The entropy formula for the Ricci flow and its geometric applications

Grisha Perelman

Source-linked AI summary

The entropy formula for the Ricci flow and its geometric applications

Grisha Perelman

arXiv:math/0211159v1Nov 11, 2002math.DG

TL;DR

The paper asks how Ricci flow can be understood as a gradient-like evolution and used in Hamilton’s program for geometrization. It develops monotonicity and related geometric arguments, proving no local collapsing at finite singular times and supporting a Ricci-flow-based geometrization strategy.

Problem
Hamilton’s program required further Ricci-flow details, including control of collapsing and singular regions, to support geometrization of closed three-manifolds.
Method
The paper develops a monotonicity formula, interprets Ricci flow as gradient-like, and applies it with logarithmic Sobolev and comparison arguments.
Results
The paper proves finite-time no-local-collapsing results and uses them to control injectivity radius and advance a Ricci-flow-based geometrization argument.
Takeaways & Limitations
Ricci flow supplies geometric control near finite-time singularities and supports topological conclusions within Hamilton’s geometrization program.
Takeaways & Limitations
The paper does not confirm bounded normalized curvature for all-time solutions, and some arguments assume unique shortest L-geodesics or require barrier interpretations.

Abstract

from arXiv · show

We present a monotonic expression for the Ricci flow, valid in all dimensions and without curvature assumptions. It is interpreted as an entropy for a certain canonical ensemble. Several geometric applications are given. In particular, (1) Ricci flow, considered on the space of riemannian metrics modulo diffeomorphism and scaling, has no nontrivial periodic orbits (that is, other than fixed points); (2) In a region, where singularity is forming in finite time, the injectivity radius is controlled by the curvature; (3) Ricci flow can not quickly turn an almost euclidean region into a very curved one, no matter what happens far away. We also verify several assertions related to Richard Hamilton's program for the proof of Thurston geometrization conjecture for closed three-manifolds, and give a sketch of an eclectic proof of this conjecture, making use of earlier results on collapsing with local lower curvature bound.

Introduction · 1 Ricci flow as a gradient flow

The paper establishes Ricci flow as a gradient-like evolution and develops monotonicity-based tools supporting Hamilton’s geometrization program. It connects the gradient formulation to diffeomorphism-gauge freedom and applications including periodic-orbit exclusion, singularity analysis, and injectivity-radius control.

Introduction: Finite-time singularities may form through pinching almost round cylindrical necks, motivating surgery by cutting the neck, attaching caps, and continuing Ricci flow [H 5].Hamilton described this procedure for four-manifolds under curvature assumptions and hoped for a three-dimensional analogue.
Introduction: The paper develops details of Hamilton’s program, showing that regions lacking bounded normalized curvature are locally collapsed with curvature bounded below, which suffices for topological conclusions.The technically more complicated surgery arguments are deferred, but the stated collapsing result advances the geometrization strategy.
Introduction: Ricci flow is presented as gradient-like, and modulo diffeomorphisms and scaling it has no nontrivial periodic orbits.The paper attributes this conclusion to a main monotonicity formula and the Gaussian logarithmic Sobolev inequality due to L.Gross.
Introduction: The monotonicity formula also controls the injectivity radius at each point by curvatures at nearby points on any smooth finite-time solution.The introduction identifies this result as removing a major stumbling block in Hamilton’s approach to geometrization.
1 Ricci flow as a gradient flow: For a fixed measure m=e^-f dV, the symmetric tensor −(R_ij+∇_i∇_j f) is the L2 gradient of F_m=∫(R+|∇f|^2)dm.The first variation is taken with respect to metric and function variations while the measure-preserving condition is v/2−h=0.
1 Ricci flow as a gradient flow: The resulting gradient flow is (g_ij)_t=−2(R_ij+∇_i∇_j f), and whenever it exists it equals Ricci flow modified by a diffeomorphism.Different choices of m produce the same flow up to diffeomorphism, so the measure acts analogously to a gauge choice.
1.2. Proposition.: The weighted functional F_m has a Bochner–Lichnerowicz interpretation for smooth measures, while DeTurck’s diffeomorphism modification makes Ricci flow strongly parabolic [H 4,§6].The weighted formulas replace the ordinary curvature terms with expressions involving the smooth measure and its associated operators.

