Introduction
Within the scope of Brown’s theory [27 ] , the configuration of a magnetoelastic body Ω \Omega Ω is described by a deformation y : Ω → R 3 \mathbf{y}:\Omega\to\mathbb{R}^{3} y : Ω → R 3 , a magnetization density m : y ( Ω ) → R 3 \mathbf{m}:y(\Omega)\to\mathbb{R}^{3} m : y ( Ω ) → R 3 , and a self magnetic field h s : R 3 → R 3 \mathbf{h}_{\mathrm{s}}:\mathbb{R}^{3}\to\mathbb{R}^{3} h s : R 3 → R 3 . The self field h s \mathbf{h}_{\mathrm{s}} h s is determined by the magnetization density m \mathbf{m} m as the distributional solution of the Maxwell constraint :
\begin{aligned}
\operatorname{curl}\mathbf{h}_{\mathrm{s}} &= 0,\\
\operatorname{div}(\mathbf{h}_{\mathrm{s}}+\chi_{y(\Omega)}\mathbf{m}) &= 0,
\end{aligned}
\qquad\text{on }\mathbb{R}^{3}, \tag*{(1)}\end{aligned}
where χ y ( Ω ) m \chi_{y(\Omega)}\mathbf{m} χ y ( Ω ) m is the trivial extension of m \mathbf{m} m to R 3 \mathbb{R}^{3} R 3 . When a current density j : R 3 → R 3 \mathbf{j}:\mathbb{R}^{3}\to\mathbb{R}^{3} j : R 3 → R 3 is generated by an external source outside Ω \Omega Ω , the body is acted upon by an applied magnetic field h a : R 3 → R 3 \mathbf{h}_{\mathrm{a}}:\mathbb{R}^{3}\to\mathbb{R}^{3} h a : R 3 → R 3 , given by
\begin{aligned}
\operatorname{curl}\mathbf{h}_{\mathrm{a}} &= \mathbf{j},\\
\operatorname{div}\mathbf{h}_{\mathrm{a}} &= 0,
\end{aligned}
\qquad\text{in }\mathbb{R}^{3}, \tag*{(2)}\end{aligned}
and its stable equilibria are local minimizers, under the constraint (1) , of the magnetoelastic energy :
E ^ ( y , m , h s ) = ∫ Ω Φ ^ ( x , ∇ y , m ∘ y ) d x + μ 0 2 ∫ R 3 ∣ h s ∣ 2 d ξ − μ 0 ∫ y ( Ω ) h a ⋅ m d ξ , (3) \widehat{\mathcal{E}}(\mathbf{y},\mathbf{m},\mathbf{h}_{\mathrm{s}})=\int_{\Omega}\widehat{\Phi}(\mathbf{x},\nabla\mathbf{y},\mathbf{m}\circ\mathbf{y})\,\mathrm{d}\mathbf{x}+\frac{\mu_{0}}{2}\int_{\mathbb{R}^{3}}|\mathbf{h}_{\mathrm{s}}|^{2}\,\mathrm{d}\xi-\mu_{0}\int_{y(\Omega)}\mathbf{h}_{\mathrm{a}}\cdot\mathbf{m}\,\mathrm{d}\xi, \tag*{(3)} E ( y , m , h s ) = ∫ Ω Φ ( x , ∇ y , m ∘ y ) d x + 2 μ 0 ∫ R 3 ∣ h s ∣ 2 d ξ − μ 0 ∫ y ( Ω ) h a ⋅ m d ξ , (3)
which we write by adapting the format used by James and Kinderlehrer in their investigation of magnetic microstructures in magnetoelastic crystalline bodies undergoing large strains [12 ] .
Equation (1.2) is set in the framework of magnetostatics, which requires two conditions on the current density j \mathbf{j} j : that induces no charge accumulation, div j = 0 \operatorname{div}\mathbf{j}=0 div j = 0 , and that its time variations are slow compared to the time light takes to traverse the body, the so-called quasi-static regimes. The first condition is necessary for (1.2) to be consistent, since div ( curl h a ) = 0 \operatorname{div}(\operatorname{curl}\mathbf{h}_{a})=0 div ( curl h a ) = 0 identically. The second justifies neglecting the displacement current in Ampère’s law [11 ] .
In writing (1.3) we use the notation
( m ∘ y ) ( x ) = m [ y ( x ) ] , x ∈ Ω , (4) (\mathbf{m}\circ\mathbf{y})(\mathbf{x})=\mathbf{m}[\mathbf{y}(\mathbf{x})],\qquad\mathbf{x}\in\Omega, \tag*{(4)} ( m ∘ y ) ( x ) = m [ y ( x )] , x ∈ Ω , (4)
which emphasizes that, while the domain of definition of the deformation y \mathbf{y} y is the reference configuration Ω \Omega Ω , the magnetization density is defined in the deformed configuration y ( Ω ) \mathbf{y}(\Omega) y ( Ω ) . Thus, when integrating the energy density Φ ˉ \bar{\Phi} Φ ˉ over the reference configuration, taking the referential point x \mathbf{x} x as independent variable, the magnetization density must be evaluated at ξ = y ( x ) \boldsymbol{\xi}=\mathbf{y}(\mathbf{x}) ξ = y ( x ) .
Within this framework, the (total) magnetic field ,
h = h a + h s , (5) \mathbf{h}=\mathbf{h}_{a}+\mathbf{h}_{s}, \tag*{(5)} h = h a + h s , (5)
is recovered as the sum of the applied- and self-field. The magnetic induction, which is the quantity that can be actually measured, is given by
b = μ 0 ( m + h ) . (6) \mathbf{b}=\mu_{0}(\mathbf{m}+\mathbf{h}). \tag*{(6)} b = μ 0 ( m + h ) . (6)
Brown’s original formulation was developed to model ferromagnetic materials at the microscopic level where domain structures can be observed [23 , 24 , 25 , 26 ] . At this level of spatial resolution, the specific magnetization density m s = ( det ∇ y ) m \mathbf{m}_{s}=(\det\nabla\mathbf{y})\mathbf{m} m s = ( det ∇ y ) m in a ferromagnetic body obeys the saturation constraint ∣ m ∣ = m s \lvert\mathbf{m}\rvert=m_{s} ∣ m ∣ = m s . The non-convexity arising from such constraint leads to the possibility of multiple local minima and formation of microstructures, whose analysis can be dealt rigorously with relaxation and Young’s measures [5 , 14 ] or by introducing exchange energy, which penalizes spatial variations of m \mathbf{m} m , as discussed in the review [6 ] .
Kankanala and Triantafyllidis Kankanala and Triantafyllidis [13 ] leveraged Brown’s approach in situations beyond classical ferromagnets, such as soft magnetoelastic bodies — specifically, magnetorheological elastomers. Although they used a similar mathematical formulation, their physical perspective is different from Brown’s: in their work, the field m \mathbf{m} m represents the magnetization density at an observation scale much larger than that of magnetic domains. In particular, the saturation constraint is no longer enforced, and, more importantly, the detailed spatial profile of the magnetization is not the quantity of main interest. This second aspect perhaps explains why, in the context of soft magnetoelastic bodies, alternative formulations have emerged, which replace magnetization with other magnetic variables. The first of these alternative formulation was proposed by Dorfmann and Ogden Dorfmann and Ogden [9 ] . In their formulation of magneto-elasticity the magnetic state of the body is described by the Lagrangian magnetic induction B \mathbf{B} B , which is related to the (Eulerian) magnetic induction (1.6) by the following transformation formula:
B = J F − 1 ( b ∘ y ) , (7) \mathbf{B}=J\mathbf{F}^{-1}(\mathbf{b}\circ\mathbf{y}), \tag*{(7)} B = J F − 1 ( b ∘ y ) , (7)
where
F = ∇ y , J = det F . (8) \mathbf{F}=\nabla\mathbf{y},\qquad J=\det\mathbf{F}. \tag*{(8)} F = ∇ y , J = det F . (8)
The Dorfmann & Ogden theory admits a variational formulation, which has been discussed by Bustamante et al. Bustamante et al. [2 ] . This formulation has been considered, with minor changes, by Šilhavý Šilhavý [22 ] in his discussion of existence theorem in non-linear electro-magneto-elasticity. With suitable modifications to align with the form of (1.3), the functional form of the energy considered by Šilhavý can be written (up to a constant) as
E ˉ ( y , B ) = ∫ Ω Φ ˉ ( x , F , B ) d x + 1 2 μ 0 ∫ R 3 { 1 J ∣ F ( B − J ( b a ∘ y ) ) ∣ 2 } d x , (9) \bar{\mathcal{E}}(\mathbf{y},\mathbf{B})=\int_{\Omega}\bar{\Phi}(\mathbf{x},\mathbf{F},\mathbf{B})\,\mathrm{d}\mathbf{x}+\frac{1}{2\mu_{0}}\int_{\mathbb{R}^{3}}\left\{\frac{1}{J}\left\lvert\mathbf{F}(\mathbf{B}-J(\mathbf{b}_{a}\circ\mathbf{y}))\right\rvert^{2}\right\}\,\mathrm{d}\mathbf{x}, \tag*{(9)} E ˉ ( y , B ) = ∫ Ω Φ ˉ ( x , F , B ) d x + 2 μ 0 1 ∫ R 3 { J 1 ∣ F ( B − J ( b a ∘ y )) ∣ 2 } d x , (9)
where y y y has been extended to R 3 \mathbb{R}^{3} R 3 in a suitable way.
Equilibrium states are identified with the critical points of E ‾ \overline{\mathcal{E}} E under the constraint
Div B = 0 , (10) \operatorname{Div}\mathbf{B}=0, \tag*{(10)} Div B = 0 , (10)
which thanks to the definition (7) is equivalent to the solenoidal constraint :
div b = 0. (11) \operatorname{div}\mathbf{b}=0. \tag*{(11)} div b = 0. (11)
Working with the Lagrangian induction field B \mathbf{B} B bypasses several difficulties inherent with the fact that the magnetic state variables are naturally defined in the physical space, here represented by R 3 \mathbb{R}^{3} R 3 , whereas the domain of definition of the deformation is a subdomain Ω \Omega Ω of a fictitious space still modeled with R 3 \mathbb{R}^{3} R 3 , thus simplifying actual calculations, be they analytical or numerical. However, it introduces the conceptual difficulty of extending the deformation gradient outside Ω \Omega Ω , which is not always possible, since a sufficiently regular extension of the deformation may not exist. Since in this paper we are concerned with theoretical aspects, we shall consider critical points, subjected to the constraint (11) , of the following functional:
E ( y , b ) = ∫ Ω Φ ( x , ∇ y , b ∘ y ) d x + 1 2 μ 0 ∫ R 3 ∣ b − b a ∣ 2 d ξ . (12) \mathcal{E}(y,\mathbf{b})=\int_{\Omega}\Phi(x,\nabla y,\mathbf{b}\circ y)\,dx+\frac{1}{2\mu_{0}}\int_{\mathbb{R}^{3}}|\mathbf{b}-\mathbf{b}_{a}|^{2}\,d\xi. \tag*{(12)} E ( y , b ) = ∫ Ω Φ ( x , ∇ y , b ∘ y ) d x + 2 μ 0 1 ∫ R 3 ∣ b − b a ∣ 2 d ξ . (12)
If the function Φ \Phi Φ is related to Φ ‾ \overline{\Phi} Φ by the transformation formula
Φ ( x , F , b ) = Φ ‾ ( x , F , J F − 1 b ) , (13) \Phi(x,F,\mathbf{b})=\overline{\Phi}(x,F,JF^{-1}\mathbf{b}), \tag*{(13)} Φ ( x , F , b ) = Φ ( x , F , J F − 1 b ) , (13)
then by (7) , E \mathcal{E} E and E ‾ \overline{\mathcal{E}} E coincide.
For many cases of engineering interest, at least one of the two functionals b ↦ E ( y , b ) \mathbf{b}\mapsto\mathcal{E}(y,\mathbf{b}) b ↦ E ( y , b ) and ( m , h s ) ↦ E ^ ( y , m , h s ) (\mathbf{m},\mathbf{h}_{s})\mapsto\widehat{\mathcal{E}}(y,\mathbf{m},\mathbf{h}_{s}) ( m , h s ) ↦ E ( y , m , h s ) is convex and coercive in their respective magnetic variables. In particular, for any fixed y y y that functionals then admits a unique minimizer. However, while the formulation based on (12) is local, the one based on m \mathbf{m} m and h s \mathbf{h}_{s} h s is not. In fact, by writing
b = curl a in R 3 , (14) \mathbf{b}=\operatorname{curl}\mathbf{a}\quad\text{in }\mathbb{R}^{3}, \tag*{(14)} b = curl a in R 3 , (14)
and adopting the vector potential a \mathbf{a} a as the unknown, the solenoidal constraint (11) is enforced identically, up to the usual gauge invariance a ↦ a + ∇ u \mathbf{a}\mapsto\mathbf{a}+\nabla u a ↦ a + ∇ u , and the resulting Euler-Lagrange equations are pointwise. By contrast, even if the irrotationality constraint curl h s = 0 \operatorname{curl}\mathbf{h}_{s}=0 curl h s = 0 in the ( m , h ) (\mathbf{m},\mathbf{h}) ( m , h ) -based formulation is enforced by introducing a scalar potential φ \varphi φ and by setting h s = − ∇ φ \mathbf{h}_{s}=-\nabla\varphi h s = − ∇ φ , the dependence of φ \varphi φ from m \mathbf{m} m is non-local.
The equivalence between the two formulations has been discussed in several papers in the engineering literature [2 , 4 , 21 ] . Although all of these papers, either implicitly or explicitly acknowledge the role convex duality, the only rigorous discussion has been given, to our knowledge, by Pedregal & Yan [17 ] . More precisely, [17 ] only considers the special case of a coercive micromagnetics energy density with a hard saturation constraint, a nonconvex but still coercive case where full model equivalence can only be expected after a suitable relaxation. Interestingly, if the hard saturation constraint or coercivity are dropped, the link to the single-field model can change significantly, see Example 4.5 (close to [17 ] ), but when starting from the ( m , h ) (\mathbf{m},\mathbf{h}) ( m , h ) -model, the hard saturation constraint is never fully recovered when passing to the ( m , h ) (\mathbf{m},\mathbf{h}) ( m , h ) -model, and softer versions of saturation are not covered in [17 ] at all, and Example 4.1 (demonstrating that in the ( m , h ) (\mathbf{m},\mathbf{h}) ( m , h ) -formulation, diamagnetic materials naturally lead to concave functionals that should be maximized, not minimized). Both are natural cases that we cover here as well.
The goal of the present paper is to provide a precise and rigorous discussion of the equivalence of the two formulations, which goes beyond the related results of [17 ] . With this, we hope to fill what we believe to be a gap in the existing literature, where the precise meaning of such equivalence is not stated clearly. In particular, we wish to make the point that a more careful treatment is needed, compared to the existing literature cited above. In doing so, we still want to keep the level of mathematical technicality accessible to the engineering community. In order to outline our line of arguments, we present our equivalence statement in a special case, and then we comment on the result. For simplicity, we restrict ourselves to the smooth case, i.e., where all densities are smooth, and we assume that the material energy density Ψ ^ ( x , F , m ) \widehat{\Psi}(x,\mathbf{F},\mathbf{m}) Ψ ( x , F , m ) is convex with respect to m \mathbf{m} m .
Equivalence statement for the smooth convex case. In discussing the equivalence between the two formulations, the deformation y : Ω → R 3 \mathbf{y}:\Omega\to\mathbb{R}^{3} y : Ω → R 3 does not interfere at all with the equivalence of the magnetic models, and hence can be treated as a fixed parameter. We write E y ( b ) : = E ( y , b ) \mathcal{E}_{\mathbf{y}}(\mathbf{b}):=\mathcal{E}(\mathbf{y},\mathbf{b}) E y ( b ) := E ( y , b ) and E ^ y ( m , h s ) : = E ^ ( y , m , h s ) \widehat{\mathcal{E}}_{\mathbf{y}}(\mathbf{m},\mathbf{h}_{s}):=\widehat{\mathcal{E}}(\mathbf{y},\mathbf{m},\mathbf{h}_{s}) E y ( m , h s ) := E ( y , m , h s ) for the two energy functionals viewed as functionals of the magnetic variables only.
The key hypothesis for the equivalence of the two formulations is that the energy densities Φ \Phi Φ and Φ ^ \widehat{\Phi} Φ are linked by the following transformation formula:
Φ ( x , F , b ) = − Ψ ^ ∗ ( x , F , b ) . (15) \Phi(x,\mathbf{F},\mathbf{b})=-\widehat{\Psi}^{*}(x,\mathbf{F},\mathbf{b}). \tag*{(15)} Φ ( x , F , b ) = − Ψ ∗ ( x , F , b ) . (15)
Here, Ψ ^ \widehat{\Psi} Ψ is the augmented material energy density , given by
Ψ ^ ( x , F , m ) : = Φ ^ ( x , F , m ) + μ 0 2 ∣ m ∣ 2 , (16) \widehat{\Psi}(x,\mathbf{F},\mathbf{m}):=\widehat{\Phi}(x,\mathbf{F},\mathbf{m})+\frac{\mu_{0}}{2}\lvert\mathbf{m}\rvert^{2}, \tag*{(16)} Ψ ( x , F , m ) := Φ ( x , F , m ) + 2 μ 0 ∣ m ∣ 2 , (16)
where Ψ ^ ∗ \widehat{\Psi}^{*} Ψ ∗ is the Legendre-Fenchel transform of Ψ ^ \widehat{\Psi} Ψ in the variable m \mathbf{m} m , given by (A.17) (still in the smooth cases, for energies with invertible gradient, Ψ ^ ∗ \widehat{\Psi}^{*} Ψ ∗ will be replaced by the generalized Legendre-Fenchel transform Ψ ^ ⋄ \widehat{\Psi}^{\diamond} Ψ ⋄ defined in (A.7)). Accordingly, Φ ( x , F , b ) \Phi(x,\mathbf{F},\mathbf{b}) Φ ( x , F , b ) is concave in b \mathbf{b} b . Moreover, the applied fields are linked by h a = μ 0 b a \mathbf{h}_{a}=\mu_{0}\mathbf{b}_{a} h a = μ 0 b a .
The reverse transformation corresponding to (15) is
Φ ^ ( x , F , m ) = ( − Φ ) ∗ ( x , F , m ) − μ 0 2 ∣ m ∣ 2 , (17) \widehat{\Phi}(x,\mathbf{F},\mathbf{m})=(-\Phi)^{*}(x,\mathbf{F},\mathbf{m})-\frac{\mu_{0}}{2}\lvert\mathbf{m}\rvert^{2}, \tag*{(17)} Φ ( x , F , m ) = ( − Φ ) ∗ ( x , F , m ) − 2 μ 0 ∣ m ∣ 2 , (17)
with the Legendre-Fenchel transform now acting in the variable b \mathbf{b} b of Φ \Phi Φ . Here we omit some technical assumptions on Ω \Omega Ω , y \mathbf{y} y , Φ ^ \widehat{\Phi} Φ and Φ \Phi Φ that we shall specify in the body of the paper. With these assumption, the equivalence statement we prove can be stated as follows in the simplest case:
(i) If ( m ‾ , h ‾ s ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_{s}) ( m , h s ) is a critical point of E ^ y \widehat{\mathcal{E}}_{\mathbf{y}} E y under the Maxwell constraint (1) , then the field b ‾ \overline{\mathbf{b}} b defined by
b ‾ = b a + μ 0 ( χ y ( Ω ) m ‾ + h ‾ s ) in R 3 . (18) \overline{\mathbf{b}}=\mathbf{b}_{a}+\mu_{0}\left(\chi_{\mathbf{y}(\Omega)}\overline{\mathbf{m}}+\overline{\mathbf{h}}_{s}\right)\quad\text{in }\mathbb{R}^{3}. \tag*{(18)} b = b a + μ 0 ( χ y ( Ω ) m + h s ) in R 3 . (18)
is a critical point of E y \mathcal{E}_{\mathbf{y}} E y under the solenoidal constraint (11) , with Φ \Phi Φ given by (15) . Moreover,
E y ( b ‾ ) = E ^ y ( m ‾ , h ‾ s ) . (19) \mathcal{E}_{\mathbf{y}}(\overline{\mathbf{b}})=\widehat{\mathcal{E}}_{\mathbf{y}}(\overline{\mathbf{m}},\overline{\mathbf{h}}_{s}). \tag*{(19)} E y ( b ) = E y ( m , h s ) . (19)
(ii) Conversely, if b ‾ \overline{\mathbf{b}} b is a critical point of E y \mathcal{E}_{\mathbf{y}} E y under the solenoidal constraint (11) , then for m ‾ \overline{\mathbf{m}} m defined via the constitutive relation
m ‾ ∘ y = − ∇ b Φ ( ⋅ , F , b ‾ ∘ y ) in Ω , (20) \overline{\mathbf{m}}\circ\mathbf{y}=-\nabla_{\mathbf{b}}\Phi(\cdot,\mathbf{F},\overline{\mathbf{b}}\circ\mathbf{y})\quad\text{in }\Omega, \tag*{(20)} m ∘ y = − ∇ b Φ ( ⋅ , F , b ∘ y ) in Ω , (20)
and h ‾ s \overline{\mathbf{h}}_{s} h s is the unique solution of (1) , equivalently, the curl-free part of − χ y ( Ω ) m ‾ -\chi_{\mathbf{y}(\Omega)}\overline{\mathbf{m}} − χ y ( Ω ) m from the Helmholtz decomposition on R 3 \mathbb{R}^{3} R 3 (cf. (31) ), the pair ( m ‾ , h ‾ s ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_{s}) ( m , h s ) is a critical point of E ^ y \widehat{\mathcal{E}}_{\mathbf{y}} E y under the Maxwell constraint (1) , with Φ ^ y \widehat{\Phi}_{\mathbf{y}} Φ y given by (17) . Moreover, (19) and (18) are again satisfied.
From the above equivalence statement, it emerges that it is crucially important to distinguish equilibrium points from general admissible states. Precisely, given a generic state ( m , h s ) (\mathbf{m},\mathbf{h}_{s}) ( m , h s ) that does not satisfy the Euler–Lagrange equations, the corresponding b \mathbf{b} b defined by (1.18) does not satisfy E y ( b ) = E ^ y ( m , h s ) \mathcal{E}_{y}(\mathbf{b})=\widehat{\mathcal{E}}_{y}(\mathbf{m},\mathbf{h}_{s}) E y ( b ) = E y ( m , h s ) as can be easily checked for concrete examples, for instance the diamagnetic case of Example 4.1 where E y \mathcal{E}_{y} E y is convex and coercive whereas E ^ y \widehat{\mathcal{E}}_{y} E y is not bounded from below. Conversely, if b \mathbf{b} b is a generic admissible state that does not satisfy the Euler–Lagrange equations, the corresponding m \mathbf{m} m defined by (1.20) and the corresponding h s \mathbf{h}_{s} h s obtained by solving (1.1) do not satisfy E y ( b ) = E ^ y ( m , h s ) \mathcal{E}_{y}(\mathbf{b})=\widehat{\mathcal{E}}_{y}(\mathbf{m},\mathbf{h}_{s}) E y ( b ) = E y ( m , h s ) either. In other words, the transformation linking the two models is such that the energies coincide only at equilibrium. In our opinion, such distinction has not been sufficiently emphasized in the engineering literature.
Fully rigorous versions of the above equivalence statement including precise assumptions are presented in Theorem 2.6 and Corollary 2.15 (smooth case, Section 2), and in Theorems 3.6 and 3.7 and Corollary 3.11 (convex/concave case, Section 3). These results also discuss cases beyond the convex one. Examples are given in Section 4.
