1: \documentstyle[fleqn,12pt]{article}

2: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

3: \newtheorem{theorem}{Theorem}

4: \newtheorem{corollary}{Corollary}

5: \newtheorem{lemma}{Lemma}

6: \newtheorem{exercise}{Exercise}

7: \newtheorem{claim}{Claim}

8: \newtheorem{definition}{Definition}

9: \newtheorem{example}{Example}

10: \newtheorem{proposition}{Proposition}

11: \newtheorem{remark}{Remark}

12: \newenvironment{notation}{\noindent\bf Notation:\em\penalty1000}{}

13: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

14: \newcommand{\blackslug}{\mbox{\hskip 1pt \vrule width 4pt height 8pt 

15: depth 1.5pt \hskip 1pt}}

16: \newcommand{\QED}{\quad\blackslug\lower 8.5pt\null\par\noindent}

17: \newcommand{\proof}{\par\penalty-1000\vskip .5 pt\noindent{\bf Proof\/: }}

18: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

19: %\newcommand{\SEP}{\makebox[0in]{\rule{.5mm}{4.5mm}}}

20: \newcommand{\ru}{\rule[-0.4mm]{.1mm}{3mm}}

21: \newcommand{\nni}{\ru\hspace{-3.5pt}}

22: \newcommand{\sni}{\ru\hspace{-1pt}}

23: \newcommand{\pre}{\hspace{0.28em}}

24: \newcommand{\post}{\hspace{0.1em}}

25: \newcommand{\NIm}{\pre\nni\sim}

26: \newcommand{\NI}{\mbox{$\: \nni\sim$}}

27: \newcommand{\notNIm}{\pre\nni\not\sim}

28: \newcommand{\notNI}{\mbox{ $\nni\not\sim$ }}

29: \newcommand{\NIVm}{\pre\nni\sim_V}

30: \newcommand{\NIV}{\mbox{ $\nni\sim_V$ }}

31: \newcommand{\notNIVm}{\pre\sni{\not\sim}_V\post}

32: \newcommand{\notNIV}{\mbox{ $\sni{\not\sim}_V$ }}

33: \newcommand{\NIWm}{\pre\nni\sim_W}

34: \newcommand{\NIW}{\mbox{ $\nni\sim_W$ }}

35: \newcommand{\NIWp}{\mbox{ $\nni\sim_{W'}$ }}

36: \newcommand{\notNIWm}{\pre\sni{\not\sim}_W\post}

37: \newcommand{\notNIW}{\mbox{ $\sni{\not\sim}_W$ }}

38: \newcommand{\eem}{\hspace{0.8mm}\rule[-1mm]{.1mm}{4mm}\hspace{-4pt}}

39: \newcommand{\EM}{\eem\equiv}

40: \newcommand{\notEM}{\eem\not\equiv}

41: \newcommand{\R}{\cal R}

42: \newcommand{\notR}{\not {\hspace{-1.5mm}{\cal R}}}

43: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

44: \newcommand{\bK}{{\bf K}}

45: \newcommand{\bKp}{${\bf K}^p$}

46: \newcommand{\oK}{$\overline {\bf K}$}

47: \newcommand{\bM}{{\bf M}}

48: \newcommand{\bP}{{\bf P}}

49: \newcommand{\ga}{\mbox{$\alpha$}}

50: \newcommand{\gb}{\mbox{$\beta$}}

51: \newcommand{\gc}{\mbox{$\gamma$}}

52: \newcommand{\gd}{\mbox{$\delta$}}

53: \newcommand{\gep}{\mbox{$\varepsilon$}}

54: \newcommand{\gf}{\mbox{$\zeta$}}

55: \newcommand{\cA}{\mbox{${\cal A}$}}

56: \newcommand{\cB}{\mbox{${\cal B}$}}

57: \newcommand{\cC}{\mbox{${\cal C}$}}

58: \newcommand{\cE}{\mbox{${\cal E}$}}

59: \newcommand{\cF}{\mbox{${\cal F}$}}

60: \newcommand{\cK}{\mbox{${\cal K}$}}

61: \newcommand{\cL}{\mbox{${\cal L}$}}

62: \newcommand{\cM}{\mbox{${\cal M}$}}

63: \newcommand{\cR}{\mbox{${\cal R}$}}

64: \newcommand{\cS}{\mbox{${\cal S}$}}

65: \newcommand{\cT}{\mbox{${\cal T}$}}

66: \newcommand{\cU}{\mbox{${\cal U}$}}

67: \newcommand{\ab}{\mbox{\ga \NI \gb}}

68: \newcommand{\cd}{\mbox{\gc \NI \gd}}

69: \newcommand{\ef}{\mbox{\gep \NI \gf}}

70: \newcommand{\xe}{\mbox{$\xi$ \NI $\eta$}}  

71: \newcommand{\pht}{\mbox{$\varphi$ \NI $\theta$}}

72: \newcommand{\rt}{\mbox{$\rho$ \NI $\tau$}}

73: \newcommand{\Cn}{\mbox{${\cal C}n$}}

74: \newcommand{\Cs}{\mbox{${\cal C}_{*}$}}

75: \newcommand{\CF}{\mbox{${\cal C}_{\cal F}$}}

76: \newcommand{\CG}{\mbox{${\cal C}_{\sim}$}}

77: \newcommand{\CW}{\mbox{${\cal C}_{W}$}}

78: \newcommand{\Pf}{\mbox{${\cal P}_{f}$}}

79: \newcommand{\leC}{\mbox{${\preceq_{\cC}}$}}

80: \newcommand{\leF}{\mbox{${\preceq_{\cF}}$}}

81: \newcommand{\lC}{\mbox{${\prec_{\cC}}$}}

82: \newcommand{\notlC}{\mbox{$\not \! \! \lC$}}

83: \newcommand{\lF}{\mbox{${\prec_{\cF}}$}}

84: \newcommand{\Z}{\mbox{$Z_{\cF}$}}

85: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

86: \newcommand{\ra}{\rightarrow}

87: \newcommand{\Ra}{\Rightarrow}

88: \newcommand{\eqdef}{\stackrel{\rm def}{=}}

89: \newcommand{\absv}[1]{\mid #1 \mid}

90: \newcommand{\vstar}{\mbox{$V\sstar_{\infty}$}}

91: \newcommand{\sumstar}{\mbox{$\sum\sstar$}}

92: \newcommand{\tilh}{\mbox{$\tilde{h}$}}

93: \newcommand{\tilep}{\mbox{$\tilde{\varepsilon}$}}

94: \newcommand{\tilf}{\mbox{$\tilde{f}$}}

95: \newcommand{\gahat}{\mbox{$\hat{\ga}$}}

96: \newcommand{\gafalse}{\mbox{$\ga\NI{\bf false}$}}

97: \newcommand{\sstar}{^{*}}

98: \newcommand{\calR}{\mbox{${\cal R}\sstar$}}

99: \newcommand{\subseteqf}{\mbox{$\subseteq_{f}$}}

100: \newcommand{\rev}[3]{\mbox{$#1$$#2$$#3$}}

101: \newcommand{\bKs}[1]{\mbox{${\bf K}$$\!{\bf *}$${\bf #1}$}}

102: \hyphenation{mono-tonicity Mono-tonicity mono-tonic Mono-tonic Mo-notonicity

103: mo-notonicity monoto-nicity Monoto-nicity}

104: 

105: \begin{document}

106: %Change the Plan:

107: %	1 Introduction: 1.1 Belief Revision

108: %			1.2 AGM Postulates

109: %			1.3 An additional postulate?

110: %			1.4 Plan of this paper

111: %	2 Previously proposed additional postulates

112: %			2.1 Katsuno-Mendelzon

113: %			2.2 Darwiche and Pearl

114: %	3 The minimal influence postulate

115: %			3.1 The postulate: K*9 and K*9', justification

116: %			3.2 First consequences: iterated revisions

117: %			3.3 K*9 and KM

118: %			3.4 K*9 and C1, C3 and C4

119: %	4 Representation of revisions by rational relations

120: %			4.1 Representation result

121: %			4.2 Conservative extension result

122: %			4.3 Ontology of revising by a fixed rational relation

123: %	5 Conclusion and future work

124: %	6 Acknowledgments

125: 	

126: \bibliographystyle{plain}

127: 

128: \title{Belief revision and rational inference

129: \thanks{This work was partially supported 

130: by the Jean and Helene Alfassa fund for 

131: research in Artificial Intelligence}

132: }

133: \author{Michael Freund \\ D\'{e}partement de Math\'{e}matiques, \\

134: Universit\'{e} d'Orl\'{e}ans, 45067 Orl\'{e}ans, C\'{e}dex 2 (France) 

135: \\ e-mail: freund@centre.univ-orleans.fr \and

136: Daniel Lehmann \\ Institute of Computer Science, \\

137: Hebrew University, Jerusalem 91904 (Israel) \\

138: e-mail: lehmann@cs.huji.ac.il

139: }

140: \date{August 12, 1994}

141: \maketitle

142: 

143: \begin{abstract}

144: The (extended) AGM postulates for belief revision seem to deal with

145: the revision of a given theory $K$ by an arbitrary formula $\varphi$,

146: but not to constrain the revisions of two different theories

147: by the same formula. 

148: A new postulate is proposed and compared with other similar postulates that

149: have been proposed in the literature.

150: The AGM revisions that satisfy this new postulate stand in one-to-one 

151: correspondence with the rational, consistency-preserving relations.

152: This correspondence is described explicitly.

153: Two viewpoints on iterative revisions are distinguished and discussed.

154: \end{abstract}

155: \section{Introduction}

156: \label{sec:intro}

157: \subsection{Belief Revision}

158: \label{subsec:beliefrev}

159: Belief revision is the study of the way an agent revises or should revise

160: its beliefs when acquiring new information.

161: A popular framework for this study has been initiated by Alchour\'{o}n,

162: Makinson and G\"{a}rdenfors in ~\cite{AM:82,AM:85,AGM:85}, in which a set

163: of rationality postulates for theory revision were put forward.

164: It assumes beliefs are sets of formulas closed under logical consequence

165: (i.e., theories) and that new information is a formula. The present work fits 

166: squarely in this framework.

167: An up-to-date description of this very active research area may be found 

168: in~\cite{Gard:beliefrev}. 

169: The basic assumptions underlying AGM's viewpoint may be summarized as follows.

170: \begin{itemize}

171: \item The agent holds beliefs: those beliefs constitute some logical theory.

172: There is no additional structure to the agent's theory. No beliefs are

173: stronger than others.

174: \item The agent's knowledge is changing but the world is not.

175: Keller and Winslett~\cite{KelWin:85} noticed that {\em updating}

176: an agent's knowledge about a changing world is a process that is altogether

177: different from theory revision.

178: Katsuno and Mendelzon~\cite{KatMend:92} proposed different postulates for 

179: belief update.

180: \item When a new piece of information $\varphi$ is presented to an agent that 

181: holds a theory $K$ with which $\varphi$ is inconsistent, 

182: the agent will give precedence to the information $\varphi$, 

183: over the theory $K$. The theory $K$ is considered as 

184: less important or less reliable

185: than the new information $\varphi$ with which it conflicts:

186: the agent revises a weakly held theory with a piece of more reliable

187: information.

188: This policy is obviously not the right one in every situation,

189: but AGM assume the agent must have some good reason to prefer $\varphi$ 

190: to $K$. As noticed just above, this reason is {\em not} that $\varphi$ 

191: describes some change in a world the previous state of which was

192: described by $K$.

193: \item When revising a theory $K$ with some formula $\varphi$, the agent

194: will try keep as much as possible of its previously held beliefs: $K$.

195: In particular, if $\varphi$ is consistent with $K$, the agent will keep

196: all of $K$ in its new belief set. This assumption will be called the

197: {\em maximal retention} assumption.

