betaFeedback

WorksMathematics

Dynamics of density patches in infinite Prandtl number convection

Abstract

This work examines the dynamics of density patches in the 2D zero-diffusivity Boussinesq system modified such that momentum is in a large Prandtl number balance. We establish the global well-posedness of this system for compactly supported and bounded initial densities, and then examine the regularity of the evolving boundary of patch solutions. For k∈{0,1,2}k \in \{0,1,2\}, we prove the global in time persistence of Ck+μC^{k+μ}-regularity, where μ∈(0,1)μ\in (0,1), for the density patch boundary via estimates of singular integrals. We conclude with a simulation of an initially circular density patch via a level-set method. The simulated patch boundary forms corner-like structures with growing curvature, and yet our analysis shows the curvature will be bounded for all finite times.

Introduction

In Earth’s mantle and many highly-pressurized gases, the rate of thermal diffusion κ\kappa is negligible compared to the rate of momentum dissipation ν\nu. We will analyze the dynamics of density patches, modeling idealized plumes, in such fluids where the nondimensional Prandtl number Pr⁡=ν/κ\operatorname{Pr} = \nu/\kappa is large.

Many scenarios of convection, including those which concern the present study, are well-modeled by the Boussinesq system:

{(∂∂t+u⋅∇)θ=κΔθ,(∂∂t+u⋅∇)u=νΔu−∇Π+θe2,div⁡u=0,(1)\begin{cases} \left(\frac{\partial}{\partial t} + u \cdot\nabla\right)\theta= \kappa\Delta\theta,\\ \left(\frac{\partial}{\partial t} + u \cdot\nabla\right)u = \nu\Delta u - \nabla\Pi+ \theta e_2,\\ \operatorname{div}u = 0, \tag*{(1)} \end{cases}

with initial temperature and velocity (θ0,u0):R2→R×R2(\theta_0,u_0): \mathbb{R}^2 \to\mathbb{R} \times\mathbb{R}^2. The unit vector e2=(0,1)e_2 = (0,1) is the vertical direction antiparallel to gravity, and the pressure Π\Pi enforces incompressibility at each instance of time.

For the system (B), global in time regularity has been well established for fixed positive Prandtl number with κ,ν\kappa,\nu both positive (see e.g. [1]). The global regularity for κ,ν\kappa,\nu both zero (degenerate Pr) remains unknown.

Known is the long time persistence of regularity for the partial viscosity scenarios—κ\kappa positive with zero ν\nu, and ν\nu positive with zero κ\kappa—given (θ0,u0)∈Hm(R2)(\theta_0,u_0)\in H^m(\mathbb{R}^2) with integer m>2m>2 ([2, 3]). For initial data of such regularity, [2] also provides the negative answer to XXI Century Problem 3 (see [4]) concerning the development of ∇θ\nabla\theta singularities in the limit κ→0\kappa\to0 with fixed ν\nu positive (formally Pr→∞\mathrm{Pr}\to\infty).

The question of well-posedness and global in time regularity for less regular initial data has been actively studied (e.g. [5, 6, 7]) for the zero diffusivity system

{system (B) with κ=0 and ν positive}.(2)\{\text{system }(B)\text{ with }\kappa=0\text{ and }\nu\text{ positive}\}. \tag*{(2)}

Without thermal diffusivity, convective plumes are idealized as density distributions initially of the form

θ0=1P0,(3)\theta_0=\mathbf{1}_{P_0}, \tag*{(3)}

where P0⊂R2P_0\subset\mathbb{R}^2 is simply connected and bounded. If u0u_0 is sufficiently regular, then the regularity results in [2] guarantee a unique solution to system (B_1) of the form

θ(t)=1P(t),(4)\theta(t)=\mathbf{1}_{P(t)}, \tag*{(4)}

where u(t)u(t) is regular enough to define the flow map X(⋅,t):R2→R2X(\cdot,t):\mathbb{R}^2\to\mathbb{R}^2 such that the evolving region P(t)=X(P0,t)P(t)=X(P_0,t) remains simply connected and bounded. However, inferring the regularity of the evolving boundary ∂P(t)\partial P(t) in time is a classical problem. Given k∈Z+={1,2,…}k\in\mathbb{Z}_+=\{1,2,\ldots\} and μ∈(0,1)\mu\in(0,1), we write ∂P(t)∈Ck+μ\partial P(t)\in C^{k+\mu} if there exists some z:S1→R2z:S^1\to\mathbb{R}^2 such that

∂P(t)={z(α)∣α∈S1}andz∈Ck+μ(S1).\partial P(t)=\{z(\alpha)\mid\alpha\in S^1\}\quad\text{and}\quad z\in C^{k+\mu}(S^1).

The question of persistence of regularity for the patch boundary is precisely whether ∂P(t)∈Ck+μ\partial P(t)\in C^{k+\mu} given ∂P(0)=∂P0∈Ck+μ\partial P(0)=\partial P_0\in C^{k+\mu}. This line of inquiry has its origin in the study of vortex patches in the two-dimensional incompressible Euler system. At some point, numerical studies like [8, 9] were done which had conflicting conclusions on whether finite time contour singularities formed in the evolution of a vortex patch. Then, [10] showed global in time persistence of smoothness using paradifferential calculus and striated regularity methods to obtain the desired regularity of ∇u\nabla u; soon after, [11] proved a similar result using geometric analysis techniques: the vortex patch boundary remains C1+μC^{1+\mu} for μ∈(0,1)\mu\in(0,1), given it was so initially. Another proof of this was given by [12].

More recently for the system (B_1), the global persistence of C1+μC^{1+\mu}-regularity ([13]), C2+μC^{2+\mu}-regularity ([14]), and Ck+μC^{k+\mu}-regularity for all k∈Z+k\in\mathbb{Z}_+ ([15]) was shown for density patches. The k=1k=1 and k∈Z+k\in\mathbb{Z}_+ results made special use of striated regularity estimates in Besov spaces.

The present study connects this inquiry of contour singularity formation to the program of XXI Century Problem 3. In the limiting dynamics of large Prandtl number, the fluid velocity is at relative equilibrium in the thermal diffusive time scale τκ=L2/κ\tau_\kappa= L^2/\kappa, where LL is the length scale. The nondimensional momentum equation is

1Pr⁡(∂∂t∗+u⋅∇∗)u=Δ∗u−∇∗Π+Ra⁡ θe2,(5)\frac{1}{\operatorname{Pr}}\left(\frac{\partial}{\partial t^*}+u\cdot\nabla^*\right)u=\Delta^*u-\nabla^*\Pi+\operatorname{Ra}\,\theta e_2, \tag*{(5)}

where the starred derivatives have been rescaled by τκ\tau_\kappa and LL, and the variables with underbars are the nondimensional counterparts to (u,θ,Π)(u,\theta,\Pi). The Rayleigh number Ra⁡\operatorname{Ra} is independent to Pr⁡\operatorname{Pr}. Thus formally the material derivative in (*) vanishes in the limit of large Prandtl number; specifically, in thermal diffusive time t∗=t/τκt^*=t/\tau_\kappa, the fluid is in equilibrium dominated by viscosity. Accordingly, we consider the (dimensional) zero diffusivity system (B1)(B_1) with the modified momentum equation

−Δu+∇Π=θe2.(6)-\Delta u+\nabla\Pi=\theta e_2. \tag*{(6)}

We thus consider patch dynamics in the system given by

{(∂∂t+u⋅∇)θ=0,−Δu+∇Π=θe2,div⁡u=0,(7)\begin{cases} \left(\dfrac{\partial}{\partial t}+u\cdot\nabla\right)\theta=0,\\ -\Delta u+\nabla\Pi=\theta e_2,\\ \operatorname{div}u=0, \tag*{(7)} \end{cases}

with initial data θ0:R2→R\theta_0:\mathbb{R}^2\to\mathbb{R}. Above, the solution θ(x,t)\theta(x,t) is an active scalar function of x∈R2x\in\mathbb{R}^2 and t∈Rt\in\mathbb{R}, where the corresponding velocity u(t):R2→R2u(t):\mathbb{R}^2\to\mathbb{R}^2 is given by the balance (6) at each instance of time.

The system (B_*) is known also as transport-Stokes, and can model the sedimentation of inertialess particles in Stokes flow [16, 17]. Let us briefly recount the literature concerning global well-posedness of this system in unbounded domains. Global well-posedness for L∞L^\infty initial data holds on the infinite strip R×(0,1)\mathbb{R}\times(0,1) with no-slip boundary conditions and a flux condition imposed [18]. On the whole space R3\mathbb{R}^3, the analogue of system (B_*) is globally well-posed for initial data in L1∩L∞L^1\cap L^\infty [19]; however, whether this is true in 2d without additional assumptions is unknown. Recent discussions of the well-posedness problem may be found in [20, 21].

The question of regularity for an interface of density in the transport-Stokes system was first addressed in [22]. On bounded domains in Rd\mathbb{R}^d with d≥2d\ge2, [22] establishes global well-posedness for piecewise constant initial densities, where the boundary conditions are no-slip, and then shows that interfaces which are initially C1+μC^{1+\mu}, will remain so for all time. More recently, [23] considers a contour dynamics equation derived from system (B_*) on the horizontally periodic strip T×R\mathbb{T}\times\mathbb{R}, wherein interfaces are shown to be locally well-posed in C1+μC^{1+\mu}, even in the Rayleigh-Taylor unstable regime.

Our contributions concern the system (B∗)(B_\ast) on the whole space R2\mathbb{R}^2. We first establish the global well-posedness of the system for initial data of Yudovich-type, meaning when θ0∈L1∩L∞\theta_0 \in L^1 \cap L^\infty is compactly supported. In particular, initial data (3) admit patch solutions (4) and we may study the regularity of ∂P(t)\partial P(t) in time. Then, we prove the global persistence of Ck+μC^{k+\mu} regularity for ∂P\partial P when k∈{0,1,2}k \in\{0,1,2\}. Further, we simulate on T2\mathbb{T}^2 the dynamics of an initially circular patch, and observe the interface develop corner-like structures with growing curvature. The results here rule out the possibility of a finite-time curvature singularity; however, corner formation at infinite time remains possible.

The content is organized as follows. The main contributions are presented in § 2. The key Lemmas 1 and 2 are proved, giving the details of the relevant singular integrals, in § 3. The global well-posedness of the system (B∗)(B_\ast), in particular Theorem 3 concerning the Yudovich class, is addressed in § 4. In § 5, the main Theorem 6 on the persistence of regularity for the evolving boundary ∂P(t)\partial P(t) is finally proved. In Section 6, the numerical simulation in Figure 1 is described in detail.

Figure 1. Snapshots of ∂P\partial P from a numerical patch solution (N=1024N = 1024) to system (B∗)(B_{\ast}) on T2\mathbb{T}^{2} for ∂P0=S1(12)\partial P_{0} = S^{1}\left(\frac{1}{2}\right). For presentation, the axes are omitted and the vertical positions of curve ∂P(t)\partial P(t) have been aligned between snapshots. Simulation details are given in Section 6.

Main Results

We suppose an initial distribution (3) in the Cauchy problem for system (B∗)(B_\ast) and show the ensuing solution velocity u(t)u(t) is regular enough such that the solution temperature has patch representation (4), where the evolving region

P(t)=X(P0,t)(8)P(t) = X(P_0,t) \tag*{(8)}

is given by the flow map X(⋅,t)X(\cdot,t) of our fluid velocity.

We start by introducing the streamfunction ψ\psi such that u=∇⊥ψu = \nabla^\perp\psi with ∇⊥=(−∂2,∂1)\nabla^\perp= (-\partial_2,\partial_1), and then the momentum equation in (B∗)(B_\ast) yields the following expression for vorticity ω≡∇⊥⋅u\omega\equiv\nabla^\perp\cdot u at each instance of time:

−Δω=∂1θ.(9)-\Delta\omega= \partial_1\theta. \tag*{(9)}

The fundamental solution of the bilaplacian Δ2\Delta^2 in R2\mathbb{R}^2 is

18π∣z∣2(log⁡∣z∣−1)(10)\frac{1}{8\pi}|z|^2(\log|z|-1) \tag*{(10)}

where z∈R2z \in\mathbb{R}^2. Given the fact Δψ=ω\Delta\psi= \omega generally, and assuming θ\theta in equation (9) has enough decay at infinity, it follows ψ=−(Δ2)−1∂1θ\psi= -(\Delta^2)^{-1}\partial_1\theta such that ψ=K∗θ\psi= K \ast\theta with kernel

K(z)=−z14π(log⁡∣z∣−12).(11)K(z) = -\frac{z_1}{4\pi}\left(\log|z|-\frac{1}{2}\right). \tag*{(11)}

More precisely, we find the expression for vorticity

ω(x)=−12π∫R2x1−y1∣x−y∣2θ(y) dy,(12)\omega(x) = -\frac{1}{2\pi}\int_{\mathbb{R}^2}\frac{x_1-y_1}{|x-y|^2}\theta(y)\,dy, \tag*{(12)}

and note that the integral converges absolutely for θ\theta in Y≡L1(R2)∩L∞(R2)Y \equiv L^1(\mathbb{R}^2) \cap L^\infty(\mathbb{R}^2). Denoting the integral as ω=Q∗θ\omega= Q * \theta, this is a convolution with a singular kernel which is homogeneous of degree −1-1. Thus, the operator (Q∗):Y→Lp(Q *) : Y \to L^p giving vorticity from temperature is bounded for 2<p≤∞2 < p \leq\infty; we give a direct proof of this claim in § 3.

Let G(z)≡∇⊥K(z)G(z) \equiv\nabla^\perp K(z), then the integral u=G∗θu = G * \theta, explicitly

u(x)=∫R2G(x−y)θ(y) dy,(13)u(x) = \int_{\mathbb{R}^2} G(x-y)\theta(y)\,\mathrm{d}y, \tag*{(13)}

converges absolutely if θ∈Y\theta\in Y is compactly supported. More quantitatively:

Lemma 1. Suppose θ∈L∞\theta\in L^\infty with spt⁡θ⊂B(0,R)\operatorname{spt}\theta\subset B(0,R). Let u=G∗θu = G * \theta. Then,

∣u(x)∣≤C(1+log⁡(∣x∣+R+1))∥θ∥Y.|u(x)| \leq C(1+\log(|x|+R+1))\|\theta\|_Y.

From here, we find the gradient as ∇u=∇G∗θ\nabla u = \nabla G * \theta, and further, we compute the principal value of the second gradient of velocity ∇∇u(x)\nabla\nabla u(x). With the kernel ∇∇G:R2→R2×2×2\nabla\nabla G : \mathbb{R}^2 \to\mathbb{R}^{2\times2\times2}, we have

∇∇u(x)=θ(x)E+14πpv⁡∫R2θ(x−z)∇∇G(z) dz.(14)\nabla\nabla u(x) = \theta(x)\mathbf{E} + \frac{1}{4\pi}\operatorname{pv}\int_{\mathbb{R}^2}\theta(x-z)\nabla\nabla G(z)\,\mathrm{d}z. \tag*{(14)}

The entries of the E∈R2×2×2\mathbf{E} \in\mathbb{R}^{2\times2\times2}, as well as GG and its gradients, ∇G\nabla G and ∇∇G\nabla\nabla G, are given in § 3. Of special note here is that the integral operator given by kernel ∇∇G\nabla\nabla G is Calderón-Zygmund in each entry, thus bounded on LpL^p for p∈(1,∞)p \in(1,\infty). Combining this with the previous observations, we deduce:

Lemma 2. Suppose θ∈Y\theta\in Y. Let ∇u=∇G∗θ\nabla u = \nabla G * \theta. Then for μ∈(0,1)\mu\in(0,1), there exists constant CμC_\mu depending only on μ\mu such that

∥∇u∥Cμ≤Cμ∥θ∥Y.\|\nabla u\|_{C^\mu} \leq C_\mu\|\theta\|_Y.

Initial data which satisfy the hypotheses of both Lemmas 1 and 2, namely compactly supported and bounded functions, are called Yudovich-type due to the classical result [24] establishing the global well-posedness of weak solutions within this class for the two-dimensional Euler system. This result allowed the study of long-time dynamics for vortex patches, in particular the program of vortex patch boundary regularity. For the system (B∗)(B_*), the initial data of our concern are clearly Yudovich-type, and the Lemmas are powerful enough to establish the following:

Theorem 3. Let θ0∈L1∩L∞\theta_0 \in L^1 \cap L^\infty and spt⁡θ0⊂B(0,R0)\operatorname{spt}\theta_0 \subset B(0,R_0). Then for arbitrary TT, the system (B∗)(B_*) has a unique weak solution

θ∈L∞(0,T;L1∩L∞).\theta\in L^\infty(0,T;L^1 \cap L^\infty).

Moreover, for t≤Tt \leq T, the following estimates hold:

  1. For 1≤p≤∞1 \le p \le\infty,

∥θ(t)∥Lp≤∥θ0∥Lp\|\theta(t)\|_{L^p} \le\|\theta_0\|_{L^p}
  1. For R(0)=R0R(0)=R_0, we have

spt⁡θ(t)⊂B(0,R(t))\operatorname{spt}\theta(t) \subset B(0,R(t))

with R(t)R(t) obeying the differential inequality

dR(t)dt≤C(1+log⁡(2R(t)+1))∥θ0∥Y\frac{dR(t)}{dt} \le C(1+\log(2R(t)+1))\|\theta_0\|_Y
  1. For μ∈(0,1)\mu\in(0,1),

∥∇u(t)∥Cμ≤Cμ∥θ0∥Y\|\nabla u(t)\|_{C^\mu} \le C_\mu\|\theta_0\|_Y

where u(t)=G∗θ(t)u(t)=G * \theta(t).

Corollary 4. Let θ0=1P0\theta_0=\mathbf{1}_{P_0} where P0⊂R2P_0 \subset\mathbb{R}^2 is simply connected and bounded. Then, the unique weak solution θ\theta to system (B∗)(B_*) has the form

θ(t)=1P(t)\theta(t)=\mathbf{1}_{P(t)}

where P(0)=P0P(0)=P_0 and P(t)P(t) is simply connected and bounded.

The corollary is due to an explicit construction: we represent the dynamics of the distribution 1P(t)\mathbf{1}_{P(t)} via the evolution of a smooth, compactly supported level-set function φ(t)∈Cc∞\varphi(t) \in C^\infty_c. Specifically, we suppose φ0∈Cc∞\varphi_0 \in C^\infty_c such that

P0={x∣φ0(x)>0}and∂P0={x∣φ0(x)=0},(15)P_0=\{x \mid\varphi_0(x)>0\} \quad\text{and} \quad\partial P_0=\{x \mid\varphi_0(x)=0\}, \tag*{(15)}

for P0P_0 in the claim. Then, we let φ\varphi be transported passively by u(t)=G∗θ(t)u(t)=G * \theta(t) where θ(t)∈Y\theta(t) \in Y is the unique weak solution with initial data 1P0\mathbf{1}_{P_0} guaranteed to exist by Theorem 3. Since u(t)u(t) is locally bounded and ∇u(t)∈Cμ\nabla u(t) \in C^\mu for all time, we have φ\varphi as the unique global solution the Cauchy problem given by

(∂∂t+u⋅∇)φ=0,(16)\left(\frac{\partial}{\partial t}+u\cdot\nabla\right)\varphi=0, \tag*{(16)}

with initial data φ0\varphi_0. Further, we have explicitly φ(t)=φ0∘X−1(⋅,t)\varphi(t)=\varphi_0\circ X^{-1}(\cdot,t) where XX is the well-defined flow map given by uu. It follows that the patch representation θ(t)=1P(t)\theta(t)=\mathbf{1}_{P(t)} holds for all time, where P(t)=X(P0,t)P(t)=X(P_0,t) is defined by

P(t)={x∣φ(x,t)>0}and∂P(t)={x∣φ(x,t)=0}.(17)P(t)=\{x \mid\varphi(x,t)>0\} \quad\text{and} \quad\partial P(t)=\{x \mid\varphi(x,t)=0\}. \tag*{(17)}

With global well-posedness of patches solutions to system (B∗)(B_*) established, we may now examine the question of global in time regularity for the evolving boundary ∂P(t)\partial P(t). We show first the local propogation of Hölder continuity:

Corollary 5. Suppose that ∂P0\partial P_0 is μ\mu-Hölder continuous at x~\tilde{x} with μ∈(0,1)\mu\in(0,1). Then, ∂P(t)\partial P(t) is μ\mu-Hölder continuous at X(x~,t)X(\tilde{x},t) at any given time.

This result follows from a Lagrangian construction of ∂P(t)\partial P(t). Rather than the construction (17), we consider some initial parametrization z0:S1↦∂P0\mathbf{z}_0:S^1\mapsto\partial P_0. We observe that the flow map X(a,t)X(a,t) for fluid velocity u=G∗1P\mathbf{u}=G*1_P then gives us the evolving patch boundary

∂P(t)={X(z0(α),t)∣α∈S1},(18)\partial P(t)=\{X(\mathbf{z}_0(\alpha),t)\mid\alpha\in S^1\}, \tag*{(18)}

such that

z(t)=X(⋅,t)∘z0(19)\mathbf{z}(t)=X(\cdot,t)\circ\mathbf{z}_0 \tag*{(19)}

parametrizes ∂P(t)\partial P(t) for any and all time. We note that for all times the image of z(t)\mathbf{z}(t) must be in

B(0,R(t))={∣x∣≤R(t)}(20)B(0,R(t))=\{|\mathbf{x}|\le R(t)\} \tag*{(20)}

wherein u(t)\mathbf{u}(t) is C1+μ(B(0,R(t)))C^{1+\mu}(B(0,R(t))), by Theorem 3. Thus if z0\mathbf{z}_0 is μ\mu-Hölder continuous at α~\tilde{\alpha}, then the composition of maps giving z(t)\mathbf{z}(t) is μ\mu-Hölder continuous at α~\tilde{\alpha}. Arranging z0(α~)=x~\mathbf{z}_0(\tilde{\alpha})=\tilde{x}, we have established Corollary 5.

Let us briefly demonstrate the difficulty in proving the persistence of higher regularity in the Lagrangian construction by examining the contour dynamics equation (CDE) for system (B∗)(B_*), which we proceed to derive.

By its construction (19), z(t)\mathbf{z}(t) obeys the evolution equation

dzdt=u(z(t),t).(21)\frac{d\mathbf{z}}{dt}=\mathbf{u}(\mathbf{z}(t),t). \tag*{(21)}

Suppose z0\mathbf{z}_0 is a parametrization by arc length, such that z(t)\mathbf{z}(t) is also. Examining the expression for velocity u(t)=∇⊥K∗1P(t)\mathbf{u}(t)=\nabla^\perp K*1_{P(t)}, we apply Green’s theorem to discover

u(x,t)=∫P(t)∇⊥K(x−y) dy=−∫02πK(x−z(σ,t))∂z∂α(σ,t) dσ,(22)\begin{aligned} \mathbf{u}(x,t)&=\int_{P(t)}\nabla^\perp K(x-y)\,dy\\ &=-\int_0^{2\pi}K(x-\mathbf{z}(\sigma,t))\frac{\partial\mathbf{z}}{\partial\alpha}(\sigma,t)\,d\sigma, \tag*{(22)} \end{aligned}

where z(t)\mathbf{z}(t) is assumed to be clockwise oriented. Suppressing the time argument, we now have the contour dynamics equation

∂z∂t(α)=−∫02πK(z(α)−z(σ))∂z∂α(σ) dσ.\frac{\partial\mathbf{z}}{\partial t}(\alpha)=-\int_0^{2\pi}K(\mathbf{z}(\alpha)-\mathbf{z}(\sigma))\frac{\partial\mathbf{z}}{\partial\alpha}(\sigma)\,d\sigma.

So, we differentiate the above equation in α\alpha to find the evolution equation

∂2z∂t∂α(α)=−∫02π[∇K(z(α)−z(σ))⋅∂z∂α(α)]∂z∂α(σ) dσ(23)\frac{\partial^2\mathbf{z}}{\partial t\partial\alpha}(\alpha)=-\int_0^{2\pi}\left[\nabla K(\mathbf{z}(\alpha)-\mathbf{z}(\sigma))\cdot\frac{\partial\mathbf{z}}{\partial\alpha}(\alpha)\right]\frac{\partial\mathbf{z}}{\partial\alpha}(\sigma)\,d\sigma \tag*{(23)}

where explicitly

∇K(z)=18π[(−2log⁡∣z∣0)−1∣z∣2(z12−z222z1z2)].(24)\nabla K(z) = \frac{1}{8\pi}\left[\begin{pmatrix}-2\log|z| & 0\end{pmatrix} - \frac{1}{|z|^2}\begin{pmatrix}z_1^2-z_2^2 \\ 2z_1z_2\end{pmatrix}\right]. \tag*{(24)}

From here, we see that proving the global in time well-posedness of the CDE by itself is difficult, even more so is showing the global in time persistence of CμC^\mu regularity for ∂αz(t)\partial_\alpha z(t) evolving via (23). This challenge is typical of contour dynamics models.

Theorem 6. Let k∈{0,1,2}k \in\{0,1,2\} and suppose that ∂P0∈Ck+μ\partial P_0 \in C^{k+\mu} with μ∈(0,1)\mu\in(0,1). Then, ∂P(t)∈Ck+μ\partial P(t) \in C^{k+\mu} for all time.

