Introduction
Classical micromagnetics describes the magnetisation by a vector field of unit length, leading to the Landau–Lifshitz–Gilbert (LLG) equation. This description is appropriate at sufficiently low temperatures, where longitudinal variations of the magnetisation are negligible, but it becomes inadequate near and above the Curie temperature. In that regime, the magnetisation represents a local thermal average of microscopic spins and its magnitude is no longer fixed. The Landau–Lifshitz–Bloch (LLB) equation introduced by Garanin in [14 , 15 ] removes this constraint and allows the magnetisation to take values in the closed unit ball. It has since been extensively studied in the physics literature [8 , 16 , 20 , 21 ] to model ultrafast heat-assisted magnetisation dynamics relevant in magnetic memory at high temperatures.
Starting from a Fokker–Planck equation for a classical spin, Garanin derives an evolution equation for the averaged spin polarisation m m m in which the relaxation coefficients depend nonlinearly on ∣ m ∣ |m| ∣ m ∣ through the inverse Langevin function; see [15 Eq. (2.16)] . For a ferromagnet, the magnetic field in this equation is subsequently replaced by a molecular mean field, whose continuum approximation consists of a dominant homogeneous exchange contribution together with the usual spatial exchange, anisotropy, and applied-field contributions.
Motivated by the preceding discussion, we re-examine the nonlinear magnetisation equation originally derived by Garanin [15 Eq. (2.16)] , before the further approximation leading to the reduced macroscopic LLB equation [15 Eqs. (4.13)–(4.15)] . Specifically, after rescaling the physical parameters, we consider a magnetisation m : [ 0 , ∞ ) × D → R 3 m \colon[0,\infty) \times\mathcal{D} \to\mathbb{R}^{3} m : [ 0 , ∞ ) × D → R 3 , where ∣ m ∣ ≤ 1 |m| \le1 ∣ m ∣ ≤ 1 , satisfying
∂ t m = γ m × H M F A ( m ) + M ( m ) H M F A ( m ) − 2 θ m in ( 0 , ∞ ) × D , (1) \partial_t m = \gamma m \times H_{\mathrm{MFA}}(m) + \mathbb{M}(m)H_{\mathrm{MFA}}(m) - 2\theta m
\quad\text{in } (0,\infty) \times\mathcal{D}, \tag*{(1)} ∂ t m = γ m × H MFA ( m ) + M ( m ) H MFA ( m ) − 2 θ m in ( 0 , ∞ ) × D , (1)
H M F A ( m ) = J m + α Δ m + A m + H d [ m ] in ( 0 , ∞ ) × D , (2) H_{\mathrm{MFA}}(m) = Jm + \alpha\Delta m + Am + H_{\mathrm{d}}[m]
\quad\text{in } (0,\infty) \times\mathcal{D}, \tag*{(2)} H MFA ( m ) = J m + α Δ m + A m + H d [ m ] in ( 0 , ∞ ) × D , (2)
curl H d [ m ] = 0 , div ( H d [ m ] + m 1 D ) = 0 in ( 0 , ∞ ) × R 3 , (3) \operatorname{curl} H_{\mathrm{d}}[m] = 0,\qquad\operatorname{div}\left(H_{\mathrm{d}}[m] + m1_{\mathcal{D}}\right) = 0
\quad\text{in } (0,\infty) \times\mathbb{R}^{3}, \tag*{(3)} curl H d [ m ] = 0 , div ( H d [ m ] + m 1 D ) = 0 in ( 0 , ∞ ) × R 3 , (3)
∂ ν m = 0 on ( 0 , ∞ ) × ∂ D , (4) \partial_{\nu}m = 0
\quad\text{on } (0,\infty) \times\partial\mathcal{D}, \tag*{(4)} ∂ ν m = 0 on ( 0 , ∞ ) × ∂ D , (4)
m ( 0 ) = m 0 in D . (5) m(0) = m_{0}
\quad\text{in } \mathcal{D}. \tag*{(5)} m ( 0 ) = m 0 in D . (5)
The constants γ \gamma γ and α > 0 \alpha> 0 α > 0 denote the gyromagnetic ratio and the exchange coefficient, respectively, while ν \nu ν denotes the outward-pointing normal to ∂ D \partial\mathcal{D} ∂ D . The term J m Jm J m represents the homogeneous molecular exchange, where J > 0 J > 0 J > 0 . The matrix A ∈ R 3 × 3 A \in\mathbb{R}^{3\times3} A ∈ R 3 × 3 contains the local anisotropy contributions and satisfies A = A ⊤ A = A^{\top} A = A ⊤ , while H d [ m ] H_{\mathrm{d}}[m] H d [ m ] is the physical stray field. The field H M F A ( m ) H_{\mathrm{MFA}}(m) H MFA ( m ) is the molecular field arising from the mean-field approximation and should be distinguished from the usual micromagnetic field. Accordingly, for emphasis we write
H M F A ( m ) = J m + H m i c ( m ) , where H m i c ( m ) = α Δ m + A m + H d [ m ] . (6) H_{\mathrm{MFA}}(m) = Jm + H_{\mathrm{mic}}(m), \qquad\text{where } H_{\mathrm{mic}}(m) = \alpha\Delta m + Am + H_{\mathrm{d}}[m]. \tag*{(6)} H MFA ( m ) = J m + H mic ( m ) , where H mic ( m ) = α Δ m + A m + H d [ m ] . (6)
Here, H m i c H_{\mathrm{mic}} H mic is the usual micromagnetic effective field containing the exchange, anisotropy, and magnetostatic contributions [7 , 22 ] .
The mobility matrix M ( m ) ∈ R 3 × 3 \mathbb{M}(m) \in\mathbb{R}^{3\times3} M ( m ) ∈ R 3 × 3 in (1) encodes the nonlinear longitudinal and transverse relaxation coefficients in [15 ] . Let
q ( s ) : = s L − 1 ( s ) , where L ( ξ ) : = coth ξ − 1 ξ , (7) q(s) := \frac{s}{\mathcal{L}^{-1}(s)}, \qquad\text{where } \mathcal{L}(\xi) := \coth\xi- \frac{1}{\xi}, \tag*{(7)} q ( s ) := L − 1 ( s ) s , where L ( ξ ) := coth ξ − ξ 1 , (7)
and L − 1 \mathcal{L}^{-1} L − 1 is the inverse function of L \mathcal{L} L . We define, for m ≠ 0 m \ne0 m = 0 ,
M ( m ) : = ( 1 − q ( ∣ m ∣ ) ) ( I − m ⊗ m ∣ m ∣ 2 ) + 2 q ( ∣ m ∣ ) m ⊗ m ∣ m ∣ 2 . \mathbb{M}(m) := \left(1-q(|m|)\right)\left(I-\frac{m\otimes m}{|m|^{2}}\right)+2q(|m|)\frac{m\otimes m}{|m|^{2}}. M ( m ) := ( 1 − q ( ∣ m ∣ ) ) ( I − ∣ m ∣ 2 m ⊗ m ) + 2 q ( ∣ m ∣ ) ∣ m ∣ 2 m ⊗ m .
Since q ( s ) → 1 / 3 q(s) \to1/3 q ( s ) → 1/3 as s ↓ 0 s \downarrow0 s ↓ 0 , the apparent directional singularity is removable and in fact M ( 0 ) : = 2 3 I \mathbb{M}(0) := \frac{2}{3}I M ( 0 ) := 3 2 I . On the other hand, q ( s ) → 0 q(s) \to0 q ( s ) → 0 as s ↑ 1 s \uparrow1 s ↑ 1 , so the part of the diffusion acting in the direction of m m m degenerates at saturation ∣ m ∣ = 1 |m| = 1 ∣ m ∣ = 1 .
In Garanin’s mean-field description, m m m is the local average of classical unit spins s ∈ S 2 s \in\mathbb{S}^2 s ∈ S 2 . Thus, the natural state space of the LLB model is the closed unit ball ∣ m ∣ ≤ 1 |m| \le1 ∣ m ∣ ≤ 1 , in contrast to the sphere-valued constraint ∣ m ∣ = 1 |m| = 1 ∣ m ∣ = 1 in LLG. Overall, the full nonlinear LLB equation (1.1) is a nonlocal quasilinear parabolic system with anisotropic degeneracy at saturation and a singular inverse-Langevin thermodynamic structure . A more detailed discussion of (1.1) and its mathematical structure will be given in § 2 .
The LLB equation commonly studied in the mathematical literature [17 , 18 , 25 , 26 , 28 ] is obtained only after a further near-equilibrium approximation of (1.1). In regimes where the homogeneous molecular-exchange field dominates the remaining effective field, the magnetisation is expanded about the corresponding equilibrium state; see [15 Sec. IV] . This leads to the reduced equation
∂ t m = γ m × h r e d ( m ) + β ∥ ∣ m ∣ 2 ( m ⋅ h r e d ( m ) ) m − β ⊥ ∣ m ∣ 2 m × ( m × h r e d ( m ) ) . (8) \partial_t m = \gamma m \times h_{\mathrm{red}}(m) + \frac{\beta_{\parallel}}{|m|^2}(m \cdot h_{\mathrm{red}}(m))m - \frac{\beta_{\perp}}{|m|^2}m \times(m \times h_{\mathrm{red}}(m)). \tag*{(8)} ∂ t m = γ m × h red ( m ) + ∣ m ∣ 2 β ∥ ( m ⋅ h red ( m )) m − ∣ m ∣ 2 β ⊥ m × ( m × h red ( m )) . (8)
A more detailed derivation of this is given in § 2.5 . Below the Curie temperature, this formulation is not defined at m = 0 m = 0 m = 0 because of its dependence on m / ∣ m ∣ m/|m| m /∣ m ∣ . The weighted formulation used in [17 ] removes the explicit division by ∣ m ∣ 2 |m|^2 ∣ m ∣ 2 , but degenerates on the zero set and is equivalent to (8) only where m ≠ 0 m \ne0 m = 0 . By contrast, the directional singularity of the full nonlinear model is removable and the equation remains well defined at vanishing magnetisation.
In this work, we study the full nonlinear LLB equation (1.1) before this reduction, which we regard as a more general mean-field starting point whenever the near-equilibrium assumptions underlying (8) are not justified. Its analysis is complicated by three interacting features: the longitudinal mobility degenerates as ∣ m ∣ ↑ 1 |m| \uparrow1 ∣ m ∣ ↑ 1 , the inverse-Langevin thermodynamic field and free-energy potential are singular at saturation, and the equation is coupled to the nonlocal whole-space stray field. In particular, saturated initial data may have infinite thermodynamic energy, so neither uniform parabolicity nor finite initial energy can be assumed.
Our analysis begins with an equivalent closed-ball formulation, where the singular thermodynamic force and the degenerate mobility cancel exactly. A viscous regularisation and Faedo–Galerkin approximation yield weighted second-order and inverse-distance estimates. Combined with positive-time W 2 , r W^{2,r} W 2 , r -regularity and a scalar comparison argument, these give uniform separation from ∣ m ∣ = 1 |m| = 1 ∣ m ∣ = 1 for every positive time. Thus solutions may emanate from saturated, possibly infinite-energy, initial data and enter the finite-energy regime instantaneously. Similar separation mechanisms are central in the analysis of Cahn–Hilliard equations with singular potentials [1 , 9 , 13 ] , where separation removes the singularity and yields higher regularity and long-time analysis. Here, it has the additional role of restoring uniform ellipticity of the principal part. However, the precession term and the nonlocal stray field introduce further difficulties absent from the scalar gradient-flow setting.
Once separation is established, the thermodynamic formulation and exact energy identity become available. We then combine asymptotic compactness with a Łojasiewicz–Simon inequality, in the spirit of [1 , 9 ] , to prove convergence to a single stationary state. The same structure yields a spectral Curie-type threshold and exponential relaxation in the paramagnetic regime. We also prove a direct quantitative low-temperature limit from the full nonlinear LLB equation to the Landau–Lifshitz equation, and develop a strong-solution and nonlinear stability theory for uniformly separated data.
The main results of the paper are as follows:
Global existence for arbitrary m 0 ∈ H 1 ( J ) m_0 \in H^1(\mathcal{J}) m 0 ∈ H 1 ( J ) with ∣ m 0 ∣ ≤ 1 |m_0| \le1 ∣ m 0 ∣ ≤ 1 , including saturated initial data; see Theorem 3.2 ;
Quantitative low-temperature convergence of the full LLB model to the Landau–Lifshitz–Gilbert equation; see Theorem 4.1 ;
Convergence of every global solution to an equilibrium as t → ∞ t \to\infty t → ∞ ; see Theorem 4.4 ;
Precise characterisation of the stationary state via the Curie-type threshold θ c = Λ ∗ / 3 \theta_{\mathrm{c}} = \Lambda_*/3 θ c = Λ ∗ /3 . Above it, 0 0 0 is the unique equilibrium and all solutions converge to it exponentially. Below
the threshold, the energy admits nonzero states of lower energy, and trajectories attaining negative energy converge to a nonzero stationary state; see Theorem 4.5 ;
(5) Local strong well-posedness for uniformly separated initial data m 0 ∈ H s ( D ) m_0 \in H^s(\mathcal{D}) m 0 ∈ H s ( D ) , s ∈ ( 3 / 2 , 2 ] s \in(3/2,2] s ∈ ( 3/2 , 2 ] , and global strong well-posedness for d ≤ 2 d \le2 d ≤ 2 ; see Theorems 5.1 .
(6) Exponential stability of nondegenerate equilibria, which implies global strong well-posedness in d = 3 d=3 d = 3 for sufficiently small initial data; see Theorem 5.2 and Corollary 5.3 .
The full nonlinear mean-field LLB model
In this section, we summarise the derivation of (1) in [15 ] and elucidate its mathematical structures. Let D ⊂ R 3 \mathcal{D} \subset\mathbb{R}^{3} D ⊂ R 3 be a bounded connected domain with sufficiently smooth boundary. We write m : [ 0 , ∞ ) × D → R 3 m\colon[0,\infty) \times\mathcal{D} \to\mathbb{R}^{3} m : [ 0 , ∞ ) × D → R 3 for the dimensionless spin polarisation. In Garanin’s classical-spin description, the microscopic spin s s s has unit length, ∣ s ∣ = 1 \lvert s\rvert=1 ∣ s ∣ = 1 , while m = ⟨ s ⟩ m=\langle s\rangle m = ⟨ s ⟩ denotes averaging with respect to the local distribution of spin orientations. It follows that the physically admissible range is ∣ m ∣ ≤ 1 \lvert m\rvert\le1 ∣ m ∣ ≤ 1 . This differs from the usual LLG convention, in which the magnetisation is constrained to take values in S 2 \mathbb{S}^{2} S 2 .
The nonlinear spin-polarisation equation
Let L : [ 0 , ∞ ) → [ 0 , 1 ) \mathcal{L}\colon[0,\infty) \to[0,1) L : [ 0 , ∞ ) → [ 0 , 1 ) given by
L ( ξ ) : = coth ξ − 1 ξ , L ( 0 ) : = 0 , (9) \mathcal{L}(\xi) := \coth\xi- \frac{1}{\xi}, \qquad\mathcal{L}(0):=0, \tag*{(9)} L ( ξ ) := coth ξ − ξ 1 , L ( 0 ) := 0 , (9)
denote the strictly increasing Langevin function , whose inverse is denoted by L − 1 \mathcal{L}^{-1} L − 1 . We set
q ( s ) : = s L − 1 ( s ) , s ∈ ( 0 , 1 ) . (10) q(s):=\frac{s}{\mathcal{L}^{-1}(s)}, \qquad s\in(0,1). \tag*{(10)} q ( s ) := L − 1 ( s ) s , s ∈ ( 0 , 1 ) . (10)
As argued physically in [15 Sec. II] , Garanin introduces a nonequilibrium reduced field ξ \xi ξ through the relation ∣ m ∣ = L ( ξ ) \lvert m\rvert=\mathcal{L}(\xi) ∣ m ∣ = L ( ξ ) , and derives the nonlinear LLB equation for the spin polarisation m m m . In his notation, using the physical relations Λ N = 2 γ 0 λ θ \Lambda_{\mathrm{N}}=2\gamma_{0}\lambda\theta Λ N = 2 γ 0 λ θ and ξ 0 = H / θ \xi_{0}=H/\theta ξ 0 = H / θ , where θ \theta θ denotes the rescaled temperature, we arrive at [15 Eq. (2.16)] :
∂ t m = γ 0 m × H − 2 γ 0 λ θ m + γ 0 λ [ ( 1 − q ( ∣ m ∣ ) ) ( I − m ⊗ m ∣ m ∣ 2 ) + 2 q ( ∣ m ∣ ) m ⊗ m ∣ m ∣ 2 ] H . \partial_t m=\gamma_0m\times H-2\gamma_0\lambda\theta m+\gamma_0\lambda\left[\left(1-q(\lvert m\rvert)\right)\left(I-\frac{m\otimes m}{\lvert m\rvert^2}\right)+2q(\lvert m\rvert)\frac{m\otimes m}{\lvert m\rvert^2}\right]H. ∂ t m = γ 0 m × H − 2 γ 0 λ θ m + γ 0 λ [ ( 1 − q (∣ m ∣) ) ( I − ∣ m ∣ 2 m ⊗ m ) + 2 q (∣ m ∣) ∣ m ∣ 2 m ⊗ m ] H .
This is the starting point of our analysis.
The form of the equation stated above motivates the definition of the mobility matrix :
M ( v ) : = ( 1 − q ( ∣ v ∣ ) ) ( I − v ⊗ v ∣ v ∣ 2 ) + 2 q ( ∣ v ∣ ) v ⊗ v ∣ v ∣ 2 , v ≠ 0. (11) \mathsf{M}(v):=\left(1-q(\lvert v\rvert)\right)\left(I-\frac{v\otimes v}{\lvert v\rvert^2}\right)+2q(\lvert v\rvert)\frac{v\otimes v}{\lvert v\rvert^2}, \qquad v\ne0. \tag*{(11)} M ( v ) := ( 1 − q (∣ v ∣) ) ( I − ∣ v ∣ 2 v ⊗ v ) + 2 q (∣ v ∣) ∣ v ∣ 2 v ⊗ v , v = 0. (11)
After rescaling time by the positive factor γ 0 λ \gamma_0\lambda γ 0 λ and renaming the resulting precession coefficient, the equation takes the form
∂ t m = γ m × H + M ( m ) H − 2 θ m . (12) \partial_t m=\gamma m\times H+\mathsf{M}(m)H-2\theta m. \tag*{(12)} ∂ t m = γ m × H + M ( m ) H − 2 θ m . (12)
Here, H H H is a prescribed magnetic field acting on a single-spin ensemble.
The continuum molecular mean field
For a ferromagnet, the model replaces the single-spin magnetic field H H H in (12) by the molecular mean field H M F A H_{\mathrm{MFA}} H MFA . Under the mean-field and weak-anisotropy approximations, this field consists of the homogeneous molecular exchange together with the usual exchange and anisotropy contributions; see [15 Sec. IV] . In the present work, we write
H M F A ( m ) : = α Δ m + B m + H d [ m ] , (13) H_{\mathrm{MFA}}(m):=\alpha\Delta m+Bm+H_{\mathrm{d}}[m], \tag*{(13)} H MFA ( m ) := α Δ m + B m + H d [ m ] , (13)
where α > 0 \alpha>0 α > 0 is the exchange coefficient and B = B T ∈ R 3 × 3 B=B^{\mathsf{T}}\in\mathbb{R}^{3\times3} B = B T ∈ R 3 × 3 .
It is useful to regard B = J I + A B=J I+A B = J I + A , where J > 0 J>0 J > 0 represents the homogeneous molecular exchange and A = A T A=A^{\mathsf{T}} A = A T the local anisotropy. Thus,
H M F A ( m ) = J m + H m i c ( m ) , H m i c ( m ) : = α Δ m + A m + H d [ m ] . H_{\mathrm{MFA}}(m)=Jm+H_{\mathrm{mic}}(m), \qquad H_{\mathrm{mic}}(m):=\alpha\Delta m+Am+H_{\mathrm{d}}[m]. H MFA ( m ) = J m + H mic ( m ) , H mic ( m ) := α Δ m + A m + H d [ m ] .
In particular, H M F A H_{\mathrm{MFA}} H MFA should be distinguished from the usual micromagnetic effective field H m i c H_{\mathrm{mic}} H mic . The term J m Jm J m is irrelevant for fixed-length LLG dynamics, since it is parallel to m m m , but it plays an essential role in the longitudinal relaxation of the LLB model.
Closed-ball and thermodynamic formulations
We proceed with a further reformulation of the problem. First, we note the endpoint behaviour of the coefficient q q q given in (10) . The Taylor expansion of the Langevin function at 0 0 0 gives
L − 1 ( s ) = 3 s + 9 5 s 3 + O ( s 5 ) , q ( s ) = 1 3 − 1 5 s 2 + O ( s 4 ) as s → 0 , \mathcal{L}^{-1}(s)=3s+\frac{9}{5}s^{3}+O(s^{5}),\qquad q(s)=\frac{1}{3}-\frac{1}{5}s^{2}+O(s^{4})\qquad\text{as }s\to0, L − 1 ( s ) = 3 s + 5 9 s 3 + O ( s 5 ) , q ( s ) = 3 1 − 5 1 s 2 + O ( s 4 ) as s → 0 ,
where L \mathcal{L} L and q q q were defined in (9) and (10) , respectively. Therefore, we can extend q q q continuously by q ( 0 ) : = 1 3 q(0):=\frac{1}{3} q ( 0 ) := 3 1 . Furthermore, since L − 1 ( s ) → ∞ \mathcal{L}^{-1}(s)\to\infty L − 1 ( s ) → ∞ as s ↑ 1 s\uparrow1 s ↑ 1 , we set q ( 1 ) : = 0 q(1):=0 q ( 1 ) := 0 . It follows from (11) that the apparent singularity at v = 0 v=0 v = 0 is removable, with M ( 0 ) : = 2 3 I \mathbb{M}(0):=\frac{2}{3}\mathbb{I} M ( 0 ) := 3 2 I . Moreover, for ∣ v ∣ = 1 |v|=1 ∣ v ∣ = 1 , we have M ( v ) = I − v ⊗ v \mathbb{M}(v)=\mathbb{I}-v\otimes v M ( v ) = I − v ⊗ v . Hence, M \mathbb{M} M extends continuously to the closed unit ball B 1 ( 0 ) ‾ \overline{B_1(0)} B 1 ( 0 ) .
Consequently, the equation
∂ t m = γ m × H M F A ( m ) + M ( m ) H M F A ( m ) − 2 θ m (14) \partial_t m=\gamma m\times H_{\mathrm{MFA}}(m)+\mathbb{M}(m)H_{\mathrm{MFA}}(m)-2\theta m \tag*{(14)} ∂ t m = γ m × H MFA ( m ) + M ( m ) H MFA ( m ) − 2 θ m (14)
is well-defined for all ∣ m ∣ ≤ 1 |m|\le1 ∣ m ∣ ≤ 1 . We shall refer to (14) as the closed-ball formulation . This is the natural formulation for the existence theory, since the initial datum is only assumed to satisfy ∣ m 0 ∣ ≤ 1 |m_0|\le1 ∣ m 0 ∣ ≤ 1 and may attain ∣ m 0 ∣ = 1 |m_0|=1 ∣ m 0 ∣ = 1 on a set of positive measure.
For ∣ v ∣ < 1 |v|<1 ∣ v ∣ < 1 , define
G ( v ) : = L − 1 ( ∣ v ∣ ) ∣ v ∣ v = v q ( ∣ v ∣ ) , G ( 0 ) : = 0. (15) G(v):=\frac{\mathcal{L}^{-1}(|v|)}{|v|}v=\frac{v}{q(|v|)},\qquad G(0):=0. \tag*{(15)} G ( v ) := ∣ v ∣ L − 1 ( ∣ v ∣ ) v = q ( ∣ v ∣ ) v , G ( 0 ) := 0. (15)
Since G ( v ) G(v) G ( v ) is parallel to v v v , one has v × G ( v ) = 0 v\times G(v)=0 v × G ( v ) = 0 and M ( v ) G ( v ) = 2 v \mathbb{M}(v)G(v)=2v M ( v ) G ( v ) = 2 v . Thus, the closed-ball formulation (14) can be equivalently written, whenever ∣ m ∣ < 1 |m|<1 ∣ m ∣ < 1 , as
∂ t m = γ m × h ( m ) + M ( m ) h ( m ) , (16) \partial_t m=\gamma m\times h(m)+\mathbb{M}(m)h(m), \tag*{(16)} ∂ t m = γ m × h ( m ) + M ( m ) h ( m ) , (16)
h ( m ) : = H M F A ( m ) − θ G ( m ) , (17) h(m):=H_{\mathrm{MFA}}(m)-\theta G(m), \tag*{(17)} h ( m ) := H MFA ( m ) − θ G ( m ) , (17)
where h ( m ) h(m) h ( m ) is called the thermodynamic effective field.
These two formulations play complementary roles. On the open unit ball, they are algebraically equivalent, and this equivalence remains valid at the level of weak solutions. The closed-ball formulation (14) extends continuously to ∣ m ∣ = 1 |m|=1 ∣ m ∣ = 1 and is therefore used to construct solutions from possibly saturated initial data. Once positive-time separation from saturation has been established, the same weak solution satisfies (16) on every interval [ τ , T ] [\tau,T] [ τ , T ] , τ > 0 \tau>0 τ > 0 , where the thermodynamic field is well-defined. This formulation is more natural for the energy and long-time analysis.
Energy structure
Define the inverse-Langevin potential
Φ ( s ) : = ∫ 0 s L − 1 ( σ ) d σ , 0 ≤ s < 1. (18) \Phi(s):=\int_{0}^{s}\mathcal{L}^{-1}(\sigma)\,\mathrm{d}\sigma,\qquad0\le s<1. \tag*{(18)} Φ ( s ) := ∫ 0 s L − 1 ( σ ) d σ , 0 ≤ s < 1. (18)
Then, we have a variational derivative identity ∇ v Φ ( ∣ v ∣ ) = G ( v ) \nabla_v\Phi(|v|)=G(v) ∇ v Φ ( ∣ v ∣ ) = G ( v ) . Note that Φ ( s ) → + ∞ \Phi(s)\to+\infty Φ ( s ) → + ∞ as s ↑ 1 s\uparrow1 s ↑ 1 . Recalling θ \theta θ is the rescaled temperature, we define the corresponding finite-temperature energy
E θ ( m ) : = α 2 ∥ ∇ m ∥ L 2 ( D ) 2 − 1 2 ∫ D m ⋅ B m d x + 1 2 ∥ H d [ m ] ∥ L 2 ( R 3 ) 2 + θ ∫ D Φ ( ∣ m ∣ ) d x . (19) \mathcal{E}_{\theta}(m):=\frac{\alpha}{2}\|\nabla m\|_{L^{2}(\mathcal{D})}^{2}-\frac{1}{2}\int_{\mathcal{D}}m\cdot Bm\,\mathrm{d}x+\frac{1}{2}\|H_{\mathrm{d}}[m]\|_{L^{2}(\mathbb{R}^{3})}^{2}+\theta\int_{\mathcal{D}}\Phi(|m|)\,\mathrm{d}x. \tag*{(19)} E θ ( m ) := 2 α ∥∇ m ∥ L 2 ( D ) 2 − 2 1 ∫ D m ⋅ B m d x + 2 1 ∥ H d [ m ] ∥ L 2 ( R 3 ) 2 + θ ∫ D Φ ( ∣ m ∣ ) d x . (19)
The negative variational derivative of the energy functional gives the thermodynamic effective field:
− δ E θ δ m = α Δ m + B m + H d [ m ] − θ G ( m ) = h ( m ) . -\frac{\delta\mathcal{E}_{\theta}}{\delta m}=\alpha\Delta m+Bm+H_{\mathrm{d}}[m]-\theta G(m)=h(m). − δ m δ E θ = α Δ m + B m + H d [ m ] − θ G ( m ) = h ( m ) .
Therefore, sufficiently regular solutions of (16) formally satisfy
d d t E θ ( m ( t ) ) + ∫ D h ( m ) ⋅ M ( m ) h ( m ) d x = 0 , (20) \frac{\mathrm{d}}{\mathrm{d}t}\mathcal{E}_{\theta}(m(t))+\int_{\mathcal{D}}h(m)\cdot\mathbb{M}(m)h(m)\,\mathrm{d}x=0, \tag*{(20)} d t d E θ ( m ( t )) + ∫ D h ( m ) ⋅ M ( m ) h ( m ) d x = 0 , (20)
where the precessional term makes no contribution since ( m × h ) ⋅ h = 0 (m \times h) \cdot h = 0 ( m × h ) ⋅ h = 0 .
The singularity of Φ \Phi Φ at ∣ m ∣ = 1 |m| = 1 ∣ m ∣ = 1 is important for the analysis. If ∣ m 0 ∣ = 1 |m_0| = 1 ∣ m 0 ∣ = 1 on a set of positive measure, then E θ ( m 0 ) = + ∞ E_\theta(m_0) = +\infty E θ ( m 0 ) = + ∞ , even though the closed-ball formulation (14) remains well-defined. The existence argument will thus be based initially on the closed-ball equation (14) . We shall prove that the solution enters the open unit ball instantaneously, after which the thermodynamic field, the finite-temperature energy, and the dissipation identity above are well-defined for every positive time.
Relation with the standard reduced LLB equations
It is useful to clarify how the full nonlinear mean-field equation considered here is related to the reduced LLB equations more commonly used in the mathematical literature. To this end, we recall that B = J I + A \mathrm{B} = JI + A B = J I + A , where J > 0 J > 0 J > 0 denotes the homogeneous molecular exchange and A A A contains the remaining anisotropy contribution. We also recall the expressions for the micromagnetic effective field H m i c ( m ) H_{\mathrm{mic}}(m) H mic ( m ) and the thermodynamic effective field h ( m ) h(m) h ( m ) from (6) and (17) , respectively.
The reduced LLB equations arise by expanding this thermodynamic field and the mobility about the appropriate homogeneous equilibrium. The discussion below is an asymptotic derivation of the usual reduced models, rather than a rigorous convergence theorem from the full equation. Rigorous limiting results, when available, are stated and proved separately later in § 4 .
Reduction below the bulk mean-field Curie temperature
We first explain how the standard reduced LLB equation below the Curie temperature is recovered from the full nonlinear model. Recall the thermodynamic field in (17) can be written as
h ( m ) = H m i c ( m ) + ( J − θ L − 1 ( ∣ m ∣ ) ∣ m ∣ ) m . (21) h(m) = H_{\mathrm{mic}}(m) + \left(J - \theta\frac{\mathcal{L}^{-1}(|m|)}{|m|}\right)m. \tag*{(21)} h ( m ) = H mic ( m ) + ( J − θ ∣ m ∣ L − 1 ( ∣ m ∣ ) ) m . (21)
The reference magnitude for the below-Curie reduction is determined by the corresponding homogeneous bulk mean-field problem, obtained from (21) by neglecting H m i c H_{\mathrm{mic}} H mic . The homogeneous dynamics then simplifies to
∂ t m = 2 ( J q ( ∣ m ∣ ) − θ ) m . (22) \partial_t m = 2\left(Jq(|m|) - \theta\right)m. \tag*{(22)} ∂ t m = 2 ( J q ( ∣ m ∣ ) − θ ) m . (22)
A nonzero homogeneous equilibrium of magnitude m e ∈ ( 0 , 1 ) m_e \in(0,1) m e ∈ ( 0 , 1 ) then satisfies q ( m e ) = θ / J q(m_e) = \theta/J q ( m e ) = θ / J . Since q q q decreases from q ( 0 ) = 1 / 3 q(0) = 1/3 q ( 0 ) = 1/3 to q ( 1 ) = 0 q(1) = 0 q ( 1 ) = 0 , such an equilibrium exists precisely when θ / J ∈ ( 0 , 1 / 3 ) \theta/J \in(0,1/3) θ / J ∈ ( 0 , 1/3 ) . Thus, θ c b u l k : = J / 3 \theta_{\mathrm{c}}^{\mathrm{bulk}} := J/3 θ c bulk := J /3 is the Curie threshold of the homogeneous bulk mean-field problem. This threshold arises here in the asymptotic reduction to the usual model and should be distinguished from the rigorous spectral threshold Λ ∗ / 3 \Lambda_*/3 Λ ∗ /3 introduced later in (129) for the full micromagnetic PDE. In the homogeneous isotropic setting where Λ ∗ = J \Lambda_* = J Λ ∗ = J , the two thresholds coincide.
We now return to the full thermodynamic field (21) , and consider states for which θ < J / 3 \theta< J/3 θ < J /3 and ∣ m ∣ |m| ∣ m ∣ remains close to m e ∈ ( 0 , 1 ) m_e \in(0,1) m e ∈ ( 0 , 1 ) . Introduce the longitudinal susceptibility
χ ∥ : = [ θ ( L − 1 ) ′ ( m e ) − J ] − 1 > 0. (23) \chi_{\parallel} := \left[\theta(\mathcal{L}^{-1})'(m_e) - J\right]^{-1} > 0. \tag*{(23)} χ ∥ := [ θ ( L − 1 ) ′ ( m e ) − J ] − 1 > 0. (23)
Consider the scalar longitudinal coefficient
f ( r ) : = J − θ L − 1 ( r ) r . f(r) := J - \theta\frac{\mathcal{L}^{-1}(r)}{r}. f ( r ) := J − θ r L − 1 ( r ) .
Since f ( m e ) = 0 f(m_e) = 0 f ( m e ) = 0 , a Taylor expansion about r = m e r = m_e r = m e , together with q ( m e ) = θ / J q(m_e) = \theta/J q ( m e ) = θ / J , gives
f ( r ) = 1 2 χ ∥ ( 1 − r 2 m e 2 ) + O ( ( r − m e ) 2 ) . f(r) = \frac{1}{2\chi_{\parallel}}\left(1 - \frac{r^2}{m_e^2}\right) + O\left((r - m_e)^2\right). f ( r ) = 2 χ ∥ 1 ( 1 − m e 2 r 2 ) + O ( ( r − m e ) 2 ) .
Consequently, for ∣ m ∣ |m| ∣ m ∣ close to m e m_e m e ,
h ( m ) = h r e d ( m ) + O ( ( ∣ m ∣ − m e ) 2 ) m , (24) h(m) = h_{\mathrm{red}}(m) + O\left((|m| - m_e)^2\right)m, \tag*{(24)} h ( m ) = h red ( m ) + O ( ( ∣ m ∣ − m e ) 2 ) m , (24)
where
h r e d ( m ) ≔ H m i c ( m ) + 1 2 χ ∥ ( 1 − ∣ m ∣ 2 m e 2 ) m . (25) h_{\mathrm{red}}(m) \coloneqq H_{\mathrm{mic}}(m) + \frac{1}{2\chi_{\parallel}}\left(1-\frac{|m|^2}{m_{\mathrm{e}}^2}\right)m. \tag*{(25)} h red ( m ) : = H mic ( m ) + 2 χ ∥ 1 ( 1 − m e 2 ∣ m ∣ 2 ) m . (25)
This is the first approximation made in the full nonlinear problem, leading to the usual effective field for this temperature regime considered in the mathematical literature [17 , 26 , 27 , 30 ] .
The second approximation consists in freezing the longitudinal and transverse mobilities at the equilibrium magnitude m e m_{\mathrm{e}} m e . Recall the mobility matrix M ( m ) \mathbf{M}(m) M ( m ) defined in (11) . Since q ( m e ) = θ / J q(m_{\mathrm{e}})=\theta/J q ( m e ) = θ / J , this approximation at m e m_{\mathrm{e}} m e gives
M ( m ) = β ⊥ ( I − m ⊗ m ∣ m ∣ 2 ) + β ∥ m ⊗ m ∣ m ∣ 2 + O ( ∣ m ∣ − m e ) , \mathbf{M}(m)=\beta_{\perp}\left(\mathbf{I}-\frac{m\otimes m}{|m|^2}\right)+\beta_{\parallel}\frac{m\otimes m}{|m|^2}+O(|m|-m_{\mathrm{e}}), M ( m ) = β ⊥ ( I − ∣ m ∣ 2 m ⊗ m ) + β ∥ ∣ m ∣ 2 m ⊗ m + O ( ∣ m ∣ − m e ) ,
where β ⊥ ≔ 1 − θ / J \beta_{\perp}\coloneqq1-\theta/J β ⊥ : = 1 − θ / J and β ∥ ≔ 2 θ / J \beta_{\parallel}\coloneqq2\theta/J β ∥ : = 2 θ / J .
Substituting these two near-equilibrium approximations into the full thermodynamic equation (16) and neglecting the higher-order terms yields
∂ t m = γ m × h r e d ( m ) + β ∥ ∣ m ∣ 2 ( m ⋅ h r e d ( m ) ) m − β ⊥ ∣ m ∣ 2 m × ( m × h r e d ( m ) ) . (26) \partial_t m=\gamma m\times h_{\mathrm{red}}(m)+\frac{\beta_{\parallel}}{|m|^2}(m\cdot h_{\mathrm{red}}(m))m-\frac{\beta_{\perp}}{|m|^2}m\times(m\times h_{\mathrm{red}}(m)). \tag*{(26)} ∂ t m = γ m × h red ( m ) + ∣ m ∣ 2 β ∥ ( m ⋅ h red ( m )) m − ∣ m ∣ 2 β ⊥ m × ( m × h red ( m )) . (26)
This is the standard reduced LLB equation below the bulk mean-field Curie temperature [17 , 26 ] .
2.5.2. Reduction above the bulk mean-field Curie temperature. For θ > J / 3 \theta>J/3 θ > J /3 , it is clear from (22) that the homogeneous equilibrium is m = 0 m=0 m = 0 . We recall the expansions
q ( r ) = 1 3 − 1 5 r 2 + O ( r 4 ) , M ( m ) = 2 3 I + O ( ∣ m ∣ 2 ) , q(r)=\frac{1}{3}-\frac{1}{5}r^2+O(r^4),\qquad\mathbf{M}(m)=\frac{2}{3}\mathbf{I}+O(|m|^2), q ( r ) = 3 1 − 5 1 r 2 + O ( r 4 ) , M ( m ) = 3 2 I + O ( ∣ m ∣ 2 ) ,
together with the exact identity
M ( m ) ( J m − θ G ( m ) ) = ( 2 J q ( ∣ m ∣ ) − 2 θ ) m . \mathbf{M}(m)(Jm-\theta G(m))=(2Jq(|m|)-2\theta)m. M ( m ) ( J m − θ G ( m )) = ( 2 J q ( ∣ m ∣ ) − 2 θ ) m .
Substituting these into the equation (16) , and noting that m × ( J m − θ G ( m ) ) = 0 m\times(Jm-\theta G(m))=0 m × ( J m − θ G ( m )) = 0 , we obtain
∂ t m = γ m × H m i c ( m ) + 2 3 H m i c ( m ) − 2 ( θ − J 3 ) m − 2 J 5 ∣ m ∣ 2 m + R ( m ) , (27) \partial_t m=\gamma m\times H_{\mathrm{mic}}(m)+\frac{2}{3}H_{\mathrm{mic}}(m)-2\left(\theta-\frac{J}{3}\right)m-\frac{2J}{5}|m|^2m+\mathcal{R}(m), \tag*{(27)} ∂ t m = γ m × H mic ( m ) + 3 2 H mic ( m ) − 2 ( θ − 3 J ) m − 5 2 J ∣ m ∣ 2 m + R ( m ) , (27)
where, formally near m = 0 m=0 m = 0 ,
∣ R ( m ) ∣ ≤ C ( ∣ m ∣ 2 ∣ H m i c ( m ) ∣ + ∣ m ∣ 5 ) . |\mathcal{R}(m)|\leq C\left(|m|^2|H_{\mathrm{mic}}(m)|+|m|^5\right). ∣ R ( m ) ∣ ≤ C ( ∣ m ∣ 2 ∣ H mic ( m ) ∣ + ∣ m ∣ 5 ) .
If anisotropy and stray-field contributions are omitted for simplicity, then (27) reduces, to leading order, to
∂ t m = γ ~ m × Δ m + κ 1 Δ m − κ 2 ( 1 + μ C ∣ m ∣ 2 ) m , (28) \partial_t m=\widetilde{\gamma}m\times\Delta m+\kappa_1\Delta m-\kappa_2(1+\mu_C|m|^2)m, \tag*{(28)} ∂ t m = γ m × Δ m + κ 1 Δ m − κ 2 ( 1 + μ C ∣ m ∣ 2 ) m , (28)
with
γ ~ = α γ , κ 1 = 2 α 3 , κ 2 = 2 ( θ − J 3 ) , μ C = J 5 ( θ − J / 3 ) . \widetilde{\gamma}=\alpha\gamma,\qquad\kappa_1=\frac{2\alpha}{3},\qquad\kappa_2=2\left(\theta-\frac{J}{3}\right),\qquad\mu_C=\frac{J}{5(\theta-J/3)}. γ = α γ , κ 1 = 3 2 α , κ 2 = 2 ( θ − 3 J ) , μ C = 5 ( θ − J /3 ) J .
Up to the normalisation of the material coefficients, (28) is the simplified LLB equation commonly analysed in the regime above the Curie temperature [6 , 18 , 19 , 25 , 28 ] .
Remark 2.1 (Low-temperature consistency). Consider the reduced below-Curie equation (26) . This equation was itself obtained by a near-equilibrium approximation of the full nonlinear model, so the following low-temperature argument is only formal. Since q ( m e ) = θ / J q(m_{\mathrm{e}})=\theta/J q ( m e ) = θ / J and q q q is strictly decreasing with q ( 1 ) = 0 q(1)=0 q ( 1 ) = 0 , we have m e → 1 m_{\mathrm{e}}\to1 m e → 1 as θ / J → 0 \theta/J\to0 θ / J → 0 . Moreover,
β ∥ = 2 θ J → 0 , β ⊥ = 1 − θ J → 1. \beta_{\parallel}=\frac{2\theta}{J}\to0,\qquad\beta_{\perp}=1-\frac{\theta}{J}\to1. β ∥ = J 2 θ → 0 , β ⊥ = 1 − J θ → 1.
Noting (25) , the longitudinal part of (26) is
β ∥ ∣ m ∣ 2 ( m ⋅ H m i c ( m ) ) m + β ∥ 2 χ ∥ ( 1 − ∣ m ∣ 2 m e 2 ) m . \frac{\beta_{\parallel}}{|m|^2}(m\cdot H_{\mathrm{mic}}(m))m+\frac{\beta_{\parallel}}{2\chi_{\parallel}}\left(1-\frac{|m|^2}{m_{\mathrm{e}}^2}\right)m. ∣ m ∣ 2 β ∥ ( m ⋅ H mic ( m )) m + 2 χ ∥ β ∥ ( 1 − m e 2 ∣ m ∣ 2 ) m .
Using L − 1 ( s ) = s / q ( s ) \mathcal{L}^{-1}(s)=s/q(s) L − 1 ( s ) = s / q ( s ) and q ( m e ) = θ / J q(m_e)=\theta/J q ( m e ) = θ / J , while noting the definition of χ ∥ \chi_{\parallel} χ ∥ in (23) , we find
β ∥ 2 χ ∥ = − J m e q ′ ( m e ) → J . \frac{\beta_{\parallel}}{2\chi_{\parallel}}=-Jm_e q'(m_e)\to J. 2 χ ∥ β ∥ = − J m e q ′ ( m e ) → J .
Hence, formally for near-equilibrium families satisfying ∣ m ∣ − m e → 0 |m|-m_e\to0 ∣ m ∣ − m e → 0 , both longitudinal terms vanish as θ → 0 \theta\to0 θ → 0 . Since the radial part of h r e d ( m ) h_{\mathrm{red}}(m) h red ( m ) does not contribute to the cross-product terms, the remaining equation is
∂ t m = γ m × H m i c ( m ) − m × ( m × H m i c ( m ) ) , ∣ m ∣ = 1 , \partial_t m=\gamma m\times H_{\mathrm{mic}}(m)-m\times\left(m\times H_{\mathrm{mic}}(m)\right),\qquad|m|=1, ∂ t m = γ m × H mic ( m ) − m × ( m × H mic ( m ) ) , ∣ m ∣ = 1 ,
which is the Landau–Lifshitz–Gilbert (LLG) equation. Thus, the standard reduced model is formally consistent with the low-temperature Landau–Lifshitz limit. A direct rigorous passage from the full nonlinear LLB equation to the LLG equation is proved independently in § 4.1 .
Global existence and instantaneous separation
In this section, we establish the existence of a global weak solution to (16) –(17) . Since G ( m ) G(m) G ( m ) is not defined at ∣ m ∣ = 1 |m|=1 ∣ m ∣ = 1 , the construction is first carried out using the equivalent closed-ball formulation (14) . We then prove that the solution is strictly separated from the saturation boundary ∣ m ∣ = 1 |m|=1 ∣ m ∣ = 1 for every positive time. This recovers (16) and its energy structure on every interval bounded away from t = 0 t=0 t = 0 .
Weak solutions and the main theorem
3.1. Weak solutions and the main theorem. The notion of weak solution employed in this paper is as follows.
Definition 3.1. Let m 0 ∈ H 1 ( D ) m_0\in H^1(\mathcal{D}) m 0 ∈ H 1 ( D ) such that ∣ m 0 ∣ ≤ 1 |m_0|\le1 ∣ m 0 ∣ ≤ 1 a.e. in D \mathcal{D} D , and let T > 0 T>0 T > 0 . A function m m m is called a global weak solution if
(i) m ∈ H 1 ( 0 , T ; L 2 ( D ) ) ∩ L ∞ ( 0 , T ; H 1 ( D ) ) m\in H^1(0,T;L^2(\mathcal{D}))\cap L^\infty(0,T;H^1(\mathcal{D})) m ∈ H 1 ( 0 , T ; L 2 ( D )) ∩ L ∞ ( 0 , T ; H 1 ( D )) ;
(ii) ∣ m ∣ ≤ 1 |m|\le1 ∣ m ∣ ≤ 1 a.e. in ( 0 , T ) × D (0,T)\times\mathcal{D} ( 0 , T ) × D ;
(iii) m ( 0 ) = m 0 m(0)=m_0 m ( 0 ) = m 0 in L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) ;
(iv) for every τ ∈ ( 0 , T ) \tau\in(0,T) τ ∈ ( 0 , T ) , m ∈ L 2 ( τ , T ; H 2 ( D ) ) m\in L^2(\tau,T;H^2(\mathcal{D})) m ∈ L 2 ( τ , T ; H 2 ( D )) and h ( m ) ∈ L 2 ( τ , T ; L 2 ( D ) ) h(m)\in L^2(\tau,T;L^2(\mathcal{D})) h ( m ) ∈ L 2 ( τ , T ; L 2 ( D )) ; and
(v) the following equations hold for almost every t ∈ ( τ , T ) t\in(\tau,T) t ∈ ( τ , T ) :
⟨ ∂ t m ( t ) , φ ⟩ = γ ⟨ m ( t ) × h ( m ( t ) ) , φ ⟩ + ⟨ M ( m ( t ) ) h ( m ( t ) ) , φ ⟩ , (29) \langle\partial_t m(t),\varphi\rangle=\gamma\langle m(t)\times h(m(t)),\varphi\rangle+\langle\mathcal{M}(m(t))h(m(t)),\varphi\rangle, \tag*{(29)} ⟨ ∂ t m ( t ) , φ ⟩ = γ ⟨ m ( t ) × h ( m ( t )) , φ ⟩ + ⟨ M ( m ( t )) h ( m ( t )) , φ ⟩ , (29) ⟨ h ( m ( t ) ) , ψ ⟩ = − α ⟨ ∇ m ( t ) , ∇ ψ ⟩ + ⟨ B m ( t ) + H d [ m ( t ) ] , ψ ⟩ − θ ⟨ G ( m ( t ) ) , ψ ⟩ , (30) \langle h(m(t)),\psi\rangle=-\alpha\langle\nabla m(t),\nabla\psi\rangle+\langle Bm(t)+H_{\mathrm{d}}[m(t)],\psi\rangle-\theta\langle G(m(t)),\psi\rangle, \tag*{(30)} ⟨ h ( m ( t )) , ψ ⟩ = − α ⟨ ∇ m ( t ) , ∇ ψ ⟩ + ⟨ B m ( t ) + H d [ m ( t )] , ψ ⟩ − θ ⟨ G ( m ( t )) , ψ ⟩ , (30) for every φ ∈ L 2 ( D ) \varphi\in L^2(\mathcal{D}) φ ∈ L 2 ( D ) and ψ ∈ H 1 ( D ) \psi\in H^1(\mathcal{D}) ψ ∈ H 1 ( D ) .
We have the following main theorem of this section.
Theorem 3.2. Let m 0 ∈ H 1 ( D ) m_0\in H^1(\mathcal{D}) m 0 ∈ H 1 ( D ) such that ∣ m 0 ∣ ≤ 1 |m_0|\le1 ∣ m 0 ∣ ≤ 1 a.e. in D \mathcal{D} D . Then there exists a global weak solution in the sense of Definition 3.1. This solution enjoys the following additional properties:
(i) for every p ≥ 1 p\ge1 p ≥ 1 , the solution satisfies
∫ D 1 ( 1 − ∣ m ( t , x ) ∣ 2 ) p d x ≤ C p ( 1 + t − p ) for every t > 0 ; (31) \int_{\mathcal{D}}\frac{1}{(1-|m(t,x)|^2)^p}\,\mathrm{d}x\le C_p(1+t^{-p})\qquad\text{for every }t>0; \tag*{(31)} ∫ D ( 1 − ∣ m ( t , x ) ∣ 2 ) p 1 d x ≤ C p ( 1 + t − p ) for every t > 0 ; (31) (ii) for every τ > 0 \tau>0 τ > 0 , there exists ρ τ ∈ ( 0 , 1 ) \rho_\tau\in(0,1) ρ τ ∈ ( 0 , 1 ) such that
ess sup x ∈ D ∣ m ( t , x ) ∣ ≤ ρ τ for every t ∈ [ τ , ∞ ) ; (32) \operatorname*{ess\,sup}_{x\in\mathcal{D}}|m(t,x)|\le\rho_\tau\qquad\text{for every }t\in[\tau,\infty); \tag*{(32)} x ∈ D ess sup ∣ m ( t , x ) ∣ ≤ ρ τ for every t ∈ [ τ , ∞ ) ; (32) (iii) for every T > 0 T>0 T > 0 , m ∈ C ( [ 0 , T ] ; H 1 ( D ) ) m\in C([0,T];H^1(\mathcal{D})) m ∈ C ([ 0 , T ] ; H 1 ( D )) . In particular,
m ( t ) → m 0 strongly in H 1 ( D ) as t ↓ 0 ; (33) m(t)\to m_0\qquad\text{strongly in }H^1(\mathcal{D})\text{ as }t\downarrow0; \tag*{(33)} m ( t ) → m 0 strongly in H 1 ( D ) as t ↓ 0 ; (33)
(iv) the energy identity
E θ ( m ( t ) ) + ∫ s t ∫ D h ( m ) ⋅ M ( m ) h ( m ) d x d r = E θ ( m ( s ) ) (34) \mathcal{E}_{\theta}(m(t))+\int_s^t\int_{\mathcal{D}}h(m)\cdot\mathcal{M}(m)h(m)\,\mathrm{d}x\,\mathrm{d}r=\mathcal{E}_{\theta}(m(s)) \tag*{(34)} E θ ( m ( t )) + ∫ s t ∫ D h ( m ) ⋅ M ( m ) h ( m ) d x d r = E θ ( m ( s )) (34)
holds for every 0 < s < t 0<s<t 0 < s < t . Furthermore, if ∣ { x ∈ D : ∣ m 0 ( x ) ∣ = 1 } ∣ > 0 \lvert\{x\in\mathcal{D}: \lvert m_0(x)\rvert=1\}\rvert>0 ∣{ x ∈ D : ∣ m 0 ( x )∣ = 1 }∣ > 0 , then
E θ ( m ( t ) ) → + ∞ as t ↓ 0. (35) \mathcal{E}_{\theta}(m(t))\to+\infty\qquad\text{as }t\downarrow0. \tag*{(35)} E θ ( m ( t )) → + ∞ as t ↓ 0. (35)
Remark 3.3. The solution is global and weak in the analytical sense, but no uniqueness is asserted. Nevertheless, on every interval [ τ , T ] [\tau,T] [ τ , T ] , τ > 0 \tau>0 τ > 0 , it has the strong parabolic regularity
m ∈ H 1 ( τ , T ; L 2 ( D ) ) ∩ L 2 ( τ , T ; H 2 ( D ) ) . m\in H^1(\tau,T;L^2(\mathcal{D}))\cap L^2(\tau,T;H^2(\mathcal{D})). m ∈ H 1 ( τ , T ; L 2 ( D )) ∩ L 2 ( τ , T ; H 2 ( D )) .
Viscous Faedo–Galerkin approximation
This section introduces a viscous regularisation of the problem and its Faedo–Galerkin approximation. Let
π ( v ) : = { v , ∣ v ∣ ≤ 1 , v ∣ v ∣ , ∣ v ∣ > 1 , (36) \pi(v):=
\begin{cases}
v, & |v|\leq1,\\
\dfrac{v}{|v|}, & |v|>1,
\end{cases} \tag*{(36)} π ( v ) := ⎩ ⎨ ⎧ v , ∣ v ∣ v , ∣ v ∣ ≤ 1 , ∣ v ∣ > 1 , (36)
be the metric projection onto B 1 ( 0 ) ‾ \overline{B_1(0)} B 1 ( 0 ) , and set
M ^ ( v ) : = M ( π ( v ) ) . \widehat{\mathcal{M}}(v):=\mathcal{M}(\pi(v)). M ( v ) := M ( π ( v )) .
Both maps π \pi π and M ^ \widehat{\mathcal{M}} M are globally Lipschitz. Notice also that
M ^ ( v ) v = { 2 q ( ∣ v ∣ ) v , ∣ v ∣ ≤ 1 , 0 , ∣ v ∣ > 1. (37) \widehat{\mathcal{M}}(v)v=
\begin{cases}
2q(|v|)v, & |v|\leq1,\\
0, & |v|>1.
\end{cases} \tag*{(37)} M ( v ) v = { 2 q ( ∣ v ∣ ) v , 0 , ∣ v ∣ ≤ 1 , ∣ v ∣ > 1. (37)
For ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , we consider a regularisation of (14) , namely
∂ t m ε = γ π ( m ε ) × H M F A ( m ε ) + M ^ ( m ε ) H M F A ( m ε ) − 2 θ m ε + ε Δ m ε , (38) \partial_t m_\varepsilon=\gamma\pi(m_\varepsilon)\times H_{\mathrm{MFA}}(m_\varepsilon)+\widehat{\mathcal{M}}(m_\varepsilon)H_{\mathrm{MFA}}(m_\varepsilon)-2\theta m_\varepsilon+\varepsilon\Delta m_\varepsilon, \tag*{(38)} ∂ t m ε = γ π ( m ε ) × H MFA ( m ε ) + M ( m ε ) H MFA ( m ε ) − 2 θ m ε + ε Δ m ε , (38)
with ∂ ν m ε = 0 \partial_\nu m_\varepsilon=0 ∂ ν m ε = 0 on ∂ D \partial\mathcal{D} ∂ D and m ε ( 0 ) = m 0 m_\varepsilon(0)=m_0 m ε ( 0 ) = m 0 .
Let { e j } j ≥ 1 \{e_j\}_{j\geq1} { e j } j ≥ 1 be an L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) -orthonormal basis of Neumann Laplacian eigenfunctions:
− Δ e j = λ j e j in D , ∂ ν e j = 0 on ∂ D , -\Delta e_j=\lambda_j e_j\quad\text{in }\mathcal{D},\qquad\partial_\nu e_j=0\quad\text{on }\partial\mathcal{D}, − Δ e j = λ j e j in D , ∂ ν e j = 0 on ∂ D ,
and set V n : = span { e 1 , … , e n } V_n:=\operatorname{span}\{e_1,\ldots,e_n\} V n := span { e 1 , … , e n } . and let P n P_n P n denote the L 2 L^2 L 2 -orthogonal projection onto V n V_n V n . We seek m ε , n : [ 0 , ∞ ) → V n m_{\varepsilon,n}:[0,\infty)\to V_n m ε , n : [ 0 , ∞ ) → V n such that
⟨ ∂ t m ε , n , φ n ⟩ = γ ⟨ π ( m ε , n ) × H M F A ( m ε , n ) , φ n ⟩ + ⟨ M ^ ( m ε , n ) H M F A ( m ε , n ) , φ n ⟩ − 2 θ ⟨ m ε , n , φ n ⟩ − ε ⟨ ∇ m ε , n , ∇ φ n ⟩ (39) \begin{aligned}
\langle\partial_t m_{\varepsilon,n},\varphi_n\rangle
={}&\gamma\langle\pi(m_{\varepsilon,n})\times H_{\mathrm{MFA}}(m_{\varepsilon,n}),\varphi_n\rangle\\
&+\langle\widehat{\mathcal{M}}(m_{\varepsilon,n})H_{\mathrm{MFA}}(m_{\varepsilon,n}),\varphi_n\rangle
-2\theta\langle m_{\varepsilon,n},\varphi_n\rangle
-\varepsilon\langle\nabla m_{\varepsilon,n},\nabla\varphi_n\rangle \tag*{(39)}
\end{aligned} ⟨ ∂ t m ε , n , φ n ⟩ = γ ⟨ π ( m ε , n ) × H MFA ( m ε , n ) , φ n ⟩ + ⟨ M ( m ε , n ) H MFA ( m ε , n ) , φ n ⟩ − 2 θ ⟨ m ε , n , φ n ⟩ − ε ⟨ ∇ m ε , n , ∇ φ n ⟩ (39)
for all φ n ∈ V n \varphi_n\in V_n φ n ∈ V n , with m ε , n ( 0 ) = P n m 0 m_{\varepsilon,n}(0)=P_nm_0 m ε , n ( 0 ) = P n m 0 . Equivalently, we solve
∂ t m ε , n = γ P n ( π ( m ε , n ) × H M F A ( m ε , n ) ) + P n ( M ^ ( m ε , n ) H M F A ( m ε , n ) ) − 2 θ m ε , n + ε Δ m ε , n . (40) \partial_t m_{\varepsilon,n}=\gamma P_n\bigl(\pi(m_{\varepsilon,n})\times H_{\mathrm{MFA}}(m_{\varepsilon,n})\bigr)+P_n\bigl(\widehat{\mathcal{M}}(m_{\varepsilon,n})H_{\mathrm{MFA}}(m_{\varepsilon,n})\bigr)-2\theta m_{\varepsilon,n}+\varepsilon\Delta m_{\varepsilon,n}. \tag*{(40)} ∂ t m ε , n = γ P n ( π ( m ε , n ) × H MFA ( m ε , n ) ) + P n ( M ( m ε , n ) H MFA ( m ε , n ) ) − 2 θ m ε , n + ε Δ m ε , n . (40)
Here, H M F A H_{\mathrm{MFA}} H MFA is given by (6) .
We begin deriving the necessary uniform estimates in the following proposition.
Proposition 3.4. For every ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] and every n ∈ N n\in\mathbb{N} n ∈ N , problem (39) has a unique global solution. For every T > 0 T>0 T > 0 ,
sup 0 ≤ t ≤ T ∥ m ε , n ( t ) ∥ H 1 ( D ) 2 + ∫ 0 T ∫ D Δ m ε , n ⋅ M ^ ( m ε , n ) Δ m ε , n d x d t + ε ∫ 0 T ∥ Δ m ε , n ∥ L 2 ( D ) 2 d t + ∫ 0 T ∥ ∂ t m ε , n ∥ L 2 ( D ) 2 d t ≤ C T , (41) \sup_{0\leq t\leq T}\|m_{\varepsilon,n}(t)\|_{H^1(\mathcal{D})}^2+\int_0^T\int_{\mathcal{D}}\Delta m_{\varepsilon,n}\cdot\widehat{\mathcal{M}}(m_{\varepsilon,n})\Delta m_{\varepsilon,n}\,\mathrm{d}x\,\mathrm{d}t+\varepsilon\int_0^T\|\Delta m_{\varepsilon,n}\|_{L^2(\mathcal{D})}^2\,\mathrm{d}t+\int_0^T\|\partial_t m_{\varepsilon,n}\|_{L^2(\mathcal{D})}^2\,\mathrm{d}t\leq C_T, \tag*{(41)} 0 ≤ t ≤ T sup ∥ m ε , n ( t ) ∥ H 1 ( D ) 2 + ∫ 0 T ∫ D Δ m ε , n ⋅ M ( m ε , n ) Δ m ε , n d x d t + ε ∫ 0 T ∥Δ m ε , n ∥ L 2 ( D ) 2 d t + ∫ 0 T ∥ ∂ t m ε , n ∥ L 2 ( D ) 2 d t ≤ C T , (41) where C T C_T C T is independent of n n n and ε \varepsilon ε .
Proof. Since all coefficients in the finite-dimensional system are locally Lipschitz, the Cauchy–Lipschitz theorem gives a unique local solution. Taking φ n = m ε , n \varphi_n=m_{\varepsilon,n} φ n = m ε , n in (39) , noting that π ( v ) \pi(v) π ( v ) is parallel to v v v , gives
1 2 d d t ∥ m ε , n ∥ L 2 2 + 2 θ ∥ m ε , n ∥ L 2 2 + ε ∥ ∇ m ε , n ∥ L 2 2 = ⟨ H M F A ( m ε , n ) , M ^ ( m ε , n ) m ε , n ⟩ . (42) \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\lVert m_{\varepsilon,n}\rVert_{L^2}^2+2\theta\lVert m_{\varepsilon,n}\rVert_{L^2}^2+\varepsilon\lVert\nabla m_{\varepsilon,n}\rVert_{L^2}^2=\left\langle H_{\mathrm{MFA}}(m_{\varepsilon,n}),\widehat{\mathrm{M}}(m_{\varepsilon,n})m_{\varepsilon,n}\right\rangle. \tag*{(42)} 2 1 d t d ∥ m ε , n ∥ L 2 2 + 2 θ ∥ m ε , n ∥ L 2 2 + ε ∥ ∇ m ε , n ∥ L 2 2 = ⟨ H MFA ( m ε , n ) , M ( m ε , n ) m ε , n ⟩ . (42)
The map v ↦ M ^ ( v ) v v\mapsto\widehat{\mathrm{M}}(v)v v ↦ M ( v ) v is globally Lipschitz and vanishes for ∣ v ∣ > 1 \lvert v\rvert>1 ∣ v ∣ > 1 by (37) . By the Sobolev chain rule, ∣ ∇ M ^ ( m ε , n ) m ε , n ∣ ≤ C ∣ ∇ m ε , n ∣ \lvert\nabla\widehat{\mathrm{M}}(m_{\varepsilon,n})m_{\varepsilon,n}\rvert\le C\lvert\nabla m_{\varepsilon,n}\rvert ∣ ∇ M ( m ε , n ) m ε , n ∣ ≤ C ∣ ∇ m ε , n ∣ a.e. in D \mathcal{D} D . Hence, after integrating by parts,
∣ ( Δ m ε , n , M ^ ( m ε , n ) m ε , n ) D ∣ ≤ C ∥ ∇ m ε , n ∥ L 2 2 . \left\lvert\left(\Delta m_{\varepsilon,n},\widehat{\mathrm{M}}(m_{\varepsilon,n})m_{\varepsilon,n}\right)_{\mathcal{D}}\right\rvert\le C\lVert\nabla m_{\varepsilon,n}\rVert_{L^2}^2. ( Δ m ε , n , M ( m ε , n ) m ε , n ) D ≤ C ∥ ∇ m ε , n ∥ L 2 2 .
Furthermore, by (A.8),
∥ B m ε , n + H d [ m ε , n ] ∥ L 2 ≤ C ∥ m ε , n ∥ L 2 , \lVert Bm_{\varepsilon,n}+H_{\mathrm{d}}[m_{\varepsilon,n}]\rVert_{L^2}\le C\lVert m_{\varepsilon,n}\rVert_{L^2}, ∥ B m ε , n + H d [ m ε , n ] ∥ L 2 ≤ C ∥ m ε , n ∥ L 2 ,
Therefore, continuing from (42) , we obtain
d d t ∥ m ε , n ∥ L 2 2 + 4 θ ∥ m ε , n ∥ L 2 2 + 2 ε ∥ ∇ m ε , n ∥ L 2 2 ≤ C ∥ m ε , n ∥ H 1 2 . (43) \frac{\mathrm{d}}{\mathrm{d}t}\lVert m_{\varepsilon,n}\rVert_{L^2}^2+4\theta\lVert m_{\varepsilon,n}\rVert_{L^2}^2+2\varepsilon\lVert\nabla m_{\varepsilon,n}\rVert_{L^2}^2\le C\lVert m_{\varepsilon,n}\rVert_{H^1}^2. \tag*{(43)} d t d ∥ m ε , n ∥ L 2 2 + 4 θ ∥ m ε , n ∥ L 2 2 + 2 ε ∥ ∇ m ε , n ∥ L 2 2 ≤ C ∥ m ε , n ∥ H 1 2 . (43)
We next test (39) with − Δ m ε , n -\Delta m_{\varepsilon,n} − Δ m ε , n . In this case, the highest-order precessional term vanishes. Using (A.4) and (A.5), we obtain
1 2 d d t ∥ ∇ m ε , n ∥ L 2 2 + α 2 ∫ D Δ m ε , n ⋅ M ^ ( m ε , n ) Δ m ε , n d x + 2 θ ∥ ∇ m ε , n ∥ L 2 2 + ε ∥ Δ m ε , n ∥ L 2 2 ≤ C ∥ B m ε , n + H d [ m ε , n ] ∥ L 2 2 ≤ C ∥ m ε , n ∥ L 2 2 . (44) \begin{aligned}
\frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\lVert\nabla m_{\varepsilon,n}\rVert_{L^2}^2+\frac{\alpha}{2}\int_{\mathcal{D}}\Delta m_{\varepsilon,n}\cdot\widehat{\mathrm{M}}(m_{\varepsilon,n})\Delta m_{\varepsilon,n}\,\mathrm{d}x+2\theta\lVert\nabla m_{\varepsilon,n}\rVert_{L^2}^2+\varepsilon\lVert\Delta m_{\varepsilon,n}\rVert_{L^2}^2
\le C\lVert Bm_{\varepsilon,n}+H_{\mathrm{d}}[m_{\varepsilon,n}]\rVert_{L^2}^2 \\
&\le C\lVert m_{\varepsilon,n}\rVert_{L^2}^2. \tag*{(44)}
\end{aligned} 2 1 d t d ∥ ∇ m ε , n ∥ L 2 2 + 2 α ∫ D Δ m ε , n ⋅ M ( m ε , n ) Δ m ε , n d x + 2 θ ∥ ∇ m ε , n ∥ L 2 2 + ε ∥ Δ m ε , n ∥ L 2 2 ≤ C ∥ B m ε , n + H d [ m ε , n ] ∥ L 2 2 ≤ C ∥ m ε , n ∥ L 2 2 . (44)
Combining (43) and (44) , we obtain a bound for the first three terms in (41) by the Gronwall lemma. In particular, the finite-dimensional solution cannot blow up and is therefore global. Finally, the Galerkin-projected equation (40) , together with (A.4) and (A.5), gives
∥ ∂ t m ε , n ∥ L 2 2 ≤ C ∫ D Δ m ε , n ⋅ M ^ ( m ε , n ) Δ m ε , n d x + C ∥ m ε , n ∥ L 2 2 + C ε 2 ∥ Δ m ε , n ∥ L 2 2 . (45) \lVert\partial_t m_{\varepsilon,n}\rVert_{L^2}^2\le C\int_{\mathcal{D}}\Delta m_{\varepsilon,n}\cdot\widehat{\mathrm{M}}(m_{\varepsilon,n})\Delta m_{\varepsilon,n}\,\mathrm{d}x+C\lVert m_{\varepsilon,n}\rVert_{L^2}^2+C\varepsilon^2\lVert\Delta m_{\varepsilon,n}\rVert_{L^2}^2. \tag*{(45)} ∥ ∂ t m ε , n ∥ L 2 2 ≤ C ∫ D Δ m ε , n ⋅ M ( m ε , n ) Δ m ε , n d x + C ∥ m ε , n ∥ L 2 2 + C ε 2 ∥ Δ m ε , n ∥ L 2 2 . (45)
Integration in time and using the bound for the first three terms in (41) derived earlier then completes the proof. □
For fixed ε > 0 \varepsilon>0 ε > 0 , the preceding estimates imply, after extraction of a subsequence,
m ε , n ⇀ ∗ m ε in L ∞ ( 0 , T ; H 1 ( D ) ) , m_{\varepsilon,n}\mathrel{\overset{*}{\rightharpoonup}}m_\varepsilon\quad\text{in }L^\infty(0,T;H^1(\mathcal{D})), m ε , n ⇀ ∗ m ε in L ∞ ( 0 , T ; H 1 ( D )) ,
m ε , n ⇀ m ε in L 2 ( 0 , T ; H 2 ( D ) ) , m_{\varepsilon,n}\rightharpoonup m_\varepsilon\quad\text{in }L^2(0,T;H^2(\mathcal{D})), m ε , n ⇀ m ε in L 2 ( 0 , T ; H 2 ( D )) ,
∂ t m ε , n ⇀ ∂ t m ε in L 2 ( 0 , T ; L 2 ( D ) ) . \partial_t m_{\varepsilon,n}\rightharpoonup\partial_t m_\varepsilon\quad\text{in }L^2(0,T;L^2(\mathcal{D})). ∂ t m ε , n ⇀ ∂ t m ε in L 2 ( 0 , T ; L 2 ( D )) .
Aubin–Lions compactness theorem gives
m ε , n → m ε strongly in L 2 ( 0 , T ; H 1 ( D ) ) ∩ C ( [ 0 , T ] ; L 2 ( D ) ) . (46) m_{\varepsilon,n}\to m_\varepsilon\quad\text{strongly in }L^2(0,T;H^1(\mathcal{D}))\cap C([0,T];L^2(\mathcal{D})). \tag*{(46)} m ε , n → m ε strongly in L 2 ( 0 , T ; H 1 ( D )) ∩ C ([ 0 , T ] ; L 2 ( D )) . (46)
The boundedness and continuity of π \pi π and M ^ \widehat{\mathrm{M}} M , together with the linear continuity of the stray field, permit passage to the limit of (39) by a standard argument. Hence, for fixed ε > 0 \varepsilon>0 ε > 0 ,
m ε ∈ H 1 ( 0 , T ; L 2 ( D ) ) ∩ L ∞ ( 0 , T ; H 1 ( D ) ) ∩ L 2 ( 0 , T ; H 2 ( D ) ) (47) m_\varepsilon\in H^1(0,T;L^2(\mathcal{D}))\cap L^\infty(0,T;H^1(\mathcal{D}))\cap L^2(0,T;H^2(\mathcal{D})) \tag*{(47)} m ε ∈ H 1 ( 0 , T ; L 2 ( D )) ∩ L ∞ ( 0 , T ; H 1 ( D )) ∩ L 2 ( 0 , T ; H 2 ( D )) (47)
is a strong solution of (38) , or equivalently (38) holds in L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) for a.e. t > 0 t>0 t > 0 .
We now derive further estimates which are uniform in ε \varepsilon ε . The following lemma shows invariance of the physical state space.
Lemma 3.5. For every ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] ,
∣ m ε ( t , x ) ∣ ≤ 1 for a.e. x ∈ D and every t ≥ 0. (48) |m_\varepsilon(t,x)| \le1 \qquad\text{for a.e. }x \in\mathcal{D}\text{ and every }t \ge0. \tag*{(48)} ∣ m ε ( t , x ) ∣ ≤ 1 for a.e. x ∈ D and every t ≥ 0. (48)
Proof. Define
q ^ ( s ) : = { q ( s ) , 0 ≤ s ≤ 1 , 0 , s > 1. \widehat{q}(s) :=
\begin{cases}
q(s), & 0 \le s \le1,\\
0, & s > 1.
\end{cases} q ( s ) := { q ( s ) , 0 , 0 ≤ s ≤ 1 , s > 1.
Using (37) and taking the scalar product of (38) with 2 m ε 2m_\varepsilon 2 m ε , we obtain
∂ t ∣ m ε ∣ 2 = ( 2 α q ^ ( ∣ m ε ∣ ) + ε ) Δ ∣ m ε ∣ 2 − 2 ( 2 α q ^ ( ∣ m ε ∣ ) + ε ) ∣ ∇ m ε ∣ 2 + 4 q ^ ( ∣ m ε ∣ ) m ε ⋅ ( B m ε + H d [ m ε ] ) − 4 θ ∣ m ε ∣ 2 . (49) \partial_t|m_\varepsilon|^2 = (2\alpha\widehat{q}(|m_\varepsilon|)+\varepsilon)\Delta|m_\varepsilon|^2 - 2(2\alpha\widehat{q}(|m_\varepsilon|)+\varepsilon)|\nabla m_\varepsilon|^2 + 4\widehat{q}(|m_\varepsilon|)m_\varepsilon\cdot(Bm_\varepsilon+ H_d[m_\varepsilon]) - 4\theta|m_\varepsilon|^2. \tag*{(49)} ∂ t ∣ m ε ∣ 2 = ( 2 α q ( ∣ m ε ∣ ) + ε ) Δ∣ m ε ∣ 2 − 2 ( 2 α q ( ∣ m ε ∣ ) + ε ) ∣∇ m ε ∣ 2 + 4 q ( ∣ m ε ∣ ) m ε ⋅ ( B m ε + H d [ m ε ]) − 4 θ ∣ m ε ∣ 2 . (49)
Here, we have used the fact that 2 m ε ⋅ Δ m ε = Δ ∣ m ε ∣ 2 − 2 ∣ ∇ m ε ∣ 2 2m_\varepsilon\cdot\Delta m_\varepsilon= \Delta|m_\varepsilon|^2 - 2|\nabla m_\varepsilon|^2 2 m ε ⋅ Δ m ε = Δ∣ m ε ∣ 2 − 2∣∇ m ε ∣ 2 .
Now, set
w ε : = ( ∣ m ε ∣ 2 − 1 ) + . w_\varepsilon:= (|m_\varepsilon|^2-1)_+. w ε := ( ∣ m ε ∣ 2 − 1 ) + .
We first justify that w ε w_\varepsilon w ε is an admissible test function. For fixed ε > 0 \varepsilon>0 ε > 0 , the regularity (47) implies m ε ∈ L 2 ( 0 , T ; L ∞ ( D ) ) m_\varepsilon\in L^2(0,T;L^\infty(\mathcal{D})) m ε ∈ L 2 ( 0 , T ; L ∞ ( D )) since H 2 ( D ) ↪ L ∞ ( D ) H^2(\mathcal{D}) \hookrightarrow L^\infty(\mathcal{D}) H 2 ( D ) ↪ L ∞ ( D ) . Consequently,
∂ t ∣ m ε ∣ 2 = 2 m ε ⋅ ∂ t m ε ∈ L 1 ( 0 , T ; L 2 ( D ) ) , \partial_t|m_\varepsilon|^2 = 2m_\varepsilon\cdot\partial_t m_\varepsilon\in L^1(0,T;L^2(\mathcal{D})), ∂ t ∣ m ε ∣ 2 = 2 m ε ⋅ ∂ t m ε ∈ L 1 ( 0 , T ; L 2 ( D )) ,
∇ ∣ m ε ∣ 2 = 2 m ε ⋅ ∇ m ε ∈ L 2 ( 0 , T ; L 2 ( D ) ) . \nabla|m_\varepsilon|^2 = 2m_\varepsilon\cdot\nabla m_\varepsilon\in L^2(0,T;L^2(\mathcal{D})). ∇∣ m ε ∣ 2 = 2 m ε ⋅ ∇ m ε ∈ L 2 ( 0 , T ; L 2 ( D )) .
Hence,
∣ m ε ∣ 2 − 1 ∈ W 1 , 1 ( 0 , T ; L 2 ( D ) ) ∩ L 2 ( 0 , T ; H 1 ( D ) ) . |m_\varepsilon|^2-1 \in W^{1,1}(0,T;L^2(\mathcal{D})) \cap L^2(0,T;H^1(\mathcal{D})). ∣ m ε ∣ 2 − 1 ∈ W 1 , 1 ( 0 , T ; L 2 ( D )) ∩ L 2 ( 0 , T ; H 1 ( D )) .
Since the map s ↦ s + s \mapsto s_+ s ↦ s + is Lipschitz, the Sobolev chain rule gives
w ε ∈ W 1 , 1 ( 0 , T ; L 2 ( D ) ) ∩ L 2 ( 0 , T ; H 1 ( D ) ) , w_\varepsilon\in W^{1,1}(0,T;L^2(\mathcal{D})) \cap L^2(0,T;H^1(\mathcal{D})), w ε ∈ W 1 , 1 ( 0 , T ; L 2 ( D )) ∩ L 2 ( 0 , T ; H 1 ( D )) ,
with
∂ t w ε = 1 { ∣ m ε ∣ 2 > 1 } ∂ t ∣ m ε ∣ 2 , ∇ w ε = 1 { ∣ m ε ∣ 2 > 1 } ∇ ∣ m ε ∣ 2 a.e. (50) \partial_t w_\varepsilon= \mathbf{1}_{\{|m_\varepsilon|^2>1\}}\partial_t|m_\varepsilon|^2,\qquad
\nabla w_\varepsilon= \mathbf{1}_{\{|m_\varepsilon|^2>1\}}\nabla|m_\varepsilon|^2
\quad\text{a.e.} \tag*{(50)} ∂ t w ε = 1 { ∣ m ε ∣ 2 > 1 } ∂ t ∣ m ε ∣ 2 , ∇ w ε = 1 { ∣ m ε ∣ 2 > 1 } ∇∣ m ε ∣ 2 a.e. (50)
In particular, t ↦ ∥ w ε ( t ) ∥ L 2 2 t \mapsto\|w_\varepsilon(t)\|_{L^2}^2 t ↦ ∥ w ε ( t ) ∥ L 2 2 is absolutely continuous and
1 2 d d t ∥ w ε ∥ L 2 2 = ∫ D ∂ t ∣ m ε ∣ 2 w ε d x for a.e. t . (51) \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\|w_\varepsilon\|_{L^2}^2 = \int_{\mathcal{D}}\partial_t|m_\varepsilon|^2w_\varepsilon\,\mathrm{d}x
\qquad\text{for a.e. }t. \tag*{(51)} 2 1 d t d ∥ w ε ∥ L 2 2 = ∫ D ∂ t ∣ m ε ∣ 2 w ε d x for a.e. t . (51)
Indeed, this is the chain rule applied to s ↦ 1 2 ( s + ) 2 s \mapsto\frac{1}{2}(s_+)^2 s ↦ 2 1 ( s + ) 2 , whose derivative is s + s_+ s + .
We may therefore test (49) by w ε w_\varepsilon w ε . On the set where w ε > 0 w_\varepsilon>0 w ε > 0 , one has ∣ m ε ∣ > 1 |m_\varepsilon|>1 ∣ m ε ∣ > 1 , and thus q ^ ( ∣ m ε ∣ ) = 0 \widehat{q}(|m_\varepsilon|)=0 q ( ∣ m ε ∣ ) = 0 . Therefore, all terms multiplied by q ^ ( ∣ m ε ∣ ) \widehat{q}(|m_\varepsilon|) q ( ∣ m ε ∣ ) vanish after multiplication by w ε w_\varepsilon w ε . Moreover, using (50) , we have
− ε ∫ D ∇ ∣ m ε ∣ 2 ⋅ ∇ w ε d x = − ε ∥ ∇ w ε ∥ L 2 2 , -\varepsilon\int_{\mathcal{D}}\nabla|m_\varepsilon|^2\cdot\nabla w_\varepsilon\,\mathrm{d}x
= -\varepsilon\|\nabla w_\varepsilon\|_{L^2}^2, − ε ∫ D ∇∣ m ε ∣ 2 ⋅ ∇ w ε d x = − ε ∥∇ w ε ∥ L 2 2 ,
where there is no boundary contribution since ∂ ν ∣ m ε ∣ 2 = 2 m ε ⋅ ∂ ν m ε = 0 \partial_\nu|m_\varepsilon|^2=2m_\varepsilon\cdot\partial_\nu m_\varepsilon=0 ∂ ν ∣ m ε ∣ 2 = 2 m ε ⋅ ∂ ν m ε = 0 on ∂ D \partial\mathcal{D} ∂ D . Hence, for almost every t > 0 t>0 t > 0 , we obtain
1 2 d d t ∥ w ε ∥ L 2 ( D ) 2 + ε ∥ ∇ w ε ∥ L 2 ( D ) 2 + 2 ε ∫ D ∣ ∇ m ε ∣ 2 w ε d x + 4 θ ∫ D ∣ m ε ∣ 2 w ε d x = 0. (52) \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\|w_\varepsilon\|_{L^2(\mathcal{D})}^2
+\varepsilon\|\nabla w_\varepsilon\|_{L^2(\mathcal{D})}^2
+2\varepsilon\int_{\mathcal{D}}|\nabla m_\varepsilon|^2w_\varepsilon\,\mathrm{d}x
+4\theta\int_{\mathcal{D}}|m_\varepsilon|^2w_\varepsilon\,\mathrm{d}x
=0. \tag*{(52)} 2 1 d t d ∥ w ε ∥ L 2 ( D ) 2 + ε ∥∇ w ε ∥ L 2 ( D ) 2 + 2 ε ∫ D ∣∇ m ε ∣ 2 w ε d x + 4 θ ∫ D ∣ m ε ∣ 2 w ε d x = 0. (52)
Since ∣ m 0 ∣ ≤ 1 |m_0|\le1 ∣ m 0 ∣ ≤ 1 almost everywhere,
w ε ( 0 ) = ( ∣ m 0 ∣ 2 − 1 ) + = 0. w_\varepsilon(0)=(|m_0|^2-1)_+=0. w ε ( 0 ) = ( ∣ m 0 ∣ 2 − 1 ) + = 0.
Integrating (52) over ( 0 , t ) (0,t) ( 0 , t ) and dropping all non-negative terms on the left-hand side yields ∥ w ε ( t ) ∥ L 2 ( D ) 2 = 0 \|w_\varepsilon(t)\|_{L^2(\mathcal{D})}^2=0 ∥ w ε ( t ) ∥ L 2 ( D ) 2 = 0 for every t ≥ 0 t\ge0 t ≥ 0 , which implies ∣ m ε ( t , x ) ∣ ≤ 1 |m_\varepsilon(t,x)|\le1 ∣ m ε ( t , x ) ∣ ≤ 1 for a.e. ( t , x ) (t,x) ( t , x ) . Moreover, since m ε ∈ C ( [ 0 , T ] ; L 2 ( D ) ) m_\varepsilon\in C([0,T];L^2(\mathcal{D})) m ε ∈ C ([ 0 , T ] ; L 2 ( D )) , the bound extends from a.e. t t t to every t ≥ 0 t \ge0 t ≥ 0 by approximation with a sequence of times for which (48) holds and strong L 2 L^2 L 2 -convergence. This proves the assertion. □
After Lemma 3.5 is established, the artificial extensions are inactive, so that
π ( m ε ) = m ε , M ^ ( m ε ) = M ( m ε ) for a.e. x ∈ D and every t ≥ 0. (53) \pi(m_\varepsilon)=m_\varepsilon,\qquad\widehat{\mathrm{M}}(m_\varepsilon)=\mathrm{M}(m_\varepsilon)\qquad\text{for a.e. }x\in\mathcal{D}\text{ and every }t\ge0. \tag*{(53)} π ( m ε ) = m ε , M ( m ε ) = M ( m ε ) for a.e. x ∈ D and every t ≥ 0. (53)
The following uniform estimates will be used repeatedly in subsequent argument.
Lemma 3.6. Let m ε m_\varepsilon m ε be a solution of (3.10), with ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] . There exists a constant C > 0 C>0 C > 0 such that
sup t ≥ 0 ∥ m ε ( t ) ∥ H 1 ( D ) 2 + sup t ≥ 0 ∫ t t + 1 ∫ D Δ m ε ⋅ M ( m ε ) Δ m ε d x d s + ε sup t ≥ 0 ∫ t t + 1 ∥ Δ m ε ( s ) ∥ L 2 ( D ) 2 d s + sup t ≥ 0 ∫ t t + 1 ∥ ∂ t m ε ( s ) ∥ L 2 ( D ) 2 d s ≤ C . (54) \sup_{t\ge0}\|m_\varepsilon(t)\|_{H^1(\mathcal{D})}^2+\sup_{t\ge0}\int_t^{t+1}\int_{\mathcal{D}}\Delta m_\varepsilon\cdot\mathrm{M}(m_\varepsilon)\Delta m_\varepsilon\,\mathrm{d}x\,\mathrm{d}s+\varepsilon\sup_{t\ge0}\int_t^{t+1}\|\Delta m_\varepsilon(s)\|_{L^2(\mathcal{D})}^2\,\mathrm{d}s+\sup_{t\ge0}\int_t^{t+1}\|\partial_t m_\varepsilon(s)\|_{L^2(\mathcal{D})}^2\,\mathrm{d}s\le C. \tag*{(54)} t ≥ 0 sup ∥ m ε ( t ) ∥ H 1 ( D ) 2 + t ≥ 0 sup ∫ t t + 1 ∫ D Δ m ε ⋅ M ( m ε ) Δ m ε d x d s + ε t ≥ 0 sup ∫ t t + 1 ∥Δ m ε ( s ) ∥ L 2 ( D ) 2 d s + t ≥ 0 sup ∫ t t + 1 ∥ ∂ t m ε ( s ) ∥ L 2 ( D ) 2 d s ≤ C . (54) The constant C > 0 C>0 C > 0 depends only on D \mathcal{D} D , α \alpha α , θ \theta θ , γ \gamma γ , B B B , and m 0 m_0 m 0 , but is independent of ε \varepsilon ε , t t t , and the final time T T T .
Proof. By Lemma 3.5, we have ∥ m ε ( t ) ∥ L 2 ( D ) 2 ≤ ∣ D ∣ \|m_\varepsilon(t)\|_{L^2(\mathcal{D})}^2\le|\mathcal{D}| ∥ m ε ( t ) ∥ L 2 ( D ) 2 ≤ ∣ D ∣ for all $t\ge 0. Furthermore, we have the regularity (3.19) so that the viscous equation (3.10) holds in L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) for a.e. t > 0 t>0 t > 0 . Thus, by the same argument leading to (3.16), but applied directly to (3.10), we obtain for a.e. t > 0 t>0 t > 0 ,
1 2 d d t ∥ ∇ m ε ∥ L 2 ( D ) 2 + α 2 ∫ D Δ m ε ⋅ M ( m ε ) Δ m ε d x + 2 θ ∥ ∇ m ε ∥ L 2 ( D ) 2 + ε ∥ Δ m ε ∥ L 2 ( D ) 2 ≤ C , (55) \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\|\nabla m_\varepsilon\|_{L^2(\mathcal{D})}^2+\frac{\alpha}{2}\int_{\mathcal{D}}\Delta m_\varepsilon\cdot\mathrm{M}(m_\varepsilon)\Delta m_\varepsilon\,\mathrm{d}x+2\theta\|\nabla m_\varepsilon\|_{L^2(\mathcal{D})}^2+\varepsilon\|\Delta m_\varepsilon\|_{L^2(\mathcal{D})}^2\le C, \tag*{(55)} 2 1 d t d ∥∇ m ε ∥ L 2 ( D ) 2 + 2 α ∫ D Δ m ε ⋅ M ( m ε ) Δ m ε d x + 2 θ ∥∇ m ε ∥ L 2 ( D ) 2 + ε ∥Δ m ε ∥ L 2 ( D ) 2 ≤ C , (55)
where C C C is independent of ε \varepsilon ε . Discarding the non-negative second-order dissipation terms yields
d d t ∥ ∇ m ε ∥ L 2 ( D ) 2 + 4 θ ∥ ∇ m ε ∥ L 2 ( D ) 2 ≤ C . \frac{\mathrm{d}}{\mathrm{d}t}\|\nabla m_\varepsilon\|_{L^2(\mathcal{D})}^2+4\theta\|\nabla m_\varepsilon\|_{L^2(\mathcal{D})}^2\le C. d t d ∥∇ m ε ∥ L 2 ( D ) 2 + 4 θ ∥∇ m ε ∥ L 2 ( D ) 2 ≤ C .
By the Gronwall lemma, we obtain
∥ ∇ m ε ( t ) ∥ L 2 ( D ) 2 ≤ e − 4 θ t ∥ ∇ m 0 ∥ L 2 ( D ) 2 + C ( 1 − e − 4 θ t ) , (56) \|\nabla m_\varepsilon(t)\|_{L^2(\mathcal{D})}^2\le e^{-4\theta t}\|\nabla m_0\|_{L^2(\mathcal{D})}^2+C(1-e^{-4\theta t}), \tag*{(56)} ∥∇ m ε ( t ) ∥ L 2 ( D ) 2 ≤ e − 4 θ t ∥∇ m 0 ∥ L 2 ( D ) 2 + C ( 1 − e − 4 θ t ) , (56)
and hence
sup t ≥ 0 ∥ m ε ( t ) ∥ H 1 ( D ) 2 ≤ C . (57) \sup_{t\ge0}\|m_\varepsilon(t)\|_{H^1(\mathcal{D})}^2\le C. \tag*{(57)} t ≥ 0 sup ∥ m ε ( t ) ∥ H 1 ( D ) 2 ≤ C . (57)
Next, integrating (55) over ( t , t + 1 ) (t,t+1) ( t , t + 1 ) gives
1 2 ∥ ∇ m ε ( t + 1 ) ∥ L 2 2 − 1 2 ∥ ∇ m ε ( t ) ∥ L 2 2 + α 2 ∫ t t + 1 ∫ D Δ m ε ⋅ M ( m ε ) Δ m ε d x d s + 2 θ ∫ t t + 1 ∥ ∇ m ε ( s ) ∥ L 2 2 d s + ε ∫ t t + 1 ∥ Δ m ε ( s ) ∥ L 2 2 d s ≤ C . \frac{1}{2}\|\nabla m_\varepsilon(t+1)\|_{L^2}^2-\frac{1}{2}\|\nabla m_\varepsilon(t)\|_{L^2}^2+\frac{\alpha}{2}\int_t^{t+1}\int_{\mathcal{D}}\Delta m_\varepsilon\cdot\mathrm{M}(m_\varepsilon)\Delta m_\varepsilon\,\mathrm{d}x\,\mathrm{d}s+2\theta\int_t^{t+1}\|\nabla m_\varepsilon(s)\|_{L^2}^2\,\mathrm{d}s+\varepsilon\int_t^{t+1}\|\Delta m_\varepsilon(s)\|_{L^2}^2\,\mathrm{d}s\le C. 2 1 ∥∇ m ε ( t + 1 ) ∥ L 2 2 − 2 1 ∥∇ m ε ( t ) ∥ L 2 2 + 2 α ∫ t t + 1 ∫ D Δ m ε ⋅ M ( m ε ) Δ m ε d x d s + 2 θ ∫ t t + 1 ∥∇ m ε ( s ) ∥ L 2 2 d s + ε ∫ t t + 1 ∥Δ m ε ( s ) ∥ L 2 2 d s ≤ C .
Using (57) and dropping the non-negative terms gives
sup t ≥ 0 ∫ t t + 1 ∫ D Δ m ε ⋅ M ( m ε ) Δ m ε d x d s + ε sup t ≥ 0 ∫ t t + 1 ∥ Δ m ε ( s ) ∥ L 2 ( D ) 2 d s ≤ C . (58) \sup_{t\ge0}\int_t^{t+1}\int_{\mathcal{D}}\Delta m_\varepsilon\cdot\mathrm{M}(m_\varepsilon)\Delta m_\varepsilon\,\mathrm{d}x\,\mathrm{d}s+\varepsilon\sup_{t\ge0}\int_t^{t+1}\|\Delta m_\varepsilon(s)\|_{L^2(\mathcal{D})}^2\,\mathrm{d}s\le C. \tag*{(58)} t ≥ 0 sup ∫ t t + 1 ∫ D Δ m ε ⋅ M ( m ε ) Δ m ε d x d s + ε t ≥ 0 sup ∫ t t + 1 ∥Δ m ε ( s ) ∥ L 2 ( D ) 2 d s ≤ C . (58)
By the same argument leading to (3.17), but using (3.10) and noting that ∣ m ε ∣ ≤ 1 |m_\varepsilon|\le1 ∣ m ε ∣ ≤ 1 , we obtain
∥ ∂ t m ε ∥ L 2 ( D ) 2 ≤ C ∫ D Δ m ε ⋅ M ( m ε ) Δ m ε d x + C + C ε 2 ∥ Δ m ε ∥ L 2 ( D ) 2 . \|\partial_t m_\varepsilon\|_{L^2(\mathcal{D})}^2\le C\int_{\mathcal{D}}\Delta m_\varepsilon\cdot\mathrm{M}(m_\varepsilon)\Delta m_\varepsilon\,\mathrm{d}x+C+C\varepsilon^2\|\Delta m_\varepsilon\|_{L^2(\mathcal{D})}^2. ∥ ∂ t m ε ∥ L 2 ( D ) 2 ≤ C ∫ D Δ m ε ⋅ M ( m ε ) Δ m ε d x + C + C ε 2 ∥Δ m ε ∥ L 2 ( D ) 2 .
Integrating this inequality over ( t , t + 1 ) (t,t+1) ( t , t + 1 ) and using (58) , together with ε ≤ 1 \varepsilon\le1 ε ≤ 1 , yields
sup t ≥ 0 ∫ t t + 1 ∥ ∂ t m ε ( s ) ∥ L 2 ( D ) 2 d s ≤ C . \sup_{t\ge0}\int_t^{t+1}\|\partial_t m_\varepsilon(s)\|_{L^2(\mathcal{D})}^2\,\mathrm{d}s\le C. t ≥ 0 sup ∫ t t + 1 ∥ ∂ t m ε ( s ) ∥ L 2 ( D ) 2 d s ≤ C .
Combining this with the preceding estimates (57) and (58) proves (54) . □
Instantaneous inverse-distance and regularity estimates
Next, we derive an entropy-type estimate controlling the approach of m ε m_\varepsilon m ε to saturation. Since
d ε ( s ) : = 2 α q ( s ) + ε ≃ ε + 1 − s , d_\varepsilon(s):=2\alpha q(\sqrt{s})+\varepsilon\simeq\varepsilon+1-s, d ε ( s ) := 2 α q ( s ) + ε ≃ ε + 1 − s ,
negative powers of d ε ( ∣ m ε ∣ 2 ) d_\varepsilon(|m_\varepsilon|^2) d ε ( ∣ m ε ∣ 2 ) control the inverse regularised gap to ∣ m ε ∣ = 1 |m_\varepsilon|=1 ∣ m ε ∣ = 1 . Testing the scalar equation for ∣ m ε ∣ 2 |m_\varepsilon|^2 ∣ m ε ∣ 2 by d ε ( ∣ m ε ∣ 2 ) − p − 1 d_\varepsilon(|m_\varepsilon|^2)^{-p-1} d ε ( ∣ m ε ∣ 2 ) − p − 1 yields the following ε \varepsilon ε -uniform estimate, which will later be used to prove strict separation.
Proposition 3.7. For every p ≥ 1 p\geq1 p ≥ 1 , there exists C p > 0 C_p>0 C p > 0 , independent of ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , such that
∫ D d x ( ε + 1 − ∣ m ε ( t , x ) ∣ 2 ) p ≤ C p ( 1 + t − p ) for a.e. t > 0. (59) \int_{\mathcal{D}} \frac{\mathrm{d}x}{(\varepsilon+1-|m_\varepsilon(t,x)|^2)^p}\leq C_p(1+t^{-p})\qquad\text{for a.e. }t>0. \tag*{(59)} ∫ D ( ε + 1 − ∣ m ε ( t , x ) ∣ 2 ) p d x ≤ C p ( 1 + t − p ) for a.e. t > 0. (59)
Proof. Let q q q be defined by (7) and p ≥ 1 p\geq1 p ≥ 1 . For s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] , set
d ε ( s ) : = 2 α q ( s ) + ε , Ψ p , ε ( s ) : = ∫ 0 s d ε ( r ) − p − 1 d r . d_\varepsilon(s):=2\alpha q(\sqrt{s})+\varepsilon,\qquad\Psi_{p,\varepsilon}(s):=\int_0^s d_\varepsilon(r)^{-p-1}\,\mathrm{d}r. d ε ( s ) := 2 α q ( s ) + ε , Ψ p , ε ( s ) := ∫ 0 s d ε ( r ) − p − 1 d r .
By (A.1), there exist constants c , C > 0 c,C>0 c , C > 0 , independent of ε ∈ ( 0 , 1 ] \varepsilon\in(0,1] ε ∈ ( 0 , 1 ] , such that
c ( ε + 1 − s ) ≤ d ε ( s ) ≤ C ( ε + 1 − s ) for all s ∈ [ 0 , 1 ] . (60) c(\varepsilon+1-s)\leq d_\varepsilon(s)\leq C(\varepsilon+1-s)\qquad\text{for all }s\in[0,1]. \tag*{(60)} c ( ε + 1 − s ) ≤ d ε ( s ) ≤ C ( ε + 1 − s ) for all s ∈ [ 0 , 1 ] . (60)
A direct integration gives
c d ε ( s ) − p − C ≤ Ψ p , ε ( s ) ≤ C d ε ( s ) − p for all s ∈ [ 0 , 1 ] , (61) c d_\varepsilon(s)^{-p}-C\leq\Psi_{p,\varepsilon}(s)\leq C d_\varepsilon(s)^{-p}\qquad\text{for all }s\in[0,1], \tag*{(61)} c d ε ( s ) − p − C ≤ Ψ p , ε ( s ) ≤ C d ε ( s ) − p for all s ∈ [ 0 , 1 ] , (61)
with constants independent of ε \varepsilon ε .
By Lemma 3.5, ∣ m ε ∣ ≤ 1 |m_\varepsilon|\leq1 ∣ m ε ∣ ≤ 1 , and hence q ^ ( ∣ m ε ∣ ) = q ( ∣ m ε ∣ ) \widehat{q}(|m_\varepsilon|)=q(|m_\varepsilon|) q ( ∣ m ε ∣ ) = q ( ∣ m ε ∣ ) in (49) . We now multiply that equation by d ε ( ∣ m ε ∣ 2 ) − p − 1 = Ψ p , ε ′ ( ∣ m ε ∣ 2 ) d_\varepsilon(|m_\varepsilon|^2)^{-p-1}=\Psi'_{p,\varepsilon}(|m_\varepsilon|^2) d ε ( ∣ m ε ∣ 2 ) − p − 1 = Ψ p , ε ′ ( ∣ m ε ∣ 2 ) and integrate over D \mathcal{D} D . First, note that this nonlinear testing is legitimate. Indeed, for fixed ε > 0 \varepsilon>0 ε > 0 , the regularity (47) and (48) imply
∣ m ε ∣ 2 ∈ H 1 ( 0 , T ; L 2 ( D ) ) ∩ L ∞ ( 0 , T ; H 1 ( D ) ) . |m_\varepsilon|^2\in H^1(0,T;L^2(\mathcal{D}))\cap L^\infty(0,T;H^1(\mathcal{D})). ∣ m ε ∣ 2 ∈ H 1 ( 0 , T ; L 2 ( D )) ∩ L ∞ ( 0 , T ; H 1 ( D )) .
Moreover, d ε ≥ ε > 0 d_\varepsilon\geq\varepsilon>0 d ε ≥ ε > 0 , and s ↦ q ( s ) s\mapsto q(\sqrt{s}) s ↦ q ( s ) is continuously differentiable on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Hence Ψ p , ε ′ \Psi'_{p,\varepsilon} Ψ p , ε ′ is Lipschitz on [ 0 , 1 ] [0,1] [ 0 , 1 ] for fixed ε \varepsilon ε . The temporal Sobolev chain rule thus gives
d d t ∫ D Ψ p , ε ( ∣ m ε ∣ 2 ) d x = ∫ D d ε ( ∣ m ε ∣ 2 ) − p − 1 ∂ t ∣ m ε ∣ 2 d x for a.e. t > 0. \frac{\mathrm{d}}{\mathrm{d}t}\int_{\mathcal{D}}\Psi_{p,\varepsilon}(|m_\varepsilon|^2)\,\mathrm{d}x
=
\int_{\mathcal{D}}d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\partial_t|m_\varepsilon|^2\,\mathrm{d}x
\qquad\text{for a.e. }t>0. d t d ∫ D Ψ p , ε ( ∣ m ε ∣ 2 ) d x = ∫ D d ε ( ∣ m ε ∣ 2 ) − p − 1 ∂ t ∣ m ε ∣ 2 d x for a.e. t > 0.
Consequently, using (49) , we obtain
d d t ∫ D Ψ p , ε ( ∣ m ε ∣ 2 ) d x = ∫ D d ε ( ∣ m ε ∣ 2 ) − p Δ ∣ m ε ∣ 2 d x − 2 ∫ D d ε ( ∣ m ε ∣ 2 ) − p ∣ ∇ m ε ∣ 2 d x + 4 ∫ D q ( ∣ m ε ∣ ) d ε ( ∣ m ε ∣ 2 ) − p − 1 m ε ⋅ B m ε d x + 4 ∫ D q ( ∣ m ε ∣ ) d ε ( ∣ m ε ∣ 2 ) − p − 1 m ε ⋅ H d [ m ε ] d x − 4 θ ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x = − ∫ D d d s [ d ε ( s ) − p ] ∣ s = ∣ m ε ∣ 2 ∣ ∇ ∣ m ε ∣ ∣ 2 d x − 2 ∫ D d ε ( ∣ m ε ∣ 2 ) − p ∣ ∇ m ε ∣ 2 d x + 4 ∫ D q ( ∣ m ε ∣ ) d ε ( ∣ m ε ∣ 2 ) − p − 1 m ε ⋅ B m ε d x ⏟ anisotropy term + 4 ∫ D q ( ∣ m ε ∣ ) d ε ( ∣ m ε ∣ 2 ) − p − 1 m ε ⋅ H d [ m ε ] d x ⏟ stray-field term − 4 θ ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x . (62) \begin{aligned}
\frac{\mathrm{d}}{\mathrm{d}t}\int_{\mathcal{D}}\Psi_{p,\varepsilon}(|m_\varepsilon|^2)\,\mathrm{d}x
={}&\int_{\mathcal{D}}d_\varepsilon(|m_\varepsilon|^2)^{-p}\Delta|m_\varepsilon|^2\,\mathrm{d}x
-2\int_{\mathcal{D}}d_\varepsilon(|m_\varepsilon|^2)^{-p}|\nabla m_\varepsilon|^2\,\mathrm{d}x\\
&+4\int_{\mathcal{D}}q(|m_\varepsilon|)d_\varepsilon(|m_\varepsilon|^2)^{-p-1}m_\varepsilon\cdot Bm_\varepsilon\,\mathrm{d}x\\
&+4\int_{\mathcal{D}}q(|m_\varepsilon|)d_\varepsilon(|m_\varepsilon|^2)^{-p-1}m_\varepsilon\cdot H_{\mathrm d}[m_\varepsilon]\,\mathrm{d}x\\
&-4\theta\int_{\mathcal{D}}|m_\varepsilon|^2d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\,\mathrm{d}x\\
={}&-\int_{\mathcal{D}}\left.\frac{\mathrm{d}}{\mathrm{d}s}[d_\varepsilon(s)^{-p}]\right|_{s=|m_\varepsilon|^2}|\nabla|m_\varepsilon||^2\,\mathrm{d}x
-2\int_{\mathcal{D}}d_\varepsilon(|m_\varepsilon|^2)^{-p}|\nabla m_\varepsilon|^2\,\mathrm{d}x\\
&+\underbrace{4\int_{\mathcal{D}}q(|m_\varepsilon|)d_\varepsilon(|m_\varepsilon|^2)^{-p-1}m_\varepsilon\cdot Bm_\varepsilon\,\mathrm{d}x}_{\text{anisotropy term}}\\
&+\underbrace{4\int_{\mathcal{D}}q(|m_\varepsilon|)d_\varepsilon(|m_\varepsilon|^2)^{-p-1}m_\varepsilon\cdot H_{\mathrm d}[m_\varepsilon]\,\mathrm{d}x}_{\text{stray-field term}}\\
&-4\theta\int_{\mathcal{D}}|m_\varepsilon|^2d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\,\mathrm{d}x. \tag*{(62)}
\end{aligned} d t d ∫ D Ψ p , ε ( ∣ m ε ∣ 2 ) d x = = ∫ D d ε ( ∣ m ε ∣ 2 ) − p Δ∣ m ε ∣ 2 d x − 2 ∫ D d ε ( ∣ m ε ∣ 2 ) − p ∣∇ m ε ∣ 2 d x + 4 ∫ D q ( ∣ m ε ∣ ) d ε ( ∣ m ε ∣ 2 ) − p − 1 m ε ⋅ B m ε d x + 4 ∫ D q ( ∣ m ε ∣ ) d ε ( ∣ m ε ∣ 2 ) − p − 1 m ε ⋅ H d [ m ε ] d x − 4 θ ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x − ∫ D d s d [ d ε ( s ) − p ] s = ∣ m ε ∣ 2 ∣∇∣ m ε ∣ ∣ 2 d x − 2 ∫ D d ε ( ∣ m ε ∣ 2 ) − p ∣∇ m ε ∣ 2 d x + anisotropy term 4 ∫ D q ( ∣ m ε ∣ ) d ε ( ∣ m ε ∣ 2 ) − p − 1 m ε ⋅ B m ε d x + stray-field term 4 ∫ D q ( ∣ m ε ∣ ) d ε ( ∣ m ε ∣ 2 ) − p − 1 m ε ⋅ H d [ m ε ] d x − 4 θ ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x . (62)
Here in the last equality, the first term on the right-hand side is obtained using integration by parts and the fact that ∂ ν ∣ m ε ∣ 2 = 0 \partial_\nu|m_\varepsilon|^2=0 ∂ ν ∣ m ε ∣ 2 = 0 on ∂ D \partial\mathcal{D} ∂ D . Since q q q is decreasing, d ε d_\varepsilon d ε is decreasing, and hence
d d s [ d ε ( s ) − p ] = − p d ε ( s ) − p − 1 d ε ′ ( s ) ≥ 0. \frac{\mathrm{d}}{\mathrm{d}s}\left[d_\varepsilon(s)^{-p}\right]=-pd_\varepsilon(s)^{-p-1}d_\varepsilon'(s)\ge0. d s d [ d ε ( s ) − p ] = − p d ε ( s ) − p − 1 d ε ′ ( s ) ≥ 0.
Thus the first two terms and the last term on the right-hand side of (62) are nonpositive. It remains only to estimate the anisotropy and stray-field terms.
For the anisotropy term, using the fact that B B B is symmetric and the definition of d ε d_\varepsilon d ε , we have
m ε ⋅ B m ε ≤ ∥ B ∥ ∣ m ε ∣ 2 , q ( ∣ m ε ∣ ) ≤ 1 2 α d ε ( ∣ m ε ∣ 2 ) . m_\varepsilon\cdot Bm_\varepsilon\le\|B\||m_\varepsilon|^2,\qquad q(|m_\varepsilon|)\le\frac{1}{2\alpha}d_\varepsilon(|m_\varepsilon|^2). m ε ⋅ B m ε ≤ ∥ B ∥∣ m ε ∣ 2 , q ( ∣ m ε ∣ ) ≤ 2 α 1 d ε ( ∣ m ε ∣ 2 ) .
Consequently, for every η > 0 \eta>0 η > 0 ,
4 ∫ D q ( ∣ m ε ∣ ) d ε ( ∣ m ε ∣ 2 ) − p − 1 m ε ⋅ B m ε d x ≤ C ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p d x ≤ η ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x + C η , p . (63) \begin{aligned}
4\int_{\mathcal{D}}q(|m_\varepsilon|)d_\varepsilon(|m_\varepsilon|^2)^{-p-1}m_\varepsilon\cdot Bm_\varepsilon\,\mathrm{d}x
\le C\int_{\mathcal{D}}|m_\varepsilon|^2d_\varepsilon(|m_\varepsilon|^2)^{-p}\,\mathrm{d}x \\
&\le\eta\int_{\mathcal{D}}|m_\varepsilon|^2d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\,\mathrm{d}x+C_{\eta,p}. \tag*{(63)}
\end{aligned} 4 ∫ D q ( ∣ m ε ∣ ) d ε ( ∣ m ε ∣ 2 ) − p − 1 m ε ⋅ B m ε d x ≤ C ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p d x ≤ η ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x + C η , p . (63)
Indeed, on the set { ∣ m ε ∣ 2 ≤ 1 / 2 } \{|m_\varepsilon|^2\le1/2\} { ∣ m ε ∣ 2 ≤ 1/2 } , we have
d ε ( ∣ m ε ∣ 2 ) ≥ 2 α q ( 1 / 2 ) > 0 , d_\varepsilon(|m_\varepsilon|^2)\ge2\alpha q(1/\sqrt{2})>0, d ε ( ∣ m ε ∣ 2 ) ≥ 2 α q ( 1/ 2 ) > 0 ,
so the corresponding integrand is uniformly bounded. On { ∣ m ε ∣ 2 > 1 / 2 } \{|m_\varepsilon|^2>1/2\} { ∣ m ε ∣ 2 > 1/2 } , thanks to the elementary inequality x p ≤ η x p + 1 + C η , p x^p\le\eta x^{p+1}+C_{\eta,p} x p ≤ η x p + 1 + C η , p for any x ≥ 0 x\ge0 x ≥ 0 , applied to x = d ε ( ∣ m ε ∣ 2 ) − 1 x=d_\varepsilon(|m_\varepsilon|^2)^{-1} x = d ε ( ∣ m ε ∣ 2 ) − 1 , we obtain (63) .
For the stray-field term in (62) , we similarly infer that
4 ∫ D q ( ∣ m ε ∣ ) d ε ( ∣ m ε ∣ 2 ) − p − 1 m ε ⋅ H d [ m ε ] d x ≤ C ∫ D d ε ( ∣ m ε ∣ 2 ) − p ∣ H d [ m ε ] ∣ d x ≤ η ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x + C η , p ( 1 + ∥ H d [ m ε ] ∥ L p + 1 ( R 3 ) p + 1 ) ≤ η ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x + C η , p . (64) \begin{aligned}
4\int_{\mathcal{D}}q(|m_\varepsilon|)d_\varepsilon(|m_\varepsilon|^2)^{-p-1}m_\varepsilon\cdot H_d[m_\varepsilon]\,\mathrm{d}x
\le C\int_{\mathcal{D}}d_\varepsilon(|m_\varepsilon|^2)^{-p}|H_d[m_\varepsilon]|\,\mathrm{d}x \\
&\le\eta\int_{\mathcal{D}}|m_\varepsilon|^2d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\,\mathrm{d}x+C_{\eta,p}\left(1+\|H_d[m_\varepsilon]\|_{L^{p+1}(\mathbb{R}^3)}^{p+1}\right) \\
&\le\eta\int_{\mathcal{D}}|m_\varepsilon|^2d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\,\mathrm{d}x+C_{\eta,p}. \tag*{(64)}
\end{aligned} 4 ∫ D q ( ∣ m ε ∣ ) d ε ( ∣ m ε ∣ 2 ) − p − 1 m ε ⋅ H d [ m ε ] d x ≤ C ∫ D d ε ( ∣ m ε ∣ 2 ) − p ∣ H d [ m ε ] ∣ d x ≤ η ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x + C η , p ( 1 + ∥ H d [ m ε ] ∥ L p + 1 ( R 3 ) p + 1 ) ≤ η ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x + C η , p . (64)
To justify the penultimate inequality, we see that on the set { ∣ m ε ∣ 2 ≤ 1 / 2 } \{|m_\varepsilon|^2\le1/2\} { ∣ m ε ∣ 2 ≤ 1/2 } , the factor d ε ( ∣ m ε ∣ 2 ) − p d_\varepsilon(|m_\varepsilon|^2)^{-p} d ε ( ∣ m ε ∣ 2 ) − p is uniformly bounded, so Hölder’s inequality controls this part by C ∥ H d [ m ε ] ∥ L p + 1 C\|H_d[m_\varepsilon]\|_{L^{p+1}} C ∥ H d [ m ε ] ∥ L p + 1 . On { ∣ m ε ∣ 2 > 1 / 2 } \{|m_\varepsilon|^2>1/2\} { ∣ m ε ∣ 2 > 1/2 } , Young’s inequality with conjugate exponents ( p + 1 ) / p (p+1)/p ( p + 1 ) / p and p + 1 p+1 p + 1 gives
d ε ( ∣ m ε ∣ 2 ) − p ∣ H d [ m ε ] ∣ ≤ η ~ d ε ( ∣ m ε ∣ 2 ) − p − 1 + C η ~ , p ∣ H d [ m ε ] ∣ p + 1 , d_\varepsilon(|m_\varepsilon|^2)^{-p}|H_d[m_\varepsilon]|
\le\widetilde{\eta}d_\varepsilon(|m_\varepsilon|^2)^{-p-1}+C_{\widetilde{\eta},p}|H_d[m_\varepsilon]|^{p+1}, d ε ( ∣ m ε ∣ 2 ) − p ∣ H d [ m ε ] ∣ ≤ η d ε ( ∣ m ε ∣ 2 ) − p − 1 + C η , p ∣ H d [ m ε ] ∣ p + 1 ,
and, since ∣ m ε ∣ 2 > 1 / 2 |m_\varepsilon|^2>1/2 ∣ m ε ∣ 2 > 1/2 , the first term is bounded by 2 η ~ ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 2\widetilde{\eta}|m_\varepsilon|^2d_\varepsilon(|m_\varepsilon|^2)^{-p-1} 2 η ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 . Finally, Lemma A.3 and ∣ m ε ∣ ≤ 1 |m_\varepsilon|\le1 ∣ m ε ∣ ≤ 1 imply
∥ H d [ m ε ] ∥ L p + 1 ( R 3 ) p + 1 ≤ C p ∥ m ε ∥ L p + 1 ( D ) p + 1 ≤ C p . \|H_d[m_\varepsilon]\|_{L^{p+1}(\mathbb{R}^3)}^{p+1}\le C_p\|m_\varepsilon\|_{L^{p+1}(\mathcal{D})}^{p+1}\le C_p. ∥ H d [ m ε ] ∥ L p + 1 ( R 3 ) p + 1 ≤ C p ∥ m ε ∥ L p + 1 ( D ) p + 1 ≤ C p .
Thus, (64) follows after renaming η \eta η .
Substituting (63) and (64) into (62) , and choosing η > 0 \eta>0 η > 0 sufficiently small compared with θ \theta θ , we obtain
d d t ∫ D Ψ p , ε ( ∣ m ε ∣ 2 ) d x + ∫ D d d s [ d ε ( s ) − p ] ∣ s = ∣ m ε ∣ 2 ∣ ∇ ∣ m ε ∣ 2 ∣ 2 d x + 2 ∫ D d ε ( ∣ m ε ∣ 2 ) − p ∣ ∇ m ε ∣ 2 d x + 2 θ ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x ≤ C p . (65) \frac{\mathrm{d}}{\mathrm{d}t}\int_{\mathcal{D}}\Psi_{p,\varepsilon}(|m_\varepsilon|^2)\,\mathrm{d}x
+\int_{\mathcal{D}}\left.\frac{\mathrm{d}}{\mathrm{d}s}\left[d_\varepsilon(s)^{-p}\right]\right|_{s=|m_\varepsilon|^2}|\nabla|m_\varepsilon|^2|^2\,\mathrm{d}x
+2\int_{\mathcal{D}}d_\varepsilon(|m_\varepsilon|^2)^{-p}|\nabla m_\varepsilon|^2\,\mathrm{d}x
+2\theta\int_{\mathcal{D}}|m_\varepsilon|^2d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\,\mathrm{d}x
\le C_p. \tag*{(65)} d t d ∫ D Ψ p , ε ( ∣ m ε ∣ 2 ) d x + ∫ D d s d [ d ε ( s ) − p ] s = ∣ m ε ∣ 2 ∣∇∣ m ε ∣ 2 ∣ 2 d x + 2 ∫ D d ε ( ∣ m ε ∣ 2 ) − p ∣∇ m ε ∣ 2 d x + 2 θ ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x ≤ C p . (65)
We now convert the last term on the left-hand side into the full singular weight. On { ∣ m ε ∣ 2 > 1 / 2 } \{|m_\varepsilon|^2>1/2\} { ∣ m ε ∣ 2 > 1/2 } ,
d ε ( ∣ m ε ∣ 2 ) − p − 1 ≤ 2 ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 , d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\le2|m_\varepsilon|^2d_\varepsilon(|m_\varepsilon|^2)^{-p-1}, d ε ( ∣ m ε ∣ 2 ) − p − 1 ≤ 2∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 ,
whereas on { ∣ m ε ∣ 2 ≤ 1 / 2 } \{|m_\varepsilon|^2 \le1/2\} { ∣ m ε ∣ 2 ≤ 1/2 } , we have
d ε ( ∣ m ε ∣ 2 ) − p − 1 ≤ ( 2 α q ( 1 / 2 ) ) − p − 1 . d_\varepsilon(|m_\varepsilon|^2)^{-p-1} \le(2\alpha q(1/\sqrt{2}))^{-p-1}. d ε ( ∣ m ε ∣ 2 ) − p − 1 ≤ ( 2 α q ( 1/ 2 ) ) − p − 1 .
Hence,
∫ D d ε ( ∣ m ε ∣ 2 ) − p − 1 d x ≤ C p + 2 ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x . \int_{\mathcal{D}} d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\,dx \le C_p+2\int_{\mathcal{D}} |m_\varepsilon|^2d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\,dx. ∫ D d ε ( ∣ m ε ∣ 2 ) − p − 1 d x ≤ C p + 2 ∫ D ∣ m ε ∣ 2 d ε ( ∣ m ε ∣ 2 ) − p − 1 d x .
Therefore, dropping the nonnegative gradient terms from (65) , we obtain
d d t ∫ D Ψ p , ε ( ∣ m ε ∣ 2 ) d x + c p ∫ D d ε ( ∣ m ε ∣ 2 ) − p − 1 d x ≤ C p . (66) \frac{d}{dt}\int_{\mathcal{D}}\Psi_{p,\varepsilon}(|m_\varepsilon|^2)\,dx+c_p\int_{\mathcal{D}}d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\,dx\le C_p. \tag*{(66)} d t d ∫ D Ψ p , ε ( ∣ m ε ∣ 2 ) d x + c p ∫ D d ε ( ∣ m ε ∣ 2 ) − p − 1 d x ≤ C p . (66)
It remains to express the above inequality as a scalar differential inequality. From (61) ,
∫ D Ψ p , ε ( ∣ m ε ∣ 2 ) d x ≤ C ∫ D d ε ( ∣ m ε ∣ 2 ) − p d x ≤ C ∣ D ∣ 1 p + 1 ( ∫ D d ε ( ∣ m ε ∣ 2 ) − p − 1 d x ) p p + 1 , \int_{\mathcal{D}}\Psi_{p,\varepsilon}(|m_\varepsilon|^2)\,dx\le C\int_{\mathcal{D}}d_\varepsilon(|m_\varepsilon|^2)^{-p}\,dx\le C|\mathcal{D}|^{\frac{1}{p+1}}\left(\int_{\mathcal{D}}d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\,dx\right)^{\frac{p}{p+1}}, ∫ D Ψ p , ε ( ∣ m ε ∣ 2 ) d x ≤ C ∫ D d ε ( ∣ m ε ∣ 2 ) − p d x ≤ C ∣ D ∣ p + 1 1 ( ∫ D d ε ( ∣ m ε ∣ 2 ) − p − 1 d x ) p + 1 p ,
where in the last step we used Hölder’s inequality. Consequently, we have
∫ D d ε ( ∣ m ε ∣ 2 ) − p − 1 d x ≥ c p ( ∫ D Ψ p , ε ( ∣ m ε ∣ 2 ) d x ) 1 + 1 p . (67) \int_{\mathcal{D}}d_\varepsilon(|m_\varepsilon|^2)^{-p-1}\,dx\ge c_p\left(\int_{\mathcal{D}}\Psi_{p,\varepsilon}(|m_\varepsilon|^2)\,dx\right)^{1+\frac{1}{p}}. \tag*{(67)} ∫ D d ε ( ∣ m ε ∣ 2 ) − p − 1 d x ≥ c p ( ∫ D Ψ p , ε ( ∣ m ε ∣ 2 ) d x ) 1 + p 1 . (67)
Thus, by writing
E ε , p ( t ) : = ∫ D Ψ p , ε ( ∣ m ε ( t ) ∣ 2 ) d x , E_{\varepsilon,p}(t):=\int_{\mathcal{D}}\Psi_{p,\varepsilon}(|m_\varepsilon(t)|^2)\,dx, E ε , p ( t ) := ∫ D Ψ p , ε ( ∣ m ε ( t ) ∣ 2 ) d x ,
and inserting (67) into (66) , we arrive at
E ε , p ′ ( t ) + c p [ E ε , p ( t ) ] 1 + 1 p ≤ C p . (68) E'_{\varepsilon,p}(t)+c_p[E_{\varepsilon,p}(t)]^{1+\frac{1}{p}}\le C_p. \tag*{(68)} E ε , p ′ ( t ) + c p [ E ε , p ( t ) ] 1 + p 1 ≤ C p . (68)
By a form of the Gronwall lemma with algebraic dissipation [31 Lemma 5.1] , we obtain
E ε , p ( t ) ≤ C p ( 1 + t − p ) , t > 0. (69) E_{\varepsilon,p}(t)\le C_p(1+t^{-p}),\qquad t>0. \tag*{(69)} E ε , p ( t ) ≤ C p ( 1 + t − p ) , t > 0. (69)
Finally, (61) and (60) imply
( ε + 1 − ∣ m ε ∣ 2 ) − p ≤ C p ( 1 + Ψ p , ε ( ∣ m ε ∣ 2 ) ) . (\varepsilon+1-|m_\varepsilon|^2)^{-p}\le C_p(1+\Psi_{p,\varepsilon}(|m_\varepsilon|^2)). ( ε + 1 − ∣ m ε ∣ 2 ) − p ≤ C p ( 1 + Ψ p , ε ( ∣ m ε ∣ 2 )) .
Integrating over D \mathcal{D} D and applying (69) then yields (59) , as claimed. □
The inverse-distance estimate obtained above controls the approach to saturation only in an integral sense. We next complement it with a positive-time spatial regularity estimate. In particular, choosing r > 3 / 2 r>3/2 r > 3/2 will later give the spatial continuity needed to turn the integral control of 1 − ∣ m ε ∣ 2 1-|m_\varepsilon|^2 1 − ∣ m ε ∣ 2 into a uniform pointwise separation from saturation.
Lemma 3.8. Let τ > 0 \tau>0 τ > 0 and r ∈ ( 1 , 2 ) r\in(1,2) r ∈ ( 1 , 2 ) . For every time interval I ⊂ [ τ , ∞ ) I\subset[\tau,\infty) I ⊂ [ τ , ∞ ) with ∣ I ∣ ≤ 1 |I|\le1 ∣ I ∣ ≤ 1 ,
∫ I ∥ m ε ( t ) ∥ W 2 , r ( D ) r d t ≤ C τ , r , (70) \int_I\|m_\varepsilon(t)\|_{W^{2,r}(\mathcal{D})}^r\,dt\le C_{\tau,r}, \tag*{(70)} ∫ I ∥ m ε ( t ) ∥ W 2 , r ( D ) r d t ≤ C τ , r , (70)
where C τ , r C_{\tau,r} C τ , r is independent of ε \varepsilon ε .
Proof. By (A.3), for a.e. ( t , x ) ∈ ( 0 , T ) × D (t,x)\in(0,T)\times\mathcal{D} ( t , x ) ∈ ( 0 , T ) × D ,
α Δ m ε ⋅ M ( m ε ) Δ m ε + ε ∣ Δ m ε ∣ 2 ≥ c ( q ( ∣ m ε ∣ ) + ε ) ∣ Δ m ε ∣ 2 , \alpha\Delta m_\varepsilon\cdot M(m_\varepsilon)\Delta m_\varepsilon+\varepsilon|\Delta m_\varepsilon|^2\ge c(q(|m_\varepsilon|)+\varepsilon)|\Delta m_\varepsilon|^2, α Δ m ε ⋅ M ( m ε ) Δ m ε + ε ∣Δ m ε ∣ 2 ≥ c ( q ( ∣ m ε ∣ ) + ε ) ∣Δ m ε ∣ 2 ,
where c > 0 c>0 c > 0 is independent of ε \varepsilon ε . Hence, by Hölder’s inequality with exponents 2 / r 2/r 2/ r and 2 / ( 2 − r ) 2/(2-r) 2/ ( 2 − r ) ,
∫ I × D ∣ Δ m ε ∣ r d x d t = ∫ I × D [ ( q ( ∣ m ε ∣ ) + ε ) 1 2 ∣ Δ m ε ∣ ] r ( q ( ∣ m ε ∣ ) + ε ) − r 2 d x d t ≤ C ( ∫ I × D ( q ( ∣ m ε ∣ ) + ε ) ∣ Δ m ε ∣ 2 d x d t ) r 2 ( ∫ I × D ( q ( ∣ m ε ∣ ) + ε ) − r 2 − r d x d t ) 1 − r 2 ≤ C τ , r . \begin{aligned}
\int_{I\times\mathcal{D}}|\Delta m_\varepsilon|^r\,dx\,dt
&=\int_{I\times\mathcal{D}}\left[(q(|m_\varepsilon|)+\varepsilon)^{\frac12}|\Delta m_\varepsilon|\right]^r(q(|m_\varepsilon|)+\varepsilon)^{-\frac r2}\,dx\,dt\\
&\le C\left(\int_{I\times\mathcal{D}}(q(|m_\varepsilon|)+\varepsilon)|\Delta m_\varepsilon|^2\,dx\,dt\right)^{\frac r2}
\left(\int_{I\times\mathcal{D}}(q(|m_\varepsilon|)+\varepsilon)^{-\frac r{2-r}}\,dx\,dt\right)^{1-\frac r2}\\
&\le C_{\tau,r}.
\end{aligned} ∫ I × D ∣Δ m ε ∣ r d x d t = ∫ I × D [ ( q ( ∣ m ε ∣ ) + ε ) 2 1 ∣Δ m ε ∣ ] r ( q ( ∣ m ε ∣ ) + ε ) − 2 r d x d t ≤ C ( ∫ I × D ( q ( ∣ m ε ∣ ) + ε ) ∣Δ m ε ∣ 2 d x d t ) 2 r ( ∫ I × D ( q ( ∣ m ε ∣ ) + ε ) − 2 − r r d x d t ) 1 − 2 r ≤ C τ , r .
Indeed, since ∣ I ∣ ≤ 1 |I| \le1 ∣ I ∣ ≤ 1 , the first factor is uniformly bounded by (54) . For the second factor, (A.1) gives $q(|m_\varepsilon|)+\varepsilon \simeq \varepsilon+1-|m_\varepsilon|^2$. On the other hand, Proposition 3.7` with p = r / ( 2 − r ) p=r/(2-r) p = r / ( 2 − r ) , yields for t ∈ I ⊂ [ τ , ∞ ) t \in I \subset[\tau,\infty) t ∈ I ⊂ [ τ , ∞ ) ,
∫ D ( q ( ∣ m ε ∣ ) + ε ) − r / ( 2 − r ) d x ≤ C τ , r . \int_{\mathcal{D}}\left(q(|m_\varepsilon|)+\varepsilon\right)^{-r/(2-r)}\,\mathrm{d}x \le C_{\tau,r}. ∫ D ( q ( ∣ m ε ∣ ) + ε ) − r / ( 2 − r ) d x ≤ C τ , r .
Thus, since ∣ I ∣ ≤ 1 |I| \le1 ∣ I ∣ ≤ 1 , we obtain
∫ I ∥ Δ m ε ( t ) ∥ L r ( D ) r d t ≤ C τ , r . \int_I \|\Delta m_\varepsilon(t)\|_{L^r(\mathcal{D})}^r\,\mathrm{d}t \le C_{\tau,r}. ∫ I ∥Δ m ε ( t ) ∥ L r ( D ) r d t ≤ C τ , r .
The standard W 2 , r W^{2,r} W 2 , r -regularity estimate for the Neumann Laplacian gives
∥ m ε ( t ) ∥ W 2 , r ( D ) ≤ C ( ∥ Δ m ε ( t ) ∥ L r ( D ) + ∥ m ε ( t ) ∥ L r ( D ) ) . \|m_\varepsilon(t)\|_{W^{2,r}(\mathcal{D})} \le C\left(\|\Delta m_\varepsilon(t)\|_{L^r(\mathcal{D})}+\|m_\varepsilon(t)\|_{L^r(\mathcal{D})}\right). ∥ m ε ( t ) ∥ W 2 , r ( D ) ≤ C ( ∥Δ m ε ( t ) ∥ L r ( D ) + ∥ m ε ( t ) ∥ L r ( D ) ) .
Since ∣ m ε ∣ ≤ 1 |m_\varepsilon| \le1 ∣ m ε ∣ ≤ 1 a.e. by Lemma 3.5, integration over I I I proves (70) . □
Uniform positive-time separation
We now upgrade the preceding positive-time estimates to a uniform pointwise separation from saturation. Given τ > 0 \tau>0 τ > 0 , the inverse-distance bound (59) and the W 2 , r W^{2,r} W 2 , r -regularity (70) first yield a time s τ ∈ ( 0 , τ ) s_\tau\in(0,\tau) s τ ∈ ( 0 , τ ) at which the solution is uniformly separated from ∣ m ∣ = 1 |m|=1 ∣ m ∣ = 1 . A scalar comparison argument, initiated at s τ s_\tau s τ , then propagates this separation forward in time and therefore yields a uniform gap on [ τ , ∞ ) [\tau,\infty) [ τ , ∞ ) .
Proposition 3.9. For every τ > 0 \tau>0 τ > 0 , there exist ρ τ ∈ ( 0 , 1 ) \rho_\tau\in(0,1) ρ τ ∈ ( 0 , 1 ) and ε τ ∈ ( 0 , 1 ] \varepsilon_\tau\in(0,1] ε τ ∈ ( 0 , 1 ] such that
ess sup x ∈ D ∣ m ε ( t , x ) ∣ ≤ ρ τ for every t ∈ [ τ , ∞ ) (71) \operatorname*{ess\,sup}_{x\in\mathcal{D}} |m_\varepsilon(t,x)| \le\rho_\tau\qquad\text{for every }t\in[\tau,\infty) \tag*{(71)} x ∈ D ess sup ∣ m ε ( t , x ) ∣ ≤ ρ τ for every t ∈ [ τ , ∞ ) (71) for every 0 < ε ≤ ε τ 0<\varepsilon\le\varepsilon_\tau 0 < ε ≤ ε τ .
Proof. We first obtain a uniform pointwise gap at a suitable time preceding each t ≥ τ t\ge\tau t ≥ τ , and then propagate this gap to t t t by a scalar comparison argument.
Fix r ∈ ( 3 2 , 2 ) r\in\left(\frac{3}{2},2\right) r ∈ ( 2 3 , 2 ) and μ : = 2 − 3 r > 0 \mu:=2-\frac{3}{r}>0 μ := 2 − r 3 > 0 , and choose p > 3 μ p>\frac{3}{\mu} p > μ 3 . By the Sobolev–Morrey embedding, we have W 2 , r ( D ) ↪ C 0 , μ ( D ‾ ) W^{2,r}(\mathcal{D})\hookrightarrow C^{0,\mu}(\overline{\mathcal{D}}) W 2 , r ( D ) ↪ C 0 , μ ( D ) . For arbitrary t ≥ τ t\ge\tau t ≥ τ , define the closed interval
I t : = [ max { τ / 4 , t − 1 } , t ] . I_t := [\max\{\tau/4,t-1\},t]. I t := [ max { τ /4 , t − 1 } , t ] .
Then I t ⊂ [ τ / 4 , ∞ ) I_t\subset[\tau/4,\infty) I t ⊂ [ τ /4 , ∞ ) and min { 1 , 3 τ 4 } ≤ ∣ I t ∣ ≤ 1 \min\{1,\frac{3\tau}{4}\}\le|I_t|\le1 min { 1 , 4 3 τ } ≤ ∣ I t ∣ ≤ 1 . By Lemma 3.8, applied with τ / 4 \tau/4 τ /4 , we have
∫ I t ∥ m ε ( s ) ∥ W 2 , r ( D ) r d s ≤ C τ / 4 , r . \int_{I_t}\|m_\varepsilon(s)\|_{W^{2,r}(\mathcal{D})}^r\,\mathrm{d}s \le C_{\tau/4,r}. ∫ I t ∥ m ε ( s ) ∥ W 2 , r ( D ) r d s ≤ C τ /4 , r .
Set ℓ τ : = min { 1 , 3 τ 4 } \ell_\tau:=\min\{1,\frac{3\tau}{4}\} ℓ τ := min { 1 , 4 3 τ } , so that ∣ I t ∣ ≥ ℓ τ > 0 |I_t|\ge\ell_\tau>0 ∣ I t ∣ ≥ ℓ τ > 0 . By Chebyshev’s inequality,
∣ { s ∈ I t : ∥ m ε ( s ) ∥ W 2 , r ( D ) r > 2 C τ / 4 , r ℓ τ } ∣ ≤ ℓ τ 2 C τ / 4 , r ∫ I t ∥ m ε ( s ) ∥ W 2 , r ( D ) r d s ≤ ℓ τ 2 . \left|\left\{s\in I_t:\|m_\varepsilon(s)\|_{W^{2,r}(\mathcal{D})}^r>\frac{2C_{\tau/4,r}}{\ell_\tau}\right\}\right|
\le\frac{\ell_\tau}{2C_{\tau/4,r}}\int_{I_t}\|m_\varepsilon(s)\|_{W^{2,r}(\mathcal{D})}^r\,\mathrm{d}s
\le\frac{\ell_\tau}{2}. { s ∈ I t : ∥ m ε ( s ) ∥ W 2 , r ( D ) r > ℓ τ 2 C τ /4 , r } ≤ 2 C τ /4 , r ℓ τ ∫ I t ∥ m ε ( s ) ∥ W 2 , r ( D ) r d s ≤ 2 ℓ τ .
Hence, the set
{ s ∈ I t : ∥ m ε ( s ) ∥ W 2 , r ( D ) ≤ ( 2 C τ / 4 , r ℓ τ ) 1 / r } \left\{s\in I_t:\|m_\varepsilon(s)\|_{W^{2,r}(\mathcal{D})}\le\left(\frac{2C_{\tau/4,r}}{\ell_\tau}\right)^{1/r}\right\} { s ∈ I t : ∥ m ε ( s ) ∥ W 2 , r ( D ) ≤ ( ℓ τ 2 C τ /4 , r ) 1/ r }
has measure at least ℓ τ / 2 \ell_\tau/2 ℓ τ /2 . On the other hand, Proposition 3.7 implies, for almost every s ∈ I t s\in I_t s ∈ I t ,
∫ D ( ε + 1 − ∣ m ε ( s ) ∣ 2 ) − p d x ≤ C p [ 1 + ( 4 τ ) p ] , \int_{\mathcal{D}}\left(\varepsilon+1-|m_\varepsilon(s)|^2\right)^{-p}\,\mathrm{d}x
\le C_p\left[1+\left(\frac{4}{\tau}\right)^p\right], ∫ D ( ε + 1 − ∣ m ε ( s ) ∣ 2 ) − p d x ≤ C p [ 1 + ( τ 4 ) p ] ,
since I t ⊂ [ τ / 4 , ∞ ) I_t\subset[\tau/4,\infty) I t ⊂ [ τ /4 , ∞ ) . The exceptional set where this estimate may fail has measure zero. We may therefore choose s ε , t ∈ I t s_{\varepsilon,t}\in I_t s ε , t ∈ I t such that
∥ m ε ( s ε , t ) ∥ W 2 , r ( D ) ≤ ( 2 C τ / 4 , r ℓ τ ) 1 / r and ∫ D ( ε + 1 − ∣ m ε ( s ε , t ) ∣ 2 ) − p d x ≤ C p [ 1 + ( 4 τ ) p ] . (72) \|m_\varepsilon(s_{\varepsilon,t})\|_{W^{2,r}(\mathcal{D})}\le\left(\frac{2C_{\tau/4,r}}{\ell_\tau}\right)^{1/r}
\qquad\text{and}\qquad
\int_{\mathcal{D}}\left(\varepsilon+1-|m_\varepsilon(s_{\varepsilon,t})|^2\right)^{-p}\,\mathrm{d}x
\le C_p\left[1+\left(\frac{4}{\tau}\right)^p\right]. \tag*{(72)} ∥ m ε ( s ε , t ) ∥ W 2 , r ( D ) ≤ ( ℓ τ 2 C τ /4 , r ) 1/ r and ∫ D ( ε + 1 − ∣ m ε ( s ε , t ) ∣ 2 ) − p d x ≤ C p [ 1 + ( τ 4 ) p ] . (72)
By the choice of s ε , t s_{\varepsilon,t} s ε , t and the Sobolev–Morrey embedding,
∥ m ε ( s ε , t ) ∥ C 0 , μ ( D ‾ ) ≤ C ( 2 C τ / 4 , r ℓ τ ) 1 / r = : K τ . \|m_\varepsilon(s_{\varepsilon,t})\|_{C^{0,\mu}(\overline{\mathcal D})}
\le C\left(\frac{2C_{\tau/4,r}}{\ell_\tau}\right)^{1/r}
=:K_\tau. ∥ m ε ( s ε , t ) ∥ C 0 , μ ( D ) ≤ C ( ℓ τ 2 C τ /4 , r ) 1/ r =: K τ .
where K τ K_\tau K τ is independent of ε \varepsilon ε and t t t . Since m ε ( s ε , t ) m_\varepsilon(s_{\varepsilon,t}) m ε ( s ε , t ) is continuous in x x x and ∣ m ε ( s ε , t ) ∣ ≤ 1 |m_\varepsilon(s_{\varepsilon,t})|\le1 ∣ m ε ( s ε , t ) ∣ ≤ 1 a.e., the latter inequality holds in fact everywhere on D ‾ \overline{\mathcal D} D . Therefore, for any x , y ∈ D ‾ x,y\in\overline{\mathcal D} x , y ∈ D ,
∣ ( ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 ) − ( ε + 1 − ∣ m ε ( s ε , t , y ) ∣ 2 ) ∣ ≤ 2 ∣ m ε ( s ε , t , x ) − m ε ( s ε , t , y ) ∣ \left|(\varepsilon+1-|m_\varepsilon(s_{\varepsilon,t},x)|^2)-(\varepsilon+1-|m_\varepsilon(s_{\varepsilon,t},y)|^2)\right|
\le2|m_\varepsilon(s_{\varepsilon,t},x)-m_\varepsilon(s_{\varepsilon,t},y)| ( ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 ) − ( ε + 1 − ∣ m ε ( s ε , t , y ) ∣ 2 ) ≤ 2∣ m ε ( s ε , t , x ) − m ε ( s ε , t , y ) ∣
≤ 2 K τ ∣ x − y ∣ μ . (73) \le2K_\tau|x-y|^\mu. \tag*{(73)} ≤ 2 K τ ∣ x − y ∣ μ . (73)
Moreover, 0 < ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 ≤ 2 0<\varepsilon+1-|m_\varepsilon(s_{\varepsilon,t},x)|^2\le2 0 < ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 ≤ 2 . Thus, the C 0 , μ ( D ‾ ) C^{0,\mu}(\overline{\mathcal D}) C 0 , μ ( D ) -norm of x ↦ ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 x\mapsto\varepsilon+1-|m_\varepsilon(s_{\varepsilon,t},x)|^2 x ↦ ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 is bounded independently of ε \varepsilon ε and t t t .
Now, set
δ ε , t : = min x ∈ D ‾ { ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 } > 0 , (74) \delta_{\varepsilon,t}:=\min_{x\in\overline{\mathcal D}}\{\varepsilon+1-|m_\varepsilon(s_{\varepsilon,t},x)|^2\}>0, \tag*{(74)} δ ε , t := x ∈ D min { ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 } > 0 , (74)
and let x ε , t ∈ D ‾ x_{\varepsilon,t}\in\overline{\mathcal D} x ε , t ∈ D be a point at which this minimum is attained. Since D \mathcal D D has sufficiently smooth boundary, it satisfies a uniform interior cone condition: there exist a height R D > 0 R_{\mathcal D}>0 R D > 0 and an opening angle ϑ D ∈ ( 0 , π / 2 ) \vartheta_{\mathcal D}\in(0,\pi/2) ϑ D ∈ ( 0 , π /2 ) , such that, for every x ∈ D ‾ x\in\overline{\mathcal D} x ∈ D and every R ∈ ( 0 , R D ] R\in(0,R_{\mathcal D}] R ∈ ( 0 , R D ] , the set D ∩ B R ( x ) \mathcal D\cap B_R(x) D ∩ B R ( x ) contains a cone of height R R R and opening angle ϑ D \vartheta_{\mathcal D} ϑ D . Consequently, there exists c D > 0 c_{\mathcal D}>0 c D > 0 , depending only on ϑ D \vartheta_{\mathcal D} ϑ D , such that
∣ D ∩ B R ( x ) ∣ ≥ c D R 3 for every x ∈ D ‾ and R ∈ ( 0 , R D ] . (75) |\mathcal D\cap B_R(x)|\ge c_{\mathcal D}R^3
\qquad\text{for every }x\in\overline{\mathcal D}\text{ and }R\in(0,R_{\mathcal D}]. \tag*{(75)} ∣ D ∩ B R ( x ) ∣ ≥ c D R 3 for every x ∈ D and R ∈ ( 0 , R D ] . (75)
From the Hölder estimate (73) ,
ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 ≤ δ ε , t + 2 K τ ∣ x − x ε , t ∣ μ for every x ∈ D ‾ . \varepsilon+1-|m_\varepsilon(s_{\varepsilon,t},x)|^2
\le\delta_{\varepsilon,t}+2K_\tau|x-x_{\varepsilon,t}|^\mu
\qquad\text{for every }x\in\overline{\mathcal D}. ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 ≤ δ ε , t + 2 K τ ∣ x − x ε , t ∣ μ for every x ∈ D .
Suppose first that δ ε , t ≤ 2 K τ R D μ \delta_{\varepsilon,t}\le2K_\tau R_{\mathcal D}^{\mu} δ ε , t ≤ 2 K τ R D μ . We then choose the radius R : = ( δ ε , t / ( 2 K τ ) ) 1 / μ R:=(\delta_{\varepsilon,t}/(2K_\tau))^{1/\mu} R := ( δ ε , t / ( 2 K τ ) ) 1/ μ . By the preceding assumption, 0 < R ≤ R D 0<R\le R_{\mathcal D} 0 < R ≤ R D , and thus, for every x ∈ D ∩ B R ( x ε , t ) x\in\mathcal D\cap B_R(x_{\varepsilon,t}) x ∈ D ∩ B R ( x ε , t ) ,
ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 ≤ δ ε , t + 2 K τ R μ = 2 δ ε , t . \varepsilon+1-|m_\varepsilon(s_{\varepsilon,t},x)|^2
\le\delta_{\varepsilon,t}+2K_\tau R^\mu
=2\delta_{\varepsilon,t}. ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 ≤ δ ε , t + 2 K τ R μ = 2 δ ε , t .
Using (75) and the inverse-distance estimate (72) , we obtain
C p ( 1 + ( 4 τ ) p ) ≥ ∫ D ( ε + 1 − ∣ m ε ( s ε , t ) ∣ 2 ) − p d x ≥ ∫ D ∩ B R ( x ε , t ) ( ε + 1 − ∣ m ε ( s ε , t ) ∣ 2 ) − p d x ≥ ( 2 δ ε , t ) − p ∣ D ∩ B R ( x ε , t ) ∣ ≥ c D 2 − p R 3 δ ε , t − p = c D 2 − p ( 2 K τ ) − 3 / μ δ ε , t 3 / μ − p . \begin{aligned}
C_p\left(1+\left(\frac{4}{\tau}\right)^p\right)
\ge\int_{\mathcal D}\left(\varepsilon+1-|m_\varepsilon(s_{\varepsilon,t})|^2\right)^{-p}\,dx \\
&\ge\int_{\mathcal D\cap B_R(x_{\varepsilon,t})}
\left(\varepsilon+1-|m_\varepsilon(s_{\varepsilon,t})|^2\right)^{-p}\,dx \\
&\ge(2\delta_{\varepsilon,t})^{-p}|\mathcal D\cap B_R(x_{\varepsilon,t})| \\
&\ge c_{\mathcal D}2^{-p}R^3\delta_{\varepsilon,t}^{-p} \\
&=c_{\mathcal D}2^{-p}(2K_\tau)^{-3/\mu}
\delta_{\varepsilon,t}^{3/\mu-p}.
\end{aligned} C p ( 1 + ( τ 4 ) p ) ≥ ∫ D ( ε + 1 − ∣ m ε ( s ε , t ) ∣ 2 ) − p d x ≥ ∫ D ∩ B R ( x ε , t ) ( ε + 1 − ∣ m ε ( s ε , t ) ∣ 2 ) − p d x ≥ ( 2 δ ε , t ) − p ∣ D ∩ B R ( x ε , t ) ∣ ≥ c D 2 − p R 3 δ ε , t − p = c D 2 − p ( 2 K τ ) − 3/ μ δ ε , t 3/ μ − p .
Since p − 3 / μ > 0 p-3/\mu>0 p − 3/ μ > 0 , we infer that
δ ε , t ≥ [ c D 2 − p ( 2 K τ ) − 3 / μ C p ( 1 + ( 4 / τ ) p ) ] 1 p − 3 / μ . \delta_{\varepsilon,t}\ge
\left[
\frac{c_{\mathcal D}2^{-p}(2K_\tau)^{-3/\mu}}
{C_p(1+(4/\tau)^p)}
\right]^{\frac{1}{p-3/\mu}}. δ ε , t ≥ [ C p ( 1 + ( 4/ τ ) p ) c D 2 − p ( 2 K τ ) − 3/ μ ] p − 3/ μ 1 .
If instead δ ε , t > 2 K τ R D μ \delta_{\varepsilon,t}>2K_\tau R_{\mathcal D}^{\mu} δ ε , t > 2 K τ R D μ , then a positive lower bound is immediate. Hence, in either case, we have δ ε , t ≥ δ τ > 0 \delta_{\varepsilon,t}\ge\delta_\tau>0 δ ε , t ≥ δ τ > 0 , where δ τ \delta_\tau δ τ is independent of ε \varepsilon ε and t t t . Finally, recall that from (74) ,
ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 ≥ δ τ ∀ x ∈ D ‾ . \varepsilon+1-|m_\varepsilon(s_{\varepsilon,t},x)|^2\ge\delta_\tau
\qquad\forall x\in\overline{\mathcal D}. ε + 1 − ∣ m ε ( s ε , t , x ) ∣ 2 ≥ δ τ ∀ x ∈ D .
Hence, whenever ε ≤ δ τ / 2 \varepsilon\le\delta_\tau/2 ε ≤ δ τ /2 ,
∣ m ε ( s ε , t , x ) ∣ 2 ≤ 1 + ε − δ τ ≤ 1 − δ τ 2 , ∀ x ∈ D ‾ . (76) |m_\varepsilon(s_{\varepsilon,t},x)|^2
\le1+\varepsilon-\delta_\tau
\le1-\frac{\delta_\tau}{2},
\qquad\forall x\in\overline{\mathcal D}. \tag*{(76)} ∣ m ε ( s ε , t , x ) ∣ 2 ≤ 1 + ε − δ τ ≤ 1 − 2 δ τ , ∀ x ∈ D . (76)
We now propagate (76) from s ε , t s_{\varepsilon,t} s ε , t to t t t . Define
b ε ( s ) : = ∫ 0 s 1 2 α q ( σ ) + ε d σ , s ∈ [ 0 , 1 ] . (77) b_\varepsilon(s) := \int_0^s \frac{1}{2\alpha q(\sqrt{\sigma})+\varepsilon}\,d\sigma,\qquad s\in[0,1]. \tag*{(77)} b ε ( s ) := ∫ 0 s 2 α q ( σ ) + ε 1 d σ , s ∈ [ 0 , 1 ] . (77)
Since q q q is decreasing and non-negative, we have
b ε ′ ( s ) = 1 2 α q ( s ) + ε > 0 , b ε ′ ′ ( s ) ≥ 0 , b_\varepsilon'(s)=\frac{1}{2\alpha q(\sqrt{s})+\varepsilon}>0,\qquad b_\varepsilon''(s)\geq0, b ε ′ ( s ) = 2 α q ( s ) + ε 1 > 0 , b ε ′′ ( s ) ≥ 0 ,
so b ε b_\varepsilon b ε is strictly increasing and convex on [ 0 , 1 ] [0,1] [ 0 , 1 ] .
Using (49) and the chain rule in temporal variable, we obtain, for a.e. s > 0 s>0 s > 0 ,
∂ s b ε ( ∣ m ε ∣ 2 ) − Δ ∣ m ε ∣ 2 = − 2 ∣ ∇ m ε ∣ 2 + 4 q ( ∣ m ε ∣ ) 2 α q ( ∣ m ε ∣ ) + ε m ε ⋅ ( B m ε + H d [ m ε ] ) − 4 θ ∣ m ε ∣ 2 2 α q ( ∣ m ε ∣ ) + ε ≤ 2 α ( ∥ B ∥ + ∥ H d [ m ε ] ( s ) ∥ L ∞ ( D ) ) . (78) \begin{aligned}
\partial_s b_\varepsilon(|m_\varepsilon|^2)-\Delta|m_\varepsilon|^2
={}&-2|\nabla m_\varepsilon|^2+\frac{4q(|m_\varepsilon|)}{2\alpha q(|m_\varepsilon|)+\varepsilon}m_\varepsilon\cdot\left(Bm_\varepsilon+H_d[m_\varepsilon]\right)\\
&-\frac{4\theta|m_\varepsilon|^2}{2\alpha q(|m_\varepsilon|)+\varepsilon}\\
\leq{}&\frac{2}{\alpha}\left(\|B\|+\|H_d[m_\varepsilon](s)\|_{L^\infty(\mathcal D)}\right). \tag*{(78)}
\end{aligned} ∂ s b ε ( ∣ m ε ∣ 2 ) − Δ∣ m ε ∣ 2 = ≤ − 2∣∇ m ε ∣ 2 + 2 α q ( ∣ m ε ∣ ) + ε 4 q ( ∣ m ε ∣ ) m ε ⋅ ( B m ε + H d [ m ε ] ) − 2 α q ( ∣ m ε ∣ ) + ε 4 θ ∣ m ε ∣ 2 α 2 ( ∥ B ∥ + ∥ H d [ m ε ] ( s ) ∥ L ∞ ( D ) ) . (78)
where we used the fact that ∣ m ε ∣ ≤ 1 |m_\varepsilon|\leq1 ∣ m ε ∣ ≤ 1 and
q ( ∣ m ε ∣ ) 2 α q ( ∣ m ε ∣ ) + ε ≤ 1 2 α , \frac{q(|m_\varepsilon|)}{2\alpha q(|m_\varepsilon|)+\varepsilon}\leq\frac{1}{2\alpha}, 2 α q ( ∣ m ε ∣ ) + ε q ( ∣ m ε ∣ ) ≤ 2 α 1 ,
while the first and last terms in the equality are non-positive. The purpose of (78) is to propagate the pointwise gap obtained at s ε , t s_{\varepsilon,t} s ε , t . We first show that the right-hand side has a uniformly bounded integral over [ s ε , t , t ] [s_{\varepsilon,t},t] [ s ε , t , t ] , then construct a scalar comparison function z ε z_\varepsilon z ε starting from 1 − δ τ / 2 1-\delta_\tau/2 1 − δ τ /2 whose evolution accounts for this forcing. A logarithmic lower bound for b ε b_\varepsilon b ε near 1 1 1 , uniform for sufficiently small ε \varepsilon ε , then prevents the comparison function from reaching the saturation boundary. A comparison argument then give
∣ m ε ( s ) ∣ 2 ≤ z ε ( s ) ≤ 1 − η τ on [ s ε , t , t ] . |m_\varepsilon(s)|^2\leq z_\varepsilon(s)\leq1-\eta_\tau\qquad\text{on }[s_{\varepsilon,t},t]. ∣ m ε ( s ) ∣ 2 ≤ z ε ( s ) ≤ 1 − η τ on [ s ε , t , t ] .
We now proceed with our aim. By Lemmas A.3 and 3.8, the embedding W 2 , r ( D ) ↪ L ∞ ( D ) W^{2,r}(\mathcal D)\hookrightarrow L^\infty(\mathcal D) W 2 , r ( D ) ↪ L ∞ ( D ) , and the fact that [ s ε , t , t ] ⊂ I t [s_{\varepsilon,t},t]\subset I_t [ s ε , t , t ] ⊂ I t has length at most one, we have
∫ s ε , t t ∥ H d [ m ε ] ( s ) ∥ L ∞ ( D ) d s ≤ C ∫ s ε , t t ∥ m ε ( s ) ∥ W 2 , r ( D ) d s ≤ C ∣ t − s ε , t ∣ 1 − 1 r ( ∫ s ε , t t ∥ m ε ( s ) ∥ W 2 , r ( D ) r d s ) 1 / r ≤ C ~ τ . (79) \begin{aligned}
\int_{s_{\varepsilon,t}}^t\|H_d[m_\varepsilon](s)\|_{L^\infty(\mathcal D)}\,ds
&\leq C\int_{s_{\varepsilon,t}}^t\|m_\varepsilon(s)\|_{W^{2,r}(\mathcal D)}\,ds\\
&\leq C|t-s_{\varepsilon,t}|^{1-\frac{1}{r}}\left(\int_{s_{\varepsilon,t}}^t\|m_\varepsilon(s)\|_{W^{2,r}(\mathcal D)}^r\,ds\right)^{1/r}\\
&\leq\widetilde C_\tau. \tag*{(79)}
\end{aligned} ∫ s ε , t t ∥ H d [ m ε ] ( s ) ∥ L ∞ ( D ) d s ≤ C ∫ s ε , t t ∥ m ε ( s ) ∥ W 2 , r ( D ) d s ≤ C ∣ t − s ε , t ∣ 1 − r 1 ( ∫ s ε , t t ∥ m ε ( s ) ∥ W 2 , r ( D ) r d s ) 1/ r ≤ C τ . (79)
Consequently,
2 α ∫ s ε , t t ( ∥ B ∥ + ∥ H d [ m ε ] ( s ) ∥ L ∞ ( D ) ) d s ≤ C ~ τ , (80) \frac{2}{\alpha}\int_{s_{\varepsilon,t}}^t\left(\|B\|+\|H_d[m_\varepsilon](s)\|_{L^\infty(\mathcal D)}\right)\,ds\leq\widetilde C_\tau, \tag*{(80)} α 2 ∫ s ε , t t ( ∥ B ∥ + ∥ H d [ m ε ] ( s ) ∥ L ∞ ( D ) ) d s ≤ C τ , (80)
where C ~ τ \widetilde C_\tau C τ is a highlighted constant independent of ε \varepsilon ε and t t t .
We next choose a uniform upper barrier below 1 1 1 . By (A.1),
2 α q ( s ) + ε ≤ C ( 1 − s + ε ) , ∀ s ∈ [ 0 , 1 ] . 2\alpha q(\sqrt{s})+\varepsilon\leq C(1-s+\varepsilon),\qquad\forall s\in[0,1]. 2 α q ( s ) + ε ≤ C ( 1 − s + ε ) , ∀ s ∈ [ 0 , 1 ] .
Hence, for η ∈ ( 0 , δ τ / 2 ) \eta\in(0,\delta_\tau/2) η ∈ ( 0 , δ τ /2 ) , we obtain from (77) that
b ε ( 1 − η ) − b ε ( 1 − δ τ 2 ) = ∫ 1 − δ τ / 2 1 − η 1 2 α q ( s ) + ε d s ≥ c ∫ 1 − δ τ / 2 1 − η 1 1 − s + ε d s = c log ( δ τ / 2 + ε η + ε ) . (81) \begin{aligned}
b_\varepsilon(1-\eta)-b_\varepsilon\left(1-\frac{\delta_\tau}{2}\right)
&=\int_{1-\delta_\tau/2}^{1-\eta}\frac{1}{2\alpha q(\sqrt{s})+\varepsilon}\,ds\\
&\geq c\int_{1-\delta_\tau/2}^{1-\eta}\frac{1}{1-s+\varepsilon}\,ds
=c\log\left(\frac{\delta_\tau/2+\varepsilon}{\eta+\varepsilon}\right). \tag*{(81)}
\end{aligned} b ε ( 1 − η ) − b ε ( 1 − 2 δ τ ) = ∫ 1 − δ τ /2 1 − η 2 α q ( s ) + ε 1 d s ≥ c ∫ 1 − δ τ /2 1 − η 1 − s + ε 1 d s = c log ( η + ε δ τ /2 + ε ) . (81)
Choose η τ ∈ ( 0 , δ τ / 2 ) \eta_{\tau} \in(0,\delta_{\tau}/2) η τ ∈ ( 0 , δ τ /2 ) sufficiently small such that
c log ( δ τ 4 η τ ) > C ~ τ . (82) c\log\left(\frac{\delta_{\tau}}{4\eta_{\tau}}\right)>\widetilde{C}_{\tau}. \tag*{(82)} c log ( 4 η τ δ τ ) > C τ . (82)
and then choose ε τ \varepsilon_{\tau} ε τ such that
0 < ε τ ≤ min { 1 , δ τ 2 , η τ } . 0<\varepsilon_{\tau}\leq\min\left\{1,\frac{\delta_{\tau}}{2},\eta_{\tau}\right\}. 0 < ε τ ≤ min { 1 , 2 δ τ , η τ } .
For every ε ∈ ( 0 , ε τ ] \varepsilon\in(0,\varepsilon_{\tau}] ε ∈ ( 0 , ε τ ] , we then have
δ τ / 2 + ε η τ + ε ≥ δ τ / 2 2 η τ = δ τ 4 η τ , \frac{\delta_{\tau}/2+\varepsilon}{\eta_{\tau}+\varepsilon}\geq\frac{\delta_{\tau}/2}{2\eta_{\tau}}=\frac{\delta_{\tau}}{4\eta_{\tau}}, η τ + ε δ τ /2 + ε ≥ 2 η τ δ τ /2 = 4 η τ δ τ ,
thus (81) and (82) gives
b ε ( 1 − η τ ) − b ε ( 1 − δ τ 2 ) > C ~ τ . (83) b_{\varepsilon}(1-\eta_{\tau})-b_{\varepsilon}\left(1-\frac{\delta_{\tau}}{2}\right)>\widetilde{C}_{\tau}. \tag*{(83)} b ε ( 1 − η τ ) − b ε ( 1 − 2 δ τ ) > C τ . (83)
Now, for s ∈ [ s ε , t , t ] s\in[s_{\varepsilon,t},t] s ∈ [ s ε , t , t ] , set
Y ε ( s ) : = b ε ( 1 − δ τ 2 ) + 2 α ∫ s ε , t s ( ∥ B ∥ + ∥ H d [ m ε ] ( σ ) ∥ L ∞ ( D ) ) d σ . (84) Y_{\varepsilon}(s):=b_{\varepsilon}\left(1-\frac{\delta_{\tau}}{2}\right)+\frac{2}{\alpha}\int_{s_{\varepsilon,t}}^{s}\left(\lVert B\rVert+\lVert H_{d}[m_{\varepsilon}](\sigma)\rVert_{L^{\infty}(\mathcal{D})}\right)\,\mathrm{d}\sigma. \tag*{(84)} Y ε ( s ) := b ε ( 1 − 2 δ τ ) + α 2 ∫ s ε , t s ( ∥ B ∥ + ∥ H d [ m ε ] ( σ ) ∥ L ∞ ( D ) ) d σ . (84)
For every s ∈ [ s ε , t , t ] s\in[s_{\varepsilon,t},t] s ∈ [ s ε , t , t ] , by (80) , (83) , and the fact that η τ ∈ ( 0 , δ τ / 2 ) \eta_{\tau}\in(0,\delta_{\tau}/2) η τ ∈ ( 0 , δ τ /2 ) ,
b ε ( 1 − δ τ 2 ) ≤ Y ε ( s ) < b ε ( 1 − η τ ) . b_{\varepsilon}\left(1-\frac{\delta_{\tau}}{2}\right)\leq Y_{\varepsilon}(s)<b_{\varepsilon}(1-\eta_{\tau}). b ε ( 1 − 2 δ τ ) ≤ Y ε ( s ) < b ε ( 1 − η τ ) .
Since b ε b_{\varepsilon} b ε is continuous and strictly increasing, the intermediate value theorem and strict monotonicity imply that there exists a unique z ε ( s ) ∈ [ 1 − δ τ / 2 , 1 − η τ ) z_{\varepsilon}(s)\in[1-\delta_{\tau}/2,1-\eta_{\tau}) z ε ( s ) ∈ [ 1 − δ τ /2 , 1 − η τ ) such that b ε ( z ε ( s ) ) = Y ε ( s ) b_{\varepsilon}(z_{\varepsilon}(s))=Y_{\varepsilon}(s) b ε ( z ε ( s )) = Y ε ( s ) . Equivalently, z ε ( s ) = b ε − 1 ( Y ε ( s ) ) z_{\varepsilon}(s)=b_{\varepsilon}^{-1}(Y_{\varepsilon}(s)) z ε ( s ) = b ε − 1 ( Y ε ( s )) .
Moreover, the map s ↦ Y ε ( s ) s\mapsto Y_{\varepsilon}(s) s ↦ Y ε ( s ) is absolutely continuous. Since b ε − 1 b_{\varepsilon}^{-1} b ε − 1 is continuously differentiable on the relevant interval, z ε z_{\varepsilon} z ε is absolutely continuous and
∂ s b ε ( z ε ( s ) ) = ∂ s Y ε ( s ) = 2 α ( ∥ B ∥ + ∥ H d [ m ε ] ( s ) ∥ L ∞ ( D ) ) \partial_{s}b_{\varepsilon}(z_{\varepsilon}(s))=\partial_{s}Y_{\varepsilon}(s)=\frac{2}{\alpha}\left(\lVert B\rVert+\lVert H_{d}[m_{\varepsilon}](s)\rVert_{L^{\infty}(\mathcal{D})}\right) ∂ s b ε ( z ε ( s )) = ∂ s Y ε ( s ) = α 2 ( ∥ B ∥ + ∥ H d [ m ε ] ( s ) ∥ L ∞ ( D ) )
for a.e. s ∈ [ s ε , t , t ] s\in[s_{\varepsilon,t},t] s ∈ [ s ε , t , t ] . Subtracting this identity from (78) gives
∂ s [ b ε ( ∣ m ε ∣ 2 ) − b ε ( z ε ) ] − Δ ∣ m ε ∣ 2 ≤ 0 \partial_{s}\left[b_{\varepsilon}(\lvert m_{\varepsilon}\rvert^{2})-b_{\varepsilon}(z_{\varepsilon})\right]-\Delta\lvert m_{\varepsilon}\rvert^{2}\leq0 ∂ s [ b ε (∣ m ε ∣ 2 ) − b ε ( z ε ) ] − Δ ∣ m ε ∣ 2 ≤ 0
in the distributional sense on ( s ε , t , t ) × D (s_{\varepsilon,t},t)\times\mathcal{D} ( s ε , t , t ) × D ; that is,
− ∫ s ε , t t ⟨ b ε ( ∣ m ε ∣ 2 ) − b ε ( z ε ) , ∂ s φ ⟩ d s + ∫ s ε , t t ⟨ ∇ ∣ m ε ∣ 2 , ∇ φ ⟩ d s ≤ 0 (85) -\int_{s_{\varepsilon,t}}^{t}\left\langle b_{\varepsilon}(\lvert m_{\varepsilon}\rvert^{2})-b_{\varepsilon}(z_{\varepsilon}),\partial_{s}\varphi\right\rangle\,\mathrm{d}s+\int_{s_{\varepsilon,t}}^{t}\left\langle\nabla\lvert m_{\varepsilon}\rvert^{2},\nabla\varphi\right\rangle\,\mathrm{d}s\leq0 \tag*{(85)} − ∫ s ε , t t ⟨ b ε (∣ m ε ∣ 2 ) − b ε ( z ε ) , ∂ s φ ⟩ d s + ∫ s ε , t t ⟨ ∇ ∣ m ε ∣ 2 , ∇ φ ⟩ d s ≤ 0 (85)
for every nonnegative φ ∈ C c ∞ ( ( s ε , t , t ) ; H 1 ( D ) ) \varphi\in C_{c}^{\infty}((s_{\varepsilon,t},t);H^{1}(\mathcal{D})) φ ∈ C c ∞ (( s ε , t , t ) ; H 1 ( D )) .
Set F ε : = b ε ( ∣ m ε ∣ 2 ) − b ε ( z ε ) F_{\varepsilon}:=b_{\varepsilon}(\lvert m_{\varepsilon}\rvert^{2})-b_{\varepsilon}(z_{\varepsilon}) F ε := b ε (∣ m ε ∣ 2 ) − b ε ( z ε ) . For fixed ε > 0 \varepsilon>0 ε > 0 , the boundedness of b ε ′ b_{\varepsilon}' b ε ′ on [ 0 , 1 ] [0,1] [ 0 , 1 ] , together with the regularity of ∣ m ε ∣ 2 \lvert m_{\varepsilon}\rvert^{2} ∣ m ε ∣ 2 and the absolute continuity of z ε z_{\varepsilon} z ε , gives
F ε ∈ W 1 , r ( s ε , t , t ; L 2 ( D ) ) ∩ L 2 ( s ε , t , t ; H 1 ( D ) ) . F_{\varepsilon}\in W^{1,r}(s_{\varepsilon,t},t;L^{2}(\mathcal{D}))\cap L^{2}(s_{\varepsilon,t},t;H^{1}(\mathcal{D})). F ε ∈ W 1 , r ( s ε , t , t ; L 2 ( D )) ∩ L 2 ( s ε , t , t ; H 1 ( D )) .
Hence, ( F ε ) + (F_{\varepsilon})_{+} ( F ε ) + has the same regularity and may be used, after multiplication by a standard nonnegative time cut-off, as a test function in (85) . Using the Sobolev chain rule, the monotonicity of b ε b_{\varepsilon} b ε , and the fact that z ε z_{\varepsilon} z ε is independent of x x x , we obtain for every s ∈ ( s ε , t , t ) s\in(s_{\varepsilon,t},t) s ∈ ( s ε , t , t )
1 2 ∥ ( F ε ( s ) ) + ∥ L 2 ( D ) 2 − 1 2 ∥ ( F ε ( s ε , t ) ) + ∥ L 2 ( D ) 2 + ∫ s ε , t s ∫ { F ε > 0 } b ε ′ ( ∣ m ε ∣ 2 ) ∣ ∇ ∣ m ε ∣ 2 ∣ 2 d x d s ≤ 0. (86) \frac{1}{2}\lVert(F_{\varepsilon}(s))_{+}\rVert_{L^{2}(\mathcal{D})}^{2}-\frac{1}{2}\lVert(F_{\varepsilon}(s_{\varepsilon,t}))_{+}\rVert_{L^{2}(\mathcal{D})}^{2}+\int_{s_{\varepsilon,t}}^{s}\int_{\{F_{\varepsilon}>0\}}b_{\varepsilon}'(\lvert m_{\varepsilon}\rvert^{2})\lvert\nabla\lvert m_{\varepsilon}\rvert^{2}\rvert^{2}\,\mathrm{d}x\,\mathrm{d}s\leq0. \tag*{(86)} 2 1 ∥( F ε ( s ) ) + ∥ L 2 ( D ) 2 − 2 1 ∥( F ε ( s ε , t ) ) + ∥ L 2 ( D ) 2 + ∫ s ε , t s ∫ { F ε > 0 } b ε ′ (∣ m ε ∣ 2 ) ∣ ∇ ∣ m ε ∣ 2 ∣ 2 d x d s ≤ 0. (86)
where we used the fact that ∇ ( F ε ) + = 1 { F ε > 0 } b ε ′ ( ∣ m ε ∣ 2 ) ∇ ∣ m ε ∣ 2 \nabla(F_\varepsilon)_+ = \mathbf{1}_{\{F_\varepsilon>0\}}b_\varepsilon'(|m_\varepsilon|^2)\nabla|m_\varepsilon|^2 ∇ ( F ε ) + = 1 { F ε > 0 } b ε ′ ( ∣ m ε ∣ 2 ) ∇∣ m ε ∣ 2 and that z ε z_\varepsilon z ε is independent of x x x . Moreover, since b ε b_\varepsilon b ε is strictly increasing, { F ε > 0 } \{F_\varepsilon>0\} { F ε > 0 } is equivalent with { ∣ m ε ∣ 2 > z ε } \{|m_\varepsilon|^2>z_\varepsilon\} { ∣ m ε ∣ 2 > z ε } . However, by (76) , using the injectivity of b ε b_\varepsilon b ε and the fact that b ε ( z ε ( s ε , t ) ) = Y ε ( s ε , t ) = b ε ( 1 − δ τ / 2 ) b_\varepsilon(z_\varepsilon(s_{\varepsilon,t}))=Y_\varepsilon(s_{\varepsilon,t})=b_\varepsilon(1-\delta_\tau/2) b ε ( z ε ( s ε , t )) = Y ε ( s ε , t ) = b ε ( 1 − δ τ /2 ) , we infer
∣ m ε ( s ε , t , x ) ∣ 2 ≤ 1 − δ τ 2 = z ε ( s ε , t ) for every x ∈ D , |m_\varepsilon(s_{\varepsilon,t},x)|^2\le1-\frac{\delta_\tau}{2}=z_\varepsilon(s_{\varepsilon,t})\qquad\text{for every }x\in\mathcal{D}, ∣ m ε ( s ε , t , x ) ∣ 2 ≤ 1 − 2 δ τ = z ε ( s ε , t ) for every x ∈ D ,
and hence ( F ε ( s ε , t ) ) + = 0 (F_\varepsilon(s_{\varepsilon,t}))_+=0 ( F ε ( s ε , t ) ) + = 0 . We thus conclude from (86) that
∥ ( F ε ( s ) ) + ∥ L 2 ( D ) 2 = 0 for every s ∈ [ s ε , t , t ] . \|(F_\varepsilon(s))_+\|_{L^2(\mathcal{D})}^2=0\qquad\text{for every }s\in[s_{\varepsilon,t},t]. ∥ ( F ε ( s ) ) + ∥ L 2 ( D ) 2 = 0 for every s ∈ [ s ε , t , t ] .
Therefore, b ε ( ∣ m ε ( s , x ) ∣ 2 ) ≤ b ε ( z ε ( s ) ) b_\varepsilon(|m_\varepsilon(s,x)|^2)\le b_\varepsilon(z_\varepsilon(s)) b ε ( ∣ m ε ( s , x ) ∣ 2 ) ≤ b ε ( z ε ( s )) for a.e. x ∈ D x\in\mathcal{D} x ∈ D . Since b ε b_\varepsilon b ε is strictly increasing, we obtain
∣ m ε ( s , x ) ∣ 2 ≤ z ε ( s ) < 1 − η τ |m_\varepsilon(s,x)|^2\le z_\varepsilon(s)<1-\eta_\tau ∣ m ε ( s , x ) ∣ 2 ≤ z ε ( s ) < 1 − η τ
for a.e. x ∈ D x\in\mathcal{D} x ∈ D and every s ∈ [ s ε , t , t ] s\in[s_{\varepsilon,t},t] s ∈ [ s ε , t , t ] . Since t ≥ τ t\ge\tau t ≥ τ was arbitrary,
ess sup x ∈ D ∣ m ε ( t , x ) ∣ ≤ ρ τ : = 1 − η τ < 1 for every t ≥ τ . \operatorname*{ess\,sup}_{x\in\mathcal{D}}|m_\varepsilon(t,x)|\le\rho_\tau:=\sqrt{1-\eta_\tau}<1\qquad\text{for every }t\ge\tau. x ∈ D ess sup ∣ m ε ( t , x ) ∣ ≤ ρ τ := 1 − η τ < 1 for every t ≥ τ .
This completes the proof of (71) .
Passage to the limit
Proof of Theorem 3.2. We first consolidate the compactness properties of m ε m_\varepsilon m ε and identify the limit equation, then prove properties (i)–(iv).
Fix T > 0 T>0 T > 0 . By (54) , we have
sup t ∈ [ 0 , T ] ∥ m ε ( t ) ∥ H 1 ( D ) 2 ≤ C , \sup_{t\in[0,T]}\|m_\varepsilon(t)\|_{H^1(\mathcal{D})}^2\le C, t ∈ [ 0 , T ] sup ∥ m ε ( t ) ∥ H 1 ( D ) 2 ≤ C ,
while, covering [ 0 , T ] [0,T] [ 0 , T ] by finitely many intervals of length at most one,
∫ 0 T ∥ ∂ t m ε ( s ) ∥ L 2 ( D ) 2 d s + ε ∫ 0 T ∥ Δ m ε ( s ) ∥ L 2 ( D ) 2 d s ≤ C T , \int_0^T\|\partial_t m_\varepsilon(s)\|_{L^2(\mathcal{D})}^2\,\mathrm{d}s+\varepsilon\int_0^T\|\Delta m_\varepsilon(s)\|_{L^2(\mathcal{D})}^2\,\mathrm{d}s\le C_T, ∫ 0 T ∥ ∂ t m ε ( s ) ∥ L 2 ( D ) 2 d s + ε ∫ 0 T ∥Δ m ε ( s ) ∥ L 2 ( D ) 2 d s ≤ C T ,
where C T C_T C T is independent of ε \varepsilon ε . Hence, after passing to a subsequence,
m ε ⇀ ∗ m in L ∞ ( 0 , T ; H 1 ( D ) ) , m_\varepsilon\overset{*}{\rightharpoonup}m\quad\text{in }L^\infty(0,T;H^1(\mathcal{D})), m ε ⇀ ∗ m in L ∞ ( 0 , T ; H 1 ( D )) ,
∂ t m ε ⇀ ∂ t m in L 2 ( 0 , T ; L 2 ( D ) ) . \partial_t m_\varepsilon\rightharpoonup\partial_t m\quad\text{in }L^2(0,T;L^2(\mathcal{D})). ∂ t m ε ⇀ ∂ t m in L 2 ( 0 , T ; L 2 ( D )) .
Moreover, for every s , t ∈ [ 0 , T ] s,t\in[0,T] s , t ∈ [ 0 , T ] ,
∥ m ε ( t ) − m ε ( s ) ∥ L 2 ( D ) ≤ ∣ t − s ∣ 1 2 ∥ ∂ t m ε ∥ L 2 ( 0 , T ; L 2 ( D ) ) ≤ C T ∣ t − s ∣ 1 2 . \|m_\varepsilon(t)-m_\varepsilon(s)\|_{L^2(\mathcal{D})}\le|t-s|^{\frac12}\|\partial_t m_\varepsilon\|_{L^2(0,T;L^2(\mathcal{D}))}\le C_T|t-s|^{\frac12}. ∥ m ε ( t ) − m ε ( s ) ∥ L 2 ( D ) ≤ ∣ t − s ∣ 2 1 ∥ ∂ t m ε ∥ L 2 ( 0 , T ; L 2 ( D )) ≤ C T ∣ t − s ∣ 2 1 .
Since H 1 ( D ) ⋐ L 2 ( D ) H^1(\mathcal{D})\Subset L^2(\mathcal{D}) H 1 ( D ) ⋐ L 2 ( D ) , the Arzelà–Ascoli theorem yields
m ε → m strongly in C ( [ 0 , T ] ; L 2 ( D ) ) . (87) m_\varepsilon\to m\qquad\text{strongly in }C([0,T];L^2(\mathcal{D})). \tag*{(87)} m ε → m strongly in C ([ 0 , T ] ; L 2 ( D )) . (87)
Since T > 0 T>0 T > 0 was arbitrary, by a standard diagonal argument over T ∈ N T\in\mathbb{N} T ∈ N , we may choose the above subsequence independently of T T T and obtain a single limit m m m on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . In what follows, we work with this subsequence without further relabelling. In particular, m ε → m m_\varepsilon\to m m ε → m strongly in L 2 ( ( 0 , T ) × D ) L^2((0,T)\times\mathcal{D}) L 2 (( 0 , T ) × D ) , which implies, after passing to a further subsequence,
m ε ( t , x ) → m ( t , x ) for a.e. ( t , x ) ∈ ( 0 , T ) × D . (88) m_\varepsilon(t,x)\to m(t,x)\qquad\text{for a.e. }(t,x)\in(0,T)\times\mathcal{D}. \tag*{(88)} m ε ( t , x ) → m ( t , x ) for a.e. ( t , x ) ∈ ( 0 , T ) × D . (88)
Since m ε ( 0 ) = m 0 m_\varepsilon(0)=m_0 m ε ( 0 ) = m 0 , by (87) we have m ( 0 ) = m 0 m(0)=m_0 m ( 0 ) = m 0 in L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) . Furthermore, since ∣ m ε ( t ) ∣ ≤ 1 |m_\varepsilon(t)|\le1 ∣ m ε ( t ) ∣ ≤ 1 a.e. in D \mathcal{D} D for every t ≥ 0 t\ge0 t ≥ 0 , the 1-Lipschitz continuity of v ↦ ( ∣ v ∣ − 1 ) + v\mapsto(|v|-1)_+ v ↦ ( ∣ v ∣ − 1 ) + gives
∥ ( ∣ m ( t ) ∣ − 1 ) + ∥ L 2 ( D ) ≤ ∥ m ( t ) − m ε ( t ) ∥ L 2 ( D ) → 0. \|(|m(t)|-1)_+\|_{L^2(\mathcal{D})}\le\|m(t)-m_\varepsilon(t)\|_{L^2(\mathcal{D})}\to0. ∥ ( ∣ m ( t ) ∣ − 1 ) + ∥ L 2 ( D ) ≤ ∥ m ( t ) − m ε ( t ) ∥ L 2 ( D ) → 0.
Thus,
∣ m ( t , x ) ∣ ≤ 1 for a.e. x ∈ D and every t ∈ [ 0 , T ] . (89) |m(t,x)|\le1\qquad\text{for a.e. }x\in\mathcal{D}\text{ and every }t\in[0,T]. \tag*{(89)} ∣ m ( t , x ) ∣ ≤ 1 for a.e. x ∈ D and every t ∈ [ 0 , T ] . (89)
We next obtain the stronger compactness needed to identify the thermodynamic equation. Fix τ ∈ ( 0 , T ) \tau\in(0,T) τ ∈ ( 0 , T ) . By Proposition 3.9 , for all sufficiently small ε \varepsilon ε ,
∣ m ε ( t , x ) ∣ ≤ ρ τ < 1 for a.e. x ∈ D and every t ∈ [ τ , T ] . (90) |m_\varepsilon(t,x)| \le\rho_\tau< 1 \qquad\text{for a.e. } x \in\mathcal{D} \text{ and every } t \in[\tau,T]. \tag*{(90)} ∣ m ε ( t , x ) ∣ ≤ ρ τ < 1 for a.e. x ∈ D and every t ∈ [ τ , T ] . (90)
Since q q q is decreasing, q ( ∣ m ε ∣ ) ≥ q ( ρ τ ) > 0 q(|m_\varepsilon|) \ge q(\rho_\tau)>0 q ( ∣ m ε ∣ ) ≥ q ( ρ τ ) > 0 . Therefore, by (A.3) and (54) ,
∫ τ T ∥ Δ m ε ( s ) ∥ L 2 ( D ) 2 d s ≤ 1 2 q ( ρ τ ) ∫ τ T ∫ D Δ m ε ⋅ M ( m ε ) Δ m ε d x d s ≤ C τ , T . (91) \int_\tau^T \|\Delta m_\varepsilon(s)\|_{L^2(\mathcal{D})}^2\,ds \le\frac{1}{2q(\rho_\tau)} \int_\tau^T \int_{\mathcal{D}} \Delta m_\varepsilon\cdot\mathrm{M}(m_\varepsilon)\Delta m_\varepsilon\,dx\,ds \le C_{\tau,T}. \tag*{(91)} ∫ τ T ∥Δ m ε ( s ) ∥ L 2 ( D ) 2 d s ≤ 2 q ( ρ τ ) 1 ∫ τ T ∫ D Δ m ε ⋅ M ( m ε ) Δ m ε d x d s ≤ C τ , T . (91)
The Neumann elliptic estimate and (90) imply
∥ m ε ∥ L 2 ( τ , T ; H 2 ( D ) ) ≤ C τ , T . (92) \|m_\varepsilon\|_{L^2(\tau,T;H^2(\mathcal{D}))} \le C_{\tau,T}. \tag*{(92)} ∥ m ε ∥ L 2 ( τ , T ; H 2 ( D )) ≤ C τ , T . (92)
Since the limit m m m has already been fixed by (87) , every weak accumulation point in L 2 ( τ , T ; H 2 ( D ) ) L^2(\tau,T;H^2(\mathcal{D})) L 2 ( τ , T ; H 2 ( D )) necessarily coincides with m m m . Hence,
m ε ⇀ m in L 2 ( τ , T ; H 2 ( D ) ) . m_\varepsilon\rightharpoonup m \qquad\text{in } L^2(\tau,T;H^2(\mathcal{D})). m ε ⇀ m in L 2 ( τ , T ; H 2 ( D )) .
Moreover, (92) , the uniform estimate for ∂ t m ε \partial_t m_\varepsilon ∂ t m ε , and the Aubin–Lions lemma gives relative compactness in L 2 ( τ , T ; H 1 ( D ) ) L^2(\tau,T;H^1(\mathcal{D})) L 2 ( τ , T ; H 1 ( D )) . Again, the limit is uniquely identified by (87) , so that
m ε → m strongly in L 2 ( τ , T ; H 1 ( D ) ) . (93) m_\varepsilon\to m \qquad\text{strongly in } L^2(\tau,T;H^1(\mathcal{D})). \tag*{(93)} m ε → m strongly in L 2 ( τ , T ; H 1 ( D )) . (93)
We now collect the remaining convergences needed to pass to the nonlinear equation. From (87) and the L 2 L^2 L 2 -boundedness of the stray-field operator,
H d [ m ε ] → H d [ m ] strongly in C ( [ 0 , T ] ; L 2 ( R 3 ) ) . H_{\mathrm d}[m_\varepsilon] \to H_{\mathrm d}[m] \qquad\text{strongly in } C([0,T];L^2(\mathbb{R}^3)). H d [ m ε ] → H d [ m ] strongly in C ([ 0 , T ] ; L 2 ( R 3 )) .
Furthermore, by (54) ,
∥ ε Δ m ε ∥ L 2 ( 0 , T ; L 2 ( D ) ) = ε 2 ∫ 0 T ∥ Δ m ε ∥ L 2 ( D ) 2 d t ≤ ε [ ε ∫ 0 T ∥ Δ m ε ∥ L 2 ( D ) 2 d t ] → 0. \|\varepsilon\Delta m_\varepsilon\|_{L^2(0,T;L^2(\mathcal{D}))} = \varepsilon^2 \int_0^T \|\Delta m_\varepsilon\|_{L^2(\mathcal{D})}^2\,dt \le\varepsilon\left[\varepsilon\int_0^T \|\Delta m_\varepsilon\|_{L^2(\mathcal{D})}^2\,dt\right] \to0. ∥ ε Δ m ε ∥ L 2 ( 0 , T ; L 2 ( D )) = ε 2 ∫ 0 T ∥Δ m ε ∥ L 2 ( D ) 2 d t ≤ ε [ ε ∫ 0 T ∥Δ m ε ∥ L 2 ( D ) 2 d t ] → 0.
By (90) and (88) , we also have ∣ m ∣ ≤ ρ τ |m| \le\rho_\tau ∣ m ∣ ≤ ρ τ a.e. on ( τ , T ) × D (\tau,T) \times\mathcal{D} ( τ , T ) × D . Thus G ( m ε ) G(m_\varepsilon) G ( m ε ) , G ( m ) G(m) G ( m ) , h ( m ε ) h(m_\varepsilon) h ( m ε ) , and h ( m ) h(m) h ( m ) are well defined there. Since G G G is Lipschitz on { ∣ v ∣ ≤ ρ τ } \{|v| \le\rho_\tau\} { ∣ v ∣ ≤ ρ τ } ,
G ( m ε ) → G ( m ) strongly in L 2 ( τ , T ; L 2 ( D ) ) . G(m_\varepsilon) \to G(m) \qquad\text{strongly in } L^2(\tau,T;L^2(\mathcal{D})). G ( m ε ) → G ( m ) strongly in L 2 ( τ , T ; L 2 ( D )) .
Combining these convergences with the weak convergence of Δ m ε \Delta m_\varepsilon Δ m ε gives
h ( m ε ) ⇀ h ( m ) in L 2 ( τ , T ; L 2 ( D ) ) . (94) h(m_\varepsilon) \rightharpoonup h(m) \qquad\text{in } L^2(\tau,T;L^2(\mathcal{D})). \tag*{(94)} h ( m ε ) ⇀ h ( m ) in L 2 ( τ , T ; L 2 ( D )) . (94)
Let φ ∈ L 2 ( ( τ , T ) × D ; R 3 ) \varphi\in L^2((\tau,T)\times\mathcal{D};\mathbb{R}^3) φ ∈ L 2 (( τ , T ) × D ; R 3 ) . By (88) , (90) , and the dominated convergence theorem, we have
φ × m ε → φ × m strongly in L 2 ( ( τ , T ) × D ) , \varphi\times m_\varepsilon\to\varphi\times m \qquad\text{strongly in } L^2((\tau,T)\times\mathcal{D}), φ × m ε → φ × m strongly in L 2 (( τ , T ) × D ) ,
M ( m ε ) φ → M ( m ) φ strongly in L 2 ( ( τ , T ) × D ) , \mathrm{M}(m_\varepsilon)\varphi\to\mathrm{M}(m)\varphi\qquad\text{strongly in } L^2((\tau,T)\times\mathcal{D}), M ( m ε ) φ → M ( m ) φ strongly in L 2 (( τ , T ) × D ) ,
noting that M \mathrm{M} M is continuous and bounded on the closed unit ball. Utilising the symmetry of M ( m ε ) \mathrm{M}(m_\varepsilon) M ( m ε ) , we then have
∫ τ T ⟨ ∂ t m ε , φ ⟩ d s = γ ∫ τ T ⟨ h ( m ε ) , φ × m ε ⟩ d s + ∫ τ T ⟨ h ( m ε ) , M ( m ε ) φ ⟩ d s + ∫ τ T ⟨ ε Δ m ε , φ ⟩ d s \int_\tau^T \langle\partial_t m_\varepsilon,\varphi\rangle\,ds
= \gamma\int_\tau^T \langle h(m_\varepsilon),\varphi\times m_\varepsilon\rangle\,ds
+ \int_\tau^T \langle h(m_\varepsilon),\mathrm{M}(m_\varepsilon)\varphi\rangle\,ds
+ \int_\tau^T \langle\varepsilon\Delta m_\varepsilon,\varphi\rangle\,ds ∫ τ T ⟨ ∂ t m ε , φ ⟩ d s = γ ∫ τ T ⟨ h ( m ε ) , φ × m ε ⟩ d s + ∫ τ T ⟨ h ( m ε ) , M ( m ε ) φ ⟩ d s + ∫ τ T ⟨ ε Δ m ε , φ ⟩ d s
⟶ γ ∫ τ T ⟨ h ( m ) , φ × m ⟩ d s + ∫ τ T ⟨ h ( m ) , M ( m ) φ ⟩ d s . \longrightarrow\gamma\int_\tau^T \langle h(m),\varphi\times m\rangle\,ds
+ \int_\tau^T \langle h(m),\mathrm{M}(m)\varphi\rangle\,ds. ⟶ γ ∫ τ T ⟨ h ( m ) , φ × m ⟩ d s + ∫ τ T ⟨ h ( m ) , M ( m ) φ ⟩ d s .
Hence, we infer that
∂ t m = γ m × h ( m ) + M ( m ) h ( m ) in L 2 ( τ , T ; L 2 ( D ) ) . (95) \partial_t m = \gamma m \times h(m) + \mathrm{M}(m)h(m)
\qquad\text{in } L^2(\tau,T;L^2(\mathcal{D})). \tag*{(95)} ∂ t m = γ m × h ( m ) + M ( m ) h ( m ) in L 2 ( τ , T ; L 2 ( D )) . (95)
To recover the weak formulation of the effective field, let ψ ∈ H 1 ( D ) \psi\in H^1(\mathcal{D}) ψ ∈ H 1 ( D ) and ζ ∈ C ∞ ( τ , T ) \zeta\in C^\infty(\tau,T) ζ ∈ C ∞ ( τ , T ) . Since ∂ ν m ε = 0 \partial_\nu m_\varepsilon=0 ∂ ν m ε = 0 , we have
∫ τ T ζ ( s ) ⟨ h ( m ε ) , ψ ⟩ d s = − α ∫ τ T ζ ( s ) ⟨ ∇ m ε , ∇ ψ ⟩ d s + ∫ τ T ζ ( s ) ⟨ B m ε + H d [ m ε ] − θ G ( m ε ) , ψ ⟩ d s . \int_\tau^T \zeta(s)\langle h(m_\varepsilon),\psi\rangle\,ds
= -\alpha\int_\tau^T \zeta(s)\langle\nabla m_\varepsilon,\nabla\psi\rangle\,ds
+ \int_\tau^T \zeta(s)\langle Bm_\varepsilon+H_{\mathrm d}[m_\varepsilon]-\theta G(m_\varepsilon),\psi\rangle\,ds. ∫ τ T ζ ( s ) ⟨ h ( m ε ) , ψ ⟩ d s = − α ∫ τ T ζ ( s ) ⟨ ∇ m ε , ∇ ψ ⟩ d s + ∫ τ T ζ ( s ) ⟨ B m ε + H d [ m ε ] − θ G ( m ε ) , ψ ⟩ d s .
Passing to the limit gives
⟨ h ( m ) , ψ ⟩ = − α ⟨ ∇ m , ∇ ψ ⟩ + ⟨ B m + H d [ m ] − θ G ( m ) , ψ ⟩ , (96) \langle h(m),\psi\rangle=-\alpha\langle\nabla m,\nabla\psi\rangle+\langle Bm+H_d[m]-\theta G(m),\psi\rangle, \tag*{(96)} ⟨ h ( m ) , ψ ⟩ = − α ⟨ ∇ m , ∇ ψ ⟩ + ⟨ B m + H d [ m ] − θ G ( m ) , ψ ⟩ , (96)
for every ψ ∈ H 1 ( D ) \psi\in H^1(\mathcal{D}) ψ ∈ H 1 ( D ) and a.e. s ∈ ( τ , T ) s\in(\tau,T) s ∈ ( τ , T ) . Thus, (29) and (30) hold on ( τ , T ) (\tau,T) ( τ , T ) , for every τ ∈ ( 0 , T ) \tau\in(0,T) τ ∈ ( 0 , T ) . This shows the existence of a global weak solution in the sense of Definition 3.1 .
We next prove properties (i)–(ii), beginning with the uniform separation. Fix τ > 0 \tau>0 τ > 0 . For all sufficiently small ε \varepsilon ε , Proposition 3.9 gives
∣ m ε ( t , x ) ∣ ≤ ρ τ for a.e. x ∈ D and every t ∈ [ τ , T ] . |m_\varepsilon(t,x)|\leq\rho_\tau\qquad\text{for a.e. }x\in\mathcal{D}\text{ and every }t\in[\tau,T]. ∣ m ε ( t , x ) ∣ ≤ ρ τ for a.e. x ∈ D and every t ∈ [ τ , T ] .
Hence, using again the Lipschitz continuity of the positive-part map and (87) ,
∥ ( ∣ m ( t ) ∣ − ρ τ ) + ∥ L 2 ( D ) ≤ ∥ m ( t ) − m ε ( t ) ∥ L 2 ( D ) → 0 \left\|\left(|m(t)|-\rho_\tau\right)_+\right\|_{L^2(\mathcal{D})}\leq\|m(t)-m_\varepsilon(t)\|_{L^2(\mathcal{D})}\to0 ( ∣ m ( t ) ∣ − ρ τ ) + L 2 ( D ) ≤ ∥ m ( t ) − m ε ( t ) ∥ L 2 ( D ) → 0
for every t ∈ [ τ , T ] t\in[\tau,T] t ∈ [ τ , T ] . Therefore,
ess sup x ∈ D ∣ m ( t , x ) ∣ ≤ ρ τ for every t ∈ [ τ , T ] . \operatorname*{ess\,sup}_{x\in\mathcal{D}}|m(t,x)|\leq\rho_\tau\qquad\text{for every }t\in[\tau,T]. x ∈ D ess sup ∣ m ( t , x ) ∣ ≤ ρ τ for every t ∈ [ τ , T ] .
Since T > τ T>\tau T > τ is arbitrary, this proves (32) .
We now pass the inverse-distance estimate to the limit. By (88) and Fubini’s theorem, there is a null set of times outside which
m ε ( t , x ) → m ( t , x ) for a.e. x ∈ D . m_\varepsilon(t,x)\to m(t,x)\qquad\text{for a.e. }x\in\mathcal{D}. m ε ( t , x ) → m ( t , x ) for a.e. x ∈ D .
Since the selected sequence ( ε m ) m ∈ N (\varepsilon_m)_{m\in\mathbb{N}} ( ε m ) m ∈ N is countable, enlarging this null set by the union of the exceptional time sets in Proposition 3.7, we may assume that both this convergence and (59) hold for every t t t outside one null set. Fatou’s lemma therefore gives, for every p ≥ 1 p\geq1 p ≥ 1 ,
∫ D 1 ( 1 − ∣ m ( t ) ∣ 2 ) p d x ≤ lim inf ε → 0 ∫ D 1 ( ε + 1 − ∣ m ε ( t ) ∣ 2 ) p d x ≤ C p ( 1 + t − p ) for a.e. t ∈ ( 0 , T ) . (97) \int_{\mathcal{D}}\frac{1}{(1-|m(t)|^2)^p}\,dx\leq\liminf_{\varepsilon\to0}\int_{\mathcal{D}}\frac{1}{(\varepsilon+1-|m_\varepsilon(t)|^2)^p}\,dx
\leq C_p(1+t^{-p})\qquad\text{for a.e. }t\in(0,T). \tag*{(97)} ∫ D ( 1 − ∣ m ( t ) ∣ 2 ) p 1 d x ≤ ε → 0 lim inf ∫ D ( ε + 1 − ∣ m ε ( t ) ∣ 2 ) p 1 d x ≤ C p ( 1 + t − p ) for a.e. t ∈ ( 0 , T ) . (97)
To extend this estimate to every positive time, fix τ ∈ ( 0 , T ) \tau\in(0,T) τ ∈ ( 0 , T ) . By the separation result (32) already proved, the map v ↦ ( 1 − ∣ v ∣ 2 ) − p v\mapsto(1-|v|^2)^{-p} v ↦ ( 1 − ∣ v ∣ 2 ) − p is Lipschitz on { ∣ v ∣ ≤ ρ τ } \{|v|\leq\rho_\tau\} { ∣ v ∣ ≤ ρ τ } . Hence, for s , t ∈ [ τ , T ] s,t\in[\tau,T] s , t ∈ [ τ , T ] ,
∣ ∫ D 1 ( 1 − ∣ m ( t ) ∣ 2 ) p d x − ∫ D 1 ( 1 − ∣ m ( s ) ∣ 2 ) p d x ∣ ≤ C τ , p ∥ m ( t ) − m ( s ) ∥ L 1 ( D ) . \left|\int_{\mathcal{D}}\frac{1}{(1-|m(t)|^2)^p}\,dx-\int_{\mathcal{D}}\frac{1}{(1-|m(s)|^2)^p}\,dx\right|\leq C_{\tau,p}\|m(t)-m(s)\|_{L^1(\mathcal{D})}. ∫ D ( 1 − ∣ m ( t ) ∣ 2 ) p 1 d x − ∫ D ( 1 − ∣ m ( s ) ∣ 2 ) p 1 d x ≤ C τ , p ∥ m ( t ) − m ( s ) ∥ L 1 ( D ) .
Thus, the left-hand side of (97) is continuous in time on [ τ , T ] [\tau,T] [ τ , T ] . Approximating an arbitrary t ∈ [ τ , T ] t\in[\tau,T] t ∈ [ τ , T ] by times for which (97) holds gives (31) for every t ∈ [ τ , T ] t\in[\tau,T] t ∈ [ τ , T ] . Since τ > 0 \tau>0 τ > 0 is arbitrary, property (i) follows.
Next, we verify property (iii), including (33) . The differential estimate (56) obtained in the proof of Lemma 3.6 gives
∥ ∇ m ε ( t ) ∥ L 2 ( D ) 2 ≤ e − 4 θ t ∥ ∇ m 0 ∥ L 2 ( D ) 2 + C ( 1 − e − 4 θ t ) , t ≥ 0 , \|\nabla m_\varepsilon(t)\|_{L^2(\mathcal{D})}^2\leq e^{-4\theta t}\|\nabla m_0\|_{L^2(\mathcal{D})}^2+C(1-e^{-4\theta t}),\qquad t\geq0, ∥∇ m ε ( t ) ∥ L 2 ( D ) 2 ≤ e − 4 θ t ∥∇ m 0 ∥ L 2 ( D ) 2 + C ( 1 − e − 4 θ t ) , t ≥ 0 ,
where C C C is independent of ε \varepsilon ε and t t t . For every fixed t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , the strong L 2 L^2 L 2 -convergence (87) and the uniform H 1 H^1 H 1 -bound imply m ε ( t ) ⇀ m ( t ) m_\varepsilon(t)\rightharpoonup m(t) m ε ( t ) ⇀ m ( t ) in H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) . Thus, by weak lower semicontinuity,
∥ ∇ m ( t ) ∥ L 2 ( D ) 2 ≤ e − 4 θ t ∥ ∇ m 0 ∥ L 2 ( D ) 2 + C ( 1 − e − 4 θ t ) . \|\nabla m(t)\|_{L^2(\mathcal{D})}^2\leq e^{-4\theta t}\|\nabla m_0\|_{L^2(\mathcal{D})}^2+C(1-e^{-4\theta t}). ∥∇ m ( t ) ∥ L 2 ( D ) 2 ≤ e − 4 θ t ∥∇ m 0 ∥ L 2 ( D ) 2 + C ( 1 − e − 4 θ t ) .
Hence, we obtain
lim sup t ↓ 0 ∥ ∇ m ( t ) ∥ L 2 ( D ) ≤ ∥ ∇ m 0 ∥ L 2 ( D ) . \limsup_{t\downarrow0}\|\nabla m(t)\|_{L^2(\mathcal{D})}\leq\|\nabla m_0\|_{L^2(\mathcal{D})}. t ↓ 0 lim sup ∥∇ m ( t ) ∥ L 2 ( D ) ≤ ∥∇ m 0 ∥ L 2 ( D ) .
On the other hand, (87) gives m ( t ) → m 0 m(t)\to m_0 m ( t ) → m 0 strongly in L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) as t ↓ 0 t\downarrow0 t ↓ 0 . The uniform H 1 H^1 H 1 -bound then implies m ( t ) ⇀ m 0 m(t)\rightharpoonup m_0 m ( t ) ⇀ m 0 in H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) as t ↓ 0 t\downarrow0 t ↓ 0 . Therefore,
∥ ∇ m 0 ∥ L 2 ( D ) ≤ lim inf t ↓ 0 ∥ ∇ m ( t ) ∥ L 2 ( D ) . \|\nabla m_0\|_{L^2(\mathcal{D})}\leq\liminf_{t\downarrow0}\|\nabla m(t)\|_{L^2(\mathcal{D})}. ∥∇ m 0 ∥ L 2 ( D ) ≤ t ↓ 0 lim inf ∥∇ m ( t ) ∥ L 2 ( D ) .
Consequently, m ( t ) → m 0 m(t) \to m_0 m ( t ) → m 0 strongly in H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) as t ↓ 0 t \downarrow0 t ↓ 0 . Note that we have already shown that m ∈ H 1 ( τ , T ; L 2 ( D ) ) ∩ L 2 ( τ , T ; H 2 ( D ) ) m \in H^1(\tau,T;L^2(\mathcal{D})) \cap L^2(\tau,T;H^2(\mathcal{D})) m ∈ H 1 ( τ , T ; L 2 ( D )) ∩ L 2 ( τ , T ; H 2 ( D )) . The standard Lions–Magenes interpolation result gives m ∈ C ( [ τ , T ] ; H 1 ( D ) ) m \in C([\tau,T];H^1(\mathcal{D})) m ∈ C ([ τ , T ] ; H 1 ( D )) . Combining this with the continuity at t = 0 t=0 t = 0 yields m ∈ C ( [ 0 , T ] ; H 1 ( D ) ) m \in C([0,T];H^1(\mathcal{D})) m ∈ C ([ 0 , T ] ; H 1 ( D )) , as required.
Finally, we prove property (iv). Fix 0 < s < t 0 < s < t 0 < s < t . By (32) , applied with τ = s \tau=s τ = s , there exists ρ s ∈ ( 0 , 1 ) \rho_s \in(0,1) ρ s ∈ ( 0 , 1 ) such that ∣ m ( r , x ) ∣ ≤ ρ s |m(r,x)| \leq\rho_s ∣ m ( r , x ) ∣ ≤ ρ s for a.e. x ∈ D x \in\mathcal{D} x ∈ D and every r ∈ [ s , t ] r \in[s,t] r ∈ [ s , t ] . Hence m m m takes values a.e. in the compact set B ρ s ( 0 ) ⊂ B 1 ( 0 ) B_{\rho_s}(0) \subset B_1(0) B ρ s ( 0 ) ⊂ B 1 ( 0 ) . Since G ∈ C 1 ( B 1 ( 0 ) ; R 3 ) G \in C^1(B_1(0);\mathbb{R}^3) G ∈ C 1 ( B 1 ( 0 ) ; R 3 ) , it is Lipschitz on B ρ s ( 0 ) B_{\rho_s}(0) B ρ s ( 0 ) . By the regularity m ∈ H 1 ( s , t ; L 2 ( D ) ) ∩ L 2 ( s , t ; H 2 ( D ) ) m \in H^1(s,t;L^2(\mathcal{D})) \cap L^2(s,t;H^2(\mathcal{D})) m ∈ H 1 ( s , t ; L 2 ( D )) ∩ L 2 ( s , t ; H 2 ( D )) , we can apply the chain rule, symmetry of B \mathrm{B} B , and the stray-field identities to obtain, for a.e. r ∈ ( s , t ) r \in(s,t) r ∈ ( s , t ) ,
d d r E θ ( m ( r ) ) = − α ⟨ Δ m ( r ) , ∂ t m ( r ) ⟩ − ⟨ B m ( r ) , ∂ t m ( r ) ⟩ − ⟨ H d [ m ( r ) ] , ∂ t m ( r ) ⟩ + θ ⟨ G ( m ( r ) ) , ∂ t m ( r ) ⟩ = − ⟨ h ( m ( r ) ) , ∂ t m ( r ) ⟩ . \frac{\mathrm{d}}{\mathrm{d}r}\mathcal{E}_{\theta}(m(r))=-\alpha\langle\Delta m(r),\partial_t m(r)\rangle-\langle\mathrm{B}m(r),\partial_t m(r)\rangle-\langle H_{\mathrm{d}}[m(r)],\partial_t m(r)\rangle+\theta\langle G(m(r)),\partial_t m(r)\rangle
=-\langle h(m(r)),\partial_t m(r)\rangle. d r d E θ ( m ( r )) = − α ⟨ Δ m ( r ) , ∂ t m ( r )⟩ − ⟨ B m ( r ) , ∂ t m ( r )⟩ − ⟨ H d [ m ( r )] , ∂ t m ( r )⟩ + θ ⟨ G ( m ( r )) , ∂ t m ( r )⟩ = − ⟨ h ( m ( r )) , ∂ t m ( r )⟩ .
Taking the L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) inner product of (16) with h ( m ) h(m) h ( m ) gives
⟨ h ( m ( r ) ) , ∂ t m ( r ) ⟩ = γ ⟨ h ( m ( r ) ) , m ( r ) × h ( m ( r ) ) ⟩ + ⟨ h ( m ( r ) ) , M ( m ( r ) ) h ( m ( r ) ) ⟩ = ⟨ M ( m ( r ) ) h ( m ( r ) ) , h ( m ( r ) ) ⟩ . \langle h(m(r)),\partial_t m(r)\rangle=\gamma\langle h(m(r)),m(r)\times h(m(r))\rangle+\langle h(m(r)),\mathrm{M}(m(r))h(m(r))\rangle
=\langle\mathrm{M}(m(r))h(m(r)),h(m(r))\rangle. ⟨ h ( m ( r )) , ∂ t m ( r )⟩ = γ ⟨ h ( m ( r )) , m ( r ) × h ( m ( r ))⟩ + ⟨ h ( m ( r )) , M ( m ( r )) h ( m ( r ))⟩ = ⟨ M ( m ( r )) h ( m ( r )) , h ( m ( r ))⟩ .
Hence,
d d r E θ ( m ( r ) ) + ∫ D h ( m ( r ) ) ⋅ M ( m ( r ) ) h ( m ( r ) ) d x = 0 \frac{\mathrm{d}}{\mathrm{d}r}\mathcal{E}_{\theta}(m(r))+\int_{\mathcal{D}}h(m(r))\cdot\mathrm{M}(m(r))h(m(r))\,\mathrm{d}x=0 d r d E θ ( m ( r )) + ∫ D h ( m ( r )) ⋅ M ( m ( r )) h ( m ( r )) d x = 0
for a.e. r ∈ ( s , t ) r \in(s,t) r ∈ ( s , t ) . Since r ↦ E θ ( m ( r ) ) r \mapsto\mathcal{E}_{\theta}(m(r)) r ↦ E θ ( m ( r )) is absolutely continuous on [ s , t ] [s,t] [ s , t ] , integration in time yields
E θ ( m ( t ) ) + ∫ s t ∫ D h ( m ( r ) ) ⋅ M ( m ( r ) ) h ( m ( r ) ) d x d r = E θ ( m ( s ) ) , \mathcal{E}_{\theta}(m(t))+\int_s^t\int_{\mathcal{D}}h(m(r))\cdot\mathrm{M}(m(r))h(m(r))\,\mathrm{d}x\,\mathrm{d}r=\mathcal{E}_{\theta}(m(s)), E θ ( m ( t )) + ∫ s t ∫ D h ( m ( r )) ⋅ M ( m ( r )) h ( m ( r )) d x d r = E θ ( m ( s )) ,
which proves (34) .
Now, we show (35) . Suppose that ∣ { x ∈ D : ∣ m 0 ( x ) ∣ = 1 } ∣ > 0 \lvert\{x \in\mathcal{D}:|m_0(x)|=1\}\rvert>0 ∣{ x ∈ D : ∣ m 0 ( x ) ∣ = 1 }∣ > 0 . If (35) were false, there would exist t j ↓ 0 t_j \downarrow0 t j ↓ 0 such that E θ ( m ( t j ) ) \mathcal{E}_{\theta}(m(t_j)) E θ ( m ( t j )) remains bounded. Since m ( t j ) → m 0 m(t_j) \to m_0 m ( t j ) → m 0 strongly in H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) , after passing to a subsequence, we have m ( t j , x ) → m 0 ( x ) m(t_j,x) \to m_0(x) m ( t j , x ) → m 0 ( x ) for a.e. x ∈ D x \in\mathcal{D} x ∈ D . Let Φ \Phi Φ be the function defined in (18) . Noting that Φ ≥ 0 \Phi\geq0 Φ ≥ 0 and Φ ( s ) → + ∞ \Phi(s) \to+\infty Φ ( s ) → + ∞ as s ↑ 1 s \uparrow1 s ↑ 1 , we obtain, by Fatou’s lemma,
lim inf j → ∞ ∫ D Φ ( ∣ m ( t j ) ∣ ) d x = + ∞ . (98) \liminf_{j\to\infty}\int_{\mathcal{D}}\Phi(|m(t_j)|)\,\mathrm{d}x=+\infty. \tag*{(98)} j → ∞ lim inf ∫ D Φ ( ∣ m ( t j ) ∣ ) d x = + ∞. (98)
The exchange term in E θ ( m ( t j ) ) \mathcal{E}_{\theta}(m(t_j)) E θ ( m ( t j )) converges to a finite limit by the strong H 1 H^1 H 1 -convergence, while the anisotropy and stray-field terms converge by the strong L 2 L^2 L 2 -convergence and the L 2 L^2 L 2 -continuity of H d H_{\mathrm{d}} H d . Therefore, E θ ( m ( t j ) ) → + ∞ \mathcal{E}_{\theta}(m(t_j)) \to+\infty E θ ( m ( t j )) → + ∞ , in contradiction with (98) . This shows (35) .
The proof is now complete. □
Low-temperature limit and long-time behaviour
Having established the global weak theory and the positive-time separation from saturation, we now investigate two asymptotic regimes of the full nonlinear LLB dynamics. The first concerns the low-temperature limit θ → 0 \theta\to0 θ → 0 , in which the longitudinal relaxation is expected to disappear and the dynamics approach the classical Landau–Lifshitz–Gilbert equation on the unit sphere.
The second concerns the asymptotic behaviour for fixed temperature as t → ∞ t \to\infty t → ∞ . Positive-time separation makes the mobility uniformly coercive and allows us to exploit the thermodynamic formulation and the free-energy dissipation. Combined with a Łojasiewicz–Simon inequality, this yields convergence of every global solution to a single stationary state. We then identify a spectral Curie-type threshold associated with the Hessian of the free energy and show that, above this threshold, the zero equilibrium is globally exponentially attracting.
Throughout this section, global weak solutions are understood to satisfy the additional properties established in Theorem 3.2 .
Convergence to the Landau–Lifshitz–Gilbert equation
For v ∈ H 2 ( D ) v \in H^2(\mathcal{D}) v ∈ H 2 ( D ) satisfying ∂ ν v = 0 \partial_\nu v = 0 ∂ ν v = 0 , recall from (6) that H M F A H_{\mathrm{MFA}} H MFA is a self-adjoint operator given by
H M F A ( v ) = α Δ v + B v + H d [ v ] , H_{\mathrm{MFA}}(v) = \alpha\Delta v + Bv + H_{\mathrm{d}}[v], H MFA ( v ) = α Δ v + B v + H d [ v ] ,
where B = J I + A B = JI + A B = J I + A . Since M ( n ) = I − n ⊗ n M(n) = I - n \otimes n M ( n ) = I − n ⊗ n whenever ∣ n ∣ = 1 |n| = 1 ∣ n ∣ = 1 , the corresponding normalised Landau–Lifshitz equation is
∂ t n = γ n × H M F A ( n ) + M ( n ) H M F A ( n ) , ∣ n ∣ = 1. (99) \partial_t n = \gamma n \times H_{\mathrm{MFA}}(n) + M(n)H_{\mathrm{MFA}}(n), \qquad|n| = 1. \tag*{(99)} ∂ t n = γ n × H MFA ( n ) + M ( n ) H MFA ( n ) , ∣ n ∣ = 1. (99)
Equivalently,
∂ t n = γ n × H M F A ( n ) − n × ( n × H M F A ( n ) ) . \partial_t n = \gamma n \times H_{\mathrm{MFA}}(n) - n \times\left(n \times H_{\mathrm{MFA}}(n)\right). ∂ t n = γ n × H MFA ( n ) − n × ( n × H MFA ( n ) ) .
If B = J I + A B = JI + A B = J I + A , the isotropic contribution J n Jn J n disappears from (99) , while the anisotropic contribution A n An A n remains. This is the usual form of the Landau–Lifshitz–Gilbert equation [2 , 22 ] .
We recall that the extension of M M M to B 1 ( 0 ) B_1(0) B 1 ( 0 ) belongs to C 1 , 1 ( B 1 ( 0 ) ; R 3 × 3 ) C^{1,1}(B_1(0);\mathbb{R}^{3\times3}) C 1 , 1 ( B 1 ( 0 ) ; R 3 × 3 ) ; see Lemma A.2 . Fix κ : = 1 + ∥ B ∥ \kappa:= 1 + \|B\| κ := 1 + ∥ B ∥ , and define, for v ∈ H 1 ( D ) v \in H^1(\mathcal{D}) v ∈ H 1 ( D ) ,
R ( v ) : = κ 2 ∥ v ∥ L 2 ( D ) 2 + α 2 ∥ ∇ v ∥ L 2 ( D ) 2 − 1 2 ⟨ B v , v ⟩ + 1 2 ∥ H d [ v ] ∥ L 2 ( R 3 ) 2 . (100) \mathcal{R}(v) := \frac{\kappa}{2}\|v\|_{L^2(\mathcal{D})}^2 + \frac{\alpha}{2}\|\nabla v\|_{L^2(\mathcal{D})}^2 - \frac{1}{2}\langle Bv,v\rangle+ \frac{1}{2}\|H_{\mathrm{d}}[v]\|_{L^2(\mathbb{R}^3)}^2. \tag*{(100)} R ( v ) := 2 κ ∥ v ∥ L 2 ( D ) 2 + 2 α ∥∇ v ∥ L 2 ( D ) 2 − 2 1 ⟨ B v , v ⟩ + 2 1 ∥ H d [ v ] ∥ L 2 ( R 3 ) 2 . (100)
Since κ − ∥ B ∥ ≥ 1 \kappa- \|B\| \ge1 κ − ∥ B ∥ ≥ 1 , there exist constants c , C > 0 c,C > 0 c , C > 0 such that
c ∥ v ∥ H 1 ( D ) 2 ≤ R ( v ) ≤ C ∥ v ∥ H 1 ( D ) 2 . (101) c\|v\|_{H^1(\mathcal{D})}^2 \le\mathcal{R}(v) \le C\|v\|_{H^1(\mathcal{D})}^2. \tag*{(101)} c ∥ v ∥ H 1 ( D ) 2 ≤ R ( v ) ≤ C ∥ v ∥ H 1 ( D ) 2 . (101)
The following theorem shows that, in effect, the Landau–Lifshitz equation (99) is an approximation of the full nonlinear LLB equation at low temperatures (θ → 0 + \theta\to0^+ θ → 0 + ). We also provide a quantitative approximation rate.
Theorem 4.1 (Low-temperature weak–strong convergence). Let T > 0 T > 0 T > 0 , and let n n n be a strong solution of the Landau–Lifshitz equation (99) on [ 0 , T ] [0,T] [ 0 , T ] such that
n ∈ C ( [ 0 , T ] ; H 1 ( D ) ) ∩ H 1 ( 0 , T ; L 2 ( D ) ) ∩ L 2 ( 0 , T ; H 2 ( D ) ) , n ∈ L ∞ ( 0 , T ; W 1 , ∞ ( D ) ) , H M F A ( n ) ∈ L 2 ( 0 , T ; W 1 , ∞ ( D ) ) , (102) \begin{aligned}
n &\in C([0,T];H^1(\mathcal{D})) \cap H^1(0,T;L^2(\mathcal{D})) \cap L^2(0,T;H^2(\mathcal{D})),\\
n &\in L^\infty(0,T;W^{1,\infty}(\mathcal{D})), \qquad H_{\mathrm{MFA}}(n) \in L^2(0,T;W^{1,\infty}(\mathcal{D})), \tag*{(102)}
\end{aligned} n n ∈ C ([ 0 , T ] ; H 1 ( D )) ∩ H 1 ( 0 , T ; L 2 ( D )) ∩ L 2 ( 0 , T ; H 2 ( D )) , ∈ L ∞ ( 0 , T ; W 1 , ∞ ( D )) , H MFA ( n ) ∈ L 2 ( 0 , T ; W 1 , ∞ ( D )) , (102) with initial datum n ( 0 ) = n 0 n(0) = n_0 n ( 0 ) = n 0 satisfying ∣ n 0 ∣ = 1 |n_0| = 1 ∣ n 0 ∣ = 1 a.e. in D \mathcal{D} D .
For θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] , let m θ m_\theta m θ be a global weak solution of the full nonlinear LLB equation with initial datum m θ ( 0 ) = m θ , 0 m_\theta(0) = m_{\theta,0} m θ ( 0 ) = m θ , 0 , given by Theorem 3.2. Set e θ : = m θ − n e_\theta:= m_\theta- n e θ := m θ − n . Then, there exists C T > 0 C_T > 0 C T > 0 , independent of θ \theta θ , such that for every t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] ,
R ( e θ ( t ) ) + 1 2 ∫ 0 t ⟨ M ( m θ ) H M F A ( e θ ) , H M F A ( e θ ) ⟩ d s ≤ C T ( R ( e θ ( 0 ) ) + θ 2 ) . (103) \mathcal{R}(e_\theta(t)) + \frac{1}{2}\int_0^t \langle M(m_\theta)H_{\mathrm{MFA}}(e_\theta),H_{\mathrm{MFA}}(e_\theta)\rangle\,\mathrm{d}s \le C_T\left(\mathcal{R}(e_\theta(0))+\theta^2\right). \tag*{(103)} R ( e θ ( t )) + 2 1 ∫ 0 t ⟨ M ( m θ ) H MFA ( e θ ) , H MFA ( e θ )⟩ d s ≤ C T ( R ( e θ ( 0 )) + θ 2 ) . (103) In particular, we have
sup t ∈ [ 0 , T ] ∥ m θ ( t ) − n ( t ) ∥ H 1 ( D ) ≤ C T ( ∥ m θ ( 0 ) − n ( 0 ) ∥ H 1 ( D ) + θ ) . (104) \sup_{t\in[0,T]}\|m_\theta(t)-n(t)\|_{H^1(\mathcal{D})} \le C_T\left(\|m_\theta(0)-n(0)\|_{H^1(\mathcal{D})}+\theta\right). \tag*{(104)} t ∈ [ 0 , T ] sup ∥ m θ ( t ) − n ( t ) ∥ H 1 ( D ) ≤ C T ( ∥ m θ ( 0 ) − n ( 0 ) ∥ H 1 ( D ) + θ ) . (104) Furthermore,
sup t ∈ [ 0 , T ] ∥ 1 − ∣ m θ ( t ) ∣ 2 ∥ L 2 ( D ) ≤ C T ( ∥ m θ ( 0 ) − n ( 0 ) ∥ H 1 ( D ) + θ ) . (105) \sup_{t\in[0,T]}\|1-|m_\theta(t)|^2\|_{L^2(\mathcal{D})} \le C_T\left(\|m_\theta(0)-n(0)\|_{H^1(\mathcal{D})}+\theta\right). \tag*{(105)} t ∈ [ 0 , T ] sup ∥1 − ∣ m θ ( t ) ∣ 2 ∥ L 2 ( D ) ≤ C T ( ∥ m θ ( 0 ) − n ( 0 ) ∥ H 1 ( D ) + θ ) . (105)
Proof. Let e θ = m θ − n e_\theta= m_\theta- n e θ = m θ − n . For every δ > 0 \delta> 0 δ > 0 , the solution m θ m_\theta m θ on [ δ , T ] [\delta,T] [ δ , T ] is strictly separated from saturation and has the positive-time regularity given by Definition 3.1. Hence, all calculations below are legitimate on [ δ , T ] [\delta,T] [ δ , T ] .
Since H M F A H_{\mathrm{MFA}} H MFA is linear, subtracting (99) from the closed-ball formulation of the LLB equation gives
∂ t e θ = γ m θ × H M F A ( e θ ) + M ( m θ ) H M F A ( e θ ) + γ e θ × H M F A ( n ) + ( M ( m θ ) − M ( n ) ) H M F A ( n ) − 2 θ m θ . (106) \begin{aligned}
\partial_t e_\theta={}& \gamma m_\theta\times H_{\mathrm{MFA}}(e_\theta) + M(m_\theta)H_{\mathrm{MFA}}(e_\theta)\\
&+ \gamma e_\theta\times H_{\mathrm{MFA}}(n) + \left(M(m_\theta)-M(n)\right)H_{\mathrm{MFA}}(n) - 2\theta m_\theta. \tag*{(106)}
\end{aligned} ∂ t e θ = γ m θ × H MFA ( e θ ) + M ( m θ ) H MFA ( e θ ) + γ e θ × H MFA ( n ) + ( M ( m θ ) − M ( n ) ) H MFA ( n ) − 2 θ m θ . (106)
Here, we used the fact that M ( n ) = I − n ⊗ n \mathrm{M}(n)=I-n\otimes n M ( n ) = I − n ⊗ n .
We first record some preliminary estimates that will be used in the relative-energy identity. If v ∈ W 1 , ∞ ( D ) v\in W^{1,\infty}(\mathcal{D}) v ∈ W 1 , ∞ ( D ) , integration by parts, (101) , and (A.7) give
∣ ⟨ H M F A ( e θ ) , e θ × v ⟩ ∣ ≤ α ∑ i = 1 3 ∣ ⟨ ∂ i e θ , e θ × ∂ i v ⟩ ∣ + C ∥ v ∥ L ∞ ( D ) ∥ e θ ∥ L 2 ( D ) 2 ≤ C ( 1 + ∥ v ∥ W 1 , ∞ ( D ) ) R ( e θ ) . (107) \begin{aligned}
\left|\left\langle H_{\mathrm{MFA}}(e_\theta),e_\theta\times v\right\rangle\right|\leq\alpha\sum_{i=1}^{3}\left|\left\langle\partial_i e_\theta,e_\theta\times\partial_i v\right\rangle\right|+C\|v\|_{L^\infty(\mathcal{D})}\|e_\theta\|_{L^2(\mathcal{D})}^{2} \\
&\leq C\left(1+\|v\|_{W^{1,\infty}(\mathcal{D})}\right)\mathcal{R}(e_\theta). \tag*{(107)}
\end{aligned} ∣ ⟨ H MFA ( e θ ) , e θ × v ⟩ ∣ ≤ α i = 1 ∑ 3 ∣ ⟨ ∂ i e θ , e θ × ∂ i v ⟩ ∣ + C ∥ v ∥ L ∞ ( D ) ∥ e θ ∥ L 2 ( D ) 2 ≤ C ( 1 + ∥ v ∥ W 1 , ∞ ( D ) ) R ( e θ ) . (107)
Since M ∈ C 1 , 1 ( B 1 ( 0 ) ‾ ; R 3 × 3 ) \mathrm{M}\in C^{1,1}(\overline{B_1(0)};\mathbb{R}^{3\times3}) M ∈ C 1 , 1 ( B 1 ( 0 ) ; R 3 × 3 ) , the Sobolev chain rule gives, for i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 ,
∂ i ( M ( m θ ) − M ( n ) ) = D M ( m θ ) ∂ i m θ − D M ( n ) ∂ i n = D M ( m θ ) ∂ i e θ + ( D M ( m θ ) − D M ( n ) ) ∂ i n . \begin{aligned}
\partial_i\left(\mathrm{M}(m_\theta)-\mathrm{M}(n)\right)=D\mathrm{M}(m_\theta)\partial_i m_\theta-D\mathrm{M}(n)\partial_i n \\
&=D\mathrm{M}(m_\theta)\partial_i e_\theta+\left(D\mathrm{M}(m_\theta)-D\mathrm{M}(n)\right)\partial_i n.
\end{aligned} ∂ i ( M ( m θ ) − M ( n ) ) = D M ( m θ ) ∂ i m θ − D M ( n ) ∂ i n = D M ( m θ ) ∂ i e θ + ( D M ( m θ ) − D M ( n ) ) ∂ i n .
Using the boundedness of D M D\mathrm{M} D M and the Lipschitz continuity of D M D\mathrm{M} D M on B 1 ( 0 ) ‾ \overline{B_1(0)} B 1 ( 0 ) , we obtain
∥ M ( m θ ) − M ( n ) ∥ L 2 ( D ) ≤ C ∥ e θ ∥ L 2 ( D ) , \|\mathrm{M}(m_\theta)-\mathrm{M}(n)\|_{L^2(\mathcal{D})}\leq C\|e_\theta\|_{L^2(\mathcal{D})}, ∥ M ( m θ ) − M ( n ) ∥ L 2 ( D ) ≤ C ∥ e θ ∥ L 2 ( D ) ,
∥ ∇ ( M ( m θ ) − M ( n ) ) ∥ L 2 ( D ) ≤ C ( ∥ ∇ e θ ∥ L 2 ( D ) + ∥ ∇ n ∥ L ∞ ( D ) ∥ e θ ∥ L 2 ( D ) ) . \|\nabla(\mathrm{M}(m_\theta)-\mathrm{M}(n))\|_{L^2(\mathcal{D})}\leq C\left(\|\nabla e_\theta\|_{L^2(\mathcal{D})}+\|\nabla n\|_{L^\infty(\mathcal{D})}\|e_\theta\|_{L^2(\mathcal{D})}\right). ∥∇ ( M ( m θ ) − M ( n )) ∥ L 2 ( D ) ≤ C ( ∥∇ e θ ∥ L 2 ( D ) + ∥∇ n ∥ L ∞ ( D ) ∥ e θ ∥ L 2 ( D ) ) .
Consequently, we have
∥ ( M ( m θ ) − M ( n ) ) H M F A ( n ) ∥ H 1 ( D ) ≤ C ( 1 + ∥ ∇ n ∥ L ∞ ( D ) ) ∥ H M F A ( n ) ∥ W 1 , ∞ ( D ) ∥ e θ ∥ H 1 ( D ) . (108) \|(\mathrm{M}(m_\theta)-\mathrm{M}(n))H_{\mathrm{MFA}}(n)\|_{H^1(\mathcal{D})}\leq C\left(1+\|\nabla n\|_{L^\infty(\mathcal{D})}\right)\|H_{\mathrm{MFA}}(n)\|_{W^{1,\infty}(\mathcal{D})}\|e_\theta\|_{H^1(\mathcal{D})}. \tag*{(108)} ∥ ( M ( m θ ) − M ( n )) H MFA ( n ) ∥ H 1 ( D ) ≤ C ( 1 + ∥∇ n ∥ L ∞ ( D ) ) ∥ H MFA ( n ) ∥ W 1 , ∞ ( D ) ∥ e θ ∥ H 1 ( D ) . (108)
By the symmetry of B \mathrm{B} B and of the stray-field operator,
d d t R ( e θ ) = ⟨ κ e θ − H M F A ( e θ ) , ∂ t e θ ⟩ \frac{\mathrm{d}}{\mathrm{d}t}\mathcal{R}(e_\theta)=\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),\partial_t e_\theta\right\rangle d t d R ( e θ ) = ⟨ κ e θ − H MFA ( e θ ) , ∂ t e θ ⟩
for a.e. t ∈ ( δ , T ) t\in(\delta,T) t ∈ ( δ , T ) . Substituting (106) , we obtain
d d t R ( e θ ) = γ ⟨ κ e θ − H M F A ( e θ ) , m θ × H M F A ( e θ ) ⟩ + ⟨ κ e θ − H M F A ( e θ ) , M ( m θ ) H M F A ( e θ ) ⟩ + γ ⟨ κ e θ − H M F A ( e θ ) , e θ × H M F A ( n ) ⟩ + ⟨ κ e θ − H M F A ( e θ ) , ( M ( m θ ) − M ( n ) ) H M F A ( n ) ⟩ − 2 θ ⟨ κ e θ − H M F A ( e θ ) , m θ ⟩ . (109) \begin{aligned}
\frac{\mathrm{d}}{\mathrm{d}t}\mathcal{R}(e_\theta)
={}&\gamma\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),m_\theta\times H_{\mathrm{MFA}}(e_\theta)\right\rangle\\
&+\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),\mathrm{M}(m_\theta)H_{\mathrm{MFA}}(e_\theta)\right\rangle\\
&+\gamma\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),e_\theta\times H_{\mathrm{MFA}}(n)\right\rangle\\
&+\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),\left(\mathrm{M}(m_\theta)-\mathrm{M}(n)\right)H_{\mathrm{MFA}}(n)\right\rangle\\
&-2\theta\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),m_\theta\right\rangle. \tag*{(109)}
\end{aligned} d t d R ( e θ ) = γ ⟨ κ e θ − H MFA ( e θ ) , m θ × H MFA ( e θ ) ⟩ + ⟨ κ e θ − H MFA ( e θ ) , M ( m θ ) H MFA ( e θ ) ⟩ + γ ⟨ κ e θ − H MFA ( e θ ) , e θ × H MFA ( n ) ⟩ + ⟨ κ e θ − H MFA ( e θ ) , ( M ( m θ ) − M ( n ) ) H MFA ( n ) ⟩ − 2 θ ⟨ κ e θ − H MFA ( e θ ) , m θ ⟩ . (109)
We estimate these five terms in the order in which they appear.
For the principal precession term, we note that e θ × m θ = e θ × n e_\theta\times m_\theta=e_\theta\times n e θ × m θ = e θ × n . Hence, (107) , applied with v = n v=n v = n , gives
∣ γ ⟨ κ e θ − H M F A ( e θ ) , m θ × H M F A ( e θ ) ⟩ ∣ = ∣ γ ∣ κ ∣ ⟨ H M F A ( e θ ) , e θ × n ⟩ ∣ ≤ C R ( e θ ) . (110) \left|\gamma\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),m_\theta\times H_{\mathrm{MFA}}(e_\theta)\right\rangle\right|=|\gamma|\kappa\left|\left\langle H_{\mathrm{MFA}}(e_\theta),e_\theta\times n\right\rangle\right|\leq C\mathcal{R}(e_\theta). \tag*{(110)} ∣ γ ⟨ κ e θ − H MFA ( e θ ) , m θ × H MFA ( e θ ) ⟩ ∣ = ∣ γ ∣ κ ∣ ⟨ H MFA ( e θ ) , e θ × n ⟩ ∣ ≤ C R ( e θ ) . (110)
We estimate the principal damping term next. Note that since 0 ≤ M ( m θ ) ≤ I 0\leq\mathrm{M}(m_\theta)\leq I 0 ≤ M ( m θ ) ≤ I , we have M ( m θ ) 2 ≤ M ( m θ ) \mathrm{M}(m_\theta)^2\leq\mathrm{M}(m_\theta) M ( m θ ) 2 ≤ M ( m θ ) . By (A.4), we infer that
⟨ κ e θ − H M F A ( e θ ) , M ( m θ ) H M F A ( e θ ) ⟩ = κ ⟨ e θ , M ( m θ ) H M F A ( e θ ) ⟩ − ⟨ M ( m θ ) H M F A ( e θ ) , H M F A ( e θ ) ⟩ ≤ κ ∥ e θ ∥ L 2 ( D ) ∣ ⟨ M ( m θ ) H M F A ( e θ ) , H M F A ( e θ ) ⟩ ∣ 1 / 2 − ⟨ M ( m θ ) H M F A ( e θ ) , H M F A ( e θ ) ⟩ ≤ − 1 2 ⟨ M ( m θ ) H M F A ( e θ ) , H M F A ( e θ ) ⟩ + C R ( e θ ) , (111) \begin{aligned}
\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),\mathrm{M}(m_\theta)H_{\mathrm{MFA}}(e_\theta)\right\rangle
={}&\kappa\left\langle e_\theta,\mathrm{M}(m_\theta)H_{\mathrm{MFA}}(e_\theta)\right\rangle\\
&-\left\langle\mathrm{M}(m_\theta)H_{\mathrm{MFA}}(e_\theta),H_{\mathrm{MFA}}(e_\theta)\right\rangle\\
\leq{}&\kappa\|e_\theta\|_{L^2(\mathcal{D})}\left|\left\langle\mathrm{M}(m_\theta)H_{\mathrm{MFA}}(e_\theta),H_{\mathrm{MFA}}(e_\theta)\right\rangle\right|^{1/2}\\
&-\left\langle\mathrm{M}(m_\theta)H_{\mathrm{MFA}}(e_\theta),H_{\mathrm{MFA}}(e_\theta)\right\rangle\\
&\le-\frac{1}{2}\langle\mathsf{M}(m_\theta)H_{\mathrm{MFA}}(e_\theta),H_{\mathrm{MFA}}(e_\theta)\rangle+C\mathcal{R}(e_\theta), \tag*{(111)}
\end{aligned} ⟨ κ e θ − H MFA ( e θ ) , M ( m θ ) H MFA ( e θ ) ⟩ = ≤ κ ⟨ e θ , M ( m θ ) H MFA ( e θ ) ⟩ − ⟨ M ( m θ ) H MFA ( e θ ) , H MFA ( e θ ) ⟩ κ ∥ e θ ∥ L 2 ( D ) ∣ ⟨ M ( m θ ) H MFA ( e θ ) , H MFA ( e θ ) ⟩ ∣ 1/2 − ⟨ M ( m θ ) H MFA ( e θ ) , H MFA ( e θ ) ⟩ ≤ − 2 1 ⟨ M ( m θ ) H MFA ( e θ ) , H MFA ( e θ )⟩ + C R ( e θ ) , (111)
where in the last step we used Young’s inequality and (101) .
For the lower-order precession term, the part containing κ e θ \kappa e_\theta κ e θ vanishes. Hence (107) , applied with v = H M F A ( n ) v=H_{\mathrm{MFA}}(n) v = H MFA ( n ) , yields
∣ γ ⟨ κ e θ − H M F A ( e θ ) , e θ × H M F A ( n ) ⟩ ∣ ≤ C ( 1 + ∥ H M F A ( n ) ∥ W 1 , ∞ ( D ) ) R ( e θ ) . (112) \left|\gamma\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),e_\theta\times H_{\mathrm{MFA}}(n)\right\rangle\right|
\le C\left(1+\left\|H_{\mathrm{MFA}}(n)\right\|_{W^{1,\infty}(\mathcal{D})}\right)\mathcal{R}(e_\theta). \tag*{(112)} ∣ γ ⟨ κ e θ − H MFA ( e θ ) , e θ × H MFA ( n ) ⟩ ∣ ≤ C ( 1 + ∥ H MFA ( n ) ∥ W 1 , ∞ ( D ) ) R ( e θ ) . (112)
For the mobility-difference term, the variational definition of H M F A H_{\mathrm{MFA}} H MFA gives
∣ ⟨ κ e θ − H M F A ( e θ ) , v ⟩ ∣ ≤ C ∥ e θ ∥ H 1 ( D ) ∥ v ∥ H 1 ( D ) . \left|\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),v\right\rangle\right|
\le C\left\|e_\theta\right\|_{H^1(\mathcal{D})}\left\|v\right\|_{H^1(\mathcal{D})}. ∣ ⟨ κ e θ − H MFA ( e θ ) , v ⟩ ∣ ≤ C ∥ e θ ∥ H 1 ( D ) ∥ v ∥ H 1 ( D ) .
Consequently, using (108) and (101) , we have
∣ ⟨ κ e θ − H M F A ( e θ ) , ( M ( m θ ) − M ( n ) ) H M F A ( n ) ⟩ ∣ ≤ C ∥ e θ ∥ H 1 ( D ) ∥ ( M ( m θ ) − M ( n ) ) H M F A ( n ) ∥ H 1 ( D ) ≤ C ( 1 + ∥ ∇ n ∥ L ∞ ( D ) ) ∥ H M F A ( n ) ∥ W 1 , ∞ ( D ) ∥ e θ ∥ H 1 ( D ) 2 ≤ C ( 1 + ∥ H M F A ( n ) ∥ W 1 , ∞ ( D ) 2 ) R ( e θ ) . (113) \begin{aligned}
\left|\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),\left(\mathsf{M}(m_\theta)-\mathsf{M}(n)\right)H_{\mathrm{MFA}}(n)\right\rangle\right|
&\le C\left\|e_\theta\right\|_{H^1(\mathcal{D})}\left\|\left(\mathsf{M}(m_\theta)-\mathsf{M}(n)\right)H_{\mathrm{MFA}}(n)\right\|_{H^1(\mathcal{D})}\\
&\le C\left(1+\left\|\nabla n\right\|_{L^\infty(\mathcal{D})}\right)\left\|H_{\mathrm{MFA}}(n)\right\|_{W^{1,\infty}(\mathcal{D})}\left\|e_\theta\right\|_{H^1(\mathcal{D})}^{2}\\
&\le C\left(1+\left\|H_{\mathrm{MFA}}(n)\right\|_{W^{1,\infty}(\mathcal{D})}^{2}\right)\mathcal{R}(e_\theta). \tag*{(113)}
\end{aligned} ∣ ⟨ κ e θ − H MFA ( e θ ) , ( M ( m θ ) − M ( n ) ) H MFA ( n ) ⟩ ∣ ≤ C ∥ e θ ∥ H 1 ( D ) ∥ ( M ( m θ ) − M ( n ) ) H MFA ( n ) ∥ H 1 ( D ) ≤ C ( 1 + ∥ ∇ n ∥ L ∞ ( D ) ) ∥ H MFA ( n ) ∥ W 1 , ∞ ( D ) ∥ e θ ∥ H 1 ( D ) 2 ≤ C ( 1 + ∥ H MFA ( n ) ∥ W 1 , ∞ ( D ) 2 ) R ( e θ ) . (113)
Finally, we estimate the thermal forcing term. Noting the identity
⟨ κ e θ − H M F A ( e θ ) , e θ ⟩ = κ ∥ e θ ∥ L 2 2 + α ∥ ∇ e θ ∥ L 2 2 − ⟨ B e θ , e θ ⟩ + ∥ H d [ e θ ] ∥ L 2 ( R 3 ) 2 = 2 R ( e θ ) , \left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),e_\theta\right\rangle
=\kappa\left\|e_\theta\right\|_{L^2}^{2}+\alpha\left\|\nabla e_\theta\right\|_{L^2}^{2}-\left\langle Be_\theta,e_\theta\right\rangle+\left\|H_d[e_\theta]\right\|_{L^2(\mathbb{R}^3)}^{2}
=2\mathcal{R}(e_\theta), ⟨ κ e θ − H MFA ( e θ ) , e θ ⟩ = κ ∥ e θ ∥ L 2 2 + α ∥ ∇ e θ ∥ L 2 2 − ⟨ B e θ , e θ ⟩ + ∥ H d [ e θ ] ∥ L 2 ( R 3 ) 2 = 2 R ( e θ ) ,
we obtain, by using m θ = e θ + n m_\theta=e_\theta+n m θ = e θ + n and the self-adjointness of H M F A H_{\mathrm{MFA}} H MFA ,
⟨ κ e θ − H M F A ( e θ ) , m θ ⟩ = ⟨ κ e θ − H M F A ( e θ ) , e θ ⟩ + ⟨ κ e θ − H M F A ( e θ ) , n ⟩ = 2 R ( e θ ) + ⟨ e θ , κ n − H M F A ( n ) ⟩ . \begin{aligned}
\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),m_\theta\right\rangle
&=\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),e_\theta\right\rangle+\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),n\right\rangle\\
&=2\mathcal{R}(e_\theta)+\left\langle e_\theta,\kappa n-H_{\mathrm{MFA}}(n)\right\rangle.
\end{aligned} ⟨ κ e θ − H MFA ( e θ ) , m θ ⟩ = ⟨ κ e θ − H MFA ( e θ ) , e θ ⟩ + ⟨ κ e θ − H MFA ( e θ ) , n ⟩ = 2 R ( e θ ) + ⟨ e θ , κn − H MFA ( n ) ⟩ .
Therefore, since ∣ n ∣ = 1 \lvert n\rvert=1 ∣ n ∣ = 1 , θ ≤ 1 \theta\le1 θ ≤ 1 , and ∥ e θ ∥ L 2 2 ≤ C R ( e θ ) \left\|e_\theta\right\|_{L^2}^{2}\le C\mathcal{R}(e_\theta) ∥ e θ ∥ L 2 2 ≤ C R ( e θ ) by (101) , we apply Young’s inequality to obtain
2 θ ∣ ⟨ κ e θ − H M F A ( e θ ) , m θ ⟩ ∣ ≤ 4 θ R ( e θ ) + 2 θ ∥ e θ ∥ L 2 ( κ ∥ n ∥ L 2 + ∥ H M F A ( n ) ∥ L 2 ) ≤ C R ( e θ ) + C θ 2 ( 1 + ∥ H M F A ( n ) ∥ L 2 2 ) ≤ C ( 1 + ∥ H M F A ( n ) ∥ L 2 2 ) R ( e θ ) + C θ 2 ( 1 + ∥ H M F A ( n ) ∥ L 2 2 ) . (114) \begin{aligned}
2\theta\left|\left\langle\kappa e_\theta-H_{\mathrm{MFA}}(e_\theta),m_\theta\right\rangle\right|
&\le4\theta\mathcal{R}(e_\theta)+2\theta\left\|e_\theta\right\|_{L^2}\left(\kappa\left\|n\right\|_{L^2}+\left\|H_{\mathrm{MFA}}(n)\right\|_{L^2}\right)\\
&\le C\mathcal{R}(e_\theta)+C\theta^2\left(1+\left\|H_{\mathrm{MFA}}(n)\right\|_{L^2}^{2}\right)\\
&\le C\left(1+\left\|H_{\mathrm{MFA}}(n)\right\|_{L^2}^{2}\right)\mathcal{R}(e_\theta)+C\theta^2\left(1+\left\|H_{\mathrm{MFA}}(n)\right\|_{L^2}^{2}\right). \tag*{(114)}
\end{aligned} 2 θ ∣ ⟨ κ e θ − H MFA ( e θ ) , m θ ⟩ ∣ ≤ 4 θ R ( e θ ) + 2 θ ∥ e θ ∥ L 2 ( κ ∥ n ∥ L 2 + ∥ H MFA ( n ) ∥ L 2 ) ≤ C R ( e θ ) + C θ 2 ( 1 + ∥ H MFA ( n ) ∥ L 2 2 ) ≤ C ( 1 + ∥ H MFA ( n ) ∥ L 2 2 ) R ( e θ ) + C θ 2 ( 1 + ∥ H MFA ( n ) ∥ L 2 2 ) . (114)
Substituting (110) –(114) into (109) , we deduce that, for a.e. t ∈ ( δ , T ) t\in(\delta,T) t ∈ ( δ , T ) ,
d d t R ( e θ ) + 1 2 ⟨ M ( m θ ) H M F A ( e θ ) , H M F A ( e θ ) ⟩ ≤ C ( 1 + ∥ H M F A ( n ) ∥ W 1 , ∞ ( D ) 2 ) R ( e θ ) + C θ 2 ( 1 + ∥ H M F A ( n ) ∥ W 1 , ∞ ( D ) 2 ) . (115) \frac{\mathrm{d}}{\mathrm{d}t}\mathcal{R}(e_\theta)+\frac{1}{2}\left\langle\mathsf{M}(m_\theta)H_{\mathrm{MFA}}(e_\theta),H_{\mathrm{MFA}}(e_\theta)\right\rangle
\le C\left(1+\left\|H_{\mathrm{MFA}}(n)\right\|_{W^{1,\infty}(\mathcal{D})}^{2}\right)\mathcal{R}(e_\theta)
+C\theta^2\left(1+\left\|H_{\mathrm{MFA}}(n)\right\|_{W^{1,\infty}(\mathcal{D})}^{2}\right). \tag*{(115)} d t d R ( e θ ) + 2 1 ⟨ M ( m θ ) H MFA ( e θ ) , H MFA ( e θ ) ⟩ ≤ C ( 1 + ∥ H MFA ( n ) ∥ W 1 , ∞ ( D ) 2 ) R ( e θ ) + C θ 2 ( 1 + ∥ H MFA ( n ) ∥ W 1 , ∞ ( D ) 2 ) . (115)
Integrating (115) over ( δ , t ) (\delta,t) ( δ , t ) and applying Gronwall’s lemma yields
R ( e θ ( t ) ) + 1 2 ∫ δ t ⟨ M ( m θ ) H M F A ( e θ ) , H M F A ( e θ ) ⟩ d s ≤ C T ( R ( e θ ( δ ) ) + θ 2 ) , (116) \mathcal{R}(e_\theta(t))+\frac{1}{2}\int_\delta^t\left\langle\mathsf{M}(m_\theta)H_{\mathrm{MFA}}(e_\theta),H_{\mathrm{MFA}}(e_\theta)\right\rangle\,\mathrm{d}s
\le C_T\left(\mathcal{R}(e_\theta(\delta))+\theta^2\right), \tag*{(116)} R ( e θ ( t )) + 2 1 ∫ δ t ⟨ M ( m θ ) H MFA ( e θ ) , H MFA ( e θ ) ⟩ d s ≤ C T ( R ( e θ ( δ )) + θ 2 ) , (116)
where C T C_T C T is a constant independent of δ \delta δ .
Since m θ , n ∈ C ( [ 0 , T ] ; H 1 ( D ) ) m_\theta,n\in C([0,T];H^1(\mathcal{D})) m θ , n ∈ C ([ 0 , T ] ; H 1 ( D )) , we have e θ ( δ ) → m θ , 0 − n 0 e_\theta(\delta)\to m_{\theta,0}-n_0 e θ ( δ ) → m θ , 0 − n 0 strongly in H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) as δ ↓ 0 \delta\downarrow0 δ ↓ 0 , and thus, by the continuity of R \mathcal{R} R on H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) , we get
R ( e θ ( δ ) ) → R ( m θ , 0 − n 0 ) . \mathcal{R}(e_\theta(\delta))\to\mathcal{R}(m_{\theta,0}-n_0). R ( e θ ( δ )) → R ( m θ , 0 − n 0 ) .
Moreover, since M ( m θ ) \mathsf{M}(m_\theta) M ( m θ ) is positive semi-definite,
⟨ M ( m θ ) H M F A ( e θ ) , H M F A ( e θ ) ⟩ ≥ 0 a.e. in ( 0 , T ) . \left\langle\mathsf{M}(m_\theta)H_{\mathrm{MFA}}(e_\theta),H_{\mathrm{MFA}}(e_\theta)\right\rangle\ge0
\qquad\text{a.e. in }(0,T). ⟨ M ( m θ ) H MFA ( e θ ) , H MFA ( e θ ) ⟩ ≥ 0 a.e. in ( 0 , T ) .
Therefore, by the monotone convergence theorem, as δ ↓ 0 \delta\downarrow0 δ ↓ 0
∫ δ t ⟨ M ( m θ ) H M F A ( e θ ) , H M F A ( e θ ) ⟩ d s → ∫ 0 t ⟨ M ( m θ ) H M F A ( e θ ) , H M F A ( e θ ) ⟩ d s . \int_\delta^t\left\langle\mathsf{M}(m_\theta)H_{\mathrm{MFA}}(e_\theta),H_{\mathrm{MFA}}(e_\theta)\right\rangle\,\mathrm{d}s
\to
\int_0^t\left\langle\mathsf{M}(m_\theta)H_{\mathrm{MFA}}(e_\theta),H_{\mathrm{MFA}}(e_\theta)\right\rangle\,\mathrm{d}s. ∫ δ t ⟨ M ( m θ ) H MFA ( e θ ) , H MFA ( e θ ) ⟩ d s → ∫ 0 t ⟨ M ( m θ ) H MFA ( e θ ) , H MFA ( e θ ) ⟩ d s .
Passing to the limit in (116) then shows (103) . The equivalence (101) gives (104) .
Finally, since ∣ n ∣ = 1 |n| = 1 ∣ n ∣ = 1 and ∣ m θ ∣ ≤ 1 |m_\theta| \le1 ∣ m θ ∣ ≤ 1 ,
0 ≤ 1 − ∣ m θ ∣ 2 = ( 1 − ∣ m θ ∣ ) ( 1 + ∣ m θ ∣ ) ≤ 2 ∣ m θ − n ∣ . 0 \le1 - |m_\theta|^2 = (1 - |m_\theta|)(1 + |m_\theta|) \le2|m_\theta- n|. 0 ≤ 1 − ∣ m θ ∣ 2 = ( 1 − ∣ m θ ∣ ) ( 1 + ∣ m θ ∣ ) ≤ 2∣ m θ − n ∣.
Hence, (105) also follows from (104) . □
Remark 4.2. The existence of arbitrarily regular solutions to the Landau–Lifshitz–Gilbert equation for initial data that is sufficiently close to a constant function is proven in [12 ] . In particular, the regularity hypothesis (102) is not vacuous.
Convergence to a stationary state
We begin by identifying the stationary states governing the long-time dynamics. Since every global solution is uniformly separated from saturation on intervals bounded away from t = 0 t = 0 t = 0 , the mobility M ( m ) \mathrm{M}(m) M ( m ) is uniformly positive definite there. Consequently, a stationary solution of the thermodynamic formulation satisfies
0 = γ m × h ( m ) + M ( m ) h ( m ) , 0 = \gamma m \times h(m) + \mathrm{M}(m)h(m), 0 = γ m × h ( m ) + M ( m ) h ( m ) ,
and testing with h ( m ) h(m) h ( m ) yields h ( m ) = 0 h(m) = 0 h ( m ) = 0 . Thus, the stationary states are precisely the critical points of the free energy in the open unit ball. For r ≥ 2 r \ge2 r ≥ 2 , let
H n r ( D ) : = { v ∈ H r ( D ; R 3 ) : ∂ ν v = 0 on ∂ D } , H_n^r(\mathcal{D}) := \left\{v \in H^r(\mathcal{D};\mathbb{R}^3) : \partial_\nu v = 0 \text{ on } \partial\mathcal{D}\right\}, H n r ( D ) := { v ∈ H r ( D ; R 3 ) : ∂ ν v = 0 on ∂ D } ,
and set
S θ : = { m ˉ ∈ H n 2 ( D ) : ∥ m ˉ ∥ L ∞ ( D ) < 1 , h ( m ˉ ) = 0 } . (117) S_\theta:= \left\{\bar{m} \in H_n^2(\mathcal{D}) : \|\bar{m}\|_{L^\infty(\mathcal{D})} < 1,\quad h(\bar{m}) = 0\right\}. \tag*{(117)} S θ := { m ˉ ∈ H n 2 ( D ) : ∥ m ˉ ∥ L ∞ ( D ) < 1 , h ( m ˉ ) = 0 } . (117)
The key tool for proving convergence to a single element of S θ S_\theta S θ is the following local Łojasiewicz–Simon inequality near each stationary state.
Lemma 4.3 (Łojasiewicz–Simon inequality). Let E θ \mathcal{E}_\theta E θ be the energy functional given in (19) . Let m ˉ ∈ S θ \bar{m} \in S_\theta m ˉ ∈ S θ , and choose ρ ∈ ( ∥ m ˉ ∥ L ∞ ( D ) , 1 ) \rho\in(\|\bar{m}\|_{L^\infty(\mathcal{D})},1) ρ ∈ ( ∥ m ˉ ∥ L ∞ ( D ) , 1 ) . Then, there exist constants σ ∈ ( 0 , 1 / 2 ] \sigma\in(0,1/2] σ ∈ ( 0 , 1/2 ] , C > 0 C > 0 C > 0 , and η > 0 \eta> 0 η > 0 , such that
∣ E θ ( v ) − E θ ( m ˉ ) ∣ 1 − σ ≤ C ∥ h ( v ) ∥ L 2 ( D ) (118) |\mathcal{E}_\theta(v) - \mathcal{E}_\theta(\bar{m})|^{1-\sigma} \le C\|h(v)\|_{L^2(\mathcal{D})} \tag*{(118)} ∣ E θ ( v ) − E θ ( m ˉ ) ∣ 1 − σ ≤ C ∥ h ( v ) ∥ L 2 ( D ) (118) for every v ∈ H n 2 ( D ) v \in H_n^2(\mathcal{D}) v ∈ H n 2 ( D ) satisfying ∥ v ∥ L ∞ ( D ) ≤ ρ \|v\|_{L^\infty(\mathcal{D})} \le\rho ∥ v ∥ L ∞ ( D ) ≤ ρ and ∥ v − m ˉ ∥ H 1 ( D ) < η \|v - \bar{m}\|_{H^1(\mathcal{D})} < \eta ∥ v − m ˉ ∥ H 1 ( D ) < η .
Proof. Since H 2 ( D ) ↪ L ∞ ( D ) H^2(\mathcal{D}) \hookrightarrow L^\infty(\mathcal{D}) H 2 ( D ) ↪ L ∞ ( D ) and H 2 ( D ) H^2(\mathcal{D}) H 2 ( D ) is a Banach algebra in three dimensions, a sufficiently small H 2 H^2 H 2 -neighbourhood of m ˉ \bar{m} m ˉ consists of functions taking values in B ρ ( 0 ) B_\rho(0) B ρ ( 0 ) . On this ball, v ↦ Φ ( ∣ v ∣ ) v \mapsto\Phi(|v|) v ↦ Φ ( ∣ v ∣ ) and v ↦ G ( v ) v \mapsto G(v) v ↦ G ( v ) are real analytic. Hence, the associated Nemytskii maps are real analytic on this neighbourhood, and so is E θ : H n 2 ( D ) → R \mathcal{E}_\theta: H_n^2(\mathcal{D}) \to\mathbb{R} E θ : H n 2 ( D ) → R . Its L 2 L^2 L 2 -gradient is
∇ L 2 E θ ( v ) = − h ( v ) . \nabla_{L^2}\mathcal{E}_\theta(v) = -h(v). ∇ L 2 E θ ( v ) = − h ( v ) .
For m ˉ ∈ H n 2 ( D ) \bar{m} \in H_n^2(\mathcal{D}) m ˉ ∈ H n 2 ( D ) , the Hessian is
D 2 E θ ( m ˉ ) [ v , w ] = ⟨ ( − α Δ − B − H d + θ D G ( m ˉ ) ) v , w ⟩ , D^2\mathcal{E}_\theta(\bar{m})[v,w] = \left\langle(-\alpha\Delta- \mathrm{B} - \mathrm{H}_d + \theta DG(\bar{m}))v,w\right\rangle, D 2 E θ ( m ˉ ) [ v , w ] = ⟨ ( − α Δ − B − H d + θ D G ( m ˉ )) v , w ⟩ ,
for v ∈ H n 2 ( D ) v \in H_n^2(\mathcal{D}) v ∈ H n 2 ( D ) and w ∈ L 2 ( D ) w \in L^2(\mathcal{D}) w ∈ L 2 ( D ) . Equivalently, the associated self-adjoint L 2 L^2 L 2 -operator is
L m ˉ : = − α Δ − B − H d + θ D G ( m ˉ ) : H n 2 ( D ) → L 2 ( D ) . \mathcal{L}_{\bar{m}} := -\alpha\Delta- \mathrm{B} - \mathrm{H}_d + \theta DG(\bar{m}) : H_n^2(\mathcal{D}) \to L^2(\mathcal{D}). L m ˉ := − α Δ − B − H d + θ D G ( m ˉ ) : H n 2 ( D ) → L 2 ( D ) .
Moreover, it is a Fredholm operator of index zero: the Neumann Laplacian is Fredholm from H n 2 ( D ) H_n^2(\mathcal{D}) H n 2 ( D ) to L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) , and all remaining terms are compact perturbations as maps from H n 2 ( D ) H_n^2(\mathcal{D}) H n 2 ( D ) to L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) . The abstract Łojasiewicz–Simon theorem [11 ] then yields constants σ ∈ ( 0 , 1 / 2 ] \sigma\in(0,1/2] σ ∈ ( 0 , 1/2 ] , C > 0 C > 0 C > 0 , and δ > 0 \delta> 0 δ > 0 such that
∣ E θ ( v ) − E θ ( m ˉ ) ∣ 1 − σ ≤ C ∥ h ( v ) ∥ L 2 ( D ) |\mathcal{E}_\theta(v) - \mathcal{E}_\theta(\bar{m})|^{1-\sigma} \le C\|h(v)\|_{L^2(\mathcal{D})} ∣ E θ ( v ) − E θ ( m ˉ ) ∣ 1 − σ ≤ C ∥ h ( v ) ∥ L 2 ( D )
whenever ∥ v − m ˉ ∥ H 2 ( D ) < δ \|v - \bar{m}\|_{H^2(\mathcal{D})} < \delta ∥ v − m ˉ ∥ H 2 ( D ) < δ .
It remains to replace the H 2 H^2 H 2 -neighbourhood by an H 1 H^1 H 1 -neighbourhood for functions uniformly separated from the saturation boundary. Since h ( m ˉ ) = 0 h(\bar{m}) = 0 h ( m ˉ ) = 0 , we have
α Δ ( v − m ˉ ) = h ( v ) − B ( v − m ˉ ) − H d [ v − m ˉ ] + θ ( G ( v ) − G ( m ˉ ) ) . \alpha\Delta(v - \bar{m}) = h(v) - \mathrm{B}(v - \bar{m}) - \mathrm{H}_d[v - \bar{m}] + \theta(G(v) - G(\bar{m})). α Δ ( v − m ˉ ) = h ( v ) − B ( v − m ˉ ) − H d [ v − m ˉ ] + θ ( G ( v ) − G ( m ˉ )) .
If ∥ v ∥ L ∞ ( D ) ≤ ρ \|v\|_{L^\infty(\mathcal{D})}\leq\rho ∥ v ∥ L ∞ ( D ) ≤ ρ and ∥ m ‾ ∥ L ∞ ( D ) ≤ ρ < 1 \|\overline{m}\|_{L^\infty(\mathcal{D})}\leq\rho<1 ∥ m ∥ L ∞ ( D ) ≤ ρ < 1 , then G : B ρ ( 0 ) ‾ → R 3 G:\overline{B_\rho(0)}\to\mathbb{R}^3 G : B ρ ( 0 ) → R 3 is Lipschitz, and thus
∥ G ( v ) − G ( m ‾ ) ∥ L 2 ( D ) ≤ C ρ ∥ v − m ‾ ∥ L 2 ( D ) . \|G(v)-G(\overline{m})\|_{L^2(\mathcal{D})}\leq C_\rho\|v-\overline{m}\|_{L^2(\mathcal{D})}. ∥ G ( v ) − G ( m ) ∥ L 2 ( D ) ≤ C ρ ∥ v − m ∥ L 2 ( D ) .
Therefore, the Neumann elliptic estimate and the L 2 L^2 L 2 -boundedness of H d H_d H d yield
∥ v − m ‾ ∥ H 2 ( D ) ≤ C ρ ( ∥ h ( v ) ∥ L 2 ( D ) + ∥ v − m ‾ ∥ L 2 ( D ) ) . (119) \|v-\overline{m}\|_{H^2(\mathcal{D})}\leq C_\rho\left(\|h(v)\|_{L^2(\mathcal{D})}+\|v-\overline{m}\|_{L^2(\mathcal{D})}\right). \tag*{(119)} ∥ v − m ∥ H 2 ( D ) ≤ C ρ ( ∥ h ( v ) ∥ L 2 ( D ) + ∥ v − m ∥ L 2 ( D ) ) . (119)
Now, choose η > 0 \eta>0 η > 0 sufficiently small, such that C ρ η < δ / 2 C_\rho\eta<\delta/2 C ρ η < δ /2 and ∣ E θ ( v ) − E θ ( m ‾ ) ∣ ≤ 1 |\mathcal{E}_\theta(v)-\mathcal{E}_\theta(\overline{m})|\leq1 ∣ E θ ( v ) − E θ ( m ) ∣ ≤ 1 whenever ∥ v − m ‾ ∥ H 1 ( D ) < η \|v-\overline{m}\|_{H^1(\mathcal{D})}<\eta ∥ v − m ∥ H 1 ( D ) < η . If ∥ h ( v ) ∥ L 2 ( D ) < δ / ( 2 C ρ ) \|h(v)\|_{L^2(\mathcal{D})}<\delta/(2C_\rho) ∥ h ( v ) ∥ L 2 ( D ) < δ / ( 2 C ρ ) , then (119) gives ∥ v − m ‾ ∥ H 2 < δ \|v-\overline{m}\|_{H^2}<\delta ∥ v − m ∥ H 2 < δ , and the original Łojasiewicz–Simon inequality applies. If instead ∥ h ( v ) ∥ L 2 ( D ) ≥ δ / ( 2 C ρ ) \|h(v)\|_{L^2(\mathcal{D})}\geq\delta/(2C_\rho) ∥ h ( v ) ∥ L 2 ( D ) ≥ δ / ( 2 C ρ ) , then
∣ E θ ( v ) − E θ ( m ‾ ) ∣ 1 − σ ≤ 1 ≤ 2 C ρ δ ∥ h ( v ) ∥ L 2 ( D ) . |\mathcal{E}_\theta(v)-\mathcal{E}_\theta(\overline{m})|^{1-\sigma}\leq1\leq\frac{2C_\rho}{\delta}\|h(v)\|_{L^2(\mathcal{D})}. ∣ E θ ( v ) − E θ ( m ) ∣ 1 − σ ≤ 1 ≤ δ 2 C ρ ∥ h ( v ) ∥ L 2 ( D ) .
This proves (118) . □
As a consequence of the above Łojasiewicz–Simon inequality, we have the following result on quantitative convergence to an equilibrium.
Theorem 4.4 (Convergence to a single equilibrium). Let m m m be a global weak solution of the full nonlinear LLB equation given by Theorem 3.2 and let S θ \mathcal{S}_\theta S θ be given by (117) . Then, there exists m ∞ ∈ S θ m_\infty\in\mathcal{S}_\theta m ∞ ∈ S θ such that
m ( t ) → m ∞ strongly in H 1 ( D ) as t → ∞ . (120) m(t)\to m_\infty\qquad\text{strongly in }H^1(\mathcal{D})\text{ as }t\to\infty. \tag*{(120)} m ( t ) → m ∞ strongly in H 1 ( D ) as t → ∞. (120) Moreover, there exists σ ∈ ( 0 , 1 / 2 ] \sigma\in(0,1/2] σ ∈ ( 0 , 1/2 ] such that, for all sufficiently large t t t ,
E θ ( m ( t ) ) − E θ ( m ∞ ) ≤ { C e − c t , σ = 1 2 , C ( 1 + t ) − 1 1 − 2 σ , 0 < σ < 1 2 , (121) \mathcal{E}_\theta(m(t))-\mathcal{E}_\theta(m_\infty)\leq
\begin{cases}
Ce^{-ct}, & \sigma=\frac{1}{2},\\
C(1+t)^{-\frac{1}{1-2\sigma}}, & 0<\sigma<\frac{1}{2},
\end{cases} \tag*{(121)} E θ ( m ( t )) − E θ ( m ∞ ) ≤ { C e − c t , C ( 1 + t ) − 1 − 2 σ 1 , σ = 2 1 , 0 < σ < 2 1 , (121) and
∥ m ( t ) − m ∞ ∥ L 2 ( D ) ≤ C [ E θ ( m ( t ) ) − E θ ( m ∞ ) ] σ . (122) \|m(t)-m_\infty\|_{L^2(\mathcal{D})}\leq C[\mathcal{E}_\theta(m(t))-\mathcal{E}_\theta(m_\infty)]^\sigma. \tag*{(122)} ∥ m ( t ) − m ∞ ∥ L 2 ( D ) ≤ C [ E θ ( m ( t )) − E θ ( m ∞ ) ] σ . (122)
Proof. Apply (32) with τ = 1 \tau=1 τ = 1 and write ρ ∈ ( 0 , 1 ) \rho\in(0,1) ρ ∈ ( 0 , 1 ) for the corresponding separation constant, so that ∣ m ( t ) ∣ ≤ ρ |m(t)|\leq\rho ∣ m ( t ) ∣ ≤ ρ a.e. in D \mathcal{D} D for every t ≥ 1 t\geq1 t ≥ 1 . The energy is bounded from below and, by (34) , t ↦ E θ ( m ( t ) ) t\mapsto\mathcal{E}_\theta(m(t)) t ↦ E θ ( m ( t )) is non-increasing on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) . Hence, there exists E ∞ ∈ R E_\infty\in\mathbb{R} E ∞ ∈ R such that E θ ( m ( t ) ) → E ∞ \mathcal{E}_\theta(m(t))\to E_\infty E θ ( m ( t )) → E ∞ as t → ∞ t\to\infty t → ∞ .
Moreover, by (A.3) and the monotonicity of q q q ,
2 q ( ρ ) ∥ h ( m ( t ) ) ∥ L 2 ( D ) 2 ≤ ⟨ M ( m ( t ) ) h ( m ( t ) ) , h ( m ( t ) ) ⟩ . (123) 2q(\rho)\|h(m(t))\|_{L^2(\mathcal{D})}^2\leq\langle M(m(t))h(m(t)),h(m(t))\rangle. \tag*{(123)} 2 q ( ρ ) ∥ h ( m ( t )) ∥ L 2 ( D ) 2 ≤ ⟨ M ( m ( t )) h ( m ( t )) , h ( m ( t ))⟩ . (123)
Using again the energy identity (34) , and the lower boundedness of E θ \mathcal{E}_\theta E θ , we have
∫ 1 ∞ ∥ h ( m ( t ) ) ∥ L 2 ( D ) 2 d t ≤ 1 2 q ( ρ ) ∫ 1 ∞ ⟨ M ( m ( t ) ) h ( m ( t ) ) , h ( m ( t ) ) ⟩ d t = E θ ( m ( 1 ) ) − E ∞ 2 q ( ρ ) < ∞ , (124) \int_1^\infty\|h(m(t))\|_{L^2(\mathcal{D})}^2\,\mathrm{d}t
\leq\frac{1}{2q(\rho)}\int_1^\infty\langle M(m(t))h(m(t)),h(m(t))\rangle\,\mathrm{d}t
=\frac{\mathcal{E}_\theta(m(1))-E_\infty}{2q(\rho)}<\infty, \tag*{(124)} ∫ 1 ∞ ∥ h ( m ( t )) ∥ L 2 ( D ) 2 d t ≤ 2 q ( ρ ) 1 ∫ 1 ∞ ⟨ M ( m ( t )) h ( m ( t )) , h ( m ( t ))⟩ d t = 2 q ( ρ ) E θ ( m ( 1 )) − E ∞ < ∞ , (124)
where E ∞ = lim t → ∞ E θ ( m ( t ) ) E_\infty=\lim_{t\to\infty}\mathcal{E}_\theta(m(t)) E ∞ = lim t → ∞ E θ ( m ( t )) .
Furthermore, noting (1) , we apply (A.4) and (A.5) give
∥ ∂ t m ( t ) ∥ L 2 ( D ) 2 = γ 2 ∥ m ( t ) × h ( m ( t ) ) ∥ L 2 ( D ) 2 + ∥ M ( m ( t ) ) h ( m ( t ) ) ∥ L 2 ( D ) 2 ≤ ( 3 2 γ 2 + 1 ) ⟨ M ( m ( t ) ) h ( m ( t ) ) , h ( m ( t ) ) ⟩ . \|\partial_t m(t)\|_{L^2(\mathcal{D})}^2
=\gamma^2\|m(t)\times h(m(t))\|_{L^2(\mathcal{D})}^2+\|M(m(t))h(m(t))\|_{L^2(\mathcal{D})}^2
\leq\left(\frac{3}{2}\gamma^2+1\right)\langle M(m(t))h(m(t)),h(m(t))\rangle. ∥ ∂ t m ( t ) ∥ L 2 ( D ) 2 = γ 2 ∥ m ( t ) × h ( m ( t )) ∥ L 2 ( D ) 2 + ∥ M ( m ( t )) h ( m ( t )) ∥ L 2 ( D ) 2 ≤ ( 2 3 γ 2 + 1 ) ⟨ M ( m ( t )) h ( m ( t )) , h ( m ( t ))⟩ .
Consequently, we infer that
∫ 1 ∞ ∥ ∂ t m ( t ) ∥ L 2 ( D ) 2 d t < ∞ . (125) \int_1^\infty\|\partial_t m(t)\|_{L^2(\mathcal{D})}^2\,\mathrm{d}t<\infty. \tag*{(125)} ∫ 1 ∞ ∥ ∂ t m ( t ) ∥ L 2 ( D ) 2 d t < ∞. (125)
We next prove relative compactness of the trajectory in H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) . Let t j → ∞ t_j \to\infty t j → ∞ . From (124) , we may choose s j ∈ [ t j , t j + 1 ] s_j \in[t_j,t_j+1] s j ∈ [ t j , t j + 1 ] such that ∥ h ( m ( s j ) ) ∥ L 2 ( D ) → 0 \lVert h(m(s_j))\rVert_{L^2(\mathcal{D})} \to0 ∥ h ( m ( s j )) ∥ L 2 ( D ) → 0 . The sequence m ( s j ) m(s_j) m ( s j ) is bounded in H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) , so after passing to a subsequence,
m ( s j ) ⇀ m ∗ weakly in H 1 ( D ) , and m ( s j ) → m ∗ strongly in L 2 ( D ) . m(s_j) \rightharpoonup m_* \quad\text{weakly in } H^1(\mathcal{D}), \qquad\text{and} \qquad m(s_j) \to m_* \quad\text{strongly in } L^2(\mathcal{D}). m ( s j ) ⇀ m ∗ weakly in H 1 ( D ) , and m ( s j ) → m ∗ strongly in L 2 ( D ) .
Since G G G is Lipschitz on { ∣ v ∣ ≤ ρ } \{|v| \le\rho\} { ∣ v ∣ ≤ ρ } , we deduce G ( m ( s j ) ) → G ( m ∗ ) G(m(s_j)) \to G(m_*) G ( m ( s j )) → G ( m ∗ ) in L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) . The identity
α Δ m ( s j ) = h ( m ( s j ) ) − B m ( s j ) − H d [ m ( s j ) ] + θ G ( m ( s j ) ) \alpha\Delta m(s_j) = h(m(s_j)) - Bm(s_j) - H_{\mathrm d}[m(s_j)] + \theta G(m(s_j)) α Δ m ( s j ) = h ( m ( s j )) − B m ( s j ) − H d [ m ( s j )] + θ G ( m ( s j ))
then shows that Δ m ( s j ) \Delta m(s_j) Δ m ( s j ) converges strongly in L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) . By the Neumann elliptic estimate,
m ( s j ) → m ∗ strongly in H 2 ( D ) . m(s_j) \to m_* \quad\text{strongly in } H^2(\mathcal{D}). m ( s j ) → m ∗ strongly in H 2 ( D ) .
In particular, h ( m ∗ ) = 0 h(m_*) = 0 h ( m ∗ ) = 0 , thus m ∗ ∈ S θ m_* \in S_\theta m ∗ ∈ S θ . Furthermore, by (125) , as j → ∞ j \to\infty j → ∞ ,
∥ m ( t j ) − m ( s j ) ∥ L 2 ( D ) ≤ ( ∫ t j t j + 1 ∥ ∂ t m ( t ) ∥ L 2 ( D ) 2 d t ) 1 / 2 → 0. \lVert m(t_j)-m(s_j)\rVert_{L^2(\mathcal{D})}
\le
\left(\int_{t_j}^{t_j+1}\lVert\partial_t m(t)\rVert_{L^2(\mathcal{D})}^2\,\mathrm{d}t\right)^{1/2}
\to0. ∥ m ( t j ) − m ( s j ) ∥ L 2 ( D ) ≤ ( ∫ t j t j + 1 ∥ ∂ t m ( t ) ∥ L 2 ( D ) 2 d t ) 1/2 → 0.
Hence, m ( t j ) → m ∗ m(t_j) \to m_* m ( t j ) → m ∗ strongly in L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) and weakly in H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) . Since E θ ( m ( t j ) ) → E ∞ = E θ ( m ∗ ) \mathcal{E}_\theta(m(t_j)) \to\mathcal{E}_\infty= \mathcal{E}_\theta(m_*) E θ ( m ( t j )) → E ∞ = E θ ( m ∗ ) , and all terms in the energy except the exchange term converge by the strong L 2 L^2 L 2 -convergence and the uniform separation, we obtain ∥ ∇ m ( t j ) ∥ L 2 ( D ) → ∥ ∇ m ∗ ∥ L 2 ( D ) \lVert\nabla m(t_j)\rVert_{L^2(\mathcal{D})} \to\lVert\nabla m_*\rVert_{L^2(\mathcal{D})} ∥ ∇ m ( t j ) ∥ L 2 ( D ) → ∥ ∇ m ∗ ∥ L 2 ( D ) . Thus
m ( t j ) → m ∗ strongly in H 1 ( D ) . m(t_j) \to m_* \quad\text{strongly in } H^1(\mathcal{D}). m ( t j ) → m ∗ strongly in H 1 ( D ) .
The trajectory is therefore relatively compact in H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) , and every element of its ω \omega ω -limit set belongs to S θ S_\theta S θ and has energy E ∞ \mathcal{E}_\infty E ∞ .
The ω \omega ω -limit set is compact in H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) . A finite collection of the neighbourhoods from Lemma 4.3 covers it. Taking the smallest Łojasiewicz exponent and enlarging the constant if necessary, we obtain numbers σ ∈ ( 0 , 1 / 2 ] \sigma\in(0,1/2] σ ∈ ( 0 , 1/2 ] , C > 0 C>0 C > 0 , t 0 > 0 t_0>0 t 0 > 0 , such that
[ E θ ( m ( t ) ) − E ∞ ] 1 − σ ≤ C ∥ h ( m ( t ) ) ∥ L 2 ( D ) (126) [\mathcal{E}_\theta(m(t))-\mathcal{E}_\infty]^{1-\sigma}
\le C\lVert h(m(t))\rVert_{L^2(\mathcal{D})} \tag*{(126)} [ E θ ( m ( t )) − E ∞ ] 1 − σ ≤ C ∥ h ( m ( t )) ∥ L 2 ( D ) (126)
for a.e. t ≥ t 0 t \ge t_0 t ≥ t 0 .
Now, set
e ( t ) : = E θ ( m ( t ) ) − E ∞ . e(t):=\mathcal{E}_\theta(m(t))-\mathcal{E}_\infty. e ( t ) := E θ ( m ( t )) − E ∞ .
By the energy identity, the separation estimate (A.3), and (126) ,
− e ′ ( t ) = ⟨ M ( m ( t ) ) h ( m ( t ) ) , h ( m ( t ) ) ⟩ ≥ 2 q ( ρ ) ∥ h ( m ( t ) ) ∥ L 2 ( D ) 2 ≥ c e ( t ) 2 ( 1 − σ ) (127) -e'(t)=\langle M(m(t))h(m(t)),h(m(t))\rangle
\ge2q(\rho)\lVert h(m(t))\rVert_{L^2(\mathcal{D})}^2
\ge c e(t)^{2(1-\sigma)} \tag*{(127)} − e ′ ( t ) = ⟨ M ( m ( t )) h ( m ( t )) , h ( m ( t ))⟩ ≥ 2 q ( ρ ) ∥ h ( m ( t )) ∥ L 2 ( D ) 2 ≥ ce ( t ) 2 ( 1 − σ ) (127)
for a.e. t ≥ t 0 t \ge t_0 t ≥ t 0 . Solving this scalar differential inequality gives (121) .
Moreover, we obtain
− d d t e ( t ) σ = σ e ( t ) σ − 1 ⟨ M ( m ( t ) ) h ( m ( t ) ) , h ( m ( t ) ) ⟩ ≥ c e ( t ) σ − 1 ∥ h ( m ( t ) ) ∥ L 2 ( D ) 2 ≥ c ∥ h ( m ( t ) ) ∥ L 2 ( D ) ≥ c ∥ ∂ t m ( t ) ∥ L 2 ( D ) . \begin{aligned}
-\frac{\mathrm{d}}{\mathrm{d}t}e(t)^\sigma
=\sigma e(t)^{\sigma-1}\langle M(m(t))h(m(t)),h(m(t))\rangle\\
&\ge c e(t)^{\sigma-1}\lVert h(m(t))\rVert_{L^2(\mathcal{D})}^2 \\
&\ge c\lVert h(m(t))\rVert_{L^2(\mathcal{D})}
\ge c\lVert\partial_t m(t)\rVert_{L^2(\mathcal{D})}.
\end{aligned} − d t d e ( t ) σ = σ e ( t ) σ − 1 ⟨ M ( m ( t )) h ( m ( t )) , h ( m ( t ))⟩ ≥ ce ( t ) σ − 1 ∥ h ( m ( t )) ∥ L 2 ( D ) 2 ≥ c ∥ h ( m ( t )) ∥ L 2 ( D ) ≥ c ∥ ∂ t m ( t ) ∥ L 2 ( D ) .
Here, the penultimate inequality follows from (126) . Integrating over ( t , ∞ ) (t,\infty) ( t , ∞ ) gives
∫ t ∞ ∥ ∂ t m ( s ) ∥ L 2 ( D ) d s ≤ C e ( t ) σ . \int_t^\infty\lVert\partial_t m(s)\rVert_{L^2(\mathcal{D})}\,\mathrm{d}s
\le C e(t)^\sigma. ∫ t ∞ ∥ ∂ t m ( s ) ∥ L 2 ( D ) d s ≤ C e ( t ) σ .
Hence, m ( t ) m(t) m ( t ) is Cauchy in L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) and converges to a single limit m ∞ ∈ S θ m_\infty\in S_\theta m ∞ ∈ S θ , with (122) . Relative compactness in H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) then upgrades the convergence to (120) . □
A Curie-type threshold
In this section, we discuss how temperature affects the stability of the paramagnetic equilibrium state m = 0 m = 0 m = 0 . Since
Φ ( s ) = 3 2 s 2 + O ( s 4 ) as s → 0 , (128) \Phi(s) = \frac{3}{2}s^2 + O(s^4) \quad\text{as } s \to0, \tag*{(128)} Φ ( s ) = 2 3 s 2 + O ( s 4 ) as s → 0 , (128)
the quadratic part of the energy at the origin is
D 2 E θ ( 0 ) [ v , v ] = α ∥ ∇ v ∥ L 2 ( D ) 2 − ⟨ B v , v ⟩ + ∥ H d [ v ] ∥ L 2 ( R 3 ) 2 + 3 θ ∥ v ∥ L 2 ( D ) 2 . D^2\mathcal{E}_\theta(0)[v,v] = \alpha\|\nabla v\|_{L^2(\mathcal{D})}^2 - \langle Bv,v\rangle+ \|H_d[v]\|_{L^2(\mathbb{R}^3)}^2 + 3\theta\|v\|_{L^2(\mathcal{D})}^2. D 2 E θ ( 0 ) [ v , v ] = α ∥∇ v ∥ L 2 ( D ) 2 − ⟨ B v , v ⟩ + ∥ H d [ v ] ∥ L 2 ( R 3 ) 2 + 3 θ ∥ v ∥ L 2 ( D ) 2 .
This motivates introducing the spectral quantity
Λ ∗ : = sup 0 ≠ v ∈ H 1 ( D ) ⟨ B v , v ⟩ − α ∥ ∇ v ∥ L 2 ( D ) 2 − ∥ H d [ v ] ∥ L 2 ( R 3 ) 2 ∥ v ∥ L 2 ( D ) 2 . (129) \Lambda_* := \sup_{0 \ne v \in H^1(\mathcal{D})} \frac{\langle Bv,v\rangle- \alpha\|\nabla v\|_{L^2(\mathcal{D})}^2 - \|H_d[v]\|_{L^2(\mathbb{R}^3)}^2}{\|v\|_{L^2(\mathcal{D})}^2}. \tag*{(129)} Λ ∗ := 0 = v ∈ H 1 ( D ) sup ∥ v ∥ L 2 ( D ) 2 ⟨ B v , v ⟩ − α ∥∇ v ∥ L 2 ( D ) 2 − ∥ H d [ v ] ∥ L 2 ( R 3 ) 2 . (129)
The quantity Λ ∗ \Lambda_* Λ ∗ is finite: Λ ∗ ≤ ∥ B ∥ \Lambda_* \le\|B\| Λ ∗ ≤ ∥ B ∥ . Furthermore, by definition,
D 2 E θ ( 0 ) [ v , v ] ≥ ( 3 θ − Λ ∗ ) ∥ v ∥ L 2 ( D ) 2 . D^2\mathcal{E}_\theta(0)[v,v] \ge(3\theta- \Lambda_*)\|v\|_{L^2(\mathcal{D})}^2. D 2 E θ ( 0 ) [ v , v ] ≥ ( 3 θ − Λ ∗ ) ∥ v ∥ L 2 ( D ) 2 .
Hence, when Λ ∗ > 0 \Lambda_* > 0 Λ ∗ > 0 , the value θ c : = Λ ∗ / 3 \theta_c := \Lambda_*/3 θ c := Λ ∗ /3 plays the role of a Curie-type threshold : above it the paramagnetic state m = 0 m = 0 m = 0 is a strict local energy minimizer with an L 2 L^2 L 2 -coercive Hessian, whereas below it the quadratic energy possesses an unstable direction. If Λ ∗ ≤ 0 \Lambda_* \le0 Λ ∗ ≤ 0 , then the Hessian at m = 0 m = 0 m = 0 is coercive throughout all θ > 0 \theta> 0 θ > 0 .
We remark that this threshold should be distinguished from the bulk mean-field value θ c b u l k = J / 3 \theta_c^{\mathrm{bulk}} = J/3 θ c bulk = J /3 introduced in § 2.5 , which is used there only in the asymptotic reduction to the standard LLB model. Indeed, writing B = J I + A B = JI + A B = J I + A , we have
Λ ∗ = J + sup 0 ≠ v ∈ H 1 ( D ) ⟨ A v , v ⟩ − α ∥ ∇ v ∥ L 2 ( D ) 2 − ∥ H d [ v ] ∥ L 2 ( R 3 ) 2 ∥ v ∥ L 2 ( D ) 2 . \Lambda_* = J + \sup_{0 \ne v \in H^1(\mathcal{D})} \frac{\langle Av,v\rangle- \alpha\|\nabla v\|_{L^2(\mathcal{D})}^2 - \|H_d[v]\|_{L^2(\mathbb{R}^3)}^2}{\|v\|_{L^2(\mathcal{D})}^2}. Λ ∗ = J + 0 = v ∈ H 1 ( D ) sup ∥ v ∥ L 2 ( D ) 2 ⟨ A v , v ⟩ − α ∥∇ v ∥ L 2 ( D ) 2 − ∥ H d [ v ] ∥ L 2 ( R 3 ) 2 .
Thus, J / 3 J/3 J /3 is the Curie threshold of the homogeneous bulk mean-field reference problem, whereas Λ ∗ / 3 \Lambda_*/3 Λ ∗ /3 is the corresponding sample-dependent stability threshold for the full micromagnetic PDE, including exchange, anisotropy, stray-field effects, and the geometry of D \mathcal{D} D . In the homogeneous bulk reduction, where these micromagnetic corrections are neglected, Λ ∗ \Lambda_* Λ ∗ reduces to J J J and the two thresholds coincide.
The following theorem formulates this more precisely. Above the critical temperature θ c \theta_c θ c , the only stationary state is 0 0 0 and every solution converges to it, whereas below the threshold the energy admits nonzero states of lower energy.
Theorem 4.5 (Characterisation relative to the Curie threshold). Let m m m be a global weak solution of the full nonlinear LLB equation given by Theorem 3.2 , and let θ c = Λ ∗ / 3 \theta_c = \Lambda_*/3 θ c = Λ ∗ /3 . Then, the following assertions hold.
(i) If θ ≥ θ c \theta\ge\theta_c θ ≥ θ c , then
S θ = { 0 } . S_\theta= \{0\}. S θ = { 0 } . Consequently, every global weak solution converges strongly in H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) to zero.
(ii) If θ > θ c \theta> \theta_c θ > θ c , then the convergence is exponential: there exist C , ω > 0 C,\omega> 0 C , ω > 0 such that
∥ m ( t ) ∥ H 1 ( D ) ≤ C e − ω t \|m(t)\|_{H^1(\mathcal{D})} \le Ce^{-\omega t} ∥ m ( t ) ∥ H 1 ( D ) ≤ C e − ω t for all sufficiently large t t t .
(iii) If θ < θ c \theta< \theta_c θ < θ c , then zero is not a minimiser of E θ \mathcal{E}_\theta E θ . More precisely, there exists v ∈ C ∞ ( D ‾ ) v \in C^\infty(\overline{\mathcal{D}}) v ∈ C ∞ ( D ) satisfying ∂ ν v = 0 \partial_\nu v = 0 ∂ ν v = 0 on ∂ D \partial\mathcal{D} ∂ D and, for all sufficiently small δ > 0 \delta> 0 δ > 0 ,
∥ δ v ∥ L ∞ ( D ) < 1 , E θ ( δ v ) < 0. \|\delta v\|_{L^\infty(\mathcal{D})} < 1, \qquad\mathcal{E}_\theta(\delta v) < 0. ∥ δ v ∥ L ∞ ( D ) < 1 , E θ ( δ v ) < 0. In particular, any trajectory satisfying E θ ( m ( t 0 ) ) < 0 \mathcal{E}_\theta(m(t_0)) < 0 E θ ( m ( t 0 )) < 0 at some t 0 > 0 t_0 > 0 t 0 > 0 converges to a nonzero stationary state.
Proof. Let m ˉ ∈ S θ \bar{m} \in S_{\theta} m ˉ ∈ S θ . Testing h ( m ˉ ) = 0 h(\bar{m}) = 0 h ( m ˉ ) = 0 by m ˉ \bar{m} m ˉ gives
0 = − α ∥ ∇ m ˉ ∥ L 2 ( D ) 2 + ⟨ B m ˉ , m ˉ ⟩ − ∥ H d [ m ˉ ] ∥ L 2 ( R 3 ) 2 − θ ∫ D ∣ m ˉ ∣ 2 q ( ∣ m ˉ ∣ ) d x . (130) 0 = -\alpha\|\nabla\bar{m}\|_{L^2(\mathcal{D})}^2 + \langle B\bar{m},\bar{m}\rangle- \|H_d[\bar{m}]\|_{L^2(\mathbb{R}^3)}^2 - \theta\int_{\mathcal{D}} \frac{|\bar{m}|^2}{q(|\bar{m}|)}\,\mathrm{d}x . \tag*{(130)} 0 = − α ∥∇ m ˉ ∥ L 2 ( D ) 2 + ⟨ B m ˉ , m ˉ ⟩ − ∥ H d [ m ˉ ] ∥ L 2 ( R 3 ) 2 − θ ∫ D q ( ∣ m ˉ ∣ ) ∣ m ˉ ∣ 2 d x . (130)
By the definition of Λ ∗ \Lambda_* Λ ∗ and the monotonicity of q q q ,
1 q ( s ) ≥ 3 for s ∈ [ 0 , 1 ) , \frac{1}{q(s)} \ge3 \qquad\text{for } s \in[0,1), q ( s ) 1 ≥ 3 for s ∈ [ 0 , 1 ) ,
with strict inequality for s > 0 s > 0 s > 0 . Hence
0 ≤ Λ ∗ ∥ m ˉ ∥ L 2 ( D ) 2 − θ ∫ D ∣ m ˉ ∣ 2 q ( ∣ m ˉ ∣ ) d x ≤ ( Λ ∗ − 3 θ ) ∥ m ˉ ∥ L 2 ( D ) 2 . 0 \le\Lambda_* \|\bar{m}\|_{L^2(\mathcal{D})}^2 - \theta\int_{\mathcal{D}} \frac{|\bar{m}|^2}{q(|\bar{m}|)}\,\mathrm{d}x \le(\Lambda_* - 3\theta)\|\bar{m}\|_{L^2(\mathcal{D})}^2 . 0 ≤ Λ ∗ ∥ m ˉ ∥ L 2 ( D ) 2 − θ ∫ D q ( ∣ m ˉ ∣ ) ∣ m ˉ ∣ 2 d x ≤ ( Λ ∗ − 3 θ ) ∥ m ˉ ∥ L 2 ( D ) 2 .
If 3 θ > Λ ∗ 3\theta> \Lambda_* 3 θ > Λ ∗ , this forces m ˉ = 0 \bar{m} = 0 m ˉ = 0 . If 3 θ = Λ ∗ 3\theta= \Lambda_* 3 θ = Λ ∗ , then equality would require q ( ∣ m ˉ ∣ ) = 1 / 3 q(|\bar{m}|) = 1/3 q ( ∣ m ˉ ∣ ) = 1/3 a.e. on { ∣ m ˉ ∣ > 0 } \{|\bar{m}| > 0\} { ∣ m ˉ ∣ > 0 } , which is impossible because q ( s ) < 1 / 3 q(s) < 1/3 q ( s ) < 1/3 for s > 0 s > 0 s > 0 . Thus, m ˉ = 0 \bar{m} = 0 m ˉ = 0 also at the critical value. Assertion (i) follows from Theorem 4.4 .
Next, assume that θ > θ c \theta> \theta_c θ > θ c and let μ : = 3 θ − Λ ∗ > 0 \mu:= 3\theta- \Lambda_* > 0 μ := 3 θ − Λ ∗ > 0 . We prove exponential convergence in three steps. First, the strict inequality 3 θ > Λ ∗ 3\theta> \Lambda_* 3 θ > Λ ∗ gives a gradient–energy estimate which controls the energy by h ( m ) h(m) h ( m ) . Second, the same condition implies that the energy controls the full H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) -norm. Finally, positive-time separation makes the mobility uniformly coercive, so that the energy identity yields a closed differential inequality for E θ ( m ( t ) ) E_\theta(m(t)) E θ ( m ( t )) .
We begin with the first point. Recall that a differentiable functional F \mathcal{F} F on a convex subset of a Hilbert space is called μ \mu μ -strongly convex with respect to ∥ ⋅ ∥ \|\cdot\| ∥ ⋅ ∥ if
F ( v ) ≥ F ( u ) + D F ( u ) [ v − u ] + μ 2 ∥ v − u ∥ 2 \mathcal{F}(v) \ge\mathcal{F}(u) + D\mathcal{F}(u)[v-u] + \frac{\mu}{2}\|v-u\|^2 F ( v ) ≥ F ( u ) + D F ( u ) [ v − u ] + 2 μ ∥ v − u ∥ 2
for all admissible u u u and v v v ; see, e.g., [5 Chapter 10] . For a twice-differentiable functional, a sufficient condition is the uniform Hessian bound D 2 F ( u ) [ z , z ] ≥ μ ∥ z ∥ 2 D^2\mathcal{F}(u)[z,z] \ge\mu\|z\|^2 D 2 F ( u ) [ z , z ] ≥ μ ∥ z ∥ 2 . For completeness, we verify this bound directly for E θ E_\theta E θ .
Set ϕ ( v ) : = Φ ( ∣ v ∣ ) \phi(v) := \Phi(|v|) ϕ ( v ) := Φ ( ∣ v ∣ ) for ∣ v ∣ < 1 |v| < 1 ∣ v ∣ < 1 . For s = ∣ v ∣ ∈ ( 0 , 1 ) s = |v| \in(0,1) s = ∣ v ∣ ∈ ( 0 , 1 ) , the Hessian of ϕ \phi ϕ has tangential eigenvalue
L − 1 ( s ) s = 1 q ( s ) ≥ 3 , \frac{\mathcal{L}^{-1}(s)}{s} = \frac{1}{q(s)} \ge3, s L − 1 ( s ) = q ( s ) 1 ≥ 3 ,
since q ( s ) ≤ q ( 0 ) = 1 / 3 q(s) \le q(0) = 1/3 q ( s ) ≤ q ( 0 ) = 1/3 . Its radial eigenvalue is ( L − 1 ) ′ ( s ) (\mathcal{L}^{-1})'(s) ( L − 1 ) ′ ( s ) . Since q q q is decreasing,
0 ≥ q ′ ( s ) = L − 1 ( s ) − s ( L − 1 ) ′ ( s ) ( L − 1 ( s ) ) 2 , 0 \ge q'(s) = \frac{\mathcal{L}^{-1}(s) - s(\mathcal{L}^{-1})'(s)}{(\mathcal{L}^{-1}(s))^2}, 0 ≥ q ′ ( s ) = ( L − 1 ( s ) ) 2 L − 1 ( s ) − s ( L − 1 ) ′ ( s ) ,
and hence
( L − 1 ) ′ ( s ) ≥ L − 1 ( s ) s ≥ 3. (\mathcal{L}^{-1})'(s) \ge\frac{\mathcal{L}^{-1}(s)}{s} \ge3. ( L − 1 ) ′ ( s ) ≥ s L − 1 ( s ) ≥ 3.
At s = 0 s = 0 s = 0 , both eigenvalues converge to 3 3 3 by the expansion of L − 1 \mathcal{L}^{-1} L − 1 . Thus,
D 2 ϕ ( v ) [ z , z ] ≥ 3 ∣ z ∣ 2 for ∣ v ∣ < 1. D^2\phi(v)[z,z] \ge3|z|^2 \qquad\text{for } |v| < 1. D 2 ϕ ( v ) [ z , z ] ≥ 3∣ z ∣ 2 for ∣ v ∣ < 1.
Consequently,
D 2 E θ ( v ) [ z , z ] = α ∥ ∇ z ∥ L 2 ( D ) 2 − ⟨ B z , z ⟩ + ∥ H d [ z ] ∥ L 2 ( R 3 ) 2 + θ ∫ D D 2 ϕ ( v ) [ z , z ] d x ≥ α ∥ ∇ z ∥ L 2 ( D ) 2 − ⟨ B z , z ⟩ + ∥ H d [ z ] ∥ L 2 ( R 3 ) 2 + 3 θ ∥ z ∥ L 2 ( D ) 2 ≥ ( 3 θ − Λ ∗ ) ∥ z ∥ L 2 ( D ) 2 = μ ∥ z ∥ L 2 ( D ) 2 , \begin{aligned}
D^2E_\theta(v)[z,z]
&= \alpha\|\nabla z\|_{L^2(\mathcal{D})}^2 - \langle Bz,z\rangle+ \|H_d[z]\|_{L^2(\mathbb{R}^3)}^2 + \theta\int_{\mathcal{D}} D^2\phi(v)[z,z]\,\mathrm{d}x \\
&\ge\alpha\|\nabla z\|_{L^2(\mathcal{D})}^2 - \langle Bz,z\rangle+ \|H_d[z]\|_{L^2(\mathbb{R}^3)}^2 + 3\theta\|z\|_{L^2(\mathcal{D})}^2 \\
&\ge(3\theta-\Lambda_*)\|z\|_{L^2(\mathcal{D})}^2 = \mu\|z\|_{L^2(\mathcal{D})}^2,
\end{aligned} D 2 E θ ( v ) [ z , z ] = α ∥∇ z ∥ L 2 ( D ) 2 − ⟨ B z , z ⟩ + ∥ H d [ z ] ∥ L 2 ( R 3 ) 2 + θ ∫ D D 2 ϕ ( v ) [ z , z ] d x ≥ α ∥∇ z ∥ L 2 ( D ) 2 − ⟨ B z , z ⟩ + ∥ H d [ z ] ∥ L 2 ( R 3 ) 2 + 3 θ ∥ z ∥ L 2 ( D ) 2 ≥ ( 3 θ − Λ ∗ ) ∥ z ∥ L 2 ( D ) 2 = μ ∥ z ∥ L 2 ( D ) 2 ,
where the last inequality follows from (129) . Hence E θ E_\theta E θ is μ \mu μ -strongly convex with respect to the L 2 ( D ) L^2(\mathcal{D}) L 2 ( D ) -norm.
We now use this property only to relate the energy to its variational derivative. To make the argument explicit, for fixed m m m consider
f ( r ) : = E θ ( ( 1 − r ) m ) , 0 ≤ r ≤ 1. f(r) := \mathcal{E}_{\theta}((1-r)m), \qquad0 \le r \le1. f ( r ) := E θ (( 1 − r ) m ) , 0 ≤ r ≤ 1.
The segment { ( 1 − r ) m : 0 ≤ r ≤ 1 } \{(1-r)m : 0 \le r \le1\} {( 1 − r ) m : 0 ≤ r ≤ 1 } remains in the open unit ball whenever m m m is uniformly separated from saturation, and
f ′ ′ ( r ) = D 2 E θ ( ( 1 − r ) m ) [ m , m ] ≥ μ ∥ m ∥ L 2 ( D ) 2 . f''(r) = D^2\mathcal{E}_{\theta}((1-r)m)[m,m] \ge\mu\lVert m\rVert_{L^2(\mathcal{D})}^2. f ′′ ( r ) = D 2 E θ (( 1 − r ) m ) [ m , m ] ≥ μ ∥ m ∥ L 2 ( D ) 2 .
Taylor’s formula with integral remainder gives
f ( 1 ) ≥ f ( 0 ) + f ′ ( 0 ) + μ 2 ∥ m ∥ L 2 ( D ) 2 . f(1) \ge f(0) + f'(0) + \frac{\mu}{2}\lVert m\rVert_{L^2(\mathcal{D})}^2. f ( 1 ) ≥ f ( 0 ) + f ′ ( 0 ) + 2 μ ∥ m ∥ L 2 ( D ) 2 .
Since E θ ( 0 ) = 0 \mathcal{E}_{\theta}(0)=0 E θ ( 0 ) = 0 and D E θ ( m ) = − h ( m ) D\mathcal{E}_{\theta}(m)=-h(m) D E θ ( m ) = − h ( m ) , we have f ′ ( 0 ) = D E θ ( m ) [ − m ] = ⟨ h ( m ) , m ⟩ f'(0)=D\mathcal{E}_{\theta}(m)[-m]=\langle h(m),m\rangle f ′ ( 0 ) = D E θ ( m ) [ − m ] = ⟨ h ( m ) , m ⟩ . Therefore,
0 ≥ E θ ( m ) + ⟨ h ( m ) , m ⟩ + μ 2 ∥ m ∥ L 2 ( D ) 2 ≥ E θ ( m ) − 1 2 μ ∥ h ( m ) ∥ L 2 ( D ) 2 − μ 2 ∥ m ∥ L 2 ( D ) 2 + μ 2 ∥ m ∥ L 2 ( D ) 2 , \begin{aligned}
0 \ge\mathcal{E}_{\theta}(m) + \langle h(m),m\rangle+ \frac{\mu}{2}\lVert m\rVert_{L^2(\mathcal{D})}^2 \\
&\ge\mathcal{E}_{\theta}(m) - \frac{1}{2\mu}\lVert h(m)\rVert_{L^2(\mathcal{D})}^2 - \frac{\mu}{2}\lVert m\rVert_{L^2(\mathcal{D})}^2 + \frac{\mu}{2}\lVert m\rVert_{L^2(\mathcal{D})}^2,
\end{aligned} 0 ≥ E θ ( m ) + ⟨ h ( m ) , m ⟩ + 2 μ ∥ m ∥ L 2 ( D ) 2 ≥ E θ ( m ) − 2 μ 1 ∥ h ( m ) ∥ L 2 ( D ) 2 − 2 μ ∥ m ∥ L 2 ( D ) 2 + 2 μ ∥ m ∥ L 2 ( D ) 2 ,
where we used Young’s inequality. We deduce that
E θ ( m ) ≤ 1 2 μ ∥ h ( m ) ∥ L 2 ( D ) 2 . (131) \mathcal{E}_{\theta}(m) \le\frac{1}{2\mu}\lVert h(m)\rVert_{L^2(\mathcal{D})}^2. \tag*{(131)} E θ ( m ) ≤ 2 μ 1 ∥ h ( m ) ∥ L 2 ( D ) 2 . (131)
This is the estimate that will turn the energy dissipation into an exponential decay inequality. We next show that E θ \mathcal{E}_{\theta} E θ controls the full H 1 ( D ) H^1(\mathcal{D}) H 1 ( D ) -norm. Since L − 1 ( s ) ≥ 3 s \mathcal{L}^{-1}(s) \ge3s L − 1 ( s ) ≥ 3 s , we have
Φ ( s ) = ∫ 0 s L − 1 ( σ ) d σ ≥ 3 2 s 2 , 0 ≤ s < 1. \Phi(s)=\int_0^s \mathcal{L}^{-1}(\sigma)\,d\sigma\ge\frac{3}{2}s^2,\qquad0\le s<1. Φ ( s ) = ∫ 0 s L − 1 ( σ ) d σ ≥ 2 3 s 2 , 0 ≤ s < 1.
Therefore,
2 E θ ( m ) ≥ α ∥ ∇ m ∥ L 2 ( D ) 2 − ⟨ B m , m ⟩ + ∥ H d [ m ] ∥ L 2 ( R 3 ) 2 + 3 θ ∥ m ∥ L 2 ( D ) 2 ≥ ( 3 θ − Λ ∗ ) ∥ m ∥ L 2 ( D ) 2 = μ ∥ m ∥ L 2 ( D ) 2 , \begin{aligned}
2\mathcal{E}_{\theta}(m) \ge\alpha\lVert\nabla m\rVert_{L^2(\mathcal{D})}^2-\langle Bm,m\rangle+\lVert H_d[m]\rVert_{L^2(\mathbb{R}^3)}^2+3\theta\lVert m\rVert_{L^2(\mathcal{D})}^2 \\
&\ge(3\theta-\Lambda_*)\lVert m\rVert_{L^2(\mathcal{D})}^2=\mu\lVert m\rVert_{L^2(\mathcal{D})}^2,
\end{aligned} 2 E θ ( m ) ≥ α ∥ ∇ m ∥ L 2 ( D ) 2 − ⟨ B m , m ⟩ + ∥ H d [ m ] ∥ L 2 ( R 3 ) 2 + 3 θ ∥ m ∥ L 2 ( D ) 2 ≥ ( 3 θ − Λ ∗ ) ∥ m ∥ L 2 ( D ) 2 = μ ∥ m ∥ L 2 ( D ) 2 ,
and consequently,
∥ m ∥ L 2 ( D ) 2 ≤ 2 μ E θ ( m ) . (132) \lVert m\rVert_{L^2(\mathcal{D})}^2\le\frac{2}{\mu}\mathcal{E}_{\theta}(m). \tag*{(132)} ∥ m ∥ L 2 ( D ) 2 ≤ μ 2 E θ ( m ) . (132)
On the other hand, after discarding the nonnegative stray-field and potential terms,
2 E θ ( m ) ≥ α ∥ ∇ m ∥ L 2 ( D ) 2 − ∥ B ∥ ∥ m ∥ L 2 ( D ) 2 . 2\mathcal{E}_{\theta}(m)\ge\alpha\lVert\nabla m\rVert_{L^2(\mathcal{D})}^2-\lVert B\rVert\lVert m\rVert_{L^2(\mathcal{D})}^2. 2 E θ ( m ) ≥ α ∥ ∇ m ∥ L 2 ( D ) 2 − ∥ B ∥ ∥ m ∥ L 2 ( D ) 2 .
Using (132) and rearranging the terms, we obtain
E θ ( m ) ≥ α μ 2 ( μ + ∥ B ∥ ) ∥ ∇ m ∥ L 2 ( D ) 2 . \mathcal{E}_{\theta}(m)\ge\frac{\alpha\mu}{2(\mu+\lVert B\rVert)}\lVert\nabla m\rVert_{L^2(\mathcal{D})}^2. E θ ( m ) ≥ 2 ( μ + ∥ B ∥) α μ ∥ ∇ m ∥ L 2 ( D ) 2 .
Together with (132) , this implies that there exists c θ > 0 c_{\theta}>0 c θ > 0 such that
E θ ( m ) ≥ c θ ∥ m ∥ H 1 ( D ) 2 . (133) \mathcal{E}_{\theta}(m)\ge c_{\theta}\lVert m\rVert_{H^1(\mathcal{D})}^2. \tag*{(133)} E θ ( m ) ≥ c θ ∥ m ∥ H 1 ( D ) 2 . (133)
It remains to combine these two estimates with the dissipation law. Using the energy identity, estimate (123) , and (131) , we infer that, for a.e. t ≥ t 0 t\ge t_0 t ≥ t 0 ,
d d t E θ ( m ( t ) ) = − ⟨ M ( m ( t ) ) h ( m ( t ) ) , h ( m ( t ) ) ⟩ ≤ − 2 q ( ρ ) ∥ h ( m ( t ) ) ∥ L 2 ( D ) 2 ≤ − 4 μ q ( ρ ) E θ ( m ( t ) ) . \begin{aligned}
\frac{\mathrm{d}}{\mathrm{d}t}\mathcal{E}_{\theta}(m(t))=-\langle M(m(t))h(m(t)),h(m(t))\rangle\\
&\le-2q(\rho)\lVert h(m(t))\rVert_{L^2(\mathcal{D})}^2 \\
&\le-4\mu q(\rho)\mathcal{E}_{\theta}(m(t)).
\end{aligned} d t d E θ ( m ( t )) = − ⟨ M ( m ( t )) h ( m ( t )) , h ( m ( t ))⟩ ≤ − 2 q ( ρ ) ∥ h ( m ( t )) ∥ L 2 ( D ) 2 ≤ − 4 μ q ( ρ ) E θ ( m ( t )) .
Gronwall’s lemma then yields
E θ ( m ( t ) ) ≤ E θ ( m ( t 0 ) ) exp ( − 4 μ q ( ρ ) ( t − t 0 ) ) , t ≥ t 0 . \mathcal{E}_{\theta}(m(t))\le\mathcal{E}_{\theta}(m(t_0))\exp(-4\mu q(\rho)(t-t_0)),\qquad t\ge t_0. E θ ( m ( t )) ≤ E θ ( m ( t 0 )) exp ( − 4 μ q ( ρ ) ( t − t 0 )) , t ≥ t 0 .
Finally, (133) gives
∥ m ( t ) ∥ H 1 ( D ) 2 ≤ C exp ( − 4 μ q ( ρ ) t ) , \|m(t)\|_{H^1(\mathcal{D})}^2 \le C\exp(-4\mu q(\rho)t), ∥ m ( t ) ∥ H 1 ( D ) 2 ≤ C exp ( − 4 μ q ( ρ ) t ) ,
after absorbing the fixed time t 0 t_0 t 0 and E θ ( m ( t 0 ) ) E_\theta(m(t_0)) E θ ( m ( t 0 )) into the constant. This proves assertion (ii).
Finally, suppose 3 θ < Λ ∗ 3\theta< \Lambda_* 3 θ < Λ ∗ , i.e. θ < θ c \theta< \theta_c θ < θ c . By definition of Λ ∗ \Lambda_* Λ ∗ and density, there exists v ∈ C ∞ ( D ) v \in C^\infty(\mathcal{D}) v ∈ C ∞ ( D ) with ∂ ν v = 0 \partial_\nu v = 0 ∂ ν v = 0 on ∂ D \partial\mathcal{D} ∂ D such that
⟨ B v , v ⟩ − α ∥ ∇ v ∥ L 2 ( D ) 2 − ∥ H d [ v ] ∥ L 2 ( R 3 ) 2 > 3 θ ∥ v ∥ L 2 ( D ) 2 . \langle Bv,v\rangle-\alpha\|\nabla v\|_{L^2(\mathcal{D})}^2-\|H_d[v]\|_{L^2(\mathbb{R}^3)}^2>3\theta\|v\|_{L^2(\mathcal{D})}^2. ⟨ B v , v ⟩ − α ∥∇ v ∥ L 2 ( D ) 2 − ∥ H d [ v ] ∥ L 2 ( R 3 ) 2 > 3 θ ∥ v ∥ L 2 ( D ) 2 .
Using (128) , we obtain
E θ ( δ v ) = δ 2 2 [ α ∥ ∇ v ∥ L 2 ( D ) 2 − ⟨ B v , v ⟩ + ∥ H d [ v ] ∥ L 2 ( R 3 ) 2 + 3 θ ∥ v ∥ L 2 ( D ) 2 ] + O ( δ 4 ) . E_\theta(\delta v)=\frac{\delta^2}{2}\left[\alpha\|\nabla v\|_{L^2(\mathcal{D})}^2-\langle Bv,v\rangle+\|H_d[v]\|_{L^2(\mathbb{R}^3)}^2+3\theta\|v\|_{L^2(\mathcal{D})}^2\right]+O(\delta^4). E θ ( δ v ) = 2 δ 2 [ α ∥∇ v ∥ L 2 ( D ) 2 − ⟨ B v , v ⟩ + ∥ H d [ v ] ∥ L 2 ( R 3 ) 2 + 3 θ ∥ v ∥ L 2 ( D ) 2 ] + O ( δ 4 ) .
The coefficient of δ 2 \delta^2 δ 2 is negative. Hence E θ ( δ v ) < 0 E_\theta(\delta v)<0 E θ ( δ v ) < 0 for all sufficiently small δ > 0 \delta>0 δ > 0 , and δ \delta δ may also be chosen so that ∥ δ v ∥ L ∞ < 1 \|\delta v\|_{L^\infty}<1 ∥ δ v ∥ L ∞ < 1 .
If a trajectory satisfies E θ ( m ( t 0 ) ) < 0 E_\theta(m(t_0))<0 E θ ( m ( t 0 )) < 0 , then its limiting equilibrium from Theorem 4.4 satisfies
E θ ( m ∞ ) ≤ E θ ( m ( t 0 ) ) < 0. E_\theta(m_\infty)\le E_\theta(m(t_0))<0. E θ ( m ∞ ) ≤ E θ ( m ( t 0 )) < 0.
Since E θ ( 0 ) = 0 E_\theta(0)=0 E θ ( 0 ) = 0 , one has m ∞ ≠ 0 m_\infty\ne0 m ∞ = 0 . This shows assertion (iii), thus completing the proof of the theorem. □
Non-saturated case: strong solutions and uniqueness
Although the preceding weak-solution theory was formulated for the physical three-dimensional problem, the arguments extend, with only minor dimensional modifications, to d ∈ { 1 , 2 } d\in\{1,2\} d ∈ { 1 , 2 } . In this section, we explicitly allow D ⊂ R d \mathcal{D}\subset\mathbb{R}^d D ⊂ R d , d ∈ { 1 , 2 , 3 } d\in\{1,2,3\} d ∈ { 1 , 2 , 3 } , while the magnetisation remains R 3 \mathbb{R}^3 R 3 -valued. For d = 3 d=3 d = 3 , H d H_d H d denotes the physical whole-space stray-field operator introduced before. For d ∈ { 1 , 2 } d\in\{1,2\} d ∈ { 1 , 2 } , we consider a corresponding lower-dimensional operator, still denoted by H d H_d H d , which is linear, self-adjoint and non-positive on L 2 ( D ; R 3 ) L^2(\mathcal{D};\mathbb{R}^3) L 2 ( D ; R 3 ) , and satisfies (A.7)–(A.8). Under these assumptions, the global weak existence, separation results, and long-time behaviour established in § 3 remain valid for d ∈ { 1 , 2 } d\in\{1,2\} d ∈ { 1 , 2 } ; in fact, the relevant estimates are more favourable in lower dimensions.
Here, we consider initial data uniformly separated from saturation. In this case, the equation is initially uniformly parabolic, and the separation estimates established in § 3 can be combined with quasilinear parabolic theory to obtain stronger well-posedness results.
For ease of presentation, define
A ( v ) z : = M ( v ) z + γ v × z . \mathsf{A}(v)z:=\mathsf{M}(v)z+\gamma v\times z. A ( v ) z := M ( v ) z + γ v × z .
The closed-ball equation (14) in strong form can be written as
∂ t m − α A ( m ) Δ m = A ( m ) ( B m + H d [ m ] ) − 2 θ m . (134) \partial_t m-\alpha\mathsf{A}(m)\Delta m=\mathsf{A}(m)\bigl(Bm+H_d[m]\bigr)-2\theta m. \tag*{(134)} ∂ t m − α A ( m ) Δ m = A ( m ) ( B m + H d [ m ] ) − 2 θ m . (134)
For every ρ < 1 \rho<1 ρ < 1 , we have
Re ( z ‾ ⋅ A ( v ) z ) = z ‾ ⋅ M ( v ) z ≥ 2 q ( ρ ) ∣ z ∣ 2 , ∣ v ∣ ≤ ρ , z ∈ C 3 . (135) \operatorname{Re}(\overline{z}\cdot\mathsf{A}(v)z)=\overline{z}\cdot\mathsf{M}(v)z\ge2q(\rho)|z|^2,\qquad|v|\le\rho,\quad z\in\mathbb{C}^3. \tag*{(135)} Re ( z ⋅ A ( v ) z ) = z ⋅ M ( v ) z ≥ 2 q ( ρ ) ∣ z ∣ 2 , ∣ v ∣ ≤ ρ , z ∈ C 3 . (135)
Thus, the skew-symmetric precession term does not affect ellipticity.
Existence of a unique maximal strong solution
5.1. Existence of a unique maximal strong solution. The following theorem concerns the existence of a unique maximal strong solution, which is global when d ≤ 2 d\le2 d ≤ 2 . To state this theorem, we define the interpolation space
X s : = ( H s − 1 ( D ) , H n s + 1 ( D ) ) 1 2 , 2 . \mathcal{X}_s:=\left(H^{s-1}(\mathcal{D}),H_n^{s+1}(\mathcal{D})\right)_{\frac{1}{2},2}. X s := ( H s − 1 ( D ) , H n s + 1 ( D ) ) 2 1 , 2 .
Away from the borderline exponent s = 3 / 2 s=3/2 s = 3/2 , this space is concretely given by
X s = { H s ( D ) , 1 < s < 3 / 2 , H n s ( D ) , 3 / 2 < s ≤ 2. \mathcal{X}_s=
\begin{cases}
H^s(\mathcal{D}), & 1<s<3/2,\\
H_n^s(\mathcal{D}), & 3/2<s\le2.
\end{cases} X s = { H s ( D ) , H n s ( D ) , 1 < s < 3/2 , 3/2 < s ≤ 2.
while for s = 3 / 2 s = 3/2 s = 3/2 we retain the interpolation-space notation.
Theorem 5.1 (Strong solutions). Let s ∈ ( 1 , 2 ] s \in(1,2] s ∈ ( 1 , 2 ] satisfy s > d / 2 s > d/2 s > d /2 . Suppose m 0 ∈ X s m_0 \in\mathcal{X}_s m 0 ∈ X s , and there exists a constant ρ 0 > 0 \rho_0 > 0 ρ 0 > 0 such that ∥ m 0 ∥ L ∞ ( D ) ≤ ρ 0 < 1 \lVert m_0\rVert_{L^\infty(\mathcal{D})} \le\rho_0 < 1 ∥ m 0 ∥ L ∞ ( D ) ≤ ρ 0 < 1 . Then, there exists a unique maximal strong solution of (134) , with maximal existence time T max ∈ ( 0 , ∞ ] T_{\max} \in(0,\infty] T m a x ∈ ( 0 , ∞ ] , such that, for every T < T max T < T_{\max} T < T m a x ,
m ∈ C ( [ 0 , T ] ; X s ) ∩ H 1 ( 0 , T ; H s − 1 ( D ) ) ∩ L 2 ( 0 , T ; H n s + 1 ( D ) ) . (136) m \in C([0,T];\mathcal{X}_s) \cap H^1(0,T;H^{s-1}(\mathcal{D})) \cap L^2(0,T;H_n^{s+1}(\mathcal{D})). \tag*{(136)} m ∈ C ([ 0 , T ] ; X s ) ∩ H 1 ( 0 , T ; H s − 1 ( D )) ∩ L 2 ( 0 , T ; H n s + 1 ( D )) . (136) For each such T T T , there exists ρ T < 1 \rho_T < 1 ρ T < 1 satisfying
∥ m ( t ) ∥ L ∞ ( D ) ≤ ρ T for every t ∈ [ 0 , T ] . \lVert m(t)\rVert_{L^\infty(\mathcal{D})} \le\rho_T \qquad\text{for every } t \in[0,T]. ∥ m ( t ) ∥ L ∞ ( D ) ≤ ρ T for every t ∈ [ 0 , T ] . The solution depends locally Lipschitz continuously on its initial datum in the function spaces appearing in (136) .
If d ≤ 2 d \le2 d ≤ 2 , then T max = ∞ T_{\max} = \infty T m a x = ∞ . In particular, for every T > 0 T > 0 T > 0 ,
sup t ∈ [ 0 , T ] ∥ m ( t ) ∥ H s 2 + ∫ 0 T ( ∥ ∂ t m ( t ) ∥ H s − 1 2 + ∥ m ( t ) ∥ H s + 1 2 ) d t ≤ C T . (137) \sup_{t\in[0,T]} \lVert m(t)\rVert_{H^s}^2 + \int_0^T \left(\lVert\partial_t m(t)\rVert_{H^{s-1}}^2 + \lVert m(t)\rVert_{H^{s+1}}^2\right)\,\mathrm{d}t \le C_T. \tag*{(137)} t ∈ [ 0 , T ] sup ∥ m ( t ) ∥ H s 2 + ∫ 0 T ( ∥ ∂ t m ( t ) ∥ H s − 1 2 + ∥ m ( t ) ∥ H s + 1 2 ) d t ≤ C T . (137) Moreover, for every 0 < τ < T 0 < \tau< T 0 < τ < T ,
m ∈ C ( [ τ , T ] ; H n 2 ( D ) ) ∩ H 1 ( τ , T ; H 1 ( D ) ) ∩ L 2 ( τ , T ; H n 3 ( D ) ) , (138) m \in C([\tau,T];H_n^2(\mathcal{D})) \cap H^1(\tau,T;H^1(\mathcal{D})) \cap L^2(\tau,T;H_n^3(\mathcal{D})), \tag*{(138)} m ∈ C ([ τ , T ] ; H n 2 ( D )) ∩ H 1 ( τ , T ; H 1 ( D )) ∩ L 2 ( τ , T ; H n 3 ( D )) , (138) and
ess sup t ∈ [ τ , T ] ∥ ∂ t m ( t ) ∥ L 2 2 + ∫ τ T ∥ ∇ ∂ t m ( t ) ∥ L 2 2 d t ≤ C τ , T . (139) \operatorname*{ess\,sup}_{t\in[\tau,T]} \lVert\partial_t m(t)\rVert_{L^2}^2 + \int_\tau^T \lVert\nabla\partial_t m(t)\rVert_{L^2}^2\,\mathrm{d}t \le C_{\tau,T}. \tag*{(139)} t ∈ [ τ , T ] ess sup ∥ ∂ t m ( t ) ∥ L 2 2 + ∫ τ T ∥ ∇ ∂ t m ( t ) ∥ L 2 2 d t ≤ C τ , T . (139) The constants may depend on the initial datum, the fixed coefficients, the operator bounds, and the indicated time intervals.
Proof. We first apply local maximal regularity and then prove that, in dimensions one and two, a finite maximal time is impossible. The latter step uses an estimate for ∂ t m \partial_t m ∂ t m to recover a continuous H 2 H^2 H 2 -trace at the maximal time.
For local existence proof, set X 0 = H s − 1 ( D ) X_0 = H^{s-1}(\mathcal{D}) X 0 = H s − 1 ( D ) and X 1 = H n s + 1 ( D ) X_1 = H_n^{s+1}(\mathcal{D}) X 1 = H n s + 1 ( D ) . The Neumann interpolation identity gives ( X 0 , X 1 ) 1 2 , 2 = X s (X_0,X_1)_{\frac{1}{2},2} = \mathcal{X}_s ( X 0 , X 1 ) 2 1 , 2 = X s . Since X s ↪ H s ( D ) \mathcal{X}_s \hookrightarrow H^s(\mathcal{D}) X s ↪ H s ( D ) and s > d / 2 s > d/2 s > d /2 , the Sobolev multiplication theorem yields
∥ f g ∥ H s − 1 ≤ C ∥ f ∥ H s ∥ g ∥ H s − 1 . \lVert fg\rVert_{H^{s-1}} \le C\lVert f\rVert_{H^s}\lVert g\rVert_{H^{s-1}}. ∥ f g ∥ H s − 1 ≤ C ∥ f ∥ H s ∥ g ∥ H s − 1 .
Therefore, the maps
v ⟼ − α A ( v ) Δ ∈ L ( X 1 , X 0 ) and v ⟼ A ( v ) ( B v + H d [ v ] ) − 2 θ v ∈ X 0 v \longmapsto-\alpha A(v)\Delta\in\mathcal{L}(X_1,X_0) \qquad\text{and} \qquad v \longmapsto A(v)(Bv+H_d[v])-2\theta v \in X_0 v ⟼ − α A ( v ) Δ ∈ L ( X 1 , X 0 ) and v ⟼ A ( v ) ( B v + H d [ v ]) − 2 θ v ∈ X 0
are locally smooth on the open set U : = { v ∈ X s : ∥ v ∥ L ∞ < 1 } \mathcal{U}:=\{v\in\mathcal{X}_s:\lVert v\rVert_{L^\infty}<1\} U := { v ∈ X s : ∥ v ∥ L ∞ < 1 } .
We also verify maximal regularity for the frozen principal operator. For each fixed v ∈ U v\in\mathcal{U} v ∈ U , the principal symbol of − α A ( v ) Δ -\alpha A(v)\Delta − α A ( v ) Δ is α ∣ ξ ∣ 2 A ( v ( x ) ) \alpha|\xi|^2A(v(x)) α ∣ ξ ∣ 2 A ( v ( x )) . By (135) , this symbol is uniformly normally elliptic on every subset of U \mathcal{U} U that is uniformly separated from saturation. Together with the homogeneous Neumann boundary condition, standard maximal-regularity theory for normally elliptic parabolic systems thus shows that
− α A ( v ) Δ : H n s + 1 ( D ) → H s − 1 ( D ) -\alpha A(v)\Delta:H_n^{s+1}(\mathcal{D})\to H^{s-1}(\mathcal{D}) − α A ( v ) Δ : H n s + 1 ( D ) → H s − 1 ( D )
has maximal L 2 L^2 L 2 -regularity; see, e.g., [3 , 4 ] . The maximal regularity theorem for quasilinear parabolic problems then gives the unique maximal solution and continuous dependence stated above; see [4 Theorems 2.1 and 3.1] . Since H s ( D ) ↪ C ( D ‾ ) H^s(\mathcal{D})\hookrightarrow C(\overline{\mathcal{D}}) H s ( D ) ↪ C ( D ) and m ∈ C ( [ 0 , T ] ; X s ) ↪ C ( [ 0 , T ] ; H s ( D ) ) m\in C([0,T];\mathcal{X}_s)\hookrightarrow C([0,T];H^s(\mathcal{D})) m ∈ C ([ 0 , T ] ; X s ) ↪ C ([ 0 , T ] ; H s ( D )) , we have
∥ m ( t ) − m 0 ∥ L ∞ ( D ) → 0 as t ↓ 0. \lVert m(t)-m_0\rVert_{L^\infty(\mathcal{D})}\to0 \qquad\text{as }t\downarrow0. ∥ m ( t ) − m 0 ∥ L ∞ ( D ) → 0 as t ↓ 0.
Hence, the strict initial separation ∥ m 0 ∥ L ∞ ≤ ρ 0 < 1 \lVert m_0\rVert_{L^\infty}\le\rho_0<1 ∥ m 0 ∥ L ∞ ≤ ρ 0 < 1 persists on a sufficiently short time interval. Moreover, if T max < ∞ T_{\max}<\infty T m a x < ∞ and m ( t ) → m ∗ m(t)\to m_* m ( t ) → m ∗ in X s \mathcal{X}_s X s as t ↑ T max t\uparrow T_{\max} t ↑ T m a x with ∥ m ∗ ∥ L ∞ < 1 \lVert m_*\rVert_{L^\infty}<1 ∥ m ∗ ∥ L ∞ < 1 , then m ∗ m_* m ∗ belongs to the open phase space U \mathcal{U} U . Restarting the local problem from m ∗ m_* m ∗ at t = T max t=T_{\max} t = T m a x therefore extends the solution beyond T max T_{\max} T m a x , contradicting maximality.
We now assume d ≤ 2 d \le2 d ≤ 2 and suppose, for contradiction, that T max < ∞ T_{\max} < \infty T m a x < ∞ . Since s > 1 s > 1 s > 1 , we have m 0 ∈ X s ↪ H 1 ( D ) m_0 \in\mathcal{X}_s \hookrightarrow H^1(\mathcal{D}) m 0 ∈ X s ↪ H 1 ( D ) . The finite-time H 1 H^1 H 1 and time-derivative a priori estimates, as well as the positive-time separation established in Theorem 3.2 , apply on every interval [ 0 , S ] [0,S] [ 0 , S ] with S < T max S < T_{\max} S < T m a x , with constants that may be chosen uniformly for S ≤ T max S \le T_{\max} S ≤ T m a x . Consequently, for every fixed a ∈ ( 0 , T max ) a \in(0,T_{\max}) a ∈ ( 0 , T m a x ) , there exist ρ a ∈ ( 0 , 1 ) \rho_a \in(0,1) ρ a ∈ ( 0 , 1 ) and C a > 0 C_a > 0 C a > 0 such that
∣ m ∣ ≤ ρ a < 1 on [ a , T max ) × D , (140) |m| \le\rho_a < 1 \quad\text{on } [a,T_{\max}) \times\mathcal{D}, \tag*{(140)} ∣ m ∣ ≤ ρ a < 1 on [ a , T m a x ) × D , (140)
and
sup t < T max ∥ m ( t ) ∥ H 1 2 + ∫ a T max ( ∥ m ( t ) ∥ H 2 2 + ∥ ∂ t m ( t ) ∥ L 2 2 ) d t ≤ C a . (141) \sup_{t<T_{\max}} \|m(t)\|_{H^1}^2 + \int_a^{T_{\max}} \left(\|m(t)\|_{H^2}^2 + \|\partial_t m(t)\|_{L^2}^2\right)\,\mathrm{d}t \le C_a. \tag*{(141)} t < T m a x sup ∥ m ( t ) ∥ H 1 2 + ∫ a T m a x ( ∥ m ( t ) ∥ H 2 2 + ∥ ∂ t m ( t ) ∥ L 2 2 ) d t ≤ C a . (141)
Here, the constants are independent of S < T max S < T_{\max} S < T m a x . The H 2 H^2 H 2 -bound follows from the weighted dissipation and q ( ∣ m ∣ ) ≥ q ( ρ a ) > 0 q(|m|) \ge q(\rho_a) > 0 q ( ∣ m ∣ ) ≥ q ( ρ a ) > 0 as in (91) .
Now, set w : = ∂ t m w := \partial_t m w := ∂ t m . Time differentiation of (134) gives
∂ t w = α A ( m ) Δ w + D A ( m ) [ w ] H M F A ( m ) + A ( m ) ( B w + H d [ w ] ) − 2 θ w . \partial_t w = \alpha A(m)\Delta w + DA(m)[w]H_{\mathrm{MFA}}(m) + A(m)(Bw + H_{\mathrm{d}}[w]) - 2\theta w. ∂ t w = α A ( m ) Δ w + D A ( m ) [ w ] H MFA ( m ) + A ( m ) ( B w + H d [ w ]) − 2 θ w .
Testing by w w w , integrating the principal term by parts, and using the skew-symmetry of the precessional coefficient, we obtain
1 2 d d t ∥ w ∥ L 2 2 + α ∑ i = 1 d ⟨ M ( m ) ∂ i w , ∂ i w ⟩ + 2 θ ∥ w ∥ L 2 2 = − α ∑ i = 1 d ⟨ D A ( m ) [ ∂ i m ] ∂ i w , w ⟩ + ⟨ D A ( m ) [ w ] H M F A ( m ) , w ⟩ + ⟨ A ( m ) ( B w + H d [ w ] ) , w ⟩ ≤ C a ( ∥ ∇ m ∥ L 4 ∥ w ∥ L 4 ∥ ∇ w ∥ L 2 + ∥ H M F A ( m ) ∥ L 2 ∥ w ∥ L 4 2 + ∥ w ∥ L 2 2 ) . (142) \begin{aligned}
\frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\|w\|_{L^2}^2 + \alpha\sum_{i=1}^{d} \langle M(m)\partial_i w,\partial_i w\rangle+ 2\theta\|w\|_{L^2}^2 \\
&= -\alpha\sum_{i=1}^{d} \langle DA(m)[\partial_i m]\partial_i w,w\rangle+ \langle DA(m)[w]H_{\mathrm{MFA}}(m),w\rangle+ \langle A(m)(Bw + H_{\mathrm{d}}[w]),w\rangle\\
&\le C_a\left(\|\nabla m\|_{L^4}\|w\|_{L^4}\|\nabla w\|_{L^2} + \|H_{\mathrm{MFA}}(m)\|_{L^2}\|w\|_{L^4}^2 + \|w\|_{L^2}^2\right). \tag*{(142)}
\end{aligned} 2 1 d t d ∥ w ∥ L 2 2 + α i = 1 ∑ d ⟨ M ( m ) ∂ i w , ∂ i w ⟩ + 2 θ ∥ w ∥ L 2 2 = − α i = 1 ∑ d ⟨ D A ( m ) [ ∂ i m ] ∂ i w , w ⟩ + ⟨ D A ( m ) [ w ] H MFA ( m ) , w ⟩ + ⟨ A ( m ) ( B w + H d [ w ]) , w ⟩ ≤ C a ( ∥∇ m ∥ L 4 ∥ w ∥ L 4 ∥∇ w ∥ L 2 + ∥ H MFA ( m ) ∥ L 2 ∥ w ∥ L 4 2 + ∥ w ∥ L 2 2 ) . (142)
For d ≤ 2 d \le2 d ≤ 2 , the Gagliardo–Nirenberg inequality gives
∥ w ∥ L 4 2 ≤ C ∥ w ∥ L 2 ( ∥ ∇ w ∥ L 2 + ∥ w ∥ L 2 ) . \|w\|_{L^4}^2 \le C\|w\|_{L^2}\left(\|\nabla w\|_{L^2} + \|w\|_{L^2}\right). ∥ w ∥ L 4 2 ≤ C ∥ w ∥ L 2 ( ∥∇ w ∥ L 2 + ∥ w ∥ L 2 ) .
Therefore, applying Young’s inequality in (142) , and using the ellipticity (135) , we obtain
d d t ∥ w ∥ L 2 2 + c a ∥ ∇ w ∥ L 2 2 ≤ C a ( 1 + ∥ ∇ m ∥ L 4 4 + ∥ H M F A ( m ) ∥ L 2 2 ) ∥ w ∥ L 2 2 . (143) \frac{\mathrm{d}}{\mathrm{d}t}\|w\|_{L^2}^2 + c_a\|\nabla w\|_{L^2}^2 \le C_a\left(1+\|\nabla m\|_{L^4}^4+\|H_{\mathrm{MFA}}(m)\|_{L^2}^2\right)\|w\|_{L^2}^2. \tag*{(143)} d t d ∥ w ∥ L 2 2 + c a ∥∇ w ∥ L 2 2 ≤ C a ( 1 + ∥∇ m ∥ L 4 4 + ∥ H MFA ( m ) ∥ L 2 2 ) ∥ w ∥ L 2 2 . (143)
where the coefficient on the right-hand side is integrable up to T max T_{\max} T m a x . Therefore, Gronwall’s inequality yields, for every b ∈ ( a , T max ) b \in(a,T_{\max}) b ∈ ( a , T m a x ) ,
ess sup b < t < T max ∥ w ( t ) ∥ L 2 2 + ∫ b T max ∥ ∇ w ( t ) ∥ L 2 2 d t < ∞ . (144) \operatorname*{ess\,sup}_{b<t<T_{\max}}\|w(t)\|_{L^2}^2 + \int_b^{T_{\max}}\|\nabla w(t)\|_{L^2}^2\,\mathrm{d}t < \infty. \tag*{(144)} b < t < T m a x ess sup ∥ w ( t ) ∥ L 2 2 + ∫ b T m a x ∥∇ w ( t ) ∥ L 2 2 d t < ∞. (144)
We now recover two spatial bounds from this estimate. Since A ( m ) A(m) A ( m ) is uniformly invertible on [ b , T max ) [b,T_{\max}) [ b , T m a x ) , equation (134) gives
α Δ m = A ( m ) − 1 ( w + 2 θ m ) − B m − H d [ m ] . (145) \alpha\Delta m = A(m)^{-1}(w + 2\theta m) - Bm - H_{\mathrm{d}}[m]. \tag*{(145)} α Δ m = A ( m ) − 1 ( w + 2 θ m ) − B m − H d [ m ] . (145)
Hence, we have
∥ m ( t ) ∥ H 2 ≤ C a ( 1 + ∥ w ( t ) ∥ L 2 ) , \|m(t)\|_{H^2} \le C_a(1+\|w(t)\|_{L^2}), ∥ m ( t ) ∥ H 2 ≤ C a ( 1 + ∥ w ( t ) ∥ L 2 ) ,
so m m m is bounded in L ∞ ( b , T max ; H 2 ( D ) ) L^\infty(b,T_{\max};H^2(\mathcal{D})) L ∞ ( b , T m a x ; H 2 ( D )) .
We next recover one additional spatial derivative. By the uniform H 2 H^2 H 2 -bound for m m m and the separation (140) , the standard Sobolev composition estimate gives
sup b < t < T max ∥ A ( m ( t ) ) − 1 ∥ H 2 ≤ C a . \sup_{b<t<T_{\max}}\|A(m(t))^{-1}\|_{H^2} \le C_a. b < t < T m a x sup ∥ A ( m ( t ) ) − 1 ∥ H 2 ≤ C a .
Noting that the product of H 2 H^2 H 2 and H 1 H^1 H 1 functions lies in H 1 H^1 H 1 , by the elliptic recovery identity (145) and the H 1 H^1 H 1 -boundedness of H d H_d H d , we obtain
∥ Δ m ∥ H 1 ≤ C a ( 1 + ∥ w ∥ H 1 ) . \|\Delta m\|_{H^1} \le C_a(1+\|w\|_{H^1}). ∥Δ m ∥ H 1 ≤ C a ( 1 + ∥ w ∥ H 1 ) .
Therefore, by (144) and Neumann elliptic regularity,
m ∈ L 2 ( b , T max ; H n 3 ( D ) ) , ∂ t m = w ∈ L 2 ( b , T max ; H 1 ( D ) ) . m \in L^2(b,T_{\max};H_n^3(\mathcal{D})), \qquad\partial_t m = w \in L^2(b,T_{\max};H^1(\mathcal{D})). m ∈ L 2 ( b , T m a x ; H n 3 ( D )) , ∂ t m = w ∈ L 2 ( b , T m a x ; H 1 ( D )) .
The standard interpolation result then shows that m ∈ C ( [ b , T max ] ; H n 2 ( D ) ) m \in C([b,T_{\max}];H_n^2(\mathcal{D})) m ∈ C ([ b , T m a x ] ; H n 2 ( D )) . Since H n 2 ( D ) ↪ X s H_n^2(\mathcal{D}) \hookrightarrow X_s H n 2 ( D ) ↪ X s for s ∈ ( 1 , 2 ] s \in(1,2] s ∈ ( 1 , 2 ] , in particular, m ( t ) m(t) m ( t ) has a limit in X s X_s X s as t ↑ T max t \uparrow T_{\max} t ↑ T m a x . By (140) , this limit remains strictly separated from saturation. Thus, we deduce T max = ∞ T_{\max}=\infty T m a x = ∞ . The maximal-regularity class gives (137) on every finite interval. Repeating the preceding estimates on [ τ / 2 , T ] [\tau/2,T] [ τ /2 , T ] proves (138) and (139) . □
Exponential stability
We next obtain an exponential stability result. This implies the existence of a unique global strong solution for small initial data when d = 3 d=3 d = 3 .
Theorem 5.2 (Exponential stability of non-degenerate equilibria). Let d ≤ 3 d \le3 d ≤ 3 and let s ∈ ( 1 , 2 ] s \in(1,2] s ∈ ( 1 , 2 ] satisfy s > d / 2 s>d/2 s > d /2 . Let m ˉ ∈ S θ \bar{m} \in S_\theta m ˉ ∈ S θ , and define the operator
L m ˉ : = D 2 E θ ( m ˉ ) = − α Δ − B − H d + θ D G ( m ˉ ) , \mathcal{L}_{\bar{m}} := D^2\mathcal{E}_\theta(\bar{m})=-\alpha\Delta-B-H_d+\theta DG(\bar{m}), L m ˉ := D 2 E θ ( m ˉ ) = − α Δ − B − H d + θ D G ( m ˉ ) , which is the Hessian of the free energy at the equilibrium m ˉ \bar{m} m ˉ . Assume that there exists μ > 0 \mu>0 μ > 0 such that
⟨ L m ˉ z , z ⟩ ≥ μ ∥ z ∥ L 2 2 for every z ∈ H n 2 ( D ) , (146) \langle\mathcal{L}_{\bar{m}}z,z\rangle\ge\mu\|z\|_{L^2}^2 \qquad\text{for every } z\in H_n^2(\mathcal{D}), \tag*{(146)} ⟨ L m ˉ z , z ⟩ ≥ μ ∥ z ∥ L 2 2 for every z ∈ H n 2 ( D ) , (146) Then, there exist η , ω , C > 0 \eta,\omega,C>0 η , ω , C > 0 such that, whenever m 0 ∈ X s m_0\in X_s m 0 ∈ X s and ∥ m 0 − m ˉ ∥ X s ≤ η \|m_0-\bar{m}\|_{X_s}\le\eta ∥ m 0 − m ˉ ∥ X s ≤ η , the corresponding solution is global, remains uniformly separated from saturation, and satisfies
sup t ≥ 0 e 2 ω t ∥ m ( t ) − m ˉ ∥ H s 2 + ∫ 0 ∞ e 2 ω t ( ∥ ∂ t m ( t ) ∥ H s − 1 2 + ∥ m ( t ) − m ˉ ∥ H s + 1 2 ) d t ≤ C ∥ m 0 − m ˉ ∥ X s 2 . (147) \sup_{t\ge0}e^{2\omega t}\|m(t)-\bar{m}\|_{H^s}^2
+\int_0^\infty e^{2\omega t}\left(\|\partial_t m(t)\|_{H^{s-1}}^2+\|m(t)-\bar{m}\|_{H^{s+1}}^2\right)\,\mathrm{d}t
\le C\|m_0-\bar{m}\|_{X_s}^2. \tag*{(147)} t ≥ 0 sup e 2 ω t ∥ m ( t ) − m ˉ ∥ H s 2 + ∫ 0 ∞ e 2 ω t ( ∥ ∂ t m ( t ) ∥ H s − 1 2 + ∥ m ( t ) − m ˉ ∥ H s + 1 2 ) d t ≤ C ∥ m 0 − m ˉ ∥ X s 2 . (147) In particular,
∥ m ( t ) − m ˉ ∥ H s ≤ C e − ω t ∥ m 0 − m ˉ ∥ X s . \|m(t)-\bar{m}\|_{H^s}\le Ce^{-\omega t}\|m_0-\bar{m}\|_{X_s}. ∥ m ( t ) − m ˉ ∥ H s ≤ C e − ω t ∥ m 0 − m ˉ ∥ X s . The constants depend on the fixed coefficients, the domain, m ˉ \bar{m} m ˉ , and the stability bound, but not on m 0 m_0 m 0 within the stated neighbourhood.
Proof. We first verify stability of the linearised operator, then estimate the nonlinear remainder in the same maximal-regularity spaces used in Theorem 5.1 .
Since h ( m ˉ ) = 0 h(\bar{m})=0 h ( m ˉ ) = 0 and m ˉ \bar{m} m ˉ is strictly separated, the H 2 H^2 H 2 -mapping property of H d H_d H d and elliptic regularity give m ˉ ∈ H 4 ( D ) \bar{m}\in H^4(\mathcal{D}) m ˉ ∈ H 4 ( D ) . In particular, all coefficients below have the regularity required for their realizations on X 0 = H n s − 1 ( D ) X_0=H_n^{s-1}(\mathcal{D}) X 0 = H n s − 1 ( D ) and X 1 = H n s + 1 ( D ) X_1=H_n^{s+1}(\mathcal{D}) X 1 = H n s + 1 ( D ) . Set A m ˉ : = A ( m ˉ ) L m ˉ \mathcal{A}_{\bar{m}}:=A(\bar{m})\mathcal{L}_{\bar{m}} A m ˉ := A ( m ˉ ) L m ˉ . Since h ( m ˉ ) = 0 h(\bar{m})=0 h ( m ˉ ) = 0 and D h ( m ˉ ) = − L m ˉ Dh(\bar{m})=-\mathcal{L}_{\bar{m}} D h ( m ˉ ) = − L m ˉ , the linearisation of the LLB equation about m ˉ \bar{m} m ˉ is ∂ t u + A m ˉ u = 0 \partial_tu+\mathcal{A}_{\bar{m}}u=0 ∂ t u + A m ˉ u = 0 .
Choose ρ ∈ ( ∥ m ˉ ∥ L ∞ , 1 ) \rho\in(\|\bar{m}\|_{L^\infty},1) ρ ∈ ( ∥ m ˉ ∥ L ∞ , 1 ) . The operator L m ˉ \mathcal{L}_{\bar{m}} L m ˉ is self-adjoint and satisfies L m ˉ ≥ μ I \mathcal{L}_{\bar{m}}\ge\mu I L m ˉ ≥ μ I on L 2 L^2 L 2 . If
A m ˉ z = λ z , z ≠ 0 , \mathcal{A}_{\bar{m}}z=\lambda z,\qquad z\ne0, A m ˉ z = λ z , z = 0 ,
then, using the complexified L 2 L^2 L 2 inner product,
Re λ ⟨ L m ˉ z , z ⟩ = Re ⟨ A ( m ˉ ) L m ˉ z , L m ˉ z ⟩ = ⟨ M ( m ˉ ) L m ˉ z , L m ˉ z ⟩ ≥ 2 q ( ρ ) ∥ L m ˉ z ∥ L 2 2 ≥ 2 μ q ( ρ ) ⟨ L m ˉ z , z ⟩ . \operatorname{Re}\lambda\langle\mathcal{L}_{\bar{m}}z,z\rangle
=\operatorname{Re}\langle A(\bar{m})\mathcal{L}_{\bar{m}}z,\mathcal{L}_{\bar{m}}z\rangle
=\langle M(\bar{m})\mathcal{L}_{\bar{m}}z,\mathcal{L}_{\bar{m}}z\rangle
\ge2q(\rho)\|\mathcal{L}_{\bar{m}}z\|_{L^2}^2
\ge2\mu q(\rho)\langle\mathcal{L}_{\bar{m}}z,z\rangle. Re λ ⟨ L m ˉ z , z ⟩ = Re ⟨ A ( m ˉ ) L m ˉ z , L m ˉ z ⟩ = ⟨ M ( m ˉ ) L m ˉ z , L m ˉ z ⟩ ≥ 2 q ( ρ ) ∥ L m ˉ z ∥ L 2 2 ≥ 2 μ q ( ρ ) ⟨ L m ˉ z , z ⟩ .
Therefore, Re λ ≥ 2 μ q ( ρ ) > 0 \operatorname{Re}\lambda\ge2\mu_q(\rho)>0 Re λ ≥ 2 μ q ( ρ ) > 0 . Moreover, A m ˉ \mathcal{A}_{\bar m} A m ˉ is normally elliptic with compact resolvent, so its spectrum consists only of isolated eigenvalues of finite multiplicity. Elliptic regularity identifies the spectra of its L 2 L^2 L 2 - and X 0 X_0 X 0 -realizations. The preceding estimate then gives a positive spectral gap,
inf { Re λ : λ ∈ σ ( A m ˉ ) } ≥ 2 μ q ( ρ ) > 0. \inf\{\operatorname{Re}\lambda:\lambda\in\sigma(\mathcal{A}_{\bar m})\}\ge2\mu_q(\rho)>0. inf { Re λ : λ ∈ σ ( A m ˉ )} ≥ 2 μ q ( ρ ) > 0.
Consequently, − A m ˉ -\mathcal{A}_{\bar m} − A m ˉ generates an exponentially stable analytic semigroup on X 0 X_0 X 0 . For some ω > 0 \omega>0 ω > 0 , exponentially weighted maximal regularity then yields
sup t ≥ 0 e ω t ∥ u ( t ) ∥ H s + ∥ e ω t ∂ t u ∥ L 2 ( 0 , ∞ ; H s − 1 ) + ∥ e ω t u ∥ L 2 ( 0 , ∞ ; H s + 1 ) ≤ C ( ∥ u ( 0 ) ∥ X s + ∥ e ω t ( ∂ t u + A m ˉ u ) ∥ L 2 ( 0 , ∞ ; H s − 1 ) ) ; (148) \sup_{t\ge0}e^{\omega t}\|u(t)\|_{H^s}+\|e^{\omega t}\partial_tu\|_{L^2(0,\infty;H^{s-1})}+\|e^{\omega t}u\|_{L^2(0,\infty;H^{s+1})}
\le C\left(\|u(0)\|_{X_s}+\|e^{\omega t}(\partial_tu+\mathcal{A}_{\bar m}u)\|_{L^2(0,\infty;H^{s-1})}\right); \tag*{(148)} t ≥ 0 sup e ω t ∥ u ( t ) ∥ H s + ∥ e ω t ∂ t u ∥ L 2 ( 0 , ∞ ; H s − 1 ) + ∥ e ω t u ∥ L 2 ( 0 , ∞ ; H s + 1 ) ≤ C ( ∥ u ( 0 ) ∥ X s + ∥ e ω t ( ∂ t u + A m ˉ u ) ∥ L 2 ( 0 , ∞ ; H s − 1 ) ) ; (148)
see, e.g. [23 Proposition 3.3] .
Writing u = m − m ˉ u=m-\bar m u = m − m ˉ and using h ( m ˉ ) = 0 h(\bar m)=0 h ( m ˉ ) = 0 , we rewrite the full nonlinear LLB equation (14) as a perturbation of its linearisation,
∂ t u + A m ˉ u = N m ˉ ( u ) , \partial_tu+\mathcal{A}_{\bar m}u=\mathcal{N}_{\bar m}(u), ∂ t u + A m ˉ u = N m ˉ ( u ) ,
where
N m ˉ ( u ) = − ( A ( m ˉ + u ) − A ( m ˉ ) ) L m ˉ u − θ A ( m ˉ + u ) ( G ( m ˉ + u ) − G ( m ˉ ) − D G ( m ˉ ) u ) . \mathcal{N}_{\bar m}(u)=-\left(\mathcal{A}(\bar m+u)-\mathcal{A}(\bar m)\right)\mathcal{L}_{\bar m}u-\theta\mathcal{A}(\bar m+u)\left(G(\bar m+u)-G(\bar m)-DG(\bar m)u\right). N m ˉ ( u ) = − ( A ( m ˉ + u ) − A ( m ˉ ) ) L m ˉ u − θ A ( m ˉ + u ) ( G ( m ˉ + u ) − G ( m ˉ ) − D G ( m ˉ ) u ) .
For u u u in a sufficiently small H s H^s H s -neighbourhood of zero, m ˉ + u \bar m+u m ˉ + u remains in a fixed compact subset of the open unit ball. The Sobolev multiplier estimate and Taylor’s theorem then give
∥ N m ˉ ( u ) ∥ H s − 1 ≤ C ∥ u ∥ H s ∥ u ∥ H s + 1 + C ∥ u ∥ H s 2 ≤ C ∥ u ∥ H s ∥ u ∥ H s + 1 . (149) \|\mathcal{N}_{\bar m}(u)\|_{H^{s-1}}\le C\|u\|_{H^s}\|u\|_{H^{s+1}}+C\|u\|_{H^s}^2\le C\|u\|_{H^s}\|u\|_{H^{s+1}}. \tag*{(149)} ∥ N m ˉ ( u ) ∥ H s − 1 ≤ C ∥ u ∥ H s ∥ u ∥ H s + 1 + C ∥ u ∥ H s 2 ≤ C ∥ u ∥ H s ∥ u ∥ H s + 1 . (149)
The same estimates for differences yield
∥ N m ˉ ( u ) − N m ˉ ( v ) ∥ H s − 1 ≤ C ( ∥ u ∥ H s + ∥ v ∥ H s ) ∥ u − v ∥ H s + 1 + C ( ∥ u ∥ H s + 1 + ∥ v ∥ H s + 1 ) ∥ u − v ∥ H s . (150) \|\mathcal{N}_{\bar m}(u)-\mathcal{N}_{\bar m}(v)\|_{H^{s-1}}\le C(\|u\|_{H^s}+\|v\|_{H^s})\|u-v\|_{H^{s+1}}+C(\|u\|_{H^{s+1}}+\|v\|_{H^{s+1}})\|u-v\|_{H^s}. \tag*{(150)} ∥ N m ˉ ( u ) − N m ˉ ( v ) ∥ H s − 1 ≤ C ( ∥ u ∥ H s + ∥ v ∥ H s ) ∥ u − v ∥ H s + 1 + C ( ∥ u ∥ H s + 1 + ∥ v ∥ H s + 1 ) ∥ u − v ∥ H s . (150)
Introduce the exponentially weighted maximal-regularity norm
∥ u ∥ E ω : = sup t ≥ 0 e ω t ∥ u ( t ) ∥ H s + ∥ e ω t ∂ t u ∥ L 2 ( 0 , ∞ ; H s − 1 ) + ∥ e ω t u ∥ L 2 ( 0 , ∞ ; H s + 1 ) . \|u\|_{E_\omega}:=\sup_{t\ge0}e^{\omega t}\|u(t)\|_{H^s}+\|e^{\omega t}\partial_tu\|_{L^2(0,\infty;H^{s-1})}+\|e^{\omega t}u\|_{L^2(0,\infty;H^{s+1})}. ∥ u ∥ E ω := t ≥ 0 sup e ω t ∥ u ( t ) ∥ H s + ∥ e ω t ∂ t u ∥ L 2 ( 0 , ∞ ; H s − 1 ) + ∥ e ω t u ∥ L 2 ( 0 , ∞ ; H s + 1 ) .
The preceding estimates (149) and (150) imply, for u , v u,v u , v in a sufficiently small E ω E_\omega E ω -ball,
∥ e ω t N m ˉ ( u ) ∥ L 2 ( 0 , ∞ ; H s − 1 ) ≤ C ∥ u ∥ E ω 2 , \|e^{\omega t}\mathcal{N}_{\bar m}(u)\|_{L^2(0,\infty;H^{s-1})}\le C\|u\|_{E_\omega}^2, ∥ e ω t N m ˉ ( u ) ∥ L 2 ( 0 , ∞ ; H s − 1 ) ≤ C ∥ u ∥ E ω 2 ,
∥ e ω t ( N m ˉ ( u ) − N m ˉ ( v ) ) ∥ L 2 ( 0 , ∞ ; H s − 1 ) ≤ C ( ∥ u ∥ E ω + ∥ v ∥ E ω ) ∥ u − v ∥ E ω . \|e^{\omega t}(\mathcal{N}_{\bar m}(u)-\mathcal{N}_{\bar m}(v))\|_{L^2(0,\infty;H^{s-1})}\le C(\|u\|_{E_\omega}+\|v\|_{E_\omega})\|u-v\|_{E_\omega}. ∥ e ω t ( N m ˉ ( u ) − N m ˉ ( v )) ∥ L 2 ( 0 , ∞ ; H s − 1 ) ≤ C ( ∥ u ∥ E ω + ∥ v ∥ E ω ) ∥ u − v ∥ E ω .
Now, let T ( u ) \mathcal{T}(u) T ( u ) denote the solution of
∂ t T ( u ) + A m ˉ T ( u ) = N m ˉ ( u ) , T ( u ) ( 0 ) = m 0 − m ˉ . \partial_t\mathcal{T}(u)+\mathcal{A}_{\bar m}\mathcal{T}(u)=\mathcal{N}_{\bar m}(u),\qquad\mathcal{T}(u)(0)=m_0-\bar m. ∂ t T ( u ) + A m ˉ T ( u ) = N m ˉ ( u ) , T ( u ) ( 0 ) = m 0 − m ˉ .
Applying (148) to T ( u ) \mathcal{T}(u) T ( u ) , with forcing N m ˉ ( u ) \mathcal{N}_{\bar m}(u) N m ˉ ( u ) and initial datum m 0 − m ˉ m_0-\bar m m 0 − m ˉ , then subsequently to T ( u ) − T ( v ) \mathcal{T}(u)-\mathcal{T}(v) T ( u ) − T ( v ) , whose initial datum vanishes, gives
∥ T ( u ) ∥ E ω ≤ C 0 ∥ m 0 − m ˉ ∥ X s + C 1 ∥ u ∥ E ω 2 , \|\mathcal{T}(u)\|_{E_\omega}\le C_0\|m_0-\bar m\|_{X_s}+C_1\|u\|_{E_\omega}^2, ∥ T ( u ) ∥ E ω ≤ C 0 ∥ m 0 − m ˉ ∥ X s + C 1 ∥ u ∥ E ω 2 ,
∥ T ( u ) − T ( v ) ∥ E ω ≤ C 1 ( ∥ u ∥ E ω + ∥ v ∥ E ω ) ∥ u − v ∥ E ω . \|\mathcal{T}(u)-\mathcal{T}(v)\|_{E_\omega}\le C_1(\|u\|_{E_\omega}+\|v\|_{E_\omega})\|u-v\|_{E_\omega}. ∥ T ( u ) − T ( v ) ∥ E ω ≤ C 1 ( ∥ u ∥ E ω + ∥ v ∥ E ω ) ∥ u − v ∥ E ω .
We now choose the fixed-point ball. First, select R > 0 R>0 R > 0 sufficiently small that
C 1 R ≤ 1 4 and C e m b R ≤ 1 2 ( 1 − ∥ m ˉ ∥ L ∞ ( D ) ) . C_1R\le\frac{1}{4}\qquad\text{and}\qquad C_{\mathrm{emb}}R\le\frac{1}{2}(1-\|\bar m\|_{L^\infty(\mathcal{D})}). C 1 R ≤ 4 1 and C emb R ≤ 2 1 ( 1 − ∥ m ˉ ∥ L ∞ ( D ) ) .
where C e m b C_{\mathrm{emb}} C emb is the norm of the embedding H s ( D ) ↪ L ∞ ( D ) H^s(\mathcal{D})\hookrightarrow L^\infty(\mathcal{D}) H s ( D ) ↪ L ∞ ( D ) . Then, for u , v u,v u , v in the closed ball
B R : = { u ∈ E ω : ∥ u ∥ E ω ≤ R } , B_R:=\{u\in E_\omega:\|u\|_{E_\omega}\le R\}, B R := { u ∈ E ω : ∥ u ∥ E ω ≤ R } ,
we have
∥ T ( u ) − T ( v ) ∥ E ω ≤ 2 C 1 R ∥ u − v ∥ E ω ≤ 1 2 ∥ u − v ∥ E ω . \|\mathcal{T}(u)-\mathcal{T}(v)\|_{E_\omega}\le2C_1R\|u-v\|_{E_\omega}\le\frac{1}{2}\|u-v\|_{E_\omega}. ∥ T ( u ) − T ( v ) ∥ E ω ≤ 2 C 1 R ∥ u − v ∥ E ω ≤ 2 1 ∥ u − v ∥ E ω .
Hence, T \mathcal{T} T is a contraction on B R B_R B R .
Having fixed R R R , choose the initial perturbation sufficiently small that ∥ m 0 − m ˉ ∥ H s ≤ R / ( 2 C 0 ) \lVert m_0-\bar{m}\rVert_{H^s}\le R/(2C_0) ∥ m 0 − m ˉ ∥ H s ≤ R / ( 2 C 0 ) . Then, for every u ∈ B R u\in B_R u ∈ B R ,
∥ T ( u ) ∥ E ω ≤ C 0 ∥ m 0 − m ˉ ∥ X s + C 1 R 2 ≤ R 2 + R 4 < R , \lVert\mathcal{T}(u)\rVert_{E_\omega}\le C_0\lVert m_0-\bar{m}\rVert_{X^s}+C_1R^2\le\frac{R}{2}+\frac{R}{4}<R, ∥ T ( u ) ∥ E ω ≤ C 0 ∥ m 0 − m ˉ ∥ X s + C 1 R 2 ≤ 2 R + 4 R < R ,
so T \mathcal{T} T maps B R B_R B R into itself. Moreover, by the choice of R R R ,
∥ m ˉ + u ( t ) ∥ L ∞ ( D ) ≤ ∥ m ˉ ∥ L ∞ ( D ) + C e m b R < 1 for all t ≥ 0 , \lVert\bar{m}+u(t)\rVert_{L^\infty(\mathcal{D})}\le\lVert\bar{m}\rVert_{L^\infty(\mathcal{D})}+C_{\mathrm{emb}}R<1\qquad\text{for all }t\ge0, ∥ m ˉ + u ( t ) ∥ L ∞ ( D ) ≤ ∥ m ˉ ∥ L ∞ ( D ) + C emb R < 1 for all t ≥ 0 ,
so the whole fixed-point ball remains uniformly separated from saturation. Hence, it has a unique fixed point u u u on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . Moreover,
∥ u ∥ E ω ≤ C ∥ m 0 − m ˉ ∥ X s , \lVert u\rVert_{E_\omega}\le C\lVert m_0-\bar{m}\rVert_{X^s}, ∥ u ∥ E ω ≤ C ∥ m 0 − m ˉ ∥ X s ,
which is precisely (147) . The choice of R R R also ensures
∥ m ˉ + u ( t ) ∥ L ∞ ( D ) < 1 for all t ≥ 0 , \lVert\bar{m}+u(t)\rVert_{L^\infty(\mathcal{D})}<1\qquad\text{for all }t\ge0, ∥ m ˉ + u ( t ) ∥ L ∞ ( D ) < 1 for all t ≥ 0 ,
so the solution remains uniformly separated from saturation. This completes the proof. □
As discussed in § 2.5 , near the paramagnetic state m = 0 m=0 m = 0 the full model formally reduces above the bulk Curie temperature to the standard reduced LLB equation (28) , for which existence of global strong solutions for d = 3 d=3 d = 3 is typically proved under a small-data assumption; see, e.g., [24 , 29 ] . For our problem, the condition 3 θ > Λ ∗ 3\theta>\Lambda_* 3 θ > Λ ∗ provides the corresponding stability condition for the zero equilibrium. Motivated by this connection, the following corollary gives global strong well-posedness for sufficiently small perturbations of m = 0 m=0 m = 0 , in the Sobolev scale s ∈ ( 3 / 2 , 2 ] s\in(3/2,2] s ∈ ( 3/2 , 2 ] .
Corollary 5.3 (Small initial magnetization). Let d = 3 d=3 d = 3 , s ∈ ( 3 / 2 , 2 ] s\in(3/2,2] s ∈ ( 3/2 , 2 ] , and suppose 3 θ > Λ ∗ 3\theta>\Lambda_* 3 θ > Λ ∗ , where Λ ∗ \Lambda_* Λ ∗ was defined in (129) , or equivalently θ > θ c \theta>\theta_c θ > θ c . Then, sufficiently small initial data in H D s H^s_{\mathcal{D}} H D s generate a unique global strong solution satisfying (147) with m ˉ = 0 \bar{m}=0 m ˉ = 0 .
Proof. Since G ( 0 ) = 0 G(0)=0 G ( 0 ) = 0 and D G ( 0 ) = 3 I DG(0)=3I D G ( 0 ) = 3 I , zero is an equilibrium and
⟨ L 0 z , z ⟩ = α ∥ ∇ z ∥ L 2 2 − ⟨ B z , z ⟩ − ⟨ H d [ z ] , z ⟩ + 3 θ ∥ z ∥ L 2 2 ≥ ( 3 θ − Λ ∗ ) ∥ z ∥ L 2 2 . \langle\mathcal{L}_0z,z\rangle=\alpha\lVert\nabla z\rVert_{L^2}^2-\langle Bz,z\rangle-\langle H_d[z],z\rangle+3\theta\lVert z\rVert_{L^2}^2\ge(3\theta-\Lambda_*)\lVert z\rVert_{L^2}^2. ⟨ L 0 z , z ⟩ = α ∥ ∇ z ∥ L 2 2 − ⟨ B z , z ⟩ − ⟨ H d [ z ] , z ⟩ + 3 θ ∥ z ∥ L 2 2 ≥ ( 3 θ − Λ ∗ ) ∥ z ∥ L 2 2 .
Thus, (146) holds. The conclusion then follows from Theorem 5.2 . □
References
Appendix A. Properties of the nonlinear coefficients
We record several elementary properties of q q q and M \mathrm{M} M which are used in our analysis.
Lemma A.1. Let L \mathcal{L} L be the Langevin function defined in (9) and let L − 1 \mathcal{L}^{-1} L − 1 be its inverse. Then, the function
q ( s ) = s L − 1 ( s ) , 0 < s < 1 , q(s)=\frac{s}{\mathcal{L}^{-1}(s)}, \qquad0<s<1, q ( s ) = L − 1 ( s ) s , 0 < s < 1 , admits a smooth extension to [ 0 , 1 ] [0,1] [ 0 , 1 ] satisfying q ( 0 ) = 1 3 q(0)=\frac{1}{3} q ( 0 ) = 3 1 and q ( 1 ) = 0 q(1)=0 q ( 1 ) = 0 , and it is strictly decreasing on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Moreover, there exist constants c q , C q > 0 c_q,C_q>0 c q , C q > 0 such that
c q ( 1 − s 2 ) ≤ q ( s ) ≤ C q ( 1 − s 2 ) for all s ∈ [ 0 , 1 ] . (151) c_q(1-s^2)\le q(s)\le C_q(1-s^2) \qquad\text{for all }s\in[0,1]. \tag*{(151)} c q ( 1 − s 2 ) ≤ q ( s ) ≤ C q ( 1 − s 2 ) for all s ∈ [ 0 , 1 ] . (151)
Proof. The Maclaurin expansion of the Langevin function is
L ( ξ ) = ξ 3 − ξ 3 45 + O ( ξ 5 ) as ξ → 0. \mathcal{L}(\xi)=\frac{\xi}{3}-\frac{\xi^3}{45}+O(\xi^5) \qquad\text{as }\xi\to0. L ( ξ ) = 3 ξ − 45 ξ 3 + O ( ξ 5 ) as ξ → 0.
e n d s m i d = 1 {ends_mid=1} e n d s m i d = 1
Since L ′ ( 0 ) = 1 / 3 \mathcal{L}'(0)=1/3 L ′ ( 0 ) = 1/3 , its local inverse is analytic at the origin and satisfies
L − 1 ( s ) = 3 s + 9 5 s 3 + O ( s 5 ) . \mathcal{L}^{-1}(s)=3s+\frac{9}{5}s^3+O(s^5). L − 1 ( s ) = 3 s + 5 9 s 3 + O ( s 5 ) .
Consequently,
q ( s ) = 1 3 − 1 5 s 2 + O ( s 4 ) as s → 0 , q(s)=\frac{1}{3}-\frac{1}{5}s^2+O(s^4)\qquad\text{as }s\to0, q ( s ) = 3 1 − 5 1 s 2 + O ( s 4 ) as s → 0 ,
and the analyticity of L − 1 \mathcal{L}^{-1} L − 1 shows that q q q extends smoothly to s = 0 s=0 s = 0 with q ( 0 ) = 1 / 3 q(0)=1/3 q ( 0 ) = 1/3 . To verify monotonicity, we use the Mittag–Leffler expansion
coth ξ = 1 ξ + 2 ξ ∑ k = 1 ∞ 1 ξ 2 + π 2 k 2 . \coth\xi=\frac{1}{\xi}+2\xi\sum_{k=1}^{\infty}\frac{1}{\xi^2+\pi^2k^2}. coth ξ = ξ 1 + 2 ξ k = 1 ∑ ∞ ξ 2 + π 2 k 2 1 .
It follows that
L ( ξ ) ξ = 2 ∑ k = 1 ∞ 1 ξ 2 + π 2 k 2 , \frac{\mathcal{L}(\xi)}{\xi}=2\sum_{k=1}^{\infty}\frac{1}{\xi^2+\pi^2k^2}, ξ L ( ξ ) = 2 k = 1 ∑ ∞ ξ 2 + π 2 k 2 1 ,
which is strictly decreasing for ξ > 0 \xi>0 ξ > 0 . Since q ( L ( ξ ) ) = L ( ξ ) / ξ q(\mathcal{L}(\xi))=\mathcal{L}(\xi)/\xi q ( L ( ξ )) = L ( ξ ) / ξ and L \mathcal{L} L is strictly increasing, q q q is strictly decreasing. Furthermore, by noting that coth ξ = 1 + 2 e − 2 ξ ( 1 − e − 2 ξ ) − 1 \coth\xi=1+2e^{-2\xi}(1-e^{-2\xi})^{-1} coth ξ = 1 + 2 e − 2 ξ ( 1 − e − 2 ξ ) − 1 , we obtain
L ( ξ ) = 1 − 1 ξ + O ( e − 2 ξ ) as ξ → ∞ . \mathcal{L}(\xi)=1-\frac{1}{\xi}+O(e^{-2\xi})\qquad\text{as }\xi\to\infty. L ( ξ ) = 1 − ξ 1 + O ( e − 2 ξ ) as ξ → ∞.
To see that the extension at s = 1 s=1 s = 1 is smooth, set u : = 1 / ξ u:=1/\xi u := 1/ ξ and define
F ( u ) : = L ( 1 / u ) = 1 − u + r ( u ) , r ( u ) = 2 e 2 / u − 1 . F(u):=\mathcal{L}(1/u)=1-u+r(u),\qquad r(u)=\frac{2}{e^{2/u}-1}. F ( u ) := L ( 1/ u ) = 1 − u + r ( u ) , r ( u ) = e 2/ u − 1 2 .
The function r r r extends smoothly to u = 0 u=0 u = 0 by r ( 0 ) = 0 r(0)=0 r ( 0 ) = 0 , since r ( u ) = O ( e − 2 / u ) r(u)=O(e^{-2/u}) r ( u ) = O ( e − 2/ u ) and all derivatives of r r r vanish as u ↓ 0 u\downarrow0 u ↓ 0 . Since F ( 0 ) = 1 F(0)=1 F ( 0 ) = 1 and F ′ ( 0 ) = − 1 F'(0)=-1 F ′ ( 0 ) = − 1 , the inverse function theorem yields a smooth inverse u = u ( s ) u=u(s) u = u ( s ) near s = 1 s=1 s = 1 , with u ( 1 ) = 0 u(1)=0 u ( 1 ) = 0 . For s < 1 s<1 s < 1 , this inverse satisfies u ( s ) = 1 / L − 1 ( s ) u(s)=1/\mathcal{L}^{-1}(s) u ( s ) = 1/ L − 1 ( s ) . Consequently, q ( s ) = s u ( s ) q(s)=su(s) q ( s ) = s u ( s ) , so q q q extends smoothly to s = 1 s=1 s = 1 with q ( 1 ) = 0 q(1)=0 q ( 1 ) = 0 .
In particular, as s ↑ 1 s\uparrow1 s ↑ 1 , we have
L − 1 ( s ) ∼ 1 1 − s , q ( s ) ∼ s ( 1 − s ) . \mathcal{L}^{-1}(s)\sim\frac{1}{1-s},\qquad q(s)\sim s(1-s). L − 1 ( s ) ∼ 1 − s 1 , q ( s ) ∼ s ( 1 − s ) .
Hence,
q ( s ) 1 − s 2 → 1 2 as s ↑ 1. \frac{q(s)}{1-s^2}\to\frac{1}{2}\qquad\text{as }s\uparrow1. 1 − s 2 q ( s ) → 2 1 as s ↑ 1.
The quotient q ( s ) / ( 1 − s 2 ) q(s)/(1-s^2) q ( s ) / ( 1 − s 2 ) thus extends to a positive continuous function on [ 0 , 1 ] [0,1] [ 0 , 1 ] , with values 1 / 3 1/3 1/3 at s = 0 s=0 s = 0 and 1 / 2 1/2 1/2 at s = 1 s=1 s = 1 . Its positive minimum and finite maximum yield (A.1). □ \square □
Lemma A.2. The matrix M \mathrm{M} M defined by (11) admits an extension M ∈ C 1 , 1 ( B 1 ( 0 ) ‾ ; R 3 × 3 ) \mathrm{M}\in C^{1,1}(\overline{B_1(0)};\mathbb{R}^{3\times3}) M ∈ C 1 , 1 ( B 1 ( 0 ) ; R 3 × 3 ) , with
M ( 0 ) = 2 3 I , M ( v ) = I − v ⊗ v if ∣ v ∣ = 1. \mathrm{M}(0)=\frac{2}{3}\mathrm{I},\qquad\mathrm{M}(v)=\mathrm{I}-v\otimes v\quad\text{if }|v|=1. M ( 0 ) = 3 2 I , M ( v ) = I − v ⊗ v if ∣ v ∣ = 1. Furthermore, for every v ∈ B 1 ( 0 ) v\in B_1(0) v ∈ B 1 ( 0 ) and z ∈ R 3 z\in\mathbb{R}^3 z ∈ R 3 ,
z ⋅ M ( v ) z ≤ ∣ z ∣ 2 , (152) z\cdot\mathrm{M}(v)z\leq|z|^2, \tag*{(152)} z ⋅ M ( v ) z ≤ ∣ z ∣ 2 , (152) 2 q ( ∣ v ∣ ) ∣ z ∣ 2 ≤ z ⋅ M ( v ) z , (153) 2q(|v|)|z|^2\leq z\cdot\mathrm{M}(v)z, \tag*{(153)} 2 q ( ∣ v ∣ ) ∣ z ∣ 2 ≤ z ⋅ M ( v ) z , (153) ∣ M ( v ) z ∣ 2 ≤ z ⋅ M ( v ) z , (154) |\mathrm{M}(v)z|^2\leq z\cdot\mathrm{M}(v)z, \tag*{(154)} ∣ M ( v ) z ∣ 2 ≤ z ⋅ M ( v ) z , (154) ∣ v × z ∣ 2 ≤ 3 2 z ⋅ M ( v ) z . (155) |v\times z|^2\leq\frac{3}{2}z\cdot\mathrm{M}(v)z. \tag*{(155)} ∣ v × z ∣ 2 ≤ 2 3 z ⋅ M ( v ) z . (155)
Proof. For v ≠ 0 v \ne0 v = 0 , write
z = z ∥ + z ⊥ , z ∥ = v ⊗ v ∣ v ∣ 2 z , z ⊥ = z − z ∥ . z = z_{\parallel} + z_{\perp}, \qquad z_{\parallel} = \frac{v \otimes v}{|v|^2}z, \qquad z_{\perp} = z - z_{\parallel}. z = z ∥ + z ⊥ , z ∥ = ∣ v ∣ 2 v ⊗ v z , z ⊥ = z − z ∥ .
Then, we have
z ⋅ M ( v ) z = 2 q ( ∣ v ∣ ) ∣ z ∥ ∣ 2 + ( 1 − q ( ∣ v ∣ ) ) ∣ z ⊥ ∣ 2 . (156) z \cdot M(v)z = 2q(|v|)|z_{\parallel}|^2 + (1-q(|v|))|z_{\perp}|^2. \tag*{(156)} z ⋅ M ( v ) z = 2 q ( ∣ v ∣ ) ∣ z ∥ ∣ 2 + ( 1 − q ( ∣ v ∣ )) ∣ z ⊥ ∣ 2 . (156)
Since 0 ≤ q ≤ 1 / 3 0 \le q \le1/3 0 ≤ q ≤ 1/3 , the eigenvalues 2 q ( ∣ v ∣ ) 2q(|v|) 2 q ( ∣ v ∣ ) and 1 − q ( ∣ v ∣ ) 1-q(|v|) 1 − q ( ∣ v ∣ ) belong to [ 0 , 1 ] [0,1] [ 0 , 1 ] , with 2 q ( ∣ v ∣ ) ≤ 1 − q ( ∣ v ∣ ) 2q(|v|) \le1-q(|v|) 2 q ( ∣ v ∣ ) ≤ 1 − q ( ∣ v ∣ ) . This proves (A.2) and (A.3). Since λ 2 ≤ λ \lambda^2 \le\lambda λ 2 ≤ λ for λ ∈ [ 0 , 1 ] \lambda\in[0,1] λ ∈ [ 0 , 1 ] , the same decomposition yields (A.4). Moreover, (A.5) follows since
∣ v × z ∣ 2 = ∣ v ∣ 2 ∣ z ⊥ ∣ 2 ≤ ∣ z ⊥ ∣ 2 ≤ 3 2 z ⋅ M ( v ) z , |v \times z|^2 = |v|^2|z_{\perp}|^2 \le|z_{\perp}|^2 \le\frac{3}{2}z \cdot M(v)z, ∣ v × z ∣ 2 = ∣ v ∣ 2 ∣ z ⊥ ∣ 2 ≤ ∣ z ⊥ ∣ 2 ≤ 2 3 z ⋅ M ( v ) z ,
where we used 1 − q ( ∣ v ∣ ) ≥ 2 / 3 1-q(|v|) \ge2/3 1 − q ( ∣ v ∣ ) ≥ 2/3 .
It remains to verify the regularity at the apparent singularity v = 0 v=0 v = 0 . Rewriting
M ( v ) = ( 1 − q ( ∣ v ∣ ) ) I + 3 q ( ∣ v ∣ ) − 1 ∣ v ∣ 2 v ⊗ v , M(v) = (1-q(|v|))I + \frac{3q(|v|)-1}{|v|^2}v \otimes v, M ( v ) = ( 1 − q ( ∣ v ∣ )) I + ∣ v ∣ 2 3 q ( ∣ v ∣ ) − 1 v ⊗ v ,
and using
q ( s ) = 1 3 − 1 5 s 2 + O ( s 4 ) as s → 0 , q(s) = \frac{1}{3} - \frac{1}{5}s^2 + O(s^4) \qquad\text{as }s \to0, q ( s ) = 3 1 − 5 1 s 2 + O ( s 4 ) as s → 0 ,
we obtain
M ( v ) = 2 3 I + 1 5 ∣ v ∣ 2 I − 3 5 v ⊗ v + O ( ∣ v ∣ 4 ) . M(v) = \frac{2}{3}I + \frac{1}{5}|v|^2I - \frac{3}{5}v \otimes v + O(|v|^4). M ( v ) = 3 2 I + 5 1 ∣ v ∣ 2 I − 5 3 v ⊗ v + O ( ∣ v ∣ 4 ) .
In particular, the singularity is removable and M ( 0 ) = 2 3 I M(0)=\frac{2}{3}I M ( 0 ) = 3 2 I . More generally, the expansion of q q q at the origin contains only even powers of s s s , so the functions
s ↦ 1 − q ( s ) , s ↦ 3 q ( s ) − 1 s 2 s \mapsto1-q(s), \qquad s \mapsto\frac{3q(s)-1}{s^2} s ↦ 1 − q ( s ) , s ↦ s 2 3 q ( s ) − 1
extend smoothly across s = 0 s=0 s = 0 as even functions. Together with the regular extension of q q q at s = 1 s=1 s = 1 , this shows that M M M is C 2 C^2 C 2 on B 1 ( 0 ) ‾ \overline{B_1(0)} B 1 ( 0 ) , and hence M ∈ C 1 , 1 ( B 1 ( 0 ) ‾ ; R 3 × 3 ) M \in C^{1,1}(\overline{B_1(0)};\mathbb{R}^{3 \times3}) M ∈ C 1 , 1 ( B 1 ( 0 ) ; R 3 × 3 ) .
Lemma A.3. Let p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) and s ∈ { 0 , 1 , 2 } s \in\{0,1,2\} s ∈ { 0 , 1 , 2 } . The stray-field linear operator H d H_{\mathrm d} H d satisfies
∥ H d [ v ] ∥ L p ( R 3 ) ≤ C p ∥ v ∥ L p ( D ) , (157) \|H_{\mathrm d}[v]\|_{L^p(\mathbb{R}^3)} \le C_p\|v\|_{L^p(\mathcal{D})}, \tag*{(157)} ∥ H d [ v ] ∥ L p ( R 3 ) ≤ C p ∥ v ∥ L p ( D ) , (157) ∥ H d [ v ] ∥ W s , p ( D ) ≤ C s , p ∥ v ∥ W s , p ( D ) . (158) \|H_{\mathrm d}[v]\|_{W^{s,p}(\mathcal{D})} \le C_{s,p}\|v\|_{W^{s,p}(\mathcal{D})}. \tag*{(158)} ∥ H d [ v ] ∥ W s , p ( D ) ≤ C s , p ∥ v ∥ W s , p ( D ) . (158)
Proof. See [7 Lemma 2.3] and [10 Section 3.1] .
Statements and declarations
Conflict of interest. The author has no competing interests to declare that are relevant to the content of this article.
Funding statement. This research was initiated while the author was supported by the Commonwealth through an Australian Government Research Training Program (RTP) Scholarship (DOI: 10.82133/C42F-K220) at UNSW Sydney. The author has also been supported by an Australian Mathematical Society Lift-off Fellowship. The work was completed during the author’s postdoctoral appointment at the Institute of Analysis and Scientific Computing of TU Wien, supported by the Austrian Science Fund (FWF) through the international project I6802 “Functional error estimates for PDEs on unbounded domains” (DOI: 10.55776/I6802), led by Prof. Dirk Praetorius.
Institute of Analysis and Scientific Computing, TU Wien, Wiedner Hauptstrasse 8–10, 1040 Vienna, Austria
Email address: agus.soenjaya@asc.tuwien.ac.at