMAS Y. HOU

In the second correction, we use χi2 , i = 1, 2, 3 in (7.6). Note that Corij involves the nonlocal
term ū = ūN + ūe . According to this decomposition, we decompose Corij as follows
N ē
(7.8) Corij = Corij + Corij .
Note that we do not decompose u(χ1 ) since we can obtain it explicitly.
Next, we derive the equations for a(b), b(t), c(t). Using the condition
(2) (2) (2)
∂xy ε1 (0) = ∂xy ε2 (0) = ∂xx ε3 (0) = 0,
from (7.5), we obtain the following ODEs for a(t), b(t), c(t)
ȧ1 (t) = (−2c̄l + c̄ω )a1 (t) + a2 (t) − F1 (t),
(7.9) ȧ2 (t) = (−2c̄l + 2c̄ω − ūx (0))a2 (t) − F2 (t),
ȧ3 (t) = (−2c̄l + 2c̄ω − ūx (0))a3 (t) − F3 (t),
where F (t) = (F1 (t), F2 (t), F3 (t))T is the error associated to the second order derivatives of
(∂t − L)Ŵ near 0. More precisely, we have
c = d ω̂xy (t, 0) − (−2c̄l + c̄ω )ω̂xy (t, 0) − η̂xy (t, 0) − cω (t)ω̄xy (0),
F1 (t) = ∂xy (∂t − L1 )W
dt
c = d η̂xy (t, 0) − (−2c̄l + 2c̄ω − ūx (0))η̂xy (t, 0) − cω (t)θ̄xxy (0),
(7.10) F2 (t) = ∂xy (∂t − L2 )W
dt
2 d
c = ξˆxx (t, 0) − (−2c̄l + 2c̄ω − ūx (0))ξˆxx (t, 0) − cω (t)θ̄xxy (0).
F3 (t) = ∂x (∂t − L3 )W
dt
2 2 T
Denote D = (∂xy , ∂xy , ∂x ) . Then we can simplify (7.10) as
Fi = Di2 (∂t − Li )Ŵ (0).
Denote by M the coefficents in (7.9)
 
−2c̄l + c̄ω 1 0
(7.11) M = 0 −2c̄l + 2c̄ω − ūx (0) 0  , M N + M ē ,
0 0 −2c̄l + 2c̄ω − ūx (0).
where the last identity is based on the decomposition c̄ω = c̄N e N e
ω + c̄ω , ūx (0) = ūx (0) + ūx (0), and
ē e e
M only contains the contribution from c̄ω , ūx (0) . According to the normalization condition
(2.19), we have ūx (0)e = c̄eω . It follows
M ē = c̄eω I3 .
We simplify the ODE for a = (a1 , a2 , a3 )T as
(7.12) ȧi (t) = Mij aj (t) − Fi (t), ȧ(t) = M a − F.
Recall χ·2 = (χ12 , χ22 , χ32 ) from (7.6). The overall error for the approximate solution W
c+
a(t) ◦ χ·2 is
(∂t − L)(W
c + a(t) ◦ χ·2 ) = (∂t − L)W
c + ∂t a(t) ◦ χ·2 − L(a(t) ◦ χ·2 )
   
= (∂t − L)W
c − F (t) ◦ χ·2 + M a(t) ◦ χ·2 − Cor(x; χ·2 )a , I + II.

To simplify the notation, we drop the dependence of χ·2 in Cor. Denote by e1 , e2 , e3 the
canonical basis in R3 , e.g., e1 = (1, 0, 0). For II, we have
(7.13) II = ei χi2 Mij aj − ei Corij (x)aj = −ei (Corij (x) − Mij χi2 )aj .
Using (7.8) and (7.11), we yield
e ij (x, χ·2 ) , Corij (x) − Mij χi2 = CorN (x) − M N χi2 + Corē (x) − M ē χi2 .
Cor ij ij ij ij

e ij (x) vanishes to the order O(|x|3 ) near x = 0. For example,

It is not difficult to show that Cor
for i = j = 1, we have
ē ē
J , Cor11 (x) − M11 χ12 = −ūe · ∇χ12 + c̄eω χ12 − c̄eω χ12 = −ūe · ∇χ12 .
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 91

Since χ12 = xy + O(|x|4 ) (7.6), ūe = ūex (0)x + O(|x|2 ), v̄ e = −ūex (0)y near 0, we have J = O(|x|3 )
near 0. For II, we estimate the weighted norm for Cor e ij and then apply the triangle inequality
to further bound II.
For the I term, using (7.3) and (7.10), for each equation i, we get

(∂t − Li )W c − Di2 (∂t − Li )W

c − Fi χi2 = (∂t − Li )W c (0)χi2

(7.14) =(∂t − LN c 2 N c ec 2 ec
i )W − Di (∂t − Li )W (0)χi2 − (Li W − Di Li W (0)χi2 )

− (Lēi W
c − Di2 Lēi W
c (0)χi2 ) , Ii,N + Ii,e + Ii,ē .

Since Di2 (∂t −LN N c

i )W (0)χi2 is the leading order term of (∂t −Li )W near 0, we yield Ii,s = O(|x| )
c 3

near 0. We estimate each term and then apply the triangle inequality to bound Ii .
In summary, to estimate the error (∂t − L)(W c + A ◦ χ·2 ), we will use the triangle inequality
and estimate Ii,N , Ii,e , Ii,ē and II separately. The error II is due to the second correction,
I1,N is from the error of solving (7.1) numerically, Ii,e , Ii,ē are due to the error of solving the
Poisson equations for ω and ω b . Since we use a cubic spline interpolation to obtain the continuous
function W̄ (t), each error is a piecewise cubic polynomial in time, and we track the coefficients
of these polynomials to verify that they are small.

7.4. Estimate of the solutions to the ODE. Since we need to solve the ODE (7.9) exactly
to achieve the vanishing order (7.5) for all time, we cannot use a numerical method, e.g. the
Runge-Kutta method, to obtain the approximate solution. Since (7.9) is a linear ODE with
constant coefficients and some forcing terms, we can solve it exactly by diagonalizing the system.
Introduce
λ1 = −2c̄l + c̄ω , λ2 = λ3 = −2c̄l + 2c̄ω − ūx (0),
a2 F2
ã1 = a1 + , F̃1 = F1 + , ãi = ai , F̃i = Fi , i = 2, 3.
λ1 − λ2 λ1 − λ2
The approximate steady state satisfies λ1 ≈ −7, λ2 = λ3 ≈ −5.5. We diagonalize (7.9) as follows
d
ãi = λi ãi − F̃i .
dt
Let a0 be the initial data, which will be chosen to correct the vanishing order of the error
F̂i (0) − F̄i in (7.2). Then the solution ãi is given by
Z t
ãi (t) = eλi t ãi (0) − eλi (t−s) F̃i (s)ds.
0

It follows
1 − eλ i t
|ãi (t)| ≤ ||F̃i (s)||L∞ (0,t) .
−λi
Using the relation between ai and ãi , we further obtain the estimate for ai . With the estimate
on ai , we can estimate the error term II (7.13). For example, we have a L∞ (ρ) estimate

|IIρ(x)| ≤ ei ||Cor
e ij (·, χ)ρ||L∞ |aj |.

The norm ||Core ij (·, χ)ρ||L∞ is time-independent and we only need to compute it once.
We remark that we do not need to track the values of a(t) during the computation. Since
Fi is a piecewise cubic polynomial in t, to bound Fi (s) over time, we only need to track the
coefficients of the cubic polynomial.

7.5. Cubic interpolation in time. Given the numerical solution with correction Wcn = (ω̂n , η̂n , ξˆn ),
c (t, x) over (t, x) ∈ [0, T ] × R+ . For
we use a piecewise cubic interpolation to construct W 2
92 JIAJIE CHEN AND THOMAS Y. HOU

s ∈ [−3k/2, 3k/2], we construct

3k 1 1 s
W (s + tn + )= (−W0 + 9W1 + 9W2 − W3 ) + (W0 − 27W1 + 27W2 − W3 )
2 16 24 k
1 s 2 1 s
+ (W0 − W1 − W2 + W3 )( ) + (−W0 + 3W1 − 3W2 + W3 )( )3
4 k 6 k
X 1 s i
, Ci · V ( ) , V = (W0 , W1 , W2 , W3 ),
i! k
i≤3

where k is the time step, Wi = Ŵn+i for tn = nk, and Ci ∈ R4 is the coefficient determined by
the interpolation formula. A direct calculation yields
X Ci · V 1 s X 1 s
c − LW
∂t W c= ( )i−1 − L(Ci · V ) ( )i
k (i − 1)! k i! k
1≤i≤3 i≤3
X  Ci+1 · V 1 s s3
= − L(Ci · V ) ( )i − L(C4 · V ) 3 .
k i! k 6k
i≤2

To estimate ∂t Wc −LWc , we will use the triangle inequality and estimate Ci+1 ·V −L(Ci ·V ), L(C4 ·
k
W ) rigorously using the method described in the previous sections.
Applying the triangle inequality and integrating the error over s ∈ [− 3k 3k
2 , 2 ] yield
Z X Ci+1 · V Z
1 s i
Z
1 s 3
|∂t W − LW |ds ≤ − L(Ci · V ) | | + |L(C4 · V )| | |
|s|≤3k/2 k |s|≤3k/2 i! k |s|≤3k/2 6 k
i≤2
X Ci+1 · V
=k − L(Ci · V ) CI (i) + |L(C4 · V )|CI (3),
k
i≤2

where
9 9 27
CI = [3, , , ].
4 8 64
7.5.1. Decomposing the time interval for parallel computing. To verify that the posteriori error
is small, we need to estimate the error rigorously at each time step, which takes a significant
amount of time. Consider a partition of the time interval 0 = T0 < T1 < .. < Tn = T , where
T is the final time of the computation. To reduce the computational time, we first solve the
equations on [0, T ] without any rigorously verification and save the solution at Ti . Since we do
not need to perform verification at this step, the running time for each time step is short. Then
we solve the equations on a smaller time interval [Ti , Ti+1 ], i = 0, 1, 2..., n − 1 from the initial
data W (Ti ) and then perform the verification in each time interval in parallel.

7.6. Stopping criterion. To construct an approximate solution, we do not need to solve the
linearized equations (7.1) for all time. In fact, since the solution decays in certain norm as t
increases, we stop the computation at time T if Ŵ − D2 Ŵ ◦ χ is small in the energy norm. Then
we extend Ŵ (t) trivially for t > T
W
c (t) = 0, t > T.
As a result, the error satisfies
Ri = (∂t − Li )W
c = (∂t − Li )W
c 1t≤T − δT W
ci (T ).

Let F = (F1 , F2 , F3 ), Fi = Di2 (∂t − Li )W

c , where D2 = (Dxy , Dxy , Dx2 ). Then similarly, we
x=0
get
F = D2 (∂t − L)W
c 1t≤T − D2 W
c (T )δT .
Recall that the coefficients of the second correction a satisfies (7.12)
ȧ = M a − F.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 93

Using the Duhamel formula, we yield

Z t
Mt
a(t) = e a0 − exp(M (t − s))F (s)ds
0
for t < T . For t ≥ T , we derive
Z min(t,T )
Mt 2 W )(T )eM (t−T ) 1
a(t) = e a0 − \
exp(M (t − s))F (s)ds + (D t≥T .
0
Although the forcing terms only have finite support in time, a(t) has global support. More-
over, it is in L∞ and piecewise smooth. Now, with the second correction, for t > T , the residual
error for rank-one perturbation is given by
Z t
R= c(t − s)(∂t − L)(W
c + a ◦ χ)ds
0
Z T  
= c(t − s) (∂t − L)W c (T ) − D2 W
c − F (t) ◦ χ ds − (W c (t) ◦ χ)c(t − T )
0
Z t
+ c(t − s)(M a ◦ χ − Cor(x; χ)a)(s)ds.
0
Note that the last part is a linear combination of (a1 (t), a2 (t), a3 (t)). We will apply the
previous method to estimate it. Since ai (t) decays exponentially fast for t > T , the last integral
is uniformly bounded for all t > 0. The above formula can be generalized straightforward to the
finite rank perturbation.
In practice, we construct the numeric solution up to time T = 12. At that time, the solution
is very small.

7.7. Ideas of estimating the norm of the error. In this section, we discuss how to estimate
the error derived in the previous section, e.g. (∂t − Li )W c − Fi χi2 (7.14), a-posteriori. The
general idea is to first evaluate f on some grid points and estimate the higher order derivatives
of f in a domain D. Then we can construct an approximation fˆ of f by interpolating the values
of f at different points. The approximation error f − fˆ can be bounded by Ck ||f ||C k hk , where h
measures the size of the domain. For example, we have the simple second order estimate (E.5)
in 1D. If the mesh h is sufficiently small and we partition [a, b] into many small sub-intervals
with size h, applying (E.5), we obtain sharp estimate of ||f ||L∞ [a,b] .
To estimate the derivatives of f in the error term Ck ||f ||C k hk , we can use a rough estimate.
Since we have the small factor hk , the overall estimate will be small. The estimate of ||f ||C k for f
being the numeric solution or residual error can be established using the estimates in Appendix
E for piecewise polynomials or semi-analytic solutions.
In order to develop efficient method for rigorous estimates, we have the following considera-
tions. Firstly, we should evaluate as a small number of points as possible so that the method
is efficient. Secondly, most functions f in the verification are complicated and it is difficult to
obtain the sharp bound of the higher derivatives. We observe that, for a product f = pq, e.g.
f = ux ω̄x , the estimates of its higher order derivatives can be established by the triangle in-
equality, the Leibniz rule, and the estimates of the derivatives of p and q. This approach allows
us to reduce estimating some complicated functions to estimating several simpler functions. Yet,
in general, this approach overestimates the derivatives significantly. To compensate the over-
estimates of the derivatives, we use higher order interpolations and estimates, which provide
the small factor hk . We develop three estimates based on different interpolations: the Newton
interpolation, the Lagrangian interpolation, and the Hermite interpolation. We generalize these
interpolations to 2D and develop the estimates in Appendix G.
Each estimate has its own advantages and we will apply them in different situations. We note
that the approximation errors in these estimates are bounded by Chk for k = 4 or 5. We want
to estimate the constant C as sharp as possible to reduce the computational cost and improve
the efficiency. In fact, when k = 4, if we can obtain an interpolation method and reduce the
C
constant C to 16 , to achieve the same level of error, we can increase h to 2h. In this verification
94 JIAJIE CHEN AND THOMAS Y. HOU

step, since the domain is 2D, it means that we can evaluate only 14 of the grid point values of
f , which is much faster.
Based on these L∞ estimates of f , we can obtain sharp estimate of the derivatives of f . Using
the method in Section G.5, we can further estimate the weighted norm of f with singular weight
near 0. This allows us to verify the smallness of the residual error in some weighted L∞ norm.
Using these L∞ estimates of f and its derivatives, we can further develop Hölder estimate for
f . See Section G.6. We remark that the numeric solutions are regular, e.g. the approximate
steady state and the solutions to the linearized equations are C 3 ,

7.8. Posteriori estimates of the velocity. In our estimate of the error for the linearized
equations, we need to verify the error between the computed stream function φN , its related
velocity uN = ∇⊥ φN and the ground true stream function (−∆)−1 ω and velocity ∇⊥ (−∆)−1 ω
for the given vorticity. We observe that
∇⊥ (−∆)−1 ω − ∇⊥ φN = ∇⊥ (−∆)−1 (ω − (−∆)φN ),
and the error ω−∆φN depends on the numerical solution ω, φN locally. Since φN is the numerical
solution to the Poisson equations −∆φ = ω, we expect that the error ω −∆φN is very small. The
norm of ω − (−∆)φN can be estimated using the method in Appendix G. For this reason, we use
functional inequalities to estimate the operator u(F ) = ∇⊥ (−∆)−1 F, ∇u(F ) = ∇∇⊥ (−∆)−1 F ,
where F = ω − (−∆)φN is the error. If the error F is small in the energy norm. Then we can
show that u(F ) and ∇u(F ) are small.
Due to the round off error in solving the elliptic equation numerically, the error ∂x F (x, y) is
large for small x and large y, e.g. x ≤ 1, y ≥ 107 , and ∂y F (x, y) is large for small y and large x,
e.g. y ≤ 1, x ≥ 107 . To obtain a sharp estimate of u(F ) and its derivatives, we avoid using the
pointwise value of these quantities.

7.8.1. Estimates of u and ∇u. The simplest way to estimate the nonlocal term u(f ) is to use
some embedding inequality that bounds the weighted L∞ norm of u(f ) by the weighted L∞
norm of f . To estimate ∇u(ω), firstly, we rewrite them as the velocity of some function f .
Let χ : R+ → [0, 1] be some cutoff function such that χ(y) = 1 for small y and χ = 0 for large
y. Denote χ1 (y) = χ(y), χ2 (y) = 1 − χ(y). We decompose F as F χ1 + F χ2 . The first part is
supported in the region with small y. Since ∂x commutes with (−∆)−1 , we have
∂x u(F χ1 ) = u(Fx χ1 ), uy (F χ1 ) = F χ1 + v(Fx χ1 ), vy (F χ1 ) = −ux (F χ1 ) = −u(Fx χ1 ),
where we have used uy (g) − vx (g) = g.
Denote ȳ = (y1 , −y2 ). For the second part, using the definition of ux and integration by parts
in y direction, we yield
−2(x1 − y1 )(x2 − y2 ) 2(x1 − y1 )(x2 + y2 ) 
Z 
1
ux (F χ2 ) = + F (y)χ2 (y)dy
2π R2+ |x − y|4 |x − ȳ|4
x1 − y1 x1 − y1
Z
1
= (−∂y2 − ∂y2 )F (y)χ2 (y)dy
2π R2+ |x − y|4 |x − ȳ|4
x1 − y1 x1 − y1
Z
1
= ( + )∂y (F (y)χ2 (y))dy , −v e ((F χ2 )y ),
2π |x − y|2 |x − ȳ|2 2

where the boundary term vanishes since χ2 (y2 ) = 0 and e in the superscript means even
y2 =0
function. Similarly, for uy and x2 ≥ 0, we have
x2 − y2
Z
1 x2 + y2
uy (F χ2 ) = (−∂y2 2
− ∂y2 )F χ2 dy,
2π R2+ |x − y| |x − ȳ|2
where ∂y2 is the distributional derivatives. Using integration by parts, we derive
x2 − y2
Z
1 x2 + y2
uy (F χ2 ) = ( + )∂y (F χ2 )dy , ue (∂y (F χ2 )).
2π R2+ |x − y|2 |x − ȳ|2 2
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 95

For vx , vy , we get
vx (F χ2 ) = uy (F χ2 ) − F χ2 = ue ((F χ2 )y ) − F χ2 , vy (F χ2 ) = −ux (F χ2 ) = v e (F χ2 ).
Therefore, we obtain
ux (F ) = u(Fx χ1 ) − v e ((F χ2 )y ), uy (F ) = F χ1 + v(Fx χ1 ) + ue ((F χ2 )y )
e
vx (F ) = v(Fx χ1 ) + u ((F χ2 )y ) − F χ2 , vy (F ) = −ux (F ).
Thus, to estimate the error ∇u(F ) and u(F ), we can use the above decomposition and
estimate the weighted norm of F , which depends on two local terms ω, ψ N . Then we develop
functional inequalities for the nonlocal operator using the methods in Section 8.

7.8.2. Estimate near 0 and estimates of ∇2 u. Denote by F the error in the elliptic equation
F = ω − (−∆)ψ N after correction on ω, ψ N . We have F = O(|x|2 ) near x = 0 (7.4). Denote
ũ = u − ux (0)x, ṽ = v − vy (0)y. We need to estimate
1 1
ũ(F ), ṽ(F ).
|x|3 |x|3
See Section 7.8.3.
Note that in general, the operator ũ does not map L∞ (|x|−2 ) to L∞ (|x|−3 ). We use the
regularity of u(F ) and estimate ∇2 u(F ) near 0. Firstly, we solve another Poisson equation to
obtain ψ2
(7.15) (−∆)ψ2 = ωxx .
Denote ψ1 = ψ N . Then we perform the decomposition
∂xx ũ(F ) = ∂xx u(F ) = u(Fxx ) = u(ωxx − (−∆)ψ1,xx )
= u(ωxx − (−∆)ψ2 + (−∆)ψ2 − (−∆)ψ1,xx ) = u(ωxx − (−∆)ψ2 ) + ψ1,xxy − ψ2,y .
Similarly, we obtain
∂xx ṽ(F ) = v(ωxx − (−∆)ψ2 ) − ψ1,xxx + ψ2,x ,
∂xy ũ(F ) = ∂xy u(F ) = ∂x (F + vx (F )) = Fx + vxx (F )
= ∂x (ω − (−∆)ψ1 ) + v(ωxx − (−∆)ψ2 ) − ψ1,xxx + ψ2,x ,
= ∂x (ω + ψ1,yyy + ψ2 ) + v(ωxx − (−∆)ψ2 ),
∂yy ũ(F ) = ∂yy u(F ) = ∂y (F + vx (F )) = Fy + vxy (F ) = Fy − uxx (F ) = Fy − u(Fxx )
= ∂y (ω − (−∆)ψ1 ) − u(ωxx − (−∆)ψ2 ) − ψ1,xxy + ψ2,y
= ∂y (ω + ψ1,yy + ψ2 ) − u(ωxx − (−∆)ψ2 ).
The important reason we introduce ψ2 is that the error ωxx − (−∆)ψ1,xx is not small. In fact,
suppose that the numerical scheme of solving the Poisson equation has order k. Then, formally,
the error ωxx − (−∆)ψ1,xx is of order hk−4 , which is not small. Moreover, due to the slow
decay of ω and large round off error, ψ1 is not accurate for large |x|. Thus, if we use functional
inequality and some weighted norm of ωxx − (−∆)ψ1,xx to estimate u(ωxx − (−∆)ψ1,xx ), the
resulting estimate is not small. The error ωxx −(−∆)ψ2 related to solving (7.15) is much smaller
since it only involves second order derivative of the numerical solution and the source term ωxx
decays much faster. To estimate the nonlocal terms in the above decomposition, we estimate the
weighted L∞ norm of ωxx − (∆)ψ2 and use functional inequalities for u, v developed in Section
8.

7.8.3. Estimate weighted norm of ũ using ∇k ũ. In the previous Section, we discuss the estimate
of ∇2 u. Based on these estimates, we can develop the weighted estimate of ũ with singular
weights. Recall ũ = u − ux (0)x, ṽ = v − vy (0)y = v + ux (0)y. We show how to estimate |x|−k ũ
using the estimate of |x|−l ∇k ũ.
96 JIAJIE CHEN AND THOMAS Y. HOU

Basic operators. Define the average operators

1 x 1 y
Z Z
E1 F (x, y) = F (z, y)dz, E2 F (x, y) = F (x, z)dz.
x 0 y 0
For i + j = 3, define
 1 Z xZ y 1 1/2
i j 2
Ci,j (x, y) , |∂ ∂ ψ̃(x, y)| dxdy
xy 0 0 |x|2 x y
where ψ̃ is the modified stream function ψ̃ = ψ − ψxy (0)xy.
Estimate of ũ/|x|3 . Since ũ is odd in x, to estimate ũ/|x|3 using ∇2 ũ/|x|, we use the following
estimates
Z xZ y Z x
1 1
ũ = ( ∂ xy ũ + ∂xx ũ(a, 0)(x − a)da)
|x|3 |x|3 0 0 0
Z xZ y
x3 1 x ũxx 2
Z
1  1 ũxy 2 1/2 xy|x| 1/2

≤ ( ( ) dxdy) √ + √ ( ( ) dx)
|x|3 xy 0 0 |x| 3 30 x 0 x
1  xy|x| x3 
= √ C1,2 (x, y) + √ C2,1 (x, 0) .
|x|3 3 30
Similarly, for ṽ, we yield
Z xZ y Z y
1 1 1  xy|x| y3 
ṽ = ( ∂ xy ṽ + ∂ yy ṽ(y − a)da) ≤ √ C 2,1 (x, y) + √ C 1,2 (0, y) .
|x|3 |x|3 0 0 0 |x|3 3 30
Next, we estimate ∇ũ. We have
Z y Z x
ũx 1
= ( ∂xy ũdy + ũxx (z, 0)dz)
|x|2 |x|2 0 0
r
1  ∂xy ũ 2 1/2 y2 ũxx 2 1/2 x2 
≤ 2
(E2 (( ) )) y x2 + + E1 (( ) ) √ .
|x| |x| 3 x 3
Using the incompressibility condition, we obtain the estimate for ṽy = −ũx .
For ṽx = vx , we have
Z x r
2
vx 1 1  vxx 1/2 x2
2
= 2
vxx (z, y)dz ≤ 2
E1 ( 2 ) x + y2 .
|x| |x| 0 |x| |x| 3
For ũy = uy , there are two different estimates. The first one is to use uy = vx + ω and the
above estimate. The second one is the following
r
u2xy 1/2
Z x
uy 1 1  x2
2
= 2
uxy (z, y)dz ≤ 2
E1 ( 2 ) x + y2 .
|x| |x| 0 |x| |x| 3
Using these estimates, we obtain the weighted estimate for ũ, ∇ũ near x = 0.
7.9. Estimate the velocity based on the sharp functional inequalities. Another way to
estimate the weighted L∞ and C 1/2 norm of u(F ), ∇u(F ) is to use the sharp nonlocal estimate
that we develop in Sections 4, 8. Since such inequalities for uA , ∇uA only involve the norm
||ω1 ϕ1 ||∞ and ||ω1 ψ||C 1/2 , we only need to estimate these norms for F . Yet, since F does not
g

vanish to O(|x|3 ) in general, we decompose F as follows

F = (F − Fxy (0)χF ) + Fxy(0) χF , F1 + F2 ,
where χF (x, y) is some cutoff function with χF (x, y) = xy for x, y near 0. Then we can decom-
pose the velocity as follows
u(F ) = u(F2 ) + u(F1 ) = u(F2 ) + uA (F1 ) + û(F1 ).
where û is the approximation term for u (2.89). Similarly, we perform a decomposition for ∇u.
We can further require χF = −∆χF,2 for another cutoff function. For example, we can choose
x3 y
χF,2 = − χF,3
2
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 97

for some cutoff function χF,3 with χF,3 = 1 + O(|x|4 ) near 0. In this case, we have
u(F2 ) = Fxy (0)u(χF ) = Fxy (0)u(−∆χF,2 ) = Fxy (0)∇⊥ χF,2 ,
which depends on χF,2 locally. For û(F1 ), we can write it as
X n Z
û(F1 ) = ai (F1 )ḡi (x), ai (F1 ) = F1 (y)qi (y)dy
i=1 R++
2

for some functions ḡi (x) and qi (y). We can estimate ∂xi ∂yj F1 in each piecewise small domain
[a, b] × [c, d]. Gluing them together, we can estimate the weighted norm of F1 and the above
integrals on F1 .

8. Estimate the norm of the velocity in the regular case

In the energy estimate in Section 4, we need to estimate the weighted L∞ and Hölder norm
of uA , ∇uA (4.2). The most singular part in these estimates can be obtained using the sharp
Hölder estimates in Section 3. In Section 4.2, we outline some ideas of obtainting sharp weighted
estimates of the nonlocal terms with computer assistance. In this section, we discuss in details
how to estimate them. We will use the scaling symmetries of the singular integral in a crucial
way. See the discussions in Section 4.2.1.
Difficulties in the computations. In addition to the difficulties discussed in Section 4.2.1,
the singular integral introduces several technical difficulties in our estimates. To address these
difficulties, we need to consider different scenarios and decompose the domain of the integrals
carefully in our computer assisted estimates. Given ωϕ ∈ L∞ , the velocity u and the commutator
(2.34) for ∇u is only log-Lipschitz. The logarithm singularity introduces several difficulties. For
example, if u is Lipschitz, a natural approach to estimate its Hölder norm in terms of ||ωϕ||∞
is to estimate the piecewise bound of u and ∂u, which are local, and then use the method in
Section G.6. However, since u is only log-Lipschitz, we need to perform a decomposition of u
into the regular part and the singular part carefully. For different parts, we will apply different
estimates. For ∇u, the estimates are more involved since it is more singular.
Another difficulty is that the odd extension W of ω from R2+ to R2 (3.3) is not Hölder
continuous in y near the boundary. The specific quantity W to be considered is relevant since
f = Kf (s) ∗ W for f = u, v, ux etc (3.4). To overcome this difficulty and estimate ∇u(x) for x
near the boundary effectively, we need to perform careful decomposition of the kernels with the
singular region adapted to the distance between x and the boundary.

8.1. Several strategies. Recall the discussion of scaling symmetry in Section 4.2.1.

8.1.1. Integral with approximation. In our computation of uA = u − û, ∇uA = ∇u − ∇u dA ,

where the approximation terms are defined in (2.82),R (2.88), (2.89), the rescaling
R argument still
applies. We consider one approximation term c(x) 1y∈S / K(x1 , y)ω(y)dy for K(x, y)ω(y) to
illustrate the ideas, where S is the singular region associated with x1 . Such an approximation
term relates to (2.82). Suppose that K is −2-homogeneous. We want to estimate
Z
I = ρ(x) (K(x, y) − c(x)K(xi , y)1y∈S
/ )W (y)dy,
R2

where W is the odd extension of ω from R2+ to R2 . Denote

(8.1) fλ (x) , f (λx).
We choose λ  |x| and denote x = λx̂, y = λŷ, x1 = λx̂1 . Then we have
Z  
2
I = ρλ (x̂) K(λx̂, λŷ) − c(λx̂)1λŷ∈S
/ K(λx̂1 , λŷ) W (λŷ)λ dŷ
R2
(8.2) Z  
= ρλ (x̂) K(x̂, ŷ) − c(λx̂)1ŷ∈S/λ
/ K(x̂ 1 , ŷ) Wλ (ŷ)dŷ.
R2
98 JIAJIE CHEN AND THOMAS Y. HOU

The singular region becomes S/λ and close to x1 /λ = x̂1 . For example, if S = {y : max |yi −
x1,i | ≤ a}, we have S/λ = {y : max |yi − x1,i | ≤ a/λ}. For the above integral, we can still
symmetrize the kernel and then estimate it similar to that in Section 4.2.1.
The bulk and approximation. To take advantage of the scaling symmetry and overcome
the singularity, in our computation for x away from the origin and not too large, we choose
several dyadic rescaling parameters λ = 2i , i ∈ I, e.g. I = {−4, −3, .., 10}. Then for any x with
max(x1 , x2 ) ∈ [2i xc , 2i+1 xc ], we can choose λ = 2i so that the rescaled x̂ = λx satisfies
(
[xc , 2xc ] × [0, 2xc ], if x2 ≤ x1 ,
(8.3) x̂ ∈
[0, xc ] × [xc , 2xc ], if x2 > x1 .

We also choose the xi and the size of the singular region ti for the approximation term (2.82)
such that xi /λ is on the grid point of the mesh and the boundary of the singular region {y :
|xi − y1 | ∨ |y2 | ≥ ti /λ} aligns with one of the edges of a mesh cell. For example, this can be
done by choosing the following y mesh in the near-field to discretize the y-integral, xi , and ti
y1,i = ih, y2,i = ih, xi = 2ni h, ti = 2mi h.
Then when we discretize the rescaled integral in y, e.g. (8.2), the singular region is the union of
several mesh cells. For large y, it is away from the singularity x̂. Then we can use an adaptive
mesh in y1 , y2 to discretize the integral.
We remark that in (8.2), if x1 6= 0 and x1 /λ is too large or too small, since c(x) is supported
near xi (see (2.82)), c(λx̂) will be 0. This means that when we compute uA (x), ∇u dA , if the
coefficient of an approximation term with center xi and parameter ti is nonzero, e.g., c(x) 6= 0,
then λ is comparable to x1 when we rescale the integral by λ. Thus x̂1 = x1 /λ is on the grid.
We also choose ti such that ti /λ is a multiple of mesh size h for λ comparable to xi .
Remark 8.1. Using the scaling symmetry and rescaling the integral by dyadic scales, we can
compute the integral for x ∈ [0, D]2 \[0, d]2 with roughly O(log(D/d)) computational cost.
The near-field and the far-field. If x is sufficiently small, i.e. max(x1 , x2 ) < mini∈I 2i xc ,
we choose λ = max(x1 , x2 )/xc so that the rescaled x̂ = λx is on the line x1 = xc or x2 = xc .
Assuming ϕ(x) ≥ |x|−β1 |x|−β
1
2
, ρ ∼ |x|−α near x = 0, and K is −l-homogeneous, then we get
Z Z
|ρ(x) K(x, y)ω(y)dy| ≤ ||ωλ ϕλ ||L∞ ρλ (x̂) |K(x̂, ŷ)|ϕλ (ŷ)−1 λ2−l dŷ
R++
2 R++
2
(8.4) Z
≤ ||ωλ ϕλ ||L∞ λβ1 +β2 +2−l ρλ (x̂) |K(x̂, ŷ)||ŷ|β1 ŷ1β2 dŷ.
R++
2

As x → 0, λ → 0. The factor λβ1 +β2 +2−l absorbs the large factor λ−α in ψλ (x̂). In our
estimate of uA , ∇uA , we have β1 + β2 = 3 for ϕ1 (4.14), l = 2, α = 2 for ψ1 (C.1), and
β1 + β2 + 2 − l − 2 = 1 > 0.
In general, the above integral may not be integrable due to the growing weight |y|β1 y1β2 . For
uA , ∇uA with small x, it takes the form (2.84), (2.81)
Z
4 y1 y2
(8.5) f (x) − Cf 0 (x)ux (0) − Cf (x)K00 = (Kfsym + Cf 0 (x) − Cf (x)K00 (y))ω(y)dy,
R2++ π |y|4

where Cf 0 , Cf are defined in (2.80), and f = u, v, ux , vx , uy , vy . In particular, the associated

kernel has a much faster decay rate |y|−6 , which will be shown in (D.16), (D.17). Thus, the
integral is integrable.
Since λ = max(x1 , x2 )/xc is very small, ρλ (x̂) can be well approximated by the most singular
power cλ−α |x|−α for some c > 0, which can be estimated effectively after factorizing out λ−α .
Similarly, if x is sufficiently large, i.e. max(x1 , x2 ) > maxi∈I 2i+1 xc , we choose λ = max(x
xc
1 ,x2 )

so that the rescaled x̂ = x/λ is on the line x1 = xc or x2 = xc . Since λ is sufficiently large, we

can estimate the weight ρλ , ϕλ based on their asymptotic behavior.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 99

8.1.2. The scaling relations. We discuss several scaling relations, which will be useful in later
computation. For a −d-homogeneous kernel K, i.e., K(λx) = λ−d K(x), we have
Z Z
I(x) = ρ(x) K(x, y)ω(y)dy = ρλ (x̂) K(x̂, ŷ)ωλ (ŷ)λ2−d dŷ , λ2−d Iλ (x̂),

where x = λx̂, y = λŷ. To compute the derivative of I(x), using the chain rule, we have
dx̂i
∂xi I(x) = λ2−d ∂x̂ Iλ (x̂) = λ1−d ∂x̂i Iλ (x̂).
dxi i
For the L∞ part, clearly, we get |I(x)| = |Iλ (x̂)|. To compute the Hölder norm, we use the
following relation |x − z| = λ|x̂ − ẑ| and
|I(x) − I(z)| |Iλ (x̂) − Iλ (ẑ)|
1/2
= λ−1/2 .
|x − z| |x̂ − ẑ|1/2
In particular, we have
(8.6) ||ωλ ϕλ ||∞ = ||ωϕ||∞ , [ωλ ψλ ]C 1/2 = λ1/2 [ωψ]C 1/2 , i = 1, 2.
xi xi

Using these scaling relations, we can perform the estimate in a rescaled domain with any λ > 0.

8.1.3. Mesh and the Trapezoidal rule. After rescaling the integral with suitable scaling factor λ,
we can restrict the rescaled singularity x̂ ∈ [0, 2xc ]2 \[0, xc ]2 (8.2),(8.3).
If a domain Q is away from the singularity x̂ of the kernel, applying (8.6), we get
(8.7)
Z Z Z
|K(x̂, y)||ωλ (y)|dy ≤ ||ωλ ϕλ ||∞ |K(x̂, y)|ϕ−1
λ (y)dy = ||ωϕ||∞ |K(x̂, y)|ϕ−1
λ (y)dy.
Q Q Q

Then, it suffices to estimate the integral of an explicit function |K(x̂, y)|ϕ−1 λ (y). If in addition,
the region Q is small, e.g. Q is the grid [yi , yi+1 ] × [yj , yj+1 ] introduced below, we further apply
Z Z
|K(x̂, y)||ωλ (y)|dy ≤ ||ωϕ||∞ ||ϕ−1λ || ∞
L (Q) |K(x̂, y)|dy.
Q Q

Since
R the domain Q is small, the estimate is sharp. We use the following method to estimate
|K(x̂i , y)|dy for a suitable kernel K and x̂i on the grid points.
We consider the estimate of the L1 norm of some function f in R2++ , e.g. f = K(x̂i , y)
mentioned above. To discretize the integral, we design uniform mesh in the domain [0, b]2
covering Ω1 and Ω2 with mesh size h and adaptive mesh in the larger domain [0, D]2
(8.8) 0 = y0 < y1 < .. < yn = D, yi = ih, i ≤ b/h.
2
The finer mesh in the near field [0, b] allows us to estimate the integral with higher accuracy.
We choose sparser mesh in the far-field since y is away from the singularity x̂ and the kernel
decays in y. We partition the integral as follows
Z X Z Z
(8.9) |f (y)|dy = |f (y)|dy + |f (y)|dy.
R++
2 0≤i,j≤n−1 [yi ,yi+1 ]×[yj ,yj+1 ] y ∈D
/

We focus on how to estimate the first part for nonsingular f . In Section 8.4, we estimate
the integral beyond [0, D]2 using the decay of the integral. We will discuss how to estimate the
integral near the singularity of the kernel in a later section.
Denote Q = [a, b] × [d, c], h1 = b − a, h2 = d − c. We use the Trapezoidal rule
Z
|f (y)|dy ≤ T (|f |, Q) + Err(f ),
[a,b]×[c,d]

where
(b − a)(d − c)
T (f, Q) , (f (a, c) + f (a, d) + f (b, c) + f (b, d)).
4
The error estimate of the above Trapezoidal rule is not obvious due to the absolute sign. In
fact, even if f is smooth, |f | is only Lipschitz near the zeros of f . Since the set of zeros is hard
100 JIAJIE CHEN AND THOMAS Y. HOU

to characterize and that |f | can have low regularity, we do not pursue higher order quadrature
rule. We have the following error estimate.

Lemma 8.2 (Trapezoidal rule for the L1 integral). For f ∈ C 2 (Q), we have

|Q| 2
Z
|f (y)|dy ≤ T (|f |, Q) + (h ||fxx ||L∞ (Q) + h22 |fyy ||L∞ (Q) ).
Q 12 1

Remark 8.3. The above estimate shows that the Trapezoidal rule remains second order accurate
from the above. In particular, this error estimate is comparable to the case without taking the
absolute value.

Proof. Define the linear interpolation of f in Q

4
X
L(f ) = λi (x)fi , E(f ) = f − L(f ),
i=1
P
where λi (x) is linear and satisfies λi (x) = 1 and λi (x) ≥ 0 for x ∈ Q. Using the triangle
inequality, we obtain
Z Z Z Z
|f |dy ≤ |E(f )|dy + λi (x)|fi |dy = T (|f |, Q) + |E(f )|dy.
Q Q Q Q

We have the standard error bound for E(f )

||fxx ||L∞ (Q) ||fyy ||L∞ (Q)

(8.10) |E(f )| ≤ |(x − a)(x − b)| + |(y − c)(y − d)|.
2 2
It can also be established using the error estimate for the 2D Lagrangian interpolation
R1 with k = 2
(G.11) in Section G.2. Integrating the above estimate in x, y and using 12 0 t(1 − t)dt = 12 1

conclude the proof. 

2
R
To estimate the integral |K(x, y)| for all x̂ ∈ Ω1 , Ω2 (8.3), we discretize
R [0, 2a] using uniform
mesh with mesh size hx = h/2. We use the above method to estimate |K(x̂i , y)|dy for xi on
the grid points. After we estimate the derivatives of the kernel, we use the following Lemma to
estimate the integral for any x in a domain.

Lemma 8.4. Suppose that K(x, y) ∈ C 2P (P × D), P = [a1 , b1 ] × [a2 , b2 ], hi = bi − ai , i = 1, 2,

and Q = [a, b] × [c, d]. Let L(K)(x, y) = i,j=1,2 λij (x)K((ai , bj ), y) be the linear interpolation
of K(x, y) in x using K((ai , bj ), y), i, j = 1, 2. Then for any x ∈ P , we have
Z X Z  h2 h2 
1
|K(x, y)|dy ≤ λij (x) |K((ai , bj ), y)|dy + ||Kxx ||L∞ (P ×Q) + 2 ||Kyy ||L∞ (P ×Q) |Q|.
Q i,j=1,2 Q 8 8

The proof follows from (8.10), the triangle inequality and 12 |t(1 − t)| ≤ 18 for t ∈ [0, 1]. We will
P Lemma and sum Q over all the near-field domains Qij = [yi , yi+1 ] × [yj , yj+1 ]
apply the above
(8.8). Since ij λij (x) = 1, we can simplify the first term as follows
X X Z X Z
λij (x) |K((ai , bj ), y)|dy ≤ max |K((ai , bj ), y)|dy.
Qkl i,j Qkl
i,j=1,2 k,l≤n k,l≤n

Therefore, it suffices to estimate the integral for x on the grid points and the piecewise
derivative bounds of the kernel.
We apply Lemmas 8.2, 8.4 to estimate the weighted integral related to the velocity. The
integrands take the form (8.24),(8.25), (8.18). To estimate the error in the above integrals, we
need to obtain piecewise L∞ estimate of the derivatives of the integrands in P, Q. We estimate
the derivatives of the weights in Appendix C.3, the kernel in Appendix D.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 101

Parameters for the integrals. In our computation, we choose

(8.11) hx = 13 · 2−12 , h = 13 · 2−11 , xc = 13 · 2−5 ,

which can be represented exactly in binary system, to reduce the round off error. The ap-
proximate values of the above parameters are hx ≈ 0.0032, h ≈ 0.0064, xc ≈ 0.4. For x ∈
[0, 2xc ]2 \[0, xc ]2 (8.3), we have

(8.12) max(x1 , x2 ) ≥ xc = 64h = 128hx .

