Introduction
Wasserstein barycenters provide a notion of averaging probability measures that accounts for the geometry of their underlying space. Agueh and Carlier [1 ] developed their variational theory and established connections with multi-marginal optimal transport. Subsequent work has extended this framework to more general metric settings and studied its applications in imaging and statistics; see, for example, [17 , 22 , 6 , 15 , 26 ] . Regularization provides a way to impose additional structure on the barycenter. Bigot, Cazelles and Papadakis [5 ] studied penalized Wasserstein barycenters, while Carlier, Eichinger and Kroshnin [8 ] investigated entropy penalties on the barycenter density. Such penalties act on the unknown measure itself. A different approach regularizes the transport problems, for example by adding entropy to the transport plans; see [21 ] . Chizat [10 ] studies a doubly regularized formulation combining entropic transport costs with an entropy penalty on the barycenter, and establishes regularity and stability results. Our formulation retains the unregularized transport costs and penalizes the total variation of the barycenter density.
Let Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d be a nonempty bounded connected Lipschitz domain, and let μ 1 , … , μ n ∈ P ( Ω ‾ ) \mu_{1}, \ldots, \mu_{n} \in\mathcal{P}(\overline{\Omega}) μ 1 , … , μ n ∈ P ( Ω ) . We seek a barycenter whose density has bounded variation and penalize its spatial variation. To formulate this problem on P ( Ω ‾ ) \mathcal{P}(\overline{\Omega}) P ( Ω ) , we extend the total variation of densities by + ∞ +\infty + ∞ outside this class. For ν ∈ P ( Ω ‾ ) \nu\in\mathcal{P}(\overline{\Omega}) ν ∈ P ( Ω ) , define
T ( ν ) : = { ∣ D ρ ∣ ( Ω ) , if ν = ρ d x for some ρ ∈ B V ( Ω ) , + ∞ , otherwise. \mathcal{T}(\nu) :=
\begin{cases}
|D\rho|(\Omega), & \text{if } \nu=\rho\,dx \text{ for some } \rho\in BV(\Omega),\\
+\infty, & \text{otherwise.}
\end{cases} T ( ν ) := { ∣ D ρ ∣ ( Ω ) , + ∞ , if ν = ρ d x for some ρ ∈ B V ( Ω ) , otherwise.
We consider
inf ν ∈ P ( Ω ‾ ) { ∑ i = 1 n W 2 2 ( μ i , ν ) + T ( ν ) } . \inf_{\nu\in\mathcal{P}(\overline{\Omega})}\left\{\sum_{i=1}^{n} W_{2}^{2}(\mu_{i},\nu)+\mathcal{T}(\nu)\right\}. ν ∈ P ( Ω ) inf { i = 1 ∑ n W 2 2 ( μ i , ν ) + T ( ν ) } .
2020 Mathematics Subject Classification. 49Q22, 49Q20, 26B30.
Key words and phrases. Wasserstein barycenter; total variation; convex duality; functions of bounded variation; relative perimeter. The probability measure with constant density 1 / ∣ Ω ∣ 1/|\Omega| 1/∣Ω∣ on Ω \Omega Ω has finite energy, so the infimum is finite. By the definition of T \mathcal{T} T , the problem is therefore equivalent to minimization over probability densities in B V ( Ω ) BV(\Omega) B V ( Ω ) :
inf ρ ∈ A { ∑ i = 1 n W 2 2 ( μ i , ρ d x ) + ∣ D ρ ∣ ( Ω ) } , A = { ρ ∈ B V ( Ω ) : ρ ≥ 0 a.e. , ∫ Ω ρ d x = 1 } . (1) \inf_{\rho\in\mathcal{A}}\left\{\sum_{i=1}^{n} W_2^2(\mu_i,\rho\,dx)+|D\rho|(\Omega)\right\}, \qquad\mathcal{A}=\left\{\rho\in BV(\Omega):\rho\geq0\ \text{a.e.},\ \int_{\Omega}\rho\,dx=1\right\}. \tag*{(1)} ρ ∈ A inf { i = 1 ∑ n W 2 2 ( μ i , ρ d x ) + ∣ D ρ ∣ ( Ω ) } , A = { ρ ∈ B V ( Ω ) : ρ ≥ 0 a.e. , ∫ Ω ρ d x = 1 } . (1)
The prescribed measures may be singular, whereas every minimizer has a density in B V ( Ω ) BV(\Omega) B V ( Ω ) . The total variation penalty controls spatial variation while allowing jump discontinuities. Its use in image restoration goes back to Rudin, Osher and Fatemi [23 ] . In the present setting, it controls the spatial variation of the barycenter without requiring its density to be smooth or strictly positive.
The closest related variational problem is the TV-JKO scheme studied by Carlier and Poon [9 ] . For a given probability measure μ \mu μ , one step of this scheme minimizes
1 2 τ W 2 2 ( μ , ρ d x ) + ∣ D ρ ∣ ( Ω ) \frac{1}{2\tau}W_2^2(\mu,\rho\,dx)+|D\rho|(\Omega) 2 τ 1 W 2 2 ( μ , ρ d x ) + ∣ D ρ ∣ ( Ω )
over ρ ∈ A \rho\in\mathcal{A} ρ ∈ A . Thus (1.1) with n = 1 n=1 n = 1 corresponds to τ = 1 / 2 \tau=1/2 τ = 1/2 . We adapt the entropic approximation of Carlier and Poon [9 ] to derive vector-field optimality conditions for TV-regularized Wasserstein barycenters, allowing singular prescribed measures. We also extend their projection argument to obtain a maximum principle on convex domains. We further establish an attained dual formulation through relative-perimeter inequalities and relate it to the vector-field formulation. This duality characterizes almost every superlevel set of positive measure of a minimizing density as a minimizer of a perimeter ratio. In one dimension, an exact variance identity reduces the barycenter problem to a single TV-JKO step.
Main results
We first prove existence for arbitrary μ 1 , … , μ n ∈ P ( Ω ‾ ) \mu_1,\ldots,\mu_n\in\mathcal{P}(\overline{\Omega}) μ 1 , … , μ n ∈ P ( Ω ) . If at least one of these measures is absolutely continuous, the classical strict-convexity argument for the transport cost gives uniqueness. We then establish an attained dual formulation involving transport potentials, a scalar multiplier, and relative-perimeter constraints.
The duality identity yields necessary and sufficient conditions for optimality. These conditions characterize optimal transport plans and imply that almost every superlevel set of positive measure of a minimizing density minimizes the relative-perimeter ratio associated with the sum of the optimal transport potentials. They also give a transport-map representation when a prescribed measure is absolutely continuous and a criterion for the uniform density to be optimal.
Following the entropy approximation in [9 ] , we establish a second attained dual formulation in terms of bounded vector fields. The resulting optimality conditions retain the multiplier for the mass constraint and a nonnegative multiplier associated with the constraint ρ ≥ 0 \rho\geq0 ρ ≥ 0 . This construction does not require the prescribed measures to be absolutely continuous.
We further prove a maximum principle on convex domains and stability under approximation of the prescribed measures. In one dimension, a variance identity reduces the problem exactly to a TV-JKO step with the ordinary Wasserstein barycenter as reference measure. Finally, existence, uniqueness under the same absolute-continuity condition, and the perimeter dual formulation extend to costs ∣ x − y ∣ p |x-y|^p ∣ x − y ∣ p with p > 1 p>1 p > 1 .
Organization of the paper
Section 2 introduces the notation and recalls the analytical tools. Section 3 establishes existence, uniqueness, and the perimeter dual formulation. Section 4 derives the corresponding optimality conditions and transport representations. Section 5 develops the entropy approximation and proves vector-field duality. Section 6 contains the maximum principle, the one-dimensional reduction, and the stability result. Section 7 treats the extension to p p p -Wasserstein costs.
Notation and preliminaries
We introduce the notation used throughout the paper and recall the compactness, optimal transport, and minimax results needed below.
Notation
Throughout, Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d , d ≥ 1 d \ge1 d ≥ 1 , is a nonempty bounded connected open set with Lipschitz boundary. Set K = Ω ‾ K = \overline{\Omega} K = Ω and D = diam ( K ) D = \operatorname{diam}(K) D = diam ( K ) . Convexity of Ω \Omega Ω will be assumed only when explicitly stated. We denote by L d \mathcal{L}^{d} L d the Lebesgue measure, and use ∣ E ∣ |E| ∣ E ∣ and 1 E \mathbf{1}_{E} 1 E for the Lebesgue measure and characteristic function of a measurable set E E E . The notation ∣ ⋅ ∣ |\cdot| ∣ ⋅ ∣ also denotes the Euclidean norm, with the meaning determined by the argument. Unless another measure is specified, almost-everywhere statements refer to Lebesgue measure.
For 1 ≤ s < ∞ 1 \le s < \infty 1 ≤ s < ∞ , we use the norms
∥ u ∥ L s ( Ω ) = ( ∫ Ω ∣ u ∣ s d x ) 1 / s , ∥ u ∥ L ∞ ( Ω ) = ess sup x ∈ Ω ∣ u ( x ) ∣ . \|u\|_{L^{s}(\Omega)} = \left(\int_{\Omega} |u|^{s}\,dx\right)^{1/s}, \qquad\|u\|_{L^{\infty}(\Omega)} = \operatorname*{ess\,sup}_{x \in\Omega} |u(x)|. ∥ u ∥ L s ( Ω ) = ( ∫ Ω ∣ u ∣ s d x ) 1/ s , ∥ u ∥ L ∞ ( Ω ) = x ∈ Ω ess sup ∣ u ( x ) ∣.
The same notation is used for vector-valued functions, with the Euclidean norm inside the integral. We write C ( K ) C(K) C ( K ) for the real-valued continuous functions on K K K , equipped with the uniform norm
∥ f ∥ C ( K ) = max x ∈ K ∣ f ( x ) ∣ . \|f\|_{C(K)} = \max_{x \in K} |f(x)|. ∥ f ∥ C ( K ) = x ∈ K max ∣ f ( x ) ∣.
The abbreviation ∥ ⋅ ∥ ∞ \|\cdot\|_{\infty} ∥ ⋅ ∥ ∞ denotes the uniform norm for functions in C ( K ) C(K) C ( K ) and the essential supremum norm for measurable functions on Ω \Omega Ω . We use C c 1 ( Ω ; R d ) C^{1}_{c}(\Omega;\mathbb{R}^{d}) C c 1 ( Ω ; R d ) for continuously differentiable vector fields with compact support in Ω \Omega Ω . For a real-valued function u u u , set
u + = max { u , 0 } , u − = max { − u , 0 } , u ‾ = 1 ∣ Ω ∣ ∫ Ω u d x if u ∈ L 1 ( Ω ) . u_{+} = \max\{u,0\}, \qquad u_{-} = \max\{-u,0\}, \qquad\overline{u} = \frac{1}{|\Omega|}\int_{\Omega} u\,dx \quad\text{if } u \in L^{1}(\Omega). u + = max { u , 0 } , u − = max { − u , 0 } , u = ∣Ω∣ 1 ∫ Ω u d x if u ∈ L 1 ( Ω ) .
The symbols → \to → , ⇀ \rightharpoonup ⇀ , and ⇀ ∗ \rightharpoonup^{*} ⇀ ∗ denote strong, weak, and weak-star convergence in the indicated function spaces. For u ∈ L 1 ( Ω ) u \in L^{1}(\Omega) u ∈ L 1 ( Ω ) , its distributional gradient is denoted by D u Du D u . Define
TV ( u ) = sup { ∫ Ω u div z d x : z ∈ C c 1 ( Ω ; R d ) , ∥ z ∥ ∞ ≤ 1 } . (2) \operatorname{TV}(u) = \sup\left\{\int_{\Omega} u\,\operatorname{div} z\,dx : z \in C^{1}_{c}(\Omega;\mathbb{R}^{d}),\ \|z\|_{\infty} \le1\right\}. \tag*{(2)} TV ( u ) = sup { ∫ Ω u div z d x : z ∈ C c 1 ( Ω ; R d ) , ∥ z ∥ ∞ ≤ 1 } . (2)
where div z = ∑ j = 1 d ∂ j z j \operatorname{div} z = \sum_{j=1}^{d} \partial_{j}z_{j} div z = ∑ j = 1 d ∂ j z j . Then BV ( Ω ) = { u ∈ L 1 ( Ω ) : TV ( u ) < ∞ } \operatorname{BV}(\Omega) = \{u \in L^{1}(\Omega) : \operatorname{TV}(u) < \infty\} BV ( Ω ) = { u ∈ L 1 ( Ω ) : TV ( u ) < ∞ } . For u ∈ BV ( Ω ) u \in\operatorname{BV}(\Omega) u ∈ BV ( Ω ) , D u Du D u is a finite vector-valued Radon measure, and
TV ( u ) = ∣ D u ∣ ( Ω ) , ∥ u ∥ BV ( Ω ) = ∥ u ∥ L 1 ( Ω ) + ∣ D u ∣ ( Ω ) . \operatorname{TV}(u) = |Du|(\Omega), \qquad\|u\|_{\operatorname{BV}(\Omega)} = \|u\|_{L^{1}(\Omega)} + |Du|(\Omega). TV ( u ) = ∣ D u ∣ ( Ω ) , ∥ u ∥ BV ( Ω ) = ∥ u ∥ L 1 ( Ω ) + ∣ D u ∣ ( Ω ) .
Here ∣ D u ∣ |Du| ∣ D u ∣ denotes the total variation measure of D u Du D u . For functions outside BV ( Ω ) \operatorname{BV}(\Omega) BV ( Ω ) , the expression ∣ D u ∣ ( Ω ) |Du|(\Omega) ∣ D u ∣ ( Ω ) is understood as + ∞ +\infty + ∞ . For a measurable set E ⊂ Ω E \subset\Omega E ⊂ Ω , its relative perimeter is Per ( E ; Ω ) = ∣ D 1 E ∣ ( Ω ) \operatorname{Per}(E;\Omega) = |D\mathbf{1}_{E}|(\Omega) Per ( E ; Ω ) = ∣ D 1 E ∣ ( Ω ) . Thus only variation inside Ω \Omega Ω is counted.
We write P ( K ) \mathcal{P}(K) P ( K ) for the Borel probability measures on K K K . For p ≥ 1 p \ge1 p ≥ 1 , let
P p ( K ) = { μ ∈ P ( K ) : ∫ K ∣ x ∣ p d μ ( x ) < ∞ } . \mathcal{P}_{p}(K) = \left\{\mu\in\mathcal{P}(K) : \int_{K} |x|^{p}\,d\mu(x) < \infty\right\}. P p ( K ) = { μ ∈ P ( K ) : ∫ K ∣ x ∣ p d μ ( x ) < ∞ } .
Since K K K is compact, P p ( K ) = P ( K ) \mathcal{P}_p(K)=\mathcal{P}(K) P p ( K ) = P ( K ) . For μ , ν ∈ P ( K ) \mu,\nu\in\mathcal{P}(K) μ , ν ∈ P ( K ) , the set Π ( μ , ν ) \Pi(\mu,\nu) Π ( μ , ν ) consists of probability measures on K × K K\times K K × K with first marginal μ \mu μ and second marginal ν \nu ν . For a Borel map T T T , its pushforward is defined by ( T # μ ) ( B ) = μ ( T − 1 ( B ) ) (T_{\#}\mu)(B)=\mu(T^{-1}(B)) ( T # μ ) ( B ) = μ ( T − 1 ( B )) for every Borel set B B B in the target space. We denote the identity map by id \operatorname{id} id and the Dirac measure at x x x by δ x \delta_x δ x . Narrow convergence μ k ⇀ μ \mu_k\rightharpoonup\mu μ k ⇀ μ means
∫ K φ d μ k → ∫ K φ d μ for every φ ∈ C ( K ) . \int_K \varphi\,d\mu_k\to\int_K \varphi\,d\mu\qquad\text{for every }\varphi\in C(K). ∫ K φ d μ k → ∫ K φ d μ for every φ ∈ C ( K ) .
Optimal transport and analytical preliminaries
We first recall the compactness and coarea properties of B V BV B V functions, followed by the transport duality and minimax results used in the analysis.
On a bounded Lipschitz domain, every bounded sequence in B V ( Ω ) BV(\Omega) B V ( Ω ) has a subsequence converging strongly in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) to a function in B V ( Ω ) BV(\Omega) B V ( Ω ) . Moreover,
u k → u in L 1 ( Ω ) ⟹ ∣ D u ∣ ( Ω ) ≤ lim inf k → ∞ ∣ D u k ∣ ( Ω ) . u_k\to u\text{ in }L^1(\Omega)\quad\Longrightarrow\quad|Du|(\Omega)\leq\liminf_{k\to\infty}|Du_k|(\Omega). u k → u in L 1 ( Ω ) ⟹ ∣ D u ∣ ( Ω ) ≤ k → ∞ lim inf ∣ D u k ∣ ( Ω ) .
The embedding B V ( Ω ) ↪ L s ( Ω ) BV(\Omega)\hookrightarrow L^s(\Omega) B V ( Ω ) ↪ L s ( Ω ) is compact for 1 ≤ s < d / ( d − 1 ) 1\leq s<d/(d-1) 1 ≤ s < d / ( d − 1 ) when d ≥ 2 d\geq2 d ≥ 2 , and for every finite s ≥ 1 s\geq1 s ≥ 1 when d = 1 d=1 d = 1 . We also use the composition estimate ∣ D ( Φ ∘ u ) ∣ ( Ω ) ≤ Lip ( Φ ) ∣ D u ∣ ( Ω ) |D(\Phi\circ u)|(\Omega)\leq\operatorname{Lip}(\Phi)|Du|(\Omega) ∣ D ( Φ ∘ u ) ∣ ( Ω ) ≤ Lip ( Φ ) ∣ D u ∣ ( Ω ) for u ∈ B V ( Ω ) u\in BV(\Omega) u ∈ B V ( Ω ) and Lipschitz functions Φ : R → R \Phi:\mathbb{R}\to\mathbb{R} Φ : R → R . For nonnegative u ∈ B V ( Ω ) u\in BV(\Omega) u ∈ B V ( Ω ) , the coarea and layer-cake formulas give
∣ D u ∣ ( Ω ) = ∫ 0 ∞ Per ( { u > t } ; Ω ) d t . (3) |Du|(\Omega)=\int_0^\infty\operatorname{Per}(\{u>t\};\Omega)\,dt. \tag*{(3)} ∣ D u ∣ ( Ω ) = ∫ 0 ∞ Per ({ u > t } ; Ω ) d t . (3)
∫ Ω u d x = ∫ 0 ∞ ∣ { u > t } ∣ d t . (4) \int_\Omega u\,dx=\int_0^\infty|\{u>t\}|\,dt. \tag*{(4)} ∫ Ω u d x = ∫ 0 ∞ ∣ { u > t } ∣ d t . (4)
In particular, { u > t } \{u>t\} { u > t } has finite relative perimeter for almost every t > 0 t>0 t > 0 . If V ∈ L ∞ ( Ω ) V\in L^\infty(\Omega) V ∈ L ∞ ( Ω ) , Fubini’s theorem also yields
∫ Ω V u d x = ∫ 0 ∞ ∫ { u > t } V d x d t . \int_\Omega Vu\,dx=\int_0^\infty\int_{\{u>t\}}V\,dx\,dt. ∫ Ω V u d x = ∫ 0 ∞ ∫ { u > t } V d x d t .
We refer to [2 ] for these facts.
Throughout the pairing arguments, we use the conjugate exponents r : = d + 1 d r:=\frac{d+1}{d} r := d d + 1 , q : = d + 1 q:=d+1 q := d + 1 and 1 r + 1 q = 1 \frac{1}{r}+\frac{1}{q}=1 r 1 + q 1 = 1 . Thus u div z ∈ L 1 ( Ω ) u\operatorname{div}z\in L^1(\Omega) u div z ∈ L 1 ( Ω ) whenever u ∈ L r ( Ω ) u\in L^r(\Omega) u ∈ L r ( Ω ) and div z ∈ L q ( Ω ) \operatorname{div}z\in L^q(\Omega) div z ∈ L q ( Ω ) . For u ∈ B V ( Ω ) ∩ L r ( Ω ) u\in BV(\Omega)\cap L^r(\Omega) u ∈ B V ( Ω ) ∩ L r ( Ω ) and z ∈ L ∞ ( Ω ; R d ) z\in L^\infty(\Omega;\mathbb{R}^d) z ∈ L ∞ ( Ω ; R d ) with div z ∈ L q ( Ω ) \operatorname{div}z\in L^q(\Omega) div z ∈ L q ( Ω ) , the pairing ( z , D u ) (z,Du) ( z , D u ) is defined in the distributional sense by
∫ Ω φ d ( z , D u ) = − ∫ Ω u φ div z d x − ∫ Ω u z ⋅ ∇ φ d x , φ ∈ C c 1 ( Ω ) . (5) \int_\Omega\varphi\,d(z,Du)=-\int_\Omega u\varphi\operatorname{div}z\,dx-\int_\Omega uz\cdot\nabla\varphi\,dx,\qquad\varphi\in C_c^1(\Omega). \tag*{(5)} ∫ Ω φ d ( z , D u ) = − ∫ Ω u φ div z d x − ∫ Ω u z ⋅ ∇ φ d x , φ ∈ C c 1 ( Ω ) . (5)
The distribution in (5) is a finite Radon measure satisfying ∣ ( z , D u ) ∣ ≤ ∥ z ∥ ∞ ∣ D u ∣ |(z,Du)|\leq\|z\|_\infty|Du| ∣ ( z , D u ) ∣ ≤ ∥ z ∥ ∞ ∣ D u ∣ . If z z z has zero normal trace, the integration-by-parts formula gives
∫ Ω u div z d x = − ∫ Ω d ( z , D u ) ≤ ∥ z ∥ ∞ ∣ D u ∣ ( Ω ) . (6) \int_\Omega u\operatorname{div}z\,dx=-\int_\Omega d(z,Du)\leq\|z\|_\infty|Du|(\Omega). \tag*{(6)} ∫ Ω u div z d x = − ∫ Ω d ( z , D u ) ≤ ∥ z ∥ ∞ ∣ D u ∣ ( Ω ) . (6)
Here zero normal trace means that
∫ Ω ϕ div z d x = − ∫ Ω z ⋅ ∇ ϕ d x for every ϕ ∈ C 1 ( Ω ‾ ) . \int_\Omega\phi\operatorname{div}z\,dx=-\int_\Omega z\cdot\nabla\phi\,dx\qquad\text{for every }\phi\in C^1(\overline{\Omega}). ∫ Ω ϕ div z d x = − ∫ Ω z ⋅ ∇ ϕ d x for every ϕ ∈ C 1 ( Ω ) .
These pairing and trace properties are recalled from [4 ] ; see also their use in [9 ] .
For p ≥ 1 p\geq1 p ≥ 1 , set c p ( x , y ) = ∣ x − y ∣ p c_p(x,y)=|x-y|^p c p ( x , y ) = ∣ x − y ∣ p , and define the p p p -Wasserstein distance by
W p p ( μ , ν ) = min γ ∈ Π ( μ , ν ) ∫ K × K ∣ x − y ∣ p d γ ( x , y ) . W_p^p(\mu,\nu)=\min_{\gamma\in\Pi(\mu,\nu)}\int_{K\times K}|x-y|^p\,d\gamma(x,y). W p p ( μ , ν ) = γ ∈ Π ( μ , ν ) min ∫ K × K ∣ x − y ∣ p d γ ( x , y ) .
We write c = c 2 c=c_2 c = c 2 , so the quadratic cost is ∣ x − y ∣ 2 |x-y|^2 ∣ x − y ∣ 2 . For f ∈ C ( K ) f\in C(K) f ∈ C ( K ) , define
f c p ( x ) = min y ∈ K { ∣ x − y ∣ p − f ( y ) } . f^{c_p}(x)=\min_{y\in K}\{|x-y|^p-f(y)\}. f c p ( x ) = y ∈ K min { ∣ x − y ∣ p − f ( y )} .
In the quadratic case, this becomes
f c ( x ) = min y ∈ K { ∣ x − y ∣ 2 − f ( y ) } . (7) f^c(x)=\min_{y\in K}\{|x-y|^2-f(y)\}. \tag*{(7)} f c ( x ) = y ∈ K min { ∣ x − y ∣ 2 − f ( y )} . (7)
Iterated transforms are denoted by f c c f^{cc} f cc and f c c c f^{ccc} f ccc . A function is called c c c -concave if it is the c c c -transform of a function in C ( K ) C(K) C ( K ) ; the analogous convention applies to c p c_p c p . For probability measures on the compact set K K K , optimal transport plans exist. Moreover, narrow convergence is equivalent to convergence in W p W_p W p for each finite p ≥ 1 p\ge1 p ≥ 1 . Consequently, if μ k ⇀ μ \mu_k\rightharpoonup\mu μ k ⇀ μ and ν k ⇀ ν \nu_k\rightharpoonup\nu ν k ⇀ ν narrowly, then W p ( μ k , ν k ) → W p ( μ , ν ) W_p(\mu_k,\nu_k)\to W_p(\mu,\nu) W p ( μ k , ν k ) → W p ( μ , ν ) . For the quadratic cost, Kantorovich duality gives
W 2 2 ( μ , ν ) = max f ∈ C ( K ) { ∫ K f c d μ + ∫ K f d ν } . (8) W_2^2(\mu,\nu)=\max_{f\in C(K)}\left\{\int_K f^c\,d\mu+\int_K f\,d\nu\right\}. \tag*{(8)} W 2 2 ( μ , ν ) = f ∈ C ( K ) max { ∫ K f c d μ + ∫ K f d ν } . (8)
Thus, when ν = ρ d x \nu=\rho\,dx ν = ρ d x , the potential f f f is integrated against ρ d x \rho\,dx ρ d x , whereas f c f^c f c is integrated against μ \mu μ .
The c c c -transform satisfies f c c ≥ f f^{cc}\ge f f cc ≥ f , f c c c = f c f^{ccc}=f^c f ccc = f c , and
∥ f c − g c ∥ ∞ ≤ ∥ f − g ∥ ∞ , ( f + a ) c = f c − a for a ∈ R . (9) \|f^c-g^c\|_\infty\le\|f-g\|_\infty,\qquad(f+a)^c=f^c-a\quad\text{for }a\in\mathbb{R}. \tag*{(9)} ∥ f c − g c ∥ ∞ ≤ ∥ f − g ∥ ∞ , ( f + a ) c = f c − a for a ∈ R . (9)
Since
∣ ∣ x − y ∣ 2 − ∣ x − y ′ ∣ 2 ∣ ≤ 2 D ∣ y − y ′ ∣ , x , y , y ′ ∈ K , \bigl||x-y|^2-|x-y'|^2\bigr|\le2D|y-y'|,\qquad x,y,y'\in K, ∣ x − y ∣ 2 − ∣ x − y ′ ∣ 2 ≤ 2 D ∣ y − y ′ ∣ , x , y , y ′ ∈ K ,
every c c c -transform is 2 D 2D 2 D -Lipschitz. Replacing f f f by f c c f^{cc} f cc does not decrease the objective in (8) , and subtracting f c c ( x ∗ ) f^{cc}(x_*) f cc ( x ∗ ) leaves it unchanged because both measures have unit mass. The analogous statements hold for c p c_p c p , using
∣ ∣ x − y ∣ p − ∣ x − y ′ ∣ p ∣ ≤ p D p − 1 ∣ y − y ′ ∣ x , y , y ′ ∈ K . \bigl||x-y|^p-|x-y'|^p\bigr|\le pD^{p-1}|y-y'|\qquad x,y,y'\in K. ∣ x − y ∣ p − ∣ x − y ′ ∣ p ≤ p D p − 1 ∣ y − y ′ ∣ x , y , y ′ ∈ K .
