Inflation for the classical triangle: definitions #
Definitions for the manuscript Inflation for the Classical Triangle: Nontermination and
Quantitative Complexity (papers/inflation-nontermination). This file carries definitions
only; the statements live in Finner.lean, Defect.lean, Main.lean, Fan.lean,
Exponent.lean and Rate.lean.
Representational decisions #
Laws are bare real weight functions. A law on a
Fintypeαis a functionα → ℝtogether with the predicateIsLaw(nonnegative, sums to1). The unbundled form is used here because the paper's familiesQ(ε,r),P_ε,R_pare written for parameter ranges in which normalization is a hypothesis of the statement rather than part of the datum, and because the feasibility definitions quantify over witnesses on large index types where carrying proof fields through every statement is noise. No MathlibMeasureorPMFis used; all inequalities are real-valued.Latent alphabets are finite.
TriangleCompatiblequantifies over aTriangleModel, whose three latent spaces are arbitraryFintypes. The paper (Section 2) allows arbitrary measurable latent spaces. Restricting to finite latent alphabets is a formalization boundary; it is justified by the theorem of Rosset, Gisin and Wolfe (2018) that latent alphabets of bounded finite size suffice for the triangle. That theorem is not formalized here. ConsequentlyTriangleCompatibleis a subset of the paper'sC_tri, and a theorem¬ TriangleCompatible Pis the WEAKER statement: nonmembership in the finite-latent set does not by itself give nonmembership in the arbitrary-latent set. The arbitrary-latent statement needs either the Rosset–Gisin–Wolfe reduction (quoted, not formalized) or a direct measure-theoretic proof of Lemma 5.7; seeFinnerMeasure.leanfor the latter where it exists.The recursively expressible set is not formalized. Definition 2.3 of the paper (Definition 2.5,
def:gexp, theI^exp_thierarchy) needsd-separation in the inflated causal graph and the recursion of Wolfe–Spekkens–Fritz Definition 7. That machinery is deliberately out of scope here, so the membershipQ(ε,r) ∈ I^exp_tof Theorem 5.1 (Lemma 5.10,lem:expressible) is an unformalized boundary. Only the two weaker hierarchiesI^NW_t(NWFeasible) andI^AI_t(AIFeasible) are defined, and every statement that the paper phrases for all three hierarchies is formalized for those two. SinceI^exp_t ⊆ I^AI_t ⊆ I^NW_t, the membership theorems formalized here are the weaker halves of the paper's claims and the rejection theorems are the stronger halves.Bit convention. Outcomes are
Bool;falseis the paper's0andtrueis its1. SoBern(r)hasBern(r)(true) = r, matchingBern(r)(1) = r.Truncated subtraction. Exponents
t - 1areℕsubtraction; every statement using them carries the hypothesis1 ≤ t, where the two agree.
Laws as weight functions #
Pushforward of a weight function along a map of finite types.
Instances For
Three-bit laws #
The outcome type of the triangle: the three observed bits (A, B, C).
false is the paper's outcome 0.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
The bit that a party reads off a three-bit outcome.
Equations
Instances For
The copied observations of the order-t inflation #
Paper Section 2.3. A^{ij} has X-index i and Z-index j; B^{ik} has X-index i
and Y-index k; C^{jk} has Z-index j and Y-index k.
Equations
- TriangleInflation.instDecidableEqObs.decEq (TriangleInflation.Obs.A a a_1) (TriangleInflation.Obs.A b b_1) = if h : a = b then h ▸ if h : a_1 = b_1 then h ▸ isTrue ⋯ else isFalse ⋯ else isFalse ⋯
- TriangleInflation.instDecidableEqObs.decEq (TriangleInflation.Obs.A i j) (TriangleInflation.Obs.B i_1 k) = isFalse ⋯
- TriangleInflation.instDecidableEqObs.decEq (TriangleInflation.Obs.A i j) (TriangleInflation.Obs.C j_1 k) = isFalse ⋯
- TriangleInflation.instDecidableEqObs.decEq (TriangleInflation.Obs.B i k) (TriangleInflation.Obs.A i_1 j) = isFalse ⋯
- TriangleInflation.instDecidableEqObs.decEq (TriangleInflation.Obs.B a a_1) (TriangleInflation.Obs.B b b_1) = if h : a = b then h ▸ if h : a_1 = b_1 then h ▸ isTrue ⋯ else isFalse ⋯ else isFalse ⋯
- TriangleInflation.instDecidableEqObs.decEq (TriangleInflation.Obs.B i k) (TriangleInflation.Obs.C j k_1) = isFalse ⋯
- TriangleInflation.instDecidableEqObs.decEq (TriangleInflation.Obs.C j k) (TriangleInflation.Obs.A i j_1) = isFalse ⋯
- TriangleInflation.instDecidableEqObs.decEq (TriangleInflation.Obs.C j k) (TriangleInflation.Obs.B i k_1) = isFalse ⋯
- TriangleInflation.instDecidableEqObs.decEq (TriangleInflation.Obs.C a a_1) (TriangleInflation.Obs.C b b_1) = if h : a = b then h ▸ if h : a_1 = b_1 then h ▸ isTrue ⋯ else isFalse ⋯ else isFalse ⋯
Instances For
Equations
- TriangleInflation.instFintypeObs = Fintype.ofEquiv ((_ : Fin t✝) × Fin t✝ ⊕ (_ : Fin t✝) × Fin t✝ ⊕ (_ : Fin t✝) × Fin t✝) (TriangleInflation.Obs.proxyTypeEquiv t✝)
A deterministic assignment of the copied observations, an element of {0,1}^{3t²}.
Equations
Instances For
Which party a copied observation is a copy of.
Equations
Instances For
Equations
- TriangleInflation.instDecidableEqLatent.decEq (TriangleInflation.Latent.X a) (TriangleInflation.Latent.X b) = if h : a = b then h ▸ isTrue ⋯ else isFalse ⋯
- TriangleInflation.instDecidableEqLatent.decEq (TriangleInflation.Latent.X i) (TriangleInflation.Latent.Z j) = isFalse ⋯
- TriangleInflation.instDecidableEqLatent.decEq (TriangleInflation.Latent.X i) (TriangleInflation.Latent.Y k) = isFalse ⋯
- TriangleInflation.instDecidableEqLatent.decEq (TriangleInflation.Latent.Z j) (TriangleInflation.Latent.X i) = isFalse ⋯
- TriangleInflation.instDecidableEqLatent.decEq (TriangleInflation.Latent.Z a) (TriangleInflation.Latent.Z b) = if h : a = b then h ▸ isTrue ⋯ else isFalse ⋯
- TriangleInflation.instDecidableEqLatent.decEq (TriangleInflation.Latent.Z j) (TriangleInflation.Latent.Y k) = isFalse ⋯
- TriangleInflation.instDecidableEqLatent.decEq (TriangleInflation.Latent.Y k) (TriangleInflation.Latent.X i) = isFalse ⋯
- TriangleInflation.instDecidableEqLatent.decEq (TriangleInflation.Latent.Y k) (TriangleInflation.Latent.Z j) = isFalse ⋯
- TriangleInflation.instDecidableEqLatent.decEq (TriangleInflation.Latent.Y a) (TriangleInflation.Latent.Y b) = if h : a = b then h ▸ isTrue ⋯ else isFalse ⋯
Instances For
Equations
The copied latent ancestors of a copied observation:
A^{ij} ↦ {X_i, Z_j}, B^{ik} ↦ {X_i, Y_k}, C^{jk} ↦ {Z_j, Y_k}.
Equations
- (TriangleInflation.Obs.A a a_1).ancestors = {TriangleInflation.Latent.X a, TriangleInflation.Latent.Z a_1}
- (TriangleInflation.Obs.B a a_1).ancestors = {TriangleInflation.Latent.X a, TriangleInflation.Latent.Y a_1}
- (TriangleInflation.Obs.C a a_1).ancestors = {TriangleInflation.Latent.Z a, TriangleInflation.Latent.Y a_1}
Instances For
The copied latent ancestors of a set of copied observations.
Instances For
Paper Definition 2.4: two sets of copied observations are ancestrally independent when their copied latent ancestors are disjoint.
Equations
Instances For
The copied triangle Δ_{ijk} = {A^{ij}, B^{ik}, C^{jk}} (paper Section 2.3).
Equations
Instances For
The diagonal triangle Δ_{lll} (paper Section 2.3).
Equations
Instances For
The three-bit outcome that an assignment gives to the copied triangle Δ_{ijk}.
Equations
- TriangleInflation.readTriangle i j k ω = (ω (TriangleInflation.Obs.A i j), ω (TriangleInflation.Obs.B i k), ω (TriangleInflation.Obs.C j k))
Instances For
The tuple of three-bit outcomes of the t diagonal triangles (Δ_{lll})_{l ∈ [t]}.
Equations
Instances For
The symmetry group #
The action of (σ_X, σ_Z, σ_Y) ∈ S_t × S_t × S_t on copied observations. The first
component permutes X-copy indices, the second Z-copy indices, the third Y-copy
indices, exactly as in Definition 2.3(i).
Equations
- TriangleInflation.Obs.perm π (TriangleInflation.Obs.A a a_1) = TriangleInflation.Obs.A (π.1 a) (π.2.1 a_1)
- TriangleInflation.Obs.perm π (TriangleInflation.Obs.B a a_1) = TriangleInflation.Obs.B (π.1 a) (π.2.2 a_1)
- TriangleInflation.Obs.perm π (TriangleInflation.Obs.C a a_1) = TriangleInflation.Obs.C (π.2.1 a) (π.2.2 a_1)
Instances For
Relabelling an assignment by an index permutation.
Equations
- TriangleInflation.relabel π ω v = ω (TriangleInflation.Obs.perm π v)
Instances For
Paper Definition 2.3(i): invariance under independent permutations of the three index families.
Equations
- TriangleInflation.SymmetricLaw t Γ = ∀ (π : Equiv.Perm (Fin t) × Equiv.Perm (Fin t) × Equiv.Perm (Fin t)) (ω : TriangleInflation.Assign t), Γ (TriangleInflation.relabel π ω) = Γ ω
Instances For
Injectable sets #
Paper Definition 2.4 and Appendix A.
The primitive definition of an injectable set (paper Appendix A.3).
Equations
- TriangleInflation.InjectableRaw S = ((∀ u ∈ S, ∀ v ∈ S, u.party = v.party → u = v) ∧ ∀ u ∈ S, ∀ v ∈ S, TriangleInflation.SharedAgree u v)
Instances For
Paper Definition 2.4 together with Appendix A: for the triangle the injectable sets are
exactly the subsets of copied triangles. This characterization is the working definition
here; Main.injectable_iff_injectableRaw records its equivalence with InjectableRaw.
Equations
- TriangleInflation.Injectable S = ∃ (i : Fin t) (j : Fin t) (k : Fin t), S ⊆ TriangleInflation.copiedTriangle i j k
Instances For
The restriction of an assignment to a set of copied observations.
Equations
- TriangleInflation.restrictAssign S ω v = ω ↑v
Instances For
The outcome pattern that a three-bit law prescribes on an injectable set: each member reads the bit of the party it is a copy of.
Equations
- TriangleInflation.partyRead S w v = TriangleInflation.partyBit (↑v).party w
Instances For
The defect cube (paper Section 5.1) #
Cell t indexes the cube [t]³ of defect bits D_{ijk}; Obs t indexes the private bits
N_v. The root bits are indexed by Root t = Cell t ⊕ Obs t.
The root bits of the defect cube: defects on cells, private bits on observations.
Equations
Instances For
onLine v c says the cell c lies on the line Λ(v) read by the observation v:
Λ(A^{ij}) = {(i,j,k) : k ∈ [t]}, Λ(B^{ik}) = {(i,j,k) : j ∈ [t]},
Λ(C^{jk}) = {(i,j,k) : i ∈ [t]} (paper Section 5.1).
Equations
- TriangleInflation.onLine (TriangleInflation.Obs.A i j) x✝ = (x✝.1 == i && x✝.2.1 == j)
- TriangleInflation.onLine (TriangleInflation.Obs.B i k) x✝ = (x✝.1 == i && x✝.2.2 == k)
- TriangleInflation.onLine (TriangleInflation.Obs.C j k) x✝ = (x✝.2.1 == j && x✝.2.2 == k)
Instances For
The output map of paper equation (eq:outputs): an observation outputs 1 exactly when
its private bit and every defect on its line are 0.
Equations
- TriangleInflation.outputs d N v = decide (N v = false ∧ ∀ (c : TriangleInflation.Cell t), TriangleInflation.onLine v c = true → d c = false)
Instances For
The output map read off a joint root assignment.
Equations
- TriangleInflation.outputsOf x = TriangleInflation.outputs (fun (c : TriangleInflation.Cell t) => x (Sum.inl c)) fun (v : TriangleInflation.Obs t) => x (Sum.inr v)
Instances For
The root bits that a set of copied observations reads: the defect cells on their lines, together with their own private bits (paper Lemma 5.4, "disjoint inputs").
Equations
- TriangleInflation.rootSupport S = Finset.image Sum.inl {c : TriangleInflation.Cell t | ∃ v ∈ S, TriangleInflation.onLine v c = true} ∪ Finset.image Sum.inr S
Instances For
R_l, the union of the three lines of the diagonal triangle Δ_{lll}: the cells with
at least two coordinates equal to l (paper equation (eq:Rl)).
Equations
- One or more equations did not get rendered due to their size.
Instances For
The paper's explicit laws #
Paper equation (eq:Q): Q(ε,r) = ε δ_{000} + (1-ε) Bern(r)^{⊗3}.
Equations
Instances For
ε_t = 1/(2t³) (paper Theorem 5.2).
Equations
- TriangleInflation.epsFam t = 1 / (2 * ↑t ^ 3)
Instances For
r_t = (1 - ε_t)^{t-1} (paper Theorem 5.2).
Equations
- TriangleInflation.rFam t = (1 - TriangleInflation.epsFam t) ^ (t - 1)
Instances For
The nontermination family P_t = Q(ε_t, r_t) (paper Theorem 5.2).
Equations
Instances For
The divergent-rejecting-order family P_ε = Q(ε, 1 - ε^{2/3}/2) (paper Proposition
5.13).
Equations
- TriangleInflation.Peps ε = TriangleInflation.Q ε (1 - ε ^ (2 / 3) / 2)
Instances For
Triangle compatibility #
A triangle model with finite latent alphabets: three latent spaces with weight
functions, and the three response probabilities f(x,z) = Pr(A = 0 | x,z),
g(x,y) = Pr(B = 0 | x,y), h(z,y) = Pr(C = 0 | z,y).
- X : Type
Latent alphabet shared by parties A and B.
- Y : Type
Latent alphabet shared by parties B and C.
- Z : Type
Latent alphabet shared by parties A and C.
Probability weight of the source shared by A and B.
Probability weight of the source shared by B and C.
Probability weight of the source shared by A and C.
Conditional probability of outcome zero at A, given sources X and Z.
Conditional probability of outcome zero at B, given sources X and Y.
Conditional probability of outcome zero at C, given sources Z and Y.
Instances For
A triangle model is valid when the three source weights are laws and the three response
probabilities take values in [0,1].
Equations
- One or more equations did not get rendered due to their size.
Instances For
The observed law of a triangle model: the sources are independent and the responses are conditionally independent given the sources (paper Section 2.3).
Equations
- One or more equations did not get rendered due to their size.
Instances For
Paper Section 2.3: the triangle-compatible set C_tri, with the finite-latent-alphabet
formalization boundary described in the file header.
Equations
- TriangleInflation.TriangleCompatible P = ∃ (M : TriangleInflation.TriangleModel), M.Valid ∧ M.law = P
Instances For
The finite inflation tests #
Paper Definition 2.3: the Navascués–Wolfe feasible set I^NW_t. A law Γ_t on the
copied observations, invariant under S_t³, whose diagonal law is the tensor power
P^{⊗t}.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The injectable-marginal prescriptions of paper Definition 2.4: every injectable set has
the corresponding marginal of P.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The ancestral-independence prescriptions of paper Definition 2.4: every union of pairwise ancestrally independent injectable sets has the product of the corresponding marginals. The union is presented as a finite family, whose joint restriction law is required to factor.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The defect-cube witness #
The per-root Bernoulli weights: defects are Bern(ε), private bits are Bern(s).
Equations
- TriangleInflation.rootWeight t ε s (Sum.inl val) = TriangleInflation.bern ε
- TriangleInflation.rootWeight t ε s (Sum.inr val) = TriangleInflation.bern s
Instances For
The joint law of the independent root bits of the defect cube.
Equations
Instances For
Paper Section 5.1: the defect-cube law Γ_t, the pushforward of the independent defect
and private bits under the output map (eq:outputs). Unfolding pushforward and prodLaw
gives the explicit finite sum of product weights of paper equation (eq:table):
Γ_t(w) = ∑_{x : outputsOf x = w} ∏_{roots} bern _ (x _).
Equations
Instances For
First rejecting order #
Paper Section 2.3: t_min^H(P) = min {t ≥ 1 : P ∉ I^H_t}. Formalized as Nat.sInf, which
returns the junk value 0 when the set is empty. The set is nonempty exactly when some
finite order rejects P; that this happens for every incompatible P is the asymptotic
completeness of the hierarchy, which is quoted from Navascués–Wolfe in the paper and is not
formalized here. Every statement about t_min below therefore either exhibits a rejecting
order or assumes one.