Cylinder Regularity Lemma

• Cylinder Regularity Lemma is a decomposition principle that partitions product spaces into vertex cylinders, where most cells exhibit quasirandom behavior.
• It employs an energy-increment strategy to refine partitions, resulting in exponential-type bounds for graphs and tower-type bounds for 3-uniform hypergraphs.
• The lemma acts as an intermediate tool that reduces the complexity of full hypergraph regularity, enhancing induced counting methods and bounded VC/VC₂ results.

Searching arXiv for recent and foundational papers on cylinder regularity and closely related regularity lemmas. The Cylinder Regularity Lemma is a regularity principle for product partitions of multipartite discrete structures, introduced as a central intermediate device in the study of regularity for hypergraphs with bounded VC$_2$ dimension. In the graph setting, it provides a partition of a product space into vertex cylinders such that most cylinders are quasirandom, with substantially better quantitative behavior than the ordinary graph regularity lemma. In the $3$-uniform hypergraph setting, it takes the form of a cylinder-chain partition controlling hypergraph quasirandomness relative to lower-order graph structure, and it is used to reduce the quantitative cost of regularization by one level in the Ackermann hierarchy (Gishboliner et al., 13 Aug 2025). The notion is naturally compared with earlier bipartite product-partition regularity for graphs of bounded VC dimension (Towsner, 2013), with stable and distal definable regularity decompositions into product cells (Malliaris et al., 2015, Simon, 2015), and with the tower-type limitations of ordinary Szemerédi-style regularity (Fox et al., 2014).

1. Product partitions and the meaning of “cylinder”

In this context, a cylinder is a product structure on a multipartite vertex set. In the graph setting, for a bipartite graph with parts $A,B$, a vertex cylinder is a product set

$A' \times B' \subseteq A \times B,$

with $A' \subseteq A$, $B' \subseteq B$. For $t$-partite graphs, the natural generalization is a product of subsets of the $t$ vertex classes (Gishboliner et al., 13 Aug 2025).

The essential feature is that a cylinder partition is a partition of the product space $A\times B$ or $X_1\times\cdots\times X_t$, rather than a partition of the vertex set itself (Gishboliner et al., 13 Aug 2025). This separates it from the usual vertex-partition formulation of Szemerédi regularity. A plausible implication is that the cylinder framework is quantitatively easier to control because it works directly with coordinate-wise product cells.

This product-structure viewpoint already appears in bounded-VC graph regularity. For a bipartite relation

$3$0

Towsner studies partitions

$3$1

so that regularity is tested on rectangles $3$2 (Towsner, 2013). That result is not a higher-dimensional Cylinder Regularity Lemma, but it is explicitly described as a regularity lemma with product partitions, and it is comparable because the partition is a grid decomposition of $3$3 (Towsner, 2013).

The same product-cell perspective also appears in model-theoretic regularity. In distal structures, one obtains definable partitions $3$4 and $3$5 so that the total product measure of non-homogeneous cells $3$6 is small (Simon, 2015). In stable definable bipartite graphs, the partition into definable pieces yields rectangles $3$7 that are almost complete or almost empty in a measure-theoretic sense (Malliaris et al., 2015). These are not called cylinder regularity lemmas in those papers, but they realize the same structural theme: regularity via decomposition into product cells.

2. Graph cylinder regularity

For a $3$8-partite graph with parts $3$9, a vertex cylinder is a product

$A,B$0

A vertex cylinder partition is a partition of the product space $A,B$1 into such cylinders (Gishboliner et al., 13 Aug 2025).

A cylinder $A,B$2 is called $A,B$3-quasirandom if, in the graph structure under consideration, the relevant bipartite graph(s) on its coordinates are $A,B$4-quasirandom in the usual graph sense (Gishboliner et al., 13 Aug 2025). The graph-cylinder statement used in the bounded-VC$A,B$5 program is a generalized multi-graph version of the Duke–Lefmann–Rödl cylinder regularity lemma. It is stated as follows:

$A,B$6

Let $A,B$7 be $A,B$8-partite graphs on the same vertex set. For any $A,B$9, there exists a vertex cylinder partition $A' \times B' \subseteq A \times B,$0 such that if $A' \times B' \subseteq A \times B,$1 is chosen uniformly at random from $A' \times B' \subseteq A \times B,$2, then with probability at least $A' \times B' \subseteq A \times B,$3, the induced subgraph of every $A' \times B' \subseteq A \times B,$4 on the cylinder $A' \times B' \subseteq A \times B,$5 is $A' \times B' \subseteq A \times B,$6-quasirandom. Moreover,

$A' \times B' \subseteq A \times B,$7

This is the version actually used in the proof of the hypergraph result, because the edge partition must simultaneously restore quasirandomness for a family of graphs, not just one graph (Gishboliner et al., 13 Aug 2025).

The quantitative contrast with ordinary graph regularity is explicit. In the graph case, the cylinder regularity lemma has only exponential-type bounds, whereas the full regularity lemma has tower-type bounds (Gishboliner et al., 13 Aug 2025). This should be read against the sharp lower-bound theory for the ordinary regularity lemma: Fox and Lovász show that the number of parts required in a version of Szemerédi’s regularity lemma can be at least a tower of twos of height $A' \times B' \subseteq A \times B,$8, matching the upper bound in tower height order (Fox et al., 2014). In that sense, graph cylinder regularity is designed as a quantitatively weaker but more manageable intermediate object.

The graph case also relates to Towsner’s bounded-VC regularity lemma. There, one still seeks an ordinary $A' \times B' \subseteq A \times B,$9-regular partition, but the bounded VC dimension assumption allows controlled refinement size via $A' \subseteq A$0-nets for differences and the Shelah–Sauer bound (Towsner, 2013). The paper proves that if $A' \subseteq A$1 has VC dimension at most $A' \subseteq A$2, then there is a $A' \subseteq A$3-regular partition $A' \subseteq A$4 with

$A' \subseteq A$5

The partition is again a product partition, and the argument uses small sample sets and difference nets to control the number of product cells (Towsner, 2013). This is not the cylinder lemma of (Gishboliner et al., 13 Aug 2025), but it is a direct antecedent in the broader effort to replace tower-type growth by smaller bounds under VC restrictions.

3. Hypergraph cylinder regularity

The core new contribution in the bounded VC$A' \subseteq A$6 setting is a $A' \subseteq A$7-uniform hypergraph analogue of cylinder regularity (Gishboliner et al., 13 Aug 2025). For a $A' \subseteq A$8-partite $A' \subseteq A$9-graph on parts $B' \subseteq B$0, the paper defines:

• a vertex subcylinder

$B' \subseteq B$1

• a vertex cylinder partition $B' \subseteq B$2, a partition of $B' \subseteq B$3 into such subcylinders;
• an edge partition $B' \subseteq B$4 of a cylinder $B' \subseteq B$5, which for each pair $B' \subseteq B$6 partitions the complete bipartite graph $B' \subseteq B$7 into subgraphs (Gishboliner et al., 13 Aug 2025).

A cylinder chain partition is then a pair

$B' \subseteq B$8

where $B' \subseteq B$9 is a vertex cylinder partition and $t$0 assigns an edge partition to each cylinder in $t$1 (Gishboliner et al., 13 Aug 2025). For a tuple $t$2, the induced chain is written

$t$3

Quasirandomness is defined relative to this partition. The hypergraph $t$4 is $t$5-quasirandom relative to $t$6 if, for a $t$7 fraction of tuples $t$8, the chain

$t$9

is $t$0-quasirandom in the Gowers chain sense, where $t$1 controls the quasirandomness of the underlying graph part as a function of its density (Gishboliner et al., 13 Aug 2025).

The main hypergraph cylinder regularity theorem states:

Let $t$2, $t$3, and let $t$4 be an increasing polynomial function satisfying $t$5 If $t$6 is a $t$7-partite $t$8-graph, then there exists a cylinder chain partition $t$9 such that $A\times B$0 is $A\times B$1-quasirandom relative to $A\times B$2, and $A\times B$3 $A\times B$4 (Gishboliner et al., 13 Aug 2025)

A slightly stronger form adds a constant $A\times B$5, depending only on $A\times B$6, such that a $A\times B$7 fraction of tuples lie in chains whose product density is at least $A\times B$8 (Gishboliner et al., 13 Aug 2025). This lower bound is needed later in induced counting arguments.

This hypergraph version is the point at which the term “Cylinder Regularity Lemma” acquires its technically distinctive meaning. It is not a full hypergraph regularity lemma. Rather, it regularizes a product-space/cylinder-chain object strongly enough to support induced counting and a later conversion to a genuine regularity partition (Gishboliner et al., 13 Aug 2025).

4. Energy increment and quantitative mechanism

The proof of the hypergraph cylinder regularity lemma uses an energy-increment argument built around a mean-squared density $A\times B$9 (Gishboliner et al., 13 Aug 2025). For a tripartite chain $X_1\times\cdots\times X_t$0 and an edge partition $X_1\times\cdots\times X_t$1,

$X_1\times\cdots\times X_t$2

where the sum is over one part from each bipartite partition and $X_1\times\cdots\times X_t$3 denotes the set of triangles of $X_1\times\cdots\times X_t$4 (Gishboliner et al., 13 Aug 2025).

For a cylinder chain partition $X_1\times\cdots\times X_t$5,

$X_1\times\cdots\times X_t$6

The paper records the key facts: $X_1\times\cdots\times X_t$7 and $X_1\times\cdots\times X_t$8 is monotone under refinement, with these facts proved by Cauchy–Schwarz (Gishboliner et al., 13 Aug 2025).

Whenever some chain is not quasirandom, the partition is refined and $X_1\times\cdots\times X_t$9 increases by a definite amount. Since $3$00 is bounded above, the process terminates (Gishboliner et al., 13 Aug 2025). The quantitative recursion is encoded in functions $3$01 and $3$02, and for polynomial $3$03 the paper shows

$3$04

The resulting tower-type growth comes from the fact that the cylinder lemma incurs only an exponential-type loss at each refinement stage, rather than the much larger losses of full hypergraph regularity (Gishboliner et al., 13 Aug 2025).

This mechanism is directly analogous in spirit to earlier energy-increment proofs of graph regularity. In bounded-VC graph regularity, Towsner defines

$3$05

with $3$06, and proves a refinement lemma giving both controlled partition growth and a definite energy increment

$3$07

After at most $3$08 steps, the procedure terminates (Towsner, 2013). The same general proof philosophy underlies Schrijver’s Euclidean proof of Szemerédi’s regularity lemma, where the potential function is $3$09 and each irregular refinement raises the energy by at least $3$10 (Schrijver, 2012). The Cylinder Regularity Lemma can therefore be placed inside the standard “energy increment under refinement” lineage, but on a different object: cylinder partitions of product spaces.

5. Role in bounded VC$3$11 hypergraph regularity

The bounded VC$3$12 theorem is the main application motivating the Cylinder Regularity Lemma (Gishboliner et al., 13 Aug 2025). The combinatorial input is that bounded VC$3$13 means a $3$14-graph forbids some fixed tripartite $3$15-graph as a tripartitely induced subgraph (Gishboliner et al., 13 Aug 2025). The regularity strategy proceeds in four stages.

First, one applies the cylinder regularity lemma to obtain a cylinder partition in which most cylinders are quasirandom (Gishboliner et al., 13 Aug 2025). Second, bounded VC$3$16 is combined with an induced counting lemma: if a chain is sufficiently quasirandom and its relative density is bounded away from $3$17 and $3$18, then it must contain every fixed forbidden tripartite pattern, contradicting bounded VC$3$19 (Gishboliner et al., 13 Aug 2025). Hence, in a bounded VC$3$20 $3$21-graph, every sufficiently quasirandom chain must have density very close to $3$22 or $3$23.

Third, one passes to the Venn diagram partition of the cylinder partition, thereby converting the cylinder structure into a genuine vertex chain partition (Gishboliner et al., 13 Aug 2025). The paper proves that homogeneity survives this refinement by a Markov-inequality argument: if most cylinders are $3$24-homogeneous, then after refining, most parts are $3$25-homogeneous (Gishboliner et al., 13 Aug 2025).

Fourth, one performs one final Szemerédi-type regularization step to restore quasirandomness of the lower-order graph structure, since the Venn refinement may destroy the quasirandomness built into the cylinder partition (Gishboliner et al., 13 Aug 2025). The relevant simultaneous graph-regularization statement refines a chain partition $3$26 so that a $3$27 fraction of pairs lie in $3$28-quasirandom graphs, with

$3$29

where $3$30 and $3$31 is the maximum number of edge classes in the edge partitions (Gishboliner et al., 13 Aug 2025).

The quantitative payoff is the distinction between three growth regimes emphasized in the paper: tower-type, double-tower-type, and wowzer-type (Gishboliner et al., 13 Aug 2025). General $3$32-graph regularity is wowzer-type, whereas bounded VC$3$33 permits a reduction to double-tower-type bounds for the final regularity partition. The cylinder regularity lemma itself is only tower-type, and this is exactly what lowers the overall complexity by one Ackermann level (Gishboliner et al., 13 Aug 2025). The paper also states that Terry proved a tower-type lower bound even for VC$3$34-dimension $3$35, so tower-type lower bounds are unavoidable in the bounded VC$3$36 setting (Gishboliner et al., 13 Aug 2025).

6. Relation to other regularity frameworks

The Cylinder Regularity Lemma occupies an intermediate position between ordinary graph regularity, full hypergraph regularity, and definable regularity decompositions.

Comparison with ordinary Szemerédi regularity

Ordinary graph regularity partitions the vertex set and measures uniformity on pairs of vertex classes. Its quantitative complexity is tower-type in $3$37, and this dependence is sharp in tower height order (Fox et al., 2014). Schrijver’s proof presents the same phenomenon via balanced partitions, regular pairs, orthogonal projection onto block-constant matrices, and iterative energy increment (Schrijver, 2012).

By contrast, cylinder regularity partitions the product space into cylinders (Gishboliner et al., 13 Aug 2025). This makes it weaker as a decomposition statement, but quantitatively much cheaper. A plausible implication is that cylinder regularity is useful precisely when one needs only enough regularity to run induced counting or extract homogeneous structure, not a full-fledged equitable vertex partition.

Comparison with bounded-VC graph regularity

Towsner’s bounded-VC graph regularity result is fundamentally a bipartite graph regularity lemma, not a cylinder lemma in the higher-order sense (Towsner, 2013). But it is closely related because the partition

$3$38

is a product partition, and bounded VC dimension yields small $3$39-nets, $3$40-nets for differences, and a Shelah–Sauer bound controlling the number of trace patterns (Towsner, 2013). Those ingredients replace tower bounds by a doubly-exponential-type bound

$3$41

(Towsner, 2013). The later cylinder-based method in (Gishboliner et al., 13 Aug 2025) may be viewed as a further development of the same quantitative agenda: regularize via product structure while exploiting a dimension restriction.

Comparison with stable and distal definable regularity

In stable definable bipartite graphs, one partitions $3$42 and $3$43 into finitely many definable pieces so that every rectangle $3$44 is almost homogeneous: either almost all edges or almost all non-edges, measured by arbitrary Keisler measures (Malliaris et al., 2015). The result explicitly gives no quantitative bounds and no equitability (Malliaris et al., 2015).

In distal theories, one obtains definable partitions into product cells so that the total measure of non-homogeneous cells is small (Simon, 2015). The higher-arity version similarly partitions each coordinate space, and all but a small measure of tuples of cells are homogeneous (Simon, 2015). These are regularity lemmas in a cylinder-like sense: the decomposition is into rectangular or product cells, and the defect is measured by product measure.

The contrast is structural. The definable/stable/distal results derive from local stability theory, smoothness of generically stable measures, weak orthogonality, and definable approximation by unions of rectangles (Malliaris et al., 2015, Simon, 2015). The Cylinder Regularity Lemma of (Gishboliner et al., 13 Aug 2025) is instead combinatorial and quasirandomness-based. Still, all of these frameworks privilege product-cell decompositions over arbitrary vertex refinements.

7. Applications, limitations, and interpretation

The Cylinder Regularity Lemma is introduced as the “right” intermediate object for obtaining better quantitative bounds from bounded VC dimension in graphs and bounded VC$3$45 dimension in $3$46-graphs (Gishboliner et al., 13 Aug 2025). It is weaker than a full hypergraph regularity lemma but strong enough for induced counting and for conversion to a genuine regularity partition after one additional regularization step (Gishboliner et al., 13 Aug 2025).

The paper also records several further applications of the hypergraph cylinder lemma. It proves a $3$47-uniform analogue of the Conlon–Fox quasirandom subset lemma, asserting that every $3$48-graph contains a linear-sized subset $3$49 and a graph $3$50 on $3$51 such that $3$52 is $3$53-quasirandom, with $3$54 doubly exponential in a polynomial of $3$55 (Gishboliner et al., 13 Aug 2025). It also proves a hypergraph analogue of Rödl’s theorem on induced-$3$56-free graphs: if $3$57 is induced-$3$58-free, then there is a large $3$59 and a quasirandom graph $3$60 on $3$61 such that $3$62 has density close to $3$63 or $3$64 (Gishboliner et al., 13 Aug 2025).

At the same time, the cylinder framework is not a universal replacement for ordinary regularity. In the graph case, it yields a product-space partition rather than a vertex partition (Gishboliner et al., 13 Aug 2025). In the hypergraph case, one still needs the final Szemerédi-type regularization to recover quasirandom lower-order graph structure after Venn refinement (Gishboliner et al., 13 Aug 2025). This suggests that cylinder regularity is best understood as an intermediate layer of structure: weaker than full regularity, but robust enough to carry quasirandomness and induced-forbidden-subgraph information.

A common misconception is to identify any product-partition regularity statement with the modern Cylinder Regularity Lemma. The comparison papers show that this would be too broad. Towsner’s bipartite bounded-VC theorem is a graph regularity lemma with product partitions, but not a higher-dimensional cylinder lemma (Towsner, 2013). Stable and distal regularity are definable measure-theoretic decompositions into product cells, but they do not use the quasirandom cylinder-chain machinery of (Gishboliner et al., 13 Aug 2025). Conversely, the cylinder lemma of (Gishboliner et al., 13 Aug 2025) is specifically designed to sit between full hypergraph regularity and induced counting.

In one sentence, the modern Cylinder Regularity Lemma is a product-space quasirandom decomposition principle for multipartite graphs and $3$65-graphs that replaces repeated full regularization by a quantitatively cheaper cylinder-level regularization, thereby enabling improved bounds in bounded VC and bounded VC$3$66 regularity theory (Gishboliner et al., 13 Aug 2025).