Curvature–Dimension (CD) Inequality

• Curvature–Dimension Inequality (CD) is a synthetic condition that enforces lower Ricci curvature bounds and upper dimension bounds across smooth, discrete, and nonlocal spaces.
• It underpins methodologies to derive sharp functional, Poincaré, and heat kernel inequalities using tools such as Bochner's formula, entropy convexity, and distortion coefficients.
• The framework extends to various settings—including weighted manifolds, graphs, and sub-Riemannian geometries—allowing for models with even negative dimension parameters.

A curvature–dimension inequality (abbreviated here as CD inequality) is a synthetic analytic or geometric condition designed to enforce lower Ricci curvature bounds and upper dimension bounds in a unified framework, extending from smooth Riemannian manifolds to metric measure spaces, weighted settings, sub-Riemannian geometries, graphs, and nonlocal operators. The archetype is the Bakry–Émery condition, formulated in terms of the carré du champ Γ and its second iteration Γ₂, but the definition admits powerful generalizations across smooth and discrete, local and nonlocal, and even negative-dimension settings.

1. Classical Bakry–Émery CD(ρ, n) Condition

Given a smooth Riemannian manifold $(M,g)$, possibly with a weight $V\in C^2(M)$ and associated measure $\mu(dx) = Z^{-1} e^{-V(x)} d\mathrm{vol}_g(x)$, define:

• Generator: $L = \Delta_g + \nabla V \cdot \nabla$
• Carré du champ:

$$\Gamma(f, h) = \frac{1}{2}(L(fh) - f L h - h L f),\quad \Gamma(f) = \Gamma(f, f)$$

• Iterated carré du champ:

$$\Gamma_2(f, h) = \frac{1}{2}(L \Gamma(f, h) - \Gamma(f, L h) - \Gamma(h, L f)),\quad \Gamma_2(f) = \Gamma_2(f, f)$$

The $CD(\rho, n)$ condition states that for all $f\in C^\infty(M)$,

$$\Gamma_2(f) \geq \rho\, \Gamma(f) + \frac{1}{n}(L f)^2$$

with curvature parameter $\rho\in \mathbb{R}$ and dimension parameter $V\in C^2(M)$0. For $V\in C^2(M)$1 and Ricci curvature $V\in C^2(M)$2, the classical Bochner–Lichnerowicz formula ensures that the unweighted Laplacian satisfies $V\in C^2(M)$3.

Generalizations

• For $V\in C^2(M)$4 (negative dimension), the theory extends to “negative effective dimension” models, such as generalized Cauchy distributions, which satisfy $V\in C^2(M)$5 with $V\in C^2(M)$6 and $V\in C^2(M)$7.
• Weighted CD$V\in C^2(M)$8 inequalities correspond to measures of the form $V\in C^2(M)$9 with $\mu(dx) = Z^{-1} e^{-V(x)} d\mathrm{vol}_g(x)$0 and allow for models with no log-Sobolev but valid Poincaré and Beckner inequalities (Gentil et al., 2019).

2. Synthetic and Metric CD$\mu(dx) = Z^{-1} e^{-V(x)} d\mathrm{vol}_g(x)$1 Theory

In the metric measure setting, the curvature–dimension condition is formulated via the convexity of entropy along Wasserstein geodesics:

• Let $\mu(dx) = Z^{-1} e^{-V(x)} d\mathrm{vol}_g(x)$2 be a complete separable geodesic metric space with a reference measure $\mu(dx) = Z^{-1} e^{-V(x)} d\mathrm{vol}_g(x)$3.
• For $\mu(dx) = Z^{-1} e^{-V(x)} d\mathrm{vol}_g(x)$4 and $\mu(dx) = Z^{-1} e^{-V(x)} d\mathrm{vol}_g(x)$5, the $\mu(dx) = Z^{-1} e^{-V(x)} d\mathrm{vol}_g(x)$6 condition asserts that for any pairs of absolutely continuous measures, the $\mu(dx) = Z^{-1} e^{-V(x)} d\mathrm{vol}_g(x)$7-Rényi entropy $\mu(dx) = Z^{-1} e^{-V(x)} d\mathrm{vol}_g(x)$8 (where $\mu(dx) = Z^{-1} e^{-V(x)} d\mathrm{vol}_g(x)$9) evolves $L = \Delta_g + \nabla V \cdot \nabla$0-convexly along $L = \Delta_g + \nabla V \cdot \nabla$1-geodesics, with explicit comparison via distortion coefficients defined from $L = \Delta_g + \nabla V \cdot \nabla$2 and the distance between points (Ketterer, 2015, Magnabosco et al., 2022, Magnabosco et al., 2022).

The key analytic expression is:

$L = \Delta_g + \nabla V \cdot \nabla$3

where $L = \Delta_g + \nabla V \cdot \nabla$4 is an optimal transport plan, $L = \Delta_g + \nabla V \cdot \nabla$5 are absolutely continuous, and $L = \Delta_g + \nabla V \cdot \nabla$6 are distortion coefficients.

3. Main Analytic and Geometric Implications

A core consequence of the CD condition is the derivation of sharp functional inequalities with explicit constants depending on the curvature and dimension parameters (Gentil et al., 2019, Cavalletti et al., 2015, Bakry et al., 2014):

• Poincaré/Variance bound:

$L = \Delta_g + \nabla V \cdot \nabla$7

• Beckner-type inequalities:

$L = \Delta_g + \nabla V \cdot \nabla$8

with precise $L = \Delta_g + \nabla V \cdot \nabla$9 under $$\Gamma(f, h) = \frac{1}{2}(L(fh) - f L h - h L f),\quad \Gamma(f) = \Gamma(f, f)$$0 for weighted measures (Gentil et al., 2019).

• Logarithmic Sobolev and spectral gap inequalities in essentially non-branching or smooth CD$$\Gamma(f, h) = \frac{1}{2}(L(fh) - f L h - h L f),\quad \Gamma(f) = \Gamma(f, f)$$1 settings (Cavalletti et al., 2015).
• Li–Yau differential Harnack inequalities and heat kernel bounds, refined by the exact CD parameters (Bakry et al., 2014, Garofalo et al., 2013, Münch, 2014).

4. Entropy Flows, Bochner Formula, and Equivalence Properties

The information-theoretic viewpoint connects CD conditions to the concavity or convexity of entropy along Wasserstein geodesics (Li, 2024, Bolley et al., 2015):

• For $$\Gamma(f, h) = \frac{1}{2}(L(fh) - f L h - h L f),\quad \Gamma(f) = \Gamma(f, f)$$2 smooth, the CD$$\Gamma(f, h) = \frac{1}{2}(L(fh) - f L h - h L f),\quad \Gamma(f) = \Gamma(f, f)$$3 condition is equivalent to the differential inequality for entropy $$\Gamma(f, h) = \frac{1}{2}(L(fh) - f L h - h L f),\quad \Gamma(f) = \Gamma(f, f)$$4 along smooth geodesics in Wasserstein space:

$$\Gamma(f, h) = \frac{1}{2}(L(fh) - f L h - h L f),\quad \Gamma(f) = \Gamma(f, f)$$5

with a characterization of rigidity models as $$\Gamma(f, h) = \frac{1}{2}(L(fh) - f L h - h L f),\quad \Gamma(f) = \Gamma(f, f)$$6-Einstein manifolds with Hessian soliton potentials attaining equality.

• In the synthetic theory, displacement convexity of entropy functional (e.g., relative entropy or Rényi entropy) along all transport geodesics is both necessary and sufficient for the CD condition; this convexity yields local Poincaré inequalities, volume doubling, and the uniqueness of almost every geodesic (Rajala, 2011).

5. Discrete, Nonlocal, and Sub-Riemannian CD Inequalities

Discrete (Graphs and Markov Chains)

Discrete analogues of the CD condition have been formulated for graphs (undirected, directed), Markov chains, and nonlocal operators:

• On a graph $$\Gamma(f, h) = \frac{1}{2}(L(fh) - f L h - h L f),\quad \Gamma(f) = \Gamma(f, f)$$7, with normalized Laplacian $$\Gamma(f, h) = \frac{1}{2}(L(fh) - f L h - h L f),\quad \Gamma(f) = \Gamma(f, f)$$8 and carré du champ $$\Gamma(f, h) = \frac{1}{2}(L(fh) - f L h - h L f),\quad \Gamma(f) = \Gamma(f, f)$$9, the Bakry–Émery $$\Gamma_2(f, h) = \frac{1}{2}(L \Gamma(f, h) - \Gamma(f, L h) - \Gamma(h, L f)),\quad \Gamma_2(f) = \Gamma_2(f, f)$$0 inequality takes the form (Lin et al., 2015, Yamada, 2017):

$$\Gamma_2(f, h) = \frac{1}{2}(L \Gamma(f, h) - \Gamma(f, L h) - \Gamma(h, L f)),\quad \Gamma_2(f) = \Gamma_2(f, f)$$1

