betaFeedback

WorksMathematics

On the continuity of martingale optimal transport in higher dimensions

Abstract

We investigate qualitative and quantitative continuity of martingale optimal transport under perturbations of both marginals in dimensions d≥2d\geq2. We assume that each pair of marginals admits a reference martingale coupling whose density relative to the product of its marginals is bounded above and away from zero, uniformly along the sequence. If the limiting target measure is not carried by a hyperplane, we prove continuity under W1W_1-convergence for continuous costs of at most linear growth, assuming only finite first moments. For Lipschitz costs, a uniform pp-moment bound on the targets, p>1p>1, yields an explicit stability estimate with exponent (1−1/p)/2(1-1/p)/2; uniformly bounded target supports give exponent 1/21/2. The proof constructs a correction of couplings with small martingale bias while preserving both marginals. Truncated moment corrections allow the target measure to have unbounded support. The construction also shows that every martingale coupling of the limiting marginals can be approximated in W1W_1 by martingale couplings of the approximating marginals. We also discuss counterexamples showing that neither the uniform density condition nor non-degeneracy of the limiting target measure can be omitted in general.

Introduction

Let μ\mu and ν\nu be two probability measures on Rd\mathbb{R}^d with finite first moment. A coupling π(dx,dy)=μ(dx)πx(dy)\pi(dx,dy) = \mu(dx)\pi_x(dy) of μ\mu and ν\nu is a martingale coupling if

∫y πx(dy)=xfor μ-a.e. x.\int y\,\pi_x(dy) = x \qquad\text{for } \mu\text{-a.e. } x.

We denote the set of martingale couplings by ΠM(μ,ν)\Pi_M(\mu,\nu). By Strassen’s theorem [26 Theorem 8] the set ΠM(μ,ν)\Pi_M(\mu,\nu) is nonempty if and only if μ\mu and ν\nu are in convex order, denoted by μ⪯cν\mu\preceq_c \nu, which we assume throughout. For a continuous cost c:Rd×Rd→Rc : \mathbb{R}^d \times\mathbb{R}^d \to\mathbb{R} satisfying ∣c(x,y)∣≤C(1+∣x∣+∣y∣)|c(x,y)| \le C(1 + |x| + |y|), the martingale optimal transport (MOT) problem is defined as

V(μ,ν):=inf⁡π∈ΠM(μ,ν)∫c dπ.(1)V(\mu,\nu) := \inf_{\pi\in\Pi_M(\mu,\nu)} \int c\,d\pi. \tag*{(1)}

In this paper we investigate continuity of VV with respect to the marginals μ,ν\mu,\nu, i.e. we ask if the functional (μ,ν)↦V(μ,ν)(\mu,\nu) \mapsto V(\mu,\nu) is continuous in the Wasserstein metric. In dimension d=1d = 1, such continuity holds, see [1, 28]. In dimension d≥2d \ge2 continuity fails in general, see [6]. The reason is that the set ΠM(μn,νn)\Pi_M(\mu_n,\nu_n) need not approximate ΠM(μ,ν)\Pi_M(\mu,\nu). Indeed, limits of martingale couplings in ΠM(μn,νn)\Pi_M(\mu_n,\nu_n) are martingale couplings in ΠM(μ,ν)\Pi_M(\mu,\nu), see Lemma 2.2, but the converse may fail if the pairs (μn,νn)(\mu_n,\nu_n) are degenerate in a sense we specify below (cf. Assumptions 1.1 and 1.3). In the example of [6], ΠM(μn,νn)\Pi_M(\mu_n,\nu_n) consists of a single coupling for every nn, while ΠM(μ,ν)\Pi_M(\mu,\nu) contains many martingale couplings, some of them with strictly smaller cost. This is due to the fact that, for every nn, the martingale constraint forces the mass to move along prescribed directions, and this constraint disappears in the limit. The aim of this article is to identify a quantitative condition which prevents such hidden constraints, and under which VV is continuous with an explicit rate.

Main result

To exclude situations as in [6], where every point can send its mass only to a few prescribed points of the target, we require the existence of one martingale coupling which spreads the mass of every point over the whole target, with weights comparable to those of the target.

