betaFeedback

WorksMathematics

Existence, Equivalence, and Convergence of Unbalanced Optimal Transport

Abstract

We study an unbalanced optimal transport problem with a spatially independent signed source. The target density is continuous on a bounded domain or continuous and integrable on Rn\mathbb R^n. We establish sequential weak lower semicontinuity at the critical Sobolev regularity W1,nW^{1,n} and existence of minimizers under prescribed distortion and Jacobian mass bounds, provided a finite-energy admissible pair exists. For domain-preserving, uniformly bi-Lipschitz affine flows, we prove an energy-preserving static-dynamic correspondence and stability of the reconstruction. Within this class, sequential ΓΓ-convergence identifies the strong source-penalty limit as transport minimization among pairs of minimum source cost. This yields optimal-cost asymptotics and subsequential convergence of minimizers and their reconstructed paths. When a constant source is admissible, optimal sources converge strongly in L2L^2 to the constant determined by the endpoint mass difference. We establish local constant-source feasibility for Hölder endpoint densities near a fixed reference pair and give a quantitative obstruction showing that constant-source feasibility may fail even when admissible pairs exist.

Introduction

Optimal transport provides a variational framework for comparing and redistributing mass. In its classical formulation, the source and target measures have the same total mass. Unbalanced optimal transport relaxes this requirement by allowing mass creation and removal.

Let Ω⊂Rn\Omega\subset\mathbb{R}^n be a bounded domain, let ρ0\rho_0 and ρ1\rho_1 denote the nonnegative initial and target densities, and let α>0\alpha> 0. For a map M∈W1,n(Ω;Rn)M \in W^{1,n}(\Omega;\mathbb{R}^n) and a spatially independent source f∈L2(0,1)f \in L^2(0,1), the static energy is formally given by

Eα(M,f)=∫Ω∣M(x)−x∣2ρ0(x) dx+1α∫01∣f(t)∣2 dt+∫01∫0tf(s)∫Ω∣M(x)−x∣2det⁡(s∇M(x)+(1−s)I) dx ds dt.(1)E_\alpha(M,f)=\int_\Omega|M(x)-x|^2\rho_0(x)\,dx+\frac{1}{\alpha}\int_0^1|f(t)|^2\,dt+\int_0^1\int_0^t f(s)\int_\Omega|M(x)-x|^2\det\bigl(s\nabla M(x)+(1-s)I\bigr)\,dx\,ds\,dt. \tag*{(1)}

where II is the identity matrix. The first term is the displacement cost of the initial mass, the second penalizes the source, and the last accounts for its contribution along the affine interpolation from the identity to MM. The map and the source satisfy the endpoint constraint

ρ1(M(x))det⁡∇M(x)=ρ0(x)+∫01f(t)Jt[M](x) dt.(2)\rho_1(M(x))\det\nabla M(x)=\rho_0(x)+\int_0^1 f(t)J_t[M](x)\,dt. \tag*{(2)}

Its distributional interpretation and the remaining admissibility conditions are specified below.

Write Xt[M]:=(1−t)Id⁡+tMX_t[M] := (1-t)\operatorname{Id} + tM, Jt[M]:=det⁡((1−t)I+t∇M)J_t[M] := \det\left((1-t)I + t\nabla M\right) and define the intermediate state

Rt[M,f](x):=ρ0(x)+∫0tf(s)Js[M](x) ds.R_t[M,f](x) := \rho_0(x) + \int_0^t f(s)J_s[M](x)\,ds.

For pairs with nonnegative intermediate states, define

G(M,f):=∫01∫Ω∣M(x)−x∣2Rt[M,f](x) dx dt∈[0,+∞]G(M,f) := \int_0^1 \int_\Omega|M(x)-x|^2 R_t[M,f](x)\,dx\,dt \in[0,+\infty]

and

Eα(M,f):=G(M,f)+1α∥f∥L2(0,1)2.E_\alpha(M,f) := G(M,f) + \frac{1}{\alpha}\|f\|_{L^2(0,1)}^2.

