Monotone Labelling Method Overview

• Monotone Labelling Method is a family of techniques with order-based constraints used to optimize label assignments in trees, classification, permutation dynamics, and data encoding.
• It leverages monotonicity to restrict label domains, compress search spaces, and induce interval structures that enable efficient optimization and canonicalization.
• Applications span optimal integer tree labeling, active monotone classification, Rauzy dynamics, grid recognition, graph encoding, and dynamic list labeling.

Monotone Labelling Method denotes, in the literature surveyed here, not a single standardized algorithm but a family of techniques in which labels are constrained by a monotonicity principle and that constraint is then exploited for optimization, inference, classification, combinatorial normalization, or compact encoding. The phrase appears in several technically distinct settings: optimal integer labeling of trees under monotone edge penalties, active monotone classification on partially ordered point sets, canonicalization in Rauzy-type dynamics, monotone transformations of multiclass learners, query-efficient recognition on multi-valued grids, adjacency labeling of monotone graph classes, and monotonic list labeling in dynamic data structures (Bolshoy et al., 2013, Tao, 12 Jun 2025, Mourgues, 2018, Bousquet et al., 2022, Aslanyan et al., 2021, Bonnet et al., 2023, Bender et al., 2024).

1. Disambiguation and shared structural motif

The term is used for structurally different objects and objectives.

Domain	Object being labeled	Monotonicity role
Tree optimization	Integer labels on vertices	Edge penalty $f(|\Delta|)$ is monotone
Ordered classification	Binary labels or queried labels on points	Positive region is an upper set
Rauzy dynamics	Endpoint and cycle labels in permutation diagrams	Progress to canonical form is monotone
Supervised learning	Output labels of a wrapped learner	Expected risk is non-increasing in sample size
Grid recognition	Binary values on grid points	Each induced cube carries a monotone Boolean function
Graphs and data structures	Binary adjacency labels or order labels	Subgraph closure or strict order preservation

In these works, “labelling” may mean assigning integers to tree vertices, assigning cyclic or positional tags to endpoints in a permutation diagram, revealing labels of queried points, or encoding vertices by binary strings so that adjacency can be decoded locally. “Monotone” may refer to a non-decreasing cost function, an isotonic classifier with respect to a partial order, a progress functional that increases under generators, a risk curve that decreases with sample size, or a class closed under subgraphs (Bolshoy et al., 2013, Mourgues, 2018, Tao, 12 Jun 2025, Bousquet et al., 2022, Aslanyan et al., 2021, Bonnet et al., 2023).

This suggests a family resemblance rather than a single canonical construction. Across the cited papers, monotonicity is used to exclude infeasible labels, compress the search space, induce interval or convex structure, or certify that local updates move toward a canonical or optimal object.

2. Optimal integer tree labeling under monotone edge costs

For trees, the method is an exact optimization procedure. A finite tree $G=(V,E)$ is labeled by positive integers, leaves have prescribed labels, each edge $(u,v)$ incurs cost $f(|L(u)-L(v)|)$, and the objective is to minimize

$C(L)=\sum_{(u,v)\in E} f\big(|L(u)-L(v)|\big),$

where $f:\mathbb{N}\to \mathbb{R}_{\ge 0}$ is monotone non-decreasing and satisfies $f(0)=0$ (Bolshoy et al., 2013).

The basic algorithm is a rooted tree DP. If $D$ is the label domain and $T_v$ is the subtree rooted at $v$, then

$G=(V,E)$0

Leaves satisfy the boundary condition $G=(V,E)$1 for $G=(V,E)$2 and $G=(V,E)$3 otherwise. The optimal total cost is $G=(V,E)$4, and an optimal labeling is reconstructed by storing, for each child and parent label, an $G=(V,E)$5 value used in a top-down traceback.

A central structural lemma bounds the label space. If $G=(V,E)$6 and $G=(V,E)$7, then every optimal labeling satisfies $G=(V,E)$8 for every node. The proof uses only monotonicity of $G=(V,E)$9 and $(u,v)$0: pushing any out-of-range label inward toward $(u,v)$1 weakly decreases all incident edge costs. Consequently one may take $(u,v)$2 and $(u,v)$3.

For arbitrary monotone $(u,v)$4, the naïve DP runs in $(u,v)$5 time and $(u,v)$6 space. The paper then isolates the special case $(u,v)$7, where two faster exact procedures are available. On binary trees, interval propagation assigns to each node an interval $(u,v)$8 of optimal labels: leaves get singleton intervals, parents take the intersection of child intervals if nonempty, and otherwise the interval spanning the gap between them. A top-down pass chooses any root label in $(u,v)$9 and assigns to each child the closest point in its interval. This yields an $f(|L(u)-L(v)|)$0-time, $f(|L(u)-L(v)|)$1-space algorithm. On general rooted trees, the same $f(|L(u)-L(v)|)$2 cost can be accelerated by a one-dimensional distance transform, reducing the DP time to $f(|L(u)-L(v)|)$3 while keeping $f(|L(u)-L(v)|)$4 space.

The $f(|L(u)-L(v)|)$5 case also exhibits the interval structure that motivates the label “monotone” in this setting. Optimal labels at a node form a contiguous interval, the root optimum lies between child labels, and the DP profiles are discrete convex. The details are stated through a sequence of lemmas on binary trees, and more generally through min-plus convolution with convex costs. In context, this is essentially a specialized Sankoff-style tree DP for numeric states, with interval propagation supplying an exact analogue of set-based parsimony in the Manhattan case.

3. Monotone classification and monotone learning

In active monotone classification, the input is a finite multiset $f(|L(u)-L(v)|)$6 with hidden binary labels and the coordinate-wise partial order $f(|L(u)-L(v)|)$7 for all $f(|L(u)-L(v)|)$8. A classifier $f(|L(u)-L(v)|)$9 is monotone if $C(L)=\sum_{(u,v)\in E} f\big(|L(u)-L(v)|\big),$0 implies $C(L)=\sum_{(u,v)\in E} f\big(|L(u)-L(v)|\big),$1, equivalently if its positive region is an upper set. The error on $C(L)=\sum_{(u,v)\in E} f\big(|L(u)-L(v)|\big),$2 is

$C(L)=\sum_{(u,v)\in E} f\big(|L(u)-L(v)|\big),$3

and the optimum is $C(L)=\sum_{(u,v)\in E} f\big(|L(u)-L(v)|\big),$4 (Tao, 12 Jun 2025).

The paper develops two distinct algorithms. The first, Random Probes with Elimination (RPE), repeatedly samples a point uniformly at random, reveals its label, and removes dominated or dominating points according to the observed sign. It reveals $C(L)=\sum_{(u,v)\in E} f\big(|L(u)-L(v)|\big),$5 labels in expectation, where $C(L)=\sum_{(u,v)\in E} f\big(|L(u)-L(v)|\big),$6 is the dominance width, and returns a monotone classifier $C(L)=\sum_{(u,v)\in E} f\big(|L(u)-L(v)|\big),$7 with expected error at most $C(L)=\sum_{(u,v)\in E} f\big(|L(u)-L(v)|\big),$8. In the realizable case $C(L)=\sum_{(u,v)\in E} f\big(|L(u)-L(v)|\big),$9, the output is always optimal. The second algorithm constructs a weighted relative-comparison coreset $f:\mathbb{N}\to \mathbb{R}_{\ge 0}$0 of size $f:\mathbb{N}\to \mathbb{R}_{\ge 0}$1 such that, with probability at least $f:\mathbb{N}\to \mathbb{R}_{\ge 0}$2, a single $f:\mathbb{N}\to \mathbb{R}_{\ge 0}$3-independent offset $f:\mathbb{N}\to \mathbb{R}_{\ge 0}$4 satisfies

$f:\mathbb{N}\to \mathbb{R}_{\ge 0}$5

for all monotone $f:\mathbb{N}\to \mathbb{R}_{\ge 0}$6. Minimizing $f:\mathbb{N}\to \mathbb{R}_{\ge 0}$7 then yields a classifier with $f:\mathbb{N}\to \mathbb{R}_{\ge 0}$8. The same work proves lower bounds: exact optimality requires $f:\mathbb{N}\to \mathbb{R}_{\ge 0}$9 labels even in one dimension, constant-factor approximation requires $f(0)=0$0 in expectation, and for $f(0)=0$1 there are two-dimensional instances requiring $f(0)=0$2 labels.

A different use of the phrase appears in multiclass supervised learning. There the goal is not to query point labels under a partial order, but to transform any learner $f(0)=0$3 into a learner whose expected population risk is monotone in the sample size. For labels $f(0)=0$4, the construction maps each hypothesis $f(0)=0$5 to the cyclic base class

$f(0)=0$6

where $f(0)=0$7, and uses a randomized ERM $f(0)=0$8 over $f(0)=0$9, a regularization operator $D$0, and an update operator $D$1 on a block schedule $D$2 (Bousquet et al., 2022). The resulting wrapper $D$3 is monotone with respect to every distribution $D$4 in the sense that

$D$5

for all $D$6. It is also competitive: for every $D$7 there exists $D$8 with $D$9 such that

$T_v$0

The same paper derives Bayes-consistent monotone learners and monotone agnostic PAC learners from arbitrary base algorithms.

These two literatures are technically unrelated but conceptually parallel. In one, monotonicity is an order constraint on the classifier itself and the main resource is label complexity. In the other, monotonicity is a property of the expected learning curve and the main resource is black-box access to a base learner.

4. Canonicalization in Rauzy-type dynamics

In Rauzy-type dynamics, the labelling method is a combinatorial classification technique for group actions on finite objects, especially irreducible permutations associated with interval exchange transformations. The basic generators are denoted $T_v$1 and $T_v$2; they act by reindexing endpoints according to special permutations $T_v$3 and $T_v$4, and connectedness is studied in the Cayley graph generated by these moves (Mourgues, 2018).

The method assigns consistent labels to the endpoints of a permutation diagram. From the diagram one constructs closed cycles and one open rank path. If the cycle multiset is $T_v$5 and the rank is $T_v$6, then cycle arcs receive cyclic labels in $T_v$7 and the rank path receives labels in $T_v$8, together with boundary-order data. The basic invariant is the pair $T_v$9, satisfying

$v$0

and preserved by $v$1 and $v$2.

A second invariant is the Arf-type sign. With $v$3 the edge set and $v$4 the number of non-crossing pairs in $v$5, the paper defines

$v$6

and sets $v$7. The combinatorial proof shows that $v$8 and $v$9 classify Rauzy classes outside two exceptional hyperelliptic families.

The “monotone” aspect is a progress functional $G=(V,E)$00 on the label space. After fixing a canonical representative $G=(V,E)$01 for a given invariant class, one measures how many labels or anchors are in their canonical positions. The key lemma states that if the current labelled configuration is not canonical, then there exists a finite word in $G=(V,E)$02 that strictly increases $G=(V,E)$03. Since $G=(V,E)$04 is bounded above, repeated application reaches $G=(V,E)$05 in finitely many steps. This turns class membership into an explicit path construction problem.

The same framework supports the reduction-and-boosting argument used in the combinatorial proof of the Kontsevich–Zorich–Boissy classification theorem. One deletes a carefully chosen colored edge, applies induction in size $G=(V,E)$06, and reinserts the edge while tracking labels and invariants. In this setting, “labelling method” refers neither to numeric optimization nor to statistical classification, but to a systematic way of transporting invariants and forcing monotone progress toward standard representatives.

5. Cube-splitting recognition on multi-valued grids

For multi-valued grids, the method addresses oracle recognition of an unknown monotone binary function. The grid is

$G=(V,E)$07

ordered componentwise, and a binary labeling $G=(V,E)$08 is monotone when $G=(V,E)$09 implies $G=(V,E)$10 (Aslanyan et al., 2021).

The grid is partitioned into “vertical equivalence classes” indexed by the upper homogeneous area

$G=(V,E)$11

For each anchor $G=(V,E)$12, the class

$G=(V,E)$13

is a poset isomorphic to a Boolean cube of dimension $G=(V,E)$14, where $G=(V,E)$15. The bijection $G=(V,E)$16 records, on each active coordinate, whether $G=(V,E)$17 equals the upper value $G=(V,E)$18 or its complementary lower value $G=(V,E)$19.

Each anchor induces a Boolean function

$G=(V,E)$20

where $G=(V,E)$21 is the reverse map from the Boolean code to the original grid. If $G=(V,E)$22 is monotone on $G=(V,E)$23, then each $G=(V,E)$24 is monotone on the corresponding Boolean cube. Recognition therefore reduces to running Hansel’s chain-splitting method independently on each induced cube, resolving each Boolean query by mapping it back through $G=(V,E)$25 and asking the oracle for the value of $G=(V,E)$26 at the corresponding grid point.

The total query complexity is bounded by

$G=(V,E)$27

where $G=(V,E)$28 and $G=(V,E)$29 is the Hansel complexity for a Boolean cube of that dimension. A lower-oriented symmetric variant, anchored in the lower homogeneous area, is useful when zeros rather than ones are sparse.

The paper is explicit about a limitation that matters for interpretation. Monotonicity on every induced cube is necessary but not sufficient for global monotonicity on the full grid, because comparable pairs may lie in different classes. Thus the method is a query-efficient recognition procedure under the monotonicity assumption; it is not, by itself, a complete monotonicity test for arbitrary noisy data. The same cube framework extends to ordinal labels by thresholding into binary monotone functions.

6. Graph labels, list labels, and monotone encoding

In graph theory, the term enters through adjacency labeling schemes for monotone graph classes, where “monotone” means subgraph-closed. A class $G=(V,E)$30 admits labels of size $G=(V,E)$31 if every $G=(V,E)$32-vertex graph in the class can assign each vertex a binary string of length $G=(V,E)$33 from which adjacency of any two vertices is decidable using only the two labels. If $G=(V,E)$34, then every graph in $G=(V,E)$35 has at most $G=(V,E)$36 edges, every subgraph has a vertex of degree at most $G=(V,E)$37, and the class is therefore $G=(V,E)$38-degenerate. The standard degeneracy scheme then yields labels of size

$G=(V,E)$39

The same paper proves this upper bound is tight up to the $G=(V,E)$40 factor: for any decent $G=(V,E)$41 with $G=(V,E)$42, there exists a monotone class of speed $G=(V,E)$43 with no adjacency labeling scheme of size at most $G=(V,E)$44. For small monotone classes, defined by $G=(V,E)$45, bounded degeneracy implies the optimal order $G=(V,E)$46 (Bonnet et al., 2023).

In dynamic data structures, the Monotonic List Labeling Problem requires labels to remain strictly increasing with list order under online insertions. The layered technique for list labeling treats labels as array indices in a packed array and composes three algorithms to inherit worst-case, expected, and adaptive guarantees simultaneously. The full array has size $G=(V,E)$47; an $G=(V,E)$48-emulator operates on a moving subarray of size $G=(V,E)$49, while an $G=(V,E)$50-shell sees only $G=(V,E)$51 genuinely empty slots and treats the rest as occupied. Rebuilds toward fixed checkpoints are carried out in two phases—left alignment and rightward placement—and a buffered item participates in at most $G=(V,E)$52 deadweight moves total. The resulting composition theorem implies that, for suitable choices of component algorithms, one obtains worst-case cost $G=(V,E)$53, expected cost $G=(V,E)$54, and adaptive amortized bounds such as $G=(V,E)$55 for hammer-insert workloads or $G=(V,E)$56 with prediction error $G=(V,E)$57. Throughout, monotonicity means the strict order invariant

$G=(V,E)$58

which the embedding preserves by controlled slot reclassification and sliding (Bender et al., 2024).

These uses are again distinct from the tree, classification, Rauzy, and grid settings. Here the label is not a queried class value or a canonical cycle tag, but an information-bearing local code or an order-preserving array position. The commonality is that monotonicity supplies a structural invariant—subgraph closure for graph classes, strict order preservation for list labels—from which sharp bounds and compositional constructions follow.