The higher regularity results follow from an analysis of the Eulerian construction of ∂P\partial P in Corollary 4. Recalling the level set function φ\varphi, we note the direction of the vector field W=∇⊥φ\mathbf{W} = \nabla^\perp\varphi is tangent to ∂P=∂P(t)\partial P = \partial P(t) in general. We then have the parametrization

z:S1↦∂Pwith∂z∂α=W∘z,(25)\mathbf{z}: \mathbb{S}^1 \mapsto\partial P \quad\text{with} \quad\frac{\partial\mathbf{z}}{\partial\alpha} = \mathbf{W} \circ\mathbf{z}, \tag*{(25)}

given W\mathbf{W} is bounded and non-vanishing everywhere on ∂P\partial P. Thus to guarantee ∂P∈C1+μ\partial P \in C^{1+\mu} for μ∈(0,1)\mu\in(0,1), the relevant quantities to control are ∥W(t)∥L∞\|\mathbf{W}(t)\|_{L^\infty} and

∣W∣inf⁡:=inf⁡x∈∂P∣W(x)∣=inf⁡x,φ(x)=0∣W(x)∣,(26)|\mathbf{W}|_{\inf} := \inf_{x\in\partial P} |\mathbf{W}(x)| = \inf_{x,\varphi(x)=0} |\mathbf{W}(x)|, \tag*{(26)}
∣W∣μ:=sup⁡x≠x′∣W(x)−W(x′)∣∣x−x′∣μ,μ∈(0,1).(27)|\mathbf{W}|_\mu:= \sup_{x\ne x'}\frac{|\mathbf{W}(x)-\mathbf{W}(x')|}{|x-x'|^\mu}, \quad\mu\in(0,1). \tag*{(27)}

Differentiating (16), the evolution of W(t)\mathbf{W}(t) where W0=∇⊥φ0\mathbf{W}_0 = \nabla^\perp\varphi_0 obeys

(∂∂t+u⋅∇)W=∇uW.(28)\left(\frac{\partial}{\partial t}+u\cdot\nabla\right)\mathbf{W} = \nabla u\mathbf{W}. \tag*{(28)}

As transport preserves ∥θ(t)∥Lp\|\theta(t)\|_{L^p}, the estimate of Lemma 2 gives immediately ∥∇u(t)∥L∞\|\nabla u(t)\|_{L^\infty} is bounded uniformly in time by a positive constant CLC_L depending only on the initial data. This is already sufficient to achieve ∂P∈C1+μ\partial P \in C^{1+\mu} with an exp⁡(CLexp⁡(CL∣t∣))\exp(C_L\exp(C_L|t|)) bound for ∣W(t)∣μ|\mathbf{W}(t)|_\mu from Grönwall-type inequalities, see [11 Proposition 3].

With the geometrical insights of [11], we may improve this initial bound for ∣W(t)∣μ|\mathbf{W}(t)|_\mu to an exp⁡(CL∣t∣)\exp(C_L|t|) bound. Further, these geometrical methods are used to prove persistence of C2+μC^{2+\mu}-regularity. Differentiating (28) yields an evolution equation,

(∂∂t+u⋅∇)∇W=[∇W,∇u]+∇∇u⋅W,(29)\left(\frac{\partial}{\partial t}+u\cdot\nabla\right)\nabla\mathbf{W} = [\nabla\mathbf{W},\nabla u] + \nabla\nabla u\cdot\mathbf{W}, \tag*{(29)}

where we have the tensor product commutator [A,B]=A⋅B−B⋅A[A,B] = A\cdot B - B\cdot A. The term to control is the product ∇∇u⋅W\nabla\nabla u\cdot\mathbf{W}, therein we observe the expression of ∇∇u\nabla\nabla u for the patch θ=1P\theta= \mathbf{1}_P is

∇∇u(x)=1P(x)E+14πp.v.⁡∫R21P(x−z)∇∇G(z) dz.(30)\nabla\nabla u(x) = \mathbf{1}_P(x)E + \frac{1}{4\pi}\operatorname{p.v.}\int_{\mathbb{R}^2}\mathbf{1}_P(x-z)\nabla\nabla G(z)\,dz. \tag*{(30)}

Clearly estimating ∥∇∇u∥μ\lVert\nabla\nabla u\rVert_\mu from this expression is difficult, but we use the fact that the vector field W∈Cμ(R2,R2)\mathbf{W}\in C^\mu(\mathbb{R}^2,\mathbb{R}^2) is divergence-free and tangent to ∂P\partial P to reduce the problem to estimating ∥∇∇u∥L∞\lVert\nabla\nabla u\rVert_{L^\infty} (Corollary 23).

Beyond satisfying the cancellation property, the CZ kernel ∇∇G(z)\nabla\nabla G(z) has reflection symmetry such that we have H:S1→R2×2×2\mathbf{H}:\mathbb{S}^1\to\mathbb{R}^{2\times2\times2} and

∇∇G(z)=H(z)∣z∣2withH(−z)=H(z).(31)\nabla\nabla G(z)=\frac{\mathbf{H}(z)}{|z|^2}\quad\text{with}\quad\mathbf{H}(-z)=\mathbf{H}(z). \tag*{(31)}

Because small neighborhoods of ∂P\partial P look like half-circles, the reflection symmetry allows us to achieve a sufficient L∞L^\infty estimate on ∇∇u(t)\nabla\nabla u(t) in Proposition 25. Ultimately, we show ∣∇W(t)∣μ\lvert\nabla\mathbf{W}(t)\rvert_\mu may grow like ∣t∣exp⁡(C3∣t∣)\lvert t\rvert\exp(C_3\lvert t\rvert), where C3C_3 depends only on μ\mu and the initial data, and the global in time persistence of C2+μC^{2+\mu} regularity for ∂P(t)\partial P(t) is proved. The complete details are given in § 5.

While the simulations of [8] and [9] were initialized with two circular vortex patches of positive sign, a single circular patch of density is already unsteady in system (B∗)(B_*). Observe that for the initial data θ0=1B(0,1)\theta_0=\mathbf{1}_{B(0,1)}, the initial vorticity ω0\omega_0 evaluated on the contour z0(α)=(sin⁡α,cos⁡α)z_0(\alpha)=(\sin\alpha,\cos\alpha),

ω0∘z0=∫02π∫01H(2rsin⁡(α+θ)−r2)cos⁡θ dr dθ(32)\omega_0\circ z_0=\int_0^{2\pi}\int_0^1 H(2r\sin(\alpha+\theta)-r^2)\cos\theta\,dr\,d\theta \tag*{(32)}

where HH is the Heaviside function, is not symmetric in α\alpha. Thus the initially circular patch will not remain circular. More generally, for a nonnegative and nontrivial θ0\theta_0, the center of mass

q=1M∫R2θx dx(33)q=\frac{1}{M}\int_{\mathbb{R}^2}\theta x\,dx \tag*{(33)}

where M=∥θ0∥L1M=\lVert\theta_0\rVert_{L^1}, drifts upwards with vertical component q2q_2 increasing as

dq2dt=−1M∫R2div⁡(θu)x2 dx=1M∫R2θu2 dx>0(34)\frac{dq_2}{dt}=-\frac{1}{M}\int_{\mathbb{R}^2}\operatorname{div}(\theta u)x_2\,dx=\frac{1}{M}\int_{\mathbb{R}^2}\theta u_2\,dx>0 \tag*{(34)}

because u2=G2∗θu_2=G_2*\theta is given by a positive operator on θ\theta. If the initial data has a horizontal reflection symmetry, the horizontal component q1q_1 is stationary.

Let us consider the system (B∗)(B_*) on T2\mathbb{T}^2 such that the spatial domain is now computationally realizable. Observing that the problem is Galilean invariant, we choose a reference frame so that the mean velocity ∫T2u dx=uˉ(0)\int_{\mathbb{T}^2}u\,dx=\bar{u}(0) is zero. Then from integrating the momentum equation, we see that uˉ(0)\bar{u}(0) is indeed constant in time, and that we must have ∫T2θ dx=0\int_{\mathbb{T}^2}\theta\,dx=0 for compatibility. Clearly, the arguments giving the main results, in particular Lemmas 1 and 2, hold readily on the domain T2\mathbb{T}^2 wherein L1∩L∞≡L∞L^1\cap L^\infty\equiv L^\infty.

Corollary 7. Let θ0∈L∞(T2)\theta_0\in L^\infty(\mathbb{T}^2) such that

∫T2θ0 dx=0.\int_{\mathbb{T}^2}\theta_0\,dx=0.

Then, Theorem 3 holds for the system (B∗)(B_*) on T2\mathbb{T}^2.

Our definition of patches must be modified to be mean zero in density. On the flat torus, the compatible patch data is

θ0=1P0−area⁡(P0)(35)\theta_{0} = \mathbf{1}_{P_{0}} - \operatorname{area}(P_{0}) \tag*{(35)}

where P0⊂T2P_{0} \subset\mathbb{T}^{2} is simply connected. For such initial data, the unique weak solution θ\theta to system (B∗)(B_{\ast}) on T2\mathbb{T}^{2} then has the form

θ(t)=1P(t)−area⁡(P0),(36)\theta(t) = \mathbf{1}_{P(t)} - \operatorname{area}(P_{0}), \tag*{(36)}

where P(0)=P0P(0) = P_{0} and P(t)P(t) is simply connected. On this numerically tractable domain, we here implement a level-set method that approximates the evolution of the boundary ∂P(t)\partial P(t) which we know is well-defined for all time.

Corollary 8. For the system (B∗)(B_{\ast}) on T2\mathbb{T}^{2}, if the initial data θ0\theta_{0} has the form (35), then the unique weak solution θ\theta in Corollary 7 has the form (36).

The numerical method evolves a discrete level-set function φij(t)\varphi_{ij}(t) on a fixed N×NN \times N uniform grid, discretizing T2\mathbb{T}^{2}, according to (16). The scheme for solving (16) is described in [25 Part II] and is first-order in space with respect to the uniform grid spacing h=1/Nh = 1/N. The discrete fluid velocity uiju_{ij} is obtained from (6) using a spectral collocation method, and time integration is done with Heun’s method, which is second-order and strong-stability preserving. Overall, the numerical solver for (B∗)(B_{\ast}) is second-order in time, first-order in space.

Using this algorithm, we compute the dynamics of the temperature patch which is initially circular. The initial curve ∂P0\partial P_{0} is fixed as the embedded circle of radius one-half S1(12)S^{1}\left(\frac{1}{2}\right). The result of the simulation is presented in Figure 1. We observe that the curve ∂P(t)\partial P(t) forms corner-like structures in its evolution; however, our persistence of regularity results, Corollary 5 and Theorem 6, hold for ∂P(t)\partial P(t) in this setting as well. Thus, if in light of Figure 1 one asks whether there develops a curvature singularity in finite time, we provide proof that the curvature remains bounded for all time.

Proof of Lemmas 1 and 2

The expressions with singular integrals allow us to sufficiently control uu and ω\omega via the bounds that follow.

Proof of Lemma 1. We write explicitly,

G(z)=∇⊥K(z)=18π(1−2log⁡∣z∣−2z^12)(37)G(z) = \nabla^\perp K(z) = \frac{1}{8\pi}\left(1 - 2\log|z| - 2\hat{z}_1^2\right) \tag*{(37)}

where z^j=zj/∣z∣\hat{z}_j = z_j/|z|, and observe

∣G(x−y)∣≤C(1+∣log⁡∣x−y∣∣).(38)|G(x-y)| \le C(1+|\log|x-y||). \tag*{(38)}

It follows immediately

∫∣x−y∣≤1∣G(x−y)θ(y)∣ dy≤C∥θ∥L∞.(39)\int_{|x-y|\le1}|G(x-y)\theta(y)|\,\mathrm{d}y \le C\|\theta\|_{L^\infty}. \tag*{(39)}

Further if ∣x∣≥R+1|x| \ge R+1, then 1≤∣x−y∣≤2∣x∣1 \le|x-y| \le2|x|, and we can bound ∣G(x−y)∣|G(x-y)| by C(1+log⁡(2∣x∣))C(1+\log(2|x|)). Now consider ∣x∣≤R+1|x| \le R+1 with ∣x−y∣≥1|x-y| \ge1, we have instead the bound C(1+log⁡(2R+1))C(1+\log(2R+1)). Combining these estimates we deduce,

∫∣x−y∣≥1∣G(x−y)θ(y)∣ dy≤C(1+log⁡(∣x∣+R+1))∥θ∥L1.(40)\int_{|x-y|\ge1}|G(x-y)\theta(y)|\,\mathrm{d}y \le C(1+\log(|x|+R+1))\|\theta\|_{L^1}. \tag*{(40)}

The result follows from (39) and (40). □\square

Proposition 9. Suppose θ∈Y\theta\in Y. Let ω=Q∗θ\omega= Q * \theta. Then for 2<p≤∞2 < p \le\infty,

∥ω∥Lp≤Cp∥θ∥Y.\|\omega\|_{L^p} \le C_p\|\theta\|_Y.

Proof. We split the integral

−2πω(x)=∫∣x−y∣≤1x1−y1∣x−y∣2θ(y) dy+∫∣x−y∣≥1x1−y1∣x−y∣2θ(y) dy(41)-2\pi\omega(x) = \int_{|x-y|\le1}\frac{x_1-y_1}{|x-y|^2}\theta(y)\,\mathrm{d}y + \int_{|x-y|\ge1}\frac{x_1-y_1}{|x-y|^2}\theta(y)\,\mathrm{d}y \tag*{(41)}

and so deduce

∣ω(x)∣≤22π(∥θ∥L1+4∥θ∥L∞).(42)|\omega(x)| \le\frac{\sqrt{2}}{2\pi}\left(\|\theta\|_{L^1}+4\|\theta\|_{L^\infty}\right). \tag*{(42)}

The case p=∞p = \infty follows.

