Theory Suppes_Theorem

chapter ‹ Suppes' Theorem \label{sec:suppes-theorem} ›

theory Suppes_Theorem
  imports Probability_Logic
begin

unbundle no funcset_syntax

text ‹ An elementary completeness theorem for inequalities for probability logic
       is due to Patrick Suppes @{cite suppesProbabilisticInferenceConcept1966}. ›

text ‹ A consequence of this Suppes' theorem is an elementary form of ‹collapse›,
       which asserts that inequalities for probabilities are logically
       equivalent to the more restricted class of ‹Dirac measures› as
       defined in \S\ref{sec:dirac-measures}. ›

section ‹ Suppes' List Theorem \label{sec:suppes-theorem-for-lists} ›

text ‹ We first establish Suppes' theorem for lists of propositions. This
       is done by establishing our first completeness theorem using
       ‹Dirac measures›. ›

text ‹ First, we use the result from \S\ref{sec:basic-probability-inequality-results}
       that shows ⊢ φ → ⨆ Ψ› implies 𝒫 φ ≤ (∑ψ←Ψ. 𝒫 ψ)›.
       This can be understood as a ‹soundness› result. ›

text ‹ To show completeness, assume ¬ ⊢ φ → ⨆ Ψ›. From this
       obtain a maximally consistent termΩ such that φ → ⨆ Ψ ∉ Ω›. 
       We then define termδ χ = (if (χ  Ω) then 1 else 0) and 
       show termδ is a  ‹Dirac measure› such that δ φ ≤ (∑ψ←Ψ. δ ψ)›. ›

lemma (in classical_logic) dirac_list_summation_completeness:
  "( δ  dirac_measures. δ φ  (ψΨ. δ ψ)) =  φ   Ψ"
proof -
  {
    fix δ :: "'a  real"
    assume "δ  dirac_measures"
    from this interpret probability_logic "(λ φ.  φ)" "(→)" "" "δ"
      unfolding dirac_measures_def
      by auto
    assume " φ   Ψ"
    hence "δ φ  (ψΨ. δ ψ)"
      using implication_list_summation_inequality
      by auto
  }
  moreover {
    assume "¬  φ   Ψ"
    from this obtain Ω where Ω:
      "MCS Ω"
      "φ  Ω"
      " Ψ  Ω"
      by (meson
            insert_subset
            formula_consistent_def
            formula_maximal_consistency
            formula_maximally_consistent_extension
            formula_maximally_consistent_set_def_def
            set_deduction_base_theory
            set_deduction_reflection
            set_deduction_theorem)
    hence" ψ  set Ψ. ψ  Ω"
      using arbitrary_disjunction_exclusion_MCS by blast
    define δ where "δ = (λ χ . if χΩ then (1 :: real) else 0)"
    from  ψ  set Ψ. ψ  Ω have "(ψΨ. δ ψ) = 0"
      unfolding δ_def
      by (induct Ψ, simp, simp)
    hence "¬ δ φ  (ψΨ. δ ψ)"
      unfolding δ_def
      by (simp add: Ω(2))
    hence
      " δ  dirac_measures. ¬ (δ φ  (ψΨ. δ ψ))"
      unfolding δ_def
      using Ω(1) MCS_dirac_measure by auto
  }
  ultimately show ?thesis by blast
qed

theorem (in classical_logic) list_summation_completeness:
  "( 𝒫  probabilities. 𝒫 φ  (ψΨ. 𝒫 ψ)) =  φ   Ψ"
  (is "?lhs = ?rhs")
proof
  assume ?lhs
  hence " δ  dirac_measures. δ φ  (ψΨ. δ ψ)"
    unfolding dirac_measures_def probabilities_def
    by blast
  thus ?rhs
    using dirac_list_summation_completeness by blast
next
  assume ?rhs
  show ?lhs
  proof
    fix 𝒫 :: "'a  real"
    assume "𝒫  probabilities"
    from this interpret probability_logic "(λ φ.  φ)" "(→)"  𝒫
      unfolding probabilities_def
      by auto
    show "𝒫 φ  (ψΨ. 𝒫 ψ)"
      using ?rhs implication_list_summation_inequality
      by simp
  qed
qed

text ‹ The collapse theorem asserts that to prove an inequalities for all 
       probabilities in probability logic, one only needs to consider the 
       case of functions which take on values of 0 or 1. ›

lemma (in classical_logic) suppes_collapse:
  "( 𝒫  probabilities. 𝒫 φ  (ψΨ. 𝒫 ψ))
      = ( δ  dirac_measures. δ φ  (ψΨ. δ ψ))"
  by (simp add:
        dirac_list_summation_completeness
        list_summation_completeness)

lemma (in classical_logic) probability_member_neg:
  fixes 𝒫
  assumes "𝒫  probabilities"
  shows "𝒫 ( φ) = 1 - 𝒫 φ"
proof -
  from assms interpret probability_logic "(λ φ.  φ)" "(→)"  𝒫
    unfolding probabilities_def
    by auto
  show ?thesis
    by (simp add: complementation)
qed

text ‹ Suppes' theorem has a philosophical interpretation.
       It asserts that if termΨ :⊢ φ, then our ‹uncertainty›
       in termφ is bounded above by our uncertainty in termΨ.
       Here the uncertainty in the proposition termφ is 1 - 𝒫 φ›.
       Our uncertainty in termΨ, on the other hand, is
       ∑ψ←Ψ. 1 - 𝒫 ψ›. ›

theorem (in classical_logic) suppes_list_theorem:
  "Ψ :⊢ φ = ( 𝒫  probabilities. (ψΨ. 1 - 𝒫 ψ)  1 - 𝒫 φ)"
proof -
  have
    "Ψ :⊢ φ = ( 𝒫  probabilities. (ψ   Ψ. 𝒫 ψ)  𝒫 ( φ))"
    using
      list_summation_completeness
      weak_biconditional_weaken
      contra_list_curry_uncurry
      list_deduction_def
    by blast
  moreover have
    " 𝒫  probabilities. (ψ  ( Ψ). 𝒫 ψ) = (ψ  Ψ. 𝒫 ( ψ))"
    by (induct Ψ, auto)
  ultimately show ?thesis
    using probability_member_neg
    by (induct Ψ, simp+)
qed

section ‹ Suppes' Set Theorem ›

text ‹ Suppes theorem also obtains for ‹sets›. ›

lemma (in classical_logic) dirac_set_summation_completeness:
  "( δ  dirac_measures. δ φ  (ψ set Ψ. δ ψ)) =  φ   Ψ"
  by (metis
        dirac_list_summation_completeness
        modus_ponens
        arbitrary_disjunction_remdups
        biconditional_left_elimination
        biconditional_right_elimination
        hypothetical_syllogism
        sum.set_conv_list)

theorem (in classical_logic) set_summation_completeness:
  "( δ  probabilities. δ φ  (ψ set Ψ. δ ψ)) =  φ   Ψ"
  by (metis
        dirac_list_summation_completeness
        dirac_set_summation_completeness
        list_summation_completeness
        sum.set_conv_list)

lemma (in classical_logic) suppes_set_collapse:
  "( 𝒫  probabilities. 𝒫 φ  (ψ  set Ψ. 𝒫 ψ))
      = ( δ  dirac_measures. δ φ  (ψ  set Ψ. δ ψ))"
  by (simp add:
        dirac_set_summation_completeness
        set_summation_completeness)

text ‹ In our formulation of logic, there is not reason that ∼ a = ∼ b›
       while terma  b.  As a consequence the Suppes theorem
       for sets presented below is different than the one given in
       \S\ref{sec:suppes-theorem-for-lists}. ›

theorem (in classical_logic) suppes_set_theorem:
  "Ψ :⊢ φ
     = ( 𝒫  probabilities. (ψ  set ( Ψ). 𝒫 ψ)  1 - 𝒫 φ)"
proof -
  have "Ψ :⊢ φ
          = ( 𝒫  probabilities. (ψ  set ( Ψ). 𝒫 ψ)  𝒫 ( φ))"
    using
      contra_list_curry_uncurry
      list_deduction_def
      set_summation_completeness
      weak_biconditional_weaken
    by blast
  thus ?thesis
    using probability_member_neg
    by (induct Ψ, auto)
qed

section ‹ Converse Suppes' Theorem ›

text ‹ A formulation of the converse of Suppes' theorem obtains
       for lists/sets of ‹logically disjoint› propositions. ›

lemma (in probability_logic) exclusive_sum_list_identity:
  assumes "  Φ"
  shows "𝒫 ( Φ) = (φΦ. 𝒫 φ)"
  using assms
proof (induct Φ)
  case Nil
  then show ?case
    by (simp add: gaines_weatherson_antithesis)
next
  case (Cons φ Φ)
  assume "  (φ # Φ)"
  hence "  (φ   Φ)" "  Φ" by simp+
  hence "𝒫 ((φ # Φ)) = 𝒫 φ + 𝒫 ( Φ)"
        "𝒫 ( Φ) = (φΦ. 𝒫 φ)"
    using Cons.hyps probability_additivity by auto
  hence "𝒫 ((φ # Φ)) = 𝒫 φ + (φΦ. 𝒫 φ)" by auto
  thus ?case by simp
qed

lemma sum_list_monotone:
  fixes f :: "'a  real"
  assumes " x. f x  0"
     and  "set Φ  set Ψ"
     and  "distinct Φ"
   shows "(φΦ. f φ)  (ψΨ. f ψ)"
  using assms
proof -
  assume " x. f x  0"
  have "Φ. set Φ  set Ψ
              distinct Φ
              (φΦ. f φ)  (ψΨ. f ψ)"
  proof (induct Ψ)
    case Nil
    then show ?case by simp
  next
    case (Cons ψ Ψ)
    {
      fix Φ
      assume "set Φ  set (ψ # Ψ)"
         and "distinct Φ"
      have "(φΦ. f φ)  (ψ'(ψ # Ψ). f ψ')"
      proof -
        {
          assume "ψ  set Φ"
          with set Φ  set (ψ # Ψ) have "set Φ  set Ψ" by auto
          hence "(φΦ. f φ)  (ψΨ. f ψ)"
            using Cons.hyps distinct Φ by auto
          moreover have "f ψ  0" using  x. f x  0 by metis
          ultimately have ?thesis by simp
        }
        moreover
        {
          assume "ψ  set Φ"
          hence "set Φ = insert ψ (set (removeAll ψ Φ))"
            by auto
          with set Φ  set (ψ # Ψ) have "set (removeAll ψ Φ)  set Ψ"
            by (metis
                  insert_subset
                  list.simps(15)
                  set_removeAll
                  subset_insert_iff)
          moreover from distinct Φ have "distinct (removeAll ψ Φ)"
            by (meson distinct_removeAll)
          ultimately have "(φ(removeAll ψ Φ). f φ)  (ψΨ. f ψ)"
            using Cons.hyps
            by simp
          moreover from ψ  set Φ distinct Φ
          have "(φΦ. f φ) = f ψ + (φ(removeAll ψ Φ). f φ)"
            using distinct_remove1_removeAll sum_list_map_remove1
            by fastforce
          ultimately have ?thesis using  x. f x  0
            by simp
        }
        ultimately show ?thesis by blast
      qed
    }
    thus ?case by blast
  qed
  moreover assume "set Φ  set Ψ" and "distinct Φ"
  ultimately show ?thesis by blast
qed

lemma count_remove_all_sum_list:
  fixes f :: "'a  real"
  shows "real (count_list xs x) * f x + (x'(removeAll x xs). f x')
           = (xxs. f x)"
  by (induct xs, simp, simp, metis combine_common_factor mult_cancel_right1)

lemma (in classical_logic) dirac_exclusive_implication_completeness:
  "( δ  dirac_measures. (φΦ. δ φ)  δ ψ) = (  Φ     Φ  ψ)"
proof -
  {
    fix δ
    assume "δ  dirac_measures"
    from this interpret probability_logic "(λ φ.  φ)" "(→)" "" "δ"
      unfolding dirac_measures_def
      by simp
    assume "  Φ" "  Φ  ψ"
    hence "(φΦ. δ φ)  δ ψ"
      using exclusive_sum_list_identity monotonicity by fastforce
  }
  moreover
  {
    assume "¬   Φ"
    hence "( φ  set Φ.  ψ  set Φ.
              φ  ψ  ¬   (φ  ψ))  ( φ  duplicates Φ. ¬   φ)"
      using exclusive_equivalence set_deduction_base_theory by blast
    hence "¬ ( δ  dirac_measures. (φΦ. δ φ)  δ ψ)"
    proof (elim disjE)
      assume " φ  set Φ.  χ  set Φ. φ  χ  ¬   (φ  χ)"
      from this obtain φ and χ
        where φχ_properties:
          "φ  set Φ"
          "χ  set Φ"
          "φ  χ"
          "¬   (φ  χ)"
        by blast
      from this obtain Ω where Ω: "MCS Ω" " (φ  χ)  Ω"
        by (meson
              insert_subset
              formula_consistent_def
              formula_maximal_consistency
              formula_maximally_consistent_extension
              formula_maximally_consistent_set_def_def
              set_deduction_base_theory
              set_deduction_reflection
              set_deduction_theorem)
      let  = "λ χ. if χΩ then (1 :: real) else 0"
      from Ω have "φ  Ω" "χ  Ω"
        by (metis
              formula_maximally_consistent_set_def_implication
              maximally_consistent_set_def
              conjunction_def
              negation_def)+
      with φχ_properties have
          "(φ[φ, χ].  φ) = 2"
          "set [φ, χ]  set Φ"
          "distinct [φ, χ]"
          "φ.  φ  0"
        by simp+
      hence "(φΦ.  φ)  2" using sum_list_monotone by metis
      hence "¬ (φΦ.  φ)   (ψ)" by auto
      thus ?thesis
        using Ω(1) MCS_dirac_measure
        by auto
    next
      assume " φ  duplicates Φ. ¬   φ"
      from this obtain φ where φ: "φ  duplicates Φ" "¬   φ"
        using
          exclusive_equivalence [where Γ="{}"]
          set_deduction_base_theory
        by blast
      from φ obtain Ω where Ω: "MCS Ω" " φ  Ω"
        by (meson
              insert_subset
              formula_consistent_def
              formula_maximal_consistency
              formula_maximally_consistent_extension
              formula_maximally_consistent_set_def_def
              set_deduction_base_theory
              set_deduction_reflection
              set_deduction_theorem)
      hence "φ  Ω"
        using negation_def by auto
      let  = "λ χ. if χΩ then (1 :: real) else 0"
      from φ have "count_list Φ φ  2"
        using duplicates_alt_def [where xs="Φ"]
        by blast
      hence "real (count_list Φ φ) *  φ  2" using φ  Ω by simp
      moreover
      {
        fix Ψ
        have "(φΨ.  φ)  0" by (induct Ψ, simp, simp)
      }
      moreover have "(0::real)
                           (aremoveAll φ Φ. if a  Ω then 1 else 0)"
        using Ψ. 0  (φΨ. if φ  Ω then 1 else 0)
        by presburger
      ultimately have "real (count_list Φ φ) *  φ
                            + ( φ  (removeAll φ Φ).  φ)  2"
        using 2  real (count_list Φ φ) * (if φ  Ω then 1 else 0)
        by linarith
      hence "(φΦ.  φ)  2" by (metis count_remove_all_sum_list)
      hence "¬ (φΦ.  φ)   (ψ)" by auto
      thus ?thesis
        using Ω(1) MCS_dirac_measure
        by auto
    qed
  }
  moreover
  {
    assume "¬   Φ  ψ"
    from this obtain Ω φ
      where
        Ω: "MCS Ω"
        and ψ: "ψ  Ω"
        and φ: "φ  set Φ" "φ  Ω"
      by (meson
            insert_subset
            formula_consistent_def
            formula_maximal_consistency
            formula_maximally_consistent_extension
            formula_maximally_consistent_set_def_def
            arbitrary_disjunction_exclusion_MCS
            set_deduction_base_theory
            set_deduction_reflection
            set_deduction_theorem)
    let  = "λ χ. if χΩ then (1 :: real) else 0"
    from φ have "(φΦ.  φ)  1"
    proof (induct Φ)
      case Nil
      then show ?case by simp
    next
      case (Cons φ' Φ)
      obtain f :: "real list  real" where f:
        "rs. f rs  set rs  ¬ 0  f rs  0  sum_list rs"
        using sum_list_nonneg by metis
      moreover have "f (map  Φ)  set (map  Φ)  0  f (map  Φ)"
        by fastforce
      ultimately show ?case
        by (simp, metis Cons.hyps Cons.prems(1) φ(2) set_ConsD)
    qed
    hence "¬ (φΦ.  φ)   (ψ)" using ψ by auto
    hence "¬ ( δ  dirac_measures. (φΦ. δ φ)  δ ψ)"
      using Ω(1) MCS_dirac_measure
      by auto
  }
  ultimately show ?thesis by blast
qed

theorem (in classical_logic) exclusive_implication_completeness:
  "( 𝒫  probabilities. (φΦ. 𝒫 φ)  𝒫 ψ) = (  Φ     Φ  ψ)"
  (is "?lhs = ?rhs")
proof
  assume ?lhs
  thus ?rhs
    by (meson
          dirac_exclusive_implication_completeness
          dirac_measures_subset
          subset_eq)
next
  assume ?rhs
  show ?lhs
  proof
    fix 𝒫 :: "'a  real"
    assume "𝒫  probabilities"
    from this interpret probability_logic "(λ φ.  φ)" "(→)"  𝒫
      unfolding probabilities_def
      by simp
    show "(φΦ. 𝒫 φ)  𝒫 ψ"
      using
        ?rhs
        exclusive_sum_list_identity
        monotonicity
      by fastforce
  qed
qed

lemma (in classical_logic) dirac_inequality_completeness:
  "( δ  dirac_measures. δ φ  δ ψ) =  φ  ψ"
proof -
  have "  [φ]"
    by (simp add: conjunction_right_elimination negation_def)
  hence "(  [φ]     [φ]  ψ) =  φ  ψ"
    by (metis
          arbitrary_disjunction.simps(1)
          arbitrary_disjunction.simps(2)
          disjunction_def implication_equivalence
          negation_def
          weak_biconditional_weaken)
  thus ?thesis
    using dirac_exclusive_implication_completeness [where Φ="[φ]"]
    by auto
qed

section ‹ Implication Inequality Completeness ›

text ‹ The following theorem establishes the converse of
       ⊢ φ → ψ ⟹ 𝒫 φ ≤ 𝒫 ψ›, which was
       proved in \S\ref{sec:prob-logic-alt-def}. ›

theorem (in classical_logic) implication_inequality_completeness:
  "( 𝒫  probabilities. 𝒫 φ  𝒫 ψ) =  φ  ψ"
proof -
  have "  [φ]"
    by (simp add: conjunction_right_elimination negation_def)
  hence "(  [φ]     [φ]  ψ) =  φ  ψ"
    by (metis
          arbitrary_disjunction.simps(1)
          arbitrary_disjunction.simps(2)
          disjunction_def implication_equivalence
          negation_def
          weak_biconditional_weaken)
  thus ?thesis
    using exclusive_implication_completeness [where Φ="[φ]"]
    by simp
qed

section ‹ Characterizing Logical Exclusiveness In Probability Logic ›

text ‹ Finally, we can say that 𝒫 (⨆ Φ) = (∑φ←Φ. 𝒫 φ)› if and only
       if the propositions in termΦ are mutually exclusive (i.e. ⊢ ∐ Φ›).
       This result also obtains for sets. ›

lemma (in classical_logic) dirac_exclusive_list_summation_completeness:
  "( δ  dirac_measures. δ ( Φ) = (φΦ. δ φ)) =   Φ"
  by (metis
        antisym_conv
        dirac_exclusive_implication_completeness
        dirac_list_summation_completeness
        trivial_implication)

theorem (in classical_logic) exclusive_list_summation_completeness:
  "( 𝒫  probabilities. 𝒫 ( Φ) = (φΦ. 𝒫 φ)) =   Φ"
  by (metis
        antisym_conv
        exclusive_implication_completeness
        list_summation_completeness
        trivial_implication)

lemma (in classical_logic) dirac_exclusive_set_summation_completeness:
  "( δ  dirac_measures. δ ( Φ) = (φ  set Φ. δ φ))
      =   (remdups Φ)"
  by (metis
        (mono_tags)
        eq_iff
        dirac_exclusive_implication_completeness
        dirac_set_summation_completeness
        trivial_implication
        set_remdups
        sum.set_conv_list    eq_iff
        dirac_exclusive_implication_completeness
        dirac_set_summation_completeness
        trivial_implication
        set_remdups
        sum.set_conv_list
        antisym_conv)

theorem (in classical_logic) exclusive_set_summation_completeness:
  "( 𝒫  probabilities.
        𝒫 ( Φ) = (φ  set Φ. 𝒫 φ)) =   (remdups Φ)"
  by (metis
        (mono_tags, opaque_lifting)
        antisym_conv
        exclusive_list_summation_completeness
        exclusive_implication_completeness
        implication_inequality_completeness
        set_summation_completeness
        order.refl
        sum.set_conv_list)

lemma (in probability_logic) exclusive_list_set_inequality:
  assumes "  Φ"
  shows "(φΦ. 𝒫 φ) = (φset Φ. 𝒫 φ)"
proof -
  have "distinct (remdups Φ)" using distinct_remdups by auto
  hence "duplicates (remdups Φ) = {}"
    by (induct "Φ", simp+)
  moreover have "set (remdups Φ) = set Φ"
    by (induct "Φ", simp, simp add: insert_absorb)
  moreover have "(φ  duplicates Φ.   φ)
                ( φ  set Φ.  ψ  set Φ. (φ  ψ)    (φ  ψ))"
    using
      assms
      exclusive_elimination1
      exclusive_elimination2
      set_deduction_base_theory
    by blast
  ultimately have
    "(φduplicates (remdups Φ).   φ)
    ( φ  set (remdups Φ).  ψ  set (remdups Φ).
          (φ  ψ)    (φ  ψ))"
    by auto
  hence "  (remdups Φ)"
    by (meson exclusive_equivalence set_deduction_base_theory)
  hence "(φset Φ. 𝒫 φ) = 𝒫 ( Φ)"
    by (metis
          arbitrary_disjunction_remdups
          biconditional_equivalence
          exclusive_sum_list_identity
          sum.set_conv_list)
  moreover have "(φΦ. 𝒫 φ) = 𝒫 ( Φ)"
    by (simp add: assms exclusive_sum_list_identity)
  ultimately show ?thesis by metis
qed

unbundle funcset_syntax

end