Simon's Factorization Forests

• Simon's Factorization Forests are bounded-depth decompositions of words using morphisms into finite semigroups to control complexity via idempotent grouping.
• They utilize automata-theoretic techniques and Ramsey splits to guarantee uniform height bounds, enabling unique parse trees and efficient rational expression constructions.
• Extensions to forest algebras and matching problems demonstrate the method's versatility in decomposing complex structures and deriving finite summaries for algorithmic applications.

Searching arXiv for the cited and closely related papers on Simon’s factorization forests. arXiv search: Simon factorization forest theorem, unambiguous forest factorization, Simon congruence piecewise testability, matching patterns Simon congruence, forest algebras factorization. Simon's factorization forests are bounded-depth decompositions of words relative to a morphism into a finite semigroup or monoid. In the classical tree form recalled in the literature, for every word $w\in\Sigma^+$ there is a factorization tree $T(w)$ of bounded height such that every internal node of arity $>2$ has all children labeled by the same idempotent element of the target semigroup. This decomposition principle is a central tool in algebraic automata theory and has wide connections in algebra, logic, and automata; more recent work also presents it through Ramsey splits, unambiguous automata, and extensions or analogues for richer structures such as forests (Gastin et al., 2018).

1. Classical statement and bounded-height factorization

The semigroup-theoretic form used in recent applications is the following. Let $(S,\cdot)$ be a finite semigroup. Then every word $w\in S^+$ has factorization height bounded uniformly by the size of $S$: $\hrank(w)\le 3|S|.$ Here $\hrank(w)$ is defined recursively by allowing either a binary factorization into two subwords or an unranked factorization into at least two subwords whose evaluations in the semigroup are all equal to the same idempotent. In this formulation, Simon’s theorem asserts that arbitrary words over a finite semigroup admit a recursively structured decomposition of uniformly bounded height (Bourneuf et al., 27 Feb 2026).

The tree formulation recalled in automata-theoretic treatments is equivalent in spirit: for every $w\in\Sigma^+$, there is a factorization tree $T(w)$ of bounded height where internal nodes of arity $T(w)$0 have all children labeled by the same idempotent element of $T(w)$1. Simon’s original bound was $T(w)$2, with later improvements. A plausible implication is that the theorem’s content is not merely the existence of a parse tree, but the existence of one in which unbounded branching is permitted only in semigroup regions that have stabilized to idempotents (Gastin et al., 2018).

This bounded-height phenomenon is the essential structural invariant behind the theorem. It ensures that long words can be decomposed recursively with complexity controlled by the finite target algebra rather than by the word length. In applications, the idempotent case is the key locality mechanism: when many consecutive factors evaluate to the same idempotent, they can be grouped without increasing depth beyond the uniform semigroup-dependent bound.

2. Idempotent branching, Ramsey splits, and automata-theoretic reformulation

A particularly clean reformulation is given through Ramsey splits induced by runs of a specially constructed automaton. A $T(w)$3-good automaton $T(w)$4 for a morphism $T(w)$5 is defined by four properties: $T(w)$6 is unambiguous and universal; for every state $T(w)$7, there exists an idempotent $T(w)$8 with $T(w)$9, where $>2$0 is the local loop language at $>2$1; the initial state has no incoming transitions and is maximal; and there is a unique final state with no outgoing transitions (Gastin et al., 2018).

Given such an automaton and a word $>2$2, let

$>2$3

be the unique accepting run. If $>2$4 is any monotone bijection, one defines a split $>2$5 by assigning label $>2$6 to position $>2$7. Two positions $>2$8 are then $>2$9-equivalent exactly when $(S,\cdot)$0 and all intermediate states are $(S,\cdot)$1; in that case the factor $(S,\cdot)$2 belongs to $(S,\cdot)$3, so

$(S,\cdot)$4

an idempotent. This is precisely the Ramsey-split condition (Gastin et al., 2018).

The resulting height bound is state-based: $(S,\cdot)$5 This automaton-theoretic view replaces a direct combinatorial proof by a universal unambiguous automaton whose state order enforces the required idempotent loop behavior. Conceptually, it makes the theorem look less like an ad hoc tree decomposition and more like a universal parsing principle for morphisms into finite semigroups.

