MathLabs
TheoremProved

Monotonicity of Perelman's $\mathcal{F}$-entropy

Statement

Let gij(t)g_{ij}(t) solve Ricci flow ∂tgij=−2Rij\partial_t g_{ij} = -2 R_{ij} and let f(t)f(t) solve the coupled backward heat equation ∂tf=−Δf+∣∇f∣2−R\partial_t f = -\Delta f + |\nabla f|^2 - R on a closed manifold MM. Then ddtF(g,f)=2∫M∣Rij+∇i∇jf∣2e−f dV≥0\frac{d}{dt}\mathcal{F}(g,f) = 2\int_M |R_{ij} + \nabla_i \nabla_j f|^2 e^{-f}\,dV \ge 0, with equality at time tt if and only if Rij+∇i∇jf=0R_{ij}+\nabla_i\nabla_jf=0 (a steady gradient Ricci soliton).

Why is it true?

This is the "arrow of time" that makes Ricci flow behave like a gradient flow: F\mathcal{F} can only increase, so the flow can never return to a metric it has already left (no non-trivial periodic orbits), and fixed points of the flow (up to diffeomorphism and rescaling) are exactly the critical points of F\mathcal{F}, i.e. gradient Ricci solitons. It converts a system of nonlinear PDEs into something with the qualitative structure of gradient descent on an energy landscape.

Proof sketch

Step 1 (conserved measure). Under Ricci flow, ∂t dV=−R dV\partial_t\,dV=-R\,dV (since ∂tlog⁡det⁡g=12gij∂tgij=−R\partial_t\log\sqrt{\det g}=\tfrac12g^{ij}\partial_tg_{ij}=-R). Combining this with the evolution ∂tf=−Δf+∣∇f∣2−R\partial_t f = -\Delta f + |\nabla f|^2 - R for ff gives ∂t(e−fdV)=(−∂tf−R)e−fdV=(Δf−∣∇f∣2)e−fdV=−Δ(e−f) dV\partial_t(e^{-f}dV)=(-\partial_tf-R)e^{-f}dV=(\Delta f-|\nabla f|^2)e^{-f}dV=-\Delta(e^{-f})\,dV, using Δ(e−f)=e−f(∣∇f∣2−Δf)\Delta(e^{-f})=e^{-f}(|\nabla f|^2-\Delta f). Integrating over the closed manifold MM, the right side vanishes by the divergence theorem, so ∫Me−fdV\int_M e^{-f}dV is constant in time: e−fdVe^{-f}dV is a conserved measure carried along by the coupled flow.

**Step 2 (reduce to an integral of ∂tφ−Δφ\partial_t\varphi-\Delta\varphi).** Write φ=R+∣∇f∣2\varphi=R+|\nabla f|^2, so F(g,f)=∫Mφ e−fdV\mathcal F(g,f)=\int_M\varphi\,e^{-f}dV. Differentiating under the integral sign and using ∂t(e−fdV)=−Δ(e−f) dV\partial_t(e^{-f}dV)=-\Delta(e^{-f})\,dV from Step 1, ddtF=∫M(∂tφ)e−fdV−∫Mφ Δ(e−f) dV\frac{d}{dt}\mathcal F=\int_M(\partial_t\varphi)e^{-f}dV-\int_M\varphi\,\Delta(e^{-f})\,dV. Since the ordinary Laplacian is self-adjoint with respect to dVdV on the closed manifold MM, ∫Mφ Δ(e−f) dV=∫M(Δφ)e−fdV\int_M\varphi\,\Delta(e^{-f})\,dV=\int_M(\Delta\varphi)e^{-f}dV, so ddtF=∫M[∂tφ−Δφ]e−fdV\frac{d}{dt}\mathcal F=\int_M\big[\partial_t\varphi-\Delta\varphi\big]e^{-f}dV.

**Step 3 (the RR part).** Hamilton's evolution formula for scalar curvature is ∂tR=ΔR+2∣Ric∣2\partial_tR=\Delta R+2|\mathrm{Ric}|^2, hence ∂tR−ΔR=2∣Ric∣2\partial_tR-\Delta R=2|\mathrm{Ric}|^2 directly. This already contributes 2∣Ric∣22|\mathrm{Ric}|^2 to the bracket in Step 2.

**Step 4 (the ∣∇f∣2|\nabla f|^2 part).** Since ∂tgij=2Rij\partial_tg^{ij}=2R^{ij}, the chain rule gives ∂t∣∇f∣2=2Ric(∇f,∇f)+2⟨∇f,∇(∂tf)⟩\partial_t|\nabla f|^2=2\mathrm{Ric}(\nabla f,\nabla f)+2\langle\nabla f,\nabla(\partial_tf)\rangle, and substituting ∂tf=−Δf+∣∇f∣2−R\partial_tf=-\Delta f+|\nabla f|^2-R together with the identity ⟨∇f,∇∣∇f∣2⟩=2∇2f(∇f,∇f)\langle\nabla f,\nabla|\nabla f|^2\rangle=2\nabla^2f(\nabla f,\nabla f) yields ∂t∣∇f∣2=2Ric(∇f,∇f)−2⟨∇f,∇Δf⟩+4∇2f(∇f,∇f)−2⟨∇f,∇R⟩\partial_t|\nabla f|^2=2\mathrm{Ric}(\nabla f,\nabla f)-2\langle\nabla f,\nabla\Delta f\rangle+4\nabla^2f(\nabla f,\nabla f)-2\langle\nabla f,\nabla R\rangle. Bochner's formula states Δ∣∇f∣2=2∣∇2f∣2+2⟨∇f,∇Δf⟩+2Ric(∇f,∇f)\Delta|\nabla f|^2=2|\nabla^2f|^2+2\langle\nabla f,\nabla\Delta f\rangle+2\mathrm{Ric}(\nabla f,\nabla f); subtracting cancels both copies of Ric(∇f,∇f)\mathrm{Ric}(\nabla f,\nabla f) and leaves ∂t∣∇f∣2−Δ∣∇f∣2=−2∣∇2f∣2+4∇2f(∇f,∇f)−4⟨∇f,∇Δf⟩−2⟨∇f,∇R⟩\partial_t|\nabla f|^2-\Delta|\nabla f|^2=-2|\nabla^2f|^2+4\nabla^2f(\nabla f,\nabla f)-4\langle\nabla f,\nabla\Delta f\rangle-2\langle\nabla f,\nabla R\rangle.

