Stele III · 1986 · with Peter Li

李–丘估计The Li–Yau Inequality

热怎样在弯曲空间里扩散How heat spreads on a curved space
原文Paper · Acta Math. 156 (1986) 阅读约 18 分钟~18 min read
TL;DR

在一块金属板上点一下热,热量会扩散开来。如果"板"是一个弯曲的空间呢?李伟光与丘成桐在 1986 年证明:只要空间的里奇曲率非负,热方程的任何正解都满足一个与空间无关的精确梯度估计:

Touch a metal plate with heat and it spreads out. What if the "plate" is a curved space? In 1986 Peter Li and Shing-Tung Yau proved that when Ricci curvature is nonnegative, every positive solution of the heat equation obeys a sharp gradient estimate independent of the space:

$$\frac{|\nabla u|^2}{u^2}-\frac{u_t}{u}\;\le\;\frac{n}{2t}.$$

它的威力在于:从这个逐点的不等式出发,沿路径积分就能比较不同时刻、不同地点的温度(Harnack 不等式),进而得到热核的高斯型上下界。这个思路后来深刻影响了里奇流。

Its power: integrating this pointwise inequality along paths compares temperatures at different times and places (a Harnack inequality), which in turn gives Gaussian upper and lower bounds for the heat kernel. The idea later shaped Ricci flow profoundly.

Chapter 1

热方程:最自然的扩散The heat equation: the most natural diffusion

温度 $u(x,t)$ 随时间的变化率,等于它与周围平均温度的差。写成方程就是

The rate at which temperature $u(x,t)$ changes equals how much it differs from the average around it. As an equation:

$$\frac{\partial u}{\partial t}=\Delta u .$$

在弯曲的黎曼流形上,$\Delta$ 换成拉普拉斯–贝尔特拉米算子,方程的形式不变。在平直的 $\mathbb{R}^n$ 上,从一点出发的热扩散是一个越来越扁的高斯钟形,这就是热核:

On a curved Riemannian manifold, $\Delta$ becomes the Laplace–Beltrami operator and the equation keeps its form. On flat $\mathbb{R}^n$, heat released at a point spreads as an ever flatter Gaussian bell — the heat kernel:

$$H(x,t)=\frac{1}{(4\pi t)^{n/2}}\,e^{-|x|^2/4t}.$$

首屏那个起伏的金色曲面,正是二维热核随时间展开、收拢的样子。

The rising and settling golden surface at the top is exactly the two-dimensional heat kernel unfolding over time.

Chapter 2

定理The theorem

李–丘梯度估计(1986)Li–Yau gradient estimate (1986)

设 $M^n$ 是完备黎曼流形,里奇曲率 $\operatorname{Ric}\ge0$。若 $u>0$ 在 $M\times(0,\infty)$ 上满足热方程,则

Let $M^n$ be a complete Riemannian manifold with $\operatorname{Ric}\ge0$. If $u>0$ solves the heat equation on $M\times(0,\infty)$, then

$$\frac{|\nabla u|^2}{u^2}-\frac{u_t}{u}\le\frac{n}{2t}\quad\Longleftrightarrow\quad|\nabla f|^2-f_t\le\frac{n}{2t},\qquad f=\log u .$$

论文还处理了里奇曲率有下界 $-K$ 的一般情形,以及带势函数的薛定谔算子,估计的右边会多出依赖于 $K$ 的项。

The paper also treats the general case $\operatorname{Ric}\ge-K$ and Schrödinger operators with potentials; the right-hand side then gains terms depending on $K$.

为什么是 $n/2t$?等号恰好取到Why $n/2t$? Equality is attained

对欧氏热核验证一下:$f=\log H=-\frac n2\log(4\pi t)-\frac{|x|^2}{4t}$,于是 $|\nabla f|^2=\frac{|x|^2}{4t^2}$,$f_t=-\frac{n}{2t}+\frac{|x|^2}{4t^2}$,所以

Check on the Euclidean heat kernel: $f=\log H=-\frac n2\log(4\pi t)-\frac{|x|^2}{4t}$, so $|\nabla f|^2=\frac{|x|^2}{4t^2}$ and $f_t=-\frac{n}{2t}+\frac{|x|^2}{4t^2}$, giving

$$|\nabla f|^2-f_t=\frac{n}{2t}\quad\text{exactly.}$$
Sharp

平直空间的热核恰好让不等式取等号,所以常数 $n/2$ 不能再改进。直观地说,非负里奇曲率的空间里,热扩散"不会比平直空间更剧烈"。

The flat heat kernel makes the inequality an equality, so the constant $n/2$ cannot be improved. Intuitively, on a space with nonnegative Ricci curvature, heat diffuses "no more violently than in flat space".

Chapter 3

证明:Bochner 公式 + 极大值原理The proof: Bochner formula + maximum principle

令 $f=\log u$,它满足 $f_t=\Delta f+|\nabla f|^2$。考虑量

Let $f=\log u$; it satisfies $f_t=\Delta f+|\nabla f|^2$. Consider

$$F=t\big(|\nabla f|^2-f_t\big)=-t\,\Delta f .$$

目标是证明 $F\le n/2$。

The goal is $F\le n/2$.

Bochner 公式:曲率登场的地方Bochner's formula: where curvature enters

$$\tfrac12\Delta|\nabla f|^2=|\nabla^2 f|^2+\langle\nabla f,\nabla\Delta f\rangle+\operatorname{Ric}(\nabla f,\nabla f).$$

$\operatorname{Ric}\ge0$ 让最后一项可以丢掉;柯西–施瓦茨不等式给出 $|\nabla^2 f|^2\ge\frac1n(\Delta f)^2$。把这些代入 $F$ 的演化方程,得到(这里 $\mathcal L=\Delta-\partial_t$)

$\operatorname{Ric}\ge0$ lets us drop the last term; Cauchy–Schwarz gives $|\nabla^2 f|^2\ge\frac1n(\Delta f)^2$. Substituting into the evolution of $F$ yields (with $\mathcal L=\Delta-\partial_t$)

$$\mathcal L F+2\langle\nabla f,\nabla F\rangle\;\ge\;\frac{2}{nt}F^2-\frac{F}{t}=\frac{F}{t}\Big(\frac{2F}{n}-1\Big).$$

极大值原理The maximum principle

在 $F$ 取最大值的点,$\nabla F=0$,$\Delta F\le0$,$F_t\ge0$,所以左边 $\le0$。于是在那一点 $\frac{2F}{n}-1\le0$,即 $F\le\frac n2$。紧流形上这就结束了。

At a maximum point of $F$: $\nabla F=0$, $\Delta F\le0$, $F_t\ge0$, so the left side is $\le0$. Hence $\frac{2F}{n}-1\le0$ there, i.e. $F\le\frac n2$. On a compact manifold, done.

完备非紧的情形是论文技术上的难点:要乘一个精心设计的截断函数,借助拉普拉斯比较定理控制它的导数,再让截断半径趋于无穷。

The complete noncompact case is the paper's technical heart: multiply by a carefully built cutoff function, control its derivatives with the Laplacian comparison theorem, and let the radius go to infinity.

Chapter 4

从逐点估计到 Harnack 不等式From a pointwise bound to Harnack

沿着从 $(x_1,t_1)$ 到 $(x_2,t_2)$($t_1<t_2$)的时空路径对 $\frac{d}{ds}f$ 积分,用梯度估计控制被积函数,再对路径取最优,就得到抛物型 Harnack 不等式:

Integrate $\frac{d}{ds}f$ along a space-time path from $(x_1,t_1)$ to $(x_2,t_2)$ with $t_1<t_2$, bound the integrand with the gradient estimate, and optimize over paths. The result is the parabolic Harnack inequality:

$$u(x_1,t_1)\;\le\;u(x_2,t_2)\Big(\frac{t_2}{t_1}\Big)^{n/2}\exp\!\Big(\frac{d(x_1,x_2)^2}{4(t_2-t_1)}\Big).$$
◈

含义:此刻这里很热,那么稍后附近的地方也不会太冷。温度不可能在短时间、短距离内凭空消失。

Meaning: if it is hot here now, then a little later, nearby, it cannot be too cold. Heat cannot vanish over short times and distances.

进一步可以得到热核 $H(x,y,t)$ 的双边高斯估计:用 $x$ 处半径 $\sqrt t$ 的测地球体积 $V_x(\sqrt t)$ 作归一化,对任意 $\varepsilon>0$,

Further, one obtains two-sided Gaussian bounds for the heat kernel $H(x,y,t)$, normalized by the volume $V_x(\sqrt t)$ of the geodesic ball of radius $\sqrt t$: for any $\varepsilon>0$,

$$\frac{C^{-1}}{V_x(\sqrt t)}\,e^{-\frac{d(x,y)^2}{(4-\varepsilon)t}}\;\le\;H(x,y,t)\;\le\;\frac{C}{V_x(\sqrt t)}\,e^{-\frac{d(x,y)^2}{(4+\varepsilon)t}}.$$

这里 $C$ 只依赖于 $n$ 和 $\varepsilon$。弯曲空间上的热核,在定量意义上"像"平直空间的高斯核。

Here $C$ depends only on $n$ and $\varepsilon$. On a curved space the heat kernel quantitatively "looks like" the flat Gaussian.

Chapter 5

影响:通往里奇流的路Legacy: a road to Ricci flow

  • Hamilton 的矩阵 Harnack 不等式(1993):Hamilton 把李–丘估计推广到里奇流本身,得到曲率的 Harnack 估计。这是分析里奇流奇点的关键工具。
  • Hamilton's matrix Harnack inequality (1993): Hamilton extended Li–Yau to Ricci flow itself, obtaining a Harnack estimate for curvature — a key tool for analyzing singularities.
  • Perelman(2002–03):Perelman 证明庞加莱猜想时引入的约化距离与约化体积,在形式上和李–丘的路径积分一脉相承,他在论文中也明确提到了这种联系。
  • Perelman (2002–03): the reduced distance and reduced volume Perelman introduced on the way to the Poincaré conjecture are formally descended from the Li–Yau path integral, a connection he notes himself.
  • 更广的世界:李–丘型估计被推广到图、度量测度空间、非线性扩散等各种情形,"微分 Harnack 不等式"成为一个独立的研究方向。
  • Wider world: Li–Yau-type estimates have been extended to graphs, metric measure spaces, nonlinear diffusions and more; "differential Harnack inequalities" became a field of their own.
References

原文与延伸阅读Original papers & further reading

  • 1986P. Li, S.-T. Yau, On the parabolic kernel of the Schrödinger operator, Acta Math. 156, 153–201. DOI ↗
  • 1975S.-T. Yau, Harmonic functions on complete Riemannian manifolds, Comm. Pure Appl. Math. 28, 201–228.(梯度估计方法的前身)(precursor of the gradient-estimate method)
  • 1993R. Hamilton, The Harnack estimate for the Ricci flow, J. Differential Geom. 37, 225–243.
  • 2002G. Perelman, The entropy formula for the Ricci flow and its geometric applications, arXiv:math/0211159.