Cheeger Constant

• The Cheeger constant is a geometric invariant measuring the smallest boundary-to-volume ratio, critical in analyzing bottlenecks and spectral gaps in various geometric and discrete structures.
• The notion of Cheeger constant is used in multiple fields, including the comparative analysis of graphs and manifolds, where it links boundary costs to topological and spectral properties.
• In spectral graph theory, the Cheeger constant provides insights into relaxation times by bounding the Laplacian eigenvalues.

The Cheeger constant is an isoperimetric invariant measuring the smallest boundary-to-volume ratio of a nontrivial portion of a geometric, discrete, spectral, or combinatorial object. In a bounded Euclidean domain $\Omega$, it is defined by

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$

where $P(E)$ is the perimeter and $|E|$ is Lebesgue measure. A minimizing set is a Cheeger set. The same principle has discrete analogues based on edge boundaries and vertex or degree volume, as well as extensions to Riemannian manifolds, quantum graphs, simplicial complexes, hypergraphs, random geometric graphs, and sheaf-valued cochains. Its central role is the quantitative relation between bottlenecks and spectral gaps.

1. Classical geometric formulation

For a bounded domain $\Omega\subset\mathbb R^N$, $N>1$, the Cheeger problem minimizes

$\frac{|\partial E|}{|E|}$

over sufficiently regular subdomains $E\subset\Omega$. In the finite-perimeter relaxation, $|\partial E|$ is replaced by the De Giorgi perimeter. If $E\subset\Omega$, the relative perimeter is

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$0

whereas the full perimeter is

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$1

The distinction is important when a Cheeger set touches $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$2: portions of its boundary lying on $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$3 contribute to the full perimeter, although they are not counted by the relative perimeter. A Cheeger set $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$4 satisfies

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$5

The quotient is scale-sensitive. Under dilation by $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$6,

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$7

For convex planar domains, the Cheeger set is unique and has free-boundary arcs of constant curvature $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$8, equivalently radius

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$9

For a Jordan domain without necks of radius $P(E)$0, the maximal Cheeger set is the union of all radius-$P(E)$1 balls contained in the domain: $P(E)$2 where

$P(E)$3

The radius is characterized by the inner Cheeger formula

$P(E)$4

These conclusions require the no-neck condition; bow-tie, unbalanced barbell, heart, and nonsimply connected domains provide examples in which the union-of-balls formula can fail (Leonardi et al., 2017).

The Cheeger constant can also be viewed as the first eigenvalue of the Dirichlet $P(E)$5-Laplacian in the BV sense. Formally,

$P(E)$6

while the rigorous formulation uses total variation and functions in $P(E)$7. This variational interpretation places the Cheeger problem at the endpoint $P(E)$8 of the Dirichlet $P(E)$9-Laplacian family (2011.3070).

2. Torsion, $|E|$0-Laplacians, and Cheeger sets

For $|E|$1, let $|E|$2 solve the Dirichlet $|E|$3-torsion problem

$|E|$4

where

$|E|$5

Testing the equation with $|E|$6 gives

$|E|$7

The Cheeger constant is recovered from both the supremum and the integral of the torsion function: $|E|$8 Here $|E|$9 is a scalar power, not an $\Omega\subset\mathbb R^N$0 norm. A level-set and coarea argument yields

$\Omega\subset\mathbb R^N$1

while variational testing against approximations of characteristic functions gives the reverse limiting inequality.

There is also a quantitative comparison between the two torsion norms: $\Omega\subset\mathbb R^N$2 where

$\Omega\subset\mathbb R^N$3

Define the normalized torsion functions

$\Omega\subset\mathbb R^N$4

Along a sequence $\Omega\subset\mathbb R^N$5,

$\Omega\subset\mathbb R^N$6

with

$\Omega\subset\mathbb R^N$7

The limit minimizes the BV functional

$\Omega\subset\mathbb R^N$8

over nonnegative, zero-extended functions with unit $\Omega\subset\mathbb R^N$9 norm, and

$N>1$0

For superlevel sets

$N>1$1

the coarea and layer-cake formulas imply

$N>1$2

Since every nontrivial level set satisfies

$N>1$3

equality must hold for almost every nonempty level set: $N>1$4 Thus almost every level set of the limiting $N>1$5-Laplacian eigenfunction is a Cheeger set.

The limiting function also satisfies

$N>1$6

When $N>1$7 is convex, uniqueness of the Cheeger set forces

$N>1$8

Hence normalized torsion functions converge to the normalized characteristic function of the unique Cheeger set (Bueno et al., 2010).

3. Geometric structures and extremal domains

The geometry of a domain strongly influences existence, uniqueness, and the form of Cheeger sets.

For a curved planar strip of half-width $N>1$9 around a connected $\frac{|\partial E|}{|E|}$0 curve, the embeddedness condition

$\frac{|\partial E|}{|E|}$1

injective on $\frac{|\partial E|}{|E|}$2 implies $\frac{|\partial E|}{|E|}$3. A universal calibration gives

$\frac{|\partial E|}{|E|}$4

If $\frac{|\partial E|}{|E|}$5 is compact, the strip is a unique Cheeger set of itself and

$\frac{|\partial E|}{|E|}$6

The curvature terms cancel in the area and perimeter computations. For infinite or semi-infinite curves the same value holds, but no Cheeger set exists: finite truncations form an optimizing sequence, while the full strip has infinite area and perimeter. For finite non-complete curves, a connected proper Cheeger set exists and

$\frac{|\partial E|}{|E|}$7

The endpoint correction is therefore of order $\frac{|\partial E|}{|E|}$8 (Krejcirik et al., 2010).

The same phenomenon extends to thin tubes around closed curves in real space forms. For a tube $\frac{|\partial E|}{|E|}$9 of radius $E\subset\Omega$0 around a closed curve in an $E\subset\Omega$1-dimensional space form of sectional curvature $E\subset\Omega$2, the Cheeger constant is independent of the shape of the core curve: $E\subset\Omega$3 The tube is self-Cheeger. A moving-sphere calibration field has unit norm and constant divergence

$E\subset\Omega$4

which gives the lower bound for every competitor, while the tube itself gives equality. Uniformly tubular unbounded tubes in noncompact Euclidean or hyperbolic space have the same constant but no finite-volume Cheeger set (Vlachopulos, 17 Aug 2026).

Among planar convex bodies of constant width, the Reuleaux triangle maximizes the Cheeger constant. For width $E\subset\Omega$5,

$E\subset\Omega$6

where $E\subset\Omega$7 is the Reuleaux triangle. Its Cheeger radius satisfies

$E\subset\Omega$8

and therefore

$E\subset\Omega$9

This is a Cheeger analogue of the Blaschke–Lebesgue theorem. The result is proved using Blaschke deformations, optimality conditions for Reuleaux polygons, and an upper bound on the inradius of an optimal body (Henrot et al., 2020).

For closed hyperbolic surfaces, the Cheeger constant reflects global topology. The hyperbolic plane has

$|\partial E|$0

whereas for closed hyperbolic surfaces $|\partial E|$1 of genus $|\partial E|$2,

$|\partial E|$3

The proof uses Poisson–Voronoi tessellations and random black-white coloring of cells. The constant $|\partial E|$4 arises from the angular geometry of the hyperbolic plane. The theorem is an upper bound; it does not establish that $|\partial E|$5 is the exact asymptotic maximum (Budzinski et al., 2022).

For a cocompact negatively curved manifold with sectional curvature at most $|\partial E|$6, Yau’s estimate gives

$|\partial E|$7

on the universal cover. Equality is rigid: if

$|\partial E|$8

then the universal cover is isometric to hyperbolic space $|\partial E|$9. The proof analyzes vanishing Busemann-function defects and uses cocompactness and strong-stable foliations to promote asymptotic equality to global Hessian equality (Jin et al., 5 Sep 2026).

4. Graphs, spectra, and spectral algorithms

For a finite graph $E\subset\Omega$0, several normalizations are used. A common unnormalized edge-expansion constant is

$E\subset\Omega$1

For degree-weighted conductance,

$E\subset\Omega$2

For a $E\subset\Omega$3-regular graph, these differ by normalization: $E\subset\Omega$4

With the normalized Laplacian, the classical Cheeger inequality is

$E\subset\Omega$5

where $E\subset\Omega$6 is the second-smallest normalized-Laplacian eigenvalue. For the unnormalized Laplacian of a graph with maximum degree $E\subset\Omega$7,

$E\subset\Omega$8

The upper estimate is obtained by thresholding an eigenvector, while the lower estimate follows from a Rayleigh quotient.

Eigenvector norms can refine the upper bound. If $E\subset\Omega$9 is a normalized harmonic eigenvector associated with $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$00, then, under

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$01

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$02

This retains the lower bound $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$03 and yields

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$04

under the stated degree hypothesis (Kenter, 2014).

Higher-order graph Cheeger constants include both worst-case and average-case formulations. For a partition $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$05, the average-case quantity is

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$06

and

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$07

With harmonic eigenvectors $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$08 and

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$09

the linear $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$10-fold inequality is

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$11

The average of the first $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$12 spectral modes replaces dependence on only $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$13, while eigenvector localization enters through $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$14 (Kenter et al., 2015).

For distance-regular graphs, the general upper bound

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$15

is conjectured to improve to

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$16

The conjectured inequality is proved for all strongly regular graphs, principal infinite families, many bipartite and antipodal distance-regular graphs, numerous classical-parameter families, and most classified graphs of small valency. The proofs construct induced subgraphs of average valency at least the second-largest adjacency eigenvalue (Koolen et al., 2018).

The Cheeger constant can also be estimated statistically from spectral data. For random regular graphs with sizes up to $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$17, a linear model using the two largest adjacency eigenvalues was observed to fit approximately

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$18

Deep neural networks trained on smaller random regular graphs estimated the constants of larger graphs with mean deviations generally between $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$19 and $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$20 in the reported experiments. These are empirical findings for the tested ensembles, not universal theorems (Jain et al., 2020).

5. Higher-dimensional, hypergraph, and sheaf generalizations

For a $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$21-dimensional simplicial complex $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$22, the ordinary graph-cut definition does not generalize uniquely. One combinatorial definition partitions the vertex set into $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$23 nonempty blocks,

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$24

and counts $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$25-faces with exactly one vertex in each block: $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$26 For incomplete $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$27-skeleta, the completion $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$28 contains potential $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$29-faces whose entire boundary is present. The resulting inequality is

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$30

where $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$31 is the smallest nontrivial eigenvalue of the upper Laplacian restricted away from coboundaries. A degree-sensitive version is

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$32

and a second parameter $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$33 satisfies

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$34

These quantities differ from $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$35-cohomological expansion, which need not control the upper-Laplacian eigenvalue for arbitrary complexes (Gundert et al., 2014).

For the simplex $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$36, the first higher Cheeger constant is a coboundary-expansion quantity: $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$37 It can be recast as a minimization over cut-minimal graphs and their triangle expansion. The sharp lower bound is

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$38

Equality holds whenever $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$39 is not a power of $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$40, while for $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$41,

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$42

Staircase graphs associated with Ferrers diagrams provide the extremal constructions (Kozlov, 2016).

For a $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$43-uniform classical hypergraph, define

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$44

and

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$45

where $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$46 is the set of hyperedges meeting $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$47 in exactly $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$48 vertices. The generalized Cheeger constant is

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$49

For the normalized hypergraph Laplacian, the relevant eigenvalue is the second-largest, because the constant functions correspond to the top eigenvalue $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$50. The hypergraph Cheeger inequalities are

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$51

For $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$52, this becomes the classical graph inequality after identifying the positive-incidence operator with the signless normalized Laplacian (Mulas, 2020).

Cheeger expansion is also the $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$53-dimensional coboundary expansion of a graph over $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$54. For a weighted graph,

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$55

and

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$56

For any nontrivial abelian group $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$57 with constant coefficients,

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$58

For arbitrary sheaves, graph expansion alone is insufficient: restriction maps can annihilate cochain differences, producing zero sheaf coboundary expansion even on strong spectral expanders. Quotients of constant sheaves by suitably linearly disjoint subsheaves provide a positive nonconstant theory with explicit lower bounds (First et al., 2022).

6. Discrete-to-continuum limits and further variants

For a random geometric graph built from independent samples in a bounded Lipschitz domain $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$59, with interaction scale $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$60 satisfying

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$61

weighted graph cuts converge to continuum weighted perimeter problems. If $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$62 is the interaction kernel and

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$63

then for each of the four combinations of vertex or degree volume and one-sided or product balance,

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$64

almost surely. The empirical measures of optimizing graph cuts converge, along subsequences, to continuum minimizing sets. In dimension $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$65, the condition $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$66 removes an extra $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$67 requirement appearing in earlier work (Müller et al., 2018).

Higher Cheeger constants for Euclidean domains are defined by

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$68

where the $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$69 are mutually disjoint positive-measure subsets of $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$70. Minimizers can be chosen with adjustment properties, and their boundaries have prescribed regularity. The associated spectral minimal $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$71-partition energy for the $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$72-Laplacian is

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$73

and

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$74

For $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$75, this recovers the classical limit $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$76; for $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$77, it gives the corresponding limit for the second $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$78-Laplacian eigenvalue (Bobkov et al., 2017).

A further family replaces $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$79 by $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$80: $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$81 The nontrivial range is

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$82

For $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$83, the generalized constants are universally equivalent to $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$84: $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$85 At the critical exponent,

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$86

while for $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$87,

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$88

For $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$89, long thin sets and disconnected components produce loss of compactness and the equivalence with $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$90 fails (Brasco, 2024).

In graph minor theory, the condition

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$91

for a graph of maximum degree at most $$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$92 forces a complete minor of order

$$h(\Omega)=\inf_{E\subseteq\Omega,\ |E|>0}\frac{P(E)}{|E|},$$93

The proof uses remainder lemmas, connected sets with large external neighborhoods, lazy random walks, connected hitting sets, and the Kostochka–Thomason theorem. Thus a global Cheeger condition alone supplies the robust expansion needed for optimal-order complete minors; an additional restricted Cheeger condition is unnecessary (Li et al., 14 Sep 2026).

Across these settings, the Cheeger constant retains one organizing principle: it quantifies the least boundary cost required to separate a substantial region from its complement. The meaning of boundary and volume changes with the ambient category—perimeter and Lebesgue measure for domains, edge boundaries and degree volume for graphs, metric length for quantum graphs, weighted coboundaries for complexes and sheaves—but the invariant consistently links geometric bottlenecks, variational minimizers, and spectral or probabilistic structure.