In our decomposition of the integral, e.g. (8.18), (8.34), (8.37), we impose a constraint on
the size of the singular region to satisfy kh < xc such that the region does not cover the origin.

8.1.4. Decomposition, commutators and the Lipschitz norm. The most difficult part of the com-
putation is to estimate the Hölder norm of ∇u, and we discuss several strategies. In this
computation, we cannot first estimate the local Lipschitz norm of ∇u and then obtain the local
Hölder norm due to the difficulties discussed at the beginning of Section 8. We need to decom-
pose the integral related to ∇u into several parts according to the distance between y and the
singularity and use different estimates for different parts.
We focus on the integral related to ux without subtracting any approximation term and
assume that x ∈ [0, 2xc ]2 \[0, xc ]2 . The approximation term ∇udA is nonsingular and can be
estimated using the method in Section 8.1.3. Let h be the mesh size in the discretization of the
integral in y. Suppose that x ∈ [ih, (i + 1)h) × [jh, (j + 1)h). Denote by R(x, k) the rectangle
covering x

(8.13) R(x, k) , [(i − k)h, (i + 1 + k)h] × [(j − k)h, (j + 1 + k)h]

for any k > 0. If k ∈ Z + , the boundary of R(x, k) is along with the mesh grid and is at least
kh away from x. Denote by Rs , Rs,1 , Rs,2 different symmetric rectangles with respect to x

Rs (x, k) , [x1 − kh, x1 + kh] × [x2 − kh, x2 + kh],

(8.14) Rs,1 (x, k) , [x1 − kh, x1 + kh] × [(j − k)h, (j + 1 + k)h],
Rs,2 (x, k) , [(i − k)h, (i + 1 + k)h] × [x2 − kh, x2 + kh].

Clearly, we have Rs (x, k) ⊂ Rs,1 (x, k), Rs,2 (x, k) ⊂ R(x, k). We introduce the upper and
lower parts of the rectangle

(8.15) R+ (x, k) , R(x, k) ∩ {y : y2 ≥ x2 }, R− (x, k) , R(x, k) ∩ {y : y2 ≤ x2 }.

We use similar notations for Rs (x, k), Rs,1 (x, k), Rs,2 (x, k). We further introduce the intersection
of the rectangle and four half planes with reflection
R(x, k, N ) = R(x, k) ∩ {y : y2 ≥ 0}, R(x, k, S) = R2 (R(x, k) ∩ {y : y2 ≤ 0}),
(8.16)
R(x, k, E) = R(x, k) ∩ {y : y1 ≥ 0}, R(x, k, W ) = R1 (R(x, k) ∩ {y : y1 ≤ 0}),

where N, E, S, W are short for north, east, south, west, respectively and the reflection operators
R1 , R2 are given by

R1 (y1 , y2 ) = (−y1 , y2 ), R2 (y1 , y2 ) = (y1 , −y2 ).

It is clear that R(x, k, S) ⊂ R++

2 , R(x, k, W ) ⊂ {y : y1 ≥ 0}. An illustration of these domains is
given in Figure 10. If x, y ∈ R++
2 , we have the equivalence

(8.17) (y1 , −y2 ) ∈

/ R(x, k) ⇐⇒ (y1 , −y2 ) ∈
/ R(x, k) ∩ {y : y2 ≤ 0} ⇐⇒ y ∈
/ R(x, k, S).

The above notations will be very useful in our later decomposition of the symmetrized kernel.
102 JIAJIE CHEN AND THOMAS Y. HOU

Figure 10. Left: The large box is R(x, k) and the red box is Rs,1 (x, k). The
small box containing x has size h × h. Right: The upper box is R(x, k, N ), and
the shaded box is R(x, k, S), the reflection of the region below the y-axis.

Recall the odd extension W of ω in R2 (3.3). For simplicity, we drop the x variable in the R
notation. For k > k2 , k, k2 ∈ Z + , we decompose the weighted ux (x) integral as follows
(8.18) Z Z Z
ψ(x) K1 (x − y)W (y)dy = ψ(x) K1 (x − y)W (y)dy + K1 (x − y)ψ(y)W (y)dy
R(k)c Rs,1 (k)
Z Z
+ K1 (x − y)ψ(y)W (y)dy + K1 (x − y)(ψ(x) − ψ(y)W (y)dy)
R(k)\Rs,1 (k) R(k)\R(k2 )
Z
+ K1 (x − y)(ψ(x) − ψ(y))W (y)dy
R(k2 )

, I1 (x, k) + I2 (x, k) + I3 (x, k) + I4 (x, k, k2 ) + I5 (x, k2 ),

where K1 (s) is defined in (3.1), and we have dropped the constant − π1 in ux (x) at this moment
to simplify the notation. For the regular part, we use numerical computation in Section 8.1.3
to estimate the derivatives. For the singular part, we will use a change of variables y = x + s to
localize our estimate to the singularity.

8.1.5. Symmetrization. After we obtain the decomposition, we use the odd symmetry of W in
y1 , y2 to symmetrize the integral and reduce the integral over R2 to the first quadrant R++ 2 .
This enables us to exploit the cancellation in the integral and obtain a sharper estimate. In our
computation, we symmetrize the integrals in I1 (x, k) and I4 (x, k, k2 ), which are more regular.
For a given kernel K(x, y), we denote by K sym the symmetrization of K
(8.19) K sym (x, y) , K(x, y) − K(x, −y1 , y2 ) − K(x, y1 , −y2 ) + K(x, −y).
We show how to symmetrize I1 (x, k) as an example. Suppose that
(8.20) x ∈ R++
2 , x2 ≤ x1 , x ∈ Bi1 ,j1 (hx ) ⊂ Bij (h), j ≤ i,
where hx = h/2 and Blm (r) is defined as
(8.21) Blm (r) = [lr, (l + 1)r] × [mr, (m + 1)r].
Recall the notations in (8.16). We choose k < i so that R(x, k) ⊂ {y : y1 > 0} and R(x, k, W ) =
∅. By definition (8.13), the domains R(x, k), R(x, k, N ), R+ (x, k) etc are the same for all x ∈
Bi1 ,j1 (hx ). Yet, R(x, k) may cross the boundary y2 = 0, i.e. R(x, k, S) 6= ∅. See the right figure
in Figure 10 for a possible configuration. Using the equivalence (8.17) and the property that W
is odd in y1 and y2 , we can symmetrize I1 (x, k) as follows
Z 
I1 (x, k) = ψ(x) K1 (x − y)1y∈R(k)c − K1 (x1 − y1 , x2 + y2 )1y∈R(k,S)
/
R++
2

− K1 (x1 + y1 , x2 − y2 ) + K1 (x + y) ω(y)dy.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 103

For x in a general location, the symmetrization of I1 reads

Z 
I1 (x, k) = ψ(x) K1 (x − y)1y∈R(k)c − K1 (x1 − y1 , x2 + y2 )1y∈R(k,S)
/
(8.22) R++
2

− K1 (x1 + y1 , x2 − y2 )1y∈R(k,W
/ ) + K1 (x + y) ω(y)dy.

For I4 (x) (8.18), we will choose weight ψ(y) that is even in y1 , y2 . Thus, the symmetrization
of I4 is
Z 
I4 (x, k, k2 ) = K1 (x − y)1y∈R(k)\R(k2 ) − K1 (x1 − y1 , x2 + y2 )1y∈R(k,S)\R(k2 ,S)
(8.23) R++
2

− K1 (x1 + y1 , x2 − y2 )1y∈R(k,W )\R(k2 ,S) (ψ(x) − ψ(y))W (y)dy.

In the above formula, we do not have the term K1 (x + y) since for y ∈ R++2 , x+y ∈ / R(k). Thus
after symmetrizing the kernel in I4 , we do not have such a term.
Though the symmetrized kernel is complicated, since these regions R(k), R(k, α), α = N, E
(8.13), (8.16) can be decomposed into the union of the mesh girds [yi , yi+1 ] × [yj , yj+1 ], in each
grid, the indicator functions are constants. See also Remark 8.6. In particular, in each grid
y ∈ [yi , yi+1 ] × [yj , yj+1 ], we can write the integrand in I1 + I4 as
(8.24) J = K N C (x, y) · ψ(x) + K C (x, y) · (ψ(x) − ψ(y))
where N C, C are short for non-commutator, commutator, respectively, and K N C , K C can be
determined by the distance between [yi , yi+1 ] × [yj , yj+1 ] and Bi1 j1 (hx ). For y away from x, e.g.
|y1 | ∨ |y2 | ≥ 4xc in our computation, we have
(8.25) J = K sym (x, y)ψ(x).
8.1.6. Integral in domains depending on x. In the computation, we need to estimate several
integrals in the domains depending on x, e.g. I3 in (8.18). We use the L∞ estimate of I3 to
illustrate the ideas. A direct estimate yields
Z
|I3 (x)| ≤ ||W ϕ||∞ |K1 (x − y)|ψ(y)ϕ−1 (y)dy.
R(k)\Rs,1 (k)

We cannot apply the method in Section 8.1.3 to first estimate I3 (x) for x on the grid points
and then estimate ∂ 2 I3 (x) for the error since the kernel is singular and the error part associated
with ∂ 2 I3 (x) is more singular (see Lemma 8.4).
Denote f = ψϕ−1 . We consider a change of variable y = x + s to center our analysis around
the singularity x. The domain for s is
(8.26) {y ∈ R(k)\Rs,1 (k)} = {s ∈ R(k) − x} ∩ {|s1 | ≥ kh} , D(x, k).
It suffices to estimate
Z
(8.27) J= |K1 (−s)|f (x + s)dy
s∈D(x,k)

for all x ∈ Bi1 ,j1 (x) (8.20). We want to futher simplify the above domain so that it does not
depend on x. Recall the location of x (8.20). To obtain a sharp estimate, we further partition
the location of x ∈ Bi1 ,j1 (hx ) as follows
(8.28) Aa = [i1 hx + ahx /m, i1 hx + (a + 1)hx /m], Bb , [j1 hx + bhx /m, j1 hx + (b + 1)hx /m],
for some m ∈ Z + and 0 ≤ a, b ≤ m − 1. Clearly, Aa × Bb is a partition of Bi1 j1 (hx ). Recall
(8.20) and (8.13). We have
R(x, k) = [(i − k)h, (i + 1 + k)h] × [(j − k)h, (j + 1 + k)h].
Now, for x ∈ Aa × Bb , since |s1 | ≥ kh, we have
(8.29)
s1 = y1 − x1 ∈ [(i − k)h − i1 hx − (a + 1)hx /m, −kh] ∪ [kh, (i + 1 + k)h − i1 hx − ahx /m]
, Xl,a ∪ Xr,a ,
104 JIAJIE CHEN AND THOMAS Y. HOU

Figure 11. The largest box is R(x, k). Left: The left and right blue re-
gions are Xl,a × Ym,b , Xr,a × Ym,b . The four red regions correspond to
Xα,a × Yβ,b , α = l, u, β = d, u. Right: Illustration of R(x, k)\Rs (x, k) and
Rs (x, k2 ). R(x, k)\Rs (x, k) consists of the blue and the red regions.

where the subscripts l, r are short for left, right, respectively. Similarly, for s2 , we have
(8.30)
s2 = y2 − x2 ∈ [(j − k)h − j1 hx − (b + 1)hx /m, (j + k)h − j1 hx − bhx /m]
, [(j − k)h − j1 hx − (b + 1)hx /m, −kh] ∪ [−kh, kh] ∪ [kh, (j + 1 + k)h − j1 hx − bhx /m]
, Yd,b ∪ Ym,b ∪ Yu,b
where the subscripts d, m, u are short for down, middle, upper, respectively. Note that the
intervals X, Y do not depend on x. We have
(8.31) D(x, k) ⊂ (Xl,a ∪ Xr,a ) × (Yd,b ∪ Ym,b ∪ Yu,b ).
Now, we can decompose J (8.27) as follows
X Z
J≤ Jα,β , Jα,β , |K1 (−s)|f (s + x)dy, α = l, r, β = d, m, u.
α=l,r,β=d,m,u Xα,a ×Yβ,b

See the left figure in Figure 11 for different domains in the above decomposition. From the
definitions of X, Y , the total width of the left and the right domains Xα,a ×(Yd,b ∪Ym,b ∪Yu,b ), α =
l, u is
|Xl,a | + |Xr,a | = h + hx /m.
For a fixed x, from the definition (8.13), the width of R(k)\Rs,1 (k) is h. We choose a large m
and further partition the location of x so that we do not overestimate the region too much.
For a small domain Q = [a, b] × [c, d] , we can estimate the integral as follows
Z Z
(8.32) |K1 (−s)|f (x + s)ds ≤ |K1 (−s)|ds||f ||L∞ (Bi1 j1 (hx )+Q) .
Q Q

Since Q is given, K1 (s) is explicit and has scaling symmetries, we can estimate the integral
2
of |K1 (s)| Reasily. For example,R if Q = [ah, bh] , we can use the scaling symmetries of K1 (s)
β
to obtain Q |K1 (−s)| = h [a,b]2 |K1 (−s)| for some β. Moreover, for many kernels in our
computations, e.g. K(s) = s|s| 1 s2
4 , we have explicit formulas for the integral.

We apply the above method to estimate the integral in Xα,a × Yβ,b , α = l, r, β = d, u (red
region in Figure 11). Since Ym,b = [−kh, kh], for the integral in Xα,a × Ym,b (blue region), we
further decompose it
X Z
(8.33) Jα,m = |K1 (−s)|f (s + x)dy,
−k≤t≤k−1 Xα,a ×[th,(t+1)h]

and then apply the above method to estimate it.

Next, we further simplify ||f ||L∞ (Bi1 j1 (hx )+Q) in the above estimate. Firstly, from (8.20), we
get
ih ≤ i1 hx < (i1 + 1)hx ≤ (i + 1)h, jh ≤ j1 hx < (j1 + 1)hx ≤ (j + 1)h.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 105

For Xl,a (8.29) with 0 ≤ a ≤ m − 1, we have the lower bound for the endpoint

(i − k)h − i1 hx − (a + 1)hx /m ≥ (i − k)h − i1 hx − hx ≥ (i − k)h − ((i + 1)h − hx ) − hx = −kh − h.

Similarly, we can cover the intervals of X, Y (8.29), (8.30) uniformly for 0 ≤ a, b ≤ m − 1 and
obtain
Xl,a ⊂ [(i − k)h − i1 hx − hx , −kh] ⊂ [−(k + 1)h, −kh],
Xu,a ⊂ [kh, (i + 1 + k)h − i1 hx ] ⊂ [kh, (k + 1)h],
Yd,b ⊂ [−(k + 1)h, −kh], Yd,u ⊂ [kh, (k + 1)h].
Thus, we only need to estimate the L∞ norm of f in

Qi1 j1 (hx ) + [αh, (α + 1)h] × [βh, (β + 1)h], α = −k − 1, k, β = −(k + 1), −k, .., k.

These estimates are independent of the choice of m, a, b. Since the size of each domain is at
most 2h × 2h, the above estimates based on (8.32) are sharp. We will estimate the piecewise
bound of the weights ψ, ϕ in later sections.
Using the above decomposition and estimates, we obtain the estimate of J (8.27) for x ∈
Aa × Bb (8.28). Similarly, we can estimate J for any 0 ≤ a, b ≤ m − 1. Taking the maximum of
these m2 estimates, we obtain the estimate of J and I3 (x) for all x ∈ Bi1 j1 (hx ).

8.1.7. First generalization: the boundary terms. We generalize the above ideas to estimate in-
tegrals in other domains depending on x. The first generalization is to estimate the boundary
term. We estimate the x1 −derivative of I3 (x) (8.18) to illustrate the ideas. In the estimate of
the x−derivative, we have an extra boundary term I32
Z Z (j+1+k)h x1 +kh
∂1 I3 (x) = ∂x1 K1 (x−y)(W ψ)(y)dy− K1 (x−y)(W ψ)(y) dy2 , I31 +I32 ,
R(k)\Rs,1 (k) (j−k)h y1 =x1 −kh

where we have used the doamin for R(x, k) (8.13).

For I31 , we apply the previous method to estimate it. Denote Γk , [j − k)h, (j + 1 + k)h].
Using a change of variable y = x + s, we can rewrite I32 as follows
Z  
I32 = − K1 (−kh, −s2 )(W ψ)(x1 +kh, x2 +s2 )−K1 (kh, −s2 )(W ψ)(x1 −kh, x2 +s2 ) ds2 .
s2 ∈Γk −x2

We partition the location of x and assume x ∈ Aa × Bb ⊂ Bi1 ,j1 (hx ) (8.28). From (8.30), we
have
s2 ∈ Γk − x2 ⊂ Yd,b ∪ Ym,b ∪ Yu,b .
Using the above decomposition and |W ψ(x)| ≤ ||W ϕ||∞ f (x), f = ψϕ−1 , we obtain
X Z
|I32 | ≤ ||W ϕ||∞ Mα,β , Mα,b , |K1 (−αkh, −s2 )| · |f (x1 + αkh, x2 + s2 )|ds2 ,
α=±,β=d,m,u Yβ,b

for α = ±, β = u, m, d. For β = u, d, the domain Yβ,b is small |Yβ,b | ≤ h. We apply the method
in (8.32) to estimate Mα,β . The only difference is that we need consider a 1D integral here
Z
|K1 (−αkh, −s2 )|ds2
Q

for some interval Q, rather than a 2D integral in (8.32). For Mα,m , we decompose the domain
Ym,b into small intervals with length h similar to (8.33) and then apply the method in (8.32).
We combine these estimates to bound I32 for x ∈ Aa × Bb . Then, we maximize the estimates
over 0 ≤ a, b ≤ m − 1 to bound I32 for x ∈ Bi1 ,j1 (hx ).
106 JIAJIE CHEN AND THOMAS Y. HOU

8.1.8. Second generalization. In some of the computations, we need to estimate

Z
J= |K(x − y)|f (y)dy
R(k)\Rs (k2 )
+
for some k2 < k with 2k2 , k ∈ Z , where Rs (k) is defined in (8.14). Similarly, we introduce
y = x + s and use
Rs (k2 ) ⊂ Rs (k) ⊂ R(k), R(k)\Rs (k2 ) = R(k)\Rs (k) ∪ Rs (k)\Rs (k2 ),
to obtain
Z Z
J =( + )K(−s)f (x + s)dy , J1 + J2 .
s∈R(k)−x,|s1 |∨|s2 |≥kh k2 h≤|s1 |∨|s2 |≤kh

The estimate of J1 is similar to J in (8.27). See the right figure in Figure 11 for different
regions. Compared to R(k)\Rs,1 (k), the domain R(k)\Rs (k) contains two more parts
Xm,a , [−kh, kh], Xm,a × Yu,b , Xm,a × Yd,b ,
i.e., the upper and lower blue regions in the right figure in Figure 11. The estimate of the
integral in these regions is similar to that in Xα,a × Ym,b in (8.31).
For J2 , the domain is simpler. Since 2k2 ∈ Z + , we partition the domain into hx × hx grids
X Z
J2 = |K(−s)|f (s+x)ds, Sl , {−k ≤ c < k, −k ≤ d < k}.
(c,d)∈Sk \Sk2 [chx ,(c+1)hx ]×[dhx ,(d+1)hx ]

For each integral, we estimate it using the method in (8.32). The remaining steps are the same
as those of J in (8.27) studied previously.
Remark 8.5. In the estimates in Section 8.1.6-8.1.8, we use the important property that the
weights are locally smooth to move them outside the integral. Moreover, we use the fact that
the singular region depend
R on x smoothly and monotonously so that we can cover it effectively.
Note that the integral Q |K1 (s)|dy for different Q, a, b in the above estimates does not depend
on x. We can first compute these integrals once and store them. Then use them in the estimate
of different x.
8.1.9. Hölder estimate of log-Lipschitz function. In some computation, we need to perform C 1/2
estimate of some log-Lipschitz function. We consider a simple example to illustrate the ideas
Z
F (x) = K(x, y)f (y)dy, |K(x, y)| ≤ C1 |x − y|−1 , |∂K(x, y)| ≤ C2 |x − y|−2 ,
maxi |xi −yi |≤b

for some constant C1 , C2 . Given f ∈ L∞ , F is log-Lipschitz. To estimate [f ]C 1/2 , we cannot

x
1/2
first estimate the piecewise values of f and ∂x f and then combine them to obtain the Cx
estimate. Instead, given x, z, for a to be determined, we decompose F into the smooth part and
the singular part
Z Z
F1 (x) , K(x, y)f (y)dy, F2 (x) , K(x, y)f (y)dy,
a≤maxi |xi −yi |≤b maxi |xi −yi |≤a
Using the assumptions of the kernel, we have
b
|∂x1 F1 (x)| ≤ C3 log ||f ||∞ , |F2 (x)| ≤ C4 |a| · ||f ||∞ ,
a
where the constants C3 , C4 depend on b, C1 , C2 . Applying the above estimates, we obtain
|F (x) − F (z)| |F1 (x) − F1 (z)| + |F2 (x) − F2 (z)|  b 1/2 −1/2

≤ ≤ C 3 log ·|x 1 −z 1 | +2C4 |a||x 1 −z1 | ||f ||∞ .
|x1 − z1 |1/2 |x1 − z1 |1/2 a
We optimizing the estimates by choosing a = C5 |x1 − z1 | for some constant C5 depending on
C3 , C4 . Then we establish the estimate. The above simple estimates show that the choice of a
depends on |x − z|. Thus, in our later Hölder estimates, we perform decomposition guided by
the above estimates and optimize the choice of size of the singular region [−a, a]2 . On the other
hand, since for different |x − z|, we need to choose different a, it increases the technicality of the
computer-assisted estimates.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 107

8.2. L∞ estimate. Let ûx,A be the approximation term of ux (2.89). We focus on the estimate
of the piecewise L∞ norm of ux,A = ux − ûx,A (4.2), which is a representative case. For simplicity,
we assume the rescaling factor λ = 1. We assume that x satisfies (8.20) without loss of generality.
We want to estimate ux,A for all x ∈ Bi1 j1 (hx ).
We can write ux,A = ux − ûx as follows
Z
ux,A = (K(x − y) − K̂(x, y))W (y)dy, KA , K(x − y) − K̂(x, y),

where K̂(x, y) is the kernel for the approximation term. From (2.82), (2.88), it is nonsingular.
Given x with (8.20), similar to (8.18), for k ≥ k2 , we perform the following decomposition
Z Z Z Z
ux,A = ( + + )K(x − y)W (y)dy − K̂(x, y)W (y)dy
(8.34) R(k)c R(k)\Rs (k2 ) Rs (k2 )

, I1 + I2 + I3 + I4 .
where Rs (k) is the symmetric singular region (8.14). See Section 8.2.4 for the choice of k.
Since I1 + I4 is nonsingular, we use the ideas in Section 8.1.5 to symmetrize the kernels in
I1 + I4 . Then we use the method in Section 8.1.3 to estimate it.
Remark 8.6. In our computation, the domain [0, D]2 ∩ R(k)c can be decomposed into the union
of small grids [yi , yi+1 ]×[yj , yj+1 ] (8.8) since the boundary of R(x, k) aligns with the mesh (8.13).
In particular, in each grid, the indicator function is constant, and the integrand is smooth in y.
Next we consider I2 . The domain of the integral is close to the singularity. If we use the
method in Section 8.1.3 to estimate it, the error will be quite large since ∂ 2 K(x − y) is very
singular. We want to estimate I2 using ||W ϕ||∞ and the singular part I3 using [W ψ1 ]C 1/2 . Since
K(z) is singular of order −2, we expect an estimate
k −1 1/2
|I2 | + |I3 | . log ϕ (x)||W ϕ||L∞ [R(k)] + ψ −1 (x)k2 [W ψ]C 1/2 .
k2 x

Note that the weights ϕ, ψ have a different order of singularity for small x and a different rate
of decay. Moreover, we need to control the right hand side using the energy, e.g. E1 (t) (4.21),
which assigns different weights to two norms (seminorms). Thus, to obtain a sharp estimate, we
need to optimize the choice of k2 .
Firstly, we consider k2 = 2, 2 + 12 , .., k, we use the method in Section 8.1.6 to estimate I2 . We
also consider very small k2 < 2. In this case, we further decompose I2 as follows
Z Z
I2 = ( + )K(x − y)W (y)dy , I21 + I22 .
R(k)\Rs (2) Rs (2)\Rs (k2 )

For I21 , we apply the method in Section 8.1.6. For I22 , we use a change of variables y = x+sh
Z
|I22 | = K(−sh)W (x + sh)h2 ds .
k2 ≤|s1 |∨|s2 |≤2

Since the region is very small, x + sh ∈ Bi1 j1 (hx ) + [−2h, 2h], and K1 (hs) = h−2 K1 (s), we get
Z
|I22 | ≤ ||W ϕ||∞ ||ϕ−1 ||L∞ (Bi1 j1 (hx )+[−2h,2h]) |K(s)|ds.
k2 ≤|s1 |∨|s2 |≤2

The integral can be computed explicitly and has the order log k22 .
It remains to estimate the most singular part I3 for different k2 . Using a change of variables
y = x + sh, the scaling symmetries, and the above derivations, we get
Z
I3 = K(−s)W (x + sh)ds.
[−k2 ,k2 ]2

To use the Hölder norm of W ψ, we decompose it as follows

Z
1 1 (W ψ)(x + sh)
I3 = K(−s)(W ψ)(x + sh)( − ) + K(−s) ds , I31 + I32 .
[−k2 ,k2 ]2 ψ(x + sh) ψ(x) ψ(x)
108 JIAJIE CHEN AND THOMAS Y. HOU

For I32 , using the Hölder seminorm, the odd symmetry of K(s) = c s|s| 1 s2
4 in s1 , and |(W ψ)(x +
√
sh) − (W ψ)(x − sh)| ≤ 2s1 h, we get
1/2
h1/2 √ 2k h1/2 √
Z Z
|I32 | ≤ [W ψ]C 1/2 |K(s) 2s1 ds = 2 [W ψ]C 1/2 |K(s) 2s1 ds,
ψ(x) x
[0,k2 ]×[−k2 ,k2 ] ψ(x) x
[0,1]2

where we used the scaling symmetry of K and a change of variables s → k2 s in the last equality.

8.2.1. The commutator. For I31 , we apply the simple Taylor expansion to f = ψ −1
m20 s21 m02 s22
(8.35) |f (x + sh) − f (x)| ≤ |fx (x)hs1 + fy (x)hs2 | + h2 ( + m11 s1 s2 + ).
2 2
where mij is the bound for the second derivatives of ψ −1
mij (s) = max ||∂xi ∂yj (ψ −1 )||L∞ , I(1) = [0, k2 h], I(−1) = [−k2 h, 0].
Bi1 j1 (h)+I(sgn(s1 ))×I(sgn(s2 ))

Note that mij is constant in each quadrant of [−k2 , k2 ]. We plug in the expansion (8.35) to
estimate I31 . We only discuss a typical term m20 s21 h21
s2
Z
I31,02 , h2 |K(−s)(W ψ)(x + sh)|m20 (s) 1 ds.
[−k2 ,k2 ]2 2
If k2 ≥ 2, we can further partition [−k2 , k2 ]2 it into B2p,2q (1/2) = [p, p + 1/2] × [q, q +
1/2], −k2 ≤ p, q ≤ k2 − 1/2, where we use the notation (8.21). For each grid B2p,2q (1/2), the
sign of s and m20 (s) are fixed, and we have
s2 |K(s)|s21 ψ
Z Z
|K(−s)|(W ψ)(x + sh)m20 (s) 1 ds ≤ m20 ||W ϕ||∞ ( )(x + sh)ds.
B2p,2q ( 21 ) 2 B2p,2q ( 12 ) 2 ϕ
The last integral can be estimated using the method in (8.27). Combining the estimate of integral
in different regions Bp,q (1/2), we obtain the estimate of I31,02 . Similarly, we can estimate the
contributions of other terms in (8.35) to I31 .
For small k2 ≤ 2, we do not partition the domain. We denote D(k2 ) = Bi1 ,j1 (hx ) +
[−k2 h, k2 h]2 . For s ∈ [−k2 , k2 ], we use x + sh ⊂ D(k2 ) ⊂ D(2) to get
(8.36)
ψ
|f (x+sh)−f (x)| ≤ ||fx ||L∞ (D(k2 )) s1 h+||fy ||L∞ (D(k2 )) s2 h. |W ψ(x+sh)| ≤ ||W ϕ||∞ || ||L∞ (D(2)) .
ϕ
Plugging the above estimate into I31 , we get
Z
X ψ
I31 ≤ h||∂xi ∂yj (ψ −1 )||L∞ (D(k2 )) ||W ϕ||∞ || ||L∞ (D(2)) |K(s)si1 sj2 |ds.
ϕ [−k2 ,k2 ]2
(i,j)=(1,0),(0,1)

Using the scaling symmetry, we can reduce the last integral to k2i+j [−1,1]2 |K(s)si1 sj2 |ds.
R

We apply the above estimates to a list of k2 . Then by optimizing the k2 , we obtain the sharp
estimate of ux,A .
Remark 8.7. In (8.35), we do not bound f (x + sh) − f (x) directly using the estimate (8.36) since
s is large. Thus, we perform a higher order expansion.
8.2.2. Estimate of uy , vx . The estimates of uy , vx follow similar ideas and estimates. Since
vx (x, y) vanishes near y = 0, to exploit this smallness, we need more careful estimates when
(x, y) is close to the boundary. We refer the details to the upcoming supplementary material.

8.2.3. Estimate of uA . The estimate of uA is much simpler since it is more regular. Let K and
K̂ be the kernel of u, v and its approximation term, respectively. For f = u or v, we perform a
decomposition similar to (8.34)
(8.37) Z Z Z Z
fA = ( + + )K(x − y)W (y)dy − K̂(x, y)W (y)dy , I1 + I2 + I3 + I4 .
R(k)c R(k)\Rs (k) Rs (k)
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 109

The estimates of I1 + I4 follow the method for ux,A . For I2 , we use the method in Section 8.1.6.
For I3 , since K has a singularity of order |x|−1 , which is locally integrable, we use a change of
variable y = x + sh to obtain
Z
I3 = h K(−s)W (x + sh)ds.
[−k,k]2

Then we partition [−k, k]2 into small grids, and use the method in (8.32) to estimate the integral
in each grid. Here, we get a factor h in the change of variables since K(λs) = λ−1 K(s).

8.2.4. Choice of parameters. Recall the choice of several parameters a, h, hx from (8.11). We
choose 3 ≤ k ≤ 10. We choose k for the size of the singular region kh (8.34), (8.37) not so small
such that the error h2 ∂ 2 K in Lemma 8.2, which has the order h2 |x−y|−α−2 near the singularity,
is smaller than the main term K, which has the order |x − y|−α , α = 1, 2. Since we will estimate
I1 + I4 , I2 , I3 in the decomposition separately using the triangle inequality, we do not choose k
to be too large so that we can exploit the cancellation in I1 + I4 .

8.3. Hölder estimates. We want to estimate |f|x−z| (x)−f (z)|

1/2 for any x, z ∈ R++
2 with x1 = z1 or
x2 = z2 and some function f , e.g. f = ux,A . Without loss of generality, we assume |z| > |x|.
1/2 1/2
Then in the Cx estimate, we have x1 < z1 , x2 = z2 ; in the Cy estimate, we have x1 = z1 , x2 <
z2 . Applying the rescaling argument in Section 8.1, we can restrict x̂ = λx to x̂ ∈ [0, 2a]2 \[0, a]2 .
For this reason, we assume λ = 1 for simplicity. We will only estimate the Hölder difference for
comparable x, z: |x|  |z|. If |z| >> |x|, we simply apply L∞ estimate to f (x), f (z) and use the
triangle inequality.
We focus on the Hölder estimate of ux,A , which is representative and the most important
nonlocal term to estimate in our energy estimate. See (2.23), (2.39).

1/2
8.3.1. Cx estimate. Recall Ii from the decomposition (8.18) and K1 (s) = s|s|
1 s2
4 . We apply the

same decomposition to ux,A (z). We assume that the approximation term ûx (2.89) takes the
following form
Z Z
(8.38) ûx (x) = K̂1 (x, y)W (y)dy, I6 (x) , ψ(x)ûx (x) = ψ(x) K̂1 (x, y)W (y)dy,

with a nonsingular kernel K̂1 . We first discuss how to estimate the regular part I1 , I3 , I4 in
(8.18) and I6 , which are Lipschitz. We will apply Lemmas 3.1-3.5 to estimate the most singular
part I2 . The most technical part is to estimate I5 , which is log-Lipschitz since the kernel
K1 (x − y)(ψ(x) − ψ(y)) has a singularity of order −1. We assemble the estimates of different
parts to estimate [ux,A ]C 1/2 in Section 8.6.
x

8.3.2. Estimates of the regular terms I1 , I3 , I4 , I6 . Recall I1 , I3 , I4 from (8.18) and I6 from (8.38).
Since the integrands in I1 , I3 , I4 are supported at least k2 h away from the singularity x, if
W is in some suitable weighted L∞ space, we can show that I1 , I3 , I4 are Lipschitz and their
derivatives can be bounded by ||W ϕ||∞(R++ ) = ||ωϕ||∞ . In fact, I1 and I4 are piecewise smooth.
2
Their derivatives jump when R(x, k), R(x, k2 ) change, or equivalently, x moves from one grid to
another. For x ∈ Bi1 ,j1 (hx ) (8.20), these rectangle domains are the same, and these functions
are smooth. The approximation term I6 (8.38) is smooth in x. To exploit the cancellation, we
combine the estimates of I1 , I4 , I6 together. We symmetrize the kernel in I1 (x) + I4 (x) − I6 (x)
and use the method in Section 8.1.3 to estimate the derivatives of I1 (x) + I4 (x) − I6 (x). See
also (8.24), (8.25) for the form of the symmetrized integrands in these integrals.
We estimate both the L∞ and Lipschitz norm of I3 using the method in Sections 8.1.6, 8.1.7.
We will optimize two estimates to obtain a sharper Hölder norm of I3 .
We choose integer k, k2 in the decomposition (8.18) . Then in each grid [yi , yi+1 ] × [yj , yj+1 ],
the indicator functions in I1 + I4 − I6 , e.g. 1R(k)c , 1R(k)\R(k2 ) , are constant. See Remark 8.6.
110 JIAJIE CHEN AND THOMAS Y. HOU

1/2
8.3.3. Cx estimate of I2 . We first estimate the second term I2 in (8.18). Recall R(x, k), Rs (x, k)
from (8.13), (8.14) and the location of x (8.20). We have
x2 − (j − k)h ≤ (j + 1)h − (j − k)h = (k + 1)h, (j + 1 + kh) − x2 ≤ (j + 1 + kh) − jh = (k + 1)h.
Since x2 = z2 , using Lemma 3.1 with (a, b1 , b2 ) = (kh, x2 − (j − k)h, (j + 1 + k)h − x2 ) and
|b1 |, |b2 | ≤ (k + 1)h, we obtain
1 (k + 1)h (k + 1)h
|I2 (x, k) − I2 (z, k)| ≤ C1 ( )[W ψ]C 1/2 = C1 ( )[ωψ]C 1/2 .
|x − z|1/2 |x − z| x |x − z| x

For I2 (x, k) associated with other terms u, v, uy , vx , we can estimate it using similar ideas
1/2
and Lemmas 3.1-3.5. The Cy estimate of I2 (x, k) is completely similar. See Section 8.3.8 for
more details.
1/2
8.3.4. Cx estimate of I5 . For I5 (8.18), K1 (x − y)(ψ(x) − ψ(y)) is singular of order −1 near
y = x. Given W ∈ L∞ (ϕ), I5 is log-Lipschitz and thus in C 1/2 . There are several approaches
1/2
to estimate its Hölder norm. One approach is the following. We use part of the Cx norm of ω
to get a better estimate. We further decompose I5 as follows
Z
I5 (x, k2 ) = K1 (x − y)(ψ(x) − ψ(y))W (y)dy
R(k2 )\Rs,1 (k2 )
Z
+ K1 (x − y)(ψ(x) − ψ(y))W (y)dy , I5,1 (x, k2 ) + I5,2 (x, k2 ).
Rs,1 (k2 )

The estimate of I5,1 is similar to that of I3 . We estimate its L∞ and x-derivative using the
method in Sections 8.1.6, 8.1.7.
For I5,2 , we want to estimate it using a method similar to that of I2 . See the left figure in
Figure 12 for the domains of the integrals in I5,2 (x), I5,2 (z). The integrand satisfies
K1 (x − y)(ψ(x) − ψ(y))W (y) = ψ(x)K1 (x − y)(ψ −1 (y) − ψ −1 (x)(W ψ)(y)
≈ ψ(x)∂i (ψ −1 (x)) · K1 (x − y)(yi − xi )(W ψ)(y).
Thus, I5,2 (x) can be seen as a weighted version of I2 (8.18) with a weight ψ(x)∂i (ψ −1 (x)), a
more regular kernel K1 (x − y)(yi − xi ), and a smaller domain Rs (k2 ). Since the kernel is more
regular and the domain is smaller, our estimate for I5,2 is much smaller than that of I2 .
Now, we justify this approach. Using a change of variables y = x + s, s ∈ Rs,1 (k2 ) − x and
the above identity, we yield
Z
I5,2 (x, k2 ) = ψ(x) K1 (−s)(ψ −1 (x + s) − ψ −1 (x))(W ψ)(x + s)ds.
Rs,1 (k2 )−x
R1
Using Newton’s formula f (1) = f (0) + f 0 (0) + 0 (1 − t)f 0 (t)dt, we get
Z 1  
−1 −1 −1
ψ (x + s) − ψ (x) = s · ∇ψ (x) + (1 − t) s · (∇2 ψ −1 )(x + ts) · s dt
0
X X 2 Z 1
−1
= si ∂i (ψ )(x) + i 2−i
s1 s2 (1 − t)∂1i ∂22−i (ψ −1 )(x + ts)dt.
i=1,2
i 00≤i≤2

Denote
Z 1
Qij (x) = ψ(x) (1 − t)∂1i ∂2j (ψ −1 )(x + ts)dt, i + j = 2, D(x) = Rs,1 (x, k2 ) − x,
0
∂1i ∂2j ψ(x)
Z
Qij (x) = ψ(x) · ∂1i ∂2j (ψ −1 )(x) = − , i + j = 1, Pij (x) = K1 (−s)si1 sj2 (W ψ)(x + s)ds.
ψ(x) D(x)

Using the above expansion and notations, we get

X X 2
I5,2 (x, k2 ) = Pij Qij + Pij Qij .
i+j=1 i+j=2
i
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 111

Figure 12. Left: Rs,1 (x, k2 ) and Rs,1 (z, k2 ) with x2 = z2 . The small square
is a mesh grid containing x or z. x, z can have different locations relative to
the grids. Right: The large rectangle is R(k2 ), the upper part is R+ (k2 ), and
the lower part is R− (k2 ). The blue region is R− (k2 )\R− (k3 ). Γ is part of its
boundary.