Equivalent properties include gradient bounds, Poincaré and reverse Poincaré inequalities, and new variants such as CDE$$\Gamma_2(f, h) = \frac{1}{2}(L \Gamma(f, h) - \Gamma(f, L h) - \Gamma(h, L f)),\quad \Gamma_2(f) = \Gamma_2(f, f)$$2 suited to discrete settings with positive functions and the square-root transformation (Lin et al., 2015, Münch, 2015).

• Nonlinear variants involving $$\Gamma_2(f, h) = \frac{1}{2}(L \Gamma(f, h) - \Gamma(f, L h) - \Gamma(h, L f)),\quad \Gamma_2(f) = \Gamma_2(f, f)$$3-Laplacians, such as $$\Gamma_2(f, h) = \frac{1}{2}(L \Gamma(f, h) - \Gamma(f, L h) - \Gamma(h, L f)),\quad \Gamma_2(f) = \Gamma_2(f, f)$$4, are appropriate for generalizing Li–Yau inequalities to graphs (Münch, 2014).
• The discrete CD$$\Gamma_2(f, h) = \frac{1}{2}(L \Gamma(f, h) - \Gamma(f, L h) - \Gamma(h, L f)),\quad \Gamma_2(f) = \Gamma_2(f, f)$$5 inequality for Markov chains is tailored to yield entropy decay estimates and Beckner-type functional inequalities using nonlinear carré du champ (Weber et al., 2020).

Nonlocal Operators

For nonlocal operators $$\Gamma_2(f, h) = \frac{1}{2}(L \Gamma(f, h) - \Gamma(f, L h) - \Gamma(h, L f)),\quad \Gamma_2(f) = \Gamma_2(f, f)$$6 on $$\Gamma_2(f, h) = \frac{1}{2}(L \Gamma(f, h) - \Gamma(f, L h) - \Gamma(h, L f)),\quad \Gamma_2(f) = \Gamma_2(f, f)$$7 determined by a symmetric jump kernel, the CD$$\Gamma_2(f, h) = \frac{1}{2}(L \Gamma(f, h) - \Gamma(f, L h) - \Gamma(h, L f)),\quad \Gamma_2(f) = \Gamma_2(f, f)$$8 inequality controls the sums-of-squares representation of $$\Gamma_2(f, h) = \frac{1}{2}(L \Gamma(f, h) - \Gamma(f, L h) - \Gamma(h, L f)),\quad \Gamma_2(f) = \Gamma_2(f, f)$$9 and gives a dimension bound if the second moment is finite (Spener et al., 2019). Fractional Laplacians typically fail such CD inequalities for any finite $CD(\rho, n)$0.

Sub-Riemannian Settings

Generalized CD$CD(\rho, n)$1 conditions have been developed for subelliptic operators, involving an additional vertical carré du champ $CD(\rho, n)$2. Consequences include volume doubling, Poincaré, and Harnack inequalities across Sasakian manifolds, Carnot groups, and more (Baudoin et al., 2010).

6. Structural and Model Cases, Examples, and Limitations

A variety of model spaces arise that saturate or illustrate the CD condition in different parameter regimes:

• Gaussian measure: $CD(\rho, n)$3, $CD(\rho, n)$4 holds and sharp Beckner inequalities are attained (Gentil et al., 2019).
• Generalized Cauchy distributions: $CD(\rho, n)$5 or $CD(\rho, n)$6 can hold, despite absence of log-Sobolev inequality.
• Weighted manifolds and metric measure spaces: CD$CD(\rho, n)$7 is characterized pointwise by variable Ricci curvature lower bounds and extends naturally to non-constant curvature functions and product spaces, with stability under Gromov–Hausdorff convergence (Ketterer, 2015).

