Marcolli, Chomsky and Berwick (2025): Mathematical Structure of Syntactic Merge #
This file formalizes parts of Marcolli, Chomsky and Berwick's algebraic model of Merge.
Internal Merge (§1.4.4) is checked on the book's example, which raises the apple out of
T = {was, {eaten, {the, apple}}}. The unit stage M_{T₂,1} splits T into the apple beside
the quotient T/T₂, the composite with M_{T/T₂,T₂} merges the two (Proposition 1.4.2), and the
step satisfies Minimal Yield under trace counting (Proposition 1.6.4).
The core computational structure of Merge (§1.10) is a fixed point. Proposition 1.10.2 works in
V(𝔗), the free ℤ-module on the nonplanar binary trees with unlabelled leaves, graded by the
number of leaves. It generates these trees by recursively solving X = 𝔐(X, X) over formal sums
X = ∑ Xₗ, with the initial condition X₁ = x, and reads the integer coefficients of the solution
as numbers of planar embeddings. Here V(𝔗) sits in the Connes–Kreimer algebra of unlabelled
trees, Merge is the grafting operator after the disjoint union (Lemma 1.3.3), and a formal sum is a
power series whose degree counts leaves. The sum of the planar binary trees, their planar
structure forgotten, solves the equation (isMergeSolution_mergeSolution) and is its only solution
(IsMergeSolution.eq). The coefficient of each tree is its number of planar embeddings
(coeff_planarSum_succ), and the coefficients of Xₙ₊₁ sum to the Catalan number Cₙ
(aeval_planarSum_succ). On syntactic objects, Lemma 1.3.3 is toCK_merge.
The book's worked examples of externalization are stated on the SyntacticObject carrier of
the Minimalist substrate: the harmonic head-initial and head-final orders of a determiner–noun
Merge, the head-side convention flipping the yield, and exocentric elimination, two saturated
nouns determining no head and hence no order. The framework itself is the Syntax/Minimalist/
theory layer; the examples are kernel-checked against it.
FormSet (§1.16) groups components of a workspace under a new root. Definition 1.16.1 writes it as
⊔ ∘ (B ⊗ id) ∘ Π_(k) ∘ Δ_P, where the primitive coproduct Δ_P splits a workspace in every way,
so FS^(k)(F) sums B(S) ⊔ (F - S) over the k-component subworkspaces S (formSet_of'). On
workspaces of syntactic objects it lands in the extended workspaces of (1.16.3) (map_formSet_le).
The book's syntax–semantics interface (Chapter 3) replaces per-feature checking by a single
recursive map, the Birkhoff renormalization of a character of the Connes–Kreimer Hopf algebra
of the syntactic object. The Boolean parsing semiring Consistency of §3.5 is the target,
toCK embeds a syntactic object into the Hopf algebra, featureConsistency is the renormalized
character, and headConsistency instantiates it for the head-following character of Lemma 3.2.5,
whose probe reads the §1.13 selection head. The machinery is noncomputable, so these are
specifications of consistency rather than checkers. In §3.5.2 the operator is the projection of
Laurent series onto their polar part. The renormalized part of any character then lands in the
power series (truncLT_birkhoffPlus_of'), while the grading character ϕt is already nonpolar
and is its own renormalized part (birkhoffPlus_gradingChar), so it cannot single out Sideward
Merge.
Obligatory control (§3.8.2) is derived by External Merge alone. In the man tried to read a
book both inscriptions of the man merge into theta positions, as repetitions of one another,
and Form Copy restricts the object to the diagonal on which they are one
(theMan_mem_copyRel), while a book, not structurally identical to the man, is no copy of it
(not_mem_copyRel_aBook).
Correspondence with the book #
The book's algebra lives in Core/ and in Syntax/Minimalist/{Workspace,Merge,Economy}/:
- Definitions 1.2.6 and 1.2.8 (admissible cuts, the coproducts
Δ^ω):ConnesKreimer.cutSummandsNandcomulAlgHomNforΔ^ρ, which Remark 1.2.9 identifies with the Connes–Kreimer coproduct;cutSummandsCNandcomulCAlgHomNforΔ^c. - Lemma 1.2.10:
ConnesKreimer.comulCN_coassocand the edge grading ofCoproduct/TraceGrading. - Lemma 1.2.11 and (1.2.12):
ConnesKreimer.instBialgebraRho, theHopfAlgebrainstance, andantipode_ofTree. - The projections
Π_{ρ,c}andΠ_{d,ρ}and the relationsΔ^d = (id ⊗ Π_{d,ρ}) ∘ Δ^ρ = (id ⊗ Π_{d,c}) ∘ Δ^cdisplayed before (1.3.10):ConnesKreimer.eraseTraces,UnorderedTree.contractUnary,ConnesKreimer.cutSummandsDN,ConnesKreimer.comulDAlgHomN,ConnesKreimer.eraseTraces_comulCAlgHomN,contractUnary_eraseTraces_comulCAlgHomN. - Definition 1.3.4:
Minimalist.Merge.mergeOpGover a cut enumeration, withmergeOpatΔ^ρandmergeOpCatΔ^c; (1.3.11):Minimalist.Merge.mergeOpG_apply_of'. - Lemma 1.4.1:
Minimalist.Merge.mergeOpG_pair,mergeOpG_pair_residual, with the residual workspace passing through bymergeOpG_of'_mul. - Proposition 1.4.2:
Minimalist.Merge.mergeOpG_im_composition; on syntactic objectsMinimalist.SyntacticObject.mergeOpC_im, and for a derivationMinimalist.Derivation.mergeOpList_initial. - Lemmas 1.4.4 and 1.4.5:
Minimalist.Merge.mergeOpG_sideward_2b,mergeOpG_sideward_3b,mergeOpG_sideward_3a. - (1.5.1) and the weights of §1.5.2:
Minimalist.Merge.comulEps,mergeOpEps, withcomulEps_oneandcomulEps_zero(the primitive coproduct); Proposition 1.5.1 atε = 0:mergeOpEps_zero_of'_of_not_le,mergeOpEps_zero_pair_add; at everyε:mergeOpEps_pair,mergeOpEps_im_composition,mergeOpEps_of'_eq_smul_of_not_le. - Definition 1.6.1 and Proposition 1.6.4:
Minimalist.MinimalYieldover a counting,MinimalYield.em_pair,em_pair_accessibleCount,im_accessibleCount_of_cut. - Lemma 1.6.3 (1.6.8) and (1.6.10):
ConnesKreimer.cutSummandsCN_accessibleCount,cutSummandsCN_accessibleSize. Proposition 1.6.8 and Remark 1.6.7, on the terms of Merge (Minimalist.Merge.mergeOpG_apply_of'_eq_sum):Minimalist.card_singleton_add_trunk_add_numWhole,accessibleCount_merge_term,accessibleSize_merge_term,MinimalYield.merge_term_accessibleCount_iff,accessibleCount_unit_termforΔ^c, andMinimalYield.merge_term_numEdges_iffforΔ^d. - Lemma 1.6.3 (1.6.7):
ConnesKreimer.cutSummandsDN_numEdges, from the unary vertices a cut leaves (ConnesKreimer.cutSummandsN_numUnary); for Internal Merge,Minimalist.numEdges_im_of_cut,MinimalYieldWeak.im_numEdges_of_cut,MinimalYield.not_im_numEdges_of_cut. Each cut's cost over a workspace:ConnesKreimer.CutChoice.HasExtractionCost. - Definition 1.6.2 and Proposition 1.6.10:
Minimalist.NoComplexityLoss, met by every admissible derivation step (Minimalist.Step.noComplexityLoss); along the induced map of a term of Merge (Minimalist.componentMap,NoComplexityLoss.of_componentMap),NoComplexityLoss.map_componentMap_merge_iff,not_map_componentMap_merge,not_map_componentMap_unit, and at the trace cuts graded byσ,not_map_accessibleSize_componentMap_merge. - Definition 1.6.2's degree
#L:UnorderedTree.numLeaves; Lemma 1.6.3:ConnesKreimer.cutSummandsCN_numNodes. - The functionals of (1.7.5) and their Lie algebra:
Bialgebra.IsDualPrimitiveandBialgebra.dualPrimitives; they are supported on single trees (ConnesKreimer.isDualPrimitive_iff). - Lemma 1.7.3, for
Δ^ρand grafting at vertices:ConnesKreimer.isDualPrimitive_lcoeff_iff,lie_lcoeff_singleton_apply_ofTree,lie_lcoeff_singleton_apply_of', andGrossmanLarson.pairingLiefor the bracket of the grafting product. - Definition 3.1.1 and the weight
+1identity of Definition 3.1.2:IsRotaBaxterat weights-1and+1; the identity operator of Remark 3.2.2, at weight+1on the Boolean semiring:Consistency.isRotaBaxter_id. - Definitions 3.1.3, 3.1.5 and 3.1.6, Proposition 3.1.7, Remark 3.1.8:
ConnesKreimer.birkhoffMinus,birkhoffPrepTree,birkhoffPlus,birkhoffPlus_eq_convMul,birkhoffFactorization, with the operator-R; the subalgebras of Proposition 3.1.7:ConnesKreimer.birkhoffMinus_of'_mem_range,birkhoffPlus_of'_mem_range; Proposition 3.1.9:ConnesKreimer.convMul_birkhoffMinus_apply_ofTree. - Proposition 3.5.2 and (3.5.4):
HahnSeries.isRotaBaxter_truncLT. The gradings of §3.5.2.1:δb₀,δα,δσ,minimalYield_iff_gradings,minimalYieldWeak_iff_truncLT. Proposition 3.5.3 forδα, without the derivation coefficients:gradingChar; Lemma 3.5.5:truncLT_gradingChar_of'; the first step of the proof of Proposition 3.5.6:truncLT_birkhoffPlus_of'.
Implementation notes #
In §1.4.4 the book displays the result with the deletion quotient T/d T₂ = {was, eaten}, and
alternatively with the contraction quotient T/c T₂, whose contracted leaf it labels by the
extracted term {the, apple}. The trace cuts of the carrier label that leaf by the head of the
extracted term (traceEncoder), so the example's quotient is {was, {eaten, t}} with t the
trace of the.
The equation carries its initial term, X = x t + 𝔐(X, X). The book writes X = 𝔐(X, X) and
supplies X₁ = x as an initial condition, but 𝔐(X, X) has no degree-one term
(coeff_one_mergeSeries), so the leaf enters as an inhomogeneous term.
TreeAlgebra is spanned by the forests of unlabelled trees, not only by the binary trees of 𝔗.
V(𝔗) is the span of the single binary trees in it, and every term of the solution lies there.
formSet is defined for every k on all forests, with the root label of B a parameter.
Definition 1.16.1 takes k ≥ 3 and the domain V(𝔉_{SO₀}) of workspaces of syntactic objects;
map_formSet_le is stated on that domain, with the bare root label of the syntactic objects.
The proof of Proposition 1.10.2 prints X₄ = 2{x{x{xx}}} + {{xx}{xx}}. Its recursion gives
𝔐(X₁, X₃) + 𝔐(X₂, X₂) + 𝔐(X₃, X₁) = 4{x{x{xx}}} + {{xx}{xx}} (planarSum_four), and four is
the number of planar embeddings that the proof says the coefficient counts: the five planar binary
trees with four leaves split as four and one.
Definition 3.1.2's weight -1 semiring identity
R(a) ⊙ R(b) ⊡ R(a ⊙ b) = R(a ⊙ R(b)) ⊡ R(R(a) ⊙ b) is not the IsRotaBaxter identity at any
weight, and Remark 3.2.2 and Lemma 3.2.7 use R = id at that weight on the tropical semiring. On
Consistency the identity has weight +1 as well, since disjunction is idempotent, and the
consistency maps use it at that weight.
TODO #
§1.17 reads (1.10.1) as the quadratic case of the combinatorial Dyson–Schwinger equation
X = B(P(X)) of (1.17.2), with the grafting operator B of Definition 1.3.2 and the recursive
solution (1.17.3), which the book cites rather than proves. Stating it needs B on formal series
in the grading by vertices, in which B raises the degree by one. With x = B(1) the Merge case
is P(t) = 1 + t², while the book writes P(X) = X² next to its requirement a₀ = 1.
Lemma 1.10.1 identifies 𝔗 with the free nonassociative commutative magma on one generator; the
link to the syntactic-object carrier, the free commutative magma on the lexical items, is not
stated.
Lemma 1.16.5 extends the coproducts Δ^ω to extended workspaces by excluding the cuts of edges at
a root of valence at least three, so that Merge does not undo the grouping FormSet builds
(§1.16.2). It is not formalized.
Proposition 1.6.8's row for the Sideward case 3(a) under Δ^d lowers α by two, counting two
contracted edges. When the two terms are sisters, as the and apple are in the apple, the
deletion leaves their parent a leaf and contracts nothing, so α is unchanged. The term still
diverges, so the proposition's conclusion stands.
Corollary 3.5.4's character ψt sums over the pairs of intermediate derivations of a target,
which needs a Merge step relation on workspaces; it, and Proposition 3.5.6 beyond the first step of
its proof, are not formalized.
The book's section locators (§1.12.1, §1.13, §1.13.2) are transcribed from an earlier version of this file and are UNVERIFIED against the published text.
References #
Internal Merge: an example (§1.4.4) #
A token of the simple lexical item of category c selecting sel, pronounced pf.
Equations
- MarcolliChomskyBerwick2025.tok c sel pf i = { item := Minimalist.LexicalItem.of c sel pf, id := i }
Instances For
Sideward Merge, Minimal Search and economy on the workspace of §1.3.4 #
The workspace {eaten, {the, apple}} of §1.3.4 also admits Sideward Merge of type 2(b), which
merges eaten with the cut out of the apple, and the apple alone admits Internal Merge of
the and Sideward Merge of type 3(a) of the and apple. On these the general results of §1.4–1.6
give Sideward Merge a nonzero term, which Minimal Search removes at ε = 0 while it keeps
External and Internal Merge. The 2(b) term satisfies Minimal Yield under trace counting, as
Remark 1.6.9 says. No Complexity Loss excludes it at the pruning cuts graded by vertices, and at
the trace cuts graded by the size σ. At the trace cuts the vertex count does not exclude it,
because the trace of the takes the place of the.
The core computational structure of Merge (§1.10) #
The Connes–Kreimer algebra of unlabelled trees over ℤ, in which the free ℤ-module V(𝔗)
on the nonplanar binary trees sits as the span of single trees.
Equations
Instances For
Forgetting the planar structure of a planar binary tree, whose leaves are its nils.
Equations
- MarcolliChomskyBerwick2025.forgetPlanar BinaryTree.nil = UnorderedTree.leaf ()
- MarcolliChomskyBerwick2025.forgetPlanar (BinaryTree.node value l r) = UnorderedTree.node () {MarcolliChomskyBerwick2025.forgetPlanar l, MarcolliChomskyBerwick2025.forgetPlanar r}
Instances For
The single variable x, the one-leaf tree.
Instances For
Merge on V(𝔗) is the grafting operator after the disjoint union (Lemma 1.3.3), a bilinear
map.
Equations
- MarcolliChomskyBerwick2025.mergeLin = (LinearMap.mul ℤ MarcolliChomskyBerwick2025.TreeAlgebra).compr₂ (ConnesKreimer.bPlusLin ())
Instances For
Merge extends to formal series X = ∑ Xₗ degree by degree, the degree-n term of 𝔐(X, Y)
being ∑_{i+j=n} 𝔐(Xᵢ, Yⱼ).
Equations
- MarcolliChomskyBerwick2025.mergeSeries X Y = PowerSeries.mk fun (n : ℕ) => (ConnesKreimer.bPlusLin ()) ((PowerSeries.coeff n) (X * Y))
Instances For
𝔐(X, X) has no degree-one term, so the printed X = 𝔐(X, X) cannot meet X₁ = x.
A series solves (1.10.1) with its initial condition X₁ = x when it has no constant term and
X = x t + 𝔐(X, X).
Equations
- MarcolliChomskyBerwick2025.IsMergeSolution X = ((PowerSeries.coeff 0) X = 0 ∧ X = (PowerSeries.monomial 1) MarcolliChomskyBerwick2025.generator + MarcolliChomskyBerwick2025.mergeSeries X X)
Instances For
The formal sum of the planar binary trees with n leaves, planarity forgotten.
Equations
- MarcolliChomskyBerwick2025.planarSum 0 = 0
- MarcolliChomskyBerwick2025.planarSum n.succ = ∑ P ∈ BinaryTree.treesOfNumNodesEq n, ConnesKreimer.ofTree (MarcolliChomskyBerwick2025.forgetPlanar P)
Instances For
The series X = ∑ₗ Xₗ of Proposition 1.10.2.
Equations
Instances For
The sum of the planar binary trees solves (1.10.1) with X₁ = x (Proposition 1.10.2).
The solution of Proposition 1.10.2 is unique, since the recursion determines each term from the lower ones.
The coefficient of a nonplanar tree in Xₙ₊₁ is its number of planar embeddings.
The coefficients of Xₙ₊₁ sum to the Catalan number Cₙ, the number of planar binary trees
with n + 1 leaves.
X₄ = 4{x{x{xx}}} + {{xx}{xx}}; the proof of Proposition 1.10.2 prints the coefficient
2.
Externalization (§1.12–1.13) #
FormSet (§1.16) #
FormSet FS^(k) = ⊔ ∘ (B ⊗ id) ∘ Π_(k) ∘ Δ_P (Definition 1.16.1, (1.16.2)) splits the
workspace by the primitive coproduct, keeps the terms whose left factor has k components,
grafts those under a new root labelled a, and multiplies back.
Equations
- One or more equations did not get rendered due to their size.
Instances For
FormSet groups k components of a workspace in every way, summing B(S) ⊔ (F - S) over
the k-component subworkspaces S of F, counted with multiplicity (Remark 1.16.4).
With fewer than k components there is nothing to group.
With exactly k components FormSet groups the whole workspace, as in the unbounded
unstructured sequences of p. 144.
A forest is an extended workspace, in 𝔉̃ᴿ of (1.16.3), when each component is a syntactic
object or a bare root of any valence over syntactic objects, binary below that root.
Equations
- One or more equations did not get rendered due to their size.
Instances For
On workspaces of syntactic objects, the range of FormSet lies in the span of the extended workspaces (Definition 1.16.1).
Feature consistency as Birkhoff renormalization (Chapter 3) #
The Boolean consistency semiring of §3.5, the two-element idempotent commutative semiring with disjunction as addition, some decomposition being consistent, and conjunction as multiplication, all parts agreeing.
- inconsistent : Consistency
- consistent : Consistency
Instances For
Equations
- MarcolliChomskyBerwick2025.instDecidableEqConsistency x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
Equations
- One or more equations did not get rendered due to their size.
Disjunction, consistent iff at least one argument is.
Equations
Instances For
Conjunction, consistent iff both arguments are.
Equations
Instances For
Equations
- One or more equations did not get rendered due to their size.
The identity is a Rota–Baxter operator of weight +1 on Consistency, since disjunction is
idempotent and absorbs the weight term. On a Boolean target the threshold operator collapses to
the identity, disagreement already being the additive zero.
A syntactic object as an element of the Connes–Kreimer Hopf algebra over ℕ, the singleton
forest of its underlying nonplanar tree. The base ring is ℕ because every commutative
semiring, Consistency included, is an ℕ-algebra, while a Boolean target is no ℤ-algebra.
Equations
Instances For
Merge factors through the grafting operator on the disjoint union of its arguments, with the bare root label (Lemma 1.3.3).
The feature-consistency map φ₊ on a syntactic object, the value at the object of the
renormalized part of a feature character φ under an operator P on Consistency. It is the
single recursive map of §3.1.5 that checks consistency over all substructures.
Equations
Instances For
The feature-consistency map factors as the semiring Birkhoff convolution φ₊ = φ₋ ⋆ φ on
the object's Hopf-algebra image (Definition 3.1.6, Proposition 3.1.9), for a unital φ.
The head-probe value Υ_{s,h} on a tree (equation (3.2.1)), the probe Υ applied to the
tree's selection head, and inconsistent when the tree has no well-defined head.
Equations
Instances For
The head-following feature character ϕ_{Υ,s,h} of Lemma 3.2.5 is the algebra homomorphism
extending Υ_{s,h} multiplicatively to forests, so that a workspace is consistent iff each of its
trees is. It is the unrenormalized feature assignment whose Birkhoff renormalization is the
consistency map.
Equations
Instances For
The feature-consistency verdict on a syntactic object (§3.1.5, Lemmas 3.2.5 and 3.2.7), the Birkhoff renormalization with the identity operator of the head-following character, consistent iff the head-probe agreements cohere across all substructures.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The head-driven verdict factors as the semiring Birkhoff convolution φ₊ = φ₋ ⋆ φ of the
head-following character (Definition 3.1.6, Lemma 3.2.7).
Minimal Yield and the polar part of Laurent series (§3.5.2) #
δb₀ F F' = b₀ F − b₀ F', nonnegative iff F → F' does not diverge.
Equations
- MarcolliChomskyBerwick2025.δb₀ F F' = ↑(Multiset.card F) - ↑(Multiset.card F')
Instances For
δα F F' = α F' − α F, nonnegative iff F → F' loses no information.
Equations
- MarcolliChomskyBerwick2025.δα F F' = ↑(Multiset.map UnorderedTree.numEdges F').sum - ↑(Multiset.map UnorderedTree.numEdges F).sum
Instances For
δσ F F' = σ F' − σ F, equal to 1 iff F → F' has minimal yield.
Equations
- MarcolliChomskyBerwick2025.δσ F F' = ↑(Multiset.map UnorderedTree.numNodes F').sum - ↑(Multiset.map UnorderedTree.numNodes F).sum
Instances For
Weak Minimal Yield holds iff the δb₀- and δα-monomials of F → F' are both nonpolar.
The character ϕt of Proposition 3.5.3 for δ = δα, recording only the grading, sends a
tree T to t^{α(T)}, multiplicatively over disjoint union.
Equations
- MarcolliChomskyBerwick2025.gradingChar = ConnesKreimer.aeval fun (T : UnorderedTree α) => (HahnSeries.single ↑T.numEdges) 1
Instances For
ϕt(F) = t^{α(F)}, since α is additive over forests.
ϕt is nonpolar on every forest, since α(F) ≥ 0 (Lemma 3.5.5).
ϕt is its own renormalized part. Being nonpolar, it has no negative part, so the Birkhoff
factorization removes nothing and cannot separate Sideward Merge from Internal and External
Merge.
The polar projection is not multiplicative: t⁻² · t = t⁻¹ is polar while t is not, so a
product can acquire a polar part that its factors lack. Hence (1 − R) ∘ φ need not be a
character, and renormalization goes through the Birkhoff factorization.
The negative part of every character into Laurent series is polar on nonempty forests, landing
in the subalgebra R(A[t⁻¹][[t]]) of Proposition 3.1.7.
The renormalized part of every character into Laurent series is nonpolar, landing in the
subalgebra (1 − R)(A[t⁻¹][[t]]) = A[[t]] of Proposition 3.1.7: the first step of the proof of
Proposition 3.5.6.
Obligatory control by Form Copy (§3.8.2) #
The controller the man, from tokens 0 and 1.
Equations
- MarcolliChomskyBerwick2025.theMan = (↑(MarcolliChomskyBerwick2025.tok Minimalist.Cat.D [Minimalist.Cat.N] "the" 0)).merge ↑(MarcolliChomskyBerwick2025.tok Minimalist.Cat.N [] "man" 1)
Instances For
The controlled subject the man, from fresh tokens 4 and 5.
Equations
- MarcolliChomskyBerwick2025.theMan' = (↑(MarcolliChomskyBerwick2025.tok Minimalist.Cat.D [Minimalist.Cat.N] "the" 4)).merge ↑(MarcolliChomskyBerwick2025.tok Minimalist.Cat.N [] "man" 5)
Instances For
a book, the object of read.
Equations
Instances For
The syntactic object (3.8.4) of the man tried to read a book is built by External Merge alone. The controller merges into the theta position of tried and the controlled subject into that of read, since Internal Merge into a theta position would break the dichotomy of §3.8.1.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The two inscriptions of the man are repetitions, structurally identical distinct tokens.
Form Copy relates the controller to the controlled subject, the restriction of (3.8.4) to
the diagonal Diag₁,₄ of (3.8.5).
a book is no copy of the man, since Form Copy relates only structurally identical inscriptions.