OAI/Probability/DilutedSpin/PoissonVariance.lean
46 declarations, 0 contain sorry, 0 declared as axioms
- theorem
OAI.DilutedSpinGlass.physical_marked_palmline 17Source documentation
Palm identity for the actual finite selected-type marked reservoir statistic. No permutation or add-one property of the score is assumed. The base tower and base weight are unchanged on the right: all the other types remain present.
theorem physical_marked_palm {L : ℕ} (S : PrescribedTree L) (a : S.Leaf) (r : ℝ≥0) (P : FiniteLaw I) (T : KernelTower Ω L) (Q : Fin L → FiniteLaw A) (m : Fin L → ℝ) (base : FinitePath Ω L → ℝ) (factor numerator : I → FinitePath Ω L → FinitePath A L → ℝ) (f : (S.Leaf → FinitePath Ω L) → ℝ) {B : ℝ} (hf : ∀ x, |f x| ≤ B) (hn : ∀ i y z, |numerator i y z| ≤ 1) (hA : ∀ i y z, 1/2 ≤ factor i y z) : (∫ n : ℕ, (FiniteLaw.pi (fun _ : Fin n => P)).expect (fun roots => ∑ q, markedTreeScore S a T Q m base roots factor numerator f q) ∂poissonMeasure r) = (r : ℝ) * ∫ n : ℕ, P.expect (fun i => (FiniteLaw.pi (fun _ : Fin n => P)).expect (fun roots => markedTreeScore S a T Q m base (Fin.cons i roots) factor numerator f 0)) …Mentions: OAI.DilutedSpinGlass.PrescribedTree, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FinitePath, OAI.DilutedSpinGlass.FiniteLaw.pi, OAI.DilutedSpinGlass.q, OAI.DilutedSpinGlass.markedTreeScore
- def
OAI.DilutedSpinGlass.markPriorline 42Source documentation
An independent mark prior with no physical state.
noncomputable def markPrior : (L : ℕ) → (Fin L → FiniteLaw A) → KernelTower A L | 0, _ => () | L+1, Q => (Q 0, fun _ => markPrior L (fun j => Q j.succ)) /-- Move the distinguished Poisson-rooted mark out of the old mark row. -/
Mentions: OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.KernelTower
- def
OAI.DilutedSpinGlass.splitRowline 47Source documentation
Move the distinguished Poisson-rooted mark out of the old mark row.
def splitRow (k : ℕ) : (Ω × (Fin (k+1) → A)) ≃ ((Ω × (Fin k → A)) × A) where …
- theorem
OAI.DilutedSpinGlass.markedTower_splitline 55Source documentation
Counts n+1 in the Palm model have exactly the original n-mark law times one unused independent prior process. No term is removed from the base law.
theorem markedTower_split (k L : ℕ) (T : KernelTower Ω L) (Q : Fin L → FiniteLaw A) : KernelTower.transport (splitRow k) L (markedTower (k+1) L T Q) = KernelTower.prod L (markedTower k L T Q) (markPrior L Q) := by …Mentions: OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.KernelTower.transport, OAI.DilutedSpinGlass.splitRow, OAI.DilutedSpinGlass.markedTower, OAI.DilutedSpinGlass.KernelTower.prod, OAI.DilutedSpinGlass.markPrior
- theorem
OAI.DilutedSpinGlass.physicalPath_splitline 73 - theorem
OAI.DilutedSpinGlass.markPath_split_succline 84 - theorem
OAI.DilutedSpinGlass.markPath_split_zeroline 96 - def
OAI.DilutedSpinGlass.insertedMarkedLogline 109Source documentation
Logarithmic weight after one truly independent added factor on the base which still contains all original selected-type terms.
noncomputable def insertedMarkedLog (base : FinitePath Ω L → ℝ) (roots : Fin k → I) (factor : I → FinitePath Ω L → FinitePath A L → ℝ) (newFactor : FinitePath Ω L → FinitePath A L → ℝ) (y : FinitePath ((Ω × (Fin k → A)) × A) L) : ℝ := …Mentions: OAI.DilutedSpinGlass.FinitePath
- theorem
OAI.DilutedSpinGlass.insertedMarkedLog_splitline 117 - def
OAI.DilutedSpinGlass.markedRootline 132Source documentation
The actual quenched root free energy of the marked model.
noncomputable def markedRoot (T : KernelTower Ω L) (Q : Fin L → FiniteLaw A) (m : Fin L → ℝ) (base : FinitePath Ω L → ℝ) (roots : Fin k → I) (factor : I → FinitePath Ω L → FinitePath A L → ℝ) : ℝ := …Mentions: OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.FinitePath
- theorem
OAI.DilutedSpinGlass.markedRoot_cons_boundline 139Source documentation
Add/delete a marked Poisson point using a common product prior; the old mark row is not reweighted before the comparison.
theorem markedRoot_cons_bound (T : KernelTower Ω L) (Q : Fin L → FiniteLaw A) (m : Fin L → ℝ) (hm : ∀ j, 0 < m j) (base : FinitePath Ω L → ℝ) (roots : Fin k → I) (i : I) (factor : I → FinitePath Ω L → FinitePath A L → ℝ) {C : ℝ} (hfactor : ∀ x y, |Real.log (factor i x y)| ≤ C) : |markedRoot T Q m base (Fin.cons i roots) factor - markedRoot T Q m base roots factor| ≤ C := by …Mentions: OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.FinitePath, OAI.DilutedSpinGlass.markedRoot
- theorem
OAI.DilutedSpinGlass.markedRoot_base_stabilityline 161Source documentation
Replacing a physical root block acts only on the base log weight.
theorem markedRoot_base_stability (T : KernelTower Ω L) (Q : Fin L → FiniteLaw A) (m : Fin L → ℝ) (hm : ∀ j, 0 < m j) (base base' : FinitePath Ω L → ℝ) (roots : Fin k → I) (factor : I → FinitePath Ω L → FinitePath A L → ℝ) {C : ℝ} (hb : ∀ x, |base x - base' x| ≤ C) : |markedRoot T Q m base roots factor - markedRoot T Q m base' roots factor| ≤ C := by …Mentions: OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.FinitePath, OAI.DilutedSpinGlass.markedRoot
- theorem
OAI.DilutedSpinGlass.markedRoot_update_boundline 172Source documentation
One changed type/site label changes exactly one bounded log factor.
theorem markedRoot_update_bound (T : KernelTower Ω L) (Q : Fin L → FiniteLaw A) (m : Fin L → ℝ) (hm : ∀ j, 0 < m j) (base : FinitePath Ω L → ℝ) (roots : Fin k → I) (q : Fin k) (i : I) (factor : I → FinitePath Ω L → FinitePath A L → ℝ) {C : ℝ} (hf : ∀ x y, |Real.log (factor i x y) - Real.log (factor (roots q) x y)| ≤ C) : |markedRoot T Q m base (Function.update roots q i) factor - markedRoot T Q m base roots factor| ≤ C := by …Mentions: OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.FinitePath, OAI.DilutedSpinGlass.q, OAI.DilutedSpinGlass.markedRoot
- def
OAI.DilutedSpinGlass.rootArrayline 204Source documentation
The recursive independent-root product and the finite indexed root array are the same labeled data. No exchangeability is needed here.
def rootArray {I : Type*} : (n : ℕ) → RootPath I n → Fin n → I | 0, _ => Fin.elim0 | n+1, x => Fin.cons x.1 (rootArray n x.2)Mentions: OAI.DilutedSpinGlass.RootPath
- theorem
OAI.DilutedSpinGlass.rootArray_replaceline 208 - instance
OAI.DilutedSpinGlass.rootPathCountableline 225 - instance
OAI.DilutedSpinGlass.rootPathSingletonline 231 - theorem
OAI.DilutedSpinGlass.markedRoot_count_boundline 241Source documentation
A linear bound in the actual random point count, from the prior coupling.
theorem markedRoot_count_bound (T : KernelTower Ω L) (Q : Fin L → FiniteLaw A) (m : Fin L → ℝ) (hm : ∀ j, 0 < m j) (base : FinitePath Ω L → ℝ) (factor : I → FinitePath Ω L → FinitePath A L → ℝ) {C : ℝ} (hf : ∀ i x y, |Real.log (factor i x y)| ≤ C) (n : ℕ) (x : RootPath I n) : |markedRoot T Q m base (rootArray n x) factor| ≤ |markedRoot T Q m base (fun j : Fin 0 => j.elim0) factor| + (n : ℝ)*C := by …Mentions: OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.FinitePath, OAI.DilutedSpinGlass.RootPath, OAI.DilutedSpinGlass.markedRoot, OAI.DilutedSpinGlass.rootArray
- theorem
OAI.DilutedSpinGlass.markedRoot_poisson_varianceline 266Source documentation
Literal random-count marked-root concentration. The auxiliary rows change with the count, and their comparison is justified by markedRoot_cons_bound, not by an assumed concentration of an already averaged auxiliary score.
theorem markedRoot_poisson_variance [MeasurableSpace I] [Countable I] [MeasurableSingletonClass I] (μ : Measure I) [IsProbabilityMeasure μ] (r : ℝ≥0) (T : KernelTower Ω L) (Q : Fin L → FiniteLaw A) (m : Fin L → ℝ) (hm : ∀ j, 0 < m j) (base : FinitePath Ω L → ℝ) (factor : I → FinitePath Ω L → FinitePath A L → ℝ) {C : ℝ} (hC : 0 ≤ C) (hf : ∀ i x y, |Real.log (factor i x y)| ≤ C) : let F := fun n (x : RootPath I n) => markedRoot T Q m base (rootArray n x) factor …Mentions: OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.FinitePath, OAI.DilutedSpinGlass.RootPath, OAI.DilutedSpinGlass.markedRoot, OAI.DilutedSpinGlass.rootArray
- theorem
OAI.DilutedSpinGlass.FiniteLaw.measurable_expectline 305 - theorem
OAI.DilutedSpinGlass.FiniteLaw.measurable_logMeanline 310 - theorem
OAI.DilutedSpinGlass.KernelTower.measurable_backwardLogline 320 - theorem
OAI.DilutedSpinGlass.measurable_rootArrayline 334Source documentation
Coordinate evaluation is measurable for the genuine independent root product, including arbitrary (not necessarily discrete) interaction marks.
theorem measurable_rootArray {I : Type*} [MeasurableSpace I] (n : ℕ) (j : Fin n) : Measurable (fun x : RootPath I n => rootArray n x j) := by …Mentions: OAI.DilutedSpinGlass.RootPath, OAI.DilutedSpinGlass.rootArray
- theorem
OAI.DilutedSpinGlass.measurable_markedRootline 345 - theorem
OAI.DilutedSpinGlass.markedRoot_poisson_variance_measurableline 360Source documentation
Physical-root concentration for general measurable interaction labels. In particular no finite-support or countability condition is imposed on the bounded physical disorder by this application of the random-count bound.
theorem markedRoot_poisson_variance_measurable (μ : Measure I) [IsProbabilityMeasure μ] (r : ℝ≥0) (T : KernelTower Ω L) (Q : Fin L → FiniteLaw A) (m : Fin L → ℝ) (hm : ∀ j, 0 < m j) (base : FinitePath Ω L → ℝ) (factor : I → FinitePath Ω L → FinitePath A L → ℝ) (hfmeas : ∀ x y, Measurable (fun i => factor i x y)) {C : ℝ} (hC : 0 ≤ C) (hf : ∀ i x y, |Real.log (factor i x y)| ≤ C) : let F := fun n (x : RootPath I n) => markedRoot T Q m base (rootArray n x) factor …Mentions: OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.FinitePath, OAI.DilutedSpinGlass.RootPath, OAI.DilutedSpinGlass.markedRoot, OAI.DilutedSpinGlass.rootArray
- def
OAI.DilutedSpinGlass.poissonRootAverageline 400Source documentation
Root averaging followed by the genuine Poisson count average.