Next, we use the above decomposition to estimate I5,2 (x, k2 ) − I5,2 (z, k2 ). The leading order
terms are Pij Qij with i + j = 1. We observe that if x2 = z2 , we have
D(x) = Rs,1 (x, k2 ) − x = Rs,1 (z, k2 ) − z = D(z).
Suppose that x1 < z1 . We perform a decomposition
|Pij (x)Qij (x) − Pij (z)Qij (z)| ≤ J1 + J2 ,
(8.39)
J1 , |Qij (z)(Pij (x) − Pij (z))|, J2 , |Pij (x)(Qij (x) − Qij (z)).
Using D(x) = D(z), we bound J1 as follows
Z
|J1 | ≤ |Qij (z)| K1 (−s)si1 sj2 ((W ψ)(x + s) − (W ψ)(z + s))ds
D(x)
Z
≤ |Qij (z)| · |x − z|1/2 ||ωψ||C 1/2 |K1 (s)si1 sj2 |ds.
x
s∈D(x)

The term Qij only depends on the weight and is smoother than Pij . We can estimate
Qij (x) − Qij (z) by bounding ∂1 Qij since Qij is locally smooth. For Pij in J2 , we use the
method in (8.32) to bound it by C||ωϕ||∞ with some constant C. Then we obtain the estimate
|J2 | ≤ C2 |x − z| · ||ωϕ||L∞
for some constant C2 . Note that the second order term Pij Qij , i + j = 2 is much smaller than
the leading order terms. For |x − z| not too small, we can estimate its contribution trivially
1 1
|Pij (x)Qij (x) − Pij (z)Qij (z)| ≤ (|Pij (x)Qij (x)| + |Pij (z)Qij (z)|).
|x − z|1/2 |x − z|1/2
In summary, to obtain the above estimates, we need to estimate piecewise bound for |Qij (x)|,
Pij (x), |∂k Qij (x)|, and the integrals D(x) |K1 (s)si1 sj2 |ds, i + j = 1, 2.
R
1/2
The above estimate of I5 (x, k2 ) can be generalized to the Cx estimate of u, v, vx , uy . Yet,
1/2
it does not apply to the Cy estimate of u, ∇u since it requires the estimate of (W ψ)(x + s) −
(W ψ)(z +s) for s in some rectangle R = D(x) = D(z). Howover, since W is discontinuous across
1/2
the boundary y = 0, W ψ ∈ / Cy (R) if x+s, z +s are not in the same half plane. If x1 < x2 , then
the rectangles R(x, k2 ), R(z, k2 ) will not intersect the boundary and the previous estimate holds
true. If x1 > x2 , we consider two modifications for different kernels in the following subsections.
1/2 1/2
8.3.5. Ideas of the Cy estimates of I5 . The main idea in the following Cy estimates is to use
a combination of the estimates for the log-Lipschitz function in Section 8.1.9 and the estimate
in Section 8.3.4. The latter provides better estimates, and we try to use this method as much
112 JIAJIE CHEN AND THOMAS Y. HOU

as possible. Following the ideas in Section 8.1.9, we decompose I5 (x) into the singular part and
nonsingular part with different size k3 of the singular region
I5 (x) = I5,S (x, k3 ) + I5,N S (x, k3 ).
Although we cannot apply the second method to the whole I5 (x), our observation is that
we can apply it to the integrals in the upper part of the regions, e.g. R+ (k2 ), R+ (k3 ) (8.15),
since these integrals only involve W ψ in R+
2 and we have W ψ ∈ C
1/2
. Thus, we will further
decompose some of the regions into the upper part and the lower part, and then apply the first
method to the lower part, and the second method to the upper part.
1/2
8.3.6. Cy estimate of the velocity with a kernel of the first type. The kernels
y1 y2 y2
(8.40) K= 4
,
|y| |y|2
associated with ux = −∂xy (−∆)−1 ω, u = −∂y (−∆)−1 ω vanish when y2 = 0. We call them the
first type kernel.
Let K be a kernel of the first type. We use the following decomposition
Z
I5 (x, k2 ) = K(x − y)(ψ(x) − ψ(y))W (y)dy
R+ (k2 )
Z
+ K(x − y)(ψ(x) − ψ(y))W (y)dy , I5+ (x, k2 ) + I5− (x, k2 )
R− (k2 )

See the right figure in Figure 12 for R± (k2 ). Since R+ (k2 )(x, k2 ) is in R+ 2 , we can apply the
same argument as that in Section 8.3.4 to obtain the desired estimates by restricting all the
derivations in R+ (x, k2 ), R+ (z, k2 ).
For the lower part I5− (x, k2 ), it is log-Lipschitz if W ∈ L∞ (ϕ). We cannot bound its derivative
using ||W ϕ||∞ . We face the difficulty discussed at the beginning of Section 8.
Alternatively, we follow the ideas in Section 8.1.9. The kernel K(x − y)(ψ(x) − ψ(y)) satisfies
the assumptions in Section 8.1.9, at least for x, y in some region. We decompose it into the
smooth part and rough part. We introduce 0 < k3 < k2 and consider the following decomposition
Z
I5− (x, k2 ) = K(x − y)(ψ(x) − ψ(y))W (y)dy
R− (k2 )\R− (k3 )
(8.41) Z
− −
+ K(x − y)(ψ(x) − ψ(y))W (y)dy , I5,1 (x, k2 ) + I5,2 (x, k2 ).
R− (k3 )

See the right figure in Figure 12 for an illustration of different domains. Recall that k2 ∈ Z+ .
−
We choose k3 = k2 − 2i , i = 0, 1, 2.., 2k2 − 4. Since the integrand in I5,1 supports at least k3 h
− −
away from the singularity, I5,1 (x, k2 ) is Lipschitz. We can estimate the derivative of I5,1 (x, k)
− −
in the y direction. The domain R (k2 )\R (k3 ) is not piecewise constant since the upper part
of its boundary, i.e.
Γ = {(y1 , x2 ) : y1 ∈ [(i − k2 )h, (i + 1 + k2 )h]\[(i − k3 )h, (i + 1 + k3 )h]},
−
depends on x2 . See Figure 12 for an illustration of Γ. Taking x2 derivative on I5,1 , we get
Z
−
|∂x2 I5,1 (x, k2 )| ≤ ∂x2 (K(x − y)(ψ(x) − ψ(y)))W (y)dy
R− (k2 )\R− (k3 )
Z
+ K(x − y)(ψ(x) − ψ(y)))W (y)dy1 .
y∈Γ

Since y ∈ Γ ⊂ {y : y2 = x2 } and that K(y1 , 0) ≡ 0, the second term vanishes. The first term
can be estimated using a change of variables y = x + s and the method in Section 8.1.6, 8.1.7,
since its support is at least k3 h away from the singularity.
−
For I5,2 , the kernel satisfies K(x − y)(ψ(x) − ψ(y)) ∼ |x − y|−1 for small |x − y| and is
locally integrable. We estimate its piecewise L∞ bound using the method in Section 8.2.1 for
the commutator.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 113

The above decomposition can be applied to estimate

− − − −
|I5− (x, k2 ) − I5− (z, k2 )| |I5,1 (x, k3 ) − I5,1 (z, k3 )| |I5,2 (x, k3 )| + |I5,2 (z, k3 )|
1/2
≤ min 1/2
+ 1/2
|x − z| i
k3 =k2 − 2 |x − z| |x − z|
h
for |x − z| not too small, e.g. |x − z| ≥ ds = 10 . When |x − z| is sufficiently small, the second
term in the above estimate can be very large.
According to the analysis in Section 8.1.9, for |x−z| very small, we need to choose k2 h ∼ |x−z|
to get the sharp estimate. Thus, we consider one more decomposition for a ≤ 1
Z
I5− (x, k2 ) = K(x − y)(ψ(x) − ψ(y))W (y)dy
R− (k2 )\Rs− (a)
(8.42) Z
− −
+ K(x − y)(ψ(x) − ψ(y))W (y)dy , I5,3 (x, a) + I5,4 (x, a).
Rs− (a)

The above decomposition is slightly different from (8.41). We choose Rs− (a) rather than
−
R (k3 ), since we need to choose the singular region with size going to 0 as |x − z| → 0. Yet,
R− (k3 ) (8.13) does not satisfy this requirement for k3 → 0. We can estimate the derivative of
− −
I5,3 (x, a) following Sections 8.1.6, 8.1.7, and the L∞ norm of I5,4 (x, a) following Section 8.2.1.
−
Again, in the computation of ∂x2 I5,3 (x, a), the boundary term vanishes due to K(y1 , 0) ≡ 0. In
summary, we can obtain the following estimate
− −
(8.43) |∂x2 I5,3 (x, a)| ≤ A(x) + B(x) log(1/a), |I5,4 (x, a)| ≤ C(x)ah,
for any a ≤ 1, where A(x), B(x) can be estimated following the method in Appendix D.6.1, and
the estimate of C(x) follows the method in Section 8.2.1. Using the above estimates and the
ideas in Section 8.1.9, we can estimate dy (I5− (·, k2 ), x, z) for small |x − z| by optimizing a, where
dy is defined below
(8.44) dy (f, x, z) = |f (x) − f (z)||x − z|−1/2 .
We will assemble these estimates in Section 8.6.
1/2 y 2 −y 2
8.3.7. Cy estimate of the velocity with a kernel of the second type. For the kernels K2 = 1|y|4 2
y1
and |y| 2 , they do not vanish on y2 = 0 in general. We call them the second type kernel. If we use
−
the strategies in the previous subsection, the boundary term in the computation of ∂x2 I5,1 (x, k3 )
−
or ∂x2 I5,3 (x, k3 ) does not vanish and can be large.
To avoid picking up a boundary term and apply the ideas in Section 8.3.5, we consider another
estimate on I5 (x, k2 ). For k3 = k2 − 2i , i = 0, 1, .., 2k2 −4, we perform the following decomposition
Z Z
I5 (x, k2 ) = K(x − y)(ψ(x) − ψ(y)W (y)dy + K(x − y)(ψ(x) − ψ(y))W (y)dy
R(k2 )\R(k3 ) R+ (k3 )
Z
+ K(x − y)(ψ(x) − ψ(y))W (y)dy , I5,1 + I5,2 + I5,3 .
R− (k3 )

Following the ideas in Section 8.1.9, we estimate the derivative of the regular part and then the
L∞ norm of the singular part. Indeed, we can estimate the y-derivative of I5,1 using the method
in Section 8.1.3, and the L∞ norm of I5,3 following Section 8.2.1. The estimate of I5,1 is similar
to that of I4 in Section 8.3.2. For I5,2 , since R+ (k3 ) is in R+
2 , we can obtain a better estimate
following the method in the estimate of I5,2 in Section 8.3.4.
After we estimate these quantities, we can estimate dy (I5 , x, z) (8.44) for |x − z| not too small
by optimizing k3 . To estimate dy (I5 , x, z) (8.44) for sufficiently small |x − z|, following (8.42),
we use the following decomposition
(8.45) Z Z
I5 (x, k2 ) = K(x − y)(ψ(x) − ψ(y)W (y)dy + K(x − y)(ψ(x) − ψ(y))W (y)dy
R(k2 )\Rs (a) Rs+ (a)
Z
+ K(x − y)(ψ(x) − ψ(y))W (y)dy , I5,4 + I5,5 + I5,6 .
Rs− (a)
114 JIAJIE CHEN AND THOMAS Y. HOU

Then we estimate the derivative of I5,4 and the L∞ norm of I5,6 as follows
(8.46) |∂x2 I5,4 | ≤ A(x) + B(x) log(1/a), |I5,6 | ≤ C(x)ah,
where the estimates of A, B are given in Appendix D.6.1, and the estimate of C follows the
method in Section 8.2.1. The Hölder estimate of I5,5 follows the method in the estimate of
I5,2 in Section 8.3.4. With these estimates, we can further bound dy (I5 , x, z) (see (3.5)) for
sufficiently small |x − z| by optimizing a. See Section 8.6.
Remark 8.8. We do not use the later decomposition on I5 , i.e. I5 = I5,4 + I5,5 + I5,6 , to estimate
dy (f, x, z) when |x − z| is not too small since the domain of the integral in I5,4 is not piecewise
constant. As a result, we need to bound the boundary term in the computation of ∂x2 I5,4 . The
resulting estimate is worse than the estimate using the decomposition I5 = I5,1 + I5,2 + I5,3 .
1/2
We do not apply the above computation with smaller window [−ah, ah]2 in the Cx estimate,
since it leads to a worse estimate. See also the discussions in Section 8.3.5.
8.3.8. Hölder estimate of u, v, uy , vx . The ideas of the Hölder estimate for other terms are sim-
ilar. For a kernel K associated with u, ∇u, we perform another decomposition similar to (8.18)
Z Z 
ψ(x) K(x − y)W (y)dy = ψ(x)1R(k)c + 1Rs (k) ψ(y) + 1R(k)\Rs (k) ψ(y)

(8.47) + 1R(k)\R(k2 ) (ψ(x) − ψ(y)) + 1R(k2 ) (ψ(x) − ψ(y)) K(x − y)W (y)dy
, I1 (x, k) + I2 (x, k) + I3 (x, k) + I4 (x, k, k2 ) + I5 (x, k2 ),
Here, we use Rs (x, k) (8.14), which is symmetric with respect to both x1 and x2 , rather than
Rs,1 (x, k). Denote by If 6 (x, k2 ) the approximation term for f = ux , uy , vx , u, v. It takes the
form similar to (8.38).
We consider two cases of x̂ ∈ [0, 2xc ]2 \[0, xc ]2 . In the first case, we consider x̂ ∈ [xc , 2xc ] ×
[0, 2xc ] , DX1 , where we have x̂1 ≥ cx̂2 for some constant c > 0. In the second case, we consider
x̂ ∈ [0, xc ] × [xc , 2xc ] , DX2 , where we have x̂1 ≤ cx̂2 . We distinguish these two cases since in
the second case, the singular region does not touch the boundary, we can apply the method in
Section 8.3.4.
1/2 1/2
Cx estimate of uy , vx . In the Cx estimate of uy , vx , we follow Section 8.3.2 to estimate the
regular part I1 + I4 − I6 and I3 . We follow Section 8.3.3 and use Lemma 3.4 to estimate I2 . For
I5 , we follow Section 8.3.4.
1/2 1/2
Cy estimate of ux . In the Cy estimate of ux , we perform the decomposition (8.47) rather
than (8.18). The estimates of I1 + I4 − I6 , I3 follow Section 8.3.2. For I2 , we use Lemma 3.3.
We follow Section 8.3.6 to estimate I5 if x̂ ∈ DX1 , and Section 8.3.4 if x̂ ∈ DX2 .
We remark that we use the decomposition (8.47) rather than (8.18) since in Lemma 3.3, we
need to assume that the singular region around x is symmetric in both x1 and x2 .
1/2 1/2
Cx and Cy estimate of u, v. The Hölder estimates of u, v are substantially easier since
1/2 1/2
u, v are more regular. We perform Cx , Cy of ρuA for another weight ρ. We decompose the
integral as follows
(8.48)
Z Z  
ρ(x) K(x − y)W (y)dy = 1R(k)c ρ(x) + 1R(k) ρ(x) K(x − y)W (y)dy , I1 (x, k) + I2 (x, k).

We follow Section 8.3.2 to estimate I1 − I6 . For I2 , we follow the ideas in Sections 8.1.9, 8.3.6,
8.3.7 to estimate the log-Lipschitz function. We choose a list of k2 and associated region S(k2 )
and decompose I2 as follows
Z Z
I2 (x, k) , ρ(x)K(x−y)W (y)dy+ ρ(x)K(x−y)W (y)dy , I21 (x, k2 )+I22 (x, k2 ).
R(k)\S(k2 ) S(k2 )

For large k2 , we choose k2 = k, k − 1/2, .., 2 with S(k2 ) = R(k2 ). For k2 < 2, we choose
S(k2 ) = Rs (k2 ). For I21 (x, k2 ), we estimate its derivatives following the estimate of I51 when
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 115

k2 ≥ 2, and the estimate of I54 when k2 < 2 in Section 8.3.7. For I22 (x, k2 ), we estimate its L∞
norm following the estimate of I53 when k2 ≥ 2, and the estimate of I56 when k2 < 2 in Section
8.3.7. The estimate is simpler since the above kernel is much simpler than K(x−y)(ψ(x)−ψ(y))
in Section 8.3.7.
1/2
8.3.9. Special case: Cy estimate of uy , vx . In this case, we want to apply Lemma 3.5 to
estimate the most singular part. Since in Lemma 3.5, we do not localize the integral, we
perform the following decomposition
(8.49)Z Z  
ψ(x) K(x − y)W (y)dy = ψ(y) + 1R(k2 )c (ψ(x) − ψ(y)) + 1R(k2 ) (ψ(x) − ψ(y)) K(x − y)W (y)dy

, I1 (x, k) + I2 (x, k) + I3 (x, k).

For I1 , we apply Lemma 3.5. We follow Section 8.3.7 to estimate I5 if x̂ ∈ DX1 , and Section
8.3.4 if x̂ ∈ DX2 .
We follow Section 8.3.2 to estimate I2 − I6 . There are additional difficulties since the weight
ψ(y) and the symmetrized integrand J = K(x, y)(ψ(x) − ψ(y)) (see similar derivations in
(8.24),(8.25)) are singular near 0.
Estimate the integral near 0. To estimate the D1 = ∂x2 derivative, we use
|D1 I| = |D1 K(ψ(x) − ψ(y)) + K · D1 ψ(x)| ≤ |D1 K · ψ(x) + K · D1 ψ(x)| + |D1 K · ψ(y)|.
Note that for y close to 0, ψ(y) is much larger than ψ(x) and K(x, y) is not singular. The main
term in D1 I is given by D1 Kψ(y), and we expect that the above estimate is sharp. It follows
Z Z Z
 ψ 
|D1 I·W (y)|dy ≤ ||W ϕ||∞ ||ϕ−1 ||L∞ (Q) |D1 Kψ(x)+K·D1 ψ(x)|dy+|| ||L∞ (Q) |D1 K|dy ,
Q Q ϕ Q

where Q is some grid near the origin. The integrands in both integrals do not involve the singular
weight, and we can compute them using the previous methods.
Estimate in the far-field. For the tail part in this case, we have improvement if λ → 0 due
to the approximation term near 0 (2.84)
fˆ = Cf 0 (x, y)ux (0) + Cf (x, y)K00 = Cf (x, y)K00 ,
where f = uy , vx and we have used Cf 0 (x, y) = 0 (2.80). Its associated integrand is given by
Kapp , π −1 Cf (x, y)K00 (y),
where K00 is defined in (2.79). To estimate it, we use the following decomposition
D1 (J − ψ(x)Kapp ) = D1 ((K − Kapp ) · ψ(x)) − D1 K · ψ(y) , P1 + P2 .
The first part P1 is estimated using the method in Section 8.4. Due to the approximation,
(K − Kapp ) has a much faster decay for large y beyond [0, D]2 . For P2 , we have
Z Z
|P2 ||W (y)|dy ≤ ||W ϕ||∞ ||ψ(y)||L∞ (Ωc ) |D1 K|ϕ−1 (y)dy,
Ωc Ωc
2
where Ω = [0, D] with large D. The last integral is computed using the method in Section 8.4.
8.4. Estimate the integrals near 0 and in the far field. In the computation of the L1
type integral, we use a combination of uniform mesh and adaptive mesh to compute the integral
in a finite domain [0, D]2 , e.g. D = 1000. See Section 8.1.3. Since the kernel decays and the
singularity is in the near-field, the integral beyond this domain is small, and we estimate it
directly . In addition, for y near 0, we estimate the integrals related to |y|−2 or |y|−4 directly,
which come from the approximation terms ux (0), K00 (8.5).
For simplicity,
R we consider λ = 1. The estimates can be generalized to other scales λ.
To estimate R D k(y)ω(y)dy for D near 0 or D in the far-field, following (8.7), we only need
to estimate D |k(y)|ϕ−1 (y)dy. Since |y| is either very small or very large, we can use the
asymptotics of ϕ in these estimates.
116 JIAJIE CHEN AND THOMAS Y. HOU

2 2
y1 y2 y1 y2 (y1 −y2 )
8.4.1. Near-field estimate. Firstly, we estimate [0,R1 ]2 |k(y)|ϕ−1 (y)dy for k(y) =
R
|y|4 , |y|8
related to ux (0), K00 (2.79). We partition [0, R1 ] into
0 = z0 < z1 < ... < zn = R1
with z1 much smaller than R1 . Denote Qij = [zi−1 , zi ] × [zj−1 , zj ]. Clearly, we have
Z X Z
−1
|k(y)|ϕ (y)dy ≤ Iij , Iij , |k(y)|ϕ−1 (y)dy.
[0,R1 ]2 1≤i,j≤n Qij

For Iij , (i, j) 6= (1, 1), we apply a trivial bound

(8.50) Iij ≤ |Qij | · ||k||L∞ (Qij ) ||ϕ−1 ||L∞ (Qij ) .
y y (y 2 −y 2 )
For k(y) = y|y|
1 y2
4 ,
1 2 1
|y|8
2
, the estimate of ||k||L∞ (Qij ) is established in Appendix D. It
remains to estimate the first term I11 . Denote r = y1 . Suppose that
ϕ(x) ≥ q|x|a (cos β)b , b ≤ 0,
y1 y2
If k(y) = |y|4 , we yield
√ Z √
Z 2r π/2 Z 2r Z π/2
−1 sin β cos β −a
I11 ≤ q r (cos β)−β rdrdβ = q −1 r −a−1
dr sin β(cos β)−b+1 dβ
0 0 r2 0 0
√ −a Z 1 √ −a
−1 ( 2r) −b+1 −1 ( 2r) 1
=q t dt = q .
−a 0 −a 2 − b
y1 y2 (y12 −y22 )
If k(y) = |y|8 , we yield |k(y)| ≤ 14 sinr44β . Since b ≤ 0, we get ϕ ≥ qra and
Z √2r Z π/2 Z √2r
1 | sin 4β| −a 1 2π
Z
−1 1 −a−3
I11 ≤ q s sdsdβ = s ds | sin β|dβ
0 0 4 s4 4q 0 4 0
√ −a−2 Z π/2 √ −a−2
1 ( 2r) 1 ( 2r)
= sin βdβ = .
4q −2 − a 0 4q −2 − a

8.4.2. Far-field estimate. Denote a ∨ b = max(a, b). To estimate the far field integral I ,
|k(y)|ϕ−1 (y)dy, we first pick sufficient large R, and then partition the domain
R
y1 ∨y2 ≥R0

0 = z0 < z1 < .. < zm = R0 < zm+1 < ... < zn = R1 < +∞.
Denote Qij = [zi−1 , zi ] × [zj−1 , zj ]. Clearly, we have
X Z Z
I= Iij + J, Iij , |k(y)|ϕ−1 (y)dy, J= |k(y)|ϕ−1 (y)dy.
m+1≤max(i,j)≤n Qij y1 ∨y2 ≥R1

For Iij , we apply the trivial estimate (8.50). Suppose that

ϕ ≥ qra (cos β)b , |k(y)| ≤ |y|−p , β ≤ 0.
We get
∞ π/2
1 R1−p−a+2 π/2
Z Z Z
1
J≤ r−p−a (cos β)−b rdrdβ = (cos β)−b dβ.
q R1 0 q |p + a − 2| 0
Using Holder’s inequality, we get
Z π/2 Z π/2 Z π/2
−b −b
(cos β) dβ ≤ ( cos βdβ) ( 1)1+b = (π/2)1+b .
0 0 0

It follows
1 R1−p−a+2
J≤ (π/2)1+b .
q |p + a − 2|
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 117

Application. We apply the above calculations to estimate the integral and its derivatives be-
yond the mesh [0, D]2 (8.9). Since the domain is far away from the singularity, the integrand is
the symmetrized kernel (3.2), e.g., (8.25). From (D.16), (D.17), and Lemma D.2 in Appendix
D, for uA , ∇uA , ∂i (ρuA ), ∂i (ψ∇uA ), the integrand in the far-field (y is large) satisfies
|K(x, y)| ≤ C(x)Den−k
with some k ≥ 2 and coefficients C(x), where Den is defined in (D.15).
In our computation, we rescale x and restrict it to the near-field [0, b]2 with b < 2. Note that
y∈/ [0, D]2 and D >> b From (D.15), we get
|x| 2
Den ≥ max |y − z|2 ≥ max (|y| − |z|)2 ≥ (|y| − |x|)2 = |y|2 (1 − )
|z1 |≤x1 ,|z2 |≤x2 |z1 |≤x1 ,|z2 |≤x2 |y|
|x| √
Since |y| ≤ 2b/D, we yield
√
Den ≥ (1 − Cs )2 |y|2 , Cs = 2b/D.
It follows
Z Z
−1 −2k
|K(x, y)|ϕ (y)dy ≤ (1 − Cs ) C(x) |y|−2k ϕ−1 (y)dy.
y ∈[0,D]
/ 2 y ∈[0,D]
/ 2

Using the method in Section 8.4.2, we can estimate the above integral.
8.5. Estimate for very small or large x. The rescaling argument and the methods in previous
sections apply to the estimate of uA (x), ∇uA (x) for x ∈ [0, xM ]2 \[0, xm ]2 , 0 < xm < xM . For
very small or large x, we cannot use a finite number of dyadic scales λ = 2i to rescale x such that
x/λ ∈ [0, 2xc ]2 \[0, xx ]2 . Instead, we choose λ = max(x
xc
1 ,x2 )
. We want to estimate the rescaled
integral with a −d-homogeneous kernel K
Z Z
p(x) K(x − y)W (y)dy = pλ (x) K(x̂ − ŷ)λ2−d Wλ (ŷ)dy,

uniformly for all small λ << 1 or large λ >> 1, where p is some weight and pλ is defined in
(8.1). The rescaled singularity x̂ = x/λ satisfies max x̂i = xc . We simplify x̂, ŷ as x, y.
Our observation is that in the rescaled integral with scaling λ, we can use the asymptotic of
the weights in our computation, see e.g. (8.4). The new difficulty is that the estimate involves
the rescaled weight pλ (y). Since λ is not fixed and depends on x that tends to 0 or ∞, we cannot
evaluate pλ (y) and the integrand directly. In the following derivation, λ is comparable to |x|,
which is either very small or very large.
For y away from the singular region, the integrand of the regular part is given by J =
K(x, y) · pλ (x) (8.25). Here we use p for the weight. We use the following decomposition to
compute D1 J with D1 = ∂xi
|D1 J| = |D1 (K(x, y) · pλ (x))| = |D1 K(x, y) · pλ (x) + K(x, y) · D1 pλ (x)|
n D1 pλ (x) o
= pλ (x) D1 K(x, y) + Rlim K(x, y) + ( − Rlim )K(x, y) ,
pλ (x)
where Rlim is given by
D1 pλ (x)
Rlim , lim ,
x→A pλ(x)
D1 pλ (x)
with A = 0 or ∞. Since we consider very small λ or very large λ, the error term pλ (x) − Rlim
is small. Hence, we use a triangle inequality to bound D1 J
D1 pλ (x)
|D1 J| ≤ pλ (x) D1 K(x, y) + Rlim K(x, y) + pλ (x) ( − Rlim )K(x, y) .
pλ (x)
The advantage of the above decomposition is that the main term D1 K(x, y) + Rlim K(x, y) does
not depend on λ so that we can estimate it using previous methods.
Since the estimate of derivative of u, v does not involve the commutator, see, e.g. (8.48), we
can apply the above method to compute the integral of D1 u for small x or large x.
118 JIAJIE CHEN AND THOMAS Y. HOU

For y near the singular region, from (8.24), the symmetrized integrand is given by
J = K C (pλ (x) − pλ (y)) + K N C pλ (x),
where we use p for the weight. Firstly, we have
|D1 J| = |D1 K C (pλ (x) − pλ (y)) + D1 K N C pλ (x) + (K C + K N C )D1 pλ (x)|
Denote K = K C + K N C . We use the following method to bound D1 J
pλ (y) D1 pλ
|D1 J| ≤ pλ (x) D1 K C · (1 − ) + D1 K N C + K ·
pλ (x) pλ
n p lim (y) D1 plim
≤ pλ (x) D1 K C · (1 − ) + D1 K N C + K ·
plim (x) plim
pλ (y) p lim (y) D p
1 lim D 1 pλ
o
+ D1 K C ( − ) +K − .
pλ (x) plim (x) plim pλ
The second and the third term on the right hand side can be seen as an error term. The main
term D1 K C · (1 − pplim
lim (y)
(x) ) + D1 K
NC
+ K · Dp1lim
plim
does not depend on λ, and the singularity
x is in the near-field and away from 0. We can apply all the delicate decompositions developed
in previous sections to estimate D1 J.
In the Hölder estimates, we need various bounds for the weights pλ . Using the asymptotics
of p(x), we can estimate the derivatives of pλ for very small λ or very large λ uniformly. See
Appendix C.3. Once we obtain the estimates of ψλ , and the weight ϕλ in the L∞ norm ||ωλ ϕλ ||∞ ,
we can use the methods in previous sections and the scaling relations in Section 8.1.2 to perform
the Hölder estimates.
The L∞ estimate follows similar ideas and is much easier. We refer more details to the
upcoming supplementary material.
8.6. Assemble the Hölder estimates. In Section 8.3, we decompose the velocity in several
parts and estimate them separately. In this section, we assemble these estimates and estimate
|f (x) − f (z)|
δ(f, x, z) , ,
|x − z|1/2
for f = ρuA , ψ∇uA with suitable weights. One way to combine the estimates of different terms
is to use the triangle inequality. To obtain better estimates, we combine some of the estimates.
1/2
To illustrate the ideas, we focus on the Cx estimate, x ∈ [xc , 2xc ] × [0, 2xc ], i.e. x1 is large
relative to x2 , z1 ≥ x1 , and x2 = z2 . For general pairs (x, z), we can rescale (x, z) to (λx, λz)
such that λx ∈ [0, 2xc ]2 \[0, xc ]2 . Using the scaling relations in (8.1.2), we can estimate the
rescaled version of δ(f, x, z). See also the discussion at the beginning of Section 8.3.
We assume that z1 ∈ [xc , 2(1 + ν)xc ] with ν < 1. For z1 ≥ 2(1 + ν)xc , we have z1 > (1 + ν)x1 .
Since z1 , x1 are large relative to z2 , x2 , respectively, we have
|x − z| = |z1 − x1 |  |z1 | & |x|, |z|.
∞
Then, we can use the L estimate and triangle inequality to estimate δ(f, x, z). Note that we
can estimate the piecewise L∞ norm of |x|−1/2 ρ(x)uA (x) and |x|−1/2 ψ∇uA following Section
8.2, where ρ, ψ are the weights in the Hölder estimate of ρuA , ψ∇uA .
We focus on f = ψux,A . We partition the domain Dν = [xc , 2(1 + ν)xc ] × [0, 2xc ] into hx × hx
grids Dij , 1 ≤ i ≤ 2(1 + ν)xc /hx , 1 ≤ j ≤ 2xc /hx . We apply the decomposition (8.48) with the
same parameters k, k2 to x in different grids Dij . For x ∈ Dij , using the method in Section 8.3,
we obtain the estimate
f (x) = I1 (x) + I2 (x) + I3 (x) + I4 (x) + I5 (x) − I6 (x), I5 = I5,1 + I5,2
(8.51) |∂x (I1 + I4 − I6 )| ≤ aij,1 ||ωϕ||∞ , |∂x I3 | ≤ aij,2 ||ωϕ||∞ , |I3 | ≤ bij,2 ||ωϕ||∞ ,
|∂x I5,1 | ≤ aij,3 ||ωϕ||∞ , |I5,1 | ≤ bij,3 ||ωϕ||∞ ,
for some constants aij,l , bij ≥ 0, where I5,1 , I5,2 are defined and estimated in Section 8.3.4.
For x, z ∈ Dν with x2 = z2 , z1 ≤ z1 , we have x ∈ Di1 ,j , z ∈ Di2 ,j for some i1 ≤ i2 . We apply
the method in Section 8.3.3 to estimate δ(I2 , x, z) and the method in Section 8.3.4 to estimate
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 119

J1 related to δ(I52 , x, z) (8.39). These estimates contribute to the bound C[ωψ]C 1/2 for some
x
C > 0, which can be computed.
By averaging the piecewise derivative bounds and using the estimates in Appendix D.6.2, for
x ∈ Di1 ,j , z ∈ Di2 ,j , we can obtain
|(I1 + I4 − I6 )(x) − (I1 + I4 − I6 )(z)| ≤ C|x1 − z1 | · ||ωϕ||∞
for constant C depending only on {akl,1 }k,l≥1 and the mesh hx explicitly. From the above
discussions, for the remaining terms in f not estimated using the seminorm [ωψ]C 1/2 , e.g. I1 +
x
I3 + I4 + I5,1 − I6 and J2 related to I5,2 (8.39), they satisfy
X
fR (x) = fl (x), |fl (x) − fl (z)| ≤ min(pl |x1 − z1 |, ql ) · ||ωϕ||∞
1≤l≤N

for some N, where we can choose ql = ∞ if we do not have L∞ estimate for fl (x). Similar
consideration applies to pl . In our problem, there are only a few terms and N < 10. Now, for
x ∈ Di1 ,j , z ∈ Di2 ,j , we have
|fR (x) − fR (z) X
1/2
≤ min(pl δ 1/2 , ql δ −1/2 )||ωϕ||∞ ,
(8.52) |z1 − x1 | 1≤l≤N
δ = z1 − x1 ∈ [max(i2 − i1 − 1, 0)hx , (i2 − i1 + 1)hx ],
The upper bound can be obtained explicitly by partitioning the range of z1 − x1 into finite
many subintervals Ml according to the threshold δl = ql /pl . In each subinterval Ml , the bound
reduces to
P δ 1/2 + Qδ −1/2
for some constants P, Q. It is convex in δ 1/2 and can be optimized easily and explicitly in any
interval [δl , δu ], δl > 0 .
Remark 8.9. We combine the estimates of different parts in (8.51) using (8.52) to obtain a sharp
estimate. If one estimate different parts separately, the distance δ = z1 − x1 for the optimizer
may not be achieved for the same value, which leads to an overestimate. We remark that for
small distance |z1 − x1 |, such an overestimate can be significant since the ratio between the
endpoints |i2 − i1 + 1|/ max(i2 − i1 − 1, 0) varies a lot.
1/2
In some estimates, e.g. the Cy estimate of ux in Section 8.3.6, we need to decompose
I5 using different size of small singular region k3 . In such a case, we have a list of estimates
associated to different k3 for the part fR not estimated by [ωψ]C 1/2 or [ωψ]C 1/2 :
x y

|fR (x) − fR (z)| X

1/2
≤ min(pl,k3 δ 1/2 , ql,k3 δ −1/2 )||ωϕ||∞ .
|z1 − x1 | 1≤l≤N

1
For |x1 − z1 | bounded away from 0, e.g. |x1 − z1 | ≥ 10 hx , we can still partition the range of
|x1 − z1 | and optimizing the above estimates first over δ and then k3 .

1/2
8.6.1. Hölder estimate for small distance. In some Hölder estimates, e.g. the Cy estimate in
Sections 8.3.6, 8.3.7, when |x − z| is very small, e.g. |x − z| ≤ chx with c < 1, we need to choose
a singular region with size a to be arbitrary small. See also Section 8.1.9 for the estimates of a
log-Lipschitz function. In these estimates, we can decompose fR (x) that is not estimated using
the Hölder norm of ωψ as follows
fR (x) = f1 (x, a, b) + f2 (x, a),
for a < b and b is fixed. We can estimate the derivative of f1 , and the L∞ norm for f2
b Ci a
|∂x f1 (x, a, b)| ≤ (Ai + Bi log )||ωϕ||∞ , |f2 | ≤ ||ωϕ||∞
a 2
120 JIAJIE CHEN AND THOMAS Y. HOU

in each grid Dij for any a ≤ b, see e.g., (8.43) and (8.51). We drop j since we consider x, z with
x2 = z2 . For t = |x − z| ≤ hx , we get
|f (x) − f (z)| b √ Ca
(8.53) 1/2
≤ (A + B log ) t + √ , F (a, t)
|x − z| a t
where A = max(Ai , Ai+1 ), B = max(Bi , Bi+1 ), C = max(Ci , Ci+1 ). For each t ≤ chx , we can
optimize the above estimate over a ≤ b explicitly. We refer the derivations to Appendix D.6.3.

Appendix A. Some Lemmas for stability estimates

Lemma A.1. Suppose that Fi (t) ≥ 0, i = 1, 2, .., n, is Lipschitz in t and satisfies the following
estimates
d X
Fi (t) ≤ −aii Fi (t) + aij Fj (t), aii > 0.
dt
j6=i

If there exists µi > 0 such that

X
(A.1) aii µ−1
i − |aij |µ−1 −1
j ≥ λµi
j6=i

for some λ > 0. Then for E(t) , maxi (µi Fi (t)), which is Lipschitz, we have
d
E(t) ≤ −λE(t),
dt
for t almost everywhere. We can do integration to estimate E(t) since E(t) is Lipschitz. In
particular, when n = 2, condition (A.1) is equivalent to
(A.2) a11 a22 − |a12 a21 | > 0,
q
a22 |a12 |
and we can choose µ1 = 1, µ2 = a11 |a21 | .

Proof. A direct calculation yields

d X X
µi Fi (t) ≤ −aii (µi Fi ) + µi aij Fj (t) ≤ −aii µi Fi + µi |aij |µ−1
j E
(A.3) dt
j6=i j6=i

≤ −aii µi Fi + µi (aii µ−1

i − λµ−1
i )E = −λµi Fi + (aii − λ)(E − µi Fi ).
Let Ω be the domain where Fi , E are differentiable. Since Fi , E are Lipschitz, Ω is almost
everywhere. For t ∈ Ω, we assume E(t) = µj Fj (t) for some j. Then for any s > 0, we have
E(t) − E(t − s) = µj Fj (t) − max µi Fi (t − s) ≤ µj (Fj (t) − Fj (t − s)).
i

Dividing s on both sides and then taking s → 0, we get

d d
E(t) ≤ µj Fj (t) ≤ −λµj Fj + (ajj − λµj )(E − µj Fj ) = −λE(t),
dt dt
where we have used (A.3) and E(t) = µj Fj (t) in the second inequality. The desired estimate
follows. When n = 2, obtainingq a11 a22 − |a12 a21 | > 0 from (A.1) is trivial. For the other
|a12 |
direction, we choose µ1 = 1, µ2 = aa22
11 |a21 |
to obtain (A.1) for some λ > 0. 
We can generalize the above ODE Lemma to the L∞ estimates.
Lemma A.2. Suppose that fi (x, z, t) : R2++ × R2++ × [0, T ] → R, 1 ≤ i ≤ n, satisfies
(A.4) ∂t fi + vi (x, z) · ∇x,z fi = −aii (x, z, t)fi + Bi (x, z, t),
where vi (x, z, t) are some vector fields Lipschitz in x, z with vi |x1 =0 = 0, vi |z1 =0 = 0, and Bi
satisfies the following estimate
X
(A.5) |Bi (x, z, t)| ≤ |aij (x, z, t)| · ||fj ||L∞ .
j6=i
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 121

If there exists some constants M, λ, µi > 0 such that for all (x, z), we have
X X
(A.6) aii (x, z, t) − |aij |µi µ−1
j ≥ λ, µi µ−1
j |aij | ≤ M.
j6=i j6=i

Then for E(t) = maxi (µi ||fi (t)||∞ ), which is Lipschitz, and 0 ≤ t0 < t ≤ T , we have
E(t) ≤ e−λ(t−t0 ) E0 , E0 = E(t0 ).
The condition (A.5) means that the damping term is stronger than the bad terms, which
further leads to the stability. We apply f (x, z, t) = ((Si ψi )(x)−(Si ψi )(z))gi (x, z) in the weighted
Hölder estimate, and fi (x, z, t) = (Si ϕi )(x) in the weighted L∞ estimate, S1 = ω, S2 = η, S3 = ξ.
In the weighted L∞ estimate, we do not need the extra variable z and fi is constant in z. For
the Boussinesq equations (4.1), we choose
b(x, t) = c̄l x + ū(x) + u(ω)(x, t), vi (x, t) = b(x, t), or vi (x, z, t) = (b(x, t), b(z, t)).
We will also perform energy estimates on some scalars ai (t) and choose fi (x, z, t) = ai (t) in
the above Lemma. In this case, advection term is 0, and aii , aij , Bi only depend on t.

Proof. For simplicity, we assume that the condition (A.6) holds for µi = 1. Otherwise, we can
estimate the variables µi fi and introduce ãij = aij µi µ−1
j . Then the equations and estimates
(A.4), (A.5) become
∂i µi fi + vi (x, z) · ∇x,z (µi fi ) = −aii (µi fi ) + µi Bi (x, z, t)
X X
µi |Bi (x, z, t)| ≤ µi µ−1
j aij (x, z, t) · µj ||fj ||L∞ = ãij (x, z, t) · µj ||fj ||L∞ .
j6=i j6=i

In this case, then the condition (A.6) for aij becomes the condition for ãij with equal weights.
Thus, it suffices to consider the case µi = 1, i = 1, 2, .., n.
Formally, we can apply the L∞ estimate on both sides of (A.4) and then evaluate (A.4) at
the maximizer to obtain the desired result. To justify it rigorously, we use the characteristics,
Duhamel’s principle, and a bootstrap argument. We define the characteristics associated with
vi
d
(A.7) (Xi (t), Zi (t)) = vi (Xi , Zi , t), Xi (0) = x0 , Zi (0) = z0 , Fi (t) = fi (Xi (t), Zi (t), t)
dt
To simplify the notation, we drop x0 , z0 . Denote
X
(A.8) Ai (t) = aii (Xi (t), Zi (t)), Ci (t) , |aij (Xi (t), Zi (t), t)|.
j6=i

It suffices to prove that for small ε > 0, we have

2
(A.9) E(t) ≤ (1 + c(M )ε)e−λε (t−t0 ) E0 , λε = λ − ε, , c(M ) =
M
where M is the upper bound in (A.6). Then taking ε → 0 completes the proof.
We want to use a bootstrap argument to prove (A.9). Firstly, since E(0) = E0 and E(t) is
Lipschitz, the above condition holds for t ∈ [t0 , t0 + T1 ] with some T1 > 0. Now, we want to
show that under (A.9), we can obtain
(A.10) E(t) ≤ (1 + c(M )ε/2)e−λε (t−t0 ) E0 .
By definition, along the characteristics, we get
d
Fi (t) = −Ai (t)Fi (t) + Bi (Xi (t), Zi (t), t),
dt
Using the estimates (A.5) and definition (A.8), we yield
X
|Bi (Xi (t), Zi (t), t)| ≤ |aij |E(t) ≤ Ci (t)E(t).
j6=i
122 JIAJIE CHEN AND THOMAS Y. HOU

Using the Duhamel principle and the above estimate, we obtain

Rt Z t Rt
−A (s)ds
Fi (t) = e t0 i Fi (t0 ) + Bi (Xi (s), Zi (s), s)e s −Ai (τ )dτ ds
t0
(A.11) Rt Z t Rt
−A (s)ds
|Fi (t)| ≤ e t0 i Fi (t0 ) + Ci (s)E(s)e s −Ai (τ )dτ ds , I + II.
t0

For the second term, using the bootstrap assumptions (A.9), we yield
Z t Rt
|II| ≤ Ci (s)(1 + c(M )ε)e−λε (s−t0 ) E0 e−λε (t−s)− s (Ai (s)−λε )ds
t0

Using Ci (s) ≤ M, Ci (s) ≤ Ai (s) − λ (A.6) and the definition of c(M ) (A.9), we get
Ci (s) M 1 + ε/M 1 + c(M )ε/2
≤ ≤ = , Ci (s) + ε ≤ Ai (s) − λ + ε = Ai (s) − λε ,
Ci (s) + ε M +ε 1 + 2ε/M 1 + c(M )ε
which implies
Ci (s)(1 + c(M )ε) ≤ (Ci (s) + ε)(1 + c(M )ε/2) ≤ (Ai (s) − λε )(1 + c(M )ε/2).
Note that we choose c(M ) in (A.9) small enough such that the above inequality holds. Hence,
we can simplify the bound of II as follows
Z t Rt
|II| ≤ (1 + c(M )ε/2)e−λε (t−t0 ) E0 (Ai (s) − λε )e− s (Ai (τ )−λε )dτ ds
t0
Rt
−λε (t−t0 ) − (Ai (τ )−λε )dτ
= (1 + c(M )ε/2)e E0 (1 − e t0
).
The estimate of I is trivial. Since |Fi (0)| ≤ E0 , we have
Rt Rt
− (Ai (τ )−λε )dτ − (Ai (τ )−λε )dτ
|I| = |Fi (t0 )|e−λε (t−t0 ) e t0
≤ E0 e−λε (t−t0 ) e t0
,
which along with the estimate of II yields
|Fi (t)| ≤ |I| + |II| ≤ (1 + c(M )ε/2)e−λε (t−t0 ) E0 .
Since the above estimate holds for any initial data x0 , z0 and i, taking the supremum, we
prove (A.10). Then the standard bootstrap argument implies the desired estimate (A.9). 
We can generalize the previous linear stability Lemma to the nonlinear stability estimates.
Lemma A.3. Suppose that fi (x, z, t) : R2++ × R2++ × [0, T ] → R, 1 ≤ i ≤ n, satisfies
(A.12) ∂t fi + vi (x, z) · ∇x,z fi = −aii (x, z, t)fi + Bi (x, z, t) + Ni (x, z, t) + ε̄i ,
where vi (x, z, t) are some vector fields Lipschitz in x, z with vi |x1 =0 = 0, vi |z1 =0 = 0. For some
µi > 0, we define the energy
E(t) = max (µi ||fi ||L∞ ).
1≤i≤n

Suppose that Bi , Ni and ε̄i satisfy the following estimate

(A.13) X
|Bi (x, z, t)| + |Ni (x, z, t)| + |ēi | ≤ (|aij (x, z, t)|E(t) + |aij,2 (x, z, t)|E 2 (t) + |aij,3 (x, z, t)|).
j6=i

If there exists some E∗ , ε0 , M > 0 such that

X
aii (x, z, t)E∗ − (|aij |E∗ + |aij,2 |E∗2 + |aij,3 (x, z, t)|) > ε0 ,
j6=i
(A.14) X
(|aij |E∗ + |aij,2 |E∗2 + |aij,3 (x, z, t)|) < M,
j6=i

for all x, z and t ∈ [0, T ]. Then for E(0) < E∗ , we have E(t) < E∗ for t ∈ [0, T ].
We remark that the second inequality in (A.14) is only qualitative.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 123

Proof. The proof is very similar to that of Lemma A.2. We fix E(0). It suffices to prove that
under the bootstrap assumption
(A.15) E(t) < E∗ ,
on [0, T1 ], there exists ε that depends on E0 , ε0 , M, E∗ , such that we can obtain
(A.16) E(t) ≤ (1 − ε)E∗ , t ∈ [0, T1 ].
Since E(0) < E∗ and E(t) is Lipschitz, we know that the bootstrap assumption holds for
some short time T1 .
We adopt most notations from the proof of Lemma A.2 but use
X
Ci (t) , (|aij (Xi (t), Zi (t), t)|E(t) + |aij,2 (Xi (t), Zi (t), t)|E 2 (t) + |aij,3 (Xi (t), Zi (t), t)|).
j6=i

Using these notations, derivations and estimates similar to those in the proof of Lemma A.2,
we obtain Z t
Rt Rt
|Fi (t)| ≤ e− 0 −Ai (s)ds Fi (0) + Ci (s)e− s Ai (τ )dτ ds.
0
Using the bootstrap assumption and (A.14), we obtain
Ci (s) < min(M, Ai (t)E∗ − ε0 ) < (1 − δ)Ai (t)E∗ ,
for some small δ depending on ε0 , M, E∗ . Note that if Ai (t)E∗ < 2M , we pick δ such that
Ai (t)E∗ − ε0 < (1 − δ)Ai (t)E∗ . If Ai (t)E∗ > 2M , we require δ < 1/2. Now, we obtain
Rt
Z t Rt
|Fi (t)| ≤ e− 0 −Ai (s)ds |Fi (0)| + (Ai (s)E∗ − ε0 )e− s Ai (τ )dτ ds
0
Z t
− 0t −Ai (s)ds
R Rt
≤e |Fi (0)| + (1 − δ) Ai (s)E∗ e− s Ai (τ )dτ ds
0
Rt Rt
= e− 0
−Ai (s)ds
|Fi (0)| + (1 − δ)E∗ (1 − e− 0
−Ai (s)ds
)
≤ max(|Fi (0)|, (1 − δ)E∗ ) ≤ max(|E(0)|, (1 − δ)E∗ ).
Taking the supremum over the initial data of the trajectory and i, we get
E(t) ≤ max(|E(0)|, (1 − δ)E∗ ).
Since we fix E(0) and E(0) < E∗ , we can pick small δ to obtain
(1 − δ)E∗ > E(0), E(t) < (1 − δ)E∗ ,
which is (A.16). Using the bootstrap argument, we complete the proof. 