Step 5 (weighted Bochner and a vanishing divergence). Introduce the Bakry–Émery (ff-weighted) Laplacian Δfh=Δh−⟨∇f,∇h⟩\Delta_f h = \Delta h - \langle\nabla f,\nabla h\rangle; it satisfies ∫M(Δfh)e−fdV=0\int_M(\Delta_fh)e^{-f}dV=0 for every function hh, since Δfh e−f=div(e−f∇h)\Delta_fh\,e^{-f}=\mathrm{div}(e^{-f}\nabla h) is a pure divergence. Writing Δf=Δff+∣∇f∣2\Delta f=\Delta_ff+|\nabla f|^2 turns ⟨∇f,∇Δf⟩\langle\nabla f,\nabla\Delta f\rangle into ⟨∇f,∇Δff⟩+2∇2f(∇f,∇f)\langle\nabla f,\nabla\Delta_ff\rangle+2\nabla^2f(\nabla f,\nabla f), and the weighted Bochner–Weitzenböck identity 12Δf∣∇f∣2=∣∇2f∣2+⟨∇f,∇Δff⟩+(Ric+∇2f)(∇f,∇f)\tfrac12\Delta_f|\nabla f|^2=|\nabla^2f|^2+\langle\nabla f,\nabla\Delta_ff\rangle+(\mathrm{Ric}+\nabla^2f)(\nabla f,\nabla f) isolates ⟨∇f,∇Δff⟩\langle\nabla f,\nabla\Delta_ff\rangle. Substituting all of this back into the Step 4 expression, every term built from Δf∣∇f∣2\Delta_f|\nabla f|^2 integrates to zero against e−fdVe^{-f}dV by the vanishing-divergence property above, and what survives is exactly ∫M[2∣∇2f∣2+4Ric(∇f,∇f)−2⟨∇f,∇R⟩]e−fdV\int_M\big[2|\nabla^2f|^2+4\mathrm{Ric}(\nabla f,\nabla f)-2\langle\nabla f,\nabla R\rangle\big]e^{-f}dV.

Step 6 (completing the square via the contracted Bianchi identity). Integrating ⟨Ric,∇2f⟩\langle\mathrm{Ric},\nabla^2f\rangle by parts and using the contracted second Bianchi identity ∇iRij=12∇jR\nabla^iR_{ij}=\frac12\nabla_jR gives exactly 4∫M⟨Ric,∇2f⟩e−fdV=∫M[4Ric(∇f,∇f)−2⟨∇f,∇R⟩]e−fdV4\int_M\langle\mathrm{Ric},\nabla^2f\rangle e^{-f}dV=\int_M\big[4\mathrm{Ric}(\nabla f,\nabla f)-2\langle\nabla f,\nabla R\rangle\big]e^{-f}dV — precisely the cross term left over from Step 5. Substituting back and adding the 2∣Ric∣22|\mathrm{Ric}|^2 from Step 3, the integrand assembles into a perfect square: 2∣Ric∣2+4⟨Ric,∇2f⟩+2∣∇2f∣2=2∣Ric+∇2f∣22|\mathrm{Ric}|^2+4\langle\mathrm{Ric},\nabla^2f\rangle+2|\nabla^2f|^2=2|\mathrm{Ric}+\nabla^2f|^2.

Conclusion. Therefore ddtF(g,f)=2∫M∣Rij+∇i∇jf∣2e−fdV≥0\frac{d}{dt}\mathcal{F}(g,f)=2\int_M|R_{ij}+\nabla_i\nabla_jf|^2e^{-f}dV\ge0, a sum of squares against the positive weight e−fdVe^{-f}dV, so F\mathcal{F} is non-decreasing along the coupled flow, and it is constant on an interval exactly when Rij+∇i∇jf≡0R_{ij}+\nabla_i\nabla_jf\equiv0 there, i.e. exactly on steady gradient Ricci solitons. ■\blacksquare

Topics that use this theorem

Step-by-step proofs

No step-by-step proof yet for this theorem.

References

  1. Grigori Perelman (2002). The entropy formula for the Ricci flow and its geometric applications · arXiv:math/0211159
  2. Grigori Perelman (2003). Ricci flow with surgery on three-manifolds · arXiv:math/0303109
  3. John Morgan, Gang Tian (2007). Ricci Flow and the Poincaré Conjecture
  4. Clay Mathematics Institute (2000). Poincaré Conjecture — Millennium Prize Problems