WELL-POSEDNESS FOR THE CONTINUITY EQUATION FOR VECTOR FIELDS WITH SUITABLE MODULUS OF CONTINUITY ALBERT CLOP, HEIKKI JYLHA¨,JOANMATEU, AND JOAN OROBITG 7 Abstract. We prove well-posedness of linear scalar conservation laws 1 usingonlyassumptions onthegrowth andthemodulusofcontinuityof 0 2 thevelocityfield,butnotonitsdivergence. Asanapplication,weobtain uniqueness of solutions in the atomic Hardy space, H1, for the scalar n conservation law induced by a class of vector fields whose divergence is a an unboundedBMO function. J 7 1 ] 1. Introduction P A The scalar continuity equation . h d ρ+div(bρ) = 0 t (1) dt a m (ρ(0,·) = ρ0 [ appears in many conservation phenomena in nature. Among the most rep- resentative ones, we mention the so called aggregation equation (see for 1 v instance [8]), which describes the time evolution of an active scalar ρ(t,·) 3 with initial state ρ , under a velocity field 0 0 6 b(t,·) = K ∗ρ(t,·). 4 0 The dependenceof b with respect to the unknown u(t,·) makes the aggrega- . 1 tion equation nonlinear,whiletheanalytic natureof thekernelK = ∇N (N 0 is the Newtonian potential) gives it a gradient flow structure. In the present 7 paper, though, we omit both nonlinearities and gradient flow structure, and 1 : focus our attention on much simpler linear case. v i X It is known by the so-called superposition principle (see for instance [3]) r that if a vector field b admits a unique flow φ of trajectories t 7→ φ (x) then a t non-negativesolutionsto(1)areuniquelydeterminedbyρ . Forthisreason, 0 in the present paper we focus on the much more subtle question of signed solutions. Having a mass conservation structure, one may think that continuity equa- tions have L1 as the natural function space for solutions. It is then reason- able to think its adjoint equation, also known as transport equation, d ω+b·∇ω = 0, dt (ω(0,·) = ω0. Date: January 18, 2017. 1 2 ALBERTCLOP,HEIKKIJYLHA¨,JOANMATEU,ANDJOANOROBITG in the L∞ setting. Indeed, the classical theory by DiPerna and Lions [16], together with Ambrosio’s more recent developments [1], provides an ex- istence, uniqueness and stability theory for solutions u ∈ L∞(0,T;L∞) for any Sobolev vector field b which, among other assumptions, satisfies div(b) ∈ L∞. There is, though, an increasing interest in removing both boundedness assumptions on the divergence [11, 12, 13, 15, 21] as well as on the datum [14]. In this setting, the Euler system of equations deserves a special mention. Its scalar formulation in the plane, dω i (2) +v·∇ω = 0, where v(t,·) = ∗ω(t,·), dt 2πz has the structureof a nonlinear transportequation, together with theinitial condition v(0,·) = i ∗ω(0,·). After Yudovich’s proof of global existence 2πz and uniqueness of solutions ω ∈ L∞(0,T;L∞) for ω ∈ L1∩L∞ [28], there 0 has been many attempts to understand the case of unbounded vorticities. Particular attention is devoted to spaces that stay close to BMO, the space of functions of bounded mean oscillation, which arises naturally since it con- tains the image of L∞ under any Calder´on-Zygmund singular integral op- erator. It is conjectured