A.1. Proof of Lemma 2.2. In this section, we prove Lemma 2.2 related to the Hölder esti-
mates.

Proof. Using (2.29), we first derive the equation for f ϕ

∂t (f ϕ) + b(x) · ∇(f ϕ) = c(x)f ϕ + (b · ∇ϕ)f + Rϕ = d(x)f ϕ + Rϕ,
where d(x) = c(x) + b·∇ϕϕ is defined in Lemma 2.2. For x, z ∈ R+
2 , we derive the equation of
δ(f ϕ)(x, z) = f ϕ(x) − f ϕ(z):
∂t δ(x, z, t) + b(x) · ∇x (f ϕ)(x) − b(z) · ∇z (f ϕ)(z) = (df ϕ)(x) − (dηϕ)(z) + δ(Rψ).
Since
∇x (f ϕ)(x) = ∇x ((f ϕ)(x) − (f ϕ)(z)) = ∇x δ(f ϕ), ∇z (f ϕ)(z) = −∇z (δ(f ϕ)),
df ϕ(x) − df ϕ(z) = d(x)(f ϕ(x) − f ϕ(z)) + (d(x) − d(z))f ϕ(z) = d(x)δ(f ϕ)(x, z) + (d(x) − d(z))f ϕ(z),
we obtain
(A.17) ∂t δ(f ϕ) + (b(x) · ∇x + b(z) · ∇z )δ(f ϕ) = d(x)δ(f ϕ)(x, z) + (d(x) − d(z))(f ϕ)(z) + δ(Rϕ).
124 JIAJIE CHEN AND THOMAS Y. HOU

Since g(h) is even in h1 , h2 , ∂i g is odd in hi and we have

(b(x) · ∇x + b(z) · ∇z )(δ(f ϕ)g(x − z))

=g(x − z) · (b(x) · ∇x + b(z) · ∇z )δ(f ϕ) + δ(f ϕ) · (b(x) · ∇x + b(z) · ∇z )g(x − z)
=g(x − z) · (b(x) · ∇x + b(z) · ∇z )δ(f ϕ) + δ(f ϕ) · (b(x) − b(z)) · (∇g)(x − z).

We further multiply both sides of (A.17) by g(x − z) and use F (x, z, t) = δ(f ϕ)(x, z)g(x − z)
and the above identity to yield

(b(x) − b(z)) · (∇g)(x − z)

∂t F +(b(x)·∇x +b(z)·∇z )F = (d(x)+ )F +((d(x)−d(z))(f ϕ)(z)+δ(Rϕ))g(x−z),
g(x − z)

which concludes the proof of (2.30). 

Appendix B. Proof of Sharp Hölder estimates

In this Appendix, we prove the sharp Hölder estimates in Section 3, and discuss the properties
of the transport map and rigorous estimates of the sharp constants. Note that we have proved
Lemmas 3.1, 3.3 in Section 3. We will discuss the properties of the map related to ux and
estimate the constants in Section B.3.

1/2
B.1. Cx estimates of vx and uy . We follow the ideas in Section 3.2 and the argument in this
2 2
1 y1 −y2
section to estimate the Hölder seminorm of uy , vx . Recall the kernel K2 = 2 |y|4 for uy , vx .
Firstly, we need the following Lemma for the principle value of the integral.

Lemma B.1. Suppose that f ∈ L∞ , is Hölder continuous near 0. For 0 < a, b < ∞ and
Q = [0, a] × [0, b], [0, a] × [−b, 0], [−a, 0] × [0, b], or [−a, 0] × [−b, 0], we have
Z Z Z
π π
P.V. K2 (y)f (y)dy = lim K2 (y)f (y)dy − f (0) = lim K2 (y)f (y)dy + f (0).
Q ε→0 Q∩|y |≥ε
1
8 ε→0 Q∩|y |≥ε
2
8

In the strip |y1 | ≤ ε, K2 (y) < 0 if |y2 | > ε. It contributes to − π8 f (0) in the first identity. In
the strip |y2 | ≤ ε, K2 (y) > 0 if |y1 | > ε. It contributes to π8 f (0) in the second identity.

Proof. Since K2 is even in y1 , y2 , we focus on Q = [0, a] × [0, b] without loss of generality. By

definition, we have
Z Z a Z b Z ε Z b
P.V. K2 (y)f (y)dy = lim ( + )K2 (y)f (y)dy , lim (Iε + IIε )
Q ε→0 ε 0 0 ε ε→0

We just need to compute IIε . Since f is Hölder continuous near 0, we get

Z ε Z b
lim K2 (y)(f (y) − f (0))dy = 0.
ε→0 0 ε

The first identity follows from

Z ε Z b b
Z ε
1 y2
lim K2 (y)f (0)dy = f (0) lim 2 2 dy1
ε→0 0 ε ε→0 0 2 y2 + y1 ε
Z ε 
f (0) b ε π
= = lim − dy1 = − f (0).
2 ε→0 0 b2 + y12 ε2 + y12 8

The second identity follows from the same argument. 

 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 125

Next, we perform the sharp Hölder estimates for uy , vx . Without loss of generality, we assume
z1 = −1, x1 = 1 and z2 = x2 > 0. We are going to estimate
(B.1) Z
1 1
vx (z) − vx (x) = ( (K2,B (y1 + 1, y2 ) − K2,B (y1 − 1, y2 ))W (y1 , x2 − y2 )dy − π(ω(z) − ω(x)),
π 2
where K2,B (y) = K2 (y)1|y1 |≤B,|y2 |≤B is the localized version of K2 over [−B, B]2 .
To simplify the notation, we denote
A = min(B, x2 ), K + , K2,B (y1 + 1, y2 ), K− , K2,B (y1 − 1, y2 ), ∆(y) = K + − K − .
We focus on B ≥ 2. It is easy to see that ∆ is odd in y1 . Since the transportation cost in the
y direction is cheaper, we shall use the Y -transportation as much as possible to obtain a sharp
estimate. Due to the presence of the boundary and the discontinuity of W across the boundary,
we partition the domain into the inner part and the outer part
Ωin , {y2 ∈ [−A, A]}, Ωout , {y2 ∈
/ [−A, A]}.
1/2
Then we have ω(·, x2 − ·) ∈ C (Ωin ). We add the parameter A in these domains due to the
localization of the kernel. Define
Z A
(B.2) ∆1D (y1 ) = ∆(y1 , y2 )dy2 .
−A

Remark that for a fixed y1 , ∆(y1 , y2 ) may not have a fixed sign over y2 .
The estimates consist of three steps. In the first two steps, we estimate the integral in Ωin .
In the first step, we fix y1 and consider the 1D transportation problem on the vertical line
vly1 , {(y1 , y2 ) : y2 ∈ R}
by moving the positive part of ∆ to its negative part. If |y1 | ≤ 9, we move the remaining part
with total mass ∆1D to the horizontal line
hlx2 , {(y1 , x2 ) : y1 ∈ R}.
Here, vl, hl are short for the vertical line and the horizontal line, respectively. In this step, the
estimate is bounded by C[ω]C 1/2 .
y
In the second step, we study the transportation problem on hlx2 . We also move the remaining
part with total mass ∆1D (y1 ) for |y1 | ≥ 9 in the first step horizontally. The estimate will be
bounded by C[ω]C 1/2 for some constant C.
x
In the third step, we estimate the integral in the outer domain Ωout . The estimate will be
bounded by C[ω]C 1/2 for some constant C.
y
We focus on |y2 − x2 | ≤ B since otherwise ∆ = 0. We assume x2 > 0. The case x2 = 0 can
be obtained by taking limit x2 → 0.

B.1.1. Sign of ∆ and ∆1D . Due to the odd symmetry of ∆(y1 , y2 ) in y1 , we focus on y1 ≥ 0.
Solving K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 ) = 0, we get y2 = sc (y1 ) (B.43). It is easy to show that
K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 ) ≥ 0, |y2 | ≥ sc (y1 ),
(B.3)
K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 ) ≤ 0, |y2 | ≤ sc (y1 ).
For |y1 | ≤ B − 1, we get
(B.4) ∆(y) = K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 ).
The sign of ∆(y1 , ·) is given above. For |y1 | ∈ [B − 1, B + 1], we have
1 (y1 − 1)2 − y22
(B.5) ∆(y) = −K2 (y1 − 1, y2 ) = − .
2 ((y1 − 1)2 + y22 )2
Since B − 1 ≥ 1, it satisfies
(B.6) ∆(y) ≥ 0, |y2 | ≥ sc (y1 ) , |y1 | + 1, ∆(y) ≤ 0, |y2 | ≤ sc (y1 ) = |y1 | + 1.
126 JIAJIE CHEN AND THOMAS Y. HOU

For y1 ≥ B + 1, we have ∆(y) = 0. Next, we compute ∆1D defined in (B.2). Since ∆ is

singular at y = (±1, 0) and B > 2, the singularity is in [−9, 9] , J1 . In the inner part, we have
(B.7) Z Z Z
Sin , ∆(y1 , y2 )W (y1 , x2 − y2 )dy = ( + )∆(y1 , y2 )W (y1 , x2 − y2 )dy , S1 + S2 .
Ωin y1 ∈J1 y1 ∈J
/ 1

By definition, we yield
(B.8) Z Z
S1 = ∆(y1 , y2 )(W (y1 , x2 − y2 ) − W (y1 , x2 ))dy + ∆(y1 , y2 )W (y1 , x2 )dy , S11 + S12 .
y1 ∈J1 y1 ∈J1
1/2
For S11 , since |W (y1 , x2 − y2 ) − W (y1 , x2 )| . y2 , the integrand is locally integrable. We
will estimate S12 and S2 in Section B.1.2.
We should pay attention to the principle value in the singular integral in S12 near the singu-
larity (±1, 0). Since ∆(y) = K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 ) near y1 = 1, applying Lemma B.1
four times to −K2 (y1 − 1, y2 ), we yield
Z
+
S12 , ∆(y1 , y2 )W (y1 , x2 )dy
y1 ∈J1 ∩R+
(B.9) Z
π
= W (1, x2 ) + lim ∆(y1 , y2 )W (y1 , x2 )dy.
2 ε→0 y ∈J ∩R \[1−ε,1+ε]
1 1 +

Recall the definition of ∆ from (3.1), (B.4), (B.5). Denote

b A A
(B.10) gb (y) = , ∆1D (y1 ) = 1|y +1|≤B − 1|y −1|≤B .
y 2 + b2 (y1 + 1)2 + A2 1 (y1 − 1)2 + A2 1
For y1 ∈
/ [1 − ε, 1 + ε] , a direct computation yields
Z A A
1 y2 y2
∆(y1 , y2 )dy2 = ( − ) = gA (y1 + 1) − gA (y1 − 1) = ∆1D (y1 )
−A 2 (y1 + 1)2 + y22 (y1 − 1)2 + y22 −A

for |y1 | ≤ B − 1, and

Z A A
1 y2
∆(y1 , y2 )dy2 = − 2 + y2
= −gA (y1 − 1) = ∆1D (y1 )
−A 2 (y1 − 1) 2 −A

for |y1 | ∈ [B − 1, B + 1]. Plugging the above computation to the P.V. integral yields
Z
+ π
S12 = ∆1D (y1 )W (y1 , x2 )dy1 + ω(1, x2 ).
J1 ∩R+ 2
The computation of the integral over R− is similar due to symmetry. We yield
Z
π
(B.11) S12 = ∆1D (y1 )W (y1 , x2 )dy1 + (ω(1, x2 ) − ω(−1, x2 )).
J1 2
B.1.2. First step. We are in a position to estimate S2 (B.7) and S11 (B.8). Recall the sign of ∆
from (B.3), (B.6)
∆(y) ≥ 0, |y2 | ≥ sc (y1 ), ∆(y) ≤ 0, |y2 | ≤ sc (y1 ).
Since ∆ is even in y2 in Ωin and odd in y1 , we focus on the first quadrant.
For a fixed y1 ≥ 0, we transport the positive part of ∆ to its negative part on the line vly1 in
the first quadrant. We construct the transportation map T (y) > 0 by solving
Z y2
∆(y1 , s2 )ds2 = 0.
T (y)

For y1 ≤ B − 1, ∆ = K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 ). The map T can be obtained from the

cubic equation (B.45). For y1 ∈ [B − 1, B + 1], ∆ = −K2 (y1 − 1, y2 ) and we get
Z y2
1 s2 y2 (y1 − 1)2
0= K2 (y1 − 1, s2 )ds2 = , T (y) = .
T (y) 2 (y1 − 1)2 + s22 T y2
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 127

Denote
W̃ (y) = W (y1 , x2 − y2 ) − W (y1 , x2 ), J1+ , J1 ∩ R+ .
Using the above map, the estimates below,
|W̃ (y1 , y2 ) − W̃ (y1 , T (y))| = |W (y1 , y2 ) − W (y1 , T (y))| ≤ |y2 − T (y)|1/2 [ω]C 1/2 ,
y

1/2
|W̃ (y1 , y2 )| ≤ |y2 | [ω]C 1/2 ,
y

and applying Lemma 3.6 to the integral on [T (y1 , A), A], we yield
Z A Z A Z T (y1 ,A)
∆(y)W̃ (y)dy = ( + )∆(y)W̃ (y)dy2
0 T (y1 ,A) 0
Z A Z T (y1 ,A) 
≤ |∆(y)||y2 − T (y)|1/2 dy2 + |∆(y)||y2 |1/2 dy2 [ω]C 1/2 .
y
sc (y1 ) 0

Due to the symmetry of ∆ in y1 , y2 , we can estimate S11 (B.8) as follows

Z Z A Z T (y1 ,A) 
S11 ≤ 4 |y2 − T (y)|1/2 dy2 + |y2 |1/2 dy2 dy1 [ω]C 1/2 .
y
J1 sc (y1 ) 0

The factor 4 is due to the fact that we have 4 quadrants.

The estimate of S2 (B.7) is similar except that we do not further transport the remaining
negative part of ∆ to the location (y1 , x2 )
Z Z Z
(B.12) S2 = ( + )∆(y)W (y)dy , I + II.
y1 ∈J
/ 1 T (y1 ,A)≤|y2 |≤A |y2 |≤T (y1 ,A)

For I, we obtain
Z Z A Z Z A
|I| ≤ 2 |T (y) − y2 |1/2 dy = 4 |T (y) − y2 |1/2 dy.
y1 ∈J
/ 1 sc (y1 ) y1 ∈J
/ 1 ,y1 ≥0 sc (y1 )

For II, we use the odd symmetry of ∆(y1 , y2 ) in y1

(B.13) Z Z
|II| ≤ ∆(y)(W (y) − W (−y1 , y2 ))
y1 ∈J
/ 1 ,y1 ≥0 |y2 |≤T (y1 ,A)
Z Z p Z p
≤ 2y1 |∆(y)|dy[ω]C 1/2 = 2y1 |∆1D (y1 )|dy1 [ω]C 1/2 ,
x x
y1 ∈J
/ 1 ,y1 ≥0 |y2 |≤T (y1 ,A) y1 ∈J
/ 1 ,y1 ≥0
RA
where we have used T (y1 ,A) ∆(y)dy2 = 0, and for sc (y1 ) < A,
Z Z Z
|∆(y)|dy2 = −∆(y)dy2 = −∆(y)dy2 = −∆1D (y1 ).
|y2 |≤T (y1 ,A) |y2 |≤T (y1 ,A) |y2 |≤A

The reason why we do not transport the negative part ∆(y1 , y2 ) for |y2 | ≤ T (y1 , A), |y1 | ≥ 9
down to (y1 , x2 ) will be clear after we estimate S12 (B.11). See remark B.2

B.1.3. Second step: Estimate S12 . We combine the estimate of S12 (B.11) and the local part of
vx , e.g. − π2 (ω(z) − ω(x). For vx , since ω(z) − ω(x) = −ω(1, x2 ) + ω(−1, x2 ), we obtain
Z
π
(B.14) I , S12 − (ω(z) − ω(x)) = ∆1D (y1 )W (y1 , x2 ) + π(ω(1, x2 ) − ω(−1, x2 )).
2 J1

Recall the definition of ∆1D (y1 ) (B.10). Clearly, ∆1D is odd and ∆1D < 0 for y1 > 0. Note
that
Z k Z Z 1
A 1
P (k) , ∆1D dy1 ≥ ∆1D dy1 = − 2 + A2 dy1 = −2 arctan( A ) ≥ −π.
0 R+ −1 1y
128 JIAJIE CHEN AND THOMAS Y. HOU

We transport all the negative part of ∆1D on [0, ∞] to (1, x2 ). Similarly, we transport all the
positive part of ∆1D on (−∞, 0] to (−1, x2 ). We derive the following estimate
(B.15) Z
|I| = ∆1D (y1 )(W (y1 , x2 ) − W (1, x2 ))dy1
J1 ∩R+
Z
+ ∆1D (y1 )(W (y1 , x2 ) − W (−1, x2 ))dy1 + (π + P (9))(W (1, x2 ) − W (−1, x2 ))
J1 ∩R−
 Z √ 
≤ 2 |∆1D |y1 − 1|1/2 dy1 + (π + P (9)) 2 [ω]C 1/2 ,
x
J1 ∩R+

where we have used the symmetry of ∆1D to get the factor 2.

Remark B.2. Suppose that we further transport the negative part in II in S2 (B.12) to (y1 , x2 ).
Then we get
Z Z Z
II = ∆(y)(W (y) − W (y1 , x2 ))dy + ∆1D (y1 )W (y1 , x2 )dy , II1 + II2 .
y1 ∈J
/ 1 |y2 |≤T (y1 ,A) y1 ∈J
/ 1

For II2 , we need to incorporate it into the estimate of S12 , I (B.14)

Z
π
C , II2 + S12 − (ω(z) − ω(x)) = ∆1D (y1 )W (y1 , x2 )dy1 + π(ω(1, x2 ) − ω(−1, x2 ))
2 R
Z
= ∆1D (W (y1 , x2 ) − W (−y1 , x2 ) − W (1, x2 ) + W (−1, x2 )) + (π + P (∞))(ω(1, x2 ) − ω(−1, x2 )),
R+

for −π ≤ P (∞) < 0. For the integrand, we have two estimates

p √ p
|W (y1 , x2 ) − W (−y1 , x2 ) − W (1, x2 ) + W (−1, x2 )| ≤ min( 2y1 + 2, 2 |y1 − 1|)||W ||C 1/2 .
x
√ √
Note that for |y1 | ≥ 9, the estimate ( 2y1 + 2)||W ||C 1/2 is sharper. Then we yield
x

Z 9 p Z ∞ √ p √
|C| ≤ |∆1D |2 |y1 − 1|dy1 + |∆1D |( 2 + 2y1 )dy1 + (π + P (∞)) 2)[ω]C 1/2 .
x
0 9

Since Z ∞ √ √ √ Z ∞ Z ∞ √
|∆1D | 2dy1 + P (∞) 2 = 2( − )∆1D dy1 = 2P (9),
9 0 9
we further obtain
Z 9 p Z ∞ p √
|C| ≤ |∆1D |2 |y1 − 1|dy1 + |∆1D | 2y1 )dy1 + (π + P (9)) 2)[ω]C 1/2 .
x
0 9

We remark that the above constant of [ω]C 1/2 is the same as the sum of those in (B.13) and
x
(B.15). Moreover, we need to further bound II1 in the above. Therefore,
√ the overall estimate
√
in this approach is worse. It is due to the inequality 2|y1 − 1|1/2 ≥ 2 + 2y1 for y1 ≥ 9.
B.1.4. Third step. In this step, we estimate the integral in the outer part, which is much easier.
If B < x2 , since ∆(y1 , y2 ) is localized to |y2 | ≤ B = A, the contribution from outer part |y2 | > B
is 0. If B > x2 = A, using the odd symmetry of W and the even symmetry of ∆ in y2 , we yields
Z Z Z B Z Z −x2
Sout = ∆(y)W (y1 , x2 − y2 )dy = ∆(y)W (y1 , x2 − y2 ) + ∆(y)W (y1 , x2 − y2 )dy
Ωout R x2 R −B
Z Z B Z Z B
=− ∆(y)ω(y1 , y2 − x2 )dy + ∆(y)ω(y1 , x2 + y2 )dy.
R x2 R x2

It follows
Z Z B √
Z Z B
(B.16) |Sout | ≤ |∆(y)|·|ω(y1 , x2 +y2 )−ω(y1 , y2 −x2 ) dy ≤ 2x2 |∆(y)|dy[ω]C 1/2 .
y
R x2 R x2
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 129

1/2
B.1.5. Cx Estimate of uy . The estimates of uy in step 1 and 2 are similar to that of vx except
that we do not transport the negative part of ∆(y1 , y2 ) with |y2 | ≤ T (y1 , A) for any y1 > 0. The
reason is similar to that in Remark B.2 and will be clear later. See Remark B.2. The estimate
of the outer part in the third step is the same as that of vx in Section B.1.4.
Denote Jε = [−1 − ε, −1 + ε] ∪ [1 − ε, 1 + ε]. Note that ∆ = K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 ) has
singularities at (±1, 0). Applying Lemma B.1 four times to K2 (y1 + 1, y2 ) and −K2 (y1 − 1, y2 ),
respectively, we can rewrite Sin (B.7) as follows
Z Z A
Sin , ∆(y)W (y1 , x2 − y2 )dy
R −A
(B.17) Z Z A
π
= (ω(1, x2 ) − ω(−1, x2 )) + lim ∆(y)W (y1 , x2 − y2 )dy , I + lim IIε .
2 ε→0 Jεc −A ε→0

For y1 ∈/ Jε , we perform a decomposition

Z A
f (y1 ) , ∆(y)W (y1 , x2 − y2 )dy2
−A
Z Z
=( + )∆(y)W (y1 , x2 − y2 )dy2 , f1 (y1 ) + f2 (y1 ).
T (y1 ,A)<|y2 |≤A |y2 |≤T (y1 ,A)

For f1 (y1 ), we estimate it using Lemma 3.6

Z A
(B.18) |f1 (y1 )| ≤ 2 |∆(y)||y2 − T (y)|1/2 dy[ω]C 1/2
y
sc (y1 )

The part f2 (y1 ) denotes the purely negative part. If T (y1 , A) < A, we get T (y1 , A) < sc (y1 ) <
A and ∆(y) < 0. It follows
Z Z Z
|∆(y)|dy2 = −∆(y)dy2 = −∆(y)dy2 = −∆1D (y1 ) = |∆1D (y1 )|.
|y2 |<T (y1 ,A) |y2 |<T (y1 ,A) |y2 |≤A

Using this estimate and the fact that ∆ is odd in y1 , we get

Z
|f2 (y1 ) + f2 (−y1 )| = ∆(y)(W (y1 , x2 − y2 ) − W (−y1 , x2 − y2 ))dy2
|y2 |≤T (y1 ,A)
(B.19) Z p p
≤ 2y1 |∆(y)|dy2 [ω]C /12 = |∆1D (y1 )| 2y1 [ω]C 1/2 .
x x
|y2 |≤T (y1 ,A)

Integrating (B.18), (B.19) over y1 ∈

/ Jε , we establish
Z Z A Z p
|IIε | ≤ 4 |∆(y)||y2 − T |1/2 dy[ω]C 1/2 + |∆1D (y1 )| 2y1 [ω]C 1/2 .
y x
Jεc ∩R+ sc (y1 ) Jεc ∩R+

Notice that near the singularity (1, 0) of ∆(y), |y2 − T |1/2 . |y2 |1/2 + |y1 − 1|1/2 . Thus, the
integrand in the first integral is locally integrable. Plugging the the above estimate in (B.17),
we derive
π π
|Sin − (ω(1, x2 ) − ω(−1, x2 ))| ≤ lim sup |Iε − (ω(1, x2 ) − ω(−1, x2 ))| + IIε
2 ε→0 2
(B.20) Z Z A Z p
≤4 |∆(y)||y2 − T |1/2 dy[ω]C 1/2 + |∆1D (y1 )| 2y1 [ω]C 1/2 .
y x
R+ sc (y1 ) R+

Recall the definition of localized uy (3.1). Combining the above estimate and the estimate of
Sout in (B.16), we prove the estimate of uy .
Remark B.3. For a fixed y1 , if we transport ∆(y1 , y2 ) for |y2 | ≤ A to the line hlx2 , we need to
estimate the 1D integral
Z
π π
S , ( ∆1D (y)W (y1 , x2 ) + (ω(1, x2 ) − ω(−1, x2 ))) − (ω(1, x2 ) − ω(−1, x2 ))).
R 2 2
130 JIAJIE CHEN AND THOMAS Y. HOU

The first part is derived similar to (B.11), and the second part is from the local term in uy (3.1).
Since the local terms are exactly canceled, and ∆1D < 0 for y1 > 0 and ∆1D > 0 for y1 < 0, the
sharp estimate of S is trivial
Z Z p
|S| = D1D (y)(W (y1 , x2 ) − W (−y1 , x2 ))dy1 ≤ |D1D (y) 2y1 |dy1 [ω]Cx 1/2 .
R+ R+

The constant of [ω]Cx 1/2 is exactly the same as that in (B.20).

1/2
B.2. Cy estimate of vx . Since W is not continuous across the boundary y2 = 0, the localized
vx or uy is no longer Hölder continuous. Therefore, we study the estimate without localization.
Without loss of generality, we assume z2 = m + 1, x2 = m − 1, x1 = z1 = 0 with m > 1. The
case m = 1 can be obtained by taking limit. The difference vx (z) − vx (x) or uy (z) − uy (x) is
given by
Z
1
I, K2 (y1 , 1 + y2 ) − K2 (y1 , y2 − 1)W (y1 , m − y2 )dy + s(ω(z) − ω(x)),
π
1
where s = 2 for uy and s = − 21 for vx . Denote
(B.21) η(y1 , y2 ) = W (y2 , y1 ), ηm (y1 , y2 ) = η(m − y1 , y2 ).
By definition, η is odd in y1 , discontinuous across y1 , and satisfies
(B.22) [η]C 1/2 (Ω) = [ω]C 1/2 , [η]C 1/2 (Ω) = [ω]C 1/2 ,
x y y x

for Ω = {y : y1 ≥ 0} or Ω = {y : y1 ≤ 0} .
Swapping the dummy variables y1 , y2 and then using K2 (y1 , y2 ) = −K2 (y2 , y1 ), ηm (y) =
η(m − y1 , y2 ) = W (y2 , m − y1 ), we yield
Z
1
I=− (K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 ))η(m − y1 , y2 )dy + s(ω(z) − ω(x))
π
(B.23) Z
1 1 1
=− ∆(y)ηm (y)dy + s(ηm (−1, 0) − ηm (1, 0)), s = − for vx , s = for uy .
π 2 2
We perform the above reformulation so that we can adopt the analysis of
∆(y) = K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 )
in (B.3) and Section B.1.1. Since η is discontinuous across y1 = 0, which relates to y1 = m in
the integral in (B.23), and the singularity of ∆ is at y = (±1, 0), we decompose the integral into
the inner region, the middle region, and the outer region
Ωin , {y : |y1 | ≤ 1}, Ωmid , {y : |y1 | ∈ [1, m]}, Ωout , {y : |y1 | > m},
(B.24)
Z
Sα , ∆(y)ηm (y)dy, α ∈ {in, mid, out}.
Ωα

In each region, ηm is Hölder continuous. Since we can obtain a small factor from [ω]C 1/2 and
y
we can use (B.22) to obtain a sharp estimate of (B.23), we should use the X transportation as
much as possible.
Firstly, we analyze the sign of ∆(y). Since ∆ is odd in y1 and even in y2 , we can focus on
y1 , y2 ≥ 0. For a fixed y2 , we have
∆(y) < 0, y1 < h−
c (y2 ) ≤ 1, ∆(y) > 0, y1 ∈ (h−
c (y2 ), 1), y2 ≤ yc , 3−1/2 ,
(B.25)
∆(y) < 0, y1 > h+
c (y2 ) ≥ 1, ∆(y) > 0, y1 ∈ (1, h+
c (y2 )),

where h± ω
c (y2 ) solves ∆(hc (y2 ), y2 ) = 0 and is given explicitly in (B.43). Denote Qε = [1 − ε, 1 +
ε] × [−ε, ε] and Qi is the four quadrants with center at (1, 0), e.g. Q1 = {y1 ≥ 1, y2 ≥ 0}. For
y 2 −y 2
the P.V. integral, since the kernel 1|y|4 2 has mean 0 in each quadrant R± × R± , it is not difficult
to show that
Z X 4 Z
lim ∆(y)η(m − y1 , y2 )1y1 ≥0 dy = lim ∆(y)η(m − y1 , y2 )1y1 ≥0 dy.
ε→0 Qcε εi →0 Qcε ∩Qi
i=1
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 131

That is, we can compute the P.V. integral separately in each Qi .

In the following subsections, we estimate the integral in each region.
B.2.1. Inner region Ωin . In this region, we have |y1 | ≤ 1. Denote yc = 3−1/2 . Note that
∆ = K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 ) is singular at (1, 0). Applying Lemma B.1 to −K2 (y1 − 1, y2 )
yields
Z 1 Z yc Z 1 Z yc
π π
(B.26) ∆(y)ηm (y)dy = lim ∆(y)ηm (y)dy − ηm (1, 0) , lim Iε − ηm (1, 0).
0 0 ε→0 0 ε 8 ε→0 8
Let T1 (y) ≥ 0 be the map that solves
Z T1 (y)
∆(s, y2 )ds = 0,
y1

which is given in (B.46). Using the sign inequality (B.25) and applying Lemma 3.6 in the y1
direction, we yield
Z yc Z T1 (0,y2 ) Z yc Z h−
c (y2 )
dy2 ∆(y)ηm (y)dy1 ≤ dy2 |∆(y)||y1 − T1 (y)|1/2 dy1 [η]C 1/2 .
x
ε 0 ε 0
Using the symmetry of ∆ in y1 , y2 , we obtain
Z Z Z yc Z h−
c (y2 )
∆(y)ηm (y)dy ≤ 4 dy2 |∆(y)||y1 − T1 (y)|1/2 dy1 [η]C 1/2 .
x
ε≤|y2 |≤yc |y1 |≤T1 (0,|y2 |) ε 0

The remaining part of the integral in Ωin is in the following region

Rin,ε , {|y1 | ≤ 1, |y2 | ≥ yc } ∪ {T1 (0, |y2 |) ≤ y1 ≤ 1, ε ≤ |y2 | ≤ yc },
(B.27) + ++
Rin,ε = Rin,ε ∩ [0, 1] × R, Rin,ε = Rin,ε ∩ [0, 1] × R+ .
Since ∆ > 0 in R+
in,ε , we use the odd symmetry of ∆ in y1 and even symmetry in y2 to obtain
Z Z Z p
∆(y)ηm (y)dy = ∆(y)(ηm (y1 , y2 ) − ηm (−y1 , y2 ))dy ≤ 2 |∆(y)| 2y1 dy[η]C 1/2 ,
+ ++ x
Rin,ε Rin,ε Rin,ε

where we have the factor 2 since the estimates in y2 ≥ 0 and y2 ≤ 0 are the same.
Plugging the above estimate in (B.26) and using the symmetry of ∆ in y1 , y2 , we derive
(B.28)
Z
π
∆(y)ηm (y) + (ηm (1, 0) − ηm (−1, 0))
Ωin 4
 Z yc Z h−
c (y 2 ) Z p 
≤ 4 dy2 |∆(y)||y1 − T1 (y)|1/2 dy1 + 2 |∆(y)| 2y1 dy [η]C 1/2 , Cin [η]C 1/2 .
++ x x
0 0 Rin,0

B.2.2. Estimate in Ωmid . We develop two estimates for the integral (B.24)
Z
Smid , ∆(y)ηm (y)dy.
Ωmid
First estimate. The first estimate is similar to that in Section B.2.1. Notice that the sin-
gularities of ∆ are (±1, 0). We first rewrite Smid as follows using Lemma B.1 twice with
Q = [1, m] × [0, 1] and Q = [1, m] × [−1, 0]
Z
π
(B.29) Smid = lim ∆(y)ηm (y)dy − (ηm (1, 0) − ηm (−1, 0)).
ε→0 Ω
mid ∩|y2 |≥ε
4
To estimate the integral, we first study the sign of ∆(y). For y1 ∈ [1, m], we have
∆(y) > 0, |y2 | > sc (m),
where sc is given in (B.43). For |y2 | < sc (m), the sign of ∆(y) is given in (B.25). Denote
Rout , {|y1 | ∈ [1, m], |y2 | ≥ sc (m)} ∪ {|y2 | < sc (m), 1 ≤ |y1 | ≤ T (m, |y2 |)},
(B.30) + ++
Rout , Rout ∩ {y1 ≥ 0}, Rout , Rout ∩ R++
2 .
132 JIAJIE CHEN AND THOMAS Y. HOU

In Ωout \Rout , using the sign properties of ∆ (B.25) and applying Lemma 3.6 in the y1
direction, we yield
Z sc (m) Z m Z sc (m) Z m
∆(y)ηm (y)dy ≤ dy2 |∆(y)||y1 − T1 (y)|1/2 dy,
ε T (m,y2 ) ε h+
c (y2 )

where T1 is given in (B.46). For the integral in Rout , ∆(y) is positive if y1 > 0. We use the odd
symmetry of ∆ and
p
|ηm (y1 , y2 ) − η(−y1 , y2 )| ≤ 2y1 [η]C 1/2 .
x

In particular, we obtain an estimate similar to (B.28)

π  Z sc (m) Z m
|Smid + (ηm (1, 0) − ηm (−1, 0))| ≤ 4 dy2 |∆(y)||y1 − T1 (y)|1/2 dy1
4 0 h+
c (y2 )
(B.31) Z 
p
+2 |∆(y)| 2y1 dy [η]C 1/2 , Cmid,1 (m)[η]C 1/2 .
++ x x
Rmid

Second estimate. In the second estimate, instead of using transportation in the y1 direction,
we use transportation in the y2 direction. This estimate will be very useful for vx with small m.
Firstly, applying Lemma B.1 twice to K2 (y1 + 1, y2 ) and −K2 (y1 − 1, y2 ), respectively, we yield
Z
π
Smid = (ηm (1, 0) − ηm (−1, 0)) + lim ∆(y)ηm (y)dy , I + lim IIε .
4 ε→0 Ω
mid ∩||y1 |−1||≥ε
ε→0

We remark that the above decomposition and the sign of (ηm (1) − ηm (−1)) are different from
those in (B.29). Here, we restrict y1 to be ε away from ±1 in the limit, while in (B.29), we
restrict y2 to be ε away from 0. Recall that ∆ satisfies the sign condition (B.4). Note that
Z ∞ ∞
y2 y2
(B.32) ∆(y1 , y2 )dy2 = 2 − 2 = 0,
0 y2 + (y1 + 1)2 y2 + (y1 − 1)2 0
and ∆ is even in y2 and odd in y1 . Applying Lemma 3.6 in the y2 direction, we obtain
Z m Z ∞
(B.33) IIε ≤ 4 dy1 |∆(y)||y2 − T (y)|1/2 dy2 [η]C 1/2 .
y
1+ε sc (y1 )

We have a factor 4 since the same estimate applies to integral in each quadrant. It follows
Z m Z ∞
π
|Smid − (ηm (1, 0) − ηm (−1, 0))| ≤ 4 dy1 |∆(y)||y2 − T (y)|1/2 dy2 [η]C 1/2
(B.34) 4 1 sc (y1 )
y

, Cmid,2 (m)[η]C 1/2 .

y

When we combine different estimates to estimate [vx ]C 1/2 , the factor − π4 (ηm (1) − ηm (−1))
y
allows us to cancel the local term in vx . The factor π4 (ηm (1)−ηm (−1)) in (B.31) has the opposite
sign and does not offer such cancellation.

B.2.3. Estimate in the outer region. We also develop two estimates for the integral (B.24)
Z
Sout , ∆(y)ηm (y)dy.
Ωout

In Ωout , we note that ηm ∈ C 1/2 ([m, ∞) × R) and ηm ∈ C 1/2 ((−∞, m] × R). We will apply the
same estimate to the integral in Ωout ∩ {y1 ≥ m} and in Ωout ∩ {y1 ≤ −m}. Thus, we focus on
the estimate in the region with y1 ≥ m.
First estimate. The first estimate is similar to that in the second estimate of Smid in Section
B.2.2. Following the argument in (B.32), (B.33), we obtain
Z ∞Z ∞
(B.35) |Sout | ≤ 4 |∆(y)||y2 − T (y)|1/2 dy[η]C 1/2 , Cout,1 (m)[η]C 1/2 .
y y
m sc (y2 )
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 133

Second estimate. In the second estimate, we use the horizontal transportation as much as
possible. Recall the function g from (B.10). Denote
y1 − 1 ∞
Z
1 y1 + 1
∆m (y2 ) , ∆(y1 , y2 )dy1 = − 2 2 + 2 2
y1 ≥m 2 (y1 + 1) + y2 (y1 − 1) + y2 m
1  m+1 m−1  1
= − = (gm+1 (y2 ) − gm−1 (y2 )),
2 (m + 1)2 + y22 (m − 1)2 + y22 2
Z
Jm (y2 ) , ∆(y)ηm (y)dy1 , Em (y2 ) , Jm (y2 ) − ∆m (y2 )ηm (m, y2 ).
y1 ≥m
It is easy to obtain that
p
(B.36) ∆m (y2 ) < 0, |y2 | < ym , m2 − 1, ∆m (y2 ) > 0, |y2 | > ym .
For a fixed y2 , our idea is to apply Lemma 3.6 to estimate the integral on the line [m, ∞)×{y2 },
and then transport the remaining positive or negative part with total mass ∆1D (y2 ) to (m, y2 ).
Then, we study the transportation problem on the line {m} × R.
Firstly, we estimate Em (y2 ). For y2 ∈ [0, sc (m)], we have ∆(y) < 0 for y1 ≥ m and estimate
Em (y2 ) as follows
Z Z
|Em (y2 )| = ∆(y)(ηm (y) − ηm (m, y2 )dy1 ≤ |∆(y)||y1 − m|1/2 dy[η]C 1/2 .
x
y1 ≥m y1 ≥m

It is not difficult to show that sc (m) < ym = (m2 − 1)1/2 from (B.43). From the sign property
(B.25) and (B.36), for y2 ∈ [sc (m), ym ) , we get
Z T (m,y2 )
∆(y)dy1 = 0, for some T (m, y1 ) ∈ (h+
c (y2 ), ∞).
m
For y2 ∈ (ym , ∞), we get
Z ∞
∆(y)dy1 = 0, T (∞, y2 ) = (y12 − 1)1/2 ∈ (m, sc (y2 ).
T (∞,y2 )

Therefore, for y2 ∈ [sc (m), ym ], we have the following estimate

Z T (m,y2 ) Z ∞
|Em (y2 )| = ∆(y)ηm (y)dy1 + ∆(y)(ηm (y) − ηm (m, y2 ))dy1
m T (m,y2 )
Z h+
c (y2 )
Z ∞ 
≤ |∆(y)||y1 − T (y)|1/2 + |∆(y)||y1 − m|1/2 dy1 [η]C 1/2 .
x
m T (m,y2 )

For y2 ∈ [ym , ∞), we yield

Z ∞ Z T (∞,y2 )
|Em (y2 )| = ∆(y)ηm (y)dy1 + ∆(y)(ηm (y) − ηm (m, y2 ))dy1
T (∞,y2 ) m
 Z ∞ Z T (∞,y2 ) 
≤ |∆(y)||y1 − T (y)|1/2 dy1 + |∆(y)||y1 − m|1/2 dy1 [η]C 1/2 .
x
h+
c (y1 ) m

The estimate of Em (y2 ) with y2 < 0 is the same since ∆(y) is even in y2 .
Now it remains to estimate the transportation on m × R
Z
II = ∆m (y2 )ηm (m, y2 )dy2 .

From the sign properties (B.36), we contruct the 1D map as follows

Z T (y2 )
m2 − 1
∆m (s)ds = 0, Tm (y2 ) = .
y2 y2
R∞
Note that 0 ∆m (s)ds = 0. Applying Lemma 3.6, we establish
Z ∞
m2 − 1 1/2
|II| ≤ 2 |∆m (y2 )||y2 − | dy2 [η]C 1/2 .
ym y2 y
134 JIAJIE CHEN AND THOMAS Y. HOU

We can apply the above estimates to estimate the integral in Ωout ∩ {y1 ≤ −m}, which leads to
an additional factor 2 in (B.37). Denote
Rout,1 , {y2 ∈ [sc (m), ym ], m ≤ y1 ≤ h+ +
c (y2 )} ∪ {y2 ≥ ym , y1 ≥ hc (y2 )},

Rout,2 , [m, ∞] × [0, sc (m)] ∪ {y2 ∈ [sc (m), ym ], y1 ≥ T (m, y2 )} ∪ {y2 ≥ ym , m ≤ y1 ≤ T (∞, y2 )}.
Combining the above estimates, we prove
Z ∞
m2 − 1 1/2
|Sout | ≤ 4 |∆m (y2 )||y2 − | dy2 [η]C 1/2
ym y2 y

Z Z 
(B.37) +4 |∆(y)||y1 − T (y)|1/2 dy + |∆(y)||y1 − m|1/2 dy [η]C 1/2
x
Rout,1 Rout,2

, Cout,2y (m)[η]C 1/2 + Cout,2x (m)[η]C 1/2 .

y x

B.2.4. Summarize the estimates. Recall from (B.23)

Z
1 1
vx (z) − vx (x) = − ∆(y)ηm dy − (ηm (−1, 0) − ηm (1, 0))
π 2
1 π
= − (Sin + Smid + Sout − (ηm (1, 0) − ηm (−1, 0)), )
π 2
1 π
uy (z) − uy (x) = − (Sin + Smid + Sout + (ηm (1, 0) − ηm (−1, 0)).
π 2
√
Note that |(ηm (1, 0) − ηm (−1, 0))| ≤ 2[η]C 1/2 . Combining the estimates (B.28), (B.31), (B.34),
x
(B.35), (B.37), we prove
π|uy (z) − uy (x)| ≤ (Cin + Cmid,1 )[η]C 1/2 + min(Cout,1 [η]C 1/2 , Cout,2x [η]C 1/2 + Cout,2y [η]C 1/2 ),
x y x y
√ π√
π|vx (z) − vx (x)| ≤ Cin [η]C 1/2 + min(Cmid,1 [η]C 1/2 + π 2[η]C 1/2 , Cmid,2 [η]C 1/2 + 2[η]C 1/2 )
x x x y 2 x

+ min(Cout,1 [η]C 1/2 , Cout,2x [η]C 1/2 + Cout,2y [η]C 1/2 ).

y x y

Using the relation (B.22), we can rewrite the upper bound in terms of [ω]C 1/2 , [ω]C 1/2 .
x y

B.3. Estimate of the sharp constants for ux . Recall that the map T (s1 , s2 ) solves the cubic
equation (3.15)
(B.38) 0 = −1 − 8T s1 + 16T s1 (T 2 + T s1 + s21 ) − 8s22 (1 − 4s1 T + 2s22 ).
Firstly, for fixed s2 and s1 > 0, we show that it has a unique solution on [0, ∞). Then, we
study the properties of T (s1 , s2 ) so that we can estimate C1 (b) in Lemma 3.1 effectively.
Note that the above identity is symmetric in T and s1 and is a cubic equation in T . For
s1 > 0, dividing s1 on both side yields
1 (4s22 + 1)2
(B.39) g(T ) , T 3 + T 2 s1 + T (s21 −
+ 2s22 ) − = 0.
2 16s1
Since g(0) < 0 and g(∞) > 0, the above equation has at least one real root on R+ . We introduce
Z = T + s31 and can rewrite the above equation in terms of Z
s1 3 s31 2 1 (4s22 + 1)2
0 =(T + ) − + T ( s21 − + 2s22 ) −
3 27 3 2 16s1
2 2
(B.40) 2 1  (4s2 + 1) 7 s1 1 
=Z 3 + Z( s21 − + 2s22 )) − + s31 + (2s22 − )
3 2 16s1 27 3 2
3
,Z + p(s1 , s2 )Z + q(s1 , s2 ).
The discriminant is given by
∆Z (s1 , s2 ) = −(27q(s1 , s2 )2 + 4p(s1 , s2 )3 ).
Note that r
7 3 s1 1 7 1 1
−q ≥ s − + ≥ (2 · − )s1 ≥ 0,
27 1 6 16s1 27 16 6
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 135

and −q, p are increasing in s2 . We yield

(1 − 4s21 )2 (27 − 56s21 + 48s41 )
−∆Z (s1 , s2 ) ≥ −DZ (s1 , 0) = C ≥ 0.
256s21
Using the solution formula for a cubic equation, we obtain that the cubic equation for T or
Z has a unique real root given by
q
4 3
p  −q + q 2 + 27 p 1/3 s1
(B.41) Z = r1 − , r1 = , T =Z− ,
3r1 2 3
where p, q are defined in (B.40).
We have the following basic properties for the map T and the threshold f (s2 ) (3.11).
Lemma B.4. The map T (s1 , s2 ) is increasing in s2 and decreasing in s1 . Moreover, we have
f (s2 ) ≥ 21 and
3
4s22 + 1 ≥ max(4s1 T (s1 , s2 ), 1), T 2 + T s1 + s21 + 2s22 ≥ , for s1 ≥ 0,
4
(B.42) 1
|T (s1 , s2 ) − s1 | ≤ max (|Tx (x, s2 )| + 1)|s1 − |, s1 ≥ f (s2 ).
x∈[f (s2 ),s1 ] 2
1
Proof. The estimate f (s1 ) ≥ 2 follows from (3.11). Denote P = 4s1 T, Q = 4s22 + 1. Using
(B.38), we get
Q2 = (4s22 + 1)2 = 16T s1 (T 2 + T s1 + s21 ) + 32s22 s1 T − 8T s1
≥ 16T s1 · 3T s1 + 2(Q − 1)P − 2P = 3P 2 + 2P Q − 4P.
If P ≤ 1, we derive Q ≥ 1 ≥ P . If P > 1, solving the quadratic equation in Q, we yield
p p
Q ≥ P + 4P 2 − 4P , or Q ≤ P − 4P 2 − 4P .
√
Note that for P > 1, we have P − 4P 2 − 4P < 1. Thus, we must have
p
Q ≥ P + 4P 2 − 4P ≥ P.
This proves the first inequality in (B.42). Using (B.38) again, we derive
1 1 3
T 2 + T s1 + s21 + 2s22 = (8T s1 + (4s22 + 1)2 ) ≥ (8T s1 + 4T s1 ) = ,
16T s1 16T s1 4
where we have used the first inequality in (B.42) that we just proved.
The last inequality follows directly from f (s2 ) ≥ 21 and the mean value theorem.
Since (B.38) is symmetric in T and s1 , and its has a unique positive real root for any s1 > 0,
we get T (T (s1 , s2 ), s2 ) = s1 , or T ◦ T = Id. Using (3.14), we get
K(s1 , s2 ) = Tx (s1 , s2 )K(T (s1 , s2 ), s2 ).
Since K(s1 , s2 ) and K(T (s1 , s2 ), s2 ) have opposite sign, it follows Tx ≤ 0.
Taking s2 derivative on both sides of (B.39), we yield
dT 1 (4s22 + 1)2
(3T 2 + 2T s1 + s21 − + 2s22 ) + s2 (4T − ) = 0.
ds2 2 s1
dT
Using the first and the second inequality in (B.42), we prove ds2 ≥ 0. 

B.4. Some functions related to the computation of the sharp constant. Here, we list
some functions that will be used to compute the sharp constant. Recall
y1 y2 1 y12 − y22
K1 = , K2 = .
|y|4 2 |y|4
136 JIAJIE CHEN AND THOMAS Y. HOU

B.4.1. Sign functions. Solving K1 (y1 + 1/2, y2 ) − K1 (y1 − 1/2, y2 ) = 0 for y1 ≥ 0, we yield
 1/2 − 2y 2 + p16y 4 + 4y 2 + 1 1/2
2 2 2
y1 = .
6
See also (3.11).
Solving K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 ) = 0 for y2 ≥ 0, we yield
 q 1/2
y1 = h± c (y2 ) , y2
2
+ 1 ± 2y2 y 2+1
2 ,
(B.43)  −(y 2 + 1) + 2(y 4 − y 2 + 1)1/2 1/2
1 1 1
y2 = sc (y1 ) , .
3
B.4.2. Transportation maps.
Map for ux . For a fixed s2 , solving
Z s1
(K1 (s1 + 1/2, s2 ) − K1 (s1 − 1/2, s2 ))ds1 = 0,
T (s)

yields
1 (4s22 + 1)2
(B.44) T 3 + T 2 s1 + T (s21 − + 2s22 ) − = 0,
2 16s1
or equivalently
2 1  (4s2 + 1)2 7 s1 1 
2
Z 3 + Z( s21 − + 2s22 )) − + s31 + (2s22 − ) = 0,
3 2 16s1 27 3 2
for Z = T + s31 . It has a unique real root that can be obtained explicitly. See (B.38)-(B.40).
Map for [uy ]C 1/2 . For a fixed y1 ≥ 0, solving
x
Z y2
(K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 ))dy2 = 0,
T (y)

yields
(y12 − 1)2
(B.45) T 3 + T 2 y2 + T (y22 + 2 + 2y12 ) − = 0,
y2
or equivalently
2y22  (y 2 − 1)2
1 y3 y2 2y 2 
W3 + W( + 2 + 2y12 ) − + 2 + ( 2 + 2 + 2y1 ) = 0,
3 y2 27 3 3
where W = T + y32 . It has a unique real root that can be obtained explicitly.
Map for [uy ]C 1/2 . For a fixed y1 , y2 ≥ 0, solving
y
Z y2
(K2 (y1 + 1, y2 ) − K2 (y1 − 1, y2 ))dy1 = 0,
T (y)

yields
1 − T 2 − y12 + T 2 y12 − 2y22 − T 2 y22 − y12 y22 − 3y24 = 0,
or equivalently
y12 + 2y22 + y12 y22 + 3y24 − 1
(B.46) T2 = .
y12 − y22 − 1
We remark that we apply the above map to two regions
y1 ∈ [0, 1], y1 ≤ hc− (y2 ), y1 ∈ [1, ∞], y1 ≥ hc+ (y2 )
separately.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 137

B.5. Some Lemmas for the computation of the constants. We have the following Lemma
to estimate the integral near the singular region.
1/2
Lemma B.5. For 0 ≤ a < b and f ∈ Cy ([a, b] × [0, ∞]), we have
Z bZ ∞ Z ∞ 2
1/2 1/2 s −1 1
K2 (y)f (y)dy ≤ |b − a | 2 + 1)2
|s − |1/2 ds[f ]C 1/2 .
a 0 1 (s s y

Proof. For a fixed y1 , note that K2 (y) > 0 if y2 < y1 , K2 (y) < 0 if y2 > y1 and
Z y2
K2 (y1 , s)ds = 0.
y12 /y 2

Applying Lemma 3.6 in the y2 direction, we obtain

Z bZ ∞ Z bZ ∞
y2
I, K2 (y)f (y)dy ≤ |K2 (y)||y2 − 1 |1/2 dy[f ]C 1/2 .
a 0 a y1 y2 y

Using change of variable y2 = sy1 , K2 (y1 , y2 ) = y1−2 K2 (1, s), we prove

b ∞ ∞
s2 − 1
Z Z Z
1/2
I≤ y1 |K2 (1, s)||s−s−1 |1/2 ds[f ]C 1/2 = |b1/2 −a1/2 | |s−s−1 |1/2 ds[f ]C 1/2 .
a 1
y
1 (s2 + 1)2 y

Appendix C. Weights and parameters

C.1. Parameters for the weights. In our energy estimates and the estimates of the nonlocal
terms, we need various weights. For the Hölder estimates, we use the following weights

ψ1 = |x|−2 + 0.6|x|−1 + 0.3|x|−1/6 ,

ψ2 = p2,1 |x|−5/2 + p2,2 |x|−1 + p2,3 |x|−1/2 + p2,4 |x|1/6 ,
(C.1) ψ3 = ψ2 , p~2,· = (0.46, 0.245, 0.3, 0.112),
p p
gi (h) = gi0 (h)gi0 (1, 0)−1 , gi0 (h) = ( h1 + qi2 h2 + qi3 h2 + qi4 h1 )−1 ,
~q1, = (0.12, 0.01, 0.25), ~q2, = (0.14, 0.005, 0.27), ~q3, = ~q2, .

For the weighted L∞ estimates with decaying weights, we use the following weights

ϕ1 = x−1/2 (|x|−2.4 + 0.6|x|−1/2 ) + 0.3|x|−1/6 ,

ϕ2 = x−1/2 (p5,1 |x|−5/2 + p5,2 |x|−3/2 + p5,3 |x|−1/6 ) + p5,4 |x|−1/4 + p5,5 |x|1/7 ,
(C.2) ϕ3 = x−1/2 (p6,1 |x|−5/2 + p6,2 |x|−3/2 + p6,3 |x|−1/6 ) + p6,4 |x|−1/4 + p6,5 |x|1/7 ,
p5,· = (0.3851, 0.1238, 0.1981, 0.1669, 0.0328),
p6,· = (0.9628, 0.3590, 0.6170, 0.3021, 0.0890).

For the weighted L∞ estimates with the growing weights, we use the following weights

ϕg1 = ϕ1 + p71 |x|1/16 , ϕg2 = ϕ2 + p81 |x|1/4 ,

(C.3) ϕg3 = (p31 |x|−5/2 + p32 |x|−3/2 + p33 |x|−1/6 )x−1/2 + p34 |x|−1/4 + p91 |x|1/7 + p92 |x|1/4 ,
p71 = 1, p81 = 0.071, p9,1 = 0.1046, p9,2 = 0.1540.

Parameters in the energy. We choose the following parameters in our energy (4.21), (4.25)

(C.4) τ1 = 5, µ1 = 0.668, µ2 = 1.336, µ4 = 0.065, τ2 = 0.24.

138 JIAJIE CHEN AND THOMAS Y. HOU

C.2. Parameters for approximating the velocity. We choose the following parameters
xi , ti in the first approximation of velocity u, ux in (2.82), (2.89) in Section 2.11,
ux : x = (1, 2, 3, 4, 6, 8, 11, 16, 22, 32, 48) · 64hx , t = (16, 16, 20, 24, 32, 40, 56, 72, 96, 128, 256)h,
u : x = (1, 2, 4, 8, 12, 16, 22, 32, 64) · 64hx , t = (8, 8, 24, 40, 56, 72, 96, 128, 256)h,
where hx , h are chosen in (8.11).
For u, v, ux , we need the second approximation (2.88). We choose the following parameters
Ri in (2.88)
R = (8, 16, 32, 64, 128, 256, 512, 1024) · 64hx .

C.3. Estimate of the weights. In our energy estimates and the estimates of the nonlocal
terms, we need various estimates of the weights and their derivatives. From (C.1), (C.2), (C.3),
we have two types of weights. The first one is the radial weight
X
ρ(x, y) = pi rai , r = (x2 + y 2 )1/2 ,
i

where ai is increasing and pi ≥ 0. We use these weights for the Hölder estimates. See e.g. (C.1).
The second type of weights is the following
ρ(x, y) = ρ1 (r)x−α + ρ2 (r),
where ρ1 , ρ2 are the radial weights.
We use fl , fu to denote the lower and upper bound of f . We have the following simple
inequalities
(f − g)l = fl − gu , (f − g)u = fu − gl , (f + g)γ = fγ + gγ ,
(C.5)
(f g)l = min(fl gl , fu gl , fl gu , fu gu ), (f g)u = max(fl gl , fu gl , fl gu , fu gu ).
where γ = l, u. If g ≥ 0, we can simplify the formula for the product
(C.6) (f g)l = min(fl gl , fl gu ), (f g)u = max(fu gl , fu gu ).

C.4. Radial weights. The advantage of radial weights is that we can estimate them easily.

C.4.1. Bounds for the derivatives. We can easily derive the derivatives and their upper and
lower bound as follows. Firstly, we have
X
(∂xi ∂yj ρ(x, y))γ = pk (∂xi ∂yj rak )γ ,
1≤k≤n

where γ = l, u. Using induction, we any α, i, j, we can obtain

X X
+ −
∂xi ∂yj rα = Ci,j,k,l (α)xk y l rα−i−j−k−l = (Ci,j,k,l (α)−Ci,j,k,l (α))xk y l rα−i−j−k−l .
k≤i,l≤min(j,1) k≤i,l≤min(j,1)
±
The bounds for Ci,j,k,l (α)xk y l rα−i−j−k−l are simple:
± ±
(Ci,j,k,l (α)xk y l rα−i−j−k−l )γ = Ci,j,k,l (α)xkγ yγl rγα−i−j−k−l .
Using (C.5), the above identities, and linearity, we can obtain the upper and lower bounds
for ∂xi ∂yj ρ.

C.4.2. Leading order behavior of ∂ρ/ρ. In our verification, we need to bound ∂ρ(λx)/ρ(λx) as
λ → 0 or λ → ∞ uniformly. A direct calculation yields
ai
p a rai
P P
∂xi ρ xi i pi ai r xi
Pi i i a .
= 2
P a
, 2
S(x), S(x) ,
ρ |x| i pi r
i |x| i pi r
i

For x close to 0, we introduce b = a − a1 . Clearly, we get bi ≥ 0 and

pi bi rai bi
P P
i i pi bi r A(r)
S(x) = a1 + P a
= a1 + P b
, a1 + .
i pi r i pi r B(r)
i i
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 139

Using bi ≥ 0 and the Cauchy-Schwarz inequalities, we yield

(C.7)
X X X  1X
A0 B − AB 0 = r−1 ( pi b2i rbi )( p i r bi ) − ( pi bi rbi )2 = r−1 pi pj (bi − bj )2 rbi +bj ≥ 0,
2 ij

which implies that A/B is increasing. Thus for λ ≤ λ∗ , r ∈ [rl , ru ], we get the uniform bound
for S(λx)
A(λ∗ ru )
a1 ≤ S(λx) ≤ a1 + .
B(λ∗ ru )
For λ = 1, we simply obtain
A(rl ) A(ru )
a1 + ≤ S(x) ≤ a1 + .
B(rl ) B(ru )
Similarly, for λ ≥ λ∗ , r ∈ [rl , ru ], we get
pi bi rbi
P
A(λ∗ rl ) A(r)
an + ≤ S(λx) ≤ an , = Pi bi
,
B(λ∗ rl ) B(r) i pi r

where b = a − an ≤ 0. Here, we have used that A(r)/B(r) is decreasing.

From the above estimates, we yield
∂xi ρ xi ∂x ρ xi |A(λ∗ x)|
lim = 2
a1 , R0 (x), | i (λx) − R0 (λx)| ≤ λ−1 2 , λ ≤ λ∗ ,
λ→0 ρ |x| ρ |x| |B(λ∗ x)|
∂xi ρ xi ∂x ρ xi |A(λ∗ x)|
lim = an , R∞ (x), | i (λx) − R∞ (λx)| ≤ λ−1 2 , λ ≥ λ∗ .
λ→∞ ρ |x|2 ρ |x| |B(λ∗ x)|
Using (C.7), we can estimate the Lipschitz coefficient of S
1 2 bi +bj −1 bi +bj
P
|S(λx) − S(λz)| A(λr) 2 ij pi pj (bi − bj ) r λ
≤ max ∂r ≤ 2
.
||x| − |z|| r∈[|x|,|z|] B(λr) Bl (λr)
Note that we have bi ≥ 0 for λ → 0 and bi ≤ 0 for λ → ∞. We can derive the upper and lower
bounds for the denominator and the numerator.
We use the above estimates to bound the integrals for ∂x (∂ 2 φρ), where φ is the stream
functions.

C.4.3. Bounds for the derivatives of 1/ρ. The bounds for dix djy ρ−1 is more complicated since
ρ−1 is not linear in the summand pi rai . We need such estimates in the estimate of the velocity.
Firstly, using the bounds in Section C.4.1 and (C.6), we can obtain the upper and the lower
bounds for Rij
∂xi ∂yj ρ
Rij = .
ρ
For i + j = 1 and k = 2, 3, we use the estimate in Section C.4.1 to obtain the bounds for
x y
R10 = S(x), R0,1 = S(x), (Rij )k .
|x|2 |x|2
In our verification, we need ∂xi ∂yj ρ−1 for i + j ≤ 3. A direct calculation yields
ρx R10 ρxx ρ2
∂x ρ−1 = − 2
=− , ∂xx ρ−1 = − 2 + 2 x3 = ρ−1 (−R20 + 2R10 2
),
ρ ρ ρ ρ
ρxy 2ρx ρy
∂xy ρ−1 = − + = ρ−1 (−R11 + 2R10 R01 ),
ρ ρ3
ρxxx 6ρxx ρx 6ρ3x
∂xxx ρ−1 = − 2 + − = ρ−1 (−R30 + 6R20 R10 − 6R10
3
),
ρ ρ3 ρ4
ρxxy 2ρxx ρy 4ρx ρxy
∂xxy ρ−1 = − 2 + + = ρ−1 (−R21 + 2R20 R01 + 4R10 R11 − 6R10
2
R01 ).
ρ ρ3 ρ3
140 JIAJIE CHEN AND THOMAS Y. HOU

Next, we estimate ∂xi ∂yj (∂xi ρ/ρ) for i ≤ 2, j = 0 or i = 0, j ≤ 2. Denote f = ∂xi ρ. Using a
direct computation, for D2 = ∂xi2 ∂yj2 with i2 + j2 = 1, we yield
f D2 f f D2 ρ
D2 = − = ρ−1 (D2 f − f Ri2 ,j2 ).
ρ ρ ρ2
For (i2 , j2 ) = (2, 0), (0, 2), denote i3 = i2 /2, j3 = j2 /2, D3 = ∂xi3 ∂yj3 .
We yield
f D2 f 2D3 f · D3 ρ 2 1 D32 f 2D3 f · D3 ρ D32 ρ 2(D3 ρ)2
D32 = 3 − + f D3 ( ) = − + f (− + )
ρ ρ ρ2 ρ ρ ρ2 ρ2 ρ3
= ρ−1 (D32 f − 2D3 f Ri3 ,j3 − f Ri2 ,j2 + 2f Ri23 ,j3 ),
D2 ρ 2
where we have used D32 ρ1 = D3 (− Dρ32ρ ) = − ρ32 + 2(Dρ33ρ) .
Since we have estimated ∂xi ∂yj ρ and Rij , we can bound these derivatives of D1 ρ/ρ using (C.5).

C.4.4. Improved estimates for ρ−1 near x = 0. For the special case a1 = −2, we can write
X
ρ(x) = r−2 pi rai +2 = r−2 ρ̃(x), ρ−1 = (x2 + y 2 )ρ̃(x)−1
i

To obtain a better estimate of ρ−1 , we use the fact that x2 + y 2 is a polynomial. Firstly, we can
obtain the bounds for ∂xi ∂yj ρ̃−1 . The bound for S0 = x2 + y 2 is trivial, e.g.,

(∂x S0 )γ = 2xγ , ∂xy S0 = 0, ∂xx S9 = ∂yy S0 = 2.

Then using (C.5)-(C.6), we can bound ρ−1 .

C.5. The mixed weight. For the second type of weights W = ρ1 (r)x−1/2 + ρ2 (r), we can
compute its derivatives and its upper and lower bounds using linearity and the Leibniz rule. For
example, we have
1
Wl = ρ1,l x−1/2
u + ρ2,l , (W −1 )u = (Wl )−1 , Wx = ∂x ρ1 x−1/2 − ρ1 x−3/2 + ∂x ρ2 .
2
To obtain the upper bound for ∂xi ∂yj W , we use the Leibniz rule:
X i (2k − 1)!! −1/2−k
i j
|∂x ∂y W | ≤ |∂xi−k ∂yj ρ1 | x + |∂xi ∂yj ρ2 |.
k 2k
k≤i

We need to bound ρ(r)/W (x, y) in the estimate of the integrals. Suppose that the lead-
ing and the last powers of ρ is a1 , an . The leading and the last terms of W are given by
pi rbi cos(β)−αi , αi ≥ 0.
W ≥ p1 rb1 , W ≥ p n r bn .
We estimate
ρ ρ
≤ C1 ra1 −b1 , ≤ C2 ran −bn ,
W W
for all x, y ∈ R+
2 . We apply the above estimates for x near 0 or x sufficiently large.

Appendix D. Estimate the derivatives of the velocity kernel and integrands

1
In this appendix, we estimate the derivatives of the kernel − 2π log |x| associated to the
⊥ −1
velocity u = ∇ (−∆) ω and its symmetrization (3.2). These estimates are used to estimate
the error terms in Lemmas 8.2, 8.4. We will perform an additional estimate for u with weight
ϕ(x) singular along x1 = 0 in Section D.5. Some additional derivations related to the estimate
of the velocity are given in Appendix D.6.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 141

D.1. Estimate the symmetrized kernel. In this section, we estimate the symmetrized kernel.
We develop several symmetrized estimates for harmonic functions. Before we introduce the
estimates, we have a simple 1D estimate, which is useful for later estimates.
Lemma D.1. We have
x4 4
|f (x)+f (−x)−2f (0)| ≤ x2 ||fxx ||L∞ [−x,x] , |f (x)+f (−x)−2f (0)−x2 fxx (0)| ≤ ||∂ f ||L∞ [−x,x] .
12 x
Proof. Denote G(x) = f (x) + f (−x). Clearly, G is even and
(D.1) G(0) = 2f (0), G0 (0) = 0, ∂x2 G(0) = 2fxx (0), ∂x3 G(0) = 0.
Using the Taylor expansion, we obtain
∂x2 G(0)x2 ∂ 3 G( 0)x3 ∂ 4 G( ξ)x4
G(x) = G(0) + G0 (0)x + + x + x ,
2 6 24
for some ξ ∈ [0, x]. Using (D.1), we get
x2 x4 x4
|G(x) − G(0) − G00 (x) | ≤ ||∂x4 G||L∞ [0,x] ≤ ||∂x2 f ||L∞ [−x,x] .
2 24 12
Plugging the identity (D.1) into the above estimate proves the second estimate in Lemma D.1.
The first estimate is simpler. 
The following lemma is useful for estimating the symmetrized kernel (3.2) and its derivatives.
Lemma D.2. Suppose that Qx = [−x1 , x1 ] × [−x2 , x2 ] and f ∈ C 4 (Qx ) is harmonic. Denote
G1 (1, x) , f (x1 , x2 ) + f (−x1 , x2 ) + f (x1 , −x2 ) + f (−x1 , −x2 ) − 4f (0, 0),
(D.2) G2 (1, x) , f (x1 , x2 ) − f (−x1 , x2 ) − f (x1 , −x2 ) + f (−x1 , −x2 ),
Ĝ1 (x) , 2x21 fxx (0, 0) + 2x22 fyy (0, 0), Ĝ2 (x) , 4x1 x2 fxy (0, 0).
We have
(D.3) |G1 (1, x)| ≤ 2|x|2 ||fxx ||L∞ (Qx ) , |∂xi G1 (1, x)| ≤ 4|xi | · ||fxx ||L∞ (Qx ) ,
(x41 + 6x21 x22 + x42 ) 4 |x|4 4
(D.4) |G1 (1, x) − Ĝ1 (x)| ≤ ||∂ f ||L∞ (Qx ) ≤ ||∂ f ||L∞ (Qx ) ,
6 3
1
(D.5) |G1 (1, x1 , 0) − Ĝ1 (x1 , 0)| ≤ x41 ||∂ 4 f ||L∞ (Qx ) ,
6 √
2 2 3 4 2 2 3 4
(D.6) |∂xi (G1 (1, x) − Ĝ1 (x))| ≤ (3x3−i xi + xi )||∂ f ||L∞ (Qx ) ≤ |x| ||∂ f ||L∞ (Qx ) ,
3 3
where ||∂ 4 f ||L∞ = max0≤i≤4 ||∂xi ∂yj f ||L∞ (Qx ) . For G2 , we have the following estimate
(D.7) |G2 (1, x)| ≤ 4x1 x2 ||fxy ||L∞ (Qx ) , |∂xi G2 (1, x)| ≤ 4|x3−i | · ||fxy ||L∞ (Qx ) ,
2
2x1 x2 |x|
(D.8) |G2 (1, x) − Ĝ2 (x)| ≤ ||∂ 4 f ||L∞ (Qx ) ,
3 √
2 2 3 4 2 2 3 4
(D.9) |∂xi (G2 (1, x) − Ĝ2 (x))| ≤ (3xi x3−i + x3−i )||∂ f ||L∞ (Qx ) ≤ |x| ||∂ f ||L∞ (Qx ) .
3 3
Proof. Recall Qx = [−x1 , x1 ] × [−x2 , x2 ]. Denote
Mij (x) = ||∂xi ∂yj f ||L∞ (Qx ) .
Using Lemma D.1, for any t ∈ [0, 1], we obtain
|f (tx1 , x2 ) + f (tx1 , −x2 ) − 2f (tx1 , 0)| ≤ M02 x22 , |f (x1 , 0) + f (−x1 , 0) − 2f (0, 0)| ≤ M20 x21 .
Since f is harmonic function, we have ∂xi+2 ∂yj f = −∂xi ∂yj+2 f and obtain Ai+2,j = Ai,j+2 . Taking
t = ±1 in the above estimate and using the triangle inequality, we prove
|G(1, x)| ≤ 2A20 (x)x21 + 2A02 x22 = 2A20 (x21 + x22 ) = 2A20 |x|2 ,
which is the first estimate in (D.3).
142 JIAJIE CHEN AND THOMAS Y. HOU

The second estimate in (D.3) is simple. We consider i = 1 without loss of generality. We get
|∂x1 G1 (1, x)| = |(∂1 f )(x1 , x2 )−(∂1 f )(−x1 , x2 )+(∂1 f )(x1 , −x2 )−(∂1 f )(−x1 , −x2 ))| ≤ 4x1 A20 (x).
For (D.4), using Lemma D.1, we yield
x42
|f (tx1 , x2 ) + f (tx1 , −x2 ) − 2f (tx1 , 0) − x22 (∂22 f )(tx1 , 0)| ≤ A04 (x) ,
12
(D.10) |∂22 f (x1 , 0) + ∂22 f (−x1 , 0) − 2∂22 f (0, 0)| ≤ x21 A2,2 (x),
x41
|f (x1 , 0) + f (−x1 , 0) − 2f (0) − x21 ∂12 f (0)| ≤ A40 ,
12
for t = ±1. Combining the above estimates and using the triangle inequality and A40 = A22 =
A04 , we prove the first estimate in (D.4). The second estimate follows from 2|x|4 − x41 − 6x21 x22 −
x42 = (x21 − x22 )2 ≥ 0.
Estimate (D.5) follows from (D.4) by taking x2 = 0.
For (D.6), we consider the estimate of ∂x1 . The other case is similar. Using
Z x1
∂1 f (x1 , s) − (∂1 f )(−x1 , s) = (∂12 )(t, s) + (∂12 )f (−t, s)dt,
0
we obtain
∂1 (G(1, x) − Ĝ1 (x)) = ∂1 f (x1 , x2 ) − (∂1 f )(−x1 , x2 ) + ∂1 f (x1 , −x2 ) − (∂1 f )(−x1 , −x2 ) − 4x1 ∂12 f (0)
Z x1  
= (∂12 f )(z, x2 ) + (∂12 f )(−z, x2 ) + ∂12 f (z, −x2 ) + (∂12 f )(−z, x2 ) − 4∂12 f (0) dz.
0
Applying (D.3), we yield
Z x1
2
|∂1 (G(1, x) − Ĝ1 (x))| ≤ 2(z 2 + x22 )dzA4,0 (x) = ( x31 + 2x1 x22 )A4,0 (x),
0 3
and complete the proof of the first estimate in (D.6). For the second estimate, we use the
AM-GM inequality to yield
1 1 2(3x22 + x21 ) + 4x21 3
(D.11) (3x22 x1 + x31 )2 = (3x22 + x21 )2 x21 = (3x22 + x21 )2 4x21 ≤ ( ) = 2|x|4 .
4 4 3
Taking a square root completes the estimate.
To estimate G2 in (D.2), we rewrite it as follows
Z x1 Z x2
G2 (1, x) − cĜ2 (x) = ∂12 f (y1 , y2 ) − c∂12 f (0)dy
−x1 −x2
Z x1 Z x2
= (∂12 f )(y1 , y2 ) + (∂12 f )(−y1 , y2 ) + (∂12 f )(y1 , −y2 ) + (∂12 f )(−y1 , y2 ) − 4c(∂12 f )(0)dy,
0 0
for c = 0, 1. The integrand has the same form as G1 in (D.2). For c = 0, using the above
decomposition, we prove
|G2 (1, x)| ≤ 4x1 x2 A11 .
When c = 1, using (D.6), we yield
Z x1 Z x2
2 2
|G2 (1, x) − Ĝ2 (x)| ≤ A40 2 |y|2 dy = A40 (x31 x2 + x1 x32 ) = A40 x1 x2 |x|2 .
0 0 3 3
To estimate the derivatives, we focus on ∂x1 . Using the above representation, we obtain
Z x2
∂x1 G2 (1, x) − cĜ2 (x) = ((∂12 f )(x1 , y2 ) + (∂12 f )(−x1 , y2 ))dy
0
Z x2
+ ((∂12 f )(x1 , −y2 ) + (∂12 f )(−x1 , y2 ) − 4c(∂12 f )(0))dy.
0
We apply the same estimates to the integrands with c = 0, 1 and yield
Z x2
2
|∂x1 G2 (1, x)| ≤ 4x2 A11 , |G2 (1, x) − Ĝ2 (x)| ≤ A40 2 (x21 + y22 )dy2 = A40 (2x21 x2 + x32 ).
0 3
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 143

The second inequality in (D.9) follows from (D.11). The above estimates imply (D.7)-(D.9).

Recall the kernels associated with ∇u, u in (3.1). These kernels are the derivatives of the
1
Green function − 2π log |x| and are harmonic away from 0. We have the following estimates for
their derivatives.
Lemma D.3. Denote r = (x2 + y 2 )1/2 and f (x, y) = log r. For any i, j ≥ 0 with i + j ≥ 1, we
have
|∂xi ∂yj f (x, y)| ≤ (i + j − 1)! · r−i−j .
As a result, for K1 (y) = − 12 ∂12 f (y), K2 (y) = − 12 ∂12 f (y), we have
1 3 60 2520
|Ki | ≤ , |∂yj1 ∂y2−j Ki | ≤ , |∂yj1 ∂y4−j Ki | ≤ , |∂yj1 ∂y6−j Ki | ≤ .
2|y|2 2
|y|4 2
|y|6 2
|y|8

Proof. Consider the polar coordinate β = arctan(y/x), r = (x2 + y 2 )1/2 . We use induction on
n = i + j to prove
(D.12) ∂xi ∂yj f = (n − 1)! cos(nβ − βij )r−n ,

for some constant βij . Denote β = arctan xy , r = x2 + y 2 . We have the formula

p

sin β cos β
(D.13) ∂x g = (cos β∂r − ∂β )g, ∂y g = (sin β∂r + ∂β )g.
r r
Firstly, for n = 1, a direct calculation yields
x cos β y sin β cos(β − π/2)
∂x f = = , ∂y f = = = .
r2 r r2 r r
Suppose that (D.12) holds for any i, j with i + j = n and n ≥ 1. Now, since
∂x ∂xi ∂yj f = (n − 1)!∂x (cos(nβ − βij )r−n )
= (n − 1)!(−n cos β cos(nβ − βij )r−n−1 + n sin β sin(nβ − βij )r−n−1 )
= n!(− cos(nβ − βij + β)r−n−1 ) = n! cos((n + 1)β − βij − π)r−n−1 ,

using a similar computation and sin(x) = cos(x − π/2), we can obtain that ∂y ∂xi ∂yj f has the
form (D.12). Using induction, we prove (D.12). The desired estimate follows from (D.12). 
Using the above two Lemmas, we can estimate the error in the discretization of the kernels
K(x, y) in both x and y directions.
Estimate the kernels in the far field. Recall the symmetrized kernel in (3.2). The integrals
in ux , uy can be written as
Z Z
1 1
ux (x) = − K1,s (x, y)ω(y)dy, uy = K2,s (x, y)ω(y)dy.
π R++
2
π R++
2

Denote Qx = [−x1 , x1 ] × [−x2 , x2 ] and

(D.14) f (x1 , x2 ) = K1 (x1 − y1 , x2 − y2 ), Den(x, y) = min |y − z|2 .
z∈Qx

It is not difficult to obtain that for x, y ∈ we have R++

2 ,
X X
(D.15) Den(x, y) = min |yi − zi |2 = (max(yi − xi , 0))2 .
|zi |≤xi
i=1,2 i=1,2

Since K1 (y1 , y2 ) is odd in y1 , y2 , we have K1,s = f (x1 , x2 )+f (−x1 , x2 )+f (x1 , −x2 )+f (−x1 , x2 ).
A direct computation yields the following formulas
y1 y2 12y1 y2 (y12 − y22 )
f (0, 0) = K1 (−y1 , −y2 ) = , fxx (0, 0) = ∂y12 K1 (y) = .
|y|4 |y|8
144 JIAJIE CHEN AND THOMAS Y. HOU

Using the fact that K1 is a harmonic function and then applying Lemmas D.2 and D.3 with
the above f , we obtain

4y1 y2 6|x|2 12x1

|K1,s (x, y) − | ≤ , |∂x1 K1,s (x, y)| ≤ ,
|y|4 Den2 (x, y) Den2 (x, y)
√
4y1 y2 2 2 2 10 2|x|4
|K1,s (x, y) − − 2(x 1 − x )∂
2 y1 1 K (y)| ≤ ,
|y|4 Den3 (x, y)
(D.16) √
40 2|x|3
 
4y1 y2 2 2 2
∂x1 K1,s (x, y) − − 2(x1 − x2 )∂y1 K1 (y) ≤ ,
|y|4 Den3 (x, y)
4y1 y2 2 2 10x41
|K1,s (x1 , 0, y) − − 2x ∂
1 y1 K 1 (y)| ≤ .
|y|4 Den3 (x, y)

We use the following two formulas to estimate the discretization error

√ √
2 2520|x|4
 
2 4y1 y2 2 2 2 2 4 6
∂yi K1,s (x, y) − − 2(x 1 − x )∂
2 y1 K1 (y) ≤ |x| |∂ y K| ≤ · ,
|y|4 6 6 Den4 (x, y)
√ √
2 2 2520|x|3
 
2 4y1 y2 2 2 2 2 2 3 6
∂xi ∂yi K1,s (x, y) − − 2(x1 − x2 )∂y1 K1 (y) ≤ |x| |∂y K| ≤ · .
|y|4 3 3 Den4 (x, y)

Similarly, we can obtain the following estimates for K2,s

12x1 x2 12x3−i
|K2,s | ≤ 2 , |∂xi K2,s (x, y)| ≤ ,
Den (x, y) Den2 (x, y)
12y1 y2 (y12 − y22 ) 40x1 x2 |x|2
(D.17) K2,s − 4x1 x2 · ≤ ,
|y|8
Den3 (x, y)
√
12y1 y2 (y12 − y22 ) 40 2|x|3
 
∂xi K2,s − 4x1 x2 · ≤ .
|y|8 Den3 (x, y)
y1 y2
The decay estimates of the symmetrized kernel for u, |y|2 , |y|2 can be established similarly.

D.2. Estimate of the stream function in 3D Euler.

D.2.1. Proof of Lemma 6.4.

Proof. Recall r = 1 − Cl y in (6.9) and ω̃(r, z) = ω θ (r, z)/r in (6.6). Since the support size
satisfies Cl (τ )S(τ ) < 1/4, within the support of ω θ (r, z), we have r ≥ 1/2. Hence, |ω̃|  |ω θ |
and |rω θ (r, z)| . |ω̃(r, z)|. We can apply Lemma 6.1 and (6.10) to get
Z  
−2
|φ(x, y)| . Cω Cl |ω̃(r1 , z1 )| 1 + | log((r1 − (1 − Cl y))2 + (z1 − Cl x)2 )| dr1 dz1
Z  
= Cω |ω̃(1 − Cl y1 , Cl x1 )| 1 + | log(Cl2 ((y1 − y)2 + (x1 − x)2 )| dy1 dx1 ,

where we have used Lemma 6.1 and r1 ≤ 1 in the first inequality, and used change of variables
r1 = 1 − Cl y1 , z1 = Cl x1 in the second identity. Since Cω |ω̃(1 − Cl y1 , Cl x1 )| = |ω(x1 , y1 )| and
|(x, y)| > 2S, within the support of ω, we get |(x, y) − (x1 , y1 )|2  |(x, y)|2 . It follows
Z Z
|φ(x, y)| . (1 + | log(Cl2 |(x, y)|2 )|) |ω(z)|dz . (1 + | log(Cl |(x, y)|)|)||ω(1 + |z|)β ||∞ (1 + |z|)−β dz
|z|≤S
β 2−β
. ||ω(1 + |z|) ||∞ S (1 + | log(Cl |(x, y)|)|).

