Stat.Minimax.Mixture.Poisson­Latent­Sign

This module collects reusable likelihood, coefficient, and cross-moment results for a Poisson-sized cell whose records share a symmetric latent sign.

Pointwise 10 core · 12 supporting The latent sign is mixed once after multiplying all record factors. ★ subsetEnergy_one_bound

One shared latent sign in a finite cell

The latent sign is mixed once after multiplying all record factors. Cells carry bounded real scores, Boolean assignments, and an optional Boolean thinning of the positive assignments. Posterior coefficients are defined for arbitrary subsets; cancellation is stated only for subsets of positive assignments.

abbrev Cell reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

A cell of the given size consists of scores, assignments, and selection flags. The product measurable space allows arbitrary conditional laws of these data.

Definition (Lean source)
m :
ℕ
Cell m :
Type
(Fin m → ℝ) × ((Fin m → Bool) × (Fin m → Bool))
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.Cell · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:28
def sign reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The Boolean latent or assignment sign is interpreted as plus or minus one.

Definition (Lean source)
b :
sign b :
ℝ
if b then 1 else -1
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.sign · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:32
def active reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The active set contains precisely the positively assigned records.

Definition (Lean source)
m :
ℕ
c :
active c :
Finset.univ.filter (fun i => c.2.1 i = true)
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.active · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:35 · uses Cell
def selected reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The selected set contains precisely the records retained by the selection flags.

Definition (Lean source)
m :
ℕ
c :
selected c :
Finset.univ.filter (fun i => c.2.2 i = true)
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.selected · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:39 · uses Cell
def Valid reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

Valid cell data have scores bounded by one and selection contained in the active set.

Definition (Lean source)
m :
ℕ
c :
Valid c :
Prop
clause 1
i :
|c.1 i| ≤ 1
clause 2
selected c ⊆ active c
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.Valid · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:43 · uses Cell
def signLikelihood reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The sign likelihood at a latent sign ℓ is the product over all records of 1 + τ · ℓ · uᵢ · sᵢ, where uᵢ is the record's score and sᵢ = ±1 its assignment sign.

Definition (Lean source)
m :
ℕ
τ :
ℝ
c :
ell :
ℝ
signLikelihood τ c ell :
ℝ
∏ i, (1 + τ * ell * c.1 i * sign (c.2.1 i))
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.signLikelihood · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:47 · uses Cell
def denominator reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The posterior denominator is the sum of the likelihoods at the two latent signs.

Definition (Lean source)
m :
ℕ
τ :
ℝ
c :
denominator τ c :
ℝ
signLikelihood τ c 1 + signLikelihood τ c (-1)
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.denominator · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:51 · uses Cell
def posterior reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The posterior coefficient of a set S of records is the average over the shared sign, weighted by the two sign likelihoods and divided by their sum, of the product over S of uᵢ/(1 + τuᵢ) at sign +1 and of −uᵢ/(1 − τuᵢ) at sign −1, where uᵢ are the scores.

Definition (Lean source)
m :
ℕ
τ :
ℝ
c :
S :
posterior τ c S :
ℝ
(signLikelihood τ c 1 * (∏ i ∈ S, c.1 i / (1 + τ * c.1 i)) + signLikelihood τ c (-1) * (∏ i ∈ S, (-c.1 i) / (1 - τ * c.1 i))) / denominator τ c
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.posterior · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:55 · uses Cell
def subsetEnergy reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The order-d energy sums squared posterior coefficients over selected subsets of size d.

Definition (Lean source)
m :
ℕ
τ :
ℝ
c :
d :
ℕ
subsetEnergy τ c d :
ℝ
∑ S ∈ (selected c).powersetCard d, (posterior τ c S) ^ 2
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.subsetEnergy · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:299 · uses Cell
theorem subsetEnergy_one_bound reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

Given a tilt τ of absolute value at most 1/4 and valid cell data with m records, the order-one posterior subset energy is at most τ² · m · (m − 1)² · (25/9)ᵐ.

Formal statement
m :
ℕ
τ :
ℝ
c :
hτ :
|τ| ≤ 1 / 4
hc :
Valid c
subsetEnergy τ c 1 ≤ τ ^ 2 * (m : ℝ) * ((m - 1 : ℕ) : ℝ) ^ 2 * ((5 / 3 : ℝ) ^ 2) ^ m
Proof (Lean source)
theorem subsetEnergy_one_bound {m : ℕ} {τ : ℝ} {c : Cell m} (hτ : |τ| ≤ 1 / 4) (hc : Valid c) : subsetEnergy τ c 1 ≤ τ ^ 2 * (m : ℝ) * ((m - 1 : ℕ) : ℝ) ^ 2 * ((5 / 3 : ℝ) ^ 2) ^ m := by have hcount : ((selected c).powersetCard 1).card ≤ m := by simpa only [Finset.card_powersetCard, Finset.card_univ, Fintype.card_fin, Nat.choose_one_right] using (Finset.card_le_card (Finset.powersetCard_mono (n := 1) (Finset.subset_univ (selected c)))) have hterm (S : Finset (Fin m)) (hS : S ∈ (selected c).powersetCard 1) : (posterior τ c S) ^ 2 ≤ τ ^ 2 * ((m - 1 : ℕ) : ℝ) ^ 2 * ((5 / 3 : ℝ) ^ 2) ^ m := by obtain ⟨i, rfl⟩ := Finset.card_eq_one.mp (Finset.mem_powersetCard.mp hS).2 have hi : i ∈ active c := hc.2 ((Finset.mem_powersetCard.mp hS).1 (by simp)) have h := (sq_le_sq₀ (abs_nonneg (posterior τ c {i})) (by positivity : 0 ≤ |τ| * ((m - 1 : ℕ) : ℝ) * (5 / 3 : ℝ) ^ m)).2 (posterior_singleton_bound hτ hc.1 hi) simpa only [mul_pow, sq_abs, ← pow_mul, Nat.mul_comm] using h calc subsetEnergy τ c 1 ≤ ∑ S ∈ (selected c).powersetCard 1, τ ^ 2 * ((m - 1 : ℕ) : ℝ) ^ 2 * ((5 / 3 : ℝ) ^ 2) ^ m := Finset.sum_le_sum hterm _ = (((selected c).powersetCard 1).card : ℝ) * (τ ^ 2 * ((m - 1 : ℕ) : ℝ) ^ 2 * ((5 / 3 : ℝ) ^ 2) ^ m) := by simp only [Finset.sum_const, nsmul_eq_mul] _ ≤ (m : ℝ) * (τ ^ 2 * ((m - 1 : ℕ) : ℝ) ^ 2 * ((5 / 3 : ℝ) ^ 2) ^ m) := mul_le_mul_of_nonneg_right (by exact_mod_cast hcount) (by positivity) _ = τ ^ 2 * (m : ℝ) * ((m - 1 : ℕ) : ℝ) ^ 2 * ((5 / 3 : ℝ) ^ 2) ^ m := by ring
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.subsetEnergy_one_bound · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:332 · uses Cell , Valid , subsetEnergy
12 supporting declarations (lemmas, instances)
  • factor_bounds theorem — Every record likelihood factor lies between three quarters and five quarters when the score and sign bounds hold.
    τ u ell σ :
    ℝ
    hτ :
    |τ| ≤ 1 / 4
    hu :
    |u| ≤ 1
    hell :
    |ell| = 1
    hσ :
    |σ| = 1
    conclusion 1
    3 / 4 ≤ 1 + τ * ell * u * σ
    conclusion 2
    1 + τ * ell * u * σ ≤ 5 / 4
    Proof (Lean source)
    theorem factor_bounds {τ u ell σ : ℝ} (hτ : |τ| ≤ 1 / 4) (hu : |u| ≤ 1) (hell : |ell| = 1) (hσ : |σ| = 1) : 3 / 4 ≤ 1 + τ * ell * u * σ ∧ 1 + τ * ell * u * σ ≤ 5 / 4 := by have hmul : |τ * ell * u * σ| ≤ 1 / 4 := by simp only [abs_mul, hell, hσ, mul_one] exact (mul_le_mul_of_nonneg_left hu (abs_nonneg τ)).trans (by simpa using hτ) have hb := abs_le.mp hmul constructor <;> linarith
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.factor_bounds · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:61
  • denominator_lower theorem — A bounded cell has a denominator at least twice three quarters to its size.
    m :
    ℕ
    τ :
    ℝ
    c :
    hτ :
    |τ| ≤ 1 / 4
    hu :
    ∀ i, |c.1 i| ≤ 1
    2 * (3 / 4 : ℝ) ^ m ≤ denominator τ c
    Proof (Lean source)
    theorem denominator_lower {m : ℕ} {τ : ℝ} {c : Cell m} (hτ : |τ| ≤ 1 / 4) (hu : ∀ i, |c.1 i| ≤ 1) : 2 * (3 / 4 : ℝ) ^ m ≤ denominator τ c := by have hl (ell : ℝ) (hell : |ell| = 1) : (3 / 4 : ℝ) ^ m ≤ signLikelihood τ c ell := by have hs (i : Fin m) : |sign (c.2.1 i)| = 1 := by cases c.2.1 i <;> simp [sign] simpa only [signLikelihood, Finset.prod_const, Finset.card_univ, Fintype.card_fin] using (Finset.prod_le_prod (s := univ) (f := fun _ : Fin m => (3 / 4 : ℝ)) (by intros; norm_num) (fun i _ => (factor_bounds hτ (hu i) hell (hs i)).1)) have hp := hl 1 (by norm_num) have hn := hl (-1) (by norm_num) unfold denominator linarith
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.denominator_lower · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:72
  • denominator_pos theorem — The posterior denominator is strictly positive for bounded scores and tilt.
    m :
    ℕ
    τ :
    ℝ
    c :
    hτ :
    |τ| ≤ 1 / 4
    hu :
    ∀ i, |c.1 i| ≤ 1
    0 < denominator τ c
    Proof (Lean source)
    theorem denominator_pos {m : ℕ} {τ : ℝ} {c : Cell m} (hτ : |τ| ≤ 1 / 4) (hu : ∀ i, |c.1 i| ≤ 1) : 0 < denominator τ c := by have h := denominator_lower hτ hu have hp : 0 < 2 * (3 / 4 : ℝ) ^ m := by positivity linarith
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.denominator_pos · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:89
  • posterior_empty theorem — The empty posterior coefficient equals one for bounded scores and tilt.
    m :
    ℕ
    τ :
    ℝ
    c :
    hτ :
    |τ| ≤ 1 / 4
    hu :
    ∀ i, |c.1 i| ≤ 1
    posterior τ c ∅ = 1
    Proof (Lean source)
    theorem posterior_empty {m : ℕ} {τ : ℝ} {c : Cell m} (hτ : |τ| ≤ 1 / 4) (hu : ∀ i, |c.1 i| ≤ 1) : posterior τ c ∅ = 1 := by have hn := ne_of_gt (denominator_pos hτ hu) unfold posterior simp only [Finset.prod_empty, mul_one] change denominator τ c / denominator τ c = 1 exact div_self hn
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.posterior_empty · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:97
  • posterior_singleton_cancel theorem — For a positive assignment, the singleton posterior cancels its own likelihood factor, leaving the difference of the products over the other records.
    m :
    ℕ
    τ :
    ℝ
    c :
    hτ :
    |τ| ≤ 1 / 4
    hu :
    ∀ i, |c.1 i| ≤ 1
    i :
    Fin m
    hi :
    i ∈ active c
    posterior τ c {i}
    = c.1 i * ((∏ j ∈ Finset.univ.erase i, (1 + τ * c.1 j * sign (c.2.1 j))) - (∏ j ∈ Finset.univ.erase i, (1 - τ * c.1 j * sign (c.2.1 j)))) / denominator τ c
    Proof (Lean source)
    theorem posterior_singleton_cancel {m : ℕ} {τ : ℝ} {c : Cell m} (hτ : |τ| ≤ 1 / 4) (hu : ∀ i, |c.1 i| ≤ 1) {i : Fin m} (hi : i ∈ active c) : posterior τ c {i} = c.1 i * ((∏ j ∈ Finset.univ.erase i, (1 + τ * c.1 j * sign (c.2.1 j))) - (∏ j ∈ Finset.univ.erase i, (1 - τ * c.1 j * sign (c.2.1 j)))) / denominator τ c := by have hai : c.2.1 i = true := (Finset.mem_filter.mp hi).2 have hp : 1 + τ * c.1 i ≠ 0 := by have h := (factor_bounds hτ (hu i) (ell := 1) (σ := 1) (by norm_num) (by norm_num)).1 simp only [mul_one] at h linarith have hn : 1 - τ * c.1 i ≠ 0 := by have h := (factor_bounds hτ (hu i) (ell := -1) (σ := 1) (by norm_num) (by norm_num)).1 simp only [mul_one, mul_neg_one, neg_mul] at h linarith have hsign : sign (c.2.1 i) = 1 := by simp [sign, hai] have hplus : signLikelihood τ c 1 = (1 + τ * c.1 i) * ∏ j ∈ Finset.univ.erase i, (1 + τ * c.1 j * sign (c.2.1 j)) := by simpa only [signLikelihood, mul_one, hsign] using (Finset.mul_prod_erase (s := univ) (f := fun j => 1 + τ * c.1 j * sign (c.2.1 j)) (Finset.mem_univ i)).symm have hminus : signLikelihood τ c (-1) = (1 - τ * c.1 i) * ∏ j ∈ Finset.univ.erase i, (1 - τ * c.1 j * sign (c.2.1 j)) := by simpa only [signLikelihood, mul_neg_one, neg_mul, ← sub_eq_add_neg, hsign, mul_one] using (Finset.mul_prod_erase (s := univ) (f := fun j => 1 - τ * c.1 j * sign (c.2.1 j)) (Finset.mem_univ i)).symm unfold posterior simp only [Finset.prod_singleton] rw [hplus, hminus] congr 1 field_simp [hp, hn] ring
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.posterior_singleton_cancel · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:107
  • opposite_products_bound theorem — Opposite-sign products differ by at most a linear tilt times the number of factors and a five-quarters exponential envelope.
    ι :
    Type*
    s :
    v :
    ι → ℝ
    τ :
    ℝ
    hτ :
    |τ| ≤ 1 / 4
    hv :
    ∀ i ∈ s, |v i| ≤ 1
    |(∏ i ∈ s, (1 + τ * v i)) - (∏ i ∈ s, (1 - τ * v i))|
    ≤ 2 * |τ| * (s.card : ℝ) * (5 / 4 : ℝ) ^ s.card
    Proof (Lean source)
    theorem opposite_products_bound {ι : Type*} (s : Finset ι) (v : ι → ℝ) {τ : ℝ} (hτ : |τ| ≤ 1 / 4) (hv : ∀ i ∈ s, |v i| ≤ 1) : |(∏ i ∈ s, (1 + τ * v i)) - (∏ i ∈ s, (1 - τ * v i))| ≤ 2 * |τ| * (s.card : ℝ) * (5 / 4 : ℝ) ^ s.card := by classical have hfactor (i : ι) (hi : i ∈ s) : |1 + τ * v i| ≤ 5 / 4 ∧ |1 - τ * v i| ≤ 5 / 4 := by have hp := factor_bounds hτ (hv i hi) (ell := 1) (σ := 1) (by norm_num) (by norm_num) have hn := factor_bounds hτ (hv i hi) (ell := -1) (σ := 1) (by norm_num) (by norm_num) simp only [mul_one, mul_neg_one, neg_mul, ← sub_eq_add_neg] at hp hn constructor <;> rw [abs_of_nonneg (by linarith)] <;> linarith have hprod (t : Finset ι) (ht : t ⊆ s) : |∏ i ∈ t, (1 - τ * v i)| ≤ (5 / 4 : ℝ) ^ t.card := by rw [Finset.abs_prod] simpa only [Finset.prod_const] using (Finset.prod_le_prod (fun _ _ => abs_nonneg _) (fun i hi => (hfactor i (ht hi)).2)) induction s using Finset.induction_on with | empty => simp | @insert a t ha ih => have hva := hv a (Finset.mem_insert_self a t) have hvt : ∀ i ∈ t, |v i| ≤ 1 := fun i hi => hv i (Finset.mem_insert_of_mem hi) have hd := ih hvt (fun i hi => hfactor i (Finset.mem_insert_of_mem hi)) (fun u hu => hprod u (hu.trans (Finset.subset_insert a t))) have hpa := (hfactor a (Finset.mem_insert_self a t)).1 have hpt := hprod t (Finset.subset_insert a t) have hdiff : |2 * τ * v a| ≤ 2 * |τ| := by rw [abs_mul, abs_mul] norm_num exact mul_le_of_le_one_right (by positivity) hva rw [Finset.prod_insert ha, Finset.prod_insert ha, Finset.card_insert_of_notMem ha, Nat.cast_add, Nat.cast_one, pow_succ] calc |(1 + τ * v a) * (∏ i ∈ t, (1 + τ * v i)) - (1 - τ * v a) * (∏ i ∈ t, (1 - τ * v i))| = |(1 + τ * v a) * ((∏ i ∈ t, (1 + τ * v i)) - (∏ i ∈ t, (1 - τ * v i))) + (2 * τ * v a) * (∏ i ∈ t, (1 - τ * v i))| := by congr 1; ring _ ≤ |1 + τ * v a| * |(∏ i ∈ t, (1 + τ * v i)) - (∏ i ∈ t, (1 - τ * v i))| + |2 * τ * v a| * |∏ i ∈ t, (1 - τ * v i)| := by simpa only [abs_mul] using abs_add_le ((1 + τ * v a) * ((∏ i ∈ t, (1 + τ * v i)) - (∏ i ∈ t, (1 - τ * v i)))) ((2 * τ * v a) * (∏ i ∈ t, (1 - τ * v i))) _ ≤ (5 / 4) * (2 * |τ| * (t.card : ℝ) * (5 / 4 : ℝ) ^ t.card) + (2 * |τ|) * (5 / 4 : ℝ) ^ t.card := by exact add_le_add (mul_le_mul hpa hd (abs_nonneg _) (by norm_num)) (mul_le_mul hdiff hpt (abs_nonneg _) (by positivity)) _ ≤ 2 * |τ| * ((t.card : ℝ) + 1) * ((5 / 4 : ℝ) ^ t.card * (5 / 4)) := by have hp : 0 ≤ |τ| * (5 / 4 : ℝ) ^ t.card := by positivity nlinarith
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.opposite_products_bound · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:145
  • posterior_singleton_bound theorem — A singleton active coefficient is bounded by the tilt times the number of other records and five thirds to the cell size.
    m :
    ℕ
    τ :
    ℝ
    c :
    hτ :
    |τ| ≤ 1 / 4
    hu :
    ∀ i, |c.1 i| ≤ 1
    i :
    Fin m
    hi :
    i ∈ active c
    |posterior τ c {i}| ≤ |τ| * ((m - 1 : ℕ) : ℝ) * (5 / 3 : ℝ) ^ m
    Proof (Lean source)
    theorem posterior_singleton_bound {m : ℕ} {τ : ℝ} {c : Cell m} (hτ : |τ| ≤ 1 / 4) (hu : ∀ i, |c.1 i| ≤ 1) {i : Fin m} (hi : i ∈ active c) : |posterior τ c {i}| ≤ |τ| * ((m - 1 : ℕ) : ℝ) * (5 / 3 : ℝ) ^ m := by have hs (j : Fin m) : |sign (c.2.1 j)| = 1 := by cases c.2.1 j <;> simp [sign] have hv (j : Fin m) : |c.1 j * sign (c.2.1 j)| ≤ 1 := by simpa only [abs_mul, hs j, mul_one] using hu j have hb := opposite_products_bound (Finset.univ.erase i) (fun j => c.1 j * sign (c.2.1 j)) hτ (fun j _ => hv j) have hcard : (Finset.univ.erase i).card = m - 1 := by simp simp only [hcard, ← mul_assoc] at hb have hpow : (5 / 4 : ℝ) ^ (m - 1) ≤ (5 / 4 : ℝ) ^ m := pow_le_pow_right₀ (by norm_num) (sub_le m 1) have hb' := hb.trans (mul_le_mul_of_nonneg_left hpow (by positivity : 0 ≤ 2 * |τ| * ((m - 1 : ℕ) : ℝ))) rw [posterior_singleton_cancel hτ hu hi, abs_div, abs_mul, abs_of_pos (denominator_pos hτ hu)] apply (div_le_iff₀ (denominator_pos hτ hu)).2 calc |c.1 i| * |(∏ j ∈ Finset.univ.erase i, (1 + τ * c.1 j * sign (c.2.1 j))) - (∏ j ∈ Finset.univ.erase i, (1 - τ * c.1 j * sign (c.2.1 j)))| ≤ 2 * |τ| * ((m - 1 : ℕ) : ℝ) * (5 / 4 : ℝ) ^ m := (mul_le_of_le_one_left (abs_nonneg _) (hu i)).trans hb' _ = (|τ| * ((m - 1 : ℕ) : ℝ) * (5 / 3 : ℝ) ^ m) * (2 * (3 / 4 : ℝ) ^ m) := by have he : (5 / 3 : ℝ) ^ m * (3 / 4 : ℝ) ^ m = (5 / 4 : ℝ) ^ m := by rw [← mul_pow]; norm_num calc _ = 2 * |τ| * ((m - 1 : ℕ) : ℝ) * ((5 / 3 : ℝ) ^ m * (3 / 4 : ℝ) ^ m) := by rw [he] _ = _ := by ring _ ≤ (|τ| * ((m - 1 : ℕ) : ℝ) * (5 / 3 : ℝ) ^ m) * denominator τ c := mul_le_mul_of_nonneg_left (denominator_lower hτ hu) (by positivity)
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.posterior_singleton_bound · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:203
  • posterior_singleton_one theorem — A one-record cell has zero active singleton coefficient.
    τ :
    ℝ
    c :
    hτ :
    |τ| ≤ 1 / 4
    hu :
    ∀ i, |c.1 i| ≤ 1
    i :
    Fin 1
    hi :
    i ∈ active c
    posterior τ c {i} = 0
    Proof (Lean source)
    theorem posterior_singleton_one {τ : ℝ} {c : Cell 1} (hτ : |τ| ≤ 1 / 4) (hu : ∀ i, |c.1 i| ≤ 1) {i : Fin 1} (hi : i ∈ active c) : posterior τ c {i} = 0 := by have h := posterior_singleton_bound hτ hu hi simpa using abs_nonpos_iff.mp (by simpa using h)
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.posterior_singleton_one · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:240
  • posterior_subset_bound theorem — Every posterior subset coefficient is bounded by two to its cardinality. The bound holds even without the active-subset restriction.
    m :
    ℕ
    τ :
    ℝ
    c :
    hτ :
    |τ| ≤ 1 / 4
    hu :
    ∀ i, |c.1 i| ≤ 1
    S :
    |posterior τ c S| ≤ (2 : ℝ) ^ S.card
    Proof (Lean source)
    theorem posterior_subset_bound {m : ℕ} {τ : ℝ} {c : Cell m} (hτ : |τ| ≤ 1 / 4) (hu : ∀ i, |c.1 i| ≤ 1) (S : Finset (Fin m)) : |posterior τ c S| ≤ (2 : ℝ) ^ S.card := by have hf (i : Fin m) : |c.1 i / (1 + τ * c.1 i)| ≤ 2 ∧ |(-c.1 i) / (1 - τ * c.1 i)| ≤ 2 := by have hp := (factor_bounds hτ (hu i) (ell := 1) (σ := 1) (by norm_num) (by norm_num)).1 have hn := (factor_bounds hτ (hu i) (ell := -1) (σ := 1) (by norm_num) (by norm_num)).1 simp only [mul_one, mul_neg_one, neg_mul, ← sub_eq_add_neg] at hp hn have hpp : 0 < 1 + τ * c.1 i := by linarith have hnp : 0 < 1 - τ * c.1 i := by linarith constructor · rw [abs_div, abs_of_pos hpp] apply (div_le_iff₀ hpp).2 linarith [hu i] · rw [abs_div, abs_neg, abs_of_pos hnp] apply (div_le_iff₀ hnp).2 linarith [hu i] have hp : |∏ i ∈ S, c.1 i / (1 + τ * c.1 i)| ≤ (2 : ℝ) ^ S.card := by rw [Finset.abs_prod] simpa only [Finset.prod_const] using (Finset.prod_le_prod (fun _ _ => abs_nonneg _) (fun i _ => (hf i).1)) have hn : |∏ i ∈ S, (-c.1 i) / (1 - τ * c.1 i)| ≤ (2 : ℝ) ^ S.card := by rw [Finset.abs_prod] simpa only [Finset.prod_const] using (Finset.prod_le_prod (fun _ _ => abs_nonneg _) (fun i _ => (hf i).2)) have hL (ell : ℝ) (hell : |ell| = 1) : 0 ≤ signLikelihood τ c ell := by apply Finset.prod_nonneg intro i _ have hs : |sign (c.2.1 i)| = 1 := by cases c.2.1 i <;> simp [sign] have h := (factor_bounds hτ (hu i) hell hs).1 linarith have hLp := hL 1 (by norm_num) have hLn := hL (-1) (by norm_num) unfold posterior rw [abs_div, abs_of_pos (denominator_pos hτ hu)] apply (div_le_iff₀ (denominator_pos hτ hu)).2 calc _ ≤ |signLikelihood τ c 1 * (∏ i ∈ S, c.1 i / (1 + τ * c.1 i))| + |signLikelihood τ c (-1) * (∏ i ∈ S, (-c.1 i) / (1 - τ * c.1 i))| := abs_add_le _ _ _ = signLikelihood τ c 1 * |∏ i ∈ S, c.1 i / (1 + τ * c.1 i)| + signLikelihood τ c (-1) * |∏ i ∈ S, (-c.1 i) / (1 - τ * c.1 i)| := by rw [abs_mul, abs_mul, abs_of_nonneg hLp, abs_of_nonneg hLn] _ ≤ signLikelihood τ c 1 * (2 : ℝ) ^ S.card + signLikelihood τ c (-1) * (2 : ℝ) ^ S.card := add_le_add (mul_le_mul_of_nonneg_left hp hLp) (mul_le_mul_of_nonneg_left hn hLn) _ = (2 : ℝ) ^ S.card * denominator τ c := by unfold denominator; ring
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.posterior_subset_bound · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:247
  • subsetEnergy_nonneg theorem — All subset energies are nonnegative.
    m :
    ℕ
    τ :
    ℝ
    c :
    d :
    ℕ
    0 ≤ subsetEnergy τ c d
    Proof (Lean source)
    theorem subsetEnergy_nonneg {m : ℕ} (τ : ℝ) (c : Cell m) (d : ℕ) : 0 ≤ subsetEnergy τ c d := by exact Finset.sum_nonneg (fun _ _ => sq_nonneg _)
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.subsetEnergy_nonneg · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:303
  • subsetEnergy_le_choose theorem — The order-d energy is bounded by four to d times the cell-size binomial coefficient.
    m :
    ℕ
    τ :
    ℝ
    c :
    hτ :
    |τ| ≤ 1 / 4
    hu :
    ∀ i, |c.1 i| ≤ 1
    d :
    ℕ
    subsetEnergy τ c d ≤ (4 : ℝ) ^ d * (m.choose d : ℝ)
    Proof (Lean source)
    theorem subsetEnergy_le_choose {m : ℕ} {τ : ℝ} {c : Cell m} (hτ : |τ| ≤ 1 / 4) (hu : ∀ i, |c.1 i| ≤ 1) (d : ℕ) : subsetEnergy τ c d ≤ (4 : ℝ) ^ d * (m.choose d : ℝ) := by have hcount : ((selected c).powersetCard d).card ≤ m.choose d := by simpa only [Finset.card_powersetCard, Finset.card_univ, Fintype.card_fin] using (Finset.card_le_card (Finset.powersetCard_mono (Finset.subset_univ (selected c)))) have hterm (S : Finset (Fin m)) (hS : S ∈ (selected c).powersetCard d) : (posterior τ c S) ^ 2 ≤ (4 : ℝ) ^ d := by have hcard := (Finset.mem_powersetCard.mp hS).2 have h := (sq_le_sq₀ (abs_nonneg (posterior τ c S)) (by positivity : 0 ≤ (2 : ℝ) ^ S.card)).2 (posterior_subset_bound hτ hu S) rw [sq_abs, ← pow_mul, Nat.mul_comm, pow_mul, hcard] at h norm_num at h ⊢ exact h calc subsetEnergy τ c d ≤ ∑ S ∈ (selected c).powersetCard d, (4 : ℝ) ^ d := Finset.sum_le_sum hterm _ = (((selected c).powersetCard d).card : ℝ) * (4 : ℝ) ^ d := by simp only [Finset.sum_const, nsmul_eq_mul] _ ≤ (m.choose d : ℝ) * (4 : ℝ) ^ d := mul_le_mul_of_nonneg_right (by exact_mod_cast hcount) (by positivity) _ = (4 : ℝ) ^ d * (m.choose d : ℝ) := mul_comm _ _
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.subsetEnergy_le_choose · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:308
  • measurable_subsetEnergy theorem — Each fixed-order subset energy is measurable as a function of the cell data.
    m d :
    ℕ
    τ :
    ℝ
    Measurable (fun c : Cell m => subsetEnergy τ c d)
    Proof (Lean source)
    @[fun_prop] theorem measurable_subsetEnergy (m d : ℕ) (τ : ℝ) : Measurable (fun c : Cell m => subsetEnergy τ c d) := by classical have hsign : Measurable sign := measurable_of_countable sign have hp (S : Finset (Fin m)) : Measurable (fun c : Cell m => (posterior τ c S) ^ 2) := by unfold posterior denominator signLikelihood fun_prop have hsel (S : Finset (Fin m)) : MeasurableSet {c : Cell m | S ⊆ selected c} := by have hflags : Measurable (fun c : Cell m => c.2.2) := measurable_snd.comp measurable_snd have hset : MeasurableSet {b : Fin m → Bool | S ⊆ Finset.univ.filter (fun i => b i = true)} := (Set.to_countable _).measurableSet exact hset.preimage hflags have heq (c : Cell m) : subsetEnergy τ c d = ∑ S ∈ (Finset.univ : Finset (Fin m)).powersetCard d, if S ⊆ selected c then (posterior τ c S) ^ 2 else 0 := by rw [← Finset.sum_filter] unfold subsetEnergy congr 1 ext S simp only [mem_filter, Finset.mem_powersetCard, Finset.subset_univ, true_and] exact and_comm simp_rw [heq] apply Finset.measurable_sum intro S _ exact Measurable.ite (hsel S) (hp S) measurable_const
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.measurable_subsetEnergy · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Pointwise.lean:363
Series 2 core · 3 supporting A nonnegative coefficient sequence dominated by an exponential series defines a strictly positive overlap within the small-radius regime. ★ overlapSeries_pow_le_exp

Exact-linear cell overlap series envelope

A nonnegative coefficient sequence dominated by an exponential series defines a strictly positive overlap within the small-radius regime. The exponential power bound retains the exact first coefficient, including at negative overlap.

def overlapSeries reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The cell overlap series starts at one and sums all positive-order coefficient terms.

Definition (Lean source)
t :
ℕ → ℝ
z :
ℝ
overlapSeries t z :
ℝ
1 + ∑' d : ℕ, t (d + 1) * z ^ (d + 1)
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.overlapSeries · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Series.lean:23
theorem overlapSeries_pow_le_exp reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

Given a nonnegative exponential radius b, coefficients with 0 ≤ t_d ≤ bᵈ/d! for every d ≥ 1, an overlap z with b|z| ≤ 1/2, and a natural power k, the k-th power of the overlap series is at most exp(k·t₁·z + k·b²z²).

Formal statement
t :
ℕ → ℝ
b z :
ℝ
hb :
0 ≤ b
ht :
∀ d
if
1 ≤ d
then
0 ≤ t d ∧ t d ≤ b ^ d / (d.factorial : ℝ)
hr :
b * |z| ≤ 1 / 2
k :
ℕ
overlapSeries t z ^ k ≤ exp (((k : ℝ) * t 1) * z + (k : ℝ) * b ^ 2 * z ^ 2)
Proof (Lean source)
theorem overlapSeries_pow_le_exp {t : ℕ → ℝ} {b z : ℝ} (hb : 0 ≤ b) (ht : ∀ d, 1 ≤ d → 0 ≤ t d ∧ t d ≤ b ^ d / (d.factorial : ℝ)) (hr : b * |z| ≤ 1 / 2) (k : ℕ) : overlapSeries t z ^ k ≤ exp (((k : ℝ) * t 1) * z + (k : ℝ) * b ^ 2 * z ^ 2) := by -- K≤1+t₁z+b²z²≤exp(t₁z+b²z²); use positivity before raising to k. have hupper := (abs_le.mp (overlapSeries_remainder hb ht hr)).2 have hexp := Real.add_one_le_exp (t 1 * z + b ^ 2 * z ^ 2) have hle : overlapSeries t z ≤ exp (t 1 * z + b ^ 2 * z ^ 2) := by linarith calc overlapSeries t z ^ k ≤ exp (t 1 * z + b ^ 2 * z ^ 2) ^ k := pow_le_pow_left₀ (le_of_lt (overlapSeries_pos hb ht hr)) hle k _ = exp (((k : ℝ) * t 1) * z + (k : ℝ) * b ^ 2 * z ^ 2) := by rw [← Real.exp_nat_mul] congr 1 ring
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.overlapSeries_pow_le_exp · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Series.lean:111 · uses overlapSeries
3 supporting declarations (lemmas, instances)
  • overlapSeries_summable_abs theorem — An exponential majorant makes the cell overlap series absolutely convergent at every real argument.
    t :
    ℕ → ℝ
    b :
    ℝ
    ht :
    ∀ d
    if
    1 ≤ d
    then
    0 ≤ t d ∧ t d ≤ b ^ d / (d.factorial : ℝ)
    z :
    ℝ
    Summable (fun d : ℕ => |t (d + 1) * z ^ (d + 1)|)
    Proof (Lean source)
    theorem overlapSeries_summable_abs {t : ℕ → ℝ} {b : ℝ} (ht : ∀ d, 1 ≤ d → 0 ≤ t d ∧ t d ≤ b ^ d / (d.factorial : ℝ)) (z : ℝ) : Summable (fun d : ℕ => |t (d + 1) * z ^ (d + 1)|) := by have hs := (summable_nat_add_iff 1).2 (Real.summable_pow_div_factorial (b * |z|)) refine hs.of_nonneg_of_le (fun _ => abs_nonneg _) (fun d => ?_) rw [abs_mul, abs_of_nonneg (ht (d + 1) (by omega)).1, abs_pow] calc t (d + 1) * |z| ^ (d + 1) ≤ (b ^ (d + 1) / ((d + 1).factorial : ℝ)) * |z| ^ (d + 1) := mul_le_mul_of_nonneg_right (ht (d + 1) (by omega)).2 (by positivity) _ = (b * |z|) ^ (d + 1) / ((d + 1).factorial : ℝ) := by rw [mul_pow]; ring
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.overlapSeries_summable_abs · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Series.lean:27
  • overlapSeries_remainder theorem — After removing the exact linear term, the overlap remainder is bounded in absolute value by the squared exponential radius when that radius is at most one half.
    t :
    ℕ → ℝ
    b z :
    ℝ
    hb :
    0 ≤ b
    ht :
    ∀ d
    if
    1 ≤ d
    then
    0 ≤ t d ∧ t d ≤ b ^ d / (d.factorial : ℝ)
    hr :
    b * |z| ≤ 1 / 2
    |overlapSeries t z - 1 - t 1 * z| ≤ b ^ 2 * z ^ 2
    Proof (Lean source)
    theorem overlapSeries_remainder {t : ℕ → ℝ} {b z : ℝ} (hb : 0 ≤ b) (ht : ∀ d, 1 ≤ d → 0 ≤ t d ∧ t d ≤ b ^ d / (d.factorial : ℝ)) (hr : b * |z| ≤ 1 / 2) : |overlapSeries t z - 1 - t 1 * z| ≤ b ^ 2 * z ^ 2 := by -- The d≥2 tail is at most r² ∑n r^n/(n+2)! ≤ r² for r≤1/2. have hsabs := overlapSeries_summable_abs ht z have hs : Summable (fun d : ℕ => t (d + 1) * z ^ (d + 1)) := by exact (by simpa only [Real.norm_eq_abs] using hsabs : Summable (fun d : ℕ => ‖t (d + 1) * z ^ (d + 1)‖)).of_norm have hsplit := hs.sum_add_tsum_nat_add 1 simp only [Finset.sum_range_one, zero_add, pow_one, Nat.add_assoc] at hsplit have heq : overlapSeries t z - 1 - t 1 * z = ∑' d : ℕ, t (d + 2) * z ^ (d + 2) := by unfold overlapSeries linarith rw [heq] have htail : Summable (fun d : ℕ => |t (d + 2) * z ^ (d + 2)|) := by simpa only [Nat.add_assoc] using (summable_nat_add_iff 1).2 hsabs have hgeom : Summable (fun d : ℕ => (b * |z|) ^ 2 / 2 * (1 / 2 : ℝ) ^ d) := (summable_geometric_of_abs_lt_one (by norm_num : |(1 / 2 : ℝ)| < 1)).mul_left _ have hbound : ∀ d : ℕ, |t (d + 2) * z ^ (d + 2)| ≤ (b * |z|) ^ 2 / 2 * (1 / 2 : ℝ) ^ d := by intro d have hf : (2 : ℝ) ≤ ((d + 2).factorial : ℝ) := by exact_mod_cast (show 2 ≤ (d + 2).factorial from by simpa using Nat.factorial_le (show 2 ≤ d + 2 by omega)) rw [abs_mul, abs_of_nonneg (ht (d + 2) (by omega)).1, abs_pow] calc t (d + 2) * |z| ^ (d + 2) ≤ (b ^ (d + 2) / ((d + 2).factorial : ℝ)) * |z| ^ (d + 2) := mul_le_mul_of_nonneg_right (ht (d + 2) (by omega)).2 (by positivity) _ = (b * |z|) ^ (d + 2) / ((d + 2).factorial : ℝ) := by rw [mul_pow]; ring _ ≤ (b * |z|) ^ (d + 2) / 2 := div_le_div_of_nonneg_left (by positivity) (by norm_num) hf _ = (b * |z|) ^ 2 / 2 * (b * |z|) ^ d := by rw [pow_add]; ring _ ≤ (b * |z|) ^ 2 / 2 * (1 / 2 : ℝ) ^ d := mul_le_mul_of_nonneg_left (pow_le_pow_left₀ (by positivity) hr d) (by positivity) calc |∑' d : ℕ, t (d + 2) * z ^ (d + 2)| ≤ ∑' d : ℕ, |t (d + 2) * z ^ (d + 2)| := by simpa only [Real.norm_eq_abs] using norm_tsum_le_tsum_norm (f := fun d : ℕ => t (d + 2) * z ^ (d + 2)) (by simpa only [Real.norm_eq_abs] using htail) _ ≤ ∑' d : ℕ, (b * |z|) ^ 2 / 2 * (1 / 2 : ℝ) ^ d := htail.tsum_le_tsum hbound hgeom _ = b ^ 2 * z ^ 2 := by rw [tsum_mul_left, tsum_geometric_of_abs_lt_one (by norm_num)] simp only [mul_pow, sq_abs] ring
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.overlapSeries_remainder · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Series.lean:41
  • overlapSeries_pos theorem — The cell overlap series is strictly positive when the absolute exponential radius is at most one half.
    t :
    ℕ → ℝ
    b z :
    ℝ
    hb :
    0 ≤ b
    ht :
    ∀ d
    if
    1 ≤ d
    then
    0 ≤ t d ∧ t d ≤ b ^ d / (d.factorial : ℝ)
    hr :
    b * |z| ≤ 1 / 2
    Proof (Lean source)
    theorem overlapSeries_pos {t : ℕ → ℝ} {b z : ℝ} (hb : 0 ≤ b) (ht : ∀ d, 1 ≤ d → 0 ≤ t d ∧ t d ≤ b ^ d / (d.factorial : ℝ)) (hr : b * |z| ≤ 1 / 2) : 0 < overlapSeries t z := by -- Use the remainder bound and |t₁z|≤r: K≥1-r-r²≥1/4. have ht1 := ht 1 (by omega) simp only [pow_one, Nat.factorial_one, Nat.cast_one, div_one] at ht1 have hlin : |t 1 * z| ≤ b * |z| := by rw [abs_mul, abs_of_nonneg ht1.1] exact mul_le_mul_of_nonneg_right ht1.2 (abs_nonneg z) have hrem := (abs_le.mp (overlapSeries_remainder hb ht hr)).1 have hlinlow := (abs_le.mp hlin).1 have hr0 : 0 ≤ b * |z| := mul_nonneg hb (abs_nonneg z) have hsq : b ^ 2 * z ^ 2 = (b * |z|) ^ 2 := by rw [mul_pow, sq_abs] rw [hsq] at hrem nlinarith
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.overlapSeries_pos · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Series.lean:93
Averaging 4 core · 7 supporting An arbitrary probability law of scores, assignments, and selection is supplied separately at each count. ★ coefficient_one_le★ poisson_overlap_pow_le_exp

Poisson-averaged posterior subset coefficients

An arbitrary probability law of scores, assignments, and selection is supplied separately at each count. Only almost-everywhere score bounds and selection within positive assignments are required; dependence within each cell is unrestricted.

def conditionalEnergy reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The conditional order-d energy at a fixed cell count is the mean, under the cell law at that count, of the order-d subset energy at tilt τ.

Definition (Lean source)
ν :
∀ m, Measure (Cell m)
τ :
ℝ
m d :
ℕ
conditionalEnergy ν τ m d :
ℝ
∫ c, subsetEnergy τ c d ∂ν m
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.conditionalEnergy · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Averaging.lean:21 · uses Cell
def coefficient reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The order-d Poisson coefficient is γ^(2d) times the mean, over a Poisson count with mean ξ, of the conditional order-d energy at that count.

Definition (Lean source)
ν :
∀ m, Measure (Cell m)
ξ :
τ γ :
ℝ
d :
ℕ
coefficient ν ξ τ γ d :
ℝ
γ ^ (2 * d) * ∫ m : ℕ, conditionalEnergy ν τ m d ∂poissonMeasure ξ
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.coefficient · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Averaging.lean:25 · uses Cell
theorem coefficient_one_le reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

Given conditional probability laws for cells, a Poisson mean at most one ξ, a tilt τ of absolute value at most 1/4, cells that are almost surely valid at every count, and an outcome amplitude γ, the order-one Poisson coefficient is at most e^((5/3)² − 1) · ((5/3)⁶ + (5/3)⁴) · ξ² τ² γ².

Formal statement
ν :
∀ m, Measure (Cell m)
∀ m, IsProbabilityMeasure (ν m)
ξ :
hξ :
(ξ : ℝ) ≤ 1
τ :
ℝ
hτ :
|τ| ≤ 1 / 4
hν :
∀ m, ∀ᵐ c ∂ν m, Valid c
γ :
ℝ
coefficient ν ξ τ γ 1
≤ (exp ((5 / 3 : ℝ) ^ 2 - 1) * ((5 / 3 : ℝ) ^ 6 + (5 / 3 : ℝ) ^ 4)) * (ξ : ℝ) ^ 2 * τ ^ 2 * γ ^ 2
Proof (Lean source)
theorem coefficient_one_le (ν : ∀ m, Measure (Cell m)) [∀ m, IsProbabilityMeasure (ν m)] {ξ : NNReal} (hξ : (ξ : ℝ) ≤ 1) {τ : ℝ} (hτ : |τ| ≤ 1 / 4) (hν : ∀ m, ∀ᵐ c ∂ν m, Valid c) (γ : ℝ) : coefficient ν ξ τ γ 1 ≤ (exp ((5 / 3 : ℝ) ^ 2 - 1) * ((5 / 3 : ℝ) ^ 6 + (5 / 3 : ℝ) ^ 4)) * (ξ : ℝ) ^ 2 * τ ^ 2 * γ ^ 2 := by -- Integrate subsetEnergy_one_bound twice. The majorant is τ² times the -- collision tilt; integrable_poisson_collision_tilt supplies its receipt, -- and integral_poisson_collision_le supplies the explicit constant. let f : ℕ → ℝ := fun m => (m : ℝ) * ((m - 1 : ℕ) : ℝ) ^ 2 * ((5 / 3 : ℝ) ^ 2) ^ m have hf : Integrable f (poissonMeasure ξ) := integrable_poisson_collision_tilt ξ (by positivity) have hb (m : ℕ) : conditionalEnergy ν τ m 1 ≤ τ ^ 2 * f m := by unfold conditionalEnergy calc _ ≤ ∫ _c : Cell m, τ ^ 2 * f m ∂ν m := by apply integral_mono_ae (integrable_subsetEnergy ν hτ hν m 1) (integrable_const _) filter_upwards [hν m] with c hc simpa only [f, mul_assoc] using subsetEnergy_one_bound hτ hc _ = _ := by simp have hi : (∫ m : ℕ, conditionalEnergy ν τ m 1 ∂poissonMeasure ξ) ≤ τ ^ 2 * (∫ m : ℕ, f m ∂poissonMeasure ξ) := by calc _ ≤ ∫ m : ℕ, τ ^ 2 * f m ∂poissonMeasure ξ := integral_mono_ae (integrable_conditionalEnergy ν ξ hτ hν 1) (hf.const_mul _) (Filter.Eventually.of_forall hb) _ = _ := integral_const_mul _ _ have hj := mul_le_mul_of_nonneg_left (integral_poisson_collision_le hξ) (sq_nonneg τ) have hk := mul_le_mul_of_nonneg_left (hi.trans hj) (sq_nonneg γ) simpa only [coefficient, mul_one, one_mul, mul_assoc, mul_comm, mul_left_comm] using hk
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.coefficient_one_le · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Averaging.lean:106 · uses Cell , Valid , coefficient
theorem poisson_overlap_pow_le_exp reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

Given conditional probability laws for cells, a Poisson mean, a tilt of absolute value at most 1/4, cells that are almost surely valid at every count, an overlap z with 4ξγ²|z| ≤ 1/2, and a natural power k, the k-th power of the overlap series of the Poisson coefficients is at most exp(k·c₁·z + 16k·ξ²γ⁴z²), where c₁ is the order-one Poisson coefficient.

Formal statement
ν :
∀ m, Measure (Cell m)
∀ m, IsProbabilityMeasure (ν m)
ξ :
τ γ z :
ℝ
hτ :
|τ| ≤ 1 / 4
hν :
∀ m, ∀ᵐ c ∂ν m, Valid c
hr :
4 * (ξ : ℝ) * γ ^ 2 * |z| ≤ 1 / 2
k :
ℕ
overlapSeries (coefficient ν ξ τ γ) z ^ k
≤ exp (((k : ℝ) * coefficient ν ξ τ γ 1) * z + 16 * (k : ℝ) * (ξ : ℝ) ^ 2 * γ ^ 4 * z ^ 2)
Proof (Lean source)
theorem poisson_overlap_pow_le_exp (ν : ∀ m, Measure (Cell m)) [∀ m, IsProbabilityMeasure (ν m)] (ξ : NNReal) {τ γ z : ℝ} (hτ : |τ| ≤ 1 / 4) (hν : ∀ m, ∀ᵐ c ∂ν m, Valid c) (hr : 4 * (ξ : ℝ) * γ ^ 2 * |z| ≤ 1 / 2) (k : ℕ) : overlapSeries (coefficient ν ξ τ γ) z ^ k ≤ exp (((k : ℝ) * coefficient ν ξ τ γ 1) * z + 16 * (k : ℝ) * (ξ : ℝ) ^ 2 * γ ^ 4 * z ^ 2) := by -- Specialize overlapSeries_pow_le_exp with b=4ξγ². No sign assumption on z. have h := overlapSeries_pow_le_exp (b := 4 * (ξ : ℝ) * γ ^ 2) (by positivity) (fun d _ => ⟨coefficient_nonneg ν ξ τ γ d, coefficient_le_factorial ν ξ hτ hν γ d⟩) hr k convert h using 1 congr 1 ring
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.poisson_overlap_pow_le_exp · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Averaging.lean:208 · uses Cell , Valid , coefficient , overlapSeries
7 supporting declarations (lemmas, instances)
Outcome 3 core · 5 supporting The posterior outcome likelihood is the posterior average of finite products. ★ integral_outcomeLikelihood_pair_by_card

Centered outcome products in a shared-sign cell

The posterior outcome likelihood is the posterior average of finite products. Independent outcome-score pairs across records give orthogonality of distinct subsets. Integrability is derived from individual score and overlap-product integrability, rather than assumed for the full likelihood product.

def outcomeLikelihood reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The posterior outcome likelihood of a cell with scores uᵢ, at outcome values vᵢ, is the average over the shared sign ±1, weighted by the two sign likelihoods and divided by their sum, of the product over the selected records of 1 + γ · (±uᵢ)/(1 ± τuᵢ) · vᵢ.

Definition (Lean source)
m :
ℕ
τ γ :
ℝ
c :
v :
Fin m → ℝ
outcomeLikelihood τ γ c v :
ℝ
(signLikelihood τ c 1 * (∏ i ∈ selected c, (1 + γ * (c.1 i / (1 + τ * c.1 i)) * v i)) + signLikelihood τ c (-1) * (∏ i ∈ selected c, (1 + γ * ((-c.1 i) / (1 - τ * c.1 i)) * v i))) / denominator τ c
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.outcomeLikelihood · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Outcome.lean:22 · uses Cell
structure CenteredOutcomePair reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

Centered outcome-score pairs are independent across records, have integrable individual scores and pair products, zero means, and the specified common overlap.

Definition (Lean source)
ι Ω :
ρ :
v w :
ι → Ω → ℝ
z :
ℝ
independent :
iIndepFun (fun i ω => (v i ω, w i ω)) ρ

Pairs of scores are jointly independent across records.

integrable_left :
∀ i, Integrable (v i) ρ

Each first score is integrable.

integrable_right :
∀ i, Integrable (w i) ρ

Each second score is integrable.

integrable_pair :
∀ i, Integrable (fun ω => v i ω * w i ω) ρ

Each within-record overlap product is integrable.

centered_left :
∀ i, ∫ ω, v i ω ∂ρ = 0

Each first score has zero mean.

centered_right :
∀ i, ∫ ω, w i ω ∂ρ = 0

Each second score has zero mean.

overlap :
∀ i, ∫ ω, v i ω * w i ω ∂ρ = z

Every record has overlap z.

Causalean.Stat.Minimax.Mixture.PoissonLatentSign.CenteredOutcomePair · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Outcome.lean:49
theorem integral_outcomeLikelihood_pair_by_card reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

Given independent centered paired outcome scores with overlap, a bounded tilt of absolute value at most 1/4, cell scores of absolute value at most one, and an outcome amplitude γ, the mean of the product of the two posterior outcome likelihoods equals 1 + Σ over d = 1, …, m of γ^(2d) · (order-d subset energy of the cell) · zᵈ, where m is the number of records and z the common overlap.

Formal statement
m :
ℕ
Ω :
v w :
Fin m → Ω → ℝ
z :
ℝ
h :
τ :
ℝ
hτ :
|τ| ≤ 1 / 4
c :
hu :
∀ i, |c.1 i| ≤ 1
γ :
ℝ
(∫ ω, outcomeLikelihood τ γ c (fun i => v i ω) * outcomeLikelihood τ γ c (fun i => w i ω) ∂ρ)
= 1 + ∑ d ∈ range m, γ ^ (2 * (d + 1)) * subsetEnergy τ c (d + 1) * z ^ (d + 1)
Proof (Lean source)
theorem integral_outcomeLikelihood_pair_by_card {m : ℕ} {Ω : Type*} [MeasurableSpace Ω] {ρ : Measure Ω} [IsProbabilityMeasure ρ] {v w : Fin m → Ω → ℝ} {z : ℝ} (h : CenteredOutcomePair ρ v w z) {τ : ℝ} (hτ : |τ| ≤ 1 / 4) {c : Cell m} (hu : ∀ i, |c.1 i| ≤ 1) (γ : ℝ) : (∫ ω, outcomeLikelihood τ γ c (fun i => v i ω) * outcomeLikelihood τ γ c (fun i => w i ω) ∂ρ) = 1 + ∑ d ∈ range m, γ ^ (2 * (d + 1)) * subsetEnergy τ c (d + 1) * z ^ (d + 1) := by -- Import Mathlib.Algebra.BigOperators.Group.Finset.Powerset for -- Finset.sum_powerset; group the preceding identity by subset cardinality. -- Extend the range to m+1 since selected c ⊆ univ; powersetCard is empty -- above selected.card. Split order zero using posterior_empty hτ hu. classical rw [integral_outcomeLikelihood_pair h τ γ c, Finset.sum_powerset] let F : ℕ → ℝ := fun d => γ ^ (2 * d) * subsetEnergy τ c d * z ^ d have hg (d : ℕ) : (∑ S ∈ (selected c).powersetCard d, γ ^ (2 * S.card) * posterior τ c S ^ 2 * z ^ S.card) = F d := by dsimp [F] unfold subsetEnergy rw [Finset.mul_sum, Finset.sum_mul] apply Finset.sum_congr rfl intro S hS rw [(Finset.mem_powersetCard.mp hS).2] simp_rw [hg] have hc : (selected c).card ≤ m := by simpa using Finset.card_le_card (Finset.subset_univ (selected c)) have hz (d : ℕ) (hd : (selected c).card < d) : F d = 0 := by simp [F, subsetEnergy, Finset.powersetCard_eq_empty.mpr hd] have hext : (∑ d ∈ range ((selected c).card + 1), F d) = ∑ d ∈ range (m + 1), F d := by apply Finset.sum_subset (Finset.range_mono (Nat.add_le_add_right hc 1)) intro d hd hnot apply hz have hn : ¬ d < (selected c).card + 1 := by simpa only [Finset.mem_range] using hnot omega rw [hext, Finset.sum_range_succ'] have h0 : F 0 = 1 := by simp [F, subsetEnergy, posterior_empty hτ hu] rw [h0] dsimp [F] ring
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.integral_outcomeLikelihood_pair_by_card · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Outcome.lean:247 · uses Cell , CenteredOutcomePair , outcomeLikelihood , subsetEnergy
5 supporting declarations (lemmas, instances)
  • outcomeLikelihood_expansion theorem — The outcome likelihood expands over all selected subsets with their posterior coefficients.
    m :
    ℕ
    τ γ :
    ℝ
    c :
    v :
    Fin m → ℝ
    = ∑ S ∈ (selected c).powerset, γ ^ S.card * posterior τ c S * ∏ i ∈ S, v i
    Proof (Lean source)
    theorem outcomeLikelihood_expansion {m : ℕ} (τ γ : ℝ) (c : Cell m) (v : Fin m → ℝ) : outcomeLikelihood τ γ c v = ∑ S ∈ (selected c).powerset, γ ^ S.card * posterior τ c S * ∏ i ∈ S, v i := by -- Finset.prod_one_add followed by distributing the posterior weights. -- This algebraic identity needs no denominator nonzero assumption: distribute -- division across the finite sum rather than canceling the denominator. classical unfold outcomeLikelihood posterior simp only [Finset.prod_one_add, Finset.prod_mul_distrib, Finset.prod_const, Finset.mul_sum] rw [← Finset.sum_add_distrib, Finset.sum_div] apply Finset.sum_congr rfl intro S _ ring
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.outcomeLikelihood_expansion · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Outcome.lean:32
  • integrable_subset_pair theorem — Products over two finite subsets are integrable for independent centered outcome pairs.
    ι Ω :
    v w :
    ι → Ω → ℝ
    z :
    ℝ
    h :
    S R :
    Integrable (fun ω => (∏ i ∈ S, v i ω) * ∏ i ∈ R, w i ω) ρ
    Proof (Lean source)
    theorem integrable_subset_pair {ι Ω : Type*} [MeasurableSpace Ω] {ρ : Measure Ω} [IsProbabilityMeasure ρ] {v w : ι → Ω → ℝ} {z : ℝ} (h : CenteredOutcomePair ρ v w z) (S R : Finset ι) : Integrable (fun ω => (∏ i ∈ S, v i ω) * ∏ i ∈ R, w i ω) ρ := by -- Regroup as a product on S∪R of v, w, or vw; reuse independent integration. -- Use the measurable map (x,y) ↦ (if i∈S then x else 1) * -- (if i∈R then y else 1) on each jointly independent pair. The integrability -- fields cover all membership cases; restrict the resulting family to S∪R. classical simp_rw [subset_pair_regroup S R] exact integrable_independent_finset (independent_union_factors h S R) (integrable_union_factor h S R) (S ∪ R)
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.integrable_subset_pair · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Outcome.lean:122
  • integral_subset_pair theorem — Distinct subset products are orthogonal; equal subsets have cross moment z to their size.
    ι Ω :
    v w :
    ι → Ω → ℝ
    z :
    ℝ
    h :
    S R :
    (∫ ω, (∏ i ∈ S, v i ω) * ∏ i ∈ R, w i ω ∂ρ) = if S = R then z ^ S.card else 0
    Proof (Lean source)
    theorem integral_subset_pair {ι Ω : Type*} [DecidableEq ι] [MeasurableSpace Ω] {ρ : Measure Ω} [IsProbabilityMeasure ρ] {v w : ι → Ω → ℝ} {z : ℝ} (h : CenteredOutcomePair ρ v w z) (S R : Finset ι) : (∫ ω, (∏ i ∈ S, v i ω) * ∏ i ∈ R, w i ω ∂ρ) = if S = R then z ^ S.card else 0 := by -- For the same regrouped union product, use -- iIndepFun.integral_fun_prod_eq_prod_integral on the union subtype. -- When S≠R, a symmetric-difference index contributes a centered zero factor. classical let f : ι → Ω → ℝ := fun i ω => (if i ∈ S then v i ω else 1) * (if i ∈ R then w i ω else 1) have hi : ∀ i, Integrable (f i) ρ := integrable_union_factor h S R have hind : iIndepFun f ρ := independent_union_factors h S R have hfactor : (∫ ω, (∏ i ∈ S, v i ω) * ∏ i ∈ R, w i ω ∂ρ) = ∏ i ∈ S ∪ R, ∫ ω, f i ω ∂ρ := by simp_rw [subset_pair_regroup S R] have hind' := hind.precomp (g := fun i : ↥(S ∪ R) => (i : ι)) val_injective change (∫ ω, ∏ i ∈ S ∪ R, f i ω ∂ρ) = _ simp_rw [← Finset.prod_coe_sort (S ∪ R)] exact hind'.integral_fun_prod_eq_prod_integral (fun i => (hi i).aestronglyMeasurable) rw [hfactor] by_cases heq : S = R · subst R simp only [Finset.union_self] calc (∏ i ∈ S, ∫ ω, f i ω ∂ρ) = ∏ _i ∈ S, z := by apply Finset.prod_congr rfl intro i hiS simpa [f, hiS] using h.overlap i _ = z ^ S.card := Finset.prod_const _ · rw [if_neg heq] have hex : ∃ i, (i ∈ S ∧ i ∉ R) ∨ (i ∈ R ∧ i ∉ S) := by by_contra hn apply heq ext i have hh := (not_exists.mp hn) i constructor <;> intro hi · by_contra hr; exact hh (inl ⟨hi, hr⟩) · by_contra hs; exact hh (inr ⟨hi, hs⟩) obtain ⟨i, hi⟩ := hex apply Finset.prod_eq_zero (i := i) · rcases hi with hi | hi · exact mem_union_left R hi.1 · exact mem_union_right S hi.1 · rcases hi with hi | hi · simpa [f, hi.1, hi.2] using h.centered_left i · simpa [f, hi.1, hi.2] using h.centered_right i
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.integral_subset_pair · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Outcome.lean:136
  • integrable_outcomeLikelihood_pair theorem — The two posterior outcome likelihoods have an integrable product under centered independent outcome pairs, without requiring bounded outcome scores.
    m :
    ℕ
    Ω :
    v w :
    Fin m → Ω → ℝ
    z :
    ℝ
    h :
    τ γ :
    ℝ
    c :
    Integrable (fun ω => outcomeLikelihood τ γ c (fun i => v i ω) * outcomeLikelihood τ γ c (fun i => w i ω)) ρ
    Proof (Lean source)
    theorem integrable_outcomeLikelihood_pair {m : ℕ} {Ω : Type*} [MeasurableSpace Ω] {ρ : Measure Ω} [IsProbabilityMeasure ρ] {v w : Fin m → Ω → ℝ} {z : ℝ} (h : CenteredOutcomePair ρ v w z) (τ γ : ℝ) (c : Cell m) : Integrable (fun ω => outcomeLikelihood τ γ c (fun i => v i ω) * outcomeLikelihood τ γ c (fun i => w i ω)) ρ := by -- Expand both finite likelihood sums and use integrable_subset_pair for each -- double-sum term. No boundedness of outcome scores is needed. classical simp_rw [outcomeLikelihood_expansion, Finset.sum_mul, Finset.mul_sum] apply integrable_finsetSum intro S _ apply integrable_finsetSum intro R _ convert (integrable_subset_pair h S R).const_mul ((γ ^ S.card * posterior τ c S) * (γ ^ R.card * posterior τ c R)) using 1 <;> first | rfl | (funext ω; ring)
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.integrable_outcomeLikelihood_pair · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Outcome.lean:185
  • integral_outcomeLikelihood_pair theorem — The finite-cell cross moment is the sum of squared posterior subset coefficients times the outcome amplitude and overlap powers.
    m :
    ℕ
    Ω :
    v w :
    Fin m → Ω → ℝ
    z :
    ℝ
    h :
    τ γ :
    ℝ
    c :
    (∫ ω, outcomeLikelihood τ γ c (fun i => v i ω) * outcomeLikelihood τ γ c (fun i => w i ω) ∂ρ)
    = ∑ S ∈ (selected c).powerset, γ ^ (2 * S.card) * posterior τ c S ^ 2 * z ^ S.card
    Proof (Lean source)
    theorem integral_outcomeLikelihood_pair {m : ℕ} {Ω : Type*} [MeasurableSpace Ω] {ρ : Measure Ω} [IsProbabilityMeasure ρ] {v w : Fin m → Ω → ℝ} {z : ℝ} (h : CenteredOutcomePair ρ v w z) (τ γ : ℝ) (c : Cell m) : (∫ ω, outcomeLikelihood τ γ c (fun i => v i ω) * outcomeLikelihood τ γ c (fun i => w i ω) ∂ρ) = ∑ S ∈ (selected c).powerset, γ ^ (2 * S.card) * posterior τ c S ^ 2 * z ^ S.card := by -- Integrate the finite double expansion using its termwise receipts, then -- integral_subset_pair eliminates every off-diagonal term. classical let a : Finset (Fin m) → ℝ := fun S => γ ^ S.card * posterior τ c S have ht (S R : Finset (Fin m)) : Integrable (fun ω => (a S * ∏ i ∈ S, v i ω) * (a R * ∏ i ∈ R, w i ω)) ρ := by convert (integrable_subset_pair h S R).const_mul (a S * a R) using 1 funext ω ring have hint (S R : Finset (Fin m)) : (∫ ω, (a S * ∏ i ∈ S, v i ω) * (a R * ∏ i ∈ R, w i ω) ∂ρ) = (a S * a R) * (if S = R then z ^ S.card else 0) := by have heq : (fun ω => (a S * ∏ i ∈ S, v i ω) * (a R * ∏ i ∈ R, w i ω)) = (fun ω => (a S * a R) * ((∏ i ∈ S, v i ω) * ∏ i ∈ R, w i ω)) := by funext ω ring rw [heq, integral_const_mul, integral_subset_pair h S R] simp_rw [outcomeLikelihood_expansion, Finset.sum_mul, Finset.mul_sum] change (∫ ω, ∑ S ∈ (selected c).powerset, ∑ R ∈ (selected c).powerset, (a S * ∏ i ∈ S, v i ω) * (a R * ∏ i ∈ R, w i ω) ∂ρ) = _ rw [integral_finsetSum _ (fun S _ => integrable_finsetSum _ (fun R _ => ht S R))] apply Finset.sum_congr rfl intro S hS rw [integral_finsetSum _ (fun R _ => ht S R)] simp_rw [hint] rw [Finset.sum_eq_single S] · dsimp [a] simp only [if_true, pow_mul, pow_two] ring · intro R _ hRS rw [if_neg (Ne.symm hRS), mul_zero] · intro hnot exact (hnot hS).elim
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.integral_outcomeLikelihood_pair · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Outcome.lean:204
Assembly 3 core · 7 supporting Iterated conditional outcome, covariate, and Poisson-count integration gives the overlap series. ★ poisson_crossMoment_eq_series

Poisson cell cross-moment assembly

Iterated conditional outcome, covariate, and Poisson-count integration gives the overlap series. The finite polynomial and actual likelihood-product interfaces both carry integrability receipts. No paper-local model is imported.

def cellOverlap reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The finite cell overlap polynomial of a cell with m records, at tilt τ, outcome amplitude γ, and overlap z, is 1 + Σ over d = 1, …, m of γ^(2d) · (order-d subset energy of the cell) · zᵈ.

Definition (Lean source)
m :
ℕ
τ γ z :
ℝ
c :
cellOverlap τ γ z c :
ℝ
1 + ∑ d ∈ range m, γ ^ (2 * (d + 1)) * subsetEnergy τ c (d + 1) * z ^ (d + 1)
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.cellOverlap · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Assembly.lean:21 · uses Cell
def conditionalCrossMoment reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

The conditional cross moment of a cell is the integral, under that cell's outcome-score law, of the product of the two posterior outcome likelihoods evaluated at the two outcome-score families.

Definition (Lean source)
Ω :
ρ :
∀ m
if
then
v w :
∀ m
if
and
Fin m
and
Ω
then
ℝ
τ γ :
ℝ
m :
ℕ
c :
conditionalCrossMoment ρ v w τ γ m c :
ℝ
∫ ω, outcomeLikelihood τ γ c (fun i => v m c i ω) * outcomeLikelihood τ γ c (fun i => w m c i ω) ∂ρ m c
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.conditionalCrossMoment · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/Assembly.lean:176 · uses Cell
theorem poisson_crossMoment_eq_series reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

Given conditional probability laws for cells, conditional probability laws for outcome scores, two outcome-score families, a Poisson mean ξ, a tilt of absolute value at most 1/4, cells that are almost surely valid at every count, outcome-score pairs that are almost surely centered and independent across records with common overlap z, and an outcome amplitude, averaging the conditional cross moment over the cell law and then over a Poisson count with mean ξ gives the overlap series 1 + Σ over d ≥ 1 of (order-d Poisson coefficient) · zᵈ.

Formal statement
Ω :
ν :
∀ m, Measure (Cell m)
∀ m, IsProbabilityMeasure (ν m)
ρ :
∀ m
if
then
∀ m c, IsProbabilityMeasure (ρ m c)
v w :
∀ m
if
and
Fin m
and
Ω
then
ℝ
ξ :
τ z :
ℝ
hτ :
|τ| ≤ 1 / 4
hν :
∀ m, ∀ᵐ c ∂ν m, Valid c
hout :
∀ m, ∀ᵐ c ∂ν m, CenteredOutcomePair (ρ m c) (v m c) (w m c) z
γ :
ℝ
(∫ m : ℕ, ∫ c, conditionalCrossMoment ρ v w τ γ m c ∂ν m ∂poissonMeasure ξ)
= overlapSeries (coefficient ν ξ τ γ) z
Proof (Lean source)
theorem poisson_crossMoment_eq_series {Ω : Type*} [MeasurableSpace Ω] (ν : ∀ m, Measure (Cell m)) [∀ m, IsProbabilityMeasure (ν m)] (ρ : ∀ m, Cell m → Measure Ω) [∀ m c, IsProbabilityMeasure (ρ m c)] (v w : ∀ m, Cell m → Fin m → Ω → ℝ) (ξ : NNReal) {τ z : ℝ} (hτ : |τ| ≤ 1 / 4) (hν : ∀ m, ∀ᵐ c ∂ν m, Valid c) (hout : ∀ m, ∀ᵐ c ∂ν m, CenteredOutcomePair (ρ m c) (v m c) (w m c) z) (γ : ℝ) : (∫ m : ℕ, ∫ c, conditionalCrossMoment ρ v w τ γ m c ∂ν m ∂poissonMeasure ξ) = overlapSeries (coefficient ν ξ τ γ) z := by rw [← integral_cellOverlap_eq_series ν ξ hτ hν γ z] apply integral_congr_ae apply Filter.Eventually.of_forall intro m apply integral_congr_ae filter_upwards [hν m, hout m] with c hc ho exact conditionalCrossMoment_eq_cellOverlap ρ v w hτ hc.1 ho γ
7 supporting declarations (lemmas, instances)
Poisson­Moments 1 core · 6 supporting These scalar integrability and tilted moment identities support averaging pointwise posterior bounds. ★ integral_poisson_collision_le

Poisson receipts for latent-sign coefficients

These scalar integrability and tilted moment identities support averaging pointwise posterior bounds. They reuse the existing Poisson law, exponential integrability, and descending-factorial infrastructure.

theorem integral_poisson_collision_le reviewed
Causalean.Stat.Minimax.Mixture.PoissonLatentSign

When the Poisson mean ξ is at most one, the mean of m(m − 1)²(25/9)ᵐ over a Poisson count m is at most e^((5/3)² − 1) · ((5/3)⁶ + (5/3)⁴) · ξ².

Formal statement
ξ :
hξ :
(ξ : ℝ) ≤ 1
(∫ m : ℕ, (m : ℝ) * ((m - 1 : ℕ) : ℝ) ^ 2 * ((5 / 3 : ℝ) ^ 2) ^ m ∂poissonMeasure ξ)
≤ exp ((5 / 3 : ℝ) ^ 2 - 1) * ((5 / 3 : ℝ) ^ 6 + (5 / 3 : ℝ) ^ 4) * (ξ : ℝ) ^ 2
Proof (Lean source)
theorem integral_poisson_collision_le {ξ : NNReal} (hξ : (ξ : ℝ) ≤ 1) : (∫ m : ℕ, (m : ℝ) * ((m - 1 : ℕ) : ℝ) ^ 2 * ((5 / 3 : ℝ) ^ 2) ^ m ∂poissonMeasure ξ) ≤ exp ((5 / 3 : ℝ) ^ 2 - 1) * ((5 / 3 : ℝ) ^ 6 + (5 / 3 : ℝ) ^ 4) * (ξ : ℝ) ^ 2 := by rw [integral_poisson_collision_tilt ξ (by positivity)] have hx : 0 ≤ (ξ : ℝ) := ξ.coe_nonneg have hc : (ξ : ℝ) ^ 3 ≤ (ξ : ℝ) ^ 2 := by nlinarith [sq_nonneg (ξ : ℝ)] have he : exp ((ξ : ℝ) * ((5 / 3 : ℝ) ^ 2 - 1)) ≤ exp ((5 / 3 : ℝ) ^ 2 - 1) := by apply Real.exp_le_exp.mpr nlinarith calc _ ≤ (((ξ : ℝ) * ((5 / 3 : ℝ) ^ 2)) ^ 3 + ((ξ : ℝ) * ((5 / 3 : ℝ) ^ 2)) ^ 2) * exp ((5 / 3 : ℝ) ^ 2 - 1) := mul_le_mul_of_nonneg_left he (by positivity) _ ≤ _ := by have hp : (((ξ : ℝ) * ((5 / 3 : ℝ) ^ 2)) ^ 3 + ((ξ : ℝ) * ((5 / 3 : ℝ) ^ 2)) ^ 2) ≤ ((5 / 3 : ℝ) ^ 6 + (5 / 3 : ℝ) ^ 4) * (ξ : ℝ) ^ 2 := by nlinarith have hh := mul_le_mul_of_nonneg_right hp (Real.exp_pos ((5 / 3 : ℝ) ^ 2 - 1)).le simpa only [mul_assoc, mul_comm, mul_left_comm] using hh
Causalean.Stat.Minimax.Mixture.PoissonLatentSign.integral_poisson_collision_le · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/PoissonMoments.lean:154
6 supporting declarations (lemmas, instances)
  • integrable_poisson_choose theorem — Every binomial coefficient in a Poisson count is integrable.
    ξ :
    d :
    ℕ
    Integrable (fun m : ℕ => (m.choose d : ℝ)) (poissonMeasure ξ)
    Proof (Lean source)
    theorem integrable_poisson_choose (ξ : NNReal) (d : ℕ) : Integrable (fun m : ℕ => (m.choose d : ℝ)) (poissonMeasure ξ) := by have h := (poisson_descFactorial_integrable ξ d).div_const (d.factorial : ℝ) convert h using 1 funext m rw [Nat.descFactorial_eq_factorial_mul_choose, Nat.cast_mul] simp [Nat.factorial_ne_zero]
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.integrable_poisson_choose · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/PoissonMoments.lean:22
  • integral_poisson_choose theorem — The expected order-d binomial coefficient of a Poisson count is the mean to d divided by d factorial.
    ξ :
    d :
    ℕ
    (∫ m : ℕ, (m.choose d : ℝ) ∂poissonMeasure ξ) = (ξ : ℝ) ^ d / (d.factorial : ℝ)
    Proof (Lean source)
    theorem integral_poisson_choose (ξ : NNReal) (d : ℕ) : (∫ m : ℕ, (m.choose d : ℝ) ∂poissonMeasure ξ) = (ξ : ℝ) ^ d / (d.factorial : ℝ) := by calc (∫ m : ℕ, (m.choose d : ℝ) ∂poissonMeasure ξ) = ∫ m : ℕ, (m.descFactorial d : ℝ) / (d.factorial : ℝ) ∂poissonMeasure ξ := by congr 1 funext m rw [Nat.descFactorial_eq_factorial_mul_choose, Nat.cast_mul] simp [Nat.factorial_ne_zero] _ = _ := by rw [integral_div, poisson_descFactorial_moment]
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.integral_poisson_choose · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/PoissonMoments.lean:31
  • integrable_poisson_tilted_factorial theorem — A descending factorial times a nonnegative exponential tilt is Poisson-integrable.
    ξ :
    a :
    ℝ
    ha :
    0 ≤ a
    d :
    ℕ
    Integrable (fun m : ℕ => (m.descFactorial d : ℝ) * a ^ m) (poissonMeasure ξ)
    Proof (Lean source)
    theorem integrable_poisson_tilted_factorial (ξ : NNReal) {a : ℝ} (ha : 0 ≤ a) (d : ℕ) : Integrable (fun m : ℕ => (m.descFactorial d : ℝ) * a ^ m) (poissonMeasure ξ) := by let b : NNReal := ⟨a, ha⟩ have hs := integrable_poissonMeasure_iff.mp (poisson_descFactorial_integrable (ξ * b) d) rw [integrable_poissonMeasure_iff] apply (hs.mul_left (exp ((ξ : ℝ) * (a - 1)))).congr intro m simp only [Real.norm_eq_abs] rw [abs_of_nonneg (by positivity : 0 ≤ (m.descFactorial d : ℝ) * a ^ m), abs_of_nonneg (Nat.cast_nonneg _)] have hw := poisson_weight_tilt ξ b m change _ * a ^ m = exp ((ξ : ℝ) * (a - 1)) * (exp (-((ξ * b : NNReal) : ℝ)) * ((ξ * b : NNReal) : ℝ) ^ m / (m.factorial : ℝ)) at hw calc _ = (exp ((ξ : ℝ) * (a - 1)) * (exp (-((ξ * b : NNReal) : ℝ)) * ((ξ * b : NNReal) : ℝ) ^ m / (m.factorial : ℝ))) * (m.descFactorial d : ℝ) := by ring _ = _ := by rw [← hw]; ring
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.integrable_poisson_tilted_factorial · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/PoissonMoments.lean:63
  • integral_poisson_tilted_factorial theorem — The tilted factorial moment equals the tilted mean to its order times the Poisson exponential factor.
    ξ :
    a :
    ℝ
    ha :
    0 ≤ a
    d :
    ℕ
    (∫ m : ℕ, (m.descFactorial d : ℝ) * a ^ m ∂poissonMeasure ξ)
    = ((ξ : ℝ) * a) ^ d * exp ((ξ : ℝ) * (a - 1))
    Proof (Lean source)
    theorem integral_poisson_tilted_factorial (ξ : NNReal) {a : ℝ} (ha : 0 ≤ a) (d : ℕ) : (∫ m : ℕ, (m.descFactorial d : ℝ) * a ^ m ∂poissonMeasure ξ) = ((ξ : ℝ) * a) ^ d * exp ((ξ : ℝ) * (a - 1)) := by let b : NNReal := ⟨a, ha⟩ rw [integral_poissonMeasure] simp only [smul_eq_mul] calc _ = exp ((ξ : ℝ) * (a - 1)) * ∑' m : ℕ, exp (-((ξ * b : NNReal) : ℝ)) * ((ξ * b : NNReal) : ℝ) ^ m / (m.factorial : ℝ) * (m.descFactorial d : ℝ) := by rw [← tsum_mul_left] apply tsum_congr intro m have hw := poisson_weight_tilt ξ b m change _ * a ^ m = exp ((ξ : ℝ) * (a - 1)) * (exp (-((ξ * b : NNReal) : ℝ)) * ((ξ * b : NNReal) : ℝ) ^ m / (m.factorial : ℝ)) at hw calc _ = (exp (-(ξ : ℝ)) * (ξ : ℝ) ^ m / (m.factorial : ℝ) * a ^ m) * (m.descFactorial d : ℝ) := by ring _ = _ := by rw [hw]; ring _ = _ := by have hm := poisson_descFactorial_moment (ξ * b) d rw [integral_poissonMeasure] at hm simp only [smul_eq_mul] at hm rw [hm] simp only [NNReal.coe_mul] change exp ((ξ : ℝ) * (a - 1)) * ((ξ : ℝ) * a) ^ d = _ ring
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.integral_poisson_tilted_factorial · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/PoissonMoments.lean:87
  • integrable_poisson_collision_tilt theorem — The singleton collision polynomial times a nonnegative exponential tilt is integrable.
    ξ :
    a :
    ℝ
    ha :
    0 ≤ a
    Integrable (fun m : ℕ => (m : ℝ) * ((m - 1 : ℕ) : ℝ) ^ 2 * a ^ m) (poissonMeasure ξ)
    Proof (Lean source)
    theorem integrable_poisson_collision_tilt (ξ : NNReal) {a : ℝ} (ha : 0 ≤ a) : Integrable (fun m : ℕ => (m : ℝ) * ((m - 1 : ℕ) : ℝ) ^ 2 * a ^ m) (poissonMeasure ξ) := by simp_rw [collision_factorial_identity, add_mul] exact (integrable_poisson_tilted_factorial ξ ha 3).add (integrable_poisson_tilted_factorial ξ ha 2)
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.integrable_poisson_collision_tilt · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/PoissonMoments.lean:135
  • integral_poisson_collision_tilt theorem — The tilted singleton collision moment is the sum of its second and third factorial moments.
    ξ :
    a :
    ℝ
    ha :
    0 ≤ a
    (∫ m : ℕ, (m : ℝ) * ((m - 1 : ℕ) : ℝ) ^ 2 * a ^ m ∂poissonMeasure ξ)
    = (((ξ : ℝ) * a) ^ 3 + ((ξ : ℝ) * a) ^ 2) * exp ((ξ : ℝ) * (a - 1))
    Proof (Lean source)
    theorem integral_poisson_collision_tilt (ξ : NNReal) {a : ℝ} (ha : 0 ≤ a) : (∫ m : ℕ, (m : ℝ) * ((m - 1 : ℕ) : ℝ) ^ 2 * a ^ m ∂poissonMeasure ξ) = (((ξ : ℝ) * a) ^ 3 + ((ξ : ℝ) * a) ^ 2) * exp ((ξ : ℝ) * (a - 1)) := by simp_rw [collision_factorial_identity, add_mul] rw [integral_add (integrable_poisson_tilted_factorial ξ ha 3) (integrable_poisson_tilted_factorial ξ ha 2), integral_poisson_tilted_factorial ξ ha 3, integral_poisson_tilted_factorial ξ ha 2]
    Causalean.Stat.Minimax.Mixture.PoissonLatentSign.integral_poisson_collision_tilt · Causalean/Stat/Minimax/Mixture/PoissonLatentSign/PoissonMoments.lean:143