Spectral graph theory & network design

Algebraic Connectivity Maximization

An ongoing research direction combining graph-theory-guided reinforcement learning for network design with exact spectral bounds and formally verified optimality proofs.

Year
2026
Status
Ongoing
Topics
Graph theory · Reinforcement learning · Spectral methods · Formal verification · Lean
01

The question

With a fixed number of vertices and edges, how can we build highly connected graphs—and prove when no better graph exists?

02

Central insight

Finding a strong graph and proving that it is optimal are different tasks. Structural graph families guide the construction algorithms, while spectral arguments establish what any competing graph can achieve.

Study 01 · 2026 · Submitted

Graph-Theory-Guided RL for Algebraic Connectivity

A structure-first method for constructing graphs with high algebraic connectivity under strict vertex and edge budgets.

A fixed-edge-budget graph and schematic consensus-error curves for different algebraic connectivity

A conceptual illustration of graph construction and spectral connectivity, not a plot of measured experimental results.

01

The question

With a fixed number of vertices and edges, which network structure coordinates most effectively?

02

Central insight

The hard part is not only choosing individual edges. Good solutions live near different structural families at different graph densities, so initialization should itself be a mathematical design variable.

03

Approach

  1. 01

    Route the target density toward complete multipartite or Cayley graph families.

  2. 02

    Construct a low-cost, high-quality structural backbone.

  3. 03

    Use reinforcement learning only for the smaller residual edge-allocation problem.

  4. 04

    Compare quality and runtime against greedy, optimization, and learning baselines.

04

My contribution

  • Developed the density-dependent backbone construction.
  • Formulated residual edge allocation as a sequential RL problem.
  • Designed and ran the quality-runtime comparison across graph sizes and densities.
05

A closer look

Why the graph family matters

In continuous-time consensus, each node moves toward its neighbors’ values. The Laplacian describes these interactions, and its second-smallest eigenvalue determines the slowest decaying disagreement mode. A larger value means faster worst-case convergence for this model. The design problem fixes both the number of nodes and the number of edges, so improvement must come from arranging the same resources differently.

The construction begins by choosing a structural family suited to the density. Complete multipartite graphs provide dense backbones; Cayley graphs provide candidates at lower and intermediate densities. This gives the learning stage a useful initial structure and a smaller remaining edge budget.

dx/dt = −Lx, L = D − A

What the policy learns—and what the experiments establish

The policy adds one missing edge at each step until the target budget is exhausted. A graph-attention network scores candidate pairs, and PPO training rewards increases in algebraic connectivity together with a terminal-quality bonus. The learned task is sequential completion of the backbone.

The experiments compare achieved connectivity and runtime against construction baselines, including Fiedler-vector greedy methods. The reported advantage concerns the tested sizes and densities; this method does not establish global optimality for arbitrary vertex and edge budgets. That distinction separates this computational design study from the exact 2(n − 2)-edge theorem within the same research direction.

06

Result

The backbone-plus-completion strategy achieved the strongest quality/runtime scaling among the tested non-exact baselines as graph size increased.

07

What remains

Characterize additional near-optimal graph families and make density-aware completion policies lighter and more interpretable.

Study 02 · 2026 · Preprint

Maximizing Algebraic Connectivity with 2(n − 2) Edges

A proof of Kolokolnikov’s conjecture: for every n ≥ 4, the maximum is 2, attained by the complete bipartite graph K₂,n−2. The work combines a mathematical proof with a Lean 4 formalization generated using MerLean.

The extremal graph K₂,₆ and its Laplacian spectrum: eight vertices, twelve edges, and λ₂ = 2

The example is K₂,₆: eight vertices and twelve edges. Assign +1 and −1 to two vertices in the larger part and zero elsewhere. This gives Lx = 2x; the spectrum confirms λ₂ = 2. The theorem shows that no graph with the same budget can exceed it.

01

The question

How much algebraic connectivity can a graph achieve with exactly 2(n − 2) edges?

02

Central insight

A hypothetical graph with connectivity above 2 must avoid certain local configurations. Those restrictions force a degree structure in which counting requires a short cycle that the spectral estimates forbid.

03

Approach

  1. 01

    Use Rayleigh test vectors to turn local graph structure into upper bounds on algebraic connectivity.

  2. 02

    Reduce a counterexample to minimum degree at least three, with no adjacent degree-three vertices.

  3. 03

    Combine degree and incidence counts with an exact non-backtracking Moore bound, and handle the remaining small orders through local cuts.

  4. 04

    Connect the full mathematical argument to the kernel-checked Lean theorem.

04

My contribution

  • Co-first author, with equal contribution alongside Jinzheng Li and Yuanjie Ren.
  • Worked on proving the conjecture and formalizing the result in Lean 4 using MerLean.
05

A closer look

From a candidate graph to a universal bound

Algebraic connectivity is the second-smallest eigenvalue of the graph Laplacian. Computing it for one construction identifies a candidate. Proving optimality requires controlling every simple graph with the same vertex and edge counts. Here K₂,n−2 supplies the candidate: every vertex in one part is connected to every vertex in the other.

|E| = 2(n − 2), n ≥ 4 ⇒ λ₂(G) ≤ 2 = λ₂(K₂,n−2)

Why the contradiction works

The Rayleigh quotient measures the energy of a zero-sum signal on the vertices. Carefully chosen signals certify λ₂ ≤ 2. Assuming the opposite excludes local configurations, leaving a rigid graph whose degree budget can be counted.

For n ≥ 48, an exact Moore bound forces a short cycle that spectral certificates exclude. An incidence argument covers 32 ≤ n ≤ 49; local sparse-set and cut arguments cover 4 ≤ n ≤ 31. The overlapping ranges exhaust all n ≥ 4.

What the formal proof establishes

The released Lean theorem covers every n ≥ 4, rather than a finite numerical search. The repository reports no unfinished proofs and only Lean’s standard axioms. MerLean generated the formal development; Lean checks the resulting proof terms. The accompanying manuscript presents the argument as a conventional mathematical proof.

06

Result

For all n ≥ 4, every finite simple graph with n vertices and exactly 2(n − 2) edges satisfies λ₂ ≤ 2. The graph K₂,n−2 reaches this bound. Both the upper bound and attainment are included in the formal theorem.

07

What remains

Classifying all equality cases is a separate problem: the theorem establishes a maximizer, without asserting uniqueness.

P

Papers & resources

Maximizing Algebraic Connectivity with 2(n − 2) Edges: The Large Vertex Number Case

Zeru Zhu, Jinzheng Li, Yuanjie Ren, and Ji Liu

arXiv:2608.07360 · Revised September 2026 · Preprint

Graph Theory Guided Reinforcement Learning for Network Design with High Algebraic Connectivity under Vertex and Edge Constraints

Zeru Zhu, Marco Gamarra, and Ji Liu

IEEE Conference on Decision and Control (CDC), 2026 · Submitted