198: \end{itemize}

199: AGM did not describe any specific method to revise theories that would be

200: the {\em right} way of revising under the assumptions above.

201: They thought that there is probably no unique best way of revising.

202: Instead, they proposed a set of rationality postulates

203: they claimed any reasonable method for belief revision should satisfy.

204: The original postulates were extended in~\cite{Gard:intro}.

205: This work accepts, under the assumptions above, the extended AGM postulates.

206: These postulates are meant to express formally the basic assumptions described

207: above.

208: These postulates will be presented now and their adequacy in expressing those

209: basic assumptions will be discussed in section~\ref{subsec:additional}.

210: \subsection{The AGM postulates}

211: \label{subsec:AGM}

212: The AGM postulates for belief revision are listed below: 

213: \Cn\ denotes logical consequence and, for an arbitrary theory $K$

214: (i.e., $K$ is a set of formulas closed under logical consequence), 

215: \rev{K}{\!*}{\varphi} denotes the result of revising $K$ by $\varphi$.

216: The inconsistent theory, i.e., the full language, is denoted by $K_{\perp}$.

217: It is a legal argument for~$*$.

218: This paper could have been developed in the more demanding framework,

219: where the first argument of a revision, $K$, must be a consistent theory:

220: the results are essentially the same.

221: 

222: For a justification of these postulates, in addition to the papers cited above,

223: the reader may also 

224: consider~\cite{Mak:giveup,GardMak:88,Mak:recovery,GarMak:92,Rott:92}.

225: Since the notion of a revision is considered by

226: philosophers as not being primitive, but derived from that of

227: {\em theory contraction}, the justification of the postulates below is

228: generally given by justifying corresponding postulates for contractions

229: and translating back and forth with the help of the Levi and Gardenf\"{o}rs

230: identities.

231: Not all specialists are in complete agreement concerning the philosophical

232: underpinnings of theory revision, the status of the Levi and Gardenf\"{o}rs

233: identities, or of the postulates below.

234: The present work does not attack those deep questions directly, but

235: presents technical results concerning the postulates that, in turn,

236: help understand what are the processes that the postulates may reasonably be 

237: thought of as modelling and what are those they do not model satisfactorily.

238: For any theory $K$ and formulas $\varphi$ and $\psi$:

239: \[

240: \bKs{1} \ \ \ \   \rev{K}{\!*}{\varphi} {\rm \ is \ a \ theory.}

241: \]

242: \[

243: \bKs{2} \ \ \ \  \varphi \in \rev{K}{\!*}{\varphi}.

244: \]

245: \[

246: \bKs{3} \ \ \ \   \rev{K}{\!*}{\varphi} \subseteq \Cn(K , \varphi).

247: \]

248: \[

249: \bKs{4} \ \ \ \ {\rm If \ } \neg\varphi \not \in K, 

250: {\rm \ then \ } \Cn(K , \varphi) \subseteq \rev{K}{\!*}{\varphi}.

251: \]

252: \[

253: \bKs{5} \ \ \ \ {\rm If \ } \rev{K}{\!*}{\varphi} {\rm \ is \ inconsistent, \  then \ } 

254: \varphi {\rm \ is \ a \ logical \ contradiction.}

255: \]

256: \[

257: \bKs{6} \ \ \ \ {\rm If \ } \models \varphi \leftrightarrow \psi , 

258: {\rm \ then \ } \rev{K}{\!*}{\varphi} = \rev{K}{\!*}{\psi}.

259: \]

260: \[

261: \bKs{7} \ \ \ \  \rev{K}{\!*}{(\varphi \wedge \psi)} 

262: 	\subseteq \Cn( \rev{K}{\!*}{\varphi} , \psi).

263: \]

264: \[

265: \bKs{8} \ \ \ \ {\rm If \ } \neg\psi \not \in \rev{K}{\!*}{\varphi} , 

266: {\rm \ then \ } \Cn(\rev{K}{\!*}{\varphi} , \psi) \subseteq \rev{K}{\!*}{(\varphi \wedge \psi)}

267: \]

268: \noindent

269: The reader may notice that, in each of the postulates in which the revision

270: symbol ($*$) appears more than once (\bKs{6}, \bKs{7} and 

271: \bKs{8}),

272: the theory that appears on the left of the different occurrences of the 

273: revision symbol is the same.

274: We take this as an indication that those postulates do not say much on the

275: way the revised theory \mbox{$\rev{K}{\!*}{\varphi}$} depends on $K$.

276: We shall come back to this point in section~\ref{subsec:additional}.

277: We shall call a revision \rev{K}{\!*}{\varphi} a {\em mild} revision,

278: when \mbox{$\neg \varphi \not \in K$} and a {\em severe} revision,

279: if \mbox{$\neg \varphi \in K$}.

280: 

281: The AGM postulates have been found to be intimately related to

282: rational consequence relations (see~\cite{MakGar:89}).

283: The translation is precisely described by the following Theorem.

284: \begin{theorem}

285: \label{the:Kfixed}

286: If $*$ is a revision operation that satisfies the AGM postulates,

287: \bKs{1}--\bKs{8}, and $K$ is a theory,

288: the relation $\NIm^{K,*}$ defined by

289: \begin{equation}

290: \label{eq:K*}

291: \ga \NIm^{K,*} \gb {\rm \ iff \ }  \gb \in \rev{K}{\!*}{\ga}

292: \end{equation}

293: is rational and consistency-preserving.

294: \end{theorem}

295: \proof

296: We just remind the reader that consistency preservation is the following 

297: property:

298: if \mbox{$\ga \NIm {\bf false}$}, then $\ga$ is a logical contradiction.

299: The definition of all the other properties mentioned in this proof may 

300: be found in~\cite{LMAI:92}.

301: Suppose $K$ is a theory and $*$ satisfies the AGM postulates.

302: The consequence relation $\NIm^{K,*}$ satisfies Reflexivity by \bKs{2},

303: Left Logical Equivalence by \bKs{6}, Right Weakening and And 

304: by \bKs{1},

305: Conditionalization (i.e., (S)) by \bKs{7}, and Rational Monotonicity by

306: \bKs{8}.

307: It is consistency-preserving by \bKs{5}.

308: Any such relation is rational.

309: \QED

310: Notice that postulates \bKs{3} and \bKs{4} are not used in the proof

311: of Theorem~\ref{the:Kfixed}.

312: This point will be discussed after we have proved 

313: Theorem~\ref{the:representation}.

314: 

315: One may expect the converse to Theorem~\ref{the:Kfixed} to hold:

316: if \NI\ is an arbitrary, rational and consistency-preserving relation,

317: then, there is a theory $K$ and a revision $*$ satisfying

318: \bKs{1}--\bKs{8} such that \mbox{$\alpha$ \NI $\beta$} iff 

319: \mbox{$\beta \in \rev{K}{\!*}{\alpha}$}.

320: This result holds.

321: It is not difficult to present a direct proof at this stage.

322: An indirect proof will be preferred and presented in 

323: Theorem~\ref{the:converse}. 

324: \subsection{Critique of the AGM postulates}

325: \label{subsec:additional}

326: The AGM postulates have been discussed in the literature, in particular

327: in~\cite{KatMend:91}, \cite{KatMend:92} and~\cite{DarwPearl:TARK}.

328: Some authors reject some of the postulates, in particular \bKs{8},

329: the same and others propose to extend the AGM postulates by additional 

330: postulates.

331: The main line of critique of the AGM postulates seems to be that they

332: do not constrain the revisions enough: some revisions allowed by the

333: postulates do not seem reasonable.

334: There are three points, in particular, on which it seems that the

335: postulates do not say enough, or, more precisely, one could feel that

336: additional acceptable postulates may be proposed.

337: \begin{itemize}

338: \item The postulates do not seem to enforce the principle of maximal retention

339: when the formula $\varphi$ is inconsistent with the theory $K$.

340: \item They do not seem to impose enough constraints on

341: the revision of two different theories by the same formula, or by related

342: formulas.

343: \item They do not seem to enforce enough constraints on iterated revisions, 

344: e.g., concerning the relation between $K$, $\varphi$, $\psi$ and 

345: \rev{(\rev{K}{\!*}{\varphi})}{*}{\psi}.

346: \end{itemize}

347: The proposals to drop some of the AGM postulates seem to stem from

348: the desire to add some other postulates with which they are inconsistent,

349: more than from a direct critique of the postulates in question.

350: The first point raised above has been the reason for the interest in

351: maxichoice contractions manifested by AGM, but no suitable additional

352: postulate has been proposed to deal with it.

353: The second point has been very aptly discussed in~\cite{MakGar:89}.

354: The authors write: ``Revision is an operation of two

355: arguments, forming $A*x$ out of theory $A$ and proposition $x$.

356: On the other hand, nonmonotonic inference conceived as an operation

357: $\cC(x)$ defined as \mbox{$\{ y \mid x \NIm y \}$} 

358: is a function of only one argument $x$. For this reason the logic of theory

359: change is potentially more general than the logic of nonmonotonic inference,

360: in that it allows the possibility of {\em variation} in the other

361: argument $A$. {\em Potentially}, because this possibility has hardly been 

362: explored. The postulates for revision presented in~\cite{Gardenfors:Flux} all

363: concern the case where the theory $A$ is held {\em constant}''.

364: We shall show, in the sequel, that the third point is intimately related

365: to the second one.

366: This paper will present an additional postulate constraining the way

367: one may revise different theories by the same formula.

368: Its main result is that this additional postulate precisely reduces the

369: generality of revisions in a way that makes revision isomorphic to

370: nonmonotonic inference. 

371: It will be shown that some of the additional postulates proposed

372: in the literature are incompatible with the AGM axioms and that those that

373: are not, are implied by our postulate.

374: \subsection{Plan of this paper}

375: \label{subsec:plan}

376: In section~\ref{sec:previous}, the additional postulates that were

377: previously proposed in the literature, by Katsuno and Mendelzon on one

378: hand, and by Darwiche and Pearl on the other hand, are described and 

379: discussed.

380: In section~\ref{sec:minimalin} the {\em minimal influence} postulate, 

381: \bKs{9}, is presented and justified on intuitive grounds. 

382: Some first consequences

383: of \bKs{9} that concern iterated revisions are then proven.

384: In section~\ref{subsec:K9U8} and \ref{subsec:K9DP}, its relations with the 

385: postulates of sections~\ref{subsec:KM} and~\ref{subsec:DP}, respectively,

386: are analyzed.

387: In section~\ref{subsec:representation}, 

388: we prove the main result of this paper:

389: revisions that satisfy \bKs{1} to \bKs{9} stand in one-to-one

390: correspondence with rational, consistency-preserving relations.

391: In section~\ref{subsec:conservext} a conservative extension result, 

392: a model-theoretic description of revision, and 

393: the converse to Theorem~\ref{the:Kfixed} are proven.

394: We discuss, then, in section~\ref{subsec:revwith}, the meaning of this 

395: result for the ontology of theory revision.

396: Section~\ref{sec:conc} discusses the existence

397: of two conflicting views on iterated revisions, concludes and describes some 

398: open questions.

399: \section{Additional postulates previously proposed}

400: \label{sec:previous}

401: \subsection{The Katsuno and Mendelzon postulate}

402: \label{subsec:KM}

403: In~\cite{KatMend:92}, Katsuno and Mendelzon considered the following postulate.

404: \[

405: {\bf U8} \ \ \ \rev{(K \cap K')}{*}{\varphi} = 

406: 	(\rev{K}{\!*}{\varphi}) \cap (\rev{K'}{\!*}{\varphi}).

407: \]

408: \noindent

409: Reasonable as it seems, this postulate is nevertheless

410: inconsistent with the AGM postulates. 

411: This is probably the reason why Katsuno and Mendelzon dropped or weakened 

412: some of the AGM postulates in the final version of their paper.

413: The postulate that will be proposed in this paper is closely related to

414: ${\bf U8}$, therefore we shall analyze it in some detail.

