Quelle  Hahn_Banach.thy

(*  Title:      HOL/Hahn_Banach/Hahn_Banach.thy
  Author: Gertrud Bauer, TU Munich
*)

section ‹The Hahn-Banach Theorem›

theory Hahn_Banach
imports Hahn_Banach_Lemmas
begin

text ‹
  We present the proof of two different versions of the Hahn-Banach Theorem,
  closely following 🍋‹‹\S36› in "Heuser:1986"›.
 ›


subsection ‹The Hahn-Banach Theorem for vector spaces›

paragraph ‹Hahn-Banach Theorem.›
text ‹
  Let ‹F› be a subspace of a real vector space ‹E›, let ‹p› be a semi-norm on
  ‹E›, and ‹f› be a linear form defined on ‹F› such that ‹f› is bounded by
  ‹p›, i.e. ‹∀x ∈ F. f x ≤ p x›. Then ‹f› can be extended to a linear form ‹h›
  on ‹E› such that ‹h› is norm-preserving, i.e. ‹h› is also bounded by ‹p›.
 ›

paragraph ‹Proof Sketch.›
text ‹
  🪙 Define ‹M› as the set of norm-preserving extensions of ‹f› to subspaces of
  ‹E›. The linear forms in ‹M› are ordered by domain extension.
 
  🪙 We show that every non-empty chain in ‹M› has an upper bound in ‹M›.
 
  🪙 With Zorn's Lemma we conclude that there is a maximal function ‹g› in ‹M›.
 
  🪙 The domain ‹H› of ‹g› is the whole space ‹E›, as shown by classical
  contradiction:
 
  🪙 Assuming ‹g› is not defined on whole ‹E›, it can still be extended in a
  norm-preserving way to a super-space ‹H'› of ‹H›.
 
  🪙 Thus ‹g› can not be maximal. Contradiction!
 ›

theorem Hahn_Banach:
  assumes E: "vectorspace E" and "subspace F E"
    and "seminorm E p" and "linearform F f"
  assumes fp: "∀x ∈ F. f x ≤ p x"
  shows "∃h. linearform E h ∧ (∀x ∈ F. h x = f x) ∧ (∀x ∈ E. h x ≤ p x)"
    🍋 ‹Let ‹E› be a vector space, ‹F› a subspace of ‹E›, ‹p› a seminorm on ‹E›,›
    🍋 ‹and ‹f› a linear form on ‹F› such that ‹f› is bounded by ‹p›,›
    🍋 ‹then ‹f› can be extended to a linear form ‹h› on ‹E› in a norm-preserving way. 🪙›
proof -
  interpret vectorspace E by fact
  interpret subspace F E by fact
  interpret seminorm E p by fact
  interpret linearform F f by fact
  define M where "M = norm_pres_extensions E p F f"
  then have M: "M = …" by (simp only:)
  from E have F: "vectorspace F" ..
  note FE = ‹F ⊴ E›
  have "∪c ∈ M" if cM: "c ∈ chains M" and ex: "∃x. x ∈ c" for c
    🍋 ‹Show that every non-empty chain ‹c› of ‹M› has an upper bound in ‹M›:›
    🍋 ‹‹∪c› is greater than any element of the chain ‹c›, so it suffices to show ‹∪c ∈ M›.›
    unfolding M_def
  proof (rule norm_pres_extensionI)
    let ?H = "domain (∪c)"
    let ?h = "funct (∪c)"

    have a: "graph ?H ?h = ∪c"
    proof (rule graph_domain_funct)
      fix x y z assume "(x, y) ∈ ∪c" and "(x, z) ∈ ∪c"
      with M_def cM show "z = y" by (rule sup_definite)
    qed
    moreover from M cM a have "linearform ?H ?h"
      by (rule sup_lf)
    moreover from a M cM ex FE E have "?H ⊴ E"
      by (rule sup_subE)
    moreover from a M cM ex FE have "F ⊴ ?H"
      by (rule sup_supF)
    moreover from a M cM ex have "graph F f ⊆ graph ?H ?h"
      by (rule sup_ext)
    moreover from a M cM have "∀x ∈ ?H. ?h x ≤ p x"
      by (rule sup_norm_pres)
    ultimately show "∃H h. ∪c = graph H h
        ∧ linearform H h
        ∧ H ⊴ E
        ∧ F ⊴ H
        ∧ graph F f ⊆ graph H h
        ∧ (∀x ∈ H. h x ≤ p x)" by blast
  qed
  then have "∃g ∈ M. ∀x ∈ M. g ⊆ x ⟶ x = g"
  🍋 ‹With Zorn's Lemma we can conclude that there is a maximal element in ‹M›. 🪙›
  proof (rule Zorn's_Lemma)
      🍋 ‹We show that ‹M› is non-empty:›
    show "graph F f ∈ M"
      unfolding M_def
    proof (rule norm_pres_extensionI2)
      show "linearform F f" by fact
      show "F ⊴ E" by fact
      from F show "F ⊴ F" by (rule vectorspace.subspace_refl)
      show "graph F f ⊆ graph F f" ..
      show "∀x∈F. f x ≤ p x" by fact
    qed
  qed
  then obtain g where gM: "g ∈ M" and gx: "∀x ∈ M. g ⊆ x ⟶ g = x"
    by blast
  from gM obtain H h where
      g_rep: "g = graph H h"
    and linearform: "linearform H h"
    and HE: "H ⊴ E" and FH: "F ⊴ H"
    and graphs: "graph F f ⊆ graph H h"
    and hp: "∀x ∈ H. h x ≤ p x" unfolding M_def ..
      🍋 ‹‹g› is a norm-preserving extension of ‹f›, in other words:›
      🍋 ‹‹g› is the graph of some linear form ‹h› defined on a subspace ‹H› of ‹E›,›
      🍋 ‹and ‹h› is an extension of ‹f› that is again bounded by ‹p›. 🪙›
  from HE E have H: "vectorspace H"
    by (rule subspace.vectorspace)

  have HE_eq: "H = E"
    🍋 ‹We show that ‹h› is defined on whole ‹E› by classical contradiction. 🪙›
  proof (rule classical)
    assume neq: "H ≠ E"
      🍋 ‹Assume ‹h› is not defined on whole ‹E›. Then show that ‹h› can be extended›
      🍋 ‹in a norm-preserving way to a function ‹h'› with the graph ‹g'›. 🪙›
    have "∃g' ∈ M. g ⊆ g' ∧ g ≠ g'"
    proof -
      from HE have "H ⊆ E" ..
      with neq obtain x' where x'E: "x' ∈ E" and "x' ∉ H" by blast
      obtain x': "x' ≠ 0"
      proof
        show "x' ≠ 0"
        proof
          assume "x' = 0"
          with H have "x' ∈ H" by (simp only: vectorspace.zero)
          with ‹x' ∉ H› show False by contradiction
        qed
      qed

      define H' where "H' = H + lin x'"
        🍋 ‹Define ‹H'› as the direct sum of ‹H› and the linear closure of ‹x'›. 🪙›
      have HH': "H ⊴ H'"
      proof (unfold H'_def)
        from x'E have "vectorspace (lin x')" ..
        with H show "H ⊴ H + lin x'" ..
      qed

      obtain xi where
        xi: "∀y ∈ H. - p (y + x') - h y ≤ xi
          ∧ xi ≤ p (y + x') - h y"
        🍋 ‹Pick a real number ‹ξ› that fulfills certain inequality; this will›
        🍋 ‹be used to establish that ‹h'› is a norm-preserving extension of ‹h›.
  \label{ex-xi-use}🪙›
      proof -
        from H have "∃xi. ∀y ∈ H. - p (y + x') - h y ≤ xi
            ∧ xi ≤ p (y + x') - h y"
        proof (rule ex_xi)
          fix u v assume u: "u ∈ H" and v: "v ∈ H"
          with HE have uE: "u ∈ E" and vE: "v ∈ E" by auto
          from H u v linearform have "h v - h u = h (v - u)"
            by (simp add: linearform.diff)
          also from hp and H u v have "… ≤ p (v - u)"
            by (simp only: vectorspace.diff_closed)
          also from x'E uE vE have "v - u = x' + - x' + v + - u"
            by (simp add: diff_eq1)
          also from x'E uE vE have "… = v + x' + - (u + x')"
            by (simp add: add_ac)
          also from x'E uE vE have "… = (v + x') - (u + x')"
            by (simp add: diff_eq1)
          also from x'E uE vE E have "p … ≤ p (v + x') + p (u + x')"
            by (simp add: diff_subadditive)
          finally have "h v - h u ≤ p (v + x') + p (u + x')" .
          then show "- p (u + x') - h u ≤ p (v + x') - h v" by simp
        qed
        then show thesis by (blast intro: that)
      qed

      define h' where "h' x = (let (y, a) =
          SOME (y, a). x = y + a ⋅ x' ∧ y ∈ H in h y + a * xi)" for x
        🍋 ‹Define the extension ‹h'› of ‹h› to ‹H'› using ‹ξ›. 🪙›

      have "g ⊆ graph H' h' ∧ g ≠ graph H' h'"
        🍋 ‹‹h'› is an extension of ‹h› \dots 🪙›
      proof
        show "g ⊆ graph H' h'"
        proof -
          have "graph H h ⊆ graph H' h'"
          proof (rule graph_extI)
            fix t assume t: "t ∈ H"
            from E HE t have "(SOME (y, a). t = y + a ⋅ x' ∧ y ∈ H) = (t, 0)"
              using ‹x' ∉ H› ‹x' ∈ E› ‹x' ≠ 0› by (rule decomp_H'_H)
            with h'_def show "h t = h' t" by (simp add: Let_def)
          next
            from HH' show "H ⊆ H'" ..
          qed
          with g_rep show ?thesis by (simp only:)
        qed

        show "g ≠ graph H' h'"
        proof -
          have "graph H h ≠ graph H' h'"
          proof
            assume eq: "graph H h = graph H' h'"
            have "x' ∈ H'"
              unfolding H'_def
            proof
              from H show "0 ∈ H" by (rule vectorspace.zero)
              from x'E show "x' ∈ lin x'" by (rule x_lin_x)
              from x'E show "x' = 0 + x'" by simp
            qed
            then have "(x', h' x') ∈ graph H' h'" ..
            with eq have "(x', h' x') ∈ graph H h" by (simp only:)
            then have "x' ∈ H" ..
            with ‹x' ∉ H› show False by contradiction
          qed
          with g_rep show ?thesis by simp
        qed
      qed
      moreover have "graph H' h' ∈ M"
        🍋 ‹and ‹h'› is norm-preserving. 🪙›
      proof (unfold M_def)
        show "graph H' h' ∈ norm_pres_extensions E p F f"
        proof (rule norm_pres_extensionI2)
          show "linearform H' h'"
            using h'_def H'_def HE linearform ‹x' ∉ H› ‹x' ∈ E› ‹x' ≠ 0› E
            by (rule h'_lf)
          show "H' ⊴ E"
          unfolding H'_def
          proof
            show "H ⊴ E" by fact
            show "vectorspace E" by fact
            from x'E show "lin x' ⊴ E" ..
          qed
          from H ‹F ⊴ H› HH' show FH': "F ⊴ H'"
            by (rule vectorspace.subspace_trans)
          show "graph F f ⊆ graph H' h'"
          proof (rule graph_extI)
            fix x assume x: "x ∈ F"
            with graphs have "f x = h x" ..
            also have "… = h x + 0 * xi" by simp
            also have "… = (let (y, a) = (x, 0) in h y + a * xi)"
              by (simp add: Let_def)
            also have "(x, 0) =
                (SOME (y, a). x = y + a ⋅ x' ∧ y ∈ H)"
              using E HE
            proof (rule decomp_H'_H [symmetric])
              from FH x show "x ∈ H" ..
              from x' show "x' ≠ 0" .
              show "x' ∉ H" by fact
              show "x' ∈ E" by fact
            qed
            also have
              "(let (y, a) = (SOME (y, a). x = y + a ⋅ x' ∧ y ∈ H)
              in h y + a * xi) = h' x" by (simp only: h'_def)
            finally show "f x = h' x" .
          next
            from FH' show "F ⊆ H'" ..
          qed
          show "∀x ∈ H'. h' x ≤ p x"
            using h'_def H'_def ‹x' ∉ H› ‹x' ∈ E› ‹x' ≠ 0› E HE
              ‹seminorm E p› linearform and hp xi
            by (rule h'_norm_pres)
        qed
      qed
      ultimately show ?thesis ..
    qed
    then have "¬ (∀x ∈ M. g ⊆ x ⟶ g = x)" by simp
      🍋 ‹So the graph ‹g› of ‹h› cannot be maximal. Contradiction! 🪙›
    with gx show "H = E" by contradiction
  qed

  from HE_eq and linearform have "linearform E h"
    by (simp only:)
  moreover have "∀x ∈ F. h x = f x"
  proof
    fix x assume "x ∈ F"
    with graphs have "f x = h x" ..
    then show "h x = f x" ..
  qed
  moreover from HE_eq and hp have "∀x ∈ E. h x ≤ p x"
    by (simp only:)
  ultimately show ?thesis by blast
qed


subsection ‹Alternative formulation›

text ‹
  The following alternative formulation of the Hahn-Banach
  Theorem\label{abs-Hahn-Banach} uses the fact that for a real linear form ‹f›
  and a seminorm ‹p› the following inequality are equivalent:\footnote{This
  was shown in lemma @{thm [source] abs_ineq_iff} (see page
  \pageref{abs-ineq-iff}).}
  \begin{center}
  \begin{tabular}{lll}
  ‹∀x ∈ H. ∣h x∣ ≤ p x› & and & ‹∀x ∈ H. h x ≤ p x› \\
  \end{tabular}
  \end{center}
 ›

theorem abs_Hahn_Banach:
  assumes E: "vectorspace E" and FE: "subspace F E"
    and lf: "linearform F f" and sn: "seminorm E p"
  assumes fp: "∀x ∈ F. ∣f x∣ ≤ p x"
  shows "∃g. linearform E g
    ∧ (∀x ∈ F. g x = f x)
    ∧ (∀x ∈ E. ∣g x∣ ≤ p x)"
proof -
  interpret vectorspace E by fact
  interpret subspace F E by fact
  interpret linearform F f by fact
  interpret seminorm E p by fact
  have "∃g. linearform E g ∧ (∀x ∈ F. g x = f x) ∧ (∀x ∈ E. g x ≤ p x)"
    using E FE sn lf
  proof (rule Hahn_Banach)
    show "∀x ∈ F. f x ≤ p x"
      using FE E sn lf and fp by (rule abs_ineq_iff [THEN iffD1])
  qed
  then obtain g where lg: "linearform E g" and *: "∀x ∈ F. g x = f x"
      and **: "∀x ∈ E. g x ≤ p x" by blast
  have "∀x ∈ E. ∣g x∣ ≤ p x"
    using _ E sn lg **
  proof (rule abs_ineq_iff [THEN iffD2])
    show "E ⊴ E" ..
  qed
  with lg * show ?thesis by blast
qed


subsection ‹The Hahn-Banach Theorem for normed spaces›

text ‹
  Every continuous linear form ‹f› on a subspace ‹F› of a norm space ‹E›, can
  be extended to a continuous linear form ‹g› on ‹E› such that ‹∥f∥ = ∥g∥›.
 ›

theorem norm_Hahn_Banach:
  fixes V and norm (‹∥_∥›)
  fixes B defines "∧V f. B V f ≡ {0} ∪ {∣f x∣ / ∥x∥ | x. x ≠ 0 ∧ x ∈ V}"
  fixes fn_norm (‹∥_∥🍋_› [0, 1000] 999)
  defines "∧V f. ∥f∥🍋V ≡ ⊔(B V f)"
  assumes E_norm: "normed_vectorspace E norm" and FE: "subspace F E"
    and linearform: "linearform F f" and "continuous F f norm"
  shows "∃g. linearform E g
     ∧ continuous E g norm
     ∧ (∀x ∈ F. g x = f x)
     ∧ ∥g∥🍋E = ∥f∥🍋F"
proof -
  interpret normed_vectorspace E norm by fact
  interpret normed_vectorspace_with_fn_norm E norm B fn_norm
    by (auto simp: B_def fn_norm_def) intro_locales
  interpret subspace F E by fact
  interpret linearform F f by fact
  interpret continuous F f norm by fact
  have E: "vectorspace E" by intro_locales
  have F: "vectorspace F" by rule intro_locales
  have F_norm: "normed_vectorspace F norm"
    using FE E_norm by (rule subspace_normed_vs)
  have ge_zero: "0 ≤ ∥f∥🍋F"
    by (rule normed_vectorspace_with_fn_norm.fn_norm_ge_zero
      [OF normed_vectorspace_with_fn_norm.intro,
       OF F_norm ‹continuous F f norm› , folded B_def fn_norm_def])
  txt ‹We define a function ‹p› on ‹E› as follows:
  ‹p x = ∥f∥ ⋅ ∥x∥›\›
  define p where "p x = ∥f∥🍋F * ∥x∥" for x

  txt ‹‹p› is a seminorm on ‹E›:›
  have q: "seminorm E p"
  proof
    fix x y a assume x: "x ∈ E" and y: "y ∈ E"

    txt ‹‹p› is positive definite:›
    have "0 ≤ ∥f∥🍋F" by (rule ge_zero)
    moreover from x have "0 ≤ ∥x∥" ..
    ultimately show "0 ≤ p x"
      by (simp add: p_def zero_le_mult_iff)

    txt ‹‹p› is absolutely homogeneous:›

    show "p (a ⋅ x) = ∣a∣ * p x"
    proof -
      have "p (a ⋅ x) = ∥f∥🍋F * ∥a ⋅ x∥" by (simp only: p_def)
      also from x have "∥a ⋅ x∥ = ∣a∣ * ∥x∥" by (rule abs_homogenous)
      also have "∥f∥🍋F * (∣a∣ * ∥x∥) = ∣a∣ * (∥f∥🍋F * ∥x∥)" by simp
      also have "… = ∣a∣ * p x" by (simp only: p_def)
      finally show ?thesis .
    qed

    txt ‹Furthermore, ‹p› is subadditive:›

    show "p (x + y) ≤ p x + p y"
    proof -
      have "p (x + y) = ∥f∥🍋F * ∥x + y∥" by (simp only: p_def)
      also have a: "0 ≤ ∥f∥🍋F" by (rule ge_zero)
      from x y have "∥x + y∥ ≤ ∥x∥ + ∥y∥" ..
      with a have " ∥f∥🍋F * ∥x + y∥ ≤ ∥f∥🍋F * (∥x∥ + ∥y∥)"
        by (simp add: mult_left_mono)
      also have "… = ∥f∥🍋F * ∥x∥ + ∥f∥🍋F * ∥y∥" by (simp only: distrib_left)
      also have "… = p x + p y" by (simp only: p_def)
      finally show ?thesis .
    qed
  qed

  txt ‹‹f› is bounded by ‹p›.›

  have "∀x ∈ F. ∣f x∣ ≤ p x"
  proof
    fix x assume "x ∈ F"
    with ‹continuous F f norm› and linearform
    show "∣f x∣ ≤ p x"
      unfolding p_def by (rule normed_vectorspace_with_fn_norm.fn_norm_le_cong
        [OF normed_vectorspace_with_fn_norm.intro,
         OF F_norm, folded B_def fn_norm_def])
  qed

  txt ‹Using the fact that ‹p› is a seminorm and ‹f› is bounded by ‹p› we can
  apply the Hahn-Banach Theorem for real vector spaces. So ‹f› can be
  extended in a norm-preserving way to some function ‹g› on the whole vector
  space ‹E›.›

  with E FE linearform q obtain g where
      linearformE: "linearform E g"
    and a: "∀x ∈ F. g x = f x"
    and b: "∀x ∈ E. ∣g x∣ ≤ p x"
    by (rule abs_Hahn_Banach [elim_format]) iprover

  txt ‹We furthermore have to show that ‹g› is also continuous:›

  have g_cont: "continuous E g norm" using linearformE
  proof
    fix x assume "x ∈ E"
    with b show "∣g x∣ ≤ ∥f∥🍋F * ∥x∥"
      by (simp only: p_def)
  qed

  txt ‹To complete the proof, we show that ‹∥g∥ = ∥f∥›.›

  have "∥g∥🍋E = ∥f∥🍋F"
  proof (rule order_antisym)
    txt ‹
  First we show ‹∥g∥ ≤ ∥f∥›. The function norm ‹∥g∥› is defined as the
  smallest ‹c ∈ ℝ› such that
  \begin{center}
  \begin{tabular}{l}
  ‹∀x ∈ E. ∣g x∣ ≤ c ⋅ ∥x∥›
  \end{tabular}
  \end{center}
  🪙 Furthermore holds
  \begin{center}
  \begin{tabular}{l}
  ‹∀x ∈ E. ∣g x∣ ≤ ∥f∥ ⋅ ∥x∥›
  \end{tabular}
  \end{center}
  ›

    from g_cont _ ge_zero
    show "∥g∥🍋E ≤ ∥f∥🍋F"
    proof
      fix x assume "x ∈ E"
      with b show "∣g x∣ ≤ ∥f∥🍋F * ∥x∥"
        by (simp only: p_def)
    qed

    txt ‹The other direction is achieved by a similar argument.›

    show "∥f∥🍋F ≤ ∥g∥🍋E"
    proof (rule normed_vectorspace_with_fn_norm.fn_norm_least
        [OF normed_vectorspace_with_fn_norm.intro,
         OF F_norm, folded B_def fn_norm_def])
      fix x assume x: "x ∈ F"
      show "∣f x∣ ≤ ∥g∥🍋E * ∥x∥"
      proof -
        from a x have "g x = f x" ..
        then have "∣f x∣ = ∣g x∣" by (simp only:)
        also from g_cont have "… ≤ ∥g∥🍋E * ∥x∥"
        proof (rule fn_norm_le_cong [OF _ linearformE, folded B_def fn_norm_def])
          from FE x show "x ∈ E" ..
        qed
        finally show ?thesis .
      qed
    next
      show "0 ≤ ∥g∥🍋E"
        using g_cont by (rule fn_norm_ge_zero [of g, folded B_def fn_norm_def])
      show "continuous F f norm" by fact
    qed
  qed
  with linearformE a g_cont show ?thesis by blast
qed

end