Assumption 1.1. Let κ∈(0,1)\kappa\in(0,1). There exists πˉ∈ΠM(μ,ν)\bar{\pi} \in\Pi_{\mathrm{M}}(\mu,\nu) with

κ≤dπˉd(μ⊗ν)≤κ−1μ⊗ν-a.e.(2)\kappa\leq\frac{\mathrm{d}\bar{\pi}}{\mathrm{d}(\mu\otimes\nu)} \leq\kappa^{-1} \qquad\mu\otimes\nu\text{-a.e.} \tag*{(2)}

By Fubini’s theorem, (Aκ)(A_\kappa) means that the kernels of πˉ\bar{\pi} satisfy

κν≤πˉx≤κ−1νfor μ-a.e. x.\kappa\nu\leq\bar{\pi}_x \leq\kappa^{-1}\nu\qquad\text{for }\mu\text{-a.e. }x.

We note that Assumption 1.1 asks for the existence of one coupling πˉ\bar{\pi} satisfying (Aκ)(A_\kappa). It is not a condition required for the whole set ΠM(μ,ν)\Pi_{\mathrm{M}}(\mu,\nu). In particular the optimizer of V(μ,ν)V(\mu,\nu) is typically singular with respect to μ⊗ν\mu\otimes\nu. Assuming existence of martingale measures equivalent to μ⊗ν\mu\otimes\nu is classical in the literature on entropic martingale optimal transport, see e.g. [19, 8] for related variants. As we will see below, (Aκ)(A_\kappa) alone is not sufficient for continuity of VV, and it has to be complemented by a non-degeneracy condition on the target ν\nu.

Let us now explain why this assumption is natural for the stability problem. In order to bound V(μn,νn)V(\mu_n,\nu_n) from above, we construct specific martingale couplings πn\pi_n of (μn,νn)(\mu_n,\nu_n), whose optimal transport costs ∫c dπn\int c\,\mathrm{d}\pi_n are close to V(μ,ν)V(\mu,\nu). This can be achieved by the shadow construction of [12], i.e. by moving an optimizer of the limit problem onto the new marginals. Such a perturbation produces a coupling with the correct marginals and almost the correct cost. However, it is only an approximate martingale coupling. The question is therefore whether a coupling with a small martingale bias can be corrected into an exact martingale coupling of the same marginals at a small cost. The coupling πˉ\bar{\pi} of (Aκ)(A_\kappa) is exactly the object which makes such a correction possible. Since πˉx≥κν\bar{\pi}_x \geq\kappa\nu, one may add to it signed perturbations of size smaller than κν\kappa\nu without losing positivity. Since πˉx≤κ−1ν\bar{\pi}_x \leq\kappa^{-1}\nu, the errors created on the second marginal by such perturbations remain controlled. In other words, (Aκ)(A_\kappa) quantifies the room available to move mass while respecting the martingale constraint. A similar tilting construction was used in [19 Ass. 2.1, Lemma 3.2] to show boundedness of dual optimizers, see also [8 Ass. 2.3, Prop. 2.8].

Assumption (Aκ)(A_\kappa) also implies that the pair (μ,ν)(\mu,\nu) is irreducible in the sense of [11], see Remark 2.5. It is however a strictly stronger requirement, since it forces the support of μ\mu to stay at a quantitative distance from the relative boundary of the component. In particular supp⁡(μ)\operatorname{supp}(\mu) has to be bounded, see Lemma 2.4. We think of (Aκ)(A_\kappa) as a quantitative version of irreducibility. For stability of VV it is required for the sequence (μn,νn)(\mu_n,\nu_n), with a constant κ\kappa independent of nn, and it then holds for the limit pair as well, see Lemma 2.6. We now state a class of examples, where Assumption 1.1 is naturally satisfied.

Example 1.2. Let ν\nu have mean mm, invertible covariance matrix Σ\Sigma and support in B(m,R)B(m,R), where here and in the following we denote by B(x,R)B(x,R) the closed ball of radius RR centered at xx, and let μ\mu have mean mm with

R∣Σ−1(x−m)∣≤θ<1for x∈supp⁡μ.R\left|\Sigma^{-1}(x-m)\right| \leq\theta< 1 \qquad\text{for }x \in\operatorname{supp}\mu.

