Cluster-Graph Edit Distance: Metric Proxies, Multiscale Embeddings, and Complexity

Abstract: The cluster graphs on $n$ vertices, the disjoint unions of complete graphs, have the integer partitions of $n$ as their isomorphism classes, and the quotient edit distance $$q^{*(λ,μ)=\min_{σ\in} S_n}|E(G_λ)\triangleσE(G_μ)|$$ makes that set a metric space. Its metric geometry and its computational complexity both issue from one identity: $q<sup>*$ is an affine function of the maximum of $\lVert X\rVert_F<sup>2$ over the contingency tables with margins $λ$ and $μ$. Combinatorially, it yields two explicit $\ell_1$ models: the vertex-mass metric $δ_1$ on sorted degree sequences, with $\frac12δ_1\le q<sup>*&lt;\frac32δ_1$ and both constants optimal, and the block-energy metric $B$ on the vectors $\bigl(\binom{λ_i}2\bigr)_i$, with $q^*\le B\le2q<sup>*-1$ by a per-table refinement measuring how far an alignment is from a block bijection. Hence $c_1(\mathcal K_n)\le2$, and an $O(n\log n)$-time algorithm returns an alignment of cost below $2q^*$ with the certificate $q<sup>*\in[\lceil(B+1)/2\rceil,B]$. The Euclidean distortion of the class is $c_2(\mathcal K_n)=Θ(n<sup>{1/4})$; against it we measure the weighted dyadic sums $F<sup>{(γ)}$ of the Ferrers staircase, of dimension below $4n$ and computable in $O(n)$ time. The unweighted member has distortion exactly $Θ(n^{{1/4}\sqrt{\log} n})$, while the critical weight $γ=\frac14$ improves this unconditionally to $O(n^{1/4}(\log n)<sup>{1/4})$ through an inverse energy inequality proved from the quantization of staircase jumps; removing the residual $(\log n)<sup>{1/4}$ is reduced to one inverse inequality on the realizable cone. Computationally, the same identity gives a classification: deciding $q^*(λ,μ)\le Q$ is strongly NP-complete, evaluation is strongly NP-hard and admits no FPTAS unless $\mathrm P=\mathrm{NP}$, while the farthest alignment is polynomial-time solvable.

• The paper represents cluster-graph edit distance as a maximum-Frobenius-norm transportation problem, enabling sharp comparisons with vertex-mass and block-energy 1 models and a certified approximation below factor 2.
• The paper proves Euclidean distortion c(Kn)=Theta(n^{1/4}) and introduces an explicit critical dyadic embedding with distortion O(n^{1/4}(log n)^{1/4}), improving substantially on sorted-degree coordinates.
• The paper establishes that nearest alignment is strongly NP-complete with no FPTAS, whereas farthest alignment is polynomial-time solvable, and it identifies open problems around optimal

The transportation representation

The paper studies the quotient graph edit distance $$q^*(\lambda,\mu)=\min_{\sigma\in S_n}|E(G_\lambda)\triangle\sigma E(G_\mu)|$$ on the class $\mathcal K_n$ of cluster graphs — disjoint unions of complete graphs — whose isomorphism classes are exactly the integer partitions of $n$. Since cluster graphs are determined by their block-size multisets, $q^*$ is a metric on partitions, and it coincides (up to a factor of 2) with the Mirkin metric on unlabelled clustering profiles, connecting the object to pair-counting comparison of clusterings.

The central result, and the hub from which every other statement derives, is the transportation representation (2608.17990):

$$q^*(\lambda,\mu)\;=\;\frac{\lVert\lambda\rVert_2^2+\lVert\mu\rVert_2^2}{2}\;-\;\max_{X\in T(\lambda,\mu)}\lVert X\rVert_F^2,$$

where $T(\lambda,\mu)$ is the set of contingency tables with margins $\lambda$ and $\mu$. This converts a minimum over $n!$ vertex bijections into a maximum over transportation polytopes. Integrality of the maximizer follows from total unimodularity via Hoffman–Kruskal, so real-valued certificates suffice for upper bounds throughout.

The extremal problem $\max_X \lVert X\rVert_F^2$ is classical in the comparison-of-clusterings literature (Hubert–Arabie, Lerman–Peter, Messatfa, Chacón), where an exact bound was described as "a very difficult problem of combinatorial optimization." Its strong NP-hardness was previously known via entropy-of-couplings results; what this paper adds is the identification with a graph edit distance, NP-completeness of the decision version under explicit part-list encoding, parameterized lower bounds, and the direction dichotomy discussed below.

Two $\mathcal K_n$0 models with optimal constants

The combinatorial reading of the representation yields two distinct linearizations of $\mathcal K_n$1, each two-sidedly equivalent to it with optimal constants, and neither subsumes the other:

Model	Definition	Relation to $\mathcal K_n$2	Optimal constant
Vertex-mass $\mathcal K_n$3	$\mathcal K_n$4	$\mathcal K_n$5	$\mathcal K_n$6 attained; $\mathcal K_n$7 supremum only
Block-energy $\mathcal K_n$8	$\mathcal K_n$9, $n$0	$n$1	$n$2 approached along $n$3 vs. $n$4

The upper bound for the vertex-mass model routes mass by rank through elementary unit transfers in the partition lattice, each step's exact cost computed from the two-block subsystem solved by a convex-parabola endpoint argument. The reverse inequality $n$5 comes from a per-table refinement stronger than a global bound: writing each alignment's cost as splitting plus merging (via $n$6), a Wasserstein triangle inequality on spectra of block energies gives

$n$7

valid for every feasible table $n$8, whose deficiencies vanish only at block bijections. This yields automatic strictness in the sandwich and identifies the slack as a precise measure of distance from a block bijection. A spectral reformulation via Mirsky's trace-norm inequality shows that relaxing the combinatorial alignment to a unitary one changes the distance by at most a factor of 2.

Three consequences follow directly. First, an $n$9-time algorithm (in the number of vertices) outputs a feasible alignment of cost strictly below $q^*$0 together with the certificate interval $q^*$1. Second, the sorted-degree map has product distortion exactly below $q^*$2 per fixed $q^*$3 with supremum $q^*$4, while the block-energy map gives $q^*$5 — though whether $q^*$6 holds for the class remains open, with no lower bound above 1 known to the authors. Third, exhaustive enumeration up to $q^*$7 confirms $q^*$8 throughout, always attained by the $q^*$9-versus-$$q^*(\lambda,\mu)\;=\;\frac{\lVert\lambda\rVert_2^2+\lVert\mu\rVert_2^2}{2}\;-\;\max_{X\in T(\lambda,\mu)}\lVert X\rVert_F^2,$$0 family.

Euclidean distortion of the class

The paper proves $$q^*(\lambda,\mu)\;=\;\frac{\lVert\lambda\rVert_2^2+\lVert\mu\rVert_2^2}{2}\;-\;\max_{X\in T(\lambda,\mu)}\lVert X\rVert_F^2,$$1. The lower half embeds a Hamming cube of dimension $$q^*(\lambda,\mu)\;=\;\frac{\lVert\lambda\rVert_2^2+\lVert\mu\rVert_2^2}{2}\;-\;\max_{X\in T(\lambda,\mu)}\lVert X\rVert_F^2,$$2 into the space using block sizes separated by gaps of 2 (which buys both injectivity and coordinate-specific thresholds), then applies Enflo's theorem: cube distortion $$q^*(\lambda,\mu)\;=\;\frac{\lVert\lambda\rVert_2^2+\lVert\mu\rVert_2^2}{2}\;-\;\max_{X\in T(\lambda,\mu)}\lVert X\rVert_F^2,$$3 transfers to give $$q^*(\lambda,\mu)\;=\;\frac{\lVert\lambda\rVert_2^2+\lVert\mu\rVert_2^2}{2}\;-\;\max_{X\in T(\lambda,\mu)}\lVert X\rVert_F^2,$$4. The construction cannot exceed $$q^*(\lambda,\mu)\;=\;\frac{\lVert\lambda\rVert_2^2+\lVert\mu\rVert_2^2}{2}\;-\;\max_{X\in T(\lambda,\mu)}\lVert X\rVert_F^2,$$5 since the weights satisfy $$q^*(\lambda,\mu)\;=\;\frac{\lVert\lambda\rVert_2^2+\lVert\mu\rVert_2^2}{2}\;-\;\max_{X\in T(\lambda,\mu)}\lVert X\rVert_F^2,$$6 and must sum to at most $$q^*(\lambda,\mu)\;=\;\frac{\lVert\lambda\rVert_2^2+\lVert\mu\rVert_2^2}{2}\;-\;\max_{X\in T(\lambda,\mu)}\lVert X\rVert_F^2,$$7 — so this method caps out at exactly the achieved exponent.

The upper half proceeds by exhibiting $$q^*(\lambda,\mu)\;=\;\frac{\lVert\lambda\rVert_2^2+\lVert\mu\rVert_2^2}{2}\;-\;\max_{X\in T(\lambda,\mu)}\lVert X\rVert_F^2,$$8 as a bounded-distortion image of a genuine finite subset of $$q^*(\lambda,\mu)\;=\;\frac{\lVert\lambda\rVert_2^2+\lVert\mu\rVert_2^2}{2}\;-\;\max_{X\in T(\lambda,\mu)}\lVert X\rVert_F^2,$$9 (via either $T(\lambda,\mu)$0 model) and quoting Chang–Naor–Ren: every $T(\lambda,\mu)$1-point subset of $T(\lambda,\mu)$2 has Euclidean distortion $T(\lambda,\mu)$3. With $T(\lambda,\mu)$4 by Hardy–Ramanujan, this gives $T(\lambda,\mu)$5. Two caveats deserve emphasis. First, this step depends on the metric theorem of (Chang et al., 2024), which appears only in the full preprint rather than its STOC abstract; declining it falls back to Arora–Lee–Naor, weakening the upper bound to $T(\lambda,\mu)$6. Second, the resulting embedding is an existence statement: no coordinate map evaluable on a single input in polynomial time is supplied by that argument, and computing $T(\lambda,\mu)$7 via SDP would require enumerating all $T(\lambda,\mu)$8 classes together with strongly-NP-hard distance evaluations.

Against this class optimum, the natural scaled sorted-degree coordinate $T(\lambda,\mu)$9 has distortion pinned to leading order, $\lambda$0 — a factor $\lambda$1 too large. The leading constant is determined because three separate inequalities (routing, staircase transform, Cauchy–Schwarz) saturate simultaneously on one witness family; the loss traces specifically to the passage from $\lambda$2 to $\lambda$3 across $\lambda$4 coordinates.

Multiscale coordinates and the critical exponent

The paper introduces a family of weighted dyadic coordinates on the Ferrers staircases $\lambda$5:

$\lambda$6

of dimension $\lambda$7, injective, computable in $\lambda$8 time from either partition encoding. The analysis rests on two structural facts special to partitions: the difference $\lambda$9 is piecewise constant with $\mu$0 runs (since $\mu$1), and jumps obey an arithmetic rigidity — $\mu$2, so any jump at position $\mu$3 costs amplitude at least $\mu$4, within the budget $\mu$5.

The unweighted member is pinned exactly. For $\mu$6, two witnesses saturate the two Lipschitz constants separately: a spike pair (one tall block against a slightly shorter one), coherent across all $\mu$7 scales, forces expansion $\mu$8; a breathing train of nearly mean-free dipoles cancels above level 1, forcing contraction $\mu$9. Hence $n!$0 — not an artefact of the estimates but of the map itself, since Theorem 5.5(C) shows no reweighting can improve the guarantee obtainable from the two-step argument quantifying over arbitrary vectors.

The critical weight $n!$1. Super-critical weights fail on the most classical pair, $n!$2 versus the empty graph, whose sign-coherent plateau lives entirely in coarse scales: for $n!$3, $n!$4. At $n!$5 the paper proves an unconditional inverse-energy inequality,

$n!$6

where $n!$7 is the variation localized to octave $n!$8. The proof replaces single-interval capture by full Whitney tilings of sign-constant runs (recovering entire run mass instead of a quarter of it), then uses jump quantization twice — once to bound runs per octave, once across octaves — yielding $n!$9, an unconditional strict improvement over the pinned $\max_X \lVert X\rVert_F^2$0 ceiling and reportedly the best distortion known for any explicit pointwise $\max_X \lVert X\rVert_F^2$1-time coordinate on this space.

The residual $\max_X \lVert X\rVert_F^2$2 is isolated precisely: it is the gap between $\max_X \lVert X\rVert_F^2$3 and $\max_X \lVert X\rVert_F^2$4, i.e., the price of a flat octave spectrum. The critical inverse-energy conjecture states that $\max_X \lVert X\rVert_F^2$5 on the realizable cone. It is verified elementarily on three sub-cones (sign-coherent pairs, balanced pairs with $\max_X \lVert X\rVert_F^2$6, geometric octave spectra), and an exactly solvable chirp family — interleaved doubled odd ladders meeting the variation budget $\max_X \lVert X\rVert_F^2$7 with nothing to spare — caps the best possible constant at $\max_X \lVert X\rVert_F^2$8, with numerical descent of $\max_X \lVert X\rVert_F^2$9 to $\mathcal K_n$00 against $\mathcal K_n$01 at $\mathcal K_n$02. If proved, the conjecture would upgrade $\mathcal K_n$03 to class-optimal distortion $\mathcal K_n$04. The authors state plainly that they have no approach to the general case and claim none; notably, the inequality is false for arbitrary vectors (an alternating pattern defeats it), so any proof must pass through the arithmetic rigidity of realizability.

Computational classification

The computational reading exploits degree and direction of the optimization over $\mathcal K_n$05:

• Nearest alignment is hard. Deciding $\mathcal K_n$06 is strongly NP-complete, via reduction from 3-Partition with $\mathcal K_n$07, $\mathcal K_n$08: the threshold $\mathcal K_n$09 is attainable iff every column of the table is concentrated, i.e., iff a valid 3-partition exists. Hardness persists when all blocks of $\mathcal K_n$10 have equal size. Parameterized by the number $\mathcal K_n$11 of parts on the uniform side, unary encoding gives W[1]-hardness and no $\mathcal K_n$12 algorithm under ETH; binary encoding is already para-NP-hard at $\mathcal K_n$13 (reducing PARTITION). Neither $\mathcal K_n$14 nor the Frobenius maximum admits an FPTAS unless P = NP, by rounding arguments exploiting integral objectives with polynomially bounded optima on the strongly-hard subfamily.
• Farthest alignment is easy. $\mathcal K_n$15 reduces to minimizing $\mathcal K_n$16, a separable convex cost flow solvable in polynomial time via Minoux's algorithm. The dichotomy is worth stating carefully: the known polynomial algorithms for quadratic transportation all minimize energy, hence compute the farthest, not the nearest, alignment.
• Consistency. The FPTAS exclusion coexists with the strict 2-approximation of Corollary 3.12 — a constant-factor guarantee is not an approximation scheme. No PTAS exclusion or APX-hardness is claimed, since the reduction carries zero gap.

Limitations and open questions

Several dependencies and gaps are conceded explicitly. The upper bound $\mathcal K_n$17 rests on the metric theorem of Chang–Naor–Ren, available only in preprint form; without it the order degrades by a logarithmic factor. All complexity statements require the explicit part-list encoding (P1); under compressed multiplicity encoding the NP-certificate can be exponentially longer than the input, and no membership claim is made there. The behaviour of $\mathcal K_n$18 for $\mathcal K_n$19 is not classified. The value $\mathcal K_n$20 is open between 1 and 2. Whether a PTAS exists, or APX-hardness holds, has no evidence in either direction, and the optimality gap of the ratio-2 approximation is unknown. The corresponding questions for cographs and trees remain open, the present methods resting on the block–contingency-table correspondence those classes lack. Finally, whether every map linear in $\mathcal K_n$21 is stuck at $\mathcal K_n$22 — which would force any resolution of the critical conjecture through weighted or nonlinear coordinates — is unresolved.

Conclusion

For cluster graphs the paper delivers a complete picture assembled from one identity. Geometrically, $\mathcal K_n$23 is established with matching-order bounds, and the gap between existence and explicit construction is reduced from a factor $\mathcal K_n$24 (the sorted-degree coordinate) to $\mathcal K_n$25 by the critical dyadic map $\mathcal K_n$26, unconditionally. Computationally, nearest alignment is strongly NP-complete with no FPTAS while farthest alignment is polynomial — a direction dichotomy induced solely by the degree and orientation of a convex objective on a transportation polytope. The remaining central question is sharply localized: whether a realizable staircase difference can carry a flat octave spectrum, which alone separates the current best explicit embedding from class-optimal distortion.