415: Our first remark is that any revision that satisfies the AGM postulates

416: also satisfies a special case (the {\em mild} case) of ${\bf U8}$.

417: \begin{proposition}

418: \label{prop:KM1}

419: If $*$ satisfies \bKs{3} and \bKs{4}, then, 

420: for any $\varphi$ such that \mbox{$\neg \varphi \not \in K \cup K'$},

421: \mbox{$\rev{(K \cap K')}{*}{\varphi} = 

422: (\rev{K}{\!*}{\varphi}) \cap (\rev{K'}{\!*}{\varphi})$}.

423: \end{proposition}

424: The proof is obvious.

425: Our second remark is that ${\bf U8}$ may be broken into two halves.

426: \begin{proposition}

427: \label{prop:KM2}

428: The postulate ${\bf U8}$ is equivalent to the following two properties:

429: \[

430: {\bf U8.1} \ \ {\rm If \ } K \subseteq K' , {\rm \ then \ } 

431: \rev{K}{\!*}{\varphi} \subseteq \rev{K'}{\!*}{\varphi} , {\rm \ and}

432: \]

433: \[

434: {\bf U8.2} \ \ (\rev{K}{\!*}{\varphi}) \cap (\rev{K'}{\!*}{\varphi}) \subseteq 

435: \rev{(K \cap K')}{*}{\varphi}.

436: \]

437: \end{proposition}

438: The proof is obvious.

439: Property ${\bf U8.1}$ has been named the postulate of Addition Monotonicity

440: and considered by G\"{a}rdenfors, Makinson and Segerberg 

441: in~\cite{Gard:Ramsey,Segerberg:impossibility,Mak:imposs}.

442: Despite its intuitive appeal (shouldn't the revised theory depend

443: monotonically on the revised theory?), it

444: has been shown to be inconsistent with the AGM postulates.

445: \begin{proposition}

446: \label{prop:KM3}

447: There is no revision operation that satisfies \bKs{4}, \bKs{5}

448: and ${\bf U8.1}$.

449: \end{proposition}

450: \proof

451: Suppose $*$ satisfies \bKs{4} and ${\bf U8.1}$.

452: Let $\varphi$ be an arbitrary formula that is not a logical contradiction.

453: By ${\bf U8.1}$, we have:

454: \[

455: \rev{\Cn(\varphi)}{*}{\bf true} \subseteq \rev{K_{\perp}}{\!*}{\bf true}.

456: \]

457: \noindent

458: But, since $\varphi$ is not a logical contradiction, 

459: \mbox{$\neg {\bf true} \not \in \Cn(\varphi)$} and, by \bKs{4},

460: we have

461: \mbox{$\Cn(\Cn(\varphi) , {\bf true}) \subseteq 

462: \rev{\Cn(\varphi)}{\!*}{\bf true}$}.

463: We conclude that 

464: \mbox{$\varphi \in \rev{\Cn(\varphi)}{\!*}{\bf true}$}, and, therefore,

465: \mbox{$\varphi \in \rev{K_{\perp}}{\!*}{\bf true}$}.

466: But the set of all formulas that are not a logical contradiction is

467: inconsistent and \rev{K_{\perp}}{\!*}{\bf true} is inconsistent,

468: contradicting \bKs{5}.

469: \QED

470: It will be shown in section~\ref{subsec:K9U8} that the additional

471: postulate we propose, \bKs{9}, implies ${\bf U8.2}$ 

472: and a special case of ${\bf U8.1}$.

473: \subsection{The Darwiche and Pearl postulates}

474: \label{subsec:DP}

475: In~\cite{DarwPearl:TARK}, Darwiche and Pearl proposed four additional

476: postulates. Notice that \rev{\rev{K}{\!*}{\psi}}{*}{\varphi}

477: means \rev{(\rev{K}{\!*}{\psi})}{*}{\varphi}.

478: \[

479: {\rm (C1)} \ \ \ \ {\rm If \ } \varphi \models \psi , {\rm \ then \ }

480: \rev{\rev{K}{\!*}{\psi}}{*}{\varphi} = \rev{K}{\!*}{\varphi}.

481: \]

482: \[

483: {\rm (C2)} \ \ \ \ {\rm If \ } \varphi \models \neg \psi , {\rm \ then \ }

484: \rev{\rev{K}{\!*}{\psi}}{*}{\varphi} = \rev{K}{\!*}{\varphi}.

485: \]

486: \[

487: {\rm (C3)} \ \ \ \ {\rm If \ } \psi \in \rev{K}{\!*}{\varphi} , {\rm \ then \ }

488: \psi \in \rev{\rev{K}{\!*}{\psi}}{*}{\varphi}.

489: \]

490: \[

491: {\rm (C4)} \ \ \ \ {\rm If \ } \neg \psi \not \in \rev{K}{\!*}{\varphi} , 

492: 	{\rm \ then \ }

493: \neg \psi \not \in \rev{\rev{K}{\!*}{\psi}}{*}{\varphi}.

494: \]

495: \noindent

496: It will be shown in section~\ref{subsec:K9DP} that the postulates (C1), (C3) 

497: and (C4) are implied by our additional postulate \bKs{9}.

498: The justifications given by Darwiche and Pearl for those postulates are

499: therefore indirect justifications for \bKs{9}.

500: The postulate (C2), contrary to the claims of Darwiche and Pearl,

501: is inconsistent with the AGM axioms.

502: \begin{proposition}

503: \label{prop:DP}

504: There is no revision that satisfies \bKs{1}--\bKs{4} and {\rm (C2)}.

505: \end{proposition}

506: \proof

507: Suppose $*$ satisfies \bKs{1}--\bKs{4} and (C2).

508: Let $\varphi$ be any tautology, say {\bf true} and $\psi$ be any logical

509: contradiction, say {\bf false}.

510: We have \mbox{${\bf true} \models \neg {\bf false}$}.

511: The postulate (C2) therefore implies that

512: \mbox{$\rev{\rev{K}{\!*}{\bf false}}{*}{\bf true} = \rev{K}{\!*}{\bf true}$}.

513: But, for any theory $K$, 

514: \mbox{$\rev{K}{\!*}{\bf false} = K_{\perp}$} by \bKs{1} and \bKs{2}.

515: Therefore, for any $K$,

516: \mbox{$\rev{K}{\!*}{\bf true} = \rev{K_{\perp}}{\!*}{\bf true}$}.

517: But, by \bKs{3} and \bKs{4}, for any {\em consistent} theory $K$,

518: \mbox{$\rev{K}{\!*}{\bf true} = K$}.

519: Since \rev{K_{\perp}}{\!*}{\bf true} does not depend on $K$, we conclude

520: that all consistent theories are equal. A contradiction.

521: \QED

522: \section{The minimal influence postulate}

523: \label{sec:minimalin}

524: \subsection{The postulate}

525: \label{subsec:postulate}

526: The postulate we propose to add to the AGM postulates is the following.

527: \[

528: \bKs{9}   \ \ \ {\rm For \ any \ theories \ } K , K' {\rm \ such \ that \ } 

529: 	\neg \varphi \in K , \neg \varphi \in K' \ 

530: \rev{K}{\!*}{\varphi} = \rev{K'}{\!*}{\varphi}.

531: \]

532: \noindent

533: The postulate \bKs{9} is obviously equivalent to:

534: if \mbox{$\neg \varphi \in K$}, then 

535: \mbox{$\rev{K}{\!*}{\varphi} = \rev{K_{\perp}}{\!*}{\varphi}$}.

536: Its meaning is that, if $\neg \varphi$ is an element of $K$, 

537: i.e., the revision of $K$ by $\varphi$ is a severe revision, 

538: then the result of revising $K$ by $\varphi$ does not depend on $K$.

539: Any revision that satisfies \bKs{3}, \bKs{4} and \bKs{9}, satisfies:

540: \begin{equation}

541: \label{eq:new}

542: \rev{K}{\!*}{\varphi} = 

543: 	\left \{ \begin{array}{l}

544: 	\rev{K_{\perp}}{\!*}{\varphi} {\rm \ if \ } \neg \varphi \in K \\

545: 	\Cn(K , \varphi) {\rm \ otherwise.}

546: 	\end{array} \right.

547: \end{equation}

548: Such a revision is therefore determined by its restriction to $K_{\perp}$.

549: \bKs{9} characterizes those revision systems in which

550: the revision $\rev{K}{\!*}{\varphi}$ depends mainly on $\varphi$ and only 

551: minimally on $K$ (in the case of a {\em mild} revision).

552: Remark also that, for any revision $*$ that satisfies \bKs{3} and \bKs{4},

553: the new postulate \bKs{9} is equivalent to:

554: \mbox{$\rev{K}{\!*}{\varphi} = \rev{\Cn(K , \varphi)}{*}{\varphi}$}.

555: 

556: The postulate \bKs{9}, at first sight, does not seem to be what one is 

557: looking for. 

558: First, it seems in danger of being inconsistent with the

559: AGM postulates. This concern will be addressed in full in 

560: section~\ref{sec:representation}.

561: But, also, one probably feels that the revisions of $K$ should

562: depend on $K$ in a major way, even when one considers a severe revision.

563: The feeling that some postulate should be added is widespread

564: and this paper will present formal arguments as to why \bKs{9} is the right

565: postulate to add.

566: Those reasons will not completely dissipate a first negative reaction

567: to \bKs{9}. The source of this clash between formal arguments and 

568: intuitive reaction probably lies in 

569: the philosophical underpinnings of theory revision and the general

570: framework chosen by AGM to study theory revision.

571: This paper claims that, in the framework chosen by AGM, \bKs{9} is

572: the right postulate to be added.

573: The AGM framework does not seem to be the one in which to study

574: iterated revisions.

575: 

576: We shall now present direct justification for \bKs{9}.

577: The postulate \bKs{9} is a special case of Darwiche and Pearl's postulate

578: (C2) described in section~\ref{subsec:DP}, assuming \bKs{1} and \bKs{2}.

579: Take, there, $\psi$ to be a logical contradiction. 

580: Since \mbox{$\rev{K}{\!*}{\psi} = K_{\perp}$},

581: one obtains \bKs{9}.

582: Anybody convinced by their defense of (C2) will endorse

583: \bKs{9}.

584: Our defense of \bKs{9} will be presented now.

585: Notice that \bKs{9} is equivalent to the following two properties:

586: they will be justified separately.

587: \[

588: \bKs{9.1}  \ \ \ {\rm For \ any \ theory \ } K {\rm \ such \ that \ } 

589: 	\neg \varphi \in K , \ 

590: \rev{K}{\!*}{\varphi} \subseteq \rev{K_{\perp}}{\!*}{\varphi}

591: \]

592: \[

593: \bKs{9.2}  \ \ \ {\rm For \ any \ theory \ } K {\rm \ such \ that \ } 

594: 	\neg \varphi \in K , \ 

595: \rev{K_{\perp}}{\!*}{\varphi} \subseteq \rev{K}{\!*}{\varphi}

596: \]

597: The first, \bKs{9.1} is relatively easy to justify.

598: It is a special case of ${\bf U8.1}$, i.e., the postulate of 

599: Addition Monotonicity. 

600: This postulate has been favorably considered by

601: many: it is very natural to hope that the more is believed before a revision,

602: the more is believed after.

603: None of the many articles cited above that

604: discuss this postulate rejects it on the grounds its meaning seems unwanted,

605: and the only problem found with it has been that it is inconsistent

606: with the AGM postulates. 

607: The special case proposed here, \bKs{9.1}, does not fall prey to this

608: criticism, it is consistent with the AGM postulates, and should be adopted.

609: 

610: The second, \bKs{9.2}, is more difficult to justify.

611: A formally weaker postulate, \bKs{9.2'}, 

612: will be presented and justified in a direct way. 

613: It will then been shown that, in the presence of other AGM postulates, 

614: it implies \bKs{9.2}.

