Cheeger constants and spectral gaps

For a smooth separating region, divide the area of its boundary by the smaller of its volume and the complementary volume, then take the infimum over nontrivial such regions. Using only one side without restricting its volume would give a different convention. The denominator prevents a small separator from being called a strong bottleneck when it encloses only a negligible amount of mass. In dimension one, boundary area counts separating points, allowing an explicit circle calculation.

The constant has units of inverse length, so its square has the same inverse-square units as a Laplace eigenvalue. Under metric length rescaling by c, the constant scales by one over c and the spectral gap by one over c squared. This dimensional check is useful, but it does not prove an inequality or determine its constant. A normalized graph Laplacian has yet another convention and cannot be inserted into a Riemannian formula without translating its measure and energy.

The proof uses a suitable positive or negative part of a test function after a median adjustment, so that its support occupies at most half the manifold. Integrating boundary areas over level sets is related by the coarea formula to an integral involving its gradient. The Cheeger definition bounds those areas from below. Cauchy–Schwarz then converts that geometric estimate into an energy estimate with the factor one quarter. This is a proof outline; the median and admissible Sobolev approximations are necessary steps, not decorative technicalities.

Min–max transfers the resulting estimate to the first nonconstant mode. A positive lower bound for h therefore prevents arbitrarily slow mean-zero heat relaxation. The inequality's direction also gives an upper bound on h when the spectral gap is small. A small h by itself does not yield a small-gap upper bound merely by reading Cheeger's inequality backward. Constructing a test function or invoking a separately qualified reverse theorem is needed for that conclusion.

A symmetric dumbbell with a thin connecting neck suggests an explicit mean-zero trial: nearly constant positive on one side, nearly constant negative on the other, and varying gradually across the neck. Its Rayleigh quotient can be small when the neck energy is small relative to the mass of the two lobes. This supplies a direct upper-bound argument for that geometric family. It is distinct from Cheeger's lower bound and requires estimates for the actual trial and neck geometry.

Buser-type reverse inequalities on closed manifolds use a lower Ricci-curvature bound. In dimension $d\geq2$, the condition $\operatorname{Ric}\geq-(d-1)K^2g$ with $K\geq0$ permits a bound $\gamma\leq C_d(Kh+h^2)$ with a dimensional constant. Omitting curvature control would overstate that theorem. These paired estimates explain why bottlenecks and low eigenvalues are closely connected in controlled geometric classes while preserving the different hypotheses of their two directions.