Hyderabad · 19 August 2010

Cédric Villani&the Fields Medal

A derivation of the three pillars

« For his proofs of nonlinear Landau damping and convergence to equilibrium for the Boltzmann equation. »
The geometric counterpart — synthetic Ricci bounds via optimal transport — was named alongside.

Villani’s medal cites two theorems: relaxation toward Maxwellian equilibrium for the spatially inhomogeneous Boltzmann equation (with Desvillettes), and the survival of Landau damping in the nonlinear Vlasov–Poisson regime (with Mouhot). A third, geometric, strand — Lott–Villani’s synthetic Ricci curvature on metric–measure spaces — runs underneath. We sketch the analytic skeleton of each.

I.

Convergence to equilibrium for the Boltzmann equation

§ Setup

Let \(f(t,x,v)\ge 0\) denote the phase-space density of a rarefied gas, with \(x\in\mathbb{T}^3\) (periodic box, to skirt boundary issues) and \(v\in\mathbb{R}^3\). The Boltzmann equation reads

\[ \partial_t f \;+\; v\cdot\nabla_x f \;=\; Q(f,f), \]

with the bilinear collision operator

\[ Q(f,f)(v) \;=\; \int_{\mathbb{R}^3}\!\!\int_{S^2} B(v-v_*,\sigma)\, \bigl[ f'f'_* - f f_*\bigr]\, d\sigma\,dv_*, \]

where pre- and post-collisional velocities are linked by

\[ v' = \tfrac{v+v_*}{2} + \tfrac{|v-v_*|}{2}\sigma, \qquad v'_* = \tfrac{v+v_*}{2} - \tfrac{|v-v_*|}{2}\sigma, \]

and \(B\ge 0\) is the collision kernel (hard spheres: \(B = |v-v_*|\)).

Definition.Boltzmann H-functional.

\[ H[f] \;=\; \int_{\mathbb{T}^3\times\mathbb{R}^3} f \log f \, dx\,dv. \]

§ The H-theorem

Multiplying the equation by \(\log f\) and integrating, the symmetries of the collision integral give

\[ \frac{d}{dt}H[f] \;=\; -\,D[f], \qquad D[f] \;=\; \tfrac14\int B \,(f'f'_* - f f_*)\log\!\tfrac{f'f'_*}{f f_*}\, d\sigma\,dv_*\,dv\,dx \;\ge\; 0. \]

Equality \(D[f]=0\) holds iff \(f\) is a local Maxwellian. Combined with conservation of mass, momentum, and energy, the only stationary state compatible with the prescribed invariants is the global Maxwellian

\[ M(v) \;=\; \frac{\rho_\infty}{(2\pi T_\infty)^{3/2}}\, \exp\!\Bigl(-\,\frac{|v-u_\infty|^2}{2T_\infty}\Bigr). \]

§ Cercignani’s conjecture and its falsification

The natural quantitative question: does entropy production control relative entropy? Cercignani conjectured

\[ D[f] \;\ge\; \lambda \,H[f \,\vert\, M], \qquad H[f\,\vert\,M] = \int f \log\frac{f}{M}. \]

This would yield exponential relaxation: \(H(f|M)\le H_0\,e^{-\lambda t}\). Bobylev and Cercignani showed it is false in general: one can build distributions arbitrarily far from \(M\) for which \(D[f]\) is arbitrarily small relative to \(H[f|M]\).

§ Desvillettes–Villani: almost-exponential relaxation

The Desvillettes–Villani strategy (2005) trades the lost exponential rate for any polynomial rate, using a hierarchy of weakenedentropy-production estimates that tolerate the failure of Cercignani’s bound but exploit additional smoothness and moments:

Theorem.(Desvillettes–Villani, 2005)

Assume \(f\in C^\infty\) with uniform-in-\(t\) Sobolev and moment bounds, and that \(f\ge K_0 e^{-A_0 |v|^{q_0}}\) is uniformly bounded below by a Maxwellian-type tail. Then for every \(\varepsilon>0\) there exists \(C_\varepsilon\) with

\[ H\!\bigl(f(t)\,\vert\,M\bigr) \;\le\; C_\varepsilon\,(1+t)^{-1/\varepsilon}. \]

That is, convergence to Maxwellian is faster than any polynomial; symbolically \(H(f(t)|M)=O(t^{-\infty})\).