This definition agrees with (1) for bounded-range maps and whenever the terms in that expression are absolutely integrable. We study existence for this static problem, its correspondence with dynamic transport under prescribed affine-flow conditions, and the limits selected by strong source penalization. We also give sufficient conditions and explicit examples verifying the required feasibility assumptions. We first review the related literature and then describe our approach and main results.

Background and related work

The classical transport problem originates with Monge [24]. Kantorovich [18] introduced its relaxation to transport plans, while Benamou and Brenier [2] formulated quadratic optimal transport through a continuity equation and a kinetic action. Existence of optimal maps has been studied under various assumptions on the cost and the endpoint measures [11, 16, 4, 1, 29]. For background on classical optimal transport, we refer the reader to [30, 31].

Several extensions of optimal transport have been developed for finite nonnegative measures with unequal total masses. One approach to unbalanced optimal transport extends the Benamou-Brenier formulation by introducing a source into the continuity equation and penalizing mass creation and removal. Chizat et al. [5] introduced the Wasserstein-Fisher-Rao metric, which combines the quadratic Wasserstein action for spatial transport with the Fisher-Rao action for mass variation. In [6], they studied a broader class of dynamic transport costs defined by convex, positively one-homogeneous action densities and established equivalent static Kantorovich formulations in terms of semi-couplings. Static approaches include Kantorovich-type problems over transport plans and Monge-type problems involving transport maps and mass scaling factors. Guittet [17] studied extended Kantorovich norms with a fixed cost per unit of mass created or removed. Piccoli and Rossi [26] introduced generalized Wasserstein distances combining WpW_p transport with total variation penalties for mass variation, and subsequently established their dynamic characterization and duality properties in [27]. Liero, Mielke, and Savaré [22] developed the entropy-transport framework as a Kantorovich-type problem with entropy penalties on the marginals. Savaré and Sodini [28] obtained formulations over homogeneous couplings through convex relaxation and investigated the corresponding Monge formulations. For the Hellinger-Kantorovich distance, Liero, Mielke, and Savaré [21] established sufficient conditions for the solvability of the Monge problem and constructed transport-growth flows describing geodesics. Gallouët, Ghezzi, and Vialard [14] studied regularity in the static entropy-transport framework and its associated Monge formulation, and proved static-dynamic equivalence for the Wasserstein-Fisher-Rao metric on closed Riemannian manifolds.

The present problem is motivated by Gangbo et al. [15]. They introduced a spatially independent additive source with a quadratic penalty into the dynamic transport problem and derived an associated Monge formulation through Lagrangian coordinates. This leads to the energy (1.1) and the endpoint relation (1.2). A generalization allowing spatially dependent sources was studied in [20]. Gangbo et al. [15] approach the problem through its dynamic formulation and derive an associated Monge formulation in Lagrangian coordinates. We take the static formulation as our starting point and establish existence of minimizers under prescribed geometric constraints. For domain-preserving, uniformly bi-Lipschitz affine flows, we then prove equivalence with the corresponding restriction of their dynamic problem, including equality of minimum values and a correspondence between minimizers.

Outline of the approach

The direct analysis of the static problem requires compactness of minimizing sequences and stability of the nonlinear constraints under weak convergence. The arguments differ according to whether the maps have a prescribed bounded range.

For bounded targets, the main difficulty is to preserve the nonlinear endpoint constraint and establish lower semicontinuity under weak convergence of both maps and sources. At the critical regularity W1,nW^{1,n}, distributional continuity of Jacobians does not directly imply convergence of the weighted energy. We combine composition identities with weak continuity of minors to obtain convergence of localized weighted Jacobians. Uniform approximation of the continuous target density also allows passage to the limit in the endpoint constraint. The polynomial dependence of Jt[M]J_t[M] on time provides the uniform convergence needed to handle its coupling with weakly convergent sources. Nonnegativity of the intermediate states then permits a cutoff argument yielding global lower semicontinuity, even when the sources change sign. The distortion and Jacobian mass bounds control the gradients, while the bounded range controls the maps themselves. These estimates provide the compactness required by the direct method.

