Statistiques
| Révision :

root / CharacterizingPH / soundness.tex @ 257

Historique | Voir | Annoter | Télécharger (5,4 ko)

1 143 adas
\documentclass{amsart}
2 143 adas
\usepackage{amsthm}
3 143 adas
4 143 adas
5 143 adas
\input{macros}
6 143 adas
7 143 adas
\begin{document}
8 143 adas
	\section{Logical operations as programs}
9 143 adas
10 143 adas
	Any polynomial-time predicate can be conservatively added to BC with only safe inputs.
11 143 adas
	Thus all logical connectives $\neg, \vee, \wedge$ and equality testing $=$ can be implemented as programs with only safe arguments.
12 143 adas
13 143 adas
	\section{Soundness}
14 143 adas
	We show that any provably convergent function of $I\Sigma^{N}_i$ is in $\Box_i$.
15 143 adas
16 143 adas
	\subsection{Interpreting $\Pi_{i-1}$ formulae as programs}
17 143 adas
	Each $\Pi_{i-1}$ formula is directly interpreted, as in usual WFM, as a descriptive program evaluating whether it is true or not.
18 143 adas
	We define it by induction on formula structure:
19 143 adas
20 143 adas
21 143 adas
	\[
22 143 adas
	\begin{array}{rcl}
23 143 adas
	\wit{s,t}{s=t} (;x) & \dfn & =(;s,t) \\
24 143 adas
	\smallskip
25 143 adas
	\wit{t}{N(t)} (;x) & \dfn & =(;x,t)\\
26 143 adas
	\smallskip
27 144 adas
	\wit{t}{\Box N(t)} (u;) & \dfn & =(;u,t) \\
28 144 adas
	\smallskip
29 143 adas
	\wit{\vec t}{A\star B} (;x) & \dfn & \star (; \wit{\vec t}{A} (;x), \wit{\vec t}{B} (;x) ) \\
30 143 adas
	\smallskip
31 143 adas
	\wit{\vec t}{\exists x^N . A(x)} (;x) & \dfn & \begin{cases}
32 143 adas
	1 & \exists a^N . \wit{a , \vec t}{A(a)} (;x) = 1 \\
33 143 adas
	0 & \text{otherwise}
34 143 adas
	\end{cases} \\
35 143 adas
	\smallskip
36 143 adas
	\wit{\vec t}{\forall x^N . A(x)} (;x) & \dfn &
37 143 adas
	\begin{cases}
38 143 adas
	0 & \exists a^N. \wit{a, \vec t}{ A(a)} (;x) = 0 \\
39 143 adas
	1 & \text{otherwise}
40 143 adas
	\end{cases}
41 143 adas
	\end{array}
42 143 adas
	\]
43 143 adas
44 144 adas
	(Notice that, by free-cut elimination, $\Box$-formulae are always of the form $\Box N(t)$, so any composite formulae can be assumed to have only safe arguments.
45 144 adas
	We can also assume that they always occur on the LHS of a sequent.)
46 143 adas
	\subsection{Witness predicate for $\Sigma_i$ formulae}
47 143 adas
%	We give the witness predicate for $\Sigma_i$ formulae assumed to be in prenex form.
48 143 adas
	When $A,B$ are $\Sigma_{i}$ we define:
49 143 adas
		\[
50 143 adas
		\begin{array}{rcl}
51 143 adas
		\wit{\vec t}{A\star B} (;\vec x , \vec y) & \dfn & \star (; \wit{\vec t}{A} (;\vec x), \wit{\vec t}{B} (;\vec y) ) \\
52 143 adas
		\smallskip
53 143 adas
		\wit{\vec t}{\exists  x . A( x)} (; x, \vec x) & \dfn & \wit{a, \vec t}{A( a)}(;\vec x) \left[ x^N /  a\right]
54 143 adas
		\end{array}
55 143 adas
		\]
56 143 adas
	(Notice that this covers all $\Sigma_i$ formulae).
57 143 adas
58 143 adas
	\subsection{Witness extraction invariant}
59 143 adas
%	We will show the following result:
60 143 adas
61 143 adas
	We will show that, if there is a proof $\pi$ of $\Gamma \seqar \Delta$, then there are programs $\vec f$ such that:
62 143 adas
	\[
63 143 adas
	\wit{\vec t}{\bigwedge \Gamma} (; \vec x)
64 143 adas
	\quad \implies \quad
65 143 adas
	\wit{\vec t}{\bigvee \Delta } (; \vec f (; \vec x) )
66 143 adas
	\]
67 143 adas
68 143 adas
	\subsection{Defining the programs}
69 143 adas
	We define the extracted programs inductively on a free-cut free proof.
70 143 adas
71 143 adas
	\begin{itemize}
72 144 adas
	\item $\forall$ right case. Auxiliary formula $A$ must be $\Pi_{i-1}$ by free-cut elimination.
73 143 adas
		\[
74 143 adas
		\dfrac{N(a), \Gamma \seqar \Delta , A(a)}{\Gamma \seqar \Delta , \forall x^N. A(x)}
75 143 adas
		\]
76 144 adas
		By the inductive hypothesis we have functions $\vec f^\Delta , \vec f^A$ such that:
77 144 adas
		\[
78 144 adas
		\wit{a , \vec t}{N(a)} (;x)
79 144 adas
		\wedge
80 144 adas
		\wit{a , \vec t}{\bigwedge \Gamma} (\vec u; \vec x)
81 144 adas
		\implies
82 144 adas
		\wit{a , \vec t}{\bigvee \Delta}{ (; \vec f^\Delta (\vec u ; x , \vec x))}
83 144 adas
		\vee
84 144 adas
		\wit{a , \vec t}{ A} (; \vec f^A (\vec u ; x , \vec x) )
85 144 adas
		\]
86 144 adas
		Since $A$ is $\Pi_{i-1}$ we can forget about $\vec f^A$.
87 144 adas
		We define $\vec g^\Delta , \vec g^{\forall x^N . A}$ as follows:
88 144 adas
		\[
89 144 adas
		\begin{array}{rcl}
90 144 adas
		\vec g^{\forall x^N . A} (\vec u ; \vec x) & \dfn & 0 \\
91 144 adas
		\vec g^\Delta (\vec u ; \vec x) & \dfn &
92 144 adas
		\vec f^\Delta ( \vec u ;  \mu a^N . ( \wit{a , \vec t}{A} (;0) = 0 ) , \vec x )
93 144 adas
		\end{array}
94 144 adas
		\]
95 144 adas
		(Notice that the minimisation program has the same complexity as the witness predicate for $\forall x^N . A(x)$, i.e.\ $\Pi_{i-1}$.)
96 144 adas
	\item Contraction right case. Need to test the witness predicate using conditional. (No need for $\mu$s.).
97 144 adas
	Given such a step,
98 144 adas
	\[
99 144 adas
	\dfrac{\Gamma \seqar \Delta , A , A}{\Gamma \seqar \Delta , A}
100 144 adas
	\]
101 144 adas
	by the inductive hypothesis we have programs $\vec f^\Delta , \vec f^0 , \vec f^1$ such that:
102 144 adas
	\[
103 144 adas
	\wit{\vec t}{\bigwedge \Gamma} (\vec u ; \vec x)
104 144 adas
	\implies
105 144 adas
	\wit{\vec t}{\bigvee \Delta} (; \vec f^\Delta (\vec u ; \vec x) )
106 144 adas
	\vee
107 144 adas
	\wit{\vec t}{A} (;\vec f^0 (\vec u ; \vec x))
108 144 adas
	\vee
109 144 adas
	\wit{\vec t}{A} (; \vec f^1 (\vec u ; \vec x))
110 144 adas
	\]
111 144 adas
	Construct the programs $\vec g^\Delta, \vec g^A$ as follows:
112 144 adas
	\[
113 144 adas
	\begin{array}{rcl}
114 144 adas
	\vec g^\Delta (\vec u ; \vec x) & \dfn & \vec f^\Delta (\vec u ; \vec x) \\
115 144 adas
	\vec g^A (\vec u; \vec x) & \dfn &
116 144 adas
	\begin{cases}
117 144 adas
	\vec f^0 (\vec u ; \vec x) & \wit{\vec t}{A} (;\vec f^0(\vec u ; \vec x) =1 ) \\
118 144 adas
	\vec f^1 (\vec u ; \vec x) & \text{otherwise}
119 144 adas
	\end{cases}
120 144 adas
	\end{array}
121 144 adas
	\]
122 145 adas
	\item Induction. NB: Can assume no side-formulae on the right!!
123 145 adas
	\[
124 145 adas
	\dfrac{\Box N(a), \Gamma, A(a) \seqar A(s_i a)}{\Box N(s), \Gamma , A(0) \seqar A(s)}
125 145 adas
	\]
126 145 adas
	We need to keep track of all the terms in the witness predicate.
127 145 adas
	\[
128 145 adas
%	\begin{array}{rl}
129 145 adas
%	&\wit{a, t(a), t(s_i a), \vec t}{LHS} (u , \vec u ; \vec x , \vec y)
130 145 adas
%	\\
131 145 adas
%	\implies &
132 145 adas
%	\wit{a , t(a), t(s_i a), \vec t}{A(s_i a)} (; \vec f^A (u , \vec u ; \vec x , \vec y)  )
133 145 adas
%	\vee
134 145 adas
%	\wit{a , t(a), t(s_i a), \vec t}{\bigvee \Delta} (; \vec f^\Delta (u , \vec u ; \vec x , \vec y)  )
135 145 adas
%	\end{array}
136 145 adas
\wit{a, t(a), t(s_i a), \vec t}{LHS} (u , \vec u ; \vec x , \vec y)
137 145 adas
\implies
138 145 adas
\wit{a , t(a), t(s_i a), \vec t}{A(s_i a)} (; \vec f^i (u , \vec u ; \vec x , \vec y)  )
139 145 adas
	\]
140 145 adas
	Define $\vec g$ by safe recursion as follows:
141 145 adas
	\[
142 145 adas
	\begin{array}{rcl}
143 145 adas
	\vec g (0 , \vec u ; \vec x , \vec y) & \dfn & \vec y \\
144 145 adas
	\vec g (s_i u, \vec u ; \vec x , \vec y) & \dfn & \vec f^i ( u, \vec u; \vec x, \vec g (u, \vec u ; \vec x , \vec y) )
145 145 adas
	\end{array}
146 145 adas
	\]
147 143 adas
	\end{itemize}
148 143 adas
	\subsection{Proof of correctness}
149 143 adas
150 143 adas
\end{document}