Then the kernel

πˉx(dy):=(1+⟨Σ−1(x−m),y−m⟩)ν(dy)\bar{\pi}_x(dy) := \left(1 + \left\langle\Sigma^{-1}(x-m),y-m\right\rangle\right)\nu(dy)

has mean

∫yπˉx(dy)=m+∫(y−m)⟨Σ−1(x−m),y−m⟩ν(dy)+m⟨Σ−1(x−m),∫(y−m)ν(dy)⟩=m+∫(y−m)(y−m)Tν(dy)Σ−1(x−m)=m+ΣΣ−1(x−m)=x.\begin{aligned} \int y\bar{\pi}_x(dy) &= m + \int(y-m)\left\langle\Sigma^{-1}(x-m),y-m\right\rangle\nu(dy) + m\left\langle\Sigma^{-1}(x-m),\int(y-m)\nu(dy)\right\rangle\\ &= m + \int(y-m)(y-m)^{\mathsf{T}}\nu(dy)\Sigma^{-1}(x-m) \\ &= m + \Sigma\Sigma^{-1}(x-m) = x. \end{aligned}

Since μ\mu has mean mm, integrating the kernels against μ\mu gives

∫πˉxμ(dx)=(1+⟨Σ−1(∫xμ(dx)−m),y−m⟩)ν(dy)=ν(dy),\int\bar{\pi}_x\mu(dx) = \left(1 + \left\langle\Sigma^{-1}\left(\int x\mu(dx)-m\right),y-m\right\rangle\right)\nu(dy) = \nu(dy),

so that πˉ(dx,dy):=μ(dx)πˉx(dy)\bar{\pi}(dx,dy) := \mu(dx)\bar{\pi}_x(dy) has second marginal ν\nu. Moreover, ∣⟨Σ−1(x−m),y−m⟩∣≤θ\left|\left\langle\Sigma^{-1}(x-m),y-m\right\rangle\right|\leq\theta for x∈supp⁡μx\in\operatorname{supp}\mu and y∈supp⁡νy\in\operatorname{supp}\nu, so that the density of πˉx\bar{\pi}_x with respect to ν\nu lies in [1−θ,1+θ][1-\theta,1+\theta]. Hence (Aκ)(A_\kappa) holds with κ=1−θ\kappa=1-\theta. The construction is illustrated in Figure 1. Every kernel πˉx\bar{\pi}_x is a tilt of ν\nu which puts more mass than ν\nu on the half-space pointing towards xx, and less on the opposite one, in such a way that the barycenter moves from mm to xx. An analogous construction, replacing the linear function yy by its truncation ψT(y)\psi_T(y) defined in (8) gives examples with unbounded ν\nu, for instance a Gaussian ν\nu and a uniform μ\mu on a small ball.

Diagram illustrating the construction of Example 1.2

Figure 1. The construction of Example 1.2, with Σ\Sigma a multiple of the identity. The kernel πˉx\bar{\pi}_x puts more mass than ν\nu on the dashed half-space and less on the other half.

We denote the pp-th moment of ν\nu by Mp(ν):=∫∣y∣p dνM_p(\nu):=\int|y|^p\,d\nu, and for sequences μn⪯cνn\mu_n\preceq_c\nu_n converging to μ⪯cν\mu\preceq_c\nu we define the Wasserstein difference1

Δn:=W1(μn,μ)+W1(νn,ν).\Delta_n := W_1(\mu_n,\mu)+W_1(\nu_n,\nu).

As we see below, only first moments are needed for continuity of VV. Higher moments of νn\nu_n quantify the tails and enter the rate, while no additional moment assumption on μn\mu_n is needed since μn⪯cνn\mu_n \preceq_c \nu_n. Before we state our main result, we introduce a second assumption.

Assumption 1.3. The target measure ν\nu is not carried by a hyperplane.

We call ν\nu satisfying Assumption 1.3 non-degenerate.

References

  1. [1]J. Backhoff-Veraguas and G. Pammer. Stability of martingale optimal transport and weak optimal transport. Ann. Appl. Probab., 32(1):721–752, 2022.
  2. [2]M. Beiglböck, P. Henry-Labordère, and F. Penkner. Model-independent bounds for option prices—a mass transport approach. Finance Stoch., 17(3):477–501, 2013.
  3. [3]M. Beiglböck, B. Jourdain, W. Margheriti, and G. Pammer. Approximation of martingale couplings on the line in the adapted weak topology. Probab. Theory Related Fields, 183:359–413, 2022.
  4. [4]M. Beiglböck and N. Juillet. On a problem of optimal transport under marginal martingale constraints. Ann. Probab., 44(1):42–106, 2016.
  5. [5]M. Beiglböck, M. Nutz, and N. Touzi. Complete duality for martingale optimal transport on the line. Ann. Probab., 45(5):3038–3074, 2017.
  6. [6]M. Brückerhoff and N. Juillet. Instability of martingale optimal transport in dimension d ≥ 2. Electron. Commun. Probab., 27:Paper No. 24, 1–10, 2022.
  7. [7]G. Carlier, H. Mلامat, and M. Sylvestre. Weak optimal transport with moment constraints: constraint qualification, dual attainment and entropic regularization. arXiv preprint arXiv:2511.16211, 2025.arxiv.org/abs/2511.16211
  8. [8]F. Chen, G. Conforti, Z. Ren, and X. Wang. Convergence of sinkhorn’s algorithm for entropic martingale optimal transport problem. Mathematics of Operations Research, 2026.DOI
  9. [9]H. Chen, R. Xu, and X. Zhang. Entropic martingale optimal transport with Wasserstein relaxed marginal constraints. arXiv preprint arXiv:2611.08743, 2026.arxiv.org/abs/2611.08743
  10. [10]H. De March. Entropic approximation for multi-dimensional martingale optimal transport. arXiv preprint arXiv:1812.11104, 2018.arxiv.org/abs/1812.11104
  11. [11]H. De March and N. Touzi. Irreducible convex paving for decomposition of multidimensional martingale transport plans. Ann. Probab., 47(3):1726–1774, 2019.arxiv.org/abs/1702.08298
  12. [12]S. Eckstein and M. Nutz. Quantitative stability of regularized optimal transport and convergence of Sinkhorn’s algorithm. SIAM J. Math. Anal., 54(6):5922–5948, 2022.arxiv.org/abs/2110.06798
  13. [13]A. Galichon, P. Henry-Labordère, and N. Touzi. A stochastic control approach to no-arbitrage bounds given marginals, with an application to lookback options. Ann. Appl. Probab., 24(1):312–336, 2014.
  14. [14]N. Ghoussoub, Y.-H. Kim, and T. Lim. Structure of optimal martingale transport plans in general dimensions. Ann. Probab., 47(1):109–164, 2019.arxiv.org/abs/1508.01806
  15. [15]G. Guo and J. Obłój. Computational methods for martingale optimal transport problems. Ann. Appl. Probab., 29(6):3311–3347, 2019.
  16. [16]P. Henry-Labordère. From (martingale) Schrödinger bridges to a new class of stochastic volatility models. arXiv preprint arXiv:1904.04554, 2019.
  17. [17]M. M. Kay. Multi-dimensional martingales from mutual information. arXiv preprint arXiv:2512.16544, 2025.arxiv.org/abs/2512.16544
  18. [18]A. Kazeykina, Z. Ren, and H. Wang. Exponential convergence of Sinkhorn algorithm for entropy martingale optimal transport. arXiv preprint arXiv:2609.18850, 2026.arxiv.org/abs/2609.18850
  19. [19]M. Nutz and J. Wiesel. On the martingale Schrödinger bridge between two distributions. arXiv preprint arXiv:2401.05209, 2024.arxiv.org/abs/2401.05209
  20. [20]M. Nutz, J. Wiesel, and L. Zhao. Limits of semistatic trading strategies. Math. Finance, 33(1):185–205, 2023.arxiv.org/abs/2204.12251
  21. [21]M. Nutz, J. Wiesel, and L. Zhao. Martingale Schrödinger bridges and optimal semistatic portfolios. Finance Stoch., 27(1):233–254, 2023.
  22. [22]J. Obłój and P. Siorpaes. Structure of martingale transports in finite dimensions. arXiv preprint arXiv:1702.08433, 2017.
  23. [23]R. T. Rockafellar. Convex analysis, volume 28 of Princeton Mathematical Series. Princeton University Press, Princeton, NJ, 1970.
  24. [24]W. Rudin. Real and complex analysis. McGraw-Hill, New York, third edition, 1987.
  25. [25]M. Sion. On general minimax theorems. Pacific J. Math., 8:171–176, 1958.DOI
  26. [26]V. Strassen. The existence of probability measures with given marginals. Ann. Math. Statist., 36:423–439, 1965.DOI
  27. [27]C. Villani. Optimal transport: Old and new, volume 338 of Grundlehren der mathematischen Wissenschaften. Springer-Verlag, Berlin, 2009.
  28. [28]J. Wiesel. Continuity of the martingale optimal transport problem on the real line. Ann. Appl. Probab., 33:4645–4692, 2023.arxiv.org/abs/1905.04574

Paper details

vlnc:W_7b85nmjx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 1

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

  • Mathematics

AI tools

View statement
Looking for related papers…
Contents