Stability, Rigidity, and Branching

• CD conditions are stable under measured Gromov–Hausdorff convergence, including the "negative dimension" regime $CD(\rho, n)$8 when equipped with the appropriate quasi-Radon framework (Magnabosco et al., 2021).
• Topological dimension need not be constant in CD spaces (example: spaces with branching or collapsed regions) (Magnabosco, 2021).
• The weak CD$CD(\rho, n)$9 property does not in general guarantee non-branching, highlighting differences between synthetic Ricci curvature bounds and classical Riemannian structure.

7. Deep Equivalences and Reformulations

Recent advances emphasize the intrinsic equivalence between the CD condition and generalized Brunn–Minkowski type inequalities, specifically:

• On weighted Riemannian manifolds, BM$f\in C^\infty(M)$0 and CD$f\in C^\infty(M)$1 are equivalent, providing a metric–measure characterization of curvature-dimension solely in terms of volume growth of interpolated sets (Magnabosco et al., 2022).
• In essentially non-branching metric spaces, the “strong Brunn–Minkowski” inequality SBM$f\in C^\infty(M)$2 is equivalent to CD$f\in C^\infty(M)$3; the global version of the sharp Brunn–Minkowski is thus a full geometric characterization of synthetic Ricci bounds (Magnabosco et al., 2022).
• These equivalences enable direct access to sharp functional inequalities (Poincaré, log-Sobolev, Talagrand, Sobolev) with model constants, and allow for globalization arguments for local curvature-dimension conditions (Cavalletti et al., 2015).

---

Table: Formulations of the Curvature–Dimension Condition

Setting	CD Inequality	Main Analytic Implication
Riemannian (smooth)	$f\in C^\infty(M)$4	Sharp Poincaré, log-Sobolev, spectral gap
Metric measure space	$f\in C^\infty(M)$5 convex along $f\in C^\infty(M)$6-geodesic	Brunn–Minkowski, doubling, Poincaré
Discrete (graph)	$f\in C^\infty(M)$7	Spectral gap, heat kernel estimates
Nonlocal (operator)	Similar, with discrete sums	Effective dimension, decay estimates
Sub-Riemannian	Involves $f\in C^\infty(M)$8, $f\in C^\infty(M)$9, with extra coupling terms	Volume doubling, Harnack, functional ineq.

---

• "A family of Beckner inequalities under various curvature-dimension conditions" (Gentil et al., 2019)
• "On the geometry of metric measure spaces with variable curvature bounds" (Ketterer, 2015)
• "The Brunn--Minkowski inequality implies the CD condition in weighted Riemannian manifolds" (Magnabosco et al., 2022)
• "Equivalent Properties of CD Inequality on Graph" (Lin et al., 2015)
• "Curvature-dimension condition, rigidity theorems and entropy differential inequalities on Riemannian manifolds" (Li, 2024)
• "Curvature-dimension inequalities for non-local operators in the discrete setting" (Spener et al., 2019)
• "Example of an Highly Branching CD Space" (Magnabosco, 2021)
• "Li-Yau inequality on finite graphs via non-linear curvature dimension conditions" (Münch, 2014)
• "The Li-Yau inequality and applications under a curvature-dimension condition" (Bakry et al., 2014)
• "Sharp geometric and functional inequalities in metric measure spaces with lower Ricci curvature bounds" (Cavalletti et al., 2015)
• "Li-Yau and Harnack type inequalities in $$\Gamma_2(f) \geq \rho\, \Gamma(f) + \frac{1}{n}(L f)^2$$0 metric measure spaces" (Garofalo et al., 2013)
• "Equivalence between dimensional contractions in Wasserstein distance and the curvature-dimension condition" (Bolley et al., 2015)
• "The entropy method under curvature-dimension conditions in the spirit of Bakry-Émery in the discrete setting of Markov chains" (Weber et al., 2020)
• "A sub-Riemannian curvature-dimension inequality, volume doubling property and the Poincaré inequality" (Baudoin et al., 2010)
• "The strong Brunn--Minkowski inequality and its equivalence with the CD condition" (Magnabosco et al., 2022)
• "Convergence of metric measure spaces satisfying the CD condition for negative values of the dimension parameter" (Magnabosco et al., 2021)
• "Local Poincaré inequalities from stable curvature conditions on metric spaces" (Rajala, 2011)
• "Curvature dimension inequalities on directed graphs" (Yamada, 2017)
• "Remarks on curvature dimension conditions on graphs" (Münch, 2015)