noncomputable def poissonRootAverage (r : ℝ≥0) (F : (n : ℕ) → RootPath Ω n → ℝ) : ℝ := …Mentions: OAI.DilutedSpinGlass.RootPath
- theorem
OAI.DilutedSpinGlass.measurable_poissonRootAverageline 406Source documentation
A parameter in a Poisson root average stays measurable, despite the count-dependent dimension of the actual root state.
theorem measurable_poissonRootAverage (r : ℝ≥0) {F : X → (n : ℕ) → RootPath Ω n → ℝ} (hF : ∀ n, Measurable (fun z : X × RootPath Ω n => F z.1 n z.2)) : Measurable (fun x => poissonRootAverage μ r (F x)) := by …Mentions: OAI.DilutedSpinGlass.RootPath, OAI.DilutedSpinGlass.poissonRootAverage
- theorem
OAI.DilutedSpinGlass.poisson_root_second_momentline 420Source documentation
The actual random-count squared displacement around any center. Both conditional root fluctuations and random-count fluctuations are retained.
theorem poisson_root_second_moment (r : ℝ≥0) {F : (n : ℕ) → RootPath Ω n → ℝ} {B : ℕ → ℝ} {C D : ℝ} (hF : ∀ n, Measurable (F n)) (hB : ∀ n x, |F n x| ≤ B n) (hD : 0 ≤ D) (hrep : ∀ n (i : Fin n) x y, |F n x-F n (replaceRoot n x i y)| ≤ C) (hadd : ∀ n x y, |F (n+1) (x,y)-F n y| ≤ D) (c : ℝ) : (∫ n : ℕ, ∫ x, (F n x-c)^2 ∂rootLaw n (fun _ => μ) ∂poissonMeasure r) ≤ (C^2/2+D^2)*(r : ℝ) + (poissonRootAverage μ r F-c)^2 := by …Mentions: OAI.DilutedSpinGlass.RootPath, OAI.DilutedSpinGlass.replaceRoot, OAI.DilutedSpinGlass.rootLaw, OAI.DilutedSpinGlass.poissonRootAverage
- theorem
OAI.DilutedSpinGlass.independent_block_poisson_varianceline 454Source documentation
Full physical-root variance, with an extra independent root block. It is not conditioned on the Poisson count or on the other root disorder.
theorem independent_block_poisson_variance (r : ℝ≥0) (ν : Measure X) [IsProbabilityMeasure ν] {F : X → (n : ℕ) → RootPath Ω n → ℝ} {B : ℕ → ℝ} {C D : ℝ} (hF : ∀ x n, Measurable (F x n)) (hB : ∀ x n y, |F x n y| ≤ B n) (hD : 0 ≤ D) (hrep : ∀ x n (i : Fin n) y z, |F x n y-F x n (replaceRoot n y i z)| ≤ C) (hadd : ∀ x n y z, |F x (n+1) (y,z)-F x n z| ≤ D) (hmean : MemLp (fun x => poissonRootAverage μ r (F x)) 2 ν) : (∫ x, ∫ n : ℕ, ∫ y, (F x n y-(∫ z, poissonRootAverage μ r (F z) ∂ν))^2 ∂rootLaw n (fun _ => μ) ∂poissonMeasure r ∂ν) ≤ (C^2/2+D^2)*(r : ℝ) + variance (fun x => poissonRootAverage μ r (F x)) ν := by …Mentions: OAI.DilutedSpinGlass.RootPath, OAI.DilutedSpinGlass.replaceRoot, OAI.DilutedSpinGlass.poissonRootAverage, OAI.DilutedSpinGlass.rootLaw
- theorem
OAI.DilutedSpinGlass.poissonRootAverage_boundline 480Source documentation
A linear root bound is sufficient for all count averages in the physical model; no bounded support for the Poisson count is introduced.
theorem poissonRootAverage_bound (r : ℝ≥0) {F : (n : ℕ) → RootPath Ω n → ℝ} {A D : ℝ} (hB : ∀ n x, |F n x| ≤ A+D*n) : |poissonRootAverage μ r F| ≤ A+D*(r : ℝ) := by …Mentions: OAI.DilutedSpinGlass.RootPath, OAI.DilutedSpinGlass.poissonRootAverage
- theorem
OAI.DilutedSpinGlass.poissonRootAverage_stabilityline 496Source documentation
Uniform stability passes through the entire marked-Poisson root law.
theorem poissonRootAverage_stability (r : ℝ≥0) {F G : (n : ℕ) → RootPath Ω n → ℝ} {A B D : ℝ} (hF : ∀ n, Measurable (F n)) (hG : ∀ n, Measurable (G n)) (hBF : ∀ n x, |F n x| ≤ A+B*n) (hBG : ∀ n x, |G n x| ≤ A+B*n) (hdiff : ∀ n x, |F n x-G n x| ≤ D) : |poissonRootAverage μ r F-poissonRootAverage μ r G| ≤ D := by …Mentions: OAI.DilutedSpinGlass.RootPath, OAI.DilutedSpinGlass.poissonRootAverage
- theorem
OAI.DilutedSpinGlass.markedRoot_uniform_boundline 526Source documentation
A bound for the root log partition uniform also in the physical base.
theorem markedRoot_uniform_bound (T : KernelTower Ω L) (Q : Fin L → FiniteLaw A) (m : Fin L → ℝ) (hm : ∀ j, 0 < m j) (base : FinitePath Ω L → ℝ) (factor : I → FinitePath Ω L → FinitePath A L → ℝ) {B C : ℝ} (hb : ∀ y, |base y| ≤ B) (hf : ∀ i x y, |Real.log (factor i x y)| ≤ C) (n : ℕ) (x : RootPath I n) : |markedRoot T Q m base (rootArray n x) factor| ≤ B+C*n := by …Mentions: OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.FinitePath, OAI.DilutedSpinGlass.RootPath, OAI.DilutedSpinGlass.markedRoot, OAI.DilutedSpinGlass.rootArray
- theorem
OAI.DilutedSpinGlass.measurable_markedRoot_baseline 549Source documentation
Joint measurability in physical fields and in the actual independent Poisson labels; no discrete-law assumption occurs.
theorem measurable_markedRoot_base (T : KernelTower Ω L) (Q : Fin L → FiniteLaw A) (m : Fin L → ℝ) (base : X → FinitePath Ω L → ℝ) (hb : ∀ y, Measurable (fun x => base x y)) (factor : I → FinitePath Ω L → FinitePath A L → ℝ) (hf : ∀ x y, Measurable (fun i => factor i x y)) (n : ℕ) : Measurable (fun z : X × RootPath I n => markedRoot T Q m (base z.1) (rootArray n z.2) factor) := by …Mentions: OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.FinitePath, OAI.DilutedSpinGlass.RootPath, OAI.DilutedSpinGlass.markedRoot, OAI.DilutedSpinGlass.rootArray
- theorem
OAI.DilutedSpinGlass.markedRoot_full_varianceline 564Source documentation
Unconditional root concentration for the actual marked model with an independent block of physical fields. The old and fresh mark priors change with the Poisson count and are not averaged before the log partition.
theorem markedRoot_full_variance (μ : Measure I) [IsProbabilityMeasure μ] (ν : Fin M → Measure X) [∀ i, IsProbabilityMeasure (ν i)] (r : ℝ≥0) (T : KernelTower Ω L) (Q : Fin L → FiniteLaw A) (m : Fin L → ℝ) (hm : ∀ j, 0 < m j) (base : RootPath X M → FinitePath Ω L → ℝ) (hbmeas : ∀ y, Measurable (fun x => base x y)) (factor : I → FinitePath Ω L → FinitePath A L → ℝ) (hfmeas : ∀ x y, Measurable (fun i => factor i x y)) {B C : ℝ} (hC : 0 ≤ C) (hb : ∀ x y, |base x y| ≤ B) (hf : ∀ i x y, |Real.log (factor i x y)| ≤ C) (E : Fin M → ℝ) …Mentions: OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.RootPath, OAI.DilutedSpinGlass.FinitePath
- def
OAI.DilutedSpinGlass.FiniteLaw.pointline 627 - theorem
OAI.DilutedSpinGlass.FiniteLaw.expect_pointline 633 - theorem
OAI.DilutedSpinGlass.FiniteLaw.logMean_pointline 637 - def
OAI.DilutedSpinGlass.FiniteLaw.uniformline 641 - theorem
OAI.DilutedSpinGlass.FiniteLaw.logMean_uniformline 646 - def
OAI.DilutedSpinGlass.KernelTower.terminalTowerline 661Source documentation
Auxiliary levels contain no physical spin randomness. The last level is exactly the Gibbs sum under its uniform spin prior.
noncomputable def terminalTower (a : Ω) (P : FiniteLaw Ω) : (r : ℕ) → KernelTower Ω (r+1) | 0 => (P, fun _ => ()) | r+1 => (FiniteLaw.point a, fun _ => terminalTower a P r)
Mentions: OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.KernelTower, OAI.DilutedSpinGlass.FiniteLaw.point
- def
OAI.DilutedSpinGlass.KernelTower.terminalStateline 665 - theorem
OAI.DilutedSpinGlass.KernelTower.backwardLog_terminalTowerline 669 - theorem
OAI.DilutedSpinGlass.backwardLog_marked_baseline 685 - theorem
OAI.DilutedSpinGlass.markedRoot_distance_baseline 709 - theorem
OAI.DilutedSpinGlass.backwardLog_physicalline 730 - theorem
OAI.DilutedSpinGlass.markedRoot_physical_distanceline 743Source documentation
Perturbing the actual finite Ising system changes its hierarchical root logarithm by the sum of the bounded perturbation log-factors.
theorem markedRoot_physical_distance {p N k n : ℕ} {A I : Type} [Fintype A] (theta : Fin k → InteractionSample p) (h : Fin N → ℝ) (indices : Fin k → Fin p → Fin N) (r : ℕ) (Q : Fin (r+1) → FiniteLaw A) (m : Fin (r+1) → ℝ) (hm : ∀ j, 0 < m j) (hend : m (Fin.last r) = 1) (roots : Fin n → I) (factor : I → FinitePath (Fin N → Spin) (r+1) → FinitePath A (r+1) → ℝ) {C : ℝ} (hf : ∀ i x y, |Real.log (factor i x y)| ≤ C) : |markedRoot (KernelTower.terminalTower (fun _ : Fin N => false) FiniteLaw.uniform r) Q m (fun y => logWeight theta h indices (KernelTower.terminalState r y)) roots factor +N*Real.log 2-logPartition theta h indices| ≤ C*n := by …Mentions: OAI.DilutedSpinGlass.InteractionSample, OAI.DilutedSpinGlass.FiniteLaw, OAI.DilutedSpinGlass.FinitePath, OAI.DilutedSpinGlass.Spin, OAI.DilutedSpinGlass.markedRoot, OAI.DilutedSpinGlass.KernelTower.terminalTower, OAI.DilutedSpinGlass.FiniteLaw.uniform, OAI.DilutedSpinGlass.logWeight, OAI.DilutedSpinGlass.KernelTower.terminalState, OAI.DilutedSpinGlass.logPartition