- The paper develops an existence theory for spectral minimal partitions on both finite and infinite weighted combinatorial graphs, focusing on three Laplacian realisations: Dirichlet, Neumann, and boundaryless.
- Key results include the determination of existence conditions for minimizers under different Laplacian realisations and the derivation of quantitative energy bounds, demonstrating the impact of boundary effects and non-contractible paths.
- The paper also highlights that different Laplacian realisations yield structurally different optimal partitions, exemplified by the ladder graph, which shows the significance of the chosen realization in the minimization problem.
This paper develops an existence theory for spectral minimal partitions on weighted combinatorial graphs, finite or countably infinite, extending the framework previously established for Euclidean domains and metric graphs. Given a partition P=(V1,…,Vk) of a graph G into connected, pairwise disjoint clusters, the authors study functionals of the form
Ek,p⋆(P)=(k1j=1∑kλ1⋆(Vj)p)1/p,p∈[1,∞],
where λ1⋆(Vj) is the first positive eigenvalue of one of three self-adjoint realisations of the Laplacian on the induced subgraph: Dirichlet (λ1D), Neumann (λ1N), and boundaryless (λ1B). The goal is to determine when the infimum of these energies over admissible partitions is actually attained.
Framework and the three Laplacian realisations
The graph G carries vertex weights ν and edge weights ω, with local finiteness G0. Edge weights induce a shortest-path metric via G1 as edge length, making G2 a metric measure space in the sense of Sturm (2608.19962). The energy form is G3, with form domain G4; boundary conditions at infinity (a Royden boundary) are encoded by choosing a closed intermediate form domain G5 between G6 and G7.
The Dirichlet energy is defined by restricting the ambient form to functions supported in G8: it decomposes as an internal Laplacian plus a Schrödinger potential G9 measuring the weighted degree flowing out across the boundary Ek,p⋆(P)=(k1j=1∑kλ1⋆(Vj)p)1/p,p∈[1,∞],0. The boundaryless energy is the plain Laplacian of the induced subgraph Ek,p⋆(P)=(k1j=1∑kλ1⋆(Vj)p)1/p,p∈[1,∞],1, ignoring the ambient graph; its first positive eigenvalue is well-defined whenever the subgraph is connected. The Neumann energy, following Shi–Yu, uses a non-isometric extension operator that averages values over the boundary Ek,p⋆(P)=(k1j=1∑kλ1⋆(Vj)p)1/p,p∈[1,∞],2, so that Neumann conditions hold there; its associated quadratic form is shown to be closed.
A key structural result is the hierarchy: uniform Poincaré-type control (Assumption 3.4 for the boundaryless case) implies the analogous Neumann assumption (Assumption 3.6), which in turn implies canonical compactifiability (continuous embedding of Ek,p⋆(P)=(k1j=1∑kλ1⋆(Vj)p)1/p,p∈[1,∞],3 into Ek,p⋆(P)=(k1j=1∑kλ1⋆(Vj)p)1/p,p∈[1,∞],4), which yields compact embedding into Ek,p⋆(P)=(k1j=1∑kλ1⋆(Vj)p)1/p,p∈[1,∞],5 and hence pure point spectrum for any admissible realisation. Geometrically, both Poincaré-type assumptions hold if Ek,p⋆(P)=(k1j=1∑kλ1⋆(Vj)p)1/p,p∈[1,∞],6 has finite maximal non-contractible path length; sufficient conditions include finite total length Ek,p⋆(P)=(k1j=1∑kλ1⋆(Vj)p)1/p,p∈[1,∞],7, or finite diameter together with finitely many independent cycles.
An explicit prism-graph example demonstrates that the three notions genuinely diverge: for certain induced subgraphs the ordering of cluster sizes by energy differs across all three functionals, foreshadowing that minimizers depend strongly on the chosen realization.
Existence theory
For finite graphs existence is trivial, since only finitely many partitions exist. The paper's contribution concerns infinite graphs, where the strategy is to reduce to the finite case via a topology of discrete convergence: a sequence of vertex sets converges iff their characteristic functions converge pointwise, a metrizable topology strictly weaker than Hausdorff convergence.
Three ingredients drive the argument. First, every sequence of Ek,p⋆(P)=(k1j=1∑kλ1⋆(Vj)p)1/p,p∈[1,∞],8-partitions has a discretely convergent subsequence (a diagonal extraction over a countable vertex set), and limits of partitions whose limit clusters are nonempty are again partitions. Second, under the compactness assumptions above, normalized first-eigenvectors along minimizing sequences have subsequences converging in Ek,p⋆(P)=(k1j=1∑kλ1⋆(Vj)p)1/p,p∈[1,∞],9, and Fatou's lemma gives lower semicontinuity of each energy, so λ1⋆(Vj)0. Third, if between any two vertices there are only finitely many non-contractible paths — automatically true on ladders — then the class of connected-cluster partitions is sequentially closed under discrete convergence.
The resulting theorems state:
- Dirichlet case: if λ1⋆(Vj)1 is compact, spectral minimal λ1⋆(Vj)2-partitions exist for all λ1⋆(Vj)3 and all λ1⋆(Vj)4, including connected ones (disconnected clusters can be repaired without increasing the energy, using monotonicity of Dirichlet eigenvalues under inclusion).
- Boundaryless and Neumann cases: under the respective uniform Poincaré assumptions plus finite-noncontractible-paths, connected spectral minimal λ1⋆(Vj)5-partitions exist for all λ1⋆(Vj)6 and λ1⋆(Vj)7.
Consequently, on any infinite graph with finite λ1⋆(Vj)8, finite λ1⋆(Vj)9, and finitely many non-contractible paths per pair of vertices, all three problems admit minimizers for every λ1D0.
Spectral inequalities
The paper derives quantitative lower bounds on the optimal energies. Under finite λ1D1 and finite total length,
λ1D2
with the factor-λ1D3 loss in the Dirichlet bound arising because the realizing paths for the inradii may share boundary edges. For unweighted finite graphs on λ1D4 vertices, sharper bounds of the form λ1D5 and λ1D6 follow from Fiedler-type inequalities. The map λ1D7 is continuous and non-decreasing.
A notable deviation from the domain and metric-graph theory: the inequality λ1D8 familiar from Euclidean domains fails here; instead the paper proves the weaker bound λ1D9, with the factor 2 forced by boundary edges shared between two clusters. A two-vertex graph shows even the uncorrected constant-1 version cannot hold: λ1N0 while λ1N1.
Ladder graphs: divergence of minimizers and failure of existence
For the finite ladder λ1N2 with unit weights, exhaustive λ1N3-partitions necessarily consist of clusters of the form λ1N4 (a ladder of λ1N5 rungs with a pendant path of length λ1N6). A rearrangement argument based on the Perron–Frobenius positivity of the ground state shows that among such clusters with fixed vertex count, energy is minimized by concentrating rungs: λ1N7 whenever λ1N8. This yields a clean trichotomy of unique-or-characterized minimizers:
| Energy |
Minimizing 2-partition |
| Dirichlet |
two copies of λ1N9 (λ1B0 even) or two copies of λ1B1 (λ1B2 odd) |
| Boundaryless |
two path graphs λ1B3, either horizontal or λ1B4-shaped (λ1B5) |
| Neumann |
horizontal paths λ1B6 only |
Thus the three functionals select structurally different optimal partitions of the same graph — Dirichlet favors balanced ladder-like cells, boundaryless favors path-like cells in two configurations, and Neumann uniquely selects the horizontal cut, since the λ1B7-shaped cell's Neumann eigenvalue strictly exceeds its boundaryless value by the equality-case analysis of Shi–Yu.
The infinite ladder λ1B8 illustrates why the summability hypotheses are not technical artifacts. With unit weights, the boundaryless and Neumann energies are not even well defined on any infinite cluster: the variational infima equal zero but are not attained, since zero energy would require a non-λ1B9 constant. In the Dirichlet case, if one admits infimum-of-spectrum as the energy, the global infimum is G0 for all G1, attained only for G2 (any bipartition into two semi-infinite ladders); for G3 no partition attains the infimum, mirroring a known phenomenon for unbounded metric graphs. Under finite G4 and G5, by contrast, all problems are well posed and minimizers exist for all G6.
The paper leaves open whether there exist graphs G7, integers G8, and exponents G9 for which the Neumann-optimal partition differs from both the Dirichlet and boundaryless optimizers — on ladders the Neumann minimizer coincides with one of the boundaryless minimizers.
Limitations and open questions
Several restrictions frame the results. Only exhaustive partitions into connected clusters are treated; relaxed notions (non-connected clusters, partitions of unity as in Osting–White–Oudet) are excluded, though Lemma 5.9 shows connectivity is not a genuine restriction for Dirichlet energies under compactness. Robin-type realisations are explicitly not considered. The existence theorems for Neumann and boundaryless energies require the uniform Poincaré assumptions, whose geometric sufficiency conditions (finite maximal path length, or finite diameter plus finitely many cycles) are restrictive; the authors note other routes to discreteness, such as sparsity conditions, but do not integrate them into the existence theory. The general claim that a discretely convergent minimizing sequence attaining a connected limit yields a minimizer is proven only under the finite-noncontractible-paths hypothesis ensuring closedness of the connected class. Whether the factor ν0 in ν1 is sharp, and whether the higher-order comparison ν2 extends verbatim to the infinite setting beyond the first eigenvalue, are left unaddressed.
Conclusion
The paper supplies the first systematic existence theory for spectral minimal partitions of infinite weighted combinatorial graphs, built on canonical compactifiability, uniform Poincaré-type inequalities, discrete set convergence, and a closedness criterion for connected partitions expressed through non-contractible paths. It establishes quantitative lower bounds on all three optimal energies, identifies a structural correction (the factor 2) to the eigenvalue-partition inequality known from domains and quantum graphs, and demonstrates through the ladder model that the choice of Laplacian realisation qualitatively changes the optimizer — while showing precisely how, absent weight summability, the problem degenerates on infinite graphs.