Let B⊂R2B \subset\mathbb{R}^2 be the unit ball centered at the origin. Then we have

ω(x)=ω1B(x)+ω1R2∖B(x)(43)\omega(x) = \omega1_B(x) + \omega1_{\mathbb{R}^2\setminus B}(x) \tag*{(43)}

pointwise. From (42), it follows immediately that

∥ω1B∥Lp≤Cp∥θ∥Y.(44)\|\omega1_B\|_{L^p} \le C_p\|\theta\|_Y. \tag*{(44)}

and the convolution ω=Q∗θ\omega= Q * \theta converges absolutely. Accordingly, we find

ω1R2∖B=(Q1R2∖B)∗θ,(45)\omega1_{\mathbb{R}^2\setminus B} = (Q1_{\mathbb{R}^2\setminus B}) * \theta, \tag*{(45)}

where the truncated kernel is

Q1R2∖B(z)={Q(z),if z∈R2∖B,0,otherwise.(46)Q1_{\mathbb{R}^2\setminus B}(z) = \begin{cases} Q(z), & \text{if } z \in\mathbb{R}^2\setminus B,\\ 0, & \text{otherwise}. \end{cases} \tag*{(46)}

Let 2<p<∞2 < p < \infty, such that ∣Q1R2∖B∣p|Q1_{\mathbb{R}^2\setminus B}|^p is integrable. We conclude

∥ω1R2∖B∥Lp≤Cp∥θ∥L1(47)\|\omega1_{\mathbb{R}^2\setminus B}\|_{L^p} \le C_p\|\theta\|_{L^1} \tag*{(47)}

from Young’s inequality for convolutions (e.g. see [26 Appendix A]).

Proposition 10. Suppose θ∈Y\theta\in Y. Let ∇u=∇G∗θ\nabla u = \nabla G * \theta. Then for 2<p≤∞2 < p \le\infty,

∥∇u∥Lp≤Cp∥θ∥Y.\|\nabla u\|_{L^p} \le C_p\|\theta\|_Y.

Proof. The singular kernel GG away from the origin has gradient ∇G\nabla G, with homogeneity of degree −1-1, such that, for integrable and bounded θ\theta, we have the absolutely convergent integral

∇u(x)=∫R2∇G(x−y)θ(y) dy,(48)\nabla u(x) = \int_{\mathbb{R}^2} \nabla G(x-y)\theta(y)\,dy, \tag*{(48)}

where we compute explicitly as

∇G(z)=−14π∣z∣4(z2(z12−z22)z13+3z1z22z1(z22−z12)z2(z22−z12)).(49)\nabla G(z) = -\frac{1}{4\pi|z|^4} \begin{pmatrix} z_2(z_1^2-z_2^2) & z_1^3+3z_1z_2^2\\ z_1(z_2^2-z_1^2) & z_2(z_2^2-z_1^2) \end{pmatrix}. \tag*{(49)}

Each entry of ∇G\nabla G has the same cancellation property and homogeneity as QQ, so we conclude as in the proof of Proposition 9.

Proposition 11. Suppose θ∈Lp\theta\in L^p for some p∈(1,∞)p \in(1,\infty). Let u=G∗θu = G * \theta. Then,

∥∇∇u∥Lp≤Cp∥θ∥Lp.\|\nabla\nabla u\|_{L^p} \le C_p\|\theta\|_{L^p}.

:::`

Proof. For ∇∇u=(∂1∇u,∂2∇u)\nabla\nabla u = (\partial_1\nabla u,\partial_2\nabla u), we differentiate carefully to resolve the strongly singular kernel. In particular, let z=x−yz=x-y such that

∂1∇u(x)=∫R2∇G(z)∂1θ(x−z) dz(50)\partial_1\nabla u(x)=\int_{\mathbb{R}^2}\nabla G(z)\partial_1\theta(x-z)\,dz \tag*{(50)}
=lim⁡ε→0∫∣z∣≥ε−∂∂z1(∇G(z)θ(x−z))+θ(x−z)∂∂z1∇G(z) dz.(51)=\lim_{\varepsilon\to0}\int_{|z|\geq\varepsilon}-\frac{\partial}{\partial z_1}\bigl(\nabla G(z)\theta(x-z)\bigr)+\theta(x-z)\frac{\partial}{\partial z_1}\nabla G(z)\,dz. \tag*{(51)}

The first integral is

=lim⁡ε→0∫∣z∣=ε∇G(z)θ(x−z)(−n1)⋅dσ=lim⁡ε→01ε∫∣z∣=εz1∇G(z)θ(x−z) dσ=18(0−310)θ(x).(52)\begin{aligned} =\lim_{\varepsilon\to0}\int_{|z|=\varepsilon}\nabla G(z)\theta(x-z)(-n_1)\cdot d\sigma\\ &=\lim_{\varepsilon\to0}\frac{1}{\varepsilon}\int_{|z|=\varepsilon}z_1\nabla G(z)\theta(x-z)\,d\sigma\\ &=\frac{1}{8}\begin{pmatrix}0&-3\\1&0\end{pmatrix}\theta(x). \tag*{(52)} \end{aligned}

We compute

∂1∇G(z)=14πH1(z)∣z∣2and∂2∇G(z)=14πH2(z)∣z∣2(53)\partial_1\nabla G(z)=\frac{1}{4\pi}\frac{\mathbf{H}_1(z)}{|z|^2} \quad\text{and}\quad \partial_2\nabla G(z)=\frac{1}{4\pi}\frac{\mathbf{H}_2(z)}{|z|^2} \tag*{(53)}

where each entry of H:R2→R2×2×2\mathbf{H}:\mathbb{R}^2\to\mathbb{R}^{2\times2\times2}, explicitly

H1(z)=1∣z∣4(2z1z2(z12−3z22)z14+6z12z22−3z24z14−6z12z22+z242z1z2(3z22−z12)),(54)\mathbf{H}_1(z)=\frac{1}{|z|^4} \begin{pmatrix} 2z_1z_2(z_1^2-3z_2^2)&z_1^4+6z_1^2z_2^2-3z_2^4\\ z_1^4-6z_1^2z_2^2+z_2^4&2z_1z_2(3z_2^2-z_1^2) \end{pmatrix}, \tag*{(54)}

and

H2(z)=1∣z∣4(−z14+6z12z22−z242z2(3z1z22−z13)2z1z2(z22−3z12)z14−6z12z22+z24).(55)\mathbf{H}_2(z)=\frac{1}{|z|^4} \begin{pmatrix} -z_1^4+6z_1^2z_2^2-z_2^4&2z_2(3z_1z_2^2-z_1^3)\\ 2z_1z_2(z_2^2-3z_1^2)&z_1^4-6z_1^2z_2^2+z_2^4 \end{pmatrix}. \tag*{(55)}

is homogeneous of degree zero, mean zero on the unit sphere, and symmetric with respect to reflection. It follows

∂1∇u(x)=18(0−310)θ(x)+14πpv⁡∫R2H1(z)∣z∣2θ(x−z) dz,(56)\partial_1\nabla u(x)=\frac{1}{8}\begin{pmatrix}0&-3\\1&0\end{pmatrix}\theta(x)+\frac{1}{4\pi}\operatorname{pv}\int_{\mathbb{R}^2}\frac{\mathbf{H}_1(z)}{|z|^2}\theta(x-z)\,dz, \tag*{(56)}

and

∂2∇u(x)=18(100−1)θ(x)+14πpv⁡∫R2H2(z)∣z∣2θ(x−z) dz.(57)\partial_2\nabla u(x)=\frac{1}{8}\begin{pmatrix}1&0\\0&-1\end{pmatrix}\theta(x)+\frac{1}{4\pi}\operatorname{pv}\int_{\mathbb{R}^2}\frac{\mathbf{H}_2(z)}{|z|^2}\theta(x-z)\,dz. \tag*{(57)}

With the appropriate E∈R2×2×2\mathbf{E}\in\mathbb{R}^{2\times2\times2}, we may write compactly

∇∇u(x)=θ(x)E+14πpv⁡∫R2H(z)∣z∣2θ(x−z) dz.(58)\nabla\nabla u(x)=\theta(x)\mathbf{E}+\frac{1}{4\pi}\operatorname{pv}\int_{\mathbb{R}^2}\frac{\mathbf{H}(z)}{|z|^2}\theta(x-z)\,dz. \tag*{(58)}

We conclude by examining this expression as a sum of operators. The first summand is bounded like the identity, and the second is a Calderón–Zygmund operator. ∎

Corollary 12. Suppose θ∈Y\theta\in Y. Let ∇u=∇G∗θ\nabla u = \nabla G * \theta. Then for p∈(2,∞)p \in(2,\infty)

∥∇u∥W1,p≤Cp∥θ∥Y.\|\nabla u\|_{W^{1,p}} \le C_p\|\theta\|_Y.

Proof of Lemma 2. To deduce the desired Hölder continuity of ∇u\nabla u, we observe Corollary 12 and recall Morrey’s embedding: W1,p⊂CμW^{1,p} \subset C^\mu for μ=1−2/p\mu= 1 - 2/p with p>2p > 2.

Proof of Theorem 3

Let us first define what we mean by a (Yudovich) weak solution:

Definition 1. Let θ0∈Y\theta_0 \in Y, then θ(x,t)\theta(x,t) is a weak solution of system (B∗)(B_*) given θ∈L∞(0,T;L1∩L∞)\theta\in L^\infty(0,T;L^1 \cap L^\infty) for some T>0T > 0, and that for any ϕ∈C1(0,T;Cc1)\phi\in C^1(0,T;C^1_c), the following holds:

∫R2θ(x,T)ϕ(x,T) dx−∫R2θ0ϕ(x,0) dx=∫0T∫R2θ(∂∂t+u⋅∇)ϕ dx dt,\int_{\mathbb{R}^2} \theta(x,T)\phi(x,T)\,dx-\int_{\mathbb{R}^2}\theta_0\phi(x,0)\,dx=\int_0^T\int_{\mathbb{R}^2}\theta\left(\frac{\partial}{\partial t}+u\cdot\nabla\right)\phi\,dx\,dt,

where u(t)=G∗θ(t)u(t)=G * \theta(t).

Notice from the definition that Lemma 2 plays a powerful role in proving the well-posedness of weak solutions with compactly supported initial data. If we can maintain that θ(t)\theta(t) remains compactly supported in its evolution, then the characteristics for θ\theta have sufficient regularity to preserve LpL^p norms via the transport equation. With this observation, we obtain a sequence of smooth, compactly supported initial data θ0ϵ\theta_0^\epsilon with regularization parameter ϵ\epsilon. For such data, we obtain a family of global smooth solutions θϵ\theta^\epsilon with control of

spt⁡θϵ(t)⊂B(0,R(t))anduϵ(t)∈C1+μ(B(0,R(t)))(59)\operatorname{spt}\theta^\epsilon(t) \subset B(0,R(t)) \quad\text{and}\quad u^\epsilon(t) \in C^{1+\mu}(B(0,R(t))) \tag*{(59)}

for all time, depending only on ∥θ0∥Y\|\theta_0\|_Y. Then, the limit ϵ→0\epsilon\to0 yields a Yudovich weak solution via the classical arguments in [24]. The limiting solution θ\theta (possibly non-unique) inherits these uniform controls, and these controls are strong enough to achieve uniqueness. This section is dedicated to proving the claim of global well-posedness of classical solutions with uniform in ϵ\epsilon control of (59) (Proposition 18).

We begin by considering solutions to the sequence of linear equations

(∂∂t+un⋅∇)θn+1=0,(60)\left(\frac{\partial}{\partial t}+u^n\cdot\nabla\right)\theta^{n+1}=0, \tag*{(60)}

where un(t):=G∗θn(t)u^n(t):=G * \theta^n(t) for positive integers nn, defined inductively. The initial data are fixed as θn(0)=θ0\theta^n(0)=\theta_0 uniformly in nn, where

θ0∈H1∩Y,spt⁡θ0⊂B(0,R0).(61)\theta_0 \in H^1 \cap Y,\qquad\operatorname{spt}\theta_0 \subset B(0,R_0). \tag*{(61)}

For n=0n=0, take θ0(t)=θ0\theta^0(t)=\theta_0.

Proposition 13. The sequence of functions θn\theta^n described above is well-defined. Moreover, the following uniform estimates hold.

  1. For 1≤p≤∞1 \le p \le\infty,

∥θn(t)∥Lp≤∥θ0∥Lp.\|\theta^n(t)\|_{L^p} \le\|\theta_0\|_{L^p}.
  1. For R(0)=R0R(0)=R_0, we have

spt⁡θn(t)⊂B(0,R(t))\operatorname{spt}\theta^n(t) \subset B(0,R(t))

with R(t)R(t) obeying the differential inequality

dR(t)dt≤C(1+log⁡(2R(t)+1))∥θ0∥Y.\frac{dR(t)}{dt} \le C(1+\log(2R(t)+1))\|\theta_0\|_Y.
  1. For μ∈(0,1)\mu\in(0,1),

∥∇un(t)∥Cμ≤Cμ∥θ0∥Y\|\nabla u^n(t)\|_{C^\mu} \le C_\mu\|\theta_0\|_Y

where un(t)=G∗θn(t)u^n(t)=G*\theta^n(t).

Proof. For n=0n=0, all the estimates are immediate as θ0=θ0\theta^0=\theta_0. Now letting n≥0n \ge0, suppose θn\theta^n satisfies the estimates of the theorem.

For the first estimate, the regularity of unu^n in the linear transport equation (60) produces a unique global solution θn+1\theta^{n+1}, and ∥θn+1∥L∞\|\theta^{n+1}\|_{L^\infty} is preserved along characteristics. Further, we may multiply (60) by θ∣θ∣p−2\theta|\theta|^{p-2}, for p≥2p \ge2, and integrate by parts to discover

ddt∥θn+1∥Lp≤0,(62)\frac{d}{dt}\|\theta^{n+1}\|_{L^p} \le0, \tag*{(62)}

since unu^n is divergence-free. We deduce θn+1\theta^{n+1} also satisfies Estimate 1, as desired. For the second estimate, we observe un=G∗θnu^n=G*\theta^n implies

∣un(x,t)∣≤C(1+log⁡(∣x∣+R(t)+1))∥θn(t)∥Y.(63)|u^n(x,t)| \le C(1+\log(|x|+R(t)+1))\|\theta^n(t)\|_Y. \tag*{(63)}

Indeed θn∈Y\theta^n \in Y, and the previous arguments produce

∥θn+1(t)∥Y≤∥θ0∥Y.(64)\|\theta^{n+1}(t)\|_Y \le\|\theta_0\|_Y. \tag*{(64)}

Moreover, θn+1\theta^{n+1} is transported by unu^n, which we showed is bounded. Let X(a,t)X(a,t) be the flow map generated by unu^n with labels a∈R2a \in\mathbb{R}^2, such that

dX(a,t)dt=un(X(a,t),t)(65)\frac{dX(a,t)}{dt}=u^n(X(a,t),t) \tag*{(65)}

where X(a,0)=aX(a,0)=a. Observe

spt⁡θn+1(t)⊂{x∣x=X(a,s), s∈[0,t], a∈B(0,R0)},(66)\operatorname{spt}\theta^{n+1}(t) \subset\{x \mid x=X(a,s),\ s\in[0,t],\ a\in B(0,R_0)\}, \tag*{(66)}

whereby the inequality (65) gives

dR(t)dt≤sup⁡a∈B(0,R0)∣u(X(a,t),t)∣≤C(1+log⁡(2R(t)+1))∥θn(t)∥Y,(67)\frac{dR(t)}{dt} \le\sup_{a\in B(0,R_0)} |u(X(a,t),t)| \le C(1+\log(2R(t)+1))\|\theta^n(t)\|_Y, \tag*{(67)}

such that Estimate 1 implies θn+1\theta^{n+1} satisfies Estimate 2 in the Proposition. For Estimate 3, we may simply apply Lemma 2 and the previous conclusions to deduce

∥∇un+1(t)∥Cμ≤Cμ∥θ0∥Y(68)\|\nabla u^{n+1}(t)\|_{C^\mu} \le C_\mu\|\theta_0\|_Y \tag*{(68)}

for μ∈(0,1)\mu\in(0,1) where un+1(t)=G∗θn+1(t)u^{n+1}(t)=G*\theta^{n+1}(t). The results follow from induction. □

Proposition 14. The sequence of functions θn\theta^n satisfies the uniform estimate

∥∇θn(t)∥Lp≤∥∇θ0∥Lpexp⁡(CLt)\|\nabla\theta^n(t)\|_{L^p}\le\|\nabla\theta_0\|_{L^p}\exp(C_Lt)

for 1≤p≤∞1\le p\le\infty.

Proof. Differentiating the transport equation (60) yields

(∂∂t+un−1⋅∇)∇⊥θn=(∇un−1)∇⊥θn.(69)\left(\frac{\partial}{\partial t}+u^{n-1}\cdot\nabla\right)\nabla^\perp\theta^n=(\nabla u^{n-1})\nabla^\perp\theta^n. \tag*{(69)}

We then multiply the above equation by ∇⊥θn∣∇⊥θn∣p−2\nabla^\perp\theta^n|\nabla^\perp\theta^n|^{p-2} for p≥2p\ge2, and integrate by parts to find

ddt∥∇⊥θn∥Lpp≤p∫R2∣∇un−1∣∣∇θn∣p dx.(70)\frac{d}{dt}\|\nabla^\perp\theta^n\|_{L^p}^p\le p\int_{\mathbb{R}^2}|\nabla u^{n-1}||\nabla\theta^n|^p\,dx. \tag*{(70)}

Therefore

ddt∥∇θn∥Lp≤∥∇un−1∥L∞∥∇θn∥Lp,(71)\frac{d}{dt}\|\nabla\theta^n\|_{L^p}\le\|\nabla u^{n-1}\|_{L^\infty}\|\nabla\theta^n\|_{L^p}, \tag*{(71)}

for 1≤p≤∞1\le p\le\infty, where the cases p=1p=1 and p=∞p=\infty are direct from (69). From here, Grönwall’s Lemma yields

∥∇θn(t)∥Lp≤∥∇θ0∥Lpexp⁡(∫0t∥∇un−1(s)∥L∞ ds).(72)\|\nabla\theta^n(t)\|_{L^p}\le\|\nabla\theta_0\|_{L^p}\exp\left(\int_0^t\|\nabla u^{n-1}(s)\|_{L^\infty}\,ds\right). \tag*{(72)}

By the third estimate in Proposition 13, it follows ∇un−1\nabla u^{n-1} is bounded absolutely by a constant CLC_L which depends only on ∥θ0∥Y\|\theta_0\|_Y. □

Proposition 15. For any tt, the sequence θn(t)\theta^n(t) converges strongly in L2L^2 to some function θ(t)\theta(t). Moreover, θ\theta obeys Estimates 1, 2 and 3 in Theorem 3.

Proof. We denote

ϑn+1=θn+1−θn(73)\vartheta^{n+1} = \theta^{n+1} - \theta^n \tag*{(73)}

and conclude by showing the sequence ϑn(t)\vartheta^n(t) is summable in L2L^2 for any tt. Accordingly, we define vn=G∗ϑnv^n = G * \vartheta^n such that equation obeyed by ϑn+1\vartheta^{n+1} reads

∂tϑn+1+⟨u⟩n⋅∇ϑn+1+vn⋅∇⟨θ⟩n+1=0(74)\partial_t\vartheta^{n+1} + \langle u\rangle^n \cdot\nabla\vartheta^{n+1} + v^n \cdot\nabla\langle\theta\rangle^{n+1} = 0 \tag*{(74)}

where

⟨u⟩n=12(un+un−1)(75)\langle u\rangle^n = \frac{1}{2}\left(u^n + u^{n-1}\right) \tag*{(75)}

and

⟨θ⟩n+1=12(θn+1+θn).(76)\langle\theta\rangle^{n+1} = \frac{1}{2}\left(\theta^{n+1} + \theta^n\right). \tag*{(76)}

Taking the L2L^2 inner product of (74) with ϑn+1\vartheta^{n+1} gives us

ddt∥ϑn+1(t)∥L22=−2∫R2vn⋅∇⟨θ⟩n+1ϑn+1 dx(77)\frac{d}{dt}\|\vartheta^{n+1}(t)\|_{L^2}^2 = -2\int_{\mathbb{R}^2}v^n\cdot\nabla\langle\theta\rangle^{n+1}\vartheta^{n+1}\,dx \tag*{(77)}

after integration by parts, noting ⟨u⟩n\langle u\rangle^n is divergence-free. We estimate the convolution G∗ϑnG * \vartheta^n using the fact that the supports are in the ball of radius R(t)R(t). We obtain

∣vn(x,t)∣≤∥ϑn(t)∥L2(∫∣y∣≤R(t)∣G(x−y)∣2 dy)1/2(78)|v^n(x,t)| \le\|\vartheta^n(t)\|_{L^2}\left(\int_{|y|\le R(t)}|G(x-y)|^2\,dy\right)^{1/2} \tag*{(78)}

from the Cauchy-Schwarz inequality. Using Lemma 1 and evaluating the integral, we discover

∣vn(x,t)∣≤C(1+log⁡(2R(t)+1))R(t)∥ϑn(t)∥L2.(79)|v^n(x,t)| \le C(1+\log(2R(t)+1))R(t)\|\vartheta^n(t)\|_{L^2}. \tag*{(79)}

Now the righthand side of (77) has the bound,

∣−2∫R2vn⋅∇⟨θ⟩n+1ϑn+1 dx∣≤C(1+log⁡(2R(t)+1))R(t)∥∇⟨θ⟩n+1(t)∥L2∥ϑn(t)∥L2∥ϑn+1(t)∥L2.(80)\left|-2\int_{\mathbb{R}^2}v^n\cdot\nabla\langle\theta\rangle^{n+1}\vartheta^{n+1}\,dx\right| \le C(1+\log(2R(t)+1))R(t)\|\nabla\langle\theta\rangle^{n+1}(t)\|_{L^2}\|\vartheta^n(t)\|_{L^2}\|\vartheta^{n+1}(t)\|_{L^2}. \tag*{(80)}

The estimate then becomes,

ddtΘn+1(t)≤AT∥∇⟨θ⟩n+1∥L2Θn(t),(81)\frac{d}{dt}\Theta^{n+1}(t) \le A_T\|\nabla\langle\theta\rangle^{n+1}\|_{L^2}\Theta^n(t), \tag*{(81)}

where Θn=∥ϑn∥L2\Theta^n=\|\vartheta^n\|_{L^2}, and AT=C(1+log⁡(2R(T)+1))R(T)A_T=C(1+\log(2R(T)+1))R(T). Using Proposition 14 for p=2p=2, we write the integral expression for all t≤Tt\le T

Θn+1(t)≤AT∥∇θ0∥L2exp⁡(CLT)∫0tΘn(s) ds,(82)\Theta^{n+1}(t) \le A_T\|\nabla\theta_0\|_{L^2}\exp(C_L T)\int_0^t\Theta^n(s)\,ds, \tag*{(82)}

noting that Θn(0)=0\Theta^n(0)=0 for all nn. Denoting CT=AT∥∇θ0∥L2exp⁡(CLT)C_T=A_T\lVert\nabla\theta_0\rVert_{L^2}\exp(C_LT), we compute n=1n=1:

Θ2(t)≤CT∫0tΘ1(s) ds≤CT(sup⁡t≤TΘ1(t))t≤CTC1t(83)\begin{aligned} \Theta^2(t) &\le C_T\int_0^t\Theta^1(s)\,ds\\ &\le C_T\left(\sup_{t\le T}\Theta^1(t)\right)t\\ &\le C_TC_1t \tag*{(83)} \end{aligned}

where C1C_1 depends only on ∥θ0∥L2\lVert\theta_0\rVert_{L^2}. Suppose now for some nn,

Θn(t)≤C1(CTt)nn!,(84)\Theta^n(t)\le C_1\frac{(C_Tt)^n}{n!}, \tag*{(84)}

then we have

Θn+1(t)≤C1CTn+1∫0tsnn!≤C1(CTt)n+1(n+1)!.(85)\begin{aligned} \Theta^{n+1}(t)&\le C_1C_T^{n+1}\int_0^t\frac{s^n}{n!}\\ &\le C_1\frac{(C_Tt)^{n+1}}{(n+1)!}. \tag*{(85)} \end{aligned}

It follows by induction that estimate (84) holds for all n≥1n\ge1. Indeed,

∑n=1∞Θn+1(t)≤C1exp⁡(CTt)(86)\sum_{n=1}^{\infty}\Theta^{n+1}(t)\le C_1\exp(C_Tt) \tag*{(86)}

such that Θn(t)\Theta^n(t) is summable for all t≤Tt\le T, where TT is arbitrary. Therefore, the sequence θn(t)\theta^n(t) converges strongly in L2L^2 to a function θ(t)\theta(t) for any tt, and the uniform bounds in Proposition 13 are inherited by the limit.

In view of Lemma 1, the linear operator given by kernel GG is generally not bounded in L2L^2. Towards a convergence in more regular spaces, an immediate corollary of Proposition 13 is that ∥∇∇un∥L2≤C∥θ0∥Y\lVert\nabla\nabla u^n\rVert_{L^2}\le C\lVert\theta_0\rVert_Y, uniformly in nn. We further establish the higher regularity estimate:

Lemma 16. Suppose θ∈Y∩Hm\theta\in Y\cap H^m for some positive integer mm. Let u=G∗θu=G\ast\theta. Then,

∥Dm+2u∥L2≤∥θ∥Hm.\lVert D^{m+2}u\rVert_{L^2}\le\lVert\theta\rVert_{H^m}.

Proof. We first observe that u=G∗θu=G\ast\theta is a weak solution to

−Δu+∇Π=θe2,(87)-\Delta u+\nabla\Pi=\theta e_2, \tag*{(87)}

where Π\Pi is given by ΔΠ=∂2θ\Delta\Pi=\partial_2\theta. Then, we apply DmD^m to the above equation and take the L2L^2 inner product with Dm(−Δu)D^m(-\Delta u) to arrive at

∫R2(Dm(Δu))⋅Dm(Δu) dx=∫R2(Dm(θe2))⋅Dm(θe2) dx,(88)\int_{\mathbb{R}^2}(D^m(\Delta u))\cdot D^m(\Delta u)\,dx=\int_{\mathbb{R}^2}(D^m(\theta e_2))\cdot D^m(\theta e_2)\,dx, \tag*{(88)}

whereby integration by parts twice gives

∥Dm+2u∥L2=∥Dmθ∥L2.(89)\lVert D^{m+2}u\rVert_{L^2} = \lVert D^m\theta\rVert_{L^2}. \tag*{(89)}

For higher regularity initial data, we show the limit θ\theta indeed solves our system. The precompactness for θn\theta^n we require follows from HmH^m-energy estimates, where the H0H^0 estimate is immediate from Estimate 1 in Proposition 13.

Proposition 17. Let θ0∈Y∩Hm\theta_0 \in Y \cap H^m for some integer m>2m > 2. Then, the sequence θn\theta^n satisfies the uniform estimate

∥θn(t)∥Hm≤Cm∥θ0∥Hmexp⁡(Cmexp⁡(CL∣t∣)),\lVert\theta^n(t)\rVert_{H^m} \le C_m\lVert\theta_0\rVert_{H^m}\exp(C_m\exp(C_L|t|)),

where CmC_m depends only on mm and θ0\theta_0.

