Non-stop Cut-Matching Game

• Non-stop Cut-Matching Game is a framework that continuously maintains progress across rounds using structured cuts and robust matchings to build constant-hop expanders.
• It employs a generalized cut strategy that replaces a single bisection with multiple structured cuts per iteration to overcome classical hop-length limitations.
• The approach enables efficient routing and expander decompositions with controlled congestion and constant-hop guarantees, benefiting dynamic graph algorithms.

Searching arXiv for recent and foundational papers on cut-matching games and non-stop variants. A non-stop cut-matching game is a cut-matching framework in which progress is maintained continuously across rounds, rather than relying on a single terminal expander-construction event or on a rigid stop condition after every adverse response. In its most technically developed form in the constant-hop setting, it is a generalization of the classical Khandekar–Rao–Vazirani and KKOV frameworks that replaces one bisection per round by many structured cuts per iteration, and produces a graph that is a $(t,2)$-hop $\phi$-expander for unit-demands with diameter at most $t$ (Haeupler et al., 2022). In adjacent lines of work, the same phrase or a closely related interpretation refers to trimming-based games that continue after unbalanced sparse cuts, directed games that proceed using only single-commodity max-flow computations, or continual distance-based analogues that keep accreting connectivity certificates until a well-connectedness condition is met (Agassy et al., 2022, Louis, 2010, Chuzhoy, 2022). This suggests that “non-stop” is best understood as a design principle for cut-matching machinery rather than as a single standardized formal object.

1. Classical framework and the source of the non-stop generalization

The classical cut-matching game is played on an initially empty multigraph $G$ on a vertex set $V$. Each round has three steps: the cut player chooses a bisection $(S, V \setminus S)$, the matching player returns a perfect matching across the cut, and the matching is added to $G$. With a good cut strategy, after $r$ rounds the accumulated graph is an expander with conductance at least $\Omega(1/\log n)$, and KKOV shows existence of a strategy achieving an $O(1/\log n)$-expander in $\phi$0 rounds (Haeupler et al., 2022).

For a $\phi$1-regular graph, the normalized edge expansion is

$\phi$2

This framework underlies near-linear-time approximation algorithms for sparsest cut and related problems through reductions to single-commodity flows (Haeupler et al., 2022).

Its central limitation is hop length. In the classical game, expansion suffices for congestion-competitive routing, but the hop-length is $\phi$3, and hence $\phi$4 in any expander. Even $\phi$5-conductance expanders require $\phi$6 hop-length to mix traffic. The constant-hop program starts precisely from this obstruction: classical expander-based algorithms incur unavoidable polylogarithmic losses in path length and associated $\phi$7 quantities such as hop count and costs (Haeupler et al., 2022).

The phrase “non-stop” emerged in several extensions of this classical picture. In expander decomposition, a non-stop variation means that the game does not stop upon encountering an unbalanced sparse cut, but continues to play on the trimmed part of the large side (Agassy et al., 2022). In the directed setting, “non-stop” refers to advancing or certifying a sparse cut using only single-commodity max-flow computations throughout, with no intermediate multicommodity-flow step (Louis, 2010). These usages share a common structural idea: the game continues to produce valid progress certificates under weakened or asymmetric feedback.

2. Constant-hop expanders as the target object

The constant-hop formulation in “A Cut-Matching Game for Constant-Hop Expanders” strengthens ordinary expansion by demanding short routing templates for all unit-demands (Haeupler et al., 2022). The underlying notion is the $\phi$8-hop $\phi$9-expander. The paper states a flow characterization: if $t$0 is an $t$1-hop $t$2-expander for unit-demands, then every $t$3-hop unit-demand can be routed in $t$4 along $t$5-hop paths with congestion $t$6; conversely, if $t$7 is not an $t$8-hop $t$9-expander, then some $G$0-hop unit-demand cannot be routed in $G$1 along $G$2-hop paths with congestion $G$3 (Haeupler et al., 2022).

