Displacement Convexity in Optimal Transport

• Displacement convexity is the property that entropy-like functionals are convex along optimal transport geodesics rather than linear interpolations, establishing connections between geometry and analysis.
• It links lower Ricci curvature bounds with functional inequalities and gradient flow stability, ensuring uniqueness and exponential convergence in various evolution equations.
• The concept finds applications in PDEs, statistical mechanics, machine learning, and coding theory, with extensions to discrete, tensor, and Lorentzian settings.

Displacement convexity is a central concept at the interface of optimal transport, geometry, analysis, and mathematical physics. It formalizes the idea that functionals, particularly entropy-like quantities, may exhibit convexity not along linear interpolations of measures, but along optimal transport geodesics in the space of probability measures endowed with the Wasserstein metric. This property has profound implications for characterizing geometric and analytic inequalities, establishing uniqueness and stability of minimizers, and connecting curvature properties of underlying spaces to the behavior of mass transport and thermodynamic functionals.

1. Definition and Core Formulation

The notion of displacement convexity was first formalized by McCann, originally in the context of the $L^2$-Wasserstein space $(\mathcal{P}_2(M), W_2)$, where $M$ is a complete Riemannian manifold and $W_2$ is the quadratic Kantorovich–Rubinstein distance: $$W_2(\mu_0, \mu_1)^2 = \inf_{\pi\in \Pi(\mu_0, \mu_1)} \int_{M \times M} d(x, y)^2\, d\pi(x, y).$$ A functional $$\mathcal{F}: \mathcal{P}_2(M) \to \mathbb{R} \cup \{+\infty\}$$ is said to be displacement convex if, for any pair $\mu_0, \mu_1$ and any $W_2$-geodesic $(\mu_s)_{s\in[0,1]}$ joining them (i.e., $\mu_s = ((1-s)\mathrm{Id} + s T)_\# \mu_0$ for some optimal transport map $T$), the map $s \mapsto \mathcal{F}(\mu_s)$ is convex: $$\mathcal{F}(\mu_s) \leq (1-s)\mathcal{F}(\mu_0) + s \mathcal{F}(\mu_1), \;\; \forall s\in[0,1] \tag{1}$$ This concept admits strengthening to $K$-convexity (or $\lambda$-displacement convexity) with a curvature term, and is robust to generalizations such as to the $L^p$-Wasserstein space, Finsler or Lorentzian settings, and to more general entropic or energy functionals (Ohta et al., 2010, McCann, 2018, Ohta et al., 2011).

2. Geometric and Analytic Characterization

Displacement convexity plays a pivotal role in linking geometric curvature bounds to the convexity of entropy-type functionals along optimal transport geodesics.

• Ricci curvature and entropy convexity: The foundational result (Otto–Villani, Cordero-Erausquin–McCann–Schmuckenschläger, von Renesse–Sturm) establishes that $K$-convexity of the Boltzmann–Shannon entropy $$\mathcal{S}(\mu) = \int \rho\log \rho\,d\mathrm{vol}_g$$ along $W_2$-geodesics in $\mathcal{P}_2(M)$ is equivalent to a lower Ricci curvature bound $\operatorname{Ric}_g \geq K g$ (McCann, 2018, Ohta et al., 2010, Lee, 2012). This principle extends to weighted Ricci curvature (Bakry–Émery) and generalized entropies (Ohta et al., 2011, Sakurai, 2017, Kuwae et al., 2020).
• Curvature-dimension conditions: The Lott–Sturm–Villani and Ohta–Sturm theories utilize displacement convexity of entire classes of entropy functionals (e.g., Rényi or Bregman entropies) to formulate synthetic lower curvature-dimension bounds for metric measure spaces, yielding the so-called $\mathrm{CD}(K, N)$ and later variants (e.g., $\mathrm{TwCD}(K,N,\varepsilon)$) (Ohta et al., 2010, Kuwae et al., 2020, Ohta et al., 2011, Sakurai, 2017, Cavalletti et al., 2020).
• Matrix and tensor strengthening: Recent advances have shown that matrix-valued (rather than scalar) displacement convexity—requiring positive semidefiniteness of the second derivative of certain matrix-valued functionals along geodesics—characterizes nonnegative sectional curvature, refining the classical Ricci-centric results and leading to 'intrinsic dimensional' inequalities (Aishwarya et al., 27 Sep 2025, Shenfeld, 2023). This tensorial approach captures directional and anisotropic aspects of curvature that are invisible to trace-based (scalar) conditions.

3. Examples and Applications Across Domains

Displacement convexity underpins major developments in geometric analysis, PDE, probability, coding theory, and mathematical physics:

• Phase boundaries and statistical mechanics: Displacement convexity ensures strict convexity of nonlocal or nonlinearly parameterized free energies, leading to uniqueness of phase boundaries and monotone fronts in multi-component systems without relying on direct rearrangement or dynamical arguments (0706.0133).
• Gradient flows and evolution equations: Displacement convexity for entropy-type functionals guarantees uniqueness, contractivity, and exponential convergence of gradient flows in spaces of measures—most notably for the Fokker–Planck (and nonlinear variants such as porous medium/fast diffusion) equations (Carrillo et al., 2016, Ohta et al., 2010, Bolte et al., 2016). The explicit computation of discrete displacement convexity constants provides rigorous rates of convergence in numerical schemes (Carrillo et al., 2016).
• Information-theoretic inequalities: The equivalence of Talagrand transport–entropy, logarithmic Sobolev, and Gagliardo–Nirenberg inequalities is governed by displacement convexity of suitable functionals, often via 'Łojasiewicz inequalities' in the Wasserstein metric (Bolte et al., 2016).
• Machine learning and mean-field theory: Mean-field limits of large neural networks and kernel methods have been rigorously connected to displacement convexity: risk functionals in the space of probability densities over weights become strongly displacement convex, ensuring global convergence and exponential rates for the limiting (gradient flow) dynamics (Javanmard et al., 2019).
• Coding theory and spatial coupling: In spatially coupled codes (e.g., LDPC for transmission near capacity), displacement convexity of global potential functionals yields uniqueness (up to translation) of interpolating fixed-point profiles, cleanly establishing threshold saturation and other nontrivial phenomena without requiring PDE analysis (El-Khatib et al., 2013, El-Khatib et al., 2017, El-Khatib et al., 2017).
• Lorentzian geometry, general relativity, and energy conditions: Recent work extends displacement convexity to the Lorentzian context, establishing that convexity of the Boltzmann entropy along suitably defined 'Lorentz–Wasserstein' geodesics is equivalent to the strong energy condition (timelike Ricci lower bound). Relatedly, null energy conditions and Bakry–Émery generalizations can be characterized via convexity of Rényi entropies on null hypersurfaces (McCann, 2018, Ketterer, 2023).

4. Methodologies and Technical Frameworks

Methodologies for establishing displacement convexity and deriving consequences rely on core optimal transport, geometric analysis, and PDE:

• Optimal transport geodesics: Convexity is assessed along geodesics in the Wasserstein space, constructed through Brenier–McCann maps (inverting monotone profiles for 1D/monotonic settings, or using Hamilton–Jacobi theory and Hamiltonian flows for higher-dimensional and time-dependent settings) (Lee, 2012, 0706.0133, Gomes et al., 2018).
• Second variation and matrix Riccati equations: Central to many results is the explicit calculation of the second derivative of functionals along transport geodesics, involving Jacobians of optimal maps and curvature terms (e.g., via Jacobi fields), leading to inequalities that intertwine entropy (or energy) convexity and geometric curvature (Aishwarya et al., 27 Sep 2025, Lee, 2012).
• Gradient-flow interpretation: Entropy functionals generate (metric) gradient flows in Wasserstein spaces, whose contractivity and convergence rates are controlled by displacement convexity, providing Lagrangian and Kleinian perspectives on evolving measures (Bolte et al., 2016, Carrillo et al., 2016, Ohta et al., 2010, Shenfeld, 2023).
• Extensions to discrete and semidiscrete settings: Displacement convexity methodology has been successfully ported to discrete graphs and semidiscrete PDE approximations by defining appropriate interpolation paths and non-local transportation metrics, enabling the proof of discrete Prekopa–Leindler, Talagrand, and Sobolev inequalities (Gozlan et al., 2012, Carrillo et al., 2016).

5. Strict Displacement Convexity, Uniqueness, and Variations

• Strict displacement convexity and uniqueness: For many physical and combinatorial models, strict displacement convexity (often after fixing translation symmetries) ensures that minimizing profiles or stationary points are unique (up to obvious invariances), a fact of critical importance in statistical mechanics (uniqueness of phase fronts), inference (uniqueness of BP profiles), and optimization (0706.0133, El-Khatib et al., 2013, El-Khatib et al., 2017).
• Joint and multicomponent displacement convexity: For multicomponent systems, the theory generalizes to functionals that are jointly convex along product-geodesics in product Wasserstein spaces, relevant to systems of interacting species and coupled recursions (0706.0133, El-Khatib et al., 2017).
• Extensions to synthetic geometry and non-smooth spaces: Displacement convexity is robustly defined in metric measure spaces, beyond smooth manifolds, enabling the extension of curvature-dimension theory and functional inequalities to non-smooth, synthetic spaces (Cavalletti et al., 2020, Sakurai, 2017, Kuwae et al., 2020).

6. Matrix Displacement Convexity and Higher-Order Quantitative Results

Matrix displacement convexity represents a sharp strengthening of the scalar convexity framework, relying on Loewner-order matrix inequalities for the second derivative (or matrix Riccati differential inequalities) along geodesics. This approach yields:

• Characterization of nonnegative sectional curvature via matrix displacement convexity of the entropy tensor, strengthening the classical (trace-based) displacement convexity criterion for Ricci curvature (Aishwarya et al., 27 Sep 2025).
• Intrinsic-dimension evolution variational inequalities, providing refined contraction and entropy growth bounds, with directional specificity that surpasses classical dimensional estimates (Shenfeld, 2023, Aishwarya et al., 27 Sep 2025).

Such matrix inequalities are directly relevant for PDE flows, quantum hydrodynamics, mean-field games, and entropic interpolation (Shenfeld, 2023).

7. Displacement Convexity Beyond the Riemannian Context

• Lorentzian and null displacement convexity: The synthesis of optimal transport with Lorentzian geometry has led to the identification of analogous geodesic and entropy convexity structures, allowing comparison with energy conditions fundamental to relativistic physics (McCann, 2018, Ketterer, 2023).
• Graph-theoretic and combinatorial settings: Discrete notions of displacement convexity underpin sharp discrete analytic inequalities (e.g., log-Sobolev, Prekopa–Leindler, Talagrand) and curvature bounds in graphs and hypercubes, with optimality in certain settings (e.g., the Boolean cube) (Gozlan et al., 2012).
• Applications to codes and constraint satisfaction: The paradigm effectively studies uniqueness and optimality in high-dimensional inference and communication models, from coding to compressive sensing (El-Khatib et al., 2013, El-Khatib et al., 2017).

---

Displacement convexity thus unifies curvature, entropy, and gradient flow perspectives in a wide array of mathematical and applied contexts, with deep consequences for analysis, probability, PDE, geometry, combinatorics, and statistical physics. It stands as a paradigm of the modern metric geometry–optimal transport interface.