The smooth case
Throughout this section and the rest of the paper, we regard the deformation y : Ω → R 3 \mathbf{y}:\Omega\to\mathbb{R}^{3} y : Ω → R 3 as fixed and we make the assumption that the deformation is regular enough so that the image y ( Ω ) \mathbf{y}(\Omega) y ( Ω ) is a measurable subset of R 3 \mathbb{R}^{3} R 3 . As to the applied field, we assume that b a ∈ L 2 ( R 3 ; R 3 ) \mathbf{b}_{a}\in L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) b a ∈ L 2 ( R 3 ; R 3 ) with div b a = 0 \operatorname{div}\mathbf{b}_{a}=0 div b a = 0 on R 3 \mathbb{R}^{3} R 3 . Since these assumption are common to all theorem statements, we shall leave them tacit. The two energy functionals that we compare are (1.12) and (1.3), but now the domain of integration is the deformed configuration after the change of variables ξ = y ( x ) \boldsymbol{\xi}=\mathbf{y}(\mathbf{x}) ξ = y ( x ) :
E y ( b ) : = ∫ y ( Ω ) Φ y ( ξ , b ) d ξ + 1 2 μ 0 ∫ R 3 ∣ b − b a ∣ 2 d ξ , (21) \mathcal{E}_{y}(\mathbf{b}):=\int_{\mathbf{y}(\Omega)}\Phi_{y}(\boldsymbol{\xi},\mathbf{b})\,\mathrm{d}\boldsymbol{\xi}+\frac{1}{2\mu_{0}}\int_{\mathbb{R}^{3}}\lvert\mathbf{b}-\mathbf{b}_{a}\rvert^{2}\,\mathrm{d}\boldsymbol{\xi}, \tag*{(21)} E y ( b ) := ∫ y ( Ω ) Φ y ( ξ , b ) d ξ + 2 μ 0 1 ∫ R 3 ∣ b − b a ∣ 2 d ξ , (21)
E ^ y ( m , h s ) : = ∫ y ( Ω ) Φ ^ y ( ξ , m ) d ξ + μ 0 2 ∫ R 3 ∣ h s ∣ 2 d ξ − ∫ y ( Ω ) m ⋅ b a d ξ , (22) \widehat{\mathcal{E}}_{y}(\mathbf{m},\mathbf{h}_{s}):=\int_{\mathbf{y}(\Omega)}\widehat{\Phi}_{y}(\boldsymbol{\xi},\mathbf{m})\,\mathrm{d}\boldsymbol{\xi}+\frac{\mu_{0}}{2}\int_{\mathbb{R}^{3}}\lvert\mathbf{h}_{s}\rvert^{2}\,\mathrm{d}\boldsymbol{\xi}-\int_{\mathbf{y}(\Omega)}\mathbf{m}\cdot\mathbf{b}_{a}\,\mathrm{d}\boldsymbol{\xi}, \tag*{(22)} E y ( m , h s ) := ∫ y ( Ω ) Φ y ( ξ , m ) d ξ + 2 μ 0 ∫ R 3 ∣ h s ∣ 2 d ξ − ∫ y ( Ω ) m ⋅ b a d ξ , (22)
Here the transformed material energy densities are
Φ y ( ξ , b ) : = 1 J ( y − 1 ( ξ ) ) Φ ( y − 1 ( ξ ) , F ( y − 1 ( ξ ) ) , b ) , (23) \Phi_{y}(\boldsymbol{\xi},\mathbf{b}):=\frac{1}{J(\mathbf{y}^{-1}(\boldsymbol{\xi}))}\Phi\bigl(\mathbf{y}^{-1}(\boldsymbol{\xi}),F(\mathbf{y}^{-1}(\boldsymbol{\xi})),\mathbf{b}\bigr), \tag*{(23)} Φ y ( ξ , b ) := J ( y − 1 ( ξ )) 1 Φ ( y − 1 ( ξ ) , F ( y − 1 ( ξ )) , b ) , (23)
Φ ^ y ( ξ , m ) : = 1 J ( y − 1 ( ξ ) ) Φ ^ ( y − 1 ( ξ ) , F ( y − 1 ( ξ ) ) , m ) . (24) \widehat{\Phi}_{y}(\boldsymbol{\xi},\mathbf{m}):=\frac{1}{J(\mathbf{y}^{-1}(\boldsymbol{\xi}))}\widehat{\Phi}\bigl(\mathbf{y}^{-1}(\boldsymbol{\xi}),F(\mathbf{y}^{-1}(\boldsymbol{\xi})),\mathbf{m}\bigr). \tag*{(24)} Φ y ( ξ , m ) := J ( y − 1 ( ξ )) 1 Φ ( y − 1 ( ξ ) , F ( y − 1 ( ξ )) , m ) . (24)
In this section we will show that the two models, the one based on the energy E ^ y ( m , h s ) \widehat{\mathcal{E}}_{y}(\mathbf{m},\mathbf{h}_{s}) E y ( m , h s ) and the one based on the energy E y ( b ) \mathcal{E}_{y}(\mathbf{b}) E y ( b ) , are indeed equivalent provided that the energy densities Φ ^ y ( ξ , m ) \widehat{\Phi}_{y}(\boldsymbol{\xi},\mathbf{m}) Φ y ( ξ , m ) and Φ y ( ξ , b ) \Phi_{y}(\boldsymbol{\xi},\mathbf{b}) Φ y ( ξ , b ) are related by a suitable duality transformation (see (H0), (H0′) and (2.28) below). Such equivalence holds both at the level of stationary points and at the level of energies.
For clarity of exposition, we divide this section into three distinct subsections. In the first one, we give the notion of constrained critical point. In the second one, we discuss the correspondence between constrained critical points of the two models at the level of the Euler–Lagrange equations. In the third one, we show the energies at the two models coincide at critical points.
Constrained critical points
To begin, we recapitulate the Euler–Lagrange equations of the two models. These equations are well known, and their derivation is given in Proposition A.9 in the Appendix for the reader’s convenience. If the energy functionals are differentiable, states that satisfy the
Euler-Lagrange equations are natural candidates for equilibrium states, and we call them smooth constrained critical points .
Definition 2.1 (Smooth constrained critical points).
(i) We say that ( m ‾ , h ‾ s ) ∈ L 2 ( Ω ; R 3 ) × L 2 ( R 3 ; R 3 ) (\overline{m},\overline{h}_{s}) \in L^{2}(\Omega;\mathbb{R}^{3}) \times L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) ( m , h s ) ∈ L 2 ( Ω ; R 3 ) × L 2 ( R 3 ; R 3 ) is a smooth constrained critical point of the functional E ^ y \widehat{\mathcal{E}}_{y} E y defined in (22) if the Maxwell equations (1) hold:
div ( χ y ( Ω ) m ‾ + h ‾ s ) = 0 , curl h ‾ s = 0 in R 3 , (25) \operatorname{div}\left(\chi_{y(\Omega)}\overline{m}+\overline{h}_{s}\right)=0,\qquad\operatorname{curl}\overline{h}_{s}=0\quad\text{in }\mathbb{R}^{3}, \tag*{(25)} div ( χ y ( Ω ) m + h s ) = 0 , curl h s = 0 in R 3 , (25) and the following Euler-Lagrange equation holds:
∇ m Φ ^ y ( ⋅ , m ‾ ) = μ 0 h ‾ s + b a a.e. in y ( Ω ) . (26) \nabla_{m}\widehat{\Phi}_{y}(\cdot,\overline{m})=\mu_{0}\overline{h}_{s}+b_{a}\quad\text{a.e. in }y(\Omega). \tag*{(26)} ∇ m Φ y ( ⋅ , m ) = μ 0 h s + b a a.e. in y ( Ω ) . (26) (ii) A field b ‾ ∈ L 2 ( R 3 ; R 3 ) \overline{b}\in L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) b ∈ L 2 ( R 3 ; R 3 ) is a smooth constrained critical point of the functional E y \mathcal{E}_{y} E y defined in (21) if div b ‾ = 0 \operatorname{div}\overline{b}=0 div b = 0 , and there exists a scalar potential φ ‾ ∈ W ˚ 1 , 2 ( R 3 ) \overline{\varphi}\in\mathring{W}^{1,2}(\mathbb{R}^{3}) φ ∈ W ˚ 1 , 2 ( R 3 ) (the space defined in Subsection A.1) such that the following Euler-Lagrange equation holds:
− χ y ( Ω ) ∇ b Φ y ( ⋅ , b ‾ ) = b ‾ − b a μ 0 + ∇ φ ‾ a.e. in R 3 . (27) -\chi_{y(\Omega)}\nabla_{b}\Phi_{y}(\cdot,\overline{b})=\frac{\overline{b}-b_{a}}{\mu_{0}}+\nabla\overline{\varphi}\quad\text{a.e. in }\mathbb{R}^{3}. \tag*{(27)} − χ y ( Ω ) ∇ b Φ y ( ⋅ , b ) = μ 0 b − b a + ∇ φ a.e. in R 3 . (27)
For a derivation of these Euler-Lagrange equations see Appendix § 5 .§ 3 .
Remark 2.2. An equivalent way of writing the Euler-Lagrange equation (26) is
∇ m Ψ ^ y ( ⋅ , m ‾ ) = μ 0 ( m ‾ + h ‾ s ) + b a a.e. in y ( Ω ) , (28) \nabla_{m}\widehat{\Psi}_{y}(\cdot,\overline{m})=\mu_{0}\left(\overline{m}+\overline{h}_{s}\right)+b_{a}\quad\text{a.e. in }y(\Omega), \tag*{(28)} ∇ m Ψ y ( ⋅ , m ) = μ 0 ( m + h s ) + b a a.e. in y ( Ω ) , (28)
where Ψ ^ y ( ξ , m ) = Φ ^ y ( ξ , m ) + μ 0 2 ∣ m ∣ 2 \widehat{\Psi}_{y}(\xi,m)=\widehat{\Phi}_{y}(\xi,m)+\frac{\mu_{0}}{2}|m|^{2} Ψ y ( ξ , m ) = Φ y ( ξ , m ) + 2 μ 0 ∣ m ∣ 2 is the augmented material energy density defined in (16) . The two forms are equivalent. However, the form (26) is more convenient for the comparison with the b b b -based model, since the right-hand side is the magnetic induction b b b , this equation can be interpreted as a constitutive relation linking b b b to m m m .
Remark 2.3. Equation (27) is often written in the form
curl ( χ y ( Ω ) ∇ b Φ y ( ⋅ , b ‾ ) + b ‾ − b a μ 0 ) = 0. (29) \operatorname{curl}\left(\chi_{y(\Omega)}\nabla_{b}\Phi_{y}(\cdot,\overline{b})+\frac{\overline{b}-b_{a}}{\mu_{0}}\right)=0. \tag*{(29)} curl ( χ y ( Ω ) ∇ b Φ y ( ⋅ , b ) + μ 0 b − b a ) = 0. (29)
The two forms are equivalent. However, while (27) holds pointwise (almost everywhere), (29) holds only in the sense of distributions.
Remark 2.4. In (27) , the field ∇ φ ‾ \nabla\overline{\varphi} ∇ φ is a Lagrange multiplier associated with the divergence-free constraint on b b b . Such multiplier is not present in the Euler-Lagrange equation (26) of the m m m -h h h -based model, since the constraints (1) are automatically satisfied through the projection of χ y ( Ω ) \chi_{y(\Omega)} χ y ( Ω ) on the space of curl-free vector fields, as explained in Remark A.3.
Correspondence between constrained critical points
In this subsection we show that the two models are indeed equivalent at the level of constrained critical points, provided that the energy densities Φ ^ y \widehat{\Phi}_{y} Φ y and Φ y \Phi_{y} Φ y are related by the duality relations (H0). In particular, if ( m ‾ , h ‾ s ) (\overline{m},\overline{h}_{s}) ( m , h s ) is a constrained critical point of E ^ y \widehat{\mathcal{E}}_{y} E y , then the field b ‾ \overline{b} b corresponding to the physical definition
b ‾ : = b a + μ 0 ( χ y ( Ω ) m ‾ + h ‾ s ) (30) \overline{b}:=b_{a}+\mu_{0}\left(\chi_{y(\Omega)}\overline{m}+\overline{h}_{s}\right) \tag*{(30)} b := b a + μ 0 ( χ y ( Ω ) m + h s ) (30)
is a constrained critical point of E \mathcal{E} E . Going from the b b b -model to the m m m -h h h -model, is slightly more involved and requires instead to define ( m ‾ , h ‾ s ) (\overline{m},\overline{h}_{s}) ( m , h s ) in terms of b ‾ \overline{b} b by:
m ‾ : = − ∇ b Φ y ( ⋅ , b ‾ ) in y ( Ω ) , (31) \overline{m}:=-\nabla_{b}\Phi_{y}(\cdot,\overline{b})\quad\text{in }y(\Omega), \tag*{(31)} m := − ∇ b Φ y ( ⋅ , b ) in y ( Ω ) , (31)
h ‾ s : = − P [ χ y ( Ω ) m ‾ ] in R 3 . \overline{h}_{s}:=-\mathcal{P}\left[\chi_{y(\Omega)}\overline{m}\right]\quad\text{in }\mathbb{R}^{3}. h s := − P [ χ y ( Ω ) m ] in R 3 .
The first of (31) depends on the material, and can be interpreted as a constitutive relation between magnetic induction and magnetization. In the second equation P [ ⋅ ] \mathcal{P}[\cdot] P [ ⋅ ] is the orthogonal projection of L 2 ( R 3 ; R 3 ) L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) L 2 ( R 3 ; R 3 ) on the space of curl-free vector fields.
Remark 2.5. The Helmholtz decomposition of L 2 ( R 3 ; R 3 ) L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) L 2 ( R 3 ; R 3 ) based on P \mathcal{P} P (see Appendix § 5 .§ 1 ) reveals a partial link between (30) and (31) . It allows us to decompose any L 2 ( R 3 ; R 3 ) L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) L 2 ( R 3 ; R 3 ) vector field into the sum of a curl-free and a divergence-free vector field, in the present case reads
χ y ( Ω ) m ‾ = P [ χ y ( Ω ) m ‾ ] + ( I − P ) [ χ y ( Ω ) m ‾ ] . (32) \chi_{y(\Omega)}\overline{\mathbf{m}}=\mathcal{P}[\chi_{y(\Omega)}\overline{\mathbf{m}}]+(I-\mathcal{P})[\chi_{y(\Omega)}\overline{\mathbf{m}}]. \tag*{(32)} χ y ( Ω ) m = P [ χ y ( Ω ) m ] + ( I − P ) [ χ y ( Ω ) m ] . (32)
Assuming that (30) with b ‾ \overline{\mathbf{b}} b divergence-free and h ‾ s \overline{\mathbf{h}}_{\mathrm{s}} h s curl-free, we necessarily have (since b a \mathbf{b}_{\mathrm{a}} b a is divergence-free by assumption)
P [ χ y ( Ω ) m ‾ ] = − h ‾ s and ( I − P ) [ χ y ( Ω ) m ‾ ] = 1 μ 0 ( b ‾ − b a ) . (33) \mathcal{P}[\chi_{y(\Omega)}\overline{\mathbf{m}}]=-\overline{\mathbf{h}}_{\mathrm{s}}\quad\text{and}\quad(I-\mathcal{P})[\chi_{y(\Omega)}\overline{\mathbf{m}}]=\frac{1}{\mu_{0}}(\overline{\mathbf{b}}-\mathbf{b}_{\mathrm{a}}). \tag*{(33)} P [ χ y ( Ω ) m ] = − h s and ( I − P ) [ χ y ( Ω ) m ] = μ 0 1 ( b − b a ) . (33)
The precise statement of the correspondence between critical points is in Theorem 2.6 below, which we prove under the following assumptions. In particular, they link the material energy densities Φ ^ y \widehat{\Phi}_{\mathbf{y}} Φ y and Φ y \Phi_{\mathbf{y}} Φ y by suitable duality relations, where the symbol ⋄ \diamond ⋄ denotes the generalized Legendre-Fenchel transform defined in (A.7). If we are coming from the ( m , h s ) (\mathbf{m},\mathbf{h}_{\mathrm{s}}) ( m , h s ) -based formulation with given Φ ^ y \widehat{\Phi}_{\mathbf{y}} Φ y , define the corresponding density function as
Φ y ( ξ , b ) = − Ψ ^ y ⋄ ( ξ , b ) , where Ψ ^ y ( ξ , m ) = Φ ^ y ( ξ , m ) + μ 0 2 ∣ m ∣ 2 (cf. (1.16)) , (34) \Phi_{\mathbf{y}}(\xi,\mathbf{b})=-\widehat{\Psi}_{\mathbf{y}}^{\diamond}(\xi,\mathbf{b}),\quad\text{where}\quad\widehat{\Psi}_{\mathbf{y}}(\xi,\mathbf{m})=\widehat{\Phi}_{\mathbf{y}}(\xi,\mathbf{m})+\frac{\mu_{0}}{2}|\mathbf{m}|^{2}\quad\text{(cf. (1.16))}, \tag*{(34)} Φ y ( ξ , b ) = − Ψ y ⋄ ( ξ , b ) , where Ψ y ( ξ , m ) = Φ y ( ξ , m ) + 2 μ 0 ∣ m ∣ 2 (cf. (1.16)) , (34)
and require that
( ξ , m ) ↦ Φ ^ y ( ξ , m ) , Φ ^ y : y ( Ω ) × R 3 → R ∪ { + ∞ } , (\xi,\mathbf{m})\mapsto\widehat{\Phi}_{\mathbf{y}}(\xi,\mathbf{m}),\quad\widehat{\Phi}_{\mathbf{y}}:y(\Omega)\times\mathbb{R}^{3}\to\mathbb{R}\cup\{+\infty\}, ( ξ , m ) ↦ Φ y ( ξ , m ) , Φ y : y ( Ω ) × R 3 → R ∪ { + ∞ } ,
is measurable in ξ \xi ξ for each m \mathbf{m} m and lower semi-continuous in m \mathbf{m} m for a.e. ξ \xi ξ ;
− C ∣ m ∣ 2 − C ≤ Φ ^ y ( ξ , m ) for all m and a.e. ξ , with a constant C > 0 ; -C|\mathbf{m}|^{2}-C\leq\widehat{\Phi}_{\mathbf{y}}(\xi,\mathbf{m})\quad\text{for all }\mathbf{m}\text{ and a.e. }\xi,\quad\text{with a constant }C>0; − C ∣ m ∣ 2 − C ≤ Φ y ( ξ , m ) for all m and a.e. ξ , with a constant C > 0 ;
for a.e. ξ ∈ y ( Ω ) , Φ ^ y ( ξ , ⋅ ) is differentiable and its gradient m ↦ ∇ m Φ ^ y ( ξ , m ) , R 3 → R 3 , \text{for a.e. }\xi\in y(\Omega),\ \widehat{\Phi}_{\mathbf{y}}(\xi,\cdot)\text{ is differentiable and its gradient }\mathbf{m}\mapsto\nabla_{\mathbf{m}}\widehat{\Phi}_{\mathbf{y}}(\xi,\mathbf{m}),\quad\mathbb{R}^{3}\to\mathbb{R}^{3}, for a.e. ξ ∈ y ( Ω ) , Φ y ( ξ , ⋅ ) is differentiable and its gradient m ↦ ∇ m Φ y ( ξ , m ) , R 3 → R 3 ,
is a continuous and continuously invertible function.
If we are the given b \mathbf{b} b -based formulation featuring Φ y \Phi_{\mathbf{y}} Φ y instead, we define
Φ ^ y ( ξ , m ) = ( − Φ y ) ⋄ ( ξ , m ) − μ 0 2 ∣ m ∣ 2 (35) \widehat{\Phi}_{\mathbf{y}}(\xi,\mathbf{m})=(-\Phi_{\mathbf{y}})^{\diamond}(\xi,\mathbf{m})-\frac{\mu_{0}}{2}|\mathbf{m}|^{2} \tag*{(35)} Φ y ( ξ , m ) = ( − Φ y ) ⋄ ( ξ , m ) − 2 μ 0 ∣ m ∣ 2 (35)
and assume that
( ξ , b ) ↦ Φ y ( ξ , b ) , y ( Ω ) × R 3 → R ∪ { + ∞ } , (\xi,\mathbf{b})\mapsto\Phi_{\mathbf{y}}(\xi,\mathbf{b}),\quad y(\Omega)\times\mathbb{R}^{3}\to\mathbb{R}\cup\{+\infty\}, ( ξ , b ) ↦ Φ y ( ξ , b ) , y ( Ω ) × R 3 → R ∪ { + ∞ } ,
is measurable in ξ \xi ξ for each b \mathbf{b} b and lower semi-continuous in b \mathbf{b} b for a.e. ξ \xi ξ ;
− C ∣ b ∣ 2 − C ≤ Φ y ( ξ , b ) for all b and a.e. ξ , with a constant C > 0 ; -C|\mathbf{b}|^{2}-C\leq\Phi_{\mathbf{y}}(\xi,\mathbf{b})\quad\text{for all }\mathbf{b}\text{ and a.e. }\xi,\quad\text{with a constant }C>0; − C ∣ b ∣ 2 − C ≤ Φ y ( ξ , b ) for all b and a.e. ξ , with a constant C > 0 ;
for a.e. ξ ∈ y ( Ω ) , Φ y ( ξ , ⋅ ) is differentiable and its gradient b ↦ ∇ b Φ y ( ξ , b ) , R 3 → R 3 , \text{for a.e. }\xi\in y(\Omega),\ \Phi_{\mathbf{y}}(\xi,\cdot)\text{ is differentiable and its gradient }\mathbf{b}\mapsto\nabla_{\mathbf{b}}\Phi_{\mathbf{y}}(\xi,\mathbf{b}),\quad\mathbb{R}^{3}\to\mathbb{R}^{3}, for a.e. ξ ∈ y ( Ω ) , Φ y ( ξ , ⋅ ) is differentiable and its gradient b ↦ ∇ b Φ y ( ξ , b ) , R 3 → R 3 ,
is a continuous and continuously invertible function.
Here and throughout, (H1)–(H2) and (H1′)–(H2′) are imposed only to ensure that the integrals in E ^ y \widehat{E}_{\mathbf{y}} E y and E y E_{\mathbf{y}} E y are well defined. As stated, these conditions readily admit convex examples; alternatively, one may impose them for − Φ ^ y -\widehat{\Phi}_{\mathbf{y}} − Φ y or − Φ y -\Phi_{\mathbf{y}} − Φ y when a strongly concave example is desired.
We emphasize that, under (H3) or (H3′), the material energy densities appearing in (H0) and (H0′) can be related using the (smooth) Legendre–Fenchel transform. Since this transform is an involution (see Lemma A.6 ), the definitions (H0) and (H0′) are mutually consistent. This equivalence does not require convexity; it only requires invertibility of the gradient.
Theorem 2.6 (Model equivalence at equilibrium: smooth case). The following holds:
Assume that (H0) and (H3) are satisfied. If ( m ‾ , h ‾ s ) ∈ L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_{\mathrm{s}})\in L^{2}(y(\Omega);\mathbb{R}^{3})\times L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) ( m , h s ) ∈ L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) is a constrained critical point of E ^ \widehat{\mathcal{E}} E in the sense of Definition 2.1, then the field b ‾ \overline{\mathbf{b}} b defined by (30) is a constrained critical point of E \mathcal{E} E .
(ii) Assume that (H0′) and (H3′) are satisfied. If b ˉ ∈ L 2 ( R 3 ; R 3 ) \bar{b} \in L^2(\mathbb{R}^3;\mathbb{R}^3) b ˉ ∈ L 2 ( R 3 ; R 3 ) is a constrained critical point of E \mathcal{E} E in the sense of Definition 2.1, then the pair ( m ˉ , h ˉ s ) (\bar{m},\bar{h}_s) ( m ˉ , h ˉ s ) defined by (31) belongs to L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) L^2(y(\Omega);\mathbb{R}^3) \times L^2(\mathbb{R}^3;\mathbb{R}^3) L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) and is a constrained critical point of E y \mathcal{E}_y E y . Moreover, it satisfies
b a + μ 0 ( χ y ( Ω ) m ˉ + h ˉ s ) = b ˉ a.e. in R 3 . (36) b_a+\mu_0\left(\chi_{y(\Omega)}\bar{m}+\bar{h}_s\right)=\bar{b}\quad\text{a.e. in }\mathbb{R}^3. \tag*{(36)} b a + μ 0 ( χ y ( Ω ) m ˉ + h ˉ s ) = b ˉ a.e. in R 3 . (36)
Remark 2.7. Under our assumptions, the Euler-Lagrange equations (26) and (27) of Definition 2.1 are well-defined, but (H3) and (H3′) alone are not strong enough to justify them as necessary conditions for local extremals via Proposition A.10. For the latter, we also need (H1) and (H1′) as well as stronger versions of (H2) and (H2′), namely, that ∇ m Ψ ^ y \nabla_m\widehat{\Psi}_y ∇ m Ψ y , ∇ b Φ y \nabla_b\Phi_y ∇ b Φ y and their inverse functions satisfy a linear growth condition.
Proof of Theorem 2.6. Proof of (i). In view of the definition of b ˉ \bar{b} b ˉ in (30) , b ˉ ∈ L 2 ( R 3 ; R 3 ) \bar{b} \in L^2(\mathbb{R}^3;\mathbb{R}^3) b ˉ ∈ L 2 ( R 3 ; R 3 ) . Moreover, the first Maxwell constraint div ( χ y ( Ω ) m ˉ + h ˉ s ) = 0 \operatorname{div}(\chi_{y(\Omega)}\bar{m}+\bar{h}_s)=0 div ( χ y ( Ω ) m ˉ + h ˉ s ) = 0 and the assumption that b a b_a b a is divergence-free yield immediately that b ˉ \bar{b} b ˉ is divergence-free:
div b ˉ = 0 in R 3 . (37) \operatorname{div}\bar{b}=0\quad\text{in }\mathbb{R}^3. \tag*{(37)} div b ˉ = 0 in R 3 . (37)
By the second Maxwell constraint, curl h ˉ s = 0 \operatorname{curl}\bar{h}_s=0 curl h ˉ s = 0 , there exists a scalar potential φ ˉ ∈ W ˚ 1 , 2 ( R 3 ) \bar{\varphi}\in\mathring{W}^{1,2}(\mathbb{R}^3) φ ˉ ∈ W ˚ 1 , 2 ( R 3 ) such that
h ˉ s = − ∇ φ ˉ . (38) \bar{h}_s=-\nabla\bar{\varphi}. \tag*{(38)} h ˉ s = − ∇ φ ˉ . (38)
Outside y ( Ω ) y(\Omega) y ( Ω ) , by definition (30) , m ˉ = 0 \bar{m}=0 m ˉ = 0 and b ˉ = μ 0 h ˉ s + b a \bar{b}=\mu_0\bar{h}_s+b_a b ˉ = μ 0 h ˉ s + b a . Hence,
b ˉ − b a μ 0 + ∇ φ ˉ = 0 a.e. in R 3 ∖ y ( Ω ) . (39) \frac{\bar{b}-b_a}{\mu_0}+\nabla\bar{\varphi}=0\quad\text{a.e. in }\mathbb{R}^3\setminus y(\Omega). \tag*{(39)} μ 0 b ˉ − b a + ∇ φ ˉ = 0 a.e. in R 3 ∖ y ( Ω ) . (39)
Thus the Euler-Lagrange equation (27) for the b-model holds in R 3 ∖ y ( Ω ) \mathbb{R}^3\setminus y(\Omega) R 3 ∖ y ( Ω ) , and it remains to show that it also holds in y ( Ω ) y(\Omega) y ( Ω ) . This is where we use the assumption that ( m ˉ , h ˉ s ) (\bar{m},\bar{h}_s) ( m ˉ , h ˉ s ) satisfies the Euler-Lagrange equation (26) . Indeed, substituting the definition of b ˉ \bar{b} b ˉ in (30) , we obtain
b ˉ = ∇ m Ψ ^ y ( ⋅ , m ˉ ) a.e. in y ( Ω ) . (40) \bar{b}=\nabla_m\widehat{\Psi}_y(\cdot,\bar{m})\quad\text{a.e. in }y(\Omega). \tag*{(40)} b ˉ = ∇ m Ψ y ( ⋅ , m ˉ ) a.e. in y ( Ω ) . (40)
By the duality properties of the Legendre transform (see Lemma A.6), equation (40) is equivalent to
m ˉ = − ∇ b Φ y ( ⋅ , b ˉ ) a.e. in y ( Ω ) . (41) \bar{m}=-\nabla_b\Phi_y(\cdot,\bar{b})\quad\text{a.e. in }y(\Omega). \tag*{(41)} m ˉ = − ∇ b Φ y ( ⋅ , b ˉ ) a.e. in y ( Ω ) . (41)
Substituting this into the definition of b ˉ \bar{b} b ˉ in R 3 \mathbb{R}^3 R 3 , and recalling that h ˉ s = − ∇ φ ˉ \bar{h}_s=-\nabla\bar{\varphi} h ˉ s = − ∇ φ ˉ , we have
− χ y ( Ω ) ∇ b Φ y ( ⋅ , b ˉ ) = b ˉ − b a μ 0 + ∇ φ ˉ a.e. in R 3 . (42) -\chi_{y(\Omega)}\nabla_b\Phi_y(\cdot,\bar{b})=\frac{\bar{b}-b_a}{\mu_0}+\nabla\bar{\varphi}\quad\text{a.e. in }\mathbb{R}^3. \tag*{(42)} − χ y ( Ω ) ∇ b Φ y ( ⋅ , b ˉ ) = μ 0 b ˉ − b a + ∇ φ ˉ a.e. in R 3 . (42)
The combination of (39) and (42) yields the Euler-Lagrange equation (27) for b ˉ \bar{b} b ˉ .
Proof of (ii). Conversely, let b ˉ \bar{b} b ˉ be a smooth constrained critical point of E y \mathcal{E}_y E y . By hypothesis, b ˉ \bar{b} b ˉ satisfies the Euler-Lagrange equation (27) . By the definition of m ˉ \bar{m} m ˉ in (31) , we obtain that
b ˉ − b a μ 0 + ∇ φ ˉ = χ y ( Ω ) m ˉ a.e. in R 3 . (43) \frac{\bar{b}-b_a}{\mu_0}+\nabla\bar{\varphi}=\chi_{y(\Omega)}\bar{m}\quad\text{a.e. in }\mathbb{R}^3. \tag*{(43)} μ 0 b ˉ − b a + ∇ φ ˉ = χ y ( Ω ) m ˉ a.e. in R 3 . (43)
(Note that we do not know yet whether φ ˉ \bar{\varphi} φ ˉ is the same as in the proof of (i).) In particular, we have that m ˉ ∈ L 2 \bar{m}\in L^2 m ˉ ∈ L 2 , as required by the formulation:
m ˉ ∈ L 2 ( y ( Ω ) ; R 3 ) . (44) \bar{m}\in L^2(y(\Omega);\mathbb{R}^3). \tag*{(44)} m ˉ ∈ L 2 ( y ( Ω ) ; R 3 ) . (44)
Applying the projection operator P \mathcal{P} P to both sides of (43) , using the orthogonality between curl-free and divergence-free fields, and the assumption that both b ˉ \bar{b} b ˉ and b a b_a b a are divergence-free, we obtain ∇ φ ˉ = P [ χ y ( Ω ) m ˉ ] \nabla\bar{\varphi}=\mathcal{P}[\chi_{y(\Omega)}\bar{m}] ∇ φ ˉ = P [ χ y ( Ω ) m ˉ ] , and hence the definition of h ˉ s \bar{h}_s h ˉ s in (31) yields
h ˉ s = − ∇ φ ˉ . (45) \bar{h}_s=-\nabla\bar{\varphi}. \tag*{(45)} h ˉ s = − ∇ φ ˉ . (45)
Substituting (2.23) into (2.21), we obtain (2.14). Moreover, since both b ‾ \overline{b} b and b a b_a b a are divergence-free, we have div ( χ y ( Ω ) m ‾ + h ‾ s ) = 0 \operatorname{div}(\chi_{y(\Omega)}\overline{m}+\overline{h}_s)=0 div ( χ y ( Ω ) m + h s ) = 0 . Thus, the pair ( m ‾ , h ‾ s ) (\overline{m},\overline{h}_s) ( m , h s ) satisfies the Maxwell constraints.
Finally, in view of (H0) and the duality properties of the Legendre transform (see Lemma A.6 ), the definition (2.11) of m ‾ \overline{m} m is equivalent to
b ‾ = ∇ m Ψ ^ y ( ⋅ , m ‾ ) , a.e. in y ( Ω ) . (46) \overline{b}=\nabla_m\widehat{\Psi}_y(\cdot,\overline{m}),\qquad\text{a.e. in }y(\Omega). \tag*{(46)} b = ∇ m Ψ y ( ⋅ , m ) , a.e. in y ( Ω ) . (46)
Substituting (2.23) and (2.24) into (2.21) (evaluated only in y ( Ω ) y(\Omega) y ( Ω ) ), we recover the Euler-Lagrange equation (2.6) for ( m ‾ , h ‾ s ) (\overline{m},\overline{h}_s) ( m , h s ) .
Remark 2.8. The proof of (ii) of Theorem 2.6 , shows that if b ‾ \overline{b} b is a constrained critical point then the Lagrange multiplier − ∇ φ -\nabla\varphi − ∇ φ in the Euler-Lagrange equation (2.7) coincides with the stray field h ‾ s \overline{h}_s h s that corresponds to b ‾ \overline{b} b through (2.11).
Remark 2.9. In the statement of Theorem 2.6 (ii), the definition of h ‾ s \overline{h}_s h s in (2.11) is not the only possible one. In particular, the second of (2.11) may be replaced by
h ‾ s : = − χ y ( Ω ) m ‾ + 1 μ 0 ( b ‾ − b a ) . \overline{h}_s:=-\chi_{y(\Omega)}\overline{m}+\frac{1}{\mu_0}(\overline{b}-b_a). h s := − χ y ( Ω ) m + μ 0 1 ( b − b a ) .
This definition, along with the assumption that b ‾ \overline{b} b and b a b_a b a are divergence-free, automatically guarantees (2.14), and div ( χ y ( Ω ) m ‾ + h ‾ s ) = 0 \operatorname{div}(\chi_{y(\Omega)}\overline{m}+\overline{h}_s)=0 div ( χ y ( Ω ) m + h s ) = 0 , but it does not guarantee the second Maxwell constraint, namely, that h ‾ s \overline{h}_s h s is curl-free, unless we assume that b ‾ \overline{b} b satisfies the Euler-Lagrange equation (2.7). In contrast, the definition of h ‾ s \overline{h}_s h s in (2.11) immediately implies that h ‾ s \overline{h}_s h s is curl-free as the projection of − χ y ( Ω ) m ‾ -\chi_{y(\Omega)}\overline{m} − χ y ( Ω ) m on the space of curl-free vector fields. However, it does not guarantee the other Maxwell constraint div ( χ y ( Ω ) m ‾ + h ‾ s ) = 0 \operatorname{div}(\chi_{y(\Omega)}\overline{m}+\overline{h}_s)=0 div ( χ y ( Ω ) m + h s ) = 0 unless we assume that b ‾ \overline{b} b satisfies the Euler-Lagrange equation (2.7).
Critical states and energy equality
We have seen in the previous subsection that to each constrained critical point of E ^ y \widehat{\mathcal{E}}_y E y corresponds a constrained critical point of E y \mathcal{E}_y E y , and vice versa. Most importantly, the correspondence between critical points is one to one (at least in the smooth case treated in this section). This correspondence allows us to define the notion of critical state, which is a triplet ( m ‾ , h ‾ s , b ‾ ) (\overline{m},\overline{h}_s,\overline{b}) ( m , h s , b ) that satisfies the Maxwell equations and the Euler-Lagrange equations of both models simultaneously. The precise definition is as follows.
Definition 2.10 (Critical states). Under the same assumptions as in Definition 2.1 and either (i) or (ii) in Theorem 2.6, a triplet ( m ‾ , h ‾ s , b ‾ ) ∈ L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) × L 2 ( R 3 ; R 3 ) (\overline{m},\overline{h}_s,\overline{b})\in L^2(y(\Omega);\mathbb{R}^3)\times L^2(\mathbb{R}^3;\mathbb{R}^3)\times L^2(\mathbb{R}^3;\mathbb{R}^3) ( m , h s , b ) ∈ L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) × L 2 ( R 3 ; R 3 ) is a critical state if
curl h ‾ s = 0 , div b ‾ = 0 , b ‾ = b a + μ 0 ( χ y ( Ω ) m ‾ + h ‾ s ) , (47) \operatorname{curl}\overline{h}_s=0,\quad\operatorname{div}\overline{b}=0,\quad\overline{b}=b_a+\mu_0\left(\chi_{y(\Omega)}\overline{m}+\overline{h}_s\right), \tag*{(47)} curl h s = 0 , div b = 0 , b = b a + μ 0 ( χ y ( Ω ) m + h s ) , (47)
and
∇ m Ψ ^ y ( ⋅ , m ‾ ) = b ‾ a.e. in y ( Ω ) , (48) \nabla_m\widehat{\Psi}_y(\cdot,\overline{m})=\overline{b}\qquad\text{a.e. in }y(\Omega), \tag*{(48)} ∇ m Ψ y ( ⋅ , m ) = b a.e. in y ( Ω ) , (48)
or, equivalently,
− ∇ b Φ ^ y ( ⋅ , b ‾ ) = m ‾ a.e. in y ( Ω ) . (49) -\nabla_b\widehat{\Phi}_y(\cdot,\overline{b})=\overline{m}\qquad\text{a.e. in }y(\Omega). \tag*{(49)} − ∇ b Φ y ( ⋅ , b ) = m a.e. in y ( Ω ) . (49)
Here, recall that Ψ ^ y ( ξ , m ) = Φ ^ y ( ξ , m ) + μ 0 2 ∣ m ∣ 2 \widehat{\Psi}_y(\xi,m)=\widehat{\Phi}_y(\xi,m)+\frac{\mu_0}{2}|m|^2 Ψ y ( ξ , m ) = Φ y ( ξ , m ) + 2 μ 0 ∣ m ∣ 2 is the augmented density introduced in (1.16).
Remark 2.11. As the smooth Legendre-Fenchel transformation is an involution by Lemma A.6 , the duality relations (H0) and (H0′ ' ′ ) are equivalent. Using either one and (A.11), we see that (2.26) and (2.27) are indeed equivalent, too.
Remark 2.12. The third of (2.25), i.e., the physical induction relation b ˉ = b a + μ 0 ( χ y ( Ω ) m ˉ + h ˉ s ) \bar{b} = b_a+\mu_0(\chi_{y(\Omega)}\bar{m}+\bar{h}_s) b ˉ = b a + μ 0 ( χ y ( Ω ) m ˉ + h ˉ s ) , holds in both directions of Theorem 2.6: in direction (i) it is the very definition (2.10) of b ˉ \bar{b} b ˉ from ( m ˉ , h ˉ s ) (\bar{m},\bar{h}_s) ( m ˉ , h ˉ s ) ; in direction (ii) it is the conclusion (2.14) of the proof once ( m ˉ , h ˉ s ) (\bar{m},\bar{h}_s) ( m ˉ , h ˉ s ) is defined via (2.11).
The duality condition (2.26), i.e., ∇ m Ψ ^ y ( ⋅ , m ˉ ) = b ˉ \nabla_m\widehat{\Psi}_y(\cdot,\bar{m})=\bar{b} ∇ m Ψ y ( ⋅ , m ˉ ) = b ˉ , is intrinsically linked to equilibrium: combined with the third of (2.25), it is equivalent to the Euler-Lagrange equation (2.6) for ( m ˉ , h ˉ s ) (\bar{m},\bar{h}_s) ( m ˉ , h ˉ s ) , while the first two of (2.25) are just the Maxwell constraint.
In direction (ii) of Theorem 2.6, the starting assumption is that the second of (2.25) holds, and that the Euler-Lagrange equation (2.7) holds. The proof then shows that the equivalent condition (2.27), i.e., − ∇ b Φ y ( ⋅ , b ˉ ) = m ˉ -\nabla_b\Phi_y(\cdot,\bar{b})=\bar{m} − ∇ b Φ y ( ⋅ , b ˉ ) = m ˉ , serves as the constitutive relation used to define m ˉ \bar{m} m ˉ from b ˉ \bar{b} b ˉ via (2.11). The proof then shows that the pair ( m ˉ , h ˉ s ) (\bar{m},\bar{h}_s) ( m ˉ , h ˉ s ) so defined automatically satisfies the Maxwell constraints and (2.6), which is exactly (2.27).
Given a critical state ( m ˉ , h ˉ s , b ˉ ) (\bar{m},\bar{h}_s,\bar{b}) ( m ˉ , h ˉ s , b ˉ ) , a natural question is whether or not the equality E y ( b ˉ ) = E ^ y ( m ˉ , h ˉ s ) \mathcal{E}_y(\bar{b})=\widehat{\mathcal{E}}_y(\bar{m},\bar{h}_s) E y ( b ˉ ) = E y ( m ˉ , h ˉ s ) holds, i.e., the energy values of the two functionals correspond. The answer to this question is positive. In particular, it hinges on the following algebraic equality which is a consequence of the duality of the Legendre-Fenchel transform.
Proposition 2.13. Assume that m ˉ ∈ L 2 ( y ( Ω ) ; R 3 ) \bar{m}\in L^2(y(\Omega);\mathbb{R}^3) m ˉ ∈ L 2 ( y ( Ω ) ; R 3 ) and b ˉ ∈ L 2 ( R 3 ; R 3 ) \bar{b}\in L^2(\mathbb{R}^3;\mathbb{R}^3) b ˉ ∈ L 2 ( R 3 ; R 3 ) are related by the relation (2.26). Then
Ψ ^ ( ⋅ , m ˉ ) + Ψ ^ ⋄ ( ⋅ , b ˉ ) = m ˉ ⋅ b ˉ on y ( Ω ) . (50) \widehat{\Psi}(\cdot,\bar{m})+\widehat{\Psi}^{\diamond}(\cdot,\bar{b})=\bar{m}\cdot\bar{b}\quad\text{on }y(\Omega). \tag*{(50)} Ψ ( ⋅ , m ˉ ) + Ψ ⋄ ( ⋅ , b ˉ ) = m ˉ ⋅ b ˉ on y ( Ω ) . (50)
Proof. This follows from Lemma A.6.
The next proposition explains why (2.28) is a key equality to obtain the equivalence of the energies at critical points.
Proposition 2.14 (Energy equality). Assume that either (H1)-(H2) or (H1′)-(H2′) hold. Moreover, let ( m ˉ , h ˉ s ) ∈ L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) (\bar{m},\bar{h}_s)\in L^2(y(\Omega);\mathbb{R}^3)\times L^2(\mathbb{R}^3;\mathbb{R}^3) ( m ˉ , h ˉ s ) ∈ L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) and b ˉ ∈ L 2 ( R 3 ; R 3 ) \bar{b}\in L^2(\mathbb{R}^3;\mathbb{R}^3) b ˉ ∈ L 2 ( R 3 ; R 3 ) be such that (2.28) holds and that
b ˉ = μ 0 ( χ y ( Ω ) m ˉ + h ˉ s ) + b a on R 3 . (51) \bar{b}=\mu_0(\chi_{y(\Omega)}\bar{m}+\bar{h}_s)+b_a\quad\text{on }\mathbb{R}^3. \tag*{(51)} b ˉ = μ 0 ( χ y ( Ω ) m ˉ + h ˉ s ) + b a on R 3 . (51) Then
E y ( b ˉ ) = E ^ y ( m ˉ , h ˉ s ) . (52) \mathcal{E}_y(\bar{b})=\widehat{\mathcal{E}}_y(\bar{m},\bar{h}_s). \tag*{(52)} E y ( b ˉ ) = E y ( m ˉ , h ˉ s ) . (52)
Corollary 2.15 (Correspondence between energy values at constrained critical points). Assume that either (H0)–(H3) or (H0′)–(H3′) are satisfied. If ( m ˉ , h ˉ s , b ˉ ) (\bar{m},\bar{h}_s,\bar{b}) ( m ˉ , h ˉ s , b ˉ ) is a critical state, then the following energy equality holds:
E ^ y ( m ˉ , h ˉ s ) = E y ( b ˉ ) . (53) \widehat{\mathcal{E}}_y(\bar{m},\bar{h}_s)=\mathcal{E}_y(\bar{b}). \tag*{(53)} E y ( m ˉ , h ˉ s ) = E y ( b ˉ ) . (53)
Proof of Proposition 2.14. From (2.29) it follows that
μ 0 2 ∣ h ˉ s ∣ 2 = 1 2 μ 0 ∣ b ˉ − b a ∣ 2 + μ 0 2 χ y ( Ω ) ∣ m ˉ ∣ 2 − ( b ˉ − b a ) ⋅ m ˉ . (54) \frac{\mu_0}{2}|\bar{h}_s|^2=\frac{1}{2\mu_0}|\bar{b}-b_a|^2+\frac{\mu_0}{2}\chi_{y(\Omega)}|\bar{m}|^2-(\bar{b}-b_a)\cdot\bar{m}. \tag*{(54)} 2 μ 0 ∣ h ˉ s ∣ 2 = 2 μ 0 1 ∣ b ˉ − b a ∣ 2 + 2 μ 0 χ y ( Ω ) ∣ m ˉ ∣ 2 − ( b ˉ − b a ) ⋅ m ˉ . (54)
This, along with (2.28) and the definitions of E ^ y ( m ˉ , h ˉ s ) \widehat{\mathcal{E}}_y(\bar{m},\bar{h}_s) E y ( m ˉ , h ˉ s ) and E y ( b ˉ ) \mathcal{E}_y(\bar{b}) E y ( b ˉ ) , gives the following chain of equalities:
E ^ y ( m ˉ , h ˉ s ) = ∫ y ( Ω ) ( Ψ ^ y ( ξ , m ˉ ) − m ˉ ⋅ b ˉ ) d ξ + ∫ R 3 1 2 μ 0 ∣ b ˉ − b a ∣ 2 d ξ = − ∫ y ( Ω ) Ψ ^ y ⋄ ( ξ , b ˉ ) d ξ + ∫ R 3 1 2 μ 0 ∣ b ˉ − b a ∣ 2 d ξ = E y ( b ˉ ) . \begin{aligned}
\widehat{\mathcal{E}}_y(\bar{m},\bar{h}_s)
&=\int_{y(\Omega)}\left(\widehat{\Psi}_y(\xi,\bar{m})-\bar{m}\cdot\bar{b}\right)\,d\xi+\int_{\mathbb{R}^3}\frac{1}{2\mu_0}|\bar{b}-b_a|^2\,d\xi\\
&=-\int_{y(\Omega)}\widehat{\Psi}^{\diamond}_y(\xi,\bar{b})\,d\xi+\int_{\mathbb{R}^3}\frac{1}{2\mu_0}|\bar{b}-b_a|^2\,d\xi\\
&=\mathcal{E}_y(\bar{b}).
\end{aligned} E y ( m ˉ , h ˉ s ) = ∫ y ( Ω ) ( Ψ y ( ξ , m ˉ ) − m ˉ ⋅ b ˉ ) d ξ + ∫ R 3 2 μ 0 1 ∣ b ˉ − b a ∣ 2 d ξ = − ∫ y ( Ω ) Ψ y ⋄ ( ξ , b ˉ ) d ξ + ∫ R 3 2 μ 0 1 ∣ b ˉ − b a ∣ 2 d ξ = E y ( b ˉ ) .
Proof of Corollary 2.15. In view of Proposition 2.14 , we only have to check (50) and (51) . The former is due to Proposition 2.13 , while the latter is given by (47) .
Remark 2.16. To verify (52) in the proof of Proposition 2.14 , we have used not only (51) and the duality property (50) (which holds by convex or concave duality in the convex/concave case, and by standard properties of the Legendre-Fenchel transform in the smooth case), but also the remarkable coincidence that their combination completely eliminates both b \mathbf{b} b and m ‾ \overline{\mathbf{m}} m in the difference E y ( b ‾ ) − E ^ y ( m ‾ , h ‾ s ) \mathcal{E}_y(\overline{\mathbf{b}})-\widehat{\mathcal{E}}_y(\overline{\mathbf{m}},\overline{\mathbf{h}}_s) E y ( b ) − E y ( m , h s ) . However, strictly speaking, energy equality does not require that the triplet ( m ‾ , h ‾ s , b ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_s,\mathbf{b}) ( m , h s , b ) is a critical state in the sense of Definition 2.10.
Remark 2.17. If the functions Ψ ^ y ( ξ , ⋅ ) \widehat{\Psi}_y(\xi,\cdot) Ψ y ( ξ , ⋅ ) and Φ y ( ξ , ⋅ ) \Phi_y(\xi,\cdot) Φ y ( ξ , ⋅ ) are convex or concave, respectively, the duality relation (50) is not only a consequence of (49) , but it in fact an equivalent condition. Thus, in view of Remark 2.11, (49) and (51) imply that the pair ( m ‾ , h ‾ s ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_s) ( m , h s ) satisfies the Euler-Lagrange equation of the functional E ^ y \widehat{\mathcal{E}}_y E y . This however still does not guarantee that the pair ( m ‾ , h ‾ s ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_s) ( m , h s ) satisfies the Maxwell constraints.
Remark 2.18. We emphasize that if ( m , h s ) (\mathbf{m},\mathbf{h}_s) ( m , h s ) is a generic pair satisfying the Maxwell constraint, the field b = μ 0 ( χ y ( Ω ) m + h s ) + b a \mathbf{b}=\mu_0(\chi_y(\Omega)\mathbf{m}+\mathbf{h}_s)+\mathbf{b}_a b = μ 0 ( χ y ( Ω ) m + h s ) + b a is divergence free, but the energies E ^ y ( m , h s ) \widehat{\mathcal{E}}_y(\mathbf{m},\mathbf{h}_s) E y ( m , h s ) and E y ( b ) \mathcal{E}_y(\mathbf{b}) E y ( b ) do not necessarily coincide for generic triplets ( m , h s , b ) (\mathbf{m},\mathbf{h}_s,\mathbf{b}) ( m , h s , b ) satisfying (47) , as can be easily seen in the examples of § 4 . For those, the energy equality is a direct consequence of either (48) or (49) which in turn are linked to equilibrium (see Remark 2.12).
Remark 2.19. The proof of Proposition 2.14 is purely algebraic: it uses only the Fenchel identity (50) and the physical induction relation (51) , neither of which requires differentiability of the energy densities. The same argument therefore applies verbatim in the convex/concave setting of § 3 (see Proposition 3.9 ).
The convex/concave case
In this section we extend the equivalence results of Section 2 to the nonsmooth setting. The key difference is that differentiability of the energy densities is no longer assumed. Instead, we require that Ψ ^ y ( ξ , ⋅ ) \widehat{\Psi}_y(\xi,\cdot) Ψ y ( ξ , ⋅ ) or Φ y ( ξ , ⋅ ) \Phi_y(\xi,\cdot) Φ y ( ξ , ⋅ ) be either convex or concave, and we replace gradients by convex subdifferentials (or concave superdifferentials). The Fenchel duality machinery, which in the smooth case relied on standard properties of the Legendre-Fenchel transform, is now applied in its classical convex-analytic form (see, e.g., [18 ] ).
As in Section 2, the deformation y : Ω → R 3 \mathbf{y}:\Omega\to\mathbb{R}^3 y : Ω → R 3 is regarded as a fixed parameter. Throughout, we again assume that y ( Ω ) \mathbf{y}(\Omega) y ( Ω ) is measurable and that b a ∈ L 2 ( R 3 ; R 3 ) \mathbf{b}_a\in L^2(\mathbb{R}^3;\mathbb{R}^3) b a ∈ L 2 ( R 3 ; R 3 ) with div b a = 0 \operatorname{div}\mathbf{b}_a=0 div b a = 0 on R 3 \mathbb{R}^3 R 3 .
We next define the notion of constrained critical point appropriate to this nonsmooth setting. In the convex (or concave) case, critical points are characterized by subdifferential (or superdifferential) inclusions that generalize the Euler-Lagrange equations of Definition 2.1 .
Remark 3.1. To treat using a single notation both the convex and the concave case, we use the symbol ∂ \partial ∂ to denote either the convex subdifferential (if it is non-empty) or the concave superdifferential (otherwise), for both functions and functionals. In the convex case, the subdifferential inclusion (55) characterizes global minimizers, while in the concave case it characterizes global maximizers. More details are given in the Appendix.
Remark 3.2. We note that in more general nonsmooth settings beyond the convex/concave framework, defining meaningful notions of stationarity becomes significantly more delicate [3 , 7 , 8 ] . Many generalized notions of differentiability formally give rise to necessary conditions for local extrema, but these can be substantially weaker than their smooth counterparts and may fail to characterize stationarity in dynamical context (for an associated gradient flow, e.g.). For convex (or concave) functionals, this is not a concern, as the subdifferential conditions characterize global minimizers (maximizers) and are thus physically meaningful. Moreover, stationarity notions for semiconvex (i.e., convex up to a quadratic perturbation) or semiconcave functionals like ours can be reduced to fully convex/concave cases and are also safe.
Definition 3.3 (Nonsmooth constrained critical points).
(i) We say that ( m ‾ , h ‾ s ) ∈ L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_{\mathrm{s}}) \in L^{2}(\mathbf{y}(\Omega);\mathbb{R}^{3}) \times L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) ( m , h s ) ∈ L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) is a constrained critical point of E ^ y \widehat{\mathcal{E}}_{\mathbf{y}} E y if h ‾ s \overline{\mathbf{h}}_{\mathrm{s}} h s is curl-free, χ y ( Ω ) m ‾ + h ‾ s \chi_{\mathbf{y}(\Omega)}\overline{\mathbf{m}}+\overline{\mathbf{h}}_{\mathrm{s}} χ y ( Ω ) m + h s is divergence-free, and the following subdifferential inclusion holds:
μ 0 m ‾ + μ 0 h ‾ s + b a ∈ ∂ m Ψ ^ y ( ⋅ , m ‾ ) a.e. in y ( Ω ) , (55) \mu_{0}\overline{\mathbf{m}}+\mu_{0}\overline{\mathbf{h}}_{\mathrm{s}}+\mathbf{b}_{\mathrm{a}} \in\partial_{\mathbf{m}}\widehat{\Psi}_{\mathbf{y}}(\cdot,\overline{\mathbf{m}}) \quad\text{a.e. in } \mathbf{y}(\Omega), \tag*{(55)} μ 0 m + μ 0 h s + b a ∈ ∂ m Ψ y ( ⋅ , m ) a.e. in y ( Ω ) , (55)
where Ψ ^ y \widehat{\Psi}_{\mathbf{y}} Ψ y is defined in (50) .
(ii) A field b ‾ ∈ L 2 ( R 3 ; R 3 ) \overline{\mathbf{b}} \in L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) b ∈ L 2 ( R 3 ; R 3 ) is a constrained critical point of E y \mathcal{E}_{\mathbf{y}} E y if div b ‾ = 0 \operatorname{div}\overline{\mathbf{b}}=0 div b = 0 in R 3 \mathbb{R}^{3} R 3 , and there exists φ ‾ ∈ W ˚ 1 , 2 ( R 3 ) \overline{\varphi} \in\mathring{W}^{1,2}(\mathbb{R}^{3}) φ ∈ W ˚ 1 , 2 ( R 3 ) such that
b ‾ − b a μ 0 + ∇ φ ‾ ∈ − χ y ( Ω ) ∂ b Φ y ( ⋅ , b ‾ ) a.e. in R 3 . (56) \frac{\overline{\mathbf{b}}-\mathbf{b}_{\mathrm{a}}}{\mu_{0}}+\nabla\overline{\varphi} \in-\chi_{\mathbf{y}(\Omega)}\partial_{\mathbf{b}}\Phi_{\mathbf{y}}(\cdot,\overline{\mathbf{b}}) \quad\text{a.e. in } \mathbb{R}^{3}. \tag*{(56)} μ 0 b − b a + ∇ φ ∈ − χ y ( Ω ) ∂ b Φ y ( ⋅ , b ) a.e. in R 3 . (56)
Here and throughout, ∂ m Ψ ^ y \partial_{\mathbf{m}}\widehat{\Psi}_{\mathbf{y}} ∂ m Ψ y denotes the convex subdifferential with respect to the variable m ∈ R 3 \mathbf{m} \in\mathbb{R}^{3} m ∈ R 3 if Ψ ^ y ( ξ , ⋅ ) \widehat{\Psi}_{\mathbf{y}}(\boldsymbol{\xi},\cdot) Ψ y ( ξ , ⋅ ) is convex, and the concave superdifferential if Ψ ^ y ( ξ , ⋅ ) \widehat{\Psi}_{\mathbf{y}}(\boldsymbol{\xi},\cdot) Ψ y ( ξ , ⋅ ) is concave. The notation for the sub- or superdifferential of Φ y \Phi_{\mathbf{y}} Φ y with respect to b ∈ R 3 \mathbf{b} \in\mathbb{R}^{3} b ∈ R 3 is analogous. For more details see Appendix § 5 .§ 2.2 .
As shown in Appendix § 5 .§ 3 for convex energy densities, the inclusions (56) and (55) are necessary conditions for constrained local minimizers of E y \mathcal{E}_{\mathbf{y}} E y and E ^ y \widehat{\mathcal{E}}_{\mathbf{y}} E y , respectively, and sufficient conditions for global minimizers. In the convex or concave case, they generalize the Euler-Lagrange equations (27) and (26) of Definition 2.1: for a differentiable convex (or concave) function the subdifferential reduces to the singleton consisting of the gradient. In the saddle case, however, the subdifferential inclusions (56) and (55) are not available, and one must rely instead on the smooth framework of Section 2.
Remark 3.4. Since Ψ ^ y ( ξ , m ) = Φ ^ y ( ξ , m ) + μ 0 2 ∣ m ∣ 2 \widehat{\Psi}_{\mathbf{y}}(\boldsymbol{\xi},\mathbf{m})=\widehat{\Phi}_{\mathbf{y}}(\boldsymbol{\xi},\mathbf{m})+\frac{\mu_{0}}{2}|\mathbf{m}|^{2} Ψ y ( ξ , m ) = Φ y ( ξ , m ) + 2 μ 0 ∣ m ∣ 2 , (55) is formally the same as
μ 0 h ‾ s + b a ∈ ∂ m Φ ^ y ( ⋅ , m ‾ ) . (57) \mu_{0}\overline{\mathbf{h}}_{\mathrm{s}}+\mathbf{b}_{\mathrm{a}} \in\partial_{\mathbf{m}}\widehat{\Phi}_{\mathbf{y}}(\cdot,\overline{\mathbf{m}}). \tag*{(57)} μ 0 h s + b a ∈ ∂ m Φ y ( ⋅ , m ) . (57)
However, our examples indicate that Ψ ^ y \widehat{\Psi}_{\mathbf{y}} Ψ y can reasonably be expected to be always either convex or concave in m \mathbf{m} m , while Φ ^ y \widehat{\Phi}_{\mathbf{y}} Φ y might be neither (cf. Example 4.5). Hence, the formulation (55) using Ψ ^ y \widehat{\Psi}_{\mathbf{y}} Ψ y is preferable in the nonsmooth setting.
Remark 3.5. In the smooth case, (55) and (56) reduce to the Euler-Lagrange equations (26) and (27) of Definition 2.1, since for a differentiable convex (or concave) function the subdifferential is the singleton consisting of the gradient.
Correspondence between constrained critical points
We now establish the correspondence between constrained critical points of the two models in both directions.
Theorem 3.6 (From the ( m , h s ) (\mathbf{m},\mathbf{h}_{\mathrm{s}}) ( m , h s ) -based formulation to the b \mathbf{b} b -based formulation). Suppose that (H1) and (H2) hold. In addition, recall that Ψ ^ y ( ξ , m ) = Φ ^ y ( ξ , m ) + μ 0 2 ∣ m ∣ 2 \widehat{\Psi}_{\mathbf{y}}(\boldsymbol{\xi},\mathbf{m})=\widehat{\Phi}_{\mathbf{y}}(\boldsymbol{\xi},\mathbf{m})+\frac{\mu_{0}}{2}|\mathbf{m}|^{2} Ψ y ( ξ , m ) = Φ y ( ξ , m ) + 2 μ 0 ∣ m ∣ 2 and assume that
either Ψ ^ y \widehat{\Psi}_{\mathbf{y}} Ψ y or − Ψ ^ y -\widehat{\Psi}_{\mathbf{y}} − Ψ y is proper (i.e. not everywhere infinite), lower semi-continuous and convex in m \mathbf{m} m , for a.e. ξ \boldsymbol{\xi} ξ .
Moreover, let Φ y \Phi_y Φ y be defined by (H0) and let ( m ‾ , h ‾ s ) (\overline{m},\overline{h}_s) ( m , h s ) be a constrained critical point of E ^ y \widehat{E}_y E y in the sense of Definition 3.3 which satisfy the Maxwell constraint (1.1). Then b ‾ \overline{b} b defined by (2.10), i.e.,
b ‾ = b a + μ 0 ( χ y ( Ω ) m ‾ + h ‾ s ) (58) \overline{b} = b_a + \mu_0\left(\chi_{y(\Omega)}\overline{m}+\overline{h}_s\right) \tag*{(58)} b = b a + μ 0 ( χ y ( Ω ) m + h s ) (58)
is a constrained critical point of E y \mathcal{E}_y E y .
Proof. As in the proof of Theorem 2.6 , we observe that b ‾ \overline{b} b automatically satisfies the Maxwell constraint (1.1) by construction, and we only need to verify that b ‾ \overline{b} b satisfies the subdifferential inclusion (3.2). Since ( m ‾ , h ‾ s ) (\overline{m},\overline{h}_s) ( m , h s ) is a constrained critical point of E ^ y \widehat{E}_y E y , the inclusion (3.1) holds. Substituting the definition of b ‾ \overline{b} b from (2.10), this inclusion reads
b ‾ ( ξ ) ∈ ∂ m Ψ ^ y ( ξ , m ‾ ( ξ ) ) for a.e. ξ in y ( Ω ) . (59) \overline{b}(\xi) \in\partial_m\widehat{\Psi}_y\left(\xi,\overline{m}(\xi)\right) \qquad\text{for a.e. }\xi\text{ in }y(\Omega). \tag*{(59)} b ( ξ ) ∈ ∂ m Ψ y ( ξ , m ( ξ ) ) for a.e. ξ in y ( Ω ) . (59)
By standard properties of the Legendre-Fenchel dual under the convexity or concavity assumption (H4) (see (A.20)), the inclusion (3.5) implies that
m ‾ ( ξ ) ∈ ∂ b Ψ ^ y ⋄ ( ξ , b ‾ ( ξ ) ) for a.e. ξ in y ( Ω ) . \overline{m}(\xi) \in\partial_b\widehat{\Psi}_y^\diamond\left(\xi,\overline{b}(\xi)\right) \qquad\text{for a.e. }\xi\text{ in }y(\Omega). m ( ξ ) ∈ ∂ b Ψ y ⋄ ( ξ , b ( ξ ) ) for a.e. ξ in y ( Ω ) .
Then, by (H0),
m ‾ ( ξ ) ∈ ∂ b ( − Φ y ) ( ξ , b ‾ ) for a.e. ξ in y ( Ω ) . (60) \overline{m}(\xi) \in\partial_b(-\Phi_y)\left(\xi,\overline{b}\right) \qquad\text{for a.e. }\xi\text{ in }y(\Omega). \tag*{(60)} m ( ξ ) ∈ ∂ b ( − Φ y ) ( ξ , b ) for a.e. ξ in y ( Ω ) . (60)
From curl h ‾ s = 0 \operatorname{curl}\overline{h}_s=0 curl h s = 0 , there exists a scalar potential φ ‾ ∈ W ˚ 1 , 2 ( R 3 ) \overline{\varphi}\in\mathring{W}^{1,2}(\mathbb{R}^3) φ ∈ W ˚ 1 , 2 ( R 3 ) such that h ‾ s = − ∇ φ ‾ \overline{h}_s=-\nabla\overline{\varphi} h s = − ∇ φ . From the definition (2.10) we then have
b ‾ − b a μ 0 + ∇ φ ‾ = χ y ( Ω ) m ‾ . (61) \frac{\overline{b}-b_a}{\mu_0}+\nabla\overline{\varphi}=\chi_{y(\Omega)}\overline{m}. \tag*{(61)} μ 0 b − b a + ∇ φ = χ y ( Ω ) m . (61)
Combining (3.6) and (3.7), we obtain
b ‾ − b a μ 0 + ∇ φ ‾ ∈ χ y ( Ω ) ∂ b ( − Φ y ) ( ⋅ , b ‾ ) a.e. in R 3 , (62) \frac{\overline{b}-b_a}{\mu_0}+\nabla\overline{\varphi}\in\chi_{y(\Omega)}\partial_b(-\Phi_y)(\cdot,\overline{b})\qquad\text{a.e. in }\mathbb{R}^3, \tag*{(62)} μ 0 b − b a + ∇ φ ∈ χ y ( Ω ) ∂ b ( − Φ y ) ( ⋅ , b ) a.e. in R 3 , (62)
which is exactly the inclusion (3.2). Hence b ‾ \overline{b} b is a constrained critical point of E y \mathcal{E}_y E y . □
Theorem 3.7 (From the b-based formulation to the ( m , h s ) (m,h_s) ( m , h s ) -based formulation). Suppose that (H1′) and (H2′) hold. In addition, assume that
either Φ y \Phi_y Φ y or − Φ y -\Phi_y − Φ y is proper, lower semi-continuous and convex in b b b , for a.e. ξ \xi ξ . (H4′)
Let Φ ^ y \widehat{\Phi}_y Φ y be defined by (H0) and let b ‾ ∈ L 2 ( R 3 ; R 3 ) \overline{b}\in L^2(\mathbb{R}^3;\mathbb{R}^3) b ∈ L 2 ( R 3 ; R 3 ) be a constrained critical point of E y \mathcal{E}_y E y in the sense of Definition 3.3 which satisfies div b ‾ = 0 \operatorname{div}\overline{b}=0 div b = 0 . Then there exists m ‾ ∈ L 2 ( y ( Ω ) ; R 3 ) \overline{m}\in L^2(y(\Omega);\mathbb{R}^3) m ∈ L 2 ( y ( Ω ) ; R 3 ) such that
m ‾ ∈ − ∂ b Φ y ( ⋅ , b ‾ ) a.e. on y ( Ω ) (63) \overline{m}\in-\partial_b\Phi_y(\cdot,\overline{b})\qquad\text{a.e. on }y(\Omega) \tag*{(63)} m ∈ − ∂ b Φ y ( ⋅ , b ) a.e. on y ( Ω ) (63) and the pair ( m ‾ , h ‾ s ) (\overline{m},\overline{h}_s) ( m , h s ) with h ‾ s \overline{h}_s h s defined by the second of (2.11), i.e., h ‾ s = − P [ χ y ( Ω ) m ‾ ] \overline{h}_s=-\mathcal{P}[\chi_{y(\Omega)}\overline{m}] h s = − P [ χ y ( Ω ) m ] , is a constrained critical point of E ^ y \widehat{E}_y E y . In addition, (3.4) holds.
Proof of Theorem 3.7. Since b ‾ \overline{b} b is a constrained critical point of E y \mathcal{E}_y E y , there exists φ ‾ ∈ W ˚ 1 , 2 ( R 3 ) \overline{\varphi}\in\mathring{W}^{1,2}(\mathbb{R}^3) φ ∈ W ˚ 1 , 2 ( R 3 ) such that
b ‾ − b a μ 0 + ∇ φ ‾ ∈ χ y ( Ω ) ∂ b ( − Φ y ) ( ⋅ , b ‾ ) a.e. in R 3 . (64) \frac{\overline{b}-b_a}{\mu_0}+\nabla\overline{\varphi}\in\chi_{y(\Omega)}\partial_b(-\Phi_y)(\cdot,\overline{b})\qquad\text{a.e. in }\mathbb{R}^3. \tag*{(64)} μ 0 b − b a + ∇ φ ∈ χ y ( Ω ) ∂ b ( − Φ y ) ( ⋅ , b ) a.e. in R 3 . (64)
We define
m ‾ : = ( b ‾ − b a μ 0 + ∇ φ ‾ ) ∣ y ( Ω ) ∈ L 2 ( y ( Ω ) ; R 3 ) . (65) \overline{m}:=\left(\frac{\overline{b}-b_a}{\mu_0}+\nabla\overline{\varphi}\right)\bigg|_{y(\Omega)}\in L^2(y(\Omega);\mathbb{R}^3). \tag*{(65)} m := ( μ 0 b − b a + ∇ φ ) y ( Ω ) ∈ L 2 ( y ( Ω ) ; R 3 ) . (65)
We proceed as in the proof of Theorem 2.6, to show that the definition (2.11), along with (3.11), implies that
h ‾ s = − ∇ ψ ‾ , (66) \overline{\mathbf h}_s=-\nabla\overline{\psi}, \tag*{(66)} h s = − ∇ ψ , (66)
which, together with the hypothesis that both b ‾ \overline{\mathbf b} b and b a \mathbf b_a b a are divergence-free, implies that the pair ( m ‾ , h ‾ s ) (\overline{\mathbf m},\overline{\mathbf h}_s) ( m , h s ) satisfies the Maxwell constraints and that the relation (3.4) holds.
By (3.11) and (3.10), we have m ‾ ∈ ∂ b ( − Φ y ) ( ⋅ , b ‾ ) \overline{\mathbf m}\in\partial_{\mathbf b}(-\Phi_y)(\cdot,\overline{\mathbf b}) m ∈ ∂ b ( − Φ y ) ( ⋅ , b ) a.e. on y ( Ω ) y(\Omega) y ( Ω ) . Since Φ y \Phi_y Φ y is either convex or concave in b \mathbf b b by (H4′), we have Φ y = ( Φ ^ y ) ⋄ = − Ψ ^ y ⋄ \Phi_y=(\widehat{\Phi}_y)^{\diamond}=-\widehat{\Psi}_y^{\diamond} Φ y = ( Φ y ) ⋄ = − Ψ y ⋄ , the latter by (H0′) with Ψ ^ y ( ξ , m ) = Φ ^ y ( ξ , m ) + μ 0 2 ∣ m ∣ 2 \widehat{\Psi}_y(\xi,\mathbf m)=\widehat{\Phi}_y(\xi,\mathbf m)+\frac{\mu_0}{2}|\mathbf m|^2 Ψ y ( ξ , m ) = Φ y ( ξ , m ) + 2 μ 0 ∣ m ∣ 2 . Hence
m ‾ ∈ ∂ b Ψ ^ y ⋄ ( ⋅ , b ‾ ) a.e. on y ( Ω ) . (67) \overline{\mathbf m}\in\partial_{\mathbf b}\widehat{\Psi}_y^{\diamond}(\cdot,\overline{\mathbf b})\quad\text{a.e. on }y(\Omega). \tag*{(67)} m ∈ ∂ b Ψ y ⋄ ( ⋅ , b ) a.e. on y ( Ω ) . (67)
By convex (or concave) duality (see (A.20)), (3.13) is equivalent to
b ‾ ∈ ∂ m Ψ ^ y ( ⋅ , m ‾ ) a.e. on y ( Ω ) . (68) \overline{\mathbf b}\in\partial_{\mathbf m}\widehat{\Psi}_y(\cdot,\overline{\mathbf m})\quad\text{a.e. on }y(\Omega). \tag*{(68)} b ∈ ∂ m Ψ y ( ⋅ , m ) a.e. on y ( Ω ) . (68)
Combined with (3.11), the inclusion (3.14) gives the inclusion (3.1). Hence ( m ‾ , h ‾ s ) (\overline{\mathbf m},\overline{\mathbf h}_s) ( m , h s ) is a constrained critical point of E ^ y \widehat{\mathcal E}_y E y .
Remark 3.8 (Alternative argument via direct global minimization). Equality of energies at the global infimum can also be shown by a direct variational argument already observed in [16 Proposition 2.1] (see also the introduction of [17 ] ). It still works for our more general models and does not even require assumptions of convexity. However, the information we get is very limited unless E ^ \widehat{\mathcal E} E is bounded from below. Assuming such a lower bound excludes all non-constant concave Φ ^ \widehat{\Phi} Φ , in particular the diamagnetic case. The argument goes as follows.
Exploiting the variational characterisation of the projection P \mathcal P P (cf. Remark A.3 in Appendix A.1), the stray field term in E ^ \widehat{\mathcal E} E can be computed by minimization of an auxiliary functional:
∫ R 3 μ 0 2 ∣ h s ∣ 2 d ξ = ∫ R 3 μ 0 2 ∣ P [ χ y ( Ω ) m ] ∣ 2 d ξ = min b ∈ L div 2 ∫ R 3 μ 0 2 ∣ b − χ y ( Ω ) m ∣ 2 d ξ , (69) \int_{\mathbb R^3}\frac{\mu_0}{2}|\mathbf h_s|^2\,\mathrm d\xi
=
\int_{\mathbb R^3}\frac{\mu_0}{2}\left|\mathcal P[\chi_{y(\Omega)}\mathbf m]\right|^2\,\mathrm d\xi
=
\min_{\mathbf b\in L^2_{\operatorname{div}}}
\int_{\mathbb R^3}\frac{\mu_0}{2}\left|\mathbf b-\chi_{y(\Omega)}\mathbf m\right|^2\,\mathrm d\xi, \tag*{(69)} ∫ R 3 2 μ 0 ∣ h s ∣ 2 d ξ = ∫ R 3 2 μ 0 P [ χ y ( Ω ) m ] 2 d ξ = b ∈ L div 2 min ∫ R 3 2 μ 0 b − χ y ( Ω ) m 2 d ξ , (69)
where L div 2 : = { b ∈ L 2 ( R 3 ; R 3 ) : div b = 0 } L^2_{\operatorname{div}}:=\{\mathbf b\in L^2(\mathbb R^3;\mathbb R^3):\operatorname{div}\mathbf b=0\} L div 2 := { b ∈ L 2 ( R 3 ; R 3 ) : div b = 0 } . Hence, for any ( m , h s ) (\mathbf m,\mathbf h_s) ( m , h s ) satisfying (1.1), E ^ y ( m , h s ) = min b ∈ L div 2 F ^ y ( m , b ) \widehat{\mathcal E}_y(\mathbf m,\mathbf h_s)=\min_{\mathbf b\in L^2_{\operatorname{div}}}\widehat{\mathcal F}_y(\mathbf m,\mathbf b) E y ( m , h s ) = min b ∈ L div 2 F y ( m , b ) , where
F ^ y ( m , b ) : = ∫ y ( Ω ) ( Φ ^ y ( ξ , m ) − b a ⋅ m ) d ξ + ∫ R 3 μ 0 2 ∣ b − χ y ( Ω ) m ∣ 2 d ξ . (70) \widehat{\mathcal F}_y(\mathbf m,\mathbf b):=
\int_{y(\Omega)}\left(\widehat{\Phi}_y(\xi,\mathbf m)-\mathbf b_a\cdot\mathbf m\right)\,\mathrm d\xi
+
\int_{\mathbb R^3}\frac{\mu_0}{2}\left|\mathbf b-\chi_{y(\Omega)}\mathbf m\right|^2\,\mathrm d\xi. \tag*{(70)} F y ( m , b ) := ∫ y ( Ω ) ( Φ y ( ξ , m ) − b a ⋅ m ) d ξ + ∫ R 3 2 μ 0 b − χ y ( Ω ) m 2 d ξ . (70)
Since the order of consecutive minimizations can be exchanged,
inf ( m , h s ) : ( 1.1 ) h o l d s E ^ y ( m , h s ) = inf b ∈ L div 2 inf m F ^ y ( m , b ) . (71) \inf_{\substack{(\mathbf m,\mathbf h_s):\\(1.1)\ \mathrm{holds}}}\widehat{\mathcal E}_y(\mathbf m,\mathbf h_s)
=
\inf_{\mathbf b\in L^2_{\operatorname{div}}}\inf_{\mathbf m}\widehat{\mathcal F}_y(\mathbf m,\mathbf b). \tag*{(71)} ( m , h s ) : ( 1.1 ) holds inf E y ( m , h s ) = b ∈ L div 2 inf m inf F y ( m , b ) . (71)
As m ∈ L 2 ( y ( Ω ) ; R 3 ) \mathbf m\in L^2(y(\Omega);\mathbb R^3) m ∈ L 2 ( y ( Ω ) ; R 3 ) is unconstrained, the inner infimum on the right hand side can be computed pointwise via the Fenchel transform of Ψ ^ y \widehat{\Psi}_y Ψ y . This leads to inf m F ^ y ( m , b ) = E y ( b ) \inf_{\mathbf m}\widehat{\mathcal F}_y(\mathbf m,\mathbf b)=\mathcal E_y(\mathbf b) inf m F y ( m , b ) = E y ( b ) , with Φ y \Phi_y Φ y defined by (H0′) as before. Consequently,
inf ( m , h s ) : ( 1.1 ) h o l d s E ^ y ( m , h s ) = inf b ∈ L div 2 E y ( b ) . (72) \inf_{\substack{(\mathbf m,\mathbf h_s):\\(1.1)\ \mathrm{holds}}}\widehat{\mathcal E}_y(\mathbf m,\mathbf h_s)
=
\inf_{\mathbf b\in L^2_{\operatorname{div}}}\mathcal E_y(\mathbf b). \tag*{(72)} ( m , h s ) : ( 1.1 ) holds inf E y ( m , h s ) = b ∈ L div 2 inf E y ( b ) . (72)
Energy equality
We now show that as in the smooth case, at any pair of corresponding critical points, the two energy functionals take the same value.
Proposition 3.9 (Energy equality). Assume that either (H1)–(H2) or (H1′)–(H2′) hold, let m ‾ ∈ L 2 ( y ( Ω ) ; R 3 ) \overline{\mathbf m} \in L^2(y(\Omega);\mathbb{R}^3) m ∈ L 2 ( y ( Ω ) ; R 3 ) and let b ‾ , h s ∈ L 2 ( R 3 ; R 3 ) \overline{\mathbf b},\mathbf h_s \in L^2(\mathbb{R}^3;\mathbb{R}^3) b , h s ∈ L 2 ( R 3 ; R 3 ) . In addition, suppose that the Fenchel identity
Ψ ^ y ( ⋅ , m ‾ ) + Ψ ^ y ⋄ ( ⋅ , b ‾ ) = m ‾ ⋅ b ‾ a.e. on y ( Ω ) , (73) \widehat{\Psi}_y(\cdot,\overline{\mathbf m})+\widehat{\Psi}^{\diamond}_y(\cdot,\overline{\mathbf b})=\overline{\mathbf m}\cdot\overline{\mathbf b}\quad\text{a.e. on }y(\Omega), \tag*{(73)} Ψ y ( ⋅ , m ) + Ψ y ⋄ ( ⋅ , b ) = m ⋅ b a.e. on y ( Ω ) , (73) the physical induction relation
b ‾ = μ 0 ( χ y ( Ω ) m ‾ + h s ) + b a on R 3 , (74) \overline{\mathbf b}=\mu_0\left(\chi_{y(\Omega)}\overline{\mathbf m}+\mathbf h_s\right)+\mathbf b_a\quad\text{on }\mathbb{R}^3, \tag*{(74)} b = μ 0 ( χ y ( Ω ) m + h s ) + b a on R 3 , (74) and the density relations
Φ y ( ξ , b ) = − Ψ ^ y ⋄ ( ξ , b ) and Φ ^ y ( ξ , m ) + μ 0 2 ∣ m ∣ 2 = Ψ ^ y ( ξ , m ) , (75) \Phi_y(\xi,\mathbf b)=-\widehat{\Psi}^{\diamond}_y(\xi,\mathbf b)\quad\text{and}\quad\widehat{\Phi}_y(\xi,\mathbf m)+\frac{\mu_0}{2}|\mathbf m|^2=\widehat{\Psi}_y(\xi,\mathbf m), \tag*{(75)} Φ y ( ξ , b ) = − Ψ y ⋄ ( ξ , b ) and Φ y ( ξ , m ) + 2 μ 0 ∣ m ∣ 2 = Ψ y ( ξ , m ) , (75) hold. Then E y ( b ‾ ) = E ^ y ( m ‾ , h s ) \mathcal E_y(\overline{\mathbf b})=\widehat{\mathcal E}_y(\overline{\mathbf m},\mathbf h_s) E y ( b ) = E y ( m , h s ) .
Proof. The proof is identical to that of Proposition 2.14: it expands (74) to rewrite μ 0 2 ∣ h s ∣ 2 \frac{\mu_0}{2}|\mathbf h_s|^2 2 μ 0 ∣ h s ∣ 2 , substitutes (75) into the definitions of E y \mathcal E_y E y and E ^ y \widehat{\mathcal E}_y E y , and applies the Fenchel identity (73) to conclude. No smoothness of the energy densities is used.
Remark 3.10. The same observations as in the remarks following Proposition 2.14 apply here: the proof uses only (73) and (74) , neither of which requires differentiability. Verifying both conditions simultaneously requires, in practice, the equilibrium condition.
Corollary 3.11 (Energy equality at constrained critical points). Assume that either (H0)–(H2) and (H4) or (H0′)–(H2′) and (H4′) hold. If ( m ‾ , h s ) (\overline{\mathbf m},\mathbf h_s) ( m , h s ) and b ‾ \overline{\mathbf b} b are related by the transition formulas (3.4) or (3.9), respectively, then
E ^ y ( m ‾ , h s ) = E y ( b ‾ ) . (76) \widehat{\mathcal E}_y(\overline{\mathbf m},\mathbf h_s)=\mathcal E_y(\overline{\mathbf b}). \tag*{(76)} E y ( m , h s ) = E y ( b ) . (76)
Proof. The proofs of Theorems 3.6 and 3.7 establish b ‾ ∈ ∂ m Ψ ^ y ( ⋅ , m ‾ ) \overline{\mathbf b}\in\partial_{\mathbf m}\widehat{\Psi}_y(\cdot,\overline{\mathbf m}) b ∈ ∂ m Ψ y ( ⋅ , m ) a.e. on y ( Ω ) y(\Omega) y ( Ω ) , which by convex/concave duality implies the Fenchel identity (73) . The physical relation (74) and density relations (75) hold by construction. Proposition 3.9 (equivalently, Proposition 2.14) then gives the result.
Examples
We illustrate the equivalence results of the preceding sections with some representative examples of standard magnetic materials. We will only discuss the case of identity deformation y = i d \mathbf y=\mathrm{id} y = id , but the same examples can be easily adapted to nontrivial deformations as well.
Smooth cases
Example 4.1 (Diamagnetic or paramagnetic materials). Paramagnetic and diamagnetic materials are described by the following linear relation between magnetic induction and magnetic field:
b = μ h (77) \mathbf b=\mu\mathbf h \tag*{(77)} b = μ h (77)
The constant μ > 0 \mu>0 μ > 0 is the magnetic permeability of the material. The material is diamagnetic if μ < μ 0 \mu<\mu_0 μ < μ 0 and paramagnetic if μ > μ 0 \mu>\mu_0 μ > μ 0 . Using the relation b = μ 0 ( m + h ) \mathbf b=\mu_0(\mathbf m+\mathbf h) b = μ 0 ( m + h ) , the equation (77) can be equivalently rewritten as
m = ( 1 μ 0 − 1 μ ) b (78) \mathbf m=\left(\frac{1}{\mu_0}-\frac{1}{\mu}\right)\mathbf b \tag*{(78)} m = ( μ 0 1 − μ 1 ) b (78)
or
μ 0 μ − μ 0 m = h . (79) \frac{\mu_0}{\mu-\mu_0}\mathbf m=\mathbf h. \tag*{(79)} μ − μ 0 μ 0 m = h . (79)
Since we are in the smooth case, we can use the results from Theorem 2.6 . Observe that (78) coincides with the first equation in (31) , i.e., the explicit constitutive relation
m = − ∇ b Φ y ( ⋅ , b ) a.e. in y ( Ω ) , (80) \mathbf{m}=-\nabla_{\mathbf{b}}\Phi_{\mathbf{y}}(\cdot,\mathbf{b}) \quad\text{a.e. in } y(\Omega), \tag*{(80)} m = − ∇ b Φ y ( ⋅ , b ) a.e. in y ( Ω ) , (80)
at the identity deformation y = id \mathbf{y}=\operatorname{id} y = id if
Φ id ( ξ , b ) = 1 2 ( 1 μ − 1 μ 0 ) ∣ b ∣ 2 . (81) \Phi_{\operatorname{id}}(\boldsymbol{\xi},\mathbf{b})=\frac{1}{2}\left(\frac{1}{\mu}-\frac{1}{\mu_{0}}\right)|\mathbf{b}|^{2}. \tag*{(81)} Φ id ( ξ , b ) = 2 1 ( μ 1 − μ 0 1 ) ∣ b ∣ 2 . (81)
As to the ( m , h s ) (\mathbf{m},\mathbf{h}_{s}) ( m , h s ) -based model, its implicit constitutive relation given at equilibrium by the Euler-Lagrange equation (26) is consistent with (79) (where h = h s + 1 μ 0 b a \mathbf{h}=\mathbf{h}_{s}+\frac{1}{\mu_{0}}\mathbf{b}_{a} h = h s + μ 0 1 b a ) for
Φ ^ id ( ξ , m ) = 1 2 μ 0 2 μ − μ 0 ∣ m ∣ 2 = 1 2 ( 1 μ 0 − 1 μ ) − 1 ∣ m ∣ 2 − μ 0 2 ∣ m ∣ 2 = ( − Φ id ) ⋄ ( ξ , m ) − μ 0 2 ∣ m ∣ 2 . (82) \widehat{\Phi}_{\operatorname{id}}(\boldsymbol{\xi},\mathbf{m})=\frac{1}{2}\frac{\mu_{0}^{2}}{\mu-\mu_{0}}|\mathbf{m}|^{2}=\frac{1}{2}\left(\frac{1}{\mu_{0}}-\frac{1}{\mu}\right)^{-1}|\mathbf{m}|^{2}-\frac{\mu_{0}}{2}|\mathbf{m}|^{2}=(-\Phi_{\operatorname{id}})^{\diamond}(\boldsymbol{\xi},\mathbf{m})-\frac{\mu_{0}}{2}|\mathbf{m}|^{2}. \tag*{(82)} Φ id ( ξ , m ) = 2 1 μ − μ 0 μ 0 2 ∣ m ∣ 2 = 2 1 ( μ 0 1 − μ 1 ) − 1 ∣ m ∣ 2 − 2 μ 0 ∣ m ∣ 2 = ( − Φ id ) ⋄ ( ξ , m ) − 2 μ 0 ∣ m ∣ 2 . (82)
At equilibrium, the inverse relation of (31) is indeed given by (30) , and the two densities are related by (H0) and (H0′), as expected from Theorem 2.6 .
The energy functional of the b-model is
E id ( b ) = ∫ Ω 1 2 ( 1 μ − 1 μ 0 ) ∣ b ∣ 2 + 1 2 μ 0 ∣ b − b a ∣ 2 d ξ + ∫ R 3 ∖ Ω 1 2 μ 0 ∣ b − b a ∣ 2 d ξ . (83) \mathcal{E}_{\operatorname{id}}(\mathbf{b})=\int_{\Omega}\frac{1}{2}\left(\frac{1}{\mu}-\frac{1}{\mu_{0}}\right)|\mathbf{b}|^{2}+\frac{1}{2\mu_{0}}|\mathbf{b}-\mathbf{b}_{a}|^{2}\,\mathrm{d}\xi+\int_{\mathbb{R}^{3}\setminus\Omega}\frac{1}{2\mu_{0}}|\mathbf{b}-\mathbf{b}_{a}|^{2}\,\mathrm{d}\xi. \tag*{(83)} E id ( b ) = ∫ Ω 2 1 ( μ 1 − μ 0 1 ) ∣ b ∣ 2 + 2 μ 0 1 ∣ b − b a ∣ 2 d ξ + ∫ R 3 ∖ Ω 2 μ 0 1 ∣ b − b a ∣ 2 d ξ . (83)
Since μ > 0 \mu>0 μ > 0 , it is strictly convex in b \mathbf{b} b in both the paramagnetic and diamagnetic cases.
The energy functional of the ( m , h s ) (\mathbf{m},\mathbf{h}_{s}) ( m , h s ) -based model is
E ^ id ( m , h s ) = 1 2 μ 0 2 μ − μ 0 ∫ Ω ∣ m ∣ 2 d ξ + μ 0 2 ∫ R 3 ∣ h s ∣ 2 d ξ − ∫ Ω m ⋅ b a d ξ . (84) \widehat{\mathcal{E}}_{\operatorname{id}}(\mathbf{m},\mathbf{h}_{s})=\frac{1}{2}\frac{\mu_{0}^{2}}{\mu-\mu_{0}}\int_{\Omega}|\mathbf{m}|^{2}\,\mathrm{d}\xi+\frac{\mu_{0}}{2}\int_{\mathbb{R}^{3}}|\mathbf{h}_{s}|^{2}\,\mathrm{d}\xi-\int_{\Omega}\mathbf{m}\cdot\mathbf{b}_{a}\,\mathrm{d}\xi. \tag*{(84)} E id ( m , h s ) = 2 1 μ − μ 0 μ 0 2 ∫ Ω ∣ m ∣ 2 d ξ + 2 μ 0 ∫ R 3 ∣ h s ∣ 2 d ξ − ∫ Ω m ⋅ b a d ξ . (84)
In the paramagnetic case μ > μ 0 \mu>\mu_{0} μ > μ 0 , the coefficient μ 0 2 2 ( μ − μ 0 ) \frac{\mu_{0}^{2}}{2(\mu-\mu_{0})} 2 ( μ − μ 0 ) μ 0 2 is positive, so both quadratic terms are convex and E ^ id \widehat{\mathcal{E}}_{\operatorname{id}} E id is jointly convex in ( m , h s ) (\mathbf{m},\mathbf{h}_{s}) ( m , h s ) . In the diamagnetic case μ < μ 0 \mu<\mu_{0} μ < μ 0 , however, μ 0 2 μ − μ 0 \frac{\mu_{0}^{2}}{\mu-\mu_{0}} μ − μ 0 μ 0 2 is negative, whence Φ ^ id \widehat{\Phi}_{\operatorname{id}} Φ id is strictly concave in m \mathbf{m} m . The stray-field term is still positive but not strong enough to compensate: since h s = − P [ χ Ω m ] \mathbf{h}_{s}=-\mathcal{P}[\chi_{\Omega}\mathbf{m}] h s = − P [ χ Ω m ] by the constraints (cf. Remark A.3 in Appendix § 5 .§ 1 ) and the fact that P : L 2 → L 2 \mathcal{P}:L^{2}\to L^{2} P : L 2 → L 2 is an orthogonal projection, we have ∫ R 3 h s ⋅ ( χ Ω m − h s ) d ξ = 0 \int_{\mathbb{R}^{3}}\mathbf{h}_{s}\cdot(\chi_{\Omega}\mathbf{m}-\mathbf{h}_{s})\,\mathrm{d}\xi=0 ∫ R 3 h s ⋅ ( χ Ω m − h s ) d ξ = 0 . Consequently,
μ 0 2 ∫ R 3 ∣ h s ∣ 2 d ξ = μ 0 2 ∫ R 3 ( ∣ χ Ω m ∣ 2 − ∣ χ Ω m − h s ∣ 2 ) d ξ . (85) \frac{\mu_{0}}{2}\int_{\mathbb{R}^{3}}|\mathbf{h}_{s}|^{2}\,\mathrm{d}\xi=\frac{\mu_{0}}{2}\int_{\mathbb{R}^{3}}\left(|\chi_{\Omega}\mathbf{m}|^{2}-|\chi_{\Omega}\mathbf{m}-\mathbf{h}_{s}|^{2}\right)\,\mathrm{d}\xi. \tag*{(85)} 2 μ 0 ∫ R 3 ∣ h s ∣ 2 d ξ = 2 μ 0 ∫ R 3 ( ∣ χ Ω m ∣ 2 − ∣ χ Ω m − h s ∣ 2 ) d ξ . (85)
Substituting back and using Φ ^ id + μ 0 2 ∣ m ∣ 2 = 1 2 ( 1 μ 0 − 1 μ ) − 1 ∣ m ∣ 2 \widehat{\Phi}_{\operatorname{id}}+\frac{\mu_{0}}{2}|\mathbf{m}|^{2}=\frac{1}{2}\left(\frac{1}{\mu_{0}}-\frac{1}{\mu}\right)^{-1}|\mathbf{m}|^{2} Φ id + 2 μ 0 ∣ m ∣ 2 = 2 1 ( μ 0 1 − μ 1 ) − 1 ∣ m ∣ 2 yields the representation
E ^ id ( m , h s ) = 1 2 ( 1 μ 0 − 1 μ ) − 1 ∫ Ω ∣ m ∣ 2 d ξ − μ 0 2 ∫ R 3 ∣ χ Ω m − h s ∣ 2 d ξ − ∫ Ω m ⋅ b a d ξ . (86) \widehat{\mathcal{E}}_{\operatorname{id}}(\mathbf{m},\mathbf{h}_{s})=\frac{1}{2}\left(\frac{1}{\mu_{0}}-\frac{1}{\mu}\right)^{-1}\int_{\Omega}|\mathbf{m}|^{2}\,\mathrm{d}\xi-\frac{\mu_{0}}{2}\int_{\mathbb{R}^{3}}|\chi_{\Omega}\mathbf{m}-\mathbf{h}_{s}|^{2}\,\mathrm{d}\xi-\int_{\Omega}\mathbf{m}\cdot\mathbf{b}_{a}\,\mathrm{d}\xi. \tag*{(86)} E id ( m , h s ) = 2 1 ( μ 0 1 − μ 1 ) − 1 ∫ Ω ∣ m ∣ 2 d ξ − 2 μ 0 ∫ R 3 ∣ χ Ω m − h s ∣ 2 d ξ − ∫ Ω m ⋅ b a d ξ . (86)
For 0 < μ < μ 0 0<\mu<\mu_{0} 0 < μ < μ 0 , we still have 1 2 ( 1 μ 0 − 1 μ ) − 1 = μ μ 0 2 ( μ − μ 0 ) < 0 \frac{1}{2}\left(\frac{1}{\mu_{0}}-\frac{1}{\mu}\right)^{-1}=\frac{\mu\mu_{0}}{2(\mu-\mu_{0})}<0 2 1 ( μ 0 1 − μ 1 ) − 1 = 2 ( μ − μ 0 ) μ μ 0 < 0 and the second term is non-positive. Hence, E ^ id \widehat{\mathcal{E}}_{\operatorname{id}} E id is strictly concave and unbounded from below as ∥ m ∥ L 2 ( Ω ) → ∞ \|\mathbf{m}\|_{L^{2}(\Omega)}\to\infty ∥ m ∥ L 2 ( Ω ) → ∞ .
Example 4.2 (Anisotropic mixed diamagnetic-paramagnetic materials). As a variant of Example 4.1 we can also describe a material which changes its behavior from diamagnetic to paramagnetic, depending on the direction of the applied field. While we do not know whether such behavior can be be found in nature, it is at least conceivable that this might effectively occur in artificially designed microstructures (see, e.g., [15 , 20 ] ) or more complex biological materials. Fix an orthonormal basis { e 1 , e 2 , e 3 } \{\mathbf{e}_{1},\mathbf{e}_{2},\mathbf{e}_{3}\} { e 1 , e 2 , e 3 } , and consider the constitutive equation the constitutive relation
b = μ p ( h ⋅ e 1 ) e 1 + μ d ( h ⋅ e 2 ) e 2 + μ d ( h ⋅ e 3 ) e 3 , with 0 < μ d < μ 0 < μ p , (87) \mathbf{b}=\mu_{p}(\mathbf{h}\cdot\mathbf{e}_{1})\mathbf{e}_{1}+\mu_{d}(\mathbf{h}\cdot\mathbf{e}_{2})\mathbf{e}_{2}+\mu_{d}(\mathbf{h}\cdot\mathbf{e}_{3})\mathbf{e}_{3}, \quad\text{with } 0<\mu_{d}<\mu_{0}<\mu_{p}, \tag*{(87)} b = μ p ( h ⋅ e 1 ) e 1 + μ d ( h ⋅ e 2 ) e 2 + μ d ( h ⋅ e 3 ) e 3 , with 0 < μ d < μ 0 < μ p , (87)
where the permeabilities μ p \mu_p μ p and μ d \mu_d μ d are constant. The relation (87) describes a material whose behavior in the direction e 1 \mathbf{e}_1 e 1 and diamagnetic behavior in the directions e 2 , e 3 \mathbf{e}_2,\mathbf{e}_3 e 2 , e 3 . For instance, for a paramagnetic behavior in a preferred direction e 1 \mathbf{e}_1 e 1 , and diamagnetic behavior in the directions e 2 , e 3 \mathbf{e}_2,\mathbf{e}_3 e 2 , e 3 .
For the single field model at the identity deformation y = id \mathbf{y}=\operatorname{id} y = id , (87) coincides with (80) for the saddle-type density function
Φ id ( ξ , b ) = 1 2 ( 1 μ p − 1 μ 0 ) ( b ⋅ e 1 ) 2 + 1 2 ( 1 μ d − 1 μ 0 ) [ ( b ⋅ e 2 ) 2 + ( b ⋅ e 3 ) 2 ] . (88) \Phi_{\operatorname{id}}(\boldsymbol{\xi},\mathbf{b})=\frac{1}{2}\left(\frac{1}{\mu_p}-\frac{1}{\mu_0}\right)(\mathbf{b}\cdot\mathbf{e}_1)^2+\frac{1}{2}\left(\frac{1}{\mu_d}-\frac{1}{\mu_0}\right)\left[(\mathbf{b}\cdot\mathbf{e}_2)^2+(\mathbf{b}\cdot\mathbf{e}_3)^2\right]. \tag*{(88)} Φ id ( ξ , b ) = 2 1 ( μ p 1 − μ 0 1 ) ( b ⋅ e 1 ) 2 + 2 1 ( μ d 1 − μ 0 1 ) [ ( b ⋅ e 2 ) 2 + ( b ⋅ e 3 ) 2 ] . (88)
Notice that the e i \mathbf{e}_i e i , representing a material anisotropy, are of course natural to prescribe in reference configuration and thus expected to change once a nontrivial deformation y ≠ id \mathbf{y}\ne\operatorname{id} y = id enters the picture. Since we are again in the smooth case, we can use (H0′) to compute the corresponding density
Φ ^ id ( ξ , m ) = 1 2 ( 1 μ 0 − 1 μ p ) − 1 ( m ⋅ e 1 ) 2 + 1 2 ( 1 μ 0 − 1 μ d ) − 1 [ ( m ⋅ e 2 ) 2 + ( m ⋅ e 3 ) 2 ] − μ 0 2 ∣ m ∣ 2 . (89) \widehat{\Phi}_{\operatorname{id}}(\boldsymbol{\xi},\mathbf{m})=\frac{1}{2}\left(\frac{1}{\mu_0}-\frac{1}{\mu_p}\right)^{-1}(\mathbf{m}\cdot\mathbf{e}_1)^2+\frac{1}{2}\left(\frac{1}{\mu_0}-\frac{1}{\mu_d}\right)^{-1}\left[(\mathbf{m}\cdot\mathbf{e}_2)^2+(\mathbf{m}\cdot\mathbf{e}_3)^2\right]-\frac{\mu_0}{2}|\mathbf{m}|^2. \tag*{(89)} Φ id ( ξ , m ) = 2 1 ( μ 0 1 − μ p 1 ) − 1 ( m ⋅ e 1 ) 2 + 2 1 ( μ 0 1 − μ d 1 ) − 1 [ ( m ⋅ e 2 ) 2 + ( m ⋅ e 3 ) 2 ] − 2 μ 0 ∣ m ∣ 2 . (89)
While E id \mathcal{E}_{\operatorname{id}} E id is strictly convex even here, E ^ id \widehat{\mathcal{E}}_{\operatorname{id}} E id now has a saddle structure. Since E id \mathcal{E}_{\operatorname{id}} E id has a unique critical point (its global minimizer), so has E ^ id \widehat{\mathcal{E}}_{\operatorname{id}} E id by Theorem 2.6 . However, the latter now is neither a local minimum nor a local maximum.
Non-smooth cases
Example 4.3 (Permanently magnetic material). Let m 0 ∈ L 2 ( Ω ; R 3 ) \mathbf{m}_0\in L^2(\Omega;\mathbb{R}^3) m 0 ∈ L 2 ( Ω ; R 3 ) be a prescribed magnetization state. For the b \mathbf{b} b -based model, if we set
Φ id ( ξ , b ) = − b ⋅ m 0 , (90) \Phi_{\operatorname{id}}(\boldsymbol{\xi},\mathbf{b})=-\mathbf{b}\cdot\mathbf{m}_0, \tag*{(90)} Φ id ( ξ , b ) = − b ⋅ m 0 , (90)
then the constitutive equation (see (63) ):
m ∈ ∂ b Φ id ( ξ , b ) , (91) \mathbf{m}\in\partial_{\mathbf{b}}\Phi_{\operatorname{id}}(\boldsymbol{\xi},\mathbf{b}), \tag*{(91)} m ∈ ∂ b Φ id ( ξ , b ) , (91)
becomes m = m 0 \mathbf{m}=\mathbf{m}_0 m = m 0 a.e. on Ω \Omega Ω , i.e., the material is permanently magnetized with fixed magnetization m 0 \mathbf{m}_0 m 0 in its undeformed state y = id \mathbf{y}=\operatorname{id} y = id .
The corresponding energy density for the ( m , h s ) (\mathbf{m},\mathbf{h}_s) ( m , h s ) -based model is Φ ^ ( ξ , m ) = Ψ ^ ( ξ , m ) − μ 0 2 ∣ m ∣ 2 \widehat{\Phi}(\boldsymbol{\xi},\mathbf{m})=\widehat{\Psi}(\boldsymbol{\xi},\mathbf{m})-\frac{\mu_0}{2}|\mathbf{m}|^2 Φ ( ξ , m ) = Ψ ( ξ , m ) − 2 μ 0 ∣ m ∣ 2 , where Ψ ^ \widehat{\Psi} Ψ is given by (recall (H0), now to be used with the nonsmooth Legendre-Fenchel transform of Definition A.7):
Ψ ^ ( ξ , m ) = ( − Φ ) ⋄ ( ξ , m ) = { 0 if m = m 0 , + ∞ if m ≠ m 0 . (92) \widehat{\Psi}(\boldsymbol{\xi},\mathbf{m})=(-\Phi)^{\diamond}(\boldsymbol{\xi},\mathbf{m})=
\begin{cases}
0 & \text{if }\mathbf{m}=\mathbf{m}_0,\\
+\infty& \text{if }\mathbf{m}\ne\mathbf{m}_0.
\end{cases} \tag*{(92)} Ψ ( ξ , m ) = ( − Φ ) ⋄ ( ξ , m ) = { 0 + ∞ if m = m 0 , if m = m 0 . (92)
Indeed,
( − Φ id ) ⋄ ( m ) = sup b ( m − m 0 ) ⋅ b = { 0 if m = m 0 , + ∞ if m ≠ m 0 . (93) (-\Phi_{\operatorname{id}})^{\diamond}(\mathbf{m})=\sup_{\mathbf{b}}(\mathbf{m}-\mathbf{m}_0)\cdot\mathbf{b}=
\begin{cases}
0 & \text{if }\mathbf{m}=\mathbf{m}_0,\\
+\infty& \text{if }\mathbf{m}\ne\mathbf{m}_0.
\end{cases} \tag*{(93)} ( − Φ id ) ⋄ ( m ) = b sup ( m − m 0 ) ⋅ b = { 0 + ∞ if m = m 0 , if m = m 0 . (93)
Both E id \mathcal{E}_{\operatorname{id}} E id and E ^ id \widehat{\mathcal{E}}_{\operatorname{id}} E id are now strictly convex, although the latter is degenerate. Here, degeneracy means that E ^ id \widehat{\mathcal{E}}_{\operatorname{id}} E id is convex but its effective domain consists of a single point (the singleton { m 0 } \{\mathbf{m}_0\} { m 0 } ), so that strict convexity holds trivially but the functional carries no information about variations of m \mathbf{m} m away from m 0 \mathbf{m}_0 m 0 .
Remark 4.4. Smooth approximations of the above example can be obtained if we modify (90) with a small diamagnetic or paramagnetic perturbation, i.e., an added quadratic term. This leads to a constitutive equation of the form m = m 0 − ε μ 0 b \mathbf{m}=\mathbf{m}_0-\frac{\varepsilon}{\mu_0}\mathbf{b} m = m 0 − μ 0 ε b for a small ε ≠ 0 \varepsilon\ne0 ε = 0 .
Example 4.5 (Ferromagnetic materials with soft saturation). A natural model for magnetic saturation, consistent with the convex framework of this paper, is to subject the magnetization to the convex constraint ∣ m ∣ ≤ m s \lvert\mathbf{m}\rvert\le m_s ∣ m ∣ ≤ m s , where m s > 0 m_s > 0 m s > 0 is the saturation intensity (typically of the order of 10 6 10^6 1 0 6 A/m). This soft saturation model arises as the large-body-limit relaxation of micromagnetics [5 ] : in the no-exchange limit, the hard saturation constraint ∣ m ∣ = m s \lvert\mathbf{m}\rvert= m_s ∣ m ∣ = m s is replaced by ∣ m ∣ ≤ m s \lvert\mathbf{m}\rvert\le m_s ∣ m ∣ ≤ m s upon relaxation, and the resulting convex variational problem correctly predicts the observed magnetization curves of soft ferromagnets, exhibiting linear response at low fields followed by saturation [5 ] . We define the energy density as the indicator function of the closed ball:
Φ ^ i d ( c ) ( ξ , m ) : = { 0 if ∣ m ∣ ≤ m s , + ∞ otherwise. (94) \widehat{\Phi}_{\mathrm{id}}^{(c)}(\boldsymbol{\xi},\mathbf{m}) :=
\begin{cases}
0 & \text{if } \lvert\mathbf{m}\rvert\le m_s,\\
+\infty& \text{otherwise.}
\end{cases} \tag*{(94)} Φ id ( c ) ( ξ , m ) := { 0 + ∞ if ∣ m ∣ ≤ m s , otherwise. (94) The corresponding augmented density is
Ψ ^ i d ( s ) ( ξ , m ) : = { μ 0 2 ∣ m ∣ 2 if ∣ m ∣ ≤ m s , + ∞ otherwise. (95) \widehat{\Psi}_{\mathrm{id}}^{(s)}(\boldsymbol{\xi},\mathbf{m}) :=
\begin{cases}
\frac{\mu_0}{2}\lvert\mathbf{m}\rvert^2 & \text{if } \lvert\mathbf{m}\rvert\le m_s,\\
+\infty& \text{otherwise.}
\end{cases} \tag*{(95)} Ψ id ( s ) ( ξ , m ) := { 2 μ 0 ∣ m ∣ 2 + ∞ if ∣ m ∣ ≤ m s , otherwise. (95) Applying (H0) to Ψ ^ i d ( s ) \widehat{\Psi}_{\mathrm{id}}^{(s)} Ψ id ( s ) we get
Φ i d ( s ) ( ξ , b ) = − ( Φ ^ i d ( s ) ) ⋄ ( ξ , b ) = − ( Ψ ^ i d ( s ) ) ∗ ( ξ , b ) , (96) \Phi_{\mathrm{id}}^{(s)}(\boldsymbol{\xi},\mathbf{b}) = -\left(\widehat{\Phi}_{\mathrm{id}}^{(s)}\right)^{\diamond}(\boldsymbol{\xi},\mathbf{b}) = -\left(\widehat{\Psi}_{\mathrm{id}}^{(s)}\right)^{*}(\boldsymbol{\xi},\mathbf{b}), \tag*{(96)} Φ id ( s ) ( ξ , b ) = − ( Φ id ( s ) ) ⋄ ( ξ , b ) = − ( Ψ id ( s ) ) ∗ ( ξ , b ) , (96) with the Legendre-Fenchel transform given by Definition A.7 (in all cases, one can use (A.17) instead, which is also defined for Ψ ^ i d ( s a t ) \widehat{\Psi}_{\mathrm{id}}^{(\mathrm{sat})} Ψ id ( sat ) ). Computed explicitly, this yields that
Φ i d ( s ) ( ξ , b ) = − sup ∣ m ∣ ≤ m s ( b ⋅ m − μ 0 2 ∣ m ∣ 2 ) = { − 1 2 μ 0 ∣ b ∣ 2 if ∣ b ∣ ≤ μ 0 m s , − ∣ b ∣ m s + μ 0 2 m s 2 else. (97) \Phi_{\mathrm{id}}^{(s)}(\boldsymbol{\xi},\mathbf{b}) = -\sup_{\lvert\mathbf{m}\rvert\le m_s}\left(\mathbf{b}\cdot\mathbf{m}-\frac{\mu_0}{2}\lvert\mathbf{m}\rvert^2\right)
=
\begin{cases}
-\frac{1}{2\mu_0}\lvert\mathbf{b}\rvert^2 & \text{if } \lvert\mathbf{b}\rvert\le\mu_0m_s,\\
-\lvert\mathbf{b}\rvert m_s+\frac{\mu_0}{2}m_s^2 & \text{else.}
\end{cases} \tag*{(97)} Φ id ( s ) ( ξ , b ) = − ∣ m ∣ ≤ m s sup ( b ⋅ m − 2 μ 0 ∣ m ∣ 2 ) = { − 2 μ 0 1 ∣ b ∣ 2 − ∣ b ∣ m s + 2 μ 0 m s 2 if ∣ b ∣ ≤ μ 0 m s , else. (97)
Auxiliary results
In this appendix we collect some auxiliary results that are used in the main text. They are mostly well known but provided here for the reader’s convenience. In particular, we recall the L 2 L^2 L 2 Helmholtz decomposition on R 3 \mathbb{R}^3 R 3 , which is a key tool in our analysis.
Helmholtz decomposition
Here and throughout the paper, we use the quotient space
W ∘ 1 , 2 ( R d ) = { [ u ] : = { u + λ : λ ∈ R } | u ∈ L l o c 2 ( R d ) and ∇ u exists weakly in L 2 ( R d ; R d ) } . (105) \overset{\circ}{W}{}^{1,2}(\mathbb{R}^{d})=\left\{[u]:=\{u+\lambda:\lambda\in\mathbb{R}\}\,\middle|\,u\in L^{2}_{\mathrm{loc}}(\mathbb{R}^{d})\ \text{and}\ \nabla u\ \text{exists weakly in }L^{2}(\mathbb{R}^{d};\mathbb{R}^{d})\right\}. \tag*{(105)} W ∘ 1 , 2 ( R d ) = { [ u ] := { u + λ : λ ∈ R } u ∈ L loc 2 ( R d ) and ∇ u exists weakly in L 2 ( R d ; R d ) } . (105)
It is a Hilbert space with the norm ∥ [ u ] ∥ W ∘ 1 , 2 = ( ∫ R d ∣ ∇ u ∣ 2 ) 1 2 \|[u]\|_{\overset{\circ}{W}{}^{1,2}}=\left(\int_{\mathbb{R}^{d}}|\nabla u|^{2}\right)^{\frac{1}{2}} ∥ [ u ] ∥ W ∘ 1 , 2 = ( ∫ R d ∣∇ u ∣ 2 ) 2 1 , the W 1 , 2 W^{1,2} W 1 , 2 -gradient semi-norm. As this should be clear from the context, we do not distinguish equivalence classes and their representatives in W ∘ 1 , 2 ( R d ) \overset{\circ}{W}{}^{1,2}(\mathbb{R}^{d}) W ∘ 1 , 2 ( R d ) . In particular, we write u ∈ W ∘ 1 , 2 ( R d ) u\in\overset{\circ}{W}{}^{1,2}(\mathbb{R}^{d}) u ∈ W ∘ 1 , 2 ( R d ) instead of [ u ] ∈ W ∘ 1 , 2 ( R d ) [u]\in\overset{\circ}{W}{}^{1,2}(\mathbb{R}^{d}) [ u ] ∈ W ∘ 1 , 2 ( R d ) . An important property is that if d > 2 d>2 d > 2 , for any u ∈ W ∘ 1 , 2 ( R d ) u\in\overset{\circ}{W}{}^{1,2}(\mathbb{R}^{d}) u ∈ W ∘ 1 , 2 ( R d ) there exists a constant u 0 ∈ R u_{0}\in\mathbb{R} u 0 ∈ R such that u − u 0 u-u_{0} u − u 0 can be approximated with respect to the W 1 , 2 W^{1,2} W 1 , 2 -gradient semi-norm by smooth functions with compact support, see [10 Thms. II.6.3 and II.7.1] . In this sense, C 0 ∞ ( R d ) C^{\infty}_{0}(\mathbb{R}^{d}) C 0 ∞ ( R d ) is dense in W ∘ 1 , 2 ( R d ) \overset{\circ}{W}{}^{1,2}(\mathbb{R}^{d}) W ∘ 1 , 2 ( R d ) for d > 2 d>2 d > 2 , in particular for our case here where d = 3 d=3 d = 3 .
Lemma A.1 (L 2 L^{2} L 2 Helmholtz decomposition). Every u ∈ L 2 ( R 3 ; R 3 ) \mathbf{u}\in L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) u ∈ L 2 ( R 3 ; R 3 ) admits a unique orthogonal decomposition
u = ∇ ϕ + w , (106) \mathbf{u}=\nabla\phi+\mathbf{w}, \tag*{(106)} u = ∇ ϕ + w , (106) where ϕ ∈ W ∘ 1 , 2 ( R 3 ) \phi\in\overset{\circ}{W}{}^{1,2}(\mathbb{R}^{3}) ϕ ∈ W ∘ 1 , 2 ( R 3 ) and div w = 0 \operatorname{div}\mathbf{w}=0 div w = 0 in the distributional sense on R 3 \mathbb{R}^{3} R 3 . The two components are orthogonal in L 2 ( R 3 ; R 3 ) L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) L 2 ( R 3 ; R 3 ) :
⟨ ∇ ϕ , w ⟩ L 2 ( R 3 ) = 0. (107) \langle\nabla\phi,\mathbf{w}\rangle_{L^{2}(\mathbb{R}^{3})}=0. \tag*{(107)} ⟨ ∇ ϕ , w ⟩ L 2 ( R 3 ) = 0. (107) The map P : = ∇ Δ − 1 div : L 2 ( R 3 ; R 3 ) → L 2 ( R 3 ; R 3 ) \mathcal{P}:=\nabla\Delta^{-1}\operatorname{div}:L^{2}(\mathbb{R}^{3};\mathbb{R}^{3})\to L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) P := ∇ Δ − 1 div : L 2 ( R 3 ; R 3 ) → L 2 ( R 3 ; R 3 ) , where the Laplacian Δ : W ∘ 1 , 2 ( R 3 ) → ( W ∘ 1 , 2 ( R 3 ) ) ∗ \Delta:\overset{\circ}{W}{}^{1,2}(\mathbb{R}^{3})\to(\overset{\circ}{W}{}^{1,2}(\mathbb{R}^{3}))^{*} Δ : W ∘ 1 , 2 ( R 3 ) → ( W ∘ 1 , 2 ( R 3 ) ) ∗ acts as an isomorphism, is linear and bounded and has the following additional properties:
P 2 = P (projection) ; P ∗ = P (self-adjoint in L 2 ( R 3 ; R 3 ) ) ; Range ( P ) = { ∇ ϕ : ϕ ∈ W ∘ 1 , 2 ( R 3 ) } ; ker ( P ) = L div 2 ( R 3 ; R 3 ) : = { w ∈ L 2 ( R 3 ; R 3 ) : div w = 0 } . \begin{aligned}
\mathcal{P}^{2}&=\mathcal{P}\quad\text{(projection)};\\
\mathcal{P}^{*}&=\mathcal{P}\quad\text{(self-adjoint in }L^{2}(\mathbb{R}^{3};\mathbb{R}^{3})\text{)};\\
\operatorname{Range}(\mathcal{P})&=\{\nabla\phi:\phi\in\overset{\circ}{W}{}^{1,2}(\mathbb{R}^{3})\};\\
\operatorname{ker}(\mathcal{P})&=L^{2}_{\operatorname{div}}(\mathbb{R}^{3};\mathbb{R}^{3}):=\{\mathbf{w}\in L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}):\operatorname{div}\mathbf{w}=0\}.
\end{aligned} P 2 P ∗ Range ( P ) ker ( P ) = P (projection) ; = P (self-adjoint in L 2 ( R 3 ; R 3 ) ) ; = { ∇ ϕ : ϕ ∈ W ∘ 1 , 2 ( R 3 )} ; = L div 2 ( R 3 ; R 3 ) := { w ∈ L 2 ( R 3 ; R 3 ) : div w = 0 } .
Proof. This is well-known; we provide a short proof for the convenience of the reader. The Laplacian Δ : W ∘ 1 , 2 ( R 3 ) → ( W ∘ 1 , 2 ( R 3 ) ) ∗ \Delta:\overset{\circ}{W}{}^{1,2}(\mathbb{R}^{3})\to(\overset{\circ}{W}{}^{1,2}(\mathbb{R}^{3}))^{*} Δ : W ∘ 1 , 2 ( R 3 ) → ( W ∘ 1 , 2 ( R 3 ) ) ∗ is an isomorphism by the Lax–Milgram theorem, since the norm ∥ [ u ] ∥ W ∘ 1 , 2 2 = ∫ R 3 ∣ ∇ u ∣ 2 d x \|[u]\|_{\overset{\circ}{W}{}^{1,2}}^{2}=\int_{\mathbb{R}^{3}}|\nabla u|^{2}\,\mathrm{d}x ∥ [ u ] ∥ W ∘ 1 , 2 2 = ∫ R 3 ∣∇ u ∣ 2 d x makes W ∘ 1 , 2 ( R 3 ) \overset{\circ}{W}{}^{1,2}(\mathbb{R}^{3}) W ∘ 1 , 2 ( R 3 ) a Hilbert space and the bilinear form ( [ u ] , [ v ] ) ↦ ∫ R 3 ∇ u ⋅ ∇ v d x ([u],[v])\mapsto\int_{\mathbb{R}^{3}}\nabla u\cdot\nabla v\,\mathrm{d}x ([ u ] , [ v ]) ↦ ∫ R 3 ∇ u ⋅ ∇ v d x is continuous and coercive on it.
Given u ∈ L 2 ( R 3 ; R 3 ) \mathbf{u}\in L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) u ∈ L 2 ( R 3 ; R 3 ) , the functional [ v ] ↦ ∫ R 3 u ⋅ ∇ v d x [v]\mapsto\int_{\mathbb{R}^{3}}\mathbf{u}\cdot\nabla v\,\mathrm{d}x [ v ] ↦ ∫ R 3 u ⋅ ∇ v d x is bounded on W ∘ 1 , 2 ( R 3 ) \overset{\circ}{W}{}^{1,2}(\mathbb{R}^{3}) W ∘ 1 , 2 ( R 3 ) (by Cauchy–Schwarz), hence by the Riesz theorem there exists a unique [ ϕ ] ∈ W ∘ 1 , 2 ( R 3 ) [\phi]\in\overset{\circ}{W}{}^{1,2}(\mathbb{R}^{3}) [ ϕ ] ∈ W ∘ 1 , 2 ( R 3 ) such that ∫ R 3 ∇ ϕ ⋅ ∇ v d x = ∫ R 3 u ⋅ ∇ v d x \int_{\mathbb{R}^{3}}\nabla\phi\cdot\nabla v\,\mathrm{d}x=\int_{\mathbb{R}^{3}}\mathbf{u}\cdot\nabla v\,\mathrm{d}x ∫ R 3 ∇ ϕ ⋅ ∇ v d x = ∫ R 3 u ⋅ ∇ v d x for all [ v ] ∈ W ∘ 1 , 2 ( R 3 ) [v]\in\overset{\circ}{W}{}^{1,2}(\mathbb{R}^{3}) [ v ] ∈ W ∘ 1 , 2 ( R 3 ) , i.e., Δ ϕ = div u \Delta\phi=\operatorname{div}\mathbf{u} Δ ϕ = div u in ( W ∘ 1 , 2 ( R 3 ) ) ∗ (\overset{\circ}{W}{}^{1,2}(\mathbb{R}^{3}))^{*} ( W ∘ 1 , 2 ( R 3 ) ) ∗ . Setting w : = u − ∇ ϕ \mathbf{w}:=\mathbf{u}-\nabla\phi w := u − ∇ ϕ , we have div w = div u − Δ ϕ = 0 \operatorname{div}\mathbf{w}=\operatorname{div}\mathbf{u}-\Delta\phi=0 div w = div u − Δ ϕ = 0 distributionally.
The orthogonality ⟨ ∇ ϕ , w ⟩ L 2 = 0 \langle\nabla\phi,\mathbf{w}\rangle_{L^{2}}=0 ⟨ ∇ ϕ , w ⟩ L 2 = 0 follows by integration by parts on R 3 \mathbb{R}^{3} R 3 (since C 0 ∞ ( R 3 ) C^{\infty}_{0}(\mathbb{R}^{3}) C 0 ∞ ( R 3 ) is dense W ∘ 1 , 2 \overset{\circ}{W}{}^{1,2} W ∘ 1 , 2 , the integration by parts can be effectively carried out on a bounded domain without boundary terms appearing):
∫ R 3 ∇ ϕ ⋅ w d x = − ∫ R 3 ϕ div w d x = 0. (108) \int_{\mathbb{R}^{3}}\nabla\phi\cdot\mathbf{w}\,\mathrm{d}x=-\int_{\mathbb{R}^{3}}\phi\,\operatorname{div}\mathbf{w}\,\mathrm{d}x=0. \tag*{(108)} ∫ R 3 ∇ ϕ ⋅ w d x = − ∫ R 3 ϕ div w d x = 0. (108)
Uniqueness of the decomposition follows from the uniqueness of [ ϕ ] [\phi] [ ϕ ] as the solution of Δ ϕ = div u \Delta\phi=\operatorname{div}\mathbf{u} Δ ϕ = div u .
Self-adjointness of P \mathcal{P} P : for any u , v ∈ L 2 ( R 3 ; R 3 ) \mathbf{u},\mathbf{v}\in L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) u , v ∈ L 2 ( R 3 ; R 3 ) , write u = ∇ ϕ u + w u \mathbf{u}=\nabla\phi_{u}+\mathbf{w}_{u} u = ∇ ϕ u + w u and v = ∇ ϕ v + w v \mathbf{v}=\nabla\phi_{v}+\mathbf{w}_{v} v = ∇ ϕ v + w v . Then, using orthogonality of the decomposition,
⟨ P u , v ⟩ L 2 = ⟨ ∇ ϕ u , ∇ ϕ v + w v ⟩ L 2 = ⟨ ∇ ϕ u , ∇ ϕ v ⟩ L 2 = ⟨ u , ∇ ϕ v ⟩ L 2 = ⟨ u , P v ⟩ L 2 . (109) \langle\mathcal{P}\mathbf{u},\mathbf{v}\rangle_{L^{2}}=\langle\nabla\phi_{u},\nabla\phi_{v}+\mathbf{w}_{v}\rangle_{L^{2}}=\langle\nabla\phi_{u},\nabla\phi_{v}\rangle_{L^{2}}=\langle\mathbf{u},\nabla\phi_{v}\rangle_{L^{2}}=\langle\mathbf{u},\mathcal{P}\mathbf{v}\rangle_{L^{2}}. \tag*{(109)} ⟨ P u , v ⟩ L 2 = ⟨ ∇ ϕ u , ∇ ϕ v + w v ⟩ L 2 = ⟨ ∇ ϕ u , ∇ ϕ v ⟩ L 2 = ⟨ u , ∇ ϕ v ⟩ L 2 = ⟨ u , P v ⟩ L 2 . (109)
The projection property P 2 = P \mathcal{P}^{2}=\mathcal{P} P 2 = P is immediate since ∇ ϕ u \nabla\phi_{u} ∇ ϕ u is already a gradient field. □
Remark A.2. Given m ∈ L 2 ( y ( Ω ) ; R 3 ) \mathbf{m} \in L^2(y(\Omega);\mathbb{R}^3) m ∈ L 2 ( y ( Ω ) ; R 3 ) , Lemma A.1 applied to u = χ y ( Ω ) m \mathbf{u}=\chi_{y(\Omega)}\mathbf{m} u = χ y ( Ω ) m shows that there is a unique [ ϕ ] ∈ W ˚ 1 , 2 ( R 3 ) [\phi]\in\mathring{W}^{1,2}(\mathbb{R}^3) [ ϕ ] ∈ W ˚ 1 , 2 ( R 3 ) such that Δ ϕ = − div ( χ y ( Ω ) m ) \Delta\phi=-\operatorname{div}(\chi_{y(\Omega)}\mathbf{m}) Δ ϕ = − div ( χ y ( Ω ) m ) in ( W ˚ 1 , 2 ( R 3 ) ) ∗ (\mathring{W}^{1,2}(\mathbb{R}^3))^* ( W ˚ 1 , 2 ( R 3 ) ) ∗ . Setting h s : = ∇ ϕ \mathbf{h}_{\mathrm{s}}:=\nabla\phi h s := ∇ ϕ , we have curl h s = 0 \operatorname{curl}\mathbf{h}_{\mathrm{s}}=0 curl h s = 0 and div ( h s + χ y ( Ω ) m ) = 0 \operatorname{div}(\mathbf{h}_{\mathrm{s}}+\chi_{y(\Omega)}\mathbf{m})=0 div ( h s + χ y ( Ω ) m ) = 0 , i.e., the Maxwell constraint (1.1) is satisfied. Moreover, since ϕ = − Δ − 1 div ( χ y ( Ω ) m ) \phi=-\Delta^{-1}\operatorname{div}(\chi_{y(\Omega)}\mathbf{m}) ϕ = − Δ − 1 div ( χ y ( Ω ) m ) , we have h s = ∇ ϕ = − ∇ Δ − 1 div ( χ y ( Ω ) m ) = − P [ χ y ( Ω ) m ] \mathbf{h}_{\mathrm{s}}=\nabla\phi=-\nabla\Delta^{-1}\operatorname{div}(\chi_{y(\Omega)}\mathbf{m})=-\mathcal{P}[\chi_{y(\Omega)}\mathbf{m}] h s = ∇ ϕ = − ∇ Δ − 1 div ( χ y ( Ω ) m ) = − P [ χ y ( Ω ) m ] , confirming the formula of Remark A.3.
Remark A.3. Perturbation of h s \mathbf{h}_{\mathrm{s}} h s for fixed m \mathbf{m} m is effectively impossible under the constraints (1.1), as these fully determine h s \mathbf{h}_{\mathrm{s}} h s as a function of m \mathbf{m} m : As a curl-free field, h ∈ L 2 ( R 3 ; R 3 ) \mathbf{h}\in L^2(\mathbb{R}^3;\mathbb{R}^3) h ∈ L 2 ( R 3 ; R 3 ) has a potential η ∈ W ˚ 1 , 2 ( R 3 ) \eta\in\mathring{W}^{1,2}(\mathbb{R}^3) η ∈ W ˚ 1 , 2 ( R 3 ) so that h s = ∇ η \mathbf{h}_{\mathrm{s}}=\nabla\eta h s = ∇ η , and the other constraint implies that Δ η = − div ( χ y ( Ω ) m ) \Delta\eta=-\operatorname{div}(\chi_{y(\Omega)}\mathbf{m}) Δ η = − div ( χ y ( Ω ) m ) in the weak sense in ( W ˚ 1 , 2 ) ∗ (\mathring{W}^{1,2})^* ( W ˚ 1 , 2 ) ∗ , so that
h s = h s m : = − P [ χ y ( Ω ) m ] , where P : = ∇ Δ − 1 div , P : L 2 ( R 3 ; R 3 ) → L 2 ( R 3 ; R 3 ) . (110) \mathbf{h}_{\mathrm{s}}=\mathbf{h}_{\mathrm{s}\mathbf{m}}:=-\mathcal{P}[\chi_{y(\Omega)}\mathbf{m}],\qquad\text{where }\mathcal{P}:=\nabla\Delta^{-1}\operatorname{div},\ \mathcal{P}:L^2(\mathbb{R}^3;\mathbb{R}^3)\to L^2(\mathbb{R}^3;\mathbb{R}^3). \tag*{(110)} h s = h s m := − P [ χ y ( Ω ) m ] , where P := ∇ Δ − 1 div , P : L 2 ( R 3 ; R 3 ) → L 2 ( R 3 ; R 3 ) . (110)
Properties of the Legendre-Fenchel transform
Smooth case: Functions with invertible gradient
Definition A.4 (Smooth Legendre-Fenchel transform). Let f : R 3 → R f:\mathbb{R}^3\to\mathbb{R} f : R 3 → R be continuously differentiable and suppose that ∇ f : R 3 → R 3 \nabla f:\mathbb{R}^3\to\mathbb{R}^3 ∇ f : R 3 → R 3 is continuous and invertible. The smooth Legendre-Fenchel transform of f f f is defined by
f ⋄ ( z ∗ ) : = z ∗ ⋅ ( ∇ f ) − 1 ( z ∗ ) − f ( ( ∇ f ) − 1 ( z ∗ ) ) for all z ∗ ∈ R 3 . (111) f^{\diamond}(\mathbf{z}^*):=\mathbf{z}^*\cdot(\nabla f)^{-1}(\mathbf{z}^*)-f\bigl((\nabla f)^{-1}(\mathbf{z}^*)\bigr)\quad\text{for all }\mathbf{z}^*\in\mathbb{R}^3. \tag*{(111)} f ⋄ ( z ∗ ) := z ∗ ⋅ ( ∇ f ) − 1 ( z ∗ ) − f ( ( ∇ f ) − 1 ( z ∗ ) ) for all z ∗ ∈ R 3 . (111) This definition requires only the invertibility of ∇ f \nabla f ∇ f and applies whether f f f is convex, concave, or neither (e.g., f ( z ) : = z 1 2 + z 2 2 − z 3 2 f(\mathbf{z}):=z_1^2+z_2^2-z_3^2 f ( z ) := z 1 2 + z 2 2 − z 3 2 for z = ( z 1 , z 2 , z 3 ) ∈ R 3 \mathbf{z}=(z_1,z_2,z_3)\in\mathbb{R}^3 z = ( z 1 , z 2 , z 3 ) ∈ R 3 ).
Remark A.5. When ⋄ \diamond ⋄ is applied to a function f ( ξ , ⋅ ) f(\xi,\cdot) f ( ξ , ⋅ ) that also depends on a parameter ξ ∈ y ( Ω ) \xi\in y(\Omega) ξ ∈ y ( Ω ) , it is understood that the transform is taken with respect to the second argument, with ξ \xi ξ fixed.
Lemma A.6 (Properties of the Legendre-Fenchel transform for smooth functions). Let f : R 3 → R f:\mathbb{R}^3\to\mathbb{R} f : R 3 → R , z ↦ f ( z ) \mathbf{z}\mapsto f(\mathbf{z}) z ↦ f ( z ) be continuously differentiable. In addition, assume that ∇ f : R 3 → R 3 \nabla f:\mathbb{R}^3\to\mathbb{R}^3 ∇ f : R 3 → R 3 is invertible with a globally Lipschitz-continuous inverse ( ∇ f ) − 1 (\nabla f)^{-1} ( ∇ f ) − 1 . Then the function f ⋄ f^\diamond f ⋄ defined as in (A.7), i.e.,
f ⋄ ( z ∗ ) = z ∗ ⋅ ( ∇ f ) − 1 ( z ∗ ) − f ( ( ∇ f ) − 1 ( z ∗ ) ) for z ∗ ∈ R 3 , (112) f^\diamond(\mathbf{z}^*)=\mathbf{z}^*\cdot(\nabla f)^{-1}(\mathbf{z}^*)-f\bigl((\nabla f)^{-1}(\mathbf{z}^*)\bigr)\quad\text{for }\mathbf{z}^*\in\mathbb{R}^3, \tag*{(112)} f ⋄ ( z ∗ ) = z ∗ ⋅ ( ∇ f ) − 1 ( z ∗ ) − f ( ( ∇ f ) − 1 ( z ∗ ) ) for z ∗ ∈ R 3 , (112) is also differentiable and its derivative satisfies
∇ f ⋄ = ( ∇ f ) − 1 on R 3 . (113) \nabla f^\diamond=(\nabla f)^{-1}\quad\text{on }\mathbb{R}^3. \tag*{(113)} ∇ f ⋄ = ( ∇ f ) − 1 on R 3 . (113) In particular, we have the linear growth condition
∣ ∇ f ⋄ ( z ∗ ) ∣ ≤ ∥ ∇ ( ∇ f ) − 1 ∥ L ∞ ( R 3 ; R 3 × 3 ) ∣ z ∗ ∣ + ∣ ( ∇ f ) − 1 ( 0 ) ∣ for all z ∗ ∈ R 3 . (114) \left|\nabla f^\diamond(\mathbf{z}^*)\right|\leq\left\|\nabla(\nabla f)^{-1}\right\|_{L^\infty(\mathbb{R}^3;\mathbb{R}^{3\times3})}|\mathbf{z}^*|+\left|(\nabla f)^{-1}(0)\right|\quad\text{for all }\mathbf{z}^*\in\mathbb{R}^3. \tag*{(114)} ∣ ∇ f ⋄ ( z ∗ ) ∣ ≤ ∇ ( ∇ f ) − 1 L ∞ ( R 3 ; R 3 × 3 ) ∣ z ∗ ∣ + ( ∇ f ) − 1 ( 0 ) for all z ∗ ∈ R 3 . (114) Moreover, for any pair ( z , z ∗ ) ∈ R 3 × R 3 (\mathbf{z},\mathbf{z}^*)\in\mathbb{R}^3\times\mathbb{R}^3 ( z , z ∗ ) ∈ R 3 × R 3 ,
z = ∇ f ⋄ ( z ∗ ) if and only if z ∗ = ∇ f ( z ) , (115) \mathbf{z}=\nabla f^\diamond(\mathbf{z}^*)\quad\text{if and only if}\quad\mathbf{z}^*=\nabla f(\mathbf{z}), \tag*{(115)} z = ∇ f ⋄ ( z ∗ ) if and only if z ∗ = ∇ f ( z ) , (115) any of the two equations in (A.11) implies the Fenchel identity
f ⋄ ( z ∗ ) + f ( z ) = z ∗ ⋅ z , (116) f^\diamond(\mathbf{z}^*)+f(\mathbf{z})=\mathbf{z}^*\cdot\mathbf{z}, \tag*{(116)} f ⋄ ( z ∗ ) + f ( z ) = z ∗ ⋅ z , (116) and ( f ⋄ ) ⋄ ( z ) = f ( z ) (f^\diamond)^\diamond(\mathbf{z})=f(\mathbf{z}) ( f ⋄ ) ⋄ ( z ) = f ( z ) for all z ∈ R 3 \mathbf{z}\in\mathbb{R}^3 z ∈ R 3 , i.e., the smooth Legendre-Fenchel transform is an involution.
Proof. We adopt the convention that z z z , ∇ f ∘ ( z ∗ ) \nabla f^\circ(z^*) ∇ f ∘ ( z ∗ ) and ( ∇ f ) − 1 ( z ∗ ) (\nabla f)^{-1}(z^*) ( ∇ f ) − 1 ( z ∗ ) are column vectors, while z ∗ z^* z ∗ and ∇ f ( z ) \nabla f(z) ∇ f ( z ) are row vectors.
Proof of (A.9). Differentiating the definition (A.8) by the chain rule, if ( ∇ f ) − 1 (\nabla f)^{-1} ( ∇ f ) − 1 is of class C 1 C^1 C 1 ,
∇ f ∘ ( z ∗ ) = ( ∇ f ) − 1 ( z ∗ ) + [ z ∗ − ( ∇ f ) ( ( ∇ f ) − 1 ( z ∗ ) ) ] ( ∇ ( ∇ f ) − 1 ( z ∗ ) ) = ( ∇ f ) − 1 ( z ∗ ) , (117) \nabla f^\circ(z^*)=(\nabla f)^{-1}(z^*)+\left[z^*-(\nabla f)((\nabla f)^{-1}(z^*))\right](\nabla(\nabla f)^{-1}(z^*))=(\nabla f)^{-1}(z^*), \tag*{(117)} ∇ f ∘ ( z ∗ ) = ( ∇ f ) − 1 ( z ∗ ) + [ z ∗ − ( ∇ f ) (( ∇ f ) − 1 ( z ∗ )) ] ( ∇ ( ∇ f ) − 1 ( z ∗ )) = ( ∇ f ) − 1 ( z ∗ ) , (117)
since the term in the square brackets vanishes by definition of ( ∇ f ) − 1 (\nabla f)^{-1} ( ∇ f ) − 1 . Using difference quotients instead of the chain rule, the same conclusion holds if ( ∇ f ) − 1 (\nabla f)^{-1} ( ∇ f ) − 1 is only Lipschitz continuous (and hence differentiable almost everywhere by Rademacher’s theorem), since the Lipschitz constant provides a uniform bound for its difference quotients. This proves (A.9). The bound (A.10) is a straightforward consequence of (A.9) and the Lipschitz property of ( ∇ f ) − 1 (\nabla f)^{-1} ( ∇ f ) − 1 .
Proof of (A.11). By (A.9) and the invertibility of ∇ f \nabla f ∇ f , we have
z = ∇ f ∘ ( z ∗ ) = ( ∇ f ) − 1 ( z ∗ ) ⟺ z ∗ = ∇ f ( z ) , (118) z=\nabla f^\circ(z^*)=(\nabla f)^{-1}(z^*)\quad\Longleftrightarrow\quad z^*=\nabla f(z), \tag*{(118)} z = ∇ f ∘ ( z ∗ ) = ( ∇ f ) − 1 ( z ∗ ) ⟺ z ∗ = ∇ f ( z ) , (118)
which is exactly (A.11).
Proof of (A.12). Assuming either equation in (A.11) holds, i.e., z = ( ∇ f ) − 1 ( z ∗ ) z=(\nabla f)^{-1}(z^*) z = ( ∇ f ) − 1 ( z ∗ ) , we substitute directly into (A.8):
f ∘ ( z ∗ ) = z ∗ ⋅ ( ∇ f ) − 1 ( z ∗ ) − f ( ( ∇ f ) − 1 ( z ∗ ) ) = z ∗ ⋅ z − f ( z ) , (119) f^\circ(z^*)=z^*\cdot(\nabla f)^{-1}(z^*)-f((\nabla f)^{-1}(z^*))=z^*\cdot z-f(z), \tag*{(119)} f ∘ ( z ∗ ) = z ∗ ⋅ ( ∇ f ) − 1 ( z ∗ ) − f (( ∇ f ) − 1 ( z ∗ )) = z ∗ ⋅ z − f ( z ) , (119)
which gives (A.12).
Proof of the involution. Since ∇ f ∘ = ( ∇ f ) − 1 \nabla f^\circ=(\nabla f)^{-1} ∇ f ∘ = ( ∇ f ) − 1 by (A.9), applying Definition A.4 to f ∘ f^\circ f ∘ (whose gradient ( ∇ f ) − 1 (\nabla f)^{-1} ( ∇ f ) − 1 has inverse ∇ f \nabla f ∇ f ) gives, for any w ∈ R 3 \mathbf{w}\in\mathbb{R}^3 w ∈ R 3 ,
( f ∘ ) ∘ ( w ) = w ⋅ ( ( ∇ f ) − 1 ) − 1 ( w ) − f ∘ ( ( ( ∇ f ) − 1 ) − 1 ( w ) ) = w ⋅ ∇ f ( w ) − f ∘ ( ∇ f ( w ) ) . (120) (f^\circ)^\circ(\mathbf{w})=\mathbf{w}\cdot((\nabla f)^{-1})^{-1}(\mathbf{w})-f^\circ(((\nabla f)^{-1})^{-1}(\mathbf{w}))=\mathbf{w}\cdot\nabla f(\mathbf{w})-f^\circ(\nabla f(\mathbf{w})). \tag*{(120)} ( f ∘ ) ∘ ( w ) = w ⋅ (( ∇ f ) − 1 ) − 1 ( w ) − f ∘ ((( ∇ f ) − 1 ) − 1 ( w )) = w ⋅ ∇ f ( w ) − f ∘ ( ∇ f ( w )) . (120)
Setting z = w z=\mathbf{w} z = w and z ∗ = ∇ f ( w ) z^*=\nabla f(\mathbf{w}) z ∗ = ∇ f ( w ) in (A.12) yields f ∘ ( ∇ f ( w ) ) = ∇ f ( w ) ⋅ w − f ( w ) f^\circ(\nabla f(\mathbf{w}))=\nabla f(\mathbf{w})\cdot\mathbf{w}-f(\mathbf{w}) f ∘ ( ∇ f ( w )) = ∇ f ( w ) ⋅ w − f ( w ) . Substituting back gives ( f ∘ ) ∘ ( w ) = f ( w ) (f^\circ)^\circ(\mathbf{w})=f(\mathbf{w}) ( f ∘ ) ∘ ( w ) = f ( w ) .
Nonsmooth case: Convex or concave functions
For completeness, we start with a recapitulation of the Fenchel-Legendre transform and relevant notions from convex analysis (see, e.g., [18 ] ). For a proper (i.e., not infinite everywhere) function f : R d → R ∪ { + ∞ } f:\mathbb{R}^d\to\mathbb{R}\cup\{+\infty\} f : R d → R ∪ { + ∞ } , its (convex) Fenchel-Legendre conjugate is
f ∗ ( z ∗ ) : = sup z ∈ R d { z ∗ ⋅ z − f ( z ) } , (121) f^*(z^*):=\sup_{z\in\mathbb{R}^d}\{z^*\cdot z-f(z)\}, \tag*{(121)} f ∗ ( z ∗ ) := z ∈ R d sup { z ∗ ⋅ z − f ( z )} , (121)
and its convex subdifferential at z z z is
∂ f ( z ) : = { z ∗ ∈ R d : f ( w ) ≥ f ( z ) + z ∗ ⋅ ( w − z ) ∀ w ∈ R d } . (122) \partial f(z):=\{z^*\in\mathbb{R}^d:f(w)\ge f(z)+z^*\cdot(w-z)\ \forall w\in\mathbb{R}^d\}. \tag*{(122)} ∂ f ( z ) := { z ∗ ∈ R d : f ( w ) ≥ f ( z ) + z ∗ ⋅ ( w − z ) ∀ w ∈ R d } . (122)
The Fenchel duality relation states
z ∗ ∈ ∂ f ( z ) ⟺ f ( z ) + f ∗ ( z ∗ ) = z ⋅ z ∗ . (123) z^*\in\partial f(z)\quad\Longleftrightarrow\quad f(z)+f^*(z^*)=z\cdot z^*. \tag*{(123)} z ∗ ∈ ∂ f ( z ) ⟺ f ( z ) + f ∗ ( z ∗ ) = z ⋅ z ∗ . (123)
Moreover, if f f f is convex, proper and lower semicontinuous, then f = f ∗ ∗ : = ( f ∗ ) ∗ f=f^{**}:=(f^*)^* f = f ∗∗ := ( f ∗ ) ∗ and we even have that
z ∗ ∈ ∂ f ( z ) ⟺ f ( z ) + f ∗ ( z ∗ ) = z ⋅ z ∗ ⟺ z ∈ ∂ f ∗ ( z ∗ ) . (124) z^*\in\partial f(z)\quad\Longleftrightarrow\quad f(z)+f^*(z^*)=z\cdot z^*\quad\Longleftrightarrow\quad z\in\partial f^*(z^*). \tag*{(124)} z ∗ ∈ ∂ f ( z ) ⟺ f ( z ) + f ∗ ( z ∗ ) = z ⋅ z ∗ ⟺ z ∈ ∂ f ∗ ( z ∗ ) . (124)
For a proper function f : R d → R ∪ { − ∞ } f:\mathbb{R}^d\to\mathbb{R}\cup\{-\infty\} f : R d → R ∪ { − ∞ } , the concave conjugate is defined as
f ∗ ( z ∗ ) : = inf z ∈ R d { z ∗ ⋅ z − f ( z ) } , (125) f_*(z^*):=\inf_{z\in\mathbb{R}^d}\{z^*\cdot z-f(z)\}, \tag*{(125)} f ∗ ( z ∗ ) := z ∈ R d inf { z ∗ ⋅ z − f ( z )} , (125)
see [18 Sections 12, 30] and [19 ] . It satisfies f ∗ = − ( − f ) ∗ f_*=-(-f)^* f ∗ = − ( − f ) ∗ , so no separate theory is needed beyond sign changes. The concave superdifferential is defined via ∂ f ( z ) : = − ∂ ( − f ) ( z ) \partial f(z):=-\partial(-f)(z) ∂ f ( z ) := − ∂ ( − f ) ( z ) , and satisfies the duality relation z ∗ ∈ ∂ f ( z ) z^*\in\partial f(z) z ∗ ∈ ∂ f ( z ) if and only if f ( z ) + f ∗ ( z ∗ ) = z ⋅ z ∗ f(z)+f_*(z^*)=z\cdot z^* f ( z ) + f ∗ ( z ∗ ) = z ⋅ z ∗ .
Definition A.7 (Convex/concave Fenchel-Legendre transform). For d ∈ N d \in\mathbb{N} d ∈ N and proper f : R d × R d → [ − ∞ , + ∞ ] f:\mathbb{R}^{d}\times\mathbb{R}^{d}\to[-\infty,+\infty] f : R d × R d → [ − ∞ , + ∞ ] , we define the convex/concave Fenchel-Legendre transform of f f f with respect to the second variable as
f ⋄ ( ξ , z ∗ ) : = { f ∗ ( ξ , z ∗ ) if z ↦ f ( ξ , z ) is convex , − ( − f ) ∗ ( ξ , − z ∗ ) if z ↦ f ( ξ , z ) is concave , for all ξ , z ∗ ∈ R d . (126) f^{\diamond}(\xi,z^{*}) :=
\begin{cases}
f^{*}(\xi,z^{*}) & \text{if } z\mapsto f(\xi,z)\text{ is convex},\\
-(-f)^{*}(\xi,-z^{*}) & \text{if } z\mapsto f(\xi,z)\text{ is concave},
\end{cases}
\qquad\text{for all }\xi,z^{*}\in\mathbb{R}^{d}. \tag*{(126)} f ⋄ ( ξ , z ∗ ) := { f ∗ ( ξ , z ∗ ) − ( − f ) ∗ ( ξ , − z ∗ ) if z ↦ f ( ξ , z ) is convex , if z ↦ f ( ξ , z ) is concave , for all ξ , z ∗ ∈ R d . (126)
Remark A.8. When f f f is differentiable, strictly convex or strictly concave and ∇ f \nabla f ∇ f is continuous and invertible (invertibility actually automatically follows from strict convexity or strict concavity), the two definitions coincide: f ⋄ f^{\diamond} f ⋄ given by (A.22) equals f ⋄ f^{\diamond} f ⋄ given by (A.7).
Necessary conditions for critical points or extremals
Smooth case
We begin with the smooth case, which provides the Euler-Lagrange equations that characterize constrained critical points in the sense of Definition 2.1.
Proposition A.9 (Euler-Lagrange equations: smooth case). Assume that Φ y ( ξ , ⋅ ) \Phi_y(\xi,\cdot) Φ y ( ξ , ⋅ ) and Φ ^ y ( ξ , ⋅ ) \widehat{\Phi}_y(\xi,\cdot) Φ y ( ξ , ⋅ ) are differentiable with respect to b \mathbf{b} b and m \mathbf{m} m respectively, for a.e. ξ ∈ y ( Ω ) \xi\in y(\Omega) ξ ∈ y ( Ω ) . In addition, assume that the gradients ∇ b Φ y \nabla_{\mathbf{b}}\Phi_y ∇ b Φ y and ∇ m Φ ^ y \nabla_{\mathbf{m}}\widehat{\Phi}_y ∇ m Φ y (and hence ∇ m Ψ ^ y \nabla_{\mathbf{m}}\widehat{\Psi}_y ∇ m Ψ y ) satisfy a linear growth condition in their second argument, uniformly in ξ \xi ξ , so that the functionals E y \mathcal{E}_y E y and E ^ y \widehat{\mathcal{E}}_y E y are Gâteaux-differentiable in their respective magnetic variables on L 2 L^2 L 2 . Then the following holds.
(i) Let ( m ‾ , h ‾ s ) ∈ L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_s)\in L^2(y(\Omega);\mathbb{R}^3)\times L^2(\mathbb{R}^3;\mathbb{R}^3) ( m , h s ) ∈ L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) satisfy (1.1). If
d d ε ∣ ε = 0 E ^ y ( m ‾ + ε η , h ‾ s + ε ζ ) = 0 (127) \left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0}\widehat{\mathcal{E}}_y(\overline{\mathbf{m}}+\varepsilon\boldsymbol{\eta},\overline{\mathbf{h}}_s+\varepsilon\boldsymbol{\zeta})=0 \tag*{(127)} d ε d ε = 0 E y ( m + ε η , h s + ε ζ ) = 0 (127) for every perturbation direction ( η , ζ ) ∈ L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) (\boldsymbol{\eta},\boldsymbol{\zeta})\in L^2(y(\Omega);\mathbb{R}^3)\times L^2(\mathbb{R}^3;\mathbb{R}^3) ( η , ζ ) ∈ L 2 ( y ( Ω ) ; R 3 ) × L 2 ( R 3 ; R 3 ) such that curl ζ = 0 \operatorname{curl}\boldsymbol{\zeta}=0 curl ζ = 0 and div ( χ y ( Ω ) η + ζ ) = 0 \operatorname{div}(\chi_{y(\Omega)}\boldsymbol{\eta}+\boldsymbol{\zeta})=0 div ( χ y ( Ω ) η + ζ ) = 0 , then ( m ‾ , h ‾ s ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_s) ( m , h s ) is a smooth constrained critical point of E ^ y \widehat{\mathcal{E}}_y E y in the sense of Definition 2.1, i.e., (26) holds.
(ii) Let b ‾ ∈ L 2 ( R 3 ; R 3 ) \overline{\mathbf{b}}\in L^2(\mathbb{R}^3;\mathbb{R}^3) b ∈ L 2 ( R 3 ; R 3 ) with div b ‾ = 0 \operatorname{div}\overline{\mathbf{b}}=0 div b = 0 . If
d d ε ∣ ε = 0 E y ( b ‾ + ε ψ ) = 0 (128) \left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0}\mathcal{E}_y(\overline{\mathbf{b}}+\varepsilon\boldsymbol{\psi})=0 \tag*{(128)} d ε d ε = 0 E y ( b + ε ψ ) = 0 (128) for every perturbation direction ψ ∈ L 2 ( R 3 ; R 3 ) \boldsymbol{\psi}\in L^2(\mathbb{R}^3;\mathbb{R}^3) ψ ∈ L 2 ( R 3 ; R 3 ) such that div ψ = 0 \operatorname{div}\boldsymbol{\psi}=0 div ψ = 0 , then b ‾ \overline{\mathbf{b}} b is a smooth constrained critical point of E y \mathcal{E}_y E y in the sense of Definition 2.1, i.e., there exists φ ‾ ∈ W ˚ 1 , 2 ( R 3 ) \overline{\varphi}\in\mathring{W}^{1,2}(\mathbb{R}^3) φ ∈ W ˚ 1 , 2 ( R 3 ) such that (27) holds.
Proof. Part (i). The conditions curl ζ = 0 \operatorname{curl}\boldsymbol{\zeta}=0 curl ζ = 0 and div ( χ y ( Ω ) η + ζ ) = 0 \operatorname{div}(\chi_{y(\Omega)}\boldsymbol{\eta}+\boldsymbol{\zeta})=0 div ( χ y ( Ω ) η + ζ ) = 0 are equivalent to ζ = − P [ χ y ( Ω ) η ] \boldsymbol{\zeta}=-\mathcal{P}[\chi_{y(\Omega)}\boldsymbol{\eta}] ζ = − P [ χ y ( Ω ) η ] by Lemma A.1 , so the perturbation directions are exactly those of the form ( η , − P [ χ y ( Ω ) η ] ) (\boldsymbol{\eta},-\mathcal{P}[\chi_{y(\Omega)}\boldsymbol{\eta}]) ( η , − P [ χ y ( Ω ) η ]) for η ∈ L 2 ( y ( Ω ) ; R 3 ) \boldsymbol{\eta}\in L^2(y(\Omega);\mathbb{R}^3) η ∈ L 2 ( y ( Ω ) ; R 3 ) . For any perturbation direction η ∈ L 2 ( y ( Ω ) ; R 3 ) \boldsymbol{\eta}\in L^2(y(\Omega);\mathbb{R}^3) η ∈ L 2 ( y ( Ω ) ; R 3 ) , differentiate at ε = 0 \varepsilon=0 ε = 0 the map
ε ↦ E ^ y ( m ‾ + ε η , − P [ χ y ( Ω ) ( m ‾ + ε η ) ] ) . \varepsilon\mapsto\widehat{\mathcal{E}}_y\bigl(\overline{\mathbf{m}}+\varepsilon\boldsymbol{\eta},-\mathcal{P}[\chi_{y(\Omega)}(\overline{\mathbf{m}}+\varepsilon\boldsymbol{\eta})]\bigr). ε ↦ E y ( m + ε η , − P [ χ y ( Ω ) ( m + ε η )] ) .
The derivative of the material term gives ∫ y ( Ω ) ∇ m Φ ^ y ( ⋅ , m ‾ ) ⋅ η d ξ \int_{y(\Omega)}\nabla_{\mathbf{m}}\widehat{\Phi}_y(\cdot,\overline{\mathbf{m}})\cdot\boldsymbol{\eta}\,\mathrm{d}\xi ∫ y ( Ω ) ∇ m Φ y ( ⋅ , m ) ⋅ η d ξ . The derivative of the stray-field term μ 0 2 ∫ R 3 ∣ P [ χ y ( Ω ) m ‾ ] ∣ 2 d ξ \frac{\mu_0}{2}\int_{\mathbb{R}^3}|\mathcal{P}[\chi_{y(\Omega)}\overline{\mathbf{m}}]|^2\,\mathrm{d}\xi 2 μ 0 ∫ R 3 ∣ P [ χ y ( Ω ) m ] ∣ 2 d ξ at m ‾ \overline{\mathbf{m}} m in direction η \boldsymbol{\eta} η is
μ 0 ∫ R 3 P [ χ y ( Ω ) m ‾ ] ⋅ P [ χ y ( Ω ) η ] d ξ = − μ 0 ∫ y ( Ω ) h ‾ s ⋅ η d ξ , \mu_0\int_{\mathbb{R}^3}\mathcal{P}[\chi_{y(\Omega)}\overline{\mathbf{m}}]\cdot\mathcal{P}[\chi_{y(\Omega)}\boldsymbol{\eta}]\,\mathrm{d}\xi
=-\mu_0\int_{y(\Omega)}\overline{\mathbf{h}}_s\cdot\boldsymbol{\eta}\,\mathrm{d}\xi, μ 0 ∫ R 3 P [ χ y ( Ω ) m ] ⋅ P [ χ y ( Ω ) η ] d ξ = − μ 0 ∫ y ( Ω ) h s ⋅ η d ξ ,
where we used self-adjointness of P \mathcal{P} P , the identity h ‾ s = − P [ χ y ( Ω ) m ‾ ] \overline{\mathbf{h}}_s=-\mathcal{P}[\chi_{y(\Omega)}\overline{\mathbf{m}}] h s = − P [ χ y ( Ω ) m ] , and P h ‾ s = h ‾ s \mathcal{P}\overline{\mathbf{h}}_s=\overline{\mathbf{h}}_s P h s = h s (since h ‾ s \overline{\mathbf{h}}_s h s is already in the range of P \mathcal{P} P ). The derivative of the forcing term − ∫ y ( Ω ) m ⋅ b a d ξ -\int_{y(\Omega)}\mathbf{m}\cdot\mathbf{b}_a\,\mathrm{d}\xi − ∫ y ( Ω ) m ⋅ b a d ξ gives − ∫ y ( Ω ) η ⋅ b a d ξ -\int_{y(\Omega)} \eta\cdot\mathbf{b}_{\mathrm{a}}\,\mathrm{d}\xi − ∫ y ( Ω ) η ⋅ b a d ξ . Setting the total derivative to zero for all η \eta η and using the definition Ψ ^ y = Φ ^ y + μ 0 2 ∣ m ∣ 2 \widehat{\Psi}_{\mathbf{y}}=\widehat{\Phi}_{\mathbf{y}}+\frac{\mu_{0}}{2}|\mathbf{m}|^{2} Ψ y = Φ y + 2 μ 0 ∣ m ∣ 2 gives (2.6).
Part (ii). By definition of E y \mathcal{E}_{\mathbf{y}} E y in (2.1), we have E y ( b ) = F y ( b ) + G ( b ) \mathcal{E}_{\mathbf{y}}(\mathbf{b})=\mathcal{F}_{\mathbf{y}}(\mathbf{b})+\mathcal{G}(\mathbf{b}) E y ( b ) = F y ( b ) + G ( b ) for all b ∈ L 2 ( R 3 ; R 3 ) \mathbf{b}\in L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) b ∈ L 2 ( R 3 ; R 3 ) , where F y \mathcal{F}_{\mathbf{y}} F y and G \mathcal{G} G are as in (2.1). The Gâteaux derivative of F y \mathcal{F}_{\mathbf{y}} F y at b \mathbf{b} b in direction ψ \boldsymbol{\psi} ψ is
D F y ( b ) [ ψ ] = ∫ R 3 ( χ y ( Ω ) ∇ b Φ y ( ⋅ , b ) + b − b a μ 0 ) ⋅ ψ d ξ . D\mathcal{F}_{\mathbf{y}}(\mathbf{b})[\boldsymbol{\psi}]=\int_{\mathbb{R}^{3}}\left(\chi_{\mathbf{y}(\Omega)}\nabla_{\mathbf{b}}\Phi_{\mathbf{y}}(\cdot,\mathbf{b})+\frac{\mathbf{b}-\mathbf{b}_{\mathrm{a}}}{\mu_{0}}\right)\cdot\boldsymbol{\psi}\,\mathrm{d}\xi. D F y ( b ) [ ψ ] = ∫ R 3 ( χ y ( Ω ) ∇ b Φ y ( ⋅ , b ) + μ 0 b − b a ) ⋅ ψ d ξ .
Since b \mathbf{b} b is a local minimizer of E y \mathcal{E}_{\mathbf{y}} E y subject to div b = 0 \operatorname{div}\mathbf{b}=0 div b = 0 , the first-order condition D F y ( b ) [ ψ ] = 0 D\mathcal{F}_{\mathbf{y}}(\mathbf{b})[\boldsymbol{\psi}]=0 D F y ( b ) [ ψ ] = 0 must hold for every divergence-free ψ ∈ L 2 ( R 3 ; R 3 ) \boldsymbol{\psi}\in L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) ψ ∈ L 2 ( R 3 ; R 3 ) . The space of divergence-free fields is the kernel of P \mathcal{P} P (cf. Lemma A.1), so its orthogonal complement is the range of P \mathcal{P} P , which consists of gradient fields ∇ φ \nabla\varphi ∇ φ . Hence there exists φ ‾ ∈ W ˚ 1 , 2 ( R 3 ) \overline{\varphi}\in\mathring{W}^{1,2}(\mathbb{R}^{3}) φ ∈ W ˚ 1 , 2 ( R 3 ) such that
χ y ( Ω ) ∇ b Φ y ( ⋅ , b ‾ ) + b ‾ − b a μ 0 + ∇ φ ‾ = 0 a.e. in R 3 , \chi_{\mathbf{y}(\Omega)}\nabla_{\mathbf{b}}\Phi_{\mathbf{y}}(\cdot,\overline{\mathbf{b}})+\frac{\overline{\mathbf{b}}-\mathbf{b}_{\mathrm{a}}}{\mu_{0}}+\nabla\overline{\varphi}=0\quad\text{a.e. in }\mathbb{R}^{3}, χ y ( Ω ) ∇ b Φ y ( ⋅ , b ) + μ 0 b − b a + ∇ φ = 0 a.e. in R 3 ,
which is exactly (2.7).
Convex/concave (nonsmooth) case
We now turn to the nonsmooth setting. The following propositions show that the subdifferential inclusions of Definition 3.3 are the natural generalization of the Euler-Lagrange equations above, characterizing global minimizers in the case of convex functionals.
In this section, we briefly demonstrate that like any classical notion of critical points, our Definition 3.3 of magnetic critical points extrapolates the natural first order necessary condition for local extrema, here in case of particular constrained functionals. The necessity of conditions like (3.2) and (3.1) for local minimizers of nonsmooth functionals is apparently not as comprehensively studied in the literature as one might like, except for single integrals and for integrands which are a convex function of gradients in the leading order term (see, e.g., [1 ] ). To the best of our knowledge, the only available results beyond that, in particular for constrained problems, were given in [7 , 8 ] but are not immediately applicable either. Fortunately, our situation is relatively simple and does not require overly technical proofs. We, too, will restrict our attention to convex integrands here, and fully concave cases are of course analogous. The results below could be suitably generalized to semiconvex of semiconcave integrands, but the discussion of anything beyond that would require significantly more effort.
Proposition A.10 (Characterization of magnetic critical points of convex E ^ \widehat{\mathcal{E}} E ). Suppose that (H1) and (H2) hold with Φ ^ y \widehat{\Phi}_{\mathbf{y}} Φ y convex in m \mathbf{m} m , that ( m ‾ , h ‾ s ) ∈ L 2 ( Ω ; R 3 ) × L 2 ( R 3 ; R 3 ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_{\mathrm{s}})\in L^{2}(\Omega;\mathbb{R}^{3})\times L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) ( m , h s ) ∈ L 2 ( Ω ; R 3 ) × L 2 ( R 3 ; R 3 ) satisfies the constraints (1.1) and that E ^ ( y , m ‾ , h ‾ s ) < ∞ \widehat{\mathcal{E}}(\mathbf{y},\overline{\mathbf{m}},\overline{\mathbf{h}}_{\mathrm{s}})<\infty E ( y , m , h s ) < ∞ . If ( m ‾ , h ‾ s ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_{\mathrm{s}}) ( m , h s ) is a constrained local minimizer of ( m , h s ) ↦ E ^ ( y , m , h s ) (\mathbf{m},\mathbf{h}_{\mathrm{s}})\mapsto\widehat{\mathcal{E}}(\mathbf{y},\mathbf{m},\mathbf{h}_{\mathrm{s}}) ( m , h s ) ↦ E ( y , m , h s ) subject to (1.1), then ( m ‾ , h ‾ s ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_{\mathrm{s}}) ( m , h s ) is a magnetic critical point of E ^ \widehat{\mathcal{E}} E in the sense of Definition 3.3, i.e., (3.1) holds. Conversely, if ( m ‾ , h ‾ s ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_{\mathrm{s}}) ( m , h s ) satisfies (3.1), then ( m ‾ , h ‾ s ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_{\mathrm{s}}) ( m , h s ) is a constrained global minimizer of ( m , h s ) ↦ E ^ ( y , m , h s ) (\mathbf{m},\mathbf{h}_{\mathrm{s}})\mapsto\widehat{\mathcal{E}}(\mathbf{y},\mathbf{m},\mathbf{h}_{\mathrm{s}}) ( m , h s ) ↦ E ( y , m , h s ) .
Proof. First notice that by our assumptions, ( m , h s ) ↦ E ^ ( y , m , h s ) (\mathbf{m},\mathbf{h}_{\mathrm{s}})\mapsto\widehat{\mathcal{E}}(\mathbf{y},\mathbf{m},\mathbf{h}_{\mathrm{s}}) ( m , h s ) ↦ E ( y , m , h s ) is convex. Hence, any constrained local minimizer ( m ‾ , h ‾ s ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_{\mathrm{s}}) ( m , h s ) of ( m , h s ) ↦ E ^ ( y , m , h s ) (\mathbf{m},\mathbf{h}_{\mathrm{s}})\mapsto\widehat{\mathcal{E}}(\mathbf{y},\mathbf{m},\mathbf{h}_{\mathrm{s}}) ( m , h s ) ↦ E ( y , m , h s ) with finite energy is automatically a constrained global minimizer and it suffices to discuss global minimizers.
Given (1.1), h s = − P [ χ y ( Ω ) m ] \mathbf{h}_{\mathrm{s}}=-\mathcal{P}[\chi_{\mathbf{y}(\Omega)}\mathbf{m}] h s = − P [ χ y ( Ω ) m ] by Remark A.3. By the definition of E ^ \widehat{\mathcal{E}} E in (2.2), we therefore have that
E ^ ( y , m , h s ) = F ^ y ( m ) + G ^ ( m ) , (129) \widehat{\mathcal{E}}(\mathbf{y},\mathbf{m},\mathbf{h}_{\mathrm{s}})=\widehat{\mathcal{F}}_{\mathbf{y}}(\mathbf{m})+\widehat{\mathcal{G}}(\mathbf{m}), \tag*{(129)} E ( y , m , h s ) = F y ( m ) + G ( m ) , (129)
where
F ^ y ( m ) : = ∫ y ( Ω ) ( Φ ^ y ( ξ , m ) − m ⋅ b a ) d ξ and G ^ ( m ) : = ∫ R 3 μ 0 2 ∣ P [ χ y ( Ω ) m ] ∣ 2 d ξ . (130) \widehat{\mathcal{F}}_{\mathbf{y}}(\mathbf{m}):=\int_{y(\Omega)}\left(\widehat{\Phi}_{\mathbf{y}}(\xi,\mathbf{m})-\mathbf{m}\cdot\mathbf{b}_{\mathrm{a}}\right)\,\mathrm{d}\xi
\quad\text{and}\quad
\widehat{\mathcal{G}}(\mathbf{m}):=\int_{\mathbb{R}^{3}}\frac{\mu_{0}}{2}\left|\mathcal{P}[\chi_{\mathbf{y}(\Omega)}\mathbf{m}]\right|^{2}\,\mathrm{d}\xi. \tag*{(130)} F y ( m ) := ∫ y ( Ω ) ( Φ y ( ξ , m ) − m ⋅ b a ) d ξ and G ( m ) := ∫ R 3 2 μ 0 P [ χ y ( Ω ) m ] 2 d ξ . (130)
Since P \mathcal{P} P is a bounded self-adjoint linear projection operator on L 2 ( R 3 ; R 3 ) L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) L 2 ( R 3 ; R 3 ) , G ^ y : L 2 ( R 3 ; R 3 ) → R \widehat{\mathcal{G}}_{y}:L^{2}(\mathbb{R}^{3};\mathbb{R}^{3})\to\mathbb{R} G y : L 2 ( R 3 ; R 3 ) → R is convex and Frechet-differentiable with derivative
D G ^ y ( m ) [ m ˙ ] = ∫ R 3 μ 0 P [ χ y ( Ω ) m ] ⋅ P [ χ y ( Ω ) m ˙ ] d ξ = ∫ y ( Ω ) μ 0 P [ χ y ( Ω ) m ] ⋅ m ˙ d ξ , (131) D\widehat{\mathcal{G}}_{y}(\mathbf{m})[\dot{\mathbf{m}}]=\int_{\mathbb{R}^{3}}\mu_{0}\mathcal{P}[\chi_{y(\Omega)}\mathbf{m}]\cdot\mathcal{P}[\chi_{y(\Omega)}\dot{\mathbf{m}}]\,\mathrm{d}\xi=\int_{y(\Omega)}\mu_{0}\mathcal{P}[\chi_{y(\Omega)}\mathbf{m}]\cdot\dot{\mathbf{m}}\,\mathrm{d}\xi, \tag*{(131)} D G y ( m ) [ m ˙ ] = ∫ R 3 μ 0 P [ χ y ( Ω ) m ] ⋅ P [ χ y ( Ω ) m ˙ ] d ξ = ∫ y ( Ω ) μ 0 P [ χ y ( Ω ) m ] ⋅ m ˙ d ξ , (131)
for all m ∈ L 2 ( y ( Ω ) ; R 3 ) \mathbf{m}\in L^{2}(y(\Omega);\mathbb{R}^{3}) m ∈ L 2 ( y ( Ω ) ; R 3 ) . Moreover, F ^ y \widehat{\mathcal{F}}_{y} F y is convex and its subdifferential is given by
∂ F ^ y ( m ) = { m ˙ ↦ ∫ y ( Ω ) g ⋅ m ˙ d ξ ∣ g ∈ L 2 ( y ( Ω ) ; R 3 ) , g ∈ μ 0 P [ χ y ( Ω ) m ] − b a + ∂ m Φ ^ y ( ⋅ , m ( ξ ) ) a.e. on y ( Ω ) } . (132) \partial\widehat{\mathcal{F}}_{y}(\mathbf{m})=\left\{\left.\dot{\mathbf{m}}\mapsto\int_{y(\Omega)}\mathbf{g}\cdot\dot{\mathbf{m}}\,\mathrm{d}\xi\ \right|\ \begin{array}{l}
\mathbf{g}\in L^{2}(y(\Omega);\mathbb{R}^{3}),\\
\mathbf{g}\in\mu_{0}\mathcal{P}[\chi_{y(\Omega)}\mathbf{m}]-\mathbf{b}_{a}+\partial_{\mathbf{m}}\widehat{\Phi}_{y}(\cdot,\mathbf{m}(\xi))\ \text{a.e. on }y(\Omega)
\end{array}\right\}. \tag*{(132)} ∂ F y ( m ) = { m ˙ ↦ ∫ y ( Ω ) g ⋅ m ˙ d ξ g ∈ L 2 ( y ( Ω ) ; R 3 ) , g ∈ μ 0 P [ χ y ( Ω ) m ] − b a + ∂ m Φ y ( ⋅ , m ( ξ )) a.e. on y ( Ω ) } . (132)
Here, we have used Riesz representation to identify the elements of the subdifferential naturally given by continuous linear functionals in the dual of L 2 ( y ( Ω ) ; R 3 ) L^{2}(y(\Omega);\mathbb{R}^{3}) L 2 ( y ( Ω ) ; R 3 ) with functions g ∈ L 2 ( y ( Ω ) ; R 3 ) \mathbf{g}\in L^{2}(y(\Omega);\mathbb{R}^{3}) g ∈ L 2 ( y ( Ω ) ; R 3 ) . Combined, we see that m ↦ F ^ y ( m ) + G ^ ( m ) \mathbf{m}\mapsto\widehat{\mathcal{F}}_{y}(\mathbf{m})+\widehat{\mathcal{G}}(\mathbf{m}) m ↦ F y ( m ) + G ( m ) , L 2 ( y ( Ω ) ; R 3 ) → R L^{2}(y(\Omega);\mathbb{R}^{3})\to\mathbb{R} L 2 ( y ( Ω ) ; R 3 ) → R , is convex with subdifferential
∂ ( F ^ y + G ^ ) ( m ) = { m ˙ ↦ ∫ y ( Ω ) g ⋅ m ˙ d ξ ∣ g ∈ L 2 ( y ( Ω ) ; R 3 ) , g ∈ μ 0 P [ χ y ( Ω ) m ] − b a + ∂ m Φ ^ y ( ⋅ , m ) a.e. on y ( Ω ) } . (133) \partial(\widehat{\mathcal{F}}_{y}+\widehat{\mathcal{G}})(\mathbf{m})=\left\{\left.\dot{\mathbf{m}}\mapsto\int_{y(\Omega)}\mathbf{g}\cdot\dot{\mathbf{m}}\,\mathrm{d}\xi\ \right|\ \begin{array}{l}
\mathbf{g}\in L^{2}(y(\Omega);\mathbb{R}^{3}),\\
\mathbf{g}\in\mu_{0}\mathcal{P}[\chi_{y(\Omega)}\mathbf{m}]-\mathbf{b}_{a}+\partial_{\mathbf{m}}\widehat{\Phi}_{y}(\cdot,\mathbf{m})\ \text{a.e. on }y(\Omega)
\end{array}\right\}. \tag*{(133)} ∂ ( F y + G ) ( m ) = { m ˙ ↦ ∫ y ( Ω ) g ⋅ m ˙ d ξ g ∈ L 2 ( y ( Ω ) ; R 3 ) , g ∈ μ 0 P [ χ y ( Ω ) m ] − b a + ∂ m Φ y ( ⋅ , m ) a.e. on y ( Ω ) } . (133)
Finally, ( m ‾ , h ‾ s ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_{s}) ( m , h s ) is a constrained global minimizer ( m ‾ , h ‾ s ) (\overline{\mathbf{m}},\overline{\mathbf{h}}_{s}) ( m , h s ) of ( m , h s ) ↦ E ^ ( y , m , h s ) (\mathbf{m},\mathbf{h}_{s})\mapsto\widehat{\mathcal{E}}(y,\mathbf{m},\mathbf{h}_{s}) ( m , h s ) ↦ E ( y , m , h s ) if and only if m ‾ \overline{\mathbf{m}} m is a global minimizer of F ^ y + G ^ \widehat{\mathcal{F}}_{y}+\widehat{\mathcal{G}} F y + G . The latter is in turn equivalent to
0 ∈ ∂ ( F ^ y + G ^ ) . (134) \mathbf{0}\in\partial(\widehat{\mathcal{F}}_{y}+\widehat{\mathcal{G}}). \tag*{(134)} 0 ∈ ∂ ( F y + G ) . (134)
By our characterizations of ∂ F y \partial\mathcal{F}_{y} ∂ F y and ∂ G \partial\mathcal{G} ∂ G , we see that this in turn is equivalent to (55) .
Auxiliary results
Proposition A.11 (Characterization of magnetic critical points of convex E \mathcal{E} E ). Suppose that (H1′) and (H2′) hold, that
Ψ y ( ξ , b ) = χ y ( Ω ) ( ξ ) Φ y ( ξ , b ) + 1 2 μ 0 ∣ b − b a ( ξ ) ∣ 2 (135) \Psi_{y}(\xi,\mathbf{b})=\chi_{y(\Omega)}(\xi)\Phi_{y}(\xi,\mathbf{b})+\frac{1}{2\mu_{0}}\lvert\mathbf{b}-\mathbf{b}_{a}(\xi)\rvert^{2} \tag*{(135)} Ψ y ( ξ , b ) = χ y ( Ω ) ( ξ ) Φ y ( ξ , b ) + 2 μ 0 1 ∣ b − b a ( ξ ) ∣ 2 (135) is convex in b \mathbf{b} b for a.e. ξ \xi ξ , that b ‾ ∈ L 2 ( R 3 ; R 3 ) \overline{\mathbf{b}}\in L^{2}(\mathbb{R}^{3};\mathbb{R}^{3}) b ∈ L 2 ( R 3 ; R 3 ) satisfies div b ‾ = 0 \operatorname{div}\overline{\mathbf{b}}=0 div b = 0 on R 3 \mathbb{R}^{3} R 3 and that E ( y , b ‾ ) < ∞ \mathcal{E}(y,\overline{\mathbf{b}})<\infty E ( y , b ) < ∞ . If b ‾ \overline{\mathbf{b}} b is a local minimizer of b ↦ E ( y , b ) \mathbf{b}\mapsto\mathcal{E}(y,\mathbf{b}) b ↦ E ( y , b ) , L 2 ( R 3 ; R 3 ) → R L^{2}(\mathbb{R}^{3};\mathbb{R}^{3})\to\mathbb{R} L 2 ( R 3 ; R 3 ) → R under the constraint div b = 0 \operatorname{div}\mathbf{b}=0 div b = 0 , then b ‾ \overline{\mathbf{b}} b is a magnetic critical point of E \mathcal{E} E in the sense of Definition 3.3, i.e., (56) holds with some φ ‾ ∈ W ˚ 1 , 2 ( R 3 ) \overline{\varphi}\in\mathring{W}^{1,2}(\mathbb{R}^{3}) φ ∈ W ˚ 1 , 2 ( R 3 ) . Conversely, if (56) holds for some φ ‾ ∈ W ˚ 1 , 2 ( R 3 ) \overline{\varphi}\in\mathring{W}^{1,2}(\mathbb{R}^{3}) φ ∈ W ˚ 1 , 2 ( R 3 ) , then b ‾ \overline{\mathbf{b}} b is a constrained global minimizer of b ↦ E ( y , b ) \mathbf{b}\mapsto\mathcal{E}(y,\mathbf{b}) b ↦ E ( y , b ) .
Proof of Proposition A.11. Since b ↦ E ( y , b ) \mathbf{b}\mapsto\mathcal{E}(y,\mathbf{b}) b ↦ E ( y , b ) is convex, it suffices to discuss global minimizers, because any local minimizers with finite energy are automatically global. By definition of E \mathcal{E} E in (21) , we have that
E ( y , b ) = F y ( b ) + G ( b ) for all b ∈ L 2 ( R 3 ; R 3 ) with div b = 0 on R 3 , (137) \mathcal{E}(y,\mathbf{b})=\mathcal{F}_{y}(\mathbf{b})+\mathcal{G}(\mathbf{b})\quad\text{for all }\mathbf{b}\in L^{2}(\mathbb{R}^{3};\mathbb{R}^{3})\text{ with }\operatorname{div}\mathbf{b}=0\text{ on }\mathbb{R}^{3}, \tag*{(137)} E ( y , b ) = F y ( b ) + G ( b ) for all b ∈ L 2 ( R 3 ; R 3 ) with div b = 0 on R 3 , (137)
where F y , G : L 2 ( R 3 ; R 3 ) → R ∪ { + ∞ } \mathcal{F}_y,\mathcal{G}:L^2(\mathbb{R}^3;\mathbb{R}^3)\to\mathbb{R}\cup\{+\infty\} F y , G : L 2 ( R 3 ; R 3 ) → R ∪ { + ∞ } ,
F y ( b ) : = ∫ R 3 Ψ y ( ξ , b ) d ξ , G ( b ) : = { 0 if div b = 0 on R 3 , + ∞ else. (138) \mathcal{F}_y(\mathbf{b}):=\int_{\mathbb{R}^3}\Psi_y(\xi,\mathbf{b})\,\mathrm{d}\xi,\qquad\mathcal{G}(\mathbf{b}):=\begin{cases}0 & \text{if }\operatorname{div}\mathbf{b}=0\text{ on }\mathbb{R}^3,\\+\infty& \text{else.}\end{cases} \tag*{(138)} F y ( b ) := ∫ R 3 Ψ y ( ξ , b ) d ξ , G ( b ) := { 0 + ∞ if div b = 0 on R 3 , else. (138)
Both F y \mathcal{F}_y F y and G \mathcal{G} G are convex and lower semicontinuous, the latter because the set of divergence-free fields in L 2 L^2 L 2 is closed. Moreover, for any b ∈ L 2 ( R 3 ; R 3 ) \mathbf{b}\in L^2(\mathbb{R}^3;\mathbb{R}^3) b ∈ L 2 ( R 3 ; R 3 ) , their subdifferentials (as unconstrained functionals on L 2 ( R 3 ; R 3 ) L^2(\mathbb{R}^3;\mathbb{R}^3) L 2 ( R 3 ; R 3 ) ) are given by
∂ F y ( b ) = { b ˙ ↦ ∫ R 3 g ⋅ b ˙ d ξ | g ∈ L 2 ( R 3 ; R 3 ) , g ( ξ ) ∈ ∂ b Ψ y ( ξ , b ( ξ ) ) for a.e. ξ ∈ R 3 } , \partial\mathcal{F}_y(\mathbf{b})=\left\{\dot{\mathbf b}\mapsto\int_{\mathbb{R}^3}\mathbf{g}\cdot\dot{\mathbf b}\,\mathrm{d}\xi\ \middle|\ \mathbf{g}\in L^2(\mathbb{R}^3;\mathbb{R}^3),\ \mathbf{g}(\xi)\in\partial_{\mathbf b}\Psi_y(\xi,\mathbf{b}(\xi))\text{ for a.e. }\xi\in\mathbb{R}^3\right\}, ∂ F y ( b ) = { b ˙ ↦ ∫ R 3 g ⋅ b ˙ d ξ g ∈ L 2 ( R 3 ; R 3 ) , g ( ξ ) ∈ ∂ b Ψ y ( ξ , b ( ξ )) for a.e. ξ ∈ R 3 } ,
∂ G ( b ) = { b ˙ ↦ ∫ y ( Ω ) g ⋅ b ˙ d ξ | g = ∇ φ ‾ for a φ ‾ ∈ W ˚ 1 , 2 ( R 3 ) } if div b = 0 , \partial\mathcal{G}(\mathbf{b})=\left\{\dot{\mathbf b}\mapsto\int_{y(\Omega)}\mathbf{g}\cdot\dot{\mathbf b}\,\mathrm{d}\xi\ \middle|\ \mathbf{g}=\nabla\overline{\varphi}\text{ for a }\overline{\varphi}\in\mathring{W}^{1,2}(\mathbb{R}^3)\right\}\quad\text{if }\operatorname{div}\mathbf b=0, ∂ G ( b ) = { b ˙ ↦ ∫ y ( Ω ) g ⋅ b ˙ d ξ g = ∇ φ for a φ ∈ W ˚ 1 , 2 ( R 3 ) } if div b = 0 ,
∂ G ( b ) = ∅ otherwise. (139) \partial\mathcal{G}(\mathbf{b})=\varnothing\quad\text{otherwise.} \tag*{(139)} ∂ G ( b ) = ∅ otherwise. (139)
Here, we have used Riesz representation to identify the elements of the subdifferential naturally given by linear functionals in the dual of L 2 ( R 3 ; R 3 ) L^2(\mathbb{R}^3;\mathbb{R}^3) L 2 ( R 3 ; R 3 ) with functions g ∈ L 2 ( R 3 ; R 3 ) \mathbf{g}\in L^2(\mathbb{R}^3;\mathbb{R}^3) g ∈ L 2 ( R 3 ; R 3 ) . Moreover, for stating ∂ G ( b ) \partial\mathcal{G}(\mathbf{b}) ∂ G ( b ) , we have exploited that
g = ∇ φ ‾ for a φ ‾ ∈ W ˚ 1 , 2 if and only if ∫ R 3 g ⋅ b ˙ d ξ = 0 for all b ˙ ∈ L 2 with div b ˙ = 0. (140) \mathbf{g}=\nabla\overline{\varphi}\text{ for a }\overline{\varphi}\in\mathring{W}^{1,2}\quad\text{if and only if}\quad\int_{\mathbb{R}^3}\mathbf{g}\cdot\dot{\mathbf b}\,\mathrm{d}\xi=0\text{ for all }\dot{\mathbf b}\in L^2\text{ with }\operatorname{div}\dot{\mathbf b}=0. \tag*{(140)} g = ∇ φ for a φ ∈ W ˚ 1 , 2 if and only if ∫ R 3 g ⋅ b ˙ d ξ = 0 for all b ˙ ∈ L 2 with div b ˙ = 0. (140)
(In terms of the projection of Appendix A.1: the range of P \mathcal{P} P , i.e., the gradient fields, is the orthogonal complement of its kernel, i.e., the divergence-free fields.)
Finally, we observe that b ‾ \overline{\mathbf b} b is a constrained minimizer of b ↦ E ( y , b ) \mathbf b\mapsto\mathcal{E}(\mathbf y,\mathbf b) b ↦ E ( y , b ) if and only if b ‾ \overline{\mathbf b} b is an (unconstrained) minimizer of F y + G \mathcal{F}_y+\mathcal{G} F y + G . In view of the definition G \mathcal{G} G , the latter is equivalent to
div b ‾ = 0 and 0 ∈ ∂ ( F y + G ) ( b ‾ ) = ∂ F y ( b ‾ ) + ∂ G ( b ‾ ) . (141) \operatorname{div}\overline{\mathbf b}=0\quad\text{and}\quad\mathbf0\in\partial(\mathcal{F}_y+\mathcal{G})(\overline{\mathbf b})=\partial\mathcal{F}_y(\overline{\mathbf b})+\partial\mathcal{G}(\overline{\mathbf b}). \tag*{(141)} div b = 0 and 0 ∈ ∂ ( F y + G ) ( b ) = ∂ F y ( b ) + ∂ G ( b ) . (141)
By our characterizations of ∂ G \partial\mathcal{G} ∂ G and ∂ F y \partial\mathcal{F}_y ∂ F y , we see that the inclusion in (A.37) is in turn equivalent to the existence of some φ ‾ ∈ W ˚ 1 , 2 ( R 3 ) \overline{\varphi}\in\mathring{W}^{1,2}(\mathbb{R}^3) φ ∈ W ˚ 1 , 2 ( R 3 ) such that (3.2) holds (recall (A.32)).