In the constant-hop specialization, the graph has constant hop-diameter $G$4 and ensures that any global unit-demand can be routed on paths of length at most $G$5 with congestion bounded independently of the hop length, though generally polynomial in $G$6 for constant $G$7. Formally, the output graph is a $G$8-hop $G$9-expander for unit-demands with diameter at most $V$0. All flow-paths have length at most $V$1, and the edge congestion is at most $V$2, where typically, with $V$3, one can achieve

$V$4

Setting $V$5 gives a hop-constrained conductance proxy for routing guarantees (Haeupler et al., 2022).

The same paper treats hop-constrained expander decomposition as both a subroutine and a guarantee. An $V$6-hop $V$7-expander decomposition for unit-demands is a pure cut $V$8 of size at most $V$9 such that $(S, V \setminus S)$0 is an $(S, V \setminus S)$1-hop $(S, V \setminus S)$2-expander for unit-demands. The existence theorem states that for $(S, V \setminus S)$3, $(S, V \setminus S)$4, and $(S, V \setminus S)$5, such a decomposition exists with congestion slack $(S, V \setminus S)$6 (Haeupler et al., 2022).

The significance of this shift is algorithmic rather than merely structural. The constant-hop formulation is designed for settings in which approximation quality depends directly on path length. The paper explicitly links the resulting expanders to constant-approximate $(S, V \setminus S)$7-commodity flows in $(S, V \setminus S)$8 time and to the optimal constant-approximate deterministic worst-case fully-dynamic APSP-distance oracle, with the constant-approximation factor directly tracing to constant-hop routing paths (Haeupler et al., 2022).

3. Generalized cut strategy: many structured cuts per iteration

The generalized cut strategy departs from the single-bisection format of KRV and KKOV. It produces constant-hop expanders deterministically, against adversarial matchings, within $(S, V \setminus S)$9 iterations, by playing many structured cuts per iteration (Haeupler et al., 2022).

The algorithm has a main phase and a final phase. In the main phase, the game starts with $G$0 on $G$1. At each iteration, it computes an $G$2-hop $G$3-expander decomposition of the current graph with congestion slack $G$4, obtaining a cut $G$5 so that $G$6 is an $G$7-hop $G$8-expander. It then computes a neighborhood cover of $G$9 with small diameter $r$0 and large separation $r$1, where $r$2 is the number of main-phase iterations. If a large cluster $r$3 with $r$4 is found, the main phase stops. Otherwise, the cover is decomposed into $r$5 groups, each containing $r$6 equal-sized, pairwise $r$7-separated blocks, and the cut player presents all pairwise cuts within each group. The matching player returns the union $r$8 of perfect matchings across these cuts, and $r$9 is added to $\Omega(1/\log n)$0 (Haeupler et al., 2022).

The final phase is a completion step. Once a large cluster $\Omega(1/\log n)$1 exists, the algorithm partitions $\Omega(1/\log n)$2 into sets $\Omega(1/\log n)$3 with $\Omega(1/\log n)$4, requests matchings between $\Omega(1/\log n)$5 and each $\Omega(1/\log n)$6, and adds them to $\Omega(1/\log n)$7. The routing interpretation is explicit: any demand $\Omega(1/\log n)$8 enters $\Omega(1/\log n)$9 by a final matching edge, moves inside $O(1/\log n)$0, and exits by another final matching edge (Haeupler et al., 2022).

Neighborhood covers are the main combinatorial device that makes this possible. The paper uses deterministic constructions with covering radius, separation, width, and load parameters chosen so that the clusters have small diameter and large separation, while width and load remain bounded by

$O(1/\log n)$1

The cover decomposition drops only a small fraction of vertices; with suitable choices of $O(1/\log n)$2 and $O(1/\log n)$3, at most $O(1/\log n)$4 of vertices are dropped, and $O(1/\log n)$5 (Haeupler et al., 2022).

This design changes what a “round” means. In the classical game, one round contributes one perfect matching across one bisection. Here, one main-phase iteration contributes an entire structured batch of pairwise cuts and their corresponding matchings. That batch structure is what allows the later entropy analysis to exploit locality inside separated blocks while still forcing large-scale mixing.

4. Two-step mixing, entropy potential, and non-stop robustness

The analytical core of the constant-hop game is a two-step mixing process and an entropy potential on pseudo-distributions (Haeupler et al., 2022). Each main-phase iteration applies a lazy random walk on the union of matchings. First, if $O(1/\log n)$6 is the number of matchings involving $O(1/\log n)$7, the commodity mass at $O(1/\log n)$8 is split evenly among the matchings touching $O(1/\log n)$9. Second, within each matching, mass is sent evenly among neighbors and then mixed back to vertices proportional to weights

$\phi$00

where $\phi$01 is the block size per group (Haeupler et al., 2022).

The paper proves stable entropy: entropy is non-decreasing under this two-step process. It then decomposes each commodity flow as $\phi$02, where leaked flow accounts for three failure modes: flow traversing edges in the decomposition cut $\phi$03, flow sitting on vertices with $\phi$04, and flow traversing removed matching edges when incomplete matchings are allowed. The typical part $\phi$05 remains local because large separation and bounded hop-length force it to stay concentrated in a single block per matching (Haeupler et al., 2022).

For an $\phi$06-typical commodity, meaning $\phi$07, the entropy increase per iteration is

$\phi$08

For the joint distribution $\phi$09 of all commodities, the paper obtains

$\phi$10

with a constant fraction $\phi$11 of commodities remaining $\phi$12-typical with $\phi$13 before each iteration, provided $\phi$14 and the cover parameters are chosen to control leakage (Haeupler et al., 2022).

Since always $\phi$15, choosing

$\phi$16

implies that the main phase requires only $\phi$17 iterations (Haeupler et al., 2022).

This is also the section in which the “non-stop” aspect becomes mathematically explicit. The analysis is robust to incomplete matchings and large matchings. The potential argument still works if the matching player returns large matchings instead of perfect ones, and it explicitly allows the matching player to remove an arbitrary $\phi$18-fraction of edges in each batch. The leaked-flow accounting shows that entropy progress still holds after parameter adjustment (Haeupler et al., 2022). A plausible implication is that the framework is not merely a batch construction for a fixed terminal object; it is a continuously progressing process whose correctness degrades gracefully under partial or online matching responses.

5. Main theorem, routing decomposition, and complexity

The central theorem fixes $\phi$19 and assumes access to an algorithm that computes an $\phi$20-hop $\phi$21-expander decomposition with congestion slack $\phi$22. Against any matching player, the cut strategy returns a graph $\phi$23 that is a $\phi$24-hop $\phi$25-expander for unit-demands with diameter at most $\phi$26, where $\phi$27, the number of main-phase iterations is $\phi$28, and the number of cuts presented to the matching player satisfies $\phi$29 (Haeupler et al., 2022).

The congestion parameter is specified by setting $\phi$30, where

$\phi$31

so in particular $\phi$32. The maximum degree obeys

$\phi$33

more precisely,

$\phi$34

By the flow characterization, any unit-demand is routed along $\phi$35-hop paths with edge congestion at most $\phi$36 (Haeupler et al., 2022).

The routing decomposition is explicit. If both terminals lie in the large cluster $\phi$37, routing stays entirely within $\phi$38 in at most $\phi$39 hops because $\phi$40 is an $\phi$41-hop $\phi$42-expander and $\phi$43. If neither terminal lies in $\phi$44, the route uses one final matching edge into $\phi$45, an internal route within $\phi$46, and one final matching edge out, for total hop length at most $\phi$47. If one terminal already lies in $\phi$48, the hop length is at most $\phi$49 (Haeupler et al., 2022).

The complexity guarantees are tied to the availability of the decomposition subroutine. Neighborhood covers are computed deterministically with near-linear work in CONGEST or PRAM, the cover decomposition per iteration runs in $\phi$50 time, and the overall number of cuts is $\phi$51. The cut strategy is efficient if given an efficient expander decomposition; the cyclic dependence can be resolved via recursion, passing subproblems of size $\phi$52 for at most $\phi$53 calls and constant recursion depth when $\phi$54 (Haeupler et al., 2022).

The paper’s high-level proof sketch summarizes the logic. The main-phase random walk increases entropy by at least a constant per commodity per iteration, once typical commodities are identified, until a large low-diameter cluster $\phi$55 emerges. The final-phase matchings then finish global routing: enter $\phi$56, route within $\phi$57 in at most $\phi$58 hops, and exit. Congestion inside $\phi$59 grows by at most $\phi$60, and overall edge congestion becomes $\phi$61 (Haeupler et al., 2022).

6. Related non-stop formulations, applications, and scope

The constant-hop game is one member of a broader family of non-stop or continual cut-matching constructions. In “Expander Decomposition with Fewer Inter-Cluster Edges Using a Spectral Cut Player,” the non-stop variant is explicit: the game maintains a partition $\phi$62, does not stop when it encounters an unbalanced sparse cut, and instead trims the large side and continues on the shrinking domain $\phi$63. The spectral potential is adapted to this shrinking domain through

$\phi$64

and the resulting expander decomposition improves the inter-cluster edge bound from $\phi$65 to $\phi$66 (Agassy et al., 2022).

In directed graphs, the non-stop interpretation emphasizes oracle structure rather than trimming. “Cut-Matching Games on Directed Graphs” alternates cut-player and matching-player moves, where the matching player uses single-commodity max-flow to either produce embeddable directed matchings or return a sparse cut, and after $\phi$67 successful rounds the union of matchings becomes a directed $\phi$68-expander. The paper states that there is no intermediate multicommodity-flow step; the algorithm advances or certifies a sparse cut using only single-commodity max-flows throughout (Louis, 2010).

A nearby but distinct line replaces cuts by distances. “A Distanced Matching Game, Decremental APSP in Expanders, and Faster Deterministic Algorithms for Graph Cut Problems” does not explicitly define a “Non-stop Cut-Matching Game,” but its distanced matching game can be interpreted as a continual process: as long as a $\phi$69-distancing exists, the algorithm adds a large matching; when no distancing remains, a hierarchical support structure certifies well-connectedness on a large supported set with short paths (Chuzhoy, 2022). The same paper notes that an independent work proposes a cut-matching-game variant for $\phi$70-hop expanders with small diameter, aligning this distance-based perspective with the constant-hop expander program (Chuzhoy, 2022).

Hypergraph generalizations extend the same design pattern beyond graph conductance. “Cut-matching Games for Generalized Hypergraph Ratio Cuts” is round-based rather than explicitly non-stop, but it supports a natural “non-stop until threshold” interpretation: keep embedding $\phi$71-regular bipartite graphs via repeated $\phi$72–$\phi$73 max-flow computations in an augmented cut-preserver until a mixing or certificate threshold is reached (Veldt, 2023). “Submodular Hypergraph Partitioning: Metric Relaxations and Fast Algorithms via an Improved Cut-Matching Game” relaxes both players’ actions, using Matrix Multiplicative Weights for the cut player and approximate dual graph certificates for the matching player; the paper explicitly describes this as a continuous, streaming-style iteration without pauses for global multicommodity flows or exact matchings (Chen et al., 2023). More recently, an adaptation of cut-matching to the bipartiteness ratio uses a skew-symmetric auxiliary graph, a density-matrix state, and one single-commodity undirected max-flow per round, and describes the procedure as non-stop because the rounds proceed consecutively without embedding interruptions or reinitializations of the potential (Soma et al., 17 Jul 2025).

Across these settings, the persistent theme is that cut-matching is no longer viewed solely as a way to prove the existence of an expander after a bounded number of idealized rounds. It becomes a robust algorithmic scaffold for routing, certification, and decomposition under partial matchings, adversarial responses, asymmetric objectives, or shrinking domains. In the constant-hop setting, this scaffold produces structures with hop-diameter $\phi$74 and routing templates of length at most $\phi$75 with controlled congestion, thereby extending expander-based machinery to objectives that are sensitive to path length and cost (Haeupler et al., 2022).