Introduction
Optimal transport provides a variational framework for comparing and redistributing mass. In its classical formulation, the source and target measures have the same total mass. Unbalanced optimal transport relaxes this requirement by allowing mass creation and removal.
Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be a bounded domain, let ρ 0 \rho_0 ρ 0 and ρ 1 \rho_1 ρ 1 denote the nonnegative initial and target densities, and let α > 0 \alpha> 0 α > 0 . For a map M ∈ W 1 , n ( Ω ; R n ) M \in W^{1,n}(\Omega;\mathbb{R}^n) M ∈ W 1 , n ( Ω ; R n ) and a spatially independent source f ∈ L 2 ( 0 , 1 ) f \in L^2(0,1) f ∈ L 2 ( 0 , 1 ) , the static energy is formally given by
E α ( M , f ) = ∫ Ω ∣ M ( x ) − x ∣ 2 ρ 0 ( x ) d x + 1 α ∫ 0 1 ∣ f ( t ) ∣ 2 d t + ∫ 0 1 ∫ 0 t f ( s ) ∫ Ω ∣ M ( x ) − x ∣ 2 det ( s ∇ M ( x ) + ( 1 − s ) I ) d x d s d t . (1) E_\alpha(M,f)=\int_\Omega|M(x)-x|^2\rho_0(x)\,dx+\frac{1}{\alpha}\int_0^1|f(t)|^2\,dt+\int_0^1\int_0^t f(s)\int_\Omega|M(x)-x|^2\det\bigl(s\nabla M(x)+(1-s)I\bigr)\,dx\,ds\,dt. \tag*{(1)} E α ( M , f ) = ∫ Ω ∣ M ( x ) − x ∣ 2 ρ 0 ( x ) d x + α 1 ∫ 0 1 ∣ f ( t ) ∣ 2 d t + ∫ 0 1 ∫ 0 t f ( s ) ∫ Ω ∣ M ( x ) − x ∣ 2 det ( s ∇ M ( x ) + ( 1 − s ) I ) d x d s d t . (1)
where I I I is the identity matrix. The first term is the displacement cost of the initial mass, the second penalizes the source, and the last accounts for its contribution along the affine interpolation from the identity to M M M . The map and the source satisfy the endpoint constraint
ρ 1 ( M ( x ) ) det ∇ M ( x ) = ρ 0 ( x ) + ∫ 0 1 f ( t ) J t [ M ] ( x ) d t . (2) \rho_1(M(x))\det\nabla M(x)=\rho_0(x)+\int_0^1 f(t)J_t[M](x)\,dt. \tag*{(2)} ρ 1 ( M ( x )) det ∇ M ( x ) = ρ 0 ( x ) + ∫ 0 1 f ( t ) J t [ M ] ( x ) d t . (2)
Its distributional interpretation and the remaining admissibility conditions are specified below.
Write X t [ M ] : = ( 1 − t ) Id + t M X_t[M] := (1-t)\operatorname{Id} + tM X t [ M ] := ( 1 − t ) Id + tM , J t [ M ] : = det ( ( 1 − t ) I + t ∇ M ) J_t[M] := \det\left((1-t)I + t\nabla M\right) J t [ M ] := det ( ( 1 − t ) I + t ∇ M ) and define the intermediate state
R t [ M , f ] ( x ) : = ρ 0 ( x ) + ∫ 0 t f ( s ) J s [ M ] ( x ) d s . R_t[M,f](x) := \rho_0(x) + \int_0^t f(s)J_s[M](x)\,ds. R t [ M , f ] ( x ) := ρ 0 ( x ) + ∫ 0 t f ( s ) J s [ M ] ( x ) d s .
For pairs with nonnegative intermediate states, define
G ( M , f ) : = ∫ 0 1 ∫ Ω ∣ M ( x ) − x ∣ 2 R t [ M , f ] ( x ) d x d t ∈ [ 0 , + ∞ ] G(M,f) := \int_0^1 \int_\Omega|M(x)-x|^2 R_t[M,f](x)\,dx\,dt \in[0,+\infty] G ( M , f ) := ∫ 0 1 ∫ Ω ∣ M ( x ) − x ∣ 2 R t [ M , f ] ( x ) d x d t ∈ [ 0 , + ∞ ]
and
E α ( M , f ) : = G ( M , f ) + 1 α ∥ f ∥ L 2 ( 0 , 1 ) 2 . E_\alpha(M,f) := G(M,f) + \frac{1}{\alpha}\|f\|_{L^2(0,1)}^2. E α ( M , f ) := G ( M , f ) + α 1 ∥ f ∥ L 2 ( 0 , 1 ) 2 .
This definition agrees with (1) for bounded-range maps and whenever the terms in that expression are absolutely integrable. We study existence for this static problem, its correspondence with dynamic transport under prescribed affine-flow conditions, and the limits selected by strong source penalization. We also give sufficient conditions and explicit examples verifying the required feasibility assumptions. We first review the related literature and then describe our approach and main results.
Background and related work
The classical transport problem originates with Monge [24 ] . Kantorovich [18 ] introduced its relaxation to transport plans, while Benamou and Brenier [2 ] formulated quadratic optimal transport through a continuity equation and a kinetic action. Existence of optimal maps has been studied under various assumptions on the cost and the endpoint measures [11 , 16 , 4 , 1 , 29 ] . For background on classical optimal transport, we refer the reader to [30 , 31 ] .
Several extensions of optimal transport have been developed for finite nonnegative measures with unequal total masses. One approach to unbalanced optimal transport extends the Benamou-Brenier formulation by introducing a source into the continuity equation and penalizing mass creation and removal. Chizat et al. [5 ] introduced the Wasserstein-Fisher-Rao metric, which combines the quadratic Wasserstein action for spatial transport with the Fisher-Rao action for mass variation. In [6 ] , they studied a broader class of dynamic transport costs defined by convex, positively one-homogeneous action densities and established equivalent static Kantorovich formulations in terms of semi-couplings. Static approaches include Kantorovich-type problems over transport plans and Monge-type problems involving transport maps and mass scaling factors. Guittet [17 ] studied extended Kantorovich norms with a fixed cost per unit of mass created or removed. Piccoli and Rossi [26 ] introduced generalized Wasserstein distances combining W p W_p W p transport with total variation penalties for mass variation, and subsequently established their dynamic characterization and duality properties in [27 ] . Liero, Mielke, and Savaré [22 ] developed the entropy-transport framework as a Kantorovich-type problem with entropy penalties on the marginals. Savaré and Sodini [28 ] obtained formulations over homogeneous couplings through convex relaxation and investigated the corresponding Monge formulations. For the Hellinger-Kantorovich distance, Liero, Mielke, and Savaré [21 ] established sufficient conditions for the solvability of the Monge problem and constructed transport-growth flows describing geodesics. Gallouët, Ghezzi, and Vialard [14 ] studied regularity in the static entropy-transport framework and its associated Monge formulation, and proved static-dynamic equivalence for the Wasserstein-Fisher-Rao metric on closed Riemannian manifolds.
The present problem is motivated by Gangbo et al. [15 ] . They introduced a spatially independent additive source with a quadratic penalty into the dynamic transport problem and derived an associated Monge formulation through Lagrangian coordinates. This leads to the energy (1.1) and the endpoint relation (1.2). A generalization allowing spatially dependent sources was studied in [20 ] . Gangbo et al. [15 ] approach the problem through its dynamic formulation and derive an associated Monge formulation in Lagrangian coordinates. We take the static formulation as our starting point and establish existence of minimizers under prescribed geometric constraints. For domain-preserving, uniformly bi-Lipschitz affine flows, we then prove equivalence with the corresponding restriction of their dynamic problem, including equality of minimum values and a correspondence between minimizers.
Outline of the approach
The direct analysis of the static problem requires compactness of minimizing sequences and stability of the nonlinear constraints under weak convergence. The arguments differ according to whether the maps have a prescribed bounded range.
For bounded targets, the main difficulty is to preserve the nonlinear endpoint constraint and establish lower semicontinuity under weak convergence of both maps and sources. At the critical regularity W 1 , n W^{1,n} W 1 , n , distributional continuity of Jacobians does not directly imply convergence of the weighted energy. We combine composition identities with weak continuity of minors to obtain convergence of localized weighted Jacobians. Uniform approximation of the continuous target density also allows passage to the limit in the endpoint constraint. The polynomial dependence of J t [ M ] J_t[M] J t [ M ] on time provides the uniform convergence needed to handle its coupling with weakly convergent sources. Nonnegativity of the intermediate states then permits a cutoff argument yielding global lower semicontinuity, even when the sources change sign. The distortion and Jacobian mass bounds control the gradients, while the bounded range controls the maps themselves. These estimates provide the compactness required by the direct method.
For whole-space targets, the absence of a prescribed bounded range creates two additional difficulties. The weights depending on the maps are no longer uniformly bounded, and gradient bounds alone do not control translations. To address the first issue, local Hölder estimates under the distortion bound yield locally uniform convergence along weakly convergent sequences. This preserves the weighted endpoint constraint and allows the preceding energy argument to be applied locally. To obtain a full Sobolev bound, we use positive initial mass to find a set of uniformly positive measure on which the intermediate states remain bounded below for a short time. The transport energy then controls the displacement on this set. Combined with Poincaré’s inequality and the gradient bound, this controls the mean of the map and establishes coercivity.
For the static-dynamic correspondence, we restrict attention to domain-preserving, uniformly bi-Lipschitz affine flows. The change-of-variables formula reconstructs dynamic densities and velocities from static pairs and preserves the energy. Conversely, testing the continuity equation with functions transported by the inverse flow recovers the static endpoint constraint. Uniform control of the inverse flows, together with weak continuity of minors and the polynomial structure of the interpolation Jacobians, also establishes stability of the reconstruction. The source-penalty limit is studied within this same geometric class. After subtracting the minimum source cost divided by α \alpha α , compactness and lower semicontinuity give the Γ \Gamma Γ -liminf inequality, while constant sequences provide recovery sequences. Comparison with admissible pairs of minimum source cost identifies the limiting transport problem and yields convergence of minimum values and subsequential convergence of minimizers. Convergence of the source norms upgrades their weak convergence to strong convergence in L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) . The reconstruction stability then transfers these limits to the corresponding dynamic paths. When a constant source is admissible, the limiting source is the constant determined by the endpoint mass difference.
Finally, we investigate the existence of admissible pairs with constant source. The Jacobian bounds give a necessary condition for constant-source admissibility. For densities close to a fixed reference pair, a solution operator for the divergence equation and Schauder’s fixed point theorem solve the nonlinear endpoint constraint. This construction produces constant-source pairs satisfying the prescribed distortion and flow bounds, so the existence, equivalence, and convergence results apply without a separate nonemptiness assumption in this local setting.
Main results
Throughout this paper, let n ≥ 2 n \ge2 n ≥ 2 , let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be a bounded open set with Lipschitz boundary, and let 0 ≤ ρ 0 ∈ L 1 ( Ω ) 0 \le\rho_0 \in L^1(\Omega) 0 ≤ ρ 0 ∈ L 1 ( Ω ) . We equip Y : = ( W 1 , n ( Ω ; R n ) ∩ L ∞ ( Ω ; R n ) ) × L 2 ( 0 , 1 ) Y := (W^{1,n}(\Omega;\mathbb{R}^n) \cap L^\infty(\Omega;\mathbb{R}^n)) \times L^2(0,1) Y := ( W 1 , n ( Ω ; R n ) ∩ L ∞ ( Ω ; R n )) × L 2 ( 0 , 1 ) with the topology τ \tau τ . Here ( M j , f j ) → τ ( M , f ) (M_j,f_j) \xrightarrow{\tau} (M,f) ( M j , f j ) τ ( M , f ) means that M j ⇀ M M_j \rightharpoonup M M j ⇀ M in W 1 , n ( Ω ; R n ) W^{1,n}(\Omega;\mathbb{R}^n) W 1 , n ( Ω ; R n ) , M j ⇀ ∗ M M_j \rightharpoonup^* M M j ⇀ ∗ M in L ∞ ( Ω ; R n ) L^\infty(\Omega;\mathbb{R}^n) L ∞ ( Ω ; R n ) , and f j ⇀ f f_j \rightharpoonup f f j ⇀ f in L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) . We equip Y 1 : = W 1 , n ( Ω ; R n ) × L 2 ( 0 , 1 ) Y_1 := W^{1,n}(\Omega;\mathbb{R}^n) \times L^2(0,1) Y 1 := W 1 , n ( Ω ; R n ) × L 2 ( 0 , 1 ) with the product weak topology. We use the following four sets of assumptions and specify the corresponding admissible classes.
(H1) Bounded targets. Assume that Ω \Omega Ω is convex and 0 ≤ ρ 1 ∈ C ( Ω ‾ ) 0 \le\rho_1 \in C(\overline{\Omega}) 0 ≤ ρ 1 ∈ C ( Ω ) . Set
X ( Ω ; Ω ‾ ) : = { M ∈ W 1 , n ( Ω ; R n ) ∩ L ∞ ( Ω ; R n ) : M ( x ) ∈ Ω ‾ a.e. } . X(\Omega;\overline{\Omega}) := \{M \in W^{1,n}(\Omega;\mathbb{R}^n) \cap L^\infty(\Omega;\mathbb{R}^n) : M(x) \in\overline{\Omega}\ \text{a.e.}\}. X ( Ω ; Ω ) := { M ∈ W 1 , n ( Ω ; R n ) ∩ L ∞ ( Ω ; R n ) : M ( x ) ∈ Ω a.e. } .
Let V ρ 1 V^{\rho_1} V ρ 1 consist of pairs ( M , f ) ∈ X ( Ω ; Ω ‾ ) × L 2 ( 0 , 1 ) (M,f) \in X(\Omega;\overline{\Omega}) \times L^2(0,1) ( M , f ) ∈ X ( Ω ; Ω ) × L 2 ( 0 , 1 ) satisfying J t [ M ] ≥ 0 J_t[M] \ge0 J t [ M ] ≥ 0 and R t [ M , f ] ≥ 0 R_t[M,f] \ge0 R t [ M , f ] ≥ 0 a.e. in Ω \Omega Ω and for every fixed t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] , and ρ 1 ( M ) det ∇ M = R 1 [ M , f ] \rho_1(M) \det\nabla M = R_1[M,f] ρ 1 ( M ) det ∇ M = R 1 [ M , f ] in D ′ ( Ω ) \mathcal{D}'(\Omega) D ′ ( Ω ) . For K > 0 K > 0 K > 0 , define
V K ρ 1 : = { ( M , f ) ∈ V ρ 1 : ∣ ∇ M ∣ n ≤ K det ∇ M a.e. in Ω , ∫ Ω det ∇ M d x ≤ ∣ Ω ∣ } . V_K^{\rho_1} := \left\{(M,f) \in V^{\rho_1} :
\begin{array}{l}
|\nabla M|^n \le K\det\nabla M\ \text{a.e. in }\Omega,\\
\int_\Omega\det\nabla M\,dx \le|\Omega|
\end{array}
\right\}. V K ρ 1 := { ( M , f ) ∈ V ρ 1 : ∣∇ M ∣ n ≤ K det ∇ M a.e. in Ω , ∫ Ω det ∇ M d x ≤ ∣Ω∣ } .
(H2) Whole-space targets. Assume that Ω \Omega Ω is connected, 0 ≤ ρ 1 ∈ C ( R n ) ∩ L 1 ( R n ) 0 \le\rho_1 \in C(\mathbb{R}^n) \cap L^1(\mathbb{R}^n) 0 ≤ ρ 1 ∈ C ( R n ) ∩ L 1 ( R n ) , and ∫ Ω ρ 0 d x > 0 \int_\Omega\rho_0\,dx > 0 ∫ Ω ρ 0 d x > 0 .
Let V ~ ρ 1 \widetilde{V}^{\rho_1} V ρ 1 satisfy the same source, nonnegativity, and endpoint constraints as V ρ 1 V^{\rho_1} V ρ 1 , with the map requirement replaced by M ∈ W 1 , n ( Ω ; R n ) M \in W^{1,n}(\Omega;\mathbb{R}^n) M ∈ W 1 , n ( Ω ; R n ) . For K , Λ > 0 K,\Lambda> 0 K , Λ > 0 , define
V K , Λ ρ 1 : = { ( M , f ) ∈ V ~ ρ 1 : ∣ ∇ M ∣ n ≤ K det ∇ M a.e. in Ω , ∫ Ω det ∇ M d x ≤ Λ } . V_{K,\Lambda}^{\rho_1} := \left\{(M,f) \in\widetilde{V}^{\rho_1} :
\begin{array}{l}
|\nabla M|^n \le K\det\nabla M\ \text{a.e. in }\Omega,\\
\int_\Omega\det\nabla M\,dx \le\Lambda
\end{array}
\right\}. V K , Λ ρ 1 := { ( M , f ) ∈ V ρ 1 : ∣∇ M ∣ n ≤ K det ∇ M a.e. in Ω , ∫ Ω det ∇ M d x ≤ Λ } .
The distributional endpoint constraint includes the requirement ρ 1 ( M ) det ∇ M ∈ L l o c 1 ( Ω ) \rho_1(M)\det\nabla M \in L^1_{\mathrm{loc}}(\Omega) ρ 1 ( M ) det ∇ M ∈ L loc 1 ( Ω ) .
(H3) Static-dynamic setting. Assume that Ω \Omega Ω is convex and 0 ≤ ρ 1 ∈ C ( Ω ‾ ) 0 \le\rho_1 \in C(\overline{\Omega}) 0 ≤ ρ 1 ∈ C ( Ω ) . For L ≥ 1 L \ge1 L ≥ 1 , let A L ρ 1 A_L^{\rho_1} A L ρ 1 consist of pairs ( M , f ) ∈ V ρ 1 (M,f) \in V^{\rho_1} ( M , f ) ∈ V ρ 1 with M ∈ C ( Ω ‾ ; Ω ‾ ) M \in C(\overline{\Omega};\overline{\Omega}) M ∈ C ( Ω ; Ω ) satisfying
X t [ M ] ( Ω ‾ ) = Ω ‾ , X_t[M](\overline{\Omega}) = \overline{\Omega}, X t [ M ] ( Ω ) = Ω ,
L − 1 ∣ x − y ∣ ≤ ∣ X t [ M ] ( x ) − X t [ M ] ( y ) ∣ ≤ L ∣ x − y ∣ (3) L^{-1}|x-y| \le|X_t[M](x)-X_t[M](y)| \le L|x-y| \tag*{(3)} L − 1 ∣ x − y ∣ ≤ ∣ X t [ M ] ( x ) − X t [ M ] ( y ) ∣ ≤ L ∣ x − y ∣ (3)
for every x , y ∈ Ω ‾ x,y \in\overline{\Omega} x , y ∈ Ω and t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] .
(H4) Local Hölder setting. Fix K > n n / 2 K > n^{n/2} K > n n /2 , L > 1 L > 1 L > 1 , and 0 < β < γ < 1 0 < \beta< \gamma< 1 0 < β < γ < 1 and assume that Ω \Omega Ω is a bounded convex domain with C 3 , β C^{3,\beta} C 3 , β boundary. Let ρ ‾ ∈ C 1 , γ ( Ω ‾ ) \overline{\rho} \in C^{1,\gamma}(\overline{\Omega}) ρ ∈ C 1 , γ ( Ω ) and a ∈ R a \in\mathbb{R} a ∈ R satisfy ρ ‾ ≥ 0 \overline{\rho} \ge0 ρ ≥ 0 , ρ ‾ − a ≥ 0 \overline{\rho}-a \ge0 ρ − a ≥ 0 , and min Ω ‾ ( ρ ‾ − a / 2 ) > 0 \min_{\overline{\Omega}}(\overline{\rho}-a/2) > 0 min Ω ( ρ − a /2 ) > 0 . The prescribed densities satisfy 0 ≤ ρ 0 , ρ 1 ∈ C 0 , β ( Ω ‾ ) 0 \le\rho_0,\rho_1 \in C^{0,\beta}(\overline{\Omega}) 0 ≤ ρ 0 , ρ 1 ∈ C 0 , β ( Ω ) and ∫ Ω ( ρ 1 − ρ 0 ) d x = a ∣ Ω ∣ \int_\Omega(\rho_1-\rho_0)\,dx = a|\Omega| ∫ Ω ( ρ 1 − ρ 0 ) d x = a ∣Ω∣ . Write d : = ∥ ρ 0 − ρ ‾ + a ∥ C 0 , β ( Ω ‾ ) + ∥ ρ 1 − ρ ‾ ∥ C 0 , β ( Ω ‾ ) d := \|\rho_0-\overline{\rho}+a\|_{C^{0,\beta}(\overline{\Omega})}+\|\rho_1-\overline{\rho}\|_{C^{0,\beta}(\overline{\Omega})} d := ∥ ρ 0 − ρ + a ∥ C 0 , β ( Ω ) + ∥ ρ 1 − ρ ∥ C 0 , β ( Ω ) .
We first establish lower semicontinuity and existence in each of the settings (H1) and (H2).
Theorem 1.1 (Lower semicontinuity). Under (H1), E α E_\alpha E α is sequentially lower semicontinuous on V ρ 1 \mathcal{V}^{\rho_1} V ρ 1 with respect to τ \tau τ for every $\alpha > 0.
Theorem 1.2 (Existence for bounded targets). Under (H1), let K > 0 K > 0 K > 0 and assume that V K ρ 1 ≠ ∅ \mathcal{V}^{\rho_1}_K \ne\varnothing V K ρ 1 = ∅ . Then E α E_\alpha E α attains its minimum on V K ρ 1 \mathcal{V}^{\rho_1}_K V K ρ 1 for every α > 0 \alpha> 0 α > 0 .
Theorem 1.3 (Whole-space targets). Under (H2), let K , Λ > 0 K,\Lambda> 0 K , Λ > 0 . For every α > 0 \alpha> 0 α > 0 , E α E_\alpha E α is sequentially weakly lower semicontinuous on V K , Λ ρ 1 \mathcal{V}^{\rho_1}_{K,\Lambda} V K , Λ ρ 1 . If this class contains a pair ( M ∗ , f ∗ ) (M_\ast,f_\ast) ( M ∗ , f ∗ ) with G ( M ∗ , f ∗ ) < ∞ G(M_\ast,f_\ast) < \infty G ( M ∗ , f ∗ ) < ∞ , then E α E_\alpha E α attains its minimum on V K , Λ ρ 1 \mathcal{V}^{\rho_1}_{K,\Lambda} V K , Λ ρ 1 .
We next state the dynamic correspondence in the bounded-target setting. Let D L ρ 0 , ρ 1 \mathcal{D}^{\rho_0,\rho_1}_L D L ρ 0 , ρ 1 denote the dynamic class defined precisely in § 5.1 . Its triples satisfy 0 ≤ ρ ∈ C ( [ 0 , 1 ] ; L 1 ( Ω ) ) 0 \le\rho\in C([0,1];L^1(\Omega)) 0 ≤ ρ ∈ C ([ 0 , 1 ] ; L 1 ( Ω )) , f ∈ L 2 ( 0 , 1 ) f \in L^2(0,1) f ∈ L 2 ( 0 , 1 ) , the continuity equation ∂ t ρ + div ( ρ v ) = f ( t ) \partial_t\rho+\operatorname{div}(\rho v)=f(t) ∂ t ρ + div ( ρ v ) = f ( t ) , the prescribed endpoint conditions, and the no-flux boundary condition. The velocity has a jointly Lipschitz representative whose flow satisfies X t = ( 1 − t ) Id + t X 1 X_t=(1-t)\operatorname{Id}+tX_1 X t = ( 1 − t ) Id + t X 1 and (3) with M = X 1 M=X_1 M = X 1 . The dynamic energy is
E α ( ρ , v , f ) : = ∫ 0 1 ∫ Ω ρ ( t , y ) ∣ v ( t , y ) ∣ 2 d y d t + 1 α ∥ f ∥ L 2 ( 0 , 1 ) 2 . \mathcal{E}_\alpha(\rho,v,f):=\int_0^1\int_\Omega\rho(t,y)|v(t,y)|^2\,dy\,dt+\frac{1}{\alpha}\|f\|_{L^2(0,1)}^2. E α ( ρ , v , f ) := ∫ 0 1 ∫ Ω ρ ( t , y ) ∣ v ( t , y ) ∣ 2 d y d t + α 1 ∥ f ∥ L 2 ( 0 , 1 ) 2 .
Theorem 1.4 (Static-dynamic equivalence). Under (H3), let L ≥ 1 L\ge1 L ≥ 1 and assume that A L ρ 1 ≠ ∅ \mathcal{A}^{\rho_1}_L\ne\varnothing A L ρ 1 = ∅ . Every ( M , f ) ∈ A L ρ 1 (M,f)\in\mathcal{A}^{\rho_1}_L ( M , f ) ∈ A L ρ 1 determines a triple ( ρ , v , f ) ∈ D L ρ 0 , ρ 1 (\rho,v,f)\in\mathcal{D}^{\rho_0,\rho_1}_L ( ρ , v , f ) ∈ D L ρ 0 , ρ 1 through
ρ ( t , X t [ M ] ( x ) ) = R t [ M , f ] ( x ) J t [ M ] ( x ) , v ( t , X t [ M ] ( x ) ) = M ( x ) − x . (4) \rho(t,X_t[M](x))=\frac{R_t[M,f](x)}{J_t[M](x)},\qquad v(t,X_t[M](x))=M(x)-x. \tag*{(4)} ρ ( t , X t [ M ] ( x )) = J t [ M ] ( x ) R t [ M , f ] ( x ) , v ( t , X t [ M ] ( x )) = M ( x ) − x . (4) Conversely, every triple in D L ρ 0 , ρ 1 \mathcal{D}^{\rho_0,\rho_1}_L D L ρ 0 , ρ 1 arises from these formulas with M = X 1 M=X_1 M = X 1 , where X t X_t X t is its flow. The correspondence preserves the energy: E α ( ρ , v , f ) = E α ( M , f ) \mathcal{E}_\alpha(\rho,v,f)=E_\alpha(M,f) E α ( ρ , v , f ) = E α ( M , f ) . For every α > 0 \alpha>0 α > 0 , both minima are attained and
min ( M , f ) ∈ A L ρ 1 E α ( M , f ) = min ( ρ , v , f ) ∈ D L ρ 0 , ρ 1 E α ( ρ , v , f ) . (5) \min_{(M,f)\in\mathcal{A}^{\rho_1}_L}E_\alpha(M,f)=\min_{(\rho,v,f)\in\mathcal{D}^{\rho_0,\rho_1}_L}\mathcal{E}_\alpha(\rho,v,f). \tag*{(5)} ( M , f ) ∈ A L ρ 1 min E α ( M , f ) = ( ρ , v , f ) ∈ D L ρ 0 , ρ 1 min E α ( ρ , v , f ) . (5) Static and dynamic minimizers correspond under this construction.
We next describe the limit under strong source penalization within the affine-flow class. For A L ρ 1 ≠ ∅ \mathcal{A}^{\rho_1}_L\ne\varnothing A L ρ 1 = ∅ , set
b L : = inf ( M , f ) ∈ A L ρ 1 ∥ f ∥ L 2 ( 0 , 1 ) 2 , S L ρ 1 : = { ( M , f ) ∈ A L ρ 1 : ∥ f ∥ L 2 ( 0 , 1 ) 2 = b L } . b_L:=\inf_{(M,f)\in\mathcal{A}^{\rho_1}_L}\|f\|_{L^2(0,1)}^2,\qquad
\mathcal{S}^{\rho_1}_L:=\left\{(M,f)\in\mathcal{A}^{\rho_1}_L:\|f\|_{L^2(0,1)}^2=b_L\right\}. b L := ( M , f ) ∈ A L ρ 1 inf ∥ f ∥ L 2 ( 0 , 1 ) 2 , S L ρ 1 := { ( M , f ) ∈ A L ρ 1 : ∥ f ∥ L 2 ( 0 , 1 ) 2 = b L } .
Define the extended functionals on Y Y Y by
F ^ α ( M , f ) : = { E α ( M , f ) − b L / α , ( M , f ) ∈ A L ρ 1 , + ∞ , otherwise . \widehat{F}_\alpha(M,f):=
\begin{cases}
E_\alpha(M,f)-b_L/\alpha, & (M,f)\in\mathcal{A}^{\rho_1}_L,\\
+\infty, & \text{otherwise}.
\end{cases} F α ( M , f ) := { E α ( M , f ) − b L / α , + ∞ , ( M , f ) ∈ A L ρ 1 , otherwise .
and
F ^ 0 ( M , f ) : = { G ( M , f ) , ( M , f ) ∈ S L ρ 1 , + ∞ , otherwise . \widehat{F}_0(M,f):=
\begin{cases}
G(M,f), & (M,f)\in\mathcal{S}^{\rho_1}_L,\\
+\infty, & \text{otherwise}.
\end{cases} F 0 ( M , f ) := { G ( M , f ) , + ∞ , ( M , f ) ∈ S L ρ 1 , otherwise .
Theorem 1.5 (Minimum-source-cost limit). Under (H3), let L ≥ 1 L \ge1 L ≥ 1 and assume that A L ρ 1 ≠ ∅ \mathcal{A}^{\rho_{1}}_{L} \ne\varnothing A L ρ 1 = ∅ . The infimum defining b L b_{L} b L is attained, and G G G attains its minimum on S L ρ 1 \mathcal{S}^{\rho_{1}}_{L} S L ρ 1 . Write g L : = min ( M , f ) ∈ S L ρ 1 G ( M , f ) g_{L}:=\min_{(M,f)\in\mathcal{S}^{\rho_{1}}_{L}}G(M,f) g L := min ( M , f ) ∈ S L ρ 1 G ( M , f ) . Then F ^ α → τ F ^ 0 \widehat{F}_{\alpha}\mathop{\to}\limits^{\tau}\widehat{F}_{0} F α → τ F 0 as α → 0 + \alpha\to0^{+} α → 0 + in the sequential sense with respect to τ \tau τ , and
min A L ρ 1 E α = b L α + g L + o ( 1 ) . (6) \min_{\mathcal{A}^{\rho_{1}}_{L}} E_{\alpha}=\frac{b_{L}}{\alpha}+g_{L}+o(1). \tag*{(6)} A L ρ 1 min E α = α b L + g L + o ( 1 ) . (6) For every sequence α j → 0 + \alpha_{j}\to0^{+} α j → 0 + and corresponding minimizers ( M α j , f α j ) (M_{\alpha_{j}},f_{\alpha_{j}}) ( M α j , f α j ) of E α j E_{\alpha_{j}} E α j , there exist a subsequence, not relabeled, and a pair ( M , f ) ∈ S L ρ 1 (M,f)\in\mathcal{S}^{\rho_{1}}_{L} ( M , f ) ∈ S L ρ 1 with G ( M , f ) = g L G(M,f)=g_{L} G ( M , f ) = g L such that
M α j ⇀ M in W 1 , n ( Ω ; R n ) , M α j → M uniformly on Ω ‾ , M_{\alpha_{j}}\rightharpoonup M\quad\text{in }W^{1,n}(\Omega;\mathbb{R}^{n}),\qquad M_{\alpha_{j}}\to M\quad\text{uniformly on }\overline{\Omega}, M α j ⇀ M in W 1 , n ( Ω ; R n ) , M α j → M uniformly on Ω , and
f α j → f strongly in L 2 ( 0 , 1 ) . f_{\alpha_{j}}\to f\quad\text{strongly in }L^{2}(0,1). f α j → f strongly in L 2 ( 0 , 1 ) . Moreover, for any choice of minimizers,
G ( M α , f α ) → g L , ∥ f α ∥ L 2 ( 0 , 1 ) 2 − b L α → 0. (7) G(M_{\alpha},f_{\alpha})\to g_{L},\qquad\frac{\lVert f_{\alpha}\rVert_{L^{2}(0,1)}^{2}-b_{L}}{\alpha}\to0. \tag*{(7)} G ( M α , f α ) → g L , α ∥ f α ∥ L 2 ( 0 , 1 ) 2 − b L → 0. (7)
The following theorem combines the preceding results with the local feasibility construction in Section 6.1. It gives a unified existence, static-dynamic equivalence, and convergence result for Hölder densities near a fixed reference pair, without a separate nonemptiness assumption.
Theorem 1.6 (Existence, equivalence, and convergence for locally feasible densities). Under (H4), there exists ε > 0 \varepsilon>0 ε > 0 , depending only on Ω \Omega Ω , n n n , β \beta β , γ \gamma γ , ρ \rho ρ , a a a , K K K , L L L , such that the following conclusions hold whenever d < ε d<\varepsilon d < ε . There exists a constant-source pair ( M ∗ , a ) ∈ V K ρ 1 ∩ A L ρ 1 (M_{*},a)\in\mathcal{V}^{\rho_{1}}_{K}\cap\mathcal{A}^{\rho_{1}}_{L} ( M ∗ , a ) ∈ V K ρ 1 ∩ A L ρ 1 . For every α > 0 \alpha>0 α > 0 , E α E_{\alpha} E α attains its minimum on both V K ρ 1 \mathcal{V}^{\rho_{1}}_{K} V K ρ 1 and A L ρ 1 \mathcal{A}^{\rho_{1}}_{L} A L ρ 1 . Minimizers on A L ρ 1 \mathcal{A}^{\rho_{1}}_{L} A L ρ 1 correspond to minimizers on D L ρ 0 , ρ 1 \mathcal{D}^{\rho_{0},\rho_{1}}_{L} D L ρ 0 , ρ 1 through the energy-preserving reconstruction. The conclusions of Theorem 1.5 apply with b L = a 2 b_{L}=a^{2} b L = a 2 and g L = m a : = min { G ( M , a ) : ( M , a ) ∈ A L ρ 1 } g_{L}=m_{a}:=\min\{G(M,a):(M,a)\in\mathcal{A}^{\rho_{1}}_{L}\} g L = m a := min { G ( M , a ) : ( M , a ) ∈ A L ρ 1 } . In particular,
min A L ρ 1 E α = min D L ρ 0 , ρ 1 E α = a 2 α + m a + o ( 1 ) as α → 0 + . \min_{\mathcal{A}^{\rho_{1}}_{L}}E_{\alpha}=\min_{\mathcal{D}^{\rho_{0},\rho_{1}}_{L}}\mathcal{E}_{\alpha}=\frac{a^{2}}{\alpha}+m_{a}+o(1)\qquad\text{as }\alpha\to0^{+}. A L ρ 1 min E α = D L ρ 0 , ρ 1 min E α = α a 2 + m a + o ( 1 ) as α → 0 + . For any choice of minimizers ( M α , f α ) (M_{\alpha},f_{\alpha}) ( M α , f α ) on A L ρ 1 \mathcal{A}^{\rho_{1}}_{L} A L ρ 1 , ∥ f α − a ∥ L 2 ( 0 , 1 ) = o ( α ) \lVert f_{\alpha}-a\rVert_{L^{2}(0,1)}=o(\sqrt{\alpha}) ∥ f α − a ∥ L 2 ( 0 , 1 ) = o ( α ) .
Organization of the paper
The paper is organized as follows. Section 2 recalls the weak continuity and variational convergence results used in the proofs. Sections 3 and 4 establish lower semicontinuity and existence for bounded and whole-space targets, respectively. Section 5 proves the static-dynamic correspondence, its stability, and the source-penalty limits. Section 6 gives a local feasibility result for Hölder densities and explicit examples illustrating nonemptiness and the role of constant-source admissibility.
Notation and preliminaries
Notation
Throughout the paper, Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n denotes the bounded open domain, with the geometric assumptions specified in each setting, and ∣ Ω ∣ |\Omega| ∣Ω∣ denotes its n n n -dimensional Lebesgue measure. ∣ A ∣ |A| ∣ A ∣ denotes the Frobenius norm of a matrix A A A , while ∣ x ∣ |x| ∣ x ∣ denotes the Euclidean norm of a vector x x x . Let C c ( Ω ) C_{c}(\Omega) C c ( Ω ) denote the space of continuous real-valued functions on Ω \Omega Ω with compact support. We denote by D ( Ω ) \mathcal{D}(\Omega) D ( Ω ) the space of smooth functions with compact support in Ω \Omega Ω . supp ϕ \operatorname{supp}\phi supp ϕ denotes the support of ϕ \phi ϕ . Let U U U and V V V be open subsets of R n \mathbb{R}^n R n . We write V ⋐ U V \Subset U V ⋐ U , if V ‾ ⊂ U \overline{V} \subset U V ⊂ U and V ‾ \overline{V} V is compact, and say V V V is compactly contained in U U U . χ U \chi_U χ U denotes the characteristic function of U U U , namely,
χ U ( x ) = { 1 , x ∈ U , 0 , x ∈ Ω ∖ U . \chi_U(x) =
\begin{cases}
1, & x \in U,\\
0, & x \in\Omega\setminus U.
