---
title: Higher Order Cheeger Inequality
url: https://www.emergentmind.com/topics/higher-order-cheeger-inequality
type: topic
---

# Higher Order Cheeger Inequality

Higher-order Cheeger inequality is the multiway extension of classical Cheeger theory in spectral graph analysis: it relates higher Laplacian eigenvalues to the existence of multiple disjoint sparse sets, rather than a single sparse cut. In its standard graph-theoretic form, it upgrades the classical correspondence between the second eigenvalue and two-way expansion to a correspondence between the \(k\)-th eigenvalue and \(k\)-way expansion, and it underlies the spectral-clustering paradigm based on the bottom \(k\) eigenvectors [1107.2686, 1111.1055]. The phrase also names a broader family of results for dual bipartite expansion, \(p\)-Laplacians, domains, graphons, and simplicial complexes, where the central problem is to identify the correct higher-order or higher-dimensional isoperimetric quantity.

## 1. Classical graph formulation and the standard higher-order inequality

For a finite undirected weighted graph \(G=(V,E,w)\), the normalized Laplacian is
\[
\mathcal L_G = I - D^{-1/2} A D^{-1/2},
\]
with eigenvalues
\[
0=\lambda_1 \le \lambda_2 \le \cdots \le \lambda_n \le 2.
\]
For a subset \(S\subseteq V\), the Dirichlet conductance is
\[
\phi_G(S)=\frac{w(E(S,\overline S))}{w(S)},
\]
and the \(k\)-way expansion constant is
\[
\rho_G(k)=\min_{S_1,\dots,S_k}\max_{i=1,\dots,k}\phi_G(S_i),
\]
where the minimum is over all collections of \(k\) nonempty pairwise disjoint subsets [1111.1055].

The standard higher-order Cheeger inequality states that for every graph \(G\) and every \(k\in\mathbb N\),
\[
\frac{\lambda_k}{2}\le \rho_G(k)\le O(k^2)\sqrt{\lambda_k}.
\]
At \(k=2\), this recovers the classical Cheeger inequality in the form
\[
\frac{\lambda_2}{2}\le \rho_G(2)\le \sqrt{2\lambda_2}.
\]
The lower bound is the direct variational direction: \(k\) disjoint sparse sets imply a small \(k\)-th eigenvalue. The upper bound is the substantive direction: a small \(k\)-th eigenvalue forces the existence of \(k\) disjoint sets of small expansion [1111.1055].

Conceptually, the theorem is an approximate multiplicity statement. In the exact disconnected case, the multiplicity of the zero Laplacian eigenvalue equals the number of connected components. Higher-order Cheeger theory replaces exact components by \(k\) “almost disconnected” clusters and exact zero eigenvalues by \(k\) eigenvalues close to zero. This is the formulation emphasized in the original 2011 conjecture-resolution papers, which also frame the result as a theoretical justification for clustering from the bottom \(k\) eigenvectors [1107.2686, 1111.1055].

## 2. Spectral embedding, localization, and the multiway partitioning mechanism

The hard direction of the inequality is proved by passing from eigenfunctions to geometry. If \(f_1,\dots,f_k\) are orthonormal eigenfunctions for the first \(k\) eigenvalues, the spectral embedding is
\[
F(v)=(f_1(v),\dots,f_k(v))\in \mathbb R^k.
\]
The analysis then uses the radial projection distance
\[
d_F(u,v)=\left\|\frac{F(u)}{\|F(u)\|}-\frac{F(v)}{\|F(v)\|}\right\|,
\]
together with random metric partitions and smooth localization, to decompose the embedding into many separated regions [1111.1055].

A key intermediate statement is functional rather than combinatorial: there exist disjointly supported functions
\[
\psi_1,\dots,\psi_k:V\to\mathbb R
\]
such that
\[
\mathcal R_G(\psi_i)\le O(k^6)\lambda_k,
\]
where
\[
\mathcal R_G(f)=\frac{\sum_{\{u,v\}\in E}w(u,v)(f(u)-f(v))^2}{\sum_{v\in V}w(v)f(v)^2}.
\]
These localized functions are then converted into sparse sets by scalar sweep arguments. In this formulation, higher-order Cheeger theory is a localization theorem for low-energy eigenspaces [1111.1055].

The same framework yields refined statements. One such theorem is
\[
\rho_G(k)\le O\!\left(\sqrt{\lambda_{2k}\log k}\right),
\]
and it implies the small-set-expansion bound
\[
\varphi_G(k/2)\le O\!\left(\sqrt{\lambda_k\log k}\right),
\]
where \(\varphi_G(k)\) minimizes expansion over sets of size at most \(|V|/k\). The noisy hypercube shows that the \(\sqrt{\log k}\) dependence is tight up to constant factors for sets of size about \(n/k\) [1111.1055].