§ Skeleton of the proof

Two functionals must be controlled simultaneously: the local relative entropy

\[ H_{\mathrm{loc}}(t) \;=\; \int_{\mathbb{T}^3} H\!\bigl(f(t,x,\cdot)\,\vert\,M_f(t,x,\cdot)\bigr)\,dx, \]

where \(M_f\) is the local Maxwellian with the same hydrodynamic moments as \(f(t,x,\cdot)\); and the fluid part

\[ H_{\mathrm{fluid}}(t) \;=\; \int_{\mathbb{T}^3} H\!\bigl(M_f(t,x,\cdot)\,\vert\,M\bigr)\,dx. \]

Three coupled inequalities drive the argument:

(i) a. A weakened Cercignani bound: under the smoothness hypotheses,

\[ D[f] \;\ge\; K_\varepsilon\,H_{\mathrm{loc}}^{1+\varepsilon}. \]

(ii) b. A differential inequality on the fluid part, derived from the second-order moments equation, gives a Korn-type estimate

\[ \frac{d^2}{dt^2}H_{\mathrm{fluid}} \;\ge\; -\,C\bigl(H_{\mathrm{loc}}\bigr)^{\theta} \;+\; c\,H_{\mathrm{fluid}}^{\,\beta}. \]

(iii) c. A system of ODEs in \((H_{\mathrm{loc}},H_{\mathrm{fluid}})\)that, when closed, forces both to decay faster than any polynomial.

The crucial novelty is (ii): it shows the oscillation of the fluid moments along trajectories is itself penalized by entropy production. This couples the kinetic (microscopic) and hydrodynamic (macroscopic) layers in a single Lyapunov hierarchy.

Remark.The polynomial loss is intrinsic: explicit constructions of perturbations of the Maxwellian whose entropy decays only polynomially are known. The result is in this sense optimal modulo the \(\varepsilon\).

II.

Nonlinear Landau damping

§ The phenomenon

In 1946 Lev Landau predicted that, in a hot collisionless plasma described by Vlasov–Poisson, small perturbations of a homogeneous equilibrium decay despite the system being time-reversible and non-dissipative. The mechanism is purely kinematic: phase mixing in velocity space shuffles the perturbation into ever-finer oscillations, and macroscopic averages relax. Landau’s argument was linear; the question of whether the prediction survives nonlinear self-consistent feedback remained open for sixty-five years.

§ The equation

On \(\mathbb{T}^d\times\mathbb{R}^d\), Vlasov–Poisson reads

\[ \partial_t f + v\cdot\nabla_x f + F[f]\cdot\nabla_v f = 0, \qquad F[f] = -\nabla W *_x \rho,\qquad \rho(t,x) = \int f\,dv, \]

with \(W\) the interaction potential (Coulomb, Newton, or any analytic kernel of suitable decay).

§ Linear Landau damping

Linearize around a homogeneous equilibrium \(f^0(v)\): write \(f = f^0(v) + h(t,x,v)\) and drop quadratic terms,

\[ \partial_t h + v\cdot\nabla_x h - (\nabla W *_x \rho_h)\cdot\nabla_v f^0 = 0. \]

Decompose into Fourier modes in \(x\), so \(\hat h(t,k,v)\) and \(\hat\rho_h(t,k)\) satisfy a Volterra equation

\[ \hat\rho_h(t,k) \;=\; \hat\rho_h^{\,\mathrm{free}}(t,k) \;+\; \int_0^t K(t-\tau,k)\,\hat\rho_h(\tau,k)\,d\tau, \]

with kernel \(K(s,k) = -\widehat W(k)\,(2\pi i)\,k\cdot\widehat{\nabla_v f^0}\,(ks)\). The solvability of this convolution equation in a half-plane is governed by the Penrose stability condition

\[ \inf_{k\ne 0}\;\inf_{\Re z\ge 0}\;\bigl|1 - \widehat{K}(z,k)\bigr| \;>\; 0. \]

If \(f^0\) is analytic and Penrose-stable, one obtains

\[ \bigl|\hat\rho_h(t,k)\bigr| \;\le\; C\, e^{-\lambda |k|\,t}, \]

and the same exponential bound on the field \(F\). Linear damping is thus field-level decay; \(h\) itself does notconverge in any strong sense — it filaments.

§ The nonlinear obstruction: plasma echoes

Quadratic resonances between modes \(k\) and \(\ell\)produce an “echo” excitation of mode \(k+\ell\) at a delayed time \(t_{\text{echo}}\sim \frac{|k+\ell|}{|k|}\bigl|t_*\bigr|\). Echoes can cascade and reinforce. In any norm based on fixed regularity, this yields a loss of derivatives at each iteration step — fatal to Picard or contraction arguments.

§ Mouhot–Villani: the analytic-norm framework

The breakthrough is to track regularity that moves with free transport, then absorb the echo loss in a Newton-type quadratic scheme.

Definition.Hybrid analytic norm.

For \(\lambda,\mu>0\) and \(\tau\ge 0\),

\[ \|h\|_{\mathcal{Z}^{\lambda,\mu;\tau}} \;=\; \sum_{k\in\mathbb{Z}^d}\sum_{n\in\mathbb{N}^d} \frac{\lambda^{|n|}\mu^{n!}}{n!}\, e^{2\pi\mu|k|}\,\bigl\|(\nabla_v + 2\pi i\tau k)^n\,\widehat h(k,\cdot)\bigr\|_{L^2}. \]

The “gliding” derivative \(\nabla_v + 2\pi i \tau k\) is the velocity gradient as seen in the frame of free transport at time \(\tau\); it captures filamentation as smoothness rather than oscillation.

§ The Newton scheme

Write the solution as \(f = f^0 + h^{(1)} + h^{(2)} + \cdots\), where each \(h^{(n+1)}\) solves the linearization about \(f^0 + h^{(1)} + \cdots + h^{(n)}\) with a quadratic source. Two estimates make the scheme close:

Estimate.(linear damping, perturbed background)

For Penrose-stable analytic backgrounds within distance \(\delta\)of \(f^0\):

\[ \|F[h^{(n+1)}]\|_{\lambda',\mu';t} \;\le\; C\,\|S^{(n)}\|_{\lambda,\mu;t}\, e^{-\,\min(\lambda,\mu)\,t}, \]

where \(S^{(n)}\) is the Newton source built from \(h^{(1)},\ldots,h^{(n)}\), and \(\lambda',\mu'<\lambda,\mu\).

Estimate.(echo control)

With shrinking analyticity radii \(\lambda_n\downarrow\lambda_\infty>0\), \(\mu_n\downarrow\mu_\infty>0\) chosen geometrically, the echo-induced derivative loss is summable and absorbed quadratically:

\[ \delta_{n+1} \;\le\; C\,\delta_n^{\,2}\,(\lambda_n-\lambda_{n+1})^{-A}, \qquad \delta_n := \|h^{(n)}\|_{\mathcal{Z}^{\lambda_n,\mu_n}}. \]

For \(\delta_0\) small enough the recursion converges super-exponentially.

Theorem.(Mouhot–Villani, 2011)

Let \(W\) be a regular interaction (e.g. Coulomb on \(\mathbb{T}^d\)), \(f^0\) analytic and Penrose-stable, and let \(f_i\) be initial data with

\[ \|f_i - f^0\|_{\mathcal{Z}^{\lambda_0,\mu_0;0}} \;\le\; \varepsilon,\qquad \varepsilon \ll 1. \]

Then there exists \(f^\infty(v)\), close to \(f^0\), and \(\lambda',\mu'>0\), such that the Vlasov–Poisson solution \(f(t,x,v)\) satisfies, in the gliding analytic norm,

\[ \bigl\|\,f(t,x+vt,v) \;-\; f^\infty(v)\,\bigr\|_{\mathcal{Z}^{\lambda',\mu';t}} \;\le\; C\,e^{-\,\lambda'' t}. \]

In particular, the force \(F(t,x)\) and density fluctuations decay exponentially.

The translation \(x\mapsto x+vt\) places the observer in the free-streaming frame: in that frame \(f\) doesconverge, in a strong analytic sense, to a velocity profile \(f^\infty(v)\). In the laboratory frame \(f\)filaments forever; the macroscopic field is what relaxes.

III.

Optimal transport & synthetic Ricci curvature

§ The Monge–Kantorovich problem

Given probability measures \(\mu,\nu\) on a Polish space \(X\) with a cost \(c:X\times X\to\mathbb{R}_+\), Kantorovich’s relaxation of Monge’s problem reads

\[ \mathcal{W}_c(\mu,\nu) \;=\; \inf_{\pi\in\Pi(\mu,\nu)} \int_{X\times X} c(x,y)\,d\pi(x,y), \]

where \(\Pi(\mu,\nu)\) is the set of couplings with marginals \(\mu,\nu\). For \(c(x,y)=d(x,y)^2\) on a Riemannian manifold \((M,g)\) the square root

\[ W_2(\mu,\nu) \;=\; \Bigl(\inf_{\pi}\!\int d(x,y)^2\,d\pi\Bigr)^{1/2} \]

defines a metric on the space \(\mathcal{P}_2(M)\) of probability measures with finite second moment.

Theorem.(Brenier / McCann)

On a smooth Riemannian manifold, if \(\mu\) is absolutely continuous with respect to volume, the optimal coupling is concentrated on the graph of a unique map \(T:M\to M\) of the form

\[ T(x) \;=\; \exp_x(-\,\nabla\varphi(x)), \]

for some \(d^2/2\)-concave potential \(\varphi\). The geodesic in \((\mathcal{P}_2,W_2)\) joining \(\mu\) to \(\nu\) is

\[ \mu_t \;=\; \bigl(\,\exp_x(-t\,\nabla\varphi(x))\,\bigr)_\#\,\mu,\qquad t\in[0,1]. \]

§ Otto’s formal Riemannian structure

Felix Otto observed that \(W_2\) endows \(\mathcal{P}_2(M)\) with a formal Riemannian metric: the tangent space at \(\mu\) is \(\{\nabla\psi : \psi\in C^\infty\}\) with norm

\[ \|\nabla\psi\|_\mu^2 \;=\; \int_M |\nabla\psi|^2\,d\mu. \]

Many evolution PDEs become \(W_2\)-gradient flows of natural functionals. In particular,

\[ \partial_t \rho \;=\; \Delta\rho \;+\; \nabla\!\cdot(\rho\,\nabla V) \quad\Longleftrightarrow\quad \dot\mu \;=\; -\,\nabla_{W_2}\!\bigl(\mathrm{Ent}(\mu)+\textstyle\int V d\mu\bigr). \]

§ Curvature ↔ entropy convexity

Define the Boltzmann–Shannon entropy

\[ \mathrm{Ent}(\mu\,\vert\,\mathrm{vol}) \;=\; \int \rho\log\rho\,d\mathrm{vol} \quad\text{if } \mu = \rho\,\mathrm{vol};\qquad +\infty \text{ otherwise.} \]

Theorem.(Otto–Villani, Cordero-Erausquin–McCann–SchmuckenschlĂ€ger, von Renesse–Sturm)

For a smooth complete Riemannian manifold \(M\) and \(K\in\mathbb{R}\), the following are equivalent:

(i) \(\mathrm{Ric}_M \;\ge\; K\,g\);

(ii) \(\mathrm{Ent}(\,\cdot\,\vert\,\mathrm{vol})\) is \(K\)-convex along every \(W_2\)-geodesic, i.e. for every \(W_2\)-geodesic \((\mu_t)_{t\in[0,1]}\),

\[ \mathrm{Ent}(\mu_t) \;\le\; (1-t)\,\mathrm{Ent}(\mu_0) + t\,\mathrm{Ent}(\mu_1) - \tfrac{K}{2}\,t(1-t)\,W_2(\mu_0,\mu_1)^2. \]

§ The Lott–Villani synthetic definition

Now invert the logic: take (ii) — a property of \(\mathrm{Ent}\)and \(W_2\) — as a definition of “Ricci curvature \(\ge K\)” on a metric–measure space \((X,d,m)\), where no smooth structure exists. Sturm independently proposed the same definition. To handle dimension, one refines to the curvature–dimension condition:

Definition.CD(K,N) (Lott–Sturm–Villani)

A metric–measure space \((X,d,m)\) satisfies \(\mathrm{CD}(K,N)\) for \(K\in\mathbb{R}\), \(N\in[1,\infty]\), if for every pair \(\mu_0,\mu_1\in\mathcal{P}_2(X)\) absolutely continuous w.r.t. \(m\), there is a \(W_2\)-geodesic \((\mu_t)\) along which the RĂ©nyi-type entropy

\[ S_N(\mu \,\vert\, m) \;=\; -\,\int \rho^{1-1/N}\,dm \qquad (\mu = \rho\,m) \]

satisfies a distortion-coefficient convexity inequality with parameters \((K,N)\) — specializing to \(K\)-convexity of \(\mathrm{Ent}\) when \(N=\infty\).

§ Why this matters

The synthetic notion is stable under measured Gromov–Hausdorff limits, where smoothness is not preserved. It thereby provides a framework for: (a) limits of Riemannian manifolds with uniform Ricci bounds, (b) singular spaces such as Alexandrov spaces and Ricci-limit spaces, (c) geometric analysis on graphs and discrete networks (with appropriate adaptation), and (d) sharp functional inequalities (Brunn–Minkowski, log-Sobolev, HWI) that follow uniformly from the abstract definition. Many classical theorems — Bonnet–Myers, Bishop–Gromov, Lichnerowicz — extend verbatim to \(\mathrm{CD}(K,N)\) spaces.

Remark.Villani’s Optimal Transport: Old and New (Springer, 2009, ~1000 pp.) catalogs this entire theory and is the canonical reference. The synthetic curvature framework was subsequently refined by Ambrosio–Gigli–SavarĂ© into the \(\mathrm{RCD}(K,N)\) condition, isolating the “Riemannian-like” part of \(\mathrm{CD}\).

IV.

Canonical expansions — sharper versions and surrounding theorems

The three pillars above are expanded here with sharper statements and structural theorems drawn from the primary literature: Otto–Villani (J. Funct. Anal. 2000), Mouhot–Villani (Acta Math. 2011), Desvillettes–Villani (Invent. Math. 2005), and Villani’s monograph Optimal Transport: Old and New (Grundlehren 338, Springer 2009).

§ 1. Brenier’s polar factorization — the foundation stone

The optimal transport map for quadratic cost was identified by Brenier (1991) in a theorem now central to all later theory. Let \(\mu,\nu\) be probability measures on \(\mathbb{R}^n\) with \(\mu\) absolutely continuous w.r.t. Lebesgue measure, and \(\int|x|^2 d\mu, d\nu < \infty\). Then there exists a \(\mu\)-a.e. unique map

\[ T = \nabla \varphi, \qquad \varphi:\mathbb{R}^n\to\mathbb{R}\ \text{convex},\ T_\#\mu=\nu, \]

attaining the Wasserstein-2 distance: \(W_2^2(\mu,\nu) = \int |x-T(x)|^2 d\mu(x)\). The convex potential \(\varphi\) is unique up to additive constants. Polar factorization: every Borel map \(s:\Omega\to\mathbb{R}^n\) with non-degenerate image factors as

\[ s = (\nabla\varphi)\circ u, \qquad u\ \text{measure-preserving},\ \varphi\ \text{convex}, \]

the convex part being unique. This is a nonlinear, infinite-dimensional analogue of the matrix polar decomposition \(M = QS\) (orthogonal \(\times\)symmetric-positive). Under Brenier’s map the Monge–Ampùre equation \(\det D^2\varphi = \rho_\mu/\rho_\nu(\nabla\varphi)\) encodes the transport.

§ 2. McCann’s displacement convexity

Given two probabilities \(\mu_0,\mu_1\) on \(\mathbb{R}^n\)with Brenier map \(T=\nabla\varphi\) from \(\mu_0\) to \(\mu_1\), the displacement interpolation is

\[ \mu_t = \big((1-t)\,\mathrm{Id} + t\,\nabla\varphi\big)_\# \mu_0, \qquad t\in[0,1]. \]

A functional \(\mathcal{F}:\mathcal{P}_2\to\mathbb{R}\) is displacement convex if \(t\mapsto\mathcal{F}(\mu_t)\) is convex along every such path. McCann (1997) proved that the internal-energy functionals

\[ \mathcal{U}(\mu) = \int U(\rho(x))\,dx, \qquad \mu = \rho\,dx, \]

are displacement convex iff \(r\mapsto r^n U(r^{-n})\) is convex non-increasing on \((0,\infty)\). Examples: Boltzmann entropy \(U(\rho)=\rho\log\rho\) (always convex); \(U(\rho)=\rho^p/(p-1)\) for \(p\ge 1-1/n\)(Rényi-type); and the negative Newtonian energy \(U(\rho)=-\rho^{1-1/n}\)which controls the \(\mathrm{CD}(0,n)\) criterion in Section III.

§ 3. Otto’s Riemannian calculus on Wasserstein space

On \(\mathcal{P}_2(\mathbb{R}^n)\) with the Wasserstein metric, Otto (2001) introduced a formal Riemannian structure: the tangent space at \(\mu=\rho\,dx\) consists of perturbations \(\delta\mu=-\nabla\!\cdot(\rho\,\nabla\psi)\) parametrized by potentials \(\psi\), with inner product

\[ \langle \nabla\psi_1,\,\nabla\psi_2\rangle_\mu \;=\; \int \nabla\psi_1\cdot\nabla\psi_2 \,d\mu. \]

In this metric, the heat equation \(\partial_t\rho = \Delta\rho\) is the gradient flow of the Boltzmann entropy \(\mathrm{Ent}(\mu)=\int\rho\log\rho\,dx\): this is the celebrated JKO scheme (Jordan–Kinderlehrer–Otto 1998) viewed through the Otto lens. Higher-order PDEs (porous-medium, drift-diffusion, fourth-order DLSS) likewise become gradient flows of explicit functionals on \((\mathcal{P}_2,W_2)\).

The Hessian of \(\mathrm{Ent}\) in this geometry is, on a manifold \((M,g)\), formally \(\mathrm{Hess}_\mu\mathrm{Ent}(\nabla\psi,\nabla\psi)=\int(\Vert\mathrm{Hess}\,\psi\Vert^2+\mathrm{Ric}(\nabla\psi,\nabla\psi))d\mu\). Bochner identifies this with the Bakry–Émery \(\Gamma_2\)-tensor; the lower bound by \(K\)corresponds exactly to \(\mathrm{Ric}\ge K g\). This is the dictionary that makes Lott–Sturm–Villani possible.

§ 4. The HWI inequality (Otto–Villani 2000)

Let \(d\nu = e^{-V}dx\) be a probability measure on \(\mathbb{R}^n\) with \(\mathrm{Hess}\,V\ge K I\) for some \(K\in\mathbb{R}\). For \(\mu\) absolutely continuous w.r.t. \(\nu\), define

  • \(\mathbf{H}(\mu\,\vert\,\nu) \;=\; \int \tfrac{d\mu}{d\nu}\log\tfrac{d\mu}{d\nu}\,d\nu\)  —  relative entropy.
  • \(\mathbf{W}(\mu,\nu) \;=\; W_2(\mu,\nu)\)  —  Wasserstein-2 distance.
  • \(\mathbf{I}(\mu\,\vert\,\nu) \;=\; \int \big|\nabla\log\tfrac{d\mu}{d\nu}\big|^2 d\mu\)  —  relative Fisher information.

Then the HWI inequality asserts

\[ \boxed{\;H(\mu\,\vert\,\nu) \;\le\; W(\mu,\nu)\sqrt{I(\mu\,\vert\,\nu)} \;-\; \tfrac{K}{2}\,W^2(\mu,\nu)\;} \]

When \(K>0\), applying the elementary inequality \(ab\le \tfrac{a^2}{2K}+\tfrac{Kb^2}{2}\) to the right-hand side eliminates \(W\) and recovers the log-Sobolev inequality

\[ H(\mu\,\vert\,\nu) \;\le\; \frac{1}{2K}\,I(\mu\,\vert\,\nu), \]

originally due to Bakry–Émery. The HWI inequality unifies log-Sobolev, Talagrand transport (\(W^2\le 2H/K\)), and PoincarĂ© into a single displacement-convex framework, and admits curved-space versions (\(K\)becomes the Bakry–Émery Ricci lower bound).

§ 5. Talagrand T2 and concentration of measure

A probability measure \(\nu\) satisfies a Talagrand T2(K) inequality if for every probability \(\mu\ll\nu\),

\[ W_2^2(\mu,\nu) \;\le\; \frac{2}{K}\,H(\mu\,\vert\,\nu). \]

Otto–Villani show LSI(K) \(\Rightarrow\) T2(K)by a gradient-flow argument along the Fokker–Planck equation \(\partial_t\rho = \nabla\!\cdot(\rho\,\nabla\log(\rho/e^{-V}))\), integrating the dissipation identity \(\frac{d}{dt}H = -I\) along a Wasserstein geodesic. Coupled with Marton’s tensorisation, T2 implies Gaussian concentration of measure: for any 1-Lipschitz \(F\),

\[ \nu\big(\,F\ge \mathbb{E}_\nu F + r\,\big) \;\le\; e^{-Kr^2/2}. \]

§ 6. Bakry–Émery \(\Gamma_2\) calculus and \(\mathrm{CD}(K,N)\)

For a Markov diffusion semigroup with generator \(L = \Delta - \nabla V\cdot\nabla\) on \(\mathbb{R}^n\)(or \((M,g)\)), define the carré-du-champ operators

\[ \Gamma(f,g)=\tfrac{1}{2}(L(fg)-fLg-gLf), \qquad \Gamma_2(f,f)=\tfrac{1}{2}(L\Gamma(f,f)-2\Gamma(f,Lf)). \]

By the Bochner identity, on a Riemannian manifold, \(\Gamma_2(f,f)=\Vert\mathrm{Hess}\,f\Vert^2+(\mathrm{Ric}+\mathrm{Hess}\,V)(\nabla f,\nabla f)\). The Bakry–Émery condition \(\mathrm{CD}(K,N)\) is the pointwise inequality

\[ \Gamma_2(f,f) \;\ge\; K\,\Gamma(f,f) \;+\; \frac{1}{N}(Lf)^2. \]

For the smooth setting this is equivalent to the Lott–Sturm–Villani displacement-convexity formulation of Section III — a remarkable agreement between two seemingly distant frameworks (semigroup-spectral vs. transport-geodesic). The bridge is what allowed Sturm and Lott–Villani to define Ricci bounds on metric–measure spaces with no smooth structure.

§ 7. Synthetic Bishop–Gromov volume comparison

On a metric–measure space \((X,d,m)\) satisfying \(\mathrm{CD}(K,N)\) with \(N<\infty\), the volume-of-balls function obeys

\[ r \;\longmapsto\; \frac{m(B_r(x))}{v_{K,N}(r)} \quad\text{is non-increasing in }r, \]

where \(v_{K,N}(r)\) is the volume of an \(N\)-dimensional model ball of constant sectional curvature \(K/(N-1)\). Specialised:

  • \(K=0\): \(m(B_r)/r^N\) is non-increasing — recovers Euclidean volume growth in dimension \(N\).
  • \(K>0\): diameter bound (synthetic Bonnet–Myers): \(\mathrm{diam}(X)\le \pi\sqrt{(N-1)/K}\).
  • \(K<0\): exponential growth bound, hyperbolic-model comparison.

These extend the smooth Bishop–Gromov theorem and Bonnet–Myers to limit spaces, sub-Riemannian geometries, and Alexandrov spaces — a payoff of stability under measured Gromov–Hausdorff convergence.

§ 8. Generalised Brunn–Minkowski inequality

On a \(\mathrm{CD}(K,N)\) space, for any pair of bounded measurable sets \(A_0,A_1\subset X\) of positive measure, with \(A_t\) the set of \(t\)-midpoints (i.e. points \(\gamma(t)\) of geodesics joining \(A_0\) to \(A_1\)),

\[ m(A_t)^{1/N} \;\ge\; \tau_{K,N}^{(1-t)}(\theta)\, m(A_0)^{1/N} \;+\; \tau_{K,N}^{(t)}(\theta)\, m(A_1)^{1/N}, \]

where \(\theta\) is the geodesic distance between the sets and \(\tau_{K,N}^{(t)}\) are the model distortion coefficients (\(\sin\)/\(\sinh\) ratios in the constant-curvature models). Setting \(K=0\), \(\tau^{(t)}=t\) recovers the classical Euclidean Brunn–Minkowski \(m(A_t)^{1/N}\ge (1-t)m(A_0)^{1/N}+t\,m(A_1)^{1/N}\).

§ 9. Cercignani’s conjecture: failure and weak forms

Cercignani conjectured that the entropy production \(D(f) = -\frac{d}{dt}\big|_{t=0}\!H(f_t)\) controls the relative entropy linearly: \(D(f) \ge \kappa \, H(f\,\vert\,M)\). Bobylev’s counter-construction (1988) showed this fails even for Maxwellian molecules when \(f\) concentrates on shells far from \(M\). Villani’s weak Cercignani inequality (Comm. Math. Phys. 2003) replaces it by an \(\varepsilon\)-degraded form valid under bounded entropy and moments:

\[ D(f) \;\ge\; C_\varepsilon(f)\,H(f\,\vert\,M)^{1+\varepsilon}, \quad \varepsilon>0\ \text{arbitrary,} \]

with constant \(C_\varepsilon\) depending only on a few moments and regularity of \(f\). This is what Desvillettes–Villani feed into the \(O(t^{-\infty})\) convergence proof: the loss of one power of \(\varepsilon\) is harmless after coupling with hypocoercive regularisation.

§ 10. Quantitative damping rates and Gevrey regularity

The Mouhot–Villani theorem (Acta Math. 2011) is sharp in the analytic class. For initial data \(f_0\) with deviation \(\Vert f_0-f_0^{\,\circ}\Vert_{\lambda,\mu} \le \delta_0\) from a stable equilibrium \(f_0^{\,\circ}\), the spatial density obeys

\[ \sup_{x\in\mathbb{T}^d} \big|\rho(t,x)-\rho^\circ\big| \;\le\; C\,e^{-\lambda' |t|}, \qquad \lambda' < \lambda, \]

with the rate \(\lambda'\) depending on the spectral gap of the Penrose-stable background. Bedrossian–Masmoudi–Mouhot (2016) subsequently extended to Gevrey-\(s\)regularity for \(s<3\); Bedrossian (2020) showed theborderline \(s=3\) fails — making Gevrey-3 the natural boundary, exactly the threshold suggested by the plasma-echo resonance count \(\sum k\sim k^3\) in the Mouhot–Villani estimates.

§ 11. Hypocoercivity — the missing ingredient

Convergence to equilibrium for kinetic equations is obstructed by the degeneracy of dissipation: the Boltzmann collision operator only damps velocity inhomogeneities, not spatial ones. Villani’s hypocoercivity programme (MĂ©m. AMS 2009) builds modified Lyapunov functionals

\[ \mathcal{E}(f) \;=\; \Vert f\Vert^2 + a\langle\nabla_v f,\nabla_v f\rangle + b\langle\nabla_v f,\nabla_x f\rangle + c\langle\nabla_x f,\nabla_x f\rangle \]

with \(a,b,c>0\) chosen so that the cross-term \([\nabla_v,v\cdot\nabla_x]=\nabla_x\) forces dissipation to propagate from velocity into position. The construction generalises to Fokker–Planck, kinetic Fokker–Planck, and the linearised Boltzmann equation, yielding exponential rates whenever the spatial PoincarĂ© constant is finite. This is the technical engine behind both the Boltzmann theorem of Section I and the probabilistic interpretation of optimal transport flow.

