Cut Hierarchy: Multi-Level Refinements

• Cut hierarchy is a design pattern that refines cut decisions using multi-scale structures, either via recursive decompositions or hierarchical relaxations.
• It applies to clustering with dendrogram pruning and adaptive cutting, leading to measurable improvements in metrics like partition density and modularity.
• Hierarchical methods also optimize classical, quantum, and applied cut problems, enhancing approximation quality and computational efficiency.

Search arXiv for papers on "cut hierarchy" across clustering, optimization, graph algorithms, and quantum max cut. Cut hierarchy is not a single universal construction. In current arXiv usage, the term and closely related phrases denote several families of hierarchical objects built around cut decisions: multi-level pruning rules on dendrograms, level-by-level LP/SDP/SoS relaxations for cut objectives, recursively maintained graph decompositions that preserve cut-flow structure, and domain-specific hierarchical cut mechanisms in road networks, ray tracing, MILP solvers, and low-multiplicity event mixing (Ge et al., 2022, Hopkins et al., 2019, Goranci et al., 2020, Farhan et al., 2023). The unifying theme is that a cut is not treated as a single flat decision: either the cut is chosen at multiple scales in a tree, or feasibility and optimality are refined through a hierarchy of increasingly structured constraints.

1. Core meanings of cut hierarchy

Across the literature, two broad meanings recur. First, a cut hierarchy can be a recursive decomposition: a tree, dendrogram, or multilevel partition in which different substructures are cut at different depths. Second, it can be a relaxation hierarchy: a sequence of LP, SDP, SoS, or ncSoS programs whose higher levels impose stronger consistency conditions on cut variables or cut-induced observables (Boucherie et al., 9 Dec 2025, Bhangale et al., 2018, Watts et al., 2023).

Setting	Hierarchical object	Representative papers
Clustering	Multi-level dendrogram cuts or optimal pruning paths	(Ge et al., 2022, Boucherie et al., 9 Dec 2025)
Classical cut optimization	Sherali-Adams, Lasserre, and related relaxations	(Hopkins et al., 2019, Chlamtac et al., 2010, Bhangale et al., 2018)
Quantum cut optimization	Symmetry-adapted ncSoS or NPA hierarchies	(Watts et al., 2023, Takahashi et al., 2023)
Graph data structures	Expander hierarchies and balanced tree hierarchies	(Goranci et al., 2020, Farhan et al., 2023)
Applied cut control	Hierarchy cut code, hierarchical cut selection, IMEHC cut	(Xiang et al., 2023, Wang et al., 2023, He, 2018)

A useful synthesis is that the first meaning refines where a cut is placed, while the second refines what constraints certify a cut’s quality. This suggests that “cut hierarchy” is best understood as a design pattern rather than a single formal definition.

2. Dendrogram pruning and multi-level cutting

In hierarchical clustering, the simplest pruning rule is a single horizontal cut. Two recent lines of work show that this is often suboptimal. “Weakest-link optimal pruning” formulates pruning as a cost-complexity problem on a dendrogram $T$, with within-cluster dispersion

$$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$

and penalized objective

$R_\alpha(T)=R(T)+\alpha |\tilde T|.$

At each step, the weakest link is the internal node whose collapse causes the smallest increase in loss per leaf removed. For every fixed number of clusters $k$, the size-$k$ subtree on the weakest-link path has $R(T)$ no larger than any size-$k$ subtree obtainable by a horizontal cut (Ge et al., 2022).

The adaptive-cut framework generalizes this idea from one-dimensional cut-height selection to combinatorial multi-cut selection on a dendrogram $T$. It optimizes

$\pi^\star=\arg\max_{\pi\in X(T)} f(\pi),$

where $X(T)$ is the set of partitions obtainable by multi-cutting $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$0, and $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$1 can be link-clustering partition density $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$2 or graph modularity $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$3. The search is performed by an MCMC with simulated annealing, initialized at the single-cut optimum, and driven by merge/split moves along the dendrogram (Boucherie et al., 9 Dec 2025).

A central diagnostic in the adaptive-cut paper is the balancedness score

$$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$4

where $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$5 is the entropy of the level-$$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$6 partition. The paper shows $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$7, with $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$8 for perfectly even splits and $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$9 for caterpillar dendrograms. Empirically, when $R_\alpha(T)=R(T)+\alpha |\tilde T|.$0, gains from multi-level cuts can be large: on a varying-density SBM, partition density improved from $R_\alpha(T)=R(T)+\alpha |\tilde T|.$1 to $R_\alpha(T)=R(T)+\alpha |\tilde T|.$2 (about $R_\alpha(T)=R(T)+\alpha |\tilde T|.$3), AMI improved by about $R_\alpha(T)=R(T)+\alpha |\tilde T|.$4–$R_\alpha(T)=R(T)+\alpha |\tilde T|.$5, and on more than 200 real networks, link-clustering gains were typically $R_\alpha(T)=R(T)+\alpha |\tilde T|.$6–$R_\alpha(T)=R(T)+\alpha |\tilde T|.$7 when $R_\alpha(T)=R(T)+\alpha |\tilde T|.$8; modularity gains reached $R_\alpha(T)=R(T)+\alpha |\tilde T|.$9–$k$0 on highly unbalanced trees (Boucherie et al., 9 Dec 2025).

These results directly address a common misconception: that a dendrogram is fully summarized by a single cut height. The pruning and adaptive-cut literature instead shows that the relevant hierarchy is often the tree itself, not a scalar threshold.

3. Relaxation hierarchies for classical cut problems

For classical cut optimization, “cut hierarchy” often means a sequence of progressively tighter relaxations. In Max-Cut, the degree-$k$1 Sherali-Adams relaxation introduces local distributions or pseudoexpectations over subsets of size at most $k$2, with objective

$k$3

A principal result is that for every constant $k$4, degree $k$5 Sherali-Adams, of size $k$6, achieves a $k$7-approximation with $k$8. The same work gives additive-$k$9 approximation on low threshold-rank graphs at degree controlled by the threshold rank, and extends the framework to Unique Games (Hopkins et al., 2019).

For Sparsest Cut on bounded-treewidth graphs, Sherali-Adams appears in a different role. A level-$k$0 SA relaxation is solved locally on bags of a width-$k$1 tree decomposition, and the rounding algorithm patches local distributions bag-by-bag to produce a global cut. The resulting algorithm gives the first constant-factor approximation for Sparsest Cut with general demands in bounded-treewidth graphs, running in $k$2 time (Chlamtac et al., 2010).

For simultaneous Max-Cut, a constant-level Lasserre hierarchy strengthens the natural SDP by enforcing large cut value even after conditioning on small partial assignments. Combined with a Raghavendra–Tan style rounding and a degree-7 bias function,

$k$3

this yields a polynomial-time $k$4-approximation for every constant number $k$5 of instances (Bhangale et al., 2018).

The landscape is not monotone across problem variants. For MAX BISECTION, a 2025 integrality-gap construction shows that the standard two-phase paradigm—first obtaining an $k$6-uncorrelated Basic SDP solution, then rounding it to an almost balanced cut—cannot achieve the Goemans–Williamson ratio $k$7 if it relies only on $k$8-uncorrelatedness. The paper constructs an explicit instance with ratio below $k$9 for some $R(T)$0-uncorrelated solution of the Basic SDP relaxation (Brakensiek et al., 4 Dec 2025).

Taken together, these results show that hierarchy strength is problem-sensitive. A stronger hierarchy can close gaps for Max-Cut or simultaneous Max-Cut, but analogous low-level information may still be insufficient for balanced variants.

4. Quantum Max-Cut hierarchies

Quantum Max-Cut has produced a distinct family of cut hierarchies built from operator algebras rather than classical cut variables. One approach replaces the usual Pauli-based quantum Lasserre viewpoint with a hierarchy over the algebra generated by swap operators $R(T)$1, using the identity

$R(T)$2

The resulting swap-ncSoS hierarchy enforces a finite presentation of the swap algebra, including $R(T)$3, braid relations, disjoint commutation, and the triangle-pair identity. Its level-$R(T)$4 relaxation optimizes a moment matrix over monomials in the swap generators. The hierarchy converges finitely, degree $R(T)$5 suffices, and level 2 was numerically exact up to tolerance $R(T)$6 on all unweighted instances with at most 8 vertices (Watts et al., 2023).

A parallel development builds an SU(2)-symmetric NPA hierarchy in the projector variables

$R(T)$7

Its moment matrices enforce positivity, normalization, and SU(2)-specific algebraic constraints, including

$R(T)$8

This hierarchy also converges finitely, with exactness guaranteed by level $R(T)$9, though many graph families are solved much earlier (Takahashi et al., 2023).

The two quantum papers emphasize that symmetry-adapted hierarchies can be both smaller and stronger than generic Pauli-based relaxations. The swap-based paper reports that level 2 often outperforms or matches Pauli-ncSoS at far lower SDP size, while the SU(2)-symmetric paper gives analytic exactness at level 1 for star graphs and complete bipartite graphs, exactness for even $k$0, and a level-1 gap of exactly $k$1 for odd $k$2 (Watts et al., 2023, Takahashi et al., 2023).

A plausible implication is that in quantum cut problems, the decisive hierarchy is often the symmetry hierarchy of the Hamiltonian rather than the generic hierarchy of local observables.

5. Structural graph hierarchies and parameter landscapes

A different notion of cut hierarchy arises when a graph is recursively decomposed so that cuts and flows are preserved approximately at every level. The expander hierarchy is such a construction. It starts from a boundary-linked expander decomposition, recursively contracts clusters, and forms a tree whose edges are annotated with capacities. The resulting tree is a tree-flow sparsifier of quality $k$3, and the hierarchy can be maintained fully dynamically with $k$4 worst-case update time (Goranci et al., 2020).

The consequences are extensive. The same paper gives the first fully dynamic deterministic algorithm with $k$5 worst-case update time that supports $k$6-approximate conductance, $k$7-$k$8 maximum flows, and $k$9-$T$0 minimum cuts in $T$1 query time; a deterministic fully dynamic connectivity algorithm with $T$2 worst-case update time; and a dynamic treewidth-decomposition algorithm on constant-degree graphs maintaining width $T$3 (Goranci et al., 2020).

Balanced tree hierarchy serves an analogous role in road-network indexing. In Hierarchical Cut 2-Hop Labelling, the hierarchy is a rooted binary tree whose cuts are $T$4-balanced and satisfy the cut-vertex condition that every pair $T$5 has a cut-vertex in $T$6 lying on some shortest path. Querying reduces to finding the LCA in $T$7 and scanning only the label arrays at that cut level in $T$8. On ten real-world road networks, HC2L is reported to be $T$9–$\pi^\star=\arg\max_{\pi\in X(T)} f(\pi),$0 faster per query than the baselines, with label size up to $\pi^\star=\arg\max_{\pi\in X(T)} f(\pi),$1 smaller; the parallel variant HC2L$\pi^\star=\arg\max_{\pi\in X(T)} f(\pi),$2 is $\pi^\star=\arg\max_{\pi\in X(T)} f(\pi),$3–$\pi^\star=\arg\max_{\pi\in X(T)} f(\pi),$4 faster in construction than any baseline (Farhan et al., 2023).

Another use of “cut hierarchy” is classificatory rather than algorithmic. For Two-Sets Cut-Uncut, the parameterized-complexity landscape orders graph parameters by the strength of algorithmic consequences. The paper gives a polynomial kernel for feedback-edge-set number, reducible to at most $\pi^\star=\arg\max_{\pi\in X(T)} f(\pi),$5 vertices and $\pi^\star=\arg\max_{\pi\in X(T)} f(\pi),$6 edges when $\pi^\star=\arg\max_{\pi\in X(T)} f(\pi),$7; FPT algorithms for distance to cographs in $\pi^\star=\arg\max_{\pi\in X(T)} f(\pi),$8 time and for treewidth in $\pi^\star=\arg\max_{\pi\in X(T)} f(\pi),$9; XP algorithms such as $X(T)$0 for vertex-cover number; and para-NP-hardness for several weaker parameters (Bentert et al., 2024).

Here the hierarchy is not a sequence of relaxations or cuts on a tree. It is a hierarchy of structural parameters governing when cut-uncut constraints cross from kernelizable to merely FPT, then XP, and finally intractable.

6. Operational and domain-specific cut hierarchies

Several applied systems use hierarchical cuts as operational primitives rather than as approximation hierarchies. In ray tracing, Hierarchy Cut Code encodes a ray by its intersections with a cut $X(T)$1 of a BVH: $X(T)$2 The code aligns sorting keys with the BVH rather than with Euclidean coordinates, and the multi-level compressed variant MLHCC reduces sorting cost. On secondary rays, HCC accelerates tracing by up to $X(T)$3, and replacing 32-bit HCC by 16-bit MLHCC reduces total encoding-plus-sorting overhead from $X(T)$4 ms to $X(T)$5 ms in the Breakfast scene (Xiang et al., 2023).

In mixed-integer linear programming, hierarchical cut selection appears as a decision policy. HEM decomposes the action into a high-level model that selects the cut count and a low-level sequence model that selects an ordered subset of cuts. This jointly addresses which cuts to prefer, how many to choose, and in what order to add them. On the reported benchmarks, HEM reduces solve time by about $X(T)$6–$X(T)$7 versus NoCuts on easy instances, improves over SBP by about $X(T)$8–$X(T)$9, reduces PD-integral on CORLAT by about $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$00 versus NoCuts and about $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$01 versus SBP, and improves PD-integral on MIPLIB mixed neos by about $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$02 versus Default and SBP (Wang et al., 2023).

In low-multiplicity event mixing for $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$03 final states, the IMEHC cut imposes a hierarchy correspondence on either the $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$04 invariant mass ordering or the pion-energy ordering, chosen randomly with $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$05 probability, while MMC and ESO cuts are always enforced. In numerical tests for $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$06, the new IMEHC+MMC+ESO scheme reduced the underestimation of $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$07 from $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$08 to $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$09, while the bias in $$R(t)=\sum_{x_i,x_{i'}\in t}\|x_i-x_{i'}\|^2,\qquad R(T)=\sum_{t\in \tilde T}R(t),$$10 was not significantly improved (He, 2018).

These applications show that the hierarchy idea also functions as a practical control mechanism: it restricts the combinatorial search space by respecting the latent structure of a BVH, a cut pool, or a kinematic configuration.

Cut hierarchy therefore names a family of techniques whose common purpose is to refine cut decisions without flattening the underlying structure. In clustering it replaces a single cut height by subtree-aware pruning; in approximation algorithms it replaces one-shot relaxations by levelwise consistency; in graph algorithms it turns cut structure into a maintained tree; and in systems applications it encodes, orders, or constrains local decisions according to a hierarchy already present in the problem instance.