This proves the desired result. 

 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 145

D.2.2. Proof of other estimates in Lemma 6.5. In this subsection, we prove the second and the
third estimate in Lemma 6.5. Using the Green function, we have
Z Z
1 4 y1 y2
φ=− log |x − y|W (y)dy, φxy (0) = − W (y)dy,
2π π R++
2
|y|4
where W is the odd extension of ω from R+ 2 to R2 . Clearly, we only need to prove the estimate
for the derivatives. Without loss of generality, we estimate ∂1 (φ − x1 x2 ∂12 φ(0)).
Similar to the proof of the first estimate in Lemma 6.5, we use the partition
(D.18) Q1 = {y : |y| ≥ 2|x|}, Q2 = {y : |y − x| ≤ |x|/2}, Q3 = (Q1 ∪ Q2 )c .
Denote M = ||ω(|x|−α + |x|β )||∞ . We have
|ω| ≤ M min(|x|α , |x|−β ).
In Q1 , we combine the estimate of φ and −ψxy (0)xy. Symmetrizing the kernel, we get
Z
1
φ − φxy (0)xy = − K(x, y)W (y)dy,
2π
2y1 y2 x1 x2
K(x, y) = log |x − y| + log |x + y| − log |(x1 − y1 , x2 + y2 )| − log |(x1 + y1 , x2 − y2 )| − .
|y|4
Applying (D.9) in Lemma D.2 with f (x) = log |x − y| and Lemma D.3, for |y| ≥ 2|x|, we get
|x|3
|∂x1 K(x, y)| . .
|y|4
For α < 2, we obtain
Z Z
∂x1 K(x, y)W (y)dy . M |x|3 |y|−4 min(|y|α , |y|−β )dy
(D.19) Q1 |y|≥2|x|

= M |x|3 min(|x|−2+α , |x|−2−β ) = M min(|x|1+α , |x|1−β ).

Note that if α = 2, the integral for small y becomes |x|≤|y|≤1 |y|−2 dy, which leads to a log
R

factor.
In Q2 , Q3 , we estimate two integrals separately. For φ in Q2 , we have |x − y| ≤ |x|/2 and
|x|  |y| in |x − y| ≤ |x|/2. Thus, we get
Z Z
−β
α
|∂x1 log |x − y|W (y)|dy . M min(|x| , |x| ) |x − y|−1 dy
|x−y|≤|x|/2 |x−y|≤|x|/2
α −β
. M min(|x| , |x| )|x|,
In Q3 , we get |x|/2 ≤ |x − y| ≤ 3|x|, |y| ≤ 2|x|. It follows
Z Z
|∂x1 log |x − y|W (y)|dy . M |x|−1 min(|y|α , |y|−β )dy
Q3 |y|≤2|x|

. M |x|−1 min(|x|2+α , |x|2−β ) . M min(|x|1+α , |x|1−β ).

For φxy (0)x1 x2 , we have
Z Z
x1 x2 y1 y2
∂x1 W (y) dy . M |x| |y|−2 min(|y|α , |y|−β ) = M |x| min(|x|α , |x|−β ).
|y|≤2|x| |y|4 |y|≤2|x|

Combining the above estimates, we prove the second estimate in Lemma 6.5.
For the last estimate in Lemma 6.5, if α ∈ (2, 3], the integrand in (D.19) is integrable near
|x| = 0. We yield Z
| ∂x1 K(x, y)W (y)dy| . M |x|3 .
Q1
For other terms, clearly, for α > 2, β > 0, we have
min(|x|2+α , |x|2−β ) . |x|3 .
The desired estimate follows.
146 JIAJIE CHEN AND THOMAS Y. HOU

D.2.3. Proof of Lemma 6.10. The bound by ||Ω||X follows from embedding. We focus on the
bound by |x|2 ||Ω||X and assume that |x| ≤ 1. We consider the estimate for ∂12 Ψ − Ψ12 (0).
Firstly, we have
Z Z
1 z1 z2 4
∂12 Ψ = − K(x − y)Ω(y)dy, K(z) = , ∂12 Ψ(0) = − K(y)W (y)dy,
π |z|4 π R++
2

where we extend Ω from R+ ++

2 to R2 by natural odd extension. Note that from (C.1)-(C.3)
max(ϕ1 , ϕg,1 ) & |x|−α , |x|β , α = 2.9, β = 1/16, ψ1  |x|−2 , for |x| . 1.
Following the standard partition (D.18), the symmetrization argument similar to that in the
proof of Lemma 6.5 or Section D.2.2, and estimate in Lemma D.2, we obtain
Z Z
K(x − y)W (y) − 4 K(y)W (y)dy . |x|2 ||Ω||X .
|y|≥2|x| |y|≥2|x|

For |y| ≤ 2|x|, we have

Z Z
K(y)Ω(y)dy . min(|y|α , |y|−β )|y|−2 dy||Ω||X . |x|2 ||Ω||X
|y|≤2|x| |y|.2|x|

In region Q3 ⊂ {y : |x − y|  |x|}, we have

Z Z Z
| K(x − y)Ω(y)dy| . |x|−2 |Ω(y)|dy . |x|−2 min(|y|α , |y|−β )dy||Ω||X
Q3 |y|≤3|x| |y|≤3|x|

. ||Ω||X min(|x|α , |x|−β ).

In the singular region Q2 = {y : |x − y| ≤ |x|/2}, for any |s|, |t| ≤ |x|/2, and |x| ≤ 1, we have
Ω(x + s) − Ω(x + t) = (Ωψ1 ψ1−1 )(x + s) − (Ωψ1 ψ1−1 )(x + t)
≤|Ωψ1 (x + s) − Ωψ1 (x + t)|ψ1−1 (x + s) + |ψ1−1 (x + s) − ψ1−1 (x + t)| · |Ωψ1 (x + t)| , I1 + I2 .

Since |∇ψ −1 (x)| . |x|, for |s|, |t| . |x|/2, we get

|ψ1−1 (x + s) − ψ1−1 (x + t)| . |x||s − t|.
Using the above estimate for the weight, we yield
|I1 | . ||Ω||X |x|2 |s − t|1/2 , I2 . |s − t||x|ψ1 (x + t)/ϕ1 (x + t) . |s − t||x||x|α−2 = |s − t||x|α−1 .
Using the symmetry of the kernel, we get
Z Z
| K(x − y)W (y)dy| = | K(s)W (x + s)ds|
|x−y|≤|x|/2 |s|≤|x|/2
Z Z
.|x|2 |s|−3/2 ds + |x|α−1 |s|−1 ds . |x|5/2 + |x|α . |x|2 .
|s|≤|x|/2 |s|≤|x|/2

Note that for small |x|, we can improve the above estimate by optimizing the window for the
singular region. We complete the estimate of the second inequality in Lemma 6.10.
To prove the first estimate in Lemma 6.10. For ω ∈ X, using (2.79), we have
24y1 y2 (y12 − y22 )
Z
4
∂1112 Φ0 = Ω(y)dy.
π R++
2
|y|8
The proof is completely similar. We use the above partiton of the domains. In Q1 , we use
the symmetrizing estimates (D.16), (D.17). In Q2 , Q3 , we estimate the integrals separately, and
use the above estimates and
24y1 y2 (y12 − y22 )
Z Z
|x| 2
Ω(y)dy . |x|2
|y|−4 min(|y|α , |y|−β )dy||Ω||X . min(|x|α , |x|−β ).
|y|.|x| |y|8 |y|.|x|
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 147

D.3. Piecewise L∞ estimate of derivatives of the Green function. In this section, we

develop sharp L∞ estimates of the derivatives of the Green function G(x) = − 2π 1
log |x| and
their linear combinations in a small domain [a, b] × [c, d]. They will be used in Lemmas 8.2, 8.4
to estimate the error, especially near the singularity of the kernel. We remark that the linear
combinations of ∂1i ∂2j G can be quite complicated. If we simply use the triangle inequality to
estimate it, we can overestimate some terms with cancellation significantly, especially near the
singularity of G. These sharp estimates are useful for reducing the estimate of the error term in
Lemmas 8.2, 8.4 without choosing very small mesh, which can lead to large computational cost.
D.3.1. Coefficients of the derivatives of the Green function. To simplify the notation, we drop
1 1 i j
π from G and denote fp = − 2 log |x|. Firstly, we derive the formulas of ∂1 ∂2 fp . Due to
homogenity, we assume
i j
P
k l i+j=k+l cij x1 x2
∂x1 ∂x2 fp = .
|x|2(k+l)
Next, we derive the recursive formula for cij . Using induction, we can obtain
P i+1 j
i+j=k+l cij ix1 x2 2(k + l)x1 X
k+1 l
∂x1 ∂x2 fp = 2(k+l)
− 2(k+l+1) cij xi1 xj2
|x| |x| i+j=k+l
1 X j i−1 j+2 j
= ( cij ixi+1
1 x2 + cij ix1 x2 − 2(k + l)cij xi+1
1 x2 )
|x|2(k+l+1) i+j=k+l
1 X j
= ( (cij i + ci+2,j−2 (i + 2) − 2(k + l)cij )xi+1
1 x2 ).
|x|2(k+l+1) i+j=k+l

Therefore, we obtain the recursive formula

ci+1,j = icij + (i + 2)ci+2,j−2 − 2(k + l)cij ,
for all i + j = k + l, or equivalently,
ci,j = (i − 1)ci−1,j − 2(k + l)ci−1,j + (i + 1)ci+1,j−2 ,
for all i + j = k + l + 1. Similarly, for ∂x2 , we yield
ci,j = (j − 1)ci,j−1 − 2(k + l)ci,j−1 + (j + 1)ci−2,j+1 ,
for all i + j = k + l + 1.
D.3.2. Estimates of rational functions. We use the above formulas to develop sharp estimates
of the derivatives of fp and their linear combinations in a small grid cell [y1l , y1u ] × [y2l , y2u ].
For k1 < k2 and S ⊂ {(i, j) : i + j = k}, we estimate
i j
P
(i,j)∈S cij y1 y2
IS , .
|y|k2
Denote i1 = mini∈S i, j1 = minj∈S j. We yield
P i−i1 j−j1
y1i1 y2j1 (i,j)∈S cij y1 y2
IS = i1 +j1 .
|y| |y|k2 −i1 −j1
We further introduce
X X
i−i1 j−j1 i−i1 j−j1
P , c+
ij y1 y2 , Q , c−
ij y1 y2 .
(i,j)∈S (i,j)∈S

Clearly, P and Q are monotone increasing in y1 ≥ 0 and y2 ≥ 0. As a result, on D =

[y1l , y1u ] × [y2l , y2u ] ⊂ R++
2 , we yield

max(Pu − Ql , Qu − Pl ) y1i1 y2j1

|I| ≤ max
|y|kl 2 −i1 −j1 y∈Ω |y|i1 +j1
max(Pu − Ql , Qu − Pl ) y1u y2u
= ( )i1 ( 2 )j1 ,
|y|kl 2 −i1 −j1 2 + y 2 |1/2
|y1u 2l |y 1l + 2 |1/2
y2u
148 JIAJIE CHEN AND THOMAS Y. HOU

where we have used the fact that yi /|y| is increasing in yi to obtain its upper bound.

D.4. Improved estimate of the higher order derivatives of the integrands. In the
Hölder estimate, we need to estimate the derivatives of the integrands (8.24), (8.25), (8.18),
which take the form
K C (x, y)(p(x) − p(y)) + K N C p(x),
for some weight p and kernels K C , K N C . Using the estimates of the kernels in Sections D.1, D.3
and the weights in Section C.3, the Leibniz rule, and the triangle inequality, we can estimate
the derivative of the integrands. Howover, such an estimate can lead to significant overestimates
near the singularity of the integrand. We use the estimates in (D.3) to handle the cancellations
among different terms and obtain improved estimates for the integrand and its derivatives near
the singularity:

(D.20) T00 (x, y) , K(y − x)(p(x) − p(y)), ∂xi T00 (x, y).

We choose weight p(x) that is even in x and y. The basic idea is to perform a Taylor expansion
on p(x) − p(y) and obtain the factor |x − y|, which cancels one order of singularity from K(x, y).
We use the formulas in Section D.3 to collect the terms with the same singularity and exploit
the cancellation.

D.4.1. Y-discretization. In the Y-discretization of the integral, we need to estimate the y−derivatives
of the integrand (D.20). For a, b = 1, 2, denote
x+y
(D.21) D1 = ∂a , D2 = ∂b , xm = .
2
Next, we compute ∂yjb ∂xi a T00 . The reader should be careful about the sign. Note that

∂xa (K(y − x)) = −(∂a K)(y − x) = −D1 K(y − x).

Using the Leibniz rule, we get

∂y2b ∂xa T00 = ∂y2b (−D1 K(p(x) − p(y)) + K · D1 p(x)) = ∂y2b (D1 K · (p(y) − p(x)) + K · D1 p(x))
= D22 D1 K · (p(y) − p(x)) + 2D2 D1 K · D2 p(y) + D1 K · D22 p(y) + D22 K · D1 p(x).

We use Talyor expansion at x = xm and write

(D.22) p(y)−p(x) = (y−x)·∇p(xm )+pm,2,err , ∂i p(z) = ∂i p(xm )+(∂i p(z)−∂i p(xm )), z = x, y,

and then combine the terms involving ∇p to get

(D.23)
X  
∂y2b ∂xa T00 = D22 D1 K · (yi − xi ) + 1D2 =∂i 2D2 D1 K + 1D1 =∂i D22 K · ∂xi p(xm ) + D22 D1 K · pm,2,err
i=1,2

+ 2D2 D1 K · (D2 p(y) − D2 p(xm )) + D22 K · (D1 p(x) − D1 p(xm )) + D1 K · D22 p(y)
 X 
, Ii · ∂xi p(xm ) + II,
i=1,2

Ii , D22 D1 K · (yi − xi ) + 1D2 =∂i 2D2 D1 K + 1D1 =∂i D22 K,

where ∂1i ∂2j K is evaluated at y − x, and II consists of the last four terms in the second equation.
The first term is the most singular term. We combine the most singular terms to exploit the
cancellation and improve the estimates. We estimate the kernels

(D.24) Kmix (D1 , D2 , i, s)(z1 , z2 ) , D22 D1 K(z)zi + 1D2 =∂i 2D2 D1 K(z) + s1D1 =∂i D22 K(z),

with s = ±1 and D1 , D2 ∈ {∂1 , ∂2 }. Then we can bound ∂y2b ∂xa T00 using the triangle inequality.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 149

D.4.2. The second singular term. For x = (x1 , x2 ) close to the y−axis or the x-axis, since we
have symmetrized the integral, we have another singular term in the integrand
T01 , K(y1 − x1 , y2 + x2 )(p(x) − p(y)), or T10 , K(y1 + x1 , y2 − x2 )(p(x) − p(y)).
We have the first term if x2 < x1 and x2 close to 0, and the second term if x1 < x2 and x1
close to 0. We label the former case with side = 1 and the latter side = 2. See the right figure
in Figure 10 for an illustration of the first case. The T01 term is supported in the blue region
R(x, k, S). Denote
(D.25) (s1 , s2 ) = (1, −1) if side = 1, (s1 , s2 ) = (−1, 1) if side = 2.
Case I. If (D1 , side) = (∂1 , 1) or (∂2 , 2), we obtain
∂xj K(y1 − s1 x1 , y2 − s2 x2 ) = −∂yj K(y1 − s1 x1 , y2 − s2 x2 ),
for (j, s1 , s2 ) = (1, 1, −1) or (2, −1, 1). The computations for ∂y2b ∂x1 T01 , ∂y2b ∂x2 T10 are the same
as (D.23) with K and its derivatives evaluating at z = (y1 − s1 x1 , y2 − s2 x2 ).
We estimate II in (D.23) directly using the triangle inequality and the bounds for ∂1i ∂2j K in
Section D.1, D.3 and p in Section C.3. For Ii in (D.23) in the most singular term, if i = side,
from definition (D.25), we get
si = 1, s3−i = −1, zi = yi − si xi = yi − xi , z3−i = y3−i + x3−i .
Therefore, it follows
Ii = D22 D1 K(z) · (yi − xi ) + 1D2 =∂i 2D2 D1 K(z) + 1D1 =∂i D22 K(z) = Kmix (D1 , D2 , i, 1)(z),
where Kmix is defined in (D.24). If i 6= side, we have zi = yi + xi ≥ |yi − xi |, z3−i = y3−i − x3−i .
We simply bound the summand using the triangle inequality
|Ii | ≤ |D22 D1 K(z)| · |yi − xi | + 1D2 =∂i 2|D2 D1 K(z)| + 1D1 =∂i |D22 K(z)|.
Case II. If (D1 , side) = (∂1 , 2) or (∂2 , 1), we obtain
∂xj K(y1 − s1 x1 , y2 − s2 x2 ) = (∂yj K)(y1 − s1 x1 , y2 − s2 x2 ),
for (j, s1 , s2 ) = (1, −1, 1) or (2, 1, −1). Recall the definitions of D1 , D2 (D.21). Using the above
identity, we yield
∂y2b ∂xa T = ∂y2b (D1 K · (p(x) − p(y)) + K · D1 p) = −(∂y2b (D1 K · (p(y) − p(x)) − K · D1 p)),
for T = T01 or T10 . Using an expansion similar to that in (D.23), (D.22), we get
(D.26) X
−∂y2b ∂xa T = (D22 D1 K · (yi − xi ) + 1D2 =∂i 2D2 D1 K − 1D1 =∂i D22 K) · ∂xi p(xm ) + D22 D1 K · pm,2,err
i=1,2

+ 2D2 D1 K · (D2 p(y) − D2 p(xm )) − D22 K · (D1 p(x) − D1 p(xm )) + D1 K · D22 p(y)
 X 
, Ii · ∂xi p(xm ) + II,
i=1,2

Ii , D22 D1 K · (yi − xi ) + 1D2 =∂i 2D2 D1 K − 1D1 =∂i D22 K,

where ∂1i ∂2j K is evaluted at z = (y1 − s1 x1 , y2 − s2 x2 ). We bound II using triangle inequality

and the bounds for K, its derivatives, and p in Sections D.1, D.3, and C.3.
For Ii , if i = side, from (D.25), we get si = 1 and zi = yi − si xi = yi − xi . Hence, we get
Ii = D22 D1 K · (yi − xi ) + 1D2 =∂i 2D2 D1 K − 1D1 =∂i D22 K = Kmix (D1 , D2 , i, −1)(z),
where Kmix is defined in (D.24).
If i 6= side and D1 = D2 = ∂i , we have zi = yi − si xi = yi + xi and get a cancellation between
D2 D1 K and D22 K and yield
|Ii | = |D22 D1 K · (yi − xi ) + D2 D1 K| ≤ |D22 D1 K| · |yi − xi | + |D2 D1 K|.
Otherwise, we simply bound each term in Ii using the triangle inequality.
150 JIAJIE CHEN AND THOMAS Y. HOU

s21 −s22
D.4.3. X-discretization. For K(s) = c s|s|
1 s2
4 ,c |s|4 , we have K(s) = K(−s). Denote
T = K(y − x)(p(x) − p(y)) = K(x − y)(p(x) − p(y)).
In this section, we compute ∂xi b ∂xj a T . Using the Taylor expansion at x
p(x) − p(y) = (x − y) · ∇p(x) + px,2,err ,
and calculations similar to those in Section D.4.1, we get
(D.27)
∂x2b ∂xa T = ∂x2b (D1 K · (p(x) − p(y)) + KD1 p(x)) = D22 D1 K · (p(x) − p(y)) + 2D1 D2 K · D2 p(x)
+ D1 K · D22 p(x) + D22 K · D1 p(x) + 2D2 K · D1 D2 p(x) + K · D1 D22 p(x)
X
= (D22 D1 K · (xi − yi ) + 1D2 =∂i 2D1 D2 K + 1D1 =∂i D22 K)∂i p(x) + D22 D1 K · px,2,err
i=1,2
 X 
+ D1 K · D22 p(x) + 2D2 K · D1 D2 p(x) + K · D1 D22 p(x) , Ii · ∂i p(x) + II,
i=1,2

Ii , D22 D1 K · (xi − yi ) + 1D2 =∂i 2D1 D2 K + 1D1 =∂i D22 K,

where II consists of the last four terms in the third equation, K and its derivatives are evaluated
at x − y. Since D1 , D2 = ∂xi , we get
Ii = D22 D1 K · (xi − yi ) + 1D2 =∂i 2D1 D2 K + 1D1 =∂i D22 K = Kmix (D1 , D2 , i, 1)(x − y),
where Kmix is defined in (D.24). We use the bound for Kmix , ∂1i ∂2j K and p to estimate D22 D1 T .

D.4.4. The second singular term. Similar to Section D.4.2, we have the second singular term for
x close to the x-axis or y-axis
T01 , K(x1 − y1 , x2 + y2 )(p(x) − p(y)), T10 , K(x1 + y1 , x2 − y2 )(p(x) − p(y)).
We have the former if x2 < x1 and x2 close to 0, and the latter if x1 < x2 and x1 close to 0.
Using the definition of side, s1 , s2 from Section D.4.2 and (D.25), we get
∂xa K(x1 − y1 s1 , x2 − y2 s2 ) = (D1 K)(x1 − y1 s1 , x2 − y2 s2 ).
Then the computations of D22 D1 T are the same as those in (D.27) with ∂1i ∂2j K evaluated at
z = (x1 − s1 y1 , x2 − s2 y2 ). We bound II in (D.27) directly using the triangle inequality and
the bounds for ∂1i ∂2j K and p. For Ii in (D.27), if i = side, from (D.25), we get si and zi =
xi − si yi = xi − yi . It follows
Ii = D22 D1 K · zi + 1D2 =∂i 2D1 D2 K + 1D1 =∂i D22 K = Kmix (D1 , D2 , i, 1)(z).
If i 6= side, we have zi = xi + yi > |xi − yi |. We bound each term in Ii separately by following
the previous argument.

D.5. Estimate of u(x) for small x1 . In the energy estimate, we need to estimate u(x)ϕ(x)
with weight ϕ singular along the line x1 = 0. We use the property that u vanishes on x1 = 0 to
establish such estimate.
By definition and symmetrizing the kernel using the odd symmetry of ω, we have
1
Z x −y x2 − y2  1
Z
2 2
u(x, y) = − ω(y)dy. = K(x, y)dy,
2π y1 ≥0 |x − y|2 (x1 + y1 )2 + (x2 − y2 )2 π y1 ≥0
where
(D.28)
1 x2 − y2 x2 − y2 2(x2 − y2 )y1
K= ( − ) = x1 · , x1 Kdu .
2 |x − y|2 (x1 + y1 )2 + (x2 − y2 )2 |x − y|2 ((x1 + y1 )2 + (x2 − y2 )2 )
Using a rescaling argument, for x = λx̂, y = λŷ, we have
Z Z Z
λ λ
u= K(x̂, ŷ)ωλ (ŷ)dŷ = ( + )K(x̂, ŷ)ωλ (ŷ)dŷ , I + II,
π y1 ≥0 π y1 ≥0,y∈S
/ y1 ≥0,y∈S
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 151

where S is the singular region adapted to x̂. For I, we further rewrite it and estimate it as
follows
Z Z
λ λ
|I| = x̂1 Kdx (x̂, ŷ)ωλ (ŷ)dy ≤ x̂1 ||ωϕ||L∞ |Kdx (x̂, ŷ)|ϕ−1
λ (ŷ)dŷ.
π y1 ≥0,y ∈S
/ π y1 ≥0,y ∈S
/

Since the integral is not singular, we can use the previous method to discretize the integral and
obtain its tight bound.
Large x̂1 . For the second part, if x̂1 ≥ xl > 0 away from 0, we have
1 1
Kdu (x̂, ŷ) . ,
xl |x̂ − y|
which is integrable near the singularity x̂. We can apply the similar method to estimate the
singular integral Z
λ
|II| ≤ x̂1 |Kdu (x̂, ŷ)|ϕ−1
λ (ŷ)dŷ.
π y1 ≥0,y∈S
Small x̂1 . The difficulty is to estimate II for small x. It is not difficult to obtain that
λ
(D.29) |II| . ||ωλ ||L∞ (S) x̂1 | log(x̂1 )|.
π
Thus we cannot bound II by C x̂1 for some constant C uniformly for small x̂1 . Denote by
(D.30) Ssym = [0, x̂1 + kh] × [x̂2 − kh, x̂2 + kh], ŷ = x̂ + x̂1 s.
Then ŷ ∈ Ssym is equivalent to
x̂1
s ∈ x1−1 (Ssym − x̂) = x−1
1 ([−x̂1 , kh] × [−kh, kh]) = [−1, B
−1
] × [B −1 , B −1 ] , RB , B= .
kh
We further decompose II as follows
Z Z
λ λ
II = ( + )Kdu (x̂, ŷ) , (II1 + II2 ).
π y1 ≥0,y∈S\Ssym y1 ≥0,S∈Ssym π
For II1 , the integral is not singular and we apply the previous L∞ estimate. For II2 , using
a change of variable (D.30), we derive
Z
II2 = Kdu (x̂, x̂ + x̂1 s)x̂21 ωλ (x̂ + x̂1 s)ds.
s∈RB
Note that
ŷ − x̂ = x̂1 s, ŷ + x̂ = x̂1 (2 + s1 , s2 ).
By definition (D.28), we get
2x̂1 s2 · (x̂1 + x̂1 s1 ) 2(s1 + 1)s2
Kdu (x̂, x̂ + x̂1 s)x̂21 = − x̂2 = − 2 , −Ks (s).
x̂21 |s|2 · x̂21 ((s1 + 2)2 + s22 ) 1 |s| ((s1 + 2)2 + s22 )
It follows Z
II2 = − Ks (s)ωλ (x̂ + x̂1 s)ds.
RB
Since Ks (s) is symmetric in s2 , we derive
n
|II2 | ≤||ωϕ||∞ ( max ϕ−1
λ (x̂ + z) + max ϕ−1
λ )J1 (B)
z∈[−x̂1 ,0]×[0,kh] z∈[−x̂1 ,0]×[−kh,0]
o
+( max ϕ−1
λ + max ϕ−1
λ )J2 (B) ,
z∈[0,kh]×[0,kh] s∈[0,kh]×[−kh,0]

where
Z Z Z
J1 (B) = Ks (s)ds = Ks (s)ds, J2 (B) = Ks (s)ds.
[−1,0]×[0,1/B] [0,1]×[0,1/B] [0,1/B]2

The explicit formula of Ji can be obtained, and obviously Ji is decreasing in B. Note that J1 (B)
is bounded, but J2 (B) . 1 + log(B) . 1 + | log x̂1 |, which relates to the estimate (D.29).
D.6. Additional derivations.
152 JIAJIE CHEN AND THOMAS Y. HOU

D.6.1. Estimate of the log-Lipschitz integral. In this section, we derive the coefficient in the
estimate of ∂x2 I5,4 (x) (8.45), (8.46). For I5,4 , we further decompose it as follows
Z Z 
I5,4 = + K(x − y)(ψ(x) − ψ(y)W (y)dy , I5,4,1 + I5,4,2 .
R(k2 )\Rs (k2 ) Rs (k2 )\Rs (a)

The first term is nonsingular and its derivative can be estimated using the method in Sections
8.1.6-8.1.8. For I5,4,2 , using the second order Taylor expansion to ψ(x) − ψ(y) centered at x, we
have
∂x2 (K(x − y)ψ(x) − ψ(y)) = (∂2 K)(x − y)ψ(x) − ψ(y)) + K(x − y)∂2 ψ(x)
=(∂2 K(x − y)(x2 − y2 ) + K(x − y))∂2 ψ(x) + ∂2 K(x − y)(x1 − y1 )∂1 ψ(x) + RK ,
where the remainder RK coming from the higher order term in the Taylor expansion satisfies
X
|RK | ≤ ||∂xi ∂yj ψ||L∞ (Q) |x1 − y1 |i |x2 − y2 |j cij ,
i+j=2

where Q = Bi1 j1 (hx ) + [−k2 h, k2 h]2 and c20 = c02 = 21 , c11 = 1. It follows
X
|∂x2 I5,4,2 | ≤ ||ωϕ||∞ Scoeij (x) · fij (a, b),
0≤i≤1,0≤j≤i+1

where the coefficients Scoeij (x) depend on the weight ψ, ϕ, and fij (a, b) is the upper bound of
the integral
Z
(D.31) |∂2 K(y) · y1i y2j + 1(i,j)=(0,1) K(y)|dy ≤ fij (a, b).
[−k2 ,k2 ]2 \[−a,a]2

For example, Scoe01 comes from the following estimate for I5,4,2
Z
|(∂2 K(x − y)(x2 − y2 ) + K(x − y))∂2 ψ(x)|ω(y)dy
Rs (k2 )\Rs (a)
Z
−1
≤||ωϕ||∞ ||ϕ ||L∞ (Q) · |∂2 ψ(x)| |∂2 K(s)s2 + K(s)|ds.
[−k2 ,k2 ]2 \[−a,a]2

The function fij (a, b) satisfies the following estimates

f1j (a, b) ≤ A1j + B1j log(b/a), j = 1, 2,
with some constants A1j , B1j . We will provide the formulas and estimates of fij in the upcoming
supplementary material.

D.6.2. Piecewise derivative bounds. In this section, we discuss how to obtain the sharp bound
of p(b)−p(a)
b−a using piecewise derivative bounds of p.
Suppose that |p0 (y)| ≤ Ci , y ∈ Ii = [yi , yi+1 ]. For any a ∈ Ik , b ∈ Il , a < b, we have the bound
Z y2 X
|p(b) − p(a)| ≤ |p0 (y)|dy ≤ |yk+1 − a|Ck + |b − yl |Cl + Cm (ym+1 − ym )
a k+1≤m≤l−1
=(yk+1 − a)Ck + (b − yl )Cl + Mkl (yl − yk+1 )1l≥k+1 ,
where Mkl is defined below:
 X 
Mkl = |yl − yk+1 |−1 Cm |ym+1 − ym | .
k+1≤m≤l−1

|p(b)−p(a)|
Next, we want to bound |b−a| . If l − k ≤ 1, we get
|p(b) − p(a)| ≤ (b − a) max(Ck , Cl ).
Otherwise, if l ≥ k + 2, we have
(p(b) − p(a)) ≤ (yk+1 − a)(Ck − Mkl ) + (b − yl )(Cl − Mkl ) + Mkl (b − a).
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 153

yk+1 −a b−yl
Since b−a is decreasing in a and b, b−a is increasing in b and a, we get

yk+1 − a yk+1 − yk b − yl yl+1 − yl

0≤ ≤ , 0≤ ≤ .
b−a yl − yk b−a yl+1 − yk+1
Using the above estimates, for a ∈ Ik , b ∈ Il , we obtain
|p(b) − p(a)| yk+1 − yk yl+1 − yl
≤ max(Ck − Mkl , 0) + max(Cl − Mkl , 0) + Mkl .
|b − a| yl − yk yl+1 − yk+1
For uniform mesh, i.e. yi+1 − yi = h, we can simplify the above estimate as follows
|p(b) − p(a)| (max(Ck − Mkl , 0) + max(Cl − Mkl , 0))
≤ + Mkl ,
|b − a| l−k
where
1 X
Mkl = Cm .
l−k−1
k+1≤m≤l−1

D.6.3. Optimization in the Hölder estimate. For each t = ch, we minimize the function
b √ Ca
F (a, t) = (A + B log ) t + √ ,
a t
in the upper bound in (8.53) over a ≤ b. We assume that A, B, C, b, c, h, hx are given. Denote
Cb
tu = chx , t1 = .
B
Bt
For a fixed t, since ∂a2 F < 0, ∂a F (0, t) < 0 and ∂a F (a, t) = 0 if a = C , we choose a =
min(b, Bt Cb
C ). For t ≤ B = t1 , we get

Bt bC √ √
min F (a, t) ≤ F ( , t) = (A + B log + B) t − B t log t.
a≤b C B
The right hand side can be further estimated by studying the concave function on s ≤ su
p−q
f (p, q, s) = (p − q log s)s ≤ f (p, q, min(su , exp( ))),
q
1/2 1/2
with p = A + B log( bCB ) + B, q = 2B, su = min(tu , t1 ).
If Cb
B ≤ t ≤ tu , we choose a = b and get
√ Cb
min F (a, t) ≤ F (b, t) = A t + √ ,
a≤b t
which is convex in t1/2 . Thus its maximum is achieved at the endpoints.

Appendix E. Representations and estimates of the solutions

Recall from Section 5 that we represent the approximate steady state as follows

ω̄ = ω̄1 + ω̄2 , θ̄ = θ̄1 + θ̄2 , ω̄1 = χ(r)rα g1 (β), θ̄1 = χ(r)r1+2α g2 (β),

where ω̄2 , θ̄2 have compact supports and are represented as piecewise polynomials. We also
represent the approximation of the stream function in a similar form. We have discussed how
to find the semi-analytic part in Section 5. We will discuss how to estimate the semi-analytic
part in Section E.3. In the following sections, we discuss more details about the representations
and establish rigorous estimate of the derivatives of ω̄2 , θ̄2 .
154 JIAJIE CHEN AND THOMAS Y. HOU

E.1. Representations. In the near field, we use piecewise polynomials to represent the solu-
tion. Let D = [0, L] × [0, L] be the computational domain. Given mesh points 0 = x0 < x1 <
... < xn = L < ... < xn+m , 0 = y0 < y1 < ... < yn = L < ... < yn+m , we define
x−i = −ih1 , y−i = −ih2 , h1 = x1 − x0 = x1 , h2 = y1 − y0 = y1 .
Then, we construct
X
(E.1) ω̄2 (x, y) = aij B1,i (x)Bj (y),
0≤i,j≤n−1

where aij ∈ R is the coefficient, Bi (x), Bj (y) are constructed from the 6−th order B-spline
X (sij − x)k−1
+
Y
(E.2) Bi (x) = Ci k , dj = (sij − sil ),
dj
0≤j≤k 0≤l≤k,l6=j

with k = 6. The constant Ci will be chosen later so that the stiffness matrix associated to these
Bspine basis has a better condition number. The points sij are choosen as follows
sij = xi+j−4 , 0 ≤ j ≤ k = 6.
Then the B-spline Bi is supported in [xi−3 , xi+3 ] and is centered around xi . Since ω is odd
in x, to impose this symmetry in the representation, we modify the first few basis
B1,i (x) = Bi (x) − Bi (−x), i ≤ 2.
Then Bi is odd.
For the density θ̄2 , the representation is similar
X
θ̄2 = x aij B1,i (x)Bj (y).
ij

Here, we multiply x since θ̄ is even and θ̄(x, y) vanishes O(x2 ) near x = 0.

For the stream function ψ̄N,2 , we represent it as follows
X
(E.3) ψ̄N,2 = aij B1,i (x)B2,j (y)ρp (y).
ij

We multiply ρp (y) given below to impose the Dirichlet boundary condition

(E.4) ρp (y) = arctan(1 + y) − 1.
We can obtain the exact formulas of ∂xi ρp using symbolic computation. We use induction to
obtain rigorous estimate of ∂xi ρp . See Section F.3.

E.2. Estimate of the derivatives of piecewise polynomials. Our approximate steady state
is represented as piecewise polynomials. We discuss how to estimate its derivatives. Suppose
that we can evaluate a function f on finite many points. For example, f is an explicit function
or a polynomial. To obtain a piecewise sharp bound of f on I = [xl , xu ], we use the following
estimate
h2
(E.5) max |f (x)| ≤ max(|f (xl )|, |f (xu )|) + ||fxx ||L∞ (I) , h = xu − xl ,
x∈I 8
whose proof is simple. We refer to Section G. If we can obtain a rough bound for fxx , as long
as the interval I is small, i.e., h is small, the error part is small. Similarly, if we can obtain a
rough bound for ∂xk+2 f , using induction and the above estimate recursively,
h2 i+2
max |∂xi f (x)| ≤ max(|∂xi f (xl )|, |∂xi f (xu )|) + ||∂ f ||L∞ (I) ,
x∈I 8 x
for i = k, k − 1, ..., 0, we can obtain the sharp bound for ∂xi f on I. We call the above method
the second order method since the error term is second order in h.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 155

E.2.1. Estimate a piecewise polynomial in 1D. Suppose that p(x) is a piecewise polynomials on
x0 < x1 < .. < xn with degree d, e.g. Hermite spline. Denote Ii = [xi , xi+1 ]. Then p(x) is a
polynomial in each Ii with degree ≤ d. Our goal is to estimate ∂xk p(x) in Ii for all k by only
finite many evaluations of p(x) and its derivatives.
Firstly, we have
∂xk p(x) = 0, k > d, ∂xd p(x) = cp ,
for some constant cp in Ii .
Now, using induction from k = d − 1, d − 2, .., 0, we have
h2i k+2
max |∂xk p(x)| ≤ max(|∂xk p(xi )|, |∂xk p(xi+1 )|) + ||∂ p||L∞ (Ii ) , hi = xi+1 − xi .
x∈Ii 8 x
Since we know ∂xd+1 p(x) = 0 on Ii , using the above method, we can obtain the sharp piecewise
bounds for all derivatives of p(x) on Ii .
Using the above approach, we can estimate the derivatives of the angular profile in (5.2)
rigorously.
E.2.2. Estimate a piecewise polynomial in 2D. Now, we generalize the above ideas to 2D so
that we can estimate the approximate steady state (E.1). We assume that p(x, y) is a piecewise
polynomials in the mesh Qij = [xi , xi+1 ] × [yj , yj+1 ] with degree d. That is, in Qij , p(x, y) can
be written as a linear combination of
xk y l , max(k, l) ≤ d,
e.g. (E.1). For (E.1), we have d = 5. Similar to the 1D case, we have
∂xk ∂yl p(x, y) = 0, max(k, l) > d.
Moreover, we know ∂xd−1 ∂yd−1 is linear in x, y.
We use the following generalization of (E.5) to 2d
||fxx ||L∞ (Q) (xu − xl )2 ||fyy ||L∞ (Q) (yu − yl )2
max |f (x, y| ≤ max |f (xα , yβ )| + + ,
(E.6) (x,y)∈Q α,β=l,u 8 8
Q = [xl , xu ] × [yl , yu ].
The proof follows from (8.10) or the estimates in Appendix G.2.
Denote
Akl , max ||∂xk ∂yl p||L∞ (Qij ) , Bkl , max |∂xk ∂yl p(xα , yβ )|, h1 = xi+1 − xi , h2 = yj+1 − yj .
Qij α,β=l,u

Since p is given, we can evaluate Bkl . Clearly, we have Akl = 0 for max(k, l) > d. For k = d−1, d,
using (E.6) and induction in the order l = d, d − 1, d − 2, .., 0, we can obtain
1
Akl ≤ Bkl + (h21 Ak+2,l + h22 Ak,l+2 ).
8
This allows us to bound Akl for k = d, d−1 and all l. Similarly, we can bound Akl for l = d, d−1
and all k.
For the remaining cases, we can use induction on n = max(k, l) = d − 2, d − 1, .., 0 to estimate
1
Akl ≤ Bkl + (h21 Ak+2,l + h22 Ak,l+2 ).
8
This allows us to estimate all derivatives of p(x, y) in Qij .
E.2.3. Estimate a piecewise polynomial in 2D with weights. We consider how to estimate the
derivatives of f = ρ(y)p(x, y), where ρ is a given weight in y and p(x, y) is the piecewise
polynomials in 2D. For example, our construction of the stream function (E.3) has such a form.
Firstly, we can estimate the derivatives of p(x, y) using the method in Appendix E.2.2. For
the weight ρ, we estimate its derivatives in Section F.3. Then, using the Leibniz rule and the
triangle inequality, we can estimate the derivatives f
X j 
i j
|∂x ∂y f | ≤ |∂xi ∂yl p(x, y)||∂yj−l ρ(y)|
l
l≤j
156 JIAJIE CHEN AND THOMAS Y. HOU

for high enough derivatives.

Now, we plug the above bounds for ∂xi+2 ∂jy , ∂xi ∂yj+2 f in (E.6) and evaluate ∂xi ∂yj f on the grid
points to obtain the sharp estimate of ∂xi ∂yj f .
E.3. Estimate of the far-field approximation. We estimate the derivatives of
g(x, y) = g(r, β) = A(r)B(β).
Note that the semi-analytic parts of ω̄, θ̄ have the above forms.
E.3.1. Formulas of the derivatives of g. Firstly, we use induction to establish
X
(E.7) Fi,j , ∂xi ∂yj g(r, β) = Ci,j,k,l (β)r−i−j+k ∂rk A∂βl B,
k+l≤i+j

with Ci,j,k,l = 0, for k < 0, l < 0 , or k + l > i + j. Let us motivate the above ansatz. Recall
from (D.13) that
sin β cos β
∂x = cos β∂r − ∂β , ∂y = sin β∂r + ∂β .
r r
1
For each derivative ∂x or ∂y , we get the factor r or a derivative ∂r , which leads to the form
−i−j+k k
r ∂r A. Moreover, we get a derivative ∂β and some functions depending on β, which leads
to the form Ci,j,k,l (β)∂βl B.
For D = ∂x or ∂y , a direct calculation yields
X
(E.8) DFi,j = D(Ci,j,k,l r−i−j+k )·∂rk A∂βl B +Ci,j,k,l r−i−j+k (D∂rk A·∂βl B +∂rk A·D∂βl B).
k+l≤i+j

Using the formula of ∂x , ∂y , we get

∂x (Ci,j,k,l (β)r−i−j+k ) = − sin β∂β Ci,j,k,l r−i−j−1+k + (k − i − j) cos βCi,j,k,l r−i−j−1+k ,
sin β l+1
∂x ∂rk A = cos β∂rk+1 A, ∂x ∂βl B = − ∂ B,
r β
Using ∂x Fi,j = Fi+1,j and comparing the above formulas and the ansatz (E.7), we yield
(E.9) Ci+1,j,k,l = (k − i − j) cos βCi,j,k,l − sin β∂β Ci,j,k,l + cos βCi,j,k−1,l − sin βCi,j,k,l−1 ,
for k ≤ i + j. Similarly, for D = ∂y , plugging the following identities
∂y (Ci,j,k,l (β)r−i−j+k ) = cos β∂β Ci,j,k,l r−i−j−1+k + (k − i − j) sin(β)Ci,j,k,l r−i−j−1+k ,
cos β l+1
∂y ∂rk A = sin β∂rk+1 A, ∂y ∂βl B =
∂β B
r
into (E.8) and then comparing (E.7) and (E.8), we yield
(E.10) Ci,j+1,k,l = (k − i − j) sin βCi,j,k,l + cos β∂β Ci,j,k,l + sin βCi,j,k−1,l + cos βCi,j,k,l−1 .
The based case is given by
F0,0 = A(r)g(β), C0,0,0,0 = 1.
Using induction and the above recursive formulas, we can derive Ci,j,k,l (β) in (E.7).
E.3.2. Estimates of Fi,j . To estimate Fi,j , using (E.7) and triangle inequality, we only need
to estimate ∂rk A, ∂βl B(β), and Ci,j,k,l (β). In our case, B(β) is piecewise polynomials, whose
estimates are simple. Function A(r) is some explicit function, which will be constructed and
estimated in Section F.1.
To estimate Ci,j,k,l (β) on β ∈ [β1 , β2 ], we use the second order estimate in (E.5) and the
induction ideas in Section E.2.1.
We can evalute Ci,j,k,l using its exact formula. It remains to bound ∂β2 Ci,j,k,l .
An important observation from (E.9), (E.8) is that Ci,j,k,l is a polynomial on sin β and cos β
with degree less than i + j, which can be proved easily using induction. In particular, we can
write Ci,j,k,l as follows
X X
Ci,j,k,l = ak sin(kβ)+bk cos(kβ), f , ∂β2 Ci,j,k,l = ck sin(kβ)+dk cos(kβ), n = i+j
0≤k≤n 1≤k≤n
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 157

for some ak , bk , ck , dk ∈ R. It is easy to see that Ci,j,k,l is either odd or even in β, which implies
ck ≡ 0 or dk ≡ 0. It follows
X  X 1/2  n Z 2π 1/2
2 2
||f ||∞ ≤ (|ck | + |dk |) ≤ n (ck + dk ) = f2 ,
π 0
1≤k≤n k≤n

where we have used orthgonality of sin kx, cos kx and ||f ||2L2 = π k≤n (c2k + d2k ) in the last
P

equality. It is easy to see that f 2 is again a polynomial in sin β, cos β with degree ≤ 2n. We fix
M > 2n. For any 0 ≤ k < M , it is easy to obtain
Z 2π M
1 1 X 2kj
eikx = exp(i π) = δk0 .
2π 0 M j=1 M

Using the above identity, we establish

M
2π X 2jπ 2
||g||2L2 = |g( )| ,
M j=1 M

for any polynomial g in sin β, cos β with degree < M . Hence, we prove
M
 2n X 2jπ 1/2
||f ||∞ ≤ f 2( ) .
M M
k=1

The advantage of the above estimate is that to obtain the sharp bound of Ci,j,k,l , we only
need to evaluate Ci,j,k,l , f = ∂β2 Ci,j,k,l on finite many points.

E.3.3. From polar coordinates to the Cartesion coordinate. We want to obtain the piecewise
estimate of Fp,q = ∂xp ∂yq (A(r)g(β)) on Qij = [xi , xi+1 ] × [yj , yj+1 ], 1 ≤ i, j ≤ n. Firstly, we
partition the (r, β) coordinate into r1 < r2 < .. < rn1 , 0 = β0 < b1 < ... < βn2 = π2 . Then we
apply the methods in Section E.3 to bound Fp,q on Sij , [ri , ri+1 ] × [βj , βj+1 ].
To estimate Fp,q , we cover Qij by Sk,l and transfer the bound from (r, β) coordinate to (x, y)
coordinate
max |Fp,q (x)| ≤ max ||Fp,q (r, β)||L∞ (Sk,l )
x∈Qij Sk,l ∩Qij 6=∅

For (r, β) ∈ Qi,j , we get

yj yj+1
r ∈ [(x2i + yj2 )1/2 , (x2i+1 + yj+1
2
)1/2 ], β ∈ [arctan , arctan ].
xi+1 xi
Therefore, we yield the neccessary conditions for Qi,j ∩ Sk,l 6= ∅:
yj+1 yj
x2i+1 + yj+1
2
≥ rk2 , x2i + yi2 ≤ ru2 , arctan ≥ βl , arctan ≤ βl+1 .
xi xi+1
Given Qi,j , we maximize ||Fp,q ||L∞ (Sk,l ) over (k, l) satisfying the above bounds to control
||Fp,q ||L∞ (Qi,j ) .

Appendix F. Estimate of explicit functions

In this section, we estimate the derivatives of several explicit or semi-explicit functions using
induction, including several cutoff functions used in the estimates and the weight in the stream
function (E.3).

F.1. Estimate of the radial functions.

158 JIAJIE CHEN AND THOMAS Y. HOU

F.1.1. Estimate of the cutoff function. We estimate the derivatives of the cutoff function
 1 1 −1
(F.1) χe (x) = 1 + exp( + )
x x−1
for a > 0, where e is short for exponential. In our verification, it involves high order derivatives
of χe . Although χe is explicit, its formula is complicated and is difficult to estimate. Instead,
we use the structure of ∂xi χe and induction to estimate ∂xi χe . Denote
1 1 1
p(x) = + , f= , χe = f (ep ).
x x−1 1+x
Firstly, we use induction to derive
k
X
dkx χe = (∂ i f )(ep )eip Qk,i (x),
i=1
where Qk,i = 0 for i > k, i < 0. A direct calculation yields
k
X k
X
∂ ∂ i f eip Qk,i (x) = (∂ i+1 f )(ep ) · p0 ep eip Qk,i + (∂ i f )∂x (eip Qk,i )
i=1 i=1
k
X
= (∂ i+1 f )(ep ) · e(i+1)p Qk,i + (∂ i f )eip (ip0 Qk,i + Q0k,i ).
i=1
Comparing the above two equations, we derive
Qk+1,i = p0 Qk,i−1 + ip0 Qk,i + Q0k,i .
The first few terms in Qk,i are given by
Q0,0 = 1, Q1,1 = p0 , Q1,0 = 0.
It is not difficult to see that Qk,i is a polynomial of ∂xj p, j ≤ k. Thus, using triangle inequality,
we only need to bound ∂xj p. We have
|∂xn p(x)| ≤ n!(|x|−n−1 + |x − 1|−n−1 ) ≤ n!(|z|−n−1 + 2n+1 ), z = min(|x|, |1 − x|).
Substituting the above bounds into the formula of Qk,i , we can obtain the upper bound
Quk,i (x) for Qk,i (x), which is a polynomial of z −1 with positive coefficient. Since each term in
Qm i P
Qk,i is given by ci1 ,i2 ,..,im j=1 ∂xj p with ij = k, the above estimate implies
m
Y m
Y
|ci1 ,i2 ,..,im ∂xij p| ≤ ci1 ,i2 ,..,im ij !(|z|−ij −1 + 2ij +1 ).
j=1 j=1
−1
Since m ≤ k, the highest order of z in the upper bound is bounded by 2k. Thus, we obtain
that Quk,i is a polynomial in z with deg Quk,i ≤ 2k. Next, we bound
|eip Qk,i | ≤ eip Quk,i .
For k ≤ 20, x ≥ 21 , z −1 = |x − 1|−1 ≥ 2k, a direct calculation implies that eip(x) Quk,i (x) is
decreasing. In fact, for l ≤ 2k, we have z = |x − 1| = 1 − x and
∂x (exp(ip(x))(1 − x)−l ) = exp(ip(x))(ip0 (1 − x)−l + l(1 − x)−l−1 )
 i i −1

= exp(ip(x)) − 2 − + l(1 − x) (1 − x)−l ≤ 0.
x (x − 1)2
i
In the last inequality, we have used − 1−x + l ≤ −2ki + 2k ≤ 0.
i p p −i−1
Note that |(∂x f )(e )| = i!|(1 + e ) | ≤ i!. Thus, for x ∈ [xl , xu ] with xl close to 1, we get
k
X k
X
|∂xk χe (x)| ≤ |∂ i f (x)|eip(xl ) Quk,i (xl ) ≤ i!eip(xl ) Quk,i (xl ).
i=1 i=1

Using the above derivatives bound, the symbolic formula of ∂xk χe , and the refined second
order estimate in Section E.2.1, we can obtain sharp bounds for ∂xk χe .
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 159

F.1.2. Estimate of polynomial decay functions. For cutoff function χe ( |x|−a

b ) based on the ex-
ponential cutoff function (F.1), it has rapid change from |x| ≤ a to |x| ≥ a + b, which is not very
smooth in the computational domain if there are not enough mesh for x with a ≤ |x| ≤ b. We
apply these cutoff functions to the far-field, e.g. |x| ≥ 10, where the mesh is relatively sparse.
Thus, we need another function similar to a cutoff function that has a slower change than the
exponential cutoff function. We consider
x7
(F.2) χ(x) = , x ∈ R+ .
(1 + x2 )7/2
and will use its rescaled version, e.g., χ( x−a
b ), in our verification.
Firstly, we use induction to derive
pk (x)
∂xk χ = , p0 = x7 .
(1 + x2 )7/2+k
where pk (x) is a polynomial. A direct calculation yields
p0k (x)(1 + x2 ) − ( 72 + k) · 2xpk (x)
∂xk+1 χ = .
(1 + x2 )7/2+k+1
Comparing the above two formulas, we yield
pk+1 = p0k (1 + x2 ) − (7 + 2k)xpk (x).
The first few terms are given by p0 = x7 , p1 = 7x6 . Using the recursive formula and
deg p1 = 6, we yield
(F.3) deg pk+1 ≤ deg pk + 1, deg pk ≤ k + 5, k ≥ 1.
Since pk is a polynomial, the above recursive formula shows that pk+1 is also a polynomial.
ToPestimate ∂xk χ, we decompose pk into the positive and the negative parts. Suppose that
pk = i ai xi . We have
X X
−
pk = p+k − pk , p+
k = a+ i
i x , p−k = a− i
i x .
−
For x ≥ 0, p+
k , pk are increasing. Thus, for x ∈ [xl , xu ], we get
− −
max(p+ +
k (xu ) − pk (xl ), pk (xu ) − pk (xl ))
|∂xk χ| ≤ 2
.
(1 + xl )7/2+k
Next, we estimate ∂xk χ for large x. For x ≥ 2, k ≥ 2 and any polynomial q(x) with non-
negative coeffcients and deg q ≤ k + 5, we yield
q 0 (1 + x2 ) (1 + x2 )(k + 5) 5(k + 5)
xq 0 ≤ (k + 5)q, ≤ ≤ < 1.
(7 + 2k)xq (7 + 2k)x2 4(7 + 2k)
The first inequality follows by comparing the coefficients of xq 0 and (k + 5)q, which are
nonnegative. It follows
q q 0 (1 + x2 ) − (7/2 + k)2xq
∂x = ≤ 0.
(1 + x2 )7/2+k (1 + x2 )7/2+k+1
q
Thus (1+x2 )7/2+k
is decreasing. For k ≥ 1 and x ≥ xl ≥ 2, using (F.3) and the monotonicity, we
yield
− −
p+
k (x) + pk (x) p+
k (xl ) + pk (xl )
|∂xk (x)| ≤ ≤
(1 + x2 )7/2+k (1 + x2l )7/2+k
For k = 0, the estimate is trivial: χ(x) ≤ 1. Using these higher order derivative bounds, we
can use the discrete values of ∂xk χ and the bound for ∂xk+2 χ to obtain sharp bounds of ∂xk χ.
(x−a)7+
Note that χ1 (x − a) = (1+(x−a)2 )7/2
is only C 6,1 . Suppose that a ∈ [xl , xu ]. Since χ1 is
smooth on x ≤ a and on x ≥ a, we can still use first order estimate to estimate ∂xk χ1 as follows
|∂xk χ1 (x)| ≤ max |∂xk χ1 (xα )| + ||∂xk+1 χ1 ||L∞] [xl ,xu ] |xu − xl |.
α∈{l,u}
160 JIAJIE CHEN AND THOMAS Y. HOU

F.1.3. Radial cutoff function. Now, we construct the radial cutoff functions for the far-field
approximation terms of ω and ψ as follows
x − a1 x − a2
(F.4) χ(r) = χ1 (1 − χ2 ) + χ2 , χ1 (r) = χrati ( 1/2 ), χ2 (r) = χexp ( ),
l1 9a2
where χexp and χrati are defined in (F.1) and (F.2), respectively. Using the estimates of
χrati , χexp established in the last two sections and the interval operations in (C.5), we can
evaluate χ on the grid points and estimate its bounds.
F.2. Cutoff function near the origin. For the cutoff function κ(x) used in Section 7, we
choose it as follows
x 1
(F.5) κ(x; ν1 , ν2 ) = κ1 (x/ν1 )(1 − χe ( )), κ1 (x) = , ν1 = 1/3, ν2 = 1.5,
ν2 1 + x4
where χe is the cutoff function chosen in (F.1). Since χe (y) = 1 for y ≥ 1 and χe (y) = 0 for
y ≤ 0. The above cutoff function is supported in x ≤ a2 . Using Taylor expansion, we have the
following properties for κ
κ1 (x/a1 ) = 1 + O(x4 ), κ(x) = 1 + O(x4 ).
For κ1 (x), we can use induction and the same method as that in Section F.1.2 to estimate the
derivatives of ∂xi κ1 (x). The estimate is simpler since κ1 has a simpler form. Using the Leibniz
rule and the triangle inequality, we can obtain estimate ∂xl κ(x) in [a, b]. Then we use these
derivative estimates for κl+2
x κ(x), evalutate κ(x; a1 , a2 ) on the grid points, and then use (E.5)
to obtain a sharp estimate of ∂xl κ(x) on [a, b].
F.3. Estimate of ρp (y). We estimate the weight ρp (y) (E.4) in the representation of the stream
function. Using symbolic computaiton, e.g., Matlab or Mathematica, we yield
f2 (y) − f1 (y)
∂x9 ρp (y) = , g(y) = 2 + 2y + y 2 ,
(g(y))8
f1 = 288y 2 + 672y 3 + 504y 4 , f2 = 16 + 168y 6 + 72y 7 + 9y 8 .
Since f1 , f2 , g ≥ 0 are increasing in y ≥ 0, for y ∈ [yl , yu ], we yield
max(f2 (yu ) − f1 (yl ), f1 (yu ) − f2 (yl )
|∂x9 ρp (y)| ≤ .
(g(yl ))8
We have a trivial estimate similar to (E.5)
h
(F.6) max |f (x)| ≤ max(|f (xl )|, |f (xu )|) + ||fx ||L∞ (I) ,
x∈I 2
which is useful if we do not have bound for fxx .
Based on the above estimates, using the estimates (E.5), (F.6), ideas in Section E.2.1, and
evaluating ρp on some grid points, we can obtain piecewise sharp bounds for ∂xi ρp for i ≤ 8.

Appendix G. Estimating the piecewise bounds of functions

In this appendix, we estimate ||f ||L∞ (D) in the domain D = [a, b] × [c, d] given the grid point
values f (xi , yj ) in D and the derivatives bound ||∂xi ∂yj f ||L∞ . We want to obtain an error term
as small as possible, while we do not need to evaluate f on too many grid points and its high
order derivatives, which are expensive for some complicated function f , e.g. f = (∂t − L)W c
(7.14). Based on these L∞ estimates, we further develop the Hölder estimate of a function
in Section G.6. These estimates will be used to verify the smallness of the residual error, e.g.
f = (∂t − L)W c (7.14), in suitable energy norm.
We will develop three estimates based on three interpolating polynomials: the Newton poly-
nomial, the Lagrangian interpolating polynomial, and the Hermitte interpolation. Each method
has its own advantages. For the Newton and the Lagrangian method, to obtain 4-th order error
estimates, we only need to evaluate 4 × 4 grid point values of f .
(a) For the Newton method, we have a sharp error bound with a much smaller constant than
that of the Lagraingian method.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 161

(b) For the Lagrangian method, it is easier and more efficient to estimate the Lagrangian
interpolating polynomials for grid points (xi , yj ) in a general position.
In some situation, we need to estimate both f and ∇f and evalute f and ∇f on 4 × 4
grid points. In this case, we can use f (xi , yj ) and ∇f (xi , yj ) to build fourth and fifth order
interpolating polynomials based on the Hermitte interpolation. The 4-th order error estimate is
as sharp as the Newton method. Moreover, in the 4-th order Newton or Lagrangian interpolation,
we need f ∈ C 4 [x0 , x3 ]. In the case where f is only piecewise smooth in [xi , xi+1 ], we cannot use
these two methods. Instead, we evaluate f (xi ), f 0 (xi ) and construct the Hermite interpolation in
each interval [xi , xi+1 ]. One disadvantage is that the estimate of the interpolating polynomials
is more complicated and takes longer time in practice.
We do not pursue higher order error estimates since most of these estimates are applied to
estimate the errors in Section 7, e.g. (∂t − L)W c (7.14), which is only piecewise smooth, and we
do not use very small h in the whole computational domain.
G.1. Estimates based on the Newton polynomials. Given x0 , x1 , x2 , .., xk , we first define
the divided differences recursively
f (y) − f (x) f [xi+1 , xi+2 , .., xj+1 ] − f (xi , xi+1 , .., xj )
f [x] = f (x), f [x, y] = , f [xi , x1 , .., xj+1 ] = .
y−x xj+1 − xi
G.1.1. The Newton polynomials in 1D. We first discuss how to bound f (x) in 1D. We consider
the domain [a, b] and denote
z0 = a, h = (b − a)/3, zi = z0 + ih.
Denote
D1,i f = f (zi+1 ) − f (zi ), i = 0, 1, 2, D2,i f = D1,i+1 f − D1,i f, i = 0, 1,
(G.1)
D3 f = D2,1 f − D2,0 f = f (z3 ) − 3f (z2 ) + 3f (z1 ) − f (z0 ).
Let {xi }3i=0 be a permutation of {zi }3i=0 . We construct the Newton polynomial
P (x) = (f [x0 ] + f [x0 , x1 ](x − x0 )) + f [x0 , x1 , x2 ](x − x0 )(x − x1 )
(G.2)
+ f [x0 , x1 , x2 , x3 ](x − x0 )(x − x1 )(x − x2 ) , l(x) + q(x) + c(x),
where l(x), q(x), c(x) denote the linear, quadratic, and the cubic parts, respectively.
By standard error analysis of the Newton interpolation, the error part R(x) can be bounded
as follows
1 1 (b − a)4
(G.3) |f (x) − P (x)| ≤ ||∂x4 f ||L∞ [a,b] max |Π0≤i≤3 (x − xi )| = ||∂x4 f ||L∞ [a,b] .
24 x∈[a,b] 24 81
To obtain the last equality, using the definition of xi , zi , we write z = a + th, t ∈ [0, 3] and get
max |Π0≤i≤3 (x − xi )| = max |Π0≤i≤3 (z − zi )| = max h4 |Π0≤i≤3 (t − i)| ≤ h4 ,
x∈[a,b] z∈[a,b] t∈[0,3]

where we have used (G.56) in Lemma G.3 in the last inequality.

To bound f (x), given the derivative bound of f and the above estimate, we only need to
control P (x). We choose different permutation {xi }3i=0 of {zi }3i=0 for z in different part of [a, b]:
xi = zi , z ∈ [z0 , z1 ], (x0 , x1 , x2 , x3 ) = (z2 , z1 , z0 , z3 ), z ∈ [z1 , z2 ],
(x0 , x1 , x2 , x3 ) = (z3 , z2 , z1 , z0 ), z ∈ [z2 , z3 ].
Let Iz be the interval with endpoints x0 , x1 . We have z0 ∈ Iz . Since l(x) in (G.2) is linear with
2
l(xi ) = f (xi ) and |(x − x0 )(x − x1 )| ≤ (x1 −x
4
0)
, we get
(x1 − x0 )2 h2
|l(z)| ≤ max(|f (x0 )|, |f (x1 )|), max |q(x)| ≤ |f [x0 , x1 , x2 ]|
= |f [x0 , x1 , x2 ]| .
z∈Iz 4 4
Since x0 , x1 , x2 are three consecutive points with distance h and I = [min(x0 , x1 , x2 ), max(x0 , x1 , x2 )]
covers z. In particular, we have
2
(G.4) max |(z − x0 )(z − x1 )(z − x2 )| = max |z(z − h)(z − 2h)| ≤ √ h3 ,
z∈I z∈[0,2h] 3 3
162 JIAJIE CHEN AND THOMAS Y. HOU

where we have used (G.54) in Lemma G.3 in the last inequality.

Next, we use (G.1) to simplify f [xi , xi+1 , .., xj ]. For each case, a direct calculation yields
f [z2 , z1 ] − f [z1 , z0 ] 1 1
z ∈ [z0 , z1 ] : |f [z0 , z1 , z2 ]| = = (D1,1 f − D1,0 f ) = 2 |D2,0 f |,
z2 − z0 2h2 2h
f [z0 , z1 ] − f [z1 , z2 ] 1 1
z ∈ [z1 , z2 ] : |f (z2 , z1 , z0 )| = = (D1,1 f − D1,0 f ) = 2 |D2,0 f |.
z0 − z2 2h2 2h
Similarly, for z ∈ [z2 , z3 ], we get
1
|f (z3 , z2 , z1 )| = 2 |D2,1 f |.
2h
A direct calulation yields |f [x0 , x1 , x2 , x3 ]| = 6h1 3 |D3 f |. Thus, we get
2 2 1 1
max |c(x)| ≤ √ h3 |f [x0 , x1 , x2 , x3 ]| = √ h3 · 3 |D3 f | = √ |D3 f |.
3 3
z∈[a,b] 3 3 6h 9 3
Combining the above estimates, we obtain
 
(G.5) |P (x)| ≤ max max |f (zi )| + c1 |D2,0 f |, max |f (zi )| + c1 |D2,1 f | + c2 |D3 f |.
i=0,1,2 i=1,2,3

where
1 1
c1 = , c2 = √ .
8 9 3
G.1.2. A quadratic interpolation. We also need a cubic interpolation in the Hermite interpolation
in Section G.4. Given x0 < x1 < x2 with x2 − x1 = x1 − x0 , we define
(G.6) N2 (f, x0 , x1 , x2 )(x) , f (x0 ) + f [x0 , x1 ](x − x0 ) + f [x0 , x1 , x2 ](x − x0 )(x − x1 ),
and construct P (x) = N2 (f, x0 , x1 , x2 )(x).
We have an error estimate similar to (G.3) for x ∈ [x0 , x2 ]
(G.7)
||∂ 3 f ||L∞ [x0 ,x2 ] ||∂ 3 f ||L∞ [x0 ,x2 ] 2h3 ||∂ 3 f ||L∞ [x0 ,x2 ] h3
|P (x)−f (x)| ≤ x max |Πi=0,1,2 (x−xi )| ≤ x √ = x √ ,
6 x∈[x0 ,x2 ] 6 3 3 9 3
where we have used (G.4) in the last inequality.
Using the same estimates in Section G.1.1 for the linear part and quadratic part, we obtain
1
x ∈ [x0 , x1 ] : |P (x)| ≤ max(|f (x0 )|, |f (x1 )|) + |f (x0 ) − 2f (x1 ) + f (x2 )|,
(G.8) 8
1
x ∈ [x0 , x1 ] : |P (x)| ≤ max(|f (x0 )|, |f (x1 )|) + |f (x0 ) − 2f (x1 ) + f (x2 )|.
8
Note that using the notation (G.1), we have |D2,0 f | = |f (x0 ) − 2f (x1 ) + f (x2 )|.
G.1.3. Generalization to 2D. Denote
D = [a, b] × [c, d], xi = a + ih1 , yj = c + jh2 , h1 = (b − a)/3, h2 = (d − c)/3.
Suppose that f (xi , yj ) and ||∂xk ∂ l f ||L∞ , k + l ≤ 4 are given. Firstly, we treat y as a parameter
and interpolate f (x, y) in x. Denote
Di,1 f (y) = f (xi+1 , y) − f (xi , y), 0 ≤ i ≤ 2, Di,2 f (y) = Di+1,1 f (y) − Di,1 f (y), 0 ≤ i ≤ 1,
D3 f (y) = D2,1 f (y) − D2,0 f (y).
Applying (G.3), (G.5), we get
 
max |f (x, y)| ≤ max max |f (xi , y)| + c1 |D2,0 f (y)|, max |f (xi , y)| + c1 |D2,1 f (y)|
x∈[a,b] i=0,1,2 i=1,2,3
1 4 4
+ c2 |D3 f (y)| +
h ||∂ f ||L∞ (D) .
24 1 x
Note that f (xi , y), Di,j f (y) are 1D functions in y and their grid point values on yj can
be obtained from f (xi , yj ). We further apply (G.3),(G.5) to estimating ||g||L∞ [c,d] with g =
f (xi , y), Di,j f (y). Maximizing the above estimate over y yields the bound for f .
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 163

G.2. Estimates based on the Lagrangian interpolation. If the grid points xi are not
uniform, the estimates of the Newton polynomials can be more complicated. We develop another
estimate based on the Lagragian interpolating polynomials. Although these two interpolating
polynomials for given grid points f (xi ) are equivalent, the Lagrangian formulation is sometimes
easier to estimate.
Firstly, let pi (x), qj (y) be the Lagrange interpolating polynomials associated to the points
x1 < .. < xk ∈ [a, b], y1 < y2 < ... < yk ∈ [c, d]
x − xj y − yj
(G.9) pi (x) = Πj6=i , qi (y) = Πj6=i .
xi − xj yi − yj
For any (x, y) ∈ D, we consider the following decomposition by first interpolating f in x and
then in y
(G.10)
k
X k X
X k
f (x, y) = f (xi , y)pi (x) + R1 (x, y) = ( f (xi , yj )qj (y) + R2 (xi , y))pi (x) + R1 (x, y)
i=1 i=1 j=1
k
X X 
= pi (x)qj (y)f (xi , yj ) + R2 (xi , y)pi (x) + R1 (x, y) , I + II.
i,j=1 i

By standard error analysis of the Lagrange interpolation, the error part R1 (x, y), R2 (x, y) can
be bounded as follows
1
|R1 (x, y)| ≤ ||∂xk f (x, y)||L∞ (D) max Πki=1 (x − xi ) ,
k! x∈[a,b]
(G.11)
1 k
|R2 (xi , y)| ≤ ||∂y f (x, y)||L∞ (D) max Πki=1 (y − yi ) .
k! y∈[c,d]

Denote
k
X
C1 = max |pi (x)|, aij = f (xi , yj ).
x∈[a,b]
i=1
xi+1 −xi
Note that the value C1 only depens on the ratio b−a , i = 1, .., k, since from (G.9), we have
x − xj t − tj x−a xj − a
pi (x) = Πj6=i = Πj6=i , t= , tj = .
xi − xj ti − tj b−a b−a
Pk
We will choose the grid points so that we also have C1 = maxy∈[c,d] i=1 |qj (y)|. See (G.15).
We have the following trivial estimate for any cj
X X X
(G.12) | pi (x)ci | ≤ |pi (x)| max |ci | ≤ C1 max |ci |, | qj (y)cj | ≤ C1 max |ci |.
i i j
P P
Next, we estimate I. Since i pi (x) = j qj (y) = 1, we expect that I ≈ f (xi , yj ) +
O(max(h1 , h2 )), where h1 = b − a, h2 = d − c. Thus, for some m to be chosen, we further
decompose it into the mean and the variation and apply (G.12) to obtain
k
X X X
|I| = |m + pi (x)qj (y)(aij − m)| ≤ |m| + max |aij − m| |pi (x)| |qj (y)|
i,j
i,j=1 i j

= |m| + C12 max |aij − m|.

i,j

We use the following trivial inequality for b1 , b2 , .., bn

1 1
(G.13) max |bi − b| = (max bi − min bi ), b= (max bi + min bi ),
i 2 2
which can be proved by ordering bi .
Thus, we optimize the estimate of I by choosing m = 21 (maxi,j aij + mini,j aij ).
164 JIAJIE CHEN AND THOMAS Y. HOU

We can obtain a sharper estimate as follows

X  X  X
|I| = qj (y) āj + pi (x)(aij − āj ) = qj (y)(āj + Sj (x))
j i j
(G.14) X X X
≤ |ā| + | qj (y)(āj − ā)| + | qj (y)Sj (x)|, Sj (x) = pi (x)(aij − āj ).
j j i

For a fixed j, we choose

maxi aij + mini aij maxj āj + min āj
āj = , ā = .
2 2
Applying (G.12), (G.13), and the definition of Sj (x) in (G.14), we yield
X C1
| lj (y)Sj (x)| ≤ C1 · max |Sj (x)|, |Sj (x)| ≤ C1 max |aij − āj | = | max aij − min aij |.
j
j i 2 i i

Similarly, we have
X C1
lj (x)(āj − ā) ≤ C1 max |āj − ā| = (max āj − min āj ).
j
j 2 j j

Combining two parts, we yield an improved estimate for |I|

1 C1 C2
|I| ≤ | max āj + min āj | + (max āj − min āj ) + 1 max | max aij − min aij |.
2 j j 2 j j 2 j i i

The above estimate is better if aij = f (xi , yj ) is smooth in x. Similarly, we can first sum over
pi (x) and then sum over qj (y) in (G.14) to obtain another improved estimate.
In practice, we apply the above method to the third order estimate of f on [a, b] × [c, d], we
choose
1 1 31 1 1 31
(G.15) (x1 , x2 , x3 ) = (a+ h1 , a+ h1 , a+ h1 ), (y1 , y2 , y3 ) = (c+ h2 , c+ h2 , c+ h2 ),
32 2 32 32 2 32
where h1 = b − a, h2 = d − c. Then we get
3
Y 3
Y
C1 ≤ 1.28, max |x − xi | ≤ 0.04h31 , max |y − yi | ≤ 0.04h32 .
x∈[a,b] x∈[a,b]
i=1 i=1

Using (G.11), (G.12), and the above estimate, we get

1
|II| ≤ C1 max ||R2 (xi , y)||L∞ + ||R1 ||L∞ ≤ · 0.04(C1 ||∂y3 f ||L∞ (D) h32 + ||∂x3 f ||L∞ (D) h31 ).
i 6
We can also first interpolate f in y and then in x (G.10) to obtain another form of II. Similar
estimates yield
1
|II| ≤ · 0.04(C1 ||∂x3 f ||L∞ (D) h31 + ||∂y3 f ||L∞ (D) h32 ).
6
We minimize the above two estimates to bound II.

G.3. Hermite interpolation in 1D. We first discuss the Hermite interpolation in 1D, and
then generalize it to 2D.
Consider x0 < x1 < x2 with x2 − x1 = x1 − x0 . Denote by pi , q̂i the cubic polynomials such
that
pi (xi ) = 1, pi (x1−i ) = 0, i = 0, 1, p0i (xj ) = 0, j = 0, 1,
qi0 (xi ) = 1, qi0 (x1−i ) = 0, i = 0, 1, qi (xj ) = 0, j = 0, 1,
and
(G.16) li = f (xi ), mi = hf 0 (xi ), h = x1 − x0 .
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 165

We consider the 4 − th and 5 − th order Hermite interpolations for f

X
H4 (f, x0 , x1 )(x) = (f (xi )pi (x) + f 0 (xi )qi (x)),
i=0,1
(G.17)
(x − x0 )2 (x − x1 )2
H5 (f, x0 , x1 , x2 )(x) = H4 (x) + (f (x2 ) − H4 (x2 )) .
(x2 − x0 )2 (x2 − x1 )2
For simplicity, we drop the dependence of f, xi . Note that the coefficients of the polynomials
pi , qi depend on x0 , x1 , x2 only. It is easy to see that
H4 (xi ) = H5 (xi ) = f (xi ), H40 (xi ) = H50 (xi ) = f 0 (xi ), i = 0, 1, H5 (x2 ) = f (x2 ).

G.3.1. Estimates of the interpolation error. For [x0 , x2 ], we have the following error estimate
1
(G.18) |f (x) − H4 (x)| ≤ ||∂ 4 f ||L∞ ([x0 ,x2 ]) (x − x0 )2 (x − x1 )2 , x ∈ [x0 , x2 ].
24 x
For x ∈ [x0 , x1 ], we have the following error estimates with further simplifications
(G.19)
1 h4
|f (x) − H4 (x)| ≤ ||∂x4 f ||L∞ ([x0 ,x1 ]) (x − x0 )2 (x − x1 )2 ≤ ||∂ 4 f ||L∞ ([x0 ,x1 ]) ,
24 384 x
||∂ 5 f ||L∞ ([x0 ,x2 ]) h5
|f (x) − H5 (x)| ≤ x |(x − x0 )2 (x − x1 )2 (x − x2 )| ≤ ||∂ 5 f ||L∞ ([x0 ,x2 ]) ,
120 1200 x
where we have used (G.52), (G.57) in Lemma G.3 in the last inequality. The proof of the first
inequality is standard. We consider the second estimate. For any t ∈ [x0 , x2 ], denote
(x − x0 )2 (x − x1 )2 (x − x2 )
Rt (x) = f (x) − H5 (x) − (f (t) − H5 (t)) .
(t − x0 )2 (t − x1 )2 (t − x2 )
Clearly, we have Rt (xi ) = 0, i = 0, 1, 2, Rt0 (xi ) = 0, i = 0, 1, Rt (t) = 0, and Rt has 6 zeros. For
f ∈ C 4,1 , applying the Rolle’s theorem repeatedly up to ∂x4 f , we yield ξ1 6= ξ2 ∈ (x0 , x2 ) with
(f (t) − H5 (t))(120ξi − C2 )
0 = ∂x4 f (ξi ) − C1 − ,
(t − x0 )2 (t − x1 )2 (t − x2 )
where we have used ∂x4 H5 (x) = C1 , ∂x4 (x − x0 )2 (x − x1 )2 (x − x2 ) = 120x − C2 . Rewritting the
above identities and computing the difference, we obtain
∂x4 f (ξ2 ) − ∂x4 f (ξ1 ) ||∂ 5 f ||L∞ [x0 ,x1 ]
|f (t)−H5 (t)| = (t−x0 )2 (t−x1 )2 (t−x2 ) ≤ x (t−x0 )2 (t−x1 )2 (t−x2 ) .
120(ξ2 − ξ1 ) 120
Since t is arbitrary, this proves the second estimate in (G.19). The first estimate can be proved
similarly. Next, we estimate the interpolating polynomials H4 , H5 .
We should compare the first estimate (G.19) with (G.3). In (G.3), x1 − x0 = b−a 3 = h and the
2
upper bound is 241
||∂x4 f ||∞ h4 . In (G.19), for x ∈ [x0 , x1 ], using |(x − x0 )(x − x1 )| ≤ (x1 −x
4
0)
, we
1
obtain an extra small factor 16 . This is one of the main advantages of the Hermite interpolation.
x−x0
G.3.2. Estimate H4 , H5 . We consider x ∈ [x0 , x1 ] and introduce t = h . Recall li , mi from
(G.16). We have
G4 (t) , H4 (x0 + th) = H4 (x),
   
(G.20) G4 (t) = l0 (1 − t) + l1 t + t2 (t − 1)(m1 − (l1 − l0 )) + (t − 1)2 t(m0 − (l1 − l0 ))
, I(t) + II(t).
To show that G4 defined via the first identity has the second expression, it suffices to verify
that the expression satisfies G4 (i) = li = f (xi ), ∂t G4 (i) = mi = hf 0 (xi ), which is obvious. To
estimate H4 (x) on [x0 , x1 ], we only need to estimate G4 (t) on [0, 1].
The estimate of the linear part is trivial
|I(t)| = |l0 (1 − t) + l1 t| ≤ max(|l0 |, |l1 |).
166 JIAJIE CHEN AND THOMAS Y. HOU

Denote
(G.21) M1 = max(|m1 − m0 |, |m1 − (l1 − l0 )|, |m0 − (l1 − l0 )|).
1
For 0 ≤ t ≤ 2, using (G.55) for t(t − 1)2 , we have
II(t) = t(t − 1)(t(m1 − (l1 − l0 )) + (t − 1)(m0 − (l1 − l0 ))
= t(t − 1)(t(m1 − m0 ) + (2t − 1)(m0 − (l1 − l0 ))),
4
|II(t) ≤ M1 |t(t − 1)|(t + |1 − 2t|) = M1 t(1 − t)2 ≤ M1 .
27
Similarly, for t ∈ [1/2, 1], writting m0 − (l1 − l0 ) = m0 − m1 + (m1 − (l1 − l0 )), we get
II(t) = t(t − 1)((t − 1)(m0 − m1 ) + (2t − 1)(m1 − (l1 − l0 ))),
4
|II(t)| ≤ |t(t − 1)|(|t − 1| + |2t − 1|)M1 = t(1 − t)(1 − t + 2t − 1)M1 = t2 (1 − t)M1 ≤ M1 .
27
To obtain the last inequality, we apply (G.55) with s = 1 − t. Therefore, we prove
4
(G.22) |H4 (x)| = |G(t)| ≤ max(|l0 |, |l1 |) + max(|m1 − m0 |, |m1 − (l1 − l0 )|, |m0 − (l1 − l0 )|).
27
For H5 , since x2 − x1 = x1 − x0 = h, we have
H4 (x2 ) = G4 (2) = −l0 + 2l1 + 4(m0 − (l1 − l0 )) + 2(m0 − (l1 − l0 )) = l0 + 2m0 + 4(m1 − (l1 − l0 )).
For x ∈ [x0 , x1 ] or t ∈ [0, 1], since |t(1 − t)| ≤ 14 , we have
(x − x0 )2 (x − x1 )2 t2 (t − 1)2 1
= ≤ .
(x2 − x0 )2 (x1 − x0 )2 4 64
We obtain
1
(G.23) max |H5 | ≤ max |H4 (x)| + |f (x2 ) − H4 (x2 )|.
x∈[x0 ,x1 ] x∈[x0 ,x1 ] 64

G.3.3. Estimate derivatives in 1D. In this subsection, we discuss how to estimate ∂f (x) using
the Hermite interpolation ∂H5 (x). We consider ∂x without loss of generality. Firstly, since
f (x) − H5 (x) has five zeros: two zeros at x0 , two zeros at x1 , and one zero at x2 , we know that
∂x (f (x) − H5 (x)) has four zeros: x0 < ξ < x1 < η. Using the Rolle’s theorem and an argument
similar to that in Section G.3.1, we get
1
|∂x (f (x) − H5 (x))| ≤ ||∂ 5 f ||L∞ [x0 ,x2 ] |(x − x0 )(x − x1 )(x − ξ)(x − η).
24 x
Next, for x ∈ [x0 , x1 ], we simplify the upper bound. Clearly, we have
|x − ξ| ≤ max(|x − x0 |, |x1 − x|), |x − η| ≤ |x − x2 |. We yield
p(x) , |(x − x0 )(x − x1 )(x − ξ)(x − η) ≤ |(x − x0 )(x − x1 )(x − x2 )| max(|x − x0 |, |x − x1 |)
= h4 |t(1 − t)(2 − t)| max(|1 − t|, t) , h4 q(t), t = (x − x0 )h−1 .

If t ≤ 21 , we denote s = (1 − t)2 ∈ [0, 1] and get

1
q(t) = t(1 − t)2 (2 − t) = (2t − t2 )s = (1 − s)s ≤ .
4
If t ∈ [1/2, 1], we get
1 1
q(t) = t(1 − t) · t(2 − t) ≤ ·1= .
4 4
Thus, for x ∈ [x0 , x1 ], we prove the error estimate
h4 5
(G.24) |∂x (f (x) − H5 (x))| ≤ ||∂ f ||L∞ [x0 ,x2 ] .
96 x
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 167

x−x0
Next, we estimate ∂x H5 . Recall t = h . Using (G.17) and the chain rule, we get
t2 (t − 1)2 t2 (t − 1)2
H5 (x) = H4 (x) + (f (x2 ) − H4 (x2 )) = G4 (t) + (f (x2 ) − H4 (x2 )) ,
(G.25) 4 4
1 t2 (t − 1)2 1
∂x H5 (x) = (∂t G4 (t) + (f (x2 ) − H4 (x2 ))∂t ) , (I(t) + II(t)).
h 4 h
We estimate two parts separately. Using (G.20), we get
(G.26) I(t) , ∂t G4 = (l1 − l0 ) + (3t2 − 2t)(m1 − (l1 − l0 )) + (3t2 − 4t + 1)(m0 − (l1 − l0 )).
Note that l1 − l0 , m1 , m0 are approximations of hf 0 (x) and have cancellations. We discuss
different t ∈ [0, 1] to exploit the cancellations.
Denote a ∨ b = max(a, b). If t ≤ 31 , we get
I(t) = (4t − 3t2 )(l1 − l0 ) + (3t2 − 4t + 1)m0 + (2t − 3t2 )(l1 − l0 − m1 ).
Since t ≤ 31 , we get
2 1
4t − 3t2 > 0, 3t2 − 4t + 1 = (1 − 3t)(1 − t) ≥ 0, 2t − 3t2 = 3t( − t) ≤ ,
3 3
where the last inequality is equivalent to (t − 31 )2 ≥ 0. It follows
1 1
(G.27) |I(t)| ≤ |l1 −l0 |∨m0 (4t−3t2 +3t2 −4t+1)+ |l1 −l0 −m1 | = |l1 −l0 |∨m0 + |l1 −l0 −m1 |.
3 3
The estimate of t ∈ [2/3, 1] is similar by swapping t and 1 − t, m0 and m1 , p0 and p1 . We get
1
(G.28) |I(t)| ≤ |l1 − l0 | ∨ m1 + |l1 − l0 − m0 |.
3
For t ∈ [1/3, 2/3], we rewrite I(t) as follows
I(t) = l1 − l0 + (4t − 3t2 − 1)(l1 − l0 − m0 ) + (2t − 3t2 )(l1 − l0 − m1 ).
Since t ∈ [1/3, 2/3], we get
4t − 3t2 − 1 = (3t − 1)(1 − t) ≥ 0, 2t − 3t2 = t(2 − 3t) ≥ 0,
3 1
4t − 3t2 − 1 + 2t − 3t2 = 6t − 6t2 − 1 = 6t(1 − t) − 1 ≤ − 1 ≤ .
2 2
Thus, using 4t − 3t2 − 1 + 2t − 3t2 = 6t − 6t2 − 1 ≤ 12 , we obtain
1
(G.29) I(t) ≤ |l1 − l0 | + (|l1 − l0 − m0 | ∨ |l1 − l0 − m1 |).
2
Combining three cases (G.27)-(G.28), we obtain the bound for ∂t G4 .
For the second term in (G.25), using (G.54), we get
t2 (t − 1)2 1 1 1 1 2 1
(G.30) ∂t = t(t−1)(t− ), |t(t−1)(t− )| = |2t(1−2t)(2−2t)| ≤ · √ ≤ √ ,
4 2 2 8 8 3 3 12 3
which implies
1 1
(G.31) II(t) ≤ √ |f (x2 ) − H4 (x2 )|.
h 12 3h
Combining the estimates (G.27)-(G.31), we obtain the estimate for ∂x H5 .