Proof. Further, we apply the operator DαD^\alpha with multi-index α\alpha to (60), multiply by Dαθn+1D^\alpha\theta^{n+1}, integrate over R2\mathbb{R}^2, and sum over ∣α∣≤m|\alpha| \le m to discover

12ddt∥θn+1∥Hm2=−∑∣α∣≤m∫R2(Dα(un⋅∇θn+1))Dαθn+1 dx\frac{1}{2}\frac{d}{dt}\lVert\theta^{n+1}\rVert_{H^m}^2 = -\sum_{|\alpha|\le m}\int_{\mathbb{R}^2}(D^\alpha(u^n\cdot\nabla\theta^{n+1}))D^\alpha\theta^{n+1}\,dx
=−∑∣α∣≤m∫R2(Dα(un⋅∇θn+1)−un⋅∇(Dαθn+1))Dαθn+1 dx.(90)= -\sum_{|\alpha|\le m}\int_{\mathbb{R}^2}(D^\alpha(u^n\cdot\nabla\theta^{n+1})-u^n\cdot\nabla(D^\alpha\theta^{n+1}))D^\alpha\theta^{n+1}\,dx. \tag*{(90)}

since unu^n is divergence-free. Using calculus inequalities (e.g. see [27 Lemma 3.4]) gives us that

12ddt∥θn+1∥Hm2≤∑∣α∣≤m∥Dα(un⋅∇θn+1)−un⋅∇(Dαθn+1)∥L2\frac{1}{2}\frac{d}{dt}\lVert\theta^{n+1}\rVert_{H^m}^2 \le\sum_{|\alpha|\le m}\lVert D^\alpha(u^n\cdot\nabla\theta^{n+1})-u^n\cdot\nabla(D^\alpha\theta^{n+1})\rVert_{L^2}
≤Cm(∥∇un∥L∞∥Dmθn+1∥L2+∥Dmun∥L2∥∇θn+1∥L∞).(91)\le C_m\left(\lVert\nabla u^n\rVert_{L^\infty}\lVert D^m\theta^{n+1}\rVert_{L^2}+\lVert D^m u^n\rVert_{L^2}\lVert\nabla\theta^{n+1}\rVert_{L^\infty}\right). \tag*{(91)}

With Proposition 14, Lemma 16, and the embedding Hm⊂L∞H^m \subset L^\infty for m>1m > 1, we find

ddt∥θn+1∥Hm≤Cm(∥Dmθn+1∥L2+exp⁡(CL∣t∣)∥θn∥Hm−1),(92)\frac{d}{dt}\lVert\theta^{n+1}\rVert_{H^m} \le C_m(\lVert D^m\theta^{n+1}\rVert_{L^2}+\exp(C_L|t|)\lVert\theta^n\rVert_{H^{m-1}}), \tag*{(92)}

where CmC_m depends only on mm and θ0∈Y∩Hm\theta_0 \in Y\cap H^m. The result follows from iteration.

Proposition 18. Let θ0∈Y∩Hm\theta_0 \in Y\cap H^m for some integer m>2m > 2, and suppose that spt⁡θ0⊂B(0,R0)\operatorname{spt}\theta_0 \subset B(0,R_0). Then for arbitrary TT, there exists a unique classical solution

θ∈C1(0,T;C01),\theta\in C^1(0,T;C^1_0),

to system (B∗)(B_*). Moreover, θ\theta satisfies Estimates 1, 2 and 3 in Theorem 3.

Proof. Because of Proposition 17, the sequence θn(t)\theta^{n}(t) has a subsequence θnj(t)\theta^{n_j}(t) which converges strongly in HmH^m to the limit θ(t)\theta(t) in Proposition 15, for 0≤t≤T0 \leq t \leq T. As m>2m > 2, the continuous embedding

Hm+k⊂C0k≡{space of Ck functions vanishing at infinity}(93)H^{m+k} \subset C_0^k \equiv\{\text{space of } C^k \text{ functions vanishing at infinity}\} \tag*{(93)}

implies θnj(t)\theta^{n_j}(t) converges strongly in C01C_0^1, and θ(t)∈C01\theta(t) \in C_0^1 satisfies Estimates 1, 2 and 3. Towards verifying θ\theta is a solution, we define u=G∗θu = G * \theta and observe that θnj∈Y\theta^{n_j} \in Y and θ∈Y\theta\in Y are both supported in B(0,R)B(0, R) where R=R(t)R = R(t) is given by Estimate 2. At each instance of time, Lemma 1 implies

∣unj(x)−u(x)∣≤C(1+log⁡(∣x∣+R+1))∥θnj−θ∥Y.(94)\left|u^{n_j}(x) - u(x)\right| \leq C(1 + \log(|x| + R + 1))\|\theta^{n_j} - \theta\|_Y. \tag*{(94)}

Then, we may conclude unj→uu^{n_j} \to u pointwise via the continuous embedding Lloc1⊂H1L^1_{\mathrm{loc}} \subset H^1. Estimate 3 and Azelà–Ascoli imply the convergence is uniform in B(0,R)B(0, R).

Similarly, we have also θnk−1\theta^{n_k-1} converging strongly to θ\theta in C01C_0^1 for each fixed time, where nk−1n_k - 1 indexes a subsequence of θnj−1\theta^{n_j-1}. It then follows that unk−1(t)⋅∇θnk(t)u^{n_k-1}(t) \cdot\nabla\theta^{n_k}(t) converges uniformly. Pointwise in space and time, we have

∂tθnk=−(unk−1)⋅∇θnk,(95)\partial_t\theta^{n_k} = -(u^{n_k-1}) \cdot\nabla\theta^{n_k}, \tag*{(95)}

and thus ∂tθnk\partial_t\theta^{n_k} converges strongly to −u⋅∇θ-u \cdot\nabla\theta in C(0,T;C0)C(0,T;C_0). The distributional limit of ∂tθnk\partial_t\theta^{n_k} is accordingly ∂tθ\partial_t\theta; therefore θ\theta is a classical solution to system (B∗)(B_*). □\square

Proof of Theorem 3. Existence with the estimates follows from Proposition 18 using the arguments in the classical paper of Yudovich [24], basically unmodified. Here, we prove uniqueness using the estimates from Lemma 2. Suppose with initial data θ0∈Y\theta_0 \in Y of compact support we have two weak solutions, θ1\theta_1 and θ2\theta_2. It follows that the difference ϑ=θ1−θ2\vartheta= \theta_1 - \theta_2 obeys the equations

(∂∂t+⟨u⟩⋅∇)ϑ=−v⋅∇⟨θ⟩(96)\left(\frac{\partial}{\partial t} + \langle u\rangle\cdot\nabla\right)\vartheta= -v \cdot\nabla\langle\theta\rangle \tag*{(96)}

in the sense of distributions, where v(t)=G∗ϑ(t)v(t) = G * \vartheta(t) and

⟨u⟩=12(G∗θ1+G∗θ2),⟨θ⟩=12(θ1+θ2).(97)\begin{aligned} \langle u\rangle&= \frac{1}{2}(G * \theta_1 + G * \theta_2), \\ \langle\theta\rangle&= \frac{1}{2}(\theta_1 + \theta_2). \tag*{(97)} \end{aligned}

At each instance of time, we have ϑ(t)∈Y\vartheta(t) \in Y and clearly

spt⁡ϑ(t)⊂spt⁡θ1(t)∪spt⁡θ2(t),(98)\operatorname{spt}\vartheta(t) \subset\operatorname{spt}\theta_1(t) \cup\operatorname{spt}\theta_2(t), \tag*{(98)}

since θ1\theta_1 and θ2\theta_2 obey the estimates. Through Lemma 2, the velocities ⟨u⟩\langle u\rangle and vv have gradients bounded uniformly in time with constant depending only on ∥θ1∥Y+∥θ2∥Y\|\theta_1\|_Y + \|\theta_2\|_Y. We multiply equation (96) with ϑ\vartheta and integrate by parts to discover

ddt∥ϑ∥L22=0,(99)\frac{d}{dt}\|\vartheta\|_{L^2}^2 = 0, \tag*{(99)}

since ⟨u⟩\langle u\rangle and vv are divergence-free and regular enough. Recalling that ∥ϑ(0)∥L2=0\|\vartheta(0)\|_{L^2} = 0, we have proved uniqueness. □\square

Proof of Theorem 6

We consider the dynamics of the patch solution θ(t)=1P(t)\theta(t) = \mathbf{1}_{P(t)} given by Corollary 4 and address here the question of regularity for the evolving boundary ∂P(t)\partial P(t). First, we note that θ\theta must necessarily satisfy the estimates in Theorem 3. The particle trajectories are volume-preserving such that the area of P(t)P(t) is constant in time, and so we define the length scale

L=area⁡(P0)=area⁡(P(t)).(100)L = \sqrt{\operatorname{area}(P_0)} = \sqrt{\operatorname{area}(P(t))}. \tag*{(100)}

Then, u=G∗1Pu = G * \mathbf{1}_P has the uniform bound in time:

∥∇u(t)∥Cμ≤CμCL(101)\lVert\nabla u(t) \rVert_{C^\mu} \le C_\mu C_L \tag*{(101)}

where CL=1+L2C_L = 1 + L^2 depends only on the initial data.

The regularity result for the distribution 1P\mathbf{1}_P follows from analysis of gradients of the defining level-set function φ(t)\varphi(t) satisfying (17), which is the unique global solution to the Cauchy problem

{(∂∂t+u⋅∇)φ=0,φ(x,0)=φ0(x),(102)\begin{cases} \left(\frac{\partial}{\partial t} + u \cdot\nabla\right)\varphi= 0, \\ \varphi(x,0) = \varphi_0(x), \tag*{(102)} \end{cases}

where φ0\varphi_0 satisfies (15). Above, the fluid velocity u=G∗1Pu = G * \mathbf{1}_P has the convenient expression using φ\varphi,

u(x,t)=∫R2G(x−y)H(φ(y,t)) dy,(103)u(x,t) = \int_{\mathbb{R}^2} G(x-y)H(\varphi(y,t))\,\mathrm{d}y, \tag*{(103)}

where HH is the Heaviside function. By construction, P(t)P(t) is precisely the zero level-set of φ(t)\varphi(t) such that

1P(t)=H∘φ(t),(104)\mathbf{1}_{P(t)} = H \circ\varphi(t), \tag*{(104)}

and so our regularity results for the boundary rely on an analysis of W=∇⊥φ\mathbf{W} = \nabla^\perp\varphi, which is tangent to ∂P\partial P. Recalling the evolution equation (28) for W(t)\mathbf{W}(t), we must estimate the product ∇uW\nabla u\mathbf{W}.

We provide an additional observation in estimating ∣⋅∣μ|\cdot|_\mu: the singular kernel GG away from the origin has gradient ∇G\nabla G, with homogeneity of degree −1-1, such that for each instance of time

∇u(x)=pv⁡∫P∇G(x−y) dy,(105)\nabla u(x) = \operatorname{pv}\int_P \nabla G(x-y)\,\mathrm{d}y, \tag*{(105)}

and so we write ∇u=∇G∗1P\nabla u = \nabla G * \mathbf{1}_P. Since the degree in modulus is less than the dimension, integration by parts yields

pv⁡∫P∇[G(x−y)]⋅W(y) dy=0(106)\operatorname{pv}\int_P \nabla[G(x-y)] \cdot\mathbf{W}(y)\,\mathrm{d}y = 0 \tag*{(106)}

for divergence-free W\mathbf{W} tangent to the boundary. Thus, we have commutative structure for the kernel ∇G(z)\nabla G(z),

∇u(x)W=(∇GW)∗1P,(107)\nabla u(x)\mathbf{W}=(\nabla G\mathbf{W})*1_P, \tag*{(107)}

and following estimate:

Lemma 19. Suppose u=G∗1Pu=G*1_P and that W∈Cμ(R2,R2)\mathbf{W}\in C^\mu(\mathbb{R}^2,\mathbb{R}^2) is divergence free and tangent to ∂P\partial P. Then, there exists independent constant C2C_2 such that

∣∇uW∣μ≤C2∥∇u∥L∞∣W∣μ.|\nabla u\mathbf{W}|_\mu\leq C_2\|\nabla u\|_{L^\infty}|\mathbf{W}|_\mu.

Using the particle trajectories of uu, we derive from (28) pointwise estimates on equation W\mathbf{W} (see [11 Proposition 3]) wherein Grönwall’s inequality yields bounds for quantities (26) and (27) which depend exponentially on

∫0t∥∇u(s)∥L∞ ds≤CL∣t∣.(108)\int_0^t\|\nabla u(s)\|_{L^\infty}\,ds\leq C_L|t|. \tag*{(108)}

The particle trajectories allow us to deduce the regularity result:

Proposition 20. Suppose P0P_0 bounded and φ0∈C1+μ(R2)\varphi_0\in C^{1+\mu}(\mathbb{R}^2) in (102) such that ∣W0∣inf⁡>0|\mathbf{W}_0|_{\inf}>0. Then, the unique global solution φ\varphi has W=∇⊥φ\mathbf{W}=\nabla^\perp\varphi which satisfies:

∣W(t)∣μ≤∣W0∣μexp⁡((C2+μ)CL∣t∣),|\mathbf{W}(t)|_\mu\leq|\mathbf{W}_0|_\mu\exp((C_2+\mu)C_L|t|),
∥W(t)∥L∞≤∥W0∥L∞exp⁡(CL∣t∣),\|\mathbf{W}(t)\|_{L^\infty}\leq\|\mathbf{W}_0\|_{L^\infty}\exp(C_L|t|),
∣W(t)∣inf⁡≥∣W0∣inf⁡exp⁡(−CL∣t∣).|\mathbf{W}(t)|_{\inf}\geq|\mathbf{W}_0|_{\inf}\exp(-C_L|t|).

Proof of Theorem 6. for ∂P0∈C1+μ\partial P_0\in C^{1+\mu}. Indeed, Proposition 20 guarantees the desired C1+μC^{1+\mu} parametrization zz in (25) exists at each instance of time, if φ0\varphi_0 in (16) is chosen such that ∣W0∣inf⁡=∣∇⊥φ0∣inf⁡|\mathbf{W}_0|_{\inf}=|\nabla^\perp\varphi_0|_{\inf} is non-vanishing. □

Remark 21. Lemma 19 is used only in the estimate of ∣W∣μ|\mathbf{W}|_\mu in Proposition 20 and thus unnecessary to deduce ∂P∈C1+μ\partial P\in C^{1+\mu}. The infimum and supremum bounds are direct from (28) and ∇u∈Cμ\nabla u\in C^\mu depends only on θ0\theta_0 so that we may alternatively conclude with an exp⁡(CLexp⁡(CL∣t∣))\exp(C_L\exp(C_L|t|)) bound on ∣W∣μ|\mathbf{W}|_\mu.

For C2+μC^{2+\mu}-regularity, we must examine the dynamics of ∇W\nabla\mathbf{W}. The presence of ∇∇u⋅W\nabla\nabla u\cdot\mathbf{W} in equation (29) augments the Grönwall-type exponential bounds from particle trajectories. Inspecting expression (58) for ∇∇u\nabla\nabla u in this context,

∇∇u(x)=1P(x)E+14πp.v.⁡∫PH(x−y)∣x−y∣2 dy,(109)\nabla\nabla u(x)=1_P(x)\mathbf{E}+\frac{1}{4\pi}\operatorname{p.v.}\int_P\frac{\mathbf{H}(x-y)}{|x-y|^2}\,dy, \tag*{(109)}

we see that ∣∇∇u∣μ|\nabla\nabla u|_\mu is difficult to estimate directly. From here, we recognize the expression for ∇∇u(x)\nabla\nabla u(x) has similar geometric properties properties to the strain tensor for vortex patches [11]. We adapt their arguments here.

Proposition 22. Suppose u=G∗1Pu = G * 1_P and that W∈Cμ(R2,R2)\mathbf{W} \in C^\mu(\mathbb{R}^2,\mathbb{R}^2) is divergence free and tangent to ∂P\partial P. Then

∇∇u(x)⋅W=14πpv⁡∫PH(x−y)∣x−y∣2⋅(W(x)−W(y)) dy.\nabla\nabla u(x)\cdot\mathbf{W} = \frac{1}{4\pi}\operatorname{pv}\int_P \frac{\mathbf{H}(x-y)}{|x-y|^2}\cdot\left(\mathbf{W}(x)-\mathbf{W}(y)\right)\,dy.

Proof. Since W\mathbf{W} is divergence free and tangent to ∂P\partial P, then

pv⁡∫P∇[∇G(x−y)]⋅W(y) dy=−lim⁡δ→0∫∣x−y∣=δ,y∈P(W(y)⋅(x−yδ))∇G(x−y) dy(110)\operatorname{pv}\int_P \nabla[\nabla G(x-y)]\cdot\mathbf{W}(y)\,dy = -\lim_{\delta\to0}\int_{\substack{|x-y|=\delta,\,y\in P}} \left(\mathbf{W}(y)\cdot\left(\frac{x-y}{\delta}\right)\right)\nabla G(x-y)\,dy \tag*{(110)}
=−1P(x)E⋅W(x).= -1_P(x)\mathbf{E}\cdot\mathbf{W}(x).

The last equality follows from (52) in the proof of Proposition 11.

Corollary 23. Suppose u=G∗1Pu = G * 1_P and that W∈Cμ(R2,R2)\mathbf{W} \in C^\mu(\mathbb{R}^2,\mathbb{R}^2) is divergence free and tangent to ∂P\partial P. Then, there exists independent constant C3C_3 such that

∣∇∇u⋅W∣μ≤C3∥∇∇u∥L∞∣W∣μ.|\nabla\nabla u\cdot\mathbf{W}|_\mu\le C_3\|\nabla\nabla u\|_{L^\infty}|\mathbf{W}|_\mu.

The estimation of ∥∇∇u∥L∞\|\nabla\nabla u\|_{L^\infty} is consequence of of the fact that small neighborhoods containing ∂P\partial P look like half-circles, and that the kernel H(z)/∣z∣2\mathbf{H}(z)/|z|^2 is reflection symmetric. To illustrate, consider the set of points x0x_0 with distance

d(x0)=inf⁡x∈∂P{∣x−x0∣}(111)d(x_0)=\inf_{x\in\partial P}\{|x-x_0|\} \tag*{(111)}

less than a cutoff 0<δ≤∞0<\delta\le\infty. This cutoff is explicit

δμ=∣∇φ∣inf⁡∣∇φ∣μ,(112)\delta^\mu=\frac{|\nabla\varphi|_{\inf}}{|\nabla\varphi|_\mu}, \tag*{(112)}

given φ\varphi which satisfies (17). Such a choice ensures that the boundary ∂P\partial P can be straightened near the points x0x_0 where

d(x0)<δ.(113)d(x_0)<\delta. \tag*{(113)}

Indeed, for such points x0x_0 the semicircle

Σ(x0)={z∣∣z∣=1,∇φ(x~)⋅z≥0},(114)\Sigma(x_0)=\{z\mid|z|=1,\nabla\varphi(\tilde{x})\cdot z\ge0\}, \tag*{(114)}

where above x~∈∂P\tilde{x}\in\partial P is any such that ∣x~−x0∣=d(x0)|\tilde{x}-x_0|=d(x_0) holds, is in fact well-approximated by the set of directions

Sρ(x0)={z∣∣z∣=1,x0+ρz∈P}(115)S_\rho(x_0)=\{z\mid|z|=1,x_0+\rho z\in P\} \tag*{(115)}

for every ρ≥d(x0)\rho\ge d(x_0). More quantitatively,

Lemma 24 (Geometric Lemma). Denote the symmetric difference as

Rρ(x0)=(Sρ(x0)∖Σ(x0))∪(Σ(x0)∖Sρ(x0))R_{\rho}(x_0) = (S_{\rho}(x_0) \setminus\Sigma(x_0)) \cup(\Sigma(x_0) \setminus S_{\rho}(x_0))

and the Lebesgue measure on the unit circle as H1H^1. Then,

H1(Rρ(x0))≤2π((1+2μ)d(x0)ρ+2μ(ρδ)μ)H^1(R_{\rho}(x_0)) \le2\pi\left((1+2^\mu)\frac{d(x_0)}{\rho}+2^\mu\left(\frac{\rho}{\delta}\right)^\mu\right)

for all ρ≥d(x0)\rho\ge d(x_0), μ>0\mu> 0 and x0x_0 such that d(x0)<δ=(∣∇φ∣inf⁡∣∇φ∣μ)1/μd(x_0)<\delta=\left(\frac{|\nabla\varphi|_{\inf}}{|\nabla\varphi|_{\mu}}\right)^{1/\mu}.

Proposition 25. Suppose u=G∗1Pu=G*1_P and that φ\varphi satisfies (17) for PP. Then, there exists constant C4C_4 depending only on μ\mu, LL, ∣W0∣μ|\mathbf{W}_0|_\mu and ∣W0∣inf⁡|\mathbf{W}_0|_{\inf} such that

∥∇∇u(t)∥L∞≤C4(1+∣t∣).\|\nabla\nabla u(t)\|_{L^\infty}\le C_4(1+|t|).

Proof. We need only estimate the singular integral, which we split into I1I_1 and I2I_2. For δ\delta given by (113), the latter has the bound

∣I2(x0)∣=∣14π∫P∩{∣x0−y∣≥δ}H(x0−y)∣x0−y∣2 dy∣≤1+log⁡(δL).(116)|I_2(x_0)|=\left|\frac{1}{4\pi}\int_{P\cap\{|x_0-y|\ge\delta\}}\frac{\mathbf{H}(x_0-y)}{|x_0-y|^2}\,dy\right|\le1+\log\left(\frac{\delta}{L}\right). \tag*{(116)}

The remaining term,

∣I1(x0)∣=∣14π∫P∩{∣x0−y∣<δ}H(x0−y)∣x0−y∣2 dy∣(117)|I_1(x_0)|=\left|\frac{1}{4\pi}\int_{P\cap\{|x_0-y|<\delta\}}\frac{\mathbf{H}(x_0-y)}{|x_0-y|^2}\,dy\right| \tag*{(117)}

vanishes for d(x0)>δd(x_0)>\delta. We thus assume d(x0)<δd(x_0)<\delta, and pass to polar coordinates centered at x0x_0 to find

∣I1(x0)∣≤14π∫d(x0)δdρρH1(Rρ(x0)),(118)|I_1(x_0)|\le\frac{1}{4\pi}\int_{d(x_0)}^\delta\frac{d\rho}{\rho}H^1(R_\rho(x_0)), \tag*{(118)}

using the fact ∫Σ(x0)H(z)dH1(z)\int_{\Sigma(x_0)}\mathbf{H}(z)dH^1(z) vanishes by reflection symmetry. Now applying the Geometric Lemma, we integrate to discover

∣I1∣≤12(1+2μ+2μμ)(119)|I_1|\le\frac{1}{2}\left(1+2^\mu+\frac{2^\mu}{\mu}\right) \tag*{(119)}

We then have our bound

∥∇∇u∥L∞≤(4+1μ)(1+log⁡[∣∇φ∣μLμ∣∇φ∣inf⁡])(120)\|\nabla\nabla u\|_{L^\infty}\le\left(4+\frac{1}{\mu}\right)\left(1+\log\left[\frac{|\nabla\varphi|_\mu L^\mu}{|\nabla\varphi|_{\inf}}\right]\right) \tag*{(120)}

and use the estimates in Proposition 20 to conclude.

The proof of Corollary 23 and the Geometric Lemma may be found in the Appendix of [11]. It follows that (29) admits a Grönwall-type estimate.

Proposition 26. Suppose P0P_0 bounded and that φ0∈C2+μ\varphi_0 \in C^{2+\mu} in (102). Then, the unique global solution φ\varphi has ∇W=∇∇⊥φ\nabla W = \nabla\nabla^\perp\varphi which satisfies:

∣∇W(t)∣μ≤exp⁡((C3+μ)CL∣t∣)[∣∇W0∣μ+C4∫0t(1+∣s∣)∣W(s)∣μ ds],\lvert\nabla W(t) \rvert_\mu\le\exp\left((C_3+\mu)C_L\lvert t\rvert\right)\left[\lvert\nabla W_0\rvert_\mu+C_4\int_0^t(1+\lvert s\rvert)\lvert W(s)\rvert_\mu\,\mathrm{d}s\right],
∥∇W(t)∥L∞≤exp⁡(CL∣t∣)[∥∇W0∥L∞+C4∫0t(1+∣s∣)∥W(s)∥L∞ ds],\lVert\nabla W(t)\rVert_{L^\infty}\le\exp\left(C_L\lvert t\rvert\right)\left[\lVert\nabla W_0\rVert_{L^\infty}+C_4\int_0^t(1+\lvert s\rvert)\lVert W(s)\rVert_{L^\infty}\,\mathrm{d}s\right],

for positive tt.

Proof of Theorem 6 for ∂P0∈C2+μ\partial P_0\in C^{2+\mu}. With Proposition 20 and Proposition 26, we see that

∥W(t)∥C1+μ≤C5∥W0∥C1+μ(1+∣t∣)exp⁡(CL∣t∣)(121)\lVert W(t)\rVert_{C^{1+\mu}}\le C_5\lVert W_0\rVert_{C^{1+\mu}}(1+\lvert t\rvert)\exp(C_L\lvert t\rvert) \tag*{(121)}

for some constant C5C_5 depending only on μ\mu, LL and φ0\varphi_0, as desired.

Details of Figure 1

Description of the numerical solver

In the compact domain T2\mathbb{T}^2, we develop a solver for the density patch problem in the system (B∗)(B_*) which resolves the dynamics of the patch boundary ∂P(t)\partial P(t) in Corollary 8 using a level-set method that is second-order in time and first-order in space (for background, see [25]).

The algorithm begins with the expression

u=−∇⊥(Δ2)−1∂1θ,(122)u=-\nabla^\perp(\Delta^2)^{-1}\partial_1\theta, \tag*{(122)}

where the scalar θ:T2→R\theta:\mathbb{T}^2\to\mathbb{R} has zero mean, so the operator (Δ2)−1(\Delta^2)^{-1} on ∂1θ\partial_1\theta is well-defined and the vector-field u:T2→R2u:\mathbb{T}^2\to\mathbb{R}^2 has zero mean also. The discrete Fourier coefficients u^\widehat{u} are related explicitly to θ^\widehat{\theta}

u^(k)=k⊥k1∣k∣4θ^(k)(123)\widehat{u}(k)=\frac{k^\perp k_1}{\lvert k\rvert^4}\widehat{\theta}(k) \tag*{(123)}

for nonzero k∈2πZ2k\in2\pi\mathbb{Z}^2 and k⊥=(−k2,k1)k^\perp=(-k_2,k_1). Accordingly, we set u^(0)=0\widehat{u}(0)=0.

With the spacing h=1/Nh=1/N, we discretize the space variable onto an N×NN\times N lattice with coordinates xij=h(i−N/2,j−N/2)x_{ij}=h(i-N/2,j-N/2) for i,j=0,…,N−1i,j=0,\ldots,N-1. Note that we have fixed NN to some power of two, and write φij(t):=φ(xij,t)\varphi_{ij}(t):=\varphi(x_{ij},t) such that θij=H(φij)\theta_{ij}=H(\varphi_{ij}). With this convention, we have

θ^(k)≈h2∑j=0N−1∑i=0N−1H(φij)exp⁡(−ik⋅xij),(124)\widehat{\theta}(k)\approx h^2\sum_{j=0}^{N-1}\sum_{i=0}^{N-1}H(\varphi_{ij})\exp(-\mathrm{i}k\cdot x_{ij}), \tag*{(124)}

where ι\iota here is the imaginary unit. We thus resolve the advecting velocity with spectral accuracy:

uij≈∑∣k∣≤πNk⊥k1∣k∣4θ^(k)exp⁡(ιk⋅xij).(125)u_{ij} \approx\sum_{|k|\leq\pi N} \frac{k^\perp k_1}{|k|^4}\widehat{\theta}(k)\exp(\iota k\cdot x_{ij}). \tag*{(125)}

The advection operator F(φ)=−u⋅∇φF(\varphi)=-u\cdot\nabla\varphi is discretized according to the monotone first-order upwinding scheme

F(φij)≈(uij)−⋅Dij+−(uij)+⋅Dij−,(126)F(\varphi_{ij}) \approx(u_{ij})^- \cdot D_{ij}^+-(u_{ij})^+\cdot D_{ij}^-, \tag*{(126)}

where the signed velocities (v)±=max⁡(±v,0)(v)^\pm=\max(\pm v,0) with v=(v)+−(v)−v=(v)^+-(v)^- product with the signed gradient

Dij±=±1h(φi±1,j−φijφi,j±1−φij).(127)D_{ij}^{\pm}=\pm\frac{1}{h} \begin{pmatrix} \varphi_{i\pm1,j}-\varphi_{ij}\\ \varphi_{i,j\pm1}-\varphi_{ij} \end{pmatrix}. \tag*{(127)}

With our procedure given by (120-123), we integrate the system

∂φij∂t=F(φij)(128)\frac{\partial\varphi_{ij}}{\partial t}=F(\varphi_{ij}) \tag*{(128)}

in time using the second-order SSRK (Heun’s) method where the time step is chosen such that CFL≤1/2CFL\leq1/2, to complete the algorithm.

Verification of convergence

Our verification of the numerical solver described in § 6.1 addresses the implementation of the following three routines:

  1. the level-set method for the transport equation,

  1. the spectral-collocation method for the momentum equation,

  1. the coupling of (1) and (2) to approximate patch solutions in the full system.

The implementation of first-order upwind to transport the level-set function φ\varphi was tested against various fixed divergence-free velocity fields on T2\mathbb{T}^2: the shear u(x)=(x23,0)u(x)=(x_2^3,0), and cellular flows u=∇⊥ψu=\nabla^\perp\psi where

ψ(x)=sin⁡(k1x1)sin⁡(k2x2),(129)\psi(x)=\sin(k_1x_1)\sin(k_2x_2), \tag*{(129)}

for various mode numbers k∈2πZ2k\in2\pi\mathbb{Z}^2. The spectral solver for the momentum equation was tested on sinusoidal temperature distributions

θ(x)=cos⁡(k1x1+k2x2)(130)\theta(x)=\cos(k_1x_1+k_2x_2) \tag*{(130)}

and it indeed recovers the exact solution

u(x)=k1k⊥∣k∣4cos⁡(k1x1+k2x2)(131)u(x)=\frac{k_1k^\perp}{|k|^4}\cos(k_1x_1+k_2x_2) \tag*{(131)}

up to machine precision for any ∣k∣≤πN|k| \le\pi N, as expected by the Nyquist-Shannon sampling theorem.

These two solvers are coupled together in the full algorithm, so we verified that quantities like ∥∇∇u∥L∞\lVert\nabla\nabla u\rVert_{L^\infty} and ∥∇∇φ∥L∞\lVert\nabla\nabla\varphi\rVert_{L^\infty} from simulations with the same initial data but various NN converge for short time we increase NN.

Finally, we examine if the numerical solver is appropriately handling the dynamics of the low regularity solution which has initial data φ0(x)=Φ(x)\varphi_0(x) = \Phi(x). Consider the regularized system

{(∂∂t+u⋅∇)φ=0,−Δu+∇Π=φϵ∗H(φ)e2,div⁡u=0,\left\{ \begin{aligned} \left(\frac{\partial}{\partial t}+u\cdot\nabla\right)\varphi&= 0,\\ -\Delta u+\nabla\Pi&= \varphi_\epsilon* H(\varphi)e_2,\\ \operatorname{div}u &= 0, \end{aligned} \right.

where we have ϵ>0\epsilon> 0 and the mollifier is the Gaussian

φϵ(z)=1ϵ2πexp⁡(−z22ϵ2).(132)\varphi_\epsilon(z)=\frac{1}{\epsilon\sqrt{2\pi}}\exp\left(-\frac{z^2}{2\epsilon^2}\right). \tag*{(132)}

The system (Φϵ)(\Phi_\epsilon) is solved with the algorithm given above for system (B∗)(B_*), except the Gaussian filter with standard deviation ϵ\epsilon (convolution with φϵ\varphi_\epsilon) is applied numerically to the points H(φij)H(\varphi_{ij}) before the discrete Fourier transform in (124) is taken. We observe that the tendency towards curvature singularities observed in Figure 1 is suppressed in this regularized system for any fixed ϵ\epsilon. However, as we take epsilon to machine precision, the simulations recover the results of ϵ=0\epsilon= 0 (system (B∗)(B_*)), in particular the picture in Figure 1.

The simulation in Figure 1

The curve illustrated in the Figure is the zero level set of φij(t)\varphi_{ij}(t) whose initial condition for the algorithm was specified as φ0(xij)=Φ(xij)\varphi_0(x_{ij}) = \Phi(x_{ij}), where

Φ(x)=cos⁡(2π∣x−ξ∣),ξ=(12,12)(133)\Phi(x)=\cos(2\pi|x-\xi|),\qquad\xi=\left(\frac{1}{2},\frac{1}{2}\right) \tag*{(133)}

has precisely the circle of radius one-half S1(1/2)S^1(1/2) centered at ξ\xi as its zero level set. The patch moves upwards, as does the curve ∂P(t)\partial P(t). To depict the changing shape of ∂P\partial P in time, the frame where the curve is drawn follows this movement.

While patch solutions to our problem maintain their area as they evolve in time, the sharpening of the ∂P(t)\partial P(t) in the fixed-grid simulations results in rapid decreases in the area once the variations of the curve are of the scale h=1/Nh=1/N. Thus for the resolution given by N=1024N=1024, we terminate the patch simulation at t=100t=100, before the relative error of the patch area is more than 0.01.

Acknowledgments

I thank my advisor, Peter Constantin, for suggesting this problem to me, and for his guidance thereof. I also express my gratitude to Greg Hammett, for his guidance on developing the simulation. The author thanks the anonymous reviewer #2 for their helpful comments. During this project, the author was partially supported by the Ford Foundation and the NSF Graduate Research Fellowship grant DGE-2039656.

References

  1. [1]Cannon, J. and DiBenedetto, E. (1980) The initial value problem for the Boussinesq equations with data in Lp. In Approximation Methods for Navier-Stokes Problems pp. 129–144 Springer Berlin Heidelberg Berlin, Heidelberg.DOI
  2. [2]Chae, D. (July, 2006) Global regularity for the 2D Boussinesq equation with partial viscosity terms. Advances in Mathematics, 203(2), 497–513.DOI
  3. [3]Hou, T. Y. and Li, C. (2005) Global well-posedness of the viscous Boussinesq equations. Discrete & Continuous Dynamical Systems - A, 12(1), 1–12.DOI
  4. [4]Keith Moffatt, H. (2001) Some Remarks on Topological Fluid Mechanics. In An Introduction to the Geometry and Topology of Fluid Flows pp. 3–10 Springer Netherlands Dordrecht.DOI
  5. [5]Abidi, H. and Hmidi, T. (February, 2007) On the global well-posedness for Boussinesq system. Journal of Differential Equations, 233(1), 199–220.DOI
  6. [6]Hmidi, T. and Keraani, S. (2007) On the global well-posedness of the two-dimensional Boussinesq system with a zero diffusivity. Advances in Differential Equations, 12(4), 461–480.
  7. [7]Danchin, R. and Paicu, M. (2008) Les théorèmes de Leray et de Fujita-Kato pour le système de Boussinesq partiellement visqueux. Bulletin de la Société mathématique de France, 136(2), 261–309.DOI
  8. [8]Buttke, T. F. (1989) The observation of singularities in the boundary of patches of constant vorticity. Physics of Fluids A: Fluid Dynamics, 1(7), 1283–1285.DOI
  9. [9]Dritschel, D. and McIntyre, M. (1990) Does contour dynamics go singular?. Physics of Fluids A: Fluid Dynamics, 2(5), 748–753.DOI
  10. [10]Chemin, J.-Y. (1993) Persistance de structures géométriques dans les fluides incompressibles bidimensionnels. In Annales scientifiques de l’Ecole normale supérieure Vol. 26, pp. 517–542.DOI
  11. [11]Bertozzi, A. and Constantin, P. (February, 1993) Global regularity for vortex patches. Communications in Mathematical Physics, 152(1), 19–28.DOI
  12. [12]Serfati, P. (1994) Une preuve directe d’existence globale des vortex patches 2D. Comptes rendus de l’Académie des sciences. Série 1, Mathématique, 318(6), 515–518.
  13. [13]Danchin, R. and Zhang, X. (November, 2016) Global persistence of geometrical structures for the Boussinesq equation with no diffusion. Communications in Partial Differential Equations, 42(1), 68–99.
  14. [14]Gancedo, F. and García-Juárez, E. (July, 2017) Global Regularity for 2D Boussinesq Temperature Patches with No Diffusion. Annals of PDE, 3(2), 1–34.arxiv.org/abs/1611.10260
  15. [15]Chae, D., Miao, Q., and Xue, L. (2022) Global Regularity of NonDiffusive Temperature Fronts for the Two-Dimensional Viscous Boussinesq System. SIAM Journal on Mathematical Analysis, 54(4), 4043–4103.DOI
  16. [16]Höfer, R. M. (2018) Sedimentation of inertialess particles in Stokes flows. Communications in Mathematical Physics, 360, 55–101.arxiv.org/abs/1610.03748
  17. [17]Mecherbet, A. (2019) Sedimentation of particles in Stokes flow. Kinetic and Related Models, 12(5), 995–1044.arxiv.org/abs/1806.07795
  18. [18]Leblond, A. (2022) Well-posedness of the Stokes-transport system in bounded domains and in the infinite strip. Journal de Mathématiques Pures et Appliquées, 158, 120–143.arxiv.org/abs/2103.16305
  19. [19]Höfer, R. M. and Schubert, R. (2021) The influence of Einstein’s effective viscosity on sedimentation at very small particle volume fraction. In Annales de l’Institut Henri Poincaré C, Analyse non linéaire Elsevier Vol. 38, pp. 1897–1927.
  20. [20]Mecherbet, A. and Sueur, F. (2022) A few remarks on the transport-Stokes system. arXiv preprint arXiv:2209.11637,.arxiv.org/abs/2209.11637
  21. [21]Cobb, D. (2023) On the Well-Posedness of a Fractional Stokes-Transport System. arXiv preprint arXiv:2301.10511,.arxiv.org/abs/2301.10511
  22. [22]Antontsev, S., Yurinsky, V., and Meirmanov, A. (2000) A free-boundary problem for Stokes equations: Classical solutions. Interfaces and Free Boundaries, 2(4), 413–424.DOI
  23. [23]Gancedo, F., Granero-Belinchón, R., and Salguero, E. (2022) Long time interface dynamics for gravity Stokes flow. arXiv preprint arXiv:2211.03437,.arxiv.org/abs/2211.03437
  24. [24]Yudovich, V. I. (1963) Non-stationary flows of an ideal incompressible fluid. Zhurnal Vychislitel’noi Mathematiki i Matematicheskoi Fiziki, 3(6), 1032–1066.DOI
  25. [25]Sethian, J. A. (1996) Level Set methods: evolving interfaces in geometry, fluid mechanics, computer vision, and materials Science, Cambridge University Press, Cambridge.
  26. [26]Stein, E. M. (1970) Singular Integrals and Differentiability Properties of Functions, Princeton Mathematical SeriesPrinceton University Press, Princeton.
  27. [27]Majda, A. J. and Bertozzi, A. L. (November, 2001) Vorticity and Incompressible Flow, Vol. 27, Cambridge University Press, Cambridge.DOI

Paper details

vlnc:W_w3r2cd8f
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents