Theory Invariants
section ‹The invariant sigma-algebra, Birkhoff theorem›
theory Invariants
imports Recurrence "HOL-Probability.Conditional_Expectation"
begin
subsection ‹The sigma-algebra of invariant subsets›
text ‹The invariant sigma-algebra of a qmpt is made of those sets that are invariant under the
dynamics. When the transformation is ergodic, it is made of sets of zero or full measure. In
general, the Birkhoff theorem is expressed in terms of the conditional expectation of an integrable
function with respect to the invariant sigma-algebra.›
context qmpt begin
text ‹We define the invariant sigma-algebra, as the sigma algebra of sets which are invariant
under the dynamics, i.e., they coincide with their preimage under $T$.›
definition Invariants where "Invariants = sigma (space M) {A ∈ sets M. T-`A ∩ space M = A}"
text ‹For this definition to make sense, we need to check that it really defines a sigma-algebra:
otherwise, the \verb+sigma+ operation would make garbage out of it. This is the content of the next
lemma.›
lemma Invariants_sets: "sets Invariants = {A ∈ sets M. T-`A ∩ space M = A}"
proof -
have "sigma_algebra (space M) {A ∈ sets M. T-`A ∩ space M = A}"
proof -
define I where "I = {A. T-`A ∩ space M = A}"
have i: "⋀A. A ∈ I ⟹ A ⊆ space M" unfolding I_def by auto
have "algebra (space M) I"
proof (subst algebra_iff_Un)
have a: "I ⊆ Pow (space M)" using i by auto
have b: "{} ∈ I" unfolding I_def by auto
{
fix A assume *: "A ∈ I"
then have "T-`(space M - A) = T-`(space M) - T-`A" by auto
then have "T-`(space M - A) ∩ space M = T-`(space M) ∩ (space M) - T-`A ∩ (space M)" by auto
also have "... = space M - A" using * I_def by (simp add: inf_absorb2 subsetI)
finally have "space M - A ∈ I" unfolding I_def by simp
}
then have c: "(∀a∈I. space M - a ∈ I)" by simp
have d: "(∀a∈I. ∀b∈I. a ∪ b ∈ I)" unfolding I_def by auto
show "I ⊆ Pow (space M) ∧ {} ∈ I ∧ (∀a∈I. space M - a ∈ I) ∧ (∀a∈I. ∀b∈I. a ∪ b ∈ I)"
using a b c d by blast
qed
moreover have "(∀F. range F ⊆ I ⟶ (⋃i::nat. F i) ∈ I)" unfolding I_def by auto
ultimately have "sigma_algebra (space M) I" using sigma_algebra_iff by auto
moreover have "sigma_algebra (space M) (sets M)" using measure_space measure_space_def by auto
ultimately have "sigma_algebra (space M) (I ∩ (sets M))" using sigma_algebra_intersection by auto
moreover have "I ∩ sets M = {A ∈ sets M. T-`A ∩ space M = A}" unfolding I_def by auto
ultimately show ?thesis by simp
qed
then show ?thesis unfolding Invariants_def using sigma_algebra.sets_measure_of_eq by blast
qed
text ‹By definition, the invariant subalgebra is a subalgebra of the original algebra. This is
expressed in the following lemmas.›
lemma Invariants_is_subalg: "subalgebra M Invariants"
unfolding subalgebra_def
using Invariants_sets Invariants_def by (simp add: space_measure_of_conv)
lemma Invariants_in_sets:
assumes "A ∈ sets Invariants"
shows "A ∈ sets M"
using Invariants_sets assms by blast
lemma Invariants_measurable_func:
assumes "f ∈ measurable Invariants N"
shows "f ∈ measurable M N"
using Invariants_is_subalg measurable_from_subalg assms by auto
text ‹We give several trivial characterizations of invariant sets or functions.›
lemma Invariants_vrestr:
assumes "A ∈ sets Invariants"
shows "T--`A = A"
using assms Invariants_sets Invariants_in_sets[OF assms] by auto
lemma Invariants_points:
assumes "A ∈ sets Invariants" "x ∈ A"
shows "T x ∈ A"
using assms Invariants_sets by auto
lemma Invariants_func_is_invariant:
fixes f::"_ ⇒ 'b::t2_space"
assumes "f ∈ borel_measurable Invariants" "x ∈ space M"
shows "f (T x) = f x"
proof -
have "{f x} ∈ sets borel" by simp
then have "f-`({f x}) ∩ space M ∈ Invariants" using assms(1)
by (metis (no_types, lifting) Invariants_def measurable_sets space_measure_of_conv)
moreover have "x ∈ f-`({f x}) ∩ space M" using assms(2) by blast
ultimately have "T x ∈ f-`({f x}) ∩ space M" by (rule Invariants_points)
then show ?thesis by simp
qed
lemma Invariants_func_is_invariant_n:
fixes f::"_ ⇒ 'b::t2_space"
assumes "f ∈ borel_measurable Invariants" "x ∈ space M"
shows "f ((T^^n) x) = f x"
by (induction n, auto simp add: assms Invariants_func_is_invariant)
lemma Invariants_func_charac:
assumes [measurable]: "f ∈ measurable M N"
and "⋀x. x ∈ space M ⟹ f(T x) = f x"
shows "f ∈ measurable Invariants N"
proof (rule measurableI)
fix A assume "A ∈ sets N"
have "space Invariants = space M" using Invariants_is_subalg subalgebra_def by force
show "f -` A ∩ space Invariants ∈ sets Invariants"
apply (subst Invariants_sets)
apply (auto simp add: assms ‹A ∈ sets N› ‹space Invariants = space M›)
using ‹A ∈ sets N› assms(1) measurable_sets by blast
next
fix x assume "x ∈ space Invariants"
have "space Invariants = space M" using Invariants_is_subalg subalgebra_def by force
then show "f x ∈ space N" using assms(1) ‹x ∈ space Invariants› by (metis measurable_space)
qed
lemma birkhoff_sum_of_invariants:
fixes f::" _ ⇒ real"
assumes "f ∈ borel_measurable Invariants" "x ∈ space M"
shows "birkhoff_sum f n x = n * f x"
unfolding birkhoff_sum_def using Invariants_func_is_invariant_n[OF assms] by auto
text ‹There are two possible definitions of the invariant sigma-algebra, competing in the
literature: one could also use the sets such that $T^{-1}A$ coincides with $A$ up to
a measure $0$ set. It turns out that this is equivalent to being invariant (in our sense) up
to a measure $0$ set. Therefore, for all interesting purposes, the two definitions would
give the same results.
For the proof, we start from an almost invariant set, and build a genuinely invariant set that
coincides with it by adding or throwing away null parts.
›
proposition Invariants_quasi_Invariants_sets:
assumes [measurable]: "A ∈ sets M"
shows "(∃B ∈ sets Invariants. A Δ B ∈ null_sets M) ⟷ (T--`A Δ A ∈ null_sets M)"
proof
assume "∃B ∈ sets Invariants. A Δ B ∈ null_sets M"
then obtain B where "B ∈ sets Invariants" "A Δ B ∈ null_sets M" by auto
then have [measurable]: "B ∈ sets M" using Invariants_in_sets by simp
have "B = T--` B" using Invariants_vrestr ‹B ∈ sets Invariants› by simp
then have "T--`A Δ B = T--`(A Δ B)" by simp
moreover have "T--`(A Δ B) ∈ null_sets M"
by (rule T_quasi_preserves_null2(1)[OF ‹A Δ B ∈ null_sets M›])
ultimately have "T--`A Δ B ∈ null_sets M" by simp
then show "T--`A Δ A ∈ null_sets M"
by (rule null_sym_diff_transitive) (auto simp add: ‹A Δ B ∈ null_sets M› Un_commute)
next
assume H: "T --` A Δ A ∈ null_sets M"
have [measurable]: "⋀n. (T^^n)--`A ∈ sets M" by simp
{
fix K assume [measurable]: "K ∈ sets M" and "T--`K Δ K ∈ null_sets M"
fix n::nat
have "(T^^n)--`K Δ K ∈ null_sets M"
proof (induction n)
case 0
have "(T^^0)--` K = K" using T_vrestr_0 by simp
then show ?case using Diff_cancel sup.idem by (metis null_sets.empty_sets)
next
case (Suc n)
have "T--`((T^^n)--`K Δ K) ∈ null_sets M"
using Suc.IH T_quasi_preserves_null(1)[of "((T^^n)--`K Δ K)"] by simp
then have *: "(T^^(Suc n))--`K Δ T--`K ∈ null_sets M" using T_vrestr_composed(2)[OF ‹K ∈ sets M›] by simp
then show ?case
by (rule null_sym_diff_transitive, simp add: ‹T--`K Δ K ∈ null_sets M› ‹K ∈ sets M›, measurable)
qed
} note * = this
define C where "C = (⋂n. (T^^n)--`A)"
have [measurable]: "C ∈ sets M" unfolding C_def by simp
have "C Δ A ⊆ (⋃n. (T^^n)--`A Δ A)" unfolding C_def by auto
moreover have "(⋃n. (T^^n)--`A Δ A) ∈ null_sets M"
using * null_sets_UN assms ‹T --` A Δ A ∈ null_sets M› by auto
ultimately have CA: "C Δ A ∈ null_sets M" by (meson ‹C ∈ sets M› assms sets.Diff sets.Un null_sets_subset)
then have "T--`(C Δ A) ∈ null_sets M" by (rule T_quasi_preserves_null2(1))
then have "T--`C Δ T--`A ∈ null_sets M" by simp
then have "T--`C Δ A ∈ null_sets M"
by (rule null_sym_diff_transitive, auto simp add: H)
then have TCC: "T--`C Δ C ∈ null_sets M"
apply (rule null_sym_diff_transitive) using CA by (auto simp add: Un_commute)
have "C ⊆ (⋂n∈{1..}. (T^^n)--`A)" unfolding C_def by auto
moreover have "T--`C = (⋂n∈{1..}. (T^^n)--`A)"
using T_vrestr_composed(2)[OF assms] by (simp add: C_def atLeast_Suc_greaterThan greaterThan_0)
ultimately have "C ⊆ T--`C" by blast
then have "(T^^0)--`C ⊆ (T^^1)--`C" using T_vrestr_0 by auto
moreover have "(T^^1)--`C ⊆ (⋃n∈{1..}. (T^^n)--`C)" by auto
ultimately have "(T^^0)--`C ⊆ (⋃n∈{1..}. (T^^n)--`C)" by auto
then have "(T^^0)--`C ∪ (⋃n∈{1..}. (T^^n)--`C) = (⋃n∈{1..}. (T^^n)--`C)" by auto
moreover have "(⋃n. (T^^n)--`C) = (T^^0)--`C ∪ (⋃n∈{1..}. (T^^n)--`C)" by (rule union_insert_0)
ultimately have C2: "(⋃n. (T^^n)--`C) = (⋃n∈{1..}. (T^^n)--`C)" by simp
define B where "B = (⋃n. (T^^n)--`C)"
have [measurable]: "B ∈ sets M" unfolding B_def by simp
have "B Δ C ⊆ (⋃n. (T^^n)--`C Δ C)" unfolding B_def by auto
moreover have "(⋃n. (T^^n)--`C Δ C) ∈ null_sets M"
using * null_sets_UN assms TCC by auto
ultimately have "B Δ C ∈ null_sets M" by (meson ‹B ∈ sets M› ‹C ∈ sets M› assms sets.Diff sets.Un null_sets_subset)
then have "B Δ A ∈ null_sets M"
by (rule null_sym_diff_transitive, auto simp add: CA)
then have a: "A Δ B ∈ null_sets M" by (simp add: Un_commute)
have "T--`B = (⋃n∈{1..}. (T^^n)--`C)"
using T_vrestr_composed(2)[OF ‹C ∈ sets M›] by (simp add: B_def atLeast_Suc_greaterThan greaterThan_0)
then have "T--`B = B" unfolding B_def using C2 by simp
then have "B ∈ sets Invariants" using Invariants_sets vimage_restr_def by auto
then show "∃B ∈ sets Invariants. A Δ B ∈ null_sets M" using a by blast
qed
text ‹In a conservative setting, it is enough to be included in its image or its preimage to be
almost invariant: otherwise, since the difference set has disjoint preimages, and is therefore
null by conservativity.›
lemma (in conservative) preimage_included_then_almost_invariant:
assumes [measurable]: "A ∈ sets M" and "T--`A ⊆ A"
shows "A Δ (T--`A) ∈ null_sets M"
proof -
define B where "B = A - T--`A"
then have [measurable]: "B ∈ sets M" by simp
have "(T^^(Suc n))--`A ⊆ (T^^n)--`A" for n using T_vrestr_composed(3)[OF assms(1)] vrestr_inclusion[OF assms(2)] by auto
then have "disjoint_family (λn. (T^^n)--`A - (T^^(Suc n))--`A)" by (rule disjoint_family_Suc2[where ?A = "λn. (T^^n)--`A"])
moreover have "(T^^n)--`A - (T^^(Suc n))--`A = (T^^n)--`B" for n unfolding B_def Suc_eq_plus1 using T_vrestr_composed(3)[OF assms(1)] by auto
ultimately have "disjoint_family (λn. (T^^n)--` B)" by simp
then have "⋀n. n ≠ 0 ⟹ ((T^^n)--`B) ∩ B = {}" unfolding disjoint_family_on_def by (metis UNIV_I T_vrestr_0(1)[OF ‹B ∈ sets M›])
then have "⋀n. n > 0 ⟹ (T^^n)-`B ∩ B = {}" unfolding vimage_restr_def by (simp add: Int_assoc)
then have "B ∈ null_sets M" using disjoint_then_null[OF ‹B ∈ sets M›] Int_commute by auto
then show ?thesis unfolding B_def using assms(2) by (simp add: Diff_mono Un_absorb2)
qed
lemma (in conservative) preimage_includes_then_almost_invariant:
assumes [measurable]: "A ∈ sets M" and "A ⊆ T--`A"
shows "A Δ (T--`A) ∈ null_sets M"
proof -
define B where "B = T--`A - A"
then have [measurable]: "B ∈ sets M" by simp
have "⋀n. (T^^(Suc n))--`A ⊇ (T^^n)--`A" using T_vrestr_composed(3)[OF assms(1)] vrestr_inclusion[OF assms(2)] by auto
then have "disjoint_family (λn. (T^^(Suc n))--`A - (T^^n)--`A)" by (rule disjoint_family_Suc[where ?A = "λn. (T^^n)--`A"])
moreover have "⋀n. (T^^(Suc n))--`A - (T^^n)--`A = (T^^n)--`B" unfolding B_def Suc_eq_plus1 using T_vrestr_composed(3)[OF assms(1)] by auto
ultimately have "disjoint_family (λn. (T^^n)--` B)" by simp
then have "⋀n. n ≠ 0 ⟹ ((T^^n)--`B) ∩ B = {}" unfolding disjoint_family_on_def by (metis UNIV_I T_vrestr_0(1)[OF ‹B ∈ sets M›])
then have "⋀n. n > 0 ⟹ (T^^n)-`B ∩ B = {}" unfolding vimage_restr_def by (simp add: Int_assoc)
then have "B ∈ null_sets M" using disjoint_then_null[OF ‹B ∈ sets M›] Int_commute by auto
then show ?thesis unfolding B_def using assms(2) by (simp add: Diff_mono Un_absorb1)
qed
text ‹The above properties for sets are also true for functions: if $f$ and $f \circ T$
coincide almost everywhere, i.e., $f$ is almost invariant, then $f$ coincides almost everywhere
with a true invariant function.
The idea of the proof is straightforward: throw away the orbits on
which $f$ is not really invariant (say this is the complement of the good set),
and replace it by $0$ there. However, this does not work
directly: the good set is not invariant, some points may have a non-constant value of $f$ on their
orbit but reach the good set eventually. One can however define $g$ to be equal to the
eventual value of $f$ along the orbit, if the orbit reaches the good set, and $0$ elsewhere.›
proposition Invariants_quasi_Invariants_functions:
fixes f::"_ ⇒ 'b::{second_countable_topology, t2_space}"
assumes f_meas [measurable]: "f ∈ borel_measurable M"
shows "(∃g ∈ borel_measurable Invariants. AE x in M. f x = g x) ⟷ (AE x in M. f(T x) = f x)"
proof
assume "∃g∈borel_measurable Invariants. AE x in M. f x = g x"
then obtain g where g:"g∈borel_measurable Invariants" "AE x in M. f x = g x" by blast
then have [measurable]: "g ∈ borel_measurable M" using Invariants_measurable_func by auto
define A where "A = {x ∈ space M. f x = g x}"
have [measurable]: "A ∈ sets M" unfolding A_def by simp
define B where "B = space M - A"
have [measurable]: "B ∈ sets M" unfolding B_def by simp
moreover have "AE x in M. x ∉ B" unfolding B_def A_def using g(2) by auto
ultimately have "B ∈ null_sets M" using AE_iff_null_sets by blast
then have "T--`B ∈ null_sets M" by (rule T_quasi_preserves_null2(1))
then have "B ∪ T--`B ∈ null_sets M" using ‹B ∈ null_sets M› by auto
then have "AE x in M. x ∉ (B ∪ T--`B)" using AE_iff_null_sets null_setsD2 by blast
then have i: "AE x in M. x ∈ space M - (B ∪ T--`B)" by auto
{
fix x assume *: "x ∈ space M - (B ∪ T--`B)"
then have "x ∈ A" unfolding B_def by blast
then have "f x = g x" unfolding A_def by blast
have "T x ∈ A" using * B_def by auto
then have "f(T x) = g(T x)" unfolding A_def by blast
moreover have "g(T x) = g x"
apply (rule Invariants_func_is_invariant) using * by (auto simp add: assms ‹g∈borel_measurable Invariants›)
ultimately have "f(T x) = f x" using ‹f x = g x› by simp
}
then show "AE x in M. f(T x) = f x" using i by auto
next
assume *: "AE x in M. f (T x) = f x"
text ‹\verb+good_set+ is the set of points for which $f$ is constant on their orbit. Here, we define
$g = f$. If a point ever enters \verb+good_set+, then we take $g$ to be the value of $f$ there. Otherwise,
$g$ takes an arbitrary value, say $y_0$.›
define good_set where "good_set = {x ∈ space M. ∀n. f((T^^(Suc n)) x) = f((T^^n) x)}"
define good_time where "good_time = (λx. Inf {n. (T^^n) x ∈ good_set})"
have "AE x in M. x ∈ good_set" using T_AE_iterates[OF *] by (simp add: good_set_def)
have [measurable]: "good_set ∈ sets M" unfolding good_set_def by auto
obtain y0::'b where True by auto
define g where "g = (λx. if (∃n. (T^^n) x ∈ good_set) then f((T^^(good_time x)) x) else y0)"
have [measurable]: "good_time ∈ measurable M (count_space UNIV)" unfolding good_time_def by measurable
have [measurable]: "g ∈ borel_measurable M" unfolding g_def by measurable
have "f x = g x" if "x ∈ good_set" for x
proof -
have a: "0 ∈ {n. (T^^n) x ∈ good_set}" using that by simp
have "good_time x = 0"
unfolding good_time_def apply (intro cInf_eq_non_empty) using a by blast+
moreover have "{n. (T^^n) x ∈ good_set} ≠ {}" using a by blast
ultimately show "f x = g x" unfolding g_def by auto
qed
then have "AE x in M. f x = g x" using ‹AE x in M. x ∈ good_set› by auto
have *: "f((T^^(Suc 0)) x) = f((T^^0) x)" if "x ∈ good_set" for x
using that unfolding good_set_def by blast
have good_1: "T x ∈ good_set ∧ f(T x) = f x" if "x ∈ good_set" for x
using *[OF that] that unfolding good_set_def apply (auto)
unfolding T_Tn_T_compose by blast
then have good_k: "⋀x. x ∈ good_set ⟹ (T^^k) x ∈ good_set ∧ f((T^^k) x) = f x" for k
by (induction k, auto)
have "g(T x) = g x" if "x ∈ space M" for x
proof (cases)
assume *: "∃n. (T^^n) (T x) ∈ good_set"
define n where "n = Inf {n. (T^^n) (T x) ∈ good_set}"
have "(T^^n)(T x) ∈ good_set" using * Inf_nat_def1 by (metis empty_iff mem_Collect_eq n_def)
then have a: "(T^^(n+1)) x ∈ good_set" by (metis Suc_eq_plus1 comp_eq_dest_lhs funpow.simps(2) funpow_swap1)
then have **: "∃m. (T^^m) x ∈ good_set" by blast
define m where "m = Inf {m. (T^^m) x ∈ good_set}"
then have "(T^^m) x ∈ good_set" using ** Inf_nat_def1 by (metis empty_iff mem_Collect_eq)
have "n+1 ∈ {m. (T^^m) x ∈ good_set}" using a by simp
then have "m ≤ n+1" using m_def by (simp add: Inf_nat_def Least_le)
then obtain k where "n+1 = m + k" using le_iff_add by blast
have "g x = f((T^^m) x)" unfolding g_def good_time_def using ** m_def by simp
also have "... = f((T^^k) ((T^^m) x))" using ‹(T^^m) x ∈ good_set› good_k by simp
also have "... = f((T^^(n+1))x)" using ‹n+1 = m + k›[symmetric] funpow_add by (metis add.commute comp_apply)
also have "... = f((T^^n) (T x))" using funpow_Suc_right by (metis Suc_eq_plus1 comp_apply)
also have "... = g(T x)" unfolding g_def good_time_def using * n_def by simp
finally show "g(T x) = g x" by simp
next
assume *: "¬(∃n. (T^^n) (T x) ∈ good_set)"
then have "g(T x) = y0" unfolding g_def by simp
have **: "¬(∃n. (T^^(Suc n)) x ∈ good_set)" using funpow_Suc_right * by (metis comp_apply)
have "T x ∉ good_set" using good_k * by blast
then have "x ∉ good_set" using good_1 by auto
then have "¬(∃n. (T^^n) x ∈ good_set)" using ** using good_1 by fastforce
then have "g x = y0" unfolding g_def by simp
then show "g(T x) = g x" using ‹g(T x) = y0› by simp
qed
then have "g ∈ borel_measurable Invariants" by (rule Invariants_func_charac[OF ‹g ∈ borel_measurable M›])
then show "∃g∈borel_measurable Invariants. AE x in M. f x = g x" using ‹AE x in M. f x = g x› by blast
qed
text ‹In a conservative setting, it suffices to have an almost everywhere inequality to get
an almost everywhere equality, as the set where there is strict inequality has $0$ measure
as its iterates are disjoint, by conservativity.›
proposition (in conservative) AE_decreasing_then_invariant:
fixes f::"_ ⇒ 'b::{linorder_topology, second_countable_topology}"
assumes "AE x in M. f(T x) ≤ f x"
and [measurable]: "f ∈ borel_measurable M"
shows "AE x in M. f(T x) = f x"
proof -
obtain D::"'b set" where D: "countable D" "(∀x y. x < y ⟶ (∃d ∈ D. x ≤ d ∧ d < y))"
using countable_separating_set_linorder2 by blast
define A where "A = {x ∈ space M. f(T x) ≤ f x}"
then have [measurable]: "A ∈ sets M" by simp
define B where "B = {x ∈ space M. ∀n. f((T^^(n+1)) x) ≤ f((T^^n)x)}"
then have [measurable]: "B ∈ sets M" by simp
have "space M - A ∈ null_sets M" unfolding A_def using assms by (simp add: assms(1) AE_iff_null_sets)
then have "(⋃n. (T^^n)--`(space M - A)) ∈ null_sets M" by (metis null_sets_UN T_quasi_preserves_null2(2))
moreover have "space M - B = (⋃n. (T^^n)--`(space M - A))"
unfolding B_def A_def by auto
ultimately have "space M - B ∈ null_sets M" by simp
have *: "B = (⋂n. (T^^n)--`A)"
unfolding B_def A_def by auto
then have "T--`B = (⋂n. T--` (T^^n)--`A)" by auto
also have "... = (⋂n. (T^^(n+1))--`A)" using T_vrestr_composed(2)[OF ‹A ∈ sets M›] by simp
also have "... ⊇ (⋂n. (T^^n)--`A)" by blast
finally have B1: "B ⊆ T--`B" using * by simp
have "B ⊆ A" using * T_vrestr_0[OF ‹A ∈ sets M›] by blast
then have B2: "⋀x. x ∈ B ⟹ f(T x) ≤ f x" unfolding A_def by auto
define C where "C = (λt. {x ∈ B. f x ≤ t})"
{
fix t
have "C t = B ∩ f-`{..t} ∩ space M" unfolding C_def using sets.sets_into_space[OF ‹B ∈ sets M›] by auto
then have [measurable]: "C t ∈ sets M" using assms(2) by simp
have "C t ⊆ T--`(C t)" using B1 unfolding C_def vimage_restr_def apply auto using B2 order_trans by blast
then have "T--`(C t) - C t ∈ null_sets M" by (metis Diff_mono Un_absorb1 preimage_includes_then_almost_invariant[OF ‹C t ∈ sets M›])
}
then have "(⋃d∈D. T--`(C d) - C d) ∈ null_sets M" using ‹countable D› by (simp add: null_sets_UN')
then have "(space M - B) ∪ (⋃d∈D. T--`(C d) - C d) ∈ null_sets M" using ‹space M - B ∈ null_sets M› by auto
then have "AE x in M. x ∉ (space M - B) ∪ (⋃d∈D. T--`(C d) - C d)" using AE_not_in by blast
moreover
{
fix x assume x: "x ∈ space M" "x ∉ (space M - B) ∪ (⋃d∈D. T--`(C d) - C d)"
then have "x ∈ B" by simp
then have "T x ∈ B" using B1 by auto
have "f(T x) = f x"
proof (rule ccontr)
assume "f(T x) ≠ f x"
then have "f(T x) < f x" using B2[OF ‹x ∈ B›] by simp
then obtain d where d: "d ∈ D" "f(T x) ≤ d ∧ d < f x" using D by auto
then have "T x ∈ C d" using ‹T x ∈ B› unfolding C_def by simp
then have "x ∈ T--`(C d)" using ‹x ∈ space M› by simp
then have "x ∈ C d" using x ‹d ∈ D› by simp
then have "f x ≤ d" unfolding C_def by simp
then show False using d by auto
qed
}
ultimately show ?thesis by auto
qed
proposition (in conservative) AE_increasing_then_invariant:
fixes f::"_ ⇒ 'b::{linorder_topology, second_countable_topology}"
assumes "AE x in M. f(T x) ≥ f x"
and [measurable]: "f ∈ borel_measurable M"
shows "AE x in M. f(T x) = f x"
proof -
obtain D::"'b set" where D: "countable D" "(∀x y. x < y ⟶ (∃d ∈ D. x < d ∧ d ≤ y))"
using countable_separating_set_linorder1 by blast
define A where "A = {x ∈ space M. f(T x) ≥ f x}"
then have [measurable]: "A ∈ sets M" by simp
define B where "B = {x ∈ space M. ∀n. f((T^^(n+1)) x) ≥ f((T^^n)x)}"
then have [measurable]: "B ∈ sets M" by simp
have "space M - A ∈ null_sets M" unfolding A_def using assms by (simp add: assms(1) AE_iff_null_sets)
then have "(⋃n. (T^^n)--`(space M - A)) ∈ null_sets M" by (metis null_sets_UN T_quasi_preserves_null2(2))
moreover have "space M - B = (⋃n. (T^^n)--`(space M - A))"
unfolding B_def A_def by auto
ultimately have "space M - B ∈ null_sets M" by simp
have *: "B = (⋂n. (T^^n)--`A)"
unfolding B_def A_def by auto
then have "T--`B = (⋂n. T--` (T^^n)--`A)" by auto
also have "... = (⋂n. (T^^(n+1))--`A)" using T_vrestr_composed(2)[OF ‹A ∈ sets M›] by simp
also have "... ⊇ (⋂n. (T^^n)--`A)" by blast
finally have B1: "B ⊆ T--`B" using * by simp
have "B ⊆ A" using * T_vrestr_0[OF ‹A ∈ sets M›] by blast
then have B2: "⋀x. x ∈ B ⟹ f(T x) ≥ f x" unfolding A_def by auto
define C where "C = (λt. {x ∈ B. f x ≥ t})"
{
fix t
have "C t = B ∩ f-`{t..} ∩ space M" unfolding C_def using sets.sets_into_space[OF ‹B ∈ sets M›] by auto
then have [measurable]: "C t ∈ sets M" using assms(2) by simp
have "C t ⊆ T--`(C t)" using B1 unfolding C_def vimage_restr_def apply auto using B2 order_trans by blast
then have "T--`(C t) - C t ∈ null_sets M" by (metis Diff_mono Un_absorb1 preimage_includes_then_almost_invariant[OF ‹C t ∈ sets M›])
}
then have "(⋃d∈D. T--`(C d) - C d) ∈ null_sets M" using ‹countable D› by (simp add: null_sets_UN')
then have "(space M - B) ∪ (⋃d∈D. T--`(C d) - C d) ∈ null_sets M" using ‹space M - B ∈ null_sets M› by auto
then have "AE x in M. x ∉ (space M - B) ∪ (⋃d∈D. T--`(C d) - C d)" using AE_not_in by blast
moreover
{
fix x assume x: "x ∈ space M" "x ∉ (space M - B) ∪ (⋃d∈D. T--`(C d) - C d)"
then have "x ∈ B" by simp
then have "T x ∈ B" using B1 by auto
have "f(T x) = f x"
proof (rule ccontr)
assume "f(T x) ≠ f x"
then have "f(T x) > f x" using B2[OF ‹x ∈ B›] by simp
then obtain d where d: "d ∈ D" "f(T x) ≥ d ∧ d > f x" using D by auto
then have "T x ∈ C d" using ‹T x ∈ B› unfolding C_def by simp
then have "x ∈ T--`(C d)" using ‹x ∈ space M› by simp
then have "x ∈ C d" using x ‹d ∈ D› by simp
then have "f x ≥ d" unfolding C_def by simp
then show False using d by auto
qed
}
ultimately show ?thesis by auto
qed
text ‹For an invertible map, the invariants of $T$ and $T^{-1}$ are the same.›
lemma Invariants_Tinv:
assumes "invertible_qmpt"
shows "qmpt.Invariants M Tinv = Invariants"
proof -
interpret I: qmpt M Tinv using Tinv_qmpt[OF assms] by auto
have "(T -` A ∩ space M = A) ⟷ (Tinv -` A ∩ space M = A)" if "A ∈ sets M" for A
proof
assume "T -` A ∩ space M = A"
then show "Tinv -` A ∩ space M = A"
using assms that unfolding Tinv_def invertible_qmpt_def
apply auto
apply (metis IntE UNIV_I bij_def imageE inv_f_f vimageE)
apply (metis I.T_spaceM_stable(1) Int_iff Tinv_def bij_inv_eq_iff vimageI)
done
next
assume "Tinv -` A ∩ space M = A"
then show "T -` A ∩ space M = A"
using assms that unfolding Tinv_def invertible_qmpt_def
apply auto
apply (metis IntE bij_def inv_f_f vimageE)
apply (metis T_Tinv_of_set T_meas Tinv_def assms qmpt.vrestr_of_set qmpt_axioms vrestr_image(3))
done
qed
then have "{A ∈ sets M. Tinv -` A ∩ space M = A} = {A ∈ sets M. T -` A ∩ space M = A}"
by blast
then show ?thesis unfolding Invariants_def I.Invariants_def by auto
qed
end
sublocale fmpt ⊆ finite_measure_subalgebra M Invariants
unfolding finite_measure_subalgebra_def finite_measure_subalgebra_axioms_def
using Invariants_is_subalg by (simp add: finite_measureI)
context fmpt
begin
text ‹The conditional expectation with respect to the invariant sigma-algebra is the same
for $f$ or $f \circ T$, essentially by definition.›
lemma Invariants_of_foTn:
fixes f::"'a ⇒ real"
assumes [measurable]: "integrable M f"
shows "AE x in M. real_cond_exp M Invariants (f o (T^^n)) x = real_cond_exp M Invariants f x"
proof (rule real_cond_exp_charact)
fix A assume [measurable]: "A ∈ sets Invariants"
then have [measurable]: "A ∈ sets M" using Invariants_in_sets by blast
then have ind_meas [measurable]: "((indicator A)::('a ⇒ real)) ∈ borel_measurable Invariants" by auto
have "set_lebesgue_integral M A (f ∘ (T^^n)) = (∫x. indicator A x * f((T^^n) x) ∂M)"
by (auto simp: comp_def set_lebesgue_integral_def)
also have "... = (∫x. indicator A ((T^^n) x) * f ((T^^n) x) ∂M)"
by (rule Bochner_Integration.integral_cong, auto simp add: Invariants_func_is_invariant_n[OF ind_meas])
also have "... = (∫x. indicator A x * f x ∂M)"
apply (rule Tn_integral_preserving(2)) using integrable_mult_indicator[OF ‹A ∈ sets M› assms] by auto
also have "... = (∫x. indicator A x * real_cond_exp M Invariants f x ∂M)"
apply (rule real_cond_exp_intg(2)[symmetric]) using integrable_mult_indicator[OF ‹A ∈ sets M› assms] by auto
also have "... = set_lebesgue_integral M A (real_cond_exp M Invariants f)"
by (auto simp: set_lebesgue_integral_def)
finally show "set_lebesgue_integral M A (f ∘ (T^^n)) = set_lebesgue_integral M A (real_cond_exp M Invariants f)"
by simp
qed (auto simp add: assms real_cond_exp_int Tn_integral_preserving(1)[OF assms] comp_def)
lemma Invariants_of_foT:
fixes f::"'a ⇒ real"
assumes [measurable]: "integrable M f"
shows "AE x in M. real_cond_exp M Invariants f x = real_cond_exp M Invariants (f o T) x"
using Invariants_of_foTn[OF assms, where ?n = 1] by auto
lemma birkhoff_sum_Invariants:
fixes f::"'a ⇒ real"
assumes [measurable]: "integrable M f"
shows "AE x in M. real_cond_exp M Invariants (birkhoff_sum f n) x = n * real_cond_exp M Invariants f x"
proof -
define F where "F = (λi. f o (T^^i))"
have [measurable]: "⋀i. F i ∈ borel_measurable M" unfolding F_def by auto
have *: "integrable M (F i)" for i unfolding F_def
by (subst comp_def, rule Tn_integral_preserving(1)[OF assms, of i])
have "AE x in M. n * real_cond_exp M Invariants f x = (∑i∈{..<n}. real_cond_exp M Invariants f x)" by auto
moreover have "AE x in M. (∑i∈{..<n}. real_cond_exp M Invariants f x) = (∑i∈{..<n}. real_cond_exp M Invariants (F i) x)"
apply (rule AE_symmetric[OF AE_equal_sum]) unfolding F_def using Invariants_of_foTn[OF assms] by simp
moreover have "AE x in M. (∑i∈{..<n}. real_cond_exp M Invariants (F i) x) = real_cond_exp M Invariants (λx. ∑i∈{..<n}. F i x) x"
by (rule AE_symmetric[OF real_cond_exp_sum [OF *]])
moreover have "AE x in M. real_cond_exp M Invariants (λx. ∑i∈{..<n}. F i x) x = real_cond_exp M Invariants (birkhoff_sum f n) x"
apply (rule real_cond_exp_cong) unfolding F_def using birkhoff_sum_def[symmetric] by auto
ultimately show ?thesis by auto
qed
end
subsection ‹Birkhoff theorem›
subsubsection ‹Almost everywhere version of Birkhoff theorem›
text ‹This paragraph is devoted to the proof of Birkhoff theorem, arguably
the most fundamental result of ergodic theory.
This theorem asserts that Birkhoff averages of an integrable function $f$ converge almost surely,
to the conditional expectation of $f$ with respect to the invariant sigma algebra.
This result implies for instance the strong law of large numbers (in probability theory).
There are numerous proofs of this statement, but none is really easy. We follow the very efficient
argument given in Katok-Hasselblatt. To help the reader, here is the same proof informally. The
first part of the proof is formalized in \verb+birkhoff_lemma1+, the second one in
\verb+birkhoff_lemma+, and the conclusion in \verb+birkhoff_theorem+.
Start with an integrable function $g$. let $G_n(x) = \max_{k\leq n} S_k g(x)$. Then $\limsup S_n g/n
\leq 0$ outside of $A$, the set where $G_n$ tends to infinity. Moreover, $G_{n+1} - G_n \circ T$ is
bounded by $g$, and tends to $g$ on $A$. It follows from the dominated convergence theorem that
$\int_A G_{n+1} - G_n \circ T \to \int_A g$. As $\int_A G_{n+1} - G_n \circ T = \int_A G_{n+1} - G_n
\geq 0$, we obtain $\int_A g \geq 0$.
Apply now this result to the function $g = f - E(f | I) - \epsilon$, where $\epsilon>0$ is fixed.
Then $\int_A g = -\epsilon \mu(A)$, then have $\mu(A) = 0$. Thus, almost surely, $\limsup S_n
g/n\leq 0$, i.e., $\limsup S_n f/n \leq E(f|I)+\epsilon$. Letting $\epsilon$ tend to $0$ gives
$\limsup S_n f/n \leq E(f|I)$.
Applying the same result to $-f$ gives $S_n f/n \to E(f|I)$.
›
context fmpt
begin
lemma birkhoff_aux1:
fixes f::"'a ⇒ real"
assumes [measurable]: "integrable M f"
defines "A ≡ {x ∈ space M. limsup (λn. ereal(birkhoff_sum f n x)) = ∞}"
shows "A ∈ sets Invariants" "(∫x. f x * indicator A x ∂M) ≥ 0"
proof -
let ?bsf = "birkhoff_sum f"
have [measurable]: "A ∈ sets M" unfolding A_def by simp
have Ainv: "x ∈ A ⟷ T x ∈ A" if "x ∈ space M" for x
proof -
have "ereal(?bsf (1 + n) x) = ereal(f x) + ereal(?bsf n (T x))" for n
unfolding birkhoff_sum_cocycle birkhoff_sum_1 by simp
moreover have "limsup (λn. ereal(f x) + ereal(?bsf n (T x)))
= ereal(f x) + limsup(λn. ereal(?bsf n (T x)))"
by (rule ereal_limsup_lim_add, auto)
moreover have "limsup (λn. ereal(?bsf (n+1) x)) = limsup (λn. ereal(?bsf n x))" using limsup_shift by simp
ultimately have "limsup (λn. ereal(birkhoff_sum f n x)) = ereal(f x) + limsup (λn. ereal(?bsf n (T x)))" by simp
then have "limsup (λn. ereal(?bsf n x)) = ∞ ⟷ limsup (λn. ereal(?bsf n (T x))) = ∞" by simp
then show "x ∈ A ⟷ T x ∈ A" using ‹x ∈ space M› A_def by simp
qed
then show "A ∈ sets Invariants" using assms(2) Invariants_sets by auto
define F where "F = (λn x. MAX k ∈{0..n}. ?bsf k x)"
have [measurable]: "⋀n. F n ∈ borel_measurable M" unfolding F_def by measurable
have intFn: "integrable M (F n)" for n
unfolding F_def by (rule integrable_MAX, auto simp add: birkhoff_sum_integral(1)[OF assms(1)])
have Frec: "F (n+1) x - F n (T x) = max (-F n (T x)) (f x)" for n x
proof -
have "{0..n+1} = {0} ∪ {1..n+1}" by auto
then have "(λk. ?bsf k x) ` {0..n+1} = (λk. ?bsf k x) ` {0} ∪ (λk. ?bsf k x) ` {1..n+1}" by blast
then have *: "(λk. ?bsf k x) ` {0..n+1} = {0} ∪ (λk. ?bsf k x) ` {1..n+1}" using birkhoff_sum_1(1) by simp
have b: "F (n+1) x = max (Max {0}) (MAX k ∈{1..n+1}. ?bsf k x)"
by (subst F_def, subst *, rule Max.union, auto)
have "(λk. ?bsf k x) ` {1..n+1} = (λk. ?bsf (1+k) x) ` {0..n}" using Suc_le_D by fastforce
also have "... = (λk. f x + ?bsf k (T x)) ` {0..n}"
by (subst birkhoff_sum_cocycle, subst birkhoff_sum_1(2), auto)
finally have c: "F (n+1) x = max 0 (MAX k ∈{0..n}. ?bsf k (T x) + f x)" using b by (simp add: add_ac)
have "{f x + birkhoff_sum f k (T x) |k. k ∈{0..n}} = (+) (f x) ` {birkhoff_sum f k (T x) |k. k ∈{0..n}}" by blast
have "(MAX k ∈{0..n}. ?bsf k (T x) + f x) = (MAX k ∈{0..n}. ?bsf k (T x)) + f x"
by (rule Max_add_commute) auto
also have "... = F n (T x) + f x" unfolding F_def by simp
finally have "(MAX k ∈{0..n}. ?bsf k (T x) + f x) = f x + F n (T x)" by simp
then have "F (n+1) x = max 0 (f x + F n (T x))" using c by simp
then show "F (n+1) x - F n (T x) = max (-F n (T x)) (f x)" by auto
qed
have a: "abs((F (n+1) x - F n (T x)) * indicator A x) ≤ abs(f x)" for n x
proof -
have "F (n+1) x -F n (T x) ≥ f x" using Frec by simp
then have *: "F (n+1) x -F n (T x) ≥ - abs(f x)" by simp
have "F n (T x) ≥ birkhoff_sum f 0 (T x)"
unfolding F_def apply (rule Max_ge, simp) using atLeastAtMost_iff by blast
then have "F n (T x) ≥ 0" using birkhoff_sum_1(1) by simp
then have "-F n (T x) ≤ abs (f x)" by simp
moreover have "f x ≤ abs(f x)" by simp
ultimately have "F (n+1) x -F n (T x) ≤ abs(f x)" using Frec by simp
then have "abs(F (n+1) x - F n (T x)) ≤ abs(f x)" using * by simp
then show "abs((F (n+1) x - F n (T x)) * indicator A x) ≤ abs(f x)" unfolding indicator_def by auto
qed
have b: "(λn. (F (n+1) x - F n (T x)) * indicator A x) ⇢ f x * indicator A x" for x
proof (rule tendsto_eventually, cases)
assume "x ∈ A"
then have "T x ∈ A" using Ainv A_def by auto
then have "limsup (λn. ereal(birkhoff_sum f n (T x))) > ereal(-f x)" unfolding A_def by simp
then obtain N where "ereal(?bsf N (T x)) > ereal(-f x)" using Limsup_obtain by blast
then have *: "?bsf N (T x) > -f x" by simp
{
fix n assume "n≥N"
then have "?bsf N (T x) ∈ (λk. ?bsf k (T x)) ` {0..n}" by auto
then have "F n (T x) ≥ ?bsf N (T x)" unfolding F_def by simp
then have "F n (T x) ≥ -f x" using * by simp
then have "max (-F n (T x)) (f x) = f x" by simp
then have "F (n+1) x - F n (T x) = f x" using Frec by simp
then have "(F (n+1) x - F n (T x)) * indicator A x = f x * indicator A x" by simp
}
then show "eventually (λn. (F (n+1) x - F n (T x)) * indicator A x = f x * indicator A x) sequentially"
using eventually_sequentially by blast
next
assume "¬(x ∈ A)"
then have "indicator A x = (0::real)" by simp
then show "eventually (λn. (F (n+1) x - F n (T x)) * indicator A x = f x * indicator A x) sequentially" by auto
qed
have lim: "(λn. (∫x. (F (n+1) x - F n (T x)) * indicator A x ∂M)) ⇢ (∫x. f x * indicator A x ∂M)"
proof (rule integral_dominated_convergence[where ?w = "(λx. abs(f x))"])
show "integrable M (λx. ¦f x¦)" using assms(1) by auto
show "AE x in M. (λn. (F (n + 1) x - F n (T x)) * indicator A x) ⇢ f x * indicator A x" using b by auto
show "⋀n. AE x in M. norm ((F (n + 1) x - F n (T x)) * indicator A x) ≤ ¦f x¦" using a by auto
qed (simp_all)
have "(∫x. (F (n+1) x - F n (T x)) * indicator A x ∂M) ≥ 0" for n
proof -
have "(∫x. F n (T x) * indicator A x ∂M) = (∫x. (λx. F n x * indicator A x) (T x) ∂M)"
by (rule Bochner_Integration.integral_cong, auto simp add: Ainv indicator_def)
also have "... = (∫x. F n x * indicator A x ∂M)"
by (rule T_integral_preserving, auto simp add: intFn integrable_real_mult_indicator)
finally have i: "(∫x. F n (T x) * indicator A x ∂M) = (∫x. F n x * indicator A x ∂M)" by simp
have "(∫x. (F (n+1) x - F n (T x)) * indicator A x ∂M) = (∫x. F (n+1) x * indicator A x - F n (T x) * indicator A x ∂M)"
by (simp add: mult.commute right_diff_distrib)
also have "... = (∫x. F (n+1) x * indicator A x ∂M) - (∫x. F n (T x) * indicator A x ∂M)"
by (rule Bochner_Integration.integral_diff, auto simp add: intFn integrable_real_mult_indicator T_meas T_integral_preserving(1))
also have "... = (∫x. F (n+1) x * indicator A x ∂M) - (∫x. F n x * indicator A x ∂M)"
using i by simp
also have "... = (∫x. F (n+1) x * indicator A x - F n x * indicator A x ∂M)"
by (rule Bochner_Integration.integral_diff[symmetric], auto simp add: intFn integrable_real_mult_indicator)
also have "... = (∫x. (F (n+1) x - F n x) * indicator A x ∂M)"
by (simp add: mult.commute right_diff_distrib)
finally have *: "(∫x. (F (n+1) x - F n (T x)) * indicator A x ∂M) = (∫x. (F (n+1) x - F n x) * indicator A x ∂M)"
by simp
have "F n x ≤ F (n+1) x" for x unfolding F_def by (rule Max_mono, auto)
then have "(F (n+1) x - F n x) * indicator A x ≥ 0" for x by simp
then have "integral⇧L M (λx. 0) ≤ integral⇧L M (λx. (F (n+1) x - F n x) * indicator A x)"
by (auto simp add: intFn integrable_real_mult_indicator intro: integral_mono)
then have "(∫x. (F (n+1) x - F n x) * indicator A x ∂M) ≥ 0" by simp
then show "(∫x. (F (n+1) x - F n (T x)) * indicator A x ∂M) ≥ 0" using * by simp
qed
then show "(∫x. f x * indicator A x ∂M) ≥ 0" using lim by (simp add: LIMSEQ_le_const)
qed
lemma birkhoff_aux2:
fixes f::"'a ⇒ real"
assumes [measurable]: "integrable M f"
shows "AE x in M. limsup (λn. ereal(birkhoff_sum f n x / n)) ≤ real_cond_exp M Invariants f x"
proof -
{
fix ε assume "ε > (0::real)"
define g where "g = (λx. f x - real_cond_exp M Invariants f x - ε)"
then have intg: "integrable M g" using assms real_cond_exp_int(1) assms by auto
define A where "A = {x ∈ space M. limsup (λn. ereal(birkhoff_sum g n x)) = ∞}"
have Ag: "A ∈ sets Invariants" "(∫x. g x * indicator A x ∂M) ≥ 0"
unfolding A_def by (rule birkhoff_aux1[where ?f = g, OF intg])+
then have [measurable]: "A ∈ sets M" by (simp add: Invariants_in_sets)
have eq: "(∫x. indicator A x * real_cond_exp M Invariants f x ∂M) = (∫x. indicator A x * f x ∂M)"
proof (rule real_cond_exp_intg[where ?f = "λx. (indicator A x)::real" and ?g = f])
have "(λx. indicator A x * f x) = (λx. f x * indicator A x)" by auto
then show "integrable M (λx. indicator A x * f x)"
using integrable_real_mult_indicator[OF ‹A ∈ sets M› assms] by simp
show "indicator A ∈ borel_measurable Invariants" using ‹A ∈ sets Invariants› by measurable
qed (simp)
have "0 ≤ (∫x. g x * indicator A x ∂M)" using Ag by simp
also have "... = (∫x. f x * indicator A x - real_cond_exp M Invariants f x * indicator A x - ε * indicator A x ∂M)"
unfolding g_def by (simp add: left_diff_distrib)
also have "... = (∫x. f x * indicator A x ∂M) - (∫x. real_cond_exp M Invariants f x * indicator A x ∂M) - (∫x. ε * indicator A x ∂M)"
using assms real_cond_exp_int(1)[OF assms] integrable_real_mult_indicator[OF ‹A ∈ sets M›]
by (auto simp: simp del: integrable_mult_left_iff)
also have "... = - (∫x. ε * indicator A x ∂M)"
by (auto simp add: eq mult.commute)
also have "... = - ε * measure M A" by auto
finally have "0 ≤ - ε * measure M A" by simp
then have "measure M A = 0" using ‹ε > 0› by (simp add: measure_le_0_iff mult_le_0_iff)
then have "A ∈ null_sets M" by (simp add: emeasure_eq_measure null_setsI)
then have "AE x in M. x ∈ space M - A" by (metis (no_types, lifting) AE_cong Diff_iff AE_not_in)
moreover
{
fix x assume "x ∈ space M - A"
then have "limsup (λn. ereal(birkhoff_sum g n x)) < ∞" unfolding A_def by auto
then obtain C where C: "⋀n. birkhoff_sum g n x ≤ C" using limsup_finite_then_bounded by presburger
{
fix n::nat assume "n > 0"
have "birkhoff_sum g n x = birkhoff_sum f n x - birkhoff_sum (real_cond_exp M Invariants f) n x - birkhoff_sum (λx. ε) n x"
unfolding g_def using birkhoff_sum_add birkhoff_sum_diff by auto
moreover have "birkhoff_sum (real_cond_exp M Invariants f) n x = n * real_cond_exp M Invariants f x"
using birkhoff_sum_of_invariants using ‹x ∈ space M - A› by auto
moreover have "birkhoff_sum (λx. ε) n x = n * ε" unfolding birkhoff_sum_def by auto
ultimately have "birkhoff_sum g n x = birkhoff_sum f n x - n * real_cond_exp M Invariants f x - n * ε"
by simp
then have "birkhoff_sum f n x = birkhoff_sum g n x + n * real_cond_exp M Invariants f x + n * ε"
by simp
then have "birkhoff_sum f n x / n = birkhoff_sum g n x / n + real_cond_exp M Invariants f x + ε"
using ‹n > 0› by (simp add: field_simps)
then have "birkhoff_sum f n x / n ≤ C/n + real_cond_exp M Invariants f x + ε"
using C[of n] ‹n > 0› by (simp add: divide_right_mono)
then have "ereal(birkhoff_sum f n x / n) ≤ ereal(C/n + real_cond_exp M Invariants f x + ε)"
by simp
}
then have "eventually (λn. ereal(birkhoff_sum f n x / n) ≤ ereal(C/n + real_cond_exp M Invariants f x + ε)) sequentially"
by (simp add: eventually_at_top_dense)
then have b: "limsup (λn. ereal(birkhoff_sum f n x / n)) ≤ limsup (λn. ereal(C/n + real_cond_exp M Invariants f x + ε))"
by (simp add: Limsup_mono)
have "(λn. ereal(C*(1/real n) + real_cond_exp M Invariants f x + ε)) ⇢ ereal(C * 0 + real_cond_exp M Invariants f x + ε)"
by (intro tendsto_intros)
then have "limsup (λn. ereal(C/real n + real_cond_exp M Invariants f x + ε)) = real_cond_exp M Invariants f x + ε"
using sequentially_bot tendsto_iff_Liminf_eq_Limsup by force
then have "limsup (λn. ereal(birkhoff_sum f n x / n)) ≤ real_cond_exp M Invariants f x + ε"
using b by simp
}
ultimately have "AE x in M. limsup (λn. ereal(birkhoff_sum f n x / n)) ≤ real_cond_exp M Invariants f x + ε"
by auto
then have "AE x in M. limsup (λn. ereal(birkhoff_sum f n x / n)) ≤ ereal(real_cond_exp M Invariants f x) + ε"
by auto
}
then show ?thesis
by (rule AE_upper_bound_inf_ereal)
qed
theorem birkhoff_theorem_AE_nonergodic:
fixes f::"'a ⇒ real"
assumes "integrable M f"
shows "AE x in M. (λn. birkhoff_sum f n x / n) ⇢ real_cond_exp M Invariants f x"
proof -
{
fix x assume i: "limsup (λn. ereal(birkhoff_sum f n x /n)) ≤ real_cond_exp M Invariants f x"
and ii: "limsup (λn. ereal(birkhoff_sum (λx. -f x) n x / n)) ≤ real_cond_exp M Invariants (λx. -f x) x"
and iii: "real_cond_exp M Invariants (λx. -f x) x = - real_cond_exp M Invariants f x"
have "⋀n. birkhoff_sum (λx. -f x) n x = - birkhoff_sum f n x"
using birkhoff_sum_cmult[where ?c = "-1" and ?f = f] by auto
then have "⋀n. ereal(birkhoff_sum (λx. -f x) n x / n) = - ereal(birkhoff_sum f n x / n)" by auto
moreover have "limsup (λn. - ereal(birkhoff_sum f n x / n)) = - liminf (λn. ereal(birkhoff_sum f n x /n))"
by (rule ereal_Limsup_uminus)
ultimately have "-liminf (λn. ereal(birkhoff_sum f n x /n)) = limsup (λn. ereal(birkhoff_sum (λx. -f x) n x / n))"
by simp
then have "-liminf (λn. ereal(birkhoff_sum f n x /n)) ≤ - real_cond_exp M Invariants f x"
using ii iii by simp
then have "liminf (λn. ereal(birkhoff_sum f n x /n)) ≥ real_cond_exp M Invariants f x"
by (simp add: ereal_uminus_le_reorder)
then have "(λn. birkhoff_sum f n x /n) ⇢ real_cond_exp M Invariants f x"
using i by (simp add: limsup_le_liminf_real)
} note * = this
moreover have "AE x in M. limsup (λn. ereal(birkhoff_sum f n x /n)) ≤ real_cond_exp M Invariants f x"
using birkhoff_aux2 assms by simp
moreover have "AE x in M. limsup (λn. ereal(birkhoff_sum (λx. -f x) n x / n)) ≤ real_cond_exp M Invariants (λx. -f x) x"
using birkhoff_aux2 assms by simp
moreover have "AE x in M. real_cond_exp M Invariants (λx. -f x) x = - real_cond_exp M Invariants f x"
using real_cond_exp_cmult[where ?c = "-1"] assms by force
ultimately show ?thesis by auto
qed
text ‹If a function $f$ is integrable, then $E(f\circ T - f | I) = E(f\circ T | I) - E(f|I) = 0$.
Hence, $S_n(f \circ T - f) / n$ converges almost everywhere to $0$, i.e., $f(T^n x)/n \to 0$.
It is remarkable (and sometimes useful) that this holds under the weaker condition that
$f\circ T - f$ is integrable (but not necessarily $f$), where this naive argument fails.
The reason is that the Birkhoff sum of $f \circ T - f$ is $f\circ T^n - f$. If $n$ is such that $x$
and $T^n(x)$ belong to a set where $f$ is bounded, it follows that this Birkhoff sum is also
bounded. Along such a sequence of times, $S_n(f\circ T - f)/n$ tends to $0$.
By Poincare recurrence theorem, there are such times for almost every points. As it also converges
to $E(f \circ T - f | I)$, it follows that this function is almost everywhere $0$. Then
$f (T^n x)/n = S_n(f\circ T^n - f)/n - f/n$ tends almost surely to $E(f\circ T-f |I) = 0$.
›
lemma limit_foTn_over_n:
fixes f::"'a ⇒ real"
assumes [measurable]: "f ∈ borel_measurable M"
and "integrable M (λx. f(T x) - f x)"
shows "AE x in M. real_cond_exp M Invariants (λx. f(T x) - f x) x = 0"
"AE x in M. (λn. f((T^^n) x) / n) ⇢ 0"
proof -
define E::"nat ⇒ 'a set" where "E k = {x ∈ space M. ¦f x¦ ≤ k}" for k
have [measurable]: "E k ∈ sets M" for k unfolding E_def by auto
have *: "(⋃k. E k) = space M" unfolding E_def by (auto simp add: real_arch_simple)
define F::"nat ⇒ 'a set" where "F k = recurrent_subset_infty (E k)" for k
have [measurable]: "F k ∈ sets M" for k unfolding F_def by auto
have **: "E k - F k ∈ null_sets M" for k unfolding F_def using Poincare_recurrence_thm by auto
have "space M - (⋃k. F k) ∈ null_sets M"
apply (rule null_sets_subset[of "(⋃k. E k - F k)"]) unfolding *[symmetric] using ** by auto
with AE_not_in[OF this] have "AE x in M. x ∈ (⋃k. F k)" by auto
moreover have "AE x in M. (λn. birkhoff_sum (λx. f(T x) - f x) n x / n)
⇢ real_cond_exp M Invariants (λx. f(T x) - f x) x"
by (rule birkhoff_theorem_AE_nonergodic[OF assms(2)])
moreover have "real_cond_exp M Invariants (λx. f(T x) - f x) x = 0 ∧ (λn. f((T^^n) x) / n) ⇢ 0"
if H: "(λn. birkhoff_sum (λx. f(T x) - f x) n x / n) ⇢ real_cond_exp M Invariants (λx. f(T x) - f x) x"
"x ∈ (⋃k. F k)" for x
proof -
have "f((T^^n) x) = birkhoff_sum (λx. f(T x) - f x) n x + f x" for n
unfolding birkhoff_sum_def by (induction n, auto)
then have "f((T^^n) x) / n = birkhoff_sum (λx. f(T x) - f x) n x / n + f x * (1/n)" for n
by (auto simp add: divide_simps)
moreover have "(λn. birkhoff_sum (λx. f(T x) - f x) n x / n + f x * (1/n)) ⇢ real_cond_exp M Invariants (λx. f(T x) - f x) x + f x * 0"
by (intro tendsto_intros H(1))
ultimately have lim: "(λn. f((T^^n) x) / n) ⇢ real_cond_exp M Invariants (λx. f(T x) - f x) x"
by auto
obtain k where "x ∈ F k" using H(2) by auto
then have "infinite {n. (T^^n) x ∈ E k}"
unfolding F_def recurrent_subset_infty_inf_returns by auto
with infinite_enumerate[OF this] obtain r :: "nat ⇒ nat"
where r: "strict_mono r" "⋀n. r n ∈ {n. (T^^n) x ∈ E k}"
by auto
have A: "(λn. k * (1/r n)) ⇢ real k * 0"
apply (intro tendsto_intros)
using LIMSEQ_subseq_LIMSEQ[OF lim_1_over_n ‹strict_mono r›] unfolding comp_def by auto
have B: "¦f((T^^(r n)) x) / r n¦ ≤ k / (r n)" for n
using r(2) unfolding E_def by (auto simp add: divide_simps)
have "(λn. f((T^^(r n)) x) / r n) ⇢ 0"
apply (rule tendsto_rabs_zero_cancel, rule tendsto_sandwich[of "λn. 0" _ _ "λn. k * (1/r n)"])
using A B by auto
moreover have "(λn. f((T^^(r n)) x) / r n) ⇢ real_cond_exp M Invariants (λx. f(T x) - f x) x"
using LIMSEQ_subseq_LIMSEQ[OF lim ‹strict_mono r›] unfolding comp_def by auto
ultimately have *: "real_cond_exp M Invariants (λx. f(T x) - f x) x = 0"
using LIMSEQ_unique by auto
then have "(λn. f((T^^n) x) / n) ⇢ 0" using lim by auto
then show ?thesis using * by auto
qed
ultimately show "AE x in M. real_cond_exp M Invariants (λx. f(T x) - f x) x = 0"
"AE x in M. (λn. f((T^^n) x) / n) ⇢ 0"
by auto
qed
text ‹We specialize the previous statement to the case where $f$ itself is integrable.›
lemma limit_foTn_over_n':
fixes f::"'a ⇒ real"
assumes [measurable]: "integrable M f"
shows "AE x in M. (λn. f((T^^n) x) / n) ⇢ 0"
by (rule limit_foTn_over_n, simp, rule Bochner_Integration.integrable_diff)
(auto intro: assms T_integral_preserving(1))
text ‹It is often useful to show that a function is cohomologous to a nicer function, i.e., to
prove that a given $f$ can be written as $f = g + u - u \circ T$ where $g$ is nicer than $f$. We
show below that any integrable function is cohomologous to a function which is arbitrarily close to
$E(f|I)$. This is an improved version of Lemma 2.1 in [Benoist-Quint, Annals of maths, 2011]. Note
that the function $g$ to which $f$ is cohomologous is very nice (and, in particular, integrable),
but the transfer function is only measurable in this argument. The fact that the control on
conditional expectation is nevertheless preserved throughout the argument follows from
Lemma~\verb+limit_foTn_over_n+ above.›
text ‹We start with the lemma (and the proof) of [BQ2011]. It shows that, if a function has a
conditional expectation with respect to invariants which is positive, then it is cohomologous to a
nonnegative function. The argument is the clever remark that $g = \max (0, \inf_n S_n f)$ and $u =
\min (0, \inf_n S_n f)$ work (where these expressions are well defined as $S_n f$ tends to infinity
thanks to our assumption).›
lemma cohomologous_approx_cond_exp_aux:
fixes f::"'a ⇒ real"
assumes [measurable]: "integrable M f"
and "AE x in M. real_cond_exp M Invariants f x > 0"
shows "∃u g. u ∈ borel_measurable M ∧ (integrable M g) ∧ (AE x in M. g x ≥ 0 ∧ g x ≤ max 0 (f x)) ∧ (∀x. f x = g x + u x - u (T x))"
proof -
define h::"'a ⇒ real" where "h = (λx. (INF n∈{1..}. birkhoff_sum f n x))"
define u where "u = (λx. min (h x) 0)"
define g where "g = (λx. f x - u x + u (T x))"
have [measurable]: "h ∈ borel_measurable M" "u ∈ borel_measurable M" "g ∈ borel_measurable M"
unfolding g_def h_def u_def by auto
have "f x = g x + u x - u (T x)" for x unfolding g_def by auto
{
fix x assume H: "real_cond_exp M Invariants f x > 0"
"(λn. birkhoff_sum f n x / n) ⇢ real_cond_exp M Invariants f x"
have "eventually (λn. ereal(birkhoff_sum f n x / n) * ereal n = ereal(birkhoff_sum f n x)) sequentially"
unfolding eventually_sequentially by (rule exI[of _ 1], auto)
moreover have "(λn. ereal(birkhoff_sum f n x / n) * ereal n) ⇢ ereal(real_cond_exp M Invariants f x) * ∞"
apply (intro tendsto_intros) using H by auto
ultimately have "(λn. ereal(birkhoff_sum f n x)) ⇢ ereal(real_cond_exp M Invariants f x) * ∞"
by (blast intro: Lim_transform_eventually)
then have "(λn. ereal(birkhoff_sum f n x)) ⇢ ∞"
using H by auto
then have B: "∃C. ∀n. C ≤ birkhoff_sum f n x"
by (intro liminf_finite_then_bounded_below, simp add: liminf_PInfty)
have "h x ≤ f x"
unfolding h_def apply (rule cInf_lower) using B by force+
have "{birkhoff_sum f n (T x) |n. n ∈ {1..}} = {birkhoff_sum f (1+n) (x) - f x |n. n ∈ {1..}}"
unfolding birkhoff_sum_cocycle by auto
also have "... = {birkhoff_sum f n x - f x |n. n ∈ {2..}}"
by (metis (no_types, opaque_lifting) Suc_1 Suc_eq_plus1_left Suc_le_D Suc_le_mono atLeast_iff)
finally have *: "{birkhoff_sum f n (T x) |n. n ∈ {1..}} = (λt. t - (f x))`{birkhoff_sum f n x |n. n ∈ {2..}}"
by auto
have "h(T x) = Inf {birkhoff_sum f n (T x) |n. n ∈ {1..}}"
unfolding h_def by (metis Setcompr_eq_image)
also have "... = (⨅t∈{birkhoff_sum f n x |n. n ∈ {2..}}. t - f x)"
by (simp only: *)
also have "... = (λt. t - (f x)) (Inf {birkhoff_sum f n x |n. n ∈ {2..}})"
using B by (auto intro!: monoI bijI mono_bij_cInf [symmetric])
finally have I: "Inf {birkhoff_sum f n x |n. n ∈ {2..}} = f x + h (T x)" by auto
have "max 0 (h x) + u x = h x"
unfolding u_def by auto
also have "... = Inf {birkhoff_sum f n x |n. n ∈ {1..}}"
unfolding h_def by (metis Setcompr_eq_image)
also have "... = Inf ({birkhoff_sum f n x |n. n ∈ {1}} ∪ {birkhoff_sum f n x |n. n ∈ {2..}})"
by (auto intro!: arg_cong[of _ _ Inf], metis One_nat_def Suc_1 antisym birkhoff_sum_1(2) not_less_eq_eq, force)
also have "Inf ({birkhoff_sum f n x |n. n ∈ {1}} ∪ {birkhoff_sum f n x |n. n ∈ {2..}})
= min (Inf {birkhoff_sum f n x |n. n ∈ {1}}) (Inf {birkhoff_sum f n x |n. n ∈ {2..}})"
unfolding inf_min[symmetric] apply (intro cInf_union_distrib) using B by auto
also have "... = min (f x) (f x + h (T x))" using I by auto
also have "... = f x + u (T x)" unfolding u_def by auto
finally have "max 0 (h x) = f x + u (T x) - u x" by auto
then have "g x = max 0 (h x)" unfolding g_def by auto
then have "g x ≥ 0 ∧ g x ≤ max 0 (f x)" using ‹h x ≤ f x› by auto
}
then have *: "AE x in M. g x ≥ 0 ∧ g x ≤ max 0 (f x)"
using assms(2) birkhoff_theorem_AE_nonergodic[OF assms(1)] by auto
moreover have "integrable M g"
apply (rule Bochner_Integration.integrable_bound[of _ f]) using * by (auto simp add: assms)
ultimately have "u ∈ borel_measurable M ∧ integrable M g ∧ (AE x in M. 0 ≤ g x ∧ g x ≤ max 0 (f x)) ∧ (∀x. f x = g x + u x - u (T x))"
using ‹⋀x. f x = g x + u x - u (T x)› ‹u ∈ borel_measurable M› by auto
then show ?thesis by blast
qed
text ‹To deduce the stronger version that $f$ is cohomologous to an arbitrarily good approximation
of $E(f|I)$, we apply the previous lemma twice, to control successively the negative and the
positive side. The sign control in the conclusion of the previous lemma implies that the second step
does not spoil the first one.›
lemma cohomologous_approx_cond_exp:
fixes f::"'a ⇒ real" and B::"'a ⇒ real"
assumes [measurable]: "integrable M f" "B ∈ borel_measurable M"
and "AE x in M. B x > 0"
shows "∃g u. u ∈ borel_measurable M
∧ integrable M g
∧ (∀x. f x = g x + u x - u (T x))
∧ (AE x in M. abs(g x - real_cond_exp M Invariants f x) ≤ B x)"
proof -
define C where "C = (λx. min (B x) 1)"
have [measurable]: "integrable M C"
apply (rule Bochner_Integration.integrable_bound[of _ "λ_. (1::real)"], auto)
unfolding C_def using assms(3) by auto
have "C x ≤ B x" for x unfolding C_def by auto
have "AE x in M. C x > 0" unfolding C_def using assms(3) by auto
have AECI: "AE x in M. real_cond_exp M Invariants C x > 0"
by (intro real_cond_exp_gr_c ‹integrable M C› ‹AE x in M. C x > 0›)
define f1 where "f1 = (λx. f x - real_cond_exp M Invariants f x)"
have "integrable M f1"
unfolding f1_def by (intro Bochner_Integration.integrable_diff ‹integrable M f› real_cond_exp_int(1))
have "AE x in M. real_cond_exp M Invariants f1 x = real_cond_exp M Invariants f x - real_cond_exp M Invariants (real_cond_exp M Invariants f) x"
unfolding f1_def apply (rule real_cond_exp_diff) by (intro Bochner_Integration.integrable_diff
‹integrable M f› ‹integrable M C› real_cond_exp_int(1))+
moreover have "AE x in M. real_cond_exp M Invariants (real_cond_exp M Invariants f) x = real_cond_exp M Invariants f x"
by (intro real_cond_exp_nested_subalg subalg ‹integrable M f›, auto)
ultimately have AEf1: "AE x in M. real_cond_exp M Invariants f1 x = 0" by auto
have A [measurable]: "integrable M (λx. f1 x + C x)"
by (intro Bochner_Integration.integrable_add ‹integrable M f1› ‹integrable M C›)
have "AE x in M. real_cond_exp M Invariants (λx. f1 x + C x) x = real_cond_exp M Invariants f1 x + real_cond_exp M Invariants C x"
by (intro real_cond_exp_add ‹integrable M f1› ‹integrable M C›)
then have B: "AE x in M. real_cond_exp M Invariants (λx. f1 x + C x) x > 0"
using AECI AEf1 by auto
obtain u2 g2 where H2: "u2 ∈ borel_measurable M" "integrable M g2" "AE x in M. g2 x ≥ 0 ∧ g2 x ≤ max 0 (f1 x + C x)" "⋀x. f1 x + C x = g2 x + u2 x - u2 (T x)"
using cohomologous_approx_cond_exp_aux[OF A B] by blast
define f2 where "f2 = (λx. (g2 x - C x))"
have *: "u2(T x) - u2 x = f2 x -f1 x" for x unfolding f2_def using H2(4)[of x] by auto
have "AE x in M. f2 x ≥ - C x" using H2(3) unfolding f2_def by auto
have "integrable M f2"
unfolding f2_def by (intro Bochner_Integration.integrable_diff ‹integrable M g2› ‹integrable M C›)
have "AE x in M. real_cond_exp M Invariants (λx. u2(T x) - u2 x) x = 0"
proof (rule limit_foTn_over_n)
show "integrable M (λx. u2(T x) - u2 x)"
unfolding * by (intro Bochner_Integration.integrable_diff ‹integrable M f1› ‹integrable M f2›)
qed (simp add: ‹u2 ∈ borel_measurable M›)
then have "AE x in M. real_cond_exp M Invariants (λx. f2 x - f1 x) x = 0"
unfolding * by simp
moreover have "AE x in M. real_cond_exp M Invariants (λx. f2 x - f1 x) x = real_cond_exp M Invariants f2 x - real_cond_exp M Invariants f1 x"
by (intro real_cond_exp_diff ‹integrable M f2› ‹integrable M f1›)
ultimately have AEf2: "AE x in M. real_cond_exp M Invariants f2 x = 0"
using AEf1 by auto
have A [measurable]: "integrable M (λx. C x - f2 x)"
by (intro Bochner_Integration.integrable_diff ‹integrable M f2› ‹integrable M C›)
have "AE x in M. real_cond_exp M Invariants (λx. C x - f2 x) x = real_cond_exp M Invariants C x - real_cond_exp M Invariants f2 x"
by (intro real_cond_exp_diff ‹integrable M f2› ‹integrable M C›)
then have B: "AE x in M. real_cond_exp M Invariants (λx. C x - f2 x) x > 0"
using AECI AEf2 by auto
obtain u3 g3 where H3: "u3 ∈ borel_measurable M" "integrable M g3" "AE x in M. g3 x ≥ 0 ∧ g3 x ≤ max 0 (C x - f2 x)" "⋀x. C x - f2 x = g3 x + u3 x - u3 (T x)"
using cohomologous_approx_cond_exp_aux[OF A B] by blast
define f3 where "f3 = (λx. C x - g3 x)"
have "AE x in M. f3 x ≥ min (C x) (f2 x)" unfolding f3_def using H3(3) by auto
then have "AE x in M. f3 x ≥ -C x" using ‹AE x in M. f2 x ≥ - C x› ‹AE x in M. C x > 0› by auto
moreover have "AE x in M. f3 x ≤ C x" unfolding f3_def using H3(3) by auto
ultimately have "AE x in M. abs(f3 x) ≤ C x" by auto
then have *: "AE x in M. abs(f3 x) ≤ B x" using order_trans[OF _ ‹⋀x. C x ≤ B x›] by auto
define g where "g = (λx. f3 x + real_cond_exp M Invariants f x)"
define u where "u = (λx. u2 x - u3 x)"
have "AE x in M. abs (g x - real_cond_exp M Invariants f x) ≤ B x"
unfolding g_def using * by auto
moreover have "f x = g x + u x - u(T x)" for x
using H3(4)[of x] H2(4)[of x] unfolding u_def g_def f3_def f2_def f1_def by auto
moreover have "u ∈ borel_measurable M"
unfolding u_def using ‹u2 ∈ borel_measurable M› ‹u3 ∈ borel_measurable M› by auto
moreover have "integrable M g"
unfolding g_def f3_def by (intro Bochner_Integration.integrable_add Bochner_Integration.integrable_diff
‹integrable M C› ‹integrable M g3› ‹integrable M f› real_cond_exp_int(1))
ultimately show ?thesis by auto
qed
subsubsection ‹$L^1$ version of Birkhoff theorem›
text ‹The $L^1$ convergence in Birkhoff theorem follows from the almost everywhere convergence and
general considerations on $L^1$ convergence (Scheffe's lemma) explained
in \verb+AE_and_int_bound_implies_L1_conv2+.
This argument works neatly for nonnegative functions, the general case reduces to this one by taking
the positive and negative parts of a given function.
One could also prove it by truncation: for bounded functions, the $L^1$ convergence follows
from the boundedness and almost sure convergence. The general case follows by density, but it
is a little bit tedious to write as one need to make sure that the conditional expectation
of the truncation converges to the conditional expectation of the original function. This is true
in $L^1$ as the conditional expectation is a contraction in $L^1$, it follows almost everywhere
after taking a subsequence. All in all, the argument based on Scheffe's lemma seems more
economical.›
lemma birkhoff_lemma_L1:
fixes f::"'a ⇒ real"
assumes "⋀x. f x ≥ 0"
and [measurable]: "integrable M f"
shows "(λn. ∫⇧+x. norm(birkhoff_sum f n x / n - real_cond_exp M Invariants f x) ∂M) ⇢ 0"
proof (rule Scheffe_lemma2)
show i: "integrable M (real_cond_exp M Invariants f)" using assms by (simp add: real_cond_exp_int(1))
show "AE x in M. (λn. birkhoff_sum f n x / real n) ⇢ real_cond_exp M Invariants f x"
using birkhoff_theorem_AE_nonergodic assms by simp
fix n
have [measurable]: "(λx. ennreal ¦birkhoff_sum f n x¦) ∈ borel_measurable M" by measurable
show [measurable]: "(λx. birkhoff_sum f n x / real n) ∈ borel_measurable M" by measurable
have "AE x in M. real_cond_exp M Invariants f x ≥ 0" using assms(1) real_cond_exp_pos by simp
then have *: "AE x in M. norm (real_cond_exp M Invariants f x) = real_cond_exp M Invariants f x" by auto
have **: "(∫ x. norm (real_cond_exp M Invariants f x) ∂M) = (∫ x. real_cond_exp M Invariants f x ∂M)"
apply (rule integral_cong_AE) using * by auto
have "(∫⇧+x. ennreal (norm (real_cond_exp M Invariants f x)) ∂M) = (∫ x. norm (real_cond_exp M Invariants f x) ∂M)"
by (rule nn_integral_eq_integral) (auto simp add: i)
also have "... = (∫ x. real_cond_exp M Invariants f x ∂M)"
using ** by simp
also have "... = (∫ x. f x ∂M)"
using real_cond_exp_int(2) assms(2) by auto
also have "... = (∫x. norm(f x) ∂M)" using assms by auto
also have "... = (∫⇧+x. norm(f x) ∂M)"
by (rule nn_integral_eq_integral[symmetric], auto simp add: assms(2))
finally have eq: "(∫⇧+ x. norm (real_cond_exp M Invariants f x) ∂M) = (∫⇧+ x. norm(f x) ∂M)" by simp
{
fix x
have "norm(birkhoff_sum f n x) ≤ birkhoff_sum (λx. norm(f x)) n x"
using birkhoff_sum_abs by fastforce
then have "norm(birkhoff_sum f n x) ≤ birkhoff_sum (λx. ennreal(norm(f x))) n x"
unfolding birkhoff_sum_def by auto
}
then have "(∫⇧+x. norm(birkhoff_sum f n x) ∂M) ≤ (∫⇧+x. birkhoff_sum (λx. ennreal(norm(f x))) n x ∂M)"
by (simp add: nn_integral_mono)
also have "... = n * (∫⇧+x. norm(f x) ∂M)"
by (rule birkhoff_sum_nn_integral, auto)
also have "... = n * (∫⇧+ x. norm (real_cond_exp M Invariants f x) ∂M)"
using eq by simp
finally have *: "(∫⇧+x. norm(birkhoff_sum f n x) ∂M) ≤ n * (∫⇧+ x. norm (real_cond_exp M Invariants f x) ∂M)"
by simp
show "(∫⇧+ x. ennreal (norm (birkhoff_sum f n x / real n)) ∂M) ≤ (∫⇧+ x. norm (real_cond_exp M Invariants f x) ∂M)"
proof (cases)
assume "n = 0"
then show ?thesis by auto
next
assume "¬(n = 0)"
then have "n > 0" by simp
then have "1/ennreal(real n) ≥ 0" by simp
have "(∫⇧+ x. ennreal (norm (birkhoff_sum f n x / real n)) ∂M) = (∫⇧+ x. ennreal (norm (birkhoff_sum f n x)) / ennreal(real n) ∂M)"
using ‹n > 0› by (auto simp: divide_ennreal)
also have "... = (∫⇧+ x. (1/ennreal(real n)) * ennreal (norm (birkhoff_sum f n x)) ∂M)"
by (simp add: ‹0 < n› divide_ennreal_def mult.commute)
also have "... = (1/ennreal(real n) * (∫⇧+ x. ennreal (norm (birkhoff_sum f n x)) ∂M))"
by (subst nn_integral_cmult) auto
also have "... ≤ (1/ennreal(real n)) * (ennreal(real n) * (∫⇧+ x. norm (real_cond_exp M Invariants f x) ∂M))"
using * by (intro mult_mono) (auto simp: ennreal_of_nat_eq_real_of_nat)
also have "... = (∫⇧+ x. norm (real_cond_exp M Invariants f x) ∂M)"
using ‹n > 0›
by (auto simp del: ennreal_1 simp add: ennreal_1[symmetric] divide_ennreal ennreal_mult[symmetric] mult.assoc[symmetric])
simp
finally show ?thesis by simp
qed
qed
theorem birkhoff_theorem_L1_nonergodic:
fixes f::"'a ⇒ real"
assumes [measurable]: "integrable M f"
shows "(λn. ∫⇧+x. norm(birkhoff_sum f n x / n - real_cond_exp M Invariants f x) ∂M) ⇢ 0"
proof -
define g where "g = (λx. max (f x) 0)"
have g_int [measurable]: "integrable M g" unfolding g_def using assms by auto
define h where "h = (λx. max (-f x) 0)"
have h_int [measurable]: "integrable M h" unfolding h_def using assms by auto
have "f = (λx. g x - h x)" unfolding g_def h_def by auto
{
fix n::nat assume "n > 0"
have "⋀x. birkhoff_sum f n x = birkhoff_sum g n x - birkhoff_sum h n x" using birkhoff_sum_diff ‹f = (λx. g x - h x)› by auto
then have "⋀x. birkhoff_sum f n x / n = birkhoff_sum g n x / n - birkhoff_sum h n x / n" using ‹n > 0› by (simp add: diff_divide_distrib)
moreover have "AE x in M. real_cond_exp M Invariants g x - real_cond_exp M Invariants h x = real_cond_exp M Invariants f x"
using AE_symmetric[OF real_cond_exp_diff] g_int h_int ‹f = (λx. g x - h x)› by auto
ultimately have "AE x in M. birkhoff_sum f n x / n - real_cond_exp M Invariants f x =
(birkhoff_sum g n x / n - real_cond_exp M Invariants g x) - (birkhoff_sum h n x / n - real_cond_exp M Invariants h x)"
by auto
then have *: "AE x in M. norm(birkhoff_sum f n x / n - real_cond_exp M Invariants f x) ≤
norm(birkhoff_sum g n x / n - real_cond_exp M Invariants g x) + norm(birkhoff_sum h n x / n - real_cond_exp M Invariants h x)"
by auto
have "(∫⇧+ x. norm(birkhoff_sum f n x / n - real_cond_exp M Invariants f x) ∂M) ≤
(∫⇧+ x. ennreal(norm(birkhoff_sum g n x / n - real_cond_exp M Invariants g x)) + norm(birkhoff_sum h n x / n - real_cond_exp M Invariants h x) ∂M)"
apply (rule nn_integral_mono_AE) using * by (simp add: ennreal_plus[symmetric] del: ennreal_plus)
also have "... = (∫⇧+ x. norm(birkhoff_sum g n x / n - real_cond_exp M Invariants g x) ∂M) + (∫⇧+ x. norm(birkhoff_sum h n x / n - real_cond_exp M Invariants h x) ∂M)"
apply (rule nn_integral_add) apply auto using real_cond_exp_F_meas borel_measurable_cond_exp2 by measurable
finally have "(∫⇧+ x. norm(birkhoff_sum f n x / n - real_cond_exp M Invariants f x) ∂M) ≤
(∫⇧+ x. norm(birkhoff_sum g n x / n - real_cond_exp M Invariants g x) ∂M) + (∫⇧+ x. norm(birkhoff_sum h n x / n - real_cond_exp M Invariants h x) ∂M)"
by simp
}
then have *: "eventually (λn. (∫⇧+ x. norm(birkhoff_sum f n x / n - real_cond_exp M Invariants f x) ∂M) ≤
(∫⇧+ x. norm(birkhoff_sum g n x / n - real_cond_exp M Invariants g x) ∂M) + (∫⇧+ x. norm(birkhoff_sum h n x / n - real_cond_exp M Invariants h x) ∂M))
sequentially"
using eventually_at_top_dense by auto
have **: "eventually (λn. (∫⇧+ x. norm(birkhoff_sum f n x / n - real_cond_exp M Invariants f x) ∂M) ≥ 0) sequentially"
by simp
have "(λn. (∫⇧+ x. norm(birkhoff_sum g n x / n - real_cond_exp M Invariants g x) ∂M)) ⇢ 0"
apply (rule birkhoff_lemma_L1, auto simp add: g_int) unfolding g_def by auto
moreover have "(λn. (∫⇧+ x. norm(birkhoff_sum h n x / n - real_cond_exp M Invariants h x) ∂M)) ⇢ 0"
apply (rule birkhoff_lemma_L1, auto simp add: h_int) unfolding h_def by auto
ultimately have "(λn. (∫⇧+ x. norm(birkhoff_sum g n x / n - real_cond_exp M Invariants g x) ∂M) + (∫⇧+ x. norm(birkhoff_sum h n x / n - real_cond_exp M Invariants h x) ∂M)) ⇢ 0"
using tendsto_add[of _ 0 _ _ 0] by auto
then show ?thesis
using tendsto_sandwich[OF ** *] by auto
qed
subsubsection ‹Conservativity of skew products›
text ‹The behaviour of skew-products of the form $(x, y) \mapsto (Tx, y + f x)$ is directly related
to Birkhoff theorem, as the iterates involve the Birkhoff sums in the fiber. Birkhoff theorem
implies that such a skew product is conservative when the function $f$ has vanishing conditional
expectation.
To prove the theorem, assume by contradiction that a set $A$ with positive measure does not
intersect its preimages. Replacing $A$ with a smaller set $C$, we can assume that $C$ is bounded in
the $y$-direction, by a constant $N$, and also that all its nonempty vertical fibers, above the
projection $Cx$, have a measure bounded from below. Then, by Birkhoff theorem, for any $r>0$, most
of the first $n$ preimages of $C$ are contained in the set $\{|y| \leq r n+N\}$, of measure $O(r
n)$. Hence, they can not be disjoint if $r < \mu(C)$. To make this argument rigorous, one should
only consider the preimages whose $x$-component belongs to a set $Dx$ where the Birkhoff sums are
small. This condition has a positive measure if $\mu(Cx) + \mu(Dx) > \mu(M)$, which one can ensure
by taking $Dx$ large enough.›
theorem (in fmpt) skew_product_conservative:
fixes f::"'a ⇒ real"
assumes [measurable]: "integrable M f"
and "AE x in M. real_cond_exp M Invariants f x = 0"
shows "conservative_mpt (M ⨂⇩M lborel) (λ(x,y). (T x, y + f x))"
proof (rule conservative_mptI)
let ?TS = "(λ(x,y). (T x, y + f x))"
let ?MS = "M ⨂⇩M (lborel::real measure)"
have f_meas [measurable]: "f ∈ borel_measurable M" by auto
have "mpt M T" by (simp add: mpt_axioms)
with mpt_skew_product_real[OF this f_meas] show "mpt ?MS ?TS" by simp
then interpret TS: mpt ?MS ?TS by auto
fix A::"('a × real) set"
assume A1 [measurable]: "A ∈ sets ?MS" and A2: "emeasure ?MS A > 0"
have "A = (⋃N::nat. A ∩ {(x,y). abs(y) ≤ N})" by (auto simp add: real_arch_simple)
then have *: "emeasure ?MS (⋃N::nat. A ∩ {(x,y). abs(y) ≤ N}) > 0"
using A2 by simp
have "space ?MS = space M × space (lborel::real measure)" using space_pair_measure by auto
then have A_inc: "A ⊆ space M × space (lborel::real measure)" using sets.sets_into_space[OF A1] by auto
{
fix N::nat
have "{(x, y). abs(y) ≤ real N ∧ x ∈ space M} = space M × {-(real N)..(real N)}" by auto
then have "{(x, y). ¦y¦ ≤ real N ∧ x ∈ space M} ∈ sets ?MS" by auto
then have "A ∩ {(x, y). ¦y¦ ≤ real N ∧ x ∈ space M} ∈ sets ?MS" using A1 by auto
moreover have "A ∩ {(x,y). abs(y) ≤ real N} = A ∩ {(x, y). ¦y¦ ≤ real N ∧ x ∈ space M}"
using A_inc by blast
ultimately have "A ∩ {(x,y). abs(y) ≤ real N} ∈ sets ?MS" by auto
}
then have [measurable]: "⋀N::nat. A ∩ {(x, y). ¦y¦ ≤ real N} ∈ sets (M ⨂⇩M borel)" by auto
have "∃N::nat. emeasure ?MS (A ∩ {(x,y). abs(y) ≤ N}) > 0"
apply (rule emeasure_pos_unionE) using * by auto
then obtain N::nat where N: "emeasure ?MS (A ∩ {(x,y). abs(y) ≤ N}) > 0"
by auto
define B where "B = A ∩ {(x,y). abs(y) ≤ N}"
have B_meas [measurable]: "B ∈ sets (M ⨂⇩M lborel)" unfolding B_def by auto
have "0 < emeasure (M ⨂⇩M lborel) B" unfolding B_def using N by auto
also have "... = (∫⇧+x. emeasure lborel (Pair x -` B) ∂M)"
apply (rule sigma_finite_measure.emeasure_pair_measure_alt)
using B_meas by (auto simp add: lborel.sigma_finite_measure_axioms)
finally have *: "(∫⇧+x. emeasure lborel (Pair x -` B) ∂M) > 0" by simp
have "∃Cx∈sets M. ∃e::real>0. emeasure M Cx > 0 ∧ (∀x ∈ Cx. emeasure lborel (Pair x -` B) ≥ e)"
by (rule not_AE_zero_int_ennreal_E, auto simp add: *)
then obtain Cx e where [measurable]: "Cx ∈ sets M" and Cxe: "e>(0::real)" "emeasure M Cx > 0" "⋀x. x ∈ Cx ⟹ emeasure lborel (Pair x -` B) ≥ e"
by blast
define C where "C = B ∩ (Cx × (UNIV::real set))"
have C_meas [measurable]: "C ∈ sets (M ⨂⇩M lborel)" unfolding C_def using B_meas by auto
have Cx_fibers: "⋀x. x ∈ Cx ⟹ emeasure lborel (Pair x -` C) ≥ e" using Cxe(3) C_def by auto
define c where "c = (measure M Cx)/2"
have "c > 0" unfolding c_def using Cxe(2) by (simp add: emeasure_eq_measure)
text ‹We will apply Birkhoff theorem to show that most preimages of $C$ at time $n$ are contained in a cylinder
of height roughly $r n$, for some suitably small $r$. How small $r$ should be to get a
contradiction can be determined at the end of the proof. It turns out that the good condition
is the following one -- this is by no means obvious now.›
define r where "r = (if measure M (space M) = 0 then 1 else e * c / (4 * measure M (space M)))"
have "r > 0" using ‹e > 0› ‹c > 0› unfolding r_def
apply auto using measure_le_0_iff by fastforce
have pos: "e*c-2*r*measure M (space M) > 0" using ‹e > 0› ‹c > 0› unfolding r_def by auto
define Bgood where "Bgood = {x ∈ space M. (λn. birkhoff_sum f n x / n) ⇢ 0}"
have [measurable]: "Bgood ∈ sets M" unfolding Bgood_def by auto
have *: "AE x in M. x ∈ Bgood" unfolding Bgood_def using birkhoff_theorem_AE_nonergodic[OF assms(1)] assms(2) by auto
then have "emeasure M Bgood = emeasure M (space M)"
by (intro emeasure_eq_AE) auto
{
fix x assume "x ∈ Bgood"
then have "x ∈ space M" unfolding Bgood_def by auto
have "(λn. birkhoff_sum f n x / n) ⇢ 0" using ‹x ∈ Bgood› unfolding Bgood_def by auto
moreover have "0 ∈ {-r<..<r}" "open {-r<..<r}" using ‹r>0› by auto
ultimately have "eventually (λn. birkhoff_sum f n x / n ∈ {-r<..<r}) sequentially"
using topological_tendstoD by blast
then obtain n0 where n0: "n0>0" "⋀n. n ≥ n0 ⟹ birkhoff_sum f n x / n ∈ {-r<..<r}"
using eventually_sequentially by (metis (mono_tags, lifting) le0 le_simps(3) neq0_conv)
{
fix n assume "n ≥ n0"
then have "n>0" using ‹n0>0› by auto
with n0(2)[OF ‹n ≥ n0›] have "abs(birkhoff_sum f n x / n) ≤ r" by auto
then have "abs(birkhoff_sum f n x) ≤ r * n" using ‹n>0› by (simp add: divide_le_eq)
}
then have "x ∈ (⋃n0. {x ∈ space M. ∀n∈{n0..}. abs(birkhoff_sum f n x) ≤ r * n})" using ‹x ∈ space M› by blast
}
then have "AE x in M. x ∉ space M - (⋃n0. {x ∈ space M. ∀n∈{n0..}. abs(birkhoff_sum f n x) ≤ r * n})"
using * by auto
then have eqM: "emeasure M (⋃n0. {x ∈ space M. ∀n∈{n0..}. abs(birkhoff_sum f n x) ≤ r * n}) = emeasure M (space M)"
by (intro emeasure_eq_AE) auto
have "(λn0. emeasure M {x ∈ space M. ∀n∈{n0..}. abs(birkhoff_sum f n x) ≤ r * n} + c)
⇢ emeasure M (⋃n0. {x ∈ space M. ∀n∈{n0..}. abs(birkhoff_sum f n x) ≤ r * n}) + c"
by (intro tendsto_intros Lim_emeasure_incseq) (auto simp add: incseq_def)
moreover have "emeasure M (⋃n0. {x ∈ space M. ∀n∈{n0..}. abs(birkhoff_sum f n x) ≤ r * n}) + c > emeasure M (space M)"
using eqM ‹c > 0› emeasure_eq_measure by auto
ultimately have "eventually (λn0. emeasure M {x ∈ space M. ∀n∈{n0..}. abs(birkhoff_sum f n x) ≤ r * n} + c > emeasure M (space M)) sequentially"
unfolding order_tendsto_iff by auto
then obtain n0 where n0: "emeasure M {x ∈ space M. ∀n∈{n0..}. abs(birkhoff_sum f n x) ≤ r * n} + c > emeasure M (space M)"
using eventually_sequentially by auto
define Dx where "Dx = {x ∈ space M. ∀n∈{n0..}. abs(birkhoff_sum f n x) ≤ r * n}"
have Dx_meas [measurable]: "Dx ∈ sets M" unfolding Dx_def by auto
have "emeasure M Dx + c ≥ emeasure M (space M)" using n0 Dx_def by auto
obtain n1::nat where n1: "n1 > max n0 ((measure M (space M) * 2 * N + e*c*n0 - e*c) / (e*c-2*r*measure M (space M)))"
by (metis mult.commute mult.left_neutral numeral_One reals_Archimedean3 zero_less_numeral)
then have "n1 > n0" by auto
have n1_ineq: "n1 * (e*c-2*r*measure M (space M)) > (measure M (space M) * 2 * N + e*c*n0 - e*c)"
using n1 pos by (simp add: pos_divide_less_eq)
define D where "D = (λn. Dx × {-r*n1-N..r*n1+N} ∩ (?TS^^n)-`C)"
have Dn_meas [measurable]: "D n ∈ sets (M ⨂⇩M lborel)" for n
unfolding D_def apply (rule TS.T_intersec_meas(2)) using C_meas by auto
have "emeasure ?MS (D n) ≥ e * c" if "n ∈ {n0..n1}" for n
proof -
have "n ≥ n0" "n ≤ n1" using that by auto
{
fix x assume [simp]: "x ∈ space M"
define F where "F = {y ∈ {-r*n1-N..r*n1+N}. y + birkhoff_sum f n x ∈ Pair ((T^^n)x) -`C}"
have [measurable]: "F ∈ sets lborel" unfolding F_def by measurable
{
fix y::real
have "(?TS^^n)(x,y) = ((T^^n)x, y + birkhoff_sum f n x)"
using skew_product_real_iterates by simp
then have "(indicator C ((?TS^^n) (x,y))::ennreal) = indicator Cx ((T^^n)x) * indicator (Pair ((T^^n)x) -`C) (y + birkhoff_sum f n x)"
using C_def by (simp add: indicator_def)
moreover have "(indicator (D n) (x, y)::ennreal) = indicator Dx x * indicator {-r*n1-N..r*n1+N} y * indicator C ((?TS^^n) (x,y))"
unfolding D_def by (simp add: indicator_def)
ultimately have "(indicator (D n) (x, y)::ennreal) = indicator Dx x * indicator {-r*n1-N..r*n1+N} y
* indicator Cx ((T^^n)x) * indicator (Pair ((T^^n)x) -`C) (y + birkhoff_sum f n x)"
by (simp add: mult.assoc)
then have "(indicator (D n) (x, y)::ennreal) = indicator (Dx ∩ (T^^n)-`Cx) x * indicator F y"
unfolding F_def by (simp add: indicator_def)
}
then have "(∫⇧+y. indicator (D n) (x, y) ∂lborel) = (∫⇧+y. indicator (Dx ∩ (T^^n)-`Cx) x * indicator F y ∂lborel)"
by auto
also have "... = indicator (Dx ∩ (T^^n)-`Cx) x * (∫⇧+y. indicator F y ∂lborel)"
by (rule nn_integral_cmult, auto)
also have "... = indicator (Dx ∩ (T^^n)-`Cx) x * emeasure lborel F" using ‹F ∈ sets lborel› by auto
finally have A: "(∫⇧+y. indicator (D n) (x, y) ∂lborel) = indicator (Dx ∩ (T^^n)-`Cx) x * emeasure lborel F"
by simp
have "(∫⇧+y. indicator (D n) (x, y) ∂lborel) ≥ ennreal e * indicator (Dx ∩ (T^^n)-`Cx) x"
proof (cases)
assume "indicator (Dx ∩ (T^^n)-`Cx) x = (0::ennreal)"
then show ?thesis by auto
next
assume "¬(indicator (Dx ∩ (T^^n)-`Cx) x = (0::ennreal))"
then have "x ∈ Dx ∩ (T^^n)-`Cx" by (simp add: indicator_eq_0_iff)
then have "x ∈ Dx" "(T^^n) x ∈ Cx" by auto
then have "abs(birkhoff_sum f n x) ≤ r * n" using ‹n ∈ {n0..n1}› Dx_def by auto
then have *: "abs(birkhoff_sum f n x) ≤ r * n1" using ‹n ≤ n1› ‹r>0›
by (meson of_nat_le_iff order_trans mult_le_cancel_left_pos)
have F_expr: "F = {-r*n1-N..r*n1+N} ∩ (+)(birkhoff_sum f n x) -` (Pair ((T^^n)x) -`C)"
unfolding F_def by (auto simp add: add.commute)
have "(Pair ((T^^n)x) -`C) ⊆ {real_of_int (- int N)..real N}" unfolding C_def B_def by auto
then have "((+)(birkhoff_sum f n x)) -` (Pair ((T^^n)x) -`C) ⊆ {-N-birkhoff_sum f n x..N-birkhoff_sum f n x}"
by auto
also have "... ⊆ {-r * n1 - N .. r * n1 + N}" using * by auto
finally have "F = ((+)(birkhoff_sum f n x)) -` (Pair ((T^^n)x) -`C)" unfolding F_expr by auto
then have "emeasure lborel F = emeasure lborel ((+)(birkhoff_sum f n x) -` (Pair ((T^^n)x) -`C))" by auto
also have "... = emeasure lborel (((+)(birkhoff_sum f n x) -` (Pair ((T^^n)x) -`C)) ∩ space lborel)" by simp
also have "... = emeasure (distr lborel borel ((+) (birkhoff_sum f n x))) (Pair ((T^^n)x) -`C)"
apply (rule emeasure_distr[symmetric]) using C_meas by auto
also have "... = emeasure lborel (Pair ((T^^n)x) -`C)" using lborel_distr_plus[of "birkhoff_sum f n x"] by simp
also have "... ≥ e" using Cx_fibers ‹(T^^n) x ∈ Cx› by auto
finally have "emeasure lborel F ≥ e" by auto
then show ?thesis using A by (simp add: indicator_def)
qed
}
moreover have "emeasure ?MS (D n) = (∫⇧+x. (∫⇧+y. indicator (D n) (x, y) ∂lborel) ∂M)"
using Dn_meas lborel.emeasure_pair_measure by blast
ultimately have "emeasure ?MS (D n) ≥ (∫⇧+x. ennreal e * indicator (Dx ∩ (T ^^ n) -` Cx) x ∂M)"
by (simp add: nn_integral_mono)
also have "(∫⇧+x. ennreal e * indicator (Dx ∩ (T ^^ n) -` Cx) x ∂M) = e * (∫⇧+x. indicator (Dx ∩ (T ^^ n) -` Cx) x ∂M)"
apply (rule nn_integral_cmult) using ‹e>0› by auto
also have "... = ennreal e * emeasure M (Dx ∩ (T ^^ n) -` Cx)" by simp
finally have *: "emeasure ?MS (D n) ≥ ennreal e * emeasure M (Dx ∩ (T ^^ n) -` Cx)" by auto
have "c + emeasure M (space M) ≤ emeasure M Dx + emeasure M Cx"
using ‹emeasure M Dx + c ≥ emeasure M (space M)› unfolding c_def
by (auto simp: emeasure_eq_measure ennreal_plus[symmetric] simp del: ennreal_plus)
also have "... = emeasure M Dx + emeasure M ((T^^n)--`Cx)"
by (simp add: T_vrestr_same_emeasure(2))
also have "... = emeasure M (Dx ∪ ((T^^n)--`Cx)) + emeasure M (Dx ∩ ((T^^n)--`Cx))"
by (rule emeasure_Un_Int, auto)
also have "... ≤ emeasure M (space M) + emeasure M (Dx ∩ ((T^^n)-`Cx))"
proof -
have "emeasure M (Dx ∪ ((T^^n)--`Cx)) ≤ emeasure M (space M)"
by (rule emeasure_mono, auto simp add: sets.sets_into_space)
moreover have "Dx ∩ ((T^^n)--`Cx) = Dx ∩ ((T^^n)-`Cx)"
by (simp add: vrestr_intersec_in_space)
ultimately show ?thesis by (metis add.commute add_left_mono)
qed
finally have "emeasure M (Dx ∩ ((T^^n)-`Cx)) ≥ c" by (simp add: emeasure_eq_measure)
then have "ennreal e * emeasure M (Dx ∩ ((T^^n)-`Cx)) ≥ ennreal e * c" using ‹e > 0›
using mult_left_mono by fastforce
with * show "emeasure ?MS (D n) ≥ e * c"
using ‹0<c› ‹0<e› by (auto simp: ennreal_mult[symmetric])
qed
have "¬(disjoint_family_on D {n0..n1})"
proof
assume D: "disjoint_family_on D {n0..n1}"
have "emeasure lborel {-r*n1-N..r*n1+N} = (r * real n1 + real N) - (- r * real n1 - real N)"
apply (rule emeasure_lborel_Icc) using ‹r>0› by auto
then have *: "emeasure lborel {-r*n1-N..r*n1+N} = ennreal(2 * r * real n1 + 2 * real N)"
by (auto simp: ac_simps)
have "ennreal(e * c) * (real n1 - real n0 + 1) = ennreal(e*c) * card {n0..n1}"
using ‹n1 > n0› by (auto simp: ennreal_of_nat_eq_real_of_nat Suc_diff_le ac_simps of_nat_diff)
also have "... = (∑n∈{n0..n1}. ennreal(e*c))"
by (simp add: ac_simps)
also have "... ≤ (∑n∈{n0..n1}. emeasure ?MS (D n))"
using ‹⋀n. n ∈ {n0..n1} ⟹ emeasure ?MS (D n) ≥ e * c› by (meson sum_mono)
also have "... = emeasure ?MS (⋃n∈{n0..n1}. D n)"
apply (rule sum_emeasure) using Dn_meas by (auto simp add: D)
also have "... ≤ emeasure ?MS (space M × {-r*n1-N..r*n1+N})"
apply (rule emeasure_mono) unfolding D_def using sets.sets_into_space[OF Dx_meas] by auto
also have "... = emeasure M (space M) * emeasure lborel {-r*n1-N..r*n1+N}"
by (rule sigma_finite_measure.emeasure_pair_measure_Times, auto simp add: lborel.sigma_finite_measure_axioms)
also have "... = emeasure M (space M) * ennreal(2 * r * real n1 + 2 * real N)"
using * by auto
finally have "ennreal(e * c) * (real n1- real n0+1) ≤ emeasure M (space M) * ennreal(2 * r * real n1 + 2 * real N)" by simp
then have "e*c * (real n1- real n0 + 1) ≤ measure M (space M) * (2 * r * real n1 + 2 * real N)"
using ‹0<r› ‹0<e› ‹0<c› ‹n0 < n1› emeasure_eq_measure by (auto simp: ennreal_mult'[symmetric] simp del: ennreal_plus)
then have "0 ≤ measure M (space M) * (2 * r * real n1 + 2 * real N) - e*c * (real n1- real n0 + 1)" by auto
also have "... = (measure M (space M) * 2 * N + e*c*n0 - e*c) - n1 * (e*c-2*r*measure M (space M))"
by algebra
finally have "n1 * (e*c-2*r*measure M (space M)) ≤ measure M (space M) * 2 * N + e*c*n0 - e*c"
by linarith
then show False using n1_ineq by auto
qed
then obtain n m where nm: "n<m" "D m ∩ D n ≠ {}" unfolding disjoint_family_on_def by (metis inf_sup_aci(1) linorder_cases)
define k where "k = m-n"
then have "k>0" "D (n+k) ∩ D n ≠ {}" using nm by auto
then have "((?TS^^(n+k))-`A) ∩ ((?TS^^n)-`A) ≠ {}" unfolding D_def C_def B_def by auto
moreover have "((?TS^^(n+k))-`A) ∩ ((?TS^^n)-`A) = (?TS^^n)-`(((?TS^^k)-`A) ∩ A)"
using funpow_add by (simp add: add.commute funpow_add set.compositionality)
ultimately have "((?TS^^k)-`A) ∩ A ≠ {}" by auto
then show "∃k>0. ((?TS^^k)-`A) ∩ A ≠ {}" using ‹k>0› by auto
qed
subsubsection ‹Oscillations around the limit in Birkhoff theorem›
text ‹In this paragraph, we prove that, in Birkhoff theorem with vanishing limit, the Birkhoff sums
are infinitely many times arbitrarily close to $0$, both on the positive and the negative side.
In the ergodic case, this statement implies for instance that if the Birkhoff sums of an integrable
function tend to infinity almost everywhere, then the integral of the function can not vanish, it
has to be strictly positive (while Birkhoff theorem per se does not exclude the convergence to
infinity, at a rate slower than linear). This converts a qualitative information (convergence to
infinity at an unknown rate) to a quantitative information (linear convergence to infinity). This
result (sometimes known as Atkinson's Lemma) has been reinvented many times, for instance by Kesten
and by Guivarch. It plays an important role in the study of random products of matrices.
This is essentially a consequence of the conservativity of the corresponding skew-product, proved in
\verb+skew_product_conservative+. Indeed, this implies that, starting from a small set $X \times
[-e/2, e/2]$, the skew-product comes back infinitely often to itself, which implies that the
Birkhoff sums at these times are bounded by $e$.
To show that the Birkhoff sums come back to $[0,e]$ is a little bit more tricky. Argue by
contradiction, and induce on $A \times [0,e/2]$ where $A$ is the set of points where the Birkhoff
sums don't come back to $[0,e]$. Then the second coordinate decreases strictly when one iterates the
skew product, which is not compatible with conservativity.›
lemma birkhoff_sum_small_asymp_lemma:
assumes [measurable]: "integrable M f"
and "AE x in M. real_cond_exp M Invariants f x = 0" "e>(0::real)"
shows "AE x in M. infinite {n. birkhoff_sum f n x ∈ {0..e}}"
proof -
have [measurable]: "f ∈ borel_measurable M" by auto
have [measurable]: "⋀N. {x ∈ space M. ∃N. ∀n∈{N..}. birkhoff_sum f n x ∉ {0..e}} ∈ sets M" by auto
{
fix N assume "N>(0::nat)"
define Ax where "Ax = {x ∈ space M. ∀n∈{N..}. birkhoff_sum f n x ∉ {0..e}}"
then have [measurable]: "Ax ∈ sets M" by auto
define A where "A = Ax × {0..e/2}"
then have A_meas [measurable]: "A ∈ sets (M ⨂⇩M lborel)" by auto
define TN where "TN = T^^N"
interpret TN: fmpt M TN
unfolding TN_def using fmpt_power by auto
define fN where "fN = birkhoff_sum f N"
have "TN.birkhoff_sum fN n x = birkhoff_sum f (n*N) x" for n x
proof (induction n)
case 0
then show ?case by auto
next
case (Suc n)
have "TN.birkhoff_sum fN (Suc n) x = TN.birkhoff_sum fN n x + fN ((TN^^n) x)"
using TN.birkhoff_sum_cocycle[of fN n 1] by auto
also have "... = birkhoff_sum f (n*N) x + birkhoff_sum f N ((TN^^n) x)"
using Suc.IH fN_def by auto
also have "... = birkhoff_sum f (n*N+N) x" unfolding TN_def
by (subst funpow_mult, subst mult.commute[of N n], rule birkhoff_sum_cocycle[of f "n*N" N x, symmetric])
finally show ?case by (simp add: add.commute)
qed
then have not0e: "⋀x n. x ∈ Ax ⟹ n ≠ 0 ⟹ TN.birkhoff_sum fN n x ∉ {0..e}" unfolding Ax_def by auto
let ?TS = "(λ(x,y). (T x, y + f x))"
let ?MS = "M ⨂⇩M (lborel::real measure)"
interpret TS: conservative_mpt ?MS ?TS
by (rule skew_product_conservative, auto simp add: assms)
let ?TSN = "(λ(x,y). (TN x, y + fN x))"
have *:"?TSN = ?TS^^N" unfolding TN_def fN_def using skew_product_real_iterates by auto
interpret TSN: conservative_mpt ?MS ?TSN
using * TS.conservative_mpt_power by auto
define MA TA where "MA = restrict_space ?MS A" and "TA = TSN.induced_map A"
interpret TA: conservative_mpt MA TA unfolding MA_def TA_def
by (rule TSN.induced_map_conservative_mpt, measurable)
have *: "⋀ x y. snd (TA (x,y)) = snd (x,y) + TN.birkhoff_sum fN (TSN.return_time_function A (x,y)) x"
unfolding TA_def TSN.induced_map_def using TN.skew_product_real_iterates Pair_def by auto
have [measurable]: "snd ∈ borel_measurable ?MS" by auto
then have [measurable]: "snd ∈ borel_measurable MA" unfolding MA_def using measurable_restrict_space1 by blast
have "AE z in MA. z ∈ TSN.recurrent_subset A"
unfolding MA_def using TSN.induced_map_recurrent_typical(1)[OF A_meas].
moreover
{
fix z assume z: "z ∈ TSN.recurrent_subset A"
define x y where "x = fst z" and "y = snd z"
then have "z = (x,y)" by simp
have "z ∈ A" using z "TSN.recurrent_subset_incl"(1) by auto
then have "x ∈ Ax" "y ∈ {0..e/2}" unfolding A_def x_def y_def by auto
define y2 where "y2 = y + TN.birkhoff_sum fN (TSN.return_time_function A (x,y)) x"
have "y2 = snd (TA z)" unfolding y2_def using * ‹z = (x, y)› by force
moreover have "TA z ∈ A" unfolding TA_def using ‹z ∈ A› TSN.induced_map_stabilizes_A by auto
ultimately have "y2 ∈ {0..e/2}" unfolding A_def by auto
have "TSN.return_time_function A (x,y) ≠ 0"
using ‹z = (x,y)› ‹z ∈ TSN.recurrent_subset A› TSN.return_time0[of A] by fast
then have "TN.birkhoff_sum fN (TSN.return_time_function A (x,y)) x ∉ {0..e}"
using not0e[OF ‹x ∈ Ax›] by auto
moreover have "TN.birkhoff_sum fN (TSN.return_time_function A (x,y)) x ∈ {-e..e}"
using ‹y ∈ {0..e/2}› ‹y2 ∈ {0..e/2}› y2_def by auto
ultimately have "TN.birkhoff_sum fN (TSN.return_time_function A (x,y)) x ∈ {-e..<0}"
by auto
then have "y2 < y" using y2_def by auto
then have "snd(TA z) < snd z" unfolding y_def using ‹y2 = snd (TA z)› by auto
}
ultimately have a: "AE z in MA. snd(TA z) < snd z" by auto
then have "AE z in MA. snd(TA z) ≤ snd z" by auto
then have "AE z in MA. snd(TA z) = snd z" using TA.AE_decreasing_then_invariant[of snd] by auto
then have "AE z in MA. False" using a by auto
then have "space MA ∈ null_sets MA" by (simp add: AE_iff_null_sets)
then have "emeasure MA A = 0" by (metis A_meas MA_def null_setsD1 space_restrict_space2)
then have "emeasure ?MS A = 0" unfolding MA_def
by (metis A_meas emeasure_restrict_space sets.sets_into_space sets.top space_restrict_space space_restrict_space2)
moreover have "emeasure ?MS A = emeasure M Ax * emeasure lborel {0..e/2}"
unfolding A_def by (intro lborel.emeasure_pair_measure_Times) auto
ultimately have "emeasure M {x ∈ space M. ∀n∈{N..}. birkhoff_sum f n x ∉ {0..e}} = 0" using ‹e>0› Ax_def by simp
then have "{x ∈ space M. ∀n∈{N..}. birkhoff_sum f n x ∉ {0..e}} ∈ null_sets M" by auto
}
then have "(⋃N∈{0<..}. {x ∈ space M. ∀n∈{N..}. birkhoff_sum f n x ∉ {0..e}}) ∈ null_sets M" by (auto simp: greaterThan_0)
moreover have "{x ∈ space M. ¬(infinite {n. birkhoff_sum f n x ∈ {0..e}})} ⊆ (⋃N∈{0<..}. {x ∈ space M. ∀n∈{N..}. birkhoff_sum f n x ∉ {0..e}})"
proof
fix x assume H: "x ∈ {x ∈ space M. ¬(infinite {n. birkhoff_sum f n x ∈ {0..e}})}"
then have "x ∈ space M" by auto
have *: "finite {n. birkhoff_sum f n x ∈ {0..e}}" using H by auto
then obtain N where "⋀n. n ≥ N ⟹ n ∉ {n. birkhoff_sum f n x ∈ {0..e}}"
by (metis finite_nat_set_iff_bounded not_less)
then have "x ∈ {x ∈ space M. ∀n∈{N+1..}. birkhoff_sum f n x ∉ {0..e}}" using ‹x ∈ space M› by auto
moreover have "N+1>0" by auto
ultimately show "x ∈ (⋃N∈{0<..}. {x ∈ space M. ∀n∈{N..}. birkhoff_sum f n x ∉ {0..e}})" by auto
qed
ultimately show ?thesis unfolding eventually_ae_filter by auto
qed
theorem birkhoff_sum_small_asymp_pos_nonergodic:
assumes [measurable]: "integrable M f" and "e > (0::real)"
shows "AE x in M. infinite {n. birkhoff_sum f n x ∈ {n * real_cond_exp M Invariants f x .. n * real_cond_exp M Invariants f x + e}}"
proof -
define g where "g = (λx. f x - real_cond_exp M Invariants f x)"
have g_meas [measurable]: "integrable M g" unfolding g_def using real_cond_exp_int(1)[OF assms(1)] assms(1) by auto
have "AE x in M. real_cond_exp M Invariants (real_cond_exp M Invariants f) x = real_cond_exp M Invariants f x"
by (rule real_cond_exp_F_meas, auto simp add: real_cond_exp_int(1)[OF assms(1)])
then have *: "AE x in M. real_cond_exp M Invariants g x = 0"
unfolding g_def using real_cond_exp_diff[OF assms(1) real_cond_exp_int(1)[OF assms(1)]] by auto
have "AE x in M. infinite {n. birkhoff_sum g n x ∈ {0..e}}"
by (rule birkhoff_sum_small_asymp_lemma, auto simp add: ‹e>0› * g_meas)
moreover
{
fix x assume "x ∈ space M" "infinite {n. birkhoff_sum g n x ∈ {0..e}}"
{
fix n assume H: "birkhoff_sum g n x ∈ {0..e}"
have "birkhoff_sum g n x = birkhoff_sum f n x - birkhoff_sum (real_cond_exp M Invariants f) n x"
unfolding g_def using birkhoff_sum_diff by auto
also have "... = birkhoff_sum f n x - n * real_cond_exp M Invariants f x"
using birkhoff_sum_of_invariants ‹x ∈ space M› by auto
finally have "birkhoff_sum f n x ∈ {n * real_cond_exp M Invariants f x .. n * real_cond_exp M Invariants f x + e}" using H by simp
}
then have "{n. birkhoff_sum g n x ∈ {0..e}} ⊆ {n. birkhoff_sum f n x ∈ {n * real_cond_exp M Invariants f x .. n * real_cond_exp M Invariants f x + e}}"
by auto
then have "infinite {n. birkhoff_sum f n x ∈ {n * real_cond_exp M Invariants f x .. n * real_cond_exp M Invariants f x + e}}"
using ‹infinite {n. birkhoff_sum g n x ∈ {0..e}}› finite_subset by blast
}
ultimately show ?thesis by auto
qed
theorem birkhoff_sum_small_asymp_neg_nonergodic:
assumes [measurable]: "integrable M f" and "e > (0::real)"
shows "AE x in M. infinite {n. birkhoff_sum f n x ∈ {n * real_cond_exp M Invariants f x - e .. n * real_cond_exp M Invariants f x}}"
proof -
define g where "g = (λx. real_cond_exp M Invariants f x - f x)"
have g_meas [measurable]: "integrable M g" unfolding g_def using real_cond_exp_int(1)[OF assms(1)] assms(1) by auto
have "AE x in M. real_cond_exp M Invariants (real_cond_exp M Invariants f) x = real_cond_exp M Invariants f x"
by (rule real_cond_exp_F_meas, auto simp add: real_cond_exp_int(1)[OF assms(1)])
then have *: "AE x in M. real_cond_exp M Invariants g x = 0"
unfolding g_def using real_cond_exp_diff[OF real_cond_exp_int(1)[OF assms(1)] assms(1)] by auto
have "AE x in M. infinite {n. birkhoff_sum g n x ∈ {0..e}}"
by (rule birkhoff_sum_small_asymp_lemma, auto simp add: ‹e>0› * g_meas)
moreover
{
fix x assume "x ∈ space M" "infinite {n. birkhoff_sum g n x ∈ {0..e}}"
{
fix n assume H: "birkhoff_sum g n x ∈ {0..e}"
have "birkhoff_sum g n x = birkhoff_sum (real_cond_exp M Invariants f) n x - birkhoff_sum f n x"
unfolding g_def using birkhoff_sum_diff by auto
also have "... = n * real_cond_exp M Invariants f x - birkhoff_sum f n x"
using birkhoff_sum_of_invariants ‹x ∈ space M› by auto
finally have "birkhoff_sum f n x ∈ {n * real_cond_exp M Invariants f x - e .. n * real_cond_exp M Invariants f x}" using H by simp
}
then have "{n. birkhoff_sum g n x ∈ {0..e}} ⊆ {n. birkhoff_sum f n x ∈ {n * real_cond_exp M Invariants f x - e .. n * real_cond_exp M Invariants f x}}"
by auto
then have "infinite {n. birkhoff_sum f n x ∈ {n * real_cond_exp M Invariants f x - e .. n * real_cond_exp M Invariants f x}}"
using ‹infinite {n. birkhoff_sum g n x ∈ {0..e}}› finite_subset by blast
}
ultimately show ?thesis by auto
qed
subsubsection ‹Conditional expectation for the induced map›
text ‹Thanks to Birkhoff theorem, one can relate conditional expectations with respect to the invariant
sigma algebra, for a map and for a corresponding induced map, as follows.›
proposition Invariants_cond_exp_induced_map:
fixes f::"'a ⇒ real"
assumes [measurable]: "A ∈ sets M" "integrable M f"
defines "MA ≡ restrict_space M A" and "TA ≡ induced_map A" and "fA ≡ induced_function A f"
shows "AE x in MA. real_cond_exp MA (qmpt.Invariants MA TA) fA x
= real_cond_exp M Invariants f x * real_cond_exp MA (qmpt.Invariants MA TA) (return_time_function A) x"
proof -
interpret A: fmpt MA TA unfolding MA_def TA_def by (rule induced_map_fmpt[OF assms(1)])
have "integrable M fA" unfolding fA_def using induced_function_integral_nonergodic(1) assms by auto
then have "integrable MA fA" unfolding MA_def
by (metis assms(1) integrable_mult_indicator integrable_restrict_space sets.Int_space_eq2)
then have a: "AE x in MA. (λn. A.birkhoff_sum fA n x / n) ⇢ real_cond_exp MA A.Invariants fA x"
using A.birkhoff_theorem_AE_nonergodic by auto
have "AE x in M. (λn. birkhoff_sum f n x / n) ⇢ real_cond_exp M Invariants f x"
using birkhoff_theorem_AE_nonergodic assms(2) by auto
then have b: "AE x in MA. (λn. birkhoff_sum f n x / n) ⇢ real_cond_exp M Invariants f x"
unfolding MA_def by (metis (mono_tags, lifting) AE_restrict_space_iff assms(1) eventually_mono sets.Int_space_eq2)
define phiA where "phiA = (λx. return_time_function A x)"
have "integrable M phiA" unfolding phiA_def using return_time_integrable by auto
then have "integrable MA phiA" unfolding MA_def
by (metis assms(1) integrable_mult_indicator integrable_restrict_space sets.Int_space_eq2)
then have c: "AE x in MA. (λn. A.birkhoff_sum (λx. real(phiA x)) n x / n) ⇢ real_cond_exp MA A.Invariants phiA x"
using A.birkhoff_theorem_AE_nonergodic by auto
have "A-recurrent_subset A ∈ null_sets M" using Poincare_recurrence_thm(1)[OF assms(1)] by auto
then have "A - recurrent_subset A ∈ null_sets MA" unfolding MA_def
by (metis Diff_subset assms(1) emeasure_restrict_space null_setsD1 null_setsD2 null_setsI sets.Int_space_eq2 sets_restrict_space_iff)
then have "AE x in MA. x ∈ recurrent_subset A"
by (simp add: AE_iff_null MA_def null_setsD2 set_diff_eq space_restrict_space2)
moreover have "⋀x. x ∈ recurrent_subset A ⟹ phiA x > 0" unfolding phiA_def using return_time0 by fastforce
ultimately have *: "AE x in MA. phiA x > 0" by auto
have d: "AE x in MA. real_cond_exp MA A.Invariants phiA x > 0"
by (rule A.real_cond_exp_gr_c, auto simp add: * ‹integrable MA phiA›)
{
fix x
assume A: "(λn. A.birkhoff_sum fA n x / n) ⇢ real_cond_exp MA A.Invariants fA x"
and B: "(λn. birkhoff_sum f n x / n) ⇢ real_cond_exp M Invariants f x"
and C: "(λn. A.birkhoff_sum (λx. real(phiA x)) n x / n) ⇢ real_cond_exp MA A.Invariants phiA x"
and D: "real_cond_exp MA A.Invariants phiA x > 0"
define R where "R = (λn. A.birkhoff_sum phiA n x)"
have D2: "ereal(real_cond_exp MA A.Invariants phiA x) > 0" using D by simp
have "⋀n. real(R n)/ n = A.birkhoff_sum (λx. real(phiA x)) n x / n" unfolding R_def A.birkhoff_sum_def by auto
moreover have "(λn. A.birkhoff_sum (λx. real(phiA x)) n x / n) ⇢ real_cond_exp MA A.Invariants phiA x" using C by auto
ultimately have Rnn: "(λn. real(R n)/n) ⇢ real_cond_exp MA A.Invariants phiA x" by presburger
have "⋀n. ereal(real(R n))/ n = ereal(A.birkhoff_sum (λx. real(phiA x)) n x / n)" unfolding R_def A.birkhoff_sum_def by auto
moreover have "(λn. ereal(A.birkhoff_sum (λx. real(phiA x)) n x / n)) ⇢ real_cond_exp MA A.Invariants phiA x" using C by auto
ultimately have i: "(λn. ereal(real(R n))/n) ⇢ real_cond_exp MA A.Invariants phiA x" by auto
have ii: "(λn. real n) ⇢ ∞" by (rule id_nat_ereal_tendsto_PInf)
have iii: "(λn. ereal(real(R n))/n * real n) ⇢ ∞" using tendsto_mult_ereal_PInf[OF i D2 ii] by simp
have "⋀n. n > 0 ⟹ ereal(real(R n))/n * real n = R n" by auto
then have "eventually (λn. ereal(real(R n))/n * real n = R n) sequentially" using eventually_at_top_dense by auto
then have "(λn. real(R n)) ⇢ ∞" using iii by (simp add: filterlim_cong)
then have "(λn. birkhoff_sum f (R n) x / (R n)) ⇢ real_cond_exp M Invariants f x" using limit_along_weak_subseq B by auto
then have l: "(λn. (birkhoff_sum f (R n) x / (R n)) * ((R n)/n)) ⇢ real_cond_exp M Invariants f x * real_cond_exp MA A.Invariants phiA x"
by (rule tendsto_mult, simp add: Rnn)
obtain N where N: "⋀n. n > N ⟹ R n > 0" using ‹(λn. real(R n)) ⇢ ∞›
by (metis (full_types) eventually_at_top_dense filterlim_iff filterlim_weak_subseq)
then have "⋀n. n > N ⟹ (birkhoff_sum f (R n) x / (R n)) * ((R n)/n) = birkhoff_sum f (R n) x / n"
by auto
then have "eventually (λn. (birkhoff_sum f (R n) x / (R n)) * ((R n)/n) = birkhoff_sum f (R n) x / n) sequentially"
by simp
with tendsto_cong[OF this] have main_limit: "(λn. birkhoff_sum f (R n) x / n) ⇢ real_cond_exp M Invariants f x * real_cond_exp MA A.Invariants phiA x"
using l by auto
have "⋀n. birkhoff_sum f (R n) x = A.birkhoff_sum fA n x"
unfolding R_def fA_def phiA_def TA_def using induced_function_birkhoff_sum[OF assms(1)] by simp
then have "⋀n. birkhoff_sum f (R n) x /n = A.birkhoff_sum fA n x / n" by simp
then have "(λn. A.birkhoff_sum fA n x / n) ⇢ real_cond_exp M Invariants f x * real_cond_exp MA A.Invariants phiA x"
using main_limit by presburger
then have "real_cond_exp MA A.Invariants fA x = real_cond_exp M Invariants f x * real_cond_exp MA A.Invariants phiA x"
using A LIMSEQ_unique by auto
}
then show ?thesis using a b c d unfolding phiA_def by auto
qed
corollary Invariants_cond_exp_induced_map_0:
fixes f::"'a ⇒ real"
assumes [measurable]: "A ∈ sets M" "integrable M f" and "AE x in M. real_cond_exp M Invariants f x = 0"
defines "MA ≡ restrict_space M A" and "TA ≡ induced_map A" and "fA ≡ induced_function A f"
shows "AE x in MA. real_cond_exp MA (qmpt.Invariants MA TA) fA x = 0"
proof -
have "AE x in MA. real_cond_exp M Invariants f x = 0" unfolding MA_def
apply (subst AE_restrict_space_iff) using assms(3) by auto
then show ?thesis unfolding MA_def TA_def fA_def using Invariants_cond_exp_induced_map[OF assms(1) assms(2)]
by auto
qed
end
end