§6 — summands , Newey–West HAC estimator, Assumptions 2–3, modes of convergence, mixing -algebras
DefinitionTPG_InferenceObjects of §6 (central limit theorem and variance estimation) and of Appendix A.4 of the source. Time is (the source's ); the data are generated under the experimental policy unless stated otherwise.
- Reward-time summands (
ipw,Bt,Bbar): with ,
so that .
- HAC estimator (
GammaHat,nwWeight,hacEst): for a bandwidth ,
-
Ideal HAC estimator (
Env.meanB,Env.GammaIdeal,Env.hacIdeal): the same construction with replaced by , , giving and . -
Assumption 2, asymptotic variance (
Env.normVar,Env.AsymptoticVariance k v): as , and . The number is the source's . -
Assumption 3, Cesàro- mean stability (
Env.cesaroDevSq,Env.MeanStability k): with ,
-
Bandwidth conditions (
IsBandwidth L): for all , , and . -
Modes of convergence as for statistics of the horizon- trajectory. Convergence in distribution to (
Env.TendstoNormal): for every bounded continuous . Convergence in probability to a constant (Env.TendstoInProb): for every . -
-algebras for strong mixing (
sigmaWindow,GammaTerm,sigmaGamma): , the local summand , and .
These objects are used in Theorem 2, Lemmas 2–5 and Corollary 1.
Formalization Note For each the statistics are random variables on the horizon- trajectory space with its own law; the laws for different are consistent with one another, and both modes of convergence in item 7 involve only the law of for each . In Assumption 2 the limit is a parameter of the predicate. The normal law is taken with variance .
import Mathlib
import Definitions.Def_TPG_Model
import Definitions.Def_TPG_Estimators
open MeasureTheory ProbabilityTheory
open scoped BigOperators ENNReal
namespace TPG
variable {S : Type} [Fintype S]
/-- Inverse-propensity weight `w^IPW = 2 (2 Z - 1)` under balanced randomization. -/
noncomputable def ipw (z : Bool) : ℝ :=
2 * (2 * (if z then (1 : ℝ) else 0) - 1)
/-- Reward-time summand `B_t = ∑_{max(0, t - k) ≤ i ≤ t} w_i^IPW Y_t`. -/
noncomputable def Bt {T : ℕ} (k : ℕ) (t : Fin T) (ω : Traj S T) : ℝ :=
∑ i : Fin T, if (i : ℕ) ≤ (t : ℕ) ∧ (t : ℕ) ≤ (i : ℕ) + k then ipw (Z i ω) * Y t ω else 0
/-- Sample mean `B̄ = (1/T) ∑_{t < T} B_t`. -/
noncomputable def Bbar (T k : ℕ) (ω : Traj S T) : ℝ :=
(1 / (T : ℝ)) * ∑ t : Fin T, Bt k t ω
/-- Lag-`ℓ` sample autocovariance centered at the sample mean,
`Γ̂_ℓ = (1/T) ∑_{t + ℓ < T} (B_t - B̄)(B_{t+ℓ} - B̄)`. -/
noncomputable def GammaHat (T k ℓ : ℕ) (ω : Traj S T) : ℝ :=
(1 / (T : ℝ)) * ∑ t : Fin T, ∑ t' : Fin T,
if (t : ℕ) + ℓ = (t' : ℕ) then (Bt k t ω - Bbar T k ω) * (Bt k t' ω - Bbar T k ω) else 0
/-- Newey–West (Bartlett) weight `w_ℓ^NW = 1 - ℓ / (L + 1)`. -/
noncomputable def nwWeight (L ℓ : ℕ) : ℝ :=
1 - (ℓ : ℝ) / ((L : ℝ) + 1)
/-- Nonstationary HAC variance estimator with bandwidth `L`,
`Ω̂_T = Γ̂_0 + 2 ∑_{ℓ=1}^{L} w_ℓ^NW Γ̂_ℓ`. -/
noncomputable def hacEst (T k L : ℕ) (ω : Traj S T) : ℝ :=
GammaHat T k 0 ω + 2 * ∑ ℓ ∈ Finset.Icc 1 L, nwWeight L ℓ * GammaHat T k ℓ ω
/-- Time-specific mean `μ_t = E[B_t]` under the experimental policy `π_{1/2}`. -/
noncomputable def Env.meanB [MeasurableSpace S] (E : Env S) (T k : ℕ) (t : Fin T) : ℝ :=
∫ ω, Bt k t ω ∂(E.expLaw T (1 / 2))
/-- Lag-`ℓ` sample autocovariance centered at the time-specific means,
`Γ̃_ℓ = (1/T) ∑_{t + ℓ < T} (B_t - μ_t)(B_{t+ℓ} - μ_{t+ℓ})`. -/
noncomputable def Env.GammaIdeal [MeasurableSpace S] (E : Env S) (T k ℓ : ℕ)
(ω : Traj S T) : ℝ :=
(1 / (T : ℝ)) * ∑ t : Fin T, ∑ t' : Fin T,
if (t : ℕ) + ℓ = (t' : ℕ) then
(Bt k t ω - E.meanB T k t) * (Bt k t' ω - E.meanB T k t') else 0
/-- Ideal HAC estimator `Ω̃_T = Γ̃_0 + 2 ∑_{ℓ=1}^{L} w_ℓ^NW Γ̃_ℓ`. -/
noncomputable def Env.hacIdeal [MeasurableSpace S] (E : Env S) (T k L : ℕ)
(ω : Traj S T) : ℝ :=
E.GammaIdeal T k 0 ω + 2 * ∑ ℓ ∈ Finset.Icc 1 L, nwWeight L ℓ * E.GammaIdeal T k ℓ ω
/-- Normalized variance `(1/T) Var(∑_{t < T} B_t)` under `π_{1/2}`. -/
noncomputable def Env.normVar [MeasurableSpace S] (E : Env S) (T k : ℕ) : ℝ :=
(1 / (T : ℝ)) * variance (fun ω : Traj S T => ∑ t : Fin T, Bt k t ω) (E.expLaw T (1 / 2))
/-- Cesàro-`L²` deviation of the means,
`Δ_T² = (1/T) ∑_{t < T} (μ_t - μ̄_T)²` with `μ̄_T = (1/T) ∑_{t < T} μ_t`. -/
noncomputable def Env.cesaroDevSq [MeasurableSpace S] (E : Env S) (T k : ℕ) : ℝ :=
(1 / (T : ℝ)) * ∑ t : Fin T,
(E.meanB T k t - (1 / (T : ℝ)) * ∑ s : Fin T, E.meanB T k s) ^ 2
/-- The sigma-algebra `F_a^b = σ((X_t, Z_t, Y_t) : a ≤ t ≤ b)` on trajectories. -/
@[instance_reducible]
def sigmaWindow [MeasurableSpace S] (T a b : ℕ) : MeasurableSpace (Traj S T) :=
⨆ t : Fin T, ⨆ _ : a ≤ (t : ℕ) ∧ (t : ℕ) ≤ b,
MeasurableSpace.comap (fun ω : Traj S T => ω t) inferInstance
/-- Local summand `Γ_t = 2 (2 Z_t - 1) ∑_{t ≤ u ≤ t + k, u < T} Y_u`. -/
noncomputable def GammaTerm {T : ℕ} (k : ℕ) (t : Fin T) (ω : Traj S T) : ℝ :=
ipw (Z t ω) * ∑ u : Fin T, if (t : ℕ) ≤ (u : ℕ) ∧ (u : ℕ) ≤ (t : ℕ) + k then Y u ω else 0
/-- The sigma-algebra `σ(Γ_t : a ≤ t ≤ b)` on trajectories. -/
@[instance_reducible]
noncomputable def sigmaGamma (S : Type) (T k a b : ℕ) : MeasurableSpace (Traj S T) :=
⨆ t : Fin T, ⨆ _ : a ≤ (t : ℕ) ∧ (t : ℕ) ≤ b,
MeasurableSpace.comap (fun ω : Traj S T => GammaTerm k t ω) inferInstance
/-- Assumption 2 (asymptotic variance) with limit `v`: the normalized variances
`(1/T) Var(∑_{t < T} B_t)` converge to `v` as `T → ∞`, and `0 < v`. -/
def Env.AsymptoticVariance [MeasurableSpace S] (E : Env S) (k : ℕ) (v : ℝ) : Prop :=
0 < v ∧ Filter.Tendsto (fun T : ℕ => E.normVar T k) Filter.atTop (nhds v)
/-- Assumption 3 (Cesàro-`L²` mean stability): `Δ_T² = o(T^{-2/3})` as `T → ∞`. -/
def Env.MeanStability [MeasurableSpace S] (E : Env S) (k : ℕ) : Prop :=
Asymptotics.IsLittleO Filter.atTop
(fun T : ℕ => E.cesaroDevSq T k) (fun T : ℕ => (T : ℝ) ^ (-(2 / 3) : ℝ))
/-- Bandwidth conditions on the deterministic integer sequence `L_T`:
`0 ≤ L_T < T` for every `T ≥ 1`, `L_T → ∞`, and `L_T = O(T^{1/3})`. -/
def IsBandwidth (L : ℕ → ℕ) : Prop :=
(∀ T : ℕ, 0 < T → L T < T)
∧ Filter.Tendsto L Filter.atTop Filter.atTop
∧ Asymptotics.IsBigO Filter.atTop
(fun T : ℕ => (L T : ℝ)) (fun T : ℕ => (T : ℝ) ^ ((1 : ℝ) / 3))
/-- Convergence in distribution, as the horizon `T → ∞`, of the statistics `W T` computed on a
trajectory of length `T` generated under `π_{1/2}`, to the centered normal law with variance `v`:
`E[f(W_T)] → ∫ f dN(0, v)` for every bounded continuous `f : ℝ → ℝ`. -/
def Env.TendstoNormal [MeasurableSpace S] (E : Env S) (W : (T : ℕ) → Traj S T → ℝ)
(v : ℝ) : Prop :=
∀ f : BoundedContinuousFunction ℝ ℝ,
Filter.Tendsto (fun T : ℕ => ∫ ω, f (W T ω) ∂(E.expLaw T (1 / 2))) Filter.atTop
(nhds (∫ x, f x ∂(gaussianReal 0 (Real.toNNReal v))))
/-- Convergence in probability, as the horizon `T → ∞`, of the statistics `W T` computed on a
trajectory of length `T` generated under `π_{1/2}`, to the constant `c`:
`P(|W_T - c| ≥ ε) → 0` for every `ε > 0`. -/
def Env.TendstoInProb [MeasurableSpace S] (E : Env S) (W : (T : ℕ) → Traj S T → ℝ)
(c : ℝ) : Prop :=
∀ ε : ℝ, 0 < ε →
Filter.Tendsto (fun T : ℕ => (E.expLaw T (1 / 2)) {ω | ε ≤ |W T ω - c|}) Filter.atTop
(nhds 0)
end TPG
Read-back
What the Lean code literally says, in plain math · claude-opus-5-5
Read-back: definition file TPG_Inference
One entry per declaration, in file order. Each entry is written to be read on its own, so the shared background (trajectories, the environment, the measure , the quantities ) is restated wherever it is used.
ipw — Takes a single argument, a flag , and returns the real number
where and . So and . There is no other argument: the value does not depend on any probability parameter, on a state space, or on a horizon.
Bt — Arguments: a type ; a natural number (implicit, read off from the other arguments); a natural number ; an index (indices are zero-based); and a trajectory of length , meaning a function that assigns to each index a triple with , and . (The file's standing declarations also supply a finite-type structure on ; this definition uses neither finiteness of nor any σ-algebra on , and the components do not enter.) The output is the real number
where is if the condition holds and otherwise, and , (in the code, ). Equivalently,
the real entry at index multiplied by the sum of the values attached to the flags at the indices with . That window has indices; it is cut off at index when . For the value is . For there is no index , so the function has empty domain.
Bbar — Arguments: a type ; natural numbers and ; and a trajectory of length , meaning a function that assigns to each index a triple with , and . (The file's standing declarations also supply a finite-type structure on ; this definition uses neither finiteness of nor any σ-algebra on .) The output is the real number
with and . The factor is a division of real numbers; by the proof assistant's convention , so for the value is (and the sum is empty).
GammaHat — Arguments: a type ; natural numbers , and ; and a trajectory of length , meaning a function that assigns to each index a triple with , and . (The file's standing declarations also supply a finite-type structure on ; this definition uses neither finiteness of nor any σ-algebra on .) Put , ,
The output is the real number
that is, times the sum of over those with . Both factors are centred by the same number , the average over all indices. The normalising factor is for every (not one over the number of terms, which is when ). For the value is . For no pair qualifies and the value is . For the value is (empty sums, and by the proof assistant's convention).
nwWeight — Takes two natural numbers and and returns the real number
where and are converted to real numbers before the division and subtraction (so there is no truncation). The denominator is at least and is never zero. No relation between and is required: , , , and for .
hacEst — Arguments: a type ; natural numbers , and ; and a trajectory of length , meaning a function that assigns to each index a triple with , and . (The file's standing declarations also supply a finite-type structure on ; this definition uses neither finiteness of nor any σ-algebra on .) Put , ,
and, for each natural number ,
(the sum of over with , always divided by ). The output is the real number
where the sum runs over the natural numbers with and the weight is computed in the real numbers. No relation between and is assumed: terms with are present but equal to . For the sum is empty and the value is . For the value is (by the convention ).
Env.meanB — Arguments: a finite type carrying a given σ-algebra, which is arbitrary (it is not assumed to be discrete, nor that singletons are measurable); an environment on ; natural numbers and ; and an index . An environment consists of exactly the following data: (i) for every natural number (all of , not only ) and every flag , a real matrix whose entries are all and each of whose rows sums to ; (ii) a function with all values and ; (iii) for every and every flag , a probability measure on the Borel sets of , with no boundedness or moment assumption.
A trajectory of length is a function assigning to each index a triple ; the set of all of them carries the product σ-algebra (the given σ-algebra on , all subsets of , the Borel sets of ). The measure on is the finite sum over all pairs of a state path and a flag path :
where is a Dirac mass and is the product measure over the indices. (The code builds as a product over all indices: a factor for the index ; a factor for every pair of indices with ; and for every a factor equal to if and if , with the constant fixed by this definition. Negative weights would be replaced by ; here every factor is and the weights sum to . For all products are empty, , and is the unit mass on the one-point set .)
The output is the real number
with and . The integral is a Bochner integral, which by the library's convention equals when the integrand is not -integrable; no integrability is assumed. For there is no index , so the function has empty domain.
Env.GammaIdeal — Arguments: a finite type carrying a given σ-algebra, which is arbitrary (it is not assumed to be discrete, nor that singletons are measurable); an environment on ; natural numbers , and ; and a trajectory . An environment consists of exactly the following data: (i) for every natural number (all of , not only ) and every flag , a real matrix whose entries are all and each of whose rows sums to ; (ii) a function with all values and ; (iii) for every and every flag , a probability measure on the Borel sets of , with no boundedness or moment assumption.
A trajectory of length is a function assigning to each index a triple ; the set of all of them carries the product σ-algebra (the given σ-algebra on , all subsets of , the Borel sets of ). The measure on is the finite sum over all pairs of a state path and a flag path :
where is a Dirac mass and is the product measure over the indices. (The code builds as a product over all indices: a factor for the index ; a factor for every pair of indices with ; and for every a factor equal to if and if , with the constant fixed by this definition. Negative weights would be replaced by ; here every factor is and the weights sum to . For all products are empty, , and is the unit mass on the one-point set .)
Put , , and for
the integral being a Bochner integral (equal to by convention if is not -integrable; no integrability is assumed). The output is the real number
that is, times the sum of over those with . Each factor is centred by the number attached to its own index. The normalising factor is for every . For the value is ; for the value is (empty sums, and by the proof assistant's convention). The trajectory is an arbitrary element of ; the measure enters only through the constants .
Env.hacIdeal — Arguments: a finite type carrying a given σ-algebra, which is arbitrary (it is not assumed to be discrete, nor that singletons are measurable); an environment on ; natural numbers , and ; and a trajectory . An environment consists of exactly the following data: (i) for every natural number (all of , not only ) and every flag , a real matrix whose entries are all and each of whose rows sums to ; (ii) a function with all values and ; (iii) for every and every flag , a probability measure on the Borel sets of , with no boundedness or moment assumption.
A trajectory of length is a function assigning to each index a triple ; the set of all of them carries the product σ-algebra (the given σ-algebra on , all subsets of , the Borel sets of ). The measure on is the finite sum over all pairs of a state path and a flag path :
where is a Dirac mass and is the product measure over the indices. (The code builds as a product over all indices: a factor for the index ; a factor for every pair of indices with ; and for every a factor equal to if and if , with the constant fixed by this definition. Negative weights would be replaced by ; here every factor is and the weights sum to . For all products are empty, , and is the unit mass on the one-point set .)
Put , , and for
the integral being a Bochner integral (equal to by convention if is not -integrable; no integrability is assumed). For each natural number put
(the sum of over with , always divided by ). The output is the real number
where the sum runs over the natural numbers with and the weight is computed in the real numbers. No relation between and is assumed: terms with are present but equal to . For the value is . For the value is (by the convention ).
Env.normVar — Arguments: a finite type carrying a given σ-algebra, which is arbitrary (it is not assumed to be discrete, nor that singletons are measurable); an environment on ; and natural numbers and . An environment consists of exactly the following data: (i) for every natural number (all of , not only ) and every flag , a real matrix whose entries are all and each of whose rows sums to ; (ii) a function with all values and ; (iii) for every and every flag , a probability measure on the Borel sets of , with no boundedness or moment assumption.
A trajectory of length is a function assigning to each index a triple ; the set of all of them carries the product σ-algebra (the given σ-algebra on , all subsets of , the Borel sets of ). The measure on is the finite sum over all pairs of a state path and a flag path :
where is a Dirac mass and is the product measure over the indices. (The code builds as a product over all indices: a factor for the index ; a factor for every pair of indices with ; and for every a factor equal to if and if , with the constant fixed by this definition. Negative weights would be replaced by ; here every factor is and the weights sum to . For all products are empty, , and is the unit mass on the one-point set .)
Put , , and let be the function
The output is the real number
Here is the library's real-valued variance: with (a Bochner integral, equal to by convention if is not -integrable), it is the -valued integral converted to a real number, with the value converted to . No moment assumption is made, so both conventions can be active. The sum is not divided by before taking the variance; the variance of the sum is divided by once. For the value is (by the proof assistant's convention ).
Env.cesaroDevSq — Arguments: a finite type carrying a given σ-algebra, which is arbitrary (it is not assumed to be discrete, nor that singletons are measurable); an environment on ; and natural numbers and . An environment consists of exactly the following data: (i) for every natural number (all of , not only ) and every flag , a real matrix whose entries are all and each of whose rows sums to ; (ii) a function with all values and ; (iii) for every and every flag , a probability measure on the Borel sets of , with no boundedness or moment assumption.
A trajectory of length is a function assigning to each index a triple ; the set of all of them carries the product σ-algebra (the given σ-algebra on , all subsets of , the Borel sets of ). The measure on is the finite sum over all pairs of a state path and a flag path :
where is a Dirac mass and is the product measure over the indices. (The code builds as a product over all indices: a factor for the index ; a factor for every pair of indices with ; and for every a factor equal to if and if , with the constant fixed by this definition. Negative weights would be replaced by ; here every factor is and the weights sum to . For all products are empty, , and is the unit mass on the one-point set .)
Put , , and for
the integral being a Bochner integral (equal to by convention if is not -integrable; no integrability is assumed). The output is the real number
the average over the indices of the squared difference between and the average of all the . It is a deterministic number (no trajectory argument). For the value is (empty sums, and by the proof assistant's convention).
sigmaWindow — Arguments: a type carrying a given σ-algebra, which is arbitrary; and natural numbers , and . (The file's standing declarations also supply a finite-type structure on , which this definition does not use.) Let be the set of trajectories of length : functions assigning to each index a triple . The output is a σ-algebra on :
the smallest σ-algebra on containing every set where is an index with (both ends included) and is a measurable subset of for the product σ-algebra (the given σ-algebra on , all subsets of , the Borel sets of ). In the code it is the supremum, over those , of the pull-back σ-algebras of the whole-triple coordinate maps. Only indices below exist, so the effective range is ; may exceed . If no index qualifies (for instance , or , or ), the result is the trivial σ-algebra . (The declaration also carries a reducibility attribute for the proof assistant's instance mechanism, which has no mathematical content.)
GammaTerm — Arguments: a type ; a natural number (implicit, read off from the other arguments); a natural number ; an index (indices are zero-based); and a trajectory of length , meaning a function that assigns to each index a triple with , and . (The file's standing declarations also supply a finite-type structure on ; this definition uses neither finiteness of nor any σ-algebra on , and the components do not enter.) The output is the real number
where is if the condition holds and otherwise, and , (in the code, ). That is: the value attached to the flag at index , multiplied by the sum of the real entries at the indices with . The window is cut off at the last index , so it has terms. For the value is . For there is no index , so the function has empty domain.
sigmaGamma — Arguments: a type , here given explicitly and with no further structure (no finiteness and no σ-algebra on is assumed or used); and natural numbers , , and . Let be the set of trajectories of length : functions assigning to each index a triple . For each index let be
The output is a σ-algebra on :
the smallest σ-algebra on containing every set where is an index with (both ends included) and is a Borel subset of . In the code it is the supremum, over those , of the pull-backs of the Borel σ-algebra of along the real-valued functions . Only indices below exist, so the effective range is . If no index qualifies (for instance , or , or ), the result is the trivial σ-algebra . (The declaration also carries a reducibility attribute for the proof assistant's instance mechanism, which has no mathematical content.)
Env.AsymptoticVariance — This is a property (true or false) of: a finite type carrying a given σ-algebra, which is arbitrary (it is not assumed to be discrete, nor that singletons are measurable); an environment on ; a natural number ; and a real number . An environment consists of exactly the following data: (i) for every natural number and every flag , a real matrix whose entries are all and each of whose rows sums to ; (ii) a function with all values and ; (iii) for every and every flag , a probability measure on the Borel sets of , with no boundedness or moment assumption.
For each natural number , a trajectory of length is a function assigning to each index a triple ; the set of all of them carries the product σ-algebra (the given σ-algebra on , all subsets of , the Borel sets of ). The measure on is the finite sum over all pairs of a state path and a flag path :
where is a Dirac mass and is the product measure over the indices. (The code builds as a product over all indices: a factor for the index ; a factor for every pair of indices with ; and for every a factor equal to if and if , with the constant fixed by this definition. Negative weights would be replaced by ; here every factor is and the weights sum to . For , is the unit mass on the one-point set .)
Put , , and for each
where is the library's real-valued variance: the -valued integral , with a Bochner integral (equal to by convention if is not integrable), converted to a real number with converted to ; and by the convention .
The property holds exactly when both of the following hold:
the limit being taken along the natural numbers with held fixed. Strict positivity of is part of the property itself.
Env.MeanStability — This is a property (true or false) of: a finite type carrying a given σ-algebra, which is arbitrary (it is not assumed to be discrete, nor that singletons are measurable); an environment on ; and a natural number . An environment consists of exactly the following data: (i) for every natural number and every flag , a real matrix whose entries are all and each of whose rows sums to ; (ii) a function with all values and ; (iii) for every and every flag , a probability measure on the Borel sets of , with no boundedness or moment assumption.
For each natural number , a trajectory of length is a function assigning to each index a triple ; the set of all of them carries the product σ-algebra (the given σ-algebra on , all subsets of , the Borel sets of ). The measure on is the finite sum over all pairs of a state path and a flag path :
where is a Dirac mass and is the product measure over the indices. (The code builds as a product over all indices: a factor for the index ; a factor for every pair of indices with ; and for every a factor equal to if and if , with the constant fixed by this definition. Negative weights would be replaced by ; here every factor is and the weights sum to . For , is the unit mass on the one-point set .)
Put , , and for each and each
the integral being a Bochner integral (equal to by convention if the integrand is not -integrable), and
The property holds exactly when as along the natural numbers, in the following precise sense: for every real there is a natural number such that for all ,
Here is the real power of the real number with real exponent (at the library's value of is , which does not affect a statement about all sufficiently large ). The parameter is held fixed.
IsBandwidth — This is a property (true or false) of a single argument, a function . It holds exactly when all three of the following hold:
- for every natural number with , one has (strict; nothing is required of ; for this forces );
- as , meaning: for every natural number there is a natural number such that for all ;
- as , meaning: there exist a real number and a natural number such that for all , where and are regarded as real numbers and is the real power with real exponent .
No monotonicity of is required, and the property does not mention any state space, environment or measure.
Env.TendstoNormal — This is a property (true or false) of: a finite type carrying a given σ-algebra, which is arbitrary (it is not assumed to be discrete, nor that singletons are measurable); an environment on ; a family in which, for each natural number , is an arbitrary real-valued function on the set of trajectories of length (no measurability is assumed); and a real number of either sign. An environment consists of exactly the following data: (i) for every natural number and every flag , a real matrix whose entries are all and each of whose rows sums to ; (ii) a function with all values and ; (iii) for every and every flag , a probability measure on the Borel sets of , with no boundedness or moment assumption.
A trajectory of length is a function assigning to each index a triple ; the set of all of them carries the product σ-algebra (the given σ-algebra on , all subsets of , the Borel sets of ). The measure on is the finite sum over all pairs of a state path and a flag path :
where is a Dirac mass and is the product measure over the indices. (The code builds as a product over all indices: a factor for the index ; a factor for every pair of indices with ; and for every a factor equal to if and if , with the constant fixed by this definition. Negative weights would be replaced by ; here every factor is and the weights sum to . For , is the unit mass on the one-point set .)
The property holds exactly when, for every bounded continuous function ,
the limit being along the natural numbers . Here is the Gaussian measure on with mean and variance : the value is clipped below at , and when the clipped variance is (that is, when ) is the Dirac mass at , so the right-hand side is . The integrals on the left are Bochner integrals, equal to by convention when is not -integrable (which can happen only through failure of measurability, since is bounded). The measure is always the one built with the constant ; the property has no other probability parameter. No positivity of is required here.
Env.TendstoInProb — This is a property (true or false) of: a finite type carrying a given σ-algebra, which is arbitrary (it is not assumed to be discrete, nor that singletons are measurable); an environment on ; a family in which, for each natural number , is an arbitrary real-valued function on the set of trajectories of length (no measurability is assumed); and a real number . An environment consists of exactly the following data: (i) for every natural number and every flag , a real matrix whose entries are all and each of whose rows sums to ; (ii) a function with all values and ; (iii) for every and every flag , a probability measure on the Borel sets of , with no boundedness or moment assumption.
A trajectory of length is a function assigning to each index a triple ; the set of all of them carries the product σ-algebra (the given σ-algebra on , all subsets of , the Borel sets of ). The measure on is the finite sum over all pairs of a state path and a flag path :
where is a Dirac mass and is the product measure over the indices. (The code builds as a product over all indices: a factor for the index ; a factor for every pair of indices with ; and for every a factor equal to if and if , with the constant fixed by this definition. Negative weights would be replaced by ; here every factor is and the weights sum to . For , is the unit mass on the one-point set .)
The property holds exactly when, for every real number with ,
the limit being along the natural numbers , of a sequence of elements of . The inequality inside the set is non-strict (). The set is not required to be measurable; the proof assistant's measure assigns a value to every set (for a non-measurable set, the infimum of the measures of the measurable sets containing it). The limit is a constant real number, and the measure is always the one built with the constant ; the property has no other probability parameter.
Confirmed by the mission captain (proposal self-audit).