Docs / Studies / Studies.Maslova2000

A Dynamic Approach to the Verification of Distributional Universals

Maslova, E. (2000). Linguistic Typology, 4(3), 307–333

A type-shift process moves each language among the types of a typology, a language of type $T_k$ being of type $T_i$ after a small interval with probability $p(T_k \to T_i)$. Its distributional universal is its stationary law, the one law that independently changing languages keep once they reach it, and a law is stationary exactly when the languages leaving each type balance those entering it. A typology has a universal that gives every type a positive probability exactly when its types are strongly connected. Where the shifts of a coarser typology depend on its own types alone, summing the universal of a finer typology over each coarse type gives the coarser typology's own universal; for interdependent typologies the paper takes that sum as the definition. Of two types the more probable is the one the process leaves with the lower probability. Splitting and dying that ignore type keep the expected frequency of a type among the surviving languages at its ancestral frequency.

Implementation notes

  • A typology is a finite type with measurable singletons and its type-shift process a Markov kernel; typeShift P is the process with transition probabilities P. Strong connection is irreducibility with respect to the counting measure. The paper's small time interval, over which a language most likely keeps its type, is not imposed on the definitions; the processes absorbing and cycle keep each type with probability at least nine tenths.
  • That the shifts of a coarser typology depend on the coarse type alone is the condition κ (g s) = (κS s).map g on the coarsening g of a finer typology.
  • The birth-and-death process enters only through the law of the numbers of descendants of the $N_0$ ancestral languages, which need only be exchangeable; independent lineages make it so (integral_frequency_cond_pi). Where no descendant survives, the frequency is 0 / 0 = 0, and the expectation is taken given survival.
  • The paper obtains the expected frequency $f_0$ from the independence of splitting and dying from type, and (A10) displays it, for many ancestral languages, as the ratio of the expected numbers; here it holds exactly for any exchangeable law (integral_frequency_cond).

TODO

  • The independence of the initial type that the definition rests on: (κ ^ n) a tends to the universal for a strongly connected process that keeps each type with positive probability, and with it the tests that compare the distributions within genetic groupings and within areas.
  • The Feller–Arley process of the appendix, (A1) to (A9) and (A11) to (A12), and with it the variance estimates (4) to (6), Tables 1 to 3 and Figures 1 to 3.
  • The estimation of transition probabilities from genetic groupings of small time depth and from shifts in progress.

References

  • Maslova 2000, A Dynamic Approach to the Verification of Distributional Universals

Type-shift processes and distributional universals

L69–72
L74–76
L78–80
L82–85
L87–91

A law is stationary exactly when, for each type, the languages leaving the type balance the languages entering it.

theorem invariant_iff_flow(1)L93–99
{T : Type u_1} [MeasurableSpace T] [Fintype T] [MeasurableSingletonClass T] {κ : ProbabilityTheory.Kernel T T} {π : MeasureTheory.Measure T} [ProbabilityTheory.IsMarkovKernel κ] [MeasureTheory.IsFiniteMeasure π] :
κ.Invariant π ↔ ∀ (i : T), ∫⁻ (k : T) in {i}, (κ k) {i}ᶜ ∂π = ∫⁻ (k : T) in {i}ᶜ, (κ k) {i} ∂π

Strong connection holds exactly when the typology has a distributional universal and the universal gives every type a positive probability.

theorem isIrreducible_iff_exists_isDistributionalUniversalL101–109
{T : Type u_1} [MeasurableSpace T] [Fintype T] [MeasurableSingletonClass T] {κ : ProbabilityTheory.Kernel T T} [Nonempty T] [ProbabilityTheory.IsMarkovKernel κ] :
ProbabilityTheory.Kernel.IsIrreducible MeasureTheory.Measure.count κ ↔ ∃ (π : MeasureTheory.Measure T), IsDistributionalUniversal κ π ∧ ∀ (i : T), π {i} ≠ 0

Built from

When the shifts of a coarser typology T depend on the coarse type alone, its universal gives each coarse type the sum of what the universal of the finer typology S gives to the fine types inside it.

theorem IsDistributionalUniversal.apply_eq_sum(2)L112–125
{T : Type u_1} [MeasurableSpace T] [MeasurableSingletonClass T] {κ : ProbabilityTheory.Kernel T T} {π : MeasureTheory.Measure T} [DecidableEq T] {S : Type u_2} [MeasurableSpace S] [Fintype S] [MeasurableSingletonClass S] {κS : ProbabilityTheory.Kernel S S} {πS : MeasureTheory.Measure S} {g : S → T} (hg : ∀ (s : S), κ (g s) = MeasureTheory.Measure.map g (κS s)) (hS : IsDistributionalUniversal κS πS) (hT : IsDistributionalUniversal κ π) (t : T) :
π {t} = ∑ s : S with g s = t, πS {s}

Built from

Preferred types

Of two types, the more probable under a stationary law is the one the process leaves with the lower probability.

theorem lt_iff_of_invariantL129–156
{κ : ProbabilityTheory.Kernel Bool Bool} [ProbabilityTheory.IsMarkovKernel κ] {π : MeasureTheory.Measure Bool} [MeasureTheory.IsProbabilityMeasure π] (hπ : κ.Invariant π) (h : 0 < (κ false).real {true} + (κ true).real {false}) :
π.real {false} < π.real {true} ↔ (κ true).real {false} < (κ false).real {true}
L160–178

Three processes

L183–188
L199–200
L217–225
L227–230
L233–236
L247–248
L260–279

Birth and death

(A10): when languages split and die independently of their type, the expected frequency of a type among the descendants of N₀ ancestral languages, given that some descendant survives, is its frequency #A / N₀ among the ancestors.

theorem integral_frequency_condL283–295
{N₀ : ℕ} {μ : MeasureTheory.Measure (Fin N₀ → ℕ)} [MeasureTheory.IsProbabilityMeasure μ] (hμ : ProbabilityTheory.Exchangeable μ) (A : Finset (Fin N₀)) (hS : μ {x : Fin N₀ → ℕ | ∑ i : Fin N₀, ↑(x i) ≠ 0} ≠ 0) :
∫ (x : Fin N₀ → ℕ), (∑ i ∈ A, ↑(x i)) / ∑ i : Fin N₀, ↑(x i) ∂μ[|{x : Fin N₀ → ℕ | ∑ i : Fin N₀, ↑(x i) ≠ 0}] = ↑A.card / ↑N₀
L298–310