This proof architecture explains why higher-order Cheeger theory is directly relevant to multiway spectral partitioning. The embedding by the bottom \(k\) eigenvectors is not merely heuristic; it is the object on which the geometric partitioning argument operates. The resulting viewpoint is more precise than the classical two-way sweep-cut picture because it treats the entire low-eigenvalue subspace as the primary geometric datum [1111.1055].

## 3. Refinements, alternative constants, and improved guarantees

A major refinement replaces the classical upper bound’s \(\sqrt{\lambda_2}\) dependence by a bound involving a higher spectral gap. For every \(k\ge 2\),
\[
\phi(G)=O(k)\frac{\lambda_2}{\sqrt{\lambda_k}},
\]
and the same guarantee is achieved by the spectral partitioning algorithm itself:
\[
\phi(f)=O(k)\frac{\lambda_2}{\sqrt{\lambda_k}},
\]
where \(f\) is the threshold-cut function built from the second eigenvector. This shows that the second eigenvector is more informative when \(\lambda_2\) is small but \(\lambda_k\) is large, and the dependence is optimal up to constant factors for every \(k\) [1301.5584].

The same paper extends the higher-order spectral-gap philosophy to several partitioning problems. It gives a corollary of the form
\[
\phi_k(G)\le O(lk^6)\frac{\lambda_k}{\sqrt{\lambda_l}}
\qquad (l>k\ge 2),
\]
together with strengthened variants for \((1-\delta)k\) clusters and for graphs excluding a \(K_h\) minor. It also gives an improved spectral algorithm for balanced separator or minimum bisection, producing \(S\) with
\[
\frac15 d(V)\le d(S)\le \frac45 d(V)
\quad\text{and}\quad
\phi(S)\le O\!\left(\frac{k}{\lambda_k}\right),
\]
and it adapts the same top-spectrum logic to maximum cut [1301.5584].

Not all higher-order Cheeger constants are worst-case max-over-clusters quantities. An alternative “average-case” \(k\)-fold Cheeger constant \(h_G^{(k)}\) averages intercluster boundary terms and is related to the average of the first \(k-1\) nontrivial eigenvalues and to the \(\infty\)-norms of the corresponding harmonic eigenvectors. The resulting inequalities are linear rather than square-root bounds and remain relevant even when \(\lambda_{k-1}\to 1\) [1501.01741].

More recent work modifies either the eigenvalue index or the partitioning model itself. One refinement uses the cyclomatic number
\[
\ell=|E|-|V|+1
\]
to shift the index and proves
\[
h_{k-\ell}(G)\le \sqrt{2T_G}\,\lambda_k(G),
\]
where
\[
T_G=\max_{i\in V}\frac{d(i)}{p_i}.
\]
In the normalized setting \(p_i=d(i)\), this becomes
\[
h_{k-\ell}(G)\le \sqrt{2}\,\lambda_k(G),
\]
and for trees, where \(\ell=0\), it yields
\[
h_k(G)\le \sqrt{2T_G}\,\lambda_k(G).
\]
The same work also proves a lower bound
\[
h_k(G)\ge (T_{\min}-\nu)(1-\nu)
\]
in terms of the normalized adjacency spectral radius [2409.07097].

Another refinement introduces \(\varepsilon\)-buffered \(k\)-partitions. If \(h_G^{k,\varepsilon}\) denotes the optimal buffered expansion, then for every \(\delta>0\),
\[
h_G^{k,\varepsilon} \le O_\delta(1)\cdot \Big( \frac{\log k}{ \varepsilon}\Big) \cdot \lambda_{\lfloor (1+\delta) k\rfloor},
\]
with a randomized polynomial-time algorithm achieving the bound. The complementary lower bound is
\[
h_G^{k,\varepsilon}\ge \frac{\lambda_k-\varepsilon}{2}.
\]
This buffered formulation removes the square-root loss present in standard Cheeger inequalities by allowing a controlled buffer around each part [2308.10160].

## 4. Dual higher-order Cheeger inequalities and the top of the spectrum

Higher-order Cheeger theory has a dual version at the top end of the normalized Laplacian spectrum. For disjoint subsets \(V_1,V_2\subseteq V\), the dual conductance is
\[
\overline{\phi}(V_1,V_2) = \frac{2|E(V_1,V_2)|}{\operatorname{vol}(V_1\cup V_2)},
\]
and the multi-way dual Cheeger constant \(\overline h(k)\) is defined by maximizing the minimum dual conductance over \(k\) pairwise disjoint bipartite pairs [1401.4737].

