| author | nipkow | 
| Sun, 15 Feb 2009 11:26:38 +0100 | |
| changeset 29920 | b95f5b8b93dd | 
| parent 27679 | 6392b92c3536 | 
| child 30198 | 922f944f03b2 | 
| permissions | -rw-r--r-- | 
| 19203 | 1  | 
(* Title: HOL/ZF/Games.thy  | 
2  | 
ID: $Id$  | 
|
3  | 
Author: Steven Obua  | 
|
4  | 
||
5  | 
An application of HOLZF: Partizan Games.  | 
|
6  | 
See "Partizan Games in Isabelle/HOLZF", available from http://www4.in.tum.de/~obua/partizan  | 
|
7  | 
*)  | 
|
8  | 
||
9  | 
theory Games  | 
|
10  | 
imports MainZF  | 
|
11  | 
begin  | 
|
12  | 
||
13  | 
constdefs  | 
|
14  | 
fixgames :: "ZF set \<Rightarrow> ZF set"  | 
|
15  | 
  "fixgames A \<equiv> { Opair l r | l r. explode l \<subseteq> A & explode r \<subseteq> A}"
 | 
|
16  | 
games_lfp :: "ZF set"  | 
|
17  | 
"games_lfp \<equiv> lfp fixgames"  | 
|
18  | 
games_gfp :: "ZF set"  | 
|
19  | 
"games_gfp \<equiv> gfp fixgames"  | 
|
20  | 
||
21  | 
lemma mono_fixgames: "mono (fixgames)"  | 
|
22  | 
apply (auto simp add: mono_def fixgames_def)  | 
|
23  | 
apply (rule_tac x=l in exI)  | 
|
24  | 
apply (rule_tac x=r in exI)  | 
|
25  | 
apply auto  | 
|
26  | 
done  | 
|
27  | 
||
28  | 
lemma games_lfp_unfold: "games_lfp = fixgames games_lfp"  | 
|
29  | 
by (auto simp add: def_lfp_unfold games_lfp_def mono_fixgames)  | 
|
30  | 
||
31  | 
lemma games_gfp_unfold: "games_gfp = fixgames games_gfp"  | 
|
32  | 
by (auto simp add: def_gfp_unfold games_gfp_def mono_fixgames)  | 
|
33  | 
||
34  | 
lemma games_lfp_nonempty: "Opair Empty Empty \<in> games_lfp"  | 
|
35  | 
proof -  | 
|
36  | 
  have "fixgames {} \<subseteq> games_lfp" 
 | 
|
37  | 
apply (subst games_lfp_unfold)  | 
|
38  | 
apply (simp add: mono_fixgames[simplified mono_def, rule_format])  | 
|
39  | 
done  | 
|
40  | 
  moreover have "fixgames {} = {Opair Empty Empty}"
 | 
|
41  | 
by (simp add: fixgames_def explode_Empty)  | 
|
42  | 
finally show ?thesis  | 
|
43  | 
by auto  | 
|
44  | 
qed  | 
|
45  | 
||
46  | 
constdefs  | 
|
47  | 
left_option :: "ZF \<Rightarrow> ZF \<Rightarrow> bool"  | 
|
48  | 
"left_option g opt \<equiv> (Elem opt (Fst g))"  | 
|
49  | 
right_option :: "ZF \<Rightarrow> ZF \<Rightarrow> bool"  | 
|
50  | 
"right_option g opt \<equiv> (Elem opt (Snd g))"  | 
|
51  | 
is_option_of :: "(ZF * ZF) set"  | 
|
52  | 
  "is_option_of \<equiv> { (opt, g) | opt g. g \<in> games_gfp \<and> (left_option g opt \<or> right_option g opt) }"
 | 
|
53  | 
||
54  | 
lemma games_lfp_subset_gfp: "games_lfp \<subseteq> games_gfp"  | 
|
55  | 
proof -  | 
|
56  | 
have "games_lfp \<subseteq> fixgames games_lfp"  | 
|
57  | 
by (simp add: games_lfp_unfold[symmetric])  | 
|
58  | 
then show ?thesis  | 
|
59  | 
by (simp add: games_gfp_def gfp_upperbound)  | 
|
60  | 
qed  | 
|
61  | 
||
62  | 
lemma games_option_stable:  | 
|
63  | 
assumes fixgames: "games = fixgames games"  | 
|
64  | 
and g: "g \<in> games"  | 
|
65  | 
and opt: "left_option g opt \<or> right_option g opt"  | 
|
66  | 
shows "opt \<in> games"  | 
|
67  | 
proof -  | 
|
68  | 
from g fixgames have "g \<in> fixgames games" by auto  | 
|
69  | 
then have "\<exists> l r. g = Opair l r \<and> explode l \<subseteq> games \<and> explode r \<subseteq> games"  | 
|
70  | 
by (simp add: fixgames_def)  | 
|
71  | 
then obtain l where "\<exists> r. g = Opair l r \<and> explode l \<subseteq> games \<and> explode r \<subseteq> games" ..  | 
|
72  | 
then obtain r where lr: "g = Opair l r \<and> explode l \<subseteq> games \<and> explode r \<subseteq> games" ..  | 
|
73  | 
with opt show ?thesis  | 
|
74  | 
by (auto intro: Elem_explode_in simp add: left_option_def right_option_def Fst Snd)  | 
|
75  | 
qed  | 
|
76  | 
||
77  | 
lemma option2elem: "(opt,g) \<in> is_option_of \<Longrightarrow> \<exists> u v. Elem opt u \<and> Elem u v \<and> Elem v g"  | 
|
78  | 
apply (simp add: is_option_of_def)  | 
|
79  | 
apply (subgoal_tac "(g \<in> games_gfp) = (g \<in> (fixgames games_gfp))")  | 
|
80  | 
prefer 2  | 
|
81  | 
apply (simp add: games_gfp_unfold[symmetric])  | 
|
82  | 
apply (auto simp add: fixgames_def left_option_def right_option_def Fst Snd)  | 
|
83  | 
apply (rule_tac x=l in exI, insert Elem_Opair_exists, blast)  | 
|
84  | 
apply (rule_tac x=r in exI, insert Elem_Opair_exists, blast)  | 
|
85  | 
done  | 
|
86  | 
||
87  | 
lemma is_option_of_subset_is_Elem_of: "is_option_of \<subseteq> (is_Elem_of^+)"  | 
|
88  | 
proof -  | 
|
89  | 
  {
 | 
|
90  | 
fix opt  | 
|
91  | 
fix g  | 
|
92  | 
assume "(opt, g) \<in> is_option_of"  | 
|
93  | 
then have "\<exists> u v. (opt, u) \<in> (is_Elem_of^+) \<and> (u,v) \<in> (is_Elem_of^+) \<and> (v,g) \<in> (is_Elem_of^+)"  | 
|
94  | 
apply -  | 
|
95  | 
apply (drule option2elem)  | 
|
96  | 
apply (auto simp add: r_into_trancl' is_Elem_of_def)  | 
|
97  | 
done  | 
|
98  | 
then have "(opt, g) \<in> (is_Elem_of^+)"  | 
|
99  | 
by (blast intro: trancl_into_rtrancl trancl_rtrancl_trancl)  | 
|
100  | 
}  | 
|
101  | 
then show ?thesis by auto  | 
|
102  | 
qed  | 
|
103  | 
||
104  | 
lemma wfzf_is_option_of: "wfzf is_option_of"  | 
|
105  | 
proof -  | 
|
106  | 
have "wfzf (is_Elem_of^+)" by (simp add: wfzf_trancl wfzf_is_Elem_of)  | 
|
107  | 
then show ?thesis  | 
|
108  | 
apply (rule wfzf_subset)  | 
|
109  | 
apply (rule is_option_of_subset_is_Elem_of)  | 
|
110  | 
done  | 
|
111  | 
qed  | 
|
112  | 
||
113  | 
lemma games_gfp_imp_lfp: "g \<in> games_gfp \<longrightarrow> g \<in> games_lfp"  | 
|
114  | 
proof -  | 
|
115  | 
have unfold_gfp: "\<And> x. x \<in> games_gfp \<Longrightarrow> x \<in> (fixgames games_gfp)"  | 
|
116  | 
by (simp add: games_gfp_unfold[symmetric])  | 
|
117  | 
have unfold_lfp: "\<And> x. (x \<in> games_lfp) = (x \<in> (fixgames games_lfp))"  | 
|
118  | 
by (simp add: games_lfp_unfold[symmetric])  | 
|
119  | 
show ?thesis  | 
|
120  | 
apply (rule wf_induct[OF wfzf_implies_wf[OF wfzf_is_option_of]])  | 
|
121  | 
apply (auto simp add: is_option_of_def)  | 
|
122  | 
apply (drule_tac unfold_gfp)  | 
|
123  | 
apply (simp add: fixgames_def)  | 
|
124  | 
apply (auto simp add: left_option_def Fst right_option_def Snd)  | 
|
125  | 
apply (subgoal_tac "explode l \<subseteq> games_lfp")  | 
|
126  | 
apply (subgoal_tac "explode r \<subseteq> games_lfp")  | 
|
127  | 
apply (subst unfold_lfp)  | 
|
128  | 
apply (auto simp add: fixgames_def)  | 
|
129  | 
apply (simp_all add: explode_Elem Elem_explode_in)  | 
|
130  | 
done  | 
|
131  | 
qed  | 
|
132  | 
||
133  | 
theorem games_lfp_eq_gfp: "games_lfp = games_gfp"  | 
|
134  | 
apply (auto simp add: games_gfp_imp_lfp)  | 
|
135  | 
apply (insert games_lfp_subset_gfp)  | 
|
136  | 
apply auto  | 
|
137  | 
done  | 
|
138  | 
||
139  | 
theorem unique_games: "(g = fixgames g) = (g = games_lfp)"  | 
|
140  | 
proof -  | 
|
141  | 
  {
 | 
|
142  | 
fix g  | 
|
143  | 
assume g: "g = fixgames g"  | 
|
144  | 
from g have "fixgames g \<subseteq> g" by auto  | 
|
145  | 
then have l:"games_lfp \<subseteq> g"  | 
|
146  | 
by (simp add: games_lfp_def lfp_lowerbound)  | 
|
147  | 
from g have "g \<subseteq> fixgames g" by auto  | 
|
148  | 
then have u:"g \<subseteq> games_gfp"  | 
|
149  | 
by (simp add: games_gfp_def gfp_upperbound)  | 
|
150  | 
from l u games_lfp_eq_gfp[symmetric] have "g = games_lfp"  | 
|
151  | 
by auto  | 
|
152  | 
}  | 
|
153  | 
note games = this  | 
|
154  | 
show ?thesis  | 
|
155  | 
apply (rule iff[rule_format])  | 
|
156  | 
apply (erule games)  | 
|
157  | 
apply (simp add: games_lfp_unfold[symmetric])  | 
|
158  | 
done  | 
|
159  | 
qed  | 
|
160  | 
||
161  | 
lemma games_lfp_option_stable:  | 
|
162  | 
assumes g: "g \<in> games_lfp"  | 
|
163  | 
and opt: "left_option g opt \<or> right_option g opt"  | 
|
164  | 
shows "opt \<in> games_lfp"  | 
|
165  | 
apply (rule games_option_stable[where g=g])  | 
|
166  | 
apply (simp add: games_lfp_unfold[symmetric])  | 
|
167  | 
apply (simp_all add: prems)  | 
|
168  | 
done  | 
|
169  | 
||
170  | 
lemma is_option_of_imp_games:  | 
|
171  | 
assumes hyp: "(opt, g) \<in> is_option_of"  | 
|
172  | 
shows "opt \<in> games_lfp \<and> g \<in> games_lfp"  | 
|
173  | 
proof -  | 
|
174  | 
from hyp have g_game: "g \<in> games_lfp"  | 
|
175  | 
by (simp add: is_option_of_def games_lfp_eq_gfp)  | 
|
176  | 
from hyp have "left_option g opt \<or> right_option g opt"  | 
|
177  | 
by (auto simp add: is_option_of_def)  | 
|
178  | 
with g_game games_lfp_option_stable[OF g_game, OF this] show ?thesis  | 
|
179  | 
by auto  | 
|
180  | 
qed  | 
|
181  | 
||
182  | 
lemma games_lfp_represent: "x \<in> games_lfp \<Longrightarrow> \<exists> l r. x = Opair l r"  | 
|
183  | 
apply (rule exI[where x="Fst x"])  | 
|
184  | 
apply (rule exI[where x="Snd x"])  | 
|
185  | 
apply (subgoal_tac "x \<in> (fixgames games_lfp)")  | 
|
186  | 
apply (simp add: fixgames_def)  | 
|
187  | 
apply (auto simp add: Fst Snd)  | 
|
188  | 
apply (simp add: games_lfp_unfold[symmetric])  | 
|
189  | 
done  | 
|
190  | 
||
191  | 
typedef game = games_lfp  | 
|
192  | 
by (blast intro: games_lfp_nonempty)  | 
|
193  | 
||
194  | 
constdefs  | 
|
195  | 
left_options :: "game \<Rightarrow> game zet"  | 
|
196  | 
"left_options g \<equiv> zimage Abs_game (zexplode (Fst (Rep_game g)))"  | 
|
197  | 
right_options :: "game \<Rightarrow> game zet"  | 
|
198  | 
"right_options g \<equiv> zimage Abs_game (zexplode (Snd (Rep_game g)))"  | 
|
199  | 
options :: "game \<Rightarrow> game zet"  | 
|
200  | 
"options g \<equiv> zunion (left_options g) (right_options g)"  | 
|
201  | 
Game :: "game zet \<Rightarrow> game zet \<Rightarrow> game"  | 
|
202  | 
"Game L R \<equiv> Abs_game (Opair (zimplode (zimage Rep_game L)) (zimplode (zimage Rep_game R)))"  | 
|
203  | 
||
204  | 
lemma Repl_Rep_game_Abs_game: "\<forall> e. Elem e z \<longrightarrow> e \<in> games_lfp \<Longrightarrow> Repl z (Rep_game o Abs_game) = z"  | 
|
205  | 
apply (subst Ext)  | 
|
206  | 
apply (simp add: Repl)  | 
|
207  | 
apply auto  | 
|
208  | 
apply (subst Abs_game_inverse, simp_all add: game_def)  | 
|
209  | 
apply (rule_tac x=za in exI)  | 
|
210  | 
apply (subst Abs_game_inverse, simp_all add: game_def)  | 
|
211  | 
done  | 
|
212  | 
||
213  | 
lemma game_split: "g = Game (left_options g) (right_options g)"  | 
|
214  | 
proof -  | 
|
215  | 
have "\<exists> l r. Rep_game g = Opair l r"  | 
|
216  | 
apply (insert Rep_game[of g])  | 
|
217  | 
apply (simp add: game_def games_lfp_represent)  | 
|
218  | 
done  | 
|
219  | 
then obtain l r where lr: "Rep_game g = Opair l r" by auto  | 
|
220  | 
have partizan_g: "Rep_game g \<in> games_lfp"  | 
|
221  | 
apply (insert Rep_game[of g])  | 
|
222  | 
apply (simp add: game_def)  | 
|
223  | 
done  | 
|
224  | 
have "\<forall> e. Elem e l \<longrightarrow> left_option (Rep_game g) e"  | 
|
225  | 
by (simp add: lr left_option_def Fst)  | 
|
226  | 
then have partizan_l: "\<forall> e. Elem e l \<longrightarrow> e \<in> games_lfp"  | 
|
227  | 
apply auto  | 
|
228  | 
apply (rule games_lfp_option_stable[where g="Rep_game g", OF partizan_g])  | 
|
229  | 
apply auto  | 
|
230  | 
done  | 
|
231  | 
have "\<forall> e. Elem e r \<longrightarrow> right_option (Rep_game g) e"  | 
|
232  | 
by (simp add: lr right_option_def Snd)  | 
|
233  | 
then have partizan_r: "\<forall> e. Elem e r \<longrightarrow> e \<in> games_lfp"  | 
|
234  | 
apply auto  | 
|
235  | 
apply (rule games_lfp_option_stable[where g="Rep_game g", OF partizan_g])  | 
|
236  | 
apply auto  | 
|
237  | 
done  | 
|
238  | 
let ?L = "zimage (Abs_game) (zexplode l)"  | 
|
239  | 
let ?R = "zimage (Abs_game) (zexplode r)"  | 
|
240  | 
have L:"?L = left_options g"  | 
|
241  | 
by (simp add: left_options_def lr Fst)  | 
|
242  | 
have R:"?R = right_options g"  | 
|
243  | 
by (simp add: right_options_def lr Snd)  | 
|
244  | 
have "g = Game ?L ?R"  | 
|
245  | 
apply (simp add: Game_def Rep_game_inject[symmetric] comp_zimage_eq zimage_zexplode_eq zimplode_zexplode)  | 
|
246  | 
apply (simp add: Repl_Rep_game_Abs_game partizan_l partizan_r)  | 
|
247  | 
apply (subst Abs_game_inverse)  | 
|
248  | 
apply (simp_all add: lr[symmetric] Rep_game)  | 
|
249  | 
done  | 
|
250  | 
then show ?thesis  | 
|
251  | 
by (simp add: L R)  | 
|
252  | 
qed  | 
|
253  | 
||
254  | 
lemma Opair_in_games_lfp:  | 
|
255  | 
assumes l: "explode l \<subseteq> games_lfp"  | 
|
256  | 
and r: "explode r \<subseteq> games_lfp"  | 
|
257  | 
shows "Opair l r \<in> games_lfp"  | 
|
258  | 
proof -  | 
|
259  | 
note f = unique_games[of games_lfp, simplified]  | 
|
260  | 
show ?thesis  | 
|
261  | 
apply (subst f)  | 
|
262  | 
apply (simp add: fixgames_def)  | 
|
263  | 
apply (rule exI[where x=l])  | 
|
264  | 
apply (rule exI[where x=r])  | 
|
265  | 
apply (auto simp add: l r)  | 
|
266  | 
done  | 
|
267  | 
qed  | 
|
268  | 
||
269  | 
lemma left_options[simp]: "left_options (Game l r) = l"  | 
|
270  | 
apply (simp add: left_options_def Game_def)  | 
|
271  | 
apply (subst Abs_game_inverse)  | 
|
272  | 
apply (simp add: game_def)  | 
|
273  | 
apply (rule Opair_in_games_lfp)  | 
|
274  | 
apply (auto simp add: explode_Elem Elem_zimplode zimage_iff Rep_game[simplified game_def])  | 
|
275  | 
apply (simp add: Fst zexplode_zimplode comp_zimage_eq)  | 
|
276  | 
apply (simp add: zet_ext_eq zimage_iff Rep_game_inverse)  | 
|
277  | 
done  | 
|
278  | 
||
279  | 
lemma right_options[simp]: "right_options (Game l r) = r"  | 
|
280  | 
apply (simp add: right_options_def Game_def)  | 
|
281  | 
apply (subst Abs_game_inverse)  | 
|
282  | 
apply (simp add: game_def)  | 
|
283  | 
apply (rule Opair_in_games_lfp)  | 
|
284  | 
apply (auto simp add: explode_Elem Elem_zimplode zimage_iff Rep_game[simplified game_def])  | 
|
285  | 
apply (simp add: Snd zexplode_zimplode comp_zimage_eq)  | 
|
286  | 
apply (simp add: zet_ext_eq zimage_iff Rep_game_inverse)  | 
|
287  | 
done  | 
|
288  | 
||
289  | 
lemma Game_ext: "(Game l1 r1 = Game l2 r2) = ((l1 = l2) \<and> (r1 = r2))"  | 
|
290  | 
apply auto  | 
|
291  | 
apply (subst left_options[where l=l1 and r=r1,symmetric])  | 
|
292  | 
apply (subst left_options[where l=l2 and r=r2,symmetric])  | 
|
293  | 
apply simp  | 
|
294  | 
apply (subst right_options[where l=l1 and r=r1,symmetric])  | 
|
295  | 
apply (subst right_options[where l=l2 and r=r2,symmetric])  | 
|
296  | 
apply simp  | 
|
297  | 
done  | 
|
298  | 
||
299  | 
constdefs  | 
|
300  | 
option_of :: "(game * game) set"  | 
|
301  | 
"option_of \<equiv> image (\<lambda> (option, g). (Abs_game option, Abs_game g)) is_option_of"  | 
|
302  | 
||
303  | 
lemma option_to_is_option_of: "((option, g) \<in> option_of) = ((Rep_game option, Rep_game g) \<in> is_option_of)"  | 
|
304  | 
apply (auto simp add: option_of_def)  | 
|
305  | 
apply (subst Abs_game_inverse)  | 
|
306  | 
apply (simp add: is_option_of_imp_games game_def)  | 
|
307  | 
apply (subst Abs_game_inverse)  | 
|
308  | 
apply (simp add: is_option_of_imp_games game_def)  | 
|
309  | 
apply simp  | 
|
310  | 
apply (auto simp add: Bex_def image_def)  | 
|
311  | 
apply (rule exI[where x="Rep_game option"])  | 
|
312  | 
apply (rule exI[where x="Rep_game g"])  | 
|
313  | 
apply (simp add: Rep_game_inverse)  | 
|
314  | 
done  | 
|
315  | 
||
316  | 
lemma wf_is_option_of: "wf is_option_of"  | 
|
317  | 
apply (rule wfzf_implies_wf)  | 
|
318  | 
apply (simp add: wfzf_is_option_of)  | 
|
319  | 
done  | 
|
320  | 
||
321  | 
lemma wf_option_of[recdef_wf, simp, intro]: "wf option_of"  | 
|
322  | 
proof -  | 
|
323  | 
have option_of: "option_of = inv_image is_option_of Rep_game"  | 
|
324  | 
apply (rule set_ext)  | 
|
325  | 
apply (case_tac "x")  | 
|
| 
19769
 
c40ce2de2020
Added [simp]-lemmas "in_inv_image" and "in_lex_prod" in the spirit of "in_measure".
 
krauss 
parents: 
19203 
diff
changeset
 | 
326  | 
by (simp add: option_to_is_option_of)  | 
| 19203 | 327  | 
show ?thesis  | 
328  | 
apply (simp add: option_of)  | 
|
329  | 
apply (auto intro: wf_inv_image wf_is_option_of)  | 
|
330  | 
done  | 
|
331  | 
qed  | 
|
332  | 
||
333  | 
lemma right_option_is_option[simp, intro]: "zin x (right_options g) \<Longrightarrow> zin x (options g)"  | 
|
334  | 
by (simp add: options_def zunion)  | 
|
335  | 
||
336  | 
lemma left_option_is_option[simp, intro]: "zin x (left_options g) \<Longrightarrow> zin x (options g)"  | 
|
337  | 
by (simp add: options_def zunion)  | 
|
338  | 
||
339  | 
lemma zin_options[simp, intro]: "zin x (options g) \<Longrightarrow> (x, g) \<in> option_of"  | 
|
340  | 
apply (simp add: options_def zunion left_options_def right_options_def option_of_def  | 
|
341  | 
image_def is_option_of_def zimage_iff zin_zexplode_eq)  | 
|
342  | 
apply (cases g)  | 
|
343  | 
apply (cases x)  | 
|
344  | 
apply (auto simp add: Abs_game_inverse games_lfp_eq_gfp[symmetric] game_def  | 
|
345  | 
right_option_def[symmetric] left_option_def[symmetric])  | 
|
346  | 
done  | 
|
347  | 
||
348  | 
consts  | 
|
349  | 
neg_game :: "game \<Rightarrow> game"  | 
|
350  | 
||
351  | 
recdef neg_game "option_of"  | 
|
352  | 
"neg_game g = Game (zimage neg_game (right_options g)) (zimage neg_game (left_options g))"  | 
|
353  | 
||
354  | 
declare neg_game.simps[simp del]  | 
|
355  | 
||
356  | 
lemma "neg_game (neg_game g) = g"  | 
|
357  | 
apply (induct g rule: neg_game.induct)  | 
|
358  | 
apply (subst neg_game.simps)+  | 
|
359  | 
apply (simp add: right_options left_options comp_zimage_eq)  | 
|
360  | 
apply (subgoal_tac "zimage (neg_game o neg_game) (left_options g) = left_options g")  | 
|
361  | 
apply (subgoal_tac "zimage (neg_game o neg_game) (right_options g) = right_options g")  | 
|
362  | 
apply (auto simp add: game_split[symmetric])  | 
|
363  | 
apply (auto simp add: zet_ext_eq zimage_iff)  | 
|
364  | 
done  | 
|
365  | 
||
366  | 
consts  | 
|
367  | 
ge_game :: "(game * game) \<Rightarrow> bool"  | 
|
368  | 
||
| 23771 | 369  | 
recdef ge_game "(gprod_2_1 option_of)"  | 
| 19203 | 370  | 
"ge_game (G, H) = (\<forall> x. if zin x (right_options G) then (  | 
371  | 
if zin x (left_options H) then \<not> (ge_game (H, x) \<or> (ge_game (x, G)))  | 
|
372  | 
else \<not> (ge_game (H, x)))  | 
|
373  | 
else (if zin x (left_options H) then \<not> (ge_game (x, G)) else True))"  | 
|
374  | 
(hints simp: gprod_2_1_def)  | 
|
375  | 
||
376  | 
declare ge_game.simps [simp del]  | 
|
377  | 
||
| 26304 | 378  | 
lemma ge_game_eq: "ge_game (G, H) = (\<forall> x. (zin x (right_options G) \<longrightarrow> \<not> ge_game (H, x)) \<and> (zin x (left_options H) \<longrightarrow> \<not> ge_game (x, G)))"  | 
| 19203 | 379  | 
apply (subst ge_game.simps[where G=G and H=H])  | 
380  | 
apply (auto)  | 
|
381  | 
done  | 
|
382  | 
||
383  | 
lemma ge_game_leftright_refl[rule_format]:  | 
|
384  | 
"\<forall> y. (zin y (right_options x) \<longrightarrow> \<not> ge_game (x, y)) \<and> (zin y (left_options x) \<longrightarrow> \<not> (ge_game (y, x))) \<and> ge_game (x, x)"  | 
|
385  | 
proof (induct x rule: wf_induct[OF wf_option_of])  | 
|
386  | 
case (1 "g")  | 
|
387  | 
  { 
 | 
|
388  | 
fix y  | 
|
389  | 
assume y: "zin y (right_options g)"  | 
|
390  | 
have "\<not> ge_game (g, y)"  | 
|
391  | 
proof -  | 
|
392  | 
have "(y, g) \<in> option_of" by (auto intro: y)  | 
|
393  | 
with 1 have "ge_game (y, y)" by auto  | 
|
| 26304 | 394  | 
with y show ?thesis by (subst ge_game_eq, auto)  | 
| 19203 | 395  | 
qed  | 
396  | 
}  | 
|
397  | 
note right = this  | 
|
398  | 
  { 
 | 
|
399  | 
fix y  | 
|
400  | 
assume y: "zin y (left_options g)"  | 
|
401  | 
have "\<not> ge_game (y, g)"  | 
|
402  | 
proof -  | 
|
403  | 
have "(y, g) \<in> option_of" by (auto intro: y)  | 
|
404  | 
with 1 have "ge_game (y, y)" by auto  | 
|
| 26304 | 405  | 
with y show ?thesis by (subst ge_game_eq, auto)  | 
| 19203 | 406  | 
qed  | 
407  | 
}  | 
|
408  | 
note left = this  | 
|
409  | 
from left right show ?case  | 
|
| 26304 | 410  | 
by (auto, subst ge_game_eq, auto)  | 
| 19203 | 411  | 
qed  | 
412  | 
||
413  | 
lemma ge_game_refl: "ge_game (x,x)" by (simp add: ge_game_leftright_refl)  | 
|
414  | 
||
415  | 
lemma "\<forall> y. (zin y (right_options x) \<longrightarrow> \<not> ge_game (x, y)) \<and> (zin y (left_options x) \<longrightarrow> \<not> (ge_game (y, x))) \<and> ge_game (x, x)"  | 
|
416  | 
proof (induct x rule: wf_induct[OF wf_option_of])  | 
|
417  | 
case (1 "g")  | 
|
418  | 
show ?case  | 
|
419  | 
proof (auto)  | 
|
420  | 
    {case (goal1 y) 
 | 
|
421  | 
from goal1 have "(y, g) \<in> option_of" by (auto)  | 
|
422  | 
with 1 have "ge_game (y, y)" by auto  | 
|
423  | 
with goal1 have "\<not> ge_game (g, y)"  | 
|
| 26304 | 424  | 
by (subst ge_game_eq, auto)  | 
| 19203 | 425  | 
with goal1 show ?case by auto}  | 
426  | 
note right = this  | 
|
427  | 
    {case (goal2 y)
 | 
|
428  | 
from goal2 have "(y, g) \<in> option_of" by (auto)  | 
|
429  | 
with 1 have "ge_game (y, y)" by auto  | 
|
430  | 
with goal2 have "\<not> ge_game (y, g)"  | 
|
| 26304 | 431  | 
by (subst ge_game_eq, auto)  | 
| 19203 | 432  | 
with goal2 show ?case by auto}  | 
433  | 
note left = this  | 
|
434  | 
    {case goal3
 | 
|
435  | 
from left right show ?case  | 
|
| 26304 | 436  | 
by (subst ge_game_eq, auto)  | 
| 19203 | 437  | 
}  | 
438  | 
qed  | 
|
439  | 
qed  | 
|
440  | 
||
441  | 
constdefs  | 
|
442  | 
eq_game :: "game \<Rightarrow> game \<Rightarrow> bool"  | 
|
443  | 
"eq_game G H \<equiv> ge_game (G, H) \<and> ge_game (H, G)"  | 
|
444  | 
||
445  | 
lemma eq_game_sym: "(eq_game G H) = (eq_game H G)"  | 
|
446  | 
by (auto simp add: eq_game_def)  | 
|
447  | 
||
448  | 
lemma eq_game_refl: "eq_game G G"  | 
|
449  | 
by (simp add: ge_game_refl eq_game_def)  | 
|
450  | 
||
| 23771 | 451  | 
lemma induct_game: "(\<And>x. \<forall>y. (y, x) \<in> lprod option_of \<longrightarrow> P y \<Longrightarrow> P x) \<Longrightarrow> P a"  | 
452  | 
by (erule wf_induct[OF wf_lprod[OF wf_option_of]])  | 
|
| 19203 | 453  | 
|
454  | 
lemma ge_game_trans:  | 
|
455  | 
assumes "ge_game (x, y)" "ge_game (y, z)"  | 
|
456  | 
shows "ge_game (x, z)"  | 
|
457  | 
proof -  | 
|
458  | 
  { 
 | 
|
459  | 
fix a  | 
|
460  | 
have "\<forall> x y z. a = [x,y,z] \<longrightarrow> ge_game (x,y) \<longrightarrow> ge_game (y,z) \<longrightarrow> ge_game (x, z)"  | 
|
461  | 
proof (induct a rule: induct_game)  | 
|
462  | 
case (1 a)  | 
|
463  | 
show ?case  | 
|
464  | 
proof (rule allI | rule impI)+  | 
|
465  | 
case (goal1 x y z)  | 
|
466  | 
show ?case  | 
|
467  | 
proof -  | 
|
468  | 
	  { fix xr
 | 
|
469  | 
assume xr:"zin xr (right_options x)"  | 
|
470  | 
assume "ge_game (z, xr)"  | 
|
471  | 
have "ge_game (y, xr)"  | 
|
472  | 
apply (rule 1[rule_format, where y="[y,z,xr]"])  | 
|
473  | 
apply (auto intro: xr lprod_3_1 simp add: prems)  | 
|
474  | 
done  | 
|
475  | 
moreover from xr have "\<not> ge_game (y, xr)"  | 
|
| 26304 | 476  | 
by (simp add: goal1(2)[simplified ge_game_eq[of x y], rule_format, of xr, simplified xr])  | 
| 19203 | 477  | 
ultimately have "False" by auto  | 
478  | 
}  | 
|
479  | 
note xr = this  | 
|
480  | 
	  { fix zl
 | 
|
481  | 
assume zl:"zin zl (left_options z)"  | 
|
482  | 
assume "ge_game (zl, x)"  | 
|
483  | 
have "ge_game (zl, y)"  | 
|
484  | 
apply (rule 1[rule_format, where y="[zl,x,y]"])  | 
|
485  | 
apply (auto intro: zl lprod_3_2 simp add: prems)  | 
|
486  | 
done  | 
|
487  | 
moreover from zl have "\<not> ge_game (zl, y)"  | 
|
| 26304 | 488  | 
by (simp add: goal1(3)[simplified ge_game_eq[of y z], rule_format, of zl, simplified zl])  | 
| 19203 | 489  | 
ultimately have "False" by auto  | 
490  | 
}  | 
|
491  | 
note zl = this  | 
|
492  | 
show ?thesis  | 
|
| 26304 | 493  | 
by (auto simp add: ge_game_eq[of x z] intro: xr zl)  | 
| 19203 | 494  | 
qed  | 
495  | 
qed  | 
|
496  | 
qed  | 
|
497  | 
}  | 
|
498  | 
note trans = this[of "[x, y, z]", simplified, rule_format]  | 
|
499  | 
with prems show ?thesis by blast  | 
|
500  | 
qed  | 
|
501  | 
||
502  | 
lemma eq_game_trans: "eq_game a b \<Longrightarrow> eq_game b c \<Longrightarrow> eq_game a c"  | 
|
503  | 
by (auto simp add: eq_game_def intro: ge_game_trans)  | 
|
504  | 
||
505  | 
constdefs  | 
|
506  | 
zero_game :: game  | 
|
507  | 
"zero_game \<equiv> Game zempty zempty"  | 
|
508  | 
||
509  | 
consts  | 
|
510  | 
plus_game :: "game * game \<Rightarrow> game"  | 
|
511  | 
||
| 23771 | 512  | 
recdef plus_game "gprod_2_2 option_of"  | 
| 19203 | 513  | 
"plus_game (G, H) = Game (zunion (zimage (\<lambda> g. plus_game (g, H)) (left_options G))  | 
514  | 
(zimage (\<lambda> h. plus_game (G, h)) (left_options H)))  | 
|
515  | 
(zunion (zimage (\<lambda> g. plus_game (g, H)) (right_options G))  | 
|
516  | 
(zimage (\<lambda> h. plus_game (G, h)) (right_options H)))"  | 
|
517  | 
(hints simp add: gprod_2_2_def)  | 
|
518  | 
||
519  | 
declare plus_game.simps[simp del]  | 
|
520  | 
||
521  | 
lemma plus_game_comm: "plus_game (G, H) = plus_game (H, G)"  | 
|
522  | 
proof (induct G H rule: plus_game.induct)  | 
|
523  | 
case (1 G H)  | 
|
524  | 
show ?case  | 
|
525  | 
by (auto simp add:  | 
|
526  | 
plus_game.simps[where G=G and H=H]  | 
|
527  | 
plus_game.simps[where G=H and H=G]  | 
|
528  | 
Game_ext zet_ext_eq zunion zimage_iff prems)  | 
|
529  | 
qed  | 
|
530  | 
||
531  | 
lemma game_ext_eq: "(G = H) = (left_options G = left_options H \<and> right_options G = right_options H)"  | 
|
532  | 
proof -  | 
|
533  | 
have "(G = H) = (Game (left_options G) (right_options G) = Game (left_options H) (right_options H))"  | 
|
534  | 
by (simp add: game_split[symmetric])  | 
|
535  | 
then show ?thesis by auto  | 
|
536  | 
qed  | 
|
537  | 
||
538  | 
lemma left_zero_game[simp]: "left_options (zero_game) = zempty"  | 
|
539  | 
by (simp add: zero_game_def)  | 
|
540  | 
||
541  | 
lemma right_zero_game[simp]: "right_options (zero_game) = zempty"  | 
|
542  | 
by (simp add: zero_game_def)  | 
|
543  | 
||
544  | 
lemma plus_game_zero_right[simp]: "plus_game (G, zero_game) = G"  | 
|
545  | 
proof -  | 
|
546  | 
  { 
 | 
|
547  | 
fix G H  | 
|
548  | 
have "H = zero_game \<longrightarrow> plus_game (G, H) = G "  | 
|
549  | 
proof (induct G H rule: plus_game.induct, rule impI)  | 
|
550  | 
case (goal1 G H)  | 
|
551  | 
note induct_hyp = prems[simplified goal1, simplified] and prems  | 
|
552  | 
show ?case  | 
|
553  | 
apply (simp only: plus_game.simps[where G=G and H=H])  | 
|
554  | 
apply (simp add: game_ext_eq prems)  | 
|
555  | 
apply (auto simp add:  | 
|
556  | 
zimage_cong[where f = "\<lambda> g. plus_game (g, zero_game)" and g = "id"]  | 
|
557  | 
induct_hyp)  | 
|
558  | 
done  | 
|
559  | 
qed  | 
|
560  | 
}  | 
|
561  | 
then show ?thesis by auto  | 
|
562  | 
qed  | 
|
563  | 
||
564  | 
lemma plus_game_zero_left: "plus_game (zero_game, G) = G"  | 
|
565  | 
by (simp add: plus_game_comm)  | 
|
566  | 
||
567  | 
lemma left_imp_options[simp]: "zin opt (left_options g) \<Longrightarrow> zin opt (options g)"  | 
|
568  | 
by (simp add: options_def zunion)  | 
|
569  | 
||
570  | 
lemma right_imp_options[simp]: "zin opt (right_options g) \<Longrightarrow> zin opt (options g)"  | 
|
571  | 
by (simp add: options_def zunion)  | 
|
572  | 
||
573  | 
lemma left_options_plus:  | 
|
574  | 
"left_options (plus_game (u, v)) = zunion (zimage (\<lambda>g. plus_game (g, v)) (left_options u)) (zimage (\<lambda>h. plus_game (u, h)) (left_options v))"  | 
|
575  | 
by (subst plus_game.simps, simp)  | 
|
576  | 
||
577  | 
lemma right_options_plus:  | 
|
578  | 
"right_options (plus_game (u, v)) = zunion (zimage (\<lambda>g. plus_game (g, v)) (right_options u)) (zimage (\<lambda>h. plus_game (u, h)) (right_options v))"  | 
|
579  | 
by (subst plus_game.simps, simp)  | 
|
580  | 
||
581  | 
lemma left_options_neg: "left_options (neg_game u) = zimage neg_game (right_options u)"  | 
|
582  | 
by (subst neg_game.simps, simp)  | 
|
583  | 
||
584  | 
lemma right_options_neg: "right_options (neg_game u) = zimage neg_game (left_options u)"  | 
|
585  | 
by (subst neg_game.simps, simp)  | 
|
586  | 
||
587  | 
lemma plus_game_assoc: "plus_game (plus_game (F, G), H) = plus_game (F, plus_game (G, H))"  | 
|
588  | 
proof -  | 
|
589  | 
  { 
 | 
|
590  | 
fix a  | 
|
591  | 
have "\<forall> F G H. a = [F, G, H] \<longrightarrow> plus_game (plus_game (F, G), H) = plus_game (F, plus_game (G, H))"  | 
|
592  | 
proof (induct a rule: induct_game, (rule impI | rule allI)+)  | 
|
593  | 
case (goal1 x F G H)  | 
|
594  | 
let ?L = "plus_game (plus_game (F, G), H)"  | 
|
595  | 
let ?R = "plus_game (F, plus_game (G, H))"  | 
|
596  | 
note options_plus = left_options_plus right_options_plus  | 
|
597  | 
      {
 | 
|
598  | 
fix opt  | 
|
599  | 
note hyp = goal1(1)[simplified goal1(2), rule_format]  | 
|
600  | 
have F: "zin opt (options F) \<Longrightarrow> plus_game (plus_game (opt, G), H) = plus_game (opt, plus_game (G, H))"  | 
|
601  | 
by (blast intro: hyp lprod_3_3)  | 
|
602  | 
have G: "zin opt (options G) \<Longrightarrow> plus_game (plus_game (F, opt), H) = plus_game (F, plus_game (opt, H))"  | 
|
603  | 
by (blast intro: hyp lprod_3_4)  | 
|
604  | 
have H: "zin opt (options H) \<Longrightarrow> plus_game (plus_game (F, G), opt) = plus_game (F, plus_game (G, opt))"  | 
|
605  | 
by (blast intro: hyp lprod_3_5)  | 
|
606  | 
note F and G and H  | 
|
607  | 
}  | 
|
608  | 
note induct_hyp = this  | 
|
609  | 
have "left_options ?L = left_options ?R \<and> right_options ?L = right_options ?R"  | 
|
610  | 
by (auto simp add:  | 
|
611  | 
plus_game.simps[where G="plus_game (F,G)" and H=H]  | 
|
612  | 
plus_game.simps[where G="F" and H="plus_game (G,H)"]  | 
|
613  | 
zet_ext_eq zunion zimage_iff options_plus  | 
|
614  | 
induct_hyp left_imp_options right_imp_options)  | 
|
615  | 
then show ?case  | 
|
616  | 
by (simp add: game_ext_eq)  | 
|
617  | 
qed  | 
|
618  | 
}  | 
|
619  | 
then show ?thesis by auto  | 
|
620  | 
qed  | 
|
621  | 
||
622  | 
lemma neg_plus_game: "neg_game (plus_game (G, H)) = plus_game(neg_game G, neg_game H)"  | 
|
623  | 
proof (induct G H rule: plus_game.induct)  | 
|
624  | 
case (1 G H)  | 
|
625  | 
note opt_ops =  | 
|
626  | 
left_options_plus right_options_plus  | 
|
627  | 
left_options_neg right_options_neg  | 
|
628  | 
show ?case  | 
|
629  | 
by (auto simp add: opt_ops  | 
|
630  | 
neg_game.simps[of "plus_game (G,H)"]  | 
|
631  | 
plus_game.simps[of "neg_game G" "neg_game H"]  | 
|
632  | 
Game_ext zet_ext_eq zunion zimage_iff prems)  | 
|
633  | 
qed  | 
|
634  | 
||
635  | 
lemma eq_game_plus_inverse: "eq_game (plus_game (x, neg_game x)) zero_game"  | 
|
636  | 
proof (induct x rule: wf_induct[OF wf_option_of])  | 
|
637  | 
case (goal1 x)  | 
|
638  | 
  { fix y
 | 
|
639  | 
assume "zin y (options x)"  | 
|
640  | 
then have "eq_game (plus_game (y, neg_game y)) zero_game"  | 
|
641  | 
by (auto simp add: prems)  | 
|
642  | 
}  | 
|
643  | 
note ihyp = this  | 
|
644  | 
  {
 | 
|
645  | 
fix y  | 
|
646  | 
assume y: "zin y (right_options x)"  | 
|
647  | 
have "\<not> (ge_game (zero_game, plus_game (y, neg_game x)))"  | 
|
648  | 
apply (subst ge_game.simps, simp)  | 
|
649  | 
apply (rule exI[where x="plus_game (y, neg_game y)"])  | 
|
650  | 
apply (auto simp add: ihyp[of y, simplified y right_imp_options eq_game_def])  | 
|
651  | 
apply (auto simp add: left_options_plus left_options_neg zunion zimage_iff intro: prems)  | 
|
652  | 
done  | 
|
653  | 
}  | 
|
654  | 
note case1 = this  | 
|
655  | 
  {
 | 
|
656  | 
fix y  | 
|
657  | 
assume y: "zin y (left_options x)"  | 
|
658  | 
have "\<not> (ge_game (zero_game, plus_game (x, neg_game y)))"  | 
|
659  | 
apply (subst ge_game.simps, simp)  | 
|
660  | 
apply (rule exI[where x="plus_game (y, neg_game y)"])  | 
|
661  | 
apply (auto simp add: ihyp[of y, simplified y left_imp_options eq_game_def])  | 
|
662  | 
apply (auto simp add: left_options_plus zunion zimage_iff intro: prems)  | 
|
663  | 
done  | 
|
664  | 
}  | 
|
665  | 
note case2 = this  | 
|
666  | 
  {
 | 
|
667  | 
fix y  | 
|
668  | 
assume y: "zin y (left_options x)"  | 
|
669  | 
have "\<not> (ge_game (plus_game (y, neg_game x), zero_game))"  | 
|
670  | 
apply (subst ge_game.simps, simp)  | 
|
671  | 
apply (rule exI[where x="plus_game (y, neg_game y)"])  | 
|
672  | 
apply (auto simp add: ihyp[of y, simplified y left_imp_options eq_game_def])  | 
|
673  | 
apply (auto simp add: right_options_plus right_options_neg zunion zimage_iff intro: prems)  | 
|
674  | 
done  | 
|
675  | 
}  | 
|
676  | 
note case3 = this  | 
|
677  | 
  {
 | 
|
678  | 
fix y  | 
|
679  | 
assume y: "zin y (right_options x)"  | 
|
680  | 
have "\<not> (ge_game (plus_game (x, neg_game y), zero_game))"  | 
|
681  | 
apply (subst ge_game.simps, simp)  | 
|
682  | 
apply (rule exI[where x="plus_game (y, neg_game y)"])  | 
|
683  | 
apply (auto simp add: ihyp[of y, simplified y right_imp_options eq_game_def])  | 
|
684  | 
apply (auto simp add: right_options_plus zunion zimage_iff intro: prems)  | 
|
685  | 
done  | 
|
686  | 
}  | 
|
687  | 
note case4 = this  | 
|
688  | 
show ?case  | 
|
689  | 
apply (simp add: eq_game_def)  | 
|
690  | 
apply (simp add: ge_game.simps[of "plus_game (x, neg_game x)" "zero_game"])  | 
|
691  | 
apply (simp add: ge_game.simps[of "zero_game" "plus_game (x, neg_game x)"])  | 
|
692  | 
apply (simp add: right_options_plus left_options_plus right_options_neg left_options_neg zunion zimage_iff)  | 
|
693  | 
apply (auto simp add: case1 case2 case3 case4)  | 
|
694  | 
done  | 
|
695  | 
qed  | 
|
696  | 
||
697  | 
lemma ge_plus_game_left: "ge_game (y,z) = ge_game(plus_game (x, y), plus_game (x, z))"  | 
|
698  | 
proof -  | 
|
699  | 
  { fix a
 | 
|
700  | 
have "\<forall> x y z. a = [x,y,z] \<longrightarrow> ge_game (y,z) = ge_game(plus_game (x, y), plus_game (x, z))"  | 
|
701  | 
proof (induct a rule: induct_game, (rule impI | rule allI)+)  | 
|
702  | 
case (goal1 a x y z)  | 
|
703  | 
note induct_hyp = goal1(1)[rule_format, simplified goal1(2)]  | 
|
704  | 
      { 
 | 
|
705  | 
assume hyp: "ge_game(plus_game (x, y), plus_game (x, z))"  | 
|
706  | 
have "ge_game (y, z)"  | 
|
707  | 
proof -  | 
|
708  | 
	  { fix yr
 | 
|
709  | 
assume yr: "zin yr (right_options y)"  | 
|
710  | 
from hyp have "\<not> (ge_game (plus_game (x, z), plus_game (x, yr)))"  | 
|
| 26304 | 711  | 
by (auto simp add: ge_game_eq[of "plus_game (x,y)" "plus_game(x,z)"]  | 
| 19203 | 712  | 
right_options_plus zunion zimage_iff intro: yr)  | 
713  | 
then have "\<not> (ge_game (z, yr))"  | 
|
714  | 
apply (subst induct_hyp[where y="[x, z, yr]", of "x" "z" "yr"])  | 
|
715  | 
apply (simp_all add: yr lprod_3_6)  | 
|
716  | 
done  | 
|
717  | 
}  | 
|
718  | 
note yr = this  | 
|
719  | 
	  { fix zl
 | 
|
720  | 
assume zl: "zin zl (left_options z)"  | 
|
721  | 
from hyp have "\<not> (ge_game (plus_game (x, zl), plus_game (x, y)))"  | 
|
| 26304 | 722  | 
by (auto simp add: ge_game_eq[of "plus_game (x,y)" "plus_game(x,z)"]  | 
| 19203 | 723  | 
left_options_plus zunion zimage_iff intro: zl)  | 
724  | 
then have "\<not> (ge_game (zl, y))"  | 
|
725  | 
apply (subst goal1(1)[rule_format, where y="[x, zl, y]", of "x" "zl" "y"])  | 
|
726  | 
apply (simp_all add: goal1(2) zl lprod_3_7)  | 
|
727  | 
done  | 
|
728  | 
}  | 
|
729  | 
note zl = this  | 
|
730  | 
show "ge_game (y, z)"  | 
|
| 26304 | 731  | 
apply (subst ge_game_eq)  | 
| 19203 | 732  | 
apply (auto simp add: yr zl)  | 
733  | 
done  | 
|
734  | 
qed  | 
|
735  | 
}  | 
|
736  | 
note right_imp_left = this  | 
|
737  | 
      {
 | 
|
738  | 
assume yz: "ge_game (y, z)"  | 
|
739  | 
	{
 | 
|
740  | 
fix x'  | 
|
741  | 
assume x': "zin x' (right_options x)"  | 
|
742  | 
assume hyp: "ge_game (plus_game (x, z), plus_game (x', y))"  | 
|
743  | 
then have n: "\<not> (ge_game (plus_game (x', y), plus_game (x', z)))"  | 
|
| 26304 | 744  | 
by (auto simp add: ge_game_eq[of "plus_game (x,z)" "plus_game (x', y)"]  | 
| 19203 | 745  | 
right_options_plus zunion zimage_iff intro: x')  | 
746  | 
have t: "ge_game (plus_game (x', y), plus_game (x', z))"  | 
|
747  | 
apply (subst induct_hyp[symmetric])  | 
|
748  | 
apply (auto intro: lprod_3_3 x' yz)  | 
|
749  | 
done  | 
|
750  | 
from n t have "False" by blast  | 
|
751  | 
}  | 
|
752  | 
note case1 = this  | 
|
753  | 
	{
 | 
|
754  | 
fix x'  | 
|
755  | 
assume x': "zin x' (left_options x)"  | 
|
756  | 
assume hyp: "ge_game (plus_game (x', z), plus_game (x, y))"  | 
|
757  | 
then have n: "\<not> (ge_game (plus_game (x', y), plus_game (x', z)))"  | 
|
| 26304 | 758  | 
by (auto simp add: ge_game_eq[of "plus_game (x',z)" "plus_game (x, y)"]  | 
| 19203 | 759  | 
left_options_plus zunion zimage_iff intro: x')  | 
760  | 
have t: "ge_game (plus_game (x', y), plus_game (x', z))"  | 
|
761  | 
apply (subst induct_hyp[symmetric])  | 
|
762  | 
apply (auto intro: lprod_3_3 x' yz)  | 
|
763  | 
done  | 
|
764  | 
from n t have "False" by blast  | 
|
765  | 
}  | 
|
766  | 
note case3 = this  | 
|
767  | 
	{
 | 
|
768  | 
fix y'  | 
|
769  | 
assume y': "zin y' (right_options y)"  | 
|
770  | 
assume hyp: "ge_game (plus_game(x, z), plus_game (x, y'))"  | 
|
771  | 
then have "ge_game(z, y')"  | 
|
772  | 
apply (subst induct_hyp[of "[x, z, y']" "x" "z" "y'"])  | 
|
773  | 
apply (auto simp add: hyp lprod_3_6 y')  | 
|
774  | 
done  | 
|
775  | 
with yz have "ge_game (y, y')"  | 
|
776  | 
by (blast intro: ge_game_trans)  | 
|
777  | 
with y' have "False" by (auto simp add: ge_game_leftright_refl)  | 
|
778  | 
}  | 
|
779  | 
note case2 = this  | 
|
780  | 
	{
 | 
|
781  | 
fix z'  | 
|
782  | 
assume z': "zin z' (left_options z)"  | 
|
783  | 
assume hyp: "ge_game (plus_game(x, z'), plus_game (x, y))"  | 
|
784  | 
then have "ge_game(z', y)"  | 
|
785  | 
apply (subst induct_hyp[of "[x, z', y]" "x" "z'" "y"])  | 
|
786  | 
apply (auto simp add: hyp lprod_3_7 z')  | 
|
787  | 
done  | 
|
788  | 
with yz have "ge_game (z', z)"  | 
|
789  | 
by (blast intro: ge_game_trans)  | 
|
790  | 
with z' have "False" by (auto simp add: ge_game_leftright_refl)  | 
|
791  | 
}  | 
|
792  | 
note case4 = this  | 
|
793  | 
have "ge_game(plus_game (x, y), plus_game (x, z))"  | 
|
| 26304 | 794  | 
apply (subst ge_game_eq)  | 
| 19203 | 795  | 
apply (auto simp add: right_options_plus left_options_plus zunion zimage_iff)  | 
796  | 
apply (auto intro: case1 case2 case3 case4)  | 
|
797  | 
done  | 
|
798  | 
}  | 
|
799  | 
note left_imp_right = this  | 
|
800  | 
show ?case by (auto intro: right_imp_left left_imp_right)  | 
|
801  | 
qed  | 
|
802  | 
}  | 
|
803  | 
note a = this[of "[x, y, z]"]  | 
|
804  | 
then show ?thesis by blast  | 
|
805  | 
qed  | 
|
806  | 
||
807  | 
lemma ge_plus_game_right: "ge_game (y,z) = ge_game(plus_game (y, x), plus_game (z, x))"  | 
|
808  | 
by (simp add: ge_plus_game_left plus_game_comm)  | 
|
809  | 
||
810  | 
lemma ge_neg_game: "ge_game (neg_game x, neg_game y) = ge_game (y, x)"  | 
|
811  | 
proof -  | 
|
812  | 
  { fix a
 | 
|
813  | 
have "\<forall> x y. a = [x, y] \<longrightarrow> ge_game (neg_game x, neg_game y) = ge_game (y, x)"  | 
|
814  | 
proof (induct a rule: induct_game, (rule impI | rule allI)+)  | 
|
815  | 
case (goal1 a x y)  | 
|
816  | 
note ihyp = goal1(1)[rule_format, simplified goal1(2)]  | 
|
817  | 
      { fix xl
 | 
|
818  | 
assume xl: "zin xl (left_options x)"  | 
|
819  | 
have "ge_game (neg_game y, neg_game xl) = ge_game (xl, y)"  | 
|
820  | 
apply (subst ihyp)  | 
|
821  | 
apply (auto simp add: lprod_2_1 xl)  | 
|
822  | 
done  | 
|
823  | 
}  | 
|
824  | 
note xl = this  | 
|
825  | 
      { fix yr
 | 
|
826  | 
assume yr: "zin yr (right_options y)"  | 
|
827  | 
have "ge_game (neg_game yr, neg_game x) = ge_game (x, yr)"  | 
|
828  | 
apply (subst ihyp)  | 
|
829  | 
apply (auto simp add: lprod_2_2 yr)  | 
|
830  | 
done  | 
|
831  | 
}  | 
|
832  | 
note yr = this  | 
|
833  | 
show ?case  | 
|
| 26304 | 834  | 
by (auto simp add: ge_game_eq[of "neg_game x" "neg_game y"] ge_game_eq[of "y" "x"]  | 
| 19203 | 835  | 
right_options_neg left_options_neg zimage_iff xl yr)  | 
836  | 
qed  | 
|
837  | 
}  | 
|
838  | 
note a = this[of "[x,y]"]  | 
|
839  | 
then show ?thesis by blast  | 
|
840  | 
qed  | 
|
841  | 
||
842  | 
constdefs  | 
|
843  | 
eq_game_rel :: "(game * game) set"  | 
|
844  | 
  "eq_game_rel \<equiv> { (p, q) . eq_game p q }"
 | 
|
845  | 
||
846  | 
typedef Pg = "UNIV//eq_game_rel"  | 
|
847  | 
by (auto simp add: quotient_def)  | 
|
848  | 
||
849  | 
lemma equiv_eq_game[simp]: "equiv UNIV eq_game_rel"  | 
|
850  | 
by (auto simp add: equiv_def refl_def sym_def trans_def eq_game_rel_def  | 
|
851  | 
eq_game_sym intro: eq_game_refl eq_game_trans)  | 
|
852  | 
||
| 25764 | 853  | 
instantiation Pg :: "{ord, zero, plus, minus, uminus}"
 | 
854  | 
begin  | 
|
855  | 
||
856  | 
definition  | 
|
857  | 
  Pg_zero_def: "0 = Abs_Pg (eq_game_rel `` {zero_game})"
 | 
|
858  | 
||
859  | 
definition  | 
|
860  | 
Pg_le_def: "G \<le> H \<longleftrightarrow> (\<exists> g h. g \<in> Rep_Pg G \<and> h \<in> Rep_Pg H \<and> ge_game (h, g))"  | 
|
861  | 
||
862  | 
definition  | 
|
863  | 
Pg_less_def: "G < H \<longleftrightarrow> G \<le> H \<and> G \<noteq> (H::Pg)"  | 
|
| 19203 | 864  | 
|
| 25764 | 865  | 
definition  | 
866  | 
  Pg_minus_def: "- G = contents (\<Union> g \<in> Rep_Pg G. {Abs_Pg (eq_game_rel `` {neg_game g})})"
 | 
|
867  | 
||
868  | 
definition  | 
|
869  | 
  Pg_plus_def: "G + H = contents (\<Union> g \<in> Rep_Pg G. \<Union> h \<in> Rep_Pg H. {Abs_Pg (eq_game_rel `` {plus_game (g,h)})})"
 | 
|
870  | 
||
871  | 
definition  | 
|
872  | 
Pg_diff_def: "G - H = G + (- (H::Pg))"  | 
|
873  | 
||
874  | 
instance ..  | 
|
875  | 
||
876  | 
end  | 
|
| 19203 | 877  | 
|
878  | 
lemma Rep_Abs_eq_Pg[simp]: "Rep_Pg (Abs_Pg (eq_game_rel `` {g})) = eq_game_rel `` {g}"
 | 
|
879  | 
apply (subst Abs_Pg_inverse)  | 
|
880  | 
apply (auto simp add: Pg_def quotient_def)  | 
|
881  | 
done  | 
|
882  | 
||
883  | 
lemma char_Pg_le[simp]: "(Abs_Pg (eq_game_rel `` {g}) \<le> Abs_Pg (eq_game_rel `` {h})) = (ge_game (h, g))"
 | 
|
884  | 
apply (simp add: Pg_le_def)  | 
|
885  | 
apply (auto simp add: eq_game_rel_def eq_game_def intro: ge_game_trans ge_game_refl)  | 
|
886  | 
done  | 
|
887  | 
||
888  | 
lemma char_Pg_eq[simp]: "(Abs_Pg (eq_game_rel `` {g}) = Abs_Pg (eq_game_rel `` {h})) = (eq_game g h)"
 | 
|
889  | 
apply (simp add: Rep_Pg_inject [symmetric])  | 
|
890  | 
apply (subst eq_equiv_class_iff[of UNIV])  | 
|
891  | 
apply (simp_all)  | 
|
892  | 
apply (simp add: eq_game_rel_def)  | 
|
893  | 
done  | 
|
894  | 
||
895  | 
lemma char_Pg_plus[simp]: "Abs_Pg (eq_game_rel `` {g}) + Abs_Pg (eq_game_rel `` {h}) = Abs_Pg (eq_game_rel `` {plus_game (g, h)})"
 | 
|
896  | 
proof -  | 
|
897  | 
  have "(\<lambda> g h. {Abs_Pg (eq_game_rel `` {plus_game (g, h)})}) respects2 eq_game_rel" 
 | 
|
898  | 
apply (simp add: congruent2_def)  | 
|
899  | 
apply (auto simp add: eq_game_rel_def eq_game_def)  | 
|
900  | 
apply (rule_tac y="plus_game (y1, z2)" in ge_game_trans)  | 
|
901  | 
apply (simp add: ge_plus_game_left[symmetric] ge_plus_game_right[symmetric])+  | 
|
902  | 
apply (rule_tac y="plus_game (z1, y2)" in ge_game_trans)  | 
|
903  | 
apply (simp add: ge_plus_game_left[symmetric] ge_plus_game_right[symmetric])+  | 
|
904  | 
done  | 
|
905  | 
then show ?thesis  | 
|
906  | 
by (simp add: Pg_plus_def UN_equiv_class2[OF equiv_eq_game equiv_eq_game])  | 
|
907  | 
qed  | 
|
908  | 
||
909  | 
lemma char_Pg_minus[simp]: "- Abs_Pg (eq_game_rel `` {g}) = Abs_Pg (eq_game_rel `` {neg_game g})"
 | 
|
910  | 
proof -  | 
|
911  | 
  have "(\<lambda> g. {Abs_Pg (eq_game_rel `` {neg_game g})}) respects eq_game_rel"
 | 
|
912  | 
apply (simp add: congruent_def)  | 
|
913  | 
apply (auto simp add: eq_game_rel_def eq_game_def ge_neg_game)  | 
|
914  | 
done  | 
|
915  | 
then show ?thesis  | 
|
916  | 
by (simp add: Pg_minus_def UN_equiv_class[OF equiv_eq_game])  | 
|
917  | 
qed  | 
|
918  | 
||
919  | 
lemma eq_Abs_Pg[rule_format, cases type: Pg]: "(\<forall> g. z = Abs_Pg (eq_game_rel `` {g}) \<longrightarrow> P) \<longrightarrow> P"
 | 
|
920  | 
apply (cases z, simp)  | 
|
921  | 
apply (simp add: Rep_Pg_inject[symmetric])  | 
|
922  | 
apply (subst Abs_Pg_inverse, simp)  | 
|
923  | 
apply (auto simp add: Pg_def quotient_def)  | 
|
924  | 
done  | 
|
925  | 
||
926  | 
instance Pg :: pordered_ab_group_add  | 
|
927  | 
proof  | 
|
928  | 
fix a b c :: Pg  | 
|
929  | 
show "a - b = a + (- b)" by (simp add: Pg_diff_def)  | 
|
930  | 
  {
 | 
|
931  | 
assume ab: "a \<le> b"  | 
|
932  | 
assume ba: "b \<le> a"  | 
|
933  | 
from ab ba show "a = b"  | 
|
934  | 
apply (cases a, cases b)  | 
|
935  | 
apply (simp add: eq_game_def)  | 
|
936  | 
done  | 
|
937  | 
}  | 
|
| 27679 | 938  | 
then show "(a < b) = (a \<le> b \<and> \<not> b \<le> a)" by (auto simp add: Pg_less_def)  | 
| 19203 | 939  | 
show "a + b = b + a"  | 
940  | 
apply (cases a, cases b)  | 
|
941  | 
apply (simp add: eq_game_def plus_game_comm)  | 
|
942  | 
done  | 
|
943  | 
show "a + b + c = a + (b + c)"  | 
|
944  | 
apply (cases a, cases b, cases c)  | 
|
945  | 
apply (simp add: eq_game_def plus_game_assoc)  | 
|
946  | 
done  | 
|
947  | 
show "0 + a = a"  | 
|
948  | 
apply (cases a)  | 
|
949  | 
apply (simp add: Pg_zero_def plus_game_zero_left)  | 
|
950  | 
done  | 
|
951  | 
show "- a + a = 0"  | 
|
952  | 
apply (cases a)  | 
|
953  | 
apply (simp add: Pg_zero_def eq_game_plus_inverse plus_game_comm)  | 
|
954  | 
done  | 
|
955  | 
show "a \<le> a"  | 
|
956  | 
apply (cases a)  | 
|
957  | 
apply (simp add: ge_game_refl)  | 
|
958  | 
done  | 
|
959  | 
  {
 | 
|
960  | 
assume ab: "a \<le> b"  | 
|
961  | 
assume bc: "b \<le> c"  | 
|
962  | 
from ab bc show "a \<le> c"  | 
|
963  | 
apply (cases a, cases b, cases c)  | 
|
964  | 
apply (auto intro: ge_game_trans)  | 
|
965  | 
done  | 
|
966  | 
}  | 
|
967  | 
  {
 | 
|
968  | 
assume ab: "a \<le> b"  | 
|
969  | 
from ab show "c + a \<le> c + b"  | 
|
970  | 
apply (cases a, cases b, cases c)  | 
|
971  | 
apply (simp add: ge_plus_game_left[symmetric])  | 
|
972  | 
done  | 
|
973  | 
}  | 
|
974  | 
qed  | 
|
975  | 
||
976  | 
end  |