Theorem 2, pp. 8β9 β on graph sequences admitting a strong coupling, the typical stationary opinion converges in conditional law to the root opinion π‘* of the local limit
ProvedDirectedOpinion.Typical.theorem_2Consider the random directed graph sequence with the signals and initial opinions of the setting, and suppose that it admits a strong coupling (Definition 4) with a marked directed graph . Assume and that the conditional law of the external signals given the vertex marks satisfies
for some constant ; assume the same of the kernel of the initial opinions. Let be uniformly chosen in . Then there are laws on , , and on , none depending on , such that, writing and :
- for every ,
for every continuous , where uses the sup distance on ; 2. for the stationary vector of each graph,
for every continuous ; 3. at a geometric rate:
Items 1 and 2 are the paper's statements that and for random variables whose law does not depend on . is the "typical" stationary opinion of a large graph: the root opinion of the local limit. By (3), is the conditional law of given the graph, so item 2 says that the empirical distribution of stationary opinions does not depend, asymptotically, on the realization of the graph.
Formalization Note
- Law form. "There exist , whose law does not depend on with " is stated as convergence in probability of the distance between the regular conditional law given and a fixed law. The two are equivalent: a coupling bounds , and conversely a conditionally optimal coupling exists.
- Stationary vector. "A vector having the stationary distribution" is the paper's own choice (p. 27), whose law is the stationary law by Theorem 1.
- Initial opinions (disclosed addition). The page requires only that be conditionally independent of the signals. Item 1 at is false for an arbitrary (e.g. ), and the proof (p. 26) copies from the graph onto the limit. The statement therefore assumes with i.i.d. uniform and -Lipschitz with the same . This covers the paper's simulations and constant initial vectors. Items 2 and 3 do not involve the initial opinions of the graph.
- Rate. The page says only "at a geometric rate"; the constant is the one its proof gives through Theorem 1 on (pp. 23, 28), and is the metric used for weak convergence on .
- is any function continuous on ; all opinions lie in . The -th graph of the sequence has vertices; the limit is unaffected.
- Disjoint in-components (added). The strong coupling of the setting requires the depth- in-components of the two uniform vertices to be disjoint on the coupling event, which the proof (p. 25) uses and Definition 4 does not state; without it the empirical-average claims are false (a graph whose vertices all listen to one hub, whose signals then shift every opinion by a common random amount). See the Setting.
import Mathlib import Definitions.Def_DirectedOpinion_Typical_GraphSeq
namespace DirectedOpinion.Typical
open MeasureTheory ProbabilityTheory Set Filter Topology
/-- Theorem 2, pp. 8β9, in law form. Under a strong coupling with a limit graph of law `ΞΌlim`,
with `d > 0`, a signal kernel `Ξ½(x) = law ΞΆ(x, U)` that is `dβ`-Lipschitz in the mark with
constant `K`, and initial opinions `R_i^(0) = Ο(X_i, V_i)` whose kernel is `dβ`-Lipschitz with the
same `K`, there are laws `L_k` on `β^{k+1}` (`k β₯ 0`) and `L*` on `β`, not depending on `m`, with:
1. for every `k`: the `dβ` distance (sup cost) between the conditional law given `π_m` of
`(R_{I_m}^(0), β¦, R_{I_m}^(k))` and `L_k` tends to `0` in probability, and for every `f`
continuous on `[β1, 1]`, `(1/n) Ξ£_i f(R_i^(k)) β β« f(y_k) dL_k(y)` in probability;
2. for the stationary vector `R = Ξ£_r ΞΜ^r W^(r)`: the `dβ` distance between the conditional law
given `π_m` of `R_{I_m}` and `L*` tends to `0` in probability, and for every `f` continuous on
`[β1, 1]`, `(1/n) Ξ£_i f(R_i) β β« f dL*` in probability;
3. the time-`k` marginal of `L_k` converges to `L*` at a geometric rate:
`dβ(L_k β y_kβ»ΒΉ, L*) β€ (1 β d)^k (2/d + 2)`. -/
theorem theorem_2 {Ξ© A : Type*} [MeasurableSpace Ξ©] [MetricSpace A] [MeasurableSpace A]
[BorelSpace A] [PolishSpace A] (P : Measure Ξ©) [IsProbabilityMeasure P] (S : GraphSeq Ξ© A)
(π : β β MeasurableSpace Ξ©) (ΞΆ Ο : Mark A β β β β) (c d K : β)
(ΞΌlim : Measure (LimGraph A)) (hS : GraphSeqModel P S π ΞΆ Ο c d)
(hSC : StrongCoupling P S π ΞΌlim) (hd : 0 < d) (hK : KernelLipschitz ΞΆ K)
(hK0 : KernelLipschitz Ο K) :
β (L : (k : β) β ProbabilityMeasure (Fin (k + 1) β β)) (Lstar : ProbabilityMeasure β),
(β k : β,
(β Ξ· : β, 0 < Ξ· β
Tendsto (fun m => P {Ο | ENNReal.ofReal Ξ· β€
W1vec (trajCondLaw P S ΞΆ Ο c d m k Ο) (L k : Measure (Fin (k + 1) β β))})
atTop (π 0)) β§
β f : β β β, ContinuousOn f (Icc (-1 : β) 1) β
TendstoInMeasure P (fun m Ο => empAvg f (S.R ΞΆ Ο c d m k Ο)) atTop
(fun _ => β« y, f (y (Fin.last k)) β(L k : Measure (Fin (k + 1) β β)))) β§
((β Ξ· : β, 0 < Ξ· β
Tendsto (fun m => P {Ο | ENNReal.ofReal Ξ· β€
W1 (statCondLaw P S ΞΆ c d m Ο) (Lstar : Measure β)})
atTop (π 0)) β§
β f : β β β, ContinuousOn f (Icc (-1 : β) 1) β
TendstoInMeasure P (fun m Ο => empAvg f (S.Rstat ΞΆ c d m Ο)) atTop
(fun _ => β« y, f y β(Lstar : Measure β))) β§
β k : β,
W1 ((L k : Measure (Fin (k + 1) β β)).map (fun y => y (Fin.last k))) (Lstar : Measure β) β€
ENNReal.ofReal ((1 - d) ^ k * (2 / d + 2)) := by sorry
end DirectedOpinion.Typical
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.