For whole-space targets, the absence of a prescribed bounded range creates two additional difficulties. The weights depending on the maps are no longer uniformly bounded, and gradient bounds alone do not control translations. To address the first issue, local Hölder estimates under the distortion bound yield locally uniform convergence along weakly convergent sequences. This preserves the weighted endpoint constraint and allows the preceding energy argument to be applied locally. To obtain a full Sobolev bound, we use positive initial mass to find a set of uniformly positive measure on which the intermediate states remain bounded below for a short time. The transport energy then controls the displacement on this set. Combined with Poincaré’s inequality and the gradient bound, this controls the mean of the map and establishes coercivity.

For the static-dynamic correspondence, we restrict attention to domain-preserving, uniformly bi-Lipschitz affine flows. The change-of-variables formula reconstructs dynamic densities and velocities from static pairs and preserves the energy. Conversely, testing the continuity equation with functions transported by the inverse flow recovers the static endpoint constraint. Uniform control of the inverse flows, together with weak continuity of minors and the polynomial structure of the interpolation Jacobians, also establishes stability of the reconstruction. The source-penalty limit is studied within this same geometric class. After subtracting the minimum source cost divided by α\alpha, compactness and lower semicontinuity give the Γ\Gamma-liminf inequality, while constant sequences provide recovery sequences. Comparison with admissible pairs of minimum source cost identifies the limiting transport problem and yields convergence of minimum values and subsequential convergence of minimizers. Convergence of the source norms upgrades their weak convergence to strong convergence in L2(0,1)L^2(0,1). The reconstruction stability then transfers these limits to the corresponding dynamic paths. When a constant source is admissible, the limiting source is the constant determined by the endpoint mass difference.

Finally, we investigate the existence of admissible pairs with constant source. The Jacobian bounds give a necessary condition for constant-source admissibility. For densities close to a fixed reference pair, a solution operator for the divergence equation and Schauder’s fixed point theorem solve the nonlinear endpoint constraint. This construction produces constant-source pairs satisfying the prescribed distortion and flow bounds, so the existence, equivalence, and convergence results apply without a separate nonemptiness assumption in this local setting.

Main results

Throughout this paper, let n≥2n \ge2, let Ω⊂Rn\Omega\subset\mathbb{R}^n be a bounded open set with Lipschitz boundary, and let 0≤ρ0∈L1(Ω)0 \le\rho_0 \in L^1(\Omega). We equip Y:=(W1,n(Ω;Rn)∩L∞(Ω;Rn))×L2(0,1)Y := (W^{1,n}(\Omega;\mathbb{R}^n) \cap L^\infty(\Omega;\mathbb{R}^n)) \times L^2(0,1) with the topology τ\tau. Here (Mj,fj)→τ(M,f)(M_j,f_j) \xrightarrow{\tau} (M,f) means that Mj⇀MM_j \rightharpoonup M in W1,n(Ω;Rn)W^{1,n}(\Omega;\mathbb{R}^n), Mj⇀∗MM_j \rightharpoonup^* M in L∞(Ω;Rn)L^\infty(\Omega;\mathbb{R}^n), and fj⇀ff_j \rightharpoonup f in L2(0,1)L^2(0,1). We equip Y1:=W1,n(Ω;Rn)×L2(0,1)Y_1 := W^{1,n}(\Omega;\mathbb{R}^n) \times L^2(0,1) with the product weak topology. We use the following four sets of assumptions and specify the corresponding admissible classes.

(H1) Bounded targets. Assume that Ω\Omega is convex and 0≤ρ1∈C(Ω‾)0 \le\rho_1 \in C(\overline{\Omega}). Set

X(Ω;Ω‾):={M∈W1,n(Ω;Rn)∩L∞(Ω;Rn):M(x)∈Ω‾ a.e.}.X(\Omega;\overline{\Omega}) := \{M \in W^{1,n}(\Omega;\mathbb{R}^n) \cap L^\infty(\Omega;\mathbb{R}^n) : M(x) \in\overline{\Omega}\ \text{a.e.}\}.

