Li–Yau Inequality Overview

• Li–Yau Inequality is a differential Harnack inequality that offers precise gradient estimates for positive heat equation solutions on manifolds with nonnegative Ricci curvature.
• It is established by differentiating the heat equation, applying Bochner’s identity, and utilizing the parabolic maximum principle, with implications for heat kernel bounds and Harnack inequalities.
• Extensions of Li–Yau theory include adaptations to synthetic, discrete, and non-local settings, enabling analysis on graphs, metric spaces, and singular geometric structures.

The Li–Yau inequality is a differential Harnack inequality for positive solutions of heat-type equations. In its classical form, if $u>0$ solves $\partial_t u=\Delta u$ on a complete $n$-dimensional Riemannian manifold with nonnegative Ricci curvature, then

$$|\nabla \log u|^2-\partial_t \log u \le \frac{n}{2t},$$

equivalently,

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

The estimate is sharp in Euclidean space, where equality is achieved by the heat kernel, and it is one of the basic mechanisms behind parabolic Harnack inequalities, heat kernel bounds, and Liouville theorems (Jiang, 2014, Camilli, 22 Jan 2025).

1. Classical heat-equation formulation

Li and Yau introduced the inequality in the setting of positive heat solutions on complete manifolds under lower Ricci curvature bounds. In the nonnegative-curvature case, the estimate controls the logarithmic gradient by the parabolic scale $t^{-1}$, and the right-hand side $\frac{n}{2t}$ is the canonical dimensional constant (Jiang, 2014). A standard curved version states that if $\mathrm{Ric}\ge -K$ with $K\ge 0$, then for every $\alpha>1$,

$\partial_t u=\Delta u$0

which reduces to the sharp Euclidean formula when $\partial_t u=\Delta u$1 and $\partial_t u=\Delta u$2 (Qian, 2021).

The estimate is often written in Laplacian form. Since

$\partial_t u=\Delta u$3

the classical inequality can also be stated as

$\partial_t u=\Delta u$4

or, after substituting $\partial_t u=\Delta u$5, as a control of $\partial_t u=\Delta u$6 by $\partial_t u=\Delta u$7 and $\partial_t u=\Delta u$8 (Hung, 2023). This multiplicity of equivalent formulations is one reason the Li–Yau inequality appears in several distinct analytic guises.

2. Analytical structure and proof mechanisms

The original proof is based on differentiating the heat equation, applying Bochner’s identity, and using the parabolic maximum principle on a carefully chosen quantity such as $\partial_t u=\Delta u$9 with $n$0 (Qian, 2021). In smooth settings this route is naturally encoded by the Bakry–Émery curvature-dimension formalism, where $n$1 or related conditions provide the synthetic replacement for lower Ricci curvature bounds (Bakry et al., 2014).

A major later development is that Li–Yau theory can be reformulated at semigroup level. Under a Markov diffusion triple satisfying $n$2, one obtains a global Li–Yau inequality expressed through an explicit concave function $n$3 of $n$4; the authors state that this inequality is stronger than all classical Li–Yau type inequalities known to them, and on Riemannian manifolds it is equivalent to a new parabolic Harnack inequality in both negative and positive curvature (Bakry et al., 2014). This replaces the classical pointwise maximum-principle derivation by an entropy and Sturm-comparison argument at the semigroup level.

A second major proof paradigm uses explicit heat-kernel representations rather than curvature-dimension inequalities. For very general non-local diffusion equations, a reduction principle shows that Li–Yau bounds for the heat kernel imply corresponding bounds for all positive solutions represented by that kernel (Weber et al., 2020). The same representation-based viewpoint also yields generalized second- and fourth-derivative Li–Yau type inequalities on $n$5 (Hung, 2023). This suggests that Li–Yau theory is not tied to a single differential identity; it can also emerge from precise kernel calculus.

3. Consequences for Harnack theory, heat kernels, and rigidity

The canonical consequence of the Li–Yau inequality is the parabolic Harnack inequality. Integrating the differential bound along space-time curves gives estimates relating $n$6 and $n$7 for $n$8, typically with an exponential factor involving $n$9 and a time-ratio term such as $$|\nabla \log u|^2-\partial_t \log u \le \frac{n}{2t},$$0 (Jiang, 2014). This is the standard Li–Yau mechanism in both smooth and synthetic settings.

Heat kernel bounds follow from the same scheme. In smooth and $$|\nabla \log u|^2-\partial_t \log u \le \frac{n}{2t},$$1 settings, Li–Yau and Harnack inequalities are consistent with Gaussian two-sided estimates; in the $$|\nabla \log u|^2-\partial_t \log u \le \frac{n}{2t},$$2 case the heat kernel itself satisfies the Li–Yau gradient estimate

$$|\nabla \log u|^2-\partial_t \log u \le \frac{n}{2t},$$3

for $$|\nabla \log u|^2-\partial_t \log u \le \frac{n}{2t},$$4-a.e. $$|\nabla \log u|^2-\partial_t \log u \le \frac{n}{2t},$$5 (Jiang, 2014). In the abstract Markov-semigroup framework, the global Li–Yau inequality also yields new heat-kernel Harnack bounds and, under positive curvature, ultracontractive estimates by a direct method (Bakry et al., 2014).

Rigidity phenomena are another standard consequence. In classical Li–Yau theory, bounded harmonic functions on complete manifolds with nonnegative Ricci curvature are constant; analogous Liouville properties persist in several extensions, including discrete graphs and metric star graphs (Camilli, 22 Jan 2025, Qian, 2013). Spectral applications are also typical: on graphs, Li–Yau inequalities lead to heat-kernel bounds, Buser-type inequalities, and Cheng-type eigenvalue estimates (Bauer et al., 2013, Gong et al., 2018).

4. Synthetic, discrete, and non-local generalizations

One major branch of the subject extends Li–Yau from smooth manifolds to synthetic metric-measure geometry. On $$|\nabla \log u|^2-\partial_t \log u \le \frac{n}{2t},$$6 spaces, the heat flow satisfies exactly the classical-form estimate

$$|\nabla \log u|^2-\partial_t \log u \le \frac{n}{2t},$$7

for nonnegative $$|\nabla \log u|^2-\partial_t \log u \le \frac{n}{2t},$$8 initial data, and the corresponding Harnack and heat-kernel inequalities persist for general $$|\nabla \log u|^2-\partial_t \log u \le \frac{n}{2t},$$9 through a Baudoin–Garofalo-type refinement (Jiang, 2014).

A second branch concerns graphs. Several discrete curvature frameworks support Li–Yau theory: the exponential curvature-dimension condition $$\frac{|\nabla u|^2}{u^2}-\frac{\partial_t u}{u}\le \frac{n}{2t}.$$0 yields gradient estimates, Harnack inequalities, heat kernel bounds, polynomial volume growth under nonnegative curvature, and Buser-type inequalities (Bauer et al., 2013); $$\frac{|\nabla u|^2}{u^2}-\frac{\partial_t u}{u}\le \frac{n}{2t}.$$1 supports time-dependent Li–Yau families and Hamilton-type estimates (Qian, 2013); $$\frac{|\nabla u|^2}{u^2}-\frac{\partial_t u}{u}\le \frac{n}{2t}.$$2 allows analogous results for unbounded graph Laplacians, including Cheng’s eigenvalue estimate (Gong et al., 2018). A distinct nonlinear curvature-dimension condition $$\frac{|\nabla u|^2}{u^2}-\frac{\partial_t u}{u}\le \frac{n}{2t}.$$3 recovers an exact logarithmic Li–Yau inequality on finite graphs, and Ricci-flat graphs have nonnegative curvature in this sense (Münch, 2014). Another graph-theoretic approach replaces $$\frac{|\nabla u|^2}{u^2}-\frac{\partial_t u}{u}\le \frac{n}{2t}.$$4 by a modified nonlinear flow $$\frac{|\nabla u|^2}{u^2}-\frac{\partial_t u}{u}\le \frac{n}{2t}.$$5; under $$\frac{|\nabla u|^2}{u^2}-\frac{\partial_t u}{u}\le \frac{n}{2t}.$$6 this yields

$$\frac{|\nabla u|^2}{u^2}-\frac{\partial_t u}{u}\le \frac{n}{2t}.$$7

from which volume doubling is deduced, solving a major open problem in discrete Ricci curvature, and implying that there exist no expander graphs satisfying $$\frac{|\nabla u|^2}{u^2}-\frac{\partial_t u}{u}\le \frac{n}{2t}.$$8 (Münch, 2019).

Discrete curvature is not the only route. Cayley graphs of virtually Abelian groups satisfy Li–Yau type gradient estimates even though they do not satisfy any known variant of the curvature-dimension inequality with non-negative curvature (Lippner et al., 2016). For non-local diffusions, a heat-kernel reduction principle yields Li–Yau inequalities for the fractional heat equation,

$$\frac{|\nabla u|^2}{u^2}-\frac{\partial_t u}{u}\le \frac{n}{2t}.$$9

thereby solving the problem of obtaining a $t^{-1}$0-scale Li–Yau inequality for positive solutions of the fractional heat equation (Weber et al., 2020).

5. Metric graphs and singular one-dimensional geometry

Metric graphs interpolate between manifolds and discrete graphs: edges carry continuous one-dimensional geometry, while vertices introduce singularities. For a metric star graph $t^{-1}$1 consisting of $t^{-1}$2 half-lines joined at one central vertex and equipped with weighted Kirchhoff matching

$t^{-1}$3

positive heat solutions admit an explicit kernel representation (Camilli, 22 Jan 2025).

In that setting, a Li–Yau type estimate holds on each edge: $t^{-1}$4 where $t^{-1}$5 is an explicit integral term depending on the initial data, the edge, the point, the time, and the coefficient $t^{-1}$6 (Camilli, 22 Jan 2025). The correction term is the analytic signature of the branching singularity and the weighted transmission law at the vertex.

This formula reduces sharply to the classical one-dimensional Li–Yau inequality on $t^{-1}$7 when the star graph degenerates to two symmetric half-lines with $t^{-1}$8; in that case the extra term vanishes (Camilli, 22 Jan 2025). The same estimate implies a parabolic Harnack inequality on the graph and a Liouville property: every bounded harmonic function satisfying the vertex conditions is constant (Camilli, 22 Jan 2025). A plausible implication is that, on singular one-dimensional spaces, Li–Yau theory remains valid but generally acquires an explicit defect term encoding vertex scattering.

6. Higher-order, semilinear, and geometric variants

The Li–Yau paradigm has also been extended beyond the linear heat equation. On $t^{-1}$9, a representation-formula method yields a generalized second-derivative family,

$\frac{n}{2t}$0

under the parameter constraint $\frac{n}{2t}$1, and also a fourth-order Li–Yau type inequality with $\frac{n}{2t}$2 scaling (Hung, 2023). These results show that higher-order Li–Yau structures can be derived directly from Gaussian kernel calculus.

For the semilinear heat equation

$\frac{n}{2t}$3

on complete manifolds with $\frac{n}{2t}$4, a semilinear Li–Yau inequality holds in the exponent range $\frac{n}{2t}$5. In particular, for admissible $\frac{n}{2t}$6,

$\frac{n}{2t}$7

which leads to a Harnack inequality and to monotonicity, convexity, decay estimates, and triviality statements for ancient and eternal solutions (Castorina et al., 2022).

The name “Li–Yau inequality” also appears in geometric analysis in a different but related sense. For closed immersed surfaces in $\frac{n}{2t}$8, the classical Willmore-theoretic Li–Yau inequality states that $\frac{n}{2t}$9 implies embeddedness; Simon’s monotonicity formula extends this principle to surfaces with boundary, producing boundary-dependent thresholds $\mathrm{Ric}\ge -K$0 and $\mathrm{Ric}\ge -K$1 for embeddedness or exclusion of axis self-intersections (Schlierf, 2024). In dimension one, an analogous embeddedness threshold for closed planar curves is expressed using the scale-invariant elastic energy $\mathrm{Ric}\ge -K$2, with the figure-eight elastica furnishing the sharp non-embedded threshold (Müller et al., 2021). This suggests that “Li–Yau inequality” now denotes a broader multiplicity-control paradigm: in heat-flow settings it controls oscillation through differential Harnack estimates, while in Willmore-type settings it controls self-intersection through curvature-energy thresholds.