Introduction
We consider the isentropic compressible Navier–Stokes system on the three-dimensional unit torus T 3 = R 3 / Z 3 \mathbb{T}^3 = \mathbb{R}^3/\mathbb{Z}^3 T 3 = R 3 / Z 3 :
\begin{cases}
\partial_t \rho+ \operatorname{div}(\rho u) = 0,\\
\partial_t(\rho u) + \operatorname{div}(\rho u \otimes u) + \nabla P = \operatorname{div}\mathcal{S}(u),
\end{cases}
\qquad(t,x) \in(0,\infty) \times\mathbb{T}^3. \tag*{(1)}\end{cases}
Here ρ ≥ 0 \rho\ge0 ρ ≥ 0 denotes the density, u ∈ R 3 u \in\mathbb{R}^3 u ∈ R 3 the velocity, and P = ρ γ P = \rho^\gamma P = ρ γ the pressure, whose coefficient is normalized to 1. As in [15 ] , we take into account the standard Newtonian stress tensor and viscosity conditions:
S ( u ) = μ ( ∇ u + ( ∇ u ) T ) + λ ( div u ) Id , \mathcal{S}(u) = \mu\left(\nabla u + (\nabla u)^{\mathsf{T}}\right) + \lambda(\operatorname{div} u)\operatorname{Id}, S ( u ) = μ ( ∇ u + ( ∇ u ) T ) + λ ( div u ) Id ,
μ > 0 , λ + 2 3 μ ≥ 0 , ν : = 2 μ + λ > 0. (2) \mu> 0,\qquad\lambda+ \frac{2}{3}\mu\ge0,\qquad\nu:= 2\mu+ \lambda> 0. \tag*{(2)} μ > 0 , λ + 3 2 μ ≥ 0 , ν := 2 μ + λ > 0. (2)
Thus div S ( u ) = μ Δ u + ( λ + μ ) ∇ div u \operatorname{div}\mathcal{S}(u) = \mu\Delta u + (\lambda+ \mu)\nabla\operatorname{div}u div S ( u ) = μ Δ u + ( λ + μ ) ∇ div u . We write Q T = ( 0 , T ) × T 3 Q_T = (0,T) \times\mathbb{T}^3 Q T = ( 0 , T ) × T 3 and m = ρ u m = \rho u m = ρ u .
Lions [27 ] established the global finite-energy weak-solution theory in three dimensions for γ ≥ 9 / 5 \gamma\ge9/5 γ ≥ 9/5 . His framework couples the energy inequality with weak
Keywords: compressible Navier–Stokes equations; finite-energy weak solutions; periodic optimal transport; pressure integrability; effective viscous flux.
Corresponding author: Changjiang Zhu. continuity of the effective viscous flux and renormalization of the continuity equation. These ingredients address different failures of weak convergence: the energy controls the size of the approximations, whereas the flux and renormalization arguments identify the pressure and remove density oscillations. Thus the existence theorem also provides a compactness method for nonlinear compressible flows. Feireisl, Novotný and Petzeltová [15 ] extended the range to γ > 3 / 2 \gamma> 3/2 γ > 3/2 . The compactness argument works with bounded truncations of the density and controls the oscillations before removing the truncation. Feireisl’s analysis in [13 ] for the density in L p L^p L p with p < 2 p < 2 p < 2 explains why renormalization must be recovered from this oscillation control. While the systematic approximation and stability theory is developed in [14 ] .
To see how these ingredients fit together, consider approximate solutions with the uniform energy bounds
ρ n ∈ L ∞ ( 0 , T ; L γ ) , u n ∈ L 2 ( 0 , T ; H 1 ) , ρ n u n ∈ L ∞ ( 0 , T ; L 2 ) . \rho_n \in L^\infty(0,T;L^\gamma), \qquad u_n \in L^2(0,T;H^1), \qquad\sqrt{\rho_n}u_n \in L^\infty(0,T;L^2). ρ n ∈ L ∞ ( 0 , T ; L γ ) , u n ∈ L 2 ( 0 , T ; H 1 ) , ρ n u n ∈ L ∞ ( 0 , T ; L 2 ) .
On the torus, the velocity mean is controlled using the conserved mass and the weighted kinetic energy; see Lemma 2.5 . These bounds give weakly convergent subsequences. However, they do not give spatial derivatives of the density or identify a weak limit of the pressure. The first additional step is therefore to establish an integrability estimate for the density function via testing the momentum equation. The classical test function reads
V = ∇ Δ T 3 − 1 ( ρ θ − ⟨ ρ θ ⟩ ) , div V = ρ θ − ⟨ ρ θ ⟩ . (3) V = \nabla\Delta_{\mathbb{T}^3}^{-1}\left(\rho^\theta-\langle\rho^\theta\rangle\right), \qquad\operatorname{div} V = \rho^\theta-\langle\rho^\theta\rangle. \tag*{(3)} V = ∇ Δ T 3 − 1 ( ρ θ − ⟨ ρ θ ⟩ ) , div V = ρ θ − ⟨ ρ θ ⟩ . (3)
Hereafter ⟨ f ⟩ : = ∫ T 3 f d x \langle f\rangle:= \int_{\mathbb{T}^3} f\,\mathrm{d}x ⟨ f ⟩ := ∫ T 3 f d x . For P = ρ γ P=\rho^\gamma P = ρ γ and a smooth time cutoff 0 ≤ η ∈ C c ∞ ( 0 , T ) 0\leq\eta\in C^\infty_c(0,T) 0 ≤ η ∈ C c ∞ ( 0 , T ) , its pressure pairing is
∫ Q T η ρ γ div V = ∫ Q T η ρ γ + θ − ∫ 0 T η ⟨ ρ θ ⟩ ∫ T 3 ρ γ d x d t . \int_{Q_T}\eta\rho^\gamma\operatorname{div}V=\int_{Q_T}\eta\rho^{\gamma+\theta}-\int_0^T\eta\langle\rho^\theta\rangle\int_{\mathbb{T}^3}\rho^\gamma\,\mathrm{d}x\,\mathrm{d}t. ∫ Q T η ρ γ div V = ∫ Q T η ρ γ + θ − ∫ 0 T η ⟨ ρ θ ⟩ ∫ T 3 ρ γ d x d t .
The mean correction is controlled by energy when θ ≤ γ \theta\leq\gamma θ ≤ γ . The same test also produces the unsigned convective term ∫ ( ρ u ⊗ u ) : ∇ V \int(\rho u\otimes u):\nabla V ∫ ( ρ u ⊗ u ) : ∇ V . With ρ ∈ L t ∞ L x γ \rho\in L^\infty_tL^\gamma_x ρ ∈ L t ∞ L x γ and u ∈ L t 2 L x 6 u\in L^2_tL^6_x u ∈ L t 2 L x 6 , direct Hölder estimates require
1 γ + 1 3 + θ γ ≤ 1. \frac{1}{\gamma}+\frac{1}{3}+\frac{\theta}{\gamma}\leq1. γ 1 + 3 1 + γ θ ≤ 1.
This, together with the viscous term, gives the classical range (cf. [14 , 15 ] ):
0 < θ < min { γ 2 , 2 γ 3 − 1 } , γ > 3 2 . (4) 0<\theta<\min\left\{\frac{\gamma}{2},\frac{2\gamma}{3}-1\right\},\qquad\gamma>\frac{3}{2}. \tag*{(4)} 0 < θ < min { 2 γ , 3 2 γ − 1 } , γ > 2 3 . (4)
Clearly, (4) leaves room for a positive θ \theta θ only when γ > 3 / 2 \gamma>3/2 γ > 3/2 . The time derivative of the test is handled through the renormalized continuity equation; it is not an independent regularity bound on ∂ t ρ \partial_t\rho ∂ t ρ .
The effective viscous flux combines the pressure with the compressive part of the viscous stress. For the coefficients in (2) , it is P − ν div u P-\nu\operatorname{div}u P − ν div u . Although neither factor in its product with a density truncation need converge strongly, the momentum and continuity equations provide a compensated weak-continuity relation [13 , 15 , 27 ] . On the torus one tests with a gradient inverse Laplacian, retaining the zero-mode correction. Combining the time derivative with the continuity equation leaves a commutator of the velocity and the periodic Riesz transforms. Compactness of this commutator, rather than separate strong convergence of pressure and velocity divergence, is what the flux calculation supplies. Monotonicity of the pressure then gives a sign to the remaining density oscillation. The renormalized equation, whose commutator foundation is due to DiPerna and Lions [11 ] , converts this information into an entropy comparison. A Young measure records possible oscillations of the approximate densities, see [3 ] . And the strict convexity of the entropy identifies when this measure is a Dirac mass. Concentration of an unbounded pressure product requires a separate integrability argument and is not ruled out merely by introducing a Young measure. More precisely, writing ρ n γ ⇀ ρ γ \rho_n^\gamma\rightharpoonup\rho^\gamma ρ n γ ⇀ ρ γ does not yet show ρ γ ‾ = ρ γ \overline{\rho^\gamma}=\rho^\gamma ρ γ = ρ γ . Hereafter g ‾ \overline{g} g denotes the weak limit of g g g . Therefore one has to compare the initialized renormalized equations for the sequence and the limit. It is important that the pressure covariance has a favorable sign, while the convex entropy defect starts from zero. For densities below L 2 L^2 L 2 , this comparison must first be performed with bounded density truncations and their associated entropies, and the weak source in the limiting truncated equation is kept until the oscillation estimate permits its removal. These are the classical compactness mechanisms used, after the required integrability of the density function has been achieved for γ > 3 / 2 \gamma> 3/2 γ > 3/2 in the classical theory.
Several results isolate parts of this difficulty, and thus relax the constraint upon γ \gamma γ . Under spherical symmetry, Jiang and Zhang [26 ] proved global weak existence for the Cauchy problem in R d \mathbb{R}^d R d , d = 2 , 3 d=2,3 d = 2 , 3 , with γ > 1 \gamma> 1 γ > 1 . The reduction supplied by symmetry is essential and does not yield a general three-dimensional theorem. Hu [24 ] constructed finite-energy renormalized solutions outside a closed exceptional set of zero parabolic Hausdorff measure for 6 / 5 < γ ≤ 3 / 2 6/5 < \gamma\le3/2 6/5 < γ ≤ 3/2 ; the result controls concentration away from that set. At the critical exponent γ = 3 / 2 \gamma= 3/2 γ = 3/2 , Plotnikov [34 ] obtained local logarithmic kinetic-energy estimates for a regularized problem. The background boundary-value formulation uses a smooth bounded domain and regular initial data with positive bounded density. The local estimate itself is expressed through the energy and local parameters, independently of those initial and boundary conditions. Such an estimate excludes the associated local concentration mechanism, but is distinct from identification of every nonlinear limit in a global finite-energy construction.
Recently, Bresch and Jabin [7 ] developed a quantitative compactness method for the continuity equation coupled to the momentum equation. Their work treats thermodynamically unstable pressure laws and anisotropic viscous stresses, for which the usual monotone-pressure effective-flux argument does not apply in its original form. It is a separate advance in compactness theory, rather than an unconditional low-exponent existence theorem for the constant-viscosity system considered here. These developments show that pressure integrability, density oscillation control and convergence of the inertial terms are connected but distinct tasks.
There is also a huge literature concerning the regularity of solutions or the existence of strong/classical solutions to (1) under stronger assumptions upon the initial data. Serrin [38 ] studied uniqueness of sufficiently regular compressible motions, Nash [32 ] established a local Cauchy theory, and Matsumura and Nishida [30 ] obtained global solutions for small smooth perturbations of a nonvacuum equilibrium. In the presence of vacuum, Huang, Li and Xin [25 ] proved global classical well-posedness for smooth data of small energy, allowing large oscillations and either vacuum or nonvacuum far-field states. Large oscillations in this result do not mean arbitrary initial energy. Li and Xin [28 ] obtained decay estimates in two and three dimensions and global classical existence in two dimensions for small-energy smooth data with far-field vacuum. Hong, Hou, Peng and Zhu [23 ] treated a class of large-energy classical solutions under a near-isothermal condition and a vacuum or sufficiently small far-field density. The near-isothermal restriction is different from small initial energy. The blow-up results of Xin [42 ] and Xin–Yan [43 ] concern smooth solutions, with compactly supported density or an isolated mass group in the models they study. The results in [42 , 43 ] do not exclude continuation in the weak solution class used below.
Hoff [18 , 19 , 20 ] established the intermediate-regularity theory for the compressible Navier–Stokes equations which allows discontinuities in the density while retaining estimates beyond the basic energy inequality. Specifically, for the isothermal and polytropic constructions, the initial density is a small perturbation of a positive constant in L 2 ∩ L ∞ L^2 \cap L^\infty L 2 ∩ L ∞ , and the velocity has small L 2 L^2 L 2 norm together with additional integrability. These restrictions are substantive relative to arbitrary finite-energy data with vacuum. One of the main ideas in [18 , 19 , 20 ] is to use the effective flux and estimates of the material derivative to separate the rough density from better controlled velocity quantities. The analysis of singularity curves [21 ] further describes how density and velocity-gradient jumps are transported. The uniqueness and continuous-dependence result is established in [22 ] by comparing corresponding particles in Lagrangian coordinates. This result applies to a weak-solution class with additional regularity and structural bounds, but it does not assert unconditional uniqueness for all Lions–Feireisl finite-energy solutions.
Density-dependent viscosities furnish another mechanism. Vasseur and Yu [41 ] and Li and Xin [29 ] constructed global weak solutions for degenerate viscosity laws. The entropy structure developed in the viscous shallow-water setting by Bresch and Desjardins in [6 ] supplies weighted density-gradient information. Mellet and Vasseur [31 ] established an additional logarithmic kinetic estimate and a stability theory for degenerate barotropic flows. This estimate addresses compactness of the kinetic quantities. Bresch, Vasseur and Yu [8 ] extended the entropy-weak theory to nonlinear density-dependent viscosities satisfying the BD relation. These estimates depend on the constitutive relation between the two viscosity coefficients. For positive constant μ \mu μ , that relation would impose λ = − 2 μ \lambda= -2\mu λ = − 2 μ , which is incompatible with (2) . The additional density-gradient estimate of that theory is consequently unavailable in the present problem.
For the constant-viscosity, monotone-pressure system considered here, the key of matter is to derive a pressure gain from the basic energy bounds with γ < 3 / 2 \gamma< 3/2 γ < 3/2 . the missing input is a pressure gain that can be paid for by the energy budget below γ = 3 / 2 \gamma= 3/2 γ = 3/2 . The test field must detect high density, have an L 2 L^2 L 2 Hessian affordable by the viscous stress, and control its time and convective contributions without a density-gradient estimate. Improving only the pressure weight does not resolve these simultaneous demands. This identifies the role of the geometric construction: it supplies a new test field within the same weak momentum equation, before the effective-flux and entropy comparison are used.
Our geometric starting point is Serre’s compensated integrability theory for nonnegative symmetric divergence-free tensors, see [35 ] . In gas dynamics the mass and momentum equations combine into a space–time tensor; its positivity allows determinant estimates to exploit the inertial terms together. Serre’s periodic extension in [36 ] gives estimates involving ρ 1 / 3 P \rho^{1/3}P ρ 1/3 P for three-dimensional inviscid gas flows. For the Navier–Stokes equations, placing the viscous stress inside this tensor destroys its positivity. Leaving the stress outside requires control of its divergence that is not supplied by the energy estimate. More recently, Serre [37 ] considered a mildly viscous model, relaxed symmetry constraint to positivity of the symmetric part, and obtained the corresponding estimate in two dimensions. The three-dimensional case of that model remains open there. Thus none of these results directly provides the pressure estimate for (1) .
Optimal transport suggests how to retain the geometric test while controlling its viscous cost. Brenier’s polar factorization identifies convex gradients in quadratic transport, see [5 ] . A classical transport map assigns one target point to each source point, but does not in general give the uniform Hessian cap needed for pairing with an L 2 L^2 L 2 stress. Transport planning permits a source point to distribute mass according to a probability measure. Alibert, Bouchitté and Champion [1 ] study costs depending nonlinearly on that measure, and Gozlan, Roberto, Samson and Tetali [17 ] establish a general Kantorovich duality framework. For barycentric quadratic costs, Gozlan and Juillet [16 ] relate the conditional barycenter to a convex gradient followed by a martingale coupling. The barycenter and the target variable therefore have different roles: the former is a deterministic field used in a differential equation, while the coupling preserves the prescribed target marginal. Strong convexity and subgradient monotonicity in Hilbert space provide the analytic control of the optimizer, see [4 ] .
We use these ideas to construct a periodic field with a controlled Hessian and a density-sensitive trace. The exact target marginal is retained at the level of plans, whereas a residual transport map identifies the part of the source on which the plan becomes deterministic. The area formula is applied only to this restricted part. On its complement, saturation of the Hessian bound supplies the trace estimate directly. Periodic lifts, the exact-marginal subgradient representation, the restricted Jacobian bound and the material pairing are established below; they are additional assertions needed here, rather than conclusions quoted from the transport literature.
The resulting field treats the time and convective contributions in the momentum equation together. After the kinetic-energy correction is separated, optimizer monotonicity makes this material contribution nonnegative. The mean Hessian and its upper bound give an L 2 L^2 L 2 cost of order K 1 / 2 K^{1/2} K 1/2 , while its trace detects density at scale K K K . Balancing these two estimates yields a uniform pressure tail. This gain supplies an integrability exponent strictly above 3 / 2 3/2 3/2 whenever γ > 4 / 3 \gamma> 4/3 γ > 4/3 . The effective-flux and entropy arguments can then be completed on the torus, with no positivity assumption on the density and no smallness condition on the initial energy.
Main results
Assume that the initial density ρ 0 \rho_0 ρ 0 and initial momentum m 0 m_0 m 0 satisfy
ρ 0 ≥ 0 , ρ 0 ∈ L γ ( T 3 ) , m 0 = 0 on { ρ 0 = 0 } , (5) \rho_0 \ge0,\qquad\rho_0 \in L^\gamma(\mathbb{T}^3),\qquad m_0 = 0\ \text{on }\{\rho_0 = 0\}, \tag*{(5)} ρ 0 ≥ 0 , ρ 0 ∈ L γ ( T 3 ) , m 0 = 0 on { ρ 0 = 0 } , (5)
and have finite initial total energy:
E 0 : = ∫ T 3 ( ∣ m 0 ∣ 2 2 ρ 0 + ρ 0 γ γ − 1 ) d x < ∞ , (6) E_0 := \int_{\mathbb{T}^3} \left(\frac{|m_0|^2}{2\rho_0} + \frac{\rho_0^\gamma}{\gamma- 1}\right)\,dx < \infty, \tag*{(6)} E 0 := ∫ T 3 ( 2 ρ 0 ∣ m 0 ∣ 2 + γ − 1 ρ 0 γ ) d x < ∞ , (6)
where the initial kinetic-energy density is defined to be zero where ρ 0 = 0 \rho_0 = 0 ρ 0 = 0 .
Definition 1.1 (Renormalized weak solution, see [27 ] ). A pair ( ρ , u ) (\rho,u) ( ρ , u ) is a renormalized weak solution of (1.1) with the prescribed initial data if the following conditions hold. The density satisfies ρ ≥ 0 \rho\ge0 ρ ≥ 0 , and, for every finite T > 0 T > 0 T > 0 ,
ρ ∈ L ∞ ( 0 , T ; L γ ) , u ∈ L 2 ( 0 , T ; H 1 ) , ρ u ∈ L ∞ ( 0 , T ; L 2 ) . \rho\in L^\infty(0,T;L^\gamma),\qquad u \in L^2(0,T;H^1),\qquad\sqrt{\rho}\,u \in L^\infty(0,T;L^2). ρ ∈ L ∞ ( 0 , T ; L γ ) , u ∈ L 2 ( 0 , T ; H 1 ) , ρ u ∈ L ∞ ( 0 , T ; L 2 ) . For all periodic test functions ϕ ∈ C c ∞ ( [ 0 , ∞ ) × T 3 ) \phi\in C^\infty_c([0,\infty)\times\mathbb{T}^3) ϕ ∈ C c ∞ ([ 0 , ∞ ) × T 3 ) and Φ ∈ C c ∞ ( [ 0 , ∞ ) × T 3 ; R 3 ) \Phi\in C^\infty_c([0,\infty)\times\mathbb{T}^3;\mathbb{R}^3) Φ ∈ C c ∞ ([ 0 , ∞ ) × T 3 ; R 3 ) , the mass and momentum equations hold in the form
∫ 0 ∞ ∫ T 3 [ ρ ϕ t + m ⋅ ∇ ϕ ] d x d t + ∫ T 3 ρ 0 ϕ ( 0 , x ) d x = 0 , (7) \int_0^\infty\int_{\mathbb{T}^3} [\rho\phi_t + m\cdot\nabla\phi]\,dx\,dt + \int_{\mathbb{T}^3}\rho_0\phi(0,x)\,dx = 0, \tag*{(7)} ∫ 0 ∞ ∫ T 3 [ ρ ϕ t + m ⋅ ∇ ϕ ] d x d t + ∫ T 3 ρ 0 ϕ ( 0 , x ) d x = 0 , (7) ∫ 0 ∞ ∫ T 3 [ m ⋅ Φ t + ( ρ u ⊗ u ) : ∇ Φ + ρ γ div Φ − S ( u ) : ∇ Φ ] d x d t + ∫ T 3 m 0 ⋅ Φ ( 0 , x ) d x = 0. (8) \int_0^\infty\int_{\mathbb{T}^3} [m\cdot\Phi_t + (\rho u\otimes u):\nabla\Phi+ \rho^\gamma\operatorname{div}\Phi- S(u):\nabla\Phi]\,dx\,dt + \int_{\mathbb{T}^3}m_0\cdot\Phi(0,x)\,dx = 0. \tag*{(8)} ∫ 0 ∞ ∫ T 3 [ m ⋅ Φ t + ( ρ u ⊗ u ) : ∇Φ + ρ γ div Φ − S ( u ) : ∇Φ ] d x d t + ∫ T 3 m 0 ⋅ Φ ( 0 , x ) d x = 0. (8) In addition, ρ \rho ρ satisfies the renormalized continuity equation with its initial term: for every b ∈ C 1 ( [ 0 , ∞ ) ) b \in C^1([0,\infty)) b ∈ C 1 ([ 0 , ∞ )) such that b ′ ( s ) = 0 b'(s)=0 b ′ ( s ) = 0 for all sufficiently large s s s , and every ϕ ∈ C c ∞ ( [ 0 , ∞ ) × T 3 ) \phi\in C^\infty_c([0,\infty)\times\mathbb{T}^3) ϕ ∈ C c ∞ ([ 0 , ∞ ) × T 3 ) ,
∫ 0 ∞ ∫ T 3 [ b ( ρ ) ∂ t ϕ + b ( ρ ) u ⋅ ∇ ϕ − ( ρ b ′ ( ρ ) − b ( ρ ) ) div u ϕ ] d x d t + ∫ T 3 b ( ρ 0 ) ϕ ( 0 , x ) d x = 0. (9) \int_0^\infty\int_{\mathbb{T}^3} \left[b(\rho)\partial_t\phi+ b(\rho)u\cdot\nabla\phi- \bigl(\rho b'(\rho)-b(\rho)\bigr)\operatorname{div}u\,\phi\right]\,dx\,dt + \int_{\mathbb{T}^3}b(\rho_0)\phi(0,x)\,dx = 0. \tag*{(9)} ∫ 0 ∞ ∫ T 3 [ b ( ρ ) ∂ t ϕ + b ( ρ ) u ⋅ ∇ ϕ − ( ρ b ′ ( ρ ) − b ( ρ ) ) div u ϕ ] d x d t + ∫ T 3 b ( ρ 0 ) ϕ ( 0 , x ) d x = 0. (9) The value b ( 0 ) b(0) b ( 0 ) is unrestricted. The solution also satisfies the distributional energy inequality (1.10) with initial energy E 0 E_0 E 0 from (1.6).
Theorem 1.2. Assume that γ > 4 / 3 \gamma> 4/3 γ > 4/3 , the viscosity coefficients satisfy (1.2), and the initial data ( ρ 0 , m 0 ) (\rho_0,m_0) ( ρ 0 , m 0 ) satisfy (1.5)–(1.6). Then system (1.1) admits a global finite-energy renormalized weak solution ( ρ , u ) (\rho,u) ( ρ , u ) in the sense of Definition 1.1.
Proof strategy
The proof has two distinct compactness stages. At fixed artificial-pressure coefficient the standard construction supplies enough integrability to use the geometric test. The estimates used to remove that coefficient must then depend only on mass, energy, and viscous dissipation. In particular, the enhanced density norm that makes a test admissible is not part of its uniform bound.
We first construct global weak solutions with the cubic artificial pressure P e ( ρ ) = ρ γ + e ρ 3 P_e(\rho)=\rho^\gamma+e\rho^3 P e ( ρ ) = ρ γ + e ρ 3 for fixed e > 0 e>0 e > 0 . A pressure estimate based on barycentric transport then allows us to pass to the limit e ↓ 0 e\downarrow0 e ↓ 0 .
The three-level approximation
Fix e > 0 e>0 e > 0 , G > max { 8 , γ + 1 } G>\max\{8,\gamma+1\} G > max { 8 , γ + 1 } , 0 < α ≤ 1 0<\alpha\le1 0 < α ≤ 1 , and σ > 0 \sigma>0 σ > 0 . For smooth positive initial density ρ ∗ \rho_* ρ ∗ and smooth velocity u ∗ u_* u ∗ , consider the periodic approximate system
∂ t ρ N + div ( ρ N u N ) = α Δ ρ N , ∂ t ( ρ N u N ) + div ( ρ N u N ⊗ u N ) + ∇ P ^ ( ρ N ) = div S ( u N ) − α ∑ j = 1 3 ( ∂ j ρ N ) ∂ j u N . (11) \begin{aligned}
\partial_t \rho_N+\operatorname{div}(\rho_N u_N)&=\alpha\Delta\rho_N,\\
\partial_t(\rho_Nu_N)+\operatorname{div}(\rho_Nu_N\otimes u_N)+\nabla\widehat{P}(\rho_N)&=\operatorname{div}S(u_N)-\alpha\sum_{j=1}^{3}(\partial_j\rho_N)\partial_j u_N. \tag*{(11)}
\end{aligned} ∂ t ρ N + div ( ρ N u N ) ∂ t ( ρ N u N ) + div ( ρ N u N ⊗ u N ) + ∇ P ( ρ N ) = α Δ ρ N , = div S ( u N ) − α j = 1 ∑ 3 ( ∂ j ρ N ) ∂ j u N . (11)
where
P ^ ( ρ ) = ρ γ + ϵ ρ 3 + σ ρ G , ( ρ N , u N ) ∣ t = 0 = ( ρ ∗ , u N 0 ) . \widehat{P}(\rho)=\rho^\gamma+\epsilon\rho^3+\sigma\rho^G,\qquad(\rho_N,u_N)|_{t=0}=(\rho_\ast,u_N^0). P ( ρ ) = ρ γ + ϵ ρ 3 + σ ρ G , ( ρ N , u N ) ∣ t = 0 = ( ρ ∗ , u N 0 ) .
Here u N ∈ X N u_N\in X_N u N ∈ X N , the Fourier–Galerkin space containing the constant modes, and u N 0 u_N^0 u N 0 is the L 2 ( ρ ∗ d x ) L^2(\rho_\ast\,dx) L 2 ( ρ ∗ d x ) -orthogonal projection of u ∗ u_\ast u ∗ . The momentum equation is projected onto X N X_N X N , while the density equation is not. The artificial-pressure and density-diffusion framework follows [14 , 15 , 27 ] . The added pressure improves density integrability at each construction level. The diffusion term makes the density equation parabolic, and its momentum correction cancels the additional kinetic-energy term; the cancellation is written out in Appendix § 5 . The choice of a cubic pressure at the last fixed level provides exactly the memberwise bound needed to pair m m m with ∂ t V \partial_tV ∂ t V . No priority claim is made for this exponent or for artificial pressure itself. Successive limits N → ∞ N\to\infty N → ∞ , α ↓ 0 \alpha\downarrow0 α ↓ 0 , and σ ↓ 0 \sigma\downarrow0 σ ↓ 0 , with ϵ \epsilon ϵ fixed, yield a global renormalized weak solution for P ϵ P_\epsilon P ϵ , including the initial terms; see Appendix § 5 .
We then take ϵ = ϵ n ↓ 0 \epsilon=\epsilon_n\downarrow0 ϵ = ϵ n ↓ 0 , with initial data from (A.1) and Lemma 2.1 . The energy bounds are uniform in n n n , but the L t ∞ L x 3 L_t^\infty L_x^3 L t ∞ L x 3 density bounds are not. This requires a pressure estimate independent of ϵ n \epsilon_n ϵ n .
From classical tests to barycentric relaxation
The classical test in (1.3) (with θ \theta θ obeys (1.4)) locates the loss of integrability. The geometric construction below instead arranges that the time and convective contributions, after the kinetic correction, have a favorable combined sign. Its Hessian must simultaneously be affordable in the stress pairing.
Serre’s compensated integrability suggests treating inertia through the nonnegative space–time tensor associated with the Euler equations:
M + = ( ρ m T m ρ u ⊗ u + P Id ) ⪰ 0 , Div t , x M + = 0 , ( det M + ) 1 / 3 = ρ 1 / 3 P . (12) \mathbb{M}_+=\begin{pmatrix}\rho&m^\mathsf{T}\\m&\rho u\otimes u+P\operatorname{Id}\end{pmatrix}\succeq0,\qquad\operatorname{Div}_{t,x}\mathbb{M}_+=0,\qquad(\det\mathbb{M}_+)^{1/3}=\rho^{1/3}P. \tag*{(12)} M + = ( ρ m m T ρ u ⊗ u + P Id ) ⪰ 0 , Div t , x M + = 0 , ( det M + ) 1/3 = ρ 1/3 P . (12)
In the smooth setting, put g = ( det M + ) 1 / 3 g=(\det\mathbb{M}_+)^{1/3} g = ( det M + ) 1/3 on a bounded space–time domain D ⊂ R 4 D\subset\mathbb{R}^4 D ⊂ R 4 . A convex potential satisfying det D 2 Ψ = g / ∫ D g \det D^2\Psi=g/\int_D g det D 2 Ψ = g / ∫ D g gives, by the matrix arithmetic–geometric mean inequality,
M + : D 2 Ψ ≥ 4 ( ∫ D g ) − 1 / 4 g . \mathbb{M}_+:D^2\Psi\ge4\left(\int_D g\right)^{-1/4}g. M + : D 2 Ψ ≥ 4 ( ∫ D g ) − 1/4 g .
Integration by parts converts the left-hand side into a boundary flux, controlling ρ 1 / 3 P \rho^{1/3}P ρ 1/3 P , see [35 , 36 ] .
For the Navier–Stokes equations,
Div t , x M + = ( 0 div S ) , \operatorname{Div}_{t,x}\mathbb{M}_+=\begin{pmatrix}0\\\operatorname{div}S\end{pmatrix}, Div t , x M + = ( 0 div S ) ,
and the same test incurs the viscous cost
∣ ∫ D S : D x 2 Ψ d z ∣ ≤ ∥ S ∥ L 2 ( D ) ∥ D x 2 Ψ ∥ L 2 ( D ) . (13) \left|\int_D S : D_x^2\Psi\, dz\right| \le\|S\|_{L^2(D)}\|D_x^2\Psi\|_{L^2(D)}. \tag*{(13)} ∫ D S : D x 2 Ψ d z ≤ ∥ S ∥ L 2 ( D ) ∥ D x 2 Ψ ∥ L 2 ( D ) . (13)
The Monge–Ampère determinant does not control the L 2 L^2 L 2 norm of the Hessian. We therefore relax the transport constraint, retaining a density-sensitive trace bound while controlling the Hessian.
Following weak and barycentric transport in [16 , 17 ] , particularly the mixed quadratic cost of Alibert–Bouchitté–Champion (see [1 Section 5.2] ), we allow mass splitting and use the conditional barycenter as the test map. For ρ ≥ 0 \rho\ge0 ρ ≥ 0 and a fixed background α b g > 0 \alpha_{\mathrm{bg}} > 0 α bg > 0 , set
c = ∫ T 3 ( ρ + α b g ) d x , σ = ρ + α b g c . c = \int_{\mathbb{T}^3}(\rho+\alpha_{\mathrm{bg}})\,dx,\qquad\sigma= \frac{\rho+\alpha_{\mathrm{bg}}}{c}. c = ∫ T 3 ( ρ + α bg ) d x , σ = c ρ + α bg .
Here σ \sigma σ is the normalized transport density, distinct from the auxiliary-pressure coefficient in (1.11). For plans π \pi π with source marginal ( ρ + α b g ) d x (\rho+\alpha_{\mathrm{bg}})\,dx ( ρ + α bg ) d x and target marginal c d x c\,dx c d x , write z ˉ π ( x ) = E π [ z ∣ x ] \bar{z}_{\pi}(x)=\mathbb{E}_{\pi}[z\mid x] z ˉ π ( x ) = E π [ z ∣ x ] . With periodic integer lifts, the cost is
J K ( π ) = 1 2 ∫ [ K − 1 K ∣ z − x ∣ 2 + 1 K ∣ z ˉ π ( x ) − x ∣ 2 ] d π , K ≥ 6. J_K(\pi)=\frac{1}{2}\int\left[\frac{K-1}{K}|z-x|^2+\frac{1}{K}|\bar{z}_{\pi}(x)-x|^2\right]\,d\pi,\qquad K\ge6. J K ( π ) = 2 1 ∫ [ K K − 1 ∣ z − x ∣ 2 + K 1 ∣ z ˉ π ( x ) − x ∣ 2 ] d π , K ≥ 6.
The full plan has the prescribed target marginal, while its barycenter need not. The construction in § 3 gives a unique periodic barycentric map, with the additional properties proved here,
{ T = ∇ φ = x + V , H = D T = D 2 φ , 0 ⪯ H ⪯ K Id , ∫ T 3 H d x = Id . \left\{
\begin{aligned}
T&=\nabla\varphi=x+V, & H&=DT=D^2\varphi,\\
0&\preceq H\preceq K\operatorname{Id}, & \int_{\mathbb{T}^3}H\,dx&=\operatorname{Id}.
\end{aligned}
\right. ⎩ ⎨ ⎧ T 0 = ∇ φ = x + V , ⪯ H ⪯ K Id , H ∫ T 3 H d x = D T = D 2 φ , = Id .
Since V V V is uniformly bounded, ∇ V = H − Id \nabla V=H-\operatorname{Id} ∇ V = H − Id , and H 2 ⪯ K H H^2\preceq KH H 2 ⪯ K H , we obtain
{ ∫ T 3 ∣ H − Id ∣ 2 d x ≤ 3 ( K − 1 ) , ∣ ∫ Q T S : ∇ V d x d t ∣ ≤ 3 T ( K − 1 ) ∥ S ∥ L 2 ( Q T ) . (14) \left\{
\begin{aligned}
\int_{\mathbb{T}^3}|H-\operatorname{Id}|^2\,dx&\le3(K-1),\\
\left|\int_{Q_T}S:\nabla V\,dx\,dt\right|&\le\sqrt{3T(K-1)}\,\|S\|_{L^2(Q_T)}.
\end{aligned}
\right. \tag*{(14)} ⎩ ⎨ ⎧ ∫ T 3 ∣ H − Id ∣ 2 d x ∫ Q T S : ∇ V d x d t ≤ 3 ( K − 1 ) , ≤ 3 T ( K − 1 ) ∥ S ∥ L 2 ( Q T ) . (14)
To recover the pressure weight without the Jacobian identity, use the residual map y = ( K x − T ) / ( K − 1 ) y=(Kx-T)/(K-1) y = ( K x − T ) / ( K − 1 ) . On the uncapped region the area formula gives det H ≥ σ \det H\ge\sigma det H ≥ σ ; on its complement an eigenvalue of H H H equals K K K . Hence
tr H ≥ min { K , 3 σ 1 / 3 } , (15) \operatorname{tr}H\ge\min\{K,3\sigma^{1/3}\}, \tag*{(15)} tr H ≥ min { K , 3 σ 1/3 } , (15)
as proved in Lemma 3.1 .
The time-dependent field requires a further sign property. Along a smooth particle trajectory X ˙ = u ( t , X ) \dot{X}=u(t,X) X ˙ = u ( t , X ) ,
d d t T ( t , X ( t ) ) = ∂ t V ( t , X ( t ) ) + H ( t , X ( t ) ) u ( t , X ( t ) ) . \frac{d}{dt}T(t,X(t))=\partial_tV(t,X(t))+H(t,X(t))u(t,X(t)). d t d T ( t , X ( t )) = ∂ t V ( t , X ( t )) + H ( t , X ( t )) u ( t , X ( t )) .
Applying the optimizer’s monotonicity to the moving mass and stationary background gives
M ( η ) : = ∫ Q T η [ m ⋅ ∂ t V + ( ρ u ⊗ u ) : H ] ≥ 0 , 0 ≤ η ∈ C c ∞ ( 0 , T ) . (16) \mathcal{M}(\eta):=\int_{Q_T}\eta\left[m\cdot\partial_tV+(\rho u\otimes u):H\right]\ge0,\qquad0\le\eta\in C_c^\infty(0,T). \tag*{(16)} M ( η ) := ∫ Q T η [ m ⋅ ∂ t V + ( ρ u ⊗ u ) : H ] ≥ 0 , 0 ≤ η ∈ C c ∞ ( 0 , T ) . (16)
The convective term in the momentum equation involves H − I d H-\mathrm{Id} H − Id , so the kinetic-energy contribution is separated before applying this sign. Lemma 3.2 justifies the pairing for the approximate solutions.
Pressure gain and the final limit
For P n = ρ n γ + ϵ n ρ n 3 P_n=\rho_n^\gamma+\epsilon_n\rho_n^3 P n = ρ n γ + ϵ n ρ n 3 , testing the momentum equation with η V \eta V η V gives
∫ Q T η P n ( tr H − 3 ) + M n ( η ) = − ∫ Q T η ′ m n ⋅ V + ∫ Q T η ρ n ∣ u n ∣ 2 + ∫ Q T η S n : ( H − I d ) . (17) \int_{Q_T}\eta P_n(\operatorname{tr}H-3)+\mathcal{M}_n(\eta)
=-\int_{Q_T}\eta' m_n\cdot V+\int_{Q_T}\eta\rho_n\lvert u_n\rvert^2+\int_{Q_T}\eta S_n:(H-\mathrm{Id}). \tag*{(17)} ∫ Q T η P n ( tr H − 3 ) + M n ( η ) = − ∫ Q T η ′ m n ⋅ V + ∫ Q T η ρ n ∣ u n ∣ 2 + ∫ Q T η S n : ( H − Id ) . (17)
The material term is nonnegative, the remaining inertial terms are controlled by energy, and (1.14) gives
∣ ∫ Q T η S n : ( H − I d ) ∣ ≤ C T ∥ η ∥ ∞ K 1 / 2 . \left\lvert\int_{Q_T}\eta S_n:(H-\mathrm{Id})\right\rvert\le C_T\lVert\eta\rVert_\infty K^{1/2}. ∫ Q T η S n : ( H − Id ) ≤ C T ∥ η ∥ ∞ K 1/2 .
Consequently,
∫ Q T η P n tr H ≤ C T ( ∥ η ′ ∥ L 1 + ∥ η ∥ ∞ ) ( 1 + K 1 / 2 ) . \int_{Q_T}\eta P_n\operatorname{tr}H\le C_T\left(\lVert\eta'\rVert_{L^1}+\lVert\eta\rVert_\infty\right)(1+K^{1/2}). ∫ Q T η P n tr H ≤ C T ( ∥ η ′ ∥ L 1 + ∥ η ∥ ∞ ) ( 1 + K 1/2 ) .
Combining this bound with (1.15), then removing the time cutoff and choosing the cap height, yields
sup n ∫ Q T P n min { δ − 1 , ρ n 1 / 3 } ≤ C T δ − 1 / 2 , 0 < δ ≤ 1. (18) \sup_n\int_{Q_T}P_n\min\{\delta^{-1},\rho_n^{1/3}\}\le C_T\delta^{-1/2},\qquad0<\delta\le1. \tag*{(18)} n sup ∫ Q T P n min { δ − 1 , ρ n 1/3 } ≤ C T δ − 1/2 , 0 < δ ≤ 1. (18)
Taking δ = R − 1 / 3 \delta=R^{-1/3} δ = R − 1/3 for R ≥ 1 R\ge1 R ≥ 1 gives
sup n ∫ Q T ∩ { ρ n > R } P n ≤ C T R − 1 / 6 . (19) \sup_n\int_{Q_T\cap\{\rho_n>R\}}P_n\le C_TR^{-1/6}. \tag*{(19)} n sup ∫ Q T ∩ { ρ n > R } P n ≤ C T R − 1/6 . (19)
In particular,
∣ Q T ∩ { ρ n > R } ∣ ≤ C T R − γ − 1 / 6 , \lvert Q_T\cap\{\rho_n>R\}\rvert\le C_TR^{-\gamma-1/6}, ∣ Q T ∩ { ρ n > R }∣ ≤ C T R − γ − 1/6 ,
so ρ n \rho_n ρ n is uniformly bounded in L r ( Q T ) L^r(Q_T) L r ( Q T ) for 1 ≤ r < γ + 1 / 6 1\le r<\gamma+1/6 1 ≤ r < γ + 1/6 . The detailed pressure estimate and the artificial-pressure limit are proved in Lemmas 3.3 and 4.1.
When γ > 4 / 3 \gamma>4/3 γ > 4/3 , we may choose
{ 3 2 < q < min { 2 , γ + 1 6 } , κ ( q ) = 2 q − 3 q + 1 > 0 , p ( q ) = 6 ( q + 1 ) q + 11 > 6 5 . (20) \begin{cases}
\frac{3}{2}<q<\min\left\{2,\gamma+\frac{1}{6}\right\},\\
\kappa(q)=\frac{2q-3}{q+1}>0,\\
p(q)=\frac{6(q+1)}{q+11}>\frac{6}{5}. \tag*{(20)}
\end{cases} ⎩ ⎨ ⎧ 2 3 < q < min { 2 , γ + 6 1 } , κ ( q ) = q + 1 2 q − 3 > 0 , p ( q ) = q + 11 6 ( q + 1 ) > 5 6 . (20)
These exponents control the high-frequency effective-flux remainder and give strong momentum compactness. To identify the pressure limit, the spatially integrated effective viscous flux is combined with the renormalized continuity equation. For the truncation T k T_k T k and its convex entropy L k L_k L k , the resulting defect relation is
D k ′ ( t ) + 1 ν ∫ T 3 C k d x = ∫ T 3 Q k div u d x , D k ( 0 ) = 0. (21) \mathcal{D}_k'(t)+\frac{1}{\nu}\int_{\mathbb{T}^3}C_k\,\mathrm{d}x=\int_{\mathbb{T}^3}Q_k\operatorname{div}u\,\mathrm{d}x,\qquad\mathcal{D}_k(0)=0. \tag*{(21)} D k ′ ( t ) + ν 1 ∫ T 3 C k d x = ∫ T 3 Q k div u d x , D k ( 0 ) = 0. (21)
Here D k ≥ 0 \mathcal{D}_k \ge0 D k ≥ 0 is the truncated entropy defect, C k ≥ 0 C_k \ge0 C k ≥ 0 is the pressure–truncation covariance, and Q k = T k ( ρ ) − ⟨ T k ( z ) ⟩ Q_k = T_k(\rho) - \langle T_k(z)\rangle Q k = T k ( ρ ) − ⟨ T k ( z )⟩ . Since Q k → 0 Q_k \to0 Q k → 0 in L 2 ( Q T ) L^2(Q_T) L 2 ( Q T ) , the common initial entropy and strict convexity of z log z z \log z z log z eliminate density oscillations; see Lemmas 4.3 and 4.4 .
Finally, the momentum bound with p ( q ) > 6 / 5 p(q) > 6/5 p ( q ) > 6/5 and the time regularity from the momentum equation yield m n → ρ u m_n \to\rho u m n → ρ u in L t 2 H x − 1 L^2_t H^{-1}_x L t 2 H x − 1 . Together with weak velocity convergence in L t 2 H x 1 L^2_t H^1_x L t 2 H x 1 , this identifies convection and the kinetic-energy limit. Proposition 2.3 then gives the limiting momentum equation and energy inequality.
The rest of this paper is organized as follows: In § 2 , we present the approximate solutions, and state the main estimates used to prove the main theorem. In § 3 , we establish the spatial geometry, the material-pairing sign, and the complete pressure and tail estimates. In § 4 , we establish the desired estimates of the effective viscous flux, and then show the nonlinear convergence.
Approximate Solutions and Main Estimates
We state the three results used to prove Theorem 1.2. Lemma 2.1, proved in Appendix § 5 , supplies solutions with cubic artificial pressure and the uniform energy bound (2.2). Proposition 2.2 gives the pressure and tail estimates; its proof in § 3 consists of the barycentric construction (see Lemma 3.1 ), the material-pairing estimate (see Lemma 3.2 ), and the momentum test (see Lemma 3.3 ). Proposition 2.3 is proved in § 4 : the pressure estimate first supplies the integrability for the effective-flux and entropy arguments, which identify the density and pressure limits; momentum compactness then gives the convective and energy limits. We first state these results and recall the periodic tools.
The approximation has a classical part and a requirement specific to the geometric test. The classical part consists of a Fourier–Galerkin momentum equation, artificial density diffusion, and artificial pressure, see [14 , 15 , 27 , 33 ] . Here the last retained artificial term is ϵ ρ 3 \epsilon\rho^3 ϵ ρ 3 . An auxiliary term σ ρ G \sigma\rho^G σ ρ G , with G > max { 8 , γ + 1 } G > \max\{8,\gamma+ 1\} G > max { 8 , γ + 1 } , is used while the Galerkin dimension and diffusion are removed. Neither exponent is a new constitutive law for the limiting fluid. The resulting fixed-ϵ \epsilon ϵ solution must solve the full initialized equations and have enough integrability for the time-dependent barycentric field to be an admissible momentum test.
The order of construction in Appendix § 5 is as follows. With ϵ \epsilon ϵ , σ \sigma σ , α \alpha α fixed, the limit N → ∞ N \to\infty N → ∞ uses parabolic density compactness. The uniform L 5 G / 3 ( Q T ) L^{5G/3}(Q_T) L 5 G /3 ( Q T ) bound supplies strong density convergence below this exponent and identifies P ^ \widehat{P} P . Strong convergence of ∇ ρ N \nabla\rho_N ∇ ρ N at fixed α \alpha α is separately needed for the diffusion correction in momentum; it follows from (A.9) and the terminal-layer estimate, not from weak gradient convergence.
Next, with ϵ \epsilon ϵ , σ \sigma σ fixed, α ↓ 0 \alpha\downarrow0 α ↓ 0 removes diffusion and its momentum correction. The classical test A ρ A\rho A ρ gives a uniform L G + 1 L^{G+1} L G + 1 density bound. The effective flux and initialized entropy comparison then give strong density convergence below $G+1. and identify the full pressure. Finally, with e e e fixed, σ ↓ 0 \sigma\downarrow0 σ ↓ 0 uses the retained cubic pressure and (A.22). The auxiliary pressure vanishes strongly, and a truncated flux and entropy argument identifies P e P_e P e . The constants in this last step may depend on e e e ; they are not the constants used in Proposition 2.2.
At every level the initial terms are included before taking limits. The weighted Galerkin projection recovers the initial kinetic energy; the smooth initial data stay fixed in the two subsequent limits. Strong convergence of the energy densities and lower semicontinuity of the dissipation yield the initialized energy inequality at that level. Only after these three limits are complete do we choose e = e n → 0 e=e_n\to0 e = e n → 0 and use the finite-energy preparation (A.1). This prevents the final density compactness statement from being used to construct its own input sequence.
The hypotheses of Proposition 2.2 deliberately separate two uses of regularity. For each member of the sequence,
∥ ρ n u n ∥ L t 2 L x 2 ≤ ∥ ρ n ∥ L t ∞ L x 3 ∥ u n ∥ L t 2 L x 6 < ∞ \|\rho_n u_n\|_{L_t^2L_x^2}\leq\|\rho_n\|_{L_t^\infty L_x^3}\|u_n\|_{L_t^2L_x^6}<\infty ∥ ρ n u n ∥ L t 2 L x 2 ≤ ∥ ρ n ∥ L t ∞ L x 3 ∥ u n ∥ L t 2 L x 6 < ∞
makes the time-dependent momentum test legitimate. In the resulting inequality this norm disappears: only E ∗ E_\ast E ∗ remains. Thus the order is to remove the smoothing at fixed n n n , establish the common pressure budget, and only then send n → ∞ n\to\infty n → ∞ . Proposition 2.3 uses that budget without returning to the memberwise L 3 L^3 L 3 bound.
Lemma 2.1. Assume (1.5)–(1.6) with total mass m > 0 \mathbf{m}>0 m > 0 . For any prescribed sequence e n > 0 e_n>0 e n > 0 with e n → 0 e_n\to0 e n → 0 , there exist smooth positive initial densities, smooth initial velocities, and global periodic approximate solutions ( ρ n , u n ) (\rho_n,u_n) ( ρ n , u n ) with
m n = ρ n u n , P n = ρ n γ + e n ρ n 3 . m_n=\rho_nu_n,\qquad P_n=\rho_n^\gamma+e_n\rho_n^3. m n = ρ n u n , P n = ρ n γ + e n ρ n 3 . These solutions satisfy the continuity, momentum, and renormalized continuity equations, including their initial terms, and conserve the mass m \mathbf{m} m . Their initial data satisfy
ρ 0 n → ρ 0 in L γ , m 0 n → m 0 in L r m , ρ 0 n u 0 n → W 0 in L 2 , (22) \rho_{0n}\to\rho_0\quad\text{in }L^\gamma,\qquad m_{0n}\to m_0\quad\text{in }L^{r_m},\qquad\sqrt{\rho_{0n}}u_{0n}\to W_0\quad\text{in }L^2, \tag*{(22)} ρ 0 n → ρ 0 in L γ , m 0 n → m 0 in L r m , ρ 0 n u 0 n → W 0 in L 2 , (22) e n ∫ T 3 ρ 0 n 3 d x → 0 , E n 0 → E 0 , e_n\int_{\mathbb{T}^3}\rho_{0n}^3\,dx\to0,\qquad E_{n0}\to E_0, e n ∫ T 3 ρ 0 n 3 d x → 0 , E n 0 → E 0 , where r m = 2 γ / ( γ + 1 ) r_m=2\gamma/(\gamma+1) r m = 2 γ / ( γ + 1 ) and W 0 = m 0 / ρ 0 W_0=m_0/\sqrt{\rho_0} W 0 = m 0 / ρ 0 , extended by zero on vacuum. For every finite T > 0 T>0 T > 0 and nonnegative ψ ∈ C c ∞ ( [ 0 , T ) ) \psi\in C_c^\infty([0,T)) ψ ∈ C c ∞ ([ 0 , T )) ,
− ∫ 0 T E n ψ ′ d t + ∫ 0 T D n ψ d t ≤ E n 0 ψ ( 0 ) , -\int_0^T E_n\psi'\,dt+\int_0^T D_n\psi\,dt\leq E_{n0}\psi(0), − ∫ 0 T E n ψ ′ d t + ∫ 0 T D n ψ d t ≤ E n 0 ψ ( 0 ) , where
E n = ∫ T 3 ( ∣ m n ∣ 2 2 ρ n + ρ n γ γ − 1 + e n 2 ρ n 3 ) d x , E_n=\int_{\mathbb{T}^3}\left(\frac{|m_n|^2}{2\rho_n}+\frac{\rho_n^\gamma}{\gamma-1}+\frac{e_n}{2}\rho_n^3\right)\,dx, E n = ∫ T 3 ( 2 ρ n ∣ m n ∣ 2 + γ − 1 ρ n γ + 2 e n ρ n 3 ) d x , D n = μ ∥ ∇ u n ∥ 2 2 + ( λ + μ ) ∥ div u n ∥ 2 2 . D_n=\mu\|\nabla u_n\|_2^2+(\lambda+\mu)\|\operatorname{div}u_n\|_2^2. D n = μ ∥∇ u n ∥ 2 2 + ( λ + μ ) ∥ div u n ∥ 2 2 . In particular, ρ n ∈ L ∞ ( 0 , T ; L 3 ( T 3 ) ) \rho_n\in L^\infty(0,T;L^3(\mathbb{T}^3)) ρ n ∈ L ∞ ( 0 , T ; L 3 ( T 3 )) for each n n n , and
ess sup 0 < t < T ∫ T 3 ( ρ n + ρ n γ + ρ n ∣ u n ∣ 2 + P n ) d x + ∫ 0 T ( ∥ ∇ u n ∥ 2 2 + ∥ u n ∥ 6 2 ) d t ≤ C T , (23) \operatorname*{ess\,sup}_{0<t<T}\int_{\mathbb{T}^3}\left(\rho_n+\rho_n^\gamma+\rho_n|u_n|^2+P_n\right)\,dx+\int_0^T\left(\|\nabla u_n\|_2^2+\|u_n\|_6^2\right)\,dt\leq C_T, \tag*{(23)} 0 < t < T ess sup ∫ T 3 ( ρ n + ρ n γ + ρ n ∣ u n ∣ 2 + P n ) d x + ∫ 0 T ( ∥∇ u n ∥ 2 2 + ∥ u n ∥ 6 2 ) d t ≤ C T , (23) where C T C_T C T is independent of n n n .
Proposition 2.2. Let ρ n \rho_n ρ n , P n ≥ 0 P_n \ge0 P n ≥ 0 and m n = ρ n u n m_n=\rho_n u_n m n = ρ n u n satisfy
{ ∂ t ρ n + div m n = 0 , ∂ t m n + div ( ρ n u n ⊗ u n ) + ∇ P n = div S n . \begin{cases}
\partial_t\rho_n+\operatorname{div}m_n=0,\\
\partial_t m_n+\operatorname{div}(\rho_n u_n\otimes u_n)+\nabla P_n=\operatorname{div}S_n.
\end{cases} { ∂ t ρ n + div m n = 0 , ∂ t m n + div ( ρ n u n ⊗ u n ) + ∇ P n = div S n . Assume that
sup n [ ess sup 0 < t < T ∫ T 3 ( ρ n + ρ n ∣ u n ∣ 2 + P n ) d x + ∫ Q T ∣ S n ∣ 2 d x d t ] ≤ E ∗ , \sup_n\left[\operatorname*{ess\,sup}_{0<t<T}\int_{\mathbb{T}^3}(\rho_n+\rho_n|u_n|^2+P_n)\,\mathrm{d}x+\int_{Q_T}|S_n|^2\,\mathrm{d}x\,\mathrm{d}t\right]\le E_*, n sup [ 0 < t < T ess sup ∫ T 3 ( ρ n + ρ n ∣ u n ∣ 2 + P n ) d x + ∫ Q T ∣ S n ∣ 2 d x d t ] ≤ E ∗ , and, for each fixed n n n ,
ρ n ∈ L ∞ ( 0 , T ; L 3 ( T 3 ) ) , u n ∈ L 2 ( 0 , T ; H 1 ( T 3 ) ) . \rho_n\in L^\infty(0,T;L^3(\mathbb{T}^3)),\qquad u_n\in L^2(0,T;H^1(\mathbb{T}^3)). ρ n ∈ L ∞ ( 0 , T ; L 3 ( T 3 )) , u n ∈ L 2 ( 0 , T ; H 1 ( T 3 )) . Then, for some constant C T C_T C T independent of n n n ,
sup n ∫ Q T P n min { δ − 1 , ρ n 1 / 3 } d x d t ≤ C T δ − 1 / 2 , 0 < δ ≤ 1 , (24) \sup_n\int_{Q_T}P_n\min\{\delta^{-1},\rho_n^{1/3}\}\,\mathrm{d}x\,\mathrm{d}t\le C_T\delta^{-1/2},\qquad0<\delta\le1, \tag*{(24)} n sup ∫ Q T P n min { δ − 1 , ρ n 1/3 } d x d t ≤ C T δ − 1/2 , 0 < δ ≤ 1 , (24) sup n ∫ Q T ∩ { ρ n > R } P n d x d t ≤ C T R − 1 / 6 , R ≥ 1. (25) \sup_n\int_{Q_T\cap\{\rho_n>R\}}P_n\,\mathrm{d}x\,\mathrm{d}t\le C_TR^{-1/6},\qquad R\ge1. \tag*{(25)} n sup ∫ Q T ∩ { ρ n > R } P n d x d t ≤ C T R − 1/6 , R ≥ 1. (25) The constant depends only on E ∗ E_* E ∗ , T T T , and the torus.
Proposition 2.3. For the approximate solutions of Lemma 2.1, there is a single subsequence valid on every finite time interval such that u n ⇀ u u_n\rightharpoonup u u n ⇀ u in L 2 ( 0 , T ; H 1 ( T 3 ) ) L^2(0,T;H^1(\mathbb{T}^3)) L 2 ( 0 , T ; H 1 ( T 3 )) and the following conclusions hold.
(i) The density Young measure is concentrated at ρ ( t , x ) \rho(t,x) ρ ( t , x ) almost everywhere, that is, Y t , x = δ ρ ( t , x ) Y_{t,x}=\delta_{\rho(t,x)} Y t , x = δ ρ ( t , x ) , and
ρ n → ρ in L r ( Q T ) , 1 ≤ r < γ + 1 6 , \rho_n\to\rho\quad\text{in }L^r(Q_T),\qquad1\le r<\gamma+\frac{1}{6}, ρ n → ρ in L r ( Q T ) , 1 ≤ r < γ + 6 1 , ρ n γ → ρ γ in L s ( Q T ) , 1 ≤ s < 1 + 1 6 γ , \rho_n^\gamma\to\rho^\gamma\quad\text{in }L^s(Q_T),\qquad1\le s<1+\frac{1}{6\gamma}, ρ n γ → ρ γ in L s ( Q T ) , 1 ≤ s < 1 + 6 γ 1 , P n → ρ γ in L s ( Q T ) , 1 ≤ s < 1 + 1 6 max { 3 , γ } . P_n\to\rho^\gamma\quad\text{in }L^s(Q_T),\qquad1\le s<1+\frac{1}{6\max\{3,\gamma\}}. P n → ρ γ in L s ( Q T ) , 1 ≤ s < 1 + 6 max { 3 , γ } 1 . The artificial pressure satisfies the L s L^s L s convergence estimate (54) . Moreover, m n → m = ρ u m_n\to m=\rho u m n → m = ρ u in distributions, including tests whose support meets t = 0 t=0 t = 0 . The density satisfies ρ ∈ C w ( [ 0 , T ] ; L γ ( T 3 ) ) \rho\in C_w([0,T];L^\gamma(\mathbb{T}^3)) ρ ∈ C w ([ 0 , T ] ; L γ ( T 3 )) and ρ ( 0 ) = ρ 0 \rho(0)=\rho_0 ρ ( 0 ) = ρ 0 . The continuity and renormalized continuity equations retain the initial terms ρ 0 \rho_0 ρ 0 and b ( ρ 0 ) b(\rho_0) b ( ρ 0 ) , respectively. The kinetic energy satisfies
ess sup 0 < t < T ∫ T 3 ρ ∣ u ∣ 2 d x ≤ sup n ess sup 0 < t < T ∫ T 3 ρ n ∣ u n ∣ 2 d x . \operatorname*{ess\,sup}_{0<t<T}\int_{\mathbb{T}^3}\rho|u|^2\,\mathrm{d}x\le\sup_n\operatorname*{ess\,sup}_{0<t<T}\int_{\mathbb{T}^3}\rho_n|u_n|^2\,\mathrm{d}x. 0 < t < T ess sup ∫ T 3 ρ ∣ u ∣ 2 d x ≤ n sup 0 < t < T ess sup ∫ T 3 ρ n ∣ u n ∣ 2 d x . (ii) Along the same subsequence as in (i),
m n → ρ u in C ( [ 0 , T ] ; H − 4 ( T 3 ) ) ∩ L 2 ( 0 , T ; H − 1 ( T 3 ) ) , m_n\to\rho u\quad\text{in }C([0,T];H^{-4}(\mathbb{T}^3))\cap L^2(0,T;H^{-1}(\mathbb{T}^3)), m n → ρ u in C ([ 0 , T ] ; H − 4 ( T 3 )) ∩ L 2 ( 0 , T ; H − 1 ( T 3 )) , ρ n u n → ρ u in L 2 ( Q T ) , \sqrt{\rho_n}u_n\to\sqrt{\rho}u\quad\text{in }L^2(Q_T), ρ n u n → ρ u in L 2 ( Q T ) , ρ n u n ⊗ u n → ρ u ⊗ u in L 1 ( Q T ) . \rho_nu_n\otimes u_n\to\rho u\otimes u\quad\text{in }L^1(Q_T). ρ n u n ⊗ u n → ρ u ⊗ u in L 1 ( Q T ) . The total-energy density converges strongly in L 1 ( Q T ) L^1(Q_T) L 1 ( Q T ) , and its spatial integral satisfies (80) , allowing energy tests with sign-changing time derivatives. The limiting momentum equation and energy inequality retain the initial data ( ρ 0 , m 0 ) (\rho_0,m_0) ( ρ 0 , m 0 ) and energy E 0 E_0 E 0 , with weakly continuous representatives and the stated conservation laws.
Unless otherwise stated, spatial integrals are over T 3 \mathbb{T}^{3} T 3 . We abbreviate L p L q = L p ( 0 , T ; L q ( T 3 ) ) L^{p}L^{q}=L^{p}(0,T;L^{q}(\mathbb{T}^{3})) L p L q = L p ( 0 , T ; L q ( T 3 )) and C t H − 1 = C ( [ 0 , T ] ; H − 1 ( T 3 ) ) C_{t}H^{-1}=C([0,T];H^{-1}(\mathbb{T}^{3})) C t H − 1 = C ([ 0 , T ] ; H − 1 ( T 3 )) ; notation with subscripts t , x t,x t , x has the same meaning. We use ∇ V = ( ∂ j V i ) i j \nabla V=(\partial_{j}V_{i})_{ij} ∇ V = ( ∂ j V i ) ij and A : B = ∑ i j A i j B i j A:B=\sum_{ij}A_{ij}B_{ij} A : B = ∑ ij A ij B ij , with summation over repeated indices. The space H − j H^{-j} H − j is the dual of the inhomogeneous Sobolev space H j H^{j} H j , including the constant mode. We write f ˉ = ∫ T 3 f d x \bar{f}=\int_{\mathbb{T}^{3}}f\,dx f ˉ = ∫ T 3 f d x for the spatial average and ⟨ F ⟩ = ∫ F ( z ) Y t , x ( d z ) \langle F\rangle=\int F(z)Y_{t,x}(dz) ⟨ F ⟩ = ∫ F ( z ) Y t , x ( d z ) for expectation with respect to the Young measure.
The effective-viscous-flux method of [13 , 14 , 27 ] uses a gradient right inverse of divergence. On the torus the divergence of a periodic vector field has zero mean, so the density test must retain its mean correction. Define the periodic inverse Laplacian to be zero at frequency zero, and define
A i = ∂ i Δ T 3 − 1 , R i j = ∂ j A i , ∑ i R i i f = f − f ˉ . A_{i}=\partial_{i}\Delta_{\mathbb{T}^{3}}^{-1},\qquad R_{ij}=\partial_{j}A_{i},\qquad\sum_{i}R_{ii}f=f-\bar{f}. A i = ∂ i Δ T 3 − 1 , R ij = ∂ j A i , i ∑ R ii f = f − f ˉ .
The periodic Calderón–Zygmund estimate (see [40 ] ) gives
∥ R i j f ∥ r + ∥ A f ∥ W 1 , r ≤ C r ∥ f ∥ r , 1 < r < ∞ . (26) \lVert R_{ij}f\rVert_{r}+\lVert Af\rVert_{W^{1,r}}\leq C_{r}\lVert f\rVert_{r},\qquad1<r<\infty. \tag*{(26)} ∥ R ij f ∥ r + ∥ A f ∥ W 1 , r ≤ C r ∥ f ∥ r , 1 < r < ∞. (26)
To obtain this estimate, decompose the periodic Green function locally into the Newtonian kernel and a smooth remainder. Lemma C.2 in Appendix § 7 gives the high-frequency estimates.
Choose a smooth, nondecreasing, concave function T \mathrm{T} T that equals the identity on [ 0 , 1 ] [0,1] [ 0 , 1 ] , equals 2 2 2 on [ 3 , ∞ ) [3,\infty) [ 3 , ∞ ) , and satisfies 0 ≤ T ′ ≤ 1 0\leq\mathrm{T}'\leq1 0 ≤ T ′ ≤ 1 . Define
T k ( z ) = k T ( z / k ) , b k = T k / k , a k ( z ) = z T k ′ ( z ) − T k ( z ) , c k = a k / k . (27) \begin{aligned}
T_{k}(z)&=k\mathrm{T}(z/k),& b_{k}&=T_{k}/k,\\
a_{k}(z)&=z\mathrm{T}'_{k}(z)-T_{k}(z),& c_{k}&=a_{k}/k. \tag*{(27)}
\end{aligned} T k ( z ) a k ( z ) = k T ( z / k ) , = z T k ′ ( z ) − T k ( z ) , b k c k = T k / k , = a k / k . (27)
Then 0 ≤ b k ≤ 2 0\leq b_{k}\leq2 0 ≤ b k ≤ 2 and − 2 ≤ c k ≤ 0 -2\leq c_{k}\leq0 − 2 ≤ c k ≤ 0 . We use k k k for density truncation and K K K for the Hessian bound of the convex potential; the corresponding optimizer is T = T K , α b g \mathrm{T}=T_{K,\alpha_{\mathrm{bg}}} T = T K , α bg .
Lemma 2.4. If r > 1 r>1 r > 1 , then L r ( T 3 ) ⊂ H − 4 ( T 3 ) L^{r}(\mathbb{T}^{3})\subset H^{-4}(\mathbb{T}^{3}) L r ( T 3 ) ⊂ H − 4 ( T 3 ) ; if p > 6 / 5 p>6/5 p > 6/5 , then L p ( T 3 ) ⊂ H − 1 ( T 3 ) L^{p}(\mathbb{T}^{3})\subset H^{-1}(\mathbb{T}^{3}) L p ( T 3 ) ⊂ H − 1 ( T 3 ) . A sequence with a time-uniform bound in the corresponding spatial norm and a common modulus of time continuity in the negative Sobolev space has a convergent subsequence in the corresponding space of continuous time-dependent functions. Moreover, for 6 / 5 < p < 2 6/5<p<2 6/5 < p < 2 and every ε > 0 \varepsilon>0 ε > 0 ,
∥ f ∥ H − 1 ≤ ε ∥ f ∥ p + C ε ∥ f ∥ H − 4 . \lVert f\rVert_{H^{-1}}\leq\varepsilon\lVert f\rVert_{p}+C_{\varepsilon}\lVert f\rVert_{H^{-4}}. ∥ f ∥ H − 1 ≤ ε ∥ f ∥ p + C ε ∥ f ∥ H − 4 .
Proof. By Rellich’s theorem [9 ] ,
H 4 ( T 3 ) ⊂ L r ′ ( T 3 ) ( r > 1 ) , H 1 ( T 3 ) ⊂ L p ′ ( T 3 ) ( p > 6 / 5 ) . H^{4}(\mathbb{T}^{3})\subset L^{r'}(\mathbb{T}^{3})\quad(r>1),\qquad H^{1}(\mathbb{T}^{3})\subset L^{p'}(\mathbb{T}^{3})\quad(p>6/5). H 4 ( T 3 ) ⊂ L r ′ ( T 3 ) ( r > 1 ) , H 1 ( T 3 ) ⊂ L p ′ ( T 3 ) ( p > 6/5 ) .
Duality gives the spatial embeddings. The common time modulus and Arzelà–Ascoli yield compactness in time; compact interpolation gives the final inequality [39 ] . □ \square □
Lemma 2.5. Suppose that r ≥ 0 r \ge0 r ≥ 0 , ∫ T 3 r d x = m > 0 \int_{\mathbb{T}^3} r\,dx = \mathfrak{m} > 0 ∫ T 3 r d x = m > 0 , r ∈ L 6 / 5 ( T 3 ) r \in L^{6/5}(\mathbb{T}^3) r ∈ L 6/5 ( T 3 ) , r v ∈ L 2 ( T 3 ) \sqrt{r}v \in L^2(\mathbb{T}^3) r v ∈ L 2 ( T 3 ) , and v ∈ H 1 ( T 3 ) v \in H^1(\mathbb{T}^3) v ∈ H 1 ( T 3 ) . Then
∥ v ∥ 6 ≤ m − 1 / 2 ∥ r v ∥ 2 + C ( 1 + ∥ r ∥ 6 / 5 m ) ∥ ∇ v ∥ 2 . \|v\|_6 \le\mathfrak{m}^{-1/2}\|\sqrt{r}v\|_2 + C\left(1+\frac{\|r\|_{6/5}}{\mathfrak{m}}\right)\|\nabla v\|_2. ∥ v ∥ 6 ≤ m − 1/2 ∥ r v ∥ 2 + C ( 1 + m ∥ r ∥ 6/5 ) ∥∇ v ∥ 2 . Proof. Set ⟨ v ⟩ = ∫ T 3 v d x \langle v\rangle= \int_{\mathbb{T}^3}v\,dx ⟨ v ⟩ = ∫ T 3 v d x . Applying Poincaré–Sobolev, weighted Cauchy–Schwarz, and Hölder’s inequality yields
∥ v − ⟨ v ⟩ ∥ 6 ≤ C ∥ ∇ v ∥ 2 , \|v-\langle v\rangle\|_6 \le C\|\nabla v\|_2, ∥ v − ⟨ v ⟩ ∥ 6 ≤ C ∥∇ v ∥ 2 , m ∣ ⟨ v ⟩ ∣ = ∣ ∫ T 3 r v d x − ∫ T 3 r ( v − ⟨ v ⟩ ) d x ∣ ≤ m 1 / 2 ∥ r v ∥ 2 + ∥ r ∥ 6 / 5 ∥ v − ⟨ v ⟩ ∥ 6 . (28) \mathfrak{m}|\langle v\rangle| = \left|\int_{\mathbb{T}^3}rv\,dx-\int_{\mathbb{T}^3}r(v-\langle v\rangle)\,dx\right|
\le\mathfrak{m}^{1/2}\|\sqrt{r}v\|_2+\|r\|_{6/5}\|v-\langle v\rangle\|_6. \tag*{(28)} m ∣ ⟨ v ⟩ ∣ = ∫ T 3 r v d x − ∫ T 3 r ( v − ⟨ v ⟩) d x ≤ m 1/2 ∥ r v ∥ 2 + ∥ r ∥ 6/5 ∥ v − ⟨ v ⟩ ∥ 6 . (28) Dividing (2.7) by m \mathfrak{m} m and using ∥ v ∥ 6 ≤ ∣ ⟨ v ⟩ ∣ + ∥ v − ⟨ v ⟩ ∥ 6 \|v\|_6\le|\langle v\rangle|+\|v-\langle v\rangle\|_6 ∥ v ∥ 6 ≤ ∣ ⟨ v ⟩ ∣ + ∥ v − ⟨ v ⟩ ∥ 6 , we then finish the proof of the present lemma. □
We will apply the DiPerna–Lions commutator (see [11 ] ) to bounded truncations with a weakly identified source. The following form keeps that source and the initial trace explicit; it does not identify the source with a nonlinear expression in the limiting density.
Lemma 2.6. Let v ≥ 0 v\ge0 v ≥ 0 be bounded on Q T Q_T Q T , let u ∈ L 2 ( 0 , T ; H 1 ( T 3 ) ) u\in L^2(0,T;H^1(\mathbb{T}^3)) u ∈ L 2 ( 0 , T ; H 1 ( T 3 )) and h ∈ L 2 ( Q T ) h\in L^2(Q_T) h ∈ L 2 ( Q T ) , and suppose that
∂ t v + div ( v u ) + h = 0 , \partial_t v+\operatorname{div}(vu)+h=0, ∂ t v + div ( v u ) + h = 0 , in distributions, with bounded initial data v 0 v_0 v 0 and with the initial term included. Then, for every b ∈ C 1 ( [ 0 , ∞ ) ) b\in C^1([0,\infty)) b ∈ C 1 ([ 0 , ∞ )) ,
∂ t b ( v ) + div ( b ( v ) u ) + ( v b ′ ( v ) − b ( v ) ) div u + b ′ ( v ) h = 0. (29) \partial_t b(v)+\operatorname{div}(b(v)u)+(vb'(v)-b(v))\operatorname{div}u+b'(v)h=0. \tag*{(29)} ∂ t b ( v ) + div ( b ( v ) u ) + ( v b ′ ( v ) − b ( v )) div u + b ′ ( v ) h = 0. (29) The initial value is b ( v 0 ) b(v_0) b ( v 0 ) . The case b ( 0 ) ≠ 0 b(0)\ne0 b ( 0 ) = 0 is allowed.
Proof. Extend v = v 0 v=v_0 v = v 0 and u = h = 0 u=h=0 u = h = 0 to negative times. The initial term makes the extended equation valid across t = 0 t=0 t = 0 . Spatial convolution gives
∂ t v ε + div ( v ε u ) + h ε = r ε , \partial_t v_\varepsilon+\operatorname{div}(v_\varepsilon u)+h_\varepsilon=r_\varepsilon, ∂ t v ε + div ( v ε u ) + h ε = r ε , where
r ε = v ε div u + ∫ ∇ ω ε ( y ) ⋅ [ u ( x ) − u ( x − y ) ] v ( x − y ) d y . r_\varepsilon=v_\varepsilon\operatorname{div}u+\int\nabla\omega_\varepsilon(y)\cdot[u(x)-u(x-y)]v(x-y)\,dy. r ε = v ε div u + ∫ ∇ ω ε ( y ) ⋅ [ u ( x ) − u ( x − y )] v ( x − y ) d y . On each finite time interval of the extension, the line-segment formula gives
∥ r ε ∥ L t , x 2 ≤ C ∥ v ∥ L t , x ∞ ∥ ∇ u ∥ L t , x 2 . \|r_\varepsilon\|_{L^2_{t,x}}\le C\|v\|_{L^\infty_{t,x}}\|\nabla u\|_{L^2_{t,x}}. ∥ r ε ∥ L t , x 2 ≤ C ∥ v ∥ L t , x ∞ ∥∇ u ∥ L t , x 2 . For smooth u u u , translation continuity of v v v in L 2 L^2 L 2 and ∫ y j ∂ i ω ε = − δ i j \int y_j\partial_i\omega_\varepsilon=-\delta_{ij} ∫ y j ∂ i ω ε = − δ ij cancel the local linear part against v ε div u v_\varepsilon\operatorname{div}u v ε div u . Then, in view of the DiPerna–Lions commutator argument in [11 ] , approximation of u u u in L t 2 H x 1 L^2_tH^1_x L t 2 H x 1 and the preceding bound give r ε → 0 r_\varepsilon\to0 r ε → 0 in L t , x 2 L^2_{t,x} L t , x 2 .
For fixed ε \varepsilon ε , the ordinary chain rule gives
∂ t b ( v ε ) + div ( b ( v ε ) u ) + ( v ε b ′ ( v ε ) − b ( v ε ) ) div u + b ′ ( v ε ) h ε = b ′ ( v ε ) r ε . \partial_t b(v_\varepsilon) + \operatorname{div}(b(v_\varepsilon)u) + \left(v_\varepsilon b'(v_\varepsilon) - b(v_\varepsilon)\right)\operatorname{div} u + b'(v_\varepsilon)h_\varepsilon= b'(v_\varepsilon)r_\varepsilon. ∂ t b ( v ε ) + div ( b ( v ε ) u ) + ( v ε b ′ ( v ε ) − b ( v ε ) ) div u + b ′ ( v ε ) h ε = b ′ ( v ε ) r ε .
On finite time intervals, the bounded compositions converge strongly in every finite L r L^r L r , and h ε → h h_\varepsilon\to h h ε → h and r ε → 0 r_\varepsilon\to0 r ε → 0 in L 2 L^2 L 2 . These limits yield (2.8). The extension equals b ( v 0 ) b(v_0) b ( v 0 ) on negative times, giving the initial term upon restriction to ( 0 , T ) (0,T) ( 0 , T ) .
Barycentric Transport and Pressure Estimates
This section proves Proposition 2.2. Lemmas 3.1, 3.2, and 3.3 respectively establish the spatial geometry, the material-pairing sign, and the complete pressure and tail estimates.
The classical pressure test uses
V = ∇ Δ T 3 − 1 ( ρ θ − ⟨ ρ θ ⟩ ) , div V = ρ θ − ⟨ ρ θ ⟩ . V = \nabla\Delta_{\mathbb{T}^3}^{-1}\left(\rho^\theta-\langle\rho^\theta\rangle\right), \qquad\operatorname{div} V = \rho^\theta-\langle\rho^\theta\rangle. V = ∇ Δ T 3 − 1 ( ρ θ − ⟨ ρ θ ⟩ ) , div V = ρ θ − ⟨ ρ θ ⟩ .
It produces the positive term ∫ ρ γ + θ \int\rho^{\gamma+\theta} ∫ ρ γ + θ , but its convective contribution requires control of ( ρ u ⊗ u ) : ∇ V (\rho u \otimes u) : \nabla V ( ρ u ⊗ u ) : ∇ V . The energy estimates give ρ u ⊗ u ∈ L t 1 L x 3 γ γ + 3 \rho u \otimes u \in L^1_t L_x^{\frac{3\gamma}{\gamma+3}} ρ u ⊗ u ∈ L t 1 L x γ + 3 3 γ when the displayed spatial exponent exceeds one. This requires γ > 3 / 2 \gamma> 3/2 γ > 3/2 ; the available time and space exponents then restrict θ \theta θ as in (1.4). This calculation concerns the classical test under the energy budget; it is not an obstruction to every possible density-dependent test.
We instead seek T = x + V = ∇ φ T = x + V = \nabla\varphi T = x + V = ∇ φ with H = D T H = DT H = D T satisfying 0 ⪯ H ⪯ K Id 0 \preceq H \preceq K\operatorname{Id} 0 ⪯ H ⪯ K Id and ∫ T 3 H = Id \int_{\mathbb{T}^3} H = \operatorname{Id} ∫ T 3 H = Id . The upper bound makes the test spatially admissible, while the mean identity improves its viscous cost to order K 1 / 2 K^{1/2} K 1/2 . These two conditions alone do not detect high density: the additional trace lower bound in Lemma 3.1 supplies that information. The time-dependent construction must also respect the material sign in Lemma 3.2.
Periodic Barycentric Map
Periodic Barycentric Map. Let ρ ∈ L 1 ( T 3 ) \rho\in L^1(\mathbb{T}^3) ρ ∈ L 1 ( T 3 ) be nonnegative and fix a background density α b g > 0 \alpha_{\mathrm{bg}} > 0 α bg > 0 . Set
m = ∫ T 3 ρ , c = m + α b g , σ = ρ + α b g c , ∫ T 3 σ = 1. (30) \mathbf{m} = \int_{\mathbb{T}^3} \rho, \qquad c = \mathbf{m} + \alpha_{\mathrm{bg}}, \qquad\sigma= \frac{\rho+\alpha_{\mathrm{bg}}}{c}, \qquad\int_{\mathbb{T}^3} \sigma= 1. \tag*{(30)} m = ∫ T 3 ρ , c = m + α bg , σ = c ρ + α bg , ∫ T 3 σ = 1. (30)
Since ∣ T 3 ∣ = 1 |\mathbb{T}^3| = 1 ∣ T 3 ∣ = 1 , the source μ = ( ρ + α b g ) d x \mu= (\rho+\alpha_{\mathrm{bg}})\,dx μ = ( ρ + α bg ) d x and uniform target β = c d x \beta= c\,dx β = c d x both have mass c c c . The background is part of the transport construction and allows the physical density ρ \rho ρ to vanish. As above, K ≥ 6 K \ge6 K ≥ 6 bounds the derivative, while k k k indexes the density truncation T k T_k T k .
For the quadratic Monge problem, a deterministic transport map sends the source directly to the target (see [5 ] ). A Jacobian constraint then controls the product of its eigenvalues, but gives no upper bound on individual eigenvalues. In the viscous momentum test, one needs an L 2 L^2 L 2 bound for the whole Hessian. Barycentric transport separates the prescribed target law from the field used in that test, see [1 , 16 , 17 ] . Write q : R 3 → T 3 q:\mathbb{R}^{3}\to\mathbb{T}^{3} q : R 3 → T 3 for the quotient map. Disintegrate a plan as π ( d x , d z ) = μ ( d x ) π x ( d z ) \pi(dx,dz)=\mu(dx)\pi_x(dz) π ( d x , d z ) = μ ( d x ) π x ( d z ) , using integer lifts of z z z . Its two relevant quantities are
∫ π x ( q − 1 ( B ) ) μ ( d x ) = c ∣ B ∣ , T ( x ) = ∫ z π x ( d z ) . \int\pi_x\left(q^{-1}(B)\right)\,\mu(dx)=c|B|,\qquad T(x)=\int z\,\pi_x(dz). ∫ π x ( q − 1 ( B ) ) μ ( d x ) = c ∣ B ∣ , T ( x ) = ∫ z π x ( d z ) .
The first identity prescribes the full target marginal. The second is a single vector at each source point and can serve as a momentum test. For instance, splitting between z 1 , z 2 z_1,z_2 z 1 , z 2 with weights θ , 1 − θ \theta,1-\theta θ , 1 − θ gives T = θ z 1 + ( 1 − θ ) z 2 T=\theta z_1+(1-\theta)z_2 T = θ z 1 + ( 1 − θ ) z 2 and
∫ ∣ z − x ∣ 2 π x ( d z ) = ∣ T − x ∣ 2 + ∫ ∣ z − T ∣ 2 π x ( d z ) . \int|z-x|^2\,\pi_x(dz)=|T-x|^2+\int|z-T|^2\,\pi_x(dz). ∫ ∣ z − x ∣ 2 π x ( d z ) = ∣ T − x ∣ 2 + ∫ ∣ z − T ∣ 2 π x ( d z ) .
Thus the mixed cost in (B.11) penalizes both displacement and conditional variance while permitting splitting. It does not impose T # μ = c d x T_{\#}\mu=c\,dx T # μ = c d x . The exact marginal will instead be used after restricting the residual plan to its deterministic part.
The uniform background has a second role beyond normalization. The strict positivity of ρ + α b g \rho+\alpha_{\mathrm{bg}} ρ + α bg transfers identities from source-almost-everywhere to Lebesgue-almost-everywhere statements. In Lemma 3.2, its stationary labels will also control ∂ t V \partial_tV ∂ t V in unweighted L 2 L^2 L 2 . None of these uses imposes a positive lower bound on the physical density.
Lemma 3.1. Let ρ ∈ L 1 ( T 3 ) \rho\in L^1(\mathbb{T}^3) ρ ∈ L 1 ( T 3 ) be nonnegative, α b g > 0 \alpha_{\mathrm{bg}}>0 α bg > 0 , and K ≥ 6 K\ge6 K ≥ 6 . Optimization over transport plans with the periodic uniform measure as target marginal determines a unique barycentric map
T = T K , α b g [ ρ ] = x + V = ∇ φ , φ − ∣ x ∣ 2 / 2 is periodic . T=T_{K,\alpha_{\mathrm{bg}}}[\rho]=x+V=\nabla\varphi,\qquad\varphi-|x|^2/2\text{ is periodic}. T = T K , α bg [ ρ ] = x + V = ∇ φ , φ − ∣ x ∣ 2 /2 is periodic . Here φ \varphi φ is convex and C 1 , 1 C^{1,1} C 1 , 1 on R 3 \mathbb{R}^3 R 3 , and almost everywhere
0 ≤ H : = D T ≤ K Id , ∫ T 3 V = 0 , (31) 0\le H:=DT\le K\operatorname{Id},\qquad\int_{\mathbb{T}^3}V=0, \tag*{(31)} 0 ≤ H := D T ≤ K Id , ∫ T 3 V = 0 , (31) ∥ V ∥ ∞ ≤ D , D = diam ( T 3 ) = 3 / 2 , \|V\|_{\infty}\le D,\qquad D=\operatorname{diam}(\mathbb{T}^3)=\sqrt{3}/2, ∥ V ∥ ∞ ≤ D , D = diam ( T 3 ) = 3 /2 , tr H ≥ min { K , 3 ρ 1 / 3 } . (32) \operatorname{tr}H\ge\min\{K,3\rho^{1/3}\}. \tag*{(32)} tr H ≥ min { K , 3 ρ 1/3 } . (32) Moreover, the periodic Hessian satisfies
∫ T 3 ∣ H − Id ∣ 2 d x ≤ 3 ( K − 1 ) . (33) \int_{\mathbb{T}^3}|H-\operatorname{Id}|^2\,dx\le3(K-1). \tag*{(33)} ∫ T 3 ∣ H − Id ∣ 2 d x ≤ 3 ( K − 1 ) . (33) These estimates allow ρ \rho ρ to vanish on any set.
Proof. We use Lemma B.1 and the marginal representation in Lemma B.2, proved in Appendix § 6 , to construct the potential and obtain a determinant bound where the Hessian cap is inactive. The convex-analytic ingredients are strong convexity, subgradient monotonicity, and conjugacy (cf. [4 ] ). Exact periodic lifts and the restricted determinant estimate below are additional parts of the present construction.
Step 1: Construction of the convex potential. Let μ = ( ρ + α b g ) d x \mu=(\rho+\alpha_{\mathrm{bg}})\,dx μ = ( ρ + α bg ) d x and represent its positions by X ∈ H = L 2 ( A , ℓ ; R 3 ) X\in\mathcal{H}=L^2(\mathcal{A},\ell;\mathbb{R}^3) X ∈ H = L 2 ( A , ℓ ; R 3 ) on an atomless label space of mass c c c . The convex envelope G G G in Appendix § 6 .§ 1 turns the plan problem into minimization of
G ( Y ) + K − 1 2 ∥ Y ∥ H 2 − K ⟨ X , Y ⟩ H + K 2 ∥ X ∥ H 2 . G(Y) + \frac{K-1}{2}\lVert Y\rVert_{\mathcal{H}}^2 - K\langle X,Y\rangle_{\mathcal{H}} + \frac{K}{2}\lVert X\rVert_{\mathcal{H}}^2. G ( Y ) + 2 K − 1 ∥ Y ∥ H 2 − K ⟨ X , Y ⟩ H + 2 K ∥ X ∥ H 2 .
The quadratic part is strongly convex with modulus K − 1 K-1 K − 1 ; hence its unique minimizer is characterized by a subgradient inclusion. The exact-marginal lemma represents that subgradient by an optimal plan, rather than merely by a limit of approximate target laws. For clarity, the monotonicity used below is a statement on the whole label space. If two inputs give differences Δ X \Delta X Δ X , Δ Y \Delta Y Δ Y , and Δ T = K Δ X − ( K − 1 ) Δ Y \Delta\mathcal{T}=K\Delta X-(K-1)\Delta Y Δ T = K Δ X − ( K − 1 ) Δ Y , then
K ⟨ Δ T , Δ X ⟩ = ∥ Δ T ∥ 2 + ( K − 1 ) ⟨ Δ T , Δ Y ⟩ ≥ ∥ Δ T ∥ 2 . K\langle\Delta\mathcal{T},\Delta X\rangle=\lVert\Delta\mathcal{T}\rVert^2+(K-1)\langle\Delta\mathcal{T},\Delta Y\rangle\geq\lVert\Delta\mathcal{T}\rVert^2. K ⟨ Δ T , Δ X ⟩ = ∥ Δ T ∥ 2 + ( K − 1 ) ⟨ Δ T , Δ Y ⟩ ≥ ∥ Δ T ∥ 2 .
Here the last term is nonnegative by monotonicity of ∂ G \partial G ∂ G . This is both the K K K -Lipschitz estimate and the sign that will survive when physical labels move and background labels remain stationary.
The exact-marginal representation requires a separate argument from this convex-analytic inequality. Appendix § 6 .§ 2 uses the kernel topology tested by L 1 ( A ; C ( Q ) ) L^1(A;C(Q)) L 1 ( A ; C ( Q )) to retain the target marginal for measurable residual positions. The nonnegative nearest-lift defect makes the lift constraint closed in that topology. Directional derivatives then identify the supporting functional in every bounded direction; separation gives ∂ G ( y ) = B y \partial G(y)=B_y ∂ G ( y ) = B y . Thus the plan used in the Jacobian argument is an actual plan with the exact uniform target marginal.
The Hilbert-space construction gives the residual position y = Y K ( X ) y=Y_K(X) y = Y K ( X ) and a unique periodic barycentric map T = x + V \mathcal{T}=x+V T = x + V . By (B.4) and the exact-marginal representation, there is an optimal plan π \pi π for which
y ( x ) = K x − T ( x ) K − 1 , T ( x ) = E π [ z ∣ x ] , (34) y(x)=\frac{Kx-\mathcal{T}(x)}{K-1},\qquad\mathcal{T}(x)=\mathbb{E}_{\pi}[z\mid x], \tag*{(34)} y ( x ) = K − 1 K x − T ( x ) , T ( x ) = E π [ z ∣ x ] , (34)
( q y , q z ) # π (qy,qz)_{\#}\pi ( q y , q z ) # π is optimal with target marginal c d x c\,dx c d x .
Here z z z is a nearest integer lift to y ( x ) y(x) y ( x ) , so ∣ z − y ( x ) ∣ ≤ D \lvert z-y(x)\rvert\leq D ∣ z − y ( x )∣ ≤ D . The target marginal is exact even for bounded measurable y y y .
Choose conjugate Kantorovich potentials a , b a,b a , b for d T 3 2 / 2 d_{\mathbb{T}^3}^2/2 d T 3 2 /2 . For nearest lifts the dual inequality reads
a ( q y ) + b ( q z ) ≤ 1 2 ∣ y − z ∣ 2 , a(qy)+b(qz)\leq\frac{1}{2}\lvert y-z\rvert^2, a ( q y ) + b ( q z ) ≤ 2 1 ∣ y − z ∣ 2 ,
with equality on the support of π \pi π . The c c c -transform therefore gives the convex function
h ( z ) = 1 2 ∣ z ∣ 2 − b ( q z ) = sup y ∈ R 3 { y ⋅ z − 1 2 ∣ y ∣ 2 + a ( q y ) } . h(z)=\frac{1}{2}\lvert z\rvert^2-b(qz)=\sup_{y\in\mathbb{R}^3}\left\{y\cdot z-\frac{1}{2}\lvert y\rvert^2+a(qy)\right\}. h ( z ) = 2 1 ∣ z ∣ 2 − b ( q z ) = y ∈ R 3 sup { y ⋅ z − 2 1 ∣ y ∣ 2 + a ( q y ) } .
For fixed y ( x ) y(x) y ( x ) , all conditional target points belong to the convex contact set of h ( z ) − y ( x ) ⋅ z h(z)-y(x)\cdot z h ( z ) − y ( x ) ⋅ z . Their barycenter T ( x ) \mathcal{T}(x) T ( x ) belongs to the same set, and hence y ( x ) ∈ ∂ h ( T ( x ) ) y(x)\in\partial h(\mathcal{T}(x)) y ( x ) ∈ ∂ h ( T ( x )) .
Define
h K ( z ) = ∣ z ∣ 2 2 K + K − 1 K h ( z ) , φ = h K ∗ . h_K(z)=\frac{\lvert z\rvert^2}{2K}+\frac{K-1}{K}h(z),\qquad\varphi=h_K^*. h K ( z ) = 2 K ∣ z ∣ 2 + K K − 1 h ( z ) , φ = h K ∗ .
Using (3.5), we have
x = T ( x ) K + K − 1 K y ( x ) ∈ ∂ h K ( T ( x ) ) . x = \frac{T(x)}{K} + \frac{K-1}{K}y(x) \in\partial h_K(T(x)). x = K T ( x ) + K K − 1 y ( x ) ∈ ∂ h K ( T ( x )) .
Since h K h_K h K is 1 / K 1/K 1/ K -strongly convex, φ \varphi φ is convex and C 1 , 1 C^{1,1} C 1 , 1 , with
T = ∇ φ , H = D 2 φ , 0 ≤ H ≤ K Id . (35) T = \nabla\varphi,\qquad H = D^2\varphi,\qquad0 \le H \le K\operatorname{Id}. \tag*{(35)} T = ∇ φ , H = D 2 φ , 0 ≤ H ≤ K Id . (35)
The periodicity of b b b implies
h K ( z + n ) = h K ( z ) + n ⋅ z + 1 2 ∣ n ∣ 2 , n ∈ Z 3 . h_K(z+n)=h_K(z)+n\cdot z+\frac{1}{2}|n|^2,\qquad n\in\mathbb{Z}^3. h K ( z + n ) = h K ( z ) + n ⋅ z + 2 1 ∣ n ∣ 2 , n ∈ Z 3 .
Taking Legendre transforms shows that ψ = φ − ∣ x ∣ 2 / 2 \psi=\varphi-|x|^2/2 ψ = φ − ∣ x ∣ 2 /2 is periodic. In particular, for each i , j i,j i , j ,
∫ T 3 ∂ i ψ d x = 0 , ∫ T 3 ∂ j ∂ i ψ d x = 0. \int_{\mathbb{T}^3}\partial_i\psi\,dx=0,\qquad\int_{\mathbb{T}^3}\partial_j\partial_i\psi\,dx=0. ∫ T 3 ∂ i ψ d x = 0 , ∫ T 3 ∂ j ∂ i ψ d x = 0.
These identities hold distributionally and hence for the Lipschitz gradient of ψ \psi ψ . They give the mean Hessian without any assumption on the spatial mean of the fluid velocity. Consequently,
V = ∇ ψ , ∇ V = H − Id , ∫ T 3 V d x = 0 , V=\nabla\psi,\qquad\nabla V=H-\operatorname{Id},\qquad\int_{\mathbb{T}^3}V\,dx=0, V = ∇ ψ , ∇ V = H − Id , ∫ T 3 V d x = 0 ,
∫ T 3 H d x = Id , ∫ T 3 tr H d x = 3. (36) \int_{\mathbb{T}^3}H\,dx=\operatorname{Id},\qquad\int_{\mathbb{T}^3}\operatorname{tr}H\,dx=3. \tag*{(36)} ∫ T 3 H d x = Id , ∫ T 3 tr H d x = 3. (36)
Trace lower bound
The argument uses absolute continuity twice, for different measures. First the residual map transfers the source restricted to the unsaturated set into an absolutely continuous measure; this permits differentiation of the transport potential. Only after the restricted plan has become deterministic do we use its target domination to estimate the Jacobian of T T T . Reversing these steps would assume the determinism that the proof must establish. Let
U = { x : H ( x ) exists and λ max ( H ( x ) ) < K } . U=\{x:H(x)\text{ exists and }\lambda_{\max}(H(x))<K\}. U = { x : H ( x ) exists and λ m a x ( H ( x )) < K } .
On U U U the residual map is Lipschitz and satisfies
D y = K Id − H K − 1 > 0 a.e. on U . (37) D_y=\frac{K\operatorname{Id}-H}{K-1}>0\qquad\text{a.e. on }U. \tag*{(37)} D y = K − 1 K Id − H > 0 a.e. on U . (37)
The set U U U is a measurable subset of differentiability points; it need not be open. The area formula for Lipschitz maps [12 Chapter 3] applies to such restrictions. If N ⊂ T 3 N\subset\mathbb{T}^3 N ⊂ T 3 is null and E = U ∩ ( q y ) − 1 ( N ) E=U\cap(qy)^{-1}(N) E = U ∩ ( q y ) − 1 ( N ) , it gives
∫ E det D y ( x ) d x = ∫ N # { x ∈ E : q y ( x ) = z } d z = 0. \int_E\det D_y(x)\,dx=\int_N\#\{x\in E: qy(x)=z\}\,dz=0. ∫ E det D y ( x ) d x = ∫ N # { x ∈ E : q y ( x ) = z } d z = 0.
Here periodic charts, or equivalently integer lifts, give the same counting formula. Since det D y > 0 \det D_y>0 det D y > 0 on U U U , ∣ E ∣ = 0 |E|=0 ∣ E ∣ = 0 . Thus μ ( E ) = 0 \mu(E)=0 μ ( E ) = 0 and ( q y ) # ( 1 U μ ) ≪ d x (qy)_{\#}(\mathbf{1}_U\mu)\ll dx ( q y ) # ( 1 U μ ) ≪ d x . This is the absolute continuity needed at the residual position, before any assertion about the Jacobian of T T T .
Restrict the optimal plan to source points in U U U . Its dual contact relation is unchanged. At almost every residual position, the Kantorovich potential a a a is differentiable; at a contact pair,
a ( q ( y + w ) ) ≤ 1 2 ∣ y + w − z ∣ 2 − b ( q z ) a(q(y+w))\le\frac{1}{2}|y+w-z|^2-b(qz) a ( q ( y + w )) ≤ 2 1 ∣ y + w − z ∣ 2 − b ( q z )
with equality at w = 0 w=0 w = 0 . Differentiation gives
z = y − ∇ a ( q y ) . z=y-\nabla a(qy). z = y − ∇ a ( q y ) .
The differentiability exceptional set of a a a has zero Lebesgue measure. The preceding residual absolute continuity makes its inverse image negligible for the restricted source. At every remaining contact point the displayed identity determines the Euclidean lift z z z , even if several nearest lifts were initially allowed. Thus the target lift is uniquely determined by the residual position. On U U U the conditional plan is concentrated at its barycenter T ( x ) T(x) T ( x ) . The exact target marginal of π \pi π therefore yields
( q T ) # [ ( ρ + α b g ) 1 U d x ] ≤ c d x . (38) (qT)_{\#}\left[(\rho+\alpha_{\mathrm{bg}})1_U\,dx\right]\le c\,dx. \tag*{(38)} ( q T ) # [ ( ρ + α bg ) 1 U d x ] ≤ c d x . (38)
Apply the area formula to the Lipschitz map T T T [12 Chapter 3] . Since its pushforward in (3.9) is absolutely continuous, det H > 0 \det H>0 det H > 0 at almost every source point of U U U . Indeed, for E = U ∩ { det H = 0 } E=U\cap\{\det H=0\} E = U ∩ { det H = 0 } the area formula gives ∣ q T ( E ) ∣ = 0 |qT(E)|=0 ∣ q T ( E ) ∣ = 0 . Domination in (3.9) then gives μ ( E ) ≤ μ ( U ∩ ( q T ) − 1 ( q T ( E ) ) ) = 0 \mu(E)\le\mu(U\cap(qT)^{-1}(qT(E)))=0 μ ( E ) ≤ μ ( U ∩ ( q T ) − 1 ( q T ( E ))) = 0 , and the positive background yields ∣ E ∣ = 0 |E|=0 ∣ E ∣ = 0 . Null images may be replaced by measurable null supersets in this argument. The density formula then implies
∑ x ∈ U q T ( x ) = z ρ ( x ) + α b g det H ( x ) ≤ c for a.e. z ∈ T 3 . \sum_{\substack{x\in U\\qT(x)=z}}\frac{\rho(x)+\alpha_{\mathrm{bg}}}{\det H(x)}\le c
\qquad\text{for a.e. }z\in\mathbb{T}^3. x ∈ U q T ( x ) = z ∑ det H ( x ) ρ ( x ) + α bg ≤ c for a.e. z ∈ T 3 .
The domination also makes the preimage of the exceptional target-null set a set of zero μ \mu μ -measure. Since ρ + α b g > 0 \rho+\alpha_{\mathrm{bg}}>0 ρ + α bg > 0 , each nonnegative summand is at most c c c for almost every x ∈ U x\in U x ∈ U . Hence
det H ( x ) ≥ ρ ( x ) + α b g c = σ ( x ) for a.e. x ∈ U . (39) \det H(x)\ge\frac{\rho(x)+\alpha_{\mathrm{bg}}}{c}=\sigma(x)
\qquad\text{for a.e. }x\in U. \tag*{(39)} det H ( x ) ≥ c ρ ( x ) + α bg = σ ( x ) for a.e. x ∈ U . (39)
The periodic area formula counts all preimages, using finitely many integer lifts because T − x T-x T − x is bounded. Global injectivity is unnecessary: every preimage contributes a nonnegative summand, so the bound for the sum bounds each source contribution. Nor does the argument give equality of Jacobians: the restricted target is only dominated by c d x c\,dx c d x , and the barycentric map outside U U U need not push the source onto that target.
On U U U , the arithmetic–geometric mean inequality gives tr H ≥ 3 σ 1 / 3 \operatorname{tr}H\ge3\sigma^{1/3} tr H ≥ 3 σ 1/3 . Outside U U U , an eigenvalue of H H H equals K K K almost everywhere, so tr H ≥ K \operatorname{tr}H\ge K tr H ≥ K . Therefore
tr H ≥ min { K , 3 σ 1 / 3 } a.e. on T 3 . (40) \operatorname{tr}H\ge\min\{K,3\sigma^{1/3}\}
\qquad\text{a.e. on }\mathbb{T}^3. \tag*{(40)} tr H ≥ min { K , 3 σ 1/3 } a.e. on T 3 . (40)
Step 3: The Hessian estimate. From (3.6), H 2 ⪯ K H H^2 \preceq KH H 2 ⪯ K H . Together with (3.7), this gives
∫ T 3 ∣ H − Id ∣ 2 d x = ∫ T 3 [ tr ( H 2 ) − 2 tr H + 3 ] d x = ∫ T 3 tr ( H 2 ) d x − 3 ≤ K ∫ T 3 tr H d x − 3 = 3 ( K − 1 ) . \begin{aligned}
\int_{\mathbb{T}^3} |H-\operatorname{Id}|^2\,dx
&=\int_{\mathbb{T}^3}\left[\operatorname{tr}(H^2)-2\operatorname{tr}H+3\right]\,dx\\
&=\int_{\mathbb{T}^3}\operatorname{tr}(H^2)\,dx-3\\
&\le K\int_{\mathbb{T}^3}\operatorname{tr}H\,dx-3=3(K-1).
\end{aligned} ∫ T 3 ∣ H − Id ∣ 2 d x = ∫ T 3 [ tr ( H 2 ) − 2 tr H + 3 ] d x = ∫ T 3 tr ( H 2 ) d x − 3 ≤ K ∫ T 3 tr H d x − 3 = 3 ( K − 1 ) .
This is (3.4). The matrix inequality follows eigenvalue by eigenvalue from 0 ≤ λ i ≤ K 0\le\lambda_i\le K 0 ≤ λ i ≤ K and λ i 2 ≤ K λ i \lambda_i^2\le K\lambda_i λ i 2 ≤ K λ i . A pointwise bound alone would give an L 2 L^2 L 2 cost of order K K K ; the fixed mean trace reduces that cost to order K 1 / 2 K^{1/2} K 1/2 . Finally, (B.5) gives ∥ V ∥ ∞ ≤ D \lVert V\rVert_\infty\le D ∥ V ∥ ∞ ≤ D .
The resulting estimate is independent of any density derivative. For every stress S ∈ L 2 ( Q T ) S\in L^2(Q_T) S ∈ L 2 ( Q T ) it gives the explicit payment
∣ ∫ Q T S : ∇ V d x d t ∣ ≤ 3 T ( K − 1 ) ∥ S ∥ L 2 ( Q T ) . \left|\int_{Q_T} S:\nabla V\,dx\,dt\right|\le\sqrt{3T(K-1)}\,\lVert S\rVert_{L^2(Q_T)}. ∫ Q T S : ∇ V d x d t ≤ 3 T ( K − 1 ) ∥ S ∥ L 2 ( Q T ) .
Indeed, the spatial estimate holds at almost every time and ∇ V = H − Id \nabla V=H-\operatorname{Id} ∇ V = H − Id . The factor T 1 / 2 T^{1/2} T 1/2 comes only from integrating that estimate in time. Time admissibility of V V V is the next issue.
Material Pairing and Admissibility
To make the field in Lemma 3.1 admissible in time, we represent physical mass by moving labels and the background by stationary labels. The time term m ⋅ ∂ t V m\cdot\partial_tV m ⋅ ∂ t V by itself has no sign: V V V changes when the entire density distribution moves. Monotonicity concerns the derivative along those moving positions, which is ∂ t V + H u \partial_tV+Hu ∂ t V + H u . It therefore controls the sum with the convective term contracted against H H H . The momentum equation instead contains H − Id H-\operatorname{Id} H − Id ; the resulting difference is the actual kinetic energy.
Lemma 3.2. Fix one approximate solution ( ρ , u , P , S ) (\rho,u,P,S) ( ρ , u , P , S ) from Proposition 2.2, set m = ρ u m=\rho u m = ρ u , and choose K ≥ 6 K\ge6 K ≥ 6 and a time-independent α b g > 0 \alpha_{\mathrm{bg}}>0 α bg > 0 . Let V = V K , α b g [ ρ ] V=V_{K,\alpha_{\mathrm{bg}}}[\rho] V = V K , α bg [ ρ ] and H = Id + ∇ V H=\operatorname{Id}+\nabla V H = Id + ∇ V . Then V ∈ H l o c 1 ( ( 0 , T ) ; L 2 ( T 3 ) ) V\in H^1_{\mathrm{loc}}((0,T);L^2(\mathbb{T}^3)) V ∈ H loc 1 (( 0 , T ) ; L 2 ( T 3 )) , and for every 0 ≤ η ∈ C c ∞ ( 0 , T ) 0\le\eta\in C^\infty_c(0,T) 0 ≤ η ∈ C c ∞ ( 0 , T ) ,
M K , α b g ( η ) : = ∫ Q T η [ m ⋅ ∂ t V + ( ρ u ⊗ u ) : H ] ≥ 0. (41) \mathcal{M}_{K,\alpha_{\mathrm{bg}}}(\eta):=\int_{Q_T}\eta[m\cdot\partial_tV+(\rho u\otimes u):H]\ge0. \tag*{(41)} M K , α bg ( η ) := ∫ Q T η [ m ⋅ ∂ t V + ( ρ u ⊗ u ) : H ] ≥ 0. (41)
Proof. If the physical mass vanishes, the uniform background gives V = 0 V=0 V = 0 and H = Id H=\operatorname{Id} H = Id , so the assertion is immediate. We henceforth assume positive mass.
Step 1: Space–time regularization. Choose intervals I ⋐ I ′ ⋐ ( 0 , T ) I\Subset I'\Subset(0,T) I ⋐ I ′ ⋐ ( 0 , T ) with supp η ⊂ I \operatorname{supp}\eta\subset I supp η ⊂ I . Let χ ∈ C c ∞ ( ( − 1 , 1 ) ) \chi\in C^\infty_c((-1,1)) χ ∈ C c ∞ (( − 1 , 1 )) be nonnegative, with ∫ R χ = 1 \int_{\mathbb{R}}\chi=1 ∫ R χ = 1 , and set
χ h ( s ) = h − 1 χ ( s / h ) . \chi_h(s)=h^{-1}\chi(s/h). χ h ( s ) = h − 1 χ ( s / h ) .
For the spatial regularization, use the periodic heat kernel
Θ h ( x ) = 1 ( 4 π h 2 ) 3 / 2 ∑ k ∈ Z 3 exp ( − ∣ x + k ∣ 2 4 h 2 ) . \Theta_h(x)=\frac{1}{(4\pi h^2)^{3/2}}\sum_{k\in\mathbb{Z}^3}\exp\left(-\frac{|x+k|^2}{4h^2}\right). Θ h ( x ) = ( 4 π h 2 ) 3/2 1 k ∈ Z 3 ∑ exp ( − 4 h 2 ∣ x + k ∣ 2 ) .
Then Θ h > 0 \Theta_h>0 Θ h > 0 and ∫ T 3 Θ h d x = 1 \int_{\mathbb{T}^3}\Theta_h\,dx=1 ∫ T 3 Θ h d x = 1 . For F ∈ { ρ , m , ρ u ⊗ u , P , S } F\in\{\rho,m,\rho u\otimes u,P,S\} F ∈ { ρ , m , ρ u ⊗ u , P , S } , define
F h ( t , x ) = ∫ R χ h ( s ) ∫ T 3 Θ h ( x − y ) F ( t − s , y ) d y d s . F_h(t,x)=\int_{\mathbb{R}}\chi_h(s)\int_{\mathbb{T}^3}\Theta_h(x-y)F(t-s,y)\,dy\,ds. F h ( t , x ) = ∫ R χ h ( s ) ∫ T 3 Θ h ( x − y ) F ( t − s , y ) d y d s .
where h > 0 h>0 h > 0 is sufficiently small that t − s ∈ I ′ t-s\in I' t − s ∈ I ′ for t ∈ I t\in I t ∈ I and s ∈ supp χ h s\in\operatorname{supp}\chi_h s ∈ supp χ h .
Since convolution commutes with differentiation, the smoothed fields satisfy
∂ t ρ h + div m h = 0 , ∂ t m h + div ( ρ u ⊗ u ) h + ∇ P h = div S h . \begin{aligned}
\partial_t\rho_h+\operatorname{div}m_h&=0,\\
\partial_tm_h+\operatorname{div}(\rho u\otimes u)_h+\nabla P_h&=\operatorname{div}S_h.
\end{aligned} ∂ t ρ h + div m h ∂ t m h + div ( ρ u ⊗ u ) h + ∇ P h = 0 , = div S h .
The positivity of the heat kernel and conservation of positive total mass imply ρ h > 0 \rho_h>0 ρ h > 0 . Thus v h = m h / ρ h v_h=m_h/\rho_h v h = m h / ρ h is well defined and smooth. The nonnegative convolution kernel also preserves the positivity of the inertia matrix:
( ρ h m h T m h ( ρ u ⊗ u ) h ) ⪰ 0. \begin{pmatrix}
\rho_h & m_h^{\mathsf T}\\
m_h & (\rho u\otimes u)_h
\end{pmatrix}\succeq0. ( ρ h m h m h T ( ρ u ⊗ u ) h ) ⪰ 0.
Taking its Schur complement gives
R h : = ( ρ u ⊗ u ) h − m h ⊗ m h ρ h ⪰ 0. R_h:=(\rho u\otimes u)_h-\frac{m_h\otimes m_h}{\rho_h}\succeq0. R h := ( ρ u ⊗ u ) h − ρ h m h ⊗ m h ⪰ 0.
For any a ∈ R a\in\mathbb{R} a ∈ R and b ∈ R 3 b\in\mathbb{R}^3 b ∈ R 3 , the original matrix has quadratic form ρ ( a + b ⋅ u ) 2 \rho(a+b\cdot u)^2 ρ ( a + b ⋅ u ) 2 . Using one and the same positive kernel for its entries preserves this inequality. More explicitly, at each smoothing point ( t , x ) (t,x) ( t , x ) ,
R h ( t , x ) = ∫ χ h ( s ) Θ h ( x − y ) ρ ( t − s , y ) [ u ( t − s , y ) − v h ( t , x ) ] ⊗ [ u ( t − s , y ) − v h ( t , x ) ] d y d s . R_h(t,x)=\int\chi_h(s)\Theta_h(x-y)\rho(t-s,y)
\left[u(t-s,y)-v_h(t,x)\right]\otimes\left[u(t-s,y)-v_h(t,x)\right]\,dy\,ds. R h ( t , x ) = ∫ χ h ( s ) Θ h ( x − y ) ρ ( t − s , y ) [ u ( t − s , y ) − v h ( t , x ) ] ⊗ [ u ( t − s , y ) − v h ( t , x ) ] d y d s .
The cross terms cancel because the weighted mean velocity is v h v_h v h . Different smoothing kernels for mass, momentum, and convection would not give this covariance identity. The tensor R h R_h R h is only a smoothing remainder; it adds no assumption to the original equations.
Step 2: The signed material pairing. Represent ρ h d x \rho_h\,dx ρ h d x by labels moving along the smooth velocity v h v_h v h , and α b g d x \alpha_{\mathrm{bg}}\,dx α bg d x by stationary labels. Both groups enter the same Hilbert-space optimization (B.1). The firm monotonicity of Lemma B.1, applied along these paths as in (B.3), gives
⟨ J ˙ K , X ˙ ⟩ H ≥ 1 K ∥ J ˙ K ∥ H 2 . \langle\dot{\mathcal{J}}_K,\dot{X}\rangle_{\mathcal{H}}\geq\frac{1}{K}\|\dot{\mathcal{J}}_K\|_{\mathcal{H}}^2. ⟨ J ˙ K , X ˙ ⟩ H ≥ K 1 ∥ J ˙ K ∥ H 2 .
In fact the moving labels have finite squared speed
∥ X ˙ ( t ) ∥ H 2 = ∫ T 3 ρ h ∣ v h ∣ 2 d x ≤ ∫ T 3 ( ρ ∣ u ∣ 2 ) h d x , \|\dot{X}(t)\|_{\mathcal{H}}^2=\int_{\mathbb{T}^3}\rho_h|v_h|^2\,dx\leq\int_{\mathbb{T}^3}(\rho|u|^2)_h\,dx, ∥ X ˙ ( t ) ∥ H 2 = ∫ T 3 ρ h ∣ v h ∣ 2 d x ≤ ∫ T 3 ( ρ ∣ u ∣ 2 ) h d x ,
whereas the background labels have zero speed. Local trajectory lifts may be chosen continuously on R 3 \mathbb{R}^3 R 3 ; integer-lift invariance makes the resulting displacement independent of this choice. Restricting the Hilbert-space H 1 H^1 H 1 map J K ( X ( t ) ) \mathcal{J}_K(X(t)) J K ( X ( t )) to the stationary background gives V h ∈ H l o c 1 ( ( 0 , T ) ; L 2 ( T 3 ) ) V_h\in H^1_{\mathrm{loc}}((0,T);L^2(\mathbb{T}^3)) V h ∈ H loc 1 (( 0 , T ) ; L 2 ( T 3 )) : those labels have fixed position, and their measure is α b g d x \alpha_{\mathrm{bg}}\,dx α bg d x . Together with the spatial Lipschitz bound, this justifies the chain rule on the moving labels. To see this without assuming a time-smooth optimizer, let Φ h \Phi_h Φ h be the smooth flow of v h v_h v h on a compact time interval. Its Jacobian and inverse Jacobian are bounded for this fixed h h h . Mollify V h V_h V h in space and time, apply the classical chain rule to V h ( t , Φ h ( t , a ) ) V_h(t,\Phi_h(t,a)) V h ( t , Φ h ( t , a )) , and pass to the limit in L 2 L^2 L 2 after the change of variables x = Φ h ( t , a ) x=\Phi_h(t,a) x = Φ h ( t , a ) . The convergence follows from the L 2 L^2 L 2 time derivative, the bounded spatial derivative, and the smooth fixed-h h h velocity. Hence
d d t T h ( t , Φ h ( t , a ) ) = ( ∂ t V h + H h v h ) ( t , Φ h ( t , a ) ) . \frac{d}{dt}T_h(t,\Phi_h(t,a))=(\partial_tV_h+H_hv_h)(t,\Phi_h(t,a)). d t d T h ( t , Φ h ( t , a )) = ( ∂ t V h + H h v h ) ( t , Φ h ( t , a )) .
The stationary derivative is ∂ t V h \partial_tV_h ∂ t V h . Bounds for the flow in this chain-rule justification may depend on h h h ; they are not used in the estimates below, which come directly from the Hilbert inequality. Consequently,
∫ T 3 ( m h ⋅ ∂ t V h + m h ⊗ m h ρ h : H h ) d x ≥ 1 K ∫ T 3 ρ h ∣ ∂ t V h + H h v h ∣ 2 d x + α b g K ∫ T 3 ∣ ∂ t V h ∣ 2 d x . (42) \int_{\mathbb{T}^3}\left(m_h\cdot\partial_tV_h+\frac{m_h\otimes m_h}{\rho_h}:H_h\right)\,dx
\geq\frac{1}{K}\int_{\mathbb{T}^3}\rho_h\lvert\partial_tV_h+H_hv_h\rvert^2\,dx+\frac{\alpha_{\mathrm{bg}}}{K}\int_{\mathbb{T}^3}\lvert\partial_tV_h\rvert^2\,dx. \tag*{(42)} ∫ T 3 ( m h ⋅ ∂ t V h + ρ h m h ⊗ m h : H h ) d x ≥ K 1 ∫ T 3 ρ h ∣ ∂ t V h + H h v h ∣ 2 d x + K α bg ∫ T 3 ∣ ∂ t V h ∣ 2 d x . (42)
The background contributes nothing to the left pairing because its velocity is zero, but contributes α b g ∥ ∂ t V h ∥ L 2 2 / K \alpha_{\mathrm{bg}}\lVert\partial_tV_h\rVert_{L^2}^2/K α bg ∥ ∂ t V h ∥ L 2 2 / K on the right. This provides time regularity also on vacuum regions. Since H h ⪰ 0 H_h\succeq0 H h ⪰ 0 and R h ⪰ 0 R_h\succeq0 R h ⪰ 0 , their contraction is nonnegative:
R h : H h = tr ( R h 1 / 2 H h R h 1 / 2 ) ≥ 0. R_h:H_h=\operatorname{tr}(R_h^{1/2}H_hR_h^{1/2})\geq0. R h : H h = tr ( R h 1/2 H h R h 1/2 ) ≥ 0.
Adding it to (42) and integrating against η ≥ 0 \eta\geq0 η ≥ 0 yields
M h ( η ) ≥ α b g K ∫ Q T η ∣ ∂ t V h ∣ 2 d x d t + ∫ Q T η R h : H h d x d t ≥ 0. (43) \mathcal{M}_h(\eta)\geq\frac{\alpha_{\mathrm{bg}}}{K}\int_{Q_T}\eta\lvert\partial_tV_h\rvert^2\,dx\,dt+\int_{Q_T}\eta R_h:H_h\,dx\,dt\geq0. \tag*{(43)} M h ( η ) ≥ K α bg ∫ Q T η ∣ ∂ t V h ∣ 2 d x d t + ∫ Q T η R h : H h d x d t ≥ 0. (43)
This inequality includes the covariance R h R_h R h and hence the full smoothed convective tensor ( ρ u ⊗ u ) h (\rho u\otimes u)_h ( ρ u ⊗ u ) h .
The K K K -Lipschitz bound on T K \mathcal{T}_K T K also controls the time derivative independently of the smoothing scale:
α b g ∫ T 3 ∣ ∂ t V h ∣ 2 d x ≤ ∥ T ˙ K ∥ H 2 ≤ K 2 ∥ X ˙ ∥ H 2 = K 2 ∫ T 3 ∣ m h ∣ 2 ρ h d x ≤ K 2 ∫ T 3 ( ρ ∣ u ∣ 2 ) h d x . \begin{aligned}
\alpha_{\mathrm{bg}}\int_{\mathbb{T}^3}\lvert\partial_tV_h\rvert^2\,dx\leq\lVert\dot{\mathcal{T}}_K\rVert_{\mathcal{H}}^2\leq K^2\lVert\dot{X}\rVert_{\mathcal{H}}^2 \\
&=K^2\int_{\mathbb{T}^3}\frac{\lvert m_h\rvert^2}{\rho_h}\,dx\leq K^2\int_{\mathbb{T}^3}(\rho\lvert u\rvert^2)_h\,dx.
\end{aligned} α bg ∫ T 3 ∣ ∂ t V h ∣ 2 d x ≤ ∥ T ˙ K ∥ H 2 ≤ K 2 ∥ X ˙ ∥ H 2 = K 2 ∫ T 3 ρ h ∣ m h ∣ 2 d x ≤ K 2 ∫ T 3 ( ρ ∣ u ∣ 2 ) h d x .
The energy bound therefore yields a local L 2 L^2 L 2 bound for ∂ t V h \partial_tV_h ∂ t V h .
Step 3: Removal of smoothing and admissibility. Keep the approximate solution, K K K , and α b g \alpha_{\mathrm{bg}} α bg fixed. For two nonnegative densities of equal mass, couple their common part identically and transport the remaining mass at distance at most D D D . Realizing this coupling on common atomless labels, with identical stationary-background labels, and applying (B.2), we obtain
α b g ∥ V [ ρ ] − V [ ρ ~ ] ∥ L 2 ( T 3 ) 2 ≤ K 2 W 2 , T 3 2 ( ρ d x , ρ ~ d x ) ≤ K 2 D 2 2 ∥ ρ − ρ ~ ∥ L 1 ( T 3 ) . (44) \begin{aligned}
\alpha_{\mathrm{bg}}\lVert V[\rho]-V[\widetilde{\rho}]\rVert_{L^2(\mathbb{T}^3)}^2\leq K^2W_{2,\mathbb{T}^3}^2(\rho\,dx,\widetilde{\rho}\,dx) \\
&\leq\frac{K^2D^2}{2}\lVert\rho-\widetilde{\rho}\rVert_{L^1(\mathbb{T}^3)}. \tag*{(44)}
\end{aligned} α bg ∥ V [ ρ ] − V [ ρ ] ∥ L 2 ( T 3 ) 2 ≤ K 2 W 2 , T 3 2 ( ρ d x , ρ d x ) ≤ 2 K 2 D 2 ∥ ρ − ρ ∥ L 1 ( T 3 ) . (44)
The transport cost is unnormalized. Common smoothing preserves mass at each time, and ρ h → ρ \rho_h \to\rho ρ h → ρ locally in L 1 ( Q T ) L^1(Q_T) L 1 ( Q T ) , so (44) yields V h → V V_h \to V V h → V strongly in local space–time L 2 L^2 L 2 . The spatial Lipschitz and time-derivative bounds then give
∂ t V h ⇀ ∂ t V locally in L 2 , H h ⇀ ∗ H in L ∞ . \partial_t V_h \rightharpoonup\partial_t V \quad\text{locally in }L^2,\qquad H_h \overset{*}{\rightharpoonup} H \quad\text{in }L^\infty. ∂ t V h ⇀ ∂ t V locally in L 2 , H h ⇀ ∗ H in L ∞ .
For this fixed solution, the L 3 L^3 L 3 density bound and Sobolev embedding imply
∥ m ∥ L t 2 L x 2 ≤ ∥ ρ ∥ L t ∞ L x 3 ∥ u ∥ L t 2 L x 6 < ∞ . (45) \|m\|_{L_t^2L_x^2} \le\|\rho\|_{L_t^\infty L_x^3}\|u\|_{L_t^2L_x^6}<\infty. \tag*{(45)} ∥ m ∥ L t 2 L x 2 ≤ ∥ ρ ∥ L t ∞ L x 3 ∥ u ∥ L t 2 L x 6 < ∞. (45)
On compact time intervals, the mollified fields converge as follows:
m h → m in L 2 , ( ρ u ⊗ u ) h → ρ u ⊗ u , P h → P in L 1 , S h → S in L 2 . m_h \to m \quad\text{in }L^2,\qquad(\rho u\otimes u)_h\to\rho u\otimes u,\qquad P_h\to P \quad\text{in }L^1,\qquad S_h\to S \quad\text{in }L^2. m h → m in L 2 , ( ρ u ⊗ u ) h → ρ u ⊗ u , P h → P in L 1 , S h → S in L 2 .
Pair the strong limit of m h m_h m h with the weak limit of ∂ t V h \partial_tV_h ∂ t V h , and the strong L 1 L^1 L 1 limit of ( ρ u ⊗ u ) h (\rho u\otimes u)_h ( ρ u ⊗ u ) h with the weak-star limit of H h H_h H h . Passing to the limit in (43) proves (41) and V ∈ H l o c 1 ( ( 0 , T ) ; L 2 ( T 3 ) ) V\in H^1_{\mathrm{loc}}((0,T);L^2(\mathbb{T}^3)) V ∈ H loc 1 (( 0 , T ) ; L 2 ( T 3 )) . For example, the temporal pairing splits as
∫ η ( m h ⋅ ∂ t V h − m ⋅ ∂ t V ) = ∫ η ( m h − m ) ⋅ ∂ t V h + ∫ η m ⋅ ( ∂ t V h − ∂ t V ) . \int\eta(m_h\cdot\partial_tV_h-m\cdot\partial_tV)
=\int\eta(m_h-m)\cdot\partial_tV_h
+\int\eta m\cdot(\partial_tV_h-\partial_tV). ∫ η ( m h ⋅ ∂ t V h − m ⋅ ∂ t V ) = ∫ η ( m h − m ) ⋅ ∂ t V h + ∫ η m ⋅ ( ∂ t V h − ∂ t V ) .
The first term tends to zero by the uniform smoothing-scale L 2 L^2 L 2 bound for ∂ t V h \partial_tV_h ∂ t V h ; the second does so by weak convergence. For the convective pairing the identical splitting uses strong L 1 L^1 L 1 and weak-star L ∞ L^\infty L ∞ convergence. Thus no strong convergence of the optimizer Hessians is required in this step.
To justify η V \eta V η V as a momentum test, mollify the limiting field V V V in space and time. Its time derivative converges in L 2 L^2 L 2 and pairs with m ∈ L 2 m\in L^2 m ∈ L 2 . Its spatial gradient converges in every finite L p L^p L p and remains bounded in L ∞ L^\infty L ∞ . An almost-everywhere convergent subsequence then permits dominated convergence against the L 1 L^1 L 1 pressure and convective tensors, while strong L 2 L^2 L 2 convergence handles S ∈ L 2 S\in L^2 S ∈ L 2 . This proves admissibility. The preceding sequence H h H_h H h converges weak-star; the almost-everywhere convergence here concerns mollifications of the limiting field. All smoothing limits are taken at a fixed approximate solution, before n → ∞ n\to\infty n → ∞ . □
Pressure Estimate
The Hessian bound from Lemma 3.1 and the signed pairing from Lemma 3.2 give the pressure and tail estimates through the weak momentum equation.
Lemma 3.3. Under the assumptions of Proposition 2.2 , for every finite T > 0 T>0 T > 0 there exists C T C_T C T , depending only on E ∗ E_* E ∗ , T T T and the torus, such that for all n n n , 0 < δ ≤ 1 0<\delta\le1 0 < δ ≤ 1 , and R ≥ 1 R\ge1 R ≥ 1 ,
∫ Q T P n min { δ − 1 , ρ n 1 / 3 } d x d t ≤ C T δ − 1 / 2 , \int_{Q_T} P_n\min\{\delta^{-1},\rho_n^{1/3}\}\,dx\,dt\le C_T\delta^{-1/2}, ∫ Q T P n min { δ − 1 , ρ n 1/3 } d x d t ≤ C T δ − 1/2 , ∫ Q T ∩ { ρ n > R } P n d x d t ≤ C T R − 1 / 6 . \int_{Q_T\cap\{\rho_n>R\}}P_n\,dx\,dt\le C_TR^{-1/6}. ∫ Q T ∩ { ρ n > R } P n d x d t ≤ C T R − 1/6 . For the family in Lemma 2.1 , these constants are independent of e n e_n e n and the individual L t ∞ L x 3 L_t^\infty L_x^3 L t ∞ L x 3 norms. The estimates give (24) –(25) of Proposition 2.2 .
Proof. Fix n n n and K ≥ 6 K \ge6 K ≥ 6 . Write m n = ∫ T 3 ρ n d x \mathbf{m}_n = \int_{\mathbb{T}^3} \rho_n\,\mathrm{d}x m n = ∫ T 3 ρ n d x and choose
α b g = max { 1 , sup n m n } , c n = m n + α b g , Λ n = K c n 1 / 3 . (46) \alpha_{\mathrm{bg}} = \max\left\{1,\sup_n \mathbf{m}_n\right\}, \qquad c_n = \mathbf{m}_n + \alpha_{\mathrm{bg}}, \qquad\Lambda_n = K c_n^{1/3}. \tag*{(46)} α bg = max { 1 , n sup m n } , c n = m n + α bg , Λ n = K c n 1/3 . (46)
Then 1 ≤ α b g ≤ c n ≤ 2 α b g 1 \le\alpha_{\mathrm{bg}} \le c_n \le2\alpha_{\mathrm{bg}} 1 ≤ α bg ≤ c n ≤ 2 α bg , uniformly in n n n .
Step 1: Momentum testing. Let 0 ≤ η ∈ C c ∞ ( 0 , T ) 0 \le\eta\in C^\infty_c(0,T) 0 ≤ η ∈ C c ∞ ( 0 , T ) . By Lemma 3.2 , the field η V \eta V η V is admissible in the weak momentum equation. Using ∇ V = H − Id \nabla V = H - \operatorname{Id} ∇ V = H − Id and div V = tr H − 3 \operatorname{div} V = \operatorname{tr} H - 3 div V = tr H − 3 , we obtain
0 = ∫ Q T η ′ m n ⋅ V + ∫ Q T η m n ⋅ ∂ t V + ∫ Q T η ( ρ n u n ⊗ u n ) : ( H − Id ) + ∫ Q T η P n ( tr H − 3 ) − ∫ Q T η S n : ( H − Id ) . (47) \begin{aligned}
0 ={}& \int_{Q_T} \eta' \mathbf{m}_n \cdot V + \int_{Q_T} \eta\mathbf{m}_n \cdot\partial_t V \\
&+ \int_{Q_T} \eta(\rho_n u_n \otimes u_n) : (H-\operatorname{Id}) \\
&+ \int_{Q_T} \eta P_n(\operatorname{tr} H-3) - \int_{Q_T} \eta S_n : (H-\operatorname{Id}). \tag*{(47)}
\end{aligned} 0 = ∫ Q T η ′ m n ⋅ V + ∫ Q T η m n ⋅ ∂ t V + ∫ Q T η ( ρ n u n ⊗ u n ) : ( H − Id ) + ∫ Q T η P n ( tr H − 3 ) − ∫ Q T η S n : ( H − Id ) . (47)
Since
( ρ n u n ⊗ u n ) : ( H − Id ) = ( ρ n u n ⊗ u n ) : H − ρ n ∣ u n ∣ 2 , (\rho_n u_n \otimes u_n) : (H-\operatorname{Id}) = (\rho_n u_n \otimes u_n) : H - \rho_n |u_n|^2, ( ρ n u n ⊗ u n ) : ( H − Id ) = ( ρ n u n ⊗ u n ) : H − ρ n ∣ u n ∣ 2 ,
separating the kinetic term leaves the material pairing from Lemma 3.2 . Thus
∫ Q T η P n ( tr H − 3 ) + M K , α b g ( η ) = − ∫ Q T η ′ m n ⋅ V + ∫ Q T η ρ n ∣ u n ∣ 2 + ∫ Q T η S n : ( H − Id ) , (48) \begin{aligned}
\int_{Q_T} \eta P_n(\operatorname{tr} H-3) + \mathcal{M}_{K,\alpha_{\mathrm{bg}}}(\eta)
={}&-\int_{Q_T} \eta' \mathbf{m}_n \cdot V + \int_{Q_T} \eta\rho_n |u_n|^2 \\
&+ \int_{Q_T} \eta S_n : (H-\operatorname{Id}), \tag*{(48)}
\end{aligned} ∫ Q T η P n ( tr H − 3 ) + M K , α bg ( η ) = − ∫ Q T η ′ m n ⋅ V + ∫ Q T η ρ n ∣ u n ∣ 2 + ∫ Q T η S n : ( H − Id ) , (48)
where M K , α b g ( η ) ≥ 0 \mathcal{M}_{K,\alpha_{\mathrm{bg}}}(\eta) \ge0 M K , α bg ( η ) ≥ 0 .
Step 2: Energy and Hessian bounds. To estimate the right-hand side of (48) , use mass conservation and Cauchy–Schwarz:
∥ m n ( t ) ∥ 1 ≤ m n 1 / 2 ( ∫ T 3 ρ n ∣ u n ∣ 2 d x ) 1 / 2 . \|\mathbf{m}_n(t)\|_1 \le m_n^{1/2}\left(\int_{\mathbb{T}^3} \rho_n |u_n|^2\,\mathrm{d}x\right)^{1/2}. ∥ m n ( t ) ∥ 1 ≤ m n 1/2 ( ∫ T 3 ρ n ∣ u n ∣ 2 d x ) 1/2 .
Consequently, the uniform energy bounds and ∥ V ∥ ∞ ≤ D \|V\|_\infty\le D ∥ V ∥ ∞ ≤ D give
∣ ∫ Q T η ′ m n ⋅ V ∣ ≤ D ∥ η ′ ∥ 1 ess sup t ∥ m n ( t ) ∥ 1 ≤ C T ∥ η ′ ∥ 1 , \left|\int_{Q_T} \eta' \mathbf{m}_n \cdot V\right| \le D\|\eta'\|_1 \operatorname*{ess\,sup}_t \|\mathbf{m}_n(t)\|_1 \le C_T\|\eta'\|_1, ∫ Q T η ′ m n ⋅ V ≤ D ∥ η ′ ∥ 1 t ess sup ∥ m n ( t ) ∥ 1 ≤ C T ∥ η ′ ∥ 1 ,
∫ Q T η ( ρ n ∣ u n ∣ 2 + 3 P n ) ≤ T ∥ η ∥ ∞ ess sup t ∫ T 3 ( ρ n ∣ u n ∣ 2 + 3 P n ) d x ≤ C T ∥ η ∥ ∞ . \int_{Q_T} \eta(\rho_n|u_n|^2+3P_n)
\le T\|\eta\|_\infty\operatorname*{ess\,sup}_t \int_{\mathbb{T}^3}(\rho_n|u_n|^2+3P_n)\,\mathrm{d}x
\le C_T\|\eta\|_\infty. ∫ Q T η ( ρ n ∣ u n ∣ 2 + 3 P n ) ≤ T ∥ η ∥ ∞ t ess sup ∫ T 3 ( ρ n ∣ u n ∣ 2 + 3 P n ) d x ≤ C T ∥ η ∥ ∞ .
For the viscous term, the Hessian estimate (3.4) yields
∣ ∫ Q T η S n : ( H − Id ) ∣ ≤ ∥ η 1 / 2 S n ∥ 2 ∥ η 1 / 2 ( H − Id ) ∥ 2 ≤ 3 T ( K − 1 ) ∥ η ∥ ∞ ∥ S n ∥ L 2 ( Q T ) ≤ C T ∥ η ∥ ∞ K 1 / 2 . (49) \begin{aligned}
\left|\int_{Q_T} \eta S_n : (H-\operatorname{Id})\right|
\le\|\eta^{1/2}S_n\|_2\|\eta^{1/2}(H-\operatorname{Id})\|_2 \\
&\le\sqrt{3T(K-1)}\|\eta\|_\infty\|S_n\|_{L^2(Q_T)}
\le C_T\|\eta\|_\infty K^{1/2}. \tag*{(49)}
\end{aligned} ∫ Q T η S n : ( H − Id ) ≤ ∥ η 1/2 S n ∥ 2 ∥ η 1/2 ( H − Id ) ∥ 2 ≤ 3 T ( K − 1 ) ∥ η ∥ ∞ ∥ S n ∥ L 2 ( Q T ) ≤ C T ∥ η ∥ ∞ K 1/2 . (49)
Inserting these bounds into (48) and moving the term − 3 P n -3P_n − 3 P n to the right gives
∫ Q T η P n tr H + M K , α b g ( η ) ≤ C T ( ∥ η ′ ∥ 1 + ∥ η ∥ ∞ ) ( 1 + K 1 / 2 ) . (50) \int_{Q_T} \eta P_n\operatorname{tr} H + \mathcal{M}_{K,\alpha_{\mathrm{bg}}}(\eta)
\le C_T(\|\eta'\|_1+\|\eta\|_\infty)(1+K^{1/2}). \tag*{(50)} ∫ Q T η P n tr H + M K , α bg ( η ) ≤ C T ( ∥ η ′ ∥ 1 + ∥ η ∥ ∞ ) ( 1 + K 1/2 ) . (50)
One can keep the budget constant explicit at this point:
∫ Q T η P n tr H + M K , α b g ( η ) ≤ D E ∗ ∥ η ′ ∥ 1 + ( 3 T E ∗ + 3 T ( K − 1 ) E ∗ ) ∥ η ∥ ∞ . \int_{Q_T} \eta P_n \operatorname{tr} H+\mathcal{M}_{K,\alpha_{\mathrm{bg}}}(\eta)
\le DE_* \|\eta'\|_1+\left(3T E_*+\sqrt{3T(K-1)E_*}\right)\|\eta\|_\infty. ∫ Q T η P n tr H + M K , α bg ( η ) ≤ D E ∗ ∥ η ′ ∥ 1 + ( 3 T E ∗ + 3 T ( K − 1 ) E ∗ ) ∥ η ∥ ∞ .
Indeed, both the mass and kinetic energy are at most E ∗ E_* E ∗ , ∫ ( ρ n ∣ u n ∣ 2 + 3 P n ) ≤ 3 E ∗ \int(\rho_n|u_n|^2+3P_n)\le3E_* ∫ ( ρ n ∣ u n ∣ 2 + 3 P n ) ≤ 3 E ∗ at almost every time, and ∥ S n ∥ 2 2 ≤ E ∗ \|S_n\|_2^2\le E_* ∥ S n ∥ 2 2 ≤ E ∗ . Moreover, c n ≤ 2 max { 1 , E ∗ } c_n\le2\max\{1,E_*\} c n ≤ 2 max { 1 , E ∗ } . Thus the normalizations and the small-cap case below only change constants through E ∗ E_* E ∗ and T T T ; neither the individual cubic-density bound nor the time-regularity constant of the optimizer enters the final estimate.
By the trace lower bound (3.3) and ρ n + α b g ≥ ρ n \rho_n+\alpha_{\mathrm{bg}}\ge\rho_n ρ n + α bg ≥ ρ n ,
min { Λ n , ρ n 1 / 3 } ≤ min { K c n 1 / 3 , 3 ( ρ n + α b g ) 1 / 3 } ≤ c n 1 / 3 tr H . (51) \begin{aligned}
\min\{\Lambda_n,\rho_n^{1/3}\}&\le\min\{Kc_n^{1/3},3(\rho_n+\alpha_{\mathrm{bg}})^{1/3}\}\\
&\le c_n^{1/3}\operatorname{tr}H. \tag*{(51)}
\end{aligned} min { Λ n , ρ n 1/3 } ≤ min { K c n 1/3 , 3 ( ρ n + α bg ) 1/3 } ≤ c n 1/3 tr H . (51)
Multiplying (3.21) by c n 1 / 3 c_n^{1/3} c n 1/3 and using K 1 / 2 = Λ n 1 / 2 c n − 1 / 6 K^{1/2}=\Lambda_n^{1/2}c_n^{-1/6} K 1/2 = Λ n 1/2 c n − 1/6 , we obtain
∫ Q T η P n min { Λ n , ρ n 1 / 3 } + c n 1 / 3 M K , α b g ( η ) ≤ C T ( ∥ η ′ ∥ 1 + ∥ η ∥ ∞ ) ( 1 + Λ n 1 / 2 ) . (52) \int_{Q_T}\eta P_n\min\{\Lambda_n,\rho_n^{1/3}\}+c_n^{1/3}\mathcal{M}_{K,\alpha_{\mathrm{bg}}}(\eta)
\le C_T(\|\eta'\|_1+\|\eta\|_\infty)(1+\Lambda_n^{1/2}). \tag*{(52)} ∫ Q T η P n min { Λ n , ρ n 1/3 } + c n 1/3 M K , α bg ( η ) ≤ C T ( ∥ η ′ ∥ 1 + ∥ η ∥ ∞ ) ( 1 + Λ n 1/2 ) . (52)
The constant is independent of n n n , K K K , and the individual L t ∞ L x 3 L^\infty_tL^3_x L t ∞ L x 3 bounds of ρ n \rho_n ρ n .
The normalized trace cap Λ n = K c n 1 / 3 \Lambda_n=Kc_n^{1/3} Λ n = K c n 1/3 is the density weight that can be detected by a given Hessian cap. Once a desired weight L L L is prescribed, the least admissible large cap is K = L c n − 1 / 3 K=Lc_n^{-1/3} K = L c n − 1/3 . A larger cap increases the viscous cost without improving min { L , ρ n 1 / 3 } \min\{L,\rho_n^{1/3}\} min { L , ρ n 1/3 } . Since 1 ≤ c n ≤ 2 α b g 1\le c_n\le2\alpha_{\mathrm{bg}} 1 ≤ c n ≤ 2 α bg , this is the choice K ∼ L K\sim L K ∼ L with uniform comparison constants. The power L 1 / 2 L^{1/2} L 1/2 below comes from the Hessian cost, whereas the density power 1 / 3 1/3 1/3 comes from the three-dimensional determinant inequality.
Step 3: Cutoff removal and tail decay. Choose 0 ≤ η ε ≤ 1 0\le\eta_\varepsilon\le1 0 ≤ η ε ≤ 1 in C ∞ ( 0 , T ) C^\infty(0,T) C ∞ ( 0 , T ) such that η ε → 1 \eta_\varepsilon\to1 η ε → 1 almost everywhere and ∥ η ε ′ ∥ 1 ≤ 2 \|\eta_\varepsilon'\|_1\le2 ∥ η ε ′ ∥ 1 ≤ 2 . One may use a smooth rise from zero to one near t = 0 t=0 t = 0 , a constant middle part, and a smooth fall near t = T t=T t = T ; their total variations add to two. Their derivatives need not converge in L 1 L^1 L 1 . The bound in (3.23) uses only this total variation and therefore controls both time endpoints uniformly. Dropping the nonnegative material pairing in (3.23) and applying Fatou’s lemma gives
∫ Q T P n min { Λ n , ρ n 1 / 3 } ≤ C T ( 1 + Λ n 1 / 2 ) . \int_{Q_T}P_n\min\{\Lambda_n,\rho_n^{1/3}\}\le C_T(1+\Lambda_n^{1/2}). ∫ Q T P n min { Λ n , ρ n 1/3 } ≤ C T ( 1 + Λ n 1/2 ) .
Let L = δ − 1 ≥ 1 L=\delta^{-1}\ge1 L = δ − 1 ≥ 1 . If L ≥ 6 c n 1 / 3 L\ge6c_n^{1/3} L ≥ 6 c n 1/3 , choose K = L c n − 1 / 3 ≥ 6 K=Lc_n^{-1/3}\ge6 K = L c n − 1/3 ≥ 6 . Then Λ n = L \Lambda_n=L Λ n = L , and
∫ Q T P n min { L , ρ n 1 / 3 } ≤ C T ( 1 + L 1 / 2 ) ≤ 2 C T L 1 / 2 . \int_{Q_T}P_n\min\{L,\rho_n^{1/3}\}\le C_T(1+L^{1/2})\le2C_TL^{1/2}. ∫ Q T P n min { L , ρ n 1/3 } ≤ C T ( 1 + L 1/2 ) ≤ 2 C T L 1/2 .
If 1 ≤ L < 6 c n 1 / 3 1 \le L < 6c_n^{1/3} 1 ≤ L < 6 c n 1/3 , the energy bound instead gives
∫ Q T P n min { L , ρ n 1 / 3 } ≤ L ∫ Q T P n ≤ C T L ≤ C T L 1 / 2 . \int_{Q_T} P_n \min\{L,\rho_n^{1/3}\} \le L\int_{Q_T} P_n
\le C_TL \le C_TL^{1/2}. ∫ Q T P n min { L , ρ n 1/3 } ≤ L ∫ Q T P n ≤ C T L ≤ C T L 1/2 .
where the last constant is uniform in n n n because c n c_n c n is uniformly bounded. Combining the two cases yields
∫ Q T P n min { δ − 1 , ρ n 1 / 3 } ≤ C T δ − 1 / 2 , 0 < δ ≤ 1. (53) \int_{Q_T} P_n \min\{\delta^{-1},\rho_n^{1/3}\} \le C_T\delta^{-1/2}, \qquad0<\delta\le1. \tag*{(53)} ∫ Q T P n min { δ − 1 , ρ n 1/3 } ≤ C T δ − 1/2 , 0 < δ ≤ 1. (53)
This proves (24) .
Finally, for R ≥ 1 R\ge1 R ≥ 1 , take δ = R − 1 / 3 \delta=R^{-1/3} δ = R − 1/3 . On { ρ n > R } \{\rho_n>R\} { ρ n > R } , the minimum in (53) equals R 1 / 3 R^{1/3} R 1/3 . Hence
∫ Q T ∩ { ρ n > R } P n ≤ R − 1 / 3 ∫ Q T P n min { R 1 / 3 , ρ n 1 / 3 } ≤ C T R − 1 / 6 . \int_{Q_T\cap\{\rho_n>R\}} P_n \le R^{-1/3}\int_{Q_T}P_n\min\{R^{1/3},\rho_n^{1/3}\}
\le C_TR^{-1/6}. ∫ Q T ∩ { ρ n > R } P n ≤ R − 1/3 ∫ Q T P n min { R 1/3 , ρ n 1/3 } ≤ C T R − 1/6 .
which proves (25) . In this last step the powers are R − 1 / 3 ( R 1 / 3 ) 1 / 2 = R − 1 / 6 R^{-1/3}(R^{1/3})^{1/2}=R^{-1/6} R − 1/3 ( R 1/3 ) 1/2 = R − 1/6 . The resulting uniform pressure bound depends only on the stated energy budget, although test admissibility used the fixed-member regularity. The constitutive law P n = ρ n γ + e n ρ n 3 P_n=\rho_n^\gamma+e_n\rho_n^3 P n = ρ n γ + e n ρ n 3 enters the subsequent conversion of this pressure tail into density moments. □
Proposition 2.2 follows from Lemma 3.3.
Effective Viscous Flux and Nonlinear Convergence
We prove Proposition 2.3 by combining the pressure tail estimate with effective viscous flux and entropy comparison. The tail estimate gives integrability; a high-frequency commutator bound then yields the spatially integrated flux identity. Renormalization and the common initial entropy eliminate density oscillations, after which momentum compactness gives the convective and energy limits.
Fix a finite T > 0 T>0 T > 0 and write
Q T = ( 0 , T ) × T 3 , m n = ρ n u n , P n = ρ n γ + e n ρ n 3 . Q_T=(0,T)\times\mathbb{T}^3,\qquad m_n=\rho_nu_n,\qquad P_n=\rho_n^\gamma+e_n\rho_n^3. Q T = ( 0 , T ) × T 3 , m n = ρ n u n , P n = ρ n γ + e n ρ n 3 .
The effective viscous flux method originates in the compactness theory of Lions and Feireisl [13 , 14 , 27 ] . Its use here has two separate prerequisites. The pressure tail must exclude concentration in the pressure–truncation product, and the momentum must have enough integrability to pair with a small high-frequency commutator. Neither prerequisite identifies the pressure by itself. The entropy comparison below performs that final identification.
The tail estimate (25) yields the following bounds.
Lemma 4.1. Let ( ρ n , u n ) (\rho_n,u_n) ( ρ n , u n ) be the family in Lemma 2.1 , with m n = ρ n u n m_n=\rho_nu_n m n = ρ n u n and P n = ρ n γ + e n ρ n 3 P_n=\rho_n^\gamma+e_n\rho_n^3 P n = ρ n γ + e n ρ n 3 . Set q c = γ + 1 / 6 q_c=\gamma+1/6 q c = γ + 1/6 . For every finite T > 0 T>0 T > 0 , the following estimates hold with constants independent of n n n .
(i) For 1 ≤ q < q c 1 \le q < q_c 1 ≤ q < q c and 0 < β < 1 / 6 0 < \beta< 1/6 0 < β < 1/6 ,
sup n ∫ Q T ρ n q d x d t ≤ C T , q , sup n ∫ Q T P n ρ n β d x d t ≤ C T , β . \sup_n \int_{Q_T} \rho_n^q\,dx\,dt \le C_{T,q}, \qquad\sup_n \int_{Q_T} P_n\rho_n^\beta\,dx\,dt \le C_{T,\beta}. n sup ∫ Q T ρ n q d x d t ≤ C T , q , n sup ∫ Q T P n ρ n β d x d t ≤ C T , β .
(ii) For 1 < s < 19 / 18 1 < s < 19/18 1 < s < 19/18 , the artificial pressure satisfies
∥ e n ρ n 3 ∥ L s ( Q T ) ≤ C T , s e n 1 − 1 / s . (54) \lVert e_n\rho_n^3\rVert_{L^s(Q_T)} \le C_{T,s}e_n^{1-1/s}. \tag*{(54)} ∥ e n ρ n 3 ∥ L s ( Q T ) ≤ C T , s e n 1 − 1/ s . (54)
Moreover,
sup n ∥ P n ∥ L s ( Q T ) ≤ C T , s , 1 < s < 1 + 1 6 max { 3 , γ } . \sup_n \lVert P_n\rVert_{L^s(Q_T)} \le C_{T,s}, \qquad1 < s < 1+\frac{1}{6\max\{3,\gamma\}}. n sup ∥ P n ∥ L s ( Q T ) ≤ C T , s , 1 < s < 1 + 6 max { 3 , γ } 1 .
In particular, e n ρ n 3 → 0 e_n\rho_n^3 \to0 e n ρ n 3 → 0 in the indicated L s L^s L s spaces, and { P n } \{P_n\} { P n } is uniformly integrable on Q T Q_T Q T .
(iii) Since γ > 4 / 3 \gamma> 4/3 γ > 4/3 , choose 3 / 2 < q < min { 2 , q c } 3/2 < q < \min\{2,q_c\} 3/2 < q < min { 2 , q c } , and set
p ( q ) = 6 ( q + 1 ) q + 11 , κ ( q ) = 2 q − 3 q + 1 . p(q)=\frac{6(q+1)}{q+11}, \qquad\kappa(q)=\frac{2q-3}{q+1}. p ( q ) = q + 11 6 ( q + 1 ) , κ ( q ) = q + 1 2 q − 3 .
Then
sup n ∥ m n ∥ L 2 ( 0 , T ; L p ( q ) ( T 3 ) ) ≤ C T , q , p ( q ) > 6 5 , κ ( q ) > 0. \sup_n \lVert m_n\rVert_{L^2(0,T;L^{p(q)}(\mathbb{T}^3))} \le C_{T,q}, \qquad p(q)>\frac{6}{5}, \qquad\kappa(q)>0. n sup ∥ m n ∥ L 2 ( 0 , T ; L p ( q ) ( T 3 )) ≤ C T , q , p ( q ) > 5 6 , κ ( q ) > 0.
Consequently L p ( q ) ( T 3 ) ⋐ H − 1 ( T 3 ) L^{p(q)}(\mathbb{T}^3)\Subset H^{-1}(\mathbb{T}^3) L p ( q ) ( T 3 ) ⋐ H − 1 ( T 3 ) , and κ ( q ) \kappa(q) κ ( q ) is the positive exponent used in the high-frequency estimate.
Proof. Step 1: Weighted moments. Let P ∗ : = sup n ess sup 0 < t < T ∫ T 3 P n d x P_*:=\sup_n\operatorname{ess\,sup}_{0<t<T}\int_{\mathbb{T}^3}P_n\,dx P ∗ := sup n ess sup 0 < t < T ∫ T 3 P n d x , which is finite by (2.2). From P n ≥ ρ n γ P_n\ge\rho_n^\gamma P n ≥ ρ n γ and (2.4), for R ≥ 1 R\ge1 R ≥ 1 ,
R γ ∣ Q T ∩ { ρ n > R } ∣ ≤ ∫ Q T ∩ { ρ n > R } ρ n γ ≤ ∫ Q T ∩ { ρ n > R } P n ≤ C T R − 1 / 6 . R^\gamma|Q_T\cap\{\rho_n>R\}|
\le\int_{Q_T\cap\{\rho_n>R\}}\rho_n^\gamma
\le\int_{Q_T\cap\{\rho_n>R\}}P_n
\le C_T R^{-1/6}. R γ ∣ Q T ∩ { ρ n > R } ∣ ≤ ∫ Q T ∩ { ρ n > R } ρ n γ ≤ ∫ Q T ∩ { ρ n > R } P n ≤ C T R − 1/6 .
The distribution-function formula therefore gives
∣ Q T ∩ { ρ n > R } ∣ ≤ C T R − q c , R ≥ 1 , |Q_T\cap\{\rho_n>R\}| \le C_T R^{-q_c}, \qquad R\ge1, ∣ Q T ∩ { ρ n > R } ∣ ≤ C T R − q c , R ≥ 1 ,
∫ Q T ρ n q d x d t ≤ T + C T , q ∫ 1 ∞ R q − 1 − q c d R = T + C T , q q q c − q , 1 ≤ q < q c . (55) \int_{Q_T}\rho_n^q\,dx\,dt
\le T+C_{T,q}\int_1^\infty R^{q-1-q_c}\,dR
= T+\frac{C_{T,q}q}{q_c-q}, \qquad1\le q<q_c. \tag*{(55)} ∫ Q T ρ n q d x d t ≤ T + C T , q ∫ 1 ∞ R q − 1 − q c d R = T + q c − q C T , q q , 1 ≤ q < q c . (55)
Applying the layer-cake formula with weight P n P_n P n yields
∫ Q T P n ρ n β d x d t ≤ T P ∗ + β ∫ 1 ∞ R β − 1 ∫ Q T ∩ { ρ n > R } P n d x d t d R ≤ T P ∗ + C T β 1 / 6 − β , 0 < β < 1 6 . (56) \begin{aligned}
\int_{Q_T}P_n\rho_n^\beta\,dx\,dt
\le TP_*+\beta\int_1^\infty R^{\beta-1}\int_{Q_T\cap\{\rho_n>R\}}P_n\,dx\,dt\,dR \\
&\le TP_*+\frac{C_T\beta}{1/6-\beta}, \qquad0<\beta<\frac{1}{6}. \tag*{(56)}
\end{aligned} ∫ Q T P n ρ n β d x d t ≤ T P ∗ + β ∫ 1 ∞ R β − 1 ∫ Q T ∩ { ρ n > R } P n d x d t d R ≤ T P ∗ + 1/6 − β C T β , 0 < β < 6 1 . (56)
The strict inequality q < q c q<q_c q < q c is required by the integral at infinity: at q = q c q=q_c q = q c its integrand is R − 1 R^{-1} R − 1 . Thus the tail estimate supplies all subcritical moments, without asserting a strong L q c L^{q_c} L q c endpoint. This proves (i).
Step 2: Pressure integrability. Take 1 < s < 19 / 18 1<s<19/18 1 < s < 19/18 and put β = 3 ( s − 1 ) < 1 / 6 \beta=3(s-1)<1/6 β = 3 ( s − 1 ) < 1/6 . Since e n ρ n 3 ≤ P n e_n\rho_n^3\le P_n e n ρ n 3 ≤ P n , estimate (4.3) gives
∫ Q T ( e n ρ n 3 ) s d x d t = e n s − 1 ∫ Q T e n ρ n 3 ρ n 3 ( s − 1 ) d x d t ≤ e n s − 1 ∫ Q T P n ρ n 3 ( s − 1 ) d x d t ≤ C T , s e n s − 1 . \begin{aligned}
\int_{Q_T}(e_n\rho_n^3)^s\,dx\,dt=e_n^{s-1}\int_{Q_T}e_n\rho_n^3\rho_n^{3(s-1)}\,dx\,dt \\
&\le e_n^{s-1}\int_{Q_T}P_n\rho_n^{3(s-1)}\,dx\,dt\le C_{T,s}e_n^{s-1}.
\end{aligned} ∫ Q T ( e n ρ n 3 ) s d x d t = e n s − 1 ∫ Q T e n ρ n 3 ρ n 3 ( s − 1 ) d x d t ≤ e n s − 1 ∫ Q T P n ρ n 3 ( s − 1 ) d x d t ≤ C T , s e n s − 1 .
This proves (4.1). Similarly,
∫ Q T ( ρ n γ ) s d x d t ≤ ∫ Q T P n ρ n γ ( s − 1 ) d x d t ≤ C T , s , 1 < s < 1 + 1 6 γ . \int_{Q_T}(\rho_n^\gamma)^s\,dx\,dt\le\int_{Q_T}P_n\rho_n^{\gamma(s-1)}\,dx\,dt\le C_{T,s},\qquad1<s<1+\frac{1}{6\gamma}. ∫ Q T ( ρ n γ ) s d x d t ≤ ∫ Q T P n ρ n γ ( s − 1 ) d x d t ≤ C T , s , 1 < s < 1 + 6 γ 1 .
Combining these estimates with P n s ≤ 2 s − 1 [ ( ρ n γ ) s + ( e n ρ n 3 ) s ] P_n^s\le2^{s-1}[(\rho_n^\gamma)^s+(e_n\rho_n^3)^s] P n s ≤ 2 s − 1 [( ρ n γ ) s + ( e n ρ n 3 ) s ] proves the asserted uniform L s L^s L s bound for P n P_n P n . Since s > 1 s>1 s > 1 , this also gives uniform integrability.
Step 3: Momentum interpolation. Fix 3 / 2 < q < min { 2 , q c } 3/2<q<\min\{2,q_c\} 3/2 < q < min { 2 , q c } and set ϑ = 2 / ( q + 1 ) \vartheta=2/(q+1) ϑ = 2/ ( q + 1 ) . On the vacuum set, both sides of the following identity are taken to be zero:
∣ m n ∣ = ∣ ρ n u n ∣ ϑ ∣ u n ∣ 1 − ϑ ρ n 1 − ϑ / 2 . |m_n|=|\sqrt{\rho_n}u_n|^\vartheta|u_n|^{1-\vartheta}\rho_n^{1-\vartheta/2}. ∣ m n ∣ = ∣ ρ n u n ∣ ϑ ∣ u n ∣ 1 − ϑ ρ n 1 − ϑ /2 .
The spatial Hölder exponents satisfy
1 p = ϑ 2 + 1 − ϑ 6 + 1 − ϑ / 2 q = q + 11 6 ( q + 1 ) , \frac{1}{p}=\frac{\vartheta}{2}+\frac{1-\vartheta}{6}+\frac{1-\vartheta/2}{q}=\frac{q+11}{6(q+1)}, p 1 = 2 ϑ + 6 1 − ϑ + q 1 − ϑ /2 = 6 ( q + 1 ) q + 11 ,
while ( 1 − ϑ ) + ( 2 − ϑ ) / q = 1 (1-\vartheta)+(2-\vartheta)/q=1 ( 1 − ϑ ) + ( 2 − ϑ ) / q = 1 gives the time exponents. Applying Hölder in space and then in time, we obtain
∥ m n ∥ L 2 ( 0 , T ; L p ) ≤ ( ess sup t ∫ T 3 ρ n ∣ u n ∣ 2 d x ) 1 / ( q + 1 ) ( ∫ 0 T ∥ u n ∥ 6 2 d t ) ( q − 1 ) / ( 2 ( q + 1 ) ) ( ∫ Q T ρ n q d x d t ) 1 / ( q + 1 ) . (57) \|m_n\|_{L^2(0,T;L^p)}\le\left(\operatorname*{ess\,sup}_{t}\int_{\mathbb{T}^3}\rho_n|u_n|^2\,dx\right)^{1/(q+1)}
\left(\int_0^T\|u_n\|_6^2\,dt\right)^{(q-1)/(2(q+1))}
\left(\int_{Q_T}\rho_n^q\,dx\,dt\right)^{1/(q+1)}. \tag*{(57)} ∥ m n ∥ L 2 ( 0 , T ; L p ) ≤ ( t ess sup ∫ T 3 ρ n ∣ u n ∣ 2 d x ) 1/ ( q + 1 ) ( ∫ 0 T ∥ u n ∥ 6 2 d t ) ( q − 1 ) / ( 2 ( q + 1 )) ( ∫ Q T ρ n q d x d t ) 1/ ( q + 1 ) . (57)
The right-hand side is uniformly bounded in n n n by (2.2) and (4.2); Lemma 2.5 controls velocities with arbitrary spatial mean. Finally,
p = 6 ( q + 1 ) q + 11 > 6 5 , 5 2 − 3 p = κ ( q ) = 2 q − 3 q + 1 > 0. p=\frac{6(q+1)}{q+11}>\frac{6}{5},\qquad\frac{5}{2}-\frac{3}{p}=\kappa(q)=\frac{2q-3}{q+1}>0. p = q + 11 6 ( q + 1 ) > 5 6 , 2 5 − p 3 = κ ( q ) = q + 1 2 q − 3 > 0.
Indeed,
p − 6 5 = 6 ( 4 q − 6 ) 5 ( q + 11 ) > 0 , q c − 3 2 = γ − 4 3 > 0. p-\frac{6}{5}=\frac{6(4q-6)}{5(q+11)}>0,\qquad q_c-\frac{3}{2}=\gamma-\frac{4}{3}>0. p − 5 6 = 5 ( q + 11 ) 6 ( 4 q − 6 ) > 0 , q c − 2 3 = γ − 3 4 > 0.
The second inequality provides a nonempty choice of q q q ; the first converts that choice into the spatial compactness needed for momentum. By Rellich compactness and duality [9 ] , H 1 ⋐ L p ′ H^1\Subset L^{p'} H 1 ⋐ L p ′ and L p ⋐ H − 1 L^p\Subset H^{-1} L p ⋐ H − 1 . This completes (iii). The quantity 5 / 2 − 3 / p 5/2-3/p 5/2 − 3/ p measures the remaining integrability available for the high-frequency kernel; its positivity will give a power of the smoothing scale. □
The following high-frequency effective-flux estimate is uniform in both the approximation index n n n and the density truncation level k k k . Let χ ∈ C c ∞ ( B 1 ) \chi\in C_c^\infty(B_1) χ ∈ C c ∞ ( B 1 ) be even, nonnegative, and have unit integral. For 0 < a < 1 / 8 0<a<1/8 0 < a < 1/8 , write
S a g = χ a ∗ T 3 g , H a = I − S a , A a , i = A i H a , L i j , a = R i j H a . S_ag=\chi_a*_{\mathbb{T}^3}g,\qquad H_a=I-S_a,\qquad A_{a,i}=A_iH_a,\qquad L_{ij,a}=R_{ij}H_a. S a g = χ a ∗ T 3 g , H a = I − S a , A a , i = A i H a , L ij , a = R ij H a .
The inverse Laplacian acts on mean-zero functions, so that
div A f = f − f ‾ , div A a f = H a f , S a f ‾ = f ‾ . \operatorname{div} Af = f - \overline{f}, \qquad\operatorname{div} A_a f = H_a f, \qquad S_a\overline{f} = \overline{f}. div A f = f − f , div A a f = H a f , S a f = f .
Thus A A A tests pressure through a prescribed divergence, while A a A_a A a isolates the part that must vanish with a a a . The complementary part S a f S_a f S a f retains the spatial mean. Removing that mean from both parts would change the flux calculation. The finite-exponent bounds for the Riesz transforms follow from singular-integral theory [40 ] . The quantitative periodic kernel bounds needed here are proved in Lemma C.2 of Appendix § 7 .
Lemma 4.2 (Uniform high-frequency effective-flux remainder). Let ( ρ n , u n ) (\rho_n,u_n) ( ρ n , u n ) be the family in Lemma 2.1 . Choose q q q as in Lemma 4.1 (iii) and put
p = p ( q ) = 6 ( q + 1 ) q + 11 , κ ( q ) = 2 q − 3 q + 1 > 0. p=p(q)=\frac{6(q+1)}{q+11}, \qquad\kappa(q)=\frac{2q-3}{q+1}>0. p = p ( q ) = q + 11 6 ( q + 1 ) , κ ( q ) = q + 1 2 q − 3 > 0. For every η ∈ C c ∞ ( 0 , T ) \eta\in C^\infty_c(0,T) η ∈ C c ∞ ( 0 , T ) , 0 < a < 1 / 8 0<a<1/8 0 < a < 1/8 , and truncation level k ≥ 1 k\geq1 k ≥ 1 , the field η A a b k ( ρ n ) \eta A_a b_k(\rho_n) η A a b k ( ρ n ) is an admissible test in the weak momentum equation of each fixed approximate solution. Moreover, setting
R n , k , a = ∫ Q T η ( P n − ν div u n ) H a b k ( ρ n ) d x d t , \mathcal{R}_{n,k,a}=\int_{Q_T}\eta(P_n-\nu\operatorname{div}u_n)H_a b_k(\rho_n)\,dx\,dt, R n , k , a = ∫ Q T η ( P n − ν div u n ) H a b k ( ρ n ) d x d t , we have
sup n , k ∣ R n , k , a ∣ ≤ C T , q [ ∥ η ∥ ∞ ( a κ ( q ) + a 2 ) + a ∥ η ′ ∥ 1 ] ⟶ 0 as a ↓ 0. (58) \sup_{n,k}|\mathcal{R}_{n,k,a}|\leq C_{T,q}\left[\|\eta\|_\infty\left(a^{\kappa(q)}+a^2\right)+a\|\eta'\|_1\right]\longrightarrow0\qquad\text{as }a\downarrow0. \tag*{(58)} n , k sup ∣ R n , k , a ∣ ≤ C T , q [ ∥ η ∥ ∞ ( a κ ( q ) + a 2 ) + a ∥ η ′ ∥ 1 ] ⟶ 0 as a ↓ 0. (58) The constant is independent of n n n , k k k , a a a , and the individual L t ∞ L x 3 L^\infty_tL^3_x L t ∞ L x 3 bounds of ρ n \rho_n ρ n .
Proof. Step 1: The complete commutator bound. Fix n n n , k k k , a a a and write
d n = div u n , g = b k ( ρ n ) = T k ( ρ n ) k , d_n=\operatorname{div}u_n,\qquad g=b_k(\rho_n)=\frac{T_k(\rho_n)}{k}, d n = div u n , g = b k ( ρ n ) = k T k ( ρ n ) ,
h = c k ( ρ n ) d n = a k ( ρ n ) k d n . h=c_k(\rho_n)d_n=\frac{a_k(\rho_n)}{k}d_n. h = c k ( ρ n ) d n = k a k ( ρ n ) d n .
The renormalized continuity equation takes the form
∂ t g + div ( g u n ) + h = 0. (59) \partial_t g+\operatorname{div}(g u_n)+h=0. \tag*{(59)} ∂ t g + div ( g u n ) + h = 0. (59)
Its spatial mean satisfies h ‾ = − ∂ t g ‾ \overline{h}=-\partial_t\overline{g} h = − ∂ t g . Since A a A_a A a annihilates constants, A a h = A a ( h − h ‾ ) A_a h=A_a(h-\overline{h}) A a h = A a ( h − h ) ; the source term must therefore be retained in the high-frequency pairing. This source is produced by the DiPerna–Lions renormalization [11 ] , through z g ′ ( z ) − g ( z ) = c k ( z ) zg'(z)-g(z)=c_k(z) z g ′ ( z ) − g ( z ) = c k ( z ) . It is not an error in the continuity equation and cannot be omitted when differentiating A a g A_a g A a g . In particular, the differentiated test involves the complete expression
∂ t ( A a g ) i = − L i j , a ( g u n , j ) − A a , i h . \partial_t(A_a g)_i=-L_{ij,a}(g u_{n,j})-A_{a,i}h. ∂ t ( A a g ) i = − L ij , a ( g u n , j ) − A a , i h .
By Lemma C.2, for g ∈ L ∞ g\in L^\infty g ∈ L ∞ and ∣ h ∣ ≤ C h ∣ ∇ u ∣ |h|\leq C_h|\nabla u| ∣ h ∣ ≤ C h ∣∇ u ∣ ,
∣ L i j , a ( g u j ) − u j L i j , a g + A a , i h ∣ ≤ C ( ∥ g ∥ ∞ + C h ) K a ∗ T 3 ∣ ∇ u ∣ . (60) \left|L_{ij,a}(g u_j)-u_jL_{ij,a}g+A_{a,i}h\right|
\leq C(\|g\|_\infty+C_h)K_a*_{\mathbb{T}^3}|\nabla u|. \tag*{(60)} ∣ L ij , a ( g u j ) − u j L ij , a g + A a , i h ∣ ≤ C ( ∥ g ∥ ∞ + C h ) K a ∗ T 3 ∣∇ u ∣. (60)
where repeated indices are summed and
∥ K a ∥ s ≤ C s ( a 3 / s − 2 + a 2 ) , 1 ≤ s < 3 2 . \lVert\mathcal{K}_a\rVert_s \le C_s(a^{3/s-2}+a^2), \qquad1\le s<\frac{3}{2}. ∥ K a ∥ s ≤ C s ( a 3/ s − 2 + a 2 ) , 1 ≤ s < 2 3 .
Near the diagonal, the velocity difference in the commutator removes one power of the order-zero singular kernel. Integration along the line segment leaves the majorant r − 2 ( 1 + ∣ log ( r / a ) ∣ ) r^{-2}(1+|\log(r/a)|) r − 2 ( 1 + ∣ log ( r / a ) ∣ ) for r < a r<a r < a . Its s s s th power is locally integrable exactly when s < 3 / 2 s<3/2 s < 3/2 . Away from the diagonal, evenness of χ \chi χ cancels the first-order Taylor term, leaving a 2 r − 4 a^2r^{-4} a 2 r − 4 after the same line-segment integration. The a 2 a^2 a 2 term in the periodic bound also includes the nonsingular lattice translates. These are the two contributions established in Lemma C.2; no uniform bound on a spatial derivative of g g g is used. Set
B n , k , a , i = L i j , a ( g u n , j ) − u n , j L i j , a g + A a , i h , B_{n,k,a,i}=L_{ij,a}(gu_{n,j})-u_{n,j}L_{ij,a}g+A_{a,i}h, B n , k , a , i = L ij , a ( g u n , j ) − u n , j L ij , a g + A a , i h ,
J n , k , a ( η ) = ∫ Q T η m n ⋅ B n , k , a d x d t . J_{n,k,a}(\eta)=\int_{Q_T}\eta\,m_n\cdot B_{n,k,a}\,dx\,dt. J n , k , a ( η ) = ∫ Q T η m n ⋅ B n , k , a d x d t .
Let p ′ p' p ′ be the conjugate exponent of p p p . The Young and Hölder exponents can be chosen as
1 p = q + 1 6 ( q + 1 ) , p ′ = 6 ( q + 1 ) 5 ( q − 1 ) , s = 3 ( q + 1 ) 4 q − 1 , \frac{1}{p}=\frac{q+1}{6(q+1)},\qquad p'=\frac{6(q+1)}{5(q-1)},\qquad s=\frac{3(q+1)}{4q-1}, p 1 = 6 ( q + 1 ) q + 1 , p ′ = 5 ( q − 1 ) 6 ( q + 1 ) , s = 4 q − 1 3 ( q + 1 ) ,
1 s = 3 2 − 1 p , 1 + 1 p ′ = 1 2 + 1 s , 3 s − 2 = κ ( q ) > 0. (61) \frac{1}{s}=\frac{3}{2}-\frac{1}{p},\qquad1+\frac{1}{p'}=\frac{1}{2}+\frac{1}{s},\qquad\frac{3}{s}-2=\kappa(q)>0. \tag*{(61)} s 1 = 2 3 − p 1 , 1 + p ′ 1 = 2 1 + s 1 , s 3 − 2 = κ ( q ) > 0. (61)
The convolution is paired first in space and then in time:
L x s ∗ L x 2 ⟶ L x p ′ , L t 2 L x p ∗ L t 2 L x p ′ ⟶ L t , x 1 . L_x^s*L_x^2\longrightarrow L_x^{p'},\qquad L_t^2L_x^p*L_t^2L_x^{p'}\longrightarrow L_{t,x}^1. L x s ∗ L x 2 ⟶ L x p ′ , L t 2 L x p ∗ L t 2 L x p ′ ⟶ L t , x 1 .
Since 3 / 2 < q < 2 3/2<q<2 3/2 < q < 2 , we have 1 < s < 3 / 2 1<s<3/2 1 < s < 3/2 and p ′ < 6 p'<6 p ′ < 6 . At q = 3 / 2 q=3/2 q = 3/2 , one would have s = 3 / 2 s=3/2 s = 3/2 and κ ( q ) = 0 \kappa(q)=0 κ ( q ) = 0 ; the near-diagonal integral would cease to be integrable. The strict inequality on q q q therefore enters both the kernel estimate and its decay as a ↓ 0 a\downarrow0 a ↓ 0 . Using ∥ b k ∥ ∞ , ∥ c k ∥ ∞ ≤ 2 \lVert b_k\rVert_\infty,\lVert c_k\rVert_\infty\le2 ∥ b k ∥ ∞ , ∥ c k ∥ ∞ ≤ 2 and ∣ d n ∣ ≤ 3 ∣ ∇ u n ∣ |d_n|\le\sqrt{3}|\nabla u_n| ∣ d n ∣ ≤ 3 ∣∇ u n ∣ , the kernel estimate and Young’s inequality give
∥ B n , k , a ( t ) ∥ p ′ ≤ C ( a κ ( q ) + a 2 ) ∥ ∇ u n ( t ) ∥ 2 . \lVert B_{n,k,a}(t)\rVert_{p'}\le C(a^{\kappa(q)}+a^2)\lVert\nabla u_n(t)\rVert_2. ∥ B n , k , a ( t ) ∥ p ′ ≤ C ( a κ ( q ) + a 2 ) ∥ ∇ u n ( t ) ∥ 2 .
The momentum bound in Lemma 4.1 and Cauchy–Schwarz in time now yield
∣ J n , k , a ( η ) ∣ ≤ C ∥ η ∥ ∞ ∥ m n ∥ L t 2 L x p ∥ ∇ u n ∥ L 2 ( Q T ) ( a κ ( q ) + a 2 ) ≤ C T , q ∥ η ∥ ∞ ( a κ ( q ) + a 2 ) . (62) \begin{aligned}
|J_{n,k,a}(\eta)|\le C\lVert\eta\rVert_\infty\lVert m_n\rVert_{L_t^2L_x^p}\lVert\nabla u_n\rVert_{L^2(Q_T)}(a^{\kappa(q)}+a^2) \\
&\le C_{T,q}\lVert\eta\rVert_\infty(a^{\kappa(q)}+a^2). \tag*{(62)}
\end{aligned} ∣ J n , k , a ( η ) ∣ ≤ C ∥ η ∥ ∞ ∥ m n ∥ L t 2 L x p ∥ ∇ u n ∥ L 2 ( Q T ) ( a κ ( q ) + a 2 ) ≤ C T , q ∥ η ∥ ∞ ( a κ ( q ) + a 2 ) . (62)
Step 2: Test admissibility. For the fixed solution, the cubic-pressure bound gives
m n ∈ L 2 ( 0 , T ; L 2 ) , ρ n u n ⊗ u n ∈ L 1 ( 0 , T ; L 3 / 2 ) . m_n\in L^2(0,T;L^2),\qquad\rho_nu_n\otimes u_n\in L^1(0,T;L^{3/2}). m n ∈ L 2 ( 0 , T ; L 2 ) , ρ n u n ⊗ u n ∈ L 1 ( 0 , T ; L 3/2 ) .
Indeed, pair ρ n ∈ L t ∞ L x 3 \rho_n\in L_t^\infty L_x^3 ρ n ∈ L t ∞ L x 3 with u n ∈ L t 2 L x 6 u_n\in L_t^2L_x^6 u n ∈ L t 2 L x 6 for the first bound and with ∣ u n ∣ 2 ∈ L t 1 L x 3 |u_n|^2\in L_t^1L_x^3 ∣ u n ∣ 2 ∈ L t 1 L x 3 for the second. The resulting constants may depend on n n n . Take $V=A_ag. Equation (4.6) implies
{ ∂ t V i = − L i j , a ( g u n , j ) − A a , i h ∈ L t 2 L x 2 , ∇ V = ( L i j , a g ) i j ∈ L t ∞ L x r , 1 < r < ∞ , V ∈ L ∞ ( Q T ) , div V = H a g ∈ L ∞ ( Q T ) , ∥ H a g ∥ ∞ ≤ 4. \left\{
\begin{aligned}
\partial_t V_i &= -L_{ij,a}(g u_{n,j})-A_{a,i}h \in L_t^2L_x^2,\\
\nabla V &= (L_{ij,a}g)_{ij}\in L_t^\infty L_x^r,\qquad1<r<\infty,\\
V &\in L^\infty(Q_T),\\
\operatorname{div}V &= H_ag\in L^\infty(Q_T),\\
\lVert H_ag\rVert_\infty&\le4.
\end{aligned}
\right. ⎩ ⎨ ⎧ ∂ t V i ∇ V V div V ∥ H a g ∥ ∞ = − L ij , a ( g u n , j ) − A a , i h ∈ L t 2 L x 2 , = ( L ij , a g ) ij ∈ L t ∞ L x r , 1 < r < ∞ , ∈ L ∞ ( Q T ) , = H a g ∈ L ∞ ( Q T ) , ≤ 4.
To use η V \eta V η V in the weak momentum equation, smooth V V V in space and time, and then remove these regularizations at fixed n n n , k k k , a a a . The time derivatives converge in L 2 L^2 L 2 and pair with m n ∈ L 2 m_n\in L^2 m n ∈ L 2 . The gradients converge in L 3 L^3 L 3 almost everywhere in time and remain bounded in L t ∞ L x 3 L_t^\infty L_x^3 L t ∞ L x 3 , which permits passage in the convective term. Their L 2 L^2 L 2 convergence handles the stress term, while the uniformly bounded divergences converge almost everywhere and pair with P n ∈ L 1 P_n\in L^1 P n ∈ L 1 . Thus η V \eta V η V is admissible. These smoothing limits are taken before any limit in n n n . The displayed cubic-density norms may depend on n n n ; they justify testing a fixed equation. Only (4.9) and the time-cutoff estimate below are used uniformly along the sequence.
Step 3: The flux remainder. Write V = ∇ ϕ V=\nabla\phi V = ∇ ϕ , where ϕ = Δ T 3 − 1 H a g \phi=\Delta_{\mathbb{T}^3}^{-1}H_ag ϕ = Δ T 3 − 1 H a g . Periodic integration by parts gives
∫ T 3 ∂ j u i , n ∂ i j ϕ d x = − ∫ T 3 u i , n ∂ i Δ ϕ d x = ∫ T 3 d n Δ ϕ d x , \int_{\mathbb{T}^3}\partial_j u_{i,n}\partial_{ij}\phi\,dx
=-\int_{\mathbb{T}^3}u_{i,n}\partial_i\Delta\phi\,dx
=\int_{\mathbb{T}^3}d_n\Delta\phi\,dx, ∫ T 3 ∂ j u i , n ∂ ij ϕ d x = − ∫ T 3 u i , n ∂ i Δ ϕ d x = ∫ T 3 d n Δ ϕ d x ,
∫ T 3 S ( u n ) : ∇ V d x = ν ∫ T 3 d n H a g d x , ν = 2 μ + λ . \int_{\mathbb{T}^3}S(u_n):\nabla V\,dx
=\nu\int_{\mathbb{T}^3}d_nH_ag\,dx,\qquad\nu=2\mu+\lambda. ∫ T 3 S ( u n ) : ∇ V d x = ν ∫ T 3 d n H a g d x , ν = 2 μ + λ .
Moreover, the time and convective terms combine exactly:
m i , n ∂ t V i + ρ n u i , n u j , n ∂ j V i = − m i , n L i j , a ( g u n , j ) − m i , n A a , i h + m i , n u j , n L i j , a g = − m i , n B n , k , a , i . (63) \begin{aligned}
m_{i,n}\partial_tV_i+\rho_nu_{i,n}u_{j,n}\partial_jV_i
&=-m_{i,n}L_{ij,a}(g u_{n,j})-m_{i,n}A_{a,i}h+m_{i,n}u_{j,n}L_{ij,a}g\\
&=-m_{i,n}B_{n,k,a,i}. \tag*{(63)}
\end{aligned} m i , n ∂ t V i + ρ n u i , n u j , n ∂ j V i = − m i , n L ij , a ( g u n , j ) − m i , n A a , i h + m i , n u j , n L ij , a g = − m i , n B n , k , a , i . (63)
Testing the weak momentum equation with η V \eta V η V therefore yields
J n , k , a ( η ) = ∫ Q T η ( P n − ν d n ) H a g d x d t + ∫ Q T η ′ m n ⋅ A a g d x d t . (64) J_{n,k,a}(\eta)=\int_{Q_T}\eta(P_n-\nu d_n)H_ag\,dx\,dt+\int_{Q_T}\eta' m_n\cdot A_ag\,dx\,dt. \tag*{(64)} J n , k , a ( η ) = ∫ Q T η ( P n − ν d n ) H a g d x d t + ∫ Q T η ′ m n ⋅ A a g d x d t . (64)
Finally, Lemma C.2 gives ∥ A a g ∥ ∞ ≤ C a ∥ g ∥ ∞ ≤ C a \lVert A_ag\rVert_\infty\le Ca\lVert g\rVert_\infty\le Ca ∥ A a g ∥ ∞ ≤ C a ∥ g ∥ ∞ ≤ C a . Mass conservation and the kinetic-energy bound imply
∣ ∫ Q T η ′ m n ⋅ A a g d x d t ∣ ≤ C a ∥ η ′ ∥ 1 ess sup 0 < t < T ∥ m n ( t ) ∥ 1 ≤ C T a ∥ η ′ ∥ 1 . \left|\int_{Q_T}\eta' m_n\cdot A_ag\,dx\,dt\right|
\le Ca\lVert\eta'\rVert_1\operatorname*{ess\,sup}_{0<t<T}\lVert m_n(t)\rVert_1
\le CTa\lVert\eta'\rVert_1. ∫ Q T η ′ m n ⋅ A a g d x d t ≤ C a ∥ η ′ ∥ 1 0 < t < T ess sup ∥ m n ( t ) ∥ 1 ≤ C T a ∥ η ′ ∥ 1 .
Since g = b k ( ρ n ) g=b_k(\rho_n) g = b k ( ρ n ) , the first term on the right of (4.11) is R n , k , a \mathcal{R}_{n,k,a} R n , k , a . Combining this identity with (4.9) proves (4.5). □
The bound (4.5) permits the limit n → ∞ n\to\infty n → ∞ at fixed k k k and smoothing scale a a a , followed by a ↓ 0 a\downarrow0 a ↓ 0 . The final limit k → ∞ k\to\infty k → ∞ belongs to the entropy argument. In particular, multiplication by k k k after the flux limit causes no loss: k k k is fixed throughout the first two limits.
Lemma 4.3. After passing to a common subsequence, let Y t , x Y_{t,x} Y t , x be the density Young measure, with ρ = ⟨ z ⟩ \rho= \langle z\rangle ρ = ⟨ z ⟩ , and suppose that u n ⇀ u u_n \rightharpoonup u u n ⇀ u in L 2 H 1 L^2H^1 L 2 H 1 . Set d = div u d = \operatorname{div} u d = div u and
Π = ⟨ z γ ⟩ , β k = ⟨ T k ⟩ , Γ k = ⟨ z γ T k ⟩ , C k = Γ k − Π β k , \Pi= \langle z^\gamma\rangle,\qquad\beta_k = \langle T_k\rangle,\qquad\Gamma_k = \langle z^\gamma T_k\rangle,\qquad C_k = \Gamma_k - \Pi\beta_k, Π = ⟨ z γ ⟩ , β k = ⟨ T k ⟩ , Γ k = ⟨ z γ T k ⟩ , C k = Γ k − Π β k , H k = weak lim n T k ( ρ n ) d n , h k = weak lim n a k ( ρ n ) d n , H_k = \operatorname*{weak\,lim}_{n} T_k(\rho_n)d_n,\qquad h_k = \operatorname*{weak\,lim}_{n} a_k(\rho_n)d_n, H k = n weak lim T k ( ρ n ) d n , h k = n weak lim a k ( ρ n ) d n , where the last two limits are weak in L 2 ( Q T ) L^2(Q_T) L 2 ( Q T ) for fixed k k k . Then C k ≥ 0 C_k \ge0 C k ≥ 0 almost everywhere, and
∫ T 3 C k ( t , x ) d x = ν ∫ T 3 ( H k − β k d ) ( t , x ) d x for almost every t . (65) \int_{\mathbb{T}^3} C_k(t,x)\,\mathrm{d}x = \nu\int_{\mathbb{T}^3}(H_k-\beta_kd)(t,x)\,\mathrm{d}x \qquad\text{for almost every }t. \tag*{(65)} ∫ T 3 C k ( t , x ) d x = ν ∫ T 3 ( H k − β k d ) ( t , x ) d x for almost every t . (65) Moreover,
T k ( ρ n ) → β k in C ( [ 0 , T ] ; H − 1 ( T 3 ) ) , T_k(\rho_n)\to\beta_k\quad\text{in }C([0,T];H^{-1}(\mathbb{T}^3)), T k ( ρ n ) → β k in C ([ 0 , T ] ; H − 1 ( T 3 )) , and
∂ t β k + div ( β k u ) + h k = 0 , β k ( 0 ) = T k ( ρ 0 ) . (66) \partial_t\beta_k+\operatorname{div}(\beta_k u)+h_k=0,\qquad\beta_k(0)=T_k(\rho_0). \tag*{(66)} ∂ t β k + div ( β k u ) + h k = 0 , β k ( 0 ) = T k ( ρ 0 ) . (66)
Proof. Extract a common subsequence on integer time intervals and countably many truncation levels. The uniform density bounds give a probability-valued Young measure Y t , x Y_{t,x} Y t , x , see [3 ] . Here and below ⟨ F ⟩ = ∫ [ 0 , ∞ ) F ( z ) d Y t , x ( z ) \langle F\rangle=\int_{[0,\infty)}F(z)\,\mathrm{d}Y_{t,x}(z) ⟨ F ⟩ = ∫ [ 0 , ∞ ) F ( z ) d Y t , x ( z ) . For bounded continuous F F F , the fundamental theorem identifies F ( ρ n ) F(\rho_n) F ( ρ n ) with this average in the weak limit. For unbounded F F F , one must additionally control its tails; a Young measure alone does not account for a possible concentration of F ( ρ n ) F(\rho_n) F ( ρ n ) . Set
B = sup n ess sup 0 < t < T ∫ T 3 ρ n γ d x . B=\sup_n\operatorname*{ess\,sup}_{0<t<T}\int_{\mathbb{T}^3}\rho_n^\gamma\,\mathrm{d}x. B = n sup 0 < t < T ess sup ∫ T 3 ρ n γ d x .
For a measurable I ⊂ ( 0 , T ) I\subset(0,T) I ⊂ ( 0 , T ) , apply Young-measure convergence to the bounded truncations F L ( z ) = min { z γ , L } F_L(z)=\min\{z^\gamma,L\} F L ( z ) = min { z γ , L } and G R ( z ) = min { z , R } G_R(z)=\min\{z,R\} G R ( z ) = min { z , R } . Passing to the limits L , R → ∞ L,R\to\infty L , R → ∞ gives
ess sup 0 < t < T ∫ T 3 ⟨ z γ ⟩ d x ≤ B , ∫ T 3 ⟨ z ⟩ ( t , x ) d x = m for almost every t . (67) \operatorname*{ess\,sup}_{0<t<T}\int_{\mathbb{T}^3}\langle z^\gamma\rangle\,\mathrm{d}x\le B,\qquad\int_{\mathbb{T}^3}\langle z\rangle(t,x)\,\mathrm{d}x=m\quad\text{for almost every }t. \tag*{(67)} 0 < t < T ess sup ∫ T 3 ⟨ z γ ⟩ d x ≤ B , ∫ T 3 ⟨ z ⟩ ( t , x ) d x = m for almost every t . (67)
In particular, Jensen’s inequality yields ess sup t ∫ T 3 ρ γ ≤ B \operatorname*{ess\,sup}_{t}\int_{\mathbb{T}^3}\rho^\gamma\le B ess sup t ∫ T 3 ρ γ ≤ B . The pressure tail estimate allows us to pass from bounded truncations to z γ z^\gamma z γ and z γ T k ( z ) z^\gamma T_k(z) z γ T k ( z ) , since for fixed k k k ,
sup n ∫ { ρ n > R } ( ρ n γ + ρ n γ T k ( ρ n ) ) ≤ ( 1 + 2 k ) C T R − 1 / 6 ⟶ 0. \sup_n\int_{\{\rho_n>R\}}\left(\rho_n^\gamma+\rho_n^\gamma T_k(\rho_n)\right)\le(1+2k)C_T R^{-1/6}\longrightarrow0. n sup ∫ { ρ n > R } ( ρ n γ + ρ n γ T k ( ρ n ) ) ≤ ( 1 + 2 k ) C T R − 1/6 ⟶ 0.
Together with the L s L^s L s , s > 1 s>1 s > 1 , pressure bound, this excludes a concentration measure in both limits. There is no corresponding claim for the untruncated product ρ n γ + 1 \rho_n^{\gamma+1} ρ n γ + 1 . Moreover, Lemma 4.1 gives
∥ e n ρ n 3 T k ( ρ n ) ∥ L 1 ( Q T ) ≤ 2 k ∥ e n ρ n 3 ∥ L 1 ( Q T ) ⟶ 0. \lVert e_n\rho_n^3T_k(\rho_n)\rVert_{L^1(Q_T)}\le2k\lVert e_n\rho_n^3\rVert_{L^1(Q_T)}\longrightarrow0. ∥ e n ρ n 3 T k ( ρ n ) ∥ L 1 ( Q T ) ≤ 2 k ∥ e n ρ n 3 ∥ L 1 ( Q T ) ⟶ 0.
Consequently,
P n ⇀ Π , P n T k ( ρ n ) ⇀ Γ k weakly in L 1 ( Q T ) . P_n\rightharpoonup\Pi,\qquad P_nT_k(\rho_n)\rightharpoonup\Gamma_k\quad\text{weakly in }L^1(Q_T). P n ⇀ Π , P n T k ( ρ n ) ⇀ Γ k weakly in L 1 ( Q T ) .
Since T k T_k T k and a k a_k a k are bounded by 2 k 2k 2 k , the products T k ( ρ n ) d n T_k(\rho_n)d_n T k ( ρ n ) d n and a k ( ρ n ) d n a_k(\rho_n)d_n a k ( ρ n ) d n admit the weak L 2 L^2 L 2 limits H k H_k H k and h k h_k h k . By symmetry,
C k = 1 2 ∬ ( z γ − w γ ) ( T k ( z ) − T k ( w ) ) d Y ( z ) d Y ( w ) ≥ 0 , C_k = \frac{1}{2}\iint(z^\gamma-w^\gamma)(T_k(z)-T_k(w))\,\mathrm{d}Y(z)\,\mathrm{d}Y(w) \ge0, C k = 2 1 ∬ ( z γ − w γ ) ( T k ( z ) − T k ( w )) d Y ( z ) d Y ( w ) ≥ 0 ,
where the inequality follows from the monotonicity of z γ z^\gamma z γ and T k T_k T k .
For fixed k k k , the renormalized continuity equation gives
∂ t T k ( ρ n ) + div ( T k ( ρ n ) u n ) + a k ( ρ n ) d n = 0. \partial_t T_k(\rho_n)+\operatorname{div}(T_k(\rho_n)u_n)+a_k(\rho_n)d_n=0. ∂ t T k ( ρ n ) + div ( T k ( ρ n ) u n ) + a k ( ρ n ) d n = 0.
Hence
∥ ∂ t T k ( ρ n ) ∥ L 2 H − 1 ≤ 2 k ( ∥ u n ∥ L 2 L 2 + ∥ d n ∥ 2 ) ≤ C T k , \|\partial_t T_k(\rho_n)\|_{L^2H^{-1}}\le2k\left(\|u_n\|_{L^2L^2}+\|d_n\|_2\right)\le C_Tk, ∥ ∂ t T k ( ρ n ) ∥ L 2 H − 1 ≤ 2 k ( ∥ u n ∥ L 2 L 2 + ∥ d n ∥ 2 ) ≤ C T k ,
∥ T k ( ρ n ) ( t ) − T k ( ρ n ) ( s ) ∥ H − 1 ≤ C T k ∣ t − s ∣ 1 / 2 . \|T_k(\rho_n)(t)-T_k(\rho_n)(s)\|_{H^{-1}}\le C_Tk|t-s|^{1/2}. ∥ T k ( ρ n ) ( t ) − T k ( ρ n ) ( s ) ∥ H − 1 ≤ C T k ∣ t − s ∣ 1/2 .
The compact embedding L 2 ⋐ H − 1 L^2\Subset H^{-1} L 2 ⋐ H − 1 and the time equicontinuity yield compactness in C ( [ 0 , T ] ; H − 1 ) C([0,T];H^{-1}) C ([ 0 , T ] ; H − 1 ) , by the compactness argument recalled in Lemma 2.4 ; compare [39 ] . The initial data satisfy
∥ T k ( ρ 0 n ) − T k ( ρ 0 ) ∥ 2 2 ≤ 2 k ∥ ρ 0 n − ρ 0 ∥ 1 ⟶ 0. \|T_k(\rho_{0n})-T_k(\rho_0)\|_2^2\le2k\|\rho_{0n}-\rho_0\|_1\longrightarrow0. ∥ T k ( ρ 0 n ) − T k ( ρ 0 ) ∥ 2 2 ≤ 2 k ∥ ρ 0 n − ρ 0 ∥ 1 ⟶ 0.
Thus
T k ( ρ n ) ⟶ β k in C ( [ 0 , T ] ; H − 1 ) , β k ( 0 ) = T k ( ρ 0 ) . T_k(\rho_n)\longrightarrow\beta_k\quad\text{in }C([0,T];H^{-1}),\qquad\beta_k(0)=T_k(\rho_0). T k ( ρ n ) ⟶ β k in C ([ 0 , T ] ; H − 1 ) , β k ( 0 ) = T k ( ρ 0 ) .
For every smooth vector field φ \varphi φ ,
∣ ∫ 0 T ⟨ T k ( ρ n ) − β k , φ ⋅ u n ⟩ d t ∣ ≤ C φ ∥ T k ( ρ n ) − β k ∥ C t H − 1 ∥ u n ∥ L 2 H 1 ⟶ 0. \left|\int_0^T\langle T_k(\rho_n)-\beta_k,\varphi\cdot u_n\rangle\,\mathrm{d}t\right|
\le C_\varphi\|T_k(\rho_n)-\beta_k\|_{C_tH^{-1}}\|u_n\|_{L^2H^1}\longrightarrow0. ∫ 0 T ⟨ T k ( ρ n ) − β k , φ ⋅ u n ⟩ d t ≤ C φ ∥ T k ( ρ n ) − β k ∥ C t H − 1 ∥ u n ∥ L 2 H 1 ⟶ 0.
Together with u n ⇀ u u_n\rightharpoonup u u n ⇀ u , this identifies the flux limit as β k u \beta_ku β k u . Passing to the limit in the initialized truncated equation proves (66) .
Fix k k k and a > 0 a>0 a > 0 . The preceding strong convergence gives
∥ S a [ T k ( ρ n ) − β k ] ∥ C t , x ≤ ∥ χ a p e r ∥ H 1 ∥ T k ( ρ n ) − β k ∥ C t H − 1 ⟶ 0. \|S_a[T_k(\rho_n)-\beta_k]\|_{C_{t,x}}
\le\|\chi_a^{\mathrm{per}}\|_{H^1}\|T_k(\rho_n)-\beta_k\|_{C_tH^{-1}}\longrightarrow0. ∥ S a [ T k ( ρ n ) − β k ] ∥ C t , x ≤ ∥ χ a per ∥ H 1 ∥ T k ( ρ n ) − β k ∥ C t H − 1 ⟶ 0.
Taking n → ∞ n\to\infty n → ∞ in the remainder of Lemma 4.2 yields
lim n R n , k , a = 1 k ∫ Q T η [ Γ k − ν H k − ( Π − ν d ) S a β k ] . \lim_{n}\mathcal{R}_{n,k,a}
=\frac{1}{k}\int_{Q_T}\eta\left[\Gamma_k-\nu H_k-(\Pi-\nu d)S_a\beta_k\right]. n lim R n , k , a = k 1 ∫ Q T η [ Γ k − ν H k − ( Π − ν d ) S a β k ] .
We next let a ↓ 0 a\downarrow0 a ↓ 0 , keeping k k k fixed. Since S a β k → β k S_a\beta_k\to\beta_k S a β k → β k in L 1 ( Q T ) L^1(Q_T) L 1 ( Q T ) and ∣ S a β k − β k ∣ ≤ 4 k |S_a\beta_k-\beta_k|\le4k ∣ S a β k − β k ∣ ≤ 4 k , for every F ∈ L 1 ( Q T ) F\in L^1(Q_T) F ∈ L 1 ( Q T ) and N > 0 N>0 N > 0 ,
∫ Q T ∣ F ∣ ∣ S a β k − β k ∣ ≤ 4 k ∫ { ∣ F ∣ > N } ∣ F ∣ + N ∥ S a β k − β k ∥ 1 . \int_{Q_T}|F|\,|S_a\beta_k-\beta_k|
\le4k\int_{\{|F|>N\}}|F|+N\|S_a\beta_k-\beta_k\|_1. ∫ Q T ∣ F ∣ ∣ S a β k − β k ∣ ≤ 4 k ∫ { ∣ F ∣ > N } ∣ F ∣ + N ∥ S a β k − β k ∥ 1 .
Sending first a ↓ 0 a\downarrow0 a ↓ 0 and then N → ∞ N\to\infty N → ∞ gives
∫ Q T F ( S a β k − β k ) ⟶ 0. \int_{Q_T}F(S_a\beta_k-\beta_k)\longrightarrow0. ∫ Q T F ( S a β k − β k ) ⟶ 0.
By (58) , the high-frequency remainder tends to zero uniformly in n n n and k k k . Therefore, after multiplying the limiting identity by k k k ,
∫ 0 T η ( t ) ∫ T 3 [ C k − ν ( H k − β k d ) ] d x d t = 0. \int_0^T \eta(t)\int_{\mathbb{T}^3}\left[C_k-\nu(H_k-\beta_k d)\right]\,dx\,dt=0. ∫ 0 T η ( t ) ∫ T 3 [ C k − ν ( H k − β k d ) ] d x d t = 0.
As η ∈ C c ∞ ( 0 , T ) \eta\in C^\infty_c(0,T) η ∈ C c ∞ ( 0 , T ) is arbitrary, this proves (65) . This is an identity of spatial integrals for almost every time. No spatially pointwise equality C k = ν ( H k − β k d ) C_k=\nu(H_k-\beta_kd) C k = ν ( H k − β k d ) is inferred. The distinction is harmless for the subsequent entropy argument because its transport divergence also vanishes after integration over the torus. The covariance C k ≥ 0 C_k\geq0 C k ≥ 0 , in contrast, is a pointwise consequence of the one-variable monotonicity calculation above. □
Lemma 4.4. For every T > 0 T>0 T > 0 , there exists C T C_T C T independent of k k k such that
lim n ∫ Q T ∣ T k ( ρ n ) − T k ( ρ ) ∣ γ + 1 + lim n ∫ Q T ∣ T k ( ρ n ) − β k ∣ γ + 1 ≤ C T . (68) \lim_{n}\int_{Q_T}|T_k(\rho_n)-T_k(\rho)|^{\gamma+1}+\lim_{n}\int_{Q_T}|T_k(\rho_n)-\beta_k|^{\gamma+1}\leq C_T. \tag*{(68)} n lim ∫ Q T ∣ T k ( ρ n ) − T k ( ρ ) ∣ γ + 1 + n lim ∫ Q T ∣ T k ( ρ n ) − β k ∣ γ + 1 ≤ C T . (68) Moreover, for L ≥ 0 L\geq0 L ≥ 0 and k > L k>L k > L ,
∫ Q T ∩ { β k ≤ L } ∣ h k ∣ ≤ C T k ( k − L ) − ( γ + 1 ) / 2 . (69) \int_{Q_T\cap\{\beta_k\leq L\}}|h_k|\leq C_T k(k-L)^{-(\gamma+1)/2}. \tag*{(69)} ∫ Q T ∩ { β k ≤ L } ∣ h k ∣ ≤ C T k ( k − L ) − ( γ + 1 ) /2 . (69) The mass flux converges to ρ u \rho u ρ u , including the initial terms. The limiting density satisfies the continuity and renormalized continuity equations with the prescribed initial data, for every b ∈ C 1 ( [ 0 , ∞ ) ) b\in C^1([0,\infty)) b ∈ C 1 ([ 0 , ∞ )) with compactly supported derivative.
Proof. We use the nonnegative covariance C k C_k C k and the averaged flux identity (65) . In (66) , h k h_k h k is only the weak limit of the approximate sources. In particular, neither h k = a k ( ρ ) d h_k=a_k(\rho)d h k = a k ( ρ ) d nor β k = T k ( ρ ) \beta_k=T_k(\rho) β k = T k ( ρ ) is available at this stage. The oscillation estimate first gives decay of h k h_k h k on bounded sublevel sets of β k \beta_k β k ; this is exactly where a renormalization with boundedly supported derivative will sample the source.
Step 1: Uniform oscillation control. Let Y = Y t , x Y=Y_{t,x} Y = Y t , x be the Young measure from Lemma 4.3. For a , b ≥ 0 a,b\geq0 a , b ≥ 0 , pressure monotonicity and the 1-Lipschitz property of T k T_k T k give
( a γ − b γ ) ( T k ( a ) − T k ( b ) ) ≥ ∣ T k ( a ) − T k ( b ) ∣ γ + 1 . (a^\gamma-b^\gamma)(T_k(a)-T_k(b))\geq|T_k(a)-T_k(b)|^{\gamma+1}. ( a γ − b γ ) ( T k ( a ) − T k ( b )) ≥ ∣ T k ( a ) − T k ( b ) ∣ γ + 1 .
Since T k T_k T k is concave, Jensen’s inequality gives β k ≤ T k ( ρ ) \beta_k\leq T_k(\rho) β k ≤ T k ( ρ ) and Π ≥ ρ γ \Pi\geq\rho^\gamma Π ≥ ρ γ . Taking b = ρ b=\rho b = ρ and averaging with respect to Y Y Y , we obtain
⟨ ∣ T k ( z ) − T k ( ρ ) ∣ γ + 1 ⟩ ≤ C k + ( Π − ρ γ ) ( β k − T k ( ρ ) ) ≤ C k . \langle|T_k(z)-T_k(\rho)|^{\gamma+1}\rangle\leq C_k+(\Pi-\rho^\gamma)(\beta_k-T_k(\rho))
\leq C_k. ⟨ ∣ T k ( z ) − T k ( ρ ) ∣ γ + 1 ⟩ ≤ C k + ( Π − ρ γ ) ( β k − T k ( ρ )) ≤ C k .
Set
X k = ∫ Q T ⟨ ∣ T k ( z ) − T k ( ρ ) ∣ γ + 1 ⟩ , V k = ∫ Q T ⟨ ∣ T k ( z ) − β k ∣ 2 ⟩ . X_k=\int_{Q_T}\langle|T_k(z)-T_k(\rho)|^{\gamma+1}\rangle,\qquad V_k=\int_{Q_T}\langle|T_k(z)-\beta_k|^2\rangle. X k = ∫ Q T ⟨ ∣ T k ( z ) − T k ( ρ ) ∣ γ + 1 ⟩ , V k = ∫ Q T ⟨ ∣ T k ( z ) − β k ∣ 2 ⟩ .
As β k = ⟨ T k ( z ) ⟩ \beta_k=\langle T_k(z)\rangle β k = ⟨ T k ( z )⟩ , Hölder’s inequality implies
X k ≤ ∫ Q T C k , V k ≤ ∣ Q T ∣ 1 − 2 / ( γ + 1 ) X k 2 / ( γ + 1 ) . X_k\leq\int_{Q_T}C_k,\qquad V_k\leq|Q_T|^{1-2/(\gamma+1)}X_k^{2/(\gamma+1)}. X k ≤ ∫ Q T C k , V k ≤ ∣ Q T ∣ 1 − 2/ ( γ + 1 ) X k 2/ ( γ + 1 ) .
Writing D T = sup n ∥ div u n ∥ L 2 ( Q T ) D_T=\sup_n\|\operatorname{div}u_n\|_{L^2(Q_T)} D T = sup n ∥ div u n ∥ L 2 ( Q T ) , the averaged flux identity and Cauchy–Schwarz yield
X k ≤ ν lim n ∫ Q T ( T k ( ρ n ) − β k ) div u n ≤ ν D T V k 1 / 2 ≤ ν D T ∣ Q T ∣ 1 / 2 − 1 / ( γ + 1 ) X k 1 / ( γ + 1 ) . \begin{aligned}
X_k &\le\nu\lim_{n}\int_{Q_T}(T_k(\rho_n)-\beta_k)\operatorname{div}u_n\\
&\le\nu D_T V_k^{1/2}\\
&\le\nu D_T|Q_T|^{1/2-1/(\gamma+1)}X_k^{1/(\gamma+1)}.
\end{aligned} X k ≤ ν n lim ∫ Q T ( T k ( ρ n ) − β k ) div u n ≤ ν D T V k 1/2 ≤ ν D T ∣ Q T ∣ 1/2 − 1/ ( γ + 1 ) X k 1/ ( γ + 1 ) .
Thus, including the case X k = 0 X_k=0 X k = 0 ,
X k ≤ ( ν D T ∣ Q T ∣ 1 / 2 − 1 / ( γ + 1 ) ) ( γ + 1 ) / γ . X_k\le\left(\nu D_T|Q_T|^{1/2-1/(\gamma+1)}\right)^{(\gamma+1)/\gamma}. X k ≤ ( ν D T ∣ Q T ∣ 1/2 − 1/ ( γ + 1 ) ) ( γ + 1 ) / γ .
Jensen’s and the triangle inequalities also give
∣ β k − T k ( ρ ) ∣ γ + 1 ≤ ⟨ ∣ T k ( z ) − T k ( ρ ) ∣ γ + 1 ⟩ , |\beta_k-T_k(\rho)|^{\gamma+1}\le\langle|T_k(z)-T_k(\rho)|^{\gamma+1}\rangle, ∣ β k − T k ( ρ ) ∣ γ + 1 ≤ ⟨ ∣ T k ( z ) − T k ( ρ ) ∣ γ + 1 ⟩ ,
∫ Q T ⟨ ∣ T k ( z ) − β k ∣ γ + 1 ⟩ ≤ 2 γ + 1 X k . \int_{Q_T}\langle|T_k(z)-\beta_k|^{\gamma+1}\rangle\le2^{\gamma+1}X_k. ∫ Q T ⟨ ∣ T k ( z ) − β k ∣ γ + 1 ⟩ ≤ 2 γ + 1 X k .
For fixed k k k , the corresponding Young-measure limits hold with the measurable centers T k ( ρ ) T_k(\rho) T k ( ρ ) and β k \beta_k β k : approximate these bounded centers by simple functions and use Lipschitz continuity of the powers on [ 0 , 2 k ] [0,2k] [ 0 , 2 k ] . An initial time layer of width ε \varepsilon ε contributes at most C ( k ) ε C(k)\varepsilon C ( k ) ε . Sending ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 proves (4.15), uniformly in k k k .
Step 2: Decay of the truncated source. Fix L ≥ 0 L\ge0 L ≥ 0 , k > L k>L k > L , and let F = Q T ∩ { β k ≤ L } F=Q_T\cap\{\beta_k\le L\} F = Q T ∩ { β k ≤ L } . On F ∩ { ρ n > k } F\cap\{\rho_n>k\} F ∩ { ρ n > k } , the truncation properties imply
∣ a k ( ρ n ) ∣ ≤ T k ( ρ n ) ≤ k k − L ( T k ( ρ n ) − β k ) , T k ( ρ n ) − β k ≥ k − L . |a_k(\rho_n)|\le T_k(\rho_n)\le\frac{k}{k-L}(T_k(\rho_n)-\beta_k),\qquad T_k(\rho_n)-\beta_k\ge k-L. ∣ a k ( ρ n ) ∣ ≤ T k ( ρ n ) ≤ k − L k ( T k ( ρ n ) − β k ) , T k ( ρ n ) − β k ≥ k − L .
Since a k = 0 a_k=0 a k = 0 on [ 0 , k ] [0,k] [ 0 , k ] and γ + 1 > 2 \gamma+1>2 γ + 1 > 2 , we have
1 F ∣ a k ( ρ n ) ∣ 2 ≤ k 2 ( k − L ) γ + 1 ∣ T k ( ρ n ) − β k ∣ γ + 1 . \mathbf{1}_F|a_k(\rho_n)|^2\le\frac{k^2}{(k-L)^{\gamma+1}}|T_k(\rho_n)-\beta_k|^{\gamma+1}. 1 F ∣ a k ( ρ n ) ∣ 2 ≤ ( k − L ) γ + 1 k 2 ∣ T k ( ρ n ) − β k ∣ γ + 1 .
Recall that a k ( ρ n ) div u n ⇀ h k a_k(\rho_n)\operatorname{div}u_n\rightharpoonup h_k a k ( ρ n ) div u n ⇀ h k in L 2 L^2 L 2 for fixed k k k . Testing this convergence with 1 F sign h k \mathbf{1}_F\operatorname{sign}h_k 1 F sign h k , using Cauchy–Schwarz, and applying (4.15), we obtain
∫ F ∣ h k ∣ = lim n ∫ F a k ( ρ n ) div u n sign h k ≤ D T lim sup n ∥ 1 F a k ( ρ n ) ∥ 2 ≤ C T k ( k − L ) − ( γ + 1 ) / 2 . \begin{aligned}
\int_F|h_k|&=\lim_n\int_F a_k(\rho_n)\operatorname{div}u_n\operatorname{sign}h_k\\
&\le D_T\limsup_n\|\mathbf{1}_F a_k(\rho_n)\|_2\\
&\le C_Tk(k-L)^{-(\gamma+1)/2}.
\end{aligned} ∫ F ∣ h k ∣ = n lim ∫ F a k ( ρ n ) div u n sign h k ≤ D T n lim sup ∥ 1 F a k ( ρ n ) ∥ 2 ≤ C T k ( k − L ) − ( γ + 1 ) /2 .
This is (4.16). In particular, for k ≥ 2 max { 1 , L } k\ge2\max\{1,L\} k ≥ 2 max { 1 , L } ,
∥ 1 { β k ≤ L } h k ∥ L 1 ( Q T ) ≤ C T , L k − ( γ − 1 ) / 2 ⟶ 0. \|\mathbf{1}_{\{\beta_k\le L\}}h_k\|_{L^1(Q_T)}\le C_{T,L}k^{-(\gamma-1)/2}\longrightarrow0. ∥ 1 { β k ≤ L } h k ∥ L 1 ( Q T ) ≤ C T , L k − ( γ − 1 ) /2 ⟶ 0.
Hence b ′ ( β k ) h k → 0 b'(\beta_k)h_k\to0 b ′ ( β k ) h k → 0 in L 1 ( Q T ) L^1(Q_T) L 1 ( Q T ) whenever b ′ = 0 b'=0 b ′ = 0 on [ L , ∞ ) [L,\infty) [ L , ∞ ) .
Step 3: Mass flux and initialized renormalization. First recover the kinetic bound for the limit. Put
B = sup n ess sup t ∫ T 3 ρ n γ , K k i n = sup n ess sup t ∫ T 3 ρ n ∣ u n ∣ 2 . B=\sup_n\operatorname*{ess\,sup}_t\int_{\mathbb{T}^3}\rho_n^\gamma,\qquad
K_{\mathrm{kin}}=\sup_n\operatorname*{ess\,sup}_t\int_{\mathbb{T}^3}\rho_n|u_n|^2. B = n sup t ess sup ∫ T 3 ρ n γ , K kin = n sup t ess sup ∫ T 3 ρ n ∣ u n ∣ 2 .
For each fixed j j j , Lemma 4.3 gives T j ( ρ n ) u n ⇀ β j u T_j(\rho_n)u_n \rightharpoonup\beta_j u T j ( ρ n ) u n ⇀ β j u , while
∫ T 3 T j ( ρ n ) ∣ u n ∣ 2 ≤ K k i n , ∫ T 3 ∣ T j ( ρ n ) u n ∣ 2 ≤ 2 j K k i n . \int_{\mathbb{T}^3} T_j(\rho_n)|u_n|^2 \le K_{\mathrm{kin}}, \qquad\int_{\mathbb{T}^3} |T_j(\rho_n)u_n|^2 \le2jK_{\mathrm{kin}}. ∫ T 3 T j ( ρ n ) ∣ u n ∣ 2 ≤ K kin , ∫ T 3 ∣ T j ( ρ n ) u n ∣ 2 ≤ 2 j K kin .
For completeness, weighted lower semicontinuity follows from the convex perspective representation
∣ w ∣ 2 r = sup ξ ∈ R 3 { 2 w ⋅ ξ − r ∣ ξ ∣ 2 } , r > 0 , \frac{|w|^2}{r}=\sup_{\xi\in\mathbb{R}^3}\{2w\cdot\xi-r|\xi|^2\}, \qquad r>0, r ∣ w ∣ 2 = ξ ∈ R 3 sup { 2 w ⋅ ξ − r ∣ ξ ∣ 2 } , r > 0 ,
extended by 0 0 0 at ( r , w ) = ( 0 , 0 ) (r,w)=(0,0) ( r , w ) = ( 0 , 0 ) and by + ∞ +\infty + ∞ at r = 0 r=0 r = 0 , w ≠ 0 w\ne0 w = 0 . Apply this formula with r = T j ( ρ n ) r=T_j(\rho_n) r = T j ( ρ n ) and w = T j ( ρ n ) u n w=T_j(\rho_n)u_n w = T j ( ρ n ) u n . The joint weak limits are β j \beta_j β j and β j u \beta_j u β j u ; testing the supremum with smooth vector fields and then taking the supremum proves the lower semicontinuity. On any measurable time set I ⊂ ( 0 , T ) I\subset(0,T) I ⊂ ( 0 , T ) , it yields
∫ I ∫ T 3 β j ∣ u ∣ 2 ≤ lim inf n ∫ I ∫ T 3 T j ( ρ n ) ∣ u n ∣ 2 ≤ K k i n ∣ I ∣ . \int_I\int_{\mathbb{T}^3}\beta_j|u|^2\le\liminf_{n}\int_I\int_{\mathbb{T}^3}T_j(\rho_n)|u_n|^2\le K_{\mathrm{kin}}|I|. ∫ I ∫ T 3 β j ∣ u ∣ 2 ≤ n lim inf ∫ I ∫ T 3 T j ( ρ n ) ∣ u n ∣ 2 ≤ K kin ∣ I ∣.
By concavity of the truncation, T j ( z ) = j T ( z / j ) ↑ z T_j(z)=jT(z/j)\uparrow z T j ( z ) = j T ( z / j ) ↑ z as j → ∞ j\to\infty j → ∞ ; indeed,
∂ j [ j T ( z / j ) ] = T ( z / j ) − ( z / j ) T ′ ( z / j ) ≥ 0. \partial_j[jT(z/j)]=T(z/j)-(z/j)T'(z/j)\ge0. ∂ j [ j T ( z / j )] = T ( z / j ) − ( z / j ) T ′ ( z / j ) ≥ 0.
Thus β j ↑ ρ \beta_j\uparrow\rho β j ↑ ρ , and monotone convergence gives
ess sup t ∫ T 3 ρ ∣ u ∣ 2 ≤ K k i n . (70) \operatorname*{ess\,sup}_{t}\int_{\mathbb{T}^3}\rho|u|^2\le K_{\mathrm{kin}}. \tag*{(70)} t ess sup ∫ T 3 ρ ∣ u ∣ 2 ≤ K kin . (70)
The time-slice density moments also imply, for almost every t t t ,
∫ T 3 ( ρ n − T j ( ρ n ) ) ≤ B j 1 − γ , ∫ T 3 ( ρ − β j ) ≤ B j 1 − γ . \int_{\mathbb{T}^3}(\rho_n-T_j(\rho_n))\le Bj^{1-\gamma}, \qquad\int_{\mathbb{T}^3}(\rho-\beta_j)\le Bj^{1-\gamma}. ∫ T 3 ( ρ n − T j ( ρ n )) ≤ B j 1 − γ , ∫ T 3 ( ρ − β j ) ≤ B j 1 − γ .
Weighted Cauchy–Schwarz now gives
∥ ( ρ n − T j ( ρ n ) ) u n ∥ 1 + ∥ ( ρ − β j ) u ∥ 1 ≤ 2 K k i n B j − ( γ − 1 ) / 2 . \|(\rho_n-T_j(\rho_n))u_n\|_1+\|(\rho-\beta_j)u\|_1
\le2\sqrt{K_{\mathrm{kin}}B}\,j^{-(\gamma-1)/2}. ∥ ( ρ n − T j ( ρ n )) u n ∥ 1 + ∥ ( ρ − β j ) u ∥ 1 ≤ 2 K kin B j − ( γ − 1 ) /2 .
Let n → ∞ n\to\infty n → ∞ at fixed j j j in the truncated mass flux, and then let j → ∞ j\to\infty j → ∞ in the preceding tail bound. For any smooth bounded vector field φ \varphi φ on [ 0 , T ] × T 3 [0,T]\times\mathbb{T}^3 [ 0 , T ] × T 3 , this gives
lim n ∫ Q T m n ⋅ φ = lim j → ∞ ∫ Q T β j u ⋅ φ = ∫ Q T ρ u ⋅ φ . \lim_{n}\int_{Q_T}m_n\cdot\varphi=\lim_{j\to\infty}\int_{Q_T}\beta_j u\cdot\varphi=\int_{Q_T}\rho u\cdot\varphi. n lim ∫ Q T m n ⋅ φ = j → ∞ lim ∫ Q T β j u ⋅ φ = ∫ Q T ρ u ⋅ φ .
Hence the weak momentum limit is m = ρ u m=\rho u m = ρ u . We now pass to the limit in the original initialized, untruncated continuity equation, using ρ 0 n → ρ 0 \rho_{0n}\to\rho_0 ρ 0 n → ρ 0 in L 1 L^1 L 1 . This avoids any assertion of global decay for h j h_j h j and recovers
∂ t ρ + div ( ρ u ) = 0 , ρ ( 0 ) = ρ 0 . \partial_t\rho+\operatorname{div}(\rho u)=0,\qquad\rho(0)=\rho_0. ∂ t ρ + div ( ρ u ) = 0 , ρ ( 0 ) = ρ 0 .
Moreover, β j ↑ ρ ∈ L γ ( Q T ) \beta_j\uparrow\rho\in L^\gamma(Q_T) β j ↑ ρ ∈ L γ ( Q T ) implies β j → ρ \beta_j\to\rho β j → ρ in L r ( Q T ) L^r(Q_T) L r ( Q T ) for 1 ≤ r ≤ γ 1\le r\le\gamma 1 ≤ r ≤ γ . Set r m = 2 γ / ( γ + 1 ) r_m=2\gamma/(\gamma+1) r m = 2 γ / ( γ + 1 ) . The density bound and (70) give m ∈ L ∞ ( 0 , T ; L r m ) m\in L^\infty(0,T;L^{r_m}) m ∈ L ∞ ( 0 , T ; L r m ) . The initialized continuity equation yields ρ ∈ C ( [ 0 , T ] ; H − 4 ) \rho\in C([0,T];H^{-4}) ρ ∈ C ([ 0 , T ] ; H − 4 ) , which, together with the uniform L γ L^\gamma L γ bound, implies
ρ ∈ C w ( [ 0 , T ] ; L γ ( T 3 ) ) , ρ ( 0 ) = ρ 0 . \rho\in C_w([0,T];L^\gamma(\mathbb{T}^3)),\qquad\rho(0)=\rho_0. ρ ∈ C w ([ 0 , T ] ; L γ ( T 3 )) , ρ ( 0 ) = ρ 0 .
For b ∈ C 1 ( [ 0 , ∞ ) ) b\in C^1([0,\infty)) b ∈ C 1 ([ 0 , ∞ )) with compactly supported derivative, Lemma 2.6 applied to (66) gives the source version of the DiPerna–Lions chain rule [11 ] ,
∂ t b ( β k ) + div ( b ( β k ) u ) + ( β k b ′ ( β k ) − b ( β k ) ) div u + b ′ ( β k ) h k = 0. \partial_t b(\beta_k) + \operatorname{div}(b(\beta_k)u) + \left(\beta_k b'(\beta_k) - b(\beta_k)\right)\operatorname{div} u + b'(\beta_k)h_k = 0. ∂ t b ( β k ) + div ( b ( β k ) u ) + ( β k b ′ ( β k ) − b ( β k ) ) div u + b ′ ( β k ) h k = 0.
As k → ∞ k \to\infty k → ∞ , bounded convergence handles the transport, divergence, and initial terms, while (69) gives b ′ ( β k ) h k → 0 b'(\beta_k)h_k \to0 b ′ ( β k ) h k → 0 in L 1 ( Q T ) L^1(Q_T) L 1 ( Q T ) . Thus ρ \rho ρ satisfies the renormalized continuity equation with initial data ρ 0 \rho_0 ρ 0 . The case b ( 0 ) ≠ 0 b(0) \ne0 b ( 0 ) = 0 follows from div ( b ( 0 ) u ) − b ( 0 ) div u = 0 \operatorname{div}(b(0)u) - b(0)\operatorname{div} u = 0 div ( b ( 0 ) u ) − b ( 0 ) div u = 0 . □
Momentum compactness before the entropy comparison. The following estimates use the pressure-derived momentum bound and the initialized momentum equation. They are available before strong density convergence; the identification m = ρ u m = \rho u m = ρ u was obtained by the truncation argument in Lemma 4.4.
The approximate energy bound gives m n m_n m n bounded in L ∞ L r m L^\infty L^{r_m} L ∞ L r m , where r m = 2 γ / ( γ + 1 ) > 1 r_m = 2\gamma/(\gamma+ 1) > 1 r m = 2 γ / ( γ + 1 ) > 1 . Test the momentum equation, with its initial term, against H 4 H^4 H 4 vector fields. Since H 4 ( T 3 ) ↪ W 1 , ∞ H^4(\mathbb{T}^3) \hookrightarrow W^{1,\infty} H 4 ( T 3 ) ↪ W 1 , ∞ ,
∥ m n ( t ) − m n ( s ) ∥ H − 4 ≤ C ∫ s t ( ∥ ρ n ( r ) u n ( r ) ⊗ u n ( r ) ∥ 1 + ∥ P n ( r ) ∥ 1 ) d r + C ∫ s t ∥ S ( u n ( r ) ) ∥ 2 d r ≤ C T ( ∣ t − s ∣ + ∣ t − s ∣ 1 / 2 ) . \begin{aligned}
\|m_n(t) - m_n(s)\|_{H^{-4}} &\le C\int_s^t \left(\|\rho_n(r)u_n(r) \otimes u_n(r)\|_1 + \|P_n(r)\|_1\right)\,dr \\
&\quad+ C\int_s^t \|S(u_n(r))\|_2\,dr \\
&\le C_T\left(|t-s| + |t-s|^{1/2}\right).
\end{aligned} ∥ m n ( t ) − m n ( s ) ∥ H − 4 ≤ C ∫ s t ( ∥ ρ n ( r ) u n ( r ) ⊗ u n ( r ) ∥ 1 + ∥ P n ( r ) ∥ 1 ) d r + C ∫ s t ∥ S ( u n ( r )) ∥ 2 d r ≤ C T ( ∣ t − s ∣ + ∣ t − s ∣ 1/2 ) .
This uses the L t ∞ L x 1 L^\infty_tL^1_x L t ∞ L x 1 bounds for ( ρ n u n ⊗ u n ) (\rho_nu_n \otimes u_n) ( ρ n u n ⊗ u n ) , P n P_n P n and the L 2 ( Q T ) L^2(Q_T) L 2 ( Q T ) stress bound. The initial term extends the estimate to s = 0 s = 0 s = 0 , with m n ( 0 ) = m 0 n m_n(0) = m_{0n} m n ( 0 ) = m 0 n . Lemma 2.4 yields
m n → m = ρ u in C ( [ 0 , T ] ; H − 4 ) . (71) m_n \to m = \rho u \quad\text{in } C([0,T];H^{-4}). \tag*{(71)} m n → m = ρ u in C ([ 0 , T ] ; H − 4 ) . (71)
The time compactness uses the initialized momentum equation and the energy bounds, not an energy equality for the limit. The time-uniform L r m L^{r_m} L r m bound and density of smooth test functions also imply
m ∈ C w ( [ 0 , T ] ; L r m ) , m ( 0 ) = m 0 . m \in C_w([0,T];L^{r_m}), \qquad m(0) = m_0. m ∈ C w ([ 0 , T ] ; L r m ) , m ( 0 ) = m 0 .
For the chosen p > 6 / 5 p > 6/5 p > 6/5 , we have p ′ < 6 p' < 6 p ′ < 6 and
H 1 ( T 3 ) ⋐ L p ′ ( T 3 ) , L p ( T 3 ) ⋐ H − 1 ( T 3 ) . H^1(\mathbb{T}^3) \Subset L^{p'}(\mathbb{T}^3), \qquad L^p(\mathbb{T}^3) \Subset H^{-1}(\mathbb{T}^3). H 1 ( T 3 ) ⋐ L p ′ ( T 3 ) , L p ( T 3 ) ⋐ H − 1 ( T 3 ) .
The limit m m m has the same L 2 L p L^2L^p L 2 L p bound. The compact interpolation estimate in Lemma 2.4 gives
∥ m n − m ∥ L 2 H − 1 ≤ ε ∥ m n − m ∥ L 2 L p + C ε T ∥ m n − m ∥ C t H − 4 . \|m_n - m\|_{L^2H^{-1}} \le\varepsilon\|m_n - m\|_{L^2L^p} + C_\varepsilon\sqrt{T}\|m_n - m\|_{C_tH^{-4}}. ∥ m n − m ∥ L 2 H − 1 ≤ ε ∥ m n − m ∥ L 2 L p + C ε T ∥ m n − m ∥ C t H − 4 .
First let n → ∞ n \to\infty n → ∞ , then ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , to obtain
m n → ρ u in L 2 ( 0 , T ; H − 1 ) . (72) m_n \to\rho u \quad\text{in } L^2(0,T;H^{-1}). \tag*{(72)} m n → ρ u in L 2 ( 0 , T ; H − 1 ) . (72)
The inhomogeneous norms retain the full constant Fourier mode and allow arbitrary mean momentum.
Closing the density oscillations. We now compare the initialized entropy identities. The averaged flux identity controls the spatial integral of the pressure covariance; Lemma 4.4 supplies the limiting renormalized equation without identifying $h_k prematurely. These are distinct inputs: the former supplies a nonnegative term, while the latter permits an entropy comparison with the same initial value. The remaining error is the truncation defect Q k Q_k Q k , which will tend to zero in L 2 L^2 L 2 before strict convexity is invoked.
Proof of Proposition 2.3 (i). Following [13 Section 8] , choose truncated entropies with divergence coefficient T k T_k T k and compare their spatial integrals through (65) . For k ≥ 1 k \ge1 k ≥ 1 , set
L k ( z ) = z ∫ 1 z T k ( s ) s − 2 d s ( z > 0 ) , L k ( 0 ) = 0 , L_k(z) = z \int_1^z T_k(s)s^{-2}\,\mathrm{d}s \quad(z>0), \qquad L_k(0)=0, L k ( z ) = z ∫ 1 z T k ( s ) s − 2 d s ( z > 0 ) , L k ( 0 ) = 0 ,
ℓ k = ⟨ L k ⟩ , Q k = T k ( ρ ) − β k ≥ 0. \ell_k = \langle L_k\rangle, \qquad Q_k = T_k(\rho)-\beta_k \ge0. ℓ k = ⟨ L k ⟩ , Q k = T k ( ρ ) − β k ≥ 0.
We have
z L k ′ ( z ) − L k = T k , L k ′ ′ = T k ′ z ≥ 0. zL_k'(z)-L_k=T_k, \qquad L_k''=\frac{T_k'}{z}\ge0. z L k ′ ( z ) − L k = T k , L k ′′ = z T k ′ ≥ 0.
Thus L k L_k L k is convex, agrees with z log z z\log z z log z on [ 0 , k ] [0,k] [ 0 , k ] , and is affine on [ 3 k , ∞ ) [3k,\infty) [ 3 k , ∞ ) . The affine continuation preserves convexity, which will give a nonnegative Jensen defect. A cutoff of the entropy itself would not in general preserve that property.
The derivative of z log z z\log z z log z is singular at zero. Lemma C.1 regularizes this derivative, subtracts the affine high-density part to obtain an admissible renormalization, and restores that part by the continuity equation. At fixed k k k , the vacuum regularization has errors of order ε \varepsilon ε in the entropy, the divergence coefficient, and the initial integral. They vanish for both the approximate and the limiting equations before the comparison below. No positive lower bound on either density is required.
Integrating the resulting renormalized equations over the torus removes the entropy-flux divergence. Affine growth of L k L_k L k and the γ \gamma γ -moment bound, γ > 1 \gamma>1 γ > 1 , give
L k ( ρ n ) ⇀ ℓ k in L 1 ( Q T ) , L k ( ρ 0 n ) → L k ( ρ 0 ) in L 1 ( T 3 ) . L_k(\rho_n)\rightharpoonup\ell_k \text{ in }L^1(Q_T), \qquad L_k(\rho_{0n})\to L_k(\rho_0)\text{ in }L^1(\mathbb{T}^3). L k ( ρ n ) ⇀ ℓ k in L 1 ( Q T ) , L k ( ρ 0 n ) → L k ( ρ 0 ) in L 1 ( T 3 ) .
For every ψ ∈ C c ∞ ( [ 0 , T ) ) \psi\in C_c^\infty([0,T)) ψ ∈ C c ∞ ([ 0 , T )) , we obtain
− ∫ 0 T ψ ′ ∫ T 3 ℓ k + ∫ 0 T ψ ∫ T 3 H k = ψ ( 0 ) ∫ T 3 L k ( ρ 0 ) , (73) -\int_0^T \psi'\int_{\mathbb{T}^3}\ell_k+\int_0^T\psi\int_{\mathbb{T}^3}H_k=\psi(0)\int_{\mathbb{T}^3}L_k(\rho_0), \tag*{(73)} − ∫ 0 T ψ ′ ∫ T 3 ℓ k + ∫ 0 T ψ ∫ T 3 H k = ψ ( 0 ) ∫ T 3 L k ( ρ 0 ) , (73)
− ∫ 0 T ψ ′ ∫ T 3 L k ( ρ ) + ∫ 0 T ψ ∫ T 3 T k ( ρ ) d = ψ ( 0 ) ∫ T 3 L k ( ρ 0 ) . (74) -\int_0^T\psi'\int_{\mathbb{T}^3}L_k(\rho)+\int_0^T\psi\int_{\mathbb{T}^3}T_k(\rho)\,\mathrm{d}=\psi(0)\int_{\mathbb{T}^3}L_k(\rho_0). \tag*{(74)} − ∫ 0 T ψ ′ ∫ T 3 L k ( ρ ) + ∫ 0 T ψ ∫ T 3 T k ( ρ ) d = ψ ( 0 ) ∫ T 3 L k ( ρ 0 ) . (74)
Here H k H_k H k is the weak limit of T k ( ρ n ) d n T_k(\rho_n)\,\mathrm{d}n T k ( ρ n ) d n , whereas h k h_k h k is the source in the truncated continuity equation. Their roles differ: z L k ′ ( z ) − L k ( z ) = T k ( z ) zL_k'(z)-L_k(z)=T_k(z) z L k ′ ( z ) − L k ( z ) = T k ( z ) , but z T k ′ ( z ) − T k ( z ) = a k ( z ) zT_k'(z)-T_k(z)=a_k(z) z T k ′ ( z ) − T k ( z ) = a k ( z ) . The identical right-hand sides in (73) –(74) come from strong convergence of the prepared initial data. Equality of the initial masses alone would not give equality of these entropy integrals.
Set
D k ( t ) = ∫ T 3 [ ℓ k − L k ( ρ ) ] ≥ 0. \mathcal{D}_k(t)=\int_{\mathbb{T}^3}\left[\ell_k-L_k(\rho)\right]\ge0. D k ( t ) = ∫ T 3 [ ℓ k − L k ( ρ ) ] ≥ 0.
Subtracting (74) from (73) gives
D k + ∫ T 3 ( H k − T k ( ρ ) d ) = 0 , D k ( 0 ) = 0. \mathcal{D}_k+\int_{\mathbb{T}^3}\left(H_k-T_k(\rho)\,\mathrm{d}\right)=0, \qquad\mathcal{D}_k(0)=0. D k + ∫ T 3 ( H k − T k ( ρ ) d ) = 0 , D k ( 0 ) = 0.
The derivative is integrable, so D k \mathcal{D}_k D k has an absolutely continuous representative. Substituting (4.12) into the spatial integral yields
D k ′ + 1 ∣ T 3 ∣ ∫ T 3 C k = ∫ T 3 Q k d , D k ( 0 ) = 0. (75) \mathcal{D}_k' + \frac{1}{|\mathbb{T}^3|}\int_{\mathbb{T}^3} C_k = \int_{\mathbb{T}^3} Q_k\,\mathrm{d}, \qquad\mathcal{D}_k(0)=0. \tag*{(75)} D k ′ + ∣ T 3 ∣ 1 ∫ T 3 C k = ∫ T 3 Q k d , D k ( 0 ) = 0. (75)
The covariance term controls the sign after spatial integration, while the remaining term contains the truncation defect Q k Q_k Q k , rather than an unidentified product involving h k h_k h k . To estimate this right-hand side, observe that
0 ≤ Q k ≤ ρ − β k , ∥ Q k ∥ L 1 ( Q T ) ≤ T B k 1 − γ . 0 \le Q_k \le\rho-\beta_k,\qquad\|Q_k\|_{L^1(Q_T)}\le TBk^{1-\gamma}. 0 ≤ Q k ≤ ρ − β k , ∥ Q k ∥ L 1 ( Q T ) ≤ T B k 1 − γ .
Jensen’s inequality and (4.15) also give
∫ Q T ∣ Q k ∣ γ + 1 ≤ ∫ Q T ∣ T k ( ρ ) − T k ( z ) ∣ γ + 1 ≤ C T . \int_{Q_T}|Q_k|^{\gamma+1}\le\int_{Q_T}|T_k(\rho)-T_k(z)|^{\gamma+1}\le C_T. ∫ Q T ∣ Q k ∣ γ + 1 ≤ ∫ Q T ∣ T k ( ρ ) − T k ( z ) ∣ γ + 1 ≤ C T .
Since
1 2 = θ + 1 − θ γ + 1 , θ = γ − 1 2 γ , \frac{1}{2}=\theta+\frac{1-\theta}{\gamma+1},\qquad\theta=\frac{\gamma-1}{2\gamma}, 2 1 = θ + γ + 1 1 − θ , θ = 2 γ γ − 1 ,
interpolation yields
∥ Q k ∥ 2 ≤ ∥ Q k ∥ 1 ( γ − 1 ) / ( 2 γ ) ∥ Q k ∥ γ + 1 ( γ + 1 ) / ( 2 γ ) ≤ C T k − ( γ − 1 ) 2 / ( 2 γ ) ⟶ 0. (76) \|Q_k\|_2\le\|Q_k\|_1^{(\gamma-1)/(2\gamma)}\|Q_k\|_{\gamma+1}^{(\gamma+1)/(2\gamma)}
\le C_T k^{-(\gamma-1)^2/(2\gamma)}\longrightarrow0. \tag*{(76)} ∥ Q k ∥ 2 ≤ ∥ Q k ∥ 1 ( γ − 1 ) / ( 2 γ ) ∥ Q k ∥ γ + 1 ( γ + 1 ) / ( 2 γ ) ≤ C T k − ( γ − 1 ) 2 / ( 2 γ ) ⟶ 0. (76)
To recover the full entropy z log z z\log z z log z , set h e n t ( z ) = z log z h_{\mathrm{ent}}(z)=z\log z h ent ( z ) = z log z , with value zero at z = 0 z=0 z = 0 . The error E k = h e n t − L k E_k=h_{\mathrm{ent}}-L_k E k = h ent − L k vanishes on [ 0 , k ] [0,k] [ 0 , k ] and is nonnegative and convex, since
E k ′ ′ = 1 − T k ′ z ≥ 0. E_k''=\frac{1-T_k'}{z}\ge0. E k ′′ = z 1 − T k ′ ≥ 0.
For z > k z>k z > k ,
0 ≤ E k ( z ) ≤ z log ( z / k ) ≤ k 1 − γ exp ( 1 ) ( γ − 1 ) z γ . 0\le E_k(z)\le z\log(z/k)\le\frac{k^{1-\gamma}}{\exp(1)(\gamma-1)}z^\gamma. 0 ≤ E k ( z ) ≤ z log ( z / k ) ≤ exp ( 1 ) ( γ − 1 ) k 1 − γ z γ .
Define the full entropy defect
D ( t ) = ∫ T 3 [ ⟨ h e n t ⟩ − h e n t ( ρ ) ] ≥ 0. \mathcal{D}(t)=\int_{\mathbb{T}^3}\left[\langle h_{\mathrm{ent}}\rangle-h_{\mathrm{ent}}(\rho)\right]\ge0. D ( t ) = ∫ T 3 [ ⟨ h ent ⟩ − h ent ( ρ ) ] ≥ 0.
For almost every time,
0 ≤ D − D k = ∫ T 3 [ ⟨ E k ⟩ − E k ( ρ ) ] ≤ C γ B k 1 − γ . 0\le\mathcal{D}-\mathcal{D}_k=\int_{\mathbb{T}^3}\left[\langle E_k\rangle-E_k(\rho)\right]\le C_\gamma Bk^{1-\gamma}. 0 ≤ D − D k = ∫ T 3 [ ⟨ E k ⟩ − E k ( ρ ) ] ≤ C γ B k 1 − γ .
Given 0 ≤ ζ ∈ C ∞ ( 0 , T ) 0\le\zeta\in C^\infty(0,T) 0 ≤ ζ ∈ C ∞ ( 0 , T ) , use the backward primitive ψ ( t ) = ∫ t T ζ ( s ) d s \psi(t)=\int_t^T\zeta(s)\,\mathrm{d}s ψ ( t ) = ∫ t T ζ ( s ) d s in (4.22). The zero initial difference, C k ≥ 0 C_k\ge0 C k ≥ 0 , and (4.23) give
0 ≤ ∫ 0 T ζ D k ≤ ∥ ψ ∥ ∞ ∥ Q k ∥ 2 ∥ d ∥ 2 ⟶ 0. 0\le\int_0^T\zeta\mathcal{D}_k\le\|\psi\|_\infty\|Q_k\|_2\|d\|_2\longrightarrow0. 0 ≤ ∫ 0 T ζ D k ≤ ∥ ψ ∥ ∞ ∥ Q k ∥ 2 ∥ d ∥ 2 ⟶ 0.
The test ψ \psi ψ is admissible up to t = 0 t=0 t = 0 and vanishes near T T T . Its initial contribution is zero because the two initialized entropy identities have the same initial term. Moreover,
∫ 0 T ζ ( D − D k ) ≤ C γ B ∥ ζ ∥ 1 k 1 − γ ⟶ 0. \int_0^T\zeta(\mathcal{D}-\mathcal{D}_k)\le C_\gamma B\|\zeta\|_1k^{1-\gamma}\longrightarrow0. ∫ 0 T ζ ( D − D k ) ≤ C γ B ∥ ζ ∥ 1 k 1 − γ ⟶ 0.
Letting k → ∞ k\to\infty k → ∞ gives D = 0 \mathcal{D}=0 D = 0 . The nonnegative Jensen defect therefore vanishes almost everywhere. Strict convexity gives Y t , x = δ ρ Y_{t,x}=\delta_\rho Y t , x = δ ρ where ρ > 0 \rho>0 ρ > 0 ; nonnegative support and zero mean give δ 0 \delta_0 δ 0 where ρ = 0 \rho=0 ρ = 0 . Hence
Y t , x = δ ρ ( t , x ) almost everywhere in Q T . (77) Y_{t,x}=\delta_{\rho(t,x)}\quad\text{almost everywhere in }Q_T. \tag*{(77)} Y t , x = δ ρ ( t , x ) almost everywhere in Q T . (77)
Set J ( z ) = z / ( 1 + z ) J(z)=z/(1+z) J ( z ) = z / ( 1 + z ) . Testing with J J J , J 2 J^2 J 2 and the bounded multiplier J ( ρ ) J(\rho) J ( ρ ) gives
∥ J ( ρ n ) − J ( ρ ) ∥ L 2 ( Q T ) 2 = ∫ Q T [ J ( ρ n ) 2 − 2 J ( ρ n ) J ( ρ ) + J ( ρ ) 2 ] → 0. \lVert J(\rho_n)-J(\rho)\rVert_{L^2(Q_T)}^2=\int_{Q_T}\left[J(\rho_n)^2-2J(\rho_n)J(\rho)+J(\rho)^2\right]\to0. ∥ J ( ρ n ) − J ( ρ ) ∥ L 2 ( Q T ) 2 = ∫ Q T [ J ( ρ n ) 2 − 2 J ( ρ n ) J ( ρ ) + J ( ρ ) 2 ] → 0.
The inverse of J J J is Lipschitz on bounded density ranges; the tail bound controls their complements. Thus ρ n → ρ \rho_n\to\rho ρ n → ρ in measure. Set q c = γ + 1 / 6 q_c=\gamma+1/6 q c = γ + 1/6 . For every 1 ≤ r < q c 1\le r<q_c 1 ≤ r < q c , the distribution-function formula and (4.2) give
∫ { ρ n > R } ρ n r ≤ C T ( 1 + r q c − r ) R r − q c . \int_{\{\rho_n>R\}}\rho_n^r\le C_T\left(1+\frac{r}{q_c-r}\right)R^{r-q_c}. ∫ { ρ n > R } ρ n r ≤ C T ( 1 + q c − r r ) R r − q c .
Thus ρ n r \rho_n^r ρ n r is uniformly integrable, and Vitali’s theorem yields
ρ n → ρ in L r ( Q T ) , 1 ≤ r < γ + 1 6 . \rho_n\to\rho\quad\text{in }L^r(Q_T),\qquad1\le r<\gamma+\frac{1}{6}. ρ n → ρ in L r ( Q T ) , 1 ≤ r < γ + 6 1 .
The same uniform-integrability argument and the artificial-pressure estimate give
ρ n γ → ρ γ in L s ( Q T ) , 1 ≤ s < 1 + 1 6 γ , \rho_n^\gamma\to\rho^\gamma\quad\text{in }L^s(Q_T),\qquad1\le s<1+\frac{1}{6\gamma}, ρ n γ → ρ γ in L s ( Q T ) , 1 ≤ s < 1 + 6 γ 1 ,
P n → ρ γ in L s ( Q T ) , 1 ≤ s < 1 + 1 6 max { 3 , γ } . P_n\to\rho^\gamma\quad\text{in }L^s(Q_T),\qquad1\le s<1+\frac{1}{6\max\{3,\gamma\}}. P n → ρ γ in L s ( Q T ) , 1 ≤ s < 1 + 6 max { 3 , γ } 1 .
The continuity equation supplies the initial trace. □
The negative-norm momentum compactness was proved before the entropy comparison. We now use it to identify convection, and use strong density convergence to identify the weak limit of ρ n u n \sqrt{\rho_n}u_n ρ n u n . Convergence of the kinetic norms then gives strong convergence of these weighted velocities. Only after these steps do we pass to the energy inequality using E n 0 → E 0 E_{n0}\to E_0 E n 0 → E 0 .
Proof of Proposition 2.3 (ii). We use the momentum convergence already established in (4.18)–(4.19). The entropy argument has now identified the density and pressure, so these compactness estimates can be combined with the strong density convergence.
Let Ψ \Psi Ψ be a smooth periodic matrix field compactly supported in time in ( 0 , T ) (0,T) ( 0 , T ) . Since multiplication by Ψ \Psi Ψ preserves H 1 H^1 H 1 , weak convergence of Ψ u n \Psi u_n Ψ u n in L 2 H 1 L^2H^1 L 2 H 1 and (4.19) give
∫ Q T ( ρ n u n ⊗ u n ) : Ψ = ∫ 0 T ⟨ m n , Ψ u n ⟩ d t ⟶ ∫ Q T ( ρ u ⊗ u ) : Ψ . \int_{Q_T}(\rho_nu_n\otimes u_n):\Psi=\int_0^T\langle m_n,\Psi u_n\rangle\,dt\longrightarrow\int_{Q_T}(\rho u\otimes u):\Psi. ∫ Q T ( ρ n u n ⊗ u n ) : Ψ = ∫ 0 T ⟨ m n , Ψ u n ⟩ d t ⟶ ∫ Q T ( ρ u ⊗ u ) : Ψ.
Indeed, the pairing is split into
∫ 0 T ⟨ m n − m , Ψ u n ⟩ d t + ∫ 0 T ⟨ m , Ψ ( u n − u ) ⟩ d t , \int_0^T\langle m_n-m,\Psi u_n\rangle\,dt+\int_0^T\langle m,\Psi(u_n-u)\rangle\,dt, ∫ 0 T ⟨ m n − m , Ψ u n ⟩ d t + ∫ 0 T ⟨ m , Ψ ( u n − u )⟩ d t ,
whose first term tends to zero by strong negative-norm convergence and whose second tends to zero by weak L 2 H 1 L^2H^1 L 2 H 1 convergence. Thus ρ n u n ⊗ u n → ρ u ⊗ u \rho_nu_n\otimes u_n\to\rho u\otimes u ρ n u n ⊗ u n → ρ u ⊗ u in distributions. This argument precedes the passage to the limiting energy inequality.
Set W n = ρ n u n W_n=\sqrt{\rho_n}u_n W n = ρ n u n and W = ρ u W=\sqrt{\rho}u W = ρ u , with value zero on vacuum. Strong density convergence gives
∥ ρ n − ρ ∥ L t 2 γ L x 2 γ γ − 1 2 γ ≤ ∥ ρ n − ρ ∥ L γ γ ⟶ 0 , \left\|\sqrt{\rho_n}-\sqrt{\rho}\right\|_{L^{2\gamma}_t L^{\frac{2\gamma}{\gamma-1}}_x}^{2\gamma}\leq\|\rho_n-\rho\|_{L^\gamma}^{\gamma}\longrightarrow0, ∥ ρ n − ρ ∥ L t 2 γ L x γ − 1 2 γ 2 γ ≤ ∥ ρ n − ρ ∥ L γ γ ⟶ 0 ,
and hence the square roots also converge strongly in L t 2 L x 6 / 5 L^2_tL_x^{6/5} L t 2 L x 6/5 . For a smooth vector test function φ \varphi φ ,
∣ ∫ Q T ( ρ n − ρ ) u n ⋅ φ ∣ ≤ ∥ φ ∥ ∞ ∥ ρ n − ρ ∥ L 2 L 6 / 5 ∥ u n ∥ L 2 L 6 ⟶ 0. \left|\int_{Q_T}(\sqrt{\rho_n}-\sqrt{\rho})u_n\cdot\varphi\right|\leq\|\varphi\|_\infty\|\sqrt{\rho_n}-\sqrt{\rho}\|_{L^2L^{6/5}}\|u_n\|_{L^2L^6}\longrightarrow0. ∫ Q T ( ρ n − ρ ) u n ⋅ φ ≤ ∥ φ ∥ ∞ ∥ ρ n − ρ ∥ L 2 L 6/5 ∥ u n ∥ L 2 L 6 ⟶ 0.
Weak convergence of u n u_n u n handles the fixed coefficient ρ \sqrt{\rho} ρ . The kinetic-energy bound then gives W n ⇀ W W_n\rightharpoonup W W n ⇀ W in L 2 ( Q T ) L^2(Q_T) L 2 ( Q T ) .
Take Ψ = η ( t ) Id \Psi=\eta(t)\operatorname{Id} Ψ = η ( t ) Id , η ≥ 0 \eta\geq0 η ≥ 0 , in the identified convective limit to obtain
∫ Q T η ∣ W n ∣ 2 ⟶ ∫ Q T η ∣ W ∣ 2 , η ∈ C c ∞ ( 0 , T ) . \int_{Q_T}\eta|W_n|^2\longrightarrow\int_{Q_T}\eta|W|^2,\qquad\eta\in C^\infty_c(0,T). ∫ Q T η ∣ W n ∣ 2 ⟶ ∫ Q T η ∣ W ∣ 2 , η ∈ C c ∞ ( 0 , T ) .
Weak convergence and convergence of the weighted norms yield strong convergence on interior time intervals. For 0 < ε < T / 4 0<\varepsilon<T/4 0 < ε < T /4 , choose 0 ≤ η ε ≤ 1 0\leq\eta_\varepsilon\leq1 0 ≤ η ε ≤ 1 supported in ( 0 , T ) (0,T) ( 0 , T ) and equal to one on [ ε , T − ε ] [\varepsilon,T-\varepsilon] [ ε , T − ε ] . Set K k i n = sup n ess sup t ∫ T 3 ρ n ∣ u n ∣ 2 K_{\mathrm{kin}}=\sup_n\operatorname*{ess\,sup}_t\int_{\mathbb{T}^3}\rho_n|u_n|^2 K kin = sup n ess sup t ∫ T 3 ρ n ∣ u n ∣ 2 . The approximate bound and (4.17) give
∫ Q T ( 1 − η ε ) ( ∣ W n ∣ 2 + ∣ W ∣ 2 ) ≤ 2 K k i n ∣ ( 0 , ε ) ∪ ( T − ε , T ) ∣ = 4 K k i n ε . \int_{Q_T}(1-\eta_\varepsilon)(|W_n|^2+|W|^2)\leq2K_{\mathrm{kin}}|(0,\varepsilon)\cup(T-\varepsilon,T)|=4K_{\mathrm{kin}}\varepsilon. ∫ Q T ( 1 − η ε ) ( ∣ W n ∣ 2 + ∣ W ∣ 2 ) ≤ 2 K kin ∣ ( 0 , ε ) ∪ ( T − ε , T ) ∣ = 4 K kin ε .
Let n → ∞ n\to\infty n → ∞ and then ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 . The norms converge on Q T Q_T Q T , and hence
∥ W n − W ∥ 2 2 = ∥ W n ∥ 2 2 + ∥ W ∥ 2 2 − 2 ∫ Q T W n ⋅ W ⟶ 0. (78) \|W_n-W\|_2^2=\|W_n\|_2^2+\|W\|_2^2-2\int_{Q_T}W_n\cdot W\longrightarrow0. \tag*{(78)} ∥ W n − W ∥ 2 2 = ∥ W n ∥ 2 2 + ∥ W ∥ 2 2 − 2 ∫ Q T W n ⋅ W ⟶ 0. (78)
Consequently,
∥ W n ⊗ W n − W ⊗ W ∥ 1 ≤ ( ∥ W n ∥ 2 + ∥ W ∥ 2 ) ∥ W n − W ∥ 2 ⟶ 0. (79) \|W_n\otimes W_n-W\otimes W\|_1\leq(\|W_n\|_2+\|W\|_2)\|W_n-W\|_2\longrightarrow0. \tag*{(79)} ∥ W n ⊗ W n − W ⊗ W ∥ 1 ≤ ( ∥ W n ∥ 2 + ∥ W ∥ 2 ) ∥ W n − W ∥ 2 ⟶ 0. (79)
Strong convergence of the pressure, convection, and initial data, together with weak convergence of the stress, yields the limiting momentum equation. Conservation of mass and momentum follows by spatial integration and extends to all times by weak continuity.
The same convergence gives strong L 1 L^1 L 1 convergence of the full energy density:
1 2 ∣ W n ∣ 2 + ρ n γ γ − 1 + 1 2 ϵ ρ n 3 ⟶ 1 2 ∣ W ∣ 2 + ρ γ γ − 1 \frac{1}{2}|W_n|^2+\frac{\rho_n^\gamma}{\gamma-1}+\frac{1}{2}\epsilon\rho_n^3\longrightarrow\frac{1}{2}|W|^2+\frac{\rho^\gamma}{\gamma-1} 2 1 ∣ W n ∣ 2 + γ − 1 ρ n γ + 2 1 ϵ ρ n 3 ⟶ 2 1 ∣ W ∣ 2 + γ − 1 ρ γ
in L 1 ( Q T ) L^1(Q_T) L 1 ( Q T ) . {#eq:4.27}
Integrating the energy density in space gives
E n ⟶ E in L 1 ( 0 , T ) , (80) E_n\longrightarrow E\qquad\text{in }L^1(0,T), \tag*{(80)} E n ⟶ E in L 1 ( 0 , T ) , (80)
where E n E_n E n is defined in Lemma 2.1. Thus ∫ E n ψ ′ ⟶ ∫ E ψ ′ \int E_n\psi'\longrightarrow\int E\psi' ∫ E n ψ ′ ⟶ ∫ E ψ ′ also for sign-changing ψ ′ \psi' ψ ′ . For every 0 ≤ ψ ∈ C c ∞ ( [ 0 , T ) ) 0\leq\psi\in C^\infty_c([0,T)) 0 ≤ ψ ∈ C c ∞ ([ 0 , T )) , lower semicontinuity of the dissipation and E n 0 → E 0 E_{n0}\to E_0 E n 0 → E 0 yield
− ∫ 0 T E ψ ′ + ∫ 0 T D ψ ≤ lim inf n [ − ∫ 0 T E n ψ ′ + ∫ 0 T D n ψ ] ≤ E 0 ψ ( 0 ) . \begin{aligned}
-\int_0^T E\psi'+\int_0^T\mathcal{D}\psi\leq\liminf_n\left[-\int_0^T E_n\psi'+\int_0^T\mathcal{D}_n\psi\right] \\
&\leq E_0\psi(0).
\end{aligned} − ∫ 0 T E ψ ′ + ∫ 0 T D ψ ≤ n lim inf [ − ∫ 0 T E n ψ ′ + ∫ 0 T D n ψ ] ≤ E 0 ψ ( 0 ) .
Here D ( t ) = μ ∥ ∇ u ( t ) ∥ 2 2 + ( λ + μ ) ∥ div u ( t ) ∥ 2 2 D(t)=\mu\lVert\nabla u(t)\rVert_{2}^{2}+(\lambda+\mu)\lVert\operatorname{div}u(t)\rVert_{2}^{2} D ( t ) = μ ∥ ∇ u ( t ) ∥ 2 2 + ( λ + μ ) ∥ div u ( t ) ∥ 2 2 , so the distributional energy inequality follows. The distributional energy inequality gives
E ( t ) + ∫ 0 t D ≤ E 0 for almost every t . E(t)+\int_{0}^{t}D\le E_{0}\quad\text{for almost every }t. E ( t ) + ∫ 0 t D ≤ E 0 for almost every t .
For each fixed t t t , choose times t j → t t_j\to t t j → t from the full-measure set where the inequality holds. The weakly continuous representatives of ρ \rho ρ and m m m give weak convergence at these times. The pressure potential is convex, and the kinetic energy has the convex perspective form ∣ m ∣ 2 / ( 2 ρ ) |m|^{2}/(2\rho) ∣ m ∣ 2 / ( 2 ρ ) used above. Consequently,
E ( t ) ≤ lim inf j → ∞ E ( t j ) , ∫ 0 t j D ( s ) d s ⟶ ∫ 0 t D ( s ) d s . E(t)\le\liminf_{j\to\infty}E(t_j),\qquad\int_{0}^{t_j}D(s)\,ds\longrightarrow\int_{0}^{t}D(s)\,ds. E ( t ) ≤ j → ∞ lim inf E ( t j ) , ∫ 0 t j D ( s ) d s ⟶ ∫ 0 t D ( s ) d s .
This extends the inequality to every t ≥ 0 t\ge0 t ≥ 0 , with the energy defined by these representatives. On vacuum the kinetic density is zero; finiteness of the perspective functional enforces m = 0 m=0 m = 0 there. □
Proof of Theorem 1.2. For zero initial mass, the zero solution suffices. Otherwise, Lemma 2.1 provides approximate solutions satisfying (2.2) with E n 0 → E 0 E_{n0}\to E_0 E n 0 → E 0 . Proposition 2.2 and Lemma 4.1 give the uniform estimates required by Proposition 2.3, which yields the desired weak solution on each finite time interval. A diagonal subsequence over integer time intervals, truncation levels, and test functions gives a global solution on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . □
Cubic Artificial-Pressure Approximation
We prove Lemma 2.1 by preparing the initial data and applying the fixed-e e e existence result below. The Galerkin scheme, artificial density diffusion, and auxiliary pressure follow the artificial-pressure approximation framework of [14 , 15 , 27 ] . We give the periodic implementation because the later material test requires the fixed-member bounds ρ ∈ L ∞ L 3 \rho\in L^{\infty}L^{3} ρ ∈ L ∞ L 3 and m ∈ L 2 L 2 m\in L^{2}L^{2} m ∈ L 2 L 2 . The exponent three is used for these pairings; no assertion of priority for artificial cubic pressure is needed.
The successive limits are
N ⟶ ∞ ( α , e , σ fixed ) , α ↓ 0 ( e , σ fixed ) , σ ↓ 0 ( e fixed ) . N\longrightarrow\infty\quad(\alpha,e,\sigma\text{ fixed}),\qquad\alpha\downarrow0\quad(e,\sigma\text{ fixed}),\qquad\sigma\downarrow0\quad(e\text{ fixed}). N ⟶ ∞ ( α , e , σ fixed ) , α ↓ 0 ( e , σ fixed ) , σ ↓ 0 ( e fixed ) .
Only after these limits does Section 4 send e = e n e=e_n e = e n to zero. In particular, the e e e -dependent bounds proved here justify tests on an individual approximate solution; they do not enter the uniform pressure budget of Proposition 2.2.
This construction does not use the barycentric pressure estimate. At fixed σ > 0 \sigma>0 σ > 0 , the auxiliary exponent G G G provides the integrability for the untruncated effective-flux test. At fixed e > 0 e>0 e > 0 , the cubic term provides the corresponding integrability while σ \sigma σ is removed. Thus the pressure gains (A.16) and (A.22) are classical divergence-test estimates with constants allowed to depend on the parameters that remain fixed. The later uniform estimate (2.4) is used only after this supply of approximate solutions has been constructed.
Preparation of Initial Data
Let m = ∫ T 3 ρ 0 > 0 m=\int_{\mathbb{T}^{3}}\rho_{0}>0 m = ∫ T 3 ρ 0 > 0 , A 0 = ρ 0 A_{0}=\sqrt{\rho_{0}} A 0 = ρ 0 , and W 0 = m 0 / ρ 0 W_{0}=m_{0}/\sqrt{\rho_{0}} W 0 = m 0 / ρ 0 , extended by zero on vacuum. The factorization m 0 = A 0 W 0 m_{0}=A_{0}W_{0} m 0 = A 0 W 0 separates the two quantities controlled by finite energy. Smoothing A 0 A_{0} A 0 and W 0 W_{0} W 0 separately preserves the kinetic energy through an unweighted L 2 L^{2} L 2 norm, including on the vacuum set. Truncating A 0 A_{0} A 0 also makes the initial artificial energy vanish. For any prescribed e n > 0 e_{n}>0 e n > 0 with e n → 0 e_{n}\to0 e n → 0 , set
L n = { max { 1 , e n − 1 / [ 4 ( 3 − γ ) ] } , γ < 3 , ∞ , γ ≥ 3. L_{n}=
\begin{cases}
\max\{1,e_{n}^{-1/[4(3-\gamma)]}\}, & \gamma<3,\\
\infty, & \gamma\ge3.
\end{cases} L n = { max { 1 , e n − 1/ [ 4 ( 3 − γ )] } , ∞ , γ < 3 , γ ≥ 3.
With a nonnegative periodic mollifier ω 1 / n \omega_{1/n} ω 1/ n , define
a ~ n = ω 1 / n ∗ min ( A 0 , L n ) + n − 1 , a n = m ∥ a ~ n ∥ 2 a ~ n \widetilde{a}_{n}=\omega_{1/n}*\min(A_{0},L_{n})+n^{-1},\qquad
a_{n}=\frac{\sqrt{m}}{\|\widetilde{a}_{n}\|_{2}}\widetilde{a}_{n} a n = ω 1/ n ∗ min ( A 0 , L n ) + n − 1 , a n = ∥ a n ∥ 2 m a n
and
b n = ω 1 / n ∗ W 0 , ρ 0 n = a n 2 , m 0 n = a n b n , u 0 n = b n / a n . (81) b_{n}=\omega_{1/n}*W_{0},\qquad
\rho_{0n}=a_{n}^{2},\qquad
m_{0n}=a_{n}b_{n},\qquad
u_{0n}=b_{n}/a_{n}. \tag*{(81)} b n = ω 1/ n ∗ W 0 , ρ 0 n = a n 2 , m 0 n = a n b n , u 0 n = b n / a n . (81)
Then ρ 0 n > 0 \rho_{0n}>0 ρ 0 n > 0 is smooth, ∫ T 3 ρ 0 n = m \int_{\mathbb{T}^{3}}\rho_{0n}=m ∫ T 3 ρ 0 n = m , and
a n → A 0 in L 2 ∩ L 2 γ , b n → W 0 in L 2 . a_{n}\to A_{0}\quad\text{in }L^{2}\cap L^{2\gamma},\qquad
b_{n}\to W_{0}\quad\text{in }L^{2}. a n → A 0 in L 2 ∩ L 2 γ , b n → W 0 in L 2 .
Consequently, ρ 0 n → ρ 0 \rho_{0n}\to\rho_{0} ρ 0 n → ρ 0 in L γ L^{\gamma} L γ and m 0 n → m 0 m_{0n}\to m_{0} m 0 n → m 0 in L r m L^{r_{m}} L r m , where r m = 2 γ / ( γ + 1 ) r_{m}=2\gamma/(\gamma+1) r m = 2 γ / ( γ + 1 ) . The initial kinetic energy converges since
1 2 ∫ T 3 ∣ m 0 n ∣ 2 ρ 0 n = 1 2 ∥ b n ∥ 2 2 ⟶ 1 2 ∥ W 0 ∥ 2 2 . \frac{1}{2}\int_{\mathbb{T}^{3}}\frac{|m_{0n}|^{2}}{\rho_{0n}}
=\frac{1}{2}\|b_{n}\|_{2}^{2}\longrightarrow\frac{1}{2}\|W_{0}\|_{2}^{2}. 2 1 ∫ T 3 ρ 0 n ∣ m 0 n ∣ 2 = 2 1 ∥ b n ∥ 2 2 ⟶ 2 1 ∥ W 0 ∥ 2 2 .
For γ < 3 \gamma<3 γ < 3 ,
e n ∫ T 3 a n 6 ≤ C e n L n 6 − 2 γ ∫ T 3 a n 2 γ ≤ C e n 1 / 2 ⟶ 0. e_{n}\int_{\mathbb{T}^{3}}a_{n}^{6}\le Ce_{n}L_{n}^{6-2\gamma}\int_{\mathbb{T}^{3}}a_{n}^{2\gamma}\le Ce_{n}^{1/2}\longrightarrow0. e n ∫ T 3 a n 6 ≤ C e n L n 6 − 2 γ ∫ T 3 a n 2 γ ≤ C e n 1/2 ⟶ 0.
For γ ≥ 3 \gamma\ge3 γ ≥ 3 , the same term is O ( e n ) O(e_{n}) O ( e n ) . The normalization factor in (A.1) tends to one because ∥ A 0 ∥ 2 2 = m \|A_{0}\|_{2}^{2}=m ∥ A 0 ∥ 2 2 = m . The estimates above use only an upper bound for this factor, while the lower bound a n > 0 a_{n}>0 a n > 0 is allowed to depend on n n n . Moreover, Hölder’s inequality gives the explicit momentum bound
∥ a n b n − A 0 W 0 ∥ r m ≤ ∥ a n − A 0 ∥ 2 γ ∥ b n ∥ 2 + ∥ A 0 ∥ 2 γ ∥ b n − W 0 ∥ 2 ⟶ 0. \|a_{n}b_{n}-A_{0}W_{0}\|_{r_{m}}
\le\|a_{n}-A_{0}\|_{2\gamma}\|b_{n}\|_{2}
+\|A_{0}\|_{2\gamma}\|b_{n}-W_{0}\|_{2}\longrightarrow0. ∥ a n b n − A 0 W 0 ∥ r m ≤ ∥ a n − A 0 ∥ 2 γ ∥ b n ∥ 2 + ∥ A 0 ∥ 2 γ ∥ b n − W 0 ∥ 2 ⟶ 0.
Thus E n 0 → E 0 E_{n0}\to E_{0} E n 0 → E 0 , without a positive lower bound for the limiting density.
Lemma A.1. Let e > 0 e>0 e > 0 , ρ ∗ > 0 \rho_{*}>0 ρ ∗ > 0 and u ∗ u_{*} u ∗ be smooth periodic initial data. Set m ∗ = ρ ∗ u ∗ m_{*}=\rho_{*}u_{*} m ∗ = ρ ∗ u ∗ , b ∗ = max { 3 , γ } b_{*}=\max\{3,\gamma\} b ∗ = max { 3 , γ } , and
P e ( ρ ) = ρ γ + e ρ 3 , H e ( ρ ) = ρ γ γ − 1 + e 2 ρ 3 . P_{e}(\rho)=\rho^{\gamma}+e\rho^{3},\qquad
H_{e}(\rho)=\frac{\rho^{\gamma}}{\gamma-1}+\frac{e}{2}\rho^{3}. P e ( ρ ) = ρ γ + e ρ 3 , H e ( ρ ) = γ − 1 ρ γ + 2 e ρ 3 . There exists a global periodic finite-energy weak solution with initial data ( ρ ∗ , m ∗ ) (\rho_{*},m_{*}) ( ρ ∗ , m ∗ ) , satisfying the renormalized continuity equation with its initial term for every C 1 C^{1} C 1 function with compactly supported derivative. For every $T>0,
{ ρ ∈ L ∞ ∩ L b ∗ + 1 ( Q T ) ∩ C w L b ∗ , u ∈ L 2 H 1 , ρ u ∈ L ∞ L 2 , ρ u ∈ C w L 2 b ∗ b ∗ + 1 ∩ L 2 L 2 , ρ u ⊗ u ∈ L 1 L 3 / 2 . \left\{
\begin{aligned}
\rho&\in L^\infty\cap L^{b^*+1}(Q_T) \cap C_w L^{b^*},\\
u &\in L^2 H^1,\qquad\sqrt{\rho}\,u \in L^\infty L^2,\\
\rho u &\in C_w L^{\frac{2b^*}{b^*+1}} \cap L^2 L^2,\\
\rho u\otimes u &\in L^1 L^{3/2}.
\end{aligned}
\right. ⎩ ⎨ ⎧ ρ u ρ u ρ u ⊗ u ∈ L ∞ ∩ L b ∗ + 1 ( Q T ) ∩ C w L b ∗ , ∈ L 2 H 1 , ρ u ∈ L ∞ L 2 , ∈ C w L b ∗ + 1 2 b ∗ ∩ L 2 L 2 , ∈ L 1 L 3/2 .
Total mass and momentum are conserved. Moreover,
− ∫ 0 T E e ψ ′ d t + ∫ 0 T D u ψ d t ≤ E e , 0 ψ ( 0 ) , 0 ≤ ψ ∈ C c ∞ ( [ 0 , T ) ) , (82) -\int_0^T \mathcal{E}_e\psi' \,dt+\int_0^T \mathcal{D}_u\psi\,dt\leq\mathcal{E}_{e,0}\psi(0),\qquad0\leq\psi\in C_c^\infty([0,T)), \tag*{(82)} − ∫ 0 T E e ψ ′ d t + ∫ 0 T D u ψ d t ≤ E e , 0 ψ ( 0 ) , 0 ≤ ψ ∈ C c ∞ ([ 0 , T )) , (82)
where
E e ( t ) = ∫ T 3 ( 1 2 ρ ∣ u ∣ 2 + H e ( ρ ) ) d x , \mathcal{E}_e(t)=\int_{\mathbb{T}^3}\left(\frac{1}{2}\rho|u|^2+H_e(\rho)\right)\,dx, E e ( t ) = ∫ T 3 ( 2 1 ρ ∣ u ∣ 2 + H e ( ρ ) ) d x ,
D u ( t ) = μ ∥ ∇ u ( t ) ∥ 2 2 + ( λ + μ ) ∥ div u ( t ) ∥ 2 2 , \mathcal{D}_u(t)=\mu\|\nabla u(t)\|_2^2+(\lambda+\mu)\|\operatorname{div}u(t)\|_2^2, D u ( t ) = μ ∥∇ u ( t ) ∥ 2 2 + ( λ + μ ) ∥ div u ( t ) ∥ 2 2 ,
E e , 0 = ∫ T 3 ( ∣ m ∗ ∣ 2 2 ρ ∗ + H e ( ρ ∗ ) ) d x . \mathcal{E}_{e,0}=\int_{\mathbb{T}^3}\left(\frac{|m_*|^2}{2\rho_*}+H_e(\rho_*)\right)\,dx. E e , 0 = ∫ T 3 ( 2 ρ ∗ ∣ m ∗ ∣ 2 + H e ( ρ ∗ ) ) d x .
Proof. Fix e > 0 e>0 e > 0 and smooth positive initial data.
Step 1: Galerkin approximation and the limit N → ∞ N\to\infty N → ∞ . Fix G > max { 8 , γ + 1 } G>\max\{8,\gamma+1\} G > max { 8 , γ + 1 } , σ > 0 \sigma>0 σ > 0 , and 0 < α ≤ 1 0<\alpha\leq1 0 < α ≤ 1 , and set
P ^ ( ρ ) = ρ γ + e ρ 3 + σ ρ G , H ( ρ ) = H e ( ρ ) + σ G − 1 ρ G . \widehat{P}(\rho)=\rho^\gamma+e\rho^3+\sigma\rho^G,\qquad H(\rho)=H_e(\rho)+\frac{\sigma}{G-1}\rho^G. P ( ρ ) = ρ γ + e ρ 3 + σ ρ G , H ( ρ ) = H e ( ρ ) + G − 1 σ ρ G .
Let X N X_N X N be the space of real vector Fourier modes with ∣ k ∣ ≤ N |k|\leq N ∣ k ∣ ≤ N , including the constant modes. For an L 2 L^2 L 2 -orthonormal basis ( φ a ) (\varphi_a) ( φ a ) , define
M [ ρ ] a b = ∫ T 3 ρ φ a ⋅ φ b , ( inf ρ ) ∣ c ∣ 2 ≤ c T M [ ρ ] c ≤ ( sup ρ ) ∣ c ∣ 2 . (83) \mathcal{M}[\rho]_{ab}=\int_{\mathbb{T}^3}\rho\varphi_a\cdot\varphi_b,\qquad
(\inf\rho)|c|^2\leq c^{\mathsf T}\mathcal{M}[\rho]c\leq(\sup\rho)|c|^2. \tag*{(83)} M [ ρ ] ab = ∫ T 3 ρ φ a ⋅ φ b , ( inf ρ ) ∣ c ∣ 2 ≤ c T M [ ρ ] c ≤ ( sup ρ ) ∣ c ∣ 2 . (83)
For positive ρ \rho ρ , the matrix is invertible. Choose u N 0 = Proj X N L 2 ( ρ ∗ d x ) u ∗ u_N^0=\operatorname{Proj}_{X_N}^{L^2(\rho_*\,dx)}u_* u N 0 = Proj X N L 2 ( ρ ∗ d x ) u ∗ , so that
∫ T 3 ρ ∗ u N 0 ⋅ φ a = ∫ T 3 m ∗ ⋅ φ a . \int_{\mathbb{T}^3}\rho_*u_N^0\cdot\varphi_a=\int_{\mathbb{T}^3}m_*\cdot\varphi_a. ∫ T 3 ρ ∗ u N 0 ⋅ φ a = ∫ T 3 m ∗ ⋅ φ a .
In particular, the constant modes preserve the initial total momentum. The weighted orthogonality also gives
∫ T 3 ρ ∗ ∣ u ∗ − u N 0 ∣ 2 = ∫ T 3 ρ ∗ ∣ u ∗ ∣ 2 − ∫ T 3 ρ ∗ ∣ u N 0 ∣ 2 ⟶ 0 , \int_{\mathbb{T}^3}\rho_*|u_*-u_N^0|^2
=\int_{\mathbb{T}^3}\rho_*|u_*|^2-\int_{\mathbb{T}^3}\rho_*|u_N^0|^2\longrightarrow0, ∫ T 3 ρ ∗ ∣ u ∗ − u N 0 ∣ 2 = ∫ T 3 ρ ∗ ∣ u ∗ ∣ 2 − ∫ T 3 ρ ∗ ∣ u N 0 ∣ 2 ⟶ 0 ,
since the Fourier spaces are dense and ρ ∗ \rho_* ρ ∗ is bounded above and below at this fixed stage. This both bounds the initial Galerkin energy and identifies its limit. The density diffusion and its momentum correction have the form used in [15 p. 361] . The Galerkin system is
{ ∂ t ρ N + div ( ρ N u N ) = α Δ ρ N , ∂ t ( ρ N u N ) + div ( ρ N u N ⊗ u N ) + ∇ P ^ ( ρ N ) = div S ( u N ) − D N , (84) \left\{
\begin{aligned}
\partial_t\rho_N+\operatorname{div}(\rho_Nu_N)&=\alpha\Delta\rho_N,\\
\partial_t(\rho_Nu_N)+\operatorname{div}(\rho_Nu_N\otimes u_N)+\nabla\widehat{P}(\rho_N)&=\operatorname{div}\mathcal{S}(u_N)-D_N,
\end{aligned}
\right. \tag*{(84)} { ∂ t ρ N + div ( ρ N u N ) ∂ t ( ρ N u N ) + div ( ρ N u N ⊗ u N ) + ∇ P ( ρ N ) = α Δ ρ N , = div S ( u N ) − D N , (84)
where
D N = α ∑ j = 1 3 ( ∂ j ρ N ) ∂ j u N , ( ρ N , u N ) ∣ t = 0 = ( ρ ∗ , u N 0 ) . D_N=\alpha\sum_{j=1}^{3}(\partial_j\rho_N)\partial_j u_N,\qquad
(\rho_N,u_N)|_{t=0}=(\rho_*,u_N^0). D N = α j = 1 ∑ 3 ( ∂ j ρ N ) ∂ j u N , ( ρ N , u N ) ∣ t = 0 = ( ρ ∗ , u N 0 ) .
The momentum equation is imposed in X N X_N X N , while the density equation is unprojected. For a prescribed Galerkin velocity, the maximum principle gives
( min ρ ∗ ) e − ∫ 0 t ∥ div u N ∥ ∞ ≤ ρ N ( t , x ) ≤ ( max ρ ∗ ) e ∫ 0 t ∥ div u N ∥ ∞ . (\min\rho_*)e^{-\int_0^t \|\operatorname{div}u_N\|_\infty} \le\rho_N(t,x) \le(\max\rho_*)e^{\int_0^t \|\operatorname{div}u_N\|_\infty}. ( min ρ ∗ ) e − ∫ 0 t ∥ div u N ∥ ∞ ≤ ρ N ( t , x ) ≤ ( max ρ ∗ ) e ∫ 0 t ∥ div u N ∥ ∞ .
For velocities in a bounded subset of X N X_N X N , the difference of the density equations satisfies ∥ ρ [ u ] − ρ [ w ] ∥ C t L 2 ≤ C N , α , ρ ∗ τ ∥ u − w ∥ C t X N \|\rho[u]-\rho[w]\|_{C_tL^2} \le C_{N,\alpha,\rho_*}\sqrt{\tau}\|u-w\|_{C_tX_N} ∥ ρ [ u ] − ρ [ w ] ∥ C t L 2 ≤ C N , α , ρ ∗ τ ∥ u − w ∥ C t X N . The inverse mass matrix is locally Lipschitz, and the Galerkin coefficients satisfy
u ( t ) = M [ ρ [ u ] ( t ) ] − 1 ( m ∗ ∣ X N + ∫ 0 t F [ ρ [ u ] , u ] ( s ) d s ) . u(t)=\mathcal{M}[\rho[u](t)]^{-1}\left(m_*|_{X_N}+\int_0^t\mathcal{F}[\rho[u],u](s)\,\mathrm{d}s\right). u ( t ) = M [ ρ [ u ] ( t ) ] − 1 ( m ∗ ∣ X N + ∫ 0 t F [ ρ [ u ] , u ] ( s ) d s ) .
For w ∈ X N w\in X_N w ∈ X N , integration by parts gives
− α ∫ T 3 ∂ j ρ N ∂ j u N , i w i = α ∫ T 3 ρ N ( Δ u N , i w i + ∂ j u N , i ∂ j w i ) . -\alpha\int_{\mathbb{T}^3}\partial_j\rho_N\,\partial_j u_{N,i}w_i
=\alpha\int_{\mathbb{T}^3}\rho_N\left(\Delta u_{N,i}w_i+\partial_j u_{N,i}\partial_jw_i\right). − α ∫ T 3 ∂ j ρ N ∂ j u N , i w i = α ∫ T 3 ρ N ( Δ u N , i w i + ∂ j u N , i ∂ j w i ) .
The resulting fixed-point map has Lipschitz constant C ( τ + τ ) C(\sqrt{\tau}+\tau) C ( τ + τ ) on a short interval and yields a local solution as in [15 ] .
Set K N = 1 2 ∫ T 3 ρ N ∣ u N ∣ 2 K_N=\frac{1}{2}\int_{\mathbb{T}^3}\rho_N|u_N|^2 K N = 2 1 ∫ T 3 ρ N ∣ u N ∣ 2 . Testing the Galerkin momentum equation with u N u_N u N gives
∫ T 3 u N ⋅ [ ∂ t ( ρ N u N ) + div ( ρ N u N ⊗ u N ) ] = K N ′ − ∫ T 3 D N ⋅ u N . \int_{\mathbb{T}^3}u_N\cdot\left[\partial_t(\rho_Nu_N)+\operatorname{div}(\rho_Nu_N\otimes u_N)\right]
=K_N'-\int_{\mathbb{T}^3}D_N\cdot u_N. ∫ T 3 u N ⋅ [ ∂ t ( ρ N u N ) + div ( ρ N u N ⊗ u N ) ] = K N ′ − ∫ T 3 D N ⋅ u N .
Indeed, the extra term from the density equation is
1 2 ∫ T 3 ∣ u N ∣ 2 α Δ ρ N = − α ∫ T 3 ∂ j ρ N u N ⋅ ∂ j u N = − ∫ T 3 D N ⋅ u N . \frac{1}{2}\int_{\mathbb{T}^3}|u_N|^2\alpha\Delta\rho_N
=-\alpha\int_{\mathbb{T}^3}\partial_j\rho_N\,u_N\cdot\partial_j u_N
=-\int_{\mathbb{T}^3}D_N\cdot u_N. 2 1 ∫ T 3 ∣ u N ∣ 2 α Δ ρ N = − α ∫ T 3 ∂ j ρ N u N ⋅ ∂ j u N = − ∫ T 3 D N ⋅ u N .
The correction cancels the corresponding term in the momentum equation. For the internal energy, the periodic chain rule gives
d d t ∫ T 3 H ( ρ N ) + ∫ T 3 P ^ ( ρ N ) div u N = − α ∫ T 3 H ′ ′ ( ρ N ) ∣ ∇ ρ N ∣ 2 . \frac{\mathrm{d}}{\mathrm{d}t}\int_{\mathbb{T}^3}H(\rho_N)+\int_{\mathbb{T}^3}\widehat{P}(\rho_N)\operatorname{div}u_N
=-\alpha\int_{\mathbb{T}^3}H''(\rho_N)|\nabla\rho_N|^2. d t d ∫ T 3 H ( ρ N ) + ∫ T 3 P ( ρ N ) div u N = − α ∫ T 3 H ′′ ( ρ N ) ∣∇ ρ N ∣ 2 .
Thus both artificial terms have a specified role: diffusion regularizes the density, and its momentum correction restores the kinetic balance. Using ρ H ′ ( ρ ) − H ( ρ ) = P ^ ( ρ ) \rho H'(\rho)-H(\rho)=\widehat{P}(\rho) ρ H ′ ( ρ ) − H ( ρ ) = P ( ρ ) in the density equation, we obtain
d d t ∫ T 3 ( 1 2 ρ N ∣ u N ∣ 2 + H ( ρ N ) ) + μ ∥ ∇ u N ∥ 2 2 + ( λ + μ ) ∥ div u N ∥ 2 2 + α ∫ T 3 H ′ ′ ( ρ N ) ∣ ∇ ρ N ∣ 2 = 0. (85) \begin{aligned}
\frac{\mathrm{d}}{\mathrm{d}t}\int_{\mathbb{T}^3}\left(\frac{1}{2}\rho_N|u_N|^2+H(\rho_N)\right)
+\mu\|\nabla u_N\|_2^2+(\lambda+\mu)\|\operatorname{div}u_N\|_2^2 \\
+\alpha\int_{\mathbb{T}^3}H''(\rho_N)|\nabla\rho_N|^2=0. \tag*{(85)}
\end{aligned} d t d ∫ T 3 ( 2 1 ρ N ∣ u N ∣ 2 + H ( ρ N ) ) + μ ∥∇ u N ∥ 2 2 + ( λ + μ ) ∥ div u N ∥ 2 2 + α ∫ T 3 H ′′ ( ρ N ) ∣∇ ρ N ∣ 2 = 0. (85)
Here H ′ ′ ( s ) = γ s γ − 2 + 3 e s + σ G s G − 2 > 0 H''(s)=\gamma s^{\gamma-2}+3es+\sigma Gs^{G-2}>0 H ′′ ( s ) = γ s γ − 2 + 3 es + σ G s G − 2 > 0 , and
∫ T 3 S ( u N ) : ∇ u N = μ ∥ ∇ u N ∥ 2 2 + ( λ + μ ) ∥ div u N ∥ 2 2 . \int_{\mathbb{T}^3}S(u_N):\nabla u_N
=\mu\|\nabla u_N\|_2^2+(\lambda+\mu)\|\operatorname{div}u_N\|_2^2. ∫ T 3 S ( u N ) : ∇ u N = μ ∥∇ u N ∥ 2 2 + ( λ + μ ) ∥ div u N ∥ 2 2 .
Since λ + μ ≥ μ / 3 > 0 \lambda+\mu\ge\mu/3>0 λ + μ ≥ μ /3 > 0 , the energy identity and Lemma 2.5 control u N u_N u N in L 2 ( 0 , T ; H 1 ) L^2(0,T;H^1) L 2 ( 0 , T ; H 1 ) . For fixed N N N , finite-dimensional norm equivalence bounds ∫ 0 T ∥ div u N ∥ ∞ \int_0^T\|\operatorname{div}u_N\|_\infty ∫ 0 T ∥ div u N ∥ ∞ ; positivity of ρ N \rho_N ρ N then permits global continuation.
Now let N → ∞ N\to\infty N → ∞ with α , e , σ , G \alpha,e,\sigma,G α , e , σ , G fixed. The energy identity and the equation for ρ N 2 \rho_N^2 ρ N 2 give
ρ N ∈ L ∞ L G , ρ N G / 2 ∈ L 2 H 1 , ρ N ∈ L 5 G / 3 ( Q T ) , P ^ ( ρ N ) ∈ L 5 / 3 ( Q T ) . (86) \rho_N\in L^\infty L^G,\qquad\rho_N^{G/2}\in L^2H^1,\qquad\rho_N\in L^{5G/3}(Q_T),\qquad\widehat{P}(\rho_N)\in L^{5/3}(Q_T). \tag*{(86)} ρ N ∈ L ∞ L G , ρ N G /2 ∈ L 2 H 1 , ρ N ∈ L 5 G /3 ( Q T ) , P ( ρ N ) ∈ L 5/3 ( Q T ) . (86)
1 2 d d t ∥ ρ N ∥ 2 2 + α ∥ ∇ ρ N ∥ 2 2 = − 1 2 ∫ T 3 ρ N 2 div u N (87) \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\lVert\rho_N\rVert_2^2+\alpha\lVert\nabla\rho_N\rVert_2^2=-\frac{1}{2}\int_{\mathbb{T}^3}\rho_N^2\operatorname{div}u_N \tag*{(87)} 2 1 d t d ∥ ρ N ∥ 2 2 + α ∥ ∇ ρ N ∥ 2 2 = − 2 1 ∫ T 3 ρ N 2 div u N (87)
Interpolation yields the space–time density bound, while the second identity gives a bound for α ∇ ρ N \sqrt{\alpha}\nabla\rho_N α ∇ ρ N uniform in N N N and 0 < α ≤ 1 0<\alpha\leq1 0 < α ≤ 1 . At fixed α > 0 \alpha>0 α > 0 , m N = ρ N u N ∈ L 2 L 2 m_N=\rho_Nu_N\in L^2L^2 m N = ρ N u N ∈ L 2 L 2 , and
∥ ∂ t ρ N ∥ L 2 H − 1 ≤ C ( ∥ m N ∥ L 2 L 2 + α ∥ ∇ ρ N ∥ L 2 L 2 ) ≤ C α . \lVert\partial_t\rho_N\rVert_{L^2H^{-1}}\leq C\left(\lVert m_N\rVert_{L^2L^2}+\alpha\lVert\nabla\rho_N\rVert_{L^2L^2}\right)\leq C_\alpha. ∥ ∂ t ρ N ∥ L 2 H − 1 ≤ C ( ∥ m N ∥ L 2 L 2 + α ∥ ∇ ρ N ∥ L 2 L 2 ) ≤ C α .
For the Fourier projection P L \mathbb{P}_L P L ,
∥ ( I − P L ) ρ N ∥ L 2 ( Q T ) 2 ≤ λ L + 1 − 1 ∥ ∇ ρ N ∥ L 2 ( Q T ) 2 . \lVert(I-\mathbb{P}_L)\rho_N\rVert_{L^2(Q_T)}^2\leq\lambda_{L+1}^{-1}\lVert\nabla\rho_N\rVert_{L^2(Q_T)}^2. ∥( I − P L ) ρ N ∥ L 2 ( Q T ) 2 ≤ λ L + 1 − 1 ∥ ∇ ρ N ∥ L 2 ( Q T ) 2 .
The time-compactness argument is a Fourier realization of the compactness criterion in [39 ] ; the compact spatial embeddings used here are those of [9 ] . The time derivative controls each finite set of modes, and the displayed estimate removes the remaining modes at fixed α \alpha α . Finite-mode compactness and this tail bound yield, along a subsequence,
ρ N → ρ in L q ( Q T ) , 1 ≤ q < 5 G / 3 , u N ⇀ u in L 2 H 1 . \rho_N\to\rho\quad\text{in }L^q(Q_T),\quad1\leq q<5G/3,\qquad u_N\rightharpoonup u\quad\text{in }L^2H^1. ρ N → ρ in L q ( Q T ) , 1 ≤ q < 5 G /3 , u N ⇀ u in L 2 H 1 .
Consequently, ρ N u N → ρ u \rho_Nu_N\to\rho u ρ N u N → ρ u in distributions and P ^ ( ρ N ) → P ^ ( ρ ) \widehat{P}(\rho_N)\to\widehat{P}(\rho) P ( ρ N ) → P ( ρ ) , H ( ρ N ) → H ( ρ ) H(\rho_N)\to H(\rho) H ( ρ N ) → H ( ρ ) in L 1 ( Q T ) L^1(Q_T) L 1 ( Q T ) . The limiting density belongs to L 2 H 1 L^2H^1 L 2 H 1 with ∂ t ρ ∈ L 2 H − 1 \partial_t\rho\in L^2H^{-1} ∂ t ρ ∈ L 2 H − 1 , hence ρ ∈ C ( [ 0 , T ] ; L 2 ) \rho\in C([0,T];L^2) ρ ∈ C ([ 0 , T ] ; L 2 ) and ρ ( 0 ) = ρ ∗ . \rho(0)=\rho_*. ρ ( 0 ) = ρ ∗ .
To control the momentum correction, the assumption G > 8 G>8 G > 8 yields
∥ m N ∥ L t 26 / 3 L x 2 ≤ ∥ m N ∥ L t ∞ L x 16 / 9 10 / 13 ∥ m N ∥ L t 2 L x 24 / 7 3 / 13 , m N = ρ N u N . (88) \lVert m_N\rVert_{L_t^{26/3}L_x^2}\leq\lVert m_N\rVert_{L_t^\infty L_x^{16/9}}^{10/13}\lVert m_N\rVert_{L_t^2L_x^{24/7}}^{3/13},\qquad m_N=\rho_Nu_N. \tag*{(88)} ∥ m N ∥ L t 26/3 L x 2 ≤ ∥ m N ∥ L t ∞ L x 16/9 10/13 ∥ m N ∥ L t 2 L x 24/7 3/13 , m N = ρ N u N . (88)
On the zero-mean space, set L = − Δ \mathcal{L}=-\Delta L = − Δ and Z N = L − 1 / 2 ( ρ N − m ) Z_N=\mathcal{L}^{-1/2}(\rho_N-m) Z N = L − 1/2 ( ρ N − m ) . The continuity equation becomes Z N ′ + α L Z N = N m N Z_N'+\alpha\mathcal{L}Z_N=\mathcal{N}m_N Z N ′ + α L Z N = N m N , with N : L 2 ( T 3 ; R 3 ) → L 2 ( T 3 ) \mathcal{N}:L^2(\mathbb{T}^3;\mathbb{R}^3)\to L^2(\mathbb{T}^3) N : L 2 ( T 3 ; R 3 ) → L 2 ( T 3 ) bounded. Maximal regularity [10 ] gives
∥ ∇ ρ N ∥ L t 26 / 3 L x 2 ≤ T 3 / 26 ∥ ∇ ρ ∗ ∥ 2 + C α ∥ m N ∥ L t 26 / 3 L x 2 , ∥ D N ∥ L t 13 / 8 L x 1 ≤ C α . \lVert\nabla\rho_N\rVert_{L_t^{26/3}L_x^2}\leq T^{3/26}\lVert\nabla\rho_*\rVert_2+C_\alpha\lVert m_N\rVert_{L_t^{26/3}L_x^2},\qquad\lVert D_N\rVert_{L_t^{13/8}L_x^1}\leq C_\alpha. ∥ ∇ ρ N ∥ L t 26/3 L x 2 ≤ T 3/26 ∥ ∇ ρ ∗ ∥ 2 + C α ∥ m N ∥ L t 26/3 L x 2 , ∥ D N ∥ L t 13/8 L x 1 ≤ C α .
Thus the correction has integral O ( ∣ I ∣ 5 / 13 ) O(|I|^{5/13}) O ( ∣ I ∣ 5/13 ) over a time interval I I I .
Multiplying (A.7) by T − t T-t T − t gives
α ∫ 0 T ( T − t ) ∥ ∇ ρ N ∥ 2 2 d t = T 2 ∥ ρ ∗ ∥ 2 2 − 1 2 ∫ 0 T ∥ ρ N ∥ 2 2 d t − 1 2 ∫ Q T ( T − t ) ρ N 2 div u N . (89) \alpha\int_0^T(T-t)\lVert\nabla\rho_N\rVert_2^2\,\mathrm{d}t=\frac{T}{2}\lVert\rho_*\rVert_2^2-\frac{1}{2}\int_0^T\lVert\rho_N\rVert_2^2\,\mathrm{d}t-\frac{1}{2}\int_{Q_T}(T-t)\rho_N^2\operatorname{div}u_N. \tag*{(89)} α ∫ 0 T ( T − t ) ∥ ∇ ρ N ∥ 2 2 d t = 2 T ∥ ρ ∗ ∥ 2 2 − 2 1 ∫ 0 T ∥ ρ N ∥ 2 2 d t − 2 1 ∫ Q T ( T − t ) ρ N 2 div u N . (89)
The limiting continuity equation has the same identity. Since ρ N → ρ \rho_N\to\rho ρ N → ρ in L 4 ( Q T ) L^4(Q_T) L 4 ( Q T ) and div u N ⇀ div u \operatorname{div}u_N\rightharpoonup\operatorname{div}u div u N ⇀ div u in L 2 L^2 L 2 , comparison of the two identities yields
T − t ∇ ρ N ⟶ T − t ∇ ρ in L 2 ( Q T ) . \sqrt{T-t}\nabla\rho_N\longrightarrow\sqrt{T-t}\nabla\rho\quad\text{in }L^2(Q_T). T − t ∇ ρ N ⟶ T − t ∇ ρ in L 2 ( Q T ) .
Moreover, maximal regularity gives
sup N ∫ T − h T ∥ ∇ ρ N ∥ 2 2 d t + ∫ T − h T ∥ ∇ ρ ∥ 2 2 d t ≤ C α h 10 / 13 . \sup_N\int_{T-h}^T\lVert\nabla\rho_N\rVert_2^2\,\mathrm{d}t+\int_{T-h}^T\lVert\nabla\rho\rVert_2^2\,\mathrm{d}t\leq C_\alpha h^{10/13}. N sup ∫ T − h T ∥ ∇ ρ N ∥ 2 2 d t + ∫ T − h T ∥ ∇ ρ ∥ 2 2 d t ≤ C α h 10/13 .
For 0 < h < T 0<h<T 0 < h < T , split the norm at T − h T-h T − h :
∥ ∇ ρ N − ∇ ρ ∥ L 2 ( Q T ) 2 ≤ h − 1 ∫ 0 T ( T − t ) ∥ ∇ ρ N − ∇ ρ ∥ 2 2 d t + C α h 10 / 13 . \lVert\nabla\rho_N-\nabla\rho\rVert_{L^2(Q_T)}^2 \le h^{-1}\int_0^T (T-t)\lVert\nabla\rho_N-\nabla\rho\rVert_2^2\,\mathrm{d}t+C_\alpha h^{10/13}. ∥ ∇ ρ N − ∇ ρ ∥ L 2 ( Q T ) 2 ≤ h − 1 ∫ 0 T ( T − t ) ∥ ∇ ρ N − ∇ ρ ∥ 2 2 d t + C α h 10/13 .
First let N → ∞ N\to\infty N → ∞ , then h ↓ 0 h\downarrow0 h ↓ 0 . This proves ∇ ρ N → ∇ ρ \nabla\rho_N\to\nabla\rho ∇ ρ N → ∇ ρ in L 2 ( Q T ) L^2(Q_T) L 2 ( Q T ) at fixed α \alpha α . The terminal-layer estimate is essential: the factor T − t T-t T − t alone would only give strong convergence away from T T T .
Testing the momentum equation with a fixed Fourier mode φ k ∈ X N \varphi_k\in X_N φ k ∈ X N yields, for 0 ≤ s < t ≤ T 0\le s<t\le T 0 ≤ s < t ≤ T ,
∣ ∫ T 3 [ m N ( t ) − m N ( s ) ] ⋅ φ k ∣ ≤ C k , α ( ∣ t − s ∣ + ∣ t − s ∣ 1 / 2 + ∣ t − s ∣ 5 / 13 ) . (90) \left|\int_{\mathbb{T}^3}[m_N(t)-m_N(s)]\cdot\varphi_k\right|\le C_{k,\alpha}\left(|t-s|+|t-s|^{1/2}+|t-s|^{5/13}\right). \tag*{(90)} ∫ T 3 [ m N ( t ) − m N ( s )] ⋅ φ k ≤ C k , α ( ∣ t − s ∣ + ∣ t − s ∣ 1/2 + ∣ t − s ∣ 5/13 ) . (90)
The Fourier tail estimate
∥ ( I − P L ) f ∥ H − 1 ≤ C λ L + 1 − 13 / 32 ∥ f ∥ 16 / 9 (91) \lVert(I-P_L)f\rVert_{H^{-1}}\le C\lambda_{L+1}^{-13/32}\lVert f\rVert_{16/9} \tag*{(91)} ∥( I − P L ) f ∥ H − 1 ≤ C λ L + 1 − 13/32 ∥ f ∥ 16/9 (91)
follows by interpolation. Together with Arzelà–Ascoli, it gives m N → ρ u m_N\to\rho u m N → ρ u in C ( [ 0 , T ] ; H − 1 ) C([0,T];H^{-1}) C ([ 0 , T ] ; H − 1 ) , with initial value m ∗ m_* m ∗ . For a smooth matrix field Ψ \Psi Ψ ,
∫ Q T ( ρ N u N ⊗ u N − ρ u ⊗ u ) : Ψ = ∫ 0 T ⟨ m N − ρ u , Ψ u N ⟩ d t + ∫ 0 T ⟨ ρ u , Ψ ( u N − u ) ⟩ d t ⟶ 0. (92) \int_{Q_T}(\rho_Nu_N\otimes u_N-\rho u\otimes u):\Psi=\int_0^T\langle m_N-\rho u,\Psi u_N\rangle\,\mathrm{d}t
+\int_0^T\langle\rho u,\Psi(u_N-u)\rangle\,\mathrm{d}t\longrightarrow0. \tag*{(92)} ∫ Q T ( ρ N u N ⊗ u N − ρ u ⊗ u ) : Ψ = ∫ 0 T ⟨ m N − ρ u , Ψ u N ⟩ d t + ∫ 0 T ⟨ ρ u , Ψ ( u N − u )⟩ d t ⟶ 0. (92)
The H − 1 H^{-1} H − 1 –H 1 H^1 H 1 pairing is valid for smooth Ψ \Psi Ψ . Identifying the weak limit of ρ N u N \sqrt{\rho_N}u_N ρ N u N and taking Ψ = Id \Psi=\operatorname{Id} Ψ = Id gives convergence of its L 2 L^2 L 2 norm. Hence
ρ N u N → ρ u in L 2 ( Q T ) , ρ N u N ⊗ u N → ρ u ⊗ u in L 1 ( Q T ) . (93) \sqrt{\rho_N}u_N\to\sqrt{\rho}u\quad\text{in }L^2(Q_T),\qquad\rho_Nu_N\otimes u_N\to\rho u\otimes u\quad\text{in }L^1(Q_T). \tag*{(93)} ρ N u N → ρ u in L 2 ( Q T ) , ρ N u N ⊗ u N → ρ u ⊗ u in L 1 ( Q T ) . (93)
Approximation of smooth tests by Fourier polynomials recovers the full momentum equation and its initial term. The strong convergence of ∇ ρ N \nabla\rho_N ∇ ρ N and weak convergence of ∇ u N \nabla u_N ∇ u N identify the correction as D i = α ∂ j ρ ∂ j u i D_i=\alpha\partial_j\rho\partial_j u_i D i = α ∂ j ρ ∂ j u i .
For b ∈ C 2 ( [ 0 , ∞ ) ) b\in C^2([0,\infty)) b ∈ C 2 ([ 0 , ∞ )) with compactly supported b ′ b' b ′ , the diffusive chain rule passes to the limit:
∂ t b ( ρ ) + div ( b ( ρ ) u ) + ( ρ b ′ ( ρ ) − b ( ρ ) ) div u = α div ( b ′ ( ρ ) ∇ ρ ) − α b ′ ′ ( ρ ) ∣ ∇ ρ ∣ 2 , (94) \partial_t b(\rho)+\operatorname{div}(b(\rho)u)+(\rho b'(\rho)-b(\rho))\operatorname{div}u=\alpha\operatorname{div}(b'(\rho)\nabla\rho)-\alpha b''(\rho)|\nabla\rho|^2, \tag*{(94)} ∂ t b ( ρ ) + div ( b ( ρ ) u ) + ( ρ b ′ ( ρ ) − b ( ρ )) div u = α div ( b ′ ( ρ ) ∇ ρ ) − α b ′′ ( ρ ) ∣∇ ρ ∣ 2 , (94)
with initial value b ( ρ ∗ ) b(\rho_*) b ( ρ ∗ ) . The case b ( 0 ) ≠ 0 b(0)\ne0 b ( 0 ) = 0 follows by adding the constant transport identity. Finally, convergence of the initial kinetic energy and of the energy densities, followed by lower semicontinuity of dissipation in (A.5), gives the energy inequality with the prescribed initial energy.
Step 2: Density diffusion α ↓ 0 \alpha\downarrow0 α ↓ 0 . Fix e e e , σ \sigma σ , G G G and let α ↓ 0 \alpha\downarrow0 α ↓ 0 . The energy bounds give
ρ α ∈ L ∞ L 8 , m α ∈ L ∞ L 16 / 9 ∩ L 2 L 2 , ∥ D α ∥ 1 + ∥ α ∇ ρ α ∥ 2 ≤ C α . \rho_\alpha\in L^\infty L^8,\qquad m_\alpha\in L^\infty L^{16/9}\cap L^2L^2,\qquad\lVert D_\alpha\rVert_1+\lVert\alpha\nabla\rho_\alpha\rVert_2\le C\sqrt{\alpha}. ρ α ∈ L ∞ L 8 , m α ∈ L ∞ L 16/9 ∩ L 2 L 2 , ∥ D α ∥ 1 + ∥ α ∇ ρ α ∥ 2 ≤ C α .
For example,
∥ D α ∥ 1 ≤ α ∥ α ∇ ρ α ∥ 2 ∥ ∇ u α ∥ 2 , ∥ α ∇ ρ α ∥ 2 = α ∥ α ∇ ρ α ∥ 2 . \lVert D_\alpha\rVert_1\le\sqrt{\alpha}\lVert\sqrt{\alpha}\nabla\rho_\alpha\rVert_2\lVert\nabla u_\alpha\rVert_2,\qquad\lVert\alpha\nabla\rho_\alpha\rVert_2=\sqrt{\alpha}\lVert\sqrt{\alpha}\nabla\rho_\alpha\rVert_2. ∥ D α ∥ 1 ≤ α ∥ α ∇ ρ α ∥ 2 ∥ ∇ u α ∥ 2 , ∥ α ∇ ρ α ∥ 2 = α ∥ α ∇ ρ α ∥ 2 .
The square-density identity (A.7) bounds α ∇ ρ α \sqrt{\alpha}\nabla\rho_\alpha α ∇ ρ α using ρ α ∈ L ∞ L 8 \rho_\alpha\in L^\infty L^8 ρ α ∈ L ∞ L 8 and div u α ∈ L 2 \operatorname{div}u_\alpha\in L^2 div u α ∈ L 2 . Thus the artificial diffusion and momentum correction vanish. These estimates do not require a uniform bound for ∇ ρ α \nabla\rho_\alpha ∇ ρ α itself. Set q α = A ρ α q_\alpha=A\rho_\alpha q α = A ρ α and d α = div u α d_\alpha=\operatorname{div}u_\alpha d α = div u α .
The effective-flux test follows [14 , 15 , 27 ] . Here A = ∇ Δ − 1 A=\nabla\Delta^{-1} A = ∇ Δ − 1 is the periodic inverse with zero Fourier coefficient at the origin, so
div A ρ α = ρ α − m . \operatorname{div} A\rho_\alpha=\rho_\alpha-\mathbf{m}. div A ρ α = ρ α − m .
This identity produces the spatial mean term on the right below. Testing the momentum equation with the periodic potential η q α \eta q_\alpha η q α , where η ∈ C c ∞ ( 0 , T ) \eta\in C^\infty_c(0,T) η ∈ C c ∞ ( 0 , T ) , yields
∫ η ( P ^ ( ρ α ) − ν d α ) ρ α = m ∫ η P ^ ( ρ α ) − ∫ η ′ m α ⋅ q α + ∫ η ψ α ⋅ B α + E α , (95) \int\eta\bigl(\widehat{P}(\rho_\alpha)-\nu d_\alpha\bigr)\rho_\alpha=\mathbf{m}\int\eta\widehat{P}(\rho_\alpha)-\int\eta' \mathbf{m}_\alpha\cdot q_\alpha
+\int\eta\psi_\alpha\cdot B_\alpha+\mathcal{E}_\alpha, \tag*{(95)} ∫ η ( P ( ρ α ) − ν d α ) ρ α = m ∫ η P ( ρ α ) − ∫ η ′ m α ⋅ q α + ∫ η ψ α ⋅ B α + E α , (95)
B α , i = ρ α R i j m α , j − m α , j R i j ρ α , B_{\alpha,i}=\rho_\alpha R_{ij}m_{\alpha,j}-m_{\alpha,j}R_{ij}\rho_\alpha, B α , i = ρ α R ij m α , j − m α , j R ij ρ α ,
E α = ∫ η D α ⋅ q α − α ∫ η m α ⋅ ∇ ρ α . \mathcal{E}_\alpha=\int\eta D_\alpha\cdot q_\alpha-\alpha\int\eta\mathbf{m}_\alpha\cdot\nabla\rho_\alpha. E α = ∫ η D α ⋅ q α − α ∫ η m α ⋅ ∇ ρ α .
All integrals without specified domains are over Q T Q_T Q T . At fixed α \alpha α , the integrability in (A.6) justifies the test by smooth approximation. The energy bounds and A : L 8 → W 1 , 8 ↪ L ∞ A:L^8\to W^{1,8}\hookrightarrow L^\infty A : L 8 → W 1 , 8 ↪ L ∞ control the time and convective terms, while the stress term is bounded in L 1 L^1 L 1 and E α = O ( α ) \mathcal{E}_\alpha=O(\sqrt{\alpha}) E α = O ( α ) . Removing the time cutoff gives
∫ Q T P ^ ( ρ α ) ρ α ≤ C , ρ α ∈ L G + 1 ( Q T ) , P ^ ( ρ α ) ∈ L ( G + 1 ) / G ( Q T ) , (96) \int_{Q_T}\widehat{P}(\rho_\alpha)\rho_\alpha\le C,\qquad\rho_\alpha\in L^{G+1}(Q_T),\qquad\widehat{P}(\rho_\alpha)\in L^{(G+1)/G}(Q_T), \tag*{(96)} ∫ Q T P ( ρ α ) ρ α ≤ C , ρ α ∈ L G + 1 ( Q T ) , P ( ρ α ) ∈ L ( G + 1 ) / G ( Q T ) , (96)
uniformly in α \alpha α .
The continuity equation bounds ∂ t ρ α \partial_t\rho_\alpha ∂ t ρ α in L 2 H − 1 L^2H^{-1} L 2 H − 1 . Together with the uniform density bounds and Fourier compactness, this gives ρ α → ρ \rho_\alpha\to\rho ρ α → ρ in C ( [ 0 , T ] ; H − 1 ) C([0,T];H^{-1}) C ([ 0 , T ] ; H − 1 ) . For each fixed Fourier mode φ k \varphi_k φ k ,
sup 0 ≤ t ≤ T ∣ ∫ 0 t ∫ T 3 D α ⋅ φ k ∣ ≤ C ( k ) α , \sup_{0\le t\le T}\left|\int_0^t\int_{\mathbb{T}^3}D_\alpha\cdot\varphi_k\right|\le C(k)\sqrt{\alpha}, 0 ≤ t ≤ T sup ∫ 0 t ∫ T 3 D α ⋅ φ k ≤ C ( k ) α ,
∣ ∫ T 3 [ m α ( t ) − m α ( s ) ] ⋅ φ k ∣ ≤ C ( k ) ( ∣ t − s ∣ + ∣ t − s ∣ 1 / 2 + α ) . (97) \left|\int_{\mathbb{T}^3}[\mathbf{m}_\alpha(t)-\mathbf{m}_\alpha(s)]\cdot\varphi_k\right|\le C(k)\bigl(|t-s|+|t-s|^{1/2}+\sqrt{\alpha}\bigr). \tag*{(97)} ∫ T 3 [ m α ( t ) − m α ( s )] ⋅ φ k ≤ C ( k ) ( ∣ t − s ∣ + ∣ t − s ∣ 1/2 + α ) . (97)
Combining these bounds with (A.11) and the uniform L ∞ L 16 / 9 L^\infty L^{16/9} L ∞ L 16/9 estimate yields
m α → m in C ( [ 0 , T ] ; H − 1 ) , m ( 0 ) = m ∗ . \mathbf{m}_\alpha\to\mathbf{m}\quad\text{in }C([0,T];H^{-1}),\qquad\mathbf{m}(0)=\mathbf{m}_*. m α → m in C ([ 0 , T ] ; H − 1 ) , m ( 0 ) = m ∗ .
Since u α ⇀ u u_\alpha\rightharpoonup u u α ⇀ u in L 2 H 1 L^2H^1 L 2 H 1 , the density convergence gives, for every smooth vector field ϕ \phi ϕ ,
∫ Q T ( ρ α u α − ρ u ) ⋅ ϕ = ∫ 0 T ⟨ ρ α − ρ , u α ⋅ ϕ ⟩ d t + ∫ 0 T ⟨ ρ , ( u α − u ) ⋅ ϕ ⟩ d t ⟶ 0. (98) \int_{Q_T}(\rho_\alpha u_\alpha-\rho u)\cdot\phi=\int_0^T\langle\rho_\alpha-\rho,u_\alpha\cdot\phi\rangle\,dt
+\int_0^T\langle\rho,(u_\alpha-u)\cdot\phi\rangle\,dt\longrightarrow0. \tag*{(98)} ∫ Q T ( ρ α u α − ρ u ) ⋅ ϕ = ∫ 0 T ⟨ ρ α − ρ , u α ⋅ ϕ ⟩ d t + ∫ 0 T ⟨ ρ , ( u α − u ) ⋅ ϕ ⟩ d t ⟶ 0. (98)
Hence m = ρ u \mathbf{m}=\rho u m = ρ u . The same H − 1 H^{-1} H − 1 –H 1 H^1 H 1 pairing identifies the convective term as in (A.12). Passing to the limit in the initialized equations also recovers conservation of total momentum.
We next pass to the effective-flux identity. Lemma A.2, applied with f j = ρ α f_j=\rho_\alpha f j = ρ α , g j = m α g_j=\mathbf{m}_\alpha g j = m α , a = 8 a=8 a = 8 , and b = 16 / 9 b=16/9 b = 16/9 , gives
B α → B strongly in L 2 ( 0 , T ; H − 1 ( T 3 ) ) , B_\alpha\to B\quad\text{strongly in }L^2(0,T;H^{-1}(\mathbb{T}^3)), B α → B strongly in L 2 ( 0 , T ; H − 1 ( T 3 )) ,
since 1 / 8 + 9 / 16 < 5 / 6 1/8 + 9/16 < 5/6 1/8 + 9/16 < 5/6 . This convergence permits passage to the term involving u α ⋅ B α u_\alpha\cdot B_\alpha u α ⋅ B α .
Write
z = w − l i m α ↓ 0 ρ α d α , Π = w ∗ − l i m α ↓ 0 P ^ ( ρ α ) ρ α , z = \mathop{\mathrm{w\!-\!lim}}_{\alpha\downarrow0} \rho_\alpha d_\alpha,\qquad\Pi= \mathop{\mathrm{w^*\!-\!lim}}_{\alpha\downarrow0} \widehat{P}(\rho_\alpha)\rho_\alpha, z = w − lim α ↓ 0 ρ α d α , Π = w ∗ − lim α ↓ 0 P ( ρ α ) ρ α ,
where Π \Pi Π is a nonnegative measure. Let Y t , x Y_{t,x} Y t , x be the density Young measure, obtained from the fundamental theorem for Young measures (see [3 ] ). The product P ^ ( ρ α ) ρ α \widehat{P}(\rho_\alpha)\rho_\alpha P ( ρ α ) ρ α is only bounded in L 1 L^1 L 1 ; its measure limit may contain a concentration part. Consequently the first relation below is an inequality, even though the separate pressure limit is represented by Y t , x Y_{t,x} Y t , x . By uniform integrability,
ρ = ⟨ s ⟩ , P ^ = ⟨ P ^ ( s ) ⟩ . \rho= \langle s\rangle,\qquad\widehat{P} = \langle\widehat{P}(s)\rangle. ρ = ⟨ s ⟩ , P = ⟨ P ( s )⟩ .
Pressure monotonicity gives, as measures,
Π ≥ ⟨ P ^ ( s ) s ⟩ d t d x , \Pi\ge\langle\widehat{P}(s)s\rangle\,dt\,dx, Π ≥ ⟨ P ( s ) s ⟩ d t d x ,
⟨ P ^ ( s ) s ⟩ − ⟨ P ^ ( s ) ⟩ ⟨ s ⟩ = 1 2 ∬ [ P ^ ( s ) − P ^ ( t ) ] ( s − t ) d Y ( s ) d Y ( t ) ≥ 0. \langle\widehat{P}(s)s\rangle-\langle\widehat{P}(s)\rangle\langle s\rangle
= \frac{1}{2}\iint[\widehat{P}(s)-\widehat{P}(t)](s-t)\,dY(s)\,dY(t)\ge0. ⟨ P ( s ) s ⟩ − ⟨ P ( s )⟩ ⟨ s ⟩ = 2 1 ∬ [ P ( s ) − P ( t )] ( s − t ) d Y ( s ) d Y ( t ) ≥ 0.
Passing to the limit in (A.15) and subtracting the corresponding identity for the limiting momentum equation yields
∫ η d Π − ∫ η P ^ ρ d t d x = ν ∫ η ( z − ρ d ) d t d x . \int\eta\,d\Pi-\int\eta\widehat{P}\rho\,dt\,dx
=\nu\int\eta(z-\rho d)\,dt\,dx. ∫ η d Π − ∫ η P ρ d t d x = ν ∫ η ( z − ρ d ) d t d x .
The left-hand side is nonnegative for η ≥ 0 \eta\ge0 η ≥ 0 . Since ν > 0 \nu>0 ν > 0 ,
∫ T 3 ( z − ρ d ) d x ≥ 0 for almost every t . (99) \int_{\mathbb{T}^3}(z-\rho d)\,dx\ge0\qquad\text{for almost every }t. \tag*{(99)} ∫ T 3 ( z − ρ d ) d x ≥ 0 for almost every t . (99)
We now compare the entropy equations. Regularize s log s s\log s s log s by
b ε ( s ) = ( s + ε ) log ( s + ε ) − ε log ε b_\varepsilon(s)=(s+\varepsilon)\log(s+\varepsilon)-\varepsilon\log\varepsilon b ε ( s ) = ( s + ε ) log ( s + ε ) − ε log ε
and truncate its second derivative at high densities. More explicitly, choose 0 ≤ θ R ≤ 1 0\le\theta_R\le1 0 ≤ θ R ≤ 1 equal to one on [ 0 , R ] [0,R] [ 0 , R ] and zero on [ 2 R , ∞ ) [2R,\infty) [ 2 R , ∞ ) , and replace b ε ′ ′ ( s ) = ( s + ε ) − 1 b_\varepsilon''(s)=(s+\varepsilon)^{-1} b ε ′′ ( s ) = ( s + ε ) − 1 by θ R ( s ) / ( s + ε ) \theta_R(s)/(s+\varepsilon) θ R ( s ) / ( s + ε ) . Twice integration, with the same value and derivative at zero, gives a convex function affine for s ≥ 2 R s\ge2R s ≥ 2 R . Subtracting this affine part makes its derivative compactly supported. Convexity preserves the sign of the diffusion term − α b ′ ′ ∣ ∇ ρ α ∣ 2 -\alpha b''|\nabla\rho_\alpha|^2 − α b ′′ ∣∇ ρ α ∣ 2 . At this stage ρ α ∈ L ∞ L 8 \rho_\alpha\in L^\infty L^8 ρ α ∈ L ∞ L 8 and div u α ∈ L 2 \operatorname{div}u_\alpha\in L^2 div u α ∈ L 2 , so the high-density errors in the entropy and source vanish as R → ∞ R\to\infty R → ∞ ; then ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 . The initial data are fixed throughout these two limits. Applying (A.14), restoring the affine part through the continuity equation, and removing the truncations gives
− ∫ ψ ′ ∫ T 3 ρ α log ρ α + ∫ ψ ρ α d α ≤ ψ ( 0 ) ∫ T 3 ρ ∗ log ρ ∗ , ψ ≥ 0. (100) -\int\psi'\int_{\mathbb{T}^3}\rho_\alpha\log\rho_\alpha+\int\psi\rho_\alpha d_\alpha
\le\psi(0)\int_{\mathbb{T}^3}\rho_*\log\rho_*,
\qquad\psi\ge0. \tag*{(100)} − ∫ ψ ′ ∫ T 3 ρ α log ρ α + ∫ ψ ρ α d α ≤ ψ ( 0 ) ∫ T 3 ρ ∗ log ρ ∗ , ψ ≥ 0. (100)
For the limiting continuity equation, the initialized renormalized identity follows from the spatial commutator argument [11 ] , after extension to negative times. Passing to the weak entropy limits gives
− ∫ 0 T ψ ′ ∫ T 3 ρ log ρ + ∫ Q T ψ z ≤ ψ ( 0 ) ∫ T 3 ρ ∗ log ρ ∗ , -\int_0^T\psi'\int_{\mathbb{T}^3}\rho\log\rho+\int_{Q_T}\psi z
\le\psi(0)\int_{\mathbb{T}^3}\rho_*\log\rho_*, − ∫ 0 T ψ ′ ∫ T 3 ρ log ρ + ∫ Q T ψ z ≤ ψ ( 0 ) ∫ T 3 ρ ∗ log ρ ∗ ,
− ∫ 0 T ψ ′ ∫ T 3 ρ log ρ + ∫ Q T ψ ρ d = ψ ( 0 ) ∫ T 3 ρ ∗ log ρ ∗ . (101) -\int_0^T\psi'\int_{\mathbb{T}^3}\rho\log\rho+\int_{Q_T}\psi\rho d
=\psi(0)\int_{\mathbb{T}^3}\rho_*\log\rho_*. \tag*{(101)} − ∫ 0 T ψ ′ ∫ T 3 ρ log ρ + ∫ Q T ψ ρ d = ψ ( 0 ) ∫ T 3 ρ ∗ log ρ ∗ . (101)
Set
D ( t ) = ∫ T 3 ( ρ log ρ ‾ − ρ log ρ ) d x ≥ 0. \mathcal{D}(t)=\int_{\mathbb{T}^3}\left(\overline{\rho\log\rho}-\rho\log\rho\right)\,dx\ge0. D ( t ) = ∫ T 3 ( ρ log ρ − ρ log ρ ) d x ≥ 0.
Subtracting the two entropy relations and taking ψ ( t ) = ∫ t T ζ ( s ) d s \psi(t)=\int_t^T\zeta(s)\,ds ψ ( t ) = ∫ t T ζ ( s ) d s with 0 ≤ ζ ∈ C c ∞ ( 0 , T ) 0\le\zeta\in C^\infty_c(0,T) 0 ≤ ζ ∈ C c ∞ ( 0 , T ) gives
0 ≥ ∫ 0 T ζ ( t ) D ( t ) d t + ∫ Q T ψ ( z − ρ d ) d x d t ≥ 0 , 0\ge\int_0^T\zeta(t)\mathcal{D}(t)\,dt+\int_{Q_T}\psi(z-\rho d)\,dx\,dt\ge0, 0 ≥ ∫ 0 T ζ ( t ) D ( t ) d t + ∫ Q T ψ ( z − ρ d ) d x d t ≥ 0 ,
where the last inequality follows from (A.19). Thus D = 0 \mathcal{D}=0 D = 0 almost everywhere. Strict convexity of s log s s\log s s log s eliminates density oscillations, and the uniform L G + 1 L^{G+1} L G + 1 bound yields
ρ α → ρ in L q ( Q T ) , 1 ≤ q < G + 1 , \rho_\alpha\to\rho\quad\text{in }L^q(Q_T),\qquad1\le q<G+1, ρ α → ρ in L q ( Q T ) , 1 ≤ q < G + 1 ,
and
P ^ ( ρ α ) → P ^ ( ρ ) in L s ( Q T ) , 1 ≤ s < G + 1 G . \widehat{P}(\rho_\alpha)\to\widehat{P}(\rho)\quad\text{in }L^s(Q_T),\qquad1\le s<\frac{G+1}{G}. P ( ρ α ) → P ( ρ ) in L s ( Q T ) , 1 ≤ s < G G + 1 .
Finally, the momentum pairing (A.12) gives strong convergence of the weighted velocities in L 2 ( Q T ) L^2(Q_T) L 2 ( Q T ) . Convergence of the kinetic and internal energies, together with lower semicontinuity of the viscous dissipation, yields the initialized energy inequality.
Step 3: Auxiliary pressure σ ↓ 0 \sigma\downarrow0 σ ↓ 0 . Fix e e e , G G G and the smooth initial data, and let σ j ↓ 0 \sigma_j\downarrow0 σ j ↓ 0 . The energy inequality gives, uniformly in j j j ,
ρ j ∈ L ∞ L b ∗ , m j ∈ L ∞ L 3 / 2 ∩ L 2 L 2 , u j ∈ L 2 H 1 , ρ j u j ⊗ u j ∈ L 1 L 3 / 2 . \rho_j\in L^\infty L^{b^*},\qquad m_j\in L^\infty L^{3/2}\cap L^2L^2,\qquad u_j\in L^2H^1,\qquad\rho_j u_j\otimes u_j\in L^1L^{3/2}. ρ j ∈ L ∞ L b ∗ , m j ∈ L ∞ L 3/2 ∩ L 2 L 2 , u j ∈ L 2 H 1 , ρ j u j ⊗ u j ∈ L 1 L 3/2 .
Testing the momentum equation with η A ρ j \eta A\rho_j η A ρ j , using A : L 3 → W 1 , 3 ↪ L 3 A:L^3\to W^{1,3}\hookrightarrow L^3 A : L 3 → W 1 , 3 ↪ L 3 and the uniform energy bounds, gives
∫ Q T ( ρ j γ + 1 + e ρ j 4 + σ j ρ j G + 1 ) ≤ C , ∥ σ j ρ j G ∥ ( G + 1 ) / G ≤ C σ j 1 / ( G + 1 ) ⟶ 0. (102) \int_{Q_T}\left(\rho_j^{\gamma+1}+e\rho_j^4+\sigma_j\rho_j^{G+1}\right)\le C,\qquad\|\sigma_j\rho_j^G\|_{(G+1)/G}\le C\sigma_j^{1/(G+1)}\longrightarrow0. \tag*{(102)} ∫ Q T ( ρ j γ + 1 + e ρ j 4 + σ j ρ j G + 1 ) ≤ C , ∥ σ j ρ j G ∥ ( G + 1 ) / G ≤ C σ j 1/ ( G + 1 ) ⟶ 0. (102)
The artificial-pressure norm in (A.22) follows directly from
∫ Q T ∣ σ j ρ j G ∣ ( G + 1 ) / G = σ j 1 / G ∫ Q T σ j ρ j G + 1 ≤ C σ j 1 / G . \int_{Q_T}|\sigma_j\rho_j^G|^{(G+1)/G}=\sigma_j^{1/G}\int_{Q_T}\sigma_j\rho_j^{G+1}\le C\sigma_j^{1/G}. ∫ Q T ∣ σ j ρ j G ∣ ( G + 1 ) / G = σ j 1/ G ∫ Q T σ j ρ j G + 1 ≤ C σ j 1/ G .
Thus the pressure itself vanishes strongly, rather than merely acquiring a small coefficient in front of a possibly unbounded density power. In particular, ρ j \rho_j ρ j is bounded in L b ∗ + 1 ( Q T ) L^{b^*+1}(Q_T) L b ∗ + 1 ( Q T ) and P e ( ρ j ) P_e(\rho_j) P e ( ρ j ) in L ( b ∗ + 1 ) / b ∗ ( Q T ) L^{(b^*+1)/b^*}(Q_T) L ( b ∗ + 1 ) / b ∗ ( Q T ) .
The continuity equation bounds ∂ t ρ j \partial_t\rho_j ∂ t ρ j in L 2 H − 1 L^2H^{-1} L 2 H − 1 . The Fourier-mode estimates for momentum, now without the diffusion correction, and
∥ ( I − P L ) f ∥ H − 1 ≤ C λ L + 1 − 1 / 4 ∥ f ∥ 3 / 2 \|(I-\mathbb{P}_L)f\|_{H^{-1}}\le C\lambda_{L+1}^{-1/4}\|f\|_{3/2} ∥ ( I − P L ) f ∥ H − 1 ≤ C λ L + 1 − 1/4 ∥ f ∥ 3/2
yield, after extraction, ρ j → ρ \rho_j\to\rho ρ j → ρ and m j → m m_j\to m m j → m in C ( [ 0 , T ] ; H − 1 ) C([0,T];H^{-1}) C ([ 0 , T ] ; H − 1 ) , with initial values ρ ∗ , m ∗ \rho_*,m_* ρ ∗ , m ∗ . The pairings in (A.18) and (A.12) give m = ρ u m=\rho u m = ρ u and identify the convective limit. Denote the weak pressure limit by P ‾ e \overline{P}_e P e .
For fixed k k k , set g j = T k ( ρ j ) g_j=T_k(\rho_j) g j = T k ( ρ j ) , h j = a k ( ρ j ) d j h_j=a_k(\rho_j)d_j h j = a k ( ρ j ) d j , and q j = A g j q_j=Ag_j q j = A g j . The initialized truncated equation satisfies
\begin{cases}
\partial_tg_j+\operatorname{div}(g_ju_j)+h_j=0,\\
g_j(0)=T_k(\rho_*),
\end{cases}
\qquad\|\partial_tg_j\|_{L^2H^{-1}}+\|h_j\|_2\le C(k). \tag*{(103)}\end{cases}
Compactness gives g j → β k g_j \to\beta_k g j → β k in C ( [ 0 , T ] ; H − 1 ) C([0,T];H^{-1}) C ([ 0 , T ] ; H − 1 ) . Taking h j ⇀ h k h_j \rightharpoonup h_k h j ⇀ h k in L 2 L^2 L 2 and passing to the limit, we obtain
∂ t β k + div ( β k u ) + h k = 0 , β k ( 0 ) = T k ( ρ ∗ ) . \partial_t\beta_k+\operatorname{div}(\beta_k u)+h_k=0,\qquad\beta_k(0)=T_k(\rho_*). ∂ t β k + div ( β k u ) + h k = 0 , β k ( 0 ) = T k ( ρ ∗ ) .
Since A , A ∗ : H − 1 → L 2 A,A^*:H^{-1}\to L^2 A , A ∗ : H − 1 → L 2 are bounded, q j → A β k q_j\to A\beta_k q j → A β k and A ∗ m j → A ∗ m A^*m_j\to A^*m A ∗ m j → A ∗ m in C t L 2 C_tL^2 C t L 2 . The uniform L ∞ W 1 , p L^\infty W^{1,p} L ∞ W 1 , p bounds for q j q_j q j , with p > 3 p>3 p > 3 , also give uniform convergence of q j q_j q j . Hence
∫ Q T η m j ⋅ A h j = ∫ Q T η ( A ∗ m j ) h j ⟶ ∫ Q T η ( A ∗ m ) h k . \int_{Q_T}\eta m_j\cdot Ah_j=\int_{Q_T}\eta(A^*m_j)h_j\longrightarrow\int_{Q_T}\eta(A^*m)h_k. ∫ Q T η m j ⋅ A h j = ∫ Q T η ( A ∗ m j ) h j ⟶ ∫ Q T η ( A ∗ m ) h k .
For B j , l = g j R i l m j , i − m j , i R i l g j B_{j,l}=g_jR_{il}m_{j,i}-m_{j,i}R_{il}g_j B j , l = g j R i l m j , i − m j , i R i l g j , Lemma A.2, with a = 12 a=12 a = 12 and b = 3 / 2 b=3/2 b = 3/2 , gives
B j ⟶ B in L t 2 H x − 1 , B_j\longrightarrow B\quad\text{in }L_t^2H_x^{-1}, B j ⟶ B in L t 2 H x − 1 ,
since 1 / 12 + 2 / 3 < 5 / 6 1/12+2/3<5/6 1/12 + 2/3 < 5/6 . Testing the momentum equation with η q j \eta q_j η q j yields
∫ η ( P ^ j − ν d j ) ( g j − g ˉ j ) = − ∫ η ′ m j ⋅ q j + ∫ η u j , l B j , l + ∫ η m j ⋅ A h j . (104) \int\eta(\widehat P_j-\nu d_j)(g_j-\bar g_j)=-\int\eta'm_j\cdot q_j+\int\eta u_{j,l}B_{j,l}+\int\eta m_j\cdot Ah_j. \tag*{(104)} ∫ η ( P j − ν d j ) ( g j − g ˉ j ) = − ∫ η ′ m j ⋅ q j + ∫ η u j , l B j , l + ∫ η m j ⋅ A h j . (104)
Regularization justifies this test at fixed j , k j,k j , k . The limiting test η A β k \eta A\beta_k η A β k is admissible since its time derivative lies in L 2 L 2 L^2L^2 L 2 L 2 and its gradient in L ∞ L p L^\infty L^p L ∞ L p for finite p p p . Passing to the limit and subtracting the two identities cancels the right-hand sides; the mean terms cancel because g ˉ j → β ˉ k \bar g_j\to\bar\beta_k g ˉ j → β ˉ k uniformly in time. Moreover, ∥ σ j ρ j G T k ( ρ j ) ∥ 1 ≤ 2 k ∥ σ j ρ j G ∥ 1 → 0 \|\sigma_j\rho_j^G T_k(\rho_j)\|_1\le2k\|\sigma_j\rho_j^G\|_1\to0 ∥ σ j ρ j G T k ( ρ j ) ∥ 1 ≤ 2 k ∥ σ j ρ j G ∥ 1 → 0 . Thus
lim j ∫ η [ P ^ e ( ρ j ) − ν d j ] T k ( ρ j ) = ∫ η [ P ‾ e − ν d ] β k . (105) \lim_{j}\int\eta[\widehat P_e(\rho_j)-\nu d_j]T_k(\rho_j)=\int\eta[\overline P_e-\nu d]\beta_k. \tag*{(105)} j lim ∫ η [ P e ( ρ j ) − ν d j ] T k ( ρ j ) = ∫ η [ P e − ν d ] β k . (105)
Let Y Y Y be the density Young measure and write Γ k = ⟨ P e T k ⟩ \Gamma_k=\langle P_eT_k\rangle Γ k = ⟨ P e T k ⟩ and β k = ⟨ T k ⟩ \beta_k=\langle T_k\rangle β k = ⟨ T k ⟩ . Monotonicity of P e P_e P e and T k T_k T k gives
Γ k − P ‾ e β k = 1 2 ∬ [ P e ( s ) − P e ( t ) ] [ T k ( s ) − T k ( t ) ] d Y ( s ) d Y ( t ) ≥ 0. \Gamma_k-\overline P_e\beta_k=\frac{1}{2}\iint[P_e(s)-P_e(t)][T_k(s)-T_k(t)]\,\mathrm{d}Y(s)\,\mathrm{d}Y(t)\ge0. Γ k − P e β k = 2 1 ∬ [ P e ( s ) − P e ( t )] [ T k ( s ) − T k ( t )] d Y ( s ) d Y ( t ) ≥ 0.
By (A.25), its spatial integral is ν ∫ ( z k − β k d ) \nu\int(z_k-\beta_kd) ν ∫ ( z k − β k d ) , where z k = w-lim j T k ( ρ j ) d j z_k=\operatorname*{w-lim}_j T_k(\rho_j)d_j z k = w-lim j T k ( ρ j ) d j . The L 4 L^4 L 4 bound in (A.22) gives
∥ ρ j − T k ( ρ j ) ∥ 2 ≤ C / k , ∥ ( ρ j − T k ( ρ j ) ) d j ∥ 1 ≤ C / k . \|\rho_j-T_k(\rho_j)\|_2\le C/k,\qquad\|(\rho_j-T_k(\rho_j))d_j\|_1\le C/k. ∥ ρ j − T k ( ρ j ) ∥ 2 ≤ C / k , ∥ ( ρ j − T k ( ρ j )) d j ∥ 1 ≤ C / k .
Writing z = w-lim j ρ j d j z=\operatorname*{w-lim}_j\rho_jd_j z = w-lim j ρ j d j in L 4 / 3 ( Q T ) L^{4/3}(Q_T) L 4/3 ( Q T ) , we also have
∥ ρ − β k ∥ 2 ≤ C / k , ∥ z − z k ∥ 1 + ∥ ( β k − ρ ) d ∥ 1 ≤ C / k . \|\rho-\beta_k\|_2\le C/k,\qquad\|z-z_k\|_1+\|(\beta_k-\rho)d\|_1\le C/k. ∥ ρ − β k ∥ 2 ≤ C / k , ∥ z − z k ∥ 1 + ∥ ( β k − ρ ) d ∥ 1 ≤ C / k .
First letting j → ∞ j\to\infty j → ∞ at fixed k k k and then k → ∞ k\to\infty k → ∞ recovers (A.19). The initialized renormalized entropy equations for ρ j \rho_j ρ j and ρ \rho ρ (using the negative-time extension for the limit), as in (A.21), have the same initial term. Subtracting them and testing in time with ψ ( t ) = ∫ t T ζ ( s ) d s \psi(t)=\int_t^T\zeta(s)\,\mathrm{d}s ψ ( t ) = ∫ t T ζ ( s ) d s , ζ ≥ 0 \zeta\ge0 ζ ≥ 0 , the integrated sign forces the entropy defect to vanish. Strict convexity yields
ρ j ⟶ ρ in L q ( Q T ) , 1 ≤ q < b ∗ + 1 , \rho_j\longrightarrow\rho\quad\text{in }L^q(Q_T),\qquad1\le q<b_*+1, ρ j ⟶ ρ in L q ( Q T ) , 1 ≤ q < b ∗ + 1 ,
and
P e ( ρ j ) ⟶ P e ( ρ ) in L s ( Q T ) , 1 ≤ s < b ∗ + 1 b ∗ . P_e(\rho_j)\longrightarrow P_e(\rho)\quad\text{in }L^s(Q_T),\qquad1\le s<\frac{b_*+1}{b_*}. P e ( ρ j ) ⟶ P e ( ρ ) in L s ( Q T ) , 1 ≤ s < b ∗ b ∗ + 1 .
The convective and weighted-velocity limits follow from (A.13); the auxiliary internal energy vanishes by (A.22).
It remains to pass to the energy inequality and recover the initial traces. At each of the three approximation limits, the weak limit of W j = ρ j u j W_j=\sqrt{\rho_j}u_j W j = ρ j u j in L 2 ( Q T ) L^2(Q_T) L 2 ( Q T ) is W = ρ u W=\sqrt{\rho}\,u W = ρ u . Strong momentum convergence in C t H x − 1 C_tH_x^{-1} C t H x − 1 and weak convergence of u j u_j u j in L 2 H 1 L^2H^1 L 2 H 1 give
∥ W j ∥ L 2 ( Q T ) 2 = ∫ 0 T ⟨ m j , u j ⟩ d t ⟶ ∫ 0 T ⟨ ρ u , u ⟩ d t = ∥ W ∥ L 2 ( Q T ) 2 , (106) \|W_j\|_{L^2(Q_T)}^2=\int_0^T\langle m_j,u_j\rangle\,\mathrm{d}t\longrightarrow\int_0^T\langle\rho u,u\rangle\,\mathrm{d}t=\|W\|_{L^2(Q_T)}^2, \tag*{(106)} ∥ W j ∥ L 2 ( Q T ) 2 = ∫ 0 T ⟨ m j , u j ⟩ d t ⟶ ∫ 0 T ⟨ ρ u , u ⟩ d t = ∥ W ∥ L 2 ( Q T ) 2 , (106)
W j ⟶ W in L 2 ( Q T ) , W_j\longrightarrow W\quad\text{in }L^2(Q_T), W j ⟶ W in L 2 ( Q T ) ,
∣ W j ∣ 2 ⟶ ∣ W ∣ 2 in L 1 ( Q T ) . |W_j|^2\longrightarrow|W|^2\quad\text{in }L^1(Q_T). ∣ W j ∣ 2 ⟶ ∣ W ∣ 2 in L 1 ( Q T ) .
The internal energies converge in L 1 ( Q T ) L^1(Q_T) L 1 ( Q T ) by
N ⟶ ∞ : ρ N ⟶ ρ in L q , G < q < 5 G 3 , N\longrightarrow\infty:\quad\rho_N\longrightarrow\rho\quad\text{in }L^q,\quad G<q<\frac{5G}{3}, N ⟶ ∞ : ρ N ⟶ ρ in L q , G < q < 3 5 G ,
α ↓ 0 : ρ α ⟶ ρ in L q , G < q < G + 1 , \alpha\downarrow0:\quad\rho_\alpha\longrightarrow\rho\quad\text{in }L^q,\quad G<q<G+1, α ↓ 0 : ρ α ⟶ ρ in L q , G < q < G + 1 ,
σ j ↓ 0 : ρ j ⟶ ρ in L q , b ∗ < q < b ∗ + 1 , σ j ρ j G ⟶ 0 in L 1 . \sigma_j\downarrow0:\quad\rho_j\longrightarrow\rho\quad\text{in }L^q,\quad b_*<q<b_*+1,\quad\sigma_j\rho_j^G\longrightarrow0\quad\text{in }L^1. σ j ↓ 0 : ρ j ⟶ ρ in L q , b ∗ < q < b ∗ + 1 , σ j ρ j G ⟶ 0 in L 1 .
The initial kinetic energy converges under the weighted Galerkin projection, the initial data remain fixed as α ↓ 0 \alpha\downarrow0 α ↓ 0 , and
E σ , 0 = E ε , 0 + σ G − 1 ∫ T 3 ρ ∗ G ⟶ E ε , 0 . \mathcal{E}_{\sigma,0}=\mathcal{E}_{\varepsilon,0}+\frac{\sigma}{G-1}\int_{\mathbb{T}^3}\rho_*^G\longrightarrow\mathcal{E}_{\varepsilon,0}. E σ , 0 = E ε , 0 + G − 1 σ ∫ T 3 ρ ∗ G ⟶ E ε , 0 .
Thus, at each stage, E j ⟶ E \mathcal{E}_j\longrightarrow\mathcal{E} E j ⟶ E in L 1 ( 0 , T ) L^1(0,T) L 1 ( 0 , T ) and E j , 0 ⟶ E 0 \mathcal{E}_{j,0}\longrightarrow\mathcal{E}_0 E j , 0 ⟶ E 0 . For 0 ≤ ψ ∈ C c ∞ ( [ 0 , T ) ) 0\leq\psi\in C_c^\infty([0,T)) 0 ≤ ψ ∈ C c ∞ ([ 0 , T )) ,
− ∫ E ψ ′ + ∫ D ψ ≤ lim inf j [ − ∫ E j ψ ′ + ∫ D j ψ ] ≤ E 0 ψ ( 0 ) . (107) -\int\mathcal{E}\psi'+\int\mathcal{D}\psi\leq\liminf_j\left[-\int\mathcal{E}_j\psi'+\int\mathcal{D}_j\psi\right]\leq\mathcal{E}_0\psi(0). \tag*{(107)} − ∫ E ψ ′ + ∫ D ψ ≤ j lim inf [ − ∫ E j ψ ′ + ∫ D j ψ ] ≤ E 0 ψ ( 0 ) . (107)
The uniform energy and higher-integrability bounds control the endpoint time layers. The initialized equations and negative-norm time compactness give ρ ∈ C w L b ∗ \rho\in C_wL^{b_*} ρ ∈ C w L b ∗ and ρ u ∈ C w L 2 b ∗ / ( b ∗ + 1 ) \rho u\in C_wL^{2b_*/(b_*+1)} ρ u ∈ C w L 2 b ∗ / ( b ∗ + 1 ) , with initial values ρ ∗ , m ∗ \rho_*,m_* ρ ∗ , m ∗ ; the commutator argument provides the initial terms in the renormalized equation. Finally, a diagonal subsequence on integer time intervals gives the required global solution on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . □
Auxiliary Periodic Commutator Lemma
The following lemma was used in Steps 2 and 3 of the proof of Lemma A.1. Its hypotheses are supplied by the negative-norm compactness of density and momentum at those fixed-pressure stages. It passes a bilinear Riesz commutator against the weak velocity by first obtaining a distributional limit at each time in a common full-measure set, then using spatial compactness and dominated convergence in time. The quantitative high-frequency lemma used in Section 4 has a different role: it gives a remainder uniform in the final approximation index, where these fixed-pressure spatial bounds are unavailable.
Lemma A.2. Let 1 < a , b < ∞ 1<a,b<\infty 1 < a , b < ∞ satisfy 1 / a + 1 / b < 5 / 6 1/a+1/b<5/6 1/ a + 1/ b < 5/6 . Suppose that
f j ⟶ f , g j ⟶ g strongly in C ( [ 0 , T ] ; H − 1 ( T 3 ) ) , f_j\longrightarrow f,\qquad g_j\longrightarrow g\quad\text{strongly in }C([0,T];H^{-1}(\mathbb{T}^3)), f j ⟶ f , g j ⟶ g strongly in C ([ 0 , T ] ; H − 1 ( T 3 )) , sup j ( ∥ f j ∥ L t ∞ L x a + ∥ g j ∥ L t ∞ L x b ) < ∞ , \sup_j\left(\|f_j\|_{L^\infty_tL^a_x}+\|g_j\|_{L^\infty_tL^b_x}\right)<\infty, j sup ( ∥ f j ∥ L t ∞ L x a + ∥ g j ∥ L t ∞ L x b ) < ∞ ,
where g j = ( g j , i ) g_j=(g_{j,i}) g j = ( g j , i ) is vector-valued. For R i l = ∂ i ∂ l Δ − 1 R_{il}=\partial_i\partial_l\Delta^{-1} R i l = ∂ i ∂ l Δ − 1 with its periodic zero-mode convention, set
B j , l = f j R i l g j , i − g j , i R i l f j . B_{j,l}=f_jR_{il}g_{j,i}-g_{j,i}R_{il}f_j. B j , l = f j R i l g j , i − g j , i R i l f j .
Then B j → B = f R i l g i − g i R i l f B_j\to B=fR_{il}g_i-g_iR_{il}f B j → B = f R i l g i − g i R i l f strongly in L 2 ( 0 , T ; H − 1 ( T 3 ) ) L^2(0,T;H^{-1}(\mathbb{T}^3)) L 2 ( 0 , T ; H − 1 ( T 3 )) , with the limit initially understood in distributions.
Proof. Put 1 / r = 1 / a + 1 / b < 5 / 6 1/r=1/a+1/b<5/6 1/ r = 1/ a + 1/ b < 5/6 , so r > 6 / 5 r>6/5 r > 6/5 and a > b ′ a>b' a > b ′ , where 1 / b + 1 / b ′ = 1 1/b+1/b'=1 1/ b + 1/ b ′ = 1 . Boundedness of the periodic Riesz operators, obtained from the Calderón–Zygmund estimates in [40 ] , gives
∥ B j ( t ) ∥ r ≤ C ∥ f j ( t ) ∥ a ∥ g j ( t ) ∥ b , sup j ∥ B j ∥ L t ∞ L x r < ∞ . \lVert B_j(t)\rVert_r\le C\lVert f_j(t)\rVert_a\lVert g_j(t)\rVert_b,\qquad\sup_j\lVert B_j\rVert_{L^\infty_tL^r_x}<\infty. ∥ B j ( t ) ∥ r ≤ C ∥ f j ( t ) ∥ a ∥ g j ( t ) ∥ b , j sup ∥ B j ∥ L t ∞ L x r < ∞.
Choose the continuous H − 1 H^{-1} H − 1 representatives. Remove the union of the null sets where either spatial bound fails, over the countable sequence, and denote the remaining full-measure set by E E E . At each time in E E E , weak compactness in L a L^a L a and L b L^b L b , together with convergence in H − 1 H^{-1} H − 1 , identifies every cluster point. Thus
f j ( t ) ⇀ f ( t ) in L a , g j ( t ) ⇀ g ( t ) in L b , t ∈ E . f_j(t)\rightharpoonup f(t)\ \text{in }L^a,\qquad g_j(t)\rightharpoonup g(t)\ \text{in }L^b,\qquad t\in E. f j ( t ) ⇀ f ( t ) in L a , g j ( t ) ⇀ g ( t ) in L b , t ∈ E .
The limits satisfy the same spatial bounds on E E E , which is independent of the spatial test function.
Fix t ∈ E t\in E t ∈ E and ϕ ∈ C ∞ ( T 3 ) \phi\in C^\infty(\mathbb{T}^3) ϕ ∈ C ∞ ( T 3 ) . Self-adjointness of R i l R_{il} R i l gives
⟨ B j , l ( t ) , ϕ ⟩ = ∫ T 3 g j , i ( t ) [ R i l , ϕ ] f j ( t ) d x . (108) \langle B_{j,l}(t),\phi\rangle=\int_{\mathbb{T}^3}g_{j,i}(t)[R_{il},\phi]f_j(t)\,dx. \tag*{(108)} ⟨ B j , l ( t ) , ϕ ⟩ = ∫ T 3 g j , i ( t ) [ R i l , ϕ ] f j ( t ) d x . (108)
The contact term in the distribution kernel of R i l R_{il} R i l cancels, because multiplication by ϕ \phi ϕ commutes with that term. Near the diagonal the remaining kernel satisfies
∣ K i l ( x − y ) ( ϕ ( y ) − ϕ ( x ) ) ∣ ≤ C ϕ ∣ x − y ∣ − 2 . \lvert K_{il}(x-y)(\phi(y)-\phi(x))\rvert\le C_\phi\lvert x-y\rvert^{-2}. ∣ K i l ( x − y ) ( ϕ ( y ) − ϕ ( x ))∣ ≤ C ϕ ∣ x − y ∣ − 2 .
Decompose [ R i l , ϕ ] = C h n e a r + C h f a r [R_{il},\phi]=C_h^{\mathrm{near}}+C_h^{\mathrm{far}} [ R i l , ϕ ] = C h near + C h far with a smooth cutoff supported in ∣ x − y ∣ < 2 h \lvert x-y\rvert<2h ∣ x − y ∣ < 2 h . Young’s inequality and L a ( T 3 ) ↪ L b ′ ( T 3 ) L^a(\mathbb{T}^3)\hookrightarrow L^{b'}(\mathbb{T}^3) L a ( T 3 ) ↪ L b ′ ( T 3 ) give
∥ C h n e a r f ∥ b ′ ≤ C ∥ C h n e a r f ∥ a ≤ C ϕ h ∥ f ∥ a , \lVert C_h^{\mathrm{near}}f\rVert_{b'}\le C\lVert C_h^{\mathrm{near}}f\rVert_a\le C_\phi h\lVert f\rVert_a, ∥ C h near f ∥ b ′ ≤ C ∥ C h near f ∥ a ≤ C ϕ h ∥ f ∥ a ,
∣ ∫ T 3 g C h n e a r f ∣ ≤ C ϕ h ∥ g ∥ b ∥ f ∥ a . (109) \left\lvert\int_{\mathbb{T}^3}gC_h^{\mathrm{near}}f\right\rvert\le C_\phi h\lVert g\rVert_b\lVert f\rVert_a. \tag*{(109)} ∫ T 3 g C h near f ≤ C ϕ h ∥ g ∥ b ∥ f ∥ a . (109)
For fixed h > 0 h>0 h > 0 , the far kernel is smooth on T 3 × T 3 \mathbb{T}^3\times\mathbb{T}^3 T 3 × T 3 . It maps bounded subsets of L a L^a L a to bounded subsets of C 1 ( T 3 ) C^1(\mathbb{T}^3) C 1 ( T 3 ) , and is therefore compact into L b ′ ( T 3 ) L^{b'}(\mathbb{T}^3) L b ′ ( T 3 ) . Thus
C h f a r f j ( t ) ⟶ C h f a r f ( t ) in L b ′ . C_h^{\mathrm{far}}f_j(t)\longrightarrow C_h^{\mathrm{far}}f(t)\quad\text{in }L^{b'}. C h far f j ( t ) ⟶ C h far f ( t ) in L b ′ .
Weak convergence of g j ( t ) g_j(t) g j ( t ) in L b L^b L b then gives convergence of the far pairing. Apply (A.29) to the sequence and its limit and let h ↓ 0 h\downarrow0 h ↓ 0 . This identifies the limit of (A.28) with the pairing for f , g f,g f , g , so B j ( t ) → B ( t ) B_j(t)\to B(t) B j ( t ) → B ( t ) in distributions for every t ∈ E t\in E t ∈ E .
Since r > 6 / 5 r>6/5 r > 6/5 , its conjugate exponent satisfies r ′ < 6 r'<6 r ′ < 6 . The compact Sobolev embedding H 1 ( T 3 ) ⋐ L r ′ ( T 3 ) H^1(\mathbb{T}^3)\Subset L^{r'}(\mathbb{T}^3) H 1 ( T 3 ) ⋐ L r ′ ( T 3 ) [9 ] , followed by duality, gives L r ( T 3 ) ⋐ H − 1 ( T 3 ) L^r(\mathbb{T}^3)\Subset H^{-1}(\mathbb{T}^3) L r ( T 3 ) ⋐ H − 1 ( T 3 ) . This strict inequality explains the restriction 1 / a + 1 / b < 5 / 6 1/a+1/b<5/6 1/ a + 1/ b < 5/6 in the lemma. The compact embedding and the identified distributional limit imply strong H − 1 H^{-1} H − 1 convergence at each such time: every subsequence has a strongly convergent further subsequence, and its limit must be B ( t ) B(t) B ( t ) . Finally,
∥ B j ( t ) − B ( t ) ∥ H − 1 2 ≤ C ( t ∈ E ′ ) , ∫ 0 T ∥ B j ( t ) − B ( t ) ∥ H − 1 2 d t ⟶ 0 \|B_j(t)-B(t)\|_{H^{-1}}^2 \le C \quad(t\in E'),
\qquad
\int_0^T \|B_j(t)-B(t)\|_{H^{-1}}^2\,dt \longrightarrow0 ∥ B j ( t ) − B ( t ) ∥ H − 1 2 ≤ C ( t ∈ E ′ ) , ∫ 0 T ∥ B j ( t ) − B ( t ) ∥ H − 1 2 d t ⟶ 0
by dominated convergence. □ \square □
Proof of Lemma 2.1. Apply the fixed-e e e existence result to the data in (A.1), whose full initial energies converge to E 0 E_0 E 0 . Coercivity of the periodic viscous term and Lemma 2.5 give the uniform bounds. For each approximate solution, the cubic density norm may grow as e n − 1 / 3 e_n^{-1/3} e n − 1/3 ; it justifies the special test field, whereas the pressure bound is uniform in n n n . □ \square □
Hilbert-Space Construction
We establish the Hilbert-space monotonicity and exact-marginal representation used in Lemmas 3.1 and 3.2 . Transport planning allows a prescribed target law to be retained under relaxation [1 , 16 , 17 ] ; the strong-convexity and subgradient arguments use the Hilbert-space framework of [4 ] . The periodic lifts and the exact representation needed by the restricted Jacobian argument are proved below. The deterministic infimum first identifies the transport value. Its convex envelope yields a unique residual optimizer, and an optimal plan then realizes its subgradient as a conditional barycenter. § 3.1 converts that barycenter into a periodic convex gradient.
Hilbert-Space Envelope
Set D = diam ( T 3 ) = 3 / 2 D=\operatorname{diam}(\mathbb{T}^3)=\sqrt{3}/2 D = diam ( T 3 ) = 3 /2 . Let ( A , ℓ ) (\mathcal{A},\ell) ( A , ℓ ) be a standard atomless measure space of mass c c c , with positions X ( a ) X(a) X ( a ) , and set H = L 2 ( A , ℓ ; R 3 ) \mathcal{H}=L^2(\mathcal{A},\ell;\mathbb{R}^3) H = L 2 ( A , ℓ ; R 3 ) with inner product ⟨ X , Y ⟩ = ∫ A X ⋅ Y d ℓ \langle X,Y\rangle=\int_{\mathcal{A}}X\cdot Y\,d\ell ⟨ X , Y ⟩ = ∫ A X ⋅ Y d ℓ . Write q : R 3 → T 3 q:\mathbb{R}^3\to\mathbb{T}^3 q : R 3 → T 3 for the quotient map and define
K l a w = { Z ∈ H : q # ( Z # ℓ ) = c d x } . \mathcal{K}_{\mathrm{law}}=\{Z\in\mathcal{H}:q_{\#}(Z_{\#}\ell)=c\,dx\}. K law = { Z ∈ H : q # ( Z # ℓ ) = c d x } .
For equal-mass measures, let W 2 , T 3 2 W_{2,\mathbb{T}^3}^2 W 2 , T 3 2 denote the unnormalized quadratic transport cost. Define
F ( X ) = − 1 2 inf Z ∈ K l a w ∥ X − Z ∥ H 2 = − 1 2 W 2 , T 3 2 ( q # X # ℓ , c d x ) , F(X)=-\frac{1}{2}\inf_{Z\in\mathcal{K}_{\mathrm{law}}}\|X-Z\|_{\mathcal{H}}^2
=-\frac{1}{2}W_{2,\mathbb{T}^3}^2(q_{\#}X_{\#}\ell,c\,dx), F ( X ) = − 2 1 Z ∈ K law inf ∥ X − Z ∥ H 2 = − 2 1 W 2 , T 3 2 ( q # X # ℓ , c d x ) ,
F K ( X ) = inf Y ∈ H { F ( Y ) + K 2 ∥ X − Y ∥ H 2 } , H K ( X ) = 1 2 ∥ X ∥ H 2 + K − 1 K F K ( X ) , T K ( X ) = ∇ H K ( X ) . (110) F_K(X)=\inf_{Y\in\mathcal{H}}\left\{F(Y)+\frac{K}{2}\|X-Y\|_{\mathcal{H}}^2\right\},
\qquad
\mathcal{H}_K(X)=\frac{1}{2}\|X\|_{\mathcal{H}}^2+\frac{K-1}{K}F_K(X),
\qquad
\mathcal{T}_K(X)=\nabla\mathcal{H}_K(X). \tag*{(110)} F K ( X ) = Y ∈ H inf { F ( Y ) + 2 K ∥ X − Y ∥ H 2 } , H K ( X ) = 2 1 ∥ X ∥ H 2 + K K − 1 F K ( X ) , T K ( X ) = ∇ H K ( X ) . (110)
Every deterministic target map induces an admissible plan, giving one inequality. Conversely, partition the source and target tori into cells of diameter at most h h h . Atomlessness allows the source labels to be split according to the mass assigned to each cell pair and mapped uniformly onto the corresponding target cell, preserving the marginal c d x c\,dx c d x . The costs converge as h ↓ 0 h\downarrow0 h ↓ 0 , and nearest integer lifts give the Euclidean infimum. The infimum need not be attained by a deterministic map.
Lemma B.1 (Hilbert-space envelope). For K > 1 K > 1 K > 1 and X ∈ H X \in\mathcal{H} X ∈ H , the minimizer Y K ( X ) Y_K(X) Y K ( X ) in the definition of F K ( X ) F_K(X) F K ( X ) is unique, with
T K ( X ) = K X − ( K − 1 ) Y K ( X ) . \mathcal{T}_K(X) = KX - (K - 1)Y_K(X). T K ( X ) = K X − ( K − 1 ) Y K ( X ) . The map T K \mathcal{T}_K T K is K K K -Lipschitz and satisfies
⟨ T K ( X ) − T K ( X ′ ) , X − X ′ ⟩ ≥ 1 K ∥ T K ( X ) − T K ( X ′ ) ∥ 2 . (111) \langle\mathcal{T}_K(X) - \mathcal{T}_K(X'), X - X'\rangle\ge\frac{1}{K} \lVert\mathcal{T}_K(X) - \mathcal{T}_K(X') \rVert^2. \tag*{(111)} ⟨ T K ( X ) − T K ( X ′ ) , X − X ′ ⟩ ≥ K 1 ∥ T K ( X ) − T K ( X ′ ) ∥ 2 . (111) The displacement T K ( X ) − X \mathcal{T}_K(X) - X T K ( X ) − X is uniformly bounded and depends only on the source law and the current torus position. For every H 1 H^1 H 1 curve X ( t ) X(t) X ( t ) in H \mathcal{H} H ,
⟨ ∂ t T K ( X ( t ) ) , X ˙ ( t ) ⟩ ≥ 1 K ∥ ∂ t T K ( X ( t ) ) ∥ 2 for almost every t . (112) \langle\partial_t\mathcal{T}_K(X(t)), \dot{X}(t)\rangle\ge\frac{1}{K}\lVert\partial_t\mathcal{T}_K(X(t))\rVert^2 \quad\text{for almost every } t. \tag*{(112)} ⟨ ∂ t T K ( X ( t )) , X ˙ ( t )⟩ ≥ K 1 ∥ ∂ t T K ( X ( t )) ∥ 2 for almost every t . (112)
Proof. Expanding the square gives
G ( Y ) : = F ( Y ) + 1 2 ∥ Y ∥ 2 = sup Z ∈ K l a w { ⟨ Y , Z ⟩ − 1 2 ∥ Z ∥ 2 } . G(Y) := F(Y) + \frac{1}{2}\lVert Y\rVert^2 = \sup_{Z \in\mathcal{K}_{\mathrm{law}}} \left\{\langle Y,Z\rangle- \frac{1}{2}\lVert Z\rVert^2\right\}. G ( Y ) := F ( Y ) + 2 1 ∥ Y ∥ 2 = Z ∈ K law sup { ⟨ Y , Z ⟩ − 2 1 ∥ Z ∥ 2 } .
Thus G G G is convex and weakly lower semicontinuous. Since − c D 2 / 2 ≤ F ≤ 0 -cD^2/2 \le F \le0 − c D 2 /2 ≤ F ≤ 0 , the objective is coercive and equals
G ( Y ) + K − 1 2 ∥ Y ∥ 2 − K ⟨ X , Y ⟩ + K 2 ∥ X ∥ 2 . G(Y) + \frac{K-1}{2}\lVert Y\rVert^2 - K\langle X,Y\rangle+ \frac{K}{2}\lVert X\rVert^2. G ( Y ) + 2 K − 1 ∥ Y ∥ 2 − K ⟨ X , Y ⟩ + 2 K ∥ X ∥ 2 .
A minimizing sequence is bounded, since its objective is at least K ∥ X − Y ∥ 2 / 2 − c D 2 / 2 K\lVert X-Y\rVert^2/2 - cD^2/2 K ∥ X − Y ∥ 2 /2 − c D 2 /2 . A weakly convergent subsequence exists in H \mathcal{H} H , and weak lower semicontinuity of the displayed expression gives a minimizer. Convexity of G G G and K − 1 > 0 K - 1 > 0 K − 1 > 0 imply uniqueness. More precisely, at the minimizer the objective satisfies
F ( Y ) + K 2 ∥ X − Y ∥ 2 ≥ F ( Y K ( X ) ) + K 2 ∥ X − Y K ( X ) ∥ 2 + K − 1 2 ∥ Y − Y K ( X ) ∥ 2 . F(Y) + \frac{K}{2}\lVert X-Y\rVert^2 \ge F(Y_K(X)) + \frac{K}{2}\lVert X-Y_K(X)\rVert^2 + \frac{K-1}{2}\lVert Y-Y_K(X)\rVert^2. F ( Y ) + 2 K ∥ X − Y ∥ 2 ≥ F ( Y K ( X )) + 2 K ∥ X − Y K ( X ) ∥ 2 + 2 K − 1 ∥ Y − Y K ( X ) ∥ 2 .
The Hilbert-space subgradient rule [4 ] gives the optimality condition
K X − ( K − 1 ) Y K ( X ) ∈ ∂ G ( Y K ( X ) ) . (113) KX - (K - 1)Y_K(X) \in\partial G(Y_K(X)). \tag*{(113)} K X − ( K − 1 ) Y K ( X ) ∈ ∂ G ( Y K ( X )) . (113)
For two inputs, write Δ X = X − X ′ \Delta X = X - X' Δ X = X − X ′ , Δ Y = Y K ( X ) − Y K ( X ′ ) \Delta Y = Y_K(X) - Y_K(X') Δ Y = Y K ( X ) − Y K ( X ′ ) , and Δ T = T K ( X ) − T K ( X ′ ) \Delta\mathcal{T} = \mathcal{T}_K(X) - \mathcal{T}_K(X') Δ T = T K ( X ) − T K ( X ′ ) . Monotonicity of ∂ G \partial G ∂ G gives
⟨ Δ T , Δ Y ⟩ ≥ 0 , \langle\Delta\mathcal{T}, \Delta Y\rangle\ge0, ⟨ Δ T , Δ Y ⟩ ≥ 0 ,
K ⟨ Δ T , Δ X ⟩ = ∥ Δ T ∥ 2 + ( K − 1 ) ⟨ Δ T , Δ Y ⟩ ≥ ∥ Δ T ∥ 2 . K\langle\Delta\mathcal{T}, \Delta X\rangle= \lVert\Delta\mathcal{T}\rVert^2 + (K - 1)\langle\Delta\mathcal{T}, \Delta Y\rangle\ge\lVert\Delta\mathcal{T}\rVert^2. K ⟨ Δ T , Δ X ⟩ = ∥ Δ T ∥ 2 + ( K − 1 ) ⟨ Δ T , Δ Y ⟩ ≥ ∥ Δ T ∥ 2 .
This gives (B.2) and, by Cauchy–Schwarz, the K K K -Lipschitz bound. Monotonicity also gives
( K − 1 ) ∥ Δ Y ∥ 2 ≤ K ⟨ Δ X , Δ Y ⟩ , ∥ Δ Y ∥ ≤ K K − 1 ∥ Δ X ∥ . (K - 1)\lVert\Delta Y\rVert^2 \le K\langle\Delta X, \Delta Y\rangle,\qquad\lVert\Delta Y\rVert\le\frac{K}{K - 1}\lVert\Delta X\rVert. ( K − 1 ) ∥ Δ Y ∥ 2 ≤ K ⟨ Δ X , Δ Y ⟩ , ∥ Δ Y ∥ ≤ K − 1 K ∥ Δ X ∥ .
For an increment U U U , use Y K ( X ) Y_K(X) Y K ( X ) as competitor at X + U X+U X + U and Y K ( X + U ) Y_K(X+U) Y K ( X + U ) as competitor at X X X . This gives
K ⟨ X − Y K ( X + U ) , U ⟩ + K 2 ∥ U ∥ 2 ≤ F K ( X + U ) − F K ( X ) , K\langle X-Y_K(X+U),U\rangle+\frac{K}{2}\lVert U\rVert^2\le F_K(X+U)-F_K(X), K ⟨ X − Y K ( X + U ) , U ⟩ + 2 K ∥ U ∥ 2 ≤ F K ( X + U ) − F K ( X ) ,
F K ( X + U ) − F K ( X ) ≤ K ⟨ X − Y K ( X ) , U ⟩ + K 2 ∥ U ∥ 2 . F_K(X+U)-F_K(X)\le K\langle X-Y_K(X),U\rangle+\frac{K}{2}\lVert U\rVert^2. F K ( X + U ) − F K ( X ) ≤ K ⟨ X − Y K ( X ) , U ⟩ + 2 K ∥ U ∥ 2 .
The Lipschitz bound for Y K Y_K Y K makes the difference between the two linear terms O ( ∥ U ∥ 2 ) O(\lVert U\rVert^2) O (∥ U ∥ 2 ) . Thus differentiating the minimum value is justified and yields
∇ F K ( X ) = K ( X − Y K ( X ) ) , ∇ H K ( X ) = K X − ( K − 1 ) Y K ( X ) , \nabla F_K(X)=K(X-Y_K(X)),\qquad\nabla\mathcal{H}_K(X)=KX-(K-1)Y_K(X), ∇ F K ( X ) = K ( X − Y K ( X )) , ∇ H K ( X ) = K X − ( K − 1 ) Y K ( X ) ,
as required.
The quadratic torus cost is D D D -Lipschitz in its first variable, so
∣ F ( X + U ) − F ( X ) ∣ ≤ D ∥ U ∥ L 1 ( ℓ ) . \lvert F(X+U)-F(X)\rvert\le D\lVert U\rVert_{L^1(\ell)}. ∣ F ( X + U ) − F ( X )∣ ≤ D ∥ U ∥ L 1 ( ℓ ) .
Simultaneously translating the input and competitor shows that F K F_K F K has the same bound. Localized variations then give ∣ K ( X − Y K ( X ) ) ∣ ≤ D \lvert K(X-Y_K(X))\rvert\le D ∣ K ( X − Y K ( X ))∣ ≤ D almost everywhere, and hence
∣ Y K ( X ) − X ∣ ≤ D / K , ∣ T K ( X ) − X ∣ ≤ ( K − 1 ) D / K ≤ D . (114) \lvert Y_K(X)-X\rvert\le D/K,\qquad\lvert\mathcal{T}_K(X)-X\rvert\le(K-1)D/K\le D. \tag*{(114)} ∣ Y K ( X ) − X ∣ ≤ D / K , ∣ T K ( X ) − X ∣ ≤ ( K − 1 ) D / K ≤ D . (114)
For integer-valued N ∈ H N\in\mathcal{H} N ∈ H , uniqueness and translation invariance give
F ( X + N ) = F ( X ) , Y K ( X + N ) = Y K ( X ) + N , T K ( X + N ) = T K ( X ) + N . F(X+N)=F(X),\qquad Y_K(X+N)=Y_K(X)+N,\qquad\mathcal{T}_K(X+N)=\mathcal{T}_K(X)+N. F ( X + N ) = F ( X ) , Y K ( X + N ) = Y K ( X ) + N , T K ( X + N ) = T K ( X ) + N .
The displacement is therefore invariant under integer lifts and measure-preserving relabelings. To prove that it depends only on position, we use conditional Jensen’s inequality for the convex law-invariant functional G G G . Here is a direct justification. For Y ∈ H Y\in\mathcal{H} Y ∈ H , realize conditionally independent copies Y 1 , Y 2 , … Y_1,Y_2,\ldots Y 1 , Y 2 , … of Y Y Y given X X X on a product label space. All Y j Y_j Y j have the law of Y Y Y , and
G ( 1 N ∑ j = 1 N Y j ) ≤ 1 N ∑ j = 1 N G ( Y j ) = G ( Y ) , G\left(\frac{1}{N}\sum_{j=1}^{N}Y_j\right)\le\frac{1}{N}\sum_{j=1}^{N}G(Y_j)=G(Y), G ( N 1 j = 1 ∑ N Y j ) ≤ N 1 j = 1 ∑ N G ( Y j ) = G ( Y ) ,
∥ 1 N ∑ j = 1 N Y j − E [ Y ∣ X ] ∥ 2 2 = 1 N ∥ Y − E [ Y ∣ X ] ∥ 2 2 ⟶ 0. \left\lVert\frac{1}{N}\sum_{j=1}^{N}Y_j-\mathbb{E}[Y\mid X]\right\rVert_2^2=\frac{1}{N}\lVert Y-\mathbb{E}[Y\mid X]\rVert_2^2\longrightarrow0. N 1 j = 1 ∑ N Y j − E [ Y ∣ X ] 2 2 = N 1 ∥ Y − E [ Y ∣ X ] ∥ 2 2 ⟶ 0.
The same formula defining G G G applies on every standard atomless label space and depends only on the law of its argument. Since F F F is Lipschitz in L 2 L^2 L 2 and the squared norm is continuous, G G G is continuous in L 2 L^2 L 2 . Passing to the limit proves G ( E [ Y ∣ X ] ) ≤ G ( Y ) G(\mathbb{E}[Y\mid X])\le G(Y) G ( E [ Y ∣ X ]) ≤ G ( Y ) . This argument is an evaluation of the law-invariant functional; it does not assume that minimizers persist when the label space is enlarged.
Set Y ‾ = E [ Y ∣ X ] \overline{Y}=\mathbb{E}[Y\mid X] Y = E [ Y ∣ X ] . Orthogonality of conditional expectation and X X X -measurability of X X X give
∥ Y ∥ 2 = ∥ Y ‾ ∥ 2 + ∥ Y − Y ‾ ∥ 2 , \lVert Y\rVert^2=\lVert\overline{Y}\rVert^2+\lVert Y-\overline{Y}\rVert^2, ∥ Y ∥ 2 = ∥ Y ∥ 2 + ∥ Y − Y ∥ 2 ,
⟨ X , Y ⟩ = ⟨ X , Y ‾ ⟩ . \langle X,Y\rangle=\langle X,\overline{Y}\rangle. ⟨ X , Y ⟩ = ⟨ X , Y ⟩ .
In the objective written in terms of G G G , replacing Y Y Y by Y ‾ \overline{Y} Y therefore lowers the value by at least ( K − 1 ) ∥ Y − Y ‾ ∥ 2 / 2 (K-1)\lVert Y-\overline{Y}\rVert^2/2 ( K − 1 ) ∥ Y − Y ∥ 2 /2 . At the minimizer this quantity must vanish, so Y K ( X ) Y_K(X) Y K ( X ) is a Borel function of X X X . Once competitors have this form, their costs depend only on the source law. Uniqueness then gives the same Borel function for any realization of that law, almost everywhere for the law. Integer-lift invariance gives the asserted dependence on the current torus position.
Lipschitz continuity gives J K ( X ( t ) ) ∈ H 1 \mathcal{J}_K(X(t))\in H^1 J K ( X ( t )) ∈ H 1 whenever X ( t ) ∈ H 1 X(t)\in H^1 X ( t ) ∈ H 1 . Applying (B.2) at neighboring times and passing to the difference-quotient limit proves (B.3). □ \square □
Exact-Marginal Representation
Lemma B.2 (Exact-marginal representation). Let y ∈ H y\in\mathcal{H} y ∈ H be bounded and let G G G be as in Lemma B.1. Denote by P y \mathcal{P}_y P y the optimal plans for ∫ ∣ y ( a ) − z ∣ 2 d π ( a , z ) \int|y(a)-z|^2\,d\pi(a,z) ∫ ∣ y ( a ) − z ∣ 2 d π ( a , z ) , with source marginal ℓ \ell ℓ , target marginal c d x c\,dx c d x modulo integers, and nearest integer lifts. Set
B y = { B π ( a ) = E π [ z ∣ a ] : π ∈ P y } ⊂ L 2 ( A , ℓ ; R 3 ) . B_y=\{B_\pi(a)=\mathbb{E}_\pi[z\mid a]:\pi\in\mathcal{P}_y\}\subset L^2(\mathcal{A},\ell;\mathbb{R}^3). B y = { B π ( a ) = E π [ z ∣ a ] : π ∈ P y } ⊂ L 2 ( A , ℓ ; R 3 ) . The set B y B_y B y is nonempty, convex, and weakly compact, and
∂ G ( y ) = B y . (115) \partial G(y)=B_y. \tag*{(115)} ∂ G ( y ) = B y . (115) Thus every subgradient is represented by an optimal plan with the prescribed target marginal.
Proof. Choose a compact cube Q ⊂ R 3 Q\subset\mathbb{R}^3 Q ⊂ R 3 containing all nearest lifts to the essential range of y y y . The compact set contains lifted targets, not just their barycenters: the exact target law must be preserved before taking a conditional expectation. In general the barycenter itself does not have the prescribed target law. Consider probability kernels κ a ∈ P ( Q ) \kappa_a\in\mathcal{P}(Q) κ a ∈ P ( Q ) satisfying
π ( d a , d z ) = ℓ ( d a ) κ a ( d z ) , \pi(da,dz)=\ell(da)\kappa_a(dz), π ( d a , d z ) = ℓ ( d a ) κ a ( d z ) ,
∫ A ∫ Q ζ ( q z ) κ a ( d z ) d ℓ ( a ) = c ∫ T 3 ζ d x ( ζ ∈ C ( T 3 ) ) . (116) \int_{\mathcal{A}}\int_Q \zeta(qz)\kappa_a(dz)\,d\ell(a)=c\int_{\mathbb{T}^3}\zeta\,dx
\qquad(\zeta\in C(\mathbb{T}^3)). \tag*{(116)} ∫ A ∫ Q ζ ( q z ) κ a ( d z ) d ℓ ( a ) = c ∫ T 3 ζ d x ( ζ ∈ C ( T 3 )) . (116)
Let P Q \mathfrak{P}_Q P Q denote this set. It is nonempty: assign the uniform torus law to each label and choose nearest lifts measurably, with a fixed rule for ties.
Set E Q = L 1 ( A , ℓ ; C ( Q ) ) \mathcal{E}_Q=L^1(\mathcal{A},\ell;C(Q)) E Q = L 1 ( A , ℓ ; C ( Q )) and associate to a kernel the functional
L π ( Φ ) = ∫ A ∫ Q Φ ( a , z ) κ a ( d z ) d ℓ ( a ) , ∣ L π ( Φ ) ∣ ≤ ∥ Φ ∥ E Q . L_\pi(\Phi)=\int_{\mathcal{A}}\int_Q\Phi(a,z)\kappa_a(dz)\,d\ell(a),
\qquad|L_\pi(\Phi)|\leq\|\Phi\|_{\mathcal{E}_Q}. L π ( Φ ) = ∫ A ∫ Q Φ ( a , z ) κ a ( d z ) d ℓ ( a ) , ∣ L π ( Φ ) ∣ ≤ ∥Φ ∥ E Q .
Finite sums f ( a ) g ( z ) f(a)g(z) f ( a ) g ( z ) are dense in E Q \mathcal{E}_Q E Q , and separability gives weak-star sequential compactness of the unit dual ball. A limit functional again defines a kernel: on a countable dense subspace of C ( Q ) C(Q) C ( Q ) , the maps f ↦ L ( f g ) f\mapsto L(fg) f ↦ L ( f g ) yield measurable k g ∈ L ∞ ( ℓ ) k_g\in L^\infty(\ell) k g ∈ L ∞ ( ℓ ) satisfying positivity, k 1 = 1 k_1=1 k 1 = 1 , and ∣ k g ∣ ≤ ∥ g ∥ ∞ |k_g|\leq\|g\|_\infty ∣ k g ∣ ≤ ∥ g ∥ ∞ . Extension and the Riesz representation theorem give measurable probability measures κ a \kappa_a κ a . The marginal identities
L π ( f 1 ) = ∫ A f d ℓ , L π ( ζ ∘ q ) = c ∫ T 3 ζ d x L_\pi(f\,1)=\int_{\mathcal{A}}f\,d\ell,
\qquad
L_\pi(\zeta\circ q)=c\int_{\mathbb{T}^3}\zeta\,dx L π ( f 1 ) = ∫ A f d ℓ , L π ( ζ ∘ q ) = c ∫ T 3 ζ d x
are preserved in the limit. Hence P Q \mathcal{P}_Q P Q is convex and weak-star compact.
To impose the nearest-lift constraint, define
d y ( a , z ) = ∣ y ( a ) − z ∣ 2 − d T 3 ( q y ( a ) , q z ) 2 ≥ 0. (117) d_y(a,z)=\lvert y(a)-z\rvert^2-d_{\mathbb{T}^3}(q y(a),qz)^2\ge0. \tag*{(117)} d y ( a , z ) = ∣ y ( a ) − z ∣ 2 − d T 3 ( q y ( a ) , q z ) 2 ≥ 0. (117)
Both d y d_y d y and c y ( a , z ) = ∣ y ( a ) − z ∣ 2 c_y(a,z)=\lvert y(a)-z\rvert^2 c y ( a , z ) = ∣ y ( a ) − z ∣ 2 belong to E Q \mathcal{E}_Q E Q . The nearest-lift constraint L π ( d y ) = 0 L_\pi(d_y)=0 L π ( d y ) = 0 is closed in this weak-star topology, and the cost L π ( c y ) L_\pi(c_y) L π ( c y ) attains its minimum on P Q \mathcal{P}_Q P Q . Measurability in a a a and continuity in z z z suffice: the topology is defined against L 1 ( A ; C ( Q ) ) L^1(A;C(Q)) L 1 ( A ; C ( Q )) , so no continuity of y y y on the label space is being assumed. The nonnegative defect turns the potentially discontinuous choice of a nearest lift at ties into a closed integral constraint.
Replacing lifts by nearest ones preserves the marginals and decreases the cost by L π ( d y ) L_\pi(d_y) L π ( d y ) . Thus minimizers are supported on nearest lifts. Disintegration of a torus coupling and measurable lift selection give
min π ∈ P Q L π ( c y ) = W 2 , T 3 2 ( q # y # ℓ , c d x ) . \min_{\pi\in\mathcal{P}_Q}L_\pi(c_y)=W_{2,\mathbb{T}^3}^2(q_{\#}y_{\#}\ell,c\,dx). π ∈ P Q min L π ( c y ) = W 2 , T 3 2 ( q # y # ℓ , c d x ) .
These minimizers form P y \mathcal{P}_y P y . Since their lifts lie in a fixed bounded set, P y \mathcal{P}_y P y is nonempty, convex, and compact, independently of any further enlargement of Q Q Q .
For W ∈ H W\in\mathcal{H} W ∈ H , we have W ( a ) ⋅ z ∈ E Q W(a)\cdot z\in\mathcal{E}_Q W ( a ) ⋅ z ∈ E Q and
⟨ B π , W ⟩ = L π ( W ⋅ z ) , ∥ B π ∥ H ≤ c 1 / 2 max z ∈ Q ∣ z ∣ . \langle B_\pi,W\rangle=L_\pi(W\cdot z),\qquad\lVert B_\pi\rVert_{\mathcal{H}}\le c^{1/2}\max_{z\in Q}\lvert z\rvert. ⟨ B π , W ⟩ = L π ( W ⋅ z ) , ∥ B π ∥ H ≤ c 1/2 z ∈ Q max ∣ z ∣ .
Thus the barycenter map is weak-star-to-weak continuous, and B y B_y B y is convex and weakly compact.
Writing
Λ w ( π ) = ∫ ( w ( a ) ⋅ z − 1 2 ∣ z ∣ 2 ) d π , \Lambda_w(\pi)=\int\left(w(a)\cdot z-\frac{1}{2}\lvert z\rvert^2\right)\,d\pi, Λ w ( π ) = ∫ ( w ( a ) ⋅ z − 2 1 ∣ z ∣ 2 ) d π ,
optimality gives G ( y ) = Λ y ( π ) G(y)=\Lambda_y(\pi) G ( y ) = Λ y ( π ) for π ∈ P y \pi\in\mathcal{P}_y π ∈ P y , while G ( w ) ≥ Λ w ( π ) G(w)\ge\Lambda_w(\pi) G ( w ) ≥ Λ w ( π ) for every w ∈ H w\in\mathcal{H} w ∈ H . Hence
G ( y + W ) ≥ G ( y ) + ⟨ B π , W ⟩ , W ∈ H , G(y+W)\ge G(y)+\langle B_\pi,W\rangle,\qquad W\in\mathcal{H}, G ( y + W ) ≥ G ( y ) + ⟨ B π , W ⟩ , W ∈ H ,
so B y ⊂ ∂ G ( y ) B_y\subset\partial G(y) B y ⊂ ∂ G ( y ) .
For the reverse inclusion, fix bounded W ∈ H W\in\mathcal{H} W ∈ H and put y ε = y + ε W y_\varepsilon=y+\varepsilon W y ε = y + ε W , 0 < ε ≤ 1 0<\varepsilon\le1 0 < ε ≤ 1 . Enlarging Q Q Q to contain the corresponding nearest lifts gives
∥ c y ε − c y ∥ E Q + ∥ d y ε − d y ∥ E Q ≤ C ε , \lVert c_{y_\varepsilon}-c_y\rVert_{\mathcal{E}_Q}+\lVert d_{y_\varepsilon}-d_y\rVert_{\mathcal{E}_Q}\le C\varepsilon, ∥ c y ε − c y ∥ E Q + ∥ d y ε − d y ∥ E Q ≤ C ε ,
with C C C depending on c c c , Q Q Q , ∥ y ∥ ∞ \lVert y\rVert_\infty ∥ y ∥ ∞ , ∥ W ∥ ∞ \lVert W\rVert_\infty ∥ W ∥ ∞ . For π ε ∈ P y ε \pi_\varepsilon\in\mathcal{P}_{y_\varepsilon} π ε ∈ P y ε and any π ∈ P y \pi\in\mathcal{P}_y π ∈ P y , optimality gives
L π ε ( c y ) ≤ L π ε ( c y ε ) + C ε ≤ L π ( c y ε ) + C ε ≤ L π ( c y ) + 2 C ε , L_{\pi_\varepsilon}(c_y)\le L_{\pi_\varepsilon}(c_{y_\varepsilon})+C\varepsilon\le L_\pi(c_{y_\varepsilon})+C\varepsilon\le L_\pi(c_y)+2C\varepsilon, L π ε ( c y ) ≤ L π ε ( c y ε ) + C ε ≤ L π ( c y ε ) + C ε ≤ L π ( c y ) + 2 C ε ,
L π ε ( d y ) ≤ C ε . L_{\pi_\varepsilon}(d_y)\le C\varepsilon. L π ε ( d y ) ≤ C ε .
Every weak-star cluster point of π ε \pi_\varepsilon π ε therefore lies in P y \mathcal{P}_y P y , with both marginals and the nearest-lift constraint preserved.
The affine maximization formula yields
⟨ B π , W ⟩ ≤ G ( y + ε W ) − G ( y ) ε ≤ ⟨ B π ε , W ⟩ . (118) \langle B_\pi,W\rangle\le\frac{G(y+\varepsilon W)-G(y)}{\varepsilon}\le\langle B_{\pi_\varepsilon},W\rangle. \tag*{(118)} ⟨ B π , W ⟩ ≤ ε G ( y + ε W ) − G ( y ) ≤ ⟨ B π ε , W ⟩ . (118)
Choose π ∈ P y \pi\in\mathcal{P}_y π ∈ P y maximizing ⟨ B π , W ⟩ \langle B_\pi, W\rangle ⟨ B π , W ⟩ ; the maximum exists by weak compactness. For the upper bound, take a subsequence of π ε \pi_\varepsilon π ε realizing the upper limit and use its weak-star cluster point in P y \mathcal{P}_y P y . The two sides of (B.9) then coincide. Convexity of G G G ensures existence of the one-sided derivative, and we obtain
G ′ ( y ; W ) = max B ∈ B y ⟨ B , W ⟩ . (119) G'(y;W)=\max_{B\in\mathcal{B}_y}\langle B,W\rangle. \tag*{(119)} G ′ ( y ; W ) = B ∈ B y max ⟨ B , W ⟩ . (119)
If B 0 ∈ ∂ G ( y ) ∖ B y B_0 \in\partial G(y) \setminus\mathcal{B}_y B 0 ∈ ∂ G ( y ) ∖ B y , separation provides W ∈ H W \in\mathcal{H} W ∈ H such that ⟨ B 0 , W ⟩ > sup B ∈ B y ⟨ B , W ⟩ \langle B_0,W\rangle> \sup_{B\in\mathcal{B}_y}\langle B,W\rangle ⟨ B 0 , W ⟩ > sup B ∈ B y ⟨ B , W ⟩ . Since B y \mathcal{B}_y B y is bounded, truncation preserves this strict separation with bounded W W W . The subgradient inequality then contradicts (B.10), proving (B.6).
For x = X ( a ) x=X(a) x = X ( a ) and y ∗ = Y K ( X ) y_* = Y_K(X) y ∗ = Y K ( X ) , the optimal plan π ∗ \pi_* π ∗ of (34) also minimizes the mixed quadratic cost
J K ( π ) = 1 2 ∫ [ K − 1 K ∣ z − x ∣ 2 + 1 K ∣ E [ z ∣ x ] − x ∣ 2 ] d π ( x , z ) , (120) \mathcal{J}_K(\pi)=\frac{1}{2}\int\left[\frac{K-1}{K}|z-x|^2+\frac{1}{K}|\mathbb{E}[z\mid x]-x|^2\right]\,\mathrm{d}\pi(x,z), \tag*{(120)} J K ( π ) = 2 1 ∫ [ K K − 1 ∣ z − x ∣ 2 + K 1 ∣ E [ z ∣ x ] − x ∣ 2 ] d π ( x , z ) , (120)
over plans with prescribed torus target marginal and variable integer lifts. Set
L ( y , π ) = K − 1 2 ∥ y ∥ 2 − K ⟨ x , y ⟩ + ∫ y ⋅ z d π + K 2 ∥ x ∥ 2 − 1 2 ∫ ∣ z ∣ 2 d π . L(y,\pi)=\frac{K-1}{2}\|y\|^2-K\langle x,y\rangle+\int y\cdot z\,\mathrm{d}\pi+\frac{K}{2}\|x\|^2-\frac{1}{2}\int|z|^2\,\mathrm{d}\pi. L ( y , π ) = 2 K − 1 ∥ y ∥ 2 − K ⟨ x , y ⟩ + ∫ y ⋅ z d π + 2 K ∥ x ∥ 2 − 2 1 ∫ ∣ z ∣ 2 d π .
Optimality of π ∗ \pi_* π ∗ and ( K − 1 ) y ∗ − K x + E π ∗ [ z ∣ x ] = 0 (K-1)y_*-Kx+\mathbb{E}_{\pi_*}[z\mid x]=0 ( K − 1 ) y ∗ − K x + E π ∗ [ z ∣ x ] = 0 give a saddle point ( y ∗ , π ∗ ) (y_*,\pi_*) ( y ∗ , π ∗ ) . Completing the square yields
inf y L ( y , π ) = − K K − 1 J K ( π ) , \inf_y L(y,\pi)=-\frac{K}{K-1}\mathcal{J}_K(\pi), y inf L ( y , π ) = − K − 1 K J K ( π ) ,
and the saddle-point inequalities yield
inf y L ( y , π ) ≤ L ( y ∗ , π ) ≤ L ( y ∗ , π ∗ ) = inf y L ( y , π ∗ ) . \inf_y L(y,\pi)\leq L(y_*,\pi)\leq L(y_*,\pi_*)=\inf_y L(y,\pi_*). y inf L ( y , π ) ≤ L ( y ∗ , π ) ≤ L ( y ∗ , π ∗ ) = y inf L ( y , π ∗ ) .
Hence π ∗ \pi_* π ∗ minimizes J K \mathcal{J}_K J K , with the prescribed target marginal. The first term of J K \mathcal{J}_K J K is linear in the plan and the second is strictly convex in its conditional barycenter. If two minimizers had barycenters T 1 ≠ T 2 T_1\ne T_2 T 1 = T 2 in L 2 ( μ ) L^2(\mu) L 2 ( μ ) , their equally weighted mixture would lower the cost by
1 8 K ∫ ∣ T 1 − T 2 ∣ 2 d μ > 0. \frac{1}{8K}\int|T_1-T_2|^2\,\mathrm{d}\mu>0. 8 K 1 ∫ ∣ T 1 − T 2 ∣ 2 d μ > 0.
Thus the barycenter is unique even if the optimal plan is not.
Auxiliary Entropy and High-Frequency Estimates
Lemma C.1 supplies the initialized identities used in the entropy comparison of Section 4. Lemma C.2 supplies the common kernel for the truncated continuity source and convective commutator in Lemma 4.2. These statements concern different limiting operations: removal of the vacuum regularization at fixed truncation level, and decay as the spatial smoothing scale tends to zero. In the latter case, boundedness of the normalized truncation and its renormalization coefficient makes the same kernel estimate uniform in the truncation level. In the former, the truncation level stays fixed until the initialized entropy identities have been obtained and compared.
Lemma C.1 (Initialized truncated entropy identity). Let ( ϱ , v ) (\varrho,v) ( ϱ , v ) satisfy the initialized renormalized continuity equation on Q T Q_T Q T , with ϱ ≥ 0 \varrho\ge0 ϱ ≥ 0 , ϱ ∈ L ∞ ( 0 , T ; L γ ( T 3 ) ) \varrho\in L^\infty(0,T;L^\gamma(\mathbb{T}^3)) ϱ ∈ L ∞ ( 0 , T ; L γ ( T 3 )) , v ∈ L 2 ( 0 , T ; H 1 ( T 3 ) ) v\in L^2(0,T;H^1(\mathbb{T}^3)) v ∈ L 2 ( 0 , T ; H 1 ( T 3 )) , and γ > 1 \gamma>1 γ > 1 . For every k ≥ 1 k\ge1 k ≥ 1 and ψ ∈ C c ∞ ( [ 0 , T ) ) \psi\in C^\infty_c([0,T)) ψ ∈ C c ∞ ([ 0 , T )) ,
− ∫ 0 T ψ ′ ( t ) ∫ T 3 L k ( ϱ ) d x d t + ∫ 0 T ψ ( t ) ∫ T 3 T k ( ϱ ) div v d x d t = ψ ( 0 ) ∫ T 3 L k ( ϱ 0 ) d x . -\int_0^T\psi'(t)\int_{\mathbb{T}^3}L_k(\varrho)\,dx\,dt+\int_0^T\psi(t)\int_{\mathbb{T}^3}T_k(\varrho)\operatorname{div}v\,dx\,dt
=\psi(0)\int_{\mathbb{T}^3}L_k(\varrho_0)\,dx. − ∫ 0 T ψ ′ ( t ) ∫ T 3 L k ( ϱ ) d x d t + ∫ 0 T ψ ( t ) ∫ T 3 T k ( ϱ ) div v d x d t = ψ ( 0 ) ∫ T 3 L k ( ϱ 0 ) d x . The identity applies to the approximating and limiting densities in § 4 and yields ( 4.20 ) (4.20) ( 4.20 ) –( 4.21 ) (4.21) ( 4.21 ) after passage to the weak limits.
Proof. Fix k ≥ 1 k\ge1 k ≥ 1 . Since L k L_k L k is affine on [ 3 k , ∞ ) [3k,\infty) [ 3 k , ∞ ) , its slope is
α k = ∫ 1 3 k T k ( s ) s − 2 d s + 2 3 . \alpha_k=\int_1^{3k}T_k(s)s^{-2}\,ds+\frac{2}{3}. α k = ∫ 1 3 k T k ( s ) s − 2 d s + 3 2 . Choose a smooth nonnegative w w w on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , supported in [ 0 , 1 ] [0,1] [ 0 , 1 ] , constant and positive near zero, with ∫ 0 1 w = 1 \int_0^1w=1 ∫ 0 1 w = 1 . Set ω ( s ) = ∫ 0 s w ( r ) d r \omega(s)=\int_0^s w(r)\,dr ω ( s ) = ∫ 0 s w ( r ) d r and, for 0 < ε < 1 0<\varepsilon<1 0 < ε < 1 , define
a k , ε ( s ) = T k ( s ) ω ( s / ε ) , a_{k,\varepsilon}(s)=T_k(s)\omega(s/\varepsilon), a k , ε ( s ) = T k ( s ) ω ( s / ε ) , L k , ε ( z ) = z ∫ 1 z a k , ε ( s ) s − 2 d s ( z > 0 ) , L k , ε ( 0 ) = 0. L_{k,\varepsilon}(z)=z\int_1^z a_{k,\varepsilon}(s)s^{-2}\,ds\quad(z>0),\qquad L_{k,\varepsilon}(0)=0. L k , ε ( z ) = z ∫ 1 z a k , ε ( s ) s − 2 d s ( z > 0 ) , L k , ε ( 0 ) = 0. The algebraic identities are
z L k , ε ′ ( z ) − L k , ε ( z ) = a k , ε ( z ) , L k , ε ′ ′ ( z ) = a k , ε ′ ( z ) z ≥ 0 ( z > 0 ) . zL'_{k,\varepsilon}(z)-L_{k,\varepsilon}(z)=a_{k,\varepsilon}(z),\qquad L''_{k,\varepsilon}(z)=\frac{a'_{k,\varepsilon}(z)}{z}\ge0\quad(z>0). z L k , ε ′ ( z ) − L k , ε ( z ) = a k , ε ( z ) , L k , ε ′′ ( z ) = z a k , ε ′ ( z ) ≥ 0 ( z > 0 ) . Here T k T_k T k and ω \omega ω are nonnegative and nondecreasing. Near zero, a k , ε ( z ) = O ( z 2 / ε ) a_{k,\varepsilon}(z)=O(z^2/\varepsilon) a k , ε ( z ) = O ( z 2 / ε ) , so the first derivative extends continuously to zero. Thus the regularization is compatible with vacuum and preserves convexity. Then L k , ε − α k z L_{k,\varepsilon}-\alpha_kz L k , ε − α k z is bounded and C 1 C^1 C 1 , with derivative vanishing above 3 k 3k 3 k . Moreover,
0 ≤ L k , ε ( z ) − L k ( z ) ≤ z log ( ε / z ) 1 { 0 < z < ε } ≤ ε exp ( 1 ) , 0\le L_{k,\varepsilon}(z)-L_k(z)\le z\log(\varepsilon/z)\mathbf{1}_{\{0<z<\varepsilon\}}\le\frac{\varepsilon}{\exp(1)}, 0 ≤ L k , ε ( z ) − L k ( z ) ≤ z log ( ε / z ) 1 { 0 < z < ε } ≤ exp ( 1 ) ε , ∣ a k , ε ( z ) − T k ( z ) ∣ ≤ ε . \left|a_{k,\varepsilon}(z)-T_k(z)\right|\le\varepsilon. ∣ a k , ε ( z ) − T k ( z ) ∣ ≤ ε . Writing d n = div u n d_n=\operatorname{div}u_n d n = div u n and D T = sup n ∥ d n ∥ L 2 ( Q T ) D_T=\sup_n\|d_n\|_{L^2(Q_T)} D T = sup n ∥ d n ∥ L 2 ( Q T ) , the source, time-derivative, and initial errors for a time test ψ \psi ψ are bounded by
ε ∥ ψ ∥ ∞ ∣ Q T ∣ 1 / 2 D T , ε exp ( 1 ) ∥ ψ ′ ∥ 1 , ε exp ( 1 ) ∣ ψ ( 0 ) ∣ . \varepsilon\|\psi\|_\infty|Q_T|^{1/2}D_T,\qquad\frac{\varepsilon}{\exp(1)}\|\psi'\|_1,\qquad\frac{\varepsilon}{\exp(1)}|\psi(0)|. ε ∥ ψ ∥ ∞ ∣ Q T ∣ 1/2 D T , exp ( 1 ) ε ∥ ψ ′ ∥ 1 , exp ( 1 ) ε ∣ ψ ( 0 ) ∣. Apply initialized renormalization to L k , ε − α k z L_{k,\varepsilon}-\alpha_kz L k , ε − α k z and add α k \alpha_k α k times the continuity equation. For the approximating solution,
− ∫ 0 T ψ ′ ( t ) ∫ T 3 L k , ε ( ϱ n ) d x d t + ∫ 0 T ψ ( t ) ∫ T 3 a k , ε ( ϱ n ) d n d x d t = ψ ( 0 ) ∫ T 3 L k , ε ( ϱ 0 n ) d x . -\int_0^T\psi'(t)\int_{\mathbb{T}^3}L_{k,\varepsilon}(\varrho_n)\,dx\,dt+\int_0^T\psi(t)\int_{\mathbb{T}^3}a_{k,\varepsilon}(\varrho_n)d_n\,dx\,dt
=\psi(0)\int_{\mathbb{T}^3}L_{k,\varepsilon}(\varrho_{0n})\,dx. − ∫ 0 T ψ ′ ( t ) ∫ T 3 L k , ε ( ϱ n ) d x d t + ∫ 0 T ψ ( t ) ∫ T 3 a k , ε ( ϱ n ) d n d x d t = ψ ( 0 ) ∫ T 3 L k , ε ( ϱ 0 n ) d x . The same argument applies to ( ϱ , u , ϱ 0 ) (\varrho,u,\varrho_0) ( ϱ , u , ϱ 0 ) . Letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 proves the stated identity, including its initial term. The uniform error bounds and the weak limits of L k ( ϱ n ) L_k(\varrho_n) L k ( ϱ n ) and T k ( ϱ n ) d n T_k(\varrho_n)d_n T k ( ϱ n ) d n then give ( 4.20 ) (4.20) ( 4.20 ) ; the limiting equation gives ( 4.21 ) (4.21) ( 4.21 ) .
Let A = ∇ Δ T 3 − 1 A=\nabla\Delta_{\mathbb{T}^3}^{-1} A = ∇ Δ T 3 − 1 and R i j = ∂ j ∂ i Δ T 3 − 1 R_{ij}=\partial_j\partial_i\Delta_{\mathbb{T}^3}^{-1} R ij = ∂ j ∂ i Δ T 3 − 1 . Choose an even χ ∈ C ∞ ( B 1 ) \chi\in C^\infty(B_1) χ ∈ C ∞ ( B 1 ) with χ ≥ 0 \chi\ge0 χ ≥ 0 and ∫ χ = 1 \int\chi=1 ∫ χ = 1 , and use the Euclidean Fourier convention χ ^ ( ξ ) = ∫ R 3 e − 2 π i x ⋅ ξ χ ( x ) d x \widehat{\chi}(\xi)=\int_{\mathbb{R}^3}e^{-2\pi i x\cdot\xi}\chi(x)\,\mathrm{d}x χ ( ξ ) = ∫ R 3 e − 2 π i x ⋅ ξ χ ( x ) d x . Define S a S_a S a by convolution with the periodization of χ a = a − 3 χ ( ⋅ / a ) \chi_a=a^{-3}\chi(\cdot/a) χ a = a − 3 χ ( ⋅ / a ) , where 0 < a < 1 / 8 0<a<1/8 0 < a < 1/8 , and set H a = I − S a H_a=I-S_a H a = I − S a , A a , i = A i H a A_{a,i}=A_iH_a A a , i = A i H a , and L i j , a = R i j H a L_{ij,a}=R_{ij}H_a L ij , a = R ij H a . Since H a H_a H a annihilates constants,
H a f ‾ = 0 , ∑ i L i i , a = H a . (121) \overline{H_af}=0,\qquad\sum_i L_{ii,a}=H_a. \tag*{(121)} H a f = 0 , i ∑ L ii , a = H a . (121)
The unfiltered identity is ∑ i R i i f = f − f ‾ \sum_i R_{ii}f=f-\overline{f} ∑ i R ii f = f − f .
Lemma C.2. The periodic kernel A a A_a A a of A a A_a A a satisfies ∥ A a ∥ 1 ≤ C a \|A_a\|_1\le Ca ∥ A a ∥ 1 ≤ C a . There is a nonnegative periodic kernel K a K_a K a such that
∥ K a ∥ s ≤ C s ( a 3 / s − 2 + a 2 ) , 1 ≤ s < 3 / 2. (122) \|K_a\|_s\le C_s(a^{3/s-2}+a^2),\qquad1\le s<3/2. \tag*{(122)} ∥ K a ∥ s ≤ C s ( a 3/ s − 2 + a 2 ) , 1 ≤ s < 3/2. (122) If g ∈ L ∞ g\in L^\infty g ∈ L ∞ , u ∈ H 1 ( T 3 ) u\in H^1(\mathbb{T}^3) u ∈ H 1 ( T 3 ) , and ∣ h ∣ ≤ C h ∣ ∇ u ∣ |h|\le C_h|\nabla u| ∣ h ∣ ≤ C h ∣∇ u ∣ , then
∣ L i j , a ( g u j ) − u j L i j , a g + A a , i h ∣ ≤ C ( ∥ g ∥ ∞ + C h ) K a ∗ T 3 ∣ ∇ u ∣ almost everywhere . (123) |L_{ij,a}(gu_j)-u_jL_{ij,a}g+A_{a,i}h|\le C(\|g\|_\infty+C_h)K_a*_{\mathbb{T}^3}|\nabla u|\quad\text{almost everywhere}. \tag*{(123)} ∣ L ij , a ( g u j ) − u j L ij , a g + A a , i h ∣ ≤ C ( ∥ g ∥ ∞ + C h ) K a ∗ T 3 ∣∇ u ∣ almost everywhere . (123) The source h h h need not have zero spatial mean.
Proof. The local potential and principal-value kernels are those of the classical singular-integral calculus [40 ] . We retain the contact term and periodize the integrable high-frequency difference, so that the periodic zero mode is fixed by the construction itself. Begin in R 3 \mathbb{R}^3 R 3 with A R ( z ) = z / ( 4 π ∣ z ∣ 3 ) A^{\mathbb{R}}(z)=z/(4\pi|z|^3) A R ( z ) = z / ( 4 π ∣ z ∣ 3 ) and form the integrable difference A a R = A R − A R ∗ χ a A_a^{\mathbb{R}}=A^{\mathbb{R}}-A^{\mathbb{R}}*\chi_a A a R = A R − A R ∗ χ a for periodization. Write r = ∣ z ∣ r=|z| r = ∣ z ∣ . For r ≤ 2 a r\le2a r ≤ 2 a , the difference is bounded by C r − 2 Cr^{-2} C r − 2 . For r > 2 a r>2a r > 2 a , evenness of χ a \chi_a χ a cancels the first-order Taylor term, giving
A a R ( z ) = − ∫ χ a ( y ) ∫ 0 1 ( 1 − t ) D 2 A R ( z − t y ) [ y , y ] d t d y , ∣ A a R ( z ) ∣ ≤ C a 2 r − 4 . A_a^{\mathbb{R}}(z)=-\int\chi_a(y)\int_0^1(1-t)D^2A^{\mathbb{R}}(z-ty)[y,y]\,\mathrm{d}t\,\mathrm{d}y,\qquad|A_a^{\mathbb{R}}(z)|\le Ca^2r^{-4}. A a R ( z ) = − ∫ χ a ( y ) ∫ 0 1 ( 1 − t ) D 2 A R ( z − t y ) [ y , y ] d t d y , ∣ A a R ( z ) ∣ ≤ C a 2 r − 4 .
Thus
∥ A a R ∥ 1 ≤ C ( ∫ 0 2 a 1 d r + a 2 ∫ 2 a ∞ r − 2 d r ) ≤ C a . (124) \|A_a^{\mathbb{R}}\|_1\le C\left(\int_0^{2a}1\,\mathrm{d}r+a^2\int_{2a}^{\infty}r^{-2}\,\mathrm{d}r\right)\le Ca. \tag*{(124)} ∥ A a R ∥ 1 ≤ C ( ∫ 0 2 a 1 d r + a 2 ∫ 2 a ∞ r − 2 d r ) ≤ C a . (124)
Since A a R A_a^{\mathbb{R}} A a R is integrable and odd, its periodization has zero Fourier coefficient at frequency zero. Its nonzero coefficients are − i ℓ [ 1 − χ ^ ( a ℓ ) ] / ( 2 π ∣ ℓ ∣ 2 ) -i\ell[1-\widehat{\chi}(a\ell)]/(2\pi|\ell|^2) − i ℓ [ 1 − χ ( a ℓ )] / ( 2 π ∣ ℓ ∣ 2 ) , those of A a A_a A a . It is therefore the periodic kernel of A a A_a A a .
Differentiation of the Euclidean kernel produces a Dirac term at the origin as well as a principal-value kernel:
{ ∂ j A i R = pv F i j + δ i j 3 δ 0 , F i j ( z ) = δ i j ∣ z ∣ 2 − 3 z i z j 4 π ∣ z ∣ 5 . \begin{cases}
\partial_jA_i^{\mathbb{R}}=\operatorname{pv}F_{ij}+\dfrac{\delta_{ij}}{3}\delta_0,\\
F_{ij}(z)=\dfrac{\delta_{ij}|z|^2-3z_iz_j}{4\pi|z|^5}.
\end{cases} ⎩ ⎨ ⎧ ∂ j A i R = pv F ij + 3 δ ij δ 0 , F ij ( z ) = 4 π ∣ z ∣ 5 δ ij ∣ z ∣ 2 − 3 z i z j .
The corresponding high-frequency part is pv F i j − ( pv F i j ) ∗ χ a + δ i j ( δ 0 − χ a ) \operatorname{pv}F_{ij}-(\operatorname{pv}F_{ij})*\chi_a+\delta_{ij}(\delta_0-\chi_a) pv F ij − ( pv F ij ) ∗ χ a + δ ij ( δ 0 − χ a ) . The Dirac term cancels in the commutator because the velocity difference vanishes at the origin, but the convolved term involving χ a \chi_a χ a must remain. Away from the origin, the required kernel is
W i j , a = F i j − ( p.v. F i j ) ∗ χ a − δ i j 3 χ a , ∣ W a ( z ) ∣ ≤ { C ( r − 3 + a − 3 ) , r ≤ 2 a , C a 2 r − 5 , r > 2 a . W_{ij,a}=F_{ij}-(\operatorname{p.v.} F_{ij})*\chi_a-\frac{\delta_{ij}}{3}\chi_a,\qquad|W_a(z)|\le\begin{cases} C(r^{-3}+a^{-3}), & r\le2a, \\ Ca^2r^{-5}, & r>2a. \end{cases} W ij , a = F ij − ( p.v. F ij ) ∗ χ a − 3 δ ij χ a , ∣ W a ( z ) ∣ ≤ { C ( r − 3 + a − 3 ) , C a 2 r − 5 , r ≤ 2 a , r > 2 a .
For r ≤ 2 a r\le2a r ≤ 2 a , rescale to a = 1 a=1 a = 1 . Since F i j F_{ij} F ij has zero integral on every spherical shell, the singular part of the convolution can be written as ∫ ∣ y ∣ < 1 F i j ( y ) [ χ ( z − y ) − χ ( z ) ] d y \int_{|y|<1}F_{ij}(y)[\chi(z-y)-\chi(z)]\,\mathrm{d}y ∫ ∣ y ∣ < 1 F ij ( y ) [ χ ( z − y ) − χ ( z )] d y . The difference estimate for the smooth kernel bounds its absolute value by C ∫ 0 1 d r C\int_0^1\mathrm{d}r C ∫ 0 1 d r . For r > 2 a r>2a r > 2 a , evenness of χ a \chi_a χ a again cancels the first-order Taylor term, and the point lies outside the support of χ a \chi_a χ a .
First let u u u be smooth and periodic and g g g bounded and periodic, and extend both periodically to R 3 \mathbb{R}^3 R 3 . The commutator is
[ L i j , a , u j ] g ( x ) = ∫ R 3 W i j , a ( z ) [ u j ( x − z ) − u j ( x ) ] g ( x − z ) d z . [L_{ij,a},u_j]g(x)=\int_{\mathbb{R}^3}W_{ij,a}(z)[u_j(x-z)-u_j(x)]g(x-z)\,\mathrm{d}z. [ L ij , a , u j ] g ( x ) = ∫ R 3 W ij , a ( z ) [ u j ( x − z ) − u j ( x )] g ( x − z ) d z .
Apply the line-segment formula to u j ( x − z ) − u j ( x ) u_j(x-z)-u_j(x) u j ( x − z ) − u j ( x ) and change variables w = t z w=tz w = t z . The absolute value of the commutator is bounded by
∥ g ∥ ∞ ∫ R 3 ∣ ∇ u ( x − w ) ∣ ∣ w ∣ ∫ 0 1 ∣ W a ( w / t ) ∣ t − 4 d t d w . \|g\|_\infty\int_{\mathbb{R}^3}|\nabla u(x-w)|\,|w|\int_0^1|W_a(w/t)|t^{-4}\,\mathrm{d}t\,\mathrm{d}w. ∥ g ∥ ∞ ∫ R 3 ∣∇ u ( x − w ) ∣ ∣ w ∣ ∫ 0 1 ∣ W a ( w / t ) ∣ t − 4 d t d w .
For r = ∣ w ∣ ≤ 2 a r=|w|\le2a r = ∣ w ∣ ≤ 2 a , split the integral at t 0 = r / ( 2 a ) t_0=r/(2a) t 0 = r / ( 2 a ) to obtain the bound
C a 2 r − 4 ∫ 0 t 0 t d t + C r − 2 ∫ t 0 1 t − 1 d t + C r a − 3 ∫ t 0 1 t − 4 d t ≤ C r − 2 [ 1 + log ( 2 a / r ) ] . Ca^2r^{-4}\int_0^{t_0}t\,\mathrm{d}t+Cr^{-2}\int_{t_0}^1t^{-1}\,\mathrm{d}t+Cra^{-3}\int_{t_0}^1t^{-4}\,\mathrm{d}t\le Cr^{-2}[1+\log(2a/r)]. C a 2 r − 4 ∫ 0 t 0 t d t + C r − 2 ∫ t 0 1 t − 1 d t + C r a − 3 ∫ t 0 1 t − 4 d t ≤ C r − 2 [ 1 + log ( 2 a / r )] .
For r > 2 a r>2a r > 2 a , only the far-field estimate occurs, giving C a 2 r − 4 Ca^2r^{-4} C a 2 r − 4 . Adjusting the constant on a < r < 2 a a<r<2a a < r < 2 a , we may take
K a ( w ) = { ∣ w ∣ − 2 [ 1 + log ( a / ∣ w ∣ ) ] , 0 < ∣ w ∣ ≤ a , a 2 ∣ w ∣ − 4 , ∣ w ∣ > a . K_a(w)=\begin{cases}|w|^{-2}[1+\log(a/|w|)], & 0<|w|\le a, \\ a^2|w|^{-4}, & |w|>a.\end{cases} K a ( w ) = { ∣ w ∣ − 2 [ 1 + log ( a /∣ w ∣ )] , a 2 ∣ w ∣ − 4 , 0 < ∣ w ∣ ≤ a , ∣ w ∣ > a .
The near- and far-field estimates also bound the potential kernel A a R A_a^{\mathbb{R}} A a R by C K a CK_a C K a , giving ∣ A a h ∣ ≤ C C h K K a ∗ R 3 ∣ ∇ u ∣ |A_a h|\le CCh_K K_a*_{\mathbb{R}^3}|\nabla u| ∣ A a h ∣ ≤ C C h K K a ∗ R 3 ∣∇ u ∣ . Thus K a K_a K a controls both the continuity source and the convective commutator.
Set K a ( x ) = ∑ ℓ ∈ Z 3 K a ( x + ℓ ) K_a(x)=\sum_{\ell\in\mathbb{Z}^3}K_a(x+\ell) K a ( x ) = ∑ ℓ ∈ Z 3 K a ( x + ℓ ) . Radial integration gives
∥ K a ∥ L s ( R 3 ) s = 4 π a 3 − 2 s [ ∫ 0 1 r 2 − 2 s ( 1 + ∣ log r ∣ ) s d r + ∫ 1 ∞ r 2 − 4 s d r ] . \|K_a\|_{L^s(\mathbb{R}^3)}^s=4\pi a^{3-2s}\left[\int_0^1r^{2-2s}(1+|\log r|)^s\,\mathrm{d}r+\int_1^\infty r^{2-4s}\,\mathrm{d}r\right]. ∥ K a ∥ L s ( R 3 ) s = 4 π a 3 − 2 s [ ∫ 0 1 r 2 − 2 s ( 1 + ∣ log r ∣ ) s d r + ∫ 1 ∞ r 2 − 4 s d r ] .
The first integral is finite precisely when 2 − 2 s > − 1 2-2s>-1 2 − 2 s > − 1 , or s < 3 / 2 s<3/2 s < 3/2 ; the second is finite for s ≥ 1 s\ge1 s ≥ 1 . On the centered fundamental cube, nonzero lattice points satisfy ∣ x + ℓ ∣ ≥ c ∣ ℓ ∣ |x+\ell|\ge c|\ell| ∣ x + ℓ ∣ ≥ c ∣ ℓ ∣ , so the remaining periodic terms are bounded by C a 2 ∑ ℓ ≠ 0 ∣ ℓ ∣ − 4 ≤ C a 2 Ca^2\sum_{\ell\ne0}|\ell|^{-4}\le Ca^2 C a 2 ∑ ℓ = 0 ∣ ℓ ∣ − 4 ≤ C a 2 . This proves (C.2), including the contribution from periodization.
Finally, approximate u ∈ H 1 u\in H^1 u ∈ H 1 by smooth periodic functions. Young’s inequality passes the convolution bound to the limit, and boundedness of R i j R_{ij} R ij on finite-exponent L r L^r L r spaces identifies the distributional limit. An almost-everywhere convergent subsequence gives (C.3). □ \square □
Acknowledgements
The research of G.Y. Hong was supported by the National Natural Science Foundation of China (Grant No. 12522111), the Guangdong Provincial Pearl River Talents Program (Grant No. 2023QN10X436), the Guangdong Basic and Applied Basic Research Foundation (Grant No. 2024A1515012306), and the Fundamental Research Funds for the Central Universities (Grant No. 2025ZYGXZR040). The research of H.Y. Peng was supported by the National Natural Science Foundation of China #12271112 and the Basic and Applied Basic Research Project of Guangzhou #SL2023A04J01554. The research of C.J. Zhu was supported by the National Natural Science Foundation of China #12571233 and the Basic and Applied Basic Research Foundation of Guangdong Province #2026A1515012081.
Data Availability: No datasets were generated or analysed for this study. Conflicts of Interest: The authors declare that they have no conflicts of interest. AI declaration: The authors acknowledge the use of generative AI tools in developing and checking mathematical arguments and in drafting and revising the manuscript. The authors take full responsibility for the mathematical content.
References
Guangyi Hong
School of Mathematics, South China University of Technology Guangzhou 510641, P. R. China magyhong@scut.edu.cn
Hongyun Peng
School of Mathematics and Statistics, Guangdong University of Technology Guangzhou 510006, P. R. China hypengmath@gdut.edu.cn
Changjiang Zhu
School of Mathematics, South China University of Technology Guangzhou 510641, P. R. China machjzhu@scut.edu.cn