Theory Expander_Graphs_Algebra
section ‹Algebra-only Theorems›
text ‹This section verifies the linear algebraic counter-parts of the graph-theoretic theorems
about Random walks. The graph-theoretic results are then derived in Section~\ref{sec:random_walks}.›
theory Expander_Graphs_Algebra
imports
"HOL-Library.Monad_Syntax"
Expander_Graphs_TTS
begin
lemma pythagoras:
fixes v w :: "'a::real_inner"
assumes "v ∙ w = 0"
shows "norm (v+w)^2 = norm v^2 + norm w^2"
using assms by (simp add:power2_norm_eq_inner algebra_simps inner_commute)
definition diag :: "('a :: zero)^'n ⇒ 'a^'n^'n"
where "diag v = (χ i j. if i = j then (v $ i) else 0)"
definition ind_vec :: "'n set ⇒ real^'n"
where "ind_vec S = (χ i. of_bool( i ∈ S))"
lemma diag_mult_eq: "diag x ** diag y = diag (x * y)"
unfolding diag_def
by (vector matrix_matrix_mult_def)
(auto simp add:if_distrib if_distribR sum.If_cases)
lemma diag_vec_mult_eq: "diag x *v y = x * y"
unfolding diag_def matrix_vector_mult_def
by (simp add:if_distrib if_distribR sum.If_cases times_vec_def)
definition matrix_norm_bound :: "real^'n^'m ⇒ real ⇒ bool"
where "matrix_norm_bound A l = (∀x. norm (A *v x) ≤ l * norm x)"
lemma matrix_norm_boundI:
assumes "⋀x. norm (A *v x) ≤ l * norm x"
shows "matrix_norm_bound A l"
using assms unfolding matrix_norm_bound_def by simp
lemma matrix_norm_boundD:
assumes "matrix_norm_bound A l"
shows "norm (A *v x) ≤ l * norm x"
using assms unfolding matrix_norm_bound_def by simp
lemma matrix_norm_bound_nonneg:
fixes A :: "real^'n^'m"
assumes "matrix_norm_bound A l"
shows "l ≥ 0"
proof -
have "0 ≤ norm (A *v 1)" by simp
also have "... ≤ l * norm (1::real^'n)"
using assms(1) unfolding matrix_norm_bound_def by simp
finally have "0 ≤ l * norm (1::real^'n)"
by simp
moreover have "norm (1::real^'n) > 0"
by simp
ultimately show ?thesis
by (simp add: zero_le_mult_iff)
qed
lemma matrix_norm_bound_0:
assumes "matrix_norm_bound A 0"
shows "A = (0::real^'n^'m)"
proof (intro iffD2[OF matrix_eq] allI)
fix x :: "real^'n"
have "norm (A *v x) = 0"
using assms unfolding matrix_norm_bound_def by simp
thus "A *v x = 0 *v x"
by simp
qed
lemma matrix_norm_bound_diag:
fixes x :: "real^'n"
assumes "⋀i. ¦x $ i¦ ≤ l"
shows "matrix_norm_bound (diag x) l"
proof (rule matrix_norm_boundI)
fix y :: "real^'n"
have l_ge_0: "l ≥ 0" using assms by fastforce
have a: "¦x $ i * v¦ ≤ ¦l * v¦" for v i
using l_ge_0 assms by (simp add:abs_mult mult_right_mono)
have "norm (diag x *v y) = sqrt (∑i ∈ UNIV. (x $ i * y $ i)^2)"
unfolding matrix_vector_mult_def diag_def norm_vec_def L2_set_def
by (auto simp add:if_distrib if_distribR sum.If_cases)
also have "... ≤ sqrt (∑i ∈ UNIV. (l * y $ i)^2)"
by (intro real_sqrt_le_mono sum_mono iffD1[OF abs_le_square_iff] a)
also have "... = l * norm y"
using l_ge_0 by (simp add:norm_vec_def L2_set_def algebra_simps
sum_distrib_left[symmetric] real_sqrt_mult)
finally show "norm (diag x *v y) ≤ l * norm y" by simp
qed
lemma vector_scaleR_matrix_ac_2: "b *⇩R (A::real^'n^'m) *v x = b *⇩R (A *v x)"
unfolding vector_transpose_matrix[symmetric] transpose_scalar
by (intro vector_scaleR_matrix_ac)
lemma matrix_norm_bound_scale:
assumes "matrix_norm_bound A l"
shows "matrix_norm_bound (b *⇩R A) (¦b¦ * l)"
proof (intro matrix_norm_boundI)
fix x
have "norm (b *⇩R A *v x) = norm (b *⇩R (A *v x))"
by (metis transpose_scalar vector_scaleR_matrix_ac vector_transpose_matrix)
also have "... = ¦b¦ * norm (A *v x)"
by simp
also have "... ≤ ¦b¦ * (l * norm x)"
using assms matrix_norm_bound_def by (intro mult_left_mono) auto
also have "... ≤ (¦b¦ * l) * norm x" by simp
finally show "norm (b *⇩R A *v x) ≤ (¦b¦ * l) * norm x" by simp
qed
definition nonneg_mat :: "real^'n^'m ⇒ bool"
where "nonneg_mat A = (∀i j. A $ i $ j ≥ 0)"
lemma nonneg_mat_1:
shows "nonneg_mat (mat 1)"
unfolding nonneg_mat_def mat_def by auto
lemma nonneg_mat_prod:
assumes "nonneg_mat A" "nonneg_mat B"
shows "nonneg_mat (A ** B)"
using assms unfolding nonneg_mat_def matrix_matrix_mult_def
by (auto intro:sum_nonneg)
lemma nonneg_mat_transpose:
"nonneg_mat (transpose A) = nonneg_mat A"
unfolding nonneg_mat_def transpose_def
by auto
definition spec_bound :: "real^'n^'n ⇒ real ⇒ bool"
where "spec_bound M l = (l ≥ 0 ∧ (∀v. v ∙ 1 = 0 ⟶ norm (M *v v) ≤ l * norm v))"
lemma spec_boundD1:
assumes "spec_bound M l"
shows "0 ≤ l"
using assms unfolding spec_bound_def by simp
lemma spec_boundD2:
assumes "spec_bound M l"
assumes "v ∙ 1 = 0 "
shows "norm (M *v v) ≤ l * norm v"
using assms unfolding spec_bound_def by simp
lemma spec_bound_mono:
assumes "spec_bound M α" "α ≤ β"
shows "spec_bound M β"
proof -
have "norm (M *v v) ≤ β * norm v" if "inner v 1 = 0" for v
proof -
have "norm (M *v v) ≤ α * norm v"
by (intro spec_boundD2[OF assms(1)] that)
also have "... ≤ β * norm v"
by (intro mult_right_mono assms(2)) auto
finally show ?thesis by simp
qed
moreover have "β ≥ 0"
using assms(2) spec_boundD1[OF assms(1)] by simp
ultimately show ?thesis
unfolding spec_bound_def by simp
qed
definition markov :: "real^'n^'n ⇒ bool"
where "markov M = (nonneg_mat M ∧ M *v 1 = 1 ∧ 1 v* M = 1)"
lemma markov_symI:
assumes "nonneg_mat A" "transpose A = A" "A *v 1 = 1"
shows "markov A"
proof -
have "1 v* A = transpose A *v 1"
unfolding vector_transpose_matrix[symmetric] by simp
also have "... = 1" unfolding assms(2,3) by simp
finally have "1 v* A = 1" by simp
thus ?thesis
unfolding markov_def using assms by auto
qed
lemma markov_apply:
assumes "markov M"
shows "M *v 1 = 1" "1 v* M = 1"
using assms unfolding markov_def by auto
lemma markov_transpose:
"markov A = markov (transpose A)"
unfolding markov_def nonneg_mat_transpose by auto
fun matrix_pow where
"matrix_pow M 0 = mat 1" |
"matrix_pow M (Suc n) = M ** (matrix_pow M n)"
lemma markov_orth_inv:
assumes "markov A"
shows "inner (A *v x) 1 = inner x 1"
proof -
have "inner (A *v x) 1 = inner x (1 v* A)"
using dot_lmul_matrix inner_commute by metis
also have "... = inner x 1"
using markov_apply[OF assms(1)] by simp
finally show ?thesis by simp
qed
lemma markov_id:
"markov (mat 1)"
unfolding markov_def using nonneg_mat_1 by simp
lemma markov_mult:
assumes "markov A" "markov B"
shows "markov (A ** B)"
proof -
have "nonneg_mat (A ** B)"
using assms unfolding markov_def by (intro nonneg_mat_prod) auto
moreover have "(A ** B) *v 1 = 1"
using assms unfolding markov_def
unfolding matrix_vector_mul_assoc[symmetric] by simp
moreover have "1 v* (A ** B) = 1"
using assms unfolding markov_def
unfolding vector_matrix_mul_assoc[symmetric] by simp
ultimately show ?thesis
unfolding markov_def by simp
qed
lemma markov_matrix_pow:
assumes "markov A"
shows "markov (matrix_pow A k)"
using markov_id assms markov_mult
by (induction k, auto)
lemma spec_bound_prod:
assumes "markov A" "markov B"
assumes "spec_bound A la" "spec_bound B lb"
shows "spec_bound (A ** B) (la*lb)"
proof -
have la_ge_0: "la ≥ 0" using spec_boundD1[OF assms(3)] by simp
have "norm ((A ** B) *v x) ≤ (la * lb) * norm x" if "inner x 1 = 0" for x
proof -
have "norm ((A ** B) *v x) = norm (A *v (B *v x))"
by (simp add:matrix_vector_mul_assoc)
also have "... ≤ la * norm (B *v x)"
by (intro spec_boundD2[OF assms(3)]) (simp add:markov_orth_inv that assms(2))
also have "... ≤ la * (lb * norm x)"
by (intro spec_boundD2[OF assms(4)] mult_left_mono that la_ge_0)
finally show ?thesis by simp
qed
moreover have "la * lb ≥ 0"
using la_ge_0 spec_boundD1[OF assms(4)] by simp
ultimately show ?thesis
using spec_bound_def by auto
qed
lemma spec_bound_pow:
assumes "markov A"
assumes "spec_bound A l"
shows "spec_bound (matrix_pow A k) (l^k)"
proof (induction k)
case 0
then show ?case unfolding spec_bound_def by simp
next
case (Suc k)
have "spec_bound (A ** matrix_pow A k) (l * l ^ k)"
by (intro spec_bound_prod assms Suc markov_matrix_pow)
thus ?case by simp
qed
fun intersperse :: "'a ⇒ 'a list ⇒ 'a list"
where
"intersperse x [] = []" |
"intersperse x (y#[]) = y#[]" |
"intersperse x (y#z#zs) = y#x#intersperse x (z#zs)"
lemma intersperse_snoc:
assumes "xs ≠ []"
shows "intersperse z (xs@[y]) = intersperse z xs@[z,y]"
using assms
proof (induction xs rule:list_nonempty_induct)
case (single x)
then show ?case by simp
next
case (cons x xs)
then obtain xsh xst where t:"xs = xsh#xst"
by (metis neq_Nil_conv)
have "intersperse z ((x # xs) @ [y]) = x#z#intersperse z (xs@[y])"
unfolding t by simp
also have "... = x#z#intersperse z xs@[z,y]"
using cons by simp
also have "... = intersperse z (x#xs)@[z,y]"
unfolding t by simp
finally show ?case by simp
qed
lemma foldl_intersperse:
assumes "xs ≠ []"
shows "foldl f a ((intersperse x xs)@[x]) = foldl (λy z. f (f y z) x) a xs"
using assms by (induction xs rule:rev_nonempty_induct) (auto simp add:intersperse_snoc)
lemma foldl_intersperse_2:
shows "foldl f a (intersperse y (x#xs)) = foldl (λx z. f (f x y) z) (f a x) xs"
proof (induction xs rule:rev_induct)
case Nil
then show ?case by simp
next
case (snoc xst xs)
have "foldl f a (intersperse y ((x # xs) @ [xst])) = foldl (λx. f (f x y)) (f a x) (xs @ [xst])"
by (subst intersperse_snoc, auto simp add:snoc)
then show ?case by simp
qed
context regular_graph_tts
begin
definition stat :: "real^'n"
where "stat = (1 / real CARD('n)) *⇩R 1"
definition J :: "('c :: field)^'n^'n"
where "J = (χ i j. of_nat 1 / of_nat CARD('n))"
lemma inner_1_1: "1 ∙ (1::real^'n) = CARD('n)"
unfolding inner_vec_def by simp
definition proj_unit :: "real^'n ⇒ real^'n"
where "proj_unit v = (1 ∙ v) *⇩R stat"
definition proj_rem :: "real^'n ⇒ real^'n"
where "proj_rem v = v - proj_unit v"
lemma proj_rem_orth: "1 ∙ (proj_rem v) = 0"
unfolding proj_rem_def proj_unit_def inner_diff_right stat_def
by (simp add:inner_1_1)
lemma split_vec: "v = proj_unit v + proj_rem v"
unfolding proj_rem_def by simp
lemma apply_J: "J *v x = proj_unit x"
proof (intro iffD2[OF vec_eq_iff] allI)
fix i
have "(J *v x) $ i = inner (χ j. 1 / real CARD('n)) x"
unfolding matrix_vector_mul_component J_def by simp
also have "... = inner stat x"
unfolding stat_def scaleR_vec_def by auto
also have "... = (proj_unit x) $ i"
unfolding proj_unit_def stat_def by simp
finally show "(J *v x) $ i = (proj_unit x) $ i" by simp
qed
lemma spec_bound_J: "spec_bound (J :: real^'n^'n) 0"
proof -
have "norm (J *v v) = 0" if "inner v 1 = 0" for v :: "real^'n"
proof -
have "inner (proj_unit v + proj_rem v) 1 = 0"
using that by (subst (asm) split_vec[of "v"], simp)
hence "inner (proj_unit v) 1 = 0"
using proj_rem_orth inner_commute unfolding inner_add_left
by (metis add_cancel_left_right)
hence "proj_unit v = 0"
unfolding proj_unit_def stat_def by simp
hence "J *v v = 0"
unfolding apply_J by simp
thus ?thesis by simp
qed
thus ?thesis
unfolding spec_bound_def by simp
qed
lemma matrix_decomposition_lemma_aux:
fixes A :: "real^'n^'n"
assumes "markov A"
shows "spec_bound A l ⟷ matrix_norm_bound (A - (1-l) *⇩R J) l" (is "?L ⟷ ?R")
proof
assume a:"?L"
hence l_ge_0: "l ≥ 0" using spec_boundD1 by auto
show "?R"
proof (rule matrix_norm_boundI)
fix x :: "real^'n"
have "(A - (1-l) *⇩R J) *v x = A *v x - (1-l) *⇩R (proj_unit x)"
by (simp add:algebra_simps vector_scaleR_matrix_ac_2 apply_J)
also have "... = A *v proj_unit x + A *v proj_rem x - (1-l) *⇩R (proj_unit x)"
by (subst split_vec[of "x"], simp add:algebra_simps)
also have "... = proj_unit x + A *v proj_rem x - (1-l) *⇩R (proj_unit x)"
using markov_apply[OF assms(1)]
unfolding proj_unit_def stat_def by (simp add:algebra_simps)
also have "... = A *v proj_rem x + l *⇩R proj_unit x" (is "_ = ?R1")
by (simp add:algebra_simps)
finally have d:"(A - (1-l) *⇩R J) *v x = ?R1" by simp
have "inner (l *⇩R proj_unit x) (A *v proj_rem x) =
inner ((l * inner 1 x / real CARD('n)) *⇩R 1 v* A) (proj_rem x)"
by (subst dot_lmul_matrix[symmetric]) (simp add:proj_unit_def stat_def)
also have "... = (l * inner 1 x / real CARD('n)) * inner 1 (proj_rem x)"
unfolding scaleR_vector_matrix_assoc markov_apply[OF assms] by simp
also have "... = 0"
unfolding proj_rem_orth by simp
finally have b:"inner (l *⇩R proj_unit x) (A *v proj_rem x) = 0" by simp
have c: "inner (proj_rem x) (proj_unit x) = 0"
using proj_rem_orth[of "x"]
unfolding proj_unit_def stat_def by (simp add:inner_commute)
have "norm (?R1)^2 = norm (A *v proj_rem x)^2 + norm (l *⇩R proj_unit x)^2"
using b by (intro pythagoras) (simp add:inner_commute)
also have "... ≤ (l * norm (proj_rem x))^2 + norm (l *⇩R proj_unit x)^2"
using proj_rem_orth[of "x"]
by (intro add_mono power_mono spec_boundD2 a) (auto simp add:inner_commute)
also have "... = l^2 * (norm (proj_rem x)^2 + norm (proj_unit x)^2)"
by (simp add:algebra_simps)
also have "... = l^2 * (norm (proj_rem x + proj_unit x)^2)"
using c by (subst pythagoras) auto
also have "... = l^2 * norm x^2"
by (subst (3) split_vec[of "x"]) (simp add:algebra_simps)
also have "... = (l * norm x)^2"
by (simp add:algebra_simps)
finally have "norm (?R1)^2 ≤ (l * norm x)^2" by simp
hence "norm (?R1) ≤ l * norm x"
using l_ge_0 by (subst (asm) power_mono_iff) auto
thus "norm ((A - (1-l) *⇩R J) *v x) ≤ l * norm x"
unfolding d by simp
qed
next
assume a:"?R"
have "norm (A *v x) ≤ l * norm x" if "inner x 1 = 0" for x
proof -
have "(1 - l) *⇩R J *v x = (1 - l) *⇩R (proj_unit x)"
by (simp add:vector_scaleR_matrix_ac_2 apply_J)
also have "... = 0"
unfolding proj_unit_def using that by (simp add:inner_commute)
finally have b: "(1 - l) *⇩R J *v x = 0" by simp
have "norm (A *v x) = norm ((A - (1-l) *⇩R J) *v x + ((1-l) *⇩R J) *v x)"
by (simp add:algebra_simps)
also have "... ≤ norm ((A - (1-l) *⇩R J) *v x) + norm (((1-l) *⇩R J) *v x)"
by (intro norm_triangle_ineq)
also have "... ≤ l * norm x + 0"
using a b unfolding matrix_norm_bound_def by (intro add_mono, auto)
also have "... = l * norm x"
by simp
finally show ?thesis by simp
qed
moreover have "l ≥ 0"
using a matrix_norm_bound_nonneg by blast
ultimately show "?L"
unfolding spec_bound_def by simp
qed
lemma matrix_decomposition_lemma:
fixes A :: "real^'n^'n"
assumes "markov A"
shows "spec_bound A l ⟷ (∃E. A = (1-l) *⇩R J + l *⇩R E ∧ matrix_norm_bound E 1 ∧ l ≥ 0)"
(is "?L ⟷ ?R")
proof -
have "?L ⟷ matrix_norm_bound (A - (1-l) *⇩R J) l"
using matrix_decomposition_lemma_aux[OF assms] by simp
also have "... ⟷ ?R"
proof
assume a:"matrix_norm_bound (A - (1 - l) *⇩R J) l"
hence l_ge_0: "l ≥ 0" using matrix_norm_bound_nonneg by auto
define E where "E = (1/l) *⇩R (A - (1-l) *⇩R J)"
have "A = J" if "l = 0"
proof -
have "matrix_norm_bound (A - J) 0"
using a that by simp
hence "A - J = 0" using matrix_norm_bound_0 by blast
thus "A = J" by simp
qed
hence "A = (1-l) *⇩R J + l *⇩R E"
unfolding E_def by simp
moreover have "matrix_norm_bound E 1"
proof (cases "l = 0")
case True
hence "E = 0" if "l = 0"
unfolding E_def by simp
thus "matrix_norm_bound E 1" if "l = 0"
using that unfolding matrix_norm_bound_def by auto
next
case False
hence "l > 0" using l_ge_0 by simp
moreover have "matrix_norm_bound E (¦1 / l¦* l)"
unfolding E_def
by (intro matrix_norm_bound_scale a)
ultimately show ?thesis by auto
qed
ultimately show ?R using l_ge_0 by auto
next
assume a:?R
then obtain E where E_def: "A = (1 - l) *⇩R J + l *⇩R E" "matrix_norm_bound E 1" "l ≥ 0"
by auto
have "matrix_norm_bound (l *⇩R E) (abs l*1)"
by (intro matrix_norm_bound_scale E_def(2))
moreover have "l ≥ 0" using E_def by simp
moreover have " l *⇩R E = (A - (1 - l) *⇩R J)"
using E_def(1) by simp
ultimately show "matrix_norm_bound (A - (1 - l) *⇩R J) l"
by simp
qed
finally show ?thesis by simp
qed
lemma hitting_property_alg:
fixes S :: "('n :: finite) set"
assumes l_range: "l ∈ {0..1}"
defines "P ≡ diag (ind_vec S)"
defines "μ ≡ card S / CARD('n)"
assumes "⋀M. M ∈ set Ms ⟹ spec_bound M l ∧ markov M"
shows "foldl (λx M. P *v (M *v x)) (P *v stat) Ms ∙ 1 ≤ (μ + l * (1-μ))^(length Ms+1)"
proof -
define t :: "real^'n" where "t = (χ i. of_bool (i ∈ S))"
define r where "r = foldl (λx M. P *v (M *v x)) (P *v stat) Ms"
have P_proj: "P ** P = P"
unfolding P_def diag_mult_eq ind_vec_def by (intro arg_cong[where f="diag"]) (vector)
have P_1_left: "1 v* P = t"
unfolding P_def diag_def ind_vec_def vector_matrix_mult_def t_def by simp
have P_1_right: "P *v 1 = t"
unfolding P_def diag_def ind_vec_def matrix_vector_mult_def t_def by simp
have P_norm :"matrix_norm_bound P 1"
unfolding P_def ind_vec_def by (intro matrix_norm_bound_diag) simp
have norm_t: "norm t = sqrt (real (card S))"
unfolding t_def norm_vec_def L2_set_def of_bool_def
by (simp add:sum.If_cases if_distrib if_distribR)
have μ_range: "μ ≥ 0" "μ ≤ 1"
unfolding μ_def by (auto simp add:card_mono)
define condition :: "real^'n ⇒ nat ⇒ bool"
where "condition = (λx n. norm x ≤ (μ + l * (1-μ))^n * sqrt (card S)/CARD('n) ∧ P *v x = x)"
have a:"condition r (length Ms)"
unfolding r_def using assms(4)
proof (induction Ms rule:rev_induct)
case Nil
have "norm (P *v stat) = (1 / real CARD('n)) * norm t"
unfolding stat_def matrix_vector_mult_scaleR P_1_right by simp
also have "... ≤ (1 / real CARD('n)) * sqrt (real (card S))"
using norm_t by (intro mult_left_mono) auto
also have "... = sqrt (card S)/CARD('n)" by simp
finally have "norm (P *v stat) ≤ sqrt (card S)/CARD('n)" by simp
moreover have "P *v (P *v stat) = P *v stat"
unfolding matrix_vector_mul_assoc P_proj by simp
ultimately show ?case unfolding condition_def by simp
next
case (snoc M xs)
hence "spec_bound M l ∧ markov M"
using snoc(2) by simp
then obtain E where E_def: "M = (1-l) *⇩R J + l *⇩R E" "matrix_norm_bound E 1"
using iffD1[OF matrix_decomposition_lemma] by auto
define y where "y = foldl (λx M. P *v (M *v x)) (P *v stat) xs"
have b:"condition y (length xs)"
using snoc unfolding y_def by simp
hence a:"P *v y = y" using condition_def by simp
have "norm (P *v (M *v y)) = norm (P *v ((1-l)*⇩R J *v y) + P *v (l *⇩R E *v y))"
by (simp add:E_def algebra_simps)
also have "... ≤ norm (P *v ((1-l)*⇩R J *v y)) + norm (P *v (l *⇩R E *v y)) "
by (intro norm_triangle_ineq)
also have "... = (1 - l) * norm (P *v (J *v y)) + l * norm (P *v (E *v y))"
using l_range
by (simp add:vector_scaleR_matrix_ac_2 matrix_vector_mult_scaleR)
also have "... = (1-l) * ¦1 ∙ (P *v y)/real CARD('n)¦ * norm t + l * norm (P *v (E *v y))"
by (subst a[symmetric])
(simp add:apply_J proj_unit_def stat_def P_1_right matrix_vector_mult_scaleR)
also have "... = (1-l) * ¦t ∙ y¦/real CARD('n) * norm t + l * norm (P *v (E *v y))"
by (subst dot_lmul_matrix[symmetric]) (simp add:P_1_left)
also have "... ≤ (1-l) * (norm t * norm y) / real CARD('n) * norm t + l * (1 * norm (E *v y))"
using P_norm Cauchy_Schwarz_ineq2 l_range
by (intro add_mono mult_right_mono mult_left_mono divide_right_mono matrix_norm_boundD) auto
also have "... = (1-l) * μ * norm y + l * norm (E *v y)"
unfolding μ_def norm_t by simp
also have "... ≤ (1-l) * μ * norm y + l * (1 * norm y)"
using μ_range l_range
by (intro add_mono matrix_norm_boundD mult_left_mono E_def) auto
also have "... = (μ + l * (1-μ)) * norm y"
by (simp add:algebra_simps)
also have "... ≤ (μ + l * (1-μ)) * ((μ + l * (1-μ))^length xs * sqrt (card S)/CARD('n))"
using b μ_range l_range unfolding condition_def
by (intro mult_left_mono) auto
also have "... = (μ + l * (1-μ))^(length xs +1) * sqrt (card S)/CARD('n)"
by simp
finally have "norm (P *v (M *v y)) ≤ (μ + l * (1-μ))^(length xs +1) * sqrt (card S)/CARD('n)"
by simp
moreover have "P *v (P *v (M *v y)) = P *v (M *v y)"
unfolding matrix_vector_mul_assoc matrix_mul_assoc P_proj
by simp
ultimately have "condition (P *v (M *v y)) (length (xs@[M]))"
unfolding condition_def by simp
then show ?case
unfolding y_def by simp
qed
have "inner r 1 = inner (P *v r) 1"
using a condition_def by simp
also have "... = inner (1 v* P) r"
unfolding dot_lmul_matrix by (simp add:inner_commute)
also have "... = inner t r"
unfolding P_1_left by simp
also have "... ≤ norm t * norm r"
by (intro norm_cauchy_schwarz)
also have "... ≤ sqrt (card S) * ((μ + l * (1-μ))^(length Ms) * sqrt(card S)/CARD('n))"
using a unfolding condition_def norm_t
by (intro mult_mono) auto
also have "... = (μ + 0) * ((μ + l * (1-μ))^(length Ms))"
by (simp add:μ_def)
also have "... ≤ (μ + l * (1-μ)) * (μ + l * (1-μ))^(length Ms)"
using μ_range l_range
by (intro mult_right_mono zero_le_power add_mono) auto
also have "... = (μ + l * (1-μ))^(length Ms+1)" by simp
finally show ?thesis
unfolding r_def by simp
qed
lemma upto_append:
assumes "i ≤ j" "j ≤ k"
shows "[i..<j]@[j..<k] = [i..<k]"
using assms by (metis less_eqE upt_add_eq_append)
definition bool_list_split :: "bool list ⇒ (nat list × nat)"
where "bool_list_split xs = foldl (λ(ys,z) x. (if x then (ys@[z],0) else (ys,z+1))) ([],0) xs"
lemma bool_list_split:
assumes "bool_list_split xs = (ys,z)"
shows "xs = concat (map (λk. replicate k False@[True]) ys)@replicate z False"
using assms
proof (induction xs arbitrary: ys z rule:rev_induct)
case Nil
then show ?case unfolding bool_list_split_def by simp
next
case (snoc x xs)
obtain u v where uv_def: "bool_list_split xs = (u,v)"
by (metis surj_pair)
show ?case
proof (cases x)
case True
have a:"ys = u@[v]" "z = 0"
using snoc(2) True uv_def unfolding bool_list_split_def by auto
have "xs@[x] = concat (map (λk. replicate k False@[True]) u)@replicate v False@[True]"
using snoc(1)[OF uv_def] True by simp
also have "... = concat (map (λk. replicate k False@[True]) (u@[v]))@replicate 0 False"
by simp
also have "... = concat (map (λk. replicate k False@[True]) (ys))@replicate z False"
using a by simp
finally show ?thesis by simp
next
case False
have a:"ys = u" "z = v+1"
using snoc(2) False uv_def unfolding bool_list_split_def by auto
have "xs@[x] = concat (map (λk. replicate k False@[True]) u)@replicate (v+1) False"
using snoc(1)[OF uv_def] False unfolding replicate_add by simp
also have "... = concat (map (λk. replicate k False@[True]) (ys))@replicate z False"
using a by simp
finally show ?thesis by simp
qed
qed
lemma bool_list_split_count:
assumes "bool_list_split xs = (ys,z)"
shows "length (filter id xs) = length ys"
unfolding bool_list_split[OF assms(1)] by (simp add:filter_concat comp_def)
lemma foldl_concat:
"foldl f a (concat xss) = foldl (λy xs. foldl f y xs) a xss"
by (induction xss rule:rev_induct, auto)
lemma hitting_property_alg_2:
fixes S :: "('n :: finite) set" and l :: nat
fixes M :: "real^'n^'n"
assumes α_range: "α ∈ {0..1}"
assumes "I ⊆ {..<l}"
defines "P i ≡ (if i ∈ I then diag (ind_vec S) else mat 1)"
defines "μ ≡ real (card S) / real (CARD('n))"
assumes "spec_bound M α" "markov M"
shows
"foldl (λx M. M *v x) stat (intersperse M (map P [0..<l])) ∙ 1 ≤ (μ+α*(1-μ))^card I"
(is "?L ≤ ?R")
proof (cases "I ≠ {}")
case True
define xs where "xs = map (λi. i ∈ I) [0..<l]"
define Q where "Q = diag (ind_vec S)"
define P' where "P' = (λx. if x then Q else mat 1)"
let ?rep = "(λx. replicate x (mat 1))"
have P_eq: "P i = P' (i ∈ I)" for i
unfolding P_def P'_def Q_def by simp
have "l > 0"
using True assms(2) by auto
hence xs_ne: "xs ≠ []"
unfolding xs_def by simp
obtain ys z where ys_z: "bool_list_split xs = (ys,z)"
by (metis surj_pair)
have "length ys = length (filter id xs)"
using bool_list_split_count[OF ys_z] by simp
also have "... = card (I ∩ {0..<l})"
unfolding xs_def filter_map by (simp add:comp_def distinct_length_filter)
also have "... = card I"
using Int_absorb2[OF assms(2)] unfolding atLeast0LessThan by simp
finally have len_ys: "length ys = card I" by simp
hence "length ys > 0"
using True assms(2) by (metis card_gt_0_iff finite_nat_iff_bounded)
then obtain yh yt where ys_split: "ys = yh#yt"
by (metis length_greater_0_conv neq_Nil_conv)
have a:"foldl (λx N. M *v (N *v x)) x (?rep z) ∙ 1 = x ∙ 1" for x
proof (induction z)
case 0
then show ?case by simp
next
case (Suc z)
have "foldl (λx N. M *v (N *v x)) x (?rep (z+1)) ∙ 1 = x ∙ 1"
unfolding replicate_add using Suc
by (simp add:markov_orth_inv[OF assms(6)])
then show ?case by simp
qed
have "M *v stat = stat"
using assms(6) unfolding stat_def matrix_vector_mult_scaleR markov_def by simp
hence b: "foldl (λx N. M *v (N *v x)) stat (?rep yh) = stat"
by (induction yh, auto)
have "foldl (λx N. N *v (M *v x)) a (?rep x) = matrix_pow M x *v a" for x a
proof (induction x)
case 0
then show ?case by simp
next
case (Suc x)
have "foldl (λx N. N *v (M *v x)) a (?rep (x+1)) = matrix_pow M (x+1) *v a"
unfolding replicate_add using Suc by (simp add: matrix_vector_mul_assoc)
then show ?case by simp
qed
hence c: "foldl (λx N. N *v (M *v x)) a (?rep x @ [Q]) = Q *v (matrix_pow M (x+1) *v a)" for x a
by (simp add:matrix_vector_mul_assoc matrix_mul_assoc)
have d: "spec_bound N α ∧ markov N" if t1: "N ∈ set (map (λx. matrix_pow M (x + 1)) yt)" for N
proof -
obtain y where N_def: "N = matrix_pow M (y+1)"
using t1 by auto
hence d1: "spec_bound N (α^(y+1))"
unfolding N_def using spec_bound_pow assms(5,6) by blast
have "spec_bound N (α^1)"
using α_range by (intro spec_bound_mono[OF d1] power_decreasing) auto
moreover have "markov N"
unfolding N_def by (intro markov_matrix_pow assms(6))
ultimately show ?thesis by simp
qed
have "?L = foldl (λx M. M *v x) stat (intersperse M (map P' xs)) ∙ 1"
unfolding P_eq xs_def map_map by (simp add:comp_def)
also have "... = foldl (λx M. M *v x) stat (intersperse M (map P' xs)@[M]) ∙ 1"
by (simp add:markov_orth_inv[OF assms(6)])
also have "... = foldl (λx N. M *v (N *v x)) stat (map P' xs) ∙ 1"
using xs_ne by (subst foldl_intersperse) auto
also have "... = foldl (λx N. M *v (N *v x)) stat ((ys ⤜ (λx. ?rep x @ [Q])) @ ?rep z) ∙ 1"
unfolding bool_list_split[OF ys_z] P'_def List.bind_def by (simp add: comp_def map_concat)
also have "... = foldl (λx N. M *v (N *v x)) stat (ys ⤜ (λx. ?rep x @ [Q])) ∙ 1"
by (simp add: a)
also have "... = foldl (λx N. M *v (N *v x)) stat (?rep yh @[Q]@(yt ⤜(λx. ?rep x @ [Q]))) ∙ 1"
unfolding ys_split by simp
also have "... = foldl (λx N. M *v (N *v x)) stat ([Q]@(yt ⤜(λx. ?rep x @ [Q]))) ∙ 1"
by (simp add:b)
also have "... = foldl (λx N. N *v x) stat (intersperse M (Q#(yt ⤜(λx.?rep x@[Q])))@[M])∙1"
by (subst foldl_intersperse, auto)
also have "... = foldl (λx N. N *v x) stat (intersperse M (Q#(yt ⤜(λx.?rep x@[Q])))) ∙ 1"
by (simp add:markov_orth_inv[OF assms(6)])
also have "... = foldl (λx N. N *v (M *v x)) (Q *v stat) (yt ⤜(λx.?rep x@[Q])) ∙ 1"
by (subst foldl_intersperse_2, simp)
also have "... = foldl (λa x. foldl (λx N. N *v (M *v x)) a (?rep x @ [Q])) (Q *v stat) yt ∙ 1"
unfolding List.bind_def foldl_concat foldl_map by simp
also have "... = foldl (λa x. Q *v (matrix_pow M (x+1) *v a)) (Q *v stat) yt ∙ 1"
unfolding c by simp
also have "... = foldl (λa N. Q *v (N *v a)) (Q *v stat) (map (λx. matrix_pow M (x+1)) yt) ∙ 1"
by (simp add:foldl_map)
also have "... ≤ (μ + α*(1-μ))^(length (map (λx. matrix_pow M (x+1)) yt)+1)"
unfolding μ_def Q_def by (intro hitting_property_alg α_range d) simp
also have "... = (μ + α*(1-μ))^(length ys)"
unfolding ys_split by simp
also have "... = ?R" unfolding len_ys by simp
finally show ?thesis by simp
next
case False
hence I_empty: "I = {}" by simp
have "?L = stat ∙ (1 :: real^'n)"
proof (cases "l > 0")
case True
have "?L = foldl (λx M. M *v x) stat ((intersperse M (map P [0..<l]))@[M]) ∙ 1"
by (simp add:markov_orth_inv[OF assms(6)])
also have "... = foldl (λx N. M *v (N *v x)) stat (map P [0..<l]) ∙ 1"
using True by (subst foldl_intersperse, auto)
also have "... = foldl (λx N. M *v (N *v x)) stat (map (λ_. mat 1) [0..<l]) ∙ 1"
unfolding P_def using I_empty by simp
also have "... = foldl (λx _. M *v x) stat [0..<l] ∙ 1"
unfolding foldl_map by simp
also have "... = stat ∙ (1 :: real^'n)"
by (induction l, auto simp add:markov_orth_inv[OF assms(6)])
finally show ?thesis by simp
next
case False
then show ?thesis by simp
qed
also have "... = 1"
unfolding stat_def by (simp add:inner_vec_def)
also have "... ≤ ?R" unfolding I_empty by simp
finally show ?thesis by simp
qed
lemma uniform_property_alg:
fixes x :: "('n :: finite)" and l :: nat
assumes "i < l"
defines "P j ≡ (if j = i then diag (ind_vec {x}) else mat 1)"
assumes "markov M"
shows "foldl (λx M. M *v x) stat (intersperse M (map P [0..<l])) ∙ 1 = 1 / CARD('n)"
(is "?L = ?R")
proof -
have a:"l > 0" using assms(1) by simp
have 0: "foldl (λx N. M *v (N *v x)) y (xs) ∙ 1 = y ∙ 1" if "set xs ⊆ {mat 1}" for xs y
using that
proof (induction xs rule:rev_induct)
case Nil
then show ?case by simp
next
case (snoc x xs)
have "x = mat 1"
using snoc(2) by simp
hence "foldl (λx N. M *v (N *v x)) y (xs @ [x]) ∙ 1 = foldl (λx N. M *v (N *v x)) y xs ∙ 1"
by (simp add:markov_orth_inv[OF assms(3)])
also have "... = y ∙ 1"
using snoc(2) by (intro snoc(1)) auto
finally show ?case by simp
qed
have M_stat: "M *v stat = stat"
using assms(3) unfolding stat_def matrix_vector_mult_scaleR markov_def by simp
hence 1: "(foldl (λx N. M *v (N *v x)) stat xs) = stat" if "set xs ⊆ {mat 1}" for xs
using that by (induction xs, auto)
have "?L = foldl (λx M. M *v x) stat ((intersperse M (map P [0..<l]))@[M]) ∙ 1"
by (simp add:markov_orth_inv[OF assms(3)])
also have "... = foldl (λx N. M *v (N *v x)) stat (map P [0..<l]) ∙ 1"
using a by (subst foldl_intersperse) auto
also have "... = foldl (λx N. M *v (N *v x)) stat (map P ([0..<i+1]@[i+1..<l])) ∙ 1"
using assms(1) by (subst upto_append) auto
also have "... = foldl (λx N. M *v (N *v x)) stat (map P [0..<i + 1]) ∙ 1"
unfolding map_append foldl_append P_def by (subst 0) auto
also have "... = foldl (λx N. M *v (N *v x)) stat (map P ([0..<i]@[i])) ∙ 1"
by simp
also have "... = (M *v (diag (ind_vec {x}) *v stat)) ∙ 1"
unfolding map_append foldl_append P_def by (subst 1) auto
also have "... = (diag (ind_vec {x}) *v stat) ∙ 1"
by (simp add:markov_orth_inv[OF assms(3)])
also have "... = ((1/CARD('n)) *⇩R ind_vec {x}) ∙ 1"
unfolding diag_def ind_vec_def stat_def matrix_vector_mult_def
by (intro arg_cong2[where f="(∙)"] refl)
(vector of_bool_def sum.If_cases if_distrib if_distribR)
also have "... = (1/CARD('n)) * (ind_vec {x} ∙ 1)"
by simp
also have "... = (1/CARD('n)) * 1"
unfolding inner_vec_def ind_vec_def of_bool_def
by (intro arg_cong2[where f="(*)"] refl) (simp)
finally show ?thesis by simp
qed
end
lemma foldl_matrix_mult_expand:
fixes Ms :: "(('r::{semiring_1,comm_monoid_mult})^'a^'a) list"
shows "(foldl (λx M. M *v x) a Ms) $ k = (∑x | length x = length Ms+1 ∧ x! length Ms = k.
(∏ i< length Ms. (Ms ! i) $ (x ! (i+1)) $ (x ! i)) * a $ (x ! 0))"
proof (induction Ms arbitrary: k rule:rev_induct)
case Nil
have "length x = Suc 0 ⟹ x = [x!0]" for x :: "'a list"
by (cases x, auto)
hence "{x. length x = Suc 0 ∧ x ! 0 = k} = {[k]}"
by auto
thus ?case by auto
next
case (snoc M Ms)
let ?l = "length Ms"
have 0: "finite {w. length w = Suc (length Ms) ∧ w ! length Ms = i}" for i :: 'a
using finite_lists_length_eq[where A="UNIV::'a set" and n="?l +1"] by simp
have "take (?l+1) x @ [x ! (?l+1)] = x" if "length x = ?l+2" for x :: "'a list"
proof -
have "take (?l+1) x @ [x ! (?l+1)] = take (Suc (?l+1)) x"
using that by (intro take_Suc_conv_app_nth[symmetric], simp)
also have "... = x"
using that by simp
finally show ?thesis by simp
qed
hence 1: "bij_betw (take (?l+1)) {w. length w=?l+2 ∧ w!(?l+1) =k} {w. length w = ?l+1}"
by (intro bij_betwI[where g="λx. x@[k]"]) (auto simp add:nth_append)
have "foldl (λx M. M *v x) a (Ms @ [M]) $ k = (∑j∈UNIV. M$k$j *(foldl (λx M. M *v x) a Ms $ j))"
by (simp add:matrix_vector_mult_def)
also have "... =
(∑j∈UNIV. M$k$j * (∑w|length w=?l+1∧w!?l=j. (∏i<?l. Ms!i $ w!(i+1) $ w!i) * a $ w!0))"
unfolding snoc by simp
also have "... =
(∑j∈UNIV. (∑w|length w=?l+1∧w!?l=j. M$k$w!?l * (∏i<?l. Ms!i $ w!(i+1) $ w!i) * a $ w!0))"
by (intro sum.cong refl) (simp add: sum_distrib_left algebra_simps)
also have "... = (∑w∈ (⋃j ∈ UNIV. {w. length w=?l+1 ∧ w!?l =j}).
M$k$w!?l*(∏i<?l. Ms!i $ w!(i+1) $ w!i) * a $ w!0)"
using 0 by (subst sum.UNION_disjoint, simp, simp) auto
also have "... = (∑w | length w=?l+1. M$k$(w!?l)*(∏i<?l. Ms!i $ w!(i+1) $ w!i) * a $ w!0)"
by (intro sum.cong arg_cong2[where f="(*)"] refl) auto
also have "... = (∑w ∈ take (?l+1) ` {w. length w=?l+2 ∧ w!(?l+1) =k}.
M$k$w!?l*(∏i<?l. Ms!i $ w!(i+1) $ w!i) * a $ w!0)"
using 1 unfolding bij_betw_def by (intro sum.cong refl, auto)
also have "... = (∑w|length w=?l+2∧w!(?l+1)=k. M$k$w!?l*(∏i<?l. Ms!i $ w!(i+1) $ w!i)* a$w!0)"
using 1 unfolding bij_betw_def by (subst sum.reindex, auto)
also have "... = (∑w|length w=?l+2∧w!(?l+1)=k.
(Ms@[M])!?l$k$w!?l*(∏i<?l. (Ms@[M])!i $ w!(i+1) $ w!i)* a$w!0)"
by (intro sum.cong arg_cong2[where f="(*)"] prod.cong refl) (auto simp add:nth_append)
also have "... = (∑w|length w=?l+2∧w!(?l+1)=k. (∏i<(?l+1). (Ms@[M])!i $ w!(i+1) $ w!i)* a$w!0)"
by (intro sum.cong, auto simp add:algebra_simps)
finally have "foldl (λx M. M *v x) a (Ms @ [M]) $ k =
(∑ w | length w = ?l+2 ∧ w ! (?l+1) = k. (∏i<(?l+1). (Ms@[M])!i $ w!(i+1) $ w!i)* a$w!0)"
by simp
then show ?case by simp
qed
lemma foldl_matrix_mult_expand_2:
fixes Ms :: "(real^'a^'a) list"
shows "(foldl (λx M. M *v x) a Ms) ∙ 1 = (∑x | length x = length Ms+1.
(∏ i< length Ms. (Ms ! i) $ (x ! (i+1)) $ (x ! i)) * a $ (x ! 0))"
(is "?L = ?R")
proof -
let ?l = "length Ms"
have "?L = (∑j ∈ UNIV. (foldl (λx M. M *v x) a Ms) $ j)"
by (simp add:inner_vec_def)
also have "... = (∑j∈UNIV. ∑x|length x=?l+1 ∧ x!?l=j.(∏i<?l. Ms!i $ x!(i+1) $ x!i) * a $ x!0)"
unfolding foldl_matrix_mult_expand by simp
also have "... = (∑x ∈ (⋃j∈ UNIV.{w. length w = length Ms+1 ∧ w ! length Ms = j}).
(∏ i< length Ms. (Ms ! i) $ (x ! (i+1)) $ (x ! i)) * a $ (x ! 0))"
using finite_lists_length_eq[where A="UNIV::'a set" and n="?l +1"]
by (intro sum.UNION_disjoint[symmetric]) auto
also have "... = ?R"
by (intro sum.cong, auto)
finally show ?thesis by simp
qed
end