The universal higher-order dual Cheeger inequality is
\[
\frac{2-\lambda_{N-k+1}}{2}\le 1-\overline h(k)\le Ck^3\sqrt{2-\lambda_{N-k+1}},
\]
or equivalently
\[
\frac{1}{C^2k^6}(1-\overline h(k))^2 \le 2-\lambda_{N-k+1} \le 2(1-\overline h(k)).
\]
This is the exact dual counterpart of the ordinary higher-order inequality: small bottom eigenvalues detect \(k\) disjoint sparse clusters, whereas small top spectral gaps \(2-\lambda_{N-k+1}\) detect \(k\) disjoint subgraphs that are close to bipartite [1401.4737].

The geometry of the dual theory is sign-insensitive. The top \(k\) eigenfunctions are assembled into a map
\[
F:V\to \mathbb R^k,\qquad F(v)=\bigl(f_{N-k+1}(v),\dots,f_N(v)\bigr),
\]
and clustering is performed not on the sphere but on the real projective space
\[
P^{k-1}\mathbb R=\mathbb S^{k-1}/\{\pm 1\},
\]
because vertices with opposite vectors should be regarded as close in the bipartite setting. This projective-space formulation is the distinguishing geometric feature of the dual theory [1401.4737].

The top-spectrum viewpoint also interacts with cut problems. In particular, a higher-order gap near the top improves the approximation ratio for maximum cut, in direct analogy with the way a higher-order bottom-spectrum gap sharpens sparse-cut guarantees [1301.5584]. Sharpness phenomena at the dual end are exhibited by the construction of the Bipartite Noisy Hypercube, which was introduced to prove the sharpness of the gap between spectral bipartite expansion and bipartite edge expansion in the dual version of the higher-order inequality [1504.06561].

## 5. Nonlinear, continuum, and limit-object generalizations

For the graph \(p\)-Laplacian, higher-order Cheeger inequalities involve variational eigenvalues and nodal domains. If \(f\) is an eigenfunction for the \(k\)-th variational eigenvalue \(\lambda_k^{(p)}\), \(p>1\), and \(m\) is the number of its strong nodal domains, then
\[
\left(\frac{2}{\tau(G)}\right)^{p-1}\left(\frac{h_m(G)}{p}\right)^p \le \lambda_k^{(p)} \le 2^{p-1}h_k(G),
\]
where
\[
\tau(G)=\max_{u\in V}\frac{d(u)}{\mu(u)}.
\]
If the eigenfunction associated to \(\lambda_k^{(p)}\) has exactly \(k\) strong nodal domains, the inequality becomes tight as \(p\to 1\) [1602.05567].

On bounded domains, the higher-order Cheeger constant
\[
h_k(\Omega) = \inf \left\{ \max_{1\le i\le k}\frac{P(E_i)}{|E_i|} : E_1,\dots,E_k \subset \Omega,\ E_i \cap E_j = \varnothing \right\}
\]
is linked to the \(k\)-th variational eigenvalue of the Dirichlet \(p\)-Laplacian. If \(\Omega\subset\mathbb R^n\) is bounded and has a comparable inscribed rectangle, then
\[
h_k(\Omega)\le C\, k^p\, \lambda_k(p,\Omega)^{1/p}.
\]
For bounded convex domains,
\[
C_1\, k^{1/n} \le h_k(\Omega) \le C_2\, k^{1/n},
\]
and hence \(h_k(\Omega)\simeq \lambda_k(p,\Omega)^{1/p}\) [1411.4737].

A related Euclidean theory studies the higher Cheeger problem for measurable sets \(\Omega\subset\mathbb R^N\), defining
\[
h_k(\Omega)=\inf\max\{h_1(E_1),\dots,h_1(E_k)\}
\]
over mutually disjoint subsets. This framework proves the existence of adjusted minimizers and establishes the \(p\to1\) spectral limit
\[
h_k(\Omega)=\lim_{p\to1}\mathfrak L_k(p;\Omega),
\]
where \(\mathfrak L_k(p;\Omega)\) is the spectral minimal \(k\)-partition functional for the Dirichlet \(p\)-Laplacian [1706.07282].

For Steklov problems, the \(k\)-th Steklov eigenvalue admits a higher-order Cheeger-type lower bound in finite spaces, measurable spaces, and compact Riemannian manifolds with boundary. In the manifold formulation,
\[
\sigma_k(M)\ge c\, h_k(M),
\]
more precisely
\[
\sigma_k(M) \ge c_2\, L_k(M),
\]
where \(L_k\) is the \(k\)-th Cheeger–Steklov constant. The proof factors through the Dirichlet–Steklov connectivity spectrum \(K_k\), with
\[
\sigma_k \ge c\,K_k \ge c\,L_k
\]
[1705.08643].

The finite-graph theory also extends to graphons. For a connected graphon \(W\), if \(h_W(k)\) denotes the \(k\)-way expansion constant and \(\lambda_k\) the \(k\)-th variational eigenvalue of the graphon Laplacian \(\Delta_W\), then
\[
\frac{\lambda_k}{2} \le h_W(k) \le O(k^{3.5})\sqrt{\lambda_k}.
\]
This is a graph-limit analogue of the Lee–Oveis Gharan–Trevisan theorem [2511.07016].

## 6. Higher-dimensional and simplicial-complex analogues

In the simplicial-complex literature, a central distinction is that “higher-order” in graph theory refers to multiple eigenvalues and multiple clusters, whereas “higher-dimensional” refers to Laplacians on chains or cochains of dimension \(k>0\). One paper states this distinction explicitly: the higher-order graph inequalities of Lee–Oveis Gharan–Trevisan concern higher-order graph partitions and higher Laplacian eigenvalues, not higher-dimensional simplicial Laplacians [1209.5091].

For finite simplicial complexes, two different expansion notions arise from cochains and chains over \(\mathbb Z_2\). In top dimension \(m\), the chain Cheeger number satisfies a genuine Cheeger/Buser-type inequality under natural geometric assumptions:
\[
h_m \ge \lambda_m \ge \frac{h_m^2}{2(m+1)}.
\]
By contrast, the analogous cochain statement fails in general: there exist families of simplicial \(m\)-balls \(X_k\) and \(Y_k\) for which \(h^{m-1}\) and \(\lambda^{m-1}\) decouple in both directions. The conclusion is that top-dimensional chain expansion behaves like a Dirichlet-boundary theory, whereas the coboundary version is too sensitive to topology and orientation for a universal Cheeger/Buser inequality [1209.5091].

A different line of work proves lower Cheeger inequalities for arbitrary finite \(k\)-dimensional simplicial complexes using the upper Laplacian. If \(\lambda(X)\) is the smallest nontrivial eigenvalue of the \((k-1)\)-dimensional upper Laplacian, then
\[
\lambda(X)\le h'(X),
\]
and also
\[
\lambda(X)\le \frac{C(X)}{|V|}\,h(X),
\]
where \(h(X)\) is a combinatorial expansion quantity adapted to the actual \((k-1)\)-skeleton, \(h'(X)\) is a refined cochain-based parameter, and \(C(X)\) is a completion-dependent incidence constant. These inequalities extend the earlier complete-skeleton result \(\lambda(X)\le h(X)\) to arbitrary complexes [1401.2290].

More recently, a new signed, oriented, \(\mathbb Z\)-valued Cheeger constant for simplicial complexes was introduced as the quantity compatible with the up-Laplacian or Eckmann Laplacian. In that formulation the higher-order inequalities control the spectral gap from the maximal possible eigenvalue \(d+2\):
\[
\frac{h_k(E_d)^2}{C\,k^6(d+1)} \le d+2-\lambda_{n+1-k}(\Delta^{\mathrm{up}}) \le 2\,h_k(E_d),
\]
and the one-set version is
\[
\frac{h_1(E_d)^2}{2(d+1)} \le d+2-\lambda_n(\Delta^{\mathrm{up}}) \le 2\,h_1(E_d).
\]
This program emphasizes that earlier \(\mathbb Z_2\)-based constants do not fully match the spectral behavior of the higher-dimensional up-Laplacian, whereas the signed \(\mathbb Z\)-valued constant does [2302.01069].

There is also a purely combinatorial higher-dimensional analogue that does not study multiway partitions in the usual graph-theoretic sense. For a finite connected pure \(n\)-dimensional simplicial complex, the quantity
\[
H(X)=\min_{0<|A|<|V|} \frac{|V|\cdot |F(A,V\setminus A)|}{|A|\cdot |V\setminus A|}
\]
counts top-dimensional faces crossing a two-way vertex cut. Its lower bounds are expressed in terms of the spectral gap of an embedded graph built from \((n-1)\)-faces, rather than in terms of a \(k\)-way partition problem. This is therefore a higher-dimensional Cheeger analogue, not a higher-order graph Cheeger inequality in the Lee–Oveis Gharan–Trevisan sense [2309.11785].

Across these variants, the stable core of the subject is the same: spectral multiplicity, or approximate multiplicity, is interpreted isoperimetrically. What changes from one setting to another is the correct notion of expansion—conductance, dual conductance, buffered expansion, chain expansion, signed simplicial expansion, or Steklov-type boundary expansion—and the success of a higher-order Cheeger theory depends on choosing that notion correctly.

Source: https://www.emergentmind.com/topics/higher-order-cheeger-inequality