Let Vρ1V^{\rho_1} consist of pairs (M,f)∈X(Ω;Ω‾)×L2(0,1)(M,f) \in X(\Omega;\overline{\Omega}) \times L^2(0,1) satisfying Jt[M]≥0J_t[M] \ge0 and Rt[M,f]≥0R_t[M,f] \ge0 a.e. in Ω\Omega and for every fixed t∈[0,1]t \in[0,1], and ρ1(M)det⁡∇M=R1[M,f]\rho_1(M) \det\nabla M = R_1[M,f] in D′(Ω)\mathcal{D}'(\Omega). For K>0K > 0, define

VKρ1:={(M,f)∈Vρ1:∣∇M∣n≤Kdet⁡∇M a.e. in Ω,∫Ωdet⁡∇M dx≤∣Ω∣}.V_K^{\rho_1} := \left\{(M,f) \in V^{\rho_1} : \begin{array}{l} |\nabla M|^n \le K\det\nabla M\ \text{a.e. in }\Omega,\\ \int_\Omega\det\nabla M\,dx \le|\Omega| \end{array} \right\}.

(H2) Whole-space targets. Assume that Ω\Omega is connected, 0≤ρ1∈C(Rn)∩L1(Rn)0 \le\rho_1 \in C(\mathbb{R}^n) \cap L^1(\mathbb{R}^n), and ∫Ωρ0 dx>0\int_\Omega\rho_0\,dx > 0.

Let V~ρ1\widetilde{V}^{\rho_1} satisfy the same source, nonnegativity, and endpoint constraints as Vρ1V^{\rho_1}, with the map requirement replaced by M∈W1,n(Ω;Rn)M \in W^{1,n}(\Omega;\mathbb{R}^n). For K,Λ>0K,\Lambda> 0, define

VK,Λρ1:={(M,f)∈V~ρ1:∣∇M∣n≤Kdet⁡∇M a.e. in Ω,∫Ωdet⁡∇M dx≤Λ}.V_{K,\Lambda}^{\rho_1} := \left\{(M,f) \in\widetilde{V}^{\rho_1} : \begin{array}{l} |\nabla M|^n \le K\det\nabla M\ \text{a.e. in }\Omega,\\ \int_\Omega\det\nabla M\,dx \le\Lambda \end{array} \right\}.

The distributional endpoint constraint includes the requirement ρ1(M)det⁡∇M∈Lloc1(Ω)\rho_1(M)\det\nabla M \in L^1_{\mathrm{loc}}(\Omega).

(H3) Static-dynamic setting. Assume that Ω\Omega is convex and 0≤ρ1∈C(Ω‾)0 \le\rho_1 \in C(\overline{\Omega}). For L≥1L \ge1, let ALρ1A_L^{\rho_1} consist of pairs (M,f)∈Vρ1(M,f) \in V^{\rho_1} with M∈C(Ω‾;Ω‾)M \in C(\overline{\Omega};\overline{\Omega}) satisfying

Xt[M](Ω‾)=Ω‾,X_t[M](\overline{\Omega}) = \overline{\Omega},
L−1∣x−y∣≤∣Xt[M](x)−Xt[M](y)∣≤L∣x−y∣(3)L^{-1}|x-y| \le|X_t[M](x)-X_t[M](y)| \le L|x-y| \tag*{(3)}

for every x,y∈Ω‾x,y \in\overline{\Omega} and t∈[0,1]t \in[0,1].