2 No breathers theorem I

The section uses monotone functionals to rule out nontrivial steady and expanding Ricci-flow breathers on closed manifolds. It also shows that steady and expanding Ricci solitons must have constant Ricci curvature.

2.1: Consequently, Ricci flow has no nontrivial steady or expanding breathers on closed M, and the corresponding dynamical-system periodic orbits are absent.Breathers correspond to periodic orbits after quotienting metrics by diffeomorphism and scaling, while Ricci solitons correspond to fixed points.
2.2: Steady breathers are necessarily steady solitons because λ(g) is nondecreasing, with equality only when Rij + ∇i∇jf = 0.Here λ is the lowest eigenvalue of −4△ + R, obtained by minimizing F under the normalization ∫_M e^−f dV = 1.
2.3: For expanding breathers, the scale-invariant functional λ̄(g) = λ(g)V(g)^(2/n) is nondecreasing when nonpositive, strictly unless the flow is a gradient soliton.An expanding breather must have dV/dt > 0 somewhere, while −∫_M R dV ≥ λ(t), forcing λ̄ to be nonpositive at some time.
2.4*: Equality conditions show that every steady or expanding Ricci soliton on closed M has constant Ricci curvature.The equality case gives R + △f constant, while the Euler–Lagrange equation yields △f − |∇f|2 = 0.

3 No breathers theorem II

Introducing a scale-dependent W-functional yields a monotone ν along Ricci flow and reduces the shrinking no-breathers result to the behavior of µ as τ approaches zero. The argument also connects this monotonicity to logarithmic Sobolev inequalities on shrinking solitons.

3.1: The scale-dependent W-functional is invariant under simultaneous scaling of τ and g_ij, and ν(g_ij(t)) is nondecreasing along Ricci flow.For closed manifolds, µ(g_ij,τ) has a smooth minimizer; ν is obtained by infimizing µ over positive τ.
3.1: µ(g_ij,τ) is negative for sufficiently small positive τ and tends to zero as τ tends to zero, establishing the key claim for shrinking no-breathers.The proof uses short-time Ricci flow and the conjugate heat equation, while the limiting argument invokes the Gaussian logarithmic Sobolev inequality due to L.Gross.
3.2: The monotonicity formula implies that shrinking Ricci solitons satisfy a logarithmic Sobolev inequality, with minimizers obeying Ric_ij + ∇_i∇_j f − g_ij = 0.This requires minimizer existence and justified integration by parts; it is straightforward on closed manifolds and requires additional work on some complete manifolds.
3.3*: In dimension three, Ivey proved the no-breathers theorem and ruled out nontrivial Ricci solitons using an almost nonnegative curvature estimate.Logarithmic Sobolev inequalities are surveyed in, with curvature effects discussed by Bakry-Emery and geometric-evolution applications in Ecker [E 1].

4 No local collapsing theorem I

The monotonicity formula rules out local collapsing at finite-time singularities on closed manifolds. Consequently, Ricci-flow solutions remain uniformly noncollapsed at controlled scales, and suitable curvature blow-ups converge to complete ancient, κ-noncollapsed solutions.

4.1: Finite-time Ricci flow on a closed manifold is not locally collapsing: no sequence of balls can satisfy r_k^−2 Vol(B_k) → 0 as t_k → T.Local collapsing is defined using balls B_k = B(p_k,r_k) at times t_k approaching T.
4.1: The monotonicity formula (3.4) yields the noncollapsing theorem by converting hypothetical collapsing balls into a contradiction through an entropy quantity tending to −∞.The proof assumes collapsing balls and applies the monotonicity formula to compare the entropy across times.
4.2: For every closed-manifold initial metric and finite T, the flow is κ-noncollapsed on scale T^1/2 for all t ∈ [0,T), for some κ(g_ij,T)>0.κ-noncollapsed means balls with |Rm| ≤ r^−2 have volume at least κr^n.
4.2: If curvature blows up at points approaching T while remaining bounded by C Q_k earlier, rescaled time slices subconverge to complete ancient solutions κ-noncollapsed on every scale.Here Q_k = |Rm|(p_k,t_k) → ∞, and the convergence follows using Hamilton’s convergence theorem.

5 A statistical analogy

The section interprets the Ricci-flow functional W as analogous to minus entropy in a canonical statistical ensemble. It establishes nonnegativity properties for the associated variance, average energy, and entropy under specified flow and concentration conditions.

5 A statistical analogy: W is interpreted as analogous to minus entropy through a canonical-ensemble formulation of Ricci flow.The analogy is developed using the partition function, average energy, entropy, and variance of statistical mechanics.
5.1: The statistical quantities are constructed from a temperature-dependent metric, probability measure, and partition function associated with Ricci flow.The metric evolves according to (gij)τ = 2(Rij + ∇i∇jf), with dm = udV and u = (4πτ)^−n/2 e−f.
5.1: σ is nonnegative and vanishes only on a gradient shrinking soliton, while < E > is nonnegative when the flow exists for all sufficiently small τ > 0.These properties connect the statistical quantities to geometric rigidity and long-time flow existence.
5.1: S is nonnegative when u concentrates to a δ-function as τ → 0, or through concentrating approximating functions; its limiting behavior differs between the two cases.In the first case, < E >, S, and σ all tend to zero; in the second, S may approach a positive limit as the metric becomes singular.
5.1: For ancient Ricci-flow solutions, the section asks whether bounded entropy S as τ → ∞ forces the flow to be a gradient shrinking soliton.This is posed as a natural question rather than established as a result.

6 Riemannian formalism in potentially infinite dimensions

The section embeds backward Ricci flow into a high-dimensional Riemannian manifold whose curvature and Laplacian encode Harnack expressions and heat equations. This construction also identifies entropy with hypersurface scalar curvature and motivates reduced-volume monotonicity.

6.1: The manifold ˜M = M × S^N × R+ uses a metric evolving by backward Ricci flow on M and constant-curvature geometry on S^N.Its curvature components agree modulo N^-1 with Hamilton’s Harnack expression, while the full Ricci tensor vanishes modulo N^-1.
6.1: The heat and conjugate heat equations on M become Laplace equations on ˜M for extended functions and volume forms, respectively.Both equivalences hold modulo N^-1.
6.2: After diffeomorphic modification, the construction yields a metric representing backward m-preserving Ricci flow, with the hypersurface volume form weighted by τ^(N/2)e^-f.The induced hypersurface scalar curvature contains the entropy integrand, so βS is essentially minus its total scalar curvature modulo N^-1.
6.3: The Ricci-flat embedding converts metric spheres near τ = 0 into hypersurfaces of nearly constant τ, whose geometry is expressible through quantities on M.Geodesics are orthogonal to the S^N fibres, and their lengths involve the reduced-distance integrand.
6.3: The resulting reduced volume ˜V(τ) is suggested to increase as τ decreases, with rigorous monotonicity deferred to the next section.The formula resembles Huisken’s mean-curvature-flow monotonicity expression, but its monotonicity direction is opposite.

7 A comparison geometry approach to the Ricci flow

The section develops L-geodesic comparison theory for Ricci flow and derives monotonicity of reduced volume, with strictness except on gradient shrinking solitons. These inequalities yield controls on reduced distance and applications to noncollapsing and ancient solutions.

7.1: L-shortest curves exist between any two points, are L-geodesics, and support barrier or distributional interpretations when minimizers are nonunique.The construction assumes a closed manifold or complete metrics with uniformly bounded curvature.
7.1: The quantity 2 exp(−l(τ))J(τ) is nonincreasing along L-geodesics, strictly unless the flow is a gradient shrinking soliton, implying reduced-volume monotonicity.The result follows from Jacobian comparison and integration over the manifold.
7.1: The minimum of l(·,τ) is at most n because the minimum of L̄(·,τ) − 2nτ is nonincreasing.This provides a pointwise upper bound on reduced distance at each time.
7.2: Under nonnegative curvature operator, Hamilton’s Harnack inequalities provide additional control of reduced distance and L-Jacobi fields away from the terminal time.The estimates apply when τ ≤ (1−c)τ0 for c > 0.
7.3: The comparison inequalities give a weakened no-local-collapsing theorem by contradicting reduced-volume monotonicity when curvature is controlled on the relevant parabolic region.The argument uses the L-exponential map, separates short and long tangent vectors, and obtains a contradiction from incompatible reduced-volume bounds.
7.3: For an ancient blow-up solution, the reduced volume based at any point and any t0 > 0 remains bounded below by κ.This strengthens the corresponding corollary for ancient solutions defined on (−∞,0).

8 No local collapsing theorem II

Theorem 8.2 shows that an initially noncollapsed Ricci-flow region with controlled curvature cannot become κ-collapsed on smaller scales within a controlled surrounding region. The proof combines distance-function estimates, reduced volume, and a maximum-principle argument.

8.2 Theorem: An initially noncollapsed ball with curvature bounded by r_0^-2 cannot become κ-collapsed at nearby points on scales below r_0, for κ depending on A.The assumptions include initial volume at least A^-1r_0^n and a curvature bound in the initial r_0-ball; the conclusion applies at points within distance Ar_0 at time r_0^2.
8.2 Theorem: The argument converts hypothetical collapse into a very small reduced volume, while the initial reduced volume remains non-small unless the reduced distance is sufficiently large.This reduces the theorem to estimating the reduced distance, equivalently the reduced length, in the relevant ball.
8.3 Lemma: A distance-function lemma supplies the differential inequalities needed to control evolving distances under a local Ricci upper bound.The inequalities are understood in the barrier sense when necessary, and the lemma is proved using second variation along geodesics.
8.2 Theorem: Applying the maximum principle to a cutoff involving distance and reduced length shows that its minimum cannot decrease too rapidly, yielding the required estimate.The cutoff is chosen to equal one on an inner region and increase rapidly outside it; the proof uses the distance lemma together with the reduced-length differential inequality.

9 Differential Harnack inequality for solutions of the conjugate heat equation

This section derives a differential Harnack quantity for conjugate heat solutions under Ricci flow and establishes its monotonicity and consequences. It also relates the inequality to reduced distance and characterizes Ricci flow through conjugate heat kernels.

9 Differential Harnack inequality for solutions of the conjugate heat equation: For a conjugate heat solution u=(4π(T−t))^-n/2 e^-f, the quantity v=[(T−t)(2Δf−|∇f|^2+R)+f−n]u yields the differential Harnack inequality.The proposition is obtained by direct computation and is especially useful for local applications.
9 Differential Harnack inequality for solutions of the conjugate heat equation: On a closed manifold, or whenever the maximum principle applies, min v/u is nondecreasing in time; if u converges to a δ-function at T, then v≤0 before T.The latter conclusion follows by applying the heat equation to a positive auxiliary function and checking that the relevant terminal limit is zero.
9 Differential Harnack inequality for solutions of the conjugate heat equation: The δ-function estimate implies a curve inequality and, for concentration at p, bounds f(q,t) above by the reduced distance l(q,T−t).The reduced distance is defined using p and the backward time parameter τ(t)=T−t.
9.6 Remark: Ricci flow is characterized among metric evolution equations by the infinitesimal behavior of conjugate heat fundamental solutions, with an O(|pq|^2+|t|) error near the initial pole.For general metric evolution, the corresponding expression is only O(1), whereas Ricci flow improves it to o(1).
9.7* Remark: Differential Harnack inequalities extend earlier work by Li and Yau and Hamilton [H 7,8], who used them for linear equations and parabolic-flow monotonicity formulas.The section notes related local monotonicity results for mean curvature flow.

10 Pseudolocality theorem

The pseudolocality theorem shows that an initially almost-Euclidean region with controlled scalar curvature retains a curvature bound for a short time, independently of distant singular behavior. The proof uses conjugate heat kernels, monotonicity, and a contradiction with the logarithmic Sobolev inequality, yielding volume, curvature, and noncollapsing consequences.

10 Pseudolocality theorem: If R(x) ≥ −r0^−2 and every subregion of B(x0,r0) is almost Euclidean isoperimetrically, then |Rm|(x,t) ≤ αt^−1 + (εr0)^−2 near x0 for 0 < t ≤ (εr0)^2.The estimate holds where dist_t(x,x0) < εr0, for ε and δ depending on α.
10 Pseudolocality theorem: Almost singular regions cannot instantly transmit substantial curvature into almost-Euclidean regions under Ricci flow.Equivalently, a region that looks trivial at a higher energy scale cannot suddenly become highly nontrivial at a slightly lower scale.
10 Pseudolocality theorem: The contradiction proof rescales a potential high-curvature point and uses conjugate heat kernels whose monotonicity conflicts with the Gaussian logarithmic Sobolev inequality.If injectivity radii remain positive, the limit would be incompatible with a gradient shrinking soliton; if they collapse, rescaling produces a flat limit, and both cases lead to contradiction.
10 Pseudolocality theorem: Under the same hypotheses, the theorem also supplies a lower volume estimate for balls centered near x0 at positive times.This volume control is stated as Corollary 10.2 with a universal dimensional constant.
10 Pseudolocality theorem: A bounded initial curvature and almost-Euclidean volume imply |Rm|(x,t) ≤ (εr0)^−2 near x0 throughout the short-time interval, while related assumptions prevent κ-collapse on controlled scales.These consequences appear in Theorem 10.3 and Corollary 10.4.

11 Ancient solutions with nonnegative curvature operator and bounded entropy

This section studies complete, non-flat ancient Ricci flows with bounded nonnegative curvature operator and κ-noncollapsing, equivalently bounded entropy. It proves asymptotic and compactness properties, including shrinking-soliton blow-downs, zero asymptotic volume ratio, curvature estimates, and structural control of non-neck regions.

11.2: Under the stated curvature and entropy assumptions, parabolic rescalings at suitable points converge subsequentially to a non-flat gradient shrinking soliton.κ-noncollapsing prevents collapse, while monotonicity of the reduced volume forces equality in the limiting inequality and yields the soliton structure.
11.4: Every solution satisfying these assumptions has zero asymptotic volume ratio at each time.Positive asymptotic volume ratio would produce either a dimension-reducible split limit, a forbidden non-flat metric cone, or flatness.
Local noncollapsing: For every ε > 0, sufficiently controlled local ancient flows have a uniform lower volume bound at curvature scale.The estimate gives Vol B(x_k,A/√Q_k)^n ≥ ε for large k under the stated curvature, time-depth, and spatial-scale conditions.
Local curvature estimates: A lower volume bound on a ball yields the curvature estimate R(x,t) ≤ C r_0^-2 + B(t−t_0)^-1 nearby, even when the bound is assumed only at the final time.The estimate applies within distance 1/4 r_0, with a modified backward time range in the final-time-only case.
Global structure: The class of noncompact ancient solutions is compact modulo scaling, and points outside ε-necks form a compact region of controlled diameter and comparable curvature.After normalizing R(x_k,0)=1, sequences admit smoothly convergent subsequences; the non-neck set M_ε satisfies diam M_ε ≤ C Q^-1/2 and C^-1Q ≤ R ≤ CQ.
Uniqueness conjecture: The text conjectures that the unique noncompact three-dimensional κ-noncollapsed ancient solution with bounded positive curvature is the rotationally symmetric Bryant steady soliton.The proposed uniqueness is not fully established, although uniqueness is claimed within the class of gradient steady solitons.

12 Almost nonnegative curvature in dimension three

In dimension three, φ-almost nonnegative curvature combined with noncollapsing yields local ancient-solution models and curvature control. These results provide scale-dependent curvature estimates under volume and sectional-curvature hypotheses, culminating in a corollary for small metric balls.

12.1: High-curvature neighborhoods become, after scaling by Q, ε-close to corresponding regions of ancient solutions under φ-almost nonnegative curvature and κ-noncollapsing.This applies on a parabolic neighborhood around (x0,t0) when t0 ≥ 1 and Q = R(x0,t0) ≥ r0^-2.
12.1: The blow-up argument obtains a smooth ancient limit by controlling curvature at bounded distances, using κ-noncollapsing, extension across earlier times, and uniform curvature bounds.Claim 2 prevents arbitrarily large curvature from accumulating too close to a base point, while Claim 1 supports extension past a putative earliest limit time.
12.2: For every A > 0, scalar curvature at time 1 is bounded by K(A) within distance A of x0 under the theorem 8.2 assumptions and φ-almost nonnegative curvature.The proof reduces large-curvature points to the local ancient-solution conclusion of Theorem 12.1, with noncollapsing supplied by Theorem 8.2.
12.3: Given volume at least w r0^n on B(x0,r0), sectional curvature at least −r0^-2, and φ-almost nonnegative curvature, scalar curvature satisfies R(x,t) < K r0^-2.The constants τ(w), K(w), and ρ(w) depend only on w, as stated in Theorem 12.3.
12.4: Theorems 12.2 and 12.3 yield a small-scale corollary for φ-almost nonnegatively curved flows on closed three-manifolds when a metric ball has prescribed lower curvature scale.The supplied passage states the corollary’s setup but does not include its concluding estimate.

13 The global picture of the Ricci flow in dimension three

At large times, three-dimensional Ricci flow admits a thick–thin decomposition whose thick part approaches finite-volume hyperbolic geometry and whose thin part is a graph manifold. Finite-time singularities decompose the topology into geometrized pieces, including connected sums with spherical and S2 × R quotients.

13.1: For sufficiently large t, M splits into thick and thin regions with controlled curvature and volume on the thick part and locally collapsed, lower-curvature-bounded balls on the thin part.The decomposition depends on w, with constants K(w) and ρ(w); thick balls satisfy |˜Rm| ≤ K and have volume at least a scale-dependent lower bound, while thin balls have volume < wr^n.
13.1: Along suitable sequences with t →∞ and w →0, the thick part converges to a finite-volume complete hyperbolic manifold, while the thin part is homeomorphic to a graph manifold.Any cusps of the hyperbolic limit are incompressible in M, and the graph-manifold conclusion uses three-dimensional collapsing with a lower curvature bound.
13.2: At a finite-time singularity, either curvature diverges everywhere and M is a quotient of S3 or S2 × R, or high curvature concentrates in necks and capped necks forming horns.The horn description permits replacing horn tips at the singular time, although finite horn diameter remains unproved.
13.2: Every maximal horn has volume at least cT^n, implying eventual smoothness and allowing M’s topology to be reconstructed as a connected sum of thick–thin pieces and quotients of S3 and S2 × R.The thick–thin pieces are those described in 13.1.
13.3*: Anderson develops an alternative differential-geometric approach to geometrization through elliptic Euler–Lagrange equations for perturbed total scalar curvature functionals.The paper notes a parallel with Hamilton’s work because Ricci flow is the gradient flow of a closely related functional.