§ 12. Heat flow on \(\mathrm{RCD}(K,N)\) spaces — the modern synthesis

Ambrosio, Gigli & Savaré (2014) showed that on \(\mathrm{CD}(K,\infty)\) spaces the gradient flow of the entropy in \((\mathcal{P}_2,W_2)\) agrees with the heat flow defined via the Cheeger energy \(\mathrm{Ch}(f)=\tfrac{1}{2}\int|\nabla f|^2 dm\). The additional linearity of this heat flow defines the Riemannian curvature-dimension \(\mathrm{RCD}(K,N)\) class. On RCD spaces:

  • The Bochner inequality \(\Gamma_2(f)\ge K\Gamma(f)+(Lf)^2/N\) holds in a weak form.
  • Sobolev and PoincarĂ© inequalities admit sharp model-comparison constants.
  • Splitting theorems (Cheeger–Gromoll, Cheeger–Colding) generalise to limit spaces.
  • The rectifiable structure (Mondino–Naber 2019) is partially recovered.

The Lott–Sturm–Villani definition is therefore not an isolated curiosity: it sits inside a self-consistent calculus that recovers, in the smooth case, every classical Riemannian theorem driven by Ricci curvature, and in the singular case furnishes the only known framework where those theorems still apply.

Coda — canonical sources.These twelve expansions trace the proximate canonical literature: Brenier 1991 (Comm. Pure Appl. Math.), McCann 1997 (Adv. Math.), Otto 2001 (Comm. PDEs), Otto–Villani 2000, Jordan–Kinderlehrer–Otto 1998, Villani 2003 (Topics in Optimal Transportation) and 2009 (Optimal Transport: Old and New), Sturm 2006 (Acta Math.), Lott–Villani 2009, Ambrosio–Gigli–SavarĂ© 2014 (Invent. Math.), Bedrossian–Masmoudi–Mouhot 2016 (Ann. PDE), Bedrossian 2020 (Comm. AMS). The unifying instinct: information, geometry and dynamics are three faces of one displacement-convex object.

Selected references

  1. L. Desvillettes & C. Villani, On the trend to global equilibrium for spatially inhomogeneous kinetic systems: the Boltzmann equation, Invent. Math. 159 (2005), 245–316.
  2. C. Mouhot & C. Villani, On Landau damping, Acta Math. 207 (2011), 29–201.
  3. J. Lott & C. Villani, Ricci curvature for metric-measure spaces via optimal transport, Ann. of Math. 169 (2009), 903–991.
  4. C. Villani, Optimal Transport: Old and New, Grundlehren 338, Springer, 2009.
  5. F. Otto & C. Villani, Generalization of an inequality by Talagrand and links with the logarithmic Sobolev inequality, J. Funct. Anal. 173 (2000), 361–400.
  6. Y. Brenier, Polar factorization and monotone rearrangement of vector-valued functions, Comm. Pure Appl. Math. 44 (1991), 375–417.
  7. R. J. McCann, A convexity principle for interacting gases, Adv. Math. 128 (1997), 153–179.
  8. F. Otto, The geometry of dissipative evolution equations: the porous medium equation, Comm. PDEs 26 (2001), 101–174.
  9. R. Jordan, D. Kinderlehrer & F. Otto, The variational formulation of the Fokker–Planck equation, SIAM J. Math. Anal. 29 (1998), 1–17.
  10. K.-T. Sturm, On the geometry of metric measure spaces I & II, Acta Math. 196 (2006), 65–131 & 133–177.
  11. L. Ambrosio, N. Gigli & G. SavarĂ©, Metric measure spaces with Riemannian Ricci curvature bounded from below, Duke Math. J. 163 (2014), 1405–1490.
  12. D. Bakry & M. Émery, Diffusions hypercontractives, SĂ©m. de ProbabilitĂ©s XIX, Lect. Notes Math. 1123, Springer, 1985, 177–206.
  13. C. Villani, Cercignani’s conjecture is sometimes true and always almost true, Comm. Math. Phys. 234 (2003), 455–490.
  14. C. Villani, Hypocoercivity, Mem. Amer. Math. Soc. 202 (2009), no. 950.
  15. J. Bedrossian, N. Masmoudi & C. Mouhot, Landau damping in finite regularity for unconfined systems with screened interactions, Comm. Pure Appl. Math. 71 (2018), 537–576.
  16. J. Bedrossian, Nonlinear echoes and Landau damping with insufficient regularity, Tunis. J. Math. 3 (2021), 121–205.


Three theorems, one geometric instinct: the world prefers to be smooth, in some norm, in some frame.


Share:XRedditLinkedIn