If μ \mu μ is absolutely continuous with respect to Lebesgue measure, the optimal quadratic transport plan from μ \mu μ to ν \nu ν is unique and has the form γ = ( id , T ) # μ \gamma=(\operatorname{id},T)_\#\mu γ = ( id , T ) # μ and T # μ = ν T_\#\mu=\nu T # μ = ν . Let ψ \psi ψ be a c c c -concave optimal potential integrated against μ \mu μ . With our cost convention,
T ( x ) = x − 1 2 ∇ ψ ( x ) for μ -a.e. x . (10) T(x)=x-\frac{1}{2}\nabla\psi(x)\qquad\text{for }\mu\text{-a.e. }x. \tag*{(10)} T ( x ) = x − 2 1 ∇ ψ ( x ) for μ -a.e. x . (10)
Indeed, the optimality relation ψ ( x ) + ψ c ( T ( x ) ) = ∣ x − T ( x ) ∣ 2 \psi(x)+\psi^c(T(x))=|x-T(x)|^2 ψ ( x ) + ψ c ( T ( x )) = ∣ x − T ( x ) ∣ 2 . For c p ( x , y ) = ∣ x − y ∣ p c_p(x,y)=|x-y|^p c p ( x , y ) = ∣ x − y ∣ p , p > 1 p>1 p > 1 , absolute continuity of the first marginal likewise guarantees a unique optimal plan induced by a map. We refer to [24 Theorem 1.17 and Remark 1.18] .
Finally, we recall the convex-concave form of Sion’s minimax theorem.
Lemma 2.1 (Sion’s minimax theorem [25 ] ). Let X X X and Y Y Y be nonempty convex subsets of Hausdorff topological vector spaces, and suppose that Y Y Y is compact. Let G : X × Y → R G:X\times Y\to\mathbb{R} G : X × Y → R satisfy the following conditions:
(1) For every y ∈ Y y\in Y y ∈ Y , the function x ↦ G ( x , y ) x\mapsto G(x,y) x ↦ G ( x , y ) is convex and lower semicontinuous on X X X .
(2) For every x ∈ X x\in X x ∈ X , the function y ↦ G ( x , y ) y\mapsto G(x,y) y ↦ G ( x , y ) is concave and upper semicontinuous on Y Y Y .
Then
inf x ∈ X sup y ∈ Y G ( x , y ) = sup y ∈ Y inf x ∈ X G ( x , y ) . \inf_{x\in X}\sup_{y\in Y}G(x,y)=\sup_{y\in Y}\inf_{x\in X}G(x,y). x ∈ X inf y ∈ Y sup G ( x , y ) = y ∈ Y sup x ∈ X inf G ( x , y ) . :::`
The primal problem and its dual
We first establish existence for arbitrary prescribed probability measures and uniqueness when at least one of them is absolutely continuous. We then derive an attained dual formulation in terms of transport potentials and relative perimeters.
Given an integer n ≥ 2 n \ge2 n ≥ 2 , let μ 1 , … , μ n ∈ P ( K ) \mu_1,\ldots,\mu_n \in\mathcal{P}(K) μ 1 , … , μ n ∈ P ( K ) , consider
inf ρ ∈ A J ( ρ ) , J ( ρ ) : = ∑ i = 1 n W 2 2 ( μ i , ρ d x ) + ∣ D ρ ∣ ( Ω ) . (11) \inf_{\rho\in\mathcal{A}} J(\rho), \qquad J(\rho) := \sum_{i=1}^{n} W_2^2(\mu_i,\rho\,dx) + |D\rho|(\Omega). \tag*{(11)} ρ ∈ A inf J ( ρ ) , J ( ρ ) := i = 1 ∑ n W 2 2 ( μ i , ρ d x ) + ∣ D ρ ∣ ( Ω ) . (11)
Existence and uniqueness
Theorem 3.1 (Existence and uniqueness). Problem (3.1) admits a minimizer for arbitrary prescribed measures in P ( K ) \mathcal{P}(K) P ( K ) . If at least one μ i \mu_i μ i is absolutely continuous with respect to Lebesgue measure, the minimizer is unique.
Proof. To see that the infimum is finite, consider the constant density ρ 0 = ∣ Ω ∣ − 1 \rho_0 = |\Omega|^{-1} ρ 0 = ∣Ω ∣ − 1 . Since Ω \Omega Ω is a nonempty bounded open set, we have 0 < ∣ Ω ∣ < ∞ 0 < |\Omega| < \infty 0 < ∣Ω∣ < ∞ . Thus ρ 0 ≥ 0 \rho_0 \ge0 ρ 0 ≥ 0 and ∫ Ω ρ 0 ( y ) d y = 1 \int_{\Omega} \rho_0(y)\,dy = 1 ∫ Ω ρ 0 ( y ) d y = 1 . Moreover, ρ 0 \rho_0 ρ 0 is constant in Ω \Omega Ω , so its distributional derivative in Ω \Omega Ω vanishes. Consequently, ρ 0 ∈ B V ( Ω ) \rho_0 \in BV(\Omega) ρ 0 ∈ B V ( Ω ) and ∣ D ρ 0 ∣ ( Ω ) = 0 |D\rho_0|(\Omega) = 0 ∣ D ρ 0 ∣ ( Ω ) = 0 , which shows that ρ 0 ∈ A \rho_0 \in\mathcal{A} ρ 0 ∈ A . For each i = 1 , … , n i = 1,\ldots,n i = 1 , … , n , the product measure π i : = μ i ⊗ ρ 0 d y \pi_i := \mu_i \otimes\rho_0\,dy π i := μ i ⊗ ρ 0 d y belongs to Π ( μ i , ρ 0 d x ) \Pi(\mu_i,\rho_0\,dx) Π ( μ i , ρ 0 d x ) . By the definition of the quadratic Wasserstein distance and the bound ∣ x − y ∣ ≤ D |x-y| \le D ∣ x − y ∣ ≤ D for x , y ∈ Ω x,y \in\Omega x , y ∈ Ω , we obtain
W 2 2 ( μ i , ρ 0 d y ) = inf π ∈ Π ( μ i , ρ 0 d x ) ∫ Ω ‾ × Ω ‾ ∣ x − y ∣ 2 d π ( x , y ) ≤ ∫ Ω ‾ × Ω ‾ ∣ x − y ∣ 2 d ( μ i ⊗ ρ 0 d y ) ( x , y ) ≤ D 2 . \begin{aligned}
W_2^2(\mu_i,\rho_0\,dy)
= \inf_{\pi\in\Pi(\mu_i,\rho_0\,dx)} \int_{\overline{\Omega}\times\overline{\Omega}} |x-y|^2\,d\pi(x,y) \\
&\le\int_{\overline{\Omega}\times\overline{\Omega}} |x-y|^2\,d(\mu_i \otimes\rho_0\,dy)(x,y) \le D^2.
\end{aligned} W 2 2 ( μ i , ρ 0 d y ) = π ∈ Π ( μ i , ρ 0 d x ) inf ∫ Ω × Ω ∣ x − y ∣ 2 d π ( x , y ) ≤ ∫ Ω × Ω ∣ x − y ∣ 2 d ( μ i ⊗ ρ 0 d y ) ( x , y ) ≤ D 2 .
Therefore,
0 ≤ inf ρ ∈ A J ( ρ ) ≤ J ( ρ 0 ) = ∑ i = 1 n W 2 2 ( μ i , ρ 0 d y ) + ∣ D ρ 0 ∣ ( Ω ) ≤ n D 2 < ∞ . 0 \le\inf_{\rho\in\mathcal{A}} J(\rho) \le J(\rho_0) = \sum_{i=1}^{n} W_2^2(\mu_i,\rho_0\,dy) + |D\rho_0|(\Omega) \le nD^2 < \infty. 0 ≤ ρ ∈ A inf J ( ρ ) ≤ J ( ρ 0 ) = i = 1 ∑ n W 2 2 ( μ i , ρ 0 d y ) + ∣ D ρ 0 ∣ ( Ω ) ≤ n D 2 < ∞.
Let ( ρ k ) ⊂ A (\rho_k) \subset\mathcal{A} ( ρ k ) ⊂ A be a minimizing sequence. Since the infimum is finite, we may assume that sup k J ( ρ k ) < ∞ \sup_k J(\rho_k) < \infty sup k J ( ρ k ) < ∞ . The Wasserstein terms are nonnegative, and hence ∣ D ρ k ∣ ( Ω ) ≤ J ( ρ k ) |D\rho_k|(\Omega) \le J(\rho_k) ∣ D ρ k ∣ ( Ω ) ≤ J ( ρ k ) . Moreover, the nonnegativity and unit mass of ρ k \rho_k ρ k imply
sup k ∥ ρ k ∥ B V ( Ω ) = sup k ( ∥ ρ k ∥ L 1 ( Ω ) + ∣ D ρ k ∣ ( Ω ) ) < ∞ . \sup_k \|\rho_k\|_{BV(\Omega)} = \sup_k \left(\|\rho_k\|_{L^1(\Omega)} + |D\rho_k|(\Omega)\right) < \infty. k sup ∥ ρ k ∥ B V ( Ω ) = k sup ( ∥ ρ k ∥ L 1 ( Ω ) + ∣ D ρ k ∣ ( Ω ) ) < ∞.
Since Ω \Omega Ω is a bounded Lipschitz domain, the compact embedding of B V ( Ω ) BV(\Omega) B V ( Ω ) into L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) yields a subsequence, not relabeled, and a function ρ ∈ B V ( Ω ) \rho\in BV(\Omega) ρ ∈ B V ( Ω ) such that ρ k → ρ \rho_k \to\rho ρ k → ρ strongly in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) . Thus, the mass constraint also passes to the limit. In particular, writing ρ − : = max { − ρ , 0 } \rho^- := \max\{-\rho,0\} ρ − := max { − ρ , 0 } and using ρ k ≥ 0 \rho_k \ge0 ρ k ≥ 0 , we obtain
∥ ρ − ∥ L 1 ( Ω ) ≤ ∥ ρ − ρ k ∥ L 1 ( Ω ) → 0. \|\rho^-\|_{L^1(\Omega)} \le\|\rho-\rho_k\|_{L^1(\Omega)} \to0. ∥ ρ − ∥ L 1 ( Ω ) ≤ ∥ ρ − ρ k ∥ L 1 ( Ω ) → 0.
Thus ρ ≥ 0 \rho\ge0 ρ ≥ 0 almost everywhere. Furthermore, the lower semicontinuity of the total variation under L 1 L^1 L 1 convergence gives
∣ D ρ ∣ ( Ω ) ≤ lim inf k → ∞ ∣ D ρ k ∣ ( Ω ) . |D\rho|(\Omega) \le\liminf_{k\to\infty} |D\rho_k|(\Omega). ∣ D ρ ∣ ( Ω ) ≤ k → ∞ lim inf ∣ D ρ k ∣ ( Ω ) .
Therefore, ρ ∈ A \rho\in\mathcal{A} ρ ∈ A . The strong convergence ρ k → ρ \rho_k \to\rho ρ k → ρ in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) implies that ρ k d x \rho_k\,dx ρ k d x converges narrowly to ρ d x \rho\,dx ρ d x as probability measures on K K K . By the lower semicontinuity of the quadratic Wasserstein cost under narrow convergence,
W 2 2 ( μ i , ρ d x ) ≤ lim inf k → ∞ W 2 2 ( μ i , ρ k d x ) , for i = 1 , … , n . W_2^2(\mu_i,\rho\,dx) \le\liminf_{k\to\infty} W_2^2(\mu_i,\rho_k\,dx), \qquad\text{for } i=1,\ldots,n. W 2 2 ( μ i , ρ d x ) ≤ k → ∞ lim inf W 2 2 ( μ i , ρ k d x ) , for i = 1 , … , n .
Together with the lower semicontinuity of the total variation, this yields
J ( ρ ) = ∑ i = 1 n W 2 2 ( μ i , ρ d x ) + ∣ D ρ ∣ ( Ω ) ≤ lim inf k → ∞ [ ∑ i = 1 n W 2 2 ( μ i , ρ k d x ) + ∣ D ρ k ∣ ( Ω ) ] = inf η ∈ A J ( η ) . J(\rho)=\sum_{i=1}^{n}W_2^2(\mu_i,\rho\,dx)+|D\rho|(\Omega)
\le\liminf_{k\to\infty}\left[\sum_{i=1}^{n}W_2^2(\mu_i,\rho_k\,dx)+|D\rho_k|(\Omega)\right]=\inf_{\eta\in\mathcal{A}}J(\eta). J ( ρ ) = i = 1 ∑ n W 2 2 ( μ i , ρ d x ) + ∣ D ρ ∣ ( Ω ) ≤ k → ∞ lim inf [ i = 1 ∑ n W 2 2 ( μ i , ρ k d x ) + ∣ D ρ k ∣ ( Ω ) ] = η ∈ A inf J ( η ) .
Therefore, ρ \rho ρ is a minimizer.
For uniqueness, suppose μ j \mu_j μ j is absolutely continuous. Let ν 0 ≠ ν 1 \nu_0\ne\nu_1 ν 0 = ν 1 be probability measures and let γ a = ( id , T a ) # μ j \gamma_a=(\operatorname{id},T_a)_{\#}\mu_j γ a = ( id , T a ) # μ j be the optimal plans from μ j \mu_j μ j to ν a \nu_a ν a , a = 0 , 1 a=0,1 a = 0 , 1 . For 0 < s < 1 0<s<1 0 < s < 1 , the plan γ s : = ( 1 − s ) γ 0 + s γ 1 \gamma_s:=(1-s)\gamma_0+s\gamma_1 γ s := ( 1 − s ) γ 0 + s γ 1 transports μ j \mu_j μ j to ν s = ( 1 − s ) ν 0 + s ν 1 \nu_s=(1-s)\nu_0+s\nu_1 ν s = ( 1 − s ) ν 0 + s ν 1 . Hence
W 2 2 ( μ j , ν s ) ≤ ( 1 − s ) W 2 2 ( μ j , ν 0 ) + s W 2 2 ( μ j , ν 1 ) . W_2^2(\mu_j,\nu_s)\le(1-s)W_2^2(\mu_j,\nu_0)+sW_2^2(\mu_j,\nu_1). W 2 2 ( μ j , ν s ) ≤ ( 1 − s ) W 2 2 ( μ j , ν 0 ) + s W 2 2 ( μ j , ν 1 ) .
Suppose that equality holds. Then γ s \gamma_s γ s attains the minimum transport cost from μ j \mu_j μ j to ν s \nu_s ν s , and hence is an optimal plan. Since μ j \mu_j μ j is absolutely continuous, Brenier’s theorem implies that γ s \gamma_s γ s is induced by a measurable map T s T_s T s , γ s = ( id , T s ) # μ j \gamma_s=(\operatorname{id},T_s)_{\#}\mu_j γ s = ( id , T s ) # μ j . By uniqueness of disintegration,
( 1 − s ) δ T 0 ( x ) + s δ T 1 ( x ) = δ T s ( x ) for μ j -almost every x . (1-s)\delta_{T_0(x)}+s\delta_{T_1(x)}=\delta_{T_s(x)} \qquad\text{for }\mu_j\text{-almost every }x. ( 1 − s ) δ T 0 ( x ) + s δ T 1 ( x ) = δ T s ( x ) for μ j -almost every x .
Both coefficients 1 − s 1-s 1 − s and s s s are strictly positive. If T 0 ( x ) ≠ T 1 ( x ) T_0(x)\ne T_1(x) T 0 ( x ) = T 1 ( x ) , the measure on the left assigns positive mass to two distinct points and therefore cannot be a Dirac mass. Consequently, T 0 ( x ) = T 1 ( x ) T_0(x)=T_1(x) T 0 ( x ) = T 1 ( x ) for μ j \mu_j μ j -almost every x x x . This implies that ν 0 = ( T 0 ) # μ j = ( T 1 ) # μ j = ν 1 \nu_0=(T_0)_{\#}\mu_j=(T_1)_{\#}\mu_j=\nu_1 ν 0 = ( T 0 ) # μ j = ( T 1 ) # μ j = ν 1 , contrary to our assumption. Therefore,
W 2 2 ( μ j , ν s ) < ( 1 − s ) W 2 2 ( μ j , ν 0 ) + s W 2 2 ( μ j , ν 1 ) whenever ν 0 ≠ ν 1 . W_2^2(\mu_j,\nu_s)<(1-s)W_2^2(\mu_j,\nu_0)+sW_2^2(\mu_j,\nu_1) \qquad\text{whenever }\nu_0\ne\nu_1. W 2 2 ( μ j , ν s ) < ( 1 − s ) W 2 2 ( μ j , ν 0 ) + s W 2 2 ( μ j , ν 1 ) whenever ν 0 = ν 1 .
Thus the displayed inequality is strict. Since total variation is convex, J J J is strictly convex on the convex set A \mathcal{A} A , which proves uniqueness. □
Remark 3.1. The uniqueness argument is the classical transport argument underlying Wasserstein barycenter uniqueness [1 ] . It is independent of the dual formulation. Without an absolutely continuous prescribed measure, the theorem asserts existence but makes no general uniqueness claim.
A dual formulation through relative perimeter
We now seek a dual expression for the minimum in (11) . The mass and nonnegativity constraints lead naturally to an inequality involving the sum of the transport potentials. Before establishing the dual formula, we explain why imposing pointwise cancellation can give a smaller value than the primal minimum. To motivate a dual formulation, let f 1 , … , f n ∈ C ( K ) f_1,\ldots,f_n\in C(K) f 1 , … , f n ∈ C ( K ) and z ∈ C c 1 ( Ω ; R d ) z\in C_c^1(\Omega;\mathbb{R}^d) z ∈ C c 1 ( Ω ; R d ) with ∥ z ∥ ∞ ≤ 1 \lVert z\rVert_\infty\le1 ∥ z ∥ ∞ ≤ 1 . For every ρ ∈ A \rho\in\mathcal{A} ρ ∈ A , Kantorovich duality gives
W 2 2 ( μ i , ρ d x ) ≥ ∫ K f i c d μ i + ∫ Ω f i ρ d x , i = 1 , … , n . W_2^2(\mu_i,\rho\,dx)\ge\int_K f_i^c\,d\mu_i+\int_\Omega f_i\rho\,dx,\qquad i=1,\ldots,n. W 2 2 ( μ i , ρ d x ) ≥ ∫ K f i c d μ i + ∫ Ω f i ρ d x , i = 1 , … , n .
Moreover, the definition of the total variation yields ∣ D ρ ∣ ( Ω ) ≥ ∫ Ω ρ div z d x |D\rho|(\Omega) \ge\int_{\Omega} \rho\operatorname{div} z\,dx ∣ D ρ ∣ ( Ω ) ≥ ∫ Ω ρ div z d x . Adding these inequalities, we obtain
J ( ρ ) ≥ ∑ i = 1 n ∫ K f i c d μ i + ∫ Ω ( ∑ i = 1 n f i + div z ) ρ d x . J(\rho) \ge\sum_{i=1}^{n} \int_{K} f_i^c\,d\mu_i + \int_{\Omega} \left(\sum_{i=1}^{n} f_i + \operatorname{div} z\right)\rho\,dx. J ( ρ ) ≥ i = 1 ∑ n ∫ K f i c d μ i + ∫ Ω ( i = 1 ∑ n f i + div z ) ρ d x .
Since every ρ ∈ A \rho\in\mathcal{A} ρ ∈ A is nonnegative and has unit mass, the condition ∑ i = 1 n f i + div z ≥ λ \sum_{i=1}^{n} f_i + \operatorname{div} z \ge\lambda ∑ i = 1 n f i + div z ≥ λ in Ω \Omega Ω implies
J ( ρ ) ≥ λ + ∑ i = 1 n ∫ K f i c d μ i . J(\rho) \ge\lambda+ \sum_{i=1}^{n} \int_{K} f_i^c\,d\mu_i. J ( ρ ) ≥ λ + i = 1 ∑ n ∫ K f i c d μ i .
Thus a lower bound independent of ρ \rho ρ can be obtained without requiring ∑ i = 1 n f i + div z \sum_{i=1}^{n} f_i + \operatorname{div} z ∑ i = 1 n f i + div z to vanish identically. Imposing this identity gives a more restrictive class of competitors. The following example shows that the resulting supremum can be strictly smaller than the primal minimum.
Example 3.1. Let Ω = ( 0 , 1 ) \Omega= (0,1) Ω = ( 0 , 1 ) , n = 2 n=2 n = 2 , and μ 1 = μ 2 = 101 ( 0 , 1 / 10 ) d x \mu_1 = \mu_2 = 10\mathbf{1}_{(0,1/10)}\,dx μ 1 = μ 2 = 10 1 ( 0 , 1/10 ) d x . In this case, the candidate dual value is
D 0 = sup { ∑ i = 1 2 ∫ [ 0 , 1 ] f i c d μ i : f i ∈ C ( [ 0 , 1 ] ) , z ∈ C c 1 ( ( 0 , 1 ) ) , ∥ z ∥ ∞ ≤ 1 , f 1 + f 2 = − z ′ in ( 0 , 1 ) } . D_0 = \sup\left\{ \sum_{i=1}^{2} \int_{[0,1]} f_i^c\,d\mu_i : f_i \in C([0,1]),\ z \in C_c^1((0,1)),\ \|z\|_{\infty} \le1,\ f_1+f_2=-z' \text{ in }(0,1) \right\}. D 0 = sup { i = 1 ∑ 2 ∫ [ 0 , 1 ] f i c d μ i : f i ∈ C ([ 0 , 1 ]) , z ∈ C c 1 (( 0 , 1 )) , ∥ z ∥ ∞ ≤ 1 , f 1 + f 2 = − z ′ in ( 0 , 1 ) } . Since z z z is compactly supported in ( 0 , 1 ) (0,1) ( 0 , 1 ) , its derivative vanishes in a neighborhood of 0. The constraint and continuity of the potentials therefore give f 1 ( 0 ) + f 2 ( 0 ) = 0 f_1(0) + f_2(0) = 0 f 1 ( 0 ) + f 2 ( 0 ) = 0 . Taking y = 0 y=0 y = 0 in the definition
f i c ( x ) = min y ∈ [ 0 , 1 ] { ∣ x − y ∣ 2 − f i ( y ) } , f_i^c(x) = \min_{y\in[0,1]} \left\{|x-y|^2-f_i(y)\right\}, f i c ( x ) = y ∈ [ 0 , 1 ] min { ∣ x − y ∣ 2 − f i ( y ) } , we obtain that f i c ( x ) ≤ x 2 − f i ( 0 ) f_i^c(x) \le x^2-f_i(0) f i c ( x ) ≤ x 2 − f i ( 0 ) . Since μ 1 = μ 2 \mu_1=\mu_2 μ 1 = μ 2 , every admissible choice satisfies
∑ i = 1 2 ∫ [ 0 , 1 ] f i c d μ i ≤ 10 ∫ 0 1 / 10 ( 2 x 2 − f 1 ( 0 ) − f 2 ( 0 ) ) d x \sum_{i=1}^{2} \int_{[0,1]} f_i^c\,d\mu_i \le10\int_{0}^{1/10} \left(2x^2-f_1(0)-f_2(0)\right)\,dx i = 1 ∑ 2 ∫ [ 0 , 1 ] f i c d μ i ≤ 10 ∫ 0 1/10 ( 2 x 2 − f 1 ( 0 ) − f 2 ( 0 ) ) d x = 20 ∫ 0 1 / 10 x 2 d x = 1 150 . =20\int_{0}^{1/10}x^2\,dx=\frac{1}{150}. = 20 ∫ 0 1/10 x 2 d x = 150 1 . Hence D 0 ≤ 1 / 150 D_0 \le1/150 D 0 ≤ 1/150 . We next derive a strictly larger lower bound for the primal problem. Let ρ ∈ A \rho\in\mathcal{A} ρ ∈ A and set M : = ∣ D ρ ∣ ( ( 0 , 1 ) ) M:=|D\rho|((0,1)) M := ∣ D ρ ∣ (( 0 , 1 )) . If M ≤ 1 M\le1 M ≤ 1 , the one-dimensional B V BV B V oscillation estimate, together with ∫ 0 1 ρ d x = 1 \int_{0}^{1}\rho\,dx=1 ∫ 0 1 ρ d x = 1 , yields
ess sup ( 0 , 1 ) ρ ≤ ess inf ( 0 , 1 ) ρ + M ≤ 1 + M ≤ 2. \operatorname*{ess\,sup}_{(0,1)}\rho\le\operatorname*{ess\,inf}_{(0,1)}\rho+M\le1+M\le2. ( 0 , 1 ) ess sup ρ ≤ ( 0 , 1 ) ess inf ρ + M ≤ 1 + M ≤ 2. It follows that
∫ 1 / 5 1 ρ ( x ) d x = 1 − ∫ 0 1 / 5 ρ ( x ) d x ≥ 1 − 2 5 = 3 5 . \int_{1/5}^{1}\rho(x)\,dx=1-\int_{0}^{1/5}\rho(x)\,dx\ge1-\frac{2}{5}=\frac{3}{5}. ∫ 1/5 1 ρ ( x ) d x = 1 − ∫ 0 1/5 ρ ( x ) d x ≥ 1 − 5 2 = 5 3 . For each prescribed measure, all the source mass lies in [ 0 , 1 / 10 ] [0,1/10] [ 0 , 1/10 ] . Any transport plan to ρ d x \rho\,dx ρ d x must therefore move at least 3 / 5 3/5 3/5 of its mass over a distance of at least 1 / 10 1/10 1/10 . Consequently,
W 2 2 ( μ i , ρ d x ) ≥ 3 5 ( 1 10 ) 2 , i = 1 , 2 , W_2^2(\mu_i,\rho\,dx)\ge\frac{3}{5}\left(\frac{1}{10}\right)^2,\qquad i=1,2, W 2 2 ( μ i , ρ d x ) ≥ 5 3 ( 10 1 ) 2 , i = 1 , 2 , and thus
J ( ρ ) ≥ 2 ⋅ 3 5 ( 1 10 ) 2 = 3 250 . J(\rho)\ge2\cdot\frac{3}{5}\left(\frac{1}{10}\right)^2=\frac{3}{250}. J ( ρ ) ≥ 2 ⋅ 5 3 ( 10 1 ) 2 = 250 3 .
If M > 1 M>1 M > 1 , then J ( ρ ) ≥ M > 1 J(\rho)\ge M>1 J ( ρ ) ≥ M > 1 , so the same lower bound holds. We conclude that
D 0 ≤ 1 150 < 3 250 ≤ inf ρ ∈ A J ( ρ ) . D_0\le\frac{1}{150}<\frac{3}{250}\le\inf_{\rho\in\mathcal{A}}J(\rho). D 0 ≤ 150 1 < 250 3 ≤ ρ ∈ A inf J ( ρ ) .
Thus the supremum in the proposed dual formulation is strictly smaller than the primal minimum. Restricting the potentials further to C c ( Ω ) C_c(\Omega) C c ( Ω ) can only decrease this supremum, so the strict duality gap persists under this additional restriction.
This example concerns the simultaneous imposition of compact support on the vector field and the pointwise cancellation condition ∑ i f i + div z = 0 \sum_i f_i+\operatorname{div}z=0 ∑ i f i + div z = 0 . It does not apply to the weak vector-field formulation of Carlier and Poon [9 Theorem 3.3] , which uses zero normal trace and a nonnegative multiplier. To retain the mass and nonnegativity constraints explicitly, for V ∈ C ( Ω ‾ ) V\in C(\overline{\Omega}) V ∈ C ( Ω ) , we introduce the auxiliary minimization problem
m ( V ) = inf ρ ∈ A { ∣ D ρ ∣ ( Ω ) + ∫ Ω V ρ d x } . (12) m(V)=\inf_{\rho\in\mathcal{A}}\left\{|D\rho|(\Omega)+\int_{\Omega}V\rho\,dx\right\}. \tag*{(12)} m ( V ) = ρ ∈ A inf { ∣ D ρ ∣ ( Ω ) + ∫ Ω V ρ d x } . (12)
The following lemma expresses this value in terms of relative perimeters and provides the basis for the dual formulation.
Lemma 3.1. The infimum in (3.2) is attained, and
m ( V ) = inf E ⊂ Ω , ∣ E ∣ > 0 Per ( E ; Ω ) < ∞ Per ( E ; Ω ) + ∫ E V d x ∣ E ∣ . (13) m(V)=\inf_{\substack{E\subset\Omega,\ |E|>0\\ \operatorname{Per}(E;\Omega)<\infty}}\frac{\operatorname{Per}(E;\Omega)+\int_E V\,dx}{|E|}. \tag*{(13)} m ( V ) = E ⊂ Ω , ∣ E ∣ > 0 Per ( E ; Ω ) < ∞ inf ∣ E ∣ Per ( E ; Ω ) + ∫ E V d x . (13) Moreover, m m m is concave and ∣ m ( V ) − m ( W ) ∣ ≤ ∥ V − W ∥ ∞ |m(V)-m(W)|\le\|V-W\|_\infty ∣ m ( V ) − m ( W ) ∣ ≤ ∥ V − W ∥ ∞ .
Proof. Since V ∈ C ( K ) V\in C(K) V ∈ C ( K ) and K K K is compact, V V V is bounded. For every ρ ∈ A \rho\in\mathcal{A} ρ ∈ A , nonnegativity and unit mass imply
∣ ∫ Ω V ρ d x ∣ ≤ ∥ V ∥ ∞ . \left|\int_{\Omega}V\rho\,dx\right|\le\|V\|_\infty. ∫ Ω V ρ d x ≤ ∥ V ∥ ∞ .
In particular, the functional is bounded below by − ∥ V ∥ ∞ -\|V\|_\infty − ∥ V ∥ ∞ . Testing with the constant density ρ 0 = ∣ Ω ∣ − 1 \rho_0=|\Omega|^{-1} ρ 0 = ∣Ω ∣ − 1 gives
− ∥ V ∥ ∞ ≤ m ( V ) ≤ 1 ∣ Ω ∣ ∫ Ω V d x ≤ ∥ V ∥ ∞ . -\|V\|_\infty\le m(V)\le\frac{1}{|\Omega|}\int_{\Omega}V\,dx\le\|V\|_\infty. − ∥ V ∥ ∞ ≤ m ( V ) ≤ ∣Ω∣ 1 ∫ Ω V d x ≤ ∥ V ∥ ∞ .
Let ( ρ k ) ⊂ A (\rho_k)\subset\mathcal{A} ( ρ k ) ⊂ A be a minimizing sequence. After discarding finitely many terms, we may assume that
∣ D ρ k ∣ ( Ω ) + ∫ Ω V ρ k d x ≤ m ( V ) + 1. |D\rho_k|(\Omega)+\int_{\Omega}V\rho_k\,dx\le m(V)+1. ∣ D ρ k ∣ ( Ω ) + ∫ Ω V ρ k d x ≤ m ( V ) + 1.
Consequently, ∣ D ρ k ∣ ( Ω ) ≤ 2 ∥ V ∥ ∞ + 1 |D\rho_k|(\Omega)\le2\|V\|_\infty+1 ∣ D ρ k ∣ ( Ω ) ≤ 2∥ V ∥ ∞ + 1 . Together with ∥ ρ k ∥ L 1 ( Ω ) = 1 \|\rho_k\|_{L^1(\Omega)}=1 ∥ ρ k ∥ L 1 ( Ω ) = 1 , this shows that ( ρ k ) (\rho_k) ( ρ k ) is bounded in B V ( Ω ) BV(\Omega) B V ( Ω ) . The compactness theorem for B V BV B V functions therefore yields a subsequence converging strongly in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) to some ρ ∈ A \rho\in\mathcal{A} ρ ∈ A , as in the preceding existence proof. Since
∣ ∫ Ω V ( ρ k − ρ ) d x ∣ ≤ ∥ V ∥ ∞ ∥ ρ k − ρ ∥ L 1 ( Ω ) → 0 , \left|\int_{\Omega}V(\rho_k-\rho)\,dx\right|\le\|V\|_\infty\|\rho_k-\rho\|_{L^1(\Omega)}\to0, ∫ Ω V ( ρ k − ρ ) d x ≤ ∥ V ∥ ∞ ∥ ρ k − ρ ∥ L 1 ( Ω ) → 0 ,
the lower semicontinuity of the total variation gives
∣ D ρ ∣ ( Ω ) + ∫ Ω V ρ d x ≤ lim inf k → ∞ ( ∣ D ρ k ∣ ( Ω ) + ∫ Ω V ρ k d x ) = m ( V ) . |D\rho|(\Omega)+\int_{\Omega}V\rho\,dx\le\liminf_{k\to\infty}\left(|D\rho_k|(\Omega)+\int_{\Omega}V\rho_k\,dx\right)=m(V). ∣ D ρ ∣ ( Ω ) + ∫ Ω V ρ d x ≤ k → ∞ lim inf ( ∣ D ρ k ∣ ( Ω ) + ∫ Ω V ρ k d x ) = m ( V ) .
Since ρ ∈ A \rho\in\mathcal{A} ρ ∈ A , equality holds, and the infimum defining m ( V ) m(V) m ( V ) is attained.
Set
a : = inf E ⊂ Ω , ∣ E ∣ > 0 Per ( E ; Ω ) < ∞ Per ( E ; Ω ) + ∫ E V d x ∣ E ∣ . a:=\inf_{\substack{E\subset\Omega,\ |E|>0\\ \operatorname{Per}(E;\Omega)<\infty}}\frac{\operatorname{Per}(E;\Omega)+\int_E V\,dx}{|E|}. a := E ⊂ Ω , ∣ E ∣ > 0 Per ( E ; Ω ) < ∞ inf ∣ E ∣ Per ( E ; Ω ) + ∫ E V d x .
Since the relative perimeter is nonnegative and Per ( Ω ; Ω ) = 0 \operatorname{Per}(\Omega;\Omega)=0 Per ( Ω ; Ω ) = 0 . Taking E = Ω E=\Omega E = Ω gives min K V ≤ a ≤ 1 ∣ Ω ∣ ∫ Ω V d x \min_K V\le a\le\frac{1}{|\Omega|}\int_\Omega V\,dx min K V ≤ a ≤ ∣Ω∣ 1 ∫ Ω V d x . In particular, a a a is finite. For every set E E E in the above infimum, the normalized characteristic function ρ E : = 1 ∣ E ∣ 1 E \rho_E:=\frac{1}{|E|}\mathbf{1}_E ρ E := ∣ E ∣ 1 1 E belongs to A \mathcal{A} A . By homogeneity of the total variation,
∣ D ρ E ∣ ( Ω ) + ∫ Ω V ρ E d x = Per ( E ; Ω ) + ∫ E V d x ∣ E ∣ . |D\rho_E|(\Omega)+\int_\Omega V\rho_E\,dx=\frac{\operatorname{Per}(E;\Omega)+\int_E V\,dx}{|E|}. ∣ D ρ E ∣ ( Ω ) + ∫ Ω V ρ E d x = ∣ E ∣ Per ( E ; Ω ) + ∫ E V d x .
Testing the definition of m ( V ) m(V) m ( V ) with ρ E \rho_E ρ E and then taking the infimum over E E E yields m ( V ) ≤ a m(V)\le a m ( V ) ≤ a . To prove the reverse inequality, fix ρ ∈ A \rho\in\mathcal{A} ρ ∈ A and let E t : = { x ∈ Ω : ρ ( x ) > t } E_t:=\{x\in\Omega:\rho(x)>t\} E t := { x ∈ Ω : ρ ( x ) > t } , t > 0 t>0 t > 0 . Since ρ ∈ B V ( Ω ) \rho\in BV(\Omega) ρ ∈ B V ( Ω ) , coarea formula gives Per ( E t ; Ω ) < ∞ \operatorname{Per}(E_t;\Omega)<\infty Per ( E t ; Ω ) < ∞ for almost every t > 0 t>0 t > 0 . The layer-cake representation and the mass constraint give ∫ 0 ∞ ∣ E t ∣ d t = 1 \int_0^\infty|E_t|\,dt=1 ∫ 0 ∞ ∣ E t ∣ d t = 1 . Since ∫ Ω ∣ V ∣ ρ d x ≤ ∥ V ∥ ∞ < ∞ \int_\Omega|V|\rho\,dx\le\|V\|_\infty<\infty ∫ Ω ∣ V ∣ ρ d x ≤ ∥ V ∥ ∞ < ∞ , Fubini’s theorem and the layer-cake representation yield
∫ Ω V ρ d x = ∫ 0 ∞ ∫ E t V d x d t . \int_\Omega V\rho\,dx=\int_0^\infty\int_{E_t}V\,dx\,dt. ∫ Ω V ρ d x = ∫ 0 ∞ ∫ E t V d x d t .
For almost every t t t with ∣ E t ∣ > 0 |E_t|>0 ∣ E t ∣ > 0 , the definition of a a a gives Per ( E t ; Ω ) + ∫ E t V d x ≥ a ∣ E t ∣ \operatorname{Per}(E_t;\Omega)+\int_{E_t}V\,dx\ge a|E_t| Per ( E t ; Ω ) + ∫ E t V d x ≥ a ∣ E t ∣ . The same inequality holds when ∣ E t ∣ = 0 |E_t|=0 ∣ E t ∣ = 0 , since then 1 E t = 0 \mathbf{1}_{E_t}=0 1 E t = 0 almost everywhere and its relative perimeter is zero. Integrating in t t t , we obtain
∣ D ρ ∣ ( Ω ) + ∫ Ω V ρ d x = ∫ 0 ∞ ( Per ( E t ; Ω ) + ∫ E t V d x ) d t ≥ a ∫ 0 ∞ ∣ E t ∣ d t = a . \begin{aligned}
|D\rho|(\Omega)+\int_\Omega V\rho\,dx
&=\int_0^\infty\left(\operatorname{Per}(E_t;\Omega)+\int_{E_t}V\,dx\right)\,dt\\
&\ge a\int_0^\infty|E_t|\,dt=a.
\end{aligned} ∣ D ρ ∣ ( Ω ) + ∫ Ω V ρ d x = ∫ 0 ∞ ( Per ( E t ; Ω ) + ∫ E t V d x ) d t ≥ a ∫ 0 ∞ ∣ E t ∣ d t = a .
Taking the infimum over ρ ∈ A \rho\in\mathcal{A} ρ ∈ A proves m ( V ) ≥ a m(V)\ge a m ( V ) ≥ a , and hence the claimed formula. Concavity follows because (3.2) is an infimum of affine functions of V V V . Finally, nonnegativity and unit mass imply
∣ ∫ Ω ( V − W ) ρ d x ∣ ≤ ∥ V − W ∥ ∞ for every ρ ∈ A . \left|\int_\Omega(V-W)\rho\,dx\right|\le\|V-W\|_\infty\qquad\text{for every }\rho\in\mathcal{A}. ∫ Ω ( V − W ) ρ d x ≤ ∥ V − W ∥ ∞ for every ρ ∈ A .
It follows by taking infima that m ( V ) ≤ m ( W ) + ∥ V − W ∥ ∞ m(V)\le m(W)+\|V-W\|_\infty m ( V ) ≤ m ( W ) + ∥ V − W ∥ ∞ . Interchanging V V V and W W W yields ∣ m ( V ) − m ( W ) ∣ ≤ ∥ V − W ∥ ∞ |m(V)-m(W)|\le\|V-W\|_\infty ∣ m ( V ) − m ( W ) ∣ ≤ ∥ V − W ∥ ∞ . □
Fix x ∗ ∈ K x_*\in K x ∗ ∈ K and introduce the compact convex set
K = { f ∈ C ( K ) : f ( x ∗ ) = 0 , Lip ( f ) ≤ 2 D } . (14) \mathcal{K}=\{f\in C(K):f(x_*)=0,\ \operatorname{Lip}(f)\le2D\}. \tag*{(14)} K = { f ∈ C ( K ) : f ( x ∗ ) = 0 , Lip ( f ) ≤ 2 D } . (14)
Compactness is with respect to the uniform norm. Notice that ∥ f ∥ ∞ ≤ 2 D 2 \|f\|_\infty\le2D^2 ∥ f ∥ ∞ ≤ 2 D 2 for f ∈ K f\in\mathcal{K} f ∈ K . For f = ( f 1 , … , f n ) ∈ C ( K ) n f=(f_1,\ldots,f_n)\in C(K)^n f = ( f 1 , … , f n ) ∈ C ( K ) n , write
V f : = ∑ i = 1 n f i , B ( f ) : = ∑ i = 1 n ∫ K f i c d μ i . V_f:=\sum_{i=1}^n f_i,\qquad B(f):=\sum_{i=1}^n\int_K f_i^c\,d\mu_i. V f := i = 1 ∑ n f i , B ( f ) := i = 1 ∑ n ∫ K f i c d μ i .
Theorem 3.2. For arbitrary μ 1 , … , μ n ∈ P ( K ) \mu_1,\ldots,\mu_n\in\mathcal{P}(K) μ 1 , … , μ n ∈ P ( K ) ,
min ρ ∈ A J ( ρ ) = max f ∈ K n { B ( f ) + m ( V f ) } . (15) \min_{\rho\in\mathcal{A}}J(\rho)=\max_{f\in\mathcal{K}^n}\{B(f)+m(V_f)\}. \tag*{(15)} ρ ∈ A min J ( ρ ) = f ∈ K n max { B ( f ) + m ( V f )} . (15) Equivalently, the minimum equals
max f 1 , … , f n ∈ C ( K ) , λ ∈ R λ ∣ E ∣ ≤ Per ( E ; Ω ) + ∫ E ∑ i f i d x for all finite-perimeter E ⊂ Ω { λ + ∑ i = 1 n ∫ K f i c d μ i } . (16) \max_{\substack{f_1,\ldots,f_n\in C(K),\ \lambda\in\mathbb{R}\\
\lambda|E|\le\operatorname{Per}(E;\Omega)+\int_E\sum_i f_i\,dx\ \text{for all finite-perimeter }E\subset\Omega}}
\left\{\lambda+\sum_{i=1}^n\int_K f_i^c\,d\mu_i\right\}. \tag*{(16)} f 1 , … , f n ∈ C ( K ) , λ ∈ R λ ∣ E ∣ ≤ Per ( E ; Ω ) + ∫ E ∑ i f i d x for all finite-perimeter E ⊂ Ω max { λ + i = 1 ∑ n ∫ K f i c d μ i } . (16) The maximum in (3.6) can be attained with f i ∈ K f_i\in\mathcal{K} f i ∈ K and λ = m ( V f ) \lambda=m(V_f) λ = m ( V f ) .
Proof. For ρ ∈ A \rho\in\mathcal{A} ρ ∈ A and f = ( f 1 , … , f n ) ∈ K n \mathbf{f}=(f_1,\ldots,f_n)\in\mathcal{K}^n f = ( f 1 , … , f n ) ∈ K n , define
L ( ρ , f ) : = ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x + B ( f ) . L(\rho,\mathbf{f}):=|D\rho|(\Omega)+\int_{\Omega}V_{\mathbf{f}}\rho\,dx+B(\mathbf{f}). L ( ρ , f ) := ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x + B ( f ) .
We first justify restricting the Kantorovich dual problem to K \mathcal{K} K . For ρ ∈ A \rho\in\mathcal{A} ρ ∈ A , we regard ν : = ρ d x \nu:=\rho\,dx ν := ρ d x as a probability measure on K K K by setting ν ( A ) : = ∫ A ∩ Ω ρ ( x ) d x \nu(A):=\int_{A\cap\Omega}\rho(x)\,dx ν ( A ) := ∫ A ∩ Ω ρ ( x ) d x for every Borel set A ⊂ K A\subset K A ⊂ K . In particular, for f ∈ C ( K ) f\in C(K) f ∈ C ( K ) , ν ( ∂ Ω ) = 0 \nu(\partial\Omega)=0 ν ( ∂ Ω ) = 0 and ∫ K f d ν = ∫ Ω f ρ d x \int_K f\,d\nu=\int_\Omega f\rho\,dx ∫ K f d ν = ∫ Ω f ρ d x . For each i i i , Kantorovich duality gives
W 2 2 ( μ i , ν ) = sup f ∈ C ( K ) { ∫ K f c d μ i + ∫ K f d ν } . W_2^2(\mu_i,\nu)=\sup_{f\in C(K)}\left\{\int_K f^c\,d\mu_i+\int_K f\,d\nu\right\}. W 2 2 ( μ i , ν ) = f ∈ C ( K ) sup { ∫ K f c d μ i + ∫ K f d ν } .
Given f ∈ C ( K ) f\in C(K) f ∈ C ( K ) , its double c c c -transform satisfies f c c ≥ f f^{cc}\ge f f cc ≥ f and ( f c c ) c = f c (f^{cc})^c=f^c ( f cc ) c = f c . Consequently, replacing f f f by f c c f^{cc} f cc does not decrease the dual objective
∫ K ( f c c ) c d μ i + ∫ K f c c d ν ≥ ∫ K f c d μ i + ∫ K f d ν . \int_K (f^{cc})^c\,d\mu_i+\int_K f^{cc}\,d\nu\ge\int_K f^c\,d\mu_i+\int_K f\,d\nu. ∫ K ( f cc ) c d μ i + ∫ K f cc d ν ≥ ∫ K f c d μ i + ∫ K f d ν .
Since
f c c ( y ) = inf x ∈ K { ∣ x − y ∣ 2 − f c ( x ) } f^{cc}(y)=\inf_{x\in K}\{|x-y|^2-f^c(x)\} f cc ( y ) = x ∈ K inf { ∣ x − y ∣ 2 − f c ( x )}
and
∣ ∣ x − y ∣ 2 − ∣ x − y ′ ∣ 2 ∣ ≤ 2 D ∣ y − y ′ ∣ , x , y , y ′ ∈ K , \bigl||x-y|^2-|x-y'|^2\bigr|\le2D|y-y'|,\qquad x,y,y'\in K, ∣ x − y ∣ 2 − ∣ x − y ′ ∣ 2 ≤ 2 D ∣ y − y ′ ∣ , x , y , y ′ ∈ K ,
the function f c c f^{cc} f cc is 2 D 2D 2 D -Lipschitz. The Kantorovich objective is invariant under the normalization f ~ : = f − f ( x ∗ ) \widetilde{f}:=f-f(x_*) f := f − f ( x ∗ ) , since f ~ c = f c + f ( x ∗ ) \widetilde{f}^c=f^c+f(x_*) f c = f c + f ( x ∗ ) and both marginal measures have unit mass. Thus optimal potentials can be chosen in K \mathcal{K} K , and Kantorovich duality yields
J ( ρ ) = max f ∈ K n L ( ρ , f ) for every ρ ∈ A . J(\rho)=\max_{\mathbf{f}\in\mathcal{K}^n}L(\rho,\mathbf{f})\qquad\text{for every }\rho\in\mathcal{A}. J ( ρ ) = f ∈ K n max L ( ρ , f ) for every ρ ∈ A .
We next verify the hypotheses of Lemma 2.1 . Equip A \mathcal{A} A with the topology inherited from L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) and K n \mathcal{K}^n K n with the product uniform topology. Both sets are convex, and K n \mathcal{K}^n K n is compact. For fixed f \mathbf{f} f , the map ρ ↦ L ( ρ , f ) \rho\mapsto L(\rho,\mathbf{f}) ρ ↦ L ( ρ , f ) is convex and lower semicontinuous, while the linear term is continuous because V f V_{\mathbf{f}} V f is bounded. For fixed ρ \rho ρ , the map f ↦ L ( ρ , f ) \mathbf{f}\mapsto L(\rho,\mathbf{f}) f ↦ L ( ρ , f ) is concave. Indeed, for f , g ∈ C ( K ) f,g\in C(K) f , g ∈ C ( K ) and s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] ,
( ( 1 − s ) f + s g ) c ( x ) = min y ∈ K { ( 1 − s ) ( ∣ x − y ∣ 2 − f ( y ) ) + s ( ∣ x − y ∣ 2 − g ( y ) ) } ≥ ( 1 − s ) f c ( x ) + s g c ( x ) . \bigl((1-s)f+sg\bigr)^c(x)=\min_{y\in K}\{(1-s)(|x-y|^2-f(y))+s(|x-y|^2-g(y))\}
\ge(1-s)f^c(x)+sg^c(x). ( ( 1 − s ) f + s g ) c ( x ) = y ∈ K min {( 1 − s ) ( ∣ x − y ∣ 2 − f ( y )) + s ( ∣ x − y ∣ 2 − g ( y ))} ≥ ( 1 − s ) f c ( x ) + s g c ( x ) .
It is also continuous, since ∥ f c − g c ∥ ∞ ≤ ∥ f − g ∥ ∞ \|f^c-g^c\|_\infty\le\|f-g\|_\infty ∥ f c − g c ∥ ∞ ≤ ∥ f − g ∥ ∞ and ∫ Ω ρ d x = 1 \int_\Omega\rho\,dx=1 ∫ Ω ρ d x = 1 . Therefore, we have
inf ρ ∈ A J ( ρ ) = inf ρ ∈ A sup f ∈ K n L ( ρ , f ) = sup f ∈ K n inf ρ ∈ A L ( ρ , f ) = sup f ∈ K n { B ( f ) + m ( V f ) } . \inf_{\rho\in\mathcal{A}}J(\rho)=\inf_{\rho\in\mathcal{A}}\sup_{\mathbf{f}\in\mathcal{K}^n}L(\rho,\mathbf{f})=\sup_{\mathbf{f}\in\mathcal{K}^n}\inf_{\rho\in\mathcal{A}}L(\rho,\mathbf{f})=\sup_{\mathbf{f}\in\mathcal{K}^n}\{B(\mathbf{f})+m(V_{\mathbf{f}})\}. ρ ∈ A inf J ( ρ ) = ρ ∈ A inf f ∈ K n sup L ( ρ , f ) = f ∈ K n sup ρ ∈ A inf L ( ρ , f ) = f ∈ K n sup { B ( f ) + m ( V f )} .
The last supremum is attained. Indeed, the stability of the c c c -transform and the Lipschitz continuity of m m m established in Lemma 3.1 imply
∣ B ( f ) − B ( g ) ∣ ≤ ∑ i = 1 n ∥ f i − g i ∥ ∞ , |B(\mathbf{f})-B(\mathbf{g})|\le\sum_{i=1}^{n}\|f_i-g_i\|_\infty, ∣ B ( f ) − B ( g ) ∣ ≤ i = 1 ∑ n ∥ f i − g i ∥ ∞ ,
∣ m ( V f ) − m ( V g ) ∣ ≤ ∥ V f − V g ∥ ∞ ≤ ∑ i = 1 n ∥ f i − g i ∥ ∞ . |m(V_{\mathbf{f}})-m(V_{\mathbf{g}})|\le\|V_{\mathbf{f}}-V_{\mathbf{g}}\|_\infty\le\sum_{i=1}^{n}\|f_i-g_i\|_\infty. ∣ m ( V f ) − m ( V g ) ∣ ≤ ∥ V f − V g ∥ ∞ ≤ i = 1 ∑ n ∥ f i − g i ∥ ∞ .
Thus the objective is continuous on the compact set K n \mathcal{K}^n K n . Together with primal attainment, this proves (15) . It remains to prove (16) . By Lemma 3.1, the constraints λ ∣ E ∣ ≤ Per ( E ; Ω ) + ∫ E V f d x \lambda|E|\le\operatorname{Per}(E;\Omega)+\int_E V_{\mathbf{f}}\,dx λ ∣ E ∣ ≤ Per ( E ; Ω ) + ∫ E V f d x for all measurable sets E ⊂ Ω E \subset\Omega E ⊂ Ω of positive measure and finite relative perimeter are equivalent to λ ≤ m ( V f ) \lambda\le m(V_f) λ ≤ m ( V f ) . Hence, for any admissible pair ( f , λ ) ∈ C ( K ) n × R (f,\lambda) \in C(K)^n \times\mathbb{R} ( f , λ ) ∈ C ( K ) n × R and any ρ ∈ A \rho\in\mathcal{A} ρ ∈ A ,
λ + B ( f ) ≤ m ( V f ) + B ( f ) ≤ ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x + B ( f ) ≤ J ( ρ ) . \lambda+ B(f) \le m(V_f) + B(f) \le|D\rho|(\Omega) + \int_{\Omega} V_f\rho\,dx + B(f) \le J(\rho). λ + B ( f ) ≤ m ( V f ) + B ( f ) ≤ ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x + B ( f ) ≤ J ( ρ ) .
where the last inequality follows from Kantorovich duality. Finally, let f ∗ ∈ K n f^* \in\mathcal{K}^n f ∗ ∈ K n attain the maximum in (3.5) and set λ ∗ : = m ( V f ∗ ) \lambda^* := m(V_{f^*}) λ ∗ := m ( V f ∗ ) . The pair ( f ∗ , λ ∗ ) (f^*,\lambda^*) ( f ∗ , λ ∗ ) satisfies all the perimeter constraints and λ ∗ + B ( f ∗ ) = min ρ ∈ A J ( ρ ) \lambda^* + B(f^*) = \min_{\rho\in\mathcal{A}} J(\rho) λ ∗ + B ( f ∗ ) = min ρ ∈ A J ( ρ ) . This proves the stated dual equality (3.6). □
Remark 3.2. The dual constraints and objective are invariant under
f i ↦ f i + a i , λ ↦ λ + ∑ i = 1 n a i , a i ∈ R . f_i \mapsto f_i+a_i,\qquad\lambda\mapsto\lambda+\sum_{i=1}^{n}a_i,\qquad a_i\in\mathbb{R}. f i ↦ f i + a i , λ ↦ λ + i = 1 ∑ n a i , a i ∈ R .
Thus the normalization f i ( x ∗ ) = 0 f_i(x_*)=0 f i ( x ∗ ) = 0 only fixes the additive constants and imposes no boundary condition. Moreover, the relative perimeter Per ( E ; Ω ) \operatorname{Per}(E;\Omega) Per ( E ; Ω ) counts only the perimeter inside Ω \Omega Ω , consistently with the total variation ∣ D ρ ∣ ( Ω ) |D\rho|(\Omega) ∣ D ρ ∣ ( Ω ) in the primal problem.
Optimality conditions and transport maps
The attained dual formulation provides a direct characterization of the minimizers. For a fixed dual maximizer, optimality is equivalent to equality in each transport inequality and in the total variation bound. We first state these conditions and their consequences for superlevel sets, and then derive a transport-map representation and a criterion for the uniform density to be optimal.
Theorem 4.1 (Transport and TV optimality). Let ( f , λ ) (f,\lambda) ( f , λ ) be a maximizer of (3.6). A density ρ ∈ A \rho\in\mathcal{A} ρ ∈ A minimizes J J J if and only if
W 2 2 ( μ i , ρ d x ) = ∫ K f i c d μ i + ∫ Ω f i ρ d x , i = 1 , … , n . (17) W_2^2(\mu_i,\rho\,dx)=\int_K f_i^c\,d\mu_i+\int_\Omega f_i\rho\,dx,\qquad i=1,\ldots,n. \tag*{(17)} W 2 2 ( μ i , ρ d x ) = ∫ K f i c d μ i + ∫ Ω f i ρ d x , i = 1 , … , n . (17) ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x = λ . (18) |D\rho|(\Omega)+\int_\Omega V_f\rho\,dx=\lambda. \tag*{(18)} ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x = λ . (18) For such a density, almost every t > 0 t>0 t > 0 with ∣ E t ∣ > 0 |E_t|>0 ∣ E t ∣ > 0 , where E t = { ρ > t } E_t=\{\rho>t\} E t = { ρ > t } , satisfies
Per ( E t ; Ω ) + ∫ E t V f d x = λ ∣ E t ∣ . (19) \operatorname{Per}(E_t;\Omega)+\int_{E_t}V_f\,dx=\lambda|E_t|. \tag*{(19)} Per ( E t ; Ω ) + ∫ E t V f d x = λ ∣ E t ∣. (19) Thus these superlevel sets minimize the ratio in (3.3) for the same potential V f V_f V f .
Proof. Fix a dual maximizer ( f , λ ) (f,\lambda) ( f , λ ) . By the strong duality, λ + B ( f ) = min ρ ∈ A J ( ρ ) \lambda+B(f)=\min_{\rho\in\mathcal{A}}J(\rho) λ + B ( f ) = min ρ ∈ A J ( ρ ) . For any ρ ∈ A \rho\in\mathcal{A} ρ ∈ A , Kantorovich duality gives
W 2 2 ( μ i , ρ d x ) − ∫ K f i c d μ i − ∫ Ω f i ρ d x ≥ 0 , i = 1 , … , n . W_2^2(\mu_i,\rho\,dx)-\int_K f_i^c\,d\mu_i-\int_\Omega f_i\rho\,dx\ge0,\qquad i=1,\ldots,n. W 2 2 ( μ i , ρ d x ) − ∫ K f i c d μ i − ∫ Ω f i ρ d x ≥ 0 , i = 1 , … , n .
Since ( f , λ ) (f,\lambda) ( f , λ ) is a dual maximizer, we have λ = m ( V f ) \lambda=m(V_f) λ = m ( V f ) . Consequently, by the definition of m ( V ) m(V) m ( V ) ,
∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x − λ = ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x − m ( V f ) ≥ 0 (20) |D\rho|(\Omega)+\int_\Omega V_f\rho\,dx-\lambda=|D\rho|(\Omega)+\int_\Omega V_f\rho\,dx-m(V_f)\ge0 \tag*{(20)} ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x − λ = ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x − m ( V f ) ≥ 0 (20)
for every ρ ∈ A \rho\in\mathcal{A} ρ ∈ A . Thus the identity
J ( ρ ) − ( λ + B ( f ) ) = ∑ i = 1 n [ W 2 2 ( μ i , ρ d x ) − ∫ K f i c d μ i − ∫ Ω f i ρ d x ] + [ ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x − λ ] J(\rho)-(\lambda+B(f))=\sum_{i=1}^{n}\left[W_{2}^{2}(\mu_i,\rho\,dx)-\int_K f_i^c\,d\mu_i-\int_\Omega f_i\rho\,dx\right]
+\left[|D\rho|(\Omega)+\int_\Omega V_f\rho\,dx-\lambda\right] J ( ρ ) − ( λ + B ( f )) = i = 1 ∑ n [ W 2 2 ( μ i , ρ d x ) − ∫ K f i c d μ i − ∫ Ω f i ρ d x ] + [ ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x − λ ]
expresses the primal–dual gap as a sum of nonnegative terms. Since λ + B ( f ) \lambda+B(f) λ + B ( f ) is the primal minimum, ρ \rho ρ is a minimizer if and only if this sum vanishes. This occurs precisely when every term is zero, giving (4.1) and (4.2). Now let ρ \rho ρ be a minimizer and set E t : = { ρ > t } E_t:=\{\rho>t\} E t := { ρ > t } for t > 0 t>0 t > 0 . By the coarea formula, E t E_t E t has finite relative perimeter for almost every t > 0 t>0 t > 0 . Dual feasibility gives
G ( t ) : = Per ( E t ; Ω ) + ∫ E t V f d x − λ ∣ E t ∣ ≥ 0 G(t):=\operatorname{Per}(E_t;\Omega)+\int_{E_t}V_f\,dx-\lambda|E_t|\ge0 G ( t ) := Per ( E t ; Ω ) + ∫ E t V f d x − λ ∣ E t ∣ ≥ 0
for almost every such t t t ; when ∣ E t ∣ = 0 |E_t|=0 ∣ E t ∣ = 0 , all three terms vanish. Using the coarea and layer-cake formulas, we obtain
∫ 0 ∞ G ( t ) d t = ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x − λ ∫ Ω ρ d x = ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x − λ = 0 , \int_0^\infty G(t)\,dt=|D\rho|(\Omega)+\int_\Omega V_f\rho\,dx-\lambda\int_\Omega\rho\,dx
=|D\rho|(\Omega)+\int_\Omega V_f\rho\,dx-\lambda=0, ∫ 0 ∞ G ( t ) d t = ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x − λ ∫ Ω ρ d x = ∣ D ρ ∣ ( Ω ) + ∫ Ω V f ρ d x − λ = 0 ,
where the last equality is (4.2). The term involving V V V is integrable because ∫ Ω ∣ V f ∣ ρ d x ≤ ∥ V f ∥ ∞ \int_\Omega|V_f|\rho\,dx\le\|V_f\|_\infty ∫ Ω ∣ V f ∣ ρ d x ≤ ∥ V f ∥ ∞ . Since G ≥ 0 G\ge0 G ≥ 0 almost everywhere and its integral is zero, G ( t ) = 0 G(t)=0 G ( t ) = 0 for almost every t > 0 t>0 t > 0 , which proves (4.3). Since λ = m ( V f ) \lambda=m(V_f) λ = m ( V f ) , the level-set identity (4.3) gives
Per ( E t ; Ω ) + ∫ E t V f d x ∣ E t ∣ = m ( V f ) \frac{\operatorname{Per}(E_t;\Omega)+\int_{E_t}V_f\,dx}{|E_t|}=m(V_f) ∣ E t ∣ Per ( E t ; Ω ) + ∫ E t V f d x = m ( V f )
for almost every t > 0 t>0 t > 0 with ∣ E t ∣ > 0 |E_t|>0 ∣ E t ∣ > 0 . By Lemma 3.1, m ( V ) m(V) m ( V ) is the infimum of this ratio over all measurable subsets of Ω \Omega Ω with positive measure and finite relative perimeter. Hence each such superlevel set E t E_t E t attains this infimum. □
When at least one of the prescribed measures is absolutely continuous, the transport optimality conditions yield the following representation of the unique minimizing measure.
Corollary 4.1 (Map representation). Suppose that μ j \mu_j μ j is absolutely continuous for some j ∈ { 1 , … , n } j\in\{1,\ldots,n\} j ∈ { 1 , … , n } , and fix any dual maximizer ( f , λ ) (f,\lambda) ( f , λ ) . The unique minimizer satisfies
ρ d x = ( T j ) # μ j , T j ( x ) = x − 1 2 ∇ f j c ( x ) for μ j -a.e. x . (21) \rho\,dx=(T_j)_\#\mu_j,\qquad T_j(x)=x-\frac{1}{2}\nabla f_j^c(x)\ \text{for }\mu_j\text{-a.e. }x. \tag*{(21)} ρ d x = ( T j ) # μ j , T j ( x ) = x − 2 1 ∇ f j c ( x ) for μ j -a.e. x . (21) Equivalently, setting Φ j ( x ) = 1 2 ( ∣ x ∣ 2 − f j c ( x ) ) \Phi_j(x)=\frac{1}{2}(|x|^2-f_j^c(x)) Φ j ( x ) = 2 1 ( ∣ x ∣ 2 − f j c ( x )) , one has ρ d x = ( ∇ Φ j ) # μ j \rho\,dx=(\nabla\Phi_j)_\#\mu_j ρ d x = ( ∇ Φ j ) # μ j , and Φ j \Phi_j Φ j is convex.
Proof. (17) makes ( f j c , f j ) (f_j^c,f_j) ( f j c , f j ) an optimal Kantorovich pair. The transport theorem and (2.9) yield (4.4). For the convexity assertion, extend the transform to R d \mathbb{R}^d R d by the same infimum over y ∈ K y\in K y ∈ K and observe that
Φ j ( x ) = sup y ∈ K { x ⋅ y − ∣ y ∣ 2 2 + f j ( y ) 2 } . \Phi_j(x)=\sup_{y\in K}\left\{x\cdot y-\frac{|y|^2}{2}+\frac{f_j(y)}{2}\right\}. Φ j ( x ) = y ∈ K sup { x ⋅ y − 2 ∣ y ∣ 2 + 2 f j ( y ) } . This is a supremum of affine functions. □
For the uniform density, the total variation vanishes, and the optimality conditions reduce to the following criterion.
Corollary 4.2 (A criterion for the uniform density). The density ρ 0 = ∣ Ω ∣ − 1 \rho_0 = |\Omega|^{-1} ρ 0 = ∣Ω ∣ − 1 minimizes J J J if and only if there exist potentials f 1 , … , f n ∈ C ( K ) f_1,\ldots,f_n \in C(K) f 1 , … , f n ∈ C ( K ) such that ( f i c , f i ) (f_i^c,f_i) ( f i c , f i ) is optimal for ( μ i , ρ 0 d x ) (\mu_i,\rho_0\,dx) ( μ i , ρ 0 d x ) for each i i i , and, writing V = ∑ i f i V=\sum_i f_i V = ∑ i f i and V ‾ = ∣ Ω ∣ − 1 ∫ Ω V d x \overline{V}=|\Omega|^{-1}\int_\Omega V\,dx V = ∣Ω ∣ − 1 ∫ Ω V d x ,
Per ( E ; Ω ) + ∫ E ( V − V ‾ ) d x ≥ 0 for every finite-perimeter E ⊂ Ω . (22) \operatorname{Per}(E;\Omega)+\int_E(V-\overline{V})\,dx\ge0\qquad\text{for every finite-perimeter }E\subset\Omega. \tag*{(22)} Per ( E ; Ω ) + ∫ E ( V − V ) d x ≥ 0 for every finite-perimeter E ⊂ Ω. (22) Proof. If ρ 0 \rho_0 ρ 0 minimizes J J J , choose a dual maximizer and apply Theorem 4.1. Since TV ( ρ 0 ) = 0 \operatorname{TV}(\rho_0)=0 TV ( ρ 0 ) = 0 , it gives λ = V ‾ \lambda=\overline{V} λ = V , and the dual constraints give (22) . Conversely, the stated conditions make ( f , V ‾ ) (f,\overline{V}) ( f , V ) dual feasible and make its value equal to J ( ρ 0 ) J(\rho_0) J ( ρ 0 ) . Weak duality then proves optimality. □
Vector-field duality via entropic approximation
Following Carlier and Poon [9 Section 3] , we use an entropic approximation to establish an attained vector-field dual. The positivity of the regularized minimizers allows us to differentiate the transport costs with respect to the density, even when the prescribed measures are singular. The resulting optimality conditions involve the sum of the Kantorovich potentials and an explicit multiplier for the mass constraint. We pass to the limit using uniform estimates and then relate the resulting formulation to the perimeter dual.
Vector fields and total variation
We introduce the vector fields used to represent the total variation and recall the associated integration-by-parts formula. The lemma below will allow us to recover such a field from a variational inequality. The embedding B V ( Ω ) ↪ L r ( Ω ) BV(\Omega)\hookrightarrow L^r(\Omega) B V ( Ω ) ↪ L r ( Ω ) is compact, including when d = 1 d=1 d = 1 . For z ∈ L ∞ ( Ω ; R d ) z\in L^\infty(\Omega;\mathbb{R}^d) z ∈ L ∞ ( Ω ; R d ) with div z = ξ ∈ L q ( Ω ) \operatorname{div}z=\xi\in L^q(\Omega) div z = ξ ∈ L q ( Ω ) , we say that its normal trace is zero if
∫ Ω ξ u d x = − ∫ Ω z ⋅ ∇ u d x for every u ∈ C 1 ( Ω ‾ ) . (23) \int_\Omega\xi u\,dx=-\int_\Omega z\cdot\nabla u\,dx\qquad\text{for every }u\in C^1(\overline{\Omega}). \tag*{(23)} ∫ Ω ξ u d x = − ∫ Ω z ⋅ ∇ u d x for every u ∈ C 1 ( Ω ) . (23)
This identity includes both the distributional divergence relation and the zero boundary trace. Write
Γ q = { ξ ∈ L q ( Ω ) : ξ = div z for some z ∈ L ∞ ( Ω ; R d ) , ∥ z ∥ ∞ ≤ 1 and (5.1) holds } . \Gamma_q=\left\{\xi\in L^q(\Omega):\ \xi=\operatorname{div}z\text{ for some }z\in L^\infty(\Omega;\mathbb{R}^d),\ \|z\|_\infty\le1\text{ and (5.1) holds}\right\}. Γ q = { ξ ∈ L q ( Ω ) : ξ = div z for some z ∈ L ∞ ( Ω ; R d ) , ∥ z ∥ ∞ ≤ 1 and (5.1) holds } .
Lemma 5.1. The set Γ q \Gamma_q Γ q is convex and norm-closed in L q ( Ω ) L^q(\Omega) L q ( Ω ) . With total variation extended by + ∞ +\infty + ∞ outside B V ( Ω ) BV(\Omega) B V ( Ω ) ,
∣ D u ∣ ( Ω ) = sup ξ ∈ Γ q ∫ Ω ξ u d x for u ∈ L r ( Ω ) . (24) |Du|(\Omega)=\sup_{\xi\in\Gamma_q}\int_\Omega\xi u\,dx\qquad\text{for }u\in L^r(\Omega). \tag*{(24)} ∣ D u ∣ ( Ω ) = ξ ∈ Γ q sup ∫ Ω ξ u d x for u ∈ L r ( Ω ) . (24) If ξ ∈ L q ( Ω ) \xi\in L^q(\Omega) ξ ∈ L q ( Ω ) satisfies ∫ Ω ξ u d x ≤ ∣ D u ∣ ( Ω ) \int_\Omega\xi u\,dx\le|Du|(\Omega) ∫ Ω ξ u d x ≤ ∣ D u ∣ ( Ω ) for every u ∈ B V ( Ω ) ∩ L ∞ ( Ω ) u\in BV(\Omega)\cap L^\infty(\Omega) u ∈ B V ( Ω ) ∩ L ∞ ( Ω ) , then ξ ∈ Γ q \xi\in\Gamma_q ξ ∈ Γ q . Proof. Convexity is immediate. If ξ k → ξ \xi_k\to\xi ξ k → ξ in L q L^q L q and z k z_k z k are corresponding fields, weak-* compactness gives z k ⇀ ∗ z z_k\rightharpoonup^* z z k ⇀ ∗ z in L ∞ L^\infty L ∞ along a subsequence, with ∥ z ∥ ∞ ≤ 1 \|z\|_\infty\le1 ∥ z ∥ ∞ ≤ 1 . Passing to the limit in (23) proves closedness. The upper bound in (24) follows from (6) ; the opposite bound follows from (2) , since compactly supported smooth fields are allowed. If u ∈ L r ( Ω ) ∖ B V ( Ω ) u\in L^r(\Omega)\setminus BV(\Omega) u ∈ L r ( Ω ) ∖ B V ( Ω ) , then u ∈ L 1 ( Ω ) u\in L^1(\Omega) u ∈ L 1 ( Ω ) because Ω \Omega Ω is bounded. By the definition of B V ( Ω ) BV(\Omega) B V ( Ω ) ,
sup z ∈ C c 1 ( Ω ; R d ) ∥ z ∥ ∞ ≤ 1 ∫ Ω u div z d x = + ∞ . \sup_{\substack{z\in C_c^1(\Omega;\mathbb{R}^d)\\ \|z\|_\infty\le1}}\int_\Omega u\,\operatorname{div}z\,dx=+\infty. z ∈ C c 1 ( Ω ; R d ) ∥ z ∥ ∞ ≤ 1 sup ∫ Ω u div z d x = + ∞.
Since div z ∈ Γ q \operatorname{div} z \in\Gamma_q div z ∈ Γ q for every such vector field, the supremum in (24) is also infinite. This proves (24) for all u ∈ L r ( Ω ) u \in L^r(\Omega) u ∈ L r ( Ω ) . Finally, suppose that ξ ∈ L q ( Ω ) \xi\in L^q(\Omega) ξ ∈ L q ( Ω ) satisfies ∫ Ω ξ u d x ≤ ∣ D u ∣ ( Ω ) \int_\Omega\xi u\,dx \le|Du|(\Omega) ∫ Ω ξ u d x ≤ ∣ D u ∣ ( Ω ) for every u ∈ B V ( Ω ) ∩ L ∞ ( Ω ) u \in BV(\Omega) \cap L^\infty(\Omega) u ∈ B V ( Ω ) ∩ L ∞ ( Ω ) . Let u ∈ B V ( Ω ) ∩ L r ( Ω ) u \in BV(\Omega) \cap L^r(\Omega) u ∈ B V ( Ω ) ∩ L r ( Ω ) , where 1 ≤ r < ∞ 1 \le r < \infty 1 ≤ r < ∞ . For M > 0 M > 0 M > 0 , define T M ( s ) : = max { − M , min { s , M } } T_M(s) := \max\{-M,\min\{s,M\}\} T M ( s ) := max { − M , min { s , M }} and u M : = T M ( u ) u_M := T_M(u) u M := T M ( u ) . Then ∣ u M ∣ ≤ M |u_M| \le M ∣ u M ∣ ≤ M almost everywhere, so u M ∈ L ∞ ( Ω ) u_M \in L^\infty(\Omega) u M ∈ L ∞ ( Ω ) . We next verify that u M ∈ B V ( Ω ) u_M \in BV(\Omega) u M ∈ B V ( Ω ) and that truncation does not increase the total variation. By smooth approximation in B V ( Ω ) BV(\Omega) B V ( Ω ) , there exists a sequence v k ∈ C ∞ ( Ω ) ∩ W 1 , 1 ( Ω ) v_k \in C^\infty(\Omega) \cap W^{1,1}(\Omega) v k ∈ C ∞ ( Ω ) ∩ W 1 , 1 ( Ω ) such that v k → u v_k \to u v k → u in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) and ∫ Ω ∣ ∇ v k ∣ d x → ∣ D u ∣ ( Ω ) \int_\Omega|\nabla v_k|\,dx \to|Du|(\Omega) ∫ Ω ∣∇ v k ∣ d x → ∣ D u ∣ ( Ω ) . Since T M T_M T M is 1 1 1 -Lipschitz,
∥ T M ( v k ) − u M ∥ L 1 ( Ω ) ≤ ∥ v k − u ∥ L 1 ( Ω ) → 0. \|T_M(v_k)-u_M\|_{L^1(\Omega)} \le\|v_k-u\|_{L^1(\Omega)} \to0. ∥ T M ( v k ) − u M ∥ L 1 ( Ω ) ≤ ∥ v k − u ∥ L 1 ( Ω ) → 0.
The Sobolev chain rule also gives T M ( v k ) ∈ W 1 , 1 ( Ω ) T_M(v_k) \in W^{1,1}(\Omega) T M ( v k ) ∈ W 1 , 1 ( Ω ) and ∣ ∇ T M ( v k ) ∣ ≤ ∣ ∇ v k ∣ |\nabla T_M(v_k)| \le|\nabla v_k| ∣∇ T M ( v k ) ∣ ≤ ∣∇ v k ∣ almost everywhere in Ω \Omega Ω . Therefore, by the lower semicontinuity of the total variation under L 1 L^1 L 1 convergence,
∣ D u M ∣ ( Ω ) ≤ lim inf k → ∞ ∫ Ω ∣ ∇ T M ( v k ) ∣ d x ≤ lim k → ∞ ∫ Ω ∣ ∇ v k ∣ d x = ∣ D u ∣ ( Ω ) < ∞ . |Du_M|(\Omega) \le\liminf_{k\to\infty} \int_\Omega|\nabla T_M(v_k)|\,dx
\le\lim_{k\to\infty} \int_\Omega|\nabla v_k|\,dx = |Du|(\Omega) < \infty. ∣ D u M ∣ ( Ω ) ≤ k → ∞ lim inf ∫ Ω ∣∇ T M ( v k ) ∣ d x ≤ k → ∞ lim ∫ Ω ∣∇ v k ∣ d x = ∣ D u ∣ ( Ω ) < ∞.
Thus u M ∈ B V ( Ω ) ∩ L ∞ ( Ω ) u_M \in BV(\Omega) \cap L^\infty(\Omega) u M ∈ B V ( Ω ) ∩ L ∞ ( Ω ) and ∣ D u M ∣ ( Ω ) ≤ ∣ D u ∣ ( Ω ) |Du_M|(\Omega) \le|Du|(\Omega) ∣ D u M ∣ ( Ω ) ≤ ∣ D u ∣ ( Ω ) . Finally, the definition of T M T_M T M yields
∣ u − u M ∣ = { 0 , ∣ u ∣ ≤ M , ∣ u ∣ − M , ∣ u ∣ > M . |u-u_M| =
\begin{cases}
0, & |u| \le M,\\
|u|-M, & |u|>M.
\end{cases} ∣ u − u M ∣ = { 0 , ∣ u ∣ − M , ∣ u ∣ ≤ M , ∣ u ∣ > M .
Consequently,
∥ u − u M ∥ L r ( Ω ) r = ∫ { ∣ u ∣ > M } ( ∣ u ∣ − M ) r d x ≤ ∫ Ω ∣ u ∣ r 1 { ∣ u ∣ > M } d x . \|u-u_M\|_{L^r(\Omega)}^r = \int_{\{|u|>M\}} (|u|-M)^r\,dx \le\int_\Omega|u|^r 1_{\{|u|>M\}}\,dx. ∥ u − u M ∥ L r ( Ω ) r = ∫ { ∣ u ∣ > M } ( ∣ u ∣ − M ) r d x ≤ ∫ Ω ∣ u ∣ r 1 { ∣ u ∣ > M } d x .
As M → ∞ M \to\infty M → ∞ , the integrand on the right converges to zero almost everywhere and is bounded by ∣ u ∣ r ∈ L 1 ( Ω ) |u|^r \in L^1(\Omega) ∣ u ∣ r ∈ L 1 ( Ω ) . The dominated convergence theorem therefore implies u M → u u_M \to u u M → u strongly in L r ( Ω ) L^r(\Omega) L r ( Ω ) . Thus ∫ Ω ξ u d x ≤ ∣ D u ∣ ( Ω ) \int_\Omega\xi u\,dx \le|Du|(\Omega) ∫ Ω ξ u d x ≤ ∣ D u ∣ ( Ω ) . For u ∈ L r ( Ω ) ∖ B V ( Ω ) u \in L^r(\Omega) \setminus BV(\Omega) u ∈ L r ( Ω ) ∖ B V ( Ω ) , this inequality holds trivially because ∣ D u ∣ ( Ω ) = + ∞ |Du|(\Omega)=+\infty ∣ D u ∣ ( Ω ) = + ∞ . It therefore holds for every u ∈ L r ( Ω ) u \in L^r(\Omega) u ∈ L r ( Ω ) . If ξ ∉ Γ q \xi\notin\Gamma_q ξ ∈ / Γ q , the Hahn–Banach separation theorem, applied to the nonempty closed convex set Γ q ⊂ L r ( Ω ) \Gamma_q \subset L^r(\Omega) Γ q ⊂ L r ( Ω ) , yields some u ∈ L r ( Ω ) = ( L q ( Ω ) ) ∗ u \in L^r(\Omega)=(L^q(\Omega))^* u ∈ L r ( Ω ) = ( L q ( Ω ) ) ∗ such that
∫ Ω ξ u d x > sup η ∈ Γ q ∫ Ω η u d x = ∣ D u ∣ ( Ω ) , \int_\Omega\xi u\,dx > \sup_{\eta\in\Gamma_q}\int_\Omega\eta u\,dx = |Du|(\Omega), ∫ Ω ξ u d x > η ∈ Γ q sup ∫ Ω η u d x = ∣ D u ∣ ( Ω ) ,
contradicting the preceding inequality. Hence ξ ∈ Γ q \xi\in\Gamma_q ξ ∈ Γ q . □
Entropy regularization and estimates
We approximate J J J by adding an entropy term. The resulting minimizers are bounded away from zero for each fixed regularization parameter. The estimates established below also provide the compactness needed to pass to the limit. Write
F ( ρ ) = ∑ i = 1 n W 2 2 ( μ i , ρ d x ) , H ( ρ ) = ∫ Ω ρ log ρ d x , J ε ( ρ ) : = F ( ρ ) + ∣ D ρ ∣ ( Ω ) + ε H ( ρ ) , F(\rho)=\sum_{i=1}^{n} W_2^2(\mu_i,\rho\,dx), \qquad H(\rho)=\int_\Omega\rho\log\rho\,dx, \qquad J_\varepsilon(\rho):=F(\rho)+|D\rho|(\Omega)+\varepsilon H(\rho), F ( ρ ) = i = 1 ∑ n W 2 2 ( μ i , ρ d x ) , H ( ρ ) = ∫ Ω ρ log ρ d x , J ε ( ρ ) := F ( ρ ) + ∣ D ρ ∣ ( Ω ) + ε H ( ρ ) ,
where 0 log 0 = 0 0\log0=0 0 log 0 = 0 and ε > 0 \varepsilon>0 ε > 0 . Since B V ( Ω ) ⊂ L r ( Ω ) BV(\Omega)\subset L^r(\Omega) B V ( Ω ) ⊂ L r ( Ω ) with r > 1 r>1 r > 1 , H H H is finite on A \mathcal{A} A . We shall use the elementary transport estimate
∣ F ( η ) − F ( ρ ) ∣ ≤ 2 n D 2 ∥ η − ρ ∥ L 1 ( Ω ) , ρ , η ∈ A . (25) |F(\eta)-F(\rho)| \le2nD^2\|\eta-\rho\|_{L^1(\Omega)}, \qquad\rho,\eta\in\mathcal{A}. \tag*{(25)} ∣ F ( η ) − F ( ρ ) ∣ ≤ 2 n D 2 ∥ η − ρ ∥ L 1 ( Ω ) , ρ , η ∈ A . (25)
Indeed, for each i i i , choose an optimal potential f i ∈ K f_i \in\mathcal{K} f i ∈ K for ( μ i , η d x ) (\mu_i,\eta\,dx) ( μ i , η d x ) . Using the same potential as a dual competitor for ( μ i , ρ d x ) (\mu_i,\rho\,dx) ( μ i , ρ d x ) and subtracting gives
W 2 2 ( μ i , η d x ) − W 2 2 ( μ i , ρ d x ) ≤ ∫ Ω f i ( η − ρ ) d x ≤ 2 D 2 ∥ η − ρ ∥ L 1 ( Ω ) , W_2^2(\mu_i,\eta\,dx)-W_2^2(\mu_i,\rho\,dx)\leq\int_{\Omega}f_i(\eta-\rho)\,dx\leq2D^2\|\eta-\rho\|_{L^1(\Omega)}, W 2 2 ( μ i , η d x ) − W 2 2 ( μ i , ρ d x ) ≤ ∫ Ω f i ( η − ρ ) d x ≤ 2 D 2 ∥ η − ρ ∥ L 1 ( Ω ) ,
where we used ∣ f i ( x ) ∣ = ∣ f i ( x ) − f i ( x ∗ ) ∣ ≤ 2 D ∣ x − x ∗ ∣ ≤ 2 D 2 |f_i(x)|=|f_i(x)-f_i(x_*)|\leq2D|x-x_*|\leq2D^2 ∣ f i ( x ) ∣ = ∣ f i ( x ) − f i ( x ∗ ) ∣ ≤ 2 D ∣ x − x ∗ ∣ ≤ 2 D 2 . Summing over i i i yields the corresponding upper bound for F ( η ) − F ( ρ ) F(\eta)-F(\rho) F ( η ) − F ( ρ ) . Interchanging ρ \rho ρ and η \eta η gives the reverse bound and proves the stated estimate.
Lemma 5.2. For every ε > 0 \varepsilon>0 ε > 0 , J ε J_\varepsilon J ε has a unique minimizer ρ ε ∈ A \rho_\varepsilon\in\mathcal{A} ρ ε ∈ A , ∣ D ρ ε ∣ ( Ω ) ≤ n D 2 |D\rho_\varepsilon|(\Omega)\leq nD^2 ∣ D ρ ε ∣ ( Ω ) ≤ n D 2 , and ess inf Ω ρ ε > 0 \operatorname*{ess\,inf}_{\Omega}\rho_\varepsilon>0 ess inf Ω ρ ε > 0 . Along a sequence ε → 0 \varepsilon\to0 ε → 0 there is ρ ∗ ∈ A \rho_*\in\mathcal{A} ρ ∗ ∈ A such that ρ ε → ρ ∗ \rho_\varepsilon\to\rho_* ρ ε → ρ ∗ strongly in L r L^r L r , ρ ∗ \rho_* ρ ∗ minimizes J J J , and
b ε − : = − ε log min { 1 , ρ ε } is bounded in L ∞ , b ε + : = ε log max { 1 , ρ ε } → 0 in every finite L s . (26) \begin{aligned}
b_\varepsilon^-&:=-\varepsilon\log\min\{1,\rho_\varepsilon\} &&\text{is bounded in }L^\infty,\\
b_\varepsilon^+&:=\varepsilon\log\max\{1,\rho_\varepsilon\} &&\to0\text{ in every finite }L^s. \tag*{(26)}
\end{aligned} b ε − b ε + := − ε log min { 1 , ρ ε } := ε log max { 1 , ρ ε } is bounded in L ∞ , → 0 in every finite L s . (26)
Here ess inf \operatorname*{ess\,inf} ess inf denotes the essential infimum. No positive lower bound for ρ ε \rho_\varepsilon ρ ε independent of ε \varepsilon ε is asserted.
Proof. Our proof is divided into three steps.
Step 1. By Jensen’s inequality, H ( ρ ) ≥ − log ∣ Ω ∣ H(\rho)\geq-\log|\Omega| H ( ρ ) ≥ − log ∣Ω∣ for every ρ ∈ A \rho\in\mathcal{A} ρ ∈ A . Hence J ε ( ρ ) ≥ ∣ D ρ ∣ ( Ω ) − ε log ∣ Ω ∣ J_\varepsilon(\rho)\geq|D\rho|(\Omega)-\varepsilon\log|\Omega| J ε ( ρ ) ≥ ∣ D ρ ∣ ( Ω ) − ε log ∣Ω∣ . A minimizing sequence is therefore bounded in B V ( Ω ) BV(\Omega) B V ( Ω ) . The compact embedding into L r ( Ω ) L^r(\Omega) L r ( Ω ) , the lower semicontinuity of the total variation and entropy, and the continuity of F F F under L r L^r L r convergence yield a minimizer ρ ε ∈ A \rho_\varepsilon\in\mathcal{A} ρ ε ∈ A . Uniqueness follows from the strict convexity of H H H and the convexity of the remaining terms. Let ρ 0 = ∣ Ω ∣ − 1 \rho_0=|\Omega|^{-1} ρ 0 = ∣Ω ∣ − 1 . Since H ( ρ 0 ) = − log ∣ Ω ∣ H(\rho_0)=-\log|\Omega| H ( ρ 0 ) = − log ∣Ω∣ , comparison with ρ 0 \rho_0 ρ 0 gives
∣ D ρ ε ∣ ( Ω ) ≤ J ε ( ρ ε ) − ε H ( ρ ε ) ≤ F ( ρ 0 ) − ε log ∣ Ω ∣ − ε H ( ρ ε ) ≤ n D 2 . |D\rho_\varepsilon|(\Omega)\leq J_\varepsilon(\rho_\varepsilon)-\varepsilon H(\rho_\varepsilon)
\leq F(\rho_0)-\varepsilon\log|\Omega|-\varepsilon H(\rho_\varepsilon)\leq nD^2. ∣ D ρ ε ∣ ( Ω ) ≤ J ε ( ρ ε ) − ε H ( ρ ε ) ≤ F ( ρ 0 ) − ε log ∣Ω∣ − ε H ( ρ ε ) ≤ n D 2 .
Together with the unit-mass constraint, this gives a uniform B V BV B V bound and hence a uniform L r L^r L r bound. Since ∣ s log s ∣ ≤ C r ( 1 + s r ) |s\log s|\leq C_r(1+s^r) ∣ s log s ∣ ≤ C r ( 1 + s r ) for s ≥ 0 s\geq0 s ≥ 0 , the entropies H ( ρ ε ) H(\rho_\varepsilon) H ( ρ ε ) are uniformly bounded. By compactness, along a sequence ε → 0 \varepsilon\to0 ε → 0 , ρ ε → ρ ∗ \rho_\varepsilon\to\rho_* ρ ε → ρ ∗ strongly in L r ( Ω ) L^r(\Omega) L r ( Ω ) , where ρ ∗ ∈ A \rho_*\in\mathcal{A} ρ ∗ ∈ A . For every η ∈ A \eta\in\mathcal{A} η ∈ A , optimality yields
J ( ρ ε ) ≤ J ( η ) + ε ( H ( η ) − H ( ρ ε ) ) . J(\rho_\varepsilon)\leq J(\eta)+\varepsilon\bigl(H(\eta)-H(\rho_\varepsilon)\bigr). J ( ρ ε ) ≤ J ( η ) + ε ( H ( η ) − H ( ρ ε ) ) .
Passing to the limit by lower semicontinuity gives J ( ρ ∗ ) ≤ J ( η ) J(\rho_*)\leq J(\eta) J ( ρ ∗ ) ≤ J ( η ) , so ρ ∗ \rho_* ρ ∗ minimizes J J J .
Step 2. We next establish a positive lower bound for each fixed ε > 0 \varepsilon>0 ε > 0 , following the perturbation argument of [9 Proposition 3.1] . For fixed ε \varepsilon ε , choose t > 0 t>0 t > 0 such that E : = { ρ ε > t } E:=\{\rho_\varepsilon>t\} E := { ρ ε > t } , has positive measure and finite relative perimeter. For 0 < δ ≤ t ∣ E ∣ 2 ∣ Ω ∣ 0<\delta\leq\frac{t|E|}{2|\Omega|} 0 < δ ≤ 2∣Ω∣ t ∣ E ∣ and m δ : = ∫ Ω ( δ − ρ ε ) + d x m_\delta:=\int_{\Omega}(\delta-\rho_\varepsilon)_+\,dx m δ := ∫ Ω ( δ − ρ ε ) + d x , define η δ : = max { ρ ε , δ } − m δ ∣ E ∣ 1 E \eta_\delta:=\max\{\rho_\varepsilon,\delta\}-\frac{m_\delta}{|E|}\mathbf{1}_E η δ := max { ρ ε , δ } − ∣ E ∣ m δ 1 E . Indeed, m δ ≤ δ ∣ Ω ∣ m_\delta\leq\delta|\Omega| m δ ≤ δ ∣Ω∣ implies m δ ∣ E ∣ ≤ t 2 \frac{m_\delta}{|E|}\leq\frac{t}{2} ∣ E ∣ m δ ≤ 2 t . Since δ ≤ t / 2 \delta\leq t/2 δ ≤ t /2 , the sets { ρ ε < δ } \{\rho_\varepsilon<\delta\} { ρ ε < δ } and E = { ρ ε > t } E=\{\rho_\varepsilon>t\} E = { ρ ε > t } are disjoint. The construction adds mass m δ m_\delta m δ on the former and removes the same amount from the latter. Moreover,
m δ ∣ E ∣ ≤ δ ∣ Ω ∣ ∣ E ∣ ≤ t 2 , \frac{m_\delta}{|E|}\leq\frac{\delta|\Omega|}{|E|}\leq\frac{t}{2}, ∣ E ∣ m δ ≤ ∣ E ∣ δ ∣Ω∣ ≤ 2 t ,
so η δ ≥ t / 2 \eta_\delta\geq t/2 η δ ≥ t /2 on E E E . Truncation preserves B V BV B V , and 1 E ∈ B V ( Ω ) \mathbf{1}_E\in BV(\Omega) 1 E ∈ B V ( Ω ) because E E E has finite relative perimeter. Hence η δ ∈ A \eta_\delta\in\mathcal{A} η δ ∈ A and ∥ η δ − ρ ε ∥ L 1 ( Ω ) = 2 m δ \|\eta_\delta-\rho_\varepsilon\|_{L^1(\Omega)}=2m_\delta ∥ η δ − ρ ε ∥ L 1 ( Ω ) = 2 m δ . Since s ↦ max { s , δ } s\mapsto\max\{s,\delta\} s ↦ max { s , δ } is 1-Lipschitz, ∣ D max { ρ ε , δ } ∣ ( Ω ) ≤ ∣ D ρ ε ∣ ( Ω ) |D\max\{\rho_\varepsilon,\delta\}|(\Omega)\leq|D\rho_\varepsilon|(\Omega) ∣ D max { ρ ε , δ } ∣ ( Ω ) ≤ ∣ D ρ ε ∣ ( Ω ) . Thus, by the triangle inequality for the total variation,
∣ D η δ ∣ ( Ω ) ≤ ∣ D max { ρ ε , δ } ∣ ( Ω ) + m δ ∣ E ∣ ∣ D 1 E ∣ ( Ω ) ≤ ∣ D ρ ε ∣ ( Ω ) + m δ ∣ E ∣ Per ( E ; Ω ) . \begin{aligned}
|D\eta_\delta|(\Omega)\leq|D\max\{\rho_\varepsilon,\delta\}|(\Omega)+\frac{m_\delta}{|E|}|D\mathbf{1}_E|(\Omega) \\
&\leq|D\rho_\varepsilon|(\Omega)+\frac{m_\delta}{|E|}\operatorname{Per}(E;\Omega).
\end{aligned} ∣ D η δ ∣ ( Ω ) ≤ ∣ D max { ρ ε , δ } ∣ ( Ω ) + ∣ E ∣ m δ ∣ D 1 E ∣ ( Ω ) ≤ ∣ D ρ ε ∣ ( Ω ) + ∣ E ∣ m δ Per ( E ; Ω ) .
By (5.3), we have F ( η δ ) − F ( ρ ε ) ≤ 4 n D 2 m δ F(\eta_\delta)-F(\rho_\varepsilon)\leq4nD^2m_\delta F ( η δ ) − F ( ρ ε ) ≤ 4 n D 2 m δ . To estimate the entropy, write h ( s ) = s log s h(s)=s\log s h ( s ) = s log s . On E E E , set a : = m δ / ∣ E ∣ a:=m_\delta/|E| a := m δ /∣ E ∣ . Since η δ = ρ ε − a ≥ t / 2 \eta_\delta=\rho_\varepsilon-a\geq t/2 η δ = ρ ε − a ≥ t /2 , we have
h ( η δ ) − h ( ρ ε ) = − ∫ η δ ρ ε ( 1 + log s ) d s ≤ a [ − 1 − log ( t / 2 ) ] ≤ a C ( t ) , \begin{aligned}
h(\eta_\delta)-h(\rho_\varepsilon)=-\int_{\eta_\delta}^{\rho_\varepsilon}(1+\log s)\,\mathrm{d}s \\
&\leq a[-1-\log(t/2)]\leq aC(t),
\end{aligned} h ( η δ ) − h ( ρ ε ) = − ∫ η δ ρ ε ( 1 + log s ) d s ≤ a [ − 1 − log ( t /2 )] ≤ a C ( t ) ,
where C ( t ) : = max { 0 , − 1 − log ( t / 2 ) } C(t):=\max\{0,-1-\log(t/2)\} C ( t ) := max { 0 , − 1 − log ( t /2 )} . Integrating over E E E yields ∫ E [ h ( η δ ) − h ( ρ ε ) ] d x ≤ m δ C ( t ) \int_E[h(\eta_\delta)-h(\rho_\varepsilon)]\,\mathrm{d}x\leq m_\delta C(t) ∫ E [ h ( η δ ) − h ( ρ ε )] d x ≤ m δ C ( t ) . Since η δ = δ \eta_\delta=\delta η δ = δ on A δ : = { ρ ε < δ } A_\delta:=\{\rho_\varepsilon<\delta\} A δ := { ρ ε < δ } and η δ = ρ ε \eta_\delta=\rho_\varepsilon η δ = ρ ε outside A δ ∪ E A_\delta\cup E A δ ∪ E , the disjointness of A δ A_\delta A δ and E E E gives
H ( η δ ) − H ( ρ ε ) = ∫ A δ [ h ( δ ) − h ( ρ ε ) ] d x + ∫ E [ h ( η δ ) − h ( ρ ε ) ] d x ≤ ( 1 + log δ ) ∫ A δ ( δ − ρ ε ) d x + m δ C ( t ) = m δ ( 1 + log δ + C ( t ) ) . \begin{aligned}
H(\eta_\delta)-H(\rho_\varepsilon)=\int_{A_\delta}[h(\delta)-h(\rho_\varepsilon)]\,\mathrm{d}x+\int_E[h(\eta_\delta)-h(\rho_\varepsilon)]\,\mathrm{d}x \\
&\leq(1+\log\delta)\int_{A_\delta}(\delta-\rho_\varepsilon)\,\mathrm{d}x+m_\delta C(t) \\
&=m_\delta(1+\log\delta+C(t)).
\end{aligned} H ( η δ ) − H ( ρ ε ) = ∫ A δ [ h ( δ ) − h ( ρ ε )] d x + ∫ E [ h ( η δ ) − h ( ρ ε )] d x ≤ ( 1 + log δ ) ∫ A δ ( δ − ρ ε ) d x + m δ C ( t ) = m δ ( 1 + log δ + C ( t )) .
Combining these estimates with optimality gives
0 ≤ J ε ( η δ ) − J ε ( ρ ε ) ≤ m δ [ 4 n D 2 + Per ( E ; Ω ) ∣ E ∣ + ε ( 1 + log δ + C ( t ) ) ] . (27) 0\leq J_\varepsilon(\eta_\delta)-J_\varepsilon(\rho_\varepsilon)\leq m_\delta\left[4nD^2+\frac{\operatorname{Per}(E;\Omega)}{|E|}+\varepsilon(1+\log\delta+C(t))\right]. \tag*{(27)} 0 ≤ J ε ( η δ ) − J ε ( ρ ε ) ≤ m δ [ 4 n D 2 + ∣ E ∣ Per ( E ; Ω ) + ε ( 1 + log δ + C ( t )) ] . (27)
For sufficiently small δ > 0 \delta>0 δ > 0 , the bracket is strictly negative. Hence m δ = 0 m_\delta=0 m δ = 0 , which implies ρ ε ≥ δ \rho_\varepsilon\geq\delta ρ ε ≥ δ almost everywhere and proves ess inf Ω ρ ε > 0 \operatorname*{ess\,inf}_{\Omega}\rho_\varepsilon>0 ess inf Ω ρ ε > 0 .
Step 3. To obtain a uniform bound for b ε − b_\varepsilon^- b ε − along the selected sequence, choose 0 < a < b < c 0<a<b<c 0 < a < b < c such that ∣ { ρ ∗ > c } ∣ > 0 |\{\rho_*>c\}|>0 ∣ { ρ ∗ > c } ∣ > 0 . Set v : = 1 2 ∣ { ρ ∗ > c } ∣ > 0 v:=\frac{1}{2}|\{\rho_*>c\}|>0 v := 2 1 ∣ { ρ ∗ > c } ∣ > 0 . Convergence in measure implies that, for all sufficiently small ε \varepsilon ε and every t ∈ [ a , b ] t\in[a,b] t ∈ [ a , b ] , ∣ { ρ ε > t } ∣ ≥ v > 0 |\{\rho_\varepsilon>t\}|\geq v>0 ∣ { ρ ε > t } ∣ ≥ v > 0 with v v v independent of ε \varepsilon ε . By the coarea formula and the uniform variation bound, we can choose t ε ∈ ( a , b ) t_\varepsilon\in(a,b) t ε ∈ ( a , b ) such that E ε : = { ρ ε > t ε } E_\varepsilon:=\{\rho_\varepsilon>t_\varepsilon\} E ε := { ρ ε > t ε } and Per ( E ε ; Ω ) ≤ n D 2 b − a \operatorname{Per}(E_\varepsilon;\Omega)\leq\frac{nD^2}{b-a} Per ( E ε ; Ω ) ≤ b − a n D 2 .
Thus
Per ( E ε ; Ω ) ∣ E ε ∣ ≤ n D 2 ( b − a ) v . \frac{\operatorname{Per}(E_\varepsilon;\Omega)}{|E_\varepsilon|}\leq\frac{nD^2}{(b-a)v}. ∣ E ε ∣ Per ( E ε ; Ω ) ≤ ( b − a ) v n D 2 .
Set c 0 : = min { 1 , a v / ( 2 ∣ Ω ∣ ) } c_0:=\min\{1,av/(2|\Omega|)\} c 0 := min { 1 , a v / ( 2∣Ω∣ )} . Using E ε E_\varepsilon E ε in (5.5) and choosing δ ε : = min { c 0 , exp ( − C / ε ) } \delta_\varepsilon:=\min\{c_0,\exp(-C/\varepsilon)\} δ ε := min { c 0 , exp ( − C / ε )} with C C C sufficiently large and independent of ε \varepsilon ε , the bracket is strictly negative. Therefore, ρ ε ≥ δ ε \rho_\varepsilon\geq\delta_\varepsilon ρ ε ≥ δ ε almost everywhere and hence 0 ≤ b ε − ≤ max { − ε log c 0 , C } 0\leq b_\varepsilon^-\leq\max\{-\varepsilon\log c_0,C\} 0 ≤ b ε − ≤ max { − ε log c 0 , C } . This proves the uniform L ∞ L^\infty L ∞ bound. Finally, for every 1 ≤ s < ∞ 1\leq s<\infty 1 ≤ s < ∞ and t ≥ 0 t\geq0 t ≥ 0 , ( log max { 1 , t } ) s ≤ C r , s ( 1 + t r ) (\log\max\{1,t\})^s\leq C_{r,s}(1+t^r) ( log max { 1 , t } ) s ≤ C r , s ( 1 + t r ) . The uniform L r L^r L r bound therefore yields
∥ b ε + ∥ L s ( Ω ) s ≤ C r , s ε s ( ∣ Ω ∣ + ∥ ρ ε ∥ L r ( Ω ) r ) ≤ C s ε s . \|b_\varepsilon^+\|_{L^s(\Omega)}^s\leq C_{r,s}\varepsilon^s\left(|\Omega|+\|\rho_\varepsilon\|_{L^r(\Omega)}^r\right)\leq C_s\varepsilon^s. ∥ b ε + ∥ L s ( Ω ) s ≤ C r , s ε s ( ∣Ω∣ + ∥ ρ ε ∥ L r ( Ω ) r ) ≤ C s ε s .
Thus b ε + → 0 b_\varepsilon^+\to0 b ε + → 0 in every finite L s ( Ω ) L^s(\Omega) L s ( Ω ) .
First variation and optimality conditions
The positivity obtained above allows us to differentiate the transport terms with respect to the density. We first establish the required first-variation formula and then combine it with the entropy derivative and the total variation inequality to obtain the regularized optimality conditions.
Lemma 5.3 (First variation at a positive density). Let σ ∈ A \sigma\in\mathcal{A} σ ∈ A have positive essential infimum. For arbitrary μ ∈ P ( K ) \mu\in\mathcal{P}(K) μ ∈ P ( K ) , there exists a unique c c c -concave potential f ∈ C ( K ) f\in C(K) f ∈ C ( K ) satisfying f ( x ∗ ) = 0 f(x_*)=0 f ( x ∗ ) = 0 and W 2 2 ( μ , σ d x ) = ∫ K f c d μ + ∫ Ω f σ d x W_2^2(\mu,\sigma\,dx)=\int_K f^c\,d\mu+\int_\Omega f\sigma\,dx W 2 2 ( μ , σ d x ) = ∫ K f c d μ + ∫ Ω f σ d x . For any probability density η ∈ L 1 ( Ω ) \eta\in L^1(\Omega) η ∈ L 1 ( Ω ) , set σ t = ( 1 − t ) σ + t η \sigma_t=(1-t)\sigma+t\eta σ t = ( 1 − t ) σ + t η for 0 ≤ t < 1 0\le t<1 0 ≤ t < 1 . Then
lim t → 0 W 2 2 ( μ , σ t d x ) − W 2 2 ( μ , σ d x ) t = ∫ Ω f ( η − σ ) d x . (28) \lim_{t\to0}\frac{W_2^2(\mu,\sigma_t\,dx)-W_2^2(\mu,\sigma\,dx)}{t}=\int_\Omega f(\eta-\sigma)\,dx. \tag*{(28)} t → 0 lim t W 2 2 ( μ , σ t d x ) − W 2 2 ( μ , σ d x ) = ∫ Ω f ( η − σ ) d x . (28)
Moreover, if w ∈ L ∞ ( Ω ) w\in L^\infty(\Omega) w ∈ L ∞ ( Ω ) satisfies ∫ Ω w d x = 0 \int_\Omega w\,dx=0 ∫ Ω w d x = 0 , then
lim t → 0 W 2 2 ( μ , ( σ + t w ) d x ) − W 2 2 ( μ , σ d x ) t = ∫ Ω f w d x , \lim_{t\to0}\frac{W_2^2(\mu,(\sigma+tw)\,dx)-W_2^2(\mu,\sigma\,dx)}{t}=\int_\Omega fw\,dx, t → 0 lim t W 2 2 ( μ , ( σ + tw ) d x ) − W 2 2 ( μ , σ d x ) = ∫ Ω f w d x ,
where σ + t w \sigma+tw σ + tw is nonnegative for all sufficiently small ∣ t ∣ |t| ∣ t ∣ .
Proof. Let f f f and g g g be two c c c -concave optimal potentials associated with σ d x \sigma\,dx σ d x , both normalized by f ( x ∗ ) = g ( x ∗ ) = 0 f(x_\ast)=g(x_\ast)=0 f ( x ∗ ) = g ( x ∗ ) = 0 . Since σ d x \sigma\,dx σ d x is absolutely continuous, Brenier’s theorem gives a unique optimal transport map from σ d x \sigma\,dx σ d x to μ \mu μ . For the cost ∣ x − y ∣ 2 |x-y|^2 ∣ x − y ∣ 2 , this map is represented by both
T ( x ) = x − 1 2 ∇ f ( x ) = x − 1 2 ∇ g ( x ) T(x)=x-\frac{1}{2}\nabla f(x)=x-\frac{1}{2}\nabla g(x) T ( x ) = x − 2 1 ∇ f ( x ) = x − 2 1 ∇ g ( x )
for σ d x \sigma\,dx σ d x -almost every x x x . As σ > 0 \sigma>0 σ > 0 almost everywhere, it follows that ∇ f = ∇ g \nabla f=\nabla g ∇ f = ∇ g Lebesgue-almost everywhere in Ω \Omega Ω . Poincaré’s inequality on the connected domain Ω \Omega Ω implies that f − g f-g f − g is constant almost everywhere. Continuity extends this equality to K K K , and the normalization f ( x ∗ ) = g ( x ∗ ) = 0 f(x_\ast)=g(x_\ast)=0 f ( x ∗ ) = g ( x ∗ ) = 0 then gives f = g f=g f = g on K K K .
For 0 < t < 1 0<t<1 0 < t < 1 , choose a normalized c c c -concave optimal potential f t ∈ K f_t\in K f t ∈ K for ( μ , σ t d x ) (\mu,\sigma_t\,dx) ( μ , σ t d x ) . By compactness of K K K , every sequence t k → 0 t_k\to0 t k → 0 admits a subsequence along which f t k f_{t_k} f t k converges uniformly. Since σ t → σ \sigma_t\to\sigma σ t → σ in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) and the c c c -transform is continuous with respect to the supremum norm on C ( K ) C(K) C ( K ) , passing to the limit in the dual identity shows that each such limit is a normalized c c c -concave optimal potential for ( μ , σ d x ) (\mu,\sigma\,dx) ( μ , σ d x ) . By the uniqueness proved above, every limit equals f f f . Hence f t → f f_t\to f f t → f uniformly on K K K as t → 0 t\to0 t → 0 . Using f f f as a dual competitor at σ t \sigma_t σ t and f t f_t f t as a dual competitor at σ \sigma σ , we obtain
∫ Ω f ( η − σ ) d x ≤ W 2 2 ( μ , σ t d x ) − W 2 2 ( μ , σ d x ) t ≤ ∫ Ω f t ( η − σ ) d x . \int_\Omega f(\eta-\sigma)\,dx\le\frac{W_2^2(\mu,\sigma_t\,dx)-W_2^2(\mu,\sigma\,dx)}{t}\le\int_\Omega f_t(\eta-\sigma)\,dx. ∫ Ω f ( η − σ ) d x ≤ t W 2 2 ( μ , σ t d x ) − W 2 2 ( μ , σ d x ) ≤ ∫ Ω f t ( η − σ ) d x .
Uniform convergence of f t f_t f t implies that both bounds converge to ∫ Ω f ( η − σ ) d x \int_\Omega f(\eta-\sigma)\,dx ∫ Ω f ( η − σ ) d x , proving (28) . The same argument applies to σ + t w \sigma+tw σ + tw for bounded zero integral perturbations w w w , since the positive lower bound of σ \sigma σ ensures admissibility for sufficiently small ∣ t ∣ |t| ∣ t ∣ . □
We apply this formula at the regularized minimizer ρ ε \rho_\varepsilon ρ ε and use the representation of total variation established above.
Proposition 5.1 (The approximate Euler-Lagrange equation). Let f i , ε ∈ K f_{i,\varepsilon}\in K f i , ε ∈ K be normalized c c c -concave potentials for ( μ i , ρ ε d x ) (\mu_i,\rho_\varepsilon\,dx) ( μ i , ρ ε d x ) and set V ε = ∑ i f i , ε V_\varepsilon=\sum_i f_{i,\varepsilon} V ε = ∑ i f i , ε . There exist λ ε ∈ R \lambda_\varepsilon\in\mathbb{R} λ ε ∈ R and z ε ∈ L ∞ z_\varepsilon\in L^\infty z ε ∈ L ∞ with ∥ z ε ∥ ∞ ≤ 1 \|z_\varepsilon\|_\infty\le1 ∥ z ε ∥ ∞ ≤ 1 , zero normal trace, and
div z ε = λ ε − V ε − ε log ρ ε , (29) \operatorname{div}z_\varepsilon=\lambda_\varepsilon-V_\varepsilon-\varepsilon\log\rho_\varepsilon, \tag*{(29)} div z ε = λ ε − V ε − ε log ρ ε , (29) ∫ Ω ρ ε div z ε d x = TV ( ρ ε ) . (30) \int_\Omega\rho_\varepsilon\operatorname{div}z_\varepsilon\,dx=\operatorname{TV}(\rho_\varepsilon). \tag*{(30)} ∫ Ω ρ ε div z ε d x = TV ( ρ ε ) . (30) The divergence belongs to every finite L s ( Ω ) L^s(\Omega) L s ( Ω ) .
Proof. Take w ∈ B V ( Ω ) ∩ L ∞ ( Ω ) w \in BV(\Omega) \cap L^\infty(\Omega) w ∈ B V ( Ω ) ∩ L ∞ ( Ω ) with ∫ Ω w d x = 0 \int_\Omega w\,dx = 0 ∫ Ω w d x = 0 . Since ess inf Ω ρ ε > 0 \operatorname{ess\,inf}_\Omega\rho_\varepsilon> 0 ess inf Ω ρ ε > 0 (Lemma 5.2), we have ρ ε + t w ∈ A \rho_\varepsilon+ tw \in\mathcal{A} ρ ε + tw ∈ A for all sufficiently small t > 0 t > 0 t > 0 . For such t t t , the triangle inequality gives ∣ D ( ρ ε + t w ) ∣ ( Ω ) ≤ ∣ D ρ ε ∣ ( Ω ) + t ∣ D w ∣ ( Ω ) |D(\rho_\varepsilon+ tw)|(\Omega) \le|D\rho_\varepsilon|(\Omega) + t|Dw|(\Omega) ∣ D ( ρ ε + tw ) ∣ ( Ω ) ≤ ∣ D ρ ε ∣ ( Ω ) + t ∣ D w ∣ ( Ω ) . Minimality of ρ ε \rho_\varepsilon ρ ε therefore yields
0 ≤ J ε ( ρ ε + t w ) − J ε ( ρ ε ) t ≤ F ( ρ ε + t w ) − F ( ρ ε ) t + ∣ D w ∣ ( Ω ) + ε H ( ρ ε + t w ) − H ( ρ ε ) t . 0 \le\frac{J_\varepsilon(\rho_\varepsilon+ tw) - J_\varepsilon(\rho_\varepsilon)}{t}
\le\frac{F(\rho_\varepsilon+ tw) - F(\rho_\varepsilon)}{t} + |Dw|(\Omega) + \varepsilon\frac{H(\rho_\varepsilon+ tw) - H(\rho_\varepsilon)}{t}. 0 ≤ t J ε ( ρ ε + tw ) − J ε ( ρ ε ) ≤ t F ( ρ ε + tw ) − F ( ρ ε ) + ∣ D w ∣ ( Ω ) + ε t H ( ρ ε + tw ) − H ( ρ ε ) .
To differentiate the entropy, write h ( s ) = s log s h(s) = s \log s h ( s ) = s log s . The positive lower bound for ρ ε \rho_\varepsilon ρ ε and boundedness of w w w give, for sufficiently small t > 0 t > 0 t > 0 ,
∣ h ( ρ ε + t w ) − h ( ρ ε ) t ∣ ≤ C ε , w ( 1 + log ( 1 + ρ ε ) ) . \left|\frac{h(\rho_\varepsilon+ tw) - h(\rho_\varepsilon)}{t}\right| \le C_{\varepsilon,w}(1+\log(1+\rho_\varepsilon)). t h ( ρ ε + tw ) − h ( ρ ε ) ≤ C ε , w ( 1 + log ( 1 + ρ ε )) .
The right-hand side is integrable and independent of t t t . Dominated convergence therefore yields
lim t ↓ 0 H ( ρ ε + t w ) − H ( ρ ε ) t = ∫ Ω ( 1 + log ρ ε ) w d x . \lim_{t\downarrow0}\frac{H(\rho_\varepsilon+ tw) - H(\rho_\varepsilon)}{t}
= \int_\Omega(1+\log\rho_\varepsilon)w\,dx. t ↓ 0 lim t H ( ρ ε + tw ) − H ( ρ ε ) = ∫ Ω ( 1 + log ρ ε ) w d x .
Using the first-variation formula for F F F and ∫ Ω w d x = 0 \int_\Omega w\,dx = 0 ∫ Ω w d x = 0 , we conclude that
∣ D w ∣ ( Ω ) ≥ − ∫ Ω ( V ε + ε log ρ ε ) w d x . (31) |Dw|(\Omega) \ge-\int_\Omega(V_\varepsilon+ \varepsilon\log\rho_\varepsilon)w\,dx. \tag*{(31)} ∣ D w ∣ ( Ω ) ≥ − ∫ Ω ( V ε + ε log ρ ε ) w d x . (31)
The constant 1 in the derivative of ρ log ρ \rho\log\rho ρ log ρ cancels because ∫ w = 0 \int w = 0 ∫ w = 0 . Define
λ ε = 1 ∣ Ω ∣ ∫ Ω ( V ε + ε log ρ ε ) d x , ξ ε = λ ε − V ε − ε log ρ ε . \lambda_\varepsilon= \frac{1}{|\Omega|}\int_\Omega(V_\varepsilon+ \varepsilon\log\rho_\varepsilon)\,dx,\qquad
\xi_\varepsilon= \lambda_\varepsilon- V_\varepsilon- \varepsilon\log\rho_\varepsilon. λ ε = ∣Ω∣ 1 ∫ Ω ( V ε + ε log ρ ε ) d x , ξ ε = λ ε − V ε − ε log ρ ε .
By construction, ∫ Ω ξ ε d x = 0 \int_\Omega\xi_\varepsilon\,dx = 0 ∫ Ω ξ ε d x = 0 . For any w ∈ B V ( Ω ) ∩ L ∞ ( Ω ) w \in BV(\Omega) \cap L^\infty(\Omega) w ∈ B V ( Ω ) ∩ L ∞ ( Ω ) , apply the preceding inequality (5.9) to w − w ‾ w-\overline{w} w − w with w ‾ : = 1 ∣ Ω ∣ ∫ Ω w d x \overline{w} := \frac{1}{|\Omega|}\int_\Omega w\,dx w := ∣Ω∣ 1 ∫ Ω w d x . Subtracting a constant changes neither the total variation nor the integral against ξ ε \xi_\varepsilon ξ ε , so ∫ Ω ξ ε w d x ≤ ∣ D w ∣ ( Ω ) \int_\Omega\xi_\varepsilon w\,dx \le|Dw|(\Omega) ∫ Ω ξ ε w d x ≤ ∣ D w ∣ ( Ω ) . The positive lower bound for ρ ε \rho_\varepsilon ρ ε and its logarithmic integrability imply log ρ ε ∈ L q ( Ω ) \log\rho_\varepsilon\in L^q(\Omega) log ρ ε ∈ L q ( Ω ) , hence ξ ε ∈ L q ( Ω ) \xi_\varepsilon\in L^q(\Omega) ξ ε ∈ L q ( Ω ) . Lemma 5.1 therefore yields a vector field z ε ∈ L ∞ ( Ω ; R d ) z_\varepsilon\in L^\infty(\Omega;\mathbb{R}^d) z ε ∈ L ∞ ( Ω ; R d ) with ∥ z ε ∥ ∞ ≤ 1 \|z_\varepsilon\|_\infty\le1 ∥ z ε ∥ ∞ ≤ 1 , zero normal trace, and div z ε = ξ ε \operatorname{div} z_\varepsilon= \xi_\varepsilon div z ε = ξ ε .
It remains to prove (5.8). Let ρ 0 : = ∣ Ω ∣ − 1 \rho_0 := |\Omega|^{-1} ρ 0 := ∣Ω ∣ − 1 and consider σ t : = ( 1 − t ) ρ ε + t ρ 0 \sigma_t := (1-t)\rho_\varepsilon+ t\rho_0 σ t := ( 1 − t ) ρ ε + t ρ 0 , where 0 ≤ t < 1 0 \le t < 1 0 ≤ t < 1 . This path lies in A \mathcal{A} A . Since ρ 0 \rho_0 ρ 0 is constant, ∣ D σ t ∣ ( Ω ) = ( 1 − t ) ∣ D ρ ε ∣ ( Ω ) |D\sigma_t|(\Omega) = (1-t)|D\rho_\varepsilon|(\Omega) ∣ D σ t ∣ ( Ω ) = ( 1 − t ) ∣ D ρ ε ∣ ( Ω ) . Lemma 5.3 gives
d d t F ( σ t ) ∣ t = 0 + = ∫ Ω V ε ( ρ 0 − ρ ε ) d x . \left.\frac{d}{dt}F(\sigma_t)\right|_{t=0+}
= \int_\Omega V_\varepsilon(\rho_0-\rho_\varepsilon)\,dx. d t d F ( σ t ) t = 0 + = ∫ Ω V ε ( ρ 0 − ρ ε ) d x .
For 0 < t < 1 / 2 0 < t < 1/2 0 < t < 1/2 , the positive lower bound for ρ ε \rho_\varepsilon ρ ε and the mean value theorem give
∣ h ( σ t ) − h ( ρ ε ) t ∣ ≤ C ε , r ( 1 + ρ ε r ) , \left|\frac{h(\sigma_t)-h(\rho_\varepsilon)}{t}\right| \le C_{\varepsilon,r}(1+\rho_\varepsilon^r), t h ( σ t ) − h ( ρ ε ) ≤ C ε , r ( 1 + ρ ε r ) ,
where the right-hand side belongs to L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) and is independent of t t t . Dominated convergence therefore yields
d d t H ( σ t ) ∣ t = 0 + = ∫ Ω ( 1 + log ρ ε ) ( ρ 0 − ρ ε ) d x . \left.\frac{d}{dt}H(\sigma_t)\right|_{t=0+}
= \int_\Omega(1+\log\rho_\varepsilon)(\rho_0-\rho_\varepsilon)\,dx. d t d H ( σ t ) t = 0 + = ∫ Ω ( 1 + log ρ ε ) ( ρ 0 − ρ ε ) d x .
The constant term vanishes because both densities have unit mass. By minimality of ρ ε \rho_\varepsilon ρ ε ,
0 ≤ d d t J ε ( σ t ) ∣ t = 0 + = − ∣ D ρ ε ∣ ( Ω ) + ∫ Ω ( V ε + ε log ρ ε ) ( ρ 0 − ρ ε ) d x = − ∣ D ρ ε ∣ ( Ω ) + ∫ Ω ξ ε ρ ε d x . \begin{aligned}
0 \leq\left.\frac{d}{dt}J_\varepsilon(\sigma_t)\right|_{t=0^+} \\
&= -|D\rho_\varepsilon|(\Omega) + \int_\Omega(V_\varepsilon+\varepsilon\log\rho_\varepsilon)(\rho_0-\rho_\varepsilon)\,dx \\
&= -|D\rho_\varepsilon|(\Omega) + \int_\Omega\xi_\varepsilon\rho_\varepsilon\,dx.
\end{aligned} 0 ≤ d t d J ε ( σ t ) t = 0 + = − ∣ D ρ ε ∣ ( Ω ) + ∫ Ω ( V ε + ε log ρ ε ) ( ρ 0 − ρ ε ) d x = − ∣ D ρ ε ∣ ( Ω ) + ∫ Ω ξ ε ρ ε d x .
In the last equality, we used V ε + ε log ρ ε = λ ε − ξ ε V_\varepsilon+\varepsilon\log\rho_\varepsilon=\lambda_\varepsilon-\xi_\varepsilon V ε + ε log ρ ε = λ ε − ξ ε , the equality of masses, and ∫ Ω ξ ε ρ 0 d x = 0 \int_\Omega\xi_\varepsilon\rho_0\,dx=0 ∫ Ω ξ ε ρ 0 d x = 0 . Conversely, (6) and div z ε = ξ ε \operatorname{div}z_\varepsilon=\xi_\varepsilon div z ε = ξ ε give ∫ Ω ξ ε ρ ε d x ≤ ∣ D ρ ε ∣ ( Ω ) \int_\Omega\xi_\varepsilon\rho_\varepsilon\,dx\leq|D\rho_\varepsilon|(\Omega) ∫ Ω ξ ε ρ ε d x ≤ ∣ D ρ ε ∣ ( Ω ) . Hence ∫ Ω ρ ε div z ε d x = ∣ D ρ ε ∣ ( Ω ) \int_\Omega\rho_\varepsilon\operatorname{div}z_\varepsilon\,dx=|D\rho_\varepsilon|(\Omega) ∫ Ω ρ ε div z ε d x = ∣ D ρ ε ∣ ( Ω ) , which proves (30) . Finally, (29) and the logarithmic integrability estimates imply div z ε ∈ L s ( Ω ) \operatorname{div}z_\varepsilon\in L^s(\Omega) div z ε ∈ L s ( Ω ) for every 1 ≤ s < ∞ 1\leq s<\infty 1 ≤ s < ∞ .
The vector-field dual problem
We now pass to the limit in the regularized optimality conditions. This produces a maximizing tuple for the vector-field dual and yields necessary and sufficient conditions for a density to minimize J J J . Let Z \mathcal{Z} Z be the fields z ∈ L ∞ ( Ω ; R d ) z\in L^\infty(\Omega;\mathbb{R}^d) z ∈ L ∞ ( Ω ; R d ) with ∥ z ∥ ∞ ≤ 1 \|z\|_\infty\leq1 ∥ z ∥ ∞ ≤ 1 , div z ∈ L ∞ ( Ω ) \operatorname{div}z\in L^\infty(\Omega) div z ∈ L ∞ ( Ω ) , and zero normal trace in the sense of (23) .
Theorem 5.1 (Vector-field duality and optimality). For arbitrary prescribed measures μ i ∈ P ( K ) \mu_i\in\mathcal{P}(K) μ i ∈ P ( K ) ,
min ρ ∈ A J ( ρ ) = max { λ + ∑ i = 1 n ∫ K f i c d μ i : f i ∈ C ( K ) , z ∈ Z , λ ∈ R , ∑ i f i + div z ≥ λ a.e. } . (32) \min_{\rho\in\mathcal{A}}J(\rho)=\max\left\{\lambda+\sum_{i=1}^{n}\int_K f_i^c\,d\mu_i:f_i\in C(K),\ z\in\mathcal{Z},\ \lambda\in\mathbb{R},\ \sum_i f_i+\operatorname{div}z\geq\lambda\quad\text{a.e.}\right\}. \tag*{(32)} ρ ∈ A min J ( ρ ) = max { λ + i = 1 ∑ n ∫ K f i c d μ i : f i ∈ C ( K ) , z ∈ Z , λ ∈ R , i ∑ f i + div z ≥ λ a.e. } . (32) There is a maximizing tuple with f i ∈ K f_i\in\mathcal{K} f i ∈ K . For any fixed maximizing tuple, set β = ∑ i f i + div z − λ \beta=\sum_i f_i+\operatorname{div}z-\lambda β = ∑ i f i + div z − λ . A density ρ ∈ A \rho\in\mathcal{A} ρ ∈ A is a minimizer if and only if
W 2 2 ( μ i , ρ d x ) = ∫ K f i c d μ i + ∫ Ω f i ρ d x ( i = 1 , … , n ) , (33) W_2^2(\mu_i,\rho\,dx)=\int_K f_i^c\,d\mu_i+\int_\Omega f_i\rho\,dx\qquad(i=1,\ldots,n), \tag*{(33)} W 2 2 ( μ i , ρ d x ) = ∫ K f i c d μ i + ∫ Ω f i ρ d x ( i = 1 , … , n ) , (33) β ρ = 0 a.e. and ∫ Ω ρ div z d x = ∣ D ρ ∣ ( Ω ) . (34) \beta\rho=0\ \text{a.e.}\quad\text{and}\quad\int_\Omega\rho\operatorname{div}z\,dx=|D\rho|(\Omega). \tag*{(34)} β ρ = 0 a.e. and ∫ Ω ρ div z d x = ∣ D ρ ∣ ( Ω ) . (34) In particular, β ∈ L ∞ \beta\in L^\infty β ∈ L ∞ , β ≥ 0 \beta\geq0 β ≥ 0 , and the TV condition is equivalently − ( z , D ρ ) = ∣ D ρ ∣ -(z,D\rho)=|D\rho| − ( z , D ρ ) = ∣ D ρ ∣ as measures.
Proof. Let ( f , z , λ ) (f,z,\lambda) ( f , z , λ ) be any admissible tuple for the dual problem, and set V : = ∑ i = 1 n f i V:=\sum_{i=1}^{n}f_i V := ∑ i = 1 n f i . For every η ∈ A \eta\in\mathcal{A} η ∈ A , Kantorovich duality gives
∑ i = 1 n W 2 2 ( μ i , η d x ) ≥ B ( f ) + ∫ Ω V η d x , \sum_{i=1}^{n}W_2^2(\mu_i,\eta\,dx)\geq B(f)+\int_\Omega V\eta\,dx, i = 1 ∑ n W 2 2 ( μ i , η d x ) ≥ B ( f ) + ∫ Ω V η d x ,
while (6) yields ∣ D η ∣ ( Ω ) ≥ ∫ Ω η div z d x |D\eta|(\Omega)\geq\int_\Omega\eta\operatorname{div}z\,dx ∣ D η ∣ ( Ω ) ≥ ∫ Ω η div z d x . Combining these inequalities with the dual constraint V + div z ≥ λ V+\operatorname{div}z\geq\lambda V + div z ≥ λ , we obtain
J ( η ) ≥ B ( f ) + ∫ Ω ( V + div z ) η d x ≥ B ( f ) + λ ∫ Ω η d x = B ( f ) + λ , \begin{aligned}
J(\eta)\geq B(f)+\int_\Omega(V+\operatorname{div}z)\eta\,dx \\
&\geq B(f)+\lambda\int_\Omega\eta\,dx \\
&=B(f)+\lambda,
\end{aligned} J ( η ) ≥ B ( f ) + ∫ Ω ( V + div z ) η d x ≥ B ( f ) + λ ∫ Ω η d x = B ( f ) + λ ,
where we used η ≥ 0 \eta\geq0 η ≥ 0 and ∫ Ω η d x = 1 \int_\Omega\eta\,dx=1 ∫ Ω η d x = 1 . Taking the infimum over η ∈ A \eta\in\mathcal{A} η ∈ A and the supremum over all dual admissible tuples shows that the dual supremum does not exceed the primal minimum. We next pass to the limit in the regularized optimality conditions. Along the sequence provided by Lemma 5.2, ρ ε → ρ ∗ \rho_\varepsilon\to\rho_\ast ρ ε → ρ ∗ strongly in L r ( Ω ) L^r(\Omega) L r ( Ω ) , where ρ ∗ \rho_\ast ρ ∗ minimizes J J J . Since f i , ε ∈ K f_{i,\varepsilon} \in\mathcal{K} f i , ε ∈ K and K \mathcal{K} K is compact in C ( K ) C(\mathcal{K}) C ( K ) , we may extract a subsequence such that f i , ε → f i ∈ K f_{i,\varepsilon} \to f_i \in\mathcal{K} f i , ε → f i ∈ K uniformly on K \mathcal{K} K , for i = 1 , … , n i=1,\ldots,n i = 1 , … , n . Consequently, V ε : = ∑ i f i , ε V_\varepsilon:= \sum_i f_{i,\varepsilon} V ε := ∑ i f i , ε converges uniformly to V : = ∑ i f i V := \sum_i f_i V := ∑ i f i . Using ε log ρ ε = b ε + − b ε − \varepsilon\log\rho_\varepsilon= b_\varepsilon^+ - b_\varepsilon^- ε log ρ ε = b ε + − b ε − , we can write
λ ε = 1 ∣ Ω ∣ ∫ Ω ( V ε + b ε + − b ε − ) d x . \lambda_\varepsilon= \frac{1}{\lvert\Omega\rvert}\int_\Omega\left(V_\varepsilon+b_\varepsilon^+-b_\varepsilon^-\right)\,dx. λ ε = ∣ Ω ∣ 1 ∫ Ω ( V ε + b ε + − b ε − ) d x .
The uniform bound for V ε V_\varepsilon V ε and (5.4) therefore imply that ( λ ε ) (\lambda_\varepsilon) ( λ ε ) is bounded. Since ∥ z ε ∥ ∞ ≤ 1 \lVert z_\varepsilon\rVert_\infty\le1 ∥ z ε ∥ ∞ ≤ 1 and ( b ε − ) (b_\varepsilon^-) ( b ε − ) is bounded in L ∞ ( Ω ) L^\infty(\Omega) L ∞ ( Ω ) , a further subsequence satisfies λ ε → λ \lambda_\varepsilon\to\lambda λ ε → λ , z ε ⇀ ∗ z z_\varepsilon\overset{*}{\rightharpoonup} z z ε ⇀ ∗ z in L ∞ ( Ω ; R d ) L^\infty(\Omega;\mathbb{R}^d) L ∞ ( Ω ; R d ) , and b ε − ⇀ ∗ β b_\varepsilon^- \overset{*}{\rightharpoonup} \beta b ε − ⇀ ∗ β in L ∞ ( Ω ) L^\infty(\Omega) L ∞ ( Ω ) . These limits satisfy ∥ z ∥ ∞ ≤ 1 \lVert z\rVert_\infty\le1 ∥ z ∥ ∞ ≤ 1 and β ≥ 0 \beta\ge0 β ≥ 0 almost everywhere. By (5.7),
ξ ε : = div z ε = λ ε − V ε − b ε + + b ε − . \xi_\varepsilon:= \operatorname{div}z_\varepsilon=\lambda_\varepsilon-V_\varepsilon-b_\varepsilon^++b_\varepsilon^-. ξ ε := div z ε = λ ε − V ε − b ε + + b ε − .
Since b ε + → 0 b_\varepsilon^+ \to0 b ε + → 0 strongly in L q ( Ω ) L^q(\Omega) L q ( Ω ) , the preceding convergences give
ξ ε ⇀ λ − V + β = : ξ in L q ( Ω ) . \xi_\varepsilon\rightharpoonup\lambda-V+\beta=:\xi\qquad\text{in }L^q(\Omega). ξ ε ⇀ λ − V + β =: ξ in L q ( Ω ) .
For every φ ∈ C 1 ( Ω ‾ ) \varphi\in C^1(\overline{\Omega}) φ ∈ C 1 ( Ω ) , the zero normal trace condition gives ∫ Ω ξ ε φ d x = − ∫ Ω z ε ⋅ ∇ φ d x \int_\Omega\xi_\varepsilon\varphi\,dx=-\int_\Omega z_\varepsilon\cdot\nabla\varphi\,dx ∫ Ω ξ ε φ d x = − ∫ Ω z ε ⋅ ∇ φ d x . Passing to the limit yields ∫ Ω ξ φ d x = − ∫ Ω z ⋅ ∇ φ d x \int_\Omega\xi\varphi\,dx=-\int_\Omega z\cdot\nabla\varphi\,dx ∫ Ω ξ φ d x = − ∫ Ω z ⋅ ∇ φ d x . Thus div z = ξ \operatorname{div}z=\xi div z = ξ in the distributional sense, and z z z has zero normal trace. Moreover, ξ = λ − V + β ∈ L ∞ ( Ω ) \xi=\lambda-V+\beta\in L^\infty(\Omega) ξ = λ − V + β ∈ L ∞ ( Ω ) , so z ∈ Z z\in\mathcal{Z} z ∈ Z and V + div z = λ + β ≥ λ V+\operatorname{div}z=\lambda+\beta\ge\lambda V + div z = λ + β ≥ λ almost everywhere in Ω \Omega Ω . Hence ( f , z , λ ) (f,z,\lambda) ( f , z , λ ) is admissible for the dual problem. The elementary bound − t log t ≤ e − 1 -t\log t\le e^{-1} − t log t ≤ e − 1 for 0 < t ≤ 1 0<t\le1 0 < t ≤ 1 gives
0 ≤ ∫ Ω ρ ε b ε − d x ≤ ε ∣ Ω ∣ e . 0\le\int_\Omega\rho_\varepsilon b_\varepsilon^-\,dx\le\frac{\varepsilon\lvert\Omega\rvert}{e}. 0 ≤ ∫ Ω ρ ε b ε − d x ≤ e ε ∣ Ω ∣ .
Strong L 1 L^1 L 1 convergence of ρ ε \rho_\varepsilon ρ ε and weak-* convergence of b ε − b_\varepsilon^- b ε − imply ∫ ρ ∗ β = 0 \int\rho_\ast\beta=0 ∫ ρ ∗ β = 0 , hence ρ ∗ β = 0 \rho_\ast\beta=0 ρ ∗ β = 0 . Moreover, strong convergence in L r L^r L r and weak convergence in L q L^q L q yield
TV ( ρ ∗ ) ≤ lim inf ε TV ( ρ ε ) = lim ε ∫ Ω ρ ε ξ ε d x = ∫ Ω ρ ∗ div z d x ≤ TV ( ρ ∗ ) . \operatorname{TV}(\rho_\ast)\le\liminf_{\varepsilon}\operatorname{TV}(\rho_\varepsilon)=\lim_{\varepsilon}\int_\Omega\rho_\varepsilon\xi_\varepsilon\,dx=\int_\Omega\rho_\ast\operatorname{div}z\,dx\le\operatorname{TV}(\rho_\ast). TV ( ρ ∗ ) ≤ ε lim inf TV ( ρ ε ) = ε lim ∫ Ω ρ ε ξ ε d x = ∫ Ω ρ ∗ div z d x ≤ TV ( ρ ∗ ) .
All terms are therefore equal. For each i i i , the estimate used to prove the L 1 L^1 L 1 continuity of the transport costs gives
∣ W 2 2 ( μ i , ρ ε d x ) − W 2 2 ( μ i , ρ ∗ d x ) ∣ ≤ 2 D 2 ∥ ρ ε − ρ ∗ ∥ L 1 ( Ω ) → 0. \left|W_2^2(\mu_i,\rho_\varepsilon\,dx)-W_2^2(\mu_i,\rho_\ast\,dx)\right|\le2D^2\lVert\rho_\varepsilon-\rho_\ast\rVert_{L^1(\Omega)}\to0. W 2 2 ( μ i , ρ ε d x ) − W 2 2 ( μ i , ρ ∗ d x ) ≤ 2 D 2 ∥ ρ ε − ρ ∗ ∥ L 1 ( Ω ) → 0.
For each i i i , the transport estimate gives W 2 2 ( μ i , ρ ε d x ) → W 2 2 ( μ i , ρ ∗ d x ) W_2^2(\mu_i,\rho_\varepsilon\,dx)\to W_2^2(\mu_i,\rho_\ast\,dx) W 2 2 ( μ i , ρ ε d x ) → W 2 2 ( μ i , ρ ∗ d x ) . Moreover, uniform convergence of f i , ε f_{i,\varepsilon} f i , ε and stability of the c c c -transform imply ∫ K f i , ε c d μ i → ∫ K f i c d μ i \int_K f_{i,\varepsilon}^c\,d\mu_i\to\int_K f_i^c\,d\mu_i ∫ K f i , ε c d μ i → ∫ K f i c d μ i . Since ρ ε \rho_\varepsilon ρ ε has unit mass,
∣ ∫ Ω f i , ε ρ ε d x − ∫ Ω f i ρ ∗ d x ∣ ≤ ∥ f i , ε − f i ∥ ∞ + ∥ f i ∥ ∞ ∥ ρ ε − ρ ∗ ∥ L 1 ( Ω ) → 0. \left|\int_\Omega f_{i,\varepsilon}\rho_\varepsilon\,dx-\int_\Omega f_i\rho_\ast\,dx\right|\le\lVert f_{i,\varepsilon}-f_i\rVert_\infty+\lVert f_i\rVert_\infty\lVert\rho_\varepsilon-\rho_\ast\rVert_{L^1(\Omega)}\to0. ∫ Ω f i , ε ρ ε d x − ∫ Ω f i ρ ∗ d x ≤ ∥ f i , ε − f i ∥ ∞ + ∥ f i ∥ ∞ ∥ ρ ε − ρ ∗ ∥ L 1 ( Ω ) → 0.
Passing to the limit in the optimality identity
W 2 2 ( μ i , ρ ε d x ) = ∫ K f i , ε c d μ i + ∫ Ω f i , ε ρ ε d x W_2^2(\mu_i,\rho_\varepsilon\,dx)=\int_K f_{i,\varepsilon}^c\,d\mu_i+\int_\Omega f_{i,\varepsilon}\rho_\varepsilon\,dx W 2 2 ( μ i , ρ ε d x ) = ∫ K f i , ε c d μ i + ∫ Ω f i , ε ρ ε d x
therefore yields
W 2 2 ( μ i , ρ ∗ d x ) = ∫ K f i c d μ i + ∫ Ω f i ρ ∗ d x . W_2^2(\mu_i,\rho_\ast\,dx)=\int_K f_i^c\,d\mu_i+\int_\Omega f_i\rho_\ast\,dx. W 2 2 ( μ i , ρ ∗ d x ) = ∫ K f i c d μ i + ∫ Ω f i ρ ∗ d x .
Thus ( f i c , f i ) (f_i^c,f_i) ( f i c , f i ) is optimal for ( μ i , ρ ∗ d x ) (\mu_i,\rho_\ast\,dx) ( μ i , ρ ∗ d x ) . Multiplying V + div z = λ + β V+\operatorname{div}z=\lambda+\beta V + div z = λ + β by ρ ∗ \rho_\ast ρ ∗ now gives J ( ρ ∗ ) = B ( f ) + λ J(\rho_\ast)=B(f)+\lambda J ( ρ ∗ ) = B ( f ) + λ . This proves strong duality and attainment.
We now characterize all primal minimizers. Fix any dual maximizer ( f , z , λ ) (f,z,\lambda) ( f , z , λ ) and set
V : = ∑ i = 1 n f i , β : = V + div z − λ ≥ 0. V := \sum_{i=1}^{n} f_i,\qquad\beta:= V+\operatorname{div}z-\lambda\ge0. V := i = 1 ∑ n f i , β := V + div z − λ ≥ 0.
For every ρ ∈ A \rho\in\mathcal{A} ρ ∈ A , the unit-mass constraint gives
J ( ρ ) − λ − B ( f ) = ∑ i = 1 n [ W 2 2 ( μ i , ρ d x ) − ∫ K f i c d μ i − ∫ Ω f i ρ d x ] + [ ∣ D ρ ∣ ( Ω ) − ∫ Ω ρ div z d x ] + ∫ Ω β ρ d x . \begin{aligned}
J(\rho)-\lambda-B(f)
&=\sum_{i=1}^{n}\left[W_2^2(\mu_i,\rho\,dx)-\int_K f_i^c\,d\mu_i-\int_\Omega f_i\rho\,dx\right]\\
&\quad+\left[|D\rho|(\Omega)-\int_\Omega\rho\operatorname{div}z\,dx\right]+\int_\Omega\beta\rho\,dx.
\end{aligned} J ( ρ ) − λ − B ( f ) = i = 1 ∑ n [ W 2 2 ( μ i , ρ d x ) − ∫ K f i c d μ i − ∫ Ω f i ρ d x ] + [ ∣ D ρ ∣ ( Ω ) − ∫ Ω ρ div z d x ] + ∫ Ω β ρ d x .
The transport terms are nonnegative by Kantorovich duality, the second bracket is nonnegative by (6) , and the last integral is nonnegative because β , ρ ≥ 0 \beta,\rho\ge0 β , ρ ≥ 0 . Since λ + B ( f ) = min A J \lambda+B(f)=\min_{\mathcal{A}}J λ + B ( f ) = min A J , a density ρ \rho ρ is a minimizer if and only if every term on the right vanishes. This is equivalent to (33) together with
∫ Ω ρ div z d x = ∣ D ρ ∣ ( Ω ) , β ρ = 0 almost everywhere . \int_\Omega\rho\operatorname{div}z\,dx=|D\rho|(\Omega),\qquad\beta\rho=0\quad\text{almost everywhere}. ∫ Ω ρ div z d x = ∣ D ρ ∣ ( Ω ) , β ρ = 0 almost everywhere .
Thus these conditions are necessary and sufficient for every primal minimizer. Finally, the estimate ∣ ( z , D ρ ) ∣ ≤ ∣ D ρ ∣ |(z,D\rho)|\le|D\rho| ∣ ( z , D ρ ) ∣ ≤ ∣ D ρ ∣ shows that ∣ D ρ ∣ + ( z , D ρ ) |D\rho|+(z,D\rho) ∣ D ρ ∣ + ( z , D ρ ) is a nonnegative measure. By (6) and the optimality condition,
∣ D ρ ∣ ( Ω ) + ∫ Ω d ( z , D ρ ) = ∣ D ρ ∣ ( Ω ) − ∫ Ω ρ div z d x = 0. |D\rho|(\Omega)+\int_\Omega d(z,D\rho)=|D\rho|(\Omega)-\int_\Omega\rho\operatorname{div}z\,dx=0. ∣ D ρ ∣ ( Ω ) + ∫ Ω d ( z , D ρ ) = ∣ D ρ ∣ ( Ω ) − ∫ Ω ρ div z d x = 0.
Hence − ( z , D ρ ) = ∣ D ρ ∣ -(z,D\rho)=|D\rho| − ( z , D ρ ) = ∣ D ρ ∣ as measures. Conversely, this measure identity implies the integral equality by the same integration-by-parts formula. □
Remark 5.1 (Relation between the two dual formulations). Let ( f , z , λ ) (f,z,\lambda) ( f , z , λ ) satisfy the constraints in (32) , and write V = ∑ i f i V=\sum_i f_i V = ∑ i f i and β = V + div z − λ ≥ 0 \beta=V+\operatorname{div}z-\lambda\ge0 β = V + div z − λ ≥ 0 . For every set E ⊂ Ω E\subset\Omega E ⊂ Ω of finite relative perimeter, applying (6) to 1 E 1_E 1 E gives Per ( E ; Ω ) ≥ ∫ E div z d x \operatorname{Per}(E;\Omega)\ge\int_E\operatorname{div}z\,dx Per ( E ; Ω ) ≥ ∫ E div z d x . Consequently,
Per ( E ; Ω ) + ∫ E ( V − λ ) d x ≥ ∫ E β d x ≥ 0. \operatorname{Per}(E;\Omega)+\int_E(V-\lambda)\,dx\ge\int_E\beta\,dx\ge0. Per ( E ; Ω ) + ∫ E ( V − λ ) d x ≥ ∫ E β d x ≥ 0.
Thus ( f , λ ) (f,\lambda) ( f , λ ) also satisfies the constraints in (16) . Both dual problems attain the same value, namely min ρ ∈ A J ( ρ ) \min_{\rho\in\mathcal{A}}J(\rho) min ρ ∈ A J ( ρ ) .
Bounds, one-dimensional reduction and stability
We next study further properties of the minimizers and of the variational problem. We prove an upper bound for the minimizing density, identify an exact TV-JKO formulation in one dimension, and establish stability under approximation of the prescribed measures.
Maximum principle
In this subsection, assume additionally that Ω \Omega Ω is convex. We show that the minimizing density is bounded above by the largest essential supremum of the prescribed densities. The proof combines the projection argument of [9 Section 4] with the BV estimate of [11 ] .
Theorem 6.1 (Maximum principle). Assume that Ω \Omega Ω is convex and that μ i = ρ i d x \mu_i=\rho_i\,dx μ i = ρ i d x with ρ i ∈ L ∞ ( Ω ) \rho_i\in L^\infty(\Omega) ρ i ∈ L ∞ ( Ω ) for every i = 1 , … , n i=1,\ldots,n i = 1 , … , n . Let ρ ∗ \rho_* ρ ∗ be the unique minimizer of J ( ρ ) J(\rho) J ( ρ ) . Then
∥ ρ ∗ ∥ L ∞ ( Ω ) ≤ max 1 ≤ i ≤ n ∥ ρ i ∥ L ∞ ( Ω ) . (35) \|\rho_*\|_{L^\infty(\Omega)}\le\max_{1\le i\le n}\|\rho_i\|_{L^\infty(\Omega)}. \tag*{(35)} ∥ ρ ∗ ∥ L ∞ ( Ω ) ≤ 1 ≤ i ≤ n max ∥ ρ i ∥ L ∞ ( Ω ) . (35)
Proof. Set M = max i ∥ ρ i ∥ ∞ M=\max_i\lVert\rho_i\rVert_\infty M = max i ∥ ρ i ∥ ∞ and let
K M = { η d x : η ≥ 0 , ∫ Ω η d x = 1 , η ≤ M a.e. } . \mathcal{K}_M=\left\{\eta\,dx:\eta\geq0,\ \int_{\Omega}\eta\,dx=1,\ \eta\leq M\ \text{a.e.}\right\}. K M = { η d x : η ≥ 0 , ∫ Ω η d x = 1 , η ≤ M a.e. } .
Since each ρ i \rho_i ρ i has unit mass, M ≥ ∣ Ω ∣ − 1 M\geq|\Omega|^{-1} M ≥ ∣Ω ∣ − 1 , so the uniform probability measure belongs to K M \mathcal{K}_M K M . Moreover, K M \mathcal{K}_M K M is the set of probability measures on K K K dominated by M 1 Ω d x M1_\Omega\,dx M 1 Ω d x . This domination is preserved under narrow convergence. Thus K M \mathcal{K}_M K M is closed in the compact space P ( K ) \mathcal{P}(K) P ( K ) , and hence narrowly compact. Let ρ ^ d x \widehat{\rho}\,dx ρ d x be a probability measure in K M \mathcal{K}_M K M minimizing the Wasserstein distance to ρ ∗ d x \rho_*\,dx ρ ∗ d x , that is, ρ ^ d x ∈ arg min ν ∈ K M W 2 2 ( ρ ∗ d x , ν ) \widehat{\rho}\,dx\in\operatorname*{arg\,min}_{\nu\in\mathcal{K}_M}W_2^2(\rho_*\,dx,\nu) ρ d x ∈ arg min ν ∈ K M W 2 2 ( ρ ∗ d x , ν ) . Define the convex lower semicontinuous function
h M ( s ) : = { 0 , 0 ≤ s ≤ M , + ∞ , s > M . h_M(s):=
\begin{cases}
0, & 0\leq s\leq M,\\
+\infty, & s>M.
\end{cases} h M ( s ) := { 0 , + ∞ , 0 ≤ s ≤ M , s > M .
The projection ρ ^ \widehat{\rho} ρ minimizes 1 2 W 2 2 ( ρ ∗ d x , η d x ) + ∫ Ω h M ( η ) d x \frac{1}{2}W_2^2(\rho_*\,dx,\eta\,dx)+\int_\Omega h_M(\eta)\,dx 2 1 W 2 2 ( ρ ∗ d x , η d x ) + ∫ Ω h M ( η ) d x over probability densities η \eta η . Since Ω \Omega Ω is convex and ρ ∗ ∈ B V ( Ω ) \rho_*\in BV(\Omega) ρ ∗ ∈ B V ( Ω ) , [11 Theorem 1.1] applies and gives
∣ D ρ ^ ∣ ( Ω ) ≤ ∣ D ρ ∗ ∣ ( Ω ) . (36) |D\widehat{\rho}|(\Omega)\leq|D\rho_*|(\Omega). \tag*{(36)} ∣ D ρ ∣ ( Ω ) ≤ ∣ D ρ ∗ ∣ ( Ω ) . (36)
For completeness, we recall the transport comparison used in [9 Lemma 4.1] . Fix i i i , and let T 0 , T 1 T_0,T_1 T 0 , T 1 be the optimal maps from ρ ∗ d x \rho_*\,dx ρ ∗ d x to μ i \mu_i μ i and ρ ^ d x \widehat{\rho}\,dx ρ d x , respectively. Define the generalized geodesic
ν t = ( ( 1 − t ) T 0 + t T 1 ) # ( ρ ∗ d x ) , 0 ≤ t ≤ 1. \nu_t=\bigl((1-t)T_0+tT_1\bigr)_{\#}(\rho_*\,dx),\qquad0\leq t\leq1. ν t = ( ( 1 − t ) T 0 + t T 1 ) # ( ρ ∗ d x ) , 0 ≤ t ≤ 1.
For every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , the base measure ρ ∗ d x \rho_*\,dx ρ ∗ d x is absolutely continuous, and both endpoint densities are bounded above by M M M . The L ∞ L^\infty L ∞ bound along generalized geodesics [9 Section 4.1] therefore yields
ν t = η t d x , 0 ≤ η t ≤ M a.e. , t ∈ [ 0 , 1 ] . \nu_t=\eta_t\,dx,\qquad0\leq\eta_t\leq M\ \text{a.e.},\qquad t\in[0,1]. ν t = η t d x , 0 ≤ η t ≤ M a.e. , t ∈ [ 0 , 1 ] .
Convexity of Ω \Omega Ω ensures that the interpolated maps take values in Ω \Omega Ω almost everywhere. Consequently, ν t ∈ K M \nu_t\in\mathcal{K}_M ν t ∈ K M . Using the transport plan ( id , ( 1 − t ) T 0 + t T 1 ) # ( ρ ∗ d x ) (\operatorname{id},(1-t)T_0+tT_1)_{\#}(\rho_*\,dx) ( id , ( 1 − t ) T 0 + t T 1 ) # ( ρ ∗ d x ) , we obtain
W 2 2 ( ρ ∗ d x , ν t ) ≤ ∫ Ω ∣ x − ( 1 − t ) T 0 ( x ) − t T 1 ( x ) ∣ 2 ρ ∗ ( x ) d x = ( 1 − t ) W 2 2 ( ρ ∗ d x , μ i ) + t W 2 2 ( ρ ∗ d x , ρ ^ d x ) − t ( 1 − t ) ∫ Ω ∣ T 0 − T 1 ∣ 2 ρ ∗ d x ≤ ( 1 − t ) W 2 2 ( ρ ∗ d x , μ i ) + t W 2 2 ( ρ ∗ d x , ρ ^ d x ) − t ( 1 − t ) W 2 2 ( μ i , ρ ^ d x ) , \begin{aligned}
W_2^2(\rho_*\,dx,\nu_t)
&\leq\int_\Omega\left|x-(1-t)T_0(x)-tT_1(x)\right|^2\rho_*(x)\,dx\\
&=(1-t)W_2^2(\rho_*\,dx,\mu_i)+tW_2^2(\rho_*\,dx,\widehat{\rho}\,dx)\\
&\quad-t(1-t)\int_\Omega|T_0-T_1|^2\rho_*\,dx\\
&\leq(1-t)W_2^2(\rho_*\,dx,\mu_i)+tW_2^2(\rho_*\,dx,\widehat{\rho}\,dx)\\
&\quad-t(1-t)W_2^2(\mu_i,\widehat{\rho}\,dx),
\end{aligned} W 2 2 ( ρ ∗ d x , ν t ) ≤ ∫ Ω ∣ x − ( 1 − t ) T 0 ( x ) − t T 1 ( x ) ∣ 2 ρ ∗ ( x ) d x = ( 1 − t ) W 2 2 ( ρ ∗ d x , μ i ) + t W 2 2 ( ρ ∗ d x , ρ d x ) − t ( 1 − t ) ∫ Ω ∣ T 0 − T 1 ∣ 2 ρ ∗ d x ≤ ( 1 − t ) W 2 2 ( ρ ∗ d x , μ i ) + t W 2 2 ( ρ ∗ d x , ρ d x ) − t ( 1 − t ) W 2 2 ( μ i , ρ d x ) ,
where the last inequality follows from ( T 0 , T 1 ) # ( ρ ∗ d x ) ∈ Π ( μ i , ρ ^ d x ) (T_0,T_1)_{\#}(\rho_*\,dx)\in\Pi(\mu_i,\widehat{\rho}\,dx) ( T 0 , T 1 ) # ( ρ ∗ d x ) ∈ Π ( μ i , ρ d x ) . Projection optimality bounds the left-hand side below by W 2 2 ( ρ ∗ d x , ρ ^ d x ) W_2^2(\rho_*\,dx,\widehat{\rho}\,dx) W 2 2 ( ρ ∗ d x , ρ d x ) . Dividing the resulting inequality by 1 − t 1-t 1 − t and letting t → 1 t\to1 t → 1 yields
W 2 2 ( μ i , ρ ^ d x ) ≤ W 2 2 ( μ i , ρ ∗ d x ) − W 2 2 ( ρ ∗ d x , ρ ^ d x ) . W_2^2(\mu_i,\widehat{\rho}\,dx)\leq W_2^2(\mu_i,\rho_*\,dx)-W_2^2(\rho_*\,dx,\widehat{\rho}\,dx). W 2 2 ( μ i , ρ d x ) ≤ W 2 2 ( μ i , ρ ∗ d x ) − W 2 2 ( ρ ∗ d x , ρ d x ) .
Summing over i i i and using (6.2) gives
J ( ρ ^ ) ≤ J ( ρ ∗ ) − n W 2 2 ( ρ ∗ d x , ρ ^ d x ) . J(\widehat{\rho})\leq J(\rho_*)-nW_2^2(\rho_*\,dx,\widehat{\rho}\,dx). J ( ρ ) ≤ J ( ρ ∗ ) − n W 2 2 ( ρ ∗ d x , ρ d x ) .
Optimality of ρ ∗ \rho_* ρ ∗ forces ρ ∗ = ρ ^ \rho_*=\widehat{\rho} ρ ∗ = ρ a.e., proving (6.1). ∎
Remark 6.1. The convexity assumption and the quadratic cost are used in this projection argument. We do not infer a higher-dimensional minimum principle, or a corresponding maximum principle for every power cost, from this proof.
TV-JKO formulation in one dimension
In one dimension, the representation of Wasserstein distances through generalized inverses of distribution functions yields an exact variance identity. Consequently, the sum of transport costs can be replaced, up to an additive constant, by a single transport cost to the ordinary Wasserstein barycenter.
Proposition 6.1. Let Ω = ( a , b ) ⊂ R \Omega=(a,b)\subset\mathbb{R} Ω = ( a , b ) ⊂ R and μ 1 , … , μ n ∈ P ( [ a , b ] ) \mu_1,\ldots,\mu_n\in\mathcal{P}([a,b]) μ 1 , … , μ n ∈ P ([ a , b ]) . For each i i i , let Q i : ( 0 , 1 ) → [ a , b ] Q_i:(0,1)\to[a,b] Q i : ( 0 , 1 ) → [ a , b ] denote the generalized inverse of the distribution function of μ i \mu_i μ i . Define
Q ‾ = 1 n ∑ i = 1 n Q i , μ ‾ = ( Q ‾ ) # L ( 0 , 1 ) 1 , \overline{Q}=\frac{1}{n}\sum_{i=1}^{n}Q_i,\qquad\overline{\mu}=(\overline{Q})_{\#}\mathcal{L}^{1}_{(0,1)}, Q = n 1 i = 1 ∑ n Q i , μ = ( Q ) # L ( 0 , 1 ) 1 , where L ( 0 , 1 ) 1 \mathcal{L}^{1}_{(0,1)} L ( 0 , 1 ) 1 denotes Lebesgue probability measure on ( 0 , 1 ) (0,1) ( 0 , 1 ) . Then, for any ν ∈ P ( [ a , b ] ) \nu\in\mathcal{P}([a,b]) ν ∈ P ([ a , b ]) , μ ‾ \overline{\mu} μ minimizes ν ↦ ∑ i = 1 n W 2 2 ( μ i , ν ) \nu\mapsto\sum_{i=1}^{n}W_2^2(\mu_i,\nu) ν ↦ ∑ i = 1 n W 2 2 ( μ i , ν ) over P ( [ a , b ] ) \mathcal{P}([a,b]) P ([ a , b ]) , and
∑ i = 1 n W 2 2 ( μ i , ν ) = ∑ i = 1 n W 2 2 ( μ i , μ ‾ ) + n W 2 2 ( μ ‾ , ν ) . (37) \sum_{i=1}^{n}W_2^2(\mu_i,\nu)=\sum_{i=1}^{n}W_2^2(\mu_i,\overline{\mu})+nW_2^2(\overline{\mu},\nu). \tag*{(37)} i = 1 ∑ n W 2 2 ( μ i , ν ) = i = 1 ∑ n W 2 2 ( μ i , μ ) + n W 2 2 ( μ , ν ) . (37) Consequently the TV-regularized problem is exactly, up to an additive constant, the problem with the single reference measure μ ‾ \overline{\mu} μ
min ρ ∈ A { n W 2 2 ( μ ‾ , ρ d x ) + ∣ D ρ ∣ ( Ω ) } . (38) \min_{\rho\in\mathcal{A}}\{nW_2^2(\overline{\mu},\rho\,dx)+|D\rho|(\Omega)\}. \tag*{(38)} ρ ∈ A min { n W 2 2 ( μ , ρ d x ) + ∣ D ρ ∣ ( Ω )} . (38) Its TV-JKO time step is τ = 1 / ( 2 n ) \tau=1/(2n) τ = 1/ ( 2 n ) .
Proof. Since each Q i Q_i Q i is nondecreasing, so is Q ‾ = n − 1 ∑ i = 1 n Q i \overline{Q}=n^{-1}\sum_{i=1}^{n}Q_i Q = n − 1 ∑ i = 1 n Q i . By the definition of μ ‾ \overline{\mu} μ , the generalized inverse Q μ ‾ Q_{\overline{\mu}} Q μ of the distribution function of μ ‾ \overline{\mu} μ satisfies Q μ ‾ = Q ‾ Q_{\overline{\mu}}=\overline{Q} Q μ = Q a.e. on ( 0 , 1 ) (0,1) ( 0 , 1 ) . For any ν ∈ P ( [ a , b ] ) \nu\in\mathcal{P}([a,b]) ν ∈ P ([ a , b ]) , let Q ν Q_\nu Q ν denote the generalized inverse of its distribution function. In one dimension, the coupling ( Q i , Q ν ) # L ( 0 , 1 ) 1 (Q_i,Q_\nu)_{\#}\mathcal{L}^{1}_{(0,1)} ( Q i , Q ν ) # L ( 0 , 1 ) 1 is optimal for the quadratic cost, since both maps are nondecreasing. Consequently,
W 2 2 ( μ i , ν ) = ∫ 0 1 ∣ Q i ( s ) − Q ν ( s ) ∣ 2 d s . W_2^2(\mu_i,\nu)=\int_{0}^{1}|Q_i(s)-Q_\nu(s)|^2\,ds. W 2 2 ( μ i , ν ) = ∫ 0 1 ∣ Q i ( s ) − Q ν ( s ) ∣ 2 d s .
Writing Q i − Q ν = ( Q i − Q ‾ ) + ( Q ‾ − Q ν ) Q_i-Q_\nu=(Q_i-\overline{Q})+(\overline{Q}-Q_\nu) Q i − Q ν = ( Q i − Q ) + ( Q − Q ν ) and using ∑ i = 1 n ( Q i − Q ‾ ) = 0 \sum_{i=1}^{n}(Q_i-\overline{Q})=0 ∑ i = 1 n ( Q i − Q ) = 0 , we obtain
∑ i = 1 n W 2 2 ( μ i , ν ) = ∑ i = 1 n ∫ 0 1 ∣ Q i − Q ‾ ∣ 2 d s + n ∫ 0 1 ∣ Q ‾ − Q ν ∣ 2 d s = ∑ i = 1 n W 2 2 ( μ i , μ ‾ ) + n W 2 2 ( μ ‾ , ν ) . \begin{aligned}
\sum_{i=1}^{n}W_2^2(\mu_i,\nu)=\sum_{i=1}^{n}\int_{0}^{1}|Q_i-\overline{Q}|^2\,ds+n\int_{0}^{1}|\overline{Q}-Q_\nu|^2\,ds \\
&=\sum_{i=1}^{n}W_2^2(\mu_i,\overline{\mu})+nW_2^2(\overline{\mu},\nu).
\end{aligned} i = 1 ∑ n W 2 2 ( μ i , ν ) = i = 1 ∑ n ∫ 0 1 ∣ Q i − Q ∣ 2 d s + n ∫ 0 1 ∣ Q − Q ν ∣ 2 d s = i = 1 ∑ n W 2 2 ( μ i , μ ) + n W 2 2 ( μ , ν ) .
The last term is nonnegative and vanishes precisely when ν = μ ‾ \nu=\overline{\mu} ν = μ . Hence μ ‾ \overline{\mu} μ is the unique minimizer of ν ↦ ∑ i W 2 2 ( μ i , ν ) \nu\mapsto\sum_i W_2^2(\mu_i,\nu) ν ↦ ∑ i W 2 2 ( μ i , ν ) . Taking ν = ρ d x \nu=\rho\,dx ν = ρ d x and adding ∣ D ρ ∣ ( Ω ) |D\rho|(\Omega) ∣ D ρ ∣ ( Ω ) yields
J ( ρ ) = C + n W 2 2 ( μ ‾ , ρ d x ) + ∣ D ρ ∣ ( Ω ) , C : = ∑ i = 1 n W 2 2 ( μ i , μ ‾ ) . J(\rho)=C+nW_2^2(\overline{\mu},\rho\,dx)+|D\rho|(\Omega),\qquad C:=\sum_{i=1}^{n}W_2^2(\mu_i,\overline{\mu}). J ( ρ ) = C + n W 2 2 ( μ , ρ d x ) + ∣ D ρ ∣ ( Ω ) , C := i = 1 ∑ n W 2 2 ( μ i , μ ) .
Since C C C is independent of ρ \rho ρ , the two minimization problems have the same minimizers, and their minimum values differ by C C C . Finally, comparison with the TV-JKO functional 1 2 τ W 2 2 ( μ ‾ , ρ d x ) + ∣ D ρ ∣ ( Ω ) \frac{1}{2\tau}W_2^2(\overline{\mu},\rho\,dx)+|D\rho|(\Omega) 2 τ 1 W 2 2 ( μ , ρ d x ) + ∣ D ρ ∣ ( Ω ) gives τ = 1 / ( 2 n ) \tau=1/(2n) τ = 1/ ( 2 n ) . □ \square □
Remark 6.2. The identity (37) follows from the one-dimensional formula expressing W 2 2 ( μ , ν ) W_2^2(\mu,\nu) W 2 2 ( μ , ν ) as the squared L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) distance between the generalized inverses of their distribution functions. In higher dimensions, optimal couplings generally do not admit such a common representation, so this identity and the resulting reduction to a TV-JKO step do not hold in general.
Stability
We examine how the minimum value and the minimizing densities depend on the prescribed measures. The estimate below controls the change in the minimum value, while BV compactness yields convergence of minimizers. Write J μ J_\mu J μ for (11) with data μ = ( μ 1 , … , μ n ) \mu=(\mu_1,\ldots,\mu_n) μ = ( μ 1 , … , μ n ) , and M ( μ ) = min A J μ M(\mu)=\min_{\mathcal A}J_\mu M ( μ ) = min A J μ . Let W 1 ( μ , ν ) = min γ ∈ Π ( μ , ν ) ∫ ∣ x − y ∣ d γ W_1(\mu,\nu)=\min_{\gamma\in\Pi(\mu,\nu)}\int|x-y|\,d\gamma W 1 ( μ , ν ) = min γ ∈ Π ( μ , ν ) ∫ ∣ x − y ∣ d γ .
Proposition 6.2. For any μ , ν ∈ P ( K ) n \mu,\nu\in\mathcal P(K)^n μ , ν ∈ P ( K ) n ,
∣ M ( μ ) − M ( ν ) ∣ ≤ 2 D ∑ i = 1 n W 1 ( μ i , ν i ) . (39) |M(\mu)-M(\nu)|\le2D\sum_{i=1}^{n}W_1(\mu_i,\nu_i). \tag*{(39)} ∣ M ( μ ) − M ( ν ) ∣ ≤ 2 D i = 1 ∑ n W 1 ( μ i , ν i ) . (39) Suppose that μ i k \mu_i^k μ i k converges narrowly to μ i \mu_i μ i for each i i i , and let ρ k \rho_k ρ k minimize J μ k J_{\mu^k} J μ k . Then ( ρ k ) (\rho_k) ( ρ k ) is relatively compact in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) , and every limit ρ \rho ρ minimizes J μ J_\mu J μ . Moreover, along any subsequence ρ k j → ρ \rho_{k_j}\to\rho ρ k j → ρ in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) ,
∣ D ρ k j ∣ ( Ω ) → ∣ D ρ ∣ ( Ω ) . (40) |D\rho_{k_j}|(\Omega)\to|D\rho|(\Omega). \tag*{(40)} ∣ D ρ k j ∣ ( Ω ) → ∣ D ρ ∣ ( Ω ) . (40) If at least one limiting measure μ i \mu_i μ i is absolutely continuous, the limiting problem has a unique minimizer, and both convergences hold for the whole sequence.
Proof. Fix η ∈ A \eta\in\mathcal A η ∈ A . For each i i i , choose an optimal Kantorovich potential f i f_i f i such that
W 2 2 ( μ i , η d x ) = ∫ K f i c d μ i + ∫ Ω f i η d x . W_2^2(\mu_i,\eta\,dx)=\int_K f_i^c\,d\mu_i+\int_\Omega f_i\eta\,dx. W 2 2 ( μ i , η d x ) = ∫ K f i c d μ i + ∫ Ω f i η d x .
Using f i f_i f i in the dual problem for ( ν i , η d x ) (\nu_i,\eta\,dx) ( ν i , η d x ) and choosing a W 1 W_1 W 1 -optimal coupling γ i ∈ Π ( μ i , ν i ) \gamma_i\in\Pi(\mu_i,\nu_i) γ i ∈ Π ( μ i , ν i ) , we obtain
W 2 2 ( μ i , η d x ) − W 2 2 ( ν i , η d x ) ≤ ∫ K f i c d ( μ i − ν i ) = ∫ K × K ( f i c ( x ) − f i c ( y ) ) d γ i ( x , y ) \begin{aligned}
W_2^2(\mu_i,\eta\,dx)-W_2^2(\nu_i,\eta\,dx)
\le\int_K f_i^c\,d(\mu_i-\nu_i) \\
&= \int_{K\times K}\left(f_i^c(x)-f_i^c(y)\right)\,d\gamma_i(x,y)
\end{aligned} W 2 2 ( μ i , η d x ) − W 2 2 ( ν i , η d x ) ≤ ∫ K f i c d ( μ i − ν i ) = ∫ K × K ( f i c ( x ) − f i c ( y ) ) d γ i ( x , y )
≤ 2 D W 1 ( μ i , ν i ) , \le2D W_1(\mu_i,\nu_i), ≤ 2 D W 1 ( μ i , ν i ) ,
since f i c f_i^c f i c is 2 D 2D 2 D -Lipschitz. Interchanging μ i \mu_i μ i and ν i \nu_i ν i gives the same bound for the absolute difference. Summing over i i i and noting that the total variation terms cancel yields
∣ J μ ( η ) − J ν ( η ) ∣ ≤ 2 D ∑ i = 1 n W 1 ( μ i , ν i ) . |J_\mu(\eta)-J_\nu(\eta)|\le2D\sum_{i=1}^{n}W_1(\mu_i,\nu_i). ∣ J μ ( η ) − J ν ( η ) ∣ ≤ 2 D i = 1 ∑ n W 1 ( μ i , ν i ) .
The bound is independent of η \eta η . Taking infima over η ∈ A \eta\in\mathcal A η ∈ A in the two corresponding one-sided inequalities proves (39) . Let ρ 0 = ∣ Ω ∣ − 1 \rho_0=|\Omega|^{-1} ρ 0 = ∣Ω ∣ − 1 . By the minimality of ρ k \rho_k ρ k ,
∣ D ρ k ∣ ( Ω ) ≤ J μ k ( ρ k ) ≤ J μ k ( ρ 0 ) ≤ n D 2 . |D\rho_k|(\Omega)\le J_{\mu^k}(\rho_k)\le J_{\mu^k}(\rho_0)\le nD^2. ∣ D ρ k ∣ ( Ω ) ≤ J μ k ( ρ k ) ≤ J μ k ( ρ 0 ) ≤ n D 2 .
Together with ∥ ρ k ∥ L 1 ( Ω ) = 1 \|\rho_k\|_{L^1(\Omega)}=1 ∥ ρ k ∥ L 1 ( Ω ) = 1 , this gives a uniform bound in B V ( Ω ) BV(\Omega) B V ( Ω ) . Hence, after passing to a subsequence, ρ k → ρ \rho_k\to\rho ρ k → ρ in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) . Nonnegativity and unit mass pass to the limit, and lower semicontinuity gives ρ ∈ A \rho\in\mathcal A ρ ∈ A . Since K K K is compact, the narrow convergences μ i k → μ i \mu_i^k\to\mu_i μ i k → μ i and ρ k d x → ρ d x \rho_k\,dx\to\rho\,dx ρ k d x → ρ d x also hold in W 2 W_2 W 2 . The triangle inequality therefore gives
∣ W 2 ( μ i k , ρ k d x ) − W 2 ( μ i , ρ d x ) ∣ ≤ W 2 ( μ i k , μ i ) + W 2 ( ρ k d x , ρ d x ) → 0. \left|W_2(\mu_i^k,\rho_k\,dx)-W_2(\mu_i,\rho\,dx)\right|
\le W_2(\mu_i^k,\mu_i)+W_2(\rho_k\,dx,\rho\,dx)\to0. W 2 ( μ i k , ρ k d x ) − W 2 ( μ i , ρ d x ) ≤ W 2 ( μ i k , μ i ) + W 2 ( ρ k d x , ρ d x ) → 0.
Moreover, (6.5) implies M ( μ k ) → M ( μ ) M(\mu^k) \to M(\mu) M ( μ k ) → M ( μ ) . Combining these facts with lower semicontinuity of the total variation, we obtain
M ( μ ) ≤ J μ ( ρ ) ≤ lim inf k → ∞ J μ k ( ρ k ) = M ( μ ) . M(\mu) \le J_\mu(\rho) \le\liminf_{k\to\infty} J_{\mu^k}(\rho_k) = M(\mu). M ( μ ) ≤ J μ ( ρ ) ≤ k → ∞ lim inf J μ k ( ρ k ) = M ( μ ) .
Thus ρ \rho ρ minimizes J μ J_\mu J μ . Consequently,
∣ D ρ k ∣ ( Ω ) = M ( μ k ) − ∑ i = 1 n W 2 2 ( μ i k , ρ k d x ) → M ( μ ) − ∑ i = 1 n W 2 2 ( μ i , ρ d x ) = ∣ D ρ ∣ ( Ω ) . |D\rho_k|(\Omega) = M(\mu^k) - \sum_{i=1}^{n} W_2^2(\mu_i^k,\rho_k\,dx) \to M(\mu) - \sum_{i=1}^{n} W_2^2(\mu_i,\rho\,dx) = |D\rho|(\Omega). ∣ D ρ k ∣ ( Ω ) = M ( μ k ) − i = 1 ∑ n W 2 2 ( μ i k , ρ k d x ) → M ( μ ) − i = 1 ∑ n W 2 2 ( μ i , ρ d x ) = ∣ D ρ ∣ ( Ω ) .
The same argument applies to any L 1 L^1 L 1 -convergent subsequence. If the limiting minimizer is unique, relative compactness implies that the whole sequence converges to it in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) , and the preceding identity gives convergence of the total variations. □
Extension to p p p -Wasserstein costs for p > 1 p>1 p > 1
Brizzi and Portinale [7 ] establish L q L^q L q estimates for unregularized p p p -Wasserstein barycenters under geometric assumptions on the supports, and show that such integrability need not be preserved in general when p ≠ 2 p \ne2 p = 2 . Here we consider the problem with a total variation penalty and extend the existence, uniqueness, and perimeter duality results to the cost c p ( x , y ) = ∣ x − y ∣ p c_p(x,y)=|x-y|^p c p ( x , y ) = ∣ x − y ∣ p , where p > 1 p>1 p > 1 . The corresponding optimality conditions also yield a transport-map representation when a prescribed measure is absolutely continuous. We state these conclusions together and explain the changes required in the quadratic-cost arguments.
For p > 1 p>1 p > 1 , consider
J p ( ρ ) = ∑ i = 1 n W p p ( μ i , ρ d x ) + ∣ D ρ ∣ ( Ω ) , ρ ∈ A . J_p(\rho)=\sum_{i=1}^{n}W_p^p(\mu_i,\rho\,dx)+|D\rho|(\Omega),\qquad\rho\in\mathcal{A}. J p ( ρ ) = i = 1 ∑ n W p p ( μ i , ρ d x ) + ∣ D ρ ∣ ( Ω ) , ρ ∈ A .
Recall that P p ( K ) = P ( K ) \mathcal{P}_p(K)=\mathcal{P}(K) P p ( K ) = P ( K ) because K K K is compact.
Theorem 7.1. The problem min ρ ∈ A J p ( ρ ) \min_{\rho\in\mathcal{A}}J_p(\rho) min ρ ∈ A J p ( ρ ) has a solution for arbitrary μ i ∈ P p ( K ) \mu_i\in\mathcal{P}_p(K) μ i ∈ P p ( K ) . The minimizer is unique if at least one μ i \mu_i μ i is absolutely continuous with respect to Lebesgue measure. It also has the attained dual representation
min ρ ∈ A J p ( ρ ) = max f ∈ K p n { ∑ i = 1 n ∫ K f i c p d μ i + m ( ∑ i f i ) } . (41) \min_{\rho\in\mathcal{A}}J_p(\rho)=\max_{f\in\mathcal{K}_p^n}\left\{\sum_{i=1}^{n}\int_K f_i^{c_p}\,d\mu_i+m\left(\sum_i f_i\right)\right\}. \tag*{(41)} ρ ∈ A min J p ( ρ ) = f ∈ K p n max { i = 1 ∑ n ∫ K f i c p d μ i + m ( i ∑ f i ) } . (41) where K p = { f ∈ C ( K ) : f ( x ∗ ) = 0 , Lip ( f ) ≤ p D p − 1 } \mathcal{K}_p=\{f\in C(K):f(x_*)=0,\ \operatorname{Lip}(f)\le pD^{p-1}\} K p = { f ∈ C ( K ) : f ( x ∗ ) = 0 , Lip ( f ) ≤ p D p − 1 } . For any dual maximizer f ∈ K p n f\in\mathcal{K}_p^n f ∈ K p n , set V : = ∑ i = 1 n f i V:=\sum_{i=1}^{n}f_i V := ∑ i = 1 n f i , λ : = m ( V ) \lambda:=m(V) λ := m ( V ) . The optimality characterization and superlevel-set identities established for the quadratic cost hold with this value of λ \lambda λ , with W 2 2 W_2^2 W 2 2 replaced by W p p W_p^p W p p and c c c replaced by c p c_p c p . If μ j \mu_j μ j is absolutely continuous for some j ∈ { 1 , … , n } j\in\{1,\ldots,n\} j ∈ { 1 , … , n } , let ρ \rho ρ denote the unique minimizer of J p J_p J p . For every dual maximizer f ∈ K p n f\in\mathcal{K}_p^n f ∈ K p n ,
ρ d x = ( T j ) # μ j , T j ( x ) = x − ( ∇ h p ) − 1 ( ∇ f j c p ( x ) ) for μ j -a.e. x , (42) \rho\,dx=(T_j)_\#\mu_j,\qquad T_j(x)=x-(\nabla h_p)^{-1}\left(\nabla f_j^{c_p}(x)\right)\quad\text{for }\mu_j\text{-a.e. }x, \tag*{(42)} ρ d x = ( T j ) # μ j , T j ( x ) = x − ( ∇ h p ) − 1 ( ∇ f j c p ( x ) ) for μ j -a.e. x , (42) where h p ( v ) = ∣ v ∣ p h_p(v)=|v|^p h p ( v ) = ∣ v ∣ p . Here ( ∇ h p ) − 1 ( 0 ) = 0 (\nabla h_p)^{-1}(0)=0 ( ∇ h p ) − 1 ( 0 ) = 0 , and
( ∇ h p ) − 1 ( w ) = p − 1 / ( p − 1 ) ∣ w ∣ ( 2 − p ) / ( p − 1 ) w , w ∈ R d ∖ { 0 } . (\nabla h_p)^{-1}(w)=p^{-1/(p-1)}|w|^{(2-p)/(p-1)}w,\qquad w\in\mathbb{R}^d\setminus\{0\}. ( ∇ h p ) − 1 ( w ) = p − 1/ ( p − 1 ) ∣ w ∣ ( 2 − p ) / ( p − 1 ) w , w ∈ R d ∖ { 0 } .
Proof. The constant competitor gives J p ( ρ 0 ) ≤ n D p J_p(\rho_0) \le nD^p J p ( ρ 0 ) ≤ n D p . Therefore every bounded-energy minimizing sequence has a subsequence converging in L 1 L^1 L 1 to an element of A \mathcal{A} A . On the compact space K K K , this implies convergence in W p W_p W p . Lower semicontinuity of TV proves existence. Suppose that μ j \mu_j μ j is absolutely continuous. Since h p ( v ) = ∣ v ∣ p h_p(v)=|v|^p h p ( v ) = ∣ v ∣ p is differentiable and strictly convex for p > 1 p>1 p > 1 , the optimal plan from μ j \mu_j μ j to any ν ∈ P ( K ) \nu\in\mathcal{P}(K) ν ∈ P ( K ) is unique and induced by a map; see [24 Theorem 1.17 and Remark 1.18] . The argument in Theorem 3.1, based on convex combinations of optimal plans, therefore shows that ν ↦ W p p ( μ j , ν ) \nu\mapsto W_p^p(\mu_j,\nu) ν ↦ W p p ( μ j , ν ) is strictly convex under linear interpolation of measures. The remaining transport terms and the total variation are convex, so J p J_p J p has a unique minimizer.
For x , y , y ′ ∈ K x,y,y'\in K x , y , y ′ ∈ K ,
∣ ∣ x − y ∣ p − ∣ x − y ′ ∣ p ∣ ≤ p D p − 1 ∣ y − y ′ ∣ . \lvert|x-y|^p-|x-y'|^p\rvert\le pD^{p-1}|y-y'|. ∣ ∣ x − y ∣ p − ∣ x − y ′ ∣ p ∣ ≤ p D p − 1 ∣ y − y ′ ∣.
Thus every c p c_p c p -transform is p D p − 1 pD^{p-1} p D p − 1 -Lipschitz. Replacing a potential f f f by f c p c p f^{c_pc_p} f c p c p does not decrease the Kantorovich objective, and subtracting f c p c p ( x ∗ ) f^{c_pc_p}(x_*) f c p c p ( x ∗ ) leaves it unchanged. Hence the dual potentials may be restricted to K p \mathcal{K}_p K p , which is compact in C ( K ) C(K) C ( K ) and convex. Set
L p ( ρ , f ) = ∣ D ρ ∣ ( Ω ) + ∫ Ω ρ ∑ i = 1 n f i d x + ∑ i = 1 n ∫ K f i c p d μ i . L_p(\rho,f)=|D\rho|(\Omega)+\int_\Omega\rho\sum_{i=1}^{n}f_i\,dx+\sum_{i=1}^{n}\int_K f_i^{c_p}\,d\mu_i. L p ( ρ , f ) = ∣ D ρ ∣ ( Ω ) + ∫ Ω ρ i = 1 ∑ n f i d x + i = 1 ∑ n ∫ K f i c p d μ i .
Kantorovich duality yields J p ( ρ ) = max f ∈ K p n L p ( ρ , f ) J_p(\rho)=\max_{f\in\mathcal{K}_p^n}L_p(\rho,f) J p ( ρ ) = max f ∈ K p n L p ( ρ , f ) . For fixed f f f , L p L_p L p is convex and lower semicontinuous in ρ \rho ρ with respect to L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) convergence. For fixed ρ \rho ρ , it is concave and continuous in f f f , using ∥ f c p − g c p ∥ ∞ ≤ ∥ f − g ∥ ∞ \lVert f^{c_p}-g^{c_p}\rVert_\infty\le\lVert f-g\rVert_\infty ∥ f c p − g c p ∥ ∞ ≤ ∥ f − g ∥ ∞ . Sion’s minimax theorem therefore gives
min ρ ∈ A J p ( ρ ) = sup f ∈ K p n inf ρ ∈ A L p ( ρ , f ) = sup f ∈ K p n { ∑ i = 1 n ∫ K f i c p d μ i + m ( ∑ i = 1 n f i ) } . \begin{aligned}
\min_{\rho\in\mathcal{A}}J_p(\rho)=\sup_{f\in\mathcal{K}_p^n}\inf_{\rho\in\mathcal{A}}L_p(\rho,f) \\
&=\sup_{f\in\mathcal{K}_p^n}\left\{\sum_{i=1}^{n}\int_K f_i^{c_p}\,d\mu_i+m\left(\sum_{i=1}^{n}f_i\right)\right\}.
\end{aligned} ρ ∈ A min J p ( ρ ) = f ∈ K p n sup ρ ∈ A inf L p ( ρ , f ) = f ∈ K p n sup { i = 1 ∑ n ∫ K f i c p d μ i + m ( i = 1 ∑ n f i ) } .
The objective is continuous by the preceding estimate and the Lipschitz continuity of m m m . Compactness of K p n \mathcal{K}_p^n K p n ensures that the supremum is attained, proving (41) .
Let ρ \rho ρ be a primal minimizer and f f f a dual maximizer, and set V = ∑ i = 1 n f i V=\sum_{i=1}^{n}f_i V = ∑ i = 1 n f i . The duality identity gives
0 = ∑ i = 1 n [ W p p ( μ i , ρ d x ) − ∫ K f i c p d μ i − ∫ Ω f i ρ d x ] 0=\sum_{i=1}^{n}\left[W_p^p(\mu_i,\rho\,dx)-\int_K f_i^{c_p}\,d\mu_i-\int_\Omega f_i\rho\,dx\right] 0 = i = 1 ∑ n [ W p p ( μ i , ρ d x ) − ∫ K f i c p d μ i − ∫ Ω f i ρ d x ]
+ ∣ D ρ ∣ ( Ω ) + ∫ Ω V ρ d x − m ( V ) . \qquad+|D\rho|(\Omega)+\int_\Omega V\rho\,dx-m(V). + ∣ D ρ ∣ ( Ω ) + ∫ Ω V ρ d x − m ( V ) .
Each term is nonnegative by Kantorovich duality and the definition of m ( V ) m(V) m ( V ) , so every term vanishes. Conversely, these equalities make the primal objective equal to the attained dual value, and hence imply optimality of ρ \rho ρ . This gives the transport optimality identities and
∣ D ρ ∣ ( Ω ) + ∫ Ω V ρ d x = m ( V ) . |D\rho|(\Omega)+\int_\Omega V\rho\,dx=m(V). ∣ D ρ ∣ ( Ω ) + ∫ Ω V ρ d x = m ( V ) .
Writing E t = { ρ > t } E_t=\{\rho>t\} E t = { ρ > t } , the coarea and layer-cake formulas imply
∫ 0 ∞ [ Per ( E t ; Ω ) + ∫ E t V d x − m ( V ) ∣ E t ∣ ] d t = 0. \int_0^\infty\left[\operatorname{Per}(E_t;\Omega)+\int_{E_t}V\,dx-m(V)|E_t|\right]\,dt=0. ∫ 0 ∞ [ Per ( E t ; Ω ) + ∫ E t V d x − m ( V ) ∣ E t ∣ ] d t = 0.
The integrand is nonnegative by Lemma 3.1. Hence, for almost every t > 0 t > 0 t > 0 with ∣ E t ∣ > 0 |E_t| > 0 ∣ E t ∣ > 0 ,
Per ( E t ; Ω ) + ∫ E t V d x ∣ E t ∣ = m ( V ) . \frac{\operatorname{Per}(E_t;\Omega)+\int_{E_t} V\,dx}{|E_t|}=m(V). ∣ E t ∣ Per ( E t ; Ω ) + ∫ E t V d x = m ( V ) .
Finally, suppose that μ j \mu_j μ j is absolutely continuous, and let T j T_j T j be the optimal transport map from μ j \mu_j μ j to ρ d x \rho\,dx ρ d x . The transport optimality identity implies
f j c p ( x ) + f j ( T j ( x ) ) = h p ( x − T j ( x ) ) for μ j -a.e. x . f_j^{c_p}(x)+f_j(T_j(x))=h_p(x-T_j(x)) \quad\text{for }\mu_j\text{-a.e. }x. f j c p ( x ) + f j ( T j ( x )) = h p ( x − T j ( x )) for μ j -a.e. x .
At any such interior point where f j c p f_j^{c_p} f j c p is differentiable, the function
x ′ ⟼ h p ( x ′ − T j ( x ) ) − f j ( T j ( x ) ) − f j c p ( x ′ ) x' \longmapsto h_p(x'-T_j(x))-f_j(T_j(x))-f_j^{c_p}(x') x ′ ⟼ h p ( x ′ − T j ( x )) − f j ( T j ( x )) − f j c p ( x ′ )
is nonnegative and vanishes at x ′ = x x'=x x ′ = x . Therefore,
∇ f j c p ( x ) = ∇ h p ( x − T j ( x ) ) . \nabla f_j^{c_p}(x)=\nabla h_p(x-T_j(x)). ∇ f j c p ( x ) = ∇ h p ( x − T j ( x )) .
Since f j c p f_j^{c_p} f j c p is Lipschitz and μ j \mu_j μ j is absolutely continuous, this holds for μ j \mu_j μ j -almost every x x x . Applying ( ∇ h p ) − 1 (\nabla h_p)^{-1} ( ∇ h p ) − 1 proves (42) .
Remark 7.1 (The case p = 1 p=1 p = 1 ). Existence and the attained dual remain valid for p = 1 p=1 p = 1 , using the normalized 1-Lipschitz class of potentials. The strict-convexity argument and the inverse-gradient formula above do not extend to this case. Their failure alone does not imply nonuniqueness of the TV-regularized barycenter.
Acknowledgements
This work was supported by the National Natural Science Foundation of China (Grant Nos. 12301661 and 12431018).
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, Nanjing University, Nanjing 210093, China Email address: nieziewei@nju.edu.cn