1.2. Distance cost functions 35 convergence in distance dBL∗ is equivalent to weak convergence, while con- vergence in distance Td is just slightly stronger when X is unbounded, as we shall discuss in Chapter 7. From the Kantorovich-Rubinstein theorem we will now deduce a corol- lary which is intuitively clear, but not so trivial to prove. To state it in a clear way, it will be convenient to consider mass transportation between nonnegative measures whose mass is not normalized to unity (this extension is straightforward). Corollary 1.16 (Invariance of Kantorovich-Rubinstein distance un- der mass subtraction). Let X = Y be a Polish space and let d be a lower semi-continuous distance on X. Let μ, ν and σ be three Borel (nonnegative) measures on X, such that μ[X] = ν[X] +∞, σ[X] +∞. Then Td(μ + σ, ν + σ) = Td(μ, ν). Of course the bound Td(μ + σ, ν + σ) Td(μ, ν) is immediate (exercise) however the converse bound would be more tricky to establish! This corol- lary can be reformulated in the following way: whenever μ and ν are two probability measures on X, then Td(μ, ν) = Td μ ν]+,ν μ]+ . In other words, in a mass transportation problem with a distance cost func- tion, one can assume that all the mass shared between the two probability measures does stay in place. In the above the notation ρ+ stands for the nonnegative part of the Radon measure ρ it is defined by the characteristic property that ρ can be written as the Hahn decomposition ρ+ ρ−, where ρ+ and ρ− are nonnegative Borel measures which are singular to each other, i.e. concentrated on disjoint sets. We now come to the proof of the Kantorovich-Rubinstein theorem. We note that a more general version (without the completeness assumption) is proven in [119, Section 11.8]. Proof of Theorem 1.14. Let dn = d/(1+n−1d): this is a distance satisfy- ing dn d, and for all x, y the quantity dn(x, y) converges monotonically to d(x, y) as n ∞. In particular, the set of 1-Lipschitz functions for dn is in- cluded in the set of 1-Lipschitz functions for d. Reasoning as in Step 3 of the proof of Theorem 1.3, we see that we just have to prove Theorem 1.14 with d replaced by dn. Hence, in the sequel we shall assume that d is bounded. In this case, all Lipschitz functions are bounded, and therefore integrable with respect to μ, ν. So, in view of Theorem 1.3, the only thing to check is
Previous Page Next Page