Minimal Search as a weighting of Merge #
Marcolli, Chomsky and Berwick implement Minimal Search by weighting the terms of the coproduct
(their §1.5). Extracting an accessible term at distance d from its root costs ε^d, the costs
of simultaneous extractions multiply, and Merge pays the weight of the terms it grafts, their
M^ε_{S,S'} = ⊔ ∘ (B^ε ⊗ id) ∘ δ_{S,S'} ∘ Δ of (1.5.1) and (1.5.2).
The weighted coproduct comulEps ε scales a tree by ε for each level it is moved down on its
way to the left channel. At ε = 1 it is the pruning coproduct, so the weighted Merge
mergeOpEps 1 is Merge. At ε = 0 only whole components reach the left channel and the coproduct
is primitive, so mergeOpEps 0 merges two components of the workspace and nothing else. At every
ε, weighted Merge agrees with its value at ε = 0 up to a multiple of ε, and External and
Internal Merge have weight 1 (Proposition 1.5.1): they cost nothing, every Sideward term costs at
least one, and in the limit only External and Internal Merge remain.
Main definitions #
Minimalist.Merge.comulEps: the depth-weighted coproduct.Minimalist.Merge.mergeOpEps: Merge weighted by Minimal Search.
Main results #
Minimalist.Merge.comulEps_one,comulEps_zero: atε = 1the weighted coproduct is the pruning coproduct, and atε = 0it is the primitive coproduct.Minimalist.Merge.mergeOpEps_pair,mergeOpEps_im_composition: External and Internal Merge have weight1at everyε.Minimalist.Merge.mergeOpEps_eq_add_smul,mergeOpEps_of'_eq_smul_of_not_le: weighted Merge is its value atε = 0plus a multiple ofε, so every form of Sideward Merge is a multiple ofε.Minimalist.Merge.mergeOpEps_zero_of'_of_not_le,mergeOpEps_zero_pair_add: atε = 0, Merge vanishes unless both objects are components, and External Merge survives beside any spectators that do not repeat its objects.
Implementation notes #
- The book weights the left channel by
ε^dand the right channel byε^{-d}, and letsB^εread the left weight. In a tensor product overRthe two weights would cancel, so the weight is the coefficient of the left channel only. The book uses the right weight only to compute the costd - d = 0of Internal Merge, which here is the cost0of merging two components. - The unit stage
M_{β,1}grafts nothing and costs nothing (the book's rule 4), so Internal Merge ismergeOpEps ε lbl Q β ∘ mergeOpUnit β, as the book writes it in proving Proposition 1.5.1. - The weights are polynomial in
ε, and the limitε → 0is evaluation atε = 0, the book'sM^ε_{S,S'}|_{ε=0}.
References #
The empty-cut enumeration #
The weighted coproduct #
The step of the fold computing the weighted coproduct pairs the node rebuilt from its children with its weighted coproduct.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The depth-weighted coproduct of a tree ([MCB25] §1.5.2), in which an
extraction at distance d from the root has weight ε^d.
Equations
Instances For
The weighted coproduct of a node keeps the node whole, or takes the weighted coproduct of its children with every extracted tree one level further down.
The depth-weighted coproduct on workspaces.
Equations
Instances For
At ε = 1 every extraction has weight 1, and the weighted coproduct is the pruning
coproduct.
At ε = 0 every tree is primitive for the weighted coproduct, since only whole components reach
the left channel.
At ε = 0 the weighted coproduct is the primitive coproduct.
The weighted coproduct modulo ε #
Every term of the weighted coproduct but those of ε = 0 extracts a tree and so carries a factor
of ε, and every term but T ⊗ 1 has a left factor with fewer vertices than T.
The weighted coproduct agrees with its value at ε = 0 modulo ε.
Weighted Merge #
Merge weighted by Minimal Search ([MCB25] (1.5.1)) grafts each matching term with the weight of its extractions.
Equations
- Minimalist.Merge.mergeOpEps ε lbl S S' = Minimalist.Merge.mergePost lbl S S' ∘ₗ (Minimalist.Merge.comulEps ε).toLinearMap
Instances For
At ε = 1 weighted Merge is Merge.
At ε = 0 weighted Merge is Merge at the empty-cut enumeration.
At ε = 0 Merge vanishes on a workspace unless S and S' are both components of it
([MCB25] Proposition 1.5.1). This covers each form of Sideward Merge,
in which S or S' is an accessible term of a component.
At ε = 0 Merge commutes with a spectator tree U other than S and S', whatever
accessible terms U has.
At ε = 0 a spectator workspace passes through Merge whenever neither S nor S' is one of
its components.
External Merge survives at ε = 0 ([MCB25] Proposition 1.5.1), beside
any spectator workspace that does not repeat S or S' as a component.
Weighted Merge agrees with its value at ε = 0 modulo ε, so External Merge and Internal
Merge, which survive at ε = 0, are its leading terms ([MCB25]
Proposition 1.5.1).
Unless S and S' are both components of the workspace, weighted Merge is a multiple of ε,
so every form of Sideward Merge costs at least one ([MCB25]
Proposition 1.5.1).
External Merge has weight 1 at every ε ([MCB25]
Proposition 1.5.1), since its two objects are taken whole, at no cost.
Internal Merge has weight 1 at every ε ([MCB25]
Proposition 1.5.1), since its unit stage is weightless and its second stage merges two
components.