that well-posedness of (2) fails in BMO. Also, strong ill-posedness has been proven in certain Sobolev spaces included in BMO [9]. However, there exist classes of unbounded vorticities for which (2) is well-posed [6, 7]. In a similar way, the present paper deals with the well-posedness of (1) in the atomic Hardy space H1, which is the predual of BMO, and which consists of L1 functions having L1 image under any Calder´on-Zygmund operator. Cauchy problems for the linear continuity equation with non-smooth ve- locity fields were successfully treated with the DiPerna-Lions scheme and the notion of renormalized solution, as well as the more recent extensions by Ambrosio in the BV setting. In both approaches, the starting point is the classical Cauchy-Lipschitz theory, which allows to write the solution ρ= ρ(t,x) of d ρ+div(bρ) = 0 (3) dt (ρ(0,·) = ρ0 as the adjoint composition operator, (4) ρ(t,x) = ρ ◦φ−1(x)J(x,φ−1) 0 t t where φ :Rn → Rn is the flow generated by the velocity field b, t φ˙ (x) = b(t,φ (x)), t t (5) (φ0(x) = x, and J(x,φ−1) denotes the jacobian determinant of the inverse flow φ−1 at t t the point x, at least for smooth enough b. Abusing of the classical mass transport notation, one could equivalently write ρ(t,·) = (φ ) ρ . Towards t ♯ 0 finding explicit solutions ρ ∈ L∞(0,T;H1(Rn)) of the problem (3) for a givenρ ∈ H1(Rn),therearetwothingstobeanalyzed. First,describingthe 0 class Qof diffeomorphismsφ underwhich(4)definesaboundedoperator in t H1(Rn). Second, describing the class of velocity fields b such that (5) has a WELL-POSEDNESS FOR CONTINUITY EQUATION 3 solution φ that falls into Q. Both questions drive us naturally to Reimann, t who partially solved them both in two papers in the 70’s [23, 22]. In the first one, quasiconformality was found to be the fundamental notion. In the second, uniform bounds for the anticonformal part of Db were proven to be enough. Let us recall that the anticonformal part of Db is, by definition, Db(t,x)+DTb(t,x) divb(t,x) S b(t,x) = − I , A n 2 n where I denotes the n-dimensional identity matrix. n Theorem1. Letb :[0,T)×Rn → Rn beavectorfield, b ∈ L1(0,T;W1,1(Rn)), loc such that S b ∈L1(0,T;L∞(Rn)). Assume also that A |b(t,x)| ∈ L1(0,T;L∞(Rn)). 1+|x| log(e+|x|) For every h ∈ H1(Rn), the problem 0 ∂ h+div(bh) = 0 in(0,T)×Rn (6) t h(0,·) = h inRn. 0 (cid:26) admits a unique weak solution h ∈L∞(0,T;H1(Rn)). The essential point for existence is the fact that quasiconformal maps trans- portboundedlymeasures with H1(Rn) density into themselves. For this, we rely on previous results by Reimann [23, 14] and Fefferman’s H1 −BMO duality. In the proof of uniqueness in Theorem 1, most of the available literature cannot be used, due to the unboundedness of div(b). However, we remark two exceptions. First, a work by Ambrosio - Bernard [2], where uniqueness is obtained from diagonal modulus of continuity assumptions on b, together with global boundedness. Unfortunately, this was not enough for proving Theorem 1, which includes many unbounded vector fields. We found a sec- ond exception in a paper by Seis [25], which provides a detailed stability estimate for solutions u ∈ L1(0,T;Lq(Rn)) of continuity equations with a velocity field b ∈ W1,p(Rn;Rn), 1 + 1 = 1, p ∈ (1,∞]. It turns out that a loc p q modification of the argument in [25] brings the following uniqueness result for solutions in L1(0,T;L1(Rn)) (actually a more general one, see Section 4 for details). Proposition 2. Let b : [0,T]×Rn → Rn be a vector field. Assume that (A1) |b(t,x)| ≤ C(1+|x|), and that (A2) |b(t,x)−b(t,y)| ≤ C|x−y| log(e+1/|x−y|), of all x,y ∈ Rn and almost every t ∈ [0,T]. If ρ ∈ L1(Rn), then the Cauchy 0 problem ∂ ρ+div(bρ) = 0 in (0,T)×Rn t ρ(0,·) = ρ in Rn 0 (cid:26) has at most one weak solution ρ∈ L1(0,T;L1(Rn)). 4 ALBERTCLOP,HEIKKIJYLHA¨,JOANMATEU,ANDJOANOROBITG The above result still does not provide a complete proof for uniqueness in Theorem 1. Indeed, there are vector fields b admissible for Theorem 1 for which (A1) and (A2) may both fail at the same time. It is worth mention- ing too that existence of solutions may fail in the setting of Proposition 2, since flows arising from the above velocity fields b do not preserve Lebesgue measurable sets in general. See [11, 12, 13] for optimal conditions in this direction. In particular, these flows could move an initial datum ρ away 0 from L1(Rn) and transform it into a non absolutely continuous measure. It is then natural to look at the case where ρ(t,·) belongs to the space of signed Borel measures on Rn with finite total variation, which is denoted by M(Rn). It turns out that this is the right choice for a full well-posedness theorem. Theorem 3. Let b :[0,T]×Rn → Rn be a vector field suchthat the following two conditions hold: There exists a continuous and nondecreasing function G :[0,∞) → (0,∞) satisfying ∞ ds = ∞ for some r > 0 such that r G(s) ´ |b(t,x)| (B1) sup ∈ L∞(0,T), x∈Rn G(|x|) andthere existsacontinuousandnondecreasing functionω : [0,∞) → [0,∞) satisfying r ds = ∞ for some r >0 such that 0 ω(s) ´ |b(t,x)−b(t,y)| (B2) sup ∈ L∞(0,T). ω(|x−y|) x,y∈B(0,R) for any radius R > 0. Then for any ρ ∈ M(Rn) there exists a unique 0 solution (as defined in Definition 4) ρ ∈ L1(0,T;M(Rn)) for the Cauchy problem ∂ ρ+div(bρ) =0 in(0,T)×Rn (7) t ρ(0,·) =ρ inRn. 0 (cid:26) Moreover, this solution is of the form ρ(t,·) = (φ ) ρ , where φ is the flow t ♯ 0 t of the vector field b. It is worth mentioning that Theorem 3 holds under slightly more general- ity. Namely, if one further knows that ρ ∈ L∞(0,T;M(Rn)), then unique- ness still holds if one replaces L∞(0,T) by L1(0,T) in both (B1) and (B2). The argument for obtaining existence in Theorem 3 is straightforward, be- cause of to the measure-valued setting and the continuity of b. Concerning uniqueness, the proof is much trickier. Exactly as in Proposition 2, the proof expands even further the methods used by Seis in [25] using opti- mal transport. Among the consequences, uniqueness in Theorem 1 now follows from Theorem 3. Moreover, it also provides uniqueness of solutions in L1(0,T;L1(Rn)) for the continuity equation in these cases where one can guaranteethepreservationofLebesguemeasurablesetsundertheflow. This is the case for instance if L div(b) ∈Exp , logLloglogL··· (cid:18) (cid:19) WELL-POSEDNESS FOR CONTINUITY EQUATION 5 see also [13]. Nevertheless, conditions (B1), (B2) are neither a consequence of, nor a reason for any condition on the divergence of b. Some counterex- amples in this direction are given at the end of Section 5. The paper is structured as follows. In Section 2 we give some definitions and basic results for continuity equations. In Section 3 we overview some results from optimal transport theory. In Section 4 we prove a slightly more general version of Proposition 2, which allows us to present the ideas behind theproofofTheorem3withoutgettinghunguponallthedetails. InSection 5 we prove Theorem 3 and give some counterexamples for the optimality of its assumptions. In Section 6 we prove Theorem 1. All along the paper we will be abusing notation, so that we identify ρ(t,·) = ρ and φ(t,·) = φ . t t Acknowledgements. A.C., J.M. and J. O. were partially supportedby re- search grants 2014SGR75 (Generalitat deCatalunya) andMTM2016-75390- P (spanishgovernment). Allnamed authorswere partially supportedbythe research grant FP7-607647 (European Union). 2. Basic properties of continuity equations Let us first introduce our notation. By M (Rn) we denote the set of finite + nonnegativeBorelmeasuresonRn. WedenotebyM(Rn)thespaceofsigned Borel measures on Rn with finite total variation, that is, µ ∈ M(Rn) if µ is a signed Borel measure and kµk := sup |µ(A )| :A ∈ B(Rn), A ∩A = ∅ fori6= j < ∞, TV j j i j nXj∈N o where B(Rn) is the Borel σ-algebra. For µ ∈ M(Rn) we denote by |µ| the total variation measureof µ. We use theJordan decomposition of measures: Any µ ∈ M(Rn) can be written as µ = µ+−µ− for some mutually singular measures µ+,µ− ∈ M (Rn) Given µ ∈ M(Rn) and a homeomorphism + φ: Rn → Rn we define the push-forward φ µ by ♯ φ µ(A) := µ(φ−1(A)) for any A∈ B(Rn). ♯ Equivalently, one can define φ µ by duality with the space C (Rn) of con- ♯ 0 tinuous functions vanishing at infinity, hφ µ,ϕi := ϕ(φ(x))dµ(x), ∀ϕ∈ C (Rn). ♯ ˆ 0 Then it is easy to see that φ µ ∈ M(Rn) with kφ µk ≤ kµk . ♯ ♯ TV TV Definition 4. A measure valued map ρ ∈ L1(0,T;M(Rn)) is a weak solu- tion of the Cauchy problem (7) for the continuity equation if the equality T ∂ ϕ(t,x)+hb(t,x),∇ϕ(t,x)i dρ (x)dt+ ϕ(0,x)dρ (x) = 0 ˆ ˆ t t ˆ 0 0 Rn Rn holds for a(cid:2)ll functions ϕ ∈ C∞([0,T)×(cid:3)Rn). c Let’s point out some simple properties of solutions, which we will prove for completeness. First is related to the weak continuity of solutions. 6 ALBERTCLOP,HEIKKIJYLHA¨,JOANMATEU,ANDJOANOROBITG Lemma 5. Let b ∈ L∞(0,T;L∞(Rn)). Let ρ ∈ L1(0,T;M(Rn)) be a loc solution for the Cauchy problem (7). Then we have the convergence t→0 ϕ(x)dρ (x) −→ ϕ(x)dρ (x) ˆ t ˆ 0 Rn Rn for every ϕ∈ C∞(Rn). c Proof. Let ϕ ∈ C∞(Rn). Fix t ∈ (0,T) and ε ∈ (0,t ). Choose a nonnega- c 0 0 tive function η ∈ C∞(t −ε,t +ε) such that η (s)ds = 1, and define t0,ε c 0 0 t0,ε ψ (t):= 1− tη (s)ds. Then using ψ ϕ a´s a test function we get t0,ε 0 t0,ε t0,ε ´ T T ψ′ (t) ϕdρ dt+ ψ (t) b(t,x),∇ϕ(x) dρ (x)dt ˆ t0,ε ˆ t ˆ t0,ε ˆ t 0 Rn 0 Rn (cid:10) (cid:11) + ϕdρ = 0. ˆ 0 Rn Taking ε → 0 we obtain (for a.e. t , if we are precise) 0 t0 (8) − ϕdρ + b(t,x),∇ϕ(x) dρ (x)dt+ ϕdρ = 0. ˆ t0 ˆ ˆ t ˆ 0 Rn 0 Rn Rn Thus we can deduce that (cid:10) (cid:11) t0 ϕdρ − ϕdρ ≤ b(t,x),∇ϕ(x) dρ (x)dt → 0, ˆ t0 ˆ 0 ˆ ˆ t Rn Rn 0 Rn (cid:12) (cid:12) (cid:12) (cid:12) as t (cid:12)→ 0, which is what we(cid:12)wa(cid:12)nted to pr(cid:10)ove. (cid:11) (cid:12) (cid:3) 0(cid:12) (cid:12) (cid:12) (cid:12) Before the next Lemma we need to introduce speciffic cut-off functions that will be used throughout the paper. Let k ∈ N and consider the function 1, if |x| ≤ k χ˜k(x) = max{0,1− |x| dr }, if |x| > k. ( k G(r) ´ Modifying this function we find a smooth, compactly supported function χ ∈ C∞(Rn) such that k c χ =1 in B(0,k), k χ =0 in Rn\B(0,R ) for some k <R < ∞ k k k (9) 2 and |∇χ (x)| ≤ . k G(|x|) Here theexistence of R follows fromthe assumption ∞ ds = ∞ for some k r G(s) (and thus any) r > 0. ´ The second important property of solutions is the conservation of mass, or from the point of view of this paper, the conservation of mass balance. Lemma 6. Suppose the vector field b satisfies the condition (B1) and that ρ∈ L1(0,T;M(Rn)) is a weak solution for the Cauchy problem ∂ ρ+div(bρ) = 0 (10) t ρ(0,·) = 0 (cid:26) Then for a.e. t ∈ [0,T] it holds that ρ (Rn) = dρ = 0, or in other words: t Rn t ρ+(Rn) = ρ−(Rn). ´ t t WELL-POSEDNESS FOR CONTINUITY EQUATION 7 Proof. Use a test function ϕ(t,x) = ψ(t)χ (x), where ψ ∈ C∞(0,T). This k c gives T T ψ′(t) χ (x)dρ (x)dt = − ψ(t) b(t,x),∇χ (x) dρ (x). ˆ ˆ k t ˆ ˆ k t 0 Rn 0 Rn Since this works for every ψ ∈ C∞(0,T), we obta(cid:10)in (cid:11) c ∂ χ (x)dρ (x) = b(t,x),∇χ (x) dρ (x). t k t k t ˆ ˆ Rn Rn (cid:16) (cid:17) Combining this with Lemma 5 we get (cid:10) (cid:11) |ρ (Rn)| = lim χ (x)dρ (x) t k t k→∞ ˆRn (cid:12) (cid:12) t (cid:12) (cid:12) ≤limsup (cid:12) b(s,x),∇χ (x) (cid:12)dρ (x) ds k s ˆ ˆ k→∞ 0 Rn (cid:12) (cid:12) |b(t,x)| (cid:12) (cid:10) (cid:11)t (cid:12) ≤2 (cid:12) limsup (cid:12) d|ρ |(x)ds = 0, s G(|x|) L∞((0,T)×Rn) k→∞ ˆ0 ˆRn\B(0,k) (cid:13) (cid:13) and thus c(cid:13)onclude t(cid:13)he proof. (cid:3) (cid:13) (cid:13) In the smooth case, the solution to the continuity equation can be found using the method of characteristics. Actually, this works also in our non- smooth case. Proof of the existence part of Theorem 3. Let φ be the flow of b. This flow exists by the classical Cauchy-Lipschitz theory, and moreover it satisfies t (11) φ(t,x) = x+ b s,φ(s,x) ds for any x ∈Rn. ˆ 0 (cid:0) (cid:1) We need to show that ρ := φ(t,·) ρ is a solution to the Cauchy problem t ♯ 0 (7). First, the properties of the push-forward ensure that we actually have ρ ∈ L∞(0,T;M(Rn)), since kρ k is bounded by kρ k . t t TV 0 TV Let ϕ ∈C∞([0,T)×Rn). First notice that ϕ is Lipschitz and t 7→ φ(t,x) c is absolutely continuous for any x ∈ Rn. This implies that the function f (t) := ϕ(t,φ(t,x)) is absolutely continuous for any x ∈ Rn and x T (12) f′(t)dt = f (T)−f (0) = −ϕ(0,x). ˆ x x x 0 On the other hand if we fix x ∈ Rn, then for L1-a.e. t ∈[0,T) we have f′(t) = ∂ ϕ(t,φ(t,x))+ ∇ϕ(t,φ(t,x)),∂ φ(t,x) x t t (13) = ∂tϕ(t,φ(t,x))+(cid:10)b(t,φ(t,x)),∇ϕ(t,φ(t,x(cid:11))) . Now we can check that ρt is a solu(cid:10)tion. Using the definition(cid:11)of the push- forward measure and applying Fubini’s Theorem, (13) and (12) we calculate T ∂ ϕ(t,x)+ b(t,x),∇ϕ(t,x) dρ (x)dt+ ϕ(0,x)dρ (x) ˆ ˆ t t ˆ 0 0 Rn Rn T (cid:2) (cid:10) (cid:11)(cid:3) = ∂ ϕ(t,φ(t,x))+ b(t,φ(t,x)),∇ϕ(t,φ(t,x)) dρ (x)dt+ ϕ(0,x)dρ (x) ˆ ˆ t 0 ˆ 0 0 Rn Rn T (cid:2) (cid:10) (cid:11)(cid:3) = f′(t)dtdρ (x)+ ϕ(0,x)dρ (x) = 0. ˆ ˆ x 0 ˆ 0 Rn 0 Rn 8 ALBERTCLOP,HEIKKIJYLHA¨,JOANMATEU,ANDJOANOROBITG Sinceϕ ∈ C∞([0,T)×Rn)wasarbitrary,φ(t,·) ρ isasolutiontotheCauchy c ♯ 0 problem (7). (cid:3) 3. Useful results from the theory of optimal transport In this section we present the part of the theory of optimal transport which we will use in the uniqueness proof. For further interest to this topic we refer to the books [27], [26] by Villani, as well as [24] by Santambrogio and [4] by Ambrosio, Gigli and Savar´e. We want to transport mass in Rn but our use of cut-off functions might lead to a difference between the initial and final mass. This could be a problem. The typical way (see e.g. [17], [10] and [19]) to avoid this mass inbalance is to add an isolated point ♦ to Rn. We writeRˆn := Rn∪{♦}. We also ”extend” the euclidean distance to Rˆn by setting |x−♦|= |♦−x| = ∞ whenever x ∈ Rn and obviously |♦−♦|= 0. Given two Borel measures µ,ν ∈ M (Rˆn) with µ(Rˆn) = ν(Rˆn) we de- + note the set of transport plans between µ and ν by Π(µ,ν). In other words, a measure λ ∈ M (Rˆn ×Rˆn) belongs to Π(µ,ν), if + λ(A×Rˆn) = µ(A) and λ(Rˆn ×A)= ν(A) for all Borel sets A⊂ Rˆn. Equivalently, λ ∈ Π(µ,ν) if u(x)+v(y) dλ(x,y) = u(x)dµ(x)+ v(y)dν(y) ˆ ˆ ˆ Rˆn×Rˆn Rˆn Rˆn (cid:0) (cid:1) for all u ∈ L1(µ), v ∈ L1(ν). If in addition to the measures µ,ν ∈ M (Rˆn) + we are given a continuous cost function c: Rˆn×Rˆn → [0,∞], we can study the optimal transport problem inf c(x,y)dλ(x,y). ˆ λ∈Π(µ,ν) Rˆn×Rˆn The existence of minimizers, called optimal transport plans, can be proved with the direct method in the calculus of variations (see e.g. [26, Theorem 4.1]). In our case, we are interested in cost functions of the form c(x,y) = c(|x−y|). With our notation, this means that c(x,♦) = c(♦,x) = c(∞) is a constant among all x ∈ Rn. We assume that c :[0,∞] → [0,∞] satisfies the following conditions: c(0) = 0, c(s) > 0 for every s > 0 and c is nondecreasing c is bounded and continuous (14)  c is concave    c is Lipschitz w.r.t. the euclidean distance in [0,∞). We set    C (µ,ν) := min c(|x−y|)dλ(x,y) c ˆ λ∈Π(µ,ν) Rˆn×Rˆn We will study the total transport cost C (µ,ν) for a certain family of cost c functions and then compare the estimates for these costs to the special case W(µ,ν):= min min{|x−y|,1}dλ(x,y), λ∈Π(µ,ν)ˆRˆn×Rˆn WELL-POSEDNESS FOR CONTINUITY EQUATION 9 which is used as a reference cost in our considerations. The key comparison between transportation costs C (µ,ν) and W(µ,ν) is given by the following c Lemma, which is a generalization of [25, Lemma 5] . Lemma 7. Let µ,ν ∈ M (Rˆn) with µ(Rˆn) = ν(Rˆn) and let c : [0,∞] → + [0,∞] be continuous and strictly increasing. Given any ε > 0 we have the upper bound C (µ,ν) C (µ,ν) W(µ,ν)≤ c−1 c µ(Rˆn)+ε+ c , ε c(1) (cid:18) (cid:19) if Cc(µ,ν) ∈ c [0,∞] . ε Proof. Define(cid:0)the fo(cid:1)llowing sets in Rˆn×Rˆn: C (µ,ν) c K = |x−y|≤ 1 :c(|x−y|) ≤ , 1 ε (cid:26) (cid:27) C (µ,ν) c K = |x−y|≤ 1 :c(|x−y|) > , 2 ε (cid:26) (cid:27) and K = |x−y|> 1 . 3 Let λ be an optimal plan(cid:8) for C (µ,ν(cid:9)). Using λ as a test plan for W(µ,ν) c we see that W(µ,ν)≤ min{|x−y|,1}dλ(x,y) ˆ Rˆn×Rˆn = min{|x−y|,1}dλ(x,y)+ min{|x−y|,1}dλ(x,y) ˆ ˆ K1 K2 + min{|x−y|,1}dλ(x,y) ˆ K3 C (µ,ν) ≤ c−1 c λ(K )+λ(K )+λ(K ). 1 2 3 ε (cid:18) (cid:19) Here it is easy to see that c(|x−y|) ε λ(K )= dλ(x,y) ≤ c(|x−y|)dλ(x,y) ≤ ε 2 ˆ c(|x−y|) C (µ,ν) ˆ K2 c K2 and similarly c(|x−y|) C (µ,ν) c λ(K ) = dλ(x,y) ≤ , 3 ˆ c(|x−y|) c(1) K3 which concludes the proof. (cid:3) If c satisfies the conditions (14), then c(|x−y|) gives a metric in Rˆn. For this kind of cost functions the optimal transport problem admits a dual formulation due to the celebrated Kantorovich Duality Theorem (see e.g. [26, Theorem 5.10]). 10 ALBERTCLOP,HEIKKIJYLHA¨,JOANMATEU,ANDJOANOROBITG Theorem 8. Let µ,ν ∈ M (Rˆn) with µ(Rˆn) = ν(Rˆn) > 0. Suppose c : + [0,∞] → [0,∞) satisfies conditions (14). Then we have the duality: min c(|x−y|)dλ(x,y) : λ ∈ Π(µ,ν) ˆ (cid:26) Rˆn×Rˆn (cid:27) |v(x)−v(y)| =max vdµ− vdν : sup ≤ 1, v(♦) = 0 . (ˆRˆn ˆRˆn x6=y c(|x−y|) ) In addition, the maximizers of the dual problem, called Kantorovich poten- tials, satisfy v(x)−v(y) = c(|x−y|) for every(x,y) ∈ suppλ, where λ ∈ Π(µ,ν) is an optimal plan for C (µ,ν). c Since it is important for us, we emphasize that in our case any test func- tion v for the dual problem is bounded and Lipschitz in the euclidean dis- tance: (15) |v(x)−v(y)| |c(t)−c(s)| sup |v(x)| ≤ c(∞) and sup ≤ sup , x∈Rˆn x,y∈Rn |x−y| t,s∈[0,∞) |t−s| In addition, information about the gradient of the Kantorovich potentials can also be extracted from the Duality Theorem. Corollary 9. Suppose c : [0,∞] → [0,∞) is C1 and satisfies (14). Let µ,ν ∈ M (Rˆn) with µ(Rˆn) = ν(Rˆn). Let λ ∈ Π(µ,ν) be an optimal plan + for C (µ,ν) and let v be a corresponding Kantorovich potential. c Then there exists a set E ⊂ Rn such that Ln(Rn \E) = 0 and for any (x,y) ∈ suppλ∩(E ×E) such that x 6= y we have x−y ∇v(x) = ∇v(y) = c′(|x−y|) |x−y| In addition, ∇v(x) = 0, if x ∈ E and (x,♦)∈ suppλ or (♦,x) ∈suppλ. Proof. Since v is Lipschitz in the euclidean distance, we can apply the Rademacher Theorem to find a set E ⊂ Rn such that Ln(Rn \ E) = 0 and v is differentiable in E. Now, given (x,y) ∈ suppλ∩(E ×E) we have by the Duality Theorem c(|x−y|)+v(y) = v(x) = inf {c(|x−y′|)+v(y′)}. y′∈Rn Thus, if x 6= y we can differentiate the function f(y′) = c(|x−y′|)+v(y′) at its minimum point y ∈ E, which gives y−x x−y c′(|x−y|) +∇v(y) = 0, i.e. ∇v(y) = c′(|x−y|) . |x−y| |x−y| Similarly, the Duality Theorem implies c(|x−y|)−v(x) = −v(y) = inf {c(|x′ −y|)−v(x′)}, x′∈Rn from which we obtain by differentiation x−y ∇v(x) = c′(|x−y|) . |x−y|