\end{cases} χ U ( x ) = { 1 , 0 , x ∈ U , x ∈ Ω ∖ U .
Weak continuity of determinants and minors
Since the formulation of the unbalanced optimal transport problem involves an integral containing the determinant, the convergence of determinants under nonsmooth mappings is essential for establishing lower semicontinuity. We therefore recall the notion of weak continuity of the determinant. Let u = ( u 1 , ⋯ , u n ) ∈ C 2 ( Ω ; R n ) u=(u^1,\cdots,u^n)\in C^2(\Omega;\mathbb{R}^n) u = ( u 1 , ⋯ , u n ) ∈ C 2 ( Ω ; R n ) . We denote
det ∇ u = ∂ ( u 1 , ⋯ , u n ) ∂ ( x 1 , ⋯ , x n ) \det\nabla u=\frac{\partial(u^1,\cdots,u^n)}{\partial(x_1,\cdots,x_n)} det ∇ u = ∂ ( x 1 , ⋯ , x n ) ∂ ( u 1 , ⋯ , u n )
by the classical determinant of the Jacobian matrix. It can also be written as
det ∇ u = ∑ α = 1 n ( − 1 ) α + 1 ∂ u 1 ∂ x α ∂ ( u 2 , ⋯ , u n ) ∂ ( x 1 , ⋯ , x α − 1 , x α + 1 , ⋯ , x n ) . \det\nabla u=\sum_{\alpha=1}^{n}(-1)^{\alpha+1}\frac{\partial u^1}{\partial x_\alpha}\frac{\partial(u^2,\cdots,u^n)}{\partial(x_1,\cdots,x_{\alpha-1},x_{\alpha+1},\cdots,x_n)}. det ∇ u = α = 1 ∑ n ( − 1 ) α + 1 ∂ x α ∂ u 1 ∂ ( x 1 , ⋯ , x α − 1 , x α + 1 , ⋯ , x n ) ∂ ( u 2 , ⋯ , u n ) .
Here ∂ ( u 2 , ⋯ , u n ) / ∂ ( x 1 , ⋯ , x α − 1 , x α + 1 , ⋯ , x n ) \partial(u^2,\cdots,u^n)/\partial(x_1,\cdots,x_{\alpha-1},x_{\alpha+1},\cdots,x_n) ∂ ( u 2 , ⋯ , u n ) / ∂ ( x 1 , ⋯ , x α − 1 , x α + 1 , ⋯ , x n ) denotes the determinant of the corresponding minor matrix. For 2 ≤ s ≤ n 2\leq s\leq n 2 ≤ s ≤ n , we define the generalized adjugate operator adj s ∇ u \operatorname{adj}_s\nabla u adj s ∇ u by
adj s ∇ u = ( ( adj s ∇ u ) α i ) 1 ≤ i ≤ ( n s ) 1 ≤ α ≤ ( n s ) , \operatorname{adj}_s\nabla u=\left((\operatorname{adj}_s\nabla u)^i_\alpha\right)_{\substack{1\leq i\leq\binom{n}{s}\\1\leq\alpha\leq\binom{n}{s}}}, adj s ∇ u = ( ( adj s ∇ u ) α i ) 1 ≤ i ≤ ( s n ) 1 ≤ α ≤ ( s n ) ,
where
( adj s ∇ u ) α i = ( − 1 ) i + α ∂ ( u i 1 , … , u i s ) ∂ ( x α 1 , … , x α s ) . (\operatorname{adj}_s\nabla u)^i_\alpha=(-1)^{i+\alpha}\frac{\partial(u^{i_1},\ldots,u^{i_s})}{\partial(x_{\alpha_1},\ldots,x_{\alpha_s})}. ( adj s ∇ u ) α i = ( − 1 ) i + α ∂ ( x α 1 , … , x α s ) ∂ ( u i 1 , … , u i s ) .
Here ( i 1 , … , i s ) (i_1,\ldots,i_s) ( i 1 , … , i s ) and ( α 1 , … , α s ) (\alpha_1,\ldots,\alpha_s) ( α 1 , … , α s ) are the s s s -tuples corresponding to the indices i i i and α \alpha α , respectively.
Definition 2.1. [7 ] Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be an open set and let f ν , f ∈ L l o c 1 ( Ω ) f_\nu,f\in L^1_{\mathrm{loc}}(\Omega) f ν , f ∈ L loc 1 ( Ω ) . We say that f ν f_\nu f ν converges to f f f in the sense of distributions, and write
f ν ⇀ f in D ′ ( Ω ) , f_\nu\rightharpoonup f \quad\text{in } \mathcal{D}'(\Omega), f ν ⇀ f in D ′ ( Ω ) , if
∫ Ω f ν ( x ) φ ( x ) d x ⟶ ∫ Ω f ( x ) φ ( x ) d x \int_\Omega f_\nu(x)\varphi(x)\,dx\longrightarrow\int_\Omega f(x)\varphi(x)\,dx ∫ Ω f ν ( x ) φ ( x ) d x ⟶ ∫ Ω f ( x ) φ ( x ) d x for every $\varphi\in\mathcal{D}(\Omega).
We now state the main theorem (see [7 ] for details), which establishes the weak continuity of determinants.
Theorem 2.1. Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be a bounded open set, 1 < p < ∞ 1<p<\infty 1 < p < ∞ , and let
u ν ⇀ u in W 1 , p ( Ω ; R n ) . u_\nu\rightharpoonup u \quad\text{in } W^{1,p}(\Omega;\mathbb{R}^n). u ν ⇀ u in W 1 , p ( Ω ; R n ) . (1) If p ≥ n p\geq n p ≥ n , then
det ∇ u ν ⇀ det ∇ u in D ′ ( Ω ) \det\nabla u_\nu\rightharpoonup\det\nabla u \quad\text{in } \mathcal{D}'(\Omega) det ∇ u ν ⇀ det ∇ u in D ′ ( Ω ) and if p > n p>n p > n , then
det ∇ u ν ⇀ det ∇ u in L p / n ( Ω ) . \det\nabla u_\nu\rightharpoonup\det\nabla u \quad\text{in } L^{p/n}(\Omega). det ∇ u ν ⇀ det ∇ u in L p / n ( Ω ) .
(2) Let 2 ≤ s ≤ n 2 \le s \le n 2 ≤ s ≤ n and p ≥ s p \ge s p ≥ s . Then
adj s ∇ u ν ⇀ adj s ∇ u in D ′ ( Ω ; R σ ( s ) ) . \operatorname{adj}_{s}\nabla u_{\nu} \rightharpoonup\operatorname{adj}_{s}\nabla u \quad\text{in } \mathcal{D}'\left(\Omega;\mathbb{R}^{\sigma(s)}\right). adj s ∇ u ν ⇀ adj s ∇ u in D ′ ( Ω ; R σ ( s ) ) .
Furthermore, if p > s p>s p > s , then
adj s ∇ u ν ⇀ adj s ∇ u in L p / s ( Ω ; R σ ( s ) ) . \operatorname{adj}_{s}\nabla u_{\nu} \rightharpoonup\operatorname{adj}_{s}\nabla u \quad\text{in } L^{p/s}\left(\Omega;\mathbb{R}^{\sigma(s)}\right). adj s ∇ u ν ⇀ adj s ∇ u in L p / s ( Ω ; R σ ( s ) ) .
Here σ ( s ) : = ( n s ) 2 \sigma(s):=\binom{n}{s}^{2} σ ( s ) := ( s n ) 2 .
The following lemma extends the weak continuity of Jacobians to continuous weights depending on the maps. It will be used in the lower semicontinuity and closedness arguments below.
Lemma 2.1 (Weak continuity of weighted Jacobians). Let n ≥ 2 n \ge2 n ≥ 2 and let Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n be a bounded open set with Lipschitz boundary. Suppose that v k , v ∈ W 1 , n ( Ω ; R n ) ∩ L ∞ ( Ω ; R n ) v_k,v \in W^{1,n}(\Omega;\mathbb{R}^{n}) \cap L^{\infty}(\Omega;\mathbb{R}^{n}) v k , v ∈ W 1 , n ( Ω ; R n ) ∩ L ∞ ( Ω ; R n ) satisfy
v k ⇀ v in W 1 , n ( Ω ; R n ) , v k ⇀ ∗ v in L ∞ ( Ω ; R n ) . v_k \rightharpoonup v \quad\text{in } W^{1,n}(\Omega;\mathbb{R}^{n}), \qquad v_k \overset{*}{\rightharpoonup} v \quad\text{in } L^{\infty}(\Omega;\mathbb{R}^{n}). v k ⇀ v in W 1 , n ( Ω ; R n ) , v k ⇀ ∗ v in L ∞ ( Ω ; R n ) . Then, for every h ∈ C ( R n ) h \in C(\mathbb{R}^{n}) h ∈ C ( R n ) ,
h ( v k ) det ∇ v k ⇀ h ( v ) det ∇ v in D ′ ( Ω ) . h(v_k)\det\nabla v_k \rightharpoonup h(v)\det\nabla v \quad\text{in } \mathcal{D}'(\Omega). h ( v k ) det ∇ v k ⇀ h ( v ) det ∇ v in D ′ ( Ω ) .
Proof. By compactness of the embedding W 1 , n ( Ω ) ↪ L n ( Ω ) W^{1,n}(\Omega) \hookrightarrow L^{n}(\Omega) W 1 , n ( Ω ) ↪ L n ( Ω ) , we obtain that v k → v v_k \to v v k → v strongly in L n ( Ω ; R n ) L^{n}(\Omega;\mathbb{R}^{n}) L n ( Ω ; R n ) . The weak-* convergence in L ∞ L^{\infty} L ∞ gives a constant R > 0 R>0 R > 0 such that ∣ v k ( x ) ∣ ≤ R |v_k(x)| \le R ∣ v k ( x ) ∣ ≤ R and ∣ v ( x ) ∣ ≤ R |v(x)| \le R ∣ v ( x ) ∣ ≤ R for almost every x ∈ Ω x \in\Omega x ∈ Ω and every k k k . We first consider h ∈ C 1 ( R n ) h \in C^{1}(\mathbb{R}^{n}) h ∈ C 1 ( R n ) . For y = ( y 1 , … , y n ) ∈ R n y=(y_1,\ldots,y_n) \in\mathbb{R}^{n} y = ( y 1 , … , y n ) ∈ R n , define
H h ( y ) : = ∫ 0 y 1 h ( t , y 2 , … , y n ) d t and Ψ h ( y ) : = ( H h ( y ) , y 2 , … , y n ) . H_h(y):=\int_{0}^{y_1} h(t,y_2,\ldots,y_n)\,\mathrm{d}t \quad\text{and} \quad\Psi_h(y):=(H_h(y),y_2,\ldots,y_n). H h ( y ) := ∫ 0 y 1 h ( t , y 2 , … , y n ) d t and Ψ h ( y ) := ( H h ( y ) , y 2 , … , y n ) .
Then Ψ h ∈ C 1 ( R n ; R n ) \Psi_h \in C^{1}(\mathbb{R}^{n};\mathbb{R}^{n}) Ψ h ∈ C 1 ( R n ; R n ) and
D Ψ h ( y ) = ( h ( y ) ∂ y 2 H h ( y ) ⋯ ∂ y n H h ( y ) 0 1 ⋯ 0 ⋮ ⋮ ⋱ ⋮ 0 0 ⋯ 1 ) . D\Psi_h(y)=
\begin{pmatrix}
h(y) & \partial_{y_2}H_h(y) & \cdots& \partial_{y_n}H_h(y)\\
0 & 1 & \cdots& 0\\
\vdots& \vdots& \ddots& \vdots\\
0 & 0 & \cdots& 1
\end{pmatrix}. D Ψ h ( y ) = h ( y ) 0 ⋮ 0 ∂ y 2 H h ( y ) 1 ⋮ 0 ⋯ ⋯ ⋱ ⋯ ∂ y n H h ( y ) 0 ⋮ 1 .
Thus,
det D Ψ h ( y ) = h ( y ) . \det D\Psi_h(y)=h(y). det D Ψ h ( y ) = h ( y ) .
We next verify the weak convergence of the compositions. Since D Ψ h D\Psi_h D Ψ h is continuous,
C h , R : = sup y ∈ B R ‾ ∣ D Ψ h ( y ) ∣ < ∞ . C_{h,R}:=\sup_{y\in\overline{B_R}}|D\Psi_h(y)|<\infty. C h , R := y ∈ B R sup ∣ D Ψ h ( y ) ∣ < ∞.
In particular, Ψ h \Psi_h Ψ h is Lipschitz on B R ‾ \overline{B_R} B R . The Sobolev chain rule gives ∇ ( Ψ h ∘ v k ) = D Ψ h ( v k ) ∇ v k \nabla(\Psi_h\circ v_k)=D\Psi_h(v_k)\nabla v_k ∇ ( Ψ h ∘ v k ) = D Ψ h ( v k ) ∇ v k a.e. in Ω \Omega Ω . The chain rule is applicable because the maps v k v_k v k take values in B R ‾ \overline{B_R} B R , on which the derivative of Ψ h \Psi_h Ψ h is bounded. Equivalently, one may first replace Ψ h \Psi_h Ψ h by a globally Lipschitz C 1 C^1 C 1 map agreeing with it on a neighborhood of B R ‾ \overline{B_R} B R . It follows that
∥ ∇ ( Ψ h ∘ v k ) ∥ L n ( Ω ) ≤ C h , R ∥ ∇ v k ∥ L n ( Ω ) . \|\nabla(\Psi_h\circ v_k)\|_{L^n(\Omega)}\le C_{h,R}\|\nabla v_k\|_{L^n(\Omega)}. ∥∇ ( Ψ h ∘ v k ) ∥ L n ( Ω ) ≤ C h , R ∥∇ v k ∥ L n ( Ω ) .
Also, Ψ h \Psi_h Ψ h is bounded on B R ‾ \overline{B_R} B R , so
∥ Ψ h ∘ v k ∥ L n ( Ω ) ≤ ∣ Ω ∣ 1 / n sup y ∈ B R ‾ ∣ Ψ h ( y ) ∣ . \|\Psi_h\circ v_k\|_{L^n(\Omega)}\le|\Omega|^{1/n}\sup_{y\in\overline{B_R}}|\Psi_h(y)|. ∥ Ψ h ∘ v k ∥ L n ( Ω ) ≤ ∣Ω ∣ 1/ n y ∈ B R sup ∣ Ψ h ( y ) ∣.
Hence { Ψ h ∘ v k } \{\Psi_h \circ v_k\} { Ψ h ∘ v k } is uniformly bounded in W 1 , n ( Ω ; R n ) W^{1,n}(\Omega;\mathbb{R}^n) W 1 , n ( Ω ; R n ) . On the other hand, the Lipschitz continuity of Ψ h \Psi_h Ψ h on B R ‾ \overline{B_R} B R yields
∥ Ψ h ( v k ) − Ψ h ( v ) ∥ L n ( Ω ) ≤ C h , R ∥ v k − v ∥ L n ( Ω ) → 0. \|\Psi_h(v_k)-\Psi_h(v)\|_{L^n(\Omega)} \le C_{h,R}\|v_k-v\|_{L^n(\Omega)} \to0. ∥ Ψ h ( v k ) − Ψ h ( v ) ∥ L n ( Ω ) ≤ C h , R ∥ v k − v ∥ L n ( Ω ) → 0.
Therefore, Ψ h ∘ v k ⇀ Ψ h ∘ v \Psi_h \circ v_k \rightharpoonup\Psi_h \circ v Ψ h ∘ v k ⇀ Ψ h ∘ v in W 1 , n ( Ω ; R n ) W^{1,n}(\Omega;\mathbb{R}^n) W 1 , n ( Ω ; R n ) . Indeed, every subsequence admits a further subsequence converging weakly in W 1 , n W^{1,n} W 1 , n . The strong L n L^n L n convergence identifies every such weak limit with Ψ h ∘ v \Psi_h \circ v Ψ h ∘ v . This proves weak convergence of the whole sequence. Applying Theorem 2.1 to these compositions, we obtain that det ∇ ( Ψ h ∘ v k ) ⇀ det ∇ ( Ψ h ∘ v ) \det\nabla(\Psi_h \circ v_k) \rightharpoonup\det\nabla(\Psi_h \circ v) det ∇ ( Ψ h ∘ v k ) ⇀ det ∇ ( Ψ h ∘ v ) in D ′ ( Ω ) \mathcal{D}'(\Omega) D ′ ( Ω ) . The chain rule and multiplicativity of the determinant yield
det ∇ ( Ψ h ∘ v k ) = det D Ψ h ( v k ) det ∇ v k = h ( v k ) det ∇ v k , \det\nabla(\Psi_h \circ v_k)=\det D\Psi_h(v_k)\det\nabla v_k=h(v_k)\det\nabla v_k, det ∇ ( Ψ h ∘ v k ) = det D Ψ h ( v k ) det ∇ v k = h ( v k ) det ∇ v k ,
and the same identity holds for v v v . This proves the assertion for h ∈ C 1 ( R n ) h \in C^1(\mathbb{R}^n) h ∈ C 1 ( R n ) . Now let h ∈ C ( R n ) h \in C(\mathbb{R}^n) h ∈ C ( R n ) . Choose h m ∈ C ∞ ( R n ) h_m \in C^\infty(\mathbb{R}^n) h m ∈ C ∞ ( R n ) such that
∥ h m − h ∥ C ( B R ‾ ) → 0. \|h_m-h\|_{C(\overline{B_R})} \to0. ∥ h m − h ∥ C ( B R ) → 0.
The uniform W 1 , n W^{1,n} W 1 , n bound gives
C J : = sup k ∥ det ∇ v k ∥ L 1 ( Ω ) + ∥ det ∇ v ∥ L 1 ( Ω ) < ∞ . C_J:=\sup_k\|\det\nabla v_k\|_{L^1(\Omega)}+\|\det\nabla v\|_{L^1(\Omega)}<\infty. C J := k sup ∥ det ∇ v k ∥ L 1 ( Ω ) + ∥ det ∇ v ∥ L 1 ( Ω ) < ∞.
For every φ ∈ C c ∞ ( Ω ) \varphi\in C^\infty_c(\Omega) φ ∈ C c ∞ ( Ω ) ,
∣ ∫ Ω φ [ h ( v k ) det ∇ v k − h ( v ) det ∇ v ] d x ∣ ≤ ∣ ∫ Ω φ [ h m ( v k ) det ∇ v k − h m ( v ) det ∇ v ] d x ∣ + C J ∥ φ ∥ L ∞ ( Ω ) ∥ h m − h ∥ C ( B R ‾ ) . \begin{aligned}
\left|\int_\Omega\varphi\left[h(v_k)\det\nabla v_k-h(v)\det\nabla v\right]\,dx\right|
&\le\left|\int_\Omega\varphi\left[h_m(v_k)\det\nabla v_k-h_m(v)\det\nabla v\right]\,dx\right|\\
&\quad+C_J\|\varphi\|_{L^\infty(\Omega)}\|h_m-h\|_{C(\overline{B_R})}.
\end{aligned} ∫ Ω φ [ h ( v k ) det ∇ v k − h ( v ) det ∇ v ] d x ≤ ∫ Ω φ [ h m ( v k ) det ∇ v k − h m ( v ) det ∇ v ] d x + C J ∥ φ ∥ L ∞ ( Ω ) ∥ h m − h ∥ C ( B R ) .
For fixed m m m , the first term tends to zero as k → ∞ k \to\infty k → ∞ by the preceding argument, and then m → ∞ m \to\infty m → ∞ proves the conclusion. □
-convergence
We recall sequential Γ \Gamma Γ -convergence and its consequences for minimization. These results provide the variational framework for the source-penalty limits studied in § 5.2 .
Definition 2.2. Let Y Y Y be a topological space and let g k , g : Y → R ∪ { + ∞ } g_k,g:Y\to\mathbb{R}\cup\{+\infty\} g k , g : Y → R ∪ { + ∞ } . We say that g k g_k g k sequentially Γ \Gamma Γ -converges to g g g in Y Y Y , and write
g k → Γ g , g_k \xrightarrow{\Gamma} g, g k Γ g , if for every y ∈ Y y\in Y y ∈ Y the following two conditions hold:
(i) For every sequence y k → y y_k\to y y k → y in Y Y Y ,
g ( y ) ≤ lim inf k → ∞ g k ( y k ) . g(y)\le\liminf_{k\to\infty}g_k(y_k). g ( y ) ≤ k → ∞ lim inf g k ( y k ) . (ii) There exists a sequence y k → y y_k\to y y k → y in Y Y Y such that
lim sup k → ∞ g k ( y k ) ≤ g ( y ) . \limsup_{k\to\infty}g_k(y_k)\le g(y). k → ∞ lim sup g k ( y k ) ≤ g ( y ) .
Proposition 2.1. [9 ] Let Y Y Y be a topological space and let g k , g : Y → R ∪ { + ∞ } g_k,g:Y\to\mathbb{R}\cup\{+\infty\} g k , g : Y → R ∪ { + ∞ } . Assume that
(i) g k → Γ g g_k \xrightarrow{\Gamma} g g k Γ g in Y Y Y ;
(ii) the family { g k } \{g_k\} { g k } is equicoercive, in the sense that for every t ∈ R t\in\mathbb{R} t ∈ R there exists a sequentially compact set K t ⊂ Y K_t\subset Y K t ⊂ Y such that
{ g k ≤ t } ⊂ K t for all k . \{g_k\le t\}\subset K_t\qquad\text{for all }k. { g k ≤ t } ⊂ K t for all k .
Then
inf Y g = lim k → ∞ inf Y g k . \inf_{Y} g = \lim_{k \to\infty} \inf_{Y} g_k. Y inf g = k → ∞ lim Y inf g k .
Moreover, if y k y_k y k minimizes g k g_k g k for every k k k , then every subsequential limit of { y k } \{y_k\} { y k } is a minimizer of g g g .
Existence of Minimizers for Bounded Targets
In this section, we prove Theorems 1.1 and 1.2. We retain the assumptions of the bounded-target setting, together with the ambient space Y Y Y and the topology τ \tau τ introduced above. We first establish sequential closedness of the admissible classes and lower semicontinuity of E α E_\alpha E α on V ρ 1 \mathcal{V}^{\rho_1} V ρ 1 . We then obtain minimizers on V K ρ 1 \mathcal{V}^{\rho_1}_K V K ρ 1 using the compactness provided by the distortion and Jacobian mass bounds. The distortion inequality ∣ ∇ M ∣ n ≤ K det ∇ M |\nabla M|^n \le K \det\nabla M ∣∇ M ∣ n ≤ K det ∇ M controls local stretching relative to volume change and restricts anisotropic deformation. For an orientation-preserving bi-Lipschitz map, the change-of-variables formula gives
∫ Ω det ∇ M d x = ∣ M ( Ω ) ∣ . \int_{\Omega} \det\nabla M\,dx = |M(\Omega)|. ∫ Ω det ∇ M d x = ∣ M ( Ω ) ∣.
In particular, this integral equals ∣ Ω ∣ |\Omega| ∣Ω∣ when M M M maps Ω \Omega Ω onto itself. Thus the Jacobian mass bound in V K ρ 1 \mathcal{V}^{\rho_1}_K V K ρ 1 is consistent with such transport maps.
Sequential closedness of the admissible class
We first prove that the admissible classes V ρ 1 \mathcal{V}^{\rho_1} V ρ 1 and V K ρ 1 \mathcal{V}^{\rho_1}_K V K ρ 1 are closed with respect to the weak topology τ \tau τ .
Lemma 3.1. The sets V ρ 1 \mathcal{V}^{\rho_1} V ρ 1 and V K ρ 1 \mathcal{V}^{\rho_1}_K V K ρ 1 are sequentially closed with respect to τ \tau τ .
Proof. Let ( M k , f k ) ∈ V ρ 1 (M_k,f_k) \in\mathcal{V}^{\rho_1} ( M k , f k ) ∈ V ρ 1 satisfy ( M k , f k ) ⟶ τ ( M , f ) (M_k,f_k) \mathrel{\mathop{\longrightarrow}^{\tau}} (M,f) ( M k , f k ) ⟶ τ ( M , f ) . We first verify that ( M , f ) ∈ V ρ 1 (M,f) \in\mathcal{V}^{\rho_1} ( M , f ) ∈ V ρ 1 . Since Ω \Omega Ω is convex, Ω ‾ \overline{\Omega} Ω is a closed convex subset of R n \mathbb{R}^n R n . We claim that
L ∞ ( Ω ; Ω ‾ ) : = { M ∈ L ∞ ( Ω ; R n ) : M ( x ) ∈ Ω ‾ for a.e. x ∈ Ω } L^\infty(\Omega;\overline{\Omega}) := \{M \in L^\infty(\Omega;\mathbb{R}^n): M(x) \in\overline{\Omega}\text{ for a.e. }x \in\Omega\} L ∞ ( Ω ; Ω ) := { M ∈ L ∞ ( Ω ; R n ) : M ( x ) ∈ Ω for a.e. x ∈ Ω }
is weak-∗ * ∗ closed in L ∞ ( Ω ; R n ) L^\infty(\Omega;\mathbb{R}^n) L ∞ ( Ω ; R n ) . Suppose, by contradiction, that E : = { x ∈ Ω : M ( x ) ∉ Ω ‾ } E := \{x \in\Omega: M(x) \notin\overline{\Omega}\} E := { x ∈ Ω : M ( x ) ∈ / Ω } has positive measure. Since Ω ‾ \overline{\Omega} Ω is a closed convex subset of R n \mathbb{R}^n R n , it can be represented as an intersection of countably many closed half-spaces,
Ω ‾ = ⋂ j = 1 ∞ { y ∈ R n : ξ j ⋅ y ≤ c j } \overline{\Omega} = \bigcap_{j=1}^{\infty}\{y \in\mathbb{R}^n : \xi_j \cdot y \le c_j\} Ω = j = 1 ⋂ ∞ { y ∈ R n : ξ j ⋅ y ≤ c j }
for some ξ j ∈ R n \xi_j \in\mathbb{R}^n ξ j ∈ R n and c j ∈ R c_j \in\mathbb{R} c j ∈ R . Hence
E = ⋃ j = 1 ∞ { x ∈ Ω : ξ j ⋅ M ( x ) > c j } . E = \bigcup_{j=1}^{\infty}\{x \in\Omega: \xi_j \cdot M(x) > c_j\}. E = j = 1 ⋃ ∞ { x ∈ Ω : ξ j ⋅ M ( x ) > c j } .
Since ∣ E ∣ > 0 |E| > 0 ∣ E ∣ > 0 , there exist j 0 j_0 j 0 and a measurable set E 0 : = { x ∈ Ω : ξ j 0 ⋅ M ( x ) > c j 0 } E_0 := \{x \in\Omega: \xi_{j_0} \cdot M(x) > c_{j_0}\} E 0 := { x ∈ Ω : ξ j 0 ⋅ M ( x ) > c j 0 } with ∣ E 0 ∣ > 0 |E_0| > 0 ∣ E 0 ∣ > 0 . Set h : = ξ j 0 χ E 0 ∈ L 1 ( Ω ; R n ) h := \xi_{j_0}\chi_{E_0} \in L^1(\Omega;\mathbb{R}^n) h := ξ j 0 χ E 0 ∈ L 1 ( Ω ; R n ) . By the weak-∗ * ∗ convergence of M k M_k M k ,
∫ Ω h ⋅ M k d x ⟶ ∫ Ω h ⋅ M d x . \int_{\Omega} h \cdot M_k\,dx \longrightarrow\int_{\Omega} h \cdot M\,dx. ∫ Ω h ⋅ M k d x ⟶ ∫ Ω h ⋅ M d x .
Since M k ( x ) ∈ Ω ‾ M_k(x) \in\overline{\Omega} M k ( x ) ∈ Ω for a.e. x ∈ Ω x \in\Omega x ∈ Ω , we have ξ j 0 ⋅ M k ( x ) ≤ c j 0 \xi_{j_0} \cdot M_k(x) \le c_{j_0} ξ j 0 ⋅ M k ( x ) ≤ c j 0 for a.e. x ∈ E 0 x \in E_0 x ∈ E 0 . Consequently,
∫ Ω h ⋅ M k d x = ∫ E 0 ξ j 0 ⋅ M k d x ≤ c j 0 ∣ E 0 ∣ \int_{\Omega} h \cdot M_k\,dx = \int_{E_0} \xi_{j_0} \cdot M_k\,dx \le c_{j_0}|E_0| ∫ Ω h ⋅ M k d x = ∫ E 0 ξ j 0 ⋅ M k d x ≤ c j 0 ∣ E 0 ∣
for every k k k . On the other hand,
∫ Ω h ⋅ M d x = ∫ E 0 ξ j 0 ⋅ M d x > c j 0 ∣ E 0 ∣ , \int_{\Omega} h \cdot M\,dx = \int_{E_0} \xi_{j_0} \cdot M\,dx > c_{j_0}|E_0|, ∫ Ω h ⋅ M d x = ∫ E 0 ξ j 0 ⋅ M d x > c j 0 ∣ E 0 ∣ ,
which contradicts the weak-∗ * ∗ convergence. Therefore M ( x ) ∈ Ω ‾ M(x) \in\overline{\Omega} M ( x ) ∈ Ω for a.e. x ∈ Ω x \in\Omega x ∈ Ω . Thus M ∈ L ∞ ( Ω ; Ω ‾ ) M \in L^\infty(\Omega;\overline{\Omega}) M ∈ L ∞ ( Ω ; Ω ) , and M ∈ X ( Ω ; Ω ‾ ) M \in X(\Omega;\overline{\Omega}) M ∈ X ( Ω ; Ω ) . For 0 ≤ j ≤ n 0 \le j \le n 0 ≤ j ≤ n , let σ j ( ∇ M k ) \sigma_j(\nabla M_k) σ j ( ∇ M k ) denote the sum of all principal j × j j \times j j × j minors, with σ 0 ≡ 1 \sigma_0 \equiv1 σ 0 ≡ 1 . Then
det ( s ∇ M k + ( 1 − s ) I ) = ∑ j = 0 n s j ( 1 − s ) n − j σ j ( ∇ M k ) . \det\left(s\nabla M_k+(1-s)I\right)=\sum_{j=0}^{n}s^j(1-s)^{n-j}\sigma_j(\nabla M_k). det ( s ∇ M k + ( 1 − s ) I ) = j = 0 ∑ n s j ( 1 − s ) n − j σ j ( ∇ M k ) .
Fix t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] and φ ∈ C ∞ ( Ω ) \varphi\in C^\infty(\Omega) φ ∈ C ∞ ( Ω ) . By Fubini’s theorem and the finite expansion above,
∫ 0 t f k ( s ) ∫ Ω det ( s ∇ M k + ( 1 − s ) I ) φ ( x ) d x d s = ∑ j = 0 n ( ∫ 0 t f k ( s ) s j ( 1 − s ) n − j d s ) ( ∫ Ω σ j ( ∇ M k ) φ d x ) . \int_0^t f_k(s)\int_\Omega\det\left(s\nabla M_k+(1-s)I\right)\varphi(x)\,dx\,ds=\sum_{j=0}^{n}\left(\int_0^t f_k(s)s^j(1-s)^{n-j}\,ds\right)\left(\int_\Omega\sigma_j(\nabla M_k)\varphi\,dx\right). ∫ 0 t f k ( s ) ∫ Ω det ( s ∇ M k + ( 1 − s ) I ) φ ( x ) d x d s = j = 0 ∑ n ( ∫ 0 t f k ( s ) s j ( 1 − s ) n − j d s ) ( ∫ Ω σ j ( ∇ M k ) φ d x ) .
For every fixed t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] and 0 ≤ j ≤ n 0 \le j \le n 0 ≤ j ≤ n , s j ( 1 − s ) n − j ∈ L 2 ( 0 , 1 ) s^j(1-s)^{n-j} \in L^2(0,1) s j ( 1 − s ) n − j ∈ L 2 ( 0 , 1 ) . Hence the weak convergence of f k f_k f k gives
∫ 0 t f k ( s ) s j ( 1 − s ) n − j d s → ∫ 0 t f ( s ) s j ( 1 − s ) n − j d s . \int_0^t f_k(s)s^j(1-s)^{n-j}\,ds\to\int_0^t f(s)s^j(1-s)^{n-j}\,ds. ∫ 0 t f k ( s ) s j ( 1 − s ) n − j d s → ∫ 0 t f ( s ) s j ( 1 − s ) n − j d s .
For 2 ≤ j ≤ n 2 \le j \le n 2 ≤ j ≤ n , the distributional continuity of minors (Theorem 2.1) yields σ j ( ∇ M k ) ⇀ σ j ( ∇ M ) \sigma_j(\nabla M_k)\rightharpoonup\sigma_j(\nabla M) σ j ( ∇ M k ) ⇀ σ j ( ∇ M ) in D ′ ( Ω ) \mathcal{D}'(\Omega) D ′ ( Ω ) . For j = 1 j=1 j = 1 , this follows directly from the weak convergence of the gradients, while the term j = 0 j=0 j = 0 is independent of k k k . Consequently,
∫ 0 t f k ( s ) ∫ Ω det ( s ∇ M k + ( 1 − s ) I ) φ ( x ) d x d s ⟶ ∫ 0 t f ( s ) ∫ Ω det ( s ∇ M + ( 1 − s ) I ) φ ( x ) d x d s . (8) \int_0^t f_k(s)\int_\Omega\det\left(s\nabla M_k+(1-s)I\right)\varphi(x)\,dx\,ds
\longrightarrow
\int_0^t f(s)\int_\Omega\det\left(s\nabla M+(1-s)I\right)\varphi(x)\,dx\,ds. \tag*{(8)} ∫ 0 t f k ( s ) ∫ Ω det ( s ∇ M k + ( 1 − s ) I ) φ ( x ) d x d s ⟶ ∫ 0 t f ( s ) ∫ Ω det ( s ∇ M + ( 1 − s ) I ) φ ( x ) d x d s . (8)
Hence, for every fixed t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] , R t [ M k , f k ] ⇀ R t [ M , f ] R_t[M_k,f_k]\rightharpoonup R_t[M,f] R t [ M k , f k ] ⇀ R t [ M , f ] in D ′ ( Ω ) \mathcal{D}'(\Omega) D ′ ( Ω ) . For every nonnegative φ ∈ C ∞ ( Ω ) \varphi\in C^\infty(\Omega) φ ∈ C ∞ ( Ω ) , we have
∫ Ω R t [ M , f ] φ d x = lim k → ∞ ∫ Ω R t [ M k , f k ] φ d x ≥ 0. \int_\Omega R_t[M,f]\varphi\,dx=\lim_{k\to\infty}\int_\Omega R_t[M_k,f_k]\varphi\,dx\ge0. ∫ Ω R t [ M , f ] φ d x = k → ∞ lim ∫ Ω R t [ M k , f k ] φ d x ≥ 0.
Since R t [ M , f ] ∈ L 1 ( Ω ) R_t[M,f]\in L^1(\Omega) R t [ M , f ] ∈ L 1 ( Ω ) , it follows that R t [ M , f ] ≥ 0 R_t[M,f]\ge0 R t [ M , f ] ≥ 0 a.e. in Ω \Omega Ω for every fixed t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . It remains to verify that det ( t ∇ M ( x ) + ( 1 − t ) I ) ≥ 0 \det(t\nabla M(x)+(1-t)I)\ge0 det ( t ∇ M ( x ) + ( 1 − t ) I ) ≥ 0 a.e. in Ω \Omega Ω . For each fixed t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , define
g k t ( x ) : = det ( t ∇ M k ( x ) + ( 1 − t ) I ) . g_k^t(x):=\det(t\nabla M_k(x)+(1-t)I). g k t ( x ) := det ( t ∇ M k ( x ) + ( 1 − t ) I ) .
By the definition of V ρ 1 \mathcal{V}^{\rho_1} V ρ 1 , g k t ≥ 0 g_k^t\ge0 g k t ≥ 0 a.e. in Ω \Omega Ω . Furthermore, by Theorem 2.1,
g k t ⇀ g t : = det ( t ∇ M + ( 1 − t ) I ) in D ′ ( Ω ) . g_k^t\rightharpoonup g^t:=\det(t\nabla M+(1-t)I)\quad\text{in }\mathcal{D}'(\Omega). g k t ⇀ g t := det ( t ∇ M + ( 1 − t ) I ) in D ′ ( Ω ) .
Therefore, for every nonnegative φ ∈ D ( Ω ) \varphi\in\mathcal{D}(\Omega) φ ∈ D ( Ω ) ,
∫ Ω g t φ d x = lim k → ∞ ∫ Ω g k t φ d x ≥ 0. \int_\Omega g^t\varphi\,dx=\lim_{k\to\infty}\int_\Omega g_k^t\varphi\,dx\ge0. ∫ Ω g t φ d x = k → ∞ lim ∫ Ω g k t φ d x ≥ 0.
Thus g t g^t g t is a positive distribution. Since g t ∈ L l o c 1 ( Ω ) g^t\in L^1_{\mathrm{loc}}(\Omega) g t ∈ L loc 1 ( Ω ) , we conclude that, for every fixed t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] ,
det ( t ∇ M ( x ) + ( 1 − t ) I ) ≥ 0 for a.e. x ∈ Ω . \det(t\nabla M(x)+(1-t)I)\ge0\quad\text{for a.e. }x\in\Omega. det ( t ∇ M ( x ) + ( 1 − t ) I ) ≥ 0 for a.e. x ∈ Ω.
It remains to prove that ρ 1 ( M k ) det ∇ M k ⇀ ρ 1 ( M ) det ∇ M \rho_1(M_k)\det\nabla M_k\rightharpoonup\rho_1(M)\det\nabla M ρ 1 ( M k ) det ∇ M k ⇀ ρ 1 ( M ) det ∇ M in D ′ ( Ω ) \mathcal{D}'(\Omega) D ′ ( Ω ) . Since ρ 1 ∈ C ( Ω ‾ ) \rho_1\in C(\overline{\Omega}) ρ 1 ∈ C ( Ω ) and Ω ‾ \overline{\Omega} Ω is closed, choose h ∈ C ( R n ) h\in C(\mathbb{R}^n) h ∈ C ( R n ) such that h = ρ 1 h=\rho_1 h = ρ 1 on Ω ‾ \overline{\Omega} Ω . The weighted-Jacobian continuity Lemma 2.1 gives h ( M k ) det ∇ M k ⇀ h ( M ) det ∇ M h(M_k)\det\nabla M_k\rightharpoonup h(M)\det\nabla M h ( M k ) det ∇ M k ⇀ h ( M ) det ∇ M in D ′ ( Ω ) \mathcal{D}'(\Omega) D ′ ( Ω ) . Since M k M_k M k and M M M take values in Ω ‾ \overline{\Omega} Ω almost everywhere, this is precisely
ρ 1 ( M k ) det ∇ M k ⇀ ρ 1 ( M ) det ∇ M in D ′ ( Ω ) . \rho_1(M_k)\det\nabla M_k\rightharpoonup\rho_1(M)\det\nabla M\quad\text{in }\mathcal{D}'(\Omega). ρ 1 ( M k ) det ∇ M k ⇀ ρ 1 ( M ) det ∇ M in D ′ ( Ω ) .
Taking t = 1 t=1 t = 1 in the source-term convergence (8) established above and passing to the limit in the terminal constraint, we obtain
∫ Ω ρ 1 ( M ) det ∇ M φ d x = ∫ Ω ρ 0 φ d x + ∫ 0 1 f ( s ) ∫ Ω det ( s ∇ M + ( 1 − s ) I ) φ d x d s \int_{\Omega} \rho_1(M)\det\nabla M\,\varphi\,dx=\int_{\Omega}\rho_0\varphi\,dx+\int_0^1 f(s)\int_{\Omega}\det\left(s\nabla M+(1-s)I\right)\varphi\,dx\,ds ∫ Ω ρ 1 ( M ) det ∇ M φ d x = ∫ Ω ρ 0 φ d x + ∫ 0 1 f ( s ) ∫ Ω det ( s ∇ M + ( 1 − s ) I ) φ d x d s
for every φ ∈ C c ∞ ( Ω ) \varphi\in C^\infty_c(\Omega) φ ∈ C c ∞ ( Ω ) . These arguments prove that ( M , f ) ∈ V ρ 1 (M,f)\in\mathcal{V}^{\rho_1} ( M , f ) ∈ V ρ 1 and hence establish the sequential closedness of V ρ 1 \mathcal{V}^{\rho_1} V ρ 1 .
Suppose now that ( M k , f k ) ∈ V K ρ 1 (M_k,f_k)\in\mathcal{V}^{\rho_1}_K ( M k , f k ) ∈ V K ρ 1 . It remains to verify the bounded-distortion inequality and the Jacobian mass bound for the limit. First, from ∣ ∇ M k ∣ n ≤ K det ∇ M k |\nabla M_k|^n\le K\det\nabla M_k ∣∇ M k ∣ n ≤ K det ∇ M k we obtain, for every nonnegative φ ∈ C c ∞ ( Ω ) \varphi\in C^\infty_c(\Omega) φ ∈ C c ∞ ( Ω ) ,
∫ Ω φ ∣ ∇ M k ∣ n d x ≤ K ∫ Ω φ det ∇ M k d x \int_{\Omega}\varphi|\nabla M_k|^n\,dx\le K\int_{\Omega}\varphi\det\nabla M_k\,dx ∫ Ω φ ∣∇ M k ∣ n d x ≤ K ∫ Ω φ det ∇ M k d x
By the weak lower semicontinuity of the convex functional M ↦ ∫ Ω φ ∣ ∇ M ∣ n d x M\mapsto\int_{\Omega}\varphi|\nabla M|^n\,dx M ↦ ∫ Ω φ ∣∇ M ∣ n d x , and the distributional convergence det ∇ M k ⇀ det ∇ M \det\nabla M_k\rightharpoonup\det\nabla M det ∇ M k ⇀ det ∇ M in D ′ ( Ω ) \mathcal{D}'(\Omega) D ′ ( Ω ) , we obtain
∫ Ω φ ∣ ∇ M ∣ n d x ≤ K ∫ Ω φ det ∇ M d x \int_{\Omega}\varphi|\nabla M|^n\,dx\le K\int_{\Omega}\varphi\det\nabla M\,dx ∫ Ω φ ∣∇ M ∣ n d x ≤ K ∫ Ω φ det ∇ M d x
Since the above inequality holds for every nonnegative φ ∈ C c ∞ ( Ω ) \varphi\in C^\infty_c(\Omega) φ ∈ C c ∞ ( Ω ) , ∣ ∇ M ∣ n ≤ K det ∇ M |\nabla M|^n\le K\det\nabla M ∣∇ M ∣ n ≤ K det ∇ M a.e. in Ω \Omega Ω . Finally, we verify the Jacobian mass bound. Choose χ m ∈ C c ∞ ( Ω ) \chi_m\in C^\infty_c(\Omega) χ m ∈ C c ∞ ( Ω ) such that 0 ≤ χ m ≤ 1 0\le\chi_m\le1 0 ≤ χ m ≤ 1 , χ m ( x ) → 1 \chi_m(x)\to1 χ m ( x ) → 1 for every x ∈ Ω x\in\Omega x ∈ Ω . For each fixed m m m , the distributional convergence of the determinants gives
∫ Ω χ m det ∇ M d x = lim k → ∞ ∫ Ω χ m det ∇ M k d x \int_{\Omega}\chi_m\det\nabla M\,dx=\lim_{k\to\infty}\int_{\Omega}\chi_m\det\nabla M_k\,dx ∫ Ω χ m det ∇ M d x = k → ∞ lim ∫ Ω χ m det ∇ M k d x
Since det ∇ M k ≥ 0 \det\nabla M_k\ge0 det ∇ M k ≥ 0 and 0 ≤ χ m ≤ 1 0\le\chi_m\le1 0 ≤ χ m ≤ 1 , we obtain
∫ Ω χ m det ∇ M d x ≤ lim inf k → ∞ ∫ Ω det ∇ M k d x ≤ ∣ Ω ∣ . \int_{\Omega}\chi_m\det\nabla M\,dx\le\liminf_{k\to\infty}\int_{\Omega}\det\nabla M_k\,dx\le|\Omega|. ∫ Ω χ m det ∇ M d x ≤ k → ∞ lim inf ∫ Ω det ∇ M k d x ≤ ∣Ω∣.
Letting m → ∞ m\to\infty m → ∞ and applying the dominated convergence theorem, we conclude that
∫ Ω det ∇ M d x ≤ lim inf k → ∞ ∫ Ω det ∇ M k d x ≤ ∣ Ω ∣ . \int_{\Omega}\det\nabla M\,dx\le\liminf_{k\to\infty}\int_{\Omega}\det\nabla M_k\,dx\le|\Omega|. ∫ Ω det ∇ M d x ≤ k → ∞ lim inf ∫ Ω det ∇ M k d x ≤ ∣Ω∣.
Thus ( M , f ) ∈ V K ρ 1 (M,f)\in\mathcal{V}^{\rho_1}_K ( M , f ) ∈ V K ρ 1 , which proves the sequential closedness of this class as well. □
Lower semicontinuity
In this section, we establish the weak lower semicontinuity of the functional E α E_\alpha E α on the admissible class V ρ 1 \mathcal{V}^{\rho_1} V ρ 1 . The main difficulty arises from the determinant term in the energy, which does not fit directly into the standard weak lower semicontinuity theory for quasiconvex functionals [7 ] . We first establish the distributional continuity of determinant-weighted expressions. This provides convergence of the localized transport terms. The nonnegativity of the intermediate states, together with an interior cutoff argument, then yields lower semicontinuity over the whole domain.
The following lemma is the main ingredient in the proof of the weak lower semicontinuity of E α E_\alpha E α .
Lemma 3.2. Let n ≥ 2 n\ge2 n ≥ 2 and let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be a bounded open set with Lipschitz boundary. Suppose that u k , u ∈ W 1 , n ( Ω ; R n ) ∩ L ∞ ( Ω ; R n ) u_k,u\in W^{1,n}(\Omega;\mathbb{R}^n)\cap L^\infty(\Omega;\mathbb{R}^n) u k , u ∈ W 1 , n ( Ω ; R n ) ∩ L ∞ ( Ω ; R n ) satisfy
u k ⇀ u in W 1 , n ( Ω ; R n ) , u k ⇀ ∗ u in L ∞ ( Ω ; R n ) . u_k\rightharpoonup u\quad\text{in }W^{1,n}(\Omega;\mathbb{R}^n),\qquad u_k\rightharpoonup^*u\quad\text{in }L^\infty(\Omega;\mathbb{R}^n). u k ⇀ u in W 1 , n ( Ω ; R n ) , u k ⇀ ∗ u in L ∞ ( Ω ; R n ) . Then, for every fixed s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] ,
∣ u k ∣ 2 det ( I + s ∇ u k ) ⇀ ∣ u ∣ 2 det ( I + s ∇ u ) in D ′ ( Ω ) . |u_k|^2\det(I+s\nabla u_k)\rightharpoonup|u|^2\det(I+s\nabla u)\quad\text{in }\mathcal{D}'(\Omega). ∣ u k ∣ 2 det ( I + s ∇ u k ) ⇀ ∣ u ∣ 2 det ( I + s ∇ u ) in D ′ ( Ω ) .
Proof. By compactness of the embedding W 1 , n ( Ω ) ↪ L n ( Ω ) W^{1,n}(\Omega) \hookrightarrow L^n(\Omega) W 1 , n ( Ω ) ↪ L n ( Ω ) , u k → u u_k \to u u k → u strongly in L n ( Ω ; R n ) L^n(\Omega;\mathbb{R}^n) L n ( Ω ; R n ) . Since n ≥ 2 n \ge2 n ≥ 2 and Ω \Omega Ω is bounded, this also gives strong convergence in L 2 L^2 L 2 . Thus the conclusion holds for s = 0 s=0 s = 0 .
Fix s > 0 s>0 s > 0 and set
v k ( x ) : = x + s u k ( x ) , v ( x ) : = x + s u ( x ) , v_k(x):=x+su_k(x), \qquad v(x):=x+su(x), v k ( x ) := x + s u k ( x ) , v ( x ) := x + s u ( x ) ,
and
g k : = det ∇ v k = det ( I + s ∇ u k ) , g : = det ∇ v = det ( I + s ∇ u ) . g_k:=\det\nabla v_k=\det(I+s\nabla u_k), \qquad g:=\det\nabla v=\det(I+s\nabla u). g k := det ∇ v k = det ( I + s ∇ u k ) , g := det ∇ v = det ( I + s ∇ u ) .
Then
v k ⇀ v in W 1 , n ( Ω ; R n ) , v k ⇀ ∗ v in L ∞ ( Ω ; R n ) . v_k \rightharpoonup v \quad\text{in } W^{1,n}(\Omega;\mathbb{R}^n), \qquad v_k \rightharpoonup^* v \quad\text{in } L^\infty(\Omega;\mathbb{R}^n). v k ⇀ v in W 1 , n ( Ω ; R n ) , v k ⇀ ∗ v in L ∞ ( Ω ; R n ) .
Applying Lemma 2.1 with
h ( y ) = 1 , h ( y ) = y i ( 1 ≤ i ≤ n ) , h ( y ) = ∣ y ∣ 2 , h(y)=1, \qquad h(y)=y_i\ (1\le i\le n), \qquad h(y)=|y|^2, h ( y ) = 1 , h ( y ) = y i ( 1 ≤ i ≤ n ) , h ( y ) = ∣ y ∣ 2 ,
we obtain
g k ⇀ g , v k i g k ⇀ v i g , ∣ v k ∣ 2 g k ⇀ ∣ v ∣ 2 g in D ′ ( Ω ) . g_k \rightharpoonup g, \qquad v_k^i g_k \rightharpoonup v^i g, \qquad|v_k|^2g_k \rightharpoonup|v|^2g \quad\text{in }\mathcal{D}'(\Omega). g k ⇀ g , v k i g k ⇀ v i g , ∣ v k ∣ 2 g k ⇀ ∣ v ∣ 2 g in D ′ ( Ω ) .
For every φ ∈ C c ∞ ( Ω ) \varphi\in C^\infty_c(\Omega) φ ∈ C c ∞ ( Ω ) , the identity u k = ( v k − x ) / s u_k=(v_k-x)/s u k = ( v k − x ) / s gives
∫ Ω φ ∣ u k ∣ 2 g k d x = 1 s 2 ( ∫ Ω φ ∣ v k ∣ 2 g k d x − 2 ∑ i = 1 n ∫ Ω ( x i φ ) v k i g k d x + ∫ Ω ( ∣ x ∣ 2 φ ) g k d x ) . \int_\Omega\varphi|u_k|^2g_k\,dx
=
\frac{1}{s^2}\left(\int_\Omega\varphi|v_k|^2g_k\,dx
-2\sum_{i=1}^n\int_\Omega(x_i\varphi)v_k^i g_k\,dx
+\int_\Omega(|x|^2\varphi)g_k\,dx\right). ∫ Ω φ ∣ u k ∣ 2 g k d x = s 2 1 ( ∫ Ω φ ∣ v k ∣ 2 g k d x − 2 i = 1 ∑ n ∫ Ω ( x i φ ) v k i g k d x + ∫ Ω ( ∣ x ∣ 2 φ ) g k d x ) .
Since φ \varphi φ , x i φ x_i\varphi x i φ , and ∣ x ∣ 2 φ |x|^2\varphi ∣ x ∣ 2 φ belong to C c ∞ ( Ω ) C^\infty_c(\Omega) C c ∞ ( Ω ) , we may pass to the limit in each term. Using u = ( v − x ) / s u=(v-x)/s u = ( v − x ) / s , we conclude that
∫ Ω φ ∣ u k ∣ 2 g k d x → ∫ Ω φ ∣ u ∣ 2 g d x . \int_\Omega\varphi|u_k|^2g_k\,dx \to\int_\Omega\varphi|u|^2g\,dx. ∫ Ω φ ∣ u k ∣ 2 g k d x → ∫ Ω φ ∣ u ∣ 2 g d x .
This proves the assertion. □
Proof of Theorem 1.1. Let { ( M k , f k ) } k = 1 ∞ ⊂ V ρ 1 \{(M_k,f_k)\}_{k=1}^\infty\subset\mathcal{V}^{\rho_1} {( M k , f k ) } k = 1 ∞ ⊂ V ρ 1 satisfy
M k ⇀ M in W 1 , n ( Ω ; R n ) , M k ⇀ ∗ M in L ∞ ( Ω ; R n ) , M_k \rightharpoonup M \quad\text{in } W^{1,n}(\Omega;\mathbb{R}^n), \qquad M_k \rightharpoonup^* M \quad\text{in } L^\infty(\Omega;\mathbb{R}^n), M k ⇀ M in W 1 , n ( Ω ; R n ) , M k ⇀ ∗ M in L ∞ ( Ω ; R n ) ,
and
f k ⇀ f in L 2 ( 0 , 1 ) . f_k \rightharpoonup f \quad\text{in } L^2(0,1). f k ⇀ f in L 2 ( 0 , 1 ) .
By Lemma 3.1, we have ( M , f ) ∈ V ρ 1 (M,f)\in\mathcal{V}^{\rho_1} ( M , f ) ∈ V ρ 1 . We prove that
E α ( M , f ) ≤ lim inf k → ∞ E α ( M k , f k ) . E_\alpha(M,f)\le\liminf_{k\to\infty}E_\alpha(M_k,f_k). E α ( M , f ) ≤ k → ∞ lim inf E α ( M k , f k ) .
If the lower limit is infinite, there is nothing to prove. Otherwise, passing to a subsequence and relabeling, we may assume that E α ( M k , f k ) → lim inf j → ∞ E α ( M j , f j ) < ∞ E_\alpha(M_k,f_k)\to\liminf_{j\to\infty}E_\alpha(M_j,f_j)<\infty E α ( M k , f k ) → lim inf j → ∞ E α ( M j , f j ) < ∞ . Set u k : = M k − x u_k:=M_k-x u k := M k − x and u : = M − x u:=M-x u := M − x . Then u k ⇀ u u_k\rightharpoonup u u k ⇀ u in W 1 , n ( Ω ; R n ) W^{1,n}(\Omega;\mathbb{R}^n) W 1 , n ( Ω ; R n ) and u k ⇀ ∗ u u_k\rightharpoonup^*u u k ⇀ ∗ u in L ∞ ( Ω ; R n ) L^\infty(\Omega;\mathbb{R}^n) L ∞ ( Ω ; R n ) . By compactness of the embedding W 1 , n ( Ω ) ↪ L n ( Ω ) W^{1,n}(\Omega)\hookrightarrow L^n(\Omega) W 1 , n ( Ω ) ↪ L n ( Ω ) , and after passing to a further subsequence, u k → u u_k\to u u k → u strongly in L n ( Ω ; R n ) L^n(\Omega;\mathbb{R}^n) L n ( Ω ; R n ) . The prescribed bounded range also gives
sup k ∥ u k ∥ L ∞ ( Ω ) + ∥ u ∥ L ∞ ( Ω ) ≤ C 2 \sup_k\|u_k\|_{L^\infty(\Omega)}+\|u\|_{L^\infty(\Omega)}\le C_2 k sup ∥ u k ∥ L ∞ ( Ω ) + ∥ u ∥ L ∞ ( Ω ) ≤ C 2
for some constant C 2 > 0 C_2>0 C 2 > 0 . Since R t [ M , f ] ≥ 0 R_t[M,f]\ge0 R t [ M , f ] ≥ 0 , we have G ( M , f ) ≥ 0 G(M,f)\ge0 G ( M , f ) ≥ 0 . In the bounded-range setting, the determinant estimate and f ∈ L 2 ( 0 , 1 ) f\in L^2(0,1) f ∈ L 2 ( 0 , 1 ) justify Fubini’s theorem, which gives
G ( M , f ) = ∫ Ω ∣ u ∣ 2 ρ 0 d x + ∫ 0 1 ( 1 − s ) f ( s ) ∫ Ω ∣ u ∣ 2 det ( I + s ∇ u ) d x d s . G(M,f)=\int_\Omega|u|^2\rho_0\,dx+\int_0^1(1-s)f(s)\int_\Omega|u|^2\det(I+s\nabla u)\,dx\,ds. G ( M , f ) = ∫ Ω ∣ u ∣ 2 ρ 0 d x + ∫ 0 1 ( 1 − s ) f ( s ) ∫ Ω ∣ u ∣ 2 det ( I + s ∇ u ) d x d s .
Consequently,
E α ( M , f ) = G ( M , f ) + 1 α ∥ f ∥ L 2 ( 0 , 1 ) 2 . E_\alpha(M,f)=G(M,f)+\frac{1}{\alpha}\|f\|_{L^2(0,1)}^2. E α ( M , f ) = G ( M , f ) + α 1 ∥ f ∥ L 2 ( 0 , 1 ) 2 .
Choose χ m ∈ C c ∞ ( Ω ) \chi_m\in C^\infty_c(\Omega) χ m ∈ C c ∞ ( Ω ) such that 0 ≤ χ m ≤ χ m + 1 ≤ 1 0\leq\chi_m\leq\chi_{m+1}\leq1 0 ≤ χ m ≤ χ m + 1 ≤ 1 and χ m ( x ) → 1 \chi_m(x)\to1 χ m ( x ) → 1 for every x ∈ Ω x\in\Omega x ∈ Ω . For each m ∈ N m\in\mathbb{N} m ∈ N , define
G m ( M , f ) : = ∫ 0 1 ∫ Ω χ m ∣ u ∣ 2 R t [ M , f ] d x d t . G_m(M,f):=\int_0^1\int_\Omega\chi_m|u|^2R_t[M,f]\,dx\,dt. G m ( M , f ) := ∫ 0 1 ∫ Ω χ m ∣ u ∣ 2 R t [ M , f ] d x d t .
We first show that, for every fixed m m m , G m ( M k , f k ) → G m ( M , f ) G_m(M_k,f_k)\to G_m(M,f) G m ( M k , f k ) → G m ( M , f ) . For s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] , set
Φ k , m ( s ) : = ∫ Ω χ m ∣ u k ∣ 2 det ( I + s ∇ u k ) d x \Phi_{k,m}(s):=\int_\Omega\chi_m|u_k|^2\det(I+s\nabla u_k)\,dx Φ k , m ( s ) := ∫ Ω χ m ∣ u k ∣ 2 det ( I + s ∇ u k ) d x
and
Φ m ( s ) : = ∫ Ω χ m ∣ u ∣ 2 det ( I + s ∇ u ) d x . \Phi_m(s):=\int_\Omega\chi_m|u|^2\det(I+s\nabla u)\,dx. Φ m ( s ) := ∫ Ω χ m ∣ u ∣ 2 det ( I + s ∇ u ) d x .
For every fixed s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] , Lemma 3.2 gives ∣ u k ∣ 2 det ( I + s ∇ u k ) ⇀ ∣ u ∣ 2 det ( I + s ∇ u ) |u_k|^2\det(I+s\nabla u_k)\rightharpoonup|u|^2\det(I+s\nabla u) ∣ u k ∣ 2 det ( I + s ∇ u k ) ⇀ ∣ u ∣ 2 det ( I + s ∇ u ) in D ′ ( Ω ) \mathcal{D}'(\Omega) D ′ ( Ω ) . Testing this convergence against χ m \chi_m χ m , we obtain that Φ k , m ( s ) → Φ m ( s ) \Phi_{k,m}(s)\to\Phi_m(s) Φ k , m ( s ) → Φ m ( s ) for every s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] . We next establish uniform convergence in s s s . The weak convergence implies that sup k ∥ ∇ u k ∥ L n ( Ω ) < ∞ \sup_k\|\nabla u_k\|_{L^n(\Omega)}<\infty sup k ∥∇ u k ∥ L n ( Ω ) < ∞ . Using ∣ det ( I + s A ) ∣ ≤ C n ( 1 + ∣ A ∣ n ) |\det(I+sA)|\leq C_n(1+|A|^n) ∣ det ( I + s A ) ∣ ≤ C n ( 1 + ∣ A ∣ n ) for s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] and A ∈ R n × n A\in\mathbb{R}^{n\times n} A ∈ R n × n , we obtain
sup k sup s ∈ [ 0 , 1 ] ∫ Ω ∣ det ( I + s ∇ u k ) ∣ d x ≤ C 3 . \sup_k\sup_{s\in[0,1]}\int_\Omega|\det(I+s\nabla u_k)|\,dx\leq C_3. k sup s ∈ [ 0 , 1 ] sup ∫ Ω ∣ det ( I + s ∇ u k ) ∣ d x ≤ C 3 .
Since 0 ≤ χ m ≤ 1 0\leq\chi_m\leq1 0 ≤ χ m ≤ 1 and the interpolation Jacobians are nonnegative, we obtain that 0 ≤ Φ k , m ( s ) ≤ C 2 2 C 3 0\leq\Phi_{k,m}(s)\leq C_2^2C_3 0 ≤ Φ k , m ( s ) ≤ C 2 2 C 3 . Moreover, the polynomial structure of the determinant gives ∣ det ( I + s A ) − det ( I + r A ) ∣ ≤ C n ∣ s − r ∣ ( 1 + ∣ A ∣ n ) |\det(I+sA)-\det(I+rA)|\leq C_n|s-r|(1+|A|^n) ∣ det ( I + s A ) − det ( I + r A ) ∣ ≤ C n ∣ s − r ∣ ( 1 + ∣ A ∣ n ) . Therefore,
∣ Φ k , m ( s ) − Φ k , m ( r ) ∣ ≤ C 2 2 C n ∣ s − r ∣ ∫ Ω ( 1 + ∣ ∇ u k ∣ n ) d x ≤ C 5 ∣ s − r ∣ , |\Phi_{k,m}(s)-\Phi_{k,m}(r)|
\leq C_2^2C_n|s-r|\int_\Omega(1+|\nabla u_k|^n)\,dx
\leq C_5|s-r|, ∣ Φ k , m ( s ) − Φ k , m ( r ) ∣ ≤ C 2 2 C n ∣ s − r ∣ ∫ Ω ( 1 + ∣∇ u k ∣ n ) d x ≤ C 5 ∣ s − r ∣ ,
where C 5 C_5 C 5 is independent of k k k and m m m . Thus, for fixed m m m , the sequence { Φ k , m } k \{\Phi_{k,m}\}_k { Φ k , m } k is uniformly bounded and equicontinuous on [ 0 , 1 ] [0,1] [ 0 , 1 ] . By the Arzelà–Ascoli theorem and the pointwise convergence established above, Φ k , m → Φ m \Phi_{k,m}\to\Phi_m Φ k , m → Φ m uniformly on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Since { f k } \{f_k\} { f k } is bounded in L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) ,
∣ ∫ 0 1 ( 1 − s ) f k ( s ) ( Φ k , m ( s ) − Φ m ( s ) ) d s ∣ ≤ ∥ f k ∥ L 2 ( 0 , 1 ) ∥ Φ k , m − Φ m ∥ L 2 ( 0 , 1 ) → 0. \left|\int_0^1(1-s)f_k(s)(\Phi_{k,m}(s)-\Phi_m(s))\,ds\right|
\leq\|f_k\|_{L^2(0,1)}\|\Phi_{k,m}-\Phi_m\|_{L^2(0,1)}\to0. ∫ 0 1 ( 1 − s ) f k ( s ) ( Φ k , m ( s ) − Φ m ( s )) d s ≤ ∥ f k ∥ L 2 ( 0 , 1 ) ∥ Φ k , m − Φ m ∥ L 2 ( 0 , 1 ) → 0.
Also, weak convergence of f k f_k f k gives
∫ 0 1 ( 1 − s ) f k ( s ) Φ m ( s ) d s → ∫ 0 1 ( 1 − s ) f ( s ) Φ m ( s ) d s . \int_0^1(1-s)f_k(s)\Phi_m(s)\,ds\to\int_0^1(1-s)f(s)\Phi_m(s)\,ds. ∫ 0 1 ( 1 − s ) f k ( s ) Φ m ( s ) d s → ∫ 0 1 ( 1 − s ) f ( s ) Φ m ( s ) d s .
Combining these two convergences, we obtain
∫ 0 1 ( 1 − s ) f k ( s ) Φ k , m ( s ) d s → ∫ 0 1 ( 1 − s ) f ( s ) Φ m ( s ) d s . \int_0^1(1-s)f_k(s)\Phi_{k,m}(s)\,ds\to\int_0^1(1-s)f(s)\Phi_m(s)\,ds. ∫ 0 1 ( 1 − s ) f k ( s ) Φ k , m ( s ) d s → ∫ 0 1 ( 1 − s ) f ( s ) Φ m ( s ) d s .
Furthermore, u k → u u_k\to u u k → u almost everywhere and 0 ≤ χ m ∣ u k ∣ 2 ρ 0 ≤ C 2 2 ρ 0 ∈ L 1 ( Ω ) 0\leq\chi_m|u_k|^2\rho_0\leq C_2^2\rho_0\in L^1(\Omega) 0 ≤ χ m ∣ u k ∣ 2 ρ 0 ≤ C 2 2 ρ 0 ∈ L 1 ( Ω ) . The dominated convergence theorem therefore yields
∫ Ω χ m ∣ u k ∣ 2 ρ 0 d x → ∫ Ω χ m ∣ u ∣ 2 ρ 0 d x . \int_\Omega\chi_m|u_k|^2\rho_0\,dx\to\int_\Omega\chi_m|u|^2\rho_0\,dx. ∫ Ω χ m ∣ u k ∣ 2 ρ 0 d x → ∫ Ω χ m ∣ u ∣ 2 ρ 0 d x .
By Fubini’s theorem,
G m ( M k , f k ) = ∫ Ω χ m ∣ u k ∣ 2 ρ 0 d x + ∫ 0 1 ( 1 − s ) f k ( s ) Φ k , m ( s ) d s . G_m(M_k,f_k)=\int_{\Omega}\chi_m|u_k|^2\rho_0\,dx+\int_0^1(1-s)f_k(s)\Phi_{k,m}(s)\,ds. G m ( M k , f k ) = ∫ Ω χ m ∣ u k ∣ 2 ρ 0 d x + ∫ 0 1 ( 1 − s ) f k ( s ) Φ k , m ( s ) d s .
Hence G m ( M k , f k ) → G m ( M , f ) G_m(M_k,f_k)\to G_m(M,f) G m ( M k , f k ) → G m ( M , f ) . We now remove the cutoff. Since R t [ M k , f k ] ≥ 0 R_t[M_k,f_k]\ge0 R t [ M k , f k ] ≥ 0 and 0 ≤ χ m ≤ 1 0\le\chi_m\le1 0 ≤ χ m ≤ 1 , G ( M k , f k ) ≥ G m ( M k , f k ) G(M_k,f_k)\ge G_m(M_k,f_k) G ( M k , f k ) ≥ G m ( M k , f k ) . For every fixed m m m , it follows that
lim inf k → ∞ G ( M k , f k ) ≥ G m ( M , f ) . \liminf_{k\to\infty}G(M_k,f_k)\ge G_m(M,f). k → ∞ lim inf G ( M k , f k ) ≥ G m ( M , f ) .
Since R t [ M , f ] ≥ 0 R_t[M,f]\ge0 R t [ M , f ] ≥ 0 and χ m → 1 \chi_m\to1 χ m → 1 , the monotone convergence theorem gives G m ( M , f ) → G ( M , f ) G_m(M,f)\to G(M,f) G m ( M , f ) → G ( M , f ) . Consequently,
G ( M , f ) ≤ lim inf k → ∞ G ( M k , f k ) . G(M,f)\le\liminf_{k\to\infty}G(M_k,f_k). G ( M , f ) ≤ k → ∞ lim inf G ( M k , f k ) .
Finally, weak lower semicontinuity of the L 2 L^2 L 2 norm yields
1 α ∥ f ∥ L 2 ( 0 , 1 ) 2 ≤ lim inf k → ∞ 1 α ∥ f k ∥ L 2 ( 0 , 1 ) 2 . \frac{1}{\alpha}\|f\|_{L^2(0,1)}^2\le\liminf_{k\to\infty}\frac{1}{\alpha}\|f_k\|_{L^2(0,1)}^2. α 1 ∥ f ∥ L 2 ( 0 , 1 ) 2 ≤ k → ∞ lim inf α 1 ∥ f k ∥ L 2 ( 0 , 1 ) 2 .
Combining the last two inequalities and using the nonnegativity of both terms, we conclude that
E α ( M , f ) ≤ lim inf k → ∞ E α ( M k , f k ) . E_\alpha(M,f)\le\liminf_{k\to\infty}E_\alpha(M_k,f_k). E α ( M , f ) ≤ k → ∞ lim inf E α ( M k , f k ) .
Thus E α E_\alpha E α is sequentially weakly lower semicontinuous on V K ρ 1 \mathcal V^{\rho_1}_K V K ρ 1 .
Existence of minimizers
We now prove the existence of minimizers in V K ρ 1 \mathcal V^{\rho_1}_K V K ρ 1 . The bounded-distortion condition yields the coercive estimate
∥ ∇ M ∥ L n ( Ω ) n ≤ K ∫ Ω det ∇ M d x ≤ K ∣ Ω ∣ \|\nabla M\|_{L^n(\Omega)}^n\le K\int_{\Omega}\det\nabla M\,dx\le K|\Omega| ∥∇ M ∥ L n ( Ω ) n ≤ K ∫ Ω det ∇ M d x ≤ K ∣Ω∣
for every ( M , f ) ∈ V K ρ 1 (M,f)\in\mathcal V^{\rho_1}_K ( M , f ) ∈ V K ρ 1 .
Proof of Theorem 1.2. Let
m : = inf ( M , f ) ∈ V K ρ 1 E α ( M , f ) . m:=\inf_{(M,f)\in\mathcal V^{\rho_1}_K}E_\alpha(M,f). m := ( M , f ) ∈ V K ρ 1 inf E α ( M , f ) .
Since V K ρ 1 ≠ ∅ \mathcal V^{\rho_1}_K\ne\varnothing V K ρ 1 = ∅ , choose ( M 0 , f 0 ) ∈ V K ρ 1 (M_0,f_0)\in\mathcal V^{\rho_1}_K ( M 0 , f 0 ) ∈ V K ρ 1 . The bounded range of M 0 M_0 M 0 , the determinant estimate, and f 0 ∈ L 2 ( 0 , 1 ) f_0\in L^2(0,1) f 0 ∈ L 2 ( 0 , 1 ) ensure that E α ( M 0 , f 0 ) < ∞ E_\alpha(M_0,f_0)<\infty E α ( M 0 , f 0 ) < ∞ . Moreover, the nonnegativity of G ( M , f ) G(M,f) G ( M , f ) gives
E α ( M , f ) = G ( M , f ) + 1 α ∥ f ∥ L 2 ( 0 , 1 ) 2 ≥ 0 E_\alpha(M,f)=G(M,f)+\frac{1}{\alpha}\|f\|_{L^2(0,1)}^2\ge0 E α ( M , f ) = G ( M , f ) + α 1 ∥ f ∥ L 2 ( 0 , 1 ) 2 ≥ 0
for every ( M , f ) ∈ V K ρ 1 (M,f)\in\mathcal V^{\rho_1}_K ( M , f ) ∈ V K ρ 1 . Thus 0 ≤ m ≤ E α ( M 0 , f 0 ) < ∞ 0\le m\le E_\alpha(M_0,f_0)<\infty 0 ≤ m ≤ E α ( M 0 , f 0 ) < ∞ . Choose a minimizing sequence ( M k , f k ) ∈ V K ρ 1 (M_k,f_k)\in\mathcal V^{\rho_1}_K ( M k , f k ) ∈ V K ρ 1 such that E α ( M k , f k ) → m E_\alpha(M_k,f_k)\to m E α ( M k , f k ) → m . We may assume that E α ( M k , f k ) ≤ E α ( M 0 , f 0 ) E_\alpha(M_k,f_k)\le E_\alpha(M_0,f_0) E α ( M k , f k ) ≤ E α ( M 0 , f 0 ) for every k k k . Since G ( M k , f k ) ≥ 0 G(M_k,f_k)\ge0 G ( M k , f k ) ≥ 0 ,
1 α ∥ f k ∥ L 2 ( 0 , 1 ) 2 ≤ E α ( M k , f k ) ≤ E α ( M 0 , f 0 ) , \frac{1}{\alpha}\|f_k\|_{L^2(0,1)}^2\le E_\alpha(M_k,f_k)\le E_\alpha(M_0,f_0), α 1 ∥ f k ∥ L 2 ( 0 , 1 ) 2 ≤ E α ( M k , f k ) ≤ E α ( M 0 , f 0 ) ,
{ f k } \{f_k\} { f k } is bounded in L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) . The bounded-distortion condition and the Jacobian mass bound imply
∥ ∇ M k ∥ L n ( Ω ) n ≤ K ∫ Ω det ∇ M k d x ≤ K ∣ Ω ∣ . \|\nabla M_k\|_{L^n(\Omega)}^n\le K\int_{\Omega}\det\nabla M_k\,dx\le K|\Omega|. ∥∇ M k ∥ L n ( Ω ) n ≤ K ∫ Ω det ∇ M k d x ≤ K ∣Ω∣.
Since M k ( x ) ∈ Ω ‾ M_k(x) \in\overline{\Omega} M k ( x ) ∈ Ω almost everywhere and Ω \Omega Ω is bounded, we also have sup k ∥ M k ∥ L ∞ ( Ω ) < ∞ \sup_k \lVert M_k\rVert_{L^\infty(\Omega)} < \infty sup k ∥ M k ∥ L ∞ ( Ω ) < ∞ . Consequently, { M k } \{M_k\} { M k } is bounded in both W 1 , n ( Ω ; R n ) W^{1,n}(\Omega;\mathbb{R}^n) W 1 , n ( Ω ; R n ) and L ∞ ( Ω ; R n ) L^\infty(\Omega;\mathbb{R}^n) L ∞ ( Ω ; R n ) . By weak compactness and weak-∗ * ∗ sequential compactness, after passing to a subsequence,
M k ⇀ M in W 1 , n ( Ω ; R n ) , M_k \rightharpoonup M \quad\text{in } W^{1,n}(\Omega;\mathbb{R}^n), M k ⇀ M in W 1 , n ( Ω ; R n ) ,
M k ⇀ ∗ M in L ∞ ( Ω ; R n ) , f k ⇀ f in L 2 ( 0 , 1 ) . M_k \rightharpoonup^* M \quad\text{in } L^\infty(\Omega;\mathbb{R}^n), \qquad f_k \rightharpoonup f \quad\text{in } L^2(0,1). M k ⇀ ∗ M in L ∞ ( Ω ; R n ) , f k ⇀ f in L 2 ( 0 , 1 ) .
By Lemma 3.1, ( M , f ) ∈ V K ρ 1 (M,f) \in\mathcal{V}^{\rho_1}_K ( M , f ) ∈ V K ρ 1 . The lower semicontinuity established in Theorem 1.1 gives
E α ( M , f ) ≤ lim inf k → ∞ E α ( M k , f k ) = m . E_\alpha(M,f) \leq\liminf_{k\to\infty} E_\alpha(M_k,f_k) = m. E α ( M , f ) ≤ k → ∞ lim inf E α ( M k , f k ) = m .
Since ( M , f ) (M,f) ( M , f ) is admissible, the definition of m m m gives the reverse inequality. Thus E α ( M , f ) = m E_\alpha(M,f)=m E α ( M , f ) = m , which proves the existence of a minimizer. □
Existence of Minimizers for Whole-Space Targets
We extend the preceding results to nonnegative target densities ρ 1 ∈ C ( R n ) ∩ L 1 ( R n ) \rho_1 \in C(\mathbb{R}^n) \cap L^1(\mathbb{R}^n) ρ 1 ∈ C ( R n ) ∩ L 1 ( R n ) , allowing the admissible maps to take values in R n \mathbb{R}^n R n . Throughout this section, Ω \Omega Ω is a bounded connected Lipschitz domain, 0 ≤ ρ 0 ∈ L 1 ( Ω ) 0 \leq\rho_0 \in L^1(\Omega) 0 ≤ ρ 0 ∈ L 1 ( Ω ) and 0 < ∫ Ω ρ 0 d x 0 < \int_\Omega\rho_0\,dx 0 < ∫ Ω ρ 0 d x . The prescribed bounded range is replaced by local compactness under a fixed bounded-distortion condition. We also replace the Jacobian bound ∣ Ω ∣ |\Omega| ∣Ω∣ by an independent constant Λ > 0 \Lambda>0 Λ > 0 . Together, these constraints give
∥ ∇ M ∥ L n ( Ω ) n ≤ K ∫ Ω det ∇ M d x ≤ K Λ . (9) \lVert\nabla M\rVert_{L^n(\Omega)}^n \leq K \qquad\int_\Omega\det\nabla M\,dx \leq K\Lambda. \tag*{(9)} ∥ ∇ M ∥ L n ( Ω ) n ≤ K ∫ Ω det ∇ M d x ≤ K Λ. (9)
Lemma 4.1. The set V K , Λ ρ 1 \mathcal{V}^{\rho_1}_{K,\Lambda} V K , Λ ρ 1 is sequentially closed with respect to the convergence
M k ⇀ M in W 1 , n ( Ω ; R n ) and f k ⇀ f in L 2 ( 0 , 1 ) . M_k \rightharpoonup M \quad\text{in } W^{1,n}(\Omega;\mathbb{R}^n) \quad\text{and} \quad f_k \rightharpoonup f \quad\text{in } L^2(0,1). M k ⇀ M in W 1 , n ( Ω ; R n ) and f k ⇀ f in L 2 ( 0 , 1 ) . Moreover, if ( M k , f k ) ∈ V K , Λ ρ 1 (M_k,f_k) \in\mathcal{V}^{\rho_1}_{K,\Lambda} ( M k , f k ) ∈ V K , Λ ρ 1 converges weakly to ( M , f ) (M,f) ( M , f ) , after passing to a subsequence,
M k → M locally uniformly in Ω . M_k \to M \quad\text{locally uniformly in } \Omega. M k → M locally uniformly in Ω.
Proof. The sequence { M k } \{M_k\} { M k } is bounded in W 1 , n W^{1,n} W 1 , n and satisfies a fixed bounded-distortion inequality. The quantitative local Hölder estimates in [25 Lemmas 5.1 and 5.2] control the local Hölder seminorms by the L n L^n L n norm of the gradient. Combined with the uniform L n L^n L n bound for the maps themselves, these estimates imply that, for every U ⋐ Ω U \Subset\Omega U ⋐ Ω ,
sup k ∥ M k ∥ C 0 , β ( U ‾ ) < ∞ , \sup_k \lVert M_k\rVert_{C^{0,\beta}(\overline{U})} < \infty, k sup ∥ M k ∥ C 0 , β ( U ) < ∞ ,
where β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) depends only on the fixed distortion bound K K K . Consequently, the Arzelà–Ascoli theorem and a diagonal argument yield a subsequence converging locally uniformly in Ω \Omega Ω . The nonnegativity of det ( t ∇ M + ( 1 − t ) I ) \det(t\nabla M+(1-t)I) det ( t ∇ M + ( 1 − t ) I ) and R t [ M , f ] R_t[M,f] R t [ M , f ] , and the two additional inequalities defining V K , Λ ρ 1 \mathcal{V}^{\rho_1}_{K,\Lambda} V K , Λ ρ 1 follow as in Lemma 3.1, with ∣ Ω ∣ |\Omega| ∣Ω∣ replaced by Λ \Lambda Λ . It only remains to pass to the limit in the weighted Jacobian ρ 1 ( M k ) det ∇ M k \rho_1(M_k)\det\nabla M_k ρ 1 ( M k ) det ∇ M k . Fix φ ∈ C c ∞ ( Ω ) \varphi\in C^\infty_c(\Omega) φ ∈ C c ∞ ( Ω ) . Local uniform convergence of M k M_k M k and the continuity of ρ 1 \rho_1 ρ 1 imply ρ 1 ( M k ) → ρ 1 ( M ) \rho_1(M_k) \to\rho_1(M) ρ 1 ( M k ) → ρ 1 ( M ) uniformly on supp φ \operatorname{supp}\varphi supp φ . On the other hand, the distributional convergence of det ∇ M k \det\nabla M_k det ∇ M k , together with its uniform L 1 L^{1} L 1 bound, implies convergence against every function in C c ( Ω ) C_{c}(\Omega) C c ( Ω ) , by the weak-∗ * ∗ convergence of measure. Consequently,
∫ Ω φ [ ρ 1 ( M k ) det ∇ M k − ρ 1 ( M ) det ∇ M ] d x = ∫ Ω φ [ ρ 1 ( M k ) − ρ 1 ( M ) ] det ∇ M k d x + ∫ Ω φ ρ 1 ( M ) [ det ∇ M k − det ∇ M ] d x → 0. \int_{\Omega}\varphi\left[\rho_{1}(M_{k})\det\nabla M_{k}-\rho_{1}(M)\det\nabla M\right]\,dx
=
\int_{\Omega}\varphi\left[\rho_{1}(M_{k})-\rho_{1}(M)\right]\det\nabla M_{k}\,dx
+\int_{\Omega}\varphi\rho_{1}(M)\left[\det\nabla M_{k}-\det\nabla M\right]\,dx\to0. ∫ Ω φ [ ρ 1 ( M k ) det ∇ M k − ρ 1 ( M ) det ∇ M ] d x = ∫ Ω φ [ ρ 1 ( M k ) − ρ 1 ( M ) ] det ∇ M k d x + ∫ Ω φ ρ 1 ( M ) [ det ∇ M k − det ∇ M ] d x → 0.
Indeed, the first term is controlled by uniform convergence and the bound Λ \Lambda Λ , while the second uses the continuous test function φ ρ 1 ( M ) ∈ C c ( Ω ) \varphi\rho_{1}(M)\in C_{c}(\Omega) φ ρ 1 ( M ) ∈ C c ( Ω ) . The source term passes to the limit in D ′ ( Ω ) D'(\Omega) D ′ ( Ω ) by the same finite expansion in principal minors as in Lemma 3.1. Thus the limiting pair satisfies the constraint and belongs to V K , Λ ρ 1 \mathcal{V}^{\rho_{1}}_{K,\Lambda} V K , Λ ρ 1 .
Proof of Theorem 1.3. Let ( M k , f k ) ∈ V K , Λ ρ 1 (M_{k},f_{k})\in\mathcal{V}^{\rho_{1}}_{K,\Lambda} ( M k , f k ) ∈ V K , Λ ρ 1 converge weakly to ( M , f ) (M,f) ( M , f ) in Y 1 Y_{1} Y 1 . By Lemma 4.1, ( M , f ) ∈ V K , Λ ρ 1 (M,f)\in\mathcal{V}^{\rho_{1}}_{K,\Lambda} ( M , f ) ∈ V K , Λ ρ 1 . We prove that
E α ( M , f ) ≤ lim inf k → ∞ E α ( M k , f k ) . E_{\alpha}(M,f)\leq\liminf_{k\to\infty}E_{\alpha}(M_{k},f_{k}). E α ( M , f ) ≤ k → ∞ lim inf E α ( M k , f k ) .
If the lower limit is infinite, there is nothing to prove. Otherwise, we pass to a subsequence realizing this finite lower limit. By Lemma 4.1, after passing to a further subsequence, we have M k → M M_{k}\to M M k → M locally uniformly in Ω \Omega Ω . Set u k : = M k − x u_{k}:=M_{k}-x u k := M k − x and u : = M − x u:=M-x u := M − x . Recall that
E α ( M , f ) = G ( M , f ) + 1 α ∥ f ∥ L 2 ( 0 , 1 ) 2 , E_{\alpha}(M,f)=G(M,f)+\frac{1}{\alpha}\|f\|_{L^{2}(0,1)}^{2}, E α ( M , f ) = G ( M , f ) + α 1 ∥ f ∥ L 2 ( 0 , 1 ) 2 ,
where
G ( M , f ) = ∫ 0 1 ∫ Ω ∣ u ∣ 2 R t [ M , f ] d x d t G(M,f)=\int_{0}^{1}\int_{\Omega}|u|^{2}R_{t}[M,f]\,dx\,dt G ( M , f ) = ∫ 0 1 ∫ Ω ∣ u ∣ 2 R t [ M , f ] d x d t
is nonnegative. Since the source penalty is weakly lower semicontinuous in L 2 ( 0 , 1 ) L^{2}(0,1) L 2 ( 0 , 1 ) , it remains to treat G G G . Choose χ m ∈ C c ∞ ( Ω ) \chi_{m}\in C_{c}^{\infty}(\Omega) χ m ∈ C c ∞ ( Ω ) such that 0 ≤ χ m ≤ χ m + 1 ≤ 1 0\leq\chi_{m}\leq\chi_{m+1}\leq1 0 ≤ χ m ≤ χ m + 1 ≤ 1 and χ m ( x ) → 1 \chi_{m}(x)\to1 χ m ( x ) → 1 for every x ∈ Ω x\in\Omega x ∈ Ω , and define
G m ( M , f ) : = ∫ 0 1 ∫ Ω χ m ∣ u ∣ 2 R t [ M , f ] d x d t . G_{m}(M,f):=\int_{0}^{1}\int_{\Omega}\chi_{m}|u|^{2}R_{t}[M,f]\,dx\,dt. G m ( M , f ) := ∫ 0 1 ∫ Ω χ m ∣ u ∣ 2 R t [ M , f ] d x d t .
Fix m m m and choose a bounded Lipschitz open set U m U_{m} U m such that supp χ m ⋐ U m ⋐ Ω \operatorname{supp}\chi_{m}\Subset U_{m}\Subset\Omega supp χ m ⋐ U m ⋐ Ω . Then u k ⇀ u u_{k}\rightharpoonup u u k ⇀ u in W 1 , n ( U m ; R n ) W^{1,n}(U_{m};\mathbb{R}^{n}) W 1 , n ( U m ; R n ) , and u k → u u_{k}\to u u k → u uniformly on U m ‾ \overline{U_{m}} U m . In particular,
sup k ∥ u k ∥ L ∞ ( U m ) + ∥ u ∥ L ∞ ( U m ) ≤ C m < ∞ . \sup_{k}\|u_{k}\|_{L^{\infty}(U_{m})}+\|u\|_{L^{\infty}(U_{m})}\leq C_{m}<\infty. k sup ∥ u k ∥ L ∞ ( U m ) + ∥ u ∥ L ∞ ( U m ) ≤ C m < ∞.
Thus Lemma 3.2, applied on U m U_{m} U m , gives
∫ Ω χ m ∣ u k ∣ 2 det ( I + s ∇ u k ) d x → ∫ Ω χ m ∣ u ∣ 2 det ( I + s ∇ u ) d x \int_{\Omega}\chi_{m}|u_{k}|^{2}\det(I+s\nabla u_{k})\,dx\to\int_{\Omega}\chi_{m}|u|^{2}\det(I+s\nabla u)\,dx ∫ Ω χ m ∣ u k ∣ 2 det ( I + s ∇ u k ) d x → ∫ Ω χ m ∣ u ∣ 2 det ( I + s ∇ u ) d x
for every s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] . As in the proof of Theorem 1.1, the polynomial dependence on s s s and the uniform gradient bounds imply that this convergence is uniform on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Moreover, local uniform convergence yields
∫ Ω χ m ∣ u k ∣ 2 ρ 0 d x → ∫ Ω χ m ∣ u ∣ 2 ρ 0 d x . \int_{\Omega}\chi_{m}|u_{k}|^{2}\rho_{0}\,dx\to\int_{\Omega}\chi_{m}|u|^{2}\rho_{0}\,dx. ∫ Ω χ m ∣ u k ∣ 2 ρ 0 d x → ∫ Ω χ m ∣ u ∣ 2 ρ 0 d x .
Using Fubini’s theorem and f k ⇀ f f_{k}\rightharpoonup f f k ⇀ f in L 2 ( 0 , 1 ) L^{2}(0,1) L 2 ( 0 , 1 ) , the same localized argument gives G m ( M k , f k ) → G m ( M , f ) G_{m}(M_{k},f_{k})\to G_{m}(M,f) G m ( M k , f k ) → G m ( M , f ) . Since R t [ M k , f k ] ≥ 0 R_{t}[M_{k},f_{k}]\geq0 R t [ M k , f k ] ≥ 0 and 0 ≤ χ m ≤ 1 0\leq\chi_{m}\leq1 0 ≤ χ m ≤ 1 , we have G ( M k , f k ) ≥ G m ( M k , f k ) G(M_{k},f_{k})\geq G_{m}(M_{k},f_{k}) G ( M k , f k ) ≥ G m ( M k , f k ) . Thus, for every fixed m m m
lim inf k → ∞ G ( M k , f k ) ≥ G m ( M , f ) . \liminf_{k\to\infty}G(M_{k},f_{k})\geq G_{m}(M,f). k → ∞ lim inf G ( M k , f k ) ≥ G m ( M , f ) .
Letting m → ∞ m \to\infty m → ∞ and using R t [ M , f ] ≥ 0 R_t[M,f] \ge0 R t [ M , f ] ≥ 0 , the monotone convergence theorem gives
G ( M , f ) ≤ lim inf k → ∞ G ( M k , f k ) . G(M,f) \le\liminf_{k\to\infty} G(M_k,f_k). G ( M , f ) ≤ k → ∞ lim inf G ( M k , f k ) .
Combining this inequality with the weak lower semicontinuity of the source penalty, we obtain
E α ( M , f ) ≤ lim inf k → ∞ G ( M k , f k ) + lim inf k → ∞ 1 α ∥ f k ∥ L 2 ( 0 , 1 ) ≤ lim inf k → ∞ E α ( M k , f k ) . E_\alpha(M,f) \le\liminf_{k\to\infty} G(M_k,f_k) + \liminf_{k\to\infty} \frac{1}{\alpha}\|f_k\|_{L^2(0,1)} \le\liminf_{k\to\infty} E_\alpha(M_k,f_k). E α ( M , f ) ≤ k → ∞ lim inf G ( M k , f k ) + k → ∞ lim inf α 1 ∥ f k ∥ L 2 ( 0 , 1 ) ≤ k → ∞ lim inf E α ( M k , f k ) .
Since the subsequence was chosen to realize the original lower limit, this proves the sequential weak lower semicontinuity of E α E_\alpha E α on V K , A ρ 1 \mathcal{V}_{K,A}^{\rho_1} V K , A ρ 1 .
We next establish a coercive estimate on energy sublevel sets. Let ( M , f ) (M,f) ( M , f ) be admissible with E α ( M , f ) < ∞ E_\alpha(M,f) < \infty E α ( M , f ) < ∞ . Because ∫ Ω ρ 0 d x > 0 \int_\Omega\rho_0\,dx > 0 ∫ Ω ρ 0 d x > 0 , there exists c > 0 c > 0 c > 0 such that D : = { x ∈ Ω : ρ 0 ( x ) ≥ c } D := \{x \in\Omega: \rho_0(x) \ge c\} D := { x ∈ Ω : ρ 0 ( x ) ≥ c } has positive measure. Choose
Q : = max { 1 , ( 2 K Λ ∣ D ∣ ) 1 / n } . Q := \max\left\{1,\left(\frac{2K\Lambda}{|D|}\right)^{1/n}\right\}. Q := max { 1 , ( ∣ D ∣ 2 K Λ ) 1/ n } .
This choice depends only on the fixed data and is independent of the admissible pair ( M , f ) (M,f) ( M , f ) . For each admissible map M M M , set
D M : = D ∩ { ∣ ∇ M ∣ ≤ Q } . D_M := D \cap\{|\nabla M| \le Q\}. D M := D ∩ { ∣∇ M ∣ ≤ Q } .
Since D ∖ D M = D ∩ { ∣ ∇ M ∣ > Q } D \setminus D_M = D \cap\{|\nabla M| > Q\} D ∖ D M = D ∩ { ∣∇ M ∣ > Q } , the gradient bound gives
Q n ∣ D ∖ D M ∣ ≤ ∫ D ∖ D M ∣ ∇ M ∣ n d x ≤ ∫ Ω ∣ ∇ M ∣ n d x ≤ K Λ . Q^n|D \setminus D_M| \le\int_{D\setminus D_M}|\nabla M|^n\,dx \le\int_\Omega|\nabla M|^n\,dx \le K\Lambda. Q n ∣ D ∖ D M ∣ ≤ ∫ D ∖ D M ∣∇ M ∣ n d x ≤ ∫ Ω ∣∇ M ∣ n d x ≤ K Λ.
By the choice of Q Q Q , it follows that
∣ D M ∣ = ∣ D ∣ − ∣ D ∖ D M ∣ ≥ ∣ D ∣ − K Λ Q n ≥ ∣ D ∣ 2 . |D_M| = |D| - |D \setminus D_M| \ge|D| - \frac{K\Lambda}{Q^n} \ge\frac{|D|}{2}. ∣ D M ∣ = ∣ D ∣ − ∣ D ∖ D M ∣ ≥ ∣ D ∣ − Q n K Λ ≥ 2 ∣ D ∣ .
On D M D_M D M , for every s ∈ [ 0 , 1 ] s \in[0,1] s ∈ [ 0 , 1 ] ,
0 ≤ det ( s ∇ M + ( 1 − s ) I ) ≤ ( 1 + Q ) n = : C Q . 0 \le\det(s\nabla M + (1-s)I) \le(1+Q)^n =: C_Q. 0 ≤ det ( s ∇ M + ( 1 − s ) I ) ≤ ( 1 + Q ) n =: C Q .
In particular, C Q C_Q C Q is independent of M M M and f f f , although the set D M D_M D M depends on M M M . Consequently, R t [ M , f ] ≥ c − C Q t ∥ f ∥ L 2 ( 0 , 1 ) R_t[M,f] \ge c-C_Q\sqrt{t}\|f\|_{L^2(0,1)} R t [ M , f ] ≥ c − C Q t ∥ f ∥ L 2 ( 0 , 1 ) a.e. on D M D_M D M . Set
δ : = min { 1 , c 2 4 C Q 2 ( 1 + ∥ f ∥ L 2 ( 0 , 1 ) 2 ) } . \delta:= \min\left\{1,\frac{c^2}{4C_Q^2(1+\|f\|_{L^2(0,1)}^2)}\right\}. δ := min { 1 , 4 C Q 2 ( 1 + ∥ f ∥ L 2 ( 0 , 1 ) 2 ) c 2 } .
Then, for 0 ≤ t ≤ δ 0 \le t \le\delta 0 ≤ t ≤ δ , R t [ M , f ] ≥ c 2 R_t[M,f] \ge\frac{c}{2} R t [ M , f ] ≥ 2 c a.e. on D M D_M D M . Since the R t [ M , f ] R_t[M,f] R t [ M , f ] are nonnegative,
G ( M , f ) ≥ ∫ 0 δ ∫ D M ∣ M − x ∣ 2 R t [ M , f ] d x d t ≥ c δ 2 ∫ D M ∣ M − x ∣ 2 d x . G(M,f) \ge\int_0^\delta\int_{D_M}|M-x|^2R_t[M,f]\,dx\,dt \ge\frac{c\delta}{2}\int_{D_M}|M-x|^2\,dx. G ( M , f ) ≥ ∫ 0 δ ∫ D M ∣ M − x ∣ 2 R t [ M , f ] d x d t ≥ 2 cδ ∫ D M ∣ M − x ∣ 2 d x .
It follows that
∥ M ∥ L 2 ( D M ) ≤ C ( 1 + ( 1 + ∥ f ∥ L 2 ( 0 , 1 ) ) G ( M , f ) 1 / 2 ) . \|M\|_{L^2(D_M)} \le C\left(1+(1+\|f\|_{L^2(0,1)})G(M,f)^{1/2}\right). ∥ M ∥ L 2 ( D M ) ≤ C ( 1 + ( 1 + ∥ f ∥ L 2 ( 0 , 1 ) ) G ( M , f ) 1/2 ) .
Let
M ‾ : = 1 ∣ Ω ∣ ∫ Ω M d x . \overline{M} := \frac{1}{|\Omega|}\int_\Omega M\,dx. M := ∣Ω∣ 1 ∫ Ω M d x .
Since ∣ D M ∣ ≥ ∣ D ∣ / 2 |D_M| \ge|D|/2 ∣ D M ∣ ≥ ∣ D ∣/2 , Poincaré’s inequality gives
∣ M ‾ ∣ ∣ D M ∣ 1 / 2 ≤ ∥ M ∥ L 2 ( D M ) + ∥ M − M ‾ ∥ L 2 ( D M ) ≤ ∥ M ∥ L 2 ( D M ) + C ∥ ∇ M ∥ L n ( Ω ) . \begin{aligned}
|\overline{M}|\,|D_M|^{1/2} \le\|M\|_{L^2(D_M)}+\|M-\overline{M}\|_{L^2(D_M)} \\
&\le\|M\|_{L^2(D_M)}+C\|\nabla M\|_{L^n(\Omega)}.
\end{aligned} ∣ M ∣ ∣ D M ∣ 1/2 ≤ ∥ M ∥ L 2 ( D M ) + ∥ M − M ∥ L 2 ( D M ) ≤ ∥ M ∥ L 2 ( D M ) + C ∥∇ M ∥ L n ( Ω ) .
Using (4.1) and Poincaré’s inequality again, we obtain
∥ M ∥ W 1 , n ( Ω ) ≤ C ( 1 + ( 1 + ∥ f ∥ L 2 ( 0 , 1 ) ) G ( M , f ) 1 / 2 ) (10) \lVert M\rVert_{W^{1,n}(\Omega)} \le C\left(1+(1+\lVert f\rVert_{L^2(0,1)})G(M,f)^{1/2}\right) \tag*{(10)} ∥ M ∥ W 1 , n ( Ω ) ≤ C ( 1 + ( 1 + ∥ f ∥ L 2 ( 0 , 1 ) ) G ( M , f ) 1/2 ) (10)
where C C C depends only on Ω \Omega Ω , n n n , ρ 0 \rho_0 ρ 0 , K K K , and Λ \Lambda Λ . The finite-energy competitor ensures that the infimum is finite. Thus a minimizing sequence can be chosen with uniformly bounded energy. Since
G ( M k , f k ) ≤ E α ( M k , f k ) , ∥ f k ∥ L 2 ( 0 , 1 ) 2 ≤ α E α ( M k , f k ) , G(M_k,f_k)\le E_\alpha(M_k,f_k),\qquad\lVert f_k\rVert_{L^2(0,1)}^2\le\alpha E_\alpha(M_k,f_k), G ( M k , f k ) ≤ E α ( M k , f k ) , ∥ f k ∥ L 2 ( 0 , 1 ) 2 ≤ α E α ( M k , f k ) ,
estimate (4.2) gives a uniform W 1 , n W^{1,n} W 1 , n bound for M k M_k M k . Weak compactness, the sequential closedness of the admissible class, and the lower semicontinuity proved above yield an admissible minimizer. □
Static-Dynamic Equivalence and Source-Penalty Limits
This section gives a dynamic interpretation of the static problem under prescribed affine-flow conditions. We establish an energy-preserving correspondence and its stability, which transfers the convergence of static minimizers to reconstructed dynamic paths. We then study strong source penalization and identify the role of constant-source feasibility in determining the leading optimal cost.
Static-dynamic equivalence for affine flows
5.1. Static-dynamic equivalence for affine flows. We first prove the static-dynamic correspondence and establish stability of the reconstruction. Throughout this subsection, let n ≥ 2 n\ge2 n ≥ 2 and let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be a bounded convex open set with Lipschitz boundary. Assume that 0 ≤ ρ 0 ∈ L 1 ( Ω ) 0\le\rho_0\in L^1(\Omega) 0 ≤ ρ 0 ∈ L 1 ( Ω ) and 0 ≤ ρ 1 ∈ C ( Ω ‾ ) 0\le\rho_1\in C(\overline{\Omega}) 0 ≤ ρ 1 ∈ C ( Ω ) . We retain the admissible class V ρ 1 \mathcal{V}^{\rho_1} V ρ 1 and the topology τ \tau τ defined above. Write
u : = M − Id , X t [ M ] ( x ) : = x + t u ( x ) , J t [ M ] ( x ) : = det ( I + t ∇ u ( x ) ) , u:=M-\operatorname{Id},\qquad X_t[M](x):=x+tu(x),\qquad J_t[M](x):=\det(I+t\nabla u(x)), u := M − Id , X t [ M ] ( x ) := x + t u ( x ) , J t [ M ] ( x ) := det ( I + t ∇ u ( x )) ,
and recall that
R t [ M , f ] ( x ) : = ρ 0 ( x ) + ∫ 0 t f ( s ) J s [ M ] ( x ) d s , R_t[M,f](x):=\rho_0(x)+\int_0^t f(s)J_s[M](x)\,\mathrm{d}s, R t [ M , f ] ( x ) := ρ 0 ( x ) + ∫ 0 t f ( s ) J s [ M ] ( x ) d s ,
E α ( M , f ) = ∫ 0 1 ∫ Ω ∣ u ( x ) ∣ 2 R t [ M , f ] ( x ) d x d t + 1 α ∫ 0 1 ∣ f ( t ) ∣ 2 d t . E_\alpha(M,f)=\int_0^1\int_\Omega|u(x)|^2R_t[M,f](x)\,\mathrm{d}x\,\mathrm{d}t+\frac{1}{\alpha}\int_0^1|f(t)|^2\,\mathrm{d}t. E α ( M , f ) = ∫ 0 1 ∫ Ω ∣ u ( x ) ∣ 2 R t [ M , f ] ( x ) d x d t + α 1 ∫ 0 1 ∣ f ( t ) ∣ 2 d t .
Lemma 5.1. The set A L ρ 1 \mathcal{A}^{\rho_1}_L A L ρ 1 is sequentially closed with respect to τ \tau τ . For every ( M , f ) ∈ A L ρ 1 (M,f)\in\mathcal{A}^{\rho_1}_L ( M , f ) ∈ A L ρ 1 and every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , X t [ M ] X_t[M] X t [ M ] is an orientation-preserving bi-Lipschitz homeomorphism of Ω \Omega Ω onto itself. Moreover,
L − n ≤ J t [ M ] ≤ L n for all t ∈ [ 0 , 1 ] , (11) L^{-n}\le J_t[M]\le L^n\qquad\text{for all }t\in[0,1], \tag*{(11)} L − n ≤ J t [ M ] ≤ L n for all t ∈ [ 0 , 1 ] , (11) and
∫ Ω J t [ M ] d x = ∣ Ω ∣ for all t ∈ [ 0 , 1 ] . (12) \int_\Omega J_t[M]\,\mathrm{d}x=|\Omega|\qquad\text{for all }t\in[0,1]. \tag*{(12)} ∫ Ω J t [ M ] d x = ∣Ω∣ for all t ∈ [ 0 , 1 ] . (12) Using the Frobenius norm for matrices, one also has
∣ ∇ X t [ M ] ∣ n ≤ n n / 2 L 2 n J t [ M ] a.e. in Ω . (13) |\nabla X_t[M]|^n\le n^{n/2}L^{2n}J_t[M]\qquad\text{a.e. in }\Omega. \tag*{(13)} ∣∇ X t [ M ] ∣ n ≤ n n /2 L 2 n J t [ M ] a.e. in Ω. (13) In particular,
∣ ∇ M ∣ n ≤ n n / 2 L 2 n det ∇ M a.e. in Ω and ∫ Ω det ∇ M d x = ∣ Ω ∣ . |\nabla M|^n\le n^{n/2}L^{2n}\det\nabla M\quad\text{a.e. in }\Omega
\qquad\text{and}\qquad
\int_\Omega\det\nabla M\,\mathrm{d}x=|\Omega|. ∣∇ M ∣ n ≤ n n /2 L 2 n det ∇ M a.e. in Ω and ∫ Ω det ∇ M d x = ∣Ω∣.
Proof. Let ( M j , f j ) ∈ A L ρ 1 (M_j,f_j)\in\mathcal{A}^{\rho_1}_L ( M j , f j ) ∈ A L ρ 1 converge to ( M , f ) (M,f) ( M , f ) with respect to τ \tau τ . The sequential closedness of V ρ 1 \mathcal{V}^{\rho_1} V ρ 1 established by Lemma 3.1 gives ( M , f ) ∈ V ρ 1 (M,f)\in\mathcal{V}^{\rho_1} ( M , f ) ∈ V ρ 1 . Taking t = 1 t=1 t = 1 in (3) , we see that the M j M_j M j are uniformly L L L -Lipschitz and take values in the compact set Ω ‾ \overline{\Omega} Ω . By the Arzelà–Ascoli theorem, a subsequence converges uniformly on Ω ‾ \overline{\Omega} Ω . Its uniform limit agrees almost everywhere with the weak limit M M M , providing a continuous representative of M M M . Consequently,
X t [ M j ] → X t [ M ] uniformly on [ 0 , 1 ] × Ω ‾ . X_t[M_j]\to X_t[M]\quad\text{uniformly on }[0,1]\times\overline{\Omega}. X t [ M j ] → X t [ M ] uniformly on [ 0 , 1 ] × Ω .
To verify surjectivity, fix t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] and y ∈ Ω ‾ y\in\overline{\Omega} y ∈ Ω . Choose x j ∈ Ω ‾ x_j\in\overline{\Omega} x j ∈ Ω with X t [ M j ] ( x j ) = y X_t[M_j](x_j)=y X t [ M j ] ( x j ) = y . After taking a further subsequence, x j → x ∈ Ω ‾ x_j\to x\in\overline{\Omega} x j → x ∈ Ω , and uniform convergence yields X t [ M ] ( x ) = y X_t[M](x)=y X t [ M ] ( x ) = y . Thus ( M , f ) ∈ A L ρ 1 (M,f)\in\mathcal{A}^{\rho_1}_L ( M , f ) ∈ A L ρ 1 . For any map satisfying (3) , each X t [ M ] X_t[M] X t [ M ] is a bi-Lipschitz homeomorphism of Ω ‾ \overline{\Omega} Ω onto itself. Such a homeomorphism preserves the interior and the boundary, so X t [ M ] ( Ω ) = Ω X_t[M](\Omega)=\Omega X t [ M ] ( Ω ) = Ω . By Rademacher’s theorem, M M M is differentiable outside a single null set. At every such point, the distance bounds imply that, for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , all singular values of I + t ∇ u I+t\nabla u I + t ∇ u lie in [ L − 1 , L ] [L^{-1},L] [ L − 1 , L ] . Consequently,
L − n ≤ ∣ det ( I + t ∇ u ) ∣ ≤ L n . L^{-n}\leq\lvert\det(I+t\nabla u)\rvert\leq L^n. L − n ≤ ∣ det ( I + t ∇ u )∣ ≤ L n .
The determinant is continuous in t t t , never vanishes, and equals one at t = 0 t=0 t = 0 . It therefore remains positive on [ 0 , 1 ] [0,1] [ 0 , 1 ] . There exists a null set N ⊂ Ω N\subset\Omega N ⊂ Ω such that L − n ≤ J t [ M ] ( x ) ≤ L n L^{-n}\leq J_t[M](x)\leq L^n L − n ≤ J t [ M ] ( x ) ≤ L n for every x ∈ Ω ∖ N x\in\Omega\setminus N x ∈ Ω ∖ N and every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . This also proves orientation preservation. The change-of-variables formula then yields
∫ Ω J t [ M ] d x = ∣ X t [ M ] ( Ω ) ∣ = ∣ Ω ∣ , \int_{\Omega}J_t[M]\,dx=\lvert X_t[M](\Omega)\rvert=\lvert\Omega\rvert, ∫ Ω J t [ M ] d x = ∣ X t [ M ] ( Ω )∣ = ∣ Ω ∣ ,
which proves (12) . The singular-value upper bound also gives ∣ ∇ X t [ M ] ∣ ≤ n L \lvert\nabla X_t[M]\rvert\leq\sqrt{n}L ∣ ∇ X t [ M ]∣ ≤ n L . Combining this with J t [ M ] ≥ L − n J_t[M]\geq L^{-n} J t [ M ] ≥ L − n yields
∣ ∇ X t [ M ] ∣ n ≤ n n / 2 L n ≤ n n / 2 L 2 n J t [ M ] . \lvert\nabla X_t[M]\rvert^n\leq n^{n/2}L^n\leq n^{n/2}L^{2n}J_t[M]. ∣ ∇ X t [ M ] ∣ n ≤ n n /2 L n ≤ n n /2 L 2 n J t [ M ] .
Taking t = 1 t=1 t = 1 proves the assertions concerning M M M . □
Writing Y t : = X t [ M ] − 1 Y_t:=X_t[M]^{-1} Y t := X t [ M ] − 1 and D : = diam ( Ω ) D:=\operatorname{diam}(\Omega) D := diam ( Ω ) , we have
∣ Y t ( y ) − Y s ( z ) ∣ ≤ L ∣ y − z ∣ + L D ∣ t − s ∣ . (14) \lvert Y_t(y)-Y_s(z)\rvert\leq L\lvert y-z\rvert+LD\lvert t-s\rvert. \tag*{(14)} ∣ Y t ( y ) − Y s ( z )∣ ≤ L ∣ y − z ∣ + L D ∣ t − s ∣ . (14)
Indeed, the spatial estimate follows from (3) , while
∣ Y t ( y ) − Y s ( y ) ∣ ≤ L ∣ X t [ M ] ( Y s ( y ) ) − X s [ M ] ( Y s ( y ) ) ∣ ≤ L D ∣ t − s ∣ . \lvert Y_t(y)-Y_s(y)\rvert\leq L\lvert X_t[M](Y_s(y))-X_s[M](Y_s(y))\rvert\leq LD\lvert t-s\rvert. ∣ Y t ( y ) − Y s ( y )∣ ≤ L ∣ X t [ M ] ( Y s ( y )) − X s [ M ] ( Y s ( y ))∣ ≤ L D ∣ t − s ∣ .
In particular, the inverse flow is jointly Lipschitz. Let D ρ 0 , ρ 1 \mathcal{D}^{\rho_0,\rho_1} D ρ 0 , ρ 1 consist of triples ( ρ , v , f ) (\rho,v,f) ( ρ , v , f ) such that
0 ≤ ρ ∈ C ( [ 0 , 1 ] ; L 1 ( Ω ) ) , v ∈ L ∞ ( ( 0 , 1 ) × Ω ; R n ) , f ∈ L 2 ( 0 , 1 ) , 0\leq\rho\in C([0,1];L^1(\Omega)),\qquad v\in L^\infty((0,1)\times\Omega;\mathbb{R}^n),\qquad f\in L^2(0,1), 0 ≤ ρ ∈ C ([ 0 , 1 ] ; L 1 ( Ω )) , v ∈ L ∞ (( 0 , 1 ) × Ω ; R n ) , f ∈ L 2 ( 0 , 1 ) ,
and
∫ Ω ρ 1 ( y ) φ ( 1 , y ) d y − ∫ Ω ρ 0 ( y ) φ ( 0 , y ) d y = ∫ 0 1 ∫ Ω ( ρ ∂ t φ + ρ v ⋅ ∇ φ + f ( t ) φ ) d y d t (15) \int_{\Omega}\rho_1(y)\varphi(1,y)\,dy-\int_{\Omega}\rho_0(y)\varphi(0,y)\,dy
=
\int_0^1\int_{\Omega}\bigl(\rho\,\partial_t\varphi+\rho v\cdot\nabla\varphi+f(t)\varphi\bigr)\,dy\,dt \tag*{(15)} ∫ Ω ρ 1 ( y ) φ ( 1 , y ) d y − ∫ Ω ρ 0 ( y ) φ ( 0 , y ) d y = ∫ 0 1 ∫ Ω ( ρ ∂ t φ + ρ v ⋅ ∇ φ + f ( t ) φ ) d y d t (15)
for every φ ∈ C 1 ( [ 0 , 1 ] × Ω ‾ ) \varphi\in C^1([0,1]\times\overline{\Omega}) φ ∈ C 1 ([ 0 , 1 ] × Ω ) . This identity expresses the continuity equation, the prescribed endpoint conditions, and the no-flux boundary condition. Define
E α ( ρ , v , f ) : = ∫ 0 1 ∫ Ω ρ ( t , y ) ∣ v ( t , y ) ∣ 2 d y d t + 1 α ∫ 0 1 ∣ f ( t ) ∣ 2 d t . \mathcal{E}_{\alpha}(\rho,v,f):=\int_0^1\int_{\Omega}\rho(t,y)\lvert v(t,y)\rvert^2\,dy\,dt+\frac{1}{\alpha}\int_0^1\lvert f(t)\rvert^2\,dt. E α ( ρ , v , f ) := ∫ 0 1 ∫ Ω ρ ( t , y ) ∣ v ( t , y ) ∣ 2 d y d t + α 1 ∫ 0 1 ∣ f ( t ) ∣ 2 d t .
The dynamic problem considered here is posed on D ρ 0 , ρ 1 \mathcal{D}^{\rho_0,\rho_1} D ρ 0 , ρ 1 , with the stated bounded-velocity and strong L 1 L^1 L 1 continuity requirements. The correspondence below concerns the subclass whose velocities generate uniformly bi-Lipschitz affine flows. Let D L ρ 0 , ρ 1 \mathcal{D}_{L}^{\rho_0,\rho_1} D L ρ 0 , ρ 1 consist of triples in D ρ 0 , ρ 1 \mathcal{D}^{\rho_0,\rho_1} D ρ 0 , ρ 1 for which v v v has a jointly Lipschitz representative on [ 0 , 1 ] × Ω ‾ [0,1] \times\overline{\Omega} [ 0 , 1 ] × Ω whose flow X t X_t X t satisfies
d d t X t ( x ) = v ( t , X t ( x ) ) , X 0 ( x ) = x , \frac{d}{dt}X_t(x)=v(t,X_t(x)), \qquad X_0(x)=x, d t d X t ( x ) = v ( t , X t ( x )) , X 0 ( x ) = x ,
together with
X t ( x ) = ( 1 − t ) x + t X 1 ( x ) (16) X_t(x)=(1-t)x+tX_1(x) \tag*{(16)} X t ( x ) = ( 1 − t ) x + t X 1 ( x ) (16)
and
X t ( Ω ‾ ) = Ω ‾ , L − 1 ∣ x − y ∣ ≤ ∣ X t ( x ) − X t ( y ) ∣ ≤ L ∣ x − y ∣ X_t(\overline{\Omega})=\overline{\Omega}, \qquad L^{-1}|x-y|\leq|X_t(x)-X_t(y)|\leq L|x-y| X t ( Ω ) = Ω , L − 1 ∣ x − y ∣ ≤ ∣ X t ( x ) − X t ( y ) ∣ ≤ L ∣ x − y ∣
for all x , y ∈ Ω ‾ x,y\in\overline{\Omega} x , y ∈ Ω and t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] .
Proof of Theorem 1.4. We first prove existence. Every pair in A L ρ 1 \mathcal{A}_{L}^{\rho_1} A L ρ 1 has finite energy, since ∣ M − Id ∣ ≤ D |M-\operatorname{Id}|\leq D ∣ M − Id ∣ ≤ D , J t ≤ L n J_t\leq L^n J t ≤ L n , and
∥ R t [ M , f ] ∥ L 1 ( Ω ) ≤ ∥ ρ 0 ∥ L 1 ( Ω ) + L n ∣ Ω ∣ ∥ f ∥ L 1 ( 0 , 1 ) . \|R_t[M,f]\|_{L^1(\Omega)}\leq\|\rho_0\|_{L^1(\Omega)}+L^n|\Omega|\|f\|_{L^1(0,1)}. ∥ R t [ M , f ] ∥ L 1 ( Ω ) ≤ ∥ ρ 0 ∥ L 1 ( Ω ) + L n ∣Ω∣∥ f ∥ L 1 ( 0 , 1 ) .
Moreover, R t [ M , f ] ≥ 0 R_t[M,f]\geq0 R t [ M , f ] ≥ 0 , so E α ( M , f ) ≥ α − 1 ∥ f ∥ L 2 ( 0 , 1 ) 2 E_\alpha(M,f)\geq\alpha^{-1}\|f\|_{L^2(0,1)}^2 E α ( M , f ) ≥ α − 1 ∥ f ∥ L 2 ( 0 , 1 ) 2 . For a minimizing sequence, the sources are therefore bounded in L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) , while the maps are uniformly Lipschitz with values in Ω ‾ \overline{\Omega} Ω . After passing to a subsequence, the pairs converge with respect to τ \tau τ , and the maps also converge uniformly. By Lemma 5.1, the limit belongs to A L ρ 1 \mathcal{A}_{L}^{\rho_1} A L ρ 1 . The lower semicontinuity of E α E_\alpha E α on V ρ 1 \mathcal{V}^{\rho_1} V ρ 1 , together with the sequential closedness of A L ρ 1 \mathcal{A}_{L}^{\rho_1} A L ρ 1 , shows that the limit attains the minimum. No additional bounded-distortion or Jacobian mass constraint is required, since Lemma 5.1 shows that the geometric conditions (3) already imply
∣ ∇ M ∣ n ≤ n n / 2 L 2 n det ∇ M a.e. in Ω and ∫ Ω det ∇ M d x = ∣ Ω ∣ . |\nabla M|^n\leq n^{n/2}L^{2n}\det\nabla M \quad\text{a.e. in }\Omega
\qquad\text{and}\qquad
\int_{\Omega}\det\nabla M\,dx=|\Omega|. ∣∇ M ∣ n ≤ n n /2 L 2 n det ∇ M a.e. in Ω and ∫ Ω det ∇ M d x = ∣Ω∣.
Fix ( M , f ) ∈ A L ρ 1 (M,f)\in\mathcal{A}_{L}^{\rho_1} ( M , f ) ∈ A L ρ 1 and abbreviate X t = X t [ M ] X_t=X_t[M] X t = X t [ M ] , J t = J t [ M ] J_t=J_t[M] J t = J t [ M ] , and R t = R t [ M , f ] R_t=R_t[M,f] R t = R t [ M , f ] . Define ρ \rho ρ and v v v by (4) . Since R t ≥ 0 R_t\geq0 R t ≥ 0 and J t ≥ L − n > 0 J_t\geq L^{-n}>0 J t ≥ L − n > 0 , the reconstructed density satisfies ρ ≥ 0 \rho\geq0 ρ ≥ 0 . Moreover, M ( x ) M(x) M ( x ) , x ∈ Ω ‾ x\in\overline{\Omega} x ∈ Ω imply ∣ u ( x ) ∣ = ∣ M ( x ) − x ∣ ≤ D |u(x)|=|M(x)-x|\leq D ∣ u ( x ) ∣ = ∣ M ( x ) − x ∣ ≤ D , and hence ∣ v ( t , y ) ∣ ≤ D |v(t,y)|\leq D ∣ v ( t , y ) ∣ ≤ D . Taking t = 1 t=1 t = 1 in (3) , we see that M M M is L L L -Lipschitz. Therefore,
∣ u ( x ) − u ( z ) ∣ ≤ ( L + 1 ) ∣ x − z ∣ for all x , z ∈ Ω ‾ . |u(x)-u(z)|\leq(L+1)|x-z| \qquad\text{for all }x,z\in\overline{\Omega}. ∣ u ( x ) − u ( z ) ∣ ≤ ( L + 1 ) ∣ x − z ∣ for all x , z ∈ Ω .
Together with (14) , this gives
∣ v ( t , y ) − v ( s , z ) ∣ = ∣ u ( Y t ( y ) ) − u ( Y s ( z ) ) ∣ ≤ ( L + 1 ) ∣ Y t ( y ) − Y s ( z ) ∣ ≤ L ( L + 1 ) ∣ y − z ∣ + L D ( L + 1 ) ∣ t − s ∣ . \begin{aligned}
|v(t,y)-v(s,z)|=|u(Y_t(y))-u(Y_s(z))| \\
&\leq(L+1)|Y_t(y)-Y_s(z)| \\
&\leq L(L+1)|y-z|+LD(L+1)|t-s|.
\end{aligned} ∣ v ( t , y ) − v ( s , z ) ∣ = ∣ u ( Y t ( y )) − u ( Y s ( z )) ∣ ≤ ( L + 1 ) ∣ Y t ( y ) − Y s ( z ) ∣ ≤ L ( L + 1 ) ∣ y − z ∣ + L D ( L + 1 ) ∣ t − s ∣.
Thus v v v is jointly Lipschitz on [ 0 , 1 ] × Ω ‾ [0,1]\times\overline{\Omega} [ 0 , 1 ] × Ω . Finally, the definition X t ( x ) = x + t u ( x ) X_t(x)=x+tu(x) X t ( x ) = x + t u ( x ) yields
∂ t X t ( x ) = u ( x ) = u ( Y t ( X t ( x ) ) ) = v ( t , X t ( x ) ) , X 0 ( x ) = x . \partial_tX_t(x)=u(x)=u(Y_t(X_t(x)))=v(t,X_t(x)), \qquad X_0(x)=x. ∂ t X t ( x ) = u ( x ) = u ( Y t ( X t ( x ))) = v ( t , X t ( x )) , X 0 ( x ) = x .
Hence X t X_t X t solves the flow equation associated with v v v . Since X 1 = M X_1=M X 1 = M , we have X t ( x ) = ( 1 − t ) x + t X 1 ( x ) X_t(x)=(1-t)x+tX_1(x) X t ( x ) = ( 1 − t ) x + t X 1 ( x ) , which is precisely (16) . We verify strong continuity of the density. The polynomial dependence of J t J_t J t on t t t , together with ∇ M ∈ L ∞ \nabla M\in L^\infty ∇ M ∈ L ∞ , gives J t ∈ C ( [ 0 , 1 ] ; L ∞ ( Ω ) ) J_t\in C([0,1];L^\infty(\Omega)) J t ∈ C ([ 0 , 1 ] ; L ∞ ( Ω )) . By (12) ,
∥ R t − R s ∥ L 1 ( Ω ) ≤ ∣ Ω ∣ ∫ min { s , t } max { s , t } ∣ f ( r ) ∣ d r . \|R_t-R_s\|_{L^1(\Omega)}\leq|\Omega|\int_{\min\{s,t\}}^{\max\{s,t\}}|f(r)|\,dr. ∥ R t − R s ∥ L 1 ( Ω ) ≤ ∣Ω∣ ∫ m i n { s , t } m a x { s , t } ∣ f ( r ) ∣ d r .
Thus R t R_t R t is continuous in L 1 L^1 L 1 . Fix s ∈ [ 0 , 1 ] s \in[0,1] s ∈ [ 0 , 1 ] and choose h ∈ C ( Ω ‾ ) h \in C(\overline{\Omega}) h ∈ C ( Ω ) with ∥ ρ ( s , ⋅ ) − h ∥ L 1 < ε \|\rho(s,\cdot)-h\|_{L^1}<\varepsilon ∥ ρ ( s , ⋅ ) − h ∥ L 1 < ε . The change-of-variables formula yields
∥ ρ ( t , ⋅ ) − h ∥ L 1 ( Ω ) = ∥ R t − h ( X t ) J t ∥ L 1 ( Ω ) . \|\rho(t,\cdot)-h\|_{L^1(\Omega)}=\|R_t-h(X_t)J_t\|_{L^1(\Omega)}. ∥ ρ ( t , ⋅ ) − h ∥ L 1 ( Ω ) = ∥ R t − h ( X t ) J t ∥ L 1 ( Ω ) .
The right-hand side is continuous in t t t . Hence
lim sup t → s ∥ ρ ( t , ⋅ ) − ρ ( s , ⋅ ) ∥ L 1 ( Ω ) ≤ 2 ∥ ρ ( s , ⋅ ) − h ∥ L 1 ( Ω ) < 2 ε . \limsup_{t\to s}\|\rho(t,\cdot)-\rho(s,\cdot)\|_{L^1(\Omega)}\le2\|\rho(s,\cdot)-h\|_{L^1(\Omega)}<2\varepsilon. t → s lim sup ∥ ρ ( t , ⋅ ) − ρ ( s , ⋅ ) ∥ L 1 ( Ω ) ≤ 2∥ ρ ( s , ⋅ ) − h ∥ L 1 ( Ω ) < 2 ε .
This proves ρ ∈ C ( [ 0 , 1 ] ; L 1 ( Ω ) ) \rho\in C([0,1];L^1(\Omega)) ρ ∈ C ([ 0 , 1 ] ; L 1 ( Ω )) . Let φ ∈ C 1 ( [ 0 , 1 ] × Ω ‾ ) \varphi\in C^1([0,1]\times\overline{\Omega}) φ ∈ C 1 ([ 0 , 1 ] × Ω ) and set
F ( t ) : = ∫ Ω R t ( x ) φ ( t , X t ( x ) ) d x . F(t):=\int_{\Omega}R_t(x)\varphi(t,X_t(x))\,dx. F ( t ) := ∫ Ω R t ( x ) φ ( t , X t ( x )) d x .
Since ∂ t R t = f ( t ) J t \partial_tR_t=f(t)J_t ∂ t R t = f ( t ) J t is integrable and φ ( t , X t ( x ) ) \varphi(t,X_t(x)) φ ( t , X t ( x )) has a bounded time derivative, F F F is absolutely continuous. For almost every t ∈ ( 0 , 1 ) t\in(0,1) t ∈ ( 0 , 1 ) ,
F ′ ( t ) = ∫ Ω f ( t ) J t ( x ) φ ( t , X t ( x ) ) d x + ∫ Ω R t ( x ) [ ∂ t φ ( t , X t ( x ) ) + u ( x ) ⋅ ∇ φ ( t , X t ( x ) ) ] d x . \begin{aligned}
F'(t)
&=\int_{\Omega}f(t)J_t(x)\varphi(t,X_t(x))\,dx\\
&\quad+\int_{\Omega}R_t(x)\left[\partial_t\varphi(t,X_t(x))+u(x)\cdot\nabla\varphi(t,X_t(x))\right]\,dx.
\end{aligned} F ′ ( t ) = ∫ Ω f ( t ) J t ( x ) φ ( t , X t ( x )) d x + ∫ Ω R t ( x ) [ ∂ t φ ( t , X t ( x )) + u ( x ) ⋅ ∇ φ ( t , X t ( x )) ] d x .
Integrating this identity over t ∈ ( 0 , 1 ) t\in(0,1) t ∈ ( 0 , 1 ) gives
F ( 1 ) − F ( 0 ) = ∫ 0 1 ∫ Ω f ( t ) J t ( x ) φ ( t , X t ( x ) ) d x d t + ∫ 0 1 ∫ Ω R t ( x ) [ ∂ t φ ( t , X t ( x ) ) + u ( x ) ⋅ ∇ φ ( t , X t ( x ) ) ] d x d t . \begin{aligned}
F(1)-F(0)
&=\int_0^1\int_{\Omega}f(t)J_t(x)\varphi(t,X_t(x))\,dx\,dt\\
&\quad+\int_0^1\int_{\Omega}R_t(x)\left[\partial_t\varphi(t,X_t(x))+u(x)\cdot\nabla\varphi(t,X_t(x))\right]\,dx\,dt.
\end{aligned} F ( 1 ) − F ( 0 ) = ∫ 0 1 ∫ Ω f ( t ) J t ( x ) φ ( t , X t ( x )) d x d t + ∫ 0 1 ∫ Ω R t ( x ) [ ∂ t φ ( t , X t ( x )) + u ( x ) ⋅ ∇ φ ( t , X t ( x )) ] d x d t .
For each t t t , make the change of variables y = X t ( x ) y=X_t(x) y = X t ( x ) . Using d y = J t ( x ) d x dy=J_t(x)\,dx d y = J t ( x ) d x , R t ( x ) = ρ ( t , X t ( x ) ) J t ( x ) R_t(x)=\rho(t,X_t(x))J_t(x) R t ( x ) = ρ ( t , X t ( x )) J t ( x ) , u ( x ) = v ( t , X t ( x ) ) u(x)=v(t,X_t(x)) u ( x ) = v ( t , X t ( x )) , and X t ( Ω ) = Ω X_t(\Omega)=\Omega X t ( Ω ) = Ω , we obtain
F ( 1 ) − F ( 0 ) = ∫ 0 1 ∫ Ω [ ρ ∂ t φ + ρ v ⋅ ∇ φ + f ( t ) φ ] d y d t . F(1)-F(0)=\int_0^1\int_{\Omega}\left[\rho\,\partial_t\varphi+\rho v\cdot\nabla\varphi+f(t)\varphi\right]\,dy\,dt. F ( 1 ) − F ( 0 ) = ∫ 0 1 ∫ Ω [ ρ ∂ t φ + ρ v ⋅ ∇ φ + f ( t ) φ ] d y d t .
Finally, X 0 = Id X_0=\operatorname{Id} X 0 = Id and R 0 = ρ 0 R_0=\rho_0 R 0 = ρ 0 yield
F ( 0 ) = ∫ Ω ρ 0 ( y ) φ ( 0 , y ) d y . F(0)=\int_{\Omega}\rho_0(y)\varphi(0,y)\,dy. F ( 0 ) = ∫ Ω ρ 0 ( y ) φ ( 0 , y ) d y .
Since X 1 = M X_1=M X 1 = M and R 1 = ρ 1 ( M ) det ∇ M R_1=\rho_1(M)\det\nabla M R 1 = ρ 1 ( M ) det ∇ M , another change of variables gives
F ( 1 ) = ∫ Ω ρ 1 ( M ( x ) ) det ∇ M ( x ) φ ( 1 , M ( x ) ) d x = ∫ Ω ρ 1 ( y ) φ ( 1 , y ) d y . \begin{aligned}
F(1)
&=\int_{\Omega}\rho_1(M(x))\det\nabla M(x)\varphi(1,M(x))\,dx\\
&=\int_{\Omega}\rho_1(y)\varphi(1,y)\,dy.
\end{aligned} F ( 1 ) = ∫ Ω ρ 1 ( M ( x )) det ∇ M ( x ) φ ( 1 , M ( x )) d x = ∫ Ω ρ 1 ( y ) φ ( 1 , y ) d y .
Substituting these endpoint identities proves (5.5). Thus ( ρ , v , f ) ∈ D L ρ 0 , ρ 1 (\rho,v,f)\in\mathcal{D}_{L}^{\rho_0,\rho_1} ( ρ , v , f ) ∈ D L ρ 0 , ρ 1 .
We next prove the converse. Let ( ρ , v , f ) ∈ D L ρ 0 , ρ 1 (\rho,v,f)\in\mathcal{D}_{L}^{\rho_0,\rho_1} ( ρ , v , f ) ∈ D L ρ 0 , ρ 1 and let X t X_t X t be its flow. Set M = X 1 M=X_1 M = X 1 . By (5.6),
X t = X t [ M ] , v ( t , X t ( x ) ) = M ( x ) − x . X_t=X_t[M],\qquad v(t,X_t(x))=M(x)-x. X t = X t [ M ] , v ( t , X t ( x )) = M ( x ) − x .
The geometric argument in Lemma 5.1 gives J t [ M ] > 0 J_t[M]>0 J t [ M ] > 0 a.e. for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . We justify the density identity directly from the weak equation. For η ∈ C c 1 ( 0 , 1 ) \eta\in C_c^1(0,1) η ∈ C c 1 ( 0 , 1 ) and ζ ∈ C 1 ( Ω ‾ ) \zeta\in C^1(\overline{\Omega}) ζ ∈ C 1 ( Ω ) , set
φ ( t , y ) : = η ( t ) ζ ( Y t ( y ) ) . \varphi(t,y):=\eta(t)\zeta(Y_t(y)). φ ( t , y ) := η ( t ) ζ ( Y t ( y )) .
Although φ \varphi φ is only Lipschitz, it is an admissible test function by smooth approximation. Indeed, the approximations converge uniformly, with uniformly bounded derivatives converging almost everywhere, so the weak identity passes to the limit by dominated convergence. The space-time map ( t , x ) ↦ ( t , X t ( x ) ) (t,x) \mapsto(t,X_t(x)) ( t , x ) ↦ ( t , X t ( x )) and its inverse are bi-Lipschitz. Applying the Lipschitz chain rule to Y t ( X t ( x ) ) = x Y_t(X_t(x))=x Y t ( X t ( x )) = x gives
∂ t Y t + D y Y t v = 0 a.e. in ( 0 , 1 ) × Ω . \partial_tY_t+D_yY_t\,v=0 \quad\text{a.e. in }(0,1)\times\Omega. ∂ t Y t + D y Y t v = 0 a.e. in ( 0 , 1 ) × Ω.
Thus
∂ t φ + v ⋅ ∇ φ = η ′ ( t ) ζ ( Y t ( y ) ) . \partial_t\varphi+v\cdot\nabla\varphi=\eta'(t)\zeta(Y_t(y)). ∂ t φ + v ⋅ ∇ φ = η ′ ( t ) ζ ( Y t ( y )) .
It follows that
d d t ∫ Ω ρ ( t , y ) ζ ( Y t ( y ) ) d y = f ( t ) ∫ Ω ζ ( Y t ( y ) ) d y = f ( t ) ∫ Ω ζ ( x ) J t ( x ) d x \frac{d}{dt}\int_{\Omega}\rho(t,y)\zeta(Y_t(y))\,dy=f(t)\int_{\Omega}\zeta(Y_t(y))\,dy=f(t)\int_{\Omega}\zeta(x)J_t(x)\,dx d t d ∫ Ω ρ ( t , y ) ζ ( Y t ( y )) d y = f ( t ) ∫ Ω ζ ( Y t ( y )) d y = f ( t ) ∫ Ω ζ ( x ) J t ( x ) d x
in distributions in time. The left-hand integral is continuous in t t t , and the right-hand side is integrable. For every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] ,
∫ Ω ρ ( t , X t ( x ) ) J t ( x ) ζ ( x ) d x = ∫ Ω ( ρ 0 ( x ) + ∫ 0 t f ( s ) J s ( x ) d s ) ζ ( x ) d x . \int_{\Omega}\rho(t,X_t(x))J_t(x)\zeta(x)\,dx=\int_{\Omega}\left(\rho_0(x)+\int_0^t f(s)J_s(x)\,ds\right)\zeta(x)\,dx. ∫ Ω ρ ( t , X t ( x )) J t ( x ) ζ ( x ) d x = ∫ Ω ( ρ 0 ( x ) + ∫ 0 t f ( s ) J s ( x ) d s ) ζ ( x ) d x .
Since ζ \zeta ζ is arbitrary, we obtain ρ ( t , X t ( x ) ) J t ( x ) = R t [ M , f ] ( x ) \rho(t,X_t(x))J_t(x)=R_t[M,f](x) ρ ( t , X t ( x )) J t ( x ) = R t [ M , f ] ( x ) for a.e. x x x . In particular, R t [ M , f ] ≥ 0 R_t[M,f]\ge0 R t [ M , f ] ≥ 0 for every fixed t t t . The terminal condition gives
ρ 1 ( M ) det ∇ M = ρ 0 + ∫ 0 1 f ( s ) J s [ M ] d s . \rho_1(M)\det\nabla M=\rho_0+\int_0^1f(s)J_s[M]\,ds. ρ 1 ( M ) det ∇ M = ρ 0 + ∫ 0 1 f ( s ) J s [ M ] d s .
Thus ( M , f ) ∈ V ρ 1 (M,f)\in\mathcal{V}^{\rho_1} ( M , f ) ∈ V ρ 1 and, by the flow geometry, ( M , f ) ∈ A L ρ 1 (M,f)\in\mathcal{A}^{\rho_1}_L ( M , f ) ∈ A L ρ 1 . Finally, in either direction the change-of-variables formula gives
E α ( ρ , v , f ) = ∫ 0 1 ∫ Ω ∣ M ( x ) − x ∣ 2 R t [ M , f ] ( x ) d x d t + 1 α ∫ 0 1 ∣ f ( t ) ∣ 2 d t = E α ( M , f ) . \mathcal{E}_\alpha(\rho,v,f)=\int_0^1\int_{\Omega}|M(x)-x|^2R_t[M,f](x)\,dx\,dt+\frac{1}{\alpha}\int_0^1|f(t)|^2\,dt
=\mathcal{E}_\alpha(M,f). E α ( ρ , v , f ) = ∫ 0 1 ∫ Ω ∣ M ( x ) − x ∣ 2 R t [ M , f ] ( x ) d x d t + α 1 ∫ 0 1 ∣ f ( t ) ∣ 2 d t = E α ( M , f ) .
The two constructions and static existence prove (1.5), attainment of the dynamic minimum, and the assertions about minimizers. □
Corollary 5.1. Assume that the minimum of E α \mathcal{E}_\alpha E α on D ρ 0 , ρ 1 \mathcal{D}^{\rho_0,\rho_1} D ρ 0 , ρ 1 is attained by a triple belonging to D L ρ 0 , ρ 1 \mathcal{D}^{\rho_0,\rho_1}_L D L ρ 0 , ρ 1 . Then
min ( ρ , v , f ) ∈ D L ρ 0 , ρ 1 E α ( ρ , v , f ) = min ( M , f ) ∈ A L ρ 1 E α ( M , f ) . \min_{(\rho,v,f)\in\mathcal{D}^{\rho_0,\rho_1}_L}\mathcal{E}_\alpha(\rho,v,f)=\min_{(M,f)\in\mathcal{A}^{\rho_1}_L}\mathcal{E}_\alpha(M,f). ( ρ , v , f ) ∈ D L ρ 0 , ρ 1 min E α ( ρ , v , f ) = ( M , f ) ∈ A L ρ 1 min E α ( M , f ) . Every minimizer on A L ρ 1 \mathcal{A}^{\rho_1}_L A L ρ 1 reconstructs to a minimizer on D ρ 0 , ρ 1 \mathcal{D}^{\rho_0,\rho_1} D ρ 0 , ρ 1 .
Proof. The assumed dynamic minimizer yields a pair in A L ρ 1 \mathcal{A}^{\rho_1}_L A L ρ 1 by Theorem 1.4 . Since D L ρ 0 , ρ 1 ⊂ D ρ 0 , ρ 1 \mathcal{D}^{\rho_0,\rho_1}_L\subset\mathcal{D}^{\rho_0,\rho_1} D L ρ 0 , ρ 1 ⊂ D ρ 0 , ρ 1 and this subclass contains a minimizer of the larger problem, the two dynamic minimum values coincide. The conclusion now follows from (5) . □
Remark 5.1. Gangbo et al. [15 ] derive the affine interpolation from the Hamilton-Jacobi optimality equation. However, their analysis does not establish the domain-preserving bi-Lipschitz flow properties required here, nor the prescribed uniform bound L L L . Thus equality between the minimum of E α \mathcal{E}_\alpha E α on A L ρ 1 \mathcal{A}^{\rho_1}_L A L ρ 1 and the minimum of E α \mathcal{E}_\alpha E α on D ρ 0 , ρ 1 \mathcal{D}^{\rho_0,\rho_1} D ρ 0 , ρ 1 requires the additional hypothesis of Corollary 5.1.
Proposition 5.1 (Stability of the reconstruction). Let ( M j , f j ) , ( M , f ) ∈ A L ρ 1 (M_j,f_j),(M,f) \in\mathcal{A}^{\rho_1}_L ( M j , f j ) , ( M , f ) ∈ A L ρ 1 satisfy
M j → M uniformly on Ω ‾ , f j ⇀ f in L 2 ( 0 , 1 ) . M_j \to M \quad\text{uniformly on } \overline{\Omega}, \qquad f_j \rightharpoonup f \quad\text{in } L^2(0,1). M j → M uniformly on Ω , f j ⇀ f in L 2 ( 0 , 1 ) . Let ( ρ j , v j , f j ) (\rho_j,v_j,f_j) ( ρ j , v j , f j ) and ( ρ , v , f ) (\rho,v,f) ( ρ , v , f ) be their reconstructed dynamic triples. Then v j → v v_j \to v v j → v uniformly on [ 0 , 1 ] × Ω ‾ [0,1] \times\overline{\Omega} [ 0 , 1 ] × Ω , and, for every ψ ∈ C ( Ω ‾ ) \psi\in C(\overline{\Omega}) ψ ∈ C ( Ω ) ,
sup t ∈ [ 0 , 1 ] ∣ ∫ Ω ψ ( y ) ( ρ j ( t , y ) − ρ ( t , y ) ) d y ∣ → 0. \sup_{t\in[0,1]}\left|\int_{\Omega}\psi(y)\bigl(\rho_j(t,y)-\rho(t,y)\bigr)\,dy\right| \to0. t ∈ [ 0 , 1 ] sup ∫ Ω ψ ( y ) ( ρ j ( t , y ) − ρ ( t , y ) ) d y → 0. Moreover, ρ j ⇀ ∗ ρ \rho_j \overset{*}{\rightharpoonup} \rho ρ j ⇀ ∗ ρ in L ∞ ( ( 0 , 1 ) × Ω ) L^\infty((0,1)\times\Omega) L ∞ (( 0 , 1 ) × Ω ) and G ( M j , f j ) → G ( M , f ) G(M_j,f_j) \to G(M,f) G ( M j , f j ) → G ( M , f ) .
Proof. Write X j , t : = X t [ M j ] X_{j,t}:=X_t[M_j] X j , t := X t [ M j ] , X t : = X t [ M ] X_t:=X_t[M] X t := X t [ M ] , Y j , t : = X j , t − 1 Y_{j,t}:=X_{j,t}^{-1} Y j , t := X j , t − 1 , Y t : = X t − 1 Y_t:=X_t^{-1} Y t := X t − 1 , and set u j : = M j − Id u_j:=M_j-\operatorname{Id} u j := M j − Id , u : = M − Id u:=M-\operatorname{Id} u := M − Id . We first prove convergence of the velocities. The inverse Lipschitz bound gives
∣ Y j , t ( y ) − Y t ( y ) ∣ ≤ L ∣ X j , t ( Y j , t ( y ) ) − X j , t ( Y t ( y ) ) ∣ = L ∣ X t ( Y t ( y ) ) − X j , t ( Y t ( y ) ) ∣ ≤ L ∥ M j − M ∥ L ∞ ( Ω ‾ ) . \begin{aligned}
\left|Y_{j,t}(y)-Y_t(y)\right| \le L\left|X_{j,t}(Y_{j,t}(y))-X_{j,t}(Y_t(y))\right| \\
&= L\left|X_t(Y_t(y))-X_{j,t}(Y_t(y))\right| \\
&\le L\|M_j-M\|_{L^\infty(\overline{\Omega})}.
\end{aligned} ∣ Y j , t ( y ) − Y t ( y ) ∣ ≤ L ∣ X j , t ( Y j , t ( y )) − X j , t ( Y t ( y )) ∣ = L ∣ X t ( Y t ( y )) − X j , t ( Y t ( y )) ∣ ≤ L ∥ M j − M ∥ L ∞ ( Ω ) .
Since u u u is ( L + 1 ) (L+1) ( L + 1 ) -Lipschitz and v j ( t , y ) = u j ( Y j , t ( y ) ) v_j(t,y)=u_j(Y_{j,t}(y)) v j ( t , y ) = u j ( Y j , t ( y )) , v ( t , y ) = u ( Y t ( y ) ) v(t,y)=u(Y_t(y)) v ( t , y ) = u ( Y t ( y )) , we obtain
∥ v j − v ∥ L ∞ ( [ 0 , 1 ] × Ω ‾ ) ≤ ( 1 + L ( L + 1 ) ) ∥ M j − M ∥ L ∞ ( Ω ‾ ) → 0. \|v_j-v\|_{L^\infty([0,1]\times\overline{\Omega})} \le\bigl(1+L(L+1)\bigr)\|M_j-M\|_{L^\infty(\overline{\Omega})} \to0. ∥ v j − v ∥ L ∞ ([ 0 , 1 ] × Ω ) ≤ ( 1 + L ( L + 1 ) ) ∥ M j − M ∥ L ∞ ( Ω ) → 0.
We next establish a uniform bound for the densities. The endpoint constraint for ( M , f ) (M,f) ( M , f ) and the Jacobian bounds imply
0 ≤ ρ 0 = ρ 1 ( M ) J 1 [ M ] − ∫ 0 1 f ( s ) J s [ M ] d s ≤ L n ( ∥ ρ 1 ∥ L ∞ ( Ω ) + ∥ f ∥ L 1 ( 0 , 1 ) ) . 0\le\rho_0=\rho_1(M)J_1[M]-\int_0^1 f(s)J_s[M]\,ds\le L^n\bigl(\|\rho_1\|_{L^\infty(\Omega)}+\|f\|_{L^1(0,1)}\bigr). 0 ≤ ρ 0 = ρ 1 ( M ) J 1 [ M ] − ∫ 0 1 f ( s ) J s [ M ] d s ≤ L n ( ∥ ρ 1 ∥ L ∞ ( Ω ) + ∥ f ∥ L 1 ( 0 , 1 ) ) .
Thus ρ 0 ∈ L ∞ ( Ω ) \rho_0\in L^\infty(\Omega) ρ 0 ∈ L ∞ ( Ω ) . Since weak convergence of ( f j ) (f_j) ( f j ) gives B : = sup j ∥ f j ∥ L 2 ( 0 , 1 ) < ∞ B:=\sup_j\|f_j\|_{L^2(0,1)}<\infty B := sup j ∥ f j ∥ L 2 ( 0 , 1 ) < ∞ , the reconstruction formulas yield
0 ≤ R t [ M j , f j ] ≤ ∥ ρ 0 ∥ L ∞ ( Ω ) + L n B 0\le R_t[M_j,f_j]\le\|\rho_0\|_{L^\infty(\Omega)}+L^nB 0 ≤ R t [ M j , f j ] ≤ ∥ ρ 0 ∥ L ∞ ( Ω ) + L n B
and
sup j sup t ∈ [ 0 , 1 ] ∥ ρ j ( t , ⋅ ) ∥ L ∞ ( Ω ) ≤ L n ∥ ρ 0 ∥ L ∞ ( Ω ) + L 2 n B . \sup_j\sup_{t\in[0,1]}\|\rho_j(t,\cdot)\|_{L^\infty(\Omega)}\le L^n\|\rho_0\|_{L^\infty(\Omega)}+L^{2n}B. j sup t ∈ [ 0 , 1 ] sup ∥ ρ j ( t , ⋅ ) ∥ L ∞ ( Ω ) ≤ L n ∥ ρ 0 ∥ L ∞ ( Ω ) + L 2 n B .
The same type of bound holds for ρ \rho ρ . To prove convergence of the intermediate states, expand
J t [ M j ] = ∑ k = 0 n c j , k t k , J t [ M ] = ∑ k = 0 n c k t k . J_t[M_j]=\sum_{k=0}^{n}c_{j,k}t^k,\qquad J_t[M]=\sum_{k=0}^{n}c_kt^k. J t [ M j ] = k = 0 ∑ n c j , k t k , J t [ M ] = k = 0 ∑ n c k t k .
The coefficients are sums of minors of ∇ u j \nabla u_j ∇ u j and ∇ u \nabla u ∇ u , respectively. The uniform Lipschitz bounds and uniform convergence of the maps imply that M j ⇀ M M_j\rightharpoonup M M j ⇀ M in W 1 , p ( Ω ; R n ) W^{1,p}(\Omega;\mathbb{R}^n) W 1 , p ( Ω ; R n ) for every 1 < p < ∞ 1<p<\infty 1 < p < ∞ . By weak continuity of minors and the uniform L ∞ L^\infty L ∞ bounds on these coefficients, c j , k ⇀ ∗ c k c_{j,k}\overset{*}{\rightharpoonup}c_k c j , k ⇀ ∗ c k in L ∞ ( Ω ) L^\infty(\Omega) L ∞ ( Ω ) , k = 0 , … , n k=0,\ldots,n k = 0 , … , n . Define
F j , k ( t ) : = ∫ 0 t f j ( s ) s k d s , F k ( t ) : = ∫ 0 t f ( s ) s k d s . F_{j,k}(t):=\int_0^t f_j(s)s^k\,ds,\qquad F_k(t):=\int_0^t f(s)s^k\,ds. F j , k ( t ) := ∫ 0 t f j ( s ) s k d s , F k ( t ) := ∫ 0 t f ( s ) s k d s .
Weak convergence of the sources gives F j , k ( t ) → F k ( t ) F_{j,k}(t)\to F_k(t) F j , k ( t ) → F k ( t ) for every t t t . Moreover, ∣ F j , k ( t ) − F j , k ( s ) ∣ ≤ B ∣ t − s ∣ 1 / 2 \left|F_{j,k}(t)-F_{j,k}(s)\right|\le B|t-s|^{1/2} ∣ F j , k ( t ) − F j , k ( s ) ∣ ≤ B ∣ t − s ∣ 1/2 . Consequently, F j , k → F k F_{j,k}\to F_k F j , k → F k uniformly on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Writing R j , t : = R t [ M j , f j ] R_{j,t}:=R_t[M_j,f_j] R j , t := R t [ M j , f j ] and R t : = R t [ M , f ] R_t:=R_t[M,f] R t := R t [ M , f ] , we therefore have
R j , t = ρ 0 + ∑ k = 0 n c j , k F j , k ( t ) , R t = ρ 0 + ∑ k = 0 n c k F k ( t ) . R_{j,t}=\rho_0+\sum_{k=0}^{n}c_{j,k}F_{j,k}(t),\qquad R_t=\rho_0+\sum_{k=0}^{n}c_kF_k(t). R j , t = ρ 0 + k = 0 ∑ n c j , k F j , k ( t ) , R t = ρ 0 + k = 0 ∑ n c k F k ( t ) .
It follows that, for every h ∈ L 1 ( Ω ) h \in L^{1}(\Omega) h ∈ L 1 ( Ω ) ,
sup t ∈ [ 0 , 1 ] ∣ ∫ Ω h ( x ) ( R j , t ( x ) − R t ( x ) ) d x ∣ → 0. (17) \sup_{t\in[0,1]}\left|\int_{\Omega}h(x)(R_{j,t}(x)-R_t(x))\,dx\right|\to0. \tag*{(17)} t ∈ [ 0 , 1 ] sup ∫ Ω h ( x ) ( R j , t ( x ) − R t ( x )) d x → 0. (17)
Indeed, each term can be estimated using the uniform convergence of F j , k F_{j,k} F j , k and the weak-* convergence of c j , k c_{j,k} c j , k . Fix ψ ∈ C ( Ω ‾ ) \psi\in C(\overline{\Omega}) ψ ∈ C ( Ω ) . By change of variables,
∫ Ω ψ ( y ) ( ρ j ( t , y ) − ρ ( t , y ) ) d y = ∫ Ω ( ψ ( X j , t ( x ) ) − ψ ( X t ( x ) ) ) R j , t ( x ) d x + ∫ Ω ψ ( X t ( x ) ) ( R j , t ( x ) − R t ( x ) ) d x . \int_{\Omega}\psi(y)(\rho_j(t,y)-\rho(t,y))\,dy
= \int_{\Omega}(\psi(X_{j,t}(x))-\psi(X_t(x)))R_{j,t}(x)\,dx
+ \int_{\Omega}\psi(X_t(x))(R_{j,t}(x)-R_t(x))\,dx. ∫ Ω ψ ( y ) ( ρ j ( t , y ) − ρ ( t , y )) d y = ∫ Ω ( ψ ( X j , t ( x )) − ψ ( X t ( x ))) R j , t ( x ) d x + ∫ Ω ψ ( X t ( x )) ( R j , t ( x ) − R t ( x )) d x .
The first term tends to zero uniformly in t t t , because X j , t → X t X_{j,t}\to X_t X j , t → X t uniformly in space and time and the intermediate states are uniformly bounded. For the second term, since { ψ ∘ X t : t ∈ [ 0 , 1 ] } \{\psi\circ X_t:t\in[0,1]\} { ψ ∘ X t : t ∈ [ 0 , 1 ]} is compact in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) and R j , t − R t R_{j,t}-R_t R j , t − R t is uniformly bounded in L ∞ ( Ω ) L^\infty(\Omega) L ∞ ( Ω ) , a finite approximation argument using (5.7) gives
sup t ∈ [ 0 , 1 ] ∣ ∫ Ω ψ ( X t ( x ) ) ( R j , t ( x ) − R t ( x ) ) d x ∣ → 0. \sup_{t\in[0,1]}\left|\int_{\Omega}\psi(X_t(x))(R_{j,t}(x)-R_t(x))\,dx\right|\to0. t ∈ [ 0 , 1 ] sup ∫ Ω ψ ( X t ( x )) ( R j , t ( x ) − R t ( x )) d x → 0.
Hence
sup t ∈ [ 0 , 1 ] ∣ ∫ Ω ψ ( y ) ( ρ j ( t , y ) − ρ ( t , y ) ) d y ∣ → 0. \sup_{t\in[0,1]}\left|\int_{\Omega}\psi(y)(\rho_j(t,y)-\rho(t,y))\,dy\right|\to0. t ∈ [ 0 , 1 ] sup ∫ Ω ψ ( y ) ( ρ j ( t , y ) − ρ ( t , y )) d y → 0.
This convergence implies convergence against space-time test functions of the form η ( t ) ψ ( y ) \eta(t)\psi(y) η ( t ) ψ ( y ) , with η ∈ L 1 ( 0 , 1 ) \eta\in L^1(0,1) η ∈ L 1 ( 0 , 1 ) and ψ ∈ C ( Ω ‾ ) \psi\in C(\overline{\Omega}) ψ ∈ C ( Ω ) . Finite sums of such functions are dense in L 1 ( ( 0 , 1 ) × Ω ) L^1((0,1)\times\Omega) L 1 (( 0 , 1 ) × Ω ) . Together with the uniform density bounds, this proves ρ j ⇀ ∗ ρ \rho_j\rightharpoonup^{*}\rho ρ j ⇀ ∗ ρ in L ∞ ( ( 0 , 1 ) × Ω ) L^\infty((0,1)\times\Omega) L ∞ (( 0 , 1 ) × Ω ) . Finally, the energy-preserving reconstruction gives
G ( M j , f j ) = ∫ 0 1 ∫ Ω ρ j ∣ v j ∣ 2 d y d t , G ( M , f ) = ∫ 0 1 ∫ Ω ρ ∣ v ∣ 2 d y d t . G(M_j,f_j)=\int_0^1\int_{\Omega}\rho_j|v_j|^2\,dy\,dt,\qquad
G(M,f)=\int_0^1\int_{\Omega}\rho|v|^2\,dy\,dt. G ( M j , f j ) = ∫ 0 1 ∫ Ω ρ j ∣ v j ∣ 2 d y d t , G ( M , f ) = ∫ 0 1 ∫ Ω ρ ∣ v ∣ 2 d y d t .
Thus
G ( M j , f j ) − G ( M , f ) = ∫ 0 1 ∫ Ω ρ j ( ∣ v j ∣ 2 − ∣ v ∣ 2 ) d y d t + ∫ 0 1 ∫ Ω ( ρ j − ρ ) ∣ v ∣ 2 d y d t . G(M_j,f_j)-G(M,f)
=\int_0^1\int_{\Omega}\rho_j(|v_j|^2-|v|^2)\,dy\,dt
+\int_0^1\int_{\Omega}(\rho_j-\rho)|v|^2\,dy\,dt. G ( M j , f j ) − G ( M , f ) = ∫ 0 1 ∫ Ω ρ j ( ∣ v j ∣ 2 − ∣ v ∣ 2 ) d y d t + ∫ 0 1 ∫ Ω ( ρ j − ρ ) ∣ v ∣ 2 d y d t .
The first term tends to zero by uniform convergence of the velocities and the uniform density bounds. The second tends to zero by weak-* convergence of the densities. Therefore G ( M j , f j ) → G ( M , f ) G(M_j,f_j)\to G(M,f) G ( M j , f j ) → G ( M , f ) . □ \square □
Strong source penalization and the constant-source limit
We work in the bounded-target setting and use A L ρ 1 \mathcal{A}^{\rho_1}_{L} A L ρ 1 . As α → 0 + \alpha\to0^+ α → 0 + , minimization first selects the smallest admissible source cost and then the least transport energy among the resulting pairs. The endpoint mass difference fixes the mean source. The question is whether this mean can be realized by a constant admissible source. We establish the general limit first and then characterize the constant-source case.
Set
a : = ∫ Ω ρ 1 ( x ) d x − ∫ Ω ρ 0 ( x ) d x ∣ Ω ∣ . a:=\frac{\int_{\Omega}\rho_1(x)\,dx-\int_{\Omega}\rho_0(x)\,dx}{|\Omega|}. a := ∣Ω∣ ∫ Ω ρ 1 ( x ) d x − ∫ Ω ρ 0 ( x ) d x .
Every ( M , f ) ∈ A L ρ 1 (M,f) \in\mathcal{A}^{\rho_1}_L ( M , f ) ∈ A L ρ 1 determines a dynamic admissible triple by the reconstruction established above. Taking φ ≡ 1 \varphi\equiv1 φ ≡ 1 in the weak formulation of the continuity equation gives
∫ Ω ρ 1 ( y ) d y − ∫ Ω ρ 0 ( y ) d y = ∣ Ω ∣ ∫ 0 1 f ( t ) d t . \int_{\Omega} \rho_1(y)\,dy-\int_{\Omega}\rho_0(y)\,dy=|\Omega|\int_0^1 f(t)\,dt. ∫ Ω ρ 1 ( y ) d y − ∫ Ω ρ 0 ( y ) d y = ∣Ω∣ ∫ 0 1 f ( t ) d t .
Thus ∫ 0 1 f ( t ) d t = a \int_0^1 f(t)\,dt=a ∫ 0 1 f ( t ) d t = a , and consequently
∥ f ∥ L 2 ( 0 , 1 ) 2 = a 2 + ∥ f − a ∥ L 2 ( 0 , 1 ) 2 . \|f\|_{L^2(0,1)}^2=a^2+\|f-a\|_{L^2(0,1)}^2. ∥ f ∥ L 2 ( 0 , 1 ) 2 = a 2 + ∥ f − a ∥ L 2 ( 0 , 1 ) 2 .
In particular,
E α ( M , f ) − a 2 α = G ( M , f ) + 1 α ∥ f − a ∥ L 2 ( 0 , 1 ) 2 . (18) E_\alpha(M,f)-\frac{a^2}{\alpha}=G(M,f)+\frac{1}{\alpha}\|f-a\|_{L^2(0,1)}^2. \tag*{(18)} E α ( M , f ) − α a 2 = G ( M , f ) + α 1 ∥ f − a ∥ L 2 ( 0 , 1 ) 2 . (18)
The source cost is therefore bounded below by a 2 a^2 a 2 , with equality precisely when f = a f=a f = a a.e. Whether this lower bound is attained in A L ρ 1 \mathcal{A}^{\rho_1}_L A L ρ 1 also depends on the endpoint densities and the prescribed geometric constraints. b L b_L b L is the least source cost compatible with the admissibility conditions.
Proof of Theorem 1.5. Since A L ρ 1 ≠ ∅ \mathcal{A}^{\rho_1}_L \ne\varnothing A L ρ 1 = ∅ and every admissible source belongs to L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) , we have 0 ≤ b L < ∞ 0 \le b_L < \infty 0 ≤ b L < ∞ . Choose a minimizing sequence ( M j , f j ) ∈ A L ρ 1 (M_j,f_j) \in\mathcal{A}^{\rho_1}_L ( M j , f j ) ∈ A L ρ 1 such that ∥ f j ∥ L 2 ( 0 , 1 ) → b L \|f_j\|_{L^2(0,1)} \to b_L ∥ f j ∥ L 2 ( 0 , 1 ) → b L . In particular, ( f j ) (f_j) ( f j ) is bounded in L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) . Since Ω \Omega Ω is bounded and M j ( Ω ‾ ) ⊂ Ω ‾ M_j(\overline{\Omega}) \subset\overline{\Omega} M j ( Ω ) ⊂ Ω , the maps are uniformly bounded in L ∞ ( Ω ) L^\infty(\Omega) L ∞ ( Ω ) . Their common Lipschitz bound also gives ∥ ∇ M j ∥ L ∞ ( Ω ) ≤ n L \|\nabla M_j\|_{L^\infty(\Omega)} \le\sqrt{n}L ∥∇ M j ∥ L ∞ ( Ω ) ≤ n L . Thus ( M j ) (M_j) ( M j ) is bounded in W 1 , n ( Ω ; R n ) W^{1,n}(\Omega;\mathbb{R}^n) W 1 , n ( Ω ; R n ) , weak compactness in W 1 , n W^{1,n} W 1 , n and L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) yields a subsequence such that
M j ⇀ M in W 1 , n ( Ω ; R n ) , f j ⇀ f in L 2 ( 0 , 1 ) . M_j \rightharpoonup M \quad\text{in } W^{1,n}(\Omega;\mathbb{R}^n), \qquad f_j \rightharpoonup f \quad\text{in } L^2(0,1). M j ⇀ M in W 1 , n ( Ω ; R n ) , f j ⇀ f in L 2 ( 0 , 1 ) .
Moreover, the maps are uniformly bounded and equicontinuous on the compact set Ω ‾ \overline{\Omega} Ω . The Arzelà–Ascoli theorem therefore gives a further subsequence converging uniformly on Ω ‾ \overline{\Omega} Ω . The uniform limit agrees almost everywhere with the weak Sobolev limit M M M , and we use this continuous representative of M M M . Hence M j → M M_j \to M M j → M uniformly on Ω ‾ \overline{\Omega} Ω . Finally, the sequential closedness established in Lemma 5.1 implies ( M , f ) ∈ A L ρ 1 (M,f) \in\mathcal{A}^{\rho_1}_L ( M , f ) ∈ A L ρ 1 . Applying this compactness to a minimizing sequence for b L b_L b L , and using weak lower semicontinuity of the L 2 L^2 L 2 norm, proves that b L b_L b L is attained. The same argument shows that S L ρ 1 \mathcal{S}^{\rho_1}_L S L ρ 1 is sequentially τ \tau τ -compact. Lower semicontinuity of G G G therefore gives a minimizer on S L ρ 1 \mathcal{S}^{\rho_1}_L S L ρ 1 . Since every admissible pair has finite transport energy, 0 ≤ g L < ∞ 0 \le g_L < \infty 0 ≤ g L < ∞ .
For the liminf inequality, let α j → 0 + \alpha_j \to0^+ α j → 0 + and ( M j , f j ) ⟶ τ ( M , f ) (M_j,f_j) \mathrel{\mathop{\longrightarrow}^{\tau}} (M,f) ( M j , f j ) ⟶ τ ( M , f ) . If the lower limit of F ^ α j ( M j , f j ) \widehat{F}_{\alpha_j}(M_j,f_j) F α j ( M j , f j ) is infinite, there is nothing to prove. Otherwise, pass to a subsequence realizing the finite lower limit, along which these energies are bounded. The pairs then belong to A L ρ 1 \mathcal{A}^{\rho_1}_L A L ρ 1 , and
0 ≤ ∥ f j ∥ L 2 ( 0 , 1 ) 2 − b L ≤ α j F ^ α j ( M j , f j ) → 0. 0 \le\|f_j\|_{L^2(0,1)}^2-b_L \le\alpha_j\widehat{F}_{\alpha_j}(M_j,f_j) \to0. 0 ≤ ∥ f j ∥ L 2 ( 0 , 1 ) 2 − b L ≤ α j F α j ( M j , f j ) → 0.
Sequential closedness and weak lower semicontinuity give ( M , f ) ∈ A L ρ 1 (M,f) \in\mathcal{A}^{\rho_1}_L ( M , f ) ∈ A L ρ 1 and
b L ≤ ∥ f ∥ L 2 ( 0 , 1 ) 2 ≤ lim inf j → ∞ ∥ f j ∥ L 2 ( 0 , 1 ) 2 = b L . b_L \le\|f\|_{L^2(0,1)}^2 \le\liminf_{j\to\infty}\|f_j\|_{L^2(0,1)}^2=b_L. b L ≤ ∥ f ∥ L 2 ( 0 , 1 ) 2 ≤ j → ∞ lim inf ∥ f j ∥ L 2 ( 0 , 1 ) 2 = b L .
Thus ( M , f ) ∈ S L ρ 1 (M,f) \in\mathcal{S}^{\rho_1}_L ( M , f ) ∈ S L ρ 1 , and
lim inf j → ∞ F ^ α j ( M j , f j ) ≥ lim inf j → ∞ G ( M j , f j ) ≥ G ( M , f ) = F ^ 0 ( M , f ) . \liminf_{j\to\infty}\widehat{F}_{\alpha_j}(M_j,f_j)\ge\liminf_{j\to\infty}G(M_j,f_j)\ge G(M,f)=\widehat{F}_0(M,f). j → ∞ lim inf F α j ( M j , f j ) ≥ j → ∞ lim inf G ( M j , f j ) ≥ G ( M , f ) = F 0 ( M , f ) .
For ( M , f ) ∈ S L ρ 1 (M,f) \in\mathcal{S}^{\rho_1}_L ( M , f ) ∈ S L ρ 1 , the constant sequence is a recovery sequence because
F ^ α ( M , f ) = G ( M , f ) = F ^ 0 ( M , f ) . \widehat{F}_\alpha(M,f)=G(M,f)=\widehat{F}_0(M,f). F α ( M , f ) = G ( M , f ) = F 0 ( M , f ) .
Outside this set, the recovery inequality is automatic. This proves sequential Γ \Gamma Γ -convergence.
If 0 < α ≤ 1 0 < \alpha\le1 0 < α ≤ 1 and F ^ α ( M , f ) ≤ T \widehat{F}_{\alpha}(M,f) \le T F α ( M , f ) ≤ T , then
∥ f ∥ L 2 ( 0 , 1 ) 2 ≤ b L + α T ≤ b L + T . \lVert f\rVert_{L^{2}(0,1)}^{2} \le b_{L}+\alpha T \le b_{L}+T. ∥ f ∥ L 2 ( 0 , 1 ) 2 ≤ b L + α T ≤ b L + T .
Together with the map bounds and sequential closedness, this gives sequential equicoercivity. Let ( M ^ , f ^ ) ∈ S L ρ 1 (\widehat{M},\widehat{f}) \in S_{L}^{\rho_{1}} ( M , f ) ∈ S L ρ 1 satisfy G ( M ^ , f ^ ) = g L G(\widehat{M},\widehat{f})=g_{L} G ( M , f ) = g L . By minimality of ( M α , f α ) (M_{\alpha},f_{\alpha}) ( M α , f α ) and ∥ f ^ ∥ L 2 ( 0 , 1 ) 2 = b L \lVert\widehat{f}\rVert_{L^{2}(0,1)}^{2}=b_{L} ∥ f ∥ L 2 ( 0 , 1 ) 2 = b L , we have
0 ≤ G ( M α , f α ) ≤ F ^ α ( M α , f α ) ≤ F ^ α ( M ^ , f ^ ) = g L . (19) 0 \le G(M_{\alpha},f_{\alpha}) \le\widehat{F}_{\alpha}(M_{\alpha},f_{\alpha}) \le\widehat{F}_{\alpha}(\widehat{M},\widehat{f})=g_{L}. \tag*{(19)} 0 ≤ G ( M α , f α ) ≤ F α ( M α , f α ) ≤ F α ( M , f ) = g L . (19)
Consequently,
0 ≤ ∥ f α ∥ L 2 ( 0 , 1 ) 2 − b L ≤ α g L . 0 \le\lVert f_{\alpha}\rVert_{L^{2}(0,1)}^{2}-b_{L} \le\alpha g_{L}. 0 ≤ ∥ f α ∥ L 2 ( 0 , 1 ) 2 − b L ≤ α g L .
Thus the sources are bounded in L 2 ( 0 , 1 ) L^{2}(0,1) L 2 ( 0 , 1 ) for 0 < α ≤ 1 0 < \alpha\le1 0 < α ≤ 1 , and their squared norms converge to b L b_{L} b L . For any sequence α j → 0 + \alpha_{j}\to0^{+} α j → 0 + , the compactness established above yields a subsequence such that
( M α j , f α j ) τ ⇀ ( M , f ) , M α j → M uniformly on Ω ‾ . (M_{\alpha_{j}},f_{\alpha_{j}})\mathrel{\tau\!\rightharpoonup}(M,f), \qquad M_{\alpha_{j}}\to M \text{ uniformly on }\overline{\Omega}. ( M α j , f α j ) τ ⇀ ( M , f ) , M α j → M uniformly on Ω .
Sequential closedness gives ( M , f ) ∈ A L ρ 1 (M,f)\in A_{L}^{\rho_{1}} ( M , f ) ∈ A L ρ 1 . By the definition of b L b_{L} b L and weak lower semicontinuity,
b L ≤ ∥ f ∥ L 2 ( 0 , 1 ) 2 ≤ lim inf j → ∞ ∥ f α j ∥ L 2 ( 0 , 1 ) 2 = b L . b_{L}\le\lVert f\rVert_{L^{2}(0,1)}^{2}\le\liminf_{j\to\infty}\lVert f_{\alpha_{j}}\rVert_{L^{2}(0,1)}^{2}=b_{L}. b L ≤ ∥ f ∥ L 2 ( 0 , 1 ) 2 ≤ j → ∞ lim inf ∥ f α j ∥ L 2 ( 0 , 1 ) 2 = b L .
Hence ( M , f ) ∈ S L ρ 1 (M,f)\in S_{L}^{\rho_{1}} ( M , f ) ∈ S L ρ 1 . Moreover, weak convergence of the sources and convergence of their norms imply f α j → f f_{\alpha_{j}}\to f f α j → f strongly in L 2 ( 0 , 1 ) L^{2}(0,1) L 2 ( 0 , 1 ) . Since g L g_{L} g L is the minimum of G G G on S L ρ 1 S_{L}^{\rho_{1}} S L ρ 1 , lower semicontinuity and (19) give
g L ≤ G ( M , f ) ≤ lim inf j → ∞ G ( M α j , f α j ) ≤ lim sup j → ∞ G ( M α j , f α j ) ≤ g L . g_{L}\le G(M,f)\le\liminf_{j\to\infty}G(M_{\alpha_{j}},f_{\alpha_{j}})\le\limsup_{j\to\infty}G(M_{\alpha_{j}},f_{\alpha_{j}})\le g_{L}. g L ≤ G ( M , f ) ≤ j → ∞ lim inf G ( M α j , f α j ) ≤ j → ∞ lim sup G ( M α j , f α j ) ≤ g L .
Thus G ( M , f ) = g L G(M,f)=g_{L} G ( M , f ) = g L and G ( M α j , f α j ) → g L G(M_{\alpha_{j}},f_{\alpha_{j}})\to g_{L} G ( M α j , f α j ) → g L . Using (19) once more, we obtain F ^ α j ( M α j , f α j ) → g L \widehat{F}_{\alpha_{j}}(M_{\alpha_{j}},f_{\alpha_{j}})\to g_{L} F α j ( M α j , f α j ) → g L and
∥ f α j ∥ L 2 ( 0 , 1 ) 2 − b L α j = F ^ α j ( M α j , f α j ) − G ( M α j , f α j ) → 0. \frac{\lVert f_{\alpha_{j}}\rVert_{L^{2}(0,1)}^{2}-b_{L}}{\alpha_{j}}=\widehat{F}_{\alpha_{j}}(M_{\alpha_{j}},f_{\alpha_{j}})-G(M_{\alpha_{j}},f_{\alpha_{j}})\to0. α j ∥ f α j ∥ L 2 ( 0 , 1 ) 2 − b L = F α j ( M α j , f α j ) − G ( M α j , f α j ) → 0.
Since every sequence α j → 0 + \alpha_{j}\to0^{+} α j → 0 + admits a subsequence along which the transport energies converge to g L g_{L} g L and the rescaled source penalties converge to zero, both limits in (7) hold as α → 0 + \alpha\to0^{+} α → 0 + for any choice of minimizers. Finally,
min A L ρ 1 E α − b L α = G ( M α , f α ) + ∥ f α ∥ L 2 ( 0 , 1 ) 2 − b L α → g L , \min_{A_{L}^{\rho_{1}}} E_{\alpha}-\frac{b_{L}}{\alpha}
=G(M_{\alpha},f_{\alpha})+\frac{\lVert f_{\alpha}\rVert_{L^{2}(0,1)}^{2}-b_{L}}{\alpha}\to g_{L}, A L ρ 1 min E α − α b L = G ( M α , f α ) + α ∥ f α ∥ L 2 ( 0 , 1 ) 2 − b L → g L ,
which proves (6) .
Corollary 5.2 (Convergence of reconstructed paths). Let α j → 0 + \alpha_{j}\to0^{+} α j → 0 + and let ( M α j , f α j ) (M_{\alpha_{j}},f_{\alpha_{j}}) ( M α j , f α j ) be minimizers as in Theorem 1.5. Along the subsequence provided by that theorem, let ( ρ α j , v α j , f α j ) (\rho_{\alpha_{j}},v_{\alpha_{j}},f_{\alpha_{j}}) ( ρ α j , v α j , f α j ) and ( ρ , v , f ) (\rho,v,f) ( ρ , v , f ) be the reconstructed triples. Then v α j → v v_{\alpha_{j}}\to v v α j → v uniformly on [ 0 , 1 ] × Ω ‾ [0,1]\times\overline{\Omega} [ 0 , 1 ] × Ω , and, for every ψ ∈ C ( Ω ‾ ) \psi\in C(\overline{\Omega}) ψ ∈ C ( Ω ) ,
sup t ∈ [ 0 , 1 ] ∣ ∫ Ω ψ ( y ) ( ρ α j ( t , y ) − ρ ( t , y ) ) d y ∣ ⟶ 0. \sup_{t\in[0,1]}\left|\int_{\Omega}\psi(y)(\rho_{\alpha_{j}}(t,y)-\rho(t,y))\,dy\right|\longrightarrow0. t ∈ [ 0 , 1 ] sup ∫ Ω ψ ( y ) ( ρ α j ( t , y ) − ρ ( t , y )) d y ⟶ 0. Moreover, ρ α j ⇀ ∗ ρ \rho_{\alpha_{j}}\rightharpoonup^{*}\rho ρ α j ⇀ ∗ ρ in L ∞ ( ( 0 , 1 ) × Ω ) L^{\infty}((0,1)\times\Omega) L ∞ (( 0 , 1 ) × Ω ) , and ( ρ , v , f ) (\rho,v,f) ( ρ , v , f ) minimizes
∫ 0 1 ∫ Ω ρ ∣ v ∣ 2 d y d t \int_{0}^{1}\int_{\Omega}\rho|v|^{2}\,dy\,dt ∫ 0 1 ∫ Ω ρ ∣ v ∣ 2 d y d t among triples in D L ρ 0 , ρ 1 \mathcal{D}_{L}^{\rho_{0},\rho_{1}} D L ρ 0 , ρ 1 satisfying ∥ f ∥ L 2 ( 0 , 1 ) 2 = b L \lVert f\rVert_{L^{2}(0,1)}^{2}=b_{L} ∥ f ∥ L 2 ( 0 , 1 ) 2 = b L .
Proof. The convergence follows from Proposition 5.1. The minimizing property follows from G ( M , f ) = g L G(M,f)=g_L G ( M , f ) = g L and the energy-preserving correspondence. □
We now identify when the preceding limit is a constant-source problem. The mass constraint gives b L ≥ a 2 b_L\ge a^2 b L ≥ a 2 . Since b L b_L b L is attained, equality holds if and only if there exists ( M ∗ , a ) ∈ A L ρ 1 (M_*,a)\in\mathcal{A}^{\rho_1}_L ( M ∗ , a ) ∈ A L ρ 1 . In that case, S L ρ 1 = { ( M , a ) ∈ A L ρ 1 } \mathcal{S}^{\rho_1}_L=\{(M,a)\in\mathcal{A}^{\rho_1}_L\} S L ρ 1 = {( M , a ) ∈ A L ρ 1 } . Thus the constant-source feasibility assumption means that the lower bound imposed by the total mass difference can be attained while satisfying the endpoint and geometric constraints. Retaining the notation for the constant-source problem, define
F α , a ( M , f ) : = { E α ( M , f ) − a 2 / α , ( M , f ) ∈ A L ρ 1 , + ∞ , otherwise . \mathcal{F}_{\alpha,a}(M,f):=
\begin{cases}
E_\alpha(M,f)-a^2/\alpha, & (M,f)\in\mathcal{A}^{\rho_1}_L,\\
+\infty, & \text{otherwise}.
\end{cases} F α , a ( M , f ) := { E α ( M , f ) − a 2 / α , + ∞ , ( M , f ) ∈ A L ρ 1 , otherwise .
and
F 0 , a ( M , f ) : = { G ( M , a ) , f = a a.e. and ( M , a ) ∈ A L ρ 1 , + ∞ , otherwise . \mathcal{F}_{0,a}(M,f):=
\begin{cases}
G(M,a), & f=a\ \text{a.e. and }(M,a)\in\mathcal{A}^{\rho_1}_L,\\
+\infty, & \text{otherwise}.
\end{cases} F 0 , a ( M , f ) := { G ( M , a ) , + ∞ , f = a a.e. and ( M , a ) ∈ A L ρ 1 , otherwise .
Corollary 5.3 (Constant-source limit). Assume that there exists ( M ∗ , a ) ∈ A L ρ 1 (M_*,a)\in\mathcal{A}^{\rho_1}_L ( M ∗ , a ) ∈ A L ρ 1 . Then
F α , a ⟶ Γ F 0 , a as α → 0 + \mathcal{F}_{\alpha,a}\mathrel{\mathop{\longrightarrow}^{\Gamma}}\mathcal{F}_{0,a}
\qquad\text{as }\alpha\to0^+ F α , a ⟶ Γ F 0 , a as α → 0 + in the sequential sense with respect to τ \tau τ . The constant-source problem admits a minimizer, and, writing
m a : = min { G ( M , a ) : ( M , a ) ∈ A L ρ 1 } , m_a:=\min\{G(M,a):(M,a)\in\mathcal{A}^{\rho_1}_L\}, m a := min { G ( M , a ) : ( M , a ) ∈ A L ρ 1 } , we have
min A L ρ 1 E α = a 2 α + m a + o ( 1 ) . (20) \min_{\mathcal{A}^{\rho_1}_L}E_\alpha=\frac{a^2}{\alpha}+m_a+o(1). \tag*{(20)} A L ρ 1 min E α = α a 2 + m a + o ( 1 ) . (20) For every sequence α j → 0 + \alpha_j\to0^+ α j → 0 + and corresponding minimizers ( M α j , f α j ) (M_{\alpha_j},f_{\alpha_j}) ( M α j , f α j ) , there exist a subsequence, not relabeled, and a minimizer ( M , a ) (M,a) ( M , a ) of F 0 , a \mathcal{F}_{0,a} F 0 , a such that
M α j ⇀ M in W 1 , n ( Ω ; R n ) , M α j → M uniformly on Ω ‾ , M_{\alpha_j}\rightharpoonup M\quad\text{in }W^{1,n}(\Omega;\mathbb{R}^n),
\qquad
M_{\alpha_j}\to M\quad\text{uniformly on }\overline{\Omega}, M α j ⇀ M in W 1 , n ( Ω ; R n ) , M α j → M uniformly on Ω , and
f α j → a strongly in L 2 ( 0 , 1 ) . f_{\alpha_j}\to a\quad\text{strongly in }L^2(0,1). f α j → a strongly in L 2 ( 0 , 1 ) . Moreover, for any choice of minimizers,
G ( M α , f α ) → m a , 1 α ∥ f α − a ∥ L 2 ( 0 , 1 ) 2 → 0. (21) G(M_\alpha,f_\alpha)\to m_a,
\qquad
\frac{1}{\alpha}\lVert f_\alpha-a\rVert_{L^2(0,1)}^2\to0. \tag*{(21)} G ( M α , f α ) → m a , α 1 ∥ f α − a ∥ L 2 ( 0 , 1 ) 2 → 0. (21) In particular, ∥ f α − a ∥ L 2 ( 0 , 1 ) = o ( α ) \lVert f_\alpha-a\rVert_{L^2(0,1)}=o(\sqrt{\alpha}) ∥ f α − a ∥ L 2 ( 0 , 1 ) = o ( α ) .
Proof. The assumption gives b L = a 2 b_L=a^2 b L = a 2 and g L = m a g_L=m_a g L = m a . Hence F ^ α = F α , a \widehat{\mathcal{F}}_\alpha=\mathcal{F}_{\alpha,a} F α = F α , a and F ^ 0 = F 0 , a \widehat{\mathcal{F}}_0=\mathcal{F}_{0,a} F 0 = F 0 , a . All convergence assertions follow from Theorem 1.5 and the identity ∥ f ∥ L 2 ( 0 , 1 ) 2 − a 2 = ∥ f − a ∥ L 2 ( 0 , 1 ) 2 \lVert f\rVert_{L^2(0,1)}^2-a^2=\lVert f-a\rVert_{L^2(0,1)}^2 ∥ f ∥ L 2 ( 0 , 1 ) 2 − a 2 = ∥ f − a ∥ L 2 ( 0 , 1 ) 2 . In particular, comparison with a constant-source minimizer ( M ^ , a ) (\widehat{M},a) ( M , a ) gives
0 ≤ G ( M α , f α ) ≤ F α , a ( M α , f α ) ≤ F α , a ( M ^ , a ) = m a . (22) 0\le G(M_\alpha,f_\alpha)\le\mathcal{F}_{\alpha,a}(M_\alpha,f_\alpha)\le\mathcal{F}_{\alpha,a}(\widehat{M},a)=m_a. \tag*{(22)} 0 ≤ G ( M α , f α ) ≤ F α , a ( M α , f α ) ≤ F α , a ( M , a ) = m a . (22) Both middle terms converge to m a m_a m a , so their difference is α − 1 ∥ f α − a ∥ L 2 ( 0 , 1 ) 2 → 0 \alpha^{-1}\lVert f_\alpha-a\rVert_{L^2(0,1)}^2\to0 α − 1 ∥ f α − a ∥ L 2 ( 0 , 1 ) 2 → 0 . □
Proposition 5.2 (A lower bound for the excess source cost). Assume that A L ρ 1 ≠ ∅ \mathcal{A}^{\rho_1}_{L}\ne\varnothing A L ρ 1 = ∅ and a < 0 a<0 a < 0 . Set
c 0 : = ess inf Ω ρ 0 , c 1 : = min Ω ρ 1 . c_0:=\operatorname{ess\,inf}_{\Omega}\rho_0,\qquad c_1:=\min_{\Omega}\rho_1. c 0 := ess inf Ω ρ 0 , c 1 := Ω min ρ 1 . Then
b L − a 2 ≥ L − 2 n [ L − n ( c 1 + ∣ a ∣ ) − c 0 ] + 2 , b_L-a^2\ge L^{-2n}\left[L^{-n}(c_1+|a|)-c_0\right]_+^2, b L − a 2 ≥ L − 2 n [ L − n ( c 1 + ∣ a ∣ ) − c 0 ] + 2 , where [ r ] + : = max { r , 0 } [r]_+:=\max\{r,0\} [ r ] + := max { r , 0 } . In particular, a necessary condition for the existence of a constant-source admissible pair is
c 0 ≥ L − n ( c 1 + ∣ a ∣ ) . c_0\ge L^{-n}(c_1+|a|). c 0 ≥ L − n ( c 1 + ∣ a ∣ ) .
Proof. Let ( M , f ) ∈ A L ρ 1 (M,f)\in\mathcal{A}^{\rho_1}_{L} ( M , f ) ∈ A L ρ 1 and write h : = f − a h:=f-a h := f − a . The endpoint constraint gives
ρ 0 = ρ 1 ( M ) J 1 [ M ] + ∣ a ∣ ∫ 0 1 J t [ M ] d t − ∫ 0 1 h ( t ) J t [ M ] d t . \rho_0=\rho_1(M)J_1[M]+|a|\int_0^1J_t[M]\,\mathrm{d}t-\int_0^1h(t)J_t[M]\,\mathrm{d}t. ρ 0 = ρ 1 ( M ) J 1 [ M ] + ∣ a ∣ ∫ 0 1 J t [ M ] d t − ∫ 0 1 h ( t ) J t [ M ] d t .
Using L − n ≤ J t [ M ] ≤ L n L^{-n}\le J_t[M]\le L^n L − n ≤ J t [ M ] ≤ L n and Cauchy–Schwarz, we obtain
ρ 0 ≥ L − n ( c 1 + ∣ a ∣ ) − L n ∥ h ∥ L 2 ( 0 , 1 ) a.e. in Ω . \rho_0\ge L^{-n}(c_1+|a|)-L^n\|h\|_{L^2(0,1)}\quad\text{a.e. in }\Omega. ρ 0 ≥ L − n ( c 1 + ∣ a ∣ ) − L n ∥ h ∥ L 2 ( 0 , 1 ) a.e. in Ω.
Taking the essential infimum yields
∥ f − a ∥ L 2 ( 0 , 1 ) ≥ L − n [ L − n ( c 1 + ∣ a ∣ ) − c 0 ] + . \|f-a\|_{L^2(0,1)}\ge L^{-n}\left[L^{-n}(c_1+|a|)-c_0\right]_+. ∥ f − a ∥ L 2 ( 0 , 1 ) ≥ L − n [ L − n ( c 1 + ∣ a ∣ ) − c 0 ] + .
Since
∥ f ∥ L 2 ( 0 , 1 ) 2 = a 2 + ∥ f − a ∥ L 2 ( 0 , 1 ) 2 , \|f\|_{L^2(0,1)}^2=a^2+\|f-a\|_{L^2(0,1)}^2, ∥ f ∥ L 2 ( 0 , 1 ) 2 = a 2 + ∥ f − a ∥ L 2 ( 0 , 1 ) 2 ,
taking the infimum over A L ρ 1 \mathcal{A}^{\rho_1}_{L} A L ρ 1 proves the assertion. □
When the lower bound in Proposition 5.2 is positive, the prescribed densities and flow bounds force every admissible source to deviate from its mean. Consequently,
min A L ρ 1 E α − a 2 α = b L − a 2 α + g L + o ( 1 ) ⟶ + ∞ . \min_{\mathcal{A}^{\rho_1}_{L}}E_\alpha-\frac{a^2}{\alpha}=\frac{b_L-a^2}{\alpha}+g_L+o(1)\longrightarrow+\infty. A L ρ 1 min E α − α a 2 = α b L − a 2 + g L + o ( 1 ) ⟶ + ∞.
Thus the total mass difference alone does not determine the leading optimal cost. Example 6.4 shows that this phenomenon occurs even when the reconstruction class is nonempty.
Nonemptiness of Admissible Classes
A local sufficient condition for nonemptiness
The following result establishes local feasibility when both endpoint densities are Hölder continuous. The construction applies to perturbations of a fixed reference pair ( ρ ‾ − a , ρ ‾ ) (\overline{\rho}-a,\overline{\rho}) ( ρ − a , ρ ) . The reference density need not be constant, and the distortion and flow bounds are prescribed in advance.
Proposition 6.1 (Local feasibility for Hölder densities). Let n ≥ 2 n \ge2 n ≥ 2 , K > n n / 2 K > n^{n/2} K > n n /2 , L > 1 L > 1 L > 1 and 0 < β < γ < 1 0 < \beta< \gamma< 1 0 < β < γ < 1 , and let Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n be a bounded convex domain with C 3 , β C^{3,\beta} C 3 , β boundary. Fix ρ ‾ ∈ C 1 , γ ( Ω ‾ ) \overline{\rho} \in C^{1,\gamma}(\overline{\Omega}) ρ ∈ C 1 , γ ( Ω ) and a ∈ R a \in\mathbb{R} a ∈ R such that ρ ‾ ≥ 0 \overline{\rho} \ge0 ρ ≥ 0 , ρ ‾ − a ≥ 0 \overline{\rho}-a \ge0 ρ − a ≥ 0 and min Ω ‾ ( ρ ‾ − a 2 ) > 0 \min_{\overline{\Omega}}(\overline{\rho}-\frac{a}{2}) > 0 min Ω ( ρ − 2 a ) > 0 . There exist ε , C > 0 \varepsilon,C>0 ε , C > 0 , depending only on Ω \Omega Ω , n n n , β \beta β , γ \gamma γ , ρ ‾ \overline{\rho} ρ , a a a , K K K , L L L , such that the following holds. Suppose that 0 ≤ ρ 0 , ρ 1 ∈ C 0 , β ( Ω ‾ ) 0 \le\rho_{0},\rho_{1} \in C^{0,\beta}(\overline{\Omega}) 0 ≤ ρ 0 , ρ 1 ∈ C 0 , β ( Ω ) and ∫ Ω ( ρ 1 − ρ 0 ) d x = a ∣ Ω ∣ \int_{\Omega}(\rho_{1}-\rho_{0})\,dx=a|\Omega| ∫ Ω ( ρ 1 − ρ 0 ) d x = a ∣Ω∣ , and
d : = ∥ ρ 0 − ρ ‾ + a ∥ C 0 , β ( Ω ‾ ) + ∥ ρ 1 − ρ ‾ ∥ C 0 , β ( Ω ‾ ) < ε . d:=\|\rho_{0}-\overline{\rho}+a\|_{C^{0,\beta}(\overline{\Omega})}+\|\rho_{1}-\overline{\rho}\|_{C^{0,\beta}(\overline{\Omega})}<\varepsilon. d := ∥ ρ 0 − ρ + a ∥ C 0 , β ( Ω ) + ∥ ρ 1 − ρ ∥ C 0 , β ( Ω ) < ε . Then there exists M ∈ C 1 , β ( Ω ‾ ; R n ) M\in C^{1,\beta}(\overline{\Omega};\mathbb{R}^{n}) M ∈ C 1 , β ( Ω ; R n ) such that M = Id M=\operatorname{Id} M = Id on ∂ Ω \partial\Omega ∂ Ω , ( M , a ) ∈ V K ρ 1 ∩ A L ρ 1 (M,a)\in\mathcal{V}_{K}^{\rho_{1}}\cap\mathcal{A}_{L}^{\rho_{1}} ( M , a ) ∈ V K ρ 1 ∩ A L ρ 1 , and
∥ M − Id ∥ C 1 , β ( Ω ‾ ) ≤ C d . \|M-\operatorname{Id}\|_{C^{1,\beta}(\overline{\Omega})}\le Cd. ∥ M − Id ∥ C 1 , β ( Ω ) ≤ C d .
Proof. Step 1. Fix 0 < η < β 0<\eta<\beta 0 < η < β . For s ∈ { η , β } s\in\{\eta,\beta\} s ∈ { η , β } , set
Z s : = { g ∈ C 0 , s ( Ω ‾ ) : ∫ Ω g d x = 0 } , Z_{s}:=\left\{g\in C^{0,s}(\overline{\Omega}):\int_{\Omega}g\,dx=0\right\}, Z s := { g ∈ C 0 , s ( Ω ) : ∫ Ω g d x = 0 } ,
and
W s : = { u ∈ C 1 , s ( Ω ‾ ; R n ) : u = 0 on ∂ Ω } . W_{s}:=\left\{u\in C^{1,s}(\overline{\Omega};\mathbb{R}^{n}):u=0\text{ on }\partial\Omega\right\}. W s := { u ∈ C 1 , s ( Ω ; R n ) : u = 0 on ∂ Ω } .
By the construction in [8 Theorem 2 and Remark (ii)] , there exists a linear operator B \mathcal{B} B such that, for each s ∈ { η , β } s\in\{\eta,\beta\} s ∈ { η , β } and g ∈ Z s g\in Z_{s} g ∈ Z s , the vector field B g ∈ W s \mathcal{B}g\in W_{s} B g ∈ W s satisfies
div ( B g ) = g in Ω , B g = 0 on ∂ Ω , \operatorname{div}(\mathcal{B}g)=g\quad\text{in }\Omega,\qquad\mathcal{B}g=0\quad\text{on }\partial\Omega, div ( B g ) = g in Ω , B g = 0 on ∂ Ω ,
with the estimate ∥ B g ∥ C 1 , s ≤ C s ∥ g ∥ C 0 , s \|\mathcal{B}g\|_{C^{1,s}}\le C_{s}\|g\|_{C^{0,s}} ∥ B g ∥ C 1 , s ≤ C s ∥ g ∥ C 0 , s , where C s C_{s} C s depends only on Ω \Omega Ω , n n n , and s s s , and is independent of g g g . The same operator is used at both exponents. Write σ : = ρ ‾ − a 2 \sigma:=\overline{\rho}-\frac{a}{2} σ := ρ − 2 a . Since σ \sigma σ is bounded away from zero, the operator B σ g : = B g σ \mathcal{B}_{\sigma}g:=\frac{\mathcal{B}g}{\sigma} B σ g := σ B g satisfies
div ( σ B σ g ) = g and ∥ B σ g ∥ C 1 , s ≤ C ~ s ∥ g ∥ C 0 , s , for s ∈ { η , β } , \operatorname{div}(\sigma\mathcal{B}_{\sigma}g)=g\quad\text{and}\quad\|\mathcal{B}_{\sigma}g\|_{C^{1,s}}\le\widetilde{C}_{s}\|g\|_{C^{0,s}},\qquad\text{for }s\in\{\eta,\beta\}, div ( σ B σ g ) = g and ∥ B σ g ∥ C 1 , s ≤ C s ∥ g ∥ C 0 , s , for s ∈ { η , β } ,
where C ~ s \widetilde{C}_{s} C s depends only on Ω \Omega Ω , n n n , s s s , ∥ σ − 1 ∥ C 1 , s \|\sigma^{-1}\|_{C^{1,s}} ∥ σ − 1 ∥ C 1 , s , and the chosen Hölder norms. In particular, it is independent of g g g .
Step 2. Choose 0 < θ < 1 0<\theta<1 0 < θ < 1 such that
1 − θ > L − 1 , 1 + θ < L , n n / 2 ( 1 + θ 1 − θ ) n < K . 1-\theta>L^{-1},\qquad1+\theta<L,\qquad n^{n/2}\left(\frac{1+\theta}{1-\theta}\right)^{n}<K. 1 − θ > L − 1 , 1 + θ < L , n n /2 ( 1 − θ 1 + θ ) n < K .
If u ∈ W β u\in W_{\beta} u ∈ W β and ∥ ∇ u ∥ L ∞ ≤ θ \|\nabla u\|_{L^{\infty}}\le\theta ∥∇ u ∥ L ∞ ≤ θ , its zero extension u ~ \widetilde{u} u to R n \mathbb{R}^{n} R n is θ \theta θ -Lipschitz. Thus, for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] ,
( 1 − θ ) ∣ x − y ∣ ≤ ∣ x + t u ~ ( x ) − y − t u ~ ( y ) ∣ ≤ ( 1 + θ ) ∣ x − y ∣ . (1-\theta)|x-y|\le\left|x+t\widetilde{u}(x)-y-t\widetilde{u}(y)\right|\le(1+\theta)|x-y|. ( 1 − θ ) ∣ x − y ∣ ≤ ∣ x + t u ( x ) − y − t u ( y ) ∣ ≤ ( 1 + θ ) ∣ x − y ∣.
For x ∈ R n x\in\mathbb{R}^{n} x ∈ R n , let F t ( x ) : = x + t u ~ ( x ) F_{t}(x):=x+t\widetilde{u}(x) F t ( x ) := x + t u ( x ) . For every fixed t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] and y ∈ R n y\in\mathbb{R}^{n} y ∈ R n , the map x ↦ y − t u ~ ( x ) x\mapsto y-t\widetilde{u}(x) x ↦ y − t u ( x ) is a contraction on R n \mathbb{R}^{n} R n , with Lipschitz constant at most t θ < 1 t\theta<1 tθ < 1 . It therefore has a unique fixed point x x x , which satisfies F t ( x ) = y F_{t}(x)=y F t ( x ) = y . Thus F t F_{t} F t is bijective on R n \mathbb{R}^{n} R n . Since u ~ = 0 \widetilde{u}=0 u = 0 on R n ∖ Ω \mathbb{R}^{n}\setminus\Omega R n ∖ Ω , every point outside Ω \Omega Ω is fixed by F t F_{t} F t . If x ∈ Ω x\in\Omega x ∈ Ω and F t ( x ) = y ∉ Ω F_{t}(x)=y\notin\Omega F t ( x ) = y ∈ / Ω , then F t ( y ) = y = F t ( x ) F_{t}(y)=y=F_{t}(x) F t ( y ) = y = F t ( x ) , contradicting injectivity. Hence F t ( Ω ) ⊂ Ω F_{t}(\Omega)\subset\Omega F t ( Ω ) ⊂ Ω . Conversely, the preimage of any y ∈ Ω y\in\Omega y ∈ Ω cannot lie outside Ω \Omega Ω , where F t F_{t} F t is the identity. Therefore F t ( Ω ) = Ω F_t(\Omega)=\Omega F t ( Ω ) = Ω . Since F t F_t F t also fixes ∂ Ω \partial\Omega ∂ Ω pointwise, ( Id + t u ) ( Ω ‾ ) = Ω ‾ (\operatorname{Id}+tu)(\overline{\Omega})=\overline{\Omega} ( Id + t u ) ( Ω ) = Ω . Since ∥ ∇ u ∥ L ∞ ( Ω ) ≤ θ < 1 \|\nabla u\|_{L^\infty(\Omega)}\leq\theta<1 ∥∇ u ∥ L ∞ ( Ω ) ≤ θ < 1 , we have ∣ ( I + t ∇ u ( x ) ) ξ ∣ ≥ ( 1 − t θ ) ∣ ξ ∣ ≥ ( 1 − θ ) ∣ ξ ∣ \lvert(I+t\nabla u(x))\xi\rvert\geq(1-t\theta)\lvert\xi\rvert\geq(1-\theta)\lvert\xi\rvert ∣( I + t ∇ u ( x )) ξ ∣ ≥ ( 1 − tθ ) ∣ ξ ∣ ≥ ( 1 − θ ) ∣ ξ ∣ for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] and ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n . Thus I + t ∇ u ( x ) I+t\nabla u(x) I + t ∇ u ( x ) is invertible for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . Thus,
det ( I + t ∇ u ) > 0 and ∫ Ω det ( I + t ∇ u ) d x = ∣ Ω ∣ . \det(I+t\nabla u)>0\qquad\text{and}\qquad\int_{\Omega}\det(I+t\nabla u)\,dx=\lvert\Omega\rvert. det ( I + t ∇ u ) > 0 and ∫ Ω det ( I + t ∇ u ) d x = ∣ Ω ∣ .
Step 3. For every u ∈ W β u\in W_\beta u ∈ W β satisfying ∥ u ∥ C 1 , β ( Ω ‾ ) < r \|u\|_{C^{1,\beta}(\overline{\Omega})}<r ∥ u ∥ C 1 , β ( Ω ) < r , where r > 0 r>0 r > 0 is sufficiently small, define
H ( u ) : = ρ ‾ ( Id + u ) det ( I + ∇ u ) − a ∫ 0 1 det ( I + t ∇ u ) d t − ( ρ ‾ − a ) . \mathcal{H}(u):=\overline{\rho}(\operatorname{Id}+u)\det(I+\nabla u)-a\int_0^1\det(I+t\nabla u)\,dt-(\overline{\rho}-a). H ( u ) := ρ ( Id + u ) det ( I + ∇ u ) − a ∫ 0 1 det ( I + t ∇ u ) d t − ( ρ − a ) .
The change-of-variables formula gives ∫ Ω H ( u ) d x = 0 \int_{\Omega}\mathcal{H}(u)\,dx=0 ∫ Ω H ( u ) d x = 0 . Its linearization at zero is
D H ( 0 ) [ u ] = ∇ ρ ‾ ⋅ u + ( ρ ‾ − a 2 ) div u = div ( σ u ) . D\mathcal{H}(0)[u]=\nabla\overline{\rho}\cdot u+\left(\overline{\rho}-\frac{a}{2}\right)\operatorname{div}u=\operatorname{div}(\sigma u). D H ( 0 ) [ u ] = ∇ ρ ⋅ u + ( ρ − 2 a ) div u = div ( σ u ) .
Define the nonlinear remainder N ( u ) : = H ( u ) − div ( σ u ) \mathcal{N}(u):=\mathcal{H}(u)-\operatorname{div}(\sigma u) N ( u ) := H ( u ) − div ( σ u ) . Both terms belong to C 0 , β ( Ω ‾ ) C^{0,\beta}(\overline{\Omega}) C 0 , β ( Ω ) and have zero mean, so N ( u ) ∈ Z β \mathcal{N}(u)\in Z_\beta N ( u ) ∈ Z β . We next estimate its C 0 , β C^{0,\beta} C 0 , β norm. Throughout the following estimates, all Hölder norms are taken over Ω ‾ \overline{\Omega} Ω . Set h s ( x ) : = ∇ ρ ‾ ( x + s u ( x ) ) − ∇ ρ ‾ ( x ) h_s(x):=\nabla\overline{\rho}(x+su(x))-\nabla\overline{\rho}(x) h s ( x ) := ∇ ρ ( x + s u ( x )) − ∇ ρ ( x ) . Since ∇ ρ ‾ ∈ C 0 , γ \nabla\overline{\rho}\in C^{0,\gamma} ∇ ρ ∈ C 0 , γ , ∥ h s ∥ C 0 ≤ [ ∇ ρ ‾ ] C 0 , γ s γ ∥ u ∥ C 0 \|h_s\|_{C^0}\leq[\nabla\overline{\rho}]_{C^{0,\gamma}}s^\gamma\|u\|_{C^0} ∥ h s ∥ C 0 ≤ [ ∇ ρ ] C 0 , γ s γ ∥ u ∥ C 0 . Moreover, the maps Id + s u \operatorname{Id}+su Id + s u have uniformly bounded Lipschitz constants. Hence
∣ h s ( x ) − h s ( y ) ∣ ≤ [ ∇ ρ ‾ ] C 0 , γ ( ∣ x + s u ( x ) − y − s u ( y ) ∣ γ + ∣ x − y ∣ γ ) ≤ C ∣ x − y ∣ γ , \lvert h_s(x)-h_s(y)\rvert\leq[\nabla\overline{\rho}]_{C^{0,\gamma}}\left(\lvert x+su(x)-y-su(y)\rvert^\gamma+\lvert x-y\rvert^\gamma\right)
\leq C\lvert x-y\rvert^\gamma, ∣ h s ( x ) − h s ( y )∣ ≤ [ ∇ ρ ] C 0 , γ ( ∣ x + s u ( x ) − y − s u ( y ) ∣ γ + ∣ x − y ∣ γ ) ≤ C ∣ x − y ∣ γ ,
and therefore ∥ h s ∥ C 0 , γ ≤ C \|h_s\|_{C^{0,\gamma}}\leq C ∥ h s ∥ C 0 , γ ≤ C . The constants are independent of s s s and of u u u in the chosen neighborhood. Since 0 < β < γ 0<\beta<\gamma 0 < β < γ , Hölder interpolation gives
∥ h s ∥ C 0 , β ≤ C ∥ h s ∥ C 0 1 − β / γ ∥ h s ∥ C 0 , γ β / γ ≤ C s γ − β ∥ u ∥ C 0 γ − β . \|h_s\|_{C^{0,\beta}}\leq C\|h_s\|_{C^0}^{1-\beta/\gamma}\|h_s\|_{C^{0,\gamma}}^{\beta/\gamma}\leq Cs^{\gamma-\beta}\|u\|_{C^0}^{\gamma-\beta}. ∥ h s ∥ C 0 , β ≤ C ∥ h s ∥ C 0 1 − β / γ ∥ h s ∥ C 0 , γ β / γ ≤ C s γ − β ∥ u ∥ C 0 γ − β .
Taylor’s formula in integral form yields ρ ‾ ( Id + u ) − ρ ‾ − ∇ ρ ‾ ⋅ u = ∫ 0 1 h s ⋅ u d s \overline{\rho}(\operatorname{Id}+u)-\overline{\rho}-\nabla\overline{\rho}\cdot u=\int_0^1h_s\cdot u\,ds ρ ( Id + u ) − ρ − ∇ ρ ⋅ u = ∫ 0 1 h s ⋅ u d s . Using the product estimate in C 0 , β C^{0,\beta} C 0 , β , we obtain
∥ ρ ‾ ( Id + u ) − ρ ‾ − ∇ ρ ‾ ⋅ u ∥ C 0 , β ≤ C ∥ u ∥ C 0 , β ∫ 0 1 ∥ h s ∥ C 0 , β d s ≤ C ∥ u ∥ C 0 , β ∥ u ∥ C 0 γ − β ∫ 0 1 s γ − β d s ≤ C ∥ u ∥ C 1 , β 1 + γ − β . \begin{aligned}
\|\overline{\rho}(\operatorname{Id}+u)-\overline{\rho}-\nabla\overline{\rho}\cdot u\|_{C^{0,\beta}}
\leq C\|u\|_{C^{0,\beta}}\int_0^1\|h_s\|_{C^{0,\beta}}\,ds \\
&\leq C\|u\|_{C^{0,\beta}}\|u\|_{C^0}^{\gamma-\beta}\int_0^1s^{\gamma-\beta}\,ds \\
&\leq C\|u\|_{C^{1,\beta}}^{1+\gamma-\beta}.
\end{aligned} ∥ ρ ( Id + u ) − ρ − ∇ ρ ⋅ u ∥ C 0 , β ≤ C ∥ u ∥ C 0 , β ∫ 0 1 ∥ h s ∥ C 0 , β d s ≤ C ∥ u ∥ C 0 , β ∥ u ∥ C 0 γ − β ∫ 0 1 s γ − β d s ≤ C ∥ u ∥ C 1 , β 1 + γ − β .
Write δ : = γ − β ∈ ( 0 , 1 ) \delta:=\gamma-\beta\in(0,1) δ := γ − β ∈ ( 0 , 1 ) and set q : = ∥ u ∥ C 1 , β q:=\|u\|_{C^{1,\beta}} q := ∥ u ∥ C 1 , β . We work in a fixed neighborhood of zero on which q ≤ r 0 ≤ 1 q\leq r_0\leq1 q ≤ r 0 ≤ 1 . All constants below are independent of u u u in this neighborhood. For t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , define P t ( u ) : = det ( I + t ∇ u ) − 1 − t div u P_t(u):=\det(I+t\nabla u)-1-t\operatorname{div}u P t ( u ) := det ( I + t ∇ u ) − 1 − t div u . Every term in P t ( u ) P_t(u) P t ( u ) has degree at least two in ∇ u \nabla u ∇ u . The product estimate in C 0 , β C^{0,\beta} C 0 , β therefore gives
sup t ∈ [ 0 , 1 ] ∥ P t ( u ) ∥ C 0 , β ≤ C ∑ k = 2 n ∥ ∇ u ∥ C 0 , β k ≤ C q 2 . \sup_{t\in[0,1]}\|P_t(u)\|_{C^{0,\beta}}\leq C\sum_{k=2}^n\|\nabla u\|_{C^{0,\beta}}^k\leq Cq^2. t ∈ [ 0 , 1 ] sup ∥ P t ( u ) ∥ C 0 , β ≤ C k = 2 ∑ n ∥∇ u ∥ C 0 , β k ≤ C q 2 .
Set e ( u ) : = ρ ‾ ( Id + u ) − ρ ‾ − ∇ ρ ‾ ⋅ u e(u):=\overline{\rho}(\operatorname{Id}+u)-\overline{\rho}-\nabla\overline{\rho}\cdot u e ( u ) := ρ ( Id + u ) − ρ − ∇ ρ ⋅ u . The preceding Taylor estimate yields ∥ e ( u ) ∥ C 0 , β ≤ C q 1 + δ \|e(u)\|_{C^{0,\beta}}\leq Cq^{1+\delta} ∥ e ( u ) ∥ C 0 , β ≤ C q 1 + δ . Consequently, ∥ ρ ‾ ( Id + u ) − ρ ‾ ∥ C 0 , β ≤ C q \|\overline{\rho}(\operatorname{Id}+u)-\overline{\rho}\|_{C^{0,\beta}}\leq Cq ∥ ρ ( Id + u ) − ρ ∥ C 0 , β ≤ C q and ∥ ρ ‾ ( Id + u ) ∥ C 0 , β ≤ C \|\overline{\rho}(\operatorname{Id}+u)\|_{C^{0,\beta}}\leq C ∥ ρ ( Id + u ) ∥ C 0 , β ≤ C . Using det ( I + ∇ u ) = 1 + div u + P 1 ( u ) \det(I+\nabla u)=1+\operatorname{div}u+P_1(u) det ( I + ∇ u ) = 1 + div u + P 1 ( u ) and
∫ 0 1 det ( I + t ∇ u ) d t = 1 + 1 2 div u + ∫ 0 1 P t ( u ) d t , \int_0^1\det(I+t\nabla u)\,dt=1+\frac{1}{2}\operatorname{div}u+\int_0^1P_t(u)\,dt, ∫ 0 1 det ( I + t ∇ u ) d t = 1 + 2 1 div u + ∫ 0 1 P t ( u ) d t ,
we obtain
N ( u ) = e ( u ) + ( ρ ‾ ( Id + u ) − ρ ‾ ) div u + ρ ‾ ( Id + u ) P 1 ( u ) − a ∫ 0 1 P t ( u ) d t . \mathcal{N}(u)=e(u)+\left(\overline{\rho}(\operatorname{Id}+u)-\overline{\rho}\right)\operatorname{div}u
+\overline{\rho}(\operatorname{Id}+u)P_1(u)-a\int_0^1 P_t(u)\,\mathrm{d}t. N ( u ) = e ( u ) + ( ρ ( Id + u ) − ρ ) div u + ρ ( Id + u ) P 1 ( u ) − a ∫ 0 1 P t ( u ) d t .
Hence
∥ N ( u ) ∥ C 0 , β ≤ C ( q 1 + δ + q 2 ) ≤ C q 1 + δ , \|\mathcal{N}(u)\|_{C^{0,\beta}}\leq C\left(q^{1+\delta}+q^2\right)\leq Cq^{1+\delta}, ∥ N ( u ) ∥ C 0 , β ≤ C ( q 1 + δ + q 2 ) ≤ C q 1 + δ ,
because q ≤ 1 q\leq1 q ≤ 1 and 1 + δ < 2 1+\delta<2 1 + δ < 2 . Thus
∥ N ( u ) ∥ C 0 , β ≤ C ∥ u ∥ C 1 , β 1 + γ − β . \|\mathcal{N}(u)\|_{C^{0,\beta}}\leq C\|u\|_{C^{1,\beta}}^{1+\gamma-\beta}. ∥ N ( u ) ∥ C 0 , β ≤ C ∥ u ∥ C 1 , β 1 + γ − β .
Step 4. Now write p : = ρ 1 − ρ ‾ p:=\rho_1-\overline{\rho} p := ρ 1 − ρ and h : = ρ 0 − ρ 1 + a h:=\rho_0-\rho_1+a h := ρ 0 − ρ 1 + a . The mass condition gives h ∈ Z β h\in Z_\beta h ∈ Z β , ∥ h ∥ C 0 , β ≤ d \|h\|_{C^{0,\beta}}\leq d ∥ h ∥ C 0 , β ≤ d and ∥ p ∥ C 0 , β ≤ d \|p\|_{C^{0,\beta}}\leq d ∥ p ∥ C 0 , β ≤ d . Define
Q p ( u ) : = p ( Id + u ) det ( I + ∇ u ) − p . Q_p(u):=p(\operatorname{Id}+u)\det(I+\nabla u)-p. Q p ( u ) := p ( Id + u ) det ( I + ∇ u ) − p .
Again, changing variables gives Q p ( u ) ∈ Z β Q_p(u)\in Z_\beta Q p ( u ) ∈ Z β . On a fixed sufficiently small ball in W β W_\beta W β , the composition and product estimates imply ∥ Q p ( u ) ∥ C 0 , β ≤ C ∥ p ∥ C 0 , β \|Q_p(u)\|_{C^{0,\beta}}\leq C\|p\|_{C^{0,\beta}} ∥ Q p ( u ) ∥ C 0 , β ≤ C ∥ p ∥ C 0 , β . This estimate requires no derivatives of p p p .
Consider
T ( u ) : = B σ ( h − N ( u ) − Q p ( u ) ) . T(u):=\mathcal{B}_\sigma\left(h-\mathcal{N}(u)-Q_p(u)\right). T ( u ) := B σ ( h − N ( u ) − Q p ( u ) ) .
The preceding estimates give a constant C 0 > 0 C_0>0 C 0 > 0 , independent of ρ 0 , ρ 1 \rho_0,\rho_1 ρ 0 , ρ 1 , such that ∥ T ( u ) ∥ C 1 , β ≤ C 0 ( d + ∥ u ∥ C 1 , β 1 + γ − β ) \|T(u)\|_{C^{1,\beta}}\leq C_0\left(d+\|u\|_{C^{1,\beta}}^{1+\gamma-\beta}\right) ∥ T ( u ) ∥ C 1 , β ≤ C 0 ( d + ∥ u ∥ C 1 , β 1 + γ − β ) on that ball. Choose r > 0 r>0 r > 0 sufficiently small that r ≤ θ r\leq\theta r ≤ θ , the preceding estimates hold for ∥ u ∥ C 1 , β ≤ r \|u\|_{C^{1,\beta}}\leq r ∥ u ∥ C 1 , β ≤ r , and C 0 r γ − β ≤ 1 2 C_0r^{\gamma-\beta}\leq\frac{1}{2} C 0 r γ − β ≤ 2 1 . Then choose ε > 0 \varepsilon>0 ε > 0 such that C 0 ε ≤ r / 2 C_0\varepsilon\leq r/2 C 0 ε ≤ r /2 . It follows that T T T maps
K r : = { u ∈ W β : ∥ u ∥ C 1 , β ≤ r } K_r:=\{u\in W_\beta:\|u\|_{C^{1,\beta}}\leq r\} K r := { u ∈ W β : ∥ u ∥ C 1 , β ≤ r }
into itself. We apply the Schauder fixed point theorem in W η W_\eta W η . The set K r K_r K r is nonempty and convex. The compact embedding C 1 , β ( Ω ‾ ) ↪ C 1 , η ( Ω ‾ ) C^{1,\beta}(\overline{\Omega})\hookrightarrow C^{1,\eta}(\overline{\Omega}) C 1 , β ( Ω ) ↪ C 1 , η ( Ω ) makes it relatively compact in W η W_\eta W η . It is also closed in W η W_\eta W η : under convergence in C 1 , η C^{1,\eta} C 1 , η , the zero boundary values are preserved, and the C 1 , β C^{1,\beta} C 1 , β norm bound passes to the limit. Hence K r K_r K r is compact in W η W_\eta W η . To verify continuity, let u j , u ∈ K r u_j,u\in K_r u j , u ∈ K r satisfy u j → u u_j\to u u j → u in C 1 , η C^{1,\eta} C 1 , η . Then N ( u j ) → N ( u ) \mathcal{N}(u_j)\to\mathcal{N}(u) N ( u j ) → N ( u ) in C 0 , η C^{0,\eta} C 0 , η . Also, p ( Id + u j ) p(\operatorname{Id}+u_j) p ( Id + u j ) converges uniformly to p ( Id + u ) p(\operatorname{Id}+u) p ( Id + u ) , and these compositions have uniformly bounded C 0 , β C^{0,\beta} C 0 , β norms. Since η < β \eta<\beta η < β , interpolation gives p ( Id + u j ) → p ( Id + u ) p(\operatorname{Id}+u_j)\to p(\operatorname{Id}+u) p ( Id + u j ) → p ( Id + u ) in C 0 , η C^{0,\eta} C 0 , η . The determinants converge in C 0 , η C^{0,\eta} C 0 , η as well. Consequently, Q p ( u j ) → Q p ( u ) Q_p(u_j)\to Q_p(u) Q p ( u j ) → Q p ( u ) in C 0 , η C^{0,\eta} C 0 , η . The boundedness of B σ \mathcal{B}_\sigma B σ at exponent η \eta η proves that T T T is continuous on K r K_r K r in the W η W_\eta W η topology. Schauder’s theorem therefore gives u ∈ K r u\in K_r u ∈ K r satisfying T ( u ) = u T(u)=u T ( u ) = u . At this fixed point,
∥ u ∥ C 1 , β ≤ C 0 d + C 0 ∥ u ∥ C 1 , β 1 + γ − β ≤ C 0 d + 1 2 ∥ u ∥ C 1 , β , \|u\|_{C^{1,\beta}}\leq C_0d+C_0\|u\|_{C^{1,\beta}}^{1+\gamma-\beta}
\leq C_0d+\frac{1}{2}\|u\|_{C^{1,\beta}}, ∥ u ∥ C 1 , β ≤ C 0 d + C 0 ∥ u ∥ C 1 , β 1 + γ − β ≤ C 0 d + 2 1 ∥ u ∥ C 1 , β ,
and hence ∥ u ∥ C 1 , β ≤ 2 C 0 d \|u\|_{C^{1,\beta}}\leq2C_0d ∥ u ∥ C 1 , β ≤ 2 C 0 d . Set M : = Id + u M:=\operatorname{Id}+u M := Id + u . Applying div ( σ ⋅ ) \operatorname{div}(\sigma\cdot) div ( σ ⋅ ) to the fixed point equation gives div ( σ u ) = h − N ( u ) − Q p ( u ) \operatorname{div}(\sigma u)=h-\mathcal{N}(u)-Q_p(u) div ( σ u ) = h − N ( u ) − Q p ( u ) . Thus H ( u ) + Q p ( u ) = h \mathcal{H}(u)+Q_p(u)=h H ( u ) + Q p ( u ) = h , which is equivalent to
ρ 1 ( M ) det ∇ M = ρ 0 + a ∫ 0 1 J t [ M ] d t = R 1 [ M , a ] . \rho_1(M)\det\nabla M=\rho_0+a\int_0^1 J_t[M]\,\mathrm{d}t=R_1[M,a]. ρ 1 ( M ) det ∇ M = ρ 0 + a ∫ 0 1 J t [ M ] d t = R 1 [ M , a ] .
Step 5. Since ∥ ∇ u ∥ L ∞ ≤ r ≤ θ < 1 \|\nabla u\|_{L^\infty}\leq r\leq\theta<1 ∥∇ u ∥ L ∞ ≤ r ≤ θ < 1 and Ω \Omega Ω is convex, we have
∣ u ( x ) − u ( y ) ∣ ≤ θ ∣ x − y ∣ for all x , y ∈ Ω ‾ . |u(x)-u(y)|\leq\theta|x-y|\qquad\text{for all }x,y\in\overline{\Omega}. ∣ u ( x ) − u ( y ) ∣ ≤ θ ∣ x − y ∣ for all x , y ∈ Ω .
Consequently, for every t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] ,
∣ X t [ M ] ( x ) − X t [ M ] ( y ) ∣ ≥ ( 1 − t θ ) ∣ x − y ∣ ≥ ( 1 − θ ) ∣ x − y ∣ , \lvert X_t[M](x)-X_t[M](y)\rvert\ge(1-t\theta)\lvert x-y\rvert\ge(1-\theta)\lvert x-y\rvert, ∣ X t [ M ] ( x ) − X t [ M ] ( y )∣ ≥ ( 1 − tθ ) ∣ x − y ∣ ≥ ( 1 − θ ) ∣ x − y ∣ ,
∣ X t [ M ] ( x ) − X t [ M ] ( y ) ∣ ≤ ( 1 + t θ ) ∣ x − y ∣ ≤ ( 1 + θ ) ∣ x − y ∣ . \lvert X_t[M](x)-X_t[M](y)\rvert\le(1+t\theta)\lvert x-y\rvert\le(1+\theta)\lvert x-y\rvert. ∣ X t [ M ] ( x ) − X t [ M ] ( y )∣ ≤ ( 1 + tθ ) ∣ x − y ∣ ≤ ( 1 + θ ) ∣ x − y ∣ .
By the choice of θ \theta θ , these estimates imply the prescribed bi-Lipschitz bounds with constant L L L . Moreover, u = 0 u=0 u = 0 on ∂ Ω \partial\Omega ∂ Ω , so the zero-extension argument above gives
X t [ M ] ( Ω ‾ ) = Ω ‾ for every t ∈ [ 0 , 1 ] . X_t[M](\overline{\Omega})=\overline{\Omega}\qquad\text{for every }t\in[0,1]. X t [ M ] ( Ω ) = Ω for every t ∈ [ 0 , 1 ] .
Thus the required flow geometry holds. It remains to verify the nonnegativity of the intermediate states. For each fixed x ∈ Ω x\in\Omega x ∈ Ω ,
∂ t R t [ M , a ] ( x ) = a J t [ M ] ( x ) . \partial_tR_t[M,a](x)=aJ_t[M](x). ∂ t R t [ M , a ] ( x ) = a J t [ M ] ( x ) .
Since J t [ M ] ( x ) > 0 J_t[M](x)>0 J t [ M ] ( x ) > 0 , the function t ↦ R t [ M , a ] ( x ) t\mapsto R_t[M,a](x) t ↦ R t [ M , a ] ( x ) is nondecreasing if a ≥ 0 a\ge0 a ≥ 0 and nonincreasing if a ≤ 0 a\le0 a ≤ 0 . Its endpoint values satisfy
R 0 [ M , a ] = ρ 0 ≥ 0 , R 1 [ M , a ] = ρ 1 ( M ) det ∇ M ≥ 0. R_0[M,a]=\rho_0\ge0,\qquad R_1[M,a]=\rho_1(M)\det\nabla M\ge0. R 0 [ M , a ] = ρ 0 ≥ 0 , R 1 [ M , a ] = ρ 1 ( M ) det ∇ M ≥ 0.
Therefore
R t [ M , a ] ( x ) ≥ min { ρ 0 ( x ) , ρ 1 ( M ( x ) ) det ∇ M ( x ) } ≥ 0 R_t[M,a](x)\ge\min\{\rho_0(x),\rho_1(M(x))\det\nabla M(x)\}\ge0 R t [ M , a ] ( x ) ≥ min { ρ 0 ( x ) , ρ 1 ( M ( x )) det ∇ M ( x )} ≥ 0
for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . Hence ( M , a ) ∈ A L ρ 1 (M,a)\in\mathcal{A}^{\rho_1}_L ( M , a ) ∈ A L ρ 1 . Finally,
∣ ∇ M ∣ n det ∇ M ≤ n n / 2 ( 1 + θ 1 − θ ) n < K , ∫ Ω det ∇ M d x = ∣ Ω ∣ . \frac{\lvert\nabla M\rvert^n}{\det\nabla M}\le n^{n/2}\left(\frac{1+\theta}{1-\theta}\right)^n<K,\qquad\int_{\Omega}\det\nabla M\,dx=\lvert\Omega\rvert. det ∇ M ∣ ∇ M ∣ n ≤ n n /2 ( 1 − θ 1 + θ ) n < K , ∫ Ω det ∇ M d x = ∣ Ω ∣ .
Hence ( M , a ) ∈ V K ρ 1 (M,a)\in\mathcal{V}^{\rho_1}_K ( M , a ) ∈ V K ρ 1 . The asserted estimate follows from ∥ u ∥ C 1 , β ≤ 2 C 0 d \lVert u\rVert_{C^{1,\beta}}\le2C_0d ∥ u ∥ C 1 , β ≤ 2 C 0 d . □
Remark 6.1. The regularity assumption ρ ∈ C 1 , γ ( Ω ‾ ) \rho\in C^{1,\gamma}(\overline{\Omega}) ρ ∈ C 1 , γ ( Ω ) concerns only the fixed reference density. Both prescribed endpoint densities may be merely C 0 , β C^{0,\beta} C 0 , β . The proposition provides a local sufficient condition and does not assert feasibility for arbitrary Hölder densities.
Proof of Theorem 1.6. Choose ε \varepsilon ε as in Proposition 6.1. If d < ε d<\varepsilon d < ε , that proposition provides a finite-energy pair ( M ∗ , a ) ∈ V K ρ 1 ∩ A L ρ 1 (M_*,a)\in\mathcal{V}^{\rho_1}_K\cap\mathcal{A}^{\rho_1}_L ( M ∗ , a ) ∈ V K ρ 1 ∩ A L ρ 1 . Existence and static-dynamic equivalence follow from Theorems 1.2 and 1.4 . The mass constraint gives
∥ f ∥ L 2 ( 0 , 1 ) 2 = a 2 + ∥ f − a ∥ L 2 ( 0 , 1 ) 2 \lVert f\rVert_{L^2(0,1)}^2=a^2+\lVert f-a\rVert_{L^2(0,1)}^2 ∥ f ∥ L 2 ( 0 , 1 ) 2 = a 2 + ∥ f − a ∥ L 2 ( 0 , 1 ) 2
for every ( M , f ) ∈ A L ρ 1 (M,f)\in\mathcal{A}^{\rho_1}_L ( M , f ) ∈ A L ρ 1 . The constant-source competitor therefore implies
b L = a 2 , S L ρ 1 = { ( M , a ) ∈ A L ρ 1 } , g L = m a . b_L=a^2,\qquad\mathcal{S}^{\rho_1}_L=\{(M,a)\in\mathcal{A}^{\rho_1}_L\},\qquad g_L=ma. b L = a 2 , S L ρ 1 = {( M , a ) ∈ A L ρ 1 } , g L = ma .
The convergence and minimum-value expansion now follow from Theorem 1.5 . □
Examples of nonemptiness
We give explicit examples verifying the feasibility assumptions used in the existence and convergence results. The first two constructions establish nonemptiness of V K ρ 1 \mathcal{V}^{\rho_1}_K V K ρ 1 and V K , Λ ρ 1 \mathcal{V}^{\rho_1}_{K,\Lambda} V K , Λ ρ 1 . For the reconstruction class, we exhibit endpoint densities admitting both constant and nonconstant sources, followed by an example for which admissible sources exist but no constant source is feasible. These examples distinguish the constant-source limit from the general minimum-source-cost limit.
Example 6.1 (Nonemptiness of V K ρ 1 \mathcal{V}_{K}^{\rho_1} V K ρ 1 ). Let Ω = ( 0 , 1 ) n \Omega=(0,1)^n Ω = ( 0 , 1 ) n and suppose that
ρ 0 ( x ) = r ( x 1 ) , ρ 1 ( y ) = q ( y 1 ) , \rho_0(x)=r(x_1),\qquad\rho_1(y)=q(y_1), ρ 0 ( x ) = r ( x 1 ) , ρ 1 ( y ) = q ( y 1 ) , where r ∈ L ∞ ( 0 , 1 ) r\in L^\infty(0,1) r ∈ L ∞ ( 0 , 1 ) is nonnegative and q ∈ C ( [ 0 , 1 ] ) q\in C([0,1]) q ∈ C ([ 0 , 1 ]) is strictly positive. Write
r ∗ : = e s s i n f ( 0 , 1 ) r , r ∗ : = ∥ r ∥ L ∞ ( 0 , 1 ) , r_*:=\mathop{\mathrm{ess\,inf}}_{(0,1)}r,\qquad r^*:=\lVert r\rVert_{L^\infty(0,1)}, r ∗ := ess inf ( 0 , 1 ) r , r ∗ := ∥ r ∥ L ∞ ( 0 , 1 ) , and
q ∗ : = min [ 0 , 1 ] q , q ∗ : = max [ 0 , 1 ] q . q_*:=\min_{[0,1]}q,\qquad q^*:=\max_{[0,1]}q. q ∗ := [ 0 , 1 ] min q , q ∗ := [ 0 , 1 ] max q . Set
a : = ∫ 0 1 q ( s ) d s − ∫ 0 1 r ( s ) d s a:=\int_0^1 q(s)\,\mathrm{d}s-\int_0^1 r(s)\,\mathrm{d}s a := ∫ 0 1 q ( s ) d s − ∫ 0 1 r ( s ) d s and assume that
r ∗ + a 2 > 0 , q ∗ − a 2 > 0. (23) r_*+\frac{a}{2}>0,\qquad q_*-\frac{a}{2}>0. \tag*{(23)} r ∗ + 2 a > 0 , q ∗ − 2 a > 0. (23) Define
Q a ( y ) : = ∫ 0 y ( q ( ξ ) − a 2 ) d ξ Q_a(y):=\int_0^y\left(q(\xi)-\frac{a}{2}\right)\,\mathrm{d}\xi Q a ( y ) := ∫ 0 y ( q ( ξ ) − 2 a ) d ξ and
F ( s ) : = Q a − 1 ( ∫ 0 s ( r ( ξ ) + a 2 ) d ξ ) . F(s):=Q_a^{-1}\left(\int_0^s\left(r(\xi)+\frac{a}{2}\right)\,\mathrm{d}\xi\right). F ( s ) := Q a − 1 ( ∫ 0 s ( r ( ξ ) + 2 a ) d ξ ) . The assumptions (6.1) ensure that both cumulative functions are strictly increasing. Furthermore,
∫ 0 1 ( r + a 2 ) d s = ∫ 0 1 ( q − a 2 ) d s . \int_0^1\left(r+\frac{a}{2}\right)\,\mathrm{d}s=\int_0^1\left(q-\frac{a}{2}\right)\,\mathrm{d}s. ∫ 0 1 ( r + 2 a ) d s = ∫ 0 1 ( q − 2 a ) d s . Thus F F F is a Lipschitz strictly increasing map of [ 0 , 1 ] [0,1] [ 0 , 1 ] onto itself. For almost every s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , its derivative is
g ( s ) : = F ′ ( s ) = r ( s ) + a / 2 q ( F ( s ) ) − a / 2 . g(s):=F'(s)=\frac{r(s)+a/2}{q(F(s))-a/2}. g ( s ) := F ′ ( s ) = q ( F ( s )) − a /2 r ( s ) + a /2 . In particular,
0 < ℓ : = r ∗ + a / 2 q ∗ − a / 2 ≤ g ( s ) ≤ l : = r ∗ + a / 2 q ∗ − a / 2 < ∞ . 0<\ell:=\frac{r_*+a/2}{q^*-a/2}\leq g(s)\leq l:=\frac{r^*+a/2}{q_*-a/2}<\infty. 0 < ℓ := q ∗ − a /2 r ∗ + a /2 ≤ g ( s ) ≤ l := q ∗ − a /2 r ∗ + a /2 < ∞. Set
M ( x ) : = ( F ( x 1 ) , x 2 , … , x n ) , f ( t ) : = a . M(x):=(F(x_1),x_2,\ldots,x_n),\qquad f(t):=a. M ( x ) := ( F ( x 1 ) , x 2 , … , x n ) , f ( t ) := a . Then M M M maps Ω \Omega Ω onto itself and belongs to W 1 , ∞ ( Ω ; R n ) W^{1,\infty}(\Omega;\mathbb{R}^n) W 1 , ∞ ( Ω ; R n ) . Moreover,
∇ M = diag ( g ( x 1 ) , 1 , … , 1 ) , det ∇ M = g ( x 1 ) , \nabla M=\operatorname{diag}(g(x_1),1,\ldots,1),\qquad\det\nabla M=g(x_1), ∇ M = diag ( g ( x 1 ) , 1 , … , 1 ) , det ∇ M = g ( x 1 ) , and
det ( t ∇ M + ( 1 − t ) I ) = 1 − t + t g ( x 1 ) ≥ min { 1 , ℓ } > 0. \det\bigl(t\nabla M+(1-t)I\bigr)=1-t+tg(x_1)\geq\min\{1,\ell\}>0. det ( t ∇ M + ( 1 − t ) I ) = 1 − t + t g ( x 1 ) ≥ min { 1 , ℓ } > 0. The definition of g g g gives
q ( F ( x 1 ) ) g ( x 1 ) = r ( x 1 ) + a 2 ( 1 + g ( x 1 ) ) . q(F(x_1))g(x_1)=r(x_1)+\frac{a}{2}(1+g(x_1)). q ( F ( x 1 )) g ( x 1 ) = r ( x 1 ) + 2 a ( 1 + g ( x 1 )) . Consequently,
R 1 [ M , f ] ( x ) = r ( x 1 ) + ∫ 0 1 a ( 1 − s + s g ( x 1 ) ) d s R_1[M,f](x)=r(x_1)+\int_0^1 a(1-s+sg(x_1))\,\mathrm{d}s R 1 [ M , f ] ( x ) = r ( x 1 ) + ∫ 0 1 a ( 1 − s + s g ( x 1 )) d s = q ( F ( x 1 ) ) g ( x 1 ) =q(F(x_1))g(x_1) = q ( F ( x 1 )) g ( x 1 ) = ρ 1 ( M ( x ) ) det ∇ M ( x ) . =\rho_1(M(x))\det\nabla M(x). = ρ 1 ( M ( x )) det ∇ M ( x ) .
For intermediate times,
R t [ M , f ] ( x ) = r ( x 1 ) + a ( t + t 2 2 ( g ( x 1 ) − 1 ) ) . R_t[M,f](x)=r(x_1)+a\left(t+\frac{t^2}{2}(g(x_1)-1)\right). R t [ M , f ] ( x ) = r ( x 1 ) + a ( t + 2 t 2 ( g ( x 1 ) − 1 ) ) .
Since det ( t ∇ M + ( 1 − t ) I ) > 0 \det(t\nabla M+(1-t)I)>0 det ( t ∇ M + ( 1 − t ) I ) > 0 and a a a is constant, this function is monotone in t t t . Its endpoint values are r ( x 1 ) r(x_1) r ( x 1 ) and q ( F ( x 1 ) ) g ( x 1 ) q(F(x_1))g(x_1) q ( F ( x 1 )) g ( x 1 ) , both nonnegative. Hence
R t [ M , f ] ( x ) ≥ 0 for every t ∈ [ 0 , 1 ] . R_t[M,f](x)\geq0 \quad\text{for every }t\in[0,1]. R t [ M , f ] ( x ) ≥ 0 for every t ∈ [ 0 , 1 ] .
The Jacobian mass bound follows from
∫ Ω det ∇ M d x = ∫ 0 1 g ( s ) d s = F ( 1 ) − F ( 0 ) = 1 = ∣ Ω ∣ . \int_{\Omega}\det\nabla M\,dx=\int_0^1g(s)\,ds=F(1)-F(0)=1=|\Omega|. ∫ Ω det ∇ M d x = ∫ 0 1 g ( s ) d s = F ( 1 ) − F ( 0 ) = 1 = ∣Ω∣.
Also,
∣ ∇ M ∣ n = ( g ( x 1 ) 2 + n − 1 ) n / 2 ≤ K ∗ det ∇ M , |\nabla M|^n=(g(x_1)^2+n-1)^{n/2}\leq K_*\det\nabla M, ∣∇ M ∣ n = ( g ( x 1 ) 2 + n − 1 ) n /2 ≤ K ∗ det ∇ M ,
where
K ∗ : = ( ℓ 2 + n − 1 ) n / 2 ℓ . K_*:=\frac{(\ell^2+n-1)^{n/2}}{\ell}. K ∗ := ℓ ( ℓ 2 + n − 1 ) n /2 .
Therefore,
( M , f ) ∈ V K ρ 1 for every K ≥ K ∗ . (M,f)\in\mathcal{V}^{\rho_1}_K \quad\text{for every }K\geq K_*. ( M , f ) ∈ V K ρ 1 for every K ≥ K ∗ .
Its energy is finite because the map has bounded range.
Example 6.2 (Nonemptiness of V K , Λ ρ 1 \mathcal{V}^{\rho_1}_{K,\Lambda} V K , Λ ρ 1 ). Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be a bounded connected open set with Lipschitz boundary. Fix K ≥ n n / 2 K\geq n^{n/2} K ≥ n n /2 and Λ > 0 \Lambda>0 Λ > 0 , and choose s > 0 s>0 s > 0 such that
s n ∣ Ω ∣ ≤ Λ . s^n|\Omega|\leq\Lambda. s n ∣Ω∣ ≤ Λ. For any b ∈ R n b\in\mathbb{R}^n b ∈ R n , set
M ( x ) : = b + s x , j s ( t ) : = ( 1 − t + t s ) n , c s : = ∫ 0 1 j s ( t ) d t > 0. M(x):=b+sx,\qquad j_s(t):=(1-t+ts)^n,\qquad c_s:=\int_0^1j_s(t)\,dt>0. M ( x ) := b + s x , j s ( t ) := ( 1 − t + t s ) n , c s := ∫ 0 1 j s ( t ) d t > 0. Choose c > 0 c>0 c > 0 and q ≥ 0 q\geq0 q ≥ 0 , and take ψ ∈ C c ∞ ( R n ) \psi\in C^\infty_c(\mathbb{R}^n) ψ ∈ C c ∞ ( R n ) satisfying
0 ≤ ψ ≤ 1 , ψ = 1 on b + s Ω ‾ . 0\leq\psi\leq1,\qquad\psi=1\quad\text{on }b+s\overline{\Omega}. 0 ≤ ψ ≤ 1 , ψ = 1 on b + s Ω . Define
ρ 0 ( x ) : = c , ρ 1 ( y ) : = c + q c s s n ψ ( y ) , f ( t ) : = q . \rho_0(x):=c,\qquad\rho_1(y):=\frac{c+qc_s}{s^n}\psi(y),\qquad f(t):=q. ρ 0 ( x ) := c , ρ 1 ( y ) := s n c + q c s ψ ( y ) , f ( t ) := q . Then ρ 1 ∈ C c ∞ ( R n ) \rho_1\in C^\infty_c(\mathbb{R}^n) ρ 1 ∈ C c ∞ ( R n ) is nonnegative, and
∫ Ω ρ 0 d x = c ∣ Ω ∣ > 0. \int_{\Omega}\rho_0\,dx=c|\Omega|>0. ∫ Ω ρ 0 d x = c ∣Ω∣ > 0. Since ∇ M = s I \nabla M=sI ∇ M = s I , we have
det ∇ M = s n , J t [ M ] = j s ( t ) > 0 , \det\nabla M=s^n,\qquad J_t[M]=j_s(t)>0, det ∇ M = s n , J t [ M ] = j s ( t ) > 0 , and
R t [ M , q ] = c + q ∫ 0 t j s ( r ) d r ≥ c > 0. R_t[M,q]=c+q\int_0^t j_s(r)\,dr\geq c>0. R t [ M , q ] = c + q ∫ 0 t j s ( r ) d r ≥ c > 0. As ψ ( M ( x ) ) = 1 \psi(M(x))=1 ψ ( M ( x )) = 1 , the endpoint constraint holds pointwise:
ρ 1 ( M ( x ) ) det ∇ M ( x ) = c + q c s = R 1 [ M , q ] ( x ) . \rho_1(M(x))\det\nabla M(x)=c+qc_s=R_1[M,q](x). ρ 1 ( M ( x )) det ∇ M ( x ) = c + q c s = R 1 [ M , q ] ( x ) .
Moreover,
∣ ∇ M ∣ n = n n / 2 s n ≤ K det ∇ M , |\nabla M|^n = n^{n/2}s^n \le K\det\nabla M, ∣∇ M ∣ n = n n /2 s n ≤ K det ∇ M ,
and
∫ Ω det ∇ M d x = s n ∣ Ω ∣ ≤ Λ . \int_{\Omega}\det\nabla M\,dx=s^n|\Omega|\le\Lambda. ∫ Ω det ∇ M d x = s n ∣Ω∣ ≤ Λ.
Therefore,
( M , q ) ∈ V K , Λ ρ 1 . (M,q)\in\mathcal{V}_{K,\Lambda}^{\rho_1}. ( M , q ) ∈ V K , Λ ρ 1 .
This pair has finite energy, since
E α ( M , q ) = ( c + q ∫ 0 1 ( 1 − t ) j s ( t ) d t ) ∫ Ω ∣ b + ( s − 1 ) x ∣ 2 d x + q 2 α < ∞ . E_\alpha(M,q)=\left(c+q\int_0^1(1-t)j_s(t)\,dt\right)\int_\Omega|b+(s-1)x|^2\,dx+\frac{q^2}{\alpha}<\infty. E α ( M , q ) = ( c + q ∫ 0 1 ( 1 − t ) j s ( t ) d t ) ∫ Ω ∣ b + ( s − 1 ) x ∣ 2 d x + α q 2 < ∞.
Taking q > 0 q>0 q > 0 gives a positive constant-source pair, whereas q = 0 q=0 q = 0 gives a finite-cost zero-source pair. The translation b b b is unrestricted, so M ( Ω ) M(\Omega) M ( Ω ) need not lie in Ω \Omega Ω .
Example 6.3 (Constant and nonconstant sources for the same endpoint densities). Let Ω = ( 0 , 1 ) n \Omega=(0,1)^n Ω = ( 0 , 1 ) n , fix L > 1 L>1 L > 1 and 0 < a < r 0<a<r 0 < a < r , and choose
0 < ε < min { 1 − L − 1 , r − a r − a / 2 } . 0<\varepsilon<\min\left\{1-L^{-1},\frac{r-a}{r-a/2}\right\}. 0 < ε < min { 1 − L − 1 , r − a /2 r − a } .
Set
ρ 1 ≡ r , ρ 0 ( x ) = r − a + ε ( r − a 2 ) cos ( 2 π x 1 ) , \rho_1\equiv r,\qquad\rho_0(x)=r-a+\varepsilon\left(r-\frac{a}{2}\right)\cos(2\pi x_1), ρ 1 ≡ r , ρ 0 ( x ) = r − a + ε ( r − 2 a ) cos ( 2 π x 1 ) ,
and define
M ∗ ( x ) = ( x 1 + ε 2 π sin ( 2 π x 1 ) , x 2 , … , x n ) . M_*(x)=\left(x_1+\frac{\varepsilon}{2\pi}\sin(2\pi x_1),x_2,\ldots,x_n\right). M ∗ ( x ) = ( x 1 + 2 π ε sin ( 2 π x 1 ) , x 2 , … , x n ) .
Then ρ 0 \rho_0 ρ 0 is strictly positive and
∫ Ω ( ρ 1 − ρ 0 ) d x = a . \int_\Omega(\rho_1-\rho_0)\,dx=a. ∫ Ω ( ρ 1 − ρ 0 ) d x = a .
Each interpolation map changes only the first coordinate, whose derivative is 1 + t ε cos ( 2 π x 1 ) 1+t\varepsilon\cos(2\pi x_1) 1 + tε cos ( 2 π x 1 ) . This coordinate map fixes 0 0 0 and 1 1 1 and is strictly increasing. Hence X t [ M ∗ ] X_t[M_*] X t [ M ∗ ] maps Ω \Omega Ω onto itself. Moreover,
L − 1 < 1 − ε ≤ 1 + t ε cos ( 2 π x 1 ) ≤ 1 + ε < L , L^{-1}<1-\varepsilon\le1+t\varepsilon\cos(2\pi x_1)\le1+\varepsilon<L, L − 1 < 1 − ε ≤ 1 + tε cos ( 2 π x 1 ) ≤ 1 + ε < L ,
so the prescribed bi-Lipschitz bounds hold. The interpolation Jacobian is
J t [ M ∗ ] ( x ) = 1 + t ε cos ( 2 π x 1 ) . J_t[M_*](x)=1+t\varepsilon\cos(2\pi x_1). J t [ M ∗ ] ( x ) = 1 + tε cos ( 2 π x 1 ) .
For the constant source a a a , we have
R t [ M ∗ , a ] ( x ) = ρ 0 ( x ) + a ( t + ε t 2 2 cos ( 2 π x 1 ) ) . R_t[M_*,a](x)=\rho_0(x)+a\left(t+\frac{\varepsilon t^2}{2}\cos(2\pi x_1)\right). R t [ M ∗ , a ] ( x ) = ρ 0 ( x ) + a ( t + 2 ε t 2 cos ( 2 π x 1 ) ) .
Since a > 0 a>0 a > 0 and J t [ M ∗ ] > 0 J_t[M_*]>0 J t [ M ∗ ] > 0 ,
R t [ M ∗ , a ] ≥ ρ 0 > 0. R_t[M_*,a]\ge\rho_0>0. R t [ M ∗ , a ] ≥ ρ 0 > 0.
At t = 1 t=1 t = 1 ,
R 1 [ M ∗ , a ] = r ( 1 + ε cos ( 2 π x 1 ) ) = ρ 1 ( M ∗ ) det ∇ M ∗ . R_1[M_*,a]=r(1+\varepsilon\cos(2\pi x_1))=\rho_1(M_*)\det\nabla M_*. R 1 [ M ∗ , a ] = r ( 1 + ε cos ( 2 π x 1 )) = ρ 1 ( M ∗ ) det ∇ M ∗ .
Thus ( M ∗ , a ) ∈ A L ρ 1 (M_*,a)\in\mathcal{A}_L^{\rho_1} ( M ∗ , a ) ∈ A L ρ 1 .
Now consider the nonconstant source f ( t ) : = 6 a t ( 1 − t ) f(t):=6at(1-t) f ( t ) := 6 a t ( 1 − t ) . It is nonnegative and satisfies
∫ 0 1 f ( t ) d t = a , ∫ 0 1 t f ( t ) d t = a 2 . \int_0^1 f(t)\,dt=a,\qquad\int_0^1tf(t)\,dt=\frac{a}{2}. ∫ 0 1 f ( t ) d t = a , ∫ 0 1 t f ( t ) d t = 2 a .
Consequently,
R t [ M ∗ , f ] = ρ 0 + ∫ 0 t f ( s ) ( 1 + s ε cos ( 2 π x 1 ) ) d s ≥ ρ 0 > 0 , R_t[M_\ast,f]=\rho_0+\int_0^t f(s)(1+s\varepsilon\cos(2\pi x_1))\,ds\geq\rho_0>0, R t [ M ∗ , f ] = ρ 0 + ∫ 0 t f ( s ) ( 1 + s ε cos ( 2 π x 1 )) d s ≥ ρ 0 > 0 ,
and the two moment identities give
R 1 [ M ∗ , f ] = ρ 0 + a + a ε 2 cos ( 2 π x 1 ) = r ( 1 + ε cos ( 2 π x 1 ) ) = ρ 1 ( M ∗ ) det ∇ M ∗ . \begin{aligned}
R_1[M_\ast,f]&=\rho_0+a+\frac{a\varepsilon}{2}\cos(2\pi x_1)\\
&=r(1+\varepsilon\cos(2\pi x_1))\\
&=\rho_1(M_\ast)\det\nabla M_\ast.
\end{aligned} R 1 [ M ∗ , f ] = ρ 0 + a + 2 a ε cos ( 2 π x 1 ) = r ( 1 + ε cos ( 2 π x 1 )) = ρ 1 ( M ∗ ) det ∇ M ∗ .
Therefore ( M ∗ , f ) ∈ A L ρ 1 (M_\ast,f)\in\mathcal{A}^{\rho_1}_L ( M ∗ , f ) ∈ A L ρ 1 as well. Both pairs have finite energy. The constant-source problem also has m a > 0 m_a>0 m a > 0 . Indeed, let c 0 : = min Ω ‾ ρ 0 > 0 c_0:=\min_{\overline{\Omega}}\rho_0>0 c 0 := min Ω ρ 0 > 0 . For every ( M , a ) ∈ A L ρ 1 (M,a)\in\mathcal{A}^{\rho_1}_L ( M , a ) ∈ A L ρ 1 , the positivity of a a a and J t [ M ] J_t[M] J t [ M ] gives
G ( M , a ) ≥ c 0 ∫ Ω ∣ M ( x ) − x ∣ 2 d x . G(M,a)\geq c_0\int_\Omega|M(x)-x|^2\,dx. G ( M , a ) ≥ c 0 ∫ Ω ∣ M ( x ) − x ∣ 2 d x .
If m a = 0 m_a=0 m a = 0 , an attained minimizer would satisfy M = Id M=\operatorname{Id} M = Id a.e. Its endpoint constraint would then imply ρ 1 = ρ 0 + a \rho_1=\rho_0+a ρ 1 = ρ 0 + a , contrary to the definition of ρ 0 \rho_0 ρ 0 . Thus a constant source is feasible, but nonzero spatial transport is still necessary.
Example 6.4
Example 6.4 (A nonempty reconstruction class without a constant source). Let Ω = ( 0 , 1 ) n \Omega=(0,1)^n Ω = ( 0 , 1 ) n , L = 2 L=2 L = 2 , and set
ρ 1 ≡ 1 , ρ 0 ( x ) = 2 + 2 cos ( 2 π x 1 ) . \rho_1\equiv1,\qquad\rho_0(x)=2+2\cos(2\pi x_1). ρ 1 ≡ 1 , ρ 0 ( x ) = 2 + 2 cos ( 2 π x 1 ) .
Then ρ 0 ≥ 0 \rho_0\geq0 ρ 0 ≥ 0 and the prescribed mean source is
a = ∫ Ω ρ 1 d x − ∫ Ω ρ 0 d x ∣ Ω ∣ = − 1. a=\frac{\int_\Omega\rho_1\,dx-\int_\Omega\rho_0\,dx}{|\Omega|}=-1. a = ∣Ω∣ ∫ Ω ρ 1 d x − ∫ Ω ρ 0 d x = − 1.
Define
M ∗ ( x ) = ( x 1 + 1 4 π sin ( 2 π x 1 ) , x 2 , … , x n ) , f ∗ ( t ) = 14 − 30 t . M_\ast(x)=\left(x_1+\frac{1}{4\pi}\sin(2\pi x_1),x_2,\ldots,x_n\right),\qquad f_\ast(t)=14-30t. M ∗ ( x ) = ( x 1 + 4 π 1 sin ( 2 π x 1 ) , x 2 , … , x n ) , f ∗ ( t ) = 14 − 30 t .
The first coordinate of X t [ M ∗ ] X_t[M_\ast] X t [ M ∗ ] fixes 0 and 1 and has derivative
1 + t 2 cos ( 2 π x 1 ) ∈ [ 1 2 , 3 2 ] . 1+\frac{t}{2}\cos(2\pi x_1)\in\left[\frac{1}{2},\frac{3}{2}\right]. 1 + 2 t cos ( 2 π x 1 ) ∈ [ 2 1 , 2 3 ] .
The remaining coordinates are unchanged. Thus X t [ M ∗ ] X_t[M_\ast] X t [ M ∗ ] maps Ω ‾ \overline{\Omega} Ω onto itself and satisfies the prescribed bi-Lipschitz bounds with L = 2 L=2 L = 2 . Moreover,
J t [ M ∗ ] ( x ) = 1 + t 2 cos ( 2 π x 1 ) . J_t[M_\ast](x)=1+\frac{t}{2}\cos(2\pi x_1). J t [ M ∗ ] ( x ) = 1 + 2 t cos ( 2 π x 1 ) .
The source satisfies
∫ 0 1 f ∗ ( t ) d t = − 1 , ∫ 0 1 t f ∗ ( t ) d t = − 3. \int_0^1 f_\ast(t)\,dt=-1,\qquad\int_0^1tf_\ast(t)\,dt=-3. ∫ 0 1 f ∗ ( t ) d t = − 1 , ∫ 0 1 t f ∗ ( t ) d t = − 3.
Hence
R 1 [ M ∗ , f ∗ ] = ρ 0 − 1 − 3 2 cos ( 2 π x 1 ) = 1 + 1 2 cos ( 2 π x 1 ) = ρ 1 ( M ∗ ) det ∇ M ∗ . \begin{aligned}
R_1[M_\ast,f_\ast]&=\rho_0-1-\frac{3}{2}\cos(2\pi x_1)\\
&=1+\frac{1}{2}\cos(2\pi x_1)\\
&=\rho_1(M_\ast)\det\nabla M_\ast.
\end{aligned} R 1 [ M ∗ , f ∗ ] = ρ 0 − 1 − 2 3 cos ( 2 π x 1 ) = 1 + 2 1 cos ( 2 π x 1 ) = ρ 1 ( M ∗ ) det ∇ M ∗ .
Since J t [ M ∗ ] > 0 J_t[M_*] > 0 J t [ M ∗ ] > 0 and f ∗ f_* f ∗ changes sign once, from positive to negative, R t [ M ∗ , f ∗ ] ( x ) R_t[M_*, f_*](x) R t [ M ∗ , f ∗ ] ( x ) first increases and then decreases as a function of t t t . Therefore,
R t [ M ∗ , f ∗ ] ( x ) ≥ min { ρ 0 ( x ) , R 1 [ M ∗ , f ∗ ] ( x ) } ≥ 0. R_t[M_*, f_*](x) \ge\min\{\rho_0(x), R_1[M_*, f_*](x)\} \ge0. R t [ M ∗ , f ∗ ] ( x ) ≥ min { ρ 0 ( x ) , R 1 [ M ∗ , f ∗ ] ( x )} ≥ 0.
It follows that
( M ∗ , f ∗ ) ∈ A 2 ρ 1 , (M_*, f_*) \in\mathcal{A}_2^{\rho_1}, ( M ∗ , f ∗ ) ∈ A 2 ρ 1 ,
and this pair has finite energy. On the other hand, if ( M , − 1 ) ∈ A 2 ρ 1 (M, -1) \in\mathcal{A}_2^{\rho_1} ( M , − 1 ) ∈ A 2 ρ 1 , the endpoint constraint and the Jacobian lower bound would give
ρ 0 = J 1 [ M ] + ∫ 0 1 J t [ M ] d t ≥ 2 1 − n a.e. in Ω . \rho_0 = J_1[M] + \int_0^1 J_t[M]\,dt \ge2^{1-n}\quad\text{a.e. in }\Omega. ρ 0 = J 1 [ M ] + ∫ 0 1 J t [ M ] d t ≥ 2 1 − n a.e. in Ω.
This contradicts
e s s i n f Ω ρ 0 = 0. \mathop{\mathrm{ess\,inf}}_{\Omega}\rho_0 = 0. ess inf Ω ρ 0 = 0.
Thus A 2 ρ 1 \mathcal{A}_2^{\rho_1} A 2 ρ 1 is nonempty but contains no constant-source admissible pair. Finally, Proposition 5.2, with c 0 = 0 c_0=0 c 0 = 0 , c 1 = 1 c_1=1 c 1 = 1 , and a = − 1 a=-1 a = − 1 , yields b 2 ≥ 1 + 2 2 − 4 n b_2 \ge1+2^{2-4n} b 2 ≥ 1 + 2 2 − 4 n . The constructed competitor gives
b 2 ≤ ∥ f ∗ ∥ L 2 ( 0 , 1 ) 2 = ∫ 0 1 ( 14 − 30 t ) 2 d t = 76. b_2 \le\|f_*\|_{L^2(0,1)}^2 = \int_0^1 (14-30t)^2\,dt = 76. b 2 ≤ ∥ f ∗ ∥ L 2 ( 0 , 1 ) 2 = ∫ 0 1 ( 14 − 30 t ) 2 d t = 76.
Consequently,
1 + 2 2 − 4 n ≤ b 2 ≤ 76. 1+2^{2-4n} \le b_2 \le76. 1 + 2 2 − 4 n ≤ b 2 ≤ 76.
This example shows that admissibility may require temporal variation of the source even when its mean is fixed by the endpoint masses.
Acknowledgements. The first author is grateful to Prof. Xiao-Ping Yang for insightful discussions. Both authors also wish to express their sincere gratitude to him for his long-term encouragement and guidance in this research direction. This work has been supported by the National Natural Science Foundation of China (Grant No. 12271093), the Jiangsu Provincial Scientific Research Center of Applied Mathematics (Grant No. BK20233002), and Shanghai Institute for Mathematics and Interdisciplinary Sciences (SIMIS) under grant number SIMIS-ID-2025-AD.
References
School of Mathematics and Shing-Tung Yau Center, Southeast University, Nanjing 211189, P. R. China
Email address: zhao-qing@seu.edu.cn
School of Mathematics and Shing-Tung Yau Center, Southeast University, Nanjing 211189, P. R. China; Shanghai Institute for Mathematics and Interdisciplinary Sciences, Shanghai 200433, P. R. China
Email address: jiangfeida@seu.edu.cn