(* Title: HOL/Probability/Characteristic_Functions.thy
Authors: Jeremy Avigad (CMU), Luke Serafin (CMU), Johannes Hölzl (TUM)
*)
section \<open>Characteristic Functions\<close>
theory Characteristic_Functions
imports Weak_Convergence Independent_Family Distributions
begin
lemma mult_min_right: "a \<ge> 0 \<Longrightarrow> (a :: real) * min b c = min (a * b) (a * c)"
by (metis min.absorb_iff2 min_def mult_left_mono)
lemma sequentially_even_odd:
assumes E: "eventually (\<lambda>n. P (2 * n)) sequentially" and O: "eventually (\<lambda>n. P (2 * n + 1)) sequentially"
shows "eventually P sequentially"
proof -
from E obtain n_e where [intro]: "\<And>n. n \<ge> n_e \<Longrightarrow> P (2 * n)"
by (auto simp: eventually_sequentially)
moreover
from O obtain n_o where [intro]: "\<And>n. n \<ge> n_o \<Longrightarrow> P (Suc (2 * n))"
by (auto simp: eventually_sequentially)
show ?thesis
unfolding eventually_sequentially
proof (intro exI allI impI)
fix n assume "max (2 * n_e) (2 * n_o + 1) \<le> n" then show "P n"
by (cases "even n") (auto elim!: evenE oddE )
qed
qed
lemma limseq_even_odd:
assumes "(\<lambda>n. f (2 * n)) \<longlonglongrightarrow> (l :: 'a :: topological_space)"
and "(\<lambda>n. f (2 * n + 1)) \<longlonglongrightarrow> l"
shows "f \<longlonglongrightarrow> l"
using assms by (auto simp: filterlim_iff intro: sequentially_even_odd)
subsection \<open>Application of the FTC: integrating $e^ix$\<close>
abbreviation iexp :: "real \<Rightarrow> complex" where
"iexp \<equiv> (\<lambda>x. exp (\<i> * complex_of_real x))"
lemma isCont_iexp [simp]: "isCont iexp x"
by (intro continuous_intros)
lemma has_vector_derivative_iexp[derivative_intros]:
"(iexp has_vector_derivative \<i> * iexp x) (at x within s)"
by (auto intro!: derivative_eq_intros simp: Re_exp Im_exp has_vector_derivative_complex_iff)
lemma interval_integral_iexp:
fixes a b :: real
shows "(CLBINT x=a..b. iexp x) = \<i> * iexp a - \<i> * iexp b"
by (subst interval_integral_FTC_finite [where F = "\<lambda>x. -\<i> * iexp x"])
(auto intro!: derivative_eq_intros continuous_intros)
subsection \<open>The Characteristic Function of a Real Measure.\<close>
definition
char :: "real measure \<Rightarrow> real \<Rightarrow> complex"
where
"char M t = CLINT x|M. iexp (t * x)"
lemma (in real_distribution) char_zero: "char M 0 = 1"
unfolding char_def by (simp del: space_eq_univ add: prob_space)
lemma (in prob_space) integrable_iexp:
assumes f: "f \<in> borel_measurable M" "\<And>x. Im (f x) = 0"
shows "integrable M (\<lambda>x. exp (\<i> * (f x)))"
proof (intro integrable_const_bound [of _ 1])
from f have "\<And>x. of_real (Re (f x)) = f x"
by (simp add: complex_eq_iff)
then show "AE x in M. cmod (exp (\<i> * f x)) \<le> 1"
using norm_exp_i_times[of "Re (f x)" for x] by simp
qed (insert f, simp)
lemma (in real_distribution) cmod_char_le_1: "norm (char M t) \<le> 1"
proof -
have "norm (char M t) \<le> (\<integral>x. norm (iexp (t * x)) \<partial>M)"
unfolding char_def by (intro integral_norm_bound)
also have "\<dots> \<le> 1"
by (simp del: of_real_mult)
finally show ?thesis .
qed
lemma (in real_distribution) isCont_char: "isCont (char M) t"
unfolding continuous_at_sequentially
proof safe
fix X assume X: "X \<longlonglongrightarrow> t"
show "(char M \<circ> X) \<longlonglongrightarrow> char M t"
unfolding comp_def char_def
by (rule integral_dominated_convergence[where w="\<lambda>_. 1"]) (auto intro!: tendsto_intros X)
qed
lemma (in real_distribution) char_measurable [measurable]: "char M \<in> borel_measurable borel"
by (auto intro!: borel_measurable_continuous_onI continuous_at_imp_continuous_on isCont_char)
subsection \<open>Independence\<close>
(* the automation can probably be improved *)
lemma (in prob_space) char_distr_add:
fixes X1 X2 :: "'a \<Rightarrow> real" and t :: real
assumes "indep_var borel X1 borel X2"
shows "char (distr M borel (\<lambda>\<omega>. X1 \<omega> + X2 \<omega>)) t =
char (distr M borel X1) t * char (distr M borel X2) t"
proof -
from assms have [measurable]: "random_variable borel X1" by (elim indep_var_rv1)
from assms have [measurable]: "random_variable borel X2" by (elim indep_var_rv2)
have "char (distr M borel (\<lambda>\<omega>. X1 \<omega> + X2 \<omega>)) t = (CLINT x|M. iexp (t * (X1 x + X2 x)))"
by (simp add: char_def integral_distr)
also have "\<dots> = (CLINT x|M. iexp (t * (X1 x)) * iexp (t * (X2 x))) "
by (simp add: field_simps exp_add)
also have "\<dots> = (CLINT x|M. iexp (t * (X1 x))) * (CLINT x|M. iexp (t * (X2 x)))"
by (intro indep_var_lebesgue_integral indep_var_compose[unfolded comp_def, OF assms])
(auto intro!: integrable_iexp)
also have "\<dots> = char (distr M borel X1) t * char (distr M borel X2) t"
by (simp add: char_def integral_distr)
finally show ?thesis .
qed
lemma (in prob_space) char_distr_sum:
"indep_vars (\<lambda>i. borel) X A \<Longrightarrow>
char (distr M borel (\<lambda>\<omega>. \<Sum>i\<in>A. X i \<omega>)) t = (\<Prod>i\<in>A. char (distr M borel (X i)) t)"
proof (induct A rule: infinite_finite_induct)
case (insert x F) with indep_vars_subset[of "\<lambda>_. borel" X "insert x F" F] show ?case
by (auto simp add: char_distr_add indep_vars_sum)
qed (simp_all add: char_def integral_distr prob_space del: distr_const)
subsection \<open>Approximations to $e^{ix}$\<close>
text \<open>Proofs from Billingsley, page 343.\<close>
lemma CLBINT_I0c_power_mirror_iexp:
fixes x :: real and n :: nat
defines "f s m \<equiv> complex_of_real ((x - s) ^ m)"
shows "(CLBINT s=0..x. f s n * iexp s) =
x^Suc n / Suc n + (\<i> / Suc n) * (CLBINT s=0..x. f s (Suc n) * iexp s)"
proof -
have 1:
"((\<lambda>s. complex_of_real(-((x - s) ^ (Suc n) / (Suc n))) * iexp s)
has_vector_derivative complex_of_real((x - s)^n) * iexp s + (\<i> * iexp s) * complex_of_real(-((x - s) ^ (Suc n) / (Suc n))))
(at s within A)" for s A
by (intro derivative_eq_intros) auto
let ?F = "\<lambda>s. complex_of_real(-((x - s) ^ (Suc n) / (Suc n))) * iexp s"
have "x^(Suc n) / (Suc n) = (CLBINT s=0..x. (f s n * iexp s + (\<i> * iexp s) * -(f s (Suc n) / (Suc n))))" (is "?LHS = ?RHS")
proof -
have "?RHS = (CLBINT s=0..x. (f s n * iexp s + (\<i> * iexp s) *
complex_of_real(-((x - s) ^ (Suc n) / (Suc n)))))"
by (cases "0 \<le> x") (auto intro!: simp: f_def[abs_def])
also have "... = ?F x - ?F 0"
unfolding zero_ereal_def using 1
by (intro interval_integral_FTC_finite)
(auto simp: f_def add_nonneg_eq_0_iff complex_eq_iff
intro!: continuous_at_imp_continuous_on continuous_intros)
finally show ?thesis
by auto
qed
show ?thesis
unfolding \<open>?LHS = ?RHS\<close> f_def interval_lebesgue_integral_mult_right [symmetric]
by (subst interval_lebesgue_integral_add(2) [symmetric])
(auto intro!: interval_integrable_isCont continuous_intros simp: zero_ereal_def complex_eq_iff)
qed
lemma iexp_eq1:
fixes x :: real
defines "f s m \<equiv> complex_of_real ((x - s) ^ m)"
shows "iexp x =
(\<Sum>k \<le> n. (\<i> * x)^k / (fact k)) + ((\<i> ^ (Suc n)) / (fact n)) * (CLBINT s=0..x. (f s n) * (iexp s))" (is "?P n")
proof (induction n)
show "?P 0"
by (auto simp add: field_simps interval_integral_iexp f_def zero_ereal_def)
next
fix n assume ih: "?P n"
have **: "\<And>a b :: real. a = -b \<longleftrightarrow> a + b = 0"
by linarith
have *: "of_nat n * of_nat (fact n) \<noteq> - (of_nat (fact n)::complex)"
unfolding of_nat_mult[symmetric]
by (simp add: complex_eq_iff ** of_nat_add[symmetric] del: of_nat_mult of_nat_fact of_nat_add)
show "?P (Suc n)"
unfolding sum.atMost_Suc ih f_def CLBINT_I0c_power_mirror_iexp[of _ n]
by (simp add: divide_simps add_eq_0_iff *) (simp add: field_simps)
qed
lemma iexp_eq2:
fixes x :: real
defines "f s m \<equiv> complex_of_real ((x - s) ^ m)"
shows "iexp x = (\<Sum>k\<le>Suc n. (\<i>*x)^k/fact k) + \<i>^Suc n/fact n * (CLBINT s=0..x. f s n*(iexp s - 1))"
proof -
have isCont_f: "isCont (\<lambda>s. f s n) x" for n x
by (auto simp: f_def)
let ?F = "\<lambda>s. complex_of_real (-((x - s) ^ (Suc n) / real (Suc n)))"
have calc1: "(CLBINT s=0..x. f s n * (iexp s - 1)) =
(CLBINT s=0..x. f s n * iexp s) - (CLBINT s=0..x. f s n)"
unfolding zero_ereal_def
by (subst interval_lebesgue_integral_diff(2) [symmetric])
(simp_all add: interval_integrable_isCont isCont_f field_simps)
have calc2: "(CLBINT s=0..x. f s n) = x^Suc n / Suc n"
unfolding zero_ereal_def
proof (subst interval_integral_FTC_finite [where a = 0 and b = x and f = "\<lambda>s. f s n" and F = ?F])
show "(?F has_vector_derivative f y n) (at y within {min 0 x..max 0 x})" for y
unfolding f_def
by (intro has_vector_derivative_of_real)
(auto intro!: derivative_eq_intros simp del: power_Suc simp add: add_nonneg_eq_0_iff)
qed (auto intro: continuous_at_imp_continuous_on isCont_f)
have calc3: "\<i> ^ (Suc (Suc n)) / (fact (Suc n)) = (\<i> ^ (Suc n) / (fact n)) * (\<i> / (Suc n))"
by (simp add: field_simps)
show ?thesis
unfolding iexp_eq1 [where n = "Suc n" and x=x] calc1 calc2 calc3 unfolding f_def
by (subst CLBINT_I0c_power_mirror_iexp [where n = n]) auto
qed
lemma abs_LBINT_I0c_abs_power_diff:
"\<bar>LBINT s=0..x. \<bar>(x - s)^n\<bar>\<bar> = \<bar>x ^ (Suc n) / (Suc n)\<bar>"
proof -
have "\<bar>LBINT s=0..x. \<bar>(x - s)^n\<bar>\<bar> = \<bar>LBINT s=0..x. (x - s)^n\<bar>"
proof cases
assume "0 \<le> x \<or> even n"
then have "(LBINT s=0..x. \<bar>(x - s)^n\<bar>) = LBINT s=0..x. (x - s)^n"
by (auto simp add: zero_ereal_def power_even_abs power_abs min_absorb1 max_absorb2
intro!: interval_integral_cong)
then show ?thesis by simp
next
assume "\<not> (0 \<le> x \<or> even n)"
then have "(LBINT s=0..x. \<bar>(x - s)^n\<bar>) = LBINT s=0..x. -((x - s)^n)"
by (auto simp add: zero_ereal_def power_abs min_absorb1 max_absorb2
ereal_min[symmetric] ereal_max[symmetric] power_minus_odd[symmetric]
simp del: ereal_min ereal_max intro!: interval_integral_cong)
also have "\<dots> = - (LBINT s=0..x. (x - s)^n)"
by (subst interval_lebesgue_integral_uminus, rule refl)
finally show ?thesis by simp
qed
also have "LBINT s=0..x. (x - s)^n = x^Suc n / Suc n"
proof -
let ?F = "\<lambda>t. - ((x - t)^(Suc n) / Suc n)"
have "LBINT s=0..x. (x - s)^n = ?F x - ?F 0"
unfolding zero_ereal_def
by (intro interval_integral_FTC_finite continuous_at_imp_continuous_on
has_field_derivative_iff_has_vector_derivative[THEN iffD1])
(auto simp del: power_Suc intro!: derivative_eq_intros simp add: add_nonneg_eq_0_iff)
also have "\<dots> = x ^ (Suc n) / (Suc n)" by simp
finally show ?thesis .
qed
finally show ?thesis .
qed
lemma iexp_approx1: "cmod (iexp x - (\<Sum>k \<le> n. (\<i> * x)^k / fact k)) \<le> \<bar>x\<bar>^(Suc n) / fact (Suc n)"
proof -
have "iexp x - (\<Sum>k \<le> n. (\<i> * x)^k / fact k) =
((\<i> ^ (Suc n)) / (fact n)) * (CLBINT s=0..x. (x - s)^n * (iexp s))" (is "?t1 = ?t2")
by (subst iexp_eq1 [of _ n], simp add: field_simps)
then have "cmod (?t1) = cmod (?t2)"
by simp
also have "\<dots> = (1 / of_nat (fact n)) * cmod (CLBINT s=0..x. (x - s)^n * (iexp s))"
by (simp add: norm_mult norm_divide norm_power)
also have "\<dots> \<le> (1 / of_nat (fact n)) * \<bar>LBINT s=0..x. cmod ((x - s)^n * (iexp s))\<bar>"
by (intro mult_left_mono interval_integral_norm2)
(auto simp: zero_ereal_def intro: interval_integrable_isCont)
also have "\<dots> \<le> (1 / of_nat (fact n)) * \<bar>LBINT s=0..x. \<bar>(x - s)^n\<bar>\<bar>"
by (simp add: norm_mult del: of_real_diff of_real_power)
also have "\<dots> \<le> (1 / of_nat (fact n)) * \<bar>x ^ (Suc n) / (Suc n)\<bar>"
by (simp add: abs_LBINT_I0c_abs_power_diff)
also have "1 / real_of_nat (fact n::nat) * \<bar>x ^ Suc n / real (Suc n)\<bar> =
\<bar>x\<bar> ^ Suc n / fact (Suc n)"
by (simp add: abs_mult power_abs)
finally show ?thesis .
qed
lemma iexp_approx2: "cmod (iexp x - (\<Sum>k \<le> n. (\<i> * x)^k / fact k)) \<le> 2 * \<bar>x\<bar>^n / fact n"
proof (induction n) \<comment> \<open>really cases\<close>
case (Suc n)
have *: "\<And>a b. interval_lebesgue_integrable lborel a b f \<Longrightarrow> interval_lebesgue_integrable lborel a b g \<Longrightarrow>
\<bar>LBINT s=a..b. f s\<bar> \<le> \<bar>LBINT s=a..b. g s\<bar>"
if f: "\<And>s. 0 \<le> f s" and g: "\<And>s. f s \<le> g s" for f g :: "_ \<Rightarrow> real"
using order_trans[OF f g] f g
unfolding interval_lebesgue_integral_def interval_lebesgue_integrable_def set_lebesgue_integral_def set_integrable_def
by (auto simp: integral_nonneg_AE[OF AE_I2] intro!: integral_mono mult_mono)
have "iexp x - (\<Sum>k \<le> Suc n. (\<i> * x)^k / fact k) =
((\<i> ^ (Suc n)) / (fact n)) * (CLBINT s=0..x. (x - s)^n * (iexp s - 1))" (is "?t1 = ?t2")
unfolding iexp_eq2 [of x n] by (simp add: field_simps)
then have "cmod (?t1) = cmod (?t2)"
by simp
also have "\<dots> = (1 / (fact n)) * cmod (CLBINT s=0..x. (x - s)^n * (iexp s - 1))"
by (simp add: norm_mult norm_divide norm_power)
also have "\<dots> \<le> (1 / (fact n)) * \<bar>LBINT s=0..x. cmod ((x - s)^n * (iexp s - 1))\<bar>"
by (intro mult_left_mono interval_integral_norm2)
(auto intro!: interval_integrable_isCont simp: zero_ereal_def)
also have "\<dots> = (1 / (fact n)) * \<bar>LBINT s=0..x. abs ((x - s)^n) * cmod((iexp s - 1))\<bar>"
by (simp add: norm_mult del: of_real_diff of_real_power)
also have "\<dots> \<le> (1 / (fact n)) * \<bar>LBINT s=0..x. abs ((x - s)^n) * 2\<bar>"
by (intro mult_left_mono * order_trans [OF norm_triangle_ineq4])
(auto simp: mult_ac zero_ereal_def intro!: interval_integrable_isCont)
also have "\<dots> = (2 / (fact n)) * \<bar>x ^ (Suc n) / (Suc n)\<bar>"
by (simp add: abs_LBINT_I0c_abs_power_diff abs_mult)
also have "2 / fact n * \<bar>x ^ Suc n / real (Suc n)\<bar> = 2 * \<bar>x\<bar> ^ Suc n / (fact (Suc n))"
by (simp add: abs_mult power_abs)
finally show ?case .
qed (insert norm_triangle_ineq4[of "iexp x" 1], simp)
lemma (in real_distribution) char_approx1:
assumes integrable_moments: "\<And>k. k \<le> n \<Longrightarrow> integrable M (\<lambda>x. x^k)"
shows "cmod (char M t - (\<Sum>k \<le> n. ((\<i> * t)^k / fact k) * expectation (\<lambda>x. x^k))) \<le>
(2*\<bar>t\<bar>^n / fact n) * expectation (\<lambda>x. \<bar>x\<bar>^n)" (is "cmod (char M t - ?t1) \<le> _")
proof -
have integ_iexp: "integrable M (\<lambda>x. iexp (t * x))"
by (intro integrable_const_bound) auto
define c where [abs_def]: "c k x = (\<i> * t)^k / fact k * complex_of_real (x^k)" for k x
have integ_c: "\<And>k. k \<le> n \<Longrightarrow> integrable M (\<lambda>x. c k x)"
unfolding c_def by (intro integrable_mult_right integrable_of_real integrable_moments)
have "k \<le> n \<Longrightarrow> expectation (c k) = (\<i>*t) ^ k * (expectation (\<lambda>x. x ^ k)) / fact k" for k
unfolding c_def integral_mult_right_zero integral_complex_of_real by simp
then have "norm (char M t - ?t1) = norm (char M t - (CLINT x | M. (\<Sum>k \<le> n. c k x)))"
by (simp add: integ_c)
also have "\<dots> = norm ((CLINT x | M. iexp (t * x) - (\<Sum>k \<le> n. c k x)))"
unfolding char_def by (subst Bochner_Integration.integral_diff[OF integ_iexp]) (auto intro!: integ_c)
also have "\<dots> \<le> expectation (\<lambda>x. cmod (iexp (t * x) - (\<Sum>k \<le> n. c k x)))"
by (intro integral_norm_bound)
also have "\<dots> \<le> expectation (\<lambda>x. 2 * \<bar>t\<bar> ^ n / fact n * \<bar>x\<bar> ^ n)"
proof (rule integral_mono)
show "integrable M (\<lambda>x. cmod (iexp (t * x) - (\<Sum>k\<le>n. c k x)))"
by (intro integrable_norm Bochner_Integration.integrable_diff integ_iexp Bochner_Integration.integrable_sum integ_c) simp
show "integrable M (\<lambda>x. 2 * \<bar>t\<bar> ^ n / fact n * \<bar>x\<bar> ^ n)"
unfolding power_abs[symmetric]
by (intro integrable_mult_right integrable_abs integrable_moments) simp
show "cmod (iexp (t * x) - (\<Sum>k\<le>n. c k x)) \<le> 2 * \<bar>t\<bar> ^ n / fact n * \<bar>x\<bar> ^ n" for x
using iexp_approx2[of "t * x" n] by (auto simp add: abs_mult field_simps c_def)
qed
finally show ?thesis
unfolding integral_mult_right_zero .
qed
lemma (in real_distribution) char_approx2:
assumes integrable_moments: "\<And>k. k \<le> n \<Longrightarrow> integrable M (\<lambda>x. x ^ k)"
shows "cmod (char M t - (\<Sum>k \<le> n. ((\<i> * t)^k / fact k) * expectation (\<lambda>x. x^k))) \<le>
(\<bar>t\<bar>^n / fact (Suc n)) * expectation (\<lambda>x. min (2 * \<bar>x\<bar>^n * Suc n) (\<bar>t\<bar> * \<bar>x\<bar>^Suc n))"
(is "cmod (char M t - ?t1) \<le> _")
proof -
have integ_iexp: "integrable M (\<lambda>x. iexp (t * x))"
by (intro integrable_const_bound) auto
define c where [abs_def]: "c k x = (\<i> * t)^k / fact k * complex_of_real (x^k)" for k x
have integ_c: "\<And>k. k \<le> n \<Longrightarrow> integrable M (\<lambda>x. c k x)"
unfolding c_def by (intro integrable_mult_right integrable_of_real integrable_moments)
have *: "min (2 * \<bar>t * x\<bar>^n / fact n) (\<bar>t * x\<bar>^Suc n / fact (Suc n)) =
\<bar>t\<bar>^n / fact (Suc n) * min (2 * \<bar>x\<bar>^n * real (Suc n)) (\<bar>t\<bar> * \<bar>x\<bar>^(Suc n))" for x
apply (subst mult_min_right)
apply simp
apply (rule arg_cong2[where f=min])
apply (simp_all add: field_simps abs_mult del: fact_Suc)
apply (simp_all add: field_simps)
done
have "k \<le> n \<Longrightarrow> expectation (c k) = (\<i>*t) ^ k * (expectation (\<lambda>x. x ^ k)) / fact k" for k
unfolding c_def integral_mult_right_zero integral_complex_of_real by simp
then have "norm (char M t - ?t1) = norm (char M t - (CLINT x | M. (\<Sum>k \<le> n. c k x)))"
by (simp add: integ_c)
also have "\<dots> = norm ((CLINT x | M. iexp (t * x) - (\<Sum>k \<le> n. c k x)))"
unfolding char_def by (subst Bochner_Integration.integral_diff[OF integ_iexp]) (auto intro!: integ_c)
also have "\<dots> \<le> expectation (\<lambda>x. cmod (iexp (t * x) - (\<Sum>k \<le> n. c k x)))"
by (rule integral_norm_bound)
also have "\<dots> \<le> expectation (\<lambda>x. min (2 * \<bar>t * x\<bar>^n / fact n) (\<bar>t * x\<bar>^(Suc n) / fact (Suc n)))"
(is "_ \<le> expectation ?f")
proof (rule integral_mono)
show "integrable M (\<lambda>x. cmod (iexp (t * x) - (\<Sum>k\<le>n. c k x)))"
by (intro integrable_norm Bochner_Integration.integrable_diff integ_iexp Bochner_Integration.integrable_sum integ_c) simp
show "integrable M ?f"
by (rule Bochner_Integration.integrable_bound[where f="\<lambda>x. 2 * \<bar>t * x\<bar> ^ n / fact n"])
(auto simp: integrable_moments power_abs[symmetric] power_mult_distrib)
show "cmod (iexp (t * x) - (\<Sum>k\<le>n. c k x)) \<le> ?f x" for x
using iexp_approx1[of "t * x" n] iexp_approx2[of "t * x" n]
by (auto simp add: abs_mult field_simps c_def intro!: min.boundedI)
qed
also have "\<dots> = (\<bar>t\<bar>^n / fact (Suc n)) * expectation (\<lambda>x. min (2 * \<bar>x\<bar>^n * Suc n) (\<bar>t\<bar> * \<bar>x\<bar>^Suc n))"
unfolding *
proof (rule Bochner_Integration.integral_mult_right)
show "integrable M (\<lambda>x. min (2 * \<bar>x\<bar> ^ n * real (Suc n)) (\<bar>t\<bar> * \<bar>x\<bar> ^ Suc n))"
by (rule Bochner_Integration.integrable_bound[where f="\<lambda>x. 2 * \<bar>x\<bar> ^ n * real (Suc n)"])
(auto simp: integrable_moments power_abs[symmetric] power_mult_distrib)
qed
finally show ?thesis
unfolding integral_mult_right_zero .
qed
lemma (in real_distribution) char_approx3:
fixes t
assumes
integrable_1: "integrable M (\<lambda>x. x)" and
integral_1: "expectation (\<lambda>x. x) = 0" and
integrable_2: "integrable M (\<lambda>x. x^2)" and
integral_2: "variance (\<lambda>x. x) = \<sigma>2"
shows "cmod (char M t - (1 - t^2 * \<sigma>2 / 2)) \<le>
(t^2 / 6) * expectation (\<lambda>x. min (6 * x^2) (abs t * (abs x)^3) )"
proof -
note real_distribution.char_approx2 [of M 2 t, simplified]
have [simp]: "prob UNIV = 1" by (metis prob_space space_eq_univ)
from integral_2 have [simp]: "expectation (\<lambda>x. x * x) = \<sigma>2"
by (simp add: integral_1 numeral_eq_Suc)
have 1: "k \<le> 2 \<Longrightarrow> integrable M (\<lambda>x. x^k)" for k
using assms by (auto simp: eval_nat_numeral le_Suc_eq)
note char_approx1
note 2 = char_approx1 [of 2 t, OF 1, simplified]
have "cmod (char M t - (\<Sum>k\<le>2. (\<i> * t) ^ k * (expectation (\<lambda>x. x ^ k)) / (fact k))) \<le>
t\<^sup>2 * expectation (\<lambda>x. min (6 * x\<^sup>2) (\<bar>t\<bar> * \<bar>x\<bar> ^ 3)) / fact (3::nat)"
using char_approx2 [of 2 t, OF 1] by simp
also have "(\<Sum>k\<le>2. (\<i> * t) ^ k * expectation (\<lambda>x. x ^ k) / (fact k)) = 1 - t^2 * \<sigma>2 / 2"
by (simp add: complex_eq_iff numeral_eq_Suc integral_1 Re_divide Im_divide)
also have "fact 3 = 6" by (simp add: eval_nat_numeral)
also have "t\<^sup>2 * expectation (\<lambda>x. min (6 * x\<^sup>2) (\<bar>t\<bar> * \<bar>x\<bar> ^ 3)) / 6 =
t\<^sup>2 / 6 * expectation (\<lambda>x. min (6 * x\<^sup>2) (\<bar>t\<bar> * \<bar>x\<bar> ^ 3))" by (simp add: field_simps)
finally show ?thesis .
qed
text \<open>
This is a more familiar textbook formulation in terms of random variables, but
we will use the previous version for the CLT.
\<close>
lemma (in prob_space) char_approx3':
fixes \<mu> :: "real measure" and X
assumes rv_X [simp]: "random_variable borel X"
and [simp]: "integrable M X" "integrable M (\<lambda>x. (X x)^2)" "expectation X = 0"
and var_X: "variance X = \<sigma>2"
and \<mu>_def: "\<mu> = distr M borel X"
shows "cmod (char \<mu> t - (1 - t^2 * \<sigma>2 / 2)) \<le>
(t^2 / 6) * expectation (\<lambda>x. min (6 * (X x)^2) (\<bar>t\<bar> * \<bar>X x\<bar>^3))"
using var_X unfolding \<mu>_def
apply (subst integral_distr [symmetric, OF rv_X], simp)
apply (intro real_distribution.char_approx3)
apply (auto simp add: integrable_distr_eq integral_distr)
done
text \<open>
this is the formulation in the book -- in terms of a random variable *with* the distribution,
rather the distribution itself. I don't know which is more useful, though in principal we can
go back and forth between them.
\<close>
lemma (in prob_space) char_approx1':
fixes \<mu> :: "real measure" and X
assumes integrable_moments : "\<And>k. k \<le> n \<Longrightarrow> integrable M (\<lambda>x. X x ^ k)"
and rv_X[measurable]: "random_variable borel X"
and \<mu>_distr : "distr M borel X = \<mu>"
shows "cmod (char \<mu> t - (\<Sum>k \<le> n. ((\<i> * t)^k / fact k) * expectation (\<lambda>x. (X x)^k))) \<le>
(2 * \<bar>t\<bar>^n / fact n) * expectation (\<lambda>x. \<bar>X x\<bar>^n)"
unfolding \<mu>_distr[symmetric]
apply (subst (1 2) integral_distr [symmetric, OF rv_X], simp, simp)
apply (intro real_distribution.char_approx1[of "distr M borel X" n t] real_distribution_distr rv_X)
apply (auto simp: integrable_distr_eq integrable_moments)
done
subsection \<open>Calculation of the Characteristic Function of the Standard Distribution\<close>
abbreviation
"std_normal_distribution \<equiv> density lborel std_normal_density"
(* TODO: should this be an instance statement? *)
lemma real_dist_normal_dist: "real_distribution std_normal_distribution"
using prob_space_normal_density by (auto simp: real_distribution_def real_distribution_axioms_def)
lemma std_normal_distribution_even_moments:
fixes k :: nat
shows "(LINT x|std_normal_distribution. x^(2 * k)) = fact (2 * k) / (2^k * fact k)"
and "integrable std_normal_distribution (\<lambda>x. x^(2 * k))"
using integral_std_normal_moment_even[of k]
by (subst integral_density)
(auto simp: normal_density_nonneg integrable_density
intro: integrable.intros std_normal_moment_even)
lemma integrable_std_normal_distribution_moment: "integrable std_normal_distribution (\<lambda>x. x^k)"
by (auto simp: normal_density_nonneg integrable_std_normal_moment integrable_density)
lemma integral_std_normal_distribution_moment_odd:
"odd k \<Longrightarrow> integral\<^sup>L std_normal_distribution (\<lambda>x. x^k) = 0"
using integral_std_normal_moment_odd[of "(k - 1) div 2"]
by (auto simp: integral_density normal_density_nonneg elim: oddE)
lemma std_normal_distribution_even_moments_abs:
fixes k :: nat
shows "(LINT x|std_normal_distribution. \<bar>x\<bar>^(2 * k)) = fact (2 * k) / (2^k * fact k)"
using integral_std_normal_moment_even[of k]
by (subst integral_density) (auto simp: normal_density_nonneg power_even_abs)
lemma std_normal_distribution_odd_moments_abs:
fixes k :: nat
shows "(LINT x|std_normal_distribution. \<bar>x\<bar>^(2 * k + 1)) = sqrt (2 / pi) * 2 ^ k * fact k"
using integral_std_normal_moment_abs_odd[of k]
by (subst integral_density) (auto simp: normal_density_nonneg)
theorem char_std_normal_distribution:
"char std_normal_distribution = (\<lambda>t. complex_of_real (exp (- (t^2) / 2)))"
proof (intro ext LIMSEQ_unique)
fix t :: real
let ?f' = "\<lambda>k. (\<i> * t)^k / fact k * (LINT x | std_normal_distribution. x^k)"
let ?f = "\<lambda>n. (\<Sum>k \<le> n. ?f' k)"
show "?f \<longlonglongrightarrow> exp (-(t^2) / 2)"
proof (rule limseq_even_odd)
have "(\<i> * complex_of_real t) ^ (2 * a) / (2 ^ a * fact a) = (- ((complex_of_real t)\<^sup>2 / 2)) ^ a / fact a" for a
by (subst power_mult) (simp add: field_simps uminus_power_if power_mult)
then have *: "?f (2 * n) = complex_of_real (\<Sum>k < Suc n. (1 / fact k) * (- (t^2) / 2)^k)" for n :: nat
unfolding of_real_sum
by (intro sum.reindex_bij_witness_not_neutral[symmetric, where
i="\<lambda>n. n div 2" and j="\<lambda>n. 2 * n" and T'="{i. i \<le> 2 * n \<and> odd i}" and S'="{}"])
(auto simp: integral_std_normal_distribution_moment_odd std_normal_distribution_even_moments)
show "(\<lambda>n. ?f (2 * n)) \<longlonglongrightarrow> exp (-(t^2) / 2)"
unfolding * using exp_converges[where 'a=real]
by (intro tendsto_of_real LIMSEQ_Suc) (auto simp: inverse_eq_divide sums_def [symmetric])
have **: "?f (2 * n + 1) = ?f (2 * n)" for n
proof -
have "?f (2 * n + 1) = ?f (2 * n) + ?f' (2 * n + 1)"
by simp
also have "?f' (2 * n + 1) = 0"
by (subst integral_std_normal_distribution_moment_odd) simp_all
finally show "?f (2 * n + 1) = ?f (2 * n)"
by simp
qed
show "(\<lambda>n. ?f (2 * n + 1)) \<longlonglongrightarrow> exp (-(t^2) / 2)"
unfolding ** by fact
qed
have **: "(\<lambda>n. x ^ n / fact n) \<longlonglongrightarrow> 0" for x :: real
using summable_LIMSEQ_zero [OF summable_exp] by (auto simp add: inverse_eq_divide)
let ?F = "\<lambda>n. 2 * \<bar>t\<bar> ^ n / fact n * (LINT x|std_normal_distribution. \<bar>x\<bar> ^ n)"
show "?f \<longlonglongrightarrow> char std_normal_distribution t"
proof (rule metric_tendsto_imp_tendsto[OF limseq_even_odd])
show "(\<lambda>n. ?F (2 * n)) \<longlonglongrightarrow> 0"
proof (rule Lim_transform_eventually)
show "\<forall>\<^sub>F n in sequentially. 2 * ((t^2 / 2)^n / fact n) = ?F (2 * n)"
unfolding std_normal_distribution_even_moments_abs by (simp add: power_mult power_divide)
qed (intro tendsto_mult_right_zero **)
have *: "?F (2 * n + 1) = (2 * \<bar>t\<bar> * sqrt (2 / pi)) * ((2 * t^2)^n * fact n / fact (2 * n + 1))" for n
unfolding std_normal_distribution_odd_moments_abs
by (simp add: field_simps power_mult[symmetric] power_even_abs)
have "norm ((2 * t\<^sup>2) ^ n * fact n / fact (2 * n + 1)) \<le> (2 * t\<^sup>2) ^ n / fact n" for n
using mult_mono[OF _ square_fact_le_2_fact, of 1 "1 + 2 * real n" n]
by (auto simp add: divide_simps intro!: mult_left_mono)
then show "(\<lambda>n. ?F (2 * n + 1)) \<longlonglongrightarrow> 0"
unfolding * by (intro tendsto_mult_right_zero Lim_null_comparison [OF _ ** [of "2 * t\<^sup>2"]]) auto
show "\<forall>\<^sub>F n in sequentially. dist (?f n) (char std_normal_distribution t) \<le> dist (?F n) 0"
using real_distribution.char_approx1[OF real_dist_normal_dist integrable_std_normal_distribution_moment]
by (auto simp: dist_norm integral_nonneg_AE norm_minus_commute)
qed
qed
end