References

  1. [1]L. Ambrosio, B. Kirchheim, A. Pratelli, Existence of optimal transport maps for crystalline norms, Duke Mathematical Journal, 125 (2004), 207-241.DOI
  2. [2]J.D. Benamou, Y. Brenier, A computational fluid mechanics solution to the Monge-Kantorovich mass transfer problem, Numer. Math., 84 (2000), 375-393.DOI
  3. [4]L.A. Caffarelli, M. Feldman, R.J. McCann, Constructing optimal maps for Monge’s transport problem as a limit of strictly convex costs, J. Amer. Math. Soc., 15 (2002), 1-26.DOI
  4. [5]L. Chizat, G. Peyré, B. Schmitzer, F.-X. Vialard, An interpolating distance between optimal transport and Fisher-Rao metrics, Found. Comput. Math., 18 (2018), 1-44.
  5. [6]L. Chizat, G. Peyré, B. Schmitzer, F.-X. Vialard, Unbalanced optimal transport: Dynamic and Kantorovich formulations, J. Funct. Anal., 274 (2018), 3090-3123.arxiv.org/abs/1508.05216
  6. [7]B. Dacorogna, Direct methods in the calculus of variations, Applied Mathematical Sciences, Springer, Berlin, 1989.
  7. [8]B. Dacorogna, J. Moser, On a partial differential equation involving the Jacobian determinant, Ann. Inst. H. Poincaré Anal. Non Linéaire, 7 (1990), 1-26.DOI
  8. [9]G. Dal Maso, An introduction to Γ-convergence. Progress in Nonlinear Differential Equations and their Applications, Birkhäuser Boston, MA, 1993.DOI
  9. [11]L.C. Evans, W. Gangbo, Differential equations methods for the Monge-Kantorovich mass transfer problem, Mem. Amer. Math. Soc., 137 (1999), 1-66.DOI
  10. [14]T. Gallouët, R. Ghezzi, F.-X. Vialard, Regularity theory and geometry of unbalanced optimal transport, J. Funct. Anal., 289 (2025), 111042.arxiv.org/abs/2112.11056
  11. [15]W. Gangbo, W. Li, S. Osher, M. Puthawala, Unnormalized optimal transport, J. Comput. Phys., 399 (2019), 108940.DOI
  12. [16]W. Gangbo, R.J. McCann, Optimal maps in Monge’s mass transport problem, C. R. Acad. Sci. Paris Sér. I Math., 321 (1995), 1653-1658.
  13. [17]K. Guittet, Extended Kantorovich norms: a tool for optimization, Research Report 4402, INRIA, 2002.
  14. [18]L. Kantorovich, On the translocation of masses, C.R. (Doklady) Acad. Sci. URSS (N.S.), 37 (1942), 199-201.
  15. [20]W. Lee, R. Lai, W. Li, S. Osher, Generalized unnormalized optimal transport and its fast algorithms, J. Comput. Phys., 436 (2021), 110041.DOI
  16. [21]M. Liero, A. Mielke, G. Savaré, Fine properties of geodesics and geodesic λ-convexity for the Hellinger–Kantorovich distance, Arch. Ration. Mech. Anal., 247 (2023), 112.
  17. [22]M. Liero, A. Mielke, G. Savaré, Optimal entropy-transport problems and a new Hellinger-Kantorovich distance between positive measures, Invent. Math., 211 (2018), 969-1117.
  18. [24]G. Monge, Mémoire sur la théorie des déblais et de remblais, Histoire de l’Académie Royale des Sciences de Paris, (1781), 666-704.
  19. [25]J. Onninen, P. Pankka, Quasiregular curves: Hölder continuity and higher integrability, Complex Analysis and its Synergies, 7 (2021), 26.
  20. [26]B. Piccoli, F. Rossi, Generalized Wasserstein distance and its application to transport equations with source, Arch. Ration. Mech. Anal., 211 (2014), 335-358.
  21. [27]B. Piccoli, F. Rossi, On properties of the generalized Wasserstein distance, Arch. Ration. Mech. Anal., 222 (2016), 1339-1365.arxiv.org/abs/1304.7014
  22. [28]G. Savaré, G.E. Sodini, A relaxation viewpoint to unbalanced optimal transport: Duality, optimality and Monge formulation, J. Math. Pures Appl., 188 (2024), 114-178.
  23. [29]N.S. Trudinger, X.J. Wang, On the Monge mass transfer problem, Calc. Var. Partial Differential Equations, 13 (2001), 19-31.DOI
  24. [30]C. Villani, Optimal transport: old and new, Springer-Verlag, Berlin, 2008.
  25. [31]C. Villani, Topics in optimal transportation, American Mathematical Society, Providence, RI, 2003.DOI

Paper details

vlnc:W_r8tfneym
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents