Corollary 1 — -mixing of the local summands :
ProvedTPG.corollary_1_gamma_alpha_mixingSetting. is a finite state space and the action is control or treatment. Time is , which is the source's time . Under action at time the state moves according to a transition matrix , i.e. ; the matrices may depend on arbitrarily. The initial state has distribution , and the reward has distribution on , with for all and conditional independence from everything else given . Under the policy every action is treatment with probability , independently of everything else; , and refer to the law of the trajectory under .
Assumption 1 (mixing time) with constant : for every time , every action and every pair of probability vectors on , , where . The kernel deviation is .
Fix a truncation size and define the local summands
Corollary. Under Assumption 1 and a Bernoulli randomization policy , : for all integers with , every and every ,
In the source's notation, with lag , this is : the bound is for and for .
The summands are the terms of the TPG estimator, ; the corollary transfers the mixing rate of the process to them, with a delay equal to the window length.
Formalization Note The supremum defining is unfolded into a statement for each split point and each pair of events; the exponent is a truncated subtraction of natural numbers.
import Mathlib import Definitions.Def_TPG_Model import Definitions.Def_TPG_Estimators import Definitions.Def_TPG_Inference open MeasureTheory ProbabilityTheory open scoped BigOperators ENNReal
namespace TPG
theorem corollary_1_gamma_alpha_mixing {S : Type} [Fintype S] [MeasurableSpace S] [MeasurableSingletonClass S]
(E : Env S) (T k : ℕ) (θ γ M : ℝ)
(hθ0 : 0 < θ) (hθ1 : θ < 1) (hγ0 : 0 < γ) (hγ1 : γ < 1)
(hmix : E.Mixing T γ) (hM : E.RewardBound M)
(s h : ℕ) (hsh : s + h + 1 < T)
(A B : Set (Traj S T))
(hA : MeasurableSet[sigmaGamma S T k 0 s] A)
(hB : MeasurableSet[sigmaGamma S T k (s + h + 1) T] B) :
|(E.expLaw T θ).real (A ∩ B) - (E.expLaw T θ).real A * (E.expLaw T θ).real B|
≤ γ ^ (h - k) := by sorry
end TPGRead-back
What the Lean code literally says, in plain math · claude-opus-5-5
Throughout, denotes the two-element set of truth values.
Objects quantified over. The statement is asserted for all of the following.
- A finite set , equipped with a -algebra in which every one-point set is measurable. (As is finite, every subset of is then measurable.)
- An environment on , which is a bundle of the following data together with the listed conditions:
- for every natural number (all of , not only ) and every , a real matrix with rows and columns indexed by , such that every entry satisfies and every row satisfies ;
- a function with for all and (so cannot be empty);
- for every and , a probability measure on with its Borel -algebra.
- Natural numbers and , real numbers , , , natural numbers , (natural numbers include ), and two sets , of trajectories, as described next.
The trajectory space. Let be the set of all functions from to . For write
carries the product -algebra, where each factor carries the product of the given -algebra on , the -algebra of all subsets of , and the Borel -algebra of . The sets and are subsets of .
The measure on . For a state path and a label path , define the real number as the product of three factors:
- the product over of the quantity that equals if and equals otherwise;
- the product over all ordered pairs of indices in of the quantity that equals if and equals otherwise;
- the product over of the quantity that equals if and equals if .
For this is
So the step from time to time uses the matrix with time index and the label at time , and the same number is used at every time . The measure is the finite sum
where denotes a point mass, the inner is the product measure on , and the outer is the product measure over the time indices. In words: is a mixture over the finitely many pairs , with weight , of the law under which and for every and the real coordinates are independent with distributed according to . Below, denotes the measure of a set read as a real number (value in converted to , with sent to ).
The real-valued functions . For each define by
where , that is, and . The sum runs over the times from up to that exist in the trajectory, i.e. over . Thus depends on only through and , and not on any state coordinate .
Hypotheses.
-
and .
-
and .
-
(One-step contraction, for each label separately.) For every natural number with , every , and every pair of functions that are probability vectors (all values and values summing to ),
-
(Bounded support of the real coordinate.) For every and , , where is the closed interval of reals with .
-
.
-
belongs to the -algebra on generated by the real-valued functions for those with ; that is, the smallest -algebra containing every set with such a and a Borel subset of .
-
belongs to the -algebra on generated in the same way by the functions for those with .
Conclusion.
The exponent is the natural-number (truncated) difference of and : it equals when and equals when . The inequality is non-strict and there is no multiplicative constant in front of the power of .
Edge cases and literal details.
- Whenever the exponent is and the right-hand side is . In particular this is the case for with any .
- Hypothesis 5 forces ; , and are all allowed, and may exceed . For , .
- The index set in hypothesis 6 is . In hypothesis 7 the upper bound is written as , but trajectories only have indices , so that bound never excludes anything and the index set is , which is non-empty by hypothesis 5.
- The two -algebras are generated by the scalar functions only (each a signed sum ), not by the individual coordinates , , .
- The generators of are functions of the coordinates () and (); the generators of are functions of and with . When these two ranges of -times overlap (both contain time ); the statement is asserted in that case too, with right-hand side .
- The number occurs only in hypothesis 4 and not in the conclusion. No sign condition on is written; if the interval is empty and hypothesis 4 cannot hold.
- Hypothesis 3 is imposed on the matrices for and for each of the two labels separately (not on a -weighted average of the two matrices). The measure involves only for ; the matrix with index is constrained by hypothesis 3 but does not enter , and the matrices with index enter neither.
- The clipping is part of the definition of . Under hypothesis 1 and the non-negativity of and of the matrix entries, every is already , so the clipping changes nothing here.
- That is a probability measure is neither a hypothesis nor a conclusion of the statement; it follows from the construction (each product measure has total mass and, by the row-sum and normalisation conditions, the weights sum to ).
Confirmed by the mission captain (proposal self-audit).