3. Unambiguous forest factorization and good rational expressions

An unambiguous version of the theorem strengthens the classical conclusion by constructing universal rational expressions that are semigroup-compatible. For a morphism $(S,\cdot)$6 into a finite monoid, the unambiguous forest factorization theorem states, first, that for each $(S,\cdot)$7 there is an $(S,\cdot)$8-free good rational expression $(S,\cdot)$9 such that

$w\in S^+$0

and therefore

$w\in S^+$1

is an unambiguous rational expression over $w\in S^+$2 with $w\in S^+$3. Second, there is an unambiguous rational expression

$w\in S^+$4

over $w\in S^+$5 such that $w\in S^+$6, where each $w\in S^+$7 and $w\in S^+$8 is $w\in S^+$9-free and $S$0-good, and each corresponding $S$1 is idempotent with

$S$2

In these “good” expressions, every subexpression maps to a single semigroup element and every Kleene-plus subexpression maps to an idempotent (Gastin et al., 2018).

The construction proceeds by synthesizing a $S$3-good automaton from the morphism $S$4, using induction on the lexicographic measure

$S$5

Base cases cover the group case and the one-generator case. The inductive cases split according to whether some $S$6 satisfies $S$7 or $S$8, producing combined automata whose state orders enforce the loop-idempotent condition. The first inductive construction can remain deterministic if its components are deterministic, while the second introduces essential nondeterminism (Gastin et al., 2018).

This unambiguous strengthening is important because it yields a unique parse tree for every accepted word. The paper explicitly uses this property for the construction of regular transducer expressions corresponding to deterministic two-way transducers. The expressions are synthesized from the good automaton by state elimination, and the unique parse is what makes the function-like behavior of deterministic transducers compatible with the factorization framework (Gastin et al., 2018).

4. Relation to Simon’s congruence and piecewise testability

Simon's factorization forests are closely related to, but distinct from, Simon’s congruence. The congruence $S$9 on words over a finite alphabet $\hrank(w)\le 3|S|.$0 is defined by equality of scattered subwords of length at most $\hrank(w)\le 3|S|.$1: $\hrank(w)\le 3|S|.$2 Here “subword” means subsequence, not necessarily contiguous. This is the standard equivalence underlying piecewise testable languages: $\hrank(w)\le 3|S|.$3 is $\hrank(w)\le 3|S|.$4-piecewise testable iff membership in $\hrank(w)\le 3|S|.$5 depends only on the $\hrank(w)\le 3|S|.$6-class of a word; equivalently, piecewise testable languages are exactly those definable in $\hrank(w)\le 3|S|.$7 (Karandikar et al., 2013).

The quantitative study of this congruence shows that, for fixed alphabet size $\hrank(w)\le 3|S|.$8,

$\hrank(w)\le 3|S|.$9

where $\hrank(w)$0. The proof does not develop Simon’s factorization forest theorem directly. Instead, it introduces a different recursive decomposition, the “rich factorization,” based on words that are rich or $\hrank(w)$1-rich, and derives combinatorial recurrences such as

$\hrank(w)$2

and

$\hrank(w)$3

The conceptual connection is that both frameworks exploit recursive decompositions of words into pieces whose behavior can be controlled locally and composed globally (Karandikar et al., 2013).

A common misconception is that Simon’s factorization forests and Simon’s congruence are the same object. They are not. The congruence is an equivalence relation defined by short subsequences, whereas the factorization-forest theorem is a bounded-depth decomposition theorem for words under arbitrary morphisms into finite semigroups. The relationship is structural rather than definitional: both belong to the same combinatorial tradition of recursively summarizing words by semigroup- or subsequence-controlled behavior.

5. Arch factorizations, bounded summaries, and matching under Simon’s congruence

Algorithmic work on matching problems under Simon’s congruence further illustrates the factorization-forest viewpoint without reproving the classical theorem. For a word $\hrank(w)$4, let

$\hrank(w)$5

Then $\hrank(w)$6 iff $\hrank(w)$7. The paper interprets $\hrank(w)$8 together with $\hrank(w)$9 as a succinct representation of $w\in\Sigma^+$0 and relies heavily on the classical arch factorization of Hébrard, a canonical decomposition into $w\in\Sigma^+$1 arches plus a rest, where $w\in\Sigma^+$2 is the universality index defined by $w\in\Sigma^+$3 (Fleischmann et al., 2023).

The main structural tools are arch factorization, subsequence universality signatures, and marginal sequences. The signature of a word is a triple

$w\in\Sigma^+$4

where $w\in\Sigma^+$5 is a permutation of the letters appearing in the word ordered by first appearance, $w\in\Sigma^+$6 stores arch-count information, and $w\in\Sigma^+$7 stores residual alphabet information after the arches. A crucial compositional proposition states that for $w\in\Sigma^+$8, one can compute the signature of $w\in\Sigma^+$9 from the signatures of $T(w)$0 and $T(w)$1 in time polynomial in $T(w)$2 and $T(w)$3, where $T(w)$4 is the maximum entry of the $T(w)$5-arrays. The paper also proves bounded-witness statements, including the corollary that a valid signature has a representative word of length at most

$T(w)$6

up to a uniform shift in the $T(w)$7-values (Fleischmann et al., 2023).

These bounded summaries support sharp complexity results. The paper proves that $T(w)$8, $T(w)$9, and $T(w)$00 are $T(w)$01-complete; $T(w)$02 is $T(w)$03-complete; and $T(w)$04 is $T(w)$05-complete for $T(w)$06. It also isolates tractable cases, including regular patterns for $T(w)$07 and $T(w)$08, and cases where some variable occurs only once or the number of variables is constant for $T(w)$09 (Fleischmann et al., 2023).

The significance for Simon-style factorization is methodological. The paper replaces explicit manipulated words by finite structural summaries, proves pumping and bounded-witness lemmas, and composes summaries through patterns. This suggests that the practical force of factorization-forest ideas often lies less in the tree theorem itself than in the availability of canonical decompositions and finite invariants that preserve the relevant congruence behavior.

6. Extensions beyond words and representative applications

A direct generalization from words to forests requires new machinery. In the setting of forest algebras, the main theorem states: $T(w)$10 Here $T(w)$11 measures the least uniform bound on the depth of full general decompositions of forests recognized by $T(w)$12. The decomposition formalism includes not only leaves and binary nodes but also idempotent nodes and centipede nodes. The essential new difficulty is that cutting a subforest leaves a context with a hole rather than another forest, so the proof augments the free forest algebra with default holes $T(w)$13, tracks references, and uses rotations to reorder cuts while preserving semantics (Almagor et al., 11 May 2026).

The semantic restriction of $T(w)$14-alignment is decisive. A forest $T(w)$15 is $T(w)$16-aligned over $T(w)$17 if, whenever

$T(w)$18

for distinct contexts $T(w)$19 with

$T(w)$20

there exists $T(w)$21 such that

$T(w)$22

Without a restriction of this kind, bounded-depth decomposition can fail: the paper proves that there exists a morphism $T(w)$23 that is not $T(w)$24-aligned and for which $T(w)$25 (Almagor et al., 11 May 2026).

A graph-theoretic application shows how factorization forests function as a general structural method. For every connected graph $T(w)$26 of pathwidth less than $T(w)$27, there is a tree decomposition of width at most $T(w)$28 indexed by a spanning tree $T(w)$29 of $T(w)$30, with every vertex belonging to its own bag. The proof starts from a nice path decomposition, encodes it as a word

$T(w)$31

over a finite semigroup $T(w)$32 of abstractions of $T(w)$33-interface graphs, applies Simon’s theorem to obtain

$T(w)$34

and then proves inductively that

$T(w)$35

This yields the explicit bound

$T(w)$36

with $T(w)$37 (Bourneuf et al., 27 Feb 2026).

This application also clarifies a limitation. The width bound is controlled by pathwidth, not treewidth. The stated reason is structural: the argument starts from a path decomposition and uses a theorem about words or sequences. Bounded treewidth does not supply the same linear order with tight interface control, and the paper explicitly notes that the negative result of Blanco, Cook, Hatzel, Hilaire, Illingworth, and McCarty rules out an analogous bound in terms of treewidth alone (Bourneuf et al., 27 Feb 2026).