G.4. Hermite interpolation in 2D. The estimate in 2D is more involved. Consider x0 < x1 <
x2 , y0 < y1 < y2 with x2 −x1 = x1 −x0 , y2 −y1 = y1 −y0 . The domain D can be decomposed into
4 blocks with size h1 × h2 . For the 5-th order interpolation, we assume (x, y) ∈ [x0 , x1 ] × [y0 , y1 ]
without loss of generality. For the 4-th order interpolation, we only need 1 block. Denote
D = [x0 , x2 ] × [y0 , y2 ], h1 = x1 − x0 , h2 = y1 − y0 ,
(G.32)
t = (x − x0 )h1−1 , s = (y − y0 )h−1
2 .
168 JIAJIE CHEN AND THOMAS Y. HOU

G.4.1. Estimate the L∞ norm. We first consider the estimate of ||f ||L∞ . We treat y as a param-
eter and use (G.17) to construct the 4-th and 5-th order interpolations in x: H4 (x, y), H5 (x, y).
Using the formula in (G.20), we have
t2 (t − 1)2
G5 (t, s) , H5 (x, y) = H4 (x, y) + (f (x2 , y) − H4 (x2 , y)) ,
4
(G.33) H4 (x, y) = f (x0 , y)(1 − t) + f (x1 , y)t + t2 (t − 1)(hf 0 (x1 , y) − (f (x1 , y) − f (x0 , y)))
+ (t − 1)2 t(hf 0 (x0 , y) − (f (x1 , y) − f (x0 , y))).
Using (G.19), we have the error bound in x
1
(G.34) |H5 (x, y) − f (x, y)| ≤ ||∂ 5 f ||L∞ (D) h51 .
1200 x
We need to further interpolate G5 (t, s), H5 (x, y) in the y direction. To achieve the overall
5-th order error, we do not need to apply a high order interpolation to each coefficient. In
particular, for the linear term, we apply the 5-th order Hermite interpolation to f (xi , y) in y
using f (xi , yj ), j = 0, 1, 2 and ∂y f (xi , yj ), j = 0, 1 for i = 0, 1 and denote it by Ai (y), i.e.
(G.35) Ai (y) = H5 (f (xi , ·), y0 , y1 , y2 )(y)
using the notation in (G.17). Applying (G.22), we have the error bound in y
1
(G.36) |f (xi , y) − Ai (y)| ≤ ||∂ 5 f ||L∞ (D) h52 .
1200 y
Denote
(G.37) Mi (y) , hf 0 (xi , y) − (f (x1 , y) − f (x0 , y)).
For Mi (y), i = 1, 2, it is of order h21 . Thus, we apply the cubic interpolation in Section G.1.2
to these functions in y direction on grids y0 , y1 , y2 and denote it by Qi (y). Using the notation
(G.6), we have
(G.38) Qi (y) = N2 (Mi , y0 , y1 , y2 )(y).
Applying (G.7), we have the error bound
h3 h3 h2
(G.39) |Mi (y) − Qi (y)| ≤ √2 ||∂y3 Mi (y))||L∞ ([y0 ,y2 ]) ≤ 2√1 ||∂x2 ∂y3 f ||L∞ (D) ,
9 3 18 3
where we have used the following estimate with c = a, b, g = ∂y3 f (·, y) in the last inequality
Z b Z b
0 0 0 00 (b − a)2 00
(b−a)g (c)−(g(b)−g(a)) = g (c)−g (s)ds ≤ ||g ||L∞ [a,b] |c−a|ds = ||g ||L∞ [a,b] .
a a 2
For the last term f (x2 , y) − H4 (x2 , y), it is already very small and of order h41 . We use first
order estimate and then (G.18) with x = x2 to bound it directly
h2
|f (x2 , y) − H4 (x2 , y)| ≤ max |f (x2 , yj ) − H4 (x2 , yj )| + |∂y (f (x2 , y) − H4 (x2 , y))|
j=0,1 2
(G.40)
h2 h41
≤ max |f (x2 , yj ) − H4 (x2 , yj )| + ||∂y ∂x4 f ||L∞ (D) .
j=0,1 2·6

Estimate the 2D interpolating polynomials for f . We obtain the 2D interpolating poly-

nomials in x, y for H4 (x, y) as follows
   
(G.41) P4 (x, y) = A0 (y)(1−t)+A1 (y)t + t2 (t−1)Q1 (y)+t(t−1)2 Q0 (y) = I(t, y)+II(t, y).

Combining (G.34)-(G.40) and using t + (1 − t) = 1, |t2 (1 − t)| + |(1 − t)2 t| = (1 − t)t ≤ 41 , we

have the following error bound
1 1
|P4 (x, y) − f (x, y)| ≤ max |f (x2 , yj ) − H4 (x2 , yj )| + √ h21 h32 ||∂x2 ∂y3 f ||L∞ (D)
64 j=0,1 72 3
(G.42)
1 h2 h41
+ (h51 ||∂x5 f ||L∞ (D) + h52 ||∂y5 f ||L∞ (D) ) + ||∂y ∂x4 f ||L∞ (D) ,
1200 768
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 169

1 1 1 1
where the constant 768 comes from (G.40) and 768 = 64 · 16 .
To estimate P4 (x, y) (G.41), we use estimates similar to (G.22). The estimate of the linear
part is trivial
|I(t, y)| ≤ max ||Ai (y)||L∞ [y0 ,y1 ] , y ∈ [y0 , y1 ].
i=0,1

Since Ai (y) (G.35) is the Hermite polynomial in y , we can use the method in Section G.3 to
estimate it. For II, following the derivations between (G.21) to (G.22) and by considering two
cases: t ≤ 12 and t > 21 , we obtain

4
|II(t, t)| ≤ max(|Q1 (y) − Q0 (y)|, |Q1 (y)|, |Q0 (y)|).
27
Since Qi , Q1 − Q0 are quadratic interpolating polynomials of Mi , M1 − M0 (G.37) on y0 , y1 , y2 ,
we can use (G.8) to estimate the L∞ [y0 , y1 ] norm.

G.4.2. Estimates of the ∂f (x, y) in 2D. Now, we consider how to estimate ∂f (x, y) using the
Hermite interpolation. We consider ∂x without loss of generality. Recall the notation (G.32).
We first fix y as a parameter and interpolate f (x, y) in x using the same method as (G.33).
Using the error estimate (G.24), we yield

h41 5
(G.43) |∂x (f (x, y) − H5 (x, y))| ≤ ||∂ f ||L∞ (D) .
96 x
Using the computation (G.25), (G.26), (G.30) in Section G.3.3 and the notation (G.16) for
mi , li , we have
1
h1 ∂x H5 (x, y) = h1 ∂x H4 (x, y) + (f (x2 , y) − H4 (x2 , y))t(t − 1)(t − ),
2
h1 ∂x H4 (x, y) = f (x1 , y) − f (x0 , y) + (3t2 − 2t)(hf 0 (x1 , y) − (f (x1 , y) − f (x0 , y)))
+ (3t2 − 4t + 1)(hf 0 (x0 , y) − (f (x1 , y) − f (x0 , y)))
= f (x0 , y) − f (x1 , y) + (3t2 − 2t)M1 (y) + (3t2 − 4t + 1)(M0 (y),

where t = x−x
h1 (G.32), and we have used (G.37) to simplify the presentation.
0

Next, we interpolate the above functions in y. We want to achieve an overall 4-th order
approximation for ∂x H4 (x, y). For f (x1 , y)−f (x0 , y), we use the 4-th order Hermite interpolation
in y based on the grid point values f (x1 , yj ) − f (x0 , yj ), ∂y (f (x1 , yj ) − f (x0 , yj )), j = 0, 1 and
denote it as B(y), i.e.

(G.44) B(y) , H4 (f (x1 , ·) − f (x0 , ·), y0 , y1 )

using the notation (G.17). By (G.19), we have the error estimate

(G.45)
B(y) f (x1 , y) − f (x0 , y) 1 f (x1 , y) − f (x0 , y) 1
− ≤ ∂y4 ≤ ∂ 4 ∂x f .
h1 h1 384 h1 ∞
L ([y0 ,y1 ]) 384 y L∞ (D)

For Mi , we apply the same quadratic interpolation Qi in y (G.38). It satisfies the error bound
(G.39).
For f (x2 , y) − H4 (x2 , y), we use the same estimate (G.40), which along with (G.30) implies
1 1 1
(f (x2 , y) − H4 (x2 , y))t(t − 1)(t − ) ≤ max |f (x2 , yj ) − H4 (x2 , yj )|
h1 2 12h1 j=0,1
(G.46)
h2 h31
+ √ ||∂y ∂x4 f ||L∞ (D) .
12 · 12 3
170 JIAJIE CHEN AND THOMAS Y. HOU

Estimate the 2D interpolating polynomials for ∂x f . Now, we construct the interpolating

polynomials for h1 ∂x H4 as follows
S4 (x, y) = B(y) + (3t2 − 2t)Q1 (y) + (3t2 − 4t + 1)Q0 (y).
Combining the estimate (G.45) and using the triangle inequality, we can estimate the error
1
h1 S4 (x, y) − f (x, y).
It remains to estimate S4 (x, y). We further decompose the above approximation as the linear
part and the nonlinear part. The linear pat is the main term, and we want to obtain a sharper
estimate. The nonlinear part is smaller, which will be estimated using the triangle inequality.
Since B is the 4 − th order Hermite interpolation in y (G.44), we can apply the decomposition
(G.20) to B. Since Qi is the quadratic interpolation of Mi (G.37), (G.38), we can apply the
decomposition (G.6) to Qi .
(G.47) S4 = Slin + Snlin , Sσ (y) , Bσ (y) + (3t2 − 2t)Q1,σ (y) + (3t2 − 4(t + 1))Q0,σ (y)
where σ ∈ {lin, nlin}, flin denotes the linear part, and fnlin denotes the nonlinear part.
Since Slin is linear in y and B(yj ) = f (x1 , yj ) − f (x0 , yj ), Qi (yj ) = Mi (yj ) for j = 0, 1 (the
interpolating polynomials agree with the functions on the grid points), we get

|Slin (y)| ≤ max Blin (y) + (3t2 − 2t)Q1,lin (yi ) + (3t2 − 4(t + 1))Q0,lin (yi ) ,
i=0,1

= max f (x1 , yi ) − f (x0 , yi ) + (3t2 − 2t)M1 (yi ) + (3t2 − 4(t + 1))M0 (yi ) .
i=0,1

Recall the definition of Mi in (G.37). The polynomials inside | · | is the 1D polynomial in t with
the same form as I(t) (G.26). Thus it can be estimated using the method in Section G.3.3 for
1D Hermite interpotion of the derivative.
For the nonlinear part,
(G.48) Snlin = Bnlin (y) + (3t2 − 2t)Q1,nlin (y) + (3t2 − 4t + 1)Q0,nlin (y),
we recall the definition of B (G.44) and Q (G.38). The estimate of Bnlin follows the method of
estimating II(t) in (G.20), (G.22) with li , mi replacing by
˜li = f (x1 , yi ) − f (x0 , yi ), m̃i = h2 ∂y (f (x1 , yi ) − f (x0 , yi )),
which gives
4
|Bnlin (y)| ≤ max(|m̃0 , m̃1 |, |m̃0 − (˜l1 − ˜l0 )|, |m̃1 − (˜l1 − ˜l0 )|).
27
The estimate of the Qnlin follows that in (G.8) and Section G.1.2, which gives
1
|Qi,nlin | ≤ |Mi (y0 ) − 2Mi (y1 ) + Mi (y2 )|.
8
Using (G.53) in Lemma G.3, for t ∈ [0, 1], we prove
|(3t2 − 2t)Q1,nlin (y) + (3t2 − 4t + 1)Q0,nlin (y)|
1
≤(|3t2 − 2t| + |3t2 − 4t + 1|) max |Qi,nlin (y)| ≤ max |Mi (y0 ) − 2Mi (y1 ) + Mi (y2 )|.
i=0,1 8 i=0,1
Thus, we yield the estimate of Snlin . Combining the above estimates of Slin and Snlin and using
(G.47), we obtain the estimate of S4 . We remark that one needs to further divide the above
bounds by h11 to get the bound for Sh14 .

G.5. Weighted estimates of a function using derivatives. In our estimate of the residual
error or some norms, we need to estimate F (x)ρ(x) near x = 0 with singular weight ρ rigorously.
In this section, we discuss how to use the estimate of the derivatives of F to estimate the weighted
norm of F . Note that the estimate of ∂xi ∂yj F can be established using the methods in Sections
G.1-G.4.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 171

−1/2
Typical behaviors of ρ near x = 0 is |x|−k , k = 2, 52 , 3 or |x|−k x1 . By decomposing ρ into
−j/2
the singular part |x|−i x1 and the regular part, we only need to estimate F |x|−i x−j/2 . Denote
Z x Z y
1 1
Ex (F, i)(x, y) , i+1 F (z, y)(x − z)i dz, Ey (F, i)(x, y) , i+1 F (x, z)(y − z)i dz.
x 0 y 0
Due to the Fubini’s Theorem, we get
Z xZ y
1
Ex (Ey (F, j), i) = Ey (Ex (F, i), j) = i+1 j+1 F (x − t)i (y − s)j dtds , Eij (F )(x, y).
x y 0 0
Using the piecewise bound of F , we can bound these functions easily.
For F (x, y) odd in x and ∇k F (0) = 0 for k ≤ 2, we have the following identities
Z y
F (x, y) = Fy (x, z)dz + F (x, 0)
Z0 x Z y Z x
1 x
Z
= Fxyy (t, s)(y − s)dtds + y Fxxy (t, 0)(x − t)dt + Fxxx (t, 0)(x − t)2 dt.
0 0 0 2 0
Using the average operators, we yield
x3
|F (x, y)| ≤ Exy (|Fxyy |, 0, 1)xy 2 + x2 yEx (|Fxxy |, 1) +
Ex (|Fxxx |, 2).
2
Denote β = arctan( xy ), r = (x2 + y 2 )1/2 . Using these estimates, for a + b = 3, b ≤ 1, we get
|F (x, y)| 1
≤ Ex (|Fxxx |, 1) cos3−b (β) + Exy (|Fxxy |, 0, 1) cos β 1−b sin2 β
(G.49) r a xb 2
+ Ex (|Fxxy |, 1) cos2−b sin β.
Since we have the bounds for these coefficients, e.g., Exy (|Fxxy |, 0, 1), by maximizing β ∈ [0, π/2],
we obtain the bounds for rF3 and r5/2Fx1/2 .
∂i ∂j F
Similarly, we can bound xrky , i + j + k ≤ 3. For odd F , we have
Z x
|F | = Fx (z, y)dz ≤ Ex (|Fx |, 0)x,
Z0 y Z xZ y Z x
(G.50)
|F | = Fy (x, z)dz + F (x, 0) = Fxy (t, s)dxdy + Fxx (t, 0)(x − t)dt
0 0 0 0
≤ Exy (|Fxy |, 0, 0)xy + Ex (|Fxx |, 1)(x, 0)x2 , ∇k F (0) = 0, k = 0, 1.
F F
Using estimate similar to (G.49), we can bound |x|2 , |x| .
1/2
∂xi F ∂xi F xj
Weighted derivatives. Similarly, we estimate |x|2 , |x|5/2 . For odd F with ∇k F = 0, k ≤ 2,
using (G.49) with F replaced by Fy , we get
|Fy | ≤ Exy (|Fxyy |, 0, 0)xy + Ex (|Fxxy , 1)(x, 0)x2 .
Then, we can use the method in (G.49) to estimate
|xj |α ∂xi F ∂x F
2+α
= (g(β))α i 2 , g(β) = cos β, or sin β.
|x| |x|
Fx Fy
We also need to estimate |x| and |x| . Using (G.50) with F replaced by Fy , we get
|Fy | ≤ Ex (|Fxy |, 0)x.
Fx
For |x| , we have two cases. If F (x, 0) ≡ 0, e.g., F satisfies the Dirichlet boundary condition,
we yield
|Fx | ≤ Ey (|Fxy , 0|)y.
Without the vanishing conditions, we require ∇F = 0 and yield
Z y Z y Z x
|Fx (x, y)| = Fxy (x, z)dz + Fx (x, 0) = Fxy (x, z)dz + Fxx (z, 0)dz
0 0 0
≤ Ey (|Fxy |, 0)y + Ex (|Fxx |, 0)(x, 0)x.
172 JIAJIE CHEN AND THOMAS Y. HOU

∂xi F
Then we apply the method in (G.49) to estimate |x| .

G.6. Hölder estimate of the functions. In the following two sections, we estimate the Hölder
seminorms [f ]C 1/2 or [f ]C 1/2 of some function f , e.g. f = (∂t − L)W c in (7.14), based on the
x y
∞
previous L estimates. We will develop two approaches.
1/2
Suppose that we have bounds for ∂x f, ∂y f and f . Firstly, we consider the Cx estimate. For
x1 < y1 and x2 = y2 , we have
Z y1
|f (x) − f (y)| 1/2 1
I= ≤ |x − y| |fx (z1 , x2 )|dz1 .
|x − y|1/2 |x − y| x1
We further bound the average of fx to obtain the first estimate. We have a second estimate
Z y1 Z y1
1 1/2 −1/2 1
|I| = fx (z1 , x2 )dz · 1/2
≤ ||f x x ||∞ z1 dz1 ·
x1 |x − y| x1 |x − y|1/2
1/2 1/2 √
y − x1 2 y1 − x1
≤ ||fx x1/2 ||∞ 2 1 1/2
= ||fx x1/2 ||∞ √ √ .
|x − y| x1 + y1
We also have a trivial L∞ estimate
1/2 1/2
−1/2 x1 + y1 2
|I| ≤ ||f x1 ||∞ , |I| ≤ ||f ||∞ .
|x − y|1/2 |x − y|1/2
Similar L∞ and Lipschitz estimates apply to ||f ||C 1/2 .
y
Near the origin, optimizing the above estimates, for x2 = y2 , we obtain
√
f (x) − f (y) 1/2 −1/2 −1 y1 − x1
1/2
≤ min(||fx x ||∞ 2t, ||f x1 ||∞ t ), t = √ √ .
|x − y| x1 + y1
In the Y −direction, x1 = y1 , x2 ≤ y2 , we use
Z y2
f (x) − f (y) 1 1/2 −1/2 1/2 |x2 − y2 |1/2
IY = ≤ |fy (x1 , z2 )||z| · |z| dz2 ≤ ||f y |x| ||∞ ,
|x − y|1/2 |x2 − y2 |1/2 x2 |x|1/2
1/2
2x1
IY ≤ ||f x−1/2 ||∞ .
|x2 − y2 |1/2
Since x1 ≤ |x|, the minimum of these two estimates are not singular near x = 0. In particular,
we optimize two estimates to estimate IY .
From the above estimates, to obtain sharp Hölder estimate of f , we estimate the piecewise
−1/2
bounds of f, f x1 , f |x|−1/2 , fx , fy , fx |x1 |1/2 , fy |x|1/2 , which are local quantities. These esti-
mates can be established using the methods in Sections G.1-G.4.
G.7. The second approach of Hölder estimate of a given function. We develop an
additional approach to estimate I(f ) = |f|x−z|
(x)−f (z)|
1/2 that is sharper if |x − z| is not small and f
is smooth. Suppose that we can evaluate the grid point value of f and have derivatives bound
of f .
We estimate I(f ) = |f|x−z|
(x)−f (z)|
1/2 for x ∈ [xl , xu ], z ∈ [zl , zu ]. Denote by fˆ the linear approxi-
mation of f with fˆ(xi ) = f (xi ) on the grid point xi . We have the following Lemma.
Lemma G.1. Suppose that f is linear on [xl , xu ], [zl , zu ] and xl ≤ xu ≤ zl ≤ zu . Then we have
|f (x) − f (z)| |f (xα ) − f (zβ )|
max = max .
x∈[xl ,xu ],z∈[zl ,zu ] |x − z|1/2 α,β∈{l,u} |xα − zβ |1/2

The above Lemma shows that for the linear interpolation of f , the maximum of the Holder
norm is achieved at the grid point.
Proof. Denote by M the right hand side in the Lemma. Clearly, it suffices to prove that the left
hand side is bounded by M . We fix x ∈ [xl , xu ], z ∈ [zl , zu ]. Suppose that
x = al xl + au xu , z = bl zl + bu zu , au + al = 1, bl + bu = 1,
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 173

for al , bl ∈ [0, 1]. Denote

mαβ = aα bβ , α, β ∈ {l, u}.
Since f (x) is linear on [xl , xu ] and [zl , zu ], we get
f (x) = al f (xl ) + au f (xu ), f (z) = bl f (zl ) + bu f (zu ).
For any function g linear on [xl , xu ], [zl , zu ], e.g., g(x) = 1, g(x) = x, g(x) = f (x), we have
X X
(G.51) g(x) = mαβ g(xα ), g(z) − g(x) = mαβ (g(zβ ) − g(xα )),
α,β∈{l,u} α,β∈{l,u}

Using the above identities and the triangle inequality and the definition of M , we yield
X X
|f (x) − f (z)| = mαβ (f (xα ) − f (zβ )) ≤ mαβ M |xα − zβ |1/2 .
α,β∈{l,u} α,β∈{l,u}

Using the Cauchy-Schwarz inequality, |xα − zβ | = zβ − xα and (G.51), we establish

X X X
|f (x) − f (z)| ≤ mαβ mαβ M |xα − zβ |1/2 = mαβ M |xα − zβ |1/2
α,β∈{l,u} α,β∈{l,u} α,β∈{l,u}
 X 1/2
=M mαβ (zβ − xα ) = M (z − x)1/2 .
α,β∈{l,u}

The desired result follows. 

We generalize Lemma G.1 to 2D as follows.
Lemma G.2. Let Ix = [xl , xu ], Iz = [zl , zu ], Iy = [yl , yu ] with xl ≤ xu ≤ zl ≤ zu . Suppose that
f is linear on Ix × Iy and Iz × Iy . Then we have
|f (x, y) − f (z, y)| |f (xα , yγ ) − f (zβ , yγ )|
max = max .
x∈Ix ,z∈Iz ,y∈Iy |x − z|1/2 α,β,γ∈{l,u} |xα − zβ |1/2
f (x,y)−f (z,y)
Proof. Note that the function I(x, z, y) = |x−z|1/2
is linear in y. We get
|I(x, z, y| = max(|I(x, z, yl )|, |I(x, z, yu )|).
Applying Lemma G.1 completes the proof. 
Let fˆ be the linear interpolation of f . Suppose that x ∈ Ix , z ∈ Iz , y ∈ Iy with xu ≤ zl .
Using the above estimates and notations, we can bound I(f ) as follows
|f (z, y) − f (x, y)| |fˆ(x, y) − f (x, y)| + |fˆ(z, y) − f (z, y)| |f (xα , yγ ) − f (zβ , yγ )|
I(f ) = ≤ + max
|x − z|1/2 |x − z|1/2 α,β,γ∈{l,u} |xα − zβ |1/2
 h2 2 2
hy h 
≤ x
||fxx ||Ix ×Iy + (||fyy ||Ix ×Iy + ||fyy ||Iz ×Iy ) + z ||fxx ||Iz ×Iy |x − z|−1/2 + M.
8 8 8
G.8. Estimate of some explicit polynomials. We use the following bounds for some poly-
nomials in the error estimate of the interpolation.
Lemma G.3. We have the following estimates
(b − a)2
(G.52) |(t − a)(t − b)| ≤ , t ∈ [a, b],
4
(G.53) |3t2 − 2t| + |3t2 − 4t + 1| ≤ 1, t ∈ [0, 1],
2
(G.54) |t(t − 1)(t − 2)| ≤ √ , t ∈ [0, 2],
3 3
4
(G.55) t(t − 1)2 ≤ , t ∈ [0, 1/2],
27
(G.56) |t(t − 1)(t − 2)(t − 3)| ≤ 1, t ∈ [0, 3],
1
(G.57) |t2 (t − 1)2 (t − 2)| ≤ , t ∈ [0, 1].
10
174 JIAJIE CHEN AND THOMAS Y. HOU

We will use the computer assisted method to verify the following estimate
X 1 1 31
|pi (x)| ≤ 1.28, max |(x − )(x − )(x − )| ≤ 0.04,
x∈[0,1] 32 2 32
1≤i≤3
1 1 31
where pi is the Lagrangian interpolating polynomials associated with ( 32 , 2 , 32 ).
Proof. The proof of (G.52) follows from the inequality of arithmetic and geometric means (AM-
GM) or a direct calculation.
For (G.53), firstly, we note that |a| + |b| = |a + b| or |a − b|. It suffices to prove |a + b| ≤ 1 and
|a − b| ≤ 1 for a = 3t2 − 2t, b = 3t2 − 4t + 1. Since t2 − t ∈ [−1/4, 0], 2t − 1 ∈ [−1, 1], we have
a + b = 6t2 − 6t + 1 ∈ [−1/2, 1], a − b = 2t − 1 ∈ [−1, 1],
which implies |a + b|, |a − b| ≤ 1. We prove the desired result.
Denote s = t − 1 ∈ [−1, 1]. Then using the AM-GM inequality, we have
1 2 1 2s2 + 2(1 − s2 ) 3 4
2s (1 − s2 )2 ≤ (
t2 (t − 1)2 (t − 2)2 = (t − 1)2 (t2 − 2t)2 = ) = .
2 2 3 27
Taking the squart root on both sides proves (G.54).
To prove (G.55), applying the AM-GM inequality, we get
1 1 2t + 2(1 − t) 3 4
t(1 − t)2 = 2t(1 − t)2 ≤ ( ) = .
2 2 3 27
Denote s = t(3 − t) ∈ [0, 94 ]. Then we have |s − 1|2 ∈ [0, 2] and
|(t − 1)(t − 2)t(t − 3)| = |s(t2 − 3t + 2)| = |(2 − s)s| = |1 − (s − 1)2 | ≤ 1,
which implies (G.56).
To prove (G.57), we use (G.52) with a = 0, b = 1 and (G.54) to obtain
1 2 1 1
|t2 (t − 1)2 (t − 2)| ≤ √ = √ < ,
43 3 6 3 10
√
where the last inequality follows from (6 3)2 = 108 > 100 = 102 . 

Acknowledgments. The research was in part supported by NSF Grants DMS-1907977

and DMS-2205590. We would like to acknowledge the generous support from Mr. K. C. Choi
through the Choi Family Gift Fund and the Choi Family Postdoc Gift Fund. We would also like
to thank Drs. Pengfei Liu and De Huang for a number of stimulating discussions in the early
stage of this project, and Dr. Tarek Elgindi for his suggestion to include the 3D Euler blowup
result in this paper. Part of the computation in this paper was performed using the Caltech
IMSS High Performance Computing Service. The support from its staff is greatly appreciated.

References
[1] Franck Barthe. On a reverse form of the brascamp-lieb inequality. Inventiones mathematicae, 134(2):335–
361, 1998.
[2] Tristan Buckmaster, Steve Shkoller, and Vlad Vicol. Formation of shocks for 2D isentropic compressible
Euler. Communications on Pure and Applied Mathematics.
[3] Tristan Buckmaster, Steve Shkoller, and Vlad Vicol. Formation of point shocks for 3D compressible Euler.
arXiv preprint arXiv:1912.04429, 2019.
[4] Roberto Castelli, Marcio Gameiro, and Jean-Philippe Lessard. Rigorous numerics for ill-posed PDEs: pe-
riodic orbits in the Boussinesq equation. Archive for Rational Mechanics and Analysis, 228(1):129–157,
2018.
[5] A Castro and D Córdoba. Infinite energy solutions of the surface quasi-geostrophic equation. Advances in
Mathematics, 225(4):1820–1829, 2010.
[6] Angel Castro, Diego Córdoba, and Javier Gómez-Serrano. Global smooth solutions for the inviscid sqg
equation. 2020.
[7] Angel Castro, Diego Córdoba, Javier Gómez-Serrano, and Alberto Martı́n Zamora. Remarks on geometric
properties of SQG sharp fronts and α-patches. arXiv preprint arXiv:1401.5376, 2014.
[8] Jiajie Chen. Singularity formation and global well-posedness for the generalized Constantin–Lax–Majda
equation with dissipation. Nonlinearity, 33(5):2502, 2020.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 175

[9] Jiajie Chen. On the regularity of the De gregorio model for the 3D Euler equations. To appear in J. Eur.
Math. Soc., arXiv preprint arXiv:2107.04777, 2021.
[10] Jiajie Chen. On the slightly perturbed De Gregorio model on S 1 . Arch. Ration. Mech. Anal., 241(3):1843–
1869, 2021.
[11] Jiajie Chen and Thomas Y Hou. On stability and instability of c1,α singular solutions to the 3D Euler and
2D Boussinesq equations. arXiv preprint: arXiv:2206.01296 [math.AP].
[12] Jiajie Chen and Thomas Y Hou. Finite time blowup of 2D Boussinesq and 3D Euler equations with C 1,α
velocity and boundary. Communications in Mathematical Physics, 383(3):1559–1667, 2021.
[13] Jiajie Chen, Thomas Y Hou, and De Huang. Asymptotically self-similar blowup of the Hou–Luo model for
the 3D Euler equations. arXiv preprint arXiv:2106.05422.
[14] Jiajie Chen, Thomas Y Hou, and De Huang. On the finite time blowup of the De Gregorio model for the
3D Euler equations. Communications on Pure and Applied Mathematics, 74(6):1282–1350, 2021.
[15] K Choi, TY Hou, A Kiselev, G Luo, V Sverak, and Y Yao. On the finite-time blowup of a 1D model for the
3D axisymmetric Euler equations. CPAM, 70(11):2218–2243, 2017.
[16] K Choi, A Kiselev, and Y Yao. Finite time blow up for a 1D model of 2D Boussinesq system. Comm. Math.
Phys., 334(3):1667–1679, 2015.
[17] P Constantin. On the Euler equations of incompressible fluids. Bulletin of the American Mathematical
Society, 44(4):603–621, 2007.
[18] P Constantin, C Fefferman, and AJ Majda. Geometric constraints on potentially singular solutions for the
3D Euler equations. Communications in Partial Differential Equations, 21(3-4), 1996.
[19] P Constantin, P. D. Lax, and A. Majda. A simple one-dimensional model for the three-dimensional vorticity
equation. CPAM, 38(6):715–724, 1985.
[20] Dario Cordero-Erausquin, Bruno Nazaret, and Cédric Villani. A mass-transportation approach to sharp
sobolev and gagliardo–nirenberg inequalities. Advances in Mathematics, 182(2):307–332, 2004.
[21] Diego Córdoba, Javier Gómez-Serrano, and Andrej Zlatoš. A note on stability shifting for the Muskat
problem, II: From stable to unstable and back to stable. Analysis & PDE, 10(2):367–378, 2017.
[22] Guy David and Jean-Lin Journé. A boundedness criterion for generalized calderón-zygmund operators.
Annals of Mathematics, pages 371–397, 1984.
[23] Guy David, Jean-Lin Journé, and Stephen Semmes. Opérateurs de calderón-zygmund, fonctions para-
accrétives et interpolation. Revista Matemática Iberoamericana, 1(4):1–56, 1985.
[24] S De Gregorio. On a one-dimensional model for the three-dimensional vorticity equation. Journal of Statis-
tical Physics, 59(5-6):1251–1263, 1990.
[25] S De Gregorio. A partial differential equation arising in a 1D model for the 3D vorticity equation. Mathe-
matical Methods in the Applied Sciences, 19(15):1233–1255, 1996.
[26] J Deng, TY Hou, and X Yu. Geometric properties and nonblowup of 3D incompressible Euler flow. Com-
munications in Partial Difference Equations, 30(1-2):225–243, 2005.
[27] Tarek M Elgindi. Finite-time singularity formation for C 1,α solutions to the incompressible Euler equations
on R3 . Annals of Mathematics, 194(3):647–727, 2021.
[28] Tarek M Elgindi, Tej-Eddine Ghoul, and Nader Masmoudi. On the stability of self-similar blow-up for C 1,α
solutions to the incompressible Euler equations on R3 . arXiv preprint arXiv:1910.14071, 2019.
[29] Tarek M Elgindi, Tej-eddine Ghoul, and Nader Masmoudi. Stable self-similar blow-up for a family of nonlocal
transport equations. Analysis & PDE, 14(3):891–908, 2021.
[30] Tarek M Elgindi and In-Jee Jeong. Finite-time singularity formation for strong solutions to the axi-symmetric
3D Euler equations. Annals of PDE, 5(2):1–51, 2019.
[31] Tarek M. Elgindi and In-Jee Jeong. On the effects of advection and vortex stretching. Archive for Rational
Mechanics and Analysis, Oct 2019.
[32] Tarek M Elgindi and In-Jee Jeong. Finite-time singularity formation for strong solutions to the Boussinesq
system. Annals of PDE, 6:1–50, 2020.
[33] Alberto Enciso, Javier Gómez-Serrano, and Bruno Vergara. Convexity of Whitham’s highest cusped wave.
arXiv preprint arXiv:1810.10935, 2018.
[34] Alessio Figalli, Francesco Maggi, and Aldo Pratelli. A mass transportation approach to quantitative isoperi-
metric inequalities. Inventiones mathematicae, 182(1):167–211, 2010.
[35] JD Gibbon. The three-dimensional Euler equations: Where do we stand? Physica D: Nonlinear Phenomena,
237(14):1894–1904, 2008.
[36] Javier Gómez-Serrano. Computer-assisted proofs in pde: a survey. SeMA Journal, 76(3):459–484, 2019.
[37] Javier Gómez-Serrano and Rafael Granero-Belinchón. On turning waves for the inhomogeneous Muskat
problem: a computer-assisted proof. Nonlinearity, 27(6):1471, 2014.
[38] Vu Hoang, Betul Orcan-Ekmekci, Maria Radosz, and Hang Yang. Blowup with vorticity control for a 2D
model of the Boussinesq equations. Journal of Differential Equations, 264(12):7328–7356, 2018.
[39] Vu Hoang and Maria Radosz. Singular solutions for nonlocal systems of evolution equations with vorticity
stretching. SIAM Journal on Mathematical Analysis, 52(2):2158–2178, 2020.
[40] T Y Hou. The potentially singular behavior of the 3D Navier–Stokes equations. Foundation of Computational
Mathematics, published online on 9/7/2022, DOI: https://doi.org/10.1007/s10208-022-09578-4, 2021.
176 JIAJIE CHEN AND THOMAS Y. HOU

[41] T Y Hou. Potential singularity of the 3D Euler equations in the interior domain. Foundation of Compu-
tational Mathematics, published online on 9/7/2022, DOI: https://doi.org/10.1007/s10208-022-09585-5,
2022.
[42] T Y Hou and D Huang. Potential singularity formation of 3D axisymmetric Euler equations with degenerate
variable viscosity coefficients. MMS, accepted, 2022, arXiv:2102.06663, 2021.
[43] T. Y. Hou and D Huang. A potential two-scale traveling wave asingularity for 3D incompressible Euler
equations. Physica D, 435:133257, 2022.
[44] TY Hou. Blow-up or no blow-up? a unified computational and analytic approach to 3D incompressible Euler
and Navier-Stokes equations. Acta Numerica, 18(1):277–346, 2009.
[45] TY Hou and C Li. Dynamic stability of the three-dimensional axisymmetric Navier-Stokes equations with
swirl. Communications on Pure and Applied Mathematics, 61(5):661–697, 2008.
[46] TY Hou and R Li. Dynamic depletion of vortex stretching and non-blowup of the 3D incompressible Euler
equations. Journal of Nonlinear Science, 16(6):639–664, 2006.
[47] Carlos E Kenig and Frank Merle. Global well-posedness, scattering and blow-up for the energy-critical,
focusing, non-linear Schrödinger equation in the radial case. Inventiones mathematicae, 166(3):645–675,
2006.
[48] A Kiselev and V Sverak. Small scale creation for solutions of the incompressible two dimensional Euler
equation. Annals of Mathematics, 180:1205–1220, 2014.
[49] Alexander Kiselev. Small scales and singularity formation in fluid dynamics. In Proceedings of the Interna-
tional Congress of Mathematicians, volume 3, 2018.
[50] Alexander Kiselev and Changhui Tan. Finite time blow up in the hyperbolic Boussinesq system. Adv. Math.,
325:34–55, 2018.
[51] Laurent Lafleche, Alexis F Vasseur, and Misha Vishik. Instability for axisymmetric blow-up solutions to
incompressible Euler equations. Journal de Mathématiques Pures et Appliquées, 155:140–154, 2021.
[52] Michael J. Landman, George C. Papanicolaou, Catherine Sulem, and Pierre-Louis Sulem. Rate of blowup
for solutions of the nonlinear Schrödinger equation at critical dimension. Phys. Rev. A (3), 38(8):3837–3843,
1988.
[53] Pengfei Liu. Spatial Profiles in the Singular Solutions of the 3D Euler Equations and Simplified Mod-
els. PhD thesis, California Institute of Technology, 2017. https://resolver.caltech.edu/CaltechTHESIS:
09092016-000915850.
[54] G Luo and TY Hou. Toward the finite-time blowup of the 3D incompressible Euler equations: a numerical
investigation. SIAM Multiscale Modeling and Simulation, 12(4):1722–1776, 2014.
[55] Guo Luo and Thomas Y Hou. Potentially singular solutions of the 3D axisymmetric Euler equations. Pro-
ceedings of the National Academy of Sciences, 111(36):12968–12973, 2014.
[56] AJ Majda and AL Bertozzi. Vorticity and incompressible flow, volume 27. Cambridge University Press,
2002.
[57] Yvan Martel, Frank Merle, and Pierre Raphaël. Blow up for the critical generalized Korteweg–de Vries
equation. I: Dynamics near the soliton. Acta Mathematica, 212(1):59–140, 2014.
[58] Nader Masmoudi and Hatem Zaag. Blow-up profile for the complex Ginzburg–Landau equation. Journal of
Functional Analysis, 255(7):1613–1666, 2008.
[59] Alan McIntosh. Algèbres d’opérateurs définis par des intégrales singulières. CR Acad. Sci. Paris Sér. I
Math., 301:395–397, 1985.
[60] David W. McLaughlin, George C. Papanicolaou, Catherine Sulem, and Pierre-Louis Sulem. Focusing singu-
larity of the cubic Schrödinger equation. Phys. Rev. A, 34(2):1200, 1986.
[61] Frank Merle and Pierre Raphael. The blow-up dynamic and upper bound on the blow-up rate for critical
nonlinear Schrödinger equation. Annals of mathematics, pages 157–222, 2005.
[62] Frank Merle, Pierre Raphaël, Igor Rodnianski, and Jeremie Szeftel. On blow up for the energy super critical
defocusing nonlinear Schrödinger equations. Inventiones mathematicae, 227(1):247–413, 2022.
[63] Frank Merle, Pierre Raphaël, Igor Rodnianski, and Jeremie Szeftel. On the implosion of a compressible fluid
i: smooth self-similar inviscid profile. Ann. of Math. (2), 196(2):567–778, 2022.
[64] Frank Merle, Pierre Raphaël, Igor Rodnianski, and Jeremie Szeftel. On the implosion of a compressible fluid
ii: singularity formation. Ann. of Math. (2), 196(2):779–889, 2022.
[65] Frank Merle and Hatem Zaag. Stability of the blow-up profile for equations of the type ut = ∆u + |u|p−1 u.
Duke Math. J, 86(1):143–195, 1997.
[66] Frank Merle and Hatem Zaag. On the stability of the notion of non-characteristic point and blow-up profile
for semilinear wave equations. Communications in Mathematical Physics, 333(3):1529–1562, 2015.
[67] Ramon E Moore, R Baker Kearfott, and Michael J Cloud. Introduction to interval analysis, volume 110.
Siam, 2009.
[68] H Okamoto, T Sakajo, and M Wunsch. On a generalization of the constantin–lax–majda equation. Nonlin-
earity, 21(10):2447–2461, 2008.
[69] Siegfried M Rump. Verification methods: Rigorous results using floating-point arithmetic. Acta Numerica,
19:287–449, 2010.
[70] Steven Schochet. Explicit solutions of the viscous model vorticity equation. Communications on pure and
applied mathematics, 39(4):531–537, 1986.
 STABLE BLOWUP OF 2D BOUSSINESQ EQUATIONS 177

[71] Jack Sherman and Winifred J Morrison. Adjustment of an inverse matrix corresponding to a change in one
element of a given matrix. The Annals of Mathematical Statistics, 21(1):124–127, 1950.
[72] Alexis F Vasseur and Misha Vishik. Blow-up solutions to 3D Euler are hydrodynamically unstable. Com-
munications in Mathematical Physics, 378(1):557–568, 2020.
[73] Cédric Villani. Optimal transport: old and new, volume 338. Springer, 2009.
[74] Cédric Villani. Topics in optimal transportation, volume 58. American Mathematical Soc., 2021.
[75] Yongji Wang, Ching-Yao Lai, Javier Gomez-Serrano, and Tristan Buckmaster. Self-similar blow-up profile
for the Boussinesq equations via a physics-informed neural network. arXiv:2201.06780v1 [math.AP], 2022.

Courant Institute and Applied and Computational Mathematics, Caltech, Pasadena, CA 91125.
Emails: [email protected], [email protected],