615: The consideration of a similar property was suggested by Isaac Levi.

616: \[

617: \bKs{9.2'} \ \ \ \ 

618: 	{\rm If \ } \psi \in K {\rm \ and \ }

619: 	\psi \in \rev{K_{\perp}}{\!*}{\varphi} , {\rm \ then \ }

620: 	\psi \in \rev{K}{\!*}{\varphi}.

621: \]

622: \noindent

623: Our justification for \bKs{9.2'} is the following.

624: If \mbox{$\psi \in K$}, then the principle of maximal retention implies

625: we should try to keep $\psi$ in \rev{K}{\!*}{\varphi}.

626: If \mbox{$\psi \in \rev{K_{\perp}}{\!*}{\varphi}$}, we know that

627: $\psi$ may be kept in \rev{K}{\!*}{\varphi} (in particular $\psi$ does 

628: not contradict $\varphi$),

629: \bKs{9.2'} says that, in this case, we should have 

630: $\psi$ in \rev{K}{\!*}{\varphi}.

631: It is easy to see that \bKs{9.2} implies \bKs{9.2'}, 

632: in the presence of \bKs{4}.

633: It will now be shown that \bKs{9.2'} implies \bKs{9.2}.

634: \begin{proposition}

635: \label{prop:9'9}

636: Any revision $*$ that satisfies \bKs{1}, \bKs{2} and \bKs{9.2'}

637: satisfies \bKs{9.2}.

638: \end{proposition}

639: \proof

640: Let $*$ satisfy \bKs{1}, \bKs{2} and \bKs{9.2'}.

641: Suppose \mbox{$\neg \varphi \in K$}, and 

642: \mbox{$\psi \in \rev{K_{\perp}}{\!*}{\varphi}$}.

643: We must show that

644: \mbox{$\psi \in \rev{K}{\!*}{\varphi}$}.

645: But, \mbox{$\varphi \rightarrow \psi \in \rev{K_{\perp}}{\!*}{\varphi}$}

646: and 

647: \mbox{$\varphi \rightarrow \psi \in K$}.

648: By \bKs{9.2'}, we conclude that

649: \mbox{$\varphi \rightarrow \psi \in \rev{K}{\!*}{\varphi}$}.

650: By \bKs{1} and \bKs{2}, then,

651: \mbox{$\psi \in \rev{L}{\!*}{\varphi}$}.

652: \QED

653: 

654: The rest of this paper is devoted to proving consequences of \bKs{9}:

655: the intuitive appeal of those consequences provides indirect justification

656: for it. In particular, in sections~\ref{subsec:K9U8} and~\ref{subsec:K9DP},

657: it will be shown that \bKs{9} implies most of the additional postulates 

658: that have been previously proposed in the literature.

659: All the arguments in favor of those postulates, in particular those developed

660: in~\cite{DarwPearl:TARK} for (C1), (C3) and (C4), provide indirect support

661: for \bKs{9}.

662: Our representation result, Theorem~\ref{the:representation}, and the analysis

663: of section~\ref{subsec:revwith} provide, both, a soundness result that

664: shows that many revisions satisfy \bKs{1}--\bKs{9}, and an ontology

665: for those revisions. This provides additional indirect justification

666: for \bKs{9}.

667: 

668: It seems that \bKs{9} contradicts, in a certain degree, the

669: assumption of maximal retention. If \mbox{$\neg \varphi \in K \cap K'$},

670: and \mbox{$\psi \in K$} but \mbox{$\psi \not \in K'$}, then

671: a revision $*$ that satisfies \bKs{9} will have to let go of

672: $\psi$ when revising $K$ by $\varphi$ or to let go of $\neg \psi$ when

673: revising $K'$ by $\varphi$.

674: The meaning of this remark is that one cannot retain maximally always,

675: one must compromise with the assumption of maximal retention in certain

676: cases to be able to apply it in other cases, as, e.g., in the justification

677: of \bKs{9.2'}.

678: \subsection{First consequences: iterated revisions}

679: \label{subsec:iterated}

680: Our understanding of the AGM framework, 

681: in which revisions are operations of two arguments,

682: a theory and a formula, does not require any special mention of iterated

683: revisions: the result of revising $K$ first by $\psi$ and then by $\varphi$

684: is the result of revising by $\varphi$ the theory that is the result of 

685: revising $K$ by $\psi$.

686: Some authors, for example~\cite{BouGold:AAAI93}, \cite{Bou:IJCAI93}

687: and~\cite{MaryAnneW:trans}, 

688: take a different view and prefer to treat revisions as operating on

689: a fixed theory and treat iterated revisions as a special case of varying

690: this theory. This attitude is methodologically at odds with the AGM point of 

691: view, and we shall argue, in section~\ref{subsec:viewpoints}, 

692: that it tries to answer a different question.

693: Discussion about this approach and its relations

694: to the present work is postponed to section~\ref{subsec:viewpoints}.

695: 

696: It is interesting to consider the meaning of our postulate

697: for iterated revisions.

698: There are two fundamental cases to consider: 

699: the case where revising by $\psi$

700: and then by $\varphi$ is equivalent to revising by the conjunction

701: $\psi \wedge \varphi$, and the case where it is equivalent to revising by

702: the second formula alone, $\varphi$.

703: Our result concerning the first case does not use \bKs{9}:

704: if the second revision is mild then the iterated revision is equivalent

705: to the direct revision by the conjunction.

706: \begin{proposition}

707: \label{prop:phiandpsi}

708: Let $*$ be a revision operation that satisfies \bKs{1}--\bKs{8}.

709: If \mbox{$\neg \varphi \not \in \rev{K}{\!*}{\psi}$}, then

710: \mbox{$\rev{(\rev{K}{\!*}{\psi})}{*}{\varphi} = 

711: \rev{K}{\!*}{(\psi \wedge \varphi)}$}.

712: \end{proposition}

713: \proof

714: Assume \mbox{$\neg \varphi \not \in \rev{K}{\!*}{\psi}$}.

715: By \bKs{7} and \bKs{8}, 

716: \mbox{$\Cn(\rev{K}{\!*}{\psi} , \varphi) = 

717: \rev{K}{\!*}{(\psi \wedge \varphi)}$}.

718: By \bKs{3} and \bKs{4},

719: \mbox{$\Cn(\rev{K}{\!*}{\psi} , \varphi) = 

720: \rev{(\rev{K}{\!*}{\psi})}{*}{\varphi}$}.

721: \QED

722: For the second case, we have three different results.

723: First, \bKs{9} alone implies that 

724: \mbox{$\rev{\rev{K}{\!*}{\psi}}{*}{\varphi} = \rev{K}{\!*}{\varphi}$}

725: if one has \mbox{$\neg \varphi \in \rev{K}{\!*}{\psi}$}

726: and \mbox{$\neg \varphi \in K$}.

727: Secondly, we have the following.

728: \begin{proposition}

729: \label{prop:psi}

730: Let $*$ be a revision operation that satisfies \bKs{1}--\bKs{9}.

731: If \mbox{$\neg \varphi \in \rev{K}{\!*}{(\psi \vee \varphi)}$}, then

732: \mbox{$\rev{(\rev{K}{\!*}{\psi})}{*}{\varphi} = \rev{K}{\!*}{\varphi}$}.

733: \end{proposition}

734: \proof

735: Suppose \mbox{$\neg \varphi \in \rev{K}{\!*}{(\psi \vee \varphi)}$}.

736: We are going to show that $\neg \varphi$ is an element of both

737: \rev{K}{\!*}{\psi} and $K$ and we shall conclude by \bKs{9}.

738: To show that \mbox{$\neg \varphi \in \rev{K}{\!*}{\psi}$}, we consider

739: two cases.

740: First, if \mbox{$\neg \psi \not \in \rev{K}{\!*}{(\psi \vee \varphi)}$},

741: by \bKs{8} and \bKs{6}, we have 

742: \[

743: \neg \varphi \in \Cn(\rev{K}{\!*}{(\psi \vee \varphi)} , \psi) \subseteq

744: \rev{K}{\!*}{((\psi \vee \varphi) \wedge \psi)}

745: = \rev{K}{\!*}{\psi}.

746: \]

747: \noindent

748: Secondly, if \mbox{$\neg \psi \in \rev{K}{\!*}{(\psi \vee \varphi)}$},

749: since $\neg \varphi$ and $\psi \vee \varphi$ are also elements of

750: \rev{K}{\!*}{(\psi \vee \varphi)} (by \bKs{2}), we conclude that

751: \rev{K}{\!*}{(\psi \vee \varphi)} is inconsistent.

752: By \bKs5, $\psi \vee \varphi$ is a logical contradiction and

753: $\neg \varphi$ is a tautology and, by \bKs1, an element of \rev{K}{\!*}{\psi}.

754: It is left to us to show that \mbox{$\neg \varphi \in K$}.

755: But, by \bKs{3}, 

756: \[

757: \neg \varphi \in \rev{K}{\!*}{(\psi \vee \varphi)} \subseteq 

758: \Cn(K , \psi \vee \varphi) \subseteq \Cn(K , \varphi).

759: \]

760: \noindent

761: We conclude that

762: \mbox{$\neg \varphi \in K$}.

763: \QED

764: The third result of this family is perhaps the most striking.

765: It generalizes the postulates (C1) and (C3) of Darwiche and Pearl.

766: \begin{proposition}

767: \label{prop:gen}

768: Let $*$ be a revision operation that satisfies \bKs{1}--\bKs{9}.

769: If \mbox{$\psi \in \rev{K}{\!*}{\varphi}$}, then

770: \mbox{$\rev{(\rev{K}{\!*}{\psi})}{*}{\varphi} = \rev{K}{\!*}{\varphi}$}.

771: \end{proposition}

772: The meaning of Proposition~\ref{prop:gen} is very natural: if revising $K$ 

773: by $\varphi$ would convince the agent that $\psi$ is true, then revising

774: $K$, first by $\psi$, and then by $\varphi$ amounts to revising, first

775: by $\psi$, and then by some information, $\varphi$, that conditionally 

776: implies the first information. This boils down to revising first by some

777: partial information and then by the full information. 

778: This is indeed expected to

779: be equivalent to revising directly by the full information.

780: \proof

781: Suppose \mbox{$\psi \in \rev{K}{\!*}{\varphi}$}.

782: Two cases will be considered.

783: Suppose, first, that \mbox{$\neg \varphi \not \in \rev{K}{\!*}{\psi}$}.

784: By Proposition~\ref{prop:phiandpsi}, 

785: \mbox{$\rev{(\rev{K}{\!*}{\psi})}{*}{\varphi} = 

786: \rev{K}{\!*}{(\psi \wedge \varphi)}$}.

787: But \mbox{$\psi \in \rev{K}{\!*}{\varphi}$}, and, by Theorem~\ref{the:Kfixed}

788: and the properties of rational relations,  

789: \mbox{$\varphi \NIm^{K , *} \psi$} implies that, for any $\chi$,

790: \mbox{$\varphi \NIm^{K , *} \chi$} iff 

791: \mbox{$\varphi \wedge \psi \NIm^{K , *} \chi$}.

792: Therefore 

793: \mbox{$\rev{K}{\!*}{\varphi} = \rev{K}{\!*}{(\varphi \wedge \psi)}$}.

794: We conclude that 

795: \mbox{$\rev{(\rev{K}{\!*}{\psi})}{*}{\varphi} = \rev{K}{\!*}{\varphi}$}.

796: 

797: Suppose, now, that \mbox{$\neg \varphi \in \rev{K}{\!*}{\psi}$}.

798: By \bKs{3}, \mbox{$\neg \varphi \in \Cn(K , \psi)$}

799: and \mbox{$\psi \rightarrow \neg \varphi \in K$}.

800: But \mbox{$\psi \in \rev{K}{\!*}{\varphi}$}, and, similarly,

801: \mbox{$\varphi \rightarrow \psi \in K$}.

802: We conclude that \mbox{$\neg \varphi \in K$}.

803: By \bKs{9}, then, 

804: \mbox{$\rev{(\rev{K}{\!*}{\psi})}{*}{\varphi} = \rev{K}{\!*}{\varphi}$}.

805: \QED

806: In the next sections, we shall examine in detail the relation between 

807: \bKs{9} and the postulates previously proposed by, both, 

808: Katsuno and Mendelzon, and Darwiche and Pearl.

809: \subsection{\bKs{9} and ${\bf U8}$}

810: \label{subsec:K9U8}

811: Our first result is that any revision that satisfies \bKs{9} satisfies

812: a special case of ${\bf U8}$.

813: \begin{proposition}

814: \label{prop:K9U8.1}

815: Let $*$ be any revision operation that satisfies \bKs{9}.

816: If \mbox{$\neg \varphi \in K \cap K'$}, then 

817: \mbox{$(\rev{K}{\!*}{\varphi}) \cap (\rev{K'}{\!*}{\varphi}) = 

818: \rev{(K \cap K')}{*}{\varphi}$}.

819: \end{proposition}

820: The proof is obvious. 

821: Our second result deals with ${\bf U8.2}$.

822: \begin{proposition}

823: \label{prop:K9U8.2}

824: Any revision operation $*$ that satisfies \bKs{3}, \bKs{4} 

825: and \bKs{9} satisfies ${\bf U8.2}$.

826: \end{proposition}

827: \proof

828: For any revision $*$ that satisfies the assumptions, we must show that

829: \mbox{$(\rev{K}{\!*}{\varphi}) \cap (\rev{K'}{\!*}{\varphi}) \subseteq 

830: \rev{(K \cap K')}{*}{\varphi}$}.

831: Suppose, first, that \mbox{$\neg \varphi \not \in K \cap K'$}.

832: By \bKs{4}, 

833: \mbox{$\Cn(K \cap K' , \varphi) \subseteq \rev{(K \cap K')}{*}{\varphi}$}.

834: By \bKs{3},

835: \[

836: (\rev{K}{\!*}{\varphi}) \cap (\rev{K'}{\!*}{\varphi}) 

837: \subseteq \Cn(K , \varphi) \cap 

838: \Cn(K' , \varphi) = \Cn(K \cap K' , \varphi).

839: \]

840: \noindent

841: Suppose, now, that \mbox{$\neg \varphi \in K \cap K'$}.

842: We conclude easily by Proposition~\ref{prop:K9U8.1}.

843: \QED

844: Our conclusion is that \bKs{9} implies all the cases of ${\bf U8}$

845: that are consistent with the AGM postulates,

846: the case that is inconsistent with the AGM postulates being, as shown 

847: in Proposition~\ref{prop:KM3}, the case in which $\neg \varphi$

848: is in $K$ but not in $K'$. 

849: In this case the revision of the intersection

850: may be larger than the intersection of the revisions.

851: \subsection{\bKs{9} and the Darwiche and Pearl postulates}

852: \label{subsec:K9DP}

853: Let us, now, consider the postulates proposed by Darwiche and Pearl.

854: %Special case for (C2)

855: We have seen, in Proposition~\ref{prop:DP}, that no revision may satisfy

856: (C2), but the following special case of (C2) is a consequence of

857: \bKs{9}.

858: \begin{proposition}

859: \label{prop:specC2}

860: Any revision operation $*$ that satisfies \bKs{1}, \bKs{2} and \bKs{9}

861: satisfies:

862: \[

863: {\rm (C2') \ \ \ \  If\ } \neg \varphi \in K {\rm \ and \ }

864: \varphi \models \neg \psi , {\rm \ then \ }

865: \rev{\rev{K}{\!*}{\psi}}{*}{\varphi} = \rev{K}{\!*}{\varphi}.

866: \]

867: \end{proposition}

868: \proof

869: Suppose \mbox{$\neg \varphi \in K$} and \mbox{$\varphi \models \neg \psi$}.

870: By \bKs{9}, it is enough to prove that 

871: \mbox{$\neg \varphi \in \rev{K}{\!*}{\psi}$}, but this is the case since

872: \mbox{$\psi \models \neg \varphi$}.

873: \QED

874: We shall now show that the other postulates considered by Darwiche and Pearl

875: are consequences of \bKs{9}.

876: \begin{proposition}

877: \label{prop:K9C}

878: Any revision operation $*$ that satisfies \bKs{1}--\bKs{9} satisfies

879: {\rm (C1)}, {\rm (C3)} and {\rm (C4)}.

880: \end{proposition}

881: \proof

882: For (C1), suppose \mbox{$\varphi \models \psi$}.

883: By \bKs{1} and \bKs{2}, \mbox{$\varphi \in \rev{K}{\!*}{\varphi}$}

884: and, by Propostion~\ref{prop:gen}, we conclude that

885: \mbox{$\rev{\rev{K}{\!*}{\psi}}{*}{\varphi} = \rev{K}{\!*}{\varphi}$}.

886: 

887: For (C3), assume that \mbox{$\psi \in \rev{K}{\!*}{\varphi}$}.

888: By Propostion~\ref{prop:gen},

889: \mbox{$\rev{\rev{K}{\!*}{\psi}}{*}{\varphi} = \rev{K}{\!*}{\varphi}$}.

890: We conclude that \mbox{$\psi \in \rev{\rev{K}{\!*}{\psi}}{*}{\varphi}$}.

891: 

892: For (C4), assume that \mbox{$\neg \psi \not \in \rev{K}{\!*}{\varphi}$}.

893: We must show that 

894: \mbox{$\neg \psi \not \in \rev{\rev{K}{\!*}{\psi}}{*}{\varphi}$}.

895: Assume, first, that \mbox{$\neg \varphi \not \in \rev{K}{\!*}{\psi}$}.

896: In this case, by Proposition~\ref{prop:phiandpsi}, 

897: \mbox{$\rev{\rev{K}{\!*}{\psi}}{*}{\varphi} = 

898: \rev{K}{\!*}{(\varphi \wedge \psi)}$}.

899: But \mbox{$\neg \psi \not \in \rev{K}{\!*}{\varphi}$}

900: and therefore, by \bKs{1}, \mbox{$ \varphi \not \models \neg \psi$}

901: and $\varphi \wedge \psi$ is not a logical contradiction.

902: By \bKs{5}, then, \mbox{$\rev{K}{\!*}{(\varphi \wedge \psi)}$} is consistent

903: and \mbox{$\neg \psi \not \in \rev{K}{\!*}{(\varphi \wedge \psi)}$}.

904: We conclude that 

905: \mbox{$\neg \psi \not \in \rev{\rev{K}{\!*}{\psi}}{*}{\varphi}$}.

906: Assume, then, that \mbox{$\neg \varphi \in \rev{K}{\!*}{\psi}$}.

907: By, \bKs{3}, \mbox{$\psi \rightarrow \neg \varphi \in K$}.

908: We shall show that \mbox{$\neg \varphi \in K$} and conclude, by \bKs{9},

909: that \mbox{$\rev{\rev{K}{\!*}{\psi}}{*}{\varphi} = \rev{K}{\!*}{\varphi}$}.

910: To show that \mbox{$\neg \varphi \in K$}, it is enough, by \bKs{4}, 

911: to show that \mbox{$K \not \subseteq \rev{K}{\!*}{\varphi}$}.

912: We shall show that 

913: \mbox{$\psi \rightarrow \neg \varphi \not \in \rev{K}{\!*}{\varphi}$}, i.e.,

914: \mbox{$\varphi \rightarrow \neg \psi \not \in \rev{K}{\!*}{\varphi}$}.

915: Since \mbox{$\varphi \in \rev{K}{\!*}{\varphi}$}, it is enough to show

916: that \mbox{$\neg \psi \not \in \rev{K}{\!*}{\varphi}$}, which holds 

917: by hypothesis.

918: \QED

919: As a corollary, we realize that (C1), (C3) and (C4) follow from (C2').

920: \begin{corollary}

921: \label{co:C}

922: Any revision operation $*$ that satisfies \bKs{1}--\bKs{8} and (C2'),

923: satisfies (C1), (C3) and (C4).

924: \end{corollary}

925: \proof

926: We shall show that $*$ satisfies \bKs{9} and conclude by 

927: Proposition~\ref{prop:K9C}.

928: Suppose \mbox{$\neg \varphi \in K$}.

929: Since, \mbox{$\varphi \models \neg {\bf false}$}, by (C2'),

930: we have 

931: \mbox{$\rev{K}{\!*}{\varphi} = \rev{\rev{K}{\!*}{\bf false}}{*}{\varphi}$}.

932: But, by \bKs{1} and \bKs{2}, \mbox{$\rev{K}{\!*}{\bf false} = K_{\perp}$}.

933: \QED

934: We have brought additional indirect justification for \bKs{9}:

935: it implies all the postulates proposed by Darwiche an Pearl, except (C2),

936: that has been proved inconsistent. Some special case of (C2) is also

937: implied by \bKs{9}. 

938: \section{Representation of revisions by rational relations}

939: \label{sec:representation}

940: \subsection{Representation result}

941: \label{subsec:representation}

942: We shall now prove the main result of this paper:

943: revision operations that satisfy \bKs{1}--\bKs{9} may be represented

944: by rational, consistency-preserving relations.

945: There is, even, a bijection between the former and the latter.

946: This result is an improvement on the results of~\cite{MakGar:89}, 

947: that described a bijection between AGM revisions of a fixed theory

948: $L$ and rational, consistency-presrving relations \NI\ such that

949: \mbox{$L = \{ \ga \mid {\bf true} \NIm \ga \}$}.

950: This correspondence shows that the postulate \bKs{9} is consistent

951: with the AGM postulates: there are revisions that satisfy \bKs{1}--\bKs{9},

952: as many as rational, consistency-preserving relations.

953: \begin{theorem}

954: \label{the:representation}

955: There exists a bijection between the set of revisions that satisfy the

956: postulates \bKs{1}--\bKs{9} and the set of 

957: consistency-preserving, rational relations.

958: This map associates to every such revision $*$ the relation 

959: $\NIm_{*}$ defined by:

960: \begin{equation}

961: \label{eq:NIm*}

962: \varphi \NIm_{*} \psi {\rm \ iff \ } \psi \in \rev{K_{\perp}}{\!*}{\varphi}. 

963: \end{equation}

964: The inverse map associates to any rational, consistency-preserving 

965: relation \NI\ the revision $*_{\sim}$

966: defined by: 

967: \begin{equation}

968: \label{eq:*sim}

969: \rev{K}{\!*_{\sim}}{\varphi} = \left\{ \begin{array}{l}

970: 	\{ \psi \mid \varphi \NIm \psi \} {\rm \ if \ } \neg \varphi \in K \\

971: 	\Cn(K , \varphi) {\rm \  otherwise.}

972: 			\end{array}

973: 		\right.

974: \end{equation}

975: \end{theorem}

976: Note that the relation $\NIm_{*}$ defined in~(\ref{eq:NIm*}) is none other than

977: the relation $\NIm^{K_{\perp},*}$ defined in Theorem~\ref{the:Kfixed}.

978: Note also that the revision $*_{\sim}$ defined in~(\ref{eq:*sim}) from

979: the relation \NI\ has a very natural meaning.

980: It says that, in the case of a mild revision, revise as mandated

981: by \bKs{3} and \bKs{4}, and in the case of a severe revision,

982: disregard $K$ (since it is contradicted by the more solid information

983: $\varphi$) and replace it by $\varphi$ together with all the formulas that 

984: the agent thinks are

985: usually true when $\varphi$ is.

986: \proof

987: Suppose first that $*$ is a revision that satisfies \bKs{1} to 

988: \bKs{9}, and let

989: us show that the relation $\NIm_{*}$ defined in~(\ref{eq:NIm*}) 

990: is rational and consistency-preserving.

991: This follows from Theorem~\ref{the:Kfixed} applied to $K_{\perp}$ and $*$,

992: since $\NIm_{*}$ is $\NIm^{K_{\perp},*}$. Note that \bKs{3},

993: \bKs{4} and \bKs{9} are not used in this part of the proof.

994: 

995: Suppose now that \NI\ is a consistency-preserving, rational relation

996: and let us show that the revision $*_{\sim}$ defined in~(\ref{eq:*sim}) 

997: satisfies \bKs{1}--\bKs{9}.

998: The cases of \bKs{1} (by right weakening and ``and''),  

999: \bKs{2} (by reflexivity), \bKs{3} (if \mbox{$\neg \varphi \in K$},

1000: \mbox{$\Cn(K , \varphi) = K_{\perp}$}), \bKs{4} (by definition),

1001: \bKs{6} (by left logical equivalence), and \bKs{9} 

1002: (by definition) are easily dealt with.

1003: Let us show that $*_{\sim}$ satisfies \bKs{5}.

1004: Suppose \mbox{$\rev{K}{\!*_{\sim}}{\varphi}$} is inconsistent.

1005: Then, either \mbox{$\varphi \NIm {\bf false}$} or 

1006: \mbox{$\neg \varphi \not \in K$}

1007: and \mbox{$\Cn(K , \varphi)$} is inconsistent.

1008: The latter is impossible, therefore \mbox{$\varphi \NIm {\bf false}$}

1009: and we conclude by consistency-preservation.

1010: For \bKs{7}, we must show that 

1011: \mbox{$\rev{K}{\!*_{\sim}}{(\varphi \wedge \psi)} \subseteq \Cn(\rev{K}{\!*_{\sim}}{\varphi} , \psi)$}.

1012: We shall consider two cases.

1013: Suppose first that  \mbox{$ \neg \varphi \in K$}. 

1014: Then \mbox{$\neg(\varphi \wedge \psi) \in K $} too.

1015: Suppose \mbox{$\chi \in \rev{K}{\!*_{\sim}}{(\varphi \wedge \psi)}$}.

1016: Then, \mbox{$\varphi \wedge \psi \NIm \chi$}.

1017: By conditionalization (i.e., rule S), 

1018: we have \mbox{$\varphi \NIm \psi \rightarrow \chi$}

1019: and, since \mbox{$ \neg \varphi \in K$}, 

1020: \mbox{$\psi \rightarrow \chi \in \rev{K}{\!*_{\sim}}{\varphi}$}.

1021: Therefore, \mbox{$\chi \in \Cn(\rev{K}{\!*_{\sim}}{\varphi} , \psi)$}.

1022: Suppose now  that \mbox{$ \neg \varphi \not \in K$}.

1023: Then, \mbox{$\rev{K}{\!*_{\sim}}{\varphi} = \Cn(K , \varphi)$}, and

1024: \mbox{$\Cn(\rev{K}{\!*_{\sim}}{\varphi} , \psi) = \Cn(K , \varphi \wedge \psi)$}.

1025: But we have already shown that $*_{\sim}$ satisfies \bKs{3},

1026: therefore 

1027: \[

1028: \rev{K}{\!*_{\sim}}{(\varphi \wedge \psi)} \subseteq 

1029: \Cn(K , \varphi \wedge \psi).

1030: \]

1031: \noindent

1032: Finally, for \bKs{8}, suppose

1033: \mbox{$\neg \psi \not \in \rev{K}{\!*_{\sim}}{\varphi}$}.

1034: We have to show that 

1035: \[

1036: \Cn(\rev{K}{\!*_{\sim}}{\varphi} , \psi) 

1037: \subseteq  \rev{K}{\!*_{\sim}}{(\varphi \wedge \psi)},

1038: \]

1039: \noindent

1040: or, equivalently that if \mbox{$\psi \rightarrow \chi \in \rev{K}{\!*_{\sim}}{\varphi}$},

1041: then, \mbox{$\chi \in \rev{K}{\!*_{\sim}}{(\varphi \wedge \psi)}$}.

1042: We consider again two cases.

1043: If, first, 

1044: \mbox{$\neg\varphi\in K$}, then

1045: \mbox{$\varphi \NIm \psi \rightarrow \chi $}

1046: and also \mbox{$\varphi \notNIm \neg \psi$},

1047: therefore, by rational monotonicity,

1048: \mbox{$\varphi \wedge \psi \NIm \psi \rightarrow \chi $}

1049: and, by ``and'' and right weakening,

1050: \mbox{$\varphi \wedge \psi \NIm \chi $}.

1051: But \mbox{$\neg\varphi\in K$} and therefore 

1052: \mbox{$\neg(\varphi \wedge \psi) \in K$} and we conclude that

1053: \mbox{$\chi \in \rev{K}{\!*_{\sim}}{(\varphi \wedge \psi)}$}.

1054: The second case we must consider is: 

1055: \mbox{$\neg \varphi \not \in K$}.

1056: Then \mbox{$\rev{K}{\!*_{\sim}}{\varphi} = \Cn(K , \varphi)$}

1057: and, by hypothesis, $\neg \psi$ is not a member of this set.

1058: This implies that 

1059: \mbox{$\neg \varphi \vee \neg \psi \not \in K$}.

1060: Therefore,

1061: \mbox{$\rev{K}{\!*_{\sim}}{(\varphi \wedge \psi)} 

1062: = \Cn(K , \varphi \wedge \psi)$}.

1063: Therefore 

1064: \mbox{$\rev{K}{\!*_{\sim}}{\varphi} = \Cn(K , \varphi) 

1065: \subseteq \rev{K}{\!*_{\sim}}{(\varphi \wedge \psi)}$}.

1066: 

1067: To complete the proof of the representation theorem, we have to show that

1068: the maps \mbox{$\NIm \longmapsto *_{\sim}$} and 

1069: \mbox{$* \longmapsto \NIm_{*}$} are inverse maps, i.e., that

1070: \mbox{$\NIm_{*_{\sim}} = \NIm$} and that 

1071: \mbox{$*_{\sim_{*}} = *$}.

1072: For the first equality, we have

1073: \mbox{$\varphi \NIm_{*_{\sim}} \psi$} iff

1074: \mbox{$\psi \in \rev{K_{\perp}}{\!*_{\sim}}{\varphi}$},

1075: iff \mbox{$\varphi \NIm \psi$} since \mbox{$\neg \varphi \in K_{\perp}$}.

1076: For the second equality, we have 

1077: \[

1078: \rev{K}{\!*_{\sim_{*}}}{\varphi} = \left\{ \begin{array}{l}

1079: 	\{ \psi \mid \varphi \NIm_{*} \psi \} {\rm \ if \ } 

1080: 			\neg \varphi \in K \\

1081: 	\Cn(K , \varphi) {\rm \  otherwise.}

1082: 			\end{array}

1083: 		\right.

1084: \]

1085: \noindent

1086: Therefore,

1087: \[

1088: \rev{K}{\!*_{\sim_{*}}}{\varphi} = \left\{ \begin{array}{l}

1089: 	\rev{K_{\perp}}{\!*}{\varphi} {\rm \ if \ } 

1090: 			\neg \varphi \in K \\

1091: 	\Cn(K , \varphi) {\rm \  otherwise.}

1092: 			\end{array}

1093: 		\right.

1094: \]

1095: \noindent

1096: But, if \mbox{$\neg \varphi \in K$}, by \bKs{9}, 

1097: \mbox{$\rev{K_{\perp}}{\!*}{\varphi} = \rev{K}{\!*}{\varphi}$},

1098: and if \mbox{$\neg \varphi \not \in K$},

1099: by \bKs{3} and \bKs{4}, 

1100: \mbox{$\Cn(K , \varphi) = \rev{K}{\!*}{\varphi}$}.

1101: Notice that the postulates used now, \bKs{3}, \bKs{4} 

1102: and \bKs{9} are those that were not used in the first part of this proof.

1103: \QED

1104: In section~\ref{subsec:AGM}, in the discussion following the proof of

1105: Theorem~\ref{the:Kfixed}, we noticed that any theory $K$ and any revision

1106: operation $*$ that satisfies \bKs{1}, \bKs{2} and \bKs{5}

1107: to \bKs{8} defines a rational, consistency-preserving relation,

1108: the relation $\NIm^{K , *}$ defined in formula~(\ref{eq:K*}).

1109: We have shown, in Theorem~\ref{the:representation}, that any such relation 

1110: is defined in this way by a specific theory $K_{\perp}$ and some

1111: revision operation that satisfies not only \bKs{1}, \bKs{2} and 

1112: \bKs{5}--\bKs{8}, but also \bKs{3}, \bKs{4} and \bKs{9}.

1113: It does not mean that \bKs{3}, \bKs{4} and \bKs{9} are redundant, i.e.,

1114: derivable from the other postulates, but only that the mapping

1115: \mbox{$(K , *) \longmapsto \NIm^{K , *}$} is not injective.

1116: In fact, the reader will quickly realize that no postulate is redundant, 

1117: though \bKs{3} may be weakened to 

1118: \mbox{$\rev{K}{\!*}{\bf true} \subseteq K$} 

1119: in the presence of \bKs{6} and \bKs{7}.

1120: We take this to mean that postulates \bKs{3}, \bKs{4} and \bKs{9}

1121: say nothing about the {\em nonmonotonic inference} aspect of revision:

1122: they are orthogonal to it.

1123: \subsection{Consequences of the representation theorem}

1124: \label{subsec:conservext}

1125: The first consequence that we shall draw from Theorem~\ref{the:representation}

1126: is that the postulate \bKs{9} does not constrain the ways one may revise 

1127: a fixed theory $K$ more than the AGM postulates already did:

1128: it only constrains the ways one may revise different theories by 

1129: related formulas.

1130: If we consider an AGM revision $*$ and a fixed theory $K$,

1131: there is a revision that satisfies \bKs{1}--\bKs{9} and

1132: that revises the theory $K$, for any formula $\varphi$, exactly as does $*$.

1133: \begin{theorem}

1134: \label{the:conservext}

1135: Let $*$ a revision that satisfies the AGM postulates

1136: \bKs{1} to \bKs{8}.

1137: Then, for any theory $K$,

1138: there exists a revision  $*_{K}$ that satisfies

1139: \bKs{1}--\bKs{9} and such that,

1140: for any $\varphi$, \mbox{$\rev{K}{\!*}{\varphi} = 

1141: \rev{K}{\!*_{K}}{\varphi}$}.

1142: \end {theorem}

1143: \proof

1144: Let $*$ a revision that satisfies the AGM postulates

1145: \bKs{1} to \bKs{8} and $K$ an arbitrary theory.

1146: By Theorem~\ref{the:Kfixed}, the relation $\NIm^{K , *}$, defined in

1147: equation~(\ref{eq:K*}) is rational and consistency-preserving.

1148: For simplifying notations, we shall denote $\NIm^{K , *}$ by \NI.

1149: By Theorem~\ref{the:representation}, the revision $*_{\sim}$,

1150: defined in equation~\ref{eq:NIm*},

1151: satisfies \bKs{1}--\bKs{9}. 

1152: But, for any $\varphi$,

1153: \[

1154: \rev{K}{\!*_{\sim}}{\varphi} = 

1155: 	\left \{ \begin{array}{l}

1156: 	\{ \psi \mid \varphi \NIm^{K , *} \psi \} 

1157: 		{\rm \ if \ } \neg \varphi \in K \\

1158: 	\Cn(K , \varphi) {\rm \  otherwise.}

1159: 			\end{array}

1160: 		\right.

1161: \]

1162: \noindent

1163: Therefore, by equation~(\ref{eq:K*}),

1164: \[

1165: \rev{K}{\!*_{\sim}}{\varphi} = 

1166: 	\left \{ \begin{array}{l}

1167: 	\rev{K}{\!*}{\varphi} {\rm \ if \ } \neg \varphi \in K \\

1168: 	\Cn(K , \varphi) {\rm \  otherwise.}

1169: 			\end{array}

1170: 		\right.

1171: \]

1172: \noindent

1173: But the right hand-side is equal to \rev{K}{\!*}{\varphi}.

1174: \QED

1175: From the proof above, one sees that the revision $*_{K}$ is defined by:

1176: \begin{equation}

1177: \label{eq:*L}

1178: \rev{L}{\!*_{K}}{\varphi} = \left\{ \begin{array}{l}

1179: 	\rev{K}{\!*}{\varphi} {\rm \ if \ } \neg \varphi \in L \\ 

1180: 	\Cn(L , \varphi) {\rm \ otherwise.} 

1181: 			\end{array}

1182: 		\right.

1183: \end{equation}

1184: \noindent

1185: An model-theoretic description of $*$ will be given now.

1186: Suppose $*$ satisfies \bKs{1}--\bKs{9}, and $\NIm_{*}$ is the rational, 

1187: consistency-preserving relation associated to it by equation~(\ref{eq:NIm*}).

1188: Let $M$ be any ranked model that defines $\NIm_{*}$ (see~\cite{LMAI:92}).

1189: Note, that, since $\NIm_{*}$ is consistency-preserving, every propositional

1190: model is the label in $M$ of some state.

1191: For any theory $K$, let us define $M_{K}$ as the ranked model obtained

1192: from $M$ by creating, below the states of lowest rank in $M$, a new level

1193: containing a state for each propositional model of $K$ (and labeled by this 

1194: model).

1195: Note that $M$ = $M_{K_{\perp}}$.

1196: It is clear that $M_{K}$ is a ranked model.

1197: Let $\NIm_{M_{K}}$ be the rational relation defined by $M_{K}$.

1198: \begin{proposition}

1199: \label{prop:modelth}

1200: \[

1201: \psi \in \rev{K}{\!*}{\varphi} {\rm \ iff \ } \varphi \NIm_{M_{K}} \psi

1202: \]

1203: \end{proposition}

1204: \proof

1205: The result follows from equation~(\ref{eq:new}) and the definition of

1206: $\NIm_{M_{K}}$.

1207: Note that there is some state at the lowest rank in $M_{K}$ that satisfies

1208: $\varphi$ iff \mbox{$\neg \varphi \not \in K$}.

1209: In this case \mbox{$\varphi \NIm_{M_{K}} \psi$} iff 

1210: \mbox{$\psi \in \Cn(K , \varphi)$}.

1211: \QED

1212: An important remark, that will be developed in 

1213: section~\ref{subsec:viewpoints},

1214: is that, even though $M$ (and $\NIm_{*}$) may be recuperated back 

1215: from $M_{K}$ by leaving out the states of lowest rank, 

1216: there is not enough information in $\NIm_{M_{K}}$ to determine

1217: $\NIm_{*}$, or $\NIm_{M_{L}}$ for \mbox{$L \neq K$}.

1218: Think that one may be build a model $M'_{K}$ from $M_{K}$ by leaving out

1219: all the states of $M_{K}$ that satisfy $K$ except the states of lowest rank.

1220: The model $M'_{K}$ defines the same relation as $M_{K}$, but no

1221: trace is left in it of the original ranks in $M$ of the states that satisfy

1222: $K$.

1223: Our next consequence of the representation theorem is the converse

1224: of Theorem~\ref{the:Kfixed}.

1225: \begin{theorem}

1226: \label{the:converse}

1227: Let \NI\ be an arbitrary, rational, consistency-preserving relation.

1228: There is a theory $K$ and a revision $*$ that satisfies \bKs{1}--\bKs{9}

1229: such that \mbox{$\varphi \NIm \psi$} iff 

1230: \mbox{$\psi \in \rev{K}{\!*}{\varphi}$}.

1231: \end{theorem}

1232: \proof

1233: Let \NI\ be as above.

1234: Take \mbox{$K \eqdef \{ \alpha \mid {\bf true} \NIm \alpha \}$}.

1235: It is clear that $K$ is a theory.

1236: Take \mbox{$ * \eqdef *_{\sim}$}, defined in equation~\ref{eq:*sim}.

1237: For any $\varphi$, we have:

1238: \[

1239: \rev{K}{\!*}{\varphi} = 

1240: 	\left \{ \begin{array}{l}

1241: 	\{ \psi \mid \varphi \NIm \psi \} {\rm \ if \ } \neg \varphi \in K \\

1242: 	\Cn(K , \varphi) {\rm \  otherwise.}

1243: 			\end{array}

1244: 		\right.

1245: \]

1246: \noindent

1247: By the definition of $K$:

1248: \[

1249: \rev{K}{\!*}{\varphi} = 

1250: 	\left \{ \begin{array}{l}

1251: 	\{ \psi \mid \varphi \NIm \psi \} 

1252: 		{\rm \ if \ } {\bf true} \NIm \neg \varphi \\

1253: 	\Cn(K , \varphi) {\rm \  otherwise.}

1254: 			\end{array}

1255: 		\right.

1256: \]

1257: \noindent

1258: But \NI\ is rational and: 

1259: \mbox{$\rev{K}{\!*}{\varphi} = \{ \psi \mid \varphi \NIm \psi \} $}.

1260: \QED

1261: \subsection{Revising with a rational, consistency-preserving relation}

1262: \label{subsec:revwith}

1263: The definition of the revision $*_{\sim}$ in equation~(\ref{eq:*sim})

1264: proposes a certain ontology for theory revision.

1265: The agent has a certain rational, consistency-preserving relation

1266: in mind. This relation describes, as far as the agent knows, 

1267: what follows from what by default, or in normal circumstances.

1268: When it receives some new information $\varphi$, the agent revises 

1269: its current belief set $K$ by either adding $\varphi$ to its current belief

1270: set, if $\varphi$ does not contradict $K$, or, if it contradicts $K$,

1271: by forgetting about $K$ altogether and adopting $\varphi$ and all default

1272: assumptions that go with $\varphi$ as its new belief set.

1273: In the latter situation, the agent adopts not only $\varphi$ but all its

1274: usual, normal consequences also.

1275: This process seems very reasonable: when learning $\varphi$ that contradicts

1276: its current beliefs $K$, the agent should probably not try to throw

1277: away as little of possible of $K$ to accommodate just $\varphi$ but should

1278: throw away enough of $K$

1279: to accommodate also the most plausible consequences of $\varphi$.

1280: Remember that it is a basic assumption of the AGM approach that

1281: $\varphi$ is more reliable than $K$, so we are only saying that the

1282: {\em default}, i.e. {\em normal}, consequences of $\varphi$ should

1283: also be considered more secure than $K$.

1284: If, for example, one's theory says it should rain in Paris and in Orl\'{e}ans

1285: and one learns that there are no clouds in the sky of Paris, 

1286: one has a choice between two revisions:

1287: conclude that it does not rain in Paris but rains in Orl\'{e}ans,

1288: or conclude that it rains neither in Paris nor in Orl\'{e}ans.

1289: If one holds the default assumption that when the sky

1290: is cloudless in Paris it does not rain in Orl\'{e}ans, 

1291: the second way of revising seems more reasonable: it gives precedence to

1292: the default ``when there are no clouds in Paris, it does not rain in 

1293: Orl\'{e}ans'' to the proposition ``it rains in Orl\'{e}ans'' 

1294: that was in the original theory.

1295: Note that this does not contradict the principle of maximal retention,

1296: since the original theory contained both 

1297: ``when there are no clouds in Paris, it does not rain in Orl\'{e}ans''

1298: and ``it rains in Orl\'{e}ans'', and one could not keep both of them

1299: in the face of the information that ``there are no clouds in Paris''.

1300: 

1301: In the process of iterated revisions considered here, the agent

1302: revises its beliefs with new information, using its knowledge of

1303: how things generally, or normally, behave. This knowledge is encoded

1304: in a rational, consistency-preserving relation.

1305: The agent does {\em not} revise its relation, i.e., its default

1306: knowledge.

1307: There is a lot to be said for this analysis: default assumptions

1308: are typically much more stable than beliefs, and so it should be.

1309: Consider, for example, the physicist revising his/her theory about

1310: the world in light of the results of new experiments.

1311: The method used in performing this revision, i.e., deciding what to

1312: keep and what to throw away of the old theory, is a matter of meta-principles,

1313: a methodological question, not a question of physics.

1314: There is no reason to think that the result of this new experiment will

1315: change anything in the methodological principles our physicist

1316: is using.

1317: Nevertheless, there must also be another process at work:

1318: the revision (or the creation) of default assumptions.

1319: After all, those default assumptions, or these meta-principles come

1320: from the observation of evidence.

1321: Decision theory seems to be mainly interested in this process of formation

1322: or revision of default assumptions in the light of new evidence.

1323: It operates on a longer time scale than the revisions considered here

1324: and this paper has nothing to say about it.

1325: The revisions we consider are revisions of belief sets performed under

1326: a fixed set of default assumptions.

1327: \section{Closing remarks}

1328: \label{sec:conc}

1329: \subsection{Two viewpoints on iterated revisions}

1330: \label{subsec:viewpoints}

1331: We have described and developed what we think is the original

1332: AGM viewpoint concerning revisions. Revisions have two arguments:

1333: a theory to be revised and a formula by which it is revised.

1334: Iterated revisions do not deserve any special consideration.

1335: Suppose an agent holds $*$ as its revision procedure.

1336: If, at some point in time, it has belief set $K$ 

1337: and learns $\varphi$ it will, then, acquire belief set \rev{K}{\!*}{\varphi}.

1338: If the agent later learns $\psi$, it will just use the {\em same} revision $*$

1339: to revise its current belief set \rev{K}{\!*}{\varphi}, by the new information

1340: $\psi$, thus obtaining \rev{\rev{K}{\!*}{\varphi}}{*}{\psi} as its new

1341: belief set. In this process, the belief revision process, $*$, 

1342: {\em does not change}. From this point of view, the revision process

1343: is fixed: {\em theories} are revised but the 

1344: {\em revision method} stays fixed.

1345: This is the {\em static} viewpoint.

1346: 

1347: There is a different point of view, that has been illustrated in a number

1348: of recent works on iterated revisions, notably~\cite{BouGold:AAAI93}, 

1349: \cite{Bou:IJCAI93} and~\cite{MaryAnneW:trans}.

1350: This point of view seems to be rooted in Bayesian decision theory

1351: and to draw on~\cite{Spohn:87}.

1352: This alternative, {\em dynamic}, point of view has attracted most of 

1353: the attention recently given to theory revision and can be described as

1354: follows.

1355: The agent starts with a belief set $K$ and a method for revising $K$

1356: when some new information will come along, i.e., the agent knows only how to

1357: revise $K$, its belief set, it does not know how to revise arbitrary theories.

1358: If some new information comes along, it will revise $K$ with this new 

1359: information $\varphi$, following its method and obtain a new belief set $K'$, 

1360: but it will also adapt its revision method to be able to revise $K'$

1361: when new information will eventually come along.

1362: From this point of view, the new information $\varphi$ does not only

1363: modify the belief set it {\em also modifies the revision method}.

1364: This dynamic viewpoint presents a departure from the AGM static framework.

1365: It even seems that those two points of view are incompatible.

1366: It will be shown, first, that a dynamical viewpoint does not provide

1367: a static revision operator, and then, that a static revision operator

1368: does not provide a method for dynamically adapting revisions.

1369: 

1370: Let us take the dynamical viewpoint.

1371: \begin{itemize}

1372: \item In the dynamic view the new revision procedure

1373: by which \rev{K}{\!*}{\varphi} will be revised depends on $\varphi$,

1374: whereas, in the static view it should not.

1375: \item It must not be the case that all theories may be obtained starting 

1376: from $K$ by iterated revisions, whereas the static view insists on being

1377: capable of revising arbitrary theories.

1378: \item It must not be the case that the revision of a theory by a formula

1379: is determined by the theory and the formula. It may depend on the

1380: sequence of revisions that led to the theory.

1381: In other terms, it may happen that $K$ is identical with

1382: \rev{\rev{K}{\!*}{\varphi}}{*}{\psi} without implying that

1383: \rev{K}{\!*}{\chi} is identical with 

1384: \rev{\rev{\rev{K}{\!*}{\varphi}}{*}{\psi}}{\!*}{\chi}.

1385: \end{itemize}

1386: 

1387: Let us take the static viewpoint, now.

1388: The static viewpoint proposes \rev{\rev{K}{\!*}{\psi}}{*}{\varphi}

1389: for the result of revising K, successively, first by $\psi$ and then

1390: by $\varphi$. The dynamic viewpoint says this iterated revision should

1391: be the result of revising \rev{K}{\!*}{\psi} by $\varphi$, {\em using the

1392: revision method obtained by adapting the revision 

1393: \mbox{$\mu \eqdef \chi \longmapsto \rev{K}{\!*}{\chi}$} to the new information

1394: $\psi$}.

1395: But, in section~\ref{subsec:conservext}, 

1396: following Proposition~(\ref{prop:modelth}), we noticed that $\mu$ does not

1397: contain enough information to determine $*$.

1398: The mapping 

1399: \mbox{$\nu \eqdef \varphi \longmapsto \rev{\rev{K}{\!*}{\psi}}{*}{\varphi}$} 

1400: is not determined by \rev{K}{\!*}{\psi} and $\mu$ alone.

1401: The method proposed in this paper for revising with a rational,

1402: consistency-preserving relation does not yield any method for adapting

1403: revisions, any {\em transmutation} in the terms of~\cite{MaryAnneW:trans}. 

1404: \subsection{Conclusion and future work}

1405: \label{subsec:conc}

1406: Building on previous work by many researchers, we have completed the

1407: analysis of the relation between the extended AGM postulates

1408: for theory revision and properties of nonmonotonic inference.

1409: It was known that one may associate rational, consistency-preserving relations

1410: to any revision that satisfies the AGM postulates.

1411: We have shown that any such relation is associated with exactly one revision

1412: that satisfies the AGM postulates and an additional one.

1413: The open questions on which further work is needed have to do with

1414: generalizing the results of this paper to larger classes of revisions

1415: and nonmonotonic systems.

1416: 

1417: One may ask if our results can be generalized to the revisions that

1418: satisfy only the postulates: \bKs{1}--\bKs{7}.

1419: Those revisions are known to be intimately related to preferential

1420: relations~\cite{Rott:92}. 

1421: Preferential relations may stand in one-to-one correspondence with 

1422: the revisions that satisfy \bKs{1}--\bKs{7} and \bKs{9}.

1423: 

1424: An orthogonal avenue seems also interesting. 

1425: We noticed that \bKs{3} and \bKs{4} have no influence on the nonmonotonic

1426: aspect of revision. 

1427: The revisions that satisfy \bKs{1}, \bKs{2} and \bKs{5}--\bKs{8}

1428: seem an interesting class to study.

1429: 

1430: We noticed that, in the AGM framework, the theories to be revised

1431: were unstructured sets of formulas. One would like to put some more

1432: structure on those theories. 

1433: Such additional structure, on $K$, should probably constrain the revisions

1434: of the revised theories of the form \rev{K}{\!*}{\varphi}.

1435: The analogy between (unstructured) theory revision and nonmonotonic

1436: inference should probably lift to an analogy between generalized theory

1437: revision and conditional logic. 

1438: \section{Acknowledgment}

1439: During the last stage of the elaboration of this work,

1440: Daniel Lehmann held long and fruitful conversations with Isaac Levi.

1441: They are gratefully acknowledged.

1442: David Makinson's detailed and knowledgeable comments on a first draft

1443: helped us catch up on the history of the field and improve the presentation

1444: altogether. 

1445: \begin{thebibliography}{10}

1446: 

1447: \bibitem{AM:82}

1448: C.~A. Alchourr\'{o}n and D.~Makinson.

1449: \newblock On the logic of theory change: Contraction functions and their

1450:   associated revision functions.

1451: \newblock {\em Theoria}, 48:14--37, 1982.

1452: 

1453: \bibitem{AM:85}

1454: C.~A. Alchourr\'{o}n and D.~Makinson.

1455: \newblock The logic of theory change: Safe contraction.

1456: \newblock {\em Studia Logica}, 44:405--422, 1985.

1457: 

1458: \bibitem{AGM:85}

1459: C.A. Alchourr\'{o}n, P.~G\"{a}rdenfors, and D.~Makinson.

1460: \newblock On the logic of theory change: partial meet contraction and revision

1461:   functions.

1462: \newblock {\em The Journal of Symbolic Logic}, 50:510--530, 1985.

1463: 

1464: \bibitem{Bou:IJCAI93}

1465: Craig Boutilier.

1466: \newblock Revision sequences and nested conditionals.

1467: \newblock In Ruzena Bajcsy, editor, {\em Proceedings of the 13th I.J.C.A.I.},

1468:   pages 519--525. Morgan Kaufmann, Chamb\'{e}ry, Savoie, France, August 1993.

1469: 

1470: \bibitem{BouGold:AAAI93}

1471: Craig Boutilier and Mois\'{e}s Goldszmidt.

1472: \newblock Revision by conditional beliefs.

1473: \newblock In {\em Proceedings of the 11th National Conference on Artificial

1474:   Intelligence (AAAI)}, pages 649--654. Morgan Kaufmann, Washington, D.C., July

1475:   1993.

1476: 

1477: \bibitem{DarwPearl:TARK}

1478: Adnan Darwiche and Judea Pearl.

1479: \newblock On the logic of iterated belief revision.

1480: \newblock In Ronald Fagin, editor, {\em Proceedings of the fifth Conference on

1481:   Theoretical Aspects of Reasoning about Knowledge}, pages 5--23. Morgan

1482:   Kaufmann, Pacific Grove, CA, March 1994.

1483: 

1484: \bibitem{GardMak:88}

1485: P.~G\"{a}rdenfors and D.~Makinson.

1486: \newblock Revisions of knowledge systems using epistemic entrenchment.

1487: \newblock In Moshe~Y. Vardi, editor, {\em Proceedings of the Second Conference

1488:   on Theoretical Aspects of Reasoning About Knowledge}, pages 83--95, Monterey,

1489:   California, March 1988. Morgan Kaufmann.

1490: 

1491: \bibitem{Gard:Ramsey}

1492: Peter G\"{a}rdenfors.

1493: \newblock Belief revisions and the ramsey test for conditionals.

1494: \newblock {\em The Philosophical Review}, 95:81--93, 1986.

1495: 

1496: \bibitem{Gardenfors:Flux}

1497: Peter G\"{a}rdenfors.

1498: \newblock {\em Knowledge in Flux: Modeling the Dynamics of Epistemic States}.

1499: \newblock MIT Press, Bradford Books, Cambridge, MA, 1988.

1500: 

1501: \bibitem{Gard:beliefrev}

1502: Peter G\"{a}rdenfors, editor.

1503: \newblock {\em Belief Revision}, number~29 in Cambridge Tracts in Theoretical

1504:   Computer Science. Cambridge University Press, 1992.

1505: 

1506: \bibitem{Gard:intro}

1507: Peter G\"{a}rdenfors.

1508: \newblock Belief revision: An introduction.

1509: \newblock In Peter G\"{a}rdenfors, editor, {\em Belief Revision}, number~29 in

1510:   Cambridge Tracts in Theoretical Computer Science, pages 1--28. Cambridge

1511:   University Press, 1992.

1512: 

1513: \bibitem{GarMak:92}

1514: Peter G\"{a}rdenfors and David Makinson.

1515: \newblock Nonmonotonic inference based on expectations.

1516: \newblock {\em Artificial Intelligence}, 65(1):197--245, January 1994.

1517: 

1518: \bibitem{KatMend:91}

1519: Hirofumi Katsuno and Alberto~O. Mendelzon.

1520: \newblock Propositional knowledge base revision and minimal change.

1521: \newblock {\em Journal of Artificial Intelligence}, 52:263--294, 1991.

1522: 

1523: \bibitem{KatMend:92}

1524: Hirofumi Katsuno and Alberto~O. Mendelzon.

1525: \newblock On the difference between updating a knowledge base and revising it.

1526: \newblock In Peter G\"{a}rdenfors, editor, {\em Belief Revision}, number~29 in

1527:   Cambridge Tracts in Theoretical Computer Science, pages 183--203. Cambridge

1528:   University Press, 1992.

1529: 

1530: \bibitem{KelWin:85}

1531: Arthur~M. Keller and Marianne~Winslett Wilkins.

1532: \newblock On the use of an extended relational model to handle changing

1533:   incomplete information.

1534: \newblock {\em IEEE Transactions on Software Engineering}, 11(7):620--633, July

1535:   1985.

1536: 

1537: \bibitem{LMAI:92}

1538: Daniel Lehmann and Menachem Magidor.

1539: \newblock What does a conditional knowledge base entail?

1540: \newblock {\em Artificial Intelligence}, 55(1):1--60, May 1992.

1541: 

1542: \bibitem{MakGar:89}

1543: D.~Makinson and P.~G\"{a}rdenfors.

1544: \newblock Relations between the logic of theory change and nonmonotonic logic.

1545: \newblock In Fuhrmann A. and M.~Morreau, editors, {\em The Logic of Theory

1546:   Change, Workshop, Lecture Notes in Artificial Intelligence, Volume 465},

1547:   Konstanz, FRG, October 1989. Springer Verlag.

1548: 

1549: \bibitem{Mak:giveup}

1550: David Makinson.

1551: \newblock How to give it up: a survey of some recent work on formal aspects of

1552:   the logic of theory change.

1553: \newblock {\em Synthese}, 62:347--363, 1985.

1554: 

1555: \bibitem{Mak:recovery}

1556: David Makinson.

1557: \newblock On the status of the postulate of recovery in the logic of theory

1558:   change.

1559: \newblock {\em The Journal of Philosophical Logic}, 16:383--394, 1987.

1560: 

1561: \bibitem{Mak:imposs}

1562: David Makinson.

1563: \newblock The {G}\"{a}rdenfors impossibility theorem in nonmonotonic contexts.

1564: \newblock {\em Studia Logica}, 49:1--6, 1990.

1565: 

1566: \bibitem{Rott:92}

1567: Hans Rott.

1568: \newblock Preferential belief change using generalized epistemic entrenchment.

1569: \newblock {\em Journal of Logic, Language and Information}, 1:45--78, 1992.

1570: 

1571: \bibitem{Segerberg:impossibility}

1572: Krister Segerberg.

1573: \newblock A note on the impossibility theorem of {G}\"{a}rdenfors.

1574: \newblock {\em No\^{u}s}, 23:351--354, 1989.

1575: 

1576: \bibitem{Spohn:87}

1577: W.~Spohn.

1578: \newblock Ordinal conditional functions: A dynamic theory of epistemic states.

1579: \newblock In W.~Harper and B.~Skyrms, editors, {\em Causation in Decision,

1580:   Belief Change and Statistics}, volume volume 2, pages 105--134. D. Reidel,

1581:   Dordrecht, 1987.

1582: 

1583: \bibitem{MaryAnneW:trans}

1584: Mary-Anne Williams.

1585: \newblock Transmutations of knowledge systems.

1586: \newblock In Jon Doyle, Erik Sandewall, and Pietro Torasso, editors, {\em

1587:   Proceedings of the fourth International Conference on Principles of Knowledge

1588:   Representation and Reasoning}, pages 619--629. Morgan Kaufmann, Bonn,

1589:   Germany, May 1994.

1590: 

1591: \end{thebibliography}

1592: 

1593: 

1594: \end{document}

1595: \bye

1596: