Good Cuts: Application-Specific Separators

• Good cuts are domain-specific separators and inequalities defined to optimize criteria like tightness, balance, and locality in various applications.
• Researchers develop quantitative measures in surface geometry, graph partitioning, and optimization by linking cut length, residual degree, and LP tightening to problem-specific objectives.
• Practical implementations show that refined cut selection in quantum circuits and image segmentation reduces computational burden while preserving essential structures.

“Good cuts” is a domain-dependent technical expression for separators, cycles, disjunctions, and valid inequalities that are judged not merely by existence, but by an application-specific notion of usefulness. In the literature considered here, the term denotes short closed cycles that isolate neck-like surface features, graph separators that optimize cut size, balance, locality, or residual degree, cutting planes that tighten LP relaxations or rule out logically impossible assignments, and even cut locations in quantum circuits or multigranular image partitions that preserve the relevant structure while reducing computational burden (Ruggerio et al., 13 Jan 2026, Hamann et al., 2015, Engelhardt et al., 25 Sep 2025, Chen et al., 2023).

1. Domain-dependent meanings and recurring criteria

The expression is used across several technically distinct settings, but the underlying pattern is consistent: a cut is “good” when it excludes undesirable configurations without destroying the structure that the downstream task needs.

Domain	Formal object	Quality criterion
Surface geometry	Simple closed cycle	Large tightness $\min\{A_1,A_2\}/L(C)^2$
Graphs and networks	Edge cut, separator, or partition	Small cut, acceptable balance, high locality, or residual degree guarantees
MILP and MaxSAT	Valid inequality or Clause Cut	Strong LP tightening with manageable numerical cost
Quantum and imaging	Circuit cut location or segmentation partition	Exact pruning of zero-information terms or persistent multiscale structure

In this usage, “good” is therefore technical rather than rhetorical. A cut can be good because it is short relative to separated area, Pareto-optimal in cut size and imbalance, buffered to recover stream locality, logically implied by a CNF, or topologically persistent under diffusion condensation. The quality criterion changes with the ambient model, but each formulation tries to avoid trivial separations: tiny caps on surfaces, fuzzy graph boundaries, weak LP cuts, or basis elements in a circuit cut that pass no information.

2. Surface bottlenecks and topological extremality

On genus-zero triangulated surfaces, a good cut is a short simple closed cycle that separates two relatively large patches. “In the Search for Good Neck Cuts” formalizes this via the tightness objective

$T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$

where $A_1$ and $A_2$ are the areas of the two patches induced by the cycle $C$, and $L(C)$ is its arc length. An $\alpha$-bottleneck satisfies $T(C)\ge \alpha$, and an optimal collar is any simple closed curve maximizing $T(C)$. The formulation is isoperimetric: on the unit sphere the equator attains $T=1/(2\pi)$, which the paper uses as a practical lower threshold for neck-like cuts. The theory further introduces $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$0-expanding surfaces, $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$1-well-behaved regions, salient points, and lasso cycles, and proves that within an $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$2-well-behaved neck an optimal collar can be approximated by a lasso with additive loss at most $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$3. Under the stated neck assumptions this yields an $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$4 polynomial-time approximation, while the practical pipeline computes salient points by a two-sweep diameter heuristic, builds a shortest-path skeleton, generates lasso cycles, evaluates tightness, filters by $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$5, and returns local maxima of $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$6 along each path (Ruggerio et al., 13 Jan 2026).

A different topological use of “good cuts” appears in good fractal necklaces. There the relevant objects are extremal $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$7-cuts. For a connected space $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$8, a $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$9-cut $A_1$0 is extremal when it maximizes

$A_1$1

and the paper proves that for a necklace without cut points one has $A_1$2. In a good necklace, the only extremal $A_1$3-cuts are the adjacent pairs of main nodes $A_1$4, and the corresponding extremal component is $A_1$5. These extremal cuts become rigid topological invariants: they determine the necklace iterated function system up to the dihedral group, imply that every homeomorphism between good necklaces is rigid, and force the self-homeomorphism group to be countable (Wen, 2021).

3. Graph partitioning, separators, and connectivity

In graph partitioning, good cuts are usually balanced separators or low-edge-cut partitions. FlowCutter formulates this as a bicriteria optimization problem over cut size and imbalance. For a bipartition $A_1$6 of an undirected graph, the edge-cut size is $A_1$7, imbalance is defined by

$A_1$8

and expansion is

$A_1$9

The algorithm computes a Pareto set of connected balanced $A_2$0-edge-cuts by alternating between max-flow augmentation and piercing operations, with time complexity $A_2$1 where $A_2$2 is the number of edges and $A_2$3 is the largest cut output. Multiple random $A_2$4 pairs make the method effectively $A_2$5-independent, and the resulting separators were shown to improve contraction orders and Customizable Contraction Hierarchies on road graphs (Hamann et al., 2015).

BuffCut addresses good cuts in streaming graph partitioning, where the objective is to minimize weighted edge cut subject to balance constraints

$A_2$6

Its central observation is that one-pass assignment is highly sensitive to stream order. BuffCut therefore delays assignments in a bounded priority buffer, ranks deferred vertices by the Hub-Aware Assigned Neighbors Ratio,

$A_2$7

forms high-locality batches, and applies a multilevel partitioner to each batch. Locality is quantified by the neighbor-to-neighbor average ID distance (AID) and the within-batch internal edge ratio (IER). Empirically, the method achieves $A_2$8 fewer edge cuts than the strongest prioritized buffering baseline while running $A_2$9 times faster and using $C$0 times less memory, and it achieves $C$1 fewer cuts than HeiStream with only modest runtime and memory overheads (Baumgärtner et al., 18 Feb 2026).

The graph-theoretic notion of good cuts can also encode residual robustness. A set $C$2 is a $C$3-good $C$4-component cut if $C$5 has at least $C$6 components and every remaining vertex has at least $C$7 neighbors in $C$8. The corresponding parameter

$C$9

refines classical connectivity by imposing both disconnection and local minimum-degree conditions after deletion. The paper on extremal spectral radius resolves the associated Brualdi–Solheid problem under fixed minimum degree $L(C)$0, showing that the adjacency-spectral-radius maximizers are unique and fall into six explicit regimes determined by the relations among $L(C)$1, $L(C)$2, and $L(C)$3; the extremal constructions are joins of cliques arranged so that the residual components meet the $L(C)$4-good requirement (Ding et al., 2024).

A related parameterized-algorithmic usage appears in important cuts. There, an $L(C)$5–$L(C)$6 separator $L(C)$7 dominates $L(C)$8 when $L(C)$9 and the $\alpha$0-reachable set after deleting $\alpha$1 contains that after deleting $\alpha$2. Important cuts are minimal undominated separators, and the standard enumeration bound cited in the routing context is $\alpha$3 with enumeration in $\alpha$4 time for undirected graphs. This compresses the separator search space for routing problems such as Minimum Shared Edges and Minimum Vulnerability (Sheng et al., 2022).

4. Cutting planes, no-good cuts, and polyhedral strength

In mixed-integer optimization, good cuts are valid inequalities that materially strengthen the LP relaxation without imposing excessive numerical or computational cost. In MaxSAT-oriented MILP formulations, Clause Cuts provide a logically strengthened version of no-good cuts. If a clause $\alpha$5 is implied by a CNF formula, then the associated Clause Cut is

$\alpha$6

with $\alpha$7 and $\alpha$8. Unlike a classical assignment no-good cut, which excludes one infeasible binary assignment, a Clause Cut excludes the entire set of assignments falsifying $\alpha$9. The paper introduces two separation procedures: the Integral Clause Cuts Algorithm (ICCA), which uses integral LP variables as assumptions in a SAT solver, and the Learned Clause Cuts Algorithm (LCCA), which also exploits clauses learned by CDCL. On SATLIB benchmarks, LCCA requires only $T(C)\ge \alpha$0 of Gurobi 12’s runtime on the full problem set and about $T(C)\ge \alpha$1 of RC2’s runtime, with many instances solved at the root node (Engelhardt et al., 25 Sep 2025).

A broader disjunctive-cut framework is given by $T(C)\ge \alpha$2-polyhedral disjunctive cuts. For a point-ray set $T(C)\ge \alpha$3, a cut $T(C)\ge \alpha$4 is valid for $T(C)\ge \alpha$5 exactly when $T(C)\ge \alpha$6 for all $T(C)\ge \alpha$7 and $T(C)\ge \alpha$8 for all $T(C)\ge \alpha$9. The implementation uses a compact point-ray LP with $T(C)$0 in nonbasic coordinates,

$T(C)$1

which avoids the high-dimensional size of the classical cut-generating LP. The method extracts multiterm disjunctions from the leaves of a partial branch-and-bound tree and produces strong one-round cuts without recursive application. In the reported experiments, GMICs alone close $T(C)$2 of the root gap on average, whereas VPCs combined with GMICs close $T(C)$3 (2207.13619).

The cut-generating-function literature studies good cuts as a balance between approximation quality and efficient evaluation. A generalized cross-polyhedron $T(C)$4 with $T(C)$5 induces a gauge

$T(C)$6

where $T(C)$7, and a trivial lifting

$T(C)$8

The resulting closure from generalized cross-polyhedra approximates the closure from all maximal $T(C)$9-free sets within a factor

$T=1/(2\pi)$0

and the paper gives explicit procedures with $T=1/(2\pi)$1 gauge evaluation and $T=1/(2\pi)$2 trivial-lifting computation. Empirically, these cuts yield some tangible improvement over GMI cuts on random dense instances, but perform poorly on MIPLIB 3.0 and on random stable set and vertex cover instances (Basu et al., 2018).

Two-row cuts from degenerate tableaux illustrate that cut quality is not reducible to depth alone. In the two-row model

$T=1/(2\pi)$3

the paper constructs Type 1 and Type 2 triangle cuts, split cuts, integer liftings, and wedge strengthenings. The deepest two-row triangle cuts are often deeper than the corresponding GMI cuts from the same row, but large-scale experiments show only marginal improvements over a reliable GMI baseline, while the apparent competitiveness of two-row cuts changes substantially with the aggressiveness and reliability of the underlying cut generator (Basu et al., 2017).

5. Learned and hybrid policies for selecting cuts in MILP solvers

A separate strand of work studies good cuts as a policy-selection problem. “Learning to Use Local Cuts” considers whether cuts should be added only globally at the root or also locally at interior nodes. On the reported dataset, local cuts are $T=1/(2\pi)$4 faster on average than never using local cuts, but the no-local-cuts policy is significantly faster on $T=1/(2\pi)$5 of instances. A perfect instance-level oracle would therefore give an $T=1/(2\pi)$6 overall speed-up. The paper trains a regression forest on 32 static and dynamic root-stage features, converts the prediction to an LC/NLC decision, and deploys a conservative 70% tree-majority safeguard. In industrial Xpress data, the learned rule yields an overall $T=1/(2\pi)$7 speed-up, increasing to about $T=1/(2\pi)$8 on instances requiring at least $T=1/(2\pi)$9 seconds (Berthold et al., 2022).

“Adaptive Cut Selection in Mixed-Integer Linear Programming” studies the scoring-rule parameters themselves. It analyzes weighted combinations of directed cutoff distance, efficacy, integer support, and objective parallelism,

$T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$00

and proves a negative result: for a specific infinite parametric MILP family, any finite grid search over the parameter space misses all parameter values that select integer-optimal-inducing cuts for infinitely many instances. The paper then uses a graph-convolutional model and reinforcement learning to learn instance-specific weights, showing that adaptive cut selection improves performance over diverse benchmark classes even though a single universal rule is difficult to identify (Turner et al., 2022).

Sequence models refine the notion of good cuts further by treating selection as a joint decision over which cuts to keep, how many to keep, and in what order. The original HEM formulates a high-level policy over the selection ratio $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$01 and a lower-level pointer network over an ordered subset of candidate cuts, thereby addressing the three decisions $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$02 which cuts, $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$03 how many, and $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$04 in what order (Wang et al., 2023). The later hierarchical sequence/set model generalizes this into HEM and HEM++, using a higher-level tanh-Gaussian policy for the selection ratio and a lower-level Seq2Seq or Set2Seq selector trained by policy gradient or hierarchical PPO. The paper reports substantial improvements across eleven MILP benchmarks and shows by ablation that learning cardinality and order is materially important (Wang et al., 2024).

HGTSM argues that cut-local features alone are insufficient because they ignore the LP relaxation. It represents the state as a heterogeneous tripartite graph over variables, constraints, and cuts, encodes it with a heterogeneous graph transformer, and then applies an order-robust hierarchical sequence model without positional encodings. The resulting policy outperforms heuristic methods and learning-based baselines on several medium and hard MILP datasets, while dramatically reducing sensitivity to candidate-list permutations relative to HEM (Zhang et al., 2024).

A simpler ranking-based alternative is Cut Ranking, which uses multiple instance learning: bags of selected cuts are labeled by downstream runtime reduction, a small MLP learns a score function on 14 static and dynamic cut features, and the solver then keeps the top-$T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$05 cuts. The reported industrial A/B test on large-scale product-planning problems yields an average speedup ratio of $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$06 with no loss of solution accuracy (Huang et al., 2021).

The cut/branch interface provides a final variation. “Branching via Cutting Plane Selection” observes that the elementary split associated with branching on a fractional basic integer variable is represented more faithfully by the weak-GMI cut than by the strengthened GMI. Ranking branching candidates by the efficacy of their associated weak-GMI therefore produces smaller trees than using strengthened GMI efficacy. More importantly for practice, storing the normalized efficacy of the most recent solver-generated GMI for each variable and adding a small weighted term to SCIP’s hybrid branching score yields about a $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$07 reduction in solve time and an $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$08 reduction in node count on affected MIPLIB 2017 instances (Turner et al., 2023).

6. Quantum circuit cutting, topological segmentation, and broader patterns

In quantum circuit cutting, a good cut is a “golden cutting point”: a wire cut for which some basis element provably contributes zero information to the target observable. For a single cut, if there exists a basis element $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$09 satisfying

$T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$10

then all terms involving $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$11 cancel in the reconstruction formula and can be dropped exactly. For a bipartition with $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$12 cuts and $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$13 golden cutting points, the number of classical reconstruction terms drops from $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$14 to $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$15, and the number of circuit evaluations from $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$16 to $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$17. On IBM hardware, pruning one basis element reduces average wall time from $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$18 seconds to $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$19 seconds, a $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$20 reduction, without loss of accuracy (Chen et al., 2023).

In unsupervised medical-image segmentation, CUTS uses the language of cuts more metaphorically but still in a structurally precise sense. The method first learns image-specific embeddings from pixel-centered patches by combining intra-image contrastive learning and local patch reconstruction, then partitions the embeddings by diffusion condensation into a hierarchy of metastable granularities. Persistent scales correspond to anatomically meaningful segmentations. On retinal fundus images and two brain-MRI tasks, CUTS improves Dice coefficient and Hausdorff distance by at least $T(C)=\frac{\min\{A_1,A_2\}}{L(C)^2},$21 over existing unsupervised methods and achieves performance on par with, or better than, several Segment Anything variants on parts of the evaluation suite (Liu et al., 2022).

Taken together, these works suggest that a good cut is rarely the shortest, strongest, or most aggressive separator in isolation. Rather, it is the separator, cycle, partition, or inequality that optimizes a domain-specific surrogate for usefulness while explicitly suppressing degenerate behavior: tiny surface caps, disconnected spectral partitions, weak or redundant LP cuts, basis elements that carry no information, or segmentation granularities that are not topologically persistent. That recurring design principle explains why recent work so often couples the definition of a cut to a second criterion—tightness, balance, locality, persistence